Datasets:

kilian-group
/

arxiv-classifier

Dataset card Data Studio Files Files and versions Community

Subset (4)

default · ~99.6k rows (showing the first 120k)

Split (2)

train · ~72.4k rows (showing the first 93.1k)

paper_id stringlengths 9 16	version stringclasses 26 values	yymm stringclasses 311 values	created timestamp[s]	title stringlengths 6 335	secondary_subfield sequencelengths 1 8	abstract stringlengths 25 3.93k	primary_subfield stringclasses 124 values	field stringclasses 20 values	fulltext stringlengths 0 2.84M
1404.4649	2	1404	2015-04-03T22:13:44	Zero cycles on singular varieties and their desingularisations	[ "math.AG", "math.KT" ]	We use pro cdh-descent of $K$-theory to study the relationship between the zero cycles on a singular variety $X$ and those on its desingularisation $X'$. We prove many cases of a conjecture of S. Bloch and V. Srinivas, and relate the Chow groups of $X$ to the Kerz--Saito Chow group with modulus of $X'$ relative to its exceptional fibre.	math.AG	math	Zero cycles on singular varieties and their desingularisations Matthew Morrow Abstract We use pro cdh-descent of K-theory to study the relationship between the zero cycles on a singular variety X and those on its desingularisation X ′. We prove many cases of a conjecture of S. Bloch and V. Srinivas, and relate the Chow groups of X to the Kerz -- Saito Chow group with modulus of X ′ relative to its exceptional fibre. 0 Introduction Let X ′ → X be a desingularisation of a d-dimensional, integral variety over a field k, with exceptional fibre E ֒→ X. Letting rE denote the rth infinitesimal thickening of E, we denote by F dK0(X ′, rE) the subgroup of the relative K-group K0(X ′, rE) generated by the cycle classes of closed points of X ′ \ E, for each r ≥ 1. This inverse system F dK0(X ′, E) ←− F dK0(X ′, 2E) ←− F dK0(X ′, 3E) ←− · · · was first studied by S. Bloch and V. Srinivas [16], in the case of normal surfaces, as a means of relating zero cycles on the singular variety X to zero cycles on the smooth variety X ′. They conjectured [pg. 6, op. cit.] in 1985 that this inverse system would ≃→ F dK0(X ′, (r − 1)E) for r ≫ 1, with stable eventually stabilise, i.e., F dK0(X ′, rE) value equal to F dK0(X), the subgroup of K0(X) generated by cycle classes of smooth, closed points of X. The Bloch -- Srinivas conjecture was proved for normal surfaces by A. Krishna and Srinivas [9, Thm. 1.1], and later extended to higher dimensional, Cohen -- Macaulay varieties with isolated singularities in characteristic zero by Krishna [6, Thm. 1.1] [7, Thm. 1.2]. The conjecture has not been previously verified in any case of non-isolated singularities, nor for any higher dimensional varieties in finite characteristic. The primary goal of this paper is to prove the following cases of the Bloch -- Srinivas conjecture for varieties which are regular in codimension one: Theorem 0.1. Let π : X ′ → X be a desingularisation of a d-dimensional, quasi- projective, integral variety X over an infinite, perfect field k which is assumed to have strong resolution of singularities. Let E ֒→ X be a closed embedding covering the exceptional fibre, and assume that codim(X, π(E)) ≥ 2. Then the associated Bloch -- Srinivas conjecture is (i) true up to (d − 1)!-torsion; Matthew Morrow (ii) true if X is projective, k = kalg, and char k = 0; (iii) true if X is projective, k = kalg, and d ≤ char k 6= 0; (iv) true if X is affine and k = kalg; (v) true "up to a finite group" if k = kalg and Xsing is contained in an affine open of X; (vi) true if π(E) is finite; (vii) true if the cycle class map CH0(X) → F dK0(X) is an isomorphism. The group CH0(X) appearing in part (vii) of Theorem 0.1 is the Levine -- Weibel Chow group of zero cycles of the singular variety X [10, 12]; it will be reviewed in Section 1.1. Part (iv) of the Theorem, combined with arguments of Krishna [7] and R. Murthy [15], has concrete applications to Chow groups of cones and to the structure of modules and ideals of graded algebras; see Theorem 1.17 and Corollaries 1.18 and 1.19. This paper is intended partly to justify the author's pro cdh-descent theorem for K- theory [13]; indeed, the results of Theorem 0.1 are obtained in Section 1.2 as corollaries of the following general result, which itself is an immediate consequence of pro cdh- descent: Theorem 0.2. Let π : X ′ → X be a desingularisation of a d-dimensional, quasi- projective, integral variety over an infinite, perfect field k which is assumed to have strong resolution of singularities. Let E ֒→ X be a closed embedding covering the exceptional fibre. Then: (i) There exists a unique homomorphism BSr : F dK0(X ′, rE) → F dK0(X) for r ≫ 1 which is compatible with cycle classes of closed points. (ii) The associated Bloch -- Srinivas conjecture is true if and only if the canonical map F dK0(X, rY ) → F dK0(X) is an isomorphism for r ≫ 1, where Y := π(E)red. Section 2 concerns Chow groups of zero cycles with modulus. If X is a smooth, projective variety over a field k and D is an effective divisor on X, then the Chow group with modulus CH0(X; D) is defined to be the free abelian group on the closed points of X \ D, modulo rational equivalence coming from closed curves C which are not contained in D and rational functions f ∈ k(C)× which are ≡ 1 mod D. This Chow group is central in M. Kerz and S. Saito's [5] higher dimensional class field theory. It is natural to formulate an analogue of the Bloch -- Srinivas conjecture for the Chow groups with modulus given by successive thickenings of the exceptional fibre of a desingularisation. We will explain this further in Section 2, where we prove it in the following cases: Theorem 0.3. Let π : X ′ → X be a desingularisation of a d-dimensional, quasi- projective, integral variety over an algebraically closed field k which is assumed to have strong resolution of singularities. Let D be an effective Cartier divisor on X covering the exceptional fibre, and assume that codim(X, π(D)) ≥ 2. 2 Zero cycles on singular varieties Then the inverse system CH0(X ′; D) ←− CH0(X ′; 2D) ←− CH0(X ′; 3D) ←− · · · eventually stabilises with stable value equal to CH0(X), assuming that either (i) X is projective and char k = 0; or (ii) X is projective and d ≤ char k 6= 0; or (iii) X is affine. Whenever the assertions of Theorem 0.3 can be proved for a singular, projective variety X over a finite field (e.g., for surfaces, as we shall see in Remark 2.8), it has applications to the class field theory of X; in particular, it shows that there is a reciprocity isomorphism of finite groups CH0(X)0 ≃→ πab 1 (Xreg)0. See Remark 2.7 for further details. We prove Theorem 0.3 by reducing it to the analogous assertion in K-theory, which is precisely the Bloch -- Srinivas conjecture, and then applying Theorem 0.1. This reduction is through the construction of a new cycle class homomorphism CH0(X; D) −→ F dK0(X, D), which is valid for any effective Cartier divisor D on a smooth variety X. This also allows us to prove the following result, which appears related to a special case of a conjecture of Kerz and Saito [5, Qu. V]: Theorem 0.4. With notation and assumptions as in Theorem 0.3, the cycle class homomorphism CH0(X ′; rD) −→ F dK0(X ′; rD) is an isomorphism for r ≫ 1. Notation, conventions, etc. A field k will be called good if and only if it is infinite, perfect, and has strong resolution of singularities, e.g., char k = 0 suffices. A k-variety means simply a finite type k- scheme; further assumptions will be specified when required, and the reference to k with occasionally be omitted. Our conventions about "desingularisations" can be found at the start of Section 1.2. A curve over k is a one-dimensional, integral k-variety. Given a closed point x ∈ C0, there is an associated order function ordx : k(X)× → Z characterised by the property that ordx(t) = lengthOC,x(OC,x/tOC,x) for any non-zero t ∈ OC,x; when C is smooth ordx is the usual valuation associated to x. An effective divisor D on X is by definition a closed subscheme whose defining sheaf of ideals OX (−D) is an invertible OX-module, or, equivalently, is locally defined by a single non-zero-divisor; its associated support is denoted by D, but we write X \ D in place of X \ D for simplicity. 3 Matthew Morrow Given a closed embedding Y = Spec OX /I ֒→ X, its rth infinitesimal thickening is denoted by rY = Spec OX /I r. A pro abelian group {Ar}r is an inverse system of abelian groups, with morphisms given by the rule HomPro Ab({Ar}r, {Bs}s) := lim←− s lim−→ r HomAb(Ar, Bs). The category of pro abelian groups is abelian; we refer to [1, App.] for more details. Acknowledgments Section 1 would not have been possible without discussions with V. Srinivas and M. Levine about zero cycles. Section 2 was inspired by conversations with F. Binda and S. Saito at the ´Etale and motivic homotopy theory workshop in Heidelberg, 24 -- 28 March 2014, and I thank A. Schmidt and J. Stix for organising such a pleasant event. 1 Zero cycles of desingularisations In this section we prove cases of the Bloch -- Srinivas conjecture relating zero cycles on a singular variety to those on its desingularisation. There will be an important distinction between closed subsets S ⊆ X and closed subschemes Y ֒→ X; in an attempt to keep this clear we will use the differentiating notation ⊆ and ֒→ just indicated. Any closed subscheme Y ֒→ X has an associated support Y ⊆ X, though we will continue to write X \ Y rather than X \ Y for the associated open complement, and any closed subset S ⊆ X has an associated reduced closed subscheme Sred ֒→ X. The singular locus of X is denoted by Xsing ⊆ X. 1.1 Review of the Levine -- Weibel Chow group We begin by reviewing the Levine -- Weibel Chow group of zero cycles [10, 12], restricting to the situation that the singularities of X are in codimension ≥ 2, since this is sufficient for our applications. Unless specified otherwise, k is an arbitrary field. Definition 1.1. Let X be an integral k-variety which is regular in codimension one, and S ⊆ X any closed subset containing Xsing. Then the associated Levine -- Weibel Chow group of zero cycles is CH0(X; S) := where (f )C :=Px∈C0 free abelian group on closed points of X \ S h(f )C : C ֒→ X a curve not meeting S, and f ∈ k(C)×i ordx(f ) x as usual. In particular, CH0(X) := CH0(X; Xsing). Remark 1.2. Several remarks should be made: (i) The group CH0(X; S) we have just defined can actually only reasonably be called the Levine -- Weibel Chow group of zero cycles if we assume that codim(X, S) ≥ 2. But it is convenient to introduce the notation in slightly greater generality since it will be useful in Section 2. 4 Zero cycles on singular varieties (ii) An inclusion of closed subsets S ⊆ S′ of X, both containing Xsing, induces a canonical surjection CH0(X; S′) ։ CH(X; S). This surjection is an isomorphism if X is quasi-projective and S, S′ have codimension ≥ 2, by a moving lemma [12, pg. 113]. (iii) Suppose that X is a smooth k-variety and that S ⊆ X is a closed subset. Then there is a canonical surjection CH0(X; S) ։ CH0(X; ∅) = CH0(X), which will be an isomorphism if S has codimension ≥ 2 and X is quasi-projective, by the aforementioned moving lemma. (iv) Suppose that X ′ → X is a proper morphism which restricts to an isomorphism X ′ \ S′ ≃→ X \ S for some closed subsets S ⊆ X, S′ ⊆ X ′ containing the singular loci. Then the induced map CH0(X; S) → CH0(X ′; S′) is an isomorphism. In- deed, both sides are generated by the closed points of X ′ \ S′ = X \ S, and closed curves on X not meeting S correspond to closed curves on X ′ not meeting S′. To review the relationship between CH0(X) and K-theory, we must first explain the cycle class map. Let X be a k-variety, and i : Y ֒→ X a fixed closed subscheme. If j : C ֒→ X is a closed subscheme with image disjoint from both Y and Xsing, then j is of finite Tor dimension since it factors as C ֒→ Xreg → X, and it is moreover proper; thus the pushforward map j∗ : K(C) → K(X) on the K-theory spectra is well-defined. Moreover, the projection formula [19, Prop. 3.18] associated to the pullback diagram ∅ Y C j / X i −→ K(Y ) is null-homotopic, and thus shows that the composition K(C) there is an induced pushforward j∗ : K(C) → K(X, Y ). The cycle class of C in K0(X, Y ) is defined to be j∗−→ K(X) i∗ [C] := j∗([OC ]) ∈ K0(X, Y ). Although this appears to depend a priori on a chosen null-homotopy, it was shown by K. Coombes [4] that the "obvious choices of homotopies" yield a class which is functorial with respect to both X and Y , and so we will follow Coombes' choices. A codimension filtration on K0(X, Y ) is now defined by F pK0(X, Y ) := h[C] : C ֒→ X an integral closed subscheme of X of codim ≥ p disjoint from Y and Xsingi In particular, F dK0(X, Y ) is the subgroup of K0(X, Y ) generated by the cycle classes of smooth, closed points of X \ Y . The following is standard: Lemma 1.3. Let notation be as immediately above. If j : C ֒→ X is a closed embedding ordx(f )[x] = 0 of a curve into X not meeting Y or Xsing, and f ∈ k(C)×, thenPx∈C0 in K0(X, Y ). 5 / / / Matthew Morrow ordx(f )[x] = j∗([OC ] − [f OC]) = j∗(0) = 0 . Proof. One hasPx∈C0 Now suppose that X is a d-dimensional, integral k-variety which is regular in codimension one, let Y ֒→ X be a closed subscheme, and let S ⊆ X be a closed subset containing both Y and Xsing. It follows from Lemma 1.3 that the cycle class homomorphism CH0(X; S) −→ F dK0(X, Y ), x 7−→ [x] is well-defined. homomorphism In particular, taking S = Xsing and Y = ∅ yields the cycle class [ ] : CH0(X) −→ F dK0(X), which is evidently surjective. Moreover, as part of a general Riemann -- Roch theory, M. Levine [11, 10] constructed a Chern class ch0 : F dK0(X) → CH0(X) such that the ] are both multiplcation by (−1)d−1(d − 1)!. In compositions [ particular, [ ] : CH0(X) → F dK0(X) is an isomorphism if d = 2. ] ◦ ch0 and ch0 ◦ [ We complete our review of the Levine -- Weibel Chow group of zero cycles by pre- senting the higher dimensional cases in which the cycle class homomorphism can be shown to be an isomorphism: Theorem 1.4 (Barbieri Viale, Levine, Srinivas). Let X be a d-dimensional, integral, quasi-projective variety over an algebraically closed field which is regular in codimension one. Then the cycle class homomorphism CH0(X) → F dK0(X) is (i) an isomorphism if X is projective and char k = 0; (ii) an isomorphism if X is projective and d ≤ char k 6= 0; (iii) an isomorphism if X is affine and char k is arbitrary; (iv) a surjection with finite kernel if Xsing is contained in an affine open subscheme of X and char k = 0; (v) a surjection with finite kernel if Xsing is contained in an affine open subscheme of X and d ≤ char k 6= 0; Proof. Thanks to the existence of Levine's Chern class ch0, it is enough to check that CH0(X) has no (d − 1)!-torsion in cases (i) -- (ii), that it has only a finite amount of (d − 1)!-torsion in cases (iv) -- (v), and that it has no torsion in case (iii). Then (i) and (ii) are [10, Thm. 3.2], while (iv) and (v) are [2, Thm. A]. Finally, (iii) in characteristic zero (and when d ≤ char k 6= 0) is [10, Corol. 2.7], and so it remains only to deal with the following case: assuming that X is an integral, affine variety which is regular in codimension one, over an algebraically closed field of finite characteristic, [18], and so it remains only to check that CH0(X) we must show that CH0(X) is torsion-free. This is true for the normalisation eX by ≃→ CH0(eX). But since X is assumed to be regular in codimension one, there are closed subsets S ⊆ X, S′ ⊆ eX (given by the conductor ideal, for example) of codimension ≥ 2, containing the singular loci, and 6 Zero cycles on singular varieties such that the morphism eX → X restricts to an isomorphism eX \ S′ ≃→ X \ S. Then, in the commutative diagram CH0(eX; S′) / CH0(eX) CH0(X; S) CH0(X) the horizontal arrows are isomorphisms by Remark 1.2(ii), while the left vertical arrow is an isomorphism by Remark 1.2(iv). Hence the right vertical arrow is an isomorphism, as required. 1.2 The Bloch -- Srinivas conjecture Before we can carefully state the Bloch -- Srinivas conjecture we must first fix some terminology concerning desingularisations. Given an integral variety X, a desingu- larisation is any proper, birational morphism π : X ′ → X where X ′ is smooth; in particular, we allow the desingularisation to change the smooth locus of X, though it is not clear if this is ever important in practice. There exists a smallest closed subset ≃→ X \ S, and π−1(S) is known as the S ⊆ X with the property that X ′ \ π−1(S) exceptional set of the resolution; setting E := π−1(S)red yields the exceptional fibre E ֒→ X ′. Corollaries 1.10 -- 1.15 will require that π(E) has codimension ≥ 2 in X, which in particular implies that X is regular in codimension one. If X ′ → X is a desingularisation of an integral variety X, with exceptional fibre E ֒→ X ′, then Bloch and Srinivas [16, pg. 6] made the following conjecture in 1985: Conjecture 1.5 (Bloch -- Srinivas). The inverse system F dK0(X ′, E) ←− F dK0(X ′, 2E) ←− F dK0(X ′, 3E) ←− · · · stabilises, with stable value F dK0(X). Remark 1.6. To be precise, Bloch and Srinivas stated their conjecture in the case of a normal surface X over an algebraically closed field, assuming that the desingu- larisation did not alter the smooth locus of X. If Conjecture 1.5 is false because it has been formulated in excessive generality, it is the author's fault. In fact, we will consider Conjecture 1.5 in greater generality still, by replacing the exceptional fibre E by any reduced closed subscheme E ֒→ X ′ whose support contains the exceptional set (henceforth "covers the exceptional set"). We interpret part of the Bloch -- Srinivas conjecture as an implicit statement that there exists a cycle class homomorphism BSr : F dK0(X ′, rE) −→ F dK0(X) for r ≫ 1 which is compatible with cycle classes of closed points x ∈ X ′ \ E, i.e., BSr([x]) = [x]. Such a map BSr is unique if it exists. 7 / / / O O O O Matthew Morrow Our main technical theorem, which is an immediate consequence of the author's pro cdh-descent theorem for K-theory [13], proves the existence of the maps BSr in full generality, and reduces the Bloch -- Srinivas conjecture to the study of the K-theory of X: Theorem 1.7. Let X be a d-dimensional, integral variety over a good field k; let π : X ′ → X be a desingularisation, E ֒→ X ′ any reduced closed subscheme covering the exceptional set, and set Y := π(E)red. Then: (i) For r ≫ 1, the canonical map F dK0(X, rY ) → F dK0(X) factors through the surjection F dK0(X, rY ) → F dK0(X ′, rE), i.e., there exists a commutative dia- gram F dK0(X ′, rE) ▼ ▼ ▼ ▼ ∃ BSr ▼ ▼ ▼ / F dK0(X ′) F dK0(X, rY ) ▼ &▼ / F dK0(X) (ii) The following are equivalent: (a) The associated Bloch -- Srinivas conjecture is true, i.e., BSr is an isomor- phism for r ≫ 1. (b) The canonical map F dK0(X, rY ) → F dK0(X) is an isomorphism for r ≫ 1. (c) The canonical map F dK0(X, rY ) → F dK0(X) is an isomorphism for all r ≥ 1. Proof. There is an abstract blow-up square Y ′ Y X ′ π / X where Y ′ := X ′ ×X Y ; note that Y ′ is a nilpotent thickening of E. By pro cdh-descent for K-theory [13, Thm. 0.1] (it is here that the field k is required to be good), the canonical homomorphism of pro abelian groups {K0(X, rY )}r −→ {K0(X ′, rY ′)}r ∼= {K0(X ′, rE)}r is an isomorphism. Restricting to the codimension filtration we deduce that the ho- momorphism {F dK0(X, rY )}r −→ {F dK0(X ′, rE)}r (†) is injective; but each map F dK0(X, rY ) → F dK0(X ′, rE) is evidently surjective, since both sides are generated by the closed points of X \ Y = X ′ \ E. Thus (†) is an isomorphism. 8 & / O O O O / O O / / / Zero cycles on singular varieties By definition of an isomorphism of pro abelian groups, this implies that for any s ≥ 1 there exists r ≥ s and a homomorphism F dK0(X ′, rE′) → F dK0(X, sY ) making the diagram commute: F dK0(X ′, rE) P P P ∃ P P P F dK0(X, rY ) / F dK0(X, sY ) (P Note that the vertical and horizontal arrows are surjective, since the groups are gen- erated by the closed points of X \ Y = X ′ \ E. This diagram shows that the canon- ical map F dK0(X, rY ) → F dK0(X) factors through the surjection F dK0(X, rY ) → F dK0(X ′, rE), proving (i). This gives a commutative diagram F dK0(X ′, rE) BSr (PPPPPPPPPPPPPP F dK0(X, rY ) / F dK0(X, sY ) / F dK0(X) from which a simple diagram chase yields the following implications (valid for any s ≥ 1 and r ≫ s): F dK0(X, rY ) → F dK0(X) is an isomorphism =⇒ BSr is an isomorphism. BSr is an isomorphism =⇒ F dK0(X, sY ) → F dK0(X) is an isomorphism. The equivalence of (a) -- (c) follow, completing the proof. Remark 1.8. Suppose that the desingularisation X ′ → X does not change the smooth locus of X and that E is equal to the exceptional fibre (this is probably the most important case of the conjecture). Then Theorem 1.7 states that the associated Bloch -- ≃→ F dK0(X) for r ≫ 1, where Srinivas conjecture is true if and only if F dK0(X, rY ) Y = (Xsing)red. In particular, under these additional hypotheses on X ′ and E we see that the Bloch -- Srinivas conjecture depends only on X, and not on the chosen desingularisation. Even in the case of arbitrary desingularisations and general E covering the exceptional set, Theorem 1.7 shows that the associated Bloch -- Srinivas conjecture depends only on X and π(E). Remark 1.9. The proof of Theorem 1.7 shows the following: the inverse system F dK0(X ′, rE), r ≥ 1, stabilises if and only if the inverse system F dK0(X, rY ), r ≥ 1, stabilises, in which case the canonical map F dK0(X, rY ) → F dK0(X ′, rE) is an iso- morphism for r ≫ 1. 9 ( O O O O / / / ( ( ( O O O O / / / / Matthew Morrow The following corollary recovers all previously known cases of the Bloch -- Srinivas conjecture (normal surfaces [9, Thm. 1.1]; Cohen -- Macaulay varieties with isolated singularities in characteristic zero [6, Thm. 1.1] [7, Thm. 1.2]; note that in these cases one can use the reduction ideal trick of Weibel [20] to avoid assuming that k has resolution of singularities, c.f., Remark 2.8): Corollary 1.10. Let X be a d-dimensional, integral variety over a good field k; let π : X ′ → X be a desingularisation, and E ֒→ X ′ any reduced closed subscheme covering the exceptional set. Assume π(E) is finite and d ≥ 2. Then the associated Bloch -- Srinivas conjecture is true. Proof. Set Y := π(E)red. According to Theorem 1.7, it is necessary and sufficient to show that the canonical map F dK0(X, rY ) → F dK0(X) is an isomorphism for all r ≥ 1. But this follows from [6, Lem. 3.1] since rY is zero dimensional. The next corollary proves the Bloch -- Srinivas conjecture under the assumption that the cycle class homomorphism CH0(X) → F dK0(X) is an isomorphism: Corollary 1.11. Let X be a d-dimensional, integral, quasi-projective variety over a good field k; let π : X ′ → X be a desingularisation, and E ֒→ X ′ any reduced closed subscheme covering the exceptional set. Assume codim(X, π(E)) ≥ 2 and that the cycle class map CH0(X) → F dK0(X) is an isomorphism. Then the associated Bloch -- Srinivas conjecture is true. Proof. Set Y = π(E)red. According to Theorem 1.7, it is necessary and sufficient to show that the canonical map F dK0(X, rY ) → F dK0(X) is an isomorphism for all r ≥ 1. To prove this we consider the commutative diagram F dK0(X, rY ) / F dK0(X) CH0(X; Y ) CH0(X) The right vertical arrow is an isomorphism by assumption, the bottom horizontal arrow is an isomorphism by Remark 1.2(ii), and the left vertical arrow is a surjection since the domain and codomain are generated by the closed points of X \ Y . It follows that the top horizontal arrow (and left vertical arrow -- we will need this in the proof of Theorem 2.5) is an isomorphism, as desired. In particular, we have proved the Bloch -- Srinivas conjecture for projective varieties over an algebraically closed field of characteristic zero which are regular in codimension one: Corollary 1.12. Let X be a d-dimensional, integral variety over an algebraically closed field k which has strong resolution of singularities; let π : X ′ → X be a desingularisa- tion, and E ֒→ X ′ any reduced closed subscheme covering the exceptional set. Assume codim(X, π(E)) ≥ 2 and that one of the following is true: 10 / / / O O O O Zero cycles on singular varieties (i) X is projective and char k = 0; or (ii) X is projective and d ≤ char k 6= 0; or (iii) X is affine and char k is arbitrary. Then the associated Bloch -- Srinivas conjecture is true. Proof. This follows from Corollary 1.11 and the results of Levine and Srinivas recalled in Theorem 1.4. Remark 1.13. It seems plausible that some descent or base change technique should eliminate the requirement in Corollary 1.12 that k be algebraically closed. We can also solve the Bloch -- Srinivas conjecture up to (d − 1)!-torsion whenever X is regular in codimension one: Corollary 1.14. Let X be a d-dimensional, integral, quasi-projective variety over a good field k; let π : X ′ → X be a desingularisation, and E ֒→ X ′ any reduced closed subscheme covering the exceptional set. Assume codim(X, π(E)) ≥ 2. Then the associated Bloch -- Srinivas conjecture is true up to (d − 1)!-torsion, i.e., the maps BSr : F dK0(X ′, rE) ⊗ Z[ 1 (d−1)! ] −→ F dK0(X) ⊗ Z[ 1 (d−1)! ] are isomorphisms for r ≫ 1. Proof. Set Y = π(E)red. By a trivial modification of Theorem 1.7, it is necessary and sufficient to show that the canonical map F dK0(X, rY ) → F dK0(X) is an isomorphism for all r ≥ 1 after inverting (d − 1)!. This follows exactly as in Corollary 1.11, since the cycle class map CH0(X) → F dK0(X) is an isomorphism after inverting (d − 1)!, thanks to the existence of Levine Chern class ch0 : F dK0(X) → CH0(X). The next result solves the Bloch -- Srinivas conjecture up to a finite group when the singular locus Xsing has codimension ≥ 2 and is contained in an affine open of X. Note that the "obvious" cases in which this happens, namely when Xsing is finite or X itself is affine, are already largely covered by Corollaries 1.10 and 1.12(iii) respectively: Corollary 1.15. Let X be a d-dimensional, integral, quasi-projective variety over an algebraically closed field k which has strong resolution of singularities; let π : X ′ → X be a desingularisation, and E ֒→ X ′ any reduced closed subscheme covering the exceptional set. Assume codim(X, π(E)) ≥ 2, that Xsing is contained in an affine open of X, and moreover that d ≤ char k if char k 6= 0. Then the maps BSr : F dK0(X ′, rE) −→ F dK0(X) are surjective with finite kernel for r ≫ 1, and the inverse system F dK0(X ′, rE), r ≥ 1, stabilises. 11 Matthew Morrow Proof. We concatenate commutative diagrams we have already considered in Theorem 1.7 and Corollary 1.11: F dK0(X ′, rE) BSr F dK0(X, rY ) / F dK0(X) CH0(X; Y ) ∼= CH0(X) The left vertical arrows are surjective since the groups are generated by the closed points of X \ Y = X ′ \ E; the bottom horizontal arrow is an isomorphism by Remark 1.2(ii); the right vertical arrow is surjective with finite kernel Λ by the result of Barbieri Viale recalled in Theorem 1.4. A simple diagram chase shows that BSr is surjective and that its kernel Λr is naturally a quotient of Λ. Since Λ is finite, this tower of quotients Λr must eventually stabilise, completing the proof. Remark 1.16. We finish our discussion of the Bloch -- Srinivas conjecture with a remark about SK1. Let π : X ′ → X, E, Y , k be as in the statement of Theorem 1.7, and assume X is quasi-projective and codim(X, Y ) ≥ 2. The maps F dK0(X, rY ) → F dK0(X) are surjective for all r ≥ 1 (by Remark 1.2(ii) and existence of the cycle class maps); hence we may add (b′) The canonical map F dK0(X, rY ) → F dK0(X) is injective for r ≫ 1. to the list of equivalent conditions in Theorem 1.7(ii). Next, it follows from [6, Lem. 3.1] that (b′) (hence the associated Bloch -- Srinivas conjecture) would follow from showing that ∂(SK1(rY )) = 0, where ∂ : K1(rY ) → K0(X, rY ) is the boundary map and SK1(rY ) := Ker(K1(rY ) ։ H 0(rY, O× rY )); equiv- alently, it is enough to show that SK1(X) → SK1(rY ) is surjective. Using the argu- ments of Theorem 1.7 it would even be enough to show, for each r ≫ 1, that Im(SK1(sY ) → SK1(rY )) ⊆ Im(SK1(X) → SK1(rY )) for some s ≥ r. It is not clear whether one should expect this to be true. We finish the section with some consequence of the Bloch -- Srinivas conjecture. The following result about Chow groups of cones was conjectured by Srinivas [17, §3] in 1987; it was proved by Krishna [7, Thm. 1.5] under the assumption that the cone X was normal and Cohen -- Macaulay, and we will combine his argument with Theorem 1.7 to establish the result in full generality; due to the failure of Kodaira vanishing in finite characteristic we must restrict to characteristic zero: Theorem 1.17. Let Y ֒→ PN k be a d-dimensional, smooth, projective variety over an algebraically closed field k of characteristic zero; assume d > 0 and H d(Y, OY (1)) = 0, and let X be the affine cone over Y . Then CH0(X) = 0. 12 O O O O / / / O O O O O O Zero cycles on singular varieties Proof. We may resolve X, which has a unique singular point, to obtain X ′ which is a line bundle over over Y , of which the zero section is the exceptional fibre of the resolu- tion X ′ → X. By Corollary 1.10 or 1.12(iii), we know that CH0(X) ∼= F d+1K0(X ′, rY ) for r ≫ 1; moreover, CH0(X ′) surjects onto F dK0(X ′), and CH0(X ′) = 0 since X ′ is a line bundle, so F dK0(X ′) = 0. So it is enough to show that the canonical map F d+1K0(X ′, rY ) → F d+1K0(X ′) is an isomorphism. According to Krishna's proof of [7, Cor. 8.5], this would follows from knowing that: (i) H d(X ′, Kd,X ′) ⊗ k× −→ H d(Y, Kd,Y ) ⊗ k× is surjective; and (ii) H d(cid:18)rY, (rY,Y ) Ωd dΩd−1 (rY,Y )(cid:19) = 0 for r ≫ 1. Condition (i) is satisfied since the zero section Y ֒→ X ′ is split by the line bundle structure map X ′ → Y . Condition (ii) is deduced from the Akizuki -- Nakano vanishing theorem, as explained in Lem. 9.1 and the proof of Thm. 1.5 in [7]. Corollary 1.18. Let Y, k be as in the previous theorem, and let A be its homogeneous coordinate ring. Then every projective module over A of rank at least d has a free direct summand of rank one. Proof. This follows from Theorem 1.17 using a result of R. Murthy [15, Cor. 3.9]. Corollary 1.19. Let k be an algebraically closed field of characteristic zero, and f ∈ k[t] := k[t0, . . . , td] a homogenous polynomial of degree at most d + 1 which defines a smooth hypersurface in Pd k. Then every smooth closed point of Spec k[t]/hf i is a complete intersection. In other words, if m is any maximal ideal of k[t] containing f other than the origin, then m = hf, f1, . . . , fdi for some f1, . . . , fd ∈ k[t]. Proof. This also follows from Theorem 1.17 thanks to Murthy [15, Thm. 4.4]. 2 Chow groups with modulus If X is a smooth variety over a field k, and D is an effective divisor on X, then the Chow group CH0(X; D) from Definition 1.1 may be a rather coarse invariant, as there may not be enough curves on X avoiding the codimension-one subset D. Of greater interest is CH0(X; D), the Chow group of zero cycles on X with modulus D, which we will define precisely in Definition 2.1; note the notational difference, indicating that CH0(X; D) depends not only on the support of D, but on its schematic, and possibly non-reduced, structure. According to the higher dimensional class field theory of M. Kerz and S. Saito, when k is finite and X is proper over k, the group CH0(X; D) classifies the abelian ´etale covers of X \ D whose ramification is bounded by D; we refer the reader to [5] for details since we will not require any of their results. 13 Matthew Morrow We now turn to definitions, and refer again to [op. cit.] for a more detailed exposi- tion. Let C be a smooth curve over a field k, and D an effective divisor on C; writing D =Px∈D mxx as a Weil divisor, we let k(C)× D := {f ∈ k(C)× : ordx(f − 1) ≥ mx for all x ∈ D} denote the rational functions on C which are ≡ 1 mod D. More generally, if X is a smooth variety over k and D is an effective divisor on X, then for any curve C ֒→ X which is not contained in D we write where φ : eC → C ֒→ X is the resulting map from the normalisation eC to X; evidently D = k(C)× if C does not meet D. The Chow group with modulus is defined as follows: k(C)× k(C)× φ∗D, D := k(eC)× Definition 2.1. Let X be a smooth variety over k, and D an effective divisor on X. Then the associated Chow group of zero cycles of X with modulus D is free abelian group on closed points of X \ D h(f )C : C ֒→ X a curve not contained in D, and f ∈ k(C)× Di CH0(X; D) := where (f )C =Px∈C0 If we were to define ordx(f ) x. k(C)× D :=(k(C)× if C does not meet D, if C meets D, 1 and repeat Definition 2.1 with D in place of D, then the resulting group CH0(X; D) would coincide with that defined in Definition 1.1. Since k(C)× D, we thus obtain a canonical surjection D ⊆ k(C)× CH0(X; D)−→→ CH0(X; D). One sense in which CH0(X; D) is a more refined invariant than CH0(X; D) is that the cycle class homomorphism CH0(X; D) → K0(X, D) of Section 1.1 factors through CH0(X; D). There does not appear to be a proof of this important result in the literature, so we give one here, beginning with a much stronger result in the case of curves: Lemma 2.2. Let C be a smooth curve over a field k, and D an effective divisor on C. Then the canonical map free abelian group on closed points of C \ D −→ K0(C, D), x 7−→ [x] induces an injective cycle class homomorphism CH0(C; D) −→ K0(C, D), which is an isomorphism if D 6= 0 (and has cokernel = Z if D = 0). 14 Zero cycles on singular varieties Proof. The Zariski descent spectral sequence for the K-theory of C relative to D degenerates to short exact sequences, since dim C = 1, yielding in particular 0 −→ H 1(C, K1,(C,D)) −→ K0(C, D) −→ H 0(C, K0,(C,D)) −→ 0. Here Ki,(C,D) is by definition the Zariski sheafification on C of the presheaf U 7→ Ki(U, U ×C D). To describe these terms further we make some standard comments about the long exact sequence of sheaves K2,C → K2,D → K1,(C,D) → K1,C → K1,D → K0,(C,D) → K0,C → K0,D. C and K1,D ∼= O× It follows that K1,(C,D) Firstly, K1,C ∼= O× D, so the map K1,C → K1,D is surjective; moreover, the sheaves K2,C and K2,D are generated by symbols, and so the map K2,C → K2,D (C,D) and that is also surjective. H 0(C, K0,(C,D)) = Ker(H 0(C, K0,C ) → H 0(D, K0,D)). Secondly, K0,C ∼= Z via the rank Z via the rank map. If D 6= 0, we deduce that the map H 0(C, K0,C ) → H 0(D, K0,D) is injective and so H 0(C, K0,(C,D)) = 0; while if D = 0 then evidently H 0(X, K0,(C,D)) = H 0(X, K0,C ) ∼= Z. map, and so H 0(C, K0,C ) ∼= Z; similarly, H 0(D, K0,D) ∼=Lx∈D D) =: O× ∼= Ker(O× C → O× In conclusion, it remains only to construct the cycle class isomorphism CH0(C; D) ≃−→ H 1(C, O× (C,D)). We will do this via a standard Gersten resolution. Given an open subscheme U ⊆ C containing D, let jU : U → C denote the open C,D fits into an exact sequence of (C,D) → jU ∗j∗ U O× inclusion. Then the canonical map O× sheaves 0 −→ O× (C,D) −→ jU ∗j∗ U O× (C,D) (ordx)x −−−−→ Mx∈C\U ix∗Z −→ 0, where ix∗Z is a skyscraper sheaf at the closed point x. This remains exact after taking the filtered colimit over all open U containing D, yielding 0 −→ O× (C,D) −→ k(C)× D (ordx)x −−−−→ Mx∈C0\D ix∗Z −→ 0, where k(C)× a flasque resolution of O× isomorphism D denotes a constant sheaf by abuse of notation. This latter sequence is (C,D), and using it to compute cohomology yields a natural coker(cid:0)k(C)× D (ordx)x −−−−→ Mx∈C0\D Z(cid:1) ≃−→ H 1(C, O× (C,D)). But the left side of this isomorphism is precisely CH0(C; D), thereby completing the proof. 15 Matthew Morrow Proposition 2.3. Let X be a smooth variety over a field k, and D an effective divisor on X. Then the canonical map free abelian group on closed points of X \ D −→ K0(X, D), x 7−→ [x] descends to a cycle class homomorphism CH0(X; D) −→ K0(X, D). Proof. We must show that if C ֒→ X is a curve not contained in D and f ∈ k(C)× D, ordx(f )[x] = 0 in K0(X, D). We will deduce this from Lemma 2.2 once then Px∈C0 Let φ : eC → C ֒→ X be the resulting map from the normalisation eC to X, and we have verified a suitable pushforward formalism. consider the following pullback square: j ′ φ∗D φ′ D j / eC φ / X We claim that φ and j are Tor-independent; that is, if y is a closed point of eC such that x := φ(y) lies in D, we must show that Tori OX,x(OD,x, O eC,y) = 0 for all i > 0. But since D is an effective Cartier divisor, there exists a non-zero-divisor t ∈ OX,x such that OD,x = OX,x/tOX,x; thus the only possible non-zero higher Tor is Tor1, which equals the φ∗(t)-torsion of O eC,y; this could only be non-zero if φ∗(t) = 0 in O eC,y, but this would contradict the condition that C does not lie in D. This proves the desired Tor-independence. Moreover, φ is a finite morphism and X is assumed to be smooth, whence φ is proper and of finite Tor-dimension. Therefore the projection formula [19, Prop. 3.18] (or [4, Thm. 4.4]) states that the diagram K(eC) φ∗ K(X) j ′∗ / K(φ∗D) φ′ ∗ / K(D) j ∗ is well-defined and commutes up to homotopy; so there is an induced pushforward map which by functoriality of pushforwards (as in Section 1.1 we must appeal to [4, §4 -- 5] to know that the obvious choices of homotopies yield a functorial construction) satisfies φ∗ : K(eC, φ∗D) −→ K(X, D), 16 / / / / Zero cycles on singular varieties φ∗[x] = [φ(x)] for any x ∈ eC0. Therefore Xx∈C0 ordx(f )[x] = Xx∈ eC0 = φ∗(cid:0) Xx∈ eC0 = φ∗(0) = 0, ordφ(x)(f )[φ(x)] ordx(f )[x](cid:1) wherePx∈ eC0 ordx(f )[x] ∈ K0(eC, φ∗D) vanishes by Lemma 2.2. Remark 2.4. F. Binda [3] has independently proved Proposition 2.3, as well as con- structing cycle class homomorphisms CH0(X; D; n) → Kn(X, D) for the higher Chow groups with modulus. Let X be a d-dimensional, smooth variety over k. Given effective divisors D′ ≥ D with the same support, the inclusions k(C)× D induce a canonical surjection CH0(X; D′) ։ CH0(X; D). This applies in particular when D′ = rD is a thickening of D. Combining this observation with Proposition 2.3 we obtain a commutative diagram of inverse systems of Chow groups and relative K-groups (recall the definition of F dK0 from Section 1.1) in which all maps are surjective (since every group is generated by the closed points of X \ D): D′ ⊆ k(C)× F dK0(X, D) F dK0(X, 2D) F dK0(X, 3D) F dK0(X, 4D) · · · CH0(X; D) CH0(X; 2D) l❳❳❳❳❳❳❳❳❳❳❳❳❳❳❳❳❳❳❳❳❳❳❳❳❳❳ h◗◗◗◗◗◗◗◗◗◗◗◗◗ CH0(X; 3D) CH0(X; 4D) · · · 3❣❣ 6♠♠♠♠♠♠♠♠♠♠♠♠♠ ❣❣ ❣❣❣❣❣ ❣❣❣❣ CH0(X; D) There are two natural questions to consider concerning this diagram. Firstly, a question seemingly related to a conjecture of Kerz and Saito [5, Qu. V] is whether the cycle class homomorphism {CH0(X; rD)}r −→ {F dK0(X; rD)}r is an isomorphism of pro abelian groups, perhaps at least ignoring (d − 1)!-torsion. Secondly, changing notation, now suppose that X ′ → X is a desingularisation of an integral variety X, whose exceptional fibre is an effective Cartier divisor D. Then, as a Chow-theoretic analogue of the Bloch -- Srinivas conjecture, we ask whether the inverse system CH0(X ′; D) ←− CH0(X ′; 2D) ←− CH0(X ′; 3D) ←− · · · 17 o o o o o o o o o o o o o o o o O O O O o o o o O O O O o o o o O O O O o o o o O O O O o o o o l l l h h h O O O O 6 6 6 3 3 3 Matthew Morrow eventually stabilises, with stable value most likely equal to the Levine -- Weibel Chow group CH0(X) of X. The following theorem simultaneously answers cases of these two questions, working under almost identical hypotheses to Corollary 1.11: Theorem 2.5. Let X be a d-dimensional, integral, quasi-projective variety over a good field k; let π : X ′ → X be a desingularisation, and D any effective Cartier divisor on X whose support contains the exceptional set. Assume codim(X, π(D)) ≥ 2 and that the cycle class map CH0(X) → F dK0(X) is an isomorphism. Then CH0(X) ∼= CH0(X ′; D), and the canonical maps CH0(X ′; D) −→ CH0(X ′; rD) −→ F dK0(X ′; rD) are isomorphisms for r ≫ 1. Proof. Let Y ֒→ X be the reduced closed subscheme with support π(D); this has codimension ≥ 2 and covers Xsing. Consider the following commutative diagram, which exists for any r ≫ 1: CH0(X ′; D) / CH0(X ′; rD) / F dK0(X ′; rD) CH0(X; Y ) / CH0(X) BSr / F dK0(X) The bottom right horizontal arrow is an isomorphism by assumption; the bottom left horizontal arrow is an isomorphism by Remark 1.2(ii); the left vertical arrow is an isomorphism by Remark 1.2(iv); the right vertical arrow is an isomorphism by Corollary 1.11. Since the two top horizontal arrows are surjective, it follows that they are isomorphisms. Corollary 2.6. Let X be a d-dimensional, integral variety over an algebraically closed field k which has strong resolution of singularities; let π : X ′ → X be a desingularisa- tion, and D any effective Cartier divisor on X whose support contains the exceptional set. Assume codim(X, π(D)) ≥ 2 and that one of the following is true: (i) X is projective and char k = 0; or (ii) X is projective and d ≤ char k 6= 0; or (iii) X is affine. Then CH0(X) ∼= CH0(X ′; D), and the canonical maps CH0(X ′; D) −→ CH0(X ′; rD) −→ F dK0(X ′; rD) are isomorphisms for r ≫ 1. Proof. This follows from Theorem 2.5 and the results of Levine and Srinivas recalled in Theorem 1.4. 18 / / O O / / Zero cycles on singular varieties Remark 2.7 (Class field theory of singular varieties). In this remark we explain how the CH0 isomorphism of Theorem 2.5 over a finite field Fq can be interpreted as part of an unramified class field theory for singular, projective varieties. Let X be a projective variety over Fq which is regular in codimension one; suppose that a desingularisation π : X ′ → X exists, that D is an effective Cartier divisor on X whose support contains the exceptional set, and that codim(X, π(D)) ≥ 2. Write U = X ′ \ D = X \ π(D). The Kerz -- Saito class group [5] of U is C(U ) := lim←−r CH0(X ′; rD), and their class field theory provides a reciprocity isomorphism C(U )0 ≃→ πab 1 (U )0, where the super- scripts 0 denote degree-0 subgroups. Assuming that the conclusions of Theorem 2.5 are true in this setting, we deduce that C(U ) = CH0(X ′; rD) ∼= CH0(X) for r ≫ 1. They prove moreover that each group CH0(X ′; rD)0 is finite. In particular, this would prove finiteness of CH0(X)0, which is known in the smooth case thanks to the unramified class field theory of S. Bloch, K. Kato and Saito, et al. It would also yield a reciprocity isomorphism CH0(X)0 ≃−→ πab 1 (U )0, [x] 7→ Frobx However, since the canonical map πab isomorphism, we would obtain in general only a surjective reciprocity map 1 (X) is surjective but generally not an 1 (U ) → πab CH0(X)0 −→ πab 1 (X)0, indicating that the Levine -- Weibel Chow group CH0(X) is not the correct class group for unramified class field theory of a singular variety. Remark 2.8 (The case of surfaces). If X is an integral, projective surface over Fq which is regular in codimension one, then we have actually proved the observations of Remark 2.7 unconditionally: CH0(X) is isomorphic to the Kerz -- Saito class group C(Xreg), its degree-0 subgroup is finite, and there is a reciprocity isomorphism CH0(X)0 ≃−→ πab 1 (Xreg)0 of finite groups. This was brought to the author's attention by [8], in which Krisha reproduced the argument while being unaware of the present paper. To prove this we must only check that Theorem 2.5 is true for surfaces over finite fields. In fact, we will let X be a 2-dimensional, integral, quasi-projective variety over an arbitrary field k which is regular in codimension one. Then X admits a resolution of singularities π : X ′ → X with exceptional set equal to exactly π−1(Xsing); let E := π−1(Xsing)red and Y := (Xsing)red. Then Theorem 1.7 is true for the data X ′ → X, Y , E. Indeed, it is only necessary to establish the isomorphism (†) occurring in the proof, which may be broken into the two isomorphisms {F dK0(X, rY )}r ≃→ {F dK0(eX, eX ×X rY )}r 19 ≃→ {F dK0(X ′, rE)}r, Matthew Morrow where eX → X denotes the normalisation of X. The second of these isomorphisms ≃→ {K0(eX, eX ×X rY )}r, which is a case of the author's is due to Krishna and Srinivas [9, Thm. 1.1]; the first isomorphism follows from the isomorphism {K0(X, rY )}r pro-excision theorem [14, Corol. 0.4 & E.g. 2.5], and the obvious surjectivity just as in the proof of Theorem 1.7. Now assume further (perhaps after blowing-up X ′ at finitely many points) that there is an effective divisor D on X ′ with support π−1(Xsing). Since the cycle class map CH0(X) → F dK0(X) is automatically an isomorphism (as we remarked imme- diately before Theorem 1.4), it follows that the assertions of Theorem 2.5 are also true, as required: CH0(X) ∼= CH0(X ′; D), and the canonical maps CH0(X ′; D) → CH0(X ′; rD) → F dK0(X ′; rD) are isomorphisms for r ≫ 1. References [1] Artin, M., and Mazur, B. Etale homotopy, vol. 100 of Lecture Notes in Math- ematics. Springer-Verlag, Berlin, 1986. Reprint of the 1969 original. [2] Barbieri Viale, L. Zero-cycles on singular varieties: torsion and Bloch's for- mula. J. Pure Appl. Algebra 78, 1 (1992), 1 -- 13. [3] Binda, F. Algebraic cycles with modulus and relative K-theory. Preprint (2014). [4] Coombes, K. R. Relative algebraic K-theory. Invent. Math. 70, 1 (1982/83), 13 -- 25. An appendix. [5] Kerz, M., and Saito, S. Chow group of 0-cycles with modulus and higher dimensional class field theory. arXiv:1304.4400 (2013). [6] Krishna, A. Zero cycles on a threefold with isolated singularities. J. Reine Angew. Math. 594 (2006), 93 -- 115. [7] Krishna, A. An Artin-Rees theorem in K-theory and applications to zero cycles. J. Algebraic Geom. 19, 3 (2010), 555 -- 598. [8] Krishna, A. 0-cycles on singular schemes and class field theory. arXiv:1502.01515 (2015). [9] Krishna, A., and Srinivas, V. Zero-cycles and K-theory on normal surfaces. Ann. of Math. (2) 156, 1 (2002), 155 -- 195. [10] Levine, M. Zero-cycles and K-theory on singular varieties. In Algebraic geometry, Bowdoin, 1985 (Brunswick, Maine, 1985), vol. 46 of Proc. Sympos. Pure Math. Amer. Math. Soc., Providence, RI, 1987, pp. 451 -- 462. [11] Levine, M. A geometric theory of the chow ring of a singular variety. Unpublished preprint (ca. 1983). 20 Zero cycles on singular varieties [12] Levine, M., and Weibel, C. Zero cycles and complete intersections on singular varieties. J. Reine Angew. Math. 359 (1985), 106 -- 120. [13] Morrow, M. Pro cdh-descent for cyclic homology and K-theory. J. Inst. Math. Jussieu, to appear. [14] Morrow, M. Pro unitality and pro excision in algebraic K-theory and cyclic homology. J. Reine Angew. Math., to appear. [15] Murthy, M. P. Zero cycles and projective modules. Ann. of Math. (2) 140, 2 (1994), 405 -- 434. [16] Srinivas, V. Zero cycles on a singular surface. II. J. Reine Angew. Math. 362 (1985), 4 -- 27. [17] Srinivas, V. Rational equivalence of 0-cycles on normal varieties over C. In Algebraic geometry, Bowdoin, 1985 (Brunswick, Maine, 1985), vol. 46 of Proc. Sympos. Pure Math. Amer. Math. Soc., Providence, RI, 1987, pp. 475 -- 482. [18] Srinivas, V. Torsion 0-cycles on affine varieties in characteristic p. J. Algebra 120, 2 (1989), 428 -- 432. [19] Thomason, R. W., and Trobaugh, T. Higher algebraic K-theory of schemes In The Grothendieck Festschrift, Vol. III, vol. 88 of and of derived categories. Progr. Math. Birkhauser Boston, Boston, MA, 1990, pp. 247 -- 435. [20] Weibel, C. The negative K-theory of normal surfaces. Duke Math. J. 108, 1 (2001), 1 -- 35. Matthew Morrow Mathematisches Institut Universitat Bonn Endenicher Allee 60 53115 Bonn, Germany [email protected] http://www.math.uni-bonn.de/people/morrow/ 21
1810.07850	2	1810	2018-11-11T23:41:15	Central extensions and the classifying spaces of projective linear groups	[ "math.AG" ]	If $G$ is a presheaf of groupoids on a small site, and $A$ is a sheaf of abelian groups, we prove that the sheaf cohomology group $H^2 (BG, A)$ is in bijection with a set of central extensions of $G$ by $A$. We use this result to study the motivic cohomology of the Nisnevich classifying space $BG$, when $G$ is a presheaf of groups on the smooth Nisnevich site over a field, and particularly when $G = PGL_{n}$. Finally, we show that, when $p$ is an odd prime, the Chow ring of the classifying space of $PGL_{p}$ injects into the motivic cohomology of the Nisnevich classifying space $BPGL_{p}$, over any field of characteristic zero containing a primitive $p^{th}$ root of unity.	math.AG	math	CENTRAL EXTENSIONS AND THE CLASSIFYING SPACES OF PROJECTIVE LINEAR GROUPS ALEXANDER ROLLE Introduction The first result of this paper is a generalization of the fact that the set of isomorphism classes of central extensions of a group G by an abelian group A is in bijection with the elements of the cohomology group H 2(G, A). Letting G be a presheaf of groupoids on a small site, and A a sheaf of abelian groups, we use the work of Jardine on the theory of higher stacks to show that there is a bijection H 2(BG, A) ∼= CenExt(G, A) where CenExt(G, A) is an appropriate set of equivalence classes of "central extensions" of G by A. Classes in H 2(BG, A) corresponding to central extensions of sheaves of groups can be interpreted as universal obstruction classes. The motivating example is the extension Gm → Gln → PGln of sheaves of groups on the ´etale site of a regular scheme, and in this case, our results recover the usual obstruction classes in the Brauer group of the base scheme. In the second part of this paper, we consider instead the Nisnevich topology on the site Smk of smooth schemes over a field k: for G a presheaf of groups on this site, we show that Nisnevich PGln-torsors over the Nisnevich classifying space BNisG can be inter- preted as homomorphisms G → PGln. There is a universal obstruction class, which measures when a projective representation factors through a linear representation, in the motivic cohomology group H 3(BNisPGln, Z(1)), by the results of the first section. On the Nisnevich site Smk, there is another notion of classifying space that has been studied in motivic homotopy theory: this is the ´etale classifying space BetG. The relationship between the motivic cohomology of BNisG and that of BetG is quite inter- esting. According to a result of Morel-Voevodsky, the motivic cohomology of the ´etale classifying space of a linear algebraic group G recovers the Chow ring of the classifying ∼= H 2∗(BetG, Z(∗)). In the final result of this space of G, in the sense of Totaro: A∗ G paper, we use work of Vistoli on A∗ PGlp to show that, when p is an odd prime, the canonical homomorphism in motivic cohomology H 2∗(BetPGlp, Z(∗)) → H 2∗(BNisPGlp, Z(∗)) is injective, over any field of characteristic zero containing a primitive pth root of unity. 1 1. Cohomology and central extensions The results of this section require significant parts of the theory of non-abelian coho- mology developed by Jardine [4, Chapter 9], generalizing work of Giraud [2] and Breen [1]. We begin by introducing the necessary parts of this theory. Let A be a sheaf of abelian groups on a site C. There is an associated presheaf of 2- groups Iso(A) defined as follows: for U an object of C, the 1-morphisms of Iso(A)(U ) are the isomorphisms of sheaves of groups on C/U , and the 2-morphisms are given by α : AU → AU Iso(A)(U )(α, α) = A(U ) for all 1-morphisms α ; Iso(A)(U )(α, β) = ∅ for all α 6= β . Define a sheaf of 2-groups A[2] on C such that A[2](U ) has only the identity 1-morphism, and has A(U ) for 2-morphisms. Let f : A[2] → Iso(A) be the map that includes A[2] as a subobject, and define a retraction g : Iso(A) → A[2] by for any 2-morphism h of Iso(A)(U ). g(h : α → α) = h : ∗ → ∗ , The approach to non-abelian cohomology described in [4, Chapter 9] makes heavy use of the Eilenberg-Mac Lane W functor. We will not need the details of the construction, which can be found in [4, Section 9.3] or [7], but we will use a few facts, which we now summarize. Write s0Gpd for the category of groupoids enriched in simplicial sets; then W is a functor s0Gpd → sSet. If G is a groupoid enriched in simplicial sets, [4, Corollary 9.39] says that there is a natural weak equivalence d(BG) → W (G), where d is the diagonal functor, and BG is the bisimplicial set given by viewing G as a simplicial groupoid, and applying the nerve functor B pointwise. The nerve also defines a functor B : 2 − Gpd → s0Gpd. If H is a 2-groupoid, then BH is the groupoid enriched in simplicial sets with Ob(BH) = Ob(H) and Mor(BH) = B(Mor(H)) . It is an abuse of notation, but we will write W (H) for the simplicial set given by first applying B to the 2-groupoid H, and then applying W . These functors extend to the presheaf level by applying them sectionwise. Because A[2] is a sheaf of 2-groups, we can identify B(A[2]) with the simplicial sheaf of morphisms from the unique object to itself; this simplicial sheaf is just BA. By [4, 2 Corollary 9.39], we have W A[2] ≃ d(BBA) ≃ K(A, 2) . There is a model structure on the category of presheaves of 2-groupoids on C such that a map A → B is a weak equivalence if and only if W A → W B is a local weak equivalence of simplicial presheaves; this is [4, Theorem 9.57]. The Eilenberg-Mac Lane functor W is part of a Quillen equivalence between simplicial presheaves and presheaves of groupoids enriched in simplicial sets [4, Theorem 9.50], and it follows from this that for any presheaf of groupoids G and for any presheaf of 2-groupoids H. [BG, W (H)] ∼= [G, H]2−Gpd For presheaves of 2-groupoids A, B, write h(A, B) for the category whose objects are diagrams and whose morphisms are commutative diagrams A C∼ / B A ∼ ~⑦⑦⑦⑦⑦⑦⑦⑦ `❅❅❅❅❅❅❅❅ ∼ C D ❅❅❅❅❅❅❅❅ >⑦⑦⑦⑦⑦⑦⑦⑦ B This is the category of cocycles from A to B. By [4, Theorem 6.5], we have So, there are isomorphisms π0h(A, B) ∼= [A, B]2−Gpd . H 2(BG, A) = [BG, K(A, 2)] ∼= [BG, W A[2]] ∼= [G, A[2]]2−Gpd ∼= π0h(G, A[2]). As the map f : A[2] → Iso(A) is a section, the induced map f∗ : π0h(G, A[2]) → π0h(G, Iso(A)) is a monomorphism. In [4, Theorem 9.66], Jardine constructs a bijection π0h(G, Iso(A)) ∼= π0Ext(G, A) with the set of path components of a category Ext(G, A), which we'll now define. 3 o o / ~ ` > If p : G′ → G is a map of presheaves of groupoids, let im(p) be the presheaf of groupoids that has the same objects as G′, and with im(p)(x, y) given by the image of the function G′(x, y) → G(p(x), p(y)). Say that a map p : G′ → G is essentially surjective if the canonical map im(p) → G is a local weak equivalence of presheaves of groupoids, in the sense that the induced map of nerves is a local weak equivalence of simplicial presheaves. Say that a kernel of a map p : G′ → G is a diagram j Aut(G′) K "❋❋❋❋❋❋❋❋❋ yssssssssss Ob(G′) where K is a group object in the category of presheaves over Ob(G′), and j is a homomorphism of group objects over Ob(G′), such that the following diagram is a pullback: K j Aut(G′) p∗ Ob(G′) / Aut(G) p∗·e Say that a kernel for p in A is a kernel j : K → Aut(G′) together with a map of presheaves w : K → A such that the induced map K → A×Ob(G′) is a homomorphism of group objects over Ob(G′) that induces an isomorphism of associated sheaves. The objects of Ext(G, A) are triples (p, j, w), where p : G′ → G is essentially surjective, and (j, w) is a choice of kernel in A for p. A morphism σ : (p, j, w) → (p′, j′, w′) of this category is a local weak equivalence σ : G′ → G′′ such that p′ ◦ σ = p, and w′ ◦ σ∗ = w. If G is a presheaf of groups, and p : G′ → G is an essentially surjective map of presheaves of groupoids, then G′ must be locally connected, i.e., a gerbe. The map K → Ob(G′) is an example of what Jardine calls a family of presheaves of groups F → S over the presheaf S, which is a group object in the category of presheaves over S. For an element x ∈ S(U ), the fibre Fx of the family F over x is defined by the pullback diagram Fx FU ∗ x / / SU If (p : G′ → G, j, w) is an object of Ext(G, A), and α : x → y is a morphism of G′(U ), then α defines an isomorphism Kx → Ky of presheaves of groups on C/U by conjugation; via w, we get an induced automorphism of AU . 4 / / " y / / / / / Let CenExt(G, A) ⊂ π0Ext(G, A) be the subset consisting of equivalence classes that have a representative (p : G′ → G, j, w) with the following property: for all objects U of C, and for all morphisms α of G′(U ), the automorphism of AU given by conjugation with α is the identity. If G happens to be a sheaf of groups, and (p : G′ → G, j, w) is an object of Ext(G, A) such that G′ is a sheaf of groups, then the kernel of p is necessarily isomorphic to A; the extension (p, j, w) has the property of the last paragraph if and only if the kernel of pU is contained in the centre of G′(U ) for all objects U of C. Theorem 1.1. For any presheaf of groupoids G on C, and any sheaf of abelian groups A, there is a bijection H 2(BG, A) ∼= CenExt(G, A) . Proof. We will show that CenExt(G, A) is the image of the function H 2(BG, A) ∼= / π0h(G, A[2]) f∗ / π0h(G, Iso(A)) ∼= / π0Ext(G, A). First, we'll recall the definition of the bijection ψ : π0h(G, Iso(A)) → π0Ext(G, A) of [4, Theorem 9.66]. Say G g Z F / Iso(A) is a cocycle from G to Iso(A). Define a presheaf of 2-groupoids EZF , whose objects are the objects of Z. The 1-morphisms x → y of EZ F (U ) are pairs (α, f ) with α : x → y a 1-morphism of Z(U ) and f ∈ A(U ). A 2-morphism (α, f ) → (β, g) of EZ F (U ) is a 2-morphism h : α → β of Z(U ) such that F (h) = g−1f . If α : x → y is a 1-morphism of Z(U ), then F (α) is a 1-morphism of Iso(A)(U ), ie an automorphism of AU ; write α∗ for this automorphism. The composite of (α, f ) : x → y and (β, g) : y → z in EZ F (U ) is (βα, gβ∗(f )) : x → z. Let EZ F = π0(EZ F ) be the presheaf of path component groupoids. There is a map π : EZ F → Z which is the identity on objects, on 1-morphisms is (α, f ) 7→ α, and takes the 2-morphism h : (α, f ) → (β, g) to the underlying 2-morphism h : α → β. Write g∗ for the composite EZ F π∗ / / π0(Z) ∼ / G. Let K(F ) = A × Ob(EZ F ); define j : K(F ) → Aut(EZ F ) by letting jx : AU → EZ Fx be defined in sections by the rule f 7→ [(1x, f )], for all x ∈ Ob(EZ F )(U ). Let w : K(F ) → A be projection; then (j, w) is a kernel for g∗ in A. The rule (g, F ) 7→ (g∗, j, w) defines a functor h(G, Iso(A)) → Ext(G, A), and ψ is the induced function on path components. 5 / / / o o / / An element of π0h(G, Iso(A)) is in the image of f∗ if and only if it has a representative (g : Z → G, F : Z → Iso(A)) such that F factors as Z → A[2] ⊂ Iso(A). This is equivalent to the condition that, for all 1-morphisms α of Z(U ), F (α) = α∗ is the identity on AU . Say (g, F ) is such a cocycle. We will show that for any 1-morphism σ of EZF (U ), the automorphism of AU given by conjugation with σ is the identity. Let σ = [(α, f )] : x → y be a 1-morphism of EZ F (U ). Let V → U be an object of C/U , and let x 7→ x′ and y 7→ y′ under Z(U ) → Z(V ). For any g ∈ A(V ), [(α, f )][(1x′ , g)][(α, f )]−1 = [(α, f )(1x′ , g)(α−1, f −1)] = [(α, f )(α−1, gf −1)] = [(1y′ , f α∗(gf −1))] = [(1y′ , g)] So g 7→ g and σ induces the identity on AU . We've proved that the image of our function H 2(BG, A) → π0Ext(G, A) is contained in CenExt(G, A). To see the opposite inclusion, let's briefly recall the bijection φ : π0Ext(G, A) → π0h(G, Iso(A)) , inverse to ψ. If (p : G′ → G, j, w) is an object of Ext(G, A), there is a cocycle G q R(p) F (p) / Iso(A). The presheaf of 2-groupoids R(p) has the same objects and 1-morphisms as G′, and there is a 2-morphism α → β if and only if p(α) = p(β). For a morphism α of G′(U ), F (p)(α) is the automorphism of AU defined by conjugation with α. We have φ[(p, j, w)] = [(q, F (p))]. Clearly, then, if [(p, j, w)] is an element of CenExt(G, A), then φ[(p, j, w)] is represented by a cocycle in the image of f∗, namely (q, F (p)). This completes the proof. (cid:3) The elements of H 2(BG, A) corresponding to central extensions of sheaves of groups can be interpreted as universal obstruction classes, in the following sense: Theorem 1.2. Let A → G′ p of groups. For any simplicial presheaf X, there is an exact sequence of pointed sets −→ G be a central extension, where A, G′ and G are sheaves H 1(X, G′) p∗ / H 1(X, G) π / H 2(X, A), and if F ∈ H 1(X, G), then π(F ) = F ∗(cp), where cp ∈ H 2(BG, A) classifies the extension A → G′ → G. 6 o o / / / Proof. Because A is central in G′, there is an induced action BA × BG′ → BG′ of the simplicial sheaf of abelian groups BA on the simplicial sheaf BG′. The Borel construc- tion for this action is the bisimplicial sheaf EBA ×BA BG′ with (p, q)-bisimplices given by the qth simplicial degree of the Borel construction for the action A×p ×G′×p → G′×p. The action of A×p on G′×p is free for all p ≥ 0, so the map EA×p ×A×p G′×p → G×p is a local weak equivalence. These maps induce a local weak equivalence from the diagonal Moreover, there is a sequence of bisimplical sheaves d(EBA ×BA BG′) → BG . BG′ → EBA ×BA BG′ → BBA, and, taking diagonals, the sequence BG′ → d(EBA ×BA BG′) → d(BBA) is a sectionwise fibre sequence, hence a local fibre sequence. In general, if A × X → X is an action of a connected simplicial abelian group A on a connected simplicial set X, then the sequences X → A×p × X → A×p are fibre sequences of connected simplicial sets, and so the sequence of bisimplicial sets induces a fibre sequence of simplicial sets after taking diagonals, by a theorem of Bousfield and Friedlander, which appears as [3, Theorem IV.4.9]. X → EA ×A X → BA As d(BBA) ≃ K(A, 2), we have the exact sequence H 1(X, G′) p∗ / H 1(X, G) π / H 2(X, A). This is exactly the argument of [4, Example 9.11] in our case. Say F ∈ H 1(X, G) = [X, BG]. Then π(F ) is given by the diagram in the homotopy category X F / BG ∼ d(EBA ×BA BG′) / d(BBA). To finish the proof, we need to show that the cocycle (∗) represents cp ∈ H 2(BG, A). BG ← d(EBA ×BA BG′) → d(BBA) First, note that EBA ×BA BG′ ∼= BBR(p), where R(p) is the resolution 2-groupoid that appeared in the proof of 1.1. Using the natural weak equivalence d(BH) → W (H), for H a presheaf of groupoids enriched in simplicial sets, we have a pointwise equivalence from (∗) to the cocycle (∗∗) W G ← W (R(p)) → W (A[2]). 7 / / / o o / The functor W has a left adjoint πG : s Pre → Pre(2 − Gpd), where G is the loop groupoid functor to presheaves of groupoids enriched in simplicial sets, and π is the fundamental groupoid functor to presheaves of 2-groupoids. If H is a groupoid enriched in simplicial sets, then π(H) is the 2-groupoid with Ob(π(H)) = Ob(H) and Mor(π(H)) = π(Mor(H)) . This adjunction defines a Quillen equivalence between presheaves of 2-groupoids and the 2-equivalence model structure on simplicial presheaves of [4, Theorem 5.49]. This follows from [4, Proposition 9.59] and [4, Proposition 9.61]. The 2-equivalence model structure has all monomorphisms for cofibrations, so that every object is cofibrant. Now, for M a presheaf of 2-groupoids and M a fibrant model of M , we have natural weak equivalences πGW M ∼ / πGW M ∼ / M , as the functor W preserves weak equivalences. It follows from this that the cocycle obtained by applying πG to (∗∗) is pointwise equivalent to the cocycle G ← R(p) → A[2], which represents the class in π0h(G, A[2]) corresponding to the extension A → G′ → G. (cid:3) The motivating example is the Brauer group: On the ´etale site (SchS )et over a scheme S, the central extension of sheaves of groups Gm → Gln p −→ PGln corresponds to a class cp ∈ H 2 Gln-torsor if and only if the class F ∗(cp) vanishes in H 2 et(BPGln, Gm), and a PGln-torsor F over S lifts to a et(S, Gm). 2. The Nisnevich and ´etale classifying spaces of projective linear groups 2.1. Preliminaries. We'll often make use of the following contruction: if G is a sheaf of groups on a site C, let G − tors be the presheaf of groupoids with G − tors(U ) the groupoid of G-torsors over U , for every object U of C. Then the obvious map BG → B(G − tors) is a local weak equivalence, and B(G − tors) is injective fibrant [4, Corollary 9.27]. Let k be a perfect field, and let Smk be the category of smooth, separated k-schemes. 8 / / For G a presheaf of groups on Smk, we'll use BG to denote the Nisnevich classifying space of G. Following Morel and Voevodsky [5], let BetG be the Nisnevich homotopy type of an ´etale fibrant model of BG. Explicitly, choose a map j : BG → Fet(BG) , where j is an ´etale local equivalence, and Fet(BG) is injective fibrant with respect to the ´etale topology. Then Fet(BG) is a model of BetG. As the name suggests, we have H 1 et(X, G) ∼= π(X, B(G − tors)et) ∼= [X, BetG]Nis where π(−, −) denotes simplicial homotopy classes of maps. In [5, Lemma 4.1.18], Morel and Voevodsky observe the following Proposition 2.1. Let G be a presheaf of groups. The map BG → BetG is a Nisnevich local equivalence if and only if G is an ´etale sheaf, and one of the following equivalent conditions holds: (1) for any smooth scheme S over k, one has H 1 (2) for any smooth scheme S over k and a point x of S, one has Nis(S, G) ∼= H 1 et(S, G). H 1 et(Spec(Oh S,x), G) = ∗ . And, they point out [5, Lemma 4.3.6] that general linear groups satisfy these conditions; so is a Nisnevich local equivalence for all n > 0. BGln → BetGln 2.2. Nisnevich PGln-torsors. If S is a smooth scheme over a field, then the Nisnevich cohomology group H 2 Nis(S, Gm) is zero, and so all Nisnevich PGln-torsors over S lift to Gln-torsors. The situation is more interesting over a simplicial scheme. In the following, an inner automorphism of a presheaf of groups is an automorphism given by conjugation with a global section. Proposition 2.2. Let G be a presheaf of groups on Smk, and let H be a Nisnevich sheaf of groups. Then, H 1 Nis(BG, H) ∼= hom(G, H)/ inner automorphisms of H. Proof. We identify H 1 Nis(BG, H) with the set of maps in the homotopy category, with respect to the Nisnevich local equivalences, [BG, BH]. Because B(H −tors) is a fibrant model of BH, we have an isomorphism [BG, BH] ∼= π(BG, B(H − tors)). 9 We'll begin by defining a function α : π(BG, B(H − tors)) → hom(G, H)/ inner automorphisms of H. Given a class in π(BG, B(H − tors)), choose a representative φ : BG → B(H − tors). Write T for the H-torsor over Spec k corresponding to φ0 : Spec k → Ob(H − tors). Because the groupoid of Nisnevich H-torsors over Spec k is contractible, we can choose an isomorphism τk : T → H, where H is the trivial H-torsor over Spec k; the choice of τk determines an isomorphism of H-torsors over U , τU : TU → HU , for every object U of Smk. Furthermore, the choice of τk allows us to define a map ψ : BG → B(H − tors), where ψ0 : Spec k → Ob(H − tors) corresponds to the trivial H-torsor H, and ψ1 is defined by the rule ψ1(g) = τU φ1(g)τ −1 U for every g ∈ G(U ). By construction, τ : φ ⇒ ψ defines a simplicial homotopy. In simplicial degree 1, we have ψ1 : G → Aut(H) = H; define α([φ]) = [ψ1]. To see this is well-defined, let φ′ : BG → B(H − tors) be a map with σ : φ ⇒ φ′ a simplicial homotopy. Write T ′ for the H-torsor over Spec k corresponding to φ′ 0, and choose an isomorphism τ ′ k : T ′ → H, which determines a simplicial homotopy τ ′ : φ′ ⇒ ψ′. Let µ : ψ ⇒ ψ′ be the composition τ ′ ◦ σ ◦ τ −1. The homotopy µ gives an isomorphism H → H of H-torsors over Spec k, i.e., an element h ∈ H(k), and we have 1 = hψ1h−1, so that α is well-defined. ψ′ The function α is a bijection, as it has inverse β : hom(G, H)/ inner automorphisms of H → π(BG, B(H − tors)), defined as follows: given a class in the domain, choose a representative f : G → H, and let β([f ]) = [BG B(f ) −−−→ BH → B(H − tors)]. If h ∈ H(k) and f ′ = hf h−1, then h defines a simplicial homotopy B(f ) ⇒ B(f ′), so β is well-defined. (cid:3) So, for any presheaf of groups G, a Nisnevich PGln-torsor over BG is given by a homomorphism f : G → PGln, and f lifts to a Gln-torsor over BG if and only if f factors through the canonical map p : Gln → PGln. By the results of Section 1, the central extension Gm → Gln → PGln corresponds to an element cp of the group H 2 Nis(BPGln, Gm), and a homomorphism f : G → PGln factors through Gln if and only if f ∗(cp) vanishes in H 2 Nis(BG, Gm). This is a statement about motivic cohomology, because of the isomorphism for any simplicial presheaf X. H 2 Nis(X, Gm) ∼= H 3(X, Z(1)), In contrast, the group H 3(BetG, Z(1)) is zero if G is a linear algebraic group, because BetG has a model that is a sequential colimit of smooth schemes, by [5, Proposition 4.2.6]. 10 In particular, the canonical homomorphism in motivic cohomology H ∗(BetPGln, Z(∗)) → H ∗(BNisPGln, Z(∗)) is not surjective. 2.3. Chern classes. Recall that we're using BG to denote the Nisnevich classifying space of a presheaf of groups G on the smooth site Smk. If k is a perfect field, the motivic cohomology of BGln is a polynomial algebra over the cohomology of the base field [6]: H ∗(BGln, Z(∗)) ∼= H ∗(k, Z(∗))[c1, . . . , cn] with ci ∈ H 2i(BGln, Z(i)). If G is a presheaf of groups on the Nisnevich site Smk as before, and f : G → Gln is a representation, then we can define the chern classes of f to be ci(f ) = f ∗(ci) ∈ H 2i(BG, Z(i)) . As BGln ≃ BetGln, we can define chern classes in the motivic cohomology of BetG in the same way. Because the canonical map in the homotopy category BG → BetG is natural in G, the homomorphism in motivic cohomology H 2∗(BetG, Z(∗)) → H 2∗(BG, Z(∗)) takes chern classes to chern classes. The goal of this section is to show that, when p is an odd prime, this homomorphism H 2∗(BetPGlp, Z(∗)) → H 2∗(BPGlp, Z(∗)) is injective, over any field of characteristic zero containing a primitive pth root of unity. First, we need a lemma, which says that the motivic cohomology of a constant simplicial presheaf, with coefficients in a field, is as simple as possible. Let X be a simplicial set, and let F be any field. There is an adjunction Γ∗ : s(F − vec) ⇄ s PreF (Smk) : Γ∗ where Γ∗ is the constant presheaf functor, and Γ∗ is global sections. Lemma 2.3. Let F be a field, and let X be a simplicial set such that all singular cohomology groups H r(X, F ) are finite-dimensional. Then, H ∗(Γ∗X, F (∗)) ∼= H ∗(k, F (∗)) ⊗ H ∗(X, F ) , where elements of H r(X, F ) are seen as elements of the motivic cohomology group H r(Γ∗X, F (0)). 11 Proof. The adjunction Γ∗ ⊣ Γ∗ is a Quillen adjunction for the injective model structure on s PreF (Smk) and the usual model structure on s(F − vec). So, we have H p(Γ∗X, F (q)) = [F (Γ∗X), F (q)[−p]] ∼= [F X, F (q)[−p](k)]. For any simplicial F -vector spaces C and D, [C, D] ∼= Y n≥0 hom(Hn(C), Hn(D)) . Hn(F (q)[−p](k)) = H p−n(k, F (q)), For n ≥ 0, we have and so we have H p(Γ∗X, F (q)) ∼= Y hom(Hn(X, F ), H p−n(k, F (q))) . n≥0 As H r(k, F (0)) ∼= F if r = 0, and is zero otherwise, we have H ∗(Γ∗X, F (0)) ∼= H ∗(X, F ) . If V, W are F -vector spaces with V finite-dimensional, then the canonical map is an isomorphism. So, by our assumptions on X, we have V ∨ ⊗ W → hom(V, W ) H p(Γ∗X, F (q)) ∼= Y H n(X, F ) ⊗ H p−n(k, F (q)) . n≥0 (cid:3) In order to prove our result for PGlp, we'll need its analogue for the group scheme of roots of unity µp. Lemma 2.4. Let p be prime. Over any perfect field k containing a primitive pth root of unity, the homomorphism H 2∗(Bet µp, Z(∗)) → H 2∗(Bµp, Z(∗)) is injective. Proof. The Chow ring of Bet µp is generated by the first chern class t of the embedding µp ⊂ Gm ([9, p190]): H 2∗(Bet µp, Z(∗)) ∼= Z[t]/(p · t) . Furthermore, H 2(Bµp, Z(1)) ∼= hom(µp, Gm) ∼= Z/p . Let c denote the first chern class of µp ⊂ Gm in H 2(Bµp, Z(1)): we need to show that c has infinite multiplicative order in H 2∗(Bµp, Z(∗)). For this we use Z/p coefficients. First, note that H 0(k, Z/p(i)) ∼= H 0 et(k, µ⊗i 12 p ) ∼= Z/p , and it's easy to see that a generator x ∈ H 0(k, Z/p(1)) has infinite multiplicative order. As k contains a primitive pth root of unity, the obvious map Γ∗BCp → Bµp is a Nisnevich local equivalence, where Cp denotes the cyclic group with p elements. By 2.3, H 2(Γ∗BCp, Z/p(0)) ∼= H 2(BCp, Z/p) ∼= Z/p , and a generator y of this group has infinite multiplicative order in H ∗(Γ∗BCp, Z/p(0)). Again by 2.3, x · y ∈ H 2(Γ∗BCp, Z/p(1)) has infinite multiplicative order. One checks that, under the map H 2(Γ∗BCp, Z(1)) → H 2(Γ∗BCp, Z/p(1)) we have c 7→ x · y, so that c has infinite multiplicative order as well. (cid:3) The proof of the following theorem relies on the work of Vistoli on the Chow ring of the classifying space of PGlp [9]. By a result of Morel-Voevodsky, [5, Proposition 4.2.6], we have A∗ G ∼= H 2∗(BetG, Z(∗)) , where G is a linear algebraic group, and A∗ of G, in the sense of Totaro [8]. G is the Chow ring of the classifying space Theorem 2.5. Let p be an odd prime. Over any field of characteristic zero containing a primitive pth root of unity, the canonical homomorphism in motivic cohomology H 2∗(BetPGlp, Z(∗)) → H 2∗(BPGlp, Z(∗)) is injective. Proof. In [9], Vistoli defines a subgroup CP × µp ⊂ PGlp, as follows. Let ω be a primitive pth root of unity in the base field k, and let τ be the diagonal matrix diag(ω, ω2, . . . , ωp−1, 1). Then τ generates a subgroup of PGlp isomorphic to µp. Let σ be the permutation matrix corresponding to the cycle (1 2 . . . p) ∈ Sp. Then σ generates a subgroup of PGlp isomorphic to Cp, the cyclic group of order p, viewed as a group scheme over k in the usual way. Of course, Cp ∼= µp, but the notation is meant to be suggestive. In Glp, we have τ σ = ωστ , so σ and τ commute in PGlp, and they generate a subgroup of PGlp isomorphic to Cp × µp. Just for this proof, write CH ∗X = H 2∗(X, Z(∗)) . 13 Let TP Glp be the standard maximal torus in PGlp, consisting of classes of diagonal matrices. By work of Totaro and Vistoli, [9, Proposition 9.3] and [9, Proposition 9.4], the inclusions TP Glp ⊂ PGlp and Cp × µp ⊂ PGlp induce an injective homomorphism CH ∗BetPGlp → CH ∗BetTP GLp × CH ∗Bet(Cp × µp) . We have ([9, p194]): CH ∗Bet(Cp × µp) ∼= Z[ξ, η]/(pξ, pη) , where ξ is the first chern class of the character with σ 7→ ω and τ 7→ 1, and η is the first chern class of the character with σ 7→ 1 and τ 7→ ω. Using this and an argument analogous to the proof of 2.4, one shows that the natural map CH ∗Bet(Cp × µp) → CH ∗B(Cp × µp) is injective. As group-schemes, we have TP Glp ∼= TSlp ∼= G×p−1 m , so that BetTP GLp ≃ BTP Glp. Consider the commutative diagram CH ∗BetPGlp CH ∗BPGlp CH ∗BetTP GLp × CH ∗Bet(Cp × µp) / CH ∗BTP GLp × CH ∗B(Cp × µp) The bottom route around the square is injective, and it follows that CH ∗BetPGlp → CH ∗BPGlp is injective. (cid:3) Finally, let's summarize what 2.5 tells us about H 2∗(BPGlp, Z(∗)). Let R = H ∗(k, Z(∗)). We have H ∗(BG×n m , Z(∗)) ∼= R[x1, . . . , xn] , where xi ∈ H 2(BG×n m , Z(1)), H ∗(BGln, Z(∗)) ∼= R[σ1, . . . , σn] , where σi is the ith elementary symmetric polynomial in the variables xi, and H ∗(BTP Gln, Z(∗)) ∼= R[x1 − x2, . . . , xn−1 − xn] . Furthermore, there is a commutative diagram: H ∗(BPGln, Z(∗)) R[x1 − x2, . . . , xn−1 − xn] R[σ1, . . . , σn] / R[x1, . . . , xn] 14 / / / / / / This gives a homomorphism to the symmetric part of the cohomology of the maximal torus: H ∗(BPGln, Z(∗)) → R[x1 − x2, . . . , xn−1 − xn]Sn . Restricting attention to the Chow groups, and letting n = p, we have a homomorphism H 2∗(BPGlp, Z(∗)) → Z[x1 − x2, . . . , xp−1 − xp]Sp , and [9, Theorem 3.2], and 2.5 together imply that this homomorphism has a section. Furthermore, there is an element ρ ∈ H 2p+2(BPGlp, Z(p + 1)) not in the image of this section, satisfying some relations given in [9, Theorem 3.3]. References 1. Lawrence Breen, Extensions du groupe additif, Inst. Hautes ´Etudes Sci. Publ. Math. (1978), no. 48, 39 -- 125. 2. Jean Giraud, Cohomologie non ab´elienne, Springer-Verlag, Berlin-New York, 1971, Die Grundlehren der mathematischen Wissenschaften, Band 179. 3. Paul G. Goerss and John F. Jardine, Simplicial homotopy theory, Modern Birkhauser Classics, Birkhauser Verlag, Basel, 2009, Reprint of the 1999 edition [MR1711612]. 4. John F. Jardine, Local homotopy theory, Springer Monographs in Mathematics, Springer, New York, 2015. 5. Fabien Morel and Vladimir Voevodsky, A1-homotopy theory of schemes, Inst. Hautes ´Etudes Sci. Publ. Math. (1999), no. 90, 45 -- 143 (2001). 6. Oleg Pushin, Higher Chern classes and Steenrod operations in motivic cohomology, K-Theory 31 (2004), no. 4, 307 -- 321. 7. Danny Stevenson, D´ecalage and Kan's simplicial loop group functor, Theory Appl. Categ. 26 (2012), No. 28, 768 -- 787. 8. Burt Totaro, The Chow ring of a classifying space, Algebraic K-theory (Seattle, WA, 1997), Proc. Sympos. Pure Math., vol. 67, Amer. Math. Soc., Providence, RI, 1999, pp. 249 -- 281. 9. Angelo Vistoli, On the cohomology and the Chow ring of the classifying space of PGLp, J. Reine Angew. Math. 610 (2007), 181 -- 227. 15
1509.06063	2	1509	2016-10-22T08:59:53	Wild coverings of Berkovich curves	[ "math.AG" ]	This paper is an extended version of the author's talk given at the conference "Non-Archimedean analytic geometry: theory and practice" held in August 2015 at Papeete. It gives a brief overview of recent results on the structure of wild coverings of Berkovich curves and its relation to the different and higher ramification theory.	math.AG	math	WILD COVERINGS OF BERKOVICH CURVES MICHAEL TEMKIN 1. Introduction This paper is an extended version of the author's talk given at the conference "Non-Archimedean analytic geometry: theory and practice" held in August 2015 at Papeete, and I wish to thank the organizers. It gives a brief overview of results and methods of works [CTT16] and [Tem14] on the structure of finite morphisms between Berkovich curves. The structure of tame morphisms between smooth Berkovich curves is pretty well-known and it is completely controlled by the simultaneous semistable reduction theorem, see, for example, [ABBR13]. The structure of wild morphisms was for a long time terra incognita, though one should mention some special results recently obtained by Faber in [Fab13a] and [Fab13b]. In this project we obtain a relatively complete description of the combinatorial structure of an arbitrary finite morphism f : Y → X between smooth Berkovich curves. It is divided into two parts. 1.1. The different function. In a joint work [CTT16] with A. Cohen and D. Trushin we study the different function δf : Y hyp → [0, 1] that assigns the different δH(y)/H(f (y)) to a point y of type 2, 3 or 4. In other words we study the analytic behavior of the most important invariant that measures wildness of an extension of valued fields, the different. It turns out that δf controls the minimal semistable model of f , and a balancing condition for the slopes of δf at a type 2 point y ∈ Y extends the local Riemann-Hurwitz formula to the wild case. 1.2. The multiplicity function and radialization. Both X and Y have canon- ical exponential metrics and f is piecewise monomial with respect to them. The behavior of f as a piecewise monomial function is controlled by the multiplicity function nf that assigns [H(y) : H(f (y))] to y. This function and the multiplicity loci Nf,≥d = {y ∈ Y nf (y) ≥ d} are described in [Tem14]. In particular, it is shown that nf is radial with respect to a large enough skeleton ΓY ⊂ Y . A central player in this study is a profile function φf : Y hyp → P[0,1] encoding the radii of all sets Nf,≥d around the skeleton, where P[0,1] is the set of piecewise monomial bijec- tions of [0, 1] onto itself. Furthermore, δf can be retrieved from φf via composing with a character P[0,1] → [0, 1] and φ is an analytic family of the classical Herbrand functions. Acknowledgments. I would like to thank the referee for pointing out various inaccuracies in the first version of the paper. 2. Semistable reduction and tame morphisms In this section we summarize the relatively well-known properties of curves and morphisms. 1 2 MICHAEL TEMKIN 2.1. Conventions. 2.1.1. Ground field. We work over a fixed algebraically closed non-archimedean analytic (i.e. real-valued complete) field k with a non-trivial valuation. 2.1.2. Nice curves. By a nice curve we mean a rig-smooth connected separated compact k-analytic curve. 2.1.3. Subgraphs. By a subgraph Γ of a nice curve C we mean a connected topo- logical subgraph Γ ⊂ C with finitely many vertices and edges such that the set of vertices Γ0 consists of points of C of types 1 and 2 and contains at least one point of type 2. 2.2. Semistable reduction for curves. 2.2.1. Skeletons of curves. A subgraph Γ is called a skeleton of C if C \ Γ0 is a disjoint union of open discs Di and semi-annuli A1, . . . ,An (i.e. Ai is either an open annulus or a punched open disc) and the edges of C are the skeletons of A1, . . . ,An. The following skeletal version of the semistable reduction theorem is easily seen to be equivalent to its classical versions. Theorem 2.2.2. Any nice curve possesses a skeleton. 2.2.3. Combinatorial structure of the curve. In a sense, a skeleton of a curve pro- vides the best possible combinatorial description of the curve. In particular, the complement C \ Γ of a skeleton is a disjoint union of discs and there is a canonical deformational retraction qΓ : C → Γ. 2.2.4. Genus. For any point x ∈ C we define the genus g(x) to be the genus of ]H(x)/ek if x is of type 2 and zero otherwise. The genus of C is then defined to be g(C) = h1(C) + Px∈C g(x); it is finite and equals to g(C) = h1(Γ) + Pv∈Γ0 g(v). This gives the usual genus of an algebraic curve when C is proper, but g(C) is a meaningful invariant for nice curves with boundary too. 2.2.5. Exponential metric. Let A be an open or closed annulus, i.e. A is isomorphic to the domain in A1 k given by r2 < t < r1 or r2 ≤ t ≤ r1. The number r(A) = r1/r2 depends only on A and it is called the radius of A. Given an interval I ⊂ C one defines its radius (or exponential length) by r(I) = supQn i=1 r(Ai), where the supremum is taken over all finite sets of disjoint open annuli Ai ⊆ C such that the skeleton of each Ai lies in I. It turns out that r defines an exponential metric on C whose singular points are precisely the points of type 1. In other words, r([a, c]) = r([a, b])r([b, c]) for an interval [a, c] ⊂ C with a point b ∈ [a, c], and r([a, b]) = ∞ if and only if the set {a, b} contains a point of type 1. Remark 2.2.6. (i) If I is the skeleton of A then r(I) = r(A). In fact, this is the main property of the radius one should check in order to establish all other properties. (ii) We prefer to work with the exponential metric in this paper, but one often considers its logarithm, which is a usual metric. For example, if I is the skeleton of an annulus A then the length of I is the modulus of A. The classical metric is only canonical up to rescaling since its definition involves a choice of the base of the logarithm. WILD COVERINGS OF BERKOVICH CURVES 3 2.2.7. Radius parametrization. If I is closed with an endpoint a of type different from 1 then x 7→ r([a, x])−1 provides the canonical homeomorphism I = [r(I)−1, 1] that we call radius parametrization of I. Note that I = [0, 1] if and only if the second endpoint is of type 1. Remark 2.2.8. (i) If E is a unit disc with maximal point q then for any point x ∈ E there exists a unique interval [x, q] and r(x) = r([x, q])−1 is the usual radius function of the disc. (ii) In the same way, any skeleton Γ induces a radius function rΓ : C → [0, 1] that measures the inverse exponential distance to the skeleton. 2.2.9. Enhanced skeleton. We naturally enhance a skeleton Γ of a curve to a metric genus graph in which each vertex is provided with a genus and each edge is provided with a radius (exponential length). 2.3. Semistable reduction for morphisms. 2.3.1. Morphisms and metrics. Let f : Y → X be a non-constant morphism of nice curves. It is easy to see that f is pm or piecewise monomial in the sense that for each interval I ⊂ Y the set f (I) is a graph and the map I → f (I) is pm with integral slopes with respect to the radii parameterizations. Moreover, the multiplicity function nf (see §1.2) is the absolute value of the degree of f in the sense that nf I = deg(f I ). Thus, nf completely encodes the pm (or metric) structure of f . 2.3.2. Skeletons of morphisms. Let f : Y → X be a generically ´etale morphism of nice curves. By a skeleton of f we mean a pair Γ = (ΓY , ΓX ) of skeletons of Y and X such that ΓY contains the ramification locus Ram(f ) and f −1(ΓX ) = ΓY (in particular, f −1(Γ0 X ) = Γ0 Y ). 2.3.3. Semistable reduction. It is easy to see that if Γ ⊆ Γ′ are two subgraphs and Γ is a skeleton then Γ′ is a skeleton. Using this and the semistable reduction for curves one easily obtains the simultaneous semistable reduction theorem that can also be called semistable reduction of morphisms. Theorem 2.3.4. Any generically ´etale morphism between nice curves possesses a skeleton Γ. Remark 2.3.5. On the complement of a skeleton a morphism reduces to finite ´etale coverings of open discs by open discs. In general, such a morphism may have a complicated structure and this is the reason why a skeleton provides a pretty loose control on the morphism. 2.3.6. Tame morphisms. A morphism f between curves is called tame if nf takes values invertible in ek and f is called wild otherwise. A tame ´etale covering of a disc by a disc is trivial and a tame ´etale covering of an annulus by an annulus is isomorphic to the standard Kummer covering of the form t 7→ tn. So, tame morphisms are controlled by skeletons very tightly. 4 MICHAEL TEMKIN 2.3.7. Maps of skeletons. More generally, it is easy to see that an ´etale covering of an annulus by an annulus is of the form t 7→ P∞ i=−∞ aiti where the series has a single dominant term adtd and d > 0. In particular, the map is of degree d on the skeleton. Thus, if Γ is a skeleton of f : Y → X then the map of graphs ΓY → ΓX is enhanced to a map of metric genus graphs: to each vertex v ∈ Γ0 Y one associates a multiplicity nv and to each edge e ∈ ΓY one associates the multiplicity ne such that r(f (e)) = r(e)ne . These multiplicities satisfy the natural balancing X and conditions: nv = Pe∈f −1(h)∩Br(v) ne for any vertex v ∈ Γ0 Y and an edge h ∈ Br(f (v)) of ΓX , if f is finite then Pv∈f −1(u) nv = deg(f ) for any vertex u ∈ Γ0 where Br(v) denotes the set of all edges (or branches) coming out of v. 2.3.8. Local Riemann-Hurwitz. For a finite tame f one also has the local Riemann- Hurwitz formulas: for any v ∈ Γ0 Y with u = f (v) one has that 2g(v) − 2 − 2nv(g(u) − 1) = X (ne − 1), e∈Br(v) which is proved by applying the RH formula to ]H(v)/]H(u). These formulas and the global genus formula imply the global RH formula when X is proper. Remark 2.3.9. One would like to extend the above formula to the non-tame case, and it is natural to expect that the local term at e should be equal to the local term at the point corresponding to e in the classical RH formula (e.g., see §3.1.4 below) of ]H(v)/ ^H(f (v)). For non-tame morphisms two things should be modified, and we will later see that both are dealt with using the different. (1) If f is not wild at v (i.e. nv ∈ ek×) but the ramification is wild along an edge e going out of v then the local term Re at e should be larger than ne − 1. So, one should naturally interpret Re in terms of the map of k-analytic curves. (2) If f is wild at v then it often happens that ]H(v)/ ^H(f (v)) is inseparable. In this case, there exists no RH-like formula based on the residue fields, and a new source of information is needed. Remark 2.3.10. (i) A tame f is split outside of a skeleton and the only restrictions on the multiplicity function along the skeleton are the balancing conditions and the local RH formulas. (ii) For wild maps the sets Nf,≥d are often huge. For example, for the Kummer to itself the set Nf,≥p is the metric neighborhood of [0, ∞] map t 7→ tp from P1 of radius p1/(p−1). Cp 3. The different function This section describes the results of [CTT16]. We start with recalling the defi- nition of different and then list our main results on the different function. 3.1. Different of extensions. 3.1.1. The definition. Let L/K be a finite extension of real-valued fields and assume that either K is discretely valued with perfect residue field or K is of the form H(x), where x is a point of a nice curve. With the convention that the absolute value of an ideal I ⊆ L◦ is supc∈I c, the different of L/K is defined to be δL/K = Ann(ΩL◦/K ◦) WILD COVERINGS OF BERKOVICH CURVES 5 if L/K is separable and δL/K = 0 otherwise. Remark 3.1.2. (i) The different measures wildness of extensions and it is multi- plicative in towers. (ii) In the case of discrete valuations one often considers the additive analogue, which is the length of ΩL◦/K ◦ . (iii) In general, the different is defined using the zeroth Fitting ideal rather than the annihilator. In our case, the torsion module ΩL◦/K ◦ is a subquotient of L◦ so both definitions agree. 3.1.3. The log different. The log different δlog using the module Ωlog δlog L/K = δL/KπL/πK, and δlog L/K is defined similarly to δL/K but L◦/K ◦ of logarithmic differentials. If K is discretely valued then L/K = δL/K otherwise. 3.1.4. The RH formula. If h : Y → X is a finite separable morphism of smooth proper connected ek-curves then the classical RH formula is (δlog 2g(Y ) − 2 − 2n(g(X) − 1) = X δy/x = X y/x + ny − 1) y∈Y y∈Y where n = deg(h), x = h(y) and δy/x is the (additive) different of k((y))/k((x)) for k((x)) = Frac( bOX,x) and k((y)) = Frac( bOY,y). 3.2. The different function. Let now f : Y → X be a finite generically ´etale morphism of nice k-analytic curves and let δf be the different function introduced in §1.1. 3.2.1. Slopes. Naturally, δf contains information about the classical different at points of Y of type 1 and branches of Y at points of type 2. It is retrieved from the slopes (or degrees) of δf . Theorem 3.2.2 ([CTT16, 4.1.8, 4.6.4]). (i) The different function extends uniquely to a pm function δf : Y → [0, 1]. (ii) The slope of δf at a type 1 point y equals δlog y/x. In particular, it is positive if and only if f is wildly ramified at y. (iii) If f is tame at a type 2 point y and v is a branch of Y at y then slopev(δf ) = δlog v/f (v). Remark 3.2.3. This indicates that δf is, in fact, the log different function. This does not affect its values at the points of Y but gives a better interpretation of formulas involving differents of discretely valued fields. 3.2.4. The balancing condition. Slopes of δf at a type 2 point satisfy the balancing condition of RH type which applies to the case when ]H(y)/ ^H(f (y)) is arbitrary. Theorem 3.2.5 ([CTT16, 4.5.4]). If y ∈ Y is of type 2 and x = f (y) then 2g(y) − 2 − 2ny(g(x) − 1) = X (−slopevδf + nv − 1). In particular, almost all slopes of δf at y equal to ni bility degree of ]H(y)/ ^H(f (y)). y − 1, where ni y is the insepara- v∈Br(y) 6 MICHAEL TEMKIN Remark 3.2.6. The balancing condition 3.2.5, the formula for slopes at type 1 points, and the global genus formula imply the global RH formula when Y is proper, and this can also be extended to the case with boundary. This indicates that the balancing formula is the "right" generalization of the local RH formula to the wild case. 3.2.7. The method. The different function is a family of differents, so it is not surprising that one can describe it using a sheafified version of the definition of X d(O◦ δL/K. Namely, one considers the sheaf Ω⋄ X ) which can be informally X is a torsion sheaf of k◦-modules X /k◦ . Then Ω⋄ thought of as a version of ΩO◦ whose stalk at y is cyclic with the absolute value of the annihilator δf (y). Choose a ∈ k◦ with a = δf (y). Reductions of Ω⋄ X at y induce a non- zero meromorphic map λ : ef ∗Ω eX → Ω eY , where ef : eY → eX is the map of ek-curves associated with ]H(y)/]H(x). Then the balancing condition boils down to computing the degree of Ω eY ⊗ ef ∗Ω−1 using poles and zeros of the section induced by λ. eX X = O◦ Y /f ∗Ω⋄ Y and a−1f ∗Ω⋄ 3.2.8. The different function and the skeletons. Let ΓY → ΓX be a skeleton of f . It is natural to encode the balancing condition in the combinatorics of Γ. For this we should first enhance its structure by including the pm different function δΓ = δf ΓY . In addition, one should check whether for a vertex y ∈ Γ0 Y the skeleton contains all branches v at y which are δf -non-trivial, i.e. satisfy the condition slopevδf 6= nv −1. It turns out that in this way one obtains a non-trivial characterization of skeletons. Theorem 3.2.9 ([CTT16, 6.3.4]). Let ΓX be a skeleton of X and ΓY = f −1(ΓX ). Then (ΓY , ΓX) is a skeleton of f if and only if Ram(f ) ⊆ Γ0 Y and for any point y ∈ ΓY all branches at y pointing outside of ΓY are δf -trivial. Remark 3.2.10. (i) The behavior of δf completely describes the locus Nf,p when deg(f ) = p. For example, if f maps P1 to itself by t 7→ tp then the different is minimal and equal to p on [0, ∞] and it is trivial outside, i.e. it growths in all directions outside of the skeleton [0, ∞] with slope p − 1. This explains why Nf,p is a metric neighborhood of radius p1/(p−1). Cp (ii) Even when the different δf behaves trivially on an interval I = [y, z] its slopes depend on the multiplicity function. For example, if ny = p then the value of δf (y) determines δf I , but if ny = pn then to determine δf I one should know the points xpn , xpn−1 , . . . ,xp2 where nf drops and these points can be pretty arbitrary. In particular, the skeleton of f does not control the sets Nf,≥d in any reasonable sense. 4. Radialization and the profile function 4.1. Radialization of the sets Nf,≥d. Let Γ be a skeleton of a nice curve X, qΓ : X → Γ the retraction, and rΓ : X → [0, 1] the inverse exponential distance from Γ. A closed subset S ⊆ X is called Γ-radial if there exists a function r : Γ → R≥0 such that S consists of all points x ∈ X satisfying rΓ(x) ≥ r(qΓ(x)). Remark 4.1.1. It is easy to see that if a skeleton radializes S then any larger skeleton does so. Theorem 4.1.2 ([Tem14, 3.3.7 and 3.3.9]). If f : Y → X is a finite morphism between nice curves then there exists a skeleton of Y that radializes the sets Nf,≥d. Moreover, if (ΓY , ΓX ) is an arbitrary skeleton of f then ΓY radializes these sets in WILD COVERINGS OF BERKOVICH CURVES 7 each of the following cases: (1) f is a normal covering (e.g. Galois), (2) f is tame, (3) f is of degree p. Example 4.1.3. It follows from Theorem 3.2.9 that if f is of degree p then Nf,p is Γ-radial of radius δ1/(p−1) Γ for any skeleton (Γ, ΓX ) of f . f 4.1.4. The splitting method. Many results about extensions of valued fields are proved by the following splitting method: 1) Prove the result for tame extensions and wild extensions of degree p. Often these cases are simpler and can be managed by hands. 2) Extend the result to compositions, obtaining the case of Galois extensions. 3) Use some form of descent to deduce the non-normal case. The splitting method extends to a local-analytic setting because the category of ´etale covers of a germ (X, x) of an analytic space at a point is equivalent to the category of ´etale H(x)-algebras by a theorem of Berkovich. Theorem 4.1.2 is proved easily by the splitting method since the tame case is clear and the degree-p case is controlled by the different by Example 4.1.3. 4.2. The profile function. One may wonder if the radii of the sets Nf,pn are reasonable functions analogous to the different. The answer is yes, but the best way to work with them is to combine them into a pm function from [0, 1] to itself. 4.2.1. Γ-radial morphisms. Let f : Y → X be a morphism and Γ = (ΓY , ΓX ) a skeleton of f . For a point a ∈ Y of type 1 consider the interval I = [a, qΓY (a)] and identify it with [0, 1] via the radius parametrization. Similarly, identify f (I) = [f (a), qΓX (f (a)] with [0, 1]. Then f I is interpreted as an element of P[0,1] and we say that f is Γ-radial if f I = φq depends only on q = qΓY (a). In this case we say that φ : ΓY → P[0,1] is the profile function of f . Remark 4.2.2. (i) It is easy to see that f is Γ-radial if and only if all sets Nf,d are ΓY -radial and then the breaks of φq occur at the radii of the sets Nf,pn . (ii) Thus, the radialization theorem implies that any finite morphism is Γ-radial for a large enough skeleton Γ. In particular, this gives another way to define φq: it is the map f I for a generic interval connecting q to a type 1 point. (iii) The profile function is a much more convenient invariant than the set of radii of Nf,d, mainly because it is compatible with compositions of radial morphisms. 4.2.3. Interpretation as Herbrand function. It turns out that the profile function can be interpreted using a classical invariant from the theory of valued fields. It is well-known that for a finite seprable extension l/k of discrete valuation fields with perfect residue fields, the Herband function φl/k is a multiplicative (with respect to towers of extensions) invariant which efficiently encodes nearly all information about the wild ramification properties of l/k. It is shown in [Tem14, §4] that the theory extends to extensions of the form H(y)/H(x), where y, x are points on k-analytic curves (for an algebraically closed k). The only technical obstacle is that in the classical theory one crucially uses that the extension of integers is monogeneous while H(y)◦/H(x)◦ is integral but does not have to be finite. However, it is shown in [Tem14, 4.2.8] that H(y)◦/H(x)◦ is almost monogeneous in the sense that H(y)◦ is a filtered union of subrings of the form H(x)◦[t], and it is shown in [Tem14, §4.3] that the classical theory extends to extensions of analytic fields with almost monogeneous extensions of rings of integers. 8 MICHAEL TEMKIN Theorem 4.2.4 ([Tem14, 4.5.2]). If f : Y → X is a generically ´etale morphism between nice curves then for any point y ∈ Y of type 2 with x = f (y) the pro- file function φy coincides with the Herbrand function φH(y)/H(x) of the extension H(y)/H(x). Remark 4.2.5. (i) The proof is via the splitting method using that for extensions of degree p the Herbrand function is determined by the different (it has slopes 1 and p and the break point is determined by the different). (ii) The theorem gives a natural geometric interpretation of Herbrand function which works for all extensions (even inseparable ones) on the equal footing. Note that the classical Herbrand function is defined first for Galois extensions and then extended to arbitrary ones by multiplicativity. 4.2.6. Piecewise monomiality of the profile function. It is natural to expect that the profile functions φy should discover a nice global behavior. Indeed, one can easily introduce a notion of pm functions on Y with values in P[0,1] and the following result holds. Theorem 4.2.7 ([Tem14, 3.4.8]). If f is as above then the family of profile func- tions φy extends uniquely to a pm function φ : Y hyp → P[0,1]. References [ABBR13] Omid Amini, Matthew Baker, Erwan Brugall´e, and Joseph Rabinoff, Liting harmonic morphisms I, ArXiv e-prints (2013). [CTT16] Adina Cohen, Michael Temkin, and Dmitri Trushin, Morphisms of Berkovich curves and the different function, Adv. Math. 303 (2016), 800 -- 858. [Fab13a] Xander Faber, Topology and geometry of the berkovich ramification locus for rational functions, I, Manuscripta Mathematica 142 (2013), no. 3-4, 439 -- 474 (English). [Fab13b] , Topology and geometry of the berkovich ramification locus for rational func- tions, II, Mathematische Annalen 356 (2013), no. 3, 819 -- 844 (English). [Tem14] Michael Temkin, Metric uniformization of morphisms of Berkovich curves, ArXiv e- prints (2014), http://arxiv.org/abs/1410.6892. Einstein Institute of Mathematics, The Hebrew University of Jerusalem, Giv'at Ram, Jerusalem, 91904, Israel E-mail address: [email protected]
1605.08088	4	1605	2017-01-16T23:12:08	Hodge ideals	[ "math.AG", "math.CV" ]	We use methods from birational geometry to study M. Saito's Hodge filtration on the localization along a hypersurface. This filtration leads to a sequence of ideal sheaves, called Hodge ideals, the first of which is a multiplier ideal. We analyze their local and global properties, and use them for applications related to the singularities and Hodge theory of hypersurfaces and their complements.	math.AG	math	HODGE IDEALS MIRCEA MUSTAT¸ A AND MIHNEA POPA Abstract. We use methods from birational geometry to study the Hodge filtration on the localization along a hypersurface. This filtra- tion leads to a sequence of ideal sheaves, called Hodge ideals, the first of which is a multiplier ideal. We analyze their local and global properties, and use them for applications related to the singularities and Hodge theory of hypersurfaces and their complements. . G A h t a m [ 4 v 8 8 0 8 0 . 5 0 6 1 : v i X r a Contents A. Introduction . . . . . . . . . . . . . . . . . . . . . . . . . B. Preliminaries . . . . . . . . . . . . . . . . . . . . . . . . 1. Background on filtered D-modules . . . . . . . . . . . . . . . 2. Localization along a divisor . . . . . . . . . . . . . . . . . . . 3. Filtrations on localizations and tensor products . . . . . . . . C. Saito's Hodge filtration and Hodge modules . . . . . 4. Hodge D-modules and strictness . . . . . . . . . . . . . . . . 5. Vanishing theorem . . . . . . . . . . . . . . . . . . . . . . . . 6. Localization as a Hodge D-module . . . . . . . . . . . . . . . 7. Hodge filtration on the complement . . . . . . . . . . . . . . D. Birational definition of Hodge ideals . . . . . . . . . . 8. The simple normal crossing case . . . . . . . . . . . . . . . . 9. The general case . . . . . . . . . . . . . . . . . . . . . . . . . 10. The case k = 0. . . . . . . . . . . . . . . . . . . . . . . . . . 11. Independence of resolution, and filtration property . . . . . 12. Comparison with Hodge filtration, and strictness property . 13. Chain of inclusions . . . . . . . . . . . . . . . . . . . . . . . E. Basic properties of Hodge ideals . . . . . . . . . . . . . 14. The ideals Jk(D). . . . . . . . . . . . . . . . . . . . . . . . . 2 9 9 11 13 16 16 17 18 18 20 20 21 23 24 25 26 26 27 2010 Mathematics Subject Classification. 14J17, 32S25, 14D07, 14F17. MM was partially supported by NSF grants DMS-1401227 and DMS-1265256; MP was partially supported by NSF grant DMS-1405516 and a Simons Fellowship. 1 2 M. MUSTAT¸ A AND M. POPA 15. Behavior under smooth pullback . . . . . . . . . . . . . . . 16. Restriction to hypersurfaces . . . . . . . . . . . . . . . . . . 17. Generation level of the Hodge filtration . . . . . . . . . . . 18. Behavior with respect to birational morphisms . . . . . . . F. Local study of Hodge ideals . . . . . . . . . . . . . . . 19. Order of vanishing along exceptional divisors . . . . . . . . 20. Examples . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21. Order of vanishing along a closed subset . . . . . . . . . . . G. Vanishing theorems . . . . . . . . . . . . . . . . . . . . 22. General vanishing . . . . . . . . . . . . . . . . . . . . . . . . 23. Vanishing for Hodge ideals . . . . . . . . . . . . . . . . . . . 28 30 34 40 45 45 49 58 60 61 61 24. Effective version . . . . . . . . . . . . . . . . . . . . . . . . 65 H. Vanishing on Pn and abelian varieties, with applications 66 25. Vanishing on Pn and toric varieties . . . . . . . . . . . . . . 66 26. Bounds for the subschemes associated to Hodge ideals in Pn 68 27. Singular points on hypersurfaces in Pn . . . . . . . . . . . . 70 28. Vanishing on abelian varieties . . . . . . . . . . . . . . . . . 71 29. Singularities of theta divisors . . . . . . . . . . . . . . . . . 30. Singular points on ample divisors on abelian varieties . . . . 73 76 I. Appendix: Higher direct images of forms with log poles 77 31. The case of SNC divisors on the base . . . . . . . . . . . . . 32. Akizuki-Nakano-type vanishing theorems . . . . . . . . . . . References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 77 81 83 A. Introduction Let X be a smooth complex variety of dimension n. To a reduced effective divisor D on X one associates the left DX -module of functions with poles along D, OX (∗D) = [k≥0 OX(kD), the localization of OX along D. In Saito's theory [Sai90], this D- i.e. module underlies the mixed Hodge module j∗QH U [n], where U = X r D and j : U ֒→ X is the inclusion map. It therefore comes with an attached Hodge filtration FkOX (∗D). Saito [Sai93] shows that this filtration is contained in HODGE IDEALS 3 the pole order filtration, namely (0.1) FkOX(∗D) ⊆ OX(cid:0)(k + 1)D(cid:1) for all k ≥ 0. The problem of how far these filtrations are from being equal is of great interest in the study of the singularities of D, and also in that of Deligne's Hodge filtration on the singular cohomology H•(U, C). The inclusion above leads to defining for each k ≥ 0 a coherent sheaf of ideals Ik(D) by the formula FkOX (∗D) = OX(cid:0)(k + 1)D(cid:1) ⊗ Ik(D). Our main goal in this paper is to approach the definition and study of these ideal sheaves using methods from birational geometry, and to put them to use in a number of applications regarding singularities and Hodge theory. In sequels to this article we will present a framework for Hodge ideals associated to Q-divisors and ideal sheaves, leading to further applications. Given a log resolution f : Y → X of the pair (X, D) which is an isomor- phism over X r D, we define E := (f∗D)red. We will see in §3 that there is a filtered complex of right f−1DX -modules 0 −→ f∗DX −→ Ω1 Y (log E) ⊗OY f∗DX −→ ··· ··· −→ Ωn−1 Y (log E) ⊗OY f∗DX −→ ωY (E) ⊗OY f∗DX −→ 0 which is exact except at the rightmost term, where the cohomology is DY →X; here DY →X = f∗DX is the transfer module of f . De- ωY (∗E) ⊗DY noting it by A•, its filtration is provided by the subcomplexes Fk−nA•, for every k ≥ 0, given by 0 −→ f∗Fk−nDX −→ Ω1 Y (log E) ⊗OY f∗Fk−n+1DX −→ ··· ··· −→ Ωn−1 (log E) ⊗ OY f∗Fk−1DX −→ ωY (E) ⊗OY f∗FkDX −→ 0. We define the k-th Hodge ideal Ik(D) associated to D by the formula Y ωX(cid:0)(k + 1)D(cid:1) ⊗ Ik(D) = Im(cid:2)R0f∗Fk−nA• −→ R0f∗A•(cid:3) , after proving that this image is contained in ωX(cid:0)(k + 1)D(cid:1). We show that this definition is independent of the choice of log resolution, and that it indeed coincides with the ideals defined by Saito's Hodge filtration. The 0th Hodge ideal belongs to a class of ideal sheaves that is quite well understood, and has proved to be extremely useful; it is not hard to show that I0(D) = I(cid:0)X, (1 − ǫ)D(cid:1), the multiplier ideal associated to the Q-divisor (1 − ǫ)D with 0 < ǫ ≪ 1. In particular, I0(D) = OX if and only if the pair (X, D) is log-canonical. Thus the sequence of ideals Ik(D) can be seen as a refinement of this type of multiplier ideal. Hodge ideals can be computed concretely when D is a simple normal In particular, if D is smooth, then crossing divisor; see Proposition 8.2. 4 M. MUSTAT¸ A AND M. POPA Ik(D) = OX for all k ≥ 0, which corresponds to equality between the Hodge filtration and the pole order filtration in (0.1). One of the main applications of the results below is an effective converse to this statement. Theorem A. Let X be a smooth complex variety of dimension n, and D a reduced effective divisor on X. Then the following are equivalent: (1) D is smooth. (2) the Hodge filtration and pole order filtration on OX(∗D) coincide. (3) Ik(D) = OX for all k ≥ 0. (4) Ik(D) = OX for some k ≥ n−1 2 . Saito introduced in [Sai09] a measure of the complexity of the Hodge filtration, and proved several results in the case of OX(∗D) (see e.g. Remark 20.11). Concretely, one says that the filtration F•OX(∗D) is generated at level k if FℓDX · FkOX (∗D) = Fk+ℓOX(∗D) for all ℓ ≥ 0. Equivalently, the ideal Ik(D) and the local equation of D determine all higher Hodge ideals, i.e. for ℓ ≥ 0, by the formula FℓDX ·(cid:0)OX(cid:0)(k + 1)D(cid:1) ⊗ Ik(D)(cid:1) = OX(cid:0)(k + ℓ + 1)D(cid:1) ⊗ Ik+ℓ(D). If D has simple normal crossings, the filtration is generated at level 0. It turns out that the same is true when X is a surface. This is a special case of the following general result, which is a consequence of our main criterion for detecting the generation level, Theorem 17.1 below; there exist simple examples in which one cannot do better. Theorem B. If X has dimension n ≥ 2, the Hodge filtration on OX(∗D) is generated at level n − 2. More generally, for every k ≥ 0 there exists an open subset Uk in X whose complement has codimension ≥ k + 3, such that the induced filtration on OX (∗D)Uk is generated at level k. Going back to the study of the singularities of the pair (X, D), the notion of log-canonical singularity is refined by the following: Definition. If D is a reduced effective divisor on the smooth variety X, we say that the pair (X, D) is k-log-canonical if I0(D) = I1(D) = ··· = Ik(D) = OX . We show in Proposition 13.1 that there is in fact a chain of inclusions ··· Ik(D) ⊆ ··· ⊆ I1(D) ⊆ I0(D). (Note that the definition gives automatically only an inclusion in the op- posite direction, namely Ik−1(D) ⊗ OX (−D) ⊆ Ik(D).) Thus being k-log- canonical is equivalent to Ik(D) = OX . Being log-canonical is of course equivalent to being 0-log-canonical in the above sense, while Theorem A says that (n− 1)/2-log-canonical or higher is HODGE IDEALS 5 equivalent to D being smooth. It turns out that any intermediate level of log- canonicity refines another basic notion, namely that of rational singularities. Recall that to D one can also associate the adjoint ideal adj(D), see [Laz04, §9.3.E], which is a concrete measure of the failure of D to have normal rational singularities. Theorem C. For every k ≥ 1 we have an inclusion Ik(D) ⊆ adj(D). Hence if Ik(D) = OX for some k ≥ 1, then D is normal with rational singularities. The condition of being k-log-canonical has Hodge-theoretic consequences for the cohomology H•(U, C), where U = X r D. Using the definition of Hodge ideals and Lemma 7.4 below, if X is smooth projective of dimension n we have that (0.2) D k−log − canonical =⇒ FpH i(U, C) = PpH i(U, C) ∀ p ≤ k−n, for all i, where F• is the Hodge filtration and P• is the pole order filtration on H i(U, C); see §7 for the definitions. One (difficult) calculation that we perform in §20 is the following; for the purpose of this paper, an ordinary singular point is a point whose projectivized tangent cone is smooth. Theorem D. Let D be a reduced effective divisor on a smooth variety X of dimension n, and let x ∈ D be an ordinary singular point of multiplicity m ≥ 2. Then In particular, if X is projective and D has only such singularities, then Ik(D)x = OX,x ⇐⇒ k ≤h n mi − 1. for all p ≤h n mi − n − 1. FpH•(U, C) = PpH•(U, C) When all the singularities of D are nodes, the equivalence in the theo- rem was established already in [DSW09, §1.4], where all Hodge ideals were computed concretely; see Example 20.10.1 It turns out that the nontriviality bound in Theorem D, i.e. the fact that Ik(D)x 6= OX,x for k ≥(cid:2) n m(cid:3), holds for any point x ∈ D of multiplicity m; see Example 21.4. This, as well as Theorem A, follows from the following statement, proved in §21 using deformation to ordinary singularities. In most cases, examples given in §20 show its optimality. Theorem E. Let D be a reduced effective divisor on a smooth variety X, and let W be an irreducible closed subset of codimension r, defined by the ideal sheaf IW . If m = multW (D), then for every k we have Ik(D) ⊆ I (q) W , where q = min{m − 1, (k + 1)m − r}. 1The paper [DSW09] also obtains a range where the equality FpH •(U, C) = PpH •(U, C) does not hold, for a general singular, hence nodal, hypersurface in Pn. In the case of nodal surfaces in P3, a more precise result can be found in [DS15, Theorem 5.1]. 6 M. MUSTAT¸ A AND M. POPA Here I (q) W is the qth symbolic power of IW , and I (q) W = OX if q ≤ 0. One ingredient in the proof of this result is the analogue of the Restriction Theorem for multiplier ideals, [Laz04, Theorem 9.5.1], which holds for all if H is a smooth hypersurface with H 6⊆ Supp(D), Hodge ideals as well: such that DH is reduced, then (0.3) Ik(H, DH ) ⊆ Ik(X, D) · OH, with equality for H sufficiently general. Thus there is an inversion of ad- junction for k-log-canonicity. The proof requires tools from the theory of mixed Hodge modules, and will be given in a separate paper [MP16]. We do include however a proof using the methods of this paper in the generic case, see Theorem 16.1. On the other hand, Theorem C is a consequence of another one of our main local results, Theorem 19.1, giving a lower bound for the order of vanishing of Ik(D) along exceptional divisors over X on carefully chosen log resolutions. We leave the slightly technical statement for the text, and note that it also leads to another nontriviality criterion, Corollary 19.4, that complements Theorem E. Just as in the theory of multiplier ideals, a precise measure of the nontriviality of Hodge ideals is crucial for applications, especially when combined with vanishing theorems; this is what we focus on next. Multiplier ideals satisfy the celebrated Nadel vanishing theorem; see [Laz04, Theorem 9.4.8]. For the ideal I0(D), this says that given any ample line bun- dle L, one has H i(cid:0)X, ωX (D) ⊗ L ⊗ I0(D)(cid:1) = 0 for all i ≥ 1. We obtain an analogous result for the entire sequence of Hodge ideals Ik(D). Things however necessarily get more complicated; in brief, in order to have full vanishing, higher log-canonicity conditions and borderline Nakano van- ishing type properties need to be satisfied. Theorem F. Let X be a smooth projective variety of dimension n, D a reduced effective divisor, and L a line bundle on X. Then, for each k ≥ 1, assuming that the pair (X, D) is (k − 1)-log-canonical we have: 2 , and L is a line bundle such that L(pD) is ample for all (1) If k ≤ n 0 ≤ p ≤ k, then for all i ≥ 2. Moreover, H i(cid:0)X, ωX ((k + 1)D) ⊗ L ⊗ Ik(D)(cid:1) = 0 H 1(cid:0)X, ωX ((k + 1)D) ⊗ L ⊗ Ik(D)(cid:1) = 0 holds if H j(cid:0)X, Ωn−j (2) If k ≥ n+1 X ⊗ L((k − j + 1)D)(cid:1) = 0 for all 1 ≤ j ≤ k. 2 , then D is smooth by Theorem A, and so Ik(D) = OX . In this case, if L is a line bundle such that L(kD) is ample, then H i(cid:0)X, ωX ((k + 1)D) ⊗ L(cid:1) = 0 for all i > 0. HODGE IDEALS 7 (3) If D is ample, then (1) and (2) also hold with L = OX . A slightly more precise statement for I1(D) is given in Theorem 23.2. The proof of Theorem F relies on the Kodaira-Saito vanishing theorem in the theory of mixed Hodge modules; at the moment we know how to give a more elementary proof only for I1(D). We explain in Corollary 24.1 how one can avoid the Nakano-type requirement in Theorem F by assuming that D is sufficiently positive with respect to an ample divisor A such that TX (A) is nef. It is worth noting that Hodge ideals also satisfy a local vanishing statement, Corollary 12.1, due to the strictness of the Hodge filtration. Vanishing for Hodge ideals takes a particularly simple form on Pn (see Corollary 25.3) or more generally on smooth toric varieties (see Corollary 25.1), and on abelian varieties (see Theorem 28.2), as in these cases the hypotheses are automatically satisfied or can be relaxed. As mentioned above, in combination with Theorem E and related results, this leads to interesting applications. We present a few here, while further applications, as well as theoretical statements, will be treated elsewhere. For instance, on Pn it is a consequence of Nadel vanishing that if an inte- gral hypersurface D of degree d is not log-canonical, then its singular locus has dimension ≥ n − d + 1. Our vanishing theorem leads to a simultaneous extension of this fact and of a result of Deligne on the Hodge filtration on complements of hypersurfaces with isolated singularities. When the Hodge ideals are nontrivial, it imposes further restrictions on the corresponding subschemes in Pn. Theorem G. Let D be a reduced hypersurface of degree d in Pn, and for each k denote by Zk the subscheme associated to the ideal Ik(D), and by zk its dimension. Then: (1) If zk < n − (k + 1)d + 1, then in fact Zk = ∅, i.e. (X, D) is k-log- (2) If zk ≥ n − (k + 1)d + 1, then canonical. The converse is of course true if n − (k + 1)d + 1 ≥ 0. deg Zk ≤(cid:18)(k + 1)d − 1 n − zk (cid:19), (3) The dimension 0 part of Zk imposes independent conditions on hy- with the convention that dim∅ and deg∅ are −1. persurfaces of degree at least (k + 1)d − n − 1. Part (1) says in particular that if D has isolated singularities, then Ik(D) = OX whenever n − (k + 1)d + 1 > 0; this is a result of Deligne, see [Sai93, 4.6(iii)], that was originally phrased in terms of the equality FkH i(U, C) = PkH i(U, C), where U = Pn r D. More generally, according to the theorem and (0.2), whenever dim Zk + (k + 1)d < n + 1 we have that Fk−nH i(U, C) = Pk−nH i(U, C) for all i. 8 M. MUSTAT¸ A AND M. POPA A related local result was proved by Saito [Sai93, Theorem 0.11] in terms of the roots of the Bernstein-Sato polynomial of D; see Remark 26.1. When combined with nontriviality criteria like Theorem E or Theorem D, part (3) in Theorem G has a number of basic consequences describing the behavior of isolated singular points on hypersurfaces in Pn, with n ≥ 3. These are collected in §27; here is the most easily stated: Corollary H. Let D be a reduced hypersurface of degree d in Pn, with n ≥ 3, and denote by Sm the set of isolated singular points on D of multiplicity m ≥ 2. Then Sm imposes independent conditions on hypersurfaces of degree at least ([ n m ] + 1)d − n − 1. To put this in perspective, recall that a classical theorem of Severi [Sev46] states that if a surface S ⊂ P3 of degree d has only nodes as singularities, then the set of nodes imposes independent conditions on hypersurfaces of degree 2d − 5. The bound above is one worse in this case, but it becomes better than what is known for most other n and m. For further discussion see §27. Results of a similar flavor hold on abelian varieties; see §30. On principally polarized abelian varieties (ppav's) we obtain an upper bound on the multiplicity of points on theta divisors whose singularities are isolated. Theorem I. Let (X, Θ) be ppav of dimension g such that Θ has isolated singularities. Then: (1) For every x ∈ Θ we have multx(Θ) ≤ g+1 2 , and also multx(Θ) ≤ ǫ(Θ) + 2, where ǫ(Θ) is the Seshadri constant of the principal polar- ization. (2) Moreover, there is at most one point x ∈ Θ with multx(Θ) = g+1 2 . See §29 for a detailed discussion, including the conjectural context in which this result is placed, and for further numerical bounds. Suffice it to say here that, in the case of isolated singularities, this improves a well-known bound of Koll´ar saying that the multiplicity of each point is at most g. See also Remark 29.6 (1) for related work of Codogni-Grushevsky-Sernesi and communications from Lazarsfeld. We conclude by noting that some of the statements in the text rely on local vanishing theorems of Akizuki-Nakano type for higher direct images of sheaves of differentials with log poles; some of these can already be found in [EV82] and [Sai07], while some are new and hopefully be of interest beyond the applications in this paper. We provide a uniform approach in the Appendix, using rather elementary arguments. Finally, a word about the use of results from the theory of mixed Hodge modules; in the treatment given here many of the definitions, including that of Hodge ideals, as well as the proofs of most theorems, do not a priori depend of it. However, this is not always the case. For instance, the proof HODGE IDEALS 9 of Theorem F on vanishing, and that of Proposition 13.1, rely essentially on results regarding Hodge modules. One topic that does not appear in this paper though, is the connection with the theory of the V -filtration. This is used in [MP16] in order to prove Theorem 16.9 and further results. It is treated systematically in the more recent preprint of Saito [Sai16b], where in particular it leads to a different approach to many of the results in Theorems A, C and D; see also Remark 20.11. Both points of view will continue to play an important role in the sequels mentioned above. Acknowledgements. This paper owes a great deal of inspiration to Mori- hiko Saito's work on the Hodge filtration on localizations along hypersufaces [Sai93], [Sai07], [Sai09]. We are grateful to Christian Schnell for making us aware of these papers, which is one of the reasons our project got started. He also asked whether the equivalence between (1) and (2) in Theorem A might hold. We thank him, Rob Lazarsfeld, and Morihiko Saito for many other useful comments, Lawrence Ein for discussions regarding the results in the Appendix, and H´el`ene Esnault, Sam Grushevsky, Sam Payne, and Claire Voisin for comments and references. The second author would like to thank the math departments at the University of Michigan and Stony Brook University for hospitality during the preparation of the paper, and the Simons Foundation for fellowship support. B. Preliminaries Let X be a smooth complex algebraic variety. We denote by DX the sheaf of differential operators on X. A left or right D-module on X is simply a left, respectively right, DX -module. 1. Background on filtered D-modules. We briefly recall some standard definitions from the theory of D-modules that will be used in this paper. A very good source for further details is [HTT08]. Left-right correspondence. We will work with both left and right D- modules; while most definitions and results are best stated for left D- modules, push-forwards are most natural in the context of right D-modules. The standard one-to-one correspondence between left and right DX -modules is given by M 7→ N := M ⊗ OX ωX and N 7→ M := HomOX (ωX,N ), where ωX is endowed with its natural right DX -module structure; see [HTT08, §1.2]. Filtrations and the de Rham complex. A filtered left D-module on X is a left DX -module M with an increasing filtration F = F•M by coherent OX-modules, bounded from below and satisfying Fk DX · FℓM ⊆ Fk+ℓM for all k, ℓ ∈ Z, 10 M. MUSTAT¸ A AND M. POPA where Fk DX is the sheaf of differential operators on X of order ≤ k. We use the notation (M, F ) for this data. The filtration is called good if the inclusions above are equalities for ℓ ≫ 0, which is in turn equivalent to the fact that the total associated graded object grF • M =Mk grF k M =Mk FkM/Fk−1M is finitely generated over grF DX ≃ Sym TX. With the analogous definitions • for a right DX-module N , in case it corresponds to M via the left-right operation, the corresponding rule on filtrations is FpM = Fp−nN ⊗OX ω−1 X . Given a left DX -module M, the associated de Rham complex is: X ⊗ M → ··· → Ωn DR(M) =hM → Ω1 It is a C-linear complex with differentials induced by the corresponding integrable connection ∇ : M → Ω1 X ⊗ M. We consider it to be placed in degrees −n, . . . , 0. (Strictly speaking, as such it is the de Rham complex associated to the corresponding right D-module.) A filtration F•M on M induces a filtration on the de Rham complex of M by the formula X ⊗ Mi. Fk DR(M) =hFkM → Ω1 k DR(M) =h grF k M → Ω1 grF X ⊗ Fk+nMi. k+n Mi. X ⊗ grF For any integer k, the associated graded complex for this filtration is X ⊗ Fk+1M → ··· → Ωn X ⊗ grF k+1 M → ··· → Ωn This is now a complex of coherent OX -modules in degrees −n, . . . , 0, provid- ing an object in Db(X), the bounded derived category of coherent sheaves on X. Transfer modules and pushforward. If f : Y → X is a morphism of smooth complex varieties, we consider the associated transfer module DY →X := OY ⊗f −1OX f−1DX . It has the structure of a DY − f−1DX -bimodule; moreover, it is filtered by f∗FkDX . It is simply f∗DX as an OY -module, and we will use this notation rather than DY →X when thinking of it as such. For a right DY -module M, due to the left exactness of f∗ versus the right exactness of ⊗, the appropriate pushforward for D-modules is at the level of derived categories, namely f+ : D(DY ) −→ D(DX ), M• 7→ Rf∗(cid:0)M• See [HTT08, §1.5] for more details, where this last functor is denoted byRf . DY →X(cid:1). L ⊗DY HODGE IDEALS 11 2. Localization along a divisor. In this section, X will always be a smooth complex variety of dimension n, and D a reduced effective divisor on X. We denote the localization of OX along D, as a left DX -module, by OX(∗D). If h is a local equation of D, then this is OX[ 1 h ], with the obvious action of differential operators. The associated right DX -module is denoted ωX(∗D). We will use the following well-known observation; see [HTT08, Lemma 5.2.7]. Lemma 2.1. If F is a coherent OX (∗D)-module supported on D, then F = 0. Let us now fix a proper morphism f : Y −→ X which is an isomorphism over U := X r D. We assume that Y is smooth, and let E = (f∗D)red and V = Y r E = f−1(U ). We denote by jU and jV the inclusions of U and V into X and Y respectively. Note that jU +ωU ≃ ωX(∗D) and jV +ωV ≃ ωY (∗E). Since jU = f ◦ jV , we have: Lemma 2.2. There is a natural isomorphism f+ωY (∗E) ≃ H 0f+ωY (∗E) ≃ ωX(∗D). Given the morphism f : Y → X, since DY →X is a left DY -module, we have a canonical morphism of left DY -modules ϕ : DY −→ f∗DX that maps 1 to 1. This is in fact a morphism of DY − f−1OX bimodules. It is clearly an isomorphism over Y r E. Since DY is torsion-free, we conclude that ϕ is injective, with cokernel supported on E. For each k, we have induced inclusions FkDY ֒→ f∗FkDX of OY − f−1OX bimodules. Lemma 2.3. The canonical map of right f−1OX -modules induced by ϕ is a split injection. ωY (∗E) −→ ωY (∗E) ⊗DY DY →X Proof. Denoting for simplicity j = jV , we have a commutative diagram ωY (∗E) ϕ j∗j∗ωY (∗E) α β ωY (∗E) ⊗DY DY →X ψ j∗j∗(cid:0)ωY (∗E) ⊗DY DY →X(cid:1), where ϕ and ψ are the canonical maps. Since DY →XV = DV , it is clear that β = j∗j∗(α) is an isomorphism. On the other hand, as OY -modules we have ωY (∗E) ≃ j∗ωV , hence ϕ is an isomorphism. It follows from the above diagram that the composition ψ ◦ α is an isomorphism, hence α is a split injection. (cid:3) 12 M. MUSTAT¸ A AND M. POPA The main result we are aiming for here is the following enhancement: Proposition 2.4. The canonical morphism induced by ϕ in the derived category of right f−1OX-modules ωY (∗E) −→ ωY (∗E) L ⊗DY DY →X is an isomorphism. The proof we give below is inspired in part by arguments in [HTT08, §5.2]. Lemma 2.5. The induced morphism Id ⊗ ϕ : OY (∗E) ⊗OY is an isomorphism. DY −→ OY (∗E) ⊗OY f∗DX Proof. It suffices to show that the induced mappings OY (∗E) ⊗OY FkDY −→ OY (∗E) ⊗OY f∗FkDX are all isomorphisms for k ≥ 0. But this follows immediately from Lemma 2.1 (note that since OY (∗E) is flat over OY , these maps are injective). (cid:3) We will use the notation DY (∗E) := OY (∗E) ⊗OY DY . This is a sheaf of rings, and one can identify it with the subalgebra of EndC(cid:0)OY (∗E)(cid:1) generated by DY and OY (∗E). Note that since OY (∗E) is DY . A basic fact is the a flat OY -module, we have DY (∗E) ≃ OY (∗E) following: ⊗OY L Lemma 2.6. The canonical morphism induced by ϕ is an isomorphism. DY (∗E) −→ DY (∗E) L ⊗DY DY →X L ⊗OY L L Proof. Via the isomorphism DY (∗E) ≃ OY (∗E) the statement gets identified to the morphism (2.7) ⊗OY f∗DX OY (∗E) induced by ϕ. Moreover, since OY (∗E) is a flat OY -module, the morphism (2.7) gets identified with the isomorphism in Lemma 2.5. DY →X = OY (∗E) DY −→ OY (∗E) DY , the morphism in ⊗DY ⊗OY ⊗OY DY L L (cid:3) Proof of Proposition 2.4. Via the right D-module structure on ωY , we have that ωY (∗E) has a natural right DY (∗E)-module structure. The morphism in the proposition gets identified with the morphism L L L ωY (∗E) ⊗DY (∗E) DY (∗E) −→ ωY (∗E) ⊗DY (∗E) DY (∗E) ⊗DY DY →X HODGE IDEALS 13 induced by ϕ. In turn, this is obtained by applying ωY (∗E) the isomorphism in Lemma 2.6, hence it is an isomorphism. L ⊗DY (∗E) − to (cid:3) 3. Filtrations on localizations and tensor products. We next include filtrations in the discussion. We fix again a smooth variety X of dimension n, and a reduced effective divisor D on X. Most obviously, on OX (∗D) there is a pole order filtration, whose nonzero terms are PkOX(∗D) = OX(cid:0)(k + 1)D(cid:1) for k ≥ 0. Less obvious is the Hodge filtration FkOX (∗D), again with nonzero terms for k ≥ 0. This is our main topic of study in this paper. Its existence is guaranteed by general results on Hodge modules (see §4). However, we will take a hands-on approach and describe it explicitly now in the simple normal crossings case, and later in the general case via log resolutions. For simple normal crossing divisors we fix different notation, in view of later use on log resolutions (note also the shift from left to right D-modules). Let E be a reduced simple normal crossing (SNC) divisor on a smooth n- dimensional variety Y . We define the Hodge filtration on the right DY - module ωY (∗E) to be given by For instance, the first two nonzero terms are FkωY (∗E) = ωY (E) · Fk+nDY . F−nωY (∗E) = ωY (E) and F−n+1ωY (∗E) = ωY (2E) ⊗ Jac(E), where Jac(E) is the Jacobian ideal of E, i.e. F1DY · OY (−E). See also Proposition 8.2 for a general local description. Recall that the right D-module ωY has a standard resolution 0 → DY → Ω1 Y ⊗OY DY → ··· → ωY ⊗OY DY → ωY → 0 by induced DY -modules; see [HTT08, Lemma 1.2.57]. The following gener- alization will be a crucial technical point later on; cf. also [Sai90, Proposition 3.11(ii)], where this is part of a more general picture. Proposition 3.1. The right DY -module ωY (∗E) has a filtered resolution with induced DY -modules given by 0 → DY → Ω1 Here the morphism DY → ··· → ωY (E) ⊗OY DY → ωY (∗E) → 0. Y (log E) ⊗OY is given by ω ωY (E) ⊗OY DY −→ ωY (∗E) f ⊗ P → ω f · P , and for each p the morphism (log E) ⊗OY DY −→ Ωp+1 Ωp Y (log E) ⊗OY Y DY is given by ω ⊗ P → dω ⊗ P +Pn z1, . . . , zn. i=1(dzi ∧ ω) ⊗ ∂iP , in local coordinates 14 M. MUSTAT¸ A AND M. POPA Proof. It is not hard to check that the expression in the statement is indeed a complex, which we call A•. We consider on Ωp FiΩp Y (log E) =(Ωp 0 Y (log E) Y (log E) the filtration if i ≥ −p if i < −p, DY the tensor product filtration. This filters A• by (log E) ⊗OY Fk−1DY → ωY (E) ⊗OY FkDY → FkωY (∗E) → 0 Y Y (log E) ⊗OY and on Ωp subcomplexes Fk−nA• given by ··· → Ωn−1 for each k ≥ 0. Note that they can be rewritten as ··· → ωY (E)⊗TY (− log E)⊗OY Fk−1DY where TY (− log E) is the dual of Ω1 ωY (E) ⊗ ∧iTY (− log E) ≃ Ωn−i (log E). Y βk→ ωY (E)⊗OY FkDY → FkωY (∗E) → 0, Y (log E), and we use the isomorphisms It is clear directly from the definition that every such complex is exact at the term FkωY (∗E). We now check that they are exact at the term ωY (E) ⊗OY FkDY . Let us assume that, in the local coordinates z1, . . . , zn, the divisor E is given by z1 ··· zr = 0. Using the notation ω = dz1∧···∧dzn, we consider an element u = gα∂α ω z1 ··· zr ⊗ Xα≤k α1!··· αr! · gα · z−α1 1 mapping to 0 in FkωY (∗E) = ωY (E) · Fk DY . This means that r = 0. ··· z−αr Xα≤k, αi=0 if i>r We show that u is in the image of the morphism βk by using a descending induction on α. What we need to prove is the following claim: for each α in the sum above, with α = k, there exists some i with αi > 0 such that zi divides gα. If so, an easy calculation shows that the term uα = ω z1···zr ⊗ gα∂α is in the image of βk, and hence it is enough to prove the statement for u − uα. Repeating this a finite number of times, we can reduce to the case when all α ≤ k − 1. But the claim is clear: if zi did not divide gα for all i with αi > 0, then the Laurent monomial z−α1 would appear in the term gα · z−α1 of the sum above, but in none of the other terms. ··· z−αr ··· z−αr To check the rest of the statement, note that after discarding the term on 1 1 r r the right, the associated graded complexes ··· −→ ωY (E) ⊗ 2^ TY (− log E) ⊗OY Sk−2TY −→ −→ ωY (E) ⊗ TY (− log E) ⊗OY Sk−1TY −→ ωY (E) ⊗OY SkTY −→ 0 are acyclic. Indeed, each such complex is, up to a twist, an Eagon-Northcott complex associated to the inclusion of vector bundles of the same rank ϕ : TY (− log E) → TY . HODGE IDEALS 15 Concretely, in the notation on [Laz04, p.323], the complex above is (ENk) tensored by ωY (E). According to [Laz04, Theorem B.2.2(iii)], (ENk) is acyclic provided that codim Dn−ℓ(ϕ) ≥ ℓ for all 1 ≤ ℓ ≤ min{k, n}, where are the deneracy loci of ϕ. But locally ϕ is given by the diagonal matrix Ds(ϕ) = {y ∈ Y rk(ϕy) ≤ s} so this condition is verified by a simple calculation. (cid:3) Diag(z1, . . . , zr, 1, . . . , 1) Let now X and D be as at the beginning of the section, and let f : Y → X be a log resolution of the pair (X, D) which is an isomorphism over X r D. Under the latter assumption, the log resolution condition simply means that f is a projective morphism, Y is smooth, and E := (f∗D)red has simple normal crossings. On the tensor product ωY (∗E) ⊗DY DY →X we consider the tensor product filtration, that is, Fk(cid:0)ωY (∗E) ⊗DY DY →X(cid:1) := FiωY (∗E) ⊗OY f∗Fk−iDX → ωY (∗E) ⊗DY = ImMi≥−n DY →X , where the map in the parenthesis is the natural map between the tensor product over OY and that over DY . Lemma 3.2. The definition above simplifies to Fk(cid:0)ωY (∗E)⊗DY DY →X(cid:1) = Im(cid:2)ωY (E)⊗OY f∗Fk+nDX → ωY (∗E)⊗DY DY →X(cid:3). Proof. Fix i ≥ −n and recall that FiωY (∗E) = ωY (E) · Fi+nDY . The factor Fi+nDY can be moved over the tensor product once we pass to the image in the tensor product over DY , and moreover we have an inclusion Fi+nDY · f∗Fk−iDX ⊆ f∗Fk+nDX. Therefore inside ωY (∗E)⊗DY is contained in the image of ωY (E) ⊗OY f∗Fk+nDX . DY →X, the image of FiωY (∗E)⊗OY f∗Fk−iDX (cid:3) Propositions 3.1 and 2.4 have the following immediate consequence: Corollary 3.3. On Y there is a filtered complex of right f−1DX -modules 0 → f∗DX → Ω1 Y (log E) ⊗OY f∗DX → ··· ··· → Ωn−1 which is exact (though not necessarily filtered exact). (log E)⊗OY f∗DX → ωY (E)⊗OY f∗DX → ωY (∗E)⊗DY Y DY →X → 0 16 M. MUSTAT¸ A AND M. POPA Proof. It follows from Proposition 3.1 that the complex 0 → f∗DX → Ω1 Y (log E) ⊗OY f∗DX → ··· → ωY (E) ⊗OY f∗DX → 0 represents the object ωY (∗E) DY →X in the derived category, hence Proposition 2.4 implies the exactness of the entire complex in the statement. (cid:3) ⊗DY L We record here the following lemma for later use. Lemma 3.4. The morphism Ωn−1 Y (log E) −→ ωY (E) ⊗OY f∗F1DX is injective. Proof. Note that we have a commutative diagram Ωn−1 Y (log E) Id Ωn−1 Y (log E) β α ωY (E) ⊗OY F1DY γ ωY (E) ⊗OY f∗F1DX , in which γ is the canonical inclusion. Since β is injective by Proposition 3.1, it follows that α is injective, too. (cid:3) C. Saito's Hodge filtration and Hodge modules This section, unlike the previous one, only contains review material. The reader familiar with the topics it covers can skip to Section D, and use it as a reference. The study of Hodge ideals relies in part on the fact that the filtered DX - module ωX(∗D) underlies a mixed Hodge module. For simplicity, we will call such objects Hodge D-modules. It is not the place here to give a detailed account of the theory of Hodge modules; for details we refer to the original [Sai88], [Sai90], for summaries of the results needed here to the surveys [Sai16a] and [Sch14a], and for a review of how they have been recently used for geometric applications to [Pop16]. We will however review properties of Hodge D-modules that make them special among all filtered D-modules, and that will be used here, as well as some results specific to ωX(∗D). 4. Hodge D-modules and strictness. If we denote by FM(DX ) the category of filtered DX -modules, one can construct an associated bounded derived category Db(cid:0)FM(DX )(cid:1). Assuming that we are given a projective morphism of smooth varieties f : Y → X, and that we are working with right D-modules, Saito constructs in [Sai88, §2.3] a filtered direct image functor f+ : Db(cid:0)FM(DY )(cid:1) → Db(cid:0)FM(DX )(cid:1), compatible with the usual direct image functor for right D-modules. HODGE IDEALS 17 L A fundamental result about Hodge D-modules is Saito's Stability Theo- rem for direct images under projective morphisms, [Sai88, Th´eor`eme 5.3.1]. This says that, in the above setting, if (M, F ) is a Hodge DY -module, then f+(M, F ) is strict as an object in Db(cid:0)FM(DX )(cid:1) (and moreover, each H if+(M, F ) is a Hodge DX -module). This means that the natural mapping (4.1) is injective for every i, k ∈ Z. Taking FkH if+(M, F ) to be the image of this map, we get the filtration on H if+(M, F ). Example 4.2. Strictness in this context can be seen as a generalization of the degeneration at E1 of the classical Hodge-to-de Rham spectral sequence. Concretely, let X be a smooth projective variety, and (M, F ) a Hodge D- module on X. The natural inclusion of complexes Fk DR(M) ֒→ DR(M) induces, after passing to cohomology, a morphism DY →X)(cid:1) −→ Rif∗(M Rif∗(cid:0)Fk(M DY →X) ⊗DY ⊗DY L ϕk,i : H i(cid:0)X, Fk DR(M)(cid:1) −→ H i(cid:0)X, DR(M)(cid:1). Now for the constant map f : X → pt, the definition of pushforward gives f+M ≃ RΓ(cid:0)X, DR(M)(cid:1), and by the discussion above, the image of ϕk,i is FkH i(cid:0)X, DR(M)(cid:1). Saito's result on the strictness of f+(M, F ) then implies that ϕk,i is injective for all k and i, which is in turn equivalent to grF k H i(cid:0)X, DR(M)(cid:1) ≃ Hi(cid:0)X, grF k DR(M)(cid:1). This is the same as the E1-degeneration of the Hodge-to-de Rham spectral sequence Ep,q 1 = Hp+q(cid:0)X, grF −q DR(M)(cid:1) =⇒ H p+q(cid:0)X, DR(M)(cid:1). 5. Vanishing theorem. Recall from §1 that given a DX -module with good filtration (M, F ) on a smooth variety X, for any integer k the associated graded complex for the induced filtration on the de Rham complex is (1) Hi(cid:0)X, grF (2) Hi(cid:0)X, grF k DR(M) ⊗ L(cid:1) = 0 for all i > 0. k DR(M) ⊗ L−1(cid:1) = 0 for all i < 0. grF k DR(M) =h grF k M → Ω1 X ⊗ grF k+1 M → ··· → Ωn X ⊗ grF k+n Mi, seen as a complex of coherent OX-modules in degrees −n, . . . , 0. When X is projective and (M, F ) underlies a mixed Hodge module, these complexes satisfy the following Kodaira-type vanishing theorem [Sai90, §2.g] (see also [Pop14] and [Sch14b]). A similar result can be formulated more generally, on singular projective varieties, but we will not make use of this here. Theorem 5.1 (Saito). Let (M, F ) be a Hodge D-module on a smooth pro- jective variety X, and let L be any ample line bundle. Then: 18 M. MUSTAT¸ A AND M. POPA 6. Localization as a Hodge D-module. If X is a smooth variety, and D is a reduced effective divisor on X, then the right DX -module ωX(∗D) is a Hodge D-module. Indeed, it underlies the mixed Hodge module j∗QH U [n], where j : U = X r D ֒→ X is the inclusion, and QH U [n] is the trivial Hodge module on U ; see e.g. [Sch14b, Example 5.4]. Let f : Y → X be a log resolution of the pair (X, D) which is an iso- morphism over X r D, and let E = (f∗D)red. If V = Y r E, the proof of Lemma 2.2 shows more generally that we have an isomorphism of mixed Hodge modules V [n] ≃ jU∗QH U [n], so in particular there is an isomorphism f∗jV ∗QH f+(cid:0)ωY (∗E), F(cid:1) ≃(cid:0)ωX(∗D), F(cid:1) Fk−nωX(∗D) ≃ Fk−nH 0f+ωY (∗E) = DY →X(cid:1) → R0f∗(cid:0)ωY (∗E) in Db(cid:0)FM(DX )(cid:1). According to §4, the filtration on the right hand side is = Im(cid:20)R0f∗Fk−n(cid:0)ωY (∗E) DY →X(cid:1)(cid:21) and the mapping in the parenthesis is in fact injective. given by L ⊗DY L ⊗DY We note also that, although not used in the sequel, the following result of Saito is suggestive for some of the constructions below. Theorem 6.1 ([Sai07, Theorem 1]). There is an isomorphism Rf∗(cid:0)Ω•+n Y (log E), F(cid:1) ≃ DR(cid:0)OX(∗D), F(cid:1) in the derived category of filtered complexes of C-vector spaces (and more generally of filtered differential complexes), where the filtration on the log-de Rham complex on the left is the "stupid" filtration. The log-de Rham complex on the left is of course also filtered quasi- isomorphic to DR(cid:0)OY (∗E), F(cid:1), by a well-known result of Deligne [Del70]. Remark 6.2. In particular, Saito deduces from Theorem 6.1 that the object Rf∗Ω•Y (log E) is independent of the log resolution, and that Rqf∗Ωp Y (log E) is 0 for p + q > n. A different proof, together with other results on forms with log-poles, is given in the Appendix. 7. Hodge filtration on the complement. In this section we assume that X is a smooth projective variety. In Hodge theory it is of interest to understand Deligne's Hodge filtration on H•(U, C); see for instance [DD90] and [Sai93]. Saito showed that there is a close relationship between this Hodge filtration and the ideals Ik(D), or equivalently FkOX (∗D). Lemma 7.1 ([Sai93, 4.6(ii)]). For every integer i there is a natural mor- phism ϕi : H i(cid:0)X, DR(OX (∗D))(cid:1) −→ H i+n(U, C), HODGE IDEALS 19 which is a filtered isomorphism. Here the right hand side is endowed with Deligne's Hodge filtration, and the left hand side with Saito's filtration given by the image of H i(cid:0)X, F• DR(OX (∗D))(cid:1). above is the cohomology Now according to Saito's theory, the left-hand side in the isomorphism H iaX∗(cid:0)OX (∗D), F(cid:1) of the filtered direct image, where aX : X → pt. This filtration is strict, and we saw in Example 4.2 that this is equivalent to the degeneration at E1 of the Hodge-to-de Rham spectral sequence Ep,q 1 = Hp+q(cid:0)X, grF −q DR(OX (∗D))(cid:1) =⇒ H p+q(cid:0)X, DR(OX (∗D))(cid:1). As a consequence one obtains Corollary 7.2. For every integer k there is a decomposition H i+n(U, C) ≃Mq∈Z Hi(cid:0)X, grF −q DR(OX (∗D))(cid:1). The spaces H•(U, C) also have a pole order filtration. Indeed, using the pole order filtration on OX (∗D), i.e. PkOX(∗D) := OX(cid:0)(k + 1)D(cid:1), we obtain a filtration on the de Rham complex Pk DR(OX (∗D)) = X ⊗ Pk+1OX (∗D) → ··· → Ωn =hPkOX(∗D) → Ω1 X ⊗ Pk+nOX (∗D)i. For each i ∈ Z, we define the pole order filtration on H i+n(U, C) by P•H i+n(U, C) := Im(cid:2)H i(cid:0)X, P• DR(OX (∗D))(cid:1) −→ H i+n(U, C)(cid:3) , where the image is considered via the isomorphism ϕi. This is defined in a slightly different, but equivalent fashion by Deligne-Dimca [DD90], and the main result of that paper is the inclusion FkH j(U, C) ⊆ PkH j(U, C) for all j and k. Using Lemma 7.1, this also follows from the stronger state- ment FkOX (∗D) ⊆ PkOX(∗D), which is proved in [Sai93, Proposition 0.9]; for a different proof see also Lemma 9.2 below. Remark 7.3. The lowest k for which FkH•(U, C) is not automatically zero is k = −n, and similarly for Pk. Note that P−nH i+n(U, C) = Im(cid:2)H i(cid:0)X, ωX (D)(cid:1) −→ H i+n(U, C)(cid:3) . On the other hand, we will see later that F−nH i+n(U, C) ≃ H i(cid:0)X, ωX (D) ⊗ I0(D)(cid:1), 20 M. MUSTAT¸ A AND M. POPA where I0(D) = I(cid:0)X, (1−ǫ)D(cid:1), the multiplier ideal of the Q-divisor (1−ǫ)D, with 0 < ǫ ≪ 1. By Corollary 7.2 this is a direct summand of H i+n(U, C); it may be different from P−nH i+n(U, C) if the pair (X, D) is not log-canonical. In any case, it is clear from the descriptions above that any statement re- lating the two filtrations on OX(∗D) automatically leads to a similar state- ment for those on H•(U, C). For instance: Lemma 7.4. If FkOX(∗D) = PkOX(∗D) for k ≤ ℓ + n, then FkH j(U, C) = PkH j(U, C) for all j and all k ≤ ℓ. D. Birational definition of Hodge ideals Throughout this section X is a smooth complex variety of dimension n and D is a reduced effective divisor on X. We define the Hodge ideals Ik(D) associated to D, for k ≥ 0, first explicitly in the simple normal crossing case, and then in general in terms of log resolutions. We show directly that they are independent of the choice of log resolution, and then note that they coincide with the ideals defined by Saito's Hodge filtration. We establish a few first properties of these ideals. (8.1) 8. The simple normal crossing case. When D is a simple normal cross- ing divisor, we define the ideals Ik(D) by the following expression: for all k ≥ 0, OX ((cid:0)k + 1)D(cid:1) ⊗ Ik(D) = FkDX · OX (D) where the left-hand side is considered via the natural injective image in OX(∗D). Note that this includes the statement I0(D) = OX , and that a clear from definition that if we use the filtration on ωX(∗D) introduced in §3, then simple local calculation shows that FkDX · OX(D) ⊆ OX(cid:0)(k + 1)D(cid:1). It is Fk−nωX(∗D) = ωX((k + 1)D) ⊗ Ik(D). Proposition 8.2. Suppose that around a point p ∈ X we have coordinates x1, . . . , xn such that D is defined by (x1 ··· xr = 0). Then, for every k ≥ 0, the ideal Ik(D) is generated around p by {xa1 1 ··· xar r 0 ≤ ai ≤ k,Xi ai = k(r − 1)}. In particular, if r = 1 (that is, when D is smooth), we have Ik(D) = OX and if r = 2, then Ik(D) = (x1, x2)k. Proof. It is clear that FkDX · OX (D) is generated as an OX -module by {x−b1 1 ··· x−br r bi ≥ 1, Xi bi = r + k}. HODGE IDEALS 21 According to (8.1), the expression for Ik(D) now follows by multiplying these generators by (x1 ··· xr)k+1. The assertions in the special cases r = 1 and r = 2 are clear. (cid:3) 9. The general case. When D is arbitrary, we consider a log resolution f : Y → X of the pair (X, D) which is an isomorphism over X r D, and let E = (f∗D)red. Note that by assumption E has simple normal crossings. Because we need to deal with pushforwards, we will work in the setting of right D-modules. We denote by A• the complex 0 → f∗DX → Ω1 Y (log E) ⊗OY f∗DX → ··· → ωY (E) ⊗OY f∗DX → 0 placed in degrees −n, . . . , 0. We have seen in §3 that it comes with a filtra- tion, and that as such it represents the object ωY (∗E) DY →X in the derived category of filtered right f−1DX -modules. In particular, we have ⊗DY L H 0A• ≃ ωY (∗E) ⊗DY DY →X. Moreover, by Corollary 3.3, we know that if we ignore the filtration, A• is exact everywhere except at the last term on the right, i.e. the natural mapping A• −→ H 0A• is a quasi-isomorphism. For every k ≥ 0, we also consider the subcomplex C•k−n = Fk−nA• of A•, given by 0 → f∗Fk−nDX → Ω1 Y (log E) ⊗OY f∗Fk−n+1DX → ··· ··· → Ωn−1 Y (log E) ⊗OY f∗Fk−1DX → ωY (E) ⊗OY f∗FkDX → 0. For every k ≥ 0, the inclusion C•k−n ֒→ A• induces a canonical morphism of (quasi-coherent) OX-modules (9.1) R0f∗C•k−n → R0f∗A• ≃ ωX(∗D), where we recall that Proof. Since D is reduced, we can find an open subset U ⊆ X with the property that codim(X r U, U ) ≥ 2, the induced morphism f−1(U ) → U is an isomorphism, and DU is a smooth (possibly disconnected) divisor. Let R0f∗A• ≃ R0f∗(H 0A•) ≃ f∗(cid:0)ωY (∗E) ⊗DY DY →X(cid:1) ≃ ωX(∗D). Let Fk−nωX(∗D) ⊆ ωX(∗D) be the image of this map. Since C•k−n is a complex of quasi-coherent f−1OX -modules, it follows that Fk−nωX(∗D) is a quasi-coherent OX -module. Lemma 9.2. For every k ≥ 0, we have an inclusion Fk−nωX(∗D) ⊆ ωX(cid:0)(k + 1)D(cid:1). 22 M. MUSTAT¸ A AND M. POPA j : U ֒→ X be the inclusion. By assumption, on f−1(U ) we have DY →X = DY . From Proposition 3.1, on U we obtain Fk−nωX(∗D) = ωX(cid:0)(k + 1)D(cid:1). Now Fk−nωX(∗D) is torsion-free, being a subsheaf of ωX(∗D), and so the following canonical map is injective: This completes the proof of the lemma. Fk−nωX(∗D) → j∗(Fk−nωX(∗D)U ) = j∗(cid:0)ωX(cid:0)(k+1)D(cid:1)U(cid:1) = ωX(cid:0)(k+1)D(cid:1). Remark 9.3. The inclusion Fk−nωX(∗D) ⊆ ωX(cid:0)(k+1)D(cid:1) given by Lemma 9.2 is equivalent to the inclusion of the Hodge filtration in the pole order filtra- tion, i.e. (cid:3) proved in [Sai93, Proposition 0.9] using the V -filtration; see §12. Fk OX(∗D) ⊆ OX(cid:0)(k + 1)D(cid:1), We can now introduce the main objects we are concerned with in this paper. Definition 9.4 (Hodge ideals). Given the inclusion in Lemma 9.2, for each k ≥ 0 we define the ideal sheaf Ik(D) on X by the formula Fk−nωX(∗D) = ωX(cid:0)(k + 1)D(cid:1) ⊗ Ik(D). We call Ik(D) the k-th Hodge ideal of D. We will show in Theorem 11.1 that the definition is independent of the choice of log resolution. We end this section by mentioning another sequence of ideals that can be defined in this context. We discuss them only briefly, since they will not play an important role in what follows; it is however a somewhat more intuitive definition that helps with a first approximation understanding of the Hodge ideals. For every k ≥ 0, define Fk−n := H 0C•k−n = = Coker(cid:2)Ωn−1 Y (log E) ⊗OY f∗Fk−1DX → ωY (E) ⊗OY f∗FkDX(cid:3) . The morphism C•k−n → A• induces a morphism f∗Fk−n = f∗H 0C•k−n → f∗H 0A• = ωX(∗D), whose image we denote by F k−n. An argument similar to that in Lemma 9.2 shows that hence there is a coherent ideal I f It is easy to see that F k−n ⊆ ωX(cid:0)(k + 1)D(cid:1), F k−n = ωX(cid:0)(k + 1)D(cid:1) ⊗ I f k (D) of OX such that k (D). Ik(D) ⊆ I f k (D) for every k. Indeed, we have a commutative diagram HODGE IDEALS 23 C•k−n H 0C•k−n A• ϕ H 0A• in which ϕ is a quasi-isomorphism. This induces a commutative diagram R0f∗C•k−n f∗H 0C•k−n α β R0f∗A• δ f∗H 0A•, hence via the isomorphism δ we have Fk−nωX(∗D) = Im(α) ⊆ Im(β) = F k−n. We do not know whether I f k (D) is independent of resolution. If k = 0 or k = 1, then Ik(D) = I f k (D). This is a consequence of the fact that the canonical morphism C•k−n → H 0C•k−n is a quasi-isomorphism in these cases (this is trivial for k = 0 and it is a consequence of Lemma 3.4 for k = 1). At the moment we do not know however whether this equality also holds for higher k; this is an intriguing question. 10. The case k = 0. Before engaging in a detailed study, let's note that the first ideal in the sequence can be identified with a multiplier ideal; for the general theory of multiplier ideals see [Laz04, Ch.9]. Proposition 10.1. We have I0(D) = I(cid:0)X, (1 − ǫ)D(cid:1), the multiplier ideal associated to the Q-divisor (1 − ǫ)D on X, for any 0 < ǫ ≪ 1. Proof. Recall that C•−n = ωY (E), hence F0ωX(∗D) = Im(cid:0)f∗ωY (E) → ωX(D)(cid:1) = f∗OY (KY /X + E − f∗D) ⊗ ωX(D). Therefore the statement to be proved is that On the other hand, the right-hand side is by definition f∗OY (KY /X + E − f∗D) = I(cid:0)X, (1 − ǫ)D(cid:1). f∗OY(cid:0)KY /X − ⌊(1 − ǫ)f∗D⌋(cid:1) = f∗OY(cid:0)KY /X + (f∗D)red − f∗D(cid:1), which implies the desired equality. Remark 10.2. An equivalent result for F0OX(∗D) can be found in Saito [Sai09], stated and proved using the theory of the V -filtration. (cid:3) The following is a direct consequence of the definition; see [Laz04, 9.3.9]. 24 M. MUSTAT¸ A AND M. POPA Corollary 10.3. We have I0(D) = OX if and only if the pair (X, D) is log-canonical. 11. Independence of resolution, and filtration property. We will remark in the next section that the ideals Ik(D) are the same as those defined by the Hodge filtration on ωX(∗D); thus their independence of the choice of log resolution can be deduced from Saito's results on the uniqueness of open direct images [Sai90, Proposition 2.11]. We also include below an elementary proof that does not appeal to the theory of mixed Hodge modules. Theorem 11.1. The ideals Ik(D) are independent of the choice of log res- olution. Proof. Since every two log resolutions can be dominated by a third one, it is enough to consider two morphisms f : Y → X and g : Z → Y such that both f and h = f ◦ g are log resolutions of (X, D) which are isomorphisms over X r D. We put E = (f∗D)red and G = (g∗f∗D)red. We denote by C Y,•k−n the complex Y (log E)⊗OY f∗Fk−1DX → ωY (E)⊗OY f∗FkDX → 0, 0 → f∗Fk−nDX → . . . → Ωn−1 on Y , and by C Z,•k−n the complex 0 → h∗Fk−nDX → . . . → Ωn−1 on Z, both of them placed in degrees −n, . . . , 0. Since each f∗Fj DX is a locally free OY -module, we may apply the projection formula and Theo- rem 31.1i) to deduce that Z (log G)⊗OZ h∗Fk−1DX → ωZ(G)⊗OZ h∗FkDX → 0, Rpg∗(cid:0)Ωn−q Z (log G) ⊗OZ g∗f∗Fk−qDX(cid:1) = 0 for all q ≥ 0 and all p > 0, and that we have canonical isomorphisms Ωn−q Y for all q ≥ 0. We thus obtain a canonical isomorphism (log E) ⊗OY f∗Fk−qDX ≃ R0g∗(cid:0)Ωn−q C Y,•k−n ≃ Rg∗C Z,•k−n, Z (log G) ⊗OZ g∗f∗Fk−qDX(cid:1) hence a canonical isomorphism Rf∗C Y,•k−n ≃ Rf∗Rg∗C Z,•k−n ≃ Rh∗C Z,•k−n. The assertion in the theorem is now a consequence of the fact that the induced isomorphism R0f∗C Y,•k−n ≃ R0h∗C Z,•k−n commutes with the two morphisms to ωX(∗D). Since the whole picture is compatible with restriction to open subsets, this follows by restricting to U = X r D, where the assertion is straightforward. (cid:3) It is instructive to also give an elementary proof of the fact that F•ωX(∗D) gives a filtration for the DX -module ωX(∗D) (without appealing to push- forwards of filtered DX -modules). HODGE IDEALS 25 Lemma 11.2. For every k, ℓ ∈ N, we have Fℓ−nωX(∗D) · Fk DX ⊆ Fℓ+k−nωX(∗D). Proof. We employ the usual log resolution notation. With the notation in §9, we see that using the multiplication maps FiDX ⊗OX FkDX → Fi+kDX we get a morphism of complexes C•ℓ−n ⊗f −1OX f−1FkDX −→ C•k+ℓ−n. Since FkDX is a locally free OX-module, applying Rf∗ and taking H 0, we obtain an induced morphism compatible with the canonical multiplication map R0f∗C•ℓ−n ⊗OX FkDX → R0C•k+ℓ−n ωX(∗D) ⊗OX FkDX → ωX(∗D). By taking the images of R0f∗C•ℓ−n and R0f∗C•k+ℓ−n in ωX(∗D), we conclude that right-multiplication with sections of FkDX induces the inclusion in the statement. (cid:3) In light of Lemma 11.2, it is natural to ask for which ℓ the inclusion in the lemma is an equality for all k ≥ 0, in which case the filtration is determined by F−nω(∗D), . . . , Fℓ−n(∗D). We will return to this in §17. 12. Comparison with Hodge filtration, and strictness property. It follows from the discussion in §6 and the results from §3 used in the defi- nition, that the filtration F•ωX(∗D) we introduced in §9 is the same as the Hodge filtration on ωX(∗D) ≃ H 0f+ωY (∗E). Our definition of Hodge ideals is independent of this, but equating it with the Hodge filtration highlights the following important extra consequence that comes from strictness. For k ≥ 0, recall that C•k−n are the complexes providing the filtration on A•. Corollary 12.1. With the notation in §9, the following hold for every k ≥ 0: i) The map is injective. R0f∗C•k−n −→ R0f∗A• = ωX(∗D) ii) (Local vanishing for Ik(D).) We have Rif∗C•k−n = 0 for i 6= 0. Proof. We have Rif∗A• = 0 for all i 6= 0 by Lemma 2.2 and Proposition 2.4. On the other hand, strictness implies that Rif∗C•k−n injects in Rif∗A•. (cid:3) Remark 12.2. The case k = 0 in part ii) of the corollary is local vanishing for multiplier ideals; in this case C•−n = ωY (E). We note that in fact all vanishings Rif∗C•k−n = 0 are more elementary when i > 0. (This completely takes care of the case k = 1 for instance, since C•1−n is quasi-isomorphic to 26 M. MUSTAT¸ A AND M. POPA H 0C•1−n, whose negative direct images trivially vanish.) Indeed, if C• = C•k−n, we have Rif∗C j ≃ Rif∗Ωj+n (log E) ⊗OX Fk+j DX = 0 for by Theorem 32.1. The first quadrant spectral sequence Y i + j > 0 Ep,q 1 = Rqf∗C p−n =⇒ Hp+q−n = Rp+q−nf∗C•, then implies that Rif∗C• = 0 for i > 0. 13. Chain of inclusions. It follows from the definition and Lemma 11.2 that Ik−1(D) · OX (−D) ⊆ Ik(D) for each k ≥ 1. However, the Hodge ideals also satisfy a more subtle sequence of inclusions. Proposition 13.1. For every reduced effective divisor D on the smooth variety X, and for every k ≥ 1, we have Ik(D) ⊆ Ik−1(D). Proof. We give an argument using the theory of mixed Hodge modules. Consider the canonical inclusion ι : OX ֒→ OX(∗D) of filtered left DX -modules that underlie mixed Hodge modules. Since the category MHM(X) of mixed Hodge modules on X constructed in [Sai90] is abelian, the cokernel M of ι underlies a mixed Hodge module on X too, and it is clear that M has support D. Since morphisms between Hodge D-modules preserve the filtrations and are strict, for each k ≥ 0 we have a short exact sequence 0 −→ FkOX −→ FkOX (∗D) −→ FkM −→ 0. Recall now that FkOX = OX for all k ≥ 0. On the other hand, if h is a local equation of D, then by [Sai88, Lemma 3.2.6] we have h · FkM ⊆ Fk−1M. Indeed this a general property of Hodge D-modules whose support is con- tained in D. It follows easily that h · FkOX (∗D) ⊆ Fk−1OX (∗D) as well, which implies the assertion in the theorem by definition of Hodge ideals. (cid:3) E. Basic properties of Hodge ideals As we have seen, the ideal I0(D) is a multiplier ideal. We now start a study of the properties of the ideals Ik(D) for k ≥ 1. HODGE IDEALS 27 14. The ideals Jk(D). By analogy with the simple normal crossings case (8.1), we define for each k ≥ 0 an auxiliary ideal sheaf Jk(D) by the formula ωX(cid:0)(k + 1)D(cid:1) ⊗ Jk(D) =(cid:0)ωX(D) ⊗ I0(D)(cid:1) · FkDX . Lemma 14.1. For each k ≥ 0 there is an inclusion Jk(D) ⊆ Ik(D). Proof. We need to check that (cid:0)ωX(D) ⊗ I0(D)(cid:1) · FkDX ⊆ ωX(cid:0)(k + 1)D(cid:1) ⊗ Ik(D). But this is precisely the case ℓ = 0 in Lemma 11.2. (cid:3) Remark 14.2. An a priori different looking, but in fact equivalent state- ment involving the V -filtration, was noted in [Sai09, Theorem 0.4]. Remark 14.3. For every k ≥ 0, we have OX(cid:0) − (k + 1)D(cid:1) ⊆ Jk(D), hence Lemma 14.1 implies OX(cid:0) − (k + 1)D(cid:1) ⊆ Ik(D). Indeed, the assertion when k = 0 follows from OX(−D) = I(X, D) ⊆ I(cid:0)X, (1 − ǫ)D(cid:1) = I0(D), where 0 < ǫ ≪ 1. The general case now follows from the definition of Jk(D), using the fact that OX ⊆ FkDX . When studying the connection between Hodge ideals and the singularities of D, the following estimate for the ideals Jk(D), depending on the multi- plicity of D, will prove useful. Recall first that if W ⊆ X is an irreducible closed subset, then the p-th symbolic power of IW is I (p) W := {f ∈ OX multx(f ) ≥ p for x ∈ W general}, i.e. the ideal sheaf consisting of functions that have multiplicity at least p at a general (and hence every) point of W . If W is smooth, it is well known that I (p) W = I p W . We begin with an estimate for Jk+1(D) in terms of Jk(D). Recall that for an ideal I in OX, its Jacobian ideal Jac(I) is the ideal F1DX · I ⊆ OX . Lemma 14.4. For every k ≥ 0, we have Jk+1(D) ⊆ OX (−D) · Jac(Jk(D)) + Jk(D) · Jac(OX (−D)). Proof. It follows from the definition of the ideals Jk(D) that F1DX ·(cid:0)OX ((k + 1)D) · Jk(D)(cid:1) = OX(cid:0)(k + 2)D(cid:1) · Jk+1(D). In other words, if h is a local equation of D, then Jk+1(D) is locally generated by hk+2(cid:0)P · hk+1(cid:1), where g varies over the local sections of Jk(D) and P varies over the local sections of F1DX. The assertion in the lemma now follows from the Leibniz rule. (cid:3) g 28 M. MUSTAT¸ A AND M. POPA Remark 14.5. The inclusion in Lemma 14.4 is not always an equality. Suppose, for example, that X = C2 with coordinates x and y, and D is defined by (x2 + y3 = 0). In this case, we have I0(D) = J0(D) = (x, y), J1(D) = (x2, xy, y3), J2(D) = (x3, x2y2, xy3, y5, y4 − 3x2y), while OX(−D) · Jac(J1(D)) + J1(D) · Jac(OX (−D)) = (x3, x2y, xy3, y4). Proposition 14.6. Let X be a smooth variety, W ⊆ X an irreducible closed subset defined by the ideal IW , and D a reduced effective divisor with multW (D) = m ≥ 1. i) If I0(D) ⊆ I (m′) W for some m′ ≥ 0, then Jk(D) ⊆ I (k(m−1)+m′) W In particular, we always have Jk(D) ⊆ I (k(m−1)) W for every k ≥ 0. . ii) If r = codim(W, X) and m ≥ r, then I0(D) ⊆ I (m−r) W . Proof. By restricting to an appropriate open subset intersecting W , we can assume that W is smooth, and work with usual powers. For the first as- sertion, we argue by induction on k, the case k = 0 being trivial since J0(D) = I0(D). Note that if I ⊆ I r W . By induction we also have Jk(D) ⊆ I k(m−1)+m′ I m−1 W , then Jac(I) ⊆ I r−1 W W . Therefore we have Jac(OX (−D)) ⊆ , hence We deduce from Lemma 14.4 that Jac(Jk(D)) ⊆ I k(m−1)+m′−1 W . Jk+1(D) ⊆ I (k+1)(m−1)+m′ W . The assertion in ii) follows from the fact that I0(D) = I (X, (1 − ǫ)D) (see Proposition 10.1) and well-known estimates for multiplier ideals (see [Laz04, Example 9.3.5]). (cid:3) 15. Behavior under smooth pullback. In this section we consider the behavior of the ideals Ik(D) under pull-back by a smooth morphism. As before, we assume that D is a reduced effective divisor on the smooth n- dimensional variety X. Proposition 15.1. If p : X′ → X is a smooth morphism and D′ = p∗D, then for every k ≥ 0 we have Ik(D′) = Ik(D) · OX ′. HODGE IDEALS 29 Proof. Note first that since p is smooth, the effective divisor D′ is reduced. Let f : Y → X be a log resolution of (X, D) which is an isomorphism over the complement of D. We have a commutative diagram Y ′ = Y ×X X′ f ′ q Y f X′ p X and it is clear that f′ is a log resolution of (X′, D′). Moreover, if E = (f∗D)red and E′ = (f′∗D′)red, then E′ = q∗E. The assertion in the proposition is local on X′. Therefore we may assume that we have a system of algebraic coordinates x1, . . . , xn ∈ Γ(X, OX ) on X, and x′1, . . . , x′r ∈ Γ(X′, OX ′) such that p∗(x1), . . . , p∗(xn), x′1, . . . , x′r form a system of algebraic coordinates on X′. Note that x′1, . . . , x′r define a smooth morphism u : X′ → Ar such that (p, u) : X′ → X × Ar is ´etale. Since p factors as the composition X′ → X × Ar → X, it is enough to consider separately the case when p is ´etale and when X′ = X × Ar and p is the projection. Following the notation in §9, let AD,• and AD′,• denote the complexes on Y and Y ′ that appear in the definition of Ik(D) and Ik(D′), respectively. We put C D′,•k−n = Fk−nAD′,• and C D,•k−n = Fk−nAD,•. Suppose first that p is ´etale. In this case it is clear that C D′,•k−n = q−1C D,•k−n ⊗q−1p−1OX f′−1OX ′ and since OX ′ is flat over f−1OX, it follows that we have a canonical iso- morphism R0f′ ∗C D′,•k−n ≃ p∗R0f∗C D,•k−n. Note that this isomorphism is compatible with restriction to open subsets. In order to check that the isomorphism is compatible with the corresponding maps to ωX ′(cid:0)(k + 1)D′(cid:1) ≃ p∗ωX(cid:0)(k + 1)D(cid:1), it is enough to restrict to U = X r D, over which the assertion is clear. This gives the equality in the proposition. Suppose now that X′ = X × Ar and p : X′ → X is the projection. It is easy to see, using the definition, that we have an isomorphism of filtered complexes (15.2) where B• is the complex AD′,• ≃ q−1AD,• ⊗C u−1B•, 0 −→ DAr −→ Ω1 Ar ⊗ DAr −→ . . . −→ ωAr ⊗ DAr −→ 0, 30 M. MUSTAT¸ A AND M. POPA placed in degrees −r, . . . , 0. It follows from Proposition 3.1 that the canon- ical map B• → ωAr is a filtered quasi-isomorphism, hence we deduce from (15.2) that we have a canonical filtered quasi-isomorphism AD′,• −→ q−1AD,• ⊗C u−1ωAr . In particular, we have a quasi-isomorphism C D′,•k−n −→ q−1C D,•k−n ⊗ u−1ωAr . It follows using Kunneth's formula that we have a canonical isomorphism R0f′ ∗C D′,•k−n ≃ p−1R0f∗C D,•k−n ⊗C u−1ωAr . As before, by restricting to U = X r D we see that this isomorphism is compatible with the corresponding maps to ωX ′(cid:0)(k + 1)D′(cid:1) ≃ p∗ωX(cid:0)(k + 1)D(cid:1)⊗ u∗ωAr = p−1ωX(cid:0)(k + 1)D(cid:1)⊗C u−1ωAr . The equality in the proposition now follows from the definition of Hodge ideals. This completes the proof. (cid:3) 16. Restriction to hypersurfaces. We now turn to the behavior of Hodge ideals under restriction to a general hypersurface. Theorem 16.1. Let D be a reduced effective divisor on the smooth n- dimensional variety X. For every k ≥ 0, if H is a general element of a base-point free linear system on X, then Ik(DH ) = Ik(D) · OH. Proof. Note first that since H is general, it follows from Bertini's theorem that H is smooth and DH is a reduced effective divisor on H, hence Ik(DH ) is well defined. After possibly replacing X by the open subsets in a suitable affine cover, we may assume that we have a system of global coordinates x1, . . . , xn on X such that H is defined by the ideal (x1). Note that in this case we have an isomorphism ωH ≃ ωXH such that if α is a local (n−1)-form on X, the isomorphism maps the restriction of α to H to (dx1 ∧ α)H . Let f : Y → X be a log resolution of (X, D) which is an isomorphism over X r D, and let E = (f∗D)red. We will freely use the notation in §9. Since H is general, it follows that the scheme-theoretic inverse image f−1(H) is equal to the strict transform eH of H, hence eH is a general section of a base- point free linear system on Y . In particular, the divisor E + eH is a reduced, simple normal crossing divisor. Moreover, the restriction g : eH → H of f is a log resolution of (H, DH ) which is an isomorphism over H r DH, and the relevant divisor for computing Ik(DH ) is E eH . HODGE IDEALS 31 Consider the Cartesian diagram f−1(H) j g H i Y f X. If F is a sheaf of f−1OX -modules on Y , then j∗F := F ⊗f −1OX f−1OH is a sheaf of f−1OX -modules on Y , that we identify in the usual way with a sheaf of g−1OH-modules on f−1(H). Given any object C• in the derived category of f−1OX-modules on Y we have, as usual, a canonical base-change morphism (16.2) Li∗Rf∗C• → Rg∗Lj∗C•. This is an isomorphism if C• is a complex of quasi-coherent OY -modules since OH and OY are Tor-independent over X, see [TSPA16, Lemma 35.18.3] (this holds without the genericity assumption on H). This implies that (16.2) is an isomorphism for our complexes C• = C•k−n as well. Indeed, for every k we have an exact triangle C•k−n−1 → C•k−n → C•k−n → C•k−n−1[1], where C•k−n is a complex of coherent sheaves on Y (see the proof of Proposi- tion 3.1). The assertion now follows by induction on k, starting with k = −1, when it is trivial. Note also that the morphism (16.2) is an isomorphism for C• = A•, since A• is quasi-isomorphic to the quasi-coherent sheaf ωY (∗E). We thus obtain a commutative diagram (16.3) Li∗Rf∗C•k−n α Li∗Rf∗A• φ ψ Rg∗Lj∗C•k−n β Rg∗Lj∗A•, in which φ and ψ are isomorphisms. Recall now that Rf∗C•k−n = R0f∗C•k−n ֒→ R0f∗A• = Rf∗A• = ωX(∗D), the image being ωX(cid:0)(k + 1)D(cid:1) ⊗ Ik(D). Since when F is either ωX(∗D) or ωX(cid:0)(k + 1)D(cid:1)⊗ Ik(D), it follows that the map (OH , F ) = 0 for i ≥ 1 α in (16.3) gets identified to T orOX i Ik(D) ⊗ ωX(cid:0)(k + 1)D(cid:1) ⊗ OH → ωX(∗D) ⊗ OH ≃ ωH(cid:0) ∗ (DH)(cid:1). 32 M. MUSTAT¸ A AND M. POPA On the other hand, recall that for every p we have hence Therefore and similarly C p k−n = Ωp+n Y (log E) ⊗OY f∗Fk+pDX , T orf −1OX i (f−1OH, C p k−n) = 0 for i ≥ 1. Lj∗C•k−n = C•k−n ⊗f −1OX f−1OH Lj∗A• = A• ⊗f −1OX f−1OH . Suppose now that A′• and C′•k−n are the corresponding complexes on eH, that are involved in the definition of Ik(DH). In order to complete the proof of the theorem, it is enough to show that we have quasi-isomorphisms C′•k−n −→ C•k−n ⊗f −1OX f−1OH (16.4) for every k ≥ 0, which are compatible with the inclusions C′•k−n ⊆ C′•k−n+1 and C•k−n ⊆ C•k−n+1. Indeed, in this case we get a commutative diagram (16.5) Rg∗Lj∗C•k−n β Rg∗Lj∗A• Rg∗C′•k−n Rg∗A′•, in which the horizontal maps are isomorphisms. By combining the commu- tative diagrams (16.3) and (16.5), we obtain an isomorphism Ik(D) ⊗ OH → Ik(DH ). (16.6) In order to deduce that Ik(D) · OH = Ik(DH), it is enough to show that the isomorphism (16.6) is induced by the canonical surjection OX → OH. This can be checked over the complement of D, where both sides are equal to OHrD and the map is the identity. We now define the quasi-isomorphism in (16.4). For every q ≥ 0, let GqDX = Mα2+···+αn≤q OX ∂α2 x2 ··· ∂αn xn ⊆ FqDX . We identify in the obvious way Fq DH with Gq DX · OH. Since Gq DX com- mutes with x1, we see that for every p, we have an injective map Ωp+n−1 eH (log E eH ) ⊗O eH g∗Fk+pDH ֒→ Ωp+n Y (log E) eH ⊗OY f∗Gk+pDX ֒→ Ωp+n Y (log E) ⊗OY f∗Fk+pDX ⊗f −1OX f−1OH given by α eH ⊗ g∗(QH ) −→ (dx1 ∧ α) ⊗ Q ⊗ 1 for every local sections α of Ωp+n−1 (log E) and Q of Gk+pDX . This gives the injective morphism of complexes in (16.4). In order to show that this is Y HODGE IDEALS 33 a quasi-isomorphism, arguing by induction on k, we see that it is enough to show that the induced injective morphism of complexes (16.7) C′•k−n −→ C•k−n ⊗OY O eH is a quasi-isomorphism. This can be checked locally on Y , hence we may identify the map w : TY (− log E) eH → (f∗TX) eH to the map (u, Id) : T eH(− log E eH ) ⊕ O eH → g∗TH ⊕ O eH. It follows from the proof of Proposition 3.1 that the complexes C′•k−n and C•k−n ⊗OY O eH are isomorphic to Eagon-Northcott type complexes corre- sponding to the morphisms of vector bundles u and (u, Id), respectively, with the map (16.7) being induced by the inclusions T eH(− log E eH ) ֒→ T eH (− log E eH) ⊕ O eH and g∗TH ֒→ g∗TH ⊕ O eH. The assertion to be proved now follows from the fact that given a morphism of vector bundles u : V → W on a variety Z, if K•1 is an Eagon-Northcott type complex constructed for u and K•2 is the corresponding complex con- structed for (u, Id) : V ⊕OZ → W ⊕OZ, then the natural inclusion K•1 ֒→ K•2 is a quasi-isomorphism. We leave this as an exercise for the reader. (cid:3) Remark 16.8. Note that the generality assumption on H in Theorem 16.1 was only used to guarantee that H is smooth, D 6⊆ Supp(H) and DH is reduced, and given a log resolution f : Y → X of (X, D), this is also a log resolution of (X, D + H) such that f∗H is the strict transform of H. These conditions also hold if we work simultaneously with several general divisors, hence we obtain the following more general version of Theorem 16.1. Suppose that D is a reduced effective divisor on the smooth n-dimensional variety X. For every k ≥ 0, if H1, . . . , Hr are general elements of base-point free linear systems V1, . . . , Vr on X, and if Y = H1 ∩ ··· ∩ Hr, then Ik(DY ) = Ik(D) · OY . If the divisor H in Theorem 16.1 is not general, then we only have one inclusion. This is the analogue of the Restriction Theorem for multiplier ideals, see [Laz04, Theorem 9.5.1]. Theorem 16.9 ([MP16, Theorem A]). Let D be a reduced, effective divisor on the smooth n-dimensional variety X. If H is a smooth divisor on X such that H 6⊆ Supp(D) and DH is reduced, then for every k ≥ 0 we have Ik(DH ) ⊆ Ik(D) · OH. In particular, if (H, DH ) is k-log-canonical, then (X, D) is k-log-canonical in some neighborhood of H. Remark 16.10. It is not hard to deduce from this inductively that if Y is a smooth subvariety of X and D is a reduced effective divisor such that Y 6⊆ Supp(D) and DY is reduced, then Ik(DY ) ⊆ Ik(D) · OY . 34 M. MUSTAT¸ A AND M. POPA The proof of Theorem 16.9 uses the connection between the V -filtration and the Hodge filtration. Since it is of a different flavor, it is presented separately in [MP16], where we also deduce the following consequence re- garding the behavior of Hodge ideals in a family, similar to semicontinuity for multiplier ideals (see [Laz04, Chapter 9.5.D]). To state it, we fix some notation. Let h : X → T be a smooth morphism of relative dimension n between arbitrary varieties X and T , and s : T → X a morphism such that h ◦ s = IdT . Suppose that D is a relative effective Cartier divisor on X over T , such that for every t ∈ T the restriction Dt of D to the fiber Xt = h−1(t) is reduced. For every t ∈ T , we denote by ms(t) the ideal defining s(t) in Xt. Theorem 16.11 ([MP16, Theorem E]). With the above notation, for every q ≥ 1, the set is open in T . This applies in particular to the set Vq :=(cid:8)t ∈ T Ik(Dt) 6⊆ m q s(t)(cid:9), V1 =(cid:8)t ∈ T (Xt, Dt) is k−log canonical at s(t)}. 17. Generation level of the Hodge filtration. Let D be a reduced effective divisor on the smooth, n-dimensional variety X. We are interested in estimating for which k ≥ 0 the filtration on ωX(∗D) is generated at level k, that is, we have Fk−nωX(∗D) · FℓDX = Fk+ℓ−nωX(∗D) This is of course equivalent to having for all ℓ ≥ 0. FjωX(∗D) · F1DX = Fj+1ωX(∗D) for all j ≥ k − n. Y (log E) are independent of the choice of log resolution. The main technical result of this section is the following. As usual, we consider a log resolution f : Y → X of (X, D) which is an isomorphism over X r D, and put E = (f∗D)red. We note that by Corollary 31.2, the sheaves Rqf∗Ωp Theorem 17.1. With the above notation, the filtration on ωX(∗D) is gen- erated at level k if and only if Rqf∗Ωn−q (log E) = 0 for all q > k. Y Proof. It is enough to show that given k ≥ 0, we have (17.2) if and only if Rk+1f∗Ωn−k−1 holds of course by Lemma 11.2, hence the issue is the reverse inclusion. Fk−nωX(∗D) · F1DX = Fk−n+1ωX(∗D) (log E) = 0. The inclusion "⊆" in (17.2) always Y We freely use of the notation in §9. Following the proof of Lemma 11.2, we consider the morphism of complexes Φk : C•k−n ⊗f −1OX f−1F1DX −→ C•k+1−n HODGE IDEALS 35 induced by right multiplication, and let T • = Ker(Φk). Note that (17.2) holds if and only if the morphism (17.3) R0f∗C•k−n ⊗OX F1DX → R0f∗C•k+1−n induced by Φk is surjective. For every m ≥ 0, let Rm be the kernel of the morphism induced by right multiplication FmDX ⊗OX F1DX → Fm+1DX . Note that this is a surjective morphism of locally free OX-modules, hence Rm is a locally free OX -module and for every p we have (log E) ⊗f −1OX f−1Rk+p. T p = Ωn+p Y Consider the first-quadrant hypercohomology spectral sequence Ep,q 1 = Rqf∗T p−n =⇒ Rp+q−nf∗T •. The projection formula gives Rqf∗T p−n ≃ Rqf∗Ωp Y (log E) ⊗OX Rk+p−n, and this vanishes for p + q > n by Theorem 32.1. We thus deduce from the spectral sequence that Rjf∗T • = 0 for all j > 0. We first consider the case when k ≥ n and show that (17.2) always holds. Indeed, in this case Φk is surjective. It follows from the projection formula and the long exact sequence in cohomology that we have an exact sequence R0f∗C•k−n ⊗OX F1DX → R0f∗C•k+1−n → R1f∗T •. We have seen that R1f∗T • = 0, hence (17.3) is surjective and (17.2) holds in this case. Suppose now that 0 ≤ k < n. Let B• ֒→ C•k+1−n be the subcomplex given k+1−n for all p 6= −k − 1 and B−k−1 = 0. Note that we have a by Bp = C p short exact sequence of complexes (17.4) 0 −→ B• −→ C•k+1−n −→ C−k−1 k+1−n[k + 1] −→ 0. It is clear that Φk factors as C•k−n ⊗f −1OX f−1F1DX k−→ B• ֒→ C•k+1−n. Φ′ Moreover, Φ′k is surjective and Ker(Φ′k) = T •. As before, since R1f∗T • = 0, we conclude that morphism induced by Φ′k: is surjective. This implies that (17.3) is surjective if and only if the morphism R0f∗C•k−n ⊗OX F1DX → R0f∗B• (17.5) R0f∗B• → R0f∗C•k+1−n is surjective. The exact sequence (17.4) induces an exact sequence k+1−n → R1f∗B•. R0f∗B• → R0f∗C•k+1−n → Rk+1f∗C−k−1 36 M. MUSTAT¸ A AND M. POPA We have seen that R2f∗T • = 0 and we also have R1f∗(cid:0)C•k−n ⊗f −1OX f−1F1DX(cid:1) = 0. This follows either as above, using the projection formula, the hypercohomol- ogy spectral sequence, and Theorem 32.1, or can be deduced from strictness, see Remark 12.2. We deduce from the long exact sequence associated to 0 −→ T • −→ C•k−n ⊗f −1OX f−1F1DX −→ B• −→ 0 that R1f∗B• = 0. Putting all of this together, we conclude that (17.3) is surjective if and only if Rk+1f∗C−k−1 k+1−n = 0. Since we have by definition this completes the proof of the theorem. Rk+1f∗C−k−1 k+1−n = Rk+1f∗Ωn−k−1 Y (log E), (cid:3) We now show how Theorem 17.1 implies the calculation of the generation level of the Hodge filtration on OX (∗D) stated in the Introduction. This question was first raised by Saito in [Sai09], where he also computed the pre- cise generation level in the case of isolated quasi-homogeneous singularities; see Remark 20.11. Proof of Theorem B. We begin by proving the first assertion in the theorem. Let f : Y → X be a log resolution of (X, D) which is an isomorphism over of D is smooth (possibly disconnected). X r D, and let E = (f∗D)red. We may assume that the strict transform eD It follows from Theorem 17.1 that we need to show that if n ≥ 2, then (17.6) (17.7) and Rnf∗OY = 0 Rn−1f∗Ω1 Y (log E) = 0. The vanishing (17.6) is an immediate consequence of the fact that the fibers of f have dimension at most n − 1, hence we focus on (17.7). We write E = eD + F , where eD is the strict transform of D and F is the reduced exceptional divisor. Recall that since eD is smooth, we have a short exact sequence 0 −→ Ω1 Y (log F ) −→ Ω1 It follows from Theorem 31.1ii) that Rn−1f∗Ω1 Y (log F ) = 0, Y (log E) −→ O eD −→ 0. and so in order to guarantee (17.7) it is enough to have Rn−1f∗O eD = 0. However, this is a consequence of the fact that all fibers of eD → D have dimension at most n − 2. This completes the proof of the first assertion. We prove the second assertion in the theorem by induction on n. Let k ≥ 0 be fixed. If k ≥ n − 2, then by what we have already proved we may HODGE IDEALS 37 take Uk = X. Suppose now that k ≤ n− 3. We define for ℓ ≥ k ideal sheaves I′ℓ(D) by the formula ωX(cid:0)(ℓ + 1)D(cid:1) ⊗ I′ℓ(D) =(cid:0)ωX(cid:0)(k + 1)D(cid:1) ⊗ Ik(D)(cid:1) ⊗ Fℓ−kDX. It is clear that we have I′ℓ(D) ⊆ Iℓ(D) for every ℓ ≥ k, with equality for ℓ = k. Saying that on an open subset U ⊆ X the filtration is generated at level k is equivalent to saying that Iℓ(D)U = I′ℓ(D)U for all ℓ ≥ k. Moreover, by what we have already proved, it is enough to check this for k + 1 ≤ ℓ ≤ n − 2. It is therefore enough to show that for every such ℓ we have codim(Zℓ, X) ≥ k + 3, where Zℓ = Supp(cid:0)Iℓ(D)/I′ℓ(D)(cid:1). Indeed, if this is the case we may take Uk = X r n−2[ℓ=k+1 Zℓ. In order to prove the bound on the codimension of Zℓ, we may assume that X is a subvariety of some AN . Let H be a general hyperplane in AN , and XH = X ∩ H and DH = DXH . Since H is general, we have Zℓ ∩ H = Supp(cid:0)Iℓ(D) · OXH /I′ℓ(D) · OXH(cid:1). On the other hand, it follows from Theorem 16.1 that since H is general, we have Ij(DH ) = Ij(D) · OXH for all j. The equality for j = k, together with the definition of the ideals I′ℓ(D), also gives It follows by induction that both I′ℓ(DH) ⊆ I′ℓ(D) · OXH . Supp(cid:0)Iℓ(D) · OXH /I′ℓ(D) · OXH(cid:1) ⊆ Supp(cid:0)Iℓ(DH )/I′ℓ(DH )(cid:1) have codimension ≥ k+3 in XH , hence codim(Zℓ, X) ≥ k+3. This completes the proof of the theorem. (cid:3) A basic question is to determine when the Hodge filtration on ωX(∗D) is generated by its 0th step, or equivalently, when the equality of Ik(D) and Jk(D) holds everywhere on X. This is of course the case when D is a simple normal crossings divisor. Theorem B implies that it is also always the case outside a closed subset of codimension at least 3. In particular: Corollary 17.8. If X is a smooth surface and D is a reduced effective divisor on X, then Ik(D) = Jk(D) for all k ≥ 0. 38 M. MUSTAT¸ A AND M. POPA Moreover, with the notation in Theorem 17.1, we see that Ik(D) = Jk(D) for all k if and only if Rqf∗Ωn−q (log E) = 0 for all q > 0. Based on this, it is not hard to find examples where equality does not hold, and the statement in Theorem B is sharp. Y Example 17.9. Let D be the cone in X = A3 over a smooth plane curve of degree d, and let f : Y → X be the log resolution obtained by blowing up the origin. The claim is that if d ≥ 3 = dim X, then R1f∗Ω2 Y (log E) 6= 0. eD (17.10) If we had 0 −→ Ω2 (log F eD) −→ 0. Indeed, we have E = eD + F , where F is the exceptional divisor of f , and so on Y there is a short exact sequence it would follow from the long exact sequence associated to the above sequence that it is enough to show that Y (log F ) −→ Ω2 H 2(cid:0)Y, Ω2 H 1(cid:0)eD, Ω1 Y (log E) −→ Ω1 Y (log F )(cid:1) = 0, (log F eD)(cid:1) 6= 0. Consider however the short exact sequence on eD: Since the fibers of f eD are at most one-dimensional, we have R2f∗Ω1 and so it suffices to check that (log F eD) −→ OF eD −→ 0. 0 −→ Ω1 eD −→ Ω1 = 0, eD eD eD R1f∗OF eD ≃ H 1(F eD, OF eD ) 6= 0. But F eD is isomorphic to the original plane curve of degree d ≥ 3, so this is clear. Therefore it is enough to prove (17.10). The short exact sequence 0 −→ OF (−F ) −→ Ω1 Y F −→ Ω1 F −→ 0 induces for every m ≥ 0 a short exact sequence 0 → Ω1 Y F ⊗ OF (−mF ) → ωF ⊗ OF (−mF ) → 0. Since F ≃ P2 and OF (−F ) ≃ OP2(1), it follows from the Euler exact sequence that F ⊗ OF (−(m + 1)F ) → Ω2 We conclude that H 2(F, Ω2 F ⊗ OF (−(m + 1)F )(cid:1) = 0 H 2(cid:0)F, Ω1 Y F ) ≃ C and H 2(cid:0)F, Ω2 Y (−(m + 1)F ) → Ω2 Y (−mF ) → Ω2 0 → Ω2 We now deduce from the exact sequence Y F ⊗ OF (−mF )(cid:1) = 0 for m ≥ 1. Y F ⊗ OF (−mF ) → 0 for all m ≥ 0. HODGE IDEALS 39 that for every m ≥ 1 the map is surjective. Since OY (−F ) is ample over X, we have H 2(cid:0)Y, Ω2 H 2(cid:0)Y, Ω2 H 2(cid:0)Y, Ω2 0 → Ω2 Y (−(m + 1)F )(cid:1) → H 2(cid:0)Y, Ω2 Y (−mF )(cid:1) Y (−mF )(cid:1) = 0 for all m ≫ 0, Y (−mF )(cid:1) = 0 for all m ≥ 1. Y (−F ) → Ω2 Y → Ω2 Y F → 0 and therefore gives an isomorphism H 2(Y, Ω2 Y ) ≃ H 2(Y, Ω2 Y F ) ≃ C. Finally, consider the exact sequence 0 → Ω2 Y → Ω2 Y (log F ) → Ω1 F → 0, In particular, the long exact sequence in cohomology corresponding to which induces H 1(F, Ω1 The composition F ) α→ H 2(Y, Ω2 Y ) → H 2(Y, Ω2 Y (log F )) → H 2(F, Ω1 F ) = 0. C ≃ H 1(F, Ω1 F ) → H 2(Y, Ω2 Y ) → H 2(F, Ω2 F ) ≃ C is given by cup-product with c1(OF (F )), hence it is nonzero. Therefore α is surjective, which implies (17.10). Remark 17.11. If we know that the filtration on ωX(∗D) is generated at level ℓ, then the argument in the proof of Lemma 14.4 shows that we have Ik+1(D) ⊆ OX(−D) · Jac(cid:0)Ik(D)(cid:1) + Ik(D) · Jac(cid:0)OX (−D)(cid:1) for all k ≥ ℓ. Arguing as in the proof of Proposition 14.6, we see that if W ⊆ X is an irreducible closed subset defined by the ideal IW , with multW (D) = m ≥ 1, and such that Iℓ(D) ⊆ I (m′) W for some m′ ≥ 0, then Iℓ+k(D) ⊆ I (k(m−1)+m′) W for every k ≥ 0. In particular, we always have Iℓ+k(D) ⊆ I (k(m−1)) W . Returning to the case of surfaces, Corollary 17.8 has the following appli- cation to the local study of Hodge ideals. Corollary 17.12. If X is a smooth surface and D is a reduced effective divisor on X such that multx(D) = m ≥ 2 for some x ∈ X, then for every k ≥ 0, where mx is the ideal defining x. Moreover, if m = 2, then Ik(D) ⊆ m (k+1)(m−1)−1 x Ik(D) ⊆ m k+1 x for every k ≥ 0, 40 M. MUSTAT¸ A AND M. POPA unless the singularity of D at x is a node. In particular, if D is a singular divisor, then Ik(D) 6= OX for every k ≥ 1. Proof. Both assertions follow by combining Corollary 17.8 and Proposi- tion 14.6. For the second assertion, we also use the fact that if (X, D) is log canonical at x, then the singularity of D at x is a node. This is easy (and well known): first, we must have multx(D) = 2, in which case (D, 0) is analytically equivalent to (cid:0)V (x2 + yℓ), 0(cid:1), for some ℓ ≥ 2. Since ℓ , we have (X, D) log canonical at x if and only if (cid:3) lct0(x2 + yℓ) = 1 ℓ = 2. 2 + 1 Remark 17.13. Corollary 17.12 says that on surfaces, unlike I0(D), for k ≥ 1 the ideal Ik(D) always detects singularities (for an extension to higher dimensions, see Corollary 21.3). For example, if D = V (xy) ⊂ C2, then it is immediate I0(D) = OX just as in the smooth case, while Proposition 8.2 shows that Ik(D) = (x, y)k for all k ≥ 0. This phenomenon persists for singular divisors with equal I0(D). For instance, if D = V (x2 + y3) ⊂ C2 is a cusp, it is well known (see [Laz04, Example 9.2.15]) that I0(D) = (x, y). Using Corollary 17.8 and Remark 14.5, we see that I1(D) = (x2, xy, y3). Consider now the divisor D = V(cid:0)xy(x + y)(cid:1) ⊂ C2 with a triple point at the origin. Blowing up this point gives a log resolution f : Y → C2, and if we denote by F the exceptional divisor, the formula in the proof of Proposition 10.1 gives I0(D) = f∗OY (−F ) = (x, y) as well. On the other hand, again using Corollary 17.8 and a simple calcu- lation, we obtain I1(D) = (x, y)3. Further concrete calculations can be done for higher k, but in general they become quite intricate. In any case, it is already apparent in dimension two that the sequence of ideals Ik(D) is a more refined invariant of singularities than the multiplier ideal I0(D) alone. 18. Behavior with respect to birational morphisms. Recall that multiplier ideals satisfy a birational transformation rule; see [Laz04, The- orem 9.2.33]. Given a proper morphism g : Z → X of smooth varieties, this describes the multiplier ideals of a divisor D on X in terms of the cor- responding multiplier ideals of g∗D and the exceptional divisor KZ/X. It would be very interesting to have a similar result for the Hodge ideals. The following theorem is a step in this direction, and will be crucial for later applications. Suppose that D is a reduced effective divisor on the smooth variety X and g : Z → X is a proper morphism which is an isomorphism over X r D, with Z smooth. Let DZ = (g∗D)red and TZ/X := Coker(TZ ֒→ g∗TX ). HODGE IDEALS 41 Theorem 18.1. With the above notation, for every k ≥ 0 the following hold: i) We have an inclusion of ideals iii) When k = 1, we have an exact sequence ii) If J is an ideal in OX such that J · TZ/X = 0, then g∗(cid:0)Ik(DZ ) ⊗ OZ(KZ/X + (k + 1)DZ − (k + 1)g∗D)(cid:1) ֒→ Ik(D). J k · Ik(D) ⊆ g∗(cid:0)Ik(DZ ) ⊗ OZ (KZ/X + (k + 1)DZ − (k + 1)g∗D)(cid:1). 0 → g∗(cid:0)I1(DZ ) ⊗ OZ (KZ/X + 2DZ − 2g∗D)(cid:1) ⊗ ωX(2D) → I1(D) ⊗ ωX(2D) → g∗(cid:0)TZ/X ⊗ I0(DZ ) ⊗ ωZ(DZ )(cid:1) → R1g∗(cid:0)I1(DZ ) ⊗ ωZ(2DZ )(cid:1) → 0. Proof. Let h : Y → Z be a log resolution of (Z, DZ ) which is an isomorphism over Z r DZ . Then f = g ◦ h is a log resolution of (X, D) as well, which is an isomorphism over X r D. As usual, we put E = (f∗D)red = (h∗DZ )red. Consider on Y the complexes C D,•k−n: 0 → f∗Fk−nDX → Ω1 and C DZ ,• k−n : 0 → h∗Fk−nDZ → Ω1 Y (log E)⊗h∗Fk−n+1DZ → ··· → ωY (E)⊗f∗FkDZ → 0, both placed in degrees −n, . . . , 0. Note that we have an inclusion of com- plexes Y (log E)⊗f∗Fk−n+1DX → ··· → ωY (E)⊗f∗FkDX → 0 C DZ ,• k−n ֒→ C D,•k−n. (The fact that the map is injective follows from the fact that all Ωp Y (log E) are locally free OY -modules, and the maps h∗Fj DZ → h∗g∗Fj DX are gener- ically injective morphisms of locally free left OY -modules.) Let M• be the quotient complex; this is a complex of right f−1OX -modules. Applying Rf∗ and taking the corresponding long exact sequence, we obtain an exact sequence (18.2) By definition, we have ι−→ R0f∗C D,•k−n → R0f∗M•. R0f∗C DZ ,• k−n R0f∗C D,•k−n = Ik(D) ⊗ ωX(cid:0)(k + 1)D(cid:1). On the other hand, recall that Rih∗C DZ ,• lary 12.1). It follows from the Leray spectral sequence that k−n = 0 for all i 6= 0 (see Corol- R0f∗C DZ ,• k−n = R0g∗R0h∗C DZ ,• k−n = g∗(cid:0)Ik(DZ ) ⊗ ωZ((k + 1)DZ )(cid:1) = g∗(cid:0)Ik(DZ ) ⊗ O(KZ/X + (k + 1)DZ − (k + 1)g∗D)(cid:1) ⊗ ωX(cid:0)(k + 1)D(cid:1). Finally, the map ι is compatible with restriction to open subsets of X. By restricting to an open subset U in the complement of D such that f is an isomorphism over U , it is clear that ιU is the identity on ωU . This implies 42 M. MUSTAT¸ A AND M. POPA inclusion in i). that ι is the restriction of the identity on ωX(cid:0)(k + 1)D(cid:1) and we obtain the to show that M• · J k = 0. Since we have Using (18.2), we see that in order to prove the assertion in ii), it is enough M p = Ωn+p Y (log E) ⊗ h∗(cid:0)g∗Fk+pDX/Fk+pDZ(cid:1), it follows that it is enough to show that g∗Fj DX ·J j ⊆ Fj DZ for every j ≥ 0. This is a consequence of the general Lemma 18.6 below, and completes the proof of ii). In order to prove the assertion in iii), we look a bit closer at the argument showing i). It follows from Lemma 3.4 that we have a commutative diagram with exact rows: 0 0 Ωn−1 Y (log E) Id Ωn−1 Y (log E) ωY (E) ⊗ h∗F1DZ α ωY (E) ⊗ f∗F1DX F DZ 1−n F D 1−n 0 0. As we have seen, the map α is injective. Moreover, we have It follows from the diagram that we have an induced exact sequence Coker(α) ≃ ωY (E) ⊗ h∗TZ/X . 0 −→ F DZ 1−n −→ F D 1−n −→ ωY (E) ⊗ h∗TZ/X −→ 0. Applying h∗ we obtain an exact sequence (18.3) 0 → I1(DZ ) ⊗ ωZ (2DZ ) → h∗F D We claim that the canonical morphism 1−n → h∗(cid:0)ωY (E) ⊗ h∗TZ/X(cid:1) → 0. h∗ωY (E) ⊗ TZ/X −→ h∗(cid:0)ωY (E) ⊗ h∗TZ/X(cid:1) 0 −→ TZ −→ g∗TX −→ TZ/X −→ 0, (18.4) is an isomorphism. Indeed, applying h∗ to the exact sequence (18.5) tensoring with ωY (E), and then applying h∗, we obtain using the projection formula the exact sequence β 0 → h∗ωY (E) ⊗ TZ → h∗ωY (E) ⊗ g∗TX → h∗(cid:0)ωY (E) ⊗ h∗TZ/X(cid:1) → 0. (Note that R1h∗(cid:0)ωY (E)⊗ h∗TZ(cid:1) = R1h∗ωY (E)⊗ TZ = 0 by Theorem 32.1.) On the other hand, by tensoring (18.5) with h∗ωY (E), we conclude that Coker(β) = h∗ωY (E) ⊗ TZ/X, hence (18.4) is an isomorphism. Since h∗ωY (E) = I0(DZ ) ⊗ ωZ(DZ ), applying g∗ to the exact sequence (18.4) gives the exact sequence in the proposition; note that the surjectivity of the last map follows from the in- clusion R1g∗(h∗F D 1−n) ֒→ R1f∗F D 1−n HODGE IDEALS 43 provided by the Leray spectral sequence, and the fact that R1f∗F D by Corollary 12.1. Lemma 18.6. Let g : Z → X be a birational morphism of smooth varieties. If J is an ideal on X such that J · TZ/X = 0, then 1−n = 0 (cid:3) Coker(FkDZ ֒→ g∗FkDX ) · J k = 0 for every k ≥ 0. Proof. Note that the inclusion FkDZ ֒→ g∗FkDX has the structure of a map of OZ − g−1OX bimodules; on the cokernel we use the right g−1OX -module structure. Recall that g∗FkDX (resp. FkDZ ) is locally generated as a left OZ-module by ≤ k products of sections in f∗TX (resp. TZ ). We prove by induction on k ≥ 0 that if D1, . . . , Dk are local sections of TX and τ1, . . . , τk are local sections of J, then f∗D1 . . . f∗Dkτ1 . . . τk is a section of FkDZ . The assertion is trivial for k = 0. If k ≥ 1, then we have by induction f∗D1 . . . f∗Dkτ1 . . . τk − f∗D1 . . . f∗Dk−1(τ1f∗Dk)τ2 . . . τk = f∗D1 . . . f∗Dk−1τ2 . . . τk−1Dk(τ1) ∈ Fk−1DZ . After iterating this k times, we obtain f∗D1 . . . f∗Dkτ1 . . . τk − (f∗D1 . . . f∗Dk−1τ1 . . . τk−1)(τkf∗Dk) ∈ Fk−1DZ . We know by assumption that τkf∗Dk ∈ F1DZ , and by induction we have f∗D1 . . . f∗Dk−1τ1 . . . τk−1 ∈ Fk−1DZ . We thus conclude that f∗D1 . . . f∗Dkτ1 . . . τk is a section of FkDZ . Example 18.7. If g : Z → X is the blow-up of a smooth variety X along the smooth subvariety W defined by the ideal IW , with exceptional divisor G, then IW · TZ/X = 0. In fact, we have an isomorphism (cid:3) TZ/X ≃ TG/W ⊗OG OG(−1) which clearly implies this. Indeed, an easy computation in local charts gives an isomorphism We deduce that Coker(g∗Ω1 X → Ω1 Z) ≃ Ω1 G/W . TZ/X ≃ Ext1 OZ (Ω1 G/W , OZ ) ≃ (Ω1 G/W )∨ ⊗OZ Ext1 OZ (OG, OZ ) ≃ TG/W ⊗OZ OG(G) ≃ TG/W ⊗OG OG(−1). In applications we will also make use of the following more "local" versions of the assertion in Theorem 18.1 ii). Remark 18.8. Let g : Z → X and D, DZ be as in Theorem 18.1. Consider an open subset V of Z and let g′ : V → X be the restriction of g. If J is an ideal sheaf on X such that J · TZ/X = 0 on V , then (18.9) ∗(cid:0)Ik(DZ ) ⊗ OZ(cid:0)KZ/X + (k + 1)DZ − (k + 1)g∗D(cid:1)V(cid:1) , J k · Ik(D) ⊆ g′ 44 M. MUSTAT¸ A AND M. POPA where the right-hand side is considered as an OX -submodule of the constant sheaf of rational functions on X. In particular, for every prime divisor G on Z that intersects V , we have k · ordG(J) + ordG(Ik(D)) + ordG(cid:0)KZ/X + (k + 1)DZ − (k + 1)f∗(D)(cid:1) ≥ 0. Indeed, using the notation in the proof of Theorem 18.1, note that if i : V ֒→ Z is the inclusion, then we have a commutative diagram of distinguished triangles on X: Rg∗Rh∗C DZ ,• k−n Rg∗Rh∗C D,•k−n Rg∗Rh∗M• γ Rg∗Ri∗i∗Rh∗C DZ ,• Rg∗Ri∗i∗Rh∗M•, k−n such that the exact sequence (18.2) is obtained by applying the cohomology functor H 0(−) to the top triangle. Note first that Rg∗ ◦ Ri∗ = Rg′∗ and Rg∗Ri∗i∗Rh∗C D,•k−n Moreover, we have Rh∗C DZ ,• k−n = Ik(DZ ) ⊗ ωZ(cid:0)(k + 1)DZ(cid:1). i∗Rh∗M• ≃ Rh′ ∗i′∗M•, where i′ : h−1(V ) ֒→ Y is the inclusion and h′ : h−1(V ) → V is the restriction of h. The argument in the proof of Theorem 18.1 and our hypothesis on J implies that i′∗M• · J k = 0, hence (cid:0)Rg∗Ri∗i∗Rh∗M•(cid:1) · J k = 0. We thus conclude that CokerH 0(γ) · J k = 0. Applying H 0(−) to the commutative diagram of distinguished triangles above, we obtain a commutative diagram g∗(cid:0)Ik(DZ ) ⊗ ωZ((k + 1)DZ )(cid:1) g′∗(cid:0)Ik(DZ ) ⊗ ωZ((k + 1)DZ )V(cid:1) ι H 0(γ) Ik(D) ⊗ ωX(cid:0)(k + 1)D(cid:1) H 0(cid:0)Rg′∗i∗Rh∗C D,•k−n(cid:1). Finally, note that the whole picture is compatible with restriction to open subsets of X. Suppose that U ⊆ X is an open subset of the complement of D such that f is an isomorphism over U and g−1(U ) ⊆ V . In this case we have a morphism from the above commutative diagram of sheaves on X to the push-forward of its restriction to U , which is a diagram all of whose entries are canonically identified to ϕ∗ωU , where ϕ : U → X is the inclusion. Since CokerH 0(γ) · J k = 0, we deduce that inside ϕ∗ωU we have Im(cid:0)Ik(D) ⊗ ωX((k + 1)D) → ωU(cid:1) · J k ⊆ g′ ∗(cid:0)Ik(DZ ) ⊗ ωZ((k + 1)DZ )V(cid:1), which is equivalent to the inclusion (18.9). HODGE IDEALS 45 Remark 18.10. We will also make use of the following variant of the pre- vious result. Suppose that we are in the setting of Remark 18.8 and that the following extra conditions hold: i) There is a prime g-exceptional divisor G on Z that on V meets the ii) J · OV = OV (−T ) for some g-exceptional divisor T . iii) Ik(D) = OX at the generic point of g(G). strict transform eD nontrivially and with simple normal crossings. k · ordG(J) + ordG(cid:0)KZ/X + (k + 1)DZ − (k + 1)g∗(D)(cid:1) ≥ k. Indeed, after possibly replacing V by a smaller open subset, we may assume In this case we have deduce from (18.9) that OV (−kaG) ⊆ Ik(DZ ) · OY(cid:0)KZ/X + (k + 1)DZ − (k + 1)f∗D(cid:1)V . that DZV = (eD + G)V and J · OV = OV (−aG), where a = ordG(J). We Let us choose coordinates x1, . . . , xn at some point y ∈ G ∩ eD ∩ V such that eG and eD are defined by (x1) and (x2), respectively. Note that by Proposition 8.2 we have Ik(DZ ) = (x1, x2)k around y, hence (x1)ka ⊆ (x1, x2)k · (x1)b around y, where b = (k + 1)(ordG(D) − 1) − ordG(KY /X ). This implies ka ≥ k + b, as claimed. F. Local study of Hodge ideals In this section we apply the results in §18 to obtain lower bounds for the order of vanishing of Hodge ideals along various divisors over the ambient variety. In particular, we address the question of whether the singularities of the original divisor D imply that the various Ik(D) are nontrivial or not; more generally, we are interested in lower bounds for the order of Ik(D) at a given point. This, sometimes combined with the vanishing theorems discussed in the next section, leads to the most significant applications. We will also see that estimating the order of vanishing of Ik(D) along general divisors leads to interesting structural results about Hodge ideals. 19. Order of vanishing along exceptional divisors. We aim for lower bounds on the order of vanishing of the Hodge ideals along given divisors. In order to state our main result in this direction, we first introduce some notation. Suppose that G is an exceptional divisor over X. By a result due to Zariski (see [KM98, Lemma 2.45]), we can obtain G by a sequence of blow-ups such that at each step we blow up the center of G on the respective variety. More precisely, if we define the sequence of birational transforma- tions fi : Xi → Xi−1, for i ≥ 1, as follows: i) fi is the blow-up of Xi−1 along the center Wi−1 of G on Xi−1; 46 M. MUSTAT¸ A AND M. POPA ii) X0 = X, then there is an s such that Ws = G is a prime divisor on Xs. Let s be the smallest integer with this property (note that s ≥ 1 since we assumed that G is exceptional over X). After successively replacing each Xi by a suitable open subset intersecting Wi, we may assume that all Xi and Wi are smooth; in this case, of course, the maps fi are not going to be proper anymore, but this will not cause any trouble. We denote the ideal defining Wi in Xi by IWi, and we put αj = ordG(IWj−1). Note that α1 ≥ α2 ≥ ··· ≥ αs−1 = αs = 1. We also denote by kG the coefficient of G in the relative canonical divisor KXs/X . With this notation, our main result in this direction is the following: Theorem 19.1. Given a reduced effective divisor D on the smooth variety X, for every exceptional divisor G and every k ≥ 0 we have ordG(cid:0)Ik(D)(cid:1) ≥ (k + 1)(ordG(D) − 1) − kG − α1qk, where q = ⌈(α1 + ··· + αs)/α1⌉. Proof. We may assume that ordG(D) ≥ 1, since otherwise the assertion in the theorem is trivial. We consider a sequence Xs → Xs−1 → . . . → X1 → X0 = X with s minimal, such that G is a prime divisor on Xs, and each Xi → Xi−1 is the blow-up of Xi−1 along the center Wi−1 of G on Xi−1. We denote by g : Xs → X the composition. We can find open subsets Ui ⊆ Xi that intersect Wi such that the following hold: i) we have induced morphisms Ui → Ui−1. ii) all Ui and Ui ∩ Wi are smooth. in Us of the strict transform eD of D and of all g-exceptional divisors but G. We denote by IWi the ideal defining Ui∩ Wi in Ui. Let V be the complement It is clear that g∗D has simple normal crossings on V , hence using Nagata's compactification theorem and by taking a suitable log resolution that is an isomorphism over U , we see that there is an open immersion V ֒→ Y over X, where f : Y → X is a log resolution of (X, D) which is an isomorphism over X r D. Let E = (f∗D)red. For every i with 1 ≤ i ≤ s, let fi : Ui → Ui−1 and gi : V → Ui be the corresponding maps. We claim that Coker(TV ֒→ g∗0TU0) is annihilated by , or equivalently by OV(cid:0) − qα1G(cid:1). Indeed, it follows from Example 18.7 I q W0 that IWi−1 · Coker(TUi → f∗i TUi−1) = 0, hence OU (−αiG) · Coker(g∗i TUi → g∗i−1TUi−1) = 0. This implies that the cokernel of the composition TV → g∗s−1TUs−1 → ··· → g∗0TU0 HODGE IDEALS 47 is annihilated by OV(cid:0)−(α1+···+αs)G(cid:1) ⊇ OV(cid:0)−qα1G(cid:1). Using Remark 18.8, we conclude that qk · ordG(IW0) + ordG(Ik(D)) + kG + (k + 1) − (k + 1) · ordG(D) ≥ 0, which implies the inequality in the theorem. (cid:3) Remark 19.2. With the notation in the proof of Theorem 19.1, suppose that we can choose the subsets U0, . . . , Us such that the following holds: there is y ∈ G ∩ eD ∩ Us that does not lie on any exceptional divisor over X different from G, and such that eD + G has simple normal crossings at y. In this case, if Ik(D) 6⊆ IW0, then (19.3) 0 ≥ k + (k + 1)(ordG(D) − 1) − kG − α1qk. For the argument in this case we consider a slightly different choice of V : we take an open set which contains y and such that the only exceptional divisor over X that it meets is G, and (G + eD)U has simple normal crossings. It follows that we can still find an open immersion V ֒→ Y over X, where Y → X is a log resolution of (X, D). We can now apply Remark 18.10 to conclude that (19.3) holds. Corollary 19.4. Let X be a smooth variety and D a reduced effective divisor on X. Suppose that W is an irreducible closed subset of X of codimension r ≥ 2, defined by the ideal IW . Given k, j ≥ 0, if m := multW (D) satisfies m ≥ 2 + j+r−2 k+1 , then Ik(D) ⊆ I (j) W . W = I (ℓ) Proof. After possibly replacing X by an open subset that intersects W , we may assume that W is smooth. We now apply Theorem 19.1 with G being the exceptional divisor of the blow-up of X along W . Note that ordG(D) = m, kG = r − 1, s = 1, and α1 = 1. For an ideal J we have J ⊆ I ℓ W if and only if ordG(J) ≥ ℓ, hence we obtain the assertion. (cid:3) Example 19.5. When k = 0, the criterion in Corollary 19.4 for having Ik(D) ⊆ I (j) W is sharp. Indeed, suppose that f ∈ C[x1, . . . , xr] is a homoge- neous degree m polynomial having an isolated singularity at 0. If D is the divisor in Spec C[x1, . . . , xn] defined by f , then a log resolution of (X, D) is given by the blow-up along W = V (x1, . . . , xr) and an easy computation shows that I0(D) = I m−r W . However, when k ≥ 1 the criterion is not sharp any more. For a sharp criterion for general k see Theorem E, proved below, which improves Corollary 19.4 in most cases. As another application of Theorem 19.1, we now show that the triviality of any of the higher ideals Ik(D) implies that D has rational singularities. We will show in fact that all the ideals Ik(D), for k ≥ 1, are contained in the adjoint ideal adj(D) (for the definition and basic properties of adjoint ideals, we refer to [Laz04, §9.3.E]). 48 M. MUSTAT¸ A AND M. POPA Proof of Theorem C. Note to begin with that it is enough to prove the first assertion, since adj(D) = OX implies that D is normal and has rational singularities by [Laz04, Proposition 9.3.48]. Moreover, since Ik(D) ⊆ I1(D) for every k ≥ 1 by Proposition 13.1, it is enough to show that I1(D) ⊆ adj(D). Let f : Y → X be a log resolution of (X, D) which is an isomorphism over X r D, such that the strict transform eD on Y is smooth. The adjoint ideal adj(D) is then defined by In order to prove the desired inclusion, it is enough to show that for every prime exceptional divisor G on Y we have adj(D) := f∗OY (KY /X − f∗D + eD). ordG(cid:0)I1(D)(cid:1) + kG ≥ ordG(D). (19.6) (19.7) We first note that by Theorem B, there is an open subset U ⊆ X with codim(X r U, X) ≥ 3 such that I1(D)U = J1(D)U . Since OX(−D) ⊆ it follows from Lemma 14.4 that adj(D) by definition, and Jac(cid:0)OX (−D)(cid:1) ⊆ adj(D) by [Laz04, Example 9.3.52], J1(D) ⊆ OX(−D) · Jac(cid:0)I0(D)(cid:1) + I0(D) · Jac(cid:0)OX(−D)(cid:1) ⊆ adj(D). In particular, we conclude that the inequality (19.6) holds if the center of the divisor G on X intersects U . From now on, we assume that this center is contained in X r U , hence its codimension in X is ≥ 3. By Theorem 19.1, we have where we use the notation in that theorem. We claim that it is enough to show that α1q ≤ kG − 1. Indeed, if this is the case, then (19.7) implies ordG(cid:0)I1(D)(cid:1) ≥ 2(ordG(D) − 1) − kG − α1q, ordG(cid:0)I1(D)(cid:1) ≥ 2(ordG(D) − kG) − 1. If ordG(D) ≥ kG + 1, this implies ordG(cid:0)I1(D)(cid:1) + kG − ordG(D) ≥ ordG(D) − kG − 1 ≥ 0, hence the inequality in (19.6) holds. On the other hand, if ordG(D) ≤ kG, then (19.6) clearly holds. This shows the claim, and so we are left with proving that α1q < kG. With the notation in the proof of Theorem 19.1, let ci = codim(Wi−1, Xi−1), for 1 ≤ i ≤ s. Recall that we have a sequence of maps Us −→ Us−1 −→ ··· −→ U1 −→ U0 ⊆ X such that each Ui → Ui−1 is the blow-up of the smooth subvariety Wi−1 ∩ Ui−1, followed by an open immersion. Let Gi ⊆ Ui be the corresponding HODGE IDEALS 49 exceptional divisor, so that KUi/Ui−1 = (ci − 1)Gi. If gi : Us → Ui is the induced map, then we have KUs/U0 = sXi=1 g∗i KUi/Ui−1. Therefore we have kG = ordG(KUs/U0) = sXi=1 (ci − 1) ordG(Gi) = sXi=1 (ci − 1) ordG(IWi−1) = sXi=1 (ci − 1)αi. By construction, we have ci ≥ 2 for 1 ≤ i ≤ s. Furthermore, by our assumption on G we have c1 > 2. We thus deduce that kG ≥ 2α1 + α2 + ··· + αs. On the other hand, we have α1q = α1⌈(α1+···+αs)/α1⌉ < α1(cid:18) α1 + ··· + αs α1 + 1(cid:19) = 2α1+α2+···+αs. We finally conclude that α1q < kG. (cid:3) Remark 19.8. In general it is far from being true that if D has rational singularities, then the Hodge ideal I1(D) is trivial. For example, if X = An with n ≥ 3, and D is the cone over a smooth hypersurface in Pn−1 of degree m, then it follows from Proposition 20.2 below that I1(D) = OX if and only if m ≤ n 2 . On the other hand, it is well known (and an easy exercise) that D has rational singularities if and only if m ≤ n − 1. 20. Examples. We now discuss a few examples and further useful cal- culations. The most significant is the computation of the order of k-log canonicity of an ordinary singularity, i.e. Theorem D. Example 20.1 (I1 for ordinary singularities). We begin by treating the case of the ideal I1(D), for which the argument is easier and we can obtain a more detailed result. Suppose that X has dimension n ≥ 3 and W = {x} is a point. We consider the case of an ordinary singularity, i.e. when m = multx(D) ≥ 2 and the projectivized tangent cone of D at x is smooth. For instance, D could be the cone over a smooth hypersurface of degree m in Pn−1. Proposition 20.2. With these hypotheses, around x we have: (1) If m ≤ n (2) If n 2 , then I1(D) = OX . 2 ≤ m ≤ n − 1, then I1(D) = m2m−n x . 50 M. MUSTAT¸ A AND M. POPA (3) If m ≥ n, we have OX (−D) · m + m m−n−1 x 2m−n x with dimC I1(D)/(OX (−D) · mm−n−1 x ⊆ I1(D) ⊆ OX (−D) · m + m2m−n x Note that by the proposition, for j ≤ n − 1 we have . x ⇐⇒ m ≥ For j ≥ n, we have the following implications: I1(D) ⊆ m n + j 2 j m−n−2 x + m 2m−n−1 x , ) = m(cid:0)m−2 n−2(cid:1). n + j + 2 m ≥ 2 ⇒ I1(D) ⊆ m j x ⇒ m ≥ n + j + 1 2 . Proof of Proposition 20.2. After possibly replacing X by a neighborhood of x, we may assume that the blow-up f : Y → X is a log resolution of (X, D). from Theorem 18.1 that we have an exact sequence This follows from the fact that if F is the exceptional divisor and eD is the strict transform of D, then eD ∩ F ֒→ F ≃ Pn−1 is the projectivized tangent cone of D at x, hence smooth by assumption. Let E = eD + F . It follows (20.3) 0 → ωX(2D)⊗ f∗(cid:0)I1(E)· OY(cid:0)(n + 1− 2m)F(cid:1) → ωX(2D)⊗ I1(D) → → f∗(cid:0)TY /X ⊗ ωY (E)(cid:1) → ωX(2D) ⊗ R1f∗(cid:0)I1(E) · OY ((n + 1 − 2m)F(cid:1). f∗D = eD + mG, we see that Recall now that by Example 18.7, we have TY /X ≃ TF ⊗ OF (−1). Since TY /X ⊗ ωY (E) = TY /X ⊗ ωF ⊗ OY (eD) ≃ TF ⊗ OF (m − n − 1). We deduce that sequence). On the other hand, it follows from Proposition 8.2 that I1(E) is equal to f∗(cid:0)TY /X ⊗ ωY (E)(cid:1) ≃ H 0(cid:0)Pn−1, TPn−1 (m − n − 1)(cid:1) and we distinguish two cases. When m ≤ n−1, then f∗(cid:0)TY /X ⊗ωY (E)(cid:1) = 0, while for m ≥ n, the sheaf f∗(cid:0)TY /X ⊗ ωY (E)(cid:1) is a skyscraper sheaf of length m(cid:0)m−2 n−2(cid:1) (this follows from an easy computation using the Euler exact OY (−F ) + OY (−eD). Consider the following exact sequence on Y : 0 −→ OY (−eD − F ) −→ OY (−eD) ⊕ OY (−F ) −→ I1(E) −→ 0. By tensoring with OY(cid:0)(n + 1 − 2m)F(cid:1) and applying f∗, we obtain an exact f∗OY(cid:0)−eD+(n+1−2m)F(cid:1)⊕f∗OY(cid:0)(n−2m)F(cid:1) → f∗OY(cid:0)I1(E)·OY ((n+1−2m)F )(cid:1) → R1f∗OY(cid:0)−eD+(n−2m)F(cid:1) → R1f∗OY(cid:0)−eD+(n+1−2m)F(cid:1)⊕R1f∗OY(cid:0)(n−2m)F(cid:1) → R1f∗OY(cid:0)I1(E) · OY ((n + 1 − 2m)F )(cid:1) → R2f∗OY(cid:0) − eD + (n − 2m)F(cid:1). sequence HODGE IDEALS 51 m−n−1 x Note that as n ≥ 3, we have R1f∗OY (jF ) = 0 for all j ∈ Z. We also have R2f∗OY (jF ) = 0, unless n = 3 and j ≥ 3. Since f∗OX(−D) = OY (−eD − mF ), using the projection formula we obtain , f∗OY(cid:0)(n−2m)F(cid:1) = m f∗OY(cid:0)−eD+(n+1−2m)F(cid:1) = OX(−D)·m R1f∗OY(cid:0) − eD + (n − 2m)F(cid:1) = R2f∗OY(cid:0) − eD + (n − 2m)F(cid:1) = 0, R1f∗OY(cid:0) − eD + (n + 1 − 2m)F(cid:1) = R1f∗OY(cid:0)(n − 2m)F(cid:1) = 0. f∗OY(cid:0)I1(E) · OY ((n + 1 − 2m)F )(cid:1) = OX (−D) · m R1f∗OY(cid:0)I1(E) · OY ((n + 1 − 2m)F )(cid:1) = 0. x = OX if ℓ ≤ 0.) We deduce from the above (We use the convention that mℓ exact sequence that The conclusion now follows from the exact sequence (20.3). m−n−1 x and + m 2m−n x and 2m−n x , (cid:3) Example 20.4 (Ik for ordinary singularities, I). Suppose that we are still in the case when X is a smooth n-dimensional variety, x ∈ X is a point, and D is a reduced effective divisor with multx(D) = m, whose projectivized tangent cone at x is smooth. We now show that m ≤ n k + 1 =⇒ Ik(D) = OX around x. This extends Proposition 20.2 (1). It would be very interesting to have analogues of its other statements for k ≥ 2; in this direction, we will see in Example 20.10 that the implication above is in fact an equivalence. To prove the assertion, as we have already seen, after passing to a suitable neighborhood of x we may assume that the blow-up f : Y → X at x is a log and F is the exceptional divisor, then it follows from Theorem 18.1 that we have an inclusion resolution of (X, D). If E = eD + F , where eD is the strict transform of D f∗(cid:0)Ik(E) · OY(cid:0)(n + k − (k + 1)m)F )(cid:1) ֒→ Ik(D). Therefore it is enough to show that f∗(cid:0)Ik(E) · OY (aF )(cid:1) = OX, where a = n + k − (k + 1)m. Now by Proposition 8.2 we have Ik(E) =(cid:0)OY (−F ) + OY (−eD)(cid:1)k ⊇ OY (−kF ), hence it is enough to have f∗OY(cid:0)(a − k)F(cid:1) = OX . This holds since by assumption a − k = n + k − (k + 1)m − k ≥ 0. 52 M. MUSTAT¸ A AND M. POPA Example 20.5 (Non-ordinary singularities). Using the Restriction The- orem for Hodge ideals, we deduce from the bound in Example 20.4 that if X is a smooth n-dimensional variety, x ∈ X is a point, and D is a reduced effective divisor with multx(D) = m such that the projectivized tangent cone P(CxD) of D at x has a singular locus of dimension r, then m ≤ n − r − 1 k + 1 =⇒ Ik(D) = OX around x. Indeed, we may assume that X is affine and that we have a system of algebraic coordinates x1, . . . , xn on X, centered at x. If H is defined by a general linear combination of the xi, then H is smooth, not contained in Supp(D), and DH is reduced. Furthermore, we have multx(DH ) = m and P(Cx(DH )) is a general hyperplane section of P(CxD). In particular, if r ≥ 0, then we have On the other hand, it follows from Theorem 16.9 that dim P(Cx(DH ))sing = r − 1. The assertion thus follows by induction on r, with the case r = −1 being covered by Example 20.4. Ik(DH ) ⊆ Ik(D) · OH. Example 20.6 (Ik for ordinary singularities, II). With more work, we can obtain a description for Ik(D) for a larger range of multiplicities than in Example 20.4, in a similar vein with what we did in Proposition 20.2 for k = 1. Suppose that X is a smooth variety of dimension n ≥ 3, D is a reduced effective divisor on X, and x ∈ X is a point such that m = multx(D) ≥ 2. We assume that the projectivized tangent cone of D at x is smooth. Proposition 20.7. Under the above hypotheses, for every k such that mk < n, we have around x, with the convention that m Ik(D) = m j (k+1)m−n x x = OX if j ≤ 0. Proof. Let f : Y → X be the blow-up of X at x, with exceptional divisor F ≃ Pn−1. The assumption implies that after possibly replacing X by an open neighborhood of x, we may assume that f is a log resolution of (X, D). Let E = eD + F , where eD is the strict transform of D. The key point will be to show that our condition on k implies that Ik(D) = f∗(cid:0)Ik(E) ⊗ O(KY /X + (k + 1)E − (k + 1)g∗D)(cid:1). We temporarily denote by ak the right-hand side in the above formula. Recall that we have the filtered complex A• 0 → f∗DX → Ω1 Y (log E) ⊗OY f∗DX → ··· → ωY (E) ⊗OY f∗DX → 0 placed in degrees −n, . . . , 0, and its filtered subcomplex B• DY → ··· → ωY (E) ⊗OY 0 → DY → Ω1 Y (log E) ⊗OY DY → 0. Consider the complex M• defined by the exact sequence of complexes HODGE IDEALS 53 0 → Fk−nB• → Fk−nA• → M• → 0. It follows from the proof of Theorem 18.1 that the inclusion can be identified with the induced morphism ak ⊗ ωX(cid:0)(k + 1)D(cid:1) ֒→ Ik(D) ⊗ ωX(cid:0)(k + 1)D(cid:1) R0f∗Fk−nB• → R0f∗Fk−nA•. In order to show that ak = Ik(D) it is thus enough to verify that R0f∗M• = 0. On the other hand, from the hypercohomology spectral sequence Ep,q 1 = Rqf∗M p−n =⇒ Rp+q−nf∗M• we deduce that in order to have R0f∗M• = 0 it is enough to prove that for every 0 ≤ q ≤ n we have Rqf∗M−q = 0. To this end, note first that from the definition of M• we have M−q = Ωn−q Y (log E) ⊗(cid:0)f∗Fk−qDX /Fk−qDY(cid:1). The sheaf f∗Fk−qDX /Fk−q DY has a filtration with successive quotients f∗SjTX /SjTY for 1 ≤ j ≤ k − q. On the other hand, Lemma 20.9 below implies that f∗SjTX/SjTY has a filtration with successive quotients Sj−iTF ⊗ OF (−ℓ− i), with 0 ≤ i ≤ j − 1 and 1 ≤ ℓ ≤ j − i. We thus conclude that in order to prove that ak = Ik(D) it is enough to show that Y H q(cid:0)F, Ωn−q 0 → Ωp for 0 ≤ q ≤ n, 1 ≤ j ≤ k − q, 0 ≤ i ≤ j − 1, and 1 ≤ ℓ ≤ j − i. Note now that for every p we have a short exact sequence (log E)F ⊗ Sj−iTF ⊗ OF (−ℓ − i)(cid:1) = 0 F (log eDF ) → 0. Y (log E)F → Ωp−1 Restricting this to F gives an exact sequence 0 → Ωp−1 On the other hand, the short exact sequence for sheaves of differential forms corresponding to the closed immersion F ֒→ Y induces an exact sequence F (log eDF ) → 0. Y (log E) → Ωp−1 Y (log eD)F → Ωp Y (log eD)F → Ω1 Y (log eD) → Ωp F (log eDF )⊗OF (1) → Ωp 0 → OF (1) → Ω1 F (log eDF ) ⊗ OF (1) → Ωp 0 → Ωp Y (log eD)F → Ωp Y (log E)F → Ωp−1 0 → Ωp−1 and by combining all of this we conclude that we have an exact sequence By taking pth exterior powers we obtain an exact sequence F (log eDF ) → 0. F (log eDF ) → 0, Using the corresponding long exact sequence in cohomology, we conclude that ak = Ik(D) holds if for all q, j, i, and ℓ as above, we have F (log eDF ) → Ωp (log Z) ⊗ Sj−iTF ⊗ OF (−ℓ − i)(cid:1) = 0, and F (log eDF ) → 0. (A1) H q(cid:0)F, Ωn−q F 54 M. MUSTAT¸ A AND M. POPA F (A2) H q(cid:0)F, Ωn−q−1 where Z = eDF . Now the Euler sequence on F gives rise to an exact Eagon- (log Z) ⊗ Sj−iTF ⊗ OF (−ℓ − i)(cid:1) = 0, 0 −→ Lj−i −→ . . . −→ L0 −→ Sj−iTF −→ 0, Northcott-type complex where each Ld is a direct sum of copies of OF (j − i − d). By breaking this into short exact sequences and taking the corresponding cohomology long exact sequences, we see that if A1) or A2) above fails, then there is an s with 0 ≤ s ≤ j − i such that either We now use the fact that since by assumption Z is a smooth, degree m F F (log Z) ⊗ OF (−ℓ + j − 2i − s)(cid:1) 6= 0, or (B1) H q+s(cid:0)F, Ωn−q (log Z) ⊗ OF (−ℓ + j − 2i − s)(cid:1) 6= 0. (B2) H q+s(cid:0)F, Ωn−q−1 hypersurface in F ≃ Pn−1, if H d(cid:0)F, Ωa F (log Z) ⊗ OF (b)(cid:1) 6= 0 for some d, a, and b, then one of the following conditions hold (see [BW08, Theorem 3.3]): (C1) d = 0, or (C2) d = n − 1, or (C3) 0 < d < n − 1, d + a = n − 1, and b ≥ n − m(d + 1). Suppose first that (B1) holds. F If we are in case (C1), then q = 0, in which case Ωn−q (log Z) = 0, a contradiction. If we are in case (C2), then q + s = n − 1. However, by assumption we have s ≤ j − i ≤ k − q, hence k ≥ n − 1, a contradiction with our hypothesis. Finally, we cannot be in case (C3) since s + n > n − 1. Suppose now that (B2) holds. If we are in case (C1), then q = s = 0 and Ωn−q−1 F (log Z) ⊗ OF (−ℓ + j − 2i − s) ≃ OF (−ℓ + j − 2i − n + m). Using the fact that this has nonzero sections, we conclude that 0 ≤ −ℓ + j − 2i − n + m ≤ k − 1 − n + m, contradicting the fact that km ≤ n − 1. We argue as in case (B1) that we cannot be in case (C2). Finally, if we are in case (C3), then s = 0 and On the other hand, we have by assumption −ℓ + j − 2i ≥ n − m(q + 1). −ℓ + j − 2i ≤ k − q − 1 and q ≤ k − 1, and by combining these inequalities, we obtain n ≤ mk, a contradiction. This completes the proof of the fact that Ik(D) = ak. By definition, we have ak = f∗(cid:0)Ik(E)⊗OY (eF )(cid:1), with e = n+k−m(k+1). It follows from Proposition 8.2 that Ik(E) =(cid:0)O(−eD) + O(−F )(cid:1)k. Therefore we have an exact Eagon-Northcott-type complex HODGE IDEALS 55 (20.8) with Gr = 0 −→ Gk −→ ··· −→ G0 −→ Ik(E) −→ 0. k−rMi=0 OY(cid:0) − (r + i)eD − (k − i)F(cid:1) for 0 ≤ r ≤ k. Since Rqf∗OY (pF ) = 0 for all p ∈ Z and all 1 ≤ q ≤ n − 2, and since k ≤ n − 2 by assumption, it follows easily that Rqf∗(Gq ⊗ OY (eF )) = 0 for 1 ≤ q ≤ k. By breaking the complex (20.8) into short exact sequences and using the corresponding cohomology long exact sequences, we deduce that the induced morphism kMi=0 is surjective. Note that f∗OY(cid:0)−ieD−(k−i−e)F(cid:1) → f∗(cid:0)Ik(E)⊗O(eF )(cid:1) = ak f∗(cid:0)G0⊗OY (eF )(cid:1) = f∗OY(cid:0) − ieD − (k − i − e)F(cid:1) = f∗OY(cid:0) − if∗D − (k − i − e − im)F(cid:1) = OY (−iD) · m k−i−e−im x . Since k − i − e − im = m(k + 1) − n − i(m + 1), we see that if m(k + 1) ≤ n, then ak = OX (we have of course already seen this in Example 20.4), and if m(k + 1) > n, then ak = f∗OY(cid:0) − (k − e)F(cid:1) = m m(k+1)−n x . m(k+1)−n Here we use that OX(−D) ⊆ m x completes the proof of the proposition. , due to the fact that mk < n. This (cid:3) Lemma 20.9. Let f : Y → X be the blow-up of a smooth n-dimensional variety X at a point x ∈ X, with exceptional divisor F . For every j ≥ 1, the sheaf f∗SjTX/SjTY has a filtration with successive quotients Sj−iTF ⊗ OF (−ℓ − i), with 0 ≤ i ≤ j − 1 and 1 ≤ ℓ ≤ j − i. Proof. By taking an ´etale morphism X → An mapping x to the origin, we reduce by base-change to the case when X = An = Spec(S•V ) and x is the origin. Recall that if P = Proj(S•V ), then we have a closed embedding j : Y ֒→ X × P. Let p : X × P → X and q : X × P → P be the canonical projections, so that p ◦ j = f and Y is isomorphic as a scheme over P (via g = q ◦ j) to Spec(S•OP(1)). In particular, this implies that g is smooth and TY /P ≃ g∗OP(−1) = OY (−1). 56 M. MUSTAT¸ A AND M. POPA On Y we have a commutative diagram with exact rows 0 0 TY /P α TY β OY (−1) f∗TX = V ∨ ⊗ OX g∗TP γ g∗(cid:0)TP(−1)(cid:1) 0 0, in which the top row is the exact sequence of tangent sheaves for the smooth morphism g and the bottom row is obtained by pulling back the twisted Euler exact sequence on P via g. Note that g∗(cid:0)TP(−1)(cid:1) = g∗TP ⊗ OY (F ) and α is an isomorphism, hence the Snake Lemma gives an isomorphism TY /X ≃ g∗TP ⊗ OF (F ) = TF ⊗ OF (−1) (cf. Example 18.7). filtration The bottom exact sequence in the above diagram induces on f∗SjTX a such that for every i with 0 ≤ i ≤ j we have 0 = Mj+1 ⊆ Mj ⊆ ··· ⊆ M1 ⊆ M0 = f∗SjTX Mi/Mi+1 ≃ Sj−ig∗(cid:0)TP(−1)(cid:1) ⊗ OY (−i). It follows from the top exact sequence in the diagram that the induced filtration on SjTY given by Ni = Mi ∩ SjTY has the property that Ni/Ni+1 ≃ Sj−ig∗TP ⊗ OY (−i). An easy calculation now shows that the induced quotient filtration has successive quotients for 0 ≤ i ≤ j − 1. Finally, it is straightforward to see that 0 = Mj+1 ⊆ Mj ⊆ ··· ⊆ M1 ⊆ M0 = f∗SjTX/SjTY Mi/Mi+1 ≃(cid:0)Sj−ig∗(TP(−1))/Sj−ig∗TP(cid:1) ⊗ OY (−i), ≃(cid:0)Sj−ig∗TP ⊗ OY ((j − i)F )/Sj−ig∗TP(cid:1) ⊗ OY (−i) Sj−ig∗TP ⊗ OF (−ℓ − i) ≃ Sj−iTF ⊗ OF (−ℓ − i) (cid:0)Sj−ig∗(TP(−1))/Sj−ig∗TP(cid:1) ⊗ OY (−i) has a filtration with successive quotients for 1 ≤ ℓ ≤ j − i, which gives the assertion in the lemma. Example 20.10 (Proof of Theorem D). It follows from Proposition 20.7 that the bound in Example 20.4 is sharp, that is, if m ≥ 2 and k is such that (cid:3) k + 1 ≤ n m < k + 2, then around x we have Ik+1(D) 6= OX. This completes the proof of Theorem D. The statement follows directly from the proposition if k + 1 < n m . To check the case k + 1 = n m , we consider X′ = X × A1 and the divisor D′ defined locally by h + zm, where h is a local equation of D and z is the HODGE IDEALS 57 coordinate on A1. In this case D′ has a smooth projectivized tangent cone at x′ = (x, 0), of degree m, and we use Proposition 20.7 to conclude that Ik+1(D′) ⊆ mx′ around x′. On the other hand, if we consider X = X × {0} ֒→ X′, then D = D′X and Theorem 16.9 gives Ik+1(D) ⊆ Ik+1(D′) · OX ⊆ mx around x. This applies, for example, when X has an ordinary double point at x (that is, the projectivized tangent cone of X at x is a smooth quadric) to give that in this case Ik(D) = OX if and only if k ≤ [n/2] − 1. In this case the result was already proved in [DSW09, §1.4], which in fact shows more, namely One implication follows in fact already from [Sai93]; see the next remark. Ik(D)x = m k−[n/2]+1 x for all k ≥ 0. Remark 20.11. By making use of V -filtrations, Saito gave a useful criterion for the pair (X, D) to be k-log-canonical at some x ∈ D in terms of the Bernstein-Sato polynomial of D at x. Suppose that f is a local equation of D. Recall that the Bernstein-Sato polynomial of D at x is the monic polynomial bf,x ∈ C[s] of smallest degree with the property that around x there is a relation bf,x(s)f s = P (s) • f s+1 for some nonzero P ∈ DX [s]. It is known that (s+1) divides bf,x and all roots of bf,x(s) are negative rational numbers. One defines αf,x to be −λ, where λ is the largest root of bf,x(s)/(s + 1). It is shown in [Sai93, Theorem 0.11] that around x we have Ik(D) = OX for all k ≤ αf,x − 1.2 Saito also showed in [Sai09, Theorem 0.7] that if x ∈ D is an isolated quasi- homogeneous singularity, then the Hodge filtration on ωX(∗D) is generated around x in level [n − αf,x] − 1. Example 20.12 (Diagonal hypersurfaces). Consider the case when D i , with ai ≥ 2. There is a general description for the roots of the Bernstein-Sato polynomial for quasi- homogeneous, isolated singularities (see [Yan78, §11]). In our case, this says that is the divisor in An defined by f = Pn i=1 xai bf,0(s) = (s + 1) · Yb1,...,bn s + nXi=1 bi ai! , 2Since this was written, Saito [Sai16b, Corollary 1] has shown that this is in fact an if and only if statement. In particular, this determines the k-log canonicity level at x as being ⌊αf,x⌋. Combined with the formulas in the next two examples, and with other known facts about αf,x, this provides an alternative approach to many of the results in Theorems A, C and D; see [Sai16b]. 58 M. MUSTAT¸ A AND M. POPA where the product is over those bi with 1 ≤ bi ≤ ai−1 for all i. In particular, , and it follows from Remark 20.11 that we have αf,0 =Pn 1 ai i=1 Ik(D) = OX for all k ≤ −1 + 1 ai . nXi=1 When α1 = ··· = αn = d, this also follows from Example 20.4, while Example 20.10 says that in this case the estimate is sharp. Example 20.13 (Semiquasihomogeneous isolated singularities). More generally, suppose that D ⊂ An has a semiquasihomogeneous isolated sin- gularity at x in the sense of [Sai09]. This means that we have local coor- dinates x1, . . . , xn centered at x and weights w1, . . . , wn ∈ Q>0 such that a local equation f of D at x can be written as g + h, where g only involves monomials of weighted degree 1, it has an isolated singularity at x, and h only involves monomials of weighted degree > 1. In this case, Saito showed in [Sai89] that αf,x =Pn Note that D has an ordinary singularity at p if and only if it has a semi- homogeneous isolated singularity at p (in the sense that it satisfies the above definition with w1 = ··· = wn). We thus see that in this case, if wi = 1 d for all i, then αf,x = n d . Using Remark 20.11, this gives another way of seeing Proposition 20.7. Example 20.14 (Generic determinantal hypersurface). Let X ≃ An2 be the affine space of n × n matrices, with n ≥ 2, and let D be the reduced, irreducible divisor given by i=1 wi. It is an observation that goes back to Cayley that if f = det, then D = {A det(A) = 0}. (s + 1)(s + 2)··· (s + n)f s = det(∂i,j)n i=1(s + i) for every x ∈ X. hence bf,x(s) divides Qn It follows from Re- mark 20.11 that in this case we have I1(D) = OX . This is optimal: in fact, the zero set of I2(D) is the singular locus of D: 1 • f s+1, Dsing = {A ∈ D rank(A) ≤ n − 2}. Indeed, if A ∈ Dsing is a point with rank(A) = n− 2, then D has an ordinary double point at A, and I2(D) vanishes at A by Example 20.10. 21. Order of vanishing along a closed subset. We can now prove our main criterion for the Hodge ideals of a divisor D to be contained in the symbolic power of the ideal defining an irreducible closed subset. In most cases this is a stronger statement than the criterion in Corollary 19.4. Proof of Theorem E. The assertion is trivial when m ≤ 1, hence from now on we assume that m ≥ 2. After replacing X by a suitable affine open subset intersecting W , we may assume that X is affine and that we have an algebraic system of coordinates x1, . . . , xn such that IW = (x1, . . . , xr). HODGE IDEALS 59 Moreover, we may and will assume that D is defined by a principal ideal (g). We first reduce to the case when W is a point. It is enough to show that if the theorem holds when dim W = d ≥ 0, then it also holds when dim W = d + 1. For every λ = (λ0, . . . , λn) ∈ Cn+1 r{0}, consider the smooth divisor Hλ defined by ℓλ = λ0 +Pn i=1 λixi. It follows from Theorem 16.1 that for λ general, we have Ik(DHλ ) = Ik(D) · OHλ. (21.1) Since dim W = d + 1 > 0, for general λ the subset W ∩ Hλ is non-empty and smooth, of dimension d, and multW∩Hλ(DHλ) = m. The assertion in the theorem for W ∩ Hλ ⊆ Hλ, together with the equality (21.1), implies Ik(D) ⊆ (x1, . . . , xr)q + (ℓλ). It is straightforward to see that in this case we have Ik(D) ⊆ (x1, . . . , xn)q, as required. From now on we assume that W = {x} is a point, hence r = n. Suppose first that q = (k+1)m−n or, equivalently, that km < n. Let AN be the affine space parametrizing the coefficients of homogeneous polynomials of degree ≥0, with u :=Pi ui = m. m, with coordinates cu, for u = (u1, . . . , un) ∈ Zn Let p : X × AN → AN be the second projection and s : AN → X × AN be given by s(t) = (x, t). We consider the effective divisor F on X × AN defined by g +Pu=m cuxu. Let U be the open subset of AN consisting of those t ∈ AN such that Ft := F ∩ (X ×{t}) is a divisor on X; note that the origin lies in U . Since F ∩ (X × U ) is flat over U , the set V = {(y, t) ∈ X × U y 6∈ F, or y ∈ F and Ft is reduced at y} is open in X × U by [Gro66, Th´eor`eme 12.1.6]. Moreover, we have (x, 0) ∈ q V . Arguing by contradiction, let us assume that Ik(D) 6⊆ m x. Applying Theorem 16.11 to the map h : Z = V ∩ p−1(s−1(V )) → s−1(V ) = T, the section T → Z induced by s, the divisor FZ , and the point t0 = 0 ∈ T , q we conclude that for a general t ∈ s−1(V ), we have Ik(Ft ∩ V ) 6⊆ m x. However, for t ∈ V general, the projectivized tangent cone of Ft at x is a general hypersurface of degree m in Pn−1, hence smooth. In this case, q since km < n, it follows from Proposition 20.7 that Ik(Ft ∩ V ) = m x in a neighborhood of x, a contradiction. This completes the proof of the theorem in the case mk < n. Suppose now that mk ≥ n and let d = mk−n+1. Consider the divisor D′ in X′ = X × Ad which is the inverse image of D via the first projection. We consider X embedded in X′, defined by the ideal (z1, . . . , zd). Note that D = D′X and D′ is reduced. Applying Theorem 16.9 (see also Remark 16.10) we see that Ik(D) ⊆ Ik(D′) · OX . 60 M. MUSTAT¸ A AND M. POPA On the other hand, since mult(x,0)(D′) = m, and we have mk < n + d and (k + 1)m − (n + d) = m − 1, the case we have already treated implies that Ik(D′) ⊆ m . This completes the proof of the theorem. (cid:3) m−1 (x,0) and therefore Ik(D) ⊆ mm−1 x Example 21.2. We spell out what this criterion says when k = 1 and W = {x} is a single point. If m = multx(D), then m ≥ max(cid:26)q + 1, n + q 2 (cid:27) =⇒ I1(D) ⊆ m q x. Taking m ≥ 2 and ensuring that q ≥ 1 in Theorem E gives: Corollary 21.3. Let D be a reduced effective divisor on the smooth variety X. If W is an irreducible closed subset of X of codimension r such that m = multW (D) ≥ 2, then Ik(D) ⊆ IW for all k ≥ In particular, if W ⊆ Dsing, then r + 1 − m . m Ik(D) ⊆ IW for all k ≥ r − 1 2 . Example 21.4. Corollary 21.3 implies that the nontriviality part of The- Indeed, if x ∈ D is a point orem D holds for arbitrary singular points. of multiplicity m, the corollary implies that Ik(D) becomes nontrivial at x when k ≥ n+1−m m , or equivalently Ik(D)x ⊆ mx for k ≥h n mi . The last statement in Corollary 21.3 immediately implies one of the main results stated in the Introduction, namely the fact that the smoothness of D is precisely characterized by the triviality of all Hodge ideals, or equivalently by the equality between the Hodge and pole order filtrations. Proof of Theorem A. We know that if D is smooth, then Ik(D) = OX for all k. On the other hand, it follows from Corollary 21.3 that if D is singular, then Ik(D) 6= OX for all k ≥ n−1 2 . (cid:3) G. Vanishing theorems In this section we prove the fundamental vanishing theorem for the Hodge ideals Ik(D), extending Nadel vanishing for the multiplier ideal I0(D). For k = 1 this can be done using more elementary methods, but at the moment in the general case we only know how to argue based on Saito's Kodaira-type vanishing theorem for mixed Hodge modules, recalled as Theorem 5.1. for all i > 0 and all k. Proof. If U = X r D, then by Corollary 7.2 we have Hi(cid:0)X, grF H i+n(U, C) ≃Mq∈Z k DR(OX(∗D))(cid:1) = 0 Hi(cid:0)X, grF −q DR(OX (∗D))(cid:1). HODGE IDEALS 61 22. General vanishing. Besides Theorem 5.1, we will also make use of a different vanishing result for the Hodge D-module OX(∗D). It is an im- mediate consequence of Saito's strictness results, surely well-known to the experts. Proposition 22.1. If X is a smooth projective variety and D is a reduced effective ample divisor on X, then Since U is affine, the left hand side is 0 for all i > 0 by the Andreotti- [Laz04, Theorem 3.1.1]. This implies Frankel vanishing theorem; see e.g. the vanishing of all spaces on the right-hand side. (cid:3) 23. Vanishing for Hodge ideals. For motivation, we start by recalling a well-known fact: Proposition 23.1. Let X be a smooth projective variety, D an effective divisor, and L an ample line bundle on X. Then: (1) H i(cid:0)X, ωX (D) ⊗ L ⊗ I0(D)(cid:1) = 0 for all i ≥ 1. nef, e.g. L = OX. (2) If D is ample, then the same vanishing holds if we only assume L is Proof. This is just a special case of Nadel vanishing; see [Laz04, Theorem 9.4.8]. Indeed, recall that I0(D) = I(X, (1−ǫ)D), and so one has the desired vanishing as long as L + ǫD is ample for 0 < ǫ ≪ 1. This holds under either hypothesis. (cid:3) We now move to analogous results for k ≥ 1. We first state the case k = 1. We will then provide a general result, in a slightly weaker form for simplicity; it is necessarily an inductive, and more technical, statement. Theorem 23.2. Let X be a smooth projective variety, D a reduced effective divisor such that the pair (X, D) is log-canonical, and L a line bundle on X. Then: (1) If L is an ample line bundle such that L(D) is also ample, then and there is a surjection H i(cid:0)X, ωX (2D) ⊗ L ⊗ I1(D)(cid:1) = 0 for all X (D) ⊗ L(cid:1) → H 1(cid:0)X, ωX (2D) ⊗ L ⊗ I1(D)(cid:1) → 0. i ≥ 2, H 1(cid:0)X, Ωn−1 (2) If D is ample, then the conclusion in (1) also holds for L = OX. 62 M. MUSTAT¸ A AND M. POPA Proof. Note to begin with that H i(cid:0)X, ωX (2D) ⊗ L ⊗ I1(D)(cid:1) ≃ H i(cid:0)X, ωX ⊗ L ⊗ grF 1 for all i ≥ 1, and so it suffices to prove the analogous statements for the cohomology groups on the right. Indeed, given that I0(D) = OX , we have a short exact sequence 0 −→ ωX ⊗ L(D) −→ ωX(2D)⊗ L⊗ I1(D) −→ ωX ⊗ L⊗ grF OX (∗D) −→ 0. The isomorphisms follow from the fact that the leftmost term in this se- quence satisfies Kodaira vanishing. 1 OX (∗D)(cid:1) Suppose first that the conditions in (1) hold. Let n = dim X and consider the complex Since I0(D) = OX , C• can be identified with a complex −n+1 DR(OX(∗D)) ⊗ L(cid:1)[−1]. C• :=(cid:0) grF (cid:2)Ωn−1 X ⊗ OX(D) ⊗ L −→ ωX ⊗ L ⊗ grF 1 OX(∗D)(cid:3) with terms in degrees 0 and 1. Theorem 5.1 implies that (23.3) Hj(X, C•) = 0 for all j ≥ 2. We use the spectral sequence Ep,q 1 = H q(X, C p) =⇒ Hp+q(X, C•). Note that for i ≥ 1, E2,i 1 = 0 since C 2 = 0. On the other hand, for i ≥ 2 we have E0,i 1 = 0 by Nakano vanishing, which implies for all i ≥ 2. r we have that the outgoing term is 0 because of the length of the complex, while the incoming term is clearly 0. Using (23.3), this implies that Now for every r ≥ 2 and i ≥ 1, for E1,i E1,i 1 = E1,i 2 E1,i 2 = E1,i ∞ = 0. Therefore E0,1 1 = H 1(cid:0)X, Ωn−1 X (D) ⊗ L(cid:1) surjects onto E1,1 1 = 0 for i ≥ 2. This proves (1). The proof of (2) is identical, replacing the use of Theorem 5.1 by Proposition 22.1. (cid:3) 1 = H 1(cid:0)X, ωX (2D) ⊗ L ⊗ I1(D)(cid:1), while E1,i We now prove the vanishing theorem for arbitrary k, i.e. Theorem F in the introduction. Note that for k = 1 it is implied by Theorem 23.2. Recall that the assumption is that the pair (X, D) is (k − 1)-log-canonical, which is equivalent to I0(D) = I1(D) = ··· = Ik−1(D) = OX . HODGE IDEALS 63 Passing to cohomology, by assumption Kodaira vanishing applies to the two extremes, which implies vanishing for the term in the middle. Proof of Theorem F. The statement for k ≥ n+1 2 , i.e. part (2), is simply an application of Kodaira vanishing. Indeed, D is smooth, and we have a short exact sequence 0 −→ ωX ⊗ L(kD) −→ ωX ⊗ L(cid:0)(k + 1)D(cid:1) −→ ωD ⊗ L(kD)D −→ 0. We thus concentrate on the case k ≤ n 0 → ωX ⊗L(kD) → ωX ⊗L(cid:0)(k +1)D(cid:1)⊗Ik(D) → ωX ⊗L⊗grF OX(∗D) → 0. Passing to cohomology and using Kodaira vanishing, this implies immedi- ately that the vanishing statements we are aiming for are equivalent to the same vanishing statements for Ik−1(D) = OX , we have a short exact sequence 2 , i.e. part (1). Note first that since k OX(∗D)(cid:3) We now consider the complex k OX (∗D)(cid:1). −n+k DR(OX (∗D)) ⊗ L(cid:1)[−k]. H i(cid:0)X, ωX ⊗ L ⊗ grF C• :=(cid:0) grF (cid:2)Ωn−k X ⊗ L(D) −→ Ωn−k+1 ··· −→ Ωn−1 X ⊗ L ⊗ OD(kD) −→ ωX ⊗ L ⊗ grF X k ⊗ L ⊗ OD(2D) −→ ··· Given the hypothesis on the ideals Ip(D), this can be identified with a complex of the form concentrated in degrees 0 up to k. Theorem 5.1 implies that (23.4) Hj(X, C•) = 0 for all j ≥ k + 1. We use the spectral sequence Ep,q 1 = H q(X, C p) =⇒ Hp+q(X, C•). 1 1 hand, Ek−1,i and so we have an exact sequence 1 with = 0 since C k+1 = 0. On the other The vanishing statements we are interested in are for the terms Ek,i i ≥ 1. Note to begin with that Ek+1,i = H i(cid:0)X, Ωn−1 X ⊗ L(kD)(cid:1) −→ Ek−1,i H i(cid:0)X, Ωn−1 (1) If i ≥ 2, we deduce that Ek−1,i (2) If i = 1, using Nakano vanishing we obtain a surjective morphism X ⊗ L ⊗ OD(kD)(cid:1), −→ H i+1(cid:0)X, Ωn−1 X ⊗ L((k − 1)D)(cid:1). = 0 by Nakano vanishing, and so Now there are two cases: 1 = Ek,i Ek,i 2 . 1 1 H 1(cid:0)X, Ωn−1 X ⊗ L(kD)(cid:1) −→ Ek−1,1 1 . If the extra vanishing hypothesis on the term on the left holds, then we draw the same conclusion as in (1). 64 M. MUSTAT¸ A AND M. POPA We need to analyze in a similar way the terms Ek,i we always have Ek+r,i−r+1 other hand, we will show that under our hypothesis we have Ek−r,i+r−1 from which we infer that Ek−r,i+r−1 Ek,i r with r ≥ 2. On one hand, = 0 because of the length of the complex. On the = 0, = 0 as well. Granting this, we obtain 1 r r r = Ek,i r+1. Repeating this argument for each r, we finally obtain Ek,i 1 = Ek,i ∞ = 0, where the vanishing follows from (23.4), as i ≥ 1. We are thus left with proving that Ek−r,i+r−1 1 since the complex C• starts in degree 0. If k = r, we have = 0. If r > k this is clear, E0,i+k−1 1 = H i+k−1(cid:0)X, Ωn−k X ⊗ L(D)(cid:1). If i ≥ 2 this is 0 by Nakano vanishing, while if i = 1 it is 0 because of our hypothesis. Finally, if k ≥ r + 1, we have which sits in an exact sequence Ek−r,i+r−1 1 H i+r−1(cid:0)X, Ωn−r X ⊗ L ⊗ OD((k − r + 1)D)(cid:1), = H i+r−1(cid:0)X, Ωn−r X ⊗ L((k − r + 1)D)(cid:1) −→ Ek−r,i−r+1 −→ H i+r(cid:0)X, Ωn−r X ⊗ L((k − r)D)(cid:1). −→ 1 We again have two cases: (1) If i ≥ 2, we deduce that Ek−r,i+r−1 1 = 0 by Nakano vanishing. (2) If i = 1, using Nakano vanishing we obtain a surjective morphism H r(cid:0)X, Ωn−r X ⊗ L((k − r + 1)D)(cid:1) −→ Ek−r,i+r−1 and if the extra hypothesis on the term on the left holds, then we draw the same conclusion as in (1). 1 , The proof of (3) is identical, replacing the application of Theorem 5.1 by (cid:3) that of Proposition 22.1. Remark 23.5. A more precise statement, like the surjectivity statement in Theorem 23.2, holds at each step in the spectral sequence appearing in the proof. We refrain from stating this, as it will not be needed in the sequel. Elementary approach to vanishing for I1. Combining Lemma 3.4 and local vanishing for ωY (E), we see that I1(D) sits in an exact sequence 0 → f∗Ωn−1 Y (log E) → ωX(D) ⊗ I0(D) ⊗ F1DX → ωX(2D) ⊗ I1(D) → → R1f∗Ωn−1 Y (log E) → 0. This can be seen as a lift of the quasi-ismorphism in Theorem 6.1 in the case k = 1. Since F1DX ≃ OX ⊕ TX , using Nadel vanishing it is then not too HODGE IDEALS 65 hard to recover Theorem 23.2 from Theorem 32.2 in the Appendix, when D is ample, more precisely from the vanishing H 2(cid:0)Y, Ωn−1 Y (log E) ⊗ f∗L(cid:1) = 0, without using the vanishing theorem for Hodge modules. Although we ex- pect this to be possible eventually, at the moment we do not know how to do a similar thing for Ik(D) with k ≥ 2, i.e. recover the full Theorem F. 24. Effective version. The main difficulty in applying Theorem F is the Nakano-type vanishing requirement. We will see in the next section that this difficulty does not occur in important examples, like toric or abelian varieties. Here we explain how an effective measure of the positivity of the tangent bundle of X allows one to get rid of this requirement at the expense of working with a sufficiently positive divisor. For simplicity we assume here that D is ample; a similar statement holds in general by tensoring with an appropriate ample line bundle L. Corollary 24.1. Let X be a smooth projective variety of dimension n, and D a reduced effective (k − 1)-log-canonical ample divisor on X, with k ≥ 1. If A is an ample Cartier divisor such that TX (A) is nef, then H i(cid:0)X, ωX(cid:0)(k + 1)D(cid:1) ⊗ Ik(D)(cid:1) = 0 assuming that D − k(cid:0) − KX + (n + 1)A(cid:1) is ample. for all i > 0, Proof. According to Theorem F, we need to check that the condition (24.2) H j(cid:0)X, ωX ⊗ ∧jTX ⊗ OX ((k − j + 1)D)(cid:1) = 0 holds for all 1 ≤ j ≤ k. A special case of Demailly's extension of the Griffiths vanishing theorem (see [Laz04, Theorem 7.3.14] and the preceding comments) says that for every nef vector bundle E and ample line bundle M on X, and for every m ≥ 1, one has We apply this with E = TX (A), to obtain that H i(X, ωX ⊗ ∧mE ⊗ (det E)⊗m ⊗ M ) = 0 for all H i(cid:0)X, ωX ⊗ ∧jTX ⊗ ω−j X (j(n + 1)A) ⊗ M(cid:1) = 0 for all i > 0. A small calculation shows that in order to satisfy (24.2) it is therefore enough (cid:3) to have the ampleness of D − k(cid:0) − KX + (n + 1)A(cid:1). Remark 24.3. Following [ELN96, Remark 4.5], inspired in turn by [Dem93, Corollary 12.12], an effective (but very large) bound can also be given de- pending on a line bundle A such that −KX + A is nef. i > 0. 66 M. MUSTAT¸ A AND M. POPA H. Vanishing on Pn and abelian varieties, with applications We revisit the vanishing theorems of the previous section for projective space and abelian varieties. In these cases the extra assumptions needed in Theorem F are automatically satisfied, due to special properties of the bundles of holomorphic forms, and this in turn has striking applications. A stronger result holds on toric varieties as well. 25. Vanishing on Pn and toric varieties. Theorem F takes a nice form on toric varieties, due to the fact that the extra condition on bundles of holomorphic forms is automatically satisfied by the Bott-Danilov-Steenbrink vanishing theorem; this says that for any ample line bundle A on a smooth projective toric variety X one has H i(X, Ωj X ⊗ A) = 0 for all j ≥ 0 and i > 0. Corollary 25.1. Let D be a reduced effective (k − 1)-log-canonical divisor on a smooth projective toric variety X, and let L be a line bundle on X such that L(pD) is ample for all 0 ≤ p ≤ k. Then H i(cid:0)X, ωX((k + 1)D) ⊗ L ⊗ Ik(D)(cid:1) = 0 for all i > 0. If D is ample, the same holds with L = OX. On Pn however the situation is even better, since one can eliminate the log-canonicity assumption as well. This is due to the existence, for any j ≥ 1, of the Koszul resolution (25.2) 0 →M OPn(−n−1) →M OPn(−n) → ··· →M OPn(−j−1) → Ωj Pn → 0. Theorem 25.3. Let D be a reduced hypersurface of degree d in Pn. ℓ ≥ (k + 1)d − n − 1, then If H i(cid:0)Pn, OPn (ℓ) ⊗ Ik(D)(cid:1) = 0 for all i ≥ 1. Proof. If D is (k−1)-log canonical, then this is a special example of Corollary 25.1. To see that this condition is not needed, we have to return to the proof of Theorem F, and see what happens if we do not assume the triviality of the Hodge ideals up to Ik−1(D). First, for each k ≥ 1 we have a short exact sequence 0 −→ ωPn ⊗ OPn (kd) ⊗ Ik−1(D) −→ ωPn ⊗ OPn((k + 1)d) ⊗ Ik(D) −→ −→ ωPn ⊗ grF k OPn(∗D) → 0. We can then proceed by induction: after twisting by any L = OPn(a) with a ≥ 0, assuming the vanishing in the statement for the term on the left, if we also have it for the term on the right, we obtain it for the term in the middle. The process can indeed be started, since for k = 1 vanishing for the term on the left is simply Nadel vanishing. HODGE IDEALS 67 We therefore need to prove vanishing of the type H i(cid:0)Pn, OPn(ℓ) ⊗ grF k OPn(∗D)(cid:1) = 0 for i > 0 and ℓ as in the statement. Note first that Saito vanishing applies in the exact same way as in Theorem F, in the form of vanishing for Hj(Pn, C•) for j ≥ k + 1. Without any assumptions on the Hodge ideals however, the complex C• now looks as follows: (cid:2)Ωn−k Pn ⊗ OPn(d) ⊗ I0(D) −→ Ωn−k+1 ··· −→ Ωn−1 ⊗ grF OPn(∗D) −→ ωPn ⊗ grF Pn ⊗ grF k−1 Pn k 1 OPn(∗D) −→ ··· OX(∗D)(cid:3) ⊗ OPn(a) concentrated in degrees 0 up to k. We use the spectral sequence in the proof of Theorem F, and recall that we are interested in the vanishing of the terms 1 with i ≥ 1. Again, this term is isomorphic to Ek,i Ek,i if we have vanishing for the term Ek−1,i 2 We claim that, using the inductive hypothesis, both extremal terms are equal to 0, which gives what we want. For this we use the short exact sequence 1 1 , which by definition sits in an exact sequence −→ Pn ⊗ OPn(kd + a) ⊗ Ik−1(D)(cid:1) −→ Ek−1,i Pn ⊗ OPn((k − 1)d + a) ⊗ Ik−2(D)(cid:1). H i(cid:0)Pn, Ωn−1 −→ H i+1(cid:0)Pn, Ωn−1 0 →M OPn(−n − 1) →M OPn (−n) → Ωn−1 H i(cid:0)Pn, OPn (ℓ) ⊗ Ik−1(D)(cid:1) = 0 for all i ≥ 1, H i(cid:0)Pn, OPn (ℓ) ⊗ Ik−2(D)(cid:1) = 0 for all i ≥ 1, Pn → 0 given by the Koszul complex. It shows that to have the vanishing of the term on the left, it is enough to have with ℓ ≥ kd− n− 1, i.e. exactly the inductive hypothesis for k− 1. Similarly, it shows that for the vanishing of the term on the right, it is enough to have with ℓ ≥ (k − 1)d − n − 1, i.e. the inductive hypothesis for k − 2. Now one needs to analyze the terms Ek,i r with r ≥ 2. Just as in the proof of Theorem F, to show that they are all isomorphic to each other, which leads to the statement of the theorem, it is enough to show that Ek−r,i+r−1 = 0 for all such r. When r > k this is clear, while when r = k we have 1 E0,i+k−1 1 = H i+k−1(cid:0)X, Ωn−k X ⊗ OPn(d + a) ⊗ I0(D)(cid:1). We again use the Koszul complex (25.2) for j = n − k. This gives by a simple calculation that it suffices to have and all ℓ ≥ d − n − 1, which is Nadel vanishing. When r < k, just as before we have an exact sequence H i(cid:0)Pn, OPn(ℓ) ⊗ I0(D)(cid:1) = 0 for all i > 0 Pn ⊗ OPn ((k − r + 1)d + a) ⊗ Ik−r(D)(cid:1) −→ Ek−r,i−r+1 Pn ⊗ OPn((k − r)d + a) ⊗ Ik−r−1(D)(cid:1). −→ H i+r(cid:0)Pn, Ωn−r H i+r−1(cid:0)Pn, Ωn−r −→ 1 68 M. MUSTAT¸ A AND M. POPA A completely similar use of the Koszul complex (25.2), this time for j = n−r, together with the inductive hypothesis, implies that Ek−r,i−r+1 = 0. (cid:3) 1 26. Bounds for the subschemes associated to Hodge ideals in Pn. We use Theorem 25.3 to give a numerical criterion for the triviality of the ideals Ik(D) when D is a hypersurface in projective space, or to impose restrictions on the corresponding subschemes Zk. To put things in context, recall that the log-canonical threshold of a hypersurface D of degree d with isolated singularities in Pn satisfies lct(D) ≥ min n n d , 1o (see, for example, [dFEM03, Corollary 3.6]). Consequently I0(D) = OX, i.e. the pair (X, D) is log-canonical, when n + 1 > d. More generally, for an arbitrary hypersurface D ⊂ Pn, a standard application of Nadel vanishing implies that if (X, D) is not log-canonical, then dim Sing(D) ≥ n − d + 1. Theorem G in the introduction generalizes this to Ik(D) with k ≥ 1. It also extends a result of Deligne, see the remarks after [Sai93, Theorem 0.11], which gives a numerical criterion for the triviality of Ik(D) for hypersurfaces with isolated singularities. Proof of Theorem G. If Zk is non-empty, then by intersecting D with a gen- eral linear subspace L of Pn of dimension n − zk, we obtain a reduced hypersurface DL ⊂ L = Pn−zk such that subscheme associated to Ik(DL) is non-empty and 0-dimensional. Indeed, by the generic restriction theorem for Hodge ideals, Theorem 16.1, we have that Ik(DL) = Ik(D) · OL. Denoting by B the line bundle ωL ⊗ OL(cid:0)(k + 1)DL(cid:1), there is a short exact sequence 0 −→ B ⊗ Ik(DL) −→ B −→ B ⊗ O For (1), the condition zk < n − (k + 1)d + 1 is precisely equivalent to Z(cid:0)Ik(DL)(cid:1) −→ 0. H 0(L, B) = 0. On the other hand, we have H 1(cid:0)L, B ⊗ Ik(DL)(cid:1) = 0. as a special case of Theorem 25.3, and so by passing to cohomology in the exact sequence above we deduce that Z(cid:0)Ik(DL)(cid:1) = ∅, a contradiction. not the case. We then have that Z(cid:0)Ik(DL)(cid:1) is a 0-dimensional scheme of For (2), the statement is trivial if Zk = ∅, so we can assume that this is length deg Zk. The same argument shows that we have a surjection H 0(cid:0)L, B(cid:1) −→ B ⊗ O Z(cid:0)Ik(DL)(cid:1) −→ 0, and this time the space on the left has dimension(cid:0)(k+1)d−1 n−zk (cid:1). HODGE IDEALS 69 Part (3) follows from the fact that, due to the same vanishing theorem, if Z′k is the 0-dimensional part of Zk, then for every ℓ ≥ (k + 1)d − n − 1, there is a surjection H 0(cid:0)Pn, OPn (ℓ)(cid:1) → OZ ′ k . (cid:3) Remark 26.1. In the case of hypersurfaces in Pn whose singularities are isolated and non-degenerate with respect to the corresponding Newton poly- hedra, Saito's result [Sai93, Theorem 0.11] discussed in Remark 20.11 implies Deligne's theorem mentioned above. Note also that [DD90] looks at the re- lationship between the Hodge filtration and the pole order filtration on other homogeneous varieties. Remark 26.2 (Toric analogue). The first assertion in Theorem G admits a version in the toric context. Suppose that X is a smooth projective toric variety and D is an ample, reduced effective divisor, with isolated singular- ities. If k ≥ 0 is such that (26.3) H 0(cid:0)X, ωX(cid:0)(k + 1)D(cid:1)(cid:1) = 0, then (X, D) is k-log-canonical. Indeed, we argue that (X, D) is j-log- canonical by induction on j ≤ k. For the induction step, we use the fact that (X, D) is (j − 1)-log-canonical and Corollary 25.1 to conclude that H 1(cid:0)X, ωX(cid:0)(j + 1)D(cid:1) ⊗ Ij(D)(cid:1) = 0 and then argue as in the proof of Theorem G. Note that every divisor D on a toric variety is linearly equivalent to a torus-invariant divisor G. A pair (X, G), with X a variety as above and G an ample torus-invariant divisor corresponds to a lattice polytope P , and condition (26.3) is equivalent with the fact that the interior of (k + 1)P does not contain any lattice points (see [Ful93, p. 90]). We give two examples when one can apply this toric criterion for k-log- canonicity. 1) Suppose that D is an effective divisor in Pn1×···×Pnr of multidegree If D has isolated singularities and (d1, . . . , dr), with all dj > 0. (k + 1)di < ni + 1 for some i, then (X, D) is k-log canonical. 2) Suppose that P is a smooth Gorenstein polytope of index r (see [LN15] for a discussion of such polytopes). This means that if (X, B) is the corresponding pair, with X a toric variety and B an ample torus-invariant divisor, then X is smooth and −KX = rB. If D is an effective, reduced divisor with isolated singularities, linearly equivalent with dB for some d > 0, and if k is such that (k + 1)d < r, then (X, D) is k-log-canonical. 70 M. MUSTAT¸ A AND M. POPA 27. Singular points on hypersurfaces in Pn. We now exploit part (3) in Theorem G, i.e. the fact that the isolated points of Zk impose independent conditions on hypersurfaces of degree at least (k + 1)d − n − 1 in Pn, in conjunction with the nontriviality criteria in §19 and §21. We assume that n ≥ 3, when some singularities are naturally detected by appropriate Hodge ideals Ik(D) with k ≥ 1; the method applies in P2 as well, but in this case it is known that the type of results we are aiming for can already be obtained by considering the multiplier ideal I0(D) or the adjoint ideal adj(D). As motivation, recall that when X ⊂ P3 is a reduced surface of degree d whose only singularities are nodes, i.e ordinary double points, a classical result of Severi [Sev46] says that the set of nodes on X imposes independent conditions on hypersurfaces of degree at least 2d − 5 in P3. Park and Woo [PW06] showed that in fact this holds replacing the set of nodes by that of all singular points, and gave similar bounds for isolated singular points on hypersurfaces in arbitrary Pn. Using the ideals Ik(D) for suitable k, we obtain a new result on hypersurfaces in any dimension. Proof of Corollary H. We know from Corollary 21.3 (see also Example 21.4) that Ik(D) ⊆ IS, where k =(cid:2) n from Theorem G (3), which says that there is a surjection k → 0, m(cid:3). Since S is a set of isolated points, the result then follows H 0(cid:0)Pn, OPn ((k + 1)d − n − 1)(cid:1) → OZ ′ where Z′k is the 0-dimensional part of Zk. (cid:3) When n = 3 and m = 2 for instance, the bound is one worse than the Severi bound, but at least when n ≥ 5 and m ≥ 3, in many instances this improves what comes out of [PW06] or similar methods.3 Remark 27.1. Example 21.4 explains why with this method one can do at least as well with arbitrary isolated singularities as with ordinary ones. Note however that there exist situations where the bound in Corollary H can be if D has only nodes, in [DS12, Corollary 2.2] the same bound improved: 2 ] + 1)d − n − 1 is obtained when n is odd, but the better bound n ([ n 2 · d − n is shown to hold when n is even. See also [Dim13] for further interesting applications of such bounds. We conclude with a statement analogous to Corollary H for higher jets, using estimates for the order of vanishing of the scheme Zk at each point. 3Rob Lazarsfeld has shown us a different approach, based on multiplier ideals, showing that the isolated points of multiplicity m ≥ 2 impose independent conditions on hypersur- faces of degree at least m−1 (d − 1) − n; this is often stronger than the bound in [PW06]. Since d ≥ m, it is somewhat weaker than the bound in Corollary H when m ≤ n + 1. n HODGE IDEALS 71 Recall that for a 0-dimensional subset S ⊆ Pn, the space of hypersurfaces of degree ℓ is said to separate s-jets along S if the restriction map is surjective. Thus S imposes independent conditions on such hypersurfaces if they separate 0-jets along it. For the next statement, given m ≥ 3 and j ≥ 1, we denote OPn /m s+1 x H 0(cid:0)Pn, OPn(ℓ)(cid:1) →Mx∈S km,j := ⌈ n−m+j ⌈ n−m+j m ⌉ if j ≤ m − 1 m−2 ⌉ if j ≥ m. Corollary 27.2. Let D be a reduced hypersurface of degree d in Pn, with n ≥ 3. Let Sm be the set of isolated singular points of D of multiplicity at least m ≥ 3. Then hypersurfaces of degree at least (km,j + 1)d − n− 1 in Pn separate (j − 1)-jets along Sm, for each j ≥ 1. Proof. When j ≤ m−1 we use Theorem E, while otherwise we use Corollary 19.4, in order to deduce that for every x ∈ Sm we have Ikm,j (D) ⊆ m j x. Combining this with Theorem G (3), we obtain a surjection H 0(cid:0)Pn, OPn((km,j + 1)d − n − 1)(cid:1) → Mx∈Sm OPn/m j x. (cid:3) 28. Vanishing on abelian varieties. On abelian varieties we can obtain stronger vanishing statements than those in the previous sections. In this paragraph X will always be a complex abelian variety of dimension g, and D a reduced effective ample divisor on X. Lemma 28.1. We have H i(cid:0)X, DR(OX (∗D)) ⊗ Cρ(cid:1) = 0 for all i > 0, where Cρ denotes the rank one local system associated to any ρ ∈ Char(X). Proof. Denote as always by j : U ֒→ X the inclusion, where U = X r D. If we denote by P the perverse sheaf DR(OX (∗D)), then P ≃ j∗j∗P. The projection formula then gives H i(cid:0)X, DR(OX (∗D)) ⊗ Cρ(cid:1) ≃ H i(cid:0)U, j∗(P ⊗ Cρ)(cid:1) = 0 for all i > 0, where the last equality follows by Artin vanishing (see e.g. [Dim04, Corollary 5.2.18]) since U is affine. (cid:3) Theorem 28.2. For all k ≥ 0 the following are true, and equivalent: 72 M. MUSTAT¸ A AND M. POPA Pic0(X). (1) H i(cid:0)X, OX ((k + 1)D) ⊗ Ik(D) ⊗ α(cid:1) = 0 for all i > 0 and all α ∈ (2) H i(cid:0)X, grF OX(∗D) ⊗ α(cid:1) = 0 for all i > 0 and all α ∈ Pic0(X).4 Proof. We proceed by induction on k. In the case k = 0, the equivalence is obvious. On the other hand, the result is true by Nadel vanishing, Proposi- tion 23.1. k Now for any k we have a short exact sequence 0 −→ OX (kD) ⊗ Ik−1(D) ⊗ α −→ OX ((k + 1)D) ⊗ Ik(D) ⊗ α −→ −→ grF k OX (∗D) ⊗ α −→ 0. Assuming that the result holds for k − 1, passing to cohomology gives the equivalence between (1) and (2) for k. We denote by Cα the unitary rank one local system associated to α. precisely as in the proof of Proposition 22.1 we obtain Considering the spectral sequence Ep,q 1 = Hp+q(cid:0)X, grF for all i > 0, all ℓ, and all α ∈ Pic0(X), by virtue of Lemma 28.1. We use this with ℓ = −g + k. More precisely, we look at the complex −q DR(OX(∗D))⊗α(cid:1) =⇒ H p+q(cid:0)X, DR(OX (∗D))⊗Cα(cid:1), Hi(cid:0)X, grF C• :=(cid:0) grF OX(∗D) ⊗ α →M grF ℓ DR(OX (∗D)) ⊗ α(cid:1) = 0 −g+k DR(OX (∗D)) ⊗ α(cid:1)[−k]. OX(∗D) ⊗ α → ··· → grF X is trivial, this can be identified with a complex of the form OX (∗D) ⊗ α(cid:3) k 1 concentrated in degrees 0 up to k. The vanishing above says that Given that Ω1 (cid:2)M grF 0 We use the spectral sequence Hj(X, C•) = 0 for all j ≥ k + 1. Ep,q 1 = H q(X, C p) =⇒ Hp+q(X, C•). = 0 since C k+1 = 0, while Ek−1,i = 0 by the 1 Note that for i ≥ 1, Ek+1,i inductive hypothesis. It follows that Ek,i 1 ≃ Ek,i 2 . 1 Continuing this way, for any r ≥ 2, we have that for Ek,i the outgoing term is 0 because of the length of the complex, while the incoming term is 0 by induction. The conclusion is that r Ek,i 1 ≃ Ek,i ∞ . 4In Fourier-Mukai language this theorem says that OX(cid:0)(k + 1)D(cid:1) ⊗ Ik(D), or equiva- lently grF k OX (∗D), satisfies IT0, i.e. the Index Theorem with index 0. HODGE IDEALS Since H k+i(X, C•) = 0, we obtain that Ek,i 1 = H i(cid:0)X, grF k OX(∗D) ⊗ α(cid:1) = 0. 73 (cid:3) Remark 28.3. A similar inductive argument as in Theorem 28.2 shows that when D is arbitrary and L is an ample line bundle, one has and that this is equivalent to the statement H i(cid:0)X, OX(cid:0)(k + 1)D(cid:1) ⊗ L ⊗ Ik(D)(cid:1) = 0, H i(cid:0)X, grF k OX (∗D) ⊗ L(cid:1) = 0, both for all i > 0 and all k. However this last statement is already a special case of [PS13, §2.3, Lemma 1]. 29. Singularities of theta divisors. A well-known result of Koll´ar [Kol95, Theorem 17.3], revisited by Ein-Lazarsfeld [EL97], states that if (A, Θ) is a principally polarized abelian variety (ppav) of dimension g, then the pair (A, Θ) is log-canonical, and in particular multx(Θ) ≤ g for any x ∈ Θ. By a result of Smith-Varley, it is also known that when the multiplicity is equal to g, the ppav must be reducible; for this and related results, see [EL97] and the references therein. For irreducible ppav's it is believed however that the situation should be substantially better. One has the following folklore: Conjecture 29.1. Let (A, Θ) be an irreducible ppav of dimension g. Then multx(Θ) ≤ g + 1 2 , for all x ∈ Θ. The conjecture is known in dimension up to five, and more generally for Prym varieties associated to double covers of irreducible stable curves; see [CM08, Theorem 3]. Here we make a first step towards the general result, by proving the conjecture for theta divisors with isolated singularities. Note that the main tool in [EL97] is the triviality of the multiplier ideal I0(Θ). We rely in turn on our study of the Hodge ideal I1(Θ), though not via its triviality, which in general does not hold. Note that better results hold for g ≫ 0; see Remark 29.6(1). Theorem 29.2. Let (X, Θ) be an irreducible ppav of dimension g, such that Θ has isolated singularities. Then: (1) For every x ∈ Θ we have multx(Θ) ≤ g+1 2 . (2) Moreover, there can be at most one x ∈ Θ such that multx(Θ) = g+1 2 . Proof. Note in passing that for g ≥ 3 irreducibility follows automatically Indeed, if (A, Θ) splits as from the assumption on isolated singularities. a product of ppav's, then we have dim Sing(Θ) = g − 2. For g = 2 it is necessary to assume it, as the bound fails for a product of elliptic curves. 74 M. MUSTAT¸ A AND M. POPA Assuming that multx(Θ) ≥ g+2 we obtain that 2 , by Theorem E (see also Example 21.2) I1(Θ) ⊆ m 2 x. Consider the short exact sequence 0 −→ OX(2Θ) ⊗ I1(Θ) −→ OX(2Θ) ⊗ m x/I1(Θ) −→ 0. For every α ∈ Pic0(X), by tensoring the exact sequence with α and using (1) in Theorem 28.2 and the fact that I1(Θ) has finite co-support, we obtain x −→ OX(2Θ) ⊗ m 2 2 H i(cid:0)X, OX (2Θ) ⊗ m 2 x ⊗ α(cid:1) = 0 for all i > 0. In particular, as OX (2Θ) is globally generated, the vanishing of H 1 implies that the linear system 2Θ separates tangent vectors at each point of X. To see this, note that the collection of line bundles OX(2Θ) ⊗ α is, as α varies in Pic0(X), the same as the collection of line bundles t∗a OX (2Θ) as a varies in X, where ta denotes translation by a. But this is a contradiction; indeed, it is well known that when Θ is irreducible this linear system provides a 2 : 1 map which is ramified at the 2-torsion points (and more precisely factors through the Kummer variety of X). This proves (1). For (2), assume that there are two distinct points x, y ∈ Θ having multi- 2 . According again to Example 21.2, it follows that plicity g+1 I1(Θ) ⊆ mx ⊗ my. Using the same argument as in (1), we obtain H i(cid:0)X, OX (2Θ) ⊗ mx ⊗ my ⊗ α(cid:1) = 0 for all i > 0, and therefore conclude that the linear system 2Θ separates all points of the form x − a and y − a with a ∈ A. Note however that the equation x − a = a − y does have solutions, which contradicts the fact that 2Θ does not separate nonzero points of the form z and −z (both mapping to the same point on the Kummer variety). (cid:3) Remark 29.3. (1) Mumford [Mum83] showed, developing ideas of Andreotti- Mayer, that the locus N0 of ppav's such that Sing(Θ) 6= ∅ is a divisor in the moduli space of ppav's, and moreover that for the general point in every irreducible component of N0, Θ has isolated singularities. Thus Theorem 29.2 applies on a dense open set of each component of N0. These open sets are in fact large: Ciliberto-van der Geer [CvdG08] have shown that their complements have codimension at least two in N0. (2) Equality in Conjecture 29.1 is known to be achieved for certain points on Jacobians of hyperelliptic curves and on the intermediate Jacobian of the cubic threefold. The latter example also shows optimality in Theorem 29.2: the theta divisor on the intermediate Jacobian of a smooth cubic threefold HODGE IDEALS 75 (a ppav of dimension 5) has a unique singular point, the origin, which is of multiplicity 3. Note also that in this case we have I1(Θ)0 = m0 ( OX,0 according to Example 20.1, as the projectivized tangent cone to Θ at 0 is isomorphic to the original cubic threefold. A similar argument involving the higher ideals Ik(D) can be used to give bounds on multiplicity in terms of effective jet separation. In particular, it leads to an asymptotic bound in terms of the Seshadri constant. Given x ∈ X, we denote by s(ℓ, x) the largest integer s such that the linear system ℓΘ separates s-jets at x, i.e. such that the restriction map H 0(cid:0)X, OX (ℓΘ)(cid:1) −→ H 0(cid:0)X, OX (ℓΘ) ⊗ OX /m s+1 x (cid:1) is surjective. It is a fundamental property of the Seshadri constant of Θ at x that (29.4) s(ℓ, x) ℓ ≤ ǫ(Θ, x), and that in fact ǫ(Θ, x) is the limit of these quotients as ℓ → ∞; see [Laz04, Theorem 5.1.17] and its proof. Due to the homogeneity of X, ǫ(Θ, x) does not in fact depend on x, so we will denote it ǫ(Θ). We denote also sℓ := min {s(ℓ, x) x ∈ X}. Theorem 29.5. Let (X, Θ) be ppav of dimension g, such that Θ has isolated singularities. Then for every x ∈ Θ and every k ≥ 1 we have multx(Θ) < 2 + sk+1 k + 1 + g . k + 1 In particular, for every x ∈ Θ we have multx(Θ) ≤ ǫ(Θ) + 2 ≤ gpg! + 2. Hence, if g ≫ 0, then multx(Θ) is at most roughly g Proof. We assume that e + 2. multx(Θ) ≥ 2 + sk+1 k + 1 + g k + 1 and aim for a contradiction. Under this assumption, according to Corollary 19.4 it follows that Ik(D) ⊆ m 2+sk+1 x . An argument identical to that in Theorem 29.2 then shows that for all α ∈ Pic0(X) one has for all i > 0, H i(cid:0)X, OX(cid:0)(k + 1)Θ(cid:1) ⊗ m 2+sk+1 x ⊗ α(cid:1) = 0 and so the linear system (k + 1)Θ separates (1 + sk+1)-jets at all points of X, a contradiction. 76 M. MUSTAT¸ A AND M. POPA (cid:3) The statement multx(Θ) ≤ ǫ(Θ) + 2 now follows using (29.4) and let- ting k → ∞. On the other hand, the definition of the Seshadri constant automatically implies ǫ(Θ) ≤ g√g!; see [Laz04, Proposition 5.1.9]. The last assertion follows from the well-known fact that limg→∞ Remark 29.6. (1) The last assertion in Theorem 29.5 becomes better than that given by Theorem 29.2 for g very large. Grushevsky (together with Codogni and Sernesi [CGS16]), and independently Lazarsfeld, have commu- nicated to us that they can show a similar, but slightly stronger statement, using methods from intersection theory. In [CGS16] it is shown that if m is the multiplicity of an isolated point on Θ, then m(m − 1)g−1 ≤ g! − 4. g√g! g = 1 e . (2) If m = multx(Θ), the numerical bound m ≤ g√g! + 2 in the statement above can be deduced directly, at least asymptotically, from the surjectivity of the mapping H 0(cid:0)X, OX ((k + 1)Θ)(cid:1) −→ H 0(cid:0)X, OX ((k + 1)Θ) ⊗ OX /m for s = (k + 1)(m − 2) − g, by comparing the dimensions of the two spaces. Thus a completely similar argument shows that if m1, . . . , mr are the mul- tiplicities of all the singular points of Θ, then (cid:1), s+2 x rXi=1 (mi − 2)g ≤ g!. We thank Sam Grushevsky for this observation; the same holds in [CGS16], for all g, with the better bound above. (3) Theorem 29.5 is weaker for k = 1 than Theorem 29.2, due to the fact that asymptotically we need to use Corollary 19.4 instead of Theorem E. (4) Theorem 29.5 holds, with the same proof, for any ample divisor D with isolated singularities on an abelian variety, replacing g! with Dg. 30. Singular points on ample divisors on abelian varieties. We conclude by noting that the results in §27 have immediate analogues for iso- lated singular points on hypersurfaces in abelian varieties. We fix a complex abelian variety X of dimension g, and a reduced effective divisor D on X. We assume that g ≥ 3, since again the case of curves on abelian surfaces can always be treated by using multiplier or adjoint ideals. Corollary 30.1. Let Sm be the set of isolated singular points on D of mul- tiplicity at least m ≥ 2. Then Sm imposes independent conditions on pD, for p ≥(cid:2) g m(cid:3) + 1. Proof. The proof is identical to that of Corollary H. We use Theorem 28.2 for the ideals Ik(D), and the same Ik as in that corollary. (cid:3) HODGE IDEALS 77 Remark 30.2. A statement analogous to Corollary 27.2 can be formulated as well. Moreover, in the result above one can say more generally that the respective finite set imposes independent conditions on all linear systems pD′, where D′ is a divisor numerically equivalent to D. The reason is that Theorem 28.2 allows for twisting with arbitrary α ∈ Pic0(X). I. Appendix: Higher direct images of forms with log poles In the appendix we establish a few local vanishing statements for higher direct images of bundles of forms with log poles. In this paper they are needed for the birational study of Hodge ideals, but they are statements of general interest. 31. The case of SNC divisors on the base. We first compute direct images of bundles of forms with log poles via birational morphisms that dominate log-smooth pairs. Theorem 31.1. Let X be a smooth variety and D a reduced simple normal crossing divisor on X. Suppose that f : Y → X is a proper, birational transform of D and F is the reduced exceptional divisor. We assume that E has simple normal crossings. morphism, with Y smooth, and consider E = eD + F , where eD is the strict i) If f is an isomorphism over X r D (so that E = (f∗D)red), then f∗Ωp X(log D) Y (log E) ≃ Ωp Rif∗Ωp Y (log E) = 0 for all p ≥ 0, and for all i > 0, p ≥ 0. ii) For every f , we have f∗Ω1 Y (log E) ≃ Ω1 X(log D) and Rif∗Ω1 Y (log E) = 0 for all i > 0. The assertion in i) is contained in [EV82, Lemmas 1.2 and 1.5], where the vanishing statement is deduced from a theorem of Deligne.5 We give a self-contained proof below, since it also applies in case ii), which is needed in the paper. First, one useful corollary of the theorem is the following: Corollary 31.2. Let X be a smooth variety and D an effective Cartier divi- sor on X. If f : Y → X is a log resolution of the pair (X, D) which is an iso- morphism over X r D, and E = (f∗D)red, then the sheaves Rif∗Ωp Y (log E) are independent of the choice of log resolution for all i and p. 5We thank H. Esnault for pointing this out. 78 M. MUSTAT¸ A AND M. POPA Proof. We consider a log resolution as in the statement, and to keep track of it we use the notation EY instead of E. Take another log resolution g : Z → X, with the divisor EZ . They can both be dominated by a third log resolution h : W → X, with the corresponding divisor EW . We denote by ϕ : W → Y the induced morphism. Note that EW = (ϕ∗EY )red. We deduce from Theorem 31.1 i) and the Leray spectral sequence that Rif∗Ωp Y (log EY ) ≃ Rih∗Ωp By symmetry, we also have Rig∗Ωp obtain the assertion in the corollary. W (log EW ). Z(log EZ ) ≃ Rih∗Ωp W (log EW ), and we (cid:3) Going back to the statement of Theorem 31.1, it is easy to see that we Y (log E), inducing in turn a have a canonical morphism f∗Ωp canonical morphism X(log D) → Ωp X (log D) −→ f∗Ωp Ωp Y (log E) which is generically an isomorphism. The first assertions in each of the two statements in the theorem say that this map is an isomorphism when p = 1, in case ii), and for all p, in case i). We first prove the theorem in a special case. Proposition 31.3. The assertions in Theorem 31.1 hold for the blow-up f of X along a smooth subvariety W of X having simple normal crossings with D. Recall that if X is smooth and Z1, . . . , Zr are subschemes of X, we say that Z1, . . . , Zr have simple normal crossings if locally on X there are algebraic coordinates x1, . . . , xn such that the ideal of each Zj is generated by a subset of {x1, . . . , xn}. We say that a divisor D and W have simple normal crossings if W and the components of D have simple normal crossings. Proof. We argue by induction on dim X, the assertion being trivial if this is equal to 1, when f is an isomorphism and E = D. Note first that if T is a prime divisor on X containing W , such that D′ = D + T is a reduced divisor having simple normal crossings with W, then if the proposition holds for (X, D′), it also holds for (X, D). Here is the only place where we have to distinguish between cases i) and ii). Suppose first that we are in case i), when by assumption W ⊆ Supp(D). We have E = (f∗D)red and let E′ = (f∗D′)red. Therefore E′ = E +eT , where eT is the strict transform of T . Note that if g : eT → T is the induced map, then g is the blow-up of T along W ⊆ DT and E eT = (g∗DT )red, hence we may apply the induction hypothesis for g and E eT . We have an exact sequence 0 −→ Ωp Y (log E) −→ Ωp Y (log E′) −→ Ωp−1 eT (log E eT ) −→ 0. HODGE IDEALS 79 Since we are assuming that the proposition holds for D′, and we also know that it holds for (T, DT ), we conclude using the long exact sequence in cohomology that Y (log E) = 0 for i ≥ 2, Rif∗Ωp and we have an exact sequence 0 → f∗Ωp Moreover, γ can be identified to Ωp Y (log E) → f∗Ωp Y (log E′) γ (log E eT ) → R1f∗Ωp Y (log E) → 0. eT → f∗Ωp−1 X(log D′) → Ωp−1 T (log DT ), hence ker(γ) ≃ Ωp X(log D) and Coker(γ) = 0. This shows that the proposition holds for (X, D). Suppose now that we are in the setting of ii). We have E = eD + F and let E′ = eD +eT + F . It is not true in general that E eT is the required divisor for the pair (T, DT ) and the morphism eT → T (trouble occurs precisely when codim(W, X) = 2, in which case eT → T is an isomorphism, with E eT corresponding to DT + W ). However, this issue does not come up when p = 1, when we only need to consider the exact sequence 0 −→ Ω1 Y (log E) −→ Ω1 We obtain an induced exact sequence Y (log E′) 0 −→ f∗Ω1 and isomorphisms Y (log E) −→ f∗Ω1 Y (log E′) −→ O eT −→ 0. ϕ −→ f∗O eT −→ R1f∗ΩY (log E) −→ 0 Y (log E′) Since the morphism ϕ can be identified with Y (log E) ≃ Rif∗Ω1 Rif∗Ω1 for all i ≥ 2. Ω1 X(log D′) −→ OT , and D′ satisfies the conclusion of Proposition 31.3, it follows that D satisfies it, too. This completes the proof of the fact that if the proposition holds for D′, then it also holds for D. From now on, the proof proceeds in the same way in both cases i) and ii). Since the assertion to be proved is local on X, we may assume that we have algebraic coordinates x1, . . . , xn on X such that W is defined by (x1, . . . , xr) and each irreducible component of D is defined by some (xi). Let I ⊆ {1, . . . , r} consist of those i such that the divisor Di defined by (xi) is not contained in D. Let D′ = D +Xi∈I Di. Applying repeatedly the observation at the beginning of the proof, we see that in order to show that (X, D) satisfies the proposition, it is enough to show that the same holds for (X, D′). It is easy to see by a local calculation in the coordinates x1, . . . , xn that f∗Ω1 Y (log E′), hence X(log D′) ≃ Ω1 f∗Ωp X(log D′) ≃ Ωp Y (log E′) for all p ≥ 0. 80 M. MUSTAT¸ A AND M. POPA The fact that (X, D′) satisfies the proposition is now an immediate conse- quence of the projection formula, combined with the fact that Rf∗OY ≃ OX. This completes the proof of the proposition. (cid:3) Proof of Theorem 31.1. The same argument applies in both cases i) and ii). By considering a log resolution of the ideal on X defining the indeterminacy locus of f−1, we obtain a morphism h : Z → X with the following properties: a) The birational map g = f−1 ◦ h is a morphism. b) h is a composition of smooth blow-ups, with each blow-up center having simple normal crossings with the sum of the strict transform of D and the exceptional divisor over X. Moreover, if we are in case i), then the blow-up center is contained in the inverse image of D. In particular, it follows from b) that Z is smooth. Note also that if G is the sum of the strict transform of D on Z with the h-exceptional divisor, then G has simple normal crossings and it is equal to the sum of the strict transform of E on Z with the g-exceptional divisor. In particular, whatever we prove for f and D, it will also apply to g and E. In order to fix ideas, suppose first that we are in the setting of i). By applying Proposition 31.3 to each of the blow-ups whose composition is h, we deduce that for every p, we have h∗Ωp Z (log G) ≃ Ωp X(log D), and Rih∗Ωp We first deduce that the canonical morphism j : Ωp Z(log G) = 0 for i ≥ 1. X(log D) → f∗Ωp Y (log E) is an isomorphism. Indeed, we know that the composition Ωp X(log D) j −→ f∗Ωp Y (log E) −→ h∗Ωp Z (log G) is an isomorphism, and therefore j is a split monomorphism. Since it is generically an isomorphism and f∗Ωp Y (log E) is torsion-free, we conclude that it is an isomorphism. By applying this to g as well, we conclude that We now consider the Leray spectral sequence Ekℓ Y (log E). g∗Ωp Z(log G) ≃ Ωp 2 = Rkf∗(cid:0)Rℓg∗Ωp Z(log G)(cid:1) ⇒ Rk+ℓh∗Ωp Rif∗Ωp Y (log E) = 0 Since we know that h satisfies the conclusion of the theorem, we deduce that Ekℓ ∞ = 0 for every (k, ℓ) 6= (0, 0). We prove by induction on i ≥ 1 that Z(log G). for every f as above; in particular, we will be able to use the inductive assumption for g as well. Now given r ≥ 2, we can identify Ei,0 r+1 to the cohomology of Since Ei−r,r−1 r Ei−r,r−1 r is a subquotient of Ei−r,r−1 → Ei,0 r → Ei+r,1−r r = 0. 2 Z (log G) = 0 Rr−1g∗Ωp , this is 0 for r ≤ i since HODGE IDEALS 81 by the inductive assumption. On the other hand, we have Ei−r,r−1 = 0 for r > i since this is a first quadrant spectral sequence. Therefore, recalling that i ≥ 1, we obtain Ei,0 ∞ = 0. Since 2 = Ei,0 r Ei,0 2 = Rif∗Ωp Y (log E), this completes the induction step and with this the proof of case i) in the theorem. The proof of case ii) follows verbatim for p = 1. (cid:3) Remark 31.4. The assertion in Theorem 31.1 ii) can fail (even when D = 0) for p > 1. For example, suppose that X = A2 and f : Y → X is the blow-up of the origin, with exceptional divisor F . It is easy to see that in this case R1f∗ωY (F ) ≃ C. 32. Akizuki-Nakano-type vanishing theorems. The goal in this sec- tion is to prove the following vanishing statement for higher direct images of sheaves of differentials with log poles. Under a slightly more restrictive hypothesis, this was obtained by Saito in [Sai07] using the theory of mixed Hodge modules; here we provide a proof based on more elementary meth- ods.6 Theorem 32.1. Let X be a variety and D an effective Cartier divisor on X such that X r D is smooth. If f : Y → X is a log resolution of (X, D) which is an isomorphism over X r D and E = (f∗D)red, then Rpf∗Ωq Y (log E) = 0 if p + q > n = dim X. We will deduce Theorem 32.1 from the following global result, closely related both in statement and proof to the Akizuki-Nakano vanishing theo- rem. Theorem 32.2. Let Y be a smooth, n-dimensional complete variety. If E is a reduced SNC divisor on Y such that Y r E is affine, then for every semiample line bundle L on Y we have H p(cid:0)Y, Ωq Y (log E) ⊗ L(cid:1) = 0 for p + q > n. Proof. We argue by induction on n. Note first that the assertion is clear on curves. It is also standard in general when L = OY . Indeed, recall that the Hodge-to-de Rham spectral sequence degenerates at the E1 term, hence Ekℓ 1 = H ℓ(cid:0)Y, Ωk(log E)(cid:1) ⇒ Hk+ℓ(Y, Ω•Y (log E)) Xp+q=m hp(cid:0)Y, Ωq(log E)(cid:1) = dimC Hm(cid:0)Y, Ω•Y (log E)(cid:1). 6Since this paper was written, Saito [Sai16b] has noted however that an even stronger statement than Theorem 32.1 can be obtained using mixed Hodge module theory: besides allowing X to be singular, over X r D one may assume only that the morphism f is semismall. 82 M. MUSTAT¸ A AND M. POPA On the other hand, we have and this is 0 for m > n since Y r E is an n-dimensional affine variety. This gives Hm(cid:0)Y, Ω•Y (log E)(cid:1) ≃ H m(Y r E; C) Y (log E)(cid:1) = 0 for p + q > n. H p(cid:0)Y, Ωq Since L is semiample, for some m ≥ 1 we may choose B ∈ L⊗m general such that E + B is reduced, with simple normal crossings. We consider π : Y ′ → Y the m-fold cyclic cover of Y branched along B. Denoting L′ = π∗L, there exists a divisor B′ ∈ L′ mapping isomorphically onto B, such that π∗B = mB′. Moreover, Y ′ is smooth and π∗E + B′ is an SNC divisor as well; see e.g. [Laz04, Proposition 4.1.6 and Remark 4.1.8]. Since Y r E is affine, it follows that Y r (E + B) is affine, and since π is finite, we conclude that Y ′ r (π∗E + B′) is affine as well. As we have seen at the beginning, this implies that (32.3) On the other hand, we have H p(cid:0)Y ′, Ωq Y ′(cid:0) log(π∗E + B′)(cid:1)(cid:1) = 0 for p + q > n. Y ′(cid:0) log(π∗E + B′)(cid:1) ≃ π∗Ωq Y(cid:0) log(E + B)(cid:1), and therefore Y ′(cid:0) log(π∗E + B′)(cid:1) ≃ Ωq π∗Ωq Y(cid:0) log(E + B)(cid:1) ⊗ π∗OY ′ ≃ Ωq We thus conclude from (32.3) that (32.4) ≃ Ωq Y(cid:0) log(E + B)(cid:1) ⊗ L⊗(−j). m−1Mj=0 Y(cid:0) log(E + B)(cid:1) ⊗ L⊗(1−m)(cid:1) = 0 for p + q > n. H p(cid:0)Y, Ωq Y(cid:0) log(E + B)(cid:1) ⊗OY H p(cid:0)Y, Ωq If q = 0, then p > n and the assertion we need to prove is trivial. Suppose now that q > 0 and consider the short exact sequence on Y Y (log E) → Ωq OY (−B) → Ωq 0 → Ωq B(log EB) → 0. By tensoring with L and taking the long exact sequence in cohomology, we obtain an exact sequence Y(cid:0) log(E + B)(cid:1) ⊗ L⊗(1−m)(cid:1) → H p(cid:0)Y, Ωq Y (log E) ⊗ L(cid:1) B(log EB) ⊗ LB). → H p(cid:0)Y, Ωq If p + q > n, then the first term vanishes by (32.4), while the third term vanishes by the inductive assumption. We thus obtain the vanishing of the second term, which completes the proof of the theorem. (cid:3) We can now prove the relative vanishing statement. HODGE IDEALS 83 Proof of Theorem 32.1. The assertion is local on X, hence we may also as- sume that X is affine. Since D is a Cartier divisor on X, it follows that X r D is affine as well. We choose an open embedding X ֒→ X, with X projective, smooth, and such that X r X is a (reduced) SNC divisor. If D is the closure of D in X, then we denote D′ := D + (X r X). We can also find an open embedding Y ֒→ Y such that we have a morphism g : Y → X which is identified with f over X r D. We may further assume that Y is smooth and if E is the closure of E in Y , then E′ := E + (Y r Y ) is an SNC divisor. Note that E′ = (g∗D′)red. It is of course enough to show that (32.5) Rpg∗Ωq Y (log E′) = 0 for p + q > n. Let L be an ample line bundle on X. A standard argument using the Leray spectral sequence for g and Ωq (log E′)⊗ g∗Lj shows that (32.5) holds if and Y only if (32.6) Since Y r E′ = Y r E ≃ X r D is affine, the vanishing in (32.6) follows from Theorem 32.2. This completes the proof. (log E′) ⊗ g∗Lj(cid:1) = 0 for p + q > n and j ≫ 0. H p(cid:0)Y , Ωq Y (cid:3) References [BW08] P. Bruckmann and P. Winkert, T -symmetrical tensor differential forms with logarithmic poles along a hypersurface section, Int. J. Pure Appl. Math. 46 (2008), no. 1, 111–136. [CM08] S. Casalaina-Martin, Cubic threefolds and abelian varieties of dimension five II, Math. Z. 260 (2008), no. 1, 115–125. [CvdG08] C. Ciliberto and G. van der Geer, Andreotti-Mayer loci and the Schottky prob- lem, Doc. Math. 13 (2008), 453–504. [CGS16] G. Codogni, S. Grushevsky, and E. Sernesi, The degree of the Gauss map of the theta divisor, preprint (2016). [dFEM03] T. de Fernex, L. Ein, and M. Mustat¸a, Bounds for log canonical thresholds with applications to birational rigidity, Math. Res. Lett. 10 (2003), no. 2–3, 219–236. [Del70] P. Deligne, Equations diff´erentielles `a points singuliers r´eguliers, Lect. Notes in Math., vol. 163, Springer, Berlin, 1970. [Dem93] J.-P. Demailly, A numerical criterion for very ample line bundles, J. Diff. Geom. 37 (1993), no. 2, 323–374. [DD90] P. Deligne and A. Dimca, Filtrations de Hodge et par l'ordre du pole pour les hypersurfaces singuli`eres, Ann. Sci. ´Ecole Norm. Sup. 23 (1990), no. 4, 645–656. [Dim04] A. Dimca, Sheaves in topology, Universitext, Springer-Verlag, Berlin, 2004. [Dim13] A. Dimca, On the syzygies and Hodge theory of nodal hypersurfaces, preprint arXiv:1310.5344 (2013). [DSW09] A. Dimca, M. Saito, and L. Wotzlaw, A generalization of Griffiths' theorem on rational integrals, II, Michigan Math. J. 58 (2009), 603–625. 84 M. MUSTAT¸ A AND M. POPA [DS12] A. Dimca and G. Sticlaru, On the syzygies and Alexander polynomial of nodal hypersurfaces, Math. Nachr. 285 (2012), 2120–2128. [DS15] A. Dimca and G. Sticlaru, Koszul complexes and pole order filtrations, Proc. Edinburgh Math. Soc. 58 (2015), 333-354. [EL97] L. Ein and R. Lazarsfeld, Singularities of theta divisors and the birational ge- ometry of irregular varieties, J. Amer. Math. Soc. 10 (1997), no. 1, 243–258. [ELN96] L. Ein, R. Lazarsfeld, and M. Nakamaye, Zero-estimates, intersection the- ory, and a theorem of Demailly, Higher-dimensional complex varieties (Trento, 1994), de Gruyter, Berlin, 1996. [EV82] H. Esnault and E. Viehweg, Revetements cycliques, Algebraic threefolds (Varenna, 1981), Lecture Notes in Math., vol. 947, Springer, Berlin-New York, 1982, pp. 241–250. [Ful93] W. Fulton, Introduction to toric varieties, Annals of Mathematics Studies, vol. 131, Princeton University Press, Princeton, NJ, 1993. The William H. Roever Lectures in Geometry. [Gro66] A. Grothendieck, ´El´ements de g´eom´etrie alg´ebrique. IV. ´Etude locale des sch´emas et des morphismes de sch´emas. III, Inst. Hautes ´Etudes Sci. Publ. Math. 28 (1966), 255 pp. [HTT08] R. Hotta, K. Takeuchi, and T. Tanisaki, D-modules, perverse sheaves, and representation theory, Birkhauser, Boston, 2008. [Kol95] J. Koll´ar, Shafarevich maps and automorphic forms, Princeton Univ. Press, 1995. [Kol97] J. Koll´ar, Singularities of pairs, Algebraic geometry–Santa Cruz 1995, Proc. Sympos. Pure Math., vol. 62, Amer. Math. Soc., Providence, RI, 1997, pp. 221– 287. [KM98] J. Koll´ar and S. Mori, Birational geometry of algebraic varieties, Cambridge Tracts in Mathematics, vol. 134, Cambridge University Press, Cambridge, 1998. With the collaboration of C. H. Clemens and A. Corti; Translated from the 1998 Japanese original. [Laz04] R. Lazarsfeld, Positivity in algebraic geometry II, Ergebnisse der Mathematik und ihrer Grenzgebiete, vol. 49, Springer-Verlag, Berlin, 2004. [LN15] B. Lorenz and B. Nill, On smooth Gorenstein polytopes, Tohoku Math. J. (2) 67 (2015), no. 4, 513–530. [Mum83] D. Mumford, On the Kodaira dimension of the Siegel modular variety, Alge- braic geometry–open problems, Ravello 1982, Lecture Notes in Math., vol. 997, Springer, Berlin, 1983, pp. 348–375. [MP16] M. Mustat¸a and M. Popa, Restriction, subadditivity, and semicontinuity theo- rems for Hodge ideals, preprint arXiv:1606.05659, to appear in Int. Math. Res. Not. (2016). [OSS80] C. Okonek, M. Schneider, and H. Spindler, Vector bundles on complex projective spaces, Progress in Mathematics, Birkhauser, Boston, 1980. [PW06] J. Park and Y. Woo, A remark on hypersurfaces with isolated singularities, Manuscripta Math. 121 (2006), 451–456. [Pop14] M. Popa, Kodaira-Saito vanishing and applications, preprint arXiv:1407.3294, to appear in L'Enseignement Math. (2014). [Pop16] M. Popa, Positivity for Hodge modules and geometric applications, preprint arXiv:1605.08093 (2016). [PS13] M. Popa and C. Schnell, Generic vanishing theory via mixed Hodge modules, Forum Math. Sigma 1 (2013), 60 pp. [Sai88] M. Saito, Modules de Hodge polarisables, Publ. Res. Inst. Math. Sci. 24 (1988), no. 6, 849–995. [Sai89] M. Saito, On the structure of Brieskorn lattice, Ann. Institut Fourier (Grenoble) 39 (1989), no. 1, 27–72. HODGE IDEALS 85 [Sai90] M. Saito, Mixed Hodge modules, Publ. Res. Inst. Math. Sci. 26 (1990), no. 2, 221–333. [Sai93] M. Saito, On b-function, spectrum and rational singularity, Math. Ann. 295 (1993), no. 1, 51–74. [Sai07] M. Saito, Direct image of logarithmic complexes and infinitesimal invariants of cycles, Algebraic cycles and motives. Vol. 2, London Math. Soc. Lecture Note Ser., vol. 344, Cambridge Univ. Press, Cambridge, 2007, pp. 304–318. [Sai09] M. Saito, On the Hodge filtration of Hodge modules, Mosc. Math. J. 9 (2009), no. 1, 161–191. [Sai16a] M. Saito, A young person's guide to mixed Hodge modules, preprint arXiv:1605.00435 (2016). [Sai16b] M. Saito, Hodge ideals and microlocal V -filtration, preprint arXiv:1612.08667 (2016). [Sch14a] C. Schnell, An overview of Morihiko Saito's theory of mixed Hodge modules, preprint arXiv:1405.3096 (2014). [Sch14b] C. Schnell, On Saito's vanishing theorem, preprint arXiv:1407.3295, to appear in Math. Res. Lett. (2014). [Sev46] F. Severi, Sul massimo numero di nodi di una superficie de dato ordine dello spazio ordinario o di una forma di un iperspazio, Ann. Mat. Pura Appl. 25 (1946), 1–41. [TSPA16] The Stacks Project Authors, Stacks Project (2016), available at http://stacks.math.columbia.edu. [Yan78] T. Yano, On the theory of b-functions, Publ. Res. Inst. Math. Sci. 14 (1978), no. 1, 111–202. Department of Mathematics, University of Michigan, Ann Arbor, MI 48109, USA E-mail address: [email protected] Department of Mathematics, Northwestern University, 2033 Sheridan Road, Evanston, IL 60208, USA E-mail address: [email protected]
1604.00866	4	1604	2018-02-20T13:25:27	ACM sheaves on the double plane	[ "math.AG", "math.AC", "math.RT" ]	The goal of this paper is to start a study of aCM and Ulrich sheaves on non-integral projective varieties. We show that any aCM vector bundle of rank two on the double plane is a direct sum of line bundles. As a by-product, any aCM vector bundle of rank two on a sufficiently high dimensional quadric hypersurface also splits. We consider aCM and Ulrich vector bundles on a multiple hyperplanes and prove the existence of such bundles that do not split, if the multiple hyperplane is linearly embedded into a sufficiently high dimensional projective space. Then we restrict our attention to the double plane and give a classification of aCM sheaves of rank at most $3/2$ on the double plane and describe the family of isomorphism classes of them.	math.AG	math	ACM SHEAVES ON THE DOUBLE PLANE E. BALLICO, S. HUH, F. MALASPINA AND J. PONS-LLOPIS Abstract. The goal of this paper is to start a study of aCM and Ulrich sheaves on non-integral projective varieties. We show that any aCM vector bundle of rank two on the double plane is a direct sum of line bundles. As a by- product, any aCM vector bundle of rank two on a sufficiently high dimensional quadric hypersurface also splits. We consider aCM and Ulrich vector bundles on a multiple hyperplanes and prove the existence of such bundles that do not split, if the multiple hyperplane is linearly embedded into a sufficiently high dimensional projective space. Then we restrict our attention to the double plane and give a classification of aCM sheaves of rank at most 3/2 on the double plane and describe the family of isomorphism classes of them. 1. Introduction Ever since Horrocks proved that a vector on the projective space splits as the sum of line bundles if and only if it has no intermediate cohomology, there have been two directions of study: one is to find out criterion of coherent sheaves that do not split on a given projective scheme X ⊂ Pn, that is, the equivalent condition with which a coherent sheaf is a direct sum of line bundles OX (t), and the other is to classify indecomposable coherent sheaves that have no intermediate cohomology on X, i.e. H i(X, E(t)) = 0 for any t ∈ Z and i = 1, . . . , dim X − 1; they are called arithmetically Cohen-Macaulay (for short, aCM). About the former direction, the case of hyperquadrics was studied in [36]. Madonna in [33] and Kumar, Rao and Ravindra in [30, 31] focused on criteria for vector bundles that do not split, usually of low rank, on hypersurfaces of higher degree. About the latter direction, the classification of aCM vector bundles has been done for several projective varieties such as smooth quadric hypersurfaces in [28, 29], cubic surfaces in [10, 21], prime Fano threefolds in [32], Grassmannian varieties in [15] and others. In fact, in [18] a complete list of varieties supporting a finite number of aCM sheaves is provided. Varieties that only support one dimensional families of aCM vector bundles, tame varieties, are known by the classical work of Atiyah in [3] for elliptic curves, and much more recently by work of Faenzi and Malaspina in [22] for rational scrolls of degree four. In [13] it is shown that all the Segre varieties have a wild behaviour, namely they support families of arbitrary dimension of aCM sheaves. Finally, it 2010 Mathematics Subject Classification. Primary: 14F05; Secondary: 13C14, 16G60. Key words and phrases. arithmetically Cohen-Macaulay sheaf, double plane, layered sheaf. The first and third authors are partially supported by GNSAGA of INDAM (Italy) and MIUR PRIN 2015 'Geometria delle variet`a algebriche'. The second author is supported by Basic Sci- ence Research Program 2015-037157 through NRF funded by MEST and the National Research Foundation of Korea(KRF) 2016R1A5A1008055 grant funded by the Korea government(MSIP). The forth author is supported by a FY2015 JSPS Postdoctoral Fellowship. The third and forth authors thank Sungkyunkwan University for the warm hospitality. 1 2 E. BALLICO, S. HUH, F. MALASPINA AND J. PONS-LLOPIS has been shown that the rest of aCM integral projective varieties which are not cones are wild; see [23]. Along these lines, the particular class of aCM sheaves supporting the maximum permitted number of global sections has raised the attention of many algebraic geometers in the last years. They are the so-called Ulrich sheaves; see [20]. The existence of an Ulrich sheaf on a projective integral variety X ⊂ Pn has very strong consequences. For instance, this implies that the cone of cohomology tables of vector bundles on X are the same as the one on a projective space of the same dimension; see [19]. Moreover, the Cayley-Chow form of X has a particular nice description; see [20]. Eisenbud and Schreyer stated the existence of Ulrich sheaves on any projective schemes as a problem in [20, page 543]. Until now, this problem has been solved for arbitrary curves, for some minimal smooth surfaces of Kodaira dimension less than or equal to zero and for some sporadic cases of higher dimension; see [14]. A nice up-to-date account can be found in [6]. As it can be seen from the previous paragraphs, up to now most of the research on aCM and Ulrich sheaves has been restricted to the case of integral varieties. The main goal of this paper is to start the study of the aforementioned issues for non-integral ones. For instance, we expect, relying on the results from this paper, that original and interesting behaviours for aCM sheaves on non-reduced varieties can be revealed. In this paper we work on the classification of aCM sheaves on the double plane X, i.e. the projective plane H ⊂ P3 with multiplicity two over an algebraically closed field of characteristic 0. It is known from [5, Theorem A] that any aCM sheaf E on X admits a OP3-free resolution of length one. Our first result is on the aCM vector bundles of rank two on X. Theorem 1.1. Every aCM vector bundle of rank two on X is a direct sum of two line bundles. Theorem 1.1 is not extended to higher rank; indeed, we find a family of indecom- posable aCM vector bundle of rank four on X; see Proposition 4.13. On the other hand, it is extended to higher dimensional quadric hypersurfaces, using an induc- tive argument; see Lemma 4.15 and Corollary 4.16. There is an indecomposable aCM vector bundle of rank two on any union of planes with multiplicities if at least one plane occurs with multiplicity one; see Corollary 3.10. On the other hand, see Proposition 3.6 and Remark 3.7 for a conditional existence theorem of aCM vector bundles on any configuration of hyperplanes with multiplicities. Then we focus our attention to general aCM sheaves on X of rank at most 3/2. In general the rank need not to be integer (see Definition 2.3). Our technical ingredient is to twist a given aCM sheaf E by OX (t) with t ∈ Z so that E is 0-regular, but not (−1)-regular (we call it minimally regular ). It guarantees the existence of a non-zero map u : E → IA(1) for a closed subscheme A that is cut out scheme-theoretically in X by linear equations. Thus we have candidate for a possible subscheme A and describe ker(u) in each case. The structure sheaf OH of H can be shown to be the unique aCM sheaf of rank 1/2 up to twist. For higher rank we introduce the notion of a layered sheaf which admits a filtration whose successive quotient is isomorphic to OH up to twist. It turns out that every aCM sheaf of rank one on X is layered. Theorem 1.2. If E is an aCM sheaf of rank one on X, then it is isomorphic to (i) a line bundle on X ACM SHEAVES ON THE DOUBLE PLANE 3 (ii) a direct sum of two aCM sheaves of rank 1/2, or (iii) there is t ∈ Z such that E(t) is isomorphic to the ideal sheaf of a plane curve in H. In particular, from the previous theorem, we can spot a new kind of wildness presumably that does not occur for smooth projective varieties; compare to [23]: Proposition 1.3. The double plane X is of wild type in a very strong sense, that is, there exist arbitrarily large dimensional families of pairwise non-isomorphic aCM sheaves of fixed rank one on X. Observing from Theorem 1.2 that every aCM sheaf of rank one on X is layered in a sense that each aCM sheaf admits a filtration whose successive quotients are aCM sheaves of rank 1/2. For aCM sheaves of rank 3/2 on X the situation is richer: Theorem 1.4. Let X ⊂ P3 be the double plane. (1) There exists a non-layered Ulrich stable sheaf E of rank 3/2 on X. More- over, every non-layered aCM sheaf of rank 3/2 on X is isomorphic to E(t) for some t ∈ Z. (2) For any layered aCM sheaf E on X of rank 3/2, there exists an integer t ∈ Z such that either (i) E(t) admits a filtration 0 = E0 ⊂ E1 ⊂ E2 ⊂ E3 = E(t) such that Ei/Ei−1 ∼= OH for each i = 1, 2, 3; (ii) it fits into the following sequence with a ≥ d, 0 → IC (a) → E(t) → OH → 0, for C ⊂ H a plane curve of degree d; (iii) it fits into the following sequence with 0 ≤ b < d, 0 → OH (b) → E(t) → IC (d) → 0, for C ⊂ H a plane curve of degree d. The only non-layered sheaf E in (1) of Theorem 1.4 is a non-trivial extension of Ip(1) by OH (−1) as OX -sheaves, where Ip(1) is the ideal sheaf of a point p ∈ X. It turns out that the sheaf E is independent of the choice of the point p ∈ X up to twists; we refer to Propositions 6.11 and 6.13. For the description of type (2-i) we refer to Lemma 6.7 with Remark 6.9. In case a = d = 1 of type (2-i) we get Ulrich sheaves. By Lemma 6.21 any sheaf fitting into the non-trivial sequence in (2-ii) is indecomposable. In fact, the isomorphism classes of such sheaves with a > deg(C) are parametrized by the orbits of Aut(IC (a)) acting on Ext1 X (OH , IC (a)) \ {0}. Lastly the description of type (2-iii) may be seen in Example 6.17 and 6.18. Then we describe the (non)-existence of (non)-layered Ulrich sheaves on X. It turns out that there is no layered Ulrich vector bundle on X, while there exist some non-layered Ulrich vector bundles of rank divisible by four; see Propositions 7.5 and 7.6. Indeed, there exists a layered indecomposable Ulrich sheaf with arbitrary half integral rank; see Theorem 7.7. Let us summarize here the structure of this paper. In section 2 we introduce the definition of aCM sheaves and a generalized notion of rank, possibly not an integer. In section 3 we collect several technical results on the restriction of aCM vector bundle and show the existence of an aCM vector bundle that does not split of arbitrary rank on an arbitrary generalized hyperplane arrangement, when it is 4 E. BALLICO, S. HUH, F. MALASPINA AND J. PONS-LLOPIS embedded linearly into a sufficiently high dimensional projective space. In section 4 we show that every aCM vector bundle of rank two on X splits. As a generalization we also show that any aCM vector bundle of rank two splits on any sufficiently high dimensional quadric hypersurface. In section 5 we deal with the aCM sheaves of rank 1/2 and 1 to give their complete classification, which induces the wildness of the double plane. We also show the existence of arbitrarily large dimensional family of indecomposable layered aCM sheaves of any rank at least one, which also implies the wildness. In section 6 we focus our attention to the case of rank 3/2. We start from calculating numeric data of extension groups on aCM sheaves of lower ranks. Our main result in this section is the unique existence of non-layered aCM sheaf of rank 3/2 up to a twist, which is also semistable and simple. Then we describe the family of isomorphism classes of non-layered sheaves. Finally in section 7, we prove the existence of layered indecomposable Ulrich sheaves for each half integral rank. We are deeply grateful to the anonymous referee for numerous corrections and very stimulating observations. 2. Preliminaries Throughout the article our base field k is algebraically closed of characteristic 0. We always assume that our projective schemes X ⊂ PN have pure dimension at least two and are arithmetically Cohen-Macaulay, namely, h1(IX (t)) = 0 for all t ∈ Z and hi(OX (t)) = 0 for all t ∈ Z and all i = 1, . . . , dim X − 1. Then by [37, Th´eor`eme 1 in page 268] all local rings OX,x are Cohen-Macaulay of dimension dim X. From h1(IX ) = 0 we see that Xred is connected. Since in all our results we have N = dim X + 1, the reader may just assume that X is a hypersurface. For a vector bundle E of rank r ∈ Z on X, we say that E splits if E ∼= ⊕r i=1OX (ti) for some ti ∈ Z with i = 1, . . . , r. We always fix the embedding X ⊂ PN and the associated polarization OX (1). For a coherent sheaf E on a closed subscheme X of a fixed projective space, we denote E ⊗ OX (t) by E(t) for t ∈ Z. For another coherent sheaf G, we denote by homX (F , G) the dimension of HomX (F , G), and by exti X (F , G) the dimension of Exti X (F , G). Now recall that the depth of a module M over a local ring A is defined to be the maximal length of M -regular sequence; see [27, page 4]. We say that a coherent sheaf E on X has pure depth k, if the depth of Ex over OX,x is k for all x ∈ X. We denote the pure depth by depth(E). For a full account on depth and related properties, see [8, Chapter 1]. Definition 2.1. A coherent sheaf E on X ⊂ PN is called arithmetically Cohen- Macaulay (for short, aCM) if the following hold: (i) the dimension of the support of E is equal to dim(X), (ii) the stalk Ex has positive depth for any point x on X, and (iii) H i(E(t)) = 0 for all t ∈ Z and i = 1, . . . , dim(X) − 1. Remark 2.2. (i) Since E is coherent, the condition in Definition 2.1 that the stalk Ex has positive depth for each x ∈ X is equivalent to H 0(E(−t)) = 0 for t ≫ 0 by [37, Th´eor`eme 1 in page 268]. Furthermore, if H 1(E(−t)) = 0 for all t ≫ 0, then every stalk of E has depth at least two also by [37, Th´eor`eme 1 in page 268]. This implies that it has depth dim(X), namely it ACM SHEAVES ON THE DOUBLE PLANE 5 is locally Cohen-Macaulay, again by [37, Th´eor`eme 1 in page 268] together with the vanishing of the intermediate cohomologies of E. (ii) Notice that being aCM does not depend on a twist of E by OX (1). If E 6∼= 0 is a coherent sheaf on a closed subscheme X of a fixed projective space, then we may consider its Hilbert polynomial PE(m) ∈ Q[m] with the leading coefficient µ(E)/d!, where d is the dimension of Supp(E) and µ = µ(E) is called the multiplicity of E. The normalized Hilbert polynomial of E is defined to be the Hilbert polynomial of E divided by µ(E). Definition 2.3. If dim Supp(E) = dim(X), then the rank of E is defined to be rank(E) = µ(E) µ(OX ) . Otherwise it is defined to be zero. For an integral scheme X, the rank of E is the dimension of the stalk Ex at the generic point x ∈ X. But in general rank(E) needs not be integer. Definition 2.4. An initialized coherent sheaf E on X ⊂ PN (i.e. 0 = h0(E(−1)) < h0(E)) is called an Ulrich sheaf if it is aCM and h0(E) = deg(X)rank(E). Ulrich sheaves have received a lot of attention during the last years. It is a central problem on this area to know which (if all) projective schemes support Ulrich sheaves. We refer the reader to [6] and [20] for a complete introduction to the theory of Ulrich sheaves. The following definition will be used extensively throughout the paper: Definition 2.5. A coherent sheaf E of positive depth on X ⊂ PN is called mini- mally regular if it is 0-regular but it is not (−1)-regular. For any coherent sheaf E of positive depth there exists t ∈ Z such that E(t) becomes minimally regular. Let us recall that a minimally regular sheaf is globally generated. Notice also that any Ulrich sheaf is minimally regular; see [20]. Let S = k[x0, . . . , xn] and f ∈ S be a nonzero homogeneous element in the irrelevant ideal. Then an aCM sheaf E on the hypersurface X = V (f ) is given as the cokernel of the map M: (1) 0 −→ ⊕e i=1OPn(ai) M −→ ⊕e i=1OPn (bi) −→ E −→ 0, where M = (mij) is a square matrix of order e with homogeneous entries of degree max{bi − aj, 0} in S; see [5, Theorem A]. In particular we get det(M) = f a with i=1(bi − ai)/ deg(f ). If X is irreducible, then we get a = rank(E); see [17, a = Pe Proposition 5.6]. Now we pay our attention to a special type of schemes, an arbitrary finite union of hyperplanes of Pn+1 with prescribed multiplicities: fix k positive integers m1, . . . , mk and k distinct hyperplanes M1, . . . , Mk of Pn+1 such that M1 ∪ · · · ∪ Mk is not necessarily a normal crossing divisor. Set m := m1 + · · · + mk and X = Xn[m1M1, . . . , mkMk] := m1M1 ∪ · · · ∪ mkMk as a hypersurface of degree m in Pn+1. Thus it is a polarized projective scheme with OX (1) as its polarization. Lemma 2.6. For X = Xn[m1M1, . . . , mkMk], we have Pic(X) ∼= ZhOX (1)i. 6 E. BALLICO, S. HUH, F. MALASPINA AND J. PONS-LLOPIS Proof. Set m := m1 + · · · + mk. Since Pic(Pn) ∼= ZhOPn (1)i, we may assume m ≥ 2 and use induction on m, i.e. we assume that the lemma is true for smaller ∼= OMk , multiplicities. For a fixed L ∈ Pic(X), it is sufficient to prove that if LMk then L ∼= OX . Set Y := Xn[m1M1, . . . , (mk − 1)Mk], with the convention that Mk does not appear inside the square brackets if mk = 1. Since m ≥ 2 and LMk ∼= OMk , by tensoring the exact sequence 0 −→ OMk (1 − m) −→ OX −→ OY −→ 0 with L, we get that the restriction map H 0(L) → H 0(LY ) is bijective. Assume for the moment mk ≥ 2, i.e. Mk ⊆ Y . By the inductive assumption we ∼= OY . Thus we get h0(L) = 1 and a section σ ∈ H 0(L) with no zero at have LY every point of Yred = Xred. Thus σ : OX → L is an isomorphism. Now assume mk = 1. Exchanging the labels of the planes we see that it is sufficient to prove the assertion in the case mi = 1 for all i, in which we have ∼= OMk , m = k. Fix i ∈ {1, . . . , k − 1} and set Li := Mi ∩ Mk. Since we have LMk ∼= OMi for all i. The inductive assumption on ∼= OLi, in particular LMi we get LLi ∼= OY and we may repeat the argument given for the case mk ≥ 2. (cid:3) m gives LY In particular, there is no ambiguity on the choice of the ample generator OX (1) of Pic(X) on X = Xn[m1M1, . . . , mkMk] with respect to which we consider aCM sheaves on X. Notice, moreover, than on X, as on any hypersurface, any line bundle OX (t) is aCM. In particular, the structure sheaf OMi of the reduction of any of the components of X is again aCM as an OX -sheaf. As a special case, let us assume that k = 1; for a fixed hyperplane Hn ⊂ Pn+1 with n ≥ 2, define a projective scheme Xn[m] := Xn[mHn] to be the effective ∼= OXn[m](m − Cartier divisor mHn for m > 0. It has the dualizing sheaf ωXn[m] n − 2). For example, in case n = 2, we have X = X2[2] the double plane. If w is a defining equation of Hn, then for m ≥ 2, the restriction map OXn[m] → OXn[m−1] has kernel isomorphic to OHn (1 − m): (2) 0 −→ OHn (1 − m) −→ OXn[m] −→ OXn[m−1] −→ 0. If E is an aCM sheaf on Xn[m], then it is given as the cokernel of the map defined by the matrix M in (1) with det(M) = wma for some a ∈ (cid:0) 1 m(cid:1)N. Note that by Definition 2.3, the rank of E belongs to (cid:0) 1 Remark 2.7. For an aCM sheaf E ∼= OH (a) of rank 1/2 on the double plane X = X2[2H], the corresponding matrix in (1) is M = (w). Thus E is isomorphic to the cokernel of the map OP3(−1) → OP3 given by the multiplication by w, up to twist. Recall that H is the plane given by w = 0. m(cid:1)N. Now we introduce a special type of coherent sheaves as in [9, 1st Definition at page 318] and [10, Definition 6.5]. Definition 2.8. Fix r ∈ (cid:0) 1 be layered if there exists a filtration 0 = E0 ⊂ E1 ⊂ · · · ⊂ Emr−1 ⊂ Emr = E of E with Ei/Ei−1 ∼= OHn (ai) with ai ∈ Z for all i = 1, . . . , mr. m(cid:1)N. A coherent sheaf E of rank r on Xn[m] is said to It is automatically true by definition that any layered coherent sheaf on Xn[m] is aCM. We end the section with a technical Lemma: ACM SHEAVES ON THE DOUBLE PLANE 7 Lemma 2.9. We have Ext1 Pn+1(OHn (a), OXn[m]) ∼= H 0(OHn (1 − a)) for all a, m ∈ Z with m > 0. Proof. Applying the functor HomPn+1(OHn (a), −) to (3) we get (4) 0 −→ OPn+1(−m) −→ OPn+1 −→ OXn[m] −→ 0, 0 −→ HomPn+1(OHn (a), OXn[m]) −→ Ext1 Pn+1(OHn (a), OPn+1(−m)) −→ Ext1 −→ Ext2 Pn+1(OHn (a), OPn+1) −→ Ext1 Pn+1(OHn (a), OPn+1(−m)), Pn+1(OHn (a), OXn[m]) Pn+1(OHn (a), OPn+1(−m)) ∼= H n−1(OHn (a + m − n − 2))∨ where the last term Ext2 is trivial. The first map is an isomorphism, because we have the following from (2) and Serre's duality HomPn+1(OHn (a), OXn[m]) ⊇ HomPn+1(OHn (a), OHn (1 − m)) ∼= H 0(OHn (1 − a − m)) Pn+1(OHn (a), OPn+1(−m)) ∼= H n(OHn (a + m − n − 2))∨ ∼= H 0(OHn (1 − a − m)). Ext1 Now the assertion follows from the isomorphism Ext1 Pn+1(OHn (a), OPn+1) ∼= H 0(OHn (1 − a)). (cid:3) 3. aCM vector bundles on hyperplane arrangements with multiplicities In this section we are going to consider aCM vector bundles E of rank r ≥ 2 on Xn[m] with m ≥ 2. Lemma 3.1. For two integers m, n ≥ 2, and for each integer s ≥ 2n + 2, there exist in Ps • a smooth hypersurface Y of degree m, and • an (n + 1)-dimensional linear subspace M such that Y ∩ M = Xn[m]. Proof. We fix homogeneous coordinates [x0 : . . . : xs] of Ps and write M := {xn+2 = · · · = xs = 0} and Hn := {x0 = xn+2 = · · · = xs = 0} ⊂ M , i.e. Hn is the hyperplane in M defined by x0 = 0. Consider the smooth hypersurface Y = V (f ) ∈ OPs(m) for (5) f = xm 0 + n+1 Xk=1 xm−1 k xk+n+1 + Xi≥n+2 xm i . Then we have Y ∩ M = Xn[m]. (cid:3) Remark 3.2. The assertion in Lemma 3.1 was originally proved by dimension counting, while the current proof using the explicit equation (5) for a smooth hy- persurface is given by the referee. Lemma 3.3. Fix a hypersurface Y and a hyperplane M in Ps with s ≥ 3 such that M is not a component of Y . For an aCM vector bundle E of rank r on Y , if EY ∩M i=1OY ∩M (ai) for some ai ∈ Z, then we have E ∼= ⊕r i=1OY (ai) ∼= ⊕r 8 E. BALLICO, S. HUH, F. MALASPINA AND J. PONS-LLOPIS Proof. We may assume 0 = a1 ≥ · · · ≥ ar and let e be the number of indices i with ai = 0. Then we have h0(EY ∩M ) = e and h0(EY ∩M (−1)) = 0. Since E is aCM, we get that the restriction map ρ : H 0(E) → H 0(EY ∩M ) is bijective. By Nakayama's lemma we also get that the natural map η : H 0(E) ⊗ OY → E is injective and that its image at each p ∈ (Y ∩ M )red spans an e-dimensional linear subspace of the fiber Ep. If e = r, we get that η is an isomorphism, because Y ∩ M is an ample divisor of Y and an injective map between two vector bundles with the same rank is either an isomorphism or drops rank on a hypersurface, which must intersect Y ∩ M . Now assume e < r and let e′ be the number of indices i with ai = ae+1, i.e. the number of second biggest numbers among ai's. Set F := Im(η) ∼= O⊕e Y . Since E/F is locally free at each point of (Y ∩M )red, it is locally free outside a finite set disjoint from Y ∩ M . Now the natural map ρ′ : H 0(E(−ae+1)) → H 0(EY ∩M (−ae+1)) is surjective, because E is aCM. From H 0(FY ∩M (−ae+1)) ∼= H 0(OY ∩M (−ae+1)⊕e), we see that there is an e′-dimensional linear subspace V of H 0(E(−ae+1)) such that V ∩ H 0(F (−ae+1)) = 0 and the map H 0(FY ∩M (−ae+1)) ⊕ V −→ H 0(⊕e+e′ i=1 OY ∩M (ai − ae+1)) = H 0(EY ∩M (−ae+1)) is bijective. Then we get a map η′ : OY (a1)⊕e ⊕ OY (ae+1)⊕e′ → E that is injective and of rank e + e′ at each point of (Y ∩ M )red. If r = e + e′, then we can conclude that the map ρ′ is an isomorphism. If r > e + e′, we continue in the same way using E(−ae+e′+1). (cid:3) Lemma 3.4. Let Y be a hypersurface of Ps and M ⊂ Ps a linear subspace not contained in Y . If E is a Ulrich vector bundle on Y , then so is EY ∩M . Proof. Note by assumption that we have deg(Y ∩ M ) = deg(Y ). First assume that dim M = s − 1. Let us consider the exact sequence 0 −→ E(−1) −→ E −→ EY ∩M −→ 0. Since hi(E(t)) = 0 for any integer t and for any i = 1, . . . s − 2 we get that EY ∩M is aCM. Moreover, since E is initialized, we get h0(EY ∩M (−1)) = 0 and h0(EY ∩M ) = h0(E). This implies that EY ∩M is Ulrich. If dim M ≤ s − 2, then we take a hyperplane of Ps containing M and use induction on the codimension of Y . (cid:3) Thus, when there exists an integer s ≥ 2n+2 such that each smooth hypersurface Y ⊂ Ps of degree m supports an aCM vector bundle of rank r, which is not a direct sum of line bundles, we may apply Lemmas 3.1, 3.3 and 3.4. But one quite often does not have an existence result of aCM vector bundles that do not split with prescribed rank on all smooth hypersurfaces of degree m in a given projective space; for an existence result about the general hypersurface, see [20]. It is sufficient to have the existence of an aCM (or Ulrich) vector bundle of rank r on the hypersurface whose equation is given in the proof of Lemma 3.1. If m > 2, the bounds 2n+2 in Lemma 3.1 are not enough to ensure that a general hypersurface of degree m in Ps contains some Xn[m]. In the following Lemma we require much higher bounds for dimension of the ambient projective space: Lemma 3.5. For two integers n ≥ 2 and m ≥ 3, set Ngen(n, m) := min(cid:26)k > 0 (cid:12)(cid:12)(cid:12)(cid:12) (k + 1)(k − n − 1) ≥ (cid:18)m + n + 1 n + 1 (cid:19)(cid:27) . ACM SHEAVES ON THE DOUBLE PLANE 9 For each integer s ≥ Ngen(n, m) and a general hypersurface Y ⊂ Ps of degree m, there exists an (n + 1)-dimensional linear subspace M ⊂ Ps such that Y ∩ M = Xn[m]. Proof. We fix an (n + 1)-dimensional linear subspace V ⊂ Ps and a hyperplane H n ⊂ V over which we consider Xn[m]. With the linear systems E and E′ in the proof of Lemma 3.1, we consider two natural maps induced by the action of SL(s + 1) on Ps: u : E × SL(s + 1) −→ OPs(m) , u′ : E′ × SL(s + 1) −→ OPs(m). The lemma is equivalent to saying that u is dominant. Since E is irreducible and E′ ⊂ E, it is sufficient to prove that u′ is dominant. Now we have (s+1)(s−n−1) ≥ (cid:3) (cid:0)m+n+1 n+1 (cid:1). This implies that the map u′ is dominant by [16]. Now we generalize the previous set-up further to an arbitrary finite union of hyperplanes of Pn+1 with prescribed multiplicities X = Xn[m1M1, . . . , mkMk]. We may modify the proofs of Lemmas 3.1 and 3.5 to get the following result. Proposition 3.6. For two integers n, m ≥ 2 and any k integers m1, . . . , mk whose sum is m, we get the following over every possible X = Xn[m1M1, . . . , mkMk] embedded linearly in Ps. (i) If s ≥ 2n + 2 and every smooth hypersurface of degree m in Ps has a rank r aCM (resp. Ulrich) vector bundle that does not split, then the same holds for X. (ii) If s ≥ Ngen(n, m) and a general hypersurface of degree m in Ps has a rank r aCM (resp. Ulrich) vector bundle that does not split, then the same holds for X. Remark 3.7. It has been conjectured in [7] that, for smooth hypersurfaces, the rank of aCM (or Ulrich) vector bundles should be at least ⌊ s−2 2 ⌋. The conjecture is sharp and has been proved on hyperquadrics and for rank 2 and 3 vector bun- dles; see [39] and references therein. Proposition 3.6 asserts that if for a certain r ≥ 2 there are aCM (or Ulrich) vector bundles of rank r that do not split on a general hypersurface of degree m in Ps with s ≫ 0, then the same is true for Xn[m1M1, . . . , mkMk]. Now we consider the case n = 2 and mi = 1 for at least one index i, i.e. we assume dim(X) = 2 and Xreg 6= ∅. The case m = 2, i.e. X is the union of two distinct planes in P3, of the following example appears in [4, Example 4.1]. In fact, in the following example, we require for our surface X ⊂ P3 only to be any arbitrary surface with at least one irreducible component of Xred appearing with multiplicity one in X. Example 3.8. Let X ⊂ P3 be a surface of degree m ≥ 2 with Xreg 6= ∅. For a fixed point p ∈ Xreg, we have Ext1 X (Ip,X (1), OX (m − 2)) ∼= H 1(Ip,X (−1))∨ ∼= k by Serre's duality. So up to isomorphisms there exists a unique non-trivial extension E as an OX -sheaf, fitting into the exact sequence (6) 0 −→ OX (m − 2) −→ E −→ Ip,X (1) −→ 0. Such sheaf E is uniquely determined by the point p and it is locally free outside p. 10 E. BALLICO, S. HUH, F. MALASPINA AND J. PONS-LLOPIS Claim 1: E is locally free of rank two. Proof of Claim 1: Obviously E has rank two on each of the components of X and it is locally free outside p. Note that X is Gorenstein and by assumption p is a smooth point of X. Since ωX ∼= OX (m − 4), we have h0(ωX ⊗ OX (3 − m)) = 0. So the Cayley-Bacharach condition is satisfied and E is locally free; see [11], where one only considers the case in which X is a Gorenstein normal surface, but in this particular case with p ∈ Xreg we can adapt the proof in [11] or the classical proof using the duality of the Cayley-Bacharach property; also we only need one implication of Cayley-Bacharach, not the " if and only if " statement. Claim 2: E is aCM. Proof of Claim 2: Since h1(Ip,X (i)) = 0 for all i ≥ 0, (6) gives h1(E(i)) = 0 Its dual asserts that h1(E ∨(i)) = 0 for all i ≤ 3 − m. From for all i ≥ −1. det(E) ∼= OX (m − 1) we have E ∨ ∼= E(1 − m). Thus we get h1(E(i)) = 0 for all i ≤ 2, concluding the proof of Claim 2. On the other hand, from (6) we have h0(E(1 − m)) = 0 and h0(E(2 − m)) > 0. In particular, E is not Ulrich for m ≥ 3, while in case m = 2 it is Ulrich. Claim 3: There are no integers a, b such that E ∼= OX (a) ⊕ OX (b). Proof of Claim 3: Here we use that m ≥ 2, because for m = 1 we would just get the trivial vector bundle of rank two. Assume that such a, b exist, say a ≥ b. Since h0(E(2 − m)) = 1 and h0(E(1 − m)) = 0, we get (a, b) = (m − 2, 1) and m ≥ 3. Then we get h0(E) = (cid:0)m+1 3 (cid:1) + 4, while (6) gives h0(E) = (cid:0)m+1 3 (cid:1) + 3. Remark 3.9. In Example 3.8 we do not claim that E does not split, e.g. if X is a smooth quadric, then E is the direct sum of the two spinor line bundles of X, up to a twist. If Pic(X) ∼= ZhOX (1)i, then E is indecomposable; it happens when X = X2[m1M1, . . . , mkMk] by Lemma 2.6. By Lemma 2.6 and Example 3.8 we immediately get the following result. Corollary 3.10. Let X = X2[m1M1, . . . , mkMk] ⊂ P3 be a union of planes with mi = 1 for at least one index i. Then there exists an indecomposable aCM vector bundle of rank two on X. Remark 3.11. Take X as in Corollary 3.10 with m > 3. For any p ∈ Xreg call Ep the aCM vector bundle of rank two in (6). Since m ≥ 3, we have h0(Ep(2 − m)) = 1, in particular (6) is the Harder-Narasimhan filtration of Ep. By uniqueness of the Harder-Narasimhan filtration, we have Ep 6∼= Eq for p 6= q ∈ Xreg. For large m we may find X with finite automorphism groups. Hence we get a 2-dimensional family of aCM vector bundles of rank two, even after the action of Aut(X). 4. aCM vector bundles on the double plane In this section, we discuss the aCM vector bundles on X := X2[2], i.e. the double plane in P3. Set S = k[x, y, z, w] and assume that X is given by f = w2, i.e. X is the double plane whose reduction is the plane H given by w = 0. By Lemma 2.6, we have Pic(X) = Z generated by OX (1) associated to the hyperplane class and ωX ∼= OX (−2); see also [26, Example 5.10]. Since every line bundle on H is aCM, every line bundle on X is also aCM due to the following exact sequence (7) 0 −→ OH (−1) −→ OX −→ OH −→ 0. In particular, every direct sum of line bundles on X is aCM. Note also that any extension of an aCM vector bundle E by a line bundle on X splits, because ACM SHEAVES ON THE DOUBLE PLANE 11 Ext1(E, OX (k)) ∼= H 1(E(−k − 2))∨ is trivial for any k ∈ Z. By Definition 2.3 we get rank(E) is in r ∈ (cid:0) 1 2(cid:1)Z. The ideal sheaf IH of H in X is OH (−1) in (7), in particular it is an aCM sheaf of rank 1/2. If E is a vector bundle on X, then by tensoring (7) with E we get the following exact sequence (8) 0 −→ E(−1)H −→ E −→ EH −→ 0. In the next lines we gather some crucial properties of minimally regular aCM sheaves on X that will be use for the rest of the paper. If E is such a sheaf on X := X2[2], it follows immediately from the definition that HomX (E, OX (1)) ∼= H n(E(−3)) 6= 0, so there exists an exact sequence of OX -sheaves (9) 0 −→ L −→ E π −→ IA(1) −→ 0 with A a proper closed subscheme of X and L := ker(π). The sheaf IA(1) is a nonzero subsheaf of OX (1), in particular it has positive depth. Let hAi ⊆ P3 denote the linear span of A. Since E is 0-regular, it is globally generated. This implies that IA(1) is also globally generated. Thus we get A = hAi ∩ X as schemes. Taking the possible linear subspaces hAi, we get that A is one of the schemes in the following list: Possible Cases 4.1. (a) A = ∅ and IA(1) ∼= OX (1). (b) A = H and A is cut out in X by only one linear equation; IA(1) ∼= OH . (c) A is a line L ⊂ H; A is cut out in X by the plane H and another plane that is different from H. (d) A is a connected scheme of degree two with Ared = {p} a point; A is a complete intersection of X with the line hAi 6⊆ H. (e) A = {p} for some point p ∈ H. (f) A = X ∩ M with M ⊂ P3 a plane that is different from H; IA(1) ∼= OX . Remark 4.2. If A ⊇ H, then we have IA(1) ⊂ OH . Since IA(1) is globally generated, we have IA(1) ∼= OH . So we get the case (b). Lemma 4.3. Let E be a minimally regular aCM sheaf on X and consider the associated short exact sequence (9). Then: (i) IA is a 1-regular OX -sheaf. (ii) If dim A 6= 0, then L is 0-regular and aCM sheaf. (iii) If dim A = 0, then L is 1-regular and aCM sheaf. (i) From the surjective map H 2(E(t)) → H 2(IA(t + 1)) we see that Proof. H 2(IA(t + 1)) = 0 for t ≥ −2. On the other hand, from the short exact sequence 0 −→ IA(t) −→ OX (t) −→ OA(t) −→ 0 (10) we also obtain H 1(IA(t)) = 0 for t = 0 (A is always connected), so IA is 1-regular. In particular, H 1(IA(t)) = 0 for t ≥ 0. Indeed, notice that in the cases when dim A 6= 0, the ideal sheaf IA is an aCM OX -sheaf. (ii) Again from (9) we get H 1(L(t)) = 0 for t < 0. Now, from the isomorphism 0 = H 1(IA(−1)) → H 2(L(−2)) → 0, we get at once that L is 0-regular. Since clearly L also has positive depth at any point x ∈ X, we can conclude that L is an aCM sheaf. (iii) In this case, we only get 0 = H 1(IA) ∼= H 2(L(−1)). But, on the other hand, note that IA(1) is generated by the image of H 0(E) and IA(1) needs two 12 E. BALLICO, S. HUH, F. MALASPINA AND J. PONS-LLOPIS sections to be generated. This implies h1(L) = 0 from h1(E) = 0. Therefore L is 1-regular. Since we have H 1(L(t)) = 0 for t < 0 in any case, L is also aCM. (cid:3) Remark 4.4. From the previous Lemma, we see that in cases (a) and (f), the extension should be trivial and therefore E ∼= L ⊕ IA(1). Lemma 4.5. Let E be an aCM vector bundle of rank r on X. If EH splits, then E also splits. Proof. Let E be an aCM vector bundle on X such that EH i=1OH (ai) with a1 ≥ · · · ≥ ar. Up to a twist we may assume that E is minimally regular and therefore from the long exact sequence associated to (8) we get ar ≥ 1. This implies that E(−1)H is globally generated. ∼= ⊕r Note that the restriction map H 0(E(−1)) → H 0(E(−1)H ) is surjective and H 0(E(−1)H) globally generates E(−1)H . Take a nonzero section s in H 0(E(−1)H ) induced from the last factor OH (ar − 1) ∼= OH . The section s lifts to a section σ ∈ H 0(E(−1)) such that σH = s. Since s vanishes at no point of H, σ also vanishes at no point of H. Let j : OX2[2] −→ E be the map induced by σ. Set G := coker(j). Claim: The map j is injective and G is a a vector bundle of rank r − 1 on X2[2]. Proof of Claim: For a fixed point p ∈ H, let E(−1)p and Gp denote the stalks of E and G at p, respectively. Let jp : OX2[2],p → E(−1)p be the map induced by σ. Since E is locally free, there is an isomorphism ϕ : E(−1)p → O⊕r X2[2],p of OX2[2],p- modules. The map ϕ ◦ jp : OX2[2],p −→ O⊕r X2[2],p is given by some a = (a1, . . . , ar) ∈ O⊕r X2[2],p. The condition σ(p) 6= 0 is equivalent to the existence of i ∈ {1, . . . , r} such that ai is not contained in the maximal ideal of OX2[2],p. Thus ϕ ◦ jp is injective and Gp is isomorphic to a direct factor of O⊕r X2[2],p (and hence it is locally free) with rank r − 1. (cid:3) The restriction to H of the injective map j maps OH (1) onto a factor of EH. Thus GH is isomorphic to a direct sum of line bundles on Hn, hence it is aCM. The exact sequence (8) with G instead of E gives that G is aCM. By induction on the rank we get that G is isomorphic to a direct sum of line bundles. Thus E is isomorphic to a direct sum of line bundles. (cid:3) Remark 4.6. An OP3-sheaf E is an OX -sheaf if and only if w2E = 0, i.e. wIm(fw) = 0, where the map fw : E → E(1) is induced by the multiplication by w; (11) 0 −→ ker fw −→ E −→ Imfw −→ 0. In particular, any OX -sheaf admits an extension of an OH -sheaf by another OH - sheaf. If E is a locally free OX -sheaf, then this exact sequence is exactly (8). Now consider any OP3-sheaf E that is an extension, as an OP3 -sheaf, of an OH -sheaf E1 by another OH -sheaf E2. We claim that E is an OX -sheaf. Indeed, Im(fw) is isomorphic to a subsheaf of E2(1). So we get wIm(fw) = 0. Thus E is an OX -sheaf.. i=1OH (ti) with t1 ≥ · · · ≥ ts be the kernel of the map fw,E : E → E(1) induced by the multiplication by w as in Remark 4.6. The image F2 := Im(fw,E) ⊂ E(1) is a torsion-free sheaf on H with rank 2r − s. Since w2 = 0, we have F2 ⊆ F1(1), in particular we get s ≥ r. We obviously have s ≤ 2r, and s = 2r if and only if E ∼= ⊕s Remark 4.7. Let E be an aCM sheaf of rank r ∈ (cid:0) 1 2(cid:1)Z and let F1 := ⊕s i=1OH (ti). ACM SHEAVES ON THE DOUBLE PLANE 13 Our next goal is to prove Theorem 1.1, namely that an aCM vector bundle E of rank two on X splits. Remark 4.8. We know that E fits in the exact sequence (9) with A a subscheme from the list of possible cases 4.1. In case (a) we have IA(1) ∼= OX (1). In case (f), A is an effective divisor in OX (1), in particular we have IA(1) ∼= OX . In either case, the vector bundle E corresponds to an element in Ext1 X (OX (i), OX (c1 − i)) = 0 with i ∈ {0, 1}. Thus E splits. In case (b), we have IA(1) = OH . Since the tensor product is a right exact functor, the surjection π : E → OH induces a surjection π1 : EH → OH . Since E is globally generated, EH is also globally generated. The surjection π1 implies that EH splits. Then by Lemma 4.5, E splits. In case (c), since the tensor product is a right exact functor, the surjection π induces a surjection EH → IL(1) ⊗ OH . The OH -sheaf IL(1) ⊗ OH has a torsion part τ supported on L and (IL(1) ⊗ OH )/τ ∼= OH . Thus we obtain a surjection π1 : EH → OH and again we obtain the decomposability of E. By Remark 4.8 it remains to deal with the cases (d) and (e), concerning the decomposability of E. In both cases, then the map π induces a surjection π1 : EH → Ip,H (1). Since Ip,H (1) has no torsion and EH is locally free, we get that ker(π1) has rank one with pure depth two. Thus ker(π1) is isomorphic to OH (c1−1), where det(E) ∼= OX (c1). Lemma 4.9. For each integer a ≥ 0 and a point p ∈ H, there exists a unique vector bundle Ep,a of rank two on H fitting into the exact sequence (12) 0 −→ OH (a) −→ Ep,a −→ Ip,H (1) −→ 0, up to isomorphism. Here, the vector bundle Ep,a is globally generated. Proof. There exists a vector bundle Ep,a fitting into (12), because the Cayley- Bacharach condition is satisfied. And any such sheaf is globally generated, because OH (a) and Ip,H (1) are globally generated and h1(OH (a)) = 0. Thus it remains to prove that the vector bundle Ep,a is unique, up to isomorphism, and it is sufficient to prove that the dimension of Ext1 H (Ip,H (1), OH (a)) is at most one. This is true, be- cause h1(OH (a − 1)) = 0, the local Ext1-group is the skyscraper sheaf kp supported by p and we may use the local-to-global spectral sequence of the Ext-functor. (cid:3) Remark 4.10. In case a = 0, we have Ep,0 ∼= Ω1 Lemma 4.11. For a fixed integer a > 0, we have the following: H (2) for any choice of p ∈ H. (i) we have h2(Ep,a(t)) = 0 if and only if t ≥ −2, (ii) we have h1(Ep,a(−1)) > 0, and (iii) for any p ∈ H, there is no aCM vector bundle E of rank two on X with EH ∼= Ep,a. Proof. Since p is zero-dimensional, we have h2(H, Ip,H (t)) = h2(H, OH (t)) for all t ∈ Z. Then we get (i) and (ii), by using (12). For (iii) assume that such E exists. Then we have h2(H, EH(−2)) = 0 and h1(H, EH (−1)) > 0 by (i) and (ii). Now we may use (8) to get a contradiction. (cid:3) Proof of Theorem 1.1: Let E be a minimally regular aCM vector bundle of rank two on X, with det(E) ∼= OX (c1). Since E is globally generated, we get c1 ≥ 0. If c1 = 0, then we get h0(E) = h0(E ∨) which implies E ∼= O⊕2 X . So we may assume that c1 ≥ 1. Now by Remark 4.8 it is enough to prove the assertion for the cases (d) 14 E. BALLICO, S. HUH, F. MALASPINA AND J. PONS-LLOPIS or (e) of Possible Cases 4.1 for A. Let τ be the torsion part of the sheaf IA(1)⊗OH . In both cases we have (IA(1) ⊗ OH )/τ ∼= Ip,H (1). Thus the surjection E → IA(1) ∼= Ep,c1−1 by Lemma induces a surjection EH → Ip,H (1), in particular we get EH 4.9. Lemma 4.11 excludes the case c1 > 1. So the only possibility is c1 = 1, when we have EH H (2). Then we have the exact sequence ∼= Ω1 (13) 0 −→ Ω1 H (−1) −→ E(−2) −→ Ω1 H −→ 0. Since h1(Ω1 E is not aCM. H ) = 1 and h2(Ω1 H (−1)) = h0(TH(−2)) = 0, we have h1(E(−2)) ≥ 1. So (cid:3) On the other hand the assertion in Theorem 1.1 may not hold for higher rank. Indeed, the vector bundle SX in Example 4.12 gives a counterexample in rank four; see Proposition 4.13. Example 4.12. Let Q5 ⊂ P6 be a smooth quadric hypersurface and S the spinor bundle on Q5 (of rank four); see [35]. Fix a plane H ⊂ Q5 and take a 3-dimensional linear space H ⊂ V ⊂ P6 such that the quadric Q5 ∩ V has rank one. So we write X = Q5∩V and set SX := SX . Since S(1) is Ulrich, SX (1) is also Ulrich by Lemma 3.4. Set SH := SH . By [35, Theorem 2.5] we have SH ∼= OH ⊕ OH (−1) ⊕ Ω1 H (1). Since SX is locally free, it fits into an exact sequence on X (14) 0 −→ SH (−1) −→ SX −→ SH −→ 0. Let ∆ be the set of isomorphism classes of vector bundles E with E(1) Ulrich and ∼= SH . Then each element of ∆ is an extension of SH by SH (−1), and we have a EH map ∆ → P Ext1 X (SH , SH (−1)). Since SX (1) is Ulrich, we have ∆ 6= ∅ and the im- age of this map contains a non-empty Zariski open subset of P Ext1 X (SH , SH (−1)). Proposition 4.13. For any [E] ∈ ∆, E is indecomposable. Proof. Otherwise, since there is no Ulrich line bundle on X, each summand of E(1) would be an Ulrich bundle of rank two. But by Theorem 1.1 any such bundle would split, a contradiction. (cid:3) In the previous lines we showed the existence of rank four Ulrich vector bundles on the double plane X. On the other hand, we already proved that there are no Ulrich bundles of rank one and two on X. Therefore with the next Proposition we show that four is the lowest possible rank for an Ulrich bundle on X. Proposition 4.14. If E is an Ulrich vector bundle of rank r on the double plane X ⊂ P3, then r is divisible by four. Proof. Let us suppose that E is an Ulrich vector bundle of rank r on X. We know that E is minimally regular and that h0(E(−1)) = 0 by [20, Proposition 2.1]. Therefore, from the long exact sequence of cohomology groups associated to (8) it is immediate to see that h1(EH (t)) = 0 for t ≤ −3. From the 0-regularity of E we have h2(E(−2)) = 0. Thus we get h2(EH (−2)) = 0 and this implies h1(EH(−1)) = 0 by (8). In particular, EH is 0-regular and we get h1(EH (t)) = 0 for t ≥ −1. On the other hand, we have h0(EH (−1)) = h1(EH (−2)) 6= 0; otherwise, E would split by Lemma 4.5. Thus by [1, Theorem 2.2 and Corollary 2.4] we get EH In H for a = h0(EH (−1)) and some b ∈ Z≥0. H (2)⊕a ⊕ OH (1)⊕a ⊕ O⊕b ∼= Ω1 ACM SHEAVES ON THE DOUBLE PLANE 15 particular, we have 2r = 6a + 2b = h0(E) = h0(EH (−1)) + h0(EH ) = a + (6a + b) = 7a + b. This implies that a = b, and in particular we get r = 4a. (cid:3) While the assertion of Theorem 1.1 does not extend to higher rank, it holds for higher dimensional quadrics, even with smaller corank. Lemma 4.15. Let Q ⊂ Pn+1 with n ≥ 3, be a quadric hypersurface of corank at least 3. Any aCM vector bundle of rank two on Q splits. Proof. Let E be an aCM vector bundle of rank two on Q. (a) First assume that Q is a hyperplane with multiplicity two in Pn+1, i.e. Q = Xn[2] for some n ≥ 3. We choose a three-dimensional linear subspace V ⊂ Pn+1 so that V ∩ Q is a double plane in V . Then EV ∩Q is an aCM vector bundle of rank two on the double plane. This implies that it splits. In particular, its restriction to the reduction of V ∩ Q, say H2 := (V ∩ Q)red, splits. Moreover V can be chosen so that H2 can be any plane contained in Hn, which implies that the splitting type of EH2 does not change as H2 varies in Hn. In particular, EHn is a uniform vector bundle of rank two on Hn and this implies that EHn splits. Then E also splits due to Lemma 4.5. (b) Assume n = 3. The case in which Q has corank 4, is true by step (a). In the case when Q has corank 3, i.e. Q = M1 ∪ M2 with M1, M2 two distinct hyperplanes of P4, we may apply [4, Theorem 3.13]. (c) Now assume n > 3 and that the assertion holds for a lower dimensional projective space. Due to step (a) we may assume that Q is a reduced quadric hypersurface. Take a hyperplane M ⊂ Pn+1 such that Q ∩ M has corank k + 1, where k is the corank of Q. Note that EQ∩M is also aCM. By the inductive ∼= OQ∩M (a) ⊕ OQ∩M (b) for some integers a ≤ b. Up assumption we have EQ∩M to a twist we may assume that b = 0. Since n > 3 and Q has corank at least 3, for each p ∈ Q there is a three-dimensional linear subspace W ⊂ Q such that ∼= OW ∩M (a) ⊕ OW ∩M , the argument in step (a) gives p ∈ W . Since EW ∩M ∼= OW (a) ⊕ OW . Note that the every point p ∈ Q is contained in a three- EW dimensional linear subspace W ⊂ Q. Since h1(E(t − 1)) = 0 for all t ∈ Z, the restriction map ρt : H 0(E(t)) → H 0(EQ∩M (t)) is surjective. Since h0(E(t)) = 0 for t ≪ 0 and h0(EQ∩M (t − 1)) = 0 for all t ≤ 0, we get that h0(E(−1)) = 0 and that ρ0 is bijective. Let η : H 0(E) ⊗ OQ → E denote the evaluation map. First assume a = 0. We get h0(E) = 2 and that the evaluation map η is an isomorphism at all points of the ample divisor Q ∩ M . Since H 0(E) ⊗ OQ and E are vector bundles with the same rank, η is an isomorphism. Now assume a < 0. We have h0(E) = 1. Fix p ∈ Q and take a three-dimensional ∼= OW (a) ⊕ OW , η induces a linear space W ⊂ Q such that p ∈ W . Since EW map of rank one from the fiber of H 0(E) ⊗ OQ to the fiber Ep. Thus η is injective and E/Im(η) is a line bundle whose restriction to each W is isomorphic to OW (a). Thus we get E/Im(η) ∼= OQ(a), in particular E splits. (cid:3) Corollary 4.16. Let Q ⊂ Pn+1 with n ≥ 7, be any quadric hypersurface. If E is an aCM vector bundle of rank two on Q, then it splits. 16 E. BALLICO, S. HUH, F. MALASPINA AND J. PONS-LLOPIS Proof. If Q has corank at least 3, then we may use Lemma 4.15. Thus we assume that Q has corank at most 2. In particular, there exists a linear subspace V ⊂ Pn+1 such that dim V = 6 and V ∩ Q is a smooth quadric hypersurface of V . By [36] the restriction EQ∩V splits. Now we may proceed as in step (c) of the proof of Lemma 4.15. (cid:3) 5. Wildness of the double plane Lemma 5.1. Any sheaf of rank 1/2 on X with pure depth 2, is isomorphic to OH (a) for some a ∈ Z. Proof. Let E be a sheaf of rank 1/2 on X with pure depth 2, in particular it is reflexive by [26, Theorem 1.9]. Then F := ker(fw) and G := Im(fw) ⊂ E(1) as in (11) are torsion-free (or trivial) by [8, Proposition 1.2.9]. Since w2E = 0, fw is not injective. This implies that F is non-zero. Thus by additivity of the rank, we have rank(F ) = 1/2 and G ∼= 0. Now F is a reflexive OH -sheaf of rank one. So we get F ∼= OH (a) for some a ∈ Z. (cid:3) Example 5.2. Fix a plane curve C ⊂ H and consider its ideal sheaf in X with the exact sequence 0 −→ IC −→ OX −→ OC −→ 0. (15) Since IC,H ∼= OH (−k) for k = deg(C), it is aCM, in particular the map H 0(OH (t)) → H 0(OC (t)) is surjective. Thus we get that the map H 0(OX (t)) → H 0(OC (t)) is surjective, because it factors through H 0(OH (t)). This implies that IC is a non- locally free aCM sheaf of rank one, because OX is aCM. Note that C is not a Cartier divisor of X, in particular IC is not locally free along C. For a fixed plane curve C ⊂ H of degree d, the injection OH (−1) → OX in (15) factors through IC . Now the cokernel of the map OH (−1) → IC is IC,H , which is OH (−d). So we get an exact sequence (16) 0 −→ OH (−1) −→ IC −→ OH (−d) −→ 0. P3(OH (1 − d), OH ) ∼= By case m = 1 of Lemma 2.9 for X we get that P Ext1 PH 0(OH (d)) and this space parametrizes the plane curves of degree d. Therefore IC (1) determines an element in P Ext1 P3(OH (1 − d), OH ). Notice that indeed we get Ext1 X (OH (1 − d), OH ), as it is easily checked applying the functor HomX (−, OH (d − 1)) to the surjection OX −→ OH . P3(OH (1 − d), OH ) ∼= Ext1 Proof of Proposition 1.3: Fix a positive integer k and take an integer d > 0 such C ⊂ H of degree d. ∆ is a non-empty Zariski open subset of the projective space that (cid:0)d+2 PH 0(OH (d)) of dimension (cid:0)d+2 2 (cid:1) > k. Let ∆ ⊂ PH 0(OH (d)) be the set parametrizing all smooth curves 2 (cid:1) − 1 ≥ k. Thus ∆ is a non-empty algebraic variety of dimension at least k. For any C ∈ ∆, C is the set of all p ∈ H at which IC is not locally free. In particular, if C, D ∈ ∆ and C 6= D, we have IC ≇ ID. Then we may use the family {IC}C∈∆ to get the assertion. (cid:3) Now we classify aCM sheaf of rank one on X to obtain Theorem 1.2. Proof of Theorem 1.2: Let E be a minimally regular aCM sheaf of rank one on X We get a surjective map π : E → IA(1) for a closed subscheme A ( X in Possible Cases 4.1. ACM SHEAVES ON THE DOUBLE PLANE 17 In cases (a), (c), (d), (e) and (f), the surjective map π : E → IA(1) is an isomorphism, because E has rank 1 with pure depth two, in particular it is reflexive In case (a) and (f), E is isomorphic to OX and OX (1), by [26, Theorem 1.9]. In case (c), E ∼= IL(1) and E(−1) is as in (iii) with a line as the respectively. plane curve; see Example 5.2. Cases (d) and (e) are excluded, because we have h1(IA(−1)) = deg(A) from the standard sequence for A ⊂ X; 0 −→ IA(−1) −→ OX (−1) −→ OA(−1) −→ 0, and this implies that A is not aCM. Finally in case (b), we have IA(1) ∼= OH and by Lemma 4.3 so by Example 5.2 we get (ii) or (iii). (cid:3) Now we discuss wildness in higher rank. For a fixed r ∈ (1/2)N that is at least one, take two positive integers r1 and r2 such that r1 + r2 = 2r together with two −→ k = (k1, . . . , kr1 ) ∈ Z⊕r1 and −→m = (m1, . . . , mr2) ∈ Z⊕r2. sequences of integers Define two vector bundles on H that split as follows: A := ⊕r1 j=1OH (kj ) , B := ⊕r2 h=1OH (mh). Then Γ := P Ext1 m = 1 of Lemma 2.9 for X and each element λ ∈ Γ corresponds to a unique aCM sheaf Eλ on X of rank r, given as an extension of B by A. Note that all sheaves Eλ are layered. P3(B, A) is of dimension −1 + Pj,h max{0,(cid:0)3+kj −mh (cid:1)} by case 2 Proposition 5.3. Fix r ∈ (1/2)N, r ≥ 1, and assume (r1, r2) = (2r − 1, 1) with −→ k = (k, . . . , k) and −→m = (m) such that m < k and (cid:0)3+k−m general λ ∈ Γ, the sheaf Eλ is indecomposable. 2 (cid:1) ≥ 2r − 1. Then for a P3(OH , OH (k)). Since (cid:0)3+k Proof. For a general λ ∈ Γ, set E := Eλ. Up to a twist, i.e. taking E instead of E for some t ∈ Z, we may assume that m = 0. We have λ = (ε1, . . . , ε2r−1) with εi ∈ Ext1 ε1, . . . , ε2r−1 are linearly independent. 2 (cid:1) ≥ 2r − 1 and λ is general in Γ, the extensions Assume that E ∼= F1 ⊕ F2. Here we consider sheaves such as OH (t), E, F1 and F2 as OP3-sheaves, seeing OX as a quotient of OP3. From this point of view these sheaves are pure sheaves of depth 2 on P3 and we may apply the notion of (semi- )stability for pure sheaves; see [38]. Note that 0 ⊂ A ⊂ E is the Harder-Narasimhan filtration of E, because k > m = 0 and both of A and B = OH are semistable. By uniqueness of the Harder-Narasimhan filtrations of E and Fi for each i, A must be the direct sum of the first subsheaves of F1 and F2 in their filtrations. Then, due to rank counting, one of the two factors of E, say F1, is a factor of A. So we have A ∼= F1 ⊕ G for some G, while the other one F2, is isomorphic to either OH (the case of A = F1) or an extension of OH by G. First assume F2 6∼= OH , that is, G 6= 0. Each OH (t) is simple and there is an integer s ∈ {1, . . . , 2r − 2} such that F1 ∼= OH (k)⊕s and G ∼= OH (k)⊕(2r−1−s). Taking instead of F1 a direct factor of F1 with minimal rank, it is sufficient to consider only the case s = 1 for contradiction. Since OH (k) is simple, we have Aut(A) ∼= GL(2r − 1, k). Hence, up to an element of GL(2r − 1, k) we may as- sume that F1 is the first factor of A. With this new basis of OH (k)⊕(2r−1), set λ = (ε1, . . . , ε2r−1). Then ε1 corresponds to the extension of OH by OH (k) with E/j(G) ∼= OH (k) ⊕ OH as its middle term, where j : G → F2 → E is the com- position. Thus we get that ε1 is zero, contradicting to the linear independence of ε1, . . . , ε2r−1. 18 E. BALLICO, S. HUH, F. MALASPINA AND J. PONS-LLOPIS Now assume G = 0, that is, E ∼= OH ⊕ OH (k)⊕(2r−1). The extension class λ induces a surjection E → OH . Since k is positive, the extension class λ is induced by the projection of OH ⊕ OH (k)⊕(2r−1) onto its first factor. Thus we get λ = 0, contradicting Lemma 2.9. (cid:3) Lemma 5.4. Assume the same numeric invariants as in Proposition 5.3. For general λ, λ′ ∈ Γ, we have Eλ ∼= Eλ′ if and only if there is g ∈ GL(2r − 1, k) such that g · λ = λ′. := Eλ′ . Up to shift, E and E ′ are indecomposable Proof. Set E := Eλ and E ′ extensions of B ∼= OH by A ∼= OH (k)⊕(2r−1), where we have Aut(A) ∼= GL(2r−1, k) because OH is simple. Consider all these sheaves as pure sheaves of depth two on P3 and use semistability of pure sheaves with respect to the polarization OP3(1). Then A is semistable and the first step of the Harder-Narasimhan filtration of both E and E ′. Hence every isomorphism E → E ′ induces an automorphism of A. The other implication is obvious. (cid:3) Corollary 5.5. For fixed positive integer n and r ∈ (cid:0) 1 family ∆ of indecomposable layered aCM sheaves on X of rank r with dim ∆ ≥ n, where each isomorphism class of sheaves appears only finitely many times in ∆. 2(cid:1) N, r ≥ 1, there exists a Proof. By Proposition 5.3 and Lemma 5.4, a general choice of (2r − 1)-dimensional subspace in Ext1 P3(B, A) gives an indecomposable aCM sheaf of rank r on X. But the dimension of such choices can be made arbitrarily large by taking k sufficiently large compared to m, due to Lemma 2.9. To get isomorphism classes of sheaves we need to factor by the action of GL(2r − 1, k). We take a general orbit F of this action and choose a variety ∆′ intersecting F transversally and with complementary dimension, so that it intersects F at finitely many points (at least one). Then ∆′ intersects transversally and at finitely many point all fibers near F . We take as ∆ a non-empty Zariski open subset of ∆′ intersecting no orbit in a positive dimensional variety. (cid:3) Over k = C we may take instead of ∆ a small Euclidean ball of ∆ and get a one- to-one complex analytic parametrization by a ball in an affine space of dimension equal to dim ∆. 6. aCM sheaves of rank 3/2 Let us consider the case of an aCM sheaf E on the double plane X := X2[2] of rank of 3/2. We know that E fits in the short exact sequence (9) with L an aCM sheaf and A being one the possible subschemes from the list 4.1. Since we are interested only in indecomposable sheaves, by Remark 4.4, we can exclude cases (a) and (f). The following Lemma also allows us to exclude case (d): Lemma 6.1. There is no aCM sheaf of rank 3/2 on X fitting on the sequence (9): 0 −→ L −→ E π −→ IA(1) −→ 0 for A a connected scheme of degree two with Ared = {p} a point. Proof. L ∼= ker(π) is aCM by Lemma 4.3. So it is isomorphic to OH (a) for some a ∈ Z. From the exact sequence 0 = H 1(F (−1)) −→ H 1(IA) ∼= k −→ H 2(OH (a − 1)) −→ H 2(F (−1)) = 0, ACM SHEAVES ON THE DOUBLE PLANE 19 we get a = −2. Since F is aCM with h2(F (−2)) = 0, we get h1(IA(−1)) = h2(OH (−4)) = 3. But since A is a zero-dimensional subscheme of length two, we have h1(IA(−1)) = 2, a contradiction. (cid:3) It will also be easy to deal with case (c): Lemma 6.2. Any non-trivial rank 3/2 sheaf E on X fitting into the sequence (9) for A as in (c) of Possible Cases 4.1, is a layered sheaf. In particular, it also has a presentation as in (b) of Possible Cases 4.1. Proof. If E is such a sheaf, by Lemma 4.3, L is an aCM 0-regular sheaf of rank 1/2 and therefore L ∼= OH (a) with a ≥ 0. The central vertical exact sequence from the following diagram (17) 0 / G / E 0 0 / OH (a) / OH (a) IH (1) ∼= OH / IL(1) 0 / 0 OH IL,H (1) 0 shows that E also fits in (b) of Possible Cases 4.1 with G a rank one aCM sheaf. Thus we get the statement. (cid:3) Therefore, the rest of the section will be devoted to study cases (b) and (e). Lemma 6.3. We have Ext1 P3(OX , OH (a)) ∼= H 0(OH (a + 2)) for a ≥ 0. Proof. Applying [12, Lemma 13 in §4] to H ⊂ P3 with a pair (F , G) = (OX , OH (a)), we get Ext1 P3(OX , OH (a)) ∼= HomH (T orP3 1 (OX , OH ), OH (a)), because we have Ext1 3))∨ = 0. By tensoring (7) with OH , we get T orP3 OH (−2). Thus we get Ext1 H (OH , OH (a)) = 0 and Ext2 P3(OX , OH (a)) ∼= H 0(OH (a + 2)). H (OH , OH (a)) ∼= H 0(OH (−a − 1 (OX , OH ) ∼= T orP3 1 (OH (−1), OH) ∼= (cid:3) Remark 6.4. Although we have a non-trivial extension of OH (a) by OX as OP3- X (OX , OH (a)) ∼= sheaves for a ≥ 0, it is not an OX -sheaf, because we have Ext1 H 1(OH (a)) = 0 for all a ∈ Z by [24, Proposition III.6.3]. Proposition 6.5. If E is a simple layered aCM sheaf on X, then it is either OH (a) or OX (b) for some a, b ∈ Z. Proof. Let r ∈ (cid:0) 1 2(cid:1)Z be the rank of E. The result is trivial for r = 1/2. By Theorem 1.2 it is sufficient to prove that r ≤ 1 (since the ideal sheaf IC of a plane curve C ⊂ H is not simple as it can be easily deduced composing the maps from the exact sequence 16 with any non zero morphism OH (−d) −→ OH (−1)). So assume r > 1 and fix a filtration with 0 = E0 ⊂ E1 ⊂ · · · ⊂ E2r−1 ⊂ E2r = E of E with Ei/Ei−1 ∼= OH (ai) with ai ∈ Z for i = 1, . . . , 2r. We may assume that E is minimally regular and therefore there is a non-zero map u : E → OX (1). Set / / / / / / / / / / 20 E. BALLICO, S. HUH, F. MALASPINA AND J. PONS-LLOPIS IA(1) := Im(u). If a1 > 0, then composing the inclusion IA(1) ⊆ OX (1) with the surjection OX (1) → OH (1) and a non-zero map OH (1) → OH (a1) ⊂ E, would imply that E is not simple. Thus we get a1 ≤ 0. (a) Assume for the moment aj > 0 for some j ≥ 2 and define s to be the minimum among these integers. Claim: There is another layering filtration 0 = F0 ⊂ F1 ⊂ F2 ⊂ · · · F2r = E with either (i) F1 ∼= OH (as) or (ii) as = 1, aj = 0 for some j < s and F2 ∼= OX (1). Proof of Claim: We use induction on s. By assumption Es/Es−2 is an extension of OH (as) by OH (as−1). If this extension splits, then we may find another layering filtration 0 = G0 ⊂ G1 ⊂ · · · ⊂ G2r−1 ⊂ G2r = E of E such that • Gi = Ei if i /∈ {s − 1, s}, • Gs−1/Gs−2 ∼= OH (as), and • Gs/Gs−1 ∼= OH (as−1). If s = 2, then we may take Fi = Gi for all i. If s > 2, then we use the inductive hypothesis. In other words, we set Gs−1 the kernel of the map Es → OH (as−1) so that Gs−1/Es−2 ∼= OH (as). Actually, in this way only one sheaf in the filtration changes, namely in degree s − 1, while two maps change; the ones having source and target in degree s − 1. Now assume that Es/Es−2 is a non-trivial extension of OH (as) by OH (as−1). Since as > 0 and as−1 ≤ 0, Lemma 2.9 for X and Example 5.2 give as = 1, as−1 = 0 and Es/Es−2 ∼= OX (1). If s = 2, then Claim is proved. Now assume s > 2. Since aj ≤ 0 for all j < 0, we may apply s − 2 times the twist by −1 of Lemma 6.4 to get a new filtration Fi such that F2 ∼= OX (1), Fj = Ej for all j ≥ s, and Fi/Fi−1 ∼= OH (ai−2) for i = 2, . . . , s. By Claim we get either a non-zero map IA(1) → F1 or a non-zero map IA(1) → F2. So by composing with u we get that E is not simple, a contradiction. (b) Assume aj ≤ 0 for all j. Since E is 0-regular, it is globally generated. In particular, E/E2r−1 is globally generated, i.e. a2r ≥ 0. Our assumption gives If a1 ≥ 0, then we get a non-zero map E/E2r−1 → E1, which implies a2r = 0. that E is not simple. If a1 < 0, then we get h2(OH (a1 − 2)) > 0. This implies h2(E(−2)) > 0 since E/A is aCM. But it contradicts the 0-regularity of E. (cid:3) Lemma 6.6. For a plane curve C of degree d in X, we have • ext1 • ext1 P3(OH (a), IC ) = (cid:0)2−a X (OH (a), IC ) ≥ (cid:0)2−a 2 (cid:1) + max{2 − a − d, 0}; 2 (cid:1) −(cid:0)2−a−d (cid:1), 2 where (cid:0)n 2(cid:1) is zero for n ≤ 1. Proof. Recall that IC is an extension of OH (−d) by OH (−1) with d = deg(C). Applying the functor HomP3(OH (a), −) to the extension for IC , we get (18) 0 −→ HomP3(OH (a), OH (−d)) −→ Ext1 P3(OH (a), OH (−d)) −→ Ext2 −→ Ext1 P3(OH (a), OH (−1)), P3(OH (a), OH (−1)) −→ Ext1 P3(OH (a), IC ) because we have an isomorphism HomP3(OH (a), OH (−1)) ∼= HomP3(OH (a), IC ), i.e. each morphism OH (a) → IC factors through OH (−1). Indeed, we have an injec- tion HomP3(OH (a), OH (−1)) → HomP3(OH (a), IC ) and it gives homP3 (OH (a), IC ) ≥ (cid:0)1−a 2 (cid:1). On the other hand, applying the functor HomP3(OH (a), −) to (15), we get the opposite directional inequality, because we have HomP3(OH (a), OX ) ∼= HomX (OH (a), OX ) ∼= H 2(OH (a − 2))∨ ∼= H 0(OH (−1 − a)) ACM SHEAVES ON THE DOUBLE PLANE 21 by Serre's duality. If a ≥ 1, then we get a + d ≥ 2. This implies that Ext1 P3(OH (a), OH (−1)) ∼= P3(OH (a), OH (−d)) ∼= H 0(OH (1 − a − d)) are trivial by P3(OH (a), IC ) is trivial. Now assume a ≤ 0. By Serre's H 0(OH (−a)) and Ext1 Lemma 2.9. Thus Ext1 duality and Lemma 2.9, we get Ext2 P3(OH (a), OH (−1)) ∼= Ext1 P3(OH (−1), OH (a − 4))∨ ∼= H 0(OH (a − 2))∨, which is trivial. Thus the sequence (18) becomes (19) 0 −→ H 0(OH (−a − d)) −→ H 0(OH (−a)) −→ Ext1 P3(OH (a), IC ) −→ H 0(OH (1 − a − d)) −→ 0. If a + d ≥ 2, then we get H 0(OH (−a)) ∼= Ext1 (cid:0)2−a 2 (cid:1). dim Ext1 a + d ≤ 0, then each term in (19) is non-zero. Thus we get If a + d = 1, then similarly we get the dimension (cid:0)2−a 2 (cid:1) + 1. Finally if P3(OH (a), IC ) = h0(OH (−a)) + h0(OH (1 − a − d)) − h0(OH (−a − d)) P3 (OH (a), IC ), whose dimension is = (a2 − 5a − 2d + 6)/2 = (cid:18)2 − a 2 (cid:19) + (2 − a − d) and we get the assertion for Ext1 P3(OH (a), IC ). Now consider Ext1 assertion, due to case m = 1 of Lemma 2.9. X (OH (a), IC ). From (18) with P3 replaced by X, we get the (cid:3) Now recall that an extension of OH by OH is isomorphic to either O⊕2 H or IL(1) for a line L ⊂ H. We get that homX (OH , IL(1)) = homX (IL(1), OH ) = 1 for any line L ⊂ H, which is essentially equivalent to IL(1) 6∼= O⊕2 H by the proof of Lemma 6.7. Lemma 6.7. Let E be a sheaf of rank 3/2 with the filtration (20) such that Ei/Ei−1 ∼= OH for all i ∈ {1, 2, 3}. Setting eL := homX (OH , E) and eR := homX (E, OH ), we have the following. 0 = E0 ⊂ E1 ⊂ E2 ⊂ E3 = E (i) 1 ≤ eL, eR ≤ 3. (ii) eL = 3 ⇔ eR = 3 ⇔ E ∼= O⊕3 H . (iii) eL = 2 (resp. eR = 2) if and only if E is an extension of OH by O⊕2 H (resp. O⊕2 H by OH ). (iv) eL = eR = 1 if and only if (20) is the unique filtration of E with Ei/Ei−1 ∼= OH for all i. In this case E2 ∼= IL(1) for a line L ⊂ H uniquely determined by E. (v) eL = eR = 2 if and only if E ∼= IL(1) ⊕ OH for a line L ⊂ H. Proof. Certainly we have eL, eR ≥ 1. In the exact sequence (21) 0 −→ E2 −→ E −→ OH −→ 0, the sheaf E2 is aCM of rank one, admitting an extension of OH by OH . By the H or E2 ∼= IL(1) for a line L. classification of acM sheaf of rank one, we get E2 ∼= O⊕2 In particular, we have homX (OH , E2) ≤ 2 and the equality hold only if E2 ∼= O⊕2 H . Now apply the functor HomX (OH , −) to (21) to see that eL ≤ 3 and the equality hold if and only if E ∼= O⊕3 H . We also obtain similar assertion for eR, by applying the functor HomX (−, OH ). 22 E. BALLICO, S. HUH, F. MALASPINA AND J. PONS-LLOPIS H ֒→ E and a surjection u : E ։ O⊕2 If E ∼= IL(1) ⊕ OH for some line L ⊂ H, then we have eL = eR = 2, because homX (OH , IL(1)) = homX (IL(1), OH ) = 1 as mentioned in the paragraph before Lemma 6.7. Conversely assume eL = eR = 2, in particular there exist an inclusion j : O⊕2 H , otherwise the successive quotient Ei/Ei−1 would have a negative degree. Due to the rank counting, we have u ◦ j 6= 0 and this gives OH ⊂ j(O⊕2 H ) mapped isomorphically by u onto some OH ⊂ O⊕2 H . Hence OH is a factor of E and we get E ∼= IL(1) ⊕ OH for some line L ⊂ H. (cid:3) Remark 6.8. Note that in Lemma 6.7 E is a semistable sheaf on P3 of pure depth two with the same normalized Hilbert polynomial as OH . Since O⊕k H is polystable for any positive integer k, any nonzero map u = (u1, . . . , uk) : O⊕k H → E has the image isomorphic to O⊕c H , where c is the dimension of the linear span of ui's in Hom(OH , E). Replacing eL by Hom(O⊕k H , E), we get in the same way all statements for eL, except (v). H . Take another sheaf E ′ corresponding to (e′ Remark 6.9. Consider the case (iii) in Lemma 6.7 with (eL, eR) = (2, 1). Let W be the set of all (e1, e2) ∈ Ext1 X (OH , OH )⊕2 such that e1 and e2 are linearly In particular, W is an integral variety of dimension 6. Consider independent. a sheaf E of this type, corresponding to (e1, e2). In particular, E has a unique subsheaf E2 ∼= O⊕2 2), which is H and we have f (E2) = E ′ isomorphic to E. Then E ′ also has a subsheaf E ′ 2 2 for any isomorphism f : E → E ′. We get that two sheaves E and E ′ are isomorphic if and only if there is M ∈ GL(2, k) with M (e1, e2) = (e′ 2). So the isomorphism classes are parametrized by the orbits of this action of GL(2, k) on W , i.e. each X (OH , OH ) ∼= k3. So the family isomorphism class corresponds to a plane in Ext1 of the sheaves of this type is parametrized by P2. A similar description may be applied to the case with (eL, eR) = (1, 2). ∼= O⊕2 1, e′ 1, e′ The next goal will be to show the existence of unique non-layered aCM sheaf on X of rank 3/2 up to twist. Lemma 6.10. For a point p ∈ H ⊂ X, we have ext1 X (Ip(1), OH (−1)) = 1. Proof. We have the standard isomorphism Ext1 X (Ip(1), OH (−1)) ∼= Ext1 X (OH (−1), Ip(1) ⊗ ωX ))∨ ∼= Ext1 X (OH , Ip)∨. We can apply the functor HomX (OH , −) to the short exact sequence 0 −→ Ip −→ OX −→ Op −→ 0, to obtain the following strand of the associated long exact sequence: 0 ∼= HomX (OH , OX ) −→ HomX (OH , Op) ∼= k −→ Ext1 X (OH , Ip) −→ Ext1 X (OH , OX ) ∼= Ext1 X (OX , OH (−2))∨ ∼= 0, concluding the proof. (cid:3) Proposition 6.11. Let Ep be the unique non-trivial extension of Ip(1) by OH (−1). Then Ep is a non-layered Ulrich sheaf on X of rank 3/2. Moreover, for any other point q ∈ H, we have Ep ∼= Eq. Proof. Let us consider the unique non-trivial OX -sheaf given as an extension of the form (22) 0 −→ OH (−1) −→ Ep π −→ Ip(1) −→ 0. ACM SHEAVES ON THE DOUBLE PLANE 23 Claim 1 : Ep is Ulrich. Proof of Claim 1 : For any t ∈ Z, let δt : H 1(Ip(t + 1)) → H 2(OH (t − 1)) be the coboundary map of the twist by OX (t) of (22). From the injection 0 −→ H 1(Ep(t)) −→ H 1(Ip(t + 1)), we get h1(Ep(t)) = 0 for t ≥ −1 and h1(Ep(t)) ≤ 1 for all t ≤ −2. We have h1(Ip(−1)) = h2(OH (−3)) = 1 and the coboundary map δ−2 : H 1(Ip(−1)) −→ H 2(OH (−3)) corresponds to the non-trivial extension class [Ep]. Thus δ−2 is an isomorphism and we get h2(Ep(−2)) = 0. Assume that Ep is not aCM and let t0 be the largest integer such that h1(Ep(t0)) 6= 0. We just saw that t ≤ −3. Since δ−2 6= 0, we have δt0+1 6= 0. Take an equation ℓ of a plane different from H. The multiplication by ℓ induces a maps between the twist by OX (t0) and the twist by OX (t0 + 1) of (22). The induced map α : H 1(Ip(t0 + 1)) → H 1(Ip(t0 + 2)) is an isomorphism. Call η : H 2(OH (t0 − 1)) → H 2(OH (t0)) the map induced by the multiplication by ℓ. Since δt0+1 ◦ α = η ◦ δt0, δt0+1 6= 0 and α is an isomorphism, we have δt0 6= 0, a contradiction. Finally, the definition of Ep as an extension (22) gives that Ep has positive depth and that h0(Ep(−1)) = 0 and h0(Ep) = 3. Hence Ep is Ulrich. (cid:3) Note that we have homX (Ep, OX (1)) = h2(E(−3)) = 3 and this gives a two- dimensional projective space P := P HomX (Ep, OX (1)) of morphisms Ep → OX (1). If any of such maps is surjective, then its kernel would be isomorphic to OH (l) for some l ∈ Z and we would get a different Hilbert polynomial for Ep. Thus none of these maps are surjective. Now at least one of these maps is not surjective only at a point (namely, at p); in particular this is true for a non-empty subset of P, because the map P → Z sending a morphism to the dimension of its zeros is upper semicontinuous. Since we have dim Ext1 X (Ip,H (1), OX ) = 1 but homX (E, OX (1)) = 3, for each p ∈ H there is an open neighborhood Up of p such that for every q ∈ Up there is a surjection Ep → Iq(1). Thus we get Eq ∼= Ep for all q ∈ Up. Since any two non-empty open subsets of H meet, we get Eq ∼= Ep for all q ∈ H. We also see that for every v ∈ P there is q ∈ H such that Im(v) = Iq(1). So we get an identification P ∼= H. Claim 2 : Ep is non-layered. Proof of Claim 2 : Assume that E is layered, in particular by Corollary 7.4, we have a surjection u : Ep → OH . Composing with the inclusion OH → OX (1), we get a morphism v ∈ P such that Im(v) 6∼= Iq(1) for any q ∈ H, a contradiction. Thus Ep is not layered. (cid:3) Now Claim 1 and Claim 2 conclude the proof. (cid:3) Notation 6.12. Let us denote the unique non-layered minimally regular Ulrich sheaf of rank 3/2 on X in Proposition 6.11 by E. It has the same Hilbert polynomial as OH (−1) ⊕ Ip(1), or as O⊕3 H . Proposition 6.13. For any non-layered aCM sheaf F of rank 3/2 on X, E ∼= F (t) for some t ∈ Z. Proof. We may assume that F is minimally regular and then F fits in the short exact sequence 9: 0 −→ L −→ E π −→ IA(1) −→ 0, 24 E. BALLICO, S. HUH, F. MALASPINA AND J. PONS-LLOPIS for a closed subscheme A ( X in Possible Cases 4.1. By the previous discussions we need only to consider cases (b) and (e). Take A as in case (b), i.e. IA(1) ∼= OH . We know by Lemma 4.3 that L is aCM of rank one and therefore it is one of the cases described in Theorem 1.2 and F is layered. Finally assume case (e). In this case, ker(π) ∼= OH (a) for some a ∈ Z. If a ≤ −2, we get H 1(Ip) = 0 −→ H 2(OH (a − 1)) −→ H 2(F (−1)) = 0 with h2(OH (a−1)) > 0, a contradiction. If a ≥ 0, we get h1(F (−2)) ≥ h1(Ip(−1))− h2(OH (a − 2)) = 1, a contradiction. Now assume a = −1, in particular we get the exact sequence (23) Therefore F is the nontrivial extension from Proposition 6.11, namely F ∼= E. (cid:3) 0 −→ OH (−1) −→ F π −→ Ip(1) −→ 0. Proposition 6.14. E is a stable OX -sheaf with pure depth two. Proof. We already showed in Observation at the end of the proof of Proposition 6.11 that E has pure depth two. Moreover, by [23, Lemma 7.3], E is semistable, and if it was strictly semistable it would fit on an exact sequence of OX -sheaves (24) 0 −→ A u −→ E −→ B −→ 0 such that A is Ulrich, B is torsion-free and rank(A) + rank(B) = 3/2. (a) Assume that A has rank 1/2. Then, since it is Ulrich, A ∼= OH . Let π : E → Ip(1) be the surjection in (22). Since homX (OH , OH (−1)) = 0, the inclusion u in (24) induces an injective map OH → Ip(1). Since OH ∼= ker(fw,OX (1)), we have homX (OH , Ip(1)) = 1. Thus we get u(OH ) ⊂ π−1(OH ) and π−1(OH ) ∼= OH ⊕ OH (−1). We see that E/OH is an extension of Ip,H (1) by OH (−1). By [12, Lemma 13 in §4], we get the first isomorphism of the following Ext1 P3(Ip,H (1), OH (−1)) ∼= Ext1 H (Ip,H (1), OH (−1)) ∼= Ext1 X (Ip,H (1), OH (−1)), because we have HomH (Ip,H , OH(−1)) = 0. Then the second isomorphism can be induced automatically. Thus E/OH is isomorphic to either O⊕2 H or OH (−1) ⊕ Ip,H (1). In the former case, E would be layered, while the latter case is impossible, because E/OH must be globally generated. (b) Now assume B has rank 1/2, in particular we get B∨∨ ∼= OH (b) for some b ∈ Z by Lemma 5.1. Since the map B → B∨∨ is injective, we get B ∼= IZ,H (b) for some zero-dimensional subscheme Z ⊂ H or B ∼= OH (b). Since E is globally generated, B is also globally generated, in particular we get either b > 0 or b = 0 and Z = ∅. The latter case is excluded, because there is no surjection E → OH from the argument in the last three lines of the proof of Proposition 6.11. The former case is also excluded, because if b > 0, then the Hilbert polynomial of IZ,H (b) is strictly bigger than the one of OH . (cid:3) 2(cid:1)Z at least 3/2, there exists a non-layered Ulrich sheaf of rank r on X. Corollary 6.15. For each r ∈ (cid:0) 1 Proof. Fix r ∈ (cid:0) 1 2(cid:1)Z at least 3/2 and consider the sheaf G := E ⊕O⊕(2r−3) , for E the unique rank 3/2 non-layered sheaf from the previous remark. G is an Ulrich sheaf of rank r. We are going to show that G is non-layered. Otherwise, by Corollary 7.4, there exists a filtration 0 = G0 ⊂ G1 ⊂ · · · ⊂ G2r−1 ⊂ G2r = G of G with H ACM SHEAVES ON THE DOUBLE PLANE 25 Gi/Gi−1 ∼= OH for all i. Consider the composition u of the inclusion E ֒→ G with the surjection G → G/G2r−1 ∼= OH . In the end of the proof of Proposition 6.11, we proved that u cannot be a surjection. Since E is globally generated and h0(OH ) = 1, we get that u is a zero map and E ⊆ G2r−1. By descending induction on i, we get E ⊆ Gi for all i, a contradiction. (cid:3) Remark 6.16. In the rest of the section we will offer a description of indecom- posable layered aCM sheaves E on X of rank 3/2. By the previous discussions we know that such a sheaf E should fit in (25) 0 −→ L −→ E π −→ OH −→ 0, where L is aCM and 0-regular by Lemma 4.3 (case (b) of Possible Cases 4.1); by Theorem 1.2, L is isomorphic to either OX (a), with a ∈ Z; or IC (a) for a plane curve C ⊂ H with a ≥ d; or OH (a) ⊕ OH (b), with a ≥ b ≥ 0. The first case is excluded: X (OH , OX (t)) ∼= H 1(OH (−t − 2))∨ is trivial and E would be decomposable. In Ext1 Example 6.17 and Lemma 6.18 below, we describe the latter case. It turns out in the proof of Lemma 6.20 that such sheaves fall into the case (2-iii) of Theorem 1.4. Example 6.17. For fixed nonnegative integers a ≥ b ≥ 0, set d := a + b and let A(b, d) be the set of isomorphism classes of sheaves fitting in an exact sequence (26) 0 −→ OH (a) ⊕ OH (b) i −→ E π −→ OH −→ 0. A(b, d) is parametrized, not necessarily finite-to-one, by a vector space E(b, d) := Ext1 X (OH , OH (a)) ⊕ Ext1 X (OH , OH (b)) of dimension (cid:0)a+3 2 (cid:1) + (cid:0)b+3 2 (cid:1). Every element in A(b, d) is aCM, because it is an extension of aCM sheaves. Fix [E] ∈ A(b, d). Since the case (a, b) = (0, 0) is already described in Lemma 6.7 and Remark 6.9, we assume that a > 0; so we also get d > 0. The sequence (26) gives H 2(E(−2)) = 0 and H 2(E(−3)) 6= 0. In particular, E is minimally regular. X (OH , OH (b)). Suppose that E is induced by ε ∈ E(b, d) and write ε = (e1, e2) with e1 ∈ X (OH , OH (a)) and e2 ∈ Ext1 Ext1 If e1 = 0 (resp. e2 = 0), then OH (a) (resp. OH (b)) is a factor of E. Now assume e1 6= 0 and e2 6= 0. The extension e1 (resp. e2) induces a rank one aCM sheaf IC (a + 1) (resp. ID(b + 1)) for uniquely determined plane curves C, D ⊂ H with deg(C) = a + 1 and deg(D) = b + 1. Conversely, the curves C and D determine α and β, respectively, up to a constant, but the constants may be different; for two nonzero constants c and c′, we may consider the automorphism of OH (a) ⊕ OH (b) obtained by the multiplication by c in the first factor and by the multiplication by c′ in the second factor, to show that the sheaf induced by (ce1, c′e2) is isomorphic to E. Assume b > 0, and in particular OH (a) ⊕ OH (b) is uniquely determined by the Harder-Narasimhan filtration of E. Thus ε is uniquely determined by the isomor- phism class of E, up to isomorphisms of OH (a) ⊕ OH (b). If OH (k) is a factor of E with k > 0, then it is a factor of OH (a)⊕OH (b), because π(OH (k)) = 0 and there is a surjection OH (a)⊕OH (b) → OH (k) only if either k = a or k = b. If OH is a factor of E, then we get E ∼= OH ⊕OH(a)⊕OH(b), because homX (OH (a)⊕OH (b), OH ) = 0 and we have the surjection π. 26 E. BALLICO, S. HUH, F. MALASPINA AND J. PONS-LLOPIS Lemma 6.18. Let B(b, d) ⊂ A(b, d) consist of all indecomposable [E] ∈ A(b, d). Then A(b, d) \ B(b, d) is parametrized by at most d + 1 proper linear subspaces of E(b, d). in particular, B(b, d) is not empty. Proof. Fix [E] ∈ A(b, d) \ B(b, d). Since E has rank 3/2 and it is decomposable, it has a factor of rank 1/2. Let k be the minimum among the integers such that E has a factor OH (k), i.e. E ∼= OH (k) ⊕ F with F a rank one aCM sheaf. First assume k > 0. Since π(OH (k)) = 0, we get that F fits into an exact sequence 0 → OH (d − k) → F → OH → 0. 2 X (OH , OH (d − k)) ∼= H 0(OH (d − k + 1)), whose dimension is (cid:0)d−k+3 So F is isomorphic to either OH ⊕ OH (d − k) or IE(d − k + 1) for some plane curve E ⊂ H with deg(E) = d − k + 1. The former case is impossible due to the definition of k. In the latter case, the extension classes arising as F are parameterized by Ext1 k ≥ b, then the isomorphism classes arising as E are parameterized by a proper linear subspace of E(b, d). If k < b, then each map Hom(OH (a) ⊕ OH (b), E) has image contained in F . So we get an injective map OH (a) ⊕ OH (b) → IE(d − k + 1). The composition of this injective map with the surjection IE(d−k+1) → OH cannot be trivial, otherwise we would get an injection OH (a) ⊕ OH(b) → OH (d − k). Thus we have b = 0, i.e. (a, b) = (d, 0). In this case we see that the extension classes arising as F are parameterized by Ext1 such sheaves are parameterized by a proper linear subspace of E(0, d). (cid:1). So if X (OH , OH (d)) whose dimension is (cid:0)d+3 2 (cid:1). So Now assume k = 0. So we get E ∼= OH ⊕ F with either • F ∼= OH (i) ⊕ OH (d − i) for some integer i with 0 ≤ 2i ≤ d ,or • F ∼= ID(j) for some curve D ⊂ H with z := deg(D) = 2j − d − 1. Note that F cannot be a line bundle on X by the regularity condition of E. In the former case, we get i = 0; F ∼= OH ⊕ OH (d). This implies that E ∼= O⊕2 H ⊕ OH (d), which corresponds to the trivial element of E(0, d). Now assume the latter case. Since E is 0-regular, F is globally generated. In particular, we get j ≥ z. If j > z, then (16) gives homX (ID(j), OH ) = 0. So the factor OH of E induces the splitting of (26), in particular we get F ∼= OH (a) ⊕ OH (b), a contradiction. Now assume j = z, i.e. j = d+1. If b > 0, then the map i from the exact sequence (26) has image contained in F , while (26) implies that each map in Hom(OH (a) ⊕ OH (b), ID(j)) has image contained in OH (d). So i can not be injective, a contradiction. In case b = 0, i.e. (a, b) = (d, 0), we get a proper linear subspace of E(0, d) as above. (cid:3) Remark 6.19. Fix E ∈ B(b, d) for some integers b, d and assume b > 0. For any two positive integers a′ and b′, we have Hom(OH (a′) ⊕ OH (b′), OH ) = 0. So if {a, b} 6= {a′, b′}, there is no injective map OH (a′) ⊕ OH (b′) → E. In particular, we have B(b, d) ∩ B(b′, d′) = ∅, if (b, d) 6= (b′, d′). Lemma 6.20. Let E be a minimally regular indecomposable sheaf fitting into (25). If E is not as in Lemma 6.7, then there exists a plane curve C ⊂ H of degree d such that E fits into one of the following sequences: (1) 0 → IC (a) → E → OH → 0 for a ≥ d; (2) 0 → OH (b) → E → IC (d) → 0 for 0 ≤ b < d; Proof. By Remark 6.16, we can assume that L ∼= OH (a) ⊕ OH (b) with a ≥ b ≥ 0. The quotient sheaf E/OH (b) is an extension of OH by OH (a). So it is isomorphic ACM SHEAVES ON THE DOUBLE PLANE 27 to either OH (a) ⊕ OH or IC (a + 1) for a plane curve C ⊂ H of degree a + 1 by Example 5.2. The latter case falls into case (2). In the former case, let u : E → OH (a) ⊕ OH be the quotient map and set F := u−1(OH (a)). Then F is an extension of OH (a) by OH (b) and E is an extension of F by OH . By Example 5.2 F is isomorphic to either OH (a) ⊕ OH (b) or ID(b + 1) for a plane curve D of degree b−a+1. Note that the plane curve D makes sense only if b − a + 1 ≥ 0, i.e. b ∈ {a − 1, a}, and if b = a − 1, we have ID(2b + 2 − a) ∼= OX (a). Assume first that F ∼= OH (a) ⊕ OH (b). In particular, E is induced by (e1, e2) ∈ Ext1 X (OH (a), OH ) ⊕ Ext1 X (OH (b), OH ). From the indecomposability of E, we get a ≤ 1. If a = 1, then the sheaf corre- sponding to e1 is isomorphic to OX (1) and E is an extension of OX (1) by OH (b), which is trivial by Remark 6.4. So we have (a, b) = (0, 0) and we already describe the case in Lemma 6.7 and Remark 6.9. Now assume that F ∼= ID(b + 1), where D is either empty or a line. If b = a − 1, then we get F ∼= OX (a). In this case E is decomposable by Remark 6.4. So we may assume that F ∼= IL(a + 1). In particular, E is an extension of IL(a + 1) by OH . Apply the functor HomX (−, OH ) to the exact sequence 0 −→ OH (a) −→ IL(a + 1) −→ OH (a) −→ 0, we get Ext1 X (IL(a+1), OH ) = 0 for a ≥ 2 by Example 5.2. The case a = 0 falls into case (2) with (b, d) = (0, 1), while in the case a = 1 we have homX (F , OH ) = 0. So the map π gives the splitting of (25). (cid:3) Proof of Theorem 1.4: If E is not layered, then we may use Propositions 6.11 and If E is decomposable, then we use Theorem 1.2. Finally when E is inde- 6.14. composable and layered, we may use the proof of Proposition 6.11 and Lemma 6.20. (cid:3) From the proof of Lemma 6.20, the case (2) of Lemma 6.20 occurs when E fits into the exact sequence (26), which is already described in Example 6.17 and Lemma 6.18. as in (1) of Lemma 6.20. Indeed, if a ≥ d, then we get dim Ext1 Now assume ker(π) ∼= IC (a) for some degree d curve C and a unique integer a X (OH , IC (a)) ≥ (cid:0)2+a 2 (cid:1) −(cid:0)2+a−d (cid:1), which is positive. Lemma 6.21. Any non-trivial extension of OH by IC (a) with a ≥ d is indecom- posable. 2 Proof. Let ε be the non-trivial extension class of OH by IC (a) corresponding to E and assume that E is decomposable with E ∼= OH (b) ⊕ G for some aCM sheaf G of rank 1, where G is described in Theorem 1.2. Since E is globally generated, G is also globally generated and b ≥ 0. (i) First assume b > 0, in particular we have HomX (OH (b), OH ) = 0 with a surjective map G → OH . Since G is globally generated, by Theorem 1.2, G is H or IL(1) for some line L ⊂ H. If G ∼= OX , then we isomorphic to either OX , O⊕2 get (27) 1 +(cid:18)a + 1 2 (cid:19) +(cid:18)a − d + 2 2 (cid:19) = h0(E) = 1 +(cid:18)b + 2 2 (cid:19). 28 E. BALLICO, S. HUH, F. MALASPINA AND J. PONS-LLOPIS Since HomX (OH (b), OH ) = 0, we get an injective map j : OH (b) → IC (d) with the cokernel isomorphic to OH (−1), which is the kernel of the surjection G → OH . Since IC (a) is an extension of OH (a − d) by OH (a − 1), we get b = a − 1, which is impossible by (27). Thus we get either G ∼= O⊕2 1 +(cid:18)a + 1 (28) 2 H or G ∼= IL(1) for some line L ⊂ H with the equality 2 (cid:19) +(cid:18)a − d + 2 (cid:19) = (cid:18)b + 2 2 (cid:19) + 2 as above. The nonzero map j : OH (b) → IC (d) gives that either b = a − 1 or b ≤ a − d, and this implies from (28) that a = 1. Then we get d = 1 and b = 0, a contradiction. (ii) Now assume b = 0, in particular the map E → OH induces the zero-map OH → OH , because ε 6= 0 and any non-zero map OH → OH is an isomorphism. Thus we get a surjective map G → OH . So as before G is isomorphic to either OX , H or IL(1) for some line L ⊂ H. If G ∼= OX , then h0(E) = 2 and h0(IC (a)) = 1, O⊕2 H or G ∼= IL(1), then we get h0(E) = 3 and a contradiction. h0(IC (a)) = 2, i.e. a = d = 1. Thus we are in the set-up of Lemma 6.7, and it is indecomposable. (cid:3) If either G ∼= O⊕2 The sheaf IC (a) is an extension of OH (a − d) by OH (a − 1). First assume a ≥ d + 1. Since ωC ∼= OC (d − 3), the exact sequence 0 −→ IC −→ OX −→ OC −→ 0 gives h2(IC (a − 3)) = 0. Therefore h2(E(−3)) = 1 and by Serre's duality there is a unique surjection E → OH , up to a scalar, i.e. a unique subsheaf isomorphic to IC (a) for some integer a and some curve C. Hence if f : E → E ′ is an isomorphism, then E ′ contains IC (a) and f (IC (a)) = IC (a). We get that the isomorphism classes of E are parametrized by the quotient of the family of the non-zero extensions of OH by IC (a), by the action of Aut(IC (a)). Note that by Proposition 6.22 below we have dim Aut(IC (a)) = 1 + (cid:0)d+1 2 (cid:1). In other words, we have the action of the algebraic group Aut(IC (a)) on the quasi-affine integral variety Ext1 X (OH , IC (a)) \ {0} so that the isomorphism classes of these sheaves are parametrized by the orbits of the algebraic group Aut(IC (a)). Proposition 6.22. For a plane curve C ⊂ H of degree d, the group Aut(IC ) is identified with the group of matrices 0 a−1(cid:19) (cid:12)(cid:12)(cid:12)(cid:12) (cid:26)(cid:18) a f a ∈ k \ {0} , f ∈ k[x, y, z]d−1(cid:27) . In particular, dim End(IC ) = 1 +(cid:0)d+1 2 (cid:1). Proof. Automorphisms of IC extend to automorphisms of its OP3-resolution 0 −→ OP3(−2) ⊕ OP3(−d − 1) A −→ OP3 (−1) ⊕ OP3(−d) −→ IC −→ 0, namely, to pairs of matrices (M, N ) ∈ End(OP3 (−2) ⊕ OP3(−d − 1)) ⊕ End(OP3(−1) ⊕ OP3 (−d)) such that N −1AM = A. Developing this equation, the pairs correspond to matrices as on the statement. Now, since Aut(IC ) is a non-empty open subset of End(IC ),we get the second part of the statement as well. (cid:3) ACM SHEAVES ON THE DOUBLE PLANE 29 There remains the case a = d. If h2(E(−3)) = 1, then again E fits in a unique extension of OH by IC (d). So we have the same description as in the case a ≥ d+ 1. Now assume h2(E(−3)) > 1. If a = d = 1, then we are in the set up of Lemma 6.7 with eR ≥ 2. Now assume a = d > 1. Note that E contains a unique copy of OH (d − 1) with F := E/OH (d − 1), where F is isomorphic to O⊕2 H or IL(1) for some line L ⊂ H. Thus any isomorphism f : E → E ′ sends OH (d − 1) to the corresponding copy of OH (d − 1) of E ′. So f induces an isomorphism of extensions of F by OH (d − 1). We get that the isomorphism classes of E are parametrized by the quotient of the family of non-zero extensions of F by OH (d − 1), by the action of Aut(F ). See Lemma 6.22 for the computation of Aut(F ) in the case F ∼= IL(1). In all cases E is decomposable only if either is the trivial one, or • E ∼= OH (d−1)⊕F , which is possible only if the extension of F by OH (d−1) • E ∼= OH ⊕ G for some aCM sheaf G that is 0-regular with h0(G) = (cid:0)d+1 If the inclusion j : OH → E induced from an OH -factor of E, does not split the extension E of OH by IC (d), then the surjection E → OH maps j(OH ) onto zero, in particular j(OH ) is a saturated subsheaf of IC (d), i.e. the quotient is torsion-free. Its existence would imply that IC (d) ∼= OH ⊕ OH (d − 1), contradicting the fact that IC (d) is not an OH -sheaf. 2 (cid:1) + 1. Remark 6.23. Among the aCM sheaves of rank 3/2 studied in this section, only the sheaves of the form OH (a) ⊕ OH (b) ⊕ OH (c) are OH -sheaves; for any other sheaf E there is no closed proper subscheme Y ⊂ X such that E is an OY -sheaf. 7. Ulrich sheaves In this section, we discuss the (non)-existence of Ulrich sheaves on X. Recall that ∆ is the collection of aCM vector bundles, admitting an extension of SH by SH (−1); see Example 4.12. Lemma 7.1. For every [E] ∈ ∆, its twist E(1) is Ulrich. Proof. By definition of ∆, the vector bundle E is aCM with EH OH (−1). So it fits into the exact sequence ∼= Ω1 H (1) ⊕ OH ⊕ 0 −→ Ω1 H ⊕ OH (−1) ⊕ OH (−2) −→ E −→ Ω1 (29) In particular, we have h0(E) ≤ 1 and H 0(E) is the kernel of the coboundary map δ : H 0(OX ) → H 1(Ω1 H ). The latter cohomology group is one-dimensional and δ must be an isomorphism, because H 1(E) is trivial. Again by (29) we get H(1) ⊕ OH ⊕ OH (−1) −→ 0. h0(E(1)) = h0(OH ) + h0(OH ) + h0(OH (1)) + h0(Ω1 H (2)) = 1 + 1 + 3 + 3 = 8. Since E(1) is an initialized vector bundle of rank four on X with degree two, it is Ulrich. (cid:3) Lemma 7.2. For each [E] ∈ ∆, we have 4 ≤ dim End(E) ≤ 9. Proof. From the Euler sequence 0 −→ Ω1 H −→ OH (−1)⊕3 −→ OH −→ 0, we get h1(Ω1 A := O⊕2 H ⊗ (Ω1 H )∨(−1)) = h0((Ω1 H )∨(−1)) = 3. Set H ⊕ OH (−1) ⊕ Ω1 H ⊗ (Ω1 H )∨ ⊕ Ω1 H (1) ⊕ Ω1 H (2) ⊕ (Ω1 H )∨(−1) ⊕ (Ω1 H )∨(−2). 30 E. BALLICO, S. HUH, F. MALASPINA AND J. PONS-LLOPIS We have h0(A(−1)) = 0, h0(A) = 9 and h1(A(−1)) = 2+h1(Ω1 Since End(E) is locally free, we have an exact sequence H ⊗(Ω1 H)∨(−1)) = 5. (30) 0 −→ A(−1) −→ End(E) −→ A −→ 0, in particular we get 4 ≤ dim End(E) ≤ 9. (cid:3) Proposition 7.3. Let E be a layered sheaf of rank r ≥ 1 on X with the filtration in Definition 2.8. Set ind(E) := max{t ∈ Z H 0(E(−t)) 6= 0}. Then we have (i) ind(E) = max{a1, . . . , a2r}; (ii) h0(E(−ind(E))) = ♯{i ∈ {1, . . . , 2r} ai = ind(E)}. Proof. If we let ρ = max{a1, . . . , a2r}, then the filtration of E gives H 0(E(−t)) = 0 if t > ai for all i. So we get ind(E) ≤ ρ. If ρ = a1, then we get H 0(E(−a1)) 6= 0 and ind(E) ≥ a1. Similarly, if ρ = ai for i ≥ 2, we get H 0((E/Ei−1)(−ρ)) 6= 0. Since Ei−1 is aCM, we get H 0(E(−ρ)) 6= 0 and ind(E) ≥ ρ. (cid:3) Corollary 7.4. If E is a layered Ulrich sheaf of rank r with the filtration in Defi- nition 2.8, then we have ai = 0 for all i. Proof. By Lemma 7.3, we have 0 = ind(E) = max{a1, . . . , a2r} and ai = 0 for 2r indices i, concluding the proof. (cid:3) Proposition 7.5. There is no layered Ulrich vector bundle on X. Proof. Let us suppose that there exists an Ulrich vector bundle E of rank r on X with filtration: (31) for i = 2, . . . , 2r with E1 ∼= OH and E2r ∼= E. 0 −→ Ei−1 −→ Ei −→ OH −→ 0, Claim 1: H j(EiH (−1)) = 0 for i = 1, . . . , 2r and j ∈ {0, 1, 2}. Proof of Claim 1: We are going to prove the claim by induction on i; notice that the claim is true for i = 1. On the other hand, tensoring the short exact sequence (7) by the OX -sheaf Ei we obtain 0 −→ T or1 X (Ei, OH ) −→ EiH (−1) −→ Ei −→ EiH −→ 0, X (OH , OH ) ∼= OH (−1) for i = 1 and dim Supp T or1 so we can deduce T or1 1 for i ≥ 2. Tensoring the short exact sequence (31) by OH (−1), we get X (Ei, OH ) ≤ T or1 X (Ei, OH )(−1) ϕ3 −→ OH (−2) ϕ2 −→ Ei−1H (−1) ϕ1 −→ EiH (−1) −→ OH (−1) −→ 0. Splitting the previous exact sequence into short ones, we can see first that the surjection H 2(Ei−1H (−1)) ։ H 2(Im ϕ1) ∼= H 2(EiH (−1)) gives us H 2(EiH (−1)) = 0 by the induction's hypothesis. Next H 1(Ei−1H (−1)) ∼= H 1(Im ϕ1) ∼= H 2(Im ϕ2) and the last group is zero due to the existence of the surjection 0 = H 2(OH (−2)) ։ H 2(Im ϕ2). Finally, we have the chain of equalities H 0(EiH (−1)) = H 0(Im ϕ1) = H 1(Im ϕ2) = H 2(Im ϕ3) and again the last cohomology is zero due to the surjection H 2(T or1 H 2(Im ϕ3) and that H 2(T or1 at most one-dimensional. X (Ei, OH )(−1)) = 0 since the support of this sheaf is (cid:3) X (Ei, OH )(−1)) ։ ACM SHEAVES ON THE DOUBLE PLANE 31 H ) = 0. Claim 2: H 0(E ∨ Proof of Claim 2: After tensoring the short exact sequence (7) by E ∨ (notice that, since E is a vector bundle, the operations of dualizing and restricting to H do commute), we get (32) 0 −→ E ∨(−1)H −→ E ∨ −→ E ∨ H −→ 0. Since h0(E ∨) = h1(E ∨) = h2(E ∨(−1)) = 0 by Serre duality and the fact of E being Ulrich, we deduce from the long exact sequence of cohomology groups associated to (32) H 0(E ∨ H ) ∼= H 1(E ∨ H (−1)) ∼= H 2(E ∨ H (−2)) ∼= H 0(EH (−1)) = 0, by Claim 1, where the last isomorphism is obtained applying Serre duality on H. This concludes the proof of Claim 2. (cid:3) Finally, after tensoring (31) by E ∨ for any i = 2, . . . , 2r and using Claim 1, we would obtain h0(E ⊗ E ∨) = · · · = h0(Ei ⊗ E ∨) = · · · = h0(E ∨ H ) = 0, a contradiction. (cid:3) Proposition 7.6. For any [E] ∈ ∆ in Example 4.12, E is not layered. Proof. Since E(1) is Ulrich for each [E] ∈ ∆, the result is immediate from Proposi- tion 7.5. (cid:3) Theorem 7.7. For each r ∈ (cid:0) 1 sheaf with rank r. 2(cid:1)Z>0, there exists a layered indecomposable Ulrich Proof. By Corollary 7.4 we need only to check the indecomposability of some layered sheaf E with a filtration 0 = E0 ⊂ · · · ⊂ E2r = E with Ei/Ei−1 ∼= OH for all i. The case r = 1/2 is trivial. Note that the assertion is also true for r ∈ (cid:8)1, 3 2(cid:9); for r = 1, up to isomorphism it only gives the indecomposable sheaf IL(1) with L ⊂ H a line, and the case r = 3 2 comes as a particular case of Lemma 6.21 with (a, d) = (1, 1). Set E1 := OH . For each r ∈ (cid:0) 1 2(cid:1)Z≥1, let E2r be the middle term of a general OP3-extension (33) 0 −→ E2r−1 −→ E2r u −→ OH −→ 0. By its inductive definition each E2r is an OX -Ulrich sheaf with a filtration 0 = E0 ⊂ · · · ⊂ E2t with Ei/Ei−1 ∼= OH for all i. We will show that E2r is indecomposable. P3(OH , E2r−1) = 0 for all r ∈ (cid:0) 1 Claim 1: ext2 Proof of Claim 1: for r = 1, it is well-known that ext2 ing the functor HomX (OH , −) to (33) we obtain the strain P3(OH , E2r−1) −→ Ext2 P3(OH , E2r−1) −→ Ext2 2(cid:1)Z≥1. Ext2 P3(OH , OH ). P3(OH , OH ) = 0. Apply- We get the Claim 1 by induction on r ∈ (cid:0) 1 2(cid:1)Z≥1. P3(OH , E2r−1) 6= 0 for all r ∈ (cid:0) 1 Claim 2: ext1 Proof of Claim 2: Claim is true for r = 1 by case m = 1 of Lemma 2.9. So we assume r > 1 and apply the functor HomP3 (OH , −) to (33) to get the exact sequence 2(cid:1)Z≥1. (cid:3) Ext1 P3(OH , E2r) −→ Ext1 Then we can conclude by Claim 1. P3(OH , OH ) −→ Ext2 P3(OH , E2r−1) = 0. (cid:3) 32 E. BALLICO, S. HUH, F. MALASPINA AND J. PONS-LLOPIS Claim 3: For each positive r ∈ (cid:0) 1 2(cid:1)Z, homP3(OH , E2r) = 1. Proof of Claim 3: We use induction on the integer 2r, the case 2r = 1 being H 6∼= IL(1) for any line L ⊂ H. obvious and the case 2r = 2 being true, because O⊕2 Now assume 2r ≥ 3 and that homP3(OH , E2r) > 1. Since homP3(OH , E2r−1) = 1 by the inductive assumption, there is an inclusion j : OH → E2r such that u ◦ j : OH → OH is the identity map. Hence (33) splits, contradicting Claim 2. (cid:3) To conclude the proof of Theorem 7.7 it is sufficient to prove that E2r is in- decomposable. Assume E2r ∼= F1 ⊕ F2 with each Fi nontrivial. Note that each Fi is aCM and initialized with h0(E2r) = h0(F1) + h0(F2). So each Fi is Ul- rich and by Corollary 7.4 it has a filtration starting with OH . Thus we get homX (OH , E2r) = homX (OH , F1)+homX (OH , F2) ≥ 2, contradicting Claim 3. (cid:3) References 1. V. Ancona and G. Ottaviani, Some applications of Beilinson theorem to projective spaces and quadrics, Forum Math. 3 (1991), 157–176. 2. M. F. Atiyah, On the Krull-Schmidt theorem with application to sheaves, Bull. S,M.F., tome 84 (1956), 307–317. 3. M. F. Atiyah, Vector bundles over an elliptic curve. Proc. London Math. Soc. (3) 7 (1957), 414–452. 4. E. Ballico, S. Huh and J. Pons-Llopis, aCM sheaves of pure rank two on reducible hyper- quadrics, to appear in J. Algebra, preprint, arXiv:1508.01632v2 [math.AG]. 5. A. Beauville, Determinantal hypersurfaces, Michigan Math. J. 48 (2000), 39–64. 6. A. Beauville, An introduction to Ulrich bundles, preprint, arXiv:1610.02771v2 [math.AG]. 7. R. Buchweitz, G. Greuel and F. Schreyer, Cohen-Macaulay modules on hypersurface singu- larities II, Invent. Math. 88 (1987) 165–182. 8. W. Bruns and J. Herzog, Cohen-Macaulay rings, Cambridge University Press, vol. 39, 1993. 9. M. Casanellas and R. Hartshorne, Gorenstein biliaison and ACM sheaves, J. Algebra 278 (2004), 314–341. 10. M. Casanellas and R. Hartshorne, ACM bundles on cubic surfaces, J. Eur. Math. Soc. (JEMS) 13 (2011), no. 3, 709–731. 11. F. Catanese, Footnotes to a theorem of Reider, in: Algebraic Geometry Proceedings, L'Aquila 1988 (ed. by A. J. Sommese, A. Biancofiore, E. L. Livorni), 64–74, Lecture Notes in Math. 1417, Springer, Berlin, 1990 12. J. Choi, K. Chung and M. Maican, Moduli of sheaves supported on quartic space curves, Michigan Math. J. 29 (2016), no. 3, 637–671. 13. L. Costa and R. M. Mir´o-Roig and J. Pons-Llopis, The representation type of Segre varieties, Adv. Math. 230 (2012), 1995–2013. 14. L. Costa and R. M. Mir´o-Roig, GL(V )-invariant Ulrich bundles on Grassmannians, Math. Ann. 361 (2015), no. 1, 443-457. 15. L. Costa and R. M. Mir´o-Roig, Homogeneous ACM bundles on a Grassmannian, Adv. Math. 289 (2016), 95–113. 16. O. Debarre and L. Manivel, Sur l`a vari´et´e des espaces lin´eaire contenus dans une intersection compl`ete, Math. Ann. 312 (1998), no. 3, 549–574. 17. D. Eisenbud, Homological algebra on a complete intersection, with an application to group representation, Trans. Amer. Math. Soc. 260 (1980), 35–64. 18. D. Eisenbud and J. Herzog, The classification of homogeneous Cohen-Macaulay rings of finite representation type, Math. Ann. 280 (1988), no. 2, 347–352. 19. D. Eisenbud and F.O. Schreyer,Boij-Soderberg theory, Combinatorial aspects of commutative algebra and algebraic geometry, 6 (2011), 35–48. 20. D. Eisenbud, F.O. Schreyer and J. Weyman, Resultants and Chow forms via exterior syzygies, J. Amer. Math. Soc. 16 (2003),no. 2, 537–579. 21. D. Faenzi, Rank 2 arithmetically Cohen-Macaulay bundles on a nonsingular cubic surface. J. Algebra 319 (2008), no. 1, 143–186. 22. D. Faenzi and F. Malaspina, Surfaces of minimal degree of tame representation type and mutations of Cohen-Macaulay modules, Adv. Math., vol. 310 (2017), 663–695. ACM SHEAVES ON THE DOUBLE PLANE 33 23. D. Faenzi and J. Pons-Llopis, The CM representation type of projective varieties, preprint, arXiv:1504.03819 [math.AG]. 24. R. Hartshorne, Algebraic geometry, Springer-Verlag, New York-Heidelberg, 1977, Graduate Texts in Mathematics, No. 52. 25. R. Hartshorne, Stable vector bundles of rank 2 on P3, Math. Ann. 238 (1978), 229–280. 26. R. Hartshorne, Generalized divisors on Gorenstein schemes, K-Theory 8 (1994), 287–339. 27. D. Huybrechts and M. Lehn, The geometry of moduli spaces of sheaves. Second edition, Cambridge Mathematical Library, Cambridge U.P. (2010). 28. M. M. Kapranov, On the derived categories of coherent sheaves on some homogeneous spaces, Invent. Math. 92 (1988), no. 3, 479–508. 29. H. Knorrer, Cohen-Macaulay modules on hypersurface singularities. I, Invent. Math. 88 (1987), no. 1, 153–164. 30. N. Mohan Kumar and A. P. Rao and G. V. Ravindra, Arithmetically Cohen-Macaulay bundles on hypersurfaces, Comment. Math. Helv. 82 (2007), no. 4, 829–843. 31. N. Mohan Kumar and A. P. Rao and G. V. Ravindra, Arithmetically Cohen-Macaulay bundles on three dimensional hypersurfaces, Int. Math. Res. Not. (2007), no. 8, Art. ID rnm025, 11 pp. 32. C. G. Madonna, ACM vector bundles on prime Fano threefolds and complete intersection Calabi-Yau threefolds, Rev. Roumaine Math. Pures Appl. 47 (2002), no. 2, 211–222. 33. C. G. Madonna, A splitting criterion for rank 2 vector bundles on hypersurfaces in P4, Rend. Sem. Mat. Univ. Politec. Torino 56 (1998), no. 2, 43–54. 34. C. Okonek and M. Schneider and H. Spindler, Vector Bundles on Complex Projective Spaces, Modern Birkhauser Classics. Birkhauser/Springer Basel AG, Basel, 2011. 35. G. Ottaviani, Spinor bundles on quadrics, Trans. Amer. Math. Soc. 307 (1988), 301–316. 36. G. Ottaviani, Some extension of Horrocks criterion to vector bundles on Grassmannians and quadrics, Ann. Mat. Pura Appl. 155 (1989), no. 4, 317–341. 37. J. P. Serre, Faisceaux Algebriques Coherents, Ann. Math. 61 (1955), no. 2, 197–278. 38. C. T. Simpson, Moduli of representations of the fundamental group of a smooth projective variety. I, Inst. Hautes ´Etudes Sci. Publ. Math. (1994), no. 79, 47–129. 39. A. Tripathi,Rank 3 arithmetically Cohen-Macaulay bundles over hypersurfaces, J. Alge- bra,478 (2017), 1–11. 40. C. A. Weibel, An introduction to homological algebra, Cambridge University Press, vol. 38, 1995. Universit`a di Trento, 38123 Povo (TN), Italy E-mail address: [email protected] Sungkyunkwan University, Suwon 440-746, Korea E-mail address: [email protected] Politecnico di Torino, Corso Duca degli Abruzzi 24, 10129 Torino, Italy E-mail address: [email protected] Department of Mathematics, Kyoto University, Kyoto, Japan E-mail address: [email protected]
1511.08660	2	1511	2016-04-27T08:07:55	$\mu$-constant monodromy groups and Torelli results for marked singularities, for the unimodal and some bimodal singularities	[ "math.AG" ]	This paper is a sequel to [He7]. There a notion of marking of isolated hypersurface singularities was defined, and a moduli space $M_\mu^{mar}$ for marked singularities in one $\mu$-homotopy class of isolated hypersurface singularities was established. One can consider it as a global $\mu$-constant stratum or as a Teichm\"uller space for singularities. It comes together with a $\mu$-constant monodromy group $G^{mar}\subset G_Z$. Here $G_Z$ is the group of automorphisms of a Milnor lattice which respect the Seifert form. It was conjectured that $M_\mu^{mar}$ is connected. This is equivalent to $G^{mar}= G_Z$. Also Torelli type conjectures were formulated. All conjectures were proved for the simple singularities and 22 of the exceptional unimodal and bimodal singularities. In this paper the conjectures are proved for the remaining unimodal singularities and the remaining exceptional bimodal singularities.	math.AG	math	µ-CONSTANT MONODROMY GROUPS AND TORELLI RESULTS FOR MARKED SINGULARITIES, FOR THE UNIMODAL AND SOME BIMODAL SINGULARITIES FALKO GAUSS AND CLAUS HERTLING µ Abstract. This paper is a sequel to [He7]. There a notion of marking of isolated hypersurface singularities was defined, and a moduli space M mar for marked singularities in one µ-homotopy class of isolated hypersurface singularities was established. One can consider it as a global µ-constant stratum or as a Teichmuller space for singularities. It comes together with a µ-constant mon- odromy group Gmar ⊂ GZ. Here GZ is the group of automorphisms of a Milnor lattice which respect the Seifert form. It was conjec- is connected. This is equivalent to Gmar = GZ. tured that M mar Also Torelli type conjectures were formulated. All conjectures were proved for the simple singularities and 22 of the exceptional uni- modal and bimodal singularities. In this paper the conjectures are proved for the remaining unimodal singularities and the remaining exceptional bimodal singularities. µ Contents Introduction 1. 2. Review on the topology of isolated hypersurface singularities 3. The group GZ for the simple elliptic and the hyperbolic singularities 4. The group GZ for 6 of the 28 exceptional unimodal and bimodal singularities 5. Review on µ-constant monodromy groups Gmar, marked , and Torelli type singularities, their moduli spaces M mar conjectures 6. Gmar, M mar and a strong Torelli result for the simple elliptic µ µ and the hyperbolic singularities 2 4 9 15 21 28 Date: April 11, 2016. 2000 Mathematics Subject Classification. 32S15, 32S40, 14D22, 58K70. Key words and phrases. µ-constant monodromy group, marked singularity, mod- uli space, Torelli type problem, simple elliptic singularities, hyperbolic singularities. This work was supported by the DFG grant He2287/4-1 (SISYPH). 1 2 FALKO GAUSS AND CLAUS HERTLING 7. Gmar, M mar µ and a strong Torelli result for 6 of the 28 exceptional unimodal and bimodal singularities 8. More on GZ for the simple elliptic singularities References 31 34 38 1. Introduction This paper is a sequel to [He7]. That paper studied local objects, namely holomorphic functions germs f : (Cn+1, 0) → (C, 0) with an isolated singularity at 0 (short: singularity), from a global perspective. There a notion of marking of a singularity was defined. One has to fix one singularity f0, which serves as reference singularity. Then a marked singularity is a pair (f,±ρ) where f is in the µ-homotopy class of f0 and ρ : (M l(f ), L) → (M l(f0), L) is an isomorphism. Here M l(f ) is the Milnor lattice of f , and L is the Seifert form on it (the definitions are recalled in section 2). The group GZ(f0) := Aut(M l(f0), L) will be important, too. µ A moduli space M mar of right equivalence classes of marked singu- larities in one µ-homotopy class was established. One can consider it as a global µ-constant stratum or as a Teichmuller space for singularities. The group GZ(f0) acts properly discontinuously on it, and the quo- tient is the moduli space Mµ of right equivalence classes of unmarked singularities from [He6, chapter 13]. The µ-constant monodromy group Gmar(f0) ⊂ GZ(f0) is the subgroup of automorphisms which map the (f0))0 which contains [(f0,± id)] to itself. topological component (M mar It can also be constructed as the group of all automorphisms which can be realized modulo ± id as transversal monodromies of µ-constant fam- ilies. µ Conjecture 1.1. [He7, conjecture 3.2 (a)] M mar alently: Gmar = GZ. µ is connected. Equiv- Roughly this conjecture says that all abstract automorphisms come from geometry, from coordinate changes and µ-constant families. Also Torelli type conjectures are formulated in [He7]. Any singu- larity f comes equipped with its Brieskorn lattice H(cid:48)(cid:48) 0 (f ), and any marked singularity (f,±ρ) comes equipped with a marked Brieskorn lattice BL(f,±ρ). The Gauss-Manin connection for singularities and the Brieskorn lattices had been introduced 1970 by Brieskorn and had been studied since then. The second author has a long-going project on Torelli type conjectures around them. µ-CONSTANT MONODROMY GROUPS FOR SOME SINGULARITIES 3 In [He4] a classifying space DBL(f0) for marked Brieskorn lattices was constructed. It is especially a complex manifold, and GZ(f0) acts properly discontinuously on it. The quotient DBL/GZ is a space of iso- morphism classes of Brieskorn lattices. There is a natural holomorphic period map BL : M mar µ (f0) → DBL(f0). It is an immersion [He6, theorem 12.8] (this refines a weaker result in [SaM]). And it is GZ-equivariant. Conjecture 1.2. (a) [He7, conjecture 5.3] BL is an embedding. (b) [He7, conjecture 5.4], [He6, conjecture 12.7], [He1, Kap. 2 d)]) LBL : Mµ = M mar µ /GZ → DBL/GZ is an embedding. Part (b) says that the right equivalence class of f is determined by the isomorphism class of H(cid:48)(cid:48) 0 (f ). Part (a) would imply part (b). Both are global Torelli type conjectures. Part (b) was proved in [He1] for all simple and unimodal singularities and almost all bimodal singularities, all except three subseries of the eight bimodal series. Therefore for the proof of part (a) in these cases it remains mainly to control GZ well. But that is surprisingly difficult. In [He7] the conjectures 1.1 and 1.2 were proved for all simple and those 22 of the 28 exceptional unimodal and bimodal singularities, where all eigenvalues of the monodromy have multiplicity one. In this paper they will be proved for the remaining unimodal and exceptional bimodal singularities, that means, for the simple elliptic and the hy- perbolic singularities and for those 6 of the 28 exceptional unimodal and bimodal singularities which had not been treated in [He7]. µ A priori, logically conjecture 1.1 comes before conjecture 1.2. But the results in [He1] are more concrete and give already some infor- mation about the action of GZ on DBL and M mar . Anyway, the main remaining work is a good control of the groups GZ. That presents some unexpected difficulties. For example, we need two surprising general- izations of the number theoretic fact Z[e2πi/a]∩ S1 = {±e2πik/a k ∈ Z} for a ∈ N, one is lemma 2.5, the other is related to U16, see remark 4.3. The groups GZ(f0) will be calculated in the sections 3 and 4. Section 2 collects well known background material on the topology of singulari- ties. But it contains also an algebraic lemma 2.5 about automorphisms of monodromy modules. Section 5 collects notions and results from [He7] on marked singularities, the moduli spaces M mar (f0), the groups Gmar(f0) and the Torelli type conjectures. The sections 6 and 7 give the proofs of the conjectures 1.1 and 1.2 in the cases considered. Section 8 is motivated by the paper [MS] of Milanov and Shen and complements µ 4 FALKO GAUSS AND CLAUS HERTLING their results on (transversal) monodromy groups for certain families of simple elliptic singularities. The three principal congruence subgroups Γ(3), Γ(4) and Γ(6) which turn up in [MS] for certain families are shown to turn up also in the biggest possible families. 2. Review on the topology of isolated hypersurface singularities First, we recall some classical facts and fix some notations. An isolated hypersurface singularity (short: singularity) is a holomorphic function germ f : (Cn+1, 0) → (C, 0) with an isolated singularity at 0. Its Milnor number µ := dimOCn+1,0/( ∂f ∂xi ) is finite. For the following notions and facts compare [AGV2], [Eb2] and (for the notion of an unfolding) [AGV1]. A good representative of f has to be defined with some care [Mi][AGV2][Eb2]. It is f : Y → T with Y ⊂ Cn+1 a suitable small neighborhood of 0 and T ⊂ C a small disk around 0. Then f : Y (cid:48) → T (cid:48) with Y (cid:48) = Y − f−1(0) and T (cid:48) = T −{0} is a locally trivial C∞-fibration, the Milnor fibration. Each fiber has the homotopy type of a bouquet of µ n-spheres [Mi]. n Therefore the (reduced for n = 0) middle homology groups are (f−1(τ ), Z) ∼= Zµ for τ ∈ T (cid:48). Each comes equipped with an inter- H (red) section form I, which is a datum of one fiber, a monodromy Mh and a Seifert form L, which come from the Milnor fibration, see [AGV2, I.2.3] for their definitions (for the Seifert form, there are several con- ventions in the literature, we follow [AGV2]). Mh is a quasiunipotent automorphism, I and L are bilinear forms with values in Z, I is (−1)n- symmetric, and L is unimodular. L determines Mh and I because of the formulas [AGV2, I.2.3] L(Mha, b) = (−1)n+1L(b, a), I(a, b) = −L(a, b) + (−1)n+1L(b, a). (1) (2) The Milnor lattices Hn(f−1(τ ), Z) for all Milnor fibrations f : Y (cid:48) → T (cid:48) and then all τ ∈ R>0 ∩ T (cid:48) are canonically isomorphic, and the isomorphisms respect Mh, I and L. This follows from Lemma 2.2 in [LR]. These lattices are identified and called Milnor lattice M l(f ). The group GZ is GZ = GZ(f ) := Aut(M l(f ), L) = Aut(M l(f ), Mh, I, L), (3) the second equality is true because L determines Mh and I. We will use the notation M l(f )C := M l(f ) ⊗Z C, and analogously for other µ-CONSTANT MONODROMY GROUPS FOR SOME SINGULARITIES 5 rings R with Z ⊂ R ⊂ C, and the notations := (cid:77) λ(cid:54)=1 M l(f )λ ⊂ M l(f )C, := ker((Mh − λ id)µ : M l(f )C → M l(f )C) ⊂ M l(f )C, M l(f )λ M l(f )1,Z := M l(f )1 ∩ M l(f ) ⊂ M l(f ), M l(f )(cid:54)=1 M l(f )(cid:54)=1,Z := M l(f )(cid:54)=1 ∩ M l(f ) ⊂ M l(f ). The formulas (1) and (2) show I(a, b) = L((Mh − id)a, b). Therefore the eigenspace with eigenvalue 1 of Mh is the radical Rad(I) ⊂ M l(f ) (cid:81)r of I. By (2) L is (−1)n+1-symmetric on the radical of I. In the case of a curve singularity (n = 1) with r branches, f = j=1 fj, the radical of I is a Z-lattice of rank r − 1, and it is generated by the classes lj ∈ M l(f ) which are obtained by pushing the (correctly oriented) cycles ∂Y ∩ f−1 (0) from the boundary of the fiber f−1(0) to the boundary of the fiber f−1(τ ). Then j l1 + ... + lr = 0, so L(li, lj) > 0 l1, ...,(cid:98)lj, ..., lr L(li, lj) = intersection number of (fi, fj) for i (cid:54)= j, for i (cid:54)= j, is a Z-basis of Rad(I) for any j. (4) (5) (6) Kaenders proved the following result using Selling reduction. It will be useful for the calculation of Aut(Rad(I), L), because it implies that any automorphism of (Rad(I), L) maps the set {l1, ..., lr} to itself or to minus itself. Theorem 2.1. [Ka] In the case of a curve singularity as above, the set {l1, ..., lr} is up to a common sign uniquely determined by the properties (4), (5) and (6). So it is up to a common sign determined by the pair (Rad(I), L). Furthermore, L is negative definite on Rad(I). Examples 2.2. In the following three examples l = (l1, ..., lr), and L(lt, l) = (L(li, lj)). Then (L(li, lj))1≤i,j≤r−1 is the matrix of L on Rad(I) with respect to the Z-basis l1, ..., lr−1. The examples (ii) and (iii) will be useful in section 4, the example (i) is an alternative to a calculation in the proof of theorem 8.4 in [He7]. (i) D2m : f = x2m−1 + xy2 = x(xm−1 − iy)(xm−1 + iy),  , (cid:18)−2 (cid:19) (L(li, lj))1≤i,j≤2 = 1 1 −m . −2 L(lt, l) = 1 1 −m m − 1 1 1 m − 1 −m 6 FALKO GAUSS AND CLAUS HERTLING Obviously Aut(Rad(I), L) = 12 in the case m = 2, and Aut(Rad(I), L) = 4 in the cases m ≥ 3. (ii) Z12 : f = x3y + xy4 = xy(x2 + y3), −4 −6  ,  , L(lt, l) = 3 1 1 −3 2 2 −5 3 Obviously Aut(Rad(I), L) = {± id}. (iii) Z18 : f = x3y + xy6 = xy(x2 + y5), L(lt, l) = 1 5 1 −3 2 2 −7 5 Obviously Aut(Rad(I), L) = {± id}. (cid:18)−4 (cid:18)−6 (cid:19) (cid:19) . . (L(li, lj))1≤i,j≤2 = 1 1 −3 (L(li, lj))1≤i,j≤2 = 1 1 −3 Finally, in section 3, distinguished bases will be used. Again, good references for them are [AGV2] and [Eb2]. We sketch their construction and properties. One can choose a universal unfolding of f , a good representative F of it with base space M ⊂ Cµ, and a generic parameter t ∈ M . Then Ft : Yt → T with T ⊂ C the same disk as above and Yt ⊂ Cn+1 is a morsification of f . It has µ A1-singularities, and their critical values u1, ..., uµ ∈ T are pairwise different. Their numbering is also a choice. Now choose a value τ ∈ T ∩ R>0 − {u1, ..., uµ} and a distinguished system of paths. That is a system of µ paths γj, j = 1, ..., µ, from uj to τ which do not intersect except at τ and which arrive at τ in clockwise order. Finally, shift from the A1 singularity above each value uj the (up to sign unique) vanishing cycle along γj to the Milnor fiber M l(f ) = Hn(f−1(τ ), Z), and call the image δj. The tuple (δ1, ..., δµ) forms a Z-basis of M l(f ). All such bases are called distinguished bases. They form one orbit of an action of a semidi- rect product Brµ (cid:110) {±1}µ. Here Brµ is the braid group with µ strings, see [AGV2] or [Eb2] for its action. The sign change group {±1}µ acts simply by changing the signs of the entries of the tuples (δ1, ..., δµ). The members of the distinguished bases are called vanishing cycles. triangular matrix in M (µ × µ, Z) with 1's on the diagonal and with The Stokes matrix S of a distinguished basis is defined as the upper Sij := (−1)n(n+1)/2 · I(δi, δj) for all i, j with i < j. The Coxeter-Dynkin diagram of a distinguished basis encodes S in a geometric way. It has µ vertices which are numbered from 1 to µ. Between two vertices i and j with i < j one draws µ-CONSTANT MONODROMY GROUPS FOR SOME SINGULARITIES 7 if Sij = 0, if Sij < 0, if Sij > 0. no edge Sij edges Sij dotted edges Coxeter-Dynkin diagrams of many singularities were calculcated by A'Campo, Ebeling, Gabrielov and Gusein-Zade. Some of them can be found in [Ga], [Eb1] and [Eb2]. Example 2.3. The hyperbolic singularities of type Tpqr with 1 1 and the simple elliptic singularities of types (cid:101)E6 = T333,(cid:101)E7 = T442 and (cid:101)E8 = T632 have distinguished bases with the Coxeter-Dynkin diagrams p + 1 q + 1 r < in figure 1 [Ga]. Figure 1. A Coxeter-Dynkin diagram of the singulari- ties of type Tpqr The Picard-Lefschetz transformation on M l(f ) of a vanishing cycle δ is sδ(b) := b − (−1)n(n+1)/2 · I(δ, b) · δ. The monodromy Mh is Mh = sδ1 ◦ ... ◦ sδµ (7) (8) for any distinguished basis δ = (δ1, ..., δµ). The matrices of the mon- odromy, Seifert form and intersection form with respect to a distin- guished basis δ are given by the following formulas. Mh(δ) = δ · (−1)n+1 · S−1St, I(δt, δ) = (−1)n(n+1)/2 · (S + (−1)nSt), L(δt, δ) = (−1)(n+1)(n+2)/2 · St. (9) (10) (11) 8 FALKO GAUSS AND CLAUS HERTLING Remark 2.4. The Stokes matrix S of a distinguished basis is related to the matrix V in [Eb2, Korollar 5.3 (i)] by the formula V = L(δt, δ) = (−1)(n+1)(n+2)/2 · St. Thus V is lower triangular, not upper triangular, contrary to the claim in [Eb2, Korollar 5.3 (i)]. The matrix of Var−1 for a distinguished basis δ and its dual basis δ∗ is (−1)(n+1)(n+2)/2S, namely Var−1(δ) = δ∗ · (−1)(n+1)(n+2)/2 · S. But then the matrix for the Seifert form L with L(a, b) = (Var−1(a))(b) is L(δt, δ) = Var−1(δt)(δ) = (−1)(n+1)(n+2)/2St. In [AGV2, I.2.5] this problem does not arise because there the matrix of a bilinear form with respect to a basis is the transpose of the usual matrix [AGV2, page 45]. This is applied to the matrices of I and L. A result of Thom and Sebastiani compares the Milnor lattices and monodromies of the singularities f = f (x0, ..., xn), g = g(y0, ..., ym) and f + g = f (x0, ..., xn) + g(xn+1, ..., xn+m+1). There are extensions by Deligne for the Seifert form and by Gabrielov for distinguished bases. All results can be found in [AGV2, I.2.7]. They are restated here. There is a canonical isomorphism Φ : M l(f + g) ∼=−→ M l(f ) ⊗ M l(g), with Mh(f + g) ∼= Mh(f ) ⊗ Mh(g) and L(f + g) ∼= (−1)(n+1)(m+1) · L(f ) ⊗ L(g). (12) (13) (14) If δ = (δ1, ..., δµ(f )) and γ = (γ1, ..., γµ(g)) are distinguished bases of f and g with Stokes matrices S(f ) and S(g), then Φ−1(δ1⊗γ1, ..., δ1⊗γµ(g), δ2⊗γ1, ..., δ2⊗γµ(g), ..., δµ(f )⊗γ1, ..., δµ(f )⊗γµ(g)) is a distinguished basis of M l(f + g), that means, one takes the van- ishing cycles Φ−1(δi ⊗ γj) in the lexicographic order. Then by (11) and (14), the matrix S(f + g) = S(f ) ⊗ S(g) (15) (where the tensor product is defined so that it fits to the lexicographic order) is the Stokes matrix of this distinguished basis. In the special case g = x2 n+1, the function germ f +g = f (x0, ..., xn)+ n+1 ∈ OCn+2,0 is called stabilization or suspension of f . As there are x2 n+1) → Z, and they differ by a sign, there only two isomorphisms M l(x2 are two equally canonical isomorphisms M l(f ) → M l(f + x2 n+1), and they differ just by a sign. Therefore automorphisms and bilinear forms µ-CONSTANT MONODROMY GROUPS FOR SOME SINGULARITIES 9 on M l(f ) can be identified with automorphisms and bilinear forms on M l(f + x2 L(f + x2 n+1). In this sense n+1) = (−1)n · L(f ) and Mh(f + x2 n+1) = −Mh(f ) (16) [AGV2, I.2.7], and GZ(f + x2 not change under stabilization. n+1) = GZ(f ). The Stokes matrix S does The following algebraic lemma from [He7] will be very useful in the It can be seen as a generalization of the number sections 3 and 4. theoretic fact Z[e2πi/a] ∩ S1 = {±e2πik/a k ∈ Z}. Lemma 2.5. [He7, lemma 8.2] Let H be a free Z-module of finite rank µ, and HC := H ⊗Z C. Let Mh : H → H be an automorphism of finite order, called monodromy, with three properties: four sequences (mi)i=1,..., Ord , (i) Each eigenvalue has multiplicity 1. Denote Hλ := ker(Mh − λ · id : HC → HC). := {ord λ λ eigenvalue of Mh} ⊂ Z≥1. (ii) Denote Ord (j(i))i=2,..., Ord , There exist (pi)i=2,..., Ord , (ki)i=2,..., Ord of numbers in Z≥1 and two num- bers i1, i2 ∈ Z≥1 with i1 ≤ i2 ≤ Ord and with the properties: Ord = {m1, ..., m Ord }, pi is a prime number, pi = 2 for i1 + 1 ≤ i ≤ i2, pi ≥ 3 else, j(i) = i − 1 for i1 + 1 ≤ i ≤ i2, j(i) < i else, (iii) A cyclic generator a1 ∈ H exists, that means, mi = mj(i)/pki i . µ−1(cid:77) essarily (±1)-symmetric) on(cid:76) H = i=0 Z · M i h(a1). Finally, let I be an Mh-invariant nondegenerate bilinear form (not nec- λ(cid:54)=±1 Hλ with values in C. Then Aut(H, Mh, I) = {±M k h k ∈ Z}. 3. The group GZ for the simple elliptic and the hyperbolic singularities The simple elliptic and the hyperbolic singularities are 1-parameter families of singularities, which had been classified by Arnold [AGV1]. For each triple (p, q, r) ∈ N3≥2 with p ≥ q ≥ r and κ := 1 r ≤ 1 one has one family, denoted Tpqr. The hyperbolic singularities are T333 = (cid:101)E6, T442 = (cid:101)E7, T632 = (cid:101)E8. The singularities of types Tpqr with those with κ < 1, the simple elliptic are those with κ = 1. For the three families of simple elliptic singularities also other symbols are used, p + 1 q + 1 10 FALKO GAUSS AND CLAUS HERTLING r = 2 exist as curve singularities, all others as surface singularities. Normal forms will be discussed in section 6. Here for each family the group GZ = Aut(M l(f ), L) will be analyzed. The result in Theorem 3.1 is completely explicit in the case κ < 1 and partly explicit in the case κ = 1. The whole section is devoted to its proof. Besides κ := 1 p + 1 q + 1 r , also χ := lcm(p, q, r) ∈ N will be used. A singularity of type Tpqr has Milnor number µ = p + q + r − 1, and its monodromy has the characteristic polynomial tp − 1 t − 1 · tq − 1 t − 1 · tr − 1 t − 1 · (t − 1)2. Theorem 3.1. Consider a surface singularity f of type Tpqr (with κ ≤ 1) with Milnor lattice M l(f ), monodromy Mh, intersection form I and Seifert form L. (a) Then dim M l(f )1 = 2, rank Rad(I) = 1 if κ < 1 and = 2 if κ = 1. Choose a Z-basis b1, b2 of M l(f )1,Z with b1 ∈ Rad(I) and L(b1, b2) ≤ 0. Then L(b1, b2) = −χ and Mhb2 = b2 + χ(κ − 1) · b1. (17) (b) The restriction map GZ → Aut(M l(f )1,Z, L) is surjective. De- note by T ∈ Aut(M l(f )1,Z) the automorphism with T (b1) = b1 and T (b2) = b2 + b1. Denote b := (b1, b2). Aut(M l(f )1,Z, L) = {b (cid:55)→ b · A A ∈ SL(2, Z)} (18) ∼= SL(2, Z) if κ = 1, Aut(M l(f )1,Z, L) = {±T k k ∈ Z} (19) (c) The group GZ for κ < 1 and the subgroup {g ∈ GZ g(b1) = ±b1} ⊂ GZ for κ = 1 will be described explicitly, except for the part U2, see below. There is a monodromy invariant decomposition if κ < 1. M l(f )(cid:54)=1 = M l(1) C ⊕ M l(2) C ⊕ M l(3) C such that the characteristic polynomial of MhM l(j) is tr − 1 t − 1 and such that the following holds. tp − 1 t − 1 tq − 1 t − 1 , , for j = 1, 2, 3 (cid:27) GZ {g ∈ GZ g(b1) = ±b1} for κ < 1 for κ = 1 = (U1 (cid:111) U2) × {± id}, (22) (20) (21) µ-CONSTANT MONODROMY GROUPS FOR SOME SINGULARITIES 11 where U1 is the infinite subgroup of GZ )α × (MhM l(2)C U1 = {T δ × (MhM l(1)C )β × (MhM l(3)C (δ, α, β, γ) ∈ Z × Zp × Zq × Zr with )γ γ r + ≡ δ χ (23) mod 1} α p + β q and where U2 is a finite subgroup of GZ with if p > q > r, if p = q > r or p > q = r, if p = q = r. (24)  = {id} ∼= S2 ∼= S3 U2 which consists of certain automorphisms which act trivially on M l(f )1 and which permute those of the subspaces M l(j) C which have equal di- mension. Proof: Choose a distinguished basis δ = (δ1, ..., δµ) with the Coxeter-Dynkin diagram in example 2.3. Then the monodromy ma- trix MM with Mh(δ) = δ· MM can be calculated either with (8) or with (9). It had been calculated in [He1] with (8), and it is (here all not specified entries are 0) MM = (25) M2 M8 M9 M3 M10 M5 M6 M7 M4 M1   ∈ M ((p − 1) × (p − 1), Z), with the following blocks, M1 = 1 0  (cid:18) 3 −1 −1 . . . . . . 0 −1 1 −1 (cid:19) M4 = 2 −2 −1 , M2 ∈ M ((q − 1) × (q − 1), Z) and M3 ∈ M ((r − 1) × (r − 1), Z) are defined analogously, and M5, M6, M7, M8, M9, M10 are of suitable sizes with all entries except the following being 0, (M5)11 = (M6)11 = (M7)11 = −1, (M8)11 = (M8)12 = (M9)11 = (M9)12 = (M10)11 = (M10)12 = 1. (M5)21 = (M6)21 = (M7)21 = 1, 12 FALKO GAUSS AND CLAUS HERTLING Define (cid:101)b1 (cid:101)b2 (cid:32) p−1(cid:88) := δµ−1 − δµ, := χ · p − i p i=1 q−1(cid:88) i=1 r − i r δi + r−1(cid:88) i=1 + q − i q δp−1+i (cid:33) δp+q−2+i + δµ−1 . Then one calculates (26) (27) (28) (29) and with (11), which is here L(δt, δ) = St, Mh((cid:101)b1) =(cid:101)b1, Mh((cid:101)b2) =(cid:101)b2 + χ(κ − 1) ·(cid:101)b1, (cid:32) (cid:33) (cid:19) L((cid:101)b1,(cid:101)b1) L((cid:101)b1,(cid:101)b2) L((cid:101)b2,(cid:101)b1) L((cid:101)b2,(cid:101)b2) −χ 2 (κ − 1) (cid:18)0 χ χ2 = . if κ = 1 and it has a 2 × 2 Jordan block if κ < 1 (of course, this is By (28),(cid:101)b1,(cid:101)b2 is a Q-basis of M l(f )1,Q and Mh is on M l(f )1 semisimple well known). From the coefficients one sees that(cid:101)b1,(cid:101)b2 is a Z-basis of M l(f )1,Z. Here it is important that the coefficients of(cid:101)b2 have greatest common divisor 1. As the equations (17) hold for(cid:101)b1,(cid:101)b2, they hold for any basis b1, b2 as in (a). (b) If κ = 1 then (29) shows that L is on M l(f )1,Z up to the factor χ the standard symplectic form. Therefore (18) holds. If κ < 1 then (29) shows (19). The restriction map GZ → Aut(M l(f )1,Z) contains T . This follows from (22) (whose proof below does not use this fact), because obviously (δ, α, β, γ) as in (23) with δ = 1 exist. This shows (b) in the case κ < 1. In the case κ = 1 we did not calculate lifts to GZ of other elements of Aut(M l(f )1,Z, L). In this case the surjectivity of GZ → Aut(M l(f )1,Z, L) follows in two ways: It follows from [Kl, III.2.6], and it follows from calculations in [He1], which are discussed in the proof of theorem 6.1. (c) We will prove (c) for the special choice(cid:101)b1,(cid:101)b2. Then (c) holds Aut(M l(f )1,Z, L), an element g ∈ GZ with g((cid:101)b1) = b1, g((cid:101)b2) = b2 exists. for any b1, b2 as in (a) because by the surjectivity of the map GZ → µ-CONSTANT MONODROMY GROUPS FOR SOME SINGULARITIES 13 Define M l[1] Z M l(1) C := Z(cid:101)b1 ⊕ p−1(cid:77) Zδi, M l[1] C := M l[1] Z ⊗Z C, := M l[1] i=1 C ∩ M l(f )(cid:54)=1, C , M l(2) (30) (31) and analogously M l[2] Z , M l[2] C and M l[3] Z , M l[3] C , M l(3) C . A look at the matrix MM shows the following. Mh : δ1 +(cid:101)b1 (cid:55)→ δ2 (cid:55)→ ... (cid:55)→ δp−1 (cid:55)→ −(δ1 + ... + δp−1) (cid:55)→ δ1 +(cid:101)b1. (32) Therefore M l[1] tp− 1, and M l[1] polynomial (tp − 1)/(t − 1). C = C(cid:101)b1⊕ M l(1) Z is a cyclic Mh-module with characteristic polynomial C has the characteristic C , and Mh on M l(1) Lemma 2.5 applies and shows Aut(M l[1] Finally, Mh, I and L are well defined on the quotient lattice M l[1] )α α ∈ {0, 1, ..., p − 1}}. Z , L) = {±(MhM l[1]Z Z /Z ·(cid:101)b1,−I) is a root lattice of type Ap−1. The last state- (cid:101)b1, and (M l[1] (33) Z /Z· ment follows immediately from the part of the Coxeter-Dynkin diagram which corresponds to δ1, ..., δp−1. M l[2] Z and M l[3] Z have the same properties as M l[1] Z , with q respectively r instead of p. Now M l(f )(cid:54)=1 = M l(1) C ⊕ M l(2) C ⊕ M l(3) C i=1 M l[1] Z + M l[2] Z + M l[3] is clear. The Z-lattice Z = Z(cid:101)b1 ⊕ µ−2(cid:77) is a primitive sublattice of M l(f ) of rank µ − 1. Any g ∈ GZ with Zδi = (C(cid:101)b1 ⊕ M l(f )(cid:54)=1) ∩ M l(f ) g((cid:101)b1) = ±(cid:101)b1 maps it to itself, because it maps C(cid:101)b1 and M l(f )(cid:54)=1 and Z /Z(cid:101)b1 Z )/Z(cid:101)b1 = M l[1] M l(f ) to themselves. g maps also the quotient lattice Z /Z(cid:101)b1 ⊕ M l[2] Z /Z(cid:101)b1 ⊕ M l[3] Z + M l[2] Z + M l[3] to itself. But this is together with −I an orthogonal sum of lattices of types Ap−1, Aq−1 and Ar−1. Therefore g can only permute the summands, and only those summands of equal rank. (M l[1] If p = q, a special element σ12 ∈ GZ is given by σ12(δi) = δp−1+i, σ12(δj) = δj σ12(δp−1+i) = δi for p + q − 2 ≤ j ≤ µ. for 1 ≤ i ≤ p − 1, 14 FALKO GAUSS AND CLAUS HERTLING Z ) = M l[j] σ12 ∈ GZ follows immediately from the symmetry of the Coxeter- Dynkin diagram. Similarly σ23 ∈ GZ is defined if q = r. In any case, these elements generate a subgroup U2 ⊂ GZ with the properties in (c). Therefore, starting with an arbitrary element (cid:101)g ∈ GZ if κ < 1 respectively (cid:101)g ∈ {g ∈ GZ g((cid:101)b1) = ±(cid:101)b1} if κ = 1, one can com- g ∈ GZ with g((cid:101)b1) = (cid:101)b1 and g(M l[j] pose it with ± id and an element of U2, and one obtains an element Z for j = 1, 2, 3. Then gM l[1]Z )α for a unique α ∈ {0, 1, ..., p − 1}, and similarly Also g((cid:101)b2) =(cid:101)b2 + δ(cid:101)b1 for some δ ∈ Z. One calculates, observing (32), with β ∈ {0, 1, ..., q − 1} and γ ∈ {0, 1, ..., r − 1} for M l[2] Z and M l[3] (cid:32) p−1(cid:88) (cid:32) p−1(cid:88) Z . (cid:32) p−1(cid:88) (cid:32) p−1(cid:88) The definition (27) of(cid:101)b2 shows p − i p p − i p = (MhM l[1]Z (cid:33) (cid:33) p − i p p − i p − (δ1 +(cid:101)b1) + (cid:32)(cid:101)b1 + α(cid:88) r−1(cid:88) (cid:101)b1, (cid:33) (cid:33) (cid:33) p−1(cid:88) q−1(cid:88) (cid:101)b1. M α h (35) δi + (34) α p δi δi δi δi = = Mh − 1 p i=1 i=1 i=1 i=1 i=1 δi + δp−1+i + r − i r δp+q−2+i,(36) (cid:101)b2 + − δµ−1 = − 1 χ q − i q i=1 i=1 i=1 p − i p (cid:18)−δ (cid:33) α(cid:88) χ δi (cid:32)(cid:101)b1 + (cid:19) p−1+β(cid:88) γ r + ·(cid:101)b1 (cid:33) − δi β q + α p (cid:32)(cid:101)b1 + + − i=1 i=p (cid:32)(cid:101)b1 + (37) (cid:33) δi . p+q−2+γ(cid:88) i=p+q−1 (35) gives g(−δµ−1) = −δµ−1 + − + Therefore β q α p + γ r ≡ δ χ mod 1 (38) )β × (MhM l(3)C )γ. and g = T δ × (MhM l(1)C )α × (MhM l(2)C Thus g ∈ U1, so GZ ⊂ (U1 (cid:111) U2) × {± id}. Vice versa, we have to show U1 ⊂ GZ. Fix a g ∈ U1. It respects the decomposition M l(f )C = M l(f )1 ⊕ M l(1) C ⊕ M l(2) C ⊕ M l(3) C . This is a left and right orthogonal decomposition with respect to the Seifert form L. The restriction of g to each of the four blocks respects (cid:16) (cid:17) ⊕ Zδµ−1. µ-CONSTANT MONODROMY GROUPS FOR SOME SINGULARITIES 15 L there, so g ∈ Aut(M l(f )C, L). maps the lattice M l[1] Z + M l[2] M l[3] sublattice of M l(f ) of rank µ − 1 with Z , thus also the sum M l[1] Z to itself, and analogously the lattices M l[2] It restricts on M l[1] C to M α h , so it Z and Z . This sum is a primitive Z + M l[3] M l(f ) = M l[1] Z + M l[2] Z + M l[3] Z The calculation above of g(−δµ−1) shows g(δµ−1) ∈ M l(f ) and g(δµ−1) ≡ δµ−1 modulo the sublattice. Therefore g ∈ GZ. (cid:3) 4. The group GZ for 6 of the 28 exceptional unimodal and bimodal singularities The 14 1-parameter families of exceptional unimodal singularities and the 14 2-parameter families of exceptional bimodal singularities had been classified by Arnold. Normal forms can be found in [AGV1]. In [He7, theorem 8.3] the group GZ was calculated for 22 of the 28 families, namely for those families where all eigenvalues of the monodromy have multiplicity 1. In these cases it turned out that GZ is simply {±M k h k ∈ Z}. The proof used lemma 2.5 and that the Milnor lattice is in these cases a cyclic monodromy module. In this section GZ will be determined for the remaining 6 of the 28 families. These are the families Z12, Q12, U12, Z18, Q16, U16. In these cases, some eigenvalues have multiplicity 2. This is similar to the case of the singularity D2m, which had also been treated in [He7, theorem 8.4]. Also the proof will be similar. It will again use lemma 2.5 and combine that with an additional analysis of the action of GZ on the sum of the eigenspaces with dimension = 2, see lemma 4.2 below. This lemma presents a surprise, it points at a funny generalization of the number theoretic fact Z[e2πi/n] ∩ S1 = {±e2πik/n k ∈ Z}. Also, the lemma uses at the end a calculation of Gmar ⊂ GZ, which will only come in section 7. The whole section 4 is devoted to the proof of the following theorem. Theorem 4.1. In the case of the 6 families of exceptional unimodal and bimodal singularities Z12, Q12, U12, Z18, Q16 and U16, the group GZ is GZ = {±M k h k ∈ Z} × U with U ∼= {id} S2 Z12 Q12 U12 Z18 Q16 U16 S3 {id} S2 S3 (39) (This is independent of the number of variables, i.e. it does not change under stabilization.) 16 FALKO GAUSS AND CLAUS HERTLING Proof: Here we consider all 6 families as surface singularities. Their characteristic polynomials pch have all the property pch = p1 · p2 with p2p1 and p1 having only simple roots. They are as follows. Here Φm is the cyclotomic polynomial of primitive m-th unit roots. Z12 Q12 U12 Z18 Q16 U16 2 Φ15Φ2 pch Φ22Φ2 3 Φ15Φ2 5 p1 Φ22Φ2 Φ15Φ3 Φ12Φ6Φ4Φ2 Φ34Φ2 Φ21Φ3 Φ15Φ5 p2 Φ2 3 Φ12Φ6Φ2 2 Φ21Φ2 2 Φ34Φ2 Φ4Φ2 4Φ2 Φ3 Φ5 Φ3 Φ2 (40) Orlik [Or] had conjectured that the Milnor lattice of any quasiho- mogeneous singularity is a sum of cyclic monodromy modules and that the characteristic polynomials of Mh on the cyclic pieces are p1, ..., pr where pch = p1 · ... · pr and prpr−1...p1 and p1 has simple roots (r and p1, ..., pr are uniquely determined by this). In the case of curve singular- ities, Michel and Weber [MW] have a proof of this conjecture. In [He1, 3.1] the conjecture was proved (using Coxeter-Dynkin diagrams) for all those quasihomogeneous surface singularites of modality ≤ 2 which are not stabilizations of curve singularities. So especially, the conjecture is true for the families of singularities Z12, Q12, U12, Z18, Q16, U16. There are a1, a2 ∈ M l(f ) with Z · M i (cid:32)deg p1−1(cid:77) (cid:32)deg p2−1(cid:77) =: B1 ⊕ B2.(41) Z · M i M l(f ) = (cid:33) (cid:33) ⊕ h(a1) h(a2) i=0 i=0 Denote (42) B3 := ker (p2(Mh) : M l(f )C → M l(f )C) ∩ M l(f ). It is a primitive sublattice of M l(f ) of rank 2 deg p2. Also, (B1)C = ker((p1/p2)(Mh)) ⊕ (B1 ∩ B3)C, B2 ⊂ B3, (43) and B1 ∩ B3 and B2 are both Mh-invariant primitive sublattices of B3 of rank deg p2. Together they generate B3. Any g ∈ GZ with gB3 = ±(MhB3)k for some k ∈ Z restricts because of (43) to an automorphism of B1. Lemma 2.5 applies and shows gB1 = ±(MhB1)l for some l ∈ Z. Now gB3 = ±(MhB3)k enforces k ≡ l mod lcm(m Φmp2) and g = ±M l h. Therefore {g ∈ GZ gB3 = ±(MhB3)k for some k} = {±M k Lemma 4.2 (c) determines Aut(B3, L), Aut(B3, L) = {±(MhB3)k k ∈ Z} × U with U as in (39). (45) Lemma 4.2 (d) shows that the map GZ → Aut(B3, L) is surjective. (cid:3) Together with (44) this gives (39). h k ∈ Z}. (44) µ-CONSTANT MONODROMY GROUPS FOR SOME SINGULARITIES 17 Lemma 4.2. (a) Let VZ be a Z-lattice of rank 2 with a Z-basis b = (b1, b2) and a symmetric pairing LZ given by LZ(bt, b) = for some m ∈ N≥2. (cid:18) 2 −1 (cid:19) −1 m Define ξ := e2πi/l for some l ∈ {3, 4, 5}, where we exclude the cases (m ≥ 3, l = 5). Define VC := VZ ⊗Z C, VZ[ξ] := VZ ⊗Z Z[ξ] ⊂ VC, and extend LZ sesquilinearly (=linear×semilinear) to VC. Then {r ∈ VZ[ξ] LC(r, r) = 2} = {±ξk k ∈ Z} × {r ∈ VZ LZ(r, r) = 2}. (46) (47) (48) (49) (50) In the case m ≥ 3 {r ∈ VZ[ξ] LC(r, r) = m, r /∈ Z[ξ]b1} = {±ξk k ∈ Z} × {r ∈ VZ LZ(r, r) = m, r /∈ Zb1}. Aut(VZ, LZ) ∼= {± (b) In the situation of (a) Aut(VZ[ξ], LC) = {±ξk k ∈ Z} · Aut(VZ, LZ) (± id lives on both sides, therefore not ×) (cid:18)−1 1 (cid:19) 0 1 } ,± (cid:19) (cid:18)1 0 ∼= {± id} × S2 ∼= {± id} × S3 0 1 Aut(VZ, LZ) ∼= Aut(root lattice of type A2) in the cases m ≥ 3, in the case m = 2. (c) In the situation of the proof of theorem 4.1 Aut(B3, L) = {±(MhB3)k k ∈ Z} × U (51) (d) In the situation of the proof of theorem 4.1 the map GZ → with U as in (39). Aut(B3, L) is surjective. Proof: (a) An element r = r1b1 + r2b2 with r1, r2 ∈ Z[ξ] satisfies LC(r, r) = 2r12 − (r1r2 + r1r2) + mr22 = r12 + r1 − r22 + (m − 1)r22. (52) First consider the cases l ∈ {3, 4}. Then Z[ξ] ∩ R = Z. Then the three absolute values in (52) are non-negative integers. Their sum is 2 if and only if r1 = 1, r2 = 0 in the cases m ≥ 3, (53) (cid:27) (r1,r2) ∈ {(1, 0), (0, 1), (1, 1)} and in the last case r1 = r2 in the case m = 2. (54) 18 FALKO GAUSS AND CLAUS HERTLING In the case m ≥ 3 and in the case of an r /∈ Z[ξ]b1, the sum of the three absolute values in (53) is m if and only if (r1 = 0,r2 = 1) or (r1 = r2,r1 = 1). (55) Together with Z[ξ] ∩ S1 = {±ξk k ∈ Z} this shows part (a) in the cases l ∈ {3, 4}. It rests to consider the case (m, l) = (2, 5). In that case write r1 = r10 + r11ξ + r12ξ2 + r13ξ3, r2 = r20 + r21ξ + r22ξ2 + r23ξ3 with rij ∈ Z. Then LC(r, r) = 2r12 + 2r22 − (r1r2 + r1r2) (cid:34) 3(cid:88) (cid:34) 3(cid:88) 3(cid:88) j=0 j=0 2 j=0 = 2 + 2 (cid:34) − j=1 3(cid:88) 3(cid:88) 3(cid:88) j=1 j=1 (cid:35) (cid:35) r1jr1k r2jr2k (cid:88) (cid:88) j−k≥2 j−k≥2 r2 1j + (ξ + ξ4) r2 2j + (ξ + ξ4) r1jr1,j−1 + (ξ2 + ξ3) r2jr2,j−1 + (ξ2 + ξ3) r1jr2j + (ξ + ξ4) (r1jr2,j−1 + r1,j−1r2j) (cid:88) (cid:35) + (ξ2 + ξ3) j−k≥2 √ 5 2 = A1 + A2 · (r1jr2k + r1kr2j) with A1, A2 ∈ Z. It is not so easy to find, but easy to check that A1 is equal to 2j + (r1j − r2j)2(cid:3) (cid:2)r2 (cid:2)(r1j − r1k − r2j + r2k)2 + (r1j − r1k)2 + (r2j − r2k)2(cid:3) . 1j + r2 (56) 3(cid:88) (cid:88) j=0 j<k 1 4 1 4 + Now it is an easy exercise to find the 8-tuples (r10, ..., r23) ∈ Z8 for which (56) takes the value 2. They are (here ej = (δij)i=1,...,8 for j = 1, ..., 8 is the standard basis of Z8) ± e1, ...,±e8,±(e1 + e5),±(e2 + e6),±(e3 + e7),±(e4 + e8), ±(1, 1, 1, 1, 0, 0, 0, 0),±(0, 0, 0, 0, 1, 1, 1, 1),±(1, 1, 1, 1, 1, 1, 1, 1). Observe 1 + ξ + ξ2 + ξ3 = −ξ4 and (57) {r ∈ VZ LZ(r, r) = 2} = {±b1,±b2,±(b1 + b2)}. µ-CONSTANT MONODROMY GROUPS FOR SOME SINGULARITIES 19 The coefficients (r10, ..., r23) in (57) give precisely the elements r = r1b1 + r2b2 on the right hand side of (46). This shows part (a) for (m, l) = (2, 5). (b) Any element g of Aut(VZ[ξ], LC) will map the sets in (46) and (47) to themselves. The basis elements b1 and b2 are mapped to two elements in these sets with LC(g(b1)), g(b2)) = LZ(b1, b2) = −1. Therefore g is up to a factor in {±ξk k ∈ Z} an element of Aut(VZ, LZ). This shows (48). In the case m ≥ 3 {r ∈ VZ LZ(r, r) = 2} = {±b1} and {r ∈ VZ LZ(r, r) = m} = {±b2,±(b1 + b2)}. This shows (49). The case m = 2 is the case of the root lattice of type A2. (50) is well known and easy to see. (c) In the cases Z12 and Z18 as curve singularities, the examples 2.2 (ii) and (iii) showed Aut(Rad(I), L) = {± id}. Here Rad(I) = − id). Under stabilization Rad(I) becomes B3, and L ker(M curve case changes just the sign, see (16). Thus Aut(B3, L) = {± id}. Because of MhB3 = − id and U := {id} in the cases Z12 and Z18, this shows (c) in these cases. h Now consider the cases Q12, Q16, U12 and U16. Here part (b) will be used, but that has to be prepared. The normal forms of the quasihomogeneous surface singularities, given in section 7, show that they are sums of singularities in different variables of types Al and D2m with (l, 2m) as follows, (l, 2m) Q12 (2, 6) A2 ⊗ D6 A2 ⊗ D8 A3 ⊗ D4 A4 ⊗ D4 U12 (3, 4) U16 (4, 4) Q16 (2, 8) (58) Here the singularity Al is in one variable and has the characteristic ch = (tl+1 − 1)/(t − 1), the singularity D2m is a curve polynomial pAl ch = (t2m−1−1)Φ1. singularity and has the characteristic polynomial pD2m The Thom-Sebastiani results which were cited in section 2 apply, (M l(f ), L) ∼= (M l(Al), LAl) ⊗ (M l(D2m), LD2m), (59) (60) Mh and show ∼= M Al h ⊗ M D2m h , p2 = pAl ch, ∼= M Al (B3, L) ∼= (M l(Al), LAl) ⊗ (M l(D2m)1,Z, LD2m), MhB3 h ⊗ id . 20 FALKO GAUSS AND CLAUS HERTLING (cid:19) . (cid:18)−2 LD2m(bt, b) = The pair (M l(D2m)1,Z, LD2m) was considered in example 2.2 (i). There is a Z-basis b = (b1, b2) of M l(D2m)1,Z with 1 1 −m (61) The pairings L and LAl ⊗ LD2m will be extended sesquilinearly from the Z-lattices to the C-vector spaces. The Z-lattice M l(Al) is a cyclic monodromy module. Choose a gen- erator e of it. Therefore M l(Al) ⊗ M l(D2m)1,Z is a sum of two cyclic monodromy modules, and generators are e ⊗ b1 and e ⊗ b2. For any h ⊗ id) there are unique automorphism g of (M l(Al) ⊗ M l(D2m)1,Z, M Al polynomials g1, g2, g3, g4 ∈ Z[t] of degree ≤ deg p2 − 1 such that h )(v) ⊗ b2 h )(v) ⊗ b2 h )(v) ⊗ b1 + g3(M Al h )(v) ⊗ b1 + g4(M Al (cid:18)g(v ⊗ b1) (cid:19) (cid:18)g1(M Al g2(M Al (cid:19) (62) g(v ⊗ b2) = for any v ∈ M l(Al). Now choose any eigenvalue ξ of M Al h . Then Z[ξ] is a principal ideal h − ξ id)∩ M l(Al)Z[ξ] is a free Z[ξ]-module of domain. The space ker(M Al rank 1. Choose a generating vector v. This choice gives an isomorphism from this space to Z[ξ]. The spaces ∼= (ker(M Al ker(Mh−ξ id)∩M l(f )Z[ξ] h −ξ id)∩M l(Al)Z[ξ])⊗M l(D2m)1,Z[ξ] are free Z[ξ]-modules of rank 2. The space on the right hand side has the Z[ξ]-basis (v ⊗ b1, v ⊗ b2) =: v ⊗ b. Now (62) becomes g(v ⊗ b) = v ⊗ b · The pairing satisfies (cid:18)g1(ξ) g2(ξ) (cid:19) (cid:18)−2 g3(ξ) g4(ξ) . (cid:19) (63) , 1 1 −m (LAl ⊗ LD2m)(v ⊗ b) = LAl(v, v) · (64) where LAl(v, v) ∈ Z[ξ] ∩ R>0. This space with this pairing is up to a scalar isomorphic to a pair (VZ[ξ], LC) considered in the parts (a) and (b). Therefore by part (b), its group of automorphisms is isomorphic to {±ξk k ∈ Z} · Aut(VZ, LZ). Thus any element of Aut(B3, L) restricts on ker(Mh−ξ id)∩M l(f )Z[ξ] to such an automorphism. In the cases Q12, Q16 and U16 the polyno- mial p2 = Φ3, Φ3 respectively Φ5 is irreducible, so all its zeros ξ are conjugate. Therefore then Aut(B3, L) ∼= {±(MhB3)k k ∈ Z} · Aut(VZ, LZ} ∼= {±(MhB3)k k ∈ Z} × U, which proves (c) in these cases. µ-CONSTANT MONODROMY GROUPS FOR SOME SINGULARITIES 21 reducible. Consider an automorphism g of In the case U12 the characteristic polynomial pA3 (M l(Al) ⊗ M l(D2m)1,Z, LAl ⊗ LD2m). ch = p2 = Φ4Φ2 is ,± 1 0 0 1 ,± ,± (cid:26) ± ± (65) , . It is determined by the polynomials g1, g2, g3, g4 ∈ Z[t] in (62). For ξ = i and for ξ = −1 it gives an automorphism of Z[ξ]v⊗ b1⊕Z[ξ]v⊗ b2 which is given by a matrix which is by part (b) in 1 −1 {±ξk k ∈ Z} · g3(ξ) g4(ξ) ,± (cid:18)g1(ξ) g2(ξ) (cid:19) (cid:18)0 −1 (cid:18)1 0 (cid:19) (cid:19) (cid:18)−1 1−1 0 (cid:19) (cid:19) (cid:18)0 1 (cid:18)−1 0−1 1 (cid:19) (cid:19)(cid:27) (cid:18)1 −1 (cid:18)g1 g2 (cid:18)1 + (t2 + 1)(cid:101)g1 (cid:19) (cid:19) (t2 + 1)(cid:101)g2 1 + (t2 + 1)(cid:101)g4 (t2 + 1)(cid:101)g3 (cid:18)1 + 2(cid:101)g1(−1) (cid:19) (cid:19) (cid:18)g1(−1) g2(−1) 2(cid:101)g2(−1) 1 + 2(cid:101)g4(−1) 2(cid:101)g3(−1) (cid:18)1 0 (cid:19) g3(−1) g4(−1) 0 −1 By multiplying g with a suitable automorphism we can suppose that the matrix for ξ = i is the identity matrix. Then for some(cid:101)g1,(cid:101)g2,(cid:101)g3,(cid:101)g4 ∈ Z[t], so g3 g4 = The only two possibilities are ± g ◦ ((M Al finishes the proof of (c) in the case U12. . In the case of a minus sign h )2 ⊗ idD2m) = − id, in the case of a plus sign g = id. This 0 1 , . = (d) In section 7 a subgroup Gmar of GZ will be calculated, and it h k ∈ Z} × U . This shows GZ ⊃ will be shown that Gmar = {±M k h k ∈ Z}× U and that the map GZ → Aut(B3, L) is surjectiv. (cid:3) {±M k Remark 4.3. The number theoretic fact Z[e2πi/a]∩S1 = {±e2πik/a k ∈ Z} can be interpreted as saying that in the case of the A1-lattice VZ = Z with Z-basis b1 = 1 and standard bilinear form LZ with LZ(b1, b1) = 1 and hermitian extension LC to C the analogue of (46) holds. Now (46) for m = 2 can be seen as a generalization from the case A1 to the case A2. Above it is proved only in the cases l = 3, 4, 5. 5. Review on µ-constant monodromy groups Gmar, marked singularities, their moduli spaces M mar , and Torelli µ type conjectures This paper is a sequel to [He7]. That paper studied holomorphic func- tions germs f : (Cn+1, 0) → (C, 0) with an isolated singularity at 0 22 FALKO GAUSS AND CLAUS HERTLING from a global perspective. Here we review most of the notions and results from [He7]. It defined the notions of marked singularity and strongly marked sin- gularity. The marking uses the Milnor lattice M l(f ) ∼= Zµ and the Seifert form L on it, which are explained in section 2. Definition 5.1. Fix one reference singularity f0. (a) Then a strong marking for any singularity f in the µ-homotopy there is a 1-parameter family of singularities with class of f0 (i.e. constant Milnor number connecting f and f0) is an isomorphism ρ : (M l(f ), L) → (M l(f0), L). (b) The pair (f, ρ) is a strongly marked singularity. Two strongly marked singularities (f1, ρ1) and (f2, ρ2) are right equivalent (notation: ∼R) if a coordinate change ϕ : (Cn+1, 0) → (Cn+1, 0) with f1 = f2 ◦ ϕ and ρ1 = ρ2 ◦ ϕhom exists, where ϕhom : (M l(f1), L) → (M l(f2), L) is the induced isomor- phism. (c) The notion of a marked singularity is slightly weaker. If f and ρ are as above, then the pair (f,±ρ) is a marked singularity (writing ±ρ, the set {ρ,−ρ} is meant, neither ρ nor −ρ is preferred). (notation: ∼R) if a coordinate change ϕ with (d) Two marked singularities (f1, ρ1) and (f2, ρ2) are right equivalent f1 = f2 ◦ ϕ and ρ1 = ±ρ2 ◦ ϕhom exists. Remarks 5.2. (i) The notion of a marked singularity behaves better than the notion of a strongly marked singularity, because it is not known whether all µ-homotopy families of singularities satisfy one of the following two properties: Assumption (5.1): Assumption (5.2): Any singularity in the µ-homotopy class of f0 has multiplicity ≥ 3. Any singularity in the µ-homotopy class of f0 has multiplicity 2. (66) (67) We expect that always one of two assumptions holds. For curve singu- larities and singularities right equivalent to semiquasihomogeneous sin- gularities this is true, but in general it is not known. In a µ-homotopy family where neither of the two assumptions holds, strong marking behaves badly, see (ii). (ii) If mult(f ) = 2 then (f, ρ) ∼R (f,−ρ), which is easy to see. If mult(f ) ≥ 3, then (f, ρ) (cid:54)∼R (f,−ρ), whose proof in [He7] is quite µ-CONSTANT MONODROMY GROUPS FOR SOME SINGULARITIES 23 intricate. These properties imply that the moduli space for strongly marked singularities discussed below is not Hausdorff in the case of a µ-homotopy class which satisfies neither one of the assumptions (66) or (67) In [He6] for the µ-homotopy class of any singularity f0 a moduli space Mµ(f0) was constructed. As a set it is simply the set of right equivalence classes of singularities in the µ-homotopy class of f0. But in [He6] it is constructed as an analytic geometric quotient, and it is shown that it is locally isomorphic to the µ-constant stratum of a singularity modulo the action of a finite group. The µ-constant stratum of a singularity is the germ (Sµ, 0) ⊂ (M, 0) within the germ of the base space of a universal unfolding F of f , such that for a suitable representative Sµ = {t ∈ M Ft has only one singularity x0 and Ft(x0) = 0}. (68) It comes equipped with a canonical complex structure, and Mµ inherits a canonical structure, see the chapters 12 and 13 in [He6]. In [He7] analogous results for marked singularities were proved. A better property is that M mar is locally isomorphic to a µ-constant stratum without dividing out a finite group action. Therefore one can consider it as a global µ-constant stratum or as a Teichmuller space for singularities. The following theorem collects results from [He7, theorem 4.3]. µ Theorem 5.3. Fix one reference singularity f0. Define the sets M smar µ (f0) M mar µ (f0) := {strongly marked (f, ρ) f in the µ-homotopy class of f0}/ ∼R, := {marked (f,±ρ) f in the µ-homotopy class of f0}/ ∼R . (69) (70) µ (a) M mar (f0) carries a natural canonical complex structure. It can be constructed with the underlying reduced complex structure as an an- alytic geometric quotient (see [He7, theorem 4.3] for details). (f0), [(f,±ρ)]) with its canonical complex struc- ture is isomorphic to the µ-constant stratum of f with its canonical complex structure (see [He6, chapter 12] for the definition of that). (b) The germ (M mar µ (c) For any ψ ∈ GZ(f0) =: GZ, the map ψmar : M mar µ → M mar , µ is an automorphism of M mar . The action µ → M mar GZ × M mar µ µ , [(f,±ρ)] → [(f,±ψ ◦ ρ)] (ψ, [(f,±ρ)] (cid:55)→ ψmar([(f,±ρ)]) 24 FALKO GAUSS AND CLAUS HERTLING is a group action from the left. µ (d) The action of GZ on M mar is properly discontinuous. The quo- tient M mar /GZ is the moduli space Mµ for right equivalence classes in the µ-homotopy class of f0, with its canonical complex structure. Es- pecially, [(f1,±ρ1)] and [(f2,±ρ2)] are in one GZ-orbit if and only if f1 and f2 are right equivalent. µ (e) If assumption (66) or (67) holds then (a) to (d) are also true and ψsmar with ψsmar([(f, ρ)]) := [(f, ψ ◦ ρ)]. If neither (66) for M smar nor (67) holds then the natural topology on M smar is not Hausdorff. µ µ We stick to the situation in theorem 5.3 and define two subgroups of GZ(f0). The definitions in [He7, definition 3.1] are different, they use µ-constant families. The following definitions are a part of theorem 4.4 in [He7]. µ Definition 5.4. Let (M mar )0 be the topological component of M mar (with its reduced complex structure) which contains [(f0,± id)]. Then )0 to itself} ⊂ GZ(f0).(71) )0 and Gsmar(f0) ⊂ GZ(f0) are := {ψ ∈ GZ ψ maps (M mar If assumption (66) or (67) holds, (M smar defined analogously. Gmar(f0) µ µ µ The following theorem is also proved in [He7]. Theorem 5.5. (a) In the situation above the map GZ/Gmar(f0) → {topological components of M mar ψ · Gmar(f0) (cid:55)→ the component ψmar((M mar µ )0) } µ is a bijection. (b) If assumption (66) or (67) holds then (a) is also true for M smar (c) − id ∈ GZ acts trivially on M mar and Gsmar(f0). assumption (66) holds for f0. Then {± id} acts freely on M smar and the quotient map (f0). Suppose additionally that (f0), µ µ µ M smar (f0) is a double covering. µ /{± id}−→ M mar µ (f0), [(f, ρ)] (cid:55)→ [(f,±ρ)] The first main conjecture in [He7] is part (a) of the following conjec- ture (the second main conjecture in [He7] is conjecture 5.11 (a) below). Conjecture 5.6. (a) Fix a singularity f0. Then M mar Equivalently (in view of theorem 5.5 (a)): Gmar(f0) = GZ. − id /∈ Gsmar(f0). (b) If the µ-homotopy class of f0 satisfies assumption (66), then is connected. µ µ-CONSTANT MONODROMY GROUPS FOR SOME SINGULARITIES 25 µ If (a) holds then (b) is equivalent to M smar (f0) having two compo- nents. If (a) and (b) hold, there should be a natural invariant which distinguishes the index 2 subgroup Gsmar ⊂ Gmar = GZ. Anyway, part (a) is the more important conjecture. Using the other definition of Gmar in [He7], it says that up to ± id, any element of GZ can be real- ized as transversal monodromy of a µ-constant family with parameter space S1. The whole conjecture 5.6 had been proved in [He7] for the simple singularities and those 22 of the 28 exceptional unimodal and bimodal singularities, where all eigenvalues of the monodromy have only multi- plicity one [He7, theorems 8.3 and 8.4]. In this paper it will be proved for the remaining unimodal and exceptional bimodal singularities. stabilizers StabGZ([(f, ρ)]) and (f0) and [(f,±ρ)] ∈ StabGZ([(f,±ρ)]) of points [(f, ρ)] ∈ M smar M mar (f0), we have to look at the symmetries of a single singularity. These had been discussed in [He6, chapter 13.2]. The discussion had been taken up again in [He7]. to understand the µ In order µ Definition 5.7. Let f0 = f0(x0, ..., xn) be a reference singularity and let f be any singularity in the µ-homotopy class of f0. If ρ is a marking, then GZ(f ) = ρ−1 ◦ GZ ◦ ρ. We define R := {ϕ : (Cn+1, 0) → (Cn+1, 0) biholomorphic}, (72) Rf (73) (74) Rf GsmarR (f ) GmarR (f ) := {ϕ ∈ R f ◦ ϕ = f}, := j1Rf /(j1Rf )0, := {ϕhom ϕ ∈ Rf} ⊂ GZ(f ), := {±ψ ψ ∈ GsmarR (f )}. (75) (76) Again, the definition of GsmarR is different from the definition in [He7, definition 3.1]. The characterization in (75) is [He7, theorem 3.3. (e)]. Rf is the finite group of components of the group j1Rf of 1-jets of coordinate changes which leave f invariant. The following theorem collects results from several theorems in [He7]. Theorem 5.8. Consider the data in definition 5.7. (a) If mult(f ) ≥ 3 then j1Rf = Rf . (b) The homomorphism ()hom : Rf → GZ(f ) factors through Rf . Its image is (Rf )hom = GsmarR (f ) ⊂ GZ(f ). (c) The homomorphism ()hom : Rf → GsmarR (f ) is an isomorphism. (d) − id /∈ GsmarR (f ) ⇐⇒ mult f ≥ 3. (77) 26 FALKO GAUSS AND CLAUS HERTLING Equivalently: GmarR (f ) = GsmarR (f ) if mult f = 2, and GmarR (f ) = GsmarR (f ) × {± id} if mult f ≥ 3. n+1). (e) GmarR (f ) = GmarR (f + x2 (f ) Mh ∈ Gsmar(f ). If f is quasihomogeneous then Mh ∈ GsmarR (f ). (g) For any [(f, ρ)] ∈ M smar µ StabGZ([(f, ρ)]) = ρ ◦ GsmarR (f ) ◦ ρ−1, StabGZ([(f,±ρ)]) = ρ ◦ GmarR (f ) ◦ ρ−1. (78) (79) ( (78) does not require assumption (66) or (67)). As GZ acts properly (f0), GsmarR (f ) and GmarR (f ) are finite. (But discontinuously on M mar this follows already from the finiteness of Rf and (b).) µ In the case of a quasihomogeneous singularity the group Rf has a canonical lift to Rf . It will be useful for the calculation of Rf . Theorem 5.9. [He6, theorem 13.11] Let f ∈ C[x0, ..., xn] be a quasi- homogeneous polynomial with an isolated singularity at 0 and weights 2] and weighted degree 1. Suppose that w0 ≤ ... ≤ w0, ..., wn ∈ Q ∩ (0, 1 wn−1 < 1 2). Let Gw be the alge- braic group of quasihomogeneous coordinate changes, that means, those which respect C[x0, ..., xn] and the grading by the weights w0, ..., wn on it. Then 2 (then f ∈ m3 if and only if wn < 1 ∼= StabGw(f ). Rf (80) Finally we need and we want to study period maps and Torelli type problems for singularities. This story should start with the definition of the Gauss-Manin con- nection and the Brieskorn lattice for an isolated hypersurface singular- ity. This had been developed in many papers of the second author, and also much earlier by Brieskorn, K. Saito, G.-M. Greuel, F. Pham, A. Varchenko, M. Saito and others. As we will build here on calculations done in [He1] and therefore never have to touch Brieskorn lattices explicitly, we take here a formal point of view and refer to [He1], [He2], [He4], [He6] and [He7] for the definitions of the following objects. Any singularity f comes equipped with a Brieskorn lattice H(cid:48)(cid:48) 0 (f ). It is much richer than, but still comparable to a Hodge structure of a closed Kahler manifold. After fixing a reference singularity f0, a marked singularity (f,±ρ) comes equipped with a marked Brieskorn lattice BL(f,±ρ). A clas- sifying space DBL(f0) for marked Brieskorn lattices was constructed in [He4]. It is especially a complex manifold, and GZ acts properly discontinuously on it. µ-CONSTANT MONODROMY GROUPS FOR SOME SINGULARITIES 27 Theorem 5.10. Fix one reference singularity f0. (a) There is a natural holomorphic period map BL : M mar µ (f0) → DBL(f0). (81) It is GZ-equivariant. (b) [He6, theorem 12.8] It is an immersion, here the reduced complex structure on M mar (f0) is considered. (The second author has also a proof that it is an immersion where the canonical complex structure on M mar (f0) is considered, but the proof is not written). µ µ The second main conjecture in [He7] is part (a) of the following conjecture. Part (a) and part (b) are global Torelli type conjectures. Conjecture 5.11. Fix one reference singularity f0. µ → DBL is injective. (a) The period map BL : M mar /GZ → DBL/GZ is injective. (b) The period map LBL : Mµ = M mar (c) For any singularity f in the µ-homotopy class of f0 and any µ marking ρ, StabGZ([(f,±ρ)]) = StabGZ(BL([(f,±ρ)])) (only ⊂ and the finiteness of both groups are clear). (82) The second author has a long-going project on Torelli type conjec- tures. Already in [He1], part (b) was conjectured and proved for all simple and unimodal singularities and almost all bimodal singularities (all except 3 subseries of the 8 bimodal series). This was possible with- out the general construction of Mµ and DBL, which came later in [He6] and [He4]. In the concrete cases considered in [He1], it is easy to iden- tify a posteriori the spaces Mµ and DBL. We will make use of that in the sections 6 and 7. Part (c) was conjectured in [He6]. The following lemma from [He7] clarifies the logic between the parts (a), (b) and (c) of conjecture 5.11. Lemma 5.12. In conjecture 5.11, (a) ⇐⇒ (b) and (c). Part (a) of conjecture 5.11 was proved in [He7] for the simple and those 22 of the 28 exceptional unimodal and bimodal singularities, where all eigenvalues of the monodromy have multiplicity one. Here it will be proved for the remaining unimodal and the remaining excep- tional bimodal singularities. As part (b) of conjecture 5.11 was already proved in all these cases in [He1], the main work in [He7, section 8] and here is the control of the group GZ. This is carried out here in the sections 3 and 4, and it is surprisingly difficult. 28 FALKO GAUSS AND CLAUS HERTLING 6. Gmar, M mar and a strong Torelli result for the simple µ elliptic and the hyperbolic singularities The 1-parameter families of the hyperbolic singularities of type Tpqr (p, q, r ∈ N≥2, p ≥ q ≥ r, κ := 1 r < 1) have as surface singularities the normal forms [AGV1] p + 1 q + 1 (83) xp + yq + zr + t · xyz, t ∈ X := C∗. (cid:101)E6, T442 = (cid:101)E7, T632 = (cid:101)E8 have as surface singularities different nor- The 1-parameter families of the simple elliptic singularities T333 = mal forms [SaK]. The normal form xp + yq + zr + t· xyz does in the case of T442 not contain representatives of all right equivalence classes, the class with j-invariant j = 1 is missing [SaK, 1.11, Bem. (ii)]. Therefore we work in the following also with the Legendre normal forms t ∈ X := C − {0, 1}, t ∈ X := C − {0, 1}, t ∈ X := C − {0, 1}. y(y − x)(y − λx) − xz2, yx(y − x)(y − λx) + z2, y(y − x2)(y − λx2) + z2, T333 : T442 : T632 : (84) They contain representatives of all right equivalence classes. Let X univ denote the universal covering of X, so X univ = C if κ < 1 and X univ = H if κ = 1. Theorem 6.1. (a) For the simple elliptic singularities and the hyper- bolic singularities in any number of variables, the space M mar of right ∼= X univ, so it is equivalence classes of marked singularities is M mar µ → DBL connected, and thus Gmar = GZ. The period map BL : M mar is an isomorphism, so the strong global Torelli conjecture 5.11 (a) is true. if r = 2 and as surface singularities if r ≥ 3. Then (b) Now consider the singularities of type Tpqr as curve singularities µ µ GZ = Gmar = Gsmar × {± id}, equivalently: − id /∈ Gsmar. (85) The subgroup of Gsmar, which acts trivially on M mar the surjective map µ , is the kernel of Gsmar → Aut(M l(f0)1,Z, L)/{± id}. It is equal to ρ ◦ (Rf )hom ◦ ρ−1 for a generic [(f, ρ)] ∈ M mar is 54, 16 and 6 for T333, T442 and T632. µ (86) . Its size Proof: (a) The proof uses two Torelli type results from [He1]. We choose a marked reference singularity [(f0,± id)] in X univ, then all elements of X univ become marked singularities, because X univ is simply connected. Then the period map X univ → DBL is well defined. µ-CONSTANT MONODROMY GROUPS FOR SOME SINGULARITIES 29 The first Torelli type result from [He1] is that this map is an isomor- phism. Therefore the marked Brieskorn lattices of the marked singularities in X univ are all different. Therefore the marked singularities in X univ are all not right equivalent. This gives an embedding X univ (cid:44)→ M mar (f0)0. (f0)0 → DBL, which is an immersion. As it restricts to the isomorphism X univ → DBL, finally X univ = M mar On the other hand, we have the period map BL : M mar (f0)0. µ µ nected and M mar For part (a) it rests to show that Gmar = GZ. Then M mar is con- µ → DBL is an isomorphism. We have to look closer at DBL and the action of GZ on it. In the µ = X univ, and BL : M mar µ µ case κ < 1, DBL ∼= {V ⊂ M l(f0)1 dim V = 1, V (cid:54)= ker(Mh − id)}. (87) In the case κ = 1, DBL ∼= one component of {V ⊂ M l(f0)1 dim V = 1, V (cid:54)= V }. (88) In both cases the group Aut(M l(f0)1,Z, L)/{± id} acts faithfully on DBL. In both cases the second Torelli type result from [He1] which we need is that the period map X univ/ ∼R → DBL/ Aut(M l(f0)1,Z, L) (89) is an isomorphism. Here ∼R denotes right equivalence for unmarked singularities. Because of X univ = M mar (f0)0, µ X univ/ ∼R= M mar µ (f0)0/Gmar. The isomorphism (89) and the isomorphism M mar DBL show that the map µ Gmar → Aut(M l(f0)1,Z, L) (90) (f0)0 = X univ → (91) is surjective. This completes the proof of part (b) of theorem 3.1. It also shows that for proving Gmar = GZ, it is sufficient to show that the kernels of the maps to Aut(M l(f0)1,Z, L)/{± id} coincide. The kernel of the map GZ → Aut(M l(f0)1,Z, L)/{± id} was determined in In fact, this is the only part of theorem 3.1 which we theorem 3.1. need here. It consists of those elements of (U1 (cid:111) U2) × {± id} in (22) for which δ = 0, so it is isomorphic to the group {(α, β, γ) ∈ Zp × Zq × Zr α p ≡ 0 mod 1} (cid:111) U2 × {± id} (cid:19) (cid:18) β q γ r + + =: (U 0 1 (cid:111) U2) × {± id}. (92) 30 FALKO GAUSS AND CLAUS HERTLING The kernel of the map Gmar → Aut(M l(f0)1,Z, L)/{± id} is the sub- It is the isotropy (f0)0. So by theorem group of Gmar which acts trivially on M mar (f0)0. group in Gmar of a generic point [(f,±ρ)] ∈ M mar 5.8 (g) it is the group µ µ ρ ◦ GmarR (f ) ◦ ρ−1 = ρ ◦ {±ϕhom ϕ ∈ Rf} ◦ ρ−1. (93) As f is generic, we can and will use now the normal form f = xp + yq + zr + t · xyz, also in the case κ = 1. The following coordinate changes generate a finite subgroup S ⊂ Rf ϕα,β,γ : (x, y, z) (cid:55)→ (e2πiα/px, e2πiβ/qy, e2πiγ/rz) (94) + ≡ 0 mod 1, + γ r with α p (cid:55)→ (y, x, z) (cid:55)→ (x, z, y) (cid:55)→ (x, y,−z − txy) β q if p = q, if q = r, if r = 2. ϕ1,2 : (x, y, z) ϕ2,3 : (x, y, z) ϕminus : (x, y, z) ϕminus has order 2 and commutes with the other coordinate changes (q = r = 2 is impossible because of κ ≤ 1). The group S is isomorphic (cid:111) U2) × {± id} (as an abstract group) to U 0 1 if r = 2. The map to 1-jets of coordinate changes is injective, (cid:111) U2 if r ≥ 3 and to (U 0 1 ∼=−→ j1S ⊂ j1Rf ⊂ j1R. S (95) Now we have to treat the cases r ≥ 3 and r = 2 separately. The case r ≥ 3: Then j1Rf is finite and isomorphic to Rf , the map ()hom : Rf → GZ(f ) = ρ−1 ◦ GZ ◦ ρ is injective, and the image GsmarR (f ) does not contain − id (theorem 5.8). Therefore then S ∼= (S)hom ⊂ GZ(f ) and − id /∈ (S)hom. Thus the group (S)hom×{± id} is isomorphic to (U 0 (cid:111)U2)×{± id}. Now it is clear that the group in (93) is at least as big as the group in (92). But it cannot be bigger. So they are of equal size. This implies Gmar = GZ. The case r = 2: We claim that the map S → (S)hom is injective. If (cid:111) U2) × {± id}, and this is of equal size this is true then (S)hom as the group in (92). Then again the group in (93) is at least as big as the group in (92), but it cannot be bigger. So they are of equal size. This implies Gmar = GZ. ∼= (U 0 1 1 It rests to prove the claim. For this we consider the curve singularity g := xp + yq − 1 4 tx2y2. (96) µ-CONSTANT MONODROMY GROUPS FOR SOME SINGULARITIES 31 Then g + z2 = f ◦ ψ with ψ(x, y, z) = (x, y, z − 1 2 Rg+z2 = ψ−1 ◦ Rf ◦ ψ, txy), (97) ψ−1 ◦ ϕα,β,γ ◦ ψ = ϕα,β,γ, ψ−1 ◦ ϕ1,2 ◦ ψ = ϕ1,2, if p = q, ψ−1 ◦ ϕminus ◦ ψ = ((x, y, z) (cid:55)→ (x, y,−z)). (q = r = 2 is impossible because of κ ≤ 1). The subgroup 1} (cid:111) U2 Scurve := {ϕα,β,γ ◦ (ϕminus)−γ (α, β, γ) ∈ U 0 (98) has index 2 in S, its conjugate ψ−1 ◦ Scurve ◦ ψ restricts to Rg, and it maps injectively to j1Rg ∼= Rg. By theorem 5.8 (c) the map Scurve → (Scurve)hom is injective, and − id is not in the image. But (ϕminus)hom = − id. This proves the claim. (b) By theorem 5.5 (c), the projection M smar µ = H if κ = 1, M smar µ = C if κ < 1 and M mar is a twofold covering, and − id exchanges the two sheets of the covering. Because of M mar has two components. Therefore − id /∈ Gsmar and GZ = Gmar = Gsmar × {± id}. µ µ → M mar µ used in the proof of part (a). The statements right before and after (86) were already proved and The group ρ◦ (Rf )hom ◦ ρ−1 for a generic [(f, ρ)] ∈ M mar has size 54, 16 and 6 for T333, T442 and T632, because it is isomorphic to an index 2 subgroup of the group in (92), and that group has 108, 32 and 12 (cid:3) elements in the cases T333, T442 and T632 µ 7. Gmar, M mar µ and a strong Torelli result for 6 of the 28 exceptional unimodal and bimodal singularities Normal forms for the 1-parameter families of the exceptional unimodal and bimodal singularities of types Z12, Q12, U12, Z18, Q16 and U16 in the minimal number of variables are as follows [AGV1]. Here Z12 and Z18 are curve singularities, Q12, U12, Q16 and U16 are surface singularities. The singularity for the parameter t = 0 is quasihomogeneous, the oth- ers are semiquasihomogeneous. The space of the parameter t = t1 or t = (t1, t2) is X = Cmod(f0). The weights w = (wx, wy) respectively w = (wx, wy, wz) are normalized such that degw f0 = 1. 32 FALKO GAUSS AND CLAUS HERTLING normal form Z12 : ft = x3y + xy4 + tx2y3 Q12 : ft = x3 + y5 + yz2 + txy4 U12 : ft = x3 + y3 + z4 + txyz2 Z18 : ft = x3y + xy6 + (t1 + t2y)y9 Q16 : ft = x3 + y7 + yz2 + (t1 + t2y)xy5 U16 : ft = x3 + xz2 + y5 + (t1 + t2y)x2y2 mod(f0) weights 11, 2 ( 3 11) ( 1 3, 1 5, 2 5) 3, 1 ( 1 3, 1 4) 17, 2 ( 5 17) 7, 3 ( 1 3, 1 7) ( 1 3, 1 5, 1 3) 1 1 1 2 2 2 (99) The normal form of the quasihomogeneous singularity of type Q12, Q16, U12 and U16 is a sum of an Al-singularity in one variable and a D2m-singularity in two variables with (l, 2m) as in table (100)=(58). (l, 2m) Q12 (2, 6) A2 ⊗ D6 A2 ⊗ D8 A3 ⊗ D4 A4 ⊗ D4 U12 (3, 4) U16 (4, 4) Q16 (2, 8) (100) The rest of this section is devoted to the proof of the following the- orem. Theorem 7.1. (a) For the 6 families of exceptional unimodal and bi- modal singularities of types Z12, Q12, U12, Z18, Q16 and U16 in any num- ber of variables, the space M mar of right equivalence classes of marked ∼= X = Cmod(f0), so it is connected, and thus singularities is M mar µ → DBL is an isomorphism, Gmar = GZ. The period map BL : M mar so the strong global Torelli conjecture 5.11 (a) is true. µ µ (b) Now consider the singularities of type Z12 and Z18 as curve sin- gularities and the singularities of types Q12, U12, Q16 and U16 as surface singularities. Then their multiplicities are ≥ 3. Then GZ = Gmar = Gsmar × {± id}, equivalently: − id /∈ Gsmar. (101) Proof: (a) The proof is similar to the proof of theorem 6.1, but simpler. We need only the first of the two Torelli type results from [He1], which were used in the proof of theorem 6.1. We choose as marked reference singularity the quasihomogeneous singularity with trivial marking [(f0,± id)] in X. Then all elements of X become marked singularities, because X is simply connected. Then the period map X → DBL is well defined. A Torelli type result from [He1] says that this map is an isomorphism. It is in fact easy, because the singularities here are semiquasihomogeneous and only f0 is quasihomogeneous. That makes the calculations easy. µ-CONSTANT MONODROMY GROUPS FOR SOME SINGULARITIES 33 Therefore the marked Brieskorn lattices of the marked singularities X are all different. Therefore the marked singularities in X are all not right equivalent. This gives an embedding X (cid:44)→ M mar (f0)0. (f0)0 → DBL, which is an immersion. As it restricts to the isomorphism X → DBL, finally X = M mar On the other hand, we have the period map BL : M mar µ µ (f0)0. µ For part (a) it rests to show that Gmar = GZ. Then M mar is con- nected and M mar µ = X, and BL : M mar µ → DBL is an isomorphism. µ The weights of the deformation parameter(s) t1 (and t2) equip the µ with a good C∗-action. It commutes with parameter space X = M mar the action of GZ. This gives the first equality in Gmar = StabGZ([(f0,± id)]) = GmarR (f0). (102) The second equality is part of theorem 5.8 (g). On the other hand, by theorem 5.8 (d), GsmarR (f0) × {± id} = GmarR (f0). But because f0 is a quasihomogeneous singularity of degree ≥ 3, the group GsmarR (f0) can be calculated easily via StabGw(f0), see ∼=−→ GsmarR (f0). Therefore theorem 5.8 (c) and theorem 5.9: StabGw(f0) it is sufficient to show that StabGw(f0) has half as many elements as the group GZ. We postpone its proof. If it holds, then GZ = Gmar = GmarR (f0) = GsmarR (f0) × {± id} (103) follows, and part (a) of the theorem is proved. For part (b), the same argument as in the proof of theorem 6.1 (b) works: M smar is a twofold covering of M mar , and the two sheets are exchanged by the action of − id. As X = Cmod(f0), M smar has two components, and − id /∈ M smar . h k ∈ Z} × U In theorem 4.1 it was shown that GZ is GZ = {±M k µ µ µ µ with U as in table (39)=(104). U ∼= {id} S2 Z12 Q12 U12 Z18 Q16 U16 S3 {id} S2 S3 (104) Now we compare StabGw(f0). It is sufficient to find enough elements so that the resulting group has half as many elements as GZ. The cases Z12 and Z18: Then ϕ1 : (x, y) (cid:55)→ (e2πiwxx, e2πiwy y) satisfies This is already sufficient. Here Gsmar = GsmarR = {M k The cases Q12, Q16, U12, U16: Here it is convenient to make use of the decomposition of the singularity f0 into a sum of an Al singularity g0 in one variable and a D2m singularity h0 in two variables. In all four cases the weight system w(cid:48) of the Al singularity and the weight system (ϕ1)hom = Mh. h k ∈ Z}. (105) StabGw(f0) = StabGw(cid:48) (g0) × StabGw(cid:48)(cid:48) (h0) (cid:26) Z2m−1 × S2 Z3 × S3 ∼= Zl+1 × if m ≥ 3 if m = 2. (106) (cid:3) 34 FALKO GAUSS AND CLAUS HERTLING w(cid:48)(cid:48) of the D2m singularity have denominators l + 1 and 2m − 1 with gcd(l + 1, 2m − 1) = 1. Therefore In all four cases this group has half as many elements as GZ. 8. More on GZ for the simple elliptic singularities This section is motivated by the paper [MS] of Milanov and Shen. They consider the 1-parameter families xp + yq + zr + t · xyz, t ∈ Σ ⊂ C the simple elliptic (107) ∈ of {(3, 3, 3), (4, 4, 2), (6, 3, 2)} which had also been used above in section 6. Here Σ ⊂ C is the complement of the finite set of pa- rameters where the function in (107) has a non-isolated singularity. Remark that now χ := lcm(p, q, r) = p. singularities Tpqr with (p, q, r) In [MS] the groups of the transversal monodromies of these three families are studied, more precisely, the natural representations ρ : π1(Σ) → GZ, ρ1 : π1(Σ) → Aut(M l(f )1,Z, L), ρ(cid:54)=1 : π1(Σ) → Aut(M l(f )(cid:54)=1,Z, L), ρ(cid:54)=1 : π1(Σ) → Aut(M l(f )(cid:54)=1,Z, L)/(cid:104)Mh(cid:105). By explicit computations they show ker(ρ1) ⊂ ker(ρ(cid:54)=1). (108) (109) They ask about a conceptual explanation of (109) and whether this might be true for other normal forms, i.e. other natural 1-parameter families of the simple elliptic singularities. Because of (109), there is an induced representation ρW : Im(ρ1) → Aut(M l(f )(cid:54)=1, L)/(cid:104)Mh(cid:105). (110) Then ker(ρW ) ⊂ Im(ρ1) ⊂ Aut(M l(f )1,Z, L). One main result of [MS] is the following. Theorem 8.1. Aut(M l(f )1,Z, L) ∼= SL(2, Z) as in (18), the subgroup ker(ρW ) is an isomorphism In all cases, under three isomorphic to the principal congruence subgroup Γ(p). µ-CONSTANT MONODROMY GROUPS FOR SOME SINGULARITIES 35 p , 1 q , 1 Here Γ(N ) := {A ∈ SL(2, Z) A ≡ 12 mod N} (not A ≡ ±12 mod N ) is the principal congruence subgroup of level N of SL(2, Z). In the following, we give results which complement (109) and theo- rem 8.1. We consider not a special 1-parameter family, but the biggest possible family of quasihomogeneous simple elliptic singularities. De- fine (wx, wy, wz) := ( 1 r ), define for any monomial its weighted de- gree degw(xαyβzγ) := αwx + βwy + γwz, and define C[x, y, z]1 := (cid:104)xαyβzγ degw(xαyβzγ) = 1(cid:105)C, R := {f ∈ C[x, y, z]1 f has an isolated singularity at 0}. the complement of a hypersurface in the vector R is space C[x, y, z]1. The transversal monodromies of the family of singulari- ties parametrized by R give the natural representation σ, the other representations are induced, (111) σ : π1(R) → GZ, σ1 : π1(R) → Aut(M l(f )1,Z, L), σ(cid:54)=1 : π1(R) → Aut(M l(f )(cid:54)=1,Z, L), σ(cid:54)=1 : π1(R) → Aut(M l(f )(cid:54)=1,Z, L)/(cid:104)Mh(cid:105). (112) By the definition of Gsmar in [He7, definition 3.1], Gsmar = Im(σ). The- orem 6.1 tells that the monodromy group Im(σ) is as large as possible (up to ± id in the case T333). Theorem 8.2. Im(σ) = GZ in the cases T442 and T632, and GZ = Im(σ) × {± id} in the case T333 (here it is important that the surface singularities are considered). The explicit information on GZ in theorem 3.1 allows the following conclusion. Corollary 8.3. The analogue ker(σ1) ⊂ ker(σ(cid:54)=1) of (109) does not hold in the cases T333 and T442. It holds in the case T632, and there the analogue of (109) holds for any µ-constant family. Proof: By theorem 3.1 (c), {g ∈ GZ gM l(f )1 = id} = {idM l(f )1 × (MhM l(1)C )α × (MhM l(2)C )β × (MhM l(3)C )γ (α, β, γ) ∈ Zp × Zq × Zr with α p + β q + γ r ≡ 0 mod 1} × U2. In the cases T442 and T632 this is isomorphic to ker(σ1)/ ker(σ), in the case T333 ker(σ1)/ ker(σ) is isomorphic to this group or a subgroup of 36 FALKO GAUSS AND CLAUS HERTLING this group of index 2. In the cases T442 and T333, already the factor U2 in this group is an obstruction to the analogue of (109). In the case T632, U2 is trivial and {g ∈ GZ gM l(f )1 = id} = {M α h α ∈ Z} (113) (because above α determines β and γ uniquely in the case (p, q, r) = (6, 3, 2)). Therefore the analogue of (109) holds in the case T632 for the (cid:3) family parametrized by R and for any subfamily. (109) is used in [MS] in order to define ρW and the group ker(ρW ). But because MhM l(f )1 = id, it is obvious that the group ker(ρW ) coin- cides with {gM l(f )1,Z g ∈ Im(σ), gM l(f )(cid:54)=1,Z = id} ⊂ Aut(M l(f )1,Z, L). (114) And the analogue of this group can be defined for any µ-constant family, whether or not it satisfies the analogue of (109). Our main result in this section is theorem 8.4. Our proof uses theorem 3.1. A different proof of theorem 8.4 was given by Kluitmann in [Kl, III 2.4 Satz, page 66]. Theorem 8.4 shows that the group Γ(p) turns up naturally within the maximal possible µ-constant family, which is parametrized by R, and it shows the part ker(ρW ) ⊂ Γ(p) of the equality ker(ρW ) = Γ(p) in theorem 8.1. Theorem 8.4. Aut(M l(f )1,Z, L) ∼= SL(2, Z) as in (18), the subgroup In all cases, under {gM l(f )1,Z g ∈ GZ, gM l(f )(cid:54)=1,Z = id} ⊂ Aut(M l(f )1,Z, L). three an isomorphism is isomorphic to the principal congruence subgroup Γ(p). Proof: We use the notations and objects in the proof of theorem 3.1. M l(1) C was defined in (31). Define M l(1) Z := M l(1) C ∩ M l(f ), (115) and analogously M l(2) Z and M l(3) Z . Then M l(1) Z = (Mh − id)(M l[1] Z ). Because of (32), this image is generated as a Z-lattice by δ2 − (δ1 +(cid:101)b1), δ3 − δ2, ..., δp−1 − δp−2,−(δ1 + ... + δp−1) − δp−1, (116) respectively by δ1 + (p − 1)δ2, δ3 − δ2, ..., δp−1 − δp−2, pδ2 − δµ−1 + δµ. (117) µ-CONSTANT MONODROMY GROUPS FOR SOME SINGULARITIES 37 Z and M l(3) M l(2) r = 2 then Z (if r ≥ 3) are generated by the analogous elements. If M l(3) Z = Z · (2δµ−2 − δµ−1 + δµ). (118) In any case, the sum M l(1) Z is a sublattice of finite index of the primitive sublattice M l(f )(cid:54)=1,Z in M l(f ). Observe that qp and rp in all three cases. The lattice M l(f )(cid:54)=1,Z is generated by Z ⊕ M l(2) Z ⊕ M l(3) δ1 + (p − 1)δ2, δ3 − δ2, ..., δp−1 − δp−2, pδ2 − δµ−1 + δµ, (119) δp + (q − 1)δp+1, δp+2 − δp+1, ..., δp+q−2 − δp+q−3, δp+q−1 + (r − 1)δp+q, δp+q+1 − δp+q, ..., δµ−2 − δµ−3, δ2 − δp+1, p q δ2 − δp+q, p r if r ≥ 3. If r = 2 then the third line has to be replaced by δ2 − δµ−2. p r In any case, one sees M l(f ) = M l(f )(cid:54)=1,Z ⊕ Z · δ2 ⊕ Z · δµ. One also calculates with γ1 := −(pδ2 − δµ−1 + δµ) ∈ M l(f )(cid:54)=1,Z, with γ2 (q − i)δp−1+i (cid:101)b1 = δµ−1 − δµ = γ1 + pδ2 (cid:101)b2 = γ2 + pδµ, p−1(cid:88) q−1(cid:88) r−1(cid:88) (p − i)δi + p q := i=1 i=1 p r + i=1 (120) (121) (122) (r − i)δp+q−2+i + pδµ−1 − pδµ ∈ M l(f )(cid:54)=1,Z. M l(f ) ∩ (Q · γ1 + Q · γ2) = Z · γ1 ⊕ Z · γ2. (123) ∈ SL(2, Z) define the automorphism f : One sees also (cid:18)a b f ((cid:101)b1) := a(cid:101)b1 + c(cid:101)b2, For any matrix c d M l(f )Q → M l(f )Q by (cid:19) f ((cid:101)b2) := b(cid:101)b1 + d(cid:101)b2, (124) It respects L because the decomposition M l(f )Q = M l(f )1,Q ⊕ M l(f )(cid:54)=1,Q is left and right orthogonal with respect to L. It is an fM l(f )(cid:54)=1,Z := id . 38 FALKO GAUSS AND CLAUS HERTLING automorphism of M l(f ) if and only if f (δ2) and f (δµ) are in M l(f ). One calculates f (δ2) = aδ2 + cδµ + f (δµ) = bδ2 + dδµ + bγ1 + (a − 1)γ1 + 1 p cγ2, (d − 1)γ2. 1 p 1 p 1 p (cid:18)a b (cid:19) c d In view of (123) this shows f ∈ GZ ⇐⇒ (cid:18)a b (cid:19) c d ≡ 12 mod p def.⇐⇒ ∈ Γ(p). (cid:3) References [AGV1] V.I. Arnold, S.M. Gusein-Zade, A.N. Varchenko: Singularities of differ- entiable maps, volume I. Birkhauser, Boston, Basel, Stuttgart, 1985. [AGV2] V.I. Arnold, S.M. Gusein-Zade, A.N. Varchenko: Singularities of differ- entiable maps, volume II. Birkhauser, Boston 1988. [Eb1] W. Ebeling: Milnor lattices and geometric bases of some special singulari- ties. In: Noeuds, tresses et singularit´es, Monographie No 31 de l'enseignement Math´ematique, Gen`eve, 1983, 129 -- 146. [Eb2] W. Ebeling: Functions of several complex variables and their singularities. Graduate Studies in Mathematics 83, AMS, 2007. [Ga] A.M. Gabrielov: Dynkin diagrams of unimodal singularities. Funct. Anal. Appl. 8 (1974), 192 -- 196. [He1] C. Hertling: Analytische Invarianten bei den unimodularen und bimod- ularen Hyperflachensingularitaten. Doctoral thesis. Bonner Math. Schriften 250, Bonn 1992. [He2] C. Hertling: Ein Torellisatz fur die unimodalen und bimodularen Hy- perflachensingularitaten. Math. Ann. 302 (1995), 359 -- 394. [He3] C. Hertling: Brieskorn lattices and Torelli type theorems for cubics in P3 and for Brieskorn-Pham singularities with coprime exponents. In: Sin- gularities, the Brieskorn anniversary volume. Progress in Mathematics 162. Birkhauser Verlag, Basel-Boston-Berlin 1998, pp. 167 -- 194. [He4] C. Hertling: Classifying spaces and moduli spaces for polarized mixed Hodge structures and for Brieskorn lattices. Compositio Math. 116 (1999), 1 -- 37. [He5] C. Hertling: Generic Torelli for semiquasihomogeneous singularities. In: Trends in singularities (eds. A. Libgober, M. Tibar), Birkhauser Verlag, Basel 2002, 115 -- 140. [He6] C. Hertling: Frobenius manifolds and moduli spaces for singularities. Cam- bridge Tracts in Mathematics 151, Cambridge University Press, 2002. [He7] C. Hertling: µ-constant monodromy groups and marked singularities. Ann. Inst. Fourier, Grenoble 61.7 (2011), 2643-2680. [Ka] R. Kaenders: The Seifert form of a plane curve singularity determines its intersection multiplicities. Indag. Mathem. N.S. 7.2 (1996), 185-197. [Kl] P. Kluitmann: Ausgezeichnete Basen erweiterter affiner Wurzelgitter. Doc- toral thesis. Bonner Math. Schriften 185, Bonn 1987. µ-CONSTANT MONODROMY GROUPS FOR SOME SINGULARITIES 39 [Ku] Va.S. Kulikov: Mixed Hodge structures and singularities. Cambridge tracts in mathematics 132, Cambridge University Press, 1998. [LR] Le Dung Tr´ang, C.P. Ramanujam: The invariance of Milnor's number im- plies the invariance of the topological type. Amer. J. of Math. 98 (1973), 67 -- 78. Sur le role de la monodromie enti`ere dans la topologie [MW] F. Michel, C. Weber: des singularit´es. Ann. Inst. Fourier Grenoble 36 (1986), 183 -- 218. [MS] T. Milanov, Y. Shen: The modular group for the total ancestor potential of Fermat simple elliptic singularities. Comm. Number Theor. Phys. 8 (2014), 329 -- 368. [Mi] J. Milnor: Singular points of complex hypersurfaces. Ann. Math. Stud. vol. 61, Princeton University Press 1968. [Or] P. Orlik: On the homology of weighted homogeneous polynomials. In: Lec- ture Notes in Math. 298, Springer, Berlin, 1972. [SaK] K. Saito: Einfach-elliptische Singularitaten. Invent. Math. 23 (1974), 289 -- 325. [SaM] M. Saito: Period mapping via Brieskorn modules. Bull. Soc. math. France 119 (1991), 141 -- 171. Falko Gauss, Universitat Mannheim, Lehrstuhl fur Mathematik VI, Seminargebaude A 5, 6, 68131 Mannheim, Germany E-mail address: [email protected] Claus Hertling, Universitat Mannheim, Lehrstuhl fur Mathematik VI, Seminargebaude A 5, 6, 68131 Mannheim, Germany E-mail address: [email protected]
1807.02517	1	1807	2018-07-06T10:12:37	Some remarks on Dupont contraction	[ "math.AG", "math.CT" ]	We present an alternative equivalent description of Dupont's simplicial contraction: it is an explicit example of a simplicial contraction between the simplicial differential graded algebra of polynomial differential forms on standard simplices and the space of Whitney elementary forms.	math.AG	math	SOME REMARKS ON DUPONT CONTRACTION LUIGI LUNARDON Abstract. We present an alternative equivalent description of Dupont's simplicial contraction: it is an explicit example of a simplicial contraction between the simplicial differential graded algebra of polynomial differential forms on standard simplices and the space of Whitney elementary forms. 1. Introduction In [6] Dupont gave an explicit description of simplicial contraction from the simplicial differential graded algebra Ω• of polynomial differential forms on the affine standard simplices to the subspace of Whitney elementary forms, a simplicial finite dimensional differential graded vector subspace of the former. More precisely, the Dupont contraction is a morphism of simplicial abelian groups h : Ω• → Ω• such that dh + hd = r − Id, where r is the classical Whitney's retraction of Ω• onto the subspace of elementary forms: in this paper we recall the general notion of contraction in Section 2 and we describe the simplicial map h in Section 3. Although Dupont contraction can be used to give alternative proofs of some classical results, such as the polynomial De Rham's theorem ([3, Theorem 2.2], [5] and [8, Theorem 10.15]), their most relevant use is given in combination with homological perturbation theory [13] and homotopy transfer of ∞-structures (see e.g. [2, 11, 14]). For instance, Dupont's contraction was used in [4] to construct a canonical C∞ structure on the normalized cochain complex of a cosimplicial commutative algebra over a field of characteristic 0. Similarly, in [10] the authors used it to induce a canonical L∞ structure on the normalized cochain complex of a cosimplicial differential graded Lie algebra: in the same paper some explicit computation is done in the particular case of cosimplicial Lie algebra, while the particular case of a single morphism of differential graded Lie algebras (consid- ered as a cosimplicial object via Kan extension) was previously considered and deeply investigated in [9]. It is also worth to mention the application of Dupont's contraction to Hodge theory of complex algebraic varieties [16]. Dupont's Theorem and the homotopy transfer theorem are also key tools in [12] and [1]. In these two papers the authors study the Deligne ∞-groupoid associated to an L∞-algebra, its relation with the Maurer-Cartan elements of that algebra and the behaviour of the Deligne ∞-groupoid under totalization and 1 2 LUIGI LUNARDON homotopy limits. In particular Dupont's contraction is used to construct a Kan complex that is quasi-isomorphic to the simplicial set of Maurer-Cartan elements. The original construction by Dupont provides a family of maps which is really hard to compute. An apparently different simplicial contraction k : Ω• → Ω• with the same key properties of Dupont's contraction, but somewhat easier to handle, was proposed by M. Manetti during a cycle of seminars on deformation theory given at Roma in 2011, leaving unsettled the question whether k = h. The main result of this paper is to give a positive answer to the above question, and then to give an alternative equivalent definition of Dupont's contraction. In Section 4 we describe the map k and we reproduce Manetti's (unpublished) proof that it is indeed a simplicial object in the category of contractions. Finally, in Section 5 we prove the equality k = h. Acknowledgement. I would like to thanks Prof. M. Manetti for his help during the (slow) preparation of this paper. This work was supported by the Engineering and Physical Sciences Research Council [EP/L015234/1]. The EPSRC Centre for Doctoral Training in Geometry and Number Theory (The London School of Geometry and Number Theory), University College London. 2. Simplicial contraction In this section we describe the category of contractions of DG-vector spaces and we recall the definition of simplicial and cosimplicial objects in any given category. Let K be a field of characteristic 0, a DG-vector space over K is a graded vector space endowed with a linear map d of degree 1 such that d2 = 0. Definition 2.1. A contraction of DG-vector spaces is a diagram h / N , M i πo with M and N two DG-vector spaces over K, h ∈ Hom−1 morphisms of DG-vector spaces. Moreover, we require the following relations: K (N, N ) and i, π two πi = IdM , iπ − IdN = dN h + hdN . Remark 2.2. The maps π and i are respectively injective and surjective, since πi = IdM . Moreover, it follows from the relation iπ − IdN = dN h + hdN that they are both quasi-isomorphisms. Remark 2.3. Suppose also that the additional conditions h2 = πh = 0 hold. From the identity iπ − IdN = dN h + hdN we obtain the identities: −h = hdN h; hiπ − h = hdN h. It follows that hiπ = 0 and since π is surjective, then hi = 0. Similarly the conditions hi = h2 = 0 imply πh = 0. The conditions h2 = πh = hi = 0 are called side conditions. / o SOME REMARKS ON DUPONT CONTRACTION 3 Definition 2.4. A morphism of contractions is a commutative diagram h N i π M k / B j p / A f bf where f : N → B is a morphism of DG-vector spaces such that f h = kf. We denote by bf : M → A the map bf = pf i. Remark 2.5. This definition of morphism doesn't seem natural. However we get the following identities: = f i + f (iπ − IdN )i = f i, jbf = jpf i = (IdB + dBk + kdB)f i = f i + f (dN h + hdN )i bf π = pf iπ = pf + pf (dN h + hdN ) = pf + p(dBk + kdB)f = pf + p(jp − IdB)f = pf. Using these these two identities it follows that the following diagrams commute: M bf A j i N f / B f N π B p M bf / A As a consequence our notion of morphism of contractions is compatible with a couple of morphism f : N → B and g : M → A commuting with every square. The category of contractions of DG-vector spaces over K is denoted by Contr. Definition 2.6. We denote with ∆ the category of finite ordinals. The objects of this category are the finite ordered sets [n] = {0 < · · · < n} and its morphisms are the non decreasing maps. A special role in this category is played by face maps, which are defined as: δk : [n − 1] → [n]; δk(x) =(cid:26) x x + 1 if p < k if p ≥ k , k = 0, . . . , n. Notation 2.7. We denote with I(n, m) ⊂ Mor∆([n], [m]) the subset of injective, and hence strictly monotone, maps. Definition 2.8. A cosimplicial object in a category C is a functor F : ∆ → C; a simplicial object in C is a functor F : ∆op → C. Dupont ([6], Chapter 2) proposed an explicit construction of a simplicial ob- ject in Contr. / O O / O O / / / / / / 4 LUIGI LUNARDON Remark 2.9. The notion of contraction has a few slight variants in literature. In this paper we follow [4] and [12]. The original definition given by Eilenberg and Mac Lane in [7] requires also the side conditions hi = πh = 0. In [15] the object described in Definition 2.1 is called a strong deformation data, and to be a contraction the condition h2 = 0 is required. The conditions hi = πh = h2 = 0 are almost granted: given i, π and h as in Definition 2.1, then we can replace h with h1 = (dh + hd)h(dh + hd); we still have a contraction, but now this contraction satisfies h1i = πh1 = 0. Replacing h1 with h2 = −h1dh1 it is again a contraction and now it satisfies πh2 = h2i = h2 2 = 0 3. Dupont's simplicial contraction In this section we describe the simplicial contraction which Dupont suggested in [6]. The proof that it is actually a contraction is in Section 4 and Section 5. Definition 3.1. The affine standard n-simplex on K is the set: ∆n K = {(x0, . . . , xn) ∈ Kn+1 such that x0 + · · · + xn = 1}. The vertices of ∆n K are the points ei ∈ ∆n K: e0 = (1, 0, . . . , 0), e1 = (0, 1, 0, . . . , 0), . . . , en = (0, . . . , 0, 1). The cosimplicial affine space ∆• the affine standard n-simplex ∆n the affine map K is the functor which associate to each set [n] K and to each non decreasing map f : [n] → [m] f : ∆n K → ∆m K , f (ei) = ef (i). Definition 3.2. The DG-vector space of polynomial differential forms on the affine standard n-simplex is: Ωn = Ωp n = nMp=0 (cid:18) nPk=0 xi − 1, K[x0, . . . , xn, dx0, . . . , dxn] dxi(cid:19) . nPk=0 Here Ωp n denotes the subspace of p-forms, which are the elements of degree p. The simplicial DG-vector space Ω• associates to each [n] the DG-vector space Ωn and to each map f : [n] → [m] in ∆ the pull-back f ∗ : Ωm → Ωn, induced by the affine map f : ∆n K → ∆m K . Next we define a finite dimensional vector subspace Cn ⊂ Ωn, called the space of Whitney elementary forms. As a consequence of Proposition 3.4 it follows that Cn is closed under derivation and thus it is a DG-vector subspace of Ωn. Definition 3.3 (Whitney, [17]). Fix f : [m] → [n] a morphism in ∆. The Whitney elementary form associated to f is the m-form: ωf = m! (−1)ixf (i)dxf (0) ∧ · · · ∧ \dxf (i) ∧ · · · ∧ dxf (m) ∈ Ωm n . We denote with Cn the vector space spanned by Whitney elementary forms. mXi=0 SOME REMARKS ON DUPONT CONTRACTION 5 Proposition 3.4. Let f : [n] → [m] a morphism in ∆. The followings hold: (1) If f is injective f ∗ωf = n!dx1 ∧ · · · ∧ dxn; otherwise ωf = 0; (2) If f is injective for every g : [p] → [m] we have g∗ωf = {h : [n]→[p], gh=f } ωh; P (3) dωf =Pk (−1)k {g : [n+1]→[m], gδk=f } ωg. P In particular Cn is a DG-vector subspace of Ωn and C• is a simplicial DG-vector subspace of Ω•. Proof. Denote by ωi0,...,in the differential form: ωi0,...,in = n! nXk=0 (−1)kxik dxi0 ∧ · · · ∧ ddxik ∧ · · · ∧ dxin ∈ Ωn m. (1) Since ωi0,...,in is alternating on indices, then if f is not injective it follows ωf = 0. Suppose f injective; since we are working on the affine standard simplex we have: f ∗ωf = n! nXk=0 (−1)kxkdx0 ∧ · · · ∧ddxk ∧ · · · ∧ dxn = n! x0dx1 ∧ · · · ∧ dxn − (−1)2k−1xkdx1 ∧ · · · ∧ dxn! nXk=1 = n! (x0 + · · · + xn)dx1 ∧ · · · ∧ dxn = n!dx1 ∧ · · · ∧ dxn. (2) Consider the family of sets Pi = {j ∈ [p] g(j) = f (i)} and note that: (a) since f is injective Pi ∩ Pj = ∅ if i 6= j; (b) g∗(xf (i)) =Pj∈Pi P0 × · · · × Pn. xj , and g∗(dxf (i)) =Pj∈Pi dxj; (c) the functions h : [n] → [p] such that gh = f are in bijection with (3) To prove the last point first we show that dωi0,...,in = (n + 1)!dxi0 ∧ · · · ∧ dxin = Indeed we have: ωi,i0,...,in . mXi=0 dωi0,...,in = n! nXk=0 dxi0 ∧ · · · ∧ dxik ∧ · · · ∧ dxin = (n + 1)!dxi0 ∧ · · · ∧ dxin , and for the second equality: ωi,i0,...,in = (n + 1)! xidxi0 ∧ · · · ∧ dxin − (n + 1) mXi=0 mXi=0 = (n + 1)!dxi0 ∧ · · · ∧ dxin . dxi ∧ ωi0,...,in mXi=0 6 LUIGI LUNARDON Taking f : [n] → [m] we can finally see that: dωf = mXi=0 ωi,f (0),...,f (n) = = nXk=0 nXk=0 (−1)k Xf (k−1)<i<f (k) (−1)k X{ggδk=f } ωg. ωf (0),...,f (k−1),i,f (k),...,f (n) In particular it follows from (1) that Cn is a finite dimensional DG-vector space for all n, while (2) and (3) imply that C• is a simplicial DG-vector space. (cid:3) From Proposition 3.4 we obtain, for all n ≥ 0, the inclusion of DG-vector spaces: Cn in−→ Ωn. We want to extend these inclusions to contractions. This means that we want to introduce two families of maps, πm and hm such that hm, πm and im satisfy the conditions of Definition 2.1. Moreover we want this construction to be simplicial. To define these maps we use integration on affine standard simplices. An ax- iomatic definition of integration of polynomial differential forms on affine standard simplices and a more detailed discussion on its properties is given in Chapter 10 of [8]. The integration map on affine standard simplices is the map Z∆n K : Ωn → K, defined by linearity using the two identities: Z∆n K η = 0, if η ∈ Ωp n, with p 6= n, (3.1) (3.2) Z∆n K xk0 0 · · · xkn n dx0 ∧ · · · ∧ cdxi ∧ · · · ∧ dxn = (−1)i k0! · · · kn! (k0 + · · · + kn + n)! . In Construction 3.6 we describe the family of maps hm ∈ Hom−1 by Dupont in [6]. K (Ωm, Ωm) defined Notation 3.5. We use the following notation: b∆n = {(s, t0, . . . , tn) ∈ Kn+1 s + t0 + · · · + tn = 1}; while the DG-vector space of polynomial differential forms on b∆n is denoted by bΩn. SOME REMARKS ON DUPONT CONTRACTION 7 Construction 3.6. Dupont uses R as base field, but his construction works in any field of characteristic 0. The map fj : [0] → [m] is defined as fj(0) = j; to this For any η ∈ Ωm, since s + t0 = 1 and ds + dt0 = 0, there are unique forms K → ∆m K , ∗ αη = (1 − s)asbxk0 pull-back map. First suppose bfj : b∆0 × ∆m then, following Dupont's notation, we define hj ∈ HomK(Ωm, Ωm)−1. one we associate the map bfj defined as: αη, βη ∈ Ωm[s] such that bfj hj(η) = 1Z0 =Z bfj((s, t0), v) = sej + t0v = sej + (1 − s)v. (η) = ds ∧ αη + βη, where bfj (1 − s)asbds xk0 0sbdt0 xk0 0 · · · xkn n dxc1 ∧ · · · ∧ dxcl 0 · · · xkn n dxc1 ∧ · · · ∧ dxcl . 0 · · · xkn n dxc1 ∧ · · · ∧ dxcl ta b∆0 ∗ : bΩ0 ⊗ Ωm → Ωm is the Now we can extend this by linearity. For any strictly increasing morphismf : [n] → [m] we define: hf ∈ Hom−n−1 K (Ωm, Ωm), hf = hf (n) ◦ · · · ◦ hf (0), and hm ∈ Hom−1 K (Ωm, Ωm) is: hm(η) = mXn=0 Xf ∈I(n,m) ωf ∧ hf (η). Next we describe some properties of the maps hf . These results were proved by Dupont in different parts of Chapter 2 of [6]. Lemma 3.7. Take f : [n] → [m] and g : [m] → [p] then: (1) (2) g∗ ◦ hgf = hf ◦ g∗, [hf , d](η) = hf (dη) + (−1)ndhf (η) =Z∆n K f ∗η − nXi=0 We use the convention that hf δ0 is the identity. (−1)ihf δi(η). Theorem 3.8 (Dupont, [6]). Consider for each m ≥ 0 the two operators πm ∈ Hom0 K(Ωm, Ωm), πm(η) = mXn=0 Xf ∈I(n,m) Z∆n K f ∗η! ωf , 8 LUIGI LUNARDON the following diagram is a simplicial contraction h• / Ω• C• i• π• Remark 3.9. The first definition of Whitney elementary forms can be found in Section 27 of [17]; they are defined exactly as those forms ωi0,...,in which appear in the proof of Proposition 3.4. We prefer Definition 3.3 due to Point (2) and (3) of Proposition 3.4. The notation used for the space of polynomial differential forms is the same of [12]. An other common notation present in literature is the one of [8] - here Ωm is denoted with (AP L)m. The family of maps {πm} was defined by Whitney in [17], Dupont described the family of maps {hm} explicitly in the original proof of Theorem 3.8 given in [6]. Later Getzler showed in [12] that side conditions πmhm = h2 m = 0 (and hence hmim = 0) hold. 4. The proof of Dupont's Theorem In this section we describe a family of maps km such that the diagram k• / Ω• C• i• π• is a simplicial contraction. The family of maps km and the proof of this result was shown to us by Manetti at a cycle of seminars at the University "La Sapienza". Recall that I(n, m) is the subset of Mor∆([n], [m]) of injective (and hence strictly increasing) morphisms. Construction 4.1. Take f ∈ I(n, m) we define: K → ∆m K , bf : b∆n × ∆m The operator kf ∈ Hom−n−1 K (Ωm, Ωm) is: ((s, t0, . . . , tn), v) 7→ sv + tief (i). nXi=0 kf : Ωm bf ∗ · ⊗Id −−−−−→bΩn ⊗ Ωm Rb∆n where bf ∗ : Ωm →bΩn ⊗ Ωm is the usual pull-back map. mXn=0 Xf ∈I(n,m) The map km ∈ Hom−1 K (Ωm, Ωm) is km(η) = ωf ∧ kf (η). −−−−−−−−→ Ωm, / o o / o o SOME REMARKS ON DUPONT CONTRACTION 9 The next lemma is exactly Lemma 3.7, but with the maps kf instead of the maps hf . We don't give a proof of this lemma. This result follows, a posteriori, from Lemma 5.3, Lemma 5.4 and Lemma 5.5. Lemma 4.2. Take f : [n] → [m] and g : [m] → [p] then: (1) (2) g∗ ◦ kgf = kf ◦ g∗, [kf , d](η) = kf (dη) + (−1)ndkf (η) =Z∆n K f ∗η − nXi=0 (−1)ikf δi (η). We use the convention that kf δ0 is the identity. The following theorem corresponds to Theorem 3.8, by replacing the maps hm with km. Theorem 4.3 (Dupont, [6]). Consider for each m ≥ 0 the operator km of Con- struction 4.1 and πm of Theorem 3.8 πm ∈ Hom0 K(Ωm, Ωm), πm(η) = mXn=0 Xf ∈I(n,m) Z∆n K f ∗η! ωf , (1) the operator πm is a projector onto Cm; (2) the identity kmd + dkm = imπm − IdΩm holds; (3) for every p ∈ N and every g : [p] → [m] we have kpg∗ = g∗km. Proof. From Point (1) and Point (2) of Proposition 3.4 given f ∈ I(n, m) we have: Z∆n K f ∗ωf = 1, Z∆n K f ∗ωg = 0 if f 6= g, thus πm projects to Cm. For every η ∈ Ωm we have: dωf ∧ kf (η) + ωf ∧ ((−1)ndkf (η) + kf (dη)) km(dη) + dkmη = = = mXn=0 Xf ∈I(n,m) dωf ∧ kf (η) + ωf ∧Z∆n mXn=0 Xf ∈I(n,m) mXn=0 Xf ∈I(n,m) dωf ∧ kf (η) − ωf ∧ nXr=0 K f ∗η − nXr=0 (−1)rkf δr (η) (−1)rkf δr (η)!! + πm(η) . 10 LUIGI LUNARDON ωf = Since kf δ0 = Id and Pf ∈I(0,m) ωf ∧ 0Xr=0 Xf ∈I(0,m) ti = 1 we have: mPi=0 (−1)rkf δr (η)! = tikf δ0 (η) = η. mXi=0 Thus it follows: km(dη) + dkm(η) − πm(η) + η = mXn=0 Xf ∈I(n,m) dωf ∧ kf (η) + mXn=1 Xf ∈I(n,m) −ωf ∧ nXr=0 (−1)rkf δr (η) . Using the result of Point 3 of Proposition 3.4 it is possible to show that the right hand side of this equation vanishes: mXn=0 Xf ∈I(n,m) dωf ∧ kf (η) = = + m−1Xn=0 Xf ∈I(n,m) m−1Xn=0 Xf ∈I(n,m) nXr=0 nXr=0 mXn=1 Xg∈I(n,m) dωf ∧ kf (η) (−1)r X{gf =gδr } ωg ∧ kgδr (η) (−1)rωg ∧ kgδr (η). And thus we proved the identity. Finally, from Lemma 4.2 follows that g∗(ωf ) ∧ g∗kf (η) = ωh ∧ g∗kf (η) g∗km(η) = = mXn=0 Xf ∈I(n,m) mXn=0 Xh∈I(n,p) mXn=0 Xf ∈I(n,m) Xh ∈ I(n, p), mXn=0 Xh∈I(n,p) f = gh ωh ∧ g∗kgh(η) = ωh ∧ kh(g∗η) = kp(g∗η). Remark 4.4. This proof of Theorem 4.3 was shown to a small audience by Manetti. Dupont in [6] showed that Lemma 4.2 holds also for the family of maps hm. Since the proof of Theorem 4.3 is based only on Lemma 4.2 and on some prop- erties of Whitney elementary forms, the same proof works also for Theorem 3.8. (cid:3) SOME REMARKS ON DUPONT CONTRACTION 11 5. Equivalence of the families hm and km In this section we compare the family of maps hm of Construction 3.6 and the family of maps of Construction 4.1. The main result of this section is Theorem 5.1, where we prove that the two families coincide. To make the proof more readable we will split it in many lemmas. Theorem 5.1. For every m ∈ N we have that km = hm. Proof. From the definition of hm and km if follows that it is enough to prove that for each f ∈ I(n, m) the identity kf = hf holds. We proceed by induction on n. If n = 0 this is Lemma 5.3; thus suppose n > 0 and assume the statement true for every function in I(p, m) with p < n. Fix f ∈ I(n, m), in particular the statement holds for f [n−1], and then we have the chain of equality: hf = hf (n) ◦ hf [n−1] = hf (n) ◦ kf [n−1]. The only thing left to prove is kf = hf (n) ◦kf [n−1]. Let f : [n] → [m] be an injective map. By linearity we can assume that η is the q-form η = xk0 0 · · · x0 f (n) · · · xkm m dxc1 ∧ · · · ∧ dxcq with c1 < c2 < · · · < cq and ci 6= f (n), for all i. If q ≤ n we have 0 = kf (η) = hf (η), since they are forms of negative degree, hence the equality holds. Suppose now q > n. Let C : = {c1, . . . , cq} and Im(f [n−1]) if the intersection is such that C ∩ Im(f [n−1]) < n − 1, then by the same degree argument it follows that 0 = hf (η) = kf (η). If C ∩ Im(f [n−1]) = n − 1 we are under the hypothesis of Lemma 5.4, so kf (η) = hf (η). If C ∩ Im(f [n−1]) = n then we are under the hypothesis of Lemma 5.5, so kf (η) = hf (η). (cid:3) The next lemmas provide the technicalities behind Theorem 5.1, whose in- ductive step will be Lemma 5.3 is the inductive base of the proof; Lemma 5.4 and Lemma 5.5 address the computation of the functions hf (η) and kf (η) in some key cases. Notation 5.2. When necessary, we use the notation dx0,...,xn for the differential form dx0 ∧ · · · ∧ dxn. Lemma 5.3. For each integer 0 ≤ j ≤ m, consider the map: fj : [0] → [m] fj(0) = j. Then the map hj of Construction 3.6 and the maps kfj of Construction 4.1 coin- cide. Proof. Take η ∈ Ωm and assume without loss of generality j = 0 and The general case will follow by linearity. Call f the map fj. η = xk0 0 · · · xkm m dxc1 ∧ · · · ∧ dxcl . 12 LUIGI LUNARDON Since the two identities: x0 = 1 − mXi=1 xi, dx0 = − dxi, mXi=1 hold on the affine standard simplex, we can assume η = xk1 with 0 < c1 < · · · < cl. Once again the general case follows by linearity. 1 · · · xkm m dxc1 ∧· · ·∧dxcl , Following Construction 3.6 we compute αη: αη = (1 − s) mP i=1 and then h0(η) is: h0(η) = mP i=1 ki+l−1 bf ∗(η) =s ki+l−1 xk1 1 · · · xkm xk1 1 · · · xkm m lXi=1 k1 + · · · + km + l lXi=1 m lXi=1 (−1)ixci ds ∧ dxc1 ∧ · · · ∧ ddxci ∧ · · · ∧ dxcl! , (−1)ixci dxc1 ∧ · · · ∧ ddxci ∧ · · · ∧ dxcl! . (−1)i−1xci ds ∧ dxc1 ∧ · · · ∧ ddxci ∧ · · · ∧ dxcl! xk1 1 · · · xkm m mP ki+l Following Construction 4.1 we have + s xk1 1 · · · xkm Using Equation (3.1), it follows i=1 m dxc1 ∧ · · · ∧ dxcl . mP i=1 s ki+l = 0, Z b∆0 since we are integrating a 0-form the 1-simplex bΩ0. And consequently m dxc1 ∧ · · · ∧ dxcl! = 0. xk1 1 · · · xkm ki+l i=1 Thus to conclude the proof we have just to compute kf0 . mP (cid:18)Zb∆0 kf0 (η) =(cid:18)Zb∆0 ⊗Id(cid:19) s ⊗Id(cid:19)(cid:16)bf ∗(η)(cid:17) k1 + · · · + km + l lXi=1 xk1 1 · · · xkm = m (−1)ixci dxc1 ∧ · · · ∧ ddxci ∧ · · · ∧ dxcl! . (cid:3) Lemma 5.4. Fix an integer n, a function f ∈ I(n, m) and a polynomial differen- tial form η = xk0 m dxc1 ∧ · · · ∧ dxcq . It is not restrictive to assume that c1 < c2 < · · · < cq and ci 6= f (n), for all i. Assume, moreover, that we have f (n) · · · xkm 0 · · · x0 and kf [n−1] (η) = hf [n−1](η). (cid:12)(cid:12){c1, . . . , cq} ∩ Im(f [n−1])(cid:12)(cid:12) = n − 1, SOME REMARKS ON DUPONT CONTRACTION 13 Then it follows that hf (η) = kf (η). Proof. The form kf (η) has negative degree, therefore it vanishes. It is not restric- tive to assume that n − 1 6∈ C ∩ Im(f [n−1]). Thus η has the form: η = xk0 0 · · · x0 f (n) · · · xkm m dxf (0) ∧ · · · ∧ dxf (n−2) ∧ dxb1 ∧ · · · ∧ dxbl , with 0 < b1 < b2 < · · · < bl < m, bi 6= f (n − 1), f (n) and l = q − n + 1 > 1. Define the set P = {(p0, . . . , pn−1) ∈ Nn such that 0 ≤ pi ≤ kf (i), ∀i}. Next we compute \f [n−1] ∗ (η). We have ∗ \f [n−1] (η) =Xp∈P ηp lXi=1 xbi sl−1(−1)i+nds,t0,...,tn−2,xb1 ,..., cxbi ,...,xbl! + ωp! , where ωp are forms which vanish under Z b∆n−1 the polynomial ⊗Id by a degree argument, and ηp is ηp = n−1Yi=0(cid:18)kf (i) pi (cid:19)tpi kf (i) −pi i x i mP ( i=0 s ki− n−1P i=0 pi) Yi = 0, . . . , m i 6∈ f ([n − 1]) xki i . ηp lXi=1 xbi (−1)idxb1 ∧ · · · ∧ ddxbi ∧ · · · ∧ dxbl! . Thus we have that Then αkf [n−1] (η) is equal to hf [n−1](η) = kf [n−1](η) =Xp∈P  lXi=1Xj<i lXi=1Xj>i + Xp∈P = 0. xbi xbj (−1)i+jds ∧ dxb1 ∧ · · · ∧ ddxbj ∧ · · · ∧ ddxbi ∧ · · · ∧ dxbl xbi xbj (−1)i+j−1ds ∧ dxb1 ∧ · · · ∧ ddxbi ∧ · · · ∧ ddxbj ∧ · · · ∧ dxbl! ηpsl−1 So we proved that hn ◦ kf [n−1] (η) = hf (η) = 0; and this completes the first part of the proof. (cid:3) Lemma 5.5. Fix an integer n, a function f ∈ I(n, m) and a polynomial differen- tial form η = xk0 m dxc1 ∧ · · · ∧ dxcq . It is not restrictive to assume that c1 < c2 < · · · < cq and ci 6= f (n), for all i. Assume, moreover, that we have: f (n) · · · xkm 0 · · · x0 and kf [n−1](η) = hf [n−1](η). (cid:12)(cid:12){c1, . . . , cq} ∩ Im(f [n−1])(cid:12)(cid:12) = n, Then it follows that hf (η) = kf (η). 14 LUIGI LUNARDON Proof. We can write η as η = xk0 0 · · · x0 f (n) · · · xkm m dxf (0) ∧ · · · ∧ dxf (n−1) ∧ dxb1 ∧ · · · ∧ dxbl , with 0 < b1 < b2 < · · · < bl < m, bi 6= f (n) and l = q − n ≥ 1. Consider, as in Lemma 5.4, the set: P = {(p0, . . . , pn−1) ∈ Nn such that 0 ≤ pi ≤ kf (i), ∀i}. In order to compute kf (η), observe that: ǫpθpd(sxf (0) + t0) ∧ · · · ∧ d(sxf (n−1) + tn−1) ∧ d(sxb1 ) ∧ · · · ∧ d(sxbl ). bf ∗(η) =Xp∈P Where θp and ǫp are defined as: ǫp = n−1Yi=0(cid:18)kf (i) pi (cid:19)x kf (i) −pi f (i) Yi = 0, . . . , m i 6∈ f ([n − 1]) xki i ; mP i=0 ki− n−1P i=0 pi θp = s tpi i . n−1Yi=0 Then kf (η) is equal to: b∆n ⊗Id Xp∈P ǫpθp lXi=1 Z pi + l − 1(cid:19)! (cid:18) mPi=0 n−1Pi=0 =Xp∈P ki + l + n(cid:19)! (cid:18) mPi=0 The forms ωp vanish under Z ki − b∆n−1 xbi sl−1dt0,...,tn−1,xb1 ,...,xbi−1 ,s,xbi+1 ,...,xbl! + ωp!! (pi!) n−1Qi=0 ǫp lXi=1 (−1)ixbi dxb1 ∧ · · · ∧ ddxbi ∧ · · · ∧ dxbl! . ⊗Id by a degree argument. Moreover we have θpsl−1dt0 ∧ · · · ∧ dtn = (cid:18) mPi=0 Z b∆n pi + l − 1(cid:19)! ki + l + n(cid:19)! (pi!) n−1Qi=0 . ki − n−1Pi=0 (cid:18) mPi=0 We can now compute kf [n−1] (η). Recall that η = xk0 0 . . . x0 f (n) . . . xkm m dxf (0) ∧ · · · ∧ dxf (n−1) ∧ dxb1 ∧ · · · ∧ dxbl . Thus we have: SOME REMARKS ON DUPONT CONTRACTION 15 xf (i)dt0 ∧ · · · ∧ dti−1 ∧ ds ∧ dti+1 ∧ · · · ∧ dtn−1 ∧ dxb1 ∧ · · · ∧ dxbl kf [n−1](η) + + =(cid:18)ZC n−1 n−1Xi=0 n−1Xj=0 lXi=1 (cid:18) mPi=0 =Xp∈P n−1Xi=0 lXi=1 − + γ1 =Xp∈P γ2 =Xp∈P + (pi!) ki − xf (i)dxb1 ∧ · · · ∧ dxbl ǫp dxb1 ∧ · · · ∧ dxbl ǫpθpsl(cid:18)dt0 ∧ · · · ∧ dtn−1 ∧ dxb1 ∧ · · · ∧ dxbl ⊗Id(cid:19)Xp∈P xbi dt0,...,xf (j),...,tn−1,xb1 ,...,xbi−1 ,s,xbi+1 ,...,xbl (cid:19) + ωp pi + l(cid:19)! n−1Qi=0 n−1Pi=0 (cid:18) mPi=0 ki + l + n(cid:19)! (−1)i−1xbi dxf (j) ∧ dxb1 ∧ · · · ∧ ddxbi ∧ · · · ∧ dxbl . n−1Xj=0 (cid:18) mPi=0 pi + l(cid:19)! n−1Qi=0 n−1Pi=0 (cid:18) mPi=0 ki + l + n(cid:19)! (cid:18) mPi=0 pi + l(cid:19)! n−1Qi=0 n−1Pi=0 (cid:18) mPi=0 ki + l + n(cid:19)! (−1)i−1xbi dxf (j) ∧ dxb1 ∧ · · · ∧ ddxbi ∧ · · · ∧ dxbl . n−1Xj=0 lXi=1 ǫp n−1Xi=0 ǫpdxb1 ∧ · · · ∧ dxbl ; (pi!) ki − (pi)! ki − For the sake of readability call −xf (i)dxb1 ∧ · · · ∧ dxbl Next we show that hf (n)(γ1) = kf (η) and hf (n)(γ2) = 0, which concludes the proof. The map hf (n) is the one described in Construction 3.6. The pullback of γ1 under the map fn : [0] → [m], 0 7→ f (n) is: f ∗ n(γ1) =Xp∈P mP ki− ap(1 − s) i=0 n−1P i=0 pi ǫpd((1 − s)xb1 ) ∧ · · · ∧ d((1 − s)xbl ). 16 LUIGI LUNARDON Where ap is defined for p ∈ P as apǫp(1 − s) i=0 mP ki− n−1P i=0 Then we have: αγ1 =Xp∈P hf (n)(γ1) =Xp∈P By integration we get: ap n−1P ki− i=0 mP i=0 pi+l ap = (cid:18) mPi=0 . ki − (pi!) pi + l(cid:19)! n−1Qi=0 n−1Pi=0 (cid:18) mPi=0 ki + l + n(cid:19)! −(−1)i−1xbi ds,xb1 ,..., cxbi ,...,xbl! . pi+l−1 lXi=1 (−1)ixbi dxb1 ∧ · · · ∧ ddxbi ∧ · · · ∧ dxbl! . ǫp lXi=1 To conclude the proof we need to show that hf (n)(γ2) = 0, but this follows (cid:3) from Lemma 5.4 since n 6∈ {f (1), . . . , f (n − 1), b1, . . . , bl}. Remark 5.6. A remarkable consequence of Theorem 4.3 is that we have a sim- plicial contraction k• / Ω• C• i• π• Getzler in [12] showed that k2 m = 0 and that πmkm = 0 (his proof of this latter fact works replacing km with any family of functions satisfying Point (2) of Theorem 4.3). This means that this is a simplicial contraction in the sense of [7] and of [15]. References [1] R. Bandiera: Descent of Deligne-Getzler ∞-groupoids. (May 2017) arxiv.org/abs/1705.02880 1 [2] R. Bandiera and M. Manetti: Algebraic models of local period maps and Yukawa algebras. Lett. Math. Phys. https://doi.org/10.1007/s11005-018-1064-1 Springer Netherlands (2018) 1 [3] A.K. Bousfield and V.K.A.M. Gugenheim: On PL De Rham theory and rational homotopy type. Memoirs of the Amer. Math. Soc. 179 (1976). 1 [4] X. Z. Cheng and E. Getzler: Transferring homotopy commutative algebraic structures. Jour- nal of Pure and Applied Algebra, 212, Issue 11, 2535-2542, (2008). 1, 4 [5] J. L. Dupont: Simplicial de Rham cohomology and characteristic classes of flat bundles. Topology 15 233-245, (1976). 1 [6] J.L. Dupont: Curvature and characteristic classes. Lecture Notes in Mathematics 640 Springer-Verlang (1978). 122-123 1, 3, 4, 6, 7, 8, 9, 10 [7] S. Eilenberg and S. Mac Lane: On the groups H(π, n). I. Ann. of Math. 58 (1953). 4, 16 [8] Y. Félix, S. Halperin and J.C. Thomas: Rational homotopy theory. Graduate Texts in Mathematics 205 New York: Springer-Verlag, (2001). 1, 6, 8 / o o SOME REMARKS ON DUPONT CONTRACTION 17 [9] D. Fiorenza and M. Manetti: L∞-structures on mapping cones. Algebra Number Theory, 1, (2007), 301-330. 1 [10] D. Fiorenza, M. Manetti and E. Martinengo: Cosimplicial DGLAs in deformation theory. Communications in Algebra, 40, (2012), 2243 - 2260. 1 [11] K. Fukaya: Deformation theory, homological algebra and mirror symmetry. Geometry and physics of branes (Como, 2001), Ser. High Energy Phys. Cosmol. Gravit., IOP Bristol (2003) 121-209 1 [12] E. Getzler: Lie theory for nilpotent L∞-algebras. Ann. of Math. 170 (1): 271-301 (2009). 1, 4, 8, 16 [13] J. Huebshmann and T. Kadeishvili: Small models for chain algebras. Math. Z. 207 (1991) 245-280. 246, 261 1 [14] M. Kontsevich, Y. Soibelman: Homological mirror symmetry and torus fibrations. K. Fukaya, (ed.) et al., Symplectic geometry and mirror symmetry. Proceedings of the 4th KIAS annual international conference, Seoul, South Korea, August 14-18, 2000. Singapore: World Scientific. (2001) 203-263; arXiv:math.SG/0011041. 1 [15] L. Lambe, J. Stasheff: Applications of perturbation theory to iterated fibrations. Manuscripta Math. 58 (1987), no. 3, 363–376. 4, 16 [16] V. Navarro Aznar: Sur la théorie de Hodge-Deligne. Invent. Math. 90 (1987) 11-76. 115 1 [17] H. Whitney: Geometric integration theory. Princeton University Press, Princeton, N. J., (1957). 111, 115 4, 8
1201.0734	1	1201	2012-01-03T19:58:55	Wild multidegrees of the form (d,d_2,d_3) for given d greather than or equal to 3	[ "math.AG" ]	Let d be any number greather than or equal to 3. We show that the intersection of the set mdeg(Aut(C^3))\ mdeg(Tame(C3)) with {(d_1,d_2,d_3) : d=d_1 =< d_2 =< d_3} has infinitely many elements, where mdeg h = (deg h_1,...,deg h_n) denotes the multidegree of a polynomial mapping h=(h_1,...,h_n):C^n ---> C^n. In other words, we show that there is infiniltely many wild multidegrees of the form (d,d_2,d_3), with fixed d >= 3 and d =< d_2 =< d_3, where a sequences (d_1,...,d_n) is a wild multidegree if there is a polynomial automorphism F of C}^n with mdeg F=(d_1,...,d_n), and there is no tame autmorphim of C^n with the same multidegree.	math.AG	math	WILD MULTIDEGREES OF THE FORM (d, d2, d3) FOR GIVEN d GREATHER THAN OR EQUAL TO 3 MAREK KARA´S, JAKUB ZYGAD LO Abstract. Let d be any number greather than or equal to 3. We show that the intersection of the set mdeg (Aut (C3))\mdeg (Tame (C3)) with {(d1, d2, d3) ∈ (N+)3 : d = d1 ≤ d2 ≤ d3} has infinitely many elements, where mdeg h = (deg h1, . . . , deg hn) denotes the multidegree of a polynomial mapping h = (h1, . . . , hn) : Cn → Cn. In other words, we show that there is infiniltely many wild multidegrees of the form (d, d2, d3), with fixed d ≥ 3 and d ≤ d2 ≤ d3, where a sequences (d1, . . . , dn) ∈ Nn is a wild multidegree if there is a polynomial automorphism F of Cn with mdeg F = (d1, . . . , dn), and there is no tame autmorphim of Cn with the same multidegree. 1. Introduction In the following we will write Aut (Cn) for the group of the all polynomial automorphisms of Cn and Tame (Cn) for the subgroup of Aut (Cn) containing all the tame automorphisms. Let us recall that a polynomial automorphism F is called tame if F can be expressed as a composition of linear and triangular automorphisms, where G = (G1, . . . , Gn) ∈ Aut (Cn) is called linear if deg Gi = 1 for i = 1, . . . n, and H = (H1, . . . , Hn) ∈ Aut (Cn) is called triangular if for some permutation σ of {1, . . . , n} we have Hσ(i) − ci · xσ(i) belongs to C[xσ(1), . . . , xσ(i−1)] for i = 1, . . . , n and some ci ∈ C∗ = C \ {0}. Here deg h denotes the total degree of a polynomial h ∈ C[x1, . . . , xn]. Let F = (f1, . . . , f3) ∈ Aut (Cn). By multidegree of F we mean the sequence mdeg F = (deg f1, . . . , + = (N \ {0})n. deg fn). One can consider the function (also denoted mdeg ) mapping Aut (Cn)) into Nn It is well-known [1, 2] that (1) mdeg (Aut (C2)) = mdeg (Tame (C2)) = {(d1, d2) ∈ N2 + : d1d2 or d2d1}, but in the higher dimension (even for n = 3) the situation is much more complicated and the question about the sets mdeg (Aut (Cn) and mdeg (Tame (Cn)) is still not well recognized. The very first results [4] about the sets mdeg (Tame (Cn)) for n > 2, say that (3, 4, 5) /∈ mdeg (Tame (C3)) and (d1, . . . , dn) ∈ mdeg (Tame (Cn)) for all d1 ≤ d2 ≤ . . . ≤ dn with d1 ≤ n− 1. Next, in [5] it was proved that for any prime numbers p2 > p1 ≥ 3 and d3 ≥ p2, we have (p1, p2, d3) ∈ mdeg (Tame (C3)) if and only if d3 ∈ p1N + p2N. The complete characterization of the set mdeg (Tame (C3)) ∩ {(3, d2, d3) : 3 ≤ d2 ≤ d3 } was given in [6]. The result says that (3, d2, d3), with 3 ≤ d2 ≤ d3, belongs to mdeg (Tame (C3)) if and only if 3d2 or d3 ∈ 3N + d2N. The similar result about the set mdeg (Tame (C3)) ∩ {(5, d2, d3) : 5 ≤ d2 ≤ d3 } and more other results are given in [7]. In the rest of the paper we will work with n = 3 and we will write C[x, y, z] instead of C[x1, x2, x3]. Let (2) and (3) W = mdeg (Aut (C3))) \ mdeg(Tame (C3)) Wd = W ∩ {(d1, d2, d3) ∈ N3 + : d = d1 ≤ d2 ≤ d3}. Note that for the famous Nagata automorphism (4) N : C3 ∋   x y z   7→   x − 2y(y2 + zx) − z(y2 + zx)2 y + z(y2 + zx) z 1 ∈ C3,   which is known to be wild automorphism, i.e. N /∈ Tame (C3), we have mdeg N = (5, 3, 1) ∈ mdeg (Tame (C3)). Thus, mdeg N is not an element of W (in other words, mdeg N is not a wild multidegree). Besides of this the autors proved that the set W is not empty, and even more that this set is infinite [8]. Now we show the following refinement of that result: Theorem 1.1. Let d > 2 be any number. The set Wd = (cid:2)mdeg (Aut (C3))\mdeg (Tame (C3))(cid:3) ∩ {(d1, d2, d3) ∈ (N+)3 : d = d1 ≤ d2 ≤ d3} is infinite. The proof of the theorem will be given separetly for odd numbers d ≥ 3 (section 2), even numbers d > 4 (section 3) and finally for d = 4 (section 4). Note also the following remarks: Remark 1.2. The sets W1 and W2 are empty, i.e. if d ∈ {1, 2} then for every d2, d3 ∈ N+ such that d ≤ d2 ≤ d3 one can show a tame automorphism F of C3 satisfying mdeg F = (d, d2, d3). For d = 1 one can take F (x, y, z) = (x, y + xd2, z + xd3 ), while for d = 2 one can use [7, Cor. 3.3] or [4, Cor. 2.3]. Remark 1.3. Let d ≤ e and define Wd,e = {(d1, d2, d3) ∈ N3 is finite. + : d = d1, e = d2 ≤ d3}. Then the set Wd,e The proof of the above result can be found in [11] or [7, Thm. 8.1]. 2. The case of odd number d 2.1. Elements of mdeg (Aut (C3)). In this section we show the following two lemmas. Lemma 2.1. Let r, k ∈ N+. If r ≡ 1(mod 4), then (5) (r, r + 2k, r + 4k) ∈ mdeg (Aut (C3)). Proof. Since r ≡ 1(mod 4), we have r = 4l + 1 for some l ∈ N+. Let (6) F = (T ◦ Nk) ◦ (T ◦ Nl) , where T (x, y, z) = (z, y, x) and for any m ∈ N∗ (7) Nm : C3 ∋   x y z   7→   One can see that mdeg (T ◦ Nl) = (1, 1 + 2l, 1 + 4l) . Moreover, if we put (f, g, h) := T ◦Nl, then g2+f h = Y 2 + ZX. Thus x − 2y(y2 + zx)m − z(y2 + zx)2m y + z(y2 + zx)m z ∈ Cn.   F = (T ◦ Nk) ◦ (f, g, h) Since deg h > max {deg f, deg g} , one can see that = (cid:0)h, g + h(Y 2 + ZX)k, f − 2g(Y 2 + ZX)k − h(Y 2 + ZX)2k(cid:1) . (8) (cid:3) mdeg F = (4l + 1, (4l + 1) + 2k, (4l + 1) + 4k) . Lemma 2.2. For every r, k ∈ N+, we have (9) (r, r + k(r + 1), r + 2k(r + 1)) ∈ mdeg (Aut (C3)). Proof. Assume that r > 1. Let (10) and put (11) (f, g, h) = (X, Y, Z + X r) F = (T ◦ Nk) ◦ (f, g, h) . 2 Since F = (cid:0)h, g + h(g2 + f h)k, f − 2g(g2 + f h)k − z(g2 + f h)2k(cid:1) (12) and deg h = r > max {deg f, deg g} , one can see that deg(cid:0)g2 + f h(cid:1) = r + 1 and so (13) mdeg F = (r, r + k(r + 1), r + 2k(r + 1)) . If r = 1, then one can take F = T ◦ Nk. (cid:3) 2.2. Elements outside mdeg (Tame (C3)). In this section we show the following two lemmas. Lemma 2.3. Let r, k ∈ N+. If r > 1 is odd and gcd (r, k) = 1, then (14) (r, r + 2k, r + 4k) /∈ mdeg (Tame (C3)). Lemma 2.4. Let r, k ∈ N+. If r > 1 is odd and gcd (r, k) = 1, then (15) (r, r + k(r + 1), r + 2k(r + 1)) /∈ mdeg (Tame (C3)). In the proofs of the above lemmas we will use the following Theorem 2.5 ([8], Thm. 2.1). Let d3 ≥ d2 > d1 ≥ 3 be positive integers. If d1 and d2 are odd numbers such that gcd (d1, d2) = 1, then (d1, d2, d3) ∈ mdeg (Tame (C3)) if and only if d3 ∈ d1N + d2N, i.e. if and only if d3 is a linear combination of d1 and d2 with coefficients in N. Proof of Lemma 2.3. Note that the numbers r and r + 2k are odd. Moreover, (16) and since r is odd, (17) gcd (r, r + 2k) = gcd (r, 2k) , gcd (r, 2k) = gcd (r, k) = 1. Assume that r + 4k ∈ rN + (r + 2k)N. Since 2(r + 2k) > r + 4k and r ∤ (r + 4k), we have (18) r + 4k = r + 2k + mr, for some m ∈ N. By (22), 2k = mr. Since r is odd, the last equality means that rk, a contradiction. Thus r + 4k /∈ rN + (r + 2k)N, and by Theorem 2.5 we obtain a thesis. (cid:3) Proof of Lemma 2.4. Since r + 1 is even, it follows that the numbers r and r + k(r + 1) are odd. Moreover, (19) and since gcd (r, k) = 1, (20) Similarily (21) gcd (r, r + k(r + 1)) = gcd (r, k(r + 1)) , gcd (r, k(r + 1)) = gcd (r, r + 1) = gcd (r, 1) = 1. gcd (r, r + 2k(r + 1)) = gcd (r, 2k(r + 1)) = gcd (r, r + 1) = 1. In particular r ∤ r + 2k(r + 1). Assume that r+2k(r+1) ∈ rN+(r + k(r + 1)) N. Since 2(r+k(r+1)) > r+2k(r+1) and r ∤ r+2k(r+1), we have (22) r + 2k(r + 1) = r + k(r + 1) + mr, for some m ∈ N. By (22), k(r + 1) = mr. Since gcd(r, k) = 1, the last equality means that rr + 1, a contradiction. Thus r + 2k(r + 1) /∈ rN + (r + k(r + 1))N, and by Theorem 2.5 we obtain a thesis. (cid:3) 3 2.3. Proof of the theorem in the case of odd d. Take any odd number d > 1. If d ≡ 1(mod 4), then by Lemmas 2.1 and 2.3 we have (23) {(d, d + 2k, d + 4k) : gcd(d, k) = 1} ⊂ mdeg (Aut(C3))\mdeg (Tame (C3)). If d ≡ 3(mod 4), then by Lemmas 2.2 and 2.4 we have {(d, d + k(d + 1), d + 2k(d + 1)) : gcd(d, k) = 1} ⊂ mdeg (Aut (C3))\mdeg (Tame (C3)). Since the set {k ∈ N+ : gcd(d, k) = 1} is infinite, the result follows. 3. The case of even number d > 4 3.1. Preparatory calculations. Fix even number d > 4 and take k ∈ N+ such that gcd(d, k) = 1. Consider the automorphisms of C3: (24) ∂x and σ = y2 + xz. and Nk defined as in (7). Note that Nk = exp(D · σk), where D = ∂ One can easily check that D is locally nilpotent derivation on C[x, y, z] and σ ∈ ker D, so σk · D is also locally nilpotent. We will consider automorphisms Fd,k of the form: Hd(x, y, z) = (x, y, z + xd) ∂z + z ∂ ∂y − 2y ∂ (25) Fd,k = T ◦ Nk ◦ Hd where T is defined as in the proof of Lemma 2.1. An easy calculation shows (even for d = 4) that (26) mdeg Fd,k = (d, d + k(d + 1), d + 2k(d + 1)) and writing d1 = d, d2 = d + k(d + 1) and d3 = d + 2k(d + 1) gives (27) (28) and gcd(d1, d2) = gcd(d, d + k(d + 1)) = gcd(d, k) = 1 gcd(d2, d3) = gcd(d + k(d + 1), d + 2k(d + 1)) = gcd(d + k(d + 1), d) = 1 gcd(d1, d3) = gcd(d, d + 2k(d + 1)) = gcd(d, 2k) = gcd(d, 2) = 2. (29) We will prove that no tame automorphism of C3 has the same multidegree as Fd,k. Suppose to the contrary that F = (F1, F2, F3) ∈ Tame(C3) and mdeg F = (d1, d2, d3). As F is not linear, due to the result of Shestakov and Umirbaev [9, 10], F must admit an elementary reduction or a reduction of types I-IV (see e.g. [9, Def. 1-3]). 3.2. Elementary reductions. Recall that an elementary reduction on i-th coordinate Fi of F occurs when there exists G(x, y) ∈ C[x, y] such that deg(Fi − G(Fj , Fk)) < deg Fi, where {i, j, k} = {1, 2, 3}. We will use extensively the following Proposition 3.1 (see e.g. [7, Prop. 2.7] or [9, Thm.2]). Suppose that f, g ∈ C[X1, . . . , Xn] are alge- braically independent and such that ¯f /∈ C[¯g] and ¯g /∈ C[ ¯f ] (¯h denotes the highest homogeneous part of h). Assume that deg f < deg g, put (30) p = deg f gcd(deg f, deg g) and suppose that G(x, y) ∈ C[x, y] with degy G(x, y) = pq + r, 0 ≤ r < p. Then (31) deg G(f, g) ≥ q(p deg g − deg g − deg f + deg[f, g]) + r deg g Suppose that F admits an elementary reduction on first coordinate, i.e. deg(F1−G(F2, F3)) < deg F1 = gcd(d2,d3) = d2. d1 for some G ∈ C[x, y]. Consequently deg G(F2, F3) = d1. By (28), we know that p := Thus, from the above proposition applied to f = F2 and g = F3 we get d2 (32) deg G(F2, F3) ≥ q(pd3 − d3 − d2 + deg[F2, F3]) + rd3 ≥ q(d2 − 1)(d3 − 1) + rd3 Since d1 < (d2 − 1)(d3 − 1) and d1 < d3 we obtain that q = 0 and r = 0. That is degy G(x, y) = 0 and G(x, y) = u(x). But then d1 = deg G(F2, F3) = deg u(F2) = d2 · deg u, which is a contradiction. 4 Similarly, suppose that F admits an elementary reduction on third coordinate, i.e. deg(F3−G(F1, F2)) < gcd(d1,d2) = d ≥ 3 by (27), it follows deg F3 = d3 for some G ∈ C[x, y]. So deg G(F1, F2) = d3. Since p := that applying Proposition 3.1 to f = F1 and g = F2 we get d1 (33) deg G(F1, F2) ≥ q(pd2 − d2 − d1 + deg[F1, F2]) + rd2 ≥ q(2k(d + 1) + d + 2) + rd2. Now, since d3 < 2k(d + 1) + d + 2 and d3 < 2d2, we obtain that q = 0 and r ∈ {0, 1}. If r = 0, we get degy G(x, y) = 0 and so G(x, y) = u(x). But then d3 = deg G(F1, F2) = deg u(F1) = d1 · deg u, which is a contradiction because gcd(d3, d1) ≤ 2 < d1. If r = 1, we get G(x, y) = u(x) + yv(x) and so d3 = deg G(F1, F2) = deg(u(F1) + F2v(F1)). Since deg F1 and deg F2 are coprime, deg(u(F1) + F2v(F1)) must be equal either to d1 · deg u or to d2 + d1 · deg v. Consequently, d3 = d1 · deg u or d3 = d2 + d1 · deg v. First case leads to a contradiction since gcd(d3, d1) = 2 < d1 and second since gcd(d3 − d2, d1) = gcd(k(d + 1), d) = 1 < d1. Now suppose that F admits an elementary reduction on second coordinate. Then deg(F2−G(F1, F3)) < gcd(d1,d3) and apply deg F2 = d2 for some G ∈ C[x, y] and so deg G(F1, F3) = d2. Let us put p = Proposition 3.1 to f = F1 and g = F3. We will show that degy G(x, y) = 0. By (29), p = d 2 ≥ 2 and so d1 deg G(F1, F3) ≥ q(pd3 − d3 − d1 + deg[F1, F3]) + rd3 ≥ q( d − 2 2 (2k(d + 1) + d) − d + 2) + rd3 ≥ q((d − 2)k(d + 1) + 2) + rd3 ≥ q(k(d + 1) + d + 2) + rd3 Since d2 < k(d+1)+d+2 and d2 < d3, we obtain that q = 0 and r = 0 so degy G(x, y) = 0. Consequently, we get G(x, y) = u(x). But then d2 = deg G(F1, F3) = deg u(F1) = d1 · deg u, which is a contradiction since gcd(d2, d1) = 1 < d1. To summarize: if F is a tame automorphism with multidegree equal to mdeg Fd,k, then F does not admit an elementary reduction. 3.3. Shestakov-Umirbaev reductions. By the previous subsection and the following theorem we only need to check that no autmorphism of C3 with multidegree (d1, d2, d3) = (d, d + k(d + 1), d + 2k(d + 1)) admits a reduction of type III. Theorem 3.2 ([7, Thm. 3.15]). Let (d1, d2, d3) 6= (1, 1, 1) , d1 ≤ d2 ≤ d3, be a sequence of positive integers. To prove that there is no tame automorphism F of C3 with mdeg F = (d1, d2, d3) it is enough to show that a (hypothetical) automorphism F of C3 with mdeg F = (d1, d2, d3) admits neither a reduction of type III nor an elementary reduction. Moreover, if we additionally assume that d3 2 or 3 ∤ d1, then d2 it is enough to show that no (hypothetical) automorphism of C3 with multidegree (d1, d2, d3) admits an elementary reduction. In both cases we can restrict our attention to automorphisms F : C3 → C3 such that F (0, 0, 0) = (0, 0, 0) . = 3 But, since d1 is even, it follows that 2 ∤ d2 by (27). Hence, no automorphism of C3 with multidegree (d1, d2, d3) admits a reduction of type III by the following remark. Remark 3.3 ([7, Rmk. 3.9]). If an automorphism F of C3 with mdeg F = (d1, d2, d3) , 1 ≤ d1 ≤ d2 ≤ d3, admits a reduction of type III, then (1) 2d2, (2) 3d1 or d3 d2 = 3 2 . 4. The case of d = 4 Let us consider the mapping F4,k defined as in (25) for k ∈ N+ with gcd(4, k) = 1 (in other words, for odd k). By (27) and (29), we know that d2 is odd and d3 is even. Then, since d3 − d2 = 5k > 1 and mdeg F4,k = (4, 4 + 5k, 4 + 10k) =: (d1, d2, d3) by (26), it follows that the result of Theorem 1.1, for d = 4, is a consequence of the following Theorem 4.1 ([7, Thm. 6.10]). If d2 ≥ 5 is odd and d3 ≥ d2 is even such that d3 − d2 6= 1, then (4, d2, d3) ∈ mdeg (cid:0)Tame (cid:0)C3(cid:1)(cid:1) if and only if d3 ∈ 4N + d2N. 5 In fact, if we assume that d3 ∈ 4N + d2N, then we get d3 = d2 + 4m for some m ∈ N, since 2d2 > d3 and 4 ∤ d3. Hence, 5k = d3 − d2 = 4m. Since k is odd, this is a contradiction. References [1] H.W.E. Jung, Uber ganze birationale Transformationen der Ebene, J. reine angew. Math. 184 (1942), 161-174. [2] W. van der Kulk, On polynomial rings in two variables, Nieuw Archief voor Wiskunde (3) 1 (1953), 33-41. [3] A. van den Essen, Polynomial Automorphisms and the Jacobian Conjecture, Birkhauser Verlag, Basel-Boston-Berlin (2000). [4] M. Kara´s, There is no tame automorphism of C3with multidegree (3, 4, 5), Proc. Am. Math. Soc., 139, no. 3 (2011) 769-775. [5] M. Kara´s, Tame automorphisms of C3 with multidegree of the form (p1, p2, d3), Bull. Pol. Acad. Sci., Math. 59, No. 1, 27-32 (2011). [6] M. Kara´s, Tame automorphisms of C3 with multidegree of the form (3, d2, d3), J. Pure Appl. Algebra (2010), no. 12 (2010) 2144-2147. [7] M. Kara´s, Multidegrees of tame automorphisms of Cn, Diss. Math. 477, 55 p. (2011). [8] M. Kara´s, J. Zygad lo, On multidegree of tame and wild automorphisms of C3, J. Pure Appl. Algebra, J. Pure Appl. Algebra, 215 (2011) 2843 -- 2846. [9] I.P. Shestakov, U.U. Umirbaev, The Nagata automorphism is wild, Proc. Natl. Acad. Sci. USA 100 (2003),12561-12563. [10] I.P. Shestakov, U.U. Umirbaev, The tame and the wild automorphisms of polynomial rings in three variables, J. Amer. Math. Soc. 17 (2004), 197-227. [11] J. Zygad lo, On multidegrees of polynomial automorphisms of C3, arXiv:0903.5512v1 [math.AC] 31 Mar 2009. Marek Kara´s Instytut Matematyki, Wydzia l Matematyki i Informatyki Uniwersytetu Jagiello´nskiego ul. Lojasiewicza 6 30-348 Krak´ow Poland e-mail: [email protected] and Jakub Zygad lo Instytut Informatyki, Wydzia l Matematyki i Informatyki Uniwersytetu Jagiello´nskiego ul. Lojasiewicza 6 30-348 Krak´ow Poland e-mail: [email protected] 6
1607.01410	2	1607	2016-09-23T13:24:37	Special divisors on curves on K3 surfaces carrying an Enriques involution	[ "math.AG" ]	We study the pencils of minimal degree on the smooth curves lying on a K3 surface X which carries a fixed-point free involution. Generically, the gonality of these curves is totally governed by the genus 1 fibrations of X	math.AG	math	Special divisors on curves on K3 surfaces carrying an Enriques involution MARCO RAMPONI Abstract. We study the pencils of minimal degree on the smooth curves lying on a K3 surface X which carries a fixed-point free involution. Generically, the gonality of these curves is totally governed by the genus 1 fibrations of X. Mathematics Subject Classification. 14H51, 14C20, 14J28 Keywords. K3 surfaces, Brill-Noether theory 1. Introduction The gonality and the Clifford index of a smooth algebraic curve C of genus g ≥ 2 are two natural numerical invariants which, roughly speaking, measure the "speciality" of the point [C] ∈ Mg. (For more details, see Section 2 below). In general, it is very difficult to determine them for a given curve, except in very special cases. Much work has been done regarding the behaviour of these invariants for curves lying on some special classes of surfaces. For example, for a curve C lying on a smooth projective surface, an interesting question is to ask whether its gonality, or the Clifford index, varies when C moves in its linear system. A classical result in this direction, essential in this work, is the theorem of Green and Lazarsfeld [7], which states that, when C lies on a K3 surface, (i) The Clifford index is constant among smooth curves in the linear system C. (ii) If Cliff(C) < ⌊ g−1 2 ⌋, then there exists a divisor D on the K3 surface such that A = OC(D) appears in the definition of Cliff(C) and attains the minimum. The possibility of an explicit computation of the Clifford index and the gonality of curves lying on particular K3 surfaces is an interesting and challenging question. This highly depends on the geometry of the given ambient surface. We are thus interested to study this problem for some special classes of K3 surfaces carrying rich structure, such as automorphisms. At least, we want to consider surfaces with Picard number r ≥ 2, for if the Clifford index of a curve on a K3 surface X with cyclic Picard group is cut out by a divisor D, then a simple computation shows that D, up to linear equivalence, coincides with the ample generator of Pic(X ). This situation is not very interesting. In this note, we consider a pair (X ,J ) where X is a K3 surface and J is a fixed- point-free involution. It is well-known [2, VIII.19] that there is a 10-dimensional moduli space of such pairs and the generic pair is such that there are no smooth rational curves lying on X. Our main result is the following. 1 MARCO RAMPONI 2 Theorem 1.1. Let (X ,J ) be a generic K3 surface X with an Enriques involution J . The gonality of any smooth curve C on X, with C2 > 0, is cut out by any elliptic curve E on X having minimal intersection with C, i.e. gon(C) = E ·C. Remark 1.2. If X is as in Theorem 1.1, by the results of [9] and the fact that there are no classes of square ±2 on X (see (3.3) below), any smooth curve on X is such that Cliff(C) = gon(C) − 2. Therefore, Theorem 1.1 automatically contains a statement about the Clifford index of C, and the elliptic curve E has the role of D in part (ii) of the result of Green and Lazarsfeld stated above. Remark 1.3. We point out that a result of a similar flavor is given in the pioneer work of Reid [13]. Consider a complete base-point free pencil of degree d on a curve C lying on a K3 surface. The inequality 1 4 d2 + d + 2 < g(C) is then a sufficient condition for the pencil to be induced by an elliptic pencil on the surface. Of course, even when d = gon(C), this condition fails for many curves on the K3 surfaces considered in Theorem 1.1 -- e.g., whenever C is the pull-back of a curve on the quotient Enriques surface having maximal gonality (these curves always exist, as we now recall). As well as [7], the second fundamental result for the proof of Theorem 1.1 is the work of Knutsen and Lopez [10] where the authors carry out a detailed analysis of the gonality of curves on an Enriques surface. We have [10, Theorem 1.3], for a curve C with C2 > 0 on an Enriques surface Y , (1.1) where gengon C = min{2f (C),m (C), ⌊ C2 4 ⌋ + 2}, f (C) = min{F ·C : F ∈ Pic(Y ),F 2 = 0,F 6≡ 0}, m (C) = min{B ·C − 2 : F ∈ Pic(Y ),B > 0,B2 = 4,B 6≡ C} and gengon C denotes the gonality of the general curve in C. Indeed, it may well happen that there exist linear subsystems of C whose smooth curves have lower gonality than gengon C (cf. [10, Corollary 1.6]). In light of this and of the trichotomy expressed by (1.1), the content of Theo- rem 1.1 might be somewhat surprising at a first glance, because the K3 surface X contains in particular the curves which are pulled back from the Enriques surface Y = X /hJ i, for which (1.1) holds. The point is that, while the gonality of a curve on an Enriques surface might well be lower than the minimal degree induced by genus 1 pencils of the surface, we still have the following condition: (1.2) (cf.[10, Corollary 1.5]). In section 3 below, we prove Theorem 1.1 by showing how the condition (1.2) is essentially enough to deduce that the gonality of all curves on the relative K3 cover X is governed by the elliptic pencils pulled back on X from its Enriques quotient. 2f (C) ≤ gengon C + 2 SPECIAL DIVISORS ON CURVES ON K3 SURFACES CARRYING AN ENRIQUES INVOLUTION 3 Throughout the paper we work over the field of complex numbers. Acknowledgements. Thanks to Alessandra Sarti whose remarks and suggestions have considerably improved the exposition of this paper. I wish to thank Andreas Leopold Knutsen for useful conversations under the shade of an oak in the uni- versity park of Bergen. Thanks to an anonymous referee for a suggestion which simplified and shortened an argument in the proof. 2. Preliminaries on gonality and Clifford index For all basic results and implications coming from Brill-Noether theory, in this section we refer the reader to [1]. Let C be a smooth algebraic curve of genus g ≥ 2. The gonality of C is defined as the integer gon(C) = min{deg(A) : A ∈ Pic(C), h0(A) = 2}. In particular, gon(C) = 2 if and only if C is hyperelliptic. By Brill-Noether theory, (2.1) gon(C) ≤ ⌊ g + 3 2 ⌋, with an equality for the general curve in Mg. We refer to ⌊ g+3 gonality of a curve of genus g. 2 ⌋ as the maximal For a line bundle A on C, one defines Cliff(A) = deg A−2h0(A)+2. The Clifford index of the curve C itself is then defined as Cliff(C) = min{Cliff(A) : A ∈ Pic(C), h0(A) ≥ 2, h1(A) ≥ 2}. Note that, by Brill-Noether theory, the line bundles appearing in the definition of Cliff(C) always exist for g ≥ 4. Thus, when g = 2,3 we adopt the standard convention that Cliff(C) = 0 when C is hyperelliptic and Cliff(C) = 1 otherwise. By Clifford's theorem, we have Cliff(C) ≥ 0 with equality if and only if C is hyperelliptic. Since Cliff(C) ≤ gon(C) − 2, we have Cliff(C) ≤ ⌊ g − 1 2 ⌋, and the equality holds for the general curve in Mg. Gonality and Clifford index are indeed very much related: for any curve C of Clifford index c and gonality k, one has [3] c + 2 ≤ k ≤ c + 3, and curves for which k = c + 3 are conjectured to be very rare [6]. As was recalled in the Introduction, if C lies on a K3 surface X and has non- maximal Clifford index, then by [7], there exists a line bundle M on the surface such that Cliff(C) = Cliff(M ⊗ OC). By [8, Lemma 8.3], building on [12], one can choose M such that h0(M ⊗ OC) = h0(M) and h1(M) = 0, whence (2.2) Cliff(C) = C · M − M2 − 2. 4 Moreover, (2.3) and (2.4) MARCO RAMPONI M is represented by an irreducible curve, 2M2 ≤ M ·C, with equality only if C ∼ 2M. 3. Proof of Theorem 1.1 Let (X ,J ) be a pair consisting of a K3 surface X together with a fixed-point free involution (i.e. an order 2 automorphism) J of X. We denote by Y = X /hJ i the quotient surface, which we call an Enriques surface. We let p : X → Y denote the natural projection and by p ∗ : H2(Y, Z) → H2(X , Z); p ∗ : H2(X , Z) → H2(Y, Z) the natural induced maps, satisfying p ∗p ∗(x) = x + J ∗(x); (p ∗y1,p ∗y2) = 2(y1,y2). p ∗p ∗(y) = 2y; If we let H2(X , Z) J be the set of classes in H2(X , Z) which are fixed by J J above properties easily imply that the restriction of p ∗ to H2(X , Z) phism onto its image which multiplies the intersection form by 2. That is, , the is an isomor- (3.1) J p ∗(H2(X , Z) J ) ≃ H2(X , Z) (2) We recall that J is a non-symplectic involution, in the sense that it acts by mul- tiplication by −1 on H2,0(X ). This, together with the fact that the action of J on H2(X , Z) preserves the intersection form, yields J H2(X , Z) ⊂ H2(X , Z) ∩ H1,1(X ). By the Lefschetz theorem on (1,1)-classes, we identify the member on the right hand side of the above equation with Pic(X ), the Picard group of X. As shown in [5], one can construct a 10-dimensional period domain D for the pairs (X ,J ), and for the generic such pair in D one has the equality (3.2) J H2(X , Z) = Pic(X ). From now on, we assume (X ,J ) to satisfy condition (3.2) above. This is our genericity assumption in the statement of Theorem 1.1. Let L be a big and nef line bundle on X. We deduce two immediate consequences of (3.2); one purely numerical and a second one more geometric in nature. Firstly, (3.2), together with (3.1), yield (3.3) L2 ≡ 0 mod 4. In particular, X contains no classes of self-intersection ±2. By [9] this implies that the gonality of smooth curves S in L is constant and Cliff(S ) = gon(S ) − 2. SPECIAL DIVISORS ON CURVES ON K3 SURFACES CARRYING AN ENRIQUES INVOLUTION 5 Thus, whenever Cliff(S ) is computed by the restriction of a divisor D on the sur- face, since degS (D) = D · S ∈ 2Z (again by (3.1)), we see that both the Clifford index and the gonality of S must be even. Secondly, (3.2) implies that J acts as an involution on L ≃ Pg. This lifts to an involution J ∗ : H0(X ,L) → H0(X ,L), at the level of sections. Let us denote by V± ⊂ H0(X ,L) the eigenspaces relative to the eigenvalues ±1 for this action. The sections in V+ and V− yield the effective divisors in L which are mapped to themselves by J . With respect to the covering p : X → Y , these divisors map 2 to 1 onto divisors on the Enriques quotient. In other words, we may choose an effective divisor C ⊂ Y such that p ∗C belongs to L and the linear subspace P+ = P(V+), as a subsystem of L, corresponds to p ∗C. (With respect to this choice, P− = P(V−) corresponds then to p ∗C + KY ). As L2 > 0 by assumption, we have C2 > 0, hence the general member of C is a smooth irreducible curve. In fact, if C is hyperelliptic then its general member is a smooth (hyperelliptic) curve by [4, Corollary 4.5.1 p. 248]. Else, C is basepoint free [4, Proposition 4.5.1] and we apply Bertini's theorem. We therefore assume C itself to be a smooth irreducible curve. Moreover, we choose C to be general in its linear system, so that, following [10], the gonality of C is equal to the general gonality (i.e. the greatest gonality among the smooth curves in C) gon(C) = gengon C. Let eC := p ∗C. It is well-known that the restriction of the canonical bundle KY to C is non-trivial. It follows that eC is a smooth irreducible curve of genus g in L and the restriction of the covering map p eC : eC → C exhibits eC as an unramified double covering of C. In particular, by push-forward of a pencil of minimal degree on eC, or by pull-back of gonality pencils from C, (3.4) gon(C) ≤ gon( eC) ≤ 2gon(C). Let now 2F be a genus 1 pencil on the Enriques surface Y such that f (C) = F ·C. We set eF = p ∗F and by (1.2) we obtain the following inequality (3.5) gon( eC) ≤ eF · eC = 2f (C) ≤ gon(C) + 2 We claim that the first inequality is, in fact, always an equality. Indeed, assume by contradiction gon( eC) < eF · eC. By (3.4), gon(C) < 2f (C). Applying [10, Corollary 1.5], we have (3.6) C2 ≥ 10 or (C2,f (C)) = (6,2) or (C2,f (C)) = (4,2). Claim. gon( eC) < ⌊ g( eC)+3 2 ⌋ (in particular, gon( eC) is even). 6 MARCO RAMPONI Proof. If C2 = 4, then eC2 = 8 and g( eC) = 5, so if equality holds, then it must be gon( eC) = 4 = 2f (C) = eF · eC, a contradiction. If C2 ≥ 6, then eC2 ≥ 12, so that g( eC) ≥ 7. Hence gon( eC) ≤ eF · eC − 1 ≤ gon(C) + 1 g(C) + 3 ≤ ⌊ ⌋ + 1 ⌋ + 1 2 g( eC)+1 2 + 3 2 ⌋ g( eC) + 11 4 g( eC) + 3 2 ⌋ = ⌊ = ⌊ < ⌊ where the last inequality uses g( eC) ≥ 7. (cid:3) Since, by our assumptions, gon(C) < 2f (C), we proceed as follows. If gon(C) = 2f (C) − 1, then gon(C) ≤ gon( eC) < 2f (C) is incompatible with the parity of gon( eC), whence yielding a contradiction. By (1.2), we may therefore assume gon(C) = 2f (C) − 2. Then, necessarily gon(C) = gon( eC). We pick a line bundle M on the K3 surface X, as in (2.2), i.e. such that Cliff( eC) = Cliff(M ⊗ O eC) = M · eC − M2 − 2. If M2 = 0, then it follows by (2.3) that M is represented by an elliptic curve E. By construction, the elliptic curve eF has minimal intersection with eC among all elliptic curves on X, whence gon( eC) = Cliff( eC) − 2 = E · eC = eF · eC, a contradiction. We may therefore assume M2 > 0. Then M2 ≥ 4 by (3.3). By (3.6) and (2.2) 4 ≤ M2 ≤ M · eC − M2 = Cliff( eC) + 2 = gon( eC). Assume gon( eC) = 4. Then M2 = 4 and M · eC = 8, so that (2.4) gives eC ∼ 2M, whence eC2 = 16. It follows that C2 = 8. This contradicts (3.6) and we may there- fore assume gon( eC) > 4. Arguing as above, J acts as an involution on M, and we get subsystems P± of M, corresponding to p ∗D and p ∗KY + D, where D is some effective divisor on Y , with p ∗D ∼ M. Since M2 > 0, we have D2 > 0, whence h0(D) ≥ 2. We have p ∗(C − D) ∼ eC − M, whence (C − D)2 > 0. Also, 2(C − D) ·C = p ∗(C − D) · p ∗C = N · eC = M · N + N2 > 0, SPECIAL DIVISORS ON CURVES ON K3 SURFACES CARRYING AN ENRIQUES INVOLUTION 7 so that by Riemann-Roch, h0(C − D) ≥ 2 and, similarly, h0(C − D + KY ) ≥ 2. Therefore, C · (C − D) ≥ 2 by the Hodge index theorem, so that C2 = (D +C − D)2 = D2 + (C − D)2 + 2C · (C − D) ≥ 2 + 2 + 2 = 6 We may now apply [11, Lemma 2.3] and obtain Cliff(C) ≤ D ·C − D2. By [11, Theorem 1.1], unless C is a smooth plane quintic (which has gonality 4), one has Cliff(C) = gon(C) − 2, and so the above inequality yields (3.7) 2(D ·C − D2) ≥ 2gon(C) − 4. On the other hand, gon( eC) = Cliff( eC) + 2 = eC · M − M2, thus (3.8) 2(D ·C − D2) = eC · M − M2 = gon( eC). Since gon(C) = gon( eC) > 4, the equations (3.7) and (3.8) are incompatible. Hence, our assumption that gon( eC) < eF · eC has led to a contradiction and we conclude gon( eC) = eF · eC. As we have already observed above, thanks to our genericity assumption on X, in eC = L. This concludes the proof of this holds true for all smooth curves S Theorem 1.1, q.e.d. References [1] E. Arbarello, M. Cornalba, P. Griffiths, and J.D. Harris. Geometry of Algebraic Curves. Number v. 1 in Grundlehren der mathematischen Wissenschaften. Springer New York, 2010. [2] W. Barth, K. Hulek, C. Peters, and A. van de Ven. Compact Complex Surfaces. Ergebnisse der Mathematik und ihrer Grenzgebiete. 3. Folge / A Series of Modern Surveys in Mathematics. Springer Berlin Heidelberg, 2014. [3] M. Coppens and G. Martens. Secant spaces and Clifford's theorem. Compositio Mathematica, 78(2):193 -- 212, 1991. [4] F. R. Cossec and I.V. Dolgachev. Enriques surfaces I. Birkhauser, 1989. [5] I.V. Dolgachev and S. Kondo. Moduli of K3 Surfaces and Complex Ball Quotients. In Rolf- Peter Holzapfel, A.Muhammed Uluda, and Masaaki Yoshida, editors, Arithmetic and Geometry Around Hypergeometric Functions, volume 260 of Progress in Mathematics, pages 43 -- 100. Birkhuser Basel, 2007. [6] D. Eisenbud, H. Lange, G. Martens, and F. Schreyer. The Clifford dimension of a projective curve. Compositio Mathematica, 72(2):173 -- 204, 1989. [7] M. Green and R. Lazarsfeld. Special divisors on curves on a K3 surface. Inventiones Mathe- maticae, 89(2):357 -- 370, 1987. [8] A.L. Knutsen. On kth-order embeddings of K3 surfaces and Enriques surfaces. Manuscripta Mathematica, 104(2):211 -- 237, 2001. [9] A.L. Knutsen. On two conjectures for curves on K3 surfaces. Internat. J. Math., 20(12):1547 -- 1560, 2009. 8 MARCO RAMPONI [10] A.L. Knutsen and A.F. Lopez. Brill -- Noether theory of curves on Enriques surfaces I: the posi- tive cone and gonality. Mathematische Zeitschrift, 261(3):659 -- 690, 2008. [11] A.L. Knutsen and A.F. Lopez. Brill -- Noether theory of curves on Enriques surfaces II: the Clif- ford index. Manuscripta Mathematica, 147(1):193 -- 237, 2015. [12] G. Martens. Algebraic Curves and Projective Geometry: Proceedings of the Conference held in Trento, Italy, March 21 -- 25, 1988, chapter On curves on K3 surfaces, pages 174 -- 182. Springer Berlin Heidelberg, Berlin, Heidelberg, 1989. [13] M. Reid. Special linear systems on curves lying on a K3 surface. J. London Math. Soc. (2), 13(3):454 -- 458, 1976. LABORATOIRE DE MATH ´EMATIQUES ET APPLICATIONS, UNIVERSIT ´E DE POITIERS, F-86962 POITIERS, FRANCE E-mail address: [email protected]
1012.5940	2	1012	2011-01-14T15:07:12	Torsion and cotorsion in the sheaf of K\"ahler differentials on some mild singularities	[ "math.AG", "math.AC" ]	We give a criterion for the sheaf of K\"ahler differentials on a cone over a smooth projective variety to be torsion-free. Applying this to Veronese embeddings of projective space and using known results on differentials on quotient singularities we show that even for mild, e.g. Gorenstein terminal, singularities the sheaf of K\"ahler differentials will in general have torsion and cotorsion.	math.AG	math	TORSION AND COTORSION IN THE SHEAF OF KÄHLER DIFFERENTIALS ON SOME MILD SINGULARITIES DANIEL GREB AND SÖNKE ROLLENSKE Abstract. We give a criterion for the sheaf of Kähler differentials on a cone over a smooth projective variety to be torsionfree. Applying this to Veronese embeddings of projective space and using known results about differentials on quotient singularities we show that even for mild, e. g. Gorenstein terminal, singularities the sheaf of Kähler differentials will in general have torsion and cotorsion. 1. Introduction Let Z be an algebraic variety over a field k, which we assume to be of char- acteristic 0. One of the few objects that come naturally with Z is its sheaf of Kähler differentials ΩZ = ΩZ/k. It is the sheafified version of the module of Kähler differentials, which is also an important tool in commutative algebra. In this note we stick to the geometric language. The sheaf of differentials and its higher exterior powers play an important rôle in many contexts, most prominently deformation theory, vanishing the- orems and (if Z is sufficiently nice) duality theory. For a more local example one could mention Berger's conjecture that a curve is smooth if and only if its sheaf of differentials is torsionfree (see e.g. [Ber63, Gre82, Poh91]) or the Zariski–Lipman conjecture: Z is locally free then Z is regular (see e.g. [Pla88, BLLS02]). if Ω∨ In the context of the minimal model program Greb, Kebekus, Kovács and Peternell proved strong extension theorems for differential forms [GKKP10], but instead of the sheaf of Kähler differentials itself they used its reflexive ∨∨ also called module of Zariski differentials (see e.g. [Kni73]). For hull ΩZ applications the following obvious question comes to mind: Question 1 - If Z has mild singularities, is ΩZ reflexive or at least torsion- free? Phrasing this slightly differently we ask if the natural map φ : ΩZ → ΩZ ∨∨, whose kernel is the torsion submodule Tors(ΩZ ), is bijective or at least in- jective. In the terminology of [Rei87, (1.6)] we say that ΩZ has cotorsion if φ is not surjective.1 Date: May 30, 2018. 2000 Mathematics Subject Classification. 14F10, 13N10, 14B05. 1In the case of curve singularities torsion and cotorsion in Ω were extensively studied for example by Greuel and his collaborators in [BG80] and [GMP85]. 1 2 DANIEL GREB AND SÖNKE ROLLENSKE If one translates mild singularity as being a local complete intersection, then indeed a complete answer to Question 1 is known. Theorem 2 ([Kun86], Proposition 9.7, Corollary 9.8) - Let Z be a local complete intersection. Then ΩZ satisfies Serre's condition Sd if and only if Z is regular in codimension d. In particular, if Z is a normal local complete intersection, then ΩZ is torsionfree, and it is reflexive if and only if Z is is nonsingular in codimension 2. However, in the context of modern birational geometry one usually mea- sures the singularities of a normal variety in terms of discrepancies, which give rise to the definition of terminal, canonical, log terminal and other sin- gularities [KM98, Section 2.3]. Even terminal singularities, the mildest class considered, are in general not complete intersections, so Theorem 2 does not apply. Somewhat contrary to our expectations we will show below that the an- swer to Question 1 is essentially negative if one interprets mild in the sense of birational geometry; as soon as one leaves the world of local complete intersections one should expect the sheaf of Kähler differentials to have both torsion and cotorsion. There are two cheap ways to produce non–lci singularities: quotients of finite groups and affine cones over projective varieties. In the second case the algebraic description is somewhat simpler and we give a criterion for the existence of torsion differential in Section 2. Recall that if X ⊂ Pn is a projective variety or scheme given by an ideal sheaf I then X is called projectively normal if H 0(Pn, OPn(d)) ։ H 0(X, OX (d)) for all d ≥ 0 or, equivalently, H 1(Pn, I(d)) = 0 for all d ≥ 0. Theorem 3 - Let X ⊂ Pn k be a smooth projective variety over a field k of characteristic zero and let I be the sheaf of ideals defining X. Let CX ⊂ An+1 be the affine cone over X. If H 1(Pn, I 2(d)) = 0 for all d ≥ 0 then ΩCX is torsionfree. If in addition X is projectively normal, then ΩCX is torsionfree if and k is projectively only if also the first infinitesimal neighbourhood of X in Pn normal.2 In the cone situation the study of cotorsion is more problematic but in Section 3 we recall results by Knighten and Steenbrink that give an easy sufficient criterion for cotorsion on finite quotient singularities. In the last section we study our main class of examples, namely Xr,d, the affine cone over the dth Veronese embedding of Pr. We can also describe Xr,d as a cyclic quotient singularity. Such cones have torsion differentials if 2After the publication of the first version of this preprint J. Wahl brought to our atten- tion that he had already observed this in the projectively normal case [Wah97, Proposition 1.4]. TORSION AND COTORSION IN THE SHEAF OF KÄHLER DIFFERENTIALS 3 and only if d ≥ 3 (Proposition 8) and cotorsion as soon as d ≥ 2 (Proposition 10); we have collected some significant cases in Table 13. Table 1. Some Veronese cones with torsion or cotorsion in ΩXd,r . singularity dim X1,2 (A1) X1,3 X2,2 X2,3 X3,2 X3,3 X5,3 type Gorenstein torsion cotorsion canonical log terminal terminal canonical terminal terminal terminal yes no no yes yes no yes no yes no yes no yes yes yes yes yes yes yes yes yes 2 2 3 3 4 4 6 let Z be a surface singularity. Remark 4 - In the surface case our results are optimal in the following sense: If Z is terminal then it is smooth and thus ΩZ is locally free. If Z is canonical but not terminal then it is one of the well–known ADE singularities, thus a hypersurface singularity; by Theorem 2 the sheaf of Kähler differentials ΩZ is torsionfree but not reflexive in this case. The easiest log terminal point is the cone over the twisted cubic X1,3 and in this case ΩZ has both torsion and cotorsion. Gorenstein terminal 3–fold singularities are hypersurface singularities by a result of Reid [Rei87, (3.2) Theorem]; hence Theorem 2 applies to show that the sheaf of Kähler differentials is reflexive in this case. It is possible that other classes of mild singularities in small dimensions turn out to have torsionfree or reflexive sheaf of Kähler differentials as well. Using structural results like Hilbert–Burch or Buchsbaum–Eisenbud, low– codimensional singularities might also be accessible (compare [MvS01, Sec- tion 4]). 2. Torsion differentials on cones In this section we give the proof of Theorem 3. Let X ⊂ Pn be a smooth, irreducible, non–degenerate projective variety and let CX be the affine cone over X. We denote by I the ideal sheaf of X. The question whether ΩCX has torsion is purely algebraic, because the cone is affine. Denote by S = k[x0, . . . , xn] the polynomial ring with homo- geneous maximal ideal m = (x0, . . . , xn) and by R = S/I the homogeneous coordinate ring of X whose homogeneous maximal ideal we denote by n. Note that CX is smooth outside the vertex so we are only interested in the local behaviour at the vertex. In the following we will use some facts about local cohomology all of which can be found in [Eis05, Appendix 1] or in more detail in [Har67]. 3The sheaf of Kähler differentials for X1,2 and X1,3 was also computed in [Kni73] but the examples were not widely known. 4 DANIEL GREB AND SÖNKE ROLLENSKE In the following diagram the middle column is the conormal sequence and the maps to the maximal ideal are given by the contraction with the vectorfield ξ =Pi xi ∂ ∂xi . I/I 2 I/I 2 ξy ξy / n n / 0 / 0 0 0 / M ΩS ⊗ R ΩR / N 0 By definition, the modules M and N are the respective kernels of the con- traction maps ξy. The fact that the composition I/I 2 → ΩS ⊗ R → n is zero follows from Euler's formula: for every homogeneous element f ∈ I we have f =Pi xi ∂f ∂xi . Since n is torsionfree the torsion–submodule Tors(ΩR) is isomorphic to Tors(N ). Since CX is smooth outside the vertex we can compute this torsion submodule via local cohomology Tors(N ) = H 0 n (N ), which we now relate to sheaf cohomology on Pn. Introducing the appropriate grading we can transform the above diagram into a diagram of coherent sheaves on Pn, where the middle row is the restric- tion of the Euler Sequence to X and the first column becomes the conormal sheaf sequence for X. The map I/I 2 → ΩPnX in the latter sequence is injective because X is smooth. 0 0 I/I 2 I/I 2 0 0 / ΩPnX OX(−1)⊕n+1 / OX / ΩX 0 fΩR 0 OX / 0 / 0 / / / / / / / / / / / / / / / / / / / / / / TORSION AND COTORSION IN THE SHEAF OF KÄHLER DIFFERENTIALS 5 We now use the comparison sequence for local and sheaf cohomology ([Eis05, Cor. A1.12]) for the first column of the diagram to obtain (1) 0 / I/I 2 Γ∗(I/I 2) H 1 n (I/I 2) 0 0 0 / H 0 n (N ) / M / N 0 0 0 Γ∗(ΩPnX ) / H 1 n (M ) Γ∗(ΩX) H 1 n (N ) / 0 where we already used the following Lemma. Lemma 5 - In the above situation we have (i ) H 0 (ii ) H 0 n (I/I 2) = 0, n (M ) = 0. Proof. (i ) This follows from the short exact sequence 0 → I 2 → I → I/I 2 → 0 and the fact that I and I 2 are saturated and have depth at least 2. (ii ) M is a submodule of the free module ΩS ⊗ R, hence torsionfree. (cid:3) Applying the same arguments as in the proof of the snake lemma to (1) we get an exact sequence (2) 0 → H 0 n (N ) → H 1 n (I/I 2) → H 1 n (M ) → H 1 n (N ) n (I/I 2) = 0. This latter and thus have shown that ΩX is torsionfree if H 1 group can be interpreted as follows: the local cohomology sequence for 0 → I 2 → I → I/I 2 → 0 compared with the sheaf cohomology sequence of 0 → I 2 → I → I/I 2 → 0 yields a diagram (3) I ∼= I/I 2 0 Γ∗(I) / Γ∗(I/I 2) Ld H 1(I 2(d)) 7o o o o o /Ld H 1(I(d)) o o o o o o H 1 n (I/I 2) / / / / / / / / / / / / / / / / / / / / / / / / _ / / / / * 7 6 DANIEL GREB AND SÖNKE ROLLENSKE We have thus proved the first part of Theorem 3 from the introduction which we repeat here for the convenience of the reader. Theorem 3 - Let X ⊂ Pn k be a smooth projective variety over a field k of characteristic 0 and let I be the sheaf of ideals defining X. Let CX ⊂ An+1 be the affine cone over X. If H 1(Pn, I 2 If in addition X is projectively normal, then ΩCX is torsionfree if and only X(d)) = 0 for all d ≥ 0 then ΩCX is torsionfree. if also the first infinitesimal neighbourhood of X is projectively normal. Proof of the second part. By definition, projective normality of X is equiv- alent to H 0 exact sequence in local cohomology n (R) = 0 or equivalently Ld H 1(Pn, I(d)) = 0. The n (R) = H 1 0 = H 0 n (n) → H 1 n (M ) → H 1 n (ΩS ⊗ R) = H 1 n (R)⊕n+1 = 0, induced by 0 → M → ΩS ⊗ R → n → 0, shows that H 1 with (2) and (3) this vanishing implies that the composition n (M ) = 0. Together Tors(ΩR) = H 0 n (N ) ֒→ H 1 H 1(Pn, I 2(d)) n (I/I 2) ֒→Md is an isomorphism. This establishes the desired equivalence. (cid:3) 3. Zariski differentials on quotient singularities and cotorsion We briefly recall a result of Knighten, also discovered by Steenbrink [Ste77, Lem. 1.8], describing the double dual of ΩX/G on finite quotient singularities X/G, and deduce a sufficient condition for the existence of cotorsion in the sheaf of Kähler differentials. Let A be a regular local k–algebra and let G be a finite group acting on A. We denote by π : X = Spec A → X/G = Spec AG the quotient map. Theorem 6 ([Kni73], Theorem 3) - The natural map ΩX/G is an isomorphism. ∨∨ → (π∗ΩX)G In concrete situations the module of invariant differentials is not hard to compute and and the following will turn out to be useful Corollary 7 - Assume that X/G has an isolated singularity. Let m be the maximal ideal of AG and e = dimk m/m2 be the embedding dimension of X/G. If the minimal number of generators for (ΩX )G (considered as a AG–module) is bigger than e then ΩX/G has cotorsion and is not reflexive. Proof. By the conormal sequence the sheaf ΩX/G of Kähler differentials can be generated by e elements and thus can never surject onto (π∗ΩX)G if this sheaf needs more than e generators. (cid:3) TORSION AND COTORSION IN THE SHEAF OF KÄHLER DIFFERENTIALS 7 4. Examples: Cones over Veronese embeddings In this section we study cones over Veronese embeddings, including the examples mentioned in Table 1. The first subsection collects some general properties of these cones, torsion differentials are computed in Propostion 8, and cotorsion will be discussed in Proposition 10. 4.1. Basic properties. Here we collect basic properties of the cones over the Veronese embeddings. In particular, we discuss their realisation as cyclic quotient singularities, and compute the discrepancies of the canonical resolu- tion in order to determine in which cases these cones are terminal, canonical, etc. For this we follow [Rei80, §1] and [Deb01, Sect. 7.2], where one can also find the definition of the singularities appearing in the Minimal Model Pro- gram. Let µd be the cyclic group of order d and let ρ : µd → GLr+1(k) be the representation given by choosing a primitive dth root of unity ξ and sending a generator of µd to ξ . . . ξ   ∈ GLr+1(k). k Let Xr,d = Ar+1 /µd be the resulting quotient singularity (in Reid's [Rei87, (4.2)] notation, this is a singularity of type 1 d (1, 1, . . . , 1)). Since the ring C[x0, . . . , xr]µd is generated by all monomials of total degree d in the coor- dinates x0, . . . , xn, the quotient Xr,d is isomorphic to the cone Cr,d over the image Vr,d of the dth Veronese embedding vd : Pr → Pn (so n =(cid:0)r+d This isomorphism is induced by the map d (cid:1) − 1). (x0, . . . , xn) 7→ (. . . , Ya0+···+an=d xa0 0 · · · xan n , . . . ). The blow–up π : Yr,d → Xr,d of the origin in Xr,d is smooth, isomorphic to the total space of the line bundle OXr,d(−1). One can check (see [Rei80, p.278]) that the index of Xr,d is the denomi- g.c.d(r+1,d) . In particular, the canonical nator of (r + 1)/d, i.e., index(X) = divisor KXr,d is Q–Cartier, and Xr,d is Gorenstein if and only if d divides r + 1. We next compute the discrepancy of the unique exceptional divisor E ∼= Vr,d; for simplicity we suppress the indices in the notation. Since the canoncial divisor is Q–Cartier we can write d KY ∼ π∗KX + aE for a rational number a. The normal bundle of E in Y is OV (−1) ∼= OPr (−d). By adjunction, the canonical divisor of E is OE(KY +E). Hence, by restrict- ing to E we obtain −r − 1 = (a + 1)(−d), hence a = r + 1 d − 1. 8 DANIEL GREB AND SÖNKE ROLLENSKE It follows that all Veronese cones Xr,d are log terminal, and in addition that Xr,d is terminal (canonical) if and only if r + 1 > d (r + 1 ≥ d). 4.2. Torsion differentials. We now apply Theorem 3 to the Veronese cones. Proposition 8 - Let Xr,d be the affine cone over Vr,d ⊂ Pn, the image of the dth Veronese embedding of Pr (so n =(cid:0)r+d if and only if d ≥ 3. d (cid:1)−1). Then ΩXr,d has torsion Note that by the discussion in the previous section all cones Xr,d have a description as group quotients. So torsion differentials occur in abundance also on quotient singularities. Proof. We now fix r and d and denote by I the ideal sheaf of the image of the Veronese embedding V = vd(Pr). To avoid confusion we denote by H a hyperplane on Pn, so that OPn(mH)V = OPr (md). Since V is projectively normal the second part of Theorem 3 applies and we only have to check if there is an m such that H 1(Pn, I 2(mH)) 6= 0. By a result of Wahl [Wah97, Theorem 2.1] we have H 1(I 2 V (mH)) = 0 for m 6= 2 and thus we only need to consider the case m = 2. We will start by showing that ΩXr,d is torsionfree if d = 2. Recall that there is a Gaussian map (4) ΦOPr (d) : Λ2H 0(Pr, OPr (d)) → H 0(Pr, Ω1 Pr (2d)) symbolically given by s ∧ t 7→ sdt − tds, which in our case is a homomor- phism of SL(r + 1)–representations. Wahl showed that H 1(I 2 V (2H)) = ker Φ [Wah97, Proposition 1.8]. Note that the representation on H 0(Pr, Ω1 Pr (4)) is irreducible, cf. [Wah97, Section 2]. It follows by a straightforward computation that the map ΦOPr (2) is non-trivial and hence surjective. A calculation similar to the ones in Lemma 9 below gives the equality dim Λ2H 0(Pr, OPr (2)) = h0(Pr, Ω1 Pr (4)). Consequently, ΦOPr (2) is an isomorphism. Thus Tors(ΩXr,2) = ker(ΦOPr (2)) = {0}. This proves the "only if" part of the claim. For the existence of torsion–differentials in case d ≥ 3 we give an elemen- tary dimension estimate that does not depend on Wahls results. By projective normality we have an exact sequence (5) 0 → H 0(Pn, I 2(2H)) → H 0(Pn, I(2H)) → → H 0(Pn, I/I 2(2H)) → H 1(Pn, I 2(2H)) → 0, and H 1(Pn, I 2(2H)) does not vanish if h0(Pn, I/I 2(2H)) > h0(Pn, I(2H)). This is the content of Lemma 9 below. (cid:3) Lemma 9 - In the situation above the following holds (i ) h0(Pn, I(mH)) =(cid:0)(d+r r )+m−1 m (cid:1) −(cid:0)md+r r (cid:1). TORSION AND COTORSION IN THE SHEAF OF KÄHLER DIFFERENTIALS 9 r (cid:1)(cid:0)(m−1)d+r (iii ) h0(Pn, I/I 2(2H)) − h0(Pn, I(2H)) > 0 for all r ≥ 1, d ≥ 3. (cid:1) r (cid:1) − (dm − 1)(cid:0)r+dm−1 (ii ) h0(Pn, I/I 2(mH)) ≥(cid:0)d+r Ld≥0 Sd then the Veronese embedding is induced by the homomorphism of graded rings Sym∗(Sd) → S with kernel a graded ideal I. In degree m we get Proof. If we dentote the graded polynomial ring in r + 1 variables as S = (cid:1) −(cid:0)md+r md r 0 → H 0(I(mH)) = Im → Symm(Sd) → Smd → 0, where surjectivity of the last map can easily be checked on monomials. The formula for the dimension follows from the well known formula for the di- mension of a symmetric product. This proves the first item. From the embedding Pr ∼= V ⊂ Pn we get the (twisted) normal bundle sequence 0 → I/I 2(mH) → ΩPn(mH)V → ΩPr (md) → 0. The global sections of ΩPr (md) can be either computed via the Euler se- quence or read of from Bott's formula (see e.g. [Bot57]). Pulling back the Euler–sequence on Pn with vd we get 0 → v∗ d(ΩPn)(md) → OPr ((m − 1)d) ⊗ Sym1(Sd) → OPr (md) → 0. The map on global sections S(m−1)d ⊗ Sd → Smd is surjective and thus h0(Pn, I/I 2(mH)) d(ΩPn)(md)) − h0(Pr, ΩPr (md)) ≥ h0(Pr, v∗ = (n + 1)h0(Pr, OPr ((m − 1)d)) − h0(Pr, OPr (md)) − h0(Pr, ΩPr (md)) =(cid:18)d + r r (cid:19)(cid:18)(m − 1)d + r r (cid:19) −(cid:18)md + r r (cid:19) − (md − 1)(cid:18)r + md − 1 md (cid:19), which proves (ii ). It remains to prove (iii ). Putting together the formulas from the first two items for m = 2, we compute h0(Pn, I/I 2(2H)) − h0(Pn, I(2H)) 2d − (2d − 1)(cid:18)r + 2d − 1 ≥(cid:18)d + r r (cid:19)2 − (2d − 1)(cid:18)r + 2d − 1 =(cid:18)d + r r (cid:19)2 2 (cid:18)d + r −(cid:18)d + r r (cid:19)2 r (cid:1) + 1 (cid:19) r (cid:1)((cid:0)d+r r (cid:1) + 1) (cid:19)! . r (cid:19) − 2(2d − 1)(cid:18)r + 2d − 1 (cid:19) −(cid:18)(cid:0)d+r (cid:19) −(cid:0)d+r 2d 2d = 1 2 2 To obtain (iii ) it therefore suffices to show that (6) (cid:18)d + r r (cid:19)2 −(cid:18)d + r r (cid:19) − 2(2d − 1)(cid:18)r + 2d − 1 2d (cid:19) > 0. 10 DANIEL GREB AND SÖNKE ROLLENSKE For r = 1 this formula reduces to d2 − 3d + 2 > 0 which is certainly true for d ≥ 3. We now proceed by induction on r. For convenience, note that for r = 2 and d = 3 the expression on the left–hand side of (6) gives 20, so we may assume d ≥ 4 if r = 2. The induction step follows from the computation below, in which we apply the induction hypothesis twice (in steps 3 and 5) and use some standard identities for binomial coefficients. −(cid:18)d + r + 1 r + 1(cid:19)(cid:19)2 (cid:18)d + r + 1 r + 1 (cid:19)2 r + 1 (cid:19) − 2(2d − 1)(cid:18)r + 2d 2d (cid:19) =(cid:18)(cid:18)d + r r (cid:19) +(cid:18)d + r −(cid:18)d + r r (cid:19) −(cid:18)d + r r + 1(cid:19) − 2(2d − 1)(cid:18)(cid:18)r + 2d − 1 (cid:19) +(cid:18)r + 2d − 1 2d − 1 (cid:19)(cid:19) = (cid:18)d + r (cid:19)! −(cid:18)d + r r (cid:19) − 2(2d − 1)(cid:18)r + 2d − 1 r (cid:19)2 + 2(cid:18)d + r r (cid:19)(cid:18)d + r r + 1(cid:19) +(cid:18)d + r −(cid:18)d + r r + 1(cid:19)2 r + 1(cid:19) − 2(2d − 1)(cid:18)r + 2d − 1 2d − 1 (cid:19) > 2(cid:18)d + r r (cid:19)(cid:18)d + r r + 1(cid:19) +(cid:18)d + r −(cid:18)d + r r + 1(cid:19)2 r + 1(cid:19) − 2(2d − 1)(cid:18)r + 2d − 1 2d − 1 (cid:19) r + 1(cid:18)d + r r + 1(cid:18)d + r +(cid:18) d r (cid:19)2 r (cid:19)(cid:19)2 2(2d − 1)(cid:18)r + 2d − 1 r + 1(cid:18)d + r r (cid:19) − (cid:18)d + r r + 1(cid:18)d + r r (cid:19)2 r (cid:19) − =(cid:18) d2 + 2d(r + 1) +(cid:18) 2d r (cid:19)(cid:18)d + r r (cid:19)2 r(r + 1)2 (r(d − 2) − 2)(cid:18)d + r r (cid:19)2 (cid:19) r (cid:19)! r (cid:18)d + r −(cid:18)d + r r (cid:19)2 r + 1(cid:19)(cid:18)d + r r (cid:19) d2 + 2d(r + 1) 2d r − (r + 1)2 2d = (r + 1)2 d − 2d 2d d − r 2d − > d > > 0, 2d 2d d where in the last step we used d ≥ 3, and d ≥ 4 if r = 2. This concludes the induction and the proof of (iii ). (cid:3) 4.3. Cotorsion in ΩXr,d. In this section we compute the cotorsion of the cones over the Veronese embeddings using their realisations as cyclic quotient singularities. Proposition 10 - For all r ≥ 1, d ≥ 2 the sheaf ΩXr,d has cotorsion. TORSION AND COTORSION IN THE SHEAF OF KÄHLER DIFFERENTIALS 11 Proof. By Corollary 7 we need to compare the number of generators of Ωµd and the embedding dimension of Xr,d. Ar+1 Xr,d is n =(cid:0)r+d Recall that the ring k[x0, . . . , xr]µd is generated by all monomials of total degree d in the coordinates x0, . . . , xr and thus the embedding dimension of Ar+1 has a mini- mal system of homogeneous generators given by all products of monomials of degree r − 1 with dx0, . . . , dxr. Subtracting the embedding dimension from the number of these generators we obtain d (cid:1). On the other hand, the k[Xr,d]-module Ωµr d − 1 (cid:19) −(cid:18)d + r d + r(cid:18)d + r (r + 1)(cid:18)d + r − 1 d (cid:19) = (r + 1) d d (cid:19) −(cid:18)d + r d (cid:19) (cid:18)d + r d (cid:19) d(r + 1) − (d + r) d + r r(d − 1) d + r (cid:18)d + r d (cid:19) = = > 0, where in the last step we use d ≥ 2. It follows that ΩXn,r has cotorsion. (cid:3) Acknowledgements: We thank Duco van Straten for an interesting discus- sion and Miles Reid for many helpful comments on an earlier version of this note. After posting of the first version of this paper on the arXiv Jonathan Wahl pointed us to [Wah97], which led to a strenghening of Proposition 8. The first author was partly supported by the DFG–Forschergruppe 790 "Classification of Algebraic Surfaces and Compact Complex Manifolds" (Bay- reuth – Freiburg). During the preparation of the second version of the pa- per, he enjoyed the hospitality of the Mathematics Department at Princeton University. He gratefully acknowledges the support of the "Eliteprogramm für Postdoktorandinnen und Postdoktoranden" of the Baden–Württemberg– Siftung. The second author gratefully acknowledges support by the DFG via the SFB/TR 45 "Periods, moduli spaces and arithmetic of algebraic varieties" and his Emmy–Noether project, and partial support by the Hausdorff Centre for Mathematics in Bonn. [Ber63] Robert Berger. Differentialmoduln eindimensionaler lokaler Ringe. Math. Z., 81:326–354, 1963. References [BG80] [Bot57] [BLLS02] Paulo Brumatti, Yves Lequain, Daniel Levcovitz, and Aron Simis. A note on the Nakai conjecture. Proc. Amer. Math. Soc., 130(1):15–21 (electronic), 2002. Raoul Bott. Homogeneous vector bundles. Ann. of Math. (2), 66:203–248, 1957. Ragnar-Olaf Buchweitz, Gert-Martin Greuel. The Milnor number and defor- mations of complex curve singularities. Invent. Math., 58(3):241–281, 1980. Olivier Debarre. Higher-dimensional algebraic Springer-Verlag, New York, 2001. David Eisenbud. The geometry of syzygies, volume 229 of Graduate Texts in Mathematics. Springer-Verlag, New York, 2005. geometry. Universitext. [Deb01] [Eis05] 12 DANIEL GREB AND SÖNKE ROLLENSKE [GKKP10] Daniel Greb, Sandor Kovács, Stefan Kebekus, and Thomas Peternell. Differ- [Gre82] ential Forms on Log Canonical Spaces. arXiv:1003.2913, 2010. Gert-Martin Greuel. On deformation of curves and a formula of Deligne In Algebraic geometry (La Rábida, 1981), pages 141–168. Lecture Notes in Math. 961, Springer-Verlag, Berlin, 1982. [Har67] [KM98] [GMP85] Gert-Martin Greuel, Bernd Martin, Gerhard Pfister. Numerische Charak- terisierung quasihomogener Gorenstein-Kurvensingularitäten. Math. Nachr., 124:123–131, 1985. Robin Hartshorne. Local cohomology, volume 1961 of A seminar given by A. Grothendieck, Harvard University, Fall. Springer-Verlag, Berlin, 1967. János Kollár and Shigefumi Mori. Birational geometry of algebraic varieties, volume 134 of Cambridge Tracts in Mathematics. Cambridge University Press, Cambridge, 1998. With the collaboration of C. H. Clemens and A. Corti. Carol M. Knighten. Differentials on quotients of algebraic varieties. Trans. Amer. Math. Soc., 177:65–89, 1973. Ernst Kunz. Kähler differentials. Advanced Lectures in Mathematics. Friedr. Vieweg & Sohn, Braunschweig, 1986. [Kun86] [Kni73] [MvS01] David Mond and Duco van Straten. Milnor number equals Tjurina number for [Pla88] [Poh91] [Rei80] [Rei87] [Ste77] [Wah97] functions on space curves. J. London Math. Soc. (2), 63(1):177–187, 2001. Erich Platte. Differentielle Methoden in der lokalen Algebra, volume 10 of Osnabrücker Schriften zur Mathematik. Universität Osnabrück Fachbereich Mathematik, Osnabrück, 1988. Thomas Pohl. Differential modules with maximal torsion. Arch. Math. (Basel), 57(5):438–445, 1991. Miles Reid. Canonical 3-folds. In Journées de Géometrie Algébrique d'Angers, Juillet 1979/Algebraic Geometry, Angers, 1979, pages 273–310. Sijthoff & No- ordhoff, Alphen aan den Rijn, 1980. Miles Reid. Young person's guide to canonical singularities. In Algebraic geom- etry, Bowdoin, 1985 (Brunswick, Maine, 1985), volume 46 of Proc. Sympos. Pure Math., pages 345–414. Amer. Math. Soc., Providence, RI, 1987. J. H. M. Steenbrink. Mixed Hodge structure on the vanishing cohomology. In Real and complex singularities (Proc. Ninth Nordic Summer School/NAVF Sympos. Math., Oslo, 1976), pages 525–563. Sijthoff and Noordhoff, Alphen aan den Rijn, 1977. Jonathan Wahl. On cohomology of the square of an ideal sheaf. J. Algebraic Geom., 6(3):481–511, 1997. Daniel Greb, Institut für Mathematik, Abteilung für Reine Mathematik, Albert–Ludwigs–Universität Freiburg, Eckerstrasse 1, 79104 Freiburg im Breisgau, Germany Current address: Mathematics Department, Princeton University, Fine Hall, Washing- ton Road, Princeton NJ 08544-1000, USA E-mail address: [email protected] Sönke Rollenske, Institut für Mathematik, Johannes Gutenberg Univer- sität Mainz, 55099 Mainz, Germany E-mail address: [email protected]
1008.5027	1	1008	2010-08-30T09:20:57	On some lattice computations related to moduli problems	[ "math.AG", "math.CO" ]	We show how to solve computationally a combinatorial problem about the possible number of roots orthogonal to a vector of given length in $E_8$. We show that the moduli space of K3 surfaces with polarisation of degree 2d is also of general type for d=52. This case was omitted from the earlier work of Gritsenko, Hulek and the second author. We also apply this method to some related problems. In Appendix A, V. Gritsenko shows how to arrive at the case d=52 and some others directly.	math.AG	math	On some lattice computations related to moduli problems A. Peterson and G.K. Sankaran, with an appendix by V. Gritsenko May 10, 2021 Abstract The method used in [GHS1] to prove that most moduli spaces of K3 surfaces are of general type leads to a combinatorial problem about the possible number of roots orthogonal to a vector of given length in E8. A similar problem arises for E7 in [GHS2]. Both cases were solved partly by computer methods. We use an improved computation and find one further case, omitted from [GHS1]: the moduli space F2d of K3 surfaces with polarisation of degree 2d is also of general type for d = 52. We also apply this method to some related problems. In Appendix A, V. Gritsenko shows how to arrive at the case d = 52 and some others directly. Many moduli spaces in algebraic geometry can be described as locally sym- metric varieties, i.e. quotients of a Hermitian symmetric domain D by an arithmetic group Γ. One method of understanding the birational geometry of such quotients is to use modular forms for Γ to give information about differential forms on Γ\D. In [GHS1] this method was used to prove that the moduli space F2d of polarised K3 surfaces of degree 2d is of general type in all but a few cases. The method works if there exists a modular form of sufficiently low weight with sufficiently large divisor. In [GHS1], and again in [GHS2] where a similar method was applied to certain moduli of polarised hyperkahler manifolds, the required modular form is constructed by quasi-pullback of the Borcherds form Φ12. A suitable quasi-pullback exists if a combinatorial condition is satisfied: there should exist a vector l in the root lattice E8 (or E7 in the hyperkahler case) of square 2d, orthogonal to very few roots. This is evidently the case if d is large, but for small d the search for such an l invites the use of a computer. This was done in both [GHS1] and in [GHS2] by a randomised search, relying on the large Weyl group to ensure that in practice no cases would be missed. Here we present an exhaustive search carried out by the first author. For the hyperkahler case the exhaustive search confirmed the results of the earlier randomised search, but in the K3 case one previously overlooked 1 value of d with a suitable vector was found, namely d = 52. In fact it turned out that the randomised search had indeed found this value, and the omission of the case d = 52 from [GHS1] happened because the output had been interpreted incorrectly. 1 Nevertheless the following result is true and has not previously appeared in the literature. Theorem 1 The moduli space F2·52 of K3 surfaces with polarisation of de- gree 104 is of general type. The paper is organised as follows. In Section 1 we explain briefly what the combinatorial problem is and how it arises, and give some more general combinatorial problems of the same nature. In Section 2 we describe the theoretical and computational methods used to solve it, along with some other results obtained in the same way. In Appendix A, Valery Gritsenko explains explains how the case d = 52 could have been foreseen without the help of a computer. Some of the computer code is given in Appendix B. Acknowledgements: Part of this paper forms part of A. Peterson's Masters' thesis. He would like to thank Gerard van der Geer for his supervision, and the University of Amsterdam for the nice environment it provides. The sec- ond author would like to thank the Fondazione Bruno Kessler in Trento and the Max-Planck-Institut fur Mathematik in Bonn for support, and Valery Gritsenko for helpful conversations. 1 Combinatorial problems and moduli In this section we first give a list of combinatorial questions and then explain the geometry that originally motivated them. First we fix some terminology. We say that L is a lattice of signature (a, b) if L ∼= Za+b and we fix a bilinear form ( , ) : L × L → Z of signature (a, b). If x ∈ L we refer to (x, x) as x2 and call it the length of x. If the length of x is 2 then x is called a root. If the roots of L generate L as an abelian group then L is called a root lattice. A lattice L is unimodular if it is equal to its dual L∨ = Hom(L, Z) ⊇ L. We do not assume that L is always unimodular but for simplicity we do assume that L is even, i.e. that x2 is always an even integer. E8 denotes the unique even unimodular positive-definite lattice of rank 8, i.e. with signature (8, 0): this is the sign convention of [Bou] and is also If n ∈ 2Z then hni is the rank 1 lattice spanned by a used in [GHS1]. vector of length n, and U denotes the integral hyperbolic plane Ze + Zf with e2 = f 2 = 0 and (e, f ) = 1. The symbol ⊕ denotes the orthogonal direct sum of lattices. If Λ is a lattice and n ∈ Z, then Λ(n) denotes the same lattice with the quadratic form multiplied by n. In particular, E8(−1) is the negative-definite even unimodular lattice of rank 8. 1By me. -- GKS 2 1.1 Combinatorial problems Let Λ be a root lattice (usually it will be E8 or E7) and denote by R(Λ) the set of its roots, i.e. R(Λ) = {r ∈ Λ r2 = 2}. The combinatorial questions arising in [GHS1] and [GHS2] are special cases of the following. Question 1 Given integers p > q ≥ 0, what are the values of d for which every vector of length 2d that is orthogonal to at least 2q roots is orthogonal to at least 2p roots? More generally we may ask about all possibilities. Question 2 Given an even natural number 2d, what are the possible num- bers of roots orthogonal to a vector of length 2d? If l ∈ Λ we denote by R(l⊥) the system of roots of Λ orthogonal to l. We denote the answer to Question 2 by P (Λ, d): that is P (Λ, d) := {m ∈ Z ∃l ∈ Λ l2 = 2d, #R(l⊥) = m}. (1) Thus P (Λ, d) is a finite set of even non-negative integers. We call this the root type of the non-negative even integer 2d for the lattice Λ There are some immediate restrictions on what the root type can be: for example, if Λ = E8 then the largest m that can occur is 126, when R(l⊥) ∼= E7; but in that case l ∈ (E7)⊥ E8 ∼= A1, so d must be a square. Especially for Λ = E8, the value of m0(d) = min P (E8, d) is of interest as it determines the lowest weight of modular form obtained by quasi-pullback (see Equation (2) below). If m0(d) = 0 then this form will not be a cusp form, so the value of m1(d) = min P (E8, d)∩ N is also significant. We should also like to know whether this form is unique. So we also have the following questions. Question 3 For given d and Λ, how can we compute m0(d)? Question 4 For given m, what is the smallest value d(m) of d for which m1(d) ≤ m? If in Question 4 we replace m1 by m0, then the case m = 0 asks for the length of shortest vectors in the interior of a Weyl chamber: these are the Weyl vectors, which are well known. If m ∈ P (Λ, d) there is a further natural refinement. Question 5 How many Weyl group orbits of vectors l with l2 = 2d and #R(l⊥) = m are there? 3 Some values of m are of particular interest for geometric reasons: for in- stance, if 14 ∈ P (E8, d) then quasi-pullback of Φ12 gives a canonical form on F2d (see Section 1.2 below). This leads us to the following variant of Question 1. Question 6 For given m and Λ, what are the values of d such that m ∈ P (Λ, 2d)? We can compute the answers to some cases of these questions by the methods described in Section 2. 1.2 Moduli The following construction describes several moduli spaces in algebraic ge- ometry, including the moduli of polarised K3 surfaces. Let L be an even lattice of signature (2, n). The Hermitian symmetric domain associated with L is DL, one of the two connected components of DL ∪ DL = {[w] ∈ P(L ⊗ C) w2 = 0, (w, w) > 0}. The group O(L) of isometries of L acts on this union and we denote by O+(L) the index 2 subgroup preserving DL. The action is discontinuous, with finite stabilisers, so if Γ is any finite index subgroup of O+(L) then FL(Γ) := Γ\DL is a complex analytic space. In fact it is a quasi-projective variety, hav- ing a minimal projective compactification, the Baily-Borel compactifica- tion FL(Γ)∗, obtained by adding finitely many curves (called 1-dimensional It is of- cusps) meeting at finitely many points (0-dimensional cusps). ten preferable to work with a toroidal compactification FL(Γ), which is a modification of FL(Γ)∗ depending on some combinatorial choices at the 0-dimensional cusps. A modular form for Γ of weight k and character χ : Γ → C∗ is a holomor- phic function F on the affine cone D• L ⊂ L ⊗ C such that and F (tZ) = t−kF (Z) ∀t ∈ C∗ F (gZ) = χ(g)F (Z) ∀g ∈ Γ. F is a cusp form if it vanishes at every cusp. For the cases we shall consider the only possible characters are 1 and det(g), and the order of vanishing at a cusp is an integer: see [GHS3]. The aim of [GHS1] is to show that the moduli space F2d of polarised K3 surfaces of degree 2d is of general type for most values of d ∈ N. Using the Torelli theorem for K3 surfaces one can show that + (L2d)), F2d = FL2d(eO 4 + (L) is the finite index subgroup of O+(L) that acts trivially on the where eO discriminant group L∨/L and L2d := 2U ⊕ 2E8(−1) ⊕ h−2di. Modular forms of suitable weight can be interpreted as differential forms on the moduli space provided that they have sufficiently large divisor. There- fore, to prove that the moduli space is of general type it is enough to give a sufficient supply of such modular forms. There are several technical difficul- ties here, one of which is the presence of singularities. A sufficient condition, however, was given in [GHS1]. Theorem 2 Suppose that n ≥ 9 and that there exists a nonzero cusp form Fa of weight a < n and character χ ≡ 1 or χ(g) = det(g), vanishing along any divisor H ⊂ DL fixed by reflections in Γ. Then FL(Γ) is of general type. The form Fa is then used to give many forms of high weight with suf- a F(n−a)k, and these in turn give ficiently large divisor, of the form F = F k pluricanonical forms on a smooth model of FL(Γ). To apply this in specific cases such as F2d one must therefore construct Fa. The method used in [GHS1] to do this is quasi-pullback of the Borcherds form Φ12. This construction first appeared in [BKPS]. The Borcherds form itself was constructed in [Bor] by means of a product expansion, whereby its divisor is evident. It is a modular form (not a cusp form) of weight 12 and character det for the group O+(II2,26). The lattice II2,26 of signature (2, 26) is 2U ⊕ N (−1), where N is any one of the 24 Niemeier lattices, positive definite unimodular lattices of rank 24: see [CS]. For our purposes the correct choice of N is 3E8. A choice of a (not necessarily primitive) vector l ∈ E8 of length 2d gives an embedding L2d = 2U ⊕ 2E8(−1) ⊕ h−2di ֒→ II2,26 = 2U ⊕ 3E8(−1) which in turn gives an embedding D• L2d ֒→ D• II2,26. If r ∈ L is a root it determines a Heegner divisor H• Denote the images of these embeddings by L2d[l] and D•[l] respectively. L, given by the equation (Z, r) = 0. The Borcherds form vanishes (to order 1) along all the Heegner divisors for L = II2,26 and in particular its restriction to D•[l] vanishes, as needed to apply Theorem 2. However, Φ12D•[l] may well be zero, since if r is a root of II2,26 orthogonal to L2d[l] then D•[l] ⊂ H• r. Instead we take the quasi-pullback, simply dividing by the equation of each such H• r ⊂ D• r , noting that H• −r = H• r. We put Rl = {r ∈ R(II2,26) (r, L2d[l]) = 0} ∼= {r ∈ R(E8) (r, l) = 0} 5 and define the quasi-pullback to be F [l] = . (2) Φ12 Q±r∈Rl (r, Z)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)D•[l] This is a nonzero modular form, and one can show that it is a cusp form pro- vided Rl 6= ∅. It vanishes along all the Heegner divisors fixed by reflections in O+(L2d). The weight, however, goes up by 1 every time we divide, so the weight of F [l] is 12 + 1 2 #Rl. We can therefore show that F2d is of general type if we can find an l ∈ E8 of length 2d with 2 ≤ #Rl < 2(n − 12) = 14. Moreover, if we can find a cusp form of weight precisely n = 19 then, by a result of Freitag [Fr], F2d has pg > 0 and in particular is not uniruled. This leads us to Question 1, with q = 1 and p = 7 or p = 8, for Λ = E8. In [GHS2], similar considerations about the moduli of some hyperkahler manifolds with a certain type of polarisation lead to Question 1 with q = 1 and p = 6 or p = 7, for Λ = E7. 2 Solving the combinatorial problems The specific combinatorial problems encountered in [GHS1] and [GHS2] can be solved in principle by first bounding d. It is clear that for sufficiently large d an l will exist orthogonal to a number of roots in the required range: indeed, for sufficiently large d we can find l orthogonal to exactly two roots. An explicit bound, followed by a finite calculation, will solve the problem. Neither is entirely straightforward, though. In [GHS1] a counting argument is used to show that an l ∈ E8 with l2 = 2d, orthogonal to at least two and at most 12 roots, exists (and therefore F2d is of general type) unless 28NE6 (2d) + 63ND6 (2d) ≥ 4NE7(2d), (3) where NL(2d) is the number of ways of representing 2d by the quadratic form L. The inequality (3) certainly fails for large d, but to obtain an effective bound on d one must bound NE6(2d) and ND6(2d) from above and NE7(2d) from below by explicit functions. This is a non-trivial problem in analytic number theory but it can be done, and after some refinements it gives a reasonable bound of around d = 150. It would be possible to resort to direct computation at that point, but there is no need yet. Some integers in that range are excluded from the list of possibly non-general type polarisations because the inequality (3) (or another similar inequality) in fact fails. Others can be excluded by inspection, actually producing a vector l by guessing the root system R(l⊥ E8). The root systems used in this way in [GHS1] were 4A1, 2A1 ⊕ A2, A3 and A1 ⊕ A2. The root systems 3A1 ⊕ A2 and 2A2 were not tried: see Appendix A. 6 In [GHS2] a similar procedure was used, although there is an extra dif- ficulty caused by the opposite parity of the rank: working in E7, one needs to estimate NR(2d) from above for some odd-rank root systems R, and this problem is not so well studied as in the even rank case. In either case, eventually one is left with a residual list of values of d for which the problem has not been settled. In [GHS1] it consists of most integers between 15 and 60 (for very small d the moduli space is known to be unirational). The residual problem in the hyperkahler case considered in [GHS2] is much smaller. Now, if we want to be (reasonably) sure that no cases have been missed, we do need a computer. Moreover, the methods we now use to solve this problem can also be used to give answers to question such as those posed in Section 1.1. 2.1 Algorithms We begin by representing E8 in the usual way, as the set of points l = (l1, . . . , l8) ∈ R8 such that the li are either all integers or all strict half- integers (i.e. either li ∈ Z for all i or 2li is an odd integer for all i) and P li ∈ 2Z, with the standard Euclidean quadratic form on R8. We need a very rough upper bound on NE8(2d), because we want to know whether NE8(2d) is small enough to allow a brute-force search for l ∈ E8 with l2 = 2d having 2 ≤ #R(l⊥) ≤ 12. We can easily find such a bound by noting that if l2 = 2d then each of the 8 components li of l must have i ≤ 2d, so −√2d ≤ li ≤ √2d, and must be a half-integer: that gives l2 NE8(2d) ≤ (2⌊2√2d⌋ + 1)8 (4) For d = 52, this bound is about 8 · 1012. If we are a bit more precise, and note that the components of l are either all integers, or all proper (i.e. non-integer) half-integers, we save a factor 27, giving a bound of about 5 · 1010. This is within reach of a brute-force search, but it is still high, especially considering that we have to do some substantial work for each candidate (compute the inner product with 240 different vectors2). Thus an exhaustive search of all vectors in E8 of length ≤ 60 is not computationally impossible but it would be cumbersome and would not extend to even slightly larger problems such as other cases of Question 1. The Weyl group W (E8) has order 214 · 35 · 52 · 7 = 696729600 and should be used to reduce the size of the problem. There are two approaches to doing this. 2We can be a lot more efficient than that, and skip most of these inner products, but even then we still have to compute dozens of inner products per candidate vector. 7 A. Randomised search. This is what was actually done in [GHS1] and [GHS2]. Since the non-existence of a vector l gives no information about the moduli space, we are willing to accept a very small probability of failing to detect such a vector. We therefore choose a large number of vectors of length less than 2 · 61 at random and expect that, as the Weyl group orbits are large, every orbit will be represented. This approach worked very fast, using only a laptop computer and imme- diately available software (Maple). A search of twenty thousand randomly chosen vectors found all the pairs (d, #R(l⊥)) in the ranges wanted within the first two thousand iterations, in approximately two minutes. That is fairly convincing practical evidence that there are no more. Unfortunately the output was then mistranscribed, leading to the omission of the case d = 52 and the erroneous (but not really misleading) statement in [GHS1] that "an extensive computer search for vectors orthogonal to at least 2 and at most 14 roots for other d has not found any". It is noteworthy that a similar search in the case Λ = E7 did find some cases not discovered analytically, and for which a constructive method of finding l is still not known. In other words, some cases of the main theorem of [GHS2] still have only a computer proof, although once l has been found it is easy enough to verify its properties by hand. It is not so easy to estimate the probability a priori that a Weyl orbit might be missed. The Weyl group of R(l⊥), which is a subgroup of the Weyl group of E8, obviously stabilises l and has order no more than 24 if #R(l⊥) ≤ 12, but in principle the stabiliser of l in W (E8) could be much larger. In that case the Weyl group orbit would be small and more easily missed. In practice the randomised method seems to find all the orbits. B. Exhaustive search. The first author organised an exhaustive search, exploiting the Weyl group by searching a fundamental domain for the sub- group H < W (E8) generated by permutations of the eight components li and sign changes of an even number of components. This subgroup H has size 27 · 8!, so index 135 in W (E8): it gives us most of the symmetries, with very little effort. We say that l ∈ E8 is in normal form if its components are all nonneg- ative (except possibly the first, l1) and the squares of the components are nondecreasing from low index to high index. By acting with an element of H, we can translate any l ∈ E8 to one in normal form: first permute the components, so their squares are in order; then make them all (but l1) nonnegative, by changing the sign of every negative component (except l1), and flipping the sign of l1 once for every such change. It is straightforward to enumerate the elements of length 2d in E8 that are in normal form. For brevity, we will describe this only for the ones having integer components (one can get the ones with proper half-integer components in a very similar manner). 8 Step 1. For every index i 6= 1, in descending order, we consider all the possible values of li: we require li to be a non-negative integer such that • its square, added to the sum of the squares of the coordinates that j with j > i), does not exceed 2d (otherwise have been chosen (i.e. the l2 l2 > 2d, for any further choice of coordinates); and • (unless i = 8) it is not greater than li+1 (otherwise l would not be in normal form). In other words, we let li take any value s ∈ Z such that j(cid:27) . 0 ≤ s ≤ min(cid:26)li+1,r2d −Xj>i l2 (5) j=2 l2 j is a perfect square m2. If so, let l1 take values Step 2. See if 2d −P8 −m and m; if not, discard this choice of coordinates. Step 3. Check whether the l so obtained are in E8, i.e. whetherP8 j=1 lj ∈ 2Z. Discard any that are not in E8. We must then filter these enumerated l ∈ E8 to find the ones with #R(l⊥) in the required range (2 ≤ #R(l⊥) ≤ 12 for the case considered in [GHS1]): this part of the procedure is exactly the same as for the randomised version. Since the roots come in pairs ±r it is enough to take inner products with a prepared list of positive roots (120 or them), and of course we can stop examining l as soon as we find a seventh pair of roots orthogonal to it. The first author implemented this search in a high-level programming language (Haskell). Without spending much time optimising, this runs fast enough (a second or so on commercial hardware, for each of the low values of d we are interested in, namely d ≤ 60). The partial use of the symmetries of E8 is crucial, though: to go through all the vectors of given length 2d would have taken weeks or months for a single value of d. This program discovered the lost case d = 52 and therefore Theorem 1. A variant of it for E7 reconfirmed the results obtained by the randomised method in [GHS2]. The code used for the E8 case is given in Appendix B. 2.2 Further results The exhaustive algorithm (B) from Section 2.1 can be modified to compute, in reasonable time, answers to some of the questions from Section 1.1 for small values of the parameters. We investigated Question 2 and and Ques- tion 6 for small m and d with Λ = E7 and Λ = E8. For Λ = E8 we also investigated Question 5 for the particular case m = 14, corresponding to canonical forms on F2d. Specifically, we have so far computed the root type P (Λ, 2d) for Λ = E7 and Λ = E8 and d ≤ 150, and the first part of the root type (whether m ∈ P (Λ, 2d) for 2 ≤ m ≤ 20, say) for larger d, up to about 300 (further 9 for some values of d). This part of the computation is fairly fast and only minor changes to the program are needed. A little more work, and more computer time, is needed for Question 5. We must work now with W (E8), not with H, and we first compute a transversal for W (E8) : H (representatives for each of the 135 left cosets of H) and then reduce each of the 135 translates of each l to standard form before comparing them. The outcome counts the number of ways of obtaining a canonical form on F2d by quasi-pullback of Φ12. There is no assurance either that the forms so obtained are linearly independent or that there are not more canonical forms that do not arise this way. The results are nevertheless intriguingly unpredictable. There are no such vectors for d < 40. There is such a vector for d = 40, and also for d = 42, 43, 48 (two orbits), 49, 51 -- 54, 55 and 56 (two orbits each), 57 and 59. There is no such vector for d = 60, but for 61 there are three orbits and thereafter the number of orbits drifts upwards irregularly. Without further comment, we tabulate below the number ν14 of W (E8) orbits of length 2d vectors in E8 orthogonal to exactly 14 roots for 61 ≤ d ≤ 150. ν14 3 1 2 2 0 2 1 2 2 1 2 2 1 3 3 d 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 d 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 d 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 ν14 8 7 5 11 5 6 8 3 8 8 7 11 5 10 6 d 121 122 124 124 125 126 127 128 129 130 131 132 133 134 135 ν14 4 5 5 3 6 8 6 6 7 4 9 2 8 9 5 ν14 5 3 2 4 3 4 2 3 2 4 5 5 5 4 4 d 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 d 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 ν14 1 2 1 4 2 2 2 3 5 4 4 3 2 3 2 ν14 2 6 3 6 0 6 6 5 3 7 6 2 6 9 8 A Appendix: d = 46, 50, 52, 54, 57, by V. Grit- senko In this appendix we find a vector l ∈ E8 of square 2d orthogonal to exactly 12 roots in E8, where d is as in the title of the appendix. (See [GHS1] and [GHS2] for the general context of this question.) We use below the combinatorics of the Dynkin diagram of E8. We take the Coxeter basis of 10 simple roots in E8 as in [Bou]: α4 ✲t α3 ✲t α1 t α5 ✲t α6 ✲t α7 ✲t α8 ✲t where (e1, . . . , e8) is a Euclidean basis in the lattice Z8 and ❄t α2 1 2 1 2 (e1 + e8) − α1 = (e2 + e3 + e4 + e5 + e6 + e7), α2 = e1 + e2, αk = ek−1 − ek−2 (3 ≤ k ≤ 8). The lattice E8 contains 240 roots. We recall that any root is a sum of simple roots with integral coefficients of the same sign. The fundamental weights ωj of E8 form the dual basis in E8 = E∨ 8 , so (αi, ωj) = δij. The formulae for the weights are given in [Bou, Tabl. VII]. The Cartan matrix of the dual basis is ((ωi, ωj)) = . (6)   10 4 6 8 7 8 10 15 12 6 5 9 8 10 14 20 16 12 2 4 3 5 4 7 10 15 20 30 24 18 12 6 12 16 24 20 15 10 5 8 6 9 4 3 6 4 2 3 2 12 18 15 12 8 8 4 4 12 10 6 5 8 6 3   We consider the two following cases when the orthogonal complement of a vector l in E8 contains exactly 12 roots: R(l⊥ E8) = A2 ⊕ 3A1 or A2 ⊕ A2. (We note that #R(A1) = 2 and #R(A2) = 6.) 2 2 = hα1, α3i, A(2,4) 2 = hα2, α4i, A(5,6) I: d = 46, 50, 54, 57. There are four possible choices of the subsystem A2⊕ 3A1 inside the Dynkin diagram of E8 according to the choices of simple roots of A2, namely A(1,3) 2 = hα5, α6i or A(7,8) 2 = hα7, α8i. If A2 is fixed then the three pairwise orthogonal copies of A1 in the Dynkin diagram are defined automatically. = hα5, α6i. Then 3A(5,6) First, we consider A(5,6) = hα2i ⊕ hα3i ⊕ hα8i. 2 ⊕ 3A(5,6) is the root system of the orthogonal complement of the vector l5,6 = ω1 + ω4 + ω7 ∈ E8. In fact, if r =P8 i=1 xiαi is a positive root (xi ≥ 0) then (r, l5,6) = x1 + x4 + x7 = 0. Therefore x1 = x4 = x7 = 0 and r belongs to A(5,6) . Using the Cartan matrix (6) we obtain that l2 5,6 = 2 · 46. Doing similar calculations with the other three copies of A2 given above we find 2 ⊕ 3A(5,6) Moreover A(5,6) 1 1 1 l1,3 = ω4 + ω6 + ω8, l2,4 = ω3 + ω5 + ω7, l7,8 = ω1 + ω4 + ω6 11 1,3 = 2 · 50, l2 2,4 = 2 · 54 and l2 with l2 II: d = 52. We consider the sublattice M = A2⊕A2 = hα3, α4i⊕hα6, α7i in E8. Then M is the root system of the orthogonal complement of the vector lM = ω1 + ω2 + ω5 + ω8 with l2 7,8 = 2 · 57. M = 2 · 52. V.A. Gritsenko, Universit´e Lille 1, Laboratoire Paul Painlev´e, F-59655 Vil- leneuve d'Ascq, Cedex, France [email protected] B Appendix: Computer code Below is the code used to check the combinatorial problem from [GHS1], and thus to find Theorem 1. The programs were written in the functional programming language Haskell (http://www.haskell.org). The web page http://people.bath.ac.uk/masgks/Rootcounts contains links to further code and output. {-# LANGUAGE TypeSynonymInstances,NoImplicitPrelude #-} module E8 where Control.Applicative import qualified Algebra.Ring import import qualified Data.Vector import import qualified Data.MemoCombinators as Memo import Data.Ratio Data.List as V ((<$>),(<>)) (intercalate,nubBy) (Ratio,numerator,denominator,(%)) import qualified Data.Set import import Data.Typeable Math.Combinatorics.Species as Set (Typeable) (ksubsets,set,ofSize,enumerate,Set(getSet,Set),Prod(Prod)) MyPrelude hiding (numerator,denominator,(%)) import import qualified Prelude import import qualified Algebra.Additive System.Environment (getArgs) -- Some types and helper functions for dealing with -- "vectors" (implemented as arrays of rational numbers). type Coordinate = Ratio Int type Vector = V.Vector Coordinate 12 -- Inner product. inp :: Vector -> Vector -> Coordinate inp a b = V.sum (V.zipWith () a b) half :: Coordinate half = 1 % 2 -- Product of scalar with vector. l :: Coordinate -> Vector -> Vector l = V.map . () instance Algebra.Additive.C Vector where (+) = V.zipWith (+) (-) = V.zipWith (-) negate = l (-1) zero = V.fromList [0,0,0,0,0,0,0,0] -- Some data regarding E_8 delta :: (Eq a,Algebra.Ring.C b) => a -> a -> b delta i j = if i == j then 1 else 0 -- 'e i' gives the i'th standard basis vector of R_8. e :: Int -> Vector e i = V.fromList $ map (delta i) [1 .. 8] -- This is the usual integral basis of the lattice E_8. basis :: [Vector] basis = [ l half $ (e 1 + e 8) - (sum $ map e [2 .. 7]) , e 1 + e 2 ] ++ map (\ i -> e (i - 1) - e (i - 2)) [3 .. 8] roots :: [Vector] roots = d8 ++ x118 where d8 = concatMap ((\ [a,b] -> [a + b,a - b,b - a,negate a - b]) . map e . getSet) $ enumerate (ksubsets 2) [1 .. 8] x118 = map (\ (Prod (Set neg) (Set pos)) -> l half $ sum (map (negate . e) neg) + sum (map e pos)) $ enumerate ((set 'ofSize' even) set) [1 .. 8] 13 -- 'posRoots' contains exactly one of every pair -- (a,-a) of roots. posRoots :: [Vector] posRoots = nubBy (\ a b -> a == b a == negate b) roots -- Generate elements l of the E_8 lattice with the property -- that l^2 = 2 d. We need only one element of each orbit -- under the action of the Weyl group. In particular, we -- may assume that all coordinates but one (say, the first) -- are nonnegative, and that the successive coordinates are -- nondecreasing. We generate exactly one element of each -- H-orbit, where H is the subgroup of permutations and even -- sign changes. gen :: Int -> [Vector] gen d = genInt d ++ genHalfInt d genInt :: Int -> [Vector] genInt d = map (V.fromList . map fromIntegral) $ go [] 0 where -- Given the length of a partial vector, compute the maximal -- new coordinate which does not increase the length of the -- vector beyond 2 d. maxCoord :: Int -> Int maxCoord s = floor (sqrt (fromIntegral $ dD - s) :: Double) dD :: Int dD = 2 * d -- We maintain a list of coordinates chosen so far, every -- one together with the sum of squares of the coordinates -- up to and including that coordinate. -- The generated vectors are elements of E_8, because the -- sum of the squares of their components is even, hence -- the sum of the components as well. go :: [(Int,Int)] -> Int -> [[Int]] -- We have fixed all eight coordinates. go fixed@((_,sq) : ps) 8 -- The vector has the right length; add the relevant -- solutions (using 'vary'), and continue searching. sq == dD = vary (map fst fixed) ++ lower ps 7 -- The vector has the wrong length, continue searching. otherwise = lower ps 7 go fixed n = let (m,s) = case fixed of 14 [] (c,s) : _ -> (Prelude.min (maxCoord s) c,s) -> (maxCoord 0,0) in go ((m,s + m ^ 2) : fixed) (n + 1) -- Lexicographically decrease the given vector, and continue -- the generation from there. lower :: [(Int,Int)] -> Int -> [[Int]] lower [] lower ((x,s) : ps) n _ = [] x == 0 otherwise = go ((x - 1,s + 1 - 2 * x) : ps) n = lower ps (n - 1) vary :: [Int] -> [[Int]] vary (x : xs) = if x == 0 then [0 : xs] else [x : xs,negate x : xs] -- For vectors with all coordinates half-integers, we work -- with the doubles of the coordinates. genHalfInt :: Int -> [Vector] genHalfInt d = map (V.fromList . map (% 2)) $ go [] 0 where maxCoord :: Int -> Int maxCoord = Memo.integral m where m s = f $ floor (sqrt (fromIntegral $ dE - s) :: Double) f k = if odd k then k else k - 1 dE :: Int dE = 8 * d go :: [(Int,Int)] -> Int -> [[Int]] go fixed@((_,sq) : ps) 8 sq == dE = filter e8 (vary $ map fst fixed) ++ lower ps 7 otherwise = lower ps 7 go fixed n = let (m,s) = case fixed of [] (c,s) : _ -> (Prelude.min (maxCoord s) c,s) -> (maxCoord 0,0) in go ((m,s + m ^ 2) : fixed) (n + 1) -- Decides whether a given vector is an element of E_8 e8 :: [Int] -> Bool 15 e8 = (== 0) . flip rem 4 . sum lower :: [(Int,Int)] -> Int -> [[Int]] lower [] lower ((x,s) : ps) n _ = [] x == 1 otherwise = go ((x - 2,s + 4 - 4 * x) : ps) n = lower ps (n - 1) vary :: [Int] -> [[Int]] vary (x : xs) = [x : xs,negate x : xs] References [Bor] R.E. Borcherds, Automorphic forms on Os+2,2(R) and infinite prod- ucts. Invent. Math. 120 (1995), 161 -- 213. [BKPS] R.E. Borcherds, L. Katzarkov, T. Pantev, N.I. Shepherd-Barron, Families of K3 surfaces. J. Algebraic Geom. 7 (1998), 183 -- 193. [Bou] N. Bourbaki, Groupes et alg`ebres de Lie. Chapitres IV `a VI. ´El´ements de math´ematique. Fasc. XXXIV. Actualit´es Scientifiques et Industrielles, No. 1337 Hermann, Paris, 1968. [CS] J.H. Conway, N.J.A. Sloane, Sphere packings, lattices and groups. Grundlehren der mathematischen Wissenschaften 290. Springer- Verlag, New York, 1988. [GHS1] V. Gritsenko, K. Hulek & G.K. Sankaran, The Kodaira dimension of the moduli of K3 surfaces. Invent. Math. 169 (2007), 519 -- 567. [GHS2] V. Gritsenko, K. Hulek & G.K. Sankaran, Moduli spaces of irre- ducible symplectic manifolds. Compos. Math. 146 (2010), 404 -- 434. [GHS3] V. Gritsenko, K. Hulek & G.K. Sankaran, Abelianisation of or- thogonal groups and the fundamental group of modular varieties. J. Algebra 322 (2009), 463 -- 478. [Fr] E. Freitag, Siegelsche Modulfunktionen. Grundlehren der mathema- tischen Wissenschaften 254. Springer-Verlag, Berlin, 1983. A. Peterson, Korteweg de Vries Instituut voor Wiskunde, Universiteit van Amsterdam, P.O. Box 9424, 1090 GE Amsterdam, The Netherlands [email protected] G.K. Sankaran, Department of Mathematical Sciences, University of Bath, Bath BA2 7AY, England [email protected] 16
1512.09296	4	1512	2017-06-28T05:55:05	Torsion points on theta divisors	[ "math.AG" ]	Using the irreducibility of a natural irreducible representation of the theta group of an ample line bundle on an abelian variety, we derive a bound for the number of $n$-torsion points that lie on a given theta divisor. We present also two alternate approaches to attacking the case $n=2$.	math.AG	math	TORSION POINTS ON THETA DIVISORS ROBERT AUFFARTH, GIAN PIETRO PIROLA AND RICCARDO SALVATI MANNI Abstract. Using the irreducibility of a natural irreducible representa- tion of the theta group of an ample line bundle on an abelian variety, we derive a bound for the number of n-torsion points that lie on a given theta divisor. We present also two alternate approaches to attacking the case n = 2. 1. Introduction Let A be a complex abelian variety of dimension g and let Θ be an ample divisor on A that gives a principal polarization L := OA(Θ) (i.e. dim H 0(A, L) = 1). We will use the notations (A, Θ) and (A, L) inter- changeably. For n ≥ 2, define Θ(n) := #A[n] ∩ Θ, where A[n] is the group of n-torsion points on A. It is well-known that Θ does not contain all n-torsion points; this follows easily, for example, from the irreducibility of the representation of the theta group of Ln in H 0(A, Ln) as we will discuss below. It is a classical result, [12] that the evaluation at the n-torsion points, n ≥ 4, of Riemann's theta function completely determines the abelian variety embedded in Png −1. The image is the intersection of all the quadrics containing the image of the n-torsion points. Moreover the structure of A[2]∩Θ tells us if the principally polarized abelian variety (A, Θ) is decomposable, [14] or is the Jacobian of an hyperelliptic curve, [11]. Also recently, in [2] it has been proved that (A, Θ) is decomposable if and only if the image of the Gauss map at the smooth points of Θ in A[2] ∩ Θ is contained in a quadric of Pg−1. In [13], a bound is obtained for the number of 2-torsion points on a theta divisor. Indeed, they show that Θ(2) ≤ 4g − 2g. However, this bound is far from optimal, and in the same paper it is conjectured that the actual bound is 4g − 3g and is achieved if and only if (A, L) is the polarized product of elliptic curves. One could generalize this and conjecture that for n-torsion points the bound should be n2g −(n2 −1)g, with equality if and only if (A, L) is the polarized product of elliptic curves. Let τ ∈ Hg be a matrix in the Siegel upper-half space, and for δ, ǫ ∈ Rg and z ∈ Cg define the theta function with characteristics θ(cid:20)δ ǫ(cid:21) (τ, z) := Xm∈Zg exp[πi(m + δ)tτ (m + δ) + 2πi(m + δ)t(z + ǫ)]. 2010 Mathematics Subject Classification. 14K25; 32G20. Key words and phrases. abelian variety, theta divisor, torsion. Partially supported by Fondecyt Grant 3150171, Prin 2012 "Moduli Spaces and Lie Theory" and Inadm Gnsaga. 1 2 ROBERT AUFFARTH, GIAN PIETRO PIROLA AND RICCARDO SALVATI MANNI When δ = ǫ = 0 we obtain Riemann's theta function θ(τ, z) := θ(cid:20)0 the projection of {θ(τ, ·) = 0} to Aτ := Cg/τ Zg + Zg gives a symmetric theta divisor (i.e. −Θ = Θ) that we will denote by Θτ . We remark that any complex principally polarized abelian variety is isomorphic to (Aτ , Θτ ) for some τ ∈ Hg. If we put Lτ := OAτ (Θτ ), it is well-known that the set 0(cid:21) (τ, z); (cid:26)θ(cid:20)δ 0(cid:21) (nτ, nz) : δ ∈ is a basis for H 0(Aτ , Ln τ ) and the set (cid:26)θ(cid:20)δ ǫ(cid:21) (τ, nz) : δ, ǫ ∈ 1 n Zg/Zg(cid:27) 1 n Zg/Zg(cid:27) is a basis for H 0(Aτ , Ln2 τ ). . A simple calculation shows that θ(τ, z + τ δ + ǫ) = λ(z)θ(cid:20)δ ǫ(cid:21) (τ, z) for some nowhere vanishing function λ, and it immediately follows that if Θ = Θτ , then Θ(n) is exactly the number of vanishing theta constants Zg/Zg. A similar statement holds if Θ is the pull- θ(cid:20)δ ǫ(cid:21) (τ, 0) for δ, ǫ ∈ 1 n back of Θτ by a translation (i.e. Θ any theta divisor). If n = 2 and 4δtǫ ≡ 1 (mod 2), then the associated theta constant vanishes, and so Θτ (2) ≥ 2g−1(2g − 1). In fact, this is an equality if Aτ ∈ Ag\θnull, where θnull is the divisor consisting of principally polarized abelian varieties such that one of its symmetric theta divisors has a singularity at a point of order 2. The goal of this paper is to give a stronger bound for Θ(n). Our main theorem gives the following: Theorem 1.1. Let (A, Θ) be a complex principally polarized abelian variety. Then Θ(2) ≤ 4g − g2g−1 − 2g and for n ≥ 3 Θ(n) ≤ n2g − (g + 1)ng. We can make this bound better if (A, Θ) is decomposable. After proving this theorem, we present alternative approaches to attacking the number Θ(2). One of these points of view will give a better bound than the theorem, in fact we get Proposition 1.2. Let (A, Θ) be a principally polarized abelian variety. Then Θ(2) ≤ 4g − 7g − 1 3g − 1 We observe that the methodologies involved are interesting and different from the original approach, and we believe they will be more useful in the future. In particular in the last approach that could produce the conjectural bound, a matrix M appears, induced by the Weil pairing between the points TORSION POINTS ON THETA DIVISORS 3 of order two in the abelian variety. This matrix appears also in other fields of mathematics, in coding theory as the matrix associated to the Macwilliams identity for the weight enumerator of the codes, cf.[1] page 103, and in the theory of Borcherds' additive lifting as the matrix associated to a unitary representation of the integral metaplectic group on C[F2g 2 ], cf.[4] Acknowledgements: We would like to thank Sam Grushevsky for reading a preliminary version of the paper and pointing out a counterxample to the original proof of the main theorem. We would also like to thank Corrado de Concini for some helpful discussions. 2. A bound for Θ(n) Since L is a principal polarization, we have that xLn ≃ Ln}, A[n] = {x ∈ A : t∗ where tx : A → A denotes translation by x. Recall that in this case, the theta group of Ln is a certain central extension of A[n] by Gm which we will denote by Gn: 1 → Gm → Gn → A[n] → 0. Let ϕn : A → PH 0(A, L) be the morphism associated to the linear system Ln. The vector space H 0(A, Ln) is an irreducible representation for the theta group Gn where Gm acts by scalar multiplication (see [9, Ch. 4] or [10, Theorem 2, pg. 297]), and we therefore obtain a projective representation ρ : A[n] → PGL(H 0(A, Ln)). Because of the irreducibility of the representation, we notice that there is no proper linear subspace of PH 0(A, Ln) that is invariant under the action of A[n]. Moreover, we have that ρ(x) · ϕn(y) = ϕn(x + y) for every x ∈ A[n] and y ∈ A. Let H ⊆ A[n] be a maximal isotropic subgroup with respect to the com- mutator pairing associated to the theta group of Ln. We say that H is c-isotropic if it has a complementary isotropic subgroup K. We remark that any maximal isotropic subgroup is isomorphic to (Z/nZ)g, as is its complementary isotropic subgroup if it exists. Let H be c-isotropic, let p : A → A/H =: AH be the natural projection, and let q : AH → A be the inverse isogeny. We have a commutative diagram A p AH ❇ ❇ ❇ ❇ nA ❇ ❇ ❇ ❇ q A where nA denotes multiplication by n on A. By descent theory for abelian varieties, we have that there exists a principal polarization M on AH such that Ln ≃ p∗M and q∗L ≡ Mn. We see in this case that ker q is a maximal c-isotropic subgroup of AH [n]. Let N be a complementary isotropic subspace of ker q. / / 4 ROBERT AUFFARTH, GIAN PIETRO PIROLA AND RICCARDO SALVATI MANNI Define Σ = q−1(Θ) ∈ q∗L and for a ∈ AH , define Σa := Σ + a. For every b ∈ AH[n], fix a section sb ∈ H 0(AH, q∗L) such that Σb = div(sb). Lemma 2.1. The set {sb : b ∈ N } is a basis for H 0(AH , q∗L). Proof. We see that for all a ∈ ker q and b ∈ N , Σa+b = Σb + a = q−1(Θ + q(b)) = Σb. asb = λasb. This means that for all a ∈ ker q and b ∈ N , there exists λa ∈ Gm such that t∗ In other words, AH[n] acts on the projective span of {sb : b ∈ N } in PH 0(AH , q∗L). Since the theta group representation is irreducible, we must have that the above set generates the whole space. Moreover, #N = dim H 0(AH, q∗L), and the result follows. (cid:3) Let ϕH : AH → PH 0(AH, q∗L) be the morphism associated with q∗L. Definition 2.2. For H a maximal c-isotropic subgroup of A[n], let c1 + H, . . . , cng + H be its cosets (we will assume c1 = 0). We define the integers QH,ci := dimC span{ϕH (q−1(ci))} ng QH := QH,ci Xi=1 Q(n) := max{QH : H ⊆ A[n] max. c-isotropic subgroup}. We can use these numbers to obtain a bound on the number of n-torsion points lying on Θ. Proposition 2.3. Let (A, Θ) be a principally polarized abelian variety and let n ≥ 2. Then Θ(n) ≤ n2g − ng − Q(n). Proof. We will prove that Θ(n) ≤ n2g − ng − QH for every maximal c- isotropic subgroup H ⊆ A[n]. Let S ⊆ H + ci be a subset with r ≤ QH,ci elements. We will first prove that Θ does not contain (H + ci)\S. We see that (H + ci)\S ⊆ Θ ⇔ q−1((H + ci)\S) ⊆ Σ ⇔ (AH[n] + di)\(ker q + t1 ⊔ · · · ⊔ ker q + tr) ⊆ Σ where q(di) = ci and the tj are chosen so that S = {q(tj) : j = 1, . . . , r}. Assume this occurs. Now for all b ∈ N , (AH [n] + di)\(ker q + t1 + b ⊔ · · · ⊔ ker q + tr + b) ⊆ Σb. It follows that q−1(ci) = ker q +di ⊆ Σb for every b /∈ (ker q +di −tj)∩N . We see then there are ng − r options for b. Using Lemma 2.1, this implies that ϕH (q−1(ci)) is contained in a linear subspace of PH 0(AH , q∗L) of dimension r − 1, a contradiction with the choice of r. Therefore in each coset ci + H, there are at most ng − QH,ci − 1 points that lie on Θ. By adding everything up we get the bound we were looking for. (cid:3) Remark 2.4. The proof of the previous proposition is valid over any alge- braically closed field of characteristic prime to n and for any theta divisor (i.e. not necessarily symmetric). Moreover, the proposition already gives us a better bound than the one in [13]. Indeed, there can be at most one QH,ci TORSION POINTS ON THETA DIVISORS 5 equal to 0 (this happens when (AH , M) is the polarized product of elliptic curves), and so Θ(2) ≤ 4g − 2g − (2g − 1) = 4g − 2g+1 + 1. The next proposition shows that when looking for a bound for Θ(n), we can always assume that Θ is given by the zero set of Riemann's theta function. Proposition 2.5. If Θ is Riemann's theta function, then equality holds in Proposition 2.3. Proof. Assume that Θ is Riemann's theta function, and so Θ = Θτ on Aτ for some τ ∈ Hg. Let Λτ be the lattice τ Zg + Zg and take the maximal c-isotropic subgroup H = {τ ǫ : ǫ ∈ 1 Zg/Zg} + Λτ of Aτ [n]. We have the n quotient maps Aτ p → AH = Aτ /n q → Aτ where p(z + Λτ ) = z + Λτ /n and q(z + Λτ /n) = nz + Λτ . We see that the cosets of H are precisely µ + H for µ ∈ 1 n Zg/Zg, and moreover q−1(µ + Λτ ) = 1 n µ + 1 n Zg + Λτ /n. Then But θ(cid:20)δ 0(cid:21) (τ, µ + a)(cid:21)δ∈ 1 ϕH(q−1(µ + Λτ )) =((cid:20)θ(cid:20)δ 0(cid:21) (τ, µ + a) = exp(2πiδta)θ(cid:20)δ QH,µ + 1 = rank(cid:18)exp(2πinδtǫ)θ(cid:20)δ µ(cid:21) (τ, 0). Therefore, µ(cid:21) (τ, 0)(cid:19)δ,ǫ∈ 1 n : a ∈ Zg/nZg) . Zg/Zg , Zg /Zg n and so we have n2g − ng − QH = n2g − Xµ∈ 1 n Zg/Zg rank(cid:18)exp(2πinδtǫ)θ(cid:20)δ µ(cid:21) (τ, 0)(cid:19)δ,ǫ∈ 1 n . Zg/Zg A quick check shows that the sum above is equal to the number of non- vanishing theta constants, which we know is equal to n2g − Θ(n). (cid:3) We can now obtain an explicit bound for the number of torsion points on a theta divisor. Theorem 2.6. Let (A, Θ) be a complex principally polarized abelian variety. Then and for n ≥ 3 Θ(2) ≤ 4g − g2g−1 − 2g Θ(n) ≤ n2g − (g + 1)ng. Proof. By the previous proposition, we can take Θ = Θτ on Aτ for some τ ∈ Hg. Using the notation as in the proof of the previous proposition, we have that q−1(µ + Λτ ) = µ + 1 n 1 n Zg + Λτ /n 6 ROBERT AUFFARTH, GIAN PIETRO PIROLA AND RICCARDO SALVATI MANNI for µ ∈ 1 Zg/Zg. Therefore each member of ϕH (q−1(µ+Λτ )) differs from the n other by the action of the representation ρH : ker q → P GL(H 0(AH, q∗L)). It is known that this action (for this particular subgroup) multiplies the projective coordinates of PH 0(AH , q∗L) by nth roots of unity, and so we can estimate QH,µ by the number of vanishing coordinates. Using this fact, it is easy to see that QH,ci is equal to ng − 1 − r, where r is the number of vanishing coordinates. Therefore nQH,µ ≥ #ϕH(q−1(µ)). For n = 2, when µ = 0 we have 2g different points in ϕH (ker q), and so QH,0 ≥ g. Let us assume Θτ /2 irreducible. When µ 6= 0, we have 2g−1 different points, and so QH,µ ≥ g − 1. Adding everything up we get Θ(2) ≤ 4g − 2g − g − (g − 1)(2g − 1) = 4g − g2g − 1. In the case Θτ /2 reducible we proceed in the same way, but now we have less points since the map is not injective on the Kummer variety. The worst case will be when (X, Θ) is a product of elliptic curves. In this case depending on µ we can get in the image 2k different points, k = 0, . . . g − 1. Varying µ this happens exactly (cid:0)g k(cid:1) times. Hence totally we get g Θ(2) ≤ 4g − Xk=0(cid:18)g k(cid:19)(k + 1) = 4g − g2g−1 − 2g. For n ≥ 3, we have that ϕH is an embedding, and so there are always ng points in ϕH (q−1(ci)). This means that QH,ci ≥ g. Therefore if n ≥ 3, Θ(n) ≤ n2g − ng − gng = n2g − (g + 1)ng. (cid:3) When Θ is reducible, even more can be said: Corollary 2.7. If (A, Θ) ≃Qs Θ(2) ≤ 4g − 2g i=1(Bi, Θi) and bi = dim Bi, then and for n ≥ 3 + 1(cid:19) s 2 Yi=1(cid:18) bi Yi=1 s Θ(n) ≤ n2g − ng (bi + 1). Proof. In this case, we see that the number of n-torsion points on Θ is equal to n2g − t where t is the number of n-torsion points of the form (x1, . . . , xs) such that xi /∈ Θi for all i. Therefore s s Θ(2) = 4g − (4bi − Θi(2)) ≤ 4g − Yi=1 The same technique can be applied for n ≥ 3. (bi2bi−1 + 2bi). Yi=1 (cid:3) Remark 2.8. If (X, Θ) is simple (or more generally not 2-isogenous to a product), using the action of the symplectic group we can improve the estimate for QH,0; in fact we can get QH,0 ≥ 2g − 1. Thus in this case we get Θ(2) ≤ 22g − 2g − g2g = 22g − (g + 1)2g. TORSION POINTS ON THETA DIVISORS 7 This number fits in the general picture. 3. Alternative approaches for n = 2 3.1. Alternative approach 1. The methodology in this section is different from that in the previous one, and there are changes in notation. Assume that Θ is symmetric and irreducible, and define Bn := H 0(A, OA(nΘ)). n be the eigenspace associated to 1 for the automorphism (−1)∗. It is Let B+ well-known that dimC B± n = 2g−1(mg ± 1). We will use a few results from [8]. For n ≥ 2 and m ≥ 3, the natural map Bn ⊗ Bm → Bm+n is surjective. Since B2 = B+ and therefore 2 , we have that B2 ⊗ B± m → B± m+2 is surjective, Sym2(B2) ⊗ B± m → B± m+4 is surjective. Let V2 ⊆ B+ 4 be the image of Sym2(B2) in B+ 4 . We are Zg/Zg ǫ(cid:21) (τ, 2z) for δ, ǫ ∈ 1 2 interested in a basis of V2, which is given by all θ(cid:20)δ and 4δtǫ ≡ 0 (mod 2) such that θ(cid:20)δ ǫ(cid:21) (τ, 0) 6= 0 (in this section all theta characteristics will be half-integer characteristics). Let ng be the dimension of V2. It is clear that Θ(2) = 4g − ng, since it is the number of vanishing theta constants. As an immediate con- sequence of the previous discussion we have Proposition 3.1. Θ(2) ≤ 4g − 7g − 1 3g − 1 Proof. We have that the map Sym2(B2) ⊗ B± m → B± Sym2(B2) ⊗ B± m m+4 factors as / B± m+4 'PPPPPPPPPPPP :ttttttttt V2 ⊗ B± m and since the above arrow is surjective, all the arrows are surjective. There- fore, ng ≥ dimC B± m+4/ dimC B± m = (m + 4)g ± 1 mg ± 1 for m ≥ 3. The maximum of this function in m is achieved when m = 3 and the sign is negative. (cid:3) ' / : 8 ROBERT AUFFARTH, GIAN PIETRO PIROLA AND RICCARDO SALVATI MANNI 3.2. Alternative approach 2. From the addition formula for theta func- tions with semi-integral characteristics (see [7, Theorem 2, pg. 139] we have θ(cid:20)δ ǫ(cid:21) (τ, 0)θ(cid:20)δ ǫ(cid:21) (τ, 2z) =Xσ (−1)<2ǫ,2σ>θ(cid:20)σ 0(cid:21) (2τ, 2z)θ(cid:20)δ + σ 0 (cid:21) (2τ, 2z). Moreover we can restate this saying that Θ(2) − 2g−1(2g − 1) = 2g−1(2g + 1) − ng is the dimension of the space of quadrics that vanish on the image of the Kummer variety K(A) = A/ ± 1, via the embedding K(A) ֒→ 2Θ ≃ P2g −1. Since the Kummer variety is irreducible and the map is finite, we have that the image of K(A) cannot be contained in any quadric of rank 2 in P2g −1. These quadrics form a variety of dimension 2g+1 − 1 in the space of all quadrics in P2g −1. Thus we have as a rough estimate: Lemma 3.2. ng ≥ 2g+1 − 1. Proof. The space of quadrics containing the image of K(A) does not intersect the above described variety. (cid:3) This then gives us the bound Θ(2) ≤ 4g − 2g+1 + 1. This estimate is very rough and a careful analysis could produce better re- sults. For example we know that if Θ is irreducible, the number of vanishing quadrics is equal to 1, 10 when g = 3, 4 respectively, and ≥ 66 when g = 5. All these are triangular numbers that could give the dimension of the space of quadrics of bounded rank. 3.3. Alternative approach 3. This method is different than the previous approach but gives us the same estimate. We have a short exact sequence 0 → R → V2 ⊗ B+ 4 → B+ where R is the space of relations. Let W2 ⊆ B+ 4 = V2 ⊕ W2; it has as a basis the set of theta functions with even characteristics that correspond to a point of order 2 on Θ. Recall that the Heisenberg group, given as a set 8 → 0 4 be such that B+ H = Gm × Fg 2 × Hom(F2, Gm)g, is a non-commutative group that is non-canonically isomorphic to the theta group of 2Θ. Now H acts on B+ 8 and decomposes these spaces with respect to its characters. Moreover, the characters are in one to one correspondence with the points of order 2 on A. It is known (see [6, Section 2.4]) that for a character χ 4 and B+ dim(B+ dim(B+ 2g−1 if χ is trivial if not 8 )χ = (cid:26) 2g 4 )χ = (cid:26) 1 if χ corresponds to an even characteristic 0 if not . TORSION POINTS ON THETA DIVISORS 9 Lemma 3.3. We have an exact sequence 0 → R0 →Mχ (V2)χ ⊗ (V2)χ → (B+ 8 )0 → 0, where the subscript 0 refers to the eigenspace corresponding to the trivial character. Proof. This follows from the surjectivity of Sym2(V2)⊕(W2 ⊗V2) → B+ 8 . (cid:3) Corollary 3.4. We have ng = 2g + dim R0; or in other words, Θ(2) = 4g − 2g − dim R0. In order to estimate Θ(2), we need a better grasp on what R0 or a suit- g the sets of isotropic (respectively g and K − able subspace is. Denote by K + anisotropic) elements in F2g 2 with respect to the quadratic form hX, Xi = x1xg+1 + · · · + xgx2g, and let k+ g and k− g be their orders. We introduce the matrix M (g) = M := exp"iπ g (ming+i − nimg+i)#!m,n∈Z2g /2Z2g Xi=1 . Now M has the decomposition M =(cid:18) M + N N t M − (cid:19) where M + (respectively M −) is the submatrix of M given by the restriction to K + g ). The following proposition is proven in [3, Lemma 1.1]: g (respectively K − g × K + g × K − Proposition 3.5. M has two eigenspaces of dimension k+ g with eigenvalues ±2g, while M ± has eigenspaces of dimension (1/3)(2g ±1)(2g−1± 1) and (1/3)(22g − 1) with eigenvalues ±2g and ∓2g−1. For X ∈ Ck+ g and Y ∈ Ck− g and k− g , we have M(cid:18)X Y(cid:19) = 2g(cid:18)X M(cid:18)X Y(cid:19) = −2g(cid:18)X Y(cid:19) ⇐⇒ M −Y = 2g−1Y = N tX Y(cid:19) ⇐⇒ M +X = −2g−1X = N Y M +X = 2gX ⇐⇒ N tX = 0 M −Y = −2gY ⇐⇒ N Y = 0 M +X = −2g−1X M −Y = 2g−1Y if M +X − N Y = 0 N tX − M −Y = 0 if If m = (a, b) ∈ K + g for a and b considered as elements of {0, 1}g , then we use the notation θm(τ, z) := θ(cid:20)a/2 b/2(cid:21) (τ, z). The following lemma is also proved in [3]: 10 ROBERT AUFFARTH, GIAN PIETRO PIROLA AND RICCARDO SALVATI MANNI Lemma 3.6. If X = (vm)m∈K + Moreover we have g is a column of N , then M +X = −2g−1X. vmθm(τ, 0)2θm(τ, 2z)2 = 0 Xm∈K + g where (vm)m∈K + g is a column of N . Since we have B := N N t = 2g−1(2gI − M +), 3 (4g − 1). Thus the columns of N span the it is easy to deduce that rk(N ) = 1 whole eigenspace of M + with eigenvalue −2g−1. If S0 ⊂ R0 is the subspace spanned by these relations, then we have dim S0 ≤ (4g − 1). 1 3 Obviously the dimension of S0 is 1 3 (4g − 1) if there are no theta constants vanishing. If there are theta constants that vanish then the dimension could drop. Let J be the k+ g diagonal matrix whose entries are 0 or 1 depending on whether or not the theta constant θm(τ, 0) corresponding to m ∈ K + g vanishes. We see that g × k+ dim S0 = rk(JN ) = rk(JN (JN )t) = rk(JBJ t) where B = N N t. Now deleting the 0 rows and columns, JBJ t corresponds to a certain principal submatrix Br of B of size r × r where r = ng ≥ 2g + dim S0 = 2g + rk(JBJ t). Thus to have an estimate for ng, we need to estimate the ranks of principal submatrices of B. We therefore obtain: Proposition 3.7. Θ(2) ≤ 4g − 2g − h0 where h0 = min{k ≥ 2g + rank(S) : S principal submatrix of B of order k }. Corollary 3.8. Θ(2) ≤ 4g − 2g+1 + 1 Proof. We will show that all principal minors of B of size s ≤ 2g − 1 are positive definite. The matrix B2g −1 is semi-positive definite. The entries are equal to 2g − 1 along the diagonal and ±1 out of the diagonal. For every X ∈ R2g −1 we have X tB2g −1X = X1≤i<j≤k Thus it is positive definite. (xi ± xj)2 + 2g −1 Xi=1 x2 i . Now Sp(2g, F2) acts on the set of characteristics by (cid:3) (cid:18) a b c d (cid:19) ·(cid:20)δ −b This action is double transitive on the set of even (respectively odd) char- acteristics. Therefore if we want to compute the rank of submatrices of the ǫ(cid:21) :=(cid:18) d −c a (cid:19)(cid:18)δ ǫ(cid:19) +(cid:18)diag(ctd) diag(atb)(cid:19) . 0, . . . , g. If we look at the submatrix Bk indexed by all even characteristics order 3g with eigenvalues (−1)k2g−k that have multiplicity (cid:0)g m =(cid:20)δ ǫ(cid:21) satisfying 4δtǫ = 0 in Z, then k(cid:1)2g−k for k = TORSION POINTS ON THETA DIVISORS 11 matrix B, we can consider only orbits with respect to the action of this group. The Kronecker product of g times the matrix M +(1) is a matrix L(g) of Bk = 2g−1(2gI3g − L(g)) and has rank 3g − 2g. We see that this implies the well-known result that if (A, Θ) is the product of elliptic curves, then there are 3g points of order two that are not on Θ. We finish our analysis by looking at the genus 2 case. Double transitivity of the action of the symplectic group implies that all submatrices of degree 8 of M +(2) are conjugate via the action of the symplectic group. For one of these matrices, we can prove that the rank is 5. This implies that n2 ≥ 9, which is sharp. We therefore conjecture the following that would imply that Θ(2) ≤ 4g − 3g for all g: Conjecture 3.9. The number h0 is reached at L(g). References [1] J. H. Conway, N.J.A. Conway. Sphere Packings, Lattices and Groups. Die Grundlehren der mathematischen Wissenschaften in Einzeldarstellungen Band 290. Springer-Verlag. First edition. 1988. [2] F. Dalla Piazza, A. Fiorentino, S. Grushevsky, S. Perna, R. Salvati Manni Vector- valued modular forms and the Gauss map. arXiv:1505.06370 [3] J. Fay. On the Riemann-Jacobi formula. Nachr. Akad. Wiss. Gottingen Math.- Phys. Kl. II. 1979, no. 5, 61-73. [4] R. Salvati Manni, E. Freitag: Modular forms for the even unimodular lattice of signature (2,10). Journal of Algebraic Geometry, vol. 16 ( 2007): 753-791, . [5] B. van Geemen. Schottky-Jung relations and vectorbundles on hyperelliptic curves. Math. Ann. 281 (1988), no. 3, 431449. [6] B. van Geemen. Some arxiv.org/pdf/1307.2463v2 equations for the universal Kummer variety. [7] J.I. Igusa. Theta Functions. Die Grundlehren der mathematischen Wis- senschaften in Einzeldarstellungen Band 194. Springer-Verlag. First edition. 1972. [8] G. Kempf. Equations of Kummer varieties. American Journal of Mathematics. Vol. 114, No. 1 (Feb., 1992), pp. 229-232. [9] G. Kempf. Complex Abelian Varieties and Theta Functions. Universitext, Springer. 1991. [10] D. Mumford. On the equations defining abelian varieties I. Invent. Math., 1:287- 354, 1966. [11] D. Mumford. Tata Lectures on Theta 2", Progr. Math. , 43, Birkhauser, Boston- Basel-Stuttgart (1984) [12] D. Mumford. Tata Lectures on Theta 3", Progr. Math. , 97, Birkhauser, Boston- Basel-Stuttgart (1991) [13] V. Marcucci, G. Pirola. Points of order 2 on theta divisors. Atti Accad. Naz. Lincei Cl. Sci. Fis. Mat. Natur. Rend. Lincei (9) Mat. Appl. 23 (2012) 319-323. [14] R. Salvati Manni: Modular varieties with level 2 theta structure. Amer. J. Math. 116 (1994), no. 6, 1489–1511. 12 ROBERT AUFFARTH, GIAN PIETRO PIROLA AND RICCARDO SALVATI MANNI R. Auffarth, Departamento de Matem´aticas, Facultad de Ciencias, Univer- sidad de Chile, Santiago, Chile E-mail address: [email protected] G. Pirola, Dipartimento di Matematica, Universit`a di Pavia, Italy E-mail address: [email protected] R. Salvati Manni, Dipartimento di Matematica "Guido Castelnuovo", Uni- versit`a di Roma "La Sapienza", Italy E-mail address: [email protected] 7 1 0 2 n u J 8 2 ] . G A h t a m [ 4 v 6 9 2 9 0 . 2 1 5 1 : v i X r a APPENDIX TO "TORSION POINTS ON THETA DIVISORS" ROBERT AUFFARTH, GIUSEPPE PARESCHI, GIAN PIETRO PIROLA AND RICCARDO SALVATI MANNI Abstract. In this note we revisit a method used in [1] to give a sharp bound for the number of 2-torsion points on a theta divisor. 1. Introduction Let A be a complex abelian variety and let L = OA(Θ) be a principal polarization on A. We set Θ(n) := #A[n] ∩ Θ, where A[n] is the group of 2-torsion points on A. In [5], the authors gave a bound for the number of 2-torsion points on a theta divisor. This bound has been recently improved in [1] where also a bound for the n torsion points is given. However, these bounds were not optimal . In [5], it has been conjectured that the bound for the two torsion points is 4g − 3g and is achieved if and only if (A, L) is the polarized product of elliptic curves. Similarly in [1] this conjecture has been generalized to the case of n-torsion points. In these cases the bound is n2g − (n2 − 1)g. The aim of this note is to prove the first part of the first conjecture and we will give an estimate for Θ(n) when n is even , more exacly we will prove the following: Theorem 1.1. Let (A, Θ) be a principally polarized abelian variety. Then Θ(2) ≤ 4g − 3g Θ(2m) ≤ m2g(4g − 3g) The proofs are consequence of results proved in [3] . 2. The proof We recall shortly some basic facts. Let τ ∈ Hg be a matrix in the Siegel upper-half space, and for δ, ǫ ∈ Rg and z ∈ Cg define the theta function with characteristics θ(cid:20)δ ǫ(cid:21) (τ, z) := Xm∈Zg exp[πi(m + δ)tτ (m + δ) + 2πi(m + δ)t(z + ǫ)]. 2010 Mathematics Subject Classification. 14K25; 32G20. Key words and phrases. abelian variety, theta divisor, torsion. Partially supported by Fondecyt Grant 3150171, progetto d'Ateneo ' "moduli, defor- mazioni e superfici K3 y" and Inadm Gnsaga. 1 2 R. AUFFARTH, G. PARESCHI, G. P. PIROLA, R. SALVATI MANNI When δ = ǫ = 0 we obtain Riemann's theta function θ(τ, z) := θ(cid:20)0 the projection of {θ(τ, ·) = 0} to Aτ := Cg/τ Zg + Zg gives the symmetric theta divisor Θτ . We set Lτ := OAτ (Θτ ). A basis for H 0(Aτ , Ln2 τ ) is given by 0(cid:21) (τ, z); (cid:26)θ(cid:20)δ ǫ(cid:21) (τ, nz) : δ, ǫ ∈ 1 n Zg/Zg(cid:27) . We shortly recall some basic facts about these functions. The evaluation at z = 0 of the above functions vanishes if and only if if the n torsion τ δ + ǫ belongs to Θτ . It is a well known fact that the dual torus A parametrizes isomorphism classes of line bundles in P ic0(A). The principal polarization Θτ induces an isomorphism A → A sendindg a point x to the line bundle Lx = t∗L ⊗ L−1 in A. We recall from [2] the following fundamental result (Theorem 6.8), which also appears in [6, Proposition 1.5] and [4, Proposition 7.2.2]. Theorem 2.1. Let M be an ample line bundle on A • For L in a non-empty subset of A the multiplication H 0(A, M⊗2 ⊗ L ⊗ M ) ⊗ H 0(A, M⊗2 ⊗ N ⊗ M −1) → H 0(A, M⊗4 ⊗ L ⊗ N ) is surjective for fixed M and N in A. • For L in a non-empty subset of A the multiplication H 0(A, M⊗2 ⊗ M ) ⊗ H 0(A, M⊗2 ⊗ N ⊗ L) → H 0(A, M⊗4 ⊗ L ⊗ M ⊗ N ) is surjective for fixed M and N in A. • If n ≥ 2 and m ≥ 3 then the multiplication H 0(A, M⊗m ⊗ M ) ⊗ H 0(A, M⊗n ⊗ N ) → H 0(A, M⊗n+m ⊗ M ⊗ N ) is surjective for arbitrary M and N in A. As an immediate consequence we get the following: Corollary 2.2. Let M be an ample line bundle on A. For L in a non-empty subset of A the multiplication H 0(A, M⊗2) ⊗ H 0(A, M⊗2) ⊗ H 0(A, M⊗2 ⊗ L) → H 0(A, M⊗6 ⊗ L) is surjective. Proof: It is enough in the second item of the previous theorem to take M = N = OA and then apply the third item. Relatively to the map H 0(A, M⊗2) ⊗ H 0(A, (M ⊗ L2y)⊗2) → H 0(A, (M ⊗ Ly)⊗4) we know from [3] that for M = Lτ the dimension of the image, say m(0, 2y) is equal to the number of η ∈ A[2] such that 2y + η /∈ Θ. This is equivalent to the number of non-vanishing (cid:26)θ(cid:20)δ ǫ(cid:21) (τ, 2y) : δ, ǫ ∈ 1 2 Zg/Zg(cid:27) . APPENDIX TO "TORSION POINTS ON THETA DIVISORS" 3 In particular we have that the map is surjective once 2y /∈ [η∈A[2] (Θ + η), i.e. no theta function with half integral characteristic vanishes at the point 2y. Moreover we have Proposition 2.3. For any y ∈ A, we have 3g ≤ m(0, 2y) ≤ 4g. Proof: The upper bound is obvious. The lower bound is an immediate consequence of Corollary 2.2. Now we can prove the theorem stated in the introduction. Proof of Th.1.1. Assuming y = 0 we get that Θ(2) = 4g − m(0, 0) ≤ 4g − 3g Now let n = 2m even, obviously A[2m]/A[2] ≡ (Z/mZ)2g. Now for each 2y in the above class the previous estimate holds. Taking the union on all representatives we get Θ(n) ≤ n2g − m2g3g = m2g(4g − 3g) and the theorem is proved. References [1] R. Auffarth, G. P. Pirola, R. Salvati Manni Torsion points on theta divisors Proc. Amer. Math. Soc. 145 (2017), 89-99; arXiv:1512.09296 [2] G. Kempf. Complex Abelian Varieties and Theta Functions. Universitext, Springer. 1991. [3] G. Kempf. Multiplication Over Abelian Varieties. American Journal of Mathe- matics. Vol. 110, No. 4 , pp. 765–773 [4] H. Lange, C. Birkenhake. Complex Abelian Varieties . Grund. Math Wiss.302 Springer-Verlag. 1992. [5] V. Marcucci, G. Pirola. Points of order 2 on theta divisors. Atti Accad. Naz. Lincei Cl. Sci. Fis. Mat. Natur. Rend. Lincei (9) Mat. Appl. 23 (2012) 319-323. [6] T. Sekiguchi. On the cubics defining abelian varieties. J.Math.Soc.Japan (1978) 703-721. R. Auffarth, Departamento de Matem´aticas, Facultad de Ciencias, Univer- sidad de Chile, Santiago, Chile E-mail address: [email protected] G. Pareschi, Dipartimento di Matematica, Universit`a di Tor Vergata, Italy E-mail address: [email protected] G. Pirola, Dipartimento di Matematica, Universit`a di Pavia, Italy E-mail address: [email protected] R. Salvati Manni, Dipartimento di Matematica "Guido Castelnuovo", Uni- versit`a di Roma "La Sapienza", Italy E-mail address: [email protected]
1101.2236	1	1101	2011-01-11T23:21:10	Relations in the tautological ring	[ "math.AG" ]	These notes cover our series of three lectures at Humboldt University in Berlin for the October 2010 conference "Intersection theory on moduli space" (organized by G. Farkas). The topic concerns relations among the kappa classes in the tautological ring of the moduli space of genus g curves. After a discussion of classical constructions in Wick form, we derive an explicit set of relations obtained from the virtual geometry of the moduli space of stable quotients. In a series of steps, the stable quotient relations are transformed to simpler and simpler forms. Our final result establishes a previously conjectural set of tautological relations proposed a decade ago by Faber-Zagier. The Faber-Zagier relations are defined using g and a single series in one variable with coefficients (6i)!/(3i)!(2i)!. Whether these relations span the complete set of relations among the kappa classes on the moduli space of genus g curves is an interesting question.	math.AG	math	Relations in the tautological ring R. Pandharipande and A. Pixton October 2010 The notes below cover our series of three lectures at Humboldt Uni- versity in Berlin for the October conference Intersection theory on mod- uli space (organized by G. Farkas). The topic concerns relations among the κ classes in the tautological ring of the moduli space of curves Mg. After a discussion of classical constructions ending in Theorem 1, we derive an explicit set of relations from the moduli space of stable quotients. In a series of steps, the stable quotient relations are trans- formed to simpler and simpler forms. The first step, Theorem 3, comes almost immediately from the virtual geometry of the moduli space of stable quotients. After a certain amount analysis, the simpler form of Proposition 3 is found. Our final result, Theorem 5, establishes a previously conjectural set of tautological relations proposed a decade ago by Faber-Zagier. A detailed presentation of the proof will appear in [7]. A. Chern vanishing relations Faber's original relations in Conjectural description of the tautolog- ical ring [1] are obtained from a very simple geometric construction. Let be the universal curve over the moduli space, and let π : C → Mg πd : Cd → Mg be the map associated to the dth fiber product of the universal curve. For every point [C, p1, . . . , pd] ∈ Cd, we have the restriction map (1) H 0(C, ωC) → H 0(C, ωCp1+...+pd) . 1 2 Since the canonical bundle ωC has degree 2g−2, the map (1) is injective if d > 2g − 2. Over the moduli space Cd, we obtain the exact sequence 0 → E → Ωd → Q → 0 where E is the rank g Hodge bundle, Ωd is the rank d > 2g − 2 bundle with fiber H 0(C, ωCp1+...+pd), and Q is the quotient bundle of rank d − g. Hence, ck(Q) = 0 ∈ Ak(Cd) for k > d − g . After cutting such vanishing ck(Q) down with cotangent line and di- agonal classes in Cd and pushing-forward via πd ∗ to Mg, we arrive at Faber's relations in R∗(Mg). From our point of view, at the center of Faber's relations in Conjec- tural description of the tautological ring [1] is the function Θ(t, x) = The differential equation ∞Xd=0 dYi=1 (1 + it) (−1)d d! xd td . t(x + 1) d dx Θ + (t + 1)Θ = 0 is easily found. Hence, we obtain the following result. Lemma 1. Θ = (1 + x)− t+1 t . We introduce a variable set z indexed by pairs of integers For monomials z = { zi,j i ≥ 1, j ≥ i − 1 } . zσ =Yi,j zσi,j i,j , we define ℓ(σ) =Xi,j Of course Aut(σ) =Qi,j σi,j! D =Xi,j iσi,j, . σ =Xi,j jσi,j . zi,j tj(cid:18)x d dx(cid:19)i . The variables z are used to define a differential operator After applying exp(D) to Θ, we obtain 3 ΘD = exp(D) Θ dYi=1 = Xσ ∞Xd=0 (1 + it) (−1)d d! xd td dℓ(σ)tσzσ Aut(σ) where σ runs over all monomials in the variables z. Define constants C r d(σ) by the formula log(ΘD) =Xσ ∞Xd=1 ∞Xr=−1 d(σ) tr xd C r d! zσ . By an elementary application of Wick, the t dependence of log(ΘD) has at most simple poles. Finally, we consider the following function, γ =Xi≥1 B2i 2i(2i − 1) κ2i−1t2i−1 +Xσ ∞Xd=1 ∞Xr=−1 d(σ) κrtr xd C r d! zσ . Denote the trxdzσ coefficient of exp(−γ) by (cid:2) exp(−γ)(cid:3)tr xd σ ∈ Q[κ−1, κ0, κ1, κ2, . . .] . z Our form of Faber's equations is the following result. Theorem 1. In Rr(Mg), the relation (cid:2) exp(−γ)(cid:3)tr xd σ = 0 z holds when r > −g + σ and d > 2g − 2. In the tautological ring R∗(Mg), the conventions κ−1 = 0, κ0 = 2g − 2 will always be followed. For fixed g and r, Theorem 1 provides infinitely many relations by increasing d. While the proof of Theorem 1 is appealingly simple, the relations do not seem to fit the other forms we will see later. The variables zi,j efficiently encode both the cotangent and diagonal operations studied in Conjectural description of the tautological ring [1]. In particular, the relations of Theorem 1 are equivalent to the mixing of all cotangent and diagonal operations studied there. 4 B. Stable quotient relations I. The function Φ. The relations in the tautological ring R∗(Mg) obtained from Moduli of stable quotients [4] are based on the function Φ(t, x) = ∞Xd=0 dYi=1 1 1 − it (−1)d d! xd td . Define the coefficients C r d by the logarithm, log(Φ) = ∞Xd=1 ∞Xr=−1 dtr xd C r d! . By an elementary application of Wick, the t dependence has at most a simple pole. Let γ =Xi≥1 B2i 2i(2i − 1) κ2i−1t2i−1 + ∞Xd=1 ∞Xr=−1 dκrtr xd C r d! . Denote the trxd coefficient of exp(−γ) by (cid:2) exp(−γ)(cid:3)tr xd ∈ Q[κ−1, κ0, κ1, κ2, . . .] . In fact, [exp(−γ)]tr xd is homogeneous of degree r in the κ classes. The first tautological relations of Moduli space of stable quotients [4] are given by the following result. Theorem 2. In Rr(Mg), the relation (cid:2) exp(−γ)(cid:3)tr xd = 0 holds when g − 2d − 1 < r and g ≡ r + 1 mod 2. For fixed r and d, if Theorem 2 applies in genus g, then Theorem 2 applies in genera h = g − 2δ for all natural numbers δ ∈ N. The genus shifting mod 2 property will also be present in the Faber-Zagier conjecture discussed later. 5 II. Partitions, differential operators, and logs. We will write partitions σ as (1a12a23a3 . . .) with ℓ(σ) =Xi ai and σ =Xi iai . The empty partition ∅ corresponding to (102030 . . .) is permitted. In all cases, we have Aut(σ) = a1!a2!a3!· · · . Consider the infinite set of variables p1, p2, p3, . . . . Monomials in the pi correspond to partitions pa1 1 pa2 2 pa3 3 . . . ↔ (1a12a23a3 . . .) . Given a partition σ, let pσ denote the corresponding monomial. Let Φp(t, x) =Xσ ∞Xd=0 dYi=1 1 1 − it (−1)d d! xd td dℓ(σ)tσpσ Aut(σ) where the first sum is over all partitions σ. The summand correspond- ing to the empty partition equals Φ(t, x). The function Φp is easily obtained from Φ, Let D denote the differential operator pitix d dx! Φ(t, x) . Φp(t, x) = exp ∞Xi=1 ∞Xi=1 D = pitix d dx . Expanding the exponential of D, we obtain (2) Φp = Φ + DΦ + D2Φ + 1 2 DΦ Φ + 1 2 1 6 D2Φ Φ D3Φ + . . . + 1 6 D3Φ Φ + . . .(cid:19) . = Φ(cid:18)1 + Let γ∗ = log(Φ) be the logarithm, Dγ∗ = DΦ Φ . 6 After applying the logarithm to (2), we see log(Φp) = γ∗ + log(cid:18)1 + Dγ∗ + 1 2 (D2γ∗ + (Dγ∗)2) + ...(cid:19) = γ∗ + Dγ∗ + D2γ∗ + . . . 1 2 where the dots stand for a universal expression in the Dkγ∗. In fact, a remarkable simplification occurs, log(Φp) = exp ∞Xi=1 pitix d dx! γ∗ . The result follows from a general identity. Proposition 1. If f is a function of x, then Proof. A simple computation for monomials in x shows d dx(cid:19) log(f ) . log(cid:18)exp(cid:18)λx d dx(cid:19) f(cid:19) = exp(cid:18)λx dx(cid:19) xk = (eλx)k . exp(cid:18)λx exp(cid:18)λx dx(cid:19) f (x) = f (eλx) . d d Hence, since the differential operator is additive, The Proposition follows immediately. (cid:3) The coefficients of the logarithm may be written as log(Φp) = = ∞Xd=1 ∞Xd=1 = Xσ d(p) tr xd C r d! ∞Xr=−1 ∞Xr=−1 d tr xd C r d! ∞Xr=−1 ∞Xd=1 d tr xd C r d! exp ∞Xi=1 dpiti! dℓ(σ)tσpσ Aut(σ) . 7 III. Full system of tautological relations. Following Proposition 5 of Moduli of stable quotients [4], we can obtain a much larger set of relations in the tautological ring of Mg by including several factors of π∗(saiωbi) in the integrand instead of just a single factor. We study the associated relations where the ai are always 1. The bi then form the parts of a partition σ. To state the relations we obtain, we start by enriching the function γ from Section B.I, κ2i−1t2i−1 B2i 2i(2i − 1) ∞Xd=1 ∞Xr=−1 dκr+σ tr xd C r d! dℓ(σ)tσpσ Aut(σ) . γ p = Xi≥1 +Xσ bγ p = Xi≥1 +Xσ Letbγ p be defined by a similar formula, B2i 2i(2i − 1) κ2i−1(−t)2i−1 ∞Xr=−1 dκr+σ (−t)r xd C r d! ∞Xd=1 dℓ(σ)tσpσ Aut(σ) . The sign of t in tσ does not change inbγ p. The κ−1 terms which appear The full system of relations are obtain from the coefficients of the will later be set to 0. functions exp(−γ p), exp(− ∞Xr=0 κrtrpr+1) · exp(−bγ p) 8 Theorem 3. In Rr(Mg), the relation h exp(−γ p)itr xdpσ = (−1)gh exp(− holds when g − 2d − 1 + σ < r. ∞Xr=0 κrtrpr+1) · exp(−bγ p)itr xdpσ Again, we see the genus shifting mod 2 property. If the relation holds in genus g, then the same relation holds in genera h = g − 2δ for all natural numbers δ ∈ N. In case σ = ∅, Theorem 3 specializes to the relation h exp(−γ(t, x))itr xd = (−1)gh exp(−γ(−t, x))itr xd = (−1)g+rh exp(−γ(t, x))itr xd , nontrivial only if g ≡ r + 1 mod 2. If the mod 2 condition holds, then we obtain the relations of Theorem 2. Consider the case σ = (1). The left side of the relation is then d κs+1ts+1 dxd C s d! !itr xd . The right side is h exp(−γ(t, x)) · − ∞Xd=1 ∞Xs=−1 (−1)gh exp(−γ(−t, x)) · −κ0t0 + ∞Xd=1 If g ≡ r + 1 mod 2, then the large terms cancel and we obtain C s d κs+1(−t)s+1 dxd d! !itr xd . ∞Xs=−1 −κ0 ·h exp(−γ(t, x))itr xd = 0 . Since κ0 = 2g − 2 and (g − 2d − 1 + 1 < r) =⇒ (g − 2d − 1 < r), we recover most (but not all) of the σ = ∅ equations. If g ≡ r mod 2, then the resulting equation is h exp(−γ(t, x)) · κ0 − 2 ∞Xd=1 ∞Xs=−1 d κs+1ts+1 dxd C s d! !itr xd = 0 when g − 2d < r. 9 IV. Expanded form. Let σ = (1a12a23a3 . . .) be a partition of length ℓ(σ) and size σ. We can directly write the corresponding relation in R∗(Mg) obtained from Theorem 3. A subpartition σ′ ⊂ σ is obtained by selecting a nontrivial subset of the parts of σ. A division of σ is a disjoint union (3) σ = σ(1) ∪ σ(2) ∪ σ(3) . . . of subpartitions which exhausts σ. The subpartitions in (3) are un- ordered. Let S(σ) be the set of divisions of σ. For example, S(1121) = { (1121), (11) ∪ (21) } , S(13) = { (13), (12) ∪ (11) } . We will use the notation σ• to denote a division of σ with subparti- tions σ(i). Let m(σ•) = 1 Aut(σ•) Aut(σ) i=1 Aut(σ(i)) . Qℓ(σ•) Here, Aut(σ•) is the group permuting equal subpartitions. The factor m(σ•) may be interpreted as counting the number of different ways the disjoint union can be made. To write explicitly the pσ coefficient of exp(γ p), we introduce the functions Fn,m(t, x) = − for n, m ≥ 1. Then, Aut(σ) ·h exp(−γ p)itr xd σ p ∞Xd=1 ∞Xs=−1 d κs+mts+m dnxd C s d! = h exp(−γ(t, x)) · Xσ•∈S(σ) m(σ•) ℓ(σ•)Yi=1 Fℓ(σ(i)),σ(i)itr xd . The length ℓ(σ∗,•) is the number of unmarked subpartitions. Let σ∗,• be a division of σ with a marked subpartition, (4) σ = σ∗ ∪ σ(1) ∪ σ(2) ∪ σ(3) . . . , 10 labelled by the superscript ∗. The marked subpartition is permitted to be empty. Let S∗(σ) denote the set of marked divisions of σ. Let m(σ∗,•) = 1 Aut(σ•) Aut(σ) Aut(σ∗)Qℓ(σ∗,•) i=1 . Aut(σ(i)) Then, Aut(σ) times the right side of Theorem 3 may be written as (−1)g+σAut(σ) ·h exp(−γ(−t, x))·  Xσ∗,•∈S ∗(σ) j −1(−t)σ∗ κσ∗ ℓ(σ∗)Yj=1 m(σ∗,•) j −1 ℓ(σ∗,•)Yi=1 Fℓ(σ(i)),σ(i)(−t, x)itr xd To write Theorem 3 in the simplest form, the following definition with the Kronecker δ is useful, m±(σ∗,•) = (1 ± δ0,σ∗) · m(σ∗,•). There are two cases. If g ≡ r +σ mod 2, then Theorem 3 is equivalent to the vanishing of h exp(−γ)· Xσ∗,•∈S ∗(σ) m−(σ∗,•) ℓ(σ∗)Yj=1 κσ∗ j −1tσ∗ j −1 ℓ(σ∗,•)Yi=1 Fℓ(σ(i)),σ(i)itr xd . If g ≡ r +σ + 1 mod 2, then Theorem 3 is equivalent to the vanishing of h exp(−γ)· Xσ∗,•∈S ∗(σ) m+(σ∗,•) ℓ(σ∗)Yj=1 κσ∗ j −1tσ∗ j −1 ℓ(σ∗,•)Yi=1 Fℓ(σ(i)),σ(i)itr xd . In either case, the relations are valid in the ring R∗(Mg) only if the condition g − 2d − 1 + σ < r holds. V. Further examples. If σ = (k) has a single part, then the two cases of Theorem 3 are the following. If g ≡ r + k mod 2, we have 11 h exp(−γ) · κk−1tk−1itr xd = 0 which is a consequence of Theorem 2. If g ≡ r + k + 1 mod 2, we have h exp(−γ) ·(cid:0)κk−1tk−1 + 2F1,k(cid:1)itr xd = 0 If σ = (k1k2) has two distinct parts, then the two cases of Theorem 3 are as follows. If g ≡ r + k1 + k2 mod 2, we have h exp(−γ) ·(cid:0)κk1−1κk2−1tk1+k2−2 + κk1−1tk1−1F1,k2 + κk2−1tk2−1F1,k1(cid:1)itr xd = 0 . If g ≡ r + k1 + k2 + 1 mod 2, we have h exp(−γ) ·(cid:0)κk1−1κk2−1tk1+k2−2 + κk1−1tk1−1F1,k2 + κk2−1tk2−1F1,k1 + 2F2,k1+k2 + 2F1,k1F1,k2(cid:1)itr xd = 0 . In fact, the g ≡ r + k1 + k2 mod 2 equation above is not new. The genus g and codimension r1 = r − k2 + 1 case of partition (k1) yields h exp(−γ) ·(cid:0)κk1−1tk1−1 + 2F1,k1(cid:1)itr1 xd = 0 . After multiplication with κk2−1tk2−1, we obtain h exp(−γ) ·(cid:0)κk1−1κk2−1tk1+k2−2 + 2κk2−1tk2−1F1,k1(cid:1)itr xd = 0 . Summed with the corresponding equation with k1 and k2 interchanged yields the above g ≡ r + k1 + k2 mod 2 case. 12 VI. Expanded form revisited. Consider the partition σ = (k1k2 · · · kℓ) with distinct parts. We obtain from Theorem 3, in the g ≡ r +σ mod 2 case, the vanishing of h exp(−γ)· Xσ∗,•∈S ∗(σ) (1 − δ0,σ∗) ℓ(σ∗)Yj=1 κσ∗ j −1tσ∗ j −1 ℓ(σ∗,•)Yi=1 Fℓ(σ(i)),σ(i)itr xd since all the factors m(σ∗,•) are 1. In the g ≡ r + σ + 1 mod 2 case, we obtain the vanishing of h exp(−γ)· Xσ∗,•∈S ∗(σ) for the same reason. (1 + δ0,σ∗) ℓ(σ∗)Yj=1 κσ∗ j −1tσ∗ j −1 ℓ(σ∗,•)Yi=1 Fℓ(σ(i)),σ(i)itr xd Proposition 2. The g ≡ r + σ mod 2 case is a consequence of the g ≡ r′ + σ′ + 1 mod 2 cases of smaller partitions σ′. Proof. The strategy is identical to that employed in the special cases of the result proven in Section V. (cid:3) If σ has repeated parts, the relations of Theorem 3 are obtained by viewing the parts are distinct and using the above formulas. For example, the two cases of Theorem 3 for σ = (k2) are as follows. If g ≡ r + 2k mod 2, we have h exp(−γ) ·(cid:0)κk−1κk−1t2k−2 + 2κk−1tk−1F1,k(cid:1)itr xd = 0 . If g ≡ r + 2k + 1 mod 2, we have h exp(−γ) ·(cid:0)κk−1κk−1t2k−2 + 2κk−1tk−1F1,k + 2F2,2k + 2F1,kF1,k(cid:1)itr xd = 0 . The factors occur via repetition of terms in the formulas for distinct parts. VII. Differential equations. The function Φ satisfies a basic differential equation obtained from the series definition, 13 1 Φ) = − t After expanding and dividing by Φ, we find (Φ − tx d dx d dx Φ . −tx which can be written as − t2xγ∗ (5) Φxx Φ − t Φx Φ + Φx Φ 1 t = − xx = t2x(γ∗ x)2 + t2γ∗ x − tγ∗ x − 1 where, as before, γ∗ = log(Φ). Equation (5) has been studied by Ionel in Relations in the tautological ring [3]. We present here results of hers which will be useful for us. To kill the pole and match the required constant term, we will con- sider the function Γ = −t Xi≥1 t2i−1 + γ∗! . B2i 2i(2i − 1) The differential equation (5) becomes txΓxx = x(Γx)2 + (1 − t)Γx − 1 . The differential equation is easily seen to uniquely determine Γ once the initial conditions are specified. By Ionel's first result, Γx = −1 + √1 + 4x 2x B2i 2i(2i − 1) t2i Γ(t, 0) = −Xi≥1 ∞Xk=1 1 + 4x + + t kXj=0 tk+1qk,j(−x)j(1 + 4x)−j− k 2 −1 where the postive integers qk,j (defined to vanish unless k ≥ j ≥ 0) are defined via the recursion qk,j = (2k + 4j − 2)qk−1,j−1 + (j + 1)qk−1,j + from the initial value q0,0 = 1. k−1Xm=0 j−1Xl=0 qm,lqk−1−m,j−1−l 14 Ionel's second result is obtained by integrating Γx with respect to x. She finds Γ = Γ(0, x) + t 4 log(1 + 4x) − ∞Xk=1 kXj=0 where the coefficients ck,j are determined by tk+1ck,j(−x)j(1 + 4x)−j− k 2 qk,j = (2k + 4j)ck,j + (j + 1)ck,j+1 for k ≥ 1 and k ≥ j ≥ 0. While the derivation of the formula for Γx is straightforward, the for- mula for Γ is quite subtle as the intial conditions (given by the Bernoulli numbers) are used to show the vanishing of constants of integration. Said differently, the recusions for qk,j and ck,j must be shown to imply the formula ck,0 = Bk k(k − 1) . A third result of Ionel's is the determination of the extremal ck,k, ck,kzk = log ∞Xk=1 ∞Xk=1 (6k)! (2k)!(3k)!(cid:16) z 72(cid:17)k! . The formula for Γ becomes simpler after the following very natural change of variables, (6) u = t √1 + 4x and y = −x 1 + 4x . The change of variables defines a new function The formula for Γ implies bΓ(u, y) = Γ(t, x) . 1 t Γ(u, y) = 1 t Γ(0, y) − 1 4 log(1 + 4y) − ∞Xk=1 kXj=0 ck,jukyj . Ionel's fourth result relates coefficients of series after the change of variables (6). Given any series P (t, x) ∈ Q[[t, x]], let bP (u, y) be the series obtained from the change of variables (6). Ionel proves coefficient relation r+2d−2 (cid:2)P (t, x)(cid:3)tr xd = (−1)d(cid:2)(1 + 4y) 2 · bP (u, y)(cid:3)uryd . VII. Analysis of the relations of Theorem 2 15 We now study in detail the simple relations of Theorem 2, (cid:2) exp(−γ)(cid:3)tr xd = 0 ∈ Rr(Mg) when g − 2d − 1 < r and g ≡ r + 1 mod 2. Let be obtained from the variable change (6), modulo κ−1 terms which we set to 0. Applying Ionel's coefficient result, κ0 4 bγ(u, y) = (cid:2) exp(−γ)(cid:3)tr xd = (cid:2)(1 + 4y) = "(1 + 4y) = "(1 + 4y) log(1 + 4y) + κkck,jukyj bγ(u, y) = γ(t, x) ∞Xk=1 kXj=0 · exp(−bγ)(cid:3)uryd κkck,jukyj)# ∞Xk=1 kXj=0 κkck,jukyj)# ∞Xk=1 kXj=0 4 · exp(− · exp(− r−g+2d−1 r+2d−2 2 − κ0 r+2d−2 2 2 ur yd . uryd In the last line, the substitution κ0 = 2g − 2 has been made. Consider first the exponent of 1 + 4y. By the assumptions on g and r in Theorem 2, r − g + 2d − 1 2 ≥ 0 (1 + 4y) r−g+2d−1 2 and the fraction is integral. Hence, the y degree of the prefactor is exactly r−g+2d−1 from above by the u degree. We conclude 2 . The y degree of the exponential factor is bounded "(1 + 4y) r−g+2d−1 2 · exp(− κkck,jukyj)# ∞Xk=1 kXj=0 = 0 uryd is the trivial relation unless r ≥ d − r − g + 2d − 1 2 r 2 = − + g + 1 2 . Rewriting the inequality, we obtain 3r ≥ g + 1 which is equivalent to r > ⌊ g 3⌋. The conclusion is in agreement with the proven freeness of R∗(Mg) up to (and including) degree ⌊ g 3⌋. 16 A similar connection between Theorem 2 and Ionel's relations in [3] has also been found by Shengmao Zhu [8]. VIII. Analysis of the relations of Theorem 3 For the relations of Theorem 3, we will require additional notation. To start, let γc(u, y) = ∞Xk=1 kXj=0 κkck,jukyj . By Ionel's second result, 1 t Γ = 1 t Γ(0, x) + 1 4 log(1 + 4x) − ∞Xk=1 kXj=0 tkck,j(−x)j(1 + 4x)−j− k 2 . Let c0 k,j = ck,j. We define the constants cn k,j for n ≥ 1 by (cid:18)x d dx(cid:19)n 1 t Γ =(cid:18)x d dx(cid:19)n−1(cid:18)−1 2t + 1 2t − ∞Xk=0 √1 + 4x(cid:19) k+nXj=0 tkcn k,j(−x)j(1 + 4x)−j− k 2 . Lemma 2. For n > 0, there are constants bn j satisfying (cid:18)x d dx(cid:19)n−1(cid:18) 1 2t √1 + 4x(cid:19) = n−1Xj=0 j u−1yj . bn Moreover, bn n−1 = −2n−2 · (2n− 5)!! where (−1)!! = 1 and (−3)!! = −1. Proof. The result is obtained by simple induction. The negative evalu- ations (−1)!! = 1 and (−3)!! = −1 arise from the Γ-regularization. (cid:3) Lemma 3. For n > 0, we have cn 0,n = 4n−1(n − 1)!. Lemma 4. For n > 0 and k > 0, we have cn k,k+n = (6k)(6k + 4)· · · (6k + 4(n − 1)) ck,k. Consider next the full set of equations given by Theorem 3 in the expanded form of Section VI. The function Fn,m may be rewritten as 17 Fn,m(t, x) = − ∞Xs=−1 ∞Xd=1 dx(cid:19)n ∞Xd=1 = −tm(cid:18)x d κs+mts+m dnxd C s d! d κs+mts xd C s d! ∞Xs=−1 d . We may write the result in terms of the constants bn j and cn k,j, t−(m−n)Fn,m = −δn,1 + (1 + 4y)− n Define the functions Gn,m(u, y) by κm−1bn j un−1yj − k+nXj=0 κk+mcn k,juk+nyj(cid:17) κm−1 2 2(cid:16) n−1Xj=0 n−1Xj=0 ∞Xk=0 k+nXj=0 ∞Xk=0 Gn,m(u, y) = κm−1bn j un−1yj − κk+mcn k,juk+nyj . Let σ = (1a12a23a3 . . .) be a partition of length ℓ(σ) and size σ. We assume the parity condition (7) g ≡ r + σ + 1 . Let G± σ (u, y) be the following function associated to σ, G± σ (u, y) = Xσ•∈S(σ) h(1 + 4y) ℓ(σ•)Yi=1 (cid:18)Gℓ(σ(i)),σ(i) ± exp(−γc)(cid:0)G+ 2 r−σ−g+2d−1 δℓ(σ(i)),1 2 p1 + 4y κσ(i)−1(cid:19) . σ(cid:1)iur−σ+ℓ(σ)yd = 0 σ + G− The relations of Theorem 3 written in the variables u and y is In fact, the relations of Theorem 3 can be written in a much more efficient form when the strategy of Proposition 2 is used to take out lower equations. Theorem 4. In Rr(Mg), the relation h(1+4y) r−σ−g+2d−1 2 exp−γc +Xσ6=∅ Gℓ(σ),σ pσ Aut(σ)iur−σ+ℓ(σ)yd = 0 σ p holds when g − 2d − 1 + σ < r and g ≡ r + σ + 1 mod 2. 18 Consider the exponent of 1 + 4y. By the inequality and the parity condition (7), r − σ − g + 2d − 1 2 ≥ 0 and the fraction is integral. Hence, the y degree of the prefactor (1 + 4y) r−σ−g+2d−1 2 is exactly r−σ−g+2d−1 . The y degree of the exponential factor is bounded from above by the u degree. We conclude the relation of Theorem 4 is trivial unless 2 r − σ + ℓ(σ) ≥ d − r − σ − g + 2d − 1 2 r − σ 2 = − + g + 1 2 . Rewriting the inequality, we obtain 3r ≥ g + 1 + 3σ − 2ℓ(σ) which is consistent with the proven freeness of R∗(Mg) up to (and including) degree ⌊ g 3⌋. X. Another form A subset of the equations of Theorem 4 admits an especially simple description. Consider the function 19 Hn,m(u) = 2n−2(2n − 5)!! κm−1un−1 + 4n−1(n − 1)! κmun + ∞Xk=1 (6k)(6k + 4)· · · (6k + 4(n − 1))ck,k κk+muk+n . Proposition 3. In Rr(Mg), the relation h exp− ∞Xk=1 ck,kκkuk −Xσ6=∅ Hℓ(σ),σ pσ Aut(σ)iur−σ+ℓ(σ)p = 0 σ holds when 3r ≥ g + 1 + 3σ − 2ℓ(σ) and g ≡ r + σ + 1 mod 2. The main advantage of Proposition 3 is the dependence on only the function (8) ck,kzk = log ∞Xk=1 ∞Xk=1 (6k)! (2k)!(3k)!(cid:16) z 72(cid:17)k! . Proposition 3 only provides finitely many relations for fixed g and r. In Theorem 5 of Section C.I below, a more elegant set of relations in Rr(Mg) conjectured by Faber-Zagier is presented. In fact, we show Proposition 3 is equivalent to the Faber-Zagier conjecture. 20 C. The conjecture of Faber-Zagier I. The function Ψ A third set of relations is defined as follows Let p = { p1, p3, p4, p6, p7, p9, p10, . . . } be a variable set indexed by integers not congruent to 2 mod 3. Let Ψ(t, p) = (1 + tp3 + t2p6 + t3p9 + . . .) ∞Xi=0 (6i)! (3i)!(2i)! ti + (p1 + tp4 + t2p7 + . . .) Define the constants C r(σ) by the formula ∞Xi=0 (6i)! (3i)!(2i)! 6i + 1 6i − 1 ti Here and below, σ denotes a partition which avoids all parts congruent to 2 mod 3. Let log(Ψ) =Xσ ∞Xr=0 C r(σ) trpσ . γ =Xσ ∞Xr=0 C r(σ) κrtrpσ . Our main result, starting from the stable quotient relations, is the following final form. Theorem 5. In Rr(Mg), the relation (cid:2) exp(−γ)(cid:3)tr σ = 0 p holds when g − 1 + σ < 3r and g ≡ r + σ + 1 mod 2. The relations of Theorem 5 were conjectured earlier by Faber and Zagier from data and a study of the Gorenstein quotient of R∗(Mg). To the best of our knowledge, a relation in R∗(Mg) which is not in the span of the relations of Theorem 5 has not yet been found. In particular, all relations obtained from Theorem 1 to date are in the span of Theorem 5 (and conversely). It is very reasonable to expect the spans of the relations in Theorem 1 and Theorem 5 exactly coin- cide. Whether Theorem 5 exhausts all relations in R∗(Mg) is a very interesting question. 21 Theorem 5 is much more efficient than Theorem 1 for several reasons. Theorem 5 only provides finitely many relations in Rr(Mg) for fixed g and r, and thus may be calculated completely. When the relations yield a Gorenstein ring with socle in Rg−2(Mg), no further relations are possible. However, the relations of Theorem 5 do not always yield such a Gorenstein ring (failing first in genus 24 as checked by Faber). For g < 24, Faber's calculations show Theorem 5 does provide all relations in R∗(Mg). For higher genus g ≥ 24, either Theorem 5 fails to provide all the relations in R∗(Mg) or R∗(Mg) is not Gorenstein. II. Connection to the stable quotient relations Theorem 5 is derived from Proposition 3. In fact, Proposition 3 is equivalent to Theorem 5. The derivation is obtained by a triangular transformation among distinguished generators. A certain amount of differential algebra is required. Consider the relation obtained from the partition σ = (1) in Propo- sition 3 and the Conjecture. For convenience, let A(z) = B(z) = ∞Xi=0 ∞Xi=0 , (6i)! 6i + 1 (6i)! 72(cid:17)i (3i)!(2i)!(cid:16) z 6i − 1(cid:16) z 72(cid:17)i (3i)!(2i)! The conjectures predict no room for different relations in R∗(Mg) for σ = (1), so we must have 1 2 − + z + 6z(cid:18)z d dz(cid:19) log(A) proportional to B/A. We find the equation 1 2 − + z + 6z(cid:18)z d dz(cid:19) log(A) = 1 2 B A holds. Equivalently, 1 2 A + zA + 6z2 dA dz − = 1 2 B 22 More interesting is the partition σ = (11). Here we predict, once the definitions are unwound, that d is a linear combination of 1 and B2/A2. We find the equation z + 4z2 + 36z2(cid:18)z dz(cid:19)2 z + 4z2 + 36z2(cid:18)z log(A) + 24z2(cid:18)z dz(cid:19)2 log(A) + 24z2(cid:18)z dz(cid:19) log A dz(cid:19) log A = 1 4 − d d d 1 4 B2 A2 holds. In fact, the main hypergeometric differential equation satisfied by the function A is 36z2 d2 dz2 A + (72z − 6) d dz A + 5A = 0 . In Section D below, further details describing the use of such differential equations to prove Theorem 5 from Proposition 3 are presented. III. Functions While the functions A(z) and B(z) of Section II have radius of con- vergence 0, an additional double factorial in the denominator yields convergent classical series, 3 2t 3 4t − 3 sin(cid:18)2 sin(cid:18)4 3 sin−1(t)(cid:19) = sin−1(t)(cid:19) = ∞Xi=0 ∞Xi=0 (6i)! 216(cid:19)i (3i)!(2i)!(2i + 1)!!(cid:18) t2 6i − 1(cid:18) t2 216(cid:19)i (3i)!(2i)!(2i + 1)!! 6i + 1 (6i)! , . IV. Remarks Stable quotients relations in R∗(Mg) have several advantages over other geometric constructions. We have already seen here the possi- bility of exact evaluation. Another advantage we have not explored in these lectures is the extension of the stable quotients relations over Mg. The boundary terms of the stable quotients relations are tautological. A study of the relations among the κ classes in the tautological ring of the moduli space of curves of compact type Mc g,n has been under- taken in [5, 6]. For example, the Gorenstein predictions for κ classes 23 are proven there if n ≥ 1. But even for Mg, the extension of the stable quotients relations over Mg has significant consequences. Since we know the stable quotients relations are all relations in R∗(Mg) for g < 24, the following result holds. Proposition 4. For g < 24, we have a right exact sequence R∗(∂Mg) → R∗(Mg) → R∗(Mg) → 0 . Speculations about such right exactness for tautological rings were advanced in [2]. D. The equivalence I. Notation The relations in Theorem 5 and Proposition 3 have a similar flavor. We start with formal series related to A(z) = , ∞Xi=0 (6i)! (3i)!(2i)!(cid:16) z 72(cid:17)i we insert classes κr, we exponentiate, and we extract coefficients to obtain relations among the κ classes. In order to make the similarities clearer, we will introduce additional notation. If F is a formal power series in z, F = with coefficients in a ring R, let {F}κ = be the series with κ-classes inserted. crzr ∞Xr=0 ∞Xr=0 crκrzr Let A be as above, and let B be the function defined in C.II. Let C = B A , and let E = exp(−{log(A)}κ) = exp − ck,kκkzk! . ∞Xk=1 24 We will rewrite the relations of Theorem 5 and Proposition 3 in terms of C and E. The equivalence between the two will rely on properties of the differential equations satisfied by C. II. Rewriting the relations The relations of Theorem 5 conjectured by Faber-Zagier are straight- forward to rewrite using the above notation: (9) "E · exp(cid:16) −n log(1 + p3z + p6z2 + · · · + C(p1 + p4z + p7z2 + · · · ))oκ(cid:17)#zrpσ = 0 for 3r ≥ g + σ + 1 and 3r ≡ g + σ + 1 mod 2. We call the above relations FZ. The stable quotient relations of Proposition 3 are a bit more com- plicated to rewrite in terms of C and E. Let 2−nCn = 2n−2(2n − 5)!!zn−1 + 4n−1(n − 1)!zn + We see ∞Xk=1 (6k)(6k + 4)· · · (6k + 4(n − 1))ck,kzk+n. The series Cn satisfy Hn,m(z) = 2−nzn−m{zm−nCn}κ. dz − 4iz(cid:19) Ci. Ci+1 =(cid:18)12z2 d C1 = C, Since C satisfies the differential equation 12z2 dC dz = 1 + 4zC − C 2, each Cn can be expressed as a polynomial in C and z: C1 = C, C2 = 1 − C 2, C3 = −8z − 2C + 2C 3, . . . , . 25 Proposition 3 can then be rewritten as follows (after an appropriate change of variables): (10) E · exp−Xσ6=∅ {zσ−ℓ(σ)Cℓ(σ)}κ pσ Aut(σ)zrpσ = 0 for 3r ≥ g + 3σ − 2ℓ(σ) + 1 and 3r ≡ g + 3σ − 2ℓ(σ) + 1 mod 2. We call the stable quotients relations SQ. The FZ and SQ relations now look much more similar, but the rela- tions in (9) are indexed by partitions with no parts of size 2 mod 3 and satisfy a slightly different inequality. The indexing differences can be erased by noting the variables p3k are actually not necessary in (9) if we are just interested in the ideal generated by a set of relations (rather than the linear span). If we remove the variables p3k and reindex the others, we obtain the following equivalent form of the FZ relations: (11) hE · exp(cid:0) −(cid:8) log(1 + C(p1 + p2z + p3z2 + · · · ))(cid:9)κ(cid:1)izrpσ = 0 for 3r ≥ g + 3σ − 2ℓ(σ) + 1 and 3r ≡ g + 3σ − 2ℓ(σ) + 1 mod 2. II. Comparing the relations We now explain how to write the SQ relations (10) as linear com- binations of the FZ relations (11) with coefficients in Q[κ0, κ1, κ2, . . .]. In fact, the associated matrix will be triangular with diagonal entries equal to 1. We start with further notation. For a partition σ, let and FZσ =(cid:2)exp(cid:0)−(cid:8)log(1 + C(p1 + p2z + p3z2 + · · · ))(cid:9)κ(cid:1)(cid:3)pσ Aut(σ)pσ SQσ =exp−Xσ6=∅ {zσ−ℓ(σ)Cℓ(σ)}κ pσ be power series in z with coefficients that are polynomials in the κ classes. The relations themselves are given by [E·SQσ]zr and [E·FZσ]zr. 26 For each σ, we can write SQσ in terms of the FZσ. For example, κ 1 6{C1}3 1 (κ0 − {C 2}κ){C}κ − 2 1 6{C}3 κ SQ(111) = − 4 3 = 3 3 1 6{C3}κ + κ1z + 1 2{C2}κ{C1}κ − 1 1 3{C 3}κ + 3{C}κ − 2(cid:19){C}κ(cid:19) κ0 + 1 2{C 2}κ{C}κ − 2(cid:19) FZ(1) + FZ(111) . 1 3 − 1 3{C 3}κ − κ0 =(cid:18) 4 κ1z(cid:19) +(cid:18)(cid:18) 1 +(cid:18)− κ1z FZ∅ +(cid:18)− κ1[E·FZ∅]zr−1+(cid:18)− 1 3 − 4 3 4 3 = [E·SQ(111)]zr = κ(cid:19) 1 6{C}3 We then obtain a corresponding linear relation between the relations themselves: κ0 2(cid:19) [E·FZ(1)]zr +[E·FZ(111)]zr. Constructing such linear combinations in general is not hard. When expanded in terms of C as in the above example, FZσ always contains exactly one term of the form (12) {za1C}κ{za2C}κ · · ·{zamC}κ . All the other terms involve higher powers of C. If we expand SQσ in terms of C, we can look at the terms of the form (12) which appear to determine how to write the SQσ as a linear combination of the FZbσ. We must check the terms involving higher powers of C also match up. The matching amounts to proving an identity between the coefficients of Ci when expressed a polynomial in C. Define polynomials fij ∈ Z[z] by and let fijC j, Ci = iXj=0 f = 1 +Xi,j≥1 (−1)j−1fij i!(j − 1)! xiyj. Lemma 5. There exists a power series g ∈ Q[z][[x]] such that f = eyg. The Lemma (which can be proven in straightforward fashion using the differential equation satisfied by f ) is the precise consistency state- ment needed to express the SQ relations as linear combinations of the FZ relations. The associated matrix is triangular with respect to the partial ordering of partitions by size, and the diagonal entries are eas- ily computed to be equal to 1. Hence, the matrix is invertible. We conclude the SQ relations are equivalent to the FZ relations. 27 References [1] C. Faber, A conjectural description of the tautological ring of the moduli s pace of curves, Moduli of curves and abelian varieties, 109 -- 129, Aspects Math., Vieweg, Braunschweig, 1999. [2] C. Faber and R. Pandharipande, Relative maps and tautological classes, JEMS 7 (2005), 13 -- 49. [3] E. Ionel, Relations in the tautological ring of Mg, Duke Math. J. 129 [4] A. Marian, D. Oprea, and R. Pandharipande, The moduli space of stable (2005), 157 -- 186. quotients, arXiv:0904.2992. [5] R. Pandharipande, The κ ring of the moduli of curves of compact type: I, arXiv:0906.2657. [6] R. Pandharipande, The κ ring of the moduli of curves of compact type: II, arXiv:0906.2658. [7] R. Pandharipande and A. Pixton, Stable quotients and the Faber-Zagier relations, in preparation. [8] S. Zhu, Note on the relations in the tautological ring of Mg, preprint 2010.
1508.00202	2	1508	2015-10-23T12:50:15	Duality of Multiple Root Loci	[ "math.AG", "cs.SC", "math.OC" ]	The multiple root loci among univariate polynomials of degree $n$ are indexed by partitions of $n$. We study these loci and their conormal varieties. The projectively dual varieties are joins of such loci where the partitions are hooks. Our emphasis lies on equations and parametrizations that are useful for Euclidean distance optimization. We compute the ED degrees for hooks. Among the dual hypersurfaces are those that demarcate the set of binary forms whose real rank equals the generic complex rank.	math.AG	math	Duality of Multiple Root Loci Hwangrae Lee and Bernd Sturmfels Abstract The multiple root loci among univariate polynomials of degree n are indexed by par- titions of n. We study these loci and their conormal varieties. The projectively dual varieties are joins of such loci where the partitions are hooks. Our emphasis lies on equations and parametrizations that are useful for Euclidean distance optimization. We compute the ED degrees for hooks. Among the dual hypersurfaces are those that demarcate the set of binary forms whose real rank equals the generic complex rank. 1 Introduction Univariate polynomials of degree n correspond to points in a projective space Pn. The multiple root locus ∆λ associated with a partition λ = (λ1, . . . , λd) is the subvariety of Pn given by polynomials that have d distinct roots with multiplicities λ1, . . . , λd. The dimension of ∆λ is d. The singular locus of ∆λ is the union of certain codimension one subloci ∆µ, as described in [15, §3]. The degree of ∆λ was determined by Hilbert in [11]. He showed that (1) deg(∆λ) = d! m1!m2!· · · mp! · λ1λ2 · · · λd, where mj denotes the number of parts λi in the partition λ that are equal to the integer j. The multiple root loci ∆λ have been studied in a wide range of contexts and under various different names: coincident root loci [5, 9, 15], pejorative manifolds [16, 23], strata of the discriminant [11, 14, 18], λ-Chow varieties [19], factorization manifolds [24, Definition 5.2.4], etc. Our motivation arose from the desire to understand the geometry of a model selection problem considered at the interface of symbolic computation [13] and numerical analysis [23]: given a univariate polynomial h, identify a low-dimensional ∆λ such that h is close to ∆λ. Finding a point in ∆λ that is closest to a given h is a problem of polynomial optimization [4]. We here characterize the geometric duality that underlies this optimization problem, in the sense of [4, Chapter 5]. The key player is the dual variety (∆λ)∨. This variety lives in the dual projective space Pn and it parametrizes all hyperplanes that are tangent to ∆λ. The duals to multiple root loci were studied by Oeding in [19]. He shows that (∆λ)∨ is a hypersurface if and only if m1 = 0, i.e. all parts of λ satisfy λi ≥ 2. In [19, Theorem 5.3] an explicit formula is given for the degree of the polynomial that cuts out this hypersurface: deg(cid:0)(∆λ)∨(cid:1) = (d + 1)! m2!· · · mp! · (λ1 − 1)(λ2 − 1) · · · (λd − 1). (2) 1 For the application to optimization in [4, Theorem 5.23], this is the number of complex critical points one encounters when minimizing a general linear function over an affine chart of ∆λ. For instance, consider n = 5 and λ = (3, 2). Following Example 4.2 and [6, §3], the polynomial for (∆λ)∨ is the apple invariant of degree 12. So, optimizing a linear function over quintics with a triple root and a double root leads to solving an equation of degree 12. The present paper is a continuation of the studies by Hilbert [11] and Oeding [19]. It is organized as follows. In Theorem 2.5 we parametrize the conormal variety Conλ that links ∆λ and (∆λ)∨. Theorem 2.9 offers a parametrization for the projective duals of multiple root loci. These results are derived from the apolarity theory for binary forms, as described in the book by Iarrobino and Kanev [12]. In Section 3 we study the multidegree of the conormal variety Conλ, and we summarize what is known about the ideals of ∆λ and (∆λ)∨. Table 1 offers a census of small instances up to n = 7. We also examine duality for hooks, and inclusions among the dual strata. In Section 2 we set up notation and basics. Section 4 offers an application to tensor rank over R. Theorem 4.1 characterizes the algebraic boundary of the set of binary forms whose real rank equals the generic complex rank. When n is even then this involves the use of Chow forms [10] and Hurwitz forms [21]. In Section 5 we turn to Euclidean distance (ED) optimization. Theorem 5.1 determines the ED degree of ∆λ for hook shapes λ, both for invariant coordinates and for generic coordinates. This generalizes known results for the rational normal curve ∆(n), in particular the tight connection to eigenvectors of symmetric tensors; cf. [8, Ex. 5.2 and Cor. 8.7]. Section 6 discusses implications for the ED optimization problem (21). The focus lies on locating all the critical points in ∆λ and in (∆λ)∨, and on certifying the global optimum. We work this out for several examples, including one from tensor decomposition over R. 2 Duality for Binary Forms We fix a field K of characteristic zero and we represent univariate polynomials by binary forms. Let V = K[x, y]n denote the space of binary forms of degree n and V ∨ = K[u, v]n its dual vector space. For f ∈ V and g ∈ V ∨, the pairing is defined by using partial derivatives: hf, gi = 1 n! · g(cid:18) ∂ ∂x , ∂ ∂y(cid:19) f (x, y) = 1 n! · f(cid:18) ∂ ∂u , ∂ ∂v(cid:19) g(u, v). (3) This is the scalar in K obtained by interpreting one polynomial as a differential operator and applying it to the other polynomial. Introducing coordinates on V and V ∨, we have: If f =Pn i=0(cid:0)n i(cid:1)aixiyn−i and g =Pn i=0(cid:0)n i(cid:1)biuivn−i then hf, gi = Pn i=0(cid:0)n i(cid:1)aibi. We regard f and g as elements in the projective spaces P(V ) and P(V ∨). Both of these spaces are identified with Pn using homogeneous coordinates (a0 : · · · : an) and (b0 : · · · : bn). A partition λ of n is represented alternatively as a list of integers (λ1, . . . , λd) satisfying λ1 ≥ · · · ≥ λd ≥ 1 andPd j=1 jmj = n. So, mj = #{i : λi = j}, and d is the number of parts of λ, and p is the largest part of λ. i=1 λi = n, or as a multiset {1m1, . . . , pmp} satisfyingPp (4) 2 The multiple root variety ∆λ ⊂ Pn comprises all binary forms f of degree n that have mj roots of multiplicity j. Equivalently, writing ℓi for linear forms, ∆λ is the image of (P1)d → Pn, (ℓ1, ℓ2, . . . , ℓd) 7→ f = ℓλ1 1 ℓλ2 2 · · · ℓλd d . The variety ∆λ has dimension d and degree as in (1). Among its smooth points are those given by d distinct linear forms ℓi. We determine the tangent space of ∆λ at such a point. Lemma 2.1. The tangent space of ∆λ at a general smooth point f =Qd h ∈ P(K[x, y]d)) ≃ Pd. Tf ∆λ = (h(x, y) · ℓλi−1 i d i=1 ℓλi i equals Yi=1 (x, y) (cid:12)(cid:12)(cid:12)(cid:12) ∂ ∂ai f = xf /ℓi and ∂ ∂bi Proof. Write ℓi = aix + biy for some indeterminate (ai : bi) ∈ P1. The tangent space Tf ∆λ is spanned by the binary forms f = yf /ℓi for i = 1, 2, . . . , d. These generators have the required form if we take h(x, y) to be x·(Qd j=1 ℓj)/ℓi. For the converse we note that K[x, y]d−1 is spanned by (cid:8)(Qd this is a Lagrange basis, since the ℓi are pairwise linearly independent. Hence, by multiplying these polynomials with x and y, we obtain a spanning set for the vector space K[x, y]d. Lemma 2.2. Let f be as above and g ∈ P(V ∨). Then g ⊥ Tf ∆λ if and only if j=1 ℓj)/ℓi i = 1, . . . , d(cid:9). Indeed, j=1 ℓj)/ℓi or y·(Qd " d Yi=1 ℓλi−1 i ∂ ∂u , (cid:0) ∂ ∂v(cid:1)# annihilates g(u, v). (5) Proof. Let g(u, v) be the binary form of degree d obtained from g(u, v) by applying the operator on the left of (5). Then g is zero if and only if g(u, v) is annihilated by h( ∂ ∂v ) ∂v ) for all for all h ∈ K[x, y]d if and only if g(u, v) is annihilated by (cid:0)h ·Qd h ∈ K[x, y]d. By Lemma 2.1, this means that g(u, v) is orthogonal to the space Tf ∆λ. i=1 ℓλi−1 (cid:1)( ∂ ∂u , ∂ ∂u , ∂ i The conormal variety of ∆λ, here denoted Conλ, is the Zariski closure of the set (cid:8)(f, g) (cid:12)(cid:12) f is a smooth point of ∆λ and g ⊥ Tf ∆λ(cid:9) in P(V ) × P(V ∨) = Pn × Pn. General results on projective duality ensure that Conλ is an irreducible variety of dimension n−1 in Pn × Pn. The dual variety (∆λ)∨ is the image of the conormal variety Conλ under projection onto the second factor P(V ∨). This is an irreducible variety of dimension ≤ n− 1. The Biduality Theorem [10] states that the conormal variety of (∆λ)∨ coincides with Conλ, and hence ((∆λ)∨)∨ = ∆λ. Lemma 2.2 implies the following description of the dual variety. Corollary 2.3. The points on the variety (∆λ)∨ are binary forms g(u, v) that are annihilated by some order n − d operator of the form Qd i=1 ℓλi−1 i ∂u , ∂ (cid:0) ∂ ∂v(cid:1) where ℓ1, . . . , ℓd ∈ K[x, y]1. At this point, let us pause to illustrate the concepts seen above with a small example. 3 Example 2.4. Let n = 3 and take λ = (3), the partition with a single part. The conormal variety Con(3) is the surface in P2 × P2 whose defining homogeneous prime ideal equals (6) (cid:10) a2 2 − a1a3, a1a2 − a0a3, a2 a0b0 + 2a1b1 + a2b2, a0b1 + 2a1b2 + a2b3, a1b0 + 2a2b1 + a3b2, a1b1 + 2a2b2 + a3b3(cid:11). 1b2 1 − a0a2 , 3b2 2 − 4b0b3 0b2 1b3 + 6b0b1b2b3 − b2 3 2 − 4b3 The multiple root locus ∆(3) is the twisted cubic curve in P3, consisting of cubes of linear forms, and defined by the first three quadrics in (6). Its dual is (∆(3))∨ = ∆(2,1). It consists of binary cubics with a double root, and it is the discriminant surface in P3 defined by the quartic in (6). We have Con(3) = Con(2,1), after swapping the a-variables with the b-variables. We illustrate Corollary 2.3 for the point (cid:0)(1 : 1 : 1 : 1), (1 : 2 : −5 : 8)(cid:1) in Con(3). This represents the pair (f, g) where f = (x + y)3 and g = (u − v)2(u + 8v). Then g(u, v) lies in (∆(3))∨ because ( ∂ ∂v )2g(u, v) = 0, and f (x, y) lies in (∆(2,1))∨ because ( ∂ ∂y )f (x, y) = 0. ♦ Note that the degree formulas (1) and (2) evaluate to 3 and 4 for λ = (3). ∂x− ∂ ∂u + ∂ The main result in this section is a parametric representation of the conormal variety. Theorem 2.5. The conormal variety Conλ is the set of pairs (f, g) ∈ Pn × Pn of the form f (x, y) = d Yi=1 (tix − siy)λi and g(u, v) = d Xi=1 λi6=1 (siu + tiv)n−λi+2 · gi(u, v), (7) where (si : ti) runs over P1 and gi runs over binary forms of degree λi− 2. This parametriza- tion is a finite-to-one map (P1)m1 ×Jλ 99K Conλ ⊂ Pn × Pn whose degree is m1!m2!· · · mp!. In order for this theorem to make sense, we need to define the parameter space. We set Jλ = Join(cid:0)P1 × Pλ1−2, P1 × Pλ2−2, . . . , P1 × Pλd−2(cid:1). This is the free join (or abstract join) of the d − m1 varieties in the argument. Here m1 is subtracted from d because the j-th argument is empty and gets deleted when λj = 1. Note that Jλ is the projective toric variety whose associated polytope is the free join of the product of simplices σ1 × σλj −2 for j = 1, 2, . . . , d. The dimension of this join equals dim(Jλ) = (d − m1 − 1) + d Xi=1 max(0, 1 + (λi − 2)) = n − m1 − 1. The point of the toric variety Jλ is to ensure that (7) gives a well-defined rational map. We will see in Example 2.10 that it is not a morphism. But that is not a problem since we can replace both the parameter space and image by their affine cones. The formulas in (7) are homogeneous. We use them in the next section to carry out the computations for Table 1. By projection onto the second factor, Theorem 2.5 immediately yields a parametrization (P1)m1 × Jλ 99K (∆λ)∨ ⊂ Pn of the dual variety. Indeed, (∆λ)∨ consists of all polynomials g(u, v) of the form in (7). This can be rephrased in the language of projective geometry: 4 Corollary 2.6. The dual variety (∆λ)∨ indexed by λ is the join of d − m1 multiple root loci determined by hook shapes, namely the loci ∆{1λi −2, n−λi+2}, where i = 1, . . . , d with λi ≥ 2. Recall that a partition λ is a hook if at most one part is not 1. This is special for us: Corollary 2.7. The dual variety (∆λ)∨ is also a multiple root locus ∆µ if and only if the In that case λ = {1n−a, a} and µ = {1a−2, n − a + 2} for some partition λ is a hook. a ∈ {2, . . . , n}. In particular, ∆λ is self-dual whenever n is even and λ = {1n/2−1, n/2 + 1}. We easily derive the theorem from results that are well-known in commutative algebra. ∂x, ∂ Proof of Theorem 2.5. We use apolarity as presented in the book of Iarrobino and Kanev [12]. Let f and g be binary forms, possibly of different degrees. We say that f is apolar to g if g(x, y) is annihilated by the operator f ( ∂ ∂y ). Our Lemma 2.2 says that the tangent space Tf ∆λ consists of all binary forms g of degree n such that ℓλ1−1 is apolar to g. By [12, Lemma 1.31], this condition is equivalent to g having a generalized additive decomposition of the form g = ℓn−(λ1−1)+1 gd. Here, only terms with λi ≥ 2 may appear. This is precisely the representation on the right of (7), and we conclude that the proposed parametric representation of the conormal variety Conλ is correct. The parametrization is finite-to-one because dim(Conλ) = n − 1, which is the dimension of the parameter space (P1)m1 × Jλ. Consider the fiber over a general point (f, g) in Conλ. The first entry f is the image of precisely m1!m2!· · · mp! points in (P1)d. For each such point (cid:0)(s1 : t1), . . . , (sd : td)(cid:1), the second entry gives a homogeneous system of linear equations: g1 + · · · + ℓn−(λd−1)+1 · · · ℓλd−1 ℓλ2−1 2 1 d const. · g(u, v) = (siu + tiv)n−λi+2 · gi(u, v). (8) 1 d d Xi=1 λi6=1 The unknowns are the coefficients of g1, . . . , gd. The solution set is a linear subspace of Jλ. It is non-empty since (f, g) was assumed to lie in Conλ. Since the parametrization is finite-to-one, the solution space of the linear system must be a point. We conclude that the parametrization of the conormal variety Conλ given in (7) has degree m1!m2!· · · mp!. At this point it is important to note that our approach in this section is quite classical. All the ingredients are known and have been published elsewhere. What is new is their arrangement and interpretation. For instance, the parametrization (7) appears in [12] but the connection to dual and conormal varieties was not made explicit. Likewise, our Lemma 2.1 is among the statements in [14, Theorem 7.1]. The dual variety does make an appearance in [14, Section 6] but it was not seen that the dual of a multiple loot locus for a hook is again such a locus. By putting all the known puzzle pieces together, we can go further in this paper. For instance, Katz states an inequality in [14, Corollary 6.4]. He conjectures in his introduction [14, p. 220] that equality holds. The following result proves Katz’ conjecture. Corollary 2.8. The variety of binary forms of degree n with a root of multiplicity a satisfies deg(cid:0)(∆{1n−a,a})∨(cid:1) = deg(∆{1n−a+1,a−1}) = (a − 1)(n − a + 2). 5 Proof. The dual on the left is ∆{1a−2,n−a+2} by Corollary 2.7. Using Hilbert’s formula (1), we see that both varieties have the same degree, namely (a − 1)(n − a + 2). We close this section by recording the dimension of the dual variety, and by pointing out that our parametrization is in fact always identifiable, i.e. birational modulo permutations. Theorem 2.9. For any partition λ of n, the dual variety (∆λ)∨ has dimension n−m1−1. The generalized additive decomposition in (7) represents a finite-to-one parametrization Jλ 99K (∆λ)∨ ⊂ Pn. This rational map has degree m2!· · · mp! and it is given by (cid:18)(cid:0)(ti : si))λi6=1(cid:1) , (g1, . . . , gd)(cid:19) 7→ g(u, v) = d Xi=1 λi6=1 (siu + tiv)n−λi+2 · gi(u, v). (9) Proof. It was shown in [14, Corollary 7.3] that dim((∆λ)∨) = n − m1 − 1. This is also the dimension of the parameter space Jλ. Theorem 2.5 implies that the map (9) is finite-to-one. Fix a generic point g(u, v) in the dual variety (∆λ)∨. Consider any smooth point f (x, y) of the primal ∆λ at which g is tangent. The unordered set of linear forms ℓi that appear with multiplicity ≥ 2 in f (x, y) can be recovered uniquely by Lemma 2.2. Permuting linear forms that appear with the same multiplicity accounts for the size m2!· · · mp! of the fiber when writing out the multipliers (siu + tiv)n−λi+2 in (9). At this point, we still need to recover the forms (g1, . . . , gd), but this can be done uniquely by the same argument as in the proof of Theorem 2.5. We conclude that the degree of the map (9) equals m2!· · · mp! Example 2.10. The rational parametrization in Theorem 2.9 can have base points, so it is generally not a morphism. Let λ = (3, 2), so (∆λ)∨ is the hypersurface in P5 referred to as little apple in Example 4.2. The toric fourfold Jλ is the free join of P1 × P1 and P1 × P0. In these products, we fix the points(cid:0)(s1 : t1), g1(cid:1) and(cid:0)(s2 : t2), g2(cid:1), where s1 = s2 = t1 = t2 = 1, g1 = u + v, and g2 = −1. We also fix the scalar 1 for a point on the line that joins them. These choices specify a point in Jλ. Plugging into the formula (9), we obtain g(u, v) = (s1u + t1v)4 · g1 + (s2u + t2v)5 · g2 = (u + v)4 · (u + v) + (u + v)5 · (−1) = 0. ♦ Hence our point in Jλ is a base point of the parametrization (9) for λ = (3, 2). 3 Equations, Multidegree, and More Given any parametrically represented variety, it is natural to ask for its implicitization. This concerns finding the ideals of polynomials that vanish on Conλ, on ∆λ, and on (∆λ)∨. For the multiple root loci, this is a well-studied problem [5, 15, 22]. Before reviewing what is known, we give the relevant definitions and we present a census of our varieties for n ≤ 7. We write I(Conλ) for the ideal of the conormal variety in the N2-graded polynomial ring Q[a, b] = Q[a0, a1, . . . , an, b0, b1, . . . , bn], with deg(ai) = (1, 0), deg(bi) = (0, 1). 6 λ 2 3 21 4 31 211 22 5 41 311 2111 221 32 6 51 411 3111 21111 2211 222 33 321 42 7 61 511 4111 31111 211111 22111 2221 3211 322 331 421 43 52 Eqns of ∆λ Multidegree Eqns of (∆λ)∨ Hooks ED-degrees 2 23 4 26 21, 31 6 37 210 23 46 8 510 428 215 26 21, 33, 41 41, 53, 61 10 713 445 329 41, 53, 631 21, 33, 431 221 210 23, 34 420 610 12 916 678 610, 838 6364 310 420, 642 310, 466 23, 34, 438 2, 2 0, 3, 4 4, 3, 0 0, 0, 4, 6 0, 6, 6, 0 6, 4, 0, 0 0, 4, 6, 3 0, 0, 0, 5, 8 0, 0, 8, 9, 0 0, 9, 8, 0, 0 8, 5, 0, 0, 0 0, 12, 16, 6, 0 0, 0, 12, 21, 12 0, 0, 0, 0, 6, 10 0, 0, 0, 10, 12, 0 0, 0, 12, 12, 0, 0 0, 12, 10, 0, 0, 0 10, 6, 0, 0, 0, 0 0, 24, 30, 10, 0, 0 0, 0, 8, 16, 12, 4 0, 0, 0, 9, 18, 12 0, 0, 36, 56, 24, 0 0, 0, 0, 16, 30, 18 0, 0, 0, 0, 0, 7, 12 0, 0, 0, 0, 12, 15, 0 0, 0, 0, 15, 16, 0, 0 0, 0, 16, 15, 0, 0, 0 0, 15, 12, 0, 0, 0, 0 12, 7, 0, 0, 0, 0, 0 0, 40, 48, 15, 0, 0, 0 0, 0, 32, 60, 40, 10, 0 0, 0, 72, 105, 40, 0, 0 0, 0, 0, 36, 80, 66, 24 0, 0, 0, 27, 48, 24, 0 0, 0, 0, 48, 80, 36, 0 0, 0, 0, 0, 24, 51, 36 0, 0, 0, 0, 20, 39, 24 2 4 23 6 21, 31 26 3 8 46 23 210 34 12 10 41, 53, 61 21, 33, 41 26 215 310 4 12 41, 61 18 12 610 420 23, 34 210 221 320 46 45, 610 24 78 81, 103, 112, 1249 36 24 2 21 3 211 31 4 4, 4 2111 311 41 5 5, 5 41, 5 21111 3111 411 51 6 6, 6 6, 6, 6 51, 51 51, 6 411, 6 211111 31111 4111 511 61 7 7, 7 7, 7, 7 61, 7 61, 7, 7 61, 61 511, 7 511, 61 4111, 7 2, 4 3, 7 3, 7 4, 10 4, 12 4, 10 7, 13 5, 13 5, 17 5, 17 5, 13 16, 34 21, 45 6, 16 6, 22 6, 24 6, 22 6, 16 28, 64 20, 40 19, 39 44, 116 26, 64 7, 19 7, 27 7, 31 7, 31 7, 27 7, 19 43, 103 62, 142 73, 217 94, 206 39, 99 52, 164 51, 111 31, 83 Table 1: Multiple root loci, their duals, and their conormal varieties for n ≤ 7. 7 The ideal I(Conλ) is bihomogeneous and prime. According to textbook definition [17, §8.5], the multidegree of the N2-graded algebra Q[a, b]/I(Conλ) is a binary form of degree n + 1: Cλ(t1, t2) = δ1t1tn 2 + δ2t2 1tn−1 2 + · · · + δntn 1 t2. (10) The coefficients δi are nonnegative integers, known among geometers as the polar classes of the variety ∆λ. Geometrically, δi is the number of intersection points in (Li × Mi) ∩ Conλ where Li is a generic plane of dimension i in P(V ) and Mi is a generic plane of codimension i−1 in P(V ∨). In particular, we have δi = 0 for i < codim(∆λ) and for i > dim(cid:0)(∆λ)∨) + 1. The ideal of the multiple root locus ∆λ is obtained by eliminating the variables b0, . . . , bn, and the ideal of its dual (∆λ)∨ is obtained by eliminating the variables a0, . . . , an. In symbols, I(∆λ) = I(Conλ) ∩ Q[a] and I(cid:0)(∆λ)∨(cid:1) = I(Conλ) ∩ Q[b]. (11) 1 2 1 tdim(∆λ)+1 2 These elimination ideals are N-graded and prime. Their degrees and codimensions in the respective Pn can be read off from the first term and the last term of the multidegree: Cλ(t1, t2) = deg(∆λ) · tcodim(∆λ) We computed the objects in (10) and (11) for all partitions up to n = 7. Our results are listed in Table 1. Each row corresponds to one partition λ. The second column lists the degrees of the minimal generators of I(∆λ). The third column lists the polar classes δ1, δ2, . . . , δn. The leftmost nonzero entry in that list is the degree of ∆λ and the rightmost nonzero entry is the degree of (∆λ)∨. The codimension of the variety in question is the number of consecutive left (resp. right) zeros plus one. The fourth column lists the degrees + · · · + deg((∆λ)∨) · tdim((∆λ)∨)+1 tcodim((∆λ)∨) . of the minimal generators of I(cid:0)(∆λ)∨(cid:1). The fifth column lists the corresponding collection of hooks {1λi−2, n−λi+2} that make up the dual variety in the join construction of Corollary 2.6. Finally, the last column refers to the Euclidean distance degrees in two coordinate sys- tems. This will be explained in Section 5. The second one is always Cλ(1, 1) = δ1+δ2+· · ·+δn. We now discuss some general facts that a reader might discover by looking at Table 1. Whenever all hooks dual to λ are identical then (∆λ)∨ is a secant variety of that hook variety. Proposition 3.1. Suppose n ≥ k(n− a + 2). The k-th secant variety of the variety of binary forms of degree n with a root of multiplicity a is projectively dual to the multiple root locus with partition λ = {1n−k(n−a+2), (n − a + 2)k}. In symbols, σk(cid:0)∆{1n−a,a}(cid:1) = (∆λ)∨. This secant variety is non-defective: it has the expected dimension k(n − a + 1) + k − 1. Proof. This is the special case of Corollary 2.6 where λ has the shape of a rectangle together with some singleton blocks, say λ = {1n−ku, uk}, and we set a = n − u + 2. The statement concerning the dimension follows from the first sentence in Theorem 2.9. (12) The best known special case of this statement arises when a = n, or u = 2. The dual of ∆{1n−2k ,2k} is the k-th secant variety of the rational normal curve ∆(n). Its ideal is generated in degree k + 1, namely by the minors of a Hankel matrix. We see this in Table 1 for λ = 21, 211, 22, 2111, 221, . . .. Also of interest is the case of a rectangular partition, when the secant variety (12) is a hypersurface. Oeding’s formula (2) and Proposition 3.1 imply 8 Corollary 3.2. The secant variety referred to in Proposition 3.1, under the same hypothesis, is a hypersurface if and only if n = k(n − a + 2). The degree of that hypersurface is deg(cid:0) σk( ∆{1n−a, a} )(cid:1) = (k + 1)(n − a + 1)k. Example 3.3. Let k = 2, n = 6, a = 5, so λ = (3, 3). The variety (12) is a hypersurface of degree 12 in P6. It consists of all binary sextics f = ℓ5 3ℓ4 where the ℓi are linear forms. Consider also the case n = 12. Here the construction gives four interesting hypersurfaces: 1ℓ2 + ℓ5 λ a k 8 2 3 10 4 11 6 12 (2, 2, 2, 2, 2, 2) (6, 6) (4, 4, 4) (3, 3, 3, 3) ℓ11 1 g1 + ℓ11 ℓ12 1 + ℓ12 forms ℓ8 1g1 + ℓ8 2g2 ℓ10 1 g1 + ℓ10 2 g2 + ℓ10 3 g3 2 g2 + ℓ11 3 g3 + ℓ11 2 + · · · + ℓ12 6 4 g4 degree 75 108 80 7 Of course, the last hypersurface is defined by the determinant of a 7 × 7 Hankel matrix. ♦ The containment relation among multiple root loci is the order on partitions by refine- ment. Indeed, ∆λ ⊂ ∆µ if and only if µ refines λ, i.e. every part of λ is a sum of parts of µ. Our next result characterizes the containment relation among their dual varieties. Note that (∆λ)∨ ⊂ (∆µ)∨ (13) cannot hold unless λ has more 1’s than µ, for dimension reasons. Given a partition λ = {1m1, 2m2, . . . , pmp}, we write λ′ for the partition {1m2, . . . , (p − 1)mp}. Note that if λ is a partition of n then λ′ is a partition of λ′ = n − d, where d =P mi is the number of parts. Proposition 3.4. Given two partitions λ and µ of n, the inclusion (13) holds if and only if λ′ ≤ µ′ and, by adding to the parts, λ′ can be transformed to a partition λ′′ refined by µ′. Proof. This can be seen from Corollary 2.3. Let ℓλ′ be the differential operator that is used in that corollary to characterize (∆λ′)∨. Here ℓ represents an arbitrary collection of linear forms. Our refinement condition means that every form g annihilated by ℓλ′ is also annihilated by ℓµ′. This condition is equivalent to (13). More precisely, write ma for the number of parts in λ′, and mb for that in µ′. If such a λ′′ exists, then we claim that [ℓma Ann((ℓma)λ′ ) ⊂ [ℓma Ann((ℓma)λ′′ ) ⊂ [ℓmb Ann((ℓmb)µ′ ). Here, ℓk is a k-vector of linear forms and the exponent is written in multi-index notation. The first inclusion holds since each (ℓma)λ′′ can be seen as a multiple of (ℓma)λ′, and the second inclusion holds since being annihilated by (ℓma)λ′′ can be seen as a special case of being annihilated by (ℓmb)µ′, where ℓmb is obtained by duplicating some parts of ℓma. Example 3.5. We can check the inclusion (13) for various cases in Table 1. The inclusion holds for λ = 2211 and µ = 321. Indeed, λ′ = 11, and µ′ = 21 refines λ′′ = 21. Likewise, if λ = 321 and µ = 222 then µ′ = 111 refines λ′ = λ′′ = 21. This explains the inclusions (∆2211)∨ ⊂ (∆321)∨ ⊂ (∆222)∨ ⊂ P6. It follows that the unique quartic polynomial 41 that vanishes on (∆321)∨ must be the 4× 4 Hankel determinant whose hypersurface is (∆222)∨. ♦ 9 We now discuss some of the entries in the column Eqns of ∆λ in Table 1, and thereby review the literature on the ideals I(∆λ). For λ = (n), we see the (cid:0)n 2(cid:1) quadrics that define the rational number curve, and for λ = {1n−2, 2} we get the discriminant of degree 2n− 2. If a ≥ ⌊n/2⌋ + 2 then I(∆{1n−a,a}) is generated in degree ≤ 4, as shown by Weyman [22]. The case a = ⌊n/2⌋ + 1 is of special interest: here ∆{1n−a,a} is the nullcone, i.e. the variety defined set-theoretically by all SL2-invariants of binary n-ics. Note that the nullcone is self-dual when n is even (by Corollary 2.7), and its ideal is generated by quartics when n is odd. Chipalkatti [5, Conjecture 6.1] had conjectured that I(∆λ) is generated in degree ≤ 4 whenever λ has only d = 2 parts, and this was proved by Abdesselam and Chipalkatti in [2, Proposition 20]. For further details on their approach see [1, Section 7]. The papers [1, 2, 5, 22] stress the fact that all our ideals are invariant under the action of SL2, and hence each of their graded components are direct sums of irreducible SL2-modules. Chipalkatti describes the minimal free resolutions of I(∆λ) in terms of SL2-modules for λ = (3, 2) in [5, §4.1] and for λ = (3, 3) in [5, §4.2]. He discusses the 364 sextics for λ = (3, 3, 2) in [5, §3.1]. The ideal I((∆λ)∨) of the dual variety has been studied only in two cases, namely when λ is a hook (by self-duality) and when λ has only parts 1 and 2. In the latter case, it is generated by minors of a Hankel matrix. In all other cases, little seems to be known. Of course, I((∆λ)∨) will often be principal (when m1 = 0), and we seek to find the generator of such an ideal explicitly. This is accomplished for some interesting cases in Section 4. 4 Real Rank Boundaries In this section we present an application of our duality theory to the study of real ranks of binary forms [3, 6]. Let f ∈ R[x, y]n be a general binary form of degree n with real coefficients. The complex rank of f equals r = ⌈(n + 1)/2⌉. This means that f is a sum of r powers of linear forms over C, but not fewer. We consider the set Rn of all real binary forms f that admit a rank r decomposition also over R. In other words, Rn is the set of all forms f ∈ R[x, y]n such that both the real rank of f and the complex rank of f are equal to r. The set Rn is a full-dimensional semi-algebraic set inside R[x, y]n. Its topological boundary ∂Rn is the set-theoretic difference of the closure of Rn minus the interior of the closure of Rn. Thus, if f ∈ ∂Rn then every open neighborhood of f contains a generic form of real rank equal to r and also a generic form of real rank bigger than r. The topological boundary ∂Rn is a semi-algebraic subset of pure codimension one inside the real affine space R[x, y]n. We define the real rank boundary, denoted ∂alg(Rn), to be the Zariski closure of the topological boundary ∂Rn. We view ∂alg(Rn) as a closed subvariety in the complex projective space P(C[x, y]n) = Pn. It is pure codimension one, so it is defined by a unique (up to scaling) squarefree polynomial in the coordinates a0, a1, . . . , an on Pn. We compute that polynomial: Theorem 4.1. Let n ≥ 5. If n = 2k − 1 is odd then the real rank boundary ∂alg(Rn) is an irreducible hypersurface of degree 2k(k − 1) in Pn. This hypersurface is the dual (∆λ)∨ of the 10 multiple root locus ∆λ where λ = {2k−2, 3}. Its defining polynomial is the discriminant of q(u, v) = det . (14)   uk a0 a1 a2 ... ak−1 uk−1v · · · uvk−1 ak−1 a1 ak a2 ak+1 a3 ... ... an−1 ak · · · · · · · · · . . . · · · vk ak ak+1 ak+2 ... an   If n = 2k is even then the real rank boundary ∂alg(Rn) is the union of the two irreducible hypersurfaces (∆λ)∨ where λ is {2k−3, 32} or {2k−2, 4}. Their degrees are 2k(k − 1)(k − 2) and 3k(k − 1). These arise as irreducible factors of the Hurwitz form of the discriminant ∆{1k−1,2} when evaluated at the line in P(R[x, y]k+1) that is the kernel of the k×(k+2)-matrix a0 a1 a1 a2 ... ... ak−1 ak   ak · · · ak+1 · · · ak+1 ak+2 ... ... . . . an · · · an−1   . (15) A third irreducible factor appearing in this specialized Hurwitz form is the Hankel determinant (∆{2k})∨, which has degree k + 1, but this is not part of the real rank boundary ∂alg(Rn). For the relevant background on Chow forms and Hurwitz forms we refer to [10] and [21]. Before presenting the proof of Theorem 4.1, here is an illustration of the first three cases. Example 4.2 (Little Apple and Big Apple). Let n = 5, so k = 3 and λ = (3, 2). In this first case, Theorem 4.1 was proved by Comon and Ottaviani in [6, §5]. The polynomial defining the hypersurface (∆λ)∨ ⊂ P5 is their apple invariant I12. It has 228 terms of degree 12. The next odd case is n = 7, so k = 4 and λ = (3, 2, 2). The real rank boundary (∆λ)∨ is a hypersurface of degree 24 in P7, namely the join of the surface ∆(6,1) and the threefold σ2(∆(7)) = σ2(ν7(P1)). Its defining polynomial is computed via (14). It has 38082 terms. ♦ Example 4.3 (Chow and Hurwitz forms). Let n = 6. We consider the discriminant ∆(2,1,1) of binary forms of degree k + 1 = 4. This discriminant is a threefold of degree 6 in P4. Its singular locus consists of the surfaces ∆(3,1) and ∆(2,2). The Hurwitz form of ∆(2,1,1) is a polynomial of degree 30 = 3· 6 + 2· 4 + 4 in the ten Plucker coordinates on the Grassmannian of lines in P4. It is reducible because ∆(2,1,1) is singular in codimension 1. It factors as Hur(∆(2,1,1)) = Chow(∆(3,1))3 · Chow(∆(2,2))2 · Tan. (16) Here Chow(X) denotes the Chow form of a surface X in P4, i.e. the polynomial in Plucker coordinates that vanishes when the line intersects X. The degree of the Chow form equals the degree of X, so it is 6 and 4 in our two cases. The last factor Tan is the condition for the line to be tangent at a smooth point of ∆(2,1,1). This is a quartic in the Plucker coordinates. 11 We now take our line in P4 to be the kernel of the 3× 5 matrix in (15). More precisely, in the formula (16) we substitute the 10 maximal minors of (15) for the 10 Plucker coordinates. Here signs have to be taken into consideration carefully. Each maximal minor is a cubic, so the degrees above have to be tripled. Equation (16) now has degree 90 = 3· 18 + 2· 12 + 3· 4. The polynomial obtained from Chow(∆(3,1)) is (∆(4,2))∨. It has 3140 terms of degree 18. The polynomial obtained from Chow(∆(2,2)) equals (∆(3,3))∨. It has 560 terms of degree 12. The degree 12 polynomial obtained from Tan is the third power of the Hankel determinant (∆(2,2,2))∨ = det  a0 a1 a2 a3 a1 a2 a3 a4 a2 a3 a4 a5 a3 a4 a5 a6   . (17) To illustrate the statement of Theorem 4.1, we present two explicit forms that lie in ∂R6. The binary form f = y6 + 15x4y2 lies in ∂R6∩ (∆(4,2))∨ but not in (∆(3,3))∨∪ (∆(2,2,2))∨. Its line of apolar quartics is Lf =(cid:8)s· uv3 + t· (u4− v4) (s : t) ∈ P1(cid:9). This has discriminant (27s4 + 256t4)t2, with only real root at (s : t) = (1 : 0). That quartic lies in κ = ∆(3,1), and it is a limit of quartics with four real roots, and also a limit of quartics with two real roots. From this we can construct generic sextics f±ǫ close to f whose real ranks are 4 and 5. The binary form f = y6 +5x2y4−5x4y2−x6 lies in ∂R6∩(∆(3,3))∨ but not in (∆(4,2))∨∪ (∆(2,2,2))∨. Its line of apolar quartics is Lf =(cid:8)s·(u−v)2(u+v)2 + t·uv(u2+v2) (s : t) ∈ P1(cid:9). This has discriminant (16s2 + t2)2t2, with only real root (s : t) = (1 : 0). That quartic lies in ν = ∆(2,2). We can construct nearby generic sextics f±ǫ whose real ranks are 4 and 5. The proof below will explain the relevance of Lf , κ, and ν for ∂R6. It will also show why the Hankel determinant (∆(2,2,2))∨ meets the boundary of R6 only in lower dimension. ♦ Remark 4.4. It is instructive to see the factorization (16) using iterative discriminants. Let A(x) and B(x) be univariate quartics, and consider A(x)+tB(x) where t is an unknown. Then discrt(cid:0)discrx(A(x) + tB(x))(cid:1) = (cid:0) [6, 6]A,B(cid:1)3 ·(cid:0) [4, 4]A,B(cid:1)2 ·(cid:0) [4, 4]A,B(cid:1), where [e, e]A,B stands for a big expression that has degree e in the coefficients of A and of B. Proof of Theorem 4.1. We write f for a generic form in R[x, y]n. We first consider the odd annihilates f (x, y), is generated by forms q(u, v) of degree k and r(u, v) of degree k + 1. Furthermore, the dual form q(u, v) has the Hankel determinantal representation in (14). case n = 2k − 1. The apolar ideal, consisting of all forms in R[u, v] such that g(cid:0) ∂ Since f is generic, there exists a unique decomposition f (x, y) =Pk i=1(six + tiy)n. The points (si : ti) are the complex roots of q. Hence f lies in Rn precisely when q is real-rooted. By real-rooted we mean that q is square-free and its roots are all real. Suppose now that the form f moves and passes through the boundary of Rn. Then two real roots of q(u, v) merge and become a double root. At this point, the discriminant of q(u, v) vanishes. Corollary 2.3 implies that this discriminant equals (∆λ)∨ where λ = {2k−2, 3}. Next consider the even case n = 2k. The apolar ideal of f is generated by two forms of degree k + 1. Let Lf be the line in Pk+1 spanned by these two forms. The points q on Lf ∂y(cid:1) ∂x, ∂ 12 correspond to the distinct decompositions f (x, y) =Pk+1 i=1 (six + tiy)n. Namely, the (si : ti) are the roots of q. This means that f lies in Rn if and only if some q in Lf is real-rooted. The set of all real-rooted forms is a connected full-dimensional semi-algebraic subset of Pk+1 R , and its algebraic boundary is the discriminant δ = ∆{1k−1,2}. Note that δ is a hypersurface of degree 2k. Its singular locus consists of the codimension two loci κ = ∆{1k−2,3} and ν = ∆{1k−3,22}. Their degrees are deg(κ) = 3(k − 1) and deg(ν) = 2(k − 1)(k − 2). Suppose that the form f moves along a general curve. Consider its image under the map f 7→ Lf into the Grassmannian of lines in Pk+1 R , here denoted Gr. As f crosses the boundary of Rn, the line Lf transitions from intersecting to not intersecting the subset of real-rooted forms. At the transition point, the form f is in ∂alg(Rn). One of the following three scenarios might happen: (i) Lf intersects κ, or (ii) Lf intersects ν, or (iii) Lf is tangent to δ at a smooth point. Each of these conditions defines an irreducible hypersurface in the Grassmannian Gr. Their equations are the irreducible factors in the Hurwitz form Hur(δ). The Hurwitz form of the discriminant factors as follows in the coordinate ring of Gr: Hur(δ) = Chow(κ)3 · Chow(ν)2 · Tan(δ). (18) The exponents 3 and 2 arise from the classical Plucker formula for the dual of a plane curve. Now, since δ is a hypersurface of degree 2k, its Hurwitz form has degree 2k(2k − 1). Hence 2k(2k−1) = 3deg(κ) + 2deg(ν) + deg(cid:0)Tan(δ)(cid:1) = 9(k−1) + 4(k−1)(k−2) + deg(cid:0)Tan(δ)(cid:1). This uses the fact that the degree of a Chow form in Plucker coordinates equals the degree of its variety. We conclude that Tan(δ) is a polynomial of degree k + 1 in Plucker coordinates. The map Pn 99K Gr, f 7→ Lf is birational. Indeed, every generic line has the form Lf for a unique sextic f that is recovered by solving the differential equations represented by Lf . Moreover, the base locus of this inverse map is precisely the tangential hypersurface Tan(δ). We now examine what happens to the three irreducible factors in (18) when pulled back under the map Pn 99K Gr. Algebraically, the line Lf is given as the kernel of the matrix (15). So, to compute the pullback of (18), we need to replace the Plucker coordinates by the (appropriately signed and scaled) maximal minors of (15). These minors are polynomials of degree k, so we must multiply the above degrees by k. This gives the degrees 3k(k − 1), 2k(k − 1)(k − 2) and k(k + 1) for the pullbacks of the three irreducible factors in (18). The first two pullbacks are irreducible as polynomials in R[a] = R[a0, . . . , an]. They are Chow(κ)(Lf ) = (∆{2k−2,4})∨ and Chow(ν)(Lf ) = (∆{2k−3,32})∨. (19) As before, this follows from the description of the dual hypersurfaces in Corollary 2.3. The third factor in (18) becomes a reducible polynomial in R[a]. Namely, we have In words: the pullback of Tan is the kth power of the Hankel determinant of order k + 1. Tan(Lf ) = (cid:0) (∆{2k})∨(cid:1)k . (20) This can be seen as follows. A general point f on the Hankel hypersurface satisfies rankC(f ) = k. It is annihilated by some binary form g(u, v) of degree k, and the line Lf 13 is spanned by u · g(u, v) and v · g(u, v). This means that Lf is tangent to δ at k points over C, given by the k roots of g(u, v). We see that the pullback Tan(Lf ) contains (∆{2k})∨ set-theoretically. Both are irreducible varieties, and hence they are equal as sets in Pn. By comparing degrees, we conclude that (20) holds. Our argument also shows the following fact: if a line of the form Lf is tangent to the discriminant δ at one smooth point then it is tangent to δ at a scheme of length k. We have proved that ∂alg(Rn) has at most three irreducible factors. However, the correct number is two. The two hypersurfaces that appear in ∂alg(Rn) are those in (19). This was shown already for n = 6. Towards the end of Example 4.3, we exhibited one binary form f for each of these two boundary strata. In general, a construction can be made as follows. We consider lines in the space Pk+1 of binary forms of degree k + 1 that look like Lκ = (cid:8)s · uv3f (u, v) + t · (u2 + v2)g(u, v) (s : t) ∈ P1(cid:9) Lν = (cid:8)s · (u − v)2(u + v)2f (u, v) + t · (u2 + v2)g(u, v) (s : t) ∈ P1(cid:9), and where f and g are generic real-rooted binary forms of degrees k − 3 and k − 1 respectively. The line Lκ meets the discriminant δ only in the cusp locus, and the line Lν meets δ only in the node locus. The binary form has k − 1 distinct real roots at these intersection points. Now, for suitable choices of f and g, the discriminant of the pencil of binary forms has no real roots other than (s : t) = (1 : 0). We fix such f and g, and we set L = Lκ or L = Lν. The pencil L consists of binary forms of degree k + 1 that have precisely k − 1 real roots. Moreover, L is not contained in the hypersurface Tan. Recall that the inverse to Pn 99K Gr is well-defined in a neighborbood U of L. Hence, for each U ∈ U there exists a unique form fU of degree n such that LfU = U. By construction, there exist generic points U±ǫ in U such that U+ǫ contains a real-rooted form, and U−ǫ contains no real-rooted form. Then f+ǫ = fU+ǫ has real rank k + 1 while f−ǫ = fU−ǫ has real rank ≥ k + 2. We conclude that f is in ∂Rn. It remains to be seen that the Hankel determinant ∆{2k} is not a factor of the real rank boundary ∂alg(Rn). The proof is by contradiction. Suppose that f ∈ ∂Rn has rankC(f ) = k, and write f = Pk i . The ℓi are linear forms over C, and these are unique. We can approximate f by a sequence of forms in Rn. Each is a sum of k + 1 powers of real linear forms. A convergence argument shows that ℓ1, . . . , ℓk must have real coefficients as well. 0 where ǫ > 0 is small. The square Hankel matrices corresponding to these two forms are invertible. Their determinants have opposite signs. This can be seen from the Matrix Determinant Lemma. The two forms lie on different sides of the Hankel hypersurface. Both have real rank k + 1, and hence both lie in Rn. This contradicts our assumption that f is in the topological boundary of Rn. Fix a generic real linear form ℓ0 and consider f +ǫℓn 0 and f −ǫℓn i=1 ℓn Remark 4.5. The real projective space Pn R of binary forms of degree n is stratified by real rank. Let r = ⌈(n + 1)/2⌉. Blekherman [3] showed that each integer in {r, r + 1, . . . , n− 1, n} arises as the real rank of some open stratum. It would be very interesting to determine the algebraic boundary that separates real rank i from real rank i + 1, for any i ∈ {r, . . . , n− 1}. Currently, only the two extreme cases are known. The algebraic boundary for i = n − 1 is the discriminant ∆{1n−2,2}, by [6, Prop. 3.1], and the case i = r is resolved by Theorem 4.1. 14 5 Euclidean Distance Degrees This section is concerned with the following optimization problem over the real field K = R. Let h ∈ R[x, y]n be a fixed binary form. We seek f ∈ ∆λ that is closest to h. In symbols: minimize hf − h, f − hi subject to f ∈ ∆λ. (21) We call this the special ED problem when the inner product is as in (3) and (4). Alternatively, we may also use a positive definite quadratic form that is generic. The resulting scenario (21) i=0 γiaibi, is the generic ED problem for ∆λ. For instance, if we replace (4) with hf, gi = Pn where γ0, γ1, . . . , γn are random positive reals, then this leads to a generic ED problem. The map that takes the given input h to the optimal solution f ∗ = f ∗(h) of the problem (21) is an algebraic function. The degree of that algebraic function is known as the ED degree of the variety ∆λ. For an introduction to this topic and its basic results see [8]. A follow-up study, aimed at varieties that admit a determinantal representation, was undertaken in [20]. In the algebraic approach to solving (21) one writes the critical equations using Lagrange multipliers and one removes the singular locus. We shall see concrete examples for ∆λ shortly. The ED degree is the number of complex solutions to these equations for generic data h. Of course, the optimal point f ∗ is real and it is among these complex critical points. Varieties such as ∆λ come with a natural invariant coordinate system, and our special inner product (3)-(4) gives the standard Euclidean distance for that coordinate system. The corresponding ED degree is the special ED degree of ∆λ. By contrast, the generic ED degree of ∆λ is usually larger. This is the degree for the generic ED problem described above. In the last column of Table 1 we list the special ED degree and the generic ED degree. The first pair in each box, corresponding to the rational normal curve ∆(n), equals n, 3n− 2. These numbers were derived, for arbitrary n, in [8, Example 5.12] and [8, Corollary 8.7]. Both notions of ED degrees are preserved under duality [8, Theorem 5.2], and the generic ED degree coincides with the sum of the polar classes, Cλ(1, 1) =P δi, by [8, Theorem 5.4]. Some of the ED degrees in Table 1 appear in [20]. For instance, the ED degrees 7, 13 for λ = (2, 2) appear in [20, Example 1.1]. When λ is dual to a collections of hooks of the form n, n or n, n, n, the variety (∆λ)∨ is given by Hankel matrices of rank 2 or 3. The corresponding ED degrees in Table 1 are found in [20, Table 4]. Therein, the left table for Λ = Ωn gives generic ED degree, while the right table for Λ = Θn gives special ED degree. From Table 1 we can guess the two ED degrees for the variety of binary forms of degree n that have a root of multiplicity a. Our next goal is to prove that this guess is correct. Theorem 5.1. For any hook λ = {1n−a, a}, the special ED degree of the variety ∆λ is always n, independently of a, whereas the generic ED degree of ∆λ equals (2a − 1)n − 2(a − 1)2. Proof. We begin with the generic ED degree. It is the sum of the polar classes δi. The codimension a − 1 of ∆{1n−a,a} equals the dimension of its dual (∆{1n−a,a})∨ = ∆{1a−2,n−a+2}. This means that precisely two polar classes are nonzero, and the generic ED degree equals δa−1 + δa = deg(∆{1n−a,a}) + deg(∆{1a−2,n−a+2}) = a(n − a + 1) + (n − a + 2)(a − 1), (22) by Hilbert’s formula (1). This expression equals (2a − 1)n − 2(a − 1)2, as desired. 15 Next consider the special ED degree. Our goal is to compute the complex critical points f ∗ of the optimization problem (21) where h is a fixed generic binary form in V = R[x, y]n. We now identify V with its dual space V ∨ by way of the distinguished inner product (3), and we regard the conormal variety Conλ as an affine cone of dimension n + 1 in V × V = R2n+2. By ED duality [8, §5], our problem is equivalent to solving the system of linear equations (23) f + g = h for (f, g) ∈ Conλ. These n + 1 inhomogeneous linear equations have finitely many complex solutions (f ∗, g∗) on the (n + 1)-dimensional affine variety Conλ. Their number is the special ED degree. We now write the equations (23) using the parametrization of Conλ given in Theorem 2.5: h(x, y) = (tx − sy)a · g1(x, y) + (sx + ty)n−a+2 · g2(x, y). (24) The two summand are unknown points in ∆{1n−a,a} and in ∆{1a−2,n−a+2}. Both are now regarded as affine varieties in V = Cn+1. The forms g1 and g2 are unknown and they have degrees n − a and a − 2 respectively. Furthermore, (s : t) is an unknown point in P1. Hence there are n + 1 unknown parameters in total, to match the n + 1 given coefficients of h(x, y). Thus (24) is a square polynomial system in Cn+1, and we need to count its solutions. Any representation (24) over C of the given binary form h(x, y) will be called an orthogo- nal decomposition of type a. Thus, our proof reduces to establishing the following assertion: a general binary form of degree n has precisely n orthogonal decompositions of type a. In what follows we describe the binary form whose zeros are the points (s : t) that can occur in (24). In other words, we eliminate the unknown binary forms g1 and g2 from (24). Once (s : t) is known, the coefficients of g1 and g2 can be recovered by solving a linear system of equations. So, it suffices to identify that binary form and to show it has degree n. We define an endomorphism L(k) of the vector space V = R[x, y]n as follows: (cid:0)L(k)(f )(cid:1)(x, y) = (n − k)! n! k Xi=0 (−1)i(cid:18)k i(cid:19)xk−iyi ∂k ∂xi∂yk−i f (x, y). This linear differential operator generalizes the SO2-invariant vector field and the second order operator L(1) = 1 n(cid:20)x ∂ ∂y − y ∂ ∂x(cid:21) L(2) = 1 n(n − 1)(cid:20)x2 ∂2 ∂y2 − 2xy ∂2 ∂x∂y i=0(cid:0)n The linear map L(k) : V → V can be written explicitly in terms of coordinates as follows. If f =Pn i(cid:1)aixiyn−i then the coefficients of L(k)(f ) = Pn i(cid:19)(cid:18) n − k i + j − k(cid:19)a2i+j−k. j(cid:1)bjxjyn−j are j(cid:19)bj = (cid:18)n n−j Xi=max{0,k−j} (−1)i(cid:18)k 16 + y2 ∂2 ∂x2(cid:21). j=0(cid:0)n (25) If we apply the n-th order operator then this amounts to a rotation by 90 degrees: (cid:0)L(n)(f )(cid:1)(x, y) = f (−y, x). Now, the relevance of the differential operator L(k) for our proof is as follows: Lemma 5.2. Let h be a binary form of degree n. Suppose that (s : t) ∈ P1 occurs in an orthogonal decomposition (24) of type a. Then (s : t) is a root of the binary form L(a−1)(h). Lemma 5.2 will be proved further below. We first derive the theorem from the lemma. The linear map L(a−1) is not zero. Hence, for generic h, the binary form L(a−1)(h) is nonzero and has degree n. Such a binary form has at most n distinct roots. Therefore, by Lemma 5.2, we know that the special ED degree is at most n. What we must prove is that the special ED degree is at least n. We do this by exhibiting, for every n and a, one particular binary form of degree n that has n distinct orthogonal decompositions of type a. We begin with the observation that the following two binary forms have real coefficients: (26) (27) hn(x, y) = kn(x, y) = 1 2(cid:0)(x + √−1 · y)n + (x − √−1 · y)n(cid:1), 2√−1(cid:0)(x + √−1 · y)n + (x − √−1 · y)n(cid:1). 1 These two forms are invariant under rotation by 2π/n, i.e. they are fixed by the endomor- phism ρ : V → V that maps x to (cos 2π n )y and y to (sin 2π Case 1. Suppose that n is odd. For our special binary form we take n )x − (sin 2π n )x + (cos 2π n )y. Note that all the exponents of x are odd while that of y are even. For odd a we have hn(x, y) = xn −(cid:0)n 2(cid:1)xn−2y2 +(cid:0)n 4(cid:1)xn−4y4 − · · · ±(cid:0) n n−1(cid:1)xyn−1. h = xag1 + yn−a+2g2, (28) for suitable forms g1 and g2 depending on a. Similarly, for even a we have h = yag1 + xn−a+2g2. For each decomposition, we obtain n − 1 others by acting with the rotations ρ, ρ2, . . . , ρn−1. Case 2. Suppose that n even and a is even. We also take hn(x, y) = xn −(cid:0)n 2(cid:1)xn−2y2 +(cid:0)n 4(cid:1)xn−4y4 − · · · ± yn. Then (28) also holds. Now, hn is fixed by ρn/2, so applying the rotations ρi gives only n/2 distinct orthogonal decompositions of type a. However, we have hn(x, y) = ±hn(y, x), so by permuting the two variables we get additional decompositions. These are not equal to any of the previous ones. In total, this yields n 2 · 2 = n distinct decompositions (24) for hn. Case 3. Suppose that n even and a is odd. In that case we take kn(x, y) = (cid:0)n 1(cid:1)xn−1y −(cid:0)n 3(cid:1)xn−3y3 +(cid:0)n 5(cid:1)xn−5y5 − · · · ±(cid:0) n n−1(cid:1)xyn−1, and the argument is the same as in Case 2. This completes the proof of Theorem 5.1. 17 We now turn to Lemma 5.2. Note that this result is familiar in the case a = 2, when (21) asks for the best rank 1 approximation of a symmetric 2×2×· · ·×2 tensor h. Finding that approximation amounts to computing the eigenvectors of h; see [8, Corollary 8.7]. However, ∂h/∂x ∂h/∂y(cid:19). by definition, the eigenvectors of h are the roots of L(1)(h) = (1/n) · det(cid:18) x y Proof of Lemma 5.2. We claim that the operator L(k) satisfies the following identity L(k)(cid:0)f (x, y)(cid:1) = (−1)n−k · L(n−k)(cid:0)f (−y, x)(cid:1) for k = 0, 1, . . . , n. (29) The special case k = n is (26). One can show that (29) holds by a direct computation, using the formula (25). Equivalently, we check that it holds for monomials and extend by linearity. Suppose now that (s : t) ∈ P1 occurs in an orthogonal decomposition (24) of type a. We apply the differential operator L(a−1) to both sides of that equation. This implies L(a−1)(cid:0)h(x, y)(cid:1) = L(a−1)(cid:0) (tx − sy)ag1(x, y)(cid:1) + L(a−1)(cid:0) (sx + ty)n−a+2g2(x, y)(cid:1). Applying (29) to the summand on the right, we conclude L(a−1)(cid:0)h(x, y)(cid:1) = L(a−1)(cid:0) (tx−sy)ag1(x, y)(cid:1) + (−1)n−a+1L(n−a+1)(cid:0) (tx−sy)n−a+2g2(−y, x)(cid:1). In both summands, a k-th order differential operator L(k) is applied to a binary form that has (s : t) as a root of multiplicity at least k + 1. Both of the resulting binary forms still have (s : t) among their roots. This shows that L(a−1)(cid:0)h(x, y)(cid:1) vanishes at (s : t). The difference between the two ED degrees in Theorem 5.1 is 2(n − a + 1)(a − 1). We shall explain this number and how to think about the operator L(a−1). If we fix values for the parameters s and t, then the equation (24) translates into an inhomogeneous linear system of equations whose unknowns are the coefficients of g1 and g2. We have n+1 linear equations in n = (n − a + 1) + (a − 1) unknowns. There is no solution for generic s and t. We write our inhomogeneous linear system as an (n + 1) × (n + 1)-matrix Mn,a. The first row consists of the coefficients of h. The next n − a + 1 rows contain the monomials of degree a in (s, t), and the last a− 1 rows contain the monomials of degree n− a + 2 in (s, t). We seek row vectors of the form (−1, g1, g2) that lie in the left kernel of Mn,a. The matrix has a banded structure, like Sylvester’s matrix for the resultant. For instance, for λ = (3, 1), h1 h0 h2 h3 −s3 3s2t −3st2 t3 3s2t −3st2 0 −s3 3s2t t3 3st2 s3 3s2t 3st2 t3 0 h4 0 t3 0 s3   .   In order for our linear equations to have a solution, the determinant of Mn,a must be zero. The degree of det(Mn,a) in (s, t) is precisely the generic ED degree. This is best seen from (22). If we replace (3) by a generic inner product then the rows for g1 change by a linear M4,3 = 18 transformation, whilst the rows for g2 change by the inverse linear transformation. Thus, for the generic ED problem, the critical points are precisely the roots of the determinant. However, for the special ED problem, our determinant admits the following factorization: det(cid:0)Mn,a(cid:1) = ±(cid:0)L(a−1)(h)(cid:1)(s, t) · (s2 + t2)(n−a+1)(a−1). The binary form s2 + t2 vanishes if and only if the last n rows of Mn,a are linearly dependent. The degree of the extraneous factor is 2(n − a + 1)(a − 1), and subtracting that from the generic ED degree gives n. The remaining factor is the remarkable binary form L(a−1)(h). We close this section with a conjecture, namely that the two ED degrees of ∆λ always have the same parity. According to Table 1, this holds for all partitions with n ≤ 7. For hooks, it is proved by Theorem 5.1, and the underlying reason is seen clearly in (30). In general, the extraneous components should come from isotropic quadrics like s2 + t2, and parallelities like sv− tu. These quadrics suggest that the difference in ED degrees is even for all λ. This conjecture is related to [20, eqn. (3.5)]. At present we do not know how prove it. (30) 6 ED Duality in Action We now illustrate how our results can be applied to find exact solutions to the optimization problem (21). Following [8, 20], our approach is to compute all critical points and then select the best real critical point. By ED duality [8, §5], the critical points are found by solving linear equations on the conormal variety. Given h, we need to compute all decompositions h(x, y) = f (x, y) + g(x, y) where (f, g) ∈ Conλ. (31) If h is generic then the number of such decompositions is the special ED degree. The binary forms f that arise in the decompositions (31) are precisely the critical points on ∆λ for the Euclidean distance to h, and similarly the forms g are the critical points on its dual (∆λ)∨. The proximity of the solution is reversed under duality because h− f2 +h− g2 = h2. This follows from h = f + g and hf, gi = 0. In particular, if f is the closest point to h among those on ∆λ, then it is paired with the farthest critical point g on (∆λ)∨, and vice versa. In what follows we illustrate how one might solve (31) and hence (21) in practice. We begin with n = 5 and λ = (3, 1, 1). For our given data point we take the binary quintic h(x, y) = 1 2(cid:0)x + √−1 · y(cid:1)5 + 1 2(cid:0)x − √−1 · y(cid:1)5 This is a slight variant of (27). The primal problem is to find the closest quintic f ∈ ∆(3,1,1) with a triple root, and the dual problem is to find the closest quintic g ∈ ∆(4,1) with a quadruple root. By Theorem 5.1, the equation (31) has five solutions on Conλ. They are + y5 = x5 − 10x3y2 + 5xy4 + y5. (32) 19 f = h − g g2 (1.471x2 − 5.582xy + 3.585y2)(x + 0.785y)3 3.02 (−0.263x2 − 0.211xy − 0.338y2)(x − 3.132y)3 4.92 5.09 (−0.263x2 + 0.167xy + 0.0346y2)(x + 3.020y)3 6 (1.473x2 + 5.397xy + 2.867y2)(x − 0.6732y)3 8.12 (x2 − 10y2)x3 g = h − f (−1.238x − 0.735y)(0.785x − y)4 (0.0131x − 0.403y)(3.132x + y)4 (0.0152x + 0.468y)(3.020x − y)4 (−2.301x + 1.875y)(0.6732x + y)4 (5x + y)y4 f2 13.98 12.08 11.91 11 8.88 The upper left quintic f in ∆(3,1,1) is closest to h, at distance 3.02156668059971216331/2, with triple root (−0.78519451639408253233 : 1). The lower right quintic g in ∆(4,1) is closest to h, at distance 8.88082776145888597831/2, with quadruple root (−1 : 0.67321557299682647408). We could have guessed the decomposition h = f + g in the fourth row from the input (32). This one is indeed a critical point, but it is neither primal optimal nor dual optimal. Our five solutions to (31) were found by using the matrix that was introduced in Section 5: M5,3(s, t) = 0 5 −10 1 0 −s3 3s2t −3st2 t3 0 3s2t −3st2 0 −s3 t3 3s2t −3st2 −s3 0 0 s4 4s3t 6s2t2 4st3 t4 4s3t 6s2t2 4st3 t4 0 1 0 0 t3 0 s4   .   The triple (resp. quadruple) roots (s : t) ∈ P1 of f (resp. g) are the roots of the binary form det(cid:0)M5,3(s, t)(cid:1) = (s2 + t2)6 · (s5 − 10s3t2 − s2t3 + 5st4) = −(s2 + t2)6 ·(cid:0)L(2)(h)(cid:1)(s, t). We compute the five real roots numerically, and at each of them we compute the left kernel of M5,3(s, t). The result of that computation is precisely the list of five pairs (f, g) above. This method scales well for hooks λ = {1n−a, a}. Here, the special ED degree is always n, and Lemma 5.2 furnishes the minimal polynomial for the desired a-fold root of f ∈ ∆λ. The matrix Mn,a(s, t) represents our task as a homogeneous polynomial eigenvalue problem, and this makes it amenable to well-developed methods of numerical linear algebra; see e.g. [7]. For an illustration we fix n = 15 and a = 6. The data point is the binary form h(x, y) = P15 i=0(cid:0)15 i(cid:1)uixiy15−i, with randomly chosen coefficient vector (u0, . . . , u15) = (20,−17, 3, 16, 12, 14,−16,−5, 7, 8,−13, 5,−13,−16, 7,−11). Among forms f of degree 15 with a root of multiplicity 6, the following is the closest to h: P15 i=0(cid:0)15 i(cid:1)vixiy15−i = 10−6(x − 8.70886y)6(131903x9+11375.9x8y − ··· − 552.901xy8+45.8419y9), (v0, v1, . . . , v15) = (20,−17, 3, 16, 12, 14,−16,−5, 7.04, 8.19, −12.17, 8.09, −3.78, 1.42, −0.46, 0.13). 20 All 15 critical points for this optimization problem are real because L(5)(h) is real-rooted for our h. Its 15 roots have the form (s : 1) where s ∈ R. They appear in the first column of root s of L(5)(h) distance2 8.70886 3.70567 2.19850 0.05736 −0.38870 −3.49092 −5.71229 −0.22118 1.25359 0.25811 0.48187 0.80694 −0.68808 −1.67383 −1.06515 ? ? min min local min 86791 111796 min 163470 max 476068 550056 564363 565936 657621 723240 max 727831 774941 934884 max 1058800 max 1150260 max 1256200 max ? ? ? By computing the left kernel of M15,6(s, t) at each root, we find the 15 decompositions (31). The squared norms g2 of the dual solutions g = h − f are listed in the second column. So, the first row gives the optimal solution for λ = {19, 6}, while the last row gives the optimal solution for the dual problem µ = {14, 11}. Local optima are indicated in the third column. There are four local optima on ∆λ (marked with “min”) and six local optima on ∆µ (marked with “max”). These were certified using the signature of the Hessian of the distance function. The local nature of the other five points cannot be decided with the second-order criterion because their Hessians are singular, in both the primal and the dual formulation. Our optimization problem is more challenging when λ is not a hook. A small interesting case is the partition λ = (3, 2). Let us see what happens here if we take the same input h as in (32). We seek a binary form f ∈ ∆(3,2) with a triple root and a double root that is closest to h. The desired decomposition (31) on the conormal variety Con(3,2) now takes the form h(x, y) = α(tx − sy)3(vx − uy)2 + (βx + γy)(sx + ty)4 + δ(ux + vy)5. This means that the vector (−1, α, β, γ, δ) lies in the left kernel of the 5 × 6-matrix   1 −s3u2 2s3uv+3s2tu2 −6s2tuv−s3v2−3st2u2 3s2tv2+6st2uv+t3u2 −3st2v2−2t3uv t3v2 s4 0 u5 −10 6s2t2 4s3t 10u3v2 4st3 6s2t2 10u2v3 4s3t s4 5u4v t4 4st3 5uv4 5 1 0 t4 v5 .   0 0 All 5 × 5 minors of this matrix must be 0. However, the ideal of 5 × 5 minors has some extraneous associated primes that must be removed. This is analogous to the factor s2 + t2 in the hook case, but more complicated, so we do not pursue that primary decomposition. 21 Instead we simply work directly with the squared Euclidean distance function and we solve the following system of polynomial equations in five unknowns: D = k h − α(tx − sy)3(vx − uy)2 k2, ∂D ∂α = ∂D ∂s = ∂D ∂t = ∂D ∂u = ∂D ∂v = 0 and sv − ut 6= 0. This has 20 complex solutions, while the ED degree of ∆(3,2) is 21. One checks that h lies in the ED discriminant [8, §7]. Only 4 of the 20 solutions are real. Setting t = v = 1, they are α s u 1.817238 −0.266252 −3.020572 −0.265424 1.815280 −0.785143 0.673272 −1.316853 0.274673 3.131909 −0.387044 1.428712 distance2 7.724678 8.643701 12.017703 13.105926 The worst critical point is given in the bottom row. The corresponding quintic is f = 1.81528x5 − 0.911262x4y − 5.15521x3y2 + 0.0137459x2y3 + 4.34203xy4 + 1.79341y5. ED duality gives us the optimal solution g = h− f for the dual problem on (∆(3,2))∨, namely g = −0.81528x5 + 0.911262x4y − 4.84479x3y2 − 0.0137459x2y3 + 0.657973xy4 − 0.793414y5 (βx + γy)(−0.785143x + y)4 + δ(1.428712x + y)5 = for suitable real constants β, γ, and δ, which can be found by solving a linear system. The form g is very interesting for the application described in Section 4. Recall that the generic tensor rank for binary quintics is 3. The unique rank 3 decomposition of the given quintic h is complex. It is shown on the left in (32). One might therefore ask for a best approximation to h that has real rank 3. This question is not well-posed because the set R5 from Section 4 is not closed. Instead one should ask for the closest binary quintic in the closure of R5, i.e. among those whose border rank equals 3. That optimal quintic must be our g, because it is a critical point on real rank boundary ∂alg(R5), which equals the little apple hypersurface (∆(3,2))∨. The fact that g has border rank 3 is verified by the representation (−0.785143x + y)5 + δ(1.428712x + y)5(cid:19). ((−0.785143 + ǫβ)x + (1 + ǫγ)y)5 − g = lim ǫ→0(cid:18) 1 5ǫ 1 5ǫ This computation offers a concrete illustration of Theorem 4.1 and Example 4.2. Acknowledgements. We thank Michael Burr for a conversation in March 2015 that ignited our interest in the ED problem for multiple root loci. We are grateful to Greg Blekherman for his help in getting Theorem 4.1 into final shape. This project was carried out in the summer of 2015, when both authors visited the National Institute of Mathematical Sciences, Daejeon, Korea. Hwangrae Lee was supported by the National Research Foundation of Korea (MSIP 2011-0030044). Bernd Sturmfels was supported by the US National Science Foundation (DMS-1419018). 22 References [1] A. Abdesselam and J. Chipalkatti: Brill-Gordan loci, transvectants and an analogue of the Foulkes conjecture, Adv. Math. 208 (2007) 491–520. [2] A. Abdesselam and J. Chipalkatti: The bipartite Brill-Gordan locus and angular momentum, Transform. Groups 11 (2006) 341–370. [3] G. Blekherman: Typical real ranks of binary forms, Found. Comput. Math. 15 (2015) 793–798. [4] G. Blekherman, P. Parrilo and R. Thomas: Semidefinite Optimization and Convex Algebraic Geometry, MOS-SIAM Series on Optimization, 13, 2012. [5] J. Chipalkatti: On equations defining coincident root loci, J. Algebra 267 (2003) 246–271. [6] P. Comon and G. Ottaviani: On the typical rank of real binary forms, Linear and Multilinear Algebra 60 (2012) 657–667. [7] J-P. Dedieu and F. Tisseur: Perturbation theory for homogeneous polynomial eigenvalue prob- lems, Linear Algebra Appl. 358 (2003) 71–94. [8] J. Draisma, E. Horobet¸, G. Ottaviani, B. Sturmfels, and R. Thomas: The Euclidean distance degree of an algebraic variety, Found. Comput. Math., to appear. [9] L.M. Feh´er, A. N´emethi and R. Rim´anyi: Coincident root loci of binary forms, Michigan J. Math. 54 (2006) 375–392. [10] I.M. Gel’fand, M.M. Kapranov and A.V. Zelevinsky: Discriminants, Resultants and Multidi- mensional Determinants, Birkhauser, Boston, 1994. [11] D. Hilbert: Singularitaten der Diskriminantenflache, Math. Annalen 30 (1887) 437–441. [12] A. Iarrobino and V. Kanev: Power Sums, Gorenstein Algebras, and Determinantal Loci, Lec- ture Notes in Mathematics, 1721, Springer-Verlag, Berlin, 1999. [13] E. Kaltofen, J. May, Z. Yang, and L. Zhi: Approximate factorization of multivariate polyno- mials using singular value decomposition, J. Symbolic Comput. 43 (2008) 359–376. [14] G. Katz: How tangents solve algebraic equations, or a remarkable geometry of discriminant varieties, Expo. Math. 21 (2003) 219–261. [15] S. Kurmann: Some remarks on equations defining coincident root loci, J. Algebra 352 (2012) 223-231. [16] J. McNamee: Numerical Mathods for Roots of Polynomials, Part I, Studies in Computational Mathematics, 14. Elsevier B.V., Amsterdam, 2007. [17] E. Miller and B. Sturmfels: Combinatorial Commutative Algebra, Graduate Texts in Mathe- matics, Springer Verlag, New York, 2004. [18] F. Napolitano: Topology of complements of strata of the discriminant of polynomials, Comptes Rendue Acad. Sci. Paris, S´erie I, 327 (1998) 665–670. [19] L. Oeding: Hyperdeterminants of polynomials. Adv. Math. 231 (2012) 1308–1326. [20] G. Ottaviani, P.-J. Spaenlehauer and B. Sturmfels: Exact solutions in structured low-rank approximation, SIAM J. Matrix Anal. Appl. 35 (2014) 1521–1542. [21] B. Sturmfels: The Hurwitz form of a projective variety, arXiv:1410.6703. [22] J. Weyman: The equations of strata for binary forms, J. Algebra 122 (1989) 244–249. [23] Z. Zeng: Computing multiple roots of inexact polynomials, Math. Comp. 74 (2005) 869–903. 23 [24] Z. Zeng: Regularization and matrix computation in numerical polynomial algebra, in Approxi- mate Commutative Algebra, 125–162, Texts Monogr. Symbol. Comp., Springer, Vienna, 2009. Authors’ addresses: Hwangrae Lee, Pohang University of Science and Technology, Korea, [email protected] Bernd Sturmfels, University of California, Berkeley, USA, [email protected] 24
1309.1022	2	1309	2015-01-11T13:09:06	On totally geodesic submanifolds in the Jacobian locus	[ "math.AG" ]	We study submanifolds of A_g that are totally geodesic for the locally symmetric metric and which are contained in the closure of the Jacobian locus but not in its boundary. In the first section we recall a formula for the second fundamental form of the period map due to Pirola, Tortora and the first author. We show that this result can be stated quite neatly using a line bundle over the product of the curve with itself. We give an upper bound for the dimension of a germ of a totally geodesic submanifold passing through [C] in M_g in terms of the gonality of C. This yields an upper bound for the dimension of a germ of a totally geodesic submanifold contained in the Jacobian locus, which only depends on the genus. We also study the submanifolds of A_g obtained from cyclic covers of the projective line. These have been studied by various authors. Moonen determined which of them are Shimura varieties using deep results in positive characteristic. Using our methods we show that many of the submanifolds which are not Shimura varieties are not even totally geodesic.	math.AG	math	ON TOTALLY GEODESIC SUBMANIFOLDS IN THE JACOBIAN LOCUS ELISABETTA COLOMBO, PAOLA FREDIANI AND ALESSANDRO GHIGI Abstract. We study submanifolds of Ag that are totally geodesic for the locally symmetric metric and which are contained in the closure of the Jacobian locus but not in its boundary. In the first section we recall a formula for the second fundamental form of the period map Mg ֒→ Ag due to Pirola, Tortora and the first author. We show that this result can be stated quite neatly using a line bundle over the product of the curve with itself. We give an upper bound for the dimension of a germ of a totally geodesic submanifold passing through [C] ∈ Mg in terms of the gonality of C. This yields an upper bound for the dimension of a germ of a totally geodesic submanifold contained in the Jacobian locus, which only depends on the genus. We also study the submanifolds of Ag obtained from cyclic covers of P1. These have been studied by various authors. Moonen determined which of them are Shimura varieties using deep results in positive characteristic. Using our methods we show that many of the submanifolds which are not Shimura varieties are not even totally geodesic. Contents Introduction 1. 2. Notation and preliminary results 3. The second fundamental form as a multiplication map 4. Totally geodesic submanifolds and gonality 5. Families of cyclic covers of the projective line References 2 4 8 12 14 19 2000 Mathematics Subject Classification. 14H10;14H15;14H40;32G20. The first author was partially supported by PRIN 2010-11 MIUR “Geometria delle Variet`a Proiettive”. The second and third authors were partially supported by PRIN 2009 MIUR ”Moduli, strutture geometriche e loro applicazioni”. The second author was partially supported also by FIRB 2012 ”Moduli spaces and applications”. The third author was supported also by FIRB 2012 ”Geometria differenziale e teoria geometrica delle funzioni” and also by a grant of Max-Planck Institut fur Mathematik, Bonn. The three authors were partially supported by INdAM (GNSAGA) . 1 2 ELISABETTA COLOMBO, PAOLA FREDIANI AND ALESSANDRO GHIGI 1. Introduction 1.1. Denote by Ag the moduli space of principally polarized abelian vari- eties of dimension g, by Mg the moduli space of smooth curves of genus g and by j : Mg → Ag the period mapping or Torelli mapping. Both Mg and Ag are complex orbifolds (or smooth stacks) and Ag is endowed with a locally symmetric metric, the so-called Siegel metric. One expects the Jacobian locus, that is the image j(Mg) ⊂ Ag, to be rather curved with respect to the Siegel metric. In particular, it should contain very few totally geodesic submanifolds of Ag. Another reason for this expectation comes from arithmetic geometry. Indeed for a special class of totally geodesic sub- manifolds (Shimura varieties) it has been conjectured by Coleman and Oort that for large genus no positive dimensional Shimura variety is contained in the closure of the Jacobian locus (in Ag) and meets the Jacobian locus it- self. Moonen [20] has proven that an algebraic totally geodesic submanifold is a Shimura subvariety if and only if it contains a complex multiplication point. For results on Shimura subvarieties contained in j(Mg) we refer to [13, 20, 22, 25, 21, 1, 19, 15, 18]. Outside the hyperelliptic locus the period map is an orbifold immersion. For g ≥ 4 the Jacobian locus j(Mg) has dimension strictly smaller than Ag. Therefore it makes sense to compute the second fundamental form of j(Mg) ⊂ Ag and to study its metric properties by infinitesimal methods. The second fundamental form has been studied by Pirola, Tortora and the first author [6], where an expression for it is given and it is proven that the second fundamental form lifts the second Gaussian map, as stated in an unpublished paper by Green and Griffiths [12]. In particular the computation of the second fundamental form on ξp ⊙ ξp (where ξp is a Schiffer variation at the point p on the curve) reduces to the evaluation of the second gaussian map at the point p. These results have been used in [5] to compute the curvature of the restriction to Mg of the Siegel metric. In [5] there is an explicit formula for the holomorphic sectional curvature of Mg in the direction ξp in terms of the holomorphic sectional curvature of Ag and the second Gaussian map. It is much harder to use the formula in [6] to compute the second fun- damental form on ξp ⊙ ξq when p 6= q. In fact the formula contains the evaluation at q of a meromorphic 1-form on the curve, called ηp, which has a double pole at p and is defined by Hodge theory. In general it seems rather hard to control the behaviour of ηp, in a way to get constraints on the second fundamental form. 1.2. In this paper we give a global and more intrinsic description of this form. We show that as p varies on the curve the forms ηp glue to give a holomorphic section η of the line bundle KS(2∆), where S = C × C and ∆ ⊂ S is the diagonal. With this interpretation we are able to prove that the second fundamental form coincides with the multiplication by η. More precisely, fix a genus g, which will always be assumed greater than 3, and fix [C] ∈ Mg outside the hyperelliptic locus. The conormal bundle ON TOTALLY GEODESIC SUBMANIFOLDS IN THE JACOBIAN LOCUS 3 of j : Mg ⊂ Ag at [C] can be identified with I2(KC ), which is the kernel of the multiplication map S2H 0(C, KC ) → H 0(C, 2KC ). Hence the second fundamental form can be seen as a map ρ : I2(KC ) → S2H 0(C, 2KC ). Let S = C × C and let ∆ be the diagonal. By Kunneth formula H 0(S, KS ) = H 0(C, KC ) ⊗ H 0(C, KC ) and H 0(S, 2KS ) = H 0(C, 2KC ) ⊗ H 0(C, 2KC ). In particular I2(KC) ⊂ H 0(S, KS (−2∆)). Theorem A. (See Theorem 3.13). The second fundamental form ρ is the restriction to I2 of the multiplication map H 0(S, KS (−2∆)) −→ H 0(S, 2KS ) Q 7→ Q · η. 1.3. Based on these results on the second fundamental form we get some constraints on the existence of totally geodesic submanifolds of Ag contained in Mg. Since our methods are local in nature, the results apply to germs of such submanifolds. We get upper bounds for the dimension of totally geodesic germs passing through [C] ∈ Mg in terms of the gonality of the curve C. Theorem B. (See Theorem 4.3). Assume that C is a k-gonal curve of genus g with g ≥ 4 and k ≥ 3. Let Y be a germ of a totally geodesic submanifold of Ag which is contained in the jacobian locus and passes through j([C]) = [J(C)]. Then dim(Y ) ≤ 2g + k − 4. This immediately yields a bound which only depends on g. Theorem C. (See Theorem 4.5). If g ≥ 4 and Y is a germ of a totally geodesic submanifold of Ag contained in the jacobian locus, then dim Y ≤ 5 2 (g − 1). 1.4. For low genus one can construct examples of totally geodesic submani- folds contained in Mg using cyclic covers of P1, see e.g. [7, 21, 25]. These are in fact Shimura varieties. A complete list of the Shimura varieties that can be obtained in this way has been given in [21] using deep results in positive characteristic. With our methods we check directly that these examples are indeed totally geodesic and we show that a large class of cyclic covers, which are not in the list of Shimura varieties, are not even totally geodesic (see Proposition 5.7 and Corollary 5.8). 1.5. Other works studying totally geodesic submanifolds contained in the Jacobian locus include [26, 13, 8]. In particular Hain [13] proves the fol- lowing. Let X be an irreducible symmetric domain and consider the locally symmetric variety Γ\X (where Γ is a lattice). If there is a totally geodesic immersion Γ\X → j(Mg), and if some additional conditions are satisfied, then X must be the complex ball. De Jong and Zhang [8] prove a similar result under milder conditions, but still retaining the irreducibility assump- tion on X. The techniques used in these works are global and are based on group cohomology and on a rigidity theorem for the mapping class group due to Farb and Masur [9]. Our result instead applies to germs of totally 4 ELISABETTA COLOMBO, PAOLA FREDIANI AND ALESSANDRO GHIGI geodesic submanifolds, since it is local in nature and does not require ir- reducibility assumptions. The same local point of view is present in [17], where the object of study are totally geodesic submanifolds contained in an algebraic subvariety of a complex hyperbolic space form. Acknowledgements. We wish to thank Fabrizio Andreatta and Bert van Geemen for interesting conversations. The second and third authors wish to thank the Max-Planck Institut fur Mathematik, Bonn for excellent conditions provided during their visit at this institution, where part of this paper was written. 2. Notation and preliminary results 2.1. Second fundamental form. Denote by Ag the moduli space of prin- cipally polarized abelian varieties of dimension g, by Mg the moduli space of smooth curves of genus g and by j : Mg → Ag the period mapping or Torelli mapping. By the Torelli theorem j is injective. To study j one can fix a level structure with n ≥ 3 and consider M (n) g which is a smooth map between manifolds. Since level structures play no role in what we are doing, it is more appropriate to think of Mg and Ag as complex orbifolds or smooth stacks, see e.g. [2, XII, 4]. The period map is smooth in the orbifold sense. Moreover its restriction to the set of non-hyperelliptic curves is an orbifold immersion [23]. By abuse of terminology we will henceforth omit the word ”orbifold”. The moduli space Ag is endowed with the Siegel metric, which is the metric induced on Ag from the symmetric metric on the Siegel upper halfspace Hg. Outside the hyperelliptic locus we have the sequence of tangent bundles: j(n) → A(n) g whose dual, at [C] ∈ Mg is 0 → T Mg → j∗(cid:0)T Ag(cid:1) π→ N → 0, 0 → I2 → S2H 0(C, KC ) m→ H 0(C, 2KC ) → 0, where I2 := I2(KC ) is the set of quadrics containing the canonical curve and m is the multiplication map (see [5] for more details). Denote by II : S2T[C]Mg = S2H 1(C, TC ) → N[C] the second fundamental form of the period map with respect to the Siegel metric on Ag. Denote by ρ : I2 → S2H 0(C, 2KC ) the dual of II. We will refer both to II and to ρ as second fundamental forms. 2.2. Schiffer variations. If C is a curve and x ∈ C, the coboundary of the exact sequence 0 → TC → TC(x) → TC(x)x → 0 yields an injection H 0(TC (x)x) ∼= C ֒→ H 1(C, TC ). Elements in the image are called Schiffer ON TOTALLY GEODESIC SUBMANIFOLDS IN THE JACOBIAN LOCUS 5 variations at x. If (U, z) is a chart centred at x and b ∈ C ∞ function which is equal to 1 on a neighbourhood of x, then 0 (U ) is a bump θ := ¯∂b z · ∂ ∂z is a Dolbeault representative of a Schiffer variation at x. The map ξ : T C → H 1(C, TC ) u = λ ∂ ∂z (x) 7→ ξu := λ2[θ] does not depend on the choice of the coordinates. Schiffer variations generate H 1(C, TC ) [2, p.175]. Lemma 2.3. Let β ∈ H 0(C, 2KC ) and let (U, z) be a chart centred at x ∈ C. If β = f (z)(dz)2 on U , then β(cid:0)ξ ∂ ∂z (x)(cid:1) = 2πi f (0). It is well known that Proof. ∂z (x)(cid:1) = ZC β(cid:0)ξ ∂ β ∪ θ = ZU f (z)dz ∧ ¯∂b z = −ZU −{x} ¯∂(cid:18) b(z)f (z) z dz(cid:19). If ε > 0 is small enough, b ≡ 1 on {z ≤ ε}. Using Stokes and Cauchy theorems we get β(cid:0)ξ ∂ ∂z (x)(cid:1) = − lim ε→0 Z ¯∂(cid:18)b(z)f (z) z dz(cid:19) = lim ε→0 Z f (z) z dz = 2πif (0). U ∩{z>ε} z=ε (cid:3) 2.4. Gaussian maps. We briefly recall the definition of Gaussian maps for curves. Let N and M be line bundles on C. Set S := C × C and let ∆ ⊂ S be the diagonal. For a non-negative integer k the k-th Gaussian or Wahl map associated to these data is the map given by restriction to the diagonal µk N,M−→ H 0(S, N ⊠ M (−k∆)∆) ∼= H 0(C, N ⊗ M ⊗ K k H 0(S, N ⊠ M (−k∆)) C ). We are only interested in the case N = M . In this case we set µk,M := µk M,M . With the indentification H 0(S, N ⊠ M ) ∼= H 0(C, N ) ⊗ H 0(C, M ) the map µ0,M is the multiplication map of global sections H 0(C, M ) ⊗ H 0(C, M ) → H 0(C, M 2), which obviously vanishes identically on ∧2H 0(C, M ). Consequently ker µ0,M = H 0(S, M ⊠ M (−∆)) decomposes as ∧2H 0(C, M ) ⊕ I2(M ), where I2(M ) is the kernel of S2H 0(C, M ) → H 0(C, M 2). Since µ1,M vanishes on symmetric tensors, one usually writes µ1,M : ∧2H 0(M ) → H 0(KC ⊗ M 2). If σ is a local frame for M and z is a local coordinate, given sections s1, s2 ∈ H 0(C, M ) with si = fi(z)σ, we have (2.1) µ1,M (s1 ∧ s2) = (f ′ 2f1)dz ⊗ σ2. 1f2 − f ′ 6 ELISABETTA COLOMBO, PAOLA FREDIANI AND ALESSANDRO GHIGI Consequently, the zero divisor of µ1,M (s1 ∧ s2) is twice the base locus of the pencil hs1, s2i plus the ramification divisor of the associated morphism. Again H 0(S, M ⊠ M (−2∆)) decomposes as the sum of I2(M ) and the kernel of µ1,M . Since µ2,M vanishes identically on skew-symmetric tensors, one usually writes µ2,M : I2(M ) → H 0(C, M 2 ⊗ K 2 C ). By µ2 we denote the second gaussian map of the canonical line bundle KC on C: µ2 := µ2,KC : I2(KC ) → H 0(K 4 C). 2.5. The form ηx and the second fundamental form. We now recall the definition of ηx. Let C be a smooth complex projective curve of genus g ≥ 4. Fix a point x ∈ C. The space H 0(C, KC (2x)) is contained in the space of closed 1-forms on C − {x}. The induced map H 0(C, KC (2x)) → H 1(C−{x}, C) is injective as soon as g > 0. By the Mayer-Vietoris sequence, the inclusion C − {x} ֒→ C induces an isomorphism H 1(C, C) ∼= H 1(C − {x}, C). Thus we get an injection (2.2) jx : H 0(C, KC (2x)) ֒→ H 1(C, C). H 1,0(C) is contained in the image of jx and h0(C, KC (2x)) = g + 1, so j−1 x (H 0,1(C)) is a line. If (U, z) is a chart centred at x, there is a unique element ϕ in this line such that on U − {x} ϕ = (cid:18) 1 z2 + h(z)(cid:19)dz with h ∈ OC(U ). (Applying Stokes theorem on C minus a disc around x shows that the residue at x vanishes, so there is no term in 1/z.) Define a linear map by the rule ηx : TxC → H 0(C, KC (2x)) u = λ ∂ ∂z (x) 7−→ ηx(u) := λϕ. We will often drop x and simply write ηu for ηx(u). An easy computation shows that ηx does not depend on the choice of the local coordinate. Theorem 2.6 ([6, Thm. 3.1], [5, Lemma 3.5]). Let C be a non-hyperelliptic curve of genus g ≥ 4. Given points x 6= y in C and tangent vectors u ∈ TxC and v ∈ TyC we have ρ(Q)(ξu ⊙ ξv) = −4πi · ηx(u)(v)Q(u, v), ρ(Q)(ξu ⊗ ξu) = −2πi · µ2(Q)(u⊗4). This theorem is basic to the whole paper. For the reader’s convenience we recall its proof. ON TOTALLY GEODESIC SUBMANIFOLDS IN THE JACOBIAN LOCUS 7 f → ∆ with C = f −1(0) and w = κ( ∂ Proof. First of all we take w ∈ H 1(C, TC ) and we compute ρ(Q)(w) ∈ H 0(C, 2KC ) for Q ∈ I2(KC ). We can assume that there is a a one di- mensional deformation X ∂t ), where ∆ = {t < 1} and κ is the Kodaira-Spencer map. Take a C∞ lifting Y of the holomorphic vector field ∂ ∂t . So we have a C∞ trivialization τ : ∆ × C → X , τ (t, x) = τt(x) := Φt(x), where {Φt} is the flow of the vector field Y . Then θ := ∂Y C is a ¯∂-closed form in A0,1(C, TC ) such that [θ] = w ∈ H 1(C, TC ). Denote by Ct the fibre of f over t and let ω(t) be a section of the Hodge bundle, i.e. ∀t ∈ ∆, ω(t) ∈ H 0(KCt ). Since ω(t) is closed, also τ ∗ t (ω(t)) is closed, so τ ∗ t (ω(t)) = ω + (α + dh)t + o(t), where ω := ω(0), α is harmonic and h is a C∞ function. Denote by ∇GM the flat Gauss-Manin connec- tion on R1f∗C. So we have ∇GM ∂/∂t[ω(t)]t=0 = [α]. By Griffiths’ results (see ∂t ) · [ω] = [α0,1], where α0,1 e.g. is the (0, 1) component of α. Now assume that {ωi}i=1,...,g is a basis of H 0(KC ). Take a quadric Q = Pi,j aijωi ⊗ ωj ∈ I2(KC). Denote by ∇1,0 the Gauss-Manin connection on the Hodge bundle f∗ωX /∆, i.e. ∇1,0 = π∇GM , where π is the projection of H 1(Ct, C) onto H 0(Ct, KCt ). Then for all i we have ∇1,0 1,0]. Denote by ∇ the induced connection on [29, pp. 234ff]) θ · ω = α0,1 + ∂h and κ( ∂ ∂/∂t[ωi(t)]t=0 = [αi S2f∗ωX /∆. If Q(t) = Pi,j aij(t)ωi(t) ⊗ ωj(t) ∈ I2(KCt ) is a section of the conormal bundle such that Q(0) = Q, then ρ(Q)(w) = m(∇ ∂ ∂t Qt=0) = Xi,j ij(0)ωiωj + 2Xi,j a′ aijα1,0 i ωj. Since Pi,j aij(t)ωi(t)ωj(t) ≡ 0, also its derivative at t = 0 vanishes, i.e. 2Pi,j aij(αi + dhi)ωj + Pi,j a′ ij(0)ωiωj ≡ 0, and if we take the (1, 0) part we have 2Pi,j aij(α1,0 ijωiωj ≡ 0, so (2.3) i + ∂hi)ωj +Pi,j a′ ρ(Q)(w) = −2Xi,j aijωj∂hi. This is an instance of a Hodge-Gaussian map, see [6] and also [24, §4]. ∂z (x) at x ∈ C with Dolbeault representative θ := Now fix a point x ∈ C and a chart (U, z) centred at x. Let w be the ¯∂b z · ∂ Schiffer variation ξ ∂ ∂z , where b ∈ C ∞ 0 (U ) be a bump function equal to 1 on a neighbourhood of x. Let ωi = fi(z)dz be the local expression of ωi in U . On C − {x} we have i = ¯∂gi. θ · ωi = fi z Define ηi := ∂gi. i + ¯∂hi = ¯∂(cid:16) bfi z (cid:17). Set gi := bfi z − hi. Then α0,1 ¯∂b, so α0,1 Now we will show that P aijωi∂hj = −P aijωiηj. In the first place, note that ηi is holomorphic in C −{x}. Indeed, αi is harmonic, thus ¯∂ηi = ¯∂∂gi = −∂ ¯∂gi = −∂α0,1 z ) = P aijωi∂hj + P aijωiηj is holomorphic on C − {x}, because the first term is a holomorphic section of 2KC by (2.3) and the second is holomorphic on C − {x} since ηj is holomorphic. In a neighborhood of x where b ≡ 1, this expression has the i = 0. Hence, P aij ωi∂( bfj 8 ELISABETTA COLOMBO, PAOLA FREDIANI AND ALESSANDRO GHIGI form X aijfi ∂ ∂z (cid:18) fj because P aijfifj = P aijfif ′ Thus we have f ′ j z (cid:19) dz2 = X aij fi(cid:18)− fj z2 + j = 0. So P aij ωi∂( bfj z (cid:19) dz2 = 0, z ) is identically zero. ρ(Q)(ξ ∂ ∂z (x)) = −2X aijωi∂hj = 2X aijωiηj. Now we claim that ηi = −fi(x)ηx( ∂ ∂z ). In fact ηi ∈ H 0(KC (2x)), since on U, where b ≡ 1, ηi has the form ηi = (cid:18)− fi(x) z2 + ψi(z)(cid:19) dz with ψi(z) a holomorphic function, hence ηi is a meromorphic form, with a double pole at x. By definition, ηi + α0,1 i = ∂gi + ¯∂gi = dgi, so jx(ηi) = −[α0,1 i ] ∈ jx(H 0(KC (2x))) ∩ H 0,1(C). This proves the claim. We can assume that the chart (U, z) contains also y. Applying Lemma 2.3 the first statement immediately follows for u = ∂/∂z(x) and v = ∂/∂z(y). It is clear that this is enough. For the second statement it suffices to use the local expression of µ2. (cid:3) Remark 2.7. We remark that Schiffer variations, the forms ηx, Q and µ2(Q) are sections of vector bundles, but they become functions as soon as a coordinate chart is fixed. Because of this many statements, like the one above, are usually stated for simplicity, as if these sections were evaluated at points instead of tangent vectors. We will follow this notation when it is convenient. In the next section instead it is better to stick a more formal notation. 3. The second fundamental form as a multiplication map 3.1. In this section we show that as x varies on the curve the form ηx varies holomorphically in an appriopriate sense and gives rise to a section of a vector bundle on C and to a corresponding section η of the line bundle KS(2∆) on S = C × C. The two main points are Theorem 3.13 and the invariance of η with respect to the action of Aut(C). 3.2. Let (U, z) be a chart centred at x ∈ C. Set u = ∂ ∂z (x). It is a classical result that there is a harmonic function fu ∈ C ∞(C − {x}) such that fu = −1/z + g(z) on U − {x} for some g ∈ C ∞(U ). This function is unique up to an additive constant and is called elementary potential. Its existence can be proven for example using the (real) Hodge decomposition theorem and the Weyl lemma (see e.g. [10, p. 46-48]) or using the Perron method (see e.g. [16, p. 213ff]). ON TOTALLY GEODESIC SUBMANIFOLDS IN THE JACOBIAN LOCUS 9 Lemma 3.3. If fu is an elementary potential, then ∂fu = ηu, ¯∂fu is smooth on C and jx(ηu) = [− ¯∂fu]. Proof. The (1, 0)-form ∂fu is holomorphic on C − {x} since fu is harmonic. Moreover ∂fu = z−2dz + ∂g on U − {x}. The form ∂g is holomorphic on U − {x}, but also smooth on U . Hence it is holomorphic on U . This shows that ∂fu = ηu + ω for some ω ∈ H 0(KC ). Set α := − ¯∂fu. Then α is smooth on C by the definition of fu. It is closed, since fu is harmonic and of type (0, 1). On C − {x} we have ∂fu − α = dfu, so [∂fu] = [α] in H 1(C − {x}). Therefore jx(∂fu) = [α]. Since [α] ∈ H 0,1(C), this shows that jx(∂fu) ∈ H 0,1(C). Therefore ω = 0 and ηu = ∂fu. (cid:3) Remark 3.4. Notice that one could prove the existence of fu using ηu and the fact that C − {x} is Stein. Lemma 3.5. If H 0,1(C) is identified with H 0(C, KC )∗ using Serre duality, then x 7→ Im jx ∩ H 0,1(C) ∈ P(H 0,1(C)) coincides with the canonical map. Proof. Let (U, z) be a coordinate centred at x. Set u = ∂ If ω ∈ H 0(C, KC ), let ω = ϕ(z)dz be its local expression on U . By Lemma 3.3 jx(ηu) = −[ ¯∂fu]. Therefore ∂z (x). (3.1) = lim ω ∧ jx(ηu) = −ZC ZC ε→0Zz=ε ϕ(z)(cid:16)− 1 z d(fu ω) = ω ∧ ¯∂fu = ZC−{x} + g(z)(cid:17)dz = −ϕ(0) = −ω(u). (cid:3) 3.6. Let S := C × C, let ∆ denote the diagonal and let p, q : S → C be the projections p(x, y) = x, q(x, y) = y. Then KS = p∗KC ⊗ q∗KC. Consider the line bundle L := KS(2∆) on S and set V := p∗(q∗KC (2∆)) E := p∗L. By the projection formula E = KC ⊗V . Since q∗KC (2∆){x}×C = q∗KC(2x), we have H 0(p−1(x), q∗KC (2∆)) ∼= H 0(C, KC (2x)). By Grauert semiconti- nuity theorem V is a holomorphic vector bundle on C with fibre Vx ∼= H 0(C, KC (2x)) and the map x 7→ ηx is a section of E. We call this section η. Proposition 3.7. η is a holomorphic section of E. Proof. Let W → C denote the trivial vector bundle with fibre H 1(C, C). We claim that the injection j : V ֒→ W defined in (2.2) is holomorphic. Fix α ∈ H 0(C, KC (2x)) and a smooth singular 1-cycle c on C. If x does not lie in the support of c, the integral Rc α is well-defined. It does not change if α is replaced by α + df with f ∈ C ∞(C − {x}). Therefore Rc α = h[c], j(α)i. Fix x0 ∈ C and choose smooth 1-cycles c1, . . . , c2g, that do not touch x0 and whose classes form a basis of H1(C, C). Let A be an open subset of C, such 10 ELISABETTA COLOMBO, PAOLA FREDIANI AND ALESSANDRO GHIGI that V is trivial over A and A does not intersect the supports of the cycles ci. Fix s ∈ H 0(A, V ). To show that j(s) is a holomorphic section of W over A, it is enough to prove that the functions x 7−→ h[ci], jx(s(x))i = Zci s(x) are holomorphic on A. Since V = p∗q∗KC(2∆), s corresponds to a section s ∈ H 0(A × C, q∗KC(2∆)). So s(x, ·)C−A is a holomorphic 1-form on C − A depending holomorphically on the parameter x ∈ A and its integral over the 1-cycle ci (which is contained in C − A) is a holomorphic function of x. Therefore j(s) is holomorphic. Since s is arbitrary, this proves that j is holomorphic, as claimed. Next fix a chart (U, z) on C. To show that η is holomorphic on U it is enough to prove that η( ∂ ∂z ) is holomorphic or - by the above - that j(η( ∂ ∂z )) is a holomorphic function U → H 1(C, C). Fix a basis {ω1, . . . , ωg} of H 0(C, KC ). The functionals RC(·) ∧ ωj and RC(·) ∧ ωj ∂z )) ∈ H 0,1(C) the latter g functionals form a basis of H 1(C, C)∗. Since j(η( ∂ vanish on j(η( ∂ ∂z )). Assume that ωj(z) = fj(z)dz on U . By (3.1) ∂ ∂z j(cid:16)η( ZC ∂z )) is holomorphic. )(cid:17) ∧ ωj = fj(z). This proves that j(η( ∂ 3.8. Since E = p∗L there is an isomorphism H 0(C, E) ∼= H 0(S, L) that associates to α ∈ H 0(C, E) the section α ∈ H 0(S, L) such that (cid:3) αx = α{x}×C ∈ T ∗ x C ⊗ H 0(C, KC (2x)) = Ex. It follows that αx(u)(v) = α(cid:0)(u, 0), (0, v)(cid:1) for x 6= y, u ∈ TxC and v ∈ TyC. Thus there is a well-defined section η ∈ H 0(S, L) corresponding to η and for u ∈ TxC and v ∈ TyC with x 6= y, we have ηx(u)(v) = η(u, v). Proof. Fix points x 6= y and tangent vectors u ∈ TxC, y ∈ TyC. Using the Lemma 3.9. The form η is symmetric, i.e. η(cid:0)(u, 0), (0, v)(cid:1) = η(cid:0)(v, 0), (0, u)(cid:1). identity d(cid:16)fu( ¯∂fv − ∂fv) − fv( ¯∂fu − ∂fu)(cid:17) = 2(cid:0)fu∂ ¯∂fv − fv∂ ¯∂fu(cid:1) = 0 and applying Stokes theorem on C − {z < ε} ∪ {w < ε} we get 0 = Zz=ε(cid:16)fu( ¯∂fv − ∂fv) − fv( ¯∂fu − ∂fu)(cid:17)+ +Zw=ε(cid:16)fu( ¯∂fv − ∂fv) − fv( ¯∂fu − ∂fu)(cid:17) (This is just Green formula.) Let us denote by Aε and Bε these two inte- grals. Observe that fv∂fu = ∂(fufv) − fu∂fv and fu ¯∂fv = ¯∂(fufv) − fv ¯∂fu. Therefore Aε = Zz=ε(cid:16)d(fufv) − 2fv ¯∂fu − 2fu∂fv(cid:17) = −2Zz=ε(cid:16)fv ¯∂fu + fu∂fv(cid:17). ON TOTALLY GEODESIC SUBMANIFOLDS IN THE JACOBIAN LOCUS 11 Since fv ¯∂fu is a smooth form near x, we have limε→0Rz=ε fv ¯∂fu = 0. Choose a coordinate (U, z) centred at x such that u = ∂/∂z(x) and let g(z) be as in 3.2. Near x we have ηy(v) = ∂fv = h(z)dz for some holomorphic function h. So lim ε→0 Aε = −2 lim ε→0Zz=ε fu∂fv = = −2 lim ε→0Zz=ε(cid:18)− 1 z + g(z)(cid:19)h(z)dz = 4πih(0) = 4πiηy(v)(u). By the corresponding computation, limε→0 Bε = −4πiηx(u)(v), so ηy(v)(u) = ηx(u)(v) as desired. (cid:3) 3.10. Aut(C) acts diagonally on C × C and preserves ∆. Therefore the action lifts to KS and also to L = KS(2∆). This yields a representation of Aut(C) on H 0(S, L). On the other hand, if σ ∈ Aut(C), then (σ−1)∗ is a map from H 0(C, KC (2x)) = Vx to H 0(C, KC (2σ(x))) = Vσ(x). Tensoring it with (dσ−1)∗ : T ∗ σ(x)C ⊗ Vσ(x). This yields an action of Aut(C) on the total space of E = KC ⊗ V which covers the action of Aut(C) on C. This means that E is an equivariant bundle. In this way we get a representation of Aut(C) on H 0(C, E). The map α 7→ α considered in 3.8 is an isomorphism of Aut(C)–representations. σ(x)C we get a map T ∗ x C ⊗ Vx → T ∗ x C → T ∗ Lemma 3.11. η and η are invariant with respect to the action of Aut(C). Proof. By the above it is enough to show that η is invariant. For τ ∈ Aut(C) we wish to prove that (dτ −1)∗ ⊗ (τ −1)∗(ηy) = ητ (y) for any y ∈ C. For simplicity, set σ := τ −1, x := τ (y). So we wish to prove that σ∗(ηy(dσ(u))) = ηx(u) for any u ∈ TxC. By continuity it is enough to prove this for x such that σ(x) 6= x. Choose a coordinate patch (U, z) centred at x, such that U ∩ σ(U ) 6= ∅ and u = ∂ ∂z x. Then (σ(U ), w := z ◦ σ−1) is a coordinate system centred at σ(x) and ∂ ∂w (σ(x)) = dσ(u). Assume that ηx(u) = (cid:18) 1 z2 + h(z)(cid:19)dz on U . Then τ ∗ηx(u) = (cid:18) 1 w2 + h(w)(cid:19)dw on σ(U ). Moreover jyτ ∗ = τ ∗jx. So jyτ ∗(ηx(u)) = τ ∗jx(ηx(u)). Since jx(ηx(u)) ∈ H 0,1(C), also jyτ ∗(ηx(u)) ∈ H 0,1(C). Hence τ ∗ηx(u) = ηy(dσ(u)) as desired. (cid:3) 3.12. If we identify H 0(C, KC ) ⊗ H 0(C, KC ) with H 0(S, KS), then I2 be- comes a subspace of H 0(S, KS(−∆)). Since elements of I2 are symmetric, they are in fact in H 0(S, KS (−2∆)). So if Q ∈ I2 the product section Q · η lies in H 0(S, 2KS ) ∼= H 0(C, 2KC ) ⊗ H 0(C, 2KC ). 12 ELISABETTA COLOMBO, PAOLA FREDIANI AND ALESSANDRO GHIGI Theorem 3.13. With the above identifications, if C is non-hyperelliptic and of genus g ≥ 4, then ρ : I2 → S2H 0(C, 2KC ) is the restriction to I2 of the multiplication map H 0(S, KS (−2∆)) −→ H 0(S, 2KS ) Q 7→ Q · η. Proof. The identification H 0(C, 2KC ) ⊗ H 0(C, 2KC ) ∼= H 0(S, 2KS ) is com- patible with Lemma 2.3, i.e. for α ∈ H 0(S, 2KS ), u ∈ TxC and v ∈ TyC we have hα, ξu ⊗ ξvi = α(cid:0)(u, 0), (0, v)(cid:1). If x 6= y, by Theorem 2.6 ρ(Q)(ξu ⊙ ξv) = Q(u, v)ηx(u)(v) = = (Q · η)(cid:0)(u, 0), (0, v)(cid:1) = (Q · η) (ξu ⊙ ξv). (In the last identity we use the fact that both Q and η are symmetric). So ρ(Q) − Q · η vanishes on tensors of the form ξu ⊙ ξv with x 6= y. Clearly we can choose a basis of S2H 1(C, TC ) formed by such elements. (cid:3) 3.14. Since the second fundamental form is symmetric, using this theorem we get another proof of Lemma 3.9. 4. Totally geodesic submanifolds and gonality 4.1. In this section we will give an upper bound for the dimension of a germ of a totally geodesic submanifold contained in the Jacobian locus. Theorem 4.2. Assume that C is a k-gonal curve of genus g, with g ≥ 4 and k ≥ 3. Then there exists a quadric Q ∈ I2(KC ) such that rank ρ(Q) > 2g − 2k. Proof. Here we follow the simplified notation mentioned in 2.7. So we un- derstand that a local coordinate has been fixed at the relevant points and we write ξP for a Schiffer variation at P . Let F be a line bundle on C such k and choose a basis {x, y} of H 0(F ). Set M = KC − F and that F is a g1 denote by B the base locus of M . By Clifford theorem deg(B) < k − 2. Assume that ht1, t2i is a pencil in H 0(M ), with base locus B. We can write ti = t′ 2i is a base point free pencil in M − B. Let ψ : C → P1 be the morphism induced by this pencil. is for a section s ∈ H 0(C, OC (B)) with div(s) = B. Then ht′ 1, t′ Now consider the rank 4 quadric Q := xt1 ⊙ yt2 − xt2 ⊙ yt1. Clearly Q ∈ I2(KC ). Set d := deg(M − B) = 2g − 2 − k − deg(B). We need the if {P1, ..., Pd} is a fibre of the morphism ψ over a regular following fact: value, then the Schiffer variations ξP1, ..., ξPd are linearly independent in H 1(C, TC ). In fact, denote by W := hξP1, ..., ξPdi. We want to show that the annihilator Ann(W ) of W has dimension 3g − 3 − d. By lemma 2.3, Ann(W ) = {α ∈ H 0(C, 2KC ) α(Pi) = 0, i = 1, ..., d}, hence by Riemann- Roch dim Ann(W ) = h0(2KC − P1 − · · · − Pd) = h0(2KC − M + B) = g − 1 + k+deg(B) = 3g−3−d. The claim is proven. Next denote by ϕ the morphism induced by the pencil F and consider the set E := ψ(Crit(ϕ)∪ Crit(ψ)∪ B) where Crit(ϕ) (resp. Crit(ψ)) denote the set of critical points of ϕ (resp. ψ). ON TOTALLY GEODESIC SUBMANIFOLDS IN THE JACOBIAN LOCUS 13 Let z ∈ P1 − E and let {P1, . . . , Pd} be the fibre of ψ over z. By changing coordinates on P1 we can assume z = [0, 1], i.e. t′ 1(Pi) = 0 for i = 1, . . . d. Then clearly t1(Pi) = 0, so Q(Pi, Pj ) = 0 for all i, j. Applying Theorem 2.6, one immediately obtains that the restriction of ρ(Q) to the subspace W is represented in the basis {ξP1, ..., ξPd } by a diagonal matrix with entries πi · µ2(Q)(Pi) on the diagonal. For a rank 4 quadric the second Gaussian map can be computed as follows: µ2(Q) = µ1,F (x ∧ y)µ1,M (t1 ∧ t2) see [4, Lemma 2.2]. Now µ1,F (x ∧ y)(Pi) 6= 0, because Pi 6∈ Crit(ϕ) by the choice of z, see (2.1). Moreover Pi 6∈ B. On C − B the morphism ψ coincides with the map associated to ht1, t2i. Since Pi 6∈ Crit(ψ), it is not a critical point for the latter map. Therefore also µ1,M (t1 ∧ t2)(Pi) 6= 0 see (2.1). Thus µ2(Q)(Pi) = µ1,F (x ∧ y)(Pi)µ1,M (t1 ∧ t2)(Pi) 6= 0 for every i = 1, . . . , d. This shows that in the basis {ξP1, ..., ξPd } the quadric ρ(Q)W is represented by a diagonal matrix with non-zero diagonal entries. So ρ(Q) has rank at least d = 2g − 2 − k − deg B > 2g − 2k. (cid:3) Theorem 4.3. Assume that C is a k-gonal curve of genus g with g ≥ 4 and k ≥ 3. Let Y be a germ of a totally geodesic submanifold of Ag which is contained in the jacobian locus and passes through [C]. Then dim Y ≤ 2g + k − 4. Proof. By Theorem 4.2 we know that there exists a quadric Q ∈ I2 such that the rank of ρ(Q) is at least 2g − 2k + 1. By assumption for any v ∈ T[C]Y we must have that ρ(Q)(v ⊙ v) = 0, so v is isotropic for ρ(Q), hence dim T[C]Y ≤ 3g − 3 − (2g − 2k + 1) 2 = 2g + k − 7 2 . (cid:3) Remark 4.4. In Theorem 4.2 if M is base point free – this happens in particular if C is generic in the locus of k-gonal curves – the above proof shows that rank ρ(Q) ≥ 2g − 2 − k. So if Y is a germ of a totally geodesic submanifold of Ag contained in the jacobian locus and passing through a generic k-gonal curve, then dim Y ≤ 2g − 2 + k/2. Theorem 4.5. If g ≥ 4 and Y is a germ of a totally geodesic submanifold of Ag contained in the jacobian locus, then dim Y ≤ 5 2 (g − 1). Proof. This immediately follows from Theorem 4.3, since the gonality of a genus g curve is at most [(g + 3)/2]. (cid:3) Corollary 4.6. For any g ≥ 4 and k ≥ 2 the k-gonal locus is not totally geodesic in Ag. Proof. For k = 2 this is Proposition 5.1 in [5]. If k ≥ 3 the dimension of the k-gonal locus is 2g + 2k − 5 > 2g + k − 4. Hence the statement follows immediately from Theorem 4.3. (cid:3) Remark 4.7. As it is evident from the proof, gonality is used to construct a quadric Q ∈ I2(KC) of rank 4 such that ρ(Q) has large rank. It seems 14 ELISABETTA COLOMBO, PAOLA FREDIANI AND ALESSANDRO GHIGI In fact we expect unlikely that gonality plays any role in this problem. the existence of Q ∈ I2(KC ) with image ρ(Q) a nondegenerate quadric on H 1(C, TC ). This would give the upper bound 3 2 (g − 1) for the dimension of a germ of a totally geodesic submanifold. On the other hand, the map ρ is injective. This can be deduced from [5, Cor. 3.4] or from Theorem 3.13, since the form η is non-zero. Therefore ρ(I2(KC )) is a linear system of quadrics of dimension (g−1)(g−4) on P(H 1(C, TC )) = P3g−4. This already gives an upper bound for the dimension of a submanifold Y as in Theorem 4.5. Indeed for any point [C] ∈ Y , the tangent space T[C]Y ⊂ H 1(C, TC ) is contained in the base locus of ρ(I2(KC)). This means that ρ(I2(KC )) is contained in the space of quadrics q ∈ S2H 0(C, 2KC ) that vanish on T[C]Y . This yields the bound 2 dim Y ≤ −1 +p32g2 − 40g + 1 2 . Nevertheless for any g ≥ 2 this bound is weaker than the one provided in Theorem 4.5. At any case the study of the base locus of the linear system ρ(I2(KC )) should clearly improve the understanding of totally geodesic sub- manifolds. If one could prove that the base locus is empty, one would rule out the existence of totally geodesic submanifolds passing through [C]. However to proceed in this direction it is probably necessary to better understand the form η. Remark 4.8. Observe that for g ≥ 5, the non existence of germs of totally geodesic hypersurfaces follows directly from Theorem 2.6 and [4, Thm. 6.1]. In fact if Y is a hypersurface in Mg, it passes through a non-trigonal curve [C]. Since PT[C]Y intersects the bicanonical curve in PH 1(TC ), T[C]Y con- tains a Schiffer variation ξx, for some x ∈ C. By [4, Thm. 6.1], there exists a quadric Q ∈ I2(KC ) such that µ2(Q)(x) 6= 0, so ρ(Q)(ξx ⊙ ξx) 6= 0 by Theorem 2.6. 5. Families of cyclic covers of the projective line 5.1. Let C be a curve of genus g ≥ 4 and let G be a subgroup of the group of automorphisms of C. This yields an inclusion of G in the mapping class group Γg [11]. So G acts on the Teichmuller space Tg and we denote by T G g the set of fixed points of G on Tg, which is a nonempty by the solution of Nielsen realization problem [14, 27, 3], and is a complex submanifold of Tg. The tangent space to T G g at a point [C] is given by H 1(C, TC )G. Moreover T G g parametrizes marked curves C endowed with a holomorphic action of G of the given topological type and there is a universal family C → T G g containing all such curves. If t ∈ T G g we denote by Ct the corresponding curve. 5.2. Let us now consider the period map at the level of Teichmuller spaces Tg → Hg. This is an immersion outside the hyperelliptic locus, and we will ON TOTALLY GEODESIC SUBMANIFOLDS IN THE JACOBIAN LOCUS 15 consider its restriction to T G point t ∈ T G g the exact sequence g . Denote by Z(G) its image in Ag. Given a such that Ct is non-hyperelliptic, the cotangent spaces fit in 0 → N ∗ → S2H 0(KCt ) π→ H 0(2KCt )G → 0, where N ∗ is the conormal space to Z(G) ⊂ Ag at the point J(Ct). Since π is G-invariant, N ∗ is a G-submodule. Let ρ : N ∗ → H 0(2KCt )G ⊗ H 0(2KCt )G denote the second fundamental form of Z(G) at point J(Ct). Lemma 5.3. The second fundamental form ρ is G-equivariant. of the family C Proof. Recall that by definition, given a ∈ N ∗, we have ρ(a) = π(∇(a)), f where ∇ is the Gauss-Manin connection on S2f∗ωC/T G → T G g . Since π is G-invariant, it suffices to prove that ∇ is G-invariant, i.e. g−1∇g = ∇ for every g ∈ G. Observe that the flat connection ∇GM on R1f∗C is G-invariant, since the G–action maps flat sections to flat sections. Let π1,0 : H 1(Ct, C) → H 1,0(Ct) be the projection. Then π1,0 ◦ ∇GM is the . Since the action of G on C is connection on the Hodge bundle f∗ωC/T G holomorphic, the projection π1,0 is G-equivariant, hence π1,0 ◦ ∇GM is also a G-invariant connection. Finally ∇ is the connection induced by π1,0 ◦ ∇GM on S2f∗ωC/T G (cid:3) . The result follows. g g g Proposition 5.4. If there are no nonzero quadrics in I2(KCt ), which are invariant under the action of the group G, then Z(G) is totally geodesic. Proof. Consider the cotangent sequence of the period map j : Tg → Hg at the point [Ct] and its restriction to T G g : (5.1) 0 0 ........ ...................................................... . . ... I2(KCt ) ........ ................................................................ . . ... S2H 0(KCt) .m ... ........ ........................................................................ . . . . . . ............. .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. .. . . . . . . . . . . . . . . . . . .. .. .. .. .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. .. . . . . . . . . . . . . . . . . . .. .. .. .. .. = ........ ........................................................................ . ... . H 0(2KCt ) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. .. . . . . . . .. .. . . . . . . . . . . . . . . .. .. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. . .. .. .. .. .. .. .. ........ ................................................................................. . ... .. N ∗ ........ ........................................................................................... . ... .. S2H 0(KCt) ........ ............................................................. . ... .. H 0(2KCt )G ........ ............................................................. . ... .. 0 0. (The notation is as in 5.2.) Since the map m is G-equivariant, any G- invariant element in N ∗ lies in I2(KCt ). Hence by the assumption there are no nontrivial invariant elements in N ∗, i.e. the trivial representation does not appear in the decomposition of N ∗. On the other hand the rep- resentation H 0(2KCt )G is trivial. By Lemma 5.3 ρ is a morphism of G- representations. Hence Schur lemma implies that ρ is the trivial map. (cid:3) 5.5. Now we will consider the case in which G = Z/mZ, m ≥ 3 and C/G ∼= P1. These families have been studied by various authors, e.g. [7, 21, 25], since they provide the only known examples of totally geodesic submanifolds contained in the Jacobian locus. These examples are in fact 16 ELISABETTA COLOMBO, PAOLA FREDIANI AND ALESSANDRO GHIGI Shimura varieties. A complete list of the Shimura varieties that can be ob- tained in this way has been given in [21], which we follow for the notation in the rest of the paper. that gcd(m, a1, ..., aN ) = 1, ai 6≡ 0 mod m and PN We identify G also with the group of m-th roots of unity. Fix an integer N ≥ 4, together with an N -tuple a = (a1, ..., aN ) of positive integers, such i=1 ai ≡ 0 mod m. Given distinct points t1, ..., tN ∈ P1 there is a well-defined curve Ct which is a cyclic cover of P1 with covering group Z/mZ, branch points ti and local monodromy ai at ti. It is the normalization of the affine curve (5.2) N ym = (x − ti)ai . Yi=1 Varying the branch points t = (t1, . . . , tN ) one obtains a (N −3)-dimensional family of curves C → B. There is an action of G on C given by the rule ζ · (x, y, t) := (x, ζ · y, t), ζ ∈ G. Thus triples (m, N, a) parametrize these families and two triples (m, N, a) and (m′, N ′, a′) yield equivalent families if and only if m = m′, N = N ′ and the classes of a and a′ in (Z/mZ)N are in the same orbit under the action of (Z/mZ)∗ × ΣN , where (Z/mZ)∗ acts diagonally by multiplication and the symmetric group ΣN acts by permu- tation of the indices. Notice that to fix the class of the triple (m, N, a) is equivalent to fixing the topological type of the G-action. Thus B and T G g have the same image in Mg and we will consider Z(m, N, a) := Z(G) ⊂ Ag the (N − 3)-dimensional subvariety given by the Jacobians of the curves Ct, t ∈ B. By the Hurwitz formula g = 1 + (N − 2)m −PN 2 i=1 ri := gcd(m, ai). A basis of H 0(Ct, KCt ) is given as follows. For with ri i ∈ {1, ..., N } and n ∈ Z set l(i, n) := h −nai m i dn = −1 + N Xi=1D −nai m E (Here [a] and hai denote the integral and fractional parts of a ∈ R.) Since G acts on Ct, there is a decomposition H 0(Ct, KCt ) = ⊕m−1 n=0 Vn, where Vn is the subspace of 1-forms ω such that ζ · ω = ζ nω. Then V0 = {0}, while for n = 1, . . . , m − 1 a basis of Vn is provided by the forms that have the following expression in the model (5.2): (5.3) ωn,ν := yn · (x − t1)ν · N (x − ti)l(i,n) · dx Yi=1 0 ≤ ν ≤ dn − 1. Remark 5.6. In [21], Moonen proved that there is a finite list of triples (m, N, a) such that the corresponding subvariety Z = Z(m, N, a) ⊂ Ag is a Shimura variety. By [20] a Shimura variety is a totally geodesic algebraic submanifold of Ag that contains a complex multiplication point. Therefore ON TOTALLY GEODESIC SUBMANIFOLDS IN THE JACOBIAN LOCUS 17 the families in Moonen’s list are totally geodesic. Using Proposition 5.4 we can immediately verify that all these families are totally geodesic. In fact in all those cases there are no nontrivial quadrics in I2(KCt ) which are invariant under the action of the cyclic group. Proposition 5.7. Consider the family C → B associated to the triple (m, N, a) as above and the corresponding subvariety Z = Z(m, N, a) ⊂ Ag given by the Jacobians of the curves Ct. Assume that not all the curves Ct are hyperelliptic. If there exists an integer n ∈ {1, ..., m − 1} such that dn ≥ 2, dm−n ≥ 2 and n 6= m − n, then Z is not totally geodesic. Proof. Fix an arbitrary non-hyperelliptic curve C = Ct belonging to the family. We use the representation (5.2). Since dn ≥ 2 and dm−n ≥ 2 and n 6= m − n, there are four distinct forms ωn,0, ωn,1, ωm−n,0, ωm−n,1 given in (5.3). We form the quadric Q := ωn,0 ⊙ ωm−n,1 − ωn,1 ⊙ ωm−n,0. One immediately sees from the definition that Q ∈ I2(KC) and that Q is G-invariant. Let D be the divisor of poles of the meromorphic function x ∈ M(C). Let σ0, σ1 ∈ H 0(C, OC (D)) be the sections corresponding to the meromorphic functions 1 and x. Assume for simplicity that t1 = 0, so we have ωn,1 = x · ωn,0 and ωm−n,1 = x · ωm−n,0. Hence the forms ωn,0, ωm−n,0 can be seen as elements in H 0(C, KC (−D)). Set τ0 := ωn,0, τ1 := ωm−n,0. The quadric Q can be written as follows: (5.4) Q = σ0τ0 ⊙ σ1τ1 − σ0τ1 ⊙ σ1τ0. Denote by ϕ : C → P1 our covering: it corresponds to the pencil h1, xi = hσ0, σ1i. Denote by ψ the map to P1 induced by the other pencil hτ0, τ1i. Take a point p ∈ C that does not belong to the set Crit(ϕ) ∪ G · Crit(ψ) ∪ B, where B is the base locus of ψ. Fix a nonzero vector u ∈ TpC. The vector v := Pg∈G ξdg(u) ∈ H 1(C, TC ) is clearly G-invariant. So it is a tangent vector to T G g at the point corresponding to C. Hence by diagram (5.1) we have ρ(Q)(v ⊙ v) = ρ(Q)(v ⊙ v). By Lemma (3.11) the map ρ is also G-equivariant. So, using Theorem 2.6 we get ρ(Q)(v ⊙ v) = ρ(Q)(v ⊙ v) = Xg1,g2∈G ρ(Q)(ξdg1(u) ⊙ ξdg2(u)) = = G ·(cid:18)Xg6=1 = −2πim ·(cid:18)2Xg6=1 ρ(Q)(ξu ⊙ ξdg(u)) − 2πiµ2(Q)(p)(cid:19) Q(u, dg(u)) · ηu(dg(u)) + µ2(Q)(p)(cid:19). By (5.4) Q is the quadric associated to the pencils hσ0, σ1i and hτ0, τ1i, cor- responding to the maps ϕ and ψ respectively. Since the fibre of ϕ containing 18 ELISABETTA COLOMBO, PAOLA FREDIANI AND ALESSANDRO GHIGI p is the orbit G · p, Q(u, g∗u) = 0, forall g ∈ G − {1}. So finally we get ρ(Q)(v ⊙ v) = −2πim · µ2(Q)(p), and µ2(Q)(p) 6= 0 by our choice of p. Hence Z is not totally geodesic. (cid:3) Corollary 5.8. For any m, there is only a finite number of families which can be totally geodesic. Proof. By Proposition 5.7, if there exists an integer n ∈ {1, ..., m − 1} such that dn, dm−n ≥ 2, n 6= n − m, then the family is not totally geodesic. So we can assume that for all n ∈ {1, ..., m − 1} either dn ≤ 1, or dm−n ≤ 1. In particular either d1 ≤ 1, or dm−1 ≤ 1. Denote by Ni the cardinality of the i=1 ai ≡ 0 mod m i=1 Ni = N and the relation PN i=1 iNi ≡ 0 mod m. But set {j aj = i}. Then Pm−1 becomes Pm−1 N Xi=1 D −lai m E = −1 + m−1 Xi=1 NiD −li m E, for l = 1, . . . , m − 1, dl = −1 + so d1 = −1 + m−1 Xi=1 m − i m Ni, dm−1 = −1 + m−1 Xi=1 i m Ni. Hence for l ∈ {1, m − 1} we have dl ≥ −1 +Pm−1 m . So dl ≤ 1 implies N ≤ 2m. Hence only a finite number of families can be totally geodesic. (cid:3) Now we give some examples of computations for low degree (m = 3 and m = 5), showing that by the previous results most of the families are not totally geodesic. Ni m = −1 + N i=1 Corollary 5.9. If m = 3 and g ≥ 5, then the varieties Z(3, N, a) are not totally geodesic. If g = 4 there is only one totally geodesic family given by the triple (3, 6, a) where a = (1, 1, 1, 1, 1, 1). Proof. If d1, d2 ≥ 2, we know by Proposition 5.7 that the family is not totally geodesic. So we can assume that there exists n ∈ {1, 2} such that dn = 1. In this case the space S2H 0(KCt )(0) given by the invariant elements equals hωn,0 ⊙ V3−ni. So there are no nonzero invariant elements in I2(KCt ) and by Proposition 5.4 we conclude that the family is totally geodesic. So we have to show that g = 4, N = 6 and a = (1, 1, 1, 1, 1, 1). Denote as above by Ni the cardinality of the set {j aj = i}. Then N1 + N2 = N and N1 + 2N2 ≡ 0 mod 3. So N d1 = −1 + d2 = −1 + Xi=1 D −ai Xi=1 D −2ai 3 E = −1 + N1D −1 3 E = −1 + N1D −2 3 E + N2D −2 3 E + N2D −4 3 E = −1 + 3 E = −1 + N 2 3 N1 + 1 3 N2, 1 3 N1 + 2 3 N2. ON TOTALLY GEODESIC SUBMANIFOLDS IN THE JACOBIAN LOCUS 19 By Hurwitz formula g = N −2. Since we are assume g ≥ 4 we get N1 +N2 = N ≥ 6. If d1 = 1, 6 ≤ N1 + N2 ≤ 2N1 + N2 = 6, so N = 6, g = 4, N1 = 0, N2 = 6, hence a = (2, 2, 2, 2, 2, 2) which is equivalent to (1, 1, 1, 1, 1, 1). If d2 = 1 we have 6 ≤ N1 + N2 ≤ N1 + 2N2 = 6, so N = 6, g = 4, N2 = 0, N1 = 6, a = (1, 1, 1, 1, 1, 1). (cid:3) Remark 5.10. If m = 5, the following families are totally geodesic (see the list in [21]): (1) g = 4, N = 4, a = (1, 3, 3, 3), (2) g = 6, N = 5, a = (2, 2, 2, 2, 2). (1) is the family constructed by de Jong-Noot [7]. Applying the criterion given in Proposition 5.7 we are able to prove that all other families are not totally geodesic except possibly for the following 4 cases (3) g = 4, N = 4, a = (1, 1, 4, 4), (4) g = 4, N = 4, a = (1, 2, 3, 4), (5) g = 6, N = 5, a = (1, 1, 1, 3, 4), (6) g = 6, N = 5, a = (1, 1, 2, 2, 4), (7) g = 8, N = 6, a = (1, 1, 2, 2, 2, 2). (3) is contained in the hyperelliptic locus, see [21, (5.7)]. (4) has been studied in detail in [28]. Since one can check that it contains a CM point, it is not totally geodesic by Moonen’s classification. It would be interesting to get a complete list of the families of cyclic coverings which are totally geodesic and to compare it with the list in [21]. References [1] F. Andreatta. Coleman-Oort’s conjecture for degenerate irreducible curves. Israel J. Math., 187:231–285, 2012. [2] E. Arbarello, M. Cornalba, and P. A. Griffiths. Geometry of algebraic curves. Vol- ume II, volume 268 of Grundlehren der Mathematischen Wissenschaften. Springer, Heidelberg, 2011. With a contribution by Joseph Daniel Harris. [3] F. Catanese. Fibred surfaces, varieties isogenous to a product and related moduli spaces. Amer. J. Math., 122(1):1–44, 2000. [4] E. Colombo and P. Frediani. Some results on the second Gaussian map for curves. Michigan Math. J., 58(3):745–758, 2009. [5] E. Colombo and P. Frediani. Siegel metric and curvature of the moduli space of curves. Trans. Amer. Math. Soc., 362(3):1231–1246, 2010. [6] E. Colombo, G. P. Pirola, and A. Tortora. Hodge-Gaussian maps. Ann. Scuola Norm. Sup. Pisa Cl. Sci. (4), 30(1):125–146, 2001. [7] J. de Jong and R. Noot. Jacobians with complex multiplication. In Arithmetic alge- braic geometry (Texel, 1989), volume 89 of Progr. Math., pages 177–192. Birkhauser Boston, Boston, MA, 1991. [8] J. de Jong and S.-W. Zhang. Generic abelian varieties with real multiplication are not Jacobians. In Diophantine geometry, volume 4 of CRM Series, pages 165–172. Ed. Norm., Pisa, 2007. 20 ELISABETTA COLOMBO, PAOLA FREDIANI AND ALESSANDRO GHIGI [9] B. Farb and H. Masur. Superrigidity and mapping class groups. Topology, 37(6):1169– 1176, 1998. [10] H. M. Farkas and I. Kra. Riemann surfaces, volume 71 of Graduate Texts in Mathe- matics. Springer-Verlag, New York, second edition, 1992. [11] G. Gonz´alez D´ıez and W. J. Harvey. Moduli of Riemann surfaces with symmetry. In Discrete groups and geometry (Birmingham, 1991), volume 173 of London Math. Soc. Lecture Note Ser., pages 75–93. Cambridge Univ. Press, Cambridge, 1992. [12] M. L. Green. Infinitesimal methods in Hodge theory. In Algebraic cycles and Hodge theory (Torino, 1993), volume 1594 of Lecture Notes in Math., pages 1–92. Springer, Berlin, 1994. [13] R. Hain. Locally symmetric families of curves and Jacobians. In Moduli of curves and abelian varieties, Aspects Math., E33, pages 91–108. Friedr. Vieweg, Braunschweig, 1999. [14] S. P. Kerckhoff. The Nielsen realization problem. Ann. of Math. (2), 117(2):235–265, 1983. [15] S. Kukulies. On Shimura curves in the Schottky locus. J. Algebraic Geom., 19(2):371– 397, 2010. [16] K. Lamotke. Riemannsche Flachen. Berlin: Springer, 2005. [17] N. Mok. On the Zariski closure of a germ of totally geodesic complex submanifold on a subvariety of a complex hyperbolic space form of finite volume. In Complex analysis, Trends Math., pages 279–300. Birkhauser/Springer Basel AG, Basel, 2010. [18] M. Moller. Shimura and Teichmuller curves. J. Mod. Dyn., 5(1):1–32, 2011. [19] M. Moller, E. Viehweg, and K. Zuo. Special families of curves, of abelian varieties, and of certain minimal manifolds over curves. In Global aspects of complex geometry, pages 417–450. Springer, Berlin, 2006. [20] B. Moonen. Linearity properties of Shimura varieties. I. J. Algebraic Geom., 7(3):539– 567, 1998. [21] B. Moonen. Special subvarieties arising from families of cyclic covers of the projective line. Doc. Math., 15:793–819, 2010. [22] Moonen, B., Oort, F., The Torelli locus and special subvarieties. in Handbook of Moduli: Volume II edited by G. Farkas, I. Morrison, International Press, 2013, pp. 549–94. [23] F. Oort and J. Steenbrink. The local Torelli problem for algebraic curves. In Journ´ees de G´eometrie Alg´ebrique d’Angers, Juillet 1979/Algebraic Geometry, Angers, 1979, pages 157–204. Sijthoff & Noordhoff, Alphen aan den Rijn, 1980. [24] G. P. Pirola. The infinitesimal variation of the spin abelian differentials and periodic minimal surfaces. Comm. Anal. Geom., 6(3):393–426, 1998. [25] J. C. Rohde. Cyclic coverings, Calabi-Yau manifolds and complex multiplication, vol- ume 1975 of Lecture Notes in Mathematics. Springer-Verlag, Berlin, 2009. [26] D. Toledo. Nonexistence of certain closed complex geodesics in the moduli space of curves. Pacific J. Math., 129(1):187–192, 1987. [27] A. J. Tromba. Dirichlet’s energy on Teichmuller’s moduli space and the Nielsen real- ization problem. Math. Z., 222(3):451–464, 1996. [28] B. van Geemen and M. Schutt. Two moduli spaces of abelian fourfolds with an automorphism of order five. Internat. J. Math., 23(10):1250108, 31, 2012. [29] C. Voisin. Th´eorie de Hodge et g´eom´etrie alg´ebrique complexe, volume 10 of Cours Sp´ecialis´es. Soci´et´e Math´ematique de France, Paris, 2002. Universit`a di Milano E-mail address: [email protected] Universit`a di Pavia E-mail address: [email protected] ON TOTALLY GEODESIC SUBMANIFOLDS IN THE JACOBIAN LOCUS 21 Universit`a di Milano Bicocca E-mail address: [email protected]
1303.1043	3	1303	2015-03-18T18:56:21	Relations on Mbar_{g,n} via 3-spin structures	[ "math.AG" ]	Witten's class on the moduli space of 3-spin curves defines a (non-semisimple) cohomological field theory. After a canonical modification, we construct an associated semisimple CohFT with a non-trivial vanishing property obtained from the homogeneity of Witten's class. Using the classification of semisimple CohFTs by Givental-Teleman, we derive two main results. The first is an explicit formula in the tautological ring of Mbar_{g,n} for Witten's class. The second, using the vanishing property, is the construction of relations in the tautological ring of Mbar_{g,n}. Pixton has previously conjectured a system of tautological relations on Mbar_{g,n} (which extends the established Faber-Zagier relations on M_g). Our 3-spin construction exactly yields Pixton's conjectured relations. As the classification of CohFTs is a topological result depending upon the Madsen-Weiss theorem (Mumford's conjecture), our construction proves relations in cohomology. The study of Witten's class and the associated tautological relations for r-spin curves via a parallel strategy will be taken up in a following paper.	math.AG	math	Relations on Mg,n via 3-spin structures Rahul Pandharipande, Aaron Pixton, Dimitri Zvonkine March 2015 Abstract Witten’s class on the moduli space of 3-spin curves defines a (non- semisimple) cohomological field theory. After a canonical modifica- tion, we construct an associated semisimple CohFT with a non-trivial vanishing property obtained from the homogeneity of Witten’s class. Using the classification of semisimple CohFTs by Givental-Teleman, we derive two main results. The first is an explicit formula in the tautological ring of Mg,n for Witten’s class. The second, using the vanishing property, is the construction of relations in the tautological ring of Mg,n. Pixton has previously conjectured a system of tautological rela- tions on Mg,n (which extends the established Faber-Zagier relations on Mg). Our 3-spin construction exactly yields Pixton’s conjectured relations. As the classification of CohFTs is a topological result de- pending upon the Madsen-Weiss theorem (Mumford’s conjecture), our construction proves relations in cohomology. The study of Witten’s class and the associated tautological relations for r-spin curves via a parallel strategy will be taken up in a following paper. Contents 0 Introduction 1 Ar−1 and the shifted Witten class 2 The R-matrix action 3 The R-matrix for A2 1 2 13 18 31 0 Introduction 0.1 Overview The study of relations in the cohomology of the moduli space of curves was initiated by Mumford [13] in the 1980s. While several classical approaches were applied with success before, the subject has developed rapidly in the last two decades via natural connections to topological string theory. A systematic study by Faber and Zagier of the algebra of κ classes on the moduli space Mg of nonsingular genus g curves led to a conjecture in 2000 of a concise set FZ of κ relations. A proof of the Faber-Zagier conjecture (in Chow) via the geometry of stable quotients was given in 2010 [14]. In 2012, the second author [16] conjectured a set P of tautological relations for the moduli spaces Mg,n of stable curves. The set P recovers FZ when restricted to Mg ⊂ Mg. Our main result proves the conjectured relations P in the cohomology ring H∗(Mg,n, Q). By restriction, we obtain a second proof of the Faber- Zagier conjecture in cohomology. Are there other relations? The sets FZ and P explain all presently known tautological relations on Mg and Mg,n respectively. At least in Chow, the sets FZ and P are conjectured to be complete in both cases [14, 16]. We study here the geometry of 3-spin curves. Witten’s class on the moduli space of 3-spin curves defines a non-semisimple cohomological field theory. After a canonical modification (obtained by moving to a semisimple point of the associated Frobenius manifold), we construct a semisimple CohFT with a non-trivial vanishing property obtained from the homogeneity of Witten’s class. Using the classification of semisimple CohFTs by Givental-Teleman [6, 20], we derive an explicit formula in the tautological ring of Mg,n for Witten’s 3-spin class and use the vanishing property to establish the relation set P. 0.2 Stable graphs The boundary strata of the moduli space of curves correspond to stable graphs Γ = (V, H, L, g : V → Z≥0, v : H → V, ι : H → H) satisfying the following properties: 2 (i) V is a vertex set with a genus function g : V → Z≥0, (ii) H is a half-edge set equipped with a vertex assignment v : H → V and an involution ι, (iii) E, the edge set, is defined by the 2-cycles of ι in H (self-edges at vertices are permitted), (iv) L, the set of legs, is defined by the fixed points of ι and endowed with a bijective correspondence with a set of markings, (v) the pair (V, E) defines a connected graph, (vi) for each vertex v, the stability condition holds: 2g(v) − 2 + n(v) > 0, where n(v) is the valence of Γ at v including both edges and legs. An automorphism of Γ consists of automorphisms of the sets V and H which leave invariant the structures g, ι, and v (and hence respect E and L). Let Aut(Γ) denote the automorphism group of Γ. The genus of a stable graph Γ is defined by: g(Γ) = g(v) + h1(Γ). (cid:88) v∈V (cid:89) v∈V A boundary stratum of the moduli space Mg,n of Deligne-Mumford stable curves naturally determines a stable graph of genus g with n legs by consid- ering the dual graph of a generic pointed curve parametrized by the stratum. To each stable graph Γ, we associate the moduli space MΓ = Mg(v),n(v). Let πv denote the projection from MΓ to Mg(v),n(v) associated to the vertex v. There is a canonical morphism ξΓ : MΓ → Mg,n (1) with image1 equal to the boundary stratum associated to the graph Γ. To construct ξΓ, a family of stable pointed curves over MΓ is required. Such a family is easily defined by attaching the pull-backs of the universal families over each of the Mg(v),n(v) along the sections corresponding to half-edges. 1The degree of ξΓ is Aut(Γ). 3 0.3 Strata algebra Let Γ be a stable graph. A basic class on MΓ is defined to be a product of monomials in κ classes2 at each vertex of the graph and powers of ψ classes at each half-edge (including the legs), γ = κi[v]xi[v] ψy[h] h ∈ H∗(MΓ, Q) , · (cid:89) h∈H (cid:89) (cid:89) (cid:88) v∈V i>0 h∈H[v] where κi[v] is the ith kappa class on Mg(v),n(v). We impose the condition ixi[v] + y[h] ≤ dimC Mg(v),n(v) = 3g(v) − 3 + n(v) (cid:88) i>0 at each vertex to avoid the trivial vanishing of γ. Here, H[v] ⊂ H is the set of half-edges (including the legs) incident to v. Consider the Q-vector space Sg,n whose basis is given by the isomorphism classes of pairs [Γ, γ], where Γ is a stable graph of genus g with n legs and γ is a basic class on MΓ. Since there are only finitely many pairs [Γ, γ] up to isomorphism, Sg,n is finite dimensional. A product on Sg,n is defined by intersection theory with respect to the morphisms (1) to Mg,n. Let [Γ1, γ1], [Γ2, γ2] ∈ Sg,n be two basis elements. The fiber product of ξΓ1 and ξΓ2 over Mg,n is canon- ically described as a disjoint union of ξΓ for stable graphs Γ endowed with contractions3 onto Γ1 and Γ2. More precisely, the set of edges E of Γ should be represented as a union of two (not necessarily disjoint) subsets, E = E1 ∪ E2, in such a way that Γ1 is obtained by contracting all the edges outside E1 and Γ2 is obtained by contracting all edges outside E2 (see Proposition 9 in the 2Our convention is κi = π∗(ψi+1 n+1) ∈ H 2i(Mg,n, Q) where π : Mg,n+1 → Mg,n is the map forgetting the marking n + 1. For a review of κ and and cotangent ψ classes, see [8]. 3If there are several different pairs of contractions from a given Γ, the corresponding ξΓ appears with multiplicity. 4 Appendix of [8]). The intersection of ξΓ1 and ξΓ2 in Mg,n is then canonically given by Fulton’s excess theory as a sum of elements in Sg,n. We define [Γ1, γ1] · [Γ2, γ2] = [Γ, γ1γ2εΓ] (cid:88) (cid:89) Γ −(ψ(cid:48) e + ψ(cid:48)(cid:48) e ) e∈E1∩E2 where εΓ = is the excess class. Here, ψ(cid:48) sponding to the two half-edges of the edge e. e and ψ(cid:48)(cid:48) e are the two cotangent line classes corre- A case of particular importance for us is when Γ2 has only a single edge. The set E2 must consist of a single element e, while E1 may be either E or E \ {e}. The above product then yields the restriction of a basic class to a boundary divisor. Via the above intersection product, Sg,n is a finite dimensional Q-algebra, called the strata algebra [16]. Push-forward along ξΓ defines a canonical ring homomorphism q : Sg,n → H∗(Mg,n, Q), q([Γ, γ]) = ξΓ∗(γ) from the strata algebra to the cohomology ring. By definition, the image of q is the tautological ring RH∗(Mg,n). An element of the kernel of q is called a tautological relation. Each basis element [Γ, γ] has a degree grading given by the number of edges of Γ plus the usual (complex) degree of γ, Hence, Sg,n is graded, deg[Γ, γ] = E + degC(γ) . 3g−3+n(cid:77) S d g,n . Sg,n = Since the product respects the grading, Sg,n is a graded algebra. Of course, d=0 q : S d g,n → H 2d(Mg,n, Q) . 5 0.4 The tautological relations (cid:101)P We define a set(cid:101)P consisting of elements Rd g,n associated to the data g,A ∈ S d • g, n ∈ Z≥0 in the stable range 2g − 2 + n > 0, • A = (a1, . . . , an), ai ∈ {0, 1}, • d ∈ Z≥0 satisfying d > g−1+(cid:80)n The elements Rd g,A are expressed as sums over stable graphs of genus g with n legs. We prove in Section 3.5 that the conjectured family of relations P of [16] is implied in cohomology by the family of relations i=1 ai . 3 q(cid:0)Rd (cid:1) = 0 ∈ H 2d(Mg,n, Q) g,A ∈ (cid:101)P. Before writing the formula for Rd g,A for all Rd required. tautological relations on the moduli space Mg of nonsingular curves: The following two series first arose in the study by Faber and Zagier of g,A, a few definitions are (cid:88) (cid:88) m≥0 m≥0 B0(T ) = B1(T ) = (6m)! (2m)!(3m)! (−T )m = 1 − 60T + 27720T 2 − ··· , 1 + 6m 1 − 6m (6m)! (2m)!(3m)! (−T )m = 1 + 84T − 32760T 2 + ··· . in the set (cid:101)P. In the first proof of the Faber-Zagier relations [14], the above These series control the original set FZ and continue to play a central role series appeared via differential equations satisfied by the logarithm of ∞(cid:88) d(cid:89) d=0 i=1 1 1 − it (−1)d d! xd td , Φ(t, x) = see [14, Section 5] and [10]. Here we discover a completely different source for the series B0(T ) and B1(T ) via the homogeneous calibration of the Frobenius manifold associated to A2. Let f (T ) be a power series with vanishing constant and linear terms, f (T ) ∈ T 2Q[[T ]] . 6 For each Mg,n, we define (cid:88) m≥0 κ(f ) = 1 m! pm∗ (cid:16) f (ψn+1)··· f (ψn+m) (cid:17) ∈ H∗(Mg,n, Q), (2) where pm is the forgetful map pm : Mg,n+m → Mg,n. By the vanishing in degrees 0 and 1 of f , the sum (2) is finite. Let Gg,n be the (finite) set of stable graphs of genus g with n legs (up to isomorphism). Let Γ ∈ Gg,n. For each vertex v ∈ V, we introduce an auxiliary variable ζv and impose the conditions ζvζv(cid:48) = ζv(cid:48)ζv , ζ 2 v = 1 . The variables ζv will be responsible for keeping track of a local parity condi- tion at each vertex. Γ ∈ Gg,n is a product of vertex, leg, and edge factors: g,A is a sum over Gg,n. The summand corresponding to The formula for Rd • For v ∈ V, let κv = κ(cid:0)T − T B0(ζvT )(cid:1). (cid:0)ζv(l)ψl • For l ∈ L, let Bl = ζ al which the leg is assigned. v(l)Bal (cid:1), where v(l) ∈ V is the vertex to • For e ∈ E, let ∆e = ζ(cid:48) + ζ(cid:48)(cid:48) − B0(ζ(cid:48)ψ(cid:48))ζ(cid:48)(cid:48)B1(ζ(cid:48)(cid:48)ψ(cid:48)(cid:48)) − ζ(cid:48)B1(ζ(cid:48)ψ(cid:48))B0(ζ(cid:48)(cid:48)ψ(cid:48)(cid:48)) ψ(cid:48) + ψ(cid:48)(cid:48) = (60ζ(cid:48)ζ(cid:48)(cid:48) − 84) + [32760(ζ(cid:48)ψ(cid:48) + ζ(cid:48)(cid:48)ψ(cid:48)(cid:48)) − 27720(ζ(cid:48)ψ(cid:48)(cid:48) + ζ(cid:48)(cid:48)ψ(cid:48))] + ··· , where ζ(cid:48), ζ(cid:48)(cid:48) are the ζ-variables assigned to the vertices adjacent to the edge e and ψ(cid:48), ψ(cid:48)(cid:48) are the ψ-classes corresponding to the half-edges. The numerator of ∆e is divisible by the denominator due to the identity B0(T )B1(−T ) + B0(−T )B1(T ) = 2. Obviously ∆e is symmetric in the half-edges. 7 Definition 0.1 Let A = (a1, . . . , an) ∈ {0, 1}n. We denote by Rd the degree d component of the strata algebra class g,A ∈ S d g,n (cid:20) (cid:104)(cid:89) (cid:89) (cid:89) Γ, κv Bl ∆e (cid:21) (cid:105)(cid:81) ∈ Sg,n, v ζg(v)−1 v (cid:88) Γ∈Gg,n 1 Aut(Γ) 1 2h1(Γ) graph Γ. The subscript(cid:81) (cid:81) where the products are taken over all vertices, all legs, and all edges of the indicates the coefficient of the monomial v ζ g(v)−1 Definition 0.2 We denote by(cid:101)P the set of classes Rd after the product inside the brackets is expanded. v ζ g(v)−1 v v g,A where g − 1 +(cid:80)n g,A ∈(cid:101)P lies in the kernel of the homomorphism i=1 ai 3 . d > Theorem 1 Every element Rd q : Sg,n → H∗(Mg,n, Q) . As a formal consequence of Theorem 1, we will establish the originally conjectured set of relations P. Corollary 2 The full set P of relations conjectured in [16] holds in cohomol- ogy. Furthermore, we will identify Rd class for r = 3 when d = g−1+(cid:80) ai of Witten’s class under a forgetful map when d < g−1+(cid:80) ai g,(a1,...,an) as a simple multiple of Witten’s and a simple multiple of a push-forward . 3 3 0.5 Cohomological field theories We recall here the basic definitions of a cohomological field theory by Kont- sevich and Manin [11]. Let V be a finite dimensional Q-vector space with a non-degenerate sym- metric 2-form η and a distinguished element 1 ∈ V . The data (V, η, 1) is the starting point for defining a cohomological field theory. Given a basis {ei} of V , we write the symmetric form as a matrix ηjk = η(ej, ek) . 8 The inverse matrix is denoted by ηjk as usual. A cohomological field theory consists of a system Ω = (Ωg,n)2g−2+n>0 of elements Ωg,n ∈ H∗(Mg,n, Q) ⊗ (V ∗)⊗n. We view Ωg,n as associating a cohomology class on Mg,n to elements of V assigned to the n markings. The CohFT axioms imposed on Ω are: (i) Each Ωg,n is Sn-invariant, where the action of the symmetric group Sn permutes both the marked points of Mg,n and the copies of V ∗. (ii) Denote the basic gluing maps by q : Mg−1,n+2 → Mg,n , r : Mg1,n1+1 × Mg2,n2+1 → Mg,n . The pull-backs q∗(Ωg,n) and r∗(Ωg,n) are equal to the contractions of Ωg−1,n+2 and Ωg1,n1+1 ⊗ Ωg2,n2+1 by the bi-vector (cid:88) ηjkej ⊗ ek j,k inserted at the two identified points. (iii) Let v1, . . . , vn ∈ V be any vectors and let p : Mg,n+1 → Mg,n be the forgetful map. We require Ωg,n+1(v1 ⊗ ··· ⊗ vn ⊗ 1) = p∗Ωg,n(v1 ⊗ ··· ⊗ vn) , Ω0,3(v1 ⊗ v2 ⊗ 1) = η(v1, v2) . Definition 0.3 A system Ω = (Ωg,n)2g−2+n>0 of elements Ωg,n ∈ H∗(Mg,n, Q) ⊗ (V ∗)⊗n satisfying properties (i) and (ii) is a cohomological field theory or a CohFT. If (iii) is also satisfied, Ω is a CohFT with unit. 9 A CohFT Ω yields a quantum product • on V via η(v1 • v2, v3) = Ω0,3(v1 ⊗ v2 ⊗ v3) . Associativity of • follows from (ii). The element 1 ∈ V is the identity for • by (iii). A CohFT ω composed only of degree 0 classes, ωg,n ∈ H 0(Mg,n, Q) ⊗ (V ∗)⊗n , is called a topological field theory. Via property (ii), ωg,n(v1, . . . , vn) is deter- mined by considering stable curves with a maximal number of nodes. Such a curve is obtained by identifying several rational curves with three marked points. The value of ωg,n(v1 ⊗ ··· ⊗ vn) is thus uniquely specified by the values of ω0,3 and by the quadratic form η. In other words, given V and η, a topological field theory is uniquely determined by the associated quantum product. 0.6 Witten’s r-spin class For every integer r ≥ 2, there is a beautiful CohFT obtained from Witten’s r-spin class. We review here the basic properties of the construction. The integer r is fixed once and for all. Let V be an (r − 1)-dimensional Q-vector space with basis e0, . . . , er−2, bilinear form ηab = (cid:104)ea, eb(cid:105) = δa+b,r−2 , and unit vector 1 = e0. Witten’s r-spin theory provides a family of classes Wg,n(a1, . . . , an) ∈ H∗(Mg,n, Q). for a1, . . . , an ∈ {0, . . . , r − 2}. These define a CohFT by Wg,n : V ⊗n → H∗(Mg,n, Q), Wg,n(ea1 ⊗ ··· ⊗ ean) = Wg,n(a1, . . . , an) . To emphasize r, we will often refer to V as Vr. Witten’s class Wg,n(a1, . . . , an) has (complex) degree given by the formula degC Wg,n(a1, . . . , an) = Dg,n(a1, . . . , an) (r − 2)(g − 1) +(cid:80)n (3) . i=1 ai = r 10 If Dg,n(a1, . . . , an) is not an integer, the corresponding Witten class vanishes. In genus 0, the construction was first carried out by Witten [21] using r- spin structures (rth roots of the canonical bundle) and satisfies the following initial conditions: (cid:12)(cid:12)(cid:12)(cid:12) 1 0 W0,3(a1, a2, a3) = if a1 + a2 + a3 = r − 2, otherwise. (4) W0,4(1, 1, r − 2, r − 2) = [point] ∈ H 2(M 0,4, Q) . 1 r Uniqueness of Witten’s r-spin theory in genus 0 follows easily from the initial conditions (4) and the axioms of a CohFT with unit. The genus 0 sector defines a quantum product • on V with unit e0, (cid:104)ea • eb, ec(cid:105) = W0,3(a, b, c) . The resulting algebra, even after extension to C, is not semisimple. The existence of Witten’s class in higher genus is both remarkable and highly non-trivial. An algebraic construction was first obtained by Polishchuk and Vaintrob [17] defining Wg,n(a1, . . . , an) ∈ A∗(Mg,n, Q) as an algebraic cycle class. The algebraic approach was later simplified by Chiodo [2]. Analytic constructions have been given by Mochizuki [12] and later by Fan, Jarvis, and Ruan [9]. The equivalence between the above analytic and algebraic constructions was heretofore unknown. Theorem 3 For every r ≥ 2, there is a unique CohFT which extends Wit- ten’s r-spin theory in genus 0 and has pure dimension (3). The unique ex- tension takes values in the tautological ring RH∗(Mg,n) ⊂ H∗(Mg,n, Q). As a consequence of Theorem 3, the analytic and algebraic approaches coincide and yield tautological classes in cohomology. Our proof of Theorem 3 is not valid for Chow field theories as topological results play an essential role. 11 0.7 Strategy of proof on Vr, entirely determined by the genus 0 sector of W , Theorems 1 and 3 are proven together. Let Wg,n be any CohFT with unit which extends Witten’s r-spin theory in genus 0 and has pure dimension (3). We use a canonical procedure (a shift on the Frobenius manifold) to define a new CohFT (cid:101)Wg,n satisfying the following four properties: (i) (cid:101)W is canonically constructed from W with the genus 0 sector of (cid:101)W (ii) the quantum product associated to (cid:101)W0,3 defines a semisimple algebra (iii) the component of (cid:101)Wg,n(ea1⊗···⊗ean) in complex degree Dg,n(a1, . . . , an) (iv) the class (cid:101)Wg,n(ea1⊗···⊗ean) has no components in degrees higher than In other words, (cid:101)W is constructed from W by adding only lower degree terms. By the results of Givental and Teleman, (cid:101)W is determined via a universal erty (iii), we deduce a formula for (cid:101)W in the tautological ring depending only ified CohFT (cid:101)W in the 3-spin case. The series B0 and B1 appear in the formula in the tautological ring by the semisimple genus 0 sector. By prop- To prove Theorem 1, we write explicitly Givental’s formula for the mod- upon Witten’s r-spin theory in genus 0 and obtain Theorem 3. equals Wg,n(a1, . . . , an), Dg,n(a1, . . . , an). associated Frobenius structure. By property (iv), we obtain vanishings in the tautological ring in degrees g − 1 +(cid:80)n i=1 ai 3 d > Dg,n(a1, . . . , an) = The outcome is exactly the relations(cid:101)P. for r = 3. As a further outcome of the above investigation, we obtain the following Wg,n(a1, . . . , an) = 2g 1728d q(cid:0)Rd formula for Witten’s 3-spin class. Theorem 4 Let r = 3. Then, for g, n ∈ Z≥0 in the stable range, we have when d = g−1+(cid:80)n (cid:1) ∈ H 2d(Mg,n, Q) is integral (and Wg,n(a1, . . . , an) is 0 otherwise). g,(a1,...,an) i=1 ai 3 12 0.8 Plan of the paper In Section 1, we define the shifted Witten class for the r-spin theory. Theo- rem 3 is proven as a consequence of semisimplicity and Teleman’s uniqueness result. A short review of the R-matrix action on CohFTs is presented in Section 2. In Section 3, we compute the R-matrix for the 3-spin case and prove Theorems 1 and 4. The proof of Corollary 2 is also given in Section 3. The study of the R-matrix for higher r and the exploration of the asso- ciated relations in the tautological ring will be taken up in [15]. 0.9 Acknowledgments We are grateful to A. Chiodo, P. Dunin-Barkovsky, C. Faber, J. Gu´er´e, F. Janda, A. Polishchuk, O. Randal-Williams, Y. Ruan, L. Spitz, and A. Vain- trob for useful and detailed discussions. Several of the ideas presented here grew out of discussions at the Geometry and topology of moduli conference in Berlin in October 2012 organized by G. Farkas. Special thanks to S. Shadrin for pointing out an error in the first draft of the paper and to L. Meng (of Peking University) for pointing out a dropped parity factor in the leg term. R. P. was partially supported by the Swiss National Science Foundation grant SNF 200021143274. A. P. was supported by an NSF Graduate Research Fellowship and was a guest of the Forschungsinstitut fur Mathematik (FIM) for several visits to ETH Zurich. D. Z. was supported by the grant ANR-09- JCJC-0104-01. 1 Ar−1 and the shifted Witten class 1.1 Potentials Frobenius manifolds were introduced and studied in detail in Dubrovin’s monograph [3]. For a concise summary see [7, Section 1]. As for every CohFT, the genus 0 part of Witten’s r-spin class determines a Frobenius manifold structure on the underlying vector space Vr. For Wit- ten’s class, the Frobenius manifold coincides with the canonical Frobenius structure on the versal deformation of the Ar−1 singularity [4] up to a co- ordinate change. We will denote by t0, . . . , tr−2 the coordinates in the basis e0, . . . , er−2 of Vr. 13 The structure of a Frobenius manifold is governed by the Gromov-Witten potential. The genus 0 Gromov-Witten potential of Witten’s r-spin class (without descendants) is: W0,n(a1, . . . , an) ta1 ··· tan n! . F(t0, . . . , tr−2) = (cid:88) (cid:88) (cid:90) n≥3 a1,...,an M0,n We will refer to F as the primary genus 0 potential. Example 1.1 For r = 3, the primary genus 0 potential obtained from Wit- ten’s class equals F(x, y) = where x = t0 and y = t1. For r = 4, the potential is 1 2 x2y + y4, 1 72 F(x, y, z) = 1 2 x2z + 1 2 xy2 + 1 16 y2z2 + 1 960 z5, where x = t0, y = t1, and z = t2. The third derivatives of F determine an associative algebra structure (the quantum product) in each tangent space to the Frobenius manifold. Let ∂i denote the vector field on Vr associated to differentiation by ti. Then, (cid:88) k,l ∂i • ∂j = ∂3F ∂ti∂tj∂tk ηkl∂l . The algebra on tangent spaces is semisimple outside the discriminant of Ar−1. For instance, for r = 3, the discriminant is {y = 0}. Definition 1.2 Let τ ∈ Vr. We define the shifted Witten class by g,n(v1 ⊗ ··· ⊗ vn) = Wτ (pm)∗Wg,n+m(vn ⊗ ··· ⊗ vn ⊗ τ ⊗ ··· ⊗ τ ), 1 m! (cid:88) m≥0 where pm : Mg,n+m → Mg,n is the forgetful map. 14 Remark 1.3 We have the following degree bound: (cid:104) (cid:105) ≤ (g − 1)(r − 2) +(cid:80) ai + m(r − 2) (pm)∗Wg,n+m(ea1 ⊗ ··· ⊗ ean ⊗ τ ⊗ ··· ⊗ τ ) deg r = Dg,n(a1, . . . , an) − 2m r . − m The sum in Definition 1.2 is thus finite for any given g and a1, . . . , an. The shifted Witten class is therefore well-defined. Moreover, the highest degree term of the shifted Witten class is equal to the Witten class itself – all the other terms are of smaller degrees. Remark 1.4 Let F(t) and Fτ ((cid:98)t ) be the primary genus 0 potentials of W and Wτ respectively. By elementary verification, Fτ ((cid:98)t ) = F(τ +(cid:98)t ) − (terms of degree < 3). Proposition 1.5 The shifted Witten class Wτ is a CohFT with unit. The proof is a straightforward check, and in any case, is identical to the proof of Proposition 2.7 given in Section 2 below. 1.2 The Euler field A Frobenius manifold is called conformal if it carries an affine Euler field E, a vector field satisfying the following properties: (i) in flat coordinates ti, the field has the form (cid:88) i E = (αiti + βi) ∂ ∂ti , (ii) the quantum product •, the unit 1, and the metric η are eigenfunctions of the Lie derivative LE with weights 0, −1, and 2 − δ respectively. The rational number δ is called the conformal dimension of the Frobenius manifold. 15 For instance, on the Frobenius manifold Ar−1, an Euler field is given by E = ta ∂ ∂ta , r−2(cid:88) (cid:16) a=0 δ = (cid:17) 1 − a r r − 2 r . Remark 1.6 We follow here Givental’s conventions for the Euler field. In Teleman’s conventions, the Euler vector field and hence the eigenvalues of LE have the opposite sign. Let Ω be a CohFT and V the corresponding Frobenius manifold. Given an Euler field E on V , a natural action of E on Ω is defined as follows. Let deg : H∗(Mg,n, Q) → H∗(Mg,n, Q) be the operator which acts on H 2k by multiplication by k. As usual, ∂i is the vector field4 on V associated to differentiation by the coordinate ti. Then (cid:32) (cid:33) n(cid:88) (E.Ω)g,n(∂i1 ⊗ ··· ⊗ ∂in) = (cid:16) ∂i1 ⊗ ··· ⊗ ∂in ⊗(cid:88) (cid:17) , βi∂i αil deg + Ωg,n(∂i1 ⊗ ··· ⊗ ∂in) + p∗Ωg,n+1 where p : Mg,n+1 → Mg,n is the forgetful map. l=1 Definition 1.7 A CohFT Ω is homogeneous if (E.Ω)g,n = [(g − 1)δ + n] Ωg,n for all g and n. Witten’s r-spin class is easily seen to be homogeneous. Indeed, we have Dg,n(a1, . . . , an) + (r − 2)(g − 1) +(cid:80) ai (cid:80) ai + n = (g − 1)δ + n. 4We will often use the canonical identification of V with the tangent space of 0 ∈ V . = (g − 1) + n − r r n(cid:88) (cid:16) i=1 (cid:17) = 1 − ai r r − 2 r 16 The underlying vector space Vr and basis e0, . . . , er−2 are the same for the CohFT obtained from the shifted Witten class. We denote the coordi- nates on Vr in the basis e0, . . . , er−2 for the shifted r-spin Witten theory by (cid:98)t 0, . . . ,(cid:98)t r−2. Proposition 1.8 The shifted Witten class is a homogeneous CohFT with Euler field r−2(cid:88) (cid:16) a=0 1 − a r (cid:17) (τ a +(cid:98)t a) ∂ ∂(cid:98)t a , E = of conformal dimension δ = r−2 r . Proof. Assume for simplicity τ = u∂b for some fixed b ∈ {0, . . . , r − 2}. Denote Wg,n+m(∂a1 ⊗ ··· ⊗ ∂an ⊗ ∂b ⊗ ··· ⊗ ∂b) here by just Wg,n+m. Then we have (cid:88) m≥0 um m! (E.Wτ )g,n(∂a1 ⊗ ··· ⊗ ∂an) = (cid:16) n(cid:88) (cid:40)(cid:20)(r − 2)(g − 1) +(cid:80) ai + mb (cid:19) − m (cid:18) (cid:21) i=0 + um m! u 1 − b r (pm+1)∗Wg,n+m+1 (cid:17)(cid:41) 1 − ai r (pm)∗Wg,n+m After simplifying, the above equals [(g − 1)δ + n] (pm)∗Wg,n+m r (cid:88) m≥0 + um m! −(cid:88) m≥1 (cid:88) m≥0 um (m − 1)! (cid:19) (cid:18) + 1 − b r (cid:88) m≥0 (pm)∗Wg,n+m (cid:18) (cid:19) um+1 m! 1 − b r (pm+1)∗Wg,n+m+1. The last two sums cancel each other, so we obtain [(g − 1)δ + n] (pm)∗Wg,n+m = [(g − 1)δ + n] Wτ g,n(∂a1 ⊗ ··· ⊗ ∂an). (cid:88) m≥0 um m! 17 The general case is similiar. ♦ Since the shifted r-spin Witten class is a homogeneous semisimple CohFT, we can apply the following theorem by C. Teleman [20, Theorem 1]. Theorem 5 (Teleman) Let Ω0,n be a genus 0 homogeneous semisimple Coh- FT with unit. The following results hold: (i) There exists a unique homogeneous CohFT with unit Ωg,n extending Ω0,n to higher genus. (ii) The extended CohFT Ωg,n is obtained by an R-matrix action on the topological (degree 0) sector of Ω0,n determined by Ω0,3. (iii) The R-matrix is uniquely specified by Ω0,3 and the Euler field. The unit-preserving R-matrix action in part (ii) of Theorem 5 will be re- viewed in Section 2, see Definition 2.13. We will compute the R-matrix for the 3-spin Witten class in Section 3. Since the shifted 3-spin Witten class (considered for all genera) is a homo- geneous CohFT with unit, the expressions obtained by the R-matrix action coincide with the shifted 3-spin Witten class. In particular, if we split the expression of the R-matrix action into pure degree parts, the parts of degree d > Dg,n(a1, . . . , an) vanish while the part of degree Dg,n(a1, . . . , an) coincides with Witten’s class, which proves Theorem 3. 2 The R-matrix action We present here a succinct but self-contained review of the R-matrix action on CohFTs. The action was first defined on Gromov-Witten potentials by Givental [6]. Its lifting to CohFTs was independently discovered by several authors: the papers by Teleman [20] and Shadrin [19] give an abbreviated treatment of the subject and refer to unpublished notes by Kazarian and by Katzarkov, Kontsevich, and Pantev. 18 2.1 The R-matrix action on CohFTs Let V be a vector space with basis {ei} and a symmetric bilinear form η. Consider the group of End(V )-valued power series R(z) = 1 + R1z + R2z2 + ··· (5) satisfying the symplectic condition, R(z)R∗(−z) = 1 , where R∗ is the adjoint with respect to η. Remark 2.1 Let Rk basis, j be the matrix form of an endomorphism R in the given (cid:1) = (cid:88) j,k Rk j tj ek . The symplectic condition in coordinates is Rj l (z)ηlsRk s (−z)ηku = δj u. After multiplying by η−1 on the right, we obtain an equivalent condition in bi-vector form l (z)ηlsRk s (−z) = ηjk . We conclude that the expression R(cid:0)tjej (cid:88) l,s,k Rj (cid:88) ηjk −(cid:80) l,s l (z)ηlsRk s (w) l,s Rj z + w is a well-defined power series in z and w. Associated to R(z) is the power series R−1(z) = 1 the symplectic condition.5 It follows that l (z)ηls(R−1)k l,s(R−1)j (cid:88) ηjk −(cid:80) j,k z + w 5By the symplectic condition, we have R−1(z) = R∗(−z). 19 R(z) which also satisfies s (w) ej ⊗ ek ∈ V ⊗2[[z, w]]. (6) We will denote the V ⊗2-valued power series (6) by η−1 − R−1(z)η−1R−1(w)t z + w for short (where the superscript t denotes matrix transpose). Let Ω = (Ωg,n)2g−2+n>0 be a CohFT on V , and let R be an element of the group (5). The CohFT RΩ is defined as follows. Definition 2.2 Let Gg,n be the finite set of stable graphs6 of genus g with n legs. For each Γ ∈ Gg,n, define a contribution ContΓ ∈ H∗(Mg,n, Q) ⊗ (V ∗)⊗n by the following construction: (i) place Ωg(v),n(v) at each vertex v of Γ, (ii) place R−1(ψl) at every leg l of Γ, (iii) at every edge e of Γ, place η−1 − R−1(ψ(cid:48) e)η−1R−1(ψ(cid:48)(cid:48) e )t ψ(cid:48) e + ψ(cid:48)(cid:48) e . Define (RΩ)g,n to be the sum of contributions of all stable graphs, (cid:88) Γ∈Gg,n (RΩ)g,n = 1 Aut(Γ) ContΓ . We use the inverse of the R-matrix in all of our formulas in Definition 2.2. There are two reasons for the seemingly peculiar choice. First, the result will be a left group action rather than a right group action on CohFTs. Second, the same convention is used by Givental and Teleman in their papers. A few remarks about Definition 2.2 are needed for clarification. By the symmetry property of CohFTs, the placement of Ωg(v),n(v) does not depend upon an ordering of the half-edges at v. At a leg l attached to a vertex v, we have R−1(ψl) ∈ H∗(Mg(v),n(v), Q) ⊗ End(V ). 6See Sections 0.2-0.4. 20 The first factor acts on the cohomology of the moduli space Mg(v),n(v) by multiplication. The endomorphism factor acts on the vectors which are “fed” to Ωg(v),n(v) at the legs. For an edge e attached to vertices v(cid:48) and v(cid:48)(cid:48) (possibly the same vertex), denote by Mg(cid:48),n(cid:48) and Mg(cid:48)(cid:48),n(cid:48)(cid:48) the corresponding moduli spaces. The insertion on e is an element of H∗(Mg(cid:48),n(cid:48), Q) ⊗ H∗(Mg(cid:48)(cid:48),n(cid:48)(cid:48), Q) ⊗ V ⊗2 e and w = ψ(cid:48)(cid:48) obtained by substituting z = ψ(cid:48) e in (6). Once again, the cohomol- ogy factors act on the corresponding cohomology spaces by multiplication. The bivector part is used to contract the two covectors sitting on the half- edges e(cid:48) and e(cid:48)(cid:48) in the corresponding CohFT elements at v(cid:48) and v. In the e )t, the bivector η−1 sits in the middle of the expression R−1(ψ(cid:48) edge, while the action of R−1 is directed from the middle of the edge towards the vertices. The similarity of Definition 2.2 with the form of the relations Rd e)η−1R−1(ψ(cid:48)(cid:48) g,A was the starting point of our paper. Proposition 2.3 If Ω is a CohFT, the system (RΩ)g,n is a CohFT. Proof. The symmetry of (RΩ) follows directly from the symmetry of Ω and the definition of the R-matrix action. Hence, we need only establish the pull-back property (ii) of Definition 0.3. Let Φ ∈ Gg,n be a stable graph with a single edge e. In order to compute the pull-back of tautological classes under ξΦ : MΦ → Mg,n , according to the rule given in Section 0.3, we must enumerate all stable graphs Γ with a distinguished edge e ∈ E(Γ) such that contracting all other edges yields the graph Γ2 = Φ. If E1 = E, we have Γ1 = Γ. The contribution of Γ to the pull-back is obtained from the contribution of Γ1 to RΩ after a multiplication by −(ψ(cid:48) e + ψ(cid:48)(cid:48) e ) . In other words, the contribution to the pull-back is obtained by placing the class R−1(ψ(cid:48) e)η−1R−1(ψ(cid:48)(cid:48) e )t − η−1 21 on the edge e with the usual insertions on all other edges and legs. If E1 = E \ {e} then Γ1 is obtained from Γ by contracting e. According to the CohFT rules for Ω, the contribution of Γ to the pull-back is obtained by placing η−1 on the edge e with the standard classes on all other edges. After summing, the total contribution of Γ to the pull-back is equivalent to placing R−1(ψ(cid:48) e) η−1 R−1(ψ(cid:48)(cid:48) e )t. on the edge e of Γ – precisely what we obtain by the CohFT rules applied to ♦ (RΩ)g,n. Proposition 2.4 The R-matrix action on CohFTs is a left group action. Proof. We must prove the action of Ra(z) followed by the action of Rb(z) is equal to the action of Rb(z)Ra(z). When we apply Ra(z), we sum over all stable graphs of type (g, n). Let us color the edges of these stable graphs in red. When we then apply Rb(z), we sum over all stable graphs of type (g, n), but now we replace each vertex of the stable graph by a small red graph. Let us color the edges of the large graph in blue. The result of the consecutive actions of Ra(z) and Rb(z) will be a sum over all stable graphs of type (g, n) whose edges are colored in red and blue. vertex and then R−1 On every leg l of the stable graph, we place first R−1 b (ψl) at the end of the leg. The final outcome is a (ψl) closer to the a (ψl)R−1 R−1 b (ψl) = (RbRa)−1(ψl) . (7) The result (7) is also what we place on a leg when we compute the action of the product RbRa. R(cid:48) = R(ψ(cid:48) Consider next an edge e of the stable graph. We will use the abbreviations e ). On a red edge, we have placed e) and R(cid:48)(cid:48) = R(ψ(cid:48)(cid:48) red edge: η−1 − (R(cid:48) a)−1η−1((R(cid:48) ψ(cid:48) e + ψ(cid:48)(cid:48) e a)−1)t via the first action. On a blue edge, on the other hand, we see R(cid:48) the ends of the edge and a and R(cid:48)(cid:48) a on η−1 − (R(cid:48) b)−1η−1((R(cid:48) ψ(cid:48) e + ψ(cid:48)(cid:48) e b)−1)t . 22 in the middle of the edge. The final outcome, after unwinding the definitions, is blue edge: (R(cid:48) a)−1η−1((R(cid:48) a)−1)t − (R(cid:48) a)−1(R(cid:48) e + ψ(cid:48)(cid:48) ψ(cid:48) e b)−1η−1((R(cid:48) b)−1)t((R(cid:48) a)−1)t . Since we are summing over all possible colorings, every edge in the stable graph will appear once in red and once in blue. The total contribution will be the sum of the contributions of the two colors, red + blue: η−1 − (R(cid:48) a)−1η−1((R(cid:48) bR(cid:48) ψ(cid:48) e + ψ(cid:48)(cid:48) e bR(cid:48) a)−1)t . The result is exactly what is placed on an edge when we compute the action ♦ of the product RbRa. 2.2 Action by translations Let Ω be a CohFT based on the vector space V , and let T (z) = T2z2 + T3z3 + ··· be a V -valued power series with vanishing coefficients in degrees 0 and 1. Definition 2.5 The translation of Ω by T is the CohFT T Ω defined by (cid:88) m≥0 = (cid:16) (T Ω)g,n(v1 ⊗ ··· ⊗ vn) v1 ⊗ ··· ⊗ vn ⊗ T (ψn+1) ⊗ ··· ⊗ T (ψn+m) (cid:17) , 1 m! (pm)∗Ωg,n+m where pm : Mg,n+m → Mg,n is the forgetful map. The use of T (ψi) as an argument in a CohFT is an abuse of notation. The result should be understood as Ωg,n(··· T (ψi)··· ) = i Ωg,n(··· Tk ··· ). ψk (cid:88) k≥2 Remark 2.6 The action by translations is very close to the shift of Defini- tion 1.2. However, unlike shifts, the translation action is always well-defined for degree reasons: the degree of the mth summand of the definition is at least m, so the sum is actually finite for any given g, n. 23 The action by translations can be described in terms of stable graphs. It is a summation over stable graphs with a single vertex and n + m legs for m ≥ 0. The first n legs carry the vectors v1, . . . , vn, and the last m legs carry the series T (ψi). The latter legs are then suppressed by a forgetful map. We will call the first n legs main legs and the last m legs κ-legs, since the push-forward of powers of ψ-classes gives rise to κ-classes. Proposition 2.7 If Ω is a CohFT, the system (T Ω)g,n is a CohFT. Proof. Let Φ be a stable graph with a single edge, and let MΦ the corre- sponding moduli space. We examine the pull-back of T Ω to MΦ. If Φ has a single vertex, then the κ-legs in the definition of T Ω just stay on this vertex. If Φ has two vertices, then the κ-legs are distributed among the two vertices. The automorphism coefficients match: there are (cid:0) m (cid:1) ways to distribute m1,m2 m κ-legs between two vertices, which leads to a coefficient (cid:18) m (cid:19) m1, m2 1 m! = 1 m1! 1 m2! . Thus, T Ω satisfies the axioms of a CohFT. ♦ Proposition 2.8 Translations form an abelian group action on CohFTs. Proof. The definition of (Ta + Tb)Ω contains the following sum (where we have suppressed the ψ-classes in the notation): (cid:88) (Ta + Tb)m = m≥0 m! (cid:88) (cid:88) m≥0 ma+mb=m a T mb T ma ma! mb! b = (cid:88) ma≥0 T ma a ma! (cid:88) mb≥0 T mb b mb! , ♦ which is the definition of the successive actions of Tb and Ta. Proposition 2.9 Let R(z) ∈ id + zEnd(V )[[z]] be an End(V )-valued power series satisfying the symplectic condition. Let Ta, Tb ∈ z2V [[z]] be two V - valued power series related to R by Ta(z) = R(z)Tb(z) . Then, for every CohFT Ω, we have TaRΩ = RTbΩ . 24 Remark 2.10 The equality of the proposition can be written more concisely as or RT Ω = (RT )RΩ RT R−1Ω = (RT )Ω. Thus the actions of R and T can be combined into an action of an affine group. Proof of Proposition 2.9. Both CohFTs TaRΩ and RTbΩ can be ex- pressed as sums over stable graphs. We will match the sums. Consider first TaRΩ. By definition, we start with n main legs marked by v1, . . . , vn and m κ-legs marked by Ta(ψn+1), . . . , Ta(ψn+m). We attach these legs to all possible stable graphs of genus g with n + m legs. Their vertices are marked with Ω, their legs with R−1(ψi), and their edges with η−1 − R−1(ψ(cid:48))η−1R−1(ψ(cid:48)(cid:48))t ψ(cid:48) + ψ(cid:48)(cid:48) . . The outcome is a sum over stable graphs of genus g with n + m legs whose n main legs are marked with R−1(ψi)(vi) and m κ-legs with (R−1Ta)(ψi) = Tb(ψi) . Consider next RTbΩ. We start with a sum over all stable graphs of genus g with n legs. Their vertices are marked with Ω, their legs with R−1(ψi), and their edges with η−1 − R−1(ψ(cid:48))η−1R−1(ψ(cid:48)(cid:48))t ψ(cid:48) + ψ(cid:48)(cid:48) Now we add to every vertex of such a graph an arbitrary number of κ-legs marked with Tb(ψi). The sum only runs over the graphs which remain stable when we remove the κ-legs. However, the summation can be extended to all stable graphs with n + m legs. Indeed, if a stable graph has a genus 0 vertex v with m κ-legs and less than 3 other half-edges, then the dimension of the moduli space assigned to v is less than m, while the degree of the class sitting on this moduli space is at least 2m. Thus the contribution of the graph vanishes. In conclusion, we obtain exactly the same sum as in the first case. The sums here are infinite (as m is unbounded), but only a finite number of terms are nonzero. The same issue arose in the definition of the translation ♦ action. 25 2.3 CohFTs with unit Let Ω be a CohFT with unit 1 ∈ V satisfying Ωg,n+1(v1 ⊗ ··· ⊗ vn ⊗ 1) = p∗Ωg,n(v1 ⊗ ··· ⊗ vn), Ω0,3(v1 ⊗ v2 ⊗ 1) = η(v1, v2) , where p : Mg,n+1 → Mg,n is the forgetful map. The R-matrix action and the translation action defined in the previous sections do not preserve the property of being a CohFT with unit. However, we will explain here how the two actions can be combined in a unique way so as to preserve the unit. We recall a well-known geometric result which we will implicitly use in the computations. Lemma 2.11 Consider the following commutative square of forgetful maps: Mg,n+k+m - Mg,n+m Pk Pm ? Mg,n+k pm pk - ? Mg,n The relation (pk)∗(pm)∗ = (Pm)∗(Pk)∗ holds in cohomology. Proof. Let X be the fiber product of pm and pk, with maps a : X → Mg,n+m , b : X → Mg,n+k , f : Mg,n+k+m → X . Then (pk)∗(pm)∗ = b∗a∗ is immediate, and also (Pm)∗(Pk)∗ = (b∗f∗)(f∗a∗) = b∗(f∗f∗)a∗ = b∗a∗ by birationality of f . ♦ The definition of the translation action involves push-forwards and the axioms of a CohFT with unit involve a pull-back. By the above Lemma, we will not have to worry about whether the pull-back is taken before or after the push-forward. 26 Proposition 2.12 Let Ω be a CohFT with unit 1 ∈ V . Let R(z) be an R-matrix satisfying the symplectic condition, and let Ta(z) = z · [R(1) − 1](z), Tb(z) = z · [1 − R−1(1)](z) be two elements of z2V [[z]]. Then, is also a CohFT with unit 1 ∈ V . TaRΩ = RTbΩ Proof. By Proposition 2.9, TaRΩ and RTbΩ are equal CohFTs. Hence, only the unit property must be verified. The CohFT RTbΩg,n(v1⊗···⊗vn) is expressed as a sum over stable graphs of genus g with n main legs and any number m ≥ 0 of κ-legs (see the proof of Proposition 2.9). Their vertices are marked with Ω, their main legs with R−1(ψi), their κ-legs with Tb(ψi), and their edges with η−1 − R−1(ψ(cid:48))η−1R−1(ψ(cid:48)(cid:48))t ψ(cid:48) + ψ(cid:48)(cid:48) . The expression is an infinite sum with only a finite number of nonzero terms. To calculate p∗RTbΩg,n(v1 ⊗ ··· ⊗ vn), we will study the pull-back under p∗ of the contribution of every stable graph Γ in the sum. Let us call the new leg marked n + 1 appearing on Mg,n+1 after the pull- back the special leg. The pull-back of a stable graph Γ is given by the stable graphs obtained by attaching the special leg to one of the vertices of Γ. The pull-backs of the stable graph contributions involve also the pull- backs of the cotangent line classes. The relation between the pulled-back ψ-classes p∗ψi from Mg,n+m in terms of the new classes ψi on Mg,n+m+1 is given by the well-known formula p∗(ψd h) = ψd h − ∆h,n+1 p∗(ψd−1 h ). (8) Here h is a half-edge of Γ and ∆h,n+1 is the divisor7 in the moduli space at the vertex carrying h corresponding to curves with a genus 0 component carrying only the markings h and n + 1. The pull-back of the contribution of Γ is given by a sum of two kinds of terms. A term of the first kind is obtained by attaching the special leg to 7The divisor is empty unless h and the special leg are on the same vertex. 27 one of the vertices of Γ and placing the class 1 on it. This happens if we choose the first term on the right side of (8) in the pull-back of each ψ-class. We use here also the original unit property of Ω. A term of the second kind is obtained by choosing a half-edge of Γ, placing a new vertex on it, and adding the special leg maked with 1 to this vertex. The power of the ψ-class that was written on this half-edge is then reduced by 1. This happens if we choose the second term of (8) in the pull-back of the ψ-class corresponding to the chosen half-edge. A term of the second kind occurs with a minus sign. Let us look more closely at the terms of the second kind. Suppose we have placed the new vertex on the ith main leg. Then there are two legs attached to this vertex. vi ψn+1 ψi ψ(cid:48)(cid:48) ψ(cid:48) = p∗ψi First, the ith main leg carrying a vi is attached. We can replace the vi by R−1(vi)(ψi) since anyway ψi = 0 on the moduli space M0,3 corresponding to our vertex. Second, the special leg carrying 1 is attached. We can similarly replace 1 by R−1(1)(ψn+1). Finally, there is the edge connecting our vertex to the rest of the graph with a [vi − R−1(vi)](p∗ψi) p∗ψi (9) placed8 on it. We can replace the edge insertion with the standard one η−1 − R−1(ψ(cid:48))η−1R−1(ψ(cid:48)(cid:48))t ψ(cid:48) + ψ(cid:48)(cid:48) 8The minus sign in the second term of (8) is included in (9). 28 1since we have ψ(cid:48)(cid:48) = 0 and Ω0,3(vi ⊗ v(cid:48)(cid:48) i ⊗ 1) = η(vi, v(cid:48)(cid:48) i ). We have therefore obtained a stable graph with n + 1 legs marked precisely as in the definition of (RTbΩ)g,n+1(v1 ⊗ ··· ⊗ vn ⊗ 1). Next, suppose we have placed the new vertex on a half-edge. We group the terms obtained from the two half-edges of a single edge of Γ together. ψ(cid:48)(cid:48)(cid:48) = 0 ψ(cid:48) ψn+1 = 0 ψ(cid:48)(cid:48)(cid:48)(cid:48) = 0 The standard edge insertion for Γ is η−1 − R−1(ψ(cid:48))η−1R−1(ψ(cid:48)(cid:48))t ψ(cid:48) + ψ(cid:48)(cid:48) ψ(cid:48)(cid:48) . If we place the new vertex at first half-edge, we obtain (cid:20) η−1 − R−1(ψ(cid:48))η−1R−1(ψ(cid:48)(cid:48))t ψ(cid:48) + ψ(cid:48)(cid:48) − 1 ψ(cid:48) − η−1 − η−1R−1(ψ(cid:48)(cid:48))t ψ(cid:48)(cid:48) Here, we have subtracted the ψ(cid:48)-free term from the edge insertion and divided the result by ψ(cid:48). The minus sign in front is the sign of the second term of (8). Similarly, if we place the new vertex on the second half-edge, we obtain (cid:20) η−1 − R−1(ψ(cid:48))η−1R−1(ψ(cid:48)(cid:48))t ψ(cid:48) + ψ(cid:48)(cid:48) − 1 ψ(cid:48)(cid:48) − η−1 − R−1(ψ(cid:48))η−1 ψ(cid:48) Adding the two contributions yields η−1 − R−1(ψ(cid:48))η−1 − η−1R−1(ψ(cid:48)(cid:48))t + R−1(ψ(cid:48))η−1R−1(ψ(cid:48)(cid:48)) ψ(cid:48)ψ(cid:48)(cid:48) 29 (cid:21) . (cid:21) . t = 1η−1 − R−1(ψ(cid:48))η−1 ψ(cid:48) η η−1 − η−1R−1(ψ(cid:48))t ψ(cid:48)(cid:48) . (10) The result (10) is precisely the product of the standard edge insertions for the two new edges, considering the ψ-classes at the new vertex vanish and Ω0,3(v(cid:48) ⊗ v(cid:48)(cid:48) ⊗ 1) = η(v(cid:48), v(cid:48)(cid:48)) . Finally, as before, we replace the 1 on the special leg by R−1(1)(ψn+1) with- out consequence since ψn+1 = 0 on our vertex. Once again, we have ob- tained a stable graph with n + 1 legs marked precisely as in the definition of (RTbΩ)g,n+1(v1 ⊗ ··· ⊗ vn ⊗ 1). The final case to consider is when we place the new vertex on a κ-leg. The edge joining the new vertex to the rest of the graph will then be marked by −Tb(ψi) = [R−1(1) − 1](ψi). ψi We immediately take the push-forward of our class under the partial forgetful map that forgets just the single κ-leg we are considering. We will obtain a graph on which the special leg carries the marking [R−1(1) − 1](ψn+1). Exactly the same graph also appears among what we called the terms of the first kind — when we attach the special leg without creating any new vertices. There the special leg carried the marking 1. After adding the two contributions together, we obtain R−1(1)(ψ), which is the standard leg insertion. We have shown p∗(RTbΩ)g,n(v1 ⊗ ··· ⊗ vn) is given by precisely the same sum over stable graph contributions as (RTbΩ)g,n+1(v1⊗···⊗ vn⊗ 1). There- ♦ fore the two are equal. Definition 2.13 Let Ω be a CohFT with unit 1 ∈ V . Let R(z) be an R-matrix satisfying the symplectic condition, and let T (z) = z · 1 − zR−1(z)(1) ∈ z2V [[z]]. The unit-preserving R-matrix action on Ω is R.Ω = RT Ω . Proposition 2.14 The unit-preserving R-matrix action is a left group ac- tion. 30 Ra.(Rb.Ω) = Ra Proof. By Definition 2.13, we have (cid:0)z[1 − R−1 a ](cid:1)Rb (cid:0)z[1 − R−1 (cid:16) a ](cid:1)Rb equals the action of Rb The action of(cid:0)z[1 − R−1 a ](cid:1)(cid:17)(cid:0)z[1 − R−1 (cid:0)z[1 − R−1 (cid:16) z(cid:2)R−1 Proposition 2.9. Thus, (11) equals R−1 ](cid:1)Ω. a ](cid:1)(cid:17) (cid:0)z[1 − R−1 b (1)(cid:3)(cid:17) b (1) − R−1 b R−1 ](cid:1)Ω = (RaRb)(cid:0)z[1 − (RaRb)−1(1)](cid:1)Ω . a (1) + 1 − R−1 Ω = RaRb R−1 b RaRb b b (cid:16) b (11) by Proposition 2.8 has been used in the last equality. The result is precisely the ♦ definition of the unit-preserving action of RaRb. 3 The R-matrix for A2 We compute the R-matrix for the Frobenius manifold of the A2 singularity and deduce an expression for the shifted 3-spin Witten class in terms of stable graphs. The outcome is a proof of Theorems 1 and 4. 3.1 The Frobenius manifold A2 We compute all the differential geometric data associated with the Frobenius manifold A2 for use in the following calculations. The Frobenius manifold A2 is based on the 2-dimensional vector space9 V with coordinates x = t0 and y = t1 corresponding to the remainders 0 and 1 modulo 3 respectively. The unit vector field is ∂x = ∂ ∂x. The metric is η = dx ⊗ dy + dy ⊗ dx or η = (cid:19) (cid:18) 0 1 1 0 . Since the only nonzero values of Witten’s 3-spin class in genus 0 are W0,3(0, 0, 1) = 1, W0,4(1, 1, 1, 1) = 1 3 , 9In the notation of Section 0.6, V is V3. 31 the primary genus 0 Gromov-Witten potential is The Euler field is F(x, y) = 1 2 x2y + 1 72 y4 . E = x ∂ ∂x + 2 3 y ∂ ∂y . The Lie derivatives of E on the basis vectors fields are easily calculated: LE(∂x) = [E, ∂x] = −∂x , LE(∂y) = [E, ∂y] = −2 3 ∂y . By Proposition 1.8, the conformal dimension equals r − 2 r = 1 3 . δ = Let v be a tangent vector at a point of the Frobenius manifold. We define the shifted degree operator µ(v), also called the Hodge grading operator, by µ(v) = [E, v] + (1 − δ/2)v . Here, the vector v is extended to a flat tangent vector field in order to compute the commutator. We have ∂x , µ(∂x) = −1 6 1 6 µ(∂y) = ∂y . Definition 3.1 To simplify the formulas, we will use the following notation: (cid:98)∂x = φ1/4∂x , (cid:98)∂y = φ−1/4∂y . y 3 φ = , The frame ((cid:98)∂x,(cid:98)∂y) in the tangent space of V at (x, y) is the most practical for the computations. The dual frame of the cotangent space is denoted by (cid:98)dx = φ−1/4dx , (cid:98)dy = φ1/4dy . 32 The quantum multiplication of vector fields on the Frobenius manifold is given by is expressed by the matrix Whether in basis (∂x, ∂y) or in frame ((cid:98)∂x,(cid:98)∂y), the shifted degree operator Unlike ∂x, the vector field (cid:98)∂x is not flat. However, in the definition of µ, we use the flat extension of (cid:98)∂x at a given point, which only differs from ∂x by a frame ((cid:98)∂x,(cid:98)∂y), ξ is given by We will also need the operator ξ of quantum multiplication by E. In the multiplicative constant. 1 6 (cid:98)∂x •(cid:98)∂x = φ1/4(cid:98)∂x , (cid:98)∂x •(cid:98)∂y = φ1/4(cid:98)∂y , (cid:98)∂y •(cid:98)∂y = φ1/4(cid:98)∂x . (cid:19) (cid:18)−1 0 . 0 1  x 2φ3/2 ξ =  . 2φ3/2 x Remark 3.2 The computations not involving the Euler vector field apply more generally to 2-dimensional Frobenius manifolds whose Gromov-Witten potential has the form F(x, y) = x2y + Φ(y) 1 2 with the convention φ = φ(y) = Φ(cid:48)(cid:48)(cid:48)(y). For instance, the Gromov-Witten potential of CP1 has the above form with Φ = φ = Qey. 3.2 The topological field theory A topological field theory ωg,n is a CohFT of degree 0, as discussed in Sec- tion 0.5. Teleman’s reconstruction, used to prove Theorem 5, expresses ev- ery semisimple CohFT Ω as a unit-preserving R-matrix action (see Defini- tion 2.13) on the topological field theory ωg,n with unit where ω0,3 = Ω0,3. 33 Let us start by determining the topological field theory ωg,n for Witten’s 3-spin class. Lemma 3.3 For the topological (degree 0) part of Witten’s 3-spin theory, we have where n = n0 + n1. Here, ωg,n((cid:98)∂⊗n0 y x ⊗(cid:98)∂⊗n1 (cid:12)(cid:12)(cid:12)(cid:12) 1 0 δodd g+n1 = ) = 2gφ 2g−2+n 4 · δodd g+n1, if g + n1 is odd, if g + n1 is even. Proof. The values of ω0,3 are prescribed by the quantum product: ω0((cid:98)∂x ⊗(cid:98)∂x ⊗(cid:98)∂x) = ω0((cid:98)∂x ⊗(cid:98)∂y ⊗(cid:98)∂y) = 0 ω0((cid:98)∂x ⊗(cid:98)∂x ⊗(cid:98)∂y) = ω0((cid:98)∂y ⊗(cid:98)∂y ⊗(cid:98)∂y) = φ1/4 . For other g and n, we consider a stable curve with a maximal possible number of nodes (each component is rational with 3 special points). The vectors (cid:98)∂x and (cid:98)∂y are placed in some way on the marked points, and we must place either (cid:98)∂x ⊗(cid:98)∂y or (cid:98)∂y ⊗(cid:98)∂x at each node in such a way that the number of (cid:98)∂y’s is odd on each component of the curve. If g + n1 is even, such a placement is impossible. If g + n1 is odd, the placement can be done in 2g ways, since the dual graph of the curve has g independent cycles. placement of the(cid:98)∂x’s and(cid:98)∂y’s equals φ By the factorization rules for CohFTs, the contribution of each successful , where 2g− 2 + n is the number ♦ of rational components of the curve. 2g−2+n 4 3.3 The R-matrix Givental [7, pages 4-5] gives a general method for computing the R-matrix of a Frobenius manifold without using an Euler field. The method is ambiguous: the R-matrix depends on the choice of certain integration constants. In the presence of an Euler field E, there is a unique choice of constants such that LERm = −mRm for every m. In the conformal case, Givental’s method can be simplified by substituting iE into his recursive equation. The simplified method for com- puting the R-matrix of a conformal Frobenius manifold is given, for instance, 34 by Teleman [20] in the proof of the theorem of Section 8.15. Since the 3- spin theory yields a conformal Frobenius manifold, the simplified method is suitable for us. Let ξ be the operator of quantum multiplication by the tangent vector E. The matrices Rm then satisfy the following recursive equation10: [Rm+1, ξ] = (m + µ)Rm. (12) Rm = Using the formulas of Section 3.1 for ξ and µ, we rewrite (12) as At a semisimple point of a conformal Frobenius manifold, the above equation determines the matrices Rm uniquely starting from R0 = 1. Let (cid:18)am bm (cid:19)(cid:21) (cid:18)bm+1 − cm+1 am+1 − dm+1 (cid:19) (cid:18)6m − 1 (cid:19)(cid:18)am bm (cid:19) (cid:19) (cid:18)(6m − 1)am (6m − 1)bm (cid:18) x (cid:20)(cid:18)am+1 bm+1 cm+1 dm+1 or in other words 2φ3/2 x 0 6m + 1 cm dm (cid:19) , = (cid:19) cm dm . 2φ3/2 1 6 0 2φ3/2 dm+1 − am+1 cm+1 − bm+1 = 1 6 (6m + 1)cm (6m + 1)dm , . The following formulas are easily checked to be the unique solutions: am = bm = cm = dm = 1 1728m φ3m/2 1 1728m φ3m/2 1 1 + 6m 1 − 6m 1 + 6m 1 − 6m (6m)! (3m)! (2m)! (6m)! (6m)! (3m)! (2m)! δeven m , δodd m , 1728m φ3m/2 (3m)! (2m)! 1 (6m)! 1728m φ3m/2 (3m)! (2m)! δodd m , δeven m . We now make explicit the connection with the central power series dis- B0(T ) = (6m)! (−T )m, B1(T ) = (2m)!(3m)! m≥0 1 + 6m 1 − 6m (6m)! (2m)!(3m)! (−T )m. 10In Teleman’s paper, the commutator has the opposite sign, since his Euler field is the opposite of ours. 35 covered by Faber and Zagier, (cid:88) (cid:88) m≥0 the respective even and odd degree 0 , Bodd , Beven Denote by Beven parts. The final expression for the R-matrix is: −Bodd , and Bodd z 0 1 1  Beven −Bodd 1 (cid:18) (cid:18) 0 R(z) = 1728 φ3/2 1 1728 φ3/2 z 1728 φ3/2 Beven 0 1728 φ3/2 (cid:19) (cid:19)  . z z The symplectic condition for the R-matrix follows from the identity B0(T )B1(−T ) + B0(−T )B1(T ) = 2, or, equivalently, Beven (T )Beven 0 (T ) − Bodd (T )Bodd 1 (T ) = 1 discovered previously in [16]. Using the identity, we find 1 (cid:18) (cid:18) Beven Bodd 0 0 R−1(z) = z 1728 φ3/2 z 1728 φ3/2 Bodd 1 Beven 1 z 1728 φ3/2 z 1728 φ3/2 (cid:19) (cid:19)  . (cid:19) (cid:19) 0 (cid:19) (cid:19) (cid:18) (cid:18) (cid:18) (cid:18) (13) (14) 3.4 An expression for the shifted 3-spin Witten class We combine here the expression for the topological field theory from Sec- tion 3.2 with the R-matrix action from Definition 2.13 using the explicit formulas for the R-matrix of Section 3.3. Let τ = (x, y), y (cid:54)= 0, be a point of the Frobenius manifold A2. Let a1, . . . , an ∈ {0, 1} and let g − 1 +(cid:80)n i=1 ai 3 D = be the degree of Witten’s 3-spin class. By convention, φ = y/3. Recall the expressions Rd g,(a1,...,an) of Definition 0.1. Theorem 6 Witten’s class for the shifted 3-spin theory equals g,n(∂a1 ⊗ ··· ⊗ ∂an) = 2g(cid:88) Wτ 1728d q(cid:0)Rd 2 (D−d) φ 3 d≥0 (cid:1) , g,(a1,...,an) where ∂0 = ∂x, ∂1 = ∂y. 36 The following Corollary is an immediate consequence of Theorem 6 and the equation g,n(∂a1 ⊗ ··· ⊗ ∂an) = Wg,n(a1, . . . , an) + lower degree terms. Wτ explained in Section 1. Theorems 1 and 4 are implied by the Corollary. Corollary 7 We have the evaluations: (cid:1) = 2g 1728D Wg,n(a1, . . . , an) (cid:1) = 0 q(cid:0)Rd q(cid:0)Rd g,(a1,...,an) g,(a1,...,an) for d = D, for d > D. Proof of Theorem 6. By Teleman’s reconstruction result in the conformal semisimple case, Witten’s shifted 3-spin class is given by R.ω where • R is given by (13), • ω is the topological part of the shifted 3-spin theory. The proof now just amounts to a systematic matching of all factors in the sums over stable graphs which occur in Definition 2.13 for the R-matrix action and Definition 0.1 for Rd g,(a1,...,an). Consider first the expression for the CohFT R.ω applied to a tensor prod- uct of n vectors ∂x and ∂y. As before, we denote by n0 and n1 the number of 0s and 1s among a1, . . . , an so that n0 + n1 = n. Powers of φ. Since we wrote the R-matrix in frame ((cid:98)∂x,(cid:98)∂y), we must substitute ∂x (cid:55)→ φ−1/4(cid:98)∂x , ∂y (cid:55)→ φ1/4(cid:98)∂y in the tensor product argument for R.ω. The result of the substitution is a factor of φ n1−n0 . 4 By formula (14), all coefficients of R−1 m contain a factor of φ−3m/2. Tracing through the definitions of the all the actions R.ω = RT ω , T (z) = z · [∂x − R−1(∂x)](z), (15) the R-matrix contributes a factor of φ−3d/2, where d is the degree of the class. 37 By Lemma 3.3, the topological field theory ω contributes (subject to for every vertex v. parity condition accounted for later) a factor of φ These factors combine to yield φ 2gv−2+nv . 4 Finally, each κ-leg contributes in two way. First, since we must substitute 2g−2+n 4 ∂x (cid:55)→ φ−1/4(cid:98)∂x in formula (15) for T (z), each κ-leg contributes φ−1/4. Second, because the κ-leg increases the valence of the vertex by 1, a factor of φ1/4 is contributed via the topological field theory. Thus, the contributions of each κ-leg to the power of φ cancel. Collecting all of the above factors, we obtain a final calculation of the exponent of φ: n1 − n0 4 − 3d 2 + 2g − 2 + n 4 = g − 1 + n1 − 3d 2 3D − 3d 2 = (D − d). = 3 2 Powers of 1728. All coefficients of R−1 obtain a factor of 1728−d from the R-matrix action. m contain a factor of 1/1728. Hence, as above, we Powers of 2. At each vertex the topological field theory contributes a factor of 2gv . These combine into (cid:89) v∈V (Γ) 2gv = 2g 2h1(Γ) . The factor 2−h1(Γ) is present in the definition of Rd 2g is included in the statement of Theorem 6. g,(a1,...,an), and the remaining Parity conditions at the vertices. The topological field theory ω provides a nonzero contribution at a vertex if and only if gv + n1(v) is odd. We must prove the parity condition which occurs in the definition of Rd g,(a1,...,an) exactly matches. The parity condition is imposed on Rd g,(a1,...,an) by extracting the coeffi- cient of ζ gv−1 , at each vertex v: see Definition 0.1. We may view the factors of ζv as having the following sources. A leg carrying the assignment al = 1 (corresponding to ∂y) contributes a ζv, while a leg carrying the assignment v 38 al = 0 (corresponding to ∂x) does not. The terms of Bodd (including the ef- contribute a ζv. The terms of Beven fect of the κ-legs) and the terms of Bodd and Beven do not contribute anything (because they leave the parity invari- ant). Finally, every edge insertion ∆e contributes a factor if e is adjacent to v. The edge term of Definition 0.1 can be expanded via 1 1 0 0 B0 = Beven 0 + Bodd 0 , B1 = Beven 1 + Bodd 1 and matched with the edge term of the CohFT R.ω using (6) and (14). Then the contributing factor is ζv if the bi-vector includes a factor (cid:98)∂y on the side counts the parity of entries (cid:98)∂y submitted to the topological field theory ω at of the vertex v and 1 otherwise. Hence, the power of the variable ζv correctly the vertex v. Coefficients of the series B. These coefficients simply coincide in the expression for Rd g,(a1,...,an) and the formulas of the unit-preserving R-matrix action in all instances (legs, κ-legs, ♦ and edges). (cid:101)P implies P 3.5 Our relations Rd set of relations P conjectured in [16] from the set(cid:101)P proven in Theorem 1. We present here the proof of Corollary 2: the derivation of the more complete g,A differ from the relations Rd g,A,σ of [16] in three ways. First, the signs of the coefficients in the series B0 and B1 are modified. The outcome is a global change of sign in some of the relations. Second, the range of the ai’s is different. In our relations, the ai’s are equal to 0 or 1, while in [16], the ai’s can be any integers equal to 0 or 1 modulo 3. In fact, replacing an ai by ai + 3 in the relations of P amounts to multiplying the relation by ψi. Therefore taking ai < 3 is sufficient. Finally, the relations of [16] also depend on a partition σ. In our relations, we are implicitly considering only the empty partition case. A relation with a nonempty partition σ is easily obtained from a relation with an empty σ by push-forward: g,(a1,...,an),(σ1,...,σm) = p∗Rd Rd g,(a1,...,an,σ1+3,...,σm+3), where p : Mg,n+m → Mg,n is the forgetful map. Thus the relations Rd imply all the relations Rd g,A g,A,σ. 39 The span of(cid:101)P is not an ideal, but generates an ideal in each Sg,n. The as- sociated family of ideals is closed under pull-backs by forgetful maps (because of axiom (iii) of a CohFT with unit) and gluing maps (because of axiom (ii) of a CohFT). The family of ideals is not closed under push-forwards by for- getful maps and gluing maps. After taking the closure under push-forwards by forgetful maps, we obtain the span of Rd g,A,σ, as we have just proved. Taking the closure under push-forwards by gluing maps we get the full set of ♦ relations P from [16]. 3.6 Examples Example 3.4 Let g = 0, n = 3. Here, we have 0,3(∂x ⊗ ∂x ⊗ ∂y) = 1, Wτ 0,3(∂y ⊗ ∂y ⊗ ∂y) = φ = Wτ y 3 . These values come directly from the topological field theory – the R-matrix is not needed. The first expression equals the Witten class W0,3(0, 0, 1), and the second expression is the push-forward of W0,4(∂y ⊗ ∂y ⊗ ∂y ⊗ y∂y) = yW0,4(1, 1, 1, 1). In both cases, no further y∂y insertions are possible for dimension reasons. Example 3.5 Let g = 0, n = 4. We will study all 5 cases. First case: Wτ condition imposes d = 1. Thus 3 0,4(∂x ⊗ ∂x ⊗ ∂x ⊗ ∂x). We have D = −1/3. The parity 2(D − d) = −2. By Theorem 6, we find 60κ1 − 60(cid:80)4 0,4(∂x ⊗ ∂x ⊗ ∂x ⊗ ∂x) = Wτ i=1 ψi + 60δ . 1728φ2 Since d > D, the above expression must be 0. We obtain the first nontrivial relation: κ1 − 4(cid:88) ψi + δ = 0 ∈ H 2(M0,4, Q) . The relation is true by the following basic evaluation in H 2(M0,4, Q): i=1 κ1 = [point], ψi = [point], δ = 3[point] . 40 Alternatively, we see that this expression coincides up to a factor with Mum- ford’s formula for λ1, and λ1 = 0 in genus 0. Second case: Wτ condition imposes d = 0. Thus 3 0,4(∂x ⊗ ∂x ⊗ ∂x ⊗ ∂y). We have D = 0. The parity 2(D − d) = 0. From Theorem 6, we obtain 0,4(∂x ⊗ ∂x ⊗ ∂x ⊗ ∂y) = 1. Wτ Since d = D we know that this expression should be equal to Witten’s class, which is indeed the case: W0,4(0, 0, 0, 1) = 1. Third case: Wτ 0,4(∂x ⊗ ∂x ⊗ ∂y ⊗ ∂y). We have D = 1/3. The parity condition imposes d = 1. Thus 3 2(D − d) = −1. We obtain 0,4(∂x ⊗ ∂x ⊗ ∂y ⊗ ∂y) = Wτ 60κ1 − 60(ψ1 + ψ2) + 84(ψ3 + ψ4) + 60δ[1,23,4] − 84(δ[1,32,4] + δ[1,42,3]) 1728φ . Since d > D, the expression must vanish (as is easly checked). Fourth case: Wτ 0,4(∂x ⊗ ∂y ⊗ ∂y ⊗ ∂y). We have D = 2/3. The parity condition imposes d = 0. Thus 3 2(D − d) = 1. We obtain 0,4(∂x ⊗ ∂y ⊗ ∂y ⊗ ∂y) = φ Wτ which is the push-forward of W0,5(∂x ⊗ ∂y ⊗ ∂y ⊗ ∂y ⊗ y∂y) = yW0,5(0, 1, 1, 1, 1) under the forgetful map forgetting the last marked point. Fifth case: Wτ imposes d = 1. Thus 3 2(D − d) = 0. We obtain 0,4(∂y⊗∂y⊗∂y⊗∂y). We have D = 1. The parity condition 0,4(∂y ⊗ ∂y ⊗ ∂y ⊗ ∂y) = Wτ = = 60κ1 + 84(cid:80)4 i=1 ψi + 60δ (60 + 84 · 4 + 60 · 3)[pt] 1728 1728 1 3 [point] which is the correct value of Witten’s class 1 3 W0,4(1, 1, 1, 1) = [point]. 41 The relations obtained through these computations in d = 1 (after divid- ing by 12 or by 60) are listed below: κ1 − ψ1 − ψ2 − ψ3 − ψ4 + δ[1,23,4] + δ[1,32,4] + δ[1,42,3] = 0, 5κ1 − 5ψ1 − 5ψ2 + 7ψ3 + 7ψ4 + 5δ[1,23,4] − 7δ[1,32,4] − 7δ[1,42,3] = 0, 5κ1 + 7ψ1 + 7ψ2 − 5ψ3 − 5ψ4 + 5δ[1,23,4] − 7δ[1,32,4] − 7δ[1,42,3] = 0, 5κ1 − 5ψ1 + 7ψ2 − 5ψ3 + 7ψ4 − 7δ[1,23,4] + 5δ[1,32,4] − 7δ[1,42,3] = 0, 5κ1 + 7ψ1 − 5ψ2 + 7ψ3 − 5ψ4 − 7δ[1,23,4] + 5δ[1,32,4] − 7δ[1,42,3] = 0, 5κ1 − 5ψ1 + 7ψ2 + 7ψ3 − 5ψ4 − 7δ[1,23,4] − 7δ[1,32,4] + 5δ[1,42,3] = 0, 5κ1 + 7ψ1 − 5ψ2 − 5ψ3 + 7ψ4 − 7δ[1,23,4] − 7δ[1,32,4] + 5δ[1,42,3] = 0. After some linear algebra, the system is equivalent to: κ1 = ψ1 = ψ2 = ψ3 = ψ4 = δ[1,23,4] = δ[1,32,4] = δ[1,42,3]. We have obtained a complete set of relations in RH 2(M0,4). Example 3.6 The Getzler relation [5] is a degree 2 relation in S1,4 which can not be obtained by the pull-back of any simpler relations. Since 1 − 1 + 1 + 1 + 1 + 1 2 > 1,(1,1,1,1) lies in the set (cid:101)P. 3 , = 4 3 In fact, R2 the relation R2 relation (modulo more elementary genus 0 and 1 relations). 1,(1,1,1,1) is the Getzler The Belorousski-Pandharipande relation [1] is a degree 2 relation in S2,3 which can not be obtained by the pull-back of any simpler relations. The relation R2 2,(1,1,1) lies in(cid:101)P since 2 − 1 + 1 + 1 + 1 3 = 4 3 2 > and is an equivalent form of the BP equation. The outcome of several such investigations is reported in [16]. All known relations have been explained by Theorem 6. 42 3.7 Some concluding remarks Our computations provide an instructive example of what happens to a Coh- FT as we move towards a non-semisimple point of a Frobenius manifold. Let us examine more closely the limit of our expressions for the shifted Witten 3-spin class as y → 0 or, in other words, φ → 0. The coefficients of the R-matrix involve negative powers of φ, therefore the R-matrix diverges. The topological field theory ω to which we apply the R-matrix involves positive powers of φ. As a result, each term of our expression for the shifted Witten class comes with a factor 2 3 (D−d), φ where D = Dg,n(a1, . . . , an) is the degree of Witten’s class and d is the degree of the term in question. As φ → 0, the terms of degree less then D tend to 0, the terms of degree equal to D are invariant, and the terms of degree greater than D diverge. At first sight the expression appears to diverge, but because the terms of degree greater than D combine into tautological relations the expression actually has a finite limit equal to Witten’s 3-spin class. Mg,n to formulas on the space M1/r is no: the divisibility condition A natural question is whether our formulas for Witten’s class lift from g;a1,...,an of r-spin structures. The answer (g − 1)(r − 2) +(cid:80) r i ai ∈ Z does not necessarily hold for each vertex of the dual graph. Hence, there is no natural boundary stratum in M1/r g;a1,...,an where the terms of our formula can be lifted. Moreover, in the simplest case r = 2, we have Wg,n(0, . . . , 0) = 1 −1 if the spin structure is even, if the spin structure is odd. (cid:12)(cid:12)(cid:12)(cid:12) Such an answer cannot be expressed in terms of dual graphs at all. Some more structure is required. References [1] P. Belorousski and R. Pandharipande, A descendent relation in genus 2, Ann. Scuola Norm. Sup. Pisa Cl. Sci. 29 (2000), 171-191. 43 [2] A. Chiodo, The Witten top Chern class via K-theory, J. Algebraic Geom. 15 (2006), no. 4, 681-707. arXiv:math/0210398 [3] B. Dubrovin, Geometry of 2d topological field theories, arXiv:hep-th/9407018 [4] B. Dubrovin, On almost duality for Frobenius manifolds, In: Geometry, topology, and mathematical physics, AMS Transl. Ser. 2, Vol. 212, Amer. Math. Soc., Providence, RI 2004, 75-132. arXiv:math/0307374 [5] E. Getzler, Intersection theory on M 1,4 and elliptic Gromov-Witten in- variants, JAMS 10 (1997), 973-998. [6] A. Givental, Gromov-Witten invariants and quantization of quadratic Hamiltonians, Mosc. Math. J. 1 (2001), no. 4, 551-568, 645. arXiv:math/0108100 [7] A. Givental, Semisimple Frobenius structures at higher genus, Internat. Math. Res. Notices (2001), no. 23, 1265-1286. arXiv:math/0008067 [8] T. Graber and R. Pandharipande, Constructions of nontautological classes on moduli spaces of curves, Michigan Math. J. 51 (2003), no. 1, 93-109. arXiv:math/0104057 [9] H. Fan, T. Jarvis, and Y. Ruan, The Witten equation, mirror symmetry and quantum singularity theory, arXiv:0712.4025. [10] E.-N. Ionel, Relations in the tautological ring of Mg, Duke Math. J. 129 (2005), no. 1, 157-186. [11] M. Kontsevich and Yu. Manin, Gromov-Witten classes, quantum co- homology, and enumerative geometry, Mirror symmetry, II, 607-653, AMS/IP Stud. Adv. Math., 1, Amer. Math. Soc., Providence, RI, 1997. arXiv:hep-th/9402147 [12] T. Mochizuki, The virtual class of the moduli stack of stable r-spin curves, Comm. Math. Phys. 264 (2006), no. 1, 1-40. [13] D. Mumford, Towards an enumerative geometry of the moduli space of curves, in Arithmetic and Geometry (M. Artin and J. Tate, eds.), Part II, Birkhauser, 1983, 271–328. 44 [14] R. Pandharipande and A. Pixton, Relations in the tautological ring of the moduli space of curves, arXiv:1301.4561. [15] R. Pandharipande, A. Pixton, and D. Zvonkine, in preparation. [16] A. Pixton, Conjectural relations in the tautological ring of Mg,n, arXiv:1207.1918 [17] A. Polishchuk and A. Vaintrob, Algebraic construction of Witten’s top Chern class, Advances in algebraic geometry motivated by physics (Low- ell, MA, 2000), 229-249, Contemp. Math., 276, Amer. Math. Soc., Prov- idence, RI, 2001. arXiv:math/0011032 [18] A. Polishchuk, Witten’s top Chern class on the moduli space of higher in Frobenius manifolds, 253-264, Aspects Math., E36, spin curves, Vieweg, Wiesbaden, 2004. arXiv:math/0208112 [19] S. Shadrin, BCOV theory via Givental group action on cohomological fields theories, Mosc. Math. J. 9 (2009), no. 2, 411-429, back matter. [20] C. Teleman,The structure of 2D semi-simple field theories, Invent. Math. 188 (2012), no. 3, 525-588. arXiv:0712.0160 [21] E. Witten, Algebraic geometry associated with matrix models of two- dimensional gravity, in Topological methods in modern mathematics (Stony Brook, NY, 1991), 235-269, Publish or Perish, Houston, TX, 1993. Departement Mathematik ETH Zurich [email protected] Department of Mathematics Princeton University [email protected] CNRS, Institut Math´ematique de Jussieu [email protected] 45
1103.6233	1	1103	2011-03-31T16:41:08	Countability properties of some Berkovich spaces	[ "math.AG" ]	We prove that any compact Berkovich space over the field of Laurent series over an arbitrary field is angelic. In particular, is it sequentially compact.	math.AG	math	COUNTABILITY PROPERTIES OF SOME BERKOVICH SPACES CHARLES FAVRE CONTENTS Introduction 1. 2. Riemann-Zariski spaces 3. Countability properties in Riemann-Zariski spaces 4. Proof of the main results References 1 3 5 8 11 1. INTRODUCTION Our aim is to prove some facts on the topology of a particular class of non archi- medean analytic spaces in the sense of Berkovich. One of the main feature of Berkovich analytic spaces is that their natural topology makes them both locally arcwise connected and locally compact. However, in general these spaces are far from being separable1. Nevertheless, we shall prove that analytic spaces retain some countability properties that reflect their algebraic nature. We expect these properties to be useful in a dynamical context. We shall work in the following setting. We consider an normal algebraic variety2 X defined over a field k. We fix an effective Cartier divisor D ⊂ X, and we let X be the formal completion of X along D. We then consider the generic fibre Xη of this formal scheme as defined in [T]: this is an analytic space in the sense of Berkovich over k endowed with the trivial norm. When X = Spec A is affine, and D = {f = 0} with f ∈ A, then Xη coincides with the set of bounded (hence ≤ 1) multiplicative semi- norms · : A → R+ such that 0 < f < 1. Note that replacing · by ν := − log · , we can define in an equivalent way Xη as the set of valuations ν : A → R+ ∪ {+∞} (possibly taking the value +∞ on a non-zero element) such that ∞ > ν(f ) > 0. In the general case, X is covered by affine open subsets Ui, and Xη is obtained by patching together the Ui,η's in a natural way. Date: October 11, 2018. 2000 Mathematics Subject Classification. Primary: 37F10, Secondary: 11S85, 37E25. Supported by the ANR project Berko, and by the ECOS project C07E01. 1a separable space is a topological space admitting a countable dense subset. We shall never use the notion of separable field extension so that no confusion should occur. 2this means X is irreducible reduced of finite type and k is algebraically closed. The last assumption is however inessential in our paper. 1 2 CHARLES FAVRE Multiplying a valuation by a positive constant yields an action of R∗ + on Xη. It is therefore natural to introduce the normalized generic fiber ]X[ ⊂ Xη of valuations nor- malized by the condition ν(f ) = +1. In this way, ]X[ can be identified with the quotient space of Xη by R∗ +, and the projection map Xη →]X[ is a R∗ +-fibration. When D is complete, then the space ]X[ is compact, see Lemma 4.2 below. Our main result reads as follows. Theorem A. Let X be a normal algebraic variety over k and D be an effective Cartier divisor. Let A be any subset of the normalized generic fiber ]X[ of the formal completion of X along D. For any x in the closure of A, there exists a sequence of points xn ∈ A such that xn → x. Recall that a topological space is angelic if any relatively ω-compact set3 is relatively compact, and for any subset A, any point in ¯A is the limit of a sequence of points in A, see [Fl] for more informations. This property plays an important role in the study of Banach spaces. A first consequence of the previous result is the following Corollary A. Let X be a normal algebraic variety over k and D be an effective Cartier divisor. Then ]X[ is angelic. If moreover D is complete then ]X[ is sequentially compact. In order to state yet another consequence of Theorem A, we need to introduce some terminology. When X is affine, a divisorial valuation on the ring of regular functions k[X] is a discrete valuation of rank 1 and transcendence degree dim(X) − 1. Geomet- rically, it is given by the order of vanishing along a prime divisor in a suitable birational model of X. A point x ∈ Xη is called divisorial if it is given by a divisorial valuation in the ring of regular functions of some affine chart. If we interpret x as a semi-norm, then x is divisorial iff it is a norm, and the residue field of the completion of k[X] w.r.t. the induced norm by x has transcendence degree dim(X) − 1 over k. Finally recall that we have a natural reduction map rX : Xη → D defined in the affine case by sending a semi-norm · to the prime ideal { · < 1} ⊂ k[X]. This reduction map sends a divisorial valuation to its center in X when it is non empty. Note that the center is automatically included in D if the valuation lies in Xη. We shall also obtain Corollary B. Let X be a normal algebraic variety over k and D be an effective Cartier divisor. Then for any point x ∈ Xη, there exists a sequence of divisorial points xn ∈ Xη that converges to x, and such that rX (xn) ∈ rX (x) for all n. Theorem A will be deduced from its analog on Riemann-Zariski spaces, see Theo- rem 3.1 below for details. In this case, it essentially boils down to the noetherianity of the Zariski topology on schemes. Let us now explain how one can transfer the previous results to non-archimedean analytic spaces. The normalized generic fiber ]X[ is not a Berkovich analytic space in a canonical way. However in the special situation where we have a map T : X → A1 k such that D = {T = 0}, then ]X[ turns out to be an analytic space over the non-archimedean field k((T )) (with the norm exp(−ord0)). In the case X = Spec (k[xi]/a) is affine, 3i.e. any sequence of points in this set has a cluster point in the ambient space COUNTABILITY PROPERTIES OF SOME BERKOVICH SPACES 3 (k[xi]/a)/(T n) is a k[[T ]]-algebra topologically of finite type, and ]X[ is then A = lim←−n the set of semi-norms on A whose restriction on k[[T ]] is exp(−ord0). In general, ]X[ coincides with the generic fiber of the T -adic completion of f in the sense of Berkovich (its construction in rigid geometry was previously given by Raynaud). We shall prove: Corollary C. Any compact Berkovich analytic space that is defined over the field k((T )) is angelic. It is in particular sequentially compact, and divisorial points are sequentially dense. Note that this result also holds over any (non archimedean) local field for simple reasons since any affinoid over such a field is separable. However when the residue field of k is not countable, Berkovich analytic spaces are not metrizable. For curves, the density of divisorial valuations follows from the semi-stable reduction theorem. And the sequential compactness is a consequence of the fact that any complete R-tree (in the sense of [FJ]) is sequentially compact. We refer to [M] for a proof. We note that any compact analytic space over any non-archimedean complete fields is angelic by the recent work of J. Poineau, [P]. Even though our result is much more restrictive, our approach can be directly adapted to prove the sequential compactness of special compactifications of complex affine varieties in the spirit of [Fa]. The following natural questions are related to the above results. Question 1. Let X be any Berkovich analytic space. Then any Borel measure on X is a Radon measure. Question 2. Let X be any Berkovich analytic space. Then the support of any Radon measure is separable. In dimension 1, question 2 has a positive answer. Question 1 remains open.4. Acknowledgements: we thank T. De Pauw, A. Ducros, M. Jonsson, J. Kiwi, J. Nicaise, and R. Menares for useful discussions on the material presented in this paper. Also we deeply thank J. Poineau for kindly informing the author about his proof of the sequential compactness of compact analytic spaces over an arbitrary field. 2. RIEMANN-ZARISKI SPACES Our basic references are [ZS, V]. A domain R is said to be a valuation ring if it has no divisors of zero, and for any non-zero element x in the fraction field of R, either x or x−1 belong to R. Any valuation ring is local, with maximal ideal mR consisting of those x ∈ R such that x−1 does not belong to R. A valuation on R is a function ν : R \ {0} → Γ to a totally ordered abelian group Γ such that ν(ab) = ν(a) + ν(b); and ν(a + b) ≥ max{ν(a), ν(b)}. Any valuation extends in a unique way to a valuation on the fraction field K of R, and the set Rν = {a ∈ K, ν(a) ≥ 0} is a valuation ring in K. 4there is a gap in both proofs of this fact given in [FJ] and [BR] 4 CHARLES FAVRE Conversely, to a valuation ring R ⊂ K is associated a unique valuation ν : R → Γ up to isomorphism, where Γ is the group obtained by moding out K ∗ by the multiplicative set R \ mR. In the sequel, we shall make no difference between valuation rings and valuations. Let X be a projective normal irreducible variety defined over a field k. Abhyankhar's inequality for a valuation ν on k(X) states that rat.rk(ν) + deg.tr(ν) ≤ dim(X) , where rat.rk(ν) denotes the dimension of the Q vector space Γ ⊗Z Q; and deg.tr(ν) is the degree of transcendence of the residue field over k. In particular, Γ⊗ZQ is countable. We thus have Lemma 2.1. The value group of any valuation on k(X) that is trivial on k is countable. The Riemann-Zariski space X is the set of valuation rings in k(X) containing k such that the sets U (A) = {R, A ⊂ R} where A ranges over all subrings of finite type of k(X) that contains k form a basis of open sets for its topology. A theorem of Zariski states: Theorem 2.2. The Riemann-Zariski space is quasi-compact. Note that X is never Hausdorff except in dimension 0. Suppose ν : k(X) → Γ is a valuation with valuation ring Rν. A projective birational model of X consists of a birational map φ : X ′ 99K X from a normal projective variety to X. The map φ induces an isomorphism between k(X) and k(X ′), so that ν can be viewed as a valuation on k(X ′). The set of closed points x′ ∈ X ′ such that the local ring OX ′,x′ is included in Rν, and its maximal ideal in mRν forms an irreducible subvariety called the center of ν in X ′. We denote it by C(ν, X ′). We shall view C(ν, X ′) scheme- theoretically as a (non necessarily closed) point in X ′. A birational model ψ : X ′′ 99K X dominates another one φ : X ′ 99K X iff µ := φ−1 ◦ ψ is regular. If X ′′ dominates X ′, then µ(C(ν, X ′′)) = C(ν, X ′). Consider the category B of all projective birational models of X up to natural isomor- phism, each model endowed with the Zariski topology. It is an inductive set for the re- lation of domination introduced before. We may thus consider the projective limit of all birational models of X, that is lim←−X ′∈B X ′ endowed with the projective limit topology. Concretely, a point in lim←−X ′∈B X ′ is a collection of irreducible subvarieties ZX ′ ⊂ X ′ for each projective birational model X ′ such that µ(ZX ′′) = ZX ′ if X ′′ dominates X ′. For a given valuation ν, we may attach the collection {C(ν, X ′)}X ′ ∈B. This defines a X ′. Conversely, given a point Z = {ZX ′}X ′∈B in the projective map from X to lim←−X ′∈B limit, we define the subset RZ of k(X) of those meromorphic functions that are regular at the generic point of ZX ′ for any birational model X ′ of X. It is not difficult to check that RZ is a valuation ring. These two maps are inverse one to the other. More precisely, one has the following fundamental result again due to Zariski: COUNTABILITY PROPERTIES OF SOME BERKOVICH SPACES 5 Theorem 2.3. The natural map X → lim←−X ′∈B induces a homeomorphism. X ′ given by ν 7→ {C(ν, X ′)}X ′ ∈B This result has the following useful consequence. For any model X ′, and any Weil divisor D′ in X ′, let U (X ′, D′) be the set of all valuation rings whose center is not included in D′. Then the collection of all sets of the form U (X ′, D′) gives a basis for the topology of X. Note that if we cover X ′ \ D′ by affine charts Yi, then X ′ \ Yi := Di is a divisor since X ′ is normal, and U (X ′, D′) = ∪ U (X ′, Di). Thus the collection of all U (X ′, D′)'s such that X ′ \ D′ is affine also forms a basis for the topology of X. Lemma 2.4. A sequence of valuations νn converges to ν in X iff for any f ∈ Rν, we have f ∈ Rνn for n large enough. Proof. Suppose first νn → ν, and pick f ∈ Rν. Choose a model X ′ such that f is regular on X ′, and let D′ be the set of poles of f . Since f belongs to Rν, the center of ν cannot be included in D′, and ν ∈ U (X ′, D′). By assumption νn → ν hence νn ∈ U (X ′, D′) for n large enough. This implies the center of νn not to be included in D′, and we conclude that f ∈ Rνn. Conversely, let us assume that for any f ∈ Rν, we have f ∈ Rνn for n large enough. The collection of open sets U (X ′, D′) with X ′ a birational model of X, and D′ ⊂ X ′ a divisor forms a basis for the topology on X. Therefore proving the convergence of νn to ν is equivalent to show νn ∈ U (X ′, D′) for n large enough if ν ∈ U (X ′, D′). As noted above, we may assume Y ′ = X ′ \ D′ is affine. Choose a finite set of regular functions fi on the affine space Y ′ that generate k[Y ′]. Since the center of ν in X ′ is not included in D′, we have ν(fi) ≥ 0 for all i. By assumption, we get νn(fi) ≥ 0 for all i and all large enough n. But then the center of νn cannot be in D′, hence νn ∈ U (X ′, D′). This concludes the proof. (cid:3) Finally we shall use several times the Lemma 2.5. Suppose I is a coherent sheaf of ideals on an irreducible normal variety X. Then there exists a regular birational map φ : X ′ → X such that X ′ is normal and I · OX ′ is locally principal (ie. invertible). Take X ′ to be the normalization of the blow-up of I, and use the universal property of blow ups. 3. COUNTABILITY PROPERTIES IN RIEMANN-ZARISKI SPACES In this section, X is a normal projective algebraic variety defined over a (non neces- sarily algebraically closed) field k. Our aim is to prove Theorem 3.1. Let A be any subset of X. Then for any ν in the closure of A, either ν ∈ {µ} for some µ ∈ A, or one can find a sequence of valuations νn ∈ A such that νn → ν. The proof relies on the following lemma. Recall that given an affine variety Y , a valuation ν : k[Y ] → Γ and γ ∈ Γ, the valuation ideal is defined by I(ν, γ) := {f, ν(f ) ≥ γ} ⊂ k[Y ] . 6 CHARLES FAVRE Lemma 3.2. Let Y be an irreducible affine variety. Suppose ν : k[Y ] → Γ is a valuation whose center in Y is non-empty. Pick γ ∈ Γ, with γ > 0, and consider a proper modification µ : Y ′ → Y with Y ′ normal such that I(ν, γ) · OY ′ = OY ′(−D) for some effective Cartier divisor D. Then C(ν, Y ′) is included in the support of D, and for any f ∈ k[Y ] such that ν(f ) = γ, the divisor div(f ◦ µ) ⊂ Y ′ is equal to D at the generic point of C(ν, Y ′). Proof. Suppose there exist a closed point p ∈ Y ′, and a regular function f ∈ OY such that ν(f ) = γ, and f ◦ µ does not vanish at p. Then the lift of f −1 is regular at p but does not belong to the valuation ring at ν. Whence p /∈ C(ν, Y ′). This proves the first claim. Next pick an element g in the valuation ideal so that ν(g) = γ, and the strict transform of g in Y ′ does not contain C(ν, Y ′). We may assume that the exceptional part of the divisor {g ◦ µ = 0} is equal to D. At a generic point p ∈ C(ν, Y ′), write f ◦ µ = f × f where f = 0 defines the strict transform of {f = 0} in Y ′, and { f = 0} is supported on the exceptional divisor. Since γ = ν(f ), we have div( f ) ≥ div(g) so that the quotient f /g is regular at p. Look at the equation: γ = ν(f ) = ν(g) + ν( f /g) + ν( f ) = γ + ν( f /g) + ν( f ) . Since f and f /g are regular at p that belong to the center of ν, we get ν( f ) ≥ 0, and ν( f /g) ≥ 0, whence ν( f ) = ν( f /g) = 0, and f , f /g are both non-zero at p. This proves the claim. (cid:3) Proof of Theorem 3.1. There is no loss of generality in assuming ν /∈ {µ} for any µ ∈ A. Let Y ⊂ X be an affine chart intersecting the center of ν, and write ν : k[Y ] → Γ for some totally ordered group Γ. By Lemma 2.1, ν(OY ) is countable. By Lemma 2.5, we π1−→ Y of birational models such that for may produce a sequence Yn+1 any γ ∈ Γ, the valuation ideal sheaf I(ν, γ) · OYn is locally principal for all n large enough. πn−→ Yn πn−1−→ ... For any subset B ⊂ A, we introduce the Zariski closed set Cn(B) := ∪µ∈BC(µ, Yn) . Lemma 3.3. There exists a countable subset A′ ⊂ A and irreducible subvarieties Zn ⊂ Yn such that (1) πn−1(Zn) = Zn−1 for all n; (2) the restriction maps πn−1 : Zn → Zn−1 are birational; (3) for all n, C(ν, Yn) ⊂ Zn; (4) for all n, Cn(A′) = Zn. Pick any enumeration {νm} of the elements of A′, and consider a countable field K such that all varieties Zn, C(νm, Yn), C(ν, Yn) are defined over K. Let us introduce the following terminology. A pro-divisor W = {Wn} is a collection of (possibly zero) reduced divisors Wn in Zn defined over K, such that πn(Wn+1) = Wn for all n, there exists an N for which WN is irreducible, and Wn = (πn−1 ◦ ... ◦ πN )−1(WN ) for all n ≥ N. We call the minimal N having this property the height of W and denote it by h(W ). Since for each N the set of prime divisors of ZN and defined over K is countable, we may find a sequence W j enumerating all pro-divisors. COUNTABILITY PROPERTIES OF SOME BERKOVICH SPACES 7 For any valuation µ ∈ A′ and any pro-divisor W , we write µ ⊂ W if C(µ, Yn) ⊂ Wn for all n. This condition is equivalent to impose C(µ, YN ) ⊂ WN for N = h(W ). Lemma 3.4. Suppose µ ∈ A′, and ν /∈ {µ}. Then one can find an integer j such that µ ⊂ W j. We now define by induction a sequence of valuations νm and of integers jm < jm+1 such that • νl ⊂ W jl; • νl 6⊂ W j for all j < jl. To do so we proceed as follows. Set Aj = {µ ∈ A′, µ ⊂ W j}. Set j1 = min{j, Aj 6= ∅}, and define recursively jm+1 = min{j > jm, Aj \ ∪l<jAl 6= ∅}. Let us justify the existence of jm for all m. By contradiction, assume that Aj \ ∪l<jAl = ∅ for all j > jm. Since by Lemma 3.4 any valuation in A′ belongs to some W j, we have A′ = ∪j≥0Aj. As a consequence, we conclude that A′ ⊂ ∪j≤jmAj which forces all valuations in A′ to have a center included in some fixed divisor of Z1. This contradicts property (4) of Lemma 3.3. Finally we pick any sequence νl ∈ Ajm. We now prove that νn converges to ν. Pick f ∈ Rν, write f = g h with g, h ∈ k[Y ]. Pick N sufficiently large such that I(ν(g), ν) and I(ν(h), ν) are locally principal in YN . By Lemma 3.2, the lift of f to YN is regular at the generic point of C(ν, YN ). Let Z be the set of poles of f in YN . Since f is regular along C(ν, YN ) ⊂ ZN , and ZN is irreducible, Z ∩ ZN is a divisor of ZN that is possibly empty. If so, then f is regular at the generic point of C(νl, YN ) for all l which implies f ∈ Rνl for all l. Otherwise, Z ∩ ZN determines a unique pro-divisor say W J for some J such that W J N ) for all n ≥ N. N = Z ∩ ZN , and W J By construction, for any integer l such that jl > J, we have νl 6⊂ W J. In other words, we have C(νl, YN ) 6⊂ W J N = Z ∩ ZN . We have thus proved that f is regular at the generic point of C(νl, YN ) for l large enough. This implies f ∈ Rνl for l large enough, and concludes the proof. (cid:3) n = (πn−1 ◦ ... ◦ πN )−1(W J Proof of Lemma 3.4. Pick any valuation µ ∈ A′. Assume that the center of µ in Yn is equal to Zn for all n. We need to prove that ν ∈ µ, i.e. Rν ⊂ Rµ. Pick any f ∈ Rν. As in the proof above, we can find an N such that f is regular at the generic point of the center of ν in YN . Hence f is regular at the generic point of Zn. Since the latter is the center of µ, we conclude that f ∈ Rµ. (cid:3) Proof of Lemma 3.3. We first construct varieties Zn in Yn satisfying the first three prop- erties. For that purpose, let us introduce the set Z be the set of all sequences Z• := {Zn}n of subvarieties of Yn such that Zn is an irreducible component of Cn(A) and πn(Zn+1) ⊂ Zn. Since we have πn(C(µ, Yn+1)) = C(µ, Yn) for any valuation, note that πn(Zn+1) = Zn for all n. For any fixed n, the set of irreducible components of Cn(A) is finite of cardinality d(n). The set of sequences of irreducibles components of Cn(A) is thus in natural bijection with Σ = Q∞ n=1{1, ..., d(n)}. For the product topology, Σ is a compact (totally disconneted) space, and Z is a closed (hence compact) subset of Σ. 8 CHARLES FAVRE Now consider Kn := {Z• ∈ Z, Zn ⊃ C(ν, Yn)}. If Z• belongs to Kn, then we have Zn−1 = πn−1(Zn) ⊃ πn−1(C(ν, Yn)) = C(ν, Yn−1). Hence Kn forms a decreasing sequence of non empty compact subsets of Z. The intersection ∩nKn is thus non empty, and we may pick Z• ∈ ∩nKn. This sequence of varieties Zn satisfies the properties (1) and (2) of the lemma. Since πn−1(Zn) = Zn−1, the dimension of Zn is increasing hence stationnary. Replacing the sequence Yn by Yn+N with N large enough, we may thus assume πn−1 : Zn → Zn−1 is birational for all n, so that (3) also holds. For each n, we now define An := {µ ∈ A, C(µ, Yn) ⊂ Zn}. We claim that for all k ≤ n, we have Ck(An) = Zn. Since Zn is an irreducible component of Cn(A), it is clear that Cn(An) = Zn. Now πn−1 is birational hence closed, which implies πn−1( ¯S) = πn−1(S) for any subset S ⊂ Yn. We infer Zn−1 = πn−1(Zn) = πn−1(Cn(An)) = πn−1(∪µ∈AnC(µ, Yn)) = ∪µ∈AnC(µ, Yn−1) = Cn−1(An) , and we conclude by a descending induction. Finally we construct the subset A′. First we shall construct countable subsets of A with special properties. To any count- able subset N of An, we attach the integer dn(N ) = dim Cn(N ), and let rn(N ) be the number of irreducible components of Cn(N ). Suppose one can find a sequence of count- able sets N j such that dn(N j) is constant, and rn(N j) → ∞. Then dn(N ) > dn(N j) for N := ∪jN j. This shows that there exists a countable subset N of An maxi- mizing the pair (dn(N ), rn(N )) for the lexicographic order on N2. We claim that dn(N ) = dim Zn. Indeed if it were not the case, and since ∪AnC(µ, Yn) is Zariski dense in Zn, then we could find a valuation µ ∈ An such that C(µ, Yn) 6⊂ Cn(N ) which would contradict the maximality of (dn(N ), rn(N )). Since Zn is irreducible, for each n we have found a countable subset N n ⊂ An such that Cn(N n) = Zn. We conclude the proof by setting A′ := ∪nN n. (cid:3) 4. PROOF OF THE MAIN RESULTS In this last section, we explain how Theorem A can be deduced from Theorems 3.1. Proofs of Corollaries A and B are given at the end of this section. 4.1. The projection of the Riemann-Zariski space to the generic fiber. Proposition 4.1. Pick any projective normal variety Y , and any effective Cartier divisor E in Y . Denote by Y(E) the subset of the Riemann-Zariski space Y of Y consisting of those valuations whose center in Y is included in E. Then there exists a surjective and continuous map Π : Y(E) →]Y [ such that for any ν, the center of ν in Y is included in rY (Π(ν)) (with equality when ν is divisorial), and any divisorial point in ]Y [ has a preimage in Y(E) which is divisorial too. This result is well-known but we give a proof for sake of completeness. COUNTABILITY PROPERTIES OF SOME BERKOVICH SPACES 9 Proof. To construct Π, pick a finite collection of affine charts Yj ⊂ Y , each intersecting E, and such that ∪jYj ⊃ E. For each j, pick an equation fj ∈ k[Yj] of E. Recall that ]Yj[ is the set of all multiplicative semi-norms · : k[Yj] → R+ such that − log fj = +1. Then ]Y [ is the disjoint union of the ]Yj['s patched together is a natural way. For any valuation ν ∈ Y(E) with valuation ring Rν, we take j such that the center of ν in Yj is non-empty, and define the function Π(ν) ≡ · ν : k[Yj] → R+ by setting Π(ν)(f ) = − log (cid:16)sup{p/q ∈ Q+, such that f q/f p j ∈ Rν , p ∈ N, q ∈ N∗}(cid:17) for any f ∈ k[Yj]. We claim this is a multiplicative semi-norm. To simplify notation, we shall work with µ(f ) = exp(−Π(ν)(f )). We shall use repeteadly the fact that g ∈ k(Y ) belongs to the valuation ring Rν iff gn ∈ Rν for some n ∈ N∗. Fix f ∈ k[Yj], and let I(f ) = {p/q ∈ Q+, such that f q/f p j ∈ Rν}. Then I(f ) is a j )q × = (f q′ /f p′ j )q′ segment containing 0. Indeed pick p/q > p′/q′ ∈ I(f ). Then (f q/f p f qp′−pq′ j ∈ Rν, and (f q/f p j ) ∈ Rν. Next assume f q/f p j , gq/f p 0 ≤ i ≤ q we have (f q−igi)/f p µ(f + g) ≥ min{µ(f ), µ(g)}. j ∈ Rν. The same argument as before shows that for all j ∈ Rν. This shows j ∈ Rν. Whence (f + g)q/f p /f p′ j , gq′ Finally if f q/f p j ∈ Rν, then (f g)qq′ ∈ Rν so that µ(f g) ≥ µ(f ) + µ(g). Conversely, if p0/q0 > µ(f ) + µ(g) then we may find two rational numbers p/q > µ(f ), p′/q′ > µ(g) such that p0/q0 = p/q + p′/q′. Since f p ∈ Rν, we get f p0 j /(f g)q0 ∈ Rν. This proves µ(f g) = µ(f ) + µ(g), and concludes the proof that Π(ν) is a multiplicative semi-norm. j /f q, f p′ /f pq′+p′q j /gq′ j To see the continuity of Π we pick a (net) νn converging to a valuation ν in the j ∈ Rν, j ∈ Rνn too, j /f q ∈ Rν (i.e. Π(ν)(f ) ≥ Riemann-Zariski space, and we pick f ∈ k[Yj]. Take p/q ∈ Q+ such that f q/f p i.e. Π(ν)(f ) ≤ − log(p/q). Then for an index n large enough f q/f p hence limn Π(νn)(f ) ≤ Π(ν)(f ). Conversely, suppose f p − log(p/q)). The same argument shows limn Π(νn)(f ) ≤ Π(ν)(f ). Let Z ⊂ E be the center of a valuation ν ∈ Y(E), and suppose Z ∩ Yj 6= Ø. Then rY (Π(ν)) is described by the prime ideal of functions f ∈ k[Yj] such that Π(ν)(f ) < 1, or in an equivalent way such that f q/f p , hence f too, and rY (Π(ν)) ⊃ Z. j ∈ Rν for some p/q > 0. But fj ∈ mRνj By construction, any rank 1 valuation ν ∈ Y(E) is mapped to the unique norm · on k(Y ) such that Rν = {f, log f ≤ 1} and − log fj = +1 in some chart. Since norms are dense in ]Y [, and Π is continuous, we infer that Π is surjective. We complete the proof by noting that Π maps divisorial valuations to divisorial points. (cid:3) We also collect the following result for later reference Lemma 4.2. Pick any projective normal variety Y , and any effective Cartier divisor E in Y whose support is complete. Then Y(E) is quasi-compact, and ]E[ is compact. Proof. It is a theorem of Zariski [ZS] that Y is quasi-compact. The result follows since Y(E) is a closed subset of Y. (cid:3) 10 CHARLES FAVRE 4.2. Proof of Theorem A. Recall our assumption: X is a normal algebraic variety, D is an effective Cartier divisor in X, and A is any subset of ]X[. We first cover X by finitely many affine open sets Xi, and write Di = D ∩ Xi. We let ]Xi[ be the normalized generic fiber of the formal completion of Xi along Di. This is an open (dense) subset of ]X[, and ∪i ]Xi[ = ]X[ . For each i, we fix a projective (normal) compactification Xi ⊂ ¯Xi. We write ¯Di for the closure of D in ¯Xi. It is a priori only a Weil divisor, but we may suppose it is Cartier by taking the normalized blow up associated to the ideal defining ¯Di. Now pick any point x in the closure of A in ]X[. It belongs to ]Xi[ for some i. Since ]Xi[ is open in ]X[, the point x also lie in the closure of Ai = A∩ ]Xi[ . Apply Proposition 4.1 to Y = ¯Xi, and E = ¯Di. This yields a continuous and surjective map Π : Y(E) → ] ¯Xi[ ⊃ ]Xi[ . Pick any valuation ν ∈ Π−1(x) ⊂ Y(E). Since Π is continuous, Π−1(x) is included in the closure of Π−1(Ai). Theorem 3.1 implies the existence of a sequence of valuations νn ∈ Π−1(Ai) converging to ν so that Π(νn) → x. 4.3. Proofs of Corollaries A and B. As in the previous section, X is a normal algebraic variety, and D is an effective Cartier divisor. We first prove Corollary A. Pick any subset A of ]X[. First take x in the closure of A. By Theorem A, there exists a sequence xn ∈ A such that xn → x. Now suppose A is relatively ω-compact. We need to show that it is relatively compact in ]X[. Consider ¯X any complete algebraic variety that contains X as a Zariski dense subset and such that the closure ¯D of D in ¯X is still Cartier. Such a space is given by Nagata's theorem, see [CLO] for a modern account. Let A′ be the closure of A in ] ¯X[. By Lemma 4.2, ] ¯X[ hence A′ are compact. Now suppose by contradiction that we can find x ∈ A′\ ]X[. Theorem A applied to A in ¯X implies the existence of a sequence xn ∈ A such that xn → x. Since A is relatively ω-compact, xn admits a cluster point in ]X[ which is absurd. Thus A′ is included in ]X[ which proves that A is relatively compact. These concludes the proof of Corollary A. We now prove Corollary B. Pick any x ∈ Xη. Note that for any closed subset C of D X (rX (x)) is an open neighborhood of x X (rX (x))). We the preimage r−1 in Xη. We claim that divisorial norms are dense in Xη (hence in ]X[ ∩r−1 conclude by applying Theorem A. X (C) is open. In particular, r−1 To justify our claim, we proceed as follows. Just as in the proof of Theorem A, we cover X by affine charts Xi and take projective compactifications ¯Xi of Xi. The set of divisorial valuations on k( ¯Xi) that are centered in ¯Di is dense in the subset of the Riemann-Zariski space of ¯Xi of valuations centered in ¯Di. By Proposition 4.1, this shows divisorial norms are dense in ] ¯Xi[. Since ]Xi[ is open in ] ¯Xi[, divisorial norms are also dense in ]Xi[, hence in ]X[ as required. 4.4. Proofs of Corollary C. Since a compact analytic space over k((T )) is covered by finitely many affinoids, and any affinoid is a closed subset in a ball of a suitable dimension, it is sufficient to treat the case of the unit ball in An k((T )). Let D is the hyperplane {x1 = 0} in the affine space (x1, ..., xn+1) ∈ X := An+1 . The normalized generic fiber ]X[ of the formal completion of X along D is the set k COUNTABILITY PROPERTIES OF SOME BERKOVICH SPACES 11 of multiplicative semi-norms · : k[x1, ..., xn+1] → R+ trivial on k, and such that x1 = e−1 and xi ≤ 1 for all i ≥ 2. This is precisely the unit ball in X = An k((T )). By Theorem A, the unit ball in any dimension is thus angelic. REFERENCES [BR] M. Baker, and R. Rumely. Potential theory and dynamics on the Berkovich projective line. Mathe- matical Surveys and Monographs, 159. American Mathematical Society, Providence, RI, 2010. [Ber] V.G. Berkovich. Spectral theory and analytic geometry over non-Archimedean fields. Math. Surveys Monographs 33, Amer. Math. Soc. Providence RI, 1990. [BGR] S. Bosch, U. Guntzer, R. Remmert. Non-Archimedean analysis. Grundlehren der Mathematischen Wissenschaften, 261. Springer-Verlag, Berlin, 1984. xii+436 pp. [CLO] B. Conrad, M. Lieblich and M. Olsson. Nagata compactification for algebraic spaces. arxiv.org [Fa] [FJ] C. Favre. Compactification of affine varieties with non-archimedean techniques. In preparation. C. Favre, and M. Jonsson. The valuative tree. Lecture Notes in Mathematics, 1853. Springer-Verlag, Berlin, 2004. K. Floret. Weakly compact sets. Lecture Notes in Mathematics, 801. Springer-Verlag, Berlin, 1980. N. Maınetti. Sequential compactness of some analytic spaces. J. Anal. 8 (2000), 39 -- 54. J. Poineau. Compacit´e s´equentielle. Manuscript. A. Thuillier. G´eom´etrie toroıdale et g´eom´etrie analytique non archim´edienne. Application au type d'homotopie de certains sch´emas formels. Manuscripta Math. 123 (2007), no. 4, 381 -- 451. M. Vaqui´e. Valuations. in Resolution of singularities, Progress in Math., Vol. 181, Birkhauser Ver- lag, 539 -- 590. [Fl] [M] [P] [T] [V] [ZS] O. Zariski and P. Samuel. Commutative algebra. Vol. 2. Graduate Texts in Mathematics, No. 29. Springer-Verlag, New York-Heidelberg-Berlin (1975). CNRS, ´ECOLE POLYTECHNIQUE, PALAISEAU, 91128 CEDEX FRANCE E-mail address: [email protected]
1211.4624	1	1211	2012-11-19T23:22:37	Another proof of the Semistable Reduction Theorem	[ "math.AG", "math.NT" ]	We give a new proof of the Semistable Reduction Theorem for curves. The main idea is to present a curve $Y$ over a local field $K$ as a finite cover of the projective line $X=\PP^1_K$. By successive blowups (and after replacing $K$ by a suitable finite extension) we construct a semistable model of $X$ whose normalization with respect to the cover is a semistable model of $Y$.	math.AG	math	Another proof of the Semistable Reduction Theorem Kai Arzdorf and Stefan Wewers Abstract We give a new proof of the Semistable Reduction Theorem for curves. The main idea is to present a curve Y over a local field K as a finite cover of the pro jective line X = P1 K . By successive blowups (and after replacing K by a suitable finite extension) we construct a semistable model of X whose normalization with respect to the cover is a semistable model of Y . 1 Introduction 1.1 Let K be a field which is complete with respect to a discrete valuation v . We let R denote the valuation ring of v , m R the maximal ideal of R and k := R/m the residue field. Let X be a smooth pro jective and absolutely irreducible curve over K . A model of X is a normal, flat and proper R-scheme XR such that XR ⊗R K = X . Given a model XR of X , its special fiber is denoted by Xs := XR ⊗R k . The k-scheme Xs is proper, connected and of pure dimension one. We say that XR is a semistable model of X if Xs is nodal, i.e. all singular points are ordinary double points. We say that X has semistable reduction if it has a semistable model XR . Theorem 1.1 (Semistable Reduction Theorem) There exists a finite ex- tension L/K such that the curve XL := X ⊗K L has semistable reduction (w.r.t. the unique extension of v to L). The first proof of this theorem was given by Deligne and Mumford ([11], Corollary 2.7). Since then, many more proofs have appeared in the literature, see e.g. [1]. 1.2 The question that originally motivated the present paper is: how can one explicitly determine a semistable model of a given curve? In a way the proof of Theorem 1.1 by Deligne and Mumford is constructive: choose n ≥ 3 prime to the residue characteristic of K and let L/K be the smallest field extension over which the n-torsion points of the jacobian of X become rational. Then the minimal regular model of XL is semistable. In theory, this gives an algorithm 1 to determine a semistable model. It seems, however, that several steps in this algorithm are today still computationally too expensive to be practical for curves of genus g ≥ 3. In this paper we work out a new proof of Theorem 1.1 which we hope will ultimately lead to a more practical algorithm. The starting point of our in- vestigation was a paper by M. Matignon ([16], see also [14]), which gives an algorithm to compute the semistable reduction of p-cyclic covers of the pro jec- tive line (satisfying an additional assumption). Trying to generalize Matignon’s method to a more general situation, we noticed that it could be used as a germ for a new proof of Theorem 1.1. 1.3 Let us give a brief sketch of our proof. The first idea is to view the curve under consideration as a finite cover of the pro jective line. So we start with a smooth pro jective K -curve Y and choose a nonconstant separable finite morphism φ : Y → X := P1 K . For any model XR of X we obtain a model YR of Y and a finite R-morphism φR : YR → XR by normalization of XR in Y . The goal is now to determine a semistable model XR of X such that YR is semistable as well. We show that this is possible after replacing K by a finite extension (Theorem 2.10) and obtain Theorem 1.1 as an immediate consequence. In order to prove Theorem 2.10, we may assume that the cover φ is Galois. Let G denote the Galois group of the cover φ. If the order of G is prime to the residue characteristic of K , then it is well known how to obtain a model XR with the desired properties: it suffices to take a semistable model which separates the branch points of the cover φ (see e.g. [15], §10.4). In particular, if the residue characteristic is zero, then the Semistable Reduction Theorem is relatively easy to prove. In the general case, let XR be any semistable model of X = P1 K . Let YR be the normalization of XR in Y . By a theorem of Epp ([12]) we may assume that the special fiber Ys of YR is reduced. If YR is semistable then we are done. Otherwise, there exists a singular point y ∈ Ys which is not an ordinary double point. Let x ∈ Xs denote the image of y under the map φR . The crucial step R → XR with center x of our proof is to show that there exists a blowup f : X (cid:48) which ‘improves the situation’. To make this a bit more precise, let Y (cid:48) R denote R → YR is a blowup R in Y . Then the induced map g : Y (cid:48) the normalization of X (cid:48) with center φ−1 R (x). We say that X (cid:48) R is an improvement of XR at y if the singularities of the special fiber of Y (cid:48) R which lie on the fiber g−1 (y) are ‘less bad’ than the singularity y ∈ Ys (‘badness’ of singularities can be measured by a suitable numerical invariant). Once the existence of an improvement has been shown, the proof of Theorem 2.10 is straightforward: start with some semistable model XR of X (e.g. the smooth model P1 R ) and repeatedly apply the above improvement procedure. After a finite number of steps we obtain a model X (cid:48) R whose normalization in Y is a semistable model of Y . Our proof that an improvement exists is local in the sense that it depends only on the formal completions of XR at x and of YR at y . Instead of working with formal schemes, the crucial step of the proof is phrased in the language of 2 rigid geometry, as follows. Let X rig and Y rig denote the rigid analytic spaces associated to the K -curves X and Y . We consider the formal fiber Xx :=]x[XR ⊂ X rig of the point x, i.e. the subset of points on X rig which specialize to x, and likewise the formal fiber Yy :=]y [YR ⊂ Y rig of y . Then φ induces a finite Galois cover φy : Yy → Xx of smooth rigid analytic spaces of dimension one. The fact that x is a smooth point of the special fiber Xs implies that Xx is an open disk, i.e. isomorphic to the rigid space { t ∈ A1 K t < 1 }. of K there exists a parameter t as above and some ∈ (cid:112)K × , with 0 < < 1, Let D ⊂ Xx be an affinoid disk. By this we mean that after a finite extension such that D is the open subspace of Xx defined by the condition t ≤ . To R → XR with center x whose exceptional D one can associate a blowup f : X (cid:48) R → YR induced by f is associated to divisor is a (−1)-curve. The blowup g : Y (cid:48) y (D) ⊂ Yy . the affinoid subdomain U := φ−1 We say that the affinoid disk D ⊂ Xx is exhausting if the complement Yy − y (D) decomposes as a union of open annuli. Let D denote the set of all φ−1 exhausting disks. A well known lemma (see e.g. [7], Lemma 2.4) says that every sufficiently large affinoid disk D ⊂ Xs is exhausting. In particular, D is nonempty. One easily shows: R → XR associated to D is an improvement at Proposition: The blowup f : X (cid:48) y if and only if D is a minimal element of D (with respect to inclusion). Therefore, we have reduced the proof of the Semistable Reduction Theorem to the claim that the set D has a minimal element.1 Our proof of the existence of a minimal exhausting disk is divided into two cases. We first assume that the Galois group G of the cover is solvable. In this case the proof can be easily reduced to the case that G is cyclic of prime order. Under the latter assumption, there is an explicit construction of the minimal exhausting disk, based on methods introduced by Matignon in [16]. This is worked out in detail in the first author’s thesis [2]. If G is not solvable then we argue by contradiction, and assume that the set D of all exhausting disks does not have a minimum. To each disk D in D we associate a point xD ∈ Xan on the Berkovich analytic space associated to X (essentially, xD corresponds to the maximum norm on D). A compactness argument shows that the sequence xD converges to a point x0 ∈ Xan . Points on Xan fall into four different classes, see [3], §1.4. We then show that in each of these four cases we can derive a contradiction, thus proving the claim. A crucial fact used in these arguments is that ‘inertia groups are solvable’. 1.4 The argument sketched above is really quite different from the tradi- tional proofs and requires less heavy machinery than most of them. For instance, 1The second named author has learned the idea of reducing the semistable reduction the- orem to the above statement from a lecture of Raynaud at a conference in Rennes in 2009 ([17]) 3 we do not use ´etale cohomology nor the Picard functor nor resolution of singu- larities. Our use of rigid analytic geometry is very limited and could be easily replaced by more elementary arguments. Ultimately, our proof relies on val- uation theoretic arguments. In this sense it may be considered to be similar in nature to Temkin’s proof of the stable modification theorem for families of curves ([18]), although this is a much deeper and more difficult result. The solvable case of our proof is truly constructive and gives a concrete and useful algorithm to compute semistable reduction of curves in the cases where it applies. Examples where the curve is a cyclic cover of the pro jective line of order p are worked out in [2]. In the nonsolvable case our argument is, as it is written down here, fundamentally nonconstruction. Nevertheless we believe that a future variant will yield a constructive and practical method as well. Acknowledgements: We would like to thank Andrew Obus for useful com- ments on an earlier version of this paper. 2 Semistable reduction for covers 2.1 Let K be a field which is complete with respect to a discrete valuation v : K × → R. We let R denote the valuation ring of v , m R the maximal ideal of R and k := R/m the residue field. We also let π denote a uniformizer of R; the particular choice of π will play no role. We make the additional assumption that the residue field k is algebraically closed. By [15], Lemma 10.4.5 this is no restriction of generality, as far as the Semistable Reduction Theorem is concerned. We remark that our base ring R is a complete discrete valuation ring and is therefore excellent (see [15], §8.2). As a consequence, all schemes and formal schemes occuring in this paper will be automatically excellent. This fact will be used in several places throughout the paper. For instance, if A is a localization of an R-algebra of finite type, then A is also excellent. 2.2 Reduced special fiber and permanence Let X be a smooth pro jec- tive and absolutely irreducible curve over K . A pre-model of X is a flat and proper R-scheme XR such that XR ⊗R K = X . Note that a pre-model is a model if and only if it is normal. Lemma 2.1 Let XR be a pre-model of X . Let Xs := XR ⊗R k be the special fiber of XR . Then XR is normal (i.e. a model) if and only if the following two conditions hold. (a) The special fiber Xs has no embedded points (see [15], Definition 7.1.6). (b) The local ring OXR ,η is a discrete valuation ring, for every generic point η of Xs . 4 Proof: This follows from Serre’s criterion for normality, see [15], Theorem 2.23. Indeed, (b) is equivalent to the Condition (R1), whereas (a) means that Xs satisfies Condition (S1). By loc.cit., Proposition 2.11, the latter is equivalent 2 to Condition (S2) for XR . Corollary 2.2 Let XR be a pre-model of X with special fiber Xs . (i) If Xs is reduced then XR is normal. (ii) Assume that XR is normal. Then Xs is reduced if and only if Xs is reduced in codimension zero (i.e. for every generic point η ∈ Xs the local ring OXs ,η is a field). Proof: Assume that Xs is reduced. We have to show that Condition (a) and (b) of the lemma hold true. This is obvious for Condition (a). In order to verify Condition (b), let η ∈ Xs be a generic point. Then A := OXR ,η is a noetherian local ring of dimension one. We have OXs ,η = A/πA. Therefore, our assumption that Xs is reduced shows that πA is the maximal ideal of A. It follows that A is a discrete valuation ring, i.e. Condition (b) of the lemma holds as well. This proves Assertion (i) of the corollary. (See [15], Lemma 1.18, for a direct proof which does not use the Serre criterion.) For the proof of (ii) we assume that XR is normal. Then Xs has no embedded points (Condition (b)). Since Xs has dimension one, this means that Xs is 2 reduced if and only if it is reduced in codimension zero. Let YR be a model of Y , and let L/K be a finite extension. Let S denote the integral closure of R in L. The normalized base change of YR to L is defined to be the normalization YS := (YR ⊗R S )∼ of the scheme YR ⊗R S . Clearly, YS is a model of YL . Proposition 2.3 Let YR be a model of Y . Then there exists a finite extension L/K such that the normalized base change YS := (YR ⊗R S )∼ of YR to L has a reduced special fiber. Furthermore, if L(cid:48)/L is any finite extension, then the usual base change YS ⊗S S (cid:48) is normal. (Here S and S (cid:48) denote the integral closures of R in L and L(cid:48) .) Proof: Let η be a generic points of Ys . Since YR is assumed to be normal, the local ring Aη := OYR ,η is a discrete valuation ring dominating R, by Condition (b) of Lemma 2.1. By Corollary 2.2 (ii), the special fiber is reduced if and only if OYs ,η = Aη /πAη is a field. The latter condition holds if and only if π is a uniformizer of Aη . Let L/K be a finite extension, S the valuation ring of L, π (cid:48) a uniformizer of S and YS := (YR ⊗R S )∼ the normalized base change. Let η (cid:48) be a generic 5 point of the special fiber of YS lying over η . Then the local ring Aη (cid:48) := OYS ,η (cid:48) is a discrete valuation ring dominating S , . Moreover, Aη (cid:48) is a direct factor of the integral closure of Aη ⊗R S . Now it follows from a theorem of Epp ([12]) that there exists a finite extension L/K such that π (cid:48) is a uniformizer for A(cid:48) η , for every generic point η (cid:48) of the special fiber of YS . This implies that the special fiber of YS is reduced. Recall that the residue field k of K is assumed to be algebraically closed. So every finite extension of K has residue field k . In particular, if L(cid:48)/L is a further finite extension, then YS and YS ⊗S S (cid:48) have the same special fiber, which is reduced. Now it follows from Corollary 2.2 (i) that YS ⊗S S (cid:48) is normal. This 2 completes the proof of the proposition. Definition 2.4 A model XR of X with reduced special fiber is called perma- nent.2 Proposition 2.3 says that every given model of X becomes permanent after normalized base change to a suitable finite extension of the base field. Further- more, permanent models are permanent in the sense that their special fibers are unchanged under any finite extension of the base field. Therefore, we may al- ways assume, while proving the Semistable Reduction Theorem, that any given model is permanent. Let YR be a permanent model of Y with special fiber Ys . Let Ys denote 2.3 the normalization of Ys . Note that Ys is a smooth (not necessarily connected) k-curve and that we have a finite morphism p : Ys → Ys which is an isomorphism when restricted to the smooth part of Ys . For a closed point y ∈ Ys we set /OYs )y δy := dimk (p∗O Ys and my := p−1 (y). It is easy to see that δy ≥ my − 1. Proposition 2.5 Let y be closed point of Ys . (i) The point y is a smooth point of Ys if and only if δy = 0. (iii) We have δy . (1) (ii) The point y is an ordinary double point of Ys if and only if δy = 1 and my = 2. (cid:88) (cid:88) y∈Ys V Here V runs over the irreducible components of Ys and gV denotes the genus of the normalization of V . (gV − 1) + g = 1 + 2This terminology is also inspired by Raynaud’s talk [17]. 6 Proof: This is well known. See for e.g. [15], Proposition 7.5.4 and Proposi- 2 tion 7.5.15. Let YR be a permanent model of Y . Let Ys denote the special fiber of 2.4 YR . R → YR , where Definition 2.6 A modification of YR is an R-morphism f : Y (cid:48) Y (cid:48) R is another model of Y and f is the identity on the general fiber. The modification f is called permanent if Y (cid:48) R is permanent. The subset of Ys where f is not an isomorphism is called the center of f . R → YR has connected fibers because YR Note that a modification f : Y (cid:48) is normal. An irreducible component W of the special fiber Y (cid:48) s of Y (cid:48) R is called exceptional of f (W ) is a closed point of Ys . The union of the exceptional compo- nents is called the exceptional divisor. The union of the irreducible components of Y (cid:48) s which are not exceptional is called the strict transform of Ys . The nor- malization p : Ys → Ys factors through a finite map p(cid:48) : Ys → Y (cid:48) s . The image of p(cid:48) is precisely the strict transform. R → YR is called simple if the Definition 2.7 A permanent modification f : Y (cid:48) following holds: (i) Every exceptional component intersects the strict transform. (ii) Every point of intersection of an exceptional component with the strict transform is an ordinary double point of Y (cid:48) s . R → YR be a simple modification and y ∈ Ys a singular Definition 2.8 Let f : Y (cid:48) point. Then f is called an improvement at y if for every closed point y (cid:48) ∈ f −1 (y) which does not lie on the strict transform of Ys we have δy (cid:48) < δy δy (cid:48) = δy , my (cid:48) > my . or R → YR be a simple modification and y ∈ Ys a singular Lemma 2.9 Let f : Y (cid:48) point which lies in the center of f . Assume that f is not an improvement at y . Then: (i) The fiber W := f −1 (y) has a unique singular point y (cid:48) . (ii) Every irreducible component of W intersects the strict transform of Ys in a unique point distinct from y (cid:48) . (iii) The normalization of every irreducible component of W has genus zero and contains a unique point lying over y (cid:48) . 7 gV + Proof: Let Z ⊂ Y (cid:48) s denote the strict transform of Ys . Write W ∩ Z = {y (cid:48) m }. By Condition (iii) of Definition 2.7, each point y (cid:48) 1 , . . . , y (cid:48) i is an ordinary double point of Y (cid:48) s and hence a smooth point of Z and of W . It follows that the finite map p(cid:48) : Ys → Z induces a bijection between the fiber p−1 (y) (where p : Ys → Ys is the normalization) and the set W ∩ Z . Hence m = my , in the notation of §2.3. We also have δy (cid:48) = 1 by Proposition 2.5 (ii). Set U := W − Z and let S denote the set of irreducible components of W . i Comparing the two expressions for the genus g obtained by applying Proposition (cid:88) (cid:88) 2.5 (iii) to YR and to Y (cid:48) R , one easily shows that δy = my − S + V ∈S y (cid:48)∈U It follows from Condition (ii) of Definition 2.7 that every component V ∈ S i (and this y (cid:48) contains at least one of the points y (cid:48) i lies on no other component in δy ≥ (cid:88) (cid:88) S ). Therefore, S ≤ my , and so (2) gives the inequality V ∈S y (cid:48)∈U Since by assumption f is not an improvement at y , there exists at least one point y (cid:48) ∈ U with δy (cid:48) ≥ δy . But then (3) implies that δy (cid:48) = δy ≥ 1, δy (cid:48)(cid:48) = 0 for all y (cid:48)(cid:48) ∈ U \{y (cid:48)} and gV = 0 for all V ∈ S . This prove (i) and the first half of (iii). Our argument also shows that the inequality (3) is actually an equality. It follows that my = S , and this proves (ii). Since W is connected and y (cid:48) the only singular point, every irreducible component must pass through y (cid:48) . This shows that S ≤ my (cid:48) . Finally, our assumption that f is not an improvement implies my (cid:48) ≤ my = S , which proves the second half of (iii). 2 δy (cid:48) . gV + δy (cid:48) . (2) (3) Let K (Y ) denote the function field of Y . The extension K (Y )/K is 2.5 a regular extension of transcendence degree one. Therefore, there exists an element x ∈ K (Y ) such that K (Y )/K (x) is finite and separable. The choice of x corresponds to a finite separable morphism φ : Y → X := P1 K (we identify the rational function field K (x) with the function field of P1 K ). We will prove the following ‘relative version’ of the Semistable Reduction Theorem. Theorem 2.10 Let φ : Y → X := P1 K be as above. Then (after replacing K by a finite extension) there exists a semistable model XR of X such that the normalization YR of XR in K (Y ) is a semistable model of Y . Obviously, Theorem 2.10 implies Theorem 1.1. The proof of Theorem 2.10 will occupy the rest of this paper. We start with a preliminary remark. Proposition 2.11 For the proof of Theorem 2.10 we may assume that the extension K (Y )/K (x) is Galois. 8 Proof: Let Y be the smooth pro jective curve whose function field K ( Y ) is the Galois closure of K (Y )/K (x). Let G := Gal(K ( Y )/K (x)) denote the Galois group and H ⊂ G the subgroup corresponding to K (Y ). Then G acts on Y and the natural map Y → Y identifies Y with the quotient curve Y /H . Let XR be a semistable model of X such that its normalization YR in K ( Y ) is a semistable model of Y . Then YR := YR /H is a semistable model of Y , see [15], Proposition 10.3.48. Moreover, the map φ : Y → X extends to a finite map YR → XR . Since YR is normal, YR is the normalization of XR in K (Y ). 2 This proves the proposition. We can now formulate our strategy to construct a semistable model of 2.6 Y . We choose a finite separable map φ : Y → X := P1 K . By Proposition 2.11, we may assume that φ is a Galois cover, with Galois group G. Let XR be a semistable model of X , and let YR denote the normalization of XR in Y . Then YR is a G-equivariant model of Y such that YR /G = XR . By Proposition 2.3 we may assume that XR and YR are permanent. Let φs : Ys → Xs be the finite map induced by φ. Note that φs is invariant under the induced action of G on Ys and that the induced map Ys /G → Xs is a homeomorphism. However, it may not be an isomorphism. It is an isomorphism only if G acts faithfully on each irreducible component of Ys . Definition 2.12 A closed point x ∈ Xs is called critical with respect to φ if the inverse image φ−1 s (x) contains a non-nodal point of Ys . We say that the semistable model XR is admissible with respect to φ if every critical point x is a smooth point of Xs . The following proposition is the crucial step in our proof of the Semistable Reduction Theorem. Proposition 2.13 Let XR be an admissible semistable model of X , relative to φ. Let YR be the normalization of XR in Y (which we assume is permanent). Let x ∈ Xs be a critical point. Then (after replacing K by a finite extension) R → XR with center x such that the there exists a simple modification f : X (cid:48) following holds. R denote the normalization of X (cid:48) (i) Let Y (cid:48) R in Y (which we assume is perma- R → YR is a simple modification. nent). Then the induced map g : Y (cid:48) (ii) The modification g is an improvement at every point y ∈ φ−1 s (x). The proof of this proposition is given in the remaining sections, starting with §3. 2.7 Assuming Proposition 2.13 for the moment we can give a proof of Theo- rem 2.10. Let φ : Y → X = P1 K be as above. Then the smooth model XR := P1 R is clearly admissible with respect to φ. Let YR be the normalization of XR in Y . 9 If YR is a semistable model then we are done. Otherwise, there exists a critical point x ∈ Xs . Since the inverse image φ−1 s (x) is a single G-orbit, the invariant δy defined in §2.3 is the same for all y ∈ φ−1 s (x). Hence we may write δx := δy . R → XR be a simple modification as in Proposition 2.13, relative Let f : X (cid:48) to x. Since x is a smooth point of Xs , the fiber f −1 (x) ⊂ X (cid:48) s is a smooth curve of genus zero, intersecting the strict transform of Xs transversally in a unique point x(cid:48) 0 (this follows easily from (2)). In particular, the model X (cid:48) R is semistable. If the normalization Y (cid:48) R of X (cid:48) R in Y is semistable, then we are done. Other- wise, let x(cid:48) ∈ f −1 (x) be a critical point with respect to φ. It follows from Condition (ii) in Proposition 2.13 and Definition 2.7 that x(cid:48) is a smooth point of Xs (this shows that the model X (cid:48) R is admissible for φ). By Definition 2.8 we have δx(cid:48) < δx or δx(cid:48) = δx , mx(cid:48) > mx . All in all we see that by repeated application of Proposition 2.13 we can either strictly decrease the invariant δx or keep it constant and increase mx . Since δx ≥ 0 and mx ≤ δx + 1, this process has to stop after a finite number of steps. It ends with a semistable model XR whose normalization in Y is a 2 semistable model of Y . 3 The rigid analytic point of view 3.1 We keep the assumption on our base field K . In the context of rigid analytic geometry it is more convenient to work with an absolute value instead of with an (exponential) valuation. We therefore choose a real constant 0 < q < 1 and set a := qv(a) for a ∈ K . Let X be a smooth pro jective K -curve. We let X rig denote the rigid analytic space associated to X , see e.g. [6] or [13]. Recall that the set underlying X rig is simply the set of closed points of X . Let XR be a permanent model of X . Given a point x ∈ X rig , its scheme theoretic closure in XR intersects the special fiber Xs in a unique point ¯x ∈ Xs , called the specialization of x. The resulting map spXR : X rig → Xs is surjective and is called the specialization map of the model XR . Let Z ⊂ Xs be a locally closed subscheme. Then the inverse image (Z ) ⊂ X rig ]Z [XR := sp−1 XR is an open set in the G-topology for X rig and hence is a smooth rigid analytic K -space. We call ]Z [XR the tube of Z in XR . See e.g. [5], §1. Remark 3.1 Let Z ⊂ Xs be a locally closed subscheme. Let X := XR (cid:98)Z be the formal completion of XR along Z . 10 (i) The tube ]Z [XR is canonically isomorphic to the generic fiber XK of X as constructed in [5], §1 (see also [10], §7). (ii) Let O◦ X rig denote the subsheaf of the structure sheaf on X rig consisting of functions that are bounded by 1. Then we have a canonical isomorphism ∼→ Γ(]Z [XR , O◦ Γ(X , OX ) X rig ), see [10], Theorem 7.4.1. (iii) The definition of ]Z [XR is compatible with base change to any finite ex- tension L/K . More precisely, we have a canonical isomorphism ]Z [XR ⊗K L ∼=]Z [XR⊗R S . Here S denotes the integral closure of R in L, and we identify the special fiber of XR with that of XR ⊗R S . Note that XR ⊗R S is again a permanent model by Proposition 2.3. (iv) The tube ]Z [XR is connected if and only if Z is connected. This follows immediately from (ii): any idempotent function on ]Z [XR is analytic and bounded by 1 and hence gives rise to an idempotent function on Z = X red . (v) Combining (iii) and (iv) shows that the connected components of ]Z [XR are absolutely connected. (vi) Suppose that Z = Spec ( ¯A) is an affine open subset of Xs . Then X = Spf (A) is an affine formal scheme, where A is flat and topologically of finite presentation over R, and is complete with respect to the π -adic topology (so A is admissible in the terminology of [8]). Therefore, AK := A ⊗R K is an affinoid K -algebra. Now it follows from the construction that ]Z [XR = XK = Spm(AK ) is an affinoid subdomain of X rig . In this special case, (ii) says that K := { f ∈ AK (cid:107)f (cid:107) ≤ 1 }. A = A◦ Here (cid:107) · (cid:107) denotes the maximum norm on the affinoid Spm(AK ). It follows that ¯A = A/πA and that Z = Spec ( ¯A) is the canonical reduction of the affinoid domain ]Z [XR (in the sense of [13], §4.8). Let us now consider the case where Z = {x} consists of a single closed 3.2 point of Xs . Then the tube X :=]x[XR is called the residue class of x (with respect to the model XR ). Let A := OXR ,x denote the complete local ring of the model XR at x. By Remark 3.1, X can be identified with the generic fiber of the formal R-scheme X = Spf (A). Moreover, A can be identified with the ring of analytic functions on X bounded by 1. It follows that the residue class X depends, as a rigid analytic space, only on the completion of XR at x. 11 Definition 3.2 An open analytic curve over K is a rigid analytic K -space X which becomes isomorphic, after a finite extension of K , to a residue class ]x[XR , where XR is a permanent model of a smooth pro jective K -curve X and x ∈ Xs is a closed point of the special fiber. The formal R-scheme X = Spf (A), where A = Γ(X, O◦ X ), is called the canonical formal model of X. A boundary point of X is a generic point of Spec (A/πA). The set of boundary points of X is denoted by ∂X. A boundary point η ∈ ∂X gives rise to a discrete valuation on Frac(A). Its residue field k(η) is a complete discrete valuation field containing k . It is thus isomorphic to k((t)). Suppose that X =]x[XR is a residue class as above. Then A = OXR ,x . It follows that a boundary point η ∈ ∂X corresponds to a local branch of Xs through x. We obtain a natural bijection between ∂X and the fiber p−1 (x), where p : Xs → X is the normalization of Xs . We write ¯η ∈ Xs for the point corresponding to η ∈ ∂X. Definition 3.3 An open analytic curve X over K is called an open disk if it is isomorphic to the standard open unit disk, i.e. to the rigid K -space { t ∈ A1 K t < 1 }. It is called an open annulus if it is isomorphic to for some ∈ (cid:112)K × , with < 1. { u ∈ A1 K < u < 1 }, Proposition 3.4 Let XR be a permanent model of a smooth pro jective curve over K . Let x ∈ Xs be a closed point of the special fiber and let X :=]x[XR denote the residue class of x. Then x is a smooth point (resp. an ordinary double point) of Xs if and only if X is an open disk (resp. an open annulus). Proof: (compare with [7], Proposition 2.2 and 2.3) Suppose first that XR = R is the pro jective line over R and x := 0 ∈ Xs = P1 P1 k the origin. It is then easy to see that the residue class of x is the standard open unit disk, and that OXR ,x = R[[t]], where t is the standard parameter on A1 R . Now let X be any smooth K -curve, XR a permanent model and x ∈ Xs . By ∼= R[[t]]. the above, the residue class ]x[XR is an open disk if and only if OXR ,x But the latter holds if and only if x is a smooth point. This proves the first equivalence. The proof of the second equivalence is similar, but we have to be more careful about the role of the base field K . As before, we can realize the standard open annulus { u ∈ A1 K < u < 1 } as the residue class of a point x on the special fiber of a model XR of P1 K . However, the model XR is permanent if and only if ∈ K × . If this is the case, 12 then OXR ,x = R[[u, v uv = a]], where a ∈ K × is any element with a = . With this in mind, the proof of the 2 second equivalence is analogous to the proof of the first. 3.3 We need a good notion of finite (Galois) covers of open analytic curves. Definition 3.5 Let X be an absolutely connected open analytic curve over K . An admissible cover of X is a finite and flat morphism φ : Y → X of rigid-analytic K -spaces, such that Y is also an open analytic curve. The cover φ is called a regular Galois cover if Y is absolutely connected and the automorphism group G := Aut(φ) has order deg(φ) (note that deg(φ) ∈ N is well defined). Let φ : Y → X be an admissible cover. Let A := Γ(X, O◦ X ) and B := Γ(Y, O◦ Y ). Then A is a normal and complete local domain with residue field k , and B is a finite A-algebra. Moreover, B is a normal complete semilocal ring, of the form B = ⊕iBi , where each Bi is a normal complete local domain with residue field k . After extending the base field K we may assume that A/πA and B/πB are reduced. Then φ is a regular Galois cover if and only if B is a domain and the field extension Frac(B )/Frac(A) is Galois. If this is the case, then the Galois group of φ can be identified with the Galois group of Frac(B )/Frac(A). Let φ : Y → X be a regular Galois cover, with Galois group G. Let H ⊂ G be a subgroup. Then the quotient Z := Y/H is again an absolutely irreducible open analytic curve. The induced maps Y → Z and Z → X are admissible covers. Moreover, Y → Z is a regular Galois cover with Galois group H , and Z → X is Galois if and only if H is a normal subgroup of G. 3.4 We now formulate our main result (Theorem 3.9), and show that it implies the Semistable Reduction Theorem. Let X be an open disk over K and φ : Y → X a regular Galois cover. Let G denote the Galois group of φ. (The assumption that X is a disk will be slightly relaxed in §4, but it will again be in force in §5). (i) A parameter for the open disk X is an element t ∈ A such Definition 3.6 that A = R[[t]]. (ii) A subset D ⊂ X is called a closed disk if there exists a parameter t for the open disk X and a real number , 0 < < 1, such that D = X(t ≤ ). the case iff ∈ (cid:112)K × ). If D is also an affinoid subdomain then it is called an affinoid disk (this is (iii) An affinoid disk D ⊂ Xx is called exhausting (with respect to φ) if the complement Y\φ−1 (D) is the disjoint union of open annuli. 13 Lemma 3.7 Let t be a parameter for the open disk X. Let , (cid:48) ∈ (cid:112)K × with We let D denote the set of all exhausting affinoid disks D ⊂ X. 0 < < (cid:48) < 1. (i) If X (t ≤ ) is exhausting then X (t ≤ (cid:48) ) is exhausting as well. (ii) There exists a constant 0 < 1, such that X (t ≤ ) is exhausting, for all ≥ 0 . Proof: This follows from [7], Lemma 2.4. 2 Corollary 3.8 The set D is nonempty. Moreover, if D(cid:48) ⊂ X is an affinoid disk containing an element D ∈ D, then D(cid:48) ∈ D. The following theorem is a ‘local’ version of Proposition 2.13 and is really the main result of the present paper. Theorem 3.9 Let φ : Y → X be a regular Galois cover of the open disk. Assume that Y is not an open disk. Then the set D of all affinoid disks D ⊂ X which are exhausting with respect to φ has a unique minimal element. The proof is given in §4 and §5, after some preliminary remarks in §3.5. In §3.6 we show that Theorem 3.9 implies the Semistable Reduction Theorem. We fix a regular G-Galois cover φ : Y → X of the open disk. We let 3.5 X = Spf (A) and Y = Spf (B ) denote the canonical formal models of X and Y. We let η ∈ ∂X denote the unique boundary point of X. We let k [η ] ⊂ k(η) denote the valuation ring of the residue field of η . In this section, we consider η as a morphism of formal schemes η : Spf (k [η ]) → X . Let D ⊂ X be an affinoid disk. It gives rise to a diagram of formal R-schemes (cid:121) (cid:121) Y (cid:48) −−−−→ Y X (cid:48) −−−−→ X , as follows. Let t ∈ A be a parameter and ∈ K × such that D = X(t ≤ ). Choose an element a ∈ R with v(a) = and let X (cid:48) → X be the formal blowup of the ideal I := (t, a) A. Let Z ⊂ X (cid:48) be the exceptional fiber (it is equal to the reduced subscheme (X )red , and it is isomorphic to P1 k ). The morphism ξ : Spf (k [η ]) → X lifts uniquely to a morphism ξ (cid:48) : Spf (k [η ]) → X (cid:48) . Let z ∈ Z denote the image of ξ (cid:48) and Z ◦ := Z \{z}. Then D =]Z ◦ [X (cid:48) . Let Y (cid:48) be the normalization of the formal scheme X (cid:48) in Y (see [9], §2.1). We call Y (cid:48) the formal model of Y induced by D. Let W := (Y (cid:48) )red denote 14 the reduced subscheme. Note that W is a connected pro jective k-curve. The canonical morphism Y (cid:48) → X (cid:48) restricts to a finite map W → Z . Let ∂W ⊂ W denote the inverse image of z and W ◦ := W \∂W . We have φ−1 (D) =]W ◦ [Y (cid:48) . It follows that D is exhausting with respect to φ if and only if the residue classes ]w[Y (cid:48) are open annuli, for all w ∈ ∂W . Actually, since G acts transitively on the set ∂W , it suffices that this holds for one w ∈ ∂W . Given a boundary point ξ ∈ ∂Y the morphism ξ : Spf (k [ξ ]) → Y lifts uniquely to ξ (cid:48) : Spf (k [ξ ]) → Y (cid:48) , and the image ¯ξ ∈ W of ξ (cid:48) lies in ∂W . We obtain a surjective G-equivariant map ∂Y → ∂W. (4) If D is exhausting, then this map is a bijection. Lemma 3.10 Assume that D ∈ D is exhausting. Then the following holds. (i) The affinoid E := φ−1 (D) is absolutely connected. (ii) The cover Y is an open disk if and only if E is a closed disk. (iii) Assume that Y is not an open disk. Then the set D of all exhausting closed disks is totally ordered with respect to inclusion. Proof: If D is exhausting, then W ◦ ⊂ W is the complement of a finite set of smooth points of W . Since W is connected it follows that W ◦ is connected as well. By Remark 3.1 (iv),(v) this implies that φ−1 (D) =]W ◦ [Y (cid:48) is absolutely connected, proving (i). The affinoid φ−1 (D) =]W ◦ [Y (cid:48) is a closed disk if and only if W ◦ ∼= A1 k . Under the assumption that D is exhausting, this holds if and only if W ∼= P1 k , Y has a unique boundary point ξ and the intersection of W with the corresponding formal subscheme Spf k [ξ ] ⊂ Y (cid:48) is transversal. By Castelnuovo’s criterion this holds if and only if Y ∼= Spf R[[s]], i.e. Y is an open disk. This proves (ii). To prove (iii) we let D(cid:48) ⊂ D be another exhausting disk, disjoint from D. We then have to show that Y is an open disk. We may write D = { t ≤ }, for some parameter t and some ∈ K × , 0 < < 1. Then there exists an (cid:48) > sucht that D (cid:48) is contained in a residue class of the affinoid A := {t = (cid:48)} (a closed annulus of thickness 0). Let X(cid:48) ⊂ A denote the residue class containing D(cid:48) . The inverse image Y (cid:48) := φ−1 (X(cid:48) ) is the disjoint union of residue classes of the affinoid φ−1 (A). Since D is exhausting, φ−1 (A) is a disjoint union of closed annuli of thickness 0 (this follows from [7], Lemma 2.4). In particular, φ−1 (A) is an affinoid with good reduction, and hence Y (cid:48) is the disjoint union of open disks. Applying (ii) to the connected components of the cover Y (cid:48) → X(cid:48) we see that φ−1 (D(cid:48) ) is a disjoint union of closed disks. By (i), φ−1 (D(cid:48) ) is connected and hence an open disk. Applying (ii) again shows that Y is an open disk, as 2 desired. This finishes the proof of the lemma. 15 Y = Yy , Yy :=]y [YR For the rest of this section we will show that Theorem 3.9 implies Propo- 3.6 sition 2.13 and hence the Semistable Reduction Theorem (as explained in §2.7). We return to the situation considered in §2.5. Let Y be a smooth pro jective and absolutely irreducible curve over K and let φ : Y → X := P1 K be a finite separable and nonconstant morphism to the pro jective line X . We assume that φ is a Galois cover, and let G denote its Galois group. The cover φ : Y → X induces a finite morphism of rigid analytic K -spaces φrig : Y rig → X rig . Let XR be a semistable model of X which is admissible for φ (Definition 2.12). Let YR denote the normalization of XR in Y . We assume that YR is a permanent model of Y . Let x ∈ Xs be a critical point. By assumption x ∈ Xs is a smooth point. It follows that the residue class X :=]x[XR is an open disk (Proposition 3.4). Set A := OXR ,x . Let Y := φ−1 (X) ⊂ Y rig denote the inverse image of the residue class X. This is an open rigid subspace of Y rig which is invariant under the action of the (cid:91) Galois group G. In fact, −1 y∈φ s (x) is the disjoint union of the residue classes of the points y ∈ Ys lying over x. The assumption that x is a critical point is equivalent to the statement that the residue classes Yy are not isomorphic to open disks. Therefore, for any y the induced cover φy : Yy → X is a Galois covers with Galois group Gy = StabG (y) which satisfies the hypotheses of Theorem 3.9. Let D ⊂ X be an affinoid disk. After replacing K by a finite extension we may a parameter for the open disk X. Moreover, D = X(t ≤ ) for some ∈ (cid:112)K × assume that D contains a K -rational point P . Choose an element t ∈ OXR ,x which has a simple zero at P and no other zero on the residue class X. Then t is with 0 < < 1. After a further extension of K we may assume that ∈ K × . R → XR be the blowup Choose an element a ∈ R with a = , and let f : X (cid:48) R (cid:98)Z the formal completion of X (cid:48) with center x of the ideal (t, a) OXR ,x . Let Z := f −1 (x) be the exceptional divisor and X (cid:48) := X (cid:48) R along Z . The natural morphism of formal R-schemes X (cid:48) → X = Spf (A) is the formal blowup of the ideal (t, a) A, see [8], §2. By the explicit description of formal blowups in loc.cit. one sees that Z ∼= P1 k is a smooth curve of genus zero which intersects the strict transform of Ys in a unique point z and that z is an ordinary double R → XR is a simple modification. Furthermore, s . In particular, f : X (cid:48) point of Y (cid:48) , with Z ◦ := Z − {z}. D =]Z ◦ [X (cid:48) R R → YR denote the modification induced by f (i.e. Y (cid:48) Let g : Y (cid:48) R is the normalization of X (cid:48) R in Y ). After a finite extension of the base field K we may R is permanent. Let W ⊂ Y (cid:48) assume that the model Y (cid:48) s denote the exceptional R → X (cid:48) R is finite, it restricts to a finite map W → Z . divisor of g . Since Y (cid:48) Moreover, the set ∂W ⊂ W of points where W intersects the strict transform of Y (cid:48) s is precisely the inverse image of z in W . It follows that , with W ◦ := W \∂W . φ−1 (D) =]W ◦ [Y (cid:48) R 16 We say that g is the modification of YR induced by D. Note that we have a natural surjective map ∂Y → ∂W (5) mapping a boundary point ξ ∈ ∂Y first to a point ¯ξ ∈ Ys on the normalization of Ys (see §3.2) and then to its image under the map p(cid:48) : Ys → Y (cid:48) s discussed after Definition 2.6. If the modification g is simple, then this map is a bijection. We call the affinoid disk D exhausting if it is exhausting with respect to the cover φy : Yy → X, in the sense of Definition 3.6 and for some y ∈ φ−1 (x) (in fact this condition is independent of y). As in §3.4 we let D denote the set of all exhausting affinoid disks D ⊂ X. Figure 1: Here D is minimal exhausting. Proposition 2.13 is now an immediate consequence of Theorem 3.9 and the following lemma. R → YR the induced Lemma 3.11 Let D ⊂ X be an affinoid disk and g : Y (cid:48) modification. (i) The modification g is simple if and only if D is exhausting. R → YR is an improvement at every (ii) Assume D is exhausting. Then g : Y (cid:48) point y ∈ φ−1 (x) if and only if D is a minimal element of D . Proof: Every irreducible component of W is a finite cover of Z and therefore contains a point w ∈ ∂W lying above z . As remarked above, these are precisely the points where W intersects the strict transform of Ys . It follows that g satisfies Condition (i) of Definition 2.7. Condition (ii) holds if and only if every point w ∈ ∂W is an ordinary double point of Y (cid:48) s . But the residue classes ]w[Y (cid:48) are precisely the connected components of Y\φ−1 (D). Therefore, Condition (ii) R of Definition 2.7 holds if and only if D is exhausting (here we use Proposition 3.4). This proves (i). For the proof of (ii) we assume that g is not an improvement at some y ∈ φ−1 (x). Set Wy := g−1 (y); this is a connected component of W . Set y := Wy ∩ W ◦ . By Lemma 2.9, Wy has a unique singular point y (cid:48) . All irre- W ◦ ducible components of Wy have geometric genus zero and are smooth outside 17 y (cid:48) . Moreover, they intersect the strict transform in a unique point (which is an ordinary double point distinct from y (cid:48) ) and have a unique branch passing through y (cid:48) . Let x(cid:48) ∈ Z ◦ denote the image of y (cid:48) . Then x(cid:48) is the only critical point with respect to φ which lies on Z . Let X(cid:48) :=]x(cid:48) [X (cid:48) denote the residue class of x(cid:48) and let Y (cid:48) := φ−1 (X) =]φ−1 (x(cid:48) )[Y (cid:48) R be the inverse image. Applying Lemma 3.7 to the restriction φ(cid:48) := φY(cid:48) : Y (cid:48) → X(cid:48) we find an affinoid disk D(cid:48) ⊂ X(cid:48) which R is exhausting with respect to φ(cid:48) . We claim that D(cid:48) , as an affinoid disk in X, is exhausting with respect to φ. Since D(cid:48) is strictly contained in D, this claim would prove the ‘if ’ part of (ii). R → Y (cid:48) To prove the claim, we consider the simple modification g (cid:48) : Y (cid:48)(cid:48) R induced from D (as an affinoid disk in X(cid:48) ). The center of g (cid:48) is precisely the singular locus of W . Let W (cid:48) ⊂ Y (cid:48)(cid:48) s denote the strict transform of W . By construction, X\φ−1 (D(cid:48) ) =]W (cid:48) [Y (cid:48)(cid:48) . R From the above description of W and the fact that g (cid:48) is a simple modification we see that W (cid:48) is isomorphic to the normalization of W . More precisely, W is a disjoint union of pro jective lines, each of which intersects the rest of Y (cid:48)(cid:48) in s exactly two points, which are ordinary double points Y (cid:48)(cid:48) s . It follows that the tube ]W (cid:48) [Y (cid:48)(cid:48) is the disjoint union of open annuli. This proves the claim and R hence the ‘if ’-part of (ii). The ‘only if ’ part is left to the reader (it is not used 2 in the rest of the paper). 4 The solvable case In this section we prove the existence of a minimal exhausting disk (Theorem 3.9) under the assumption that the Galois group G of the cover φ : Y → X is solvable. The proof is by induction on the order of G. The base case of the induction is when G is cyclic of prime order, and this case is treated in greater detail in [2]. To make the induction step work we actually have to consider a slightly more general situation. Namely, we allow X to be either a disk or an annulus. It is then natural to replace the notion of exhausting affinoid disk by separating boundary domain. We keep all our assumptions on the base field K . Let X be an open 4.1 analytic curve over K which is either an open disk or an open annulus. Let A = Γ(X, O◦ X ) be the ring of analytic functions on X bounded by 1. Then X is the generic fiber of the formal R-scheme X = Spf (A), see §3.1-§3.2. Let us choose a boundary point η ∈ ∂X. If X is an annulus, this amounts to choosing an ‘orientation’ of X (if X is a disk, there is no choice). A parameter for X (with respect to η) is an element t ∈ A which yields an isomorphism ∼→ { t ∈ A1 K 0 < t < 1 }, X 18 for some 0 < 0 if X is a disk and with 0 > 0 if X is an annulus. If X is a disk, then this implies A = R[[t]], and the new terminology agrees with the old one. If X is an annulus, then there exists an element a ∈ mR such that s := a/t ∈ A is a parameter for X with respect to the boundary point distinct from η , and we have A = R[[t, s ts = a]]. By a boundary domain of X (containing η) we mean an open rigid subspace U ⊂ X of the form where t is a parameter for X and ∈ (cid:112)K × , < 1. So if X is an open disk, U = X(t > ), then U = X − D, where D ⊂ X is an affinoid disk. In any case, U is an open annulus. Let φ : Y → X be a regular Galois cover of X, with Galois group G. A boundary domain U ⊂ X is called separating (with respect to φ) if the inverse image φ−1 (U) is the disjoint union of open annuli. Let U denote the set of all separating boundary domains. We consider U as a partially ordered set by inclusion. If X is an open disk, then a boundary domain U ⊂ X is separating with respect to φ if and only if the affinoid disk D := X\U is exhausting with respect to φ. It follows that the set U has a unique maximum if and only if the set D has a unique minimum. Therefore, the following proposition implies Theorem 3.9 in case the Galois group G is solvable. Proposition 4.1 Assume that (a) The open analytic curve Y is not an open disk. (b) The Galois group G of φ is solvable. Then the set U has a unique maximal element. For the proof of Proposition 4.1 we will use induction on the order of G. The case where G = {1} is trivial. Indeed, if G = {1} then X = Y is not a disk. It follows that X is an annulus, and that U := X is the unique maximal element of U . After a preliminary argument in §4.2, we prove the prime order case in §4.3–§??. Finally, the induction step is done in §4.5. Let x1 , . . . , xr ∈ X be the pairwise distinct branch points of φ (which 4.2 we assume to be K -rational). We claim that, in order to prove Proposition 4.1, we may assume that r ≤ 1 if X is a disk and r = 0 if X is an annulus. To prove this claim we assume that either r ≥ 2 and X is a disk or that r ≥ 1 and X is an annulus. Under this condition there exists a maximal boundary domain U0 ⊂ X containing none of the branch points. Assume, moreover, that U0 is not separating with respect to φ. Then φ−1 (U0 ) is not a union of open disks. Furthermore, a boundary domain U ⊂ U0 is separating with respect to φ : Y → X if and only it is separating with respect to the cover φ−1 (U0 ) → U0 . But φ−1 (U0 ) → U0 is ´etale by choice of U0 . So in this case it suffices to prove the proposition under the assumption that φ is ´etale (i.e. r = 0). 19 Now consider the case that U0 is separating. If it is maximal with this property then we are done. Hence we may assume that U0 is not maximal. For simplicity, we also assume that X is a disk (the other case is proved similarly). Then D0 := X\U0 is a non-minimal exhausting affinoid disk. In this situation it follows from Lemma 3.10 (iii) that there exists a unique residue class X(cid:48) ⊂ D0 containing all exhausting disks strictly contained in D0 . Furthermore, for any closed affinoid disk D ⊂ X(cid:48) , D is exhausting with respect to φ if and only it is exhausting with respect to the restricted cover φ−1 (X(cid:48) ) → X(cid:48) . But by the choice of U0 , D0 is the smallest closed disk containing all r ≥ 2 branch points. It follows that the residue class X(cid:48) ⊂ D0 contains strictly less then r branch points. Our claim now follows by induction on the number r of branch points. For the rest of the proof of Proposition 4.1 we may now assume that r ≤ 1, and that X is a disk if r = 1. 4.3 Let us assume that the group G is cyclic of prime order (cid:96). We have to distinguish two main cases, of which the first is divided into two subcases. We start with the assumption that the characteristic of K is prime to (cid:96). (For the time being, we make no assumption on the residue characteristic of K .) Then, after replacing K by some finite extension, we may also assume that K contains an (cid:96)th rooth of unity. Moreover, the cover φ : Y → X is given generically by a Kummer equation of the form y (cid:96) = f , where f ∈ A is not an (cid:96)th power. More precisely, the ring B := O(Y)◦ contains an element y which satisfies the above equation and B is the normalization of A in the field Frac(A)[y ]. Clearly, the ring B is unchanged if we divide f by an (cid:96)th power in A. We may therefore assume that the order of zero of f at any point x ∈ X is strictly less than (cid:96). Under this condition the zeroes of f are precisely the branch points of φ. We claim that there is no branch point, i.e. that φ is ´etale. To prove this claim we assume the converse. Then X is a disk and there is exactly one branch point, by our assumption made at the end of §4.2. We choose a parameter t for X such that the unique branch point is t = 0. Now the Weierstrass preparation theorem shows that f is of the form f = c(t + a2 t2 + . . .) ∈ A = R[[t]], with c (cid:54)= 0. After a finite extension of K , c is an (cid:96)th power, and we may therefore assume that c = 1. Applying [2], Lemma 1.28, one shows that B = A[y ] and that Y is an open disk with parameter y . But this contradicts Assumption (a) of Proposition 4.1, proving the claim. Let us now make the additional assumption that (cid:96) (cid:54)= p, i.e. that the residue characteristic of K is prime to (cid:96). Let us choose a parameter t for X. If X were a disk then A = R[[t]] and, by the above claim, f = 1 + a1 t + a2 t2 + . . .. But then Hensel’s lemma shows that f is an (cid:96)th power in A. This contradicts our assumption that φ is a regular Galois cover. We conclude that X is an open annulus and hence A = R[[t, s ts = a]]]. Using again Hensel’s Lemma we see that f = ctmu, with c ∈ R, m ≥ 1 and u a unit of A with constant coefficient 1. As before we may assume that c = 1. Dividing f by a suitable power of 20 t(cid:96) we may also assume that m < (cid:96). In this situation [2], Lemma 1.31, shows that Y is an open annulus. This means that X itself is the maximal separating boundary domain we are looking for. Therefore, Proposition 4.1 is proved in the case G = (cid:96) (cid:54)= p. We continue with the notation introduced in §4.3, but we now assume 4.4 that (cid:96) = p is the (positive) characteristic of the residue field k of K . (We keep the assumption that (cid:96) = p is prime to the characteristic of K , but this now amounts to saying that K has characteristic zero.) In this case, Proposition 4.1 is proved in [2] (Theorem 2.1 and Theorem 4.2). We briefly sketch the proof. For simplicity, we restrict to the case where X is an open disk. The proof in the case where X is an annulus uses the same methods, but is slightly more complicated. As we have seen before, we may assume that the cover φ is given generically by an equation of the form yp = f , where f is a principal unit in A = R[[t]]. Note that Hensel’s lemma is not applicable anymore, and we cannot conclude that f is a pth power. The first step is to compute the ring B . Let v0 : Frac(A) → Q ∪ {∞} denote the discrete valuation corresponding to the prime ideal πA A (normalized such that v0 (p) = 1 which implies that v0 K = v). Since ¯A := A/πA is isomorphic to a power series ring k [[t]], the residue field of v0 carries a canonical discrete valuation ¯v0 (normalized such that ¯v0 (t) = 1). Let vη : A → (Q × Z) ∪ {∞}) denote the discrete valuation of rank two obtained as the composition of v0 g = (cid:80)∞ with ¯v0 . Here we consider the target set as an ordered group with respect if t is a parameter for X and to the lexicographic ordering. More explicitly: i=0 gi ti ∈ A then vη (g) = (µ, m), where m = min{ i v(gi ) = µ}. v(gi ), µ = min i Proposition 4.2 The maximum (µ, m) := max{vη (f − hp ) h ∈ A×} exists. Furthermore: (i) 0 ≤ µ < p/(p − 1) and µ/p ∈ v(K × ). (ii) m > 1 and (m, p) = 1. (iii) Choose c ∈ K × such that v(c) = µ/p and h ∈ A× such that vη (f − hp ) = (µ, m). Then B = A[w], where w := (y − h)/c. 21 Proof: Let h ∈ A× be given and set (µ, m) := vη (f − hp ). Then µ = v0 (f − hp ) ≤ p/(p − 1) (otherwise, completness of A with respect to πA would show that f ∈ Ap ). Since µ ∈ v(K × ) takes values in a discrete group, we may assume that µ takes the maximal possible value. Let v0 be an extension of v0 to Frac(B ); it corresponds to a minimal prime ideal of ¯B := B/πB . Our running assumption says that ¯B is reduced, which means that π is a prime element for v0 . A simple calculation using µ ≤ p/(p − 1) and the equation yp = f shows that pv0 (y − h) = v0 (f − hp ) = µ. It follows that µ/p ∈ v(K × ). Choose an element c ∈ K with v(c) = µ/p and set g0 := (f − hp )/cp ∈ A. Then w := (y − h)/c satiesfies an irreducible equation over A: wp + . . . + pc1−php−1w + g0 = 0. (6) It follows that w ∈ B . If µ = p/(p − 1) then Hensel’s lemma, applied to the complete local ring A, would show that this equation is reducible. We conclude that µ < p/(p − 1). Now Part (i) of the proposition is proved. Let ¯w denote the image of w in ¯B := B/πB and ¯g0 the image of g0 in ¯A := A/πA = k [[t]]. Then m = ¯v0 (¯g0 ). It is easy to see that for any choice of h such that v0 (f − hp ) = µ we have ¯g0 (cid:54)∈ ¯Ap (otherwise µ wouldn’t be maximal). Moreover, choosing a different h results in adding to ¯g0 an element of ¯Ap . We may therefore assume that (p, m) = 1. Furthermore, m is now the maximal possible value for ¯v0 (¯g0 ). It follows that (µ, m) is the maximal possible value for vη (f − hp ). This proves the first claim of the proposition and Part (ii), except for the statement that m > 1. From (6) we see that ¯w satisfies the equation ¯wp = ¯g0 , which is irreducible because ¯g0 (cid:54)∈ ¯Ap . We conclude that v0 is the unique exten- sion of v0 to Frac(B ), with residue field extension purely inseparabel of degree p. Using Serre’s Normality Criterion (as in Lemma 2.1) we also see that A[w] is normal and hence B = A[w]. This proves (iii). Moreover, if m = 1 then we would have B = R[[w]], i.e. Y was an open disk. This contradicts our assump- 2 tion. Now the proof of the proposition is complete. We continue to use the notation (µ, m) from the above proposition. We note that the values of µ and m and do not change if we replace K by any finite extension. An element h ∈ A such that vη (f − hp ) = (µ, m) is called a best approximation of f (with respect to vη ) (cf. [2], §2.2.1). Lemma 4.3 (p-Taylor expansion) Let n ≥ 1 be given, and set νn := 1 + 1/p + . . . + 1/pn . After replacing K by a finite extension, there exists a parameter t for the disk X and an element h ∈ A = R[[t]] such that ∞(cid:88) f − hp = j=1 aj tj 22 with aj ∈ mR and for all j . v(apj ) ≥ νn , (7) Following [16], we call (t, h, aj ) a p-Taylor expansion of f of level n. Proof: See [2], Proposition 2.12. Since νn → p/(p − 1) for n → ∞, we may choose n such that − p/(p − 1) − µ p p − 1 νn > m Let (t, h, aj ) be a p-Taylor expansion of f of level n. Since νn > µ ≥ v0 (f − hp ), it follows from (8) that the minimum of the valuations v(aj ) occurs for an index j which is prime to p. By inspection of the proof of Proposition 4.2 one concludes that vη (f − hp ) = (µ, m), i.e. a p-Taylor expansion of f yields a best approximation (see [2], Corollary 2.16). We define (8) 2 . , 1 ≤ j < m }. v(aj ) − µ ρ := min { p/(p − 1) − µ m − j m Note that µ + mρ ≤ p/(p − 1). If equality holds, we set k := 0. Otherwise, we let k denote the smallest index such that 1 ≤ k < m and ρ = (v(ak ) − µ)/(m − k). Then by definition the Newton polygon of f − hp has a line segment of slope −ρ over the intervall [k , . . . , m]. Moreover, it follows from (7) and (8) that (k , p) = 1. One can check that ρ and k do not depend on the choice of h (see [2], Proposition 2.28). However, ρ and k may depend on the choice of the parameter t! Lemma 4.4 After a finite extension of K , there exists a p-Taylor expansion (t, h, aj ) of level n such that k (cid:54)= 1. ∞(cid:88) Proof: See [2], Proposition 2.31. The idea is to use a ‘generic’ p-Taylor expansion j=1 f (t + T ) − H p = Aj tj , where the Aj and the t-coefficients of H are algebraic functions in T . One has to show that, after a finite extension of K , there is a point T = ξ such that A1 (ξ ) = 0. Replacing the parameter t by t(cid:48) := t − ξ then gives a p-Taylor expansion (t(cid:48) , h(cid:48) , a(cid:48) j ) with a(cid:48) 1 = 0. With respect to this p-Taylor expansion we have k (cid:54)= 1. 2 Now the following proposition completes the proof of Proposition 4.1 in the special case considered in this subsection. 23 Proposition 4.5 Let (t, h, aj ) be a p-Taylor expansion of f such that k (cid:54)= 1 (notation as above). Then D := X(t ≤ ρ) is the minimal exhausting disk with respect to φ. Proof: Let U := X\D. Choose an element b ∈ R such that v(b) = ρ and set t1 := t/b, s1 := t−1 (in general, this requires again a finite extension of K ). 1 Then A(cid:48) := Γ(D, O◦ X ) = R{{t1 }} (the ring of convergent power series over R) and B (cid:48) := Γ(φ−1 (D), O◦ Y ) is the in- tegral closure of A(cid:48) in the extension of fraction fields given by yp = f . Similarly, A(cid:48)(cid:48) := Γ(U, O◦ X ) = R[[t, s1 ts1 = b]], By construction, the Newton polygon of the power series f − hp = (cid:80) and B (cid:48)(cid:48) := Γ(φ−1 (U), O◦ Y ) is the integral closure of A(cid:48)(cid:48) . j≥1 aj tj has a line segment of slope −ρ over the interval [k , m] and, moreover, −ρ is the largest negative slope that occurs. It follows that we can write f − hp , as an element of the ring A(cid:48)(cid:48) , in the form f − hp = cp 1 tmu1 , where c1 ∈ R is an element of valuation µ/p and u1 is a principal unit. It follows that the element w1 := (y − h)/c1 is an element of B (cid:48)(cid:48) satisfying an irreducible equation over A(cid:48)(cid:48) of the form 1 + . . . + pc1−p 1 hp−1 = tmu1 . wp Now it follows from [2], Lemma 1.31, that φ−1 (U) is an open annulus. We have shown that D is exhausting. To show that D is the minimal disk with this property we shall provide an explicit description of the canonical reduction W ◦ of φ−1 (D) (we freely use the notation set up in §3.5) . Set λ := µ + mρ. Then λ ≤ p/(p − 1) and we can write f − hp = cp 2 g , with g ∈ A(cid:48)(cid:48) and where c2 ∈ R is chosen such that v(c2 ) = λ/p. Then the element w := (y − h)/c2 satisfies the integral equation wp + . . . + pc1−p 2 hp−1 = g over A(cid:48)(cid:48) and therefore w ∈ B (cid:48)(cid:48) . Let ¯g ∈ ¯A(cid:48)(cid:48) := A(cid:48)(cid:48)/πA(cid:48)(cid:48) = k [t1 ] (resp. ¯w ∈ ¯B (cid:48)(cid:48) := B (cid:48)(cid:48)/πB (cid:48)(cid:48) ) denote the image of g (resp. of w). Suppose first that λ = p/(p − 1). Then ¯w satisfies an Artin-Schreier equation ¯wp + ¯c ¯w = ¯g . Moreover, ¯g ∈ k [t1 ] is a polynomial in t1 of degree m > 1, (m, p) = 1. It follows that W ◦ = Spec k [t1 , ¯w] is a smooth and connected affine curve of genus 24 g = (p − 1)(m − 1)/2 > 0. Using Lemma 3.11 (ii) we conclude that D is the minimal exhausting disk. Now suppose that λ < p/(p − 1). Then ¯w satisfies the equation ¯wp = ¯g , where ¯g = ¯gk tk 1 is a polynomial of degree m, divisible by tk 1 + . . . ¯gm tm 1 . The affine curve W ◦ = Spec k [t1 , ¯w] is irreducible and has a unibranched singularity precisely over each point of Z ◦ = Spec k [t1 ] where the differential d¯g has a zero. Since 1 < k < m and (p, k) = (p, m) = 1 it follows that W ◦ has at least two singular points. Using Lemma 3.11 (ii) again we conclude that D is the minimal 2 exhausting disk. 4.5 We can now prove the general case of Proposition 4.1. By the result of the previous subsections we may assume that G has a proper normal subgroup H G. Set Z := Y/H → X and consider the factorization Y H−→ Z G/H−→ X of φ into regular Galois subcovers. Suppose Z is not a disk. Then by the induction hypothesis there exists a maximal boundary component X1 ⊂ X which is separating with respect to Z → X. If Z is an open disk then we set X1 := X. Choose a connected component Y1 of the inverse image of X1 in Y. Let G1 ⊂ G denote the stabilizer of Y1 in G and set H1 := G1 ∩ H1 . Note that H1 G1 is a normal subgroup. We can identify the quotient Z1 := Y1/H1 (resp. Y1/G1 ) with a connected component of the inverse image of X1 in Z (resp. with X1 ). By construction, Z1 is either an open disk (and then Z1 = Z) or an open annulus. Applying the induction hypothesis once more to the H1 - cover Y1 → Z1 we obtain a maximal separating boundary domain Z2 ⊂ Z1 with respect to Y1 → Z1 . We claim that the subset Z2 ⊂ Z1 is fixed by the action of G1 /H1 . Indeed, any element g ∈ G1 induces an isomorphism of the cover Y1 → Z1 . It follows that g(Z2 ) ⊂ Z1 is also a maximal separating boundary domain with respect to Y1 → Z1 . Uniqueness shows that g(Z2 ) = Z2 . The quotient X2 := Z2 /(G1/H1 ) can be identified with a boundary domain of X. We claim that X2 is separating with respect to the cover φ : Y → X, and is maximal with respect to this property. Indeed, let Y2 denote a connected component of the inverse image of X2 in Y. We may assume that Y2 is contained in Y1 . Then Y2 is also a connected component of the inverse image of Z2 ⊂ Z1 in Y1 . By the choice of Z2 this means that Y2 is an open annulus. This shows that X2 is separating with respect to the cover φ. The maximality of X2 is proved in 2 a similar manner. This completes the proof of Proposition 4.1. 25 5 The nonsolvable case We return to the situation considered in §3.4: we are given a regular G- 5.1 Galois cover φ : Y → X, where X is an open disk and Y is an open analytic curve which is not an open disk. Our goal is to show that the set D of all affinoid disks D ⊂ X which are exhausting with respect to φ has a unique minimal element (Theorem 3.9). In view of Proposition 4.1 we may assume that the group G is not solvable. We let X = Spf (A) and Y = Spf (B ) denote the canonical formal mod- 5.2 els. We let η ∈ ∂X denote the unique boundary point of X. We let k [η ] ⊂ k(η) denote the valuation ring of the residue field of η . In this section, we consider η as a morphism of formal schemes η : Spf (k [η ]) → X . Let D ⊂ X be an affinoid disk. By a local variant of the procedure described in §3.6, D gives rise to a diagram of formal R-schemes (cid:121) (cid:121) Y (cid:48) −−−−→ Y X (cid:48) −−−−→ X , as follows. Let t ∈ A be a parameter and ∈ K × such that D = X(t ≤ ). Let X (cid:48) → X be the formal blowup of the ideal I := (t, a) A. Let Z ⊂ X (cid:48) be the exceptional fiber (it is equal to the reduced subscheme (X )red , and it k ). The morphism ξ : Spf (k [η ]) → X lifts uniquely to a is isomorphic to P1 morphism ξ (cid:48) : Spf (k [η ]) → X (cid:48) . Let z ∈ Z denote the image of ξ (cid:48) and Z ◦ := Z \{z}. Then D =]Z ◦ [X (cid:48) . Let Y (cid:48) be the normalization of the formal scheme X (cid:48) in Y (see ??). We call Y (cid:48) the formal model of Y induced by D. Let W := (Y (cid:48) )red denote the reduced subscheme. Note that W is a connected pro jective k-curve. The canonical morphism Y (cid:48) → X (cid:48) restricts to a finite map W → Z . Let ∂W ⊂ W denote the inverse image of z and W ◦ := W \∂W . We have φ−1 (D) =]W ◦ [Y (cid:48) . It follows that D is exhausting with respect to φ if and only if the residue classes ]w[Y (cid:48) are open annuli, for all w ∈ ∂W . Actually, since G acts transitively on the set ∂W , it suffices that this holds for one w ∈ ∂W . Given a boundary point ξ ∈ ∂Y the morphism ξ : Spf (k [ξ ]) → Y lifts uniquely to ξ (cid:48) : Spf (k [ξ ]) → Y (cid:48) , and the image ¯ξ ∈ W of ξ (cid:48) lies in ∂W . We obtain a surjective G-equivariant map ∂Y → ∂W, (9) the local analog of the map (5). If D is exhausting, then this map is a bijection. 26 We fix a boundary point ξ ∈ ∂Y. Let D ⊂ X be either an affinoid disk or 5.3 the empty set. Then we let V(D) denote the connected component of φ−1 (X\D) which ‘contains’ ξ . This means the following. If D = ∅ then V(D) is equal to Y. , where w ∈ ∂W On the other hand, if D is an affinoid disk then V(D) :=]w[Y (cid:48) is the image of ξ under the map (9). Note that D ∈ D if and and only if V(D) R is an open annulus. We let G(D) ⊂ G denote the stabilizer in G of the component V(D). Note that V(D)/G(D) ∼= X\D. Lemma 5.1 For D ∈ D the following holds. (i) The group G(D) is equal to the stabilizer Gξ in G of the boundary point ξ and is solvable. (ii) Let D(cid:48) ⊂ D be a subset which is either empty or an affinoid disk strictly contained in D. Then V(D(cid:48) )\φ−1 (D) is absolutely connected. In particular, φ−1 (D) is absolutely connected. Proof: Since D ∈ D , the map (9) is a bijection. The equality G(D) = Gξ follows immediately. Moreover, Gξ ⊂ G = Gal(Frac(B )/Frac(A)) is the decomposition group of the discrete valuation on Frac(B ) corresponding to ξ . We obtain a short exact sequence 1 → Iξ → Gξ → Gal(k(ξ )/k(η)) → 1, where Iξ is the inertia group of ξ . The residue field extension k(ξ )/k(η) is a finite extension of complete discrete valued fields with algebraic residue field. So Gal(k(ξ )/k(η)) is also an inertia group and hence solvable. We conclude that Gξ is solvable. For the proof of (ii) we first assume that D(cid:48) = ∅, and we use the notation introduced in §5.2. We have already noted that W is connected. Since D ∈ D , the subset ∂W consists of smooth points of W . It follows that the complement W ◦ = W \∂W is still connected. Now Remark 3.1 (iv),(v) shows that φ−1 (D) =]W ◦ [Y (cid:48) is absolutely connected, proving (ii) if D = ∅. The proof in the case D(cid:48) (cid:54)= ∅ is 2 similar and left to the reader. Let Xan denote the Berkovich analytic space associated to X, see [3]. As 5.4 a set, Xan consists of all continuous multiplicative seminorms · x : A → R≥0 bounded by 1 which extend the standard valuation · on R. To each point · x ∈ Xan we can associate its residue field H(x), which is defined as the completion of the fraction field of A/Ker( · x ). By construction, H(x)/K is an extension of complete valued fields. We let (cid:93)H(x) denote the residue field of H(x). 27 Any point x ∈ X gives rise to a point Xan by the formula f x := f (x). We may thus consider X as a subset of Xan (called the set of classical points). Classical points are characterized by the property that the extension H(x)/K is finite. In order to have a uniform and suggestive notation, we shall write x ∈ Xan instead of · x ∈ Xan and f (x) instead of f x , for arbitrary points on Xan . For instance, to any closed disk D ⊂ X (affinoid or not) we can associate a point xD ∈ Xan by setting f (xD ) := max f (x(cid:48) ). x(cid:48)∈D If D is affinoid, then the residue field (cid:93)H(x) can be identified with the function field of the canonical reduction of D. In particular, (cid:93)H(x)/k has transcendence degree one. Otherwise, (cid:93)H(x) = k . See [3], §1.4.4. The space Xan carries a natural topology which makes it a locally compact If D ∈ X is an affinoid disk, then Dan ⊂ Xan is a compact Hausdorff space. subset. It follows that the limit D∈D xD ∈ Xan x := lim exists. In fact, it is easy to see that for all f ∈ A we have f (x) = inf D∈Df (xD ). Proposition 5.2 Let D0 := ∩D∈DD. Exactly one of the following cases occurs. (1) The limit point x is a classical point, and D0 = {x}. (2) The set D0 is an affinoid disk, and x = xD0 . (3) The set D0 is a closed disk which is not affinoid, and x = xD0 . (4) The set D0 is empty. In Case (1), (3) and (4) we have (cid:93)H(x) = k . K )an in [3], §1.4.4. Proof: This follows from the classification of points on (A1 2 Lemma 5.3 Assume that we are in Case (1), (3) or (4) of Proposition 5.2. Then for any y ∈ φ−1 (x), the stabilizer Gy ⊂ G of y is solvable. Proof: Let OXan ,x denote the local ring of the point x on the analytic K - space Xan . By [4], Theorem 2.1.5, OXan ,x is a henselian local ring. Moreover, by n(cid:89) loc.cit., Lemma 2.1.6, we have a decomposition i=1 φ∗ (OYan )x = OYan ,yi , 28 where φ−1 (x) = {y1 , . . . , yn}. Therefore, the extension OYan ,y /OXan ,x is a Galois extension of henselian local rings with Galois group Gy . It follows that Gy sits in a short exact sequence 1 → Iy → Gy → Gal(κ(y)/κ(x)) → 1, where κ(x) and κ(y) are the residue fields of OXan ,x and OYan ,y , respectively, and Iy is the inertia group. By [4], Proposition 2.4.3 and Proposition 2.4.4, Iy is solvable. It remains to show that Gal(κ(y)/κ(x) is solvable. By [4], Theorem 2.3.3, κ(x) is a henselian valued field (called quasi-complete in loc.cit.) whose completion is the field H(x). By Proposition 5.2, the residue field of H(x) (equal to the residue field of κ(x)) is equal to k , which is algebraically closed by assumption. Using again [4], Proposition 2.4.4 we conclude that Gal(κ(y)/κ(x)) 2 and hence Gy is solvable. Lemma 5.4 (i) In Case (1) and (4) of Proposition 5.2, the affinoid disks Dan , D ∈ D, form a neighborhood basis of x in Xan . (ii) In Case (3), a neighborhood basis of x in Xan is given by the sets (D\D(cid:48) )an , where D ∈ D and D(cid:48) is an affinoid disk contained in D0 = ∩D∈DD. Proof: In Case (1), the statement in (i) is clear. Suppose we are in Case (4), i.e. ∩D∈DD = ∅. Any f ∈ A, seen as an analytic function on Xan , has finitely many zeroes, all of which are classical points. Therefore, there exists D1 ∈ D such that f Dan is an invertible analytic function on Dan , for all D ∈ D with D ⊂ D1 . Applying the maximum principle to f Dan and f −1 Dan we see that f (x) = f (xD ). The statement in (i) now follows from the definition of the topology of Xan . The proof of (ii) is similar. 2 5.5 We can now finish the proof of Theorem 3.9. We first suppose that we are in Case (1) or (4) of Proposition 5.2. Let φ−1 (x) = {y1 , . . . , yn} be the fiber above x. Then it follows from Lemma 5.4 and [4], proof of Theorem 2.1.5, that for all sufficiently small D ∈ D the inverse image φ−1 (D) decomposes into n disjoint affinoid neighborhoods of the points yi . But φ−1 (D) is connected by Lemma 5.1 (ii), hence n = 1. By Lemma 5.3 this shows that the group G is solvable, contradicting our assumption. We conclude that Case (1) and (4) of Proposition 5.2 cannot occur. We now suppose that we are in Case (2) of Proposition 5.2, i.e. D0 := ∩D∈DD is an affinoid disk. By Lemma 5.1 (i) we have g(V(D)) ∩ V(D) = ∅ for all D ∈ D and g ∈ G\Gξ . It follows easily that V(D0 ) = ∪D∈D V(D) and G(D0 ) = Gξ . Consider the Gξ -cover V(D0 ) → X\D0 . 29 If D ⊂ D0 is an affinoid disk, then D ∈ D if and only if U := X\D is a separating boundary component for this cover, see §4. Since Gξ is solvable by Lemma 5.1 (i), Proposition 4.1 shows that there exists a unique maximal separating boundary component. It follows that D0 ∈ D is the unique maximal element of D, i.e. Theorem 3.9 holds. It remains to rule out Case (3) of Proposition 5.2. Using Lemma 5.4 (ii) we find an element D ∈ D and an affinoid disk D(cid:48) ⊂ D such that x ∈ Dan\(D(cid:48) )an and such that V(D(cid:48) )an\V(D)an contains a unique point y ∈ φ−1 (x). By Lemma 5.3 this implies that the stabilizer of V(D(cid:48) )\V(D) in G is solvable. Together with Lemma 5.1 (ii) this shows that G(D(cid:48) ) is solvable. Applying Proposition 4.1 to the G(D(cid:48) )-cover V(D(cid:48) ) → X\D(cid:48) and using a similar argument as in Case (2) we conclude that D has a minimal element. But since this minimal element must be equal to the intersection D0 = ∩D∈DD which is not an affinoid disk in Case (3), we obtain a contradiction. Therefore, Case (3) of Proposition 5.2 cannot occur. This completes the proof 2 of Theorem 3.9. References [1] A. Abbes. R´eduction semi-stable des courbes d’apres Artin, Deligne, Grothendieck, Mumford, Saito, Winters,.. In Courbes semi-stables et groupes fondamental en g´eometrie alg´ebrique, pages 59–110. Birkhauser, 2000. [2] K. Arzdorf. Semistable reduction of cyclic covers of prime power degree. PhD thesis, Leibniz University Hannover, 2012. Spectral theory and analytic geometry over non- [3] V.G. Berkovich. archimedian fields. Number 33 in Mathematical Surveys and Monographs. AMS, 1990. [4] V.G. Berkovich. ´Etale cohomology of non-archimedian analytic spaces. Publ. Math. IHES, 78:5–161, 1993. [5] P. Berthelot. Cohomologie rigide et cohomologie rigid `a supports propres. Prepublication 96-03, Universit´e de Rennes 1. [6] S. Bosch, U. Guntzer, and R. Remmert. Non-Archimedian analysis: a sys- tematic approach to rigid analytic geometry. Number 261 in Grundlehren der math. Wiss. Springer-Verlag, 1984. [7] S. Bosch and W. Lutkebohmert. Stable reduction and uniformization of abelian varieties, I. Math. Ann., 270(3):349–379, 1985. [8] S. Bosch and W. Lutkebohmert. Formal and rigid gemetry: I. Rigid spaces. Math. Ann., 295:291–317, 1993. 30 [9] B. Conrad. Irreducible components of rigid spaces. Annales Inst. Fourier, 49(2):473–541, 1999. [10] A.J. deJong. Crystalline Dieudonn´e module theory via formal and rigid geometry. Publ. Math. IHES, 82:5–96, 1995. [11] P. Deligne and D. Mumford. The irreducibility of the space of curves of given genus. Publ. Math. IHES, 36:75–109, 1969. [12] H.P. Epp. Eliminating wild ramification. Inventiones math., 19:235–249, 1973. [13] J. Fresnel and M. van der Put. Rigid Analytic Geometry and its Applica- tions. Number 218 in Progress in Math. Birkhauser, 2004. [14] C. Lehr and M. Matignon. Wild monodromy and automorphisms of curves. Duke Math. J., 135(3):569–586, 2006. [15] Q. Liu. Algebraic geometry and arithmetic curves. Oxford Univ. Press, 2002. [16] M. Matignon. Vers un algorithme pour la r´eduction semistable des revetements p-cycliques de la droite pro jective sur un corps p-adique. Math. Ann., 325(2):323–354, 2003. [17] M. Raynaud. Enonc´es de permanence en g´eom´etrie relative. Journ´ees de G´eom´etrie Arithm´etiques de Rennes, July 2009. [18] M. Temkin. Stable modification of relative curves. J. Algebraic Geometry, 19:603–677, 2010. 31
1504.00147	5	1504	2016-02-08T02:01:29	Gonality and Clifford index of curves on elliptic K3 surfaces with Picard number two	[ "math.AG" ]	We compute the Clifford index of all curves on a K3 surface with Picard group isomorphic to U(m).	math.AG	math	GONALITY AND CLIFFORD INDEX OF CURVES ON ELLIPTIC K3 SURFACES WITH PICARD NUMBER TWO MARCO RAMPONI Abstract. We compute the Clifford index of all curves on K3 surfaces with Picard group isomorphic to U (m). 1. Introduction In the past years many authors have studied problems related to the gonality and Clifford index of curves lying on K3 surfaces. In this paper, by curve we always mean a smooth, reduced and irreducible curve over the field of complex numbers. The gonality and the Clifford index of a curve C are respectively defined by gon(C) = min{deg(A) : A ∈ Div(C), h0(A) = 2}, Cliff(C) = min{Cliff(A) : A ∈ Div(C), h0(A) ≥ 2, h1(A) ≥ 2} where Cliff(A) = deg A − 2h0(A) + 2. The Clifford index measures how special C is in the moduli space Mg of curves of genus g, in the following sense. One has 0 ≤ Cliff(C) ≤ ⌊ g−1 2 ⌋, where the second inequality is an equality for the generic member of Mg, and on the other hand Cliff(C) = 0 if and only if C is hyperelliptic (cf. [1]). Therefore, in this paper we say that a curve C is Clifford general if Cliff(C) = ⌊ g−1 2 ⌋. In all other cases, we say that C is Clifford special. A classical result by Saint-Donat [14, 5.8] states that given a hyperelliptic curve C on a K3 surface, all the curves in the linear system C are also hyperelliptic. This interesting fact was vastly generalized by Green and Lazarsfeld [9], who proved that indeed the Clifford index is the same for each smooth member of C. They also proved that, whenever C is Clifford special, there exists a divisor D on the ambient K3 surface X whose restriction to C computes its Clifford index1. In this case one says that the Clifford index is cut out on C by D. In [11] Knutsen showed that, if C is Clifford special, we can choose D to be a (smooth and irreducible) curve, and moreover Cliff(C) = D.(C − D) − 2. Originally, the constancy of the gonality of curves in a linear system on a K3 surface had been conjectured (unpublished) by Harris and Mumford, but Donagi and Morrison [7] found a counterexample. However, by the works of Ciliberto and Pareschi [5] and Knutsen [12] we know that this is indeed the only counterexample. On the other hand, the notions of gonality and Clifford index are very much related: for any curve C of Clifford index c one has c + 2 ≤ gon(C) ≤ c + 3, 1A divisor A ∈ Div(C) is said to compute the Clifford index of C if it appears in the definition of Cliff(C) and achieves the minimal value, i.e. Cliff(A) = Cliff(C). 1 2 MARCO RAMPONI and curves for which gon(C) = c + 3 are conjectured to be very rare (cf. [8]). When lying on K3 surfaces, these curves are completely classified by Knutsen [12]. In general, the gonality and the Clifford index are subtle invariants which are hard to compute explicitly for a given curve. In this note, we compute the Clifford index and gonality of all curves on some elliptic K3 surfaces. We prove the following. Theorem 1.1. Let X be a K3 surface with Picard group isomorphic to U (m), with m ∈ Z, m ≥ 1. Denote by E and F two generators of Pic(X), with E 2 = F 2 = 0 and E.F = m. Let C be a curve on X of genus g > 2. Then, either (i) The Clifford index of C is cut out on C by an elliptic curve EC , which is linearly equivalent to the one among E and F having minimal intersection with C. Then Cliff(C) = C.EC − 2 and C.EC is equal to the gonality of C; or (ii) m > 2 and C is linearly equivalent to E + F . Then C has maximal Clifford index Cliff(C) = ⌊m/2⌋. In the statement of the theorem U denotes the hyperbolic lattice: the lattice given by Z ⊕ Z with intersection matrix U = (cid:18)0 1 1 0(cid:19) . U (m) denotes the lattice obtained by U by multiplying the intersection matrix by a non-zero integer m. The term isomorphic means isomorphic as lattices. Notice that there always exists a class of square zero in Pic(X) ≃ U (m). There- fore a K3 surface X as in the theorem admits an elliptic fibration by [13, §3]. Part of the motivation for studying this problem came from a recent paper by Watanabe [16] in which he shows that for a K3 surface X which is a double cover of a smooth del Pezzo surface of degree 4 ≤ d ≤ 8 such that X carries a non-symplectic automorphism of order two which acts trivially on the Picard group, then for any curve C on X, either the Clifford index is cut out on C by some elliptic curve on X, or C is linearly equivalent to a multiple of a curve of genus 2. The key idea of Watanabe is that the automorphism yields some useful geometric informations which help to characterize the topological properties of the curves on X. In this work, we started to investigate the analogue situation when X carries a non-symplectic automorphism of order 3 which acts trivially on the Picard group. For ρ(X) = 2, we know by the classification results of Artebani-Sarti [2] and Taki [15], that the Picard group of X is isomorphic to either U or U (3). Therefore, our result applies to this case. However, not all K3 surfaces with Picard group U (m) admit non-symplectic automorphisms: see [2], and also Artebani-Sarti-Taki [3]. Notation and conventions. We work over the complex number field C. By surface we mean a smooth irreducible projective surface. A K3 surface is a regular surface with trivial canonical bundle. The Picard number of a surface X is by definition the rank of the Picard group Pic(X) and is denoted by the letter ρ. The symbol ∼ denotes linear equivalence between divisors and D is used to denote the complete linear system associated to a divisor D. A lattice is a free Z-module L of finite rank equipped with a non-degenerate symmetric integral bilinear form L × L → Z, (x, y) 7→ x.y. An isomorphism of lattices is a Z-module isomorphism preserving the bilinear forms. 3 2. Clifford special curves on a K3 surface Let X be a K3 surface. In this short section we recall some fundamental results which will be needed in the following and we also explain why the case ρ(X) = 1 is not interesting for our purposes. Fix a curve C of genus g on X. Let A(C) := {D ∈ Div(X) h0(OX (D)) ≥ 2, h0(OX (C − D)) ≥ 2}. Notice that C admits a decomposition C ∼ D + D′ into two moving classes D and D′ if and only if A(C) 6= ∅. When this happens, among such decompositions it is interesting to consider those with minimal intersection D.D′. Hence, one defines µC := min{D.(C − D) − 2 D ∈ A(C)}, and denotes by A0(C) the divisors in A(C) achieving this minimal value: A0(C) := {D ∈ A(C) D.(C − D) − 2 = µC} ⊂ A(C). Observe that µC ≥ 0 since the curve C (or any member of the complete linear system of a basepoint free and big line bundle on a K3 surface) is numerically 2-connected (cf. [14, (3.9.6)]). By the results in [9] and [11] we have: (cf. [10, p.11]) Cliff(C) = min{µC, ⌊ g−1 2 ⌋}. In other words, either C is Clifford general, or C is Clifford special and then A(C) is non-empty, the Clifford index of C is cut out by some divisor D ∈ A0(C) on X and Cliff(C) = µC . Then, by definition of Clifford index h0(OC (D)) ≥ 2 and h1(OC (D)) = h0(ωC ⊗ OC (−D)) ≥ 2. In particular, the linear systems on the curve C given by the line bundles OC (D) and ωC ⊗ OC (−D) = OC (C − D) contain some non-trivial effective divisors, hence of positive degree. Therefore we get degC (C − D) = C.(C − D) > 0 and also degC (D) = C.D > 0. Altogether, this yields the following inequalities: 0 < C.D < C 2. In particular, when Pic(X) = Z[H], we see that this inequalities are impossible when C ∈ H, and so C has general Clifford index in this case. On the other hand, for C ∼ kH with k ≥ 2, a direct computation shows that the Clifford index of C is cut out on C by a member of H. When ρ(X) ≥ 2, however, the situation is more interesting. In the next section we compute the Clifford index of any curve on a K3 surface with Picard group isomorphic to U (m). 3. Proof of the Theorem Let X be a K3 surface with Picard group isomorphic to U (m), with m ≥ 1. We let the Picard group of X be generated by the classes of two effective divisors E and F such that E 2 = F 2 = 0 and E.F = m. Up to the action of the Weyl group of X we may assume that E is an elliptic curve (cf. [13, §3]). When m = 1, we observe that the rational curve Γ ∼ F −E yields a section of the elliptic fibration given by E. Moreover, the linear system F = E + Γ contains a rational curve as a base component and therefore F cannot be represented by an irreducible curve (cf. Saint-Donat [14, 2.6 & 2.7]). On the other hand, when m > 1, since x2 ∈ 2mZ for x ∈ U (m) we observe that there are no rational curves on X. Thus any effective divisor is nef and basepoint 4 MARCO RAMPONI free (ibid.). computation shows that any elliptic curve on X belongs to either E or F . In particular we may assume that F is an elliptic curve. A simple For any effective divisor C on X let us define dC E 0(C) := min{E ′ · C E ′ is an elliptic curve on X}, := {elliptic curves EC such that EC · C = dC}. Lemma 3.1. Let C be a curve with C 2 > 0 and let EC be an elliptic curve in E 0(C). If (C − EC )2 = 0, then C belongs to the linear system E + F . Proof. If (C −EC )2 = 0 then C −EC is linearly equivalent to a multiple of an elliptic curve E ′, so that we can write C = EC + (C − EC ) ∼ EC + kE ′, some k ≥ 1. Since C 2 > 0 we see that E ′ is not linearly equivalent to EC . Since E ′.C = EC .C, we get k = 1 and C ∼ EC + E ′ ∼ E + F . (cid:3) Lemma 3.2. Let m ≥ 2 and let C be a curve in the linear system E + F . Then (i) If m = 2 then C is Clifford special. (ii) If m > 2 then C is Clifford general. Proof. Assume D ∈ A0(C) and let D ∼ aE + bF , with a, b ≥ 0. Then by definition of A(C) we may assume C − D effective, so that 0 ≤ (C − D).E = m(1 − b) and 0 ≤ (C − D).F = m(1 − a). Hence a, b ∈ {0, 1}. This shows that the only curves D in A0(C) are the members of E and F . Then C is Clifford special whenever µC = D.(C − D) − 2 = m − 2 < ⌊C 2/4⌋ = ⌊m/2⌋, that is for m ≤ 2. (cid:3) Remark 3.3. The case m = 1 is not to be considered here since the linear system E + F contains a rational curve Γ ∼ F − E as base component; hence there are no (irreducible) curves in E + F in this case. Lemma 3.4. Let C ⊂ X be an effective divisor with C 2 > 0. For any elliptic curve E ′ on X we have Moreover, C − E ′ is basepoint free for m > 1. (C − E ′)2 ≥ 0, h0(C − E ′) ≥ 2. Proof. Since E and F are the only effective reduced divisors with self-intersection zero, it is clear that in order to show the Lemma we may assume E ′ ∈ E, by the symmetry of the roles of E and F in Pic(X). Let C ∼ aE + bF for some positive integers a and b. Then clearly (C −E)2 ≥ 0 and also C.E > 0. Thus E.(C −E) > 0, which shows that C − E is effective. It follows h0(C − E) ≥ 2 by Riemann-Roch. Moreover, if m > 1, then C − E is basepoint free since in this case there are no rational curves on X (cf. [14, §2.7]). (cid:3) Remark 3.5. Let C be a curve on X. By the definition of A(C) and Lemma 3.4 above E 0(C) ⊂ A(C). In particular µC ≤ dC − 2. Moreover, E 0(C) ⊂ A0(C) ⇐⇒ µC = dC − 2. Indeed, let EC ∈ E 0(C) ⊂ A(C). If µC = dC − 2 then EC computes µC . Hence EC ∈ A0(C). The other implication is obvious. Proof of Theorem 1.1. The first (and longer) part of the proof is to show that Let EC ∈ E 0(C). By Lemma 3.4, EC ∈ A(C), so that A0(C) is not empty. E 0(C) ⊂ A0(C). 5 If A0(C) contains some elliptic curve F , then C.EC ≤ C.F and so C.EC −2 ≤ µC . Since EC ∈ A(C), we have C.EC − 2 = µC . Therefore EC ∈ A0(C). So we assume that A0(C) contains no elliptic curves at all. Let D be an effective divisor in A0(C). Since D ∈ A(C), and h1(D) = 0 by [10, Prop. 2.6], we have D2 ≥ 0. Let us show that, in fact, Indeed, assume by contradiction D2 = 0. D2 ≥ 2. • In the case where m ≥ 2, since X contains no rational curves, D is basepoint free, and so it is linearly equivalent to an elliptic curve, by [14, 2.6]. This contradicts the assumption that A0(C) contains no elliptic curves. • In the case where m = 1, let E and F be generators of the Picard group of X, with E 2 = F 2 = 0 and E.F = 1. Then we may assume that E is an elliptic curve and there exists a rational curve Γ on X such that Γ ∼ F − E. Since D2 = 0 and D ∈ A0(C), by [10, Prop. 2.6] we have D ∼ E or F . However, E is not in A0(C) by assumption, thus D ∼ F . Since Γ is the base locus of F , we have C.(F − E) = 0. On the other hand, µC = C · D − 2 = C · E − 2 and, since E ∈ A(C), this yields E ∈ A0(C). A contradiction. By the above discussion we always have D2 ≥ 2. Notice that C − D ∈ A0(C) and then (C − D)2 ≥ 2 by the same reason. We want to show that EC ∈ A0(C), contradicting the assumption that A0(C) contains no elliptic curves. Concretely, we need to show the following inequality EC · C ≤ D · (C − D). Rewrite this inequality as (3.1) For ED ∈ E 0(D) and ED′ ∈ E 0(D′) we let (D − EC ) · (D′ − EC ) ≥ 0, D′ := C − D nD = (D − ED).(D′ − ED′ ) rD = D.(ED′ − EC ) rD′ = D′.(ED − EC ) so that we may now rewrite (3.1) as follows: (3.2) nD + rD + rD′ ≥ ED · ED′ Claim. For any choice of elliptic curves ED ∈ E 0(D) and ED′ ∈ E 0(D′), nD = (D − ED).(D′ − ED′ ) ≥ 0. Indeed, by Lemma 3.4 the classes of (D − ED) and (D′ − ED′ ) have non-negative self-intersection and are effective, thus they lie in the closure of the positive cone and intersect non-negatively (cf. [4, IV.7]). This proves our claim. Now, consider the following inequalities: (3.3) rD ≥ rD + C.(EC − ED′ ) = D′.(EC − ED′ ) ≥ 0 rD′ ≥ rD′ + C.(EC − ED) = D.(EC − ED) ≥ 0 If we assume that either rD > 0 or rD′ > 0 then (3.2) holds, since nD ≥ 0 and EC · ED′ ≤ m and rD ≥ m or rD′ ≥ m. 6 MARCO RAMPONI (recall that x.y ∈ mZ for x, y ∈ U (m)). Hence, we assume rD = 0 and rD′ = 0. Substituting this in (3.3) we get D.EC = D.ED and D′.EC = D′ED′ . Thus, EC ∈ E 0(D) ∩ E 0(D′). Using the claim above, we can replace both ED and ED′ by EC in the definition of nD and this yields the desired inequality (3.1) and therefore E 0(C) ⊂ A0(C). Now that we know E 0(C) ⊂ A0(C), we determine all Clifford general curves. Take EC ∈ E 0(C). By Lemma 3.4 we know C − EC ∈ A(C). Moreover, we also have C − EC ∈ A0(C) since EC ∈ A0(C) by assumption. We distinguish two cases: • (C − EC )2 = 0. Then, by Lemma 3.1 and Lemma 3.2, C is Clifford general if and only if m > 2 and C ∈ E + F . • (C − EC )2 > 0. (in particular C is not linearly equivalent to E + F ). We then show that C is Clifford special. This amounts to show 2µC ≤ g − 3 which, by the definition of µC and the genus formula, is equivalent to (C − 2EC)2 ≥ −4. We may write C ∼ aEC + D, with a ≥ 1, D effective and D2 = 0. If a = 1 we get C ∈ E + F by Lemma 3.1, which is not the case. So a ≥ 2 and Therefore, C is Clifford special. (C − 2EC)2 ≥ 0. This proves that C is Clifford general if and only if m > 2 and C ∈ E + F , as in part (ii) of the Theorem. To show part (i), we can therefore assume that C is Clifford special. Then Cliff(C) = µC and since E 0(C) ⊂ A0(C) we have µC = dC − 2. Therefore, the Clifford index of C is cut out by some elliptic curve EC ∈ E 0(C). In particular, Cliff(C) is computed by a pencil: the restriction of EC to C. Therefore (cf. [8, p.174]) gon(C) = Cliff(C) + 2 = dC . Hence, the assertions of (i) follow and the Theorem is proved. (cid:3) Remark 3.6. In particular, we observe that when m = 1 or 2, any curve on X is Clifford special and its Clifford index is cut out by an elliptic curve. The same conclusion when m = 2 is implicitly contained in [16]. Acknowledgement. I am thankful to Alessandra Sarti for her guidance and sup- port. A special thank also to Flaminio Flamini and Andreas Knutsen for their warm welcome in Rome and useful conversations. I am grateful to Kenta Watanabe for a careful review and for pointing out some mistakes in a draft version of this paper. References [1] E. Arbarello, M. Cornalba, P. Griffiths, and J.D. Harris. Geometry of Algebraic Curves. Number v. 1 in Grundlehren der mathematischen Wissenschaften. Springer New York, 2010. [2] M. Artebani and A. Sarti. Non-symplectic automorphisms of order 3 on K3 surfaces. Mathe- matische Annalen, 342(4):903 -- 921, 2008. [3] M. Artebani, A. Sarti, and S. Taki. K3 surfaces with non-symplectic automorphisms of prime order. Mathematische Zeitschrift, 268(1-2):507 -- 533, 2011. [4] W. Barth, K. Hulek, C. Peters, and A. van de Ven. Compact Complex Surfaces. Ergebnisse der Mathematik und ihrer Grenzgebiete. 3. Folge / A Series of Modern Surveys in Mathematics. Springer Berlin Heidelberg, 2014. 7 [5] Ciro Ciliberto and Giuseppe Pareschi. Pencils of minimal degree on curves on a K3 surface. J. Reine Angew. Math., 460:15 -- 36, 1995. [6] I.V. Dolgachev and S. Kondo. Moduli of K3 Surfaces and Complex Ball Quotients. In Rolf- Peter Holzapfel, A.Muhammed Uluda, and Masaaki Yoshida, editors, Arithmetic and Ge- ometry Around Hypergeometric Functions, volume 260 of Progress in Mathematics, pages 43 -- 100. Birkhuser Basel, 2007. [7] Ron Donagi and David R. Morrison. Linear systems on K3-sections. J. Differential Geom., 29(1):49 -- 64, 1989. [8] D. Eisenbud, H. Lange, G. Martens, and F. Schreyer. The Clifford dimension of a projective curve. Compositio Mathematica, 72(2):173 -- 204, 1989. [9] M. Green and R. Lazarsfeld. Special divisors on curves on a K3 surface. Inventiones Mathe- maticae, 89(2):357 -- 370, 1987. [10] T. Johnsen and A.L. Knutsen. K3 Projective Models in Scrolls. Number no. 1842 [11] A.L. Knutsen. On kth-order embeddings of K3 surfaces and Enriques surfaces. Manuscripta Mathematica, 104(2):211 -- 237, 2001. [12] A.L. Knutsen. On two conjectures for curves on K3 surfaces. Internat. J. Math., 20(12):1547 -- 1560, 2009. [13] I.I. Pyatetskij-Shapiro and I.R. Shafarevich. A Torelli theorem for algebraic surfaces of type K3. Izv. Akad. Nauk SSSR, Ser. Mat., 35:530 -- 572, 1971. [14] B. Saint-Donat. Projective models of K3 surfaces. American Journal of Mathematics, 96(4):pp. 602 -- 639, 1974. [15] S. Taki. Classification of non-symplectic automorphisms of order 3 on K3 surfaces. Mathe- matische Nachrichten, 284(1):124 -- 135, 2011. [16] K. Watanabe. The Clifford index of line bundles on a 2-elementary K3 surface given by a double cover of a del Pezzo surface. Geometriae Dedicata, pages 1 -- 15, 2014. Marco Ramponi, Universit´e de Poitiers, Poitiers, France E-mail address: [email protected]
1812.00256	1	1812	2018-12-01T20:11:17	A Riemann-Hilbert correspondence for Cartier crystals	[ "math.AG" ]	For a variety $X$ separated over a perfect field of characteristic $p>0$ which admits an embedding into a smooth variety, we establish an anti-equivalence between the bounded derived categories of Cartier crystals on $X$ and constructible $\mathbb Z/p \mathbb Z$-sheaves on the \'etale site $X_{\text{\'et}}$. The key intermediate step is to extend the category of locally finitely generated unit $\mathcal O_{F,X}$-modules for smooth schemes introduced by Emerton and Kisin to embeddable schemes. On the one hand, this category is equivalent to Cartier crystals. On the other hand, by using Emerton-Kisin's Riemann-Hilbert correspondence, we show that it is equivalent to Gabber's category of perverse sheaves in $D_c^b(X_{\text{\'et}},\mathbb Z/p \mathbb Z)$.	math.AG	math	A Riemann-Hilbert correspondence for Cartier crystals Tobias Schedlmeier Abstract For a variety X separated over a perfect field of characteristic p > 0 which admits an embedding into a smooth variety, we establish an anti-equivalence between the bounded derived categories of Cartier crystals on X and constructible Z/pZ-sheaves on the étale site Xét. The key intermediate step is to extend the category of locally finitely generated unit OF,X -modules for smooth schemes introduced by Emerton and Kisin to embeddable schemes. On the one hand, this category is equivalent to Cartier crystals. On the other hand, by using Emerton-Kisin's Riemann-Hilbert cor- respondence, we show that it is equivalent to Gabber's category of perverse sheaves in Db c(Xét, Z/pZ). Introduction The Riemann-Hilbert correspondence gave a general answer to Hilbert's 21st problem from a modern point of view by connecting D-modules to constructible CX-sheaves on a smooth complex variety X. For a variety X over a field of positive characteristic, Emerton and Kisin established an analogue to this correspondence in [EK04]. However, the smoothness of X is essential for this approach because it ensures that the objects Emerton and Kisin consider instead of D-modules behave nicely. In this paper we relax the smoothness assumption. We establish a Riemann-Hilbert type correspondence for a variety X over a perfect field of positive characteristic which is possibly singular but we require that X admits an embedding into a smooth vari- ety Y . Thereby we suggest the category of so-called Cartier crystals as a replacement for D-modules because Cartier crystals seem to be more suitable for generalizations of the correspondence to singular varieties or even more general schemes. In particular, Cartier crystals are indeed closely related to perverse constructible étale sheaves on X, as conjectured by Blickle and Böckle in [BB11]. Let us take a brief look at the development of the Riemann-Hilbert correspondence throughout history. A fundamental step to the modern version of the Riemann-Hilbert correspondence was the result by Deligne in 1970 ([Del70]), which states that for a smooth variety X over the complex numbers, there is an equivalence Connreg −→ Loc(X an) 1 between the categories of regular integrable connections on X and local systems on X an, i.e. CX-modules which are locally free of finite rank for the analytic topology. Here an integrable connection is a DX-module M which is locally free of finite rank as an OX -module. This is nothing but a locally free OX -module of finite rank together with a C-linear map ∇ : M −→ Ω1 X ⊗OX M . For the above equivalence, we have to pass to a certain subcategory, namely the regular integrable connections, and the underlying functor of the equivalence is given by taking the kernel of ∇. Note that if dX denotes the dimension of X, this is the cohomology in degree −dX of the de Rham complex 0 −→ Ω1 X ⊗OX M −→ 0 X ⊗OX M −→ Ω2 X ⊗OX M −→ · · · −→ ΩdX located between the degrees −dX and 0 and whose differentials are induced by ∇. Let DRX(M ) denote this complex. Both categories Connreg and Loc(X an) are not closed under push-forwards. For in- stance, the push-forward of a local system on the origin to the affine line is obviously not a local system. The correct extensions are (regular holonomic) DX -modules on the left and constructible CX-sheaves on the right. However, the functor H −dX (DRX ( )) does not yield an equivalence between these larger categories. Again considering the C, we see that H −1(DRA1(i∗C)) = 0. This is due example of the inclusion i : {0} −→ A1 to the fact that we lose to much information by only taking into account the −dX- th cohomology of DRX( ). To avoid this problem, one considers the derived functor L ⊗OX ) = ΩX between the derived categories of DX -modules and constructible DRX( In the context of complex manifolds, Kashiwara ([Kas80] and [Kas84]) CX-sheaves. passed to a suitable subcategory called regular holonomic D-modules -- more precisely the full subcategory of the bounded derived category Db(DX ) consisting of complexes whose cohomology sheaves are regular holonomic -- and proved that the de Rham functor is an equivalence Db rh(DX ) −→ Db c(CX), which is compatible with the six operations f∗, f!, f ∗, f !, RHom• and c(CX) denotes the full subcategory of D(CX ) of bounded complexes with constructible co- homology sheaves. This result from 1980 and 1984 is known as the Riemann-Hilbert correspondence. Around the same time, Mebkhout ([Meb84b] and [Meb84a]) gave a proof, which is independent of Kashiwara's work. Later on, Beilinson and Bernstein developed the Riemann-Hilbert correspondence for algebraic D-modules on complex al- gebraic varieties. Their work is explained in the unpublished notes ([Ber]). ⊗. Here Db L Deligne's result, which is a special case of the Riemann-Hilbert correspondence, ap- plied to X = P1(C)\S, the Riemann sphere without a finite set S of points, gives an answer to Hilbert's 21st problem. For this recall that the sheaf of solutions of a system of linear differential equations is a local system. Via analytically continuing of local solutions along closed paths in X, we obtain a transition matrix and therefore a rep- resentation of the fundamental group of the Riemann sphere without S. The group of such matrices is called the monodromy group of the system of differential equations. Conversely, Hilbert's 21st problem asks for the existence of a system of linear differen- 2 tial equations on the Riemann sphere with Fuchsian singularities in S and with a given monodromy. The functor DRX( ) is closely related to the so-called solution functor SolX = RHom• DX ( , OX ): for every bounded complex M • of DX -modules, we have DRX(M •) ∼= SolX (DXM •)[dX ], where DX is a certain duality. For a coherent DX -module M , the sheaf HomDX (M, OX ) can be identified with the solutions of the system of differential equations corresponding to M . Furthermore, there is an equivalence between representations of the fundamental group of X and locally constant CX-sheaves. The Riemann-Hilbert correspondence in turn is a far reaching generalization of Deligne's result. Of course the essential image of the abelian category of regular holonomic DX -modules under the equivalence DRX is an abelian category inside Db c(CX), but it turns out that this category differs from the category of constructible CX-sheaves. The example of the immersion of the origin into the affine line from above already is a first sign of this phenomenon. The abelian subcategory of Db c(CX ) given by the essential image of regular holonomic DX -modules under the de Rham functor is called perverse sheaves. There is a general tool for describing abelian subcategories of triangulated categories: the theory of t-structures. A t-structure on a triangulated category D consists of two subcategories D≤0 and D≥0 with certain properties. The intersection D≤0 ∩ D≥0 is called the heart of the t-structure. It is an abelian category. For example, the so- called canonical t-structure of Db rh (DX ) and D≥0 rh (DX ) of complexes whose cohomology is zero in positive or negative degrees. In the same way, the category of perverse sheaves on X is obtained as the heart of a t-structure on Db Indeed, the development of the theory of perverse sheaves by Beilinson, Bernstein, Deligne and Gabber was motivated by the Riemann-Hilbert correspondence. A standard reference for this is [BBD82]. rh(DX ) is given by the two subcategories D≤0 c(CX ) which is called the perverse t-structure. At the beginning of the 21st century, the time was right for a positive characteristic version of the Riemann-Hilbert correspondence. The de Rham theory for varieties over a field of positive characteristic p differs strongly from the one on complex varieties. Instead of the Poincaré lemma, we have the Cartier isomorphism and as a consequence, for a smooth variety X, the kernel of the map OX −→ Ω1 X is not a locally constant Z/pZ-sheaf but given by the p-th powers (OX )p. Therefore, one has to find a different approach. The Frobenius endomorphism F is a major tool in characteristic p. Especially sheaves with an action of the Frobenius turned out to be very useful. The starting point of these objects is the sheaf OF,X = OX [F ] of non-commutative rings given on an affine open subset U ⊆ X by the polynomial ring OX (U )[F ] with the relation F r = rpF for local sections r ∈ OX (U ). A simple calculation shows that left OX [F ]-modules are identified with OX -modules F together with a morphism F ∗F −→ F. In [Kat73, Proposition 4.1.1], Katz proved that there is an equivalence between the category of locally free étale Fp-sheaves and the category of coherent, locally free OX -modules E together with an isomorphism F ∗ E −→ E of OX -modules. This may be considered as an analogue of Deligne's result that there is a natural equivalence Connreg −→ Loc(X an). 3 It is this result of Katz that motivated Emerton and Kisin to consider left OF,X- modules for establishing an analogue of the Riemann-Hilbert correspondence for smooth varieties over a field k of positive characteristic p. As Katz' work already suggested, certain unit left OF,X-modules, i.e. OF,X-modules F whose structural morphism F ∗F −→ F is an isomorphism together with some finiteness condition, is the subcategory to look at. In 2004, Emerton and Kisin published [EK04], where they proved that the functor Sol = RHom• ét, OXét)[dX ] yields an anti-equivalence ( OF,Xét Db lfgu(OF,X) −→ Db c(Xét, Z/pZ) L between the bounded derived categories of locally finitely generated unit (lfgu for short) left OF,X-modules on the one hand, and the bounded derived category of constructible Z/pZ-sheaves on the étale site Xét of X on the other hand. Their correspondence is shown to be compatible with half of the six cohomological operations, namely f !, f+ ⊗. They also prove that under the correspondence the abelian category µlfgu(X) of and locally finitely generated unit modules corresponds to the category of perverse sheaves Perv(Xét, Z/pZ) defined by Gabber in [Gab04] on Db In this Riemann- Hilbert type correspondence, the sheaf of partial differential operators is substituted by the sheaf OF,X. Every OF,X-module naturally has the structure of a DX -module. The crucial point is that the ring DX of arithmetic differential operators introduced by Berthelot equals the union S EndOX pe (OX ) ([Ber96], [Ber00]). The details of the DX- module structure of an OF,X-module are explained in [Bli03]. It follows that the category considered by Emerton and Kisin is a subcategory of the category of left modules over the sheaf of rings of differential operators. c(Xét, Z/pZ). The sequence 0 −→ OXét 1−F−−−→ OXét −→ 0 in some sense plays the role of the de Rham complex for varieties over C. For instance, we can compute Sol(OX ) = RHom• (OXét, OXét )[dX ] using the resolution OF,Xét 0 −→ OF,Xét 1−F−−−→ OF,Xét of OXét by free left OF,Xét-modules. As a consequence of Artin-Schreier theory, the sequence 0 −→ (Z/pZ)X −→ OXét 1−F−−−→ OXét −→ 0 is exact and therefore Solκ(OX ) ∼= (Z/pZ)X [dX ]. This observation is fundamental in the proof of Emerton and Kisin's Riemann-Hilbert correspondence. In [BB11], Blickle and Böckle show that if X is smooth and F -finite (i.e. the Frobenius morphism is a finite map), then Emerton-Kisin's category µlfgu(X) is equivalent to their category Crysκ(X) of Cartier crystals on X. This category is obtained by localizing the category of coherent sheaves M on X equipped with a right action by Frobenius, i.e. a map F∗M −→ M , at the Serre subcategory consisting of those M where the structural map is nilpotent. 4 The category of Cartier crystals is also defined on singular schemes, and a Kashiwara type equivalence holds in this context [BB13, Theorem 4.1.2], showing that Cartier crystals on a closed subscheme Z ⊆ X are "the same" as Cartier crystals on X supported in Z. This suggests that for singular schemes, the category of Cartier crystals should be a reasonable replacement for Emerton-Kisin's theory, which was only developed for X smooth. Hence one expects a natural equivalence of categories Crysκ(X) −→ Perv(Xét, Z/pZ) for any F -finite scheme X. In this paper we show this result under the assumption that X is a variety over a perfect field k, embeddable into a smooth variety. Note that a variety over a perfect field is F -finite. The closed immersion of X into a smooth variety Y enables us to employ the Kashiwara equivalence to show that the category of Cartier crystals on X is equivalent to the category of lfgu modules on Y supported in X. This equivalence on the level of abelian categories then extends to a derived equivalence Db crys(QCrysκ(X)) ∼= Db lfgu(OF,Y )X, lfgu(OF,Y )X denotes the full subcategory of Db where Db lfgu(OF,Y ) consisting of complexes whose cohomology sheaves are supported in X. The details of this equivalence are worked out in Section 2 and involve showing that the equivalence sketched by Blickle and Böckle between Cartier crystals and µlfgu(X) alluded to above is compatible with pull-back functors for immersions of smooth, F -finite schemes and push-forward functors for arbitrary morphisms between smooth, F -finite schemes. lfgu(OF,X) := Db In Section 3 we give an intrinsic proof of the fact that for a variety X over a perfect field k the category Db lfgu(OF,Y )X is well-defined, i.e. independent of If one had resolution of singularities the embedding of X into a smooth scheme Y . ∼= f! Sol for in characteristic p, one would have natural isomorphisms of functors Sol f+ every morphism f between smooth k-schemes [EK04, Theorem 9.7.1]. This would enable us to work with derived categories of constructible étale sheaves, which are defined on singular schemes as well, turning the independence of a chosen embedding into an easy exercise. As resolution of singularities is an open problem in higher dimensions, we are required to extend the adjunction between the functors f ! and f+ for proper f from Emerton-Kisin to the case that f is proper over some closed subset, which is somewhat technical. The source of this is a general adjunction statement for quasi-coherent sheaves provided in [Sch18]. It says that for a separated morphism of finite type f : X −→ Y of Noetherian schemes, closed immersions i : Z −→ Y and i′ : Z ′ −→ X and a proper morphism f ′ : Z ′ −→ Z making the diagram i′ i Z ′ f ′ Z X f / Y commutative, there exists a morphism trf : Rf∗RΓZ ′f ! −→ id which acts as the counit of an adjunction between Rf∗ and RΓZ ′f ! regarded as functors between certain derived 5 / / / categories with cohomology sheaves supported on Z and Z ′. Here RΓZ ′ is the local cohomology functor. Combining these steps, the following theorem summarizes the main results in this paper: Theorem. Let X be a variety over a perfect field and assume that X is embeddable into a smooth variety Y . Then there are natural equivalences of categories Db crys(QCrysκ(X)) ∼−→ Db lfgu(OF,Y )X ∼−→ Db c(Xét, Z/pZ). Here the middle category is independent of the embedding. These equivalences are com- patible with the respectively defined push-forward and pull-back functors for immersions. Furthermore, the standard t-structure on the left corresponds to Gabber's perverse t- structure on the right. Corollary. The abelian category Crysκ(X) of Cartier crystals on a variety X embeddable into a smooth variety is naturally equivalent to the category Perv(Xét, Z/pZ) of perverse constructible étale p-torsion sheaves. While in the final stages of writing up these results, the preprint [Ohk16] appeared. Therein the author shows that Emerton-Kisin's Riemann-Hilbert correspondence can be extended to the case that X is embeddable into a proper smooth Wn-scheme. The case n = 1 hence also implies the right half of the just stated theorem in the case that X is embeddable into a proper smooth scheme. Acknowledgments I cordially thank the supervisor of my thesis, Manuel Blickle, for his excellent guidance and various inspiring conversations. Moreover, I thank Gebhard Böckle for many use- ful comments, Axel Stäbler for advancing discussions and Sachio Ohkawa for a careful reading and commenting on an earlier draft of parts of this thesis. The author was partially supported by SFB / Transregio 45 Bonn-Essen-Mainz financed by Deutsche Forschungsgemeinschaft. Notation and conventions Unless otherwise stated, all schemes are locally Noetherian and separated over the field Fp for some fixed prime number p > 0. For such a scheme X, we let FX or F , if no ambiguity is possible, denote the Frobenius endomorphism X −→ X which is the identity on the underlying topological space and which is given by r 7→ rp on local sections. Often we will deal with F -finite schemes, i.e. F is a finite morphism. For instance, a variety over a perfect field is F -finite. Working with Emerton and Kisin's category of locally finitely generated unit modules forces us at some points to restrict to varieties, i.e. to schemes which are of finite type over a field k containing Fp. With "schemes over k" or "k-scheme" we always mean 6 schemes which are separated and of finite type over k. For a smooth scheme X over a perfect field k, the sheaf of top differential forms ωX is an invertible sheaf with a canonical morphism ωX −→ F !ωX of OX-modules given by the Cartier operator, see Example 1.2 for the affine space. One can check that it is an isomorphism. In general, if X is regular and F -finite, we will assume that there is a dualizing sheaf ωX with an isomorphism κX : ωX −→ F !ωX. For example, this assumption holds if X is a scheme over a local Gorenstein scheme S = Spec R ([BB11, Proposition 2.20]). Moreover, we assume that ωX is invertible. As in [EK04], for a smooth k-scheme X, we let dX denote the function x 7→ dimension of the component of X containing x. If f : X −→ Y is a morphism of smooth k-schemes, the relative dimension dX/Y is given by dX/Y = dX − dY ◦ f . 1 Review of Cartier crystals and locally finitely generated unit modules We begin by reviewing the definitions and results from the theory of Cartier crystals as developed by Blickle and Böckle in [BB11] and [BB13]. In short, a coherent Cartier module M on X is a coherent OX -module together with a right action of the Frobenius F . These form an abelian category and the category of Cartier crystals is obtained by localizing at the full Serre subcategory of those M on which F acts nilpotently. The resulting localized category is an abelian category, which has been shown in [BB11] to enjoy strong finiteness properties: All objects have finite length and all endomorphism sets are finite dimensional Fp-vector spaces. 1.1 Cartier modules and Cartier crystals Definition 1.1. A Cartier module on X is a quasi-coherent OX -module M together with a morphism of OX -modules κ : F∗M −→ M. Equivalently, a Cartier module M is a sheaf of right OF,X-modules whose underlying sheaf of OX -modules is quasi-coherent. Here OF,X is the sheaf of (non-commutative) rings OX [F ], defined affine locally on Spec R as the ring R[F ] := R{F }/hF r − rpF r ∈ Ri. On the level of abelian sheaves, M and F∗M are equal, hence we may view the structural map κ of a Cartier module M as an additive map κ : M −→ M which satisfies κ(rp · m) = rκ(m) for all local sections r ∈ OX and m ∈ M . In this way it is clear that defining the right action of F on M via κ defines a right action of OX [F ] on M , and vice versa. 7 Iterations of κ are defined inductively: κn := κ ◦ F∗κn−1. Considering κ as an additive map of abelian sheaves, κn is the usual n-th iteration. ∗ := f HomOY (f∗OX , For a finite morphism f : X −→ Y of schemes, the functor f∗ is left adjoint to the functor f ♭ ), where f is the flat morphism of ringed spaces (X, OX ) −→ (Y, f∗OX ), see [Har66, III. 6]. Hence the structural morphism of a Cartier module M on an F -finite scheme may also be given in the form κ : M −→ F ♭M . Example 1.2. The prototypical example of a Cartier module is the sheaf ωX of top differential forms on a smooth variety over a perfect field k. If X = Spec k[x1, . . . , xn], then ωX is the free k[x1, . . . , xn]-module of rank 1 generated by dx1 ∧ · · · ∧ dxn. This module has a natural homomorphism κ : F∗ωX −→ ωX called the Cartier operator given by the formula xi1 1 · · · · · xin n dx1 ∧ · · · ∧ dxn 7→ x (i1+1) p −1 1 (in+1) p −1 · · · · · x n dx1 ∧ · · · ∧ dxn where a non-integral exponent anywhere renders the whole expression zero. A morphism ϕ : M −→ N of Cartier modules is a morphism of the underlying quasi- coherent sheaves making the following diagram commutative: F∗M F∗ϕ F∗N κM κN M ϕ / N. As F∗ is exact, one immediately verifies that Cartier modules form an abelian category, the kernels and cokernels being just the underlying kernels and cokernels in OX -modules with the induced structural morphism. We denote the category of Cartier modules on X by QCohκ(X). The full subcategory of coherent Cartier modules Cohκ(X) consists of those Cartier modules whose underlying OX -module is coherent. A Cartier module (M, κ) is called nilpotent if some power of κ is zero; (M, κ) is called locally nilpotent if it is the union of its nilpotent Cartier submodules. By LNilκ(X) we denote the full subcategory of QCohκ(X) consisting of locally nilpotent Cartier modules, and Nilκ(X) denotes the intersection Cohκ(X) ∩ LNilκ(X). The full subcategory of QCohκ(X) con- sisting of extensions of coherent and locally nilpotent Cartier modules (in either order) we denote by LNilCohκ(X). One has the following inclusions Nilκ(X) 2❡❡❡ ,❨❨❨ LNilκ(X) Cohκ(X) -❩❩❩ 1❞❞❞ LNilCohκ(X) / QCohκ(X), and each of the full subcategories are Serre subcategories in their ambient category1. This leads us to our key construction. 1A Serre subcategory is a full abelian subcategory which is closed under extensions. 8 / / / { - $ 2 z , / # 1 Definition 1.3. The category of Cartier quasi-crystals is the localization of the category of quasi-coherent Cartier modules QCohκ(X) at its Serre subcategory LNilκ(X). It is an abelian category, which we denote by QCrysκ(X). Similarly, the category of Cartier crystals on X is the localization of the category Cohκ(X) of coherent Cartier modules at its Serre subcategory Nilκ(X). It is an abelian category, which we denote by Crysκ(X). Cartier crystals also can be obtained by local- izing LNilCohκ(X) at the subcategory LNilκ(X). In order to define derived functors we have to make sure that the considered categories have enough injectives. Proposition 1.4. The category QCohκ(X) is a Grothendieck category with enough in- jectives whose underlying OX-module is injective. Its Serre subcategory LNilκ(X) is localizing and hence QCrysκ(X) has enough injectives. Proof. The first statements were shown in [BB13, Theorem 2.0.9 and Proposition 3.3.17]. That LNilκ(X) is localizing now follows from Corollaire 1 on p. 375 of [Gab62] and from the fact that each M ∈ QCohκ(X) has a maximal locally nilpotent κ-subsheaf Mnil, see [BB13, Lemma 2.1.3]. Then Corollaire 2 of [Gab62] shows that the associated quotient category QCrysκ(X) has enough injectives. Concretely, if T : QCohκ(X) −→ QCrysκ(X) denotes the exact localization func- tor, then the fact that LNilκ(X) is localizing asserts the existence of a right adjoint V : QCrysκ(X) −→ QCohκ(X). If M/Mnil ֒−→ I is an injective hull in QCohκ(X), then it is shown in op. cit. that T I is an injective hull of T (M/Mnil). The following finiteness statements are the main results of [BB11]: Theorem 1.5 ([BB11, Corollary 4.7, Theorem 4.17]). Let X be a locally Noetherian, F -finite scheme of positive characteristic p. (a) Every object in the category of Cartier crystals Crysκ(X) satisfies the ascending and descending chain condition on its subobjects. (b) The Hom-sets in Crysκ(X) are finite dimensional Fp-vector spaces. These finiteness properties are precisely the ones one expects from a category of per- verse constructible sheaves in the topological context. It is this result (and the related statement [EK04, Theorem 11.5.4] in the smooth case) that prompted our investigation of a connection between Cartier crystals and Gabber's category of perverse constructible Z/pZ-sheaves on Xét, which is the content of this article. 1.2 Cartier crystals and morphisms of schemes Up to now, we studied different categories stemming from quasi-coherent sheaves on a In this subsection, we consider morphisms f : X −→ Y of schemes single scheme X. and construct functors between the categories of Cartier (quasi-)crystals on X and on Y . With the notation D∗(A) for an abelian category A and ∗ ∈ {+, −, b}, we mean the 9 subcategory of the derived category D(A) of bounded below, bounded above or bounded complexes. The first result is concerned with the derived functor Rf∗ for quasi-coherent OX - modules. For a large class of morphisms it behaves well with the additional structure of Cartier modules and with localization at nilpotent objects. In principle, to any quasi- coherent Cartier module M with structural map κM , we assign the quasi-coherent OY - module f∗M together with the composition FY ∗f∗M ∼−→ f∗FX∗M f∗κM−−−→ f∗M. This is the underived functor f∗ : QCohκ(X) −→ QCohκ(Y ). Theorem 1.6 ([BB13, Corollary 3.2.12]). Let f : X −→ Y be a morphism of F -finite schemes. Suppose ∗ ∈ {+, −, b}. The functor Rf∗ on quasi-coherent sheaves induces a functor Rf∗ : D∗(QCohκ(X)) −→ D∗(QCohκ(Y )). It preserves local nilpotence and hence induces a functor Rf∗ : D∗(QCrysκ(X)) −→ D∗(QCrysκ(Y )). If f is of finite type (but not necessarily proper!) then it restricts to a functor Rf∗ : D∗ crys(QCrysκ(X)) −→ D∗ crys(QCrysκ(Y )) where the subscript crys indicates that the cohomology lies in LNilCrysκ. For essentially étale morphisms and for closed immersions there are pull-back functors. Theorem 1.7. Let f : X −→ Y be a morphism of schemes. (a) Suppose ∗ ∈ {+, −, b}. If f is essentially étale, the exact functor f ∗ induces a functor f ! : D∗ crys(QCrysκ(Y )) −→ D∗ crys(QCrysκ(X)), which is left adjoint to Rf∗. (b) Suppose ∗ ∈ {+, b}. If f is a closed immersion of F -finite schemes, the functor f ♭ = ), where f denotes the flat morphism (X, OX ) −→ (Y, f∗OX ) of ∗ HomOY (f∗OX , f ringed spaces, induces a functor f ! : D∗ crys(QCrysκ(Y )) −→ D∗ crys(QCrysκ(X)), which is right adjoint to Rf∗. Proof. Let M be a quasi-coherent Cartier module on Y . For essentially étale f , there is a canonical isomorphism bc : FX∗f ∗ ∼−→ f ∗FY ∗. Hence we may equip f ∗M with the structural morphism given by the composition FX∗f ∗M bc−→ f ∗FY ∗M f ∗κM−−−→ f ∗M. 10 As f ∗ preserves coherence, we obtain a functor Cohκ(Y ) −→ Cohκ(X). It is easy to see that f ∗ preserves nilpotency. Therefore, and by exactness of f ∗, we obtain the desired functor Db crys(QCrysκ(Y )) −→ Db crys(QCrysκ(Y )). In the case of a closed immersion f , the composition f ♭M f ♭ κ −−→ f ♭F ♭ Y M ∼−→ F ♭ X f ♭M, where κ is the adjoint of κ, is a natural Cartier structure for the OX -module f ♭M . Once again it remains to check that it gives rise to a functor f !Db crys(QCrysκ(Y )) −→ crys(QCrysκ(X)). The adjunctions of Rf∗ and f ∗ or f ! follow from the corresponding Db adjunctions for quasi-coherent sheaves. For more details see [BB13, Proposition 3.3.19] and [BB13, Corollary 3.3.24]. Now let i : Z −→ X be a closed immersion and j : U −→ X the open immersion of the complement X\Z. Note that for a closed immersion i, the functor i∗ is exact and therefore we drop the R indicating derived functors. The units and counits of the adjunctions between i∗ and i! and between Rj∗ and j∗ lead to a familiar distinguished triangle. Theorem 1.8 ([BB13, Theorem 4.1.1]). In D+ triangle crys(QCrysκ(X)) there is a distinguished i∗i! −→ id −→ Rj∗j∗ −→ i∗i![1]. This theorem shows the equivalence mentioned in the following definition. crys(QCrysκ(X)) is supported on Z if j!M• = 0 Definition 1.9. A complex M• of Db or, equivalently, if the natural morphism i∗i!M• −→ M• is an isomorphism. We let Db crys(QCrysκ(X))Z denote the full triangulated subcategory consisting of complexes supported in Z. For Cartier crystals, there is a natural isomorphism of functors i∗i! ∼= RΓZ where RΓZ is the local cohomology functor, see [BB11, Proposition 2.5] for the basic result concerning the abelian categories of Cartier modules and the proof of [BB13, Theorem 4.1.1]. This isomorphism identifies the distinguished triangle of Theorem 1.8 with the fundamental triangle of local cohomology. The following theorem is a formal consequence of Theorem 1.8: Theorem 1.10 ([BB13, Theorem 4.1.2]). Let i : Z ֒−→ X be a closed immersion. The functors i∗ and i! are a pair of inverse equivalences Db crys(QCrysκ(Z))) i∗ i! / Db crys(QCrysκ(X))Z . We call this equivalence the Kashiwara equivalence. If Z is a singular scheme which is embeddable into a smooth scheme X, the Kashiwara equivalence enables us to work with objects in Db crys(QCrys(X)) instead of Db crys(QCrys(Z)). 11 / o o 1.3 Review of locally finitely generated unit modules In [EK04], Emerton and Kisin consider left OF,X-modules, i.e. OX-modules M with a structural morphism F ∗M −→ M. Instead of localizing, they pass to a certain subcat- egory. If we speak of OF,X-modules we mean left OF,X-modules. In this subsection, all schemes are separated and of finite type over a field k containing Fp. Definition 1.11. Let X be a variety over k. A quasi-coherent OF,X-module is an OF,X- module whose underlying OX -module is quasi-coherent. If the structural morphism F ∗M −→ M of a quasi-coherent OF,X-module M is an isomorphism, then M is called unit. We let µ(X) and µu(X) denote the abelian categories of quasi-coherent and quasi- coherent unit OF,X-modules. The term "locally finitely generated" for an OF,X-module M means that M is locally finitely generated as a left OF,X-module. Emerton and Kisin's focus is on locally finitely generated unit modules, lfgu for short, on smooth schemes, where they form an abelian category. Definition 1.12. We let µlfgu(X) denote the abelian category of locally finitely gener- ated unit OF,X-modules. We let Dlfgu(OF,X ) denote the derived category of complexes of OF,X-modules whose cohomology sheaves are lfgu. Proposition 1.13. Let f : X −→ Y be a morphism of smooth k-schemes. The functor f ! : D(OF,Y ) −→ D(OF,X) defined by f !M• = OF,X→Y L ⊗f −1OF,Y f −1M•[dX/Y ] restricts to a functor f ! : Db lfgu(OF,Y ) −→ Db lfgu(OF,X). Here OF,X→Y denotes OF,X with the natural (OF,X , f −1OF,Y )-bimodule structure. Proof. This is Lemma 2.3.2 and Proposition 6.7 of [EK04]. Example 1.14. Let f : U −→ X be an open immersion of smooth k-schemes. Then we have dU/X = 0 and the inverse image of OF,X is the restriction to U : f −1OF,X = OF,XU = OF,U . Hence we regard OF,U →X as OF,U with the usual (OF,U , OF,U )-bimodule structure. It follows that f !M = f ∗M with the natural structure as a left OF,U -module for every left OF,X-module M. The construction of the push-forward is more involved. Emerton and Kisin first show that OF,Y ←X = f −1OF,Y ⊗f −1OY ωX/Y is naturally an (f −1OF,Y , OF,X)-bimodule. 12 We summarize the construction of the right OF,X-module structure from Proposition- Definition 1.10.1, Proposition-Definition 3.3.1 and Appendix A.2 of [EK04]: The relative Frobenius diagram is the diagram X ❇ FX/Y ❇❇❇❇❇❇❇ f X′ f ′ Y F ′ Y FY X f / Y. (1) Here X′ is the fiber product of X and Y considered as a Y -scheme via the Frobenius and FX/Y is the map obtained from the Frobenius FX : X −→ X and the morphism f . We call FX/Y the relative Frobenius. Unlike FX , it is a morphism of Y -schemes. Locally, for X = Spec S and Y = Spec R, the structure sheaf of X′ is given by the tensor product R ⊗R S, where R is viewed as an R-module via the Frobenius FR. Globally we have an isomorphism OX′ ∼= f −1OX F ⊗f −1OY OY where OXF denotes the submodule of OX [F ] = OF,X generated as a left OX -module by F . Consequently, for any OX -module M , F ′∗ Y M may be viewed as f −1OXF ⊗f −1OY M . Let γ : OY −→ F ∗ Y OY be the canonical isomorphism. The adjoint of the composition f !OY ∼−→ F ! X/Y f ′!OY F ! X/Y f ′!γ −−−−−−→ F ! X/Y f ′!F ∗ Y OY ∼−→ F ! X/Y F ′∗ Y f !OY yields a morphism CX/Y : FX/Y ∗ωX/Y −→ F ′∗ Y ωX/Y called the relative Cartier opera- tor. Note that FX/Y is the identity on the underlying topological spaces of X and X′. Therefore CX/Y defines a map of abelian sheaves ωX/Y −→ F ′∗ Y ωX/Y . Together with ∼= f −1OX F ⊗f −1OY ωX/Y and the inclusion OXF ⊂ OF,X the identification F ′∗ the relative Cartier defines a map ωX/Y −→ f −1OF,Y ⊗f −1OY ωX/Y . Now we can state the structure of f −1OF,Y ⊗f −1OY ωX/Y as a right OF,X-module. The endomorphism on f −1OF,Y ⊗f −1OY ωX/Y induced by multiplication with F ∈ OF,X on the right is given by the composition Y ωX/Y f −1OF,Y ⊗f −1OY ωX/Y CX/Y−−−→ f −1OF,Y ⊗f −1OY f −1OF,Y ⊗f −1OY ωX/Y m−→ f −1OF,Y ⊗f −1OY ωX/Y , where m is the the multiplication a ⊗ b 7→ ab in the sheaf of rings f −1OF,Y . The functor f+ : D(OF,X) −→ D(OF,Y ) is then defined by f+M• = Rf∗(OF,Y ←X L ⊗OF,X M•). Proposition 1.15. The functor f+ : D(OF,X) −→ D(OF,Y ) restricts to a functor f+ : Db lfgu(OF,X) −→ Db lfgu(OF,Y ). Proof. This is Theorem 3.5.3 and Proposition 6.8.2 of [EK04]. 13 / / / / / Example 1.16. Once again, let f : U −→ X be an open immersion. Then FU/X is an isomorphism, identifying X′ with the open subset U of X, and ωU/X = f !OX = f ∗OX = OU . Therefore we have OF,X←U = f −1OF,X ⊗f −1OX OU = OF,U ⊗OU OU . The left OF,U -module structures of OF,U ⊗OU OU and OF,U are obviously compatible ∼= OF,U . One verifies that this isomorphism with the natural isomorphism OF,U ⊗OU OU identifies the right OF,U -module structure on OF,U ⊗OU OU with the natural one on OF,U . Depending on f , there are adjunction relations between f ! and f+. If f is a closed immersion, a Kashiwara-type equivalence for unit modules holds. Lemma 1.17 ([EK04, Lemma 4.3.1.]). If f : X −→ Y is an open immersion of smooth k-schemes, then, for any M• ∈ D−(OF,Y ) and any N • ∈ D+(OF,X ), there is a natural isomorphism RHom• OF,Y (M•, f+N •) ∼−→ Rf∗RHom• OF,X (f !M•, N •) in D+(X, Z/pZ). Theorem 1.18 ([EK04, Theorem 4.4.1]). Let f : X −→ Y be a proper morphism of smooth k-schemes. For every M• in Db qc(OF,Y ), there is a natural isomorphism in D+(X, Z/pZ): qc(OF,X ) and every N • in Db RHom• OF,Y (f+M•, N •) ∼−→ Rf∗RHom• OF,X (M•, f !N •). qc(OF,X) denotes the subcategory of Db(OF,X) of complexes whose cohomology Here Db sheaves are quasi-coherent and analogously for Db qc(OF,Y ). For the proof, Emerton and Kisin show that the trace map f∗f ∆E• −→ E• for the residual complex E• of OX is compatible with the natural map E• −→ F ∗ X E•. Here f ∆ denotes the functor f ! for residual complexes, see [Har66, VI.3]. Thus it induces a morphism f+OF,X[dX/Y ] −→ OF,Y , and with the isomorphisms f+f !F • −→ f+(OF,X ⊗OF,X f !F •) −→ f+OF,X[dX/Y ] L ⊗OF,Y F •, the second one being a projection formula ([EK04, Lemma 4.4.7]), we obtain a trace map tr : f+f !F • −→ F • for every F • ∈ Db lfgu(OF,Y ). Similarly, as in the case of the adjunction between Rf∗ and f ! in Grothendieck-Serre duality, the natural transformation of the theorem is obtained by the composition Rf∗RHom• OF,X (M•, f !N •) / RHom• OF,Y (f+M•, f+f !N •) tr RHom• OF,Y (f+M•, N •), 14 / where the horizontal arrow is a natural transformation constructed in [EK04, Proposition 4.4.2]. It is this adjunction between f+ and f ! that we want to extend to morphisms which are only proper over the support of the considered complexes. This will be done in section 4. Finally, for a closed immersion of smooth varieties, we have a Kashiwara type equiv- alence. Theorem 1.19 ([EK04, Theorem 5.10.1]). If f : X −→ Y is a closed immersion of smooth k-schemes, then the adjunction of Theorem 1.18 provides an equivalence between the category of unit OF,X-modules and the category of unit OF,Y -modules supported on X. The fact that the natural map f+f !M −→ M is an isomorphism implies that H 0(f !)M ∼= f !M. 2 From Cartier crystals to locally finitely generated unit modules In order to construct an equivalence between Cartier crystals and locally finitely gen- erated unit modules, one uses an equivalence between Cartier modules and so-called γ-sheaves. It will induce an equivalence between Cartier crystals and γ-crystals. The latter in turn are known to be equivalent to lfgu modules. 2.1 Cartier modules and γ-sheaves We note that for a regular scheme X, the Frobenius FX : X −→ X is a flat morphism and hence F ∗ X is exact ([Kun69, Theorem 2.1]). Definition 2.1. A γ-sheaf on a regular, F -finite scheme X is a quasi-coherent OX - module N together with a morphism γN : N −→ F ∗N . The theory of γ-sheaves is very similar to that of Cartier modules. With the obvious morphisms, γ-sheaves form an abelian category with the nilpotent γ-sheaves being a Serre subcategory. We obtain γ-crystals in the same way as we obtained Cartier crystals and so forth. In this section we revisit the connection between Cartier modules and γ-sheaves as explained in section 5.2.1 of [BB11] and give some details of the proof. Definition 2.2. For any isomorphism ϕ : E1 −→ E2 of invertible OX -modules, while ϕ−1 : E2 −→ E1 denotes the inverse, let ϕ∨ denote the induced isomorphism E −1 2 −→ E −1 1 between the duals. If we speak of the γ-sheaf OX we mean the structure sheaf of X together with the natural isomorphism γX : OX −→ F ∗OX . By abuse of notation we call this isomorphism the Frobenius. For a regular, F -finite scheme X, let κX , or κR if X = Spec R is affine, denote the ∼−→ F ♭ωX, which is the adjoint of the Cartier operator if X is natural isomorphism ωX a smooth variety. 15 The next lemma makes explicit a fundamental isomorphism, which will be used re- peatedly. Lemma 2.3 ([BB11, Lemma 5.7]). Let f : X −→ Y be a finite and flat morphism of schemes. For every quasi-coherent OY -module F, there is a natural isomorphism can : f ♭OX ⊗OX f ∗G ∼−→ f ♭G. Proof. It suffices to construct a natural isomorphism locally and therefore we can iden- tify f with a ring homomorphism R −→ S and F with an R-module M . Define the homomorphism can : HomR(S, R) ⊗S (M ⊗R S) −→ HomR(S, M ) of S-modules by mapping α ⊗ (m ⊗ t) to the homomorphism s 7→ α(st)m. Since f is finite flat, we can assume that S is a free R-module and choose a basis s1, s2, . . . , sn. Let ϕ1, . . . , ϕn be the dual basis, i.e. ϕi ∈ HomR(S, R) and ϕi(j) = δij. One easily checks that the map HomR(S, M ) −→ HomR(S, R) ⊗S (M ⊗R S) ϕ 7→ nX i=1 ϕi ⊗ ϕ(si). is inverse to can. The following definition is extracted from [BB11, Theorem 5.9]. Definition 2.4. Let X be a regular, F -finite scheme. (a) For every Cartier module M with structural morphism κ, the sheaf M ⊗ ω−1 X has a natural γ-structure given by the composition M ⊗ ω−1 X κM ⊗(κ∨ X )−1 F ♭M ⊗ (F ♭ωX)−1 can−1 ⊗ can∨ F ♭OX ⊗ F ∗M ⊗ (F ♭OX )−1 ⊗ F ∗ω−1 X ∼ ∼ F ∗(M ⊗ ω−1 X ) ev F ♭OX ⊗ (F ♭OX )−1 ⊗ F ∗M ⊗ F ∗ω−1 X , where the vertical arrow on the right is the permutation and evL : L ⊗OX L−1 ∼−→ OX is the evaluation map l ⊗ ϕ 7→ ϕ(l). This morphism is called the γ-structure of M ⊗ ω−1 X induced by κM . 16 / / o o (b) For every γ-module N with structural morphism γN , the sheaf N ⊗ ωX has a natural Cartier structure given by the composition N ⊗ ωX γN ⊗κX F ∗N ⊗ F ♭ωX id⊗can−1 F ∗N ⊗ F ♭OX ⊗ F ∗ωX ∼ ∼ F ♭(N ⊗ ωX) can F ♭OX ⊗ F ∗N ⊗ F ∗ωX, where the vertical arrow on the right is the permutation. This morphism is called the Cartier structure of N ⊗ ωX induced by γN . Remark 2.5. Thanks to the fact that the proof of Lemma 2.3 contains explicit formulas for the isomorphism can and its inverse, we can concretely describe the induced Cartier structure of N ⊗ ωR for a γ-module N over a regular ring R such that F∗R is free with basis s1, . . . , sr. For m ∈ ωR set ϕm := κR(m) and let ϕ1, . . . , ϕr ∈ HomR(F∗R, R) be the dual basis of s1, . . . , sr, this means ϕi(sj) = δij. Following the arrows of Definition 2.4, we see that the Cartier structure N ⊗ ωX −→ F ♭(N ⊗ ωX) is given by n ⊗ m 7→ γ(n) ⊗ ϕm γ(n) ⊗ ϕi ⊗ ϕm(si) 7→ X 7→ (s 7→ X i γ(n) ⊗ ϕi(s) ⊗ ϕm(si)). i We will need this concrete version later on to prove that, for affine schemes, assigning a Cartier module to a γ-sheaf commutes with certain pullbacks, see Lemma 2.17. The use of the isomorphism F ♭OX ⊗ (F ♭OX )−1 ∼= OX involves the concrete formula for the structural morphism of the γ-sheaf associated to a Cartier module. Lemma 2.6. Let (N, γN ) be a γ-sheaf on a regular, F -finite scheme X. The adjoint F∗(N ⊗ ωX) −→ N ⊗ ωX of the structural morphism of the Cartier module N ⊗ ωX is given by the composition F∗(N ⊗ ωX) γN−−→ F∗(F ∗N ⊗ ωX) ∼−→ N ⊗ F∗ωX κX−−→ N ⊗ ωX, where the isomorphism in the middle is given by the projection formula. Proof. By construction, the structural morphism of N ⊗ ωX is the composition of the upper horizontal and the rightmost vertical arrow of the following diagram: N ⊗ ωX adj F ♭F∗(N ⊗ ωX) γN γN F ∗N ⊗ ωX adj F ∗N ⊗ F ♭F∗ωX κX / F ∗N ⊗ F ♭ωX adj ∼ ∼ / F ♭F∗(F ∗N ⊗ ωX) proj−1 / F ♭(N ⊗ F∗ωX) κX / F ♭(N ⊗ ωX). 17 / / o o / / / / / / / / Here adj denotes the respective adjunction morphism and proj is the isomorphism from the projection formula. The third and the fourth vertical morphism are isomorphisms stemming from can. For example, the morphism F ∗N ⊗ F ♭ωX −→ F ♭(N ⊗ F∗ωX) is the composition F ∗N ⊗ F ♭ωX id ⊗ can−1 −−−−−−→ F ∗N ⊗ F ∗ωX ⊗ F ♭OX can−−→ F ♭(N ⊗ ωX). Following the leftmost vertical and the lower horizontal arrows we obtain the adjoint of the morphism which is claimed to be the adjoint of the Cartier structure of N ⊗ ωX. Hence it suffices to show that the diagram above is commutative. The first and the last square commute by functoriality. The commutativity of the square in the middle can be checked locally on affine open subsets of X because F is an affine morphism. For the proof of Proposition 2.10, we need the isomorphism ωX ⊗ ω−1 X ∼= OX of γ- sheaves, which is a consequence of the following general lemma. Lemma 2.7. Let f : X −→ Y be a morphism of schemes and L an invertible OY -module. (a) If ρ : L ∼−→ L1 ⊗OX L2 is an isomorphism with invertible OX-modules L1 and L2, then the diagram L ⊗ L−1 ρ⊗(ρ∨)−1 id ⊗ evL1 / L2 ⊗ L−1 2 L1 ⊗ L2 ⊗ L−1 1 ⊗ L−1 2 r❡❡❡❡❡❡❡❡❡❡❡❡❡❡❡❡❡❡❡❡❡❡❡❡❡❡❡❡❡❡❡❡❡❡❡❡❡❡❡❡❡❡❡❡❡ evL2 evL OX commutes. (b) The diagram of canonical isomorphisms f ∗(L ⊗OY L−1) / f ∗L ⊗OX (f ∗L)−1 f ∗OX / OY commutes. Proof. It suffices to verify the claims for an affine scheme X = Spec R and, for (b), for a morphism of affine schemes Spec S −→ Spec R and an R-module M . In this affine situation the claims follow from straightforward calculations. Example 2.8. The γ-sheaf ω := ωX ⊗ ω−1 X on a regular, F -finite scheme X is canonically isomorphic to the structure sheaf OX equipped with the natural morphism OX −→ 18 / / / r / / F ∗OX. We have to show that the diagram ωX ⊗ ω−1 X evωX / OX (can−1 ◦κX )⊗(can−1 ◦(κ∨ X )−1) F ♭OX ⊗ F ∗ωX ⊗ (F ♭OX)−1 ⊗ F ∗ω−1 X id evF ♭OX ⊗ id ∼ F ∗ωX ⊗ F ∗ω−1 X ∼ F ∗(ωX ⊗ ω−1 X ) ∼ evF ∗ωX ∼ F ∗ evωX OX γX / F ∗OX commutes. The commutativity of both the top and the bottom rectangle follows from Lemma 2.7. Note that the horizontal isomorphisms of the lower square are the inverses of the natural isomorphisms of part (b) of Lemma 2.7. By definition, the γ-structure of ωX ⊗ ω−1 X is given by the composition of the vertical arrows on the right. Hence the γ-structure of OX with respect to the natural isomorphism OX Frobenius. ev−1 ωX−−−→ ωX ⊗ ω−1 X is the Similarly, starting with the γ-module OX , the induced Cartier structure of OX ⊗ ωX is compatible with κX with respect to the isomorphism OX ⊗ ωX ∼= ωX. Definition 2.9. Let QCohγ(X) denote the category of γ-sheaves and let Cohγ(X) denote the category of γ-sheaves whose underlying OX -module is coherent. We let QCrysγ(X) and Crysγ(X) denote the corresponding categories of crystals, see Definition 1.3. Proposition 2.10. If X is a regular, F -finite scheme, then tensoring with ωX and its inverse induces inverse equivalences of categories between Cartier modules and γ-sheaves on X: and QCohκ(X) Cohκ(X) ⊗OX ω−1 X ⊗OX ωX ⊗OX ω−1 X ⊗OX ωX / QCohγ(X) / Cohγ(X). In terms of this equivalence, the Cartier module (ωX, κX ) corresponds to the γ-sheaf (OX , γX ). Proof. Let ω denote the γ-sheaf ω−1 X ⊗ ωX and γω its structural morphism. (Note that there is no considerable difference between ωX ⊗ ω−1 X ⊗ ωX.) We start with a Cartier module (M, κM ). Consider the diagram (3) on page 23. Passing through the top arrow we follow the construction of the structural morphism of M ⊗ ω−1 X ⊗ ωX while the structural morphism of κM is given by the composition of the horizontal arrows on the X and ω−1 19 / / / / / o o / o o bottom: κM = can ◦(can−1 ◦κM ). Hence we have to show that (3) is commutative. Here µ denotes a permutation of the tensor product followed by the evaluation map, similar to the top most horizontal arrow. More precisely, it is the composition F ♭OX ⊗ F ∗M ⊗ (F ♭OX )−1 ⊗ F ∗ω−1 X ⊗ F ♭OX ⊗ F ∗ωX ∼ F ♭OX ⊗ F ∗M ⊗ F ∗ω−1 X ⊗ ((F ♭OX)−1 ⊗ F ♭OX ) ⊗ F ∗ωX ev F ♭OX ⊗ F ∗M ⊗ F ∗ω−1 X ⊗ F ∗ωX ∼ F ♭OX ⊗ F ∗(M ⊗ ω−1 X ⊗ ωX) of natural isomorphisms and ev. For simplicity, we will not distinguish between F ∗(M ⊗ ω−1 X ⊗ ωX) and F ∗M ⊗ F ∗ω−1 X ⊗ F ∗ωX. The commutativity of the upper square is an easy computation. In the lower left square the map from M ⊗ ωX ⊗ ω−1 X to F ♭OX ⊗ F ∗(M ⊗ ω−1 X ⊗ ωX) is the composition M ⊗ ωX ⊗ ω−1 X can−1 ◦κM −−−−−−−→ F ♭OX ⊗ F ∗M ⊗ ω−1 X ⊗ ωX id ⊗γω−−−−→ F ♭OX ⊗ F ∗(M ⊗ ω−1 X ⊗ ωX). Hence it suffices to show that F ∗M ⊗ ω−1 X ⊗ ωX id ⊗δ ∼ F ∗M ⊗ OX id ⊗γω id ⊗F F ∗M ⊗ F ∗(ω−1 X ⊗ ωX) id ⊗F ∗δ ∼ F ∗M ⊗ F ∗OX ∼ ∼ F ∗M ∼ (2) / F ∗(M ⊗ OX ) is commutative. The commutativity of the left square is Example 2.8 tensored with F ∗M . That the right square commutes can easily be checked by hand: For an arbitrary commutative ring R, an R-algebra S and an R-module M , the diagram (M ⊗R S) ⊗S S M ⊗R S (M ⊗R S) ⊗S (R ⊗R S) / (M ⊗R R) ⊗R S of natural homomorphisms is commutative. Along both ways an element (m ⊗ s1) ⊗ s2 is mapped to (m ⊗ 1) ⊗ s1s2. Locally the right square of (2) is just a special case of this diagram. Now let (N, γN ) be a γ-sheaf. By definition, the structural morphism γ′ N of N ⊗ ωX ⊗ X is the line in the middle of the diagram (4). The upper squares of this diagram ω−1 20 o o / / o o / / / / commute by construction. Here the horizontal morphism to the top right corner is given by id ⊗(can−1 ◦κX ) and the horizontal morphism below is the unique morphism making the upper right square commute. Therefore we see that γ′ N is the tensor product of γN and γω, i.e. the bottom rectangle of (4) is commutative. Now consider the diagram N ⊗ ω id ⊗γω / N ⊗ F ∗ω γN ⊗id / F ∗N ⊗ F ∗ω id ⊗δ id ⊗F ∗δ id ⊗F ∗δ N ⊗ OX id ⊗γX / N ⊗ F ∗OX γN ⊗id / F ∗N ⊗ F ∗OX ∼ ∼ / F ∗(N ⊗ ω) F ∗(id ⊗δ) / F ∗(N ⊗ OX ) ∼ N γN id ⊗γX / F ∗N ⊗ OX ∼ ∼ / F ∗N. The upper left square is the commutative diagram of Example 2.8 tensored with N . The upper square in the middle and the bottom left rectangle are clearly commutative. The square to the left of it commutes because of the naturality of the isomorphism ) ∼−→ F ∗(N ⊗ ). We already have seen that the bottom right square F ∗N ⊗ F ∗( commutes: It is the same square as the left one of diagram (2) with M replaced by N . Moreover, the composition of the leftmost arrow is N ⊗ δ and the composition of the rightmost arrow is F ∗(N ⊗ δ). Hence the structural morphism of N is compatible with γN ⊗ γω, which turned out to be compatible with the structural morphism induced from the Cartier module N ⊗ ωX. Thus we can extract the commutative diagram N ⊗ ωX ⊗ ω−1 X id ⊗δ N γ′ N γN / F ∗(N ⊗ ωX ⊗ ω−1 X ) F ∗(id ⊗δ) / F ∗N. ⊗ ω−1 ⊗ ωX are inverse equivalences. From It follows that the functors Example 2.8 we know that ⊗ ω−1 X maps ωX with the structural morphism κX to OX with the structural morphism γX . Consequently, ⊗ ωX maps the γ-sheaf (OX , γX ) to the Cartier module (ωX, κX ). X and Corollary 2.11. Tensoring with ωX and with ω−1 X induces equivalences of categories and QCrysκ(X) Crysκ(X) for every regular, F -finite scheme X. / QCrysγ(X) / Crysγ(X) ⊗OX ω−1 X ⊗OX ωX ⊗OX ω−1 X ⊗OX ωX 21 / / / O O / O O / O O / O O O O / O O / O O / O O / O O / o o / o o Corollary 2.12. If X is regular and F -finite, the categories QCohγ(X) and QCrysγ(X) have enough injectives. Proof. This follows from Proposition 1.4 and Corollary 2.11. 22 (F ♭OX ⊗ (F ♭OX)−1) ⊗ F ∗M ⊗ F ∗ω−1 X ⊗ F ♭OX ⊗ F ∗ωX ev / F ∗M ⊗ F ∗ω−1 X ⊗ F ♭OX ⊗ F ∗ωX ∼ ∼ F ♭OX ⊗ F ∗M ⊗ (F ♭OX )−1 ⊗ F ∗ω−1 X ⊗ F ♭OX ⊗ F ∗ωX µ / F ♭OX ⊗ F ∗(M ⊗ ω−1 X ⊗ ωX) can / F ♭(M ⊗ ω−1 X ⊗ ωX) can−1(κM ⊗(κ∨ X )−1⊗κX ) M ⊗ ω−1 X ⊗ ωX id ⊗δ ∼ M id ⊗F ∗(id ⊗δ) F ♭ϕ can−1 ◦κM / F ♭OX ⊗ F ∗M can / F ♭M (3) 2 3 F ∗N ⊗ F ♭OX ⊗ F ∗ωX ⊗ ω−1 X ∼ / F ♭OX ⊗ F ∗N ⊗ F ∗ω−1 X ⊗ ω−1 X γN ⊗(can−1 ◦κX )⊗id can ∼ N ⊗ ωX ⊗ ω−1 X κN⊗ωX / F ♭(N ⊗ ωX) ⊗ ω−1 X F ♭OX ⊗ F ∗N ⊗ F ∗ω−1 X ⊗ (F ♭OX )−1 ⊗ F ∗ω−1 ∼ evF ♭OX ⊗ id X / F ∗(N ⊗ ωX ⊗ ω−1 X ) N ⊗ ωX ⊗ ω−1 X γN ⊗γω / F ∗N ⊗ F ∗(ωX ⊗ ω−1 X ) ∼ / F ∗(N ⊗ ωX ⊗ ω−1 X ) (4) / O O / / O O O O / O O / O O / / / O O / / / / 2.2 Compatibility with pull-back The pull-back of quasi-coherent sheaves defines a pull-back functor on γ-sheaves: Definition 2.13. Let f : X −→ Y be a morphism of regular schemes and N a γ-sheaf on Y with structural morphism γN . The γ-structure for f ∗N is defined as the composition f ∗N f ∗γN−−−→ f ∗F ∗ Y N ∼−→ F ∗ Y f ∗N. First we consider a closed immersion i : X −→ Y of regular F -finite schemes. The aim is to prove the following theorem: Theorem 2.14. Let i : X −→ Y be a closed immersion of regular, F -finite schemes with codimension n. Then there is a canonical isomorphism of functors ⊗ ω−1 X ◦ Rni♭ ∼= i∗ ◦ ⊗ ω−1 Y inducing a corresponding isomorphism of functors of crystals, i.e. the diagram Crysκ(Y ) Rni♭ Crysκ(X) ⊗OY ω−1 Y ⊗OY ωY ⊗OX ω−1 X ⊗OX ωX / Crysγ(Y ) i∗ / Crysγ(X) is commutative. We begin with the affine case. Noting that any closed immersion of regular schemes is a local complete intersection morphism, it suffices to consider the case of a complete intersection, where the pull-back of a Cartier module can be computed by using the Koszul complex. Lemma 2.15. Let f = f1, f2, . . . , fn be a regular sequence of elements of a commutative ring R. Let I be the ideal generated by the fi. Then for every Cartier module M with structural map κ, there is an isomorphism ϕf : Extn R(R/I, M ) ∼−→ M/IM where M/IM is viewed as an R/I-module with the Cartier structure κM/IM : F∗M/IM −→ M/IM m + IM 7→ κM ((f1 · f2 · · · fn)p−1m) + IM. Proof. By definition we have to compute RHomR(R/I, M ). The structural morphism κi♭M : F∗i♭M −→ i♭M equals the composition F∗ RHomR(R/I, M ) −→ RHomR(F∗R/I, F∗M ) −→ RHomR(R/I, M ), 24 / o o / o o where the first morphism is the canonical one and the second is induced by F : R/I −→ F∗R/I in the first and κ : F∗M −→ M in the second argument. A free resolution of the R-module R/I is given by the Koszul chain complex K(f ). It is the total tensor product complex in the sense of [Wei94, 2.7.1] of the following complexes K(fi) 0 −→ R fi−→ R −→ 0 concentrated in degrees −1 and 0. Each complex K(fi) admits a lift of the Frobenius F : R −→ F∗R in degree 0 by mapping r to rpf p−1 . This means the diagrams i 0 0 fi F fi / R f p−1 i / F∗R / R F / F∗R / 0 / 0 f p−1 i−−−→ F∗R give rise to a map of complexes F : K(f ) −→ F∗K(f ) commute. The maps R lifting the Frobenius in degree zero: In general, if ϕ : M −→ F∗M and ψ : N −→ F∗N are R-linear maps, it is easy to check that the map M ⊗R N −→ M ⊗R N m ⊗ n 7→ ϕ(n) ⊗ ψ(n) of abelian groups is p-linear, i.e. it is an R-linear map M ⊗ N −→ F∗(M ⊗ N ). Thus we inductively obtain an R-linear morphism of complexes F : K(f ) −→ F∗K(f ). Unwinding the definition of the tensor product of complexes, we see that the left end of this map is the square 0 0 / R ((−1)i+1fi)i / Rn Qn i=1 f p−1 i (Qj6=i f p−1 j )i / F∗R ((−1)i+1fi)i / F∗Rn / · · · / · · · Consequently, the n-th degree of the composition F∗ HomR(K(f ), M ) canonical −−−−−→ HomR(F∗K(f ), F∗M ) F ◦ ◦κ −−−−→ HomR(K(f ), M ) maps m to κ((Q f p−1 differential HomR(Kn−1(f ), M ) −→ HomR(Kn(f ), M ) corresponds to the map )m) by the identification HomR(R, M ) ∼= M via ϕ 7→ ϕ(1). The i M n −→ M (mi)i 7→ X fimi. The image is the submodule IM and thus the n-th cohomology of HomR(K(f ), M ) is isomorphic to M/IM , equipped with the claimed Cartier structure. 25 / / / / / / / / / / / / Remark 2.16. The isomorphism Extn R(R/I, M ) −→ M/IM of the underlying sheaves is not canonical. It depends on the choice of the regular sequence f . From the construction of this isomorphism we see that if g = g1, . . . , gn is another regular sequence of R generating I and gi = P cijfj, the automorphism τ on M/IM making the diagram Extn R(R/I, M ) τ ϕf 7♦♦♦♦♦♦♦♦♦♦♦ '❖❖❖❖❖❖❖❖❖❖❖ ϕg M/IM M/IM commutative is given by multiplication with det(cij). Nevertheless, interpreting the top-Ext-groups as quotient modules in the case we are interested in, namely Extn R(R/I, ωR)−1, leads to isomorphisms, which are independent of the regular sequence generating I, since the correcting factors from both terms cancel. R(R/I, M ) ⊗ Extn Lemma 2.17. Let R be a commutative ring such that F∗R is finite free, N a γ-module over R and I ⊆ R an ideal which is generated by a regular sequence of length n. There is a canonical isomorphism between Cartier modules Extn R(R/I, N ⊗ ωR) ∼= N/IN ⊗ ωR/I . Proof. Choose a regular sequence f = f1, . . . , fn such that I is generated by the fi. Also choose a basis r1, . . . , rt of R viewed as a free R-module via the Frobenius. The dual basis ϕ1, . . . , ϕt is given by ϕi(rj) = δij. Let κ denote the intrinsic Cartier structure of Extn R(R/I, N ⊗ ωR) and κ denote the Cartier structure of N/IN ⊗ ωR/I as explained in Proposition 2.10. Identifying Extn R(R/I, N ⊗ ωR) with (N ⊗ ωR)/I(N ⊗ ωR) via the isomorphism ϕf from Lemma 2.15, we obtain the map κ′ : F∗((N ⊗ ωR)/I(N ⊗ ωR)) −→ (N ⊗ ωR)/I(N ⊗ ωR) n ⊗ m + I(N ⊗ ωR) 7→ tX i=1 ϕi(1)γN (n) ⊗ κR(rif p−1m) + I(N ⊗ ωR) as the induced Cartier structure on (N ⊗ ωR)/I(N ⊗ ωR). Here γN is the γ-structure of N . Also identifying ωR/I R(R/I, ωR) with ωR/IωR via f , its Cartier structure κR/I is given by m + IωY 7→ κR(f p−1m) + I(ωY ). From this perspective, the Cartier structure ∼= Extn 26 7 ' of κ of N/IN ⊗ ωR/I induces the structural morphism κ′ : F∗(N/IN ⊗ ωR/IωR) −→ (N/IN ⊗ ωR/IωR) n ⊗ m 7→ nX i=1 ϕi(1) · γN (n) ⊗ κR(rif p−1m) = tX i=1 ϕi(1)γN (n) ⊗ κR(rif p−1m) + I(N ⊗ ωR) on N/IN ⊗ ωR/IωR. Finally, there is a natural isomorphism τ : (N ⊗R ωR)/I(N ⊗R ωR) ∼−→ N/IN ⊗R/I ωR/IωR of R-modules, mapping n ⊗ m to n ⊗ m. The explicit formulas for κ′ and κ′ show that τ makes the square in the middle of the diagram Extn R(R/I, N ⊗R ωR) ϕf (N ⊗ ωR)/I(N ⊗ ωR) τ N/IN ⊗ ωR/IωR id ⊗ϕf i∗N ⊗R ωR/I κ κ′ κ′ κ / F ♭ Extn R(R/I, N ⊗R ωR) F ♭ϕf F ♭((N ⊗ ωR)/I(N ⊗ ωR)) F ♭τ / F ♭(N/IN ⊗ ωR/IωR) F ♭(id ⊗ϕf ) F ♭(i∗N ⊗R ωR/I ) commutative. The squares above and below commute by construction. Let Φ be the composition (id ⊗ϕωR . We have just seen that the diagram f )−1 ◦ τ ◦ ϕN ⊗ωR f Extn R(R/I, N ⊗R ωR) Φ i∗N ⊗R ωR/I κ κ / F ♭ Extn R(R/I, N ⊗R ωR) F ♭Φ / F ♭(i∗N ⊗R ωR/I ) commutes. Furthermore, Φ is natural: Let g = g1, . . . , gn be another regular sequence generating I with gi = P cijfj. Then, by Remark 2.16, = det(cij)−1(id ⊗ϕωR = (id ⊗ϕωR f )−1 ◦ τ ◦ det(cij)ϕN ⊗ωR g )−1 ◦ τ ◦ ϕN ⊗ωR (id ⊗ϕωR . f )−1 ◦ τ ◦ ϕN ⊗ωR f g f 27 / / / / / / O O O O / / Proposition 2.18. Let i : X ֒−→ Y be a closed immersion of regular, F -finite schemes and let N be a γ-sheaf on Y . There is a canonical isomorphism Φ : i ∗ Extn OY (i∗OX, N ⊗OY ωY ) ∼−→ i∗N ⊗OX ωX of Cartier modules, which is functorial in N. Here i denotes the flat morphism of ringed spaces (X, OX ) −→ (Y, i∗OX ). Proof. Choose an affine open covering {Uk}k = Spec Rk of Y such that iUk : i−1(Uk) ֒−→ Uk is a complete intersection. By refining the covering we can assume that F∗Rk is free. Let Ik ⊆ Rk be the ideal such that iUk corresponds to the ring homomorphism R(Rk/Ik, N Uk ⊗ ωRk) ∼= Rk −→ Rk/Ik. By Lemma 2.17, we have an isomorphism Extn i∗N Uk ⊗ ωRk/Ik , which is natural and therefore, we can glue the local isomorphisms to the desired global map Φ. ωY Proposition 2.18 shows that there is a natural isomorphism of functors Rni♭ ◦ ⊗OY ∼= ⊗OX ωX ◦ i∗. This enables us to prove Theorem 2.14 because ⊗OY ωY and ⊗OX ωX are equivalences of categories. Proof of Theorem 2.14. For every γ-sheaf N on Y , there is an isomorphism Rni♭(N ⊗OY ωY ) ∼= i−1Extn OY (i∗OX , N ⊗OY ωY ) ∼= i∗N ⊗OX ωX by Proposition 2.18, which is functorial in N . As ⊗OY ωY and ⊗OX ωX are equiv- alences of categories, even the diagram Cohκ(Y ) Rni! Cohκ(X) ⊗OY ω−1 Y ⊗OY ωY ⊗OX ω−1 X ⊗OX ωX / Cohγ(Y ) i∗ / Cohγ(X) commutes. Passing to crystals finishes the proof. Now we turn to open immersions. Proposition 2.19. Let j : U −→ X be an open immersion of regular, F -finite schemes and M a Cartier module on X. Then there is a natural isomorphism of γ-sheaves j!M ⊗OU ω−1 U ∼= j∗(M ⊗ ω−1 X ). Proof. One easily checks that, for a Cartier module M on X with structural morphism κ : M −→ F ♭ X M , the Cartier structure on j∗M is the composition j∗M j∗κ −−→ j∗F ♭M ∼−→ F ♭ U j∗M. 28 / o o / o o The dualizing sheaf ωU of U is given by j∗ωX. Therefore we have j!M ⊗ ω−1 U ω−1 X ) and the diagram ∼= j∗(M ⊗ j∗M ⊗ j∗ω−1 X / F ♭j∗M ⊗ F ♭j∗ω−1 X ∼ / F ∗(j∗M ⊗ j∗ω−1 X ) ∼ ∼ ∼ j∗(M ⊗ ω−1 X ) / j∗(F ♭M ⊗ F ♭ω−1 X ) ∼ / j∗F ∗(M ⊗ ω−1 X ) commutes. Here the horizontal arrows are the γ-structures of j∗M ⊗OU ω−1 ω−1 X ). U and j∗(M ⊗ 2.3 Compatibility with push-forward For the construction of a push-forward for γ-sheaves, we follow the construction given in subsection 6.3 of [BB]. Then we show that the equivalence between Cartier modules and γ-sheaves given by tensoring with the dualizing sheaf is compatible with push-forward for morphisms of regular schemes. This proof is also mainly the one given in ibid. By abuse of notation, let κX : FX∗ωX −→ ωX be the adjoint of the Cartier structure of ωX. Definition 2.20 ([BB, Definition 6.3.1]). Let f : X −→ Y be a morphism of smooth, F -finite k-schemes. Let N be a γ-sheaf on X. Then we define the push-forward f+N as the twist of the push-forward f∗ of Cartier modules, i.e. f+N = f∗(N ⊗OX ωX) ⊗OY ω−1 Y . The push-forward for γ-crystals is the one induced by the just given push-forward of γ-sheaves. By construction, the push-forward for γ-sheaves is compatible with the push-forward for Cartier modules. In order to show that f+ is compatible with the equivalence between γ-crystals and lfgu modules, we need a different description of f+ for γ-sheaves based on the relative Cartier operator. We recall two general constructions, which are repeatedly used in this subsection. For this we consider a morphism f : X −→ Y of arbitrary schemes over Spec Z. Let F be a quasi-coherent OX -module and E a quasi-coherent OY -module. The adjoint of the composition f ∗(f∗F ⊗OX E) ∼= f ∗f∗F ⊗ f ∗E eadf ⊗id −−−−→ F ⊗ f ∗E where fadf : f ∗f∗ −→ id is the counit of the adjunction, yields a natural morphism proj : f∗F ⊗OX E −→ f∗(F ⊗ f ∗E). As a consequence of the projection formula ([Har77, Exercise III.8.3]), it is an isomor- phism if f is quasi-compact and separated and if E is locally free. 29 / / / / For two morphisms f : X −→ S and g : Y −→ S, let f ′ : X×S Y −→ Y and g′ : X×S Y −→ X be the projections such that the square X ×S Y g′ X f ′ f Y g / S is cartesian. There is a canonical morphism of functors bc: f ∗g∗ −→ g′ ∗f ′∗ of quasi-coherent sheaves given by the adjoint of the composition g∗ adf ′ −−−−→ g∗f ′ ∗f ′∗ ∼= f∗g′ ∗f ′∗, g∗ ∗f ′∗ is the unit of the adjunction. If g is affine, bc is an isomorphism. where adf ′ : id −→ f ′ To see this, we can assume that S, X and Y are affine, because g and therefore g′ is an affine morphism. Then the claim is a well known property of the tensor product. The morphism bc is also an isomorphism if X and Y are Noetherian, f is flat and g is separated of finite type ([Har77, Proposition III.9.3]). The next lemma relates these two isomorphisms. Lemma 2.21. Let f : X −→ S and g : Y −→ S be morphisms of schemes and let f ′ : X ×S X −→ Y and g′ : X ×S Y −→ X be the projections. Then, for every quasi-coherent OX- module F and every quasi-coherent OS-module E, the diagram f∗F ⊗ g∗E proj f∗F ⊗ g∗E proj f∗(F ⊗ f ∗g∗E) g∗(g∗f∗F ⊗ E) bc bc f∗(F ⊗ g′ ∗f ′∗E) g∗(f ′ ∗g′∗F ⊗ E) proj proj f∗g′ ∗(g′∗F ⊗ f ′∗E) ∼ / g∗f ′ ∗(g′∗F ⊗ f ′∗E) commutes. Proof. For a morphism h of schemes, let fadh denote the counit of adjunction h∗h∗ −→ id. The diagram of which we want to prove the commutativity is obtained by adjunction 30 / / / / from the diagram g′∗f ∗f∗F ⊗ g′∗f ∗g∗E ∼ f ′∗g∗f∗F ⊗ f ′∗g∗g∗E eadf ⊗id g′∗F ⊗ g′∗f ∗g∗E ∼ id ⊗ bc ∗f ′∗E g′∗F ⊗ g′∗g′ id ⊗eadg′ g′∗F ⊗ f ′∗E bc ⊗ id f ′∗f ′ ∗g′∗F ⊗ f ′∗g∗g∗E eadf ′ ⊗id g′∗F ⊗ f ′∗g∗g∗E id ⊗eadg g′∗F ⊗ f ′∗E. Both parts of the diagram are commutative by construction of the morphism bc. We turn back to the situation of a morphism f : X −→ Y of smooth schemes over a field k containing Fp. For simplicity, let ωf denote the relative dualizing sheaf ωX/Y = f !OY . Lemma 2.22. Let f : X −→ Y be a morphism of smooth, F -finite schemes over k. For every γ-sheaf N on X, there is a natural isomorphism f∗(N ⊗OX ωX) ⊗OY ω−1 Y −→ f∗(N ⊗OX ωf ) of quasi-coherent sheaves. Proof. Since X and Y are smooth, any morphism X −→ Y is regular, i.e. it is a com- position of a closed immersion X −→ W such that X is a local complete intersection in W , followed by a smooth morphism W −→ Y . (For a smooth morphism f , the graph factorization X (id,f ) −−−→ X ×k Y prY−−→ Y, where prY denotes the projection, satisfies this requirement.) For a closed immersion i we have the isomorphism i♭(ωY ) ∼= Li∗ωY ⊗ ωf [−n] of [Har66, Corollary III.7.3] and a smooth morphism is quasi-perfect, see [Sch18, Defi- nition 3.3]. Overall, we see that there are natural isomorphisms ωX ∼= f !ωY ∼−→ f !OY ⊗ Lf ∗ωY ∼−→ ωf ⊗ f ∗ωY . Now we obtain the desired isomorphism as the composition f∗(N ⊗ ωX) ⊗ ω−1 Y proj −−→ f∗(N ⊗ ωX ⊗ f ∗ω−1 Y ) ∼= f∗(N ⊗ ωf ). 31 / / / / With the relative Frobenius diagram X FX/Y &▲▲▲▲▲▲▲▲▲▲▲▲▲ f FX X′ f ′ Y F ′ Y FY X f / Y, see diagram (1) and below for the notation, we can define a γ-structure γN,f for f∗(N ⊗ ωf ) by the composition ∗FX/Y ∗(N ⊗ ωf ) ∗FX/Y ∗(F ∗ X/Y F ′∗ Y N ⊗ ωf ) f∗(N ⊗ ωf ) ∼−→ f ′ γN−→ f ′ proj−1 −−−−→ f ′ CX/Y−−−→ f ′ bc−→ F ∗ ∗(F ′∗ ∗(F ′∗ Y N ⊗ FX/Y ∗ωf ) Y N ⊗ F ′∗ Y ωf ) Y f∗(N ⊗ ωf ). We will show that γN,f is the structural morphism of f+N via the isomorphism of Lemma 2.22. But first, we clarify how the relative Cartier operator is related to κX and κY . Lemma 2.23. With the notation of the preceding lemma, the composition X/Y f ′∗ωY ) Y ∗FX/Y ∗(ωf ⊗ F ∗ FY ∗(ωf ⊗ f ∗ωY ) ∼−→ F ′ proj−1 −−−−→ F ′ CX/Y−−−→ F ′ Y ωf ⊗ f ′∗ωY ) proj −−→ ωf ⊗ F ′ Y ∗f ′∗ωY bc−→ ωf ⊗ f ∗FY ∗ωY κY−→ ωf ⊗ f ∗ωY Y ∗(FX/Y ∗ωf ⊗ f ′∗ωY ) Y ∗(F ′∗ is compatible with the Cartier structure of ωX under the canonical isomorphism ωX ωf ⊗ f ∗ωY . Proof. In the appendix A.2.3. (iii) of [EK04], Emerton and Kisin explain how the relative −→ Z Cartier operators CX/Y , CY /Z and CX/Z are related for a composition X of morphisms. Our lemma is the special case where Z = Spec k and g is the structural morphism of the k-scheme Y . −→ Y f g ∼= Proposition 2.24. Let f : X −→ Y be a morphism of smooth, F -finite schemes over k. Let N be a γ-sheaf on X. The canonical isomorphism f∗(N ⊗ ωf ) ∼−→ f∗(N ⊗ ωX) ⊗ ω−1 Y 32 / / & % % / / / of quasi-coherent OY -modules from Lemma 2.22 is an isomorphism of γ-sheaves. Proof. As ⊗ ωX and ⊗ ω−1 are equivalences between the categories of γ-sheaves Y and Cartier modules on X and on Y , it suffices to show that the canonical isomorphism ∼−→ f∗(N ⊗ ωX) is an isomorphism of Cartier modules on Y . The left f∗(N ⊗ ωf ) ⊗ ωY hand side of the diagram FY ∗(f∗(N ⊗ ωf ) ⊗ ωY ) γN FY ∗(f∗(F ∗ Y N ⊗ ωf ) ⊗ ωY ) proj proj FY ∗f∗(N ⊗ ωf ⊗ f ∗ωY ) γN FY ∗f∗(F ∗ X N ⊗ ωf ⊗ f ∗ωY ) ∼ ∼ FY ∗(f ′ ∗FX/Y ∗(F ∗ X/Y F ′∗ Y N ⊗ ωf ) ⊗ ωY ) FY ∗(f ′ ∗FX/Y ∗(F ∗ X/Y F ′∗ Y N ⊗ ωf ) ⊗ ωY ) FY ∗f ′ ∗FX/Y ∗(F ∗ X/Y F ′∗ Y N ⊗ F ∗ X/Y f ′∗ωY ⊗ ωf ) ∼ FY ∗f ′ ∗FX/Y ∗(F ∗ X/Y (F ′∗ Y N ⊗ f ′∗ωY ) ⊗ ωf ) proj−1 proj−1 FY ∗(f ′ ∗(F ′∗ Y N ⊗ FX/Y ∗ωf ) ⊗ ωY ) CX/Y FY ∗(f ′ ∗F ′∗ Y (N ⊗ ωf ) ⊗ ωY ) bc−1 FY ∗(F ∗ Y f∗(N ⊗ ωf ) ⊗ ωY ) proj−1 f∗(N ⊗ ωf ) ⊗ FY ∗ωY κY f∗(N ⊗ ωf ) ⊗ ωY proj proj proj proj / FY ∗f ′ ∗(F ′∗ Y N ⊗ f ′∗ωY ⊗ FX/Y ∗ωf ) CX/Y / FY ∗f ′ ∗(F ′∗ Y (N ⊗ ωf ) ⊗ f ′∗ωY ) ∼ f∗F ′ Y ∗(F ′∗ Y (N ⊗ ωf ) ⊗ f ′∗ωY ) proj−1 f∗(N ⊗ ωf ⊗ F ′ Y ∗f ′∗ωY ) bc−1 / f∗(N ⊗ ωf ⊗ f ∗FY ∗ωY ) κY / f∗(N ⊗ ωf ⊗ f ∗ωY ) is the structural morphism of the Cartier module f∗(N ⊗ ωf ) ⊗ ωY . It is easy to see that the right hand side is the structural morphism of the Cartier module f∗(N ⊗ ωX) ∼= f∗(N ⊗ωF ⊗f ∗ωY ). Hence we have to show that the diagram above commutes. The three upper squares and the bottom square commute by the functoriality and the compatibility of the projection formula with compositions of morphisms. The commutativity of the fourth square from above follows from Lemma 2.21. 33 / / / / / / / / 2.4 Cartier crystals and locally finitely generated unit modules The category of γ-sheaves was just an intermediate step on the way to locally finitely generated unit modules. Recall that there is a functorial way of associating a unit OX [F ]-module to a γ-sheaf N on X. Definition 2.25. Let µu(X) denote the category of unit left OF,X-modules whose un- derlying OX -module is quasi-coherent. For a smooth k-scheme X, let Gen be the functor QCohγ(X) −→ µu(X) which assigns to any quasi-coherent γ-sheaf N with structural morphism γ : N −→ F ∗N the direct limit N of γ −→ F ∗N N F ∗γ −−→ F 2∗N F 2∗γ −−−→ · · · together with the inverse of the induced isomorphism N ∼−→ F ∗N . Lemma 2.26. Let X be a smooth, F -finite k-scheme. The functor Gen is essentially surjective and induces an equivalence of categories QCrysγ(X) ∼−→ µu(X). Proof. Let Neg : µu(X) −→ QCohγ(X) be the functor which assigns to a quasi-coherent unit OF,X-module M with structural morphism u : F ∗M −→ M the quasi-coherent γ- sheaf M whose structural morphism is given by the inverse of u. Obviously there is a natural isomorphism Gen ◦ Neg ∼−→ id, whence the surjectivity of Gen. For a quasi-coherent γ-sheaf N , the corresponding crystals N is nil-isomorphic to the corresponding crystal of Neg ◦ Gen(N ). The reason for this is the fact that the structural morphism γ : N −→ F ∗N of a quasi-coherent γ-sheaf N is a nil-isomorphism: It is immediate that the structural map of the kernel and the cokernel of γ, interpreted as a morphism of γ-sheaves, is the zero map. The image of Gen of the subcategory Cohγ(X) of coherent γ-sheaves on X is the Indeed, after localizing at nilpotent γ-sheaves and considering γ- category µlfgu(X). crystals, Gen induces an equivalence of categories. Proposition 2.27 ([BB11, Proposition 5.12]). For a smooth, F -finite k-scheme X, the functor factors through Crysγ(X), inducing an equivalence of categories: GenX : Cohγ(X) −→ µlfgu(X) Cohγ(X) Gen &▼▼▼▼▼▼▼▼▼▼ Crysγ(X) ∼ / µlfgu(X). 34 & / Theorem 2.28. Let X be an F -finite, smooth k-scheme. Let G denote the composition of the exact functors ⊗ ω−1 X and Gen. It induces an equivalence of derived categories G : Db crys(QCrysκ(X)) −→ Db lfgu(OF,X). Proof. Combining Corollary 2.11 and Lemma 2.26, we see that G induces an equivalence of abelian categories QCrysκ(X) −→ µu(X) and therefore an equivalence of derived categories Db(QCrys(X)) −→ Db(µu(X)). Since G is exact and restricts to an equivalence Crysκ(X) −→ µlfgu(X), we obtain an equivalence Db crys(QCrysκ(X)) −→ Dlfgu(µu(X)). It remains to show that Db lfgu(OF,X). The inclusion µlfgu(X) −→ µ(X) induces an equivalence Db(µlfgu) −→ Db lfgu(OF,X) ([EK04, 11.6]). As the inclusion µlfgu(X) −→ µ(X) factors through µu(X), this implies an equiv- alence Db lfgu(µu(X)) is naturally equivalent to Db lfgu(µu(X)) −→ Db lfgu(OF,X). Finally, we prove that the equivalence G of derived functors is compatible with pull- backs. Note that for a morphism f : X −→ Y of smooth schemes, the functor f ! : Db lfgu(OF,Y ) −→ Db lfgu(OF,X) is obtained from a right-exact functor of abelian categories. Definition 2.29. Let f : X −→ Y be a morphism of smooth k-schemes. The (underived) pull-back f ∗M of an OF,Y -module M is given by f ∗M = OF,X→Y ⊗f −1OF,Y f −1M = OF,X ⊗f −1OF,Y f −1M, cf. Proposition 1.13. The pull-back f ! for complexes M• of OF,Y -modules from Defini- tion 2.3.1 of [EK04] is the left derived functor of f ∗, shifted by dX/Y : f !M• = OF,X→Y L ⊗f −1OF,Y f −1M•[dX/Y ]. Corollary 2.30. Let f : X −→ Y be a closed immersion of smooth, F -finite k-schemes of relative dimension dX/Y = n. There is a natural equivalence of functors Crysκ(Y ) −→ µlfgu(X): f ∗ ◦ GY ∼= GX ◦Rnf ♭. Proof. Consider the following diagram of functors: Crysκ(Y ) Rnf ♭ Crysκ(X) ⊗ω−1 Y ⊗ω−1 X GenY Crysγ(Y ) f ∗ µlfgu(Y ) f ∗ / Crysγ(X) GenX / µlfgu(X). The left square commutes by Theorem 2.14. The right square also commutes because there is a natural isomorphism GenX ◦f ∗ ∼= f ∗ ◦ GenY . For a γ-sheaf N on Y , let N 35 / / / / / / denote GenY (N ), which is the direct limit lim−→ F i∗ pull-back of quasi-coherent sheaves, we have a natural isomorphism Y N . As direct limits commute with GenX f ∗(N ) = lim−→(OX ⊗f −1OY f −1F i∗ Y N ) ∼−→ OX ⊗f −1OY f −1(lim−→ F i∗ = OX ⊗f −1OY f −1(N ). Y N ) X f ∗M −→ f ∗M induced by the structural morphism F ∗ One checks that for a left OF,Y -module M, the underived pullback f ∗M = OF,X⊗f −1OF,Y f −1M is the quasi-coherent sheaf f ∗M = OX ⊗f −1OY f −1M with the natural mor- phism F ∗ Y M −→ M. Hence OX ⊗f −1OY f −1(N ) is isomorphic to f ∗ GenY (N ). Lemma 2.31. Let i : X −→ Y be a closed immersion of smooth, F -finite schemes over k. Let P be a locally free left OF,Y -module. Then Rn(( ⊗ ω−1 X ) ◦ i! ◦ ( ⊗ ωY ))P = 0 for all n 6= −dX/Y , where ( ⊗ ω−1 X ) ◦ i! ◦ ( ⊗ ωY ) is understood as the composition of functors Db lfgu(OF,Y ) ⊗ωY−−−−→ Db crys(QCrysκ(Y )) i! −→ Db crys(QCrysκ(X)) ⊗ω−1 X−−−−−→ Db lfgu(OF,X). Proof. Locally free left OF,Y -modules are in particular locally free as quasi-coherent OY -modules. Thus we have Rn(( ⊗ ω−1 X ) ◦ i! ◦ ( ⊗ ωY ))P ∼= (( ⊗ ω−1 ∼= ( ⊗ ω−1 ∼= 0 X ) ◦ Rni! ◦ ( ⊗ ωY )) ⊕j∈J OY X ) ◦ Rni! ⊕j∈J ωY locally for all n 6= −dX/Y on the underlying quasi-coherent sheaves. Theorem 2.32. For closed immersions i : X −→ Y of smooth, F -finite k-schemes, the equivalences Db lfgu(OF,Y ) of derived categories induced by GX and GY are compatible with the pull-backs i!, i.e. we have a canonical isomorphism crys(QCrysκ(X)) −→ Db lfgu(OF,X) and Db crys(QCrysκ(Y )) −→ Db GX ◦i! ∼= i! ◦ GY of functors from Db crys(QCrysκ(Y )) to Db lfgu(OF,X ). Proof. This is an application of the following general result concerning derived functors: Proposition 2.33 ([Har66, Proposition I.7.4]). Lat A and B be abelian categories, where A has enough injectives, and let F1 : A −→ B be an additive functor which has cohomo- logical dimension ≤ n on A. Let P be the set of objects X of A such that RiF1(X) = 0 for all i 6= n, and assume that every object of A is a quotient of an element of P . Let F2 = RnF1. Then RF1 and LF2 exist, and there is a functorial isomorphism RF1 ∼−→ LF2[−n]. 36 HomOY (i∗OX , First we have to check the requirements. Let F denote the functor ( ⊗ ω−1 X ) ◦ R0i♭ ◦ ( ⊗ ωY ). As done in the proof of Lemma 2.15, the derived functors of R0i♭ = ∗ ) may be computed locally by resolving i∗OX by the Koszul complex. i Since this complex has length −dX/Y , the cohomological dimension of F is smaller or equal −dX/Y . As Y is smooth, every left OF,Y -module is the quotient of a locally free left OF,Y -module ([EK04, Lemma 1.6.2]). Finally, for every locally free left OF,Y -module P , Lemma 2.31 states that RnF (P ) = 0 for all n 6= −dX/Y . It follows from [Har66, Proposition I.7.4] that RF ∼= Li∗[dX/Y ] = i! because i∗ ∼= R−dX/Y F (Theorem 2.14). Thus ( ⊗ ω−1 X ) ◦ i! ◦ ( ⊗ ωY ) ∼= i!, i.e. the diagram Db crys(QCrysκ(Y )) i! Db crys(QCrysκ(X)) ⊗ω−1 Y ⊗ωY ⊗ω−1 X ⊗ωX / Db lfgu(OF,Y ) i! / Db lfgu(OF,X) is commutative. Corollary 2.34. For every closed immersion of smooth, F -finite k-schemes i : X −→ Y , there is a canonical isomorphism GY ◦i∗ ∼= i∗ ◦ GX . Proof. This follows formally as GX and GY are equivalences of categories and since i∗ is uniquely determined as a left adjoint functor of i!. Proposition 2.35. Let j : X −→ Y be an open immersion of smooth, F -finite schemes. Then there are natural isomorphisms GX ◦j∗ ∼= j! ◦ GY and GY ◦Rj∗ ∼= j+ ◦ GX . Proof. We already have seen that (Proposition 2.19) and that GenX ◦j∗ ∼= j! ◦ GenY (see the proof of Corollary 2.30, this part holds for an arbitrary flat morphism of smooth k-schemes). Therefore GX ◦j∗ ∼= j! ◦ GY . The rest follows from the adjunction of Rj∗ or j+ and j∗ or j!. −1 ◦ j∗ ∼= j∗ ◦ ⊗ ω−1 Y ⊗ ωX Up to now, we have seen that the equivalence G between Cartier crystals and lfgu modules is compatible with the (derived) push-forward for open and closed immersions by showing the compatibility for the adjoint pull-back functors. In fact, G is compatible 37 / o o / o o with push-forward for arbitrary morphisms of smooth schemes, but we can give a proof only up to the following theorem2, which is a result of Lurie, see [Lur16, Theorem 1.3.3.2]. Theorem 2.36. Let F : D(A) −→ D(B) be a functor between derived categories of abelian categories A and B, which is a morphism of triangulated categories. If F lifts to an exact functor of the stable ∞-categories whose homotopy categories are the cohomo- logically bounded below derived categories D+(A) and D+(B), if F is t-left exact for the canonical t-structure, i.e. F maps D≥0(A) to D≥0(B), and if the cohomology of F (I) is concentrated in degree 0 for every injective object I of A, then F arises as a right derived functor between the abelian categories A and B. Proposition 2.37. Let f : X −→ Y be a morphism of smooth, F -finite k-schemes. There is a natural isomorphism GY ◦Rf∗ −→ f+ ◦ GX from Db crys(QCrysκ(X)) to Db lfgu(OF,Y ). Y ◦ Rf∗ ∼= Rf+ ◦ ⊗ ω−1 Proof. As ⊗ ω−1 X by construction, it suffices to show that there is a natural isomorphism of functors Gen Rf+ −→ f+ Gen from Db crys(QCrysγ(X)) to lfgu(OF,Y ). For every complex N • of γ-sheaves, the complex Gen N • of quasi-coherent Db unit OF,X-modules has a two-term resolution by induced modules, namely the short exact sequence 0 −→ OF,X ⊗OX N • 1−β′ −−−→ OF,X ⊗OX N • −→ Gen N • −→ 0 of [EK04, Proposition 5.3.3]. Here β′ : OF,X ⊗OX N • −→ OF,X ⊗OX N • denotes the morphism corresponding to β via the identification HomOF,X (OF,X ⊗OX A, OF,X ⊗OX B) ∼−→ HomOX (A, ⊕∞ n=0(F r X )∗B) for OX -modules A and B described in 1.7.3 of ibid. First we verify that the requirements of Theorem 2.36 are satisfied. Let I • be a bounded below complex of injective γ-sheaves with H i(I •) = 0 for i < 0. Let β : I • −→ F ∗ X I • be the morphism of complexes induced by the structural morphisms of the I i. The complex f+I • represents Rf+I • and, as explained above, we have a short exact sequence 0 −→ OF,X ⊗OX f+I • 1−f+β′ −−−−−→ OF,X ⊗OX f+I • −→ Gen f+I • −→ 0. Applying f+ to the two-term resolution of Gen I • yields a distinguished triangle f+(OF,X ⊗OX I •) f+(1−β′) −−−−−−→ f+(OF,X ⊗OX I •) −→ f+(Gen I •) −→ f+(OF,X ⊗OX I •)[1]. 2We will not discuss this theorem here as its theoretical background, for example ∞-categories, goes beyond the scope of this work. We just note that the requirement that f+ Gen lifts to a functor of the corresponding stable ∞-categories is satisfied, because f+ is a composition of left and right derived functors, which have this property ([Lur16, Example 1.3.3.4]). 38 The sheaf OF,Y ←X is locally free as an OX-module. It follows that locally OF,Y ←X ⊗OX I • is a direct sum of flasque sheaves and hence flasque. We have f+(OF,X ⊗OX I •) = Rf∗(OF,Y ←X L ⊗OF,X (OF,X ⊗OX I •)) ∼−→ Rf∗(OF,Y ←X ⊗OX I •) ∼−→ f∗(OF,Y ←X ⊗OX I •), see also [EK04, Lemma 3.5.1] and its proof. In particular, the complex f+(OF,X ⊗OX I •) is represented by the complex whose i-th degree equals the sheaf f+(OF,X ⊗OX I i). The canonical isomorphism OF,X ⊗OX f+I • ∼−→ f+(OF,X ⊗OX I •) of the proof of [EK04, Theorem 3.5.3] makes the left hand square of the diagram OF,X ⊗OX f+I • 1−f+β′ OF,X ⊗OX f+I • Gen f+I • (5) ∼ ∼ ∼ f+(OF,X ⊗OX I •) f+(1−β′) / f+(OF,X ⊗OX I •) / f+(Gen I •) commutative ([EK04, Proposition 3.6.1]). This shows that the cohomology sheaves of f+(Gen I •) vanish in negative degrees, i.e. f+ Gen is left t-exact for the canonical t- structure of the bounded derived category of γ-sheaves on X. Furthermore, for a single injective γ-sheaf I on X, the upper row of the commutative diagram OF,X ⊗OX f+I 1−f+β′ OF,X ⊗OX f+I ∼ ∼ Gen f+I ∼ f+(OF,X ⊗OX I) f+(1−β′) / f+(OF,X ⊗OX I) / f+(Gen I) is a short exact sequence when adding 0 at the ends. Consequently, the cohomology of f+ Gen I is concentrated in degree 0. To see that there is an isomorphism of functors Gen f+ ∼= H 0(f+) Gen, let M be a γ-sheaf on X. Choose a resolution I • of M by injective γ-sheaves. The long ex- act cohomology sequences for the triangles of the diagram Equation 5 yield a unique isomorphism Gen f+M ∼= H 0(Gen Rf+M ) = H 0(Gen f+I •) ∼−→ H 0(f+ Gen I •) ∼= H 0(f+ Gen M ). By Theorem 2.36, the functor f+ Gen is the right derived functor of H 0(f+) Gen. Fur- thermore, as Gen is exact, Gen Rf+ is the right derived functor of Gen f+. Thus, there ∼= f+ Gen of functors from the bounded derived cat- is a natural equivalence Gen Rf+ egory of γ-sheaves on X to the bounded derived category of quasi-coherent unit left OF,Y -modules. It induces an isomorphism of functors between Db crys(QCrysγ(X)) and Db lfgu(OF,Y ) because Db(µu(Y )) ∼−→ Db lfgu(OF,Y ) ([EK04, Corollary 17.2.5]). 39 / / / / / / / / / / / / 3 Locally finitely generated unit modules on singular schemes For a proper map f : X −→ Y of smooth k-schemes, Emerton and Kisin proved that there is a natural isomorphism RHom• OF,Y (f+M•, N •) ∼−→ Rf∗RHom• OF,X (M•, f !N •) qc(OF,X) and N • ∈ Db for M• ∈ Db qc(OF,Y ) ([EK04, Theorem 4.4.1]) by constructing a trace map acting as the counit of adjunction. We generalize this trace map to sep- arated and finite type morphisms f : X −→ Y between smooth k-schemes sitting in a commutative diagram i′ i Z ′ f ′ Z X f / Y, where i and i′ are closed immersions and f ′ is proper. This generalized trace map induces an adjunction between f+ and RΓZ ′f ! considered as functors between the derived categories Db lfgu(OF,Y )Z of complexes whose cohomology sheaves are supported in Z ′ or Z. lfgu(OF,X)Z ′ and Db The base for this more general trace for lfgu modules is a corresponding general- ized trace map trZ,f = trf : Rf∗RΓZ ′f ! −→ id for quasi-coherent sheaves established in [Sch18] in the situation of the diagram above. First, let us fix some notation: Let D− qc(OX )Z denote the subcategory of the derived category Dqc(OX ) of quasi-coherent sheaves on X whose objects have bounded above cohomology supported on Z and sim- ilar for D+ qc(OY )Z . The generalized trace has many compatibilities of the classical one, for example it behaves well with residually stable base change3. But most important, it gives rise to the following adjunction: Theorem 3.1 ([Sch18, Theorem 3.2]). Let f : X −→ Y be a separated and finite type morphism of Noetherian schemes and let i : Z −→ Y and i′ : Z ′ −→ X be closed immer- sions with a proper morphism f ′ : Z ′ −→ Z such that the diagram i′ i Z ′ f ′ Z X f / Y commutes. Then there is a natural transformation trf : Rf∗RΓZ ′f ! −→ id such that, for all F • ∈ D− qc(OY )Z, the composition qc(OX )Z and G• ∈ D+ Rf∗RHom• OX (F •, RΓZ ′f !G•) / RHom• OY (Rf∗F •, Rf∗RΓZ ′f !G•) trf RHom• OY (Rf∗F •, G•) 3Here a morphism f is called residually stable if it is flat, integral and the fibers of f are Gorenstein. 40 / / / / / / / is an isomorphism. In particular, taking global sections, the functor Rf∗ is left adjoint to the functor RΓZ ′f !. 3.1 Generalization of Emerton-Kisin's adjunction Proposition 3.2. Let f : X −→ Y be a separated and finite type morphism of smooth k-schemes and let i : Z −→ Y and i′ : Z ′ −→ X be closed immersions with a proper morphism f ′ : Z ′ −→ Z such that f ◦ i′ = i ◦ f ′. (a) There is a natural morphism trF,f : f+RΓZ ′OF,X[dX/Y ] −→ OF,Y of (OF,Y , OF,Y )-bimodules which, as a morphism of left OF,Y -modules, is the trace OF,Y ⊗OY Rf∗RΓZ ′ωX/Y [dX/Y ] −→ OF,Y of Theorem 3.1. (b) For every M• ∈ Db qc(OF,Y ), the trace map trF,f induces a morphism trF,f (M•) : f+RΓZ ′f !M• −→ M• in Db qc(OF,Y ). Proof. (a) This is an analogue of [EK04, Proposition 4.4.9 (i)]. A careful reading of the proof shows that we can adopt it. Consider the relative Frobenius diagram (diagram 1 on page 13): FX/Y X ❇❇ X′ ❇❇❇❇❇❇ f f ′ Y F ′ Y FY X f / Y. Since X and Y are assumed to be smooth k-schemes, we still have flatness of the Frobe- nius FY and therefore of F ′ Y because flatness is stable under base change. Note that FX/Y is finite ([EK04, A.2]). First Emerton and Kisin explain how the relative Cartier operator CX/Y : FX/Y ∗ωX/Y −→ F ′∗ Y ωX/Y is realized for the residual complex f ∆E•. Here E• denotes the Cousin complex E•(OX ). For our result we replace f ∆E• by the subcomplex ΓZ ′f ∆E• of flasque sheaves which computes RΓZ ′ωX/Y . We obtain the relative Cartier operator with support on Z ′: X/Y : FX/Y ∗ΓZ ′f ∆E• −→ F ′∗ C Z Y ΓZ ′f ∆E•. By Proposition-Definition 1.10.1 of [EK04], f −1OF,Y ⊗f −1OY ΓZ ′f ∆E• is equipped with a (f −1OF,Y , OF,X)-bimodule structure or, after restricting scalars via the natural map 41 / / / / / f −1OY [F ] −→ OX [F ], with a (f −1OF,Y , f −1OF,Y )-bimodule structure. Finally this en- dows f∗(f −1OF,Y ⊗f −1OY ΓZ ′f ∆E•) with the structure of a (OF,Y , OF,Y )-bimodule, the one from the definition of f+RΓZ ′OF,X. But there is another way to look at this bimodule: The map C Z X/Y gives rise to a morphism by the composition σ : f∗ΓZ ′f ∆E• −→ F ∗ Y f∗ΓZ ′f ∆E• f∗ΓZ ′f ∆E• ∼−→ f ′ ∗FX/Y ∗ΓZ ′f ∆E• CZ X/Y−−−→ f ′ ∗F ′∗ Y ΓZ ′f ∆E• bc−1 −→ F ∗ Y f∗ΓZ ′f ∆E•, where the first isomorphism is deduced from f = f ′ ◦ FX/Y and the last isomorphism is flat base change. Now Proposition-Definition 1.10.1 of ibid. in the special case of the morphism idY yields a (OF,Y , OF,Y )-bimodule structure on OF,Y ⊗OY f∗ΓZ ′f ∆E•. The isomorphism f∗(f −1OF,Y ⊗f −1OY ΓZ ′f ∆E•) ∼= OF,Y ⊗OY f∗ΓZ ′f ∆E• stemming from the projection formula is compatible with the constructed bimodule structure for both complexes by Lemma 1.10.6 of ibid. Hence it suffices to show that trF,f induces a morphism between the (OF,Y , OF,Y )-bimodule OF,Y ⊗OY f∗ΓZ ′f ∆E• and E• equipped with the structure of a (OF,Y , OF,Y )-bimodule via the canonical isomorphism E• ∼−→ F ∗ Y OY . Lemma 1.10.2 of ibid. applied to the identity morphism on Y reduces to the commutativity of the diagram Y E• induced from the Frobenius OY −→ F ∗ f∗ΓZ ′f ∆E• trF,f σ E• ∼ Y f∗ΓZ ′f ∆E• F ∗ F ∗ Y trF,f / F ∗ Y E• 42 / / / of complexes. For this we have to see that the following bigger diagram commutes: f∗ΓZ ′f ∆E• ∼ f ′ ∗FX/Y ∗ΓZ ′F ∆ X/Y f ′∆E• ∼ f ′ ∗FX/Y ∗ΓZ ′F ∆ X/Y f ′∆F ∗ Y E• β ∼ f ′ ∗FX/Y ∗ΓZ ′F ∆ X/Y F ′∗ Y f ∆E• trFX/Y trFX/Y trFX/Y trf trf ′ trf ′ ∗ΓZ ′f ′∆E• f ′ ∼ ∗ΓZ ′f ′∆F ∗ f ′ Y E• β ∼ E• E• ∼ Y E• F ∗ / f ′ ∗ΓZ ′F ′∗ Y f ∆E• ∼ f ′ ∗F ′∗ Y ΓZ ′f ∆E• bc−1 ∼ Y f∗ΓZ ′f ∆E• F ∗ trf / F ∗ Y E•. The two squares in the middle and the lower left square commute by functoriality of the trace maps trFX/Y and trf ′. The other squares are commutative because the general- ized trace is compatible with compositions of morphisms and with base change by the residully stable map FY ([Sch18, Propositions 2.10 and 2.11]). (b) Once we know that trf is a morphism in Db qc(OF,Y ), we can define trf (M•) as the following composition: f+RΓZ ′f !M• ∼−→ f+(OF,X ⊗OF,X RΓZ ′f !M•) ⊗OF,X f !M•) −→ f+(RΓZ ′OF,X L L ⊗OF,Y M• −→ f+RΓZ ′OF,X[dX/Y ] trf ⊗ id −−−−→ OF,Y ⊗OF,Y M• ∼−→ M•. Here the second morphism is the one of [Sch18, Lemma 1.11] and the third morphism is the one of [EK04, Lemma 4.4.7]. Lemma 3.3. We keep the notation of the preceding proposition. For an open immersion j : U −→ Y , let f ′ and j′ denote the projections of U ′ = U ×Y X. Assume that Z and Z ′ U ′ = Z ′ ∩ U ′ in Y and in are the closures of the locally closed subsets ZU = Z ∩ U and Z ′ X. 43 / / / / / / / / / / / / (a) There is a functorial isomorphism ej,f : f+RΓZf !j+ diagram ∼−→ j+f ′ +RΓZ ′f ′! such that the f+RΓZf !j+ ej,f j+f ′ +RΓZ ′f ′! 'PPPPPPPPPPPPPP trf j+ j+ trf ′ j+ commutes. (b) Let ctrf denote the unit id −→ RΓZf !f+ of the adjunction. Then there is a func- torial isomorphism e′ j,f : RΓZ f !f+j′ + ∼−→ j′ +RΓZ ′f ′!f ′ + such that the diagram RΓZf !f+j′ + +RΓZ ′f ′!f ′ / j′ + ej,f gPPPPPPPPPPPPPP ctrf j′ + j′ + ctrf ′ j′ + commutes. Proof. Let ZU and Z ′ [Sch18, Proposition 1.13] we know that the functors j+ and j′ U ′ denote the closed subsets U ∩ Z and U ′ ∩ Z ′ of U and U ′. From + are equivalences ∼−→ Db lfgu(OF,X)Z ′\U ′ Z ′ . Db lfgu(OF,U )ZU ∼−→ Db lfgu(OF,Y )Z\U Z and Db lfgu(OF,U ′)Z ′ U ′ Furthermore, there are natural isomorphisms f+j′ + ∼−→ j+f ′ + and RΓZ ′f !j+ ∼−→ j′ +RΓZ ′ U ′ f ′!, where the second one is obtained from the composition j′!RΓZ ′f ! ∼−→ RΓZ ′ U ′ j′!f ! ∼−→ RΓZ ′ U ′ f ′!j! of natural isomorphisms. Moreover, together with the canonical isomorphism f ′ j!f+ of [EK04, Proposition 3.8], this composition yields a canonical isomorphism +j′! ∼= For M• ∈ Db ej,f : f ′ qc(OF,Y ), the diagram +RΓZ ′f ′!j! −→ j!f+RΓZf !. +RΓZ ′f ′!j!M• f ′ ∼ ej,f j!f+RΓZf !M• ∼ +(OF,U ′ ⊗OF,U ′ RΓZ ′f ′!j!M•) ∼ f ′ j!f+(OF,X ⊗OF,X RΓZf !M•) ∼ ∼ +RΓZ ′f ′!OF,U ′ f ′ L ⊗OF,U j!M• ej,f j!f+RΓZf !OF,X L ⊗OF,U j!M• s❣❣❣❣❣❣❣❣❣❣❣❣❣❣❣❣❣❣❣❣❣❣❣❣❣ trf trf ′ j!M• 44 / / ' / g O O / / / / / / s of natural isomorphisms and the trace commutes: While the first square commutes simply by functoriality, the commutativity of the second square follows from [EK04, Lemma 4.4.7 (ii)]. The commutativity of the lower triangle follows from the compatibility of the trace with residually stable base change ([Sch18, Proposition 2.10]). In summary the diagram +RΓZ ′f ′!j! f ′ ej,f j!f+RΓZf ! w♦♦♦♦♦♦♦♦♦♦♦♦♦♦ j! trf trf ′ j! j! is commutative. Since j! and j′! are quasi-inverses of j+ and j′ + and f+ and RΓZ ′f ! restrict to the functors f ′ with respect to the equivalences j! and j′!, the claims of the lemma are formal consequences. lfgu(OF,U )ZU and Db lfgu(OF,U ′)Z ′ U ′ + and RΓZ ′ U ′ f ′! between Db Theorem 3.4. Let f : X −→ Y be a separated and finite type morphism of smooth schemes and let i : Z −→ Y and i′ : Z ′ −→ X be closed immersions with a morphism f ′ : Z ′ −→ Z such that the diagram i′ i Z ′ f ′ Z X f / Y commutes. Then, for any M• ∈ Db natural isomorphism qc(OF,X)Z ′ and any N • ∈ Db qc(OF,Y )Z, there is a RHom• OF,Y (f+M•, N •) ∼−→ Rf∗RHom• OF,X (M•, RΓZ ′f !N •). In particular, f+ : Db qc(OF,X)Z ′ −→ Db qc(OF,Y )Z is left adjoint to RΓZ ′f !. Proof. The morphism f factors through the graph morphism X ×k Y , which is a closed immersion, followed by the projection X ×k Y −→ Y , which is smooth. Therefore, we may assume that f is an essentially perfect morphism. We show that the natural transformation τ given by the composition Rf∗RHom• OF,X (M•, RΓZ ′f !N •) / RHom• OF,Y (f+M•, f+RΓZ ′f !N •) trF,f RHom• OF,Y (f+M•, N •) is an isomorphism in D+(X, Z/pZ). Here the horizontal arrow is the natural morphism of [EK04, Proposition 4.4.2]. Let OF,f denote the (f −1OF,Y , OF,X)-bimodule OF,Y ←X and let ωf denote the OX -module ωX/Y . We set d = dX/Y . First we replace M• by 45 / / w / / / / a bounded above complex of quasi-coherent induced left OF,X-modules, i.e. left OF,X- modules of the form OF,X ⊗OX M with quasi-coherent OX -modules M , see Definition 1.7 and Lemma 1.7.1 of [EK04]. Now by the Lemma on Way-out Functors ([Har66, Proposition I.7.1]), we reduce to the case of a single sheaf M• = OF,X ⊗OX M . For such an induced module we have an isomorphism f+M ∼−→ OF,Y ⊗OY Rf∗(ωX/Y ⊗OX M ), (6) which is based on the projection formula, see the proof of [EK04, Theorem 3.5.3]. Note that in this proof f ! always denotes Emerton-Kisin's pull-back of left OF,Y -modules, It is connected to the functor f ! for quasi- sometimes considered as an OY -module. coherent sheaves by the canonical isomorphisms f !N • ∼−→ Lf ∗N •[d] ∼−→ ω−1 X/Y ⊗OX 'f !'N • in Dqc(X), where 'f !' denotes the classical f !. One can show that there is a commutative diagram Rf∗RHom• OX (M, RΓZ ′f !N •) ∼ / Rf∗RHom• OF,X (M, RΓZ ′f !N •) t τ RHom• OY (Rf∗(ωX/Y ⊗OX M ), N •) ∼ / RHom• OF,Y (f+M, N •) with an isomorphism t and where the horizontal arrows are the natural isomorphisms induced by the isomorphism HomOX (M, ) ∼−→ HomOF,X (OF,X ⊗OX M, ) of [EK04, 1.7.2] and (6). For this we consider the bigger diagram of natural maps on page 49. Let t be the composition of the left vertical arrows. It is an isomorphism by [Sch18, Proposition 3.4] and Theorem 3.1. Recall that OF,Y ←X is locally free as a ⊗OF,X M ∼= OF,Y ←X ⊗OX M , which is computed right OX -module and that OF,Y ←X in the proof of Lemma 3.5.1 of [EK04]. In particular, induced modules are acyclic for the functor OF,Y ←X ⊗OF,X . For the first square, we consider the diagram without the outer Rf∗, resolve M by a complex P • of locally free OX-modules and RΓZ ′f !N • by a complex J • of left OX,F -modules which are acyclic for the functor OF,Y ←X , as in the proof of Proposition 4.4.2 of ibid. Now P• = OF,X ⊗OX P • is a complex of locally ⊗OF,X L L 46 / / free OF,X-modules. We obtain a commutative diagram RHom• OX (P •, J •) ∼ Hom• OX (P •, J •) ∼ ∼ ∼ RHom• OF,X (P•, J •) ∼ Hom• OF,X (P•, J •) Hom• OX (ωf ⊗OX P •, ωf ⊗OX J •) / Hom• f −1OF,Y (OF,f ⊗OF,X P•, OF,f L ⊗OF,X J •) ∼ RHom• OX (ωf ⊗OX P •, ωf ⊗OX J •) / RHom• f −1OF,Y (OF,f ⊗OF,X P•, OF,f L ⊗OF,X J •) of canonical maps. The last two vertical arrows are the canonical morphisms from a functor to its right derived functor. Here the left one is an isomorphism because ωX/Y ⊗ P • is a locally free OX -module. For the second square, we check that the natural map RHom• OX (ωf ⊗OX M, ωf ⊗OX RΓZ ′N •) RHom• f −1OF,Y (OF,f ⊗OX M, OF,f L ⊗OF,X RΓZ ′N •) L L factors through RHom• f −1OY (ωf ⊗OX M, ωf ⊗OX RΓZ ′N •). For this we replace ωf ⊗OX ⊗OF,X RΓZ ′f !N • by RΓZ ′f !N • by a complex I • of injective f −1OY -modules and OF,f a complex I • of injective f −1OF,Y -modules. The functor f −1OF,Y ⊗f −1OY is exact because the right OY -module OF,Y is free ([EK04, Lemma 1.3.1]). Furthermore, it is left adjoint to the forgetful functor from f −1OF,Y -modules to f −1OY -modules. Hence the latter functor preserves injectives. This implies that I • is a complex of injective f −1OY - ⊗OF,X RΓZ ′f !N • yields modules and the canonical morphism ωf ⊗OX RΓZ ′f !N • −→ OF,f a map I • −→ I •. After replacing M by a complex P • of locally free OX -modules as above we have reduced the three RHom to Hom and the claimed factorization is trivial. We return to the second square of the diagram on page 49, where we replace M by a complex F • of flasque OX -sheaves. The complexes ωf ⊗OX F • and OF,f ⊗OX F • are also flasque because locally they are direct sums of flasque sheaves. Hence f∗(ωf ⊗ F •) and f∗(OF,f ⊗ F •) represent Rf∗(ωf ⊗ F •) and Rf∗(OF,f ⊗ F •). As above, we resolve ⊗OF,X RΓZ ′f !N • by I •. The injectivity of I • and I • ωf ⊗OX RΓZ ′f !N • by I • and OF,f (OF,f ⊗OX F •, I •) are flasque implies that Hom• ([God58, Lemme II.7.3.2]) and hence may be used to compute Rf∗. As f∗ is right adjoint to the exact functor f −1, the complex f∗I • is a complex of injective OY -modules and (ωf ⊗OX F •, I •) and Hom• f −1OF,Y L f −1OY 47 / / / / / / f∗ I • is a complex of injective OF,Y -modules. Therefore and RHom• OX ( , f∗I •) ∼= Hom• OX ( , f∗I •) RHom• OF,X ( , f∗ I •) ∼= Hom• OF,X ( , f∗ I •). This finishes the proof of the commutativity of the second square because the diagram f∗Hom• f −1OY (ωf ⊗OX F •, I •) / f∗Hom• f −1OF,Y (OF,f ⊗OX F •, I •) Hom• OF,Y (f∗(OF,f ⊗OX F •), f∗ I •) Hom• OY (f∗(ωf ⊗OX F •), f∗I •) / Hom• OF,Y (OF,Y ⊗OY f∗(ωf ⊗OX F •), f∗ I •) of natural morphisms commutes. The commutativity of the third and the fifth square can be shown similarly. The fourth square commutes by the functoriality of the corresponding horizontal isomorphisms. For the adjunction of f+ and RΓZ ′f ! we proceed as in the proof of [Sch18, Theorem 3.2]. 48 / / Rf∗RHom• OX (M, RΓZ ′ f !N •) ∼ Rf∗RHom• OF,X (M, RΓZ ′f !N •) Rf∗RHom• OX (ωX/Y ⊗OX M, ωX/Y ⊗OX RΓZ ′ f !N •) Rf∗RHom• f −1OF,Y (OF,Y ←X L ⊗OF,X M, OF,Y ←X L ⊗OF,X RΓZ ′ f !N •) RHom• OF,Y (Rf∗(OF,Y ←X L ⊗ M), Rf∗(OF,Y ←X L ⊗ RΓZ ′f !N •)) RHom• OY (Rf∗(ωX/Y ⊗ M ), Rf∗(ωX/Y ⊗ RΓZ ′ f !N •)) RHom• OF,Y (OF,Y ⊗OY Rf∗(ωX/Y ⊗OX M ), Rf∗(OF,Y ←X L ⊗ RΓZ ′f !N •)) 4 9 RHom• OY (Rf∗(ωX/Y ⊗OX M ), Rf∗RΓZ ′ωX/Y [d] L ⊗OY N •) RHom• OF,Y (OF,Y ⊗OY Rf∗(ωX/Y ⊗OX M ), f+RΓZ ′ OF,X[d] L ⊗OF,Y N •) tr tr RHom• OY (Rf∗(ωX/Y ⊗OX M ), OY ⊗OY N •) RHom• OF,Y (OF,Y ⊗OY Rf∗(ωX/Y ⊗OX M ), OF,Y ⊗OF,Y N •) ∼ ∼ RHom• OY (Rf∗(ωX/Y ⊗OX M ), N •) / RHom• OF,Y (f+M, N •) / / / / / / / / / / / 3.2 Definition of lfgu modules on singular schemes As mentioned earlier, for a regular scheme X, the Frobenius FX : X −→ X is a flat morphism and hence F ∗ X is exact ([Kun69, Theorem 2.1]). For varieties, the exactness of F ∗ X plays an important role in the definition of (locally finitely generated) unit OF,X- modules. For example, it implies that the category of unit OF,X-modules is abelian. In this section we define the abelian category µlfgu(X) of locally finitely generated unit OF,X-modules for schemes X which admit a closed immersion i : X −→ Y into a smooth k-scheme as a certain subcategory of µlfgu(Y ). Note that this definition generally works for unit OF,X-modules. We restrict to locally finitely generated modules due to our application to Cartier crystals and perverse constructible étale p-torsion sheaves. For the motivation of our approach to µlfgu(X) for embeddable X, recall the Kashiwara equivalence: Theorem 3.5. Let i : Z −→ X be a closed immersion of smooth k-schemes. If M is a unit OF,X-module supported on Z, the adjunction i+i!M −→ M is an isomorphism. Consequently, H 0(i!M) ∼−→ i!M and the functors i+ and i! are equivalences between the categories of unit OF,Z-modules and unit OF,X-modules supported on Z. Proof. This is Theorem 5.10.1 of [EK04]. Hence, keeping the notation of the preceding theorem, we can canonically interpret unit OF,Z -modules as a certain subcategory of unit OF,X-modules, namely the subcat- egory of unit OF,X-modules with support on (the image of) Z. If Z is not smooth this subcategory still exists because it may be characterized as the subcategory of unit OF,X- modules M with j!M ∼= 0, where j is the immersion of the open complement of Z in X. This motivates the definition of unit OF,Z-modules for Z possibly not smooth but embeddable into a smooth scheme. But first we introduce some notation. Definition 3.6. We call a k-scheme X embeddable if there is a closed immersion i : X −→ Y of k-schemes where Y is smooth. Example 3.7. Let X = Spec k[x1, . . . , xn]/I be an affine variety. Then X is embed- dable into the affine space An k by the closed immersion corresponding to the canonical projection k[x1, . . . , xn] −→ k[x1, . . . , xn]/I. Example 3.8. Let X be a quasi-projective k-scheme. By definition, there exists an open immersion j : X −→ Z and a projective morphism p : Z −→ Spec k such that f = p ◦ j. In turn, the morphism p factors into a closed immersion i : Z −→ Pn k followed by the natural morphism Pn k such that U ∩ i(Z) = i(j(X)). Then X ∼= U ×Pk Z and the projection X −→ U is a closed immersion of X into an open subset of the projective space. Thus X is embeddable. k −→ Spec k. Let U be an open subset of Pn n Definition 3.9. Assume that k is perfect. Let X be an embeddable k-scheme. Let i : X −→ Y be a closed immersion into a smooth k-scheme Y . The category of lfgu OF,X-modules is defined as the full subcategory of lfgu OF,Y -modules M supported 50 on the image of X, i.e. j!M ∼= 0, where j : Y \X −→ X is the open immersion of the complement of X. The category Db lfgu(OF,Y ) whose cohomology sheaves are supported on X. lfgu(OF,X) is the full subcategory Db Db lfgu(OF,Y )X of those objects in Remark 3.10. With the notation of the preceding definition, let M be an lfgu module on Y . Whether M is supported in X only depends on the closed subset i(X) in Y . For ex- ample, the preceding definition does not distinguish between the categories Db lfgu(OF,X) and Db lfgu(OF,Xred), where Xred is the unique closed subscheme of X whose underlying topological space equals the one of X and which is reduced. By Theorem 3.5, it is clear that this definition generalizes the already existing notion of lfgu OF,X-modules for smooth X. Of course the crucial point is to see that the definition for not-necessarily smooth X is -- up to natural equivalence -- independent of a chosen embedding into a smooth scheme. Theorem 3.11. Assume that k is a perfect field. Let f : X −→ Y be a flat morphism between smooth k-schemes and let iX : Z −→ X and iY : Z −→ Y be closed immersions of k-schemes such that the diagram Z iX ❅❅❅❅❅❅❅❅ iY X f Y commutes. Then there are natural isomorphisms of functors (i) f+ ◦ RΓZf ! ∼= idDb lfgu(OF,Y )Z (ii) RΓZ f ! ◦ f+ ∼= idDb lfgu(OF,X )Z , . Proof. The proof proceeds by an excision argument, in a similar way as the proof of [Ohk16, Theorem 4.5]. In the case of a smooth scheme Z we can use the isomorphism of functors RΓZ Y from Db ∼= iX+i! lfgu(OF,Y )Z ([EK04, Proposition 5.11.5]): lfgu(OF,X)Z and RΓZ lfgu(OF,Y ) to Db X from Db lfgu(OF,X) to Db ∼= iY +i! f+RΓZf ! ∼= f+iX+i! Xf ! ∼= iY +i! Y ∼= RΓZ ∼= id . We may assume that Z is reduced, see Remark 3.10. Since a finite set of closed points with the reduced scheme structure is always smooth, this verifies the claim if Z is 0- dimensional. For the general case, i.e. Z is not necessarily smooth, let V be a smooth and dense4 open subscheme of Z and assume that the claim holds for all closed subschemes Z ′ with dim Z ′ < dim Z . Let g denote the immersion V ֒−→ Y . After choosing an open subset U ⊆ Y with U ∩ Z = V , we can factor g as g = u ◦ i′ where u is the open 4In order to guarantee the existence of a smooth, dense subset, we assumed that k is perfect. 51 / / immersion of U into Y and i′ is the closed immersion of V into U , i.e. the base change of iY . For an object M• of Db lfgu(OF,Y ), there is a natural morphism ϕ : M• −→ g+g!M• whose cone N • is supported on Z\U ([EK04, Proposition 5.12.1]). This means that there is a distinguished triangle N • −→ M• ϕ −→ g+g!M• −→ N •[1] in Db lfgu(OF,Y ). Applying f+RΓZ f !, the trace yields a morphism of triangles f+RΓZf !N • f+RΓZ f !M• trf (N •) trf (M•) N • / M• ϕ ϕ f+RΓZf !g+g!M• trf (g+g!M•) / g+g!M•. Since f is the identity on Z, i.e. iY = f ◦ iX, we have Z ∩ f −1(Z\U ) = Z\U . Therefore, RΓZf !N • ∼= RΓZ\U f !N • and trf (N •) factors through f+RΓZ\U f !N •. This means that the diagram f+RΓZf !N • trf,Z (N •) N • ∼ 8rrrrrrrrrrr trf,Z\U (N •) ∼ (◗◗◗◗◗◗◗◗◗◗◗◗◗ f+RΓZ\U f !N • is commutative. The dimension of the support Z\U of N • is less than that of Z as V is dense in Z. By induction hypothesis, trf,Z\U (N •) is an isomorphism and hence trf,Z (N •) is an isomorphism. It remains to show that trf (g+g!M•) ∼= trf (u+i′ Kashiwara equivalence, the object M• Let f ′ denote the projection U ×Y X −→ U . The map trf (u+M• +i′!u!M• of Db U := i′ +i′!u!M•) is an isomorphism. By the lfgu(OF,U ) is supported on V . U ) equals the composition f+RΓZf !u+M• U ∼−→ u+f ′ +RΓV f ′!M• U u+ trf ′ (M• −−−−−−−−→ u+M• U U ) (Lemma 3.3 (a)). Here the second map is an isomorphism because V is smooth. Conse- quently, the map trf (M•) is an isomorphism. This proves (i). The isomorphism of (ii) can be constructed similarly, using the unit of the adjunction between f+ and RΓZf ! (i.e. the cotrace) instead of the trace map, and applying Lemma 3.3 (b). The next corollary shows that the definition of Db lfgu(OF,X) for embeddable varieties X is independent of the chosen embedding. Corollary 3.12. If i1 : X −→ Y1 and i2 : X −→ Y2 are two embeddings of a k-scheme X into smooth k-schemes Y1 and Y2, where k is perfect, then there exists a natural equivalence Db lfgu(OF,Y1)X ∼−→ Db lfgu(OF,Y2)X . 52 / / / / / / / / ( 8 Proof. The universal property of Y1 ×k Y2 yields a morphism (i1, i2) : X −→ Y1 ×k Y2. It equals the composition X (id,id) −−−−→ X ×k X (i1,id) −−−−→ Y1 ×k X (id,i2) −−−−→ Y1 ×k Y2, where all maps are closed immersions, the first one because X is assumed to be separated over k. Hence (i1, i2) is a closed immersion. We obtain a commutative diagram X Y1 ×k Y2 i1 (i1,i2) 7♥♥♥♥♥♥♥♥♥♥♥♥♥♥♥ '❖❖❖❖❖❖❖❖❖❖❖❖❖❖❖ i2 Y1 p1 p2 Y2, where p1 and p2 are the projections. By Theorem 3.11, the compositions p2+RΓZp! p1+RΓZp! 2 are inverse equivalences between Db lfgu(OF,Y1)X and Db lfgu(OF,Y2)X. 1 and 4 The Riemann-Hilbert correspondence for Cartier crystals As its title suggests, one of the main results of Emerton and Kisins "The Riemann- Hilbert correspondence for unit F-crystals" ([EK04]) is a characteristic p-analogue of the Riemann-Hilbert correspondence for D-modules. More precisely, for a smooth k- scheme X, the authors construct inverse equivalences of categories Db lfgu(OF,X) Sol M / Db c(Xét, Z/pZ) . lfgu(OF,X)) ⊆ pD≥0 and Sol(D≥0 Furthermore, Sol(D≤0 lfgu(OF,X)) ⊆ pD≤0 where pD≥0 and pD≥0 are two subcategories of Db c(Xét, Z/pZ) defining the perverse t-structure of [Gab04]. Hence Sol establishes an equivalence between the hearts of the corresponding t-structures, namely the locally finitely generated unit OF,X-modules and the so-called perverse constructible p-torsion sheaves. Using this correspondence of Emerton and Kisin, we will establish a Riemann-Hilbert correspondence between Cartier crystals and perverse constructible étale Z/pZ-sheaves on a scheme which admits an embedding into a smooth scheme. For this we ex- tend the equivalences G : Db lfgu(OF,X) −→ Db c(Xét, Z/pZ) to singular varieties embeddable into a smooth variety. crys(QCrysκ(X)) −→ Db lfgu(OF,X) and Sol : Db 4.1 Review of Emerton and Kisin's Riemann-Hilbert correspondence Let Xét denote the small étale site of a scheme X. A reference for the étale topology is, for example, [Mil80, Chapter II]. A Z/pZ-sheaf on Xét is an étale sheaf of modules over the constant sheaf Z/pZ. Let Db c(Xét, Z/pZ) denote the derived category of complexes of Z/pZ-sheaves on Xét whose cohomology sheaves are constructible. 53 / / 7 ' O O / o o Definition 4.1. A sheaf L of Z/pZ-modules on Xét is called constructible if there is a stratification X = `i∈I Si such that the restrictions of L to the Si are locally constant sheaves of Z/pZ-modules for the étale topology with finite stalks. For x ∈ X, let ix : x −→ X be the inclusion, which is the composition of the inclusion of the closed point of Spec OXét,x followed by the canonical morphism Spec OXét,x −→ X. In [Gab04], Gabber showed that the two subcategories pD≤0 = {L• ∈ Db pD≥0 = {L• ∈ Db c(Xét, Z/pZ) H i(i∗ c(Xét, Z/pZ) H i(i! xL•) = 0 for i > − dim {x}}, xL•) = 0 for i < − dim {x}} c(Xét, Z/pZ). define a t-structure on Db Remark 4.2. Indeed, Gabber shows that these subcategories define a t-structure on the ambient category Db(Xét, Z/pZ). For a closed immersion i : Z −→ X and the open immersion j : U −→ X of the complement U of Z, it is obtained from the perverse t-structures on Db(Uét, Z/pZ) and Db(Zét, Z/pZ) by recollement: pD≤0 = {L• ∈ Db(Xét, Z/pZ) i∗L• ∈ pD≤0(Zét, Z/pZ) and j∗L• ∈ pD≤0(Uét, Z/pZ)}, pD≥0 = {L• ∈ Db(Xét, Z/pZ) i!L• ∈ pD≥0(Zét, Z/pZ) and j∗L• ∈ pD≥0(Uét, Z/pZ)}. This follows directly from the construction of the perverse t-structure on Db(Xét, Z/pZ). In this subsection let X be a smooth k-scheme. The Riemann-Hilbert correspondence c(Xét, Z/pZ) is realized in two steps: first passing to the between Db étale site and then applying a certain duality functor. lfgu(OF,X) and Db Theorem 4.3. (a) For every smooth k-scheme X, the functor Sol = RHom• OF,Xét ( ét, OXét )[dX ] : Db lfgu(OF,X) −→ Db c(Xét, Fp) is an equivalence of categories. A quasi-inverse is given by M = RHom• Z/pZ( , OXét )[dX ]. (b) For a morphism f : X −→ Y of smooth k-schemes, there is a natural isomorphism of functors Sol ◦f ! ∼= f ∗ ◦ Sol . For an allowable morphism f : X −→ Y , i.e. a morphism f which factors as g ◦ h, where h is an immersion and g is a proper smooth morphism, there is also a natural isomorphism of functors Sol ◦f+ ∼= f! ◦ Sol . (c) The essential image of the full subcategory D≥0 c(Xét, Z/pZ) while the essential image of D≤0 lfgu(OF,X) is equal to the full subcat- lfgu(OF,X) is equal to egory pD≤0 of Db the full subcategory pD≥0 of Db c(Xét, Z/pZ). Proof. This is [EK04, Theorem 11.4.2 and Theorem 11.5.4]. 54 4.2 Cartier crystals and lfgu modules on singular schemes We show that the equivalence G : Db crys(QCrysκ(X)) ∼−→ Db lfgu(OF,X ) for smooth X extends to an equivalence for embeddable X. As a consequence, for a morphism f between smooth schemes, the inverse equivalences f+ and RΓZ f ! between the subcategories of complexes supported on a closed subscheme are t-exact. Proposition 4.4. Let k be a perfect field and let X be an embeddable k-scheme. The functor G induces an equivalence of categories G : Db crys(QCrysκ(X)) −→ Db lfgu(OF,X). Proof. Choose a closed immersion i : X −→ Y into a smooth k-scheme Y and let j denote the open immersion of the complement of X in Y . The Kashiwara equivalence (Theorem 1.10) identifies Db crys(QCrysκ(Y ))X of Db crys(QCrysκ(X)) with the subcategory Db crys(QCrysκ(Y )). For M• ∈ Db crys(QCrysκ(Y ))X we have (j! ◦ GY )M• ∼= (GU ◦j∗)M• ∼= 0 by Proposition 2.35. As G is an equivalence of categories, there is also a natural isomor- phism of functors j∗ ◦ G−1 U ◦j! for the inverse G−1 of G. It follows that G induces Y an equivalence of subcategories ∼= G−1 G : Db crys(QCrysκ(Y ))X −→ Db lfgu(OF,Y )X. It remains to show that this equivalence is independent of the choice of the embedding. In the same way as in the proof of Corollary 3.12 we can reduce to the case of two closed immersions i1 : X −→ Y1 and i2 : X −→ Y2 into smooth k-schemes Y1 and Y2 together with a morphism f : Y1 −→ Y2 such that i2 = f ◦ i1. The composition i2∗ ◦ i! 1 is a natural equivalence between Db crys(QCrysκ(Y1))X and Db crys(QCrysκ(Y2))X . Note that i2∗i! 1M• ∼= Rf∗i1∗i! 1M• ∼= Rf∗M• for M• ∈ Db crys(QCrysκ(Y1))X . Hence Rf∗ is a natural equivalence of categories Rf∗ : Db crys(QCrysκ(Y1))X −→ Db crys(QCrysκ(Y2))X which is compatible with G, i.e. by Proposition 2.37. f+ ◦ GY1 ∼= GY2 ◦Rf∗ Remark 4.5. This also implies that f+ provides a natural equivalence f+ : Db lfgu(OF,Y1)X −→ Db lfgu(OF,Y2)X because f+ ∼= GY2 ◦Rf∗ ◦ G−1 Y1 . 55 Keeping the notation of the proof of Proposition 4.4, the canonical t-structure of lfgu(OF,Y ) obviously induces a t-structure on the subcategory Db lfgu(OF,Y )X defined by Db the two subcategories Db lfgu(OF,Y )X ∩ D≥0 lfgu(OF,Y ) and Db lfgu(OF,Y )X ∩ D≤0 lfgu(OF,Y ). Corollary 4.6. Let f : Y1 −→ Y2 be a morphism between smooth schemes over a perfect field k. Let i1 : X −→ Y1 and i2 : X −→ Y2 be closed immersions such that i2 = f ◦ i1. The equivalence f+ of Corollary 3.12 between Db lfgu(OF,Y2)X is t- exact for the canonical t-structures of both derived categories. In particular, by taking 0-th cohomology, it gives rise to an equivalence of abelian categories lfgu(OF,Y1)X and Db {µlfgu(Y1)X } ∼−→ {µlfgu(Y2)X }. Proof. The functor f+ is a composition of t-exact functors: ∼= GY2 ◦i2∗ ◦ i! ∼= GY2 ◦Rf∗ ◦ G−1 Y1 f+ 1 ◦ G−1 Y1 , where Rf∗ denotes the restricted functor from Db It is exact because Rf∗ ∼= Rf∗i1∗i! 1 ∼= i2∗i! 1. crys(QCrysκ(Y1))X to Db crys(QCrysκ(Y2))X . 4.3 A Riemann-Hilbert correspondence on singular schemes Now we extend the Riemann-Hilbert correspondence between lfgu modules and con- structible étale Z/pZ-sheaves to embeddable schemes. The corresponding equivalence of categories Db lfgu(OF,X) ∼−→ Db c(Xét, Z/pZ) for embeddable X will be t-exact for the canonical t-structure on Db ber's perverse t-structure on Db j : U −→ X denote the open immersion of the complement of Z into X. lfgu(OF,X) and Gab- c(Xét, Z/pZ). Again, for a closed subscheme Z of X, let Recall that there are distinguished triangles and j!j∗ −→ id −→ i∗i∗ −→ j!j∗[1] i∗i! −→ id −→ j∗j∗ −→ i∗i![1] in D+(Xét, Z/pZ) ([BBD82, 1.4.1.1]). Defining ΓZ : D(Xét, Z/pZ) −→ D(Xét, Z/pZ) as the composition i∗i∗ of exact functors we obtain a fundamental triangle of local coho- mology j!j∗ −→ id −→ ΓZ −→ j!j∗[1]. Note that i! = i∗ because i is a closed immersion. Lemma 4.7. Let Z be a closed subscheme of a smooth k-scheme X. Then there is a natural isomorphism of functors Sol ◦RΓZ ∼−→ ΓZ ◦ Sol . 56 Proof. We show that there is a natural isomorphism M ◦ΓZ ∼−→ RΓZ ◦ M, ∼= where M is the quasi-inverse of Sol, see Theorem 4.3. The natural isomorphism j! ◦MX MU ◦j∗ implies that M(ΓZL•) is supported on Z for every complex L•. Consequently, the morphism M(ΓZL•) −→ M(L•) induced by the natural map L• −→ ΓZL•, which is defined by the fundamental triangle of local cohomology above, factors through RΓZ M(L•). This gives rise to a morphism of distinguished triangles M(ΓZL•) M(L•) / M (j!j∗L•) ∼ RΓZ M(L•) / M(L•) / j+j! M(L•), where the horizontal arrows are the natural morphisms and the second and third vertical arrow is an isomorphism. Hence the vertical arrow on the left is an isomorphism. Lemma 4.8. For a closed subscheme Z of a scheme X, Gabber's perverse t-structure on Db c(Xét, Z/pZ) induces a t-structure on Db c(Xét, Z/pZ)Z given by pD≥0(Xét)Z = Db pD≤0(Xét)Z = Db c(Xét, Z/pZ)Z ∩ pD≥0(Xét), c(Xét, Z/pZ)Z ∩ pD≤0(Xét). Proof. We consider the construction of the perverse truncation functor pτ≤0 in [Gab04] in more detail. It will turn out that pτ≤0L• is supported on Z for all L• ∈ Db c(Xét, Z/pZ)Z . For simplicity we write τ≤p for pτ≤0 where p is a perversity function, see the first section of [Gab04]. For a complex F •, C(F •) denotes the total complex of the double complex C •(F •), where C •(F n) is the Godement resolution of F n. Let c = − dim X. It is a lower bound for the perversity function p(x) = − dim {x}. For a complex L•, d ≥ c and pd(x) = min(d, p(x)), Gabber iteratively constructs a direct system τ≤pdL• and defines τ≤pL• as the direct limit. We start with pc = c and the usual truncation τpcL• = τ≤cL•. Clearly, if L• is supported on Z, then so is τpcL•. Now for τ≤pdF • of some complex F •, we construct τ≤pd+1F • as a subcomplex of C(F •). By the construction of the Godement resolution, C(F •) is supported on Z if F • is supported on Z. It follows that for every d ≥ c, the complex τ≤pdL• is supported on Z and therefore the direct limit τ≤pL• is supported on Z. Proposition 4.9. Let i : Z −→ X be a closed immersion of schemes. (a) The exact functors i∗ and i∗ are inverse equivalences of categories Db c(Zét, Z/pZ) i∗ i∗ / Db c(Xét, Z/pZ)Z . (b) These functors i∗ : Db c(Zét, Z/pZ) −→ Db c(Xét, Z/pZ)Z −→ c(Zét, Z/pZ) are also t-exact with respect to the perverse t-structures of both c(Xét, Z/pZ)Z and i∗ : Db Db categories. 57 / / / / / / o o Proof. (a) This is a formal consequence of the distinguished triangle j!j∗ −→ id −→ i∗i∗ −→ j!j∗[1] in Db(Xét, Z/pZ) and the fact that the natural map id −→ i∗i∗ is always an isomorphism. (b) It suffices to show that i∗ is t-exact with respect to the perverse t-structures, i.e. the essential image of pD≤0(Zét) under i∗ is contained in pD≤0 and the essential image of pD≥0(Zét) under i∗ is contained in pD≥0. For x ∈ Z let ix denote the composition {x} −→ Spec OZét,x −→ Z of canonical morphisms. For L• in pD≤0(Zét), i.e. H n(i∗ every n > − dim {x}, we have xL•) = 0 for every x ∈ Z and H n(i∗ xi∗L•) ∼= H n(i∗ ∼= H n(i∗ ∼= 0 xi∗i∗L•) xL•) xi∗L• ∼= 0 for every x ∈ Z and every n > − dim {x}. For x ∈ U = X\Z, we even have i∗ because j∗i∗L• ∼= 0 and hence τ ∗i∗L• ∼= 0, where τ : Spec OXét,x −→ X is the natural morphism. Now let L• be in pD≥0(Zét), i.e. H n(i! xL•) = 0 for every x ∈ Z and every n < − dim {x}. There is a natural isomorphism of functors i!i∗ ∼= i∗i∗ given by the composition of the natural isomorphisms i!i∗ −→ id and id −→ i∗i∗ of [BBD82, 1.4.1.2]. Whence H n(i! xi∗L•) ∼= H n(i! ∼= H n(i! ∼= H n(i! ∼= 0 xi!i∗L•) xi∗i∗L•) xL•) for every x ∈ Z and every n < − dim {x}. We have already seen that there is nothing to show for x ∈ U . Definition 4.10. For a perfect field k and a k-scheme X which admits an embedding into a smooth k-scheme Y , we define D≥0 D≤0 lfgu(OF,X) = Db lfgu(OF,X) = Db lfgu(OF,Y )X ∩ D≥0 lfgu(OF,Y )X ∩ D≤0 lfgu(OF,Y ), lfgu(OF,Y ). These subcategories of Db lfgu(OF,X ) form a natural t-structure. 58 The independence of these subcategories of the embedding into a smooth scheme follows from the fact that for a morphism f : Y1 −→ Y2 between two smooth schemes over k, together with closed immersions i1 : X −→ Y1 and i2 : X −→ Y2 with i2 = f ◦ i1, the equivalence RΓXf ! ∼= M ◦ Solκ ◦RΓXf ! ∼= M ◦ΓXf ∗ ◦ Solκ ∼= M ◦i1∗i∗ 1f ∗i2∗i∗ ∼= M ◦i1∗i∗ 2 ◦ Solκ 2 ◦ Solκ is a composition of t-exact functors, where Db c(Y2,ét, Z/pZ) are equipped with the perverse t-structures (Theorem 4.3 and Proposition 4.9). Therefore, RΓXf ! is t-exact. Theorem 4.11. Let k be a perfect field and X an embeddable k-scheme. c(Y1,ét, Z/pZ) and Db (a) The equivalence Sol for smooth schemes induces an anti-equivalence of categories Sol : Db lfgu(OF,X) −→ Db c(Xét, Z/pZ) (b) The essential image of D≥0 lfgu(OF,X) under this equivalence equals pD≤0 and the essential image of D≤0 lfgu(OF,X) equals pD≥0. Proof. After choosing a closed immersion i : X −→ Y into a smooth k-scheme Y , we see that Sol restricts to an anti-equivalence Sol : Db lfgu(OF,Y )X −→ Db c(Yét, Z/pZ)X in the same way as in the proof of Proposition 4.4. For the proof of the independence of the choice of an embedding we again reduce to the situation of two closed immersions i1 : X −→ Y1 and i2 : X −→ Y2 together with a morphism f : Y1 −→ Y2 such that i2 = f ◦i1. We obtain natural equivalences of categories RΓXf ! : Db lfgu(OF,Y2)X −→ Db lfgu(OF,Y1)X and f −1ΓX : Db c(Y2,ét, Z/pZ)X −→ Db c(Y1,ét, Z/pZ)X , which are compatible with Sol because Sol ◦RΓX f ! ∼−→ ΓX ◦ Sol ◦f ! ∼−→ ΓXf ∗ ◦ Sol by Theorem 4.3 and Lemma 4.7. This proves (a). The equivalence Sol not only restricts to an equivalence between Db c(Yét, Z/pZ)X but also between D≥0 lfgu(OF,Y )X and lfgu(OF,Y ) and pD≤0(Yét) (Theorem 4.3). Therefore, Db Sol induces an equivalence Db lfgu(OF,Y )X ∩ D≥0 lfgu(OF,Y ) ∼−→ Db c(Yét, Z/pZ)X ∩ pD≤0(Yét). By Proposition 4.9, Db c(Yét, Z/pZ)X ∩ pD≤0(Yét) is canonically equivalent to pD≤0(Xét). Analogously, one can show that the essential image of D≤0 lfgu(OF,X) equals pD≥0. 59 Theorem 4.12. Let X be an embeddable k-scheme and assume that k is perfect. (a) The equivalences G and Sol for smooth schemes induce equivalences G : Db crys(QCrysκ(X)) −→ Db lfgu(OF,X ) and Sol : Db lfgu(OF,X) −→ Db c(Xét, Z/pZ). This means that for a closed immersion i : X −→ Y with Y smooth, there is a commutative diagram Db c(Yét, Z/pZ) Db c(Xét, Z/pZ) SolY Db crys(QCrysκ(Y )) GY / Db lfgu(OF,Y ) SolX h◗◗◗◗◗◗◗◗◗◗◗◗ Db crys(QCrysκ(X)) GX / Db lfgu(OF,X). (b) Let Solκ be the composition Sol ◦ G. The essential image of D≥0 der Solκ equals the subcategory pD≤0, while the essential image of D≤0 equals the subcategory pD≥0. crys(QCrysκ(X)) un- crys(QCrysκ(X)) (c) If h : W −→ X is an open or a closed immersion, then there are natural isomor- phisms of functors Solκ ◦Rh∗ ∼−→ h! ◦ Solκ and Solκ ◦h! ∼−→ h∗ ◦ Solκ where h! denotes the functor h∗ if h is an open immersion and Rh∗ = h∗ for a closed immersion. Proof. (a) and (b) follow from Proposition 4.4 and Theorem 4.11. It remains to prove (c). Let h : W −→ X be an open immersion. We will construct the natural transformations by choosing embeddings of X. Since we have to make sure that this construction is independent of the embedding, we will consider two closed immersions i1 : X −→ Y1 and i2 : X −→ Y2 of X into smooth k-schemes ab initio. We may assume that there is a morphism f : Y1 −→ Y2 with i2 = f ◦ i1, see the proof of Corollary 3.12. Let h′ 2 : V2 −→ Y2 be an open immersion such that (i2 ◦ h)(W ) = 2(V2) ∩ i2(X) and hence W ∼= X ×Y2 V2. Let i′ h′ 2 : W −→ V2 be the closed immersion in- ∼= duced by i2, i.e. the projection X ×Y2 V2 −→ V2. We have a natural equivalence Rh′ i2∗Rh∗ of functors from Db crys(QCrysκ(Y2))X . Moreover, we have a natural equivalence h′ crys(QCrysκ(V2))W to Db crys(QCrysκ(X)) induces a lfgu(OF,X) such that GX h∗ ∼= h+ GW . Note that this functor h+ : Db In particular, to show the inde- functor h+ does not depend on the choice of V2. pendence of the embedding of X, we may choose an open immersion h′ 1 : V1 −→ Y1 with (i1 ◦ h)(W ) = h′ 2(V2). Here we set 1(V1) ∩ i1(X) and such that (f ◦ h′ lfgu(OF,Y2)X. Therefore, Rh∗ : Db crys(QCrysκ(W )) −→ Db 2∗ of functors from Db crys(QCrysκ(W )) to Db lfgu(OF,W ) −→ Db 2+ GV2 ∼= GY2 Rh′ 1)(V1) ⊆ h′ 2∗i′ 2∗ 60 ? _ o / O O ? O / 5 U O O 1 be the projection of W ∼= X ×Y2 V2 ∼= Y1 ×Y2 V2. Let f ′ : V1 −→ V2 be the projection and V1 = f −1(V2), which means V1 ∼= Y1 ×Y2 V2. It let i′ is a closed immersion because it is the base change of the morphism i1. We obtain the following commutative diagram: ∼= X ×Y1 (Y1 ×Y2 V2) to V1 W h X i′ 1 i1 f ′ f / V1 h′ 1 / Y1 / V2 h′ 2 / Y2. The following cube demonstrates the natural equivalences, which we have by Proposition 2.37: Db crys(QCrysκ(V2))W Rf ′ ∗ 5❦❦❦❦❦❦❦❦❦❦❦❦❦❦❦ Db crys(QCrysκ(V1))W GV1 Rh′ 2∗ Rh′ 1∗ Db crys(QCrysκ(Y2))X Rf∗ 5❦❦❦❦❦❦❦❦❦❦❦❦❦❦❦ GY1 Db crys(QCrysκ(Y1))X GV2 Db lfgu(OF,V2)W f ′ + 7♦♦♦♦♦♦♦♦♦♦♦ Db lfgu(OF,V1)W h′ 2+ GY2 / Db lfgu(OF,Y2)X h′ 1+ f+ 7♦♦♦♦♦♦♦♦♦♦♦ Db lfgu(OF,Y1)X ∼= h′ Here every cube face indicates a natural equivalence, for example, the front refers to the isomorphism of functors GY1 ◦Rh′ 1+ ◦ GV1. This shows the independence of 1∗ the isomorphism of functors GX h∗ ∼= h+ GW from the chosen embedding of X. By adjunction, we obtain a canonical isomorphism GW h∗ ∼= h!GX as well. Similarly, one shows that we have a natural isomorphism Sol h! ∼= h∗ Sol, using the fact that Sol commutes with the local cohomology functors (Lemma 4.7) and with pull- backs for morphisms between smooth schemes (Theorem 4.3). Again, by adjunction, ∼= h! Sol. Composing these isomorphisms of we obtain a natural isomorphism Sol h+ functors yields the desired one: Solκ ◦Rh∗ ∼= Sol ◦ GX ◦Rh∗ ∼= Sol ◦h+ ◦ GW ∼= h! ◦ Sol ◦ GW ∼= h! ◦ Solκ and analogously for Solκ ◦h∗ ∼= h∗ ◦ Solκ. If h : W −→ X is a closed immersion, we proceed similarly, but the proof is simpler because a closed immersion i : X −→ Y of X into a smooth k-scheme Y yields a closed immersion of W into Y by composing h and i. Definition 4.13. The abelian category Pervc(Xét, Z/pZ) of perverse constructible étale p-torsion sheaves is the heart pD≤0 ∩ pD≥0 of the perverse t-structure on Db c(Xét, Z/pZ). 61 / / / / / / / / 5 7 / / / 5 7 Corollary 4.14. For a perfect field k and an embeddable k-scheme X, the functor Solκ induces an anti-equivalence Crysκ(X) −→ Pervc(Xét, Z/pZ) between the abelian categories of Cartier crystals on X and perverse constructible Z/pZ- sheaves on Xét. References [BB] M. Blickle and G. Böckle, Cartier crystals and unit OF,X-modules, uncircu- lated preprint. [BB11] M. Blickle and G. Böckle, Cartier modules: Finiteness results, J. Reine Angew. Math. 661 (2011), 85 -- 123. [BB13] M. Blickle and G. Böckle, Cartier Crystals, arXiv:1309.1035 (2013). [BBD82] A. A. Beılinson, J. Bernstein, and P. Deligne, Faisceaux pervers, Astérisque, vol. 100, Soc. Math. France, Paris, 1982. [Ber] J. Bernstein, Algebraic Theory of D-modules, unpublished notes. [Ber96] P. Berthelot, D-modules arithmétiques. I. Opérateurs différentiels de niveau fini, Ann. Sci. École Norm. Sup. (4) 29 (1996), no. 2, 185 -- 272. [Ber00] , D-modules arithmétiques. II. Descente par Frobenius, Mém. Soc. Math. Fr. (N.S.) (2000), no. 81, vi+136. [Bli03] M. Blickle, The D-Module structure of R[F ]-modules, Trans. Amer. Math. Soc. 355 (2003), no. 4, 1647 -- 1668. [Del70] P. Deligne, Équations différentielles à points singuliers réguliers, Lecture Notes in Mathematics, Vol. 163, Springer-Verlag, Berlin-New York, 1970. [EK04] M. Emerton and M. Kisin, Riemann -- Hilbert correspondence for unit F- crystals, Astérisque 293 (2004), vi+257 pp. [Gab62] P. Gabriel, Des catégories abéliennes, Bull. Soc. Math. France 90 (1962), 323 -- 448. [Gab04] O. Gabber, Notes on some t-structures, Geometric aspects of Dwork theory. Vol. I, II, Walter de Gruyter GmbH & Co. KG, Berlin, 2004, pp. 711 -- 734. [God58] R. Godement, Topologie algébrique et théorie des faisceaux, Actualit'es Sci. Ind. No. 1252. Publ. Math. Univ. Strasbourg. No. 13, Hermann, Paris, 1958. 62 [Har66] R. Hartshorne, Residues and Duality, Lecture notes of a seminar on the work of A. Grothendieck, given at Harvard 1963/64. With an appendix by P. Deligne. Lecture Notes in Mathematics, No. 20, Springer-Verlag, Berlin, 1966. [Har77] , Algebraic Geometry, Springer-Verlag, New York, 1977, Graduate Texts in Mathematics, No. 52. [Kas80] M. Kashiwara, Faisceaux constructibles et systèmes holonômes d'équations aux dérivées partielles linéaires à points singuliers réguliers, Séminaire Goulaouic- Schwartz, 1979 -- 1980 (French), École Polytech., Palaiseau, 1980, pp. Exp. No. 19, 7. [Kas84] , The Riemann-Hilbert Problem for Holonomic Systems, Publ. Res. Inst. Math. Sci. 20 (1984), no. 2, 319 -- 365. [Kat73] N. M. Katz, p-adic properties of modular schemes and modular forms, Modular functions of one variable, III (Proc. Internat. Summer School, Univ. Antwerp, Antwerp, 1972), Springer, Berlin, 1973, pp. 69 -- 190. Lecture Notes in Mathe- matics, Vol. 350. [Kun69] E. Kunz, Characterizations of Regular Local Rings of Characteristic p, Amer. J. Math. 91 (1969), 772 -- 784. [Lur16] J. Lurie, Higher Algebra, Available at: ~lurie/papers/HA.pdf [Accessed 7 August 2016] (2016). http://www.math.harvard.edu/ [Meb84a] Z. Mebkhout, Une autre équivalence de catégories, Compositio Math. 51 (1984), no. 1, 63 -- 88. [Meb84b] , Une équivalence de catégories, Compositio Math. 51 (1984), no. 1, 51 -- 62. [Mil80] J. S. Milne, Étale Cohomology, Princeton Mathematical Series, vol. 33, Prince- ton University Press, Princeton, N.J., 1980. [Ohk16] S. Ohkawa, Riemann-Hilbert correspondence for unit F-crystals on embeddable algebraic varieties, arXiv:1601.01525 (2016). [Sch18] Schedlmeier, Grothendieck T. arXiv:1810.06082 (2018). duality for non-proper morphisms, [Wei94] C. A. Weibel, An introduction to homological algebra, Cambridge Studies in Advanced Mathematics, vol. 38, Cambridge University Press, Cambridge, 1994. 63
0908.3281	3	0908	2010-09-23T15:43:48	A Derived Equivalence For A Del Pezzo Surface Of Degree 6 Over An Arbitrary Field	[ "math.AG", "math.CT" ]	Let $S$ be a degree six del Pezzo surface over an arbitrary field $F$. Motivated by the first author's classification of all such $S$ up to isomorphism in terms of a separable $F$-algebra $B \times Q \times F$, and by his K-theory isomorphism $K_n(S) \cong K_n(B \times Q \times F)$ for $n \ge 0$, we prove an equivalence of derived categories $$ \sD^b(\coh S) \equiv \sD^b(\mod A) $$ where $A$ is an explicitly given finite dimensional $F$-algebra whose semisimple part is $B \times Q \times F$. Submitted to the Journal of K-theory	math.AG	math	A DERIVED EQUIVALENCE FOR A DEGREE 6 DEL PEZZO SURFACE OVER AN ARBITRARY FIELD M. BLUNK, S.J. SIERRA, AND S. PAUL SMITH Abstract. Let S be a degree six del Pezzo surface over an arbitrary field F . Motivated by the first author's classification of all such S up to isomorphism [3] in terms of a separable F -algebra B × Q × F , and by his K-theory isomorphism Kn(S) ∼= Kn(B × Q × F ) for n ≥ 0, we prove an equivalence of derived categories Db(cohS) ≡ Db(modA) where A is an explicitly given finite dimensional F -algebra whose semisimple part is B × Q × F . 1. Introduction We will work over an arbitrary field F . Throughout S denotes a degree six del Pezzo surface over F . Equivalently, S is a smooth projective surface over F whose anti-canonical sheaf is ample and has self-intersection number 6. Throughout ¯F will denote a separable closure of F and we will write ¯S = S ¯F = S ×Spec F Spec ¯F . In [3], the first author classified such S up to isomorphism by associating to S a pair of separable F -algebras B and Q, both defined as endomorphism rings of certain locally free sheaves on S. Furthermore, it was shown there that the algebraic K-theory of S is isomorphic to that of the algebra B × Q × F . Let cohS denote the category of coherent sheaves on S and let modA denote the category of noetherian right A-modules. Let ≡ denote equivalence of derived categories. Our main result (Theorem 4.5) establishes a derived equivalence (1-1) Db(cohS) ≡ Db(modA) where A is a finite dimensional F -algebra whose semi-simple quotient is B × Q × F . We prove this equivalence by constructing a tilting bundle T on S that has A as its endomorphism ring. (The definition of a tilting bundle is given in section 4.) The main novelty of our approach is that we do not make any assumptions on the base field F . Since the field F is arbitrary, we cannot assume that S is obtained by blowing up P2 F (in fact S could be a minimal surface), nor can we use exceptional collections. 1991 Mathematics Subject Classification. 116E35, 14F05, 14J26, 14J60. Key words and phrases. Derived equivalence, del Pezzo surface, arbitrary field. M. Blunk was supported by the National Science Foundation, Award No. 0902659. S.J. Sierra was supported by the National Science Foundation, Award No. 0802935. S.P. Smith was supported by the National Science Foundation, Award No. 0602347. 1 2 M. BLUNK, S.J. SIERRA, AND S. PAUL SMITH Acknowledgments. All three authors acknowledge the support of the National Science Foundation with gratitude. We would also like to thank the referee for their comments. This research was partially done while the first author was visiting the University of Washington, and he would like to thank the institution for its support and excellent working conditions. Finally, the first author would like to thank Aravind Asok, Baptise Calm`es, and Daniel Krashen for suggesting this problem. 2. Basic facts about ¯S In this section, we give basic facts about the degree 6 del Pezzo surface ¯S. Since all the results here are well-known, we do not give references. There are six (−1)-curves on ¯S, which we may take to lie in the following con- figuration: (2-1) M2 L1 ???????????????????????? M1 L3 M3 ???????????????????????? L2 The Picard group is Pic ¯S ∼= i=1(ZLi ⊕ ZMi) . (Mi + Lj = Mj + Li 1 ≤ i, j ≤ 3) L3 Usually we only care about the class of a divisor in Pic ¯S. We will write D1 ∼ D2 if D1 and D2 are linearly equivalent divisors. As remarked in the discussion after Prop. 2.1 in [3], the group of connected components of the group Aut ¯S is S2 × S3, which can be identified with the au- tomorphism group of the hexagon of (−1)-curves on ¯S. In particular, there is an element σ ∈ Aut( ¯S) that cyclically permutes the six exceptional lines. It is easy to see that (1 + σ)(1 − σ3) acts trivially on Pic ¯S. An anti-canonical divisor is −K ¯S := L1 + L2 + L3 + M1 + M2 + M3. This is ample. We define two particular divisors H := L1 + M2 + M3 ∼ L2 + M1 + M3 ∼ L3 + M1 + M2 (2-2) and H ′ (2-3) on ¯S. Note that σ(H) ∼ H ′ and σ2(H) ∼ H. := L1 + L2 + M3 ∼ L2 + L3 + M1 ∼ L3 + L1 + M2 A DERIVED EQUIVALENCE FOR A DEGREE 6 DEL PEZZO SURFACE 3 We define the degree of a divisor C on ¯S as deg C = −C · K. Each exceptional line has degree 1. There are two morphisms f, f ′ : ¯S → P2 ¯F , each of which realizes ¯S as the blowup of P2 ¯F at three non-collinear points. We choose these so that f contracts the lines L1, L2, and L3 and f ′ contracts the lines M1, M2, and M3. These two morphisms induce injective group homomorphisms f ∗, f ′∗ : Pic P2 → Pic ¯S. If ℓ is a line on P2 ¯F , then f ∗ℓ = H and f ′∗ℓ = H ′. The action of Gal( ¯F /F ) on the exceptional lines on ¯S induces actions of Gal( ¯F /F ) on and I := 5Mi=0 O ¯S(σiH) J := O ¯S(σi(L1 + M2)) 5Mi=0 that are compatible with its action on ¯S. In particular, I and J are Gal( ¯F /F )- invariant. It follows that the locally free sheaves I and J descend to locally free sheaves I and J on S. Define and T := I ⊕ J ⊕ O ¯S, T := I ⊕ J ⊕ OS, B := EndS I, Q := EndS J , A := EndS T . (B, Q). (Actually, in [3], B is defined as(cid:0) EndS I ∨(cid:1)op In [3] it is shown that S is determined up to isomorphism by the pair of F -algebras . Since sending a homomor- phism α : I → I to its transpose α∨ : I ∨ → I ∨ is an anti-isomorphism from EndS I to EndS I ∨, our B is the same as that in [3], and similarly for Q.) As discussed in [3], the algebras B and Q are Azumaya over their centers, which are respectively ´etale quadratic and cubic extensions of F . Moreover, these ´etale centers can be re- covered from the action of Gal( ¯F /F ) on the hexagon of (−1)-curves, as the action induces a 1-cocycle of Gal( ¯F /F ) with values in S2 × S3, inducing a pair of ´etale extensions of F , quadratic and cubic. We end this section with two results about the endomorphism algebra of T . Lemma 2.1. Let A := EndS T . Then A = B HomS(J , I) HomS(OS, I) 0 HomS(OS, J ) 0 Q 0 F  . It suffices to show Hom ¯S(I, J ) = Hom ¯S(I, O ¯S) = Hom ¯S(J , O ¯S) = 0. Proof. However, each of these three Hom-spaces is isomorphic to a direct sum of terms of the form H 0( ¯S, O ¯S(D)) for a divisor D with deg D < 0. But if D has a section then D ∼ D′ for some effective D′ so deg D = −D′.K ≥ 0. These Hom spaces are therefore zero. (cid:3) The projective dimension of a left T -module is denoted by pdimT M . The global homological dimension of T is defined and denoted by gldim T := sup{pdimT M M ∈ ModT }. Proposition 2.2. gldim A ≤ 2. 4 M. BLUNK, S.J. SIERRA, AND S. PAUL SMITH Proof. Let R and S be rings and X an R-S-bimodule. If S is a semisimple ring, then (See [1, Prop. III.2.7].) Applying this result twice, first to (2-4) gldim(cid:18)R X 0 S(cid:19) = max{pdimR X + 1, gldim R}. A′ :=(cid:18)B Hom(J , I) (cid:19) Q 0 then to A with R = A′ and S = F , gives the desired result. (cid:3) 3. Cohomology vanishing lemmas We will prove several results about vanishing of cohomology and Ext-groups for sheaves on S. These results will be used in Section 4 to show that T is a tilting bundle and therefore induces an equivalence of derived categories. A key step in proving that T is tilting is showing that Exti S(T , T ) = 0 for i > 0. This reduces, by flat base change, to proving that Exti ¯S(T , T ) = 0. Given the explicit description of T as a direct sum of invertible sheaves, it suffices to prove that h1(D − D′) = h2(D − D′) = 0 for all D and D′ belonging to the list (3-1) H, H ′, L1 + M2, L2 + M3, L3 + M1, 0. We will make repeated use of the relation Li + Mj ∼ Lj + Mi. Proposition 3.1. Let D and D′ be divisors on ¯S appearing in the list (3-1). Then −3 ≤ deg(D − D′) ≤ 3. Furthermore, (1) if deg(D − D′) = 1, then D − D′ is linearly equivalent to an exceptional line. (2) if deg(D − D′) = 2, then D − D′ ∼ Li + Mj for some i 6= j ∈ {1, 2, 3}. (3) if deg(D − D′) = 3, then D − D′ is linearly equivalent to either H or H ′. (4) if deg(D − D′) = 0, then D − D′ is linearly equivalent to either 0, Li − Lj, Li − Mi, or Mi − Li for some i, j ∈ {1, 2, 3}. (5) if deg(D − D′) < 0, then D − D′ is linearly equivalent to either −Li, or −Mj, or −Li − Mj, or −H, or −H ′. Proof. Exceptional lines have degree 1 so deg H = deg H ′ = 3 and deg(Li + Mj) = 2. It follows that the degree of D − D′ is between 3 and -3. (1) If deg(D−D′) = 1, then D is linearly equivalent to H or H ′ and D′ = Li+Mj for some i, j. It follows from (2-2) and (2-3) that D − D′ is linearly equivalent to an exceptional line, and every exceptional line can occur as D − D′. (2) and (3) are obvious. (4) In this case D and D′ have the same degree. If deg D = deg D′ = 2, then D = Li + Mj and D′ = Lk + Mℓ. By considering all possible i, j, k, ℓ, we see that D − D′ is linearly equivalent to a divisor of the form Li − Lj. If deg D = deg D′ = 3, then, for example, D ∼ H and D′ ∼ H ′, and D − D′ ∼ Li − Mi. Switching the roles of H and H ′, we see D − D′ ∼ Mi − Li. Finally, we may have D − D′ ∼ 0. (5) This is the mirror of the cases (1)-(3). (cid:3) A DERIVED EQUIVALENCE FOR A DEGREE 6 DEL PEZZO SURFACE 5 Corollary 3.2. Suppose D is the difference of two divisors appearing in the list (3-1). If deg D ≥ −2, then there is an exceptional line E on ¯S such that D − E is also a difference of two divisors appearing in the list (3-1) and D.E ≥ −1. Proof. This is established through case-by-case analysis using Proposition 3.1 to look at all the possibilities for D. (cid:3) A divisor D on ¯S is good if h1(D) = h2(D) = 0. Lemma 3.3. The divisors −H and −H ′ on ¯S are good. Proof. The existence of the morphisms f, f ′ : ¯S → P2 ¯F allows us to use the Leray spectral sequence. The arguments for −H and −H ′ are the same so we only prove the result for −H. Because ¯S is a blowup of P2 Since O ¯S(−H) ∼= f ∗OP2 ¯F ¯F , f∗O ¯S = OP2 ¯F and Rjf∗O ¯S = 0 if j ≥ 1. (−ℓ), the projection formula gives Rjf∗O ¯S(−H) = Rjf∗(cid:0)O ¯S ⊗ f ∗OP2 ∼= Rjf∗O ¯S ⊗ OP2 ¯F (−ℓ) (−ℓ)(cid:1) ¯F ∼=(OP2 0 ¯F (−ℓ) if j = 0 if j 6= 0. The Leray spectral sequence H i(P2 ¯F , Rjf∗O ¯S(−H)) ⇒ H i+j( ¯S, O ¯S(−H)) therefore degenerates to give H i( ¯S, O ¯S(−H)) ∼= H i(P2 ¯F , OP2 (−ℓ)) for all i. The result follows because H i(P2 ¯F , OP2(−ℓ)) = 0 for all i. (cid:3) Lemma 3.4. Let C be any divisor on ¯S, and let E be one of the (−1)-curves. If C − E is good and C.E ≥ −1, then C is good. Proof. The long exact sequence in cohomology associated to 0 → O ¯S(C − E) → O ¯S(C) → OE(C) → 0 reads in part / H 1( ¯S, O ¯S(C − E)) / H 1( ¯S, O ¯S(C)) / H 1( ¯S, OE(C)) / H 2( ¯S, O ¯S(C − E)) / H 2( ¯S, O ¯S(C)) / H 2( ¯S, OE(C)). By hypothesis, the left-most term in each row is zero. The right-most term in each row is also zero because H i( ¯S, OE(C)) ∼= H i(P1 ¯F , OP1(C.E)). Hence C is good. (cid:3) / / / / / / / / 6 M. BLUNK, S.J. SIERRA, AND S. PAUL SMITH 4. The tilting bundle T In this section, we show that T is a tilting bundle and prove our main result. Proposition 4.1. Let i ≥ 1. Then Exti Proof. By flat base change it suffices to prove this when F is separably closed so we assume that F = ¯F . In that case Exti S(T , T ) is isomorphic to a direct sum of terms of the form H i(S, OS(D − D′)) where D and D′ are divisors in the list (3-1). It therefore suffices to show that D − D′ is good whenever D and D′ are divisors S(T , T ) = 0. in the list (3-1). We argue by induction on deg(D −D′). By Proposition 3.1, −3 ≤ deg(D −D′) ≤ 3. If deg(D − D′) = −3, then D − D′ is good by Lemma 3.3. Now suppose that −2 ≤ deg(D − D′) ≤ 3. By Corollary 3.2, there is an exceptional line E such that D − D′ − E is a difference of divisors in (3-1) and (D − D′).E ≥ −1. By the induction hypothesis, D − D′ − E is good, and it then follows from Lemma 3.4 that D − D′ is good. (cid:3) Since ¯S is a del Pezzo surface of degree ≥ 6 it is a toric variety so we can, and will, make use of Cox's homogeneous coordinate ring for it [5]. Lemma 4.2. Every F ∈ coh ¯S has a finite resolution in which all terms are direct sums of invertible sheaves O ¯S(D) for various divisors D on ¯S. Proof. Let A be Cox's homogeneous coordinate ring for ¯S [5]. Then A is a polynomial ring with a grading by Pic( ¯S). Let M be a finitely generated graded A-module. Then M has a finite projective resolution in the category of graded A- modules. By [9, Lemma 2.2], every finitely generated projective graded A-module various O ¯S(D), D ∈ Div( ¯S). Given F ∈ coh ¯S, there is a finitely generated graded (cid:3) described in [5, Thm. 3.11] sends the resolution of M to an exact sequence in is a direct sum of twists of A. The exact functor Gr(A, Pic( ¯S)) → Qcoh ¯S, M fM , Qcoh ¯S in which the right-most term is fM and all other terms are direct sums of A-module M such that F ∼= fM . For the rest of this paper, we will work in the derived category. If D is a tri- angulated category, we denote the shift of an object M by M[1]. Recall that a subcategory of D is thick (´epaisse) if it is closed under isomorphisms, shifts, taking cones of morphisms, and taking direct summands of objects. Let D be a triangulated category and E a set of objects in D. Then • Dc denotes the full subcategory of D consisting of the compact objects, i.e., those objects C such that HomD(C, −) commutes with direct sums; • hEi denotes the smallest thick full triangulated subcategory of D containing E; • E ⊥ denotes the full subcategory of D consisting of objects M such that HomD(E[i], M) = 0 for all E ∈ E and all i ∈ Z. We say that • E generates D if E ⊥ = 0 and that • D is compactly generated if (Dc)⊥ = 0. Clearly, if D is compactly generated and hEi = Dc, then E generates D. Theorem 4.3 (Ravenel and Neeman [8]. Also see Thm. 2.1.2 in [4]). Let D be a compactly generated triangulated category. Then a set of objects E ⊂ Dc generates D if and only if hEi = Dc. (cid:3) A DERIVED EQUIVALENCE FOR A DEGREE 6 DEL PEZZO SURFACE 7 The unbounded derived categories D(QcohS) and D(Qcoh ¯S) are compactly gen- erated. Moreover, D(QcohS)c = Db(cohS) and D(Qcoh ¯S)c = Db(coh ¯S). Tilting bundles. Let X be a projective scheme over a field k. A locally free X (T , T ) = 0 sheaf T ∈ cohX is a tilting bundle if it generates D(QcohX) and Exti for all i > 0. Theorem 4.4. T generates D(Qcoh ¯S) and hT i = Db(coh ¯S). Proof. By Theorem 4.3, it suffices to show that hT i = Db(coh ¯S). Since hcoh ¯Si = Db(coh ¯S) it suffices to show that every coherent O ¯S-module belongs to hT i. If D is an effective divisor on ¯S we write ID for the ideal vanishing on D as a scheme. Thus ID ∼= O ¯S(−D). Whenever we write an arrow O ¯S(−D) → O ¯S it will be with the tacit understanding that this is the composition of an isomorphism O ¯S(−D) → ID followed by the inclusion ID → O ¯S. ∼= OM3 (L1 + M2 + M3). It follows from Since M3 · (L1 + M2 + M3) = 0, OM3 the exact sequences 0 → O ¯S(L1 + M2) → O ¯S(L1 + M2 + M3) → OM3 (L1 + M2 + M3) → 0 and 0 → O ¯S(−M3) → O ¯S → OM3 → 0 that OM3 and O ¯S(−M3) belong to hT i. Hence OE and O ¯S(−E) belong to hT i for all exceptional lines E. Since Li.Lk = 0 if i 6= k, there is an exact sequence 0 → O ¯S(−Li − Lk) → O ¯S(−Li) ⊕ O ¯S(−Lk) → O ¯S → 0. Twisting by Li + Mj + Lk, we obtain 0 → O ¯S(Mj) → O ¯S(Mj + Lk) ⊕ O ¯S(Li + Mj) → O ¯S(Li + Mj + Lk) → 0. Therefore, O ¯S(Mj) ∈ hT i. From the exact sequence 0 → O ¯S → O ¯S(Mj) → OMj (Mj) → 0, we deduce that OMj (Mj) ∈ hT i. It follows that OE(E) ∈ hT i for every exceptional curve E. But OE is also (−1), it follows that in hT i so, because Db(cohP1 and OP1 ¯F Db(cohE) ⊂ hT i. Hence OE(D) ∈ hT i for all divisors D on ¯S. ¯F ) is generated by OP1 ¯F Suppose O ¯S(D) ∈ hT i. Then O ¯S(D − E) ∈ hT i because there is an exact sequence 0 → O ¯S(D − E) → O ¯S(D) → OE(D) → 0. Likewise, O ¯S(D + E) ∈ hT i because there is an exact sequence 0 → O ¯S(D) → O ¯S(D + E) → OE(D + E) → 0. It follows that hT i contains O ¯S(D) for all D ∈ Div ¯S and therefore, by Lemma 4.2, contains F for every F ∈ coh ¯S. (cid:3) When F is not separably closed T need not split as a direct sum of line bun- dles so the arguments in Theorem 4.4 can not be used to prove directly that hT i = Db(cohS). Instead we will show that T generates D(QcohS) and then apply Theorem 4.3. 8 M. BLUNK, S.J. SIERRA, AND S. PAUL SMITH Theorem 4.5. Let F be an arbitrary field. Then RHomS(T , −) : Db(cohS) → Db(modA) is an equivalence of categories. Proof. We will show that T generates D(QcohS). It will then follow from Theorem 4.3 that hT i = D(QcohS)c = Db(cohS). By Proposition 4.1, Exti S(T , T ) = 0 for i > 0. By Proposition 2.2, A = EndS(T ) has finite global dimension. Thus we have shown that T is a tilting bundle and our theorem will then follow directly from [2, Thm. 3.1.2] (or [6, Thm. 7.6]). Let M ∈ D(QcohS) and suppose RHomS(T , M) = 0. We must show that M = 0. Since T is locally free, Hom S(T , −) and T ∨ ⊗S − are exact functors on QcohS. ⊗ ¯S − are exact functors on Qcoh ¯S. Thus, for ex- Likewise, Hom ¯S(T , −) and T ample, RHom S(T , M) can be computed on D(QcohS) by applying Hom S(T , −) to each individual term in M. ∨ Consider the cartesian square ¯S q Spec( ¯F ) v u S p / Spec(F ). Since u (and therefore v) is flat, the natural transformation u∗Rp∗ → Rq∗v∗ is an isomorphism of functors from D(QcohS) to D( ¯F ) [7, (3.18)]. We now have 0 = u∗ RHomS(T , M) ∼= u∗Rp∗ RHom S(T , M) by [7, p.85] ∼= Rq∗v∗ RHom S(T , M) by [7, (3.18)] ∼= Rq∗v∗(T ∨ ⊗L ∼= Rq∗(T ∼= Rq∗ RHom ¯S(T , Lv∗M) ∼= RHom ¯S(T , Lv∗M). S M) ¯S Lv∗M) ∨ ⊗L But T generates D(Qcoh ¯S) so v∗M = 0. Since v∗ is faithful, M = 0, and we are done. (cid:3) Corollary 4.6 (cf. [3], Corollary 5.2). The functor HomS(T , −) : coh(S) → modA induces an isomorphism HomS(T , −) : K∗(S) → K∗(F × B × Q). Proof. It follows from Theorem 1.98 of [10] that the equivalence of derived cate- gories found in Theorem 4.5 induces an isomorphism in K-theory HomS(T , −) : K∗(cohS) → K∗(modA). Moreover, A has a nilpotent ideal I so that A/I is isomorphic to its semi-simple quotient F × B × Q. Thus, it follows that the K-theory of A is isomorphic to that of F × B × Q, and we recover the isomorphism found in [3]. (cid:3) / / / A DERIVED EQUIVALENCE FOR A DEGREE 6 DEL PEZZO SURFACE 9 References [1] M. Auslander, I. Reiten, and S. Smalø, Representation Theory of Artin Algebras, Camb. Studies in Adv. Math., No. 36, 1995. [2] D. Baer, Tilting sheaves in representation theory of algebras, Manus. Math., 60 (1988) 323- 347. [3] M. Blunk, Del Pezzo surfaces of degree 6 over an arbitrary field, J. Algebra, 323 (2010), no. 1, 42 -- 58 [4] A. Bondal and M. Van den Bergh, Generators and representability of functors in commutative and noncommutative geometry, Moscow Mathematical Journal, 3 (2003), no. 1, 1-36. [5] D.A. Cox, The homogeneous coordinate ring of a toric variety, J. Alg. Geom., 4 (1995), no. 1, 17-50. [6] L. Hille and M. Van den Bergh, Fourier-Mukai transforms, pp. 147-173 in Handbook of Tilting Theory, Lond. Math. Soc. Lect. Note Series (No. 332), Eds. L.A. Hugel, D. Happel, and H. Krause, Camb. Univ. Press, 2007. [7] D. Huybrechts, Fourier-Mukai transforms in algebraic geometry, Oxford Univ. Press, 2006. [8] A. Neeman, The connection between the K-theory localization theorem of Thomason, Trobaugh and Yao and the smashing subcategories of Bousfield and Ravanel, Ann. Sci. de l' ´Ec. Norm. Sup., 25 (1992), no. 5, 547-566. [9] S.P. Smith, Computation of the Grothendieck and Picard groups of a toric DM stack X by using a homogeneous coordinate ring for X , Glasgow Math. Jour., to appear. arXiv:0806.0192v2. [10] R. W. Thomason and T. Trobaugh, Higher algebraic K-theory of schemes and of derived categories, pp. 247 -- 435 in The Grothendieck Festschrift, Vol. III , Progr. Math., 88, Birkhuser Boston, Boston, MA, 1990. Department of Mathematics, Univ. of British Columbia., Department of Mathemat- ics, Box 354350, Univ. Washington, Seattle, WA 98195 E-mail address: [email protected], [email protected], [email protected]
1305.7207	4	1305	2015-02-08T14:56:12	On the Average Value of the Canonical Height in Higher Dimensional Families of Elliptic curves	[ "math.AG", "math.NT" ]	Given an elliptic curve E over a function field K=Q(T_1,...,T_n), we study the behavior of the canonical height ^h_(E_w) of the specialized elliptic curve E_w with respect to the height of w in Q^n. In this paper, we prove that there exists a uniform non-zero lower bound for the average of the quotient ^h_(E_w)(P_w)/h(w) for all non-torsion P in E(K).	math.AG	math	ON THE AVERAGE VALUE OF THE CANONICAL HEIGHT IN HIGHER DIMENSIONAL FAMILIES OF ELLIPTIC CURVES WEI PIN WONG Abstract. Given an elliptic curve E over a function field K = Q(T1, . . . , Tn), we study the behavior of the canonical height hEω of the specialized elliptic curve Eω with respect to the height of ω ∈ Qn. In this paper, we prove that there exists a uniform non- zero lower bound for the average of the quotient for all non-torsion P ∈ E(K). hEω (Pω ) h(ω) 1. Introduction Let K be the function field Q(T1, . . . , Tn) and T = (T1, . . . , Tn). Let E/K be an elliptic curve with Weierstrass equation: Y 2 = X 3 + A(T)X + B(T) where by change of variable, we can assume A(T), B(T) ∈ Z[T] and there's no nonconstant g(T) ∈ Q[T] such that A(T) g(T)4 , B(T) g(T)6 ∈ Z[T]. We further assume that E/K is not split over K, i.e. E is not K- birational isomorphic to E0 ×Q K for any elliptic curve E0/Q. This implies A(T) and B(T) cannot be both constant. The discriminant ∆E(T) = −16(4A3(T) + 27B2(T)) is a non-zero element in Z[T]. Let Qn(∆E) be the set of all ω = (ω1, . . . , ωn) ∈ Qn such that ∆E(ω) 6= 0. Thus for every P ∈ E(K), for Date: January 14, 2021. 2010 Mathematics Subject Classification. Primary 11G05; Secondary: 11G50, 14G40. Key words and phrases. height function, elliptic curve, function field, average, lower bound. 1 2 W. WONG ω such that Pω := P (ω) is defined, the point Pω is a rational point on the elliptic curve Eω/Q defined by the Weierstrass equation Y 2 = X 3 + A(ω)X + B(ω). We denote the canonical height on Eω by hEω and the logarithmic height on Pn Q by h, i.e. h(ω) = log H(ω) := log H([1, ω1, . . . , ωn]), where H([ν0, . . . , νn]) = max {νi}, i if νi ∈ Z and gcd(ν0, . . . , νn) = 1. To ease the notation, we will denote ν := max i {νi} for any ν ∈ Zn. We prove the following theorems about the average value of hEω (Pω) h(ω) . Theorem 1. With notation as above, let Qn B(∆E) := {ω ∈ Qn 1 < H(ω) ≤ B and ∆E(ω) 6= 0}, and E(K)nt := {P ∈ E(K) P non-torsion }. Then there exists an L1 > 0 depending only on ∆E, such that for all P in E(K)nt, let Qn B(∆E, P ) := {ω ∈ Qn B(∆E) Pω is defined }, we have AhQ E(P ) := lim inf B→∞ #Qn 1 B(∆E, P ) Xω∈Qn B (∆E,P ) hEω(Pω) h(ω) ≥ L1. When n = 1, Silverman proved in [12] that hEω(Pω) h(ω) = hE(P ), lim ω∈ ¯Qn h(ω)→∞ where hE(P ) is the canonical height of P in E/K. One would like to obtain a similar result for general n but by a simple observation this limit cannot exist for n ≥ 2. This is because we can restrict the ω to lie on a particular algebraic curve γ for h(ω) tends to infinity, reducing this to the case of n = 1, but now the limit obtained will depend on AVERAGE VALUE OF THE CANONICAL HEIGHT 3 Pγ and the elliptic curve Eγ in which it lies. For illustration, consider the elliptic curve E/Q(S, T ) : Y 2 = X 3 − S2X + T 2 h(ω) hEω (Pω) and P = (S, T ) ∈ E(Q(S, T )). If we restrict ω to γ : S = 0, a simple calculation shows that Pγ = (0, T ) is a torsion point on Eγ(Q(T )) : Y 2 = X 3 + T 2. Thus the limit of is zero when h(ω) tends to infinity by restriciting ω ∈ γ. On the other hand, if we restrict ω on the curve γ′ : S = T , Pγ ′ = (T, T ) is in a basis of Eγ ′(Q(T )) : Y 2 = X 3 − T 2X + T 2 (this is an example given in [13]). Thus Silverman's theorem implies a non-zero limit of the quotient when h(ω) tends to infinity by restriciting ω ∈ γ′. In fact this limit is 1 6. One can also look at the restriction T = 1 (resp. S = 1) and get the limit of the quotient equal to 1 2 (resp. 1 3 ). Since the limit of the quotient hEω (Pω) h(ω) fails to exist in general for n ≥ 2, we turn our attention to look at the average of the quotient: AhQ E(P )B := #Qn 1 B(∆E, P ) Xω∈Qn B (∆E ,P ) hEω (Pω) h(ω) . Following the idea of Silverman, we would like make the following con- jecture: Conjecture 2. With the same setting as Theorem 1, for any P ∈ E(K), AhQ E(P )B = hE(P ). lim B→∞ The case n = 1 for this conjecture is true, which follows trivially from Silverman's theorem and Ces`aro mean theorem. However, proving this conjecture for n ≥ 2 appears to be difficult, so we first check whether the conjecture even makes sense, i.e., if the limit of the average exists as a function of P ∈ E(K), does it satisfy the properties of canonical height function ( [14] Chapter VIII, Theorem 9.3 or [7] Chapter 5)? One such property is that hE is a quadratic form. By linearity of average, it's straightforward that the limit of AhQ E(−)B, if it exists, is a quadratic form too. Another important propety of the canonical height on E/K is that hE(P ) = 0 if and only if P is in the subgroup generated 4 W. WONG by torsion points and the image of K/Q-trace of E ([7] Chapter 6, Theorem 5.4). Since we assume E is not split over K, then the K/Q- trace is of dimension zero, which means it's the trivial group and hence its image in E is the identity ([4] Example 2.2). In other words, if E is not split over K, then (1) hE(P ) = 0 if and only if P is a torsion point. So we investigate property (1) for the limit inferior of AhQ E(P )B. We shall prove that the limit inferior of AhQ E(P )B is zero if and only if P is a torsion point of E(K). The if part is trivial as if P is a torsion point of E(K), then Pω is a torsion point of Eω(Q) and so the average is always zero. It turns out the other direction is also true. We will first prove this by looking at the average over Zn, which is Proposition 3. Proposition 3. With notation as above, we further let Zn B(∆E) := {ν ∈ Zn 1 < ν ≤ B and ∆E(ν) 6= 0}. Then there exists an L2 > 0 depending only on ∆E, such that for all P in E(K)nt, let Zn B(∆E, P ) := {ν ∈ Zn B(∆E) Pν is defined }, we have AhZ E(P ) := lim inf B→∞ 1 (2B)n Xν∈Zn B (∆E ,P ) hEν (Pν) h(ν) ≥ L2. E(−) and AhQ Proposition 3 is the key tool used to prove Theorem 1 via a standard inclusion-exclusion argument. Notice that Proposition 3 and Theorem 1 state something stronger: there exists a uniform non-zero lower bound of AhZ E(−) for all non-torsion P in E(K). One might think that the uniform lower bound is expected once we proved that AhZ E(P ) > 0 for P in E(K)nt, due to the fact that E(K)nt is finitely generated and hEω can be extended to a positive definite quadratic form on Eω(Q) ⊗Z R. At the level of Eω(Q), one can get a uniform lower bound of hEω on the lattice Eω(Q)nt ⊂ Eω(Q) ⊗Z R in terms of the canonical height of a nice basis of Eω(Q)nt. ([7] Chapter 5, Theorem 7.7 and Corollary 7.9). However, at the average level, it is E(P ) > 0 and AhQ AVERAGE VALUE OF THE CANONICAL HEIGHT 5 not obvious at all whether one can find a basis {Pi}i∈I of E(K)nt such that the specialization {Pi(ω)}i∈I is always a nice basis in the image of specialization (E(K)nt)ω ⊆ Eω(Q)nt for all ω. Our proofs produce the uniform lower bounds without exploiting these facts. We will postpone the proofs of Proposition 3 and Theorem 1 to Section 5 and 6 respectively. On the other side, we also prove that the limit superior of AhQ E(P )B is finite. Theorem 4. With the same hypothesis as in Theorem 1 and for any P ∈ E(K), there exists a constant UP depending only on P , such that hEω(Pω) ≤ UP (1 + h(ω)) for all ω ∈ Qn B(∆E, P ). Consequently, we have Ah Q E(P ) := lim sup B→∞ #Qn 1 B(∆E, P ) Xω∈Qn B (∆E ,P ) hEω (Pω) h(ω) < ∞. In fact Theorem 4 is true in a more general setting as stated in the following theorem: Theorem 5. Let k be a number field, let S and A be nonsingular, irreducible, projective varieties defined over k, and let π : A → S be a flat morphism defined over k so that the generic fiber Aη of π is an abelian variety over k(S). Let S0 := {ω ∈ S(k) Aω is a non-singular abelian variety defined over k}. Fix a divisor D ∈ Divk(A). For each ω ∈ S0, let Dω ∈ Div(Aω) be any divisor in the restriction of the divisor class of D to Aω and the corresponding canonical height be hAω,Dω . Fix a projective embedding i : S ⊂ Pn, then for any P ∈ Aη(k(S)), there exists a constant c0 depending on hA,D, D, i and P such that hAω,Dω(Pω) < c0(1 + h(i(ω))) for all ω ∈ S0 with Pω is defined. As a consequence, if we let S0 B(P ) := {ω ∈ S0 1 < H(i(ω)) < B, P is defined.}, 6 then W. WONG Ah Q Aη(P ) := lim sup B→∞ 1 #S0 B(P ) Xω∈S0 B(P ) hAω,Dω (Pω) h(i(ω)) < ∞. Theorem 5 is easier to prove than Theorem 1, so we will prove this theorem first in Section 2. After that we will prove Theorem 4 in Section 3 by a similar fashion. The behavior of hEω (Pω) for n = 1 is well studied in the literature in a more general setting of an abelian variety A defined over a function field k(C) of a non-singular projective curve C over a number field k. In fact this is the original setting in [12] where Silverman proved hAt(Pt) h(t) = hA(P ). lim t∈C(¯k) h(t)→∞ For the special case where A = E is an elliptic surface, Tate [18] ob- tained a stronger result by showing that hEt(Pt) = hE(P )h(t) + OP (ph(t) + 1) and if C = P1, the error is only OP (1). This stronger result was extended to the general case of abelian varieties by Lang ([7] Chapter 12, Section 5) under the assumption that the N´eron model of the generic fiber has a good completion. In [2], Call reproved Lang's result using a theorem on canonical heights and further discussed cases where the good completion assumption may be weakened or eliminated. Readers can consult Chapter III of [16] for a nice introduction and other results on elliptic surfaces. Although the behavior of hEω (Pω) for n ≥ 2 is not yet well studied in the literature, we know something about the density of ω such that hEω(Pω) = 0, i.e. Pω is torsion. Again, this is known in the setting of an abelian variety A defined over a function field k(V ) of a variety V over a number field k. In [9], Masser proved that for a finitely generated subgroup Γ of A the specialization homomorphism σω : Γ → Aω(k(ω)) is injective "almost always" for ω ∈ V (¯k). AVERAGE VALUE OF THE CANONICAL HEIGHT 7 2. Proof of Theorem 5 hAω ,Dω (Pω) h(i(ω)) Notice that it suffices to prove that the quotient is bounded above uniformly for all ω ∈ S0 B(P ). This is an immediate consequence of Theorem A of [12], due to Silverman and Tate. In effect, with the given hypothesis in Theorem 5 and further let hA,D be the Weil height (defined up to equivalence) corresponding to D, Theorem A says that there exists a constant c depending on D and A, so that for all P ∈ Aη(K) (cid:12)(cid:12)(cid:12)hAω,Dω (Pω) − hA,D(Pω)(cid:12)(cid:12)(cid:12) < ch(i(ω)) + O(1), where O(1) depends on the choice of particular Weil heights hA,D and the embedding i. So we turn the problem into estimating hA,D(Pω). We remind the reader about the definition of hA,D. If D ∈ Divk(A) is very ample, then choose an embedding correspnding to the linear system D and hA,D is defined by φD : A → Pm k hA,D : A(k) −→ R p 7−→ h(φD(p)). For a general divisor D ∈ Divk(A), write D = X − Y , where X, Y ∈ Divk(A) are very ample divisors, and define hA,D(p) := hA,X(p) − hA,Y (p). For any P ∈ Aη(K), it defines a rational map ψP : S A ω 7−→ Pω. So we have hA,D(Pω) = hA,X(Pω) − hA,Y (Pω) = h(φX(ψP (ω))) − h(φY (ψP (ω))) ≤ h(φX(ψP (ω))), where fX := φX ◦ ψP is a rational map from S to Pm. By using the triangle inequality of absolute values of k, one can show the following 8 W. WONG standard property of height on projective space ([7] Chapter 4, Lemma 1.6): h(fX(ω)) ≤ dh(i(ω)) + c1 for some constant c1 and d that depend on fX only. Finally, by applying Theorem A, we get hAω,Dω (Pω) ≤ hA,D(Pω) + ch(i(ω)) + O(1) ≤ dh(i(ω)) + c1 + ch(i(ω)) + O(1), which is the first part of the theorem. Since the set of points of bounded height in Pn(k) is finite, there's a non-zero lower bound (which depends on k) for h(i(ω)) > 0. We obtain our desired uniform upper bound for hAω ,Dω (Pω) by dividing the inequality above by h(i(ω)) > 0 and hence h(i(ω)) proved the second part of the theorem. 3. Proof of Theorem 4 We remark that Theorem 4 doesn't follow trivially from Theorem 5 even if we can find a nonsingular irreducible projective variety E/Q and a flat morphsim π : E −→ Pn with generic fiber Eη isomorphic to E/K. This is because it is not true in general that we can find a divisor D ∈ DivQ(E) such that hE,D(p) = hP1([x(p), 1]) + O(1), due to the fact that the X-coordinate map φ : E P1 p 7−→ [x(p), 1] is just a rational map in general. By mimicking the idea of the proof of Theorem A in [12], one can overcome this by blowing up E and extending φ to a morphism. However, we found a more direct and elementary proof for Theorem 4, which is the one that we are going to present. Using just the definition of height on elliptic curves and triangle inequality of absolute values of Q, we first prove that there exist positive AVERAGE VALUE OF THE CANONICAL HEIGHT 9 constants c1, c2 such that for all ω ∈ Qn(∆E) and all p ∈ Eω(Q), we have (2) hEω ([2]p) − 4hEω (p) ≤ c1h(ω) + c2. Recall that Eω is defined by the Weierstrass equation: Y 2 = X 3 + A(ω)X + B(ω). For any p = (x, y) ∈ Eω(Q), we may assume [2]p 6= OEω or otherwise inequality (2) is trivially true for any positive c1, c2. The duplication formula gives x([2]p) = x4 − 2A(ω)x2 − 8B(ω) + A(ω)2 4x3 + 4A(ω)x + 4B(ω) . Thus, we have HEω([2]p) := H ([x([2]p), 1]) = H(cid:0)[x4 − 2A(ω)x2 − 8B(ω)x + A(ω)2, 4x3 + 4A(ω)x + 4B(ω)](cid:1) ≤ 4H(cid:0)[1, −2A(ω), −8B(ω), A(ω)2, 4, 4A(ω), 4B(ω)](cid:1) H ([x, 1])4 (3) ≤ 4NA,BH(ω)dA,BHEω (p)4 where the inequalities are obtained by triangle inequality of absolute values of Q. The constant NA,B depends on the coefficients and the number of monomials of A and B, whereas dA,B is the maximum of deg A2 and deg B. Inequality (2) is obtained by taking natural loga- rithm of (3). Now, we use Tate's telescoping sum trick to prove an analogy of Theorem A in [12]: hEω (p) − hEω (p) = = ≤ ∞Xn=0 ∞Xn=0 ∞Xn=0 1 4n+1(cid:0)hEω ([2n+1]p) − 4hEω ([2n]p)(cid:1) 1 4n+1 (hEω ([2] ◦ [2n]p) − 4hEω([2n]p)) 1 4n+1 (c1h(ω) + c2) (Using (2)) 10 (4) W. WONG = c1 3 h(ω) + c2 3 . Finally, given any P = (x(T), y(T)) ∈ E(K), the X-coordinate of P defines a rational map ψP : Pn P1 ω 7−→ [x(ω), 1]. Just like in the proof of Theorem 5, the standard property of height on projective space gives (5) hEω (Pω) := h([x(ω), 1]) = h(ψP (ω)) ≤ dh(ω) + c3 for some constants d, c3 that depend on ψP only. We get our conclusion of Theorem 4 by combining (4) and (5). 4. Lemmas Besides some results on elliptic curves over Q, the proof of Proposi- tion 3 requires several non-trivial facts about polynomials with integer coefficients. In this section, we will state these results and give com- plete proofs with appropriate references. We remind the reader that we continue to use all the notations that we have defined previously. In addition, for the specialized elliptic curve Eω, let ∆Eω = ∆E(ω) and ∆min Eω be the discriminant and minimum discriminant of Eω/Q respec- tively. Also, for any UFD R, whenever we say P1, . . . , Pn are relatively prime in R[T], we always mean that P1, . . . , Pn don't have a common irreducible factor in R[T]. Lemma 6. There exists an absolute constant C1 > 0 such that the following holds. Let k ≥ 4 be an integer, Nk := lcm(1, 2, 3, . . . , k) and suppose ν ∈ Zn so that ∆E(ν) is non-zero and kth-power-free (abbre- viated as k-free for the rest of the paper). Then for any non-torsion point q ∈ Eν(Q), we have hEν (q) > C1 N 2 k log ∆min Eν . AVERAGE VALUE OF THE CANONICAL HEIGHT 11 Proof. We make use of a weakened form of a conjecture of Serge Lang proved by Silverman in section 4 of [11]. We apply it to a non-torsion point q ∈ Eν(Q) such that q is in (Eν)0(Qp) := {q ∈ Eν(Qp) q (mod p) is non-singular} for every prime p in Q. This is possible by Kodaira-N´eron Theo- rem ([16] Chapter VII, Theorem 6.1) which implies that the order of Eν(Qp)/(Eν)0(Qp) is either ordp(∆min Eν ) or at most 4. So if ∆E(ν) is k-free, we have ordp(∆min Eν ) ≤ k and thus [Nk]q is in (Eν)0(Qp) for all p with the choice of Nk := lcm(1, 2, 3, . . . , k). Then the special case of the conjecture gives hEν ([Nk]q) > C1 log ∆min Eν , for an absolute constant C1 > 0. Using the fact hEν is a quadratic form will complete the proof. (cid:3) Lemma 7. 1 log ν = o(Bn), Xν∈Zn 1<ν≤B where the implicit constant in the small o depends only on n. Proof. In this proof, the implicit constants of all the big O's depend only on n. By symmetry of each quadrant in Zn, we have 1 log ν Xν∈Zn 1<ν≤B 0 = O X1<x1≤x2≤...≤xn≤B = O(cid:18)Z B · · ·Z x2 2 Z xn = O(cid:18)Z B = O Z √B = O(cid:18) Bn log B(cid:19) . tn−1 log t (n − 1)! 1 2 2 1 log xn! log xn 1 dx1 · · · dxn(cid:19) dxn(cid:19) 0 xn−1 n log xn dt +Z B √B tn−1 log t dt! (cid:3) 12 W. WONG Lemma 8. Let k, m, r ∈ N satisfy 1 m 1 k + + 1 r ≤ 1. If P, Q, R ∈ C[T] satisfy P k + Qm = Rr, then either P, Q, R are all constant or else they are not relatively prime. Proof. We first prove the case n = 1, which is an immediate conse- quence of Mason -- Stothers theorem ([6] Chapter IV, Theorem 7.1 or [17] Theorem 1.1). Suppose to the contrary that P, Q, R are not all constant and relatively prime, then by Mason -- Stothers theorem, we have max{k deg P, m deg Q, r deg R} + 1 ≤ #distinct roots of P kQmRr ≤ deg P + deg Q + deg R. Without lose of generality, suppose k deg P ≥ m deg Q, which implies k deg P ≥ r deg R, so the inequality above becomes k deg P + 1 ≤ deg P + ⇒ 1 ≤(cid:18) 1 k which is absurd. + 1 m + 1 r k m k r deg P + deg P − 1(cid:19) k deg P ≤ 0, Now, let P, Q, R ∈ C[T] satisfy the hypothesis of the lemma. Sup- pose P, Q, R not all constant and relatively prime. Without lose of generality, we can assume the degrees of Tn in P, Q are at least 1. We will make use of some standard results about the resultant of two poly- nomials in R[x], where R is a UFD. These results eventually boil down to linear algebra ([5] Chapter VIII, Theorem 8.1). Consider P, Q as element in C[T1, . . . , Tn−1][Tn] and let f ∈ C[T1, . . . , Tn−1] be the resul- tant of P, Q with respect to the variable Tn. Then there exist non-zero u, v ∈ C[T] with degTn u < degTn Q and degTn v < degTn P such that uP + vQ = f. Since P, Q have no common factor in C[T], f cannot be identically zero. We can choose y := (y1, . . . , yn−1) ∈ Cn−1 such that f (y) 6= 0 and P (y, Tn) is nonconstant. Then Py(Tn) := P (y, Tn) and Qy(Tn) := Q(y, Tn) are relatively prime in C[Tn] and Py(Tn) is nonconstant. So we AVERAGE VALUE OF THE CANONICAL HEIGHT 13 get relatively prime Py, Qy, Ry ∈ C[Tn] such that not all are constant and satisfies the hypothesis of the lemma for n = 1, which is impossible as we have shown previously. (cid:3) Lemma 9. Let k, m, r ∈ N satisfy 1 m 1 k + + 1 r ≤ 1. Let ℓ = lcm(k, m) and g = gcd(k, m), and assume that ℓr. Let P, Q, R ∈ C[T] be polynomials with R 6= 0 that satisfy P k + Qm = Rr. Then there exists α1, α2 ∈ C such that P = α1R m g r ℓ and Q = α2R k g r ℓ . Proof. The case where P, Q, R are all constant is trivial. So suppose P, Q, R are not all constant. We let S := R ℓ and we have r (6) P k + Qm = Sℓ. Let G1, . . . , Gs be the distinct irreducible factors of P QS and write with α, β, γ ∈ C. Then we can rewrite the equality (6) as (7) P = αYi αkYi Gai Gaik Gbi i , Q = βYi i + βmYi i , S = γYi i = γℓYi Gciℓ Gbim . i Gci i We claim that aik = bim = ciℓ for all i. Notice that we cannot have one exponent of Gi in equation (7) that is strictly less than the other two, otherwise by dividing by the least power Gi factor, we get a con- tradiction. So two of the exponents of Gi in equation (7) are equal and at most equal to the third one. We divide equation (7) by Gi with the common lower exponent and we do this for all i. Using the fact that ℓ = lcm(k, m), the resulting equation can be written in the form P k 1 + Qm 1 = Sℓ 1, where P1, Q1, S1 are either all constant or relatively prime. Notice that the former case corresponds to our claim aik = bim = ciℓ for all i and we are going to prove that this must be the case. Since 14 W. WONG 1 m + 1 k + 1 r ≤ 1, without lose of generality, k ≥ 2 and m ≥ 3 and one easily verifies that ℓ := lcm(k, m) ≥ 6 except for the cases (k, m, ℓ) = (3, 3, 3), (2, 4, 4), (4, 4, 4), (5, 5, 5). So we always have 1 ℓ ≤ 1 and hence we can apply lemma 8 on P1, Q1, S1 to conclude that they are all constant. So we have m + 1 k + 1 P = α1S ℓ k = α1S m g and Q = α2S ℓ m = α2S k g . Substituting back S = R r ℓ completes the proof. (cid:3) To avoid heavy notation in the proofs below, we denote ZB := Z ∩ [−B, B] and for any F ∈ Z[T], ρF (m) := {ν ∈ (Z/mZ)n F (ν) ≡ 0 (mod m)}, F := max{c c is a coefficient of F }. Note: (1) By abuse of notation, the symbol ≡ used in the proofs of lem- mas 10 and 11 has three different meanings depending on the context. When f is an element of Z[x], f ≡ 0 means f is the zero polynomial. The notation f ≡ 0 in Z/pZ[x] means the reduction mod p of f is the zero polynomial in Z/pZ[x]. If we evalaute f at x and f (x) is an integer, the notation f (x) ≡ 0 (mod p) means p divides f (x). (2) By definition, a polynomial F ∈ Z[T] consists the information of its domain. Thus, if an implicit constant in the big O or small o notation is said to be dependent on F , that means that it depends on deg F and n as well. Lemma 10. Let F ∈ Z[T] with total degree d ≥ 1. Then for all prime p bigger than F , we have Np(F, B) := #{ν ∈ Zn B F (ν) ≡ 0 (mod p)} = O(cid:18)Bn p + Bn−1(cid:19) , where the implicit constant in the big O depends only on n and d. AVERAGE VALUE OF THE CANONICAL HEIGHT 15 Proof. In this proof, the implicit constants of all the big O's depend only on n and d. We prove by induction on n. For n = 1, with the condition on p, F 6≡ 0 ∈ Z/pZ[T1]. So Np(F, B) ≤ ρF (p)(cid:18)2B p + 1(cid:19) ≤ d(cid:18)2B p + 1(cid:19) . Now let F ∈ Z[T] and for y ∈ Zn−1, Fy(Tn) := F (y, Tn) ∈ Z[Tn]. The condition Fy ≡ 0 in Z/pZ[Tn] becomes a bunch (at most d) of polynomials of degree at most d in Z[T1, . . . , Tn−1] equal zero mod p. Thus, by induction hypothesis, #{y ∈ Zn−1 B So we get Np(F, B) = Xy∈Zn−1 B Fy ≡ 0 in Z/pZ[Tn]} = O(cid:18)Bn−1 Np(Fy, B) + Xy∈Zn−1 Np(Fy, B) p B + Bn−2(cid:19) . Fy6≡0 in Z/pZ[Tn] + (2B)n−2(cid:19) (2B + 1) + (2B + 1)n−1d(cid:18)2B p Fy≡0 in Z/pZ[Tn] ≤ O(cid:18)(2B)n−1 = O(cid:18)Bn p p + Bn−1(cid:19) . + 1(cid:19) (cid:3) Lemma 11. Suppose F (T) ∈ Z[T] has total degree d ≥ 1 and has no repeating irreducible factor in Z[T]. Then except for finitely many prime p, we have Np2(F, B) := #{ν ∈ Zn B F (ν) ≡ 0 (mod p2)} = O(cid:18) Bn p2 + Bn−1(cid:19) , whenever p ≤ B. The implicit constant in the big O depends only on F . Proof. In this proof, the implicit constants of all the big O's depend only on F , which includes d and n. Since we allow finitely many ex- ception on p, we can assume F is primitive, i.e. the content of F is 1. For n = 1, if p ∤Disc(F ) 6= 0, then Np2(F, B) ≤ d(cid:16) 2B Now let F ∈ Z[T] and for y ∈ Zn−1, Fy(Tn) := F (y, Tn) ∈ Z[Tn]. Let p2 + 1(cid:17) . 16 W. WONG Y := (T1, . . . , Tn−1). By Gauss' lemma for UFDs, the fact that F has no repeating irreducible factor in Z[T] = Z[Y][Tn] implies the same holds in Q(Y)[Tn]. So D(Y) := Disc(FY) 6≡ 0 ∈ Z[Y]. Also, for p bigger than D, D(Y) 6≡ 0 ∈ Z/pZ[Y]. We write FY(Tn) = ad(Y)T d n +. . .+a0(Y) and we divide into two cases: Case 1: gcd(ad(Y), . . . , a0(Y)) = 1. We decompose Np2(F, B) into the following three sums: Xy∈Zn−1 Np2(Fy, B) + Xy∈Zn−1 B Np2(Fy, B) + Xy∈Zn−1 B Fy6≡0 in Z/pZ[Tn] (mod p) D(y)≡0 Fy≡0 in Z/pZ[Tn] D(y)6≡0 B (mod p) Np2(Fy, B) for the second sum, we apply lemma 10 on D(Y) to get an upper bound The first sum is trivially bounded by (2B + 1)n−1d(cid:16) 2B p2 + 1(cid:19) = O(cid:18)Bn + Bn−2(cid:19) dp(cid:18)2B p2 + 1(cid:17). Whereas p2 + Bn−1(cid:19) . O(cid:18)Bn−1 p Lastly, since gcd(ad(Y), . . . , a0(Y)) = 1, if we look at the y ∈ (Z/pZ)n−1 such that Fy ≡ 0 in Z/pZ[Tn], either y is a common root in (Z/pZ)n−1 of at least two polynomials that are relatively prime in Z[T1, . . . , Tn−1] or there is no such y because there is only one non-zero ai(Y) and it must be 1 by assumption of case 1. From the proof of Theorem 3.1 of Poonen in [10], the set of such y ∈ (Z/pZ)n−1 has order O(p(n−1)−2) for large p. Hence we have for p ≤ B, Xy∈Zn−1 B Fy≡0 in Z/pZ[Tn] + 1(cid:19)n−1 p Np2(Fy, B) ≤ O(pn−3)(cid:18)2B = O(pn−3)O(cid:18)Bn−1 p2 + Bn−1(cid:19) = O(cid:18)Bn pn−1 + (2B + 1) Bn−2 pn−2(cid:19) (2B + 1) and we are done. AVERAGE VALUE OF THE CANONICAL HEIGHT 17 Case 2: gcd(ad(Y), . . . , a0(Y)) = g(Y) 6= 1. Then F (T) = A(T)g(Y) for some nonconstant A(T) ∈ Z[T]. Then Np2(F, B) = Xy∈Zn−1 O(2B) + Xy∈Zn−1 B B (mod p2) g(y)≡0 pg(y) Np(Ay, B) + Xy∈Zn−1 B p∤g(y) Np2(Ay, B), where pg(y) means pg(y) but p2 ∤ g(y). Since g(Y) does not have repeating irreducible factor in Z[T1, . . . , Tn−1] too, we use the induction p2 + Bn−2(cid:17) O(2B). As for the third sum, it is trivially bounded by Np2(A, B) and this reduces to case 1. Finally, we split the middle sum as follows: on n to bound the first sum by Od,n(cid:16) Bn−1 Xy∈Zn−1 Np(Ay, B) = Xy∈Zn−1 B B pg(y) Np(Ay, B) + Xy∈Zn−1 B O(2B). pg(y) Ay6≡0 in Z/pZ[Tn] pg(y) Ay≡0 in Z/pZ[Tn] Using lemma 10, we can bound the first sum by O(cid:18)Bn−1 p + Bn−2(cid:19) d(cid:18)2B p + 1(cid:19) = O(cid:18) Bn p2 + Bn−1(cid:19) . As for the second sum, since A is of case 1, the order of y ∈ (Z/pZ)n−1 such that Ay ≡ 0 in Z/pZ[Tn] is O(p(n−1)−2) and so the sum is bounded by O(pn−3)(cid:18)2B p + 1(cid:19)n−1 for large p ≤ B. O(2B) = O pn−3(cid:18) 3B p (cid:19)n−1 B! = O(cid:18)Bn p2(cid:19) , (cid:3) Lemma 12. Suppose F (T) ∈ Z[T] has total degree d and has no re- peating irreducible factor in Z[T]. Then for all integers k ≫ d we have #{ν ∈ Zn B F (ν) is k-free} ∼ γk,F (2B)n, where γk,F := Yprime p∈Z(cid:18)1 − ρF (pk) pnk (cid:19) is a non-zero convergent Euler product. Here, we adopt the convention that 0 is k-free for all integer k ≥ 2. Proof. In this proof, we will introduce some arbitrary constants ξ, ǫ > 0, and the implicit constants of all the big O's depend only on F , k, ξ 18 W. WONG and ǫ. We adapt the idea of Browning in section 4 of [1]. Let ξ > 0 be a constant and define the following sets: N (k) fr : = {ν ∈ Zn B F (ν) is k-free}, N (k) N (k) N (k) : = {ν ∈ Zn B (cid:12)(cid:12)(cid:12)(cid:12) nfr, 1 : =(cid:26)ν ∈ Zn B (cid:12)(cid:12)(cid:12)(cid:12)(cid:12) nfr, 2 : =(ν ∈ Zn B (cid:12)(cid:12)(cid:12)(cid:12) : =(cid:26)ν ∈ Zn B (cid:12)(cid:12)(cid:12)(cid:12) nfr : =(cid:26)ν ∈ Zn M (2) : = {ν ∈ Zn M (2) fr M (2) ξ F (ν) is not k-free, and for all prime p such that pkF (ν), we have ξ < p ≤ B F (ν) is not k-free, and for all prime p such that pkF (ν), we have p > ξ, and pkF (ν) for some prime p > B (cid:27) , ) , B if pkF (ν) then p > ξ} = N (k) fr ⊔ N (k) nfr, 1 ⊔ N (k) nfr, 2, F (ν) is k-free, and p2F (ν) for some prime ξ < p ≤ B (cid:27) , p2F (ν) for some prime ξ < p ≤ B (cid:27) , F (ν) is not k-free, and B p2F (ν) for some prime ξ < p ≤ B} = M (2) fr ⊔ M (2) nfr . Then obviously we have M (2) #N (k) ξ ≥ #N (k) We first estimate #N (k) fr ⊂ N (k) fr ≥ #N (k) fr and N (k) nfr, 1 ⊂ M (2) ξ − #M (2) − #N (k) nfr, 2. nfr and thus ξ with the help of Mobius function µ. Let ρF (m) := #{ν ∈ (Z/mZ)n F (ν) ≡ 0 (mod m)}. Then we can write #N (k) ph⇒p≤ξ ξ = Xh∈N = Xh∈N = Xh∈N ph⇒p≤ξ ph⇒p≤ξ µ(h)#{ν ∈ Zn B hkF (ν)} µ(h)ρF (hk)(cid:18)2B µ(h)ρF (hk) (2B)n hk + O(1)(cid:19)n hkn + O (cid:18) 2B hk(cid:19)n−1 + 1!! . Since the summation sums only square-free h, the condition ph ⇒ p ≤ ξ implies h ≤Yp≤ξ p = exp Xp≤ξ log p! ≤ e2ξ, AVERAGE VALUE OF THE CANONICAL HEIGHT 19 where the last inequality is gotten by the prime number theorem. More- over, it follows from the proof of Theorem 3.2 of Poonen in [10] that ρF (p2) = O(p2n−2) (or we can also deduce this from lemma 11 with B = p2). Subsequent lifting will lead to ρF (pj) = O(pjn−2) for j ≥ 2. Together with the fact that ρF is multiplicative, we have for square-free h that ρF (hk) = O(hnk−2+ǫ) for any ǫ > 0. The ǫ is needed here in order to bound the product of rh copies of the implicit constant of O(p2n−2), where rh is the number of distinct prime factors of h. Using the fact that h is square-free, we have h ≥ rh! and the Stirling's formula will give us the desired bound. With this, we obtain #N (k) ξ = (2B)nYp≤ξ(cid:18)1 − ρF (pk) pnk (cid:19)+O(cid:0)(2B)n−1e2ξ(k−1+ǫ) + e2ξ(nk−1+ǫ)(cid:1) . Again because ρF (pk) = O(pkn−2), the infinite product γk,F := Yprime p∈Z(cid:18)1 − ρF (pk) pnk (cid:19) converges and we have #N (k) ξ ≥ (2B)nγk,F + O(Bn−1). Next, by lemma 11, we have #M (2) ≤ Xξ<p≤B = Xξ<p≤B ≤ c(cid:18)Bn ξ #{ν ∈ Zn B p2F (ν)} O(cid:18)Bn p2 + Bn−1(cid:19) log B(cid:19) Bn + for some constant c > 0 depending only on n and d. The first term of the upper bound is obtained by integral estimate and the second is by the Prime Number Theorem. Lastly, #N (k) nfr, 2 = 0 for B big enough. In effect, there exists a constant CF > 0 such that F (ν) ≤ CF νd. Thus, for all B > CF , prime p > B, ν ∈ Zn B and k > d, we have pk > Bk ≥ Bd+1 > CF Bd ≥ F (ν), 20 W. WONG so no such pk divides F (ν). Combining together all the estimates, we get #N (k) fr ≥ γk,F (2B)n + O(Bn−1) − c ξ (2B)n + od,n(Bn) for B > max{ξ, CF }. The fact that ρF (pk) = O(pkn−2) implies γk,F converges and γk,F is zero if and only if one of its factors is zero. So in order to make γk,F > 0, it's sufficient to choose k big enough such that ρF (pk) < pnk for all prime p. More explicitly, choose a ν0 such in Z. Then that F (ν0) 6= 0 and look at its prime factorizationYi {βi, d} will do. We fix this k and for any λ ∈ (0, 1), by pβi i any k > max choosing ξ big enough so that c i ξ ≪ γk,F , we get fr ≥ λγk,F (2B)n for B big enough. Since for all ξ ≫ 0, #N (k) #N (k) #N (k) ξ ∼ (2B)nYp≤ξ(cid:18)1 − we get fr and ξ ≥ #N (k) ρF (pk) pnk (cid:19) , Yp≤ξ(cid:18)1 − ρF (pk) pnk (cid:19) ≥ lim sup B→∞ #N (k) fr (2B)n ≥ lim inf B→∞ #N (k) fr (2B)n ≥ λγk,F for all ξ ≫ 0 and λ ∈ (0, 1). Taking λ → 1 (which forces ξ → ∞ ) will complete the proof. Corollary 13. For any F (T) ∈ Z[T], then there exists an integer k0, such that for all integer k ≥ k0 and for all λ ∈ (0, 1), there exist B0 > 0 depending on F, k, λ and ck,F > 0 depending on F, k such that #{ν ∈ Zn B F (ν) is k-free} ≥ λck,F (2B)n (cid:3) f αi i where fi are distinct irreducible factors of whenever B ≥ B0. Proof. Write F = rYi=1 F in Z[T]. Let f := rYi=1 Now it is immediate by the previous lemma that for all k ≥ k′α, where fi with total degree d and α := max {αi}. i AVERAGE VALUE OF THE CANONICAL HEIGHT 21 k′ > d big enough as in the previous lemma, we obtain our corollary with ck,F = γk′,f > 0. (cid:3) Lemma 14. For any integer N ≥ 2, we denote frN (m) to be the N th- power-free part of the integer m, i.e. the smallest positive integer ℓ such that m is a N th-power of an integer. Suppose a primitive F (T) ∈ Z[T] ℓ is not a pth-power in C[T] for all prime pN. Then for all M > 2, we have #{ν ∈ Zn B (frN (F (ν)))M > ν} ∼ (2B)n. Proof. Let the total degree of F be d and hence there exists a constant CF ≥ 1 such that F (ν) ≤ CF νd. Define SM (F, B) := (ν, y, z) ∈ Zn+2 (cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) We will prove by induction that ν ≤ B, y ≤ (CF Bd) 0 < z ≤ B M , 1 1 N , F (ν) = yN z .  #SM (F, B) ≪F,ǫ,M,N C ǫ F B(n−1)+ 1 N + 1 M +2dǫ log B ∀ǫ > 0. The implicit constants of the big O's and small o's that appear in this proof will depend only on F , ǫ, M and N. We are going to apply Theorem 15 of Heath-Brown in [3], so we try to use notations that are coherent with it. For n = 1, for all z0 ∈ Z, let N(fz0, B, (CF Bd) Then 1 N ) :=(cid:26)(ν, y) ∈ Z2 (cid:12)(cid:12)(cid:12)(cid:12) #SM (F, B) = X0<z0≤B 1 M ν ≤ B, y ≤ (CF Bd) fz0(ν, y) := F (ν) − z0yN = 0 (cid:27) . N , 1 #N(fz0, B, (CF Bd) 1 N ). Since for all prime pN, F is not a pth-power in C[T1], the same holds in C(T1). By Capelli's lemma ([6] Chapter VI, Theorem 9.1), fz0(T1, Y ) = F (T1) − z0Y N is absolutely irreducible in C(T1)[Y ] for all z0 6= 0. We need this fact for the next step. Now we apply Theorem 15 of Heath-Brown in [3] on N(fz0, B, (CF Bd) Let T := max{Bd, CF Bd} = CF Bd. Then for all ǫ > 0, there exists a 1 N ). 22 W. WONG constant D = Dd,ǫ and k ∈ N with k ≪d,ǫ T ǫ exp(log B log(CF Bd) log(CF Bd) 1 N ) log fz0 ≪d,ǫ (CF Bd)ǫB 1 N log fz0, such that there exists f1, . . . , fk ∈ Z[T1, Y ], coprime to fz0 and with de- grees at most D, such that every (ν, y) counted by N(fz0, B, (CF Bd) N ) is a zero of some polynomial fi. By B´ezout's theorem, the number of points of intersection of curves fi = 0 and fz0 = 0 is bounded by deg fi · deg fz0 ≤ D(d + N). This gives immediately 1 #N(fz0, B, (CF Bd) 1 N ) ≪d,ǫ D(d + N)(CF Bd)ǫB 1 N log fz0. So (8) #SM (F, B) ≪d,ǫ,N X0<z0≤B ≪d,ǫ,N X0<z0≤B (CF Bd)ǫB 1 N log fz0 (CF Bd)ǫB 1 N log B 1 M 1 M ≪d,ǫ,N C ǫ F B 1 N + 1 M +dǫ log B, 1 where we may choose B ≥ F and hence fz0 ≤ B for z0 ≤ B M . Now we proceed to prove for a general n ≥ 2. For all x ∈ Zn−1, let Fx(Tn) := F (x, Tn) ∈ Z[Tn]. For all pN, since F (T) is not a pth-power, we look at the pth-power-free part of F (T) in Z[T], call it Gp(T). In other words, Gp(T) is the smallest degree polynomial such that F (T) Gp(T) is a pth-power in Z[T]. So Gp(T) =Qj Gp,j(T)βj where Gp,j are distinct irreducible factors and 0 < βj < p. Let gp(T) := Qj Gp,j(T), which has no repeated irreducible factor in Z[T]. Using Gauss' lemma on UFDs and by reindexing if necessary, the discriminant of gp(X, Tn) ∈ (Z[X])[Tn] is not a zero polynomial in Z[X]. So there are at most O(Bn−2) of x ∈ Zn−1 such that gp,x(Tn) := gp(x, Tn) has repeated irreducible factor in Z[Tn]. This will imply that there are at most O(Bn−2) of x ∈ Zn−1 such that Fx(Tn) is a pth-power in Z[Tn]. So we B B AVERAGE VALUE OF THE CANONICAL HEIGHT 23 have (9) #SM (F, B) = Xx∈Zn−1 B #SM (Fx, B) + Xx∈Zn−1 B Fx non-pth-power for all pN Fx pth-power for some pN #SM (Fx, B). Notice that Fx(νn) ≤ (CF Bd)νnd for x ∈ Zn−1 using the result from the case n = 1, we get B and deg Fx ≤ d. So #SM (F, B) ≪d,ǫ,N (2B)n−1(CF Bd)ǫB N + 1 ≪F,ǫ,M,N C ǫ F B(n−1)+ 1 M +2dǫ log B, 1 N + 1 M +dǫ log B + O(Bn−2 · B · B 1 M ) where for the estimation of the second sum, we use the fact that y is determined (up to sign for the case N is even) once (νn, z0) is fixed in Fx. When ǫ is sufficiently small relative to d and M > 2, we get #SM (F, B) = o(Bn). Lastly, define BadM (F, B) := {ν ∈ Zn B (frN (F (ν)))M ≤ ν}, which is the complement of {ν ∈ Zn B. It is a simple exercise to show that BadM (F, B) injects into SM (F, B) via B (frN (F (ν)))M > ν} in Zn the map ν 7−→ (ν, NpF (ν)/ frN (F (ν)), sign(F (ν)) frN (F (ν))), hence giving us the lemma. (cid:3) Corollary 15. With the same hypothesis as in the previous lemma and further let g ∈ N such that 0 < g ≤ B N+2 , then for M big enough, we have 1 #nν ∈ Zn B g (cid:12)(cid:12)(cid:12) (frN (F (ν)))M > gM (N−1)+1νo ∼(cid:18)2 B g(cid:19)n . In particular, when N = 2 or 3, then any M > 8 is admissible. Proof. The proof is just a slight modification of the previous proof, so we will continue using all the notations from the previous proof. Again, the implicit constants of the big O's and small o's in this proof depend only on F, ǫ, M and N. We are going to show that the complement of n ν ∈ Zn B g (cid:12)(cid:12)(cid:12) (frN (F (ν)))M > gM (N−1)+1νo , 24 which is W. WONG BadM (F, B, g) :=n ν ∈ Zn has order o(cid:16)2 B g(cid:17)n into B g (cid:12)(cid:12)(cid:12) (frN (F (ν)))M ≤ gM (N−1)+1νo , . Just like the previous proof, BadM (F, B, g) injects N ν ≤ B g , SM (F, B, g) :=  (ν, y, z) ∈ Zn+2 (cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) y ≤(cid:18)CF(cid:16) B g(cid:17)d(cid:19) 1 M(cid:16) B g(cid:17) 1 g(cid:17)n Thus, it suffices to show that #SM (F, B, g) = o(cid:16)2 B . Comparing SM (F, B, g) to SM (F, B) from the previous proof, this boils down to just changing B to B g and slightly increasing the upper bound for z with a factor of gN−1+ 1 M . So for n = 1, from (8), we have 0 < z ≤ gN−1+ 1 , F (ν) = yN z M , .  #SM (F, B, g) ≪d,ǫ,N #SM (F, B, g) ≪d,ǫ,N C ǫ where we choose B ≥ F and hence fz0 ≤ max{B, gN−1B 1 N log fz0 1 M N + 1 M +dǫ g ) M ( B M log B, gN−1+ 1 M } ≤ B2. ≪d,ǫ,N C ǫ CF(cid:18) B g(cid:19)d!ǫ(cid:18)B g(cid:19) 1 N+2 , we have B X0<z0≤gN −1+ 1 F(cid:18)B g(cid:19) 1 g(cid:17) N+2 g(cid:17)2 g ≥ gN +1 and B ≤(cid:16) B N+1 <(cid:16) B F(cid:18)B g(cid:19) N −1 g(cid:19) 1 g(cid:19) , log(cid:18)B M +dǫ(cid:18) B g(cid:17) for M sufficiently big and ǫ sufficiently small. Using g(cid:19)n−1 CF(cid:18)B g(cid:19)n−2 g(cid:19)d!ǫ(cid:18)B g(cid:19) 1 M! M(cid:18) B g(cid:19) 1 log(cid:18)B g(cid:19) · gN−1+ 1 M +dǫ+ N −1 N + 1 N + 1 N+1 + B g N+1 + (N+1)M 1 1 (N+1)M , so · Since g ≤ B 1 which is o(cid:16)2 B #SM (F, B, g) B ≪d,ǫ,M,N (cid:18)2 + O (cid:18)B the same induction argument as in (9), we have for n ≥ 2, AVERAGE VALUE OF THE CANONICAL HEIGHT 25 (10) ≪F,ǫ,M,N C ǫ g(cid:19)n−1+ 1 F(cid:18)B which is also o(cid:16)2 B g(cid:17)n N + 1 M +2dǫ+ N −1 N+1 + 1 (N+1)M g(cid:19) , log(cid:18)B for M sufficiently big and ǫ sufficiently small. We remark that for the case N = 2, 3, the exponent of B g in (10) are n − 1 + 2dǫ + 5 4M respectively. For any M > 8, there exists ǫ > 0 such that this exponent is strictly less than n, hence giving us corollary 15 for the case N = 2, 3 and these are the instances where we will apply this corollary. (cid:3) 3M and n − 1 + 2dǫ + 5 6 + 5 6 + 4 5. Proof of Proposition 3 We keep all the notations as previously defined in this paper. The implicit constants of the big O's and small o's that appear in this proof will depend only on ∆E, n and P . The main idea in this proof is to first apply lemma 6. This allows us to get a lower bound of hEν (Pν) Eν , for all "nice" ν ∈ Zn. Then we try to bound ∆min in term of ∆min Eν below in term of ∆Eν and then in term of h(ν), again for all "nice" ν. The nontrivial part of the proof is to show that after we impose again and again certain niceness conditions on ν, this set of of "nice" ν has a positive density in Zn B(∆E, P ). Fix a big integer k ≥ 4, which we will specify how big it should be at the end of the proof and let Nk := lcm(1, 2, 3, . . . , k). Then by lemma 6, there is an absolute constant C1 > 0 such that for any P ∈ E(K)nt and for any ν ∈ Zn B(∆E, k, P nt ν ), where Zn B(∆E, k, P nt ν ) := {ν ∈ Zn B(∆E, P ) ∆E(ν) is k-free, Pν ∈ Eν(Q)nt}, we have (11) Xν∈Zn B (∆E ,P ) hEν (Pν) h(ν) ≥ = B (∆E,k,P nt ν ) C1 N 2 C1 N 2 k Xν∈Zn k Xν∈Zn B (∆E ,k,P nt ν ) log ∆min Eν h(ν) log ∆min Eν log ν . We obtain the second line because of the convention that we made earlier : h(ν) = log H([1, ν1, . . . , νn]). 26 W. WONG Next, we claim that ∆E(T) = −16(4A3(T) + 27B2(T)) is never a constant times a twelfth power in Z[T], otherwise lemma 9 says that there exists g(T) ∈ Q[T] such that A(T) g(T)6 ∈ Z. So using Gauss' lemma, we can write ∆E(T) = α(F (T))a+12b, where α ∈ Z, F (T) is primitive in Z[T], non-power in C[T], a ∈ {1, 2, 3, . . . , 11} and b ∈ N. Recall that for all primes p in Q, g(T)4 , B(T) 0 ≤ ordp(∆min Eν ) ≡ ordp(∆Eν ) (mod 12), so ordp(∆min Eν ) is at least the unique integer in {0, 1, 2, . . . , 11} congru- ent to ordp(∆Eν ) (mod 12). We split into two cases in order to get a lower bound of ∆min Eν . Case 1: a 6= 4, 8. If we let sqfr(m) := fr2(m) and sq(m) := m the square-free part and square part of an integer m, then we have sqfr(m) be ∆E(ν) = α sq(F (ν))a+12b sqfr(F (ν))a+12b . Notice that for every prime factor p of sqfr(F (ν)) that is relatively prime to α, its power βp in F (ν) is odd and thus aβp 6≡ 0 (mod 12) for a 6= 4, 8. So p is a factor of ∆min ν ), Eν and we have for ν ∈ Zn B(∆E, k, P nt ∆min Eν ≥ sqfr(F (ν)) α . Case 2: a = 4 or 8. The argument is similar to case 1 except that we look at cufr(F (ν)) := fr3(F (ν)), the cube-free part of F (ν). Then for every prime factor p of cufr(F (ν)) that is relatively prime to α, its power βp in F (ν) is not a multiple of 3 and thus aβp 6≡ 0 (mod 12) for a = 4 or 8. In fact, aβp ≡ 4 or 8 (mod 12). So again, for ν ∈ Zn B(∆E, k, P nt ν ), we have ∆min Eν ≥ cufr(F (ν)) α2 . Fix M > 2 and let GoodM (F, B) := (cid:8)ν ∈ Zn (cid:8)ν ∈ Zn B(∆E) sqfr(F (ν))M > ν(cid:9) if a 6= 4, 8 B(∆E) cufr(F (ν))M > ν(cid:9) if a = 4, 8. Then from inequality (11), we have AVERAGE VALUE OF THE CANONICAL HEIGHT 27 log ν 1 M − log α2 log ν o(Bn), Xν∈Zn B (∆E ,P ) hEν (Pν) h(ν) ≥ = C1 N 2 k C1 N 2 k M X X ν∈Zn B (∆E ,k,P nt ν )∩GoodM (F,B) ν∈Zn B (∆E,k,P nt 1 + ν )∩GoodM (F,B) where we use lemma 7 to bound the sum of the second term. Now we are at the final step of analyzing the asymptotic cardinal of the set Zn ν ) ∩ GoodM (F, B). It is straightforward that B(∆E, k, P nt #Zn B(∆E, P ) ∼ (2B)n as the set of points for which ∆E(T) vanishes or Pν is not defined is of order at most O(Bn−1). Next, by Mazur's theorem ([14] Chapter VIII, Theorem 7.5), the order of Eν(Q)tor is at most 12. Hence if Pν is torsion, ν must satisfy one of the twelve algebraic equations of torsion points that depends on P . Since P is non-torsion in E(K), none of the twelve equations is identically zero and so #{ν ∈ Zn H(ν) ≤ B and Pν is torsion} = O(Bn−1). This gives #Zn B(∆E, P nt ν ) := #{ν ∈ Zn B(∆E, P ) Pν ∈ Eν(Q)nt} ∼ (2B)n. We now apply corollary 13 to ∆E(T), and we specify that k is big enough such that ck,∆E > 0 as in the corollary. Then for any λ ∈ (0, 1) and for B big enough, we get #Zn B(∆E, k, P nt ν ) ≥ λck,∆E (2B)n. Lastly, since F is primitive and is neither a square nor cube in Z[T], we use lemma 14 to conclude and this give us B(∆E, k, P nt ν ) ∩ GoodM (F, B)(cid:1) ≥ λck,∆E (2B)n hEν (Pν) h(ν) ≥ C1 N 2 k M λck,∆E(2B)n + o(Bn) #(cid:0)Zn Xν∈Zn B (∆E ,P ) This proves Proposition 3 with a lower bound C1 k M λck,∆E for any λ ∈ N 2 (0, 1) and M > 2, hence we can take L2 = C1 ck,∆E . 2N 2 k 28 W. WONG 6. Proof of Theorem 1 The idea of this proof is to reduce to the case of Proposition 3, since a point in Pn Q can be represented with integers coordinates. Again, the implicit constants of the big O's and small o's that appear in this proof will depend only on ∆E, n and P , unless stated otherwise. Let vn(cid:17) ∈ Qn in the lowest form and let ℓω := lcm(v1, . . . , vn). Recall that the Weierstrass equation of Eω is ω =(cid:16) u1 v1 , . . . , un Y 2 = X 3 + A(ω)X + B(ω), which might not have integer coefficients. In order to estimate ∆min Eω , we need to look at a Weierstrass equation that is Q-isomorphic to Eω with integer coefficients. Let d be the maximum of deg A and deg B. By a change of variable Y ′ = ℓ3d ω X, we obtain an integral coefficients Weierstrass equation: ω Y and X′ = ℓ2d Y ′2 = X′3 + ℓ4d ω A(ω)X′ + ℓ6d ω B(ω) with discriminant ∆′Eω := −16ℓ12d ω (4A(ω)3 + 27B(ω)2) = ℓ12d ω ∆E(ω). Let us set up the following correspondence to ease our argument. If we write δαT α1 1 . . . T αn n , δαT 12d−α 0 T α1 1 . . . T αn n . then let ∆E(T) = Xα≤d e∆E(T0, T) := Xα≤d Notice that e∆E is a homogeneous polynomial of degree 12d. We have a one-to-one correspondence between Qn B(∆E) = {ω ∈ Qn 1 < H(ω) ≤ B and ∆E(ω) 6= 0} and {ν = (ν0, ν1, . . . , νn) ∈ Zn+1 via the map B (e∆E) gcd(ν0, ν1, . . . , νn) = 1 and ν0 > 0} ω 7−→ ν :=(cid:18)ℓω, u1ℓω v1 , . . . , unℓω vn (cid:19) AVERAGE VALUE OF THE CANONICAL HEIGHT 29 This correspondence gives Xω∈Qn B(∆E ,P ) hEω (Pω) h(ω) = hE( ν1 ′ 1 2 Xν∈Zn+1 B (e∆E ,P ) gcd ν=1 (P(cid:16) ν1 ν0 ,..., νn ν0 (cid:17)) ,..., νn ν0 ) ν0 h([ν0, . . . , νn]) , where gcd ν := gcd(ν0, . . . , νn) and the primed summation means ν0 6= 0 with the factor 1 2 taking care of the negative ν0. In order to use the inclusion-exclusion argument effectively in the later part, we need to Zn+1 which is a homogeneous polynomial too and let modify the estimate on the set of ν for which e∆E(ν) is k-free. Let sqfre∆E(T0, T) := fE(T0, T), B (e∆E, P ) (cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) B (e∆E, k, P nt) )  := the exponents of distinct irreducible factors of e∆E(T0, T), then for all B (e∆E, k, P nt) with gcd ν = 1, let ω =(cid:16) ν1 ν0(cid:17) and we have ν0 6= 0, fE(ν) is k-free, Pω ∈ Eω(Q)nt where ω = ( ν1 ν0 for some k ≥ 4 big enough as in lemma 12. If α is the maximum of ν = (ν0, . . . , νn) ∈ Zn+1 , . . . , νn ν0 ν ∈ Zn+1 is kα-free. , . . . , νn ν0 , e∆E(ν) = ∆′Eω Thus, letting Nk := lcm(1, . . . , kα) and using the same argument as in lemma 6, we get Xω∈Qn B(∆E ,P ) hEω (Pω) h(ω) ≥ 1 2 Xν∈Zn+1 ′ C1 N 2 k B (e∆E ,k,P nt) gcd ν=1 log ∆min Eω log ν . Notice that e∆E(T0, T) is not a constant times a twelfth power in Z[T0, T], otherwise it will imply the same for e∆E(1, T) = ∆E(T). Just like in the proof of Proposition 3, we writee∆E(T0, T) = β(F (T0, T))a+12b, where β ∈ Z, F (T0, T) is primitive homogeneous in Z[T0, T] and non- power in C[T0, T], b ∈ N and a ∈ {1, 2, 3, . . . , 11}. Since the same property 0 ≤ ordp(∆min Eω ) ≡ ordp(e∆E(ν)) (mod 12) 30 W. WONG still hold for all prime p in Q, we can repeat the corresponding whole argument as in section 5 and get (12) Xω∈Qn B (∆E ,P ) hEω (Pω) h(ω) ≥ where GoodM (F, B) := nν ∈ Zn+1 nν ∈ Zn+1 B 1 1 ν∈Zn+1 1 M ′ C1 N 2 k B (e∆E ,k,P nt) gcd ν=1 + O Xν∈Zn+1 log ν 2 Xν∈GoodM (F,B) B (e∆E) (cid:12)(cid:12) sqfr(F (ν))M > νo if a 6= 4, 8 B (e∆E) (cid:12)(cid:12) cufr(F (ν))M > νo if a = 4, 8, = Xν∈GoodM (F,B) BXg=1 ′ Xg gcd ν µ(g) Xν∈GoodM (F,B) ν∈Zn+1 B (e∆E,k,P nt) µ(g) 1. = ′ ν∈Zn+1 B (e∆E ,k,P nt) g gcd(ν) for any fixed M > 2. We know from lemma 7 that the second term is o(Bn+1). As for the first term, we estimate it by an inclusion-exclusion argument using the Mobius function: ′ Xν∈GoodM (F,B) 1 ν∈Zn+1 B (e∆E ,k,P nt) gcd ν=1 (13) To deal with the inner sum, we have to analyse the sets of which we are summing over. Recall that F is a homogeneous polynomial, so we have F (gν) = gtF (ν) where t = deg F and the trivial inequalities sqfr(F (gν)) ≥ cufr(F (gν)) ≥ sqfr(F (ν)) g cufr(F (ν)) g2 , . These imply the following inclusions: GoodM (F, B, g) :=nν ∈ Zn+1 B g g ·(cid:26)ν ∈ Zn+1 g ·(cid:26)ν ∈ Zn+1 B g =  B (e∆E)(cid:12)(cid:12)(cid:12) g gcd(ν)o ∩ GoodM (F, B) (e∆E)(cid:12)(cid:12)(cid:12)(cid:12) sqfr(F (gν))M > gν(cid:27) if a 6= 4, 8 (e∆E)(cid:12)(cid:12)(cid:12)(cid:12) cufr(F (gν))M > gν(cid:27) if a = 4, 8 AVERAGE VALUE OF THE CANONICAL HEIGHT 31 M M > gν) if a 6= 4, 8 > gν) if a = 4, 8, B g g ·(ν ∈ Zn+1 g ·(ν ∈ Zn+1 B g ⊇  (cid:12)(cid:12)(cid:12)(cid:12) (cid:12)(cid:12)(cid:12)(cid:12) g g2 (cid:12)(cid:12)(cid:12)(cid:12) (cid:12)(cid:12)(cid:12)(cid:12) sqfr(F (ν)) cufr(F (ν)) (e∆E) (cid:12)(cid:12)(cid:12)(cid:12)(cid:12) (e∆E) (cid:12)(cid:12)(cid:12)(cid:12)(cid:12) GoodM (F, B, g) ∼(cid:18)2 1 where the notation g · S means {gν ν ∈ S} for any set S of vectors. By Corollary 15, for g ≤ B 5 and M > 8, we have (14) B g(cid:19)n+1 . On the other hand, fE is also homogeneous. Let the degree of fE be r and we have Zn+1 B (e∆E, k, P nt, g) :=nν ∈ Zn+1 = g · ν = (ν0, . . . , νn) ∈ Zn+1 B (e∆E)(cid:12)(cid:12)(cid:12) g gcd(ν)o ∩ Zn+1 B (e∆E, k, P nt) (e∆E, P ) (cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) B (e∆E, k, P nt, g) ⊆ g · Zn+1 (e∆E, k − r, P nt) ⊆ Zn+1 For µ(g) 6= 0, i.e. g is squarefree, we have the inclusions g · Zn+1 B g B g B g .  (e∆E, k, P nt). ν0 6= 0, grfE(ν) is k-free, Pω ∈ Eω(Q)nt where ω = ( ν1 ν0 , . . . , νn ν0 ) From (14), lemma 12 and Mazur's theorem again, we have for any ǫ > 0, there exists Bǫ such that if B 5 and µ(g) 6= 0 then g ≥ Bǫ, g ≤ B 1 γk−r,fE − ǫ ≤ #(cid:16)Zn+1 B (e∆E, k, P nt, g) ∩ GoodM (F, B, g)(cid:17) (cid:16)2 B g(cid:17)n+1 ≤ γk,fE + ǫ. It is important to remark that the implicit constants of the big O's that appear in the rest of the proof depend only on n and nothing else. 32 W. WONG From (13), for B > B ′ 5 4 ǫ , Xν∈GoodM (F,B) 1 ν∈Zn+1 B (e∆E ,k,P nt) gcd ν=1 is bounded below by µ(g)(γk,fE + ǫ)(cid:18)2 B g(cid:19)n+1 1 5 ′ 1 ν∈Zn+1 B (e∆E ,k,P nt,g) µ(g) Xν∈GoodM (F,B,g) g(cid:19)n+1 5Xg=1 + B 1 B µ(g)=−1 jB 1 5k ′ 1 B ≥ µ(g)=1 1 + ν∈Zn+1 B (e∆E,k,P nt,g) BXg=B Xg=1 µ(g) Xν∈GoodM (F,B,g) µ(g)(γk−r,fE − ǫ)(cid:18)2 5Xg=1 + O g(cid:19)n+1 5(cid:18)2 BXg=B 5Xg=1 = (γk−r,fE − ǫ)(2B)n+1 B B 1 1 µ(g) gn+1 + (2ǫ + γk,fE − γk−r,fE)(2B)n+1 1 B 5Xg=1 µ(g)=−1 µ(g) gn+1 So we get 1 5 1 gn+1 . + (2B)n+1O BXg=B (2B)n+1 Xω∈Qn M (cid:18)(γk−r,fE − ǫ) C1 N 2 k B(∆E ,P ) 1 2 1 1 1 lim inf B→∞ ≥ hEω (Pω) h(ω) + (2ǫ + γk,fE − γk−r,fE)O(1)(cid:19) ζ(n + 1) where ζ is the Riemann zeta function and one possible bound for the O(1) here is ζ(n + 1). So lim inf B→∞ 2ζ(n + 1) (2B)n+1 Xω∈Qn B (∆E ,P ) hEω (Pω) h(ω) AVERAGE VALUE OF THE CANONICAL HEIGHT 33 ≥ C1 MN 2 k ((γk−r,fE − ǫ) + (2ǫ + γk,fE − γk−r,fE)O(1)) . Since this holds for all ǫ > 0, M > 8 and the same inclusion-exclusion argument will give #{ω ∈ Qn H(ω) ≤ B} ∼ (2B)n+1 2ζ(n + 1) , we have proven Theorem 1 with L1 = C1 (γk−r,fE +(γk,fE−γk−r,fE)O(1)). 8N 2 k Notice that L1 is positive for k big enough because the sequence (γk,fE)∞k=1 is increasing and bounded above by 1. 7. Discussion Our proofs of Proposition 3 and Theorem 1 use the weakened form of Lang's height conjecture proven by Silverman mentioned in lemma 6, which is a key tool in our proof that there is a positive density γk−r,fE + (γk,fE − γk−r,fE)O(1) of ω ∈ Qn(∆E) such that log ∆min Eω > M h(ω) + OE(1)(cid:1) and hence hEω (p) ≥ C1 M h(ω) + OE(1)(cid:1) for all p ∈ (cid:0) 1 k(cid:0) 1 Eω(Q)nt. If we denote N 2 µ(ω) := minnhEω (p) p ∈ Eω(Q)nto , then we can get a uniform upper bound of the quotient h(ω) µ(ω) for a positive density of ω ∈ Qn(∆E). In view of this, we can apply this to the following theorem to say something about the integral points on Eω. Theorem 16. ([15] Cor.4.2) Set the following notations: F a number field. S a finite set of places of the absolute values of F . RS(ǫ) = (cid:8)x ∈ F Pv∈S max(−v(x), 0) ≥ ǫh(x)(cid:9) , so in particular RS(1) = RS, the ring of S-integers of F . T /F a quasi-projective variety and hT the height on T correspondind to a fixed ample divisor, chosen so that hT (t) ≥ 1 for all t ∈ T (F ). C/F an algebraic family of smooth, irreducible, projective curves over T , i.e. there is a F -morphism π : C → T which is proper and smooth of relative dimension 1; each fiber Ct is a smooth irreducible projective curve. 34 W. WONG J/F the Jacobian of C/T , so J is an abelian acheme over T . Let D ∈ DivF (J) a very ample and symmetric divisor. For each t ∈ T , the fiber Jt is the Jacobian variety of the fiber Ct and let ρ(t) := rank Jt(F ), τ (t) := #Jt(F )tors, µ(t) := minnhJ,D(p) p ∈ Jt(F ), P non-torsiono . Let ǫ > 0 and f ∈ F (C) be a non-constant rational function on C with the following property: the map [f : 1] : C −→ P1 is a morphism. Then there exists a constant c depending only on [F : Q], ǫ, C, f and T , such that for all t ∈ T (F ), the size of the set {p ∈ Ct(F ) f (p) ∈ Rs(ǫ)} is at most τ (t)c1+#S+ρ(t)(cid:18)hT (t) µ(t)(cid:19) ρ(t) 2 . Corollary 17. With the setting and notations as in the proof of The- orem 1 and let 0 < δ < 1 and M > 8, then there exists a constant c depending only on E, such that the set #Eω(Z) ≤ 16c2+rank Eω(Q) s MN 2 k (1 − δ)C1!rank Eω(Q)  ω∈Qn(∆E) (cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) has density at least γk−r,fE + (γk,fE − γk−r,fE)O(1). Proof. We will apply Theorem 16 for F = Q, S = { · ∞}, ǫ = 1 and so RS(ǫ) = Z. Let T ⊂ Qn(∆E) be a quasi-projective variety such that we can define a smooth group scheme C over T associated to our elliptic curve E/K: C : Y 2Z = X 3 + A(T)XZ 2 + B(T)Z 3. Notice that since each fiber Ct = Et is an elliptic curve, it is equal to its Jacobian Jt. We will use the rational function f = x = X Z ∈ Q(C) and we need to show that [x : 1] : C −→ P1 AVERAGE VALUE OF THE CANONICAL HEIGHT 35 is a morphism. The map is clearly defined on all points with Z 6= 0. From the equation for C, we have [x : 1] = [X : Z] = [Y 2 − B(T)Z 2 : X 2 + A(T)Z 2]. Since any point of C with Z = 0 will have the form ([0 : 1 : 0], t), we see that [x : 1] will map such a point to [1 : 0]. So [x : 1] defines a morphims and Theorem 16 says that there exists a constant c depending on E, such that for all ω ∈ T (Q) ⊂ Qn(∆E), we have # {p ∈ Eω(Q) x(p) ∈ Z} ≤ τ (ω)c1+1+ρ(ω)(cid:18) h(ω) ≤ 16c2+rank Eω(Q)(cid:18) h(ω) µ(ω)(cid:19) ρ(ω) µ(ω)(cid:19) rank Eω (Q) 2 2 , (15) where we obtain the second inequality by bounding τ (ω) ≤ 16 using Mazur's theorem. From the proof (See inequality (12)) of Theorem 1, we have for any M > 8, there is a positive density γk−r,fE + (γk,fE − γk−r,fE)O(1) of ω ∈ Qn(∆E) such that µ(ω) > C1 N 2 k (cid:18) 1 M h(ω) + OE(1)(cid:19) . For any 0 < δ < 1, the set of bounded height Bδ :=(cid:26)ω ∈ Qn (cid:12)(cid:12)(cid:12)(cid:12) h(ω) ≤ k (cid:18)1 − δ h(ω) + δ M M C1 N 2 is a finite set. By excluding these finite points, we still have a positive density γk−r,fE + (γk,fE − γk−r,fE)O(1) of ω ∈ Qn(∆E)\Bδ such that h(ω). (16) µ(ω) > (1 − δ) h(ω) + OE(1)(cid:19) ≥ C1 N 2 k MOE(1) δ (cid:27) Since T is a dense Zariski open subset of Qn(∆E), from inequalities (15) and (16), we have a positive density γk−r,fE + (γk,fE − γk−r,fE )O(1) of ω ∈ (Qn(∆E)\Bδ) ∩ T (Q) such that M #Eω(Z) ≤ # {p ∈ Eω(Q) x(p) ∈ Z} ≤ 16c2+rank Eω(Q)(cid:18) h(ω) ≤ 16c2+rank Eω(Q)(cid:18) MN 2 µ(ω)(cid:19) rank Eω (Q) (1 − δ)C1(cid:19) rank Eω (Q) k 2 2 , 36 W. WONG which completes the proof of the corollary. (cid:3) We remark that if the Lang's conjecture is true, then we can im- 2 and L1 = C1 8 , prove both Proposition 3 and Theorem 1 to L2 = C1 independent of E. Also, corollary 17 will be improved to density 1. One might be interested to ask whether we can generalize our initial In order to do that, we first setting of Q to any number field F . have to replace Z to F integers OF in Proposition 3 and scrutinize all the lemmas used in the proof to see whether they are still valid in F . Lemma 6 can be easily generalized to F as both the Silverman [11] and Kodaira-N´eron Theorems [16] were originally proven for number fields. Further, lemmas 7, 9 generalize immediately, Mazur's theorem also has its generalized counterpart, Merel's Theorem. Alternatively, we can use the following Masser's bound (we thank the referee for pointing this out). Using methods from transcendence theory, Masser obtained the upper bound ([8] Corollary 2) #E(K)tor ≤ Ckph([1, g1, g2])[K : k] (h([1, g1, g2]) + log[K : k]) for elliptic curve E/k : y2 = 4x3 − g1x − g3, where Ck is an effective constant that depends only on the number field k and K/k is any finite field extension. Hence, applying this to our setting over the number field F , we can obtain easily that for all ν ∈ (OF )n B, #Eν(F )tor ≤ CF (h([1, 4A(ν), 4B(ν)])) 3 2 ≤ C′F (log B) 3 2 where C′F is an effective constant that depends on F and the polyno- mials A(T), B(T). This is sufficient for our application as it gives us the bound #{ν ∈ On F H(ν) ≤ B and Pν is torsion} = O(Bn−1(log B) 3 2 ) = o(Bn). Besides having the advantage of a computable effective constant, Masser's bound is also true for general abelian varieties ([9] Main Theorem and Scholium 2). What are left to be worked on are lemmas 12 and 14. Another, and possibly more interesting problem is to prove convergence of the average, or even better, to prove the average converges to hE(P ). AVERAGE VALUE OF THE CANONICAL HEIGHT 37 Acknowledgements. I would like to thank my advisor, Joseph Sil- verman, for many enlightening discussions. Also, many thanks to the referee for the valuble and insightful comments and suggestions. References [1] T.D.Browning, Power-free Values of Polynomials, Arch. Math. 96 (2011), 139 -- 150. [2] G.H. Call, Variation of Local Heights on an Algebraic Family of Abelian Varieties, Th´eorie des nombres (Quebec, PQ, 1987), de Gruyter, Berlin, 1989, 72 -- 96. [3] D.R. Heath-Brown, Counting Rational Points on Algebraic Vari- eties, in: Analytic Number Theory, Lecture Notes in Math. 1891, Springer-Verlag, Berlin, 2006, 51 -- 95. [4] B. Conrad, Chow's K/k-image and K/k-trace, and the Lang- N´eron Theorem, Enseign. Math. 52 (2006), 37-108. [5] A.W. Knapp, Advanced Algebra, Cornerstones, Springer, Boston, 2007. [6] S. Lang, Algebra, GTM 211, Springer-Verlag, New York, 2002. [7] S. Lang, Fundamentals of Diophantine Geometry, Springer- Verlag, New York, 1983. [8] D.W.Masser, Counting Points of Small Height on Elliptic Curves, Bull. Soc. Math. France 117 (1989), 247 -- 265. [9] D.W.Masser, Specializations of Finitely Generated Subgroups of Abelian Varieties, Trans. Amer. Math. Soc. 311 (1989), 413 -- 424. [10] B.Poonen, Squarefree Values of Multivariable Polynomials, Duke Math. J. 118 (2003), 353 -- 373. [11] J.H. Silverman, Lower Bound for the Canonical Height on Elliptic Curves, Duke Math. J. 48 (1981), 633 -- 648. [12] J.H. Silverman, Heights and the Specialization Map for Families of Abelian Varieties, J. Reine Angew. Math. 342 (1983), 197 -- 211. [13] J.H. Silverman, Divisibility of the Specialization Map for Families of Elliptic Curves, Amer. J. of Math. 107 (1985), 555 -- 565. [14] J.H. Silverman, The Arithmetic of Elliptic Curves, GTM 106, Springer-Verlag, New York, 1986. 38 W. WONG [15] J.H. Silverman, A Quantitative Version of Siegel's Theorem: In- tegral Points on Elliptic Curves and Catalan Curves, J. Reine Angew. Math. 378 (1987), 60 -- 100. [16] J.H. Silverman, Advanced Topics in the Arithmetic of Elliptic Curves, GTM 151, Springer-Verlag, New York, 2007. [17] W.W. Stothers, Polynomial identities and Hauptmoduln, Quart. J. Math. Oxford Ser. (2) 32 (1981), 349 -- 370. [18] J. Tate, Variation of the Canonical Height of a Point Depending on a Parameter, Amer. J. Math. 105 (1983), 287 -- 294. Mathematics Department, Box 1917 Brown University, Providence, RI 02912 USA E-mail address: [email protected]
1303.4564	3	1303	2013-12-18T10:59:15	Curves and cycles on K3 surfaces	[ "math.AG" ]	The notion of constant cycle curves on K3 surfaces is introduced. These are curves that do not contribute to the Chow group of the ambient K3 surface. Rational curves are the most prominent examples. We show that constant cycle curves behave in some respects like rational curves. E.g. using Hodge theory one finds that in each linear system there are at most finitely many such curves of bounded order. Over finite fields, any curve is expected to be a constant cycle curve, whereas over number fields this does not hold. The relation to the Bloch--Beilinson conjectures for K3 surfaces over global fields is discussed.	math.AG	math	CURVES AND CYCLES ON K3 SURFACES D. HUYBRECHTS WITH AN APPENDIX BY C. VOISIN Abstract. The notion of constant cycle curves on K3 surfaces is introduced. These are curves that do not contribute to the Chow group of the ambient K3 surface. Rational curves are the most prominent examples. We show that constant cycle curves behave in some respects like rational curves. E.g. using Hodge theory one finds that in each linear system there are at most finitely many such curves of bounded order. Over finite fields, any curve is expected to be a constant cycle curve, whereas over ¯Q this does not hold. The relation to the Bloch–Beilinson conjectures for K3 surfaces over global fields is discussed. Contents Introduction 1. 2. Motivation 3. Constant cycle curves 4. Constant cycle curves on other surfaces 5. Finiteness of constant cycle curves of fixed order 6. First examples of constant cycle curves 7. More examples: fixed curves 8. Bitangent correspondence 9. Finite fields 10. Further questions and remarks 11. Appendix by Claire Voisin References 2 3 7 14 16 23 26 29 34 37 38 42 This work was supported by the SFB/TR 45 ‘Periods, Moduli Spaces and Arithmetic of Algebraic Varieties’ of the DFG (German Research Foundation). 1 1. Introduction 1.1. Due to results of Mumford [Mu68] and Bloch [Bl80], the Chow group of zero-cycles CH0(X) = CH2(X) on a complex K3 surface X is known to be huge (infinite dimensional In particular, there is no curve C ⊂ X such that the natural in some well-defined sense). push-forward map (1.1) Pic(C) ≃ CH0(C) / CH2(X) is surjective. This paper studies curves C ⊂ X for which the image of Pic(C) small as possible, i.e. for which (1.1) induces a trivial map Pic0(C) / CH2(X). / CH2(X) is as For a K3 surface over an algebraically closed field k, we define an integral curve C ⊂ X to be a constant cycle curve if the class of its generic point ηC ∈ C viewed as a closed point in Xk(ηC ) = X ×k k(ηC ) satisfies (1.2) n · [ηC] = n · (cX )k(ηC ) in CH2(X ×k k(ηC )) for some positive integer n. Here, cX ∈ CH2(X) is the distinguished class of degree one introduced by Beauville and Voisin in [BV04]. The minimal such n is called the / CH2(X) order of the constant cycle curve. If C ⊂ X is a constant cycle curve, then Pic0(C) will be shown to be indeed the zero map. See Sections 3.1 and 3.2 for the definitions and Proposition 3.7 for the relation between the two notions. Finding sufficient criteria that decide whether a given curve is a constant cycle curve seems as hard as finding criteria that would ensure the opposite. The only positive criterion at the time being is that any rational curve is a constant cycle curve (of order one) and the only effective method to exclude a given curve from being a constant cycle curve uses Hodge theory (cf. Corollary 5.5). 1.2. Some of the results proved in this article are: • There are at most finitely many constant cycle curves of bounded order in each linear system L on a K3 surface X in characteristic zero (cf. Proposition 5.1). The enumerative problem that suggests itself at this point remains largely open, but see Section 10.2. To get a better idea of the notion of a constant cycle curve, we give many concrete examples with an emphasis on curves of low order and high genus, see Sections 6–8. Apart from curves of torsion points in families of elliptic curves, the following result turns out to be useful. • Every fixed curve of a non-symplectic automorphism is a constant cycle curve (see Propo- sition 7.1). This immediately leads to: • There are constant cycle curves of order one, that are not rational (see Corollary 7.4). 2 / / / / We also manage to construct a constant cycle curve in the generic quartic hypersurface X ⊂ P3 that is of order at most four and genus 201, see Proposition 8.7. The notion of constant cycle curves makes sense for arbitrary surfaces. In Section 4 we briefly study surfaces with pg = 0 satisfying Bloch’s conjecture. It can easily be shown that curves on such surfaces are all constant cycle curves of bounded order and • On an Enriques surface all curves are constant cycle curves of order at most four (see Proposition 4.2). In Section 9 we discuss K3 surfaces over finite fields and prove: • For a Kummer surface X over ¯Fp every curve in X is a constant cycle curve (cf. Proposition 9.4). This is expected to hold for all K3 surfaces over ¯Fp and is related to the conjectured finite- dimensionality in the sense of Kimura–O’Sullivan and the Bloch–Beilinson conjecture for K3 surfaces over global fields in positive characteristic (see Proposition 9.2). For arbitrary K3 surfaces over ¯Fp one can at least prove the following: • Let X be a K3 surface over ¯Fp. Then every closed point x ∈ X is contained in a constant cycle curve (see Proposition 9.6). This is expected to hold as well for X over ¯Q and would imply the Bloch–Beilinson conjecture for X. For arbitrary K3 surfaces the existence is expected for points rationally equivalent to points on rational curves. The result should also be compared with [BT05, Thm. 4.2], where it is shown that every point in a Kummer surface over ¯Fp associated to the Jacobian of a curve of genus two is contained in a rational curve (and hence in a constant cycle curve of order one). 1.3. Acknowledgements: I am grateful to Claire Voisin for many comments, in particular on the torsion problem in the Bloch–Srinivas argument, and for the example in Section 6.3. Thanks to Burt Totaro for a question that triggered the results in Section 9 and for detailed comments on the first version, to Rahul Pandharipande for bringing [GG03] to my attention, and to François Charles and Davesh Maulik for help with arguments related to Section 8. Jimmy Dillies and Alessandra Sarti patiently answered my emails concerning Section 7 and Kieran O’Grady commented on the content of Section 8. Suggestions of the referee have helped to improve the exposition. The intellectual debt to the foundational work of Mark Green, Phillip Griffiths, and Claire Voisin is gratefully acknowledged. 2. Motivation We shall try to motivate the study of constant cycle curves from two perspectives: rational curves and Chow groups. For simplicity we shall restrict to K3 surfaces over algebraically closed fields k with char(k) = 0 and in fact k = ¯Q or C. For technical details, in particular concerning the case char(k) > 0, see the later sections. 3 2.1. Let (X, H) be a polarized complex K3 surface. The following folklore conjecture has been studied intensively over the last couple of years. Conjecture 2.1. The union S C ⊂ X of all rational curves C ⊂ X is dense (in the Zariski or, stronger, in the classical topology). The stronger and less studied version would only allow integral rational curves linearly equi- valent to some multiple of the given polarization H. The motivation for this conjecture stems from the classical result that for all m > 0 there ex- ists a rational curve C in the linear system mH (Bogomolov, Mumford, Mori–Mukai [MM83]). Here, a curve is rational if the reduction of each of its components has a normalization ≃ P1. For fixed m > 0 and generic (X, H), i.e. for polarized complex K3 surfaces in a Zariski dense open subset of the moduli space of polarized K3 surfaces, C can be chosen integral and even nodal [Ch02]. Note that at the same time there are at most finitely many rational curves in any fixed linear system, e.g. in mH, as K3 surfaces in characteristic zero are not unirational. More recently, the conjecture (for the Zariski topology) has been verified for K3 surfaces with ρ(X) ≡ 1 (2) by Li and Liedtke [LL12] following an approach by Bogomolov–Hassett–Tschinkel [BHT11]. The same ideas also apply to K3 surfaces that are not defined over ¯Q (see [Hu13] for details). Note that both conditions, ‘ρ(X) ≡ 1(2)’ and ‘not defined over ¯Q’, are general but not generic, i.e. they hold for K3 surfaces in the complement of a countable union of proper algebraic subsets of the moduli space of polarized K3 surfaces. Using work of Bogomolov and Tschinkel [BT00], Chen and Lewis [CL13] settled the conjecture in the classical topology for general (X, H). Not much is known about the stronger form of the conjecture, except for ρ(X) = 1 (when the rational curves have no other choice than being linearly equivalent to multiples of H). 2.2. Assume X is a K3 surface over C (or ¯Q). In [BV04] Beauville and Voisin described a distinguished class cX ∈ CH2(X) of degree one, which in particular has the properties that (2.1) c2(X) = 24 · cX and c1(L)2 ∈ Z · cX for all line bundles L on X. The set of closed points realizing this class was subsequently studied by McLean in [Ma04], where it is shown that the set (2.2) XcX := {x ∈ X [x] = cX ∈ CH2(X)} ⊂ X is dense in the classical topology. Similarly, one can consider the set Xα of points realizing any given class α ∈ CH2(X) and again, at least for generic K3 surface, this set is dense if not empty. However, as shall become clear, the set XcX is rather special. For abstract reasons, it is a countable union of Zariski closed subsets, but one expects it to be a countable union of curves, i.e. isolated points should not occur. This would be another distinction between cX and any other class α ∈ CH2(X) (cf. [Vo12b]). See also Section 3.5 for more on the sets X[x]. 4 The distinguished class cX can also be considered from a more arithmetic point of view, as expressed by the following special case of the much more general set of conjectures due to Bloch and Beilinson. But note that even for K3 surfaces, it has not been verified in a single example. Conjecture 2.2. (Bloch–Beilinson) Suppose X is a K3 surface over ¯Q and x ∈ X( ¯Q) is a ¯Q-rational point. Then [x] = cX . Note that McLean’s proof in fact shows that for every K3 surface X defined over some subfield k ⊂ C the set of ¯k-rational points realizing cX is dense in the classical topology, i.e. XcX (¯k) ⊂ X(C) is dense. In particular, it is known that for X over ¯Q there are many points x ∈ X( ¯Q) realizing cX . The property of cX proved in [BV04] that is the most relevant for our purpose, is the following: (2.3) If x ∈ C ⊂ X with C rational, then [x] = cX . This links rational curves to the study of CH2(X) and the distinguished class cX ∈ CH2(X). In particular, McLean’s density result could be seen as (weak) evidence for the density for rational curves as in Conjecture 2.1. Also, as pointed out by Bogomolov many years ago, it might a priori be possible that any ¯Q-rational point lies on a rational curve which in turn would prove Conjecture 2.2. The fact that points on rational curves all define the same class in CH2(X), which eventually relies on the existence of ample rational curves à la Bogomolov–Mumford [MM83], also leads to the concept of constant cycle curves studied in this paper. 2.3. Let C ⊂ X be a curve in a complex K3 surface. Then C is called a (pointwise) constant cycle curve if [x] ∈ CH2(X) is constant for points x ∈ C or, equivalently, if the push-forward Pic0(C) / CH2(X) is the zero map. Voisin shows in [Vo12b], by again using the existence of ample rational curves, that the class realized by a constant cycle curve is always the same, namely cX . The most important examples of constant cycle curves are provided by rational curves. But not every (pointwise) constant cycle curve is rational, see Section 6-8 for examples. This triggers the natural question how much weaker the notion of constant cycle curves really is. As rational curves, constant cycle curves do not come in families (at least not in characteristic zero). Indeed, any family of constant cycle curves would dominate X and so points in an open dense subset would all realize the same class in CH2(X) contradicting CH2(X) 6= Z (cf. proof of Lemma 3.6, ii)). Hence, for abstract reasons, the set of constant cycle curves in a fixed linear system, e.g. in mH, consists of at most countably many points. A finiteness result as for rational curves can be proved after restricting to constant cycle curves of bounded order (see 5 / Proposition 5.1). This result is based on normal functions and the recent results of Brosnan– Pearlstein [BP09] and Saito [Sa08] showing that the zero-set of admissible normal functions is algebraic. Alternatively to the definition of the order of a constant cycle curve using (1.2) one could define it directly as the order of a certain class κC ∈ CH2(X × k(ηC )) naturally associated to any integral curve C ⊂ X with its generic point ηC ∈ C (see Section 3.2). Note that CH2(X) is torsion free for X over a separably closed field and so the subtle information needed for the finiteness is contained in the kernel of CH2(X × k(ηC )) / CH2(X × k(ηC)). It is not difficult to show that rational curves are in fact constant cycle curves of order one (see Lemma 6.1). Also, non-rational constant cycle curves can be constructed in many ways, but they usually tend to be of higher order, i.e. κC 6= 0 in CH2(X × k(ηC )). So it is natural to wonder whether constant cycle curves of order one are all rational, but this turns out to be wrong and an explicit counterexample will be described (see Corollary 7.4). As constant cycle curves of bounded order resemble rational curves in many ways, we state Conjecture 2.1 for this more flexible class of curves. Note that if in the following the order is not bounded, the result is not difficult to prove, see [Vo12b] or Lemma 6.2. Conjecture 2.3. For any K3 surface X there exists an n > 0 such that the union S C ⊂ X of all constant cycle curves C ⊂ X of order ≤ n is dense.1 Once the finiteness of constant cycle curves of bounded order has been established, it would be interesting to actually count them. Counting rational curves on K3 surfaces is a fascinating subject which recently culminated in the proof of the Yau–Zaslow conjecture in complete gene- rality in [KMPS10]. Unfortunately, it seems much harder to count constant cycle curves, see Section 10.2. There is little evidence for an affirmative answer to Bogomolov’s question whether maybe any x ∈ X( ¯Q) is contained in a rational curve. Again, one could replace rational curves by constant cycle curves and an affirmative answer to this weaker form would still imply the arithmetic Conjecture 2.2. The following could be seen as a geometric version. Conjecture 2.4. Let X be a complex K3 surface. Then any point x ∈ X with [x] = cX is contained in a constant cycle curve. As alluded to before, it is much easier to construct constant cycle curves than rational curves. E.g. for K3 surfaces over finite fields every point is in fact contained in a constant cycle curve (see Proposition 9.6). However, a general technique that would allow to settle this problem is not yet available. 1In the appendix Claire Voisin provides a proof of the conjecture for generic complex K3 surfaces. The main idea is to produce constant cycle curves as non-torsion multi-sections of dominating families of elliptic curves, similar to [BT00]. 6 / 3. Constant cycle curves 3.1. Let X be a projective K3 surface over a field k. For two reasons, one often has to assume that k is algebraically closed. Firstly, CH2(X) might have torsion otherwise and, secondly, the good behavior of the Beauville–Voisin class cX depends on the existence of rational curves for which k algebraically closed is needed. In fact, for k not algebraically closed, cX ∈ CH2(X) has the desired properties (e.g. being realized by points on rational curves) only up to torsion. We will state explicitly when k = ¯k is assumed. Definition 3.1. A curve C ⊂ X is a pointwise constant cycle curve if all closed points x ∈ C define the same class [x] ∈ CH2(X). For k = ¯k the condition is equivalent to require [x] = cX (the Beauville–Voisin class, see Section 2.2) for all closed points x ∈ C, see [Vo12b, Lem. 2.2].2 Another way of expressing the condition (still assuming k = ¯k) is as follows. A curve C ⊂ X is a pointwise constant cycle curve if and only if the natural map takes image in Z · cX or, still equivalent, that (3.1) (3.2) is zero. Here, fC : eC C ⊂ X. fC∗ : Pic(eC) fC∗ : Pic0(eC) / CH2(X) / CH2(X) / X is the composition of the normalization eC / C with the inclusion The notion of pointwise constant cycle curves is really interesting only for uncountable fields k. E.g. it is not preserved under base change when the base field k is too small (cf. Lemma 3.6 and Proposition 3.7). Also, according to Conjecture 2.2, every curve in a K3 surface over ¯Q should be a pointwise constant cycle curve which makes it a notion of little interest in this case. However, the same is true for K3 surfaces over k = ¯Fp and in this case it is a shadow of the Bloch–Beilinson conjecture for function fields, see Proposition 9.2. 3.2. In order to introduce the finer version of this notion, we define the class κC naturally associated to any integral curve C ⊂ X. To this end, denote (abusively) by ∆C ⊂ X × C the graph of the inclusion and consider the cycle ∆C − {x0} × C, where x0 ∈ X is an arbitrary point with [x0] = cX . Again, for the existence of such a point and for cX being well-defined, k has to be algebraically closed. Here, cX is the Beauville–Voisin class (cf. Section 2.2). Then let (3.3) κC ∈ CH2(X × k(ηC )) 2We tacitly assume that the standard facts on cX hold true for K3 surfaces over algebraically closed fields of positive characteristic, but all we really need is that points on rational curves all realize the same class. This follows from the existence of ample rational curves which can be shown by reduction modulo p. 7 / / / / be the class of the restriction ∆C − {x0} × C to the generic fibre Xk(ηC ) = X ×k k(ηC ) of the second projection X × C / C. In other words, the generic point ηC ∈ C is viewed as a closed point of the K3 surface Xk(ηC ) over the function field k(ηC) of C and then corrected by the ‘constant’ point {x0} × ηC. This natural class has been considered before in the literature, see e.g. [Ke06, Ex. 4.2]. Definition 3.2. Assume k algebraically closed. An integral curve C ⊂ X is a constant cycle curve if κC ∈ CH2(X × k(ηC )) is a torsion class. We call an arbitrary curve C ⊂ X a constant cycle curve if every integral component of C has this property. Note that this definition makes perfect sense for all surfaces with a distinguished class in CH2(X), e.g. for those with CH2(X) ≃ Z. In fact constant cycle curves can be defined for arbitrary surfaces by means of Lemma 3.8, but, with the exception of Section 4, we will restrict to K3 surfaces. Remark 3.3. The class κC is in fact the direct image under the push-forward CH1(C × k(ηC )) / CH2(X × k(ηC )) of the class [ηC] − [x0], where x0 ∈ C is any point with [x0] = cX in CH2(X) (e.g. a point of intersection with a rational curve). However, the class [ηC ] − [x0] ∈ CH1(C × k(ηC )) itself is never torsion except for C rational, because no non-trivial multiple of [ηC] is ever contained in the image of the base change map CH1(C) / CH1(C × k(ηC)). In fact, whether a non-rational curve C is a constant cycle curve depends on the particular / X. We will see examples of (smooth) curves that can be embedded as constant embedding C cycle curves of varying order and even as non-constant cycle curves in the same K3 surface X. For any field extension K/k the pull-back yields a map (3.4) CH2(X) / CH2(X ×k K). The image of the Beauville–Voisin class cX ∈ CH2(X) shall be denoted (cX )K. It can also be seen, at least for K algebraically closed, as the Beauville–Voisin class of XK = X ×k K, i.e. (cX )K = cXK . Compare the following result to Lemma 3.8. Lemma 3.4. Let X be a K3 surface over an algebraically closed field k. For an integral curve C ⊂ X the following conditions are equivalent: i) The curve C is a constant cycle curve. ii) There exists a positive integer n such that (3.5) n · [ηC] = n · (cX )k(ηC ) in CH2(X ×k k(ηC)), where the generic point ηC ∈ C is viewed as a closed point in X ×k k(ηC ). 8 / / / / / iii) If ηC ∈ C is viewed as a point in the geometric generic fibre X ×k k(ηC ), then in CH2(X ×k k(ηC )). [ηC] = (cX )k(ηC ) Proof. This is an immediate consequence of the fact that the pull-back (3.4) has torsion kernel and Roitman’s theorem [Ro80], and its generalizations due to Bloch [Bl80] and Milne [Mi82], showing that the group CH2(X ×k k(ηC )) is torsion free. (cid:3) The Chow group of the generic fibre is best viewed as (3.6) CH2(X × k(ηC )) = lim → CH2(X × U ), where the direct limit is over all non-empty Zariski open subsets U ⊂ C, see [Bl80, Lem. 1.I.20]. Hence, κC is a torsion class if and only if the class κC,U ∈ CH2(X × U ) of the restriction of [∆C − {x0} × C] to X × U for some non-empty open subset U ⊂ C is torsion. In order to obtain finiteness results, one needs to bound the order of the torsion class κC of a constant cycle curve. Definition 3.5. The order of an integral constant cycle curve C ⊂ X is the order of the torsion class κC ∈ CH2(X × k(ηC )). The order of an arbitrary constant cycle curve C is the maximal order of its integral components. Note that by shrinking U ⊂ C one can always assume that the order of κC and κC,U coincide. By definition the order is the minimal positive n satisfying (3.5). 3.3. We shall explain the relation between the two notions of constant cycle curves and state some basic properties. Proposition 3.6. Let X be a K3 surface over an algebraically closed field k. i) Let C ⊂ X be a curve and let K/k be an algebraically closed base field extension. Then C is a constant cycle curve of order n if and only if CK ⊂ XK is a constant cycle curve of order n. ii) Assume X is not (Artin) supersingular and char(k) 6= 2.3 If K/k is an extension with K also algebraically closed and D ⊂ XK is a constant cycle curve, then D descends, i.e. there exists a constant cycle curve C ⊂ X with D = CK . iii) If X is defined over a (finitely generated) field k0 with ¯k0 = k and C ⊂ X is a constant cycle curve, then the natural Gal(¯k/k0)-action applied to C yields only constant cycle curves. 3This assumption holds whenever char(k) = 0 and see below for a reminder on the notion of supersingular K3 surfaces. What is really needed in the proof is ρ(X) 6= 22 or CH2(Xk′ ) 6= Z for some algebraically closed extension k′/k. Thanks to Burt Totaro for pointing out how to weaken the original assumption. 9 Proof. i) The pull-back (3.7) CH2(X ×k k(ηC )) / CH2(XK ×K k(ηCK )) induced by the base change XK ×K k(ηCK ) = (X ×k k(ηC)) ×k K / kernel and maps κC to κCK . Hence, C is a constant cycle curve if and only if CK is. / X ×k k(ηC ) has torsion That the order does not change can be shown using arguments of Lecomte in [Le86], where it is proved that for any variety Y over an algebraically closed field K0 the base change CH∗(Y ) / CH∗(YK) to a larger algebraically closed field K/K0 induces an isomorphism on torsion. (Injectivity is enough for our purpose, for which only K0 algebraically closed is needed.) It cannot be applied directly, as in our case K0 = k(ηC ) is not algebraically closed, but we may apply it to Y = X ×U for open subsets U ⊂ C to obtain CH2(X × k(ηC )) ≀ CH2(XK ×K k(ηCK )) ≀ CH2(X ×k U ) lim → / lim → CH2(XK ×K UK) / lim → CH2(XK ×K V ), where the last map might a priori not be injective. Suppose that the pull-back αK of α ∈ CH2(X ×k U ) yields a trivial class in CH2(XK ×K V ) for some open set V ⊂ UK, which we can assume to be the complement of finitely many closed points p1, . . . , pm ∈ UK . (One can further reduce to the case that each of the pi dominates C, otherwise shrink U ). Now use the localization exact sequence (see [Bl80, Fu98, Vo02]): CH1(XK × {p1, . . . , pm}) / CH2(XK ×K UK ) / CH2(XK ×K V ) / 0 to conclude that αK is supported on {p1, . . . , pm}. Write CH2(XK ×K UK ) = lim CH2(X ×k U ×k W ) with the limit over all non-empty open subsets W ⊂ Spec(A) for all (finitely generated) k- algebras k ⊂ A ⊂ K (cf. [Le86]). For small W ⊂ Spec(A) represent αK by αW ∈ CH2(X ×k U ×k W ). Its restriction αt to a closed point t ∈ W gives back α. Indeed, since k is alge- / αt is given by the isomorphism CH2(X ×k braically closed, one has k(t) ≃ k and α ✤ U ) / CH2(X ×k U ×k k(t)). / αW ✤ But for small W ⊂ Spec(A) the class αt is supported on the intersection of the closure of {p1, . . . , pm} in X ×k U ×k W with the fibre over t, which is a finite set of points {p1t, . . . , pmt} in X ×k U ×k k(t). Hence, α restricted the complement of these points is trivial. This eventually shows that α represents the trivial class in CH2(X × k(ηC )). ii) If D ⊂ XK is not defined over k, then there exists a one-dimensional family of curves Ct ⊂ X which can be seen as specializations of D. In particular, their classes κCt are obtained 10 / / / / / / / / / / by specializing κD and, therefore, are also torsion. Thus, one would obtain a dominant family of pointwise constant cycle curves Ct ⊂ X. For every larger algebraically closed k′/k base changing the family {Ct} defines a dominating family of constant cycle curves for Xk′ and which would imply CH2(Xk′) ≃ Z, which is excluded for non-supersingular K3 surfaces, as e.g. CH2(X × k(ηX )) 6= Z. Recall that a K3 surface X over a field k of char(k) > 0 is called Artin supersingular if its height is infinite and Shioda supersingular if ρ(X¯k) = 22. It has been known for a long time that Shioda supersingular implies Artin supersingular and the proof of the converse has recently be completed in [Cha12, Mau12, Ma13]. (It has also been known that unirational K3 surfaces are Shioda (and hence Artin) supersingular and the converse has been established in [Li13].) Thus, a K3 surface X over k with char(k) 6= 2 is supersingular if and only if ρ(X) 6= 22, i.e. H 2 et(X, Qℓ(1)) 6= NS(X) ⊗ Qℓ. Hence, Bloch’s result [Bl80, Thm. 6, Appendix to Sec. 1] applies. iii) The last assertion is obvious, as the notion of constant cycle curves is scheme-theoretic. (cid:3) For the reader’s convenience and later use, we recall the following fact, which is a special case of a result due to Voisin (cf. [Vo02, Ch. 22]) improving upon a result of Bloch and Srinivas [BS83]. Proposition 3.7. Assume k is algebraically closed. Then a constant cycle curve C ⊂ X is also a pointwise constant cycle curve. If k is uncountable, the converse holds true as well. Proof. Clearly, we may assume that C is integral and for simplicity we also assume that C is smooth (otherwise pass to its normalization and replace C / X by the generically finite map / X). The first assertion is easy. If κC is torsion, say n · κC = 0, then there exists an open subset U := C \ {p1, . . . , pm} ⊂ C such that 0 = n · [(∆C − {x0} × C)X×U ] ∈ CH2(X × U ) (use (3.6)). By the localization exact sequence eC (3.8) CH1(X × {p1, . . . , pm}) / CH2(X × C) / CH2(X × U ) / 0, we can assume that n · (∆C − {x0} × C) is rationally equivalent to a cycle Z on X × C with support in X × {p1, . . . , pm}. As any zero-cycle on C is linearly equivalent to one disjoint to the / CH2(X) is trivial. Thus, finite set of points {p1, . . . , pm}, the induced map [Z]∗ : CH0(C) / CH2(X) is trivial and, since CH2(X) is torsion free, also n · [∆C − {x0} × C]∗ : CH0(C) [∆C − {x0} × C]∗ = 0. The latter is equivalent to saying that [x] ≡ [x0] = cX for all closed points x ∈ C. For the converse use [Vo02, Cor. 22.20], which is stated for k = C but in fact holds for any uncountable field k. Then n · [∆C − {x0} × C] = [Z] for some n > 0 with Z supported on a closed set of the form X ×{p1, . . . , pm}. But then ZXk(ηC ) = 0 and hence κC ∈ CH2(X ×k(ηC )) is torsion. (cid:3) 11 / / / / / / / 3.4. One could avoid mentioning the distinguished class cX ∈ CH2(X) in the definition of a constant cycle curve altogether by proving analogously to Lemma 3.4 the next Lemma 3.8. Let X be a K3 surface over an algebraically closed field k. For an integral curve C ⊂ X the following conditions are equivalent: i) The curve C is a constant cycle curve. ii) There exists a positive integer n such that (3.9) n · [ηC ] ∈ Im(cid:0)CH2(X) / CH2(X ×k k(ηC))(cid:1) , where the generic point ηC ∈ C is viewed as a closed point in X ×k k(ηC ). iii) If ηC ∈ C is viewed as a point in the geometric generic fibre X ×k k(ηC ), then [ηC] ∈ Im(cid:16)CH2(X) / CH2(X ×k k(ηC ))(cid:17) . Proof. Clearly, i) implies ii) and iii). Since CH2(X ×k k(ηC )) kernel and torsion free target, ii) and iii) are equivalent. / CH2(X ×k k(ηC )) has torsion It remains to show that ii) implies i). Now, for a closed point x ∈ C, the composition of the / CH2(X) (see [Fu98, Ch. 20.3]) with the pull-back specialization map sx : CH2(X × k(ηC )) (3.9) yields the identity on CH2(X). If now n · [ηC ] is in the image of (3.9), say n · [ηC ] = αk(ηC ), then n · [x] = α, as ηC clearly specializes to [x]. Thus, C is a pointwise constant cycle curve, which for an uncountable field is enough to conclude (use Proposition 3.7). If k is only countable, use base change to an uncountable algebraically closed extension K/k (e.g. a universal domain). Clearly, then n · [ηCK ] = n · [ηC ]K is contained in the image of / CH2(XK ×K k(ηCK )). Hence, CK is a constant cycle curve and, by Lemma 3.6, i), CH2(X) also C is. (cid:3) 3.5. Let K/k be an extension of algebraically closed fields. Let X be a K3 surface over k and denote by XK the K3 surface over K obtained by base change. The base change morphism / CH2(XK ), i.e. [x]K = ξ∗[x] for all ξ : XK closed points x ∈ X. It is injective, but not surjective as soon as trdegk(K) ≥ 2 and ρ(X) 6= 22 by [Bl80] or trdegk(K) ≥ 1 and char(k) = 0 (see [GGP04]). / X induces the pull-back map ξ∗ : CH2(X) Compare the following also to [Go12, Prop. 5]. Corollary 3.9. Let x ∈ XK be a closed point with e.g. [x] = cXK = (cX )K . Then one of the following is true: [x] ∈ Im(cid:0)CH2(X) / CH2(XK )(cid:1) , i) The image ξ(x) ∈ X is a closed point, i.e. x is defined over k. ii) The closure C := {ξ(x)} ⊂ X of ξ(x) ∈ X is a constant cycle curve. iii) The image ξ(x) ∈ X is the generic point of X and CH2(X) ≃ Z. 12 / / / / / / / / Proof. Suppose x is not defined over k. Then its image in X is either the generic point of X or of a curve C ⊂ X. In the second case, consider the natural inclusion k(ηC ) / k(x) ≃ K / CH2(XK ) sends [ηC] to for the generic point ηC ∈ C. The induced map CH2(X × k(ηC )) [x]. Since the kernel is torsion, one can conclude by Lemma 3.8. Similarly, if x is mapped to / CH2(X ×k k(ηX )). ηX ∈ X, then [ηX] is up to torsion contained in the image of CH2(X) And then, by specialization, in fact [y] ≡ const for all points y ∈ X. (cid:3) So, in principle, one could try to produce constant cycle curves by finding points x ∈ XK with [x] = cXK not defined over k. Then either CH2(X) ≃ Z or the closure of x in X is a constant cycle curve. Although finding points x not defined over k is in principle possible, deciding whether also [x] = cXK is difficult to verify without knowing beforehand that x is contained in a constant cycle curve C ⊂ XK which for CH2(X) 6≃ Z would automatically descend to k, cf. Lemma 3.6. Also note that by i) in Lemma 3.6 one might expect that the order of the constant cycle curve C := {ξ(x)} ⊂ X in situation ii) is an invariant of the point x ∈ XK , but how to read it off directly from x is unclear. Corollary 3.10. Let x ∈ XK be a closed point not defined over k. Then i) Either, [x] = (cX )K ∈ CH2(XK ) and there exists a constant cycle curve x ∈ C ⊂ XK or ii) the class [x] is not contained in the image of CH2(X) / CH2(XK ). Proof. Indeed, take the curve {ξ(x)} ⊂ X (or any curve contained in it) and base change it to a curve in XK. (cid:3) Rephrasing this result yields another proof of a weak form of the main result of [GGP04] for K3 surfaces and of the original result by Bloch [Bl80]. Contrary to the proof in [GGP04], the arguments here do not involve Hodge theory and, therefore, work in positive characteristic. Roughly, the result says, if CH2(X)0 6= 0, then CH2(X) grows under any base field extension to a bigger algebraically closed field. Corollary 3.11. Assume K/k is an extension of algebraically closed fields with trdegk(K) > 0. If X is a K3 surface over k with CH2(X) 6≃ Z, then CH2(X) / CH2(XK ) is not surjective. (cid:3) Remark 3.12. The following is rather speculative and probably well-know to experts: One may wonder, if the above opens a way to prove the Bloch–Beilinson conjecture for K3 surfaces (see Conjecture 2.2). Suppose X is a K3 surface over k = ¯Q. If for any x0 ∈ X(k) there exist a field extension K/k and a point x ∈ X(K) \ X(k) with [x] = [x0]K, then CH2(X) ≃ Z. Indeed, by Corollary 3.10 one would have [x] = (cX )K, as [x] is by assumption contained in the image of 13 / / / / / / CH2(XK ), and hence [x0] = cX . Unfortunately, I do not know of any method that CH2(X) could possibly construct such a point [x]. Note that it has to be important that the original surface is defined over a number field, i.e. k = ¯Q, as we do not expect CH2(X) ≃ Z for X over bigger fields. (A concrete counterexample has been given by Schoen, see [Ja07].) Similarly to (2.2) one can define for any point x0 ∈ X(k) the set (3.10) X[x0](k) := {x ∈ X(k) [x] = [x0] ∈ CH2(X)}, which due to Maclean’s result [Ma04] is dense (at least for generic complex X). If C ⊂ X is a constant cycle curve, then C(k) ⊂ XcX (k). Now consider a non-trivial extension K/k with K algebraically closed. Then CK ⊂ XK remains a constant cycle curve. Clearly, not all points in CK will be defined over k and, therefore, the natural inclusion XcX (k) ⊂ XcX (K) is strict. For any other class [x0] 6= cX the set of points realizing it does not grow under base field extension, as shown by the next result which is again just a reformulation of Corollary 3.10. Corollary 3.13. Let K/k be any extension of algebraically closed fields. If X is a K3 surface over k and x0 ∈ X is a closed point with [x0] 6= cX , then X[x0](k) = X[x0](K), i.e. all points in XK rationally equivalent to x0 are defined over k. (cid:3) 4. Constant cycle curves on other surfaces The notion of constant cycle curves makes sense for other types of surfaces, see Section 3.2. We shall briefly discuss the case of surfaces satisfying Bloch’s conjecture. Recall that Bloch’s conjecture for surfaces X with pg(X) = 0 predicts that the kernel of the Albanese map CH2(X)0 / Alb(X) is trivial. It has been verified in [BKL76] for all surfaces of Kodaira dimension ≤ 1 and for many surfaces of general type, in particular those dominated by products of curves. In [Ki05, Cor. 7.7] it has been shown that the finite-dimensionality of the Chow motive h(X) (in the sense of Kimura–O’Sullivan) implies Bloch’s conjecture. For simplicity we will restrict to the case q(X) = 0, otherwise one would have to restrict to curves in the fibres of the Albanese map. Proposition 4.1. Let X be a smooth projective surface with pg(X) = q(X) = 0 over an algebraically closed field k of characteristic zero. Then X satisfies Bloch’s conjecture, i.e. CH2(X)0 = 0, if and only if every curve in X is a constant cycle curve. 14 / / Proof. Suppose X satisfies Bloch’s conjecture. If k is uncountable, e.g. k = C, then the assertion follows from the arguments used to prove Proposition 3.7. Indeed, assuming Bloch’s conjecture, every curve is a pointwise constant cycle curve. For arbitrary field k one uses [GP03, Thm. 7] (or directly the techniques of [BS83]) which states that Bloch’s conjecture is equivalent to the finite-dimensionality of h(X). By [KMP07] the Chow motive h(X) is finite-dimensional if and only if its transcendental motive t2(X) is finite-dimensional. But for pg(X) = 0, the latter is equivalent to t2(X) = 0 and by [KMP07, Cor. 7.4.9] this implies CH2(X × k(ηX ))0 ⊗ Q = 0 (see also the discussion in [An05, Ch. 4.1]). To conclude, use that the generic point of any curve C ⊂ X is a specialization of ηX. Conversely, if every curve C ⊂ X is a constant cycle curve, then clearly CH2(X) ≃ Z. (cid:3) However, determining the order of constant cycle curves in these surfaces is a different matter. Mainly, because nothing seems to be known about the integral version of the transcendental motive t2(X) and the torsion in CH2(X × k(ηX )) is difficult to control. The following cases are instructive for our purpose. Proposition 4.2. Let X be a surface over an algebraically closed field k with pg(X) = q(X) = 0 satisfying Bloch’s conjecture CH2(XK )0 = 0 for all algebraically closed extensions K/k. i) There exists an integer n such that all curves on X are constant cycle curves of order ≤ n.4 ii) If X is rational, then every curve C ⊂ X is a constant cycle curve of order one. iii) If X is an Enriques surface, then all curves are constant cycle curves of order n4. Proof. i) For an integral curve C ⊂ X consider the specialization map CH2(X × k(ηX )) / CH2(X × k(ηC )) which sends [ηX ] to [ηC]. Thus, it suffices to show that there exists an n with n · [ηX] = n · [x0 × k(ηX )] for some x0 ∈ X or, equivalently, that [ηX ] − [x0 × k(ηX )] is contained in the / CH2(X × k(ηX )). But by assumption CH2(X × k(ηX )) ≃ Z and kernel of CH2(X × k(ηX )) [ηX] − [x0 × k(ηX )] is of degree zero. ii) As rational curves are constant cycle curves of order one (see Lemma 6.1 for the argument which is a direct consequence of the definition), we may also replace X by a minimal model. 1 × 1 + h1 × h2 + 1 × h2 So, if X is rational, we can in fact assume X ≃ P2. Now use that [∆P2] ∈ CH2(P2 ×P2) can be written as [∆P2] = h2 2, where hi, i = 1, 2, denote the hyperplane sections on the two factors. Hence, for every integral curve C ⊂ P2 one finds [∆C ] = [∆P2]P2×C = h2 1 × [C] + h1 × h2C and, therefore, κC = 0. iii) Every Enriques surface admits an elliptic fibration X / / P1 with a 2-section C0 ⊂ X. Then the base change to X × k(ηC ) for any curve C ⊂ X comes with a natural 2-section, too. Now copy the arguments in [BKL76] to show that every class in CH2(X ′)0 for an Enriques 4Thanks to Claire Voisin for suggesting this. 15 / / surface X ′ over an arbitrary field k′ is annihilated by n2, when n is the degree of a multi-section. In [BKL76] this was combined with the fact that CH2(X ′)0 is torsion free for k′ algebraically closed to deduce CH2(X ′)0 = 0. Here, we apply it to X ′ = X × k(ηC ) and k′ = k(ηC ) to conclude the assertion. (cid:3) 5. Finiteness of constant cycle curves of fixed order The aim of the section is to prove Proposition 5.1. Let X be a projective K3 surface over an algebraically closed field of charac- teristic zero. Then there are at most finitely many constant cycle curves C of fixed order n in any linear system L. It is clearly enough to prove the theorem for complex K3 surfaces. The techniques, involving Hodge theory and normal functions, do not apply to K3 surfaces over fields of positive charac- teristic. In fact, for unirational K3 surfaces over ¯Fp, which come with infinitely many rational and hence constant cycle curves of order one in a certain linear system, the result is clearly false. Whether this is the only exception is not clear. 5.1. For the Hodge theoretic considerations below, we first need to introduce a compactification of κC to a class in CH2(X × C) that is different from the naive one used e.g. in the proof of Proposition 3.7. (5.1) Let X be a K3 surface with a fixed point x0 ∈ X such that [x0] = cX. Consider an integral / C. As before, the composition shall be denoted ZC := ∆fC − C ×(cid:16)X ni · yi(cid:17) − {x0} × eC / C ⊂ X. Given C and cycle P ni · yi of degree one on eC, one defines fC which can also be seen as the pull-back of the diagonal ∆X ⊂ X × X under (id, fC) : / X × X. Note that, although not reflected by the notation, ZC and the associated curve C ⊂ X and its normalization eC fC : eC which is an integral cycle on the smooth threefold X × eC. Here, ∆fC denotes the graph of X × eC class [ZC ] ∈ CH2(X × eC) do depend on the cycle P ni · yi resp. the associated line bundle L0 := O (P ni · yi) on eC. (In Lemma 5.4 below the transcendental Abel–Jacobi class associated with ZC will be shown to be independent of P ni · yi.) Geometrically, it would seem natural to choose L0 to be of the form O(y0) for some point y0 ∈ eC. However, for arguments involving families of curves C, allowing line bundles has technical advantages, cf. Section 5.3. In fact, we will later choose L0 such that L2g−2 CL, for C ∈ L with L.L = 2g − 2. ≃ f ∗ 0 It is straightforward to check that ZC is homologously trivial, e.g. for a complex K3 surface the induced map is zero. This allows one to define its Abel–Jacobi class, see Section 5.2. [ZC]∗ : H ∗(eC, Z) / H ∗(X, Z) 16 / / / / For an integral curve C, the generic points η eC ∈ eC and ηC ∈ C can be identified. Then Lemma 5.2. For an integral curve C ⊂ X, restriction to the generic fibre maps [ZC ] to κC , i.e. CH2(X × eC) / CH2(X × k(ηC )), [ZC] ✤ / κC . Clearly, if [ZC ] is torsion, then so is κC and hence C is a constant cycle curve, but the converse is not true. If in addition L2g−2 CL is required, then at least the question whether the cycle ZC is torsion does no longer depend on the particular L0, for another choice of L0 ≃ f ∗ 0 differs by torsion in Pic0(eC). Remark 5.3. We briefly discuss various other possible choices for ZC. We restrict to the case of a complex K3 surface X. i) On X × X one often considers ZX := ∆X −X ni · (Ci × Di) − {x0} × X − X × {x0}, the transcendental part of the diagonal. The curves Ci, Di ⊂ X and the ni ∈ Q are chosen such that [Z]∗ is trivial on H 0(X, Q) ⊕ H 4(X, Q) ⊕ NS(X) ⊗ Q ⊂ H ∗(X, Q), i.e. [ZX ] ∈ T (X) ⊗ T (X) ⊗ Q ⊂ H 4(X × X, Q), where T (X) ⊂ H 2(X, Z) is the transcendental lattice. For ρ(X) = 1, this can be rewritten as (5.2) ZX = ∆X − (1/(2g − 2)) · (C × C) − {x0} × X − X × {x0}, with C.C = 2g − 2, which is defined in general and satisfies (2g − 2) · [ZX ] ∈ [C]⊥ ⊗ [C]⊥ ⊂ H 2(X, Z) × H 2(X, Z). restriction to X × k(ηC ) represents κC . The obvious advantage to be defined universally on X × X has the price, due to the coefficient 1/(2g − 2), of being only rationally defined, i.e. The pullback to X × eC would be another natural choice for ZC with the property that its [(id, fC)∗ZX] is well defined only in CH2(X × eC) ⊗ Q, which makes it more difficult to work with its Abel–Jacobi class. Similarly, the class defined by the cycle Z ′ C := ∆fC − (1/(2g(eC ) − 2)) · (C × D) − {x0} × C with D ∈ ωC would be independent of any additional choice of L0 ∈ Pic1(eC), but again it is only rational. ii) In [GG03] Green and Griffiths study a cycle introduced by Faber and Pandharipande. For a divisor D of degree d on a smooth curve C they define zD := D × D − d · D∆. 17 / / Assume now that C is contained in a K3 surface X and suppose for simplicity that C generates NS(X). Then for D ∈ ωC zD ∼ (2 − 2g(C)) · ZX C×C under C × C ⊂ X × X, where in this case ZX is as in (5.2). Similarly, zD is rationally equivalent to the restriction of (2 − 2g(C)) · Z ′ C under C × C ⊂ X × C. As pointed out in [GG03], the class [zD] ∈ CH2(C × C) ⊗ Q for D ∈ ωC is (by Bloch– Beilinson) conjectured to be trivial for curves C defined over ¯Q, but it is also shown that it is not trivial for general C of genus g(C) ≥ 4 (see also the shorter proof in [Yi13]). For 4 ≤ g(C) < 10 this can be used to show that for general C ⊂ X (with varying X) the class [Z ′ C] ∈ CH2(X × C) ⊗ Q is non-trivial. 5.2. Consider the Abel–Jacobi map for the smooth threefold X × eC: / ZΓ AJ : CH2(X × eC)hom / J 3(X × eC), γ ✤ . Here, CH2(X × eC)hom denotes the homologically trivial part and for any γ we let Γ be a real three-dimensional cycle with ∂Γ = γ. The intermediate Jacobian J 3(X × eC) := H 3(X × eC, C) F 2H 3(X × eC) + H 3(X × eC, Z) ≃ F 2H 3(X × eC)∗ H 3(X × eC, Z) is a complex torus of dimension 22 · g(eC). We shall also need the ‘transcendental part’ of the intermediate Jacobian which we define as (see also [Gr98, Vo99]): (5.3) J 3(X × eC)tr := F 2(T (X) ⊗ H 1(eC, C))∗ T (X) ⊗ H 1(eC, Z) . Here, T (X) := NS(X)⊥ ⊂ H 2(X, Z) is the transcendental lattice which equals the orthogonal complement [C]⊥ ⊂ H 2(X, Z) if ρ(X) = 1. We shall denote the composition of the projection J 3(X × eC) / J 3(X × eC)tr with the Abel–Jacobi map by AJtr : CH2(X × eC)hom / J 3(X × eC)tr. The following is well-known, see e.g. [Vo99]: Lemma 5.4. The transcendental Abel–Jacobi map AJtr factorizes naturally AJtr : CH2(X × eC)hom/(Pic(X) × CH1(eC)hom) In particular, AJtr(ZC ) is independent of the choice of L0 = O(P ni · yi) ∈ Pic1(eC) as in (5.1). Proof. One has to show that RΓ is trivial on T (X) ⊗ H 1(eC, C) whenever ∂Γ is of the form D ×P ni · xi for some curve D ⊂ X and with xi ∈ eC and P ni = 0. But in this case one can assume that Γ is of the form D × Γ0 with a path Γ0 ⊂ eC and clearly RD = 0 on T (X). (cid:3) / J 3(X × eC)tr. 18 / / / / / The Abel–Jacobi class of AJ(κC ) is the class eX,C in [Gr98], which can also be understood as an extension class in the category of mixed Hodge structures. The transcendental part is (essentially) the class studied further in [Vo99]. The second assertion of the following has been observed already by Green in [Gr98]. Corollary 5.5. Suppose C ⊂ X is an integral curve. Then (5.4) is well-defined. If C is a constant cycle curve of order n, then n · AJtr(κC ) = 0. AJtr(κC ) := AJtr(ZC) ∈ J 3(X × eC)tr Proof. One argues as in the proof of Proposition 3.7. Using (3.8) and Lemma 5.4, one finds that AJtr(ZC ) only depends on its restriction to some open subset U := eC \ {p1, . . . , pk). But if n · κC = 0, then U can be chosen such that n · [ZC X×U ] = 0 in CH2(X × U ) and hence n · AJtr(ZC) = 0. (cid:3) One can also project AJ(ZC ) onto a class AJalg(ZC) in the algebraic part of the intermediate Jacobian. (5.5) J 3(X × eC)alg := F 2(NS(X) ⊗ H 1(eC, C))∗ NS(X) ⊗ H 1(eC, Z) ≃ Pic0(eC)ρ(X). To simplify the notation we shall henceforth assume ρ(X) = 1 (but see Remark 5.10), so that Lemma 5.6. For an integral curve C ∈ L with L.L = 2g − 2, one has J 3(X × eC)alg = F 2([C] ⊗ H 1(eC, C))∗/([C] ⊗ H 1(eC, Z)) ≃ Pic0(eC). )∗ ∈ Pic0(eC). Thus, if L0 is chosen such that L2g−2 AJalg(ZC ) = f ∗ CL ⊗ (L2g−2 0 0 ≃ f ∗ CL, then AJalg(ZC ) = 0. Cf. [Gr98, p. 270]. Proof. This must be standard. As I was not able to find a reference, I will sketch the argument. Let us first explain how AJalg(ZC ) changes with L0. The difference of the cycles for two different L0 and L′ 0 is a cycle of the form C × D with D a degree zero cycle on C. Then for any one-form α on C one finds AJalg(ZC − Z ′ as predicted by the assertion, AJalg(ZC − Z ′ this allows one to restrict to the case L0 = O(y0) for some point y0 ∈ C. C)(C × α) =RC×δ[C] × α = (2g − 2)Rδ α, where ∂δ = D. Hence, Now let γ ⊂ X × C with ∂γ = ZC. Then Rγ[C] × α =Rγ∩(C ′×C) pr∗ 2α, where the deformation C ′ of C is chosen generic such that γ ∩ (C ′ × C) is indeed one-dimensional. Note that C ′ ∩ C := {x1, . . . , x2g−2} is a divisor in LC. The intersection of C ′ × C with ZC consists of the points (xi, xi) and (with negative sign) (xi, y0) and the intersection of C ′ × C with γ consists of paths connecting these points. Now project onto the second factor, which gives paths connecting the points xi = pr2(xi, xi) with y0 = pr2(xi, y0), i.e. a path with boundary LC ⊗ O((2g − 2) · y0)∗. (Alternatively, one can pull back α.) (cid:3) C) = O((2g − 2) · D) ≃ (L′ 0)2g−2. In particular, 0 ⊗ L∗ 19 The natural inclusion T (X) ⊕ NS(X) ⊂ H 2(X, Z) is of finite index. Hence (5.6) J 3(X × C) / J 3(X × eC)tr × J 3(X × eC)alg is finite of degree say N (which only depends on [C] ∈ H 2(X, Z) and not on C). In particular, if both, AJ(ZC)tr ∈ J 3(X × eC)tr and AJ(ZC )alg ∈ J 3(X × eC)alg vanish, then also N · AJ(ZC) ∈ J 3(X × eC). It is difficult to decide for any given curve C ⊂ X whether AJtr(κC ) is non-trivial. 5.3. However, when put in a family the resulting normal function is easier to control and from its non-vanishing one deduces the non-vanishing of AJtr(κC ) at least for generic C. This type of argument is standard (for hyperplane sections), but there are technical details that have to be adjusted to our situation. The key point is eventually the algebraicity of the zero-locus of normal functions recently established by Brosnan–Pearlstein [BP09] and M. Saito [Sa08], see [Cha13] for a recent overview. We start with a positive-dimensional family of curves T ⊂ L in a fixed linear system L on X with L.L = 2g − 2. We assume that all curves C ∈ T0 parametrized by a dense open subset T0 ⊂ T are integral with (analytically) constant singularity type. Under this assumption, simultaneous normalization and then compactification (for both, one may need to replace the inclusion T0 ⊂ L by a generically finite map T0 / L) yields a diagram / CT0 / X f : eCT0 ❍ ❍ ❍ ❍ p0 ❍ ❍ ❍ ❍ #❍ T0 / C ⊂ X considered before. / T0 is smooth with fibres given by the normalizations / Ct ⊂ X for t ∈ T0 is exactly as in the situation with f dominant and such that p0 : eCT0 of the curves Ct, t ∈ T0. So, ft : eCt fC : eC Recall that the class of the cycle ZC on X × eC defined in (5.1) depends on the choice of a line bundle L0 of degree one on eC. In order to define a global cycle that restricts to ZC on 0 ⊂ Pic1(eC/T0) X × eC for a fibre C = Ct, t ∈ T0, we pull-back CT0 of all line bundles L0 on fibres eC of eCT0 / T0 comes with a line bundle L0 on eCT0 0, we may assume that eCT0 of degree one on all fibres eC and such that L2g−2 compactify eCT0 morphism f : eC / T0 to projective families p : eC CL. Once this is achieved, one can / T which still come with a / C CL. Replacing T0 by (some Next, consider the closure of the cycle / T0 to the subscheme T ′ further étale cover of) T ′ / T0 such that L2g−2 0 / CT0 / C ≃ f ∗ 0 eC ≃ f ∗ / X. (5.7) ZC := ∆f − CT0 ×T [L0] − {x0} × eCT0 20 / / # / / / / / / / / / / / / / / on X ×eC. Here, ∆f is the pull-back of ∆X ⊂ X × X under id × f : X ×eC to X ×eC for C = Ct, t ∈ T0, the cycle ZC yields ZC = ∆fC −C ×[L0]−{x0}×eC with L0 ≃ L0 eC. We shall be interested in the normal function associated to this cycle. Denote by / X × X. Restricted the family of intermediate Jacobians of the family J 3 := J 3(X × eC/T0) π := p0 ◦ pr2 : X × eCT0 tr and J 3 / T0 / T0 with fibres X × eCt. Analogously, J 3 with fibres over t ∈ T0 as described by (5.3) resp. (5.5). In particular, alg denote the transcendental resp. algebraic parts The sheaf of sections of J 3, denoted by the same symbol, is part of the short exact sequence J 3 alg ≃ Pic0(eC/T0) / T0. (5.8) 0 / R3π∗Z / H3/F 2H3 / J 3 / 0, where H3 := R3π∗Ω• X× eC/T0 and F 2H3 := R3π∗Ω≥2 X× eC/T0 . Now, the fibrewise Abel–Jacobi classes AJ(ZCt ) induced by the global cycle ZC yield the normal function i.e. a holomorphic section of J 3 J 3 alg are denoted / J 3 ν ∈ Γ(T0, J 3), / T0, cf. [Vo02]. Similarly, its projections under J 3 / J 3 tr and νtr ∈ Γ(T0, J 3 tr) resp. νalg ∈ Γ(T0, J 3 alg). Corollary 5.7. Under the above choice of ZC and assuming ρ(X) = 1, the algebraic part is trivial. Moreover, for the zero sets of the normal functions and any n > 0 one has νalg ∈ Γ(T0, J 3 alg) = Γ(T0, Pic0(eC/T0)) where N is the degree of (5.6). Z(n · νtr) ⊂ Z(n · N · ν), Proof. The first part is an immediate consequence of Lemma 5.6. For the second, recall that n · AJtr(ZC) = 0 and n · AJalg(ZC) = 0 imply N · n · AJ(ZC) = 0. (cid:3) Clearly, for t ∈ T0 not contained in the zero locus of Z(N ·ν) the Abel–Jacobi class N ·AJ(ZC) In order to prove N · ν 6= 0 and thus and hence N · AJtr(ZC ) of C := Ct are non-trivial. T0 \ Z(N · ν) 6= ∅, one describes its image under the boundary map of (5.8) and shows that it is actually non-trivial. Note that R3π∗Z = H 2(X, Z) ⊗ R1p0∗Z. δ : Γ(T0, J 3) / H 1(T0, R3π∗Z) 21 / / / / / / / / / / / / Lemma 5.8. The boundary class δ(ν) ∈ H 1(T0, R3π∗Z) is non-torsion. Proof. We can restrict to the case dim(T ) = 1. (2g − 2) · [∆X ] − [C] × [C] ∈ H 2(X, Z) ⊗ H 2(X, Z), which is in fact the non-trivial class (2g − 2) · id ∈ [C]⊥ ⊗ [C]⊥ (cf. Remark 5.3). (The argument is easily adapted to the case g = 1.) – One first proves that [ZC] is a non-trivial class in [C]⊥ ⊗ H 2(eCT0, Z) ⊂ H 4(X × eCT0, Z). Note that the cohomology class (2g − 2) · [ZC] ∈ H 2(X, Z) ⊗ H 2(eC, Z) is the pull-back of Since eC / H 2(eC, Z) is injective. The kernel of / H 2(eCT0) is spanned by the divisor classes of the boundary eC \ eCT0 and, as T (X) is H 2(eC) – The further restriction to the fibres eCt, t ∈ T0, is trivial, i.e. under an irreducible Hodge structure of level two, intersects T (X) trivially. / X is dominant, the pull-back T (X) ⊂ [C]⊥ [C]⊥ ⊗ H 2(eCT0 , Z) ⊂ H 4(X × eCT0, Z) / H 0(T0, R4π∗Z) the class [ZC] vanishes. This is just rephrasing that the ZC are cohomologically trivial cycles. – The Leray spectral sequence for π : X × eCT0 d : Ker(cid:16)H 4(X × eCT0, Z) the kernel of which is a quotient of the trivial / H 0(T0, R4π∗Z)(cid:17) / T0 yields a natural (surjective) map / H 1(T0, R3π∗Z), H 2(T0, R2π∗Z) =(cid:0)H 2(X, Z) ⊗ H 2(T0, p0∗Z)(cid:1) ⊕(cid:0)H 0(X, Z) ⊗ H 2(T0, R2p0∗Z)(cid:1) = 0. Here one uses that for dimension reasons all H i(T0, Rjπ∗Z) are trivial for i > 2 and also H 2(T0, p0∗Z) ≃ H 2(T0, R2p0∗Z) = 0 (after shrinking T0, if necessary). This yields a non-trivial class d[ZC] ∈ H 1(T0, R3π∗Z) as the image of [ZC]. Since all the arguments above apply as well to multiples N · ZC, the class d[ZC] is in fact non-torsion. – Eventually, one uses the fact that the class d[ZC] is indeed δ(νZ ). See [Vo02]. (cid:3) Corollary 5.9. Under the above assumptions, there exist at most finitely many points t ∈ T0 with AJtr(ZCt) = 0. Proof. Here one uses that the vanishing locus Z(N · ν) ⊂ T0 of the normal function N · ν ∈ Γ(T0, J 3) is an algebraic set due to [BP09, Sa08]. Thus, if Z(N · ν) is zero-dimensional, then it can only be a finite set of points and hence also Z(νtr) is finite by Corollary 5.7. If Z(N · ν) has positive dimensional components, then repeat the above argument with T0 replaced by such a component. But over the new T0 the class N · δ[ZC] would be trivial, contradicting that δ[ZC] is non-torsion by Lemma 5.8. (cid:3) 22 / / / / / / / Remark 5.10. For simplicity we assumed ρ(X) = 1 in the above discussion. The case ρ(X) > 1 is dealt with similarly, either by replacing H 2(X, Z) throughout by T (X)⊕Z·[C] or by observing that the arguments above prove directly that νtr 6= 0 (without controlling the algebraic part) or by working with the cycle ZX in Remark 5.3, i). 5.4. To conclude the proof of Proposition 5.1, one stratifies L according to the singularity types of the curve and their number of integral components. Since eventually there are only finitely many strata and for each stratum finiteness of curves C := Ct with n · AJtr(κC ) = 0 is assured by Corollary 5.9, the proposition then follows from Corollary 5.5. 5.5. Suppose C ⊂ X / T is a flat family of curves Ct in K3 surfaces Xt. Standard arguments prove that the locus of t ∈ T for which Ct ⊂ Xt is a constant cycle curve is a countable union of Zariski closed subsets of T . Indeed, by using that the (relative) Hilbert scheme (of the relative symmetric products) is a countable union of projective schemes over T one proves that the set of points x ∈ Ct with [x] = cXt is a countable union of Zariski closed subsets (see e.g. [Vo02, Ch. 22.1]). Without bounding the order, the result cannot be improved (cf. Lemma 6.5). For the trivial family X = X × T , Proposition 5.1 proves that the locus of t ∈ T with Ct ⊂ X a constant cycle curve of order ≤ n is Zariski closed. However, the Hodge theoretic line of reasoning seems / T . Only the locus of t ∈ T with n · AJtr(κCt ) = 0 not to extend to non-trivial families X describes a Zariski closed subset. 6. First examples of constant cycle curves It is notoriously difficult to construct rational curves on K3 surfaces, see Section 2.1. As constant cycle curves are natural generalizations of rational curves, in particular with respect to (2.3), it seems worthwhile to work out examples. On the one hand, it will become clear that constant cycle curves are much easier to construct than rational curves, at least if we do not care about their order. On the other hand, the powerful technique developed in [BHT11, LL12] to prove existence of rational curves by reducing modulo p and then using Tate’s conjecture does not seem to apply to constant cycle curves. The main reason being that a curve could be a constant cycle curve modulo infinitely many primes without being a constant cycle curve itself. In fact, all curves over ¯Fp should be constant cycle curves (see Section 9), but not over ¯Q. If not stated otherwise, X will be a K3 surface over an arbitrary algebraically closed field k. 6.1. As mentioned before, all rational curves C ⊂ X are pointwise constant cycle curves. In fact, we have Lemma 6.1. A rational curve C ⊂ X in a K3 surface is a constant cycle curve of order one. 23 / / Proof. This is easy to see by using (3.8), which together with CH2(X × P1) ≃ CH2(X) ⊕ CH1(X) · h, where h is the hyperplane section on P1, shows CH2(X) ≃ CH2(X × U ) for every proper non-empty open subset U ⊂ P1. Hence CH2(X × k(ηC )) ≃ CH2(X), which is torsion free. (cid:3) 6.2. We shall discuss a construction inspired by [Vo12b]. In particular, as remarked in [Vo12b, Lem. 2.3], it shows that the union of all constant cycle curves is dense (and even in the classical topology for complex K3 surfaces). Consider an elliptic K3 surface with a zero-section π : X / / P1, C0 ⊂ X. We first give an ad hoc and geometric description of the constant cycle curves of torsion points on the fibres. For this assume that k is uncountable. The formal definition for arbitrary algebraically closed field k, which is also needed to determine the order of Cn, is given below. Let Cn ⊂ X be the closure of the set of n-torsion points on the smooth fibres Xt, t ∈ P1. Let xt ∈ Xt denote the origin, i.e. {xt} = C0 ∩ Xt. Thus, for any x ∈ Cn ∩ Xt the class [x] − [xt] is n-torsion in CH1(Xt) and hence trivial in CH2(X). Hence, Cn is a pointwise constant cycle curve and by Proposition 3.7 in fact a constant cycle curve. For arbitrary k (as always algebraically closed) these curves (or rather their irreducible com- ponents) are constructed as follows: Let µ ∈ P1 denote the generic point and Xµ the generic fibre of π. For x ∈ Xµ with k(x)/k(µ) finite, let Cx ⊂ X be the curve obtained as the closure of the point x ∈ Xµ ⊂ X. In particular, the generic point ηCx ∈ Cx is just x ∈ Xµ and k(ηCx ) = k(x). The diagonal of Cx as a subvariety ∆Cx ⊂ X × Cx is in fact contained in the surface X ×P1 Cx. Hence, its class [∆Cx] ∈ CH2(X × Cx) is the push-forward under CH1(X ×P1 Cx) / CH2(X × Cx) of the class of the relative diagonal ∆Cx/P1 ⊂ X ×P1 Cx. Restricting to the generic point ηCx shows that [∆CxX×k(ηCx )] ∈ CH2(X × k(ηCx )) is the push-forward of the class of the point x ∈ Xµ in CH1(X ×P1 k(x)) = CH1(Xµ ×k(µ) k(x)). This observation can now be applied to the origin x = o ∈ Xµ, i.e. the intersection of C0 with Xµ, and any point xn ∈ Xµ of order n. For the latter the associated curve Cxn := {xn} ⊂ X is an irreducible component of the curve Cn of n-torsion points in the fibres as considered above. Lemma 6.2. i) The curve Cxn is a constant cycle curve of order dn. ii) The union of all Cxn is dense and, in case k = C, even dense in the classical topology. 24 / Proof. The density statement ii) is obvious, as the set of geometric torsion points is dense in the generic fibre Xµ. For i), note first that [∆C0] ∈ CH2(X ×C0) = CH2(X ×P1) is the class of {x0}×C0+C0×{x0} (for an arbitrary point x0 ∈ C0), which restricted to X ×k(o) is [x0]×k(o). Thus, the restriction of the class of ∆Cxn − {x0} × Cxn to X × k(ηCxn ), which by definition is κCxn , is the image of [xn − o] ∈ CH1(Xµ ×k(µ) k(xn)). Here, since k(o) = k(µ), the class [o] ∈ CH1(Xµ) can be base changed to a class in CH1(Xµ ×k(µ) k(xn)). As xn ∈ Xµ is an n-torsion point, the class [xn − o] is n-torsion and hence also its image (cid:3) κCxn . Remark 6.3. Note that it could happen that Cxn is of order d < n, e.g. when xn is a k(µ)- rational point and, therefore, Cxn a rational curve. But presumably in the generic situation κCxn is of order exactly n. In any case, this elementary construction already provides many examples of constant cycle curves which are not rational. The construction will now be generalized to covering families of elliptic curves, still following [BV04, Vo12b]: Firstly, the existence of nodal rational curves on the generic K3 surface leads to the existence of a dominating family of elliptic curves for every K3 surface. More precisely, for an arbitrary K3 surface X there exists a smooth elliptic surface C / T with a surjective morphism p : C / X. See [HT00, Thm. 4.1] for details in the case of characteristic zero. In positive characteristic use a lift to characteristic zero and reduce the family of curves back to p which yields a family of elliptic curves if X is not unirational. The unirational case being trivial, we shall just ignore it. Note that then, one could moreover assume that the generic fibre Ct is mapped birationally onto its image in X, but we will not need this. Now, choose an ample rational curve C0 ⊂ X and let eC0 be its preimage in C. Replacing / T by its base change to eC0, we can assume that C / C such C that σ0(T ) maps onto a rational curve in X, namely C0. Then consider the curve Cn ⊂ X defined as the closure of the set of images of all points x ∈ Ct in the smooth fibres Ct, t ∈ T , such that n · ([x] − [xt]) = 0 in CH1(Ct), where xt is the point of intersection of Ct with σ0(T ). For the same reason as before, the curve Cn is a (possibly reducible) pointwise constant cycle curve. Note that indeed any torsion point in a smooth fibre Ct is contained in a multi-section of C / T consisting of torsion points in the smooth fibres. Also, the arguments in the proof of Lemma 6.2 still apply. / T admits a section σ0 : T The following was explicitly stated already in [Vo12b] and was also used in [Ma04]. It should be compared to Conjecture 2.3. Corollary 6.4. On any K3 surface X over an arbitrary algebraically closed field k the union [C=ccc C ⊂ X 25 / / / / / / of constant cycle curves (of unbounded order) is dense. For k = C density holds in the classical topology. (cid:3) 6.3. In Section 6.2 we explained how to produce constant cycle curves as multi-sections of elliptic fibrations. But, as was pointed out by Claire Voisin, also smooth (and hence non- rational) fibres can be constant cycle curves. Consider the Kummer surface X obtained as the minimal resolution of the standard involution on the product E1 × E2 of two elliptic curves. Let X / / (E1 × E2)/± / P1 ≃ E1/± be the elliptic fibration induced by the first projection. There are only four singular fibres all of type I ∗ 0 over the two-torsion points. Being rational, they are constant cycle curves of order one. Now consider a torsion point t ∈ E1 of order n 6= 2 and the fibre Ct ⊂ X over ¯t ∈ P1. Lemma 6.5. For a torsion point t ∈ E1 of order n 6= 2 the associated fibre Ct ⊂ X is a smooth constant cycle curve of order dn. / P1 ≃ E2/± Proof. Indeed, if Ct ⊂ X is viewed with respect to the other elliptic fibration X / then it intersects the (smooth) fibres in n-torsion points. In particular, it intersects the generic fibre in an n-torsion point xn ∈ E1 × K(E2/±). Thus, Ct is one of the curves Cxn considered in Lemma 6.2 and hence a constant cycle curve of order dn. (It seems likely that the order equals / CH2(X ×k k(µ)), n, but for proving this one would need to control the kernel of CH1(Xµ) see Section 6.2.) (cid:3) In particular, this construction yields an elliptic K3 surface X / / P1 with infinitely many fibres that are constant cycle curves (of growing order). It is not clear to me whether this holds / P1 for arbitrary elliptic K3 surfaces. Also note that a smooth fibre of an elliptic K3 surface X / can even be a constant cycle curve of order one, see Example 7.3 and the discussion in Section 10.2. The above construction dispels hope expressed in [Ke08, Sec. 17] that the Abel–Jacobi class of a smooth fibre should in particular always be non-torsion. 7. More examples: fixed curves We start this section with the branching curve of a double plane which turns out to be a constant cycle curve of order at most two. This then generalizes to fixed point curves of arbitrary non-symplectic automorphisms. For simplicity we work in characteristic zero. 7.1. Consider a generic double plane, i.e. K3 surface X given as a 2 : 1 cover X / / P2, i : X ∼ / X 26 / / / ramified over a smooth curve C ⊂ P2 of degree six with i the covering involution. Using the eigenspace decomposition one finds that i∗ = −id on CH2(X)0, for CH2(P2)0 = 0. Now consider C as a curve in X and write the class [x] of a point x ∈ C as [x] = cX + αx with αx ∈ CH2(X)0. On the one hand, i∗[x] = [i(x)] = [x] and, on the other, i∗[x] = i∗(cX + αx) = cX − αx. Hence, for x ∈ C, one has 2 · αx = 0 and, since CH2(X) is torsion free, also αx = 0, i.e. C is a (pointwise) constant cycle curve. This provides an explicit example of a constant cycle curve which is smooth and of genus ten (and so in particular not rational). As we shall see in broader generality, C is a constant cycle curve of order one or two. This shows already that in general the genus of a constant cycle curve is not determined by its order. Finding an example of a constant cycle curve of order one that is non-rational is harder, see Corollary 7.2. 7.2. This naive example is now generalized as follows. Suppose f : X of finite order n. Assume that the quotient ∼ / X is an automorphism π : X / / ¯X := X/hf i, which is possibly singular, satisfies CH2( ¯X)0 = 0. Note that due to Bloch’s conjecture (cf. Section 4), which is known for surfaces of Kodaira dimension < 2, the latter condition is equivalent to f ∗ 6= id on H 2,0(X). Suppose a curve C ⊂ X is contained in the fixed point locus Fix(f ). Then a similar trick as above shows that C is a (pointwise) constant cycle curve. Indeed, write [x] = cX + αx for x ∈ C. Then n · [x] = n · cX + n · αx, but on the other hand n · [x] is the pull-back of [π(x)] ∈ CH2( ¯X). Hence, n · [x] = n · cX, which yields n · αx = 0 and, therefore, αx = 0. The calculation in this case suggests that any curve in the fixed point locus of a non-symplectic automorphism of finite order n is a constant cycle curve of order dn. This can be shown rigorously as follows. Proposition 7.1. Let f : X over an algebraically closed field k of char(k) = 0 such that f ∗ 6= id on H 0(X, Ω2 curve C ⊂ X contained in Fix(f ) is a constant cycle curve of order dn. / X be an automorphism of finite order n of a K3 surface X X ). Then any ∼ / CH2(X) induced by the projection π : X / / ¯X. Proof. Consider the pull-back π∗ : CH2( ¯X) Let y := π(x) for a point x ∈ X. Then π∗[y] = n · cX if [x] = cX and π∗[y] = n · [x] for any fixed point x ∈ X. The same holds after base change to any field extension K/k. Apply this to K = k(ηC ) for a curve C ⊂ X in the fixed point locus of f . Then the generic point ηC ∈ C, viewed as a closed point ηC ∈ X × k(ηC ), is fixed under fk(ηC ). Similarly, the generic point η ¯C ∈ ¯C := π(C) ⊂ ¯X can be viewed as a closed point in ¯X × k(η ¯C ) and, moreover, k(η ¯C) = k(ηC ). Then [η ¯C] ✤ k(ηC ) : CH2( ¯X × k(η ¯C)) π∗ / n · [ηC ] under (7.1) / CH2(X × k(ηC )). Since f ∗ 6= id on H 0(X, Ω2 X ) and kod( ¯X) < 2, Bloch’s conjecture holds true for ¯X, i.e. CH2( ¯X) ≃ Z. In particular, there is a distinguished generator c ¯X ∈ CH2( ¯X) with π∗(c ¯X ) = 27 / / / / / n · cX . In order to conclude, it is therefore enough to prove that [η ¯C ] ∈ CH2( ¯X × k(η ¯C )) is in the image of the base change map CH2( ¯X) / CH2( ¯X × k(η ¯C)). Indeed, then the image n · κC of [η ¯C] − c ¯X × k(η ¯C ) = 0 under (7.1) would also be zero. In other words, it is enough to prove that any curve in ¯X (and so in particular ¯C) is a constant cycle curve of order one. As by assumption f has a fixed curve and hence ¯X is rational, this follows from Proposition 4.2, ii). (cid:3) If the 7.3. Non-symplectic automorphisms have been studied intensively in the literature. order is prime, only p = 2, 3, 5, 7, 11, 13, 17, and 19 can occur. Their fixed point loci can be described, which often contains apart from isolated fixed points and rational curves also smooth elliptic curves and even smooth curves of higher genus, see e.g. [AST11]. However, the genus of constant cycle curves obtained in this way is rather small, e.g. for p = 7, 11 at most elliptic curves can occur and for p = 13, 17, 19 all curves in Fix(f ) are in fact rational. The maximal genus g = 11 can be achieved for p = 2. Corollary 7.2. Consider an automorphism f ∈ Aut(X) of order p · q for two primes p 6= q. Assume that f p and f q are both non-symplectic, i.e. f ∗ acts by a primitive (p · q)-th root of unity on H 0(X, Ω2 X ). Then any curve C ⊂ X in Fix(f ) is a constant cycle curve of order one. Proof. Under the assumptions, the curve C would be in the fixed point locus of both, f p and f q. Hence, C is a constant cycle curve of order dividing q and p and, therefore, of order one. (cid:3) Example 7.3. As observed by Dillies in [Di09] and extended by Garbagnati and Sarti in [GS10], non-symplectic automorphisms of prime order p often occur as the square f 2 of an automorphism f with f p a non-symplectic involution. The corollary applies in this situation to curves in Fix(f ). Cases where a non-rational curve is contained in Fix(f ) can be found in [Di09, GS10]. It is worth pointing out that in the explicit example described in [Di09, Sect. 7] the K3 / P1 which is preserved by f and such that surface X comes with an elliptic fibration π : X / one of the smooth(!) fibres is contained in Fix(f ). Concretely, X is the elliptic surface given by the Weierstrass equation y2 = x3 + (t6 − 1)2 and f (x, y, t) = (x, y, ξ · t) with ξ a primitive sixth root of unity. As Alessandra Sarti informs me, this example can be generalized to yield a whole family of elliptic K3 surfaces with a purely non-symplectic automorphism of order six with a smooth elliptic curve in the fixed point locus and thus being a constant cycle curve of order one. For another family of examples see X3,1 described in [AS08, Prop. 4.7]. It is not difficult to write down a square root of the order three automorphism given there, but the elliptic structure in the example is not obvious. This then eventually yields an example of a constant cycle curve of order one which is not rational (but smooth elliptic). 28 / Corollary 7.4. There exist non-rational constant cycle curves of order one. (cid:3) Although it might be difficult to exhibit explicitly constant cycle curves of order one and arbitrary high genus, there does not seem to be any reason why this should not be possible. 8. Bitangent correspondence Here, we exhibit a more involved example, close in spirit to the ones described in Section 7, which leads to constant cycle curves of order at most four and geometric genus 201. These curves will be constructed as the fixed locus of the ‘bitangent correspondence’ for a generic quartic K3 surface. The bitangent correspondence maps a generic point to the second contact point of a bitangent at x. Since generically there are six bitangents at every point, this does not define a map, but we will show that it is well defined on the Chow ring. Its fixed locus is the curve of contact points of hyperflexes. Recall that for a quartic X ⊂ P3 a line ℓ ⊂ P3 is called a bitangent of X if at every x ∈ X ∩ ℓ the intersection multiplicity is at least two. A bitangent ℓ is a hyperflex if there is a unique point of intersection (and, clearly, the intersection multiplicity is four then). In this section we shall work over C. 8.1. Consider a smooth quartic X ⊂ P3 not containing a line. For generic x ∈ X the curve Cx := TxX ∩ X Choose a generic line P1 ⊂ TxX ≃ P2 and consider the linear projection ϕ : Cx / Cx be its normalization and f : eCx has exactly one singularity, a node at x. Let ν : eCx its composition with the inclusion Cx ⊂ X. Then eCx is a smooth curve of genus two. composition ϕ : eCx / P1 from the node x ∈ Cx. As deg(Cx) = 4 and x ∈ Cx is a node, ϕ is of degree two and so is the / P1. By Hurwitz formula, the ramification divisor of ϕ is of degree six. / X Thus, for generic choices there are exactly six lines ℓ1, . . . , ℓ6 ⊂ TxX passing through x and such that they are bitangent to Cx at some other point y1, . . . , y6 ∈ Cx. (This is classical and well known. All it is saying is that there are exactly six bitangents through a generic x ∈ X.) The construction in particular shows that up to two-torsion the points y1, . . . , y6 ∈ eCx are linearly equivalent, as ϕ∗O(1) ≃ O(2 · yi) for all i. Thus, for f∗ : Pic(eCx) / CH2(X) one finds f∗O(1) = 2 · [yi] ∈ CH2(X) and hence 2 · [y1] = . . . = 2 · [y6]. Since CH2(X) is torsion free, in fact [y1] = . . . = [y6] ∈ CH2(X). So the 1 : 6 correspondence x ✤ / {y1, . . . , y6} induces a well-defined involution(!) γ : CH2(X) ∼ / CH2(X), [x] ✤ / [y1]. 29 / / / / / / / / In fact, 6 · γ = [ΓX]∗, where ΓX is the closure of the locus {(x, y) x 6= y, x, y bitangent} ⊂ X × X. Note that for (x, y) ∈ ΓX, i.e. generically x, y is a bitangent, the class α := [x] − [y] satisfies γ(α) = −α. This can be proved in general. Proposition 8.1. The bitangent correspondence γ acts by −id on CH2(X)0. Proof. Write [x] = cX + αx for any x ∈ X. We have to show γ(αx) = −αx. Since any point is rationally equivalent to a cycle contained in a fixed non-empty open subset, we can assume that On the other hand, ω eCx node x ∈ Cx. Since ωCx ≃ ωTxX ⊗ OTxX(4) ≃ OP3(1)Cx , one obtains ≃ ϕ∗OP1(−2) ⊗ O(P yi). x is generic as above. For ϕ : eCx ≃ ν ∗ωCx ⊗ O(−x1 − x2), where x1, x2 ∈ eCx are the two points over the O(cid:16)X yi(cid:17) ≃ ν ∗(OP3(1)Cx ) ⊗ ϕ∗OP1(2) ⊗ O(−x1 − x2) ≃ ν ∗(OP3(3)Cx ) ⊗ O(−3 · x1 − 3 · x2), / P1, Hurwitz formula yields ω eCx where one uses ϕ∗OP1(1) ≃ ν ∗(OP3 (1)Cx ) ⊗ O(−x1 − x2). But then (using (2.1)) [ΓX]∗[x] = f∗(cid:16)X yi(cid:17) = 3 · (h.Cx) − 6 · [x] = 6 · cX − 6 · αx. For a point x with αx = 0 this shows [ΓX]∗cX = 6 · cX and then for arbitrary x ∈ X also [ΓX]∗αx = −6 · αx. (cid:3) Corollary 8.2. The bitangent correspondence γ acts by −id on H 2,0(X). Proof. This follows from the ‘easy direction’ of the general conjectures on the Bloch–Beilinson filtration, see e.g. [Vo02, Prop. 23.18]. See also Remark 8.5. (cid:3) 8.2. There is a more geometric way of defining this correspondence. Consider the universal family of bitangents: p / X BX q FX . More explicitly, bitangents ℓ of X correspond bijectively to points [ℓ] ∈ FX and BX is the variety of all (ℓ, x), with ℓ a bitangent of X and x is a point of contact, i.e. x ∈ X ∩ ℓ. By p / [ℓ]. In particular, q is of degree two and q we denote the projections (ℓ, x) ✤ with a well studied ramification divisor Dhf ⊂ FX (the curve of hyperflexes) and p is of degree six, as there are exactly six bitangents at a generic x. As shown by Tikhomirov and Welters, FX and BX are smooth irreducible surfaces (of general type), see [Ti80, We81]. / x resp. (ℓ, x) ✤ Now consider the covering involution i : BX / BX 30 / / / / / of the double cover q : BX (ℓ, x) and (ℓ, y), where X ∩ ℓ = {x, y}. / FX . So for generic bitangent ℓ, the involution i interchanges Then ΓX ⊂ X × X is the image of the natural morphism g : BX / X × X, (ℓ, x) ✤ / (x = p(ℓ, x), p(i(ℓ, x))). In other words, g(ℓ, x) = (x, y) for ℓ = x, y with x, y ∈ X. The morphism g is generically injective, as the points of contact x, y of a bitangent ℓ clearly determine ℓ when x 6= y. Lemma 8.3. The correspondence [ΓX ]∗ = 6 · γ : CH2(X) / CH2(X) coincides with p∗ ◦ i∗ ◦ p∗ : CH2(X) / CH2(BX ) / CH2(BX ) / CH2(X). The same assertion of course holds on the level of cohomology. (cid:3) Mapping a bitangent ℓ ∈ FX to its intersection X ∩ ℓ defines a closed embedding FX / Hilb2(X). The image is the fixed point locus of the Beauville involution mapping [Z] ∈ Hilb2(X) to the residual intersection of the unique line through Z with X. Corollary 8.4. The surface FX ⊂ Hilb2(X) is a rigid Lagrangian surface of general type. Proof. That FX is Lagrangian is an immediate consequence of γ acting as −id on H 2,0(X). It is rigid, because H 0(FX , NFX /Hilb2(X)) ≃ H 0(FX , ΩX) = 0 due to [We81, (3.43)]. Its canonical bundle ωFX is actually ample, as was shown in [Ti80, We81]. (cid:3) / FX of degree two with the property that V2 H 1(F, Q) ≃ Remark 8.5. Using the geometric description of the bitangent correspondence, one can give another, more roundabout, proof of Corollary 8.2. Suppose γ∗ 6= −id on H 2,0(X). Then q∗ ◦p∗ : H 2,0(X) / H 2,0(FX ) does not vanish. By the work of Tikhomirov [Ti80] and Welters [We81], there exists an étale cover π : F π∗H 2(FX , Q). Here, F is the Fano variety of lines on the double quartic solid Y / P3 branched over X ⊂ P3. This would establish a non-trivial algebraic(!) correspondence between H 2(X, Q) and H 1(F, Q) ⊗ H 1(F, Q), i.e. between the K3 surface X and the square of the abelian variety given by the weight-one Hodge structure of F . This is reminiscent of the Kuga–Satake corres- pondence which is conjectured to be algebraic. However, as François Charles explained to me, the ten-dimensional abelian variety determined by H 1(F, Q) is too small to play a part in the Kuga–Satake correspondence for the generic quartic (which would be of the form A210 with A a simple abelian variety of dimension 29), so that the uniqueness of the Kuga–Satake correspondence eventually leads to a contradiction. Hence, γ∗ = −id on H 2,0(X). 31 / / / / / / / / / / / Remark 8.6. Once Corollary 8.2 has been verified, it can in turn be used to prove Proposition 8.1 by considering the family of all quartics X / O(4). Either one uses the explicit description for the Chow ring of CH∗(X ) by viewing X as a projective bundle over P3 or a technique developed in [Vo12a] which only uses that X is rationally connected. 8.3. Consider now the curve of hyperflexes Dhf ⊂ FX and the image of the curve q−1(Dhf ) ⊂ BX under the projection p : BX / X. This yields the curve of all points x ∈ X such that there exists a hyperflex at x. Chf := p(q−1(Dhf )) ⊂ X Proposition 8.7. For a quartic X ⊂ P3 not containing a line, the curve Chf ⊂ X of contact points of hyperflexes is a constant cycle curve of order n4. Proof. We shall first give two pointwise arguments showing that Chf is a constant cycle curve. The first one seems to (wrongly?) suggest that the order should be at most two, whereas the second one can be turned into a rigorous argument proving the assertion. – Any hyperflex ℓ at a point x ∈ Chf is the limit of proper bitangents ℓt := xt, yt / ℓ with xt, yt both specializing to x. Since γ([xt]) = [yt], Proposition 8.1 shows [xt] + [yt] = 2 · cX which after specializing gives 2 · [x] = 2 · cX . Therefore, Chf is a pointwise constant cycle curve and hence, by Proposition 3.7, also a constant cycle curve. Note that the argument cannot be used to actually prove that the order is n4 (or even better n2), as only 6 · γ is a priori defined by an integral cycle and running the argument again with 6 · γ only shows that the order divides 12. – For x ∈ Chf there exists a line ℓ ⊂ TxX with ℓ and Cx only intersecting in x (with multiplicity four). But ℓ = TxX ∩ H for some hyperplane H ⊂ P3. Hence, 4 · [x] = Cx.(HX ) = 4 · cX by (2.1). – The last argument works for the base change Xk to any field k/C (not necessarily alge- braically closed), as long as the hyperflex ℓ is unique. Indeed, TxXk and Cx ⊂ Xk are defined over k and if ℓ ⊂ TxX¯k is the unique hyperflex, it is left invariant under Gal(¯k/k). Hence, ℓ is and H can be defined over k. Thus, 4 · [x] = 4 · cXk in CH2(Xk). This can be applied to the generic point η of Chf ⊂ X viewed as a closed point η ∈ X × k(η). Since q−1(D) / Chf is generically injective (see the proof of Proposition 8.8), the generic point of Chf admits a unique hyperflex. This proves that κChf is of order n4. (cid:3) Continuing the analogy between the covering involutions of X / / FX in the degree two resp. four case, the curve Chf ⊂ X should be seen as the analogue of the degree / P2. It is curious to note that as in the degree two case, the six branching curve C ⊂ X / / P2 and BX 32 / / / / / generic K3 surface of degree four contains a distinguished curve. It would be interesting to find distinguished curves in K3 surfaces of higher degree curve as well.5 8.4. A little more can be said about the curve Chf ⊂ X, which shall be recorded here. Proposition 8.8. For a quartic X ⊂ P3 not containing a line the curve Chf ⊂ X is a singular irreducible curve in the linear system OX (20) of geometric genus 201. In particular, Chf is not rational. Proof. By construction, Chf is the image of q−1(Dhf ) ≃ Dhf ⊂ FX , where q : BX / FX is a 2 : 1 cover with ramification curve Dhf ⊂ FX . By [We81, (1.6)& p. 40] the curve of hyperflexes Dhf ⊂ FX is a smooth curve of genus 201. A dimension count similar to the one in [We81, p. 17-18] shows that q−1(Dhf ) / Chf is ge- nerically injective. Also, using the notation of [We81, Sec. 3], one computes the intersection number p∗h.[q−1(Dhf )] = p∗h.(σ + p∗h) = 80, where h is the hyperplane section of X and σ the hyperplane section on the space of lines. Since h.h = 4, this shows [Chf ] = 20 · h, i.e. Chf ∈ OX (20). (cid:3) Remark 8.9. Coming back to the variety of bitangents FX ⊂ Hilb2(X). As pointed out by Kieran O’Grady, Hilb2(X) can be seen as a degeneration of an EPW-sextic, such that FX corresponds to the singular locus of the sextic. In the same spirit, whenever an EPW-sextic is of the form Hilb2(X ′) for some K3 surface X ′, then the singular locus of the sextic is likely to give an interesting antisymplectic auto-correspondence for X ′. This can be applied to a generic K3 surface of genus six, see [OG10, Sec. 4]. This should eventually lead to a picture quite similar to the one described here. In particular, one finds as an analogue of the curve Chf the curve of points in X ′ with a conic in a certain ambient Fano threefold with higher contact at this point which is expected to be a constant cycle curve. Remark 8.10. We conclude by making two remarks on the dynamical aspects of the bitangent correspondence. i) On may wonder whether the bitangent correspondence x ✤ / yi can play the role of an endomorphism in other respects as well. E.g. endomorphisms of K3 surfaces have been used to prove potential density of rational points on K3 surfaces over number fields. Although the universal family of bitangents is also defined over the same field as X, the construction seems of little use for this purpose. Indeed the surface BX is of general type and Lang’s conjecture would predict much fewer rational points on BX than on X itself. In other words, mapping x ∈ X to the other points of contact yi ∈ X of bitangents at x, increases the residue field. 5To venture a guess, any distinguished curve, i.e. a curve that is naturally defined in all generic K3 surfaces of fixed degree, should be a constant cycle curve. Compare this to O’Grady’s Franchetta conjecture in [OG13]. 33 / / / ii) If C ⊂ X is a constant cycle curve in an arbitrary K3 surface and f : X / X is any automorphism, then f (C) ⊂ X is again a constant cycle curve (of the same order). But, since the generic K3 surface does not admit non-trivial automorphisms, this method fails in the generic case. However, for the generic quartic X ⊂ P3 the bitangent correspondence can sometimes replace the missing automorphisms. Indeed, if C ⊂ X is a constant cycle curve, then p(i(p−1(C))) is again a constant cycle curve. This can be applied to e.g. C = Chf , in which case p−1(C) = q−1(Dhf ) ∪ C ′. The curve q−1(Dhf ) is of course invariant under the involution i, but C ′ is not and produces a new constant cycle curve p(i(C ′)) in X. ∼ / 9. Finite fields In this section we study K3 surfaces over finite fields Fq, q = pd, with p 6= 2. Contrary to the case of K3 surfaces over ¯Q, it is easy to see that CH2(X) ≃ Z for any X over ¯Fp. Indeed, any finite collection of points is contained in some curve C ⊂ X. As one always finds a finite field Fq ⊂ ¯Fp over which the finitely many points and the curve are defined, the finiteness of Fq-rational points of Pic0(C) and the torsion freeness of CH2(X) (over the algebraically closed field ¯Fp) are enough to conclude. But the Bloch–Beilinson philosophy in positive characteristic also predicts CH2(X) ≃ Z for X over the algebraic closure of Fp(t). Surprisingly, explicit examples of K3 surfaces over function fields verifying the conjecture can actually be worked out, unlike the case of K3 surfaces over number fields. Remark 9.1. For k = ¯Fp the Chow group is not expected to increase when passing from X over k to X ×k k(t). This does not hold for other fields. Indeed, similarly to Bloch’s argument / CH2(X ×k k(ηX )) is not surjective for a K3 surface X in arbitrary showing that CH2(X) characteristic as long as ρ(X) 6= 22, Green, Griffiths, and Paranjape showed in [GGP04] that in characteristic zero the Chow group grows already when passing to an algebraically closed field of transcendence degree one over the base field. Cf. Corollary 3.11 for a weaker version that also works in positive characteristic. 9.1. Due to CH2(X) ≃ Z for a K3 surface X over ¯Fp, any curve C ⊂ X is a pointwise constant cycle curve. But the stronger statement is also conjectured to hold due to the following Proposition 9.2. Let X be a K3 surface over ¯Fp. Then X ×Fp Beilinson conjecture, i.e. CH2(X ×Fp cycle curve. Fp(t) satisfies the Bloch– Fp(t)) ≃ Z, if and only if every curve C ⊂ X is a constant Proof. Clearly, CH2(X ×Fp Fp(t)) ≃ Z if and only if the pull-back (9.1) CH2(X) / CH2(X ×Fp Fp(t)) 34 / / is an isomorphism. Due to Lemma 3.8, a curve C ⊂ X is a constant cycle curve if and only if [ηC ] ∈ CH2(X × k(ηC )) is in the image of the pull-back. Assuming the Bloch–Beilinson conjecture, the latter now follows from choosing an embedding ¯Fp ⊂ k(ηC ) ⊂ k(ηC ) ≃ Fp(t). Conversely, a closed point x ∈ X × Fp(t) either projects to a closed point in X or the generic point of a curve C ⊂ X. In the first case, [x] is in the image of (9.1), whereas in the latter [x] = [ηC ] with the generic point ηC ∈ C viewed as a closed point of X×k(ηC ) = X×Fp(t). But if C is a constant cycle curve, then up to torsion [ηC ] is in the image of CH2(X) / CH2(X×k(ηC )) and, therefore, [x] is in the image of (9.1). (cid:3) We stress again, that due to [GGP04] the corresponding statement is false for X over ¯Q, i.e. there always exist curves C ⊂ X which are not constant cycle curves. However, according to Conjecture 2.2 all curves are expected to be pointwise constant cycle curves and it seems likely that every point is at least contained in a constant cycle curve (cf. Conjecture 2.4). Note that any linear system L could contain countably many constant cycle curves. So a priori the finiteness of constant cycle curves of bounded order does not prove the existence of curves that are not constant cycle curves. The existence for surfaces over ¯Q as proved in [GGP04] eventually relies on results of Terasoma and in fact shows the existence of infinitely many such curves in any ample linear system. 9.2. The Bloch–Beilinson conjecture for function fields is related to the conjectured finite- dimensionality in the sense of Kimura and O’Sullivan. To be more precise, consider a K3 surface X and a curve C ⊂ X over a finite field Fq. Suppose Kimura–O’Sullivan finite-dimensionality holds for X and eC and hence for X × eC (cf. [Ki05, Prop. 5.10]). Then the Tate conjecture T 1 for X (using the recent [Cha12, Ma13, Mau12]) and eC, which implies the Tate conjecture T 1 for X × eC and by duality the Tate conjecture T 2 for X × eC (see [Ta94] or [Mi07, Cor. 2.2, Thm. 1.4]), yields CH2(X × eC) ⊗Z Qℓ ∼ / H 4((X × eC)¯Fq , Qℓ(2))Gal(¯Fq /Fq), see [Ja07, Thm. 0.4, Thm. 5.2], [Ka03, Thm. 1], or [An05, Thm. 4.2]. The cycle ZC is homo- logically trivial, i.e. 0 = [ZC = ∆fC − C × [L0] − {x0} × eC] ∈ H 4((X × eC)¯Fq , Qℓ(2))Gal(¯Fq /Fq) (see Section 5.1) and, under the assumption of Kimura–O’Sullivan finite-dimensionality, the cycle ZC on X × eC must then also be rationally equivalent to zero up to torsion. Hence, its restriction κC ∈ CH2(X × k(ηC )) is also torsion. Thus, the Bloch–Beilinson conjecture holds for the K3 surface X × Fq(t), i.e. CH2(X × Fq(t)) ≃ Z. This remark immediately produces actual examples of K3 surfaces over function fields for which the Bloch–Beilinson conjecture can be confirmed. Note that this is in contrast to the number field case where not a single K3 surface is known to satisfy CH2(X ¯Q) ≃ Z. It would 35 / / also be very interesting to find an example of a non-isotrivial K3 surface over Fp(t) satisfying the Bloch–Beilinson conjecture. Example 9.3. Kimura–O’Sullivan finite-dimensionality is known for rational Chow motives in the tensor subcategory generated by Chow motives of abelian varieties, cf. [Ki05, Ex. 9.1]. Also, quotients of finite-dimensional Chow motives are again finite dimensional. Hence, Kummer surfaces are known to be finite-dimensional. Therefore, any Kummer surface X over Fq yields a K3 surface XFq(t) over Fq(t) which satisfies the Bloch–Beilinson conjecture, i.e. CH2(X×Fq(t)) ≃ Z. Finite dimensionality of Kummer surfaces is eventually deduced from the fact that they are dominated by products of curves and, in fact, in [Sch87] the Bloch–Beilinson conjecture for X × Fq(t) was proved for any variety X dominated by a product of curves. The reduction modulo p of Voisin’s example in [Vo96, Sect. 3.3] is likely to provide other examples. See [Pe12] for more on finite-dimensionality of K3 surfaces (in characteristic zero). Proposition 9.4. Let X be a K3 surface over a finite field Fq for which Kimura–O’Sullivan finite-dimensionality holds, e.g. X a Kummer surface. Then every curve in X¯Fq is a constant cycle curve. (cid:3) By Proposition 4.2, the result holds also true for unirational (and hence by [Li13] for all supersingular) K3 surfaces. Remark 9.5. In the situation of the corollary and assuming that X is not unirational, it would be interesting to decide whether among the curves in a fixed linear system L, which are all constant cycle curves, there are at most finitely many of bounded order, i.e. whether the analogue of Proposition 5.1 holds. Also, is there an explicit example (for which Kimura– O’Sullivan finite-dimensionality is known) where one can show that all curves are constant cycle curves directly (i.e. geometrically)? 9.3. We conclude this section with a result lending evidence to Conjecture 2.4. Proposition 9.6. Every closed point x ∈ X in a K3 surface over ¯Fp is contained in a constant cycle curve x ∈ C ⊂ X. Proof. If X is unirational, then every point x ∈ X is in contained in a rational curve, which yields the assertion. Thus, we may assume that X is not unirational. Next we use the construction of Section 6.2. So we pick a smooth elliptic surface C / T with a surjective morphism p : C / X. As before, by a further base change, we can assume that C / T comes with a zero-section that maps to a constant cycle curve in X. Now, all singular fibres are rational and, therefore, map to constant cycle curves in X. A point x ∈ Ct in a smooth fibre is automatically torsion and hence contained in one of the curves Cn ⊂ C of fibrewise torsion points. By Section 6.2, the curves Cn are constant cycle curves. This proves the assertion. (cid:3) 36 / / / Note that a stronger result for Kummer surfaces, namely the existence of a rational(!) curve through every point, has been proved already by Bogomolov and Tschinkel in [BT05]. The assumption that X is Kummer is in [BT05] used for an explicit geometric construction for Jacobian Kummer surfaces. 10. Further questions and remarks 10.1. Let X be a K3 surface over an arbitrary algebraically closed field k. It seems not impossible that any closed point x ∈ X with [x] = cX is in fact contained in a constant cycle curve C ⊂ X (see Conjecture 2.4). Corollary 3.10 provides some weak evidence. Note however that this should fail in general (probably whenever k is uncountable and algebraically closed) when constant cycle curves are replaced by rational curves. E.g. for a complex K3 surface and a non-rational constant cycle curve C ⊂ X (examples have been given above), only countably many of the points x ∈ C can be contained in some rational curve. Indeed, there are at most countably many rational curves contained in X. 10.2. Due to Proposition 5.1, on a complex K3 surface only finitely many curves in a fixed linear system L can be constant cycle curves of a given order n. This prompts two questions: i) Can the number be determined, e.g. in geometrically interesting situations? ii) Is this number deformation invariant? Both aspects have been addressed extensively for rational curves in the context of the Yau– Zaslow conjecture, see [KMPS10]. Once the counting is done properly, the number of rational curves (counted with multiplicities) is deformation invariant and is given by a certain generating series. For the time being it is not at all clear (to me) whether a similar picture is to be expected for constant cycle curves. In fact, the observation that apart from rational curves also non-rational curves (even smooth ones) can occur as constant cycle curves of order one is a little suspicious. One would need to add those to the (Gromov–Witten) number of rational curves. It is to be expected that an analogous picture will emerge for curves C ⊂ X that are 10.3. special with respect to O’Grady’s filtration (see [OG13]): S0(X) ⊂ S1(X) ⊂ . . . ⊂ Sd(X) ⊂ . . . ⊂ CH2(X). Here, Sd(X) is the set of cycles that can be written as [Z] + m · cX with Z an effective cycle of degree d. Equivalently, α ∈ Sd(X) if and only if there exists a possibly reducible curve C ⊂ X / CH2(X)) (see [OG13, Vo12b]). In particular, such that g(eC) ≤ d and α ∈ Im(fC∗ : Pic(eC) Note that the kernel of fC∗ : Pic0(eC) abelian subvarieties Ai ⊂ Pic0(eC). One defines dim(Ker(fC∗)) as the minimum of all dimensions dim(Ai). So, dim(Ker(fC∗)) = g if and only if C is a constant cycle curve (we work over S0(X) = Z · cX . / CH2(X) is a countable union of translates Li + Ai of 37 / / C), which is equivalent to Im(fC∗) ⊂ S0(X). More generally, it is not difficult to see that dim(Ker(fC∗)) ≥ g − d implies Im(fC∗) ⊂ Sd(X) and I would expect that with the techniques from [Vo12c] also the converse can be proved. Then in any linear system L the locus of curves C with Im(fC∗) ⊂ Sd(X) should be a countable union of Zariski closed subsets, but in order to get honest Zariski closed subsets, one would first need to introduce the analogue of the order of a constant cycle curve and then restrict to those of finite order. For a lack of a better name one could call a curve C ⊂ X, say integral, d-special if Im(fC∗) ⊂ curves would simply be constant cycle curves. So the first question one needs to address is how Sd(X). Of course, there would not be anything special about C for d ≥ g(eC) and 0-special to define the order of a d-special curve for d < g(eC). 10.4. We have not yet explored constant cycle curves from the infinitesimal point of view. As explained by Bloch, H 2(X, OX ) ⊗ ΩC/Q should be viewed as the ‘tangent space’ of CH2(X). / CH2(X) induces a natural map between the / H 2(X, OX ) ⊗ ΩC/Q. For a constant cycle curve the map should be trivial and, as addressed in [GG05], it is interesting to see how much information about fC∗ is actually encoded by the derivatives (at all points). Is there anything special about the derivative in points x ∈ C with [x] = cX ? For any curve C ⊂ X the induced fC∗ : Pic0(eC) tangent spaces H 1(eC, O eC ) 11. Appendix by Claire Voisin The aim of this appendix is to prove Conjecture 2.3 for the generic (and not only general!) K3 surface. More precisely, we prove Theorem 11.1. Let X be a complex K3 surface admitting a non-isotrivial one-parameter family of elliptic curves which is not an elliptic pencil. Then X admits infinitely many constant cycle curves of bounded order, the union of which is dense in the classical topology. Since the assumptions are Zariski open conditions and families of elliptic curves satisfying the assumptions can be found for the general K3 surface of Picard number one, the theorem does imply Conjecture 2.3 for the generic complex K3 surface. 11.1. To set up notation, we spell out the assumptions of the theorem: We assume that there exist a smooth elliptic surface q : C / X with the following properties: i) The family q : C / T and a surjective morphism p : C / T is not isotrivial, i.e. there does not exist any dominant quasi-finite / T ′ is a trivial family of / T such that the base change q′ : C′ := C ×T T ′ morphism T ′ elliptic curves. ii) For generic y ∈ X there exist at least two smooth fibres Ct, Ct′ ⊂ C such that p(Ct) and p(Ct′) intersect transversally in y. 38 / / / / / / / As in Section 6.2, we can always reduce to the situation that q : C / T comes with a zero- section, the image of which is a constant cycle curve in X (we could even assume the image to be a rational curve). The origin of a smooth fibre Ct fixed by the zero section will be called xt and torsion in the fibres is considered with respect to this choice. 11.2. The technique to produce new constant cycle curves is by multiplying a given one with respect to the additive structure of the fibres of q : C / T . Under the above hypotheses one proves: Lemma 11.2. Assume D ⊂ C is a multi-section of q : C / T such that: i) The image C := p(D) is a constant cycle curve of order n and ii) The intersection of D with the generic fibres Ct contains a non-torsion point. Then there exists an N > 0 such that the union of all constant cycle curves of order ≤ N is dense in the classical topology. Proof. If C ⊂ X is the image of a finite morphism f : D / the image κC × k(ηD) of κC under the base change map / X from an integral curve D, then can be described as the restriction of [Γf − {x0} × D] ∈ CH2(X × D). CH2(X × k(ηC )) / CH2(X × k(ηD)) Next consider fibrewise multiplication by m for the family q : C / T , which defines a mor- / U on some open set U ⊂ C. Denote the graphs of the two morphisms / X by Γ resp. Γm. We claim that, possibly after shrinking phism µm : U p : U U , there exists an Nm > 0 such that / X and pm := p ◦ µm : U Nm · ([Γm − {x0} × U ] − m · [Γ − {x0} × U ]) = 0 in CH2(X × U ). Indeed, the correspondence [Γm]∗ : CH2(U ) / CH2(X) maps the class of a point x ∈ Ct to m · [x] + (1 − m) · [xt], where xt ∈ Ct is the origin of the elliptic curve Ct. Since [p(xt)] = [x0], one finds [Γm − {x0} × U ]∗ = m · [Γ − {x0} × U ]∗ : CH2(U ) / CH2(X). Using Bloch–Srinivas [BS83], one concludes the existence of Nm (after shrinking U if necessary). Let now D ⊂ C be as assumed. Fix m and choose Nm above such that nNm. Then define m(D) (or, more precisely, µmk (D ∩ U )) and C(mk) := p(D(mk)) = D(mk) := µmk (D) = µk pmk (D). Since κC × k(ηD) is the specialization of [Γp − {x0} × U ] and similarly for all C(mk), one finds Nm ·(cid:16)κC(mk ) × k(ηD) − m · κC(mk−1) × k(ηD)(cid:17) = 0 in CH2(X × k(ηD)). Since nNm and, therefore, Nm · κC = 0, one obtains by recursion that Nm · (κC(mk ) × k(ηD)) = 0 for all k. The base change induced by pmk : D / / C(mk) factorizes via (11.1) CH2(X × k(ηC(mk ))) / CH2(X × k(ηD(mk))) / CH2(X × k(ηD)). 39 / / / / / / / / / / / / / D(mk)) ≤ (D.Ct), Since deg(p : D(mk) there exists an N ′ > 0 independent of m and k, such that the kernel of (11.1) is annihilated by N ′. This shows that the curves C(mk) are constant cycle curves of order ≤ Nm · N ′. / X) and deg(µmk : D / / C(mk)) ≤ deg(p : C If the intersection of D with the generic fibre Ct contains a non-torsion point, then the union S D(mk) ⊂ C of constant cycle curves of order ≤ Nm · N ′ is dense in the Zariski topology and so is its image S C(mk) ⊂ X. To obtain density in the classical topology one needs to prove that the integers Nm can be chosen independent of m. Consider the map σ : C ×T C / X that is given as the composition of the summation in the fibres Ct with the projection p. For points x, x′ ∈ C in a smooth fibre Ct, we have [p(x + x′)] = [p(x)] + [p(x′)] − cX in CH2(X), as by assumption the class of the image p(xt) of the origin xt ∈ Ct is cX. Now consider the codimension two cycle / C Γ := Γσ − Γp◦pr1 − Γp◦pr2 + {x0} × C ×T C on X ×C ×T C. Then by Bloch–Srinivas [BS83], there exists an integer N such that the N ·[Γ] = 0 in CH2(X × V ) for some non-empty open subset V ⊂ C ×T C. Restriction to the diagonal in C ×T C yields N · ([Γ2] − 2 · [Γ] + [{x0} × W ]) = 0 in CH2(X ×W ) for a certain non-empty open set W ⊂ C (the restriction of V with the diagonal). By induction on m and by restricting the cycle Γ to the surface that is given as the image of id × µm−1 : C / (x, (m − 1) · x), one proves similarly / C ×T C, x ✤ N · ([Γm] − m · [Γ] + (m − 1) · [{x0} × W ]) = 0 in CH2(X × W ). As before, this shows that all curves C(m) are constant cycle curves of order ≤ N · N ′ (independent of m). If the generic fibre Ct is viewed as a torus R2/Z2, then the intersection of S D(m) with Ct contains a set of the form Z · (x1, x2) ⊂ R2/Z2 with x1 or x2 non-torsion. Therefore, the closure of Ct ∩S D(m) is either the full torus Ct = R2/Z2 or a finite union of translates of circles S1 ⊂ Ct = R2/Z2. The second possibility is ruled out, since the monodromy action on H1(Ct, Z) would then act via a finite group on the class of the finite union of circles contradicting the fact that the monodromy group is of finite index in Sl(2, Z). The argument can be made more explicit as follows. If the intersection is generically not dense, then the components of its closure satisfy a linear equation of the form a · x1 + b · x2 = 0 with a, b ∈ Z not both zero. If assumed coprime, a and b are unique up to sign. Such an equation exists for a countable union of real analytic subsets of D over the regular locus of q. Thus, (a, b) is locally constant up to sign and defines a section of the pull-back of R1q∗Z to a dense open subset of D (or rather the appropriate double 40 / / / / / / cover that accommodates for the sign ambiguity). The latter contradicts the non-isotriviality of C ×T D / / D. The last part of the proof is inspired by a similar argument due to Chen and Lewis in (cid:3) [CL13]. 11.3. Let Cn ⊂ C be the closure of the set of all n-torsion points x ∈ Ct in smooth fibres Ct and let Cn := p(Cn). Then by Section 6.2 the curve Cn is a (possibly reducible) constant cycle curve of order dn. Of course, the curves Cn cannot be used as input for Lemma 11.2, but the curve eCn := p−1(p(Cn)) might contain another component D satisfying the assumptions i) (which is automatic) and ii) of Lemma 11.2. We will argue by contradiction and assume: (∗) If D ⊂ p−1(p(Cn)) = p−1(Cn) is an irreducible component, then its intersection with the generic fibres Ct contains only torsion points. Consider a multi-section D ⊂ C. We call D flat at a point x ∈ D if locally analytically D C := R1q∗C. E.g. the multi-sections Cn are flat in all points of can be lifted to a flat section of H 1 intersection with smooth fibres. Note that the notion makes sense for locally analytic sections as well. Lemma 11.3. Suppose D ⊂ C is an irreducible multi-section of q : C locally around a point x ∈ D which is a non-torsion point in a smooth fibre Ct. Then q : C is isotrivial (contradicting the assumption i) in 11.1). / T and assume it is flat / T Proof. The assumption immediately implies that D is flat in all points of intersection with smooth fibres. Consider the base change qD : C ×T D / / D. In addition to the pull-back of the zero-section, it comes with a natural flat section which is fibrewise non-torsion. This flat section of qD defines a section of R1qD∗C/R1qD∗Z (over the regular part) and, therefore, corresponds to a monodromy invariant element of C2/Q2. But for a non-isotrivial family the monodromy group is a finite index subgroup Γ ⊂ Sl(2, Z) and thus (C2/Q2)Γ = 0. (cid:3) Let now x, x′ ∈ C be generic points with p(x) = p(x′) and x ∈ D ⊂ C a local multi-section through x. Consider a component D′ ⊂ p−1(p(D)) through x′. A priori, D might be flat at x without D′ being flat at x′. However, if x is n-torsion in the fibre Ct, t = q(x), then x ∈ Cn and hence x′ ∈ eCn. So assuming (∗), the components of eCn containing x′ are again contained in / T . Using a density argument this is some Cm and are, therefore, flat multi-sections of q : C enough to conclude the general case: Lemma 11.4. Let x, x′ ∈ C be generic points with p(x) = p(x′). Assume x ∈ D ⊂ C is a local analytic section through x and D′ ⊂ p−1(p(D)) is a component through x′. Under the assumption (∗) one has: If D is flat at x, then D′ is flat at x′. Proof. Locally, flat sections are given by constant sections of H 1 C∆ = C2 × ∆, where ∆ ⊂ T is a disk around a generic point of T . For a torsion section D, which corresponds to a section 41 / / / contained in Q2 × ∆, assumption (∗) implies that D′ is again torsion and hence flat. In order to prove the assertion, it is enough to prove that this holds true on an analytically dense set of flat C∆⊗O does not admit local sections. However, since C flat sections and, therefore, Γ(∆, R1q∗C/R1q∗Z) can be identified with the set of flat sections / ∆. Under this identification, the subset of sections contained in R1q∗Q/R1q∗Z is of C∆ analytically dense in Γ(∆, R1q∗C/R1q∗Z). (cid:3) / T is not isotrivial, R0q∗(ΩC/T )∆ ⊂ H 1 End of proof of Theorem 11.1. Choose generic points x, x′ ∈ C with p(x) = p(x′) and assume x ∈ Ct, t = q(x), is not torsion. Consider the two curves E := p(Ct) and E′ := p(Ct′), t′ = q(x), in X. Then we can assume that the two curves E and E′ intersect transversally in y := p(x) = p(x′) ∈ E ∩ E′. Now choose a flat local analytic section x ∈ D ⊂ C of / T such that p(D) is tangent to E′ at y. (This can always be achieved over points t ∈ T q : C with maximal VHS. Indeed a local calculation shows that every non-vertical tangent direction v ∈ TxC in a point x ∈ C over t can be integrated to a flat local section.) Then the component D′ ⊂ p−1(p(D)) through x′ is tangent to the fibre Ct′, t′ = q(x′). Since by Lemma 11.4, D′ is also flat, in fact D′ ⊂ Ct′. But then p−1(E′) contains D and hence D can be extended to an algebraic multi-section / T through x ∈ C which, moreover, is flat locally around x and non-torsion. Then, by (cid:3) / T is isotrivial, which is excluded by assumption on C of C Lemma 11.3, C / T . References [An05] [AS08] Y. André Motifs de dimension finie (d’après S.-I. Kimura, P. O’Sullivan ...), Séminaire Bourbaki. Vol. 2003/2004. Astérisque No. 299 (2005), Exp. No. 929, viii, 115–145. M. Artebani, A. Sarti Non symplectic automorphisms of order 3 on K3 surfaces, Math. Ann. 342 (2008), 903–921. [AST11] M. Artebani, A. Sarti, S. Taki K3 surfaces with non-symplectic automorphisms of prime order, Math. [BV04] [Bl80] [BKL76] [BS83] [BHT11] [BT00] [BT05] [BP09] Z. 268 (2011), 507–533. A. Beauville, C. Voisin On the Chow ring of a K3 surface, J. Alg. Geom. 13 (2004), 417–426. S. Bloch Lectures on algebraic cycles, Duke Univ. Math. Series, IV. (1980). S. Bloch, A. Kas, D. Lieberman Zero cycles on surfaces with pg = 0, Comp. Math. 33 (1976), 135–145. S. Bloch, V. Srinivas Remarks on correspondences and algebraic cycles, Amer. J. Math. 105 (1983), 1235–1253. F. Bogomolov, B. Hassett, Y. Tschinkel Constructing rational curves on K3 surfaces, Duke Math. J. 157 (2011), 535–550. F. Bogomolov, Y. Tschinkel Density of rational points on elliptic K3 surfaces, Asian J. Math. 4 (2000), 351–368. F. Bogomolov, Y. Tschinkel Rational curves and points on K3 surfaces, Amer. J. Math. 127 (2005), 825–835. P. Brosnan, G. Pearlstein Zero loci of admissible normal functions with torsion singularities, Duke Math. J. 150 (2009), 77–100. 42 / / / / / / [Cha12] [Cha13] [Ch02] [CL13] [Di09] [Fu98] [GS10] F. Charles The Tate conjecture for K3 surfaces over finite fields, arXiv:1206.4002. Invent. math. to appear. F. Charles Progrès récents sur les fonctions normales (d’après Green–Griffiths, Brosnan–Pearlstein, M. Saito, Schnell...), arXiv:1301.7235. Séminaire Bourbaki (2013). X. Chen A simple proof that rational curves on K3 are nodal, Math. Ann. 324 (2002), 71–104. X. Chen, J. Lewis Density of rational curves on K3 surfaces, Math. Ann. 356 (2013), 331–354. J. Dillies Order 6 non-symplectic automorphisms of K3 surfaces, arXiv:0912.5228. W. Fulton Intersection theory, volume 2 of Ergebnisse der Mathematik und ihrer Grenzgebiete. 3. Folge. Springer-Verlag, Berlin, (1998). A. Garbagnati, A. Sarti On symplectic and non–symplectic automorphisms of K3 surfaces, arXiv:1006.1604. S. Gorchinskiy Generation of modules and transcendence degree of zero-cycles, arXiv:1210.0233. M. Green Higher Abel–Jacobi maps, Proc. Int. Congr. Math. Vol. II (Berlin, 1998), 267–276. [Go12] [Gr98] [GGP04] M. Green, P. Griffiths, K. Paranjape Cycles over fields of transcendence degree 1, Michigan Math. [GG05] [GG03] [GP03] [HT00] [Hu13] [Ja07] [Ka03] J. 52 (2004), 181–187. M. Green, P. Griffiths On the tangent space to the space of algebraic cycles on a smooth algebraic variety, Annals of Mathematics Studies, 157. Princeton University Press (2005). M. Green, P. Griffiths An interesting 0-cycle, Duke J. Math. 119 (2003), 261–313. V. Guletskii, C. Pedrini Finite dimensional motives and the conjecture of Beilinson and Murre, K-theory 550 (2003), 1–21. B. Hassett, Y. Tschinkel Abelian fibrations and rational points on symmetric products, Internat. J. Math. 11 (2000), 1163–1176. D. Huybrechts Lectures on K3 surfaces, http://www.math.uni-bonn.de/people/huybrech/K3.html U. Jannsen On finite-dimensional motives and Murre’s conjecture, in ‘Algebraic cycles and Motives’ Vol. 2, LMS Lecture Note Ser., vol. 344, Cambridge Univ. Press (2007), 112–142. B. Kahn Équivalences rationnelle et numérique sur certaines variétés de type abélien sur un corps fini, Ann. Sci. Éc. Norm. Sup. 36 (2003), 977–1002. [Ke06] [KMP07] B. Kahn, J. Murre and C. Pedrini On the transcendental part of the motive of a surface, in ‘Algebraic cycles and Motives’ Vol. 2, LMS Lecture Note Ser., vol. 344, Cambridge Univ. Press (2007), 143–202. M. Kerr A survey of transcendental methods in the study of Chow groups of 0-cycles, in ‘Mirror Symmetry V’ (Lewis, Yui and Yau, eds.), AMS/IP Stud. Adv. Math. 38 (2006), 295–350. M. Kerr Higher Abel–Jacobi maps for 0-cycles, J. K-Theory 2 (2008), 41–101. S. Kimura Chow groups are finite dimensional, in some sense, Math. Ann. 331 (2005), 173–201. [Ke08] [Ki05] [KMPS10] A. Klemm, D. Maulik, R. Pandharipande, E. Scheidegger Noether–Lefschetz theory and the Yau– [Le86] [LL12] [Li13] [Ma04] [Ma13] [Mau12] [Mi82] [Mi07] Zaslow conjecture, J. Amer. Math. Soc. 23 (2010), 1013–1040. F. Lecomte Rigidité des groupes de Chow, Duke Math. J. 53 (1986), 405–426. J. Li, C. Liedtke Rational curves on K3 surfaces, Invent. Math. 188 (2012), 713–727. C. Liedtke, Supersingular K3 surfaces are unirational, arXiv:1304.5623. C. Maclean Chow groups of surfaces with h2,0 ≤ 1, C. R. Math. Acad. Sci. Paris 338 (2004), 55–58. K. Madapusi Pera The Tate conjecture for K3 surfaces in odd characteristic, arXiv:1301.6326. D. Maulik Supersingular K3 surfaces for large primes, arXiv:1203.2889. J. Milne Zero cycles on algebraic varieties in nonzero characteristic: Rojtman’s theorem, Comp. Math. 47 (1982), 271–287. J. Milne The Tate conjecture over finite fields, arXiv:0709.3040. 43 [MM83] [Mu68] [OG10] [OG13] [Pe12] [Ro80] [Sa08] [Sch87] [Ta94] [Ti80] [Vo96] [Vo99] [Vo02] [Vo12a] [Vo12b] [Vo12c] [We81] [Yi13] S. Mori, S. Mukai The uniruledness of the moduli space of curves of genus 11, Alg. geometry (Tokyo/Kyoto, 1982), 334–353, Lecture Notes in Math., 1016, Springer, Berlin, (1983). D. Mumford Rational equivalence of 0-cycles on surfaces, J. Math. Kyoto Univ. 9 (1968) 195–204. K. O’Grady Double covers of EPW-sextics, arXiv:1007.4952v3. K. O’Grady Moduli of sheaves and the Chow group of K3 surfaces, arXiv:1205.4119. J. Math. Pures Appl. (2013). to appear. C. Pedrini On the finite dimensionality of a K3 surface, Manuscripta Math. 138 (2012), 59–72. A. Rojtman The torsion of the group of 0-cycles modulo rational equivalence, Ann. Math. (2) 111 (1980), 553–569. M. Saito Hausdorff property of the Néron models of Green, Griffiths and Kerr, arXiv:0803.2771v4. C. Schoen Zero cycles modulo rational equivalence for some varieties over fields of transcendence degree one, Algebraic geometry, Bowdoin, 1985 (Brunswick, Maine, 1985), Proc. Sympos. Pure Math., 46, Part 2, (1987), 463–473. J. Tate Conjectures on algebraic cycles in ℓ-adic cohomology, Motives (Seattle, WA, 1991), Proc. Sympos. Pure Math. 55, Part 1, AMS (1994), 71–83. A. Tikhomirov Geometry of the Fano surface of a double P3 branched in a quartic, Izv. Akad. Nauk SSSR Ser. Mat. 44 (1980), 415–442. C. Voisin Remarks on zero-cycles of self-products of varieties, Moduli of vector bundles (Sanda, 1994; Kyoto, 1994), Lecture Notes in Pure and Appl. Math. 179, Dekker, New York, (1996), 265–285. C. Voisin Some results on Green’s higher Abel-Jacobi map, Ann. of Math. 149 (1999), 451–473. C. Voisin Théorie de Hodge et géométrie algébrique complexe, Cours spécialisés 10. SMF (2002). C. Voisin Bloch’s conjecture for Catanese and Barlow surfaces, arXiv:1210.3935. C. Voisin Rational equivalence of 0-cycles on K3 surfaces and conjectures of Huybrechts and O’Grady, arXiv:1208.0916. C. Voisin Symplectic involutions of K3 surfaces act trivially on CH0, Documenta Math. 17 (2012) 851–860. G. Welters Abel–Jacobi isogenies for certain types of Fano threefolds, Mathematical Centre Tracts 141 (1981). Q. Yin On a result of Green and Griffiths, arXiv:1302.6378. 44
1706.05366	2	1706	2018-05-16T20:08:50	General Variational Formulas for Abelian Differentials	[ "math.AG" ]	We use the jump problem technique developed in a recent paper by Grushevsky, Krichever and the second author to compute the variational formula of any stable differential and its periods to arbitrary precision in plumbing coordinates. In particular, we give the explicit variational formula for the degeneration of the period matrices near an arbitrary stable curve, easily reproving the results of Yamada for nodal curves with one node. Concrete examples are included. We also apply the same technique to give an alternative proof of the sufficiency part of the theorem by Bainbridge, Chen, Gendron, Grushevsky and M\"oller on the closures of strata of differentials with prescribed multiplicities of zeroes and poles.	math.AG	math	GENERAL VARIATIONAL FORMULAS FOR ABELIAN DIFFERENTIALS XUNTAO HU AND CHAYA NORTON Abstract. We use the jump problem technique developed in a recent paper [GKN17] to compute the variational formula of any stable differential and its periods to arbitrary precision in plumb- ing coordinates. In particular, we give the explicit variational for- mula for the degeneration of the period matrix, easily reproving the results of Yamada [Yam80] for nodal curves with one node and extending them to an arbitrary stable curve. Concrete examples are included. We also apply the same technique to give an alternative proof of the sufficiency part of the theorem in [BCGGM16] on the closures of strata of differentials with prescribed multiplicities of zeroes and poles. 1. Introduction We work over the field of complex numbers C. Let Mg be the mod- uli space of curves, and Mg be its Deligne-Mumford compactification. Given a stable nodal curve C with n nodes, the standard plumbing construction cuts out neighborhoods of size pse on the normaliza- tion of C at the two pre-images qe and q−e of each node qe of C, and identifies their boundaries (called seams, denoted by γ±e) via a gluing map Ie sending ze to z−e := se/z−1 , where se ≪ 1 is called the plumb- ing parameter and ze and z−e are chosen local coordinates near qe and q−e respectively. The irreducible components of the nodal curve C are denoted by Cv. The plumbing construction thus constructs a family of curves C → ∆ with the central fiber identified with C, where ∆ is the small polydisc neighborhood of 0 ∈ Cn with coordinates given by the plumbing param- eters s := (s1, . . . , sn). Depending on circumstances s are also called the plumbing coordinates, as they give versal deformation coordinates on Mg to the boundary stratum containing the point C. The Riemann surface resulting from plumbing with parameters s is denoted by Cs. e 1.1. Motivations. We are interested in studying the degeneration of the periods of an Abelian differential in plumbing coordinates, which 1 2 XUNTAO HU AND CHAYA NORTON can be seen as a direct application of the study of the behavior of degenerating families of Abelian differentials in plumbing coordinates. Our interest in the degeneration of the periods is motivated by the following two sources: the degeneration of period matrices near the boundary of the Schottky locus in Ag; and the degeneration of the pe- riod coordinates (both absolute and relative periods) of strata in the Hodge bundle (also known as the moduli spaces of Abelian differen- tials with fixed multiplicities at marked points). In this paper we will only study the degeneration of period matrices. Application to period coordinates near the boundary of the strata will be treated elsewhere. 1.2. Results and structure of the paper. In this paper we give a complete answer to the question stated above. In order to state our results, we introduce some more notation. The Hodge bundle ΩMg is a rank g vector bundle over Mg. A point in the Hodge bundle corresponds to a pair (C, Ω) where Ω is an Abelian differential on C. One can extend the Hodge bundle over Mg. The fiber of the extension ΩMg over a nodal curve C in the boundary of Mg parametrizes stable differentials, that is, meromorphic differentials that have at worst simple poles at the nodes with opposite residues. The goal in our paper is to compute the variational formula for any stable differential and its period over any 1-cycle on C in plumbing coordinates. The term "variational formula" in our paper means an ex- pansion in terms of both se and lnse. Note that a variational formula in this sense is not synonymous to a power series expansion in plumbing coordinates s. We will use specifically the term "s-expansion" when we mean the latter, where no logarithmic terms are involved. Moreover, throughout the paper objects subscripted by "e" are indexed by the set of edges of the dual graph of the stable curve C, and those by "v" are indexed by the set of vertices of the dual graph. The technique we use to construct the degenerating family of Abelian differentials Ωs along the plumbing family Cs is called (solving) the jump problem, which will be properly defined in Section 2. The main idea is that given a stable differential Ω on C, we have the mis-matches {Ωγe − I∗e (Ωγ−e)} (which we call the jumps of Ω) on the seams γ±e at opposite sides of each node qe. The solution to the jump problem is a "correction" differential η that matches the jumps of Ω with opposite sign. By adding η to Ω on each irreducible component, one obtains new differentials with zero jumps, that can thus be glued to get a global holomorphic differential Ωs on Cs. The jump problem is a special version of the classical Dirichlet prob- lem. It was developed and used in the real-analytic setting in a recent GENERAL VARIATIONAL FORMULAS FOR ABELIAN DIFFERENTIALS 3 paper by Grushevsky, Krichever and the second author [GKN17]. In the classical approach, the Cauchy kernel on the plumbed surface is used, while in [GKN17] the fixed Cauchy kernels on the irreducible components of the nodal curve are used, which is crucial to obtain an L2-bound of the solution to the jump problem in plumbing parameters. Our construction of the solution to the jump problem largely follows the method in that paper. Instead of the real normalization condition used in [GKN17], we normalized the solution by requiring that it has vanishing A-periods, where A is a set of generators of a chosen La- grangian subgroup of H1(Cs, Z) containing the classes of the seams. This normalization condition allows us to work in the holomorphic set- ting (as opposed to the real-analytic setting in [GKN17]) where we can use Cauchy's integral formula. As a consequence, the construction of the solution is simpler. In Section 3 of our paper, the solution η to the jump problem is con- structed explicitly as an iterated series (see (3.3) and (3.9)). We show that such a solution depends analytically on the plumbing parameters and moreover, we bound the L2-norm of η by a power of the L∞-norm of s. For convenience we denote hol(Ω) the regular part in the Laurent expansion of the function Ω(ze) given in the local coordinate ze that was dze used to define plumbing near the node qe. Theorem 1.1. (=Theorem 3.3) Let (C, Ω) ∈ ∂ΩMg be a stable differ- ential. For each v, let Ωv be the restriction of Ω on Cv. There exists a unique solution {ηv = ηv,s} with vanishing A-periods to the jump problem for Ω, such that for any s small enough and all se > 0, {Ωv,s := Ωv + ηv} defines a holomorphic differential Ωs on Cs satis- fying Ωv = lims→0 ΩsCv uniformly on compact sets of Cv \ ∪e∈Evqe. Furthermore, we have ηv,sL2 = O(ps). The solution to the jump problem ηv is constructed explicitly in (3.3) and (3.9). In order to highlight the series construction, we compute the leading term of the s-expansion for η(k) v , which in particular gives the linear term of the s-expansion for Ωs. Let lk v = (e1, . . . , ek) be a path of length k starting at a given vertex v = v(e1) in the dual graph ΓC. Let ωv(z, w) be the fundamental normalized bidifferential (see section 3.1 for details) on Cv and βe,e′ := hol(ωv)(qe, qe′). The s-expansion of η(k) is given as follows (Proposition 3.4): v η(k) v (z) = (−1)kXlk v kYi=1 sei·ωv(z, qe1) k−1Yj=1 β−ej ,ej+1 hol(Ω)(q−ek)+O(sk+1), 4 XUNTAO HU AND CHAYA NORTON where z ∈ Cs, and the sum is over the set of paths of length k in the dual graph starting at vertex v. One can compare this formula with [Wol15, Proposition 7], where a complete expansion is given in Schiffer coordinates. We also want to point out that one can in principle compute for any k the explicit s-expansion of η(k) (and hence for Ωs) via our method. v We are grateful to Scott Wolpert for the following remark: the con- struction provides a local frame for the sheaf of Abelian differentials near a nodal curve. This implies that the push forward of the relative dualizing sheaf is locally free of the expected rank. The locally-freeness of the first and second powers was first shown in [Ma76] and then in [HK13]. Our method can be further generalized to give a local frame of k-differential for any positive k near a boundary point in the moduli space, which will not be treated in this paper. In Section 4, we compute the leading terms in the variational formula for the period of Ωs over any smooth cycle in Cs. Let the residue of Ω at qe be denoted re, so that re = −r−e. Let α be any oriented loop in C not contained completely in any irreducible component Cvi. Let {qe0, . . . , qeN−1} be the ordered collection of nodes that α passes through (with possible repetition), so that q−ei−1 and qei lie on the same component Cvi, and let qeN = qe0. Let αs be a perturbation of α such that its restrictions on each Cv minus the caps at each node glue correctly to give a loop on Cs. Theorem 1.2. (=Theorem 4.1) The variational formula of the period of Ωs over αs is given by: αs Ωs = NXi=1 (rei lnsei + ci + li) + O(s2), where ci and li are the constant and linear terms in s respectively, explicitly given as ei (√sei) (rei−1 lnsei−1 + rei lnsei)! , 1 2 Ωvi − ci = lim s→0 z−1 li := − Xe∈Evi z−1 −ei−1 (√sei−1) se hol(Ω)(q−e) · σe, GENERAL VARIATIONAL FORMULAS FOR ABELIAN DIFFERENTIALS 5 where Evi denotes the set of nodes on the component Cvi, and ωv(ei)(zei, qei) + 1 sei! ωv(ei)(z−ei−1, q−ei−1) − 1 σe := ei (sei ) lims→0 z−1 ´q−ei−1 lims→0 z−1 −ei−1 ´ qei (sei ) ωv(ei)(z, qe) ´ qei q−ei−1  if e = ei; sei if e = −ei−1; Ωs −PN otherwise. This theorem in particular shows that ´αs i=1 rei lnsei is a holomorphic function in s. Using the theorem we can compute the variational formula for the degeneration of the period matrices near an arbitrary stable curve. We choose a suitable symplectic basis {Ak,s, Bk,s}g k=1 of H1(Cs, Z) such that the first m A-cycle classes gen- erates the span of the homology classes of the seams. We take a nor- malized basis {v1, . . . , vg} of H 1(C, KC), such that after applying the jump problem construction, the resulting set of differentials {vk,s}g k=1 is a normalized basis of H 1(Cs, KCs) (see Section 2 and 4 for a proper definition). By taking Ω = vk and αs = Bh,s, the following corollary gives the leading terms in the variational formula of the period matrix. Corollary 1.3. (=Corollary 4.6) For any e and k, denote Ne,k := γe × Bk,s the intersection product. For any fixed h, k, the expansion of τh,k(s) = ´Bh,s vk,s is given by τh,k(s) =Xe∈EC · lnse ei (√sei) Ne,h · Ne,k 2 NXi=1 z−1 z−1 −ei−1 (√sei−1) + lim s→0 − Xe∈EC vk − Nei,hNei,k lnsei! se (hol(vk)(qe) hol(vh)(q−e)) + O(s2), where EC is the set of oriented edges of the dual graph of C, {qei}N−1 i=0 is the set of nodes Bh passes through. Furthermore, under our choice of symplectic basis, Ne,h · Ne,k is equal to 1 if h = k and the node qe lies on Bh and equals 0 otherwise. In the paper [Tan91] Taniguchi computed the logarithmic term Ne,hNe,k 2 lnse Xe∈EC 6 XUNTAO HU AND CHAYA NORTON in the variational formula of τh,k(s). In addition to his result, we get the constant and the linear term as above, and in principle one can also expand the error term to any order. The variational formula of the period matrix is useful in various places. For instance it is used by the first author in [Hu] to compute the degeneration of the theta constants near the boundary of A3, the (compactified) moduli space of p.p.a.v of dimension three, and furthermore via a modular form method to compute the boundary behavior of the strata H(4) of the Hodge bundle where the Abelian differentials have a quadruple zero. When restricted to the case where C has only one node, Theorem 1.1 and Corollary 1.3 imply the results in [Yam80]. The computation reproving [Yam80] is done in Section 5. In order to demonstrate the greater applicability of our method compared to [Yam80], in Section 5 we study two other examples which, to the knowledge of the authors, have not been treated in the literature previously. The first one is the so-called banana curve, that is, curves with two irreducible components meeting at two nodes. The second example is the totally degenerate curve, namely a P1 with 2g marked points glued in pairs. The last example appears to be important in the study of Teichmuller curves. All of the above are only applicable to the case where Ω is a stable differential. In Section 6 we apply the jump problem technique to give a new method to smoothen a differential with higher order zeroes and poles at the nodes. The existing terminology of such a smoothing pro- cedure is "higher order plumbing", as opposed to standard plumbing. The higher order plumbing is a crucial ingredient used by Bainbridge, Chen, Gendron, Grushevsky and Moller in the paper [BCGGM16] to construct the incidence variety compactification (IVC) of strata of dif- ferentials with prescribed zeroes and poles. In that paper, they prove that the necessary and sufficient conditions for a stable differentials (C, Ω) to lie in the boundary of the IVC is the existence of a twisted differential Ξ and a level function l on the vertices of the dual graph of C, with certain compatibility conditions. The obviously harder di- rection is the sufficiency of the conditions, which essentially requires a construction of a degenerating family of Abelian differentials in the strata to the limit differential (C, Ω) with the compatible data (Ξ, l). In [BCGGM16], the authors give two such constructions by using a plumbing argument and a flat geometry argument respectively. Both arguments only give rise to one-parameter degenerating families. In Section 6, we briefly revisit their main results, and construct via the jump problem approach a degenerating family that is different from the two mentioned above (Theorem 6.4). The number of parameters GENERAL VARIATIONAL FORMULAS FOR ABELIAN DIFFERENTIALS 7 in our degenerating family is equal to the number of levels in the level graph ΓC minus 1. In particular, we reprove the sufficiency of the conditions for a stable differential to lie in the boundary of the IVC. 1.3. Prior work. The history of studying the degeneration of period matrices using plumbing coordinates traces back to Fay [Fay73] and Yamada [Yam80]. They only study the case n = 1, that is, when C has only one node. In this case the authors give complete power series expansion in s (the plumbing parameter at the only node) of Abelian differentials and describe the degeneration of period matrices by a variational formula in terms of s and ln s. For the case where the stable curve C has multiple nodes, Taniguchi [Tan91] discusses the degeneration of the period matrix and computes the logarithmic term in the variational formula. However, the complete variational formula of neither the Abelian differentials nor the period matrix is derived in his papers. A recent paper of Wolpert [Wol15] gives a complete expansion of Abelian differentials and the second order expansion for the period matrix in terms of Schiffer deformations which are deformation of smooth Riemann surfaces. In [LZR13], Liu, Zhao, and Rao study holomorphic one-forms and the period matrix under small deformations of a smooth Riemann surfaces to give a complete variational formula using the Kuranishi coordinates on the Teichmuller space. 2. Smoothing Riemann Surfaces Let C be a stable nodal curve over the complex numbers. In this section we recall the plumbing construction and fix the notation. Definition 2.1. (Dual graph) The dual graph ΓC of a stable curve C is a graph where each unoriented edge corresponds to a node of C, and each vertex v corresponds to the normalization of an irreducible component Cv of C. The edge connecting vertices corresponds to the node between components. For future convenience we write EC for the set of oriented edges e of ΓC. We will use −e to denote the same edge as e but with the opposite orientation, and e = −e the corresponding unoriented edge. Namely, q±e are the pre-images of the node qe in the normalization of C. We write v(e) to denote the source of the oriented edge e, and write Ev for the set of edges e such that v = v(e), that is, edges pointing out of the vertex v. We denote EC = {e}e∈EC the set of unoriented edges. The cardinality of EC is half of the cardinality of EC. 8 XUNTAO HU AND CHAYA NORTON 2.1. Plumbing construction. We now recall the local smoothing procedure of a nodal curve C via plumbing. There are many equiv- alent versions of the local plumbing procedure, we follow the one used in [Yam80], [Wol13] and [GKN17]. Definition 2.2. (Standard plumbing) Let qe, q−e be the two preimages of the node qe in the normalization of C. Let z±e be fixed chosen local coordinates near q±e. Take a sufficiently small se = s−e ∈ C, we denote U±e = U se call the seams. We orient each seam γe counter-clockwise with respect to Ue. The standard plumbing Cse of C is ±e := {z±e <pse} ⊂ C, and denote γ±e := ∂U±e, which we Cse := [C\Ue ⊔ U−e]/(γe ∼ γ−e), where γe ∼ γ−e is identified via the diffeomorphism Ie : γe → γ−e send- ing ze to z−e = se/ze. We call the identified seam γe. The holomorphic structure of Cse is inherited from C\U e ⊔ U−e. Notation 2.3. (1) Since se = s−e, we can use the notation se and denote s := {se}EC . In later parts of the paper we will continue to use se (instead of se) for simplicity. (2) We write Cs for the global smoothing of C by plumbing every node qe with plumbing parameter se, so that C = C0. Let bCv := Cv\ ⊔e∈Ev Ue, then bCv has boundaries γ := {γe}e∈Ev. We use eCv to denote the interior of bCv, and bCs to denote the disjoint union of bCv for all v. We have Cs = bCs/{γe ∼ γ−e}EC . (3) Throughout this paper, in a specified component Cv, the sub- scripted ze is used to denote the chosen local coordinate near the node qe for every e ∈ Ev for the standard plumbing. The non-subscripted notation z is used to denote an arbitary local coordinate of any point in eCv. (4) For future convenience, we denote s := maxe∈EC se. Remark 2.4. Let u = (u1, . . . , uk) be some coordinates along the boundary stratum of Mg that C lies in. One can think of the boundary stratum as a Cartesian product of moduli of curves with marked points, and u is the combination of some coordinates chosen on each moduli space. It is a standard result in Teichmuller theory (see [IT], and [HK13]) that the set of plumbing parameters s together with u give local coordinates on Mg near C. We denote Cu,s a nearby curve of C, then C = Cu0,0 for some u0. Our results depend smoothly on u throughout the paper, and all the bounds we derive in this paper hold for u varying in a small neighborhood of u0, therefore we fix the coordinate u0 for C and consistently write Cs = Cu0,s. GENERAL VARIATIONAL FORMULAS FOR ABELIAN DIFFERENTIALS 9 2.2. Conditions on residues. Our goal is to express the variational formulas for Abelian differentials with at worst simple poles on C in plumbing coordinates. Take a stable differential Ω on the stable curve C in the boundary of the Deligne-Mumford compactification Mg. Denote Ωv to be the restriction of Ω to the irreducible component Cv. We require Ωv to be a meromorphic differential whose only singularities are simple poles at the nodes of Cv. We denote the residue of Ωv at qe to be re (possibly zero). We have re = −r−e for any e ∈ EC by the definition of the extended Hodge bundle ΩMg (see e.g., [HM]). 2.3. Jump problem. Given a stable differential Ω on the boundary point C = C0, we want to construct a holomorphic differential Ωs on the nearby curve Cs by an analytic procedure called solving the jump problem. Definition 2.5. (Jump problem) The initial data of the jump problem is a collection φ of complex-valued continuous 1-forms {φe} supported on the seams γ of bCs, satisfying the conditions φe = 0, (2.1) φe = −I∗e (φ−e), γe ∀e ∈ EC. We call the set {φe}e∈EC jumps. A solution to the jump problem is a continuous on the boundaries γ, satisfying the condition holomorphic differential ηs on bCs such that it is holomorphic on eCs and ηsγe − I∗e (ηsγ−e) = −φe ∀e ∈ EC. Note that by letting {φe}EC be the mis-matches {Ωv(e)γe−I∗e (Ωv(−e)γ−e)}EC of Ω, one can check that they satisfy (2.1). Therefore (Ω + ηs)Cv(e) and (Ω+ηs)Cv(−e) have no jump along γe at every node e, where ηs is the so- lution to the jump problem with jumps the mis-matches of Ω. We can thus glue them along each seam to obtain a required global differential Ωs on Cs. Notation 2.6. For simplicity, we drop the subscript s in ηs through- out the paper. But it is important to bear in mind that the solution depends on s as the size of the seams varies with s. 2.4. A-normalization. Note that the solution to the jump problem is never unique: adding any differential on Cs gives another solution. We need to impose a normalizing condition to ensure the uniqueness of the solution. On each irreducible component Cv of the nodal curve C we choose and fix a Lagrangian subspace of H1(Cv, Z), and we also choose and 10 XUNTAO HU AND CHAYA NORTON fix a basis of the subspace. In Definition 2.2, the plumbed surface bCs is seen to be a subset of C. Since the seams (as boundaries of bCs) are contractible on each Cv (without boundaries), we know that the classes of the seams {[γe]}EC together with the union of the basis of the Lagrangian subspaces on the irreducible components span a La- grangian subspace of H1(Cs, Z). We can fix this Lagrangian subspace of H1(Cs, Z) along the plumbing family {Cs}. If some γe is homol- ogous to zero on Cs, as se approaches 0 the Lagrangian subspace of H1(Cs, Z) is invariant; if the class of γe is non-zero, then the rank of the Lagrangian subspace drops by 1 as the corresponding element in the basis goes to zero. We denote this choice of basis as {A1,s, . . . , Ag,s} where the first m cycles A1,s, . . . , Am,s generate the subspace spanned by the seams {[γe]}EC . This choice of indexing will be used later in the computa- tion of the period matrices in Section 4.2. Definition 2.7. A solution to the jump problem is A-normalized if it has vanishing periods over A1,s, . . . , Ag,s. Note that this definition only depends on the choice of Lagrangian subspace of H1(Cs, Z). In particular by our choice of Lagrangian sub- space, an A-normalized solution η must have vanishing periods over It is a standard fact (see e.g., [GH]) that any holomorphic A-normalized differential is identically zero on a compact Riemann surface without η = 0. the seams: ´γe boundaries. Given two A-normalized solutions η and η′ on bCs which are both holomorphic by definition, the differential η − η′ has zero jumps on the seams and thus defines a global holomorphic A-normalized dif- ferential on Cs, which is therefore identically zero. This shows the uniqueness of an A-normalized solution. 3. Variational Formulas for Stable Differentials In this section we construct the degenerating family Ωs in a plumbing family Cs, and give the variational formula for Ωs in terms of s. As introduced in Section 1.2, we plan to construct the solution to the jump problem that matches the jumps of Ω0 = Ω. In the classical construction, such differentials are obtained by integrating the jumps against the Cauchy kernel (see the following section) on the whole Cs. In this approach the Cauchy kernel depends on s, and this dependence is implicit and hard to determine. Alternatively, following [GKN17], we fix the Cauchy kernels on each irreducible components of the normalization of the limit stable curve C. GENERAL VARIATIONAL FORMULAS FOR ABELIAN DIFFERENTIALS 11 On each component Cv we integrate the jumps {Ωv(e)γe−I∗e (Ωv(−e)γ−e)}e∈Ev against the Cauchy kernel. In this way we obtain a differential in the classical sense on each component Cv which has jumps across the seams. We then restrict it to bCv, the component minus the "caps". In this way the original jumps are compensated by the newly-constructed differen- tials, but these differentials in turn produce new jumps. However, since the L∞-norms of the newly-constructed differentials along the seams γe in local coordinates ze are controlled in an explicit way by s, the new jumps are also controlled by s. By iterating the process one obtain a sequence of differentials, each term controlled by a higher power of s. This sequence converges to the desired solution to the jump problem. 3.1. The Cauchy kernels. The construction of the A-normalized so- lution to the jump problem is parallel to the construction of the almost real-normalized solution in [GKN17], which uses a different normaliz- ing condition, and the solution differential obtained there allows one to control the reality of periods. Given a smooth Riemann surface C′, the Cauchy kernel is the unique object on C′ × C′, satisfying the following properties: (1) It is a meromorphic differential of the second kind in p whose (2) It is an A-normalized differential in p: ´p∈Ai only simple poles are at p = q and p = q0 with residue ± 1 2πi ; i = 1, . . . , g and ∀q ∈ C′. KC′(p, q) = 0, for The Cauchy kernel can be viewed as a multi-valued meromorphic function in q whose only simple pole is at p = q. Let {Ai, Bi} be a symplectic basis of H1(C′, Z), and let {vi} to be the basis of holomor- phic 1-form dual to the A-cycles. The multi-valuedness is precisely as follows (where q + γ denotes the value of the kernel at q upon extension around the loop γ): K(p, q + Ai) = K(p, q); K(p, q + Bi) = K(p, q) + vi(p). Note that the Cauchy kernel is a section of a line bundle on C′ × C′ satisfying the first two normalization conditions above, and therefore it can be written in terms of theta functions and the Abel-Jacobi map (for a reference of the theta function see [Gun76]). We also remark that KC′ depends on the choice of the Lagrangian subspace spanned by the A cycles. For completeness below we include the explicit expression for the Cauchy kernel in terms of theta functions: KC′(p, q) := 1 2πi ∂ ∂p ln θ(A(p) − A(q) − Z) θ(A(p) − Z) , 12 XUNTAO HU AND CHAYA NORTON where θ denotes the theta function on the Jacobian of C′, Z denotes a general point of the Jacobian, and A denotes the Abel-Jacobi map with some base point q0 ∈ Z. The expression does not depend on the choice of Z. We call ωC′(p, q) := 2πidqKC′(p, q) the fundamental normalized bidif- ferential of the second kind on C′ (also known as the Bergman kernel). Note that the term "normalized" here means A-normalization. Namely, p∈Ai ωC′(p, q) = 0, i = 1, . . . , g. The fundamental normalized bidifferential has its only pole of second order at p = q. It is uniquely determined by its normalization along A-cycles, symmetry in the entries, and the bi-residue coefficient along p = q. See notes [Ber06, Ch. 6] for a review of the Cauchy kernel and fundamental bidifferential. When C is a nodal curve with irreducible components Cv, we denote by Kv (resp. ωv) the Cauchy kernel (resp. bidifferential) on each Cv. Recall that ze (or we, when we need to distinguish between two distinct points in the same neighborhood) denotes the local coordinates in some neighborhood Ve of qe that contains U e. We define a local holomorphic differential Kv ∈ Ω1,0(⊔e∈EvVe × ⊔e∈EvVe), by taking the regular part of Kv: Kv(ze, we′) :=(Kv(ze, we′) Kv(ze, we) − if e 6= e′, if e = e′. dze 2πi(ze−we) Define ωv(ze, we′) = 2πidwe′ Kv(ze, we′), then similarly we have ωv(ze, we′) =(ωv(ze, we′) ωv(ze, we) − dzedwe (ze−we)2 if e 6= e′, if e = e′. For future convenience we fix the notation for the coefficients in the expansion of ωv(ze, we′): Clearly we have βv the superscript v and write simply βe,e′ instead. e,e′ = ωv(qe, qe′). When the context is clear, we drop 3.2. Approach to solving the jump problem. In this section we approach the jump problem directly in order to clarify the appearance of a series expression for local differentials (3.3) below. (3.1) ωv(ze, we′) =: dwe′dze βv e,e′ + Xi,j≥0,i+j>0 βv i,jzi ewj e′! . GENERAL VARIATIONAL FORMULAS FOR ABELIAN DIFFERENTIALS 13 We first look at the simplest example: g = 0. On P1 the Cauchy kernel is simply K(z, w) = dz w−z . Let γ be a Jordan curve bounding a region R on P1. Cauchy's integral formula implies that integrating K(z, w) against a differential f (w)dw holomorphic inside the region R along γ (negatively oriented with respect to R) vanishes when z is in the exterior of R; it is equal to f (z)dz when z ∈ R. In other words integrating f (w)dw against the Cauchy kernel defines a differential on P1 \ γ whose jump along γ is precise f (z)dz. When replicating this idea on Riemann surfaces of higher genus, integrating a differential which is holomorphic inside a contractible loop γ against the Cauchy kernel produces a holomorphic differential on that Riemann surface whose jump across γ is given by the differential. Below z+ is an point outside γ and z− is inside γ: lim z+→z′∈γγ K(z, w)f (w)dw − lim z−→z′∈γγ K(z, w)f (w)dw = f (z)dz. This follows directly from Cauchy's integral formula. In [GKN17], or in general when integrating against a jump which is not meromorphic, obtaining a result such as above would require the Sokhotski-Plemelj formula (for reference see [Ro88]). We would like to explicitly analyze the dependence of the solution to the jump problem on the plumbing parameters. Solving the jump problem on Cs by integrating against the Cauchy kernel on Cs, as described above is classical [Ro88], but it does not allow one to study the dependence on plumbing parameters. Therefore the approach we take, which was introduced in [GKN17], is integration against fixed Cauchy kernels defined individually on each irreducible component of the nodal curve, which are thus independent of s. The result will give an explicit expansion of the solution in s, and constructing this solution is more involved. The procedure of the construction of the solution is clarified below. Step 1. We denote the holomorphic part of the differential Ωv(e)(ze) as ξ(0) e (ze) := Ωv(ze) − redze . It follows from the residue condition that the singular parts of the differentials on the opposite sides of the node cancel. Thus the jumps can be written as follows: ze {Ωv(e)γe − I∗e (Ωv(−e)γ−e)}e∈Ev = {ξ(0) e γe − I∗e (ξ(0) −eγ−e)}e∈Ev . 14 XUNTAO HU AND CHAYA NORTON Step 2. We integrate the jumps against the Cauchy kernel. This integration defines a differential on the open Riemann surface Cv, η(1) v (z) :=Xe∈Ev ze∈γe = Xe∈Ev ze∈γe Kv(z, ze)(ξ(0) e γe − I∗e ξ(0) −eγ−e)(ze) Kv(z, ze)I∗e ξ(0) −eγ−e(ze) where the equality follows from Cauchy's integral formula. We also extend η(1) v (z) continuously to the boundary of the plumbing neighbor- hood. We have an important remark here: The differential η(1) v (z) can be seen as our first attempt at solving the jump problem, but it does not give the solution of the desired jump problem. There is a new jump between Ωv(e) + η(1) v(−e) on each node. The "error" comes from the holomorphic part of the Cauchy kernel. Step 3. We look at this "error" explicitly. Locally near the seam v(e) and Ωv(−e) + η(1) γe0, the differential η(1) expression: v (ze0) for pse0 < ze0 < 1 has the following Xe∈Ev ze∈γe Kv(ze0, ze)I∗e ξ(0) −e (ze) =Xe∈Ev ze∈γe 2πi we0∈γe0 1 Kv(ze0, ze)I∗e ξ(0) −e (ze)+ dze0 ze0 − we0 I∗e0ξ(0) −e0(we0). Where we recall that Kv is the holomorphic part of Kv, and the last part involves the integral of the singular part of the Cauchy kernel. The last integral can be evaluated by Cauchy's integral formula by noting that sez−1 (3.2) 1 dzewe∈γe 2πi < √se where we point out that I∗ is orientation reversing, w−eξ(0) −e (w−e) w−e − sez−1 ze w−e∈γ−e I∗e ξ(0) −e (we) ze − we = −se 1 2πi dze z2 e = − dze e e ξ(0) −e ( se ze ) = I∗ξ(0) −e (ze) . Therefore we have the following: v(e)γe − I∗e η(1) {η(1) v(−e)γ−e}e∈Ev = − {Ωv(e)γe − I∗e Ωv(−e)γ−e}e∈Ev +(cid:16) Xe′∈Ev ze′∈γe′ − I∗ Xe′∈Ev(−e) ze′∈γe′ Kv(e)(ze, ze′)I∗e′ξ(0) −e′(ze′) Kv(−e)(z−e, ze′)I∗e′ξ(0) −e′(ze′)(cid:17)e∈Ev . GENERAL VARIATIONAL FORMULAS FOR ABELIAN DIFFERENTIALS 15 Thus we see that η(1) exactly the "error". Kv(ze, ze′)I∗e′ξ(0) holomorphic differentialsPe′∈Ev ´ze′∈γe′ v (z) has the desired jump plus the jump of (local) −e′(ze′), which is Step 4. We give an estimate on the size of the "error". We show in Lemma 3.2 below that the L∞-norm of the "error" is controlled by the L∞-norm of the plumbing parameters s. We therefore apply the jump problem again to further reduce the gap. Finally, our approach to solving the jump problem is by integrating against a series constructed from the recursively appearing jump problems. We prove (in Lemma 3.2) that the "errors" produced in the k-th step of the recursion is con- trolled by the k-th power of s, and we use this to show the convergence of the desired solution of the jump problem. e (ze) in each step, and then we prove Lemma 3.2 which bounds each by a power of the plumbing parameters, thus the series defined by adding the "errors" converges. And at last we prove that the solution to the jump problem, denoted ηv, is the result of integrating this series against the Cauchy kernel on each irreducible component. In the following section we first define the "errors" ξ(k) 3.3. Construction of the A-normalized solution to the jump problem. We construct the A-normalized solution to the jump prob- lem as suggested by the computation above, namely we define local holomorphic differentials, which can be understood as the recursively appearing jumps, and show the series converges. The resulting local differentials are such that when integrated against the Cauchy kernel on each irreducible component of the nodal curve, the jump is given by the first term in the series. Let Ω be a stable differential on the stable curve C. We can define re- cursively the following collection of holomorphic differentials described locally in the neighborhood of each node : (3.3) k = 0 : ξ(0) k > 0 : ξ(k) redze ; e (ze) := Ωv(ze) − we′∈γe′ e (ze) := Xe′∈Ev ze Kv(ze, we′) · I∗e′ξ(k−1) −e′ (we′). Note ξ(k) e for k > 0 depend on s as γe and I∗ depend on s. We suppress this in the notation. Let ΓC be the dual graph of C. Let lk := (e1 . . . , ek) be an oriented path of length k in the dual graph, starting from the vertex v = v(e1). We denote Lk v the collection of all such paths starting from the vertex v. We remark that given the Cauchy kernels on each component, the 16 XUNTAO HU AND CHAYA NORTON e1 (ze1) is determined by the collection of local differentials v where ek is the ending edge of lk ∈ Lk v. e (ze). Since Kv(ze, we′) is holomorphic differential ξ(k) {ξ(0) −ek(w−ek)}Lk We define ξe(ze) :=P∞k=0 ξ(k) in the first variable, we have ´γe ξ(k) e = 0 for any e and k, therefore (3.4) γe ξe = 0, ∀e ∈ EC. The convergence of this series is ensured by the following lemma, whose proof follows very much along the lines of [GKN17]. We include the proof here for completeness. The essential ingredient of the proof is the fact that our Cauchy kernels and the bidifferentials are defined on the irreducible components, thus they are independent of the plumbing parameters s. Notation 3.1. (1) Throughout the paper we use the tilde notation to denote the function corresponding to a given differential in a given local coordinate chart, for instance K(z, w) =: eK(z, w)dz, ω(z, w) =: eω(z, w)dzdw, and also ξ(k) e (ze) := ξ(k) (2) To simplify notation, we denote e (ze)dze. ξe := ξ(0) e (qe) (3.5) at every node qe. uated in the second variable at any node qe, by an abuse of (3) When the functioneωv(z, w) of the bidifferential ωv(z, w) is eval- notation, we write ωv(z, qe) =eωv(z, qe)dz for z ∈ eCv. (4) Recall that s := maxe∈EC se. For future convenience, for any collection of holomorphic functions on the unit disks neigh- borhood at each node f := {fe ∈ O(Ve)}e∈EC , we define the following L∞-norms: fes := sup ze∈γe fe(ze); fs := max e∈EC fes. Moreover by the Schwarz lemma on Ue = {ze < pse} we have that fs ≤ f1ps Lemma 3.2. For sufficiently small s, there exists a constant M1 inde- pendent of s, such that the following estimate holds: , where ord f := mine∈EC ordqe fe. ord f (3.6) (k) ξ s ≤ (sM1)k ξ (0) s. In particular, the local differential ξe(ze) is a well-defined holomorphic differential at each node e ∈ EC. GENERAL VARIATIONAL FORMULAS FOR ABELIAN DIFFERENTIALS 17 Proof. For all v and all e, e′ ∈ Ev, the Cauchy kernel eKv(ze, we′) is analytic in both variables and independent of s. Thus there exists a uniform constant M2 (independent of s) such that for any ze ∈ Ve, . < w2 e′ max (3.7) By (3.4), we have This in turn implies eKv(ze, we′) − eKv(ze, 0) < M2we′. (cid:12)(cid:12)(cid:12)(cid:12)(cid:12) we′∈γe′(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)eK(ze, we′) − eK(ze, 0) M2pse′ (we′)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) =(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) (cid:12)(cid:12)(cid:12)(cid:12)(cid:12) we′∈γe′heKv(ze, we′) − eKv(ze, 0)i I∗e′ξ(k−1) we′∈γe′ eKv(ze, we′)I∗e′ξ(k−1) (cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)I∗e′ ξ(k−1) = se′we′∈γe′(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)eK(ze, we′) − eK(ze, 0) (we′)(cid:12)(cid:12)(cid:12) dwe′ < se′M2 · 2π ξ(k−1) The second equality is the result of pulling back dw−e′. Note that the last inequality is due to the fact that for we′ ∈ γe′, we have I∗e′ ξ(k−1) s, therefore the integration over −e′ γe′ gives a pse′ that cancels the one in (3.7). By definition of ξ(k) , there exists a constant M1 independent of s and k such that, (we′)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) (we′) = ξ(k−1) −e′ ) ≤ ξ(k−1) −e′ ( se′ we′ w2 e′ −e′ −e′ −e′ −e′ s e ξ(k) e s ≤ sM1 max e′∈Ev(e) ξ(k−1) −e′ s < sM1 ξ (k−1) s. (k) Note that the RHS is independent of e and v, we can thus pass to the s. maximum over e ∈ EC of the LHS and obtain ξ By induction, we have the desired estimate (3.6). When s < 2M−1 to a limit bounded by(cid:16)1 + sM1 s converges 1. We therefore conclude that the local differential ξe(ze) is analytic in s. (cid:3) 1 , the geometric series ξs :=P∞k=0 ξ s < 2ps 1−sM1(cid:17) ξ s < sM1 ξ s < 2 ξ (k−1) ξ ord ξ (k) (0) (0) (0) (0) We now construct the solution to the jump problem with initial data (3.8) ηv(z) = Xe∈Ev {Ωv(e)γe − I∗e (Ωv(−e)γ−e)}. We define the following differential on bCv: where z ∈ eCv. By extending continuously to the seams, the differential ηv is defined on bCv. Kv(z, ze)I∗e ξ−e(ze). ze∈γe I∗e ξ−e(ze) = 0 18 XUNTAO HU AND CHAYA NORTON η(k) (3.9) Recall ξe(ze) :=P∞k=0 ξ(k) v (z) := Xe∈Ev In this notation we have ηv :=P∞k=1 η(k) ze∈γe v . e (ze). For future use we denote, Kv(z, ze) · I∗e ξ(k−1) −e (ze). We claim the differentials ηv(z) are single-valued. This follows from noticing the multi-valuedness of K(z, ze) along Bi depends exclusively on z, and thus any multi-valuedness is canceled after integration against I∗e ξ−e by (3.4). ze∈γe Kv(z, ze))·I∗e ξ−e(ze) = vi(z)ze∈γe Kv(z+Bi, ze))·I∗e ξ−e(ze)−ze∈γe Although the Cauchy kernel Kv has a simple pole with residue (2πi)−1 at the base point q0, it follows from (3.4) that ηv(z) is holomorphic at q0 and hence defines a holomorphic differential on eCv. Let γq0 be a small loop around the point q0. Here we verify integrating ηv along γq0 is zero. The paths γq0 does not intersect any γe, and we could exchange the order of integration in z and ze. The integral of Kv(z, ze) along z ∈ γq0 is (2πi)−1 for any q. Thus by (3.4) integrating the result times I∗e ξ−e(ze) along γe is zero. We recall that the L2-norm of a holomorphic differential ω on a smooth Riemann surface C′ is given by ωL2 := i 2 ´C′ ω∧ω. Note that both ξe(ze) and ηv(z) implicitly depend on s. The following theorem establishes an L2 bound on ηv, and shows that it is the desired solution to the jump problem. Theorem 3.3. Let C be a stable nodal curve with irreducible compo- nents Cv, Ω a stable differential on C. Let Ωv be the restriction of Ω on Cv. For s small enough, {ηv} is the unique A-normalized solution to the jump problem with jump data Ωv(e)γe − I∗e (Ωv(−e)γ−e). More- over, there exists a constant M independent of v and s, such that the following L2-bound of the solution holds: 1+ord ξ (0) (0) M ξ ηvL2 <ps 1. (3.10) Therefore {Ωv,s := Ωv + ηv} defines a holomorphic differential when all se > 0, denoted Ωs on Cs, satisfying Ωv = lims→0 ΩsCv uniformly on compact sets of Cv \ ∪e∈Evqe. Proof. Step 1. We first show that the solutions ηv are A-normalized. Recall that our choice of the maximal Lagrangian subspace of H1(Cs, Z) contains the subspace generated by the classes of the seams γe. By the fact that {Kv(z, ze)}v,e are A-normalized in the first variable, the GENERAL VARIATIONAL FORMULAS FOR ABELIAN DIFFERENTIALS 19 η(k) v (z) is zero if the class [Ai,s] does not belong to the integral ´z∈Ai,s span of the seams {[γe]}. In order to compute the integrals of η(k) in the neighborhood {√se < ze < 1}. the local expression for η(k) Note that the Cauchy kernel Kv(ze, we′) is holomorphic if e 6= e′, and it has a singular part 2πi(ze−we) when both variables are in the neigh- borhood of the same nodes. Therefore we have v along the seams, we compute dz v (3.11) η(k) v (ze) = = dze dze 2πi we∈γe 2πi we∈γe =(cid:16)I∗e ξ(k−1) 1 ze − we 1 ze − we I∗e ξ(k−1) −e I∗e ξ(k−1) −e e (cid:17) (ze), −e + ξ(k) (we) + Xe′∈Ev we′∈γe′ (we) + ξ(k) e (ze) Kv(ze, we′)I∗e′ξ−e′(we′). the equality follows from Cauchy's integral formula, see (3.2) for details. Note that η(k) v admits continuous extension to the boundary γe. By this η(k) expression and property (3.4), we conclude that ´z∈γe v (z) = 0 and Step 2. We show that the differentials {Ωv,s} have zero jumps among hence the solution ηv is A-normalized. the seams γ = {γe}e∈EC . It is sufficient to prove ∞Xk=1(cid:16)η(k) −eγe(cid:17) (ze). (3.12) (cid:0)Ωv(e) − I∗e Ωv(−e)(cid:1)γe(ze) = − v(e) − I∗e η(k) v(−e)(cid:17)γe(ze). First we note that by the opposite residue condition (re = −r−e) the singular parts of Ωv(e) and I∗e Ωv(−e) cancels, therefore we have For k ≥ 1, by (3.11) the jumps along the identified seams for each v terms η(k) can be analyzed. e − I∗e ξ(0) (cid:0)Ωv(e) − I∗e Ωv(−e)(cid:1)γe(ze) =(cid:16)ξ(0) (cid:16)η(k) v(−e)(cid:17) (ze) =(cid:16)I∗e ξ(k−1) v − I∗e η(k) − I∗e ξ(k) e − ξ(k−1) −e + ξ(k) =(cid:16)ξ(k) −e(cid:17) (ze) −(cid:16)ξ(k−1) e − I∗e ξ(k) Therefore P∞k=1(cid:16)η(k) v(−e)(cid:17)γe(ze) = −(cid:16)ξ(0) v(e) − I∗e η(k) and we have shown (3.12). e e −e(cid:17) (ze) − I∗e ξ(k−1) e − I∗e ξ(0) −e (cid:17) (ze). −eγe(cid:17) (ze), Step 3. We want to prove the L2-bound (3.10) for the solution. We v (ze)/dze on the seams. By (3.11) take the L∞ norm of η(k) v (ze) := η(k) 20 XUNTAO HU AND CHAYA NORTON we have (3.13) η(k) v s := max ze=√seη(k) v (ze) ≤ I∗e ξ(k−1) −e (ze)s + ξ(k) e s. i z0 L2 = γe 2 Cv ηv ∧ ηv = Xe∈Ev By Lemma 3.2, we know that for any k ≥ 1, there exists a constant v s < (M′s)k−1 ξ(0)s. By the summing the series, we ηv. Then since M′ such that η(k) have ηvs < M′′ ξ(0)s for some constant M′′. Take any base point z0 ∈ Cv, define πv(z) := ´ z dπv = ηv, by Stokes theorem, we have πvηv < M′′ ξ(0)sXe∈Ev zeγe πv(ze)dze. ηv2 Since ηv is bounded on γe, by taking z0 ∈ γe the length of arc from z0 to ze ∈ γe is at most 2πps. Therefore we can bound πvs = πvs by 2πpsηvs = 2πM′′ps ξ(0)s. At last we have Thus by letting M := 2πM′′√maxv #Ev, since ξ we have the required L2-bound (3.10) for ηvL2. Note that for holomorphic differentials, convergence in L2 sense im- plies uniform convergence on compact sets. Therefore we conclude that Ωv = lims→0 ΩsCv uniformly on compact sets of Cv \ ∪e∈Evqe. Lastly, the holomorphicity of Ωv,s for s > 0 follows from the holo- morphicity of Ωv away from the nodes and the holomorphicity of ηv on Cv. Recall the Cauchy kernel is holomorphic in Cv except at q0, and we've verified that ηv does not have a pole at q0. (cid:3) L2 < s · (2πM′′ ξ(0)s)2 · #Ev. 1ps s ≤ ξ ηv2 ord ξ (0) (0) (0) v In fact, by construction (3.9) we can compute the s-expansion of each summand η(k) explicitly, and thus the s-expansion of the differential Ωs. For future applications and comparisons to earlier works, we compute the first term in the s-expansion for each summand. Proposition 3.4. Let lk = (e1, . . . , ek) ∈ Lk j=1 β−ej ,ej+1. Then the expansion of η(k) ΓC starting from a given vertex v = v(e1). Denote s(lk) = Qk and β(lk) =Qk−1 s(lk) · ωv(z, qe1)β(lk) ξ−ek + O(sk+1), where z ∈ bCv, βe,e′ is defined in (3.1), and ξe is defined in (3.5). v (z) = (−1)k Xlk∈Lk v be a path of length k in i=1 sei, is given by (3.14) η(k) v v GENERAL VARIATIONAL FORMULAS FOR ABELIAN DIFFERENTIALS 21 Proof. Fix a vertex v in ΓC. First for a fixed e ∈ Ev, we show the following expansion for ξ(k) e for k > 0: s(lk) · ωv(ze, qe1)β(lk) ξ−ek + O(sk+1). (3.15) ξ(k) e (ze) = (−1)k Xlk∈Lk v This is derived by induction. For k = 1, we have L1 v = Ev, and l1 = (e1) where e1 ∈ Ev. We compute (3.16) ξ(1) e (ze) = Xe1∈Ev = − Xe1∈Ev = − Xe1∈Ev −e1(we1) Kv(ze, we1)I∗e1ξ(0) we1∈γe1 we1∈γe1 ωv(ze, qe1) ξ−e1 + O(s2), Kv(ze, we1) se1 se1 e1 · ξ−e1dwe1 + O(s2 e1) w2 where the last equality follows from Cauchy's integral formula. For the general case, by applying the inductive assumption (3.15) to , we have I∗e1ξ(k−1) −e1 I∗e1ξ(k−1) −e1 = (−1)k−1I∗e1(cid:0)ωv(−e1)(w−e1, qe2)(cid:1)· Xlk−1∈Lk−1 v(−e1) s(lk−1)β(lk−1) ξ−ek+O(sk). Kv(ze, we1)I∗e1(cid:0)ωv(−e1)(w−e1, qe2)(cid:1) = −se1 Therefore it suffices to prove that for any e1 ∈ Ev we have: (3.17) we1∈γe1 This is due to I∗e1(cid:0)ωv(−e1)(w−e1, qe2)(cid:1) = I∗e1 ((β−e1,e2 + o(w−e1))dw−e1) = − se1 β−e1,e2 dwe1 the induction for (3.15). The expansion (3.14) for η(k) + O(s2) and Cauchy's integral formula. We conclude (ze1) against Kv(z, ze1), and the computation is exactly the same as (3.17): v (z) is obtained by integrating ξ(k−1) w2 e1 e1 ωv(ze, qe1)β−e1,e2+O(s2). Kv(z, we1)I∗e1(cid:0)ωv(−e1)(w−e1, qe2)(cid:1) = −se1ωv(z, qe1)β−e1,e2+O(s2), we1∈γe1 where z ∈ bCv. The proof is thus completed. Remark 3.5. It is important to point out that the expansion (3.14) is not the s-expansion of the solution ηv, while the latter is also com- putable by expanding the error term in (3.16) using the higher order (cid:3) 22 XUNTAO HU AND CHAYA NORTON coefficients βv ij of ωv. The explicit formula for the case where ΓC con- tains only one edge is given by [Yam80], and will be recomputed (up to the second order) in Section 5. and practical to consider ηv as the series P∞k=1 η(k) However, as we highlighted by the proposition, it is often more useful v , given the bound (3.13) and the recursive construction (3.9). In most cases it is already useful to know the first non-constant term of Ωv,s, which the proposition suffices to give: η(1) v (z) = −Xe∈Ev seωv(z, qe) ξ−e + O(s2 e). 4. Period Matrices in Plumbing Coordinates 4.1. General periods. Using the construction (3.8) and expansion (3.14) of the stable differential Ω, we can compute the variational for- mula of its periods. Notation 4.1. For a stable curve C and its dual graph ΓC, define the map p : H1(C, Z) → H1(ΓC, Z) as follows: for the class of a homological (oriented) 1-cycle [γ] on C, p([γ]) is the class of the oriented loop in the dual graph that contains the vertices corresponding to the components that γ passes, and the edges corresponding to the nodes contained in γ. The orientation of p([γ]) is inherited from the orientation of γ. It is easy to see that the map is surjective, but not injective unless all components have genus zero. Moreover, if γ is completely contained in some component Cv, then p([γ]) = 0. Let α be any closed oriented path on the stable curve C, such that p([α]) 6= 0 (the zero case is trivial in our discussion below). For any small enough s, there exists a small perturbation αs of α such that the requiring 1) αs ∩ γe = I−1 e (αs ∩ γ−e) for any seam γe that α passes; 2) αs does not totally contain any seam γe. By an abuse of notation, the path on Cs after the gluing is also denoted by αs. restriction of αs on bCs glues to be a path on Cs. This can be seen by Ωs. To this end, recall that Ue = {ze < pse} and The following theorem computes the leading terms in the variational formula of ´αs denote We = {ze < se} and Ve = {ze < 1}. Theorem 4.2. For any stable differential Ω on C with residue re at the node qe, let α be any closed oriented path on C such that p([α]) 6= 0 and {e0, . . . , eN−1} be the collection of oriented edges that p([α]) passes through (with possible repetition), such that v(−ei−1) = v(ei), and let GENERAL VARIATIONAL FORMULAS FOR ABELIAN DIFFERENTIALS 23 eN = e0. Then we have NXi=1 ei (√sei) (4.1) αs Ωs = (rei lnsei + ci + li) + O(s2), where ci and li are the constant and linear terms in s respectively, explicitly given as (4.2) ci = lim s→0 z−1 z−1 −ei−1 (4.3) (rei−1 lnsei−1 + rei lnsei)! , ξ−e · σe, 1 2 se Ωv − (√sei−1) li := − Xe∈Ev(ei) where ξe is defined in (3.5), and (4.4) σe := ei (sei ) lims→0 z−1 ´q−ei−1 lims→0 z−1 −ei−1 ´ qei (sei ) ωv(ei)(z, qe) ´ qei q−ei−1  ωv(ei)(zei, qei) + 1 sei! ωv(ei)(z−ei−1, q−ei−1) − 1 if e = ei; sei if e = −ei−1; otherwise. Remark 4.3. Prior to the proof of the theorem we have two remarks. Firstly, the period integral in (4.1) depends not only on p[α], but also on the class of the actual path α. The integration over α ∩ bCv gives precisely the constant term (4.2). Secondly, note that the limits of the integrals in (4.4) are singular because the integrants have a double pole on the nodes. However the singular parts are canceled by ± 1 , so the limits are indeed well-defined. Computations leading to both remarks are contain in the proofs of the following lemma and the theorem. sei To prove the theorem, it suffices to compute the integral on each component Cv(ei), i = 1 . . . N that α passes through. To simplify no- tation, throughout the proof below we consider α only passing each component once, while the proof also holds for the general case. Let us denote the intersection of αs with ∂Uei, ∂Vei, ∂U−ei−1, ∂V−ei−1 respec- tively by uei, vei, u−ei−1 and v−ei−1. Then αsCv(ei) breaks into three pieces bounded by the four points: (1) αsV−ei−1\U−ei−1 (2) αs bCv(ei)\Vei∪V−ei−1 (3) αsVei\Uei connecting u−ei−1 and v−ei−1; connecting vei and v−ei−1; connecting vei and uei; 24 XUNTAO HU AND CHAYA NORTON For convenience, by a composition of a rotation we can assume that ei (√sei), vei = z−1 −ei−1(√s−ei−1), uei = z−1 v−ei−1 = z−1 −ei−1(1). Therefore in the lemma and the proofs below, integrations in the local chart zei from uei to vei will be written as from √sei to 1, for any i = 0, . . . , N − 1. We also remark that this ei (1) and similarly u−ei−1 = z−1 assumption does not change the statement of the theorem. The following lemma simplifies the computation: Lemma 4.4. Given an edge e, let Ωv,s and ξ(k) We have the following equality: e be defined as before. √se (4.5) 1 Ωv(e),s(ze) + 1 √se Ωv(−e),s(z−e) =re lnse + 1 + se ∞Xk=0 ξ(k) e (ze) se 1 ∞Xk=0 ξ(k) −e (z−e) v , and as we are concerned with the regular part of the period, locally in the annuli e (ze) + Ve \ We, we have the following expression Ωv,s(ze) = re Proof of Lemma 4.4. We recall that Ωv,s = Ωv(ei) +Pk η(k) P∞k=1 η(k) v . The logarithmic term in (4.5) is given by + ξ(0) √se dze ze 1 re dze ze = 1 2 re lnse 1 1 √se √se ξ(0) η(k) v + ξ(0) −e + ∞Xk=1 e + 1 and re = −r−e. What is left to show is (4.6) √se ∞Xk=1 Note that for each k ≥ 0, we have ´ se√se ξ(k) −e (z−e) = ´ √se ∞Xk=0 e (ze) = ´ 1 ξ(k) √se −e (ze). This gives for k ≥ 0: η(k) v(−e) = 1 I∗e ξ(k) √se se √se 1 (4.7) √se ´ 1 ξ(k) e (ze) + 1 √se −e (z−e) = se ξ(k) − 1 1 √se ξ(k) e (ze) + 1 se ξ(k) −e (z−e) √se I∗e ξ(k) e (z−e) − 1 I∗e ξ(k) −e (ze). se 1 ξ(k) e + 1 se ξ(k) −e ∞Xk=0 I∗e ξ(k) e (z−e) and GENERAL VARIATIONAL FORMULAS FOR ABELIAN DIFFERENTIALS 25 Grouping the last two terms above with the (k+1) entries inP∞k=1´ P∞k=1´ 1 η(k) v(−e), and applying (3.11), we obtain: √se e )(z−e) √se −e )(ze) + 1 v(e) − I∗e ξ(k) (η(k+1) = ξ(k+1) e √se √se 1 (η(k+1) v(−e) − I∗e ξ(k) (ze) + 1 ξ(k+1) −e √se (z−e) (4.8) 1 √se 1 η(k) v + Summing up both (4.7) and (4.8) over all k ≥ 0 and adding the two equalities together, we immediately obtain (4.6). The lemma follows. (cid:3) Proof of Theorem 4.2. Our goal is to compute the leading terms of the Ωv(ei),s. To this end, we rearrange the terms and compute the following integrals: variational formula of PN−1 Ωv(ei),s + z−1 (4.9) ei (1) i=0 ´αv(ei) √sei 1 z−1 −ei−1 (1) Ωv(ei),s(zei) + 1 √sei Ωv(−ei),s(z−ei) 1 sei ξ(k) z−1 −ei−1 (1)(cid:16)Ωv +Pk≥1 η(k) v (cid:17). It needs to be pointed out that the first two entries above are integrals inside Cv, while the last entry is in Cv(−ei). To simplify notation, in the rest of the proof we denote v := v(ei). Using the lemma, the last two entries of (4.9) are equal to rei lnsei+ ei (zei) +P∞k=0´ 1 P∞k=0´ sei ξ(k) −ei(z−ei). By definition of Ωv,s, the first integral in (4.9) is equal to ´ z−1 ei (1) Note that by (3.14) and (3.15), for k ≥ 1 the integrals of ξ(k) only give terms of order ≥ k. Also observe that ´ vei constant term we only have to compute the integrals of ´ sei ´ 1 ξ(0) −ei(z−ei). sei Since ξ(0) ±ei(z±ei) is holomorphic in V±ei, we have sei ±ei and η(k) Ωv is a v constant independent of s. Thus to compute the remaining part of the ξ(0) ei (zei) + −ei(z−ei) = 0 ξ(0) ei (zei) + 1 ei (zei) + 1 ξ(0) −ei(z−ei) v−ei−1 ξ(0) ξ(0) 1 (4.10) 1 sei 1 0 + sei ·(cid:16)ξei − ξ−ei(cid:17) + O(s2). Summing up the constant terms on the RHS over i, we have computed the constant term (4.2). 26 XUNTAO HU AND CHAYA NORTON Now we compute the linear term. Note that ξ(k) are holomorphic in s. Again by (3.15), we only need to compute the integrals of ξ(1) v(±ei), whose expansion is already given by (3.16). Therefore we have the following: v sei 1 ξ(1) ei (zei) = −Xe∈Ev = −Xe∈Ev =sei 1 se se ξ−e sei ξ−e sei ξ−ei − Xe∈Ev,e6=ei ( sei ξ−ei lim sei→0 1 1 − sei 1 eωv(zei, qe)dzei + O(s2) ξ−ei sei eωv(zei, qe)dzei + sei se ξ−e 0 1 eωv(zei, qe)dzei eωv(zei, qei)dzei + 1 sei The existence of the limit above can be seen by integrating the 1/z2 ei ) + O(s2). dzei z2 ei + O(s2) term in the expansion ofeωv(zei, qei). The linear term in ´ 1 ξei − Xe∈Ev(−ei),e6=−ei 1 ξ(1) −ei(z−ei) = − sei sei sei ξ(1) −ei(z−ei) is computed similarly: Note that the linear terms in (4.10) have been canceled by the linear terms produced by the singular part of ωv. Moreover, se ξ−e 1 0 eωv(z−ei, qe)dz−ei ) + O(s2). 1 sei ( 1 seieωv(z−ei, q−ei)dz−ei − se · ξ−e · z−1 ei (1) z−1 −ei−1 (1) ωv(z, qe) + O(s2). ξei − sei lim sei→0 vei v−ei−1 η(1) v (z) = −Xe∈Ev Summing up all the linear terms above, then summing up over i, we (cid:3) have the desired linear term. Remark 4.5. Note that in the proof of the theorem, the function h(s) := ´αs NXi=1 ∞Xk=0 Ωs −PN sei 1 i=1 rei lnsei is computed as ξ(k) ei (zei) + 1 sei ∞Xk=0 −ei(z−ei) + z−1 ξ(k) ei (1) z−1 −ei−1 (1) Ωv(ei),s! . The analyticity of h(s) in s follows from the analyticity of each inte- grand above. The analyticity of h(s) will be used in our improvement GENERAL VARIATIONAL FORMULAS FOR ABELIAN DIFFERENTIALS 27 of the result of [Tan91] below. In [GKN17, Lem. 5.5], without com- puting any terms in h(s), an estimate of h(s) is derived in the real normalized setup. Moreover, from the proof one can see that besides the complexity of the computation, there is no obstacle in computing every higher order terms in the expansion of the periods of Ωs. 4.2. Period matrices. Recall that in Section 2.4 we have chosen a ba- sis {Ai,s}g i=1 for a Lagrangian subspace of H1(Cs, Z) along the plumbing family. We required that the first m A-cycles generate the span of the classes of the seams. In order to study degenerations of the period matrix, we now choose B1,s, . . . , Bg,s completing the A-cycles to a sym- plectic basis of H1(Cs, Z). The cycles B1,s, . . . , Bg,s are chosen such that they vary continuously in the family. One can easily see that for 1 ≤ k ≤ m, p([Bk,0]) 6= 0, while for m + 1 ≤ k ≤ g, the map p annihilates the classes of Bk,0. From now on we write Ak := Ak,0, Bk := Bk,0. Note that for 1 ≤ k ≤ m, one can also see Bk,s as constructed from Bk by applying a small perturbation as introduced in the previous section. By the Riemann bilinear relations, we define the following basis of k=1 in H 0(C, KC) where C = C0 is a stable Ak is contained in the same component. Define vk(z) to be the Abelian differential dual to Ak in H 0(Cv, KCv). Abelian differential {vk}g curve: (1) For m + 1 ≤ k ≤ g, Bk is contained in eCv for some v, thus e0, . . . eN−1. Define vk :=PN−1 (2) For 1 ≤ k ≤ m, p([Bk]) 6= 0, assume p([Bk]) passes the edges the A-normalized meromorphic differential of the third kind supported on Cv(ei) that has only simple poles at q−ei−1 and qei with residues −1 and 1 correspondingly. , where ωqei−q−ei−1 i=0 ωqei−q−ei−1 denotes vk,s = ´γe By applying the jump problem construction, we have a collection of Abelian differentials {vk,s}g k=1 for the curve Cs, which is seen to be a normalized basis of H 0(Cs, KCs). For every k and e, we have ´γe (vk + ηk,s) on Cs. Since the solution ηk,s to the jump problem with initial jumps of vk is A-normalized, this is equal to the integral ´γe vk on Cv(e). Therefore by the residue theorem, we have ´Aj,s k=1 is a normalized ba- sis of H 0(Cs, KCs). The period matrix of Cs is hence defined to be {τh,k(s)}g×g where τh,k(s) := ´Bh,s vk,s = 2πi · δjk. This shows that {vk,s}g We can apply Theorem 4.2 to compute the leading terms in the vk,s. variational formula of τh,k(s). 28 XUNTAO HU AND CHAYA NORTON Corollary 4.6. For every e and k, denote Ne,k := γe · Bk,s the intersection product. For any fixed h, k, the expansion of τh,k(s) is given by τh,k(s) = Xe∈EC (Ne,h · Ne,k) · lnse NXi=1 z−1 z−1 −ei−1 ei (√sei) (√sei−1) vk − Nei,hNei,k lnsei! se (hol(vk)(qe) hol(vh)(q−e)) + O(s2), (4.11) + lim s→0 − Xe∈EC i=0 where {ei}N−1 is the set of oriented edges p([Bh]) passes through, and hol(vk) denotes the regular part of the Laurent expansion of the function of vk near the nodes of the components where vk is not identically zero. Furthermore, under our choice of the symplectic basis, Ne,h · Ne,k is equal to 1 if h = k and the node qe lies on Bh and equals 0 otherwise. Remark 4.7. (1) For the purpose of defining the intersection product, we assign an random orientation to γe. We further remark that there is no canonical way to orient γe, and the assigned orientation does not affect the statement and the proof. (Ne,h· Ne,k) · lnse is holomorphic in s. We can see that only the logarith- mic term was computed. By Remark 4.5, our corollary in particular reproves his result, and we express more terms in the expansion. (2) The main result in [Tan91] is that h(s) := τh,k(s)−Pe∈EC (3) We want to point out that Taniguchi does not require the classes of γe to be part of the symplectic basis, therefore Ne,h · Ne,k may be any integer. Since the A, B-cycles generate H1(Cs, Z), the general case follows by linearity. Proof. We first compute the logarithmic term. Note that the intersec- tion product is independent of s. When e does not lie on p([Bh,s]), we have Ne,h = 0, otherwise Ne,h = ±1 and the sign depends on the orientation of [γe] compared to that of the corresponding generator [Ai] of the symplectic basis. We now only need to prove that vk has residue Ne,h · Ne,k at qe, which is seen as follows: if e ∈ p([Bk]), then by construction of vk, it has residue ±1 = Ne,k at qe depending again on whether [γe] = [Ai] or −[Ai]; if e does not lie on p([Bk]), both the intersection number and the residue are 0. Note that the signs of Ne,h and Ne,k are always the same, therefore Ne,h · Ne,k = δi,h · δi,k. Secondly, we compute the linear term. Note as can be verified from the normalization conditions of the fundamental bidifferential, vh(z) = GENERAL VARIATIONAL FORMULAS FOR ABELIAN DIFFERENTIALS 29 ω(w, z) for m + 1 ≤ h ≤ g, and vh(z) =PN ´Bh 1 ≤ h ≤ m, where {ei}N−1 To compute the linear term for τh,k(s) = ´Bh,s definition of σe in (4.4), we have ωv(ei)(w, z) for is the set of edges p([Bh]) passes through. vk,s, we observe that by i=0 i=1´ qei q−ei−1 σe = hol(vh)(qe). Since Ω := vk, we have ξe = hol(eΩ)(qe) = hol(vk)(qe). Note that vh is only supported on ∪v∈p([Bh])Cv, the sum in (4.11) is taken over all edges. Lastly, the constant term follows directly from Theorem 4.2. (cid:3) 5. Examples In this section we will compute four explicit examples of the varia- tional formula for Abelian differentials and for the period matrix of a stable curve C. Throughout this section the stable curve has geometric genus g, Ω is a stable differential on C. We choose the symplectic basis of holomorphic 1-cycles and its dual basis of 1-forms as in the previous sections. The notation will vary among the examples according to the structure of C. 5.1. One node. We first deal with the case where the curve C has only one node q. In [Yam80], Yamada computed the variational formula of both Abelian differentials and the period matrices to any order of the plumbing parameter s. We will reprove his result up to the second order, while the full expansion can also be found using our method. 5.1.1. One node: compact type. When C is of compact type, it has two components C1 and C2 that meet at a single separating node q, whose pre-images are denoted by q1 ∈ C1 and q2 ∈ C2. Let zi be the local coordinates near qi. Denote the restriction of Ω to Ci by Ωi (i = 1, 2). The subscripts of the Cauchy kernel and its derivative are changed correspondingly. Since the curve is of compact type, the differentials Ωi have no residue at qi, therefore they are holomorphic and we have ξ(0) i (zi) = Ωi(zi). We denote ξi := ξ(0) i (qi), and Ωs is defined on Cs by formulas (3.3) (3.9). Explicitly, by Proposition 3.4, the expansion of the restriction Ωi,s (i = 1, 2) is given by (5.1) Ωi,s(z) = Ωi(z) +(cid:16)−s · ωi(z, qi) ξi′ + s2 · ωi(z, qi)βi′ ξi(cid:17) + O(s3) 30 XUNTAO HU AND CHAYA NORTON where by convention, i′ = 2 if i = 1 and vice versa, and βi denotes the leading coefficient in the expansion of ωi as in (3.1). Let gi be the genus of Ci, with g1 + g2 = g. We take a normalized k=1 of C such that {v1, . . . , vg1} are basis of Abelian differentials {vk}g supported on C1, and {vg1+1, . . . , vg} on C2. For 1 ≤ k ≤ g1, letting Ω1 = vk, Ω2 = 0 in (5.1), we obtain vk,s(z) =(vk(z) + s2 · ω1(z, q1)β2vk(q1) + O(s3) −s · ω2(z, q2)vk(q1) + O(s3) z ∈ eC1, z ∈ eC2. For g1 + 1 ≤ k ≤ g, the formula is symmetric. We have then reproven [Yam80, Cor. 1]. 5.1.2. One node: non-compact type. In this case C is irreducible, with a single node q. We denote q1, q2 the pre-images of q in the normal- ization eC, and z1, z2 the corresponding local coordinates. Let Ω be a meromorphic differential on the normalization of C which has simple poles of residues ri at qi (i = 1, 2). By the residue theorem r1 = −r2. i (qi). By . Denote ξi = ξ(0) We now have ξ(0) i (zi) = Ω(zi) − ridzi zi Proposition 3.4, we have (5.2) Ωs(z) = Ω(z) − s ·(cid:16)ω(z, q1) ξ2 + ω(z, q2) ξ1(cid:17) + O(s2). For a symplectic basis {Ak,s, Bk,s}g k=1, we choose A1,s to be the seam, whereas B1,s is taken to intersect A1,s once, oriented from the neigh- borhood of q1 to the neighborhood of q2. As in Section 4.2, we take v1 = ωq2−q1, and {vk}g By letting Ω = vk for 2 ≤ k ≤ g, we have ri = 0, and the equation (5.2) gives [Yam80, Cor. 4]. For the case Ω = v1, we have r2 = −r1 = 1, then (5.2) gives [Yam80, Cor. 5]. Moreover, we can compute the period matrix of Cs, reproving [Yam80, Cor. 6]. By Corollary 4.6 we have k=2 to be the normalized basis of H 1(eC, C). τ1,1(s) = B1,s where c1,1 = lims→0(cid:18)´ z−1 ωq2−q1,s = lns + c1,1 + s · l1,1 + O(s2), 2 (√s) 1 (√s) τk,1(s) = τ1,k(s) = B1,s ωq2−q1 − lns(cid:19), and l1,1 = −2σ1σ2. vk,s = c1,k + s · l1,k + O(s2) We also have z−1 for 2 ≤ k ≤ g. Since vk(x) is holomorphic, we have ξ(0) and hence ξi = vk(pi). The constant term c1,k is equal to ´ q2 the linear term l1,k is seen to be −vk(q1)σ2 − vk(q2)σ1 by (4.11). i (zi) = vk(zi) vk, and q1 GENERAL VARIATIONAL FORMULAS FOR ABELIAN DIFFERENTIALS 31 Finally for 2 ≤ k, h ≤ g we have τk,h(s) = τk,h + s · lhk, where τh,k is the period matrix of the normalization of C, and lh,k = −vk(q1)vh(q2) − vh(q1)vk(q2). 5.2. Banana curves. The second example we consider is the stable genus g curve C that has two irreducible components meeting in two distinct nodes (so-called "banana curve"). This computation has not been done in the literature before. Let the two components of C be Ca, Cb with genera ga and gb where ga + gb = g − 1. The edges corresponding to the two nodes are denoted by e1 and e2. The preimages of the nodes and the local coordinates are denoted as q±ei and z±ei (i = 1, 2), where "+" corresponds to the Ca side, and "−" the Cb side. Note that in the case where C has only two components (with any number of nodes connecting them), the path lk can only go back and forth. Therefore ξ(k) b ) are ei determined by Ωa if k is even (resp. odd), and Ωb if k is odd (resp. even), as we can see from the terms in expansion (5.1). Also note that in expansion (3.14) of η(k) , there is no residue of Ω in- volved. Therefore the we can simplify our computation by assuming that Ωb = 0 and the residues of Ωa at both nodes are zero. Under these assumptions, we have ξ(0) ei (zei) = Ωa(zei) (i = 1, 2). Furthermore, for any integer k ≥ 0, we have −ei) (i = 1, 2) and η(k) (resp. η(k) (resp. ξ(k) and η(k) a a b ξ(2k) −ei = ξ(2k+1) ei = 0 (i = 1, 2), z ∈ bCa; z ∈ bCb. thus by construction (3.9), we have (z) = 0 η(2k+1) a η(2k) b (z) = 0 By Proposition 3.4, we have for z ∈ bCb: and for z ∈ bCa: 1ωa(z, qe1)βb 1,1 a (z) = s2 η(2) ξe1 + s2 + s1s2(cid:16)ωa(z, qe1)βb 1,2 jk := βb where βb as in (3.1). η(1) b (z) = −s1ωb(z, q−e1) ξe1 − s2ωb(z, q−e2) ξe2 + O(s2), 2ωa(z, qe2)βb 2,2 ξe2 + ωa(z, qe2)βb 2,1 ξe2 ξe1(cid:17) + O(s3), −ej ,−ek is the constant term in the expansion of ωb(z−ej , z−ek) 32 XUNTAO HU AND CHAYA NORTON Note that we can also assume Ωa = 0, and that the residues of Ωb at both nodes are zero. The general case follows by adding the differentials in these two cases together. + ωq−e1−q−e2 where ωqe2−qe1 We now compute the degeneration of period matrix for the ba- nana curve. For the symplectic basis of H1(Cs, Z), we let A1 := γe2, and B1 is taken to intersect each seam once, with the orien- tation from qe1 to qe2, then from q−e2 to q−e1. Thus we let v1 := is supported on Ca, and ωq−e1−q−e2 ωqe2−qe1 on Cb. Take {Ak, Bk}ga+1 j=ga+2 to be the symplectic bases of H1(Ca, Z) and H1(Cb, Z) respectively. The normalized basis of holomorphic differentials {vk}g k=2 on the two components are taken correspondingly, and we require that vk is identically zero on Cb if 2 ≤ k ≤ ga + 1, and on Ca if ga + 2 ≤ k ≤ g. lns1 + lns2 + c1,1 + l1,1 + O(s2). By (4.2), the constant term is Note that v1 has residues re2 = r−e1 = 1, thus we have τ1,1(s) = k=2 and {Aj, Bj}g c1,1 = lim s→0 z−1 (√s2) (√s2) As for the linear term l1,1, by (4.11) we obtain e2 (√s2) e1 (√s1) + z−1 ωqe2−qe1 z−1 −e2 z−1 −e1 ωq−e1−q−e2 − lns1 − lns2! . if 2 ≤ k ≤ ga + 1, if ga + 2 ≤ k ≤ g. h,k=2 of l1,1 = −2s1σ−e1σe1 − 2s2σ−e2σe2. We also see that the expansion of τk,1(s) = τ1,k(s) is given by qe1 q−e2 the period matrix is computed as: vk − s1vk(qe1)σ−e1 − s2vk(qe2)σ−e2 + O(s2) vk − s2vk(q−e2)σe2 − s1vk(q−e1)σe1 + O(s2) τ1,k(s) =(´ qe2 ´ q−e1 The remaining (g − 1) × (g − 1) minor τg−1(s) := {τh,k(s)}g τg−1(s) =(cid:18)τa − s2 ·(cid:18) τb(cid:19) − s1 ·(cid:18) (cid:19) (cid:19) + O(s2), tRb(q−e2)Ra(qe2) tRb(q−e1)Ra(qe1) tRa(qe1)Rb(q−e1) tRa(qe2)Rb(q−e2) 0 0 0 0 0 0 where τa (resp. τb) is the period matrix of Ca (resp. Cb), and Ra := (v2, . . . , vga+1), Rb := (vga+2, . . . , vg). 5.3. Totally degenerate curves. It is a fact that the stable curves that lie in the intersection of the Teichmuller curve and the boundary of Mg are of arithmetic genus zero. In this subsection we study the largest dimensional boundary stratum of such stable curves and give the variational formula of its period matrix, which to the knowledge of the authors is again not dealt with in literature before. The periods GENERAL VARIATIONAL FORMULAS FOR ABELIAN DIFFERENTIALS 33 of totally degenerate curves has been studied by Gerritzen in his series of papers [Ger90] [Ger92a] [Ger92b]. The perspectives in those papers are algebraic, mainly by studying the theta functions, the Torreli map and the Schottky problem. No analytic construction such as plumbing is involved. eC is a P1 with g pairs of marked points {q±i}g Let C be a totally degenerate stable curve, namely the normalization i=1. Let qi and q−i be the preimages of the i-th node on C, and ri = −r−i be the residue of Ω at qi. Let z be the global coordinate, then the local coordinates at the pre- images of nodes are given by z±i := z − q±i, where i = 1, . . . , g. We have as usual ξ(0) ±i (z) := Ω(z) ∓ ridz z−q±i , and ξ±i := ξ(0) ±i (q±i). explicitly: K(z, w) = The Cauchy kernel and the fundamental bidifferential on P1 are given (z−w)2 . We (z−qi)2 , for i ∈ {±1, . . . ,±g}. The expansion of Ωs 2πi(z−w); ω(z, w) = 2πi∂wK(z, w) = dzdw dz compute eω(z, qi) = 1 is thus given by Ωs(z) = Ω(z) − dz sk ξ−k (z − qk)2 + ξk (z − q−k)2! + O(s2) gXk=1 where z ∈ bC. The classes of the seams {[γi]}g dz(cid:16) 1 z−qi − 1 (qk−qi)(qk−q−i) and ξi = ξ−i = 1 qi−q−i q−i−qi i=1 generate the Lagrangian subgroup of H1(Cs, Z), thus we can take Ai := γi and Bi the path from q−i to qi. The corresponding normalized basis of 1-forms will be vi := ωqi−q−i = z−q−i(cid:17) for i = 1, . . . , g. Let Ω = vi, we have for k 6= ±i, One thus computes the period matrix as follows. ξk = . i = j : τi,i = lnsi − 2 lnqi − q−i − 2si (qi − q−i)2 2sk(qi − q−i)2 − Xk∈{1,..i,..g} (qk − q−i)(qk − qi)(q−k − q−i)(q−k − qi) + O(s2) 34 XUNTAO HU AND CHAYA NORTON i 6= j : τi,j = ln (qi, q−i; qj, q−j) −Xk6=i,j sk(cid:16) (qi − q−i)(qj − q−j) (qk − qi)(qk − q−i)(q−k − qj)(q−k − q−j) + − si − sj qj − q−j (qi − q−i)(qj − q−j) (qk − qj)(qk − q−j)(q−k − qi)(q−k − q−i)(cid:17) q−i − qi(cid:16) q−j − qj(cid:16) (qi − qj)(qi − q−j) (qj − qi)(qj − q−i) qi − q−i + + 1 1 1 (q−i − qj)(q−i − q−j)(cid:17) (q−j − qi)(q−j − q−i)(cid:17) + O(s2) 1 where (qi, q−i, qj, q−j) stands for the cross-ratio of the (ordered) four points. 6. Higher Order Plumbing In the last section we construct the solution to the jump problem with the initial data arising from the jumps of an Abelian differential that has higher order zeroes and poles at the nodes of the limit curve. Following the terminology in [Gen15] and [BCGGM16], we call the procedure of smoothing such a differential higher order plumbing. We obtain an alternative proof of the sufficiency part of the main theorem in [BCGGM16]. Moreover, our approach gives more information than the two constructions given in that paper. We first give a brief review of definitions and results in [BCGGM16]. Readers that are familiar with this material can safely skip the following subsection. 6.1. Incidence variety compactification. Let µ = (m1, . . . , mn) be a partition of 2g−2, and we assume mi > 0 for i = 1, . . . , n. We denote ΩMg,n(µ) to be the stratum whose points are Abelian differentials that inc have multiplicity mk at the marked point pk. Let PΩM g,n(µ) be the closure of the strata in the projectivized compactification of the Hodge bundle PΩMg,n over Mg,n, called the incidence variety compactifica- tion (IVC) of the stratum ΩMg,n(µ) in [BCGGM16]. Take a stable pointed differential (C, Ω) in the boundary of ΩMg,n, where C is a stable nodal curve with marked points p1, . . . , pn, and Ω is a stable differential on C. Let (C,W) → ∆ be a one parameter family in the stratum ΩMg,n(µ), where ∆ is a disk with parameter t, such that C0 = C. Note that Ω may be identically zero on some irreducible component Cv of C. By an analytic argument [BCGGM16, Lemma 4.1] one can show that there exist lv ∈ Z≤0 for each Cv such that Ξv := lim t→0 tlvΩv GENERAL VARIATIONAL FORMULAS FOR ABELIAN DIFFERENTIALS 35 is non-zero and not equal to infinity. Such differential {Ξv}v must sat- isfy the following conditions (see the proof of necessity of [BCGGM16, Theorem 1.3]): (0) If pk ∈ Cv for some k, Ξv vanishes to the correct order: ordpk Ξv = (1) The only singularities of Ξv are (possible) poles at the nodes of mk; Cv; (2) For any node qe on C, ordqe Ξv(e) + ordq−e Ξv(−e) = −2; (3) If ordqe Ξv(e) = ordq−e Ξv(−e) = −1 at some node qe, then the residues are opposite at the node: resqe Ξv(e) = − resq−e Ξv(−e). Definition 6.1 ([BCGGM16, Def. 1.1]). A differential Ξ satisfying Conditions (0) ∼ (3) is called a twisted differential of type µ. Given a one parameter family, l : v 7→ lv gives a (full) level function on the vertices of the dual graph ΓC. The function l makes ΓC into a level graph, in which the order is denoted by " < ". Moreover, the twisted differential Ξ constructed from the one-parameter family must satisfy the following conditions (again see the proof of necessity of [BCGGM16, Theorem 1.3]): (4) At a node e, v(e) < v(−e) if and only if ordqe Ξv(e) ≥ ordq−e Ξv(−e), and v(e) ≍ v(−e) if and only if ordqe Ξv(e) = ordq−e Ξv(−e) = −1 (5) For any level L in the level graph, for any v such that lv > L, let EL v be the set of all the nodes e such that v(e) = v, lv(−e) = L, we have Xe∈EL v resq−ei Ξv(−ei) = 0. The last condition is called the Global residue condition in [BCGGM16]. Definition 6.2 ([BCGGM16, Def. 1.2]). A twisted differential Ξ is called compatible with the stable differential Ω and the full level func- tion l (or equivalently the full level graph ΓC) if (i) Ξ and l satisfy the Conditions (0)∼(5); (ii) the maxima of the level graph correspond to the components Cv where Ωv is not identically zero, and on those components, Ξv = Ωv. The main result of [BCGGM16] is that the necessary and sufficient condition for a pointed stable differential (C, Ω) to lie in the bound- ary of the IVC compactification of strata is the existence of a twisted differential Ξ (on C) and a full level function l (on ΓC) such that Ξ is compatible with Ω and l. 36 XUNTAO HU AND CHAYA NORTON 6.2. Jump problem for higher order plumbing. The proof of suf- ficiency of this result requires a construction of a family of Abelian dif- ferentials in the smooth locus of the strata that degenerates to the limit differential (C, Ω), given the compatible data (Ξ, l). In [BCGGM16], the authors give two proofs to the sufficiency by: 1) constructing a one complex parameter family using plumbing; 2) constructing a one real parameter family via a flat geometry argument. We now give a third argument via the jump problem approach. Moreover, the number of parameters over C in our degenerating family is equal to the number of levels in ΓC minus 1. Similar to the plumbing argument used in [BCGGM16], we will also use a modification differential to match up the residues. Furthermore, the original argument in [BCGGM16] on the operation merging the zeroes will also be applied here to embed the family into the stratum. Assume j = lv(−e), recall that Ej Take a plumbing family {Cs} as in Definition 2.2 such that C = C0, and s = {se}e∈EC are the plumbing parameters. Denote the restriction of Ξ on the irreducible component Cv by Ξv. Let Nl be the number of levels in ΓC. Without loss of generality, we assume the range of the level function l to be {0,−1, . . . , 1 − Nl}. v = {e ∈ Ev : lv(−e) = j} as defined in condition (5). To glue the twisted differentials Ξv and Ξv(−e), we need to add a modification differential φv,j to Ξv in order to match the residue (denoted by r−e) of Ξv(−e). The modification differential φv,j is chosen to be any differential which has simple pole at qe with residue re = −r−e, where e ∈ Ej v. The global residue condition ensures that the sum of the residues of φv,j is zero. The existence of φv,j is due to the classical Mittag-Leffler problem. For e ∈ Ej v, assume Ξv has a zero of order ke at the node qe, then by Condition (2), Ξv(−e) has a pole of order ke + 2 at the node q−e. In order to apply the jump problem to obtain a global differential, the following conditions need to be imposed on the plumbing parameters s: v, we have ske+1 (i) For any e, e′ ∈ Ej (ii) For any two vertices v0, v1 at different levels (namely lv0 6= lv1), for any two paths {ei}i∈I and {ej}j∈J connecting v0, v1 with lv(ei) > lv(−ei) (∀i ∈ I) and the same for {ej}, we have e′ e = ske′ +1 ; Qi∈I skei +1 ei =Qj∈J s kej +1 ej =: s(v0, v1); (iii) If lv0 = lv1, we require that s(v0, v1) = 1. It is important to remark that for such a tuple of plumbing param- eters one can deduce that s(v0, v1) depends only on the levels of v0, v1, GENERAL VARIATIONAL FORMULAS FOR ABELIAN DIFFERENTIALS 37 namely, s(v0, v1) = s(v′0, v′1) as long as lv0 = lv′ thus choose one parameter for each level drop: 0 and lv1 = lv′ 1 . We can Definition 6.3. Let ti,j := s(v0, v1) where v0, v1 are two vertices at level i, j respectively. The tuple t := {t−1, . . . , t1−Nl} where ti := t0,i/t0,i+1 are called the scaling parameters. Note that ti,j = Qj k=i tk. The theorem below gives a degenerating family of Abelian differentials parametrized by t with central fiber the differential (C, Ω) in the boundary of the IVC. Theorem 6.4. Let (C, Ω, p1, . . . , pn) be a stable pointed differential in a given stratum ΩMg,n(µ). Given the triple (C, Ξ, l) where Ξ is a twisted differential of type µ on C and l is a full level function on ΓC, such that Ξ is compatible with Ω and l, there exists a degenerating family of Abelian differentials (Ct, Ξt) ⊂ ΩMg,n(µ) such that limt→0(Ct, Ξt) = (C, Ω), where t are the scaling parameters. Proof. The proof is completed in three steps. Firstly we construct via the jump problem method a degenerating family of Abelian differentials the stratum, that is, we show that the solution to the jump problem is uniformly controlled by some positive power of t := max1≤i≤Nl−1 ti. obtain a family contained in the stratum. (Ct,bΞt) in ΩMg,n. Then we show that the family lies sufficiently "near" Lastly we apply [BCGGM16, Lemma 4.7] to merge the zeroes ofbΞt to correct initial data {ξ(0) and (3.8). For the jump problem construction, we only need to construct the e }, the rest of the construction is given by (3.3) Assume that the vertex v lies on the i-th level. We define bΞv := Ξv +Xj<i ti,jφv,j, where φv,j is the modification differential we defined earlier. We now apply the jump problem construction to glue the differentials at the opposite sides of each node qe. Assume v(e) and v(−e) are on opposite sides of the node qe. Namely, let the levels i and j respectively. We glue t0,i ·bΞv and t0,j ·bΞv(−e) from the ξ(0) e (ze) := t0,i ·(cid:16)bΞv(ze) − ti,jI∗e P (bΞv(−e))(ze)(cid:17) ; −e (z−e) := t0,j · hol(bΞv(−e))(z−e), ξ(0) where P (·) denotes the principal part of a differential, and hol(·) de- notes the holomophic part. Conditions (i) ∼ (iii) ensures that t0,iti,j = t0,j. 38 XUNTAO HU AND CHAYA NORTON −e )(ze). e − I∗e ξ(0) In order to apply (3.3) and (3.8) to construct the A- normalized solution to the jump problem with this initial data, we need to show that ξ(0) e (ze) is holomorphic in ze. It is immediate because the Note that the initial data t0,i(bΞv − I∗ebΞv(−e))(ze) is equal to (ξ(0) pair of differentialsbΞv and ti,jbΞv(−e) have opposite residues at the node qe and the pull-back of the principal part by Ie is holomorphic. and (3.9): For k ≥ 1, we define for ze ∈ Ue : We recall here the construction of the A-normalized solution in (3.3) Kv(ze, we′) · I∗e′ξ(k−1) −e′ (we′); ξ(k) e (ze) := Xe′∈Ev we′∈γe′ v (z) := Xe∈Ev ze∈γe for z ∈ bCv : By Theorem 3.3, ηv := Pk≥1 η(k) η(k) Kv(z, ze) · I∗e ξ(k−1) −e (ze). v (z) is the A-normalized solution to the jump problem of higher order zeroes and poles. Similar to the proof of Theorem 3.3, we need to show that ηv is convergent by giving a L2-bound for the solution ηv. We can repeat the proof in Lemma 3.2 and Theorem 3.3 to get (3.10), which we recall as 1+ord ξ (0) (0) M ξ ηvL2 <ps (0) 1 for some constant M. We have shown above that ξ(0) e (ze) is holomorphic ξ(0) for every e, therefore ord ξ e ≥ 0. The only thing left (0) to show here is that ξ 1 is bounded by some power of t, in other words, the power of t in ξ(0) ±e is non-negative for any e. = mine ordqe Note that the power of t in ξ(0) need to show the same holds for ξ(0) principal part of bΞv(−e) through Ie, its lowest order term z−ke−2 −e is automatically non-negative, we only . Note that when pulling back the dz−e contributes a factor of s−ke−1 i,j by condition (ii). Since all other terms in the principal part contribute factors of lower powers of se, the power of t in ξ(0) e must be non-negative. We can thus apply the same argument as in the proof of Lemma 3.2 and Theorem 3.3 and achieve an L2-bound for ηv. , which is seen to be equal to t−1 LetbΞv,t := t0,ibΞv +ηv for any v at level i, then by the argument in the proof of Theorem 3.3, we have that {bΞv,t}v glues to a global differential bΞt on Ct such that limt→0(Ct,bΞt) = (C, Ω). Note that by adding the modification differential φv,j and the solution differential ηv to Ξv, the zeroes of multiplicity mi of Ξv at pi ∈ Cv are −e e e GENERAL VARIATIONAL FORMULAS FOR ABELIAN DIFFERENTIALS 39 broken into mi simple zeroes in a small neighborhood Ui of pi. The radius of the neighborhood is controlled by the norm of the added differentials. The modification differentials φv,j are multiples of ti,j, and the argument above gives the L2-bound on ηv. We can thus merge the zeroes of bΞt using the arguments in [BCGGM16, Lemma 4.7] to get the wanted degenerating family (C, Ξt) with ordpi Ξt = mi, and limt→0(Ct, Ξt) = (C, Ω). (cid:3) Although the differential Ξt depends on the choice of the modification v }, the existence of such a degenerating family does not differentials {φL rely on the choice of {φL Corollary 6.5. [BCGGM16, Sufficiency Part of Theorem 1.3] A stable inc g,n(µ) pointed differential (C, Ω, p1, . . . , pn) lies in the boundary of PΩM if there exist a twisted differential Ξ of type µ and a full level function l such that Ξ is compatible with Ω and l. v}. Theorem 6.4 in particular implies: Acknowledgements Both authors thank Samuel Grushevsky, the current advisor of the first author and the former advisor of the second author, for the useful discus- sions throughout the process. The second author would like to thank Marco Bertola and Igor Krichever for conversations related to this work. We thank Scott Wolpert for reading the first version of the paper carefully and his valu- able comments. We also thank the referees for their valuable advices which in specific helped us make the paper more accessible. Finally, we thank the organizers of the Cycles on Moduli Spaces, GIT, and Dynamics workshop in ICERM, Brown University, August 2016, where this collaboration started. References [BCGGM16] M. Bainbridge, D. Chen, Q. Gendron, S. Grushevsky and M. Moeller. preprint, strata of Abelian differentials." (2016): "Compactification of arXiv:1604.08834. [Ber06] M. Bertola. "Riemann Surfaces and Theta Functions." Concordia Univer- sity, 2006. [Fay73] J. Fay. "Theta Functions on Riemann Surfaces". Lecture Notes in Mathe- matics 352. Berlin-New York: Springer-Verlag, 1973. [Gen15] Q. Gendron "The Deligne-Mumford and the Incidence Variety Compacti- [Ger90] L. Gerritzen "The Torelli Map at the Boundary of the Schottky Space." J. fications of the Strata of ΩMg." Preprint, arXiv:1503.03338, 2015. reine angew. Math. 405(1990): 29-47. [Ger92a] L. Gerritzen "Theta Functions of Totally Degenerate Curves." Arch. Math 59(1992): 302-312. 40 XUNTAO HU AND CHAYA NORTON [Ger92b] L. Gerritzen "Equations Defining the Periods of Totally Degenerated Curves." Isreal Journal of Mathematics. 77(1992): 187-210. [GH] P. Griffith, J. Harris "Principles of Algebraic Geometry". Wiley Classics Li- brary 52, Wiley, 2011. [GKN17] S. Grushevsky, I. Krichever, and C. Norton. "Real-Normalized Differen- tials: Limits on Stable Curves." (2017): Preprint, arXiv:1703.07806. [Gun76] R. C. Gunning "Riemann Surfaces and Generalized Theta Functions." Springer-Verlag, 1976. [HM] J. Harris, I. Morrison "Moduli of Curves." Graduate Texts in Mathematics, 187. New York: Springer-Verlag, 1998. [Hu] X. Hu "Locus of Plane Quartics with A Hyperflex." Proceedings of the Amer- ican Mathematical Society 145 (2017), 1399-1413. [HK13] J. Hubbard, S. Koch "An Analytic Construction of the Deligne-Mumford Compactification of the Moduli Space of Curves." J. Differential Geom. 98, no.2(2014): 261-313. [IT] Y. Imayoshi, M. Taniguchi "An Introduction to Teichmuller Spaces." Tokyo: Springer-Verlag, 1992. [LZR13] Kefeng Liu, Quanting Zhao, and Sheng Rao. "New Proofs of the Torelli Theorems for Riemann Surfaces." (2013): Preprint, arXiv:1207.5697v2. [Ma76] H. Masur "The Extension of Weil-Petersson Metric to the Boundary of Teichmuller Space." Duke Mathematical Journal, 43 no.3(1976): 623-635. [Ro88] Yu. L. Rodin "The Riemann Boundary Problem on Riemann Surfaces." Mathematics and its Applications, Vol. 16, Springer, 1988. [Tan91] M. Taniguchi. "On the Singularity of the Periods of Abelian Differentials with Normal Behavior under Pinching Deformation." J. Math. Kyoto Univ., 31 no.4(1991): 1063 -- 1069. [Wol13] S. Wolpert "Infinitesimal Deformations of Nodal Stable Curves." Adv. Math., 244(2013): 413 -- 440. [Wol15] S. Wolpert "Schiffer Variations and Abelian Differentials." (2015): Preprint, arXiv:1508.01100. [Yam80] A. Yamada. "Precise Variational Formulas for Abelian Differentials." Ko- dai Math. J., 3 no.1(1980): 114 -- 143. Mathematics Department, Stony Brook University, Stony Brook, NY 11794-3651, USA E-mail address: [email protected] Concordia University, Montreal, QC, Canada, and Centre de Recherches Math´ematiques (CRM), Universit´e de Montr eal, QC, Canada E-mail address: [email protected]
1603.05104	1	1603	2016-03-16T13:58:33	Liftability of singularities and their Frobenius morphism modulo $p^2$	[ "math.AG" ]	We investigate the $W_2(k)$-liftability of singular schemes. We prove constructibility of the locus of $W_2(k)$-liftable schemes in a flat family $X \to S$. Moreover, we construct an explicit $W_2(k)$-lifting of a Frobenius split scheme $X$ over a perfect field $k$, reproving Bhatt's existential result. Furthermore, we study existence of liftings of the Frobenius morphism. In particular, we prove that in dimension $n \geq 4$ ordinary double points do not admit a $W_2(k)$-lifting compatible with Frobenius, and that canonical surface singularities are Frobenius liftable. Combined with Bhatt's results, the latter result implies that the crystalline cohomology groups over $k$ of surfaces with canonical singularities are not finite dimensional. As a corollary of our results, we provide a thorough comparison between the notions of $W_2(k)$-liftability, Frobenius liftability and classical $F$-singularity types.	math.AG	math	LIFTABILITY OF SINGULARITIES AND THEIR FROBENIUS MORPHISM MODULO p2 MACIEJ ZDANOWICZ Abstract. We investigate the W2(k)-liftability of singular schemes. We prove constructibility of the locus of W2(k)-liftable schemes in a flat family X → S. Moreover, we construct an explicit W2(k)-lifting of a Frobenius split scheme X over a perfect field k, reproving Bhatt's existential result. Furthermore, we study existence of liftings of the Frobenius morphism. In particular, we prove that in dimension n ≥ 4 ordinary double points do not admit a W2(k)-lifting compatible with Frobenius, and that canonical surface singularities are Frobenius liftable. Combined with Bhatt's results, the latter result implies that the crystalline cohomology groups over k of surfaces with canonical singularities are not finite dimensional. As a corollary of our results, we provide a thorough comparison between the notions of W2(k)-liftability, Frobenius liftability and classical F -singularity types. 1. Introduction In the seminal paper [DI87], Deligne and Illusie observed that a smooth variety X defined over a perfect field k of characteristic p possesses many desired properties, which are standard in characteristic 0, provided it admits a flat lifting to W2(k), i.e., a unique Z /p2-flat extension of k. This observation turned out to be very fruitful and leads to many further results overcoming pathological behaviour of positive characteristic geometry of smooth varieties (see, e.g., [OV07, Lan15]). The crucial part of the arguments given in the papers mentioned was the existence of a local lifting of the Frobenius morphism mod p2. In this paper, we investigate the existence of W2(k) and Frobenius liftings in the more general setting of singular schemes. In this situation, contrary to the case of smooth schemes, a W2(k)- lifting need not to exist even affine locally. The simplest instance of that phenomenon, in fact 0-dimensional, was given without a proof in [Bha12, Remark 3.16]. We present a detailed analysis of a slightly more general example in Section 3.2. Furthermore, in [Bha14] the author generalizes the results of [DI87] under the assumption of global Frobenius liftability (see Definition 1.1), and consequently proves that the crystalline coho- mology over k of Frobenius liftable locally complete intersection schemes is not finitely generated. The aforementioned examples and results suggest the following two basic question: Question 1: What are the necessary conditions for a scheme to possess a W2(k)-lifting? Question 2: For a given scheme, does there exist a W2(k)-lifting together with a compatible lifting of the Frobenius morphism? General answers to the questions were given in [Ill72, MS87, Jos07a] and stated that above problems can be expressed in purely cohomological manner. Moreover, in [LS14] the authors investigated the first question in the setting of birational geometry. Our contribution to answering the above questions is the following. As far as the first question is concerned, in Section 3.3 we analyse how the property of W2(k)- liftability behaves in families f : X → S. More precisely, we show, that under mild assumptions suitable for deformation-theoretic considerations (satisfied if e.g. f is proper or affine), the locus = f −1(s) is constructible (see Theorem 3.11). Furthermore, we of Witt vector liftable fibres Xs reprove the result that Frobenius split schemes are W2(k)-liftable. The advantage of our approach is that we provide the lifting explicitly and functorially with respect to a splitting ϕ : FX∗OX → OX , in fact as a certain subscheme of the Witt scheme (X, W2(OX )). The details of our construction are presented in Section 3.4. def The author was supported by Polish National Science Centre (NCN) contract number 2014/13/N/ST1/02673. 1 2 MACIEJ ZDANOWICZ Moreover, we give a new construction of a non-liftable affine scheme. In fact, the construction is given by a cone over a suitable embedding on a non-liftable projective scheme (for the proof see Lemma A.21 in the Appendix A). Examples of non-liftable projective schemes are given, e.g., by Raynaud [Ray78] and Mukai [Muk13]. Concerning the second question, we derive a computationally feasible criterion for Frobenius liftability of complete intersection affine schemes (see Theorem 4.6). Consequently, we apply the criterion in the case of ordinary double points (see Theorem 4.10), and thus answer the question asked by Bhatt in [Bha14, Remark 3.14]. As another application of our criterion, in Section 4.4 we prove that canonical surface singularities are Frobenius liftable. Finally, in Section 4.4.2 we present a general approach for proving Frobenius liftability of tame reductions of quotient singularities. The method is based on the spreading out technique combined with functoriality of obstruction classes for lifting morphisms. The above results allow us to address the following question: Question 3: What is the relation between W2(k)-liftability and Frobenius liftability, and classical notions defined in terms of Frobenius action? Our contribution can be summarized by the following diagram which describes all potential relations between the classical F -singularities and the notion considered in this paper: F -liftable \Ex. 5.7 F -regular F -rational \Ex. 5.8 \Ex. 5.7 \ Thm 4.10 \Thm 4.10 yes : normal k-algebra - Thm 5.5 no : non-normal - Ex. 5.6 F -pure Thm 3.18 W2(k)-liftable. The counterexamples corresponding to most of the strikethrough implications in the diagram are provided in Section 5 As a byproduct of our considerations, we formalise the foregoing functoriality properties of obstruction classes for deformation of schemes and their morphism. In particular, we analyse their behaviour in the case of affine schemes. The results are given in Appendix A. Apart from explicit statements and proofs concerning functoriality, the whole theory can be found in the standard reference [Ill72]. In the course of our consideration we faced the following problem: Open question: Does there exist a finite ´etale covering π : Y → X of schemes locally of finite type over k such that Y is W2(k)-liftable and X is not. Note that by Lemma 3.1 the degree of π would necessarily be divisible by p and by Lemma A.11 the schemes need not be affine. 1.1. Notation and basic definitions. Here we present a few general assumptions on schemes and their morphisms. Moreover, we give basic definitions concerning Frobenius morphism and Witt vector liftability. 1.1.1. Generalities on schemes and morphisms. Throughout the following work k is a perfect field of characteristic p > 0. All rings we consider are commutative, noetherian and with identity. Unless otherwise stated the schemes and morphisms between them are quasi-compact, quasi-separated and essentially of finite type over k (in particular, the schemes are noetherian). For any morphism f : X → Y , the associated mapping of sheaves of rings is denoted by f # : OY → f∗OX . For any k-scheme X by FX : X → X we mean the absolute Frobenius morphism, i.e., the identity on the level of topological spaces and F # X (f ) = f p. Moreover, for any morphism f : X → S by X : OX → FX∗OX defined by F # FX/S : X → X (S) def = X ×FS S LIFTABILITY OF SINGULARITIES AND THEIR FROBENIUS MORPHISM MODULO p2 3 we denote the relative Frobenius morphism defined by the following diagram with a cartesian right square: FX FX/S WX/S X (S) X f '❖❖❖❖❖❖❖❖❖❖❖❖❖❖ f (S) S FS / X f / S. By Wn(k) we denote the ring of Witt vectors of length n and by σ : Wn(k) → Wn(k) the associated Frobenius morphism given by the formula (a0, . . . , an−1) 7→ (ap 1.1.2. Frobenius splitting. By a Frobenius splitting of X we mean a homomorphism ϕ : FX∗OX → OX of OX -modules which splits the canonical map F # X : OX → FX∗OX . We say that a scheme X is Frobenius-split if there exists an associated Frobenius splitting. For an exhaustive treatment of Frobenius splittings the reader is referred to [BK05]. 0, . . . , ap n−1) a closed immersion given by a square-zero ideal. 1.1.3. Lifting of schemes and Frobenius morphism. Let X/S be a morphism of schemes and S → eS Definition 1.1. An eS-lifting of X/S is a flat morphism eX → eS together with an isomorphism eX ×eS S ≃ X. We say that X/S admits an eS-lifting or is eS-liftable if there exists a lifting eX/eS as described eX/W2(k). Moreover, we say that it lifts compatibly with Frobenius or is Frobenius liftable (abbr. F -liftable) if there exists a lifting eX of X together with a morphism eF : eX → eX over the Frobenius of W2(k) which restricts to FX : X → X, i.e., such that the following diagram is commutative: In particular, we say that X/k lifts to W2(k) or is W2(k)-liftable if it admits a lifting above. X i w♦♦♦♦♦♦♦♦♦♦♦♦♦♦ eF eX Spec(k) h x♣♣♣♣♣♣♣♣♣♣♣ Spec(W2(k)) σ FX eX Fk X i w♦♦♦♦♦♦♦♦♦♦♦♦♦♦ / Spec(k) h w♣♣♣♣♣♣♣♣♣♣♣ / Spec(W2(k)). For a perfect field k, we can reformulate Frobenius liftability in terms of the relative Frobenius morphism FX/k : X → X ×k,Fk Spec(k) and its lifting eFX/k : eX → eX ×W2(k),σ Spec(W2(k)). with liftings of their associated k-algebras (resp. their p-th power mappings). In the affine case, we sometimes identify liftings of k-schemes (resp. their Frobenius morphisms) 1.1.4. Derived categories and derived functors. Throughout the paper by DQCoh(X) we denote the derived category of quasi-coherent sheaves on a scheme X. We add + (respectively −) in the superscript in order to indicate that we consider the category of bounded below (respectively above) complexes. Moreover, by D− QCoh(X) consisting of complexes whose cohomology sheaves are coherent. Again, based on our assumptions, one can prove that any object in D− Coh(X) we denote the full subcategory of D− Coh(X) is quasi-isomorphic to a complex of coherent sheaves. 1.2. Structure of the paper. The paper is organized as follows. In Section 2 we give a few preliminary results concerning commutative algebra and homological algebra. In Section 3 we investigate W2(k)-liftability of schemes. In particular, we prove that under suitable assumptions Witt vector liftability is a constructible property (see Theorem 3.11) and give a construction of a W2(k)-lifting of a Frobenius split scheme (see Theorem 3.16). & & / / ' / / / / w w / / x / w / 4 MACIEJ ZDANOWICZ In Section 4 we address the problem of Frobenius liftability of singularities. We derive a compu- tational criterion for Frobenius liftability of complete intersection schemes (see Theorem 4.6) and consequently apply it to ordinary double points (see Theorem 4.10) and canonical surface singular- ities (see Theorem 4.12). Section 5 is devoted to comparing W2(k)-liftability and F -liftability with classical notions of F -regularity, F -rationality and F -purity. In Section 6 we recall the classical approach to deformation of cones with an emphasis on W2(k)-liftability. Moreover, we extend the standard results with a study of Frobenius liftability. Due to their technical nature deformation theory considerations are deferred to Appendix A. We recall a few results concerning deformation functors and present a thorough description of functoriality of obstruction classes expressed in terms of cotangent complex. Moreover, we present a handful of remarks concerning deformations of affine schemes. 2. Preliminaries 2.1. Basic commutative algebra, flatness criteria. We recall basic facts concerning flatness of W2(k)-modules and algebras. Under standard bijection W2(k) ≃ k2 the element p ∈ W2(k) corresponds to (0, 1) ∈ k2. Lemma 2.1. A W2(k)-module M is flat if and only if the annihilator AnnM (p) = {m ∈ M : pm = 0} of p in M is equal to pM , i.e., the natural mapping M/pM → pM given by [m] 7→ pm is injective. Proof. We firstly prove that pW2(k) ⊗W2(k) M is isomorphic to M/pM . This follows from right exactness of the functor − ⊗ M and a resolution of pW2(k): / pW2(k) / W2(k) W2(k) / 0. p· p· By a well-known criterion stating that M is flat over W2(k) if and only if − ⊗W2(k) M preserves any injection of ideal of W2(k) (see [AM69], Exercise 2.24), the flatness of M is now equivalent to the injectivity of the multiplication by p mapping pW2(k) ⊗W2(k) M → M . By the isomorphism above, this mapping is the same as mp : M/pM → M given by [m] 7→ pm. The kernel of mp is exactly AnnM (p)/pM ⊂ M/pM and therefore injectivity is equivalent to the hypothesis of the lemma. (cid:3) As a simple corollary we obtain: Corollary 2.2. A ring B/W2(k) is a flat lifting of A/k if and only if B/pB ≃ A and pB = AnnB(p). Moreover, for any flat lifting B/W2(k) of A the quotient B/J is a flat lifting of A/I if and only if J is a lifting of I and (p) ∩ J = pJ. Proof. The first part follows directly from Lemma 2.1. To prove the second part, we apply the following sequence of equivalences: (1) AnnB/J p = p · B/J ⇔ (J : (p)) = (J + p) ⇔ p(J : (p)) = p(J + p) ⇔ (p) ∩ (J) = pJ, where the middle one follows from the inclusions (J : (p)) ⊃ (p) ⊂ (J + p) and the injectivity of the mapping B/p → pB. Lemma 2.3. Suppose B/W2(k) is a flat lifting of B/p. Then, any lift f ∈ B of a non-zerodivisor f ∈ B/pB is a non-zerodivisor. Moreover, B/f B is a W2(k)-lifting of B/(p, f )B. Proof. Assume that g ∈ B satisfies the equality gf = 0. By reducing mod p we see that g · f = 0 and therefore g = ph for some h ∈ B. Again, using the assumption and the isomorphism B/pB ≃ pB (see Lemma 2.1), we see that h · f = 0 and consequently h = ph′. This implies that g = p2h′ = 0 yielding the first part of the lemma. For the second part, we consider the flat resolution of B/f B given by 0 −→ B −→ B −→ B/f B −→ 0. By reducing mod p we see that TorW2(k) (B/f B, k) = 0 and therefore B/f B is W2(k)-flat (see [Sta15, Tag 051C]). Alternatively, we can prove that (f ) ∩ (p) = (pf ) and use the above Corollary 2.2. (cid:3) (cid:3) 1 Consequently we get the following result. / / / LIFTABILITY OF SINGULARITIES AND THEIR FROBENIUS MORPHISM MODULO p2 5 Lemma 2.4. Let I = (f1, . . . , fm) be an ideal in k[x1, . . . , xn]. Every W2(k)-lifting of k[x1, . . . , xn]/I is given by W2(k)[x1, . . . , xn]/J, where J is generated a sequence of lifts of {fi}. Conversely, if (f1, . . . , fm) is a regular sequence then for every ideal J generated by the sequence of lifts of {fi} the ring W2(k)[x1, . . . , xn]/J is a W2(k)-lifting. Proof. For the first part see [Har09, Theorem 10.1]. For the converse, apply the above Lemma 2.3, inductively. (cid:3) 2.2. Base change for higher direct images. We now present a few results concerning derived inverse images and base change for unbounded complexes. Definition 2.5 (Morphism of finite Tor dimension). We say that a morphism of schemes f : X → Y is of finite Tor dimension if OX is a f −1OY -module of finite Tor dimension. For any morphism f : X → Y of finite Tor dimension the structural sheaf OX is in fact quasi-isomorphic to a bounded complex of flat f −1OY -modules (see [Sta15, Tag 08CI]). Remark 2.6. In the case of morphisms of quasi-projective schemes over a field by [FM81, II.1.2] we know that the following morphisms are of finite Tor dimension: a) morphism with a smooth target, b) flat morphisms, c) regular and Koszul closed immersions. Due to the lack of an adequate reference, we give a detailed proof of the following lemma. Lemma 2.7 (Spectral sequence for derived pullback). Suppose f : X → Y is a morphism of finite Tor dimension. Then, for any F • ∈ D(OX ) there exists a spectral sequence: Epq 2 = Lpf ∗ (Hq(F •)) ⇒ Lp+qf ∗F •. Proof. This is an application of the spectral sequence of a double complex. Indeed, by [Sta15, Tag 08DE] we see that the total derived pullback Lf ∗F • is given by the totalisation of the double complex f −1F • ⊗f −1OY E • where E • is a bounded f −1OY -flat resolution of OX existing by the assumption on f . By the classical result there exist a spectral sequence: Epq 2 = HpHq(f −1F • ⊗f −1OY E •) ⇒ Hp+qTot(f −1F • ⊗f −1OY E •), whose convergence follows from the fact that f −1F • ⊗f −1OY E • is a double complex supported in the bounded horizontal strip {(i, j) : j ∈ {k : E k 6= 0}}. The sheaves Hp+qTot(f −1F • ⊗f −1OY E •) are easily identified with Lp+qf ∗F •. Moreover, by the exactness of the functor f −1 and flatness of E • (in fact, we use [Sta15, Tag 08DE] again), we see that HpHq(f −1F • ⊗f −1OY E •) ≃ Hp(f −1Hq(F •) ⊗f −1OY E •) ≃ Lpf ∗Hq(F •). This gives the desired result. Lemma 2.8. Let f : X → S be a locally of finite type morphism of schemes, and E a complex in DQCoh(X). Then for every cartesian diagram: (cid:3) XT if / X fT f T i / S, the natural base change φ : Li∗Rf∗E ∼−→ RfT ∗Li∗ Proof. For the proof we refer to [Sta15, Tag 0A1D]. Lemma 2.9. Suppose f : X → Y is a morphism of schemes. Then, for every F • ∈ D− there exists a natural isomorphism Lf ∗RHom(F •,OY ) ≃ RHom(Lf ∗F •,OX ). Proof. By [Sta15, Tag 08I3] we obtain a mapping: fE is an isomorphism. (cid:3) Coh(Y ) Lf ∗RHom(F •,OY ) → RHom(Lf ∗F •,OX ). In order to prove that it is an isomorphism we can work locally and therefore we may assume that LX/S is resolved as a complex of locally free sheaves. In this case, the result follows from the / / 6 MACIEJ ZDANOWICZ definition of Lf ∗ and the fact that RHom(−,OX) might be computed by a locally free resolution of the first argument. (cid:3) 2.3. A primer in deformation theory. Due to unoriginal and tediously technical nature of the considerations given in the corresponding section we defer it to Appendix A. The reader familiar with 1) obstruction theories defined in terms of cotangent complex, 2) the notion of deformation functor and its basic properties, may comfortably proceed with the actual content of the paper. In the upcoming considerations we freely use the above notions by referring to Appendix A.1 and Appendix A.3, respectively. 3. W2(k)-liftability In this section we begin the main considerations of this work. Firstly, we present general facts concerning W2(k)-liftability and give a few examples. Subsequently, we prove that the locus of W2(k)-liftable schemes in a flat family X → S is constructible. Finally, we reprove constructively the classical result (see Proposition 3.2) that any Frobenius-split variety admits a W2(k)-lifting. We begin with functoriality of obstructions to existence of W2(k)-liftings. 3.1. General results on obstruction to liftability. Lemma 3.1. For any scheme X/k there exists an obstruction σX ∈ Ext2(LX/k,OX ) whose van- ishing is sufficient and necessary for existence of W2(k)-lifting. The obstruction is functorial, i.e., for any k-scheme morphism g : X → Y the following diagrams in DQCoh(X) and DQCoh(Y ) are commutative: Lg∗LY /k Lg∗σY/ dg LX/k Lg∗OY [2] ≃ σX / OX [2], σY LY /k dg OY [2] g#[2] Rg∗LX/k Rg∗σX/ / Rg∗OX [2]. In particular W2(k)-liftability descends along finite surjective maps of degree prime to the charac- teristic of k. Proof. This is a direct consequence of Lemma A.6. The final remark follows from existence of the trace maps splitting g#. (cid:3) As corollary we obtain a well-known result that every Frobenius split variety lifts. Proposition 3.2. If X/k is a Frobenius split scheme then X (1)/k lifts to W2(k). Proof. See [Ill96, p. 164] or [Jos07b, Corollary 9.2] for the case of smooth schemes. We reproduce general proof by Bhargav Bhatt given in [Lan15, Proposition 8.4]. The idea is to use the functo- riality of obstructions for the relative Frobenius FX/k : X → X (1). Namely, we have the following commutative diagram: L X (1)/k dFX/k σ X(1) OX (1) [2] FX/k∗LX/k FX/k∗σX / FX/k∗OX [2]. The differential dFX/k = 0 and therefore by existence of splitting σX (1) = 0. Note that in case of varieties over a perfect field the W2(k)-liftability of X (1) is in fact equivalent to the liftability of X. (cid:3) 3.2. Example of non W2(k)-liftable scheme, p-neighbourhoods of smooth quadrics. We proceed to an example of 0-dimensional scheme which is not W2(k)-liftable. In this case we give a direct computational proof which is afterwards used in the considerations concerning Frobenius liftability of ordinary double points. We shall prove that the Artinian local k-algebras: (A2n, m2n) =(cid:0)k[x1, . . . , x2n](cid:14)(x1x2 + . . . + x2n−1x2n, xp (A2n+1, m2n+1) =(cid:0)k[x1, . . . , x2n+1](cid:14)(x1x2 + . . . + x2n−1x2n + x2 1 , . . . , xp 2n) , (x1, . . . , x2n)(cid:1) 1, . . . , xp 2n+1, xp 2n+1) , (x1, . . . , x2n+1(cid:1) / / / / / / / Z Z LIFTABILITY OF SINGULARITIES AND THEIR FROBENIUS MORPHISM MODULO p2 7 are non-liftable to W2(k) for n ≥ 3. Moreover, as suggested in [BO78][page 3.3], their maximal ideals do not admit PD-structure. 3.2.1. Non-liftability to W2(k). Proposition 3.3. For any N ≥ 5 the local algebra AN does not admit a W2(k)-lifting. Proof. Let f2n denote the polynomial x1x2 + . . . + x2n−1x2n and f2n+1 the polynomial x1x2 + . . . + x2n−1x2n + x2 2n+1. We begin with the case of even number of variables N = 2n. By Lemma 2.4, 2n of f2n, xp 1, . . . , xp 2n, the ring is not flat over W2(k). By means of Lemma 2.3 we see that it suffices to show that for any choice of liftings ff2n,fxp eA = W2[x1, . . . , x2n]/(ff2n,fxp eC = W2[x1, . . . , x2n]/(fxp = eC ⊗W2(k) k = k[x1, . . . , x2n]/(xp 1, . . . ,gxp 1, . . . ,gxp 1, . . . ,gxp is a flat W2(k)-lifting of C 2n) def 1, . . . , xp 2n) applied to the commutative diagram: show that the canonical homomorphism AnneC (ff2n) → AnnC (f2n) coming from the snake lemma 2n). Consequently it suffices to is not surjective, i.e., there exists an element h in the annihilator of f2n ∈ C which does not lift is the right choice. For the p to an element in the annihilator of ff2n in eC. We claim that h = f p−1 sake of contradiction, suppose that there exists a lifting ff2n + pgf2n = 0 ∈ eC. By direct computation we see that ff2n ff2n = (x1x2 + . . . + x2n−1x2n + pu)p = (x1x2 + . . . + x2n−1x2n)p 2 + . . . + xp 2n−1xp = xp 1xp p−1 2n p + pg such that (ff2n p−1 + pg)ff2n = 2n + pPf2n = pPf2n for Pf2n defined as i1,...,in6=p (p − 1)! i1!··· in! (x1x2)i1 ··· (x2n−1x2n)in ∈ eC(cid:14)peC ≃ k[x1, . . . , x2n]/(xp Pf2n = Xi1+...+in=p Therefore, we obtain that p(Pf2n + gf2n) = 0 in eC, i.e., by Corollary 2.2 the polynomial Pf2n belongs to the ideal (f2n, xp 2n). By specializing xi = 0 for i > 6 we see that it suffices to derive the contradiction for n = 3. We prove that even the element (x5x6)p−2Pf6 does not belong 1, . . . , xp 1, . . . , xp 2n). AnneC (ff2n) eC ·ff2n eC / eA ·p ·p ·p C ·f2n C A AnnC (f2n) / 0 / 0 C eC / A, / / / / / / / / / / / / / / (−1)ii(x3x4)i−1(x3x4)j + (x5x6)p−1 Xi+j=p−1 (p − 1)! k!j! (−1)k +(x3x4x5x6)p−1 Xi+j=p−1 (p − 1)! i!j! (−1)i (p − 1)! i!j! (−1)i(x3x4)i+j } 6). This follows from the (mod xp 1, . . . , xp 6) (mod xp 1, . . . , xp 6) (mod xp 1, . . . , xp 6). 8 MACIEJ ZDANOWICZ to (f3, xp 1, . . . , xp 6). We begin with the computation: (p − 1)! i!j!k! (x1x2)i(x3x4)j (x5x6)k (−1)i(x3x4 + x5x6)i(x3x4)j (p − 1)! i!j! (−1)i(x3x4 + x5x6)i(x3x4)j (p − 1)! i!j! i,j6=p i,j,k6=p (x5x6)p−2Pf6 = (x5x6)p−2 Xi+j+k=p = (x5x6)p−2 Xi+j=p + (x5x6)p−1 Xi+j=p−1 = (x5x6)p−1 Xi+j=p = (x3x4x5x6)p−1 Xk+j=p−1 i,j6=p i!j! (p − 1)! k6=p−1 Therefore, we are left to show that (x3x4x5x6)p−1 6∈ (f6, xp following: = −(x3x4x5x6)p−1 (mod f6, xp 1, . . . , xp 6) } =−1 {z (x1x2 + x3x4 + x5x6)Xi,j 2  +Xi,j Xi,j Xi,j As k[x1, . . . , x6].(xp gi,j · xi+1 1, . . . , xp 1 xj+1 we may compare coefficients to obtain: =0 1, . . . , xp {z 2 = (x3x4x5x6)p−1 2 = (x3x4x5x6)p−1 2 = (x3x4x5x6)p−1 6) is a free k[x3, . . . , x6].(xp (gi−1,j−1 + gi,j(x3x4 + x5x6)) · xi 1xj gi,j(x3x4 + x5x6) · xi 1xj gi,j · xi 1xj 3, . . . , xp 6) -module on generators xi 1xj 2 g0,0(x3x4 + x5x6) = (x3x4x5x6)p−1 gi,i = −gi+1,i+1(x3x4 + x5x6) After using the second relation inductively we obtain: (mod xp (mod xp 3, . . . , xp 6) 3, . . . , xp 6). 0 = gp−1,p−1(x3x4 + x5x6)p = gp−1,p−1(x3x4 + x5x6)p−1(x3x4 + x5x6) = g0,0(x3x4 + x5x6) = (x3x4x5x6)p−1 (mod xp 3, . . . , xp 6), which is a contradiction. (cid:3) The proof for odd number N = 2n + 1 ≥ 7 of variables is analogously reduced to the case of 6 variables by additional specialisation x2n+1 = 0. The case N = 5 can be treated easily by an explicit computation following the lines of the above. Remark 3.4. We have implemented Macaulay2 procedure checking whether a give affine scheme is W2(k)-liftable. The code is given at . 3.2.2. Non-existence of a PD-structure on (A, m). We now prove that the ideal m does not admit PD-structure in spite of satisfying the necessary condition m[p] = 0. We precede the actual result with a necessary definition. Definition 3.5 (PD-structure). Let A be a commutative ring and let I be an ideal of A. A divided power structure is a collection of mappings γi : I → A for i ∈ N satisfying the following properties: i) γ0(x) = 1, γ1(x) = x for any x ∈ I and γn(x) ∈ I for any n ≥ 1, ii) γn(x)γm(x) = (n+m)! n!m! γn+m(x), LIFTABILITY OF SINGULARITIES AND THEIR FROBENIUS MORPHISM MODULO p2 9 (m!)nn! γnm(x), iii) γn(γm(x)) = (nm)! iv) γn(ax) = anγn(x) for any x ∈ I and a ∈ A, v) γn(x + y) =Pi+j=n γi(x)γj (y). The above notion is a necessary tool for dealing with de Rham and crystalline cohomology in characteristic p. Using ii) we see that p!γp(x) = xp, which indeed implies that any ideal I of an Fp-algebra A admitting a PD-structure satisfies the condition I [p] = 0. Proposition 3.6. For N ≥ 5 the local algebras (AN , mN ) defined above do not admit PD- structures on the maximal ideals mN . Proof. We focus on the case of even N = 2n. For the sake of contradiction, we assume an appropriate system of mappings γi exists and analyse the element γp(f2n) = γp(0) = 0. in! 1 = (p − 1)! Pf2n i1,...,in6=p i1! ··· γi1 (x1x2)··· γin (x2n−1x2n) (x1x2)i1 (x2n−1x2n)in 0 = γp(f2n) = Xi1+...+in=p = Xi1+...+in=p This means that Pf2n ∈ (f2n, xp 3.3. W2(k)-liftability in families. We now proceed to the proof that under suitable conditions Witt vector liftability is a constructible property, i.e., for a locally finite type family f : X → S the set of closed points s ∈ S such that the fibre Xs is W2(k(s))-liftable is constructible. We begin with a definition encompassing the properties of a morphism necessary to prove constructibility. Definition 3.7. We say that a morphism f : X → S is strongly equicohomological if it is locally of finite type and flat, and for any i ∈ {0, . . . , dim S+2} the higher direct image Rif∗RHom(LX/S,OX ) is a flat OS-module. 2n) which gives a contradiction as proven above. (mod f2n, xp 1, . . . , xp 2n). 1, . . . , xp (cid:3) Strongly equicohomological families satisfy the following crucial property motivated by the application for deformation theory. Lemma 3.8. Let f : X → S be a strongly equicohomological morphism with a smooth target S. For any morphism T → S the sheaf R2f∗RHom(LX/S,OX ) satisfies base change property, i.e., for any cartesian diagram: j i XT fT T / X f / S. there exists a natural isomorphism i∗R2f∗RHom(LX/S,OX ) ≃ R2fT ∗RHom(LXT /T ,OXT ). Proof. By Lemma 2.8 there exists a natural quasi-isomorphism: (2) Li∗Rf∗RHom(LX/S,OX ) → RfT ∗Lj∗RHom(LX/S,OX ). By Lemma 2.9 and the base change property of cotangent complex we obtain an isomorphism: Lj∗RHom(LX/S,OX ) ≃ RHom(Lj∗LX/S, Lj∗OX ) ≃ RHom(LXT /T ,OXT ), which implies that the right hand side of (2) is isomorphic to RfT ∗RHom(LXT /T ,OXT ). By taking cohomology, for any k ∈ Z we obtain an isomorphism: Lki∗Rf∗RHom(LX/S,OX ) → RkfT ∗RHom(LXT /T ,OXT ). By Lemma 2.7 there exists a convergent spectral sequence: Epq 2 = Lpi∗Hq(Rf∗RHom(LX/S,OX )) ⇒ Lp+qi∗Rf∗RHom(LX/S,OX ). The terms on E2 page are isomorphic to Lpi∗Rqf∗RHom(LX/S,OX )) and therefore by the as- sumption on being strongly equicohomological Epq 2 = 0 for (p, q) ∈ Z<0 ×{0, . . . , dim S + 2} and / / 10 MACIEJ ZDANOWICZ p > 0. Moreover, by the smoothness of S we see that Epq 2 = 0 for p < − dim S and con- for any p ∈ Z and r ≥ 2. This means that the natural filtra- sequently E−p,p+2 tion induced on L2i∗Rf∗RHom(LX/S,OX ) by the spectral sequence consists of a single term i∗R2f∗RHom(LX/S,OX ). This finishes the proof. ≃ E−p,p+2 r+1 (cid:3) r As a corollary we obtain: Lemma 3.9. Let f : X → S be a strongly equicohomological morphism with a smooth affine target. Then, the set of closed points s ∈ S such that Xs is W2(k(s))-liftable is closed. Proof. By the assumptions on S we see that there exists a W2(k)-lifting eS of S. Therefore, there is a relative obstruction class σf ∈ Ext2(LX/S, f ∗(pOeS)) = Ext2(LX/S ,OX ) which vanishes if and only if there exists a flat eS-scheme eX fitting into the cartesian diagram: /❴❴❴❴❴❴❴ X f S eX eS Spec(k) / Spec(W2(k)). By the formal smoothness of eS → Spec(W2(k)), any point s ∈ S can be lifted to a mapping ei : Spec(W2(k(s)) → eS. We denote by is : Xs → X the closed immersion defined by the diagram (with a cartesian square): Xs fs is z✉✉✉✉✉✉✉✉✉✉ z✈✈✈✈✈✈✈✈✈✈ i X f S Spec(k(s)) Spec(W2(k(s)) w♥♥♥♥♥♥♥♥♥♥♥♥♥♥ ei / eS, R2f∗RHom(LX/S,OX ), and by σs ∈ Ext2(LXs/s,OXs ) the associated obstruction class to lifting Xs/k(s) to W2(k(s)). The Γ(S,OS)-module Ext2(LX/S,OX ) is in fact a set of global sections of the sheaf and therefore by Lemma 3.8 we obtain a specialisation isomorphism: ψs : Ext2(LX/S,OX ) ⊗OS k(s) = i∗R2f∗RHom(LX/S,OX ) ≃ R2fs∗RHom(LXs/k(s),OXs ) = Ext2(LXs/k(s),OXs). By Lemma A.6 we see that obstructions σf ∈ Ext2(LX/S ,OX ) and σs ∈ Ext2(LXs/k(s),OXs) fit into commutative diagram: Li∗ s LX/S i∗ s σf Li∗ sOX dis LXs/k(s) σs / OXs , and therefore the fibre ψs([σf ]) of an obstruction class σf ∈ Ext2(LX/S,OX ) is equal to the obstruction σs ∈ Ext2(LXs/s,OXs). Note that here we implicitly use the fact that the second part of the isomorphism (2) in Lemma 3.8 is given by the differential dj. Constructibility then follows from the fact that the zero set of a section of a flat module is constructible (in fact it is closed for a finitely generated module). (cid:3) / ✤ ✤ ✤ / / / z / / z w / / / / LIFTABILITY OF SINGULARITIES AND THEIR FROBENIUS MORPHISM MODULO p2 11 satisfies P . In order to obtain a general statement for morphism which are not strongly equicohomological we need the following lemma preceded with a notational remark. We say that a morphism f : X → S can be stratified into morphisms satisfying property P if there exists a stratification of the target Si Si = S into locally closed subschemes Si ⊂ S such that the base change fi : X ×S Si → Si Lemma 3.10. The following classes of morphism of schemes can be stratified into strongly equico- homological morphisms: i) proper morphisms, ii) affine morphisms of finite type. Proof. We observe that stratification is a purely topological notion and therefore we may assume that S is integral. Now, it suffices to prove that for the classes in question there exists an open subset U ∈ S such that fU : X ×S U → U is strongly equicohomological. In both cases we apply the generic flatness to prove a more general statement that for any complex E • ∈ D+ Coh(X) the higher direct images are generically S-flat. Firstly, we deal with affine morphisms. In this case, the higher direct images of any complex E • ∈ D+ Coh(X) are computed by taking the pushforwards of the cohomology groups Hj(E •). Those are coherent sheaves on X and are therefore generically S-flat. Secondly, we treat the case of proper morphism. Now, higher direct images Rif∗E • are coherent S-modules and therefore are generically flat. (cid:3) We are now ready to state and prove the theorem. Theorem 3.11. Suppose f : X → S is a morphism which can be stratified into strongly equicoho- mological morphisms. Then, the set of closed points s ∈ S such that the fiber Xs = X×S Spec(k(s)) lifts to W2(k(s)) is constructible. Proof. Observe that in order to prove constructibility we may stratify the scheme S into locally closed subsets. Therefore, by the assumption on existence of strongly equicohomological stratifi- cation, we may assume that : X → S is strongly equicohomological with a smooth affine target. Then, the result follows from Lemma 3.9. (cid:3) def As a direct corollary of Lemma 3.10 and Theorem 3.11 we obtain: Corollary 3.12. For any proper or affine morphism f : X → S the set of closed points s ∈ S such that the fibre Xs lifts to W2(k(s)) is constructible. 3.4. Functorial W2(k)-lifting for Frobenius-split varieties. We now reprove the classical re- sult stating that any Frobenius-split scheme X over k is W2(k)-liftable. For the standard reference and Bhargav Bhatt's proof in singular case the reader is encouraged to see Proposition 3.2. The advantage of our approach is that we provide the lifting constructively and functorially with re- spect to the splitting ϕ : FX∗OX → OX , in fact as a certain subscheme of the Witt scheme (X, W2(OX )). 3.4.1. Functorial setting. We shall introduce our construction as a certain functor from the cate- gory described as follows. By Frobenius split scheme we mean a pair (X, ϕX ) of a scheme over k and a Frobenius splitting ϕX : F∗OX → OX . Moreover, a morphism of Frobenius split schemes (X, ϕX ) and (Y, ϕY ) is a scheme morphism π : X → Y such that the splittings satisfy compatibility conditions expressed by commutativity of the diagram: OY π# π∗OX ϕY F # π∗ϕX π∗F # F∗OY F∗π∗OX = π∗F∗OX , i.e., the relation π# ◦ ϕY = (F∗π#)◦ (π∗ϕX ) holds. In case of an affine morphism corresponding to a homomorphism f : A → B this is equivalent to the equality f ϕA = ϕBf . The notions described above allow us to introduce the category of Frobenius split varieties. Definition 3.13. We define a category Schsplit of Frobenius split schemes over k as a category with the set of objects consisting of all Frobenius split schemes (X, φX ) and with the set of morphisms consisting of morphism of Frobenius split schemes f : (X, φX ) → (Y, φY ). k / / / / O O O O 12 MACIEJ ZDANOWICZ 3.4.2. Witt vectors. We shall need the following explicit description of Witt vectors. Definition 3.14 (Witt vectors W2(A)). Suppose A is a commutative ring. We define the ring of Witt vectors W2(A) to be the set A × A equipped with addition +W and multiplication ·W given by the following formulas: (3) (4) where P (a, b) is a polynomial (a+b)p−ap−bp (a0, a1) +W (b0, b1) = (a0 + b0, a1 + b1 − P (a0, b0)) (a0, a1) ·W (b0, b1) = (a0b0, ap 0a1 + pa1b1), ∈ Z[a, b]. 0b1 + bp p The unit element of W2(A) is represented by (1, 0) and, in case of Fp-rings, the prime number p is represented by (0, 1). The natural projection (a0, a1) 7→ a0 gives a ring homomorphism π : W2(A) → A. In case of characteristic p ring A, the ring W2(A) possesses a Frobenius endomorphism σ2 : W2(A) → W2(A) given by the formula (a0, a1) 7→ (ap 1), which is compatible with the Frobenius endomorphism F : A → A, i.e., the identity F π = πσ2 holds. 3.4.3. The construction of W ϕ 2 (A). We begin our considerations of explicit liftings of Frobenius split schemes by the affine case. A simple consequence of Corollary 2.2 is that a functorial construc- tion A 7→ W2(A) does not give a flat lifting (the ideal pW2(A) is not equal to Ker[W2(A) → A]). We therefore proceed in a different manner taking the Frobenius splitting into account. Let (A, ϕ) by a k-algebra together with a Frobenius splitting ϕ. We claim that a flat lifting of Spec(A) over W2(k) is given by the following construction functorial with respect to natural mappings of Fp rings with Frobenius splitting, i.e., ring homomorphisms commuting with p−1-linear splitting operators. 0, ap Definition 3.15. The ring of twisted Witt vectors W ϕ 2 (A) associated to a characteristic p ring with a splitting (A, ϕ) is a set A × A with addition and multiplication given respectively by the formulas: def (5) (a0, a1) +ϕ (b0, b1) (a0, a1) ·ϕ (b0, b1) 1) +W (b0, bp 1) ·W (b0, bp which can also be described by the following diagrams: = (id×ϕ)[(a0, ap = (id×ϕ)[(a0, ap (6) def 1)] = (a0 + b0, a1 + b1 − (ϕ ◦ P )(a0, b0)) 1)] = (a0b0, a0b1 + b0a1), +ϕ (A × A)×2 A × A (id ×F )×2 id ×ϕ (A × A)×2 +W / / A × A ·ϕ (A × A)×2 A × A (id ×F )×2 id ×ϕ (A × A)×2 ·W / / A × A. The verification that the above definition gives indeed a structure of an associative and com- mutative ring is a straightforward computation. We are now ready to prove: Theorem 3.16. Any morphism of Frobenius-split algebras f : (A, ϕA) → (B, ϕB) induces a morphism of fW2 : W ϕA 2 (A) is a flat W2(k)-lifting of A. (B). Moreover, the ring W ϕ 2 (A) → W ϕB 2 Proof. Firstly, we observe that our construction is functorial. This follows directly from the formulas for operations and the condition f ϕA = ϕBf . Consequently, we show that W ϕ 2 (A) is flat over W2(k). Clearly the projection onto the first factor gives a surjective ring homomorphism from W ϕ 2 (A) to A whose kernel I is given by the ideal of elements of the form (0, a1) ∈ W ϕ 2 (A) for a1 ∈ A, which is generated by a single element q = (0, 1). Note that the above diagrams prove that in case of a perfect field k with a Frobenius splitting γk(α) = p√α the bijective mapping id×F gives an isomorphism between the rings W γk 2 (k) and W2(k). This consequently means that natural homomorphism W2(k) ≃ W γk 2 (k) → W ϕ 2 (A) induced by the embedding k → A sends p = (0, 1) ∈ W2(k) to (0, 1) ∈ W ϕ 2 (A) and therefore q = p. This allows us to apply Corollary 2.2 to conclude that W ϕ 2 (A) is indeed a flat lifting of A. (cid:3) / / / / O O O O LIFTABILITY OF SINGULARITIES AND THEIR FROBENIUS MORPHISM MODULO p2 13 3.4.4. Glueing to a twisted Witt scheme. In order to globalise the above construction we shall show that it behaves well with respect to localisation. For this purpose, we prove the following lemma resembling the computation given in [Blo77, Lemma 5.1]. Lemma 3.17. Let A be a k-algebra. For any lifting (f, c) ∈ W ϕ homomorphism if : W ϕ 2 (A) of f ∈ A the natural 2 (Af ) is an isomorphism. 2 (A)(f,c) → W ϕ Before proceeding to the proof, we present a series of useful equalities in W ϕ 2 (A) which follows directly from addition and multiplication formulas: (7) (8) (f, c)−1 = (1/f,−c/f 2) (a, b)n = (an, nan−1b). Proof of Lemma 3.17. The proof boils down to showing that if is bijective. To prove it, we first apply equalities (7) and (8) to obtain: if(cid:18) (a, b) (f, c)n(cid:19) = (a, b) ·ϕ(cid:18) 1 f n+1(cid:19) =(cid:18) a f n , −nc and then carry on with a straightforward inspection using the following two facts: (i) a and only if f sa = 0 for some s ∈ N, (ii) the mapping Af ∋ d 7→ d− nac of parameters (n, a, c). = (a, b) ·ϕ(cid:18) 1 f n = 0 if f n+1 is bijective for any choice (cid:3) f n+1 (cid:19) , f 2(cid:19)n f b − nac , −c f n , f We are now ready to prove the theorem. Theorem 3.18. There exists a functor W ϕ to the category SchW2(k) of schemes over the ring of second Witt vectors such that for every Frobenius split variety (X, ϕX ) the object W ϕ 2 from the category Schsplit 2 ((X, ϕX )) is a flat lifting of X to W2(k). k Proof. In case of affine schemes (Spec(A), φ) it is sufficient to use the construction of W φ 2 (A). In general, we proceed as follows. For a scheme (X,OX ) we consider a ringed space (X, W ϕ 2 (OX )) where W ϕ 2 (OX ) is a sheaf equal to OX × OX set theoretically and with ring structure defined by the assignment U 7→ W ϕ 2 (OX (U )). The sheaf transition maps are induced by the functoriality of the the construction given in Theorem 3.16. Consequently, Lemma 3.17 proves that locally over ^ the affine subset V = Spec(A) of (X,OX ), the sheaf W ϕ 2 (OX (V )) and therefore (X, W ϕ 2 (OX )) is in fact a scheme over W2(k). 2 (OX )V is isomorphic to W ϕ (cid:3) 3.5. Alternative view on the construction. As suggested by Piotr Achinger, the above con- struction might be alternatively described as a certain subscheme of the second Witt scheme (X, W2(OX )). Firstly, observe that W2(OX ) is an algebra extension of OX by the ideal FX∗OX . The splitting ϕ gives rise to a subideal of FX∗OX given by Ker(ϕ) ≃ B1 (OX ) to be the quotient sheaf W2(OX )/ Ker(ϕ). We now prove: Proposition 3.19. The constructions W ϕ 2 (OX ) and W ϕ,alt (OX ) can be described by the following diagram: (OX ) give rise to isomorphic schemes. Proof. Indeed, the construction of W ϕ,alt X . We define W φ,alt 2 2 2 B1 X ≃ Ker(φ) 0 / FX∗OX ϕ W2(OX ) cϕ / OX 0 / W ϕ,alt 2 0 / OX where cϕ : W2(OX ) → W ϕ,alt (OX ) is a ring homomorphism. Therefore, we see that the operations in W ϕ,alt (OX ) are given by the lifting to W2(OX ) (e.g., id×F given in Definition 3.15), followed by the corresponding operation in W2(OX ) and the reduction cϕ. This is exactly the description given by the diagrams in Definition 3.15. (OX ) / OX / 0, (cid:3) 2 2 / / / / / / / / / / / / 14 MACIEJ ZDANOWICZ 4. Frobenius liftability We now proceed to the investigation of Frobenius liftability of affine schemes. We begin with functoriality of obstruction classes for lifting Frobenius, and then give a few examples. Conse- quently we present a computational criterion for Frobenius liftability of affine complete intersec- tions which we then apply to the case of ordinary double points in dimension ≥ 4 and canonical surface singularities. 4.1. Functoriality of obstructions to lifting Frobenius. As a direct corollary of Lemma A.10 we obtain the following result: L X/k X (1)/k,OX ) to lifting of relative Frobenius FX/k : X → X (1) to an infinitesimal Lemma 4.1. For any scheme X/k together with a W2(k)-lifting eX there exists an obstruction X ∈ Ext1(F ∗ σF flat thickening eFX/k : eX → eX (1) which satisfies the following functoriality property. For any eg : eX → eY restricting to a given g : X → Y the obstructions σF X and σF Y satisfy: Lg∗F ∗ Lg∗σF Y F ∗ σF Y L L Y /k Y (1)/k Y /k Y (1)/k Lg∗OY [1] OY [1] d(1) g g#[1] Rg∗F ∗ X/k L X (1)/k Rg∗σF X/ / Rg∗OX [1], X/kdg(1) F ∗ ≃ F ∗ X/k L X (1)/k σF X / OX [1] L L Y /k : F ∗ Y (1)/k → Rg∗F ∗ X (1)/k is the adjoint of the mapping F ∗ where d(1) g if g is affine and g# : OY → g∗OX splits then the Frobenius lifting of X descends to Y . Proof. We apply Lemma A.10 to the case X = Y , Y = Y (1), X ′ = X, Y ′ = X (1), f = FY /k, f ′ = FX/k, and then use the adjunction as in Lemma A.6. The second part follows from the functoriality diagram and the existence of a splitting of g#. (cid:3) X/kdg(1). In particular, X/k As an exemplary application we obtain a corollary: Corollary 4.2. Let j : U → X be the inclusion of an open subset such that complement Z = X\ U is of codimension ≥ 3 and X satisfies property S3 at any point of Z. Then, X admits a Frobenius lifting if and only if U does. Proof. Clearly, it suffices to show that Frobenius liftability of U implies liftability for X. For this purpose we observe that by Lemma A.23 the deformation functors of U and X are isomorphic. Hence, any lifting eU ∈ DefU (W2(k)) extends to a lifting eX ∈ DefX (W2(k)) and consequently we may apply functoriality of obstructions Lemma 4.1 to the inclusion eU → eX. We obtain a diagram: σF X L F ∗ X/k X (1)/k OX [1] d(1) j j#[1] Rj∗F ∗ U/k L U (1)/k Rj∗σF U / / Rj∗OU [1]. By assumption σF natural mapping: U = 0 and therefore j# ◦ σF X = 0. To conclude that σF X = 0, it suffices to show a Ext2(F ∗ X/k L X (1)/k,OX ) −→ Ext2(F ∗ X/k L X (1)/k, Rj∗OU ) coming from the long exact sequence of Ext•(F ∗ guished triangle X/k L X (1)/k,−) groups associated to the distin- OX −→ Rj∗OU −→ Kj −→ OX [1] is injective. This follows from the vanishing Ext1(F ∗ X (1)/k, Kj) = 0, which we prove along the lines of the proof of Lemma A.23 (this requires that any point in the complement X \ U satisfies the property S2). X/k (cid:3) L The simplest example of a Frobenius liftable scheme is given by the following. / / / / / / / LIFTABILITY OF SINGULARITIES AND THEIR FROBENIUS MORPHISM MODULO p2 15 Example 4.3 (Frobenius liftability of toric varieties). Every affine toric variety Spec(A)/k is Frobenius liftable. Proof. It is well-known that A is isomorphic to a k-algebra k[M ] associated to a finitely generated monoid M . Its flat lifting is given by W2(k)[M ] and the corresponding Frobenius lift is induced by the monoid mapping M ∋ m 7→ pm. (cid:3) 4.2. Computational criterion. In this section we present an explicit condition for the existence of a lifting of the Frobenius morphism to a fixed W2(k)-lifting of an affine scheme Spec(k[x1, . . . , xn]/(f1, . . . , fm)) (cf. Lemma 2.4). Then, we derive a compu- tational criterion for the Frobenius liftability in the case of affine complete intersection schemes (in particular hypersurface singularities). Spec(W2(k)[x1, . . . , xn]/(ef1, . . . ,ffm)) Before stating the criteria we need some auxiliary definitions and notation. We denote by R the polynomial ring k[x1, . . . , xn], by eR its natural W2(k)-lifting W2(k)[x1, . . . , xn] (unique up to isomorphism) and by eF : eR → eR the lifting of Frobenius morphism given by the formula: X aI xI 7→X σ(aI )xpI , where I = (i1, . . . , in) is a multi-index, xI denotes the monomial xi1 1 ··· xin n and σ is the Witt vector Frobenius described in Section 1.1. Moreover, let w : eR → eR be the base change homomorphism over σ : W2(k) → W2(k) defined by: w(cid:16)X aI xI(cid:17) =X σ(aI )xI . Proof. Let I ⊳ k[x1, . . . , xn] be the ideal generated by the polynomials (f1, . . . , fm) and J ⊳ W2(k)[x1, . . . , xn] be the ideal generated by the their lifts (ef1, . . . ,ffm). Every Frobenius lift- ing of B = eR/J comes from a homomorphism h : eR → eR over σ : W2(k) → W2(k), given by the assignment xi 7→ xp In order to get an induced mapping B → B the ideal J should be mapped to itself. This boils down to the following i + ji + p · hi for ji ∈ J, and W2(k) ∋ a 7→ σ(a). element belonging to pB can be uniquely identified (by multiplication by p) with an element of For any W2(k)-flat ring B = W2(k)[x1, . . . , xn]/(ef1, . . . ,ffm) we see by Corollary 2.2 that any A = B/p. In particular for any element ef ∈ eR there exists a unique element P (ef ) ∈ R satisfying pP (ef ) = eF (ef ) − ef p. Example 4.4. It turns out we have already seen an example of a polynomial P (f ). Namely, in Proposition 3.3 we defined a polynomial: Pf2n = Xi1+...+in=p i1,...,in6=p (p − 1)! i1!··· in! (x1x2)i1 ··· (x2n−1x2n)in . This is exactly −P (x1x2 + . . . + x2n−1x2n). We are now ready to present the first computational condition for the existence of a Frobenius lift. algebra A = k[x1, . . . , xn]/(f1, . . . , fm). Then, the Frobenius morphism of A lifts if and only if there exists a sequence of elements hk ∈ R such that for every i ∈ {1, . . . , m} the element Lemma 4.5. Suppose B = W2(k)[x1, . . . , xn]/(ef1, . . . ,ffm) is a W2(k)-flat lifting of the affine P (efi) = eF (ef )−ef p ∈ R satisfies: p hk (mod f1, . . . , fm). ∂xk(cid:19)p P (efi) = X1≤k≤n(cid:18) ∂fi 16 MACIEJ ZDANOWICZ computation: (9) ∂xk (xp · hk n) · hk 1, . . . , xp ∂w(efi) n + jn + p · hn) h(efi) = w(efi)(xp 1 + j1 + p · h1, . . . , xp n) + p · X1≤k≤n = w(efi)(xp 1, . . . , xp ∂xk(cid:19)p = eF (efi) + p · X1≤k≤n(cid:18) ∂fi + p · X1≤k≤n(cid:18) ∂fi ∂xk(cid:19)p ) +efi = (eF (efi) −efi = p ·P (efi) + X1≤k≤n(cid:18) ∂fi · hk (mod J), ∂xk(cid:19)p and p2 = 0, and in (10) we used the fact that (eF (efi) −ff p i ) = p · P (efi) for a unique polynomial P (efi) ∈ R. By Corollary 2.2 we therefore see that existence of a Frobenius lifting for a W2(k)-lifting where in (9) we applied the Taylor expansion of a polynomial, together with the fact that ji ∈ J B is equivalent to existence of hi ∈ R such that: · hk (10) p p ∂xk(cid:19)p P (efi) = X1≤k≤n(cid:18) ∂fi hk (mod I), for any i ∈ {1, . . . , m}. This finishes the proof. (cid:3) We now restrict ourselves to the case of affine complete intersection schemes and derive a computationally feasible criterion to check whether they possess a W2(k)-lifting compatible with Frobenius. For this purpose we firstly observe that: P (ef + pg) − P (ef ) = (eF (ef + pg) − (ef + pg)p) p − (eF (ef ) − f p) p = eF (pg) p = gp, which implies that the choice of a different lifting ef of f leads to the change of P (ef ) by an arbitrary p-th power gp. This leads to a criterion: Theorem 4.6. The affine complete intersection algebra A = k[x1, . . . , xn]/(f1, . . . , fm) is Frobe- nius liftable if and only if there exists a sequence of elements hk ∈ R and a sequence of elements gi ∈ R such that: (11) (mod f1, . . . , fm), Pfi + gp hk ∂xk(cid:19)p i = X1≤k≤n(cid:18) ∂fi Proof. By Lemma 2.4 we know that flat liftings of complete intersection affine schemes are given by the liftings of the the associated regular sequences. Therefore, by Lemma 4.5 we need to find where Pfi is a polynomial defined up to a p-th power and computed by the formula Pfi = P (efi) for any lifting efi of fi. liftings {efi} if {fi} such that there exist {hk} ∂xk(cid:19)p P (efi) = X1≤k≤n(cid:18) ∂fi But, as we mentioned, the polynomials P (efi) for different liftings of fi differ by an arbitrary p-th As a corollary we obtain the criterion for Frobenius liftability of hypersurface singularities: power. This finishes the proof. (mod f1, . . . , fm), hk (cid:3) Corollary 4.7. Suppose Hf = Spec(k[x1, . . . , xn]/(f )) is a hypersurface in An a lift to W2(k) compatible with Frobenius if and only if there exists an element g such that k . Then Hf admits Pf + gp ∈(cid:18)f,(cid:18) ∂f ∂xk(cid:19)p , . . . ,(cid:18) ∂f ∂xn(cid:19)p(cid:19) . LIFTABILITY OF SINGULARITIES AND THEIR FROBENIUS MORPHISM MODULO p2 17 Remark 4.8. The equality above might be treated as a relation in the Frobenius push-forward F∗ (k[x1, . . . , xn]/(f )). From this perspective, the polynomial Pf ∈ F∗ (k[x1, . . . , xn]/(f )) (module structure by p-th powers) is supposed to belong to a submodule k[x1, . . . , xn]/(f )·1+( ∂f )· F∗(k[x1, . . . , xn]/(f )) (in fact, gp = g · 1 ∈ F∗ (k[x1, . . . , xn]/(f ))). , . . . , ∂f ∂xn ∂xk It turns out that the second form of the above criterion can be (quite) efficiently verified using Macaulay2. The code is given at http://www.mimuw.edu.pl/mez/Macaulay2. Remark 4.9. The existence of Frobenius lifting depends on the choice of the W2(k)-lifting. Indeed, Frobenius liftability of any lifting of Spec(k[x1, . . . , xn]/(f )) would mean that Pf + gp )p) contains all 3 in characteristic p ≡ 1 (mod 3). In )p) for any g and consequently (f, ( ∂f ∂xk )p, . . . , ( ∂f ∂xn )p, . . . , ( ∂f ∂xn belongs to (f, ( ∂f ∂xk p-th powers. This is already not true for f = x3 fact, a simple computation shows that xp 1 6∈ (x3 2 , x2p 3 ). 1 , x2p 2 + x3 1 + x3 1 + x3 2 + x3 3, x2p 4.3. Non-liftability of ordinary double points in any dimension. This section is devoted to answering the question posed in [Bha14][Remark 3.14], whether the ordinary double points, in particular cones over smooth projective quadrics, admit a lift to W2(k) compatible with Frobenius in arbitrary dimensions. The negative answer is given by following Theorem 4.10. Theorem 4.10 (Frobenius non-liftability of ordinary double points). Let n > 5 be an integer and assume p > 3. Then the ordinary double points defined by the equation f = x2 n + Q for Q ∈ k[x1, . . . , xn]≥3 does not admit a W2(k)-lifting compatible with Frobenius. Proof. For the proof, we need the lemma given in [Lan08] combined with the Proposition 3.3. Lemma 4.11 ([Lan08], Proposition 3.1). Let k be a field of characteristic p > 3 and 0 ≤ e < p be an integer. Then for any d ≤ N · p−1 NXi=1 i )p−e!d 2 − e we have = xp i )e!d (xp 1 + . . . + x2 NXi=1 1, . . . , xp 1, . . . , xp N ) : ( N , ( x2 x2 in k[x1, . . . , xN ]. By Corollary 4.7 we need to show that Pf + gp 6∈(cid:18)f,(cid:18) ∂f ∂xk(cid:19)p for any g ∈ k[x1, . . . , xn]. Clearly,(cid:16)f,(cid:16) ∂f ∂xk(cid:17)p , . . . ,(cid:16) ∂f fices to prove that Pf 6∈ (f, xp such that: + P ′ = h · (x2 1, . . . , xp 1+...+x2 n ∂xn(cid:19)p(cid:19) , . . . ,(cid:18) ∂f ∂xn(cid:17)p(cid:17) ⊆ (f, xp n) and therefore it suf- n). Assume the contrary, i.e., there exists an h ∈ k[x1, . . . , xn] 1, . . . , xp Pf = Px2 (mod xp n), where P ′ is a polynomial in k[x1, . . . , xn]>2p. The ring k[x1, . . . , xn]/(xp n) admits a grading (induced from the polynomial ring). We may assume that the lowest grade m of the polynomial h satisfies m ≥ 2p − 2. Indeed, if m < 2p − 2 then by lowest grade comparison and Lemma 4.11 for N = 5, e = 1 and d = m we see that 1, . . . , xp 1, . . . , xp 1 + . . . + x2 n + Q) h[m] ∈ (xp 1, . . . , xp n) : ( i )!m x2 = xp nXi=0 1, . . . , xp N , ( i )p−1!m x2 NXi=1 ⊂ (xp 1, . . . , xp n)m and therefore h − h[m] is also an appropriate choice of h with the lowest grade at least m + 1. Consequently, by the 2p grade comparison we see that Px2 1 + . . . + n, xp x2 n). This gives a contradiction with the proof of Proposition 3.3. Namely Pfn from Proposition 3.3 is the representation of the polynomial Px2 after standard linear transfor- mation. (cid:3) belongs to the ideal (x2 1, . . . , xp 1+...+x2 n 1+...+x2 n This in fact implies that the methods resulting from [Bha14] cannot be applied for ordinary double points. However, Bhatt proved that crystalline cohomology over k of homogeneous ordinary double points is not finitely generated by a different calculation. 18 MACIEJ ZDANOWICZ 4.4. Canonical singularities in positive characteristic are Frobenius liftable. In this sec- tion, we present two different approaches to proving that canonical singularities of surfaces are F -liftable. The first one is a direct computation using the criterion given in Corollary 4.7. The second gives a slightly weaker result in the case of tame quotient singularities and exploits the structure of the quotient together with functoriality of obstructions for lifting Frobenius morphism. We note that in any characteristic the canonical surface singularities over algebraically closed field are classified by the dual graphs of the minimal resolution, which in turn correspond to Dynkin diagrams of type A, D and E. Therefore, we shall use the terms canonical surface singularities and ADE singularities interchangeably. 4.4.1. Direct computational approach. Firstly, we approach the F -liftability of ADE singularities by a direct computation. By [Art77] under the assumption p ≥ 7 the ADE singularities are locally analytically equivalent to the germs of the following form: Table 1. Models of canonical singularities in char. p ≥ 7 Type smooth An−1 Dn+2 E6 E7 E8 Equation - xn + y2 + z2 = 0 x2 + y2z + zn+1 = 0 x2 + y3 + z4 = 0 x2 + y3 + yz3 = 0 x2 + y3 + z5 = 0 Equipped with this classification we can approach the proof of the following: Theorem 4.12. For any algebraically closed field k of characteristic p ≥ 7 any affine scheme X with canonical surface singularities is Frobenius liftable. Proof. Firstly, we observe that by Lemma A.12 it suffices to prove that for any x ∈ X the germ Spec(OX,x) is Frobenius liftable. By Artin's approximation theorem and the analytical classifica- tion of canonical singularities there exists a diagram of schemes: (U, u) ´etale yrrrrrrrrrr Spec(OX,x) ´etale &◆◆◆◆◆◆◆◆◆◆◆ Spec(OSADE,s), where OSADE,s is the local ring at 0 of one of the affine models given in Table 1. Therefore, by Lemma A.11 and Lemma A.12, to prove Frobenius liftability of Spec(OX,x) it suffices to show that the affine models given in Table 1 are F -liftable. The case of An−1 singularities simply follows from Example 4.3 as An−1 singularities are toric. Consequently, we treat the case of Dn+2 singularities. The affine model of Dn+2 is given by the equation x2 + y2z + zn+1 = 0 and therefore by Corollary 4.7 we are left to show that: x2i(y2z)jz(n+1)k ∈ (x2 + y2z + zn+1, xp, ypzp, y2p + (n + 1)pznp). (cid:0) p i,j,k(cid:1)p Xi+j+k=p i,j,k6=p (cid:0) p i,j,k(cid:1)p Xi+j+k=p i,j,k6=p By substituting u = x2 = −y2z − zn+1 we see that it suffices to prove that: (−1)i(y2z + zn+1)i(y2z)jz(n+1)k ∈ ((y2z + zn+1)⌈p/2⌉, ypzp, y2p + (n + 1)pznp). The left hand side of the above relation is a sum of elements of the form (y2z)a · z(n+1)b for a + b = p so it is enough to show that: (y2z)az(n+1)b ∈ ((y2z + zn+1)⌈p/2⌉, ypzp, y2p + (n + 1)pznp), y & LIFTABILITY OF SINGULARITIES AND THEIR FROBENIUS MORPHISM MODULO p2 19 for any a + b = p. The claim is clear for every a ≥ ⌈p/2⌉ and therefore we reason by contradiction. Suppose a ≤ ⌊p/2⌋ is maximal with respect to the relation (y2z)az(n+1)(p−a) 6∈ ((y2z + zn+1)⌈p/2⌉, ypzp, y2p + (n + 1)pznp). As p − a ≥ ⌈p/2⌉ we can write (y2z)a · z(n+1)(p−a) = (y2z)az(n+1)⌈p/2⌉z(n+1)(p−a−⌈p/2⌉) i (cid:19)(y2z)iz(n+1)(⌈p/2⌉−i) = (y2z)az(n+1)(p−a−⌈p/2⌉) · X1≤i≤p(cid:18)⌈p/2⌉ = X1≤i≤p(cid:18)⌈p/2⌉ i (cid:19)(y2z)a+iz(n+1)(p−a−i), (cid:0) p i,j,k(cid:1)p x2iy3j(yz3)k ∈ (x2 + y3 + yz3, xp, 3y2p + z3p, ypz2p). Xi+j+k=p i,j,k6=p which gives a contradiction with the definition of a. We now proceed to the case of E7. We shall prove that By the trick as above, it suffices to prove that for any a + b = p we have (yz3)ay3b ∈ ((y3 + yz3)⌈p/2⌉, 3y2p + z3p, ypz2p), which follows by analogous reasoning by contradiction. We finish with case of E2m singularities given by the affine equation x2 + y3 + zm+1 for m = 3, 4. By Corollary 4.7 to prove the existence of W2(k)-lifting compatible with Frobenius it suffices to prove that Xi+j+k=p(cid:0) p i,j,k(cid:1)p x2iy3jz(m+1)k ∈ (x2 + y3 + zm+1, xp, y2p, zmp). After substituting u = x2, v = y3, w = zm+1 and s = u + v + w it suffices to show that: Xi+j+k=p(cid:0) p i,j,k(cid:1)p uivj(s−u−v)k ∈ (s, u⌈p/2⌉, v⌈2p/3⌉, (s−u−v)⌈ mp m+1 ⌉) = (s, u⌈p/2⌉, v⌈2p/3⌉, (u+v)⌈ mp m+1 ⌉). For this purpose, we prove that any monomial uivj ∈ k[u, v] for i + j = p appearing on the left hand side belongs to (u⌈p/2⌉, u⌈2p/3⌉, (u + v)⌈ mp m+1 ⌉), which is equivalent to the surjectivity of the linear mapping defined by: m+1 ⌋ ∋ (f, g, h) 7→ f · u⌈p/2⌉ + g · u⌈2p/3⌉ + h · (u + v)⌈mp/(m+1)⌉. k[u, v]⌊p/2⌋ × k[u, v]⌊p/3⌋ × k[u, v]⌊ p After writing this in the basis given by monomials ukvj ordered lexicographically, to show sur- , for jectivity it is sufficient to show that the determinant of the matrix h(cid:0)⌈mp/(m+1)⌉ n = p − 1 − ⌊p/2⌋ − ⌊p/3⌋ is non-zero. By formula (2.1) in [Kra05], it is equal to: (cid:1)i1≤i,j≤n (⌈ mp m+1⌉ + i − 1)!(i − 1)! (s + i − 1)!(⌈ mp m+1⌉ + n − 1 < p. nYi=1 and therefore we are done by the inequality ⌈ mp 4.4.2. Quotient singularity approach for F -liftability of canonical singularities. Here, we approach the F -liftability of canonical singularities of surfaces using a different method based on functoriality of obstructions. We begin with a general result concerning F -liftability of quotients of smooth schemes by finite groups, for which the action can be lifted to characteristic 0 (in fact, lifting mod p2 would be sufficient). Subsequently, we investigate to which extent our method can be applied in the case of canonical surfaces singularities. Our analysis is based on the results of [LS14, Section 4] describing whether ADE singularities admit a structure of a quotient singularity. m+1⌉ − s + i − 1)! s+i−j det(cid:20)(cid:18) ⌈ mp s + i − j(cid:19)(cid:21)i,j≤n m+1⌉ = (cid:3) , We precede our actual considerations by a few essential and standard remarks concerning base change for invariants of actions of finite groups. The proofs are based on the existence of so-called Reynolds operator. 20 MACIEJ ZDANOWICZ Lemma 4.13. Let G be a finite group and let R be a commutative ring such that G is invertible in R. Then, any R[G]-module M projective as an R-module is projective as an R[G]-module. Proof. In order to prove that M is projective we need show that for any diagram of R[G]-modules: M f P p / N / 0 there exists an R[G]-linear homomorphism s : M → P such that f = p ◦ s. By the assumptions we obtain an R-linear homomorphism σ : M → P such that f = p ◦ α. Consequently, by a direct computation we obtain that a homomorphism defined by the formula (i.e., the Reynolds operator): s(m) = 1 GXg∈G gσ(g−1m) is R[G]-linear and satisfies f = p ◦ s. We now apply the above result to prove: (cid:3) Lemma 4.14. Let G be a finite group and let R be a commutative ring such that G is invertible in R. Let M be a projective R-module together with an action of G. Then, for any homomorphism R → S the natural mapping M G ⊗R S → (M ⊗R S)G is an isomorphism. Proof. Firstly, by Lemma 4.13 we observe that M is a projective R[G]-module. Therefore, there exists an R[G]-module N such that M ⊕ N ≃ R[G]I for some set of indices I. Consequently, we obtain the following diagram: (M ⊕ N )G ⊗R S ≃ ≃ / M G ⊗R S ⊕ N G ⊗R S (R[G]I )G ⊗R S / (M ⊗R S)G ⊕ (N ⊗R S)G ≃ (R[G] ⊗R S)G, which reduces our claim to the case of free modules. We finish by the following simple sequence of identifications: (R[G]I )G ⊗R S → RI ⊗R S → SI → (S[G]I )G → (R[G]I ⊗R S)G. (cid:3) As a corollary we obtain: Corollary 4.15. Let G be a finite group and let R be a ring such that G is invertible in R. Then, for every G-action over Spec(R) on An Spec(R)/G is flat over R. Moreover, for every ring homorphism R → S the natural map: Spec(R) the quotient An is an isomorphism. Spec(R)/G(cid:17) ×Spec(R) Spec(S) → An (cid:16)An Spec(S)/G. Proof. Using the same idea as in Lemma 4.13 we prove that the ring of invariants R[x1, .., xn]G is a direct sum of a flat R-module R[x1, . . . , xn]. This implies that An Spec(R)/G = Spec(R[x1, .., xn]G) is flat over R. The base change property is a direct consequence of Lemma 4.14. (cid:3) We are now ready to approach the following result concerning F -liftability of quotient singu- larities. Lemma 4.16 (Frobenius liftability of mod p reductions of quotients). Let G be a finite group and let T be a spectrum of finitely generated Z-algebra ´etale over Spec(Z[1/G]). Suppose, G acts on An T relatively to T . Then, for every perfect field k and a k-point t ∈ T (k) the scheme t /G ≃ (An An T /G)t is W2(k)-liftable compatibly with Frobenius. / / / / πt '❖❖❖❖❖❖❖❖❖❖❖❖❖❖ w♥♥♥♥♥♥♥♥♥♥♥♥ (An T /G)t ≃ An t /G T /G)et of (An &◆◆◆◆◆◆◆◆◆◆◆◆◆ πet w♣♣♣♣♣♣♣♣♣♣♣ / (An T /G)et LIFTABILITY OF SINGULARITIES AND THEIR FROBENIUS MORPHISM MODULO p2 21 Proof. Let t ∈ T (k) be k-point of T . By the formal lifting property of the ´etale map T → Spec(Z), any point t ∈ T (k) can be lifted to a point et ∈ T (W2(k)). Consequently, by the flatness of T /G → T (see Corollary 4.15) we obtain a W2(k)-lifting (An An T /G)t fitting into a commutative diagram with cartesian squares: et An t An Spec(k) / Spec(W2(k)). T /G)t is isomorphic to An By Corollary 4.15 the scheme (An t → An t /G. Hence by the assumptions on T the degree deg(πt) (equal to the length of a generic orbit) is coprime to the characteristic of k and therefore there exists a splitting operator of π# t , which consequently allows us to apply Lemma 4.1 (for the upper square with mappings πt and πW2(k)) to obtain our claim. More precisely, the Frobenius lifting of An T /G)W2(k), which is a W2(k)-lifting of (An (cid:3) t /G and πt is a quotient map An W2(k) descends to (An T /G)W2(k). In order to apply Lemma 4.16 to the case of canonical singularities of surfaces we need some results from [LS14]. First, let us recall the definition: Definition 4.17. A scheme over a field k has linearly reductive quotient singularities (resp. tame quotient singularities) if it is ´etale equivalent to a quotient of a smooth k-scheme by a finite linearly reductive group scheme (resp. finite ´etale group scheme of order prime to the characteristic of k). It is proven in [LS14, Prop. 4.2] that except for a few characteristics all ADE singularities fit into the above definition. Moreover, from the proof we can infer the following result concerning characteristic 0 liftability of group actions leading to ADE singularities: Proposition 4.18. For any singularity type τ ∈ {An−1, Dn+2, E6, E7, E8} there exists a finite flat generically ´etale group scheme Λτ /Spec(Z) such that for every field k the type τ singularity over k arises as the quotient A2 k/Λk. The group scheme Λ is ´etale precisely over an open subset of Spec(Z) where the corresponding group scheme quotient is tame (in the sense of the above definition). Example 4.19. For example An−1 singularity over a field k of arbitrary characteristic arises as the quotient of A2 k/Λk by the natural action of the group scheme µn,k ≃ Λk = Λ ×Spec(Z) Spec(k) for Λ = µn,Z. In this case, the scheme Λ/ Z is ´etale over Spec(Z[1/n]). Finally, the details of the behaviour of canonical surface singularities with respect to character- istic of the field k can be summarised by the following table from [LS14, Section 4]. Table 2. Classification of canonical singularities of surfaces Linearly reductive quotient singularity Tame quotient singularity An−1 Dn+2 E6 E7 E8 every p p ≥ 3 p ≥ 5 p ≥ 5 p ≥ 7 p ≥ 3, p 6 n p 6 n p ≥ 5 p ≥ 5 p ≥ 7 Note that the last column indicates the open subset where the respective group scheme Λ is ´etale. We are now ready to give an alternative proof of F -liftability of canonical singularities of surfaces in the case they are tame quotient singularities. Theorem 4.20 (Tame version of Theorem 4.12). Suppose X is an affine surface over an alge- braically closed field k such that its singularities are tame quotient canonical surface singularities. Then, X is Frobenius liftable. / / ' & w / w / 22 MACIEJ ZDANOWICZ Proof. Again by Lemma A.11 and Lemma A.12 it suffices to address the case of tame quotients. By Proposition 4.18 and the tameness assumption there exists an open subset U = Spec(Z[1/N ]) ⊂ Spec(Z) and an ´etale group scheme Λ/U of rank coprime to N such that X ≃ A2 u/Λu for a certain point u ∈ U (k). Every finite ´etale group scheme is ´etale locally constant and therefore there exists = Λ ×U T is isomorphic to a constant group scheme an ´etale morphism T → U such that ΛT associated to a finite group G. As k is algebraically closed, the point u ∈ U (k) lifts to t ∈ T (k). T the fibre A2 By Lemma 4.16 applied for G acting on A2 T /G)t is F -liftable. This finishes the proof. (cid:3) u/Λu ≃ A2 t /G ≃ (A2 def Example 4.21. A Dn+2 singularity arises as the quotient of A2 dihedral group scheme BDn of rank 4n induced by the natural operation of the matrices: k by the action of the binary (cid:20)ξ2n 0 0 ξ−1 2n(cid:21) and (cid:20)0 −1 0 (cid:21) , 1 where ξ2n denotes the 2n-th primitive root of unity. This means that the model of Dn+2 is tame over Spec(Z[1/2n]) and the associated ´etale group scheme is trivialized by the covering induced by Z → Z[1/2n][ξ2n]. For more details, the reader is referred to [LS14, Section 4]. 5. W2(k)-liftability and F -liftability compared to standard F -singularities The following section contains the comparison of the classical F -singularity types, i.e., F - regularity, F -purity and F -rationality with the notions of W2(k)-liftability and Frobenius liftability. We work with F -finite rings, i.e, characteristic p > 0 rings such that F∗R is a finite R-module. In particular, the F -finiteness assumption is satisfied in the case of essential finite type algebras over a field k of characteristic p. We begin by recalling a few necessary definitions and criteria. Note that we freely identify affine schemes with their associated rings. Definition 5.1 (F -regularity). Let R be a reduced F -finite ring of characteristic p > 0. i) We say that R is F -regular if for any c ∈ R the mapping ϕe ii) We say that R is F -pure if the mapping F : R → F∗R splits, i.e., the scheme Spec(R) is 1 7→ c splits for some e ≫ 0. Frobenius split. ∗ R defined by the formula c : R → F e It turns out that both of above properties can be verified locally, i.e., R is F -pure (respectively strongly F -regular) if and only if for any maximal m ∈ Spec(R) the local ring Rm is F -pure (respectively strongly F -regular). In the case of quotients of F -finite regular rings the above properties can be efficiently verified using the following criteria. Lemma 5.2 (Fedder's and Glassbrenner's criteria). Let S be a F-finite regular ring such that F∗S is a free S-module and let R = S/I. Then R is F -split at m ⊃ I if and only if I [p] : I 6⊂ m[p]. Moreover, let s be an element of S not in any minimal prime of I such that Rs is regular. Then, R is strongly F -regular at m if and only if there exists an e ∈ N such that s(I [pe] : I) 6⊂ m [pe]. Proof. For the proof see [Gla96, Theorem 2.3]. (cid:3) Note that the assumptions of the above lemma are satisfied for S = k[x1, . . . , xn] and S regular local. Moreover, in the special case I = (f ) the colon ideal I [pe] : I = (f pe−1) and therefore F -purity boils down to an efficiently verifiable criterion f pe−1 ∈ m[pe]. Definition 5.3 (F -rational). Let (R, m) be a d-dimensional local ring of characteristic p > 0. i) We say that R is F -rational if R is Cohen-Macaulay and if for any c ∈ R◦, there exists e ∈ N such that cF e : H d m(R) → H d m(R) is injective. We shall need the following lemma. Lemma 5.4 (Watanabe a-invariant). Let (R, m) be a Cohen-Macaulay N-graded ring of dimension d such that the punctured spectrum Spec(R) \ {m} is F -rational. Then, if the invariant a(R) = max{i ∈ Z : [H d m(R)]i 6= 0} satisfies the inequality a(R) < 0 then for any n ≫ 0 the Veronese subring R(n) is F -rational. def Proof. See [Sin98, Proposition 5.1.1]. (cid:3) LIFTABILITY OF SINGULARITIES AND THEIR FROBENIUS MORPHISM MODULO p2 23 We are now ready to describe our results. They can be summarized by the following diagram indicating possible implications and referring to counterexamples in case they do not hold: F -liftable \Ex. 5.7 F -regular F -rational \Ex. 5.8 \Ex. 5.7 \ Thm 4.10 \Thm 4.10 yes : normal k-algebra - Thm 5.5 no : non-normal - Ex. 5.6 F -pure Thm 3.18 W2(k)-liftable. Firstly, we prove that an F -liftable normal scheme over k is Frobenius split. Theorem 5.5. Let X/k be normal scheme locally of finite type over a perfect field k, such that Frobenius split. there exists a lifting eX/W2(k) together with a lifting eF of Frobenius morphism F . Then, X is Proof. The following is a simple extension of a proof given in [MS87] covering the smooth case. Let n be the dimension of X and let U be the smooth locus of X. In case of smooth schemes liftability of Frobenius is equivalent to the existence of a splitting ξ : Ω1 U of the Cartier mapping C (see U ≃ F∗Ωn [MS87]). After taking the n-th exterior power of ξ, we obtain a mapping ∧nξ : Ωn U which splits the sequence: U → Z n U → Z 1 0 / Bn U dual to the sequence / Z n U ≃ F∗Ωn U / Ωn U / 0, / B1 U / 0, 0 by applying HomOU (−, Ωn property, one may extend this to a splitting of X. / OU / F∗OU U ). Therefore, U is Frobenius split. Using normality of X, in fact S2 (cid:3) We shall now present an example which violates the hypothesis of Theorem 5.5 in case of non-normal schemes. Example 5.6. The scheme {x1x2(x1 + x2) = 0} ⊂ A2 Proof. We apply Fedder's criterion given in Lemma 5.2 and a simple fact that: k is Frobenius liftable but not F -pure. (x1x2(x1 + x2))p−1 ∈ (xp 1, xp 2). i preserves the ideal x1x2(x1 + x2). Frobenius liftability follows by observing that the canonical lifting of Frobenius of W2[x1, x2] given by xi 7→ xp (cid:3) Example 5.7. The cone over an ordinary elliptic curve is F -liftable but not strongly F -regular or F -rational. Proof. For for the sake of clarity, we focus on the case of a cone over Fermat cubic C = {x3 + y3 + z3 = 0} ⊂ A3 k in characteristic p ≡ 1 (mod 3). Firstly, F -liftability follows from Corollary 4.7 by a simple direct computation. We verify that C is strongly F -regularity by means of Lemma 5.2. Indeed, we observe that for any s ∈ m = (x1, x2, x3) we have sf pe−1 ∈ (x1, x2, x3)[pe] and consequently the cone is not strongly F -regular at m. Finally, the proof of the fact that the cone C is not F -rational is the content of [TW15, Example 2.6]. (cid:3) Example 5.8. There exists an F -rational scheme which is not W2(k)-liftable. Proof. Let X be a scheme embedded into Pn k which is not W2(k)-liftable. By [LS14, Theorem 2.3] we see that Y k ) does not lift to W2(k), either. We claim that a cone over sufficiently ample projective embedding of Y satisfies our requirements, i.e., is F -rational and does not lift to W2(k). = BlX (Pn def / / / / / / / / 24 MACIEJ ZDANOWICZ Indeed, using [AC11] and Leray spectral sequence we see that H i(Y,OY ) = 0 for i > 0, and therefore we may apply Corollary 6.5 and Proposition 6.8 to obtain a projective embedding Y ⊂ PN such that the cohomology groups H i(Y,OY (k)) vanish for i > 0 and k > 0, the cone CY = ConeY,PN is Cohen-Macaulay and does not lift to W2(k). Consequently, in order to apply Lemma 5.4 (which might require additional Veronese embedding) we are left to show that the positive degree part [H n+1 m (CY ,OCY ) is zero. This follows from the identification [H n+1 m (CY ,OCY )]≥0 of the graded module H n+1 m (CY ,OCY )]k = H n(Y,OY (k)) = 0 given in Proposition 6.3. (cid:3) def 6. W2(k) and Frobenius liftability of cones over projective schemes Here, we present a few result concerning cones over projective schemes. We begin with a series of classical results (see, e.g., [Sch71, Art76]). Consequently, in Section 6.2.1 we extend the classical results with a comparison theorem for Frobenius liftability. For the purpose of clarity we present a few definitions and results concerning cones over pro- jective schemes and their deformation. We begin by introducing the construction of a cone over a closed embedding of a scheme into the projective space Pn def k . For the sake of brevity we set R = k[x0, . . . , xn]. 6.1. Construction of a cone over a projective scheme. For any closed subscheme X ⊂ Pn given by an ideal sheaf I, we denote by IX the homogeneous ideal: = Pn From the long exact sequence of cohomology associated to the short exact sequence: def IX =Mk≥0 H 0(OPn ,I(k)) ⊳ R ≃Mk≥0 0 −→Mk≥0 I(k) −→Mk≥0 OPn (k) −→Mk≥0 we see that IX arises as the first term in: rX−−→Mk≥0 H 0(Pn,OPn (k)) H 0(Pn,OPn(k)). OX (k) −→ 0, 0 −→ IX −→ R =Mk≥0 Definition 6.1. Suppose X ⊂ Pn = Proj(R) is a closed embedding of k-schemes. We define the affine cone over X ⊂ Pn to be an affine scheme: H 0(X,OX (k)) −→Mk≥0 H 1(Pn,I(k)) −→ 0. We define the affine cone, as follows. Directly from the definition we see that: ConeX,Pn def = Spec(R(cid:14)IX ). Remark 6.2. The ring R(cid:14)IX admits a natural grading induced by the grading of R. the complement of the vertex, and by p : U → X the natural Gm-bundle given by For notational convenience, we denote by j : U = ConeX,Pn \ {m} → ConeX,Pn the inclusion of U ≃ SpecX Mk∈Z OX (k)! → X, where the isomorphism follows from the Grauert's ampleness criterion (see EGA II 8.9.1). We have the following standard proposition. Proposition 6.3. Suppose X ⊂ Pn is a closed subscheme. The affine scheme C = ConeX,Pn is normal if and only if X is itself normal and the restriction mapping rX is surjective. In this case, C ≃ Spec(cid:0)Lk∈Z H 0(X,OX (k))(cid:1) ≃ Spec(H 0(U,OU )). Moreover, for i ≥ 1 the local cohomology m (C,OC ) =Lk∈Z H i(X,OX (k)), as graded H 0(C,OC )-modules. around the vertex satisfy H i+1 Remark 6.4. Taking the cone over the d-th Veronese embedding of X ⊂ Pn into P(n rX,d :Mk≥0 H 0(Pn,OPn (dk)) →Mk≥0 to considering the morphism of Veronese subrings H 0(X,OX (dk)). d)−1 amounts def LIFTABILITY OF SINGULARITIES AND THEIR FROBENIUS MORPHISM MODULO p2 25 By the Serre's vanishing theorem we see that there exists d such that for k ≥ d we have H 1(X, IX (k)) = 0, and therefore there exists d such that rX,d is surjective. This means that for d sufficiently large the cone over the Veronese embedding of X is normal. Moreover, we have the corollary: Corollary 6.5. Let X be a smooth projective scheme of dimension dim(X) = n > 2 satisfying H i(X,OX ) = 0 for i ∈ {1, . . . , n − 1}. Then, for a sufficiently ample embedding X ⊂ PN the cone satisfies property Sn+1. Proof. This follows directly from the local cohomology part of Proposition 6.3 and Serre vanishing. (cid:3) 6.2. Deformation theory and cones. We now proceed to the summary of deformation theory of cones. Again, X denotes a closed subscheme of Pn and we assume that C = ConeX,Pn is normal at its vertex (cf. Proposition 6.3). For a thorough exposition (in fact equicharacteristic) expressed in classical terms of tangent sheaves one may take a look at [Art76, Section 11 and 12]. def We begin with a proposition expressing the results of [Art76, Theorem 12.1, p. 48; Lemma 12.1, p. 53]. −→ DefC defined Proposition 6.6. There exists a morphism of deformation functors φ : HilbX,Pn by performing cone construction relatively. The tangent and obstruction mappings of φ satisfy the properties: φ i) The tangent mapping Tφ : THilb X,Pn → TDefC can be identified with the canonical homomor- phism H 0(X,NX/ Pn) → Coker Mk∈Z H 0(X,TPn X (k)) →Mk∈Z H 0(X,NX/ Pn (k))! , coming from the long exact cohomology sequence associated to 0 −→Mk∈Z TX (k) −→Mk∈Z TPn X(k) −→Mk∈Z NX/ Pn (k) −→ 0, iii) The obstruction mapping Obφ is injective. ii) The tangent mapping Tφ is surjective ifLk6=0 H 1(X,TX (k)) = 0. Proof. For the proof, we refer to [Art76]. The identification of the tangent space TDefC = Ext1(LC/k,OC ) with the given cokernel follows from an explicit calculation of the long exact sequence of Ext(−,OC ) groups for a distinguished triangle: Li∗LAn+1/k −→ LC/k −→ LC/An+1 −→ Li∗LAn+1/k[1] associated with the inclusion of schemes i : C → An+1. (cid:3) As a corollary we obtain the following result comparing the deformation theory of a projective scheme X and the cone over its sufficiently large ample embedding. Corollary 6.7. Let X be a projective scheme. Then, there exists a sufficiently large Veronese embedding of X ⊂ Pn into PNd such that the morphism of functor φd : HilbX,PNd → Def Cone is smooth. X,PNd Proof. Firstly, by consideration following Proposition 6.3 we see that by taking sufficiently large Veronese embedding we may assume that the cone is in normal. This allows us to apply Proposi- tion 6.6. By part ii), we need to show that for sufficiently large d we haveLk6=0 H 1(X,TX (kd)) = 0. This in turn follows from Serre's vanishing and Serre duality. (cid:3) By combining Lemma A.21 with Corollary 6.7 we obtain: Proposition 6.8. For any smooth projective scheme X satisfying H 2(X,OX ) = 0 and a suffi- ciently large d, the morphisms of deformation functors φd and ψd given in a diagram: are smooth. Def Cone X,PNd φd←− HilbX,PNd ψd−−→ DefX , 26 MACIEJ ZDANOWICZ As a simple corollary, we see that a characteristic p projective scheme X admits a W2(k)-lifting if and only if an appropriate cone does. 6.2.1. Frobenius liftability of cones over sufficiently ample embeddings. We shall now relate the Frobenius liftability of a normal projective scheme and a cone over its sufficiently ample embedding. Proposition 6.9. Let X be a normal projective scheme. Then, for a sufficiently ample embedding X ⊂ Pn the Frobenius liftability of C = ConeX,Pn implies the Frobenius liftability of X. Moreover, for any smooth scheme X satisfying H 1(X,OX ) = H 2(X,OX ) = 0 Frobenius liftability of X is equivalent to the Frobenius liftability of a cone over a sufficiently ample embedding. Proof. The first part we proceed as follows. Let C = ConeX,Pn be the cone over an embedding of X such that the functor φ : HilbX,Pn → DefC is smooth (cf. Corollary 6.7). Moreover, let U be the cone without the vertex m. Assume C is Frobenius liftable, i.e, there exists a lifting eC ∈ DefC (W2(k)) admitting a lifting of Frobenius. By the smoothness of φ we obtain a (potentially non-unique) diagram: C / eC ^❃❃❃❃❃❃❃❃ _❅❅❅❅❅❅❅❅ U p X eU ep / eX. By Lemma 4.1 we see that the obstruction classes to existence of a Frobenius lifting for X and U fit into a commutative diagram in D(X): F ∗ X/k L X (1)/k σF X d(1) p OX [1] p#[1] Rp∗F ∗ U/k L U (1) /k Rp∗σF U/ / Rp∗OU [1] therefore X is Frobenius liftable. U = 0. By the existence of splitting σF where the right most arrow is a splitting of p# induced by the Gm-bundle structure of p. We X = 0 ad X = 0. By Lemma A.21 and Serre vanishing we obtain an embedding of X into Pn such that σF such that ψ : HilbX,Pn → DefX is smooth and H 1(X,OX (k)) = H 2(X,OX (k)) = 0 for any k 6= 0. assumed that eC admits a lifting a therefore σF For the second, we assume that X is Frobenius liftable, i.e., there exists a lifting eX ∈ DefX (W2(k)) By smoothness of ψ we see that eX arises as an embedded deformation and therefore induces a compatible deformation eC ∈ DefC (W2(k)) of the cone C. By Corollary 4.2 in order to prove that eC admits a Frobenius lift it suffices to infer it for eU. By functoriality of obstructions to lifting Frobenius we obtain a diagram: Lp∗σF X U/kLp(1)∗L F ∗ X (1)/k ≃ Lp∗F ∗ U/kdp(1) F ∗ L X (1)/k X/k F ∗ U/k L U (1)/k σF U Lp∗OX [1] ≃ / OU [1]. We consider a Frobenius pullback of a distinguished triangle of cotangent complexes associated to the Gm-bundle p(1) : U (1) → X (1): U/kLp(1)∗L F ∗ X (1)/k −→ F ∗ U/k L U (1)/k −→ F ∗ U/k L U (1)/X (1) ≃ OU −→ F ∗ U/kLp(1)∗L X (1)/k[1], where the isomorphism in the middle follows from the standard property of Zariski locally trivial Gm-bundles. By the long exact sequence of Ext(−,OU ) groups we obtain an exact sequence: X (1)/k,OU ) −→ ··· , . . . −→ Ext1(OU ,OU ) −→ Ext(F ∗ U (1)/k,OU ) −→ Ext1(F ∗ U/kLp(1)∗L U/k L / ^ / / _ / / / Z Z / / / LIFTABILITY OF SINGULARITIES AND THEIR FROBENIUS MORPHISM MODULO p2 27 and therefore it suffices to prove that Ext1(OU ,OU ) = 0. This follows from the isomorphisms Ext1(OU ,OU ) ≃ H 1(U,OU ) ≃ H 1(X,Lk∈Z OX (k)) and the assumptions. (cid:3) Appendix A. Deformation theory Oftentimes we resign from giving proofs and refer to the standard works, e.g, [Ill72, Sch68, FM98]. In the part concerning cotangent complex and obstruction classes, the reader may freely ignore the technical proofs and focus on the statements concerning functoriality. A.1. Deformation theory in terms of cotangent complex. We begin with the relevant topics in deformation theory (see [Ill72, Part I, Chapitre III]) based on the notion of cotangent complex LX/S. We begin with the basic properties of Kodaira-Spencer class KX/Y /S and consequently we express the deformation obstruction classes in terms of KX/Y /S and other natural extensions. Definition A.1 (Kodaira-Spencer class). For any f : X → Y an S-morphism of schemes there exists a mapping KX/Y /S : LX/Y → Lf ∗LY /S[1] in D(X) called a Kodaira-Spencer class. It is defined as the connecting homomorphism in a natural distinguished triangle of cotangent complexes associated to f : Lf ∗LY /S df / LX/S / LX/Y KX/Y /S / Lf ∗LY /S[1]. The Kodaira-Spencer class satisfies the following functoriality property. For any commutative square: X f Y g i / X ′ f ′ / Y ′ defined over a fixed scheme S, the natural diagram: Lg∗LX ′/Y ′ g∗KX′ /Y ′/S / Lg∗Lf ′∗LY ′/S[1] ≃ Lf ∗Li∗LY ′/S[1] dg LX/Y KX/Y /S Lf ∗di / Lf ∗LY /S[1] is commutative, where the right most vertical arrow is the canonical morphism Lf ∗di induced by the equality if ′ = f j. A.1.1. Lifting schemes - functoriality of obstruction classes. We now recall the obstruction theory expressed in terms of cotangent complex. We consider i : Y → eY , a square-zero S-extension of Y by an OY -module J . Apart from functoriality all the results are given in [Ill72]. We begin with a classification of S-extensions. Theorem A.2. There exists a bijection between the group ExalS(Y,J ) of isomorphism classes of S-extension of Y by an OY -module J and the group Ext1(LY /S,J ). Proof. See [Ill72, Part I, Chapitre III, Th´eor`eme 1.2.3]. (cid:3) Definition A.3 (Deformation tuple). A deformation tuple over S consists of: a) a diagram of schemes X f Y S, i xqqqqqqqqqqqqq eY where eY is a square-zero S-extension of Y by a module J . / / / / / / / / / x 28 MACIEJ ZDANOWICZ b) an OX -module I and a module homomorphism w : f ∗J → I. Theorem A.4. For any deformation tuple as above, there exists an obstruction σX ∈ Ext2(LX/Y ,I) whose vanishing is sufficient and necessary for the existence of an extension eX as in the diagram: /❴❴❴❴❴❴ X eX eY f Y S, i xqqqqqqqqqqqqq inducing w : f ∗J → I on the kernels of thickenings. The obstruction σX is given by the composi- tion: KX/Y /S LX/Y / Lf ∗LY /S[1] Lf ∗δ[1] / Lf ∗J [2] H0 / f ∗J [2] w[2] / I[2], where δ ∈ Ext1(LY /S,J ) denotes the cohomology class associated to the extension eY . If the obstruction vanishes, the liftings constitute a torsor under Ext1(LX/S,I). In particular, the ob- struction class for the existence of a flat lifting eX → eY is an element of Ext2(LX/Y , f ∗J ). Proof. For the proof see [Ill72, Part I, Chapitre III, Th´eor`eme 2.1.7]. (cid:3) Liftability obstructions satisfy the following functoriality property (for the sake of convenience we assume the ideals of thickenings are flat). Definition A.5 (Morphism of deformation tuples). A morphism of deformation tuples over S is a piece of data denoted diagramatically by: a morphism of −−−−−−−−−−−−→ deformation tuples  X ′ f ′ i′ x♣♣♣♣♣♣♣♣♣♣♣♣♣ eY ′ Y ′ S (I′,J ′, w′ : f ′∗J ′ → I′),  X f Y S, i xqqqqqqqqqqqqq eY   (I,J , w : f ∗J → I) and consisting of: a) a diagram of schemes over S: whereeh is a morphism inducing an OY -module homomorphism u : h∗J ′ → J , b) a homomorphism v : g∗I′ → I fitting into a commutative diagram: f i X g ~⑥⑥⑥⑥⑥⑥⑥⑥ ⑧⑧⑧⑧⑧⑧⑧⑧ h Y i′ X ′ f ′ Y ′ eY ⑧⑧⑧⑧⑧⑧⑧⑧ eh / eY ′, g∗w′ g∗f ′∗J ′ ≃ f ∗h∗J ′ f ∗u f ∗J w / g∗I′ v / I. / ✤ ✤ ✤ / / x / / / / / / x / / x ~ / / / / / LIFTABILITY OF SINGULARITIES AND THEIR FROBENIUS MORPHISM MODULO p2 29 Lemma A.6 (Functoriality of obstructions to lifting schemes). For any morphism of deformation tuples as above, the obstruction classes to existence of liftings fit in the commutative diagrams: Lg∗LX ′/Y ′ g∗σX′ dg LX/Y σX Lg∗I′ v / I, LX ′/Y ′ σX′ I′[2] dg Rg∗LX/Y Rg∗σX/ / Rg∗I[2]. Proof. We use scommutativity of the diagram: Lg∗ σ X′ Lg∗ L X′/Y ′ g∗K X′ /Y ′ /S / Lg∗Lf ′∗L Y ′/S [1] ≃ Lf ∗Lh∗L Y ′/S [1] Lg∗Lf ′∗J ′[2] Lg∗H0 / Lg∗f ′∗J ′[2] Lg∗ w′ / Lg∗I ′ dg L X/Y Lf ∗ di f ∗ u v◦H0 KX/Y /S / Lf ∗L Y /S[1] / Lf ∗J [2] H0 / f ∗J [2] w / I, σX which follows from: a) the description of obstructions classes, b) functoriality of Kodaira-Spencer class (left-most square), c) commutativity of the lower part of the diagram in the definition of morphism of deformation tuples (middle square), d) the assumption on the homomorphisms of flat ideals (right-most square). We consequently derive the following diagram which is constituted by boundary arrows. Lg∗LX ′/Y ′ g∗σX′ dg LX/Y σX Lg∗I′ v / I. After application of Rg∗ and the natural adjunction this yields the commutativity of: σX′ LX ′/Y ′ ηL X′ /Y ′ Rg∗Lg∗LX ′/Y ′ Rg∗Lg∗σX′ Rg∗dg Rg∗LX/Y Rg∗σX I′[2] ηI′ [2] Rg∗Lg∗I′[2] Rg∗v / Rg∗I[2], where ηA is the counit of the adjunction for an object A. This gives the desired result. (cid:3) A.1.2. Lifting morphisms. Here, we proceed with some facts concerning liftability of morphisms, in particular, we analyse the functoriality properties of obstruction classes. Firstly, we recall the following theorem describing obstructions to lifting of morphisms. Definition A.7 (Deformation of morphism tuple). A deformation of morphism tuple over a square-zero extension Z → eZ by a flat OZ -module K consists of: / / / / / * * / / / / / 2 2 / / / / / / / / / / / / 30 MACIEJ ZDANOWICZ a) a diagram of schemes: i j X f h X ′ g Z eX eX ′ eg / eZ, eh J ). b) a homomorphism w : f ∗J → I satisfying the equality wh = w ◦ f ∗wg where wg : g∗K → J where i (resp. j) is an eZ-extension of X (resp. X ′) by an OX -module I (resp. OX ′ -module and wh : h∗K → I are the natural maps induced byeg andeh, respectively. Theorem A.8. For every deformation of morphism tuple as above, there exists an obstruction σf ∈ Ext1(f ∗LX ′/Z ,I) whose vanishing is sufficient and necessary for the existence of a lifting ef : eX → eX ′ inducing w on the level of ideals of X and X ′. The obstruction is given by the unique preimage under the natural mapping Ext1(f ∗LX ′/Z ,I) −→ Ext1(f ∗L X ′/eZ ,I) of the difference w ◦ f ∗ej − ei ◦ df of arrows in the diagram: f ∗L f ∗ej df f ∗J w L X ′/eS X/eS X ′/eZ ,J ) are classes of an eZ-extension eX, resp. eX ′. X/eZ ,I), resp. ej ∈ Ext1(L / I, ei (cid:3) Proof. For the proof see [Ill72, Part I, Chapitre III, Proposition 2.2.4]. where ei ∈ Ext1(L Morphism liftability obstructions satisfy the following functoriality property. Definition A.9 (Morphism of deformation of morphism tuples). A morphism of deformation of morphism tuples over a square-zero extension S → eS by an OS-module is a piece of data denoted diagramatically by: i′ j′ X ′ f ′ h′ Y ′ g′ Z  eX ′ eY ′ eg′ / eZ  eh′ a morphism of deformation −−−−−−−−−−−−−−−−−→ of morphism tuples  i j X f h Y g Z eX eY eg / eZ  eh (I′,J ′, w′ : f ′∗J ′ → I′) (I,J , w : f ∗J → I) / / ✤ ✤ ✤ / / / / / / / / ✤ ✤ ✤ / / / / / ✤ ✤ ✤ / / / LIFTABILITY OF SINGULARITIES AND THEIR FROBENIUS MORPHISM MODULO p2 31 and given by the diagram of schemes X ′ k ⑦⑦⑦⑦⑦⑦⑦⑦ f ′ eX X f Y ′ l ⑧⑧⑧⑧⑧⑧⑧⑧ ✎✎✎✎✎✎✎✎✎✎✎✎✎✎✎ Y S eX ′ ⑦⑦⑦⑦⑦⑦⑦⑦ ek fY ′ ⑧⑧⑧⑧⑧⑧⑧⑧ ✎✎✎✎✎✎✎✎✎✎✎✎✎✎✎ el eY / eS such that the mapping qk : k∗I → I′ and ql : l∗J → J ′ fit into a commutative diagram: k∗f ∗J ≃ f ′∗l∗J k∗w k∗I f ′∗ql qk / f ′∗J ′ w′ / I′. We need the result in limited generality and therefore for the notational convenience we only state the case of flat extensions over a given eZ. Lemma A.10 (Functoriality of obstructions to lifting morphisms). For any morphism of defor- mation of morphism tuples as above where the associated ideals are flat the obstructions σf and σf ′ satisfy the relation: qk ◦ k∗σf = σf ′ ◦ f ′∗dl ◦ u−1, where u : f ′∗l∗L Proof. We observe that the obstructions fit into the following diagram: X ′/eZ → k∗f ∗L X ′/eZ is the canonical isomorphism. f ′∗l∗ej k∗f ∗ej = k∗ei ei′ / k∗f ∗J ≃ / / f ′∗l∗J k∗w f ′∗ql f ′∗J ′ / k∗I / I′, qk {✇✇✇✇✇✇✇✇✇✇✇ w′ f ′∗l∗L k∗f ∗L u ≃ / f ′∗dl Y /eZ Y ′/eZ &▲▲▲▲▲▲▲▲▲▲ df ′ f ′∗L k∗df Y /eZ X/eZ X ′/eZ k∗L L where the middle square is the pullback of σf along k and the bottom boundary mappings give rise to the obstruction σf ′ . The functoriality is a consequence of a) the description of obstructions classes, b) commutativity of the left square coming from transitivity of differentials for f k = lf ′, c) commutativity of the right square coming from the definition of morphism of deformations of morphism tuples, d) commutativity of the lower and boundary upper square which in fact constitute the obstructions e) flatness of the ideals, which allows us to omit derived pull-backs. for existence ofek andel, respectively. (cid:3) / / ✤ ✤ ✤ ✤ ✤ ✤ ✤ / / ✤ ✤ ✤ ✤ ✤ ✤ ✤ / / / / / / / + + / / & + + / { / 32 MACIEJ ZDANOWICZ A.2. Deformations of affine schemes. In this section we present a few results concerning deformation theory of affine schemes. In particular, we prove that one can check the existence of a W2(k)-lifting ´etale locally. Lemma A.11. Let X ≃ Spec(A) be an affine scheme essentially of finite type over k. Assume there exists an ´etale surjective covering p : U → X such that U is W2(k)-liftable (respectively F -liftable). Then, X is W2(k)-liftable (respectively F -liftable). Proof. We treat the case of W2(k)-liftability as the proof in the case of F -liftability is analogous. By taking an affine Zariski covering of U we may assume that U ≃ Spec(B). Let σA ∈ Ext2(LA/k, A) and σB ∈ Ext2(LB/k, B) be the obstruction classes to the existence of a W2(k)-lifting of Spec(A) and Spec(B), respectively. By Lemma A.6 and the properties of the morphism p we see that Ext2(LB/k, B) ≃ Ext2(LA/k, A) ⊗A B and σB = σA ⊗ 1 under this isomorphism. Therefore, the claim of the lemma follows from the fact that p is ´etale and surjective and hence faithfully flat. (cid:3) Along the same lines we can prove the following. Lemma A.12. Let X be an affine scheme locally of finite type over k. Assume that for any closed point x ∈ X the scheme Spec(OX,x) is W2(k)-liftable (respectively F -liftable). Then, X is W2(k)-liftable (respectively F -liftable). A.3. Deformation functors and their morphisms. Here, we recall the formalism of functors of Artin rings. The classical reference for the topic is [Sch68]. The results concerning obstruction theories are well-explained in [FM98]. Suppose k be a field, Λ a complete local ring and let Art(Λ, k) denote the category of local Artinian Λ-algebras with residue field k. We consider covariant functors F : Art(Λ, k) → Sets, which naturally represent infinitesimal deformations of algebraic and geometric objects. The case we are mostly interested in is Λ = W2(k) where k is a perfect field of positive characteristic p. Definition A.13 (Functor of Artin rings). A functor of Artin rings is a covariant functor F : Art(Λ, k) → Sets such that F (k) consists of a single element. All such functors form a category with natural transformations as morphisms. Any surjective morphism in Art(Λ, k) can be decomposed into a sequence of small extensions, i.e., ring surjections (B, mB) −→ (A, mA) with kernel I satisfying mBI = 0. Small extensions naturally form a category SmallExt(Λ, k) whose objects are small extensions and morphisms are diagrams: 0 0 / I φ / I ′ B A 0 f2 f1 / B′ / A′ / 0, where f1, f2 are rings homomorphisms and φ is an induced homomorphism of k-vector spaces. Definition A.14 (Smooth morphism of functors). A natural transformation ξ : F → G of functors of Artin rings is smooth if for every surjective morphism B → A in Art(Λ, k) the natural mapping: is surjective. F (B) → F (A) ×G(A) G(B) From the special case A = k, we see that a smooth morphism is surjective. Definition A.15 (Tangent space and morphism). Suppose F : Art(Λ, k) → Sets is a functor of = F (k[ε]) is called the tangent space of a functor F . For any morphism Artin rings. The set TF of deformation functors ψ : F → G the mapping Tψ = ψk[ε] : TF = F (k[ε]) → G(k[ε]) = TG is called the tangent map of ψ. def def Under certain condition (see [FM98]), satisfied in any of our applications, the tangent space admits a natural structure of a k-vector space such that the tangent morphism becomes k-linear. Definition A.16 (Obstruction theory). Suppose F : Art(Λ, k) → Sets is a functor of Artin rings. An obstruction theory for F is a pair (V,{νe}e∈SmallExt(Λ)) of a vector space V and a family of mappings νe : F (A) → V ⊗k I parametrised by infinitesimal extensions e : 0 −→ I −→ B −→ A −→ 0 and satisfying the following properties: / / / / / / / / / / / LIFTABILITY OF SINGULARITIES AND THEIR FROBENIUS MORPHISM MODULO p2 33 i) (functoriality) for any morphism of small extensions e : 0 / I B A 0 φ f2 f1 e′ : 0 / B′ / A′ / 0, to a different obstruction theory (V ′, i ◦ νe). ii) (completeness) for any infinitesimal extension e : 0 −→ I −→ B −→ A −→ 0 and an element / I ′ we have (idV ⊗φ) ◦ νe = νe′ ◦ F (f1). a ∈ F(A) the condition νe(a) = 0 is equivalent to existence of b ∈ F(B) lifting a. The choice of obstruction theory is not unique, e.g., any proper inclusion i : V → V ′ gives rise We now present a criterion for smoothness of morphism of functors in terms of their tangent and obstruction spaces. Let (F , (ObF , νF e )) be deformation functors together with associated obstruction theories and ψ : F → G be a morphism. We say that a linear mapping c : ObF → ObG is an obstruction map of ψ if for every infinitesimal extension e : 0 −→ I −→ B −→ A −→ 0 we have νG e . We emphasize that the notion of obstruction map of ψ does not depend solely on ψ but also on the choice of obstruction theories for F and G. e ◦ ψA = (c ⊗k idI ) ◦ νF e )) and (G, (ObG, νG We have the following criterion. Lemma A.17 (Smoothness criterion for morphism of functors). Suppose (F , (ObF , νF e )) and (G, (ObG, νG e )) are deformation functors together with associated obstruction theories. Let ψ : F → G be a morphism of functors admitting an obstruction map Obψ : ObF → ObG. Then, ψ is smooth if the following conditions hold: i) Tψ : TF → TG is surjective, ii) Obψ : ObF → ObG is injective. Proof. See [FM98, Lemma 6.1]. (cid:3) We now give two examples relevant in what follows. Example A.18 (Abstract deformations of a scheme). Suppose X is a scheme over a field k. The functor of abstract deformations DefX : Art(Λ, k) → Sets is defined by the assignment: Art(Λ, k) ∋ A 7→ { isomorphism classes of flat deformations of X over Spec(A) } , where a deformation consists of a flat scheme X over Spec(A) together with an isomorphism φX : X ×Spec(A) Spec(k) → X. Morphism of deformations (X , φX ) → (X ′, φX ′) is a morphism of schemes X → X ′ such that its restriction to the special fibre commutes with isomorphisms φ. By the results of Theorem A.4 the tangent space of the abstract deformation functors DefX is equal to Ext1(LX/k,OX ). Moreover, it admits an obstruction theory with the obstruction space equal to Ext2(LX/k,OX ). Example A.19 (Embedded deformations of a projective scheme). Suppose X ⊂ Pn k is a closed subscheme. The functor of embedded deformations HilbX,Pn : Art(Λ, k) → Sets is defined by the assignment: flat over Spec(A) and restricting to X over Spec(k)) . Art(Λ, k) ∋ A 7→( isomorphism classes of closed subschemes of Pn A By [Har09, Chapter I] the tangent space of HilbX,Pn is naturally isomorphic to H 0(X,NX/ Pn ). Furthermore, it admits an obstruction theory with the obstruction space equal to H 1(X,NX/ Pn ). We now present an exemplary application of Lemma A.17. Example A.20. For any closed embedding X ⊂ Pn of a smooth scheme X/k, we obtain a natural morphism of functors ψ : HilbX,Pn → DefX admitting an obstruction morphism defined by taking the underlying abstract deformation of an embedded deformation. Its tangent and obstruction mappings: Tψ : THilb X,Pn = H 0(X,NX/ Pn ) → H 1(X,TX ) ≃ Ext1(LX/k,OX ) = TDefX Obψ : ObHilbX,Pn = H 1(X,NX/ Pn ) → H 2(X,TX ) ≃ Ext2(LX/k,OX ) = ObDefX / / / / / / / / / / / 34 MACIEJ ZDANOWICZ are given by the canonical morphisms coming from the long exact sequence of cohomology associ- ated to: 0 −→ TX −→ TPn X −→ NX/ Pn −→ 0. As a simple corollary of Lemma A.17 we have the following lemma applicable for example to projective Calabi-Yau varieties of dim X > 3. Lemma A.21. Suppose X is a projective scheme satisfying H 2(X,OX ) = 0. For sufficiently positive embedding of X ⊂ Pn the morphism ψ : HilbX,Pn → DefX is smooth. In particular, every abstract deformation arises as an embedded one. Proof. By Lemma A.17 we need to show that H 0(X,NX/ Pn ) → H 1(X,TX ) is surjective and H 1(X,NX/ Pn ) → H 2(X,TX ) is injective for sufficiently ample embedding X → Pn. By the long exact sequence of cohomology it suffices to show that H 1(X,TPn X ) = 0. By restricting the Euler sequence to X we obtain: By another long exact sequence of cohomology we see that: 0 −→ OX −→ OX (1)⊕n+1 −→ TPn X −→ 0. ··· −→ H 1(X,OX (1)⊕n+1) −→ H 1(X,TPn X ) −→ H 2(X,OX ) −→ ··· , which gives the claim by the Serre's vanishing, i.e., for any coherent sheaf F on X we have H 1(X,F (1)) = 0 if the embedding X → Pn is sufficiently ample. Remark A.22. In particular, for sufficiently positive Veronese embedding of a smooth scheme X satisfying H 2(X,OX ) = 0, the mapping of functors ψd : HilbX,PNd → DefX is smooth. (cid:3) We finish with a lemma comparing abstract deformation functors of a scheme X and its open subset U . We give a proof based on the cotangent complex. For the classical proof the reader is encouraged to see [Art76, Proposition 9.2]. Lemma A.23. Let j : U → X be the inclusion of an open subset. Assume, that X satisfies property S3 at any point p of the complement Z = X \ U . Then, the natural morphism of deformation functors χj : DefX → DefU coming from the restriction is smooth. Proof. By the long exact sequence of Ext(LX/k,−) groups associated to a distinguished triangle: OX −→ Rj∗OU −→ Kj −→ OX [1] we obtain a sequence: . . . −→ Ext0(LX/k, Kj) −→ Ext1(LX/k,OX ) −→ Ext1(LX/k, Kj) −→ Ext2(LX/k,OX ) tj−→ Ext1(LX/k, Rj∗OU ) ≃ Ext1(LU/k,OU ) −→ oj−→ Ext2(LX/k, Rj∗OU ) ≃ Ext2(LU/k,OU ) −→ . . . , where the mappings tj and oj can be identified with Tχj and Obχj . Therefore, by Lemma A.17, it suffices to prove that Ext1(LX/k, Kj) = 0. This is a direct consequence of the spectral sequence Extp(LX/k,Hq(Kj)) ⇒ Extp+q(LX/k, Kj) and the following lemma: Lemma A.24 (Local cohomology in terms of Kj). For any i ∈ Z the sheaf Hi(Kj) is isomorphic ^ Z (U,OU ). Consequently, if X satisfies property Sn+1 at to the local cohomology sheaves any point of Z = X \ U of codimension at least n + 1 then Hi(Kj) = 0 for i ≤ n. Proof. For the proof see [Sta15, Tag 0A39] and [HK04, Proposition 3.3]. U 7→ H i+1 (cid:3) This finishes the proof of Lemma A.23. (cid:3) Acknowledgement I would like to thank my advisor Professor Adrian Langer for the choice of the research topic and numerous guiding suggestions. Moreover, I am grateful to Agnieszka Bodzenta, Joachim Jelisiejew, Lukasz Sienkiewicz and especially Piotr Achinger for their support and many useful discussions. I was supported by Polish National Science Centre (NCN) contract number 2014/13/N/ST1/02673. LIFTABILITY OF SINGULARITIES AND THEIR FROBENIUS MORPHISM MODULO p2 35 References [AC11] K. Rulling A. Chatzistamatiou. Higher direct images of the structure sheaf in positive characteristic. Algebra Number Theory, 5:693–775, 2011. [AM69] M. F. Atiyah and I. G. MacDonald. Introduction to commutative algebra. Addison-Wesley-Longman, 1969. [Art76] M. Artin. Lectures on deformations of singularities, 1976. [Art77] M. Artin. Coverings of the Rational Double Points in Characteristic p. Complex Analysis and Algebraic Geometry, 1977. [Bha12] Bhargav Bhatt. p-adic derived de Rham cohomology, 2012, arXiv:1204.6560. [Bha14] Bhargav Bhatt. Torsion in the crystalline cohomology of singular varieties. Documenta Mathematica, 19(22), 2014. [BK05] M. Brion and S. Kumar. Frobenius Splitting Methods in Geometry and Representation Theory. Progress in Mathematics. Birkhauser Boston, 2005. [Blo77] Spencer Bloch. Algebraic K-theory and crystalline cohomology. Publications Math´ematiques de l'Institut des Hautes ´Etudes Scientifiques, 47(1):188–268, 1977. [BO78] P. Berthelot and A. Ogus. Notes on Crystalline Cohomology. Mathematical Notes - Princeton University Press. Princeton University Press, 1978. [DI87] Pierre Deligne and Luc Illusie. Relevements modulo p2 et decomposition du complexe de de rham. Inven- tiones mathematicae, 89:247–270, 1987. [FM81] W. Fulton and R. MacPherson. Categorical Framework for the Study of Singular Spaces. Number no. 243 in American Mathematical Society: Memoirs of the American Mathematical Society. American Mathematical Society, 1981. [FM98] Barbara Fantechi and Marco Manetti. Obstruction calculus for functors of artin rings, i. Journal of Algebra, 202(2):541 – 576, 1998. [Gla96] Donna Glassbrenner. Strong f-regularity in images of regular rings. Proceedings of the American Mathe- matical Society, 124(2), 1996. [Har09] R. Hartshorne. Deformation Theory. Graduate Texts in Mathematics. Springer New York, 2009. [HK04] Brendan Hassett and S´andor J Kov´acs. Reflexive pull-backs and base extension. Journal of Algebraic Ge- ometry, 13(2):233–248, 2004. [Ill72] Luc Illusie. Complexe cotangent et d´eformations I and II. Lecture Notes in Mathematics, Vol. 239 and 283. Springer-Verlag, Berlin, 1971/1972. [Ill96] Luc Illusie. Frobenius et d´eg´en´erescence de Hodge. In: Introduction la th´eorie de Hodge, volume 3 of Panoramas et Synth`eses, pages 113–168. Soc. Math. France, Paris, 1996. [Jos07a] Kirti Joshi. Exotic torsion, Frobenius splitting and the slope spectral sequence. Canadian Mathematical Bulletin, 50(4):567–587, December 2007. [Jos07b] Kirti Joshi. Exotic torsion, Frobenius splitting and the slope spectral sequence. Canadian Mathematical Bulletin, 50(4):567–587, December 2007. [Kra05] Christian Krattenthaler. Advanced Determinant Calculus: A Complement. , 2005. [Lan08] Adrian Langer. D-affinity and Frobenius morphism on quadrics. International Mathematics Research No- tices, 145:1–26, 2008. [Lan15] Adrian Langer. Bogomolov's inequality for Higgs sheaves in positive characteristic. Inventiones mathemat- icae, 199(3):889–920, 2015. [LS14] Christian Liedtke and Matthew Satriano. On the birational nature of lifting. Advances in Mathematics, 254:118–137, 2014. [MS87] V. B. Mehta and V. Srinivas. Varieties in positive characteristic with trivial tangent bundle. Compositio Mathematica, 64(2):191–212, 1987. [Muk13] Shigeru Mukai. Counterexamples to Kodaira's vanishing and Yau's inequality in positive characteristics. Kyoto J. Math., 53(2):515–532, 2013. [OV07] Arthur Ogus and Vadim Vologodsky. Nonabelian Hodge theory in characteristic p. Publications math´ematiques. Institut des Hautes Etudes Scientifiques, 2007. [Ray78] Michel Raynaud. Contre-example au "vanishing de Kodaira" sur une surface lisse en caract´eristique p > 0, pages 273–278. C. P. Ramanujam - A Tribute. Springer Verlag, 1978. [Sch68] Michael Schlessinger. Functors of artin rings. TRANS. A.M.S, 130(2):208–222, 1968. [Sch71] Michael Schlessinger. Rigidity of quotient singularities. Inventiones mathematicae, 14(1):17–26, 1971. [Sin98] Anurag Singh. PhD Thesis: F-regularity, F-rationality and F-purity, 1998. [Sta15] The Stacks Project Authors. Stacks Project. , 2015. [TW15] S. Takagi and K-I. Watanabe. f -singularities: applications of characteristic p methods to singularity theory, 2015, arXiv:1409.3473. Instytut Matematyki UW, Banacha 2, 02-097 Warszawa, Poland E-mail address: [email protected]
1002.2097	1	1002	2010-02-10T14:07:39	On the connection between fundamental groups and pencils with multiple fibers	[ "math.AG", "math.CV", "math.GR", "math.GT" ]	We present two results about the relationship between fundamental groups of quasiprojective manifolds and linear systems on a projectivization. We prove the existence of a plane curve with non-abelian fundamental group of the complement which does not admit a mapping onto an orbifold with non-abelian fundamental group. We also find an affine manifold whose irreducible components of its characteristic varieties do not come from the pull-back of the characteristic varieties of an orbifold.	math.AG	math	ON THE CONNECTION BETWEEN FUNDAMENTAL GROUPS AND PENCILS WITH MULTIPLE FIBERS ENRIQUE ARTAL BARTOLO AND JOS´E IGNACIO COGOLLUDO-AGUST´IN Introduction The study of the topology of complex projective (or quasiprojective) smooth varieties depends strongly on the knowledge of the topology of the complement of hypersurfaces in a projective space. Considering a projection, any smooth projective variety is a covering of a projective space of the same dimension ramified along a hypersurface. These coverings are measured by (finite index subgroups of) the fundamental group of the complement of the hypersurface. Using Lefschetz-Zariski theory, if we take a generic plane section the fundamental group of the complement does not change. As a consequence, for fundamental group purposes, one can restrict their attention to the study of complements of curves in the projective plane, as stated in the foundational paper by O. Zariski [24]. The richness of coverings for a space depends on its fundamental group. This is why we are mostly interested in curves C ⊂ P2 whose π1(P2 \ C) is non-abelian. The first known example is probably the curve formed by three lines C := L1 ∪ L2 ∪ L3 intersecting at one point P . There is an easy way to compute this fundamental group; the pencil of lines through P is parametrized by P1; this pencil induces an epimorphism of P2 \ C onto P1 \ {p1, p2, p3} (the punctures corresponding to the three lines). Moreover, this map is a locally trivial fibration (with fiber isomorphic to C) and hence π1(P2 \ C) ∼= π1(P1 \ {p1, p2, p3}), which is a free group of two generators. The first known examples of irreducible curves whose fundamental groups are known to be non-abelian appeared in [24]. The first one corresponds to a hexacuspidal sextic, with its six cusps on a conic; the equation of such a curve is of the form f 3 3 = 0, where fj is a Its fundamental group is Z/2 ∗ Z/3; in §2 we will see homogeneous polynomial of degree j. the relation between this group and the pencil generated by f 3 3 = 0. This kind of examples have been generalized by various authors replacing (2, 3) by (p, q). In the same paper, Zariski found the irreducible curve with smallest possible degree having a non-abelian fundamental group: the tricuspidal quartic. This example and many others appearing in the literature are also connected with pencils. 2 = 0 and f 2 2 + f 2 The precise connection with pencils can be stated as follows: a pencil defines a dominant morphism to a quasi-projective curve, inducing an epimorphism at the level of fundamental Date: March 8, 2021. Partially supported by MTM2007-67908-C02-01. 1 2 E. ARTAL AND J.I. COGOLLUDO groups. The multiplicities of the fibers of the pencil induce an orbifold structure on the quasi- projective group, and the map defines an epimorphism onto the orbifold fundamental group. When such an orbifold fundamental group is non-abelian, then the original fundamental group has a surjection onto a non-abelian group. Such surjections coming from dominant maps will be referred to as geometric surjections. The tricuspidal quartic is the only irreducible curve of degree 4 with a non-abelian fundamental group. The degree-five case was studied by A. Degtyarev [8]; he found exactly two irreducible quintics with non-abelian fundamental groups. One of them, also studied by the first author [2], In §2, we will study its relationship with a pencil. The has an infinite fundamental group. question whether or not all non-abelian fundamental groups have a geometric surjection onto an orbifold group naturally arises. All the examples studied, up to now, supported an affirmative answer to this question. In this paper we will show an explicit example of a non-abelian fundamental group whose complement admits no geometric surjections. This curve is one of the quintics referred to in the previous paragraph, which will be called the projective Degtyarev curve throughout this text. As a brief description, the projective Degtyarev curve has exactly three singular points of type A4; its fundamental group is finite and non-abelian. In Proposition 4.4, we prove that this group admits no geometric surjections. Once the group is computed, the proof is rather straightforward; it depends on the orders of the group and its abelianization and on the properties of orbifold groups. If we add a tangent line to one of the singular points of the projective Degtyarev curve, the complement of the union in P2 is the complement of an affine curve, which will be called the affine Degtyarev curve. This affine curve has an infinite non-abelian fundamental group and non-trivial characteristic varieties (see §1 for the definition). Extending results of Arapura and others, it is known that irreducible components of positive dimension (for the fundamental group of a quasiprojective variety) are obtained as pull-back of irreducible components of characteristic varieties of orbifolds. A natural question arises: Is it also true for irreducible components of dimension 0 (isolated points)? Plenty of computations supported a positive answer: most quasiprojective groups satisfy the property for irreducible components of any dimension (see §2 for examples). The main Theorem 4.5 of this paper shows that the fundamental group of the complement of the affine Degtyarev curve does not satisfy this property. This is the only known example, up to now. The paper is organized as follows. In §1, the concepts of orbifold and characteristic varieties are recalled, also some orbifold groups are studied. In §2, we relate non-abelian fundamental groups of the complements of curves (which are known in the literature) with orbifold morphisms (via pencils of curves). In §3, we describe Degtyarev curves and, in order to obtain a prresentation for their fundamental groups, we compute a special braid monodromy. The fundamental groups are obtained in §4, where also the main results of the paper are stated and proved. Finally, further properties of the affine Degtyarev curve are sketched in §5. FUNDAMENTAL GROUPS AND PENCILS 3 1. Orbifold groups and characteristic varieties The fundamental groups of oriented Riemann surfaces have been extensively studied. The fundamental group of a compact Riemann surface of genus g is (cid:42) πg := ai, bi, 1 ≤ i ≤ g aibia−1 i b−1 i . (cid:43) (cid:12)(cid:12)(cid:12)(cid:12)(cid:12) g(cid:89) i=1 If C is a surface with genus g and k > 0 punctures then its fundamental group is free of rank 2g + k − 1. We are going to extend this family by considering orbifold groups. Definition 1.1. An orbifold Xϕ is a quasiprojective Riemann surface X with a function ϕ : X → N with value 1 outside a finite number of points. We may think that a neighborhood of a point P ∈ Xϕ such that ϕ(P ) = n is the quotient n . We will consider that a loop around P is of a disk (centered at P ) by a rotation of angle 2π trivial if its lifting bounds a disk. Following this idea, we define orbifold fundamental groups. Definition 1.2. For an orbifold Xϕ, let p1, . . . , pn the points such that mj := ϕ(pj) > 1. Then, the orbifold fundamental group of Xϕ is πorb 1 (Xϕ) := π1(X \ {p1, . . . , pn})/(cid:104)µmj j = 1(cid:105) where µj is a meridian of pj. We denote Xϕ by Xm1,...,mn . Example 1.3. If X is a compact surface of genus g and type Xm1,...,mn , then (cid:42) πorb 1 (Xϕ) = a1, . . . , ag, b1, . . . , bg, µ1, . . . , µn If X is not compact and π1(X) is free of rank r, then (cid:42) πorb 1 (Xϕ) = a1, . . . , ar, µ1, . . . , µn (cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) g(cid:89) i=1 n(cid:89) (cid:43) j=1 [ai, bi] = (cid:12)(cid:12)(cid:12)(cid:12)(cid:12)µmj j = 1 j=1,...,10 (cid:43) µj, µmj j = 1 j=1,...,10 . . Definition 1.4. A dominant algebraic morphism ρ : Y → X defines an orbifold morphism Y → Xϕ if for all p ∈ X, the divisor ρ∗(p) is a ϕ(p)-multiple. Proposition 1.5. Let ρ : Y → X define an orbifold morphism Y → Xϕ. Then ρ induces a morphism ρ∗ : π1(Y ) → πorb (Xϕ). Moreover, if the generic fiber is connected, then ρ∗ is surjective. Proof. Let Mϕ := {x ∈ X ϕ(x) > 1}; we consider the restriction mapping ρ := ρ : Y \ ρ−1(Mϕ) → X \ Mϕ. This map induces a morphism ρ∗ : π1(Y \ ρ−1(Mϕ)) → π1(X \ Mϕ) fitting in the following commutative diagram: 1 π1(Y \ ρ−1(Mϕ)) i∗ ↓ π1(Y ) ρ∗−→ π1(X \ Mϕ) ↓ j∗ π1(X). ρ∗−→ 4 E. ARTAL AND J.I. COGOLLUDO The vertical mappings are induced by the inclusions. They are both surjective; the kernel of j∗ is generated by the meridians of the points in Mϕ while the kernel of i∗ is generated by the meridians of the irreducible components of ρ−1(Mϕ), i.e., the components of the pull-back divisor ρ∗(Mϕ). Let us consider an irreducible component D of ρ∗(Mϕ) such that ρ(D) =: x ∈ Mϕ. Let n := ϕ(x); note that the multiplicity mD of D in ρ∗(Mϕ) is a multiple of n. We can interpret mD as follows. If µD denotes a meridian of D, then there is a meridian µx of x such that ρ∗(µD) = (µx)mD . Following Definition 1.2, it is easily seen that ρ∗ factorizes through a morphism (also called ρ∗) π1(Y ) → πorb The above argument also works if one replaces Mϕ by a finite set M ⊇ Mϕ. In particular, one can choose M to be the bifurcation locus of ρ, i.e., the mapping is a differentiable locally trivial fibration outside M . If the fiber is generically connected, the long exact homotopy sequence of (cid:3) this fibration implies the surjectivity of ρ∗ (for M ). The result follows. (Xϕ). 1 Definition 1.6. A fundamental group G := π1(Y ) is said to posses a geometric surjection if Y possesses an orbifold morphism Y → Xϕ whose generic fiber is connected, and such that πorb (Xϕ) is non-abelian. 1 We recall the notion of characteristic varieties and its relationship with orbifolds. We focus our attention on the characteristic varieties of quasiprojective manifolds, though they can be defined in general and depend only on the fundamental group. Let X be a connected topological space X, having the homotopy type of a finite CW -complex, and let G := π1(X, x0), x0 ∈ X which will be omitted if it is not necessary. Recall that the space of characters of G is H 1(X; C∗) = Hom(H1(X; Z), C∗) = Hom(π1(X), C∗) =: TG. (1.1) Remark 1.7. Since G is finitely generated, then it is also the case for H1(X; Z). Let n := rk H1(X; Z) and TorsG be the torsion subgroup of H1(X; Z). Then TG is an abelian complex Lie group with TorsG connected components (each one is isomorphic to (C∗)n) satisfying the following exact sequence: 1 → T1 G → TG → TorsG → 1, where T1 G is the connected component containing the trivial character 1. For a character ξ ∈ TG, we can construct a local system of coefficients Cξ over X. Definition 1.8. The k-th characteristic variety of X is the subvariety of TG, defined by: Vk(X) = {ξ ∈ TG dim H 1(X, Cξ) ≥ k}, where H 1(X, Cξ) is the cohomology with coefficients in the local system ξ. In some cases we will use the notation Vk(G). The following result is straightforward. Proposition 1.9. Let ϕ : G → H be a group epimorphism. Then ϕ∗ induces injections TH ≡ ϕ∗TH (cid:44)→ TG and Vj(H) ≡ ϕ∗Vj(H) (cid:44)→ Vj(G). FUNDAMENTAL GROUPS AND PENCILS 5 Remark 1.10. Let us explain how to compute these invariants. For the sake of simplicity, the twisted homology, instead of the cohomology, will be computed. Let us consider a finite CW - complex homotopy equivalent to X. Let π : X → X be the maximal abelian covering. Note that X inherits a CW -complex structure. The group of automorphisms of π is H1(X; Z). The action of this Abelian group endows the chain complex C∗( X; C) with a module structure over the ring Λ := Z[H1(X; Z)]. The differentials of the complex are Λ-homomorphisms. Moreover, C∗( X; C) is a free Λ-module of finite rank. If we fix a character ξ, C has a natural Λ-module structure which is denoted by Cξ (as the local system of coefficients). The twisted homology of X is the homology of the C∗(X; C)ξ := C∗( X; C) ⊗Λ Cξ. Following this interpretation, it is not difficult to prove that the characteristic varieties are algebraic subvarieties of TG, defined with integer equations. This i-th jumping loci of C∗( X; C) with respect to ⊗Λ Cξ can also be viewed as the zero locus of the i-th Fitting ideal of H1( X; C) or, analogously, the support of the module ∧iH1( X; C) over the ring Λ (see [15]). Following the theory developed by various authors (Beauville [6], Arapura [1], Simpson [20], Budur [7], Delzant [10], Dimca [12]), the structure of characteristic varieties for quasiprojective manifolds) can be stated as follows. Theorem 1.11 ([4]). Let Σ be an irreducible component of Vk(G), k ≥ 1. Then one of the two following statements holds: • There exists a surjective orbifold morphism ρ : X → Cϕ and an irreducible component Σ1 of Vk(πorb 1 (Cϕ)) such that Σ = ρ∗(Σ1). • Σ is an isolated torsion point. Remark 1.12. In general, both G and its characteristic varieties are difficult to compute. For the complement of hypersurfaces in a projective space, Libgober [15] gave an alternative way of computing most components of the characteristic varieties from algebraic properties of the hypersurface without computing G. Remark 1.13. Characteristic varieties can also be understood from Alexander-invariant point of view. Following Theorem 1.11, characteristic varieties are determined by finite-index abelian coverings. We compute the invariants for some orbifold groups. Proposition 1.14. Let G be the orbifold group of P1 2,5,10. Then G is a semidirect product of the fundamental group of a compact surface of genus 2 and Z/10. The torus TG is µ10, the group of 10-th roots of unity, V1(G) consists of the primitive 10-th roots of unity and V2(G) = ∅. Proof. Let us consider the short exact sequence associated with the abelianization map (Z/10 := (cid:104)t t10 = 1(cid:105) is G/G(cid:48)). This sequence corresponds to a uniformization covering of the orbifold, and using Riemann-Hurwitz one obtains a compact Riemann surface of genus 2. Since the exact sequence splits, we have a semidirect-group structure. 6 E. ARTAL AND J.I. COGOLLUDO In order to compute V1(G) we follow the construction outlined in Remark 1.10, applied to the CW -complex associated with the presentation of G given by (cid:104)x, y x2 = y5 = (xy)10 = 1(cid:105). Let us denote p the unique 0-cell, x, y the 1-cells and A, B, C the 2-cells (corresponding to the relations in the given order). Let us fix a character ξ ∈ TG. It is clear that 1 /∈ V1(G). We can assume that ζ := ξ(t) (cid:54)= 1. The complex C∗(X; C)ξ is given by 0 −→ C3 ∂2−→ C2 ∂1−→ C −→ 0. The matrix for ∂1 is . In particular, dim ker ∂1 = 1. The matrix for ∂2 equals ζ2−1 (cid:16) ζ5−1 (cid:17) ζ 5 + 1 0 0 ζ 8 − ζ 6 + ζ 4 − ζ 2 + 1 ζ 10 − 1 ζ − 1 ζ 5 ζ 10 − 1 ζ − 1  In order to have non-trivial homology, this matrix must vanish and this happens only when ζ is (cid:3) a primitive 10-th root of unity. Proposition 1.15. Let G be the orbifold group of P1 2,2,5,5. Then G is an extension of Z/10 by the fundamental group of a compact surface of genus 4. The torus TG is µ10, the group of 10-th roots of unity, and both V1(G) and V2(G) consist of the primitive 10-th roots of unity. Proof. The short exact sequence associated with the abelianization map (G/G(cid:48) = Z/10) corre- sponds to a uniformization covering of the orbifold, and using Riemann-Hurwitz one obtains a compact Riemann surface of genus 4. We compute the characteristic varieties as in the proof of Proposition 1.14 for the presentation of G given by (cid:104)x, y, z x5 = y5 = z2 = (xyz)2 = 1(cid:105). Let us denote p the unique 0-cell, x, y, z the 1-cells and A, B, C, D the 2-cells (corresponding to the relations in the given order). Let us fix a character ξ ∈ TG. It is clear that 1 /∈ V1(G). We can assume that ζ := ξ(t) (cid:54)= 1. The complex C∗(X; C)ξ is given by 0 −→ C4 ∂2−→ C3 ∂1−→ C −→ 0. The matrix for ∂1 is the transposed of ( ζ2−1 ζ2−1 ζ5−1 ). In particular, dim ker ∂1 = 2. The matrix for ∂2 equalsζ 8 − ζ 6 + ζ 4 − ζ 2 + 1 0 0 0 ¯ζ 8 − ¯ζ 6 + ¯ζ 4 − ¯ζ 2 + 1 0 0 0 ζ 5 + 1 ζ(ζ 5 + 1) ζ 5 + 1 ζ 5 + 1  In order to have non-trivial homology, this matrix must have rank less than 2 and this happens only when ζ is a primitive 10-th root of unity. Moreover, in that case, the matrix vanishes. (cid:3) 2. Examples In this section, we will present a collection of examples of curves with non-abelian fundamental groups and geometric surjections and its relationship with characteristic varieties. Remark 2.1. If Y := P2 \ C admits an orbifold morphism Y → Xϕ, then the non-singular compactification ¯X of X is P1. FUNDAMENTAL GROUPS AND PENCILS 7 Remark 2.2. The easiest examples of curves with non-abelian fundamental groups and geometric surjections come from hyperplane (or line) arrangements. If a line arrangement A has a point P of multiplicity k ≥ 3, then the pencil of lines through P defines a morphism ρ : P2 \ A → X, where X is a k-punctured projective line. We have an epimorphism ρ∗ : π1(P2 \ A) → π1(X) and the latter is a free group of rank k − 1 (hence non abelian). The following result is well known for specialists. Proposition 2.3. The following three assertions are equivalent: (1) The group π1(P2 \ A) is non abelian, (2) The arrangement A has a point of multiplicity at least 3, (3) The group π1(P2 \ A) has a geometric surjection. Proof. By the remark above, it is obvious that (2) implies (1) and (3). Also, by definition, (3) implies (1). Hence it is enough to prove that (1) implies (2). Note that, if (2) does not hold, then A is an arrangement in general position. Either we choose a particular example (e.g. a real arrangement) and a braid monodromy argument implies immediately the abelianity or we use (cid:3) Hattori's topological description of arrangements of hyperplanes in general position. The argument used in Remark 2.2 can be easily generalized when, instead of considering three (or more) incident lines, one considers three (or more) fibers of any pencil of curves in P2. Of course, any such example corresponds to curves with at least three irreducible components. The notion of orbifold allows for wider generalizations of this concept to curves with any number of irreducible components (for example to irreducible curves). As it was stated in the Introduction, the first example of this kind is rather old, see [24]. Let us consider a conic C2 of equation f2 = 0 and a cubic C3 of equation f3 = 0; we assume they do not 2 − f 2 have common components and they are not multiple lines. Let C be a curve of equation f 3 3 . Note that the mapping ρ : P2 → P1 \ {[1 : 1]} given by [x : y : z] (cid:55)→ [f2(x, y, z)3 : f3(x, y, z)2] is well defined (all the base points of the pencil are in C) and surjective. This mapping induces an orbifold map onto a 1-punctured Riemann sphere with two orbifold points of multiplicities 2 and 3 (at [0 : 1] and [1 : 0] respectively). Thus according to Proposition 1.5, one obtains an epimorphism π1(P2 \ C) onto Z/2 ∗ Z/3. Proposition 2.4. Let G be the orbifold fundamental group of C2,3. Then, TG = µ6, V1(G) consists of the 6-th primitive roots of unity and V2(G) = ∅. In particular, any curve with equation f 3 3 = 0 has non-trivial characteristic varieties. 2 − f 2 Remark 2.5. For generic choices of f2 and f3 this epimorphism is in fact an isomorphism (this is actually the case originally considered by Zariski in [24]). However, this is not the case, for instance, when C is reducible (since b1(P2 \ C) > 1). Even if C is irreducible one may also have not an isomorphism for several reasons: either there are few non-generic fibers in the pencil (e.g., a sextic with six cusps and four ordinary nodes) or there are several pencils (a sextic with nine cusps). 8 E. ARTAL AND J.I. COGOLLUDO These examples can be generalized if we replace (2, 3) by any coprimes (p, q), see Oka [19], In such cases, the fundamental group of a generic curve with q = 0 is Z/p ∗ Z/q. Also Zariski [24] considered another interesting example N´emethi [18] and Dimca [11]. equation f q where the target orbifold is compact. p + f p Let us consider the tricuspidal quartic C4 with equation f4 = 0. It is not hard to prove that we can choose (2.1) and Sing(C4) = {[1 : 0 : 0], [0 : 1 : 0], [0 : 0 : 1]}. The curve C4 is parametrized by f4 := x2y2 + y2z2 + x2z2 − 2xyz(x + y + z). [t : s] (cid:55)→ [t2s2 : (t − s)2s2 : t2(t − s)2]; (2.2) and its singular points correspond to [t : s] = [0 : 1], [1 : 1], and [1 : 0]. Let P ∈ C4 be a smooth point with parameter α ≡ [α : 1] and let L be the tangent line to C4 at P , with equation f1 = 0, where f1 := (α − 1)3x − α3y + z. Let C2 be the conic passing through the singular points of C4 and tangent to C4 at P ; since we have imposed five (non-degenerate) conditions, such a conic is unique. As before, let f2 = 0 be the equation of C2, where f2 := α(α − 1)xy − (α − 1)xz + αyz. We consider now a cubic C3 having a nodal point at P (one of the branches tangent to C4 at P ) and tangent to C4 at the three cuspidal points. Counting the conditions it is easy to prove that only one such cubic exists, with equation f3 = 0, where f3 := − (α − 2) (2 α − 1) (α + 1) xyz − α3xy2 − xz2 − (α − 1)3 x2y + yz2 + (α − 1)3 x2z + α3y2z. 3 − 4f 3 2 . 1 = f 2 Lemma 2.6. f4f 2 Proposition 2.7. The fundamental group of P2\C4 possesses a geometric surjection onto P1 2,2,3. Remark 2.8. Zariski proved in [24] that π1(P2 \ C4) is non-abelian of order 12. The above mapping induces a central extension of D6 (dihedral group of order 6) whose kernel is cyclic of order 2. Note that we also have an epimorphism from π1(P2 \ (C4 ∪ L)) onto the orbifold group of a 1-punctured Riemann sphere with two multiple points (2, 3); for a generic P it is possible to prove that π1(P2 \ (C4 ∪ L)) equals B3. There is a particular case corresponding to the bitangent line; in this case there are two such mappings and π1(P2 \ (C4 ∪ L)) is the Tits-Artin group of a triangle. In [8], Degtyarev proved that only two irreducible curves of degree 5 have non-abelian funda- mental group. One of them is extensively studied in §3. The other one was also studied by the first author in [2]. It is a rigid curve with one point of type A6 and three cuspidal points (it is the dual curve of the quartic with one A6). Let C5 be this curve (with equation f5 = 0). The way to prove that this group is non-abelian in [2] is to show that there is an epimorphism from FUNDAMENTAL GROUPS AND PENCILS 9 E1 L E2 T C C Figure 1. Cremona transformation an actual presentation of π1(P2 \ C5) onto the triangle group of type 2, 3, 7; this is the orbifold group of P1 Proposition 2.9. The fundamental group of P2\C5 possesses a geometric surjection onto P1 ϕ with three multiple points of these orders. In fact, one has the following: 2,3,7 Proof. We start with a pencil as in Lemma 2.6, where L is the bitangent. We are going to consider the Cremona transformation ρ : P2 (cid:57)(cid:57)(cid:75) P2 associated with the net of conics having three infinitely near points in common with C at P , the singular point of type A6. Let us describe this transformation. We blow-up this three infinitely near points and we obtain a rational surface X with a morphism σ1 : X → P2. Let us denote the three exceptional components (in order of appearance) by E1, E2, and T , and finally the tangent line of C at P by L, see Figure 1. Convention 2.10. For birational morphisms, we keep the notation of a curve for its strict transform unless otherwise stated. In X we have E1 · E1 = −2 and E1 · E1 = T · T = L · L = −1. Since we can interchange L and T , we consider the birational morphism σ2 : X → P2 obtained as the composition of the contractions of L, E2 and E1; the contracted surface is rational with Euler characteristic 3 and 2 ◦ σ1. Let us denote C := ρ(C). hence it is a projective plane. It is not hard to prove that ρ = σ−1 Note that C is a tricuspidal quartic and T is its unique bitangent line, one point P comes from the infinitely near point of C at P and the other one Q comes from the other intersection point of C and L. We consider the pencil defined by the orbifold map of Proposition 2.7, where the base point is P . Let C3 be the cubic such that 2C3 is in the pencil. Following C3 by σ2 and σ1, C6 := ρ∗(C3) is a sextic with only one singular point at P (with two branches, one of type A6 and a smooth branch with maximal contact with the singular branch). With the same ideas, if C2 is the conic such that 2C3 is in the pencil, then C4 := ρ∗(C2) is a quartic with an A6 singular point at P . Finally ρ∗( C + 2T ) = C + 7L. We have a pencil of degree 12 containing the fibers 2C6, 3C4 (cid:3) and C + 7L. This pencil produces the desired morphism. One can find more examples in the literature: Degtyarev [8], Flenner-Zaıdenberg [13], and Tono [22]. In what follows, the last two families will be described. We start with some definitions. 10 E. ARTAL AND J.I. COGOLLUDO Definition 2.11. A Hirzebruch surface is a rational surface X with a morphism π : X → P1 which is a holomorphic (or algebraic) fibration with fiber P1. Such a surface is either Σ0 := P1×P1 or it has a unique section Sn with negative self-intersection −n, n > 0; in that case π is unique and X is denoted by Σn (Σ1 is the blowing-up of one point in P2). For any Hirzebruch surface X there is a family of birational maps which are called elementary Nagata transformations. They are obtained as follows. Let us consider π : X → P1, P ∈ X and F := π−1(π(P )); we consider the blowing-up σ : X → X of P , with exceptional component F . Since (F · F )X = 0, we have that (F · F ) X = −1. By Castelnuovo criterion, we can blow down F and we obtain a new Hirzebruch surface X where F is a fiber. Definition 2.12. An elementary Nagata transformation is said to be positive (resp. negative) if P belongs (resp. does not belong) to a section with non-positive self-intersection. For a positive one, one goes from Σn to Σn+1; for a negative one, from Σn to Σn−1. In [3], the first author computed the fundamental group of Flenner-Zaıdenberg curves and showed when it is non-abelian using orbifold groups. We show here that this can also be ge- ometrically proved. In order to construct these curves, we start with a smooth conic C with two tangent lines L1 and L2, intersecting at some point P . After blowing-up P one obtains π : Σ1 → P1 with exceptional component E. Let L3 be another line in the pencil through P which intersects C at two points Q1 and Q2. Let us fix two positive integers a, b. After perform- ing a negative elementary Nagata transformations at the point corresponding to the fiber of L1 and b at the point corresponding to the fiber of L2 one obtains a Hirzebruch surface Σa+b+1. One can then perform a + b negative elementary Nagata transformations at the point Q2 (it corresponds to the fiber L3). After this process, E can be blown down which turns our surface into P2. The curve Ca,b obtained has degree d := a + b + 2 and three singular points of type A2a, A2b, and a third one with local equation ud−2 = vd−1. Proposition 2.13. The fundamental group of P2 \ Ca,b possesses a geometric surjection onto P1 2,a+b,c, where c := gcd(2a + 1, 2b + 1). Proof. It is enough to follow the pencil of conics generated by L1 + L2 and C through the above transformations. We obtain a pencil of curves of degree 2(d − 1), where one fiber is (2a + 1) L1 + (2b + 1) L2 (they are the lines corresponding to the fibers of L1 and L2. The fiber containing Ca,b is of the form Ca,b + (d − 2) L3. Finally the double line in the pencil becomes a double curve of degree d − 1. (cid:3) In [22], K. Tono describes all rational unicuspidal curves such that its complement in P2 has logarithmic Kodaira dimension 1. The construction given in [22, Theorem 1] shows that the complement of these curves have non-abelian fundamental group. Any other known rational unicuspidal curve has abelian fundamental group (for the complement). FUNDAMENTAL GROUPS AND PENCILS 11 Proposition 2.14. For any Tono's curve C their fundamental group possesses a geometric surjection onto P1 intersection of the strict transform of C after the minimal embedded resolution of its unique singular point. This number is at least 2. µA,µG,n(C), where µA, µG ≥ 2 and the number n(C) is the opposite of the self- Proof. It is enough to consider the construction of [22, Theorem 1] where a pencil is obtained with two multiple fibers µAA and µGG and a reducible fiber of the form C + n(C)B, where B (cid:3) is either a line (type I) or a smooth conic (type II). Example 2.15. The curves of type I are parametrized by two integers n, s ≥ 2. The curve C has degree (n + 1)2(s − 1) + 1, where n(C) = n, µA = n + 1 and µG = (n + 1)(s − 1) + 1. For n = s = 2, we obtain the multiplicities 2, 3, 4; in fact, one can compute that this group is finite. 3. Degtyarev curves Let us consider a projective Degtyarev curve, i.e., a plane projective curve of degree 5 such that Sing(C) consists of three points, and for each point P ∈ Sing(C) the germ (C, P ) is topologically equivalent to an A4-singularity, i.e. with local equation v2 − u5 = 0; note that in this case, the germs are also analytically equivalent. Most of the following properties appear in [8] and [17], but we include for the sake of com- pleteness. Properties 3.1. Let C ⊂ P2 be a projective Degtyarev curve. Then: (D1) The curve C is irreducible. (D2) The tangent line L of C at a singular point P satisfies (L · C)P = 4. (D3) Two Degtyarev projective curves are projectively equivalent. (D4) The subgroup of projective transformations preserving C is cyclic of order 3. (D5) The curve C is autodual. Proof. Since the three singular points are locally irreducible, (D1) is true. For (D2), note that 4 ≤ (L · C)P ≤ 5. Let us assume that (L · C)P = 5; considering L as the line at infinity, C \ L is an affine curve homeomorphic to C. This case is discarded using Zaıdenberg-Lin Theorem [23] and (D2) results. In order to prove (D3), there are two approaches. The direct approach consists of computing the equations of the curve C fixing the position of the singular points and some of their tangent lines. The second method is quite simple and worth describing here: Let C1, C2 be two projective Degtyarev curves. By B´ezout's Theorem, the singular points are not aligned; and hence, after a projective transformation, one may assume that Sing(C1) = Sing(C2) =: S. Assuming that S := {[1 : 0 : 0], [0 : 1 : 0], [0 : 0 : 1]}, one can perform a standard Cremona transformation ψ : P2 (cid:57)(cid:57)(cid:75) P2 based on the three singular points and defined by ψ([x : y : z]) = [yz : xz : xy]. Geometrically, this rational map is obtained by blowing-up the three vertices of S (obtaining a rational surface X(cid:96)) and then blowing down the strict transforms of the lines joining the points of S (which have self-intersection −1 in X(cid:96)). One can easily compute that Ci := ψ(Ci) is a tricuspidal quartic. It is well known that there is only one tricuspidal quartic, up to projective 12 E. ARTAL AND J.I. COGOLLUDO transformation, therefore, after a suitable change of coordinates, one may assume C1 = C2 =: C, where C is the curve with equation given in (2.1). The tricuspidal quartic satisfies the following properties. Let Sing( C) = {P1, P2, P3}; there are three points Q(cid:96) 3 ∈ C, (cid:96) = 1, 2 such that k are aligned for all the possibilities with #{i, j, k} = 3. Let A(cid:96) be the arrangements Pi, Q(cid:96) of curves given by C and the lines joining Q(cid:96) 2, Q(cid:96) 1, Q(cid:96) j, Q(cid:96) i and Q(cid:96) j. 1, Q(cid:96) 2, Q(cid:96) The curve C is parametrized as in (2.2) and the singular points P1 = [0 : 1 : 0], P2 = [1 : 0 : 0], and P3 = [0 : 0 : 1] correspond to [t : s] = [0 : 1], [1 : 1], and [1 : 0]. It is not hard to check that A(cid:96) := (α(cid:96), 2 + α(cid:96),−α(cid:96)) are affine parameters of (Q(cid:96) 3). The last condition implies that (cid:96) + α(cid:96) − 1 = 0. If α1 = α2 then A1 = A2. α2 The group of projective transformations fixing C is the group of the permutation of the coordinates. The mapping [x : y : z] σ(cid:55)→ [x : z : y] induces [t : s] (cid:55)→ [s : t] in the parametrization, and [x : y : z] τ(cid:55)→ [y : z : x] induces [t : s] (cid:55)→ [s : s − t]. operations on A(cid:96): (−α−1 Let us assume that α1 (cid:54)= α2 Applying the projective transformation σ, results into two the permutation (1, 3), and the change of parameters. Thus, σ(A1) = 1 ) = A2, which implies σ(A1) = A2. 1 , (α1 + 2)−1, α−1 Note that any projective transformation sending A1 to A2 lifts to an isomorphism X1 → X2 and this isomorphism induces a projective transformation of the source P2, hence (D3) results. In order to prove (D4) one can use a similar argument on the projective transformations fixing C (this last property was communicated to the authors by C.T.C. Wall). The property (D5) follows from (D2) and Plucker generalized formulae, see [17]. More pre- cisely, given a curve D and a point P ∈ D, the order of the curve is the degree of its dual curve of D: deg( D) = deg(D)(deg(D) − 1) − (cid:88) (µ(C, P ) − 1 + m(C, P )). P∈D This formula implies that deg( D) = 5. The dual of a singular point of type A4 is either of the same type or of type E8 (in case the tangent line has multiplicity of intersection 5 with the curve (cid:3) at the singular point). Thus (D5) holds. Remark 3.2. Note that any two projective Degtyarev curves are isotopic. Using the direct approach, we can give a symmetric equation: (cid:16) (cid:17) √ 7 + 3 5 (x3z2 + x2y3 + y2z3) + 5 + 6 (x3yz + xy3z + xyz3)+ +2(x3y2 + x2z3 + y3z2) + 33 + 11 (x2yz2 + x2y2z + xy2z2) =0. (cid:17) (cid:16) √ 2 √ (cid:17) 5 (cid:16) Note that the permutation of two variables comes from the Galois transformation in Q( 5). √ The curve also admits an equation with rational coefficients; in that case one of the singular points has rational coordinates but the other two are conjugate in Q( (3.1) z2y3−z(33xz +2x2 +8z2)y2 +(21z2 +21xz−x2)(z2 +11xz−x2)y +(x−18z)(z2 +11xz−x2)2 = 0 5): √ Properties 3.1 imply that the affine Degtyarev curve is also rigid, i.e. any two affine Degtyarev In order to study its curves are projectively equivalent, and in particular, they are isotopic. FUNDAMENTAL GROUPS AND PENCILS 13 F− F0 F+ √ x = 11−5 2 5 = a− x = 0 √ x = 11+5 2 5 = a+ Figure 2. Real picture of the affine Degtyarev curve β− γ− α− β0 γ0 α0 β+ γ+ ∗0 α+ Figure 3. Paths in C \ {0, a+, a−} complement, it is convenient to assume that the line corresponds to the line at infinity and hence it is enough to consider the complement of the affine curve whose equation is obtained from (3.1) by taking z = 1. The fundamental group of the projective Degtyarev curve was computed in [8]. Here we will compute the fundamental group of the affine curve and also show how to recover the group of the projective group. In order to compute the group we will use the braid monodromy associated with the projection (x, y) (cid:55)→ x. Note that the discriminant of the equation (3.1) (with z = 1) is (up to a constant) x(x2− 11x− 1)5. Since the three roots are real and the projection is 3 : 1 with enough real roots, the real picture in Figure 2 contains all the required information to obtain the braid monodromy (the dotted lines represent the real part of the complex conjugate roots). The braid monodromy is defined as a representation ∇0 : π1(C \ {0, a+, a−};∗0) → B3. The source is a free group of rank three generated by: µ+ := α+ · β+ · γ+ · α−1 µ− := α+ · β+ · α0 · β0 · α− · β− · γ− · α−1− · β−1 + , µ0 := α+ · β+ · α0 · β0 · γ0 · α−1 · α−1 0 0 Figure 3 shows a geometric basis of π1(C \ {0, a+, a−};∗0). The braids are obtained by consid- ering the way the roots with respect to y move when the parameters move along x. We follow these conventions: 0 · β−1 + · α−1 + · α−1 · β−1 + . + and 14 E. ARTAL AND J.I. COGOLLUDO Paths Braids α+ β+ γ+ α0 β0 γ0 α− β− γ− Table 1. Braids 1 σ2 2 σ3 2 σ−1 1 σ2 1 σ1 1 σ2 2 σ3 2 (B1) In order to draw the braids we consider the projection onto the real axis. (B2) When two points have the same real part, we perturb the projection such that positive imaginary parts go to the right and negative imaginary parts go to the right. (B3) Roots will be numbered from right to left. (B4) The above conventions give a canonical way to identify open braids with closed braids. Using the standard Artin generators of the braid groups, the braids obtained from following the paths in C \ {0, a+, a−} shown in Figure 3 are presented in Table 1. Proposition 3.3. The braid monodromy for the chosen projection of the affine Degtyarev curve is given by: ∇0(µ+) = σ5 2, ∇0(µ0) = (σ2 2σ−1 1 σ2) ∗ σ1, ∇0(µ−) = (σ2 2σ−1 1 σ2σ1) ∗ σ5 2 = σ2 2 ∗ σ5 1, where a ∗ b := aba−1. 4. Groups of Degtyarev curves In order to compute the fundamental groups we apply the Zariski-van Kampen method. Let us consider the vertical line F of equation x = ∗0. The set F \C is of the form {∗0}×C\{y1, y2, y3}, where y1, y2, y3 ∈ R, identified with the second factor. We choose a big real ∗ number as base point. The free group π1(F \ C;∗) has a free basis g1, g2, g3 constructed as in Figure 3. The natural action of B3 on the free group F3 is expressed in this case as  gi+1 gi+1 ∗ gi gi if i = j, if i = j + 1, if i (cid:54)= j, j + 1. (cid:69) (4.1) gσj i := Proposition 4.1. The fundamental group of the affine Degtyarev curve has a presentation (4.2) g1, g2, g3 ∇0(µj ) i = gi, i = 1, 2, 3, j = −, 0, + . (cid:68) (cid:12)(cid:12)(cid:12) g In this presentation, the meridian of the line at infinity is (cid:0)g3(g2g1)2(cid:1)−1 FUNDAMENTAL GROUPS AND PENCILS 15 . In particular, a pre- sentation for the projective Degtyarev curve is (4.3) (cid:10)g1, g2, g3 (cid:12)(cid:12) (4.2), g3 = (g2g1)−2(cid:11) . Proof. The first presentation is a consequence of the Zariski-van Kampen method by means of the braid monodromy. In order to prove the second one may consider a small deformation of the vertical line F . It will intersect the curve at five points. Three of them are close to (∗0, yi), i = 1, 2, 3, and the other two ones lie in the real branches which go faster to infinity. The (cid:3) boundary of a big disk in this line is the inverse of a meridian of the line at infinity. Remark 4.2. Proposition 4.1 provides right presentations of the group, but they may be quite cumbersome to work with by hand. Even if one wants to work with them with computer programs, like GAP[14], the presentations could be intractable. There are several ways around this problem (P1) The presentation (4.2) works if we replace the braid monodromy ∇0 for a conjugate. For 2 produces simpler braids and example, conjugating the braids in Proposition 3.3 by σ2 hence a simpler presentation of the group. (P2) Instead of finding a good braid to perform the conjugation in (P1) by inspection, one can try to interpret this conjugation in a geometric way. Changing the base point in C \ {0, a+, a−} might produce simpler braids. For example choosing a real number ∗1 ∈ (a−, 0) as a base point, one obtains the following as braid monodromy (for the new generators of the group): (4.4) µ+ (cid:55)→ (σ−1 2 σ1) ∗ σ5 2, µ0 (cid:55)→ σ1, µ− (cid:55)→ σ5 2. These braids have been obtained by conjugation of the ones in Proposition 3.3 by 2σ1σ−1 σ2 2 . (P3) If g is a meridian of the line at infinity obtained using a braid monodromy ∇0, then, for a braid monodromy (∇0)τ := τ−1∇0τ = (τ−1) ∗ ∇0, a meridian of the line at infinity is gτ . (P4) There is another geometric way to reduce the presentation. Note that among the rela- tions (gj)σ5 2 = gj, j = 1, 2, 3, one only needs to keep the relation given by j = 2. First of all, the relation for j = 1 is trivial; secondly (g3g2)τ = g3g2 and hence one of them is redundant. In the general case, this can be summarized as follows: where τj involves only a set of nj consecutive strings and n = (cid:80)r • Let us consider the action (4.1) (replacing 3 by n) of Bn on the free group with basis g1, . . . , gn; let us consider a braid τ ∈ Bn which can be decomposed as τ = τ1·····τr, j=1 nj. Then, j=1(nj − 1) = n − r, among the relations gτ disregarding one for each block of strings. Let Jτ be the chosen subset of indices. j = gj, j = gj, we only need to consider s :=(cid:80)r • If β = (τ )σ, and τ can be decomposed as above, then the set of relations gβ j = 1, . . . , n, is equivalent to (gσ j )τ = gσ j , j ∈ J. For example, in our case the presentation (4.2) can be reduced to have 3 relators. 16 E. ARTAL AND J.I. COGOLLUDO Proposition 4.3. The group G of the affine Degtyarev curve has a presentation: (cid:10)x, y(cid:12)(cid:12)xyxyx = yxyxy, [x, yxy−1xyxy−1xy] = 1(cid:11) (4.5) A presentation of the group GP of the projective Degtyarev curve is obtained from (4.5) by adding x5 = 1. It turns out that GP is a group of order 320 with the following properties: (GP1) GP/G(cid:48)P is cyclic of order 5. (GP2) The center Z(GP) the Klein group of order 4. (GP3) The group G/Z(GP) is a semidirect product of (Z/2)4 by Z5, where the action of a generator of Z5 cyclically permutes a generator system h1, . . . , h5 of order 2 elements of (Z/2)4 satisfying(cid:80) hi ≡ 0. Proof. The presentation of G is obtained using the braid monodromy 4.4 and Remark 4.2(P4), where x = g1, g2 and y = g3; note that x and y are conjugate. In order to obtain the presentation of GP the relation of the line at infinity needs to be added. This is a complicated product of five conjugates of x. If one types this presentation in GAP, the output is that GP has order 320 and that x is an element of order 5. Also according to GAP, the order of the quotient of G obtained by adding the relation x5 = 1 is 320. These facts give the presentation of the statement. The (cid:3) properties of GP are either trivial or easily computed using GAP. Proposition 4.4. The group GP possesses no geometric surjections. Proof. The only properties we need for this are the sizes of the group and its abelianization. Let us assume that GP possesses a geometric surjection. Since it is finite, the orbifold should be spherical, i.e. either P1 2,3,m, m = 3, 4, 5. Since the order of the orbifold group must divide 320, the only possibilities are (2, 2, n), where n160. The group is dihedral and its abelianization is either Z/2 or (Z/2)2; since the abelianization of GP is of order 5 the (cid:3) result follows. 2,2,n, n ≥ 3, or P1 We finish this section with the main result of this paper. We are going to compute the characteristic varieties of the complement of the affine Degtyarev curve and we will prove that these components cannot come from the characteristic varieties of an orbifold. Theorem 4.5. Let TG = C∗ be the character torus of G. Then V1(G) is the set containing 1 and the 10-th primitive roots of unity, whereas V2(G) = ∅. Therefore there is no geometric surjection of G onto an infinite orbifold group. Since finite group orbifolds do not have characteristic varieties, the following Corollary holds. Corollary 4.6. No irreducible component of V1(G) is obtained as the pull-back of an irreducible component of the V1(Γ) where Γ is an orbifold group. Proof of Theorem 4.5. We are going to change the presentation (4.5), by taking a new generator t satisfying y = xt: (cid:10)x, t(cid:12)(cid:12) xtx2tx = tx2tx2t, [x, txt−1xtxt−1xt] = 1(cid:11) (4.6) FUNDAMENTAL GROUPS AND PENCILS 17 It is clear that 1 ∈ V1(G) \ V2(G) since the non-twisted homology has rank 1. Let us consider a non-trivial character ξ ∈ TG, which is identified by the image 1 (cid:54)= ζ of a positive generator of Z. One can associate a CW -complex with the presentation (4.6) with one 0-cell p, two 1-cells x, t and two 2-cells A, B (corresponding to the relations). Then, the complex C∗(X; C)ξ with which to compute the twisted homology is The matrix for ∂1 is(cid:0) ζ−1 (cid:32) ∂2 equals 0 0 −→ C2 ∂2−→ C2 ∂1−→ C −→ 0. (cid:1). In particular, dim ker ∂1 = 1 and hence V2(G) = 0. The matrix for 0 0 1 − ζ + ζ 2 − ζ 3 + ζ 4 (1 − ζ + ζ 2 − ζ 3 + ζ 4)(ζ − 1) (cid:33) . The homology is non trivial if and only if the matrix vanishes and hence V1(G) is as in the statement. Since we are working with the complement of an affine (hence projective) curve, if G admits a geometric surjection onto an infinite orbifold group, the orbifold must be over a rational curve. Since the abelianization has rank 1, the rational curve must be either C or P1. Any dominant morphism with target C can be considered as dominant on P1 and we treat only this case. One needs to consider only orbifolds over P1 whose fundamental groups are infinite, have cyclic abelianizations and admit the 10-th primitive roots of unity in their characteristic varieties. In particular, the abelianization must be of the type Z/n, where 10 divides n. Any such orbifolds admit dominant morphisms onto P1 2,5,10 and P1 2,2,5,5. The properties of V2 allow us to discard P1 2,2,5,5, see Proposition 1.15. Let us assume that there is a geometric surjection onto the orbifold P1 2,5,10. Proposition 1.14 does not provide a direct obstruction in terms of V1. We know that the kernel of the abelian- ization map is the fundamental group K2 of a compact Riemann surface of genus 2, see Propo- sition 1.14. Note that (xy)5 = (x2t)5 is a central element and the group K generated by its element defines an injection in G/G(cid:48). Following [9], if G0 := G/K, the groups G(cid:48) 0 and G(cid:48) are isomorphic and hence G(cid:48) is finitely presented. Using Reidemeister-Schreier method, we find the following presentation: (4.7) G(cid:48) = (cid:104)t0, t1, t2, t3, t4 tn+1tn+3 = tntn+2tn+4, Bn = Bn+1(cid:105) , n+1tn+2t−1 n+3tn+4 and x ∗ tn = tn+1. Note that x10 ∗ tn = tn+10 = A ∗ tn, where Bn := tnt−1 where A := tntn+2tn+4tn+6tn+8 for any n. This guarantees that the above presentation is finite. Summarizing, one can deduce that the kernel K1 of the epimorphism onto Z/10 equals Z × G(cid:48). Note that the rank of K1 equals 5 and the rank of K2 equals 4, so no contradiction arises. According to GAP the following quotients of the lower central series have ranks 5 and 16 for (cid:3) K2, and 2 and 0 (order 5) for K1 and hence such an epimorphism cannot exist. 18 E. ARTAL AND J.I. COGOLLUDO 5. Further properties of the affine Degtyarev curve The affine Degtyarev curve is related with elliptic fibrations as follows. In order to work in a projective setting, one can first consider the projective Degtyarev curve, and fix a singular point P . We will denote by L the tangent line of C at P , and the remaining singular points by P±. Let σ : Σ1 → P2 be the blow-up of P where E denotes the exceptional component. Strict transforms will follow Convention 2.10. Each generic fiber of Σ intersects C at three points. There are four exceptions; three of them can be seen in Figure 2 and they are denoted by F+, F0, and F−. The fourth one is L, which intersects C at two points: one is smooth and transversal and the other one is the infinitely near point of P in E, which is of type A2. In order to separate C and E we perform a positive elementary Nagata transformation ρ : Σ1 (cid:57)(cid:57)(cid:75) Σ2 on the fiber corresponding to L. The fiber which replaces L is denoted by F∞. Note that F∞ intersects C at two points: one of them corresponds to the blow-down of L and the other one is a point with a generic tangency. In particular, the combinatorics of the intersections at F0 and F∞ coincides. Remark 5.1. Properties 3.1 imply the rigidity of this arrangement of curves in Σ2. In particular, once the four fibers are ordered the cross-ratio of their images in P1 provides an invariant of the arrangement. The existence of an automorphism of Σ2 preserving C and exchanging the two fibers containing the singular points can be easily checked. As a consequence of the cross- √ ratio argument, the two tangent fibers must also be exchanged. This automorphism defines a birational map of P2 which is related to the two solutions in Q( 5) exhibited in the proof of Property 3.1(D3). Let us consider the minimal resolution Z of the double covering of Σ2 ramified at C + E. The ruling of Σ2 induces a morphism ρ : Z → P1 such that the generic fiber is elliptic. The only singular fibers are the preimages of F+, F− (of type I5 in Kodaira notation), F0, and F∞ (of type I1). These elliptic fibrations have been extensively studied in [16]. Once a section is fixed (e.g. the preimage of E), the set of sections has an abelian group structure (inherited by the structure on the fibers) which is called the Mordell-Weil group. Note that the involution associated with the double covering is defined by taking the opposite. It is known that the Mordell-Weil group of Z is cyclic of order 5. Let us consider a conic C1 tangent to C both at P and at another singular point and transversal to the third singular point. The preimage of C1 by the double covering has two irreducible components which are denoted by E1 and −E1: they are opposite sections in the Mordell-Weil group. Interchanging the two singular points, one obtains the remaining two sections E2 and −E2 of Z. Let us recall that G denotes the fundamental group of the complement of the affine Degtyarev curve, i.e. P2 \ (C ∪ L) = Σ2 \ (C ∪ E ∪ L∞). Remark 5.2. Despite Proposition 4.4, note that its affine version, G = π1(P2 \ (C ∪ L)) does posses a geometric surjection onto the orbifold over P1 2,2,5, since G admits an epimorphism onto the dihedral group of order 10, see for instance [5]. FUNDAMENTAL GROUPS AND PENCILS 19 In order to construct this morphism, we may use the ideas in [21]. The mapping is obtained by a pencil of rational curves of degree 10, with the following non-reduced fibers: • A smooth conic C2 of multiplicity 5 such that (C · C2)P+ = 2, (C · C2)P− = 4 and (C · C2)P = 4. • A quintic C5 of multiplicity 2 such that (C · C5)P+ = 5 (P+ is a smooth point of C5), (C · C2)P− = 10 (P− is a singular point of C5 of type A4), and (C · C2)P = 10 (P is a singular point of C5 of type D6). • The curve C +L+2D2 where D2 is a smooth conic such that (C·D2)P+ = 0, (C·D2)P− = 5, and (C · D2)P = 4. We finish this section by describing some properties of the group G. For a point Q ∈ C, the 1 (C, Q) of C at Q is π1(BQ \ C), where BQ is a Milnor ball. The local fundamental group πloc inclusion BQ \ C (cid:44)→ C2 \ C induces a conjugacy class of subgroups (since the base point is not fixed) which will be called the image of the local fundamental group. Proposition 5.3. Let P± be the two singular points of the affine Degtyarev curve. (a) The images of the local fundamental groups at P+ and P− are the whole group G. (b) The center of G contains an abelian free subgroup of rank 2. Proof. The property about the image of the local fundamental group at P− is obvious from the presentation (4.5). For P+ it can be deduced using GAP. As a consequence we obtain two central elements (the images of the central elements of the local fundamental groups). The last property (cid:3) can be deduced by studying some quotients of subgroups of G. References [1] D. Arapura, Geometry of cohomology support loci for local systems. I, J. Algebraic Geom. 6 (1997), no. 3, 563 -- 597. [2] E. Artal, A curve of degree five with non-abelian fundamental group, Topology Appl. 79 (1997), no. 1, 13 -- 29. [3] , Fundamental group of a class of rational cuspidal curves, Manuscripta Math. 93 (1997), no. 3, 273 -- 281. [4] E. Artal, J.I. Cogolludo, and D. Matei, Orbifolds and characteristic varieties, In preparation. [5] E. Artal, J.I. Cogolludo, and H. Tokunaga, Pencils and infinite dihedral covers of P2, Proc. Amer. Math. Soc. 136 (2008), no. 1, 21 -- 29 (electronic). [6] A. Beauville, Annulation du H 1 pour les fibr´es en droites plats, Complex algebraic varieties (Bayreuth, 1990), Lecture Notes in Math., vol. 1507, Springer, Berlin, 1992, pp. 1 -- 15. [7] N. Budur, Unitary local systems, multiplier ideals, and polynomial periodicity of Hodge numbers, Adv. Math. 221 (2009), no. 1, 217 -- 250. [8] A.I. Degtyarev, Isotopic classification of complex plane projective curves of degree 5, Leningrad Math. J. 1 (1990), no. 4, 881 -- 904. [9] , Plane sextics via dessins d'enfants, Geom. Topol. 14 (2010), no. 1, 393 -- 433. [10] T. Delzant, Trees, valuations and the Green-Lazarsfeld set, Geom. Funct. Anal. 18 (2008), no. 4, 1236 -- 1250. [11] A. Dimca, Singularities and topology of hypersurfaces, Universitext, Springer-Verlag, New York, 1992. [12] , Characteristic varieties and constructible sheaves, Atti Accad. Naz. Lincei Cl. Sci. Fis. Mat. Natur. Rend. Lincei (9) Mat. Appl. 18 (2007), no. 4, 365 -- 389. [13] H. Flenner and M.G. Zaıdenberg, On a class of rational cuspidal plane curves, Manuscripta Math. 89 (1996), no. 4, 439 -- 459. 20 E. ARTAL AND J.I. COGOLLUDO [14] The GAP Group, GAP -- Groups, Algorithms, and Programming, Version 4.4, 2004, available at (\protect\vrule width0pt\protect\href{http://www.gap-system.org}{http://www.gap-system.org}). [15] A. Libgober, Characteristic varieties of algebraic curves, Applications of algebraic geometry to coding theory, physics and computation (Eilat, 2001), Kluwer Acad. Publ., Dordrecht, 2001, pp. 215 -- 254. [16] R. Miranda and U. Persson, On extremal rational elliptic surfaces, Math. Z. 193 (1986), 537 -- 558. [17] M. Namba, Geometry of projective algebraic curves, Marcel Dekker Inc., New York, 1984. [18] A. N´emethi, On the fundamental group of the complement of certain singular plane curves, Math. Proc. Cambridge Philos. Soc. 102 (1987), no. 3, 453 -- 457. [19] M. Oka, Some plane curves whose complements have non-abelian fundamental groups, Math. Ann. 218 (1975), no. 1, 55 -- 65. [20] C. Simpson, Subspaces of moduli spaces of rank one local systems, Ann. Sci. ´Ecole Norm. Sup. (4) 26 (1993), no. 3, 361 -- 401. [21] H. Tokunaga, Dihedral coverings of algebraic surfaces and their application, Trans. Amer. Math. Soc. 352 (2000), no. 9, 4007 -- 4017. [22] K. Tono, Rational unicuspidal plane curves with κ = 1, S¯urikaisekikenky¯usho K¯oky¯uroku (2001), no. 1233, 82 -- 89, Newton polyhedra and singularities (Japanese) (Kyoto, 2001). [23] M.G. Zaıdenberg and V.Ya. Lin, An irreducible, simply connected algebraic curve in C2 is equivalent to a quasihomogeneous curve, Dokl. Akad. Nauk SSSR 271 (1983), no. 5, 1048 -- 1052. [24] O. Zariski, On the problem of existence of algebraic functions of two variables possessing a given branch curve, Amer. J. Math. 51 (1929), 305 -- 328. Departamento de Matem´aticas, IUMA, Facultad de Ciencias, Universidad de Zaragoza, c/ Pedro Cerbuna, 12, 50009 Zaragoza, Spain. E-mail address: [email protected] URL: http://riemann.unizar.es/geotop/WebGeoTo/Profes/eartal/ Departamento de Matem´aticas, IUMA, Facultad de Ciencias, Universidad de Zaragoza, c/ Pedro Cerbuna, 12, 50009 Zaragoza, Spain. E-mail address: [email protected] URL: http://riemann.unizar.es/geotop/WebGeoTo/Profes/jicogo/
1603.07500	1	1603	2016-03-24T09:39:42	Recognizing projections of rational curves	[ "math.AG" ]	Given two rational, properly parametrized space curves ${\mathcal C}_1$ and ${\mathcal C}_2$, where $\CCC_2$ is contained in some plane $\Pi$, we provide an algorithm to check whether or not there exist perspective or parallel projections mapping $\CCC_1$ onto $\CCC_2$, i.e. to recognize $\CCC_2$ as the projection of $\CCC_1$. In the affirmative case, the algorithm provides the eye point(s) of the perspective transformation(s), or the direction(s) of the parallel projection(s). The problem is mainly discussed from a symbolic point of view, but an approximate algorithm is also included.	math.AG	math	Recognizing projections of rational curves. Juan Gerardo Alc´azara,1, Carlos Hermosoa aDepartamento de F´ısica y Matem´aticas, Universidad de Alcal´a, E-28871 Madrid, Spain Abstract Given two rational, properly parametrized space curves C1 and C2, where C2 is contained in some plane Π, we provide an algorithm to check whether or not there exist perspective or parallel projections mapping C1 onto C2, i.e. to In the affirmative case, the algorithm recognize C2 as the projection of C1. provides the eye point(s) of the perspective transformation(s), or the direction(s) of the parallel projection(s). The problem is mainly discussed from a symbolic point of view, but an approximate algorithm is also included. 1. Introduction In this paper, we address the following geometric problem: given a ratio- nal space curve C2, lying on a plane Π, and another rational space curve C1, not necessarily planar, check whether or not there exist perspective or parallel projections mapping C1 onto C2, and find them in the affirmative case. Our problem can be translated into the context of Computer Vision. For this purpose, we recall [13] the simplest camera model, known as the pinhole camera. In this model, a camera is modeled as a pair (a, Π), where a is called the eye point of the camera, and Π is the image plane: then, given an object Ω ⊂ R3, the photograph of Ω taken by the camera is the projection of Ω from the point a onto the plane Π (see Fig. 1). The eye point can be allowed to be at infinity, in which case we have a parallel projection from a certain direction. Therefore, in this context our problem can be translated as whether or not C2 can be regarded as a photograph of C1, taken with a camera where the image plane is known (it is the plane Π containing C2), but where the eye point is unknown. A more general problem is treated in [8]. In [8] the input is a pair of algebraic curves, D1 ⊂ R3 and D2 ⊂ R2, not necessarily rational,and the question is in to check if there exists some camera where D2 is the photograph of D1: other words, to find the positions of the eye point and the image plane, if any, Email addresses: [email protected] (Juan Gerardo Alc´azar), [email protected] (Carlos Hermoso) 1Supported by the Spanish Ministerio de Econom´ıa y Competitividad and by the European Regional Development Fund (ERDF), under the project MTM2014-54141-P. Member of the Research Group asynacs (Ref. ccee2011/r34) Preprint submitted to Elsevier July 13, 2021 Figure 1: Pinhole camera model. such that D2 is the photograph of D1 taken from the camera eye point. This problem is known as the object-image correspondence problem, and amounts to recognizing images without any clue on the parameters of the camera used to take the photograph. Since in our case we assume that the image plane is known, the problem here can be considered as a weak version of the problem in [8]. Because the problem is simpler, we can find a solution computationally simpler than that of [8], too. In [8] the problem is solved by deciding whether the curve D2 is equivalent to some curve in a family of planar curves, computed from D1, under an action of the projective or the affine group. In turn, this is done by using differential invariants. Computationally, the question boils down to solving a quantifier elimination problem with five variables, in the case of perspective projections, and with four variables in the case of parallel projections. Since the differential invariants used in [8] are high-order (5 and 6 for affine actions, 7 and 8 for projective actions), the elimination problem can be hard. In our case, we restrict to rational curves parametrized over Q, and use a very different approach. We observe that any projection between C1 and C2 cor- responds to a rational function ψ between the parameter spaces; furthermore, ψ is shown to be a Mobius transformation in the case of non-degenerate pro- jections between C1 and C2, i.e. projections which are injective for almost all points of C2. We show that the rational functions ψ potentially corresponding to projections from eye points with rational coordinates can be efficiently com- puted by means of standard bivariate factoring techniques over the rationals. In order to also find the projections from non-rational eye points we need bivariate factoring over the reals, i.e. an absolute factorization. Furthermore, once ψ is computed, checking if it gives rise to some projection between C1 and C2 is easy. As an alternative to computing an absolute factorization, we also provide another algorithm, both in symbolic and approximate versions, which takes advantage of the existence of ψ without actually computing it. In general, the symbolic version of this last algorithm requires to compute the primitive element of an algebraic extension Q(α, β), which can be costly. However, the approximate version of the algorithm, where algebraic numbers are numerically approximated, is fast. In this last case, we do not check if C2 is the projection of C1, but if C2 is "approximately" the projection of C1. Furthermore, this approximate algorithm is well-suited for curves whose defining parametrizations are known only up to a certain precision, i.e. with floating point coefficients, 2 ΩaΠwhich is closer to applications. Acknowledgements. We thank Ron Goldman for some fruitful discussions on the problem. We also thank Sonia Rueda for her help with the estimation of Hausdorff distances, in Example 4. 2. Projective and parallel projections. Throughout the paper, we consider two rational space curves C1,C2 ⊂ R3, C1 (cid:54)= C2. Such curves are algebraic, irreducible and can be parametrized by rational maps xj(t) =(cid:0)xj(t), yj(t), zj(t)(cid:1), j = 1, 2. (1) (cid:105) : 1 . (cid:104) x ω : y ω : z ω 3 xj : R (cid:57)(cid:57)(cid:75) Cj ⊂ R3, We will suppose that x1, x2 have coefficients in Q, and that C2 is contained in some plane Π; however, C1 is not necessarily planar. Also, we will exclude the case when C1,C2 are two planar curves contained in the same plane. The components xj, yj, zj of xj are real, rational functions of t, therefore defined for all but a finite number of values of t. Nevertheless, at certain moments we will consider xj, yj, zj as functions from C to C. For j = 1, 2, the parametrization xj generates all the points of Cj except perhaps the point P ∞ j = limt→∞xj(t), which is affine whenever the degrees of the numerators of xj, yj, zj are less or equal than the degrees of the denomina- tors. We will assume that the parametrizations in (1) are proper, i.e., birational or, equivalently, injective except for perhaps finitely many values of t. This can be assumed without loss of generality, since any rational curve can be properly reparametrized. For these claims and other results on properness, the interested reader can consult [19] for plane curves and [1, §3.1] for space curves. The parallel projection onto Π, in the direction of a nonzero vector v ∈ R3, is Now let us introduce projective and parallel projections. Let Π be a plane. the transformation in 3-space that maps every (complex or real) point p onto the intersection q of Π and the line through p parallel to v. The perspective projection onto Π from a point a, called the eye point, is the transformation in 3-space that maps every point p onto the intersection q of Π and the line connecting a and p. These definitions are illustrated by Fig 2. Parallel and perspective projections can be unified when we move to a projective setting. The (complex) projective space, P3C, is the set of 4-tuples p = [x : y : z : ω], where x, y, z, ω ∈ C and at least one them is nonzero, such that two such 4-tuples are considered equal when they are proportional to each other. If ω = 0 we have a point at infinity, while ω (cid:54)= 0 corresponds to an affine point. In particular, if p is affine then it can be represented by the 4-tuple Figure 2: Perspective projection (left), parallel projection (right). The projective closure of Cj is the curve in P3C whose affine part is Cj; the projective closure can be parametrized as x(t) = [xj(t) : yj(t) : zj(t) : ωj(t)] , ωj (t) , yj(t) = yj (t) where xj(t) = xj (t) ωj (t) . For simplicity, we will use the same notation for a curve and its projective closure; it will be clear from the context whether we are working with one or the other. ωj (t) , zj(t) = zj (t) In this context, parallel or perspective projections are treated in the same way: in the case of perspective projections the eye point is affine, and in the case of parallel projections, the eye point is at infinity. So both projections [11, §13] can be represented by a projective transformation  = x(cid:48) y(cid:48) z(cid:48) ω(cid:48) p11 (cid:124) p21 p31 p41  . x y z ω  (cid:125) · p14 p24 p34 p44 p12 p22 p32 p42 p13 p23 p33 p43 (cid:123)(cid:122) P If a := [a1 : a2 : a3 : a4] denotes the eye point of the projection, and the implicit equation of the projection plane Π is Ax + By + Cz + D = 0, an easy computation shows that the matrix P is a1B −a1A − a3C − a4D a3B a4B a1C a2C −a1A − a2B − a4D a4C a1D a2D a3D −a1A − a2B − a3C (2) In particular, not every 4 × 4 matrix represents a projection. Any matrix P representing a projection satisfies that rank(P ) = 3; therefore the dimension of Ker(P ) is 1, and any basis of Ker(P ) provides projective coordinates for the eye point of the transformation. Furthermore, the affine part of the image Im(P ) defines the projection plane Π. If we denote the matrix consisting of the first three columns of P by M , perspective projections satisfy that rank(M ) = 3, −a2B − a3C − a4D a2A a3A a4A P =  . 4 C2C2C1C1avΠΠwhile parallel projections satisfy that rank(M ) < 3. We will represent the pro- jection onto a plane Π from an eye point a as Pa, so as to make the eye point explicit. Furthermore, when moving to the affine space, parallel projections cor- respond to affine 3-D transformations, while perspective projections correspond to rational 3-D transformations, where the numerator and denominator of each component has degree one. In the paper we will work with real projections, so that the eye point, and therefore the matrix P , are real. Additionally, given a projection Pa and a curve C, the projection of C, Pa(C), is the image of C under Pa. Proposition 1. Let P be a 4× 4 matrix defining a projection Pa, and let C be a projective rational curve properly parametrized by x(t), with at least one affine point. If C is not a line going through a, then Pa(C) is a rational curve. Proof. Let x(t) = [x(t) : y(t) : z(t) : ω(t)]. If Pa(C) is a curve then P · x(t) parametrizes Pa(C), so Pa(C) is clearly rational. So suppose that Pa(C) is not a curve. Then there exists a polynomial λ(t), and αi ∈ C, i = 1, . . . , 4, such that P ·  = λ(t) ·  x(t) y(t) z(t) ω(t)  . α1 α2 α3 α4 In this situation, for i = 1, . . . , 4 we have that pi1 · x(t) + pi2 · y(t) + pi3 · z(t) + pi4 · ω(t) = λ(t) · αi. Therefore, there exist µ1, µ2, µ3 ∈ C such that (3) (p11 − µ1 · p21)x(t) + (p12 − µ1 · p22)y(t) + (p13 − µ1 · p23)z(t) + (p14 − µ1 · p24)ω(t) = 0 (p11 − µ2 · p31)x(t) + (p12 − µ2 · p32)y(t) + (p13 − µ2 · p33)z(t) + (p14 − µ2 · p34)ω(t) = 0 (p11 − µ3 · p41)x(t) + (p12 − µ3 · p42)y(t) + (p13 − µ3 · p43)z(t) + (p14 − µ3 · p44)ω(t) = 0. (4) Notice that each of the above equations expresses that C is contained in a certain plane Πk, k = 1, 2, 3. Since rank(P ) = 3, at least two of these planes are different, so C is contained in the intersection of two different planes, and therefore C is a straight line. Furthermore, since by hypothesis C has at least one affine point, C must be an affine straight line. Hence, let x0 + t · v, where x0 = [x0 : y0 : z0 : 1], v = [v1 : v2 : v3 : 0] be a parametrization of C. Then P · x(t) = P · x0 + tP · v. Now recall that P · x(t) = λ(t) · α, where α = [α1 : α2 : α3 : α4]. If λ(t) is a constant, then P · v = 0, so v is the eye point of the projection defined by P . Therefore, C goes through the projective point defined by v. However, by hypothesis this cannot happen. So λ(t) cannot be a constant, in which case P · x0 and P · v are proportional, because each one must be proportional to α. Therefore there exist a, b ∈ C such that ax0 + bv ∈ Ker(P ), which implies that C contains the eye point of the projection. Again, by hypothesis this cannot happen. We deduce that Pa(C) must be a curve, so the proposition is proven. 5 3. Statement of the problem. In order to formally state the problem we want to solve, we first need the following definition. Definition 2. Let a ∈ P3R, let C1, C2 be two space rational curves, where C2 is contained in a plane Π, and let Pa be the projection from the eye point a onto the plane Π. We say that C2 is the projection of C1 from a, i.e. C2 = Pa(C1), if every (projective) point of C2 is the projection of some (projective) point of C1. Furthermore, we will say that the projection is non-degenerate if Pa is injective for almost all points of C2; otherwise we will say that the projection is degenerate. −1C1 Observe that Definition 2 is consistent. Indeed, because of Proposition 1 and since C1, C2 are both irreducible, the projection Pa(C1) of C1 either has finitely many points in common with C2 (in which case certainly C2 is not the projection of C1 from a), or completely coincides with C2. However, this requires working over the complex projective space: for instance, according to Definition 2 the parabola {y = x2, z = 0} is the projection of the space curve parametrized by (t, t2, 1/t) from the point [0 : 0 : 1 : 0], but the origin is the image of a point at (cid:17) infinity. Also, {y = x2, z = 0} is the (degenerate) projection of the space curve (t2, t4, t) from the point [0 : 0 : 1 : 0]; however, all the affine points (a, a2, 0) of of the parabola with a < 0 come from complex, affine points the space curve. −a,a2, i(cid:112) (cid:16) a On the other hand, the notion of degeneracy arising in Definition 2 implies the existence of either two different (possibly complex) branches of C1 projecting onto a same branch of C2, or some branch of C1 collapsing onto a point of C2 under projection. The notion of degeneracy is, in general, also defined over the complex numbers. It can even happen that degeneracy occurs only over the com- plex, but not the real numbers. For instance, let C1 be the curve parametrized by x1(t) = (t3, t6, t), and let C2 be the parabola parametrized by x2(s) = (s, s2, 0). It is clear that for every real point p = (a, a2, 0) ∈ C2, with a ∈ R, there is just one real point of C1 projecting onto p, namely the point (a, a2, 3√a). However, since 3√a has three different complex values for a (cid:54)= 0, there are three complex points of C1 projecting onto p, namely the points (a, a2, 3√a · ξ), where ξ3 = 1. So every branch of C2 comes from three different complex branches of C1. Now we can state the problem we want to solve. Although in the paper we will mostly work in the affine space, the language of projective space is useful to state the problem in a clearer and more general way. • Projection Problem: given two affine, space real algebraic curves C1, C2, where C2 is contained in a plane Π, properly parametrized by x1(t), x2(s), check if there exists a ∈ P3R (and find it in the affirmative case) such that C2 is the projection of C1 from a. Observe that once a is computed we can find out the nature of the projection: if a is a point at infinity we have a parallel projection, and if a is affine we have 6 Figure 3: Right: intersection of the cylinders x2 + y2 = 1, x2 + z2 = 1. Left: the two irreducible curves we get when intersecting x2 + y2 = 1, x2 + z2 = 1 can be projected onto each other in two different directions. Figure 4: The circles {x2 + y2 = 1, z = 1} and {x2 + y2 = 2, z = 0} are related by two different perspective projections. a perspective projection. Additionally, if either C1 or C2 is a straight line, then C2 = Pa(C1) implies that both C1 and C2 are contained in a same plane, which is a case we excluded at the beginning of the section; so in the rest of the paper, we will assume that C1, C2 are not straight lines. The parallel or perspective projection mapping C1 onto C2 is not necessarily unique. For instance, let C1, C2 be the two irreducible components of the space curve obtained by intersecting the cylinders x2 + y2 = 1, x2 + z2 = 1 (see Figure 3). One can check that the projections parallel to the y-axis and the z-axis both transform C1 into C2. Similarly, let C1 be {x2 + y2 = 1, z = 1} and let C2 be {x2 + y2 = 2, z = 0} (see Fig. 4). These curves are two circles of radii 1 and √2, located in the planes z = 1 and z = 0, with centers on the z-axis. In this case there are two different perspective projections transforming C1 into C2, one . This from the point example also shows that two rational curves C1 and C2 parametrized over Q can however be related by a projection from a point with non-rational coordinates. Notice that there can be a parallel projection and a perspective projection √ 2√ 2−1 √ 2√ 2+1 0, 0, , and another one from the point 0, 0, (cid:16) (cid:17) (cid:16) (cid:17) 7 simultaneously transforming C1 into C2, too. For instance, let C1 be {x2 + y2 = 1, z = 1} and let C2 be {x2 + y2 = 1, z = 0}. It is clear that C2 is the projection of C1 parallel to the direction of the z-axis. But C2 is also the perspective projection of C1 from the point (0, 0, 1/2). 4. A factoring-based strategy. and a scalar (cid:98)λ =(cid:98)λ(t, s), such that the vector connecting the points x1(t) and Suppose that C2 = Pa(C1), where a = [a1 : a2 : a3 : a4], a4 (cid:54)= 0. Hence, for almost every affine point x2(s) ∈ C2 there exists another affine point x1(t) ∈ C1, x2(s) is parallel to the vector connecting the point x1(t) and the eye point (cid:18) a1 with (cid:98)λ being the proportionality constant between these two vectors. So for almost all s there exist t,(cid:98)λ such that (cid:19) a2 a4 a3 a4 a = a4 , , , We will say that two t, s satisfying Eq.(5) for some(cid:98)λ, are related by the projec- Calling λ = (cid:98)λ tion Pa; i.e. t, s are related by Pa iff x2(s) = Pa(x1(t)). , Eq. (5) gives rise to a4 . a4 a4 a4 (cid:16) (cid:17) (cid:16) (cid:17) x1(t) − a1 (cid:16) (cid:17) y1(t) − a2 z1(t) − a3  x2(s) − x1(t) = (cid:98)λ · y2(s) − y1(t) = (cid:98)λ · z2(s) − z1(t) = (cid:98)λ ·  x2(s) − x1(t) = λ · (a4 · x1(t) − a1) y2(s) − y1(t) = λ · (a4 · y1(t) − a2) z2(s) − z1(t) = λ · (a4 · z1(t) − a3) (5) (6) Now notice that when C2 is the projection of C1 from a projective point a = [a1 : a2 : a3 : 0] at infinity, Eq.(6) also holds. Indeed, in that case Eq.(6) expresses that the vector connecting x1(t) and x2(s), where t, s are related, is a multiple of the vector v = (−a1,−a2,−a3), which defines the projection direction; λ = λ(t, s) is the proportionality constant between x1(t) − x2(s) and v. Therefore, in order to solve the projection problem we need to study Eq. (6): every solution with a4 = 0 corresponds to a parallel projection, and every solution with a4 (cid:54)= 0 corresponds to a perspective projection. The next theorem is crucial. We recall here the definition of degree of a rational function: if ψ(t) = p(t) q(t) , where p(t), q(t) are polynomials, the degree of ψ(t) is the maximum of the degrees of p(t), q(t). Furthermore, we define the associated polynomial of a given rational function ψ(t) = p(t) q(t) as G(t, s) = p(t)− sq(t). Finally, we recall that Mobius transformation is a function ϕ(t) = at+b ct+d , with ad−bc (cid:54)= 0; it is well-known that Mobius transformations are the birational transformations of the complex line [19]. 8 Figure 5: Illustrating Theorem 3. Theorem 3. If C2 = Pa(C1), then there exists a real rational function ψ(t) such that Pa ◦ x1 = x2 ◦ ψ. Furthermore, if Pa is non-degenerate, then ψ(t) is a Mobius transformation. Proof. Since by hypothesis x2 is proper, x−1 ψ = x−1 2 ◦ Pa ◦ x1 (see Diagram (7)) defines a mapping from C to C such that the t-value generating a point in C1 via x1, is mapped onto the s-value generating the point in C2 that is the projection of x1(t). exists. Therefore, the function 2 C1 x1 C Pa ψ / C2 −1 x 2 / C (7) Since x2 is rational then x−1 is also rational, and since x1 and Pa are rational too, we deduce that ψ is a rational function. Finally, if Pa is non-degenerate then it has an inverse, and therefore ψ also has an inverse, which must be rational. Therefore ψ is a birational transformation of the complex line, and hence it is a Mobius transformation. 2 We will say that the function ψ(t) is associated with Pa. Additionally, if G(t, s) is the polynomial associated with ψ(t), we will also say that G(t, s) is associated with Pa. Observe that if C2 = Pa(C1) then the projection of x1(t) under Pa is the point x2(ψ(t)) (see Figure 5). Furthermore, Theorem 3 provides the following corollary, which follows by substituting s = ψ(t) into Eq. (6), taking into account that x2(s) and x1(t) are rational parametrizations. Corollary 4. Let C2 = Pa(C1), and let ψ(t) be associated with Pa. Then there exists a rational function λ(t) such that for each t, the values s = ψ(t), λ = λ(t) satisfy Eq. (6). Remark 1. If there are several projections Pai such that C2 = Pai (C1), i = 1, . . . , m, we can see from the diagram (7) that each one must correspond to a different ψi(t). 9 x1(t)x2(ψ(t))/ O O / In order to find the projection Pa, if any, we will focus on computing the associated function ψ(t) described in Theorem 3. More precisely, we will seek the polynomial associated with ψ(t). Now let N (t, s) = ∂x1(t) ∂t × ∂x2(s) ∂s . Then we have the following result. Proposition 5. If t, s are related by the projection Pa, then N (t, s) · (x1(t) − x2(s)) = 0, (8) (9) where · represents the usual Euclidean dot product. Proof. If C2 = Pa(C1) then C1 and C2 are contained in a developable surface S, which is cylindrical when a is at infinity, and conical when a is not. Furthermore, if t, s are related, then there is a generatrix of S connecting x1(t) and x2(s). At each point p = x1(t) the vector ∂x1(t) belongs to the tangent plane to S at p. Similarly, at each point q = x2(s) the vector ∂x2(s) belongs to the tangent plane to S at q. Since the tangent plane at a point of a developable surface contains the generatrix through the point [21, §2.4], if t, s are related then x1(t)− x2(s), belong to the same plane. Therefore, they are linearly dependent, and Condition (9) follows. , ∂x2(s) ∂s ∂t ∂x1(t) ∂t ∂s Proposition 5 provides the following corollary; we denote the square-free part of the numerator of the rational function at the left hand-side of (9), by N (t, s). Corollary 6. If t, s are related under the projection Pa, and G(t, s) is associated with Pa, then G(t, s) divides N (t, s). Corollary (6) suggests a method for computing the factors G(t, s) giving rise to projections: (a) compute the factors of N (t, s) (by means of factoring techniques); (b) pick the factors which are linear in s, and compute the cor- responding functions ψ(t); (c) find the projections by making use of Eq. (6). Observe that this strategy requires that N (t, s) is not identically zero. Lemma 7. Under the preceding hypotheses, N (t, s) cannot be identically zero. Proof. We argue by contradiction. Suppose that N (t, s) · (x1(t) − x2(s)) is identically zero. Then for all (t, s) ∈ C2, (cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) x1(t) − x2(s) x(cid:48) 1(t) x(cid:48) 2(s) y1(t) − y2(s) y(cid:48) 1(t) y(cid:48) 2(s) z1(t) − z2(s) z(cid:48) 1(t) z(cid:48) 2(s) (10) (cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) = 0. We will prove that if the above determinant is zero, C1, C2 are contained in a same plane, which was excluded at the beginning of the paper. In order to do this, we will first see that C1,C2 must both be planar, and then we will 10 prove that the plane containing both curves must coincide. We start proving 2(s0) = that C1 is planar. Let s = s0 satisfy that x2(s0) is well-defined, and x(cid:48) (u0, v0, w0) (cid:54)= 0. By substituting s = s0 in Eq. (10), we get (cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) x1(t) − x0 x(cid:48) 1(t) u0 y1(t) − y0 y(cid:48) 1(t) v0 z1(t) − z0 z(cid:48) 1(t) w0 (cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) = 0 (11) (cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) = 0. (13) for all t. Since x(cid:48) 2(s0) = (u0, v0, w0) (cid:54)= 0 we can assume that, say, w0, is nonzero. Then there exist γ, µ such that u0 = γ·w0, v0 = µ·w0. Now perform the following elementary transformations in the determinant at the left hand-side of Eq. (11): multiply the last column by −γ, and add it up to the first column; multiply the last column by −µ, and add it up to the second column. Hence, we get z1(t) − z0 y1(t) − µ · z1(t) − y0 + µ · z0 y(cid:48) 1(t) − µ · z(cid:48) 1(t) 0 z(cid:48) 1(t) w0 (12) Calling X(t) = x1(t) − γ · z1(t) − x0 + γ · z0, Y (t) = y1(t) − µ · z1(t) − y0 + µ · z0 and Z(t) = z1(t) − z0, we have 0 1(t) x(cid:48) 1(t) − γ · z(cid:48) (cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) x1(t) − γ · z1(t) − x0 + γ · z0 (cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) X(t) (cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) = 0. Y (t) Z(t) X(cid:48)(t) Y (cid:48)(t) Z(cid:48)(t) w0 0 0 If X(t) is identically zero then C1 is contained in the plane of equation x − γz − x0 + γ · z0 = 0; similarly, if Y (t) is identically zero then C1 is contained in the plane of equation y − µz − y0 + µ · z0 = 0. If either X(cid:48)(t) or Y (cid:48)(t) are identically zero, then we can also easily conclude that C1 is planar. In any other case, by expanding the determinant at the left hand-side of Eq. (13), we get X(t)Y (cid:48)(t) − X(cid:48)(t)Y (t) = 0, which leads to X(t) = ν · Y (t), with ν a constant. Hence, we deduce that C1 is contained in the plane x − νy − (γ − µν)z − x0 + νy0 = 0. Let Π1 be the plane containing C1, and let Π2 be the plane containing C2. Let P = x1(t0) ∈ C1, where the tangent vector (cid:126)vP = x(cid:48) 1(t0) to C1 is well-defined, and let Q = x2(s0) ∈ C2, where the tangent vector (cid:126)vQ to C2 is also well-defined. (cid:16) (cid:126)P Q, (cid:126)vP , (cid:126)vQ Since C1,C2 are not both straight lines, we can assume that (cid:126)vP and (cid:126)vQ are not If Π1 (cid:54)= Π2 then { (cid:126)P Q, (cid:126)vP , (cid:126)vQ} are linearly independent. Therefore parallel. 2(s0)) (cid:54)= 0. However (10); hence, Π1 = Π2, which is excluded by (cid:54)= 0, i.e. det (x1(t0) − x2(s0), x(cid:48) 1(t0), x(cid:48) (cid:17) det this is contradictory with Eq. hypothesis. Since N (t, s) is not identically zero, N (t, s) has finitely many factors; so taking Remark 1 into account, Lemma 7 provides the following corollary. Corollary 8. The number of projections Pa such that C2 = Pa(C1), is finite. 11 Figure 6: Unlucky points. We need to consider an additional question. If there are two projections Pa1 (cid:54)= Pa2 mapping C1 onto C2, it might exist points p ∈ C1 and q ∈ C2 that are mapped onto each other by both projections, i.e. such that q = Pa1 (p) = Pa2(p). For instance, consider the two lemniscatas shown in Fig. 6. These two lem- niscatas are mapped onto each other by two different projections, as shown by the picture. However, both projections map the singular point of C1 onto the singular point of C2. For these pairs (p, q), the line connecting p, q contains both eye points. The next lemma shows that we just have finitely pairs of this kind, that can be computed. Here we represent by N (cid:63) the planar algebraic curve, in the (t, s) plane, defined by the polynomial N (t, s). Lemma 9. Let a1 (cid:54)= a2 such that Pa1 and Pa2 map C1 onto C2, and let p ∈ C1 satisfy that Pa1 (p) = Pa2 (p). Let q = Pai(p), for i = 1, 2. Then one of the following situations occur: (i) p = P ∞ 2 ; (iii) q is a self-intersection of C2; (iv) p = x1(t0), where the line t = t0 contains a singular point of N (cid:63). Proof. Let ψ1(t), ψ2(t) be associated with Pa1, Pa2. Assume that there exists t0 ∈ C such that p = x1(t0), and assume also that ψ1(t0), ψ2(t0) are well defined. Notice that if any of these conditions is not satisfied, then either p = P ∞ 1 or q = P ∞ 2 . Since Pa1(x1(t0)) = Pa2(x1(t0)) and Pai (x1(t0)) = x2(ψi(t0)) for i = 1, 2, we have 1 ; (ii) q = P ∞ x2(ψ1(t0)) = x2(ψ2(t0)). Now if ψ1(t0) (cid:54)= ψ2(t0), then q is a self-intersection of C2. If ψ1(t0) = ψ2(t0), then the point (t0, s0), where s0 = ψ1(t0) = ψ2(t0), is a point of N (cid:63) where 12 pqC1C2a1pqC1C2a2two different branches of N (cid:63) (namely, the ones corresponding to the functions s = ψ1(t) and s = ψ2(t)) intersect. So (t0, s0) is a singularity of N (cid:63). Lemma 9 gives rise to the following definition. Definition 10. Let (p, q), where p ∈ C1, q ∈ C2 and Pa(p) = q. We say that the pair (p, q) is lucky, if p (cid:54)= q and p, q do not satisfy any of the conditions (i)-(iv) in Lemma 9. Otherwise, we say that (p, q) is unlucky. Notice that the set of unlucky pairs is finite. Observe that a priori we do not know the number of projections mapping C1 onto C2. Hence, the unlucky pairs correspond to the potential points in C1 and C2 such that, if there are several projections mapping C1 onto C2, are mapped onto its partner by all those projections. Unlucky pairs, as we will see later, are to be avoided, because we will be interested in pairs of points such that the line connecting them contains just one eye point, if any. In the rest of the paper, given a function ψ(t), associated with some pro- jection, we denote by Gψ the set of (finitely many) t0 values such that the pair (p, q), where p = x1(t0) and q = x2(ψ(t0)), is unlucky. 4.1. A factoring-based algorithm. As we observed after Corollary 6, in order to solve our problem, once we compute the factors of N (t, s) which are linear in s we just need to find out which of these factors, if any, correspond to projections mapping C1 onto C2. In order to do so, the notion of lucky point is essential. Let G(t, s) be a factor of N (t, s), linear in s, and let s = ψ(t) be the result of solving for s in G(t, s) = 0. By Lemma 9, given t1 (cid:54)= t2, t1, t2 /∈ Gψ there is just one projection Pa mapping x1(tj) → x2(ψ(tj)), j = 1, 2. Furthermore, if t1 (cid:54)= t2, t1, t2 /∈ Gψ there is a line Lj connecting x1(tj) and x2(ψ(tj)), j = 1, 2, and a is the (projective) intersec- tion point of L1 and L2. Then we have the following algorithm, Algorithm 1, to solve the projection problem. Algorithm 1 is illustrated in the following example. In this example and in the other examples of the paper, the computations have been done with Maple 18, running on a laptop, with 2.9 GHz i7-3520M processor and 8 Gb RAM. Example 1. Let C1 be the curve parametrized by x1(t) = (x1(t), y1(t), z1(t)), where 2t5+11t4+14t3−3t2−12t+4 x1(t) = y1(t) = z1(t) = 2t6+10t5+19t4+14t3−3t2−12t+4 −2t2(t3+3t2+3t−2) (t+2)2 t+2 (t+2)2 , , , and let C2 be the curve parametrized by x2(s) = (10s + 5,−2(s + 3)(s − 2), 2s2 + s + 2). One can check that C2 is planar, and that both C1 and C2 are contained in the cone x2 + y2 − z2 = 0; therefore, they must be related by at least one perspective 13 Algorithm 1 Require: Two proper, rational parametrizations x1(t), x2(s) defining two space curves C1,C2, where C2 is planar. Ensure: The projections, if any, mapping C1 onto C2. 1: compute the polynomial N (t, s). 2: find the real factors Gi(t, s) of N (t, s) which are linear in s. 3: for each Gi(t, s) do 4: 5: 6: 7: 8: solve Gi(t, s) for s, to find s = ψi(t). pick t1 (cid:54)= t2, t1, t2 /∈ Gψi. compute the lines Lj, j = 1, 2, connecting x1(tj) and x2(ψ(tj)). if L1 ∩ L2 intersect at an affine point (c1, c2, c3), then else ai := [c1 : c2 : c3 : 1] 9: 10: 11: 12: ai := [v1 : v2 : v3 : 0], where v = (v1, v2, v3) is parallel to L1 and L2. end if plug ai into Eq. (6), and check if x2(ψi(t)) − x1(t) a4 · x1(t) − a1 = y2(ψi(t)) − y1(t) a4 · y1(t) − a2 = z2(ψi(t)) − z1(t) a4 · z1(t) − a3 . in the affirmative case, add Pai to the list of projections (initially empty). 13: 14: end for 15: return the list of projections mapping C1 to C2, or the message No projection has been found. projection from the origin, which is the vertex of the cone. factoring over the rationals, we get In this case, by N (t, s) = −8(t6 + 9t5 + 5st3 − 34t4 + 30st2 − 330t3 + 80ts − 684t2 + 80s −544t − 160) · (st2 − 5t3 + 3ts − 12t2 − 2s − 6t + 4). Therefore, we have ψ1(t) = 5t3 + 12t2 + 6t − 4 t2 + 3t − 2 , ψ2(t) = −t6 − 9t5 + 34t4 + 330t3 + 684t2 + 544t + 160 5(t3 + 6t2 + 16t + 16) . In order to check if ψ1(t) corresponds to some projection, we observe first that t1 = 0, t2 = 1 are not in Gψ1 . The lines Li, i = 1, 2, connecting the points x1(ti) and x2(ψ1(ti)), intersect at the origin. The test in Step 12 of Algorithm 1 is positive, so we conclude that C1 is mapped onto C2 by a perspective projection from the origin, as expected. Furthermore, since ψ1(t) has degree 3, the projec- tion is degenerate. Now we proceed in the same way with ψ2(t). In this case we also take t1 = 0, t2 = 1; however, the corresponding lines Li, i = 1, 2 are skew, so ψ2(t) does not correspond to any projection. The whole computation takes 0.062 seconds. 14 If all the factors of N (t, s) which are linear in s have rational coefficients, then we can find them by applying standard bivariate factoring algorithms over the rationals, implemented in most computer algebra systems. The next result shows that, under the hypothesis that x1(t) and x2(s) are parametrizations with rational coefficients, projections from rational eye points can always be found this way. Proposition 11. If a has rational coordinates, the polynomial G(t, s) associated with Pa has rational coefficients. Proof. Since by hypothesis x2(s) has rational coefficients, the projection plane Π, which is the plane containing the curve parametrized by x2(s), has rational coefficients too. Since also by hypothesis a has rational coefficients, the elements of the matrix P associated with Pa are rational (see Eq. (2)). Therefore, from Diagram (7), ψ has rational coefficients. However, as we saw in Section 3, we can have projections mapping C1 onto C2 from a point with non-rational coefficients. In order to also find these pro- jections, we need to factor N (t, s) over the reals, i.e. one needs an absolute factorization of N (t, s) [10], [12], [15]. This is implemented, for instance, in the computer algebra system Maple 18 through the command AFactors, and works finely for moderate and medium degrees. (cid:16) p1(t) (cid:17) p4(t) , p2(t) p4(t) , p3(t) p4(t) , where Example 2. Let C1 be parametrized by x1(t) = p1(t) = 2t6 + t3 − t2 − 3, p2(t) = −t11 − t9 − 3t8 − 2t5 + t3, p3(t) = −3t11 + t5 − t3 − 3t2 + 2t + 1, (cid:17) p4(t) = −2t10 − t8 − 3t6 − 3t5 − t + 1. (cid:16) q1(s) Also, let C2 be parametrized by x2(s) = q4(s) , q2(s) q4(s) , q3(s) q4(s) , where q1(s) = −4s11 − s9 − 3s8 − s5 − 3s2 + 2s + 1, q2(s) = −3s11 + 2s6 + s5 − 4s2 + 2s − 2, q3(s) = 7s11 + s9 + 3s8 − 2s6 + 7s2 − 4s + 1, q4(s) = −4s11 + 2s10 − s9 − 2s8 + 5s6 + 2s5 + s3 − 4s2 + 3s − 3. One can check that C2 is the image of C1 under the projection defined by the matrix An absolute factorization of N (t, s) yields N (t, s) = N1(t, s)·N2(t, s). The first factor is N1(t, s) = st + s + t − 1. Therefore, it gives rise to  0 1 1 0 −1 −1 1 1 P =  . 1 0 1 0 2 0 1 −1 ψ1(t) = − t − 1 t + 1 . 15 One can check that t1 = 3, t2 = 4 satisfy that t1, t2 /∈ Gψ1 . Then we compute the line L1, connecting the points x1(3), x2(ψ1(3)), and the line L2, connecting the points x1(4), x2(ψ1(4)). These two lines intersect at the affine point (1, 1,−1). Furthermore, the test in Step 12 of Algorithm 1 is positive, so we conclude that C2 is the perspective projection of C1 from the point (1, 1,−1). On the other hand, the second factor N2(t, s) is a big polynomial, of total degree equal to 38, and infinity norm equal to 319426480. Since the degree in s of N2(t, s) is equal to 19, therefore different from 1, N2(t, s) does not correspond to any projection. The whole computation takes 3.526 seconds. 5. An alternative method. In this section we present another method for computing the projections that does not require absolute factoring, and that, additionally, can be adapted to curves which are known up to finite precision, i.e. whose parametrizations have floating point coefficients. The method uses the fact that we have a rational transformation ψ behind each projection, as we saw in the previous section, without actually computing it. We first provide a symbolic version of the method. In general, this symbolic method requires to compute the primitive element of an algebraic extension, and then carry out computations in the new simple extension. However, these operations can be time-consuming. Therefore in practice this symbolic method is only useful when C1 and C2 are low-degree curves. However, one can also have an approximate version of the method, which is really fast. In the symbolic version of the method, we first find tentative values ai for the eye points of the projections mapping C1 onto C2, if any; then, for each ai, we check whether Pai(C1) = C2. In the approximate version, we compute approximations of the ai, and then we evaluate "how close" Pai(C1) and C2 are, i.e. we check whether Pai (C1) ≈ C2, instead of Pai(C1) = C2. 5.1. Symbolic version of the method. The first step is to find tentative values ai for the eye points of the projections mapping C1 onto C2, if any. In order to do this, we need to compute two different lucky pairs, (p1, q1) and (p2, q2), which are related by the same projection. Lemma 12. Let (p1, q1) and (p2, q2) be two different lucky pairs such that pj ∈ C1, qj ∈ C2, pj = x1(tj), qj = x2(sj), tj, sj ∈ R for j = 1, 2, and the one real branch of the curve N (cid:63), such that there is no point ((cid:98)t,(cid:98)s) ∈ N (cid:63), with points (t1, s1) and (t2, s2) belong to N (cid:63). If (t1, s1) and (t2, s2) are connected by (cid:98)t ∈ [t1, t2], where the partial derivative ∂N ∂t vanishes, then there is at most one projection Pa mapping C1 onto C2, such that Pa(pj) = qj for both j = 1, 2. Proof. If (t1, s1) and (t2, s2) are connected by a real branch of N (cid:63) in the con- ditions of the lemma, then there is at most one rational function s = ψ(t), whose graph is contained in N (cid:63), connecting (t1, s1) and (t2, s2). Since different projections correspond to different ψ(t)s, the result follows. 16 Figure 7: Illustrating Lemma 12: ∂t If (p1, q1) and (p2, q2) satisfy the hypotheses of Lemma 12, we will say that the corresponding points (t1, s1), (t2, s2) of N (cid:63) are compatible. This notion, together with Lemma 12, is illustrated in Figure 7. Now we proceed in the following way. Consider a line t = t1, which is not a vertical asymptote of N (cid:63), and not containing any point of N (cid:63) where the partial derivative ∂N ∂t vanishes. Notice that since N (t, s) is square-free, there are just finitely many points of N (cid:63) where ∂N is zero. Therefore, the line t = t1 intersects all the graphs of the different functions s = ψ(t) associated with the projections mapping C1 onto C2. So let s1,1 < ··· < s1,m be the real roots of n1(s) = N (t1, s). For each s1,i, by Lemma 9 there is just one projection, if any, mapping x1(t1) → x2(s1,i). Furthermore, taking t2 > t1 such that there is no t-value in [t1, t2] corresponding to a vertical asymptote of N (cid:63), or a point of N (cid:63) where ∂N ∂t vanishes, n2(s) = N (t2, s) also has m real roots s2,1 < ··· < s2,m. Therefore, for each i = 1, . . . , m the points (t1, s1,i), (t2, s2,i) are compatible, so they correspond to the same projection Pai , if any. Hence, by intersecting the lines Li and Li, connecting the points x1(t1), x2(s1,i) and x1(t2), x2(s2,i), respectively, we get tentative values for the ai. If s1,i and s2,i are rational then we can compute ai exactly. Otherwise we need to work in an algebraic extension. However, in general we will have two different algebraic numbers β, γ for s1,i and s2,i respectively, so we need to compute the primitive element α of a double field extension Q(β, γ); an algorithm to do so can be found in page 145 of [20]. Then the intersection point Li ∩ Li can be determined doing computations in Q(α). (cid:98)x1(t) = ((cid:98)x1(t),(cid:98)y1(t),(cid:98)z1(t)) = Pai (x1(t)). whether Pai(C1) = C2. In order to do this, let Once the tentative ai have been computed, we need to check, for each i, Since Pai(C1) and C2 are rational they are irreducible, so Pai(C1) and C2 are equal iff they have infinitely many points in common. In order to check this, let us denote the numerators of(cid:98)x1(t)− x2(s),(cid:98)y1(t)− y2(s),(cid:98)z1(t)− z2(s) by n1(t, s), n2(t, s), n3(t, s). Then Pai(C1) = C2 iff gcd(n1(t, s), n2(t, s), n3(t, s)) (cid:54)= 1. problem. Algorithm 2 is illustrated in a toy example, Example 3. So we get the following algorithm, Algorithm 2, for solving the projection 17 CompatibleNon-compatibleN?Algorithm 2 Require: Two proper, rational parametrizations x1(t), x2(s) defining two space curves C1,C2, where C2 is planar. 6: Eq. (9). Ensure: The projections, if any, mapping C1 onto C2. 1: compute the polynomial N (t, s) as the numerator of the left hand-side of 2: pick t1, t2 ∈ Q such that t = t1, t = t2 are not asymptotes of N (cid:63), and such 3: Let sj,1, . . . , sj,m, j = 1, 2, be the real roots of nj(s) = N (tj, s). 4: for i = 1, . . . , m do 5: that no singularity of N (cid:63) lies in [t1, t2] × R. check if the pair (x1(t1), x2(s1,i)), is lucky; in the negative case, pick a different value for t1. check if the pair (x1(t2), x2(s2,i)), is lucky; in the negative case, pick a different value for t2. compute the line Li connecting the points x1(t1) and x2(s1,i). compute the line Li connecting the points x1(t2) and x2(s2,i). compute the intersection point ai (possibly at infinity), if any, of Li and Li; if Li ∩ Li = ∅, i := i + 1. find Pai. if Pai(C1) = C2 then 12: end if 13: 14: end for 15: return the list of projections mapping C1 to C2, or the message No add Pai to the list of projections (initially empty). 10: 11: 8: 9: 7: projection has been found. (cid:16)−t2+2t+1 , t2+2t−1 (cid:17) (cid:16) 1−s2 (cid:17) Example 3. Let C1 be parametrized by x1(t) = C2 be parametrized by x2(s) = . One can easily recognize that C1 and C2 are two circles, placed on parallel planes; therefore, we should find two different perspective projections mapping C1 onto C2. In order to check this, we first compute N (t, s), which yields 1+s2 , 2s t2+1 , 0 1+s2 , 1 , and let t2+1 N (t, s) = −4s2t2 − 8s2t + 8st2 + 4s2 − 16st + 4t2 − 8s + 8t − 4. Now we proceed with Algorithm 2. In order to do this, we pick t1 = 0 and t2 = 1, both satisfying the requirements in Step 2. For t1 = 0, we get n1(s) = N (0, s) = 4s2 − 8s − 4, which is irreducible over Q. The pair (x1(0), x2(β)), where 4β2 − 8β + 4 = 0, is lucky; observe that since the polynomial 4β2 − 8β + 4 has two different roots, in fact (x1(0), x2(β)) represents two different pairs, that we can treat at the same time. Also, for t = t2 we get n2(s) = −8s2 − 16s + 8, which is irreducible over Q too. The pair (x1(1), x2(γ)), where −8γ2−16γ +8 = 0, is also lucky. Furthermore, one can check that γ = −β, so we can write (x1(1), x2(−β)), instead, where 4β2 − 8β + 4 = 0; as before, (x1(1), x2(−β)) represent two different pairs that we treat simultaneously. Now the parametric 18 representation of the line Lβ connecting x1(0) and x2(β) is 2β Lβ ≡ 1 + µ · 1 − ,−1 + µ · −1 + β2 + 1 1 − β2 1 + β2 (cid:20) (cid:20) (cid:18) (cid:18) (cid:19) (cid:19) (cid:18) (cid:18) (cid:19) (cid:21) ,−µ, , (cid:19) (cid:21) (14) (cid:19) where µ is the parameter. The parametric representation of the line Lβ con- necting x1(1) and x2(−β) is Lβ ≡ 1 + ν · 1 − 1 − β2 1 + β2 , 1 + ν · −1 + 2β β2 + 1 ,−ν, , where ν is the parameter. Lβ and Lβ intersect at (0, 0, 3 + β). Since the two roots of 4β2 − 8β + 4 = 0 are real, we have two candidates for the eye point of a projection mapping C1 onto C2, that we can analyze simultaneously. In order to check whether or not (0, 0, 3 + β) corresponds to some projection of C1 onto C2, we substitute aβ = [0 : 0 : 3 + β : 1] into Eq. (2), to get  . 0 0 0 0 1 −(3 + β) 1 −(3 + β) P = 0 0 0 0 −(3 + β) − 1 −(3 + β) − 1 Paβ (x1(t)) = ((cid:98)x1(t),(cid:98)y1(t),(cid:98)z1(t)) = (cid:18) (−2 − β)(−t2 + 2t + 1) 0 0 = Hence, we have Finally, the gcd of the numerators of the components of (cid:98)x1(t) − x2(s) is (−3 − β)(t2 + 1) (−3 − β)(t2 + 1) (−2 − β)(t2 + 2t − 1) , , 1 . −sβ + βt + st − 2s + 2t + 1, therefore different from 1. So Paβ (C1) = C2. The whole computation takes 0.234 seconds. As mentioned before, from a computational point of view the drawback of Algorithm 2 is the fact that it requires computing first a primitive element, and then carry out computations in the new simple algebraic extension. Hence, in general it does not work better than Algorithm 1. However, the idea behind Algorithm 2 is useful to provide an approximate method, which is done in the next subsection. 5.2. Approximate version of the method. The computation of a simple algebraic extension is necessary when we want to symbolically carry out the intersection of the lines Li and Li. We can alter- natively find an approximation of Li∩ Li by proceeding in the following way: (a) approximate the values of si and si, so that we also approximate the equations 19 of the lines Li and Li; (b) approximate Li ∩ Li, to get approximations of the ai; (c) evaluate "how close" Pai(C1) and C2 are. In order to evaluate the closeness between Pai(C1) and C2, observe first that a projective parametrization of Pai (C1) is provided by P · x1(t), where P is the matrix in Eq. (2), and x1(t) = [x1(t) : y1(t) : z1(t) : 1]. Therefore, Pai(C1) and C2 are two curves contained in the same plane Π. To measure how close they are, we can use the Hausdorff distance [2] between these two curves. More precisely, given a metric space (X, d), the distance between an element a ∈ X and a non-empty set B ⊂ X is defined as d(a, B) = infb∈Bd(a, b). Furthermore, given two non-empty subsets A, B ⊂ X, the Hausdorff distance between A, B is H(A, B) = max{supa∈A{d(a, B)}, supb∈B{d(b, A)}}. In our case, d will be the usual Euclidean distance, so we will just write H(A, B). The Hausdorff distance is a natural and widely used measure to evaluate the closeness between two objects, and, in particular, to check whether two objects are approximately equal. One can find algorithms to approximate the Hausdorff distance in [4], [9], [16], [14], for the case of planar curves; in [17], [18], for space curves; and in [3], [5], [7] for surfaces. Additionally, some theoretical aspects of the question are treated in [6], for algebraic curves in n-space. Therefore, in order to decide whether Pai(C1) ≈ C2, we compute the Haus- dorff distance between Pai(C1) and C2 (by whatever method), and fix a certain tolerance . Then we will admit Pai(C1) ≈ C2 whenever H(Pai(C1),C2) < . So we get the following algorithm, Algorithm 3, to find the ai such that Pai(C1) ≈ C2. Notice that this algorithm is also applicable to the case of curves with perturbed coefficients, so that the coefficients of the definining parametriza- tions are known only up to a certain precision. (cid:16) p1(t) (cid:17) Example 4. Let C1 be parametrized by x1(t) = p4(t) , p2(t) p4(t) , p3(t) p4(t) , where p1(t) = −2t4 − t3 − 2t2 − 2t, p2(t) = −t4 − t3 − 3t2 − 2t + 1, p3(t) = t3 + 2t2 − 2t − 1, p4(t) = t4 + 1. (cid:16) q1(s) Also, let C2 be parametrized by x2(s) = q4(s) , q2(s) q4(s) , q3(s) q4(s) (cid:17) , where q1(s) = −s2(3s2 + 2s + 3), q2(s) = −(2s3 + 6s2 − 3), q3(s) = 3s4 + 4s3 + 9s2 − 3, q4(s) = 6s4 + s3 + 3s2 + 6s + 3. 20 Algorithm 3 Require: Two proper, rational parametrizations x1(t), x2(s) defining two Eq. (9). space curves C1,C2, where C2 is planar. tolerance . Ensure: The eye points ai, if any, such that Pai(C1) ≈ C2, under a certain 1: compute the polynomial N (t, s) as the numerator of the left hand-side of 2: pick t1, t2 ∈ Q such that t = t1, t = t2 are not asymptotes of N (cid:63), and such 3: Let sj,1, . . . , sj,m, j = 1, 2, be the real roots of nj(s) = N (tj, s). 4: for i = 1, . . . , m do 5: that no singularity of N (cid:63) lies in [t1, t2] × R. check if the pair (x1(t1), x2(s1,i)), is lucky; in the negative case, pick a different value for t1. check if the pair (x1(t2), x2(s2,i)), is lucky; in the negative case, pick a different value for t2. compute the line Li connecting the points x1(t1) and x2(s1,i). compute the line Li connecting the points x1(t2) and x2(s2,i). approximate the intersection point ai of Li and Li. compute Pai(C1). 10: compute or estimate H(Pai (C1),C2). 11: if H(Pai (C1),C2) < then 12: 13: end if 14: 15: end for 16: return the list of projections such that Pai(C1) ≈ C2, or the message No add Pai to the list of "approximate" projections (initially empty). projection has been found. 6: 7: 8: 9: One can check that C2 is the image of C1 under the projection defined by the matrix −2 P =  . 1 1 1 1 −2 1 1 −2 1 1 0 0 0 1 −3 Furthermore, C2 is contained in the plane x + y + z = 0. The polynomial N (t, s) has two irreducible factors over Q, the first one being t − s, the second one a dense bivariate polynomial of degree 10. The first factor gives rise to the rational function ψ(t) = t. By applying Algorithm 1, one can check that this function corresponds to a perspective projection mapping C1 onto C2 from the affine point (1, 1, 1). Furthermore, by applying Algorithm 2, one can check that the second factor of N (t, s) does not provide any other projection. 21 Now let us perturb C2 in the following way: q1(s) = −s2(3s2 + 2s + 3) + 1 q2(s) = −(2s3 + 6s2 − 3), q3(s) = 3s4 + 4s3 + 9s2 − 3 − 1 q4(s) = 6s4 + s3 + 3s2 + 6s + 3. 1000 , 1000 , For simplicity, we keep the notation C2 for this perturbed curve as well. Now C2 is not exactly the projection of C1 anymore; however, it is expectable that Pai(C1) ≈ C2 for some ai close to (1, 1, 1). In this case, the new polynomial N (t, s) is absolutely irreducible. We pick t1 = 1, t2 = 3 2 , which satisfy the conditions in step 2 of Algorithm 3. The real roots of n1(s) = N (1, s), are −16.82354557, − 1.536480878, − 0.4983244587, 1.000123802. Also, the real roots of n2(s) = N (3/2, s), are −5.921589025, − 1.022763869, − 0.5901004394, 1.499976466. One can check that the pair x1(1), x2(1.000123802), is lucky. Similarly, the pair x1(3/2), x1(1.499976466) is also lucky. Furthermore, since the points (1, 1.000123802), (3/2, 1.499976466) are connected by a real branch of N (cid:63), and no singularity of N (cid:63) lies in this branch in between the points, according to Lemma 12 the points x1(1), x2(1.000123802) and x1(3/2), x1(1.499976466) are related by at most one projection, if any. In order to find the eye point of that projection, we compute the lines L1, L1 connecting x1(1), x2(1.000123802) and x1(3/2), x1(1.499976466), respectively. One can check that these lines are skew, but the least squares method provides the solution a1 = (0.998393660, 0.998630969, 0.9998988141) (cid:16) q1(s) for the intersection point of L1, L1. Notice that this point is really very close to the point (1, 1, 1). The projection Pa1(C1) is the curve C2, parametrized by x2(s) = , where q4(s) , q2(s) q4(s) , q3(s) (cid:17) q4(s) q1(s) = 2.998665906s4 + 1.998529783s3 + 2.998665906s2 + 0.003484926s, q2(s) = 0.001030536s4 + 1.998292474s3 + 5.994877422s2 + 0.002061072s −2.996923443, +2.996923443, q3(s) = −2.999696442s4 − 3.996822257s3 − 8.993543328s2 − 0.005545998s q4(s) = −5.996923443s4 − s3 − 3s2 − 6s − 2.996923443. Although C2 (cid:54)= C2, one can check that the infinity norm of qi(t) − qi(t) is small, suggesting that Pa1(C1) ≈ C2. More precisely, in Figure 8 we have plotted C2 (in blue color) and C2 (in red color). Furthermore, we have plotted C2 over a 22 Figure 8: The curves C2 and C2 (cid:17) , a2 = (13.96606371, 26.47971848,−6.649856672). smaller parameter interval than in the case of C2, so that we can appreciate the almost perfect overlapping between C2 and C2. Furthermore, using the technique of [17], the Hausdorff distance between C2 and C2 is 9.72 · 10−6. Therefore, if we choose = 0.001, which is the size of the perturbation we introduced in the second curve, we have that Pa2(C1) ≈ C2. Similarly, one can check that the pairs x1(1), x2(−0.4983244587) and x1(3/2), x1(−0.5901004394) are also lucky. By proceeding as before, we get the tentative (cid:16)(cid:98)q1(s) eye point The projection Pa2(C1) is the curve (cid:98)C2, parametrized by(cid:98)x2(s) = (cid:98)q4(s) , (cid:98)q3(s) (cid:98)q4(s) , (cid:98)q2(s) (cid:98)q4(s) where(cid:98)q1(s) = 25.69365991s4 + 19.82986181s3 + 25.69365991s2 − 16.20453122s, (cid:98)q2(s) = −45.64322992s4 + 7.316207038s3 + 21.94862111s2 − 91.28645982s (cid:98)q3(s) = 19.94957001s4 − 27.14606885s3 − 47.64228102s2 + 107.4909911s (cid:98)q4(s) = −36.79592552s4 − s3 − 3s2 − 6s − 33.79592552. In this case, the infinity norm of qi(t) −(cid:98)qi(t) is not small. Furthermore, in Figure 9 we have plotted C2 (in blue color) and (cid:98)C2 (in red color): one can see Hausdorff distance between C2 and (cid:98)C2 is 0.253 > 0.001. Therefore, Pa2(C1) (cid:54)≈ C2. that both curves are very different. Furthermore, using the technique of [17], the −33.79592552, +33.79592552, The other tentative eye points behave in a similar way. 6. Conclusions and Further Work. We have presented three algorithms, two of them symbolic and one of them approximate, to check whether or not a planar, rational curve C2 properly is the (parallel or perspective) projection of another rational parametrized, 23 Figure 9: The curves C2 and (cid:98)C2 curve C1, also properly parametrized, non-necessarily planar. In the affirma- tive case, the algorithms compute all the projections mapping C1 onto C2. All the algorithms are based on the fact that behind each projection, we have a ra- tional mapping between the parameter spaces of C1 and C2. The first algorithm, Algorithm 1, uses bivariate factoring. In order to compute the projections from eye points with rational coordinates, standard bivariate factoring over the ratio- nals suffices. However, in the most general case we need to compute an absolute factorization of a bivariate polynomial. The second algorithm, Algorithm 2, does not use factoring, but in turn it requires, in general, to compute the prim- itive element of a double field extension. Except for easy, low degree cases, this operation can be costly. However, this second algorithm gives rise to a third, approximate, algorithm, Algorithm 3. This last algorithm avoids the computation of the primitive element by changing to a floating point setting. In this last case, we compute approximations for tentative eye points, and then we evaluate whether C2 is "aproximately" the projection of C1. In order to do this, we compute the Hausdorff distance between C2 and the projection of C1 from the tentative eye point. The algorithms strongly exploit the fact that C1 and C2 are rational curves. Therefore, as such they cannot be generalized to the case when C1 and C2 are implicit, not necessarily rational, algebraic curves. Hence, an interesting prob- lem to be attacked in the future is the development of algorithms for solving the problem in the implicit case. References References [1] Alc´azar J.G. (2012), Computing the shapes arising in a family of space rational curves depending on one parameter, Computer Aided Geometric Design vol. 29, pp. 315 -- 331 24 [2] Aliprantis C.D., Border K.C. (2006), Infinite Dimensional Analysis, Springer-Verlag. [3] Aspert N., Santa-Cruz D., Ebrahimi T. (2002), MESH: measuring errors between surfaces using the Hausdorff distance, Multimedia and Expo, 2002. ICME '02. Proceedings 2002 IEEE International Conference on Multimedia and Expo (ICME 02), pp. 705 -- 708. [4] Bai Y.-B., Yong,J.-H., Liu,C.-Y., Liu,X.-M., Meng,Y. (2011), Polyline approach for approximating Hausdorff distance between planar free-form curves, Computer Aided Design Vol. 43 (6), pp. 687698. [5] Barton M., Hanniel I., Elber G., Kim M-S. (2010), Precise Hausdorff Dis- tance Computation between Polygonal Meshes, Computer Aided Geometric Design Vol. 27, Issue 8, pp. 580 -- 591. [6] Blasco A., P´erez D´ıaz S. (2015), Characterizing the finiteness of the Haus- dorff distance between two algebraic curves, Journal of Computational and Applied Mathematics, Vol. 280, No. 1, pp. 327 -- 346. [7] Bronstein A., Bronstein M, Kimmel R. (2006), Efficient computation of isometry-invariant distances between surfaces, Siam J. Sci. Comput. Vol. 28, No 5., pp. 18121836. [8] Burdis J., Kogan I., Hong H. (2013), Object-Image Correspondence for Algebraic Curves Under Projections, SIGMA 9, 023, 31 pages. [9] Chen X.D., Ma W., Xu G., Paul J.C. (2010), Computing the Hausdorff distance between two B-spline curves, Computer Aided Design Vol. 42, pp. 11971206. [10] Corless R., Galligo A., Kotsireas I., Watt S. (2002), A Geometric-Numeric Algorithm for Absolute Factorization of Multivariate Polynomials, Proceed- ings ISSAC 2002, pp. 37 -- 45, ACM New York, NY, USA. [11] Goldman R. (2009), Computer Graphics and Geometric Modeling, CRC Press, Taylor & Francis Group, Boca Raton (USA). [12] Gao S. (2003), Factoring multivariate polynomials via partial differential equations, Mathematics of Computation Vol. 72, pp. 801 -- 822. [13] Hartley R., Zisserman A. (2003), Multiple view geometry in computer vi- sion, Cambridge University Press, New York (USA). [14] Juttler B. (2000), Bounding the Hausdorff distance of implicitly defined and/or parametric curves, Mathematical Methods for Curves and Surfaces, pp.223232. [15] Kaltofen E., May J.P., Yang Z., Zhic L. (2008), Approximate factorization of multivariate polynomials using singular value decomposition, Journal of Symbolic Computation, Vol. 43, Issue 5, pp. 359 -- 376. 25 [16] Kim Y.-J., Oh Y.-T., Yoon S.-H., Kim M.-S., Elber G. (2010), Precise Hausdorff distance computation for planar free form curves using biarcs and depth buffer, Visual Computing Vol. 26(68), pp. 10071016. [17] Rueda S., Sendra J.R., Sendra J. (2013), An algorithm to parametrize ap- proximately space curves, Journal of Symbolic Computation, Vol. 56, pp. 80 -- 106. [18] Rueda S., Sendra J.R., Sendra J. (2014), Bounding and Estimating the Hausdorff distance between real space algebraic curves, Computer Aided Geometric Design Vol. 31, Issue 34, pp. 182 -- 198. [19] Sendra J.R., Winkler F., P´erez-D´ıaz S. (2008), Rational Algebraic Curves, Springer-Verlag. [20] Winkler F. (1996), Polynomial algorithms in computer algebra, Springer- Verlag. [21] Struik D.J. (1961), Lectures On Classical Differential Geometry, Addison- Wesley Pub. Co. 26
1506.06424	1	1506	2015-06-21T23:39:45	The volume of a set of arcs on a variety	[ "math.AG" ]	In this paper, we give a definition of volume for subsets in the space of arcs of an algebraic variety, and study its properties. Our main result relates the volume of a set of arcs on a Cohen-Macaulay variety to its jet-codimension, a notion which generalizes the codimension of a cylinder in the arc space of a smooth variety.	math.AG	math	THE VOLUME OF A SET OF ARCS ON A VARIETY TOMMASO DE FERNEX AND MIRCEA MUSTAT¸ A Dedicated to Lucian Badescu on the occasion of his seventieth birthday 1. Introduction In this paper, we give a definition of volume for subsets in the space of arcs of an algebraic variety and study its properties. As our definition implies that the volume of a set of arcs is finite if and only if its projection to the variety is a finite set of closed points, we can restrict without loss of generality to the case of an affine variety. Suppose therefore that X = Spec(R) is an n-dimensional affine algebraic variety defined over an algebraically closed field of characteristic zero. For every ideal a in R we denote by ℓ(R/a) the length of the quotient ring R/a and, if the cosupport consists of one point x defined by the ideal mx, we denote by e(a) the Hilbert -- Samuel multiplicity of Rmx with respect to aRmx. Let X∞ be the arc scheme of X. Recall that for every field extension K/k, the K-valued points of X∞ are in natural bijection with the arcs Spec K[[t]] → X (see [EM09, Section 3]). For every subset C ⊆ X∞ and any integer m ≥ 0, we consider the ideal bm(C) := {f ∈ R ordγ(f ) ≥ m for all γ ∈ C}. This defines a graded sequence of ideals b•(C) = (bm(C))m≥0. We then define the volume of C by the formula vol(C) := vol(b•(C)) = lim sup m→∞ ℓ(R/bm(C)) mn/n! . It follows from [Cut14] that the limsup is, in fact, a limit. It is easy to see that vol(C) < ∞ if and only if the image of π(C) in X is a finite set of closed points. Here π : X∞ → X is the canonical projection mapping an arc γ to its base point γ(0). The volume satisfies the following inclusion/exclusion property. Proposition 1.1. If C1, C2 ⊆ X∞, then vol(C1 ∪ C2) + vol(C1 ∩ C2) ≤ vol(C1) + vol(C2). The contact locus of order at least q of an ideal a ⊆ R is defined to be Cont≥q(a) = {γ ∈ X∞ ordγ(a) ≥ q}. Contact loci form a special class of subsets in X∞. For ideals cosupported at one point, the volumes of these sets relate to the Samuel multiplicities of the ideal in the following way. Proposition 1.2. For every ideal a ⊆ R whose cosupport consists of one point and for every m, p ≥ 1, we have mn · vol(Cont≥m(a)) ≤ (mp)n · vol(Cont≥mp(a)) ≤ e(a) 2010 Mathematics Subject Classification. Primary 13H15, 14E18; Secondary 13A18, 14B05. Key words and phrases. Arc space, graded sequence of ideals, valuation, Mather discrepancy. The research of the first author was partially supported by NSF grants DMS-1402907 and DMS-1265285. The research of the second author was partially supported by NSF grants DMS-1401227 and DMS-1265256. 1 2 TOMMASO DE FERNEX AND MIRCEA MUSTAT¸ A for every m, p ≥ 1. Furthermore, both inequalities are equalities for m sufficiently divisible. Generalizing the definition of codimension of a cylinder in the space of arcs of a smooth variety, we define the jet-codimension of an irreducible closed subset C of X∞ to be jet-codim(C) := lim p→∞(cid:16)(p + 1)n − dim πp(C)(cid:17) where πp : X∞ → Xp is the truncation map to the p-jet space. This definition extends to an arbitrary set C ⊆ X∞ by taking the smallest jet-codimension of the irreducible components of the closure C of C in X∞. We will see, for instance, that if X is smooth, then the jet- codimension of a set C coincides with its Krull codimension codim(C) (which is similarly defined as the smallest Krull codimension of an irreducible component of C). Our main result relates the volume of a set of arcs on a Cohen -- Macaulay variety to its jet-codimension. Theorem 1.3. If X is Cohen -- Macaulay, of dimension n, then for every subset C ⊆ X∞ whose image in X is a closed point we have vol(C)1/n · jet-codim(C) ≥ n. In particular, if X is smooth, then vol(C)1/n · codim(C) ≥ n. The proof of this theorem requires a suitable extension of the main result of [dFEM04] to singular varieties, which we discuss next. Let a ⊆ R be an m-primary ideal, where m ⊂ R is a maximal ideal. If X is smooth, then the colength and the Hilbert -- Samuel multiplicity of a are related to the log canonical threshold lct(a) by the formulas (1) (2) (n! · ℓ(OX /a))1/n · lct(a) ≥ n, e(a)1/n · lct(a) ≥ n. We want to extend this result to all Cohen -- Macaulay varieties. If X is singular, then the log canonical threshold (even when it is defined) is not the right invariant to consider. Instead, we look at the Mather log canonical threshold of the ideal [Ish13], which is defined by where the infimum ranges over all birational morphisms f : Y → X, with Y smooth, and all prime divisors E ⊂ Y , with Jacf being the Jacobian ideal of f . clct(a) := inf f,E ordE(Jacf ) + 1 ordE(a) Theorem 1.4. With the above notation, if X is Cohen -- Macaulay, of dimension n, then we have (3) (4) (n! · ℓ(OX /a))1/n ·clct(a) ≥ n, e(a)1/n ·clct(a) ≥ n. The proofs of (1) and (2) rely on the reduction to monomial ideals via flat degenera- tion, where the inequality follows from Howard's description of log canonical thresholds of monomial ideals and the well-known inequality between arithmetic and geometric means. A slightly more general formulation of (2) is the key ingredient in the proof of a theorem of [dFEM03] on log canonical thresholds of generic projections. The proof of Theorem 1.4 fol- lows the opposite direction: we first prove a theorem on Mather log discrepancies of generic projections (see Theorem 2.5 below), and then deduce (3) and (4) from it. THE VOLUME OF A SET OF ARCS ON A VARIETY 3 The paper is organized as follows. In the next section we prove Theorem 1.4. Section 3 is devoted to a discussion of volumes of graded sequences of ideals, with emphasis on sequences associated to pseudo-valuations. Then, in the last section we define the volume of a set of arcs and prove several properties including those stated above. Acknowledgments. The results in Section 2 are heavily inspired by our work with Lawrence Ein in [dFEM03,dFEM04], and it is a pleasure to thank him for many enlightening discussions throughout these years. We would like to thank also Rob Lazarsfeld for useful discussions on these topics. 2. Mather log discrepancies Let X be a variety of dimension n defined over an algebraically closed field of characteristic zero. Recall that a divisor over X is a prime divisor E on a normal variety Y , with a birational morphism f : Y → X. Such a divisor determines a valuation ordE of k(Y ) = k(X) and as usual, we identify two divisors over X if they give the same valuation. The valuations that arise in this way are the divisorial valuations of k(X) that have center on X (recall that the center of ordE is the closure of f (E)). Given a birational morphism f : Y → X, with Y smooth, we consider Jacf := Fitt0(ΩY /X) ⊆ OY , the Jacobian ideal of the map. Definition 2.1. Given a divisor E over X, the Mather log discrepancy baE(X) of E over X is defined as follows. Suppose that f : Y → X is a birational morphism, with Y normal, such that E is a prime divisor on Y . After possibly replacing Y by its smooth locus, we may assume that Y is smooth. If Jacf ⊆ OY is the Jacobian ideal of the map, then Given a nonzero ideal sheaf a ⊂ OX and a number c ≥ 0, we define the Mather log discrepancy of E with respect to the pair (X, ac) to be baE(X) := ordE(Jacf ) + 1. baE(X, ac) := ordE(Jacf ) + 1 − c · ordE(a). log canonical threshold of the pair (X, a), with a a proper nonzero ideal of R, is defined by tively. It is clear that the definition of Mather log discrepancy only depends on the valuation ordE that E defines on the function field of X, and not on the model Y . We say that the pair When X is smooth, we write aE(X) and aE(X, ac) instead of baE(X) and baE(X, ac), respec- (X, ac) is Mather log canonical if for every E as above, we have baE(X, ac) ≥ 0. The Mather It is straightforward to check that this is equivalent to the definition of clct(a) given in Introduction. We put, by convention, clct(0) = 0 and clct(OX ) = ∞. clct(a) := sup{ c ∈ R≥0 (X, ac) is Mather log canonical }. Remark 2.2. We refer to [Ish13] for basic facts about Mather log discrepancies and Mather log canonical threshold. A useful fact is that if f : Y → X is a log resolution of (X, a) which factors through the Nash blow-up of X, then there is a divisor E on Y such that clct(a) = baE (X) ordE (a) . We will use several times the following basic fact about divisorial valuations. Lemma 2.3. Let f : X → Y be a dominant morphism of varieties. If E is a divisor over X, then the restriction of ordE to k(Y ) is a multiple of a divisorial valuation, that is, we can write ordE k(Y ) = q · ordF 4 TOMMASO DE FERNEX AND MIRCEA MUSTAT¸ A for some divisor F over Y and some positive integer q. Proof. Let v = ordE and w = vk(Y ). Note that w is a valuation with center on Y , the center being the closure of the image of the center of v on X. We denote by Rv and Rw the valuation rings corresponding to v and w, respectively, and by kv and kw the corresponding residue fields. Note that trdeg(kw/k) ≤ dim(Y ), with equality if and only if w is the trivial valuation. Furthermore, w is a multiple of a divisorial valuation if and only if trdeg(kw/k) = dim(Y ) − 1 (see [KM98, Lemma 2.45]). On the other hand, since v is a divisorial valuation, we know It follows from [ZS60, Chapter VI.6, Corollary 1] that that trdeg(kv/k) = dim(X) − 1. trdeg(kv/kw) ≤ trdeg(k(X)/k(Y )) = dim(X) − dim(Y ). We conclude that trdeg(kw/k) ≥ dim(Y ) − 1. Since it is clear that w is not the trivial valuation, we conclude that in fact trdeg(kw/k) = dim(Y ) − 1, hence w is a multiple of a divisorial valuation. Since w only takes integer values, it is immediate to see that the multiple is a positive integer. (cid:3) The next result gives an alternative way of computing Mather log discrepancies. Suppose that E is a prime divisor over a normal n-dimensional affine variety X. Given a closed immersion X ֒→ AN and a general linear projection π : AN → Y := An, we may write ordE k(Y ) = q · ordF , for a prime divisor F over Y and a positive integer q, by Lemma 2.3. Proposition 2.4. With the above notation, we have Proof. Consider a commutative diagram baE(X) = q · aF (Y ). X ′ g Y ′ f / X / Y / AN π An where X ′ → X and Y ′ → Y are resolutions such that E is a divisor on X ′ and F is a divisor on Y ′. Note that ordE(g∗F ) = q and ordE(KX ′/Y ′) = q − 1. Denoting by h : X ′ → Y the composition of f with the projection to Y , we have ordE(KX ′/Y ) = ordE(Jach). If x1, . . . , xn is a regular system of parameters in X ′ centered at a general point of E and y1, . . . , yN are affine coordinates in AN , then f is locally given by equations yi = fi(x1, . . . , xn), and Jacf is locally defined by the n × n minors of the matrix (∂fi/∂xj). For a linear projection π : AN → Y = An, Jach is locally defined by a linear combination of the n × n minors of (∂fi/∂xj). If the projection is general, then so is the linear combination, and we have baE(X) = ordE(KX ′/Y ) + 1. Writing KX ′/Y = KX ′/Y ′ + KY ′/Y , we get baE(X) = ordE(KX ′/Y ′) + ordE(g∗KY ′/Y ) + 1 = q · aF (Y ). (cid:3) The following theorem is a generalization of [dFEM03, Theorem 1.1] to Cohen -- Macaulay varieties. Theorem 2.5. Let X ⊆ AN be a Cohen -- Macaulay variety of dimension n, and E a divisor over X. For some 1 ≤ r ≤ n, consider the morphism φ : X → An−r+1 induced by restriction of a general linear projection σ : AN → An−r+1. Write ordE k(An−r+1) = q · ordG, where G is a prime divisor over An−r+1 and q is a positive integer (cf. Lemma 2.3). Let Z ֒→ X a closed Cohen -- Macaulay subscheme of pure codimension r such that φZ is a / / / THE VOLUME OF A SET OF ARCS ON A VARIETY 5 finite morphism. Note that φ∗[Z] is a cycle of codimension one in An−r+1; we regard φ∗[Z] as a Cartier divisor on An−r+1. Then, for every c ∈ R≥0 such that baE(X, cZ) ≥ 0, we have Moreover, if the ideal defining Z in X is locally generated by a regular sequence, then (5) (6) q · aG(cid:18)An−r+1, q · aG(cid:18)An−r+1, r! cr rr · φ∗[Z](cid:19) ≤baE(X, cZ). rr · φ∗[Z](cid:19) ≤baE(X, cZ). cr Proof. Our argument is similar to the one used in the proof of [dFEM03, Theorem 1.1]. We assume that ordE(Z) > 0 (the case ordE(Z) = 0 is easier and left to the reader). We factor σ as a composition of two general linear projections AN → U = An → V = An−r+1. By Lemma 2.3, we can write ordE k(U ) = p · ordF for some prime divisor F over U and some positive integer p. Note that p divides q. Let h : V ′ → V be a proper, birational morphism, with V ′ smooth, such that G is a prime divisor on V ′. Let X ′ := V ′ ×V X and U ′ := V ′ ×V U , and consider the induced commutative diagram with Cartezian squares . φ X ′ ψ′ φ′ U ′ γ ′ V ′ f g h X ψ / U γ / V Let Z ′ := f −1(Z) ֒→ X ′ and Z ′′ := ψ′(Z ′) ֒→ U ′, both defined scheme-theoretically. In general, we have Z ′′ ֒→ g−1(ψ(Z)), but this may be a proper subscheme. First, note that ψ is a finite, flat morphism. Finiteness follows from the fact that it is induced by a generic projection, while flatness follows from the fact that it is finite, U is smooth, and X is Cohen- Macaulay. Since γ is clearly flat (in fact, smooth), we conclude that φ is flat. Therefore both X ′ and U ′ are varieties and f and g are proper, birational morphisms. Furthermore, the restriction φ′Z ′ is finite by base-change, and thus both ψ′Z ′ and γ′Z ′′ are finite. Note that Z ′ is a closed subscheme of ψ′−1(Z ′′), hence (7) p · ordF (Z ′′) = ordE((ψ′)−1(Z ′′)) ≥ ordE(Z ′) = ordE(Z). Since h, being a morphism between two smooth varieties, is a locally complete intersection morphism, it follows by flat base change that f is a locally complete intersection morphism as well. More explicitly, h factors as h = h1 ◦ h2 where h1 : V ′ × V → V is the projection and h2 : V ′ ֒→ V ′ × V is the regular embedding given by the graph of h. By pulling back, we get a decomposition f = f1 ◦ f2 where f1 : V ′ × X → X is smooth and f2 : X ′ ֒→ V ′ × X is a regular embedding of codimension equal to dim V = dim V ′. Recall that the pull-back f ∗[Z] ∈ An−r(X ′) is defined as f ! 2[V ′ × Z] (see [Ful98, Section 6.6]). We now show that Z ′ is pure-dimensional, of the same dimension as Z, and f ∗[Z] is equal to the class of [Z ′] in An−r(X ′). Since φ′Z ′ is finite and φ′(Z ′) is a proper subset of V ′, we see that dim Z ′ ≤ dim V ′ −1 = n−r. On the other hand, Z ′ is locally cut out in V ′ ×Z by dim V ′ equations, hence every irreducible component of Z ′ has dimension at least dim Z = n − r. / / / / 6 TOMMASO DE FERNEX AND MIRCEA MUSTAT¸ A Therefore Z ′ is pure dimensional, of dimension dim Z. Since V ′ × X is Cohen-Macaulay, it follows from [Ful98, Proposition 7.1] that f ∗[Z] = [Z ′] in An−r(X ′). Since ψ′Z ′ : Z ′ → Z ′′ is a finite, dominant morphism of schemes, we see that Z ′′ is also pure dimensional of the same dimension as Z ′, and ψ′ ∗[Z ′] ≥ [Z ′′]. Note that h∗φ∗[Z] and φ′ ∗[Z ′] are divisors on V ′. Since f and h are locally complete intersection morphisms of the same codimension, and since we have seen that f ∗[Z] = [Z ′] in An−r(X ′), it follows from [Ful98, Example 17.4.1] that h∗φ∗[Z] ∼ φ′ ∗[Z ′] (note that while φ and φ′ are not proper morphisms, they are proper when restricted to the supports of Z and Z ′, respectively). Since the two divisors are equal away from the exceptional locus of h, we deduce that h∗φ∗[Z] = ∗[Z ′] by the Negativity Lemma (see [KM98, Lemma 3.39]). We thus conclude that φ′ h∗φ∗[Z] = φ′ ∗[Z ′] ≥ γ′ ∗[Z ′′]. On the other hand, the center C of ordF in U ′ is contained in Z ′′ and dominates G. Since φ′Z ′ is finite, it follows that the map γ′C : C → G is finite. In particular, we have dim(C) = dim(G) = n − r = dim(Z ′′), hence C is an irreducible component of Z ′′. Therefore we have (8) ordG(φ∗[Z]) = ordG(h∗φ∗[Z]) ≥ ordG(γ′ ∗[Z ′′]) ≥ eC([Z ′′]) = ℓ(OZ ′′,C). Let b := kG(V ) denote the discrepancy of G over V , and let H := (γ′)∗G. Note that p · valF (H) = q and since γ′ is smooth, H is a smooth divisor at the generic point of C. Moreover, since KU ′/U = (γ′)∗KV ′/V , we have KU ′/U = bH + R, where R is a divisor that does not contain C in its support. Then, by Proposition 2.4 and equation (7), we see that baE(X, cZ) ≥ p · aF (U ′, cZ ′′ − KU ′/U ) = p · aF (U ′, cZ ′′ − bH). Setting α :=baE(X, cZ)/q, we have aF (U ′, cZ ′′ − (b − α)H) = aF (U ′, cZ ′′ − bH) − α · ordF (H) ≤ α(1 − ordF (H)) ≤ 0, where the last inequality follows from the fact that ordF (H) ≥ 1 and, by assumption, α ≥ 0. This in turn implies (9) ℓ(OZ ′′,C) ≥ (b − α + 1)rr r! cr . Indeed, if b − α ≥ 0, then (9) follows by [dFEM03, Theorem 2.1]. The case b − α < 0 is easier, and follows from [dFEM03, Lemma 2.4] using the same degeneration to monomial ideals (see [dFEM04, Section 2]). Combining (8) and (9), we get · φ∗[Z](cid:19) = q · aG(V ) − q · aG(cid:18)V, r! cr rr as stated in (5) r! cr rr · ordG(φ∗[Z]) ≤ q(b + 1 − (b − α + 1)) =baE(X, cZ), Suppose now that the ideal of Z in X is locally generated by a regular sequence. If IZ ⊆ OX is the ideal sheaf of Z and Zi is an irreducible component of Z, then (10) ℓ(OZ,Zi) = e(IZ OX,Zi) = lim m→∞ r! mr · ℓ(OX,Zi/I m Z OX,Zi). For every m, let Zm ֒→ X be the subscheme defined by I m Z . Since IZ is locally generated by a regular sequence, I m is a locally free OZ -module, and thus is Cohen -- Macaulay (as an OZ -module, hence as an OX -module). Note that OZ is also Cohen -- Macaulay (as an OZ -module, hence as an OX -module). By applying [BH93, Proposition 1.2.9] to the exact sequences of OX -modules Z /I m+1 Z 0 → I m Z /I m+1 Z → OZm+1 → OZm → 0, THE VOLUME OF A SET OF ARCS ON A VARIETY 7 we see by induction that OZm is a Cohen -- Macaulay OX -module, and therefore Zm is a Cohen -- Macaulay scheme. Note that baE(cid:16)X, c m · Zm(cid:17) =baE(X, cZ) lim m→∞ r! mr · [Zm] = [Z] for all m, and by (10). Since φZm if finite for every m, we may apply (5) with (Z, c) replaced by (Zm, c/m) to deduce, after letting m go to infinity, the inequality in (6). (cid:3) Corollary 2.6. With the same assumptions as in the first part of Theorem 2.5, we have (11) lct(An−r+1, φ∗[Z]) ≤ clct(X, Z)r rr/r! . Moreover, if the ideal of Z in X is locally generated by a regular sequence, then (12) lct(An−r+1, φ∗[Z]) ≤ clct(X, Z)r rr . Proof. We apply Theorem 2.5 for a divisor E computing clct(X, Z). We apply the first part of the corollary to prove Theorem 1.4. (cid:3) Proof of Theorem 1.4. Let x ∈ X be the cosupport of a. After replacing X by an affine neighborhood of x, we may assume that we have a closed immersion X ֒→ AN . Let m ≥ 1 be fixed and Zm ֒→ X be the zero-dimensional scheme defined by am. Note that Zm is Cohen -- Macaulay, since it is zero dimensional. Consider a general linear projection AN → A1 and let φ : X → A1 be the induced map. Note that and since we have Then (11) gives clct(X, Zm) = 1 m ·clct(X, Z), φ∗[Zm] = ℓ(OX /am) · [f (x)], lct(A1, φ∗[Zm]) = 1 ℓ(OX /am) . Setting m = 1 and taking n-th roots, we get (3). The formula (4) follows by taking the limit as m goes to infinity and then taking n-th roots. (cid:3) ℓ(OX /am) mn/n! ·clct(X, Z)n ≥ nn. 3. The volume of a graded sequence of ideals We recall, following [ELS03] and [Mus02a], some basic facts about the volume of a graded sequence of ideals. Let k be an algebraically closed field of arbitrary characteristic and let X = Spec(R) be an n-dimensional affine variety over k (in particular, we assume that R is a domain). Recall that a sequence a• = (am)m≥0 of ideals am ⊆ R is a graded sequence of ideals if a0 = R and ap · aq ⊆ ap+q for every p, q ≥ 1. 8 TOMMASO DE FERNEX AND MIRCEA MUSTAT¸ A Definition 3.1. The volume of a graded sequence a• is defined by vol(a•) := lim sup m→∞ ℓ(R/am) mn/n! . Let a• be a graded sequence of ideals in R. The main case for understanding the notion of volume is that when there is a closed point x in X such that for every m ≥ 1, the cosupport of am is equal to {x} (we say that a• is cosupported at x). Note that in this case we have vol(a•) < ∞. Indeed, if N is a positive integer such that mN x ⊆ a1, where mx is the ideal defining x, then mpN x ⊆ ap for every p ≥ 1, hence vol(a•) ≤ N n ·e(mx). In fact, under the same assumption, it follows from [LM09, Theorem 3.8] that the volume of a• can be computed as a limit of normalized Hilbert-Samuel multiplicities. More precisely, we have (13) vol(a•) = lim m→∞ e(am) mn . Moreover, the limit superior in the definition of volume is a limit (14) by [Cut14, Theorem 1]. vol(a) = lim m→∞ ℓ(R/am) mn/n! Remark 3.2. Suppose that a• is a graded sequence of ideals such that ap ⊆ aq whenever p ≥ q. If a• is cosupported at a point x ∈ X, then e(am) mn . Indeed, this is a consequence of (13) and of the fact that vol(a•) = inf m≥1 e(am) mn = inf This equality is a consequence of Lemma 3.7 below. lim m→∞ m≥1 e(am) mn . Remark 3.3. Suppose that a• is a graded sequence of ideals and Γ = {x1, . . . , xr} is a finite set of closed points in X such that for every m ≥ 1, the ideal am has cosupport Γ. For every m ≥ 1, let us consider the primary decomposition am = r\i=1 a(i) m , where each a of ideals. Since (i) (i) m is an ideal with cosupport {xi}. It is clear that each a • is a graded sequence (15) we deduce (16) ℓ(R/am) = vol(a•) = rXi=1 rXi=1 ℓ(R/a(i) m ), vol(a (i) • ). In particular, we see that vol(a•) < ∞ and the assertion in (14) also holds for a•. THE VOLUME OF A SET OF ARCS ON A VARIETY 9 Example 3.4. Suppose that a• is a graded sequence of ideals such that each am, with m ≥ 1, has cosupport equal to a finite set Γ. If a• is such that am is the integral closure of the ideal am, then a• is a graded sequence and vol(a•) = vol(a•). The first assertion follows from the fact that ap · aq is contained in the integral closure of ap · aq, hence in ap+q. In order to see that vol(a•) = vol(a•), we may assume that all am have cosupport at the same point x ∈ X (see Remark 3.3). In this case, since e(am) = e(am) for every m, the assertion follows from (13). Under a mild condition on a• which is often satisfied, we give in the next proposition a new easy proof of the assertions (13) and (14) in the smooth case. Proposition 3.5. Suppose that X = Spec(R) is smooth. If a• is a graded sequence of ideals in R which is cosupported at a point in X, and ap ⊆ aq whenever p ≥ q, then (17) (18) vol(a•) = lim m→∞ ℓ(R/am) mn/n! = inf m≥1 ℓ(R/am) mn/n! = lim m→∞ e(am) mn = inf m≥1 e(am) mn . Note that while the proposition recovers (13) and (14) in the smooth setting, it also implies ℓ(R/am) mn/n! , which needs the smoothness assumption. For the proof the equality vol(a•) = infm≥1 of the proposition we need two lemmas. The first one is a special case of [KN14, Lemma 25]; for completeness, we include the proof of this special case. Lemma 3.6. If X = Spec(R) is smooth, x ∈ X is a closed point defined by mx, and a is an mx-primary ideal in R, then for every p ≥ 1, we have ℓ(R/a) ≥ 1 pn · ℓ(R/ap). Proof. Since X is smooth, it is straightforward to reduce to the case when X = An and a is an ideal supported at the origin. We choose a monomial order on R = k[x1, . . . , xn] and for every ideal b in R, we consider the initial ideal in(b) = (in(f ) f ∈ b). We refer to [Eis95, Chapter 15] for the basic facts about initial ideals. Note that we have ℓ(R/b) = ℓ(R/ in(b)). It is clear that in(ap) ⊇ in(a)p. It follows that if we know the assertion in the lemma for in(a), then ℓ(R/a) = ℓ(R/ in(a)) ≥ 1 pn · ℓ(R/ in(a)p) ≥ 1 pn · ℓ(R/ in(ap)) = 1 pn · ℓ(R/ap), hence we obtain the assertion for a. The above argument shows that we may assume that a is a monomial ideal. For every such ideal a, we consider the sets Q(a) := [xu∈a (u + Rn) and Qc(a) := Rn ≥0 r Q(a). Note that Qc(a) is equal, up to a set of measure zero, to the union of ℓ(R/a) disjoint open unit cubes. Therefore ℓ(R/a) is equal to vol(Qc(a)), the Euclidean volume of Qc(a). On the 10 TOMMASO DE FERNEX AND MIRCEA MUSTAT¸ A other hand, it is clear from definition that Q(ap) ⊇ p · Q(a), hence Qc(ap) ⊆ p · Qc(a). We thus conclude ℓ(R/a) = vol(Qc(a)) ≥ vol(cid:18) 1 p · Qc(ap)(cid:19) = 1 pn · vol(Qc(ap)) = 1 pn · ℓ(R/ap). This completes the proof of the lemma. (cid:3) The following is a variant of [Mus02a, Lemma 2.2]. Lemma 3.7. If (αm)m≥1 is a sequence of non-negative real numbers that satisfies the fol- lowing two conditions: i) αpq ≤ p · αq for every p, q ≥ 1, and ii) αp ≥ αq whenever p ≥ q, then lim m→∞ αm m = inf m≥1 αm m . Proof. Let λ := inf m m ≫ 1. By definition, there is d > 0 such that αd d < λ + ǫ where 0 ≤ i < d (hence j = ⌈m/d⌉). The hypotheses imply αm m . We need to show that for every ǫ > 0, we have αm m ≤ λ + ǫ for all 2 . Given m, we write m = jd − i, αm m ≤ αjd jd − i ≤ αd d · jd jd − i ≤(cid:16)λ + ǫ 2(cid:17) · jd jd − i . For m ≫ 1, we have j ≫ 1, hence jd jd−i < λ+ǫ λ+ ǫ 2 . This completes the proof of the lemma. (cid:3) Proof of Proposition 3.5. Let αm = ℓ(R/am). If p ≥ q, then by assumption ap ⊆ aq, hence αp ≥ αq. Moreover, it follows from Lemma 3.6 that αpq ≤ p · αq for all p, q ≥ 1. The two equalities in (17) now follow from the definition of volume and Lemma 3.7. Note now that Lemma 3.7 also gives the second equality in (18). Indeed, for p ≥ q, q ⊆ apq implies e(apq) ≤ q ) = pn · e(aq). In order to prove the first equality in (18), note first that by definition of we have ap ⊆ aq, hence e(ap) ≥ e(aq); moreover, the inclusion ap e(ap Hilbert -- Samuel multiplicity, for every m we have e(am) = lim q→∞ m) ℓ(R/aq qn/n! , hence using Lemma 3.6 we conclude that e(am) ≤ ℓ(R/am) limit, we obtain n! . Dividing by mn and passing to L := lim m→∞ e(am) mn ≤ vol(a•). In order to prove the reverse inequality, note that given any ǫ > 0, by definition of L and of the Hilbert -- Samuel multiplicity, we can find first m ≥ 1 and then q ≥ 1 such that L > ℓ(R/aq m) mnqn/n! −ǫ. Since aq m ⊆ amq, it follows that L > ℓ(R/amq) (mq)n/n! − ǫ ≥ inf p ℓ(R/ap) pn/n! − ǫ. Since this holds for every ǫ > 0, using (17) we conclude that L ≥ vol(a•), completing the proof of the proposition. (cid:3) THE VOLUME OF A SET OF ARCS ON A VARIETY 11 Remark 3.8. Suppose that X = Spec(R) is smooth and a is an ideal in R which is cosupported at a point. Applying Proposition 3.5 in the case of the sequence given by the powers of a, we see that e(a) = inf m≥1 ℓ(R/am) mn/n! . In this note we will be interested in graded sequences that arise from pseudo-valuations. Definition 3.9. A function v : R → R≥0 ∪ {∞} is said to be a pseudo-valuation of R if it satisfies the following conditions: (i) (ii) (iii) v(0) = ∞ and v(λ) = 0 for every λ ∈ k, v(f + g) ≥ min{v(f ), v(g)} for every f, g ∈ R, and v(f g) ≥ v(f ) + v(g) for every f, g ∈ R. We say that a pseudo-valuation v is radical if, in addition, it satisfies (iv) v(f r) = r · v(f ) for every f ∈ R, r ∈ Z>0. The support of a pseudo-valuation v is the closed subscheme Supp(v) ֒→ X defined by the ideal Given a pseudo-valuation v and an ideal a in R, we put b∞(v) := {f ∈ R v(f ) = ∞}. v(a) := inf{v(f ) f ∈ a}. We say that v has center at the closed subscheme Y defined by the ideal b in R if we have b = {f ∈ R v(f ) > 0}. Remark 3.10. Note that if b defines the center of v, then v(b) > 0. Indeed, if we put Im := {f ∈ R v(f ) ≥ 1/m}, then each Im is an ideal in R and we have Im ⊆ Im+1. Since b =Sm Im and R is Noetherian, it follows that b = Im for m ≫ 0. Remark 3.11. There are two other related notions. A semi-valuation of R is a pseudo- valuation with the property that the inequality in (iii) is an equality for all f and g (in this case, condition (iv) is automatically satisfied). A semi-valuation v is a valuation if, in addition, we have v(f ) < ∞ for all f ∈ R r {0}. It is clear that in this case we can extend v to a valuation of the function field of R by putting v(f /g) = v(f ) − v(g) for every nonzero f, g ∈ R. Note that if v is a semi-valuation, then the ideal b∞(v) is a prime ideal and we have a valuation v on R/b∞ such that v = v ◦ π, where π : R → R/b∞ is the canonical projection. Remark 3.12. If (vα)α∈Λ is a family of semi-valuations of R and we put v(f ) := inf α∈Λ vα(f ), then v is a radical pseudo-valuation. Note that the support of v is the union of the supports of the vα and if Λ is finite, then the center of v is the union of the centers of the vα. In particular, these sets are not necessarily irreducible. It is a theorem of Bergman that every radical pseudo-valuation arises in this way. More precisely, for every radical pseudo-valuation w of R, there is a family (wi)i∈I of semi-valuations of R such that w(f ) = infi wi(f ) for every f ∈ R (see [Ber71, Theorem 2]). Remark 3.13. There is a canonical way to obtain a radical pseudo-valuation of R from an arbitrary pseudo-valuation. Indeed, if v is any pseudo-valuation, then we put v(f m) m = lim m→∞ v(f m) m , ev(f ) := inf m≥1 12 TOMMASO DE FERNEX AND MIRCEA MUSTAT¸ A where the second equality follows from property (iii) and a version of Lemma 3.7 (see [Mus02a, every f ∈ R. Moreover, if w is another radical pseudo-valuation such that w(f ) ≤ v(f ) for Lemma 1.4]). It is easy to see that ev is a radical pseudo-valuation such that ev(f ) ≤ v(f ) for every f ∈ R, then w(f ) ≤ ew(f ) for every f ∈ R. Suppose that v is a pseudo-valuation of R. We define for every m ∈ Z≥0 bm(v) := {f ∈ R v(f ) ≥ m}. It follows from (ii) and (iii) that b•(v) = (bm(v))m≥0 is a graded sequence of ideals. Remark 3.14. The sequence b•(v) clearly satisfies the condition bp(v) ⊆ bq(v) for p ≥ q. Example 3.15. Suppose that I 6= R is an ideal of R. If for every f ∈ R, we put vI (f ) := min{m ≥ 0 f ∈ I m}, then vI is a pseudo-valuation of R, with support X and whose center is defined by I. It follows from definition that in this case bm(vI ) = I m. Remark 3.16. It is clear that for every pseudo-valuation v and every m ≥ 1, if b is the ideal defining the center of v, then bm(v) ⊆ b and the two ideals have the same radical. In fact, if d is an integer such that d · v(b) ≥ 1 (see Remark 3.10), then bdm ⊆ bm(v) for every m ≥ 1. We will be mostly interested in pseudo-valuations with 0-dimensional center. Definition 3.17. The volume of a pseudo-valuation v of R is defined to be the volume vol(v) := vol(b•(v)) of the graded sequence b•(v). Recall that by (13) and (14), we have vol(v) = lim m→∞ ℓ(R/bm(v)) mn/n! = lim m→∞ e(bm(v)) mn . Remark 3.18. We have vol(v) < ∞ if and only if the center of v is a finite set. Indeed, if the latter condition holds, then the finiteness of the volume follows from Remark 3.3. On the other hand, if the center of v has positive dimension, then ℓ(R/bm(v)) = ∞ for all m ≥ 1 by Remark 3.16. Example 3.19. If I 6= R is an ideal whose cosupport consists of one point and vI is the pseudo-valuation associated to I in Remark 3.15, then vol(vI ) = e(I). Example 3.20. Let I 6= R be an ideal in R. Recall that there are finitely many divisorial valuations w1, . . . , wr of R (the Rees valuations of I) with the property that for every m ≥ 0, the integral closure I m of I m is equal to {f ∈ R wi(f ) ≥ m · wi(I) for 1 ≤ i ≤ r}. In particular, we see that if We refer to [Swa11] for an introduction to Rees valuations. w is the pseudo-valuation given by w = mini In particular, it follows from Example 3.4 that if the cosupport of I consists of one point, then vol(w) = e(I). wi(I) wi, then bm(w) = I m for every m. 1 Example 3.21. Suppose that v and w are pseudo-valuations of R such that v(f ) ≥ w(f ) for all f ∈ R. In this case we have bm(w) ⊆ bm(v) for all m. By taking the colength, dividing by mn/n!, and passing to limit, we obtain vol(w) ≥ vol(v). THE VOLUME OF A SET OF ARCS ON A VARIETY 13 Example 3.22. If v is a pseudo-valuation of R and α is a positive real number, then αv is a pseudo-valuation such that vol(αv) = 1 αn · vol(v). Indeed, note that we have bm(αv) ⊇ b⌈m/α⌉(v), hence vol(αv) ≤ lim m→∞ ℓ(R/b⌈m/α⌉(v)) ⌈m/α⌉n/n! · ⌈m/α⌉n mn = vol(v) · 1 αn . By writing v = 1 vol(αv). By combining the two inequalities, we obtain vol(αv) = 1 α (αv) and applying the inequality already proved, we obtain vol(v) ≤ αn · αn · vol(v). The following proposition gives an important example of valuation with positive volume. Proposition 3.23. If v is a divisorial valuation of R having center at a closed point x ∈ X and X is analytically unramified1 at x, then val(v) > 0. Proof. This is an immediate consequence of Izumi's theorem (see for example [HS01, The- orem 1.2]). This says that since the local ring OX,x is analytically unramified, there is a constant c = c(v) such that for every other divisorial valuation v′ with center {x}, we have v(f ) ≤ c · v′(f ) for every f ∈ R. Let w1, . . . , wr be the Rees valuations corresponding to the wi(mx) wi and α = c · maxi wi(mx), then we see that maximal ideal mx defining x. If w = mini v(f ) ≤ α · w(f ) for every f ∈ R. By combining Examples 3.20, 3.21, and 3.22, we conclude that 1 vol(v) ≥ vol(α · w) = vol(w) αn = e(mx) αn > 0. (cid:3) 4. The volume of a subset in the space of arcs Suppose, as in the previous section, that X = Spec(R) is an n-dimensional, affine algebraic variety over an algebraically closed field k. We now assume that char(k) = 0. Let X∞ be the scheme of arcs of X (for an introduction to spaces of arcs, see for example [EM09]). Since X is affine, X∞ is affine as well, but in general not of finite type over k. Note that if γ ∈ X∞ is a point with residue field k(γ), then we can identify γ with a morphism Spec(k(γ)[[t]]) → X. We denote by π : X∞ → X the canonical projection taking γ to γ(0), the image by γ of the closed point. Remark 4.1. While X∞ is not a Noetherian scheme, if C is a closed subset of X∞, we may still consider the irreducible components of C: these correspond to the prime ideals in O(X∞) which are minimal over the ideal of C. Note that we can still write C as the union of its irreducible components: this is an immediate application of Zorn's Lemma. For every γ ∈ X∞, we define the function ordγ : R → Z≥0 ∪ {∞} given by ordγ(f ) = ordt(γ∗(f )). It is clear that ordγ is a semi-valuation of R. Given a subset C ⊆ X∞, we consider the function ordC : R → Z≥0 ∪ {∞} defined by ordC(f ) = min γ∈C ordγ(f ). It follows from the definition that ordC is a radical pseudo-valuation. For short, we denote bm(C) := bm(ordC) ⊆ R and, similarly, let b•(C) := b•(ordC ). 1This means that the completion [OX,x is a domain (note that it is always reduced, since OX,x is a reduced excellent ring). The condition is satisfied, for example, if X is normal. 14 TOMMASO DE FERNEX AND MIRCEA MUSTAT¸ A Lemma 4.2. If C is the closure of a subset C ⊆ X∞, then ordC = ordC . Proof. The assertion follows from the fact that for every f ∈ R and every m ∈ Z, the set {γ ∈ X∞ ordγ(f ) ≥ m} is closed. (cid:3) The assertion in the next lemma follows directly from definition. Lemma 4.3. If C =Si∈I Ci, then ordC(f ) = mini∈I ordCi(f ) for every f ∈ R. Remark 4.4. If C is irreducible, then ordC is a semi-valuation. Indeed, it follows from Lemma 4.2 that if δ is the generic point of C, then ordC = ordδ, hence ordC is a semi- valuation. Remark 4.5. The center of the pseudo-valuation ordC is equal to π(C), with the reduced scheme structure. Indeed, this follows from the fact that for f ∈ R and γ ∈ X∞, we have ordγ(f ) ≥ 1 if and only if f lies in the ideal defining π(γ). Definition 4.6. We define the volume vol(C) of a set C ⊆ X∞ to be the volume of the pseudo-valuation ordC. vol(C) := vol(ordC) = vol(b•(C)) Proposition 4.7. For every C ⊆ X∞, we have vol(C) < ∞ if and only if π(C) is a finite set of closed points. Proof. The assertion follows by combining Remarks 3.18 and 4.5. (cid:3) From now on, we restrict our attention to subsets C ⊆ X∞ whose image in X is a finite set of closed points. In the next propositions, we give some basic properties of volumes of subsets of X∞. Proposition 4.8. If C1 ⊆ C2, then vol(C1) ≤ vol(C2). Proof. If C1 ⊆ C2 then it is clear that ordC1(f ) ≥ ordC2(f ) for every f ∈ R. The assertion then follows from Example 3.21. (cid:3) The next proposition allows us to reduce to considering subsets lying in a fiber of π : X∞ → X. For every closed point x ∈ X, we denote the fiber π−1(x) by X∞(x). Proposition 4.9. Let C ⊆ X∞ be such that π(C) is a finite set of closed points. If we consider the unique decomposition C = C1 ∪ . . . ∪ Cr such that the π(Ci) are pairwise distinct points, then we have vol(C) = Proof. If π(Ci) = {xi}, then it is clear that rXi=1 vol(Ci). bm(C) = r\i=1 bm(Cj) and bm(Cj) is cosupported at xj for every m ≥ 1. Therefore the assertion follows from Remark 3.3. (cid:3) Proposition 4.10. If C ⊆ X∞(x), for some closed point x ∈ X, then vol(X) ≤ ex(X). THE VOLUME OF A SET OF ARCS ON A VARIETY 15 Proof. Note that if mx is the ideal defining x, then mx ⊆ b1(C). Therefore mp bp(C) for every p, and we obtain vol(C) ≤ e(mx) = ex(X). x ⊆ b1(C)p ⊆ (cid:3) The following definition extends the notions of thin and fat arcs introduced in [ELM04, Ish05] to arbitrary sets of arcs. Definition 4.11. A subset C of X∞ is said to be thin if there exists a proper closed subscheme Z ֒→ X such that C ⊆ Z∞. If C is not thin, then we say that C is fat. A subset C of X∞ is a cylinder if C = π−1 m (S) for some m and some constructible subset S ⊆ Xm, where πm : X∞ → Xm is the canonical projection. It is a basic fact that a cylinder C is thin if and only if C ⊆ (Xsing)∞, where Xsing is the singular locus of X (see [EM09, Lemma 5.1]). Proposition 4.12. Let C be a subset of X∞ whose image in X is a finite set of closed points. If C is thin, then vol(C) = 0, and if the closure of C is a fat cylinder and X is analytically unramified at every point, then vol(C) > 0. Proof. Suppose first that there exists a proper closed subscheme Z of X such that C ⊆ Z∞. Let IZ ⊆ OX be the ideal of Z. We have IZ ⊆ bm(C) for every m, hence ℓ(OX /bm(C)) = ℓ(OZ /bm(C)OZ ) = o(mn) since dim Z < n. This implies that vol(C) = 0. Let us assume now that C is a fat cylinder. Since ordC = ordC by Lemma 4.2, we may replace C by C and thus assume that C is closed. Since C is a cylinder, it has finitely many irreducible components (see [dFEI08, Proposition 3.5]). One of these, say C ′, has to be fat, in which case ordC ′ is a divisorial valuation by [dFEI08, Propositions 2.12 and 3.9]. Of course, the image of C ′ in X consists of one closed point. Using Propositions 4.8 and 3.23, we conclude that vol(C) ≥ vol(C ′) = vol(ordC ′) > 0. (cid:3) We now address the results stated in the introduction. We begin with the first two propo- sitions. Proof of Proposition 1.1. For every p, we have (19) (20) The exact sequence bp(C1 ∪ C2) = bp(C1) ∩ bp(C2) and bp(C1 ∩ C2) ⊇ bp(C1) + bp(C2). 0 → OX/(bp(C1) ∩ bp(C2)) → OX/bp(C1) ⊕ OX/bp(C2) → OX/(bp(C1) + bp(C2)) → 0 implies ℓ(OX /bp(C1)) + ℓ(OX /bp(C2)) = ℓ(OX /bp(C1) ∩ bp(C2)) + ℓ(OX /bp(C1) + bp(C2)). Using (19) and (20), we conclude ℓ(OX /bp(C1)) + ℓ(OX /bp(C2)) ≥ ℓ(OX /bp(C1 ∪ C2)) + ℓ(OX /bp(C1 ∩ C2)). Then the assertion follows by dividing by pn/n! and letting p go to infinity. Note that this step uses the property that the limsup in the definition of the volume is, in fact, a limit. (cid:3) 16 TOMMASO DE FERNEX AND MIRCEA MUSTAT¸ A Proof of Proposition 1.2. Let Cm = Cont≥m(a). It follows from definition that ap ⊆ bmp(Cm) for every p ≥ 1. By (13), we have mn · vol(Cm) = lim p→∞ e(bmp(Cm)) pn ≤ lim p→∞ e(ap) pn = e(a). Using the characterization of volume in Remark 3.2, we deduce from the inclusion a ⊆ bm(Cm) that Note that if γ(t) ∈ Cm, then γ(tp) ∈ Cmp. This implies that we have an inclusion vol(Cm) ≤ e(bm(Cm)) mn ≤ e(a) mn . and therefore bmpq(Cmp) ⊆ bmq(Cm) for every q, mn · e(bmq(Cm)) (mq)n ≤ (mp)n · e(bmpq(Cmp)) (mpq)n . By letting q go to infinity, we obtain mn · vol(Cm) ≤ (mp)n · vol(Cmp). In order to complete the proof, it is enough to show that when m is divisible enough, we have vol(Cm) ≥ e(a) mn . Suppose that E1, . . . , Er are the divisors over X corresponding to the Rees valuations associated to the ideal a (see Example 3.20). We put qi = ordEi(a) and assume that m is divisible by every qi. Recall that if E is a divisor over X, then there is a sequence of irreducible closed subsets C q X(E), for q ≥ 1, called the maximal divisorial sets, which are defined as follows. If π : Y → X is a birational map such that Y is smooth and E is a smooth divisor on Y , then C q X(E) is the closure of π∞(Cont≥q(E)). It is easy to see X (E) = q · ordE. For a discussion of these subsets of X∞, we refer to [ELM04] and that ordC q [dFEI08]. With this notation, we consider the closed subset Tm := r[i=1 C m/qi X (Ei). Note that we have Tm ⊆ Cm, hence bjm(Cm) ⊆ bjm(Tm) = r\i=1 {f ∈ R ordEi(f ) ≥ jqi} = aj, where we denote by c the integral closure of an ideal c. We conclude that e(bjm(Cm)) ≥ e(aj) = jn · e(a). Dividing by (jm)n and letting j go to infinity, we get vol(Cm) ≥ e(a) proof of the proposition. mn . This completes the (cid:3) Next, we review the definition of jet-codimension and prove two more preliminary prop- erties before addressing the proof of Theorem 1.3. Recall that the Krull codimension of a closed irreducible set C ⊆ X∞ is the dimension of the local ring OX∞,C, and is denoted by codim(C). The definition extends to an arbitrary set C ⊆ X∞ by taking the smallest codimenion of an irreducible component of the closure C. While the Krull codimension is computed from the local rings of X∞, the jet-codimension is computed from the finite levels Xm. In order to define it, we need the following lemma. THE VOLUME OF A SET OF ARCS ON A VARIETY 17 Lemma 4.13. For every subset C ⊆ X∞, the limit m→∞(cid:16)(m + 1)n − dim πm(C)(cid:17) lim exists. Proof. It follows from [DL99, Lemma 4.3] that for every m, the fibers of the map πm+1(X∞) → πm(X∞) have dimension ≤ n (note that both sets are constructible by a result due to Green- berg [Gre66]). It follows from Lemma 4.14 below that dim πm+1(C) ≤ dim πm(C) + n, hence the sequence (am)m≥1 with am = (m + 1)n − dim πm(C) is a non-decreasing sequence of integers. Therefore it either stabilizes or it has limit infinity. (cid:3) Lemma 4.14. Let f : V → W be a morphism of algebraic varieties over k and suppose that d is a non-negative integer and A is a constructible subset of V such that for every y ∈ f (A), we have dim(f −1(y) ∩ A) ≤ d. For every subset B ⊆ A, we have dim(B) ≤ d + dim(f (B)). Proof. We can write A =Sr i=1 Bi, B = Sr then B = Sr i=1 Ai, with each Ai a locally closed subset of V . If Bi = B ∩ Ai, i=1 f (Bi). Since it is enough to prove the assertion for each Bi, it follows that we may assume that A is a locally closed subset. In this case A is open in A, hence A ∩ B is a dense open subset of B. Since dim(B) = dim(A ∩ B) and the fibers of the morphism A ∩ B → f (B) have dimension ≤ d, we obtain the assertion in the lemma. (cid:3) i=1 Bi, and f (B) = Sr Definition 4.15. The jet-codimension of an irreducible closed subset C of X∞ is defined to be jet-codim(C) := lim m→∞(cid:16)(m + 1)n − dim πm(C)(cid:17) . For an arbitrary subset C ⊆ X∞, we define jet-codim(C) to be the smallest jet-codimension of an irreducible component of C. Remark 4.16. It follows from the proof of Lemma 4.13 that if C is closed and irreducible, then jet-codim(C) ≥ n − dim π(C) ≥ 0. This implies that for every C ⊆ X, we have jet-codim(C) ≥ 0. Remark 4.17. If C1 ⊆ C2 ⊆ X∞, then jet-codim(C1) ≥ jet-codim(C2). an irreducible component of C1, then there is an irreducible component C ′ C ′ 1 ⊆ C ′ 2. In this case, for every m we have (m + 1)n − dim πm(C ′ 1) ≥ (m + 1)n − dim πm(C ′ 2). Indeed, if C ′ 1 is 2 of C2 such that By letting m go to infinity, we conclude that jet-codim(C ′ 2) ≥ codim(C2). Since this holds for every irreducible component of C1, we conclude that jet-codim(C1) ≥ jet-codim(C2). 1) ≥ jet-codim(C ′ Remark 4.18. For any subset C ⊆ X∞, we have codim(C) = codim(C) and jet-codim(C) = jet-codim(C). If X is smooth and C ⊆ X∞ is a cylinder, then jet-codim(C) = codim(πm(C), Xm) for all m ≫ 1. As the next proposition shows, this is equal to the Krull codimension codim(C). More generally, we have the following property. Proposition 4.19. If X is smooth and C ⊆ X∞ is any set, then jet-codim(C) = codim(C). 18 TOMMASO DE FERNEX AND MIRCEA MUSTAT¸ A Proof. The proof of the proposition follows immediately by applying the next lemma to the irreducible components of C. (cid:3) Lemma 4.20. If X is smooth and C ⊆ X∞ is a closed irreducible subset, then and this number is finite if and only if C is a cylinder. jet-codim(C) = codim(C), Proof. If C is a cylinder, then it follows from [ELM04, Corollary 1.9] that jet-codim(C) = codim(πm(C), Xm) = codim(C) for m ≫ 1. Therefore it suffices to show that if C is not a cylinder then jet-codim(C) = dim(C) = ∞. In order to check this, consider the sequence of closed irreducible cylinders Fi := π−1 i (πi(C)), i ≥ 0. We have inclusions C ⊆ · · · ⊆ Fi+1 ⊆ Fi ⊆ · · · ⊆ F1 ⊆ F0 ⊆ X∞. Moreover, since C is closed, we have C =Ti≥0 Fi. Since C is not a cylinder, the sequence (Fi)i≥0 does not stabilize. Therefore we can pick a subsequence (Fim)m≥0 such that C ( Fim ( Fim−1 ( · · · ( Fi1 ( Fi0 ( X∞, which clearly implies that codim(C) = ∞. In fact, for every m, if p ≥ im, then we also have the sequence πp(C) ⊆ πp(Fim) ( πp(Fim−1 ) ( · · · ( πp(Fi1) ( πp(Fi0 ) ( Xp. Note that for every k ≤ m, the subset πp(Fik ) of Xp is irreducible and closed since p ≥ ik. Therefore codim(πp(C), Xp) ≥ m and we conclude that jet-codim(C) = ∞. (cid:3) Remark 4.21. The definition of jet-codimension generalizes to all sets the definition of codi- mension of a quasi-cylinder given in [dFEI08]. In general, if X is singular and C ⊆ X∞ is a closed irreducible set, then there is only an inequality codim(C) ≤ jet-codim(C) which can be strict (e.g., see [IR13, Example 2.8]). If E is a prime exceptional divisor over X and C q X (E) ⊆ X∞ is the maximal divisorial set associated to the divisorial valuation q · ordE, then we have (21) jet-codim(C q X(E)) = q ·baE(X) by [dFEI08, Theorem 3.8]. Using this fact, it is easy to extend [Mus02b, Corollary 0.2] to the singular setting, as follows. This proposition is also proved in [Ish13, Proposition 3.5], but since the proof is short, we include it for the convenience of the reader. Proposition 4.22. For every proper, nonzero ideal a ⊆ R and every positive integer m, we have with equality if m is sufficiently divisible. jet-codim(Cont≥m(a)) ≥ m ·clct(a), THE VOLUME OF A SET OF ARCS ON A VARIETY 19 Proof. By [dFEI08, Propositions 3.5 and 2.12], Cont≥m(a) has finitely many fat irreducible components, and any such component C is a maximal divisorial set. In particular, there is a fat irreducible component of the form C = C q X(E) for some divisorial valuation q · ordE, such that jet-codim(Cont≥m(a)) = jet-codim(C q by (21). Note that q · ordE(a) ≥ m, since C q X(E) ⊆ Cont≥m(a). On the other hand, we have X (E)) = q ·baE(X), clct(a) ≤ baE(X) ordE(a) by the definition of Mather log canonical threshold. We conclude that jet-codim(Cont≥m(a)) ≥ m ·clct(a). On the other hand, suppose that F is a divisor over X such that clct(a) = baF (X) suppose that m = q · ordF (a) for some positive integer q. In this case C q hence ordF (a) and X (F ) ⊆ Cont≥m(a), jet-codim(Cont≥m(a)) ≤ jet-codim(C q By combining this with what we have already proved, we conclude that in this case we have (cid:3) X (F )) = q ·baF (X) = m ·clct(a). jet-codim(Cont≥m(a)) = m ·clct(a). Proof of Theorem 1.3. For every p ≥ 1, we have C ⊆ Cont≥p(bp(C)). Note that if C lies over the closed point x ∈ X, defined by the maximal ideal mx, the ideal bp(C) is mx-primary. It follows from Proposition 4.22 that (22) On the other hand, Theorem 1.4 implies that jet-codim(C) ≥ jet-codim Cont≥p(bp(C))) ≥ p ·clct(bp(C)). (23) By combining (22) and (23), we get (n! · ℓ(OX /bp(C)))1/n ·clct(bp(C)) ≥ n. (cid:18) ℓ(OX /bp(C)) · jet-codim(C) ≥ n. (cid:19)1/n pn/n! We conclude that vol(C)1/n · jet-codim(C) = lim p→∞(cid:18) ℓ(OX /bp(C)) pn/n! (cid:19)1/n · jet-codim(C) ≥ n. This gives the first part of the statement of the theorem. The second part follows from Proposition 4.19. (cid:3) References [Ber71] George M. Bergman, A weak Nullstellensatz for valuations, Proc. Amer. Math. Soc. 28 (1971), 32 -- 38. ↑11 [BH93] Winfried Bruns and Jurgen Herzog, Cohen-Macaulay rings, Cambridge Studies in Advanced Math- ematics, vol. 39, Cambridge University Press, Cambridge, 1993. ↑6 [Cut14] Steven Dale Cutkosky, Asymptotic multiplicities of graded families of ideals and linear series, Adv. Math. 264 (2014), 55 -- 113. ↑1, 8 [dFEI08] Tommaso de Fernex, Lawrence Ein, and Shihoko Ishii, Divisorial valuations via arcs, Publ. Res. Inst. Math. Sci. 44 (2008), no. 2, 425 -- 448. ↑15, 16, 18, 19 [dFEM03] Tommaso de Fernex, Lawrence Ein, and Mircea Mustat¸a, Bounds for log canonical thresholds with applications to birational rigidity, Math. Res. Lett. 10 (2003), no. 2-3, 219 -- 236. ↑2, 3, 4, 5, 6 [dFEM04] , Multiplicities and log canonical threshold, J. Algebraic Geom. 13 (2004), no. 3, 603 -- 615. ↑2, 3, 6 20 TOMMASO DE FERNEX AND MIRCEA MUSTAT¸ A [DL99] Jan Denef and Fran¸cois Loeser, Germs of arcs on singular algebraic varieties and motivic integra- tion, Invent. Math. 135 (1999), no. 1, 201 -- 232. ↑17 [ELM04] Lawrence Ein, Robert Lazarsfeld, and Mircea Mustat¸a, Contact loci in arc spaces, Compos. Math. 140 (2004), no. 5, 1229 -- 1244. ↑15, 16, 18 [ELS03] Lawrence Ein, Robert Lazarsfeld, and Karen E. Smith, Uniform approximation of Abhyankar val- uation ideals in smooth function fields, Amer. J. Math. 125 (2003), no. 2, 409 -- 440. ↑7 [EM09] Lawrence Ein and Mircea Mustat¸a, Jet schemes and singularities, Algebraic geometry -- Seattle 2005. Part 2, 2009, pp. 505 -- 546. ↑1, 13, 15 [Eis95] David Eisenbud, Commutative algebra. With a view toward algebraic geometry, Graduate Texts in Mathematics, vol. 150, Springer-Verlag, New York, 1995. ↑9 [Ful98] William Fulton, Intersection theory, 2nd ed., Ergebnisse der Mathematik und ihrer Grenzgebiete. 3. Folge, Springer-Verlag, Berlin, 1998. ↑5, 6 [Gre66] Marvin J. Greenberg, Rational points in Henselian discrete valuation rings, Inst. Hautes ´Etudes Sci. Publ. Math. 31 (1966), 59 -- 64. ↑17 [HS01] Reinhold Hubl and Irena Swanson, Discrete valuations centered on local domains, J. Pure Appl. Algebra 161 (2001), no. 1-2, 145 -- 166. ↑13 [Ish05] Shihoko Ishii, Arcs, valuations and the Nash map, J. Reine Angew. Math. 588 (2005), 71 -- 92. ↑15 [Ish13] , Mather discrepancy and the arc spaces, Ann. Inst. Fourier (Grenoble) 63 (2013), no. 1, 89 -- 111. ↑2, 3, 18 [IR13] Shihoko Ishii and Ana J. Reguera, Singularities with the highest Mather minimal log discrepancy, Math. Z. 275 (2013), no. 3-4, 1255 -- 1274. ↑18 [KM98] J´anos Koll´ar and Shigefumi Mori, Birational geometry of algebraic varieties, Cambridge Tracts in Mathematics, vol. 134, Cambridge University Press, Cambridge, 1998. With the collaboration of C. H. Clemens and A. Corti; Translated from the 1998 Japanese original. ↑4, 6 [KN14] J´anos Koll´ar and Andr´as N´emethi, Durfee's conjecture on the signature of smoothings of surface singularities, with an appendix by Tommaso de Fernex (2014). arXiv:1411.1039. ↑9 [LM09] Robert Lazarsfeld and Mircea Mustat¸a, Convex bodies associated to linear series, Ann. Sci. ´Ec. Norm. Sup´er. (4) 42 (2009), no. 5, 783 -- 835. ↑8 [Mus02a] Mircea Mustat¸a, On multiplicities of graded sequences of ideals, J. Algebra 256 (2002), no. 1, 229 -- 249. ↑7, 10, 12 [Mus02b] , Singularities of pairs via jet schemes, J. Amer. Math. Soc. 15 (2002), no. 3, 599 -- 615 (electronic). ↑18 [Swa11] Irena Swanson, Rees valuations, Commutative algebra -- Noetherian and non-Noetherian perspec- tives, 2011, pp. 421 -- 440. ↑12 [ZS60] Oscar Zariski and Pierre Samuel, Commutative algebra. Vol. II, The University Series in Higher Mathematics, D. Van Nostrand Co., Inc., Princeton, N.J. -- Toronto -- London -- New York, 1960. ↑4 Department of Mathematics, University of Utah, Salt Lake City, UT 48112, USA E-mail address: [email protected] Department of Mathematics, University of Michigan, Ann Arbor, MI 48109, USA E-mail address: [email protected]
1609.05005	1	1609	2016-09-16T11:29:49	Homology of the three flag Hilbert scheme	[ "math.AG" ]	We prove the existence of an affine paving for the three-step flag Hilbert scheme $$ \text{Hilb}^{n, n+1, n+2}(0) := \left\{\mathbb{C}[[x,y]]\supset I_n\supset I_{n+1}\supset I_{n+2}: I_i \,\,\text{ ideals with } \text{dim}_{\mathbb{C}} {\mathbb{C}[x,y]}/{I_i} = i \right\} $$ of 0-dimensional subschemes that are supported at the origin of $\mathbb{C}^2$. This is done by showing that the space stratifies in smooth subvarieties, the Hilbert-Samuel's strata, each of which has an affine paving with cells of known dimension, indexed by marked Young diagrams. The affine pavings of the Hilbert-Samuel's strata allow us to prove that the Poincar\'{e} polynomials for $\text{Hilb}^{n,n+1, n+2}(0)$ satisfy:$$ \sum_{n\geq 0} P_q\left(\text{Hilb}^{n,n+1, n+2}(0)\right) z^n = \frac{q+1}{(1-zq)(1-z^2q^2)}\,\, \prod_{k\geq 1} \frac{1}{1-z^kq^{k-1}}. $$ In the process of proving this formula we relate combinatorially the homology of our spaces with that of known subspaces of $\text{Hilb}^{n+1, n+3}(0)$. As a corollary we find an affine paving and a formula for the generating function of the Poincar\'{e} polynomials of $\text{Hilb}^{n, n+2}(0)$ for all $n\in \mathbb{N}$.	math.AG	math	Homology of the three flag Hilbert scheme Daniele Boccalini July 24, 2016 6 1 0 2 p e S 6 1 ] . G A h t a m [ 1 v 5 0 0 5 0 . 9 0 6 1 : v i X r a ii Contents Acknowledgements Abstract Acknowledgements Introduction 1 Fundamental Facts 1.1 Hilbert scheme of points . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1.2 Hilbert-Samuel's strata . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1.3 Torus action and Borel-Moore homology . . . . . . . . . . . . . . . . . . . v vii ix xi 1 1 7 11 2 Tangent spaces at torus fixed points 2.1 A weight basis for TI Hilbn(C2) . . . . . . . . . . . . . . . . . . . . . . . . . . 2.1.1 The tangent space to the Hilbert-Samuel's strata, case n = n . . . 2.2 A weight basis for TI1,I2Hilbn,n+1(C2) . . . . . . . . . . . . . . . . . . . . . . . 19 20 26 29 2.2.1 The tangent space to the Hilbert-Samuel's strata, case n = (n, n + 1) 35 39 2.3.1 The tangent to the Hilbert-Samuel's strata, case n = (n, n + 1, n + 2) 48 2.3 A weight basis for TI1,I2,I3Hilbn,n+1,n+2(C2) . . . . . . . . . . . . . . . . . . . 3 Smoothness of the Hilbert-Samuel's strata 3.1 Iarrobino's standard generators . . . . . . . . . . . . . . . . . . . . . . . . . 3.2 MT1,T2,T3 is smooth. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3.2.1 GT1,T2,T3 is smooth . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3.3 Strata for longer flags are singular . . . . . . . . . . . . . . . . . . . . . . . . 3.3.1 Four flag case . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3.3.2 Five, and longer, flag case . . . . . . . . . . . . . . . . . . . . . . . . 55 56 61 72 77 77 80 4 Generating function and Hilbn,n+2(C2). 85 86 4.1 A tale of two tori . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4.2 Hilbn,n+2(C2) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 106 iii iv Résumé On prouve l'existence d'un pavage affine pour le schéma de Hilbert Hilbn,n+1,n+2(0) :=nC[[x, y]] ⊃ In ⊃ In+1 ⊃ In+2 : Ii idéaux avec dimC C[x, y].Ii = io des drapeaux de longueur trois des sous schémas 0-dimensionels qui sont supportés à l'origine de C2. On atteint ce résultat en montrant que l'espace est stratifié par des sous variétés lisses, les strata de Hilbert-Samuel. On montre que chacun de ces strata a un pavage affine en cellules de dimension connue et indexées par des diagrammes de Young marqués. Le pavage affine nous permet de montrer que les polynômes de Poincaré de Hilbn,n+1,n+2(0) sont tels que: Xn≥0 Pq³Hilbn,n+1,n+2(0)´ zn = q + 1 (1 − zq)(1 − z2q 2) Yk≥1 1 1 − zk q k−1 . (1) Dans la preuve de (1) on construit une correspondance combinatoire entre l'homologie de nos espaces et l'homologie des certains sous espaces connus de Hilbn+1,n+3(0). On obtient comme corollaire un pavage affine et une formule pour la série génératrice des polynômes de Poincaré de Hilbn,n+2(0) pour tous les n ∈ N. Mots-clés. Schéma de Hilbert, homologie, drapeaux d'idéaux, strata de Hilbert- Samuel, pavage affine. v vi Abstract We prove the existence of an affine paving for the three-step flag Hilbert scheme Hilbn,n+1,n+2(0) :=nC[[x, y]] ⊃ In ⊃ In+1 ⊃ In+2 : Ii ideals with dimC C[x, y].Ii = io of 0-dimensional subschemes that are supported at the origin of C2. This is done by showing that the space stratifies in smooth subvarieties, the Hilbert-Samuel's strata, each of which has an affine paving with cells of known dimension, indexed by marked Young diagrams. The affine pavings of the Hilbert-Samuel's strata allow us to prove that the Poincaré polynomials for Hilbn,n+1,n+2(0) satisfy: Xn≥0 Pq³Hilbn,n+1,n+2(0)´ zn = q + 1 (1 − zq)(1 − z2q 2) Yk≥1 1 1 − zk q k−1 . (2) In the process of proving (2) we relate combinatorially the homology of our spaces with that of known subspaces of Hilbn+1,n+3(0). As a corollary we find an affine paving and a formula for the generating function of the Poincaré polynomials of Hilbn,n+2(0) for all n ∈ N. Keywords. Hilbert scheme, homology, flags of ideals, Hilbert-Samuel's strata, affine paving. vii Acknowledgements I would like to thank my advisor, Tamás Hausel, for his fundamental help and encour- agement during the preparation of this thesis, and more generally during these four long years of studying and exploring. I am very grateful to him also for the extraor- dinary group he created and guided here in Epfl. It was honestly a big honor and a lot fun to be part of it. Everyone contributed immensely in making this experience a really enjoyable and interesting adventure, and that is why I want to thank, and thank a lot, present and past members: Michael Groechenig, Michael Wong, Martin Mereb, Zongbin Chen, Mario Garcia Fernandez, Ben Davison, Szilárd Szabó, Alexander Noll, Michael McBreen, and Yohan Brunebarbe. A special thank to Riccardo Grandi and Dimitri Wyss with whom not only I shared the experience of the PhD and a big messy desk, but also a deep and genuine friendship that I hope will last for many years to come. I also owe many thanks to the many visitors that I met during these years and from whom I learned so much. In particular I would like to thank Luca Migliorini for the truly invaluable teachings and support. Finally I would like to thank Pierrette Paulou that has always been the soul of our group: her help has been so fundamental and, more importantly, so kind and generous that it is impossible to overestimate it. I would like to thank Kathryn Hess, Donna Testerman, András Szenes and Balázs Szendroi for accepting to be part of the jury for my defense and devoting time to review my work. The best part of all this experience has the name of Alexandre, the name of Marie and the name of Martina. Thanks a lot. Finally I would like to thank my family and my friends: after a lot of research they are still what I like the most. During these years, my research was supported by the Project Funding num- ber 144119: Arithmetic harmonic analysis on Higgs, character and quiver varieties is- sued by the Swiss National Science Foundation; Foundation that I thank a lot for this. ix x Introduction The Hilbert scheme of points of a smooth surface is one of the most beautiful and most studied examples of a moduli space. It is an algebraic geometric object that relates to many branches of mathematics: symplectic geometry, representation the- ory, combinatorics and, recently, theoretical physics. This special position is ensured on one hand by the relative simplicity and naturalness of its definition, and, on the other hand, by the many interesting structures it is equipped with. Curiously these structures are of two natures: some are directly inherited from the base surface, oth- ers appear from the moduli problem. Given a smooth surface X , the Hilbert scheme of n points of X , denoted as Hilbn(X ), parametrizes 0-dimensional subschemes of X of length n. The most generic example is a collection of distinct points of X : in this case the length is the number of points. However it is when the points start colliding together that the spectrum of possible scheme structures becomes more and more complicated and its geome- try more and more interesting. For example when two points are infinitely close the Hilbert scheme remembers the direction along which they came together, i.e. a tan- gent vector at the collision point. Subschemes that are entirely supported on a single point form a subvariety sometimes called the punctual Hilbert scheme and denoted by Hilbn(0). This space is the same for every surface and every point. It is singular, not even normal, but projective, reduced and irreducible (Haiman [Hai98] and Brian- con [Bri77]). It precisely measures the difference between Hilbn(X ) and X (n) the n-th symmetric power of X that parametrizes n-tuples of points up to order. In this sense the punctual Hilbert scheme is of key importance for those structures of Hilbn(X ) that are inherited from X : smoothness for example (Fogarty [Fog68]), or an holomorphic symplectic form if X has one (Beauville [B+83]). In fact Hilbn(0) is the most singular fiber of the natural forgetful map Hilbn(X ) → X (n) that turns out to be a crepant res- olution of singularities. The study of the geometry of Hilbn(0) is also an important step in the work of Haiman [Hai01] to prove the combinatorial conjecture of n!. In an- other sense, Hilbn(0) is interesting as it contains a lot of the topological information of Hilbn(X ). One can prove that Hilbn(0) is a deformation retract of Hilbn(C2), and, since X is covered by open subvarieties diffeomorphic to C2, Hilbn(X ) is covered by open subvarieties diffeomorphic to Hilbn(C2). xi xii INTRODUCTION In 1987 Ellingsrud and Stromme [ES87] coronated the efforts of many pre- senting a neat description of the Borel-Moore homology of Hilbn(C2) by exploiting a natural torus action on it. The exact form for the Poincaré polynomial of Hilbn(C2) became more relevant when Goettsche [Göt90] considered all of the Hilbert schemes Hilbn(C2) for different n at once and proved the formula that bears his name +∞Xn=0Xi ≥0 dim Hi¡Hilbn(C2)¢ q i zn = +∞Yk=1 1 1 − q 2k−2zk . Bundling all the Hilbn(C2) together, not only produces prettier formulas, but it is the starting point for the study of those additional and somewhat mysterious structures that we mentioned above. Motivated by Goettsche formula (that holds more gener- ally [GS93]) Witten and Vafa [VW94] related the study of Hilbert schemes to string theory; Nakajima [Nak97] constructed a geometric representation of products of the Heisenberg and Clifford algebras on the homology of Fn Hilbn(C2); Lehn [Leh99] used a Vertex Algebra structure to study the product in cohomology. This just to cite some examples. An important geometric player in studying all the Hilbert schemes together is a space that has also an intrinsic interest: the flag Hilbert scheme. This parametrizes flags of 0-dimensional subschemes of specified lengths. Its global geometry deteri- orates quickly: Hilbn,n+1(C2) is the last one to be smooth, as Cheah [Che98] proves. Again, if we ask for all subschemes to be concentrated in only one point we get quite interesting varieties. For longer flags we get varieties with many irreducible compo- nents of different dimensions. The case we are most interested in in this thesis is Hilbn,n+1,n+2(0). The main goal is to prove the following result. Theorem 0.0.1. For every n ∈ N the space Hilbn,n+1,n+2(0) has a cellular decomposition with cells that are isomorphic to affine spaces. These affine cells are indexed by Young diagrams of size n +2 with two marked boxes. The dimension of each affine cell is read- able from its label and the homology classes of the closures of these cells give a graded basis for the homology of Hilbn,n+1,n+2(0). The Poincaré polynomials of Hilbn,n+1,n+2(0) for all n fit into a generating function: +∞Xn=0 Xi ≥0 dim Hi³Hilbn,n+1,n+2(0)´ q i zn = q + 1 (1 − zq 2)(1 − z2q 4) 1 1 − q 2k−2zk . +∞Yk=1 The techniques we utilize are by now classical in this area of studies. They were first used by Ellingsrud and Stromme [ES87, ES88] , perfected by Goettsche [Göt90, Göt94] and used by Cheah [Che98] to treat the case of Hilbn,n+1(0). To give some more details on the strategy of the proof we explain the structure of the different chapters. In Chapter I, after the general definitions, we quickly focus on flag Hilbert sche- mes of subvarieties concentrated at one point. Here we introduce a stratification due xiii to Iarrobino [Iar72, Iar77] that is key to understand the geometry of Hilbn(0) (and sim- ilar spaces). Every point of Hilbn(0) is an ideal I ⊂ C[[x, y]] and as such has a Hilbert- Samuel's type T (I ) ∈ Nn. The Hilbert-Samuel's strata MT are indexed by the possible Hilbert-Samuel's type T ∈ Nn and contain all ideals I such that T (I ) = T . It turns out that the Hilbert-Samuel's strata are smooth, as Iarrobino [Iar72] proves. In the last section we introduce the famous technique of Bialynicki-Birula [BB73] to prove that a smooth space with a torus action has, under some conditions, an affine cell de- composition with cells labeled by torus fixed points. A result of Fulton [Ful13] tells us that the closures of these cells give a graded basis for the Borel-Moore homology of the space. We show that Hilbn(C2) and all related varieties carry a natural two di- mensional torus action that comes from rescaling the coordinates of C2. We finish the chapter by studying the torus fixed points of Hilbn(0), Hilbn,n+1(0) and Hilbn,n+1,n+2(0) and by relating these fixed points with marked Young diagrams. In Chapter II we study the Zariski tangent spaces of Hilbn(C2), Hilbn,n+1(C2) and Hilbn,n+1,n+2(C2) at their fixed points. In particular, we find a basis of eigenvec- tors for the two dimensional torus action at each fixed point and we interpret these eigenvectors as combinatorial gadgets of the marked Young diagram that labels the fixed point. The bases for the tangent spaces of Hilbn,n+1,n+2(C2), that are our origi- nal contributions, are constructed extending the classical study of the similar bases for Hilbn(C2) and Hilbn,n+1(C2). The use of these bases is far reaching. We prove that Hilbn(C2) is smooth and thus we describe a cell decomposition of Hilbn,n+1(0) and its homology (Fogarty [Fog68], Ellingsrud-Stromme [ES87]). We prove that all Hilbert- Samuel's strata MT ⊂ Hilbn(C2) are smooth and describe their cell decomposition and homology (Iarrobino[Iar77, IY03], Goettsche [Göt94]). We prove that Hilbn,n+1(C2) is smooth and thus describe a cell decomposition of Hilbn,n+1(0) and its homology (Cheah [Che98]). We prove that all Hilbert-Samuel's strata MT1,T2 ⊂ Hilbn,n+1(0) are smooth and describe their cell decomposition and homology (Cheah [Che98]). All of this crucially relies on smoothness and is possible thanks to the result of Bialynicki- Birula. Recall that unfortunately Hilbn,n+1,n+2(C2) is not smooth. We end the section by giving an original description of the tangent spaces of the Hilbert-Samuel's strata MT1,T2,T3 of Hilbn,n+1,n+2(0). In Chapter III we prove that the Hilbert-Samuel's strata MT1,T2,T3 are smooth. To do so we use results of Iarrobino [Iar77] on special opens that cover MT1 and whose points have especially nice generators. We are then able to study the dimension of MT1,T2,T3 relating it with that of MT1. This, thanks to the knowledge acquired in Chap- ter 2 on the tangent spaces, is enough to prove smoothness. We can then apply Bialynicki-Birula decomposition to describe their cell decompositions and homolo- gies, and ultimately the cell decomposition and the homology of Hilbn,n+1,n+2(0). In the last section we show that for longer flag cases, i.e. starting at Hilbn,n+1,n+2,n+3(0), smoothness of the Hilbert-Samuel's strata no longer holds. This means that the case xiv INTRODUCTION of Hilbn,n+1,n+2(0) is the last case where the classical techniques we described yield in- teresting results. In Chapter IV we prove the formula for the generating function. The fact that Hilbn,n+1,n+2(C2) is not smooth implies that the cell decomposition we obtained in Chapter 3 might depend on the single appropriate choice of a one dimensional subtorus action. The resulting combinatorics of the Poincaré polynomials is not well suited to prove the formula for the generating function. In the case of Hilbn(0) and Hilbn,n+1(0), smoothness of the ambient space guarantees that we are free to use any one dimensional subtorus. In fact we obtain many different cell decompositions that give rise to different combinatorial expressions of the same Poincaré polynomials crucial to prove, combinatorially, that also in the case of Hilbn,n+1,n+2(0) we can rewrite Pq¡Hilbn(0)¢ and Pq³Hilbn,n+1(0)´. Understanding better the details of these cases is Pq³Hilbn,n+1,n+2(0)´ more conveniently. Once this is done we actually do not need to sum the results. In fact it turns out that Nakajima and Yoshioka [NY08] already consid- ered the same generating function by studying a different family of smooth subspaces Hilbn−1,n+1(C2)t r ⊂ Hilbn−1,n+1(C2). Thus we only need to match the combinatorics. It remains unclear if there is also a geometrical connection between their spaces and ours. However we manage to deduce a last original result: an affine cell decomposi- tion of the spaces Hilbn,n+2(0) and a generating function for their Poincaré polynomi- als. Chapter 1 Fundamental Facts In this chapter we introduce the geometrical spaces we are interested in and the techniques that will allow us to describe some of their geometrical properties. The starting point is the definition of the Hilbert scheme of 0-dimensional sub- schemes of length n on a smooth surface X . This variety parametrizes configurations of n points of X . We define the punctual Hilbert scheme Hilbn(0) that measures the local difference between the Hilbert scheme Hilbn(X ) and the n-th symmetric power of X that parametrizes lists of n points of X up to order. We then define the flag version of the Hilbert scheme, that parametrizes flags of subschemes of specified length. Again we will be interested in the case where the support of all the subschemes is a single point of X and thus we define the punctual flag Hilbert scheme. Following Iarrobino [Iar77], we will stratify these spaces accord- ing to the Hilbert-Samuel's type of the ideals that compose the flags. To study the topological properties of the spaces introduced we use the natural torus action induced by the rescaling action on the local coordinates of the plane C2. An action with isolated fixed points on a smooth variety Y has attracting sets that are affine cells thanks to the theorem of Bialynicki-Birula [BB73]. A result of Fulton [Ful13] shows that in this situation the homology groups are freely generated by the homology classes of the affine cells. 1.1 Hilbert scheme of points We start with some rather general definitions and then quickly specialize them to single out the spaces we are interested in. We give concrete presentations for them and work with these for the rest of the thesis. Let T be a locally noetherian scheme, X a quasiprojective variety over T and L 1 2 CHAPTER1. FUNDAMENTALFACTS a very ample invertible sheaf on X over T . Definition 1.1.1. [Gro60] Let Hilb(X /T ) be the contravariant functor from the cate- gory of locally noetherian T -schemes to the category of sets, which, for locally noethe- rian T -schemes U ,V and a morphism f : U → V , is given by: Hilb(X /T )(U ) =©Z ⊂ X ×U closed subscheme, flat over Uª , Hilb(X /T )( f ) : Hilb(X /T )(V ) → Hilb(X /T )(U ); Z 7→ Z ×U V. For U a locally noetherian T -scheme and Z ⊂ X ×T U closed subscheme, flat over U , let p : Z → X and q : Z → U be the two projections and u ∈ U . Define the Hilbert poly- nomial of Z in u as Pu(Z )(m) := χ(OZu (m)) = χ¡OZu ⊗OZ p ∗(L m)¢ , m ∈ Z where χ is the Euler characteristic and Zu = q −1(u). One can prove that Pu(Z )(m) is a polynomial in m, independent of u, if U is connected. Then we can fix the Hilbert polynomial to create a subfunctor. Let P ∈ Q[x] and define HilbP (X ) to be the subfunc- tor given by HilbP (X /T )(U ) =( Z ⊂ X ×U closed subscheme ¯¯¯¯¯ Z is flat over U and Pu(Z ) = P for all u ∈ U) . Theorem 1.1.2. [Gro60] Let X be projective over T . For every P ∈ Q[x] the functor HilbP (X /T ) is representable by a projective T -scheme HilbP (X /T ). For an open sub- scheme Y ⊂ X the functor HilbP (Y /T ) is represented by an open subscheme HilbP (Y /T ) ⊂ HilbP (X /T ). Definition 1.1.3 (Hilbert scheme of points). From now on we will be interested in the case where T = spec(C), and we will write HilbP (X ) for HilbP (X /T ). Moreover we will only be interested in the case where P = n ∈ N is a constant polynomial. Then we will write either X [n] or Hilbn(X ), and call it the Hilbert scheme of n points over X . In fact we can identify the closed X [n](C) points with the closed zero-dimensional subschemes of length n of X which are defined over C. In the most simple case such a scheme is just the set of n distinct points of X with the reduced induced structure, hence the name. Definition 1.1.4. Let Sn be the symmetric group in n letters acting on X n by permut- ing the factors. The geometric quotient X n±Sn exists and is called the n-th symmetric power of X , and is denoted as X (n). We denote the quotient map as follow: Φn : X n → X (n). The n-th symmetric power parametrizes effective zero-cycles of degree n of X , i.e. 1.1. HILBERTSCHEMEOFPOINTS 3 formal linear combinations of points of X with nonnegative integer coefficients that sum to n: Xi ni [xi ] ∈ X (n) with xi ∈ X , ni ∈ N andXi ni = n. We have a stratification into locally closed subsets given by prescribing how many of the points in the support of the zero-cycle actually coincide. Definition 1.1.5. Let ν = ν0 ≥ · · · ≥ νr be a partition of n. Denote the diagonals of X n by and define ∆νi :=©(x1, . . . , xνi )¯¯x1 = · · · = xνiª ⊂ X νi ν :=Yi ∆νi ⊂Yi X νi ⊂ X n. X n Then we set X (n) ν ν¢ := Φn¡X n and X (n) ν := X (n) ν X (n) µ , \ [µ>ν where µ > ν means that µ is a coarser partition than n. The geometric points of X (n) are ν X (n) ν =(Xi ni [xi ] ∈ X (n) ¯¯ the points xi are pairwise distinct ) . Given Z a subscheme of X of length n its support is precisely an effective zero-cycle of degree n. This gives the following celebrated relation between Hilbn(X ) and X (n). Theorem 1.1.6. [MFK94, §5.4] There is canonical morphism called the Hilbert-Chow morphism πn : Hilbn(X ) → X (n) that at the level of points is given by Z 7→ Xx∈X len(Zx)[x], where len(Zx) is the multiplicity of Z at the point x. ν := π−1 n (X (n) The above stratification of X (n) induces a stratification of Hilbn(X ). Define its strata as X [n] ν ), for each ν ⊢ n. Along a strata of X (n) the fibers of πn are constant and depend only on the dimension of X , if X is smooth. For example, in the open smooth strata X (n) (1n ) where all the points are distinct πn is an isomorphism. 4 CHAPTER1. FUNDAMENTALFACTS Example 1.1.7. Let C be a smooth curve. Then the Hilbert-Chow morphism is actually an isomorphism that shows C [n] ∼= C (n). In fact there is only a single scheme structure on n points. Suppose that X = C, then X (n) ∼= X n as Newton's theorem on symmetric functions C[x1, . . . , xn]Sn ∼= C[e1, . . . , en] proves. Here ei is the i -th elementary symmetric function. One can also show that (P1)(n) = Pn. We will only be interested in the case of a smooth surface X . In this case the symmetric power is singular: as soon as at least two points coincide the stabilizer is not trivial. Famously it turns out that in this case the Hilbert scheme is smooth, as we will see. From now on we will focus on the case where X = C2. Observe that the Theorem of existence 1.1.2 proves that for every X , with a covering of opens isomorphic to C2, there is an open cover of Hilbn(X ) with opens that are isomorphic to Hilbn(C). This is true also in the category of complex analytic spaces thanks to the definition of the Duady space. We do not need this but we give as a reference [dCM00]. Example 1.1.8. Consider the case X = C2. Denote R = C[x, y] its ring of functions. Then we can identify (C2)[n] =©I ⊂ R ¯¯ I is an ideal such that dimC (R/I ) = nª . We call the dimension of the vector space R/I the length of I . For example if p1 = (x1, y1), . . . , pn = (xn, yn) are n distinct points of C2, then there is a unique ideal I of length n of functions that vanish exactly at p1, . . . , pn. These ideals represent the generic examples. At the other end of the spectrum of examples there are the powers of the maximal ideal (x, y)d that live in Hilb (C2) for every d ∈ N. d(d+1) 2 Observation 1.1.9. [Nak99, Chapter 1] A possibly more explicit description of (C2)[n] is the one of Nakajima in terms of commuting matrices. Consider Mx, My ∈ End(Cn), and v ∈ HomC(C, Cn). Then define eH := (¡Mx, My , v¢¯¯¯¯¯ Mx My − My Mx = 0, x M l y (v (1)) k, l ≥ 0E = Cn) . DM k The first condition says that the actions of Mx and My commute. The second condi- tion, that is a stability condition, says that there does not exist an Mx, My invariant subspace of Cn that contains the vector v (1) ∈ Cn, 1 ∈ C. There is an action of GLn(C) on such triples given by g ·¡Mx , My , v¢ :=¡g Mx g −1, g My g −1, g v¢ 1.1. HILBERTSCHEMEOFPOINTS 5 for g ∈ GLn(C). The action turns out to have closed orbits and trivial stabilizers. Naka- jima [Nak99, Theorem 1.9] proves that we have the following isomorphism: Hilbn(C2) ∼= H±GLn(C) . The above map is given by the following procedure on closed points. If I ⊂ C[x, y] is an ideal of length n we define Mx as the multiplication action of x on the n dimen- sional vector space C[x, y]/I . Similarly for My . The homomorphism v ∈ HomC(C, Cn) is given by v (1) = 1 mod I . It is clear that Mx and My commute since the multiplica- tion in C[x, y] is commutative. The stability condition follows from the fact that 1 is a C[x, y] generator of C[x, y]. Conversely given a triple¡Mx, My , v¢ ∈ eH we can define a map φ : C[x, y] → Cn as φ( f ) = f (Mx , My )v (1). Stability of the triple proves that φ is surjective and that I := Ker(φ) is an ideal of length n. Since the two matrices commute we can always simultaneously conjugate both matrices to upper triangular matrices. Call (λ1, . . . , λn) and (µ1, . . . , µn) the ele- ments on the diagonal of Mx and My respectively. Then the Hilbert Chow map sends the triple¡Mx, My , v¢ to {(λ1, µ1), . . . , (λn, µn)} ∈ (C2)(n) . For a more general surface X , not only smoothness, but most of the topolog- ical and geometrical properties of X [n] are a mixture of the corresponding properties of the base surface X and of the most singular fiber of the Hilbert-Chow map, some- times called the punctual Hilbert scheme. The punctual Hilbert scheme, and its flag versions, are the object of our study. Definition 1.1.10. Consider the Chow morphism π : (C2)[n] → (C2)(n), and denote 0 = (0, . . . , 0) ∈ (C2)(n). Then the punctual Hilbert scheme is Hilbn(0) := π−1(0) with the induced scheme structure. The closed points of Hilbn(0) are the schemes of length n whose support is concen- trated at the origin 0 ∈ C2. An example is (x, y n) ∈ (C2)[n]. If m = (x, y) is the maximal ideal in R = C[x, y] and R = C[[x, y]] is the completion of R in m, then we will see that Hilbn(0) = Hilbn(R)r ed = Hilbn(R/mn )r ed ⊂ Gras(n, R/mn ) . In terms of the description of commuting matrices: Hilbn(0) =©¡Mx , My , v¢ ∈ eH ¯¯ (Mx )n = (My )n = 0ª.GLn . Example 1.1.11. Pose n = 2. Then Hilb2(0) is isomorphic to P1. In fact all ideals of length two of functions with zeros only in (0, 0) are of the form (ω1x + ω2 y) + (x2, x y, y 2) 6 CHAPTER1. FUNDAMENTALFACTS as ideals of C[x, y], with [ω1 : ω2] ∈ P1. The intuition behind this is that when two points collide at the origin (0, 0) ∈ C2 we remember the direction along which they collided i.e. the vector [ω1 : ω2] ∈ P(T(0,0)C2), where T(0,0)C2 is the tangent space at the origin. The global Hilbert scheme Hilb2(C2) is then stratified as follow (C2)[2] (C2)(2) π2 (2) (1,1)F(C2)[2] (1,1)F(C2)(2) (2) (C2)[2] π2 (C2)(2) (2), where two points coincide, and a single point { pt } over (C2)(2) where the vertical arrow is the Hilbert-Chow map and has fiber P1 = Hilb2(0) over (C2)(2) (1,1), where two points are distinct. In fact, in this case, it is easy to see that the Hilbert scheme is the blow up of the 2-symmetric product of C2 at the diagonal, that is its singular locus. Example 1.1.12. Pose n = 3. Then the punctual Hilbert scheme Hilb3(0) is a cone over P1 with an isolated singular point. The singular point is the ideal m2 = (x2, x y, y 2). It is different from all the others in the sense that, if we call Z∞ the corresponding subscheme, then T(0,0)Z∞ is two dimensional, whereas all the other points of Hilb3(0) have tangent spaces that are one dimensional. Said in another way, the other points correspond to those subschemes that are contained in the germ of a smooth curve. They are called curvilinear. They are of the form I = (y 3, x + ω2 y + αy 2) plus the ide- als I = (x3, y + αx2). It is clear that these curvilinear ideals form an affine bundle over Hilb2(0) = (ω1x + ω2 y) + (x2, x y, y 2), and that the bundle is compactified with the point m2 at infinity. One can write down explicitly a model for Hilb3(0): consider in P4 = Proj(C[a, b, c, d , e]) the projective cone over a rational normal cubic given by equations ac − b2, ad − bc, bd − c 2. Then the family of subschemes of C2 parametrized by Hilb3(0) is the zero set of the ideal (ax + b y + ex2, bx + c y + ex y, c x + d y + e y 2). The global Hilbert scheme Hilb3(C2) is then stratified as follow: (C2)[3] (C2)(3) (3) (1,1,1) F (C2)[3] (1,1,1) F (C2)(3) (2,1)F (C2)[3] (2,1) F (C2)(3) (3) π3 (C2)[3] π3 (C2)[2] (1,1,1), i.e. where the three points are distinct, to Hilb2(0) over (C2)[3] where the Hilbert Chow map π3 has fibers that are respectively isomorphic to a point over (C2)(3) (2,1), i.e. where two points coincide and the third is different, and to Hilb3(0) over (C2)[3] (3) i.e. where the three points coincide. We will give a description of Hilb4(0) in the following section. 1.2. HILBERT-SAMUEL'SSTRATA 7 The induced scheme structure of π−1 n (0) is reduced as a result of Haiman proves. A celebrated result of Briançon proves irreducibility and tells us the dimension of the punctual Hilbert scheme. Irreducibility is equivalent to say that the curvilinear ideals of the form (y n, x + ω2 y + α1 y 2 + · · · + αn−2 y n1 ) are dense in every Hilbn(0), n ∈ N. Theorem 1.1.13. [Bri77, Theorem II.2.3], [Hai98, Poposition 2.10] The punctual Hilbert scheme Hilbn(0) is projective and irreducible of dimension n − 1. It is a locally complete intersection. The fact that Hilb2(C2) is a resolution of (C2)(2) is not accidental. In fact for every n the Hilbert scheme is in some sense the best resolution of singularities of the symmetric product of C2. Theorem 1.1.14. [Fog68, Theorem 2.4], [B+83, Theorem 3] The Hilbert-Chow mor- phism πn : (C2)[n] → (C2)(n) is a symplectic resolution of singularities. We are only interested in smoothness and not in the symplectic structure. Nowadays there are many different proofs of smoothness for (C2)[n]. We will show it by describ- ing the Zariski tangent space and giving its dimension at each point. We now define the flag Hilbert scheme that parametrizes flags of ideals of points. Definition 1.1.15 (Flag Hilbert scheme). For n = (n1, . . . , nk ) ∈ Nk, with n1 < · · · < nk, define the flag Hilbert scheme and the flag punctual Hilbert scheme to be, respectively, Hilbn(C2) :=©Ini ∈ Hilbni (C2)In1 ⊃ · · · ⊃ Inkª ⊂ k× Hilbn(0) :=©Ini ∈ Hilbni (0)In1 ⊃ · · · ⊃ Inkª ⊂ k× i =1 Hilbni (0) . Hilbni (C2), i =1 It is clear that we have an Hilbert-Chow map also for the flag Hilbert scheme and thus a corresponding stratification according to the multiplicities of the supports of the schemes in the flag. In particular the dimension of Hilbn(C2) is 2nk. The geometry of Hilbn(0) becomes more complicated even for short flags. For example irreducibility holds only in the case where n = (n, n +1), see [CE12, Prop. 18, Thorem 19]. The homol- ogy of Hilbn(0) is known for n = n and n = (n, n +1). We will give a basis of the homology for the case n = (n, n + 1, n + 2). 1.2 Hilbert-Samuel's strata In the rest of the thesis we want to understand basic geometrical properties of Hilbn(0), of Hilbn,n+1(0) and of Hilbn,n+1,n+2(0). In order to do so we introduce a stratification of these spaces due to Iarrobino. 8 CHAPTER1. FUNDAMENTALFACTS Definition 1.2.1 (Hilbert-Samuel type). Let R = C[[x, y]], m = (x, y) be its maximal ideal and Ri = mi±mi +1 be the space of forms of degree i . Suppose I ⊂ R is an ideal of length n. Then we define T (I ) = (ti (I ))i ≥0 ∈ N∞, the Hilbert-Samuel type, of I as: ti (I ) := dim³mi±I ∩ mi + mi +1´ = dim Ri±Ii where Ii := I ∩ mi±I ∩ mi +1 . Denote T =Pi ≥0 ti . Call initial degree of I the first index d = d (I ) such that td < d + 1. Example 1.2.2. Let I = (x, y n) ⊂ R. Then I has length n. We have ti (I ) = 1 for 0 ≤ i ≤ n −1 and ti (I ) = 0 for i ≥ n. Its initial degree is 1. Let md ⊂ R. Then the length of md is N = d(d +1) . We have ti (md ) = i + 1 for 0 ≤ i ≤ d, and ti (md ) = 0 for d + 1 ≤ i . Its initial degree is d. 2 Lemma 1.2.3. Let I ⊂ R be an ideal of length n and T = (ti (I ))i ≥0, then (2) I ⊃ mn. (1) dim m j±I ∩ m j =Pi ≥ j ti for all j ≥ 0, in particular T = n. Proof. Call Z = R/I and Zi the image of Z under the projection map R → R±Ii . Of course we have thatTi ≥0 Zi = 0. Since Z is finite dimensional, there must exist i0 such that Zi0 = 0, i.e. mi0 ⊂ I . There are isomorphisms of vector spaces: Zi = mi±mi ∩ I ∼= i0−1Mj =i R j.I j . t j = 0 implies mi ⊂ I , (2) is a consequence of T = n. Remark 1.2.4. Since every length n ideal of R contains mn, we can see it as an ideal Then if we choose i0 to be minimal we have Ri±Ii 6= 0 for i < i0. This proves (1). Since in R/mn . Thus the Hilbert scheme Hilbn¡ R/mn¢ also parametrizes the ideals of length n in R. For the same reason all the reduced schemes¡Hilbn¡ R±mk¢¢red are naturally isomorphic for k ≥ n. We will denote one of these by¡Hilbn( R)¢red. Here¡Hilbn( R)¢red is the closed subscheme with the reduced induced structure of the Grassmannian Grass(n, R/mn ) of n dimensional quotients of R/mn whose geometric points are the ideal of length n in R/mn . Of course the intuition is that, since a point in Hilbn(C2) that is supported only at the origin is an ideal of I ⊂ C[x, y] such that m ⊂ I we can see it as a point of¡Hilbn( R)¢red and vice-versa. To be more precise we state the following Lemma that can be found in Goettsche [Göt94, Chapter 2]. Lemma 1.2.5. [Göt94, Lemmas 2.1.2, 2.1.4] The natural morphism that maps a sub- scheme of length n supported at a point to this point π : (C2)[n] → C2 factors through the (n) Hilbert Chow map and is a locally trivial fiber bundle in the Zariski topology with fiber We can regroup ideals according to their Hilbert-Samuel function to get a strat- ¡Hilbn( R)¢red = Hilbn(0). ification of Hilbn( R)red. 1.2. HILBERT-SAMUEL'SSTRATA 9 Definition 1.2.6 (Hilbert-Samuel's strata). Let T = (ti )i ≥0 be a sequence of nonnega- tive integers, with T = n, we define the Hilbert-Samuel's stratum MT and the homoge- nous Hilbert-Samuel's stratum GT to be, respectively, MT :=©I ∈ Hilbn(0) T (I ) = Tª ⊂ Hilbn(0), GT :=©I ∈ Hilbn(0)¯¯ T (I ) = T, I is homegenous ª ⊂ Hilbn(0). Let ρT : MT → GT be the surjective morphism that associates to an ideal I the associ- ated homogenous ideal, i.e. the ideal generated by all the initial forms of f ∈ I . It is surjective and the natural embedding GT ⊂ MT is a section. The fact that such a map is well defined is a classical observation, and can be proved, for example, by reasoning a way similar to the proof of Lemma ??. If T = (T1, . . . , Tk ) is a k-tuple of sequences of nonnegative integers Ti = (ti , j ) j ≥0 satisfying Ti = ni then we define MT := Hilbn(0) ∩¡MT1 × · · · × MTk¢ , GT := Hilbn(0) ∩¡GT1 × · · · ×GTk¢ . Again we have a morphism ρT : MT → GT with section GT ⊂ MT. It is clear that the strata MT which are not empty stratify Hilbn(0). Lemma 1.2.7. [Iar77, Lemma 1.3] Let T = (ti )i ≥0 be a sequence of nonnegative integers with T = n, then MT and GT are not empty if and only if T = (1, 2, . . . , d , td , td +1, . . . , tn−1, 0, . . . ) with d + 1 > td ≥ td +1 ≥ · · · ≥ tn−1 ≥ 0. Moreover, if n = (n, n + 1, . . . , n + k − 1), given T = (T1, . . . , Tk ) with Ti = ni , GT and MT are not empty if and only if (1) Ti = (1, 2, . . . , di , ti ,d , ti ,d +1, . . . ) with di + 1 > ti ,d ≥ ti ,d +1 ≥ · · · ≥ ti ,ni −1 ≥ 0 (2) For all j = 2, . . . k there exists and index m j such that t j ,m j = t j −1,m j + 1. From now on, we call such a k-tuple of Ti admissible. Example 1.2.8. Let T = (ti )i ≥0 be with ti = 1 for 0 ≤ i ≤ n − 1 and ti = 0 for i ≥ n, with n ∈ N. Then GT is isomorphic to P1 and it is parametrized as follow: GT = (ω1x + ω2 y) + m n, [ω1, ω2] ∈ P1. On the other hand MT fibers on GT with affine fibers of dimension n − 2 given, for example on the affine chart {(x + ω2 y) + mn ω2 ∈ C} ⊂ GT , by ©(x + ω2 y + α1 y 2 + · · · + αn−2 y n−1, y n) ω2, αi ∈ C i = 1, . . . n − 2ª . One can work out the transition functions on the intersection with the affine chart {(ω1x + y) + mn ω1 ∈ C} ⊂ GT to check that ρT is an affine bundle that is not a linear 10 CHAPTER1. FUNDAMENTALFACTS bundle. The ideals in MT are called curvilinear as they arise when n points collide following a trajectory that describes a smooth curve. This is equivalent of saying that the initial degree is one. The Theorem of Briançon 1.1.13 proves that MT ⊂ Hilbn(0) is a Zariski open and dense subset. In terms of commuting matrices these points can be parametrized as follow: 0 a1 a2 0 a1 0 1 0 1 0 0 ... 0 0 . . . 0 0 0 . . . . . . ... 0 . . .  0 0 1 0  ,  0 0 0 ... 0 0 . . . 0 . . . an . . . an−1 ... 0 ... a1 0 ,  0 0 ... 0 1   . Example 1.2.9. Pose n = 4. The possible admissible types for an ideal I ∈ Hilb4(0) are T = (1, 1, 1, 1) and T ′ = (1, 2, 1). As we have seen in the above example MT is an affine bundle over P1 with fiber of dimension 2. Instead MT ′ = GT ′ ∼= P2 as one can check with the following parametrization: MT ′ =©(φ1, φ2) + m 3¯¯ dimC2 〈φ1, φ2〉±m3 = 2, φi ∈ m 2ª . Remark 1.2.10. All of the points of MT ′ are singular in Hilb4(0), even though not all points of MT ′ are analytically equivalent in Hilb4(0): in fact one can see that there are two possibilities that give rise to different geometrical behaviors. Interestingly enough the two different behaviors can be describe like this: one set of points is the set of points I of MT ′ such that there exists at least a point J I in M(1,2,1,1) with (I , J I ) ∈ Hilb4,5(0). The other set is the complementary in MT ′. Of all the examples we wrote up for small n (say n ≤ 11) similar criteria to identify the analytical type of points in Hilbn(0) always hold. The tangent spaces at points of Hilbn(0) is the subject of the recent paper [BS16]. The admissible types for flags of two ideals (I1, I2) ∈ Hilb4,5(0) are the following: T1, T2 = (1, 1, 1, 1, 0), (1, 1, 1, 1, 1, 0), T1, T ′ 2 = (1, 2, 1, 0), (1, 2, 1, 1, 0) and T ′ 2 = (1, 2, 1, 0), (1, 2, 2, 0). We describe an open chart of the Hilbert-Samuel's strata, an obvious change of coordinates of the plane shows how to cover the cor- responding stratum with such charts. 2 = (1, 1, 1, 1, 0), (1, 2, 1, 1, 0), T ′ 1, T ′′ 1, T ′ (y 5, x + ω2 y + α1 y 2 + α2 y 3 + α3 y 4)) , MT1,T2 = Bundle with fiber A3 over P1 ⊃( (y 4, x + ω2 y + α1 y 2 + α2 y 3) ⊃ = Bundle with fiber A2 over P1 ⊃ ! Ãy 4, x2 + ω2x y + α1x y 2 + α2x y 3, (y 4, x + ω2 y + α1 y 2 + α2 y 3) ⊃ x y + ω2 y 2 + α1 y 3 + α2 y 4 MT1,T ′ , 2 1.3. TORUSACTIONANDBOREL-MOOREHOMOLOGY 11 (y 3, x2 + ω2x y, x y + ω2 y 2) ⊃ x y + ω2 y 2 + α2 y 3 !  = Bundle with fiber A2 over P1 ⊃ Ãy 4, x2 + ω2x y + α1 y 3 = Bundle with fiber P1 over P2 ⊃( (y 3, x2 + α1 y 2, x y + α2 y 2) ⊃ (y 3, x2 + θx y + (α1 + θα2)y 2)) . , MT ′ 1,T ′ 2 MT ′ 1,T ′′ 2 The admissible types for flags of three ideals (I1, I2, I3) ∈ Hilb4,5,6(0) are T1, T2, T3 = (1, 1, 1, 1), (1, 1, 1, 1, 1), (1, 1, 1, 1, 1, 1) , T1, T2, T ′ 2, T ′′ T1, T ′ 1, T ′ T ′ 1, T ′′ T ′ 2, T ′ T1, T ′ 3 = (1, 1, 1, 1), (1, 2, 1, 1), (1, 2, 1, 1, 1) , 2, T ′ 1, T ′ T ′ 3 = (1, 2, 1), (1, 2, 1, 1), (1, 2, 1, 1, 1) , 2 , T ′′ 1, T ′′ T ′ 3 = (1, 2, 1), (1, 2, 2), (1, 2, 2, 1) , 3 = (1, 1, 1, 1), (1, 1, 1, 1, 1), (1, 2, 1, 1, 1), 3 = (1, 1, 1, 1), (1, 2, 1, 1), (1, 2, 2, 1, 0), 2, T ′′ 2 , T ′′′ 3 = (1, 2, 1), (1, 2, 1, 1), (1, 2, 2, 1) 3 = (1, 2, 1), (1, 2, 2), (1, 2, 3) . As the list starts to be too long we give a description of only some of Hilbert-Samuel's strata. 1,T ′ 2,T ′ 3 MT1,T2,T3 = Bundle with fiber A4 over P1, = Bundle with fiber A3 over P1, MT ′ = Bundle with fiber P1 over a bundle with fiber P1 over P2 = Bundle with fiber P0 = { pt } over MT ′ 2 ,T ′′ 3 1,T ′′ . 1,T ′′ 2 MT ′ MT ′ 1,T ′′ 2 ,T ′′′ 3 Proposition 1.2.11. [Iar77, Theorem 3.13] Let T = (ti )i ≥0 be an admissible sequence of nonnegative integers. Then MT is smooth, and GT is smooth and projective. The map ρT : MT → GT is an affine fibration, Zariski locally trivial. The dimensions of MT and of GT are given in 2.1.16. We will see later a proof of this proposition. The techniques used are central to many of the discussions in this thesis. 1.3 Torus action and Borel-Moore homology In this section we describe the techniques used to compute the basic topological properties of the spaces we introduced. There are two main ingredients. The first is a result of Fulton that tells us that if we find an affine cell decomposition for a variety the homology classes of the closure of the cells form a graded basis for the Borel-Moore homology of the variety. The second is a result of Bialynicki-Birula that finds for us an affine cell decomposition for varieties with a nice enough torus action. In the rest of the section we describe a torus action on the relevant spaces and we study the fixed points. Borel-Moore homology is historically the preferred homological theory to study the topology of Hilbert schemes. We will only need Proposition 1.3.2 below, and we re- 12 CHAPTER1. FUNDAMENTALFACTS fer at Fulton [Ful13, Chapter 19] and reference therein for more details. Borel-Moore homology, indicated with H∗, is singular homology with locally finite supports and integer coefficients. For a space X that is imbedded as a closed subspace of Rn one can see that Hi (X ) ∼= H n−i (Rn , Rn \ X ) where the group on the right is relative singular cohomology with integer coefficients. For a complex scheme X there is a cycle map cl : A∗(X ) → H∗(X ) where A∗ is the Chow group of X . Definition 1.3.1. Let X be a scheme over C. A cell decomposition of X is a filtration X = Xn ⊃ Xn−1 ⊃ · · · ⊃ X0 ⊃ X−1 = ; such that Xi \ Xi −1 is a disjoint union of schemes Ui , j isomorphic to affine spaces Ani , j for all i = 0, . . . , n. We call Ui , j the cells of the decomposition. Often we stress the adjective affine and say that X has an affine cell decomposition. Proposition 1.3.2. [Ful13, Exercise 19.1.11] Let X be a scheme over C with a cell de- composition. Then (1) H2i +1 (X ) = 0 for all i . (2) H2i (X ) is the free abelian group generated by the homology classes of the closure of the i -dimensional cells, for all i . (3) The cycle map cl : A∗(X ) → H∗(X ) is an isomorphism. Definition 1.3.3. We will only study the Borel-Moore homologies of varieties that have an affine cell decomposition. In particular all odd dimensional homology groups will be zero so that we will denote bi (X ) ∈ N, and call it the i -th Betti number of the variety X , the dimension of H2i (X ). Moreover we define Pq (X ) ∈ N[q], the Poincaré polynomial of X , as: Pq (X ) :=Xi q bi (X ) =Xi q dim H 2i (X ). (1.1) Let X be a smooth projective variety over C with an action of the multiplicative group T = C∗. Denote this action with a dot " · ". If x ∈ X is a fixed point for this action, the torus acts also on the tangent space at x, denote it Tx X . Let T + x X ⊂ Tx X be the linear subspace on which all the weights of the induced action of T are positive. Theorem 1.3.4. [BB73, Theorem 4.4] Let X be a smooth projective variety over C with an action of T. Assume that the set of fixed points is the finite set { x1, . . . , xm }. For all 1.3. TORUSACTIONANDBOREL-MOOREHOMOLOGY 13 i = 1, . . . , m we define the attracting set at the fixed point xi as: Xxi = Xi :=½ x ∈ X ¯¯¯¯ lim t →0 Then we have: t · x = xi¾ . (1) X has an affine cell decomposition whose cells are the Xi . (2) Txi Xi = T + xi X . Remark 1.3.5. The condition of projectivity is only needed to ensure that all the limits limt →0 t · x, x ∈ X actually exists. The theorem remains true if we replace the latter condition with the hypothesis of projectivity of X . We have a two dimensional torus action on Hilbn(0) that comes from the rescal- ing action on R = C[[x, y]]. If τ = (τ1, τ2) ∈ T2, f = f (x, y) ∈ C[x, y], I ∈ Hilbn(C2), then τ · f = f (τ1x, τ2 y) τ · I =¡τ · f¯¯ f ∈ I¢ . Of course this action is the restriction of the two torus action on Hilbn(C2). The following observation is well known and clear. Lemma 1.3.6. The fixed points for the T2 (cid:8) Hilbn(C2) action are the monomial ideals in Hilbn(C2). In particular we have a bijection of sets: © fixed points of T2 (cid:8) Hilbn(C2)ª ←→© monomial ideals in Hilbn(C2)ª Let n ∈ Nk for some k ∈ N. The fixed points for the T2 (cid:8) Hilbn(C2) action are the flags of monomial ideals in Hilbn(C2). Every one-parameter subgroup of T2 will then act on Hilbn(0). Definition 1.3.7. Let n = (n1, . . . , nk ) ∈ Nk. Let φ : Gm → T2 be a one-parameter subgroup of the form φ(τ) = (τw1 , τw2 ) with w1, w2 ∈ Z. We say that it is generic with respect to the action on Hilbn(C2) if it has the same fixed points as T2. This means avoiding a finite set of given hyperplanes in the lattice of one-parameter subgroups of T2. We define two generic one-parameter subgroups that we will use, for different goals, in the rest of the thesis: T∞ generic, with weights w1, w2 such that 0 < w1 < w2 and 1 ≪ w2 w1 , T1+ generic, with weights w1, w2 such that 0 < w1 < w2 and nk · w1 > (nk − 1) · w2. (1.2) Here 1 ≪ w2 w1 is relative to nk and in fact it is enough that nk < w2 w1 . The action of T1+ behaves especially well with respect to the stratification in Hilbert-Samuel's strata. 14 CHAPTER1. FUNDAMENTALFACTS Lemma 1.3.8. [Göt90, Lemma 2.2.9] ?? Let T be an admissible sequence of nonnegative integers as in 1.2.7. Suppose that we are considering the T1+ torus action on Hilbn(0), i.e. with weights w1, w2 such that 0 < w1 < w2 and nw1 > (n − 1)w2. Then we have that: (1) MT is a union of attracting sets that are attracting sets of Hilbn(0). (2) ρT : MT → GT is equivariant with respect to the T1+ action. (3) The T1+ action induces an attracting sets decomposition of GT . The attracting sets are the intersection of the attracting sets of MT with GT . The same is true for n = (n1, . . . , nk ), T = (T1, . . . , Tk ) admissible k-tuple of sequences of nonnegative integers, and the T1+ action with weights w1, w2 such that 0 < w1 < w2 and nk w1 > (nk − 1)w2. Proof. We prove it in the case of Hilbn(0) since the proof is the same and we do not want to complicate it with indexes. Let then I ∈ Hilbn(0) be with Hilbert function T . For j ∈ N call s j = j + 1 − t j the dimension of I j , the space of initial forms of I of degree j . Call J = limt →0 t · I . We need to prove that J ∈ MT . Suppose that the Hilbert func- tion of J is T ′ = (t ′ i )i ≥0. Choose f1, . . . , fs j ∈ I such that their initial forms g1, . . . , g s j are a basis for I j , so that the gi are homogenous of degree j . We can assume, up to linear combinations, that the gi are of the form: gi = xl (i ) y j −l (i ) + Xm>l (i ) gi ,m xm y j −m gi ,m ∈ C with l (1) > l (2) > · · · > l (s j ). Then by the choice of weights we have that limt →0 t · fi = xl (i ) y j −l (i ), in fact other terms of fi either must have higher degree than j and then go to 0 faster, either have degree j i.e. are in the support of gi , and then have higher y degree that forces them to go to zero faster. This proves that all the xl (i ) y j −l (i ) ∈ J j , and then t ′ j = j + i − dim J j ≥ t j . Since J ∈ Hilbn(0), it is still true that T ′ = n, and thus T ′ = T . Observation 1.3.9. In the case of Hilbn(C2) ⊃ MT all the attracting sets are, as we will see, affine cells. This follows from smoothness of Hilbn(C2), that implies smoothness of Hilbn(P2) and the fact that its attracting sets are affine. It is then an easy observation that the attracting sets for MT are some of the attracting sets for Hilbn(P2). This is the only case Goettsche was interested in. For the case Hilbn,n+1(C2) ⊃ MT1,T2, always thanks to smoothness, the attracting sets are affine cells. The main geometric result of this thesis is that it is still true that the attracting cells are affine for MT1,T2,T3, since these spaces are smooth, even though the ambient space Hilbn,n+1,n+2(C2) it is not. To better deal with the fixed points of the torus action we need to introduce some notations. 1.3. TORUSACTIONANDBOREL-MOOREHOMOLOGY 15 Definition 1.3.10 (Partitions and Young diagrams). Let ν = ν0 ≥ ν1 ≥ . . . be a partition of n, i.e. the νi ∈ N are nonnegative integers weakly decreasing that sum to n. We will write ν ⊢ n to say that ν is a partition of n. We will also write ν = n. The length of the partition ν, denoted ℓ(ν), is the minimum index i for which νi = 0. We will also write a partition ν ⊢ n as ν = (1α1 , 2α2 , . . . , nαn ) where Xi αi i = n and αi ∈ N if the parts of ν are αn times n, . . ., α1 times 1. We will confuse the n-tuple of integers (α1, . . . , αn) with ν. Consider the first quadrant of R × R as covered by square (boxes) with vertices the points of integer coordinates, side 1, and indexed by the coordinate of their left-lower vertex. We denote this set of boxes ∆. A Young diagram Γ is a finite set of boxes of ∆ such that if (i , j ) ∈ Γ then (i − 1, j ) and (i , j − 1) are either in Γ or have negative co- ordinates. Young diagrams are also called Ferrers diagrams, and they are in bijection with partitions of integers. The Young diagram of ν, Γ(ν), is the set of boxes labeled by (0, 0), (0, 1), . . . (0, ν0 −1), (1, 0), (1, 1), . . . (1, ν1 −1), . . ., i.e. is the Young diagram with νi boxes in the i -th column. The Young diagram Γ(ν) associ- ated to ν = (5, 4, 3, 1). We think of Γ as living in the lattice N × N, that we called ∆. Conversely if we are given a Young diagram we write the partition associated to it as ν = #{boxes in 0-th column} ≥ · · · ≥ #{boxes in i-th column} ≥ . . .. This is a bijection between Young diagrams and partitions. From now on we will confuse the two and write Γ ⊢ n. For example the length of a Young diagram is the number of its columns. The diagonal sequence of ν is T (ν) = (t0(ν), . . . , t j (ν), . . . ), where ti (ν) := #© (l , j ) ∈ Γ(ν)¯¯ l + j = iª . 3 4 3 2 t0 = 1 The diagonal sequence of Γ is the number of boxes on each antidiagonal. In this case is T (Γ) = (1, 2, 3, 4, 3). Of course The hook difference of a box (u, v ) ∈ Γ(ν), denoted as hu,v (ν) or simply as hu,v , is T :=Pi ti = Γ. hu,v := # { (l , v ) ∈ Γ(ν) l > u} − #© (u, j ) ∈ Γ(ν)¯¯ j > vª . 16 CHAPTER1. FUNDAMENTALFACTS It is the difference between the number of boxes in Γ(ν) in the same row and to the right of (u, v ) and the number of boxes in the same column and above (u, v ). 0 −1 0 1 0 −1 0 −1 2 3 2 3 2 1 0 The boxes of Γ are each marked with their hook dif- ference. Definition 1.3.11 (Bijection between Young diagrams and torus fixed points of Hilbn(0)). We can interpret each box in N × N as a monomial in R = C[x, y]: to the box labeled by (i , j ) we associate the monomial xi y j . Then we can associate to each partition a monomial ideal. Explicitly if ν = ν0 ≥ ν1 ≥ . . . , νn−1 ≥ 0 is a partition of n, we define Iν as Iν =³y ν0, x1 y ν1, . . . , xi y νi , . . . , xℓ(n)´ . Observe that the monomials that are not in Iν are exactly the monomials labeled by boxes in Γ(ν), so it is clear that Iν ∈ Hilbn(0). Moreover to each I ∈ Hilbn(0) we can associated a partition of n by defining νi := minnk ¯¯¯ xi y k ∈ Io . This gives a bijection between monomial ideals and partitions. When we look at it as a bijection between Young diagrams and monomial ideals we write I Γ or I (Γ). For a monomial ideal in I ∈ Hilbn(0) we define the standard monomial genera- tors (α0, . . . , αs ) as the list of minimal monomial generators of I ordered with decreas- ing x power. If Γ = Γ(I ) is the corresponding Young diagram, its standard monomial generators are the boxes (i , j ) ∈ ∆ \ Γ such that (i − 1, j ) (and (i , j − 1)) is either in Γ or i − 1 < 0 (or j − 1 < 0). These are the external corners of Γ. α4 α3 α2 α1 α0 The standard monomial generators of the represented torus fixed point of Hilbn(C2). The indexes are labelled so that to a lower index corresponds a monomial with higher x degree. Lemma 1.3.12. Let ν = ν0 ≥ ν1 ≥ . . . be a partition of n. Then the monomial ideal I associated to the partition ν, i.e. I =¡y ν1, . . . , xi y νi , . . . , xℓ(ν)¢, satisfies T (I ) = T (ν). Proof. Let T (I ) = (ti )i ≥0. The monomials xi y j with i + j = l and j ≥ νi are a basis of the space Il of homogenous polynomials of degree l in I . Then we have tl = l + 1 − #©(i , j ) ∈ N2¯¯ i + j = l , j ≥ νiª = #©(i , j ) ∈ Γ(I )¯¯ i + j = lª = tl (Γ(I )). 1.3. TORUSACTIONANDBOREL-MOOREHOMOLOGY 17 Definition 1.3.13. Let Γ1 and Γ2 be two Young diagrams of size, respectively, n and n + 1, such that Γ1 ⊂ Γ2. Then the monomial ideals associated satisfy I Γ1 ⊃ I Γ2 and the couple represents a fixed point of Hilbn,n+1(0). In this case we will use the following notation: Γ = (Γ1, Γ2) ⊢ [n, n + 1]. Clearly there is a bijection between fixed points of Hilbn,n+1(0) and couples (Γ1, Γ2) ⊢ [n, n + 1]. Let Γ1, Γ2 and Γ3 be three Young diagrams of size, respectively, n, n +1 and n +2, such that Γ1 ⊂ Γ2 ⊂ Γ3. Then the monomial ideals associated satisfy I Γ1 ⊃ I Γ2 ⊃ I Γ3 and the triple represents a fixed point of Hilbn,n+1,n+2(0). In this case we will use the following notation: Γ = (Γ1, Γ2, Γ3) ⊢ [n, n + 1, n + 2]. Clearly there is a bijection between fixed points of Hilbn,n+1,n+2(0) and triples (Γ1, Γ2, Γ3) ⊢ [n, n + 1, n + 2]. 18 CHAPTER1. FUNDAMENTALFACTS Chapter 2 Tangent spaces at torus fixed points In this chapter we study the Zariski tangent space of the spaces Hilbn(C2), Hilbn,n+1(C2) , and Hilbn,n+1,n+2(C2) at a torus fixed point. The stepping stone is an homological interpretation of the tangent spaces as spaces of R-homomorphisms, where R = C[x, y], that dates back to Grothendieck. Then one can interpret these R-homomorphisms in terms of the combinatorial data that describe a torus fixed point. The chapter is divided in three sections: one for each of the spaces Hilbn(C2), Hilbn,n+1(C2) , and Hilbn,n+1,n+2(C2). The goal of each section is to define a pure weight basis for the tangent space at each fixed point and to study the weights of the elements of these bases. The study of these weights will also help to understand the tangent spaces of the Hilbert-Samuel's strata of our Hilbert schemes. In the cases of Hilbn(C2) and Hilbn,n+1(C2) we will then be able to show that the spaces are smooth, and we will thus give graded bases for the homologies of Hilbn(0) and Hilbn,n+1(0). As we will see the space Hilbn,n+1,n+2(C2) is not smooth. However we will prove in the next chapter that the Hilbert-Samuel's strata are still smooth, thus allowing us to use the same techniques to give graded bases for their homologies and for that of Hilbn,n+1,n+2(0). The rough strategy is the following: given a fixed point of Hilbn(C2) labelled by Γ1 we define a basis B (Γ1) for the tangent space at I Γ1. Then we will use B (Γ1) to build a basis B (Γ1, Γ2) for the tangent space at the fixed point (I Γ1, I Γ2) ∈ Hilbn,n+1(C2) with Γ2 ⊢ n + 1. The modifications we need to perform on B (Γ1) depend on the combina- torics of the couple of Young diagrams (Γ1, Γ2). Then we start by B (Γ1, Γ2) to build a basis B (Γ1, Γ2, Γ3) for the tangent space of Hilbn,n+1,n+2(C2) where Γ3 ⊢ n + 2. Most of the modifications needed will only depend on the couple (Γ2, Γ3), exactly as in passing from B (Γ2) to B (Γ2, Γ3). In fact only few (in the general case only one) modifications will actually depend on the full triple (Γ1, Γ2, Γ3). This is really the key philosophi- cal point of most of the arguments in this thesis: the geometrical or combinatorial 19 20 CHAPTER2. TANGENTSPACESATTORUSFIXEDPOINTS properties of a flag of three ideals (I1, I2, I3) can be understood by looking at the corre- sponding properties for two flags of two ideals, i.e. (I1, I2) and (I2, I3), independently, and then taking into account some, in our cases always manageable, properties that are truly intrinsic to the triple. We start the chapter with the interpretation of the tangent spaces of the various Hilbert schemes in terms of R-homomorphisms. Lemma 2.0.14. [Gro60] Let I ∈ Hilbn(C2) be a fixed point and denote with TI¡Hilbn(C2)¢ the Zariski tangent space of Hilbn(C2) at I . Then we have a natural T2-equivariant isomorphism TI Hilbn(C2) ∼= HomR (I , R/I ). Let n = (n, n+1, n+2, . . ., n+k). Let I = (I1, I2, . . . , Ik ) ∈ Hilbn(C2) be a fixed point, and denote TI¡Hilbn(C2)¢ the Zariski tangent space of Hilbn(C2) at the point I. For 1 ≤ i < j ≤ k we have Ii ⊃ I j so that we can define the obvious maps: φi j : HomR (Ii , R±Ii ) → HomR (I j , R±Ii ), ψi j : HomR (I j , R.I j ) → HomR (I j , R±Ii ), Define also the projection maps ( f : Ii → R±Ii ) 7→ ( f¯¯I j ( f : I j → R.I j ) 7→ (p ◦ f : I j → R.I j ։ R±Ii ) . : I j → R±Ii ), πi j : kMl =1 HomR (Ii , R±Ii ) → HomR (Ii , R±Ii ) ⊕ HomR (I j , R.I j ) . Then we have a T2-equivariant isomorphism TI¡Hilbn(C2)¢ ∼= \1≤i < j ≤k¡Ker(φi j − ψi j ) ◦ πi j¢ . Now that we know what the tangent spaces are, we only need to find weight bases for them. 2.1 A weight basis for TI Hilbn(C2) Suppose that I ∈ Hilbn(C2) is a torus fixed point. The tangent space TI Hilbn(C2) at I is then equipped with the torus action. The goal of this section is to understand the tangent space TI Hilbn(C2) in terms of R-homomorphisms HomR (I , R/I ), and in particular of those R-homomorphisms that are of pure weights with respect to the torus action. The goal is to visualize them as arrows of boxes of Γ(I ). It is clear that, to describe an f ∈ HomR (I , R/I ) we need only to prescribe the images of the generators of I , and we need only to describe these images in terms of 2.1. AWEIGHTBASISFOR TI HILBN (C2) 21 linear combinations of monomials in Γ(I ). It is good to visualize the situation in terms of Young diagrams: on the left we have, in black, the set of standard monomial gener- ators of I and on the right we have, in gray, the elements of Γ(I ) i.e. those monomials that are not in I . In black the standard monomial In gray the elements of Γ(I ). generators of I . If we only look for maps f of pure weight, the image of a generator α must be itself a scalar multiple of a monomial β in Γ(I ), so graphically α moves p boxes to the left and q boxes upward, where p and q can be negative, and reaches β inside Γ(I ). In terms of monomials f (α) = cβ = c x p y q α with x p y q ∈ C[x ±1, y ±1], c ∈ C. If the scalar c is not zero, the fact that f is an R-homomorphism forces every other α′ ∈ I to be sent either to c x p y q α′ or to zero; graphically they must move by the same exact translation, or they must go to zero. For α a standard monomial generator of I , and β ∈ Γ(I ), call Sα,β :=© f ∈ HomR (I , R/I ) f (α) = βª . It can happen that Sα,β = ;. For example, if I = m2 = (x2, x y, y 2) ∈ Hilb3(0), then there does not exist f ∈ HomR (I , R/I ) that sends x2 7→ 1. y 2 1 x y x2 y x2 If x2 7→ 1 then x2 y 7→ y, but then x y 7→ γ cannot be defined. In fact its image γ ∈ R/I should be such that xγ = y. This is impossible. To understand for which α and β we have that Sα,β 6= ; we introduce the following notations. Definition 2.1.1. Let ν = ν0 ≥ ν1 · · · ≥ νn−1 be a partition of n, let Γ be the associated Young diagram and let I = I Γ = (α0, . . . , αs ) be the monomial ideal associated to Γ, i.e. I =¡y ν1, . . . , xi y νi , . . . , xℓ(ν)¢. We define pi := degy αi +1 − degy αi = vertical distance between αi and αi +1, qi := degx αi − degx αi −1 = horizontal distance between αi and αi −1, ps := ∞, q0 := ∞. 22 CHAPTER2. TANGENTSPACESATTORUSFIXEDPOINTS For each α = αi we also define Pα = Pαi :=© β ∈ Γ¯¯ degx β < degx α and βy pi ∈ Iª , Qα = Qαi :=n β ∈ Γ¯¯ degy β < degy α and βx qi ∈ Io, Pαs := ;, Qα0 := ;. Example 2.1.2. Let I be the monomial ideal represented by the diagram below. Let α = α3 be the generator of I marked in the picture with its name. We have that p3 = 2 and q3 = 3. The elements of Pα are indicated with a p while the elements of Qα are indicated with a q. p p p p p p p p p p p p p p α q q q q q q q q q q q q The fact that β ∈ Pα ∪Qβ implies that there exists an f ∈ Sα,β, as the next Lemma proves. Lemma 2.1.3 (definition). Let I = (α0, . . . , αs ) be a monomial ideal of length n, and let α = αi be one of its standard monomial generators. Let β ∈ Pα ∪ Qα. Then Sα,β 6= ;. In this case we define fα,β ∈ Sα,β as follow: If β ∈ Pα, then β ∈ Γ(I ) is q boxes to the left and p boxes above α, where p, q ≥ 0. We define fα,β by prescribing the images of each of the generators of I as: x q¶ , if β ∈ Pαi , β = αiµ y p y p¶ , if β ∈ Qαi , β = αiµ x q 0 then fαi ,β(αk ) = then fαi ,β(αk ) = αk³ y p x q´ β if k < i , if k = i , if i < k. if k < i , if k = i , if i < k. 0 β α j³ x q y p´ If β ∈ Qα, then β ∈ Γ(I ) is q boxes to the right and p boxes below α, where p, q ≥ 0. We define fα,β by prescribing the images of each of the generators of I as: Proof. Suppose that α = α3 is the one depicted below and that β ∈ Qα is the box two boxes below it and one to the right, i.e. the box marked with the 3. Then fα,β is the homomorphism depicted in the picture, where with the index k we denote the box inside Γ(I ) that is the image of αk if and only if this image is not zero. 2.1. AWEIGHTBASISFOR TI HILBN (C2) 23 α8 α7 8 α6 7 6 α5 5 α3 x3α3 3 x3 f (α3) Of course fαi ,β(αk ) is well defined if k > i since it will be a monomial in C[x, y] (graph- ically it will not fall out of the positive quadrant). The condition β ∈ Qα3 precisely makes sure that we can send all generators with lower index to zero: the element of I with lowest degree that is both divisible by αi and αk with k < i is x qi αi = y pi −1αi −1, but, by definition of Qαi , x qi β ∈ I , and then 0 mod I . If β ∈ Pα the argument is completely analogous. Observation 2.1.4. This is slightly different from the construction of Cheah [Che98], and it is the choice of Goettsche [Göt90]. In the case of Cheah, fα,β is chosen to be the element in Sα,β that sends the biggest possible number of generators to zero. Definition 2.1.5 (B (I )). For I = (α0, . . . , αs ) a monomial ideal with standard monomial generators αi 's, we define B (I ) to be the finite subset of TI Hilbn(C2)) that contains all the fα,β as defined in Lemma 2.1.3 : B (I ) :=© fα,β ¯¯ α standard monomial generator and β ∈ Pα ∪Qαª . Lemma 2.1.6. Let I = (α0, . . . , αs ) be a monomial ideal of length n with prescribed stan- dard monomial generators. The set B (I ) defined in 2.1.5 has cardinality 2n. (2.1) for a fixed α, Pα and Qα are disjoint. Then #B (I ) =Ps Proof. By definition it is clear that fα,β = fα′,β′ if and only if (α, β) = (α′, β′). Moreover i =0(#Pαi + #Qαi ). Suppose that ν0 ≥ ν1 ≥ · · · ≥ νs−1 > νs = 0 is the associated partition of n, i.e. the number of columns of Γ(I ). Then for a generator αi the distance pi = νs−i −1 − νs−i (for i = s we can put ps = ∞, but it does not matter since Pαs = 0). The number of elements in Pαi is equal to pi times the number of columns in Γ(I ) that are on the left of αi . Thus we have that #Pαi = (νs−i −1 − νs−i )(s − i ) = Xi ≤k<s (νs−i −1 − νs−i ). (2.2) 24 CHAPTER2. TANGENTSPACESATTORUSFIXEDPOINTS Then we have X0≤i ≤s #Pαi = X0≤i ≤s X0≤k<s−i (νs−i −1 − νs−i ) = X0≤i ≤ j ≤s (νs−i −1 − νs−i ) = X0<i ≤s νs−i = n, since νs = 0. By transposing we prove the same forPi #Qαi . Lemma 2.1.7. Let I = (α0, . . . , αs ) be a monomial ideal of length n with prescribed stan- dard monomial generators. The set B (I ) is a basis for TI Hilbn(C2). Proof. We prove first that the fα,β are linearly independent. LetP(α,β) cα,β fα,β = 0 with cα,β ∈ C, and suppose by contradiction that at least one coefficient is not zero. We can suppose that the left hand side term is of pure weight, otherwise we can study it taking all the terms of given weight. If there is a couple (α, β) with cα,β 6= 0 and β ∈ Pα (resp. 6= 0, we have β′ ∈ Pα′ ( resp. β ∈ Qα), then, for all the other couples (α′, β′) with cα′,β′ β′ ∈ Qα′). Suppose then β ∈ Pα for one couple. Then take α the standard generator with cα,β 6= 0 that is on the highest row, then (P(α,β) cα,β fα,β)(α) 6= 0 since every other generator of I is sent to a lower row. This is absurd as the combination was zero. Now we prove that B (I ) generates TI Hilbn(C2). Since we know that TI Hilbn(C2) is generated by pure weight elements we only need to prove that every f ∈ HomR (I , R/I ) of pure weight is in the span of B (I ). Let n( f ) = # {αi f (αi ) 6= 0} be the number of gener- ators of I that f does not send to 0. We use induction on n( f ). If n( f ) = 0, then f = 0, and there is nothing to prove. Suppose now that all f with n( f ) < t are in the span of B (I ), and suppose f has n( f ) = t . Since f is of pure weight we know that either: 1) we have ¯ı := max{i f (αi ) 6= 0} is strictly smaller than s and f (α¯ı ) ∈ 〈β〉 with β ∈ Pα¯ı , or 2) we have ¯ı := min{i f (αi ) 6= 0} is strictly bigger than 0 and f (α¯ı ) ∈ 〈β〉 with β ∈ Qα¯ı . Suppose we are in the first case. Renormalize f so that f (α¯ı ) = β. Then thanks to the definition of ¯ı, since y p ¯ı β ∈ I , we have that f − fα¯ı ,β is well defined and sends strictly more generators to 0 mod I , i.e. n( f − fα¯ı ) < t . By induction we have that f ∈ span(B (I )) as desired. The case 2), is completely analogous. Proposition 2.1.8. [Fog68, Theorem 2.4] The Hilbert scheme Hilbn(C2) is smooth. Proof. We just proved that at every torus fixed point I the Zariski tangent space has dimension equal to the dimension of Hilbn(C2), Lemma 2.1.6 and Lemma 2.1.7. How- ever since every point in Hilbn(C2) is attracted by a fixed point, and the dimension of the tangent space can only be greater at a fixed point, we have that the same state- ment holds for every point. Thus every point of Hilbn(C2) is smooth. 2.1. AWEIGHTBASISFOR TI HILBN (C2) 25 Theorem 2.1.9. [ES87, Theorem 1.1] The space Hilbn(0) has an affine cell decomposi- tion given by the attracting sets at its torus fixed points for every generic one dimen- sional torus acting with weights 0 < w1 < w2. If we chose the T∞ action of C∗ on C[x, y], i.e. the action with weights w1, w2 such that 0 < w1 < w2 and 1 ≪ w2 , then the affine cell w1 attracted by a monomial ideal I has dimension n − d where d = min{i xi ∈ I } = ℓ(Γ(I )). In particular the Betti numbers for the Borel-Moore homology of Hilbn(0) satisfy: bi = # { ν ⊢ n ℓ(ν) = n − i } . The Poincaré polynomial of Hilbn(0) is Pq¡Hilbn(0)¢ = XΓ⊢n q n−ℓ(Γ). Proof. Consider P2 with homogenous coordinates [Z : X : Y ]. Observe that Hilbn(P2) is smooth since there is an open cover with opens that are isomorphic to Hilbn(C2) that is smooth thanks to Proposition 2.1.8. Consider the T action on it given by t · [Z : X : Y ] = [t W0 Z : t W1 X : t W2], with W0 + W1 + W2 = 0, and W0 < W1 < W2. It has finitely many fixed points, so we can apply Theorem 1.3.4 to obtain a cell decomposition of Hilbn(P2). Consider the subvariety Hilbn(P0) ⊂ Hilbn(P2) parametrizing subschemes that have support in P0 = [1 : 0 : 0]. Since under the T action every point of P2 flows away from P0 we have that Z ∈ Hilbn(P0) ⇐⇒ lim t →0 (t · Z ) ∈ Hilbn(P0). This shows that Hilbn(P0) is a union of some of the cells of the Bialynicki-Birula cell decomposition of Hilbn(P2) and we need only to calculate their dimension to find a graded basis for the BM homology of Hilbn(0). Choose W1 = 2w1−w2 and W2 = 2w2−w1 and use the identifications 3 3 x := X Z y := Y Z R := C[[x, y]], we have Hilbn(P0) = Hilbn(R)red with the torus action described in the statement. In particular T + I Hilbn(0) = T + I Hilbn(C2). Let then I = (α0, . . . , αs ) be a monomial ideal of length n, with prescribed standard monomial generators. We I Hilbn(C2). Given the choice want to study the positive part of the tangent space T + of torus, the weight of fα,β is positive if and only if β ∈ Pα and β lies in a row strictly I Hilbn(P2) = T + higher than the one of α. We know thatPi #Pαi = n from (2.2), and we know that the only β ∈ Pα that are not on a row higher than α are those on the same row of α. It is clear, by projecting down, that these are as many as the boxes in the first row, i.e. the number of columns. Then we have that dim T + I Hilbn(C2) = n − d as wanted. Then thanks to Theorem 1.3.4 we know that Hilbn(0) has an affine cell decom- 26 CHAPTER2. TANGENTSPACESATTORUSFIXEDPOINTS I Hilbn(0). position with cells labeled by monomial ideals and of dimensions dim T + Moreover, thanks to Proposition 1.3.2 we know that the cycles associated to these cells give us a graded basis for the homology of Hilbn(0). Theorem 2.1.10. [Göt90, Theorem 0.1] The generating function for the Poincaré poly- nomials of Hilbn(0) as n varies is +∞Xn=0 Pq¡Hilbn(0)¢ zn =Yk≥1 1 1 − zk q k−1 . (2.3) Proof. Call p(n, k) the number of partitions of n with l parts. Then we know bi (Hilbn(0)) = p(n, n − i ), thanks to Theorem 2.1.9. The formula in the statement then is simply the well known combinatorial identity p(n, n − i )q i zn = +∞Xn=0 +∞Xi =0 1 1 − q k−1zk +∞Yk=1 that follows easily from the famous Euler identity for the generating function of the number of partitions of an integer n. Remark 2.1.11. Goettsche proves a more general formula that holds for every smooth surface X . It is worth mentioning its ground breaking result even though we do not use it. We need to admit non zero odd homology groups to work in this generality, and the formula is: +∞Xn=0 +∞Xj =0 dim H j¡Hilbn(X )¢ q j zn = +∞Yk=1 4Yi =0³1 − zk q 2k−2+i´(−1)dim Hi (X ) . 2.1.1 The tangent space to the Hilbert-Samuel's strata, case n = n Now that we know a weight basis for the tangent space TI Hilbn(C2) at each fixed point I , we can study the tangent spaces TI MT (resp. TI GT ) for those fixed points I ∈ MT i.e such that T (I ) = T . Here we suppose that T = (ti )i ≥0 is an admissible sequence of nonnegative integers as in 1.2.7. The key observation for this study is due to Goettsche and is the following . Observation 2.1.12. [Göt94, Chapter 2] Let I be a monomial ideal with T (I ) = T and standard monomial generators (α0, . . . , αs ). With the choice of weights T1+, we have the natural identifications: TI MT =(cid:173) fα,β¯¯ fα,β preserves or raises the degree® =(cid:173) fα,β¯¯ deg β ≥ deg α® ; TIGT =(cid:173) fα,β¯¯ fα,β preserves the degree® =(cid:173) fα,β ¯¯ deg β = deg α® . 2.1. AWEIGHTBASISFOR TI HILBN (C2) 27 Moreover we can determine the subspace of the tangent space where the weights of the torus action are positive simply as T + T + I MT =(cid:173) fα,β¯¯ fα,β raises the degree, or preserves it and strictly raises it in the y® =D fα,β ¯¯¯ deg β > deg α, or deg β = deg α and degy β > degy αE ; I GT =(cid:173) fα,β¯¯ fα,β preserves the degree and strictly raises it in the y® =D fα,β ¯¯¯ deg β = deg α, degy β > degy αE . Lemma 2.1.13. [Göt94, Lemma 2.2.11] The dimension of the subspace T + where the weights of the action are positive is: I MT of TI MT dim(T + I MT ) = n − #©(u, v ) ∈ Γ(ν)hu,v = 0 or hu,v = 1ª . (2.4) Proof. We give the proof of Goettsche. Call ν = ν0 ≥ · · · > νs = 0 the partition associated to I . Call λ and µ two linearly independent characters of the two torus T2 acting on R as t · x = λ(t )x and t · y = µ(t )y. Then the existence of the basis B (I ) shows that as representation of T2 we have the identity TI Hilbn(C2) = X0≤i ≤ j <s ν j −1Xk=ν j +1 (λi − j −1µνi −k−s + λi − j µk−νi ) (2.5) as each addendum is the weight of one of the fα,β ∈ B (I ). The action of T1+ has positive weight on λa µb if and only if a + b > 0 or a+b=0 and b > 0. Then the term (λi − j −1µνi −k−s) has positive weight if and only if i +νi > j +k +1 and the term (λi − j µk−νi ) has positive weight if and only if i + νi < j + k. Denote with ν the transpose partition of ν, i.e. νi = # {(m, i )(m, i ) ∈ Γ(ν)}. Then it is clear that νk is the smallest j satisfying k ≥ ν j , so that νk − 1 is the smallest j satisfying k ≥ ν j +1. Notice also that we can reformulate the definition of the hook difference in terms of the transpose partition as hu,v (ν) = νu − u − νv + v. Then we have that 0 ≤ j + k − i − νi + 1 ≤ 1)! ν j +1 ≤ k < ν j ; (νi − # {k ∈ Z 0 ≤ k < νi , 0 ≤ νi + i − νk − k ≤ 1 }) dim T + I MT = X0≤i ≤ j <sÃν j − ν j +1 − #(k ∈ Z¯¯¯¯¯ = X0≤i <s = n − #©(u, v ) ∈ Γ(ν)¯¯ 0 ≤ hu,v (ν) ≤ 1ª . 28 CHAPTER2. TANGENTSPACESATTORUSFIXEDPOINTS For the homogenous Hilbert-Samuel's stratum GT Goettsche shows, with similar ar- guments, the following Lemma of which we omit the proof. Lemma 2.1.14. [Göt94, Lemma 2.2.12] The dimension of the subspace T + where the weights of the action are positive is: I GT of TI GT dim(T + I GT ) = n − #©(u, v ) ∈ Γ(ν)hu,v = 1ª . Thanks to smoothness, Proposition 1.2.11, we can apply Theorem 1.3.4 and Proposi- tion 1.3.2 to immediately get the following. Theorem 2.1.15. [Göt94, Theorem 2.2.7] Let T = (ti )i ≥0 be a sequence of nonnegative integers, with T = n as in 1.2.7. Then we have that: (1) The Hilbert-Samuel's strata MT and GT have an affine cell decomposition. (2) The Betti numbers of MT satisfy (3) The Betti numbers of GT satisfy bi (MT ) = #(ν ⊢ n¯¯¯¯¯ bi (GT ) = #(ν ⊢ n¯¯¯¯¯ T (ν) = T #©(u, v ) ∈ Γ(ν) hu,v ∈ {0, 1}ª = n − i) . #©(u, v ) ∈ Γ(ν) hu,v = 1ª = i) . T (ν) = T We can then utilize smoothness of the Hilbert-Samuel's strata and their homogenous counterparts to give their dimensions. Corollary 2.1.16. [Iar77, Theorem 2.12] Let T = (ti )i ≥0 be a sequence of nonnegative integers, with T = n as in 1.2.7. The dimension of MT and GT are: dim MT = n −Xj dimGT = Xj ≥d (t j −1 − t j ) (t j −1 − t j + 1) 2 , (t j −1 − t j + 1)(t j − t j +1). Remark 2.1.17. As an immediate consequence we obtain that the Poincaré polyno- mial for Hilbn(0) can be written as Pq¡Hilbn(0)¢ = XΓ⊢n posT 1+ (Γ) q where posT1+ (Γ) is the quantity described in formula (2.4). Just looking at it combina- torially, it is not immediately clear that this formula coincides with the one we wrote in Theorem 2.1.9, nor that these polynomials can be summed to give the same gen- erating function as in Proposition 2.1.10. In the last chapter we devote a section to explain in some details how this is working. 2.2. AWEIGHTBASISFOR TI1,I2HILBN ,N +1(C2) 29 2.2 A weight basis for TI1,I2Hilbn,n+1(C2) Now that we defined a basis for the tangent space of Hilbn(C2) at each of its fixed points, it is time to do the same for the tangent spaces of Hilbn,n+1(C2) at its fixed points. If I1, I2 is such a fixed point, the vector space we want to understand, thanks to Lemma 2.0.14, is TI1,I2Hilbn,n+1(C2) = Ker(φ12 − ψ12) ⊂ HomR (I1, R±I1 ) ⊕ HomR (I2, R±I2 ). In the previous section we saw how to define a weight basis for HomR (I1, R±I1 ) and for HomR (I2, R±I2 ). The idea is that the two bases share a lot in common, since I1 and I2 differ by a small amount. There is only one monomial in I1 but not in I2, call it α j . In the weight basis B (I1, I2) there will be two types of vectors: The first type un- derlines the similarities between HomR (I1, R±I1 ) and HomR (I2, R±I2 ): more precisely these are vectors of the form ( fα,β, ⋆) with fα,β ∈ B (I1) and ⋆ some appropriate vector in TI2Hilbn+1(C2) that looks like fα,β so that the difference is zero for φ12 − ψ12. The second type underlines the differences between I1 and I2: more precisely these are vectors of the form (0, (α 7→ α j )), where we use the fact that α j is 0 mod I1 and not 0 mod I2. Lets dive into the details. Notation 2.2.1. Let (I1, I2) be a fixed point of the T2 action on Hilbn,n+1(C2), i.e. I1, I2 are monomial ideals of lengths n and n +1 respectively, with I1 ⊃ I2. Call (α1, . . . , αs ) the standard monomial generators of I1 and (α′ s ′ ) the standard monomial genera- tors of I2. Call Γi = Γ(Ii ), then call j the index for which 1, . . . , α′ Γ2 = Γ1 ∪ {α j }. We will denote this configuration also as Γ = (Γ1, Γ2) ⊢ [n, n + 1]. Call pi and qi , as before, the distances between generators of I1, and p ′ i and q ′ i the distances between generators of I2: pi := degy αi +1 − degy αi i +1 − degy α′ p ′ i := degy α′ i ps := ∞, p ′ s ′ := ∞, qi := degx αi − degx αi −1 q ′ i := degx α′ i − degx α′ i −1 q0 := ∞, q ′ 0 := ∞. Analogously we call Pα = Pαi and Qα = Qαi the relevant sets for α = αi a generator of I1, and Pα′ = Pα′ i a generator of I2. Whenever confusion is possible we will clarify if fα,β is seen as an element in the basis B (I1) of TI1Hilbn(C2) or as an element in the basis B (I2) of TI2Hilbn+1(C2), and so on. the relevant sets for α′ = α′ and Qα′ = Qα′ i i 30 CHAPTER2. TANGENTSPACESATTORUSFIXEDPOINTS Definition 2.2.2 (Cases). Let I1 = (α1, . . . , αs ), I2 = (α′ s ′ ) be ideals with prescribed standard monomial generators, such that (I1, I2) is a fixed point for Hilbn,n+1(C2), and such that Γ(I2) = Γ(I1) ∪ {α j }. There are four possible different cases we need to distin- guish. In the following pictures we will indicate with a bullet • those standard mono- mial generators of I2 that are not already standard monomial generators of I1. 1, . . . , α′ Cases 1a), 1b) (s = s ′) The number of generators s + 1 and s′ + 1 is the same for I1 and I2. This can only hap- pen if p j = 1 or q j = 1 but not both. Then we distinguish the two possibilities, and we look at the generators of I2 in terms of those of I1: Case 1a) , q j = 1 α′ j = yα j α′ 1 = α1 . . . . . . α′ s ′ = α′ s = αs . • α j Case 2) (s ′ = s + 1) Case 1b) , p j = 1 α′ 1 = α1 . . . α′ j = xα j . . . α′ s ′ = α′ s = αs .  α j •   This happens if and only if p j > 1, q j > 1, or j = 0 and p0 > 1, or j = s and qs > 1. Then we have Case 2) , p j , q j > 1 α′ 1 = α1 . . . = xα j α′ j α′ j +1 = yα j j +2 = α j +1 α′ . . . α′ s ′ = α′ s+1 = αs . • α j • Case 3 (s ′ = s − 1) This happens if and only if p j = 1, q j = 1. Then we have 2.2. AWEIGHTBASISFOR TI1,I2HILBN ,N +1(C2) 31 α′ 1 = α1 . . . j −1 = α j −1 = α j +1 α′ α′ j α′ j +1 = α j +2 . . . α′ s ′ = α′ s−1 = αs . α j  Case 3) , p j , q j = 1 We now focus on the task of understanding for which fα,β ∈ B (I ) there exists h ∈ HomR (I2, R±I2 ) such that ( fα,β, h) ∈ TI1,I2Hilbn,n+1(C2), i.e. φ12( fα,β) − ψ12(h) = 0 mod I1. Such an h does not always exist, as the following example shows. Example 2.2.3. Let the following diagram represents the fixed point of Hilb3,4(C2) with I1 = (x2, x y, y 2), I2 = (x2, x y, y 3), and α j = y 2. Let fα,β be the map fx2,y ∈ B (I ). Then there does not exist a map h ∈ HomR (I2, R±I2 ) such that φ12( fα,β) − ψ12(h) = 0, because such a map would send x2 7→ y and x2 y 7→ y 2 6= 0 mod I2 which is absurd since it should also be x y 7→ 0 =⇒ x2 y 7→ x · 0 = 0. y 3 α j x y x2 y x2 If f (x2) = y then h(x2) = y, but then h(x y) cannot be defined as we discussed before. 1, . . . , α′ s ′ ) be a fixed point for Hilbn,n+1(C2), Definition 2.2.4. Let I1 = (α1, . . . , αs ), I2 = (α′ such that Γ(I2) = Γ(I1)∪{α j }, with prescribed standard monomial generators. We define a subset of indexes (α, β) of the basis B (I1) =© fα,β(α, β)ª and we denote it Obs(I1, I2). We will see that these correspond to those fα,β ∈ B (I1) that we do not want to try to extend to elements ( fα,β, h) ∈ TI1,I2Hilbn,n+1(C2). The definition of Obs(I1, I2) varies according to the possible cases of Definition 2.2.2. Obs(I1, I2) :=( (αi , Case 1a) α j x qi ) for i > j , α j y pi ) for i < j − 1.) (αi , Obs(I1, I2) :=((αi , Case 2) α j x qi ) for i > j , α j y pi ) for i < j.) (αi , Case 1b) Obs(I1, I2) :=((αi , (αi , α j x qi ) for i > j + 1, y pi ) for i < j. ) α j Obs(I1, I2) :=((αi , Case 3) α j x qi ) for i > j + 1, α j y pi ) for i < j − 1.) (αi , (2.6) 32 CHAPTER2. TANGENTSPACESATTORUSFIXEDPOINTS Example 2.2.5. The pictures below represent some of elements of Obs that we de- fined. For each picture there are two fα,β represented: the one represented with stars sends the generator of I1 marked with a star to the element of Γ1 marked with a star. The other is represented in a similar way with two bullets. Case 1a) • • α j ⋆ ⋆ Case 2) • • α j ⋆ ⋆ Case 1b) • • α j ⋆ ⋆ Case 3) • • α j ⋆ ⋆ Observation 2.2.6. Observe that in terms of s′ = s2, the highest index of a standard generator of I2, the set Obs(I1, I2) has the same number of elements for all cases. Case 1a) Case 1b) #Obs(I1, I2) = s1 + 1 − 2 = s2 + 1 − 2. #Obs(I1, I2) = s1 + 1 − 2 = s2 + 1 − 2. Case 2) Case 3) #Obs(I1, I2) = s1 + 1 − 1 = s2 + 1 − 2. #Obs(I1, I2) = s1 + 1 − 3 = s2 + 1 − 2. Lemma 2.2.7 (definition). Let I1 = (α1, . . . , αs ), I2 = (α′ s ′ ) be ideals with prescribed standard monomial generators, such that (I1, I2) is a fixed point for Hilbn,n+1(C2), and such that Γ(I2) = Γ(I1)∪{α j }. Let fα,β be one of the elements of the basis B (I1) of TI1Hilbn(C2) such that (α, β) ∉ Obs(I1, I2). Then we define 1, . . . , α′ Suiv( fα,β)(α′ k ) = fα,β(α′ k ), for all k = 0, . . . , s′. (2.7) Then Suiv( fα,β) is well defined as an element of HomR (I2, R±I2 ). Moreover it is such that ¡ fα,β,Suiv( fα,β)¢ ∈ TI1,I2Hilbn,n+1(C2) . Proof. Suppose α 6= α j , then α is also a generator of I2. Moreover the hypothesis that (α, β) is not an element of Obs(I1, I2) ensures that β ∈ P ′ α so that we can define α ∪ Q ′ 2.2. AWEIGHTBASISFOR TI1,I2HILBN ,N +1(C2) 33 f ′ α,β ∈ B (I2) and we have that f ′ α,β = Suiv( fα,β). Of course φ12( fα,β) − ψ12(Suiv( fα,β)) = 0 since it is zero on each generator α′ k of I2. Similarly if α = α j but xα (or yα) is a gener- ators of I2, we have that Suiv( fα,β) = f ′ ∈ B (I2)) so that it is well defined and it is, again obviously, in the kernel of φ12 − ψ12. If, finally, α = α j β x p j −1 ∈ Pα′ and and we are in case 3) of 2.2.2 then, if β ∈ Qα, y we have that β x q j +1 ∈ Qα′ and, if β ∈ Pα, x ∈ B (I2) (or Suiv( fα,β) = f ′ yα,yβ xα,xβ If β ∈ Qα, then Suiv( fα,β) = f ′ α′ j ,y , β q j +1 x If β ∈ Pα, then Suiv( fα,β) = f ′ α′ j −1,x . β p j −1 x This shows that also in this case Suiv( fα,β) is well defined. Again it is clear that the vector ( fα,β,Suiv( fα,β)) is in Ker(φ12 − ψ12). Definition 2.2.8. For every generator α′ i of I2 define: hα′ i ,α j ∈ HomR (I2, R±I2 ) as hα′ i ,α j (α′ 0 α j if k 6= i , if k = i . k ) = Observation 2.2.9. The hα′ i ,α j as above are well defined. Moreover (0, hαi ,α j ) ∈ Ker(φ12 − ψ12) are linearly independent vectors, as it follows recalling that α j 0 mod I1. ∼= 0 mod I1 but α j ≇ s ′ ) be a fixed point for Hilbn,n+1(C2), Definition 2.2.10. Let I1 = (α1, . . . , αs ), I2 = (α′ such that Γ(I2) = Γ(I1) ∪ {α j }, with prescribed standard monomial generators. Then we define the set: 1, . . . , α′ B (I1, I2) :=n³0, hα′ i ,α j´¯¯ i = 0, . . . , s′o ∪ ⊂ HomR (I1, R±I1 ) ⊕ HomR (I2, R±I2 ). ∪©¡ fα,β,Suiv( fα,β)¢¯¯ fα,β ∈ B (I1), (α, β) ∉ Obs(I1, I2))ª (2.8) Observation 2.2.11. Observe that # B (I1, I2) = 2n + 2. In fact #B (I1) = 2n, #Obs(I1, I2) = s′ + 1 − 2 thanks to Observation 2.2.6, and #n(0, hα′ i ,α j )i = 0, . . . , s′o = s′ + 1. Lemma 2.2.12. With the notations as in Definition 2.2.10, we have that B (I1, I2) is a basis for TI1,I2Hilbn,n+1(C2). Proof. First of all we observe that by definition all elements of B (I1, I2) are in the Ker(φ12− ψ12), and are of pure weights. Abbreviate Obs(I1, I2) to Obs. 34 CHAPTER2. TANGENTSPACESATTORUSFIXEDPOINTS Let us now prove that they are linearly independent. Suppose s ′Xi =0 ci (0, hα′ i ,α j ) + X(α,β)∉Obs cα,β ( fα,β,Suiv( fα,β)) = (0, 0). i ,α j 0, . . . , s′. i =0 ci hα′ know that the vectors fα,β are linearly independent, so that cα,β = 0 for each (α, β). Then in particular P(α,β)∉Obs cα,β fα,β = 0 as an element of HomR (I1, R±I1 ) where we ∈ HomR (I2, R±I2 ), which again implies that ci = 0 for all i = Then, also Ps ′ Let us then prove that B (I1, I2) generates TI1,I2Hilbn,n+1(C2). Since we know that TI1,I2Hilbn,n+1(C2) is generated by pure weight elements we only need to prove that every g = ( f , h) ∈ Ker(φ12 − ψ12) of pure weight is in the span of B (I1, I2). Let n( f ) be the number of generators of I1 that the first coordinate of g does not send to 0 i.e n( f ) = # {αi f (αi ) 6= 0 }. Then we use induction on n( f ). i ) ∈ 〈α j 〉. Then, clearly h ∈ 〈hα′ If n( f ) = 0, then f = 0, and then g = (0, h) is such that h(α′ i ) = 0 mod I1 so that i = 0, . . . s′〉. Suppose now that all g = ( f , h) with h(α′ n( f ) < t are in the span of B (I1, I2), and suppose g = ( f , h) with n( f ) = t . Since f is of pure weight we know that either: i ,α j 1) we have ¯ı := max{i f (αi ) 6= 0} is strictly smaller than s and f (α¯ı ) ∈ 〈β〉 with β ∈ Pα¯ı , or 2) we have ¯ı := min{i f (αi ) 6= 0} is strictly bigger than 0 and f (α¯ı ) ∈ 〈β〉 with β ∈ Qα¯ı . Suppose we are in the first case. Renormalize g = ( f , h) so that f (α¯ı ) = β. First of all notice that (α¯ı , β) ∉ Obs, otherwise we would be in the situation where ¯ı ≤ j and β = α j y p ¯ı with f (αi ) = 0 for all i > ¯ı. But all θ ∈ HomR (I2, R±I2 ) such that θ(α¯ı ) = 0, if not α j y p¯ı have θ(α¯ı +1) 6= y p ¯ı α¯ı = x q¯ı +1α¯ı +1 =⇒ α j = y p ¯ı α j y p ¯ı = θ(y p ¯ı α¯ı ) = θ(x q¯ı +1 α¯ı +1) = 0. This equality mod I2 is absurd since α j ∉ I2. Then, always thanks to the definition of ¯ı, we have that y p ¯ı β ∈ I1, and thanks to the fact that (α¯ı , β) ∉ Obs, we have that ( f − fα¯ı ,β, h − Suiv( fα¯ı ,β)) ∈ Ker(φ12 − ψ12) is well defined and has first coordinate that sends strictly more generators to 0 mod I1, i.e. n( f − fα¯ı ) < t . By induction we have that g ∈ span(B (I1, I2)) as desired. 2.2. AWEIGHTBASISFOR TI1,I2HILBN ,N +1(C2) 35 The case 2) is completely analogous. Theorem 2.2.13. [Che98, Theorem 3.2.2, 3.3.3] The space Hilbn,n+1(C2) is smooth. The space Hilbn,n+1(0) has an affine cell decomposition given by the attracting sets at torus fixed points. Moreover if the T1 action on R has weights 0 < w1 < w2 and 1 ≪ w2 then w1 the affine attracting cell in Hilbn,n+1(0) of the point (I1, I2) has dimension n +1−d where d = min{i xi ∈ I2} i.e. d = ℓ(Γ(I2)). In particular the Betti numbers of Hilbn,n+1(0) satisfy: bi = #©(Γ1, Γ2)¯¯Γ1 ⊢ n, Γ2 ⊢ n + 1, Γ1 ⊂ Γ2 and ℓ(Γ2) = n + 1 − iª . Pq³Hilbn,n+1(0)´ = X(Γ1,Γ2) ⊢ [n,n+1] The Poincaré polynomial of Hilbn,n+1(0) is q n+1−ℓ(Γ2). The generating function for these polynomials is +∞Xn=0 Pq³Hilbn,n+1(0)´ zn =µ 1 1 − zq¶ Yk≥1 1 1 − zk q k−1 . (2.9) Proof. The proof of smoothness follows immediately from Observation 2.2.11 and Lemma 2.2.12 that tell us that the dimension of the tangent space at each torus fixed point is the same as the dimension of the variety. The proof on the Betti numbers goes exactly as the proof of Theorem 2.1.9. First of all Hilbn,n+1(P2) is smooth since Hilbn,n+1(C2) is. Then Theorem 1.3.4 gives us a cell decomposition with the dimension of the cells specified by the positive part of the tangent spaces at fixed points. Call P0 ∈ P2 the fixed points with zero dimen- sional attracting cells for the T action on P2; since the limiting process preserves the support for subschemes with support concentrated in P0 we have that Hilbn,n+1(0) = Hilbn,n+1(P0) is union of cells of Hilbn,n+1(P2). We do not include the proof of the statement on the generating function. Cheah [Che98, pp 69-70] proves it directly in one page. In [NY08, Chapter 5] the authors give a slightly more indirect and more geometrical proof whose ingredients are more sim- ilar to the discussion we will give in the last section of Chapter 4. 2.2.1 The tangent space to the Hilbert-Samuel's strata, case n = (n, n + 1) For T1, T2 two admissible sequences of nonnegative integers as in 1.2.7, we want to study the tangent space of MT1,T2 ( resp. of GT1,T2 ) at a flag of monomial ideals (I1, I2) with T (Ii ) = Ti . Thanks to Observation 2.1.12 we can reduce this to the study of the weights of the elements of the basis B (I1, I2). Moreover to understand the cellular decomposi- tion of MT1,T2 induced by the action of the torus T1+ we are interested in studying the 36 CHAPTER2. TANGENTSPACESATTORUSFIXEDPOINTS positive parts of these tangent spaces. We divide the study of elements of B (I1, I2) into two observations, according to their type. Observation 2.2.14. To study the weight of the first kind of elements (0, hα′ i ,α j ), with i = 0, . . . , s′ we need only to compare the degree of a generator of I2 with that of α j the only i ,α j ) ∈ B (I1, I2) is monomial in I1 but not in I2. More precisely, we have that the (0, hα′ tangent to MT (I1),T (I2) if and only if deg α′ i ≤ deg α j ; tangent to GT (I1),T (I2) if and only if deg α′ i ≤ deg α j and degy α′ i < degy α j ; tangent to GT (I1),T (I2) and with positive weight if and only if deg α′ i i = deg α j ; tangent to MT (I1),T (I2) and with positive weight if and only if deg α′ = deg α j and degy α′ i < degy α j . Example 2.2.15. m, g m, g αj m+, g + m+ With m (resp. g ) we indicate a generator of I2 that produces a tangent vector to MT (I1),T (I2)( resp. GT (I1),T (I2)) The plus is added if the vector produced has positive weight. Observation 2.2.16. We want now to analyze the second type of elements of the basis B (I1, I2): ©¡ fα,β,Suiv( fα,β)¢¯¯ fα,β ∈ B (I1), (α, β) ∉ Obs(I1, I2))ª . From the definition of Suiv( fα,β) we have that ( fα,β,Suiv( fα,β)) has the same weight as fα,β. So ( fα,β,Suiv( fα,β)) ∈ B (I1, I2) is tangent to MT (I1),T (I2) if and only if fα,β is tangent to MT (I1) i.e. if and only if deg α ≤ deg β. Similarly for the positive part, and for the tangent to GT (I1),T (I2) and its positive part. In fact it is enough to understand when fα,β with (α, β) ∈ Obs(I1, I2) is raising or preserving degree, and so on, and then understand the dimension of the appropriate vector spaces by difference: if (α, β) ∈ Obs(I1, I2), fα,β does not extend to an element of B (I1, I2). Looking at definition 2.2.4 we have that f = fα,β with (α, β) ∈ Obs(I1, I2) is such that either case1) f (αi ) = α j y pi with i < j or 2.2. AWEIGHTBASISFOR TI1,I2HILBN ,N +1(C2) 37 case2) f (αi ) = α j x qi with i > j . α j y pi − deg αi = deg α j − y pi deg αi . The α j x qi − In case1) we have the weight of fα,β is equal to deg same formula holds for the y-degree. In case 2), similarly, we have: wt( fα,β) = deg deg αi = deg α j − x qi deg αi . Notice also that for each αi generator of I1 there is at most one β such that (αi , β) ∈ Obs(I1, I2). Then to know for which αi , i = 0, . . . , s there is a β such that fα,β ∈ TI1Hilbn(C2) does not produce an element of the basis B (I1, I2) of TI1,I2Hilbn,n+1(C2) but that is, say, raising degree (i.e. fα,β is tangent to MT (I1)), the relevant check on degrees is be- tween α j and y pi αi if i ≤ j and between α j and x qi αi if i ≥ j . Notice that we do not need to distinguish cases as in the definition of Obs(I1, I2) 2.2.4, since for the indexes i = j − 1, j , j + 1, where there is actually a difference in the definition of Obs(I1, I2), we have deg α j < deg y pi αi or deg α j < deg x qi αi so that we would not consider them any- way, since the corresponding fα,β is of negative weight. We can then summarize and say fαi ,β ∈ TI1 MT (I1) but (αi , β) ∈ Obs(I1, I2) ⇐⇒ Remark once more that these are the vectors we will need to exclude. Similarly for all the other vector spaces we are interested in. Let us visualize it in an example: deg y pi αi ≤ deg α j deg x qi αi ≤ deg α j if i ≤ j , if i ≥ j . (2.10) α j m+g + m+ With m (resp. g ) we indicate y p times a generator of I1 that produce a tangent vector to MT1 (resp. GT1) that cannot be extended to a tangent vector of MT1,T2 (resp. GT1,T2). The plus as exponent is there if the tangent to MT1 (resp. GT1) is also attracting. Definition 2.2.17. We put together what we remarked in the last two observations to give a combinatorial rule of how the dimension of the tangent space at the two step flag case grows with respect to the one step flag case. Let Γ2 = Γ1 ∪{α j } be the Young diagrams associated to monomial ideals (I1, I2) in Hilbn,n+1(C2). Then put a • at each box occupied by a standard monomial generators of I2, and a ⋆ at each vertex of Γ2, i.e. wherever on the same column below and on the same row on the left there is a •, and no others marked boxes in between. 38 CHAPTER2. TANGENTSPACESATTORUSFIXEDPOINTS • ⋆ • ⋆ • ⋆ • ⋆ • ⋆ α j • ⋆ • ⋆ • Define Λi the i -th antidiagonal of the positive quadrant, i.e. all boxes (k, m) with k + m = i for i ≥ 0. Say that Λi < Λk if i < k, i.e. if Λi is a lower antidiagonal than Λk. Call Λα j the antidiagonal that passes through α j . Then define the following numbers: • G(Γ1, Γ2) := # {• ∈ Λα j } • M (Γ1, Γ2) := # {• ∈ Λi¯¯Λi < Λα j } + # {• ∈ Λα j } • M +(Γ1, Γ2) := # {• ∈ Λi¯¯Λi < Λα j } + # {• ∈ Λα j¯¯• to the right of α j } •G +(Γ1, Γ2) := # {• ∈ Λα j¯¯• to the right of α j } ⋆ M (Γ1, Γ2) := # {⋆ ∈ Λi¯¯Λi < Λα j } + # {⋆ ∈ Λα j } ⋆ M +(Γ1, Γ2) := # {⋆ ∈ Λi¯¯Λi < Λα j } + # {⋆ ∈ Λα j¯¯⋆ to the right of α j } ⋆G +(Γ1, Γ2) := # {⋆ ∈ Λα j¯¯⋆ to the right of α j }. ⋆ G(Γ1, Γ2) := # {⋆ ∈ Λα j } Lemma 2.2.18. [Che98, Lemma 3.4.9] Suppose that (I1, I2) ∈ GT1,T2 ⊂ MT1,T2 is a fixed point of the T1+ action on Hilbn,n+1(0) with generic weights w1 < w2 and (n +1)w1 > nw2. Call (Γ1, Γ2) the couple of Young diagrams associated to (I1, I2). Then: 1) The dimension of the tangent space to MT1,T2 at (I1, I2) is equal to dim MT1 + • M (Γ1, Γ2) − ⋆ M (Γ1, Γ2). 2) The dimension of the tangent space to GT1,T2 at (I1, I2) is equal to dimGT1 + •G(Γ1, Γ2) − ⋆G(Γ1, Γ2). 3) The dimension of the positive part of tangent space to MT1,T2 at (I1, I2) is equal to dim T + I1 MT1 + •M +(Γ1, Γ2) − ⋆M +(Γ1, Γ2). 4) The dimension of the positive part of the tangent space to GT1,T2 at (I1, I2) is equal 2.3. AWEIGHTBASISFOR TI1,I2,I3HILBN ,N +1,N +2(C2) 39 to dim T + I1 GT1 + •G +(Γ1, Γ2) − ⋆G +(Γ1, Γ2). Proof. The proof is immediate thanks to the definition of B (I1, I2) 2.2.10 and Observa- tions 2.2.14 and 2.2.16. In the next chapter we will see that these strata are smooth, and we will talk about their dimension and about their homology. 2.3 A weight basis for TI1,I2,I3Hilbn,n+1,n+2(C2) If (I1, I2, I3) is a fixed point of Hilbn,n+1,n+2(C2) we denote by α j the only mono- l the only monomial in I2 but not in I3. We need to mial in I1 but not in I2, and by α′ find a weight basis for Trying to mimic what we did for the two step flag case, we want to find elements in TI1,I2,I3Hilbn,n+1,n+2(C2) ∼= \1≤i < j ≤3¡Ker(φi j − ψi j ) ◦ πi j¢ . HomR (I1, R±I1 ) ⊕ HomR (I2, R±I2 ) ⊕ HomR (I3, R±I3 ) of the form (0, 0, (α 7→ α′ l )), (0, (α 7→ α j ),Suiv((α 7→ α j )), ( fα,β,Suiv( fα,β),Suiv(Suiv( fα,β))). As before, then, most of the work is to understand for what elements we can actually define the Suiv. For example, we know already that for the last kind of vectors we need to take fα,β ∈ B (I1) with (α, β) ∉ Obs(I1, I2). Presumably we need to take care about the problem of extending vectors from B (I2) to B (I2, I3). For this reason we will define 2.3.6. In fact the real twist of the three flag case is Definition 2.3.9: this tackles the problem of making sure that we pick out those vectors of B (I2) that do not extend to vectors in B (I2, I3) but that actually matter for the construction of B (I1, I2, I3). Philo- sophically we are trying to define B (I1, I2, I3) by using twice, and independently, the two flag case and then slightly correcting with some information that is intrinsic to the three flag case. Notation 2.3.1. Let (I1, I2, I3) be a fixed point of Hilbn,n+1,n+2(C2), i.e. each Ii is a mono- mial ideal with dim R±Ii = n + i − 1 and I1 ⊃ I2 ⊃ I3. We assume that: I1 = (α0, . . . , αs ), I2 = (α′ 0, . . . , α′ Γ(I3) = Γ(I2) ∪ {α′ I3 = (α′′ 0 , . . . , α′′ s ′ ), l } = Γ(I1) ∪ {α j } ∪ {α′ s ′′ ), l }, 40 CHAPTER2. TANGENTSPACESATTORUSFIXEDPOINTS where the αi s are standard monomial generators of I1, the α′ i s are standard monomial i s are standard monomial generators of I3 and there are s′′ + 1 = generators of I2, the α′′ s3 + 1 of them. In the first step we add the box α j to Γ(I1) that was a generator of I1. In the second step we add α′ l to Γ(I2) that was a generator of I2. Definition 2.3.2. We distinguish two cases based on the relative position of α j and α′ l . The corresponding fixed points will have different geometrical properties. case a) The second box we add, α′ l , is a minimal generator also of I1. This is always the case unless α′ l is imme- diately above or on the right of α j . In this case we often drop the prime and call the second box sim- ply αl . case b) The second box we add, α′ ator of I1. This can happen only if α′ above or to the right of α j , i.e. α′ l We will keep the prime in α′ l , is not a minimal gener- l is immediately = xα j . = yα j or α′ l l in this case. αl α j αl α j Notation 2.3.3. Call pi and qi , as before, the distances between generators of I1, and p ′ i and q ′ i the distances between generators of I3: i the distances between generators of I2, and p ′′ i and q ′′ pi := degy αi +1 − degy αi i +1 − degy α′ p ′ i := degy α′ p ′′ i := degy α′′ i +1 − degy α′′ i i ps := ∞, p ′ s ′ := ∞, p ′′ s ′′ := ∞, qi := degx αi − degx αi −1 i := degx α′ q ′ q ′′ i := degx α′′ i − degx α′ i −1 i − degx α′′ i −1 q0 := ∞, q ′ 0 := ∞, q ′′ 0 := ∞. i and Qα′ = Qα′ Analogously we call Pα = Pαi and Qα = Qαi the relevant sets for α = αi generator of I1, those Pα′ = Pα′ for the generators of I3. Whenever confusion is possible we will clarify if fα,β is seen as an element in the basis B (I1) of TI1Hilbn(C2) or as an element in the basis B (I2) of TI2Hilbn+1(C2), and so on. those for the generators of I2 and Pα′′ = Pα′′ and Qα′′ = Qα′′ i i i Definition 2.3.4. We need to distinguish between the possible cases of Definition 2.2.2 on s, s′ and s′′. To give a name to all possible cases we will combine the name of the case as in 2.2.2 and the name of the boxes it refers to. For example the case l 1a) means that, we do not care about the position of α j , and when we add α′ l we add it as a minimal generator of I2 that has p ′ > 1 and q ′ = 1. Another example: j 1a), l 3) means that p j ≥ 2, q j = 1 and p ′ l = 1 = q ′ l . 2.3. AWEIGHTBASISFOR TI1,I2,I3HILBN ,N +1,N +2(C2) 41 It is clear that, between these cases and the cases for the relative position of α j and αl of 2.3.2, we have quite a few possible cases. Example 2.3.5. Before introducing Obs(I2, I3), we show with an example why we need to care about extending vectors from TI2Hilbn+1(C2) to TI2,I3Hilbn+1,n+2(C2) in order to define vectors of the third type¡ fα,β,Suiv( fα,β)),Suiv(Suiv( fα,β)¢. αl • α j • Let fα,β ∈ B(I1) be the map depicted by the bullets, i.e. that sends the bullet outside Γ1 to the bullet in- side and is zero on the other generators of I1. Then ( fα,β,Suiv( fα,β)) ∈ B(I1, I2), but there does not exist g ∈ TI3 Hilbn+2(C2) such that ( fα,β,Suiv( fα,β), g ) is tangent to TI1,I2,I3 Hilbn,n+1,n+2(C2), simply because there does not exist a map g ∈ TI3 Hilbn+2(C2) that sends • to •. Definition 2.3.6. We need to define Obs(I1, I2) and Obs(I2, I3). We define them only according to the respective cases of the last box added, and not on the relative position of α j and αl . Thus, for example, Obs(I2, I3) depends only on p ′ l and not on p j and q j . l and q ′ For Obs(I1, I2), then, the definition is exactly as in 2.6. The definition of Obs(I2, I3) is completely analogous but we add it for convenience. Case l 1a) i , αl i , αl x qi ) for i > l Obs(I2, I3) :=( (α′ Obs(I2, I3) :=( (α′ (α′ Case l 2) i , αl i , αl y pi ) for i < l − 1.) y pi ) for i < l .) x qi ) for i > l (α′ Case l 1b) Obs(I2, I3) :=((α′ Obs(I2, I3) :=( (α′ i , αl x qi ) for i > l + 1 i , αl (α′ Case l 3) i , αl i , αl y pi ) for i < l . ) y pi ) for i < l − 1.) x qi ) for i > l + 1 (α′ Observation 2.3.7. Again observe that in terms of s′′ = s3, where s′′ + 1 is the number of generators of I3, the set Obs(I2, I3) has the same number of elements along all cases. Case l 1a) Case l 1b) #Obs(I1, I2) = s2 − 2 = s3 + 1 − 2 #Obs(I2, I3) = s2 − 2 = s3 + 1 − 2 Case l 2) Case l 3) #Obs(I2, I3) = s2 − 1 = s3 + 1 − 2 #Obs(I2, I3) = s2 − 3 = s3 + 1 − 2 However, as suggested at the beggining of the section, not all elements of Obs(I2, I3) are relevant. The next example shows why this can happen. Example 2.3.8. 42 CHAPTER2. TANGENTSPACESATTORUSFIXEDPOINTS αl • α j • Let f•,• ∈ B (I2) be the map depicted by the bullets, i.e. that sends the bullet outside Γ2 to the bullet in- side and is zero on other generators of I2 that are above the bullet outside (but is not zero on the re- maining generators of I2, α′ 1). Notice that p ′ • = 2. Then (α, β) ∈ Obs(I2, I3), but there does not exist f ∈ TI1Hilbn(C2) such that fα,β = Suiv( f ) since p• = 1. Then the couple (α, β) in B (I2, I3) is actually irrelevant to our purposes. 0 and α′ To avoid counting unnecessarily obstructions at the second step we introduce this definition. Definition 2.3.9. We define a set of indexes NotP = NotP(Γ1, Γ2, Γ3) := {(α′ be only one of these, and it will be such that fα′ corresponding elementsn fα′ R-homomorphism in HomR (I1, R±I1 ) with Suiv( f ) = fα′ tween the two cases a) and b) of 2.3.2. i ,βo are in HomR (I2, R±I2 ). In the general case there will i ,β ∈ Obs(I2, I3) but there is not f an i ,β. We need to distinguish be- i , β)} whose αl not immediately above or to the right of α j case a) The definition is according to the cases 2.2.2 for α j . Case j 1a) y p′ If ( j < l ) then NotP :=½µα j −1, αl j −1¶¾ If ( j > l ) then NotP :=©¡yα j , αl x¢ª . If ( j < l ) then NotP :=n³xα j , αl y´o x¢ª . If ( j > l ) then NotP :=©¡yα j , αl Case j 2) Case j 1b) y ´o If ( j < l ) then NotP :=n³xα j , αl If ( j > l ) then NotP :=½µyα j +1, αl j +1¶¾ . j −1¶¾ If ( j < l ) then NotP :=½µα j −1, αl If ( j > l ) then NotP :=½µα j +1, αl j +1¶¾ . Case j 3) p′ q′ q′ x x y case b) α′ l is immediately above or to the right of α j . If α′ l = yα j we define If α′ l = xα j we define NotP :=©(α′ i , α j )¯¯i < j , and p ′ i = 1ª . Observation 2.3.10. In case a), i.e. αl not immediately above or on the right of α j , there is always only a single element in NotP. NotP :=©(α′ i , α j )¯¯i > j , and q ′ i = 1ª . Example 2.3.11. In the example below we depict four possible cases of nested triples of monomial ideals, and the corresponding NotP ∈ HomR (I2, R±I2 ), which is just a sin- gleton. As usual we represent an R-homomorphism with a couple of corresponding 2.3. AWEIGHTBASISFOR TI1,I2,I3HILBN ,N +1,N +2(C2) 43 symbols, the one outside Γ2 represents the generator of I2, the one inside represents its image. • α j ⋆ α j ⋆ • αl ⋄ △ α j △ α j ⋄ Lemma 2.3.12 (definition). Let (α′, β′) ∈ Obs(I2, I3) be such that (α′, β′) ∉ NotP. Then there exists fα,β ∈ B (I1) such that ( fα,β, fα′,β′) ∈ Ker(φ12 − ψ12) and (α′, β′) = (x a y bα, x a y bβ) with (a, b) = (0, 0), (0, 1) or (1, 0). We define fα,β := Prec( fα′,β′) and PObs(I1, I2, I3) = PObs =©(α, β)¯¯ fα,β = Prec( fα′,β′ ) for (α′, β′) ∈ Obs(I2, I3) \ NotPª Proof. It is enough to notice that, under the hypothesis on (α′, β′), either α is a minimal β generator of I1 (or one of α x a y b is in P α xa yb x a y b with (a, b) = (0, 1), (1, 0) is), and β ∈ Pα ∪ Qα (or ∪Q α xa yb ). that ( f , h) ∈ Ker(φ12 − ψ12) and (h, g ) ∈ Ker(φ23 − ψ23), then ( f , g ) ∈ Ker(φ13 − ψ13), so that Lemma 2.3.13. Let ( f , h, g ) ∈ HomR (I1, R±I1 ) ⊕ HomR (I2, R±I2 ) ⊕ HomR (I3, R±I3 ) be such ( f , h, g ) ∈T1≤i < j ≤3¡Ker(φi j − ψi j ) ◦ πi j¢. Proof. It is enough to check that − hI3 ) + ( hI3 − g ) = 0 mod I1 ( f¯¯I3 since the two terms in parenthesis are 0: the first directly by hypothesis, the second because by hypothesis is 0 mod I2 and I1 ⊃ I2. 44 CHAPTER2. TANGENTSPACESATTORUSFIXEDPOINTS Lemma 2.3.14 (definition). Let (α, β) with fα,β ∈ B (I1) and (α, β) ∉ Obs(I1, I2) ∪ PObs. Then we can define ¡ fα,β,Suiv( fα,β),Suiv(Suiv( fα,β))¢ ∈ \1≤i < j ≤3¡Ker(φi j − ψi j ) ◦ πi j¢ . Proof. The proof is completely analogous to the proof of Lemma 2.2.7. Now that we have dealt with vectors of type¡ fα,β,Suiv( fα,β),Suiv(Suiv( fα,β))¢, we will deal with vectors of type¡0, (α 7→ α j ),Suiv(α 7→ α j )¢. The situation is easier: if we are in case a) of 2.3.2 Suiv(α 7→ α j ) always exists, and the definition is as below. In case b) we still need some care as it might happen that (α′ i , α j ) ∈ Obs(I2, I3). Definition 2.3.15. Let (α′ 0, . . . , α′ s ′ ) be the list of standard monomial generators of I2. case a) αl is not immediately above or to the right of αj . ∈ B (I2) be the vector already defined in (2.2.8), i.e. the map that takes α′ Let hα′ to α j ∈ Γ2 and sends all other generators of I2 to zero. Then we define Suiv(hα′ i ,α j i ∈ I2 i ,α j ) ∈ HomR (I3, R±I3 ) by specifying the image of each generator of I3 as: Suiv(hα′ i ,α j )(α′′ hα′ i ,α j (α′′ k ) if α′′ k = x a y bα′ i , with a, b ≥ 0, 0 otherwise. This works for all α′ clearly linearly independent. i exactly because (α′ i , α j ) ∉ Obs(I2, I3). The hα′ i ,α j we just defined are case b) α′ l is immediately above or to the right of αj . In this case we define hα′ Then, if (α′ through Lemma 2.2.7. i , α j ) ∉ Obs(I2, I3), we also define Suiv(hα′ i ,α j := fα′ i ,α j ∈ B (I2), i.e. through the usual Definition 2.1.3. i ,α j ) with the usual definition, i.e. 0 , . . . , α′′ s ′′ ) be the list of standard monomial generators of I3. k ) = Definition 2.3.16. Let (α′′ We define hα′′ l i ,α′ ∈ HomR (I3, R±I3 ) as k ) = = α′′ i , otherwise. if α′′ k α′ l 0 hα′′ i ,α′ l (α′′ We have s′′ + 1 = s3 + 1 of these and they are linearly independents. Definition 2.3.17. (B (I1, I2, I3)) Let (I1, I2, I3) be a fixed point of Hilbn,n+1,n+2(C2) with α j the only monomial in I1 but not in I2 and α′ l the only monomial in I2 but not in I3. Then we define the subset B (I1, I2, I3) ⊂ HomR (I1, R±I1 ) ⊕ HomR (I2, R±I2 ) ⊕ HomR (I3, R±I3 ) 2.3. AWEIGHTBASISFOR TI1,I2,I3HILBN ,N +1,N +2(C2) as i ,α′ l B (I1, I2, I3) :=n(0, 0, hα′′ ∪n³0, hα′ ∪©¡ fα,β,Suiv( fα,β),Suiv(Suiv( fα,β))¢¯¯ (α, β) ∉ Obs(I1, I2) ∪ PObsª ) i = 0, . . . , s′′o ∪ i ,α j )´ (α′ i , α j ) ∉ Obs(I2, I3)o ∪ i ,α j ,Suiv(hα′ Observation 2.3.18. In case a), i.e. αl is not immediately above or to the left of α j , we have that: 45 (2.11) # B (I1, I2, I3) = 2n + 5. To see this, start observing that we have s′′ + 1 elements in the first set. Then (α′ Obs(I2, I3) for all i = 0, . . . , s′, so that we have s′ + 1 elements in the second set. i , α j ) ∉ #n(0, 0, hα′′ i ,α j ,Suiv(hα′ i αl ) i = 0, . . . , s′′o = s′′ + 1 i ,α j )´ i = 0, . . . , s′o = s′ + 1 #n³0, hα′ Moreover Obs(I1, I2) and PObs are distinct sets of cardinality s′ + 1 − 2 and s′′ + 1 − 3 respectively. So #©¡ fα,β,Suiv( fα,β),Suiv(Suiv( fα,β))¢¯¯ (αβ) ∉ Obs(I1, I2) ∪ PObsª = # B (I1)−(s′+1−2)−(s′′+1−3). Finally remembering # B (I1) = 2n and putting together the results we have # B (I1, I2, I3) = 2n + 5. Observation 2.3.19. In case b), i.e. α′ have that: l is immediately above or to the left of α j , we # B (I1, I2, I3) = 2n + 4. Suppose, for example that α′ l Moreover (α′ NotP, so that we have s′ + 1 − #NotP elements in the second set. = yα j . Then we have s′′ + 1 elements in the first set. i , α j ) ∈ i , α j ) ∉ Obs(I2, I3) if and only if i < j and p ′ i = 1 i.e. if and only if (α′ #n³0, hα′ i ,α j ,Suiv(hα′ i αl ) i = 0, . . . , s′′o = s′′ + 1 #n(0, 0, hα′′ i α j ) ∉ Obs(I2, I3)o = s′ + 1 − #NotP i ,α j )´ (α′ Finally Obs(I1, I2) and PObs are distinct sets of cardinality s′ +1−2 and s′′ +1−2−#NotP respectively. So that #©¡ fα,β,Suiv( fα,β),Suiv(Suiv( fα,β))¢¯¯ (αβ) ∉ Obs(I1, I2) ∪ PObsª = # B (I1)−(s′+1−2)−(s′′+1−3). Again # B (I1) = 2n, and putting together the results we have # B (I1, I2, I3) = 2n + 4. 46 CHAPTER2. TANGENTSPACESATTORUSFIXEDPOINTS Lemma 2.3.20. Let (I1, I2, I3) be a fixed point of Hilbn,n+1,n+2(C2). The set B (I1, I2, I3) as defined in 2.3.17 is a weight basis for TI1,I2,I3Hilbn,n+1,n+2(C2) Proof. We prove first that they are all linearly independent. Let s ′′Xi =0 c ′′ i ³0, 0, hα′′ i ,α′ s ′Xi =0 l´ + + c ′ i ³0, hα′ X i ,α j ,Suiv(hα′ i ,α j )´ + (α,β)∉Obs(I1,I2)∪Pobs i =0 c ′ i =0 c ′′ cα,β¡ fα,β,Suiv( fα,β),Suiv(Suiv( fα,β))¢ = (0, 0, 0). Then we havePα,β cα,β¡ fα,β¢ = 0 ∈ HomR (I1, R±I1 ) that implies cα,β = 0 for all (α, β). Then alsoPs ′ and finallyPs ′′ i ,α j´ ∈ HomR (I2, R±I2 ), that in turns implies c ′ i ³hα′ l´ = 0 implies c ′′ i ³hα′′ i = 0 for all i = 0, . . . , s′′. i = 0 for all i = 0, . . . , s′; We prove that they actually generate TI1,I2,I3Hilbn,n+1,n+2(C2). As we know, we can generate the tangent space with elements of pure weight, thus it is enough to show that all such elements are in the span of B (I1, I2, I3). Let τ = ( f , h, g ) be a tangent vector of pure weight. We will prove that it belongs to span(B (I1, I2, I3)) by induction on n( f ) the number of generators of I1 that f does not send to 0. i ,α′ If n( f ) = 0, then f = 0, and τ = (0, h, g ). Since in particular (0, h) ∈ Ker(φ12 − ψ12) i ) ∈ 〈α j 〉 for all i = 0, . . . , s′. Then we argue by induction on n(h), the we have that h(α′ number of generators of I2 that h does not send to 0. Either n(h) = 0, i.e. h = 0 and, reasoning in the same way, g ∈ 〈hα′′ i = 0, . . . , s′′〉 ⊂ i αl TI3Hilbn+2(C2), or n(h) = a. If we are in case a) of cases 2.3.2, we have immediately that h ∈ 〈hα′ i = 0, . . . , s′〉 ⊂ i ,α j TI2Hilbn+1(C2) and we can subtract (h − hα′ k α j , g − Suiv(hα′ k α j )) with h(α′ k ) 6= 0 to be able to use the induction step. If we are in case b), we need to be slightly more careful and use the same rea- soning as the induction step in Lemma 2.1.7. In particular if we suppose, say, α j ∈ P ′ α′ i and ¯ı := maxk {kh(α′ ¯ı , α j ) ∉ Obs(I2, I3) and we subtract k ) = 0}, we see that (α′ (h − hα′ ¯ı α j , g − Suiv(hα′ k α j )) to be able to use the induction step. If α j ∈ Q ′ α′ i the reasoning is similar. Suppose now the statement true for all τ′ = ( f ′, h′, g ′) with n( f ′) < t , and suppose that τ = ( f , h, g ) has n( f ) = t . Since f is of pure weight we know that either: 1) we have ¯ı := max{i f (αi ) 6= 0} is strictly smaller than s and f (α¯ı ) ∈ 〈β〉 with β ∈ Pα¯ı , or 2.3. AWEIGHTBASISFOR TI1,I2,I3HILBN ,N +1,N +2(C2) 47 2) we have ¯ı := min{i f (αi ) 6= 0} is strictly bigger than 0 and f (α¯ı ) ∈ 〈β〉 with β ∈ Qα¯ı . Suppose we are in the first case. Renormalize τ = ( f , h, g ) so that f (α¯ı ) = β. First of all notice that (α¯ı , β) ∉ Obs(I1, I2) ∪ PObs. In fact if (α¯ı , β) ∈ Obs(I1, I2) we would be in the situation where ¯ı ≤ j and β = α j y p ¯ı with f (αi ) = 0 for all i > ¯ı. But all θ ∈ HomR (I2, R±I2 ) such that θ(α¯ı ) = if not α j y p¯ı has θ(α¯ı +1) 6= 0, y p ¯ı α¯ı = x q¯ı +1 α¯ı +1 =⇒ α j = y p ¯ı α j y p ¯ı = θ(y p ¯ı α¯ı ) = θ(x q¯ı +1α¯ı +1) = 0 that mod I2 is absurd, since α j ∉ I2. Then, indeed, (α¯ı , β) ∉ Obs(I1, I2). Suppose, now, (α¯ı , β) ∈ PObs. This can happen only in the case a). Moreover, by definition, it means that ¯ı ≤ l and β = α j y p′ ¯ı . Then, since ( f , h) ∈ Ker(φ12 − ψ12) h(α¯ı ) = β, and h(α′ i ) ∈ 〈α j 〉 for all i ≥ ¯ı. So that we can suppose that h(α′ priate combinations of (0, hα′ i ) = 0 for all i ≥ ¯ı by changing τ = ( f , h, g ) with appro- i ,α j )) ∈ B (I1, I2, I3). But then, exactly as before, i ,α j ,Suiv(hα′ it is impossible that there exists a g ∈ HomR (I3, R±I3 ) such that (h, g ) ∈ Ker(φ23 − ψ23). So, in fact, (α¯ı , β) ∉ Obs(I1, I2) ∪ PObs. Then, always thanks to the definition of ¯ı, we have that y p ¯ı β ∈ I1, and thanks to the fact that (α¯ı , β) ∉ Obs(I1, I2) ∪ PObs, we have that ¡ f − fα¯ı ,β, h − Suiv( fα¯ı ,β), g − Suiv(Suiv( fα¯ı ,β))¢ ∈ \1≤i < j ≤3¡Ker(φi j − ψi j ) ◦ πi j¢ is well defined and has first coordinate that sends strictly more generators to 0 mod I1, i.e. n( f − fα¯ı ) < t . So that by induction we have that τ ∈ span(B (I1, I2, I3)) as desired. The case 2) is completely analogous. Lemma 2.3.21 (Cheah). The spaces Hilbn,n+1,n+2(C2) are singular for all n ≥ 1. Proof. We always have a fixed point of type described in case a) of Definition 2.3.2. For that fixed point Observation 2.3.18 and Lemma 2.3.20 tell us that the tangent space is 2n + 5 dimensional, whereas we know that the dimension of the space is 2n + 4. 48 CHAPTER2. TANGENTSPACESATTORUSFIXEDPOINTS 2.3.1 The tangent to the Hilbert-Samuel's strata, case n = (n, n + 1, n + 2) We can now study the weights of the elements of the basis B (I1, I2, I3) in order to understand the tangent space at (I1, I2, I3) of MT (I1),T (I2),T (I3) and of GT (I1),T (I2),T (I3), and their respective positive parts. We divide this task in three observations, one for each type of vectors in the basis B (I1, I2, I3). Observation 2.3.22. Let us look first at the first kind of elements in B (I1, I2, I3) i.e. n(0, 0, hα′′ ) i = 0, . . . , s′′o . i ,α′ l For these we only need to compare the degree of a generator of I3 with that of α′ only monomial in I2 but not in I3. More precisely, we have that (0, 0, hα′′ is tangent to MT (I1),T (I2),T (I3) if and only if deg α′′ i only if deg α′′ deg α′′ i ≤ deg α′ if and only if deg α′′ to the case described in Observation 2.2.14 and its graphic interpretation. l the ) ∈ B (I1, I2, I3) l ; tangent to GT (I1),T (I2),T (3) if and l ; tangent to MT (I1),T (I2),T (I3) and with positive weights if and only if l ; tangent to GT (I1),T (I2),T (I3) and with positive weight l . So everything is perfectly analogous i < degy α′ l and degy α′′ l and degy α′′ i = deg α′ i < degy α′ i = deg α′ ≤ deg α′ i ,α′ l • • α′ l The first set of elements of B (I1, I2, I3) contains as many vectors tangent to MT (I1),T (I2),T (I3) as bul- lets on the same anti-diagonal of α′ l or on lower anti-diagonals. They are positive if they are on a strictly lower anti-diagonal or are on the right of α′ l . Similarly for GT (I1),T (I2),T (I3). The relative posi- tion of α′ l and α j does not matter here. • α j • • • The bullets are generators of I3. Observation 2.3.23. Let us look now at the second kind of elements in B (I1, I2, I3) i.e. n³0, hα′ i ,α j ,Suiv(hα′ i ,α j )´ i = 0, . . . , s′, and (α′ i , α j ) ∉ Obs(I2, I3)o . We need to distinguish cases of Definition 2.3.2. In case a), i.e. αl is not immedi- ately to the left or above α j we have that (α′ i , α j ) ∉ Obs(I2, I3), for all i = 0, . . . , s′. Again (0, h,Suiv(h)) has the same weight as h, so that we can study the weight of the triple only by comparing the degree of α′ i , a generator of I2, with that of α j the only mono- mial in I1 but not in I2. More precisely, we have that³0, hα′ is tangent to MT (I1),T (I2),T (I3) if and only if deg α′ and only if deg α′ only if deg α′ i weights if and only if deg α′ i ≤ deg α j ; tangent to GT (I1),T (I2),T (3) if i = deg α j ; tangent to MT (I1),T (I2),T (I3) and with positive weights if and < degy α j ; tangent to GT (I1),T (I2),T (I3) and with positive ≤ deg α j and degy α′ i i ,α j )´ ∈ B (I1, I2, I3) i ,α j ,Suiv(hα′ i < degy α j . Graphically: i = deg α j and degy α′ case a) 2.3. AWEIGHTBASISFOR TI1,I2,I3HILBN ,N +1,N +2(C2) 49 • • • • α j Here the position of αl does not matter, as long as it is not immediately above or to the left of α j . The second set of elements of B (I1, I2, I3) contains as many vectors tangent to MT (I1),T (I2),T (I3) as bul- lets on the same anti-diagonal of α j or on lower anti-diagonals. They are positive if they are on a strictly lower anti-diagonal or are on the right of α j . Similarly for GT (I1),T (I2),T (I3). • • • The bullets are the generators of I2. In case b) i.e. if α′ l = yα j or α′ l = xα j we have, instead, that some of (α′ in Obs(I1, I2). Precisely these are©(αi , α j )i < j and p ′ = 1} if α′ l = xα j ). i = 1ª if α′ j and q ′ i However instead of not counting them we do like this: we count them, and we consider their weight as usual by comparing deg α′ i with deg α j if i < j (resp. deg α′ i with deg α j if i > j ), but then we subtract their contribution considering their weight by comparing deg yα′ l if i > j ). Graphically: l if i < j (resp. deg xα′ i with deg xα j = α′ i with deg yα j = α′ i , α j ) are actually = yα j , (resp. {(αi , α j )i > l • • α′ l α j • • ⋆ • p ′ 0 = 1 The bullets are some generators of I2. case b) = yα j . Here α′ The second set of elements l of B (I1, I2, I3) contains as many vectors tangent to MT (I1),T (I2),T (I3) as bullets on the same anti- diagonal of α j or on lower anti-diagonals. They are positive if they are on a strictly lower anti- diagonal or are on the right of α j . But we should have not counted those on the right of α j with p ′ i = 1 (in this case only α′ 0). Then we subtract all the ⋆ for these, according to their weight that we calculate by comparing the degree of the star with the degree of α′ = yα j . Similarly for GT (I1),T (I2),T (I3). l Observation 2.3.24. Finally let us look at the third type of elements of B (I1, I2, I3), those of the form ¡ fα,β,Suiv( fα,β),Suiv(Suiv( fα,β))¢ with (αβ) ∉ Obs(I1, I2) ∪ PObs. Again we need only to consider the weight of fα,β. In particular ( fα,β,Suiv( fα,β), Suiv(Suiv( fα,β))) ∈ B (I1, I2, I3) is tangent to MT (I1),T (I2),T (I3) if and only if fα,β is tangent to MT (I1) i.e. if and only if deg α ≤ deg β. Similarly for the positive part, and for the tangent to GT (I1),T (I2),T (I3) and its positive part. 50 CHAPTER2. TANGENTSPACESATTORUSFIXEDPOINTS In fact we need to record when fα,β with (α, β) ∈ Obs(I1, I2) ∪ PObs is raising or preserving degree, and so on, and then understand the dimension of the appropriate vector spaces by difference. Suppose first (α, β) ∈ Obs(I1, I2), then everything goes as in Observation 2.2.16, and for each αi generator of I1 there is at most one β such that (αi , β) ∈ Obs(I1, I2) and the relevant checks on degree are between either y pi αi and α j or between x qi αi and α j , so that fαi ,β ∈ TI1 MT (I1) but (αi , β) ∈ Obs(I1, I2) ⇐⇒ and similarly for all the other spaces we are interested in. Graphically: deg y pi αi ≤ deg α j deg x qi αi ≤ deg α j if i ≤ j , if i ≥ j (2.12) α j ⋆ ⋆ Here the position αl does not matter. We want to understand who are those (α, β) such that fα,β ∈ B (I1) but (α, β) ∈ Obs(I1, I2), and study their weights. Then exactly as in the Observation 2.2.16, we need to subtract a star according to the anti-diagonal it is in relative to the anti- diagonal of α j . The stars are in correspondence with Obs(I1, I2). Suppose now that (α, β) ∈ PObs, then necessarily (x a y bα, x a y bβ) ∈ Obs(I2, I3) with (a, b) = (0, 0), (0, 1) or (1, 0) and x a y bα = α′ i a minimal generator of I2. Then we can focus on Obs(I2, I3). For these we can reproduce, completely analogously, the same reasoning we did in the case of Obs(I1, I2), replacing, though, generators of I1 with generators of I2, and α j with αl . Case a). Remember that there is only one element in NotP and it corresponds to the index of Obs(I2, I3) that is not relevant, because there is not fα,β ∈ B (I1) that realizes it as Suiv( fα,β). Recall that NotP is either y p′ µα j −1, αl j −1¶ If ( j < l ) then NotP = ³xα j , αl y´ . If ( j > l ) then NotP = ¡yα j , αl x¢ j −1¶ . µα j +1, αl q′ x or or Now look first at one case, say ( j < l ). Then interestingly enough, the relevant test on the degrees is between y p′ j −1 α j −1 and αl or between y xα j and αl but in any case y p′ j −1 α j −1 or y xα j is the box one to the left and one above of α j , so that in the graphical 2.3. AWEIGHTBASISFOR TI1,I2,I3HILBN ,N +1,N +2(C2) 51 visualization of 2.2.17 the relevant check is always the same. The same is true if ( j > l ). Thus, in all cases, the Obs(I2, I3) that we need not to consider, is the one represented by the box one to the right and one above α j . This starts to be relevant only if it is on the same anti-diagonal as α′ l or on a lower one, i.e. only if α j is on an anti-diagonal at least two steps lower than the one of αl . Graphically we can summarize as follow: Case a) αl ⋆ α j ⋄ We need to subtract those (α, β) with fα,β ∈ B (I1) and the correct weight, such that (α, β) ∈ PObs. i , β′) ∈ Obs(I2, I3). For each such, there is an (α′ We can then proceed by studying this (α′ i , β′) ∈ Obs(I2, I3) i.e. comparing the associated star with αl . However we need not consider NotP, the cor- ner of α j , that is the box one to the left and one above of α j , is represented as a ⋄ in the picture. ⋆ ⋆ The stars are in correspondence with the Obs(I2, I3) that matter. Case b) Now either α′ = xα j . Suppose to be in the first case. When we list l the elements fα,β ∈ B (I1) according to their weight, we have still to exclude all those for which (α, β) ∈ PObs. The PObs are in one to one correspondence with the Obs(I2, I2) = yα j or α′ l except for NotP =©(α′ case b) i , α j )¯¯i < j and p ′ i = 1ª . ⋆ α′ l α j ⋆ ° Here α′ = yα j . We need to subtract Obs(I2, I3) \ l NotP according to their weights, i.e. according to the relative position of the anti-diagonal they are in and α′ l . The elements in NotP are the ° above the α′ i with p ′ i = 1. The stars represent Obs(I2, I3) \ NotP The ° represents NotP. Observation 2.3.25. The elements in NotP, the ° of the picture above, are exactly those stars we subtracted in the Observation 2.3.23 case b). This insures that the final formula is the same along all cases. 52 CHAPTER2. TANGENTSPACESATTORUSFIXEDPOINTS We are now ready to write combinatorial formulas for the dimensions of the tangent spaces at MT1,T2,T3 GT1,T2,T3 and their positive parts. To state them we need to recall Definition 2.2.17. We will apply it to the two couples of Young diagrams (Γ1, Γ2) and (Γ2, Γ3). Each couple is considered independently from the other. The only extra bit of information that remembers that we are actually dealing with a triple of nested Young diagrams is given by the following. Definition 2.3.26. Let Γ3 = Γ2 ∪ {α1} = Γ1 ∪ {α2} ∪ {α1} be a triple of nested Young diagrams. Then if α2 is not im- mediately above or on the right of α1 put a diamond ⋄ in the box immediately above and to the right of α1 i.e. in position x yα1. Then define, in all cases, α2 ⋄ α1 ⋄ G(Γ1, Γ2, Γ3) := # {⋄ ∈ Λα2} ⋄ M (Γ1, Γ2, Γ3) := # {⋄ ∈ Λi¯¯Λi < Λα2 } + # {⋄ ∈ Λα2 } ⋄ M +(Γ1, Γ2, Γ3) := # {⋄ ∈ Λi¯¯Λi < Λα2 } + # {⋄ ∈ Λα2¯¯ ⋄ to the right of α2 } ⋄ G +(Γ1, Γ2, Γ3) := # {⋄ ∈ Λα2¯¯ ⋄ to the right of α2 }. Lemma 2.3.27. Suppose that (I1, I2, I3) ∈ GT1,T2,T3 ⊂ MT1,T2,T3 is a fixed point of the T1+ action on Hilbn,n+1,n+2(0) with generic weights w1, w2 such that w1 < w2 and (n + 2)w1 > (n + 1)w2. Call (Γ1, Γ2, Γ3) the triple of Young diagrams associated to (I1, I2, I3). 1) The dimension of the tangent space to MT1,T2,T3 at (I1, I2, I3) is equal to dim MT1 + • M (Γ1, Γ2) + • M (Γ2, Γ3) − ⋆ M (Γ1, Γ2) − ⋆ M (Γ2, Γ3) + ⋄ M (Γ1, Γ2, Γ3). 2) The dimension of the tangent space to GT1,T2,T3 at (I1, I2, I3) is equal to dimGT1 + •G(Γ1, Γ2) + •G(Γ2, Γ3) − ⋆G(Γ1, Γ2) − ⋆G(Γ2, Γ3) + ⋄G(Γ1, Γ2, Γ3). 3) The dimension of the positive part of tangent space to MT1,T2,T3 at (I1, I2, I3) is equal to dim T + I1 MT1 +•M +(Γ1, Γ2)+•M +(Γ2, Γ3)−⋆M +(Γ1, Γ2)−⋆M +(Γ2, Γ3)+⋄M +(Γ1, Γ2, Γ3). 4) The dimension of the positive part of the tangent space to GT1,T2,T3 at (I1, I2, I3) is 2.3. AWEIGHTBASISFOR TI1,I2,I3HILBN ,N +1,N +2(C2) 53 equal to dim T + I1 GT1 + •G +(Γ1, Γ2) + •G +(Γ2, Γ3) − ⋆G +(Γ1, Γ2) − ⋆G +(Γ2, Γ3) + ⋄G +(Γ1, Γ2, Γ3). Proof. The statement follows immediately from the definition of B (I1, I2, I3) in 2.3.17 and Observations 2.3.22, 2.3.23, 2.3.24 and 2.3.25. In the next chapter we will prove that the Hilbert-Samuel's strata are smooth using the first two points of Lemma 2.3.27 and studying the dimensions of the strata. Once smoothness is proven the last two points of Lemma 2.3.27 will give us the homological degrees of a basis for the homology of Hilbn,n+1,n+2(0). 54 CHAPTER2. TANGENTSPACESATTORUSFIXEDPOINTS Chapter 3 Smoothness of the Hilbert-Samuel's strata In this chapter we prove the main geometric result we need, namely that the Hilbert-Samuel's strata for the Hilbert Scheme Hilbn,n+1,n+2(0) are smooth. This will allow us to use the theorem of Bialynicki-Birula to study their cell decompositions, and, ultimately, to study the homology of Hilbn,n+1,n+2(0) itself. The strategy is the following. Now that we know the dimensions of the tan- gent spaces at the fixed points we can prove that the spaces MT1,T2,T3 (and GT1,T2,T3 ) are smooth by studying their dimensions. In particular we want to prove the following. Proposition 3.0.28. Let T1, T2, T3 be three admissible sequences of nonnegative integers as in 1.2.7. Then dim MT1,T2,T3 ≥ dim TI1,I2,I3 MT1,T2,T3 dimGT1,T2,T3 ≥ dim TI1,I2,I3GT1,T2,T3 for all I1, I2, I3 fixed points of the T2 action on MT1,T2,T3. As a consequence of smoothness, we know that the attracting sets are affine cells, and their dimensions are equal to the positive part of the tangent spaces that we already calculated in the previous chapter. To prove Proposition 3.0.28 we use the results of Iarrobino [Iar77]. For every T , admissible sequence of integers, he defines a special Young diagram ΓT , whose attracting set AT he proves being affine by giving explicitly a set of special generators for each ideal in the cell. Moreover he proves that MT is covered by a finite union of spaces isomorphic to this cell. Then we see how this can be extended to MT1,T2. More precisely we introduce an affine space A• M(ΓT1 ,ΓT2 ) and ⋆ M (ΓT1 , ΓT2 ) equations that cut out of AT × A• M(ΓT1 ,ΓT2 ) exactly the attracting cell labeled by (ΓT1 , ΓT2 ). This suffices to prove that the dimen- sion of MT1,T2 is greater than or equal to the dimension of its tangent space at each point. 55 56 CHAPTER3. SMOOTHNESSOFTHEHILBERT-SAMUEL'SSTRATA The previous step can be extended without problem to MT1,T2,T3 by iterating it twice. However to get to the dimension of the tangent space in this case, we need to gain an extra dimension whenever ⋄ (ΓT1 , ΓT2 , ΓT3 ) = 1, recalling point (1) of Lemma 2.3.27. To do so we show that one of the ⋆ M (ΓT2 , ΓT3 ) equations we mentioned above is actually trivially satisfied. We repeat all the arguments to prove similar statements for GT1,T2,T3. Finally, in the last section, we present some direct computations for attracting sets of Hilbert schemes of longer flags. As a result we give sharp bounds for when Hilbert-Samuel's strata are no longer all smooth. We give the Poincaré polynomials for those few remaining Hilbert schemes whose attracting sets are all affine. 3.1 Iarrobino's standard generators We need to recall the results of Iarrobino on the punctual Hilbert scheme. The crucial step is that of the definition of normal pattern associated to a type T and a sys- tem of parameters (u, v ) = (ax + b y, c x + d y). For an ideal with normal pattern we find especially nice generators, the so called standard generators. A system of parameters is simply a linear change of coordinates of the plane (u, v ) =Ãa b c d!Ãx y! withÃa b c d! ∈ GL2(C), as GL2(C) acts by automorphisms on R = C[[x, y]]. Fixing coordinates (u, v ), we con- sider monomials in u and v. Let us begin to define a pattern for an ideal I ∈ Hilbn(0). This is a set of monomials, in some sense, as disjoint as possible from I . Definition 3.1.1. Let P be a set of monomials in (u, v ). We denote by P j the monomials in P having degree j . By type T (P ) we mean the sequence T (P ) = (t0, . . . , t j , . . . ), where t j = # P j . Let I ∈ Hilbn(0). We say that I has pattern P if one of the following equivalent condi- tions is satisfied. (i) For all j , 〈P ∩ m j 〉 ⊕ I ∩ m j = m j , (ii) For all j , 〈P j 〉 ⊕ I j = R j , (iii) 〈P 〉 ∩ I = 0, and T (P ) = T (I ). Definition 3.1.2. Let T be an admissible type as in 1.2.7 and (u, v ) = (ax + b y, c x + d y) a system of parameters. Define ΓT to be the only Young diagram that has diagonal 3.1. IARROBINO'SSTANDARDGENERATORS 57 sequence T (ΓT ) equal to T and is such that all its boxes have the lowest possible x coordinate. The normal pattern P = P ((u, v ), T ) is the set of monomials in u, v with expo- nents in ΓT . Explicitly, P =S j P j , where P j :=© the t j monomials of degree j with highest v degreeª =n u j −t j v t j +1, u j −t j −1v t j +2, . . . , uv j −1, v jo . Once a system of parameters is chosen we define IT = I(u,v),T to be the mono- mial ideal associated to ΓT in the parameters (u, v ). Most of the time there will be no possible confusion on the choice of parameters, so we simply write IT . Example 3.1.3. Graphically we can see an example as follow. Let T = (1, 2, . . . , 6, 5, 3, 2, 0). Then ΓT is: A Young diagram with normal pat- tern T = (ti )i ≥0 : all the anti-diagonals Λi are filled with ti elements starting from the left and without interrup- tions. Definition 3.1.4. Let P be a normal pattern for T and (u, v ). Then we define MP = MP ((u,v),T ) :=©I ∈ Hilb(0)¯¯ I has pattern Pª ⊂ MT and GP = GP ((u,v),T ) :=©I ∈ Hilb(0)¯¯ I is homogeneous and has pattern Pª ⊂ GT . Remark 3.1.5. In the original choice of parameters (u, v ) = (x, y) we have that MP is exactly the attracting set of MT to IT , and GP is the attracting set of GT to IT . Fixing a type T all MP (resp. GP ) for different systems of parameters are clearly isomorphic. Proposition 3.1.6. [Iar77, Proposition 3.2] Fix an admissible type T . Let N =P j t j ( j + 1 − t j ). Then whenever P runs through the normal patterns of type T in any sets of distinct systems of parameters (x, y − a0x), . . . , (x, y − aN x) with ai ∈ C we have MP (x,y −a j x) = MT , and N[j =0 GP (x,y −a j x) = GT . N[j =0 In other words, each ideal I ∈ MT (resp. any I ∈ GT ) fails to have normal pattern in at most N systems of parameters of the specified sort. Observe that the MP (x,y −a j x) (resp. the GP (x,y −a j x)) share the fixed point IT so that the above unions are connected. Iarrobino then, in order to prove that MT is smooth, proves by brute force that MP is isomorphic to an affine space by giving an explicit isomorphism. This isomor- 58 CHAPTER3. SMOOTHNESSOFTHEHILBERT-SAMUEL'SSTRATA phism specifies the coefficients of a fixed number of some polynomials that are a spe- cial set of generators of ideals in MP . Even if remark 3.1.5 and the fact that the attract- ing sets are affine cells (thanks to Byialinucky-Birula) already prove smoothness, the explicitness of the work of Iarrobino is essential to have more informations on MT and GT . To go on we need to introduce more notations. We work for the rest of the chapter only with the starting systems of parameter (u, v ) = (x, y). Suppose T = (ti )i ≥0 is chosen. Observe that for IT , the only monomial ideal of normal pattern, the standard monomial generators are especially easy. We have I = (α0, α1, . . . , αd ) where d is the initial degree of I and there exist ki ∈ N such that k0 = 0 < k1 < · · · < kd and αi = xd −i y ki . y kd x y kd−1 αi The standard monomial generators of the monomial ideal IT with normal pattern are especially easy: • There are exactly d + 1 monomial generators where d is that initial de- gree of IT i.e. s = d following previous notation. • For all i = 1, . . . , d we have qi = 1. Re- call that qi is the x distance between αi and αi +1. • The ki are computable from the pt for t ≤ i , where pt is the y distance between αt and αt +1. . . . xd −2 y k2 xd −1 y k1 xd Definition 3.1.7. Let IT be the monomial ideal with normal pattern P (T ), and let (α0, . . . , αd ) be the list of its standard monomial generators. If α = αi is one of them, then we define, using the notation of 2.1.1, the following: P ≥ P = S ≥ S = α := ©β ∈ Pα deg β ≥ deg αª , α := ©β ∈ Pα deg β = deg αª , α := ©γ ∈ ΓT \ Pα deg γ ≥ deg αª , α := ©γ ∈ ΓT \ Pα deg γ = deg αª . (3.1) 3.1. IARROBINO'SSTANDARDGENERATORS 59 P ≥ αi = P = αi = S ≥ αi = S = αi = d[i =0©(αi , β)¯¯β ∈ Pαi deg β ≥ deg αiª , d[i =0©(αi , β)¯¯β ∈ Pαi deg β = αª , d[i =0©(αi , γ)¯¯ γ ∈ ΓT \ Pα and deg γ ≥ deg αiª , d[i =0©(αi , γ)¯¯ γ ∈ ΓT \ Pα and deg γ = deg αiª . We also set: P M := PG := S M := SG := d[i =0 d[i =0 d[i =0 d[i =0 Theorem 3.1.8. [Iar77, Lemma 2.4, Prop. 2.5, Lemma 2.7, Prop. 2.8] Let T be an admissible sequence of nonnegative integers. (1) We have that dim MT = # P M and there is an isomorphism Adim MT ∼= MP . The iso- morphism is explicitly constructed by determining the coefficients of certain poly- nomials that will be generators of a point in MP . More precisely: (2) For each (aα,β)(α,β)∈P M point in Adim MT there exist unique coefficients dα,γ ∈ C for all (α, γ) ∈ S M such that the ideal I = I (aα,β) generated by the polynomials fi = αi + Xβ∈P ≥ αi aαi ,β β + Xγ∈S + αi dαi ,γ γ (3.2) is in MP . These polynomials are called the standard generators for I . The coeffi- cients dαi ,γ are polynomial expressions in the aαk ,β's, for k ≤ i . (3) We have that dimGT = # PG and there is an isomorphism AdimGT ∼= GP . The iso- morphism is explicitly constructed by determining the coefficients of certain poly- nomials that will be generators of a point in GP . More precisely: (4) For each (aα,β)(αβ)∈PG point in AdimGT there exist unique coefficients dα,γ ∈ C for all (α, γ) ∈ SG such that the ideal I = I (aα,β) generated by the polynomials fi = αi + Xβ∈P = αi aαi ,β β + Xγ∈S = αi dαi ,γ γ (3.3) is in GP . These polynomials are called the standard generators for I = I (aα,β). The coefficients dαi ,γ are polynomial expressions in the aαk ,β's, for k ≤ i . Remark 3.1.9. We will not prove the theorem. However the arguments we use in the next few lemmas can be easily adapted to prove it. As a matter of fact the arguments that follow are adapted from the proof of Iarrobino. Example 3.1.10. We would like to visualize more clearly the standard generators we introduced in (3.2). Let then I (aα,β) be the ideal in MP defined as in (3.2). We focus for 60 CHAPTER3. SMOOTHNESSOFTHEHILBERT-SAMUEL'SSTRATA example on f1: f1 = α1 + Xβ∈P ≥ α1 aα1β β + Xγ∈S ≥ α1 dα1γ γ . fd = y kd a d fd −1 a d fi a f2 f1 f0 The standard generators of an ideal with nor- mal pattern: Here we concentrate on f1. The boxes marked with an a correspond to the monomial in P ≥ α1, while the boxes marked with a d correspond to monomials in S ≥ α1. The a's are the free coefficients to be chosen in A#P ≥ α1 while the d's are the unique coefficients that depend on the free choices made for the free coefficients of f1 and f0. Notice in particular that there are no monomials in the support of f1 with degree lower than α1. Also, all other monomials have strictly bigger y degree than α1. Let now T1 = (t1,i )i ≥0, T2 = (t2,i )i ≥0, T3 = (t3,i )i ≥0 be three admissible sequences of non- negative integers, nested as in (1.2.7), with Ti = n + i − 1. We call m, m′ the indexes of where we have the jump, i.e. the indexes such that t1,i if i 6= m, t1,m + 1 if i = m and t2,i t2,m + 1 if i 6= m′, if i = m′ (3.4) Let Γi = ΓTi be the standard Young diagram associated to Ti , and let I Γi be the mono- mial ideal associated to Γi . Observation 3.1.11. One of the gifts of working with the normal patterns is that most of the cases of 2.2.2, among which we had to distinguish in the previous sections, will simply not happen now. More precisely only case 1a) and case 2) are possible. This is true at both steps j and l . Moreover case 2) is possible if and only if t1,d = d and α j = α0 or t2,d = d and αl = α′ 0. Case 1a): possible. Case 1b): not possible. α h α The boxes marked with the h for hole show why some of the cases of 2.2.2 are not standard. t2,i = t3,i = 3.2. MT1,T2,T3 ISSMOOTH. 61 Case 2): possible only with α = α0. Case 3): not possible. h α α Observation 3.1.12. Observe that for I Γ1 and I Γ2, nested monomial ideals with normal patterns, the numbers we defined in 2.2.17 are also more explicit. In fact elements on the same diagonal or on lower diagonals must all lie on their right. Explicitly: t1,m−1 − t1,m • M (Γ1, Γ2) = #©i = 0, . . . , d¯¯ deg αi ≤ deg α jª = j , ⋆G(Γ1, Γ2) = #©i = 0, . . . , d¯¯ deg αi = deg α jª = ⋆ M (Γ1, Γ2) = #©i = 0, . . . , d¯¯ deg αi ≤ deg α j − 1ª = ⋆G(Γ1, Γ2) = #©i = 0, . . . , d¯¯ deg αi = deg α j − 1ª = t1,k−2 − t1,k−1. ⋄G(Γ1, Γ2, Γ3) = ⋄ M (Γ1, Γ2, Γ3) = if deg α j ≤ deg αl − 2, m−1Xk=d otherwise. 1 0 (See next subsection for more details on the case for G(Γ1, Γ2)). If I Γ1, I Γ2 and I Γ3 are nested monomial ideals with normal patterns, the interesting observation is that: t1,m−1 − t1,m + 1 t1,k−1 − t1,k, if m = d , if m > d , (3.5) (3.6) 1 0 if deg α j = deg αl − 2, otherwise. (3.7) 3.2 MT1,T2,T3 is smooth. We start now to apply the results of Iarrobino. Thanks to Proposition 3.1.6 we have that: MP ((x,y ),T1,T2,T3) =n(I1, I2, I3) ∈ Hilbn,n+1,n+2(0)¯¯ Ii ∩ 〈P ((x, y), Ti )〉 = {0}, i = 1, 2, 3o = attracting set of MT1,T2,T3 to the fixed point¡I Γ1, I Γ2, I Γ3¢ . is open in MT1,T2,T3 (and in fact MT1,T2,T3 is covered with a finite number of opens of the form MP ((u,v),T1,T2,T3)). In particular dim MT1,T2,T3 = MP ((x,y ),T1,T2,T3). Now MP ((x,y ),T1,T2,T3) has a unique fixed point, namely (I Γ1, I Γ2, I Γ3), and we already know the dimension of the tangent space at this point thanks to Lemma 2.3.27. Thus, in order to prove Proposition 3.0.28, we need only to prove the following. 62 CHAPTER3. SMOOTHNESSOFTHEHILBERT-SAMUEL'SSTRATA Proposition 3.2.1. For T1, T2, T3 three admissible sequences of nonnegative integers as in 1.2.7, we have dim MP ((x,y ),T1,T2,T3) ≥ dim TIΓ1 ,IΓ2 ,IΓ3 MT1,T2,T3 . Remark 3.2.2. As a consequence we have that for all I1, I2, I3 fixed points of the T2 action on MT1,T2,T3 the dimension of TI1,I2,I3 MT1,T2,T3 is the same. This can also be seen as a combinatorial result. Even if the possibilities for nested Young diagrams with normal patterns are fewer, we still need to treat differently two possible cases. Definition 3.2.3. Case i) For the three monomial ideals I Γ1, I Γ2 and I Γ3 we have that the initial degrees coincide d = d ′ = d ′′. This is the generic case. It happens if we are in case j 1) and in case l 1) of the cases discussed in Definition 2.3.4. Case ii) All the other cases, i.e. either d + 1 = d ′ = d ′′ or d + 1 = d ′ + 1 = d ′′. This happens if and only if we are, respectively, in case j 2) or in case l 2): Case j 2) t1,d = d and m = d Case l 2) t2,d = d and m′ = d αl α j α j αl We start by considering the case where d = d ′ = d ′′ as it is more interesting. In fact, having understood that, we will be able to deal with the other cases with some simple ∼= A# P M given in 3.1.8, and call observations. Consider the parametrization of MP (T1) I (aα,β), (α, β) ∈ P M , the ideal I ((aα,β)) = ( f0, . . . , fd ) generated by the standard generators associated to (aα,β))(α,β)∈P M like in (3.2). We want to see what kind of ideals J are such that (I , J ) ∈ MP ((x,y ),T1,T2). For each θi ∈ C with i = 0, . . . , j − 1 we define the ideal J (aα,β, θi ) giving its generators in terms of those of I (aα,β): J (aα,β, θi ) :=  . . . f0 + θ0 f j f ′ 0 = . . . f ′ j −1 = f j −1 + θ j −1 f j f ′ j = f ′ j +1 = . . . f ′ d y f j f j +1 . . . fd = ,  where the fi , recall, are fi = αi +Pβ∈P ≥ αi aαi ,β β +Pγ∈S ≥ and αi dαi ,γ γ (3.8) I ((aα,β)) = ( f0, . . . , fd ), aα,β, θi ∈ C. 3.2. MT1,T2,T3 ISSMOOTH. Lemma 3.2.4. Call r = ∗M (Γ1, Γ2). Then there exist r equations g0, . . . , gr −1 in the ((aα,β), (θi ))(α,β)∈P M,i =0,..., j −1, defined as y pi fi − x fi +1 = gi (aα,β, θs ) f j mod ( f ′ i +1, . . . , f ′ d ), such that (I (aα,β), J (aα,β, θi )) ∈ MP ((x,y ),T1,T2) ⇐⇒ gi (aα,β, θs ) = 0 for all i = 0, . . . , r − 1. Equivalently if we define Y to be the variety cut out by the gi 's then we have A# P M × A j ⊃ Y :=©(aα,β, θs )(α,β),s ¯¯ gi (aα,β, θs ) = 0ª ,→ MP ((x,y ),T1,T2). Proof. We divide the proof in two parts: in the first part a division procedure will de- fine for us the equations gi . In the second part we will prove that satisfying these equations is a sufficient condition to be in MP ((x,y ),T1,T2). The first part, or rather the procedure that proves it, will be used also later, so we add a name for future reference. 63 (3.9) (3.10) Procedure 3.2.5. We want to prove that condition 3.26 actually defines r equations gi . We do a reduction (division), reasoning by degree. For every element h ∈ R = C[[x, y]], write iny (h) ∈ N∪{+∞} to be the lowest degree of a y power appearing in the expansion of h. Write y pi fi −x fi +1 as a linear combination of monomials: some will be in Γ1, some will be outside Γ2 and one will be α j : y pi fi − x fi +1 = Xβ∈Γ1 cβ β + Xγ∉Γ2 cγ γ + cα j α j cβ, cγ, cα j ∈ C. (3.11) Now we will use the f ′ s , s > i to eliminate the γ ∉ Γ2 in order, starting from the lowest iny (γ). Precisely: let t = iny (y pi fi ). For p ≥ t order all monomials γ ∉ Γ2 with iny (γ) = p by their ascending x degree. To eliminate them in order means to perform many suc- cessive steps. [step p = t ] Looking at the definition of fi and fi +1 there are no γ ∉ Γ2 with cγ 6= 0 and p = iny (γ) in (3.11). [step p = t + 1( if < t + pi +1)] Looking at the definition of fi , fi +1 and f ′ i +1, there is only at most one γ ∉ Γ2 with cγ 6= 0 and iny (γ) = p in the left hand side, and this is yαi +1. Then eliminate it subtracting cyαi +1 y f ′ i +1. Now we do not have any γ ∉ Γ2 with cγ 6= 0 and iny γ ≤ p. More specifically we are left with: y pi fi − x fi +1 − cyαi +1 y f ′ Gi ,β,p (aα,β, θs ) β i +1 = Xβ∈Γ1 + Xγ∉Γ2,in(γ)>p cγ,p γ + gi ,p (aα,β, θs )α j . Here Gi ,β,p (aα,β, θs ), gi ,p (aα,β, θs ) and cγ,p are polynomials in the aα,β, θs and depend on the current step p. We repeat this step similarly until we reach p = t + pi +1. 64 CHAPTER3. SMOOTHNESSOFTHEHILBERT-SAMUEL'SSTRATA [...] [step p = t + pi +1] Let (γ1, . . . , γm) be the ordered (by ascending x degree) list of mono- mials γ ∉ Γ2 with cγ 6= 0 and iny (γ) = p. Multiply y pi +1 f ′ i +1 by appropriate coefficients and monomials in x to eliminate the γe starting from the last one. If needed eliminate α j +2 by a linear multiple of f j +2. The result is: y pi fi − x fi +1 + h(x)y f ′ i +1 + c(aα,β, θs ) f ′ j +2 = Xβ∈Γ1 Gi ,β,p−1(aα,β, θs ) β + + Xγ∉Γ2,in(γ)>p cγ,p−1 γ + gi ,p−1(aα,β, θs )α j . Here again we have polynomial expressions in the (aα,β, θs ), like c(aα,β, θs ) ∈ C or the coefficients of h(x). The sum on the γ ∉ Γ2 is now only on those γ with iny (γ) > p. We repeat this step until we reach p = t +pi +1+pi +2, where we repeat this step but use also f ′ i +3 to eliminate, if necessary αi +3. [...] [step p = in(α j ) = k j ] As all the previous steps but with the difference that we eliminate multiple of α j with f j and not with f ′ j . [...] [step p=kd or p = kd + 1] Now we are left only with multiple of y kd , or y kd +1, depending = y kd , or y kd +1. Then we can eliminate all the on whether j = d i.e. on whether f ′ d remaining γ ∉ Γ2 with appropriate multiples of fd (and not f ′ d : cfr. step p = in(α j ) = k j , if f ′ = fd so there is no difference) and be d left with an expression: = y f d then j = d and we use fd , otherwise f ′ d y pi fi − x fi +1 +Xt >i hi f ′ i + gi (aα,β, θs ) f j = Xβ∈Γ1 Gi ,β(aα,β, θs ) β. (3.12) Here, again, the Gi ,β(aα,β, θs ) and gi (aα,β, θs ) are coefficients that depend polynomially on aα,β, θs ∈ A# PG × A j , and, since this is the last step, we do not write the dependence on the step p. Now since the left hand side of (3.12) is in I (aα,β) and the right hand side is in 〈Γ1〉, and by hypothesis I (aα,β) has T normal pattern, i.e. I (aα,β) ∩ 〈Γ1〉 = {0}, we know that the Gi ,β(aα,β, θs ) must all vanish. Thus we are left with gi , the equation we wanted. Example 3.2.6 (Procedure, picture below). We want to see how the equation g0, relat- ing f0 and f1, arises. The boxes marked with a1 (resp. a1,2) are the free coordinates for f1 (resp. f1 and f2, of course these coefficients can be different one from the others), and those marked with d1 are the possibly non-zero but constrained coefficients of f1. All others are zero. When we multiply f0 by y 2 the boxes of its potentially nonzero coefficients move two steps up, and those of f1 multiplied by x move one to the right. 3.2. MT1,T2,T3 ISSMOOTH. 65 First step p = t = 2: there are no boxes that get out of Γ1 on the 3-rd row (that of y degree 2). At the 4-th row (step p = 3) we eliminate what got out with y f ′ 1. Step p = t + p1 = 4, row=5, we eliminate those that we can (i.e. above and to the right, in this specific case none) with f ′ 2. When we get to step=7 = k j we use f j instead of f ′ 1, the other, only that marked with f2, with f ′ j . We finish with fd . fd = y kd a1,2 fd −1 a1,2 a1,2 f j d1 a1,2 a1,2 d1 a1,2 f2 d1 a1,2 y 2 f0 − x f1 a1 f1 a1 f0 End procedure We have defined equations gi for all i = 0, . . . , r − 1. Recall now that r − 1 = max{i deg αi < deg α j }. Then if we repeat the previous argument for i ≥ r we get that the corresponding gi are identically zero on A# P M × A j , for degree reasons. (This is a big part of Iarrobino's proof of Proposition 3.1.8). By the definition of J we have that (J (aα,β, θi ), f j ) = I (aα,β). We also claim that x f j ∈ J (aα,β, θi ). (3.13) This can be seen by applying the above Procedure 3.2.5 to x f j and see that we do not have a remainder outside J (aα,β, θi ). Then we have that either J (aα,β, θi ) = I or (I (aα,β), J (aα,β, θi )) ∈ MP ((x,y ),T1,T2) as we want. To prove the lemma we then need to check that we are in the second case. Thus it is sufficient to show that, whenever the coefficients ((aα,β)(α,β), (θi )i ) ∈ Adim MP × A j satisfy gi (aα,β, θs ) = 0, we have that J (aα,β, θi ) has normal pattern P (T2). This amounts to show that J (aα,β, θi ) ∩ 〈Γ2〉 = 0. 66 CHAPTER3. SMOOTHNESSOFTHEHILBERT-SAMUEL'SSTRATA Suppose then that we have chosen (aα,β, θs ) ∈ A# P M ×A j , such that gi (aα,β, θs ) = 0 for all i = 0, . . . , r − 1. Since by hypothesis I (aα,β) ∩ 〈Γ1〉 = 0, we need to prove that there does not exist f ∈ J (aα,β, θi ) such that f = α j + Xβ∈Γ1 cββ, cβ ∈ C. Pi ≥0 hi f ′ Suppose by absurd that there exists such a f , and find a contradiction. Write f = i for some polynomial hi ∈ R, and call inJ ( f ) the minimum index for which hi 6= 0. Then we will prove the statement by reverse induction on inJ ( f ). We have a contradiction if inJ ( f ) = d, because then f = hd f ′ hd fd = hd y kd hd y fd = hd y kd +1 if j < d if j = d , d = does not contain in its expansion α j as a monomial, as the degree is strictly bigger. This is indeed true for all inJ ( f ) ≥ j . Suppose now the statement true for all inJ ( f ) > t , j > t and lets prove it for inJ ( f ) = t . We write ht as a series in y ht = ht ,0(x) + ht ,1(x)y 1 + · · · + ht ,kd −1(x)y kd −1 + ht ,kd (x)y kd + . . . , where ht ,s ∈ C[[x]]. f = ht f ′ hi f ′ cββ. (3.14) t +Xi >t i = α j + Xβ∈Γ1 Observe that if ht ,0(x) = ht ,1(x) = · · · = ht ,kd −1(x) = ht ,kd (x) = 0 then we do arrive at a contradiction since y kd +1 = f ′ d ) and then we can rewrite f as d (or y kd +1 = y f ′ f = Xt <i <d hi f ′ i + (hd + ht (x, y/y kd +1) f ′ t ) f ′ d i.e. as an element in ( f ′ already know there is a contradiction. t +1, . . . , f ′ d ) with inJ ( f ) > t for which, by induction hypothesis, we We will then proceed by reverse induction on iny (ht ), the minimal index for which ht ,s(x) 6= 0. We just finished giving the base induction case when iny (ht ) = kd +1, and we proceed with the induction step. We observe that iny (ht ) cannot be 0, oth- erwise the monomial x aα′ t , for the appropriate a ∈ N, would appear only once in the lefthand side of (3.14) and never in the righthand side. In fact, for the same reason, it must be that ht ,m = 0 for all m ≤ p ′ i ≥ pi for i < j . Then we are finally able to use the equations gi . In particular for i = t we have: i . Observe that p ′ y pt ft − x ft +1 = g t (aα,β, θm) f j +Xi ≥t h′ i f ′ i (3.15) and since we picked (aα,β, θs ) ∈ A# P M × A j , such that gi (aα,β, θs ) = 0 for all i we can substitute (3.15) into (3.14) to obtain an f with in(ht ) strictly larger. Then we are done. 3.2. MT1,T2,T3 ISSMOOTH. 67 Corollary 3.2.7. Suppose that T1, T2 is an admissible nested couple of sequences of non- negative integers. Then dim MP ((x,y ),T1,T2) ≥ dim MP ((x,y ),T1) + • M (Γ1, Γ2) − ⋆ M (Γ1, Γ2). (3.16) Proof. We have just proven the statement in case i), i.e. whenever d = d ′ in Lemma 3.2.4. We only need to deal with the other case. Suppose then d + 1 = d ′. This implies α j = α0. Then there is actually only one ideal J (aα,β) such that ¡I (aα,β), J (aα,β)¢ ∈ MP ((x,y ),T1,T2) and it is J (aα,β) = (x f0, y f0, f1, . . . , fd ) where the fi are the standard generators of I (aα,β). Then it is still true that dim MP ((x,y ),T1,T2) ≥ dim MP ((x,y ),T1) + • M (Γ1, Γ2) − ⋆ M (Γ1, Γ2). In this case: • M (Γ1, Γ2) = ⋆ M (Γ1, Γ2) = 0 and dim MP ((x,y ),T1,T2) ≥ dim MP ((x,y ),T1). Corollary 3.2.8. [Che98, Proposition 3.4.11.] Let T1 and T2 be two sequences of non- negative integers as in 1.2.7. Call m the index such that t2,m = t1,m + 1, and d the initial degree of T1. Then MT1,T2 is smooth of dimension dim MT1,T2 = n +Xj ≥d (t j −1 − t j )(t j −1 − t j + 1) 2 + (tm−1 − tm + 1). Proof. Lemma 3.2.4 and at Lemma 2.2.18 prove that dim MP ((x,y ),T1,T2) ≥ dim TIΓ1 ,IΓ2 MT1,T2. This proves that MP ((x,y ),T1,T2) is smooth: in fact all other points have Zariski tangent space of dimension smaller or equal to that of the fixed point. The Proposition of Iar- robino 3.1.6 proves that MT1,T2 is covered by opens that are isomorphic to MP ((x,y ),T1,T2). The formula for the dimension in terms of the t j ,i is clear by looking at Lemma 2.2.18 and Proposition 2.1.16 on the dimension of MT1. Remember that in this case that we do not need this proof: in fact as we know Hilbn,n+1(P2) is smooth and has an affine paving. Moreover MT1,T2 is the union of some of those cells, and MP ((x,y ),T1,T2) is one of those cells. However, as already mentioned, [n, n + 1] is the last case where smoothness of the ambient space Hilbn,n+1(C2) holds. Observation 3.2.9. Observe that as a consequence we get that the inclusion in (3.10) is actually an isomorphism. When we write the generators of J (aα,β, θi ) we actually do not write its standard generators since the element f ′ j = y f j could have expansion with non zero coefficient for a monomial outside Γ2. We could easily remedy to this but it is not needed. In fact to prove the key Lemma 3.2.4 we used the fact that the expansions of the standard 68 CHAPTER3. SMOOTHNESSOFTHEHILBERT-SAMUEL'SSTRATA generators fi has only terms of higher or equal degree than αi and the only one with smallest possible y degree is exactly αi . This remains true for the set of generators we gave for J (aα,β, θi ). This means that we are able to iterate Lemma 3.2.4 to find enough ideals K (aα,β, θi , ηs ) such that (aα,β, θi ), K (aα,β, θi , ηs ) ∈ MP ((x,y ),T2,T3). That is exactly what we are about to do. Now that we understand how and why the dimension of MP ((x,y ),T1,T2) changes with re- spect to the dimension of MP ((x,y ),T1), we can also understand how it changes adding a further step and considering MP((x,y ),T1,T2,T3). Lemma 2.3.27 shows that the dimension of TIΓ1 ,IΓ2 ,IΓ3 MP , exactly following the rule (3.16) twice for the inclusions Γ1 ⊃ Γ2 and Γ2 ⊃ Γ3 except when deg α j ≤ deg αl − 2. In this case it grows by one more. MP ((x,y ),T1,T2,T3 grows, with respect to dim TIΓ1 Let us iterate twice the results in Lemma 3.2.4. For this let I (aα,β) ∈ MP (T1) with (aα,β)α,β ∈ A# P M be the ideal with standard generators ( f0, . . . , fd ) as in (3.2). Let (θi )0≤i < j ∈ A j , and (ηi )0≤i <l ∈ Al . We define the ideals J (aα,β, θi ) and K (aα,β, θi , ηs ) mim- icking the construction in (3.25) J (aα,β, θi ) :=  . . . f0 + θ0 f j f ′ 0 = . . . f ′ j −1 = f j −1 + θ j −1 f j f ′ j = f ′ j +1 = . . . f ′ d y f j f j +1 . . . fd =  , K (aα,β, θi , ηs ) :=  f ′′ 0 = . . . 0 + η0 f ′ f ′ l = f ′ = l −1 f ′′ l −1 f ′′ l f ′′ l +1 . . . f ′′ d = = . . . + ηl −1 f ′ l y f ′ l f ′ l +1 . . . f ′ d .  The results of Lemma (3.2.4) tell us that there exists r = ∗M (Γ1, Γ2) equations gi (aα,β, θs ) in A# P M × A j and r ′ = ∗M (Γ2, Γ3) equations g ′ i (aα,β, θt , ηs ) in A# P M × A j × Al such that: (I (aα,β), J (aα,β, θi )) ∈ MP ((x,y ),T1,T2) ⇐⇒ gi (aα,β, θs ) = 0 for all i = 0, . . . , r − 1, and (I (aα,β), J (aα,β, θi ), K (aα,β, θi , ηs )) ∈ MP ((x,y ),T1,T2,T3) with gi (aα,β, θs ) = 0 ∀ i = 0, . . . , r − 1 ⇒ ⇐ i (aα,β, θt , ηs ) = 0 for all i = 0, . . . , r ′ − 1. g ′ (3.17) Then, since j = • M (Γ1, Γ2) and l = • M (Γ2, Γ3) we have that dim MP ((x,y ),T1,T2,T3) ≥ dim MP ((x,y ),T1 +• M (Γ1, Γ2)+• M (Γ2, Γ3)− ⋆ M (Γ1, Γ2)− ⋆ M (Γ2, Γ3). (3.18) 3.2. MT1,T2,T3 ISSMOOTH. 69 This is good enough to prove dim MP ((x,y ),T1,T2,T3) ≥ dim TIΓ1 ,IΓ2 ,IΓ3 where ⋄ M (Γ1, Γ2, Γ3) = 0. MT1,T2,T3 in all cases Thus the last step is to show that when ⋄ M (Γ1, Γ2, Γ3) = 1, one of the equations we found is not necessary. More precisely. Lemma 3.2.10. Utilize all notations as above. Suppose deg α j ≤ deg αl −2. Then j −1 < r ′ and g ′ j −1(aα,β, θt , ηs ) = 0 for all¡(aα,β)α,β, (θt )0≤t < j , (ηs )0≤t <l¢ ∈ A# P M × A j × Al gi (aα,β, θs ) = 0 ∀ i = 0, . . . , r − 1 such that and g ′ i (aα,β, θt , ηs ) = 0 ∀ i = 0, . . . , r ′ − 1, for i 6= j − 1. In other words the equation g ′ deg α j ≤ deg αl − 2, even though it appears in the list (3.17), being j − 1 < r ′. j −i is not necessary because trivially satisfied whenever Proof. Since deg α j ≤ deg αl − 2, necessarily j < l and j − 1 < r ′. Then in particular we have: f ′′ j −1 = f j −1 + θ j −1 f j + η j −1 fl , f ′′ j = y f j + η j fl , f ′′ l = y fl . The equations g ′ i (aα,β, θt , ηs ) are defined by y p′ i f ′ i − x f ′ i +1 = g ′ i (aα,β, θt , ηs ) f ′ l mod ( f ′′ i +1, . . . , f ′′ d ). In fact we will see that for i = j − 1 we have y p′ j −1 f ′ j −1 − x f ′ j = 0 mod ( f ′′ j , . . . , f ′′ d ). (3.19) (3.20) (3.21) This implies that g ′ j −1(aα,β, θt , ηs ) is always equal to zero, as wanted. Call αl y = β ∈ Γ1. Consider an elimination procedure exactly as in 3.2.5 in the proof of Lemma 3.2.4 to write, relative to I1, y p j −1 f j −1 − x f j = G j −1(aα,β)β mod ( f j , . . . , fd ). Here G j −1(aα,β) is a polynomial expression in the coefficients (aα,β)α,β ∈ Adim MP . Since I1 ∩ 〈Γ1〉 = 0, G j −1(aα,β) = 0 identically on Adim MP , i.e. y p j −1 f j −1 − x f j =Pi ≥ j hi fi . Observe that iny (h j ) > 0. Multiply now both sides of the last equation by y, and add and subtract few terms to get: y p j −1¡y f j −1 + θ j −1 y f j − θ j −1 y f j¢ − x y f j = h j y f j + · · · + hl y fl + · · · + hd y fd . (3.22) 70 CHAPTER3. SMOOTHNESSOFTHEHILBERT-SAMUEL'SSTRATA Observe that p ′ j −1 = p j −1 + 1. Then rearranging we have: y p′ j −1 f ′ j −1 − x f ′ j =¡θ j −1 y p j −1 + h j¢ y f j + · · · + hl y fl + · · · + hd y fd . Since as observed iny (h j ) > 0, and p j −1 ≥ 1 we can factor a further y from the R- multiple of y f j , add and subtract η j fl , to write: y p′ j −1 f ′ j −1 − x f ′ j =¡θ j −1 y p j −1−1 + h j¢ y(y f j + η j fl − η j fl ) + · · · + hl y fl + · · · + hd y fd . Now by substituting (3.19) in (3.22), we have (3.21), as desired. Example 3.2.11. fd = y kd a1,2 fd −1 a1,2 a1,2 fl d1 a1,2 a1,2 ←−β d1 a1,2 f2 d1 a1,2 y f1 y 3 f ′ 0 − x y f1 a1 f1 a1 f0 In the situation depicted on the left αj is the green box, marked f1, while αl is the box marked with fl . Since deg αl = deg αj + 2 we need to prove that g ′ 0 is trivially satisfied. The idea is that this equation was already satisfied since there must be a relation for y 2 f0 − x f1. More explicitly in the proof we write an equation G0 that represent the coefficient of β (the gray box) once we apply the Procedure 3.2.5 to y 2 f0 − x f1: since β is in Γ1 we find that G0 must be identically zero. Now the key point is that when we do this we have an extra power of y in all the interesting terms. This allows us to relate G0 to g0 with some manipulations. Observation 3.2.12. Now we deal with case ii) of 3.2.3. If d + 1 = d ′ + 1 = d ′′ then there is nothing new to prove, as ⋄ M (Γ1, Γ2, Γ3) = 0. Moreover, as we already observed there exists only one K that sits inside J (aα,β, θi ), so the dimension is the expected one. If d + 1 = d ′ = d ′′ then we still need to prove that: if deg α j ≤ deg αl − 2 then g ′ 0, the first of the equations g ′ i i = 0, . . . , ⋆ M (Γ2, Γ3) is trivially satisfied. But this is obvious since in this case g0 is the equation relating y x f0 and x y f0: y p0 f ′ 0 + x f ′ 1 = y x f0 − x y f0 = 0. Proposition 3.2.13. Let T1, T2, T3 be three admissible sequences of nonnegative integers as in 1.2.7 that differ in the indexes m and m′ as in (3.4). Then the Hilbert-Samuel's strata MT1,T2,T3 is a smooth variety of dimension dim MT1,T2,T3 = (n − 1) +Xj ≥d (t j −1 − t j ) (t j −1 − t j + 1) 2 + (tm−1 − tm) + (tm′−1 − tm′)+ 1 0 if m′ ≥ m + 2, otherwise. + 3.2. MT1,T2,T3 ISSMOOTH. 71 It has an affine cell decomposition with cells parametrized by nested Young diagrams Γ1, Γ2, Γ3 that differ in only one box: Γ3 = Γ2 ⊔ {αl } = Γ1 ⊔ {αl } ⊔ {α j } with Γ1 ⊢ n and such that T (Γi ) = Ti . Call α j = (u j , v j ) and αl = (ul , vl ), then the affine cell indexed by (Γ1, Γ2, Γ3) has dimension given by the following formula: pos (Γ1, Γ2, Γ3) = #©(u, v ) ∈ Γ1¯¯ hu,v 6= 0, 1ª + u,v = 0 and u = ul ) or (h′ + #©(u, v ) ∈ Γ1¯¯ (hu,v = 0 and u = u j ) or (hu,v = 1 and v = v j )ª + + #©(u, v ) ∈ Γ2¯¯ (h′ + if ul + vl = u j + v j + 2and vl > v j , if ul + vl > u j + v j + 2, u,v = 1 and v = vl )ª + (3.23) 1 1 0 otherwise. Here hu,v is the hook difference of the box (u, v ) with respect to the Young diagram Γ1, and h′ u,v is the hook difference of the box (u, v ) with respect to the Young diagram Γ2. The Poincaré polynomial of MT1,T2,T3 is Pq¡MT1,T2,T3¢ = (Γ1,Γ2,Γ3)⊢[n,n+1,n+2], T (Γi )=Ti X q pos(Γ1,Γ2,Γ3) . Proof. Thanks to equation (3.18) and Lemma 3.2.10 we know that the Hilbert-Samuel stratum MT1,T2,T3 is covered by opens isomorphic to MP ((x,y ),T1,T2,T3). One of these opens is smooth since at his only torus fixed point it has dimension equal to the dimension of the Zariski tangent space, thanks to equation (3.18), Lemma 3.2.10 and Lemma 2.3.27. Thanks to Theorem 1.3.4 we know that it has a cell decomposition with cells whose dimensions are given by Lemma 2.3.27. Thanks to Theorem 1.3.2 the decom- position in affine cells gives us a basis for the homology, and we can then calculate the Poincaré polynomial. As a consequence we immediately get the following. Proposition 3.2.14. The space Hilbn,n+1,n+2(0) has an affine paving given by the attract- ing sets at the fixed points for the T1+ action. The Poincaré polynomial of Hilbn,n+1,n+2(0) is given by Pq³Hilbn,n+1,n+2(0)´ = X (Γ1,Γ2,Γ3)⊢[n,n+1,n+2] where pos(Γ1, Γ2, Γ3) is as in (3.23). q pos(Γ1,Γ2,Γ3) (3.24) Remark 3.2.15. The natural goal would now be to find a generating series for the Poincaré polynomials of Hilbn,n+1,n+2(0) as n varies. However the above expression appears hard to sum. We will deal with this problem in the next chapter. 72 CHAPTER3. SMOOTHNESSOFTHEHILBERT-SAMUEL'SSTRATA 3.2.1 GT1,T2,T3 is smooth We prove, with the same exact arguments, that the homogenous Hilbert-Samuel's strata GT1,T2,T3 are smooth. One of the main reasons why the proof is basically the same as for the case of MT1,T2,T3 is the following lemma, that dates back to Iarrobino. Lemma 3.2.16. [Iar77, Lemma 2.6.] Let I ∈ MP (T1) be an ideal with normal pattern. Then I ∈ GP (T1) if and only if each standard generator is homogenous. If the standard generators of I are ( f0, . . . , fd ) then the associated graded ideal ρT (I ) has standard gener- ators (F0, . . . , Fd ) where Fi is the initial form of fi , i.e. ρ( fi ) = Fi . For completeness we expand on notation and content. The complication on the ide- als is inversely proportional to the complications on the indexes: even though ev- erything is homogeneous so contains less terms, we need to distinguish those terms with more complicated indexes. In fact basically only the definition of J (aα,β, θi ) of 3.25, and the form of the equations gi (aα,β, θs ) of 3.26 are different but all the argu- ments are the same. Let T1, T2, T3 be admissible sequences of nonnegative integers. Let I (aα,β) ∈ GP (T1) be with standard generators ( f0, . . . , fd ) given by (3.3) for the point (aα,β)α,β ∈ AdimGT1 . Then we can read from the elements of T1 how many generators we have of each degree: we have t1,d −1 − t1,d + 1 generators of degree d, and t1,k−1 − t1,k genera- tors of degree k with k > d, where d is the initial degree of I (aα,β). Suppose now that α j , the box in ΓT2 \ ΓT1 , is of degree d + k i.e. T1 = (0, 1, 2, . . . , d , t1,d , t1,d +1, . . . , t1,d +k , t1,d +k+1, . . . ), T2 = (0, 1, 2, . . . , d , t1,d , t1,d +1, . . . , t1,d +k + 1, t1,d +k+1, . . . ). Suppose first k > 0 so that α j does not have degree d and d ′ = d. This is the gen- eral case, we will treat the others later. In particular it must be j = t1,d +k+1 − 1, since we always want Young diagrams of standard normal form T2. Then for θi with i = t1,d +k, . . . , t1,d +k+1 − 2 we can define the family of ideals J (aα,β, θi ) as J (aα,β, θi ) :=  f ′ 0 = . . . f ′ t1,d+k −1 = f ′ = t1,d+k . . . f0 . . . ft1,d+k −1 ft1,d+k + θt1,d+k ft1,d+k+1−1 f ′ t1,d+k+1−2 = ft1,d+k+1−2 + θt1,d+k+1 −2 ft1,d+k+1−1 f ′ t1,d+k+1−1 = f ′ = t1,d+k+1 y ft1,d+k+1−1 ft1,d+k+1 . . . f ′ d = . . . fd (aα,β)α,β ∈ AdimGT1 , with (θi )i ∈ At1,d+k+1−t1,d+k −1.  3.2. MT1,T2,T3 ISSMOOTH. 73 (3.25) We define this family because we are looking for enough J such that (I (aα,β), J ) ∈ GP (T1, T2). Example 3.2.17. f4 = y k4 f3 ftk+d+1−1 f2 f1 f0 In this example: T1 = (1, 2, 3, 4, 5, 5, 2), and T2 = (1, 2, 3, 4, 5, 5, 3). Here α j is marked by f j = ftk+d+1−1, its degree is d + k = d + 2 = 7, j is 3 = tk+d +1 − 1, and finally t1,d +k+1 − t1,d +k − 1 is 5 − 2 − 1 = 2. Every polynomial here is homoge- neous, i.e. it contains monomials that are only on its antidiagonal. Then, when we add the box α j we can mod- ify only f1 and f2. Observation 3.2.18. As it was happening in the case MP the possible free new coordi- nates we are adding are as many as •G(Γ1, Γ2). Now we look at the equations we have to impose. Lemma 3.2.19. Call r = ∗G(Γ1, Γ2). Then there exist r equations g t1,d+k−2 , . . . , g t1,d+k−1 in the ((aα,β), (θi ))(α,β)∈PG,i =t1,d+k ,...,t1,d+k+1−2 defined as y pi −1 fi −1 − x fi = gi (aα,β, θs ) f j mod ( f ′ i +1, . . . , f j −1), (3.26) such that (I (aα,β), J (aα,β, θi )) ∈ GP ((x,y ),T1,T2) ⇐⇒ gi (aα,β, θs ) = 0 for all i = t1,d +k−2, . . . , t1,d +k−1. Equivalently if we define Y to be the variety cut out by the gi then we have A# PG × At1,d+k+1−t1,d+k −1 ⊃ Y := {(aα,β, θs )gi (aα,β, θs ) = 0} isom −−−→ MP ((x,y ),T1,T2). Proof. The proof is exactly the same as for Lemma 3.2.4. The equations are in corre- spondence, by multiplying by x, with the free boxes on the antidiagonal immediately below α j since everything is homogenous and has to remain homogenous. Corollary 3.2.20. With the notations as above we have that: dimGP ((x,y ),T1,T2) ≥ dimGP ((x,y ),T1 + •G(Γ1, Γ2) − ⋆G(Γ1, Γ2). Proof. We need to consider the other cases i.e. when d 6= d ′ and when deg α j = d. Observe that in the homogenous case the above expression actually sais that the di- 74 CHAPTER3. SMOOTHNESSOFTHEHILBERT-SAMUEL'SSTRATA mension of GP ((x,y ),T1,T2) can be lower than that of GP ((x,y ),T1 . Suppose we are in the case d + 1 = d ′. Then α j = α0 deg α j = d. Then there is actually only one ideal J (aα,β) such that ¡I (aα,β), J (aα,β)¢ ∈ GP ((x,y ),T1,T2) and it is J (aα,β) = (x f0, y f0, f1, . . . , fd ) where the fi are the standard generators of I (aα,β). It is still true that dimGP ((x,y ),T1,T2) ≥ dimGP ((x,y ),T1) + •G(Γ1, Γ2) − ⋆G(Γ1, Γ2) because •G(Γ1, Γ2) = ⋆G(Γ1, Γ2) = 0 and dimGP ((x,y ),T1,T2) ≥ dimGP ((x,y ),T1). Suppose finally d = d ′ and deg α j = d. Then the reasoning for the case d = d ′ and deg α j > d still works, the only difference being that there is one more standard generator of I (aα,β) with degree d, i.e. f0. So up to a change of indexes everything is the same. Corollary 3.2.21. [Che98, Proposition 3.4.12] Let T1 and T2 be two admissible nested sequences of nonnegative integers as in 1.2.7. Call m the index for which t2,m = t1,m + 1 and d the initial degree of T1. Then GT1,T2 is smooth, and of dimension dimGT1,T2 =Xj ≥d (t j −1 − t j + 1)(t j − t j +1) + (tm−1 − tm) − (tm−2 − tm−1). It has the affine cell decomposition given by the Bialynicki-Birula decompositions whose cells have dimensions as specified in Lemma 2.2.18. Proof. The proof is clear by looking at Lemma 3.2.4 and at Lemma 2.2.18 on the di- mension of the tangent space and at Proposition 2.1.16 on the dimension of GT . Ob- serve that in this case, contrary to the case MT1,T2 we actually need this proof, since it is not true that GT1,T2 is union of affine cells for Hilbn,n+1(P2). Having understood how to pass from GP (T1) to GP (T1,T2) we can understand how to go from GP (T1,T2) to GP (T1,T2,T3). We start by supposing that d = d ′ = d ′′ and deg α j , deg α′ l > d, in fact say deg α j = = d + k ′, k, k ′ > 0. Let again I (aα,β) ∈ GP (T1) with (aα,β)α,β ∈ A# PG be the d + k and deg α′ l ideal with standard generators ( f0, . . . , fd ) as in (3.3). Let (θi )t1,d+k ≤i ≤t1,d+k+1−2 ∈ At1,d+k+1−t1,d+k −1, and (ηi )t2,d+k′ ≤i ≤t2,d+k′ +1−2 ∈ At2,d+k′ +1−t2,d+k′ −1. Then we define the ideals J (aα,β, θi ) as in 3.2. MT1,T2,T3 ISSMOOTH. 75 (3.25) and the ideals K (aα,β, θi , ηs ) mimicking the construction in definition (3.25). K (aα,β, θi , ηs ) :=  f ′′ 0 = . . . f ′′ t2,d+k′ −1 = f ′′ t2,d+k′ = . . . f ′ 0 . . . f ′ t2,d+k′ −1 f ′ t2,d+k′ + ηt2,d+k′ f ′ t2,d+k′ +1−1 t2,d+k′ +1−2 + ηt2,d+k′ +1−2 ft2,d+k′ +1−1 t2,d+k′ +1−2 = f ′ f ′′ f ′′ t2,d+k′ +1−1 = f ′′ = t2,d+k′ +1 . . . f ′′ d = y f ′ t2,d+k′ +1−1 ft2,d+k′ +1 . . . f ′ d .  (3.27) The results of Lemma (3.2.19) tell us that there exists r = ∗G(Γ1, Γ2) equations gi (aα,β, θs ) in A# PG ×At1,d+k+1−t1,d+k −1 and r ′ = ∗G(Γ2, Γ3) equations g ′ At2,d+k′ +1−t2,d+k′ −1 such that: i (aα,β, θt , ηs ) in A# PG ×At1,d+k+1−t1,d+k −1× (I (aα,β), J (aα,β, θi )) ∈ GP ((x,y ),T1,T2) ⇐⇒ gi (aα,β, θs ) = 0 for all i , and (I (aα,β), J (aα,β, θi ), K (aα,β, θi , ηs )) ∈ GP ((x,y ),T1,T2,T3) with gi (aα,β, θs ) = 0 ∀ i ⇒ ⇐ g ′ i (aα,β, θt , ηs ) = 0 for all i . (3.28) Then, since t1,d +k+1 − t1,d +k − 1 = •G(Γ1, Γ2) and t2,d +k ′+1 − t2,d +k ′ − 1 = •G(Γ2, Γ3) we have that dimGP ((x,y ),T1,T2,T3) ≥ dimGP ((x,y ),T1 +•G(Γ1, Γ2)+•G(Γ2, Γ3)−⋆G(Γ1, Γ2)−⋆G(Γ2, Γ3) (3.29) which is good enough to prove dimGP ((x,y ),T1,T2,T3) ≥ dim TIΓ1 ,IΓ2 ,IΓ3 where ⋄ M (Γ1, Γ2, Γ3) = 0. GT1,T2,T3 in all cases The last step needed is then to show that when ⋄ M (Γ1, Γ2, Γ3) = 1, one of the equations we found is not necessary. This is the content of the next Lemma. Lemma 3.2.22. Utilize all notations as above, and suppose deg α j = deg α′ l t2,d +k ′ ≤ j ≤ t1,d +k+1 − 2 and − 2. Then j (aα,β, θt , ηs ) = 0 for all¡(aα,β)α,β, (θt )t , (ηs )t¢ ∈ A# PG × A•G(Γ1,Γ2) × A•G(Γ2,Γ3) gi (aα,β, θs ) = 0 ∀ i = t1,d +k, . . . , t1,d +k+1 − 2 and g ′ such that g ′ i (aα,β, θt , ηs ) = 0 ∀ i = t2,d +k ′, . . . , t2,d +k ′+1 − 2, i 6= j. 76 CHAPTER3. SMOOTHNESSOFTHEHILBERT-SAMUEL'SSTRATA In other words the equation g ′ j is not necessary because trivially satisfied whenever deg α j = deg αl −2, even though it appears in the list (3.28), being t2,d +k ′ ≤ j ≤ t1,d +k+1 −2. Proof. The proof is exactly as in the case of MP ((x,y ),T1,T2,T3). Observation 3.2.23. Now we will deal with the cases left aside on d , d ′ and d ′′. If d +1 = d ′ + 1 = d ′′ then there is nothing new to prove, as observed in the proof of 3.2.20. If d + 1 = d ′ = d ′′ then we still need to prove that: if deg α j ≤ deg αl − 2 then g ′ 0 the first of the equations g ′ i i = 0, . . . , ⋆ M (Γ2, Γ3) is trivially satisfied. But this is obvious since, in this case: y p0 f ′ 0 + x f ′ 1 = y x f0 − x y f0 = 0. Proposition 3.2.24. Let T1, T2, T3 be three admissible sequences of nonnegative integers as in 1.2.7. Then the homogenous Hilbert-Samuel's stratum GT1,T2,T3 is a smooth variety of dimension + (tm−1 − tm − 1) − (tm−2 − tm−1) + (tm′−1 − tm′ ) − (tm′−2 − tm′−1) + It has an affine cell decomposition with cells parametrized by nested Young diagrams Γ1, Γ2, Γ3 that differ in only one box: Γ3 = Γ2 ⊔ {αl } = Γ1 ⊔ {αl } ⊔ {α j } with Γ1 ⊢ n and such that T (Γi ) = Ti . The affine cell indexed by (Γ1, Γ2, Γ3) has dimension given by the follow- ing formula: pos (Γ1, Γ2, Γ3) = #©(u, v ) ∈ Γ1¯¯ hu,v = −1ª + •G +(Γ1, Γ2) − ⋆G +(Γ1, Γ2) + + •G +(Γ2, Γ3) − ⋆G +(Γ2, Γ3) + ⋄G +(Γ1, Γ2, Γ3). (3.30) The Poincaré polynomial of GT1,T2,T3 is Pq¡GT1,T2,T3¢ = (Γ1,Γ2,Γ3)⊢[n,n+1,n+2], T (Γi )=Ti X q pos(Γ1,Γ2,Γ3) . Proof. Thanks to equation 3.29 and Lemma 3.2.22 we know that the Hilbert-Samuel stratum GT1,T2,T3 has everywhere dimension equal to the dimension of its Zariski tan- gent space, then it is smooth. Thanks to Theorem 1.3.4 we know that it has a cell decomposition with cells whose dimensions are given by Lemma 2.3.27. Thanks to Theorem 1.3.2 the decomposition in affine cells gives us a basis for the homology, and we can then calculate the Poincaré polynomial. dim GT1,T2,T3 = Xj ≥d + 0 1 (t j −1 − t j + 1)(t j − t j +1) + if m′ ≥ m + 2, otherwise. 3.3. STRATAFORLONGERFLAGSARESINGULAR 77 3.3 Strata for longer flags are singular In this section we deal with flags of more than three ideals, i.e. starting at Hilbn,n+1,n+2,n+3(0). The goal is to prove that the Hilbert-Samuel's strata in these cases are not all smooth, so that the techniques we used until now will not yield results to- wards the understanding of their homology. In fact we will show a bit more, namely that there are attracting sets that are not smooth, in particular not isomorphic to affine cells. We will specify at what n, the length of the biggest ideal, these singular attracting sets start appearing, while for lower n we still have an affine cell decompo- sition and a basis for the homology. Everything in this section is obtained by direct computation. 3.3.1 Four flag case Consider the punctual Hilbert scheme of flags of four nested ideals Hilbn,n+1,n+2,n+3(0). We will see that starting at n = 6 we can find an Hilbert-Samuel's stratum MT1,T2,T3,T4 ⊂ Hilb6,7,8,9(0) that is not smooth. We will see that it is not smooth by exhibiting an attracting cell that is not smooth. We can also show, by direct computation that 6 is the smallest integer for which this happens. Lemma 3.3.1. For n = 1, 2, 3, 4, 5 all the Hilbert-Samuel's strata MT1,T2,T3,T4 of the Hilbert scheme Hilbn,n+1,n+2,n+3(0) are smooth. They have an affine paving that implies that Hilbn,n+1,n+2,n+3(0) has an affine paving. The Poincaré polynomials of the total spaces are given by: Pq³Hilb1,2,3,4(0)´ = 1 + 3q + 4q 2 + 2q 3, Pq³Hilb2,3,4,5(0)´ = 1 + 4q + 8q 2 + 9q 3 + 4q 4, Pq³Hilb3,4,5,6(0)´ = 1 + 4q + 10q 2 + 14q 3 + 13q 4 + 6q 5, Pq³Hilb4,5,6,7(0)´ = 1 + 4q + 11q 2 + 22q 3 + 30q 4 + 25q 5 + 9q 6, Pq³Hilb5,6,7,8(0)´ = 1 + 4q + 11q 2 + 24q 3 + 42q 4 + 51q 5 + 36q 6 + 11q 7. For all n ≥ 6 there is at least one Hilbert-Samuel's stratum that is not smooth and with an attracting set that is not isomorphic to an affine space. For the proof of the first part we checked by hand that the all the attracting sets for 78 CHAPTER3. SMOOTHNESSOFTHEHILBERT-SAMUEL'SSTRATA n ≤ 5 are indeed affine, the computation is a bit long but straightforward. We now show that for n = 6 we have an attracting cell that is not isomorphic to an affine space, and thus proving that the corresponding Hilbert-Samuel's strata is not smooth. The fixed point is the following: 3 2 The fixed point represented on the left is the flag (x2, x y 2, y 4) ⊃ (x3, x2 y, x y 2, y 4) ⊃ ⊃ (x3, x2 y, x y 3, y 4) ⊃ (x3, x2 y, x y 3, y 5). 1 It is the flag of monomial ideals with normal patterns (1, 2, 2, 1), (1, 2, 3, 1), (1, 2, 3, 2), (1, 2, 3, 2, 1). The attracting cell is described as follow. Let I1 = ( f0, f1, f2) be in MP (T1) with standard generators f0, f1, f2 given by f0 = x2 + a1x y + a2 y 2 + a3 y 3, f1 = x y 2 + b1 y 3, f2 = y 4. The attracting cell is where a1, a2, a3 and b1 in C are the free affine coordinates. ( f0, f1, f2) ⊃ (x f0, y f0, f1, f2) ⊃ (x f0 + θ0 f1, y f0 + θ1 f1, y f1, f2) ⊃ (3.31) ⊃ (x f0 + θ0 f1 + η0 y 4, y f0 + θ1 f1 + η1 y 4, y f1 + η2 y 4, y 5) where in the last step we need to impose the equation (a1, a2, a3, b1, θ0, θ1, η0, η1, η2) ∈ A9. for G¡(a, b), (θ), (η)¢ = 0 y g0 − xg1 ≡ G¡(a, b), (θ), (η)¢ y 4 mod (g1, g2, g3), Here G is defined with a procedure similar to the Procedure 3.2.5 as where we called g0, g1, g2, g3 the generators (that need not be standard) of the last ideal in (3.31) in the order we have written them. Doing explicitly the computations we find G¡(a, b), (θ), (η)¢ = η2(θ1(b1 − a1) − θ0 − θ2 Taking partial derivatives we see that the points 1 ) + b1θ1(b1 − a1) + a2θ1. a2 = b1(a1 − b2), θ0 = θ1 = η2 = 0, a1, b1, η0, η1 ∈ C are all singular points for the equation G, so that the attracting cell is not isomorphic to an affine space. 3.3. STRATAFORLONGERFLAGSARESINGULAR 79 Observe that we just proved that also the projective strata GT1,T2,T3,T4 is not smooth, as the extra dimension of MT1,T2,T3,T4 , represented by the coordinate a3, does not play a role in what we said. If n ≥ 7, the situation is slightly different as what we describe is a phenomenon that belongs to the affine fiber of the projective map ρT1,T2,T3,T4 : MT1,T2,T3,T4 → GT1,T2,T3,T4. The reasoning is however completely similar. We give the de- tails for completness. 3 ... n − 7 boxes ... 2 1 The fixed point represented on the left is the flag (x2, x y 2, y n−2) ⊃ (x3, x2 y, x y 2, y n−2) ⊃ ⊃ (x3, x2 y, x y 3, y n−2) ⊃ (x3, x2 y, x y 3, y n−2). It is the flag of monomial ideals with normal patterns (1, 2, 2, 1, 1, . . . , 1), (1, 2, 3, 1, 1, . . . , 1), (1, 2, 3, 2, 1 . . . , 1), (1, 2, 3, 2, 1, . . . , 1, 1). The attracting cell is described as follow. Let I1 = ( f0, f1, f2) be in MP (T1) with standard generators f0, f1, f2 given by f0 = x2 + a1x y + a2y 2 + a3 y 3 + · · · + an−3 y n−3, f1 = x y 2 + b1 y 3 + b2 y 5 + · · · + bn−5 y n−3, f2 = y n−2. Now a1, an−4, an−3 and b1, b2, . . . , bn−5 in C are the free affine coordinates, while a2, . . . , an−5 depend polynomially on the previous coordinates according to equations that arise imposing y 2 f0 − x f1 ≡ 0 mod ( f1, f2). The attracting cell is ( f0, f1, f2) ⊃ (x f0, y f0, f1, f2) ⊃ (x f0 + θ0 f1, y f0 + θ1 f1, y f1, f2) ⊃ ⊃ (x f0 + θ0 f1 + η0 y n−2, y f0 + θ1 f1 + η1 y n−2, y f1 + η2 y n−2, y n−1) where in the last step we need to impose the equation (3.32) (3.33) ((ai )1≤i ≤n−3, (bi )1≤i ≤n−5, θ0, θ1, η0, η1, η2) ∈ A2n−3. for G¡(a, b), (θ), (η)¢ = 0 y g0 − xg1 ≡ G¡(a, b), (θ), (η)¢ y n−2 mod Here G is defined with a Procedure similar to 3.2.5 as (g1, g2, g3). 80 CHAPTER3. SMOOTHNESSOFTHEHILBERT-SAMUEL'SSTRATA We called g0, g1, g2, g3 the generators (that need not be standard) of the last ideal in (3.33) in the order we have written them. Doing explicitly the computations as in Procedure 3.2.5, and using the equations coming from (3.32), we find G((a, b), (θ), (η)) = η2(θ1(b1 − a1) − θ0 − θ2 1) + bn−5(2θ1b1 + θ1a1) + an−3θ1. Taking partial derivatives we see that the points an−3 = bn−5(a1 − 2b1), θ0 = θ1 = η2 = 0 are all singular points for the equation G, so that the attracting cell is not isomorphic to an affine space. 3.3.2 Five, and longer, flag case Consider the punctual Hilbert scheme of flags of five nested ideals Hilbn,n+1,n+2,n+3,n+4(0). We will see that starting at n = 3 we can find an Hilbert-Samuel's stratum MT1,T2,T3,T4,T5 ⊂ Hilb3,4,5,6,7(0) that is not smooth. Again we will see that it is not smooth by exhibiting an attracting cell that is not smooth. We can also show, by direct computation, that 3 is the smallest integer for which this happens. Lemma 3.3.2. For n = 1, 2 all the Hilbert-Samuel's strata MT1,T2,T3,T4,T5 of the Hilbert scheme Hilbn,n+1,n+2,n+3,n+4(0) are smooth. The spaces Hilbn,n+1,n+2,n+3,n+4(0) for n = 1, 2 have an affine paving and their Poincaré polynomials are given by: Pq³Hilb1,2,3,4,5(0)´ = 1 + 4q + 8q 2 + 9q 3 + 4q 4, Pq³Hilb2,3,4,5,6(0)´ = 1 + 5q + 13q 2 + 22q 3 + 23q 4 + 11q 5 + q 6. For n ≥ 3 there is at least one Hilbert-Samuel's stratum that is not smooth, and with an attracting cell that is not isomorphic to an affine space. Let n ≥ 2, then at least one of the Hilbert-Samuel's stratum of Hilbn,...,n+5(C2) is not smooth, and has an attracting cell that is not isomorphic to an affine space. For all k ≥ 6 and all n ≥ 1 at least one of the Hilbert-Samuel's stratum of Hilbn,...,n+k (C2) is not smooth, and has an attracting cell that is not isomorphic to an affine space. The only case left aside is Hilbn,...,n+5(C2) for n = 1. This is naturally isomorphic to Hilb2,3,4,5,6(C2), so we know an affine paving for it. Again, to prove the first part of 3.3. STRATAFORLONGERFLAGSARESINGULAR 81 the Lemma we wrote down explicitly the attracting sets and checked that they are all affine, in the cases where we claim that they are so. We now show that for n = 3 we have an attracting cell that is not isomorphic to an affine space, thus proving also that the corresponding Hilbert-Samuel's stratum is not smooth. Example 3.3.3. The fixed point is the following: 4 3 2 1 The fixed point represented on the left is the flag (x, y 3) ⊃ (x2,x y, y 3) ⊃ (x2, x y 2, y 3) ⊃ ⊃ (x2, x y 2, y 4) ⊃ (x2, x y 2, y 5). This is the flag of monomial ideals with normal patterns (1, 1, 1), (1, 2, 1), (1, 2, 2), (1, 2, 2, 1), (1, 2, 2, 1, 1) Let us call (ω, a, θ, α, η, β1, β2) the coordinates of A7. The attracting set is isomorphic to the subspace Y ⊂ A7 cut out by the equation (η − ω)(η − θ) = 0 i.e. it is the intersection of two linear spaces of dimension 6 in a linear space of dimension 5. The explicit isomorphism between Y and MP is given by parametrizing MP as Ãx + ωy + a y 2 y 3 ! ⊃ ⊃ x2 + ωx y x y + ωy 2 y 3  ⊃ x2 + (ω + θ)x y + ωθy 2 x y 2 y 3  ⊃  ⊃ x2(ω + θ)x y + ωθy 2 + αy 3 x2(ω + θ)x y + ωθy 2 + αy 3 + β1 y 4 x y 2 + ηy 3 y 4 x y 2 + ηy 3 + β2 y 4 y 5  and noticing that in the last inclusion we have to impose the equation 0 = det 1 ω + θ 0 0 0 0 1 0 0 0  ωθ ω + θ 1 1 0 0 0 ωθ 0 ω + θ ωθ 0 η η 1 = (η − ω)(η − θ) = 0  because otherwise y 4 ∈ I5. Remark 3.3.4. The fixed point we just described is also a fixed point of Hilb1,2,3,4,5,6,7(0), in the sense that the flag of five ideals we described can be uniquely extended to a flag of seven ideals (combinatorially there is only one way to extend the above skew stan- dard diagram to a Young standard diagram). Observe now that the projection map p2 : Hilb1,2,3,4,5,6,7(0) → Hilb2(0) is a Zariski locally trivial fibration, as GL2 acts by auto- morphisms on Hilb2(0) ∼= P1 by changing the coordinates of the plane. 82 CHAPTER3. SMOOTHNESSOFTHEHILBERT-SAMUEL'SSTRATA One can then check that the attracting set described above in Example 3.3.3 is the only one that is not an affine cell. The checking is done by writing explicitly all the remaining towhundredthirtyone cells. In particular all the sets attracted by . . . 1 2 . . . are affine cells. This mean that the fiber p −1 2 ((y, x2)) ⊂ fixed points of the form Hilb1,2,3,4,5,6,7(0) has an affine paving, and thus, since p2 is locally trivial on P1, that the entire Hilb1,2,3,4,5,6,7(0) has an affine paving, even though it is not the one coming from the attracting sets. Then we can compute by hand the Poincaré polynomial of Hilb1,2,3,4,5,6,7(0) and the result is the following: Pq³Hilb1,2,3,4,5,6,7(0)´ = 1 + 6q + 19q 2 + 41q 3 + 63q 4 + 64q 5 + 32q 6 + 5q 7. For longer flags, or for flags of five ideals but of higher lengths, unfortunately this no longer holds and I do not know if there is an affine cell decomposition. Remark 3.3.5. One can still write a weight basis for the tangent spaces at the fixed points of Hilb1,2,3,4,5,6,7(C2) and see if the weights with respect to the torus T1+ action have some geometrical meaning. The positive part, however, fails to give the right Poincaré polynomial, as direct computations show. This is why we didn't investigate it further for longer flag cases. Now we show that for flags of ideals of higher lengths the analogous stratum is not smooth, slightly differently, but in a similar way. We will also show that all longer flags have the analogous stratum that is not smooth. 4 3 ... k − 5 boxes ... 2 1 The fixed point represented on the left is the flag (x, y k ) ⊃ (x2,x y, y k) ⊃ (x2, x y 2, y k ) ⊃ ⊃ (x2, x y 2, y k+1) ⊃ (x2, x y 2, y k+1) that is the flag of monomial ideals with normal patterns (1, 1, . . . , 1), (1, 2, 1, . . . , 1), (1, 2, 2, 1, . . . , 1), (1, 2, 2, 1, . . . , 1, 1), (1, 2, 2, 1, . . . , 1, 1, 1) Here we call (ω, a1, . . . , ak−2, θ, β1, β2, γ1, γ2) the coordinates of Ak+4. Then the attracting cell is isomorphic to the subspace Y ⊂ Ak+4 cut out by the single equation (ak3 − β2)(ω − θ) = 0. Thus for every n ≥ 3 we have an Hilbert-Samuel's stratum in Hilbn,n+1,n+2,n+3,n+4(0) not smooth. Observation 3.3.6. We now show that for all longer flags the analogous stratum is not smooth. Consider the above picture. There is only one way to complete it to a skew diagram with k boxes, i.e. we need to put the number 1, 2, . . . , k − 5 in the first 3.3. STRATAFORLONGERFLAGSARESINGULAR 83 column. For the same reason there is only one attracting cell associated to this picture in HilbN −k,N −k+1,...,N and its description is exactly as before. Then this proves that for every k there is an Hilbert-Samuel's stratum in HilbN −k,N −k+1,...,N with k ≤ N that is not smooth, thus showing the claim that the three step case is the last one where all the Hilbert-Samuel's strata are smooth. 84 CHAPTER3. SMOOTHNESSOFTHEHILBERT-SAMUEL'SSTRATA Chapter 4 Generating function and Hilbn,n+2(C2). The goal of this chapter is to prove the formula for the generating function of the Poincaré polynomials of Hilbn,n+1,n+2(0): Xn≥0 Pq³Hilbn,n+1,n+2(0)´ zn = q + 1 (1 − zq)(1 − z2q 2) Yk≥1 1 1 − zk q k−1 . To do so we will need three ingredients. (4.1) (1) Study the positive part of the tangent spaces at fixed points of Hilbn,n+1,n+2(C2) with respect to the torus T∞ i.e. the generic one dimensional torus that has weights w1, w2 such that w1 < w2 and 1 ≪ w1 . As it happens for previous cases the positive w2 part with respect to this torus is much easier to understand combinatorially. Observe, however, that since Hilbn,n+1,n+2(C2) is not smooth we do not know that the attracting sets for this torus character are affine varieties, nor do we have an immediate geomet- ric interpretation of the positive part. This is why we need the following step. (2) Prove that we can actually change the one dimensional torus and take weights as described in the previous point but still obtain the same result for the Poincaré polynomial. In the process we show that the attracting sets for this torus are still affine cells. (3) Study the space Hilbn,n+2(C2) and two associated spaces as described below. One of them is smooth, the generating function for its Betti numbers is known, and we claim it has the same Betti numbers as Hilbn,n+1,n+2(0). To prove the claim, we will only need to match the combinatorial data of the fixed points and the respective positive part of the tangent spaces to obtain the wanted result for Hilbn,n+1,n+2(0). This relation is clear only at the combinatorial level, and not at the geometric one. As a byproduct we will obtain the last result of the thesis i.e. a combinatorial formula for the Poincaré 85 86 CHAPTER4. GENERATINGFUNCTIONANDHILBN ,N +2(C2). polynomials of Hilbn,n+2(0) and a closed expression for their generating function: Xn≥0 Pq³Hilbn,n+2(0)´ zn = 1 + q − q z (1 − zq)(1 − z2q 2) Ym≥1 1 1 − zm q m−1 . (4.2) 4.1 A tale of two tori We start motivating the work of this section by showing that, indeed, the pos- itive part of the tangent spaces at fixed points of Hilbn,n+1,n+2(C2) for the torus with weights 1 ≪ w1 is combinatorially easier. Then we will prove that we can use the pos- w2 itive part with respect to this torus to write down the Poincaré polynomial. Lemma 4.1.1. Consider the action of T∞ on R = C[x, y]. Let (Γ1, Γ2, Γ3) be three nested Young diagrams representing a fixed point for this torus action on Hilbn,n+1,n+2(C2). Then the positive part of the tangent space at (Γ1, Γ2, Γ3) is pos∞ (Γ1, Γ2, Γ3) = n + 2 − ℓ(Γ3) + 1 n + 2 − ℓ(Γ3) if degy α′ l otherwise. ≥ degy α j + 2, (4.3) Here ℓ(Γ3) is the number of columns of Γ3 and with the notations of the last chapter we call α′ l and α j the boxes such that Γ3 = Γ2 ∪ {α′ l } = Γ1 ∪ {α j } ∪ {α′ l }. Proof. Observe that an eigenvector for T2 with eigenvalues λr µs, where λ and µ are independent torus characters, is positive with respect to the chosen weights if and only if s is positive. Recall the definition of B = B (I Γ1, I Γ2, I Γ3) in 2.3.17. B¡I Γ1, I Γ2, I Γ3¢ = B¡I Γ1¢ \© fα,β(α, β) ∈ Obs(I Γ1, I Γ2)ª \© fα,β(α, β) ∈ PObs(I Γ1, I Γ2, I Γ3)ª ∪n³0, hα′ ∪n³0, 0, hα′′ i , α j ) ∉ Obs(I Γ2, I Γ3)o ∪ i ,α j )´¯¯ (α′ l´¯¯ i = 0, . . . s′′o . (4.4) i ,α j ,Suiv(hα′ i ,α′ A vector v ∈ B is positive with respect to the chosen weights, if and only if it represents a map from a box to a box that is in a strictly higher row. In terms of mono- mials, if v is label by α 7→ β then v is positive if and only if degy β > degy α. Let us analyze all the terms in (4.4). It is clear, and we already know, that pos(B (Γ1)) = n − ℓ(Γ1). Suppose for the moment that α′ l is not on the same row or column of α j . 4.1. ATALEOFTWOTORI 87 Consider now what we exclude and what we add by removing the set of vectors { fα,β(α, β) ∈ Obs(I Γ1, I Γ2)} and by adjoiningn³0, hα′ i ,α j ,Suiv(hα′ In the latter set we do not need to exclude any term worrying about (α′ thanks to the current hypothesis on the relative position of α′ l and α j . Looking at def- inition 2.3.6, we have that fα,β is such that (α, β) ∈ Obs(I Γ1, I Γ2) and positive if and only if α = αi , β = α j y pi with i ≤ j − 1 since we have that i ,α j )´¯¯ (α′ i , α j ) ∉ Obs(I Γ2, I Γ3)o. i , α j ) ∈ Obs(I Γ2, I Γ3), degy α j = degy αi + pk j −1Xk=i (4.5) for all i < j . Still looking at equation (4.5), we have that hα′ i ,α j is positive if and only if i ≤ j . Thus the total contribution of these two sets to the positive part of B (Γ1, Γ2, Γ3) is pos³n³0, hα′ i ,α j ,Suiv(hα′ i , α j ) ∉ Obs(I Γ2, I Γ3)o´ + i ,α j )´¯¯ (α′ −pos¡© fα,β(α, β) ∈ Obs(I Γ1, I Γ2)ª¢ =  l´¯¯ i = 0, . . . s′′o when we add it. i ,α′ Completely analogously for the contribution of© fα′,β(α′, β) ∈ Obs(I Γ2, I Γ3)ª when we re- move it, and ofn³0, 0, hα′′ The only term left to be considered is© fα′,β(α′, β) ∈ NotP(I Γ1, I Γ2, I Γ3)ª. This, un- der the current hypothesis that α′ l is not on the same row or column as α j , contains only an element given by Definition 2.3.9, and this element is positive with respect to the chosen torus action if and only if degy αl > degy α j + 2. 1 0 if 0 < j , if 0 = j. Then noticing that ℓ(Γ3) = ℓ(Γ1) +  1 0 if 0 < j , if 0 = j we obtain exactly what we wanted. 1 0 +  if 0 < l , if 0 = l If we now suppose that α′ l = yα j we can say, analogously, that ³n³0, hα′ i ,α j ,Suiv(hα′ i ,α j )´¯¯ (α′ i , α j ) ∉ Obs(I Γ2, I Γ3)o´ + l´¯¯ i = 0, . . . , s′′o´ + +pos³n³0, 0, hα′′ −pos¡© fα,β(α, β) ∈ Obs(I Γ1, I Γ2)ª¢ + −pos¡© fα′,β(α′, β) ∈ Obs(I Γ2, I Γ3)ª¢ + i ,α′ +pos¡© fα′,β(α′, β) ∈ NotPª¢ =  (4.6) 2 1 if 0 < j , if 0 = j 88 CHAPTER4. GENERATINGFUNCTIONANDHILBN ,N +2(C2). since in this case (α′ also that, since α′ l statement is proved also in this case. i , α j ) ∉ Obs(I Γ2, I Γ3) if and only if (α′ i , = yα j , we always have at least a positive vector as hα′′ α j y ) ∈ NotP(I Γ1, I Γ2, I Γ3) (notice ). Then the 0 ,α′ l Finally if α′ l = xα j then we do not have any positive vectors in either nhα′ i ,α j¯¯ (α′ i , α j ) ∉ Obs(I Γ2, I Γ3)o or in© fα′,β(α′, β) ∈ NotP(I Γ1, I Γ2, I Γ3)ª. So that the left hand side of (4.6) is either 1 if 0 < j or 0 if j = 0, that, noticing j = 0 iff ℓ(Γ3) = ℓ(Γ1) + 2, is what we wanted. As promised formula (4.3) is much easier than the formula (3.23). Now we deal with the more cumbersome task of proving that we get the same polynomial by using one or the other. We will see that the main part of the work is to understand in more details why we can use freely one formula instead of the other in the case of Hilbn(0) and in the case of Hilbn,n+1(0). Then we will be able to tackle the case Hilbn,n+1,n+2(0). We start by fixing some notation to better frame the problem. Definition 4.1.2. We want to give names to the one dimensional subtori of T2 that are most relevant for us. We will always assume that the weights w1, w2 are such that w1 < w2. We will denote TW ⊂ T2 the one dimensional subtorus given by putting w2 = W . Most of these tori are generic in the sense that the fixed points sets for their w1 action on Hilbn(C2), Hilbn+1(C2) and Hilbn+2(C2) are isolated. However we care also about those special values of W that are not generic. So we say that W ∈ Q is a wall for n if the fixed point set of Hilbn(C2) is not discrete. Similarly we say that W ∈ Q is a wall for [n, n + 1] if the fixed point set of Hilbn,n+1(C2) is not discrete. Finally we say that W ∈ Q is a wall for [n, n +1, n +2] if the fixed point set of Hilbn,n+1,n+2(C2) is not discrete. This typically happens for values like: m k with 1 ≤ m, k ≤ n + 1. If W is a wall for n (resp. for [n, n + 1], resp. for [n, n + 1, n + 2]) we will denote W + and W − two values in Q that are respectively bigger and smaller than W by a small amount, small enough that between W + and W and between W and W − there are no other walls for n (resp. for [n, n + 1], resp. for [n, n + 1, n + 2]). Recall the special names we gave the two extreme cases we are interested in with respect to n ∈ N: T∞ := and w1 < w2 n + 2 < w2 w1 T1+ := and w1 < w2 n+1 > w2 n+2 w1 . These two tori are generic. The torus T∞ is the one we would like to use, as it gives formula (4.3). However only the second one i.e. T1+ is such that the Hilbert-Samuel strata MT are union of attracting sets of Hilbn(P2), and so, in the case of Hilbn,n+1,n+2(0), only for T1+ we know that the attracting sets are affine and that the positive parts of the tangent spaces at a fixed point give the dimension of the affine cells that contracts 4.1. ATALEOFTWOTORI 89 to it. Example 4.1.3. A typical wall for n happens for value like: m For example: let n = 3, then the only wall is W = 2. For any generic torus the fixed point sets of Hilb3(0) consists of the three isolated points given by the monomial ideals of length 3, i.e. (x, y 3), (y, x3) and (x2, x y, y 2). If W = 2 then the fixed points set consists of two components, one with all the points of the form (ω1 y +ω2x2) + m3 with [ω1, ω2] ∈ P1 and the other with the isolated point (x, y 3). k with 1 ≤ m, k ≤ n + 1. y 2 x y x2 ω1 y +ω2x2 y −→ x3 The one dimensional fixed com- ponents of the torus T2 acting on Hilb3(0). Example 4.1.4. Even when we know that a torus action gives an affine cell decom- position, we cannot change freely the one parameter subgroup, hoping to still get an affine cell decomposition, or hoping to extract useful informations by looking at the positive part of the tangent spaces at the fixed points. We give an example. Consider X the variety that consists of three P1 that meet at the point P so that the intersection is planar, i.e. dim TP X = 2. Suppose that P is the point [1 : 0] for each of the P1 and that a action on X that restricts to each P1 all other points of X are smooth. Consider a T1 as the usual rescaling action with weights 1 and a, i.e. t · [ω1 : ω2] = [t ω1 : t aω2]. Then, if a < 1, the T1 a action gives a decomposition of X in affine cells parametrized by fixed points P = P1, P2, P3, P4 and given by Ai = {x ∈ X limt →0 t · x = Pi } , each of which has di- mension given by the positive part of the tangent space at the fixed point. However if a > 1 this is not longer true, as the only attracting set that is not a point is A1 and the positive part of the tangent space at P1 = P is only two dimensional thanks to the planar hypothesis. Notation 4.1.5. Denote with ∆ the positive quadrant N × N seen as a union of boxes. In the rest of the chapter we denote Γ ⊢ n, (Γ1, Γ2) ⊢ [n, n + 1] (or the shorter version Γ ⊢ [n, n + 1]) and (Γ1, Γ2, Γ3) ⊢ [n, n + 1, n + 2] (or the shorter version Γ ⊢ [n, n + 1, n + 2]) respectively Young diagram of size n, a couple of Young diagrams of size n and n + 1 that differ only in one box marked with a 1 , and a triple of Young diagrams of sizes n, n + 1 and n + 2 that differ, in order, in one box marked with a 1 , and in one box 2 . We identify these with fixed points of, respectively, Hilbn(C2), marked with a Hilbn,n+1(C2) or Hilbn,n+1,n+2(C2). In this chapter all the objects we associated to the latter we can consider as associated to the diagrams themselves. For example we will often talk about the ideal Γ to mean the ideal I Γ. Also, for example, given Γ ⊢ n we will talk about its outer corners as the standard monomial generators of the associated monomial ideal I Γ, and we will denote B (Γ) the basis of the tangent space TΓHilbn(C2) constructed in Chapter 2, and so on. Definition 4.1.6. Let n ∈ N and W be a wall for n or [n, n + 1] or [n, n + 1, n + 2]. Then we 90 CHAPTER4. GENERATINGFUNCTIONANDHILBN ,N +2(C2). define for Γ ⊢ n or Γ ⊢ [n, n + 1] or Γ ⊢ [n, n + 1, n + 2] the following integers. posW + (Γ) := #©v ∈ B (Γ)¯¯v is positive wrt to the weights of TW +ª posW − (Γ) := #©v ∈ B (Γ)¯¯v is positive wrt to the weights of TW −ª posW (Γ) := #©v ∈ B (Γ)¯¯v is positive wrt to the weights of TWª W (Γ) := #©v ∈ B (Γ)¯¯v has zero weight wrt toTW and is positive wrt to TW +ª W (Γ) := #©v ∈ B (Γ)¯¯v has zero weight wrt to TW and is positive wrt to TW −ª s+ s− We also use the same operators to count the corresponding numbers of vectors of any subset of one of the B (Γ). Proposition 4.1.7. Let n ∈ N and let W be a wall for n. Then for every k ∈ N we have In particular we have that the following two polynomials are the same: #©Γ ⊢ n¯¯posW + (Γ) = kª = #©Γ ⊢ n¯¯posW − (Γ) = kª . XΓ ⊢ n q pos∞(Γ) = XΓ ⊢ n q pos1+ (Γ) . Remark 4.1.8. The statement is equivalent to the existence of a bijection of sets φW : {Γ ⊢ n} → {Γ ⊢ n} such that posW +(Γ) = posW +(φW (Γ)). We do not claim that there is a preferred such bi- jection, as Example 4.1.23 shows. As already said we could prove this fact simply by using smoothness of Hilbn(C2). Smoothness, as we saw, implies that every generic one dimensional torus gives a de- composition of Hilbn(0) in affine cell whose dimensions are given by the positive parts of the tangent spaces at the attracting fixed points. However we need to understand better what is going on in order to prove a sim- ilar statement for Hilbn,n+1,n+2(0). In particular we will introduce some combinatorial transformations on the set {Γ ⊢ n} so to give a better description of φW and prove that a similar map, with similar properties exists for the set {Γ ⊢ [n, n + 1, n + 2]}. In 4.1.22 we will give a more ad hoc proof of Proposition 4.1.7. Before introducing the definition we present an example. Example 4.1.9. Consider the case n = 3. There is only one wall W = 2. We depict below the dimension of the positive parts (for the notation cfr. Definition 4.1.6 ) at the fixed 4.1. ATALEOFTWOTORI 91 points with respect to the two different tori TW + = T∞ and TW − = T1+. posW + = 2 posW 1 = 2 posW + = 1 posW − = 0 posW + = 0 posW − = 1 The two polynomials q 2 + q + 1 = XΓ ⊢ n q pos∞(Γ) = XΓ ⊢ n q pos1+ (Γ) = q 2 + 1 + q are clearly the same. We will see that in passing the wall W the last two fixed points "exchange roles". In more details: observe that there are two tangent vectors of weight W that are: fx2,y ∈ B  and f y,x2 ∈ B . These are the two vectors that are, respectively, positive only for TW + and positive only for TW −. They are also the tangent vectors to the TW fixed P1 that connects the corresponding two fixed points: see Example 4.1.3. The plan is the following: we will define combinatorial transformations of Young diagrams associated to these two vectors that keep track of how the dimensions of the positive parts change in crossing the wall. For example the transformation associated to the vector fx2,y is Tx2,y : Tx2,y −→ posW = 0, (s+ W , s− W ) = (1, 0) posW = 0, (s+ W , s− W ) = (0, 1) Definition 4.1.10 (Sliding boxes). Let n ∈ N, W be a wall for n and Γ ⊢ n. For every fα,β ∈ B (Γ) that has zero weight with respect to TW we want to define a new partition Tα,β(Γ) of n obtained from Γ by sliding down (or up) some boxes in the opposite direction of the translation α 7→ β. Example 4.1.11 (W = 4 2 ). β k α Tα,β −→ Here only one of the boxes of Γ is involved and the sliding is easily de- In other cases we need to scribed. slide in many steps. See Example 4.1.12 below. 92 CHAPTER4. GENERATINGFUNCTIONANDHILBN ,N +2(C2). Recall that by definition fα,β ∈ HomR (I , R/I ) with I = I Γ and α is a minimal generator for I and β ∈ Γ is such that either β ∈ Pα or β ∈ Qα. The Definition is at 2.1.3. We consider fα,β as a set theoretic map of boxes of ∆. Call l the horizontal distance in number of boxes between β and α, and m the vertical distance in number of boxes between α and β, where a negative distance means that we move either to the left or downwards, so that in Laurent monomials y m xl α = β and in terms of boxes β = α+(−l , m). The choice is such that if β ∈ Pα then l > 0 and m > 0 whereas if β ∈ Qα l < 0 and m < 0. Define B 1 =©γ ∈ Γ¯¯ ∃ δ ∈ ∆ \ Γ with fα,β(δ) = γª and B 1 =Gi B 1 i where the B 1 i 's are the connected components of B 1, and two boxes are connected if they share a side or a corner. We start numbering from top to bottom if β ∈ Pα and from bottom to top if β ∈ Qα. Call α1 the generators of I Γ such that fα,β(α1) ∈ B 1 1 in the lowest row of B 1 1 if β ∈ Pα, resp. in the highest row of B 1 1 if β ∈ Qα. Call β1 = fα,β(α1). We define Tα,β(Γ) by performing different steps. Suppose now, for the rest of the definition that β ∈ Pα. The case β ∈ Qα is completely analogous, in fact one can pass to the transpose Young diagram, perform Tα,β there, and re-transpose everything to get the result. At the step 1 we slide all of B 1 1 by (l , −m). Call what we obtain α,β(Γ) = Γ \ B 1 T 1 1ª . 1 ⊔©γ + (l , −m)γ ∈ B 1 If T 1 α,β(Γ) is Young diagram we stop and we define it Tα,β(Γ). Otherwise we look at all 1 + (l , −m) that are in a row strictly above β1 + (l , −m) and such that γ + (l , −m) is γ ∈ B 1 not already in B 1 1 + (l , −m) and slide all of these by adding again (l , −m). We call what we obtain T 2 α,β(Γ), if it is a Young diagram we stop, otherwise we continue until all the elements of B1 have been slid down the maximum, but always staying above α1. Call this step T s1 α,β(Γ) is a Young diagram, we stop. Otherwise we restart from the generator immediately below of α1 call it α2 and call β2 = fα,β(α2). We define: α,β(Γ). If T s1 B 2 =nγ ∈ T 1 α,β(Γ)¯¯¯ degy (γ) < degy (β1) and ∃ δ ∈ ∆ \ T 1 α,β(Γ) with fα,β(δ) = γo , B 2 =Gi B 2 i where B 2 i are the connected components. We perform all the previous steps for B 2 1 , stopping the first time we get a Young diagram. Again if we do not get a Young diagram when all the boxes above β2 have been slid down the most possible, we define α3 and β3 and B 3 as in the step 2 and keep going until you do reach a Young diagram. Lemma 4.1.14 makes sure that this procedure actually creates a Young diagram Tα,β(Γ). Example 4.1.12 (W = 4 steps before getting to a Young diagram. 2 ). A more complex example. Here we need to perform few 4.1. ATALEOFTWOTORI 93 • ⋆ • ⋆ T 1 •• −→ T 2 •• −→ Here (α, β) = (•, •). B 1 is in yellow, B 1 1 is the upper component of it. T 1 •,•(Γ) is not a Y diagram. To do the next step we need to slide down the (α1, β1) = (⋆, ⋆). orange part. β2 α2 T s1+1 •• −→ β3 . . . −→ α3 ••(Γ) = T s1 This is T 2 •• (Γ). Since it is still not a Y diagram we need to define B 2, in light blue, and slide β2 to α2 . ••(Γ) = T s2 This is T 4 •• (Γ). Since it is still not a Y diagram we need to define B 3, in violet, and slide β3 to α3 . ⋄ ⋄ This is T 6 ••(Γ) = T s4 •• (Γ) the last step. It is a Young diagram. We put in yellow B 1 for (α′, β′) = (⋄, ⋄). If we were to perform Tα′,β′ we would get bet to the original Young diagram Γ. See Lemma 4.1.16 below. One can check the affirmations of Lemmas 4.1.18 and 4.1.19 below in this example and appreciate their geometrical and free of indexes proof. Definition 4.1.10 [Sliding boxes continued] Suppose that Γ ⊢ n and β ∈ Pα ∪ Qα, as above, are fixed. Define ∆n ⊂ ∆ the first n × n boxes in ∆, i.e. where all the Young diagrams Γ ⊢ n live. It is convenient to see Tα,β as a bijection of ∆n+2 to itself. To this end we define Tα,β on ∆n+2 \ Γ referring to the same steps described in the first part of the definition: if δ ∈ ∆n+2 \ Γ and at any point t there is γ ∈ T t α,β (γ) = δ and t was not the last step in the definition of Tα,β, then we pose T t +1 α,β (δ) = γ, otherwise we leave δ invariant i.e. T t +1 α,β(Γ) such that T t +1 α,β (δ) = δ. Notation 4.1.13. Given Γ ⊢ n and v ∈ B (Γ) if v = fα,β we will denote, according to con- venience and context, the transformation defined with any of the following notations Tα,β = T = T fα,β = Tv , and we will say that it a transformation of weight W . We collect all 94 CHAPTER4. GENERATINGFUNCTIONANDHILBN ,N +2(C2). such transformations in a set that we denote ΘW . Then we subdivide {Γ ⊢ n} in orbits for the action of ΘW . {Γ ⊢ n} = Gi =1 Oi . (4.7) Lemma 4.1.14. With the notations as in Definition 4.1.10, T u procedure, is a Young diagram, so that the definition is well posed. α,β(Γ), the last step of the Proof. The fact that Tα,β is well defined is actually equivalent to the fact that fα,β is an R = C[x, y] homomorphism fα,β : I Γ → R±I Γ , so it is equivalent to the fact that β ∈ Pα ∪Qα. We can suppose that β ∈ Pα as the other case is completely analogous. Call T = Tα,β, and let m, l ∈ N be such that α = (−l , m) + β. At any step k of the procedure described in Definition 4.1.10 we call a box (s, t ) outside T k (Γ) an exterior corner if (s, t − 1) and (s − 1, t ) ∈ T k (Γ). Of course at each step the total number of boxes is still n, and we have something that is connected. So we only need to prove that whenever (i , j ) ∈ T u(Γ) then (i −1, j ) and (i , j −1) ∈ T u(Γ). Suppose now by contradiction that there is (i , j ) ∈ T u(Γ) but, for example (i − 1, j ) ∉ T u(Γ). Observe that this force (i − l , j + m) ∈ T u−s (Γ) for some s > 1. If (i − l − 1, j + m) ∉ T u−s (Γ) for any s > 0 we find the contradiction, since then we would have (a multiple of) fα,β((i − 1, j )) = 0 but (a multiple of) fα,β((i , j )) 6= 0 which is absurd since fα,β is an R homomorphism. Then, indeed, for an s′ > 0, (i − l − 1, j + m) ∈ T u−s ′ (Γ). Since by contradiction we supposed (i − 1, j ) ∉ T u(Γ) we have that either (i − l − 1, j + m) is still there at the last step T u or it must slide to (i − 1, j ) and then to (i − l − 1, j + m). But the latter scenario violates the hypothesis that T u was the last step because it means that there is still an attracting exterior corner to the right of (i , j ). Thus at the step T u, the box (i − l − 1, j + m) is still there. Then lets look at (i − 2, j ). It cannot be in T u(Γ) otherwise (i − 1, j ) would have been an attracting exterior corner. Repeating the above argument we find that (i −l −1, j +m) ∈ T u(Γ). Then we can keep going, and we find that (i −l −N , j +m) ∈ T u(Γ) for every N ∈ N, that is absurd. Example 4.1.15. • • Suppose we want to mimic the procedure of Definition 4.1.10 for the couple (•, •) that does not represent an f•,• ∈ B(IΓ). The result is clearly not a Young diagram, as explained in the proof. Lemma 4.1.16. Given n ∈ N, W a wall for n and Γ ⊢ n, every sliding Tα,β with fα,β ∈ B (Γ) of weight W has an inverse of the same form, meaning that there exists fα′,β′ ∈ B (Tα,β(Γ)) 4.1. ATALEOFTWOTORI 95 such that Tα′,β′¡Tα,β (Γ)¢ = Γ. If β ∈ Pα, then β′ ∈ Qα′ and vice-versa. The weight of the inverse transformation is still W . Proof. The proof is immediate from the construction of Tα,β (Γ): suppose αu is, in Def- inition 4.1.10, the last step we need to complete to obtain a Young diagram, then it is sufficient to take α′ = βu = fα,β(αu) and β′ = αu. Observation 4.1.17. If fα,β ∈ B (Γ) is of weight W , then it contributes to s+ s+ W (Γ) ) if and only if β ∈ Pα (resp. β ∈ Qα). W (Γ) (resp. to We are now ready to prove the key properties of the transformations we de- fined on the set of Young diagrams. These are combinatorial properties related to the quantities defined in 4.1.6 even though we will give an easy geometric proof for them. Lemma 4.1.18. Let n ∈ N, W be a wall for n and Γ ⊢ n. Let T be a transformation of Γ with weight W . Then the following two facts hold (1) posW (Γ) = posW (T (Γ)), (2) s+ W (Γ) + s− W (Γ) = s+ W (T (Γ)) + s− W (T (Γ)) . Lemma 4.1.19. Let n ∈ N, W be a wall for n and Γ ⊢ n. Consider the division of {Γ ⊢ n} in orbits for transformations of weight W as in (4.7). Then for every i ∈ N and every k ∈ N we have #©Γ ∈ Oi¯¯s+ W (Γ) = kª = #©Γ ∈ Oi¯¯s− W (Γ) = kª . To prove the previous two lemmas we need the following. Lemma 4.1.20. Let n ∈ N, W be a wall for n and Γ ⊢ n. Consider the division of {Γ ⊢ n} in orbits for transformations of weight W as in (4.7). Let ¡Hilbn(C2)¢TW = Gi Fi be the decomposition in connected components of the fixed points set of Hilbn(C2) for the TW action. Then there is a bijective correspondence between the components (Fi )i and the orbits (Oi )i given by: Fi 7→ {fixed points for the T+ W action on Fi }. Proof. Let Γ ∈ Oi and Tα,β be a transformation of Γ of weight W . Without loss of gener- ality we can assume that β ∈ Pα, otherwise we consider the inverse picture. We denote 96 CHAPTER4. GENERATINGFUNCTIONANDHILBN ,N +2(C2). Tα,β simply T . Observe that γ ∈ ∆n then ω1γ + ω2T (γ) with [ω1 : ω2] ∈ P1 is TW invariant by definition of T . Define ∆Γ,T := ©γ ∈ ∆n+3¯¯γ ∉ Γ or γ ∉ T (Γ)ª . It is clear that all the generators of I Γ and of IT (Γ) are in ∆Γ,T . Then consider the follow- ing P1 of ideals embedded in Hilbn(0) ¡ω1γ + ω2T (γ)¯¯γ ∈ ∆Γ,T¢ [ω1 : ω2] ∈ P1. (4.8) The fact that it is actually a family of ideals in Hilbn(0) is a consequence of the results of Iarrobino Theorem 3.2, by exchanging in the proposition the role of x and y and using the fact that W > 1. The fact that is a TW invariant family is clear since every generator of the ideal is. Finally observe that when [ω1 : ω2] = [0, 1] ∈ P1 then we get I Γ and when [ω1 : ω2] is the point [1, 0] ∈ P1 we get IT (Γ). This proves that whenever two points are in the same orbit Γ, Γ′ ∈ Oi , then they are in the same connected component Fi . Observe however that we proved more: we proved that locally around each fixed point Γ ∈ Oi the dimension of Fi is exactly W (Γ) since the P1's described in (4.8) are all different for different T 's, and the s+ W (Γ) + s− tangent space of¡Hilbn(C2)¢TW at Γ has a basis formed by those vectors that contribute W (Γ). Then, this concludes the proof, and it is worth a separate statement W (Γ) + s− to s+ for future reference. Corollary 4.1.21. With the notation of the previous Lemma we have that, locally, around each point Γ ∈ Oi the dimension of¡Hilbn(C2)¢TW is equal to s+ Now we are ready to give the proof of two main lemmas. W (Γ) + s− W (Γ). Proof of Lemma 4.1.18 and 4.1.19. The two fixed points Γ and Tα,β(Γ) are in the same O orbit, then they are, thanks to Lemma 4.1.20, fixed points in the same compo- nent F . Since Hilbn(C2) is smooth, its fixed points components are smooth, so F is smooth. Then the integers posW (Γ) and posW¡Tα,β(Γ)¢ are the dimension of the pos- itive normal bundle of F in Hilbn(C2), thus they are the same, and s+ W (Γ) and s+ W (Tα,β(Γ)) + s− W (Tα,β(Γ)) are the dimension of the tangent space at, respectively, Γ and Tα,β(Γ) in F , and since F is smooth they, also, are the same. This proves Lemma 4.1.18. W (Γ) + s− Consider F as ambient variety. It is smooth, as said before, and projective, since contained in Hilbn(0). We have an action of T2 on it, and of all its one dimensional subtori. Consider then the TW + action on F : it has finitely many fixed points i.e. the elements of O . Then we can apply Theorem 1.3.4 to obtain an affine cell decompo- sition, where each cell has dimension equal to the positive part of the tangent space at Γ ∈ O of F ; this positive part is exactly s+ W (Γ) is the W (Γ). Recall also that s+ W (Γ) + s− 4.1. ATALEOFTWOTORI 97 dimension of F . Then we can apply Poincaré duality to have for all k ∈ N #©Γ ∈ O¯¯s+ W (Γ) = kª = #©Γ ∈ Oi¯¯s+ that, rearranging, is what we wanted to prove Lemma 4.1.19. W (Γ) = s+ W (Γ) + s− W (Γ) − kª Proof 4.1.22 ( of Proposition 4.1.7). Now we can give a proof that is more convenient for us, of the fact that we can compute the positive part of the tangent spaces at the fixed points with respect to either the torus T∞ or the torus T1+ and obtain the same total polynomials XΓ⊢n q pos∞(Γ) = XΓ⊢n q pos1+ (Γ). In fact we consider the polynomial on the left and we start taking smaller W to exam- inePΓ⊢n q posW (Γ): nothing changes until we hit W1 the first wall for n. Then passing on the other side only the vectors fα,β ∈ B (Γ) of weight W1, for Γ ⊢ n, are affected. Pre- cisely those that contribute to s+ (Γ) stop contributing, i.e. they contribute only on W1 the left of the wall W1 and those that contribute to s− (Γ) start contributing i.e. they W1 contribute only on the right of the wall. Observe that posW +(Γ) = posW (Γ) + s+ posW −(Γ) = posW (Γ) + s− W (Γ) W (Γ). and Then since we were able to group the Young diagrams of size n in sets where the positive part that is not affected by passing the wall is fixed and the part that is ef- fected by passing the wall is symmetric, we are sure that the results before and after the wall are the same. We repeat this for all the walls until we reach W = 1+. Ob- serve that in this way we prove the existence of a map φW on the set {Γ ⊢ n} such that posW +(Γ) = posW −(φW (Γ)) even though we do not suggest that there is a preferred such map, as Example 4.1.23 below shows. Example 4.1.23. We look at a specific orbit O for n = 10 and the wall W = 2. We indicate with a couple •, • (resp. ⋆, ⋆ ) a vector fα,β ∈ B (Γ) of weight W that contributes to s+ W (Γ) (resp. s− W (Γ)). In this example the same box can be marked with both, or with two stars, meaning that two vectors are represented by maps that start or end there. We have the three possibilities (l , m) = (2, 1), (4, 2) and (6, 3) correponding to W = 2 2 and 6 3 . For all Γ depicted we have posW (Γ) = 3 and s+ W (Γ) = 3. In every line the vector W (Γ), s− W (Γ)¢ is constant and its value is indicated under the corresponding line. On the right we put the graph that represents the T•,• 's, between the Γ's, the T⋆,⋆, not depicted, are the inverse arrows. W (Γ)+ s− ¡s+ 1 = 4 98 CHAPTER4. GENERATINGFUNCTIONANDHILBN ,N +2(C2). • • • • • • W (Γ), s− ¡s+ W (Γ)¢ = (3, 0) • • •, ⋆ ⋆ • • ⋆ • ⋆ • W (Γ), s− W (Γ)¢ = (2, 1) ¡s+ ⋆ • ⋆ ⋆ • •, ⋆ ⋆ ⋆ • ⋆ ⋆ W (Γ), s− ¡s+ W (Γ)¢ = (1, 2) • • • • • • • • ⋆ • • ⋆ • • ⋆ ⋆ • ⋆ • ⋆, ⋆ ⋆ ⋆ ⋆ ⋆ W (Γ), s− ¡s+ W (Γ)¢ = (0, 3) Observation 4.1.24. Observe that, through this wall crossing procedure, we just re- proved the statements that says that the attracting sets of Hilbn(0) for the torus T∞ are affine, by supposing that those for the torus T1+ are. In fact an attracting set is isomor- phic to an affine space if and only if it is smooth, thanks to Bialynicki-Birula. Then our wall crossing procedure tells us that locally around each fixed point the dimension of the attracting set is always at least as big as the tangent space to the attracting set. In fact to each linearly independent vector v ∈ B (Γ) of weight W we associated another fixed point given by Tv (Γ) and an invariant P1 that connects the two fixed points and this proves the claim on the local dimension. See also 4.1.21. Remark 4.1.25. We divided the content of Lemma 4.1.18 and 4.1.19 in two different statements to underline this fact: while the first Lemma has a combinatorial proof, however less nice because full of indexes, we were not able to prove the second state- ment only combinatorially. Moreover observe that the fact that s+ W (Γ) is constant along Γ ∈ Oi is equivalent to the fact that Fi is smooth, so that Lemma 4.1.19 can be seen as a geo- metric consequence of the results of Lemma 4.1.18 if this is proved combinatorially. This is what we will do for the case [n, n + 1, n + 2]. W (Γ) + s− Now we describe the little changes we need to implement to pass to the case [n, n +1]. They are mostly about the definition of the sliding transformation associated to elements of the basis of a tangent space: in fact we need to specified how we slide the marked box 1 . 4.1. ATALEOFTWOTORI 99 Definition 4.1.26. [Sliding for [n, n + 1]] Let n ∈ N, W be a wall for [n, n + 1], and Γ = (Γ1, Γ2) ⊢ [n, n + 1] a fixed point for the TW + action on Hilbn,n+1(C2). As usual call α j the box in Γ2 \ Γ1 marked with a 1 . We will either define T (Γ) by performing T (Γ1) and moving α j as an any other element of ∆n+1 \ Γ1, or we will perform T (Γ2) and move α j as any other element of Γ2. Let us see the details. Suppose v ∈ B (Γ1, Γ2) is of weight W . Then either v is of type ( fα,β,Suiv( fα,β)) i ,α j (Γ) simply by i ,α j ). In the latter case we define Tα′ with fα,β ∈ B (Γ1) or v is of type (0, hα′ sliding α j 7→ α′ i , which is always possible since W > 1. α′ 2 1 Tα′ 2,α j −→ 1 W = 3 2 , v ∈ B(Γ1, Γ2) is of the form (0, h), so it involves only two boxes: the slid- ing and all its properties are immedi- ate. If v = ( fα,β,Suiv( fα,β)) we need to distinguish two cases: either α j is not involved in the sliding, meaning that seen as generator of Γ1 it does not attract any box of Γ1 at any step of the procedure that defines Tα,β(Γ1), or it does. In the first case we follow the rules of Tα,β(Γ2): it must be α j 6= α and α is still a minimal generator of Γ2 so we can consider α j as any other box of Γ2 and we perform the sliding Tα,β on Γ2 and define Tα,β(Γ1, Γ2) = Tα,β(Γ2) with the marked box 1 in Tα,β(α j ) seen as any other element of Γ2. 1 • • T•,• −→ 1 W = 2 1 , v ∈ B(Γ1, Γ2) is of the form ( f•,•,Suiv( f•,•)), and αj is not attracting any box in the sliding: we simply treat it as another box of Γ2 and slide everything with the rule T•,•(Γ2). In the second case, i.e. if Tα,β(Γ1) involves sliding a box β j onto α j , we preform Tα,β(Γ1) as if α j was not there and then we put α j where β j was, i.e. we look at α j as any other element of ∆n+1. • ⋆ • 1 T•,• −→ 1 As before Γ1 and f•,• are the same. However, here αj is attracting the box marked with a star, so first we do the sliding T•,•(Γ1) then we put αj where there was the star. 100 CHAPTER4. GENERATINGFUNCTIONANDHILBN ,N +2(C2). For Γ = (Γ1, Γ2) ⊢ [n, n + 1] and v ∈ B (Γ), we will denote the result of one of the transformations defined with any of the following notations: Tv (Γ) = T (Γ) = T (Γ1, Γ2) = (T (Γ1), T (Γ2)) depending on the aspect of the sliding we want to underline, and we see T as a bijec- tion on ∆n+1 to itself as in Definition 4.1.10. Having clarified the transformations we perform in this case, all the remaining steps are as before. The interpretation of the following Lemma 4.1.28 in equations (4.10) and (4.11) is what we need to deal with the case [n, n + 1, n + 2]. Observation 4.1.27. Let n ∈ N and W be a wall for [n, n + 1]. The transformations of 4.1.26 are well defined and every transformation is invertible. They partition the set of (Γ1, Γ2) ⊢ [n, n + 1] in orbits { (Γ1, Γ2) ⊢ [n, n + 1] } = Gi =1 Oi ; (4.9) where an orbit Oi is closed under the action of transformations T of weight W . Lemma 4.1.28. Let n ∈ N, W be a wall for [n, n + 1] and Γ ⊢ [n, n + 1]. Let T be a transfor- mation of Γ with weight W . Then the following two facts hold (1) posW (Γ) = posW (T (Γ)), (2) s+ W (Γ) + s− W (Γ) = s+ W (T (Γ)) + s− W (T (Γ)) . Lemma 4.1.29. Let n ∈ N, W be a wall for [n, n + 1] and Γ ⊢ [n, n + 1]. Consider the division of {Γ ⊢ [n, n + 1]} in orbits for transformations of weight W as in (4.9). Then for every i ≥ 1 and every k ∈ N we have #©Γ ∈ Oi¯¯s+ W (Γ) = kª = #©Γ ∈ Oi¯¯s− W (Γ) = kª . To prove the previous two lemmas we need the following. Lemma 4.1.30. Let n ∈ N, W be a wall for [n, n + 1] and Γ ⊢ [n, n + 1]. Consider the division of {Γ ⊢ [n, n + 1]} in orbits for transformations of weight W as in (4.9). Let ³Hilbn,n+1(C2)´TW = Gi Fi be the decomposition in connected components of the fixed points set of Hilbn,n+1(C2) for the TW action. Then there is a bijective correspondence between the components (Fi )i and the orbits (Oi )i given by: Fi 7→ {fixed points for the T+ W action on Fi }. 4.1. ATALEOFTWOTORI 101 Proof. The proof of the complete bundle of results is completely analogous to that for Hilbn(C2). In the proof of Lemma 4.1.30 for each v ∈ B (Γ1, Γ2) with weight W we can, again, easily write down a TW invariant P1 connecting (Γ1, Γ2) and Tv (Γ1, Γ2) with the same exact recipe. We can still use that the Fi are smooth, since the ambient space Hilbn,n+1(C2) is. Then the quantity posW (Γ) is the dimension of the positive normal bundle, and thus the same for each Γ ∈ Fi fixed point. The quantity s+ W (Γ) is also the same for every Γ since it is the dimension of the tangent space of Fi at Γ ∈ Fi . Moreover the number of such Γ with given s+ W (Γ) again thanks to Poincaré duality for the Fi 's. W (Γ) is equal to the number of such Γ with given s− W (Γ) + s− We want to spell out what points (1) and (2) of Lemma 4.1.28 mean in terms of Definition 2.2.10 of B (Γ1, Γ2) . Remark 4.1.31. Use the notation of Definition 4.1.26. Call αi , i = 0, . . . , s the generators of Γ1, α j the box in Γ2 \ Γ1 and βi , i = 0, . . . , t the generators of T (Γ1) and βh the box in T (Γ2) \ T (Γ1). Since B (Γ1, Γ2) is constructed by elements of the form (0, hαi ,α j ) and ( fα,β,Suiv( fα,β)) for those α, β that are not in Obs(Γ1, Γ2) and analogously for T (Γ1, Γ2) we have by Lemma 4.1.28 posW (Γ1) − posW (Obs(Γ1, Γ2)) + posW ({hαi ,α j i = 0, . . . , s}) = = posW (Γ1, Γ2) =posW (T (Γ1, Γ2)) = = posW (T (Γ1)) − posW (Obs(T (Γ1, Γ2))) + posW ({hβi ,βh i = 0, . . . , t }). Since we know from Lemma 4.1.18 that posW (Γ1) = posW (T (Γ1)) we obtain that posW (Obs(Γ1, Γ2)) − posW ({hαi ,α j i = 0, . . . , s}) =posW (Obs(T (Γ1, Γ2))) − posW ({hβi ,β j i = 0, . . . , t }). (4.10) Analogously, for s+ W and s− W we obtain that (s+ W + s− W )¡Obs(Γ1, Γ2)¢ − (s+ = (s+ W + s− W + s− W )¡{hαi ,α j i = 0, . . . , s}¢ W + s− W )¡Obs(T (Γ1, Γ2))¢ − (s+ W )³{hβi ,β j i = 0, . . . , t }´ . (4.11) Observation 4.1.32. Notice this important fact: we can write (4.10) and (4.11) in terms of the generators of Γ1 and T (Γ1) and not in term of the generators of Γ2 and T (Γ2) because we are only interested in the positive part, and since W > 1 there is not differ- ence in the two: for example if α j is of case 3), according to cases 2.2.2, then the list of αi and that of α′ i differ only around α j and in particular we have two new generators xα j and yα j , but we have that¡0, hxα j ,α j¢ and¡0, hyα j ,α j¢ cannot be possibly positive 102 CHAPTER4. GENERATINGFUNCTIONANDHILBN ,N +2(C2). for any W . Similarly for the other cases. This observation saves us a lot of work in distinguishing cases later. Now we are well placed to use equation (4.10) and (4.11) for the next step i.e. [n, n + 1, n + 2]. We start by spelling out the definitions of the transformation associ- ated to a vector of weight W in the tangent spaces of fixed points, even though the definitions are, in this case, exactly as in the previous case. Definition 4.1.33. Sliding for [n, n + 1, n + 2] Let n ∈ N, W be a wall for [n, n + 1, n + 2], and Γ = (Γ1, Γ2, Γ3) ⊢ [n, n + 1, n + 2] be a fixed point for the TW + action on Hilbn,n+1,n+2(C2). As usual call α j the box in Γ2 \ Γ1 marked with a 1 and α′ l the box in Γ3 \ Γ2 marked with 2 . Suppose v ∈ B (Γ1, Γ2, Γ3) is of weight W . We either have v of type (0, 0, hα′′ ) or v i ,α′ of type³0, hα′ fα,β ∈ B (Γ1). i ,α j ,Suiv(hα′ i ,α j )´ or, finally, v of type¡ fα,β,Suiv( fα,β),Suiv(Suiv( fα,β))¢ with l In the first case we define Tα′′ i , which is always possible since W > 1. This is exactly as if we were looking at the transformation of Γ3 without caring for the marked boxes. i ,αl (Γ) simply by sliding αl 7→ α′′ α′′ 2 2 1 Tα′′ 2 ,α′ l −→ 2 1 W = 3 1 , v ∈ B(Γ1, Γ2, Γ3) is of the form (0, 0, h), so it involves only two boxes: the sliding and all its properties are immediate. i ,α j ,Suiv(hα′ If v =³0, hα′ i ,α j )´ we need to distinguish two cases: If α′ l we sim- ply switch α j and α′ l as it would have happen following the rule described in 4.1.26: we are performing the transformation for Γ2 and looking at α′ l as any other element of ∆n+2 \ Γ2. If α′ i ,α j of Γ3 without caring for the marked boxes. Observe that we move more than one box only if it it happens that α′ l is on the same row or column of α j . In this case we might need to slide more than two boxes. i 6= αl again we slide according to the transformation Tα′ i = α′ 1 2 α′ 1 Tα′ 1,α j −→ 1 2 i ,α j W = 2 ,Suiv(hα′ for Γ3 without special meanings for 1 and 2 . 1 , v ∈ B(Γ1, Γ2, Γ3) is of the form³0, hα′ )´, we think of everything happening Finally v =¡ fα,β,Suiv( fα,β),Suiv(Suiv( fα,β))¢. Again the definition goes as for the case [n, n + 1]: we either perform T (Γ1) or T (Γ2) or T (Γ3) and look at α j as in ∆n+2 \ Γ1 or in Γ2 and at α′ l as in ∆n+2 \ Γ2 or in Γ3 depending on whether some box slide onto them or not. i ,α j 4.1. ATALEOFTWOTORI 103 1 2 • • T•,• −→ 1 2 W = 2 j are attracting any box in the sliding: we simply treat them as any another boxes and slide 1 , v ∈ B(Γ1, Γ2, Γ3) is of the form¡ f•,•,Suiv( f•,•),Suiv(Suiv( f•,•))¢, and nor αj nor α′ everything with the rule T•,•(Γ3). • • 2 1 T•,• −→ 2 1 Same situation as before meaning that Γ1 and f•,• are the same. Here αj and α′ attracting a respective box: first we do the sliding T•,•(Γ1) and then we put αj and α′ l are each l in the boxes that occupied their respective old positions. For Γ = (Γ1, Γ2, Γ3) ⊢ [n, n + 1, n + 2], and v ∈ B (Γ), we will denote the result of one of the transformations defined with any of the following notation: T (Γ) = Tv (Γ) = T (Γ1, Γ2, Γ3) = (T (Γ1), T (Γ2), T (Γ3)) = . . . depending on the aspect of the sliding we want to underline. Again we see Tv as a transformation of ∆n+2 as in definitions 4.1.10 and 4.1.26. Now we are well placed to prove all the analogous results we proved in the pre- vious cases. Observation 4.1.34. Let n ∈ N and W be a wall for [n, n + 1, n + 2]. The transformations of 4.1.33 are well defined and every transformation is invertible. They partition the set of (Γ1, Γ2, Γ3) ⊢ [n, n + 1, n + 2] in orbits { (Γ1, Γ2, Γ3) ⊢ [n, n + 1, n + 2] } = Gi =1 Oi ; (4.12) where Oi is closed under the action of transformations T of weight W . Lemma 4.1.35. Let n ∈ N, W be a wall for [n, n + 1, n + 2] and Γ ⊢ [n, n + 1, n + 2]. Let T be a transformation of Γ with weight W . Then the following two facts hold (1) posW (Γ) = posW (T (Γ)), (2) s+ W (Γ) + s− W (Γ) = s+ W (T (Γ)) + s− W (T (Γ)) . Proof. We use Observation 4.1.32 to cut the number of possible cases we need to dis- tinguish between: in fact we want to use equations (4.10) and (4.11) separately on α j and α′ l . 104 CHAPTER4. GENERATINGFUNCTIONANDHILBN ,N +2(C2). Call αi , i = 0, . . . , s the generators of Γ1, α j the box in Γ2 \ Γ1 and α′ l the box in Γ3 \ Γ2. Call as well βi , i = 0, . . . , t the generators of T (Γ1), βh the box in T (Γ2) \ T (Γ1) and β′ k the box in T (Γ3) \ T (Γ3). Suppose now that α′ l is not on the same row or column as α j . Suppose also j < l , the other cases is treated in a completely similar way (in fact one can pass to the transpose to treat it.). Given the definition of B (Γ1, Γ2, Γ3) we can write that posW (Γ1, Γ2, Γ3) = l + + + if if we are in case j 1b) or j2) if we are in case j 1a) or j3), posW (Γ1) − posW¡Obs(Γ1, Γ2)¢ − posW¡Obs(Γ1, Γ3)¢ + posW¡{hαi ,α j i = 0, . . . , s}¢ + posW³{hαi ,α′ i = 0, . . . , s}´ + posW³α j −1 7→ αl y p j −1´ − posW {¡α j 7→ αl¢}  posW¡yα j 7→ αl¢ posW¡xα j 7→ αl¢ posW¡xα j 7→ αl¢ + posW¡yα j 7→ αl¢ 0 0 + + if we are in case j 1 a), if we are in case j 1 b), if we are in case j 2), if we are in case j 3). The cases we refer to are from Definition 2.2.2. The last three terms are there to compensate the choice of writing the other terms only with respect to the generators of Γ1. Applying now Equation (4.10) twice to α j and to αl and the fact that posW (Γ1) = pos(T (Γ1)), we see that we only need to deal with the last 3 terms, and prove that they contribute the same amount before and after the sliding T . Call this contribution A( j , l ). Since positiveness of A( j , l ) depends only on the relative positions of α j and αl we can suppose that αl is fixed by T and l is higher than α j and T (α′ that only α j is affected by T . Then there are three cases: α′ l ) is higher than T (α j ), or α′ l ) and α j is higher than αl and T (α j ) is higher than T (α′ l ). We deal with the first case the others are extremely similar. l is higher than α j but T (α j ) is higher than T (α′ There are three possibilities: if α j also is fixed by T but its case is different after T ; if α j is slid by T as an element of Γ2; and finally if there is a β j ∈ Γ1 that slid to α j with T . Since A( j , l ) now depends only on the case of α j we can reduce the cases we need to analyze to 16: the cases of α j times the cases T (α j ). However, every single case is immediate and ultimately they are all similar to one another. We do for example the case j 2) and T ( j )1a). Here we have that the following vectors have weights that coincide: (in the first column we look at ∆n+2 before T , on the second column we look at ∆n+2 after T . We use the notation α − 1 to mean the 4.1. ATALEOFTWOTORI generator immediately below α.) ¡yα j 7→ αl¢ ←→¡yT (α j ) 7→ αl¢ ; ¡T (α j ) 7→ αl¢ ←→¡α j 7→ αl¢ ; ¡xα j 7→ αl¢ ←→µT (α j ) − 1 7→ y αl pT (α j )−1¶ . 105 (4.13) All other vectors remain the same before and after T . That proves that the weight of A( j , l ) remains the same after the transformation T . All other cases are completely similar. αl ... αl ... • α j ⋄ ⋆ ⋆ . . . • α j ⋄ . . . An example of the case j 2) T(j) 1a). The symbols are the generators that correspond to one another in 4.13. W and s− The proof for s+ W is completely analogous. In fact if one applies equa- tion (4.11) everything one needs to analyze is the the weight of the same term A( j , l ) as above. But then for every cases, there is a conservation of the total quantities of weights of the vectors that change before and after T . Lemma 4.1.36. Let n ∈ N, W be a wall for [n, n +1, n +2] and Γ ⊢ [n, n +1, n +2]. Consider the division of {Γ ⊢ [n, n +1, n +2]} in orbits for transformations T of weight W as in (4.7). Let ³Hilbn,n+1,n+2(C2)´TW = Gi Fi be the decomposition in connected components of the fixed points set of Hilbn,n+1,n+2(C2) for the TW action. Then there is a bijective correspondence between the components (Fi )i and the orbits (Oi )i and it is given by: Fi 7→ {fixed points for the T+ W action on Fi }. 106 CHAPTER4. GENERATINGFUNCTIONANDHILBN ,N +2(C2). Moreover the Fi are smooth. Proof. The proof that there is a bijection between orbits of transformations T of weight W and fixed points components for TW goes exactly as before, since, as before, we can construct for each such T a TW invariant P1 connecting Γ and T (Γ). This time however, we already know that s+ W (Γ) is constant for all Γ ∈ Oi thanks to the combinatorial proof of Lemma 4.1.35. Then we have smoothness of Fi , even though the ambient space is not smooth. W (Γ) + s− Lemma 4.1.37. Let n ∈ N, W be a wall for [n, n +1, n +2] and Γ ⊢ [n, n +1, n +2]. Consider the division of {Γ ⊢ [n, n + 1, n + 2]} in orbits for transformations of weight W as in (4.9). Then for every i ≥ 1 and every k ∈ N we have #©Γ ∈ Oi¯¯s+ W (Γ) = kª = #©Γ ∈ Oi¯¯s− W (Γ) = kª . Proof. Thanks to Lemma 4.1.36 we know that the Fi 's are smooth. Then we can apply Poincaré duality as in the proof of Lemma 4.1.19 to conclude. Proposition 4.1.38. Let n ∈ N and let W be a wall for [n, n +1, n +2]. Then for every k ∈ N we have In particular we have that the following two polynomials are the same: #©Γ ⊢ [n, n + 1, n + 2]¯¯posW + (Γ) = kª = #©Γ ⊢ [n, n + 1, n + 2]¯¯posW − (Γ) = kª . XΓ ⊢ [n,n+1,n+2] XΓ ⊢ [n,n+1,n+2] q pos∞(Γ) = q pos1+ (Γ) . (4.14) Proof. The proof is exactly the same as the proof of 4.1.22, using Lemmas 4.1.35 and 4.1.37. Observation 4.1.39. Observe that we just proved that the attracting sets for the torus action T∞ are affine cells as well. This follows thanks to a reasoning completely similar to the one presented in Observation 4.1.24. Now we need to find a generating function for the polynomials on the left hand side of (4.14) when we consider all n ∈ N. This is the goal of the next section. 4.2 Hilbn,n+2(C2) We turn now our attention to Hilbn,n+2(C2). By imposing a specific condition on the two ideals of length n and n +2 of the flag, we define a smooth subspace Hilbn,n+2(C2)t r of Hilbn,n+2(C2) whose cell decomposition is studied by Nakajima and Yoshioka [NY08, Chapter 5]. It turns out that the generating function for the Poincaré polynomials 4.2. HILBN ,N +2(C2) 107 of this smooth space resemble closely to the formula (4.1). The similarity was no- ticed almost accidentally: the construction of the analogue subspace in the case of Hilbn,n+1(C2) gives back the full Hilbn,n+1(C2) and reproduces the results originally of Cheah. To prove our formula (4.1) we thus simply need to match the combinatorics of the formula proved by Nakajima and Yoshioka. The question about some geometrical connections between the two families of spaces, however, remains mysterious . From the affine cell decomposition we found for Hilbn,n+1,n+2(0), we are then able to find an affine cell decomposition of Hilbn,n+2(0) and a generating function for its Poincaré polynomials (4.21). Definition 4.2.1. Let (J , I ) be a point of Hilbn,n+2(C2). We say that J , I are trivially re- lated if J /I ∼= C2 as trivial C[x, y] modules. Said it otherwise J /I ⊆ Ker"x y# : C[x, y]/I → C2 ⊗ C[x, y]/I , i.e. x f ∈ I and y f ∈ I for all f ∈ J. We then divide Hilbn,n+2(C2) into two parts, those couples that are trivially related and those that are not Hilbn,n+2(C2) = Hilbn,n+2(C2)t r G Hilbn,n+2(C2)N t r . We will see below more details of the following proposition, but we state it here to motivate the definition of this space. Proposition 4.2.2. [NY08, Proposition 5.2, Corollary 5.4] The space Hilbn,n+2(C2)t r is smooth for every n in N. Moreover it has a cellular decomposition given by the torus action and the Poincaré polynomials have generating function: Xn≥0 Pq³Hilbn,n+2(0)t r´ t n = 1 (1 − t q)(1 − t 2q 2) 1 1 − q d −1t d . ∞Yd =1 (4.15) We observe the following connection between Hilbn,n+1,n+2(0) and Hilbn,n+2(0). Note however that this is not the connection that explains the similarities between the gen- erating function: for that we would need to relate Hilbn,n+1,n+2(0) and Hilbn+1,n+3 (0) t r Lemma 4.2.3. Consider the projection on first and third factor, i.e. the map that forgets the middle element of a three steps flag p1,3 : Hilbn,n+1,n+2(0) → Hilbn,n+2(0). Then p −1 1,3 ((J , I )) ∼=  P1 {pt} if (J , I ) ∈ Hilbn,n+2(0)t r if (J , I ) ∈ Hilbn,n+2(0)N t r . 108 CHAPTER4. GENERATINGFUNCTIONANDHILBN ,N +2(C2). We call Hilbn,n+1,n+2(0)t r those triples projecting to Hilbn,n+2(0)t r , and Hilbn,n+1,n+2(0)N t r those triples projecting to Hilbn,n+2(0)N t r . The projection restricted to Hilbn,n+2(0)t r p1,3 : Hilbn,n+1,n+2(0)t r → Hilbn,n+2(0)t r is a Zariski locally trivial bundle. Proof. Let (J , I ) ∈ Hilbn,n+2(0). Write I = 〈 f1, f2, . . . , fN 〉 for a linear basis of I seen as a vector space, i.e. I ∈ Gras(N , R±mn+2 ), where R = C[[x, y]], m is the maximal ideal of R and N + n + 2 = (n+2)(n+3) look at the Grassmannian of possible vector spaces V of dimension N + 1 such that . Extend the basis of I to one of J as J = 〈g1, g2, f1, . . . , fN 〉. We 2 J ⊃ V ⊃ I . It is clear that this is isomorphic to a P1. We need to prove that if (J , I ) ∈ Hilbn,n+2(0)t r then all such vector spaces are actually ideals, and if (J , I ) ∈ Hilbn,n+2(0)N t r only one such vector space is an ideal. Remember that an ideal is a subvector space of R closed by multiplication for x and y. Suppose first (J , I ) ∈ Hilbn,n+2(0)t r . Then by definition we have xgi ∈ I and y gi ∈ I for i = 1, 2. Then if V = 〈ω1g1 + ω2g2, f1, . . . , fN 〉 we have that x(ω1g1 + ω2g2) ∈ I ⊂ V and y(ω1g1 + ω2g2) ∈ I ⊂ V , proving that all possible choices of V give an ideal. Suppose now that (J , I ) ∈ Hilbn,n+2(0)N t r .Then by definition there must be a choice of [ω1 : ω2] ∈ P1 such that x(ω1g1 + ω2g2) ∉ I or y(ω1g1 + ω2g2) ∉ I . Up to a linear chang of coordinates on g1 and g2 we can then suppose that xg1 ∉ I while y g1, xg2, y g2 ∈ I . Then we see that the only choice of V that is an ideal is V = 〈g2, f1, . . . , fN 〉, as for all ω2 ∈ C we have that 〈g1 + ω2g2, f1, . . . , fN 〉 is not closed by multiplication or it would be N + 2 dimensional. To prove the last claim observe that the map p1,3 is simply the restriction to the space of ideals of the standard projection map defined at the level of flag varieties. Then it is a Zariski fibration since the map between flag varieties is. Remark 4.2.4. We now want to prove that the Betti numbers of Hilbn,n+1,n+2(0)t r co- incide with those of Hilbn−1,n,n+1(0). Note the shift in the length of the ideals involved. The geometric relationship between the two spaces remains unclear. Observation 4.2.5 (Fixed points Hilbn,n+2(C2)t r ). The fixed points of Hilbn,n+2(C2)t r are in bijection with Young diagrams Y of n + 2 with 2 removable boxes marked. A removable box is a corner, i.e. a box (i , j ) ∈ Y such that (i + 1, j ) ∉ Y and (i , j + 1) ∉ Y . We call such an object (Y , S) where Y is the Young diagram and S is the set of marked boxes of cardinality 2. 4.2. HILBN ,N +2(C2) 109 C A Young diagram of size n + 2 with two removable boxes marked in black. The box with the C is the only box on the same column and on the same row of a marked box. The fact that the two marked boxes cannot be in the same row or column is a conse- quence of the fact that we ask the monomial ideals J , I to be trivially related. We call relevant all the boxes of Y that are not marked and are not on the same row and the same column of a marked box. In the picture above the relevant boxes are the white, empty ones. Proposition 4.2.6. [NY08, Corollary 5.3] Let T∞ acts on Hilbn,n+2(C2)t r . Then the pos- itive part of the tangent space of Hilbn,n+2(C2)t r at a fixed point (Y , S) has dimension given by pos∞ ((Y , S)) = n + 1 − ℓ(Y ). (4.16) The Poincaré polynomial of Hilbn,n+2(C2)t r is Pq³Hilbn,n+2(C2)t r´ = X(Y ,S) q pos∞((Y ,S)), where the sum is on all torus fixed points. Observe that, even though Y is a partition of n + 2, the formula compares the number of its columns with n + 1 and not n + 2. A combinatorial correspondence We want to relate the parametrization of the fixed points of Hilbn−1,n,n+1(C2) with that of Hilbn,n+2(C2)t r . More precisely consider the two sets An and Bn consisting of An: Fixed points of Hilbn−1,n,n+1(C2), Γ = (Γ1, Γ2, Γ3). We keep the notations of the l and the box previous chapters and we call the box Γ3 \ Γ2 marked with 2 as α′ Γ2 \ Γ1 marked with 1 as α j . Bn: A couple (Y , S) of a Young diagram Y of size n + 2 and S a subset of marked re- movable boxes of Y of size 2. The goal is to construct a 2 to 1 map bn : An → Bn such that if b−1 n ((Y , S)) = {Γ, H } and pos∞ ((Y , S)) = r pos∞ (Γ) = r and pos∞ (H ) = r + 1 then (4.17) 110 CHAPTER4. GENERATINGFUNCTIONANDHILBN ,N +2(C2). where the integers pos∞ ((Y , S)) and pos∞ (Γ) are calculated combinatorially as in for- mulae (4.16) and (4.3) respectively. To construct the map bn : An 2:1 −−→ Bn we actually start from an element of Bn and create two elements of An with a procedure that we show is both injective and surjec- tive. Let λ ⊢ n be a Young diagram of size n. We consider all possible (Y , S) ∈ Bn with Y \S = λ. We say that all such elements of Bn are at level λ. Call, as usual, α0, . . . , αs the standard monomial generators of Iλ. Then we denote with (αi , αk ), with i > k, the element (Y , S) of Bn given by Y = λ ∪ {αi } ∪ {αk } and S = {αi } ∪ {αk } i.e. with αi , αk marked in black. Observe that we have s(s+1) such choices. We order the elements of Bn at level λ in this way 2 (αi , αk ) > (αr , αt ) with i > k, r > t if and only if i > r or, i = r, k > t .  It is clear that varying λ ⊢ n we obtain all of the elements of Bn. α3 α2 α1 α0 > The Young diagram λ is in white and in gray there are the αi s between which we need to chose two boxes to create an element of Bn of level λ. The element of Bn represented as (α3, α0) is bigger then the one represented as (α2, α1) Given λ ⊢ n we denote β0, . . . , βs−1 the corners of λ. We start numbering from the right. β2 β1 β0 The corners of λ ⊢ n. Recall that a corner is box in a Young diagram λ such that there are no other boxes of λ on its row to the right nor on its column above. We will denote (αi , βr ) the element Γ = (Γ1, Γ2, Γ3) of An given by Γ1 = λ, Γ2 = λ ∪{βr } and Γ3 = λ ∪ {βr } ∪ {αi }. This is the same as marking with 1 the box βr and with 2 the box αr . We say that all elements in An that arise in this way are at level λ. It is clear that varying λ ⊢ n we obtain all of the elements of An. Definition 4.2.7. Let λ ⊢ n. We utilize all the notations as specified above. Given a couple of generators of Iλ , (αs−i , αs−k ), with 0 ≤ i < k ≤ s, we associate to it a couple of corners of λ, precisely (βk−i −1, βs−i −1) . Then to the element (αi , αk ) ∈ Bn at level λ we Observation 4.2.8. Maybe the above map is more clear if we interpret it in terms of the order we have introduced on Bn. More precisely, if (αs , αs−1) > (αs , αs−2) > (αs , αs−3) > . . . > (αs , α0) > > (αs−1, αs−2) > (αs−1, αs−3) > . . . > (αs−1, α0) > . . . > (α2, α1) > (α2, α0) > > (α1, α0) is the ordered list of all elements of Bn of level λ, then we associate to them, respec- tively, ©(αs , β0), (αs−1, βs−1)ª ,©(αs , β1), (αs−2, βs−1)ª , . . . ,©(αs , βs−1), (α0, βs−1)ª , ©(αs−1, β0), (αs−2, βs−2)ª , . . . ,©(αs−1, βs−2), (α0, βs−2)ª , . . . ©(α2, β0), (α1, β1)ª ,©(α2, β1)(α0, β1)ª , ©(α1, β0)(α0, β0)ª . 4.2. HILBN ,N +2(C2) associate the two elements of An at level λ given by: (αi , αk ) ∈ Bn b−1 n−−→©(αs−i , βk−i −1), (αs−k , βs−i −1)ª ⊂ An. 111 (4.18) Example 4.2.9. Below we depict an example. On the left we have the first element of Bn at level λ and on the right the two elements of An of level λ in its preimage along bn. Observe that pos∞ ((α3, α2)) = pos∞¡(α3, β0)¢ − 1 = pos∞¡(α2, β2)¢. 2 1 2 7→ 1 , The example below, instead, shows b−1 = pos∞¡(α0, β1)¢ = pos∞¡(α2, β1)¢ − 1. n ((α2, α0)) =©(α2, β1), (α0, β1)ª. Observe that pos∞ ((α2, α0)) 7→ 2 1 1 , 2 Lemma 4.2.10. The map bn : An → Bn of Definition 4.18 is two to one and satisfies the condition (4.17). Proof. It is clear that bn is well defined and two to one at level λ for each λ ⊢ n. It is also clear that it is defined on all of An and surjective. In fact all elements of Bn 112 CHAPTER4. GENERATINGFUNCTIONANDHILBN ,N +2(C2). and An arise in the way we described and are involved in the construction. To prove that is well defined at different levels λ and µ it is sufficient to observe that if there exist β1 corner of λ, β2 corner of µ, α1 generator of Iλ and α2 generator of Iµ such that (α1, β1) = (α2, β2), then necessarily α1 = α2 and so λ = µ implying that we are actually at the same λ level. Then we only need to prove that condition (4.17) is satisfied. To do so observe that: n + 1 − ℓ(λ) n − ℓ(λ) for all i > k > 0, for all i > k = 0 n + 1 − ℓ(λ) n + 1 − ℓ(λ) + 1 n − ℓ(λ) for all r ≥ i − 1 > −1, for all i > 0, i − 1 > r, for all i = 0. whereas pos∞ ((αi , αk )) = pos∞¡(αi , βr )¢ = Then looking at the definition of b−1 so that k − i − 1 ≥ s − i − 1 and then n (αs−i , αs−k ) in (4.18) it is clear that : either s − k = 0 or s − k > 0 and then s − i > k − i − 1 so that, again, pos∞¡(αs−i , βk−i −1)¢ = pos∞ ((αs−i , α0)) + 1 = pos∞¡(αs−k , βs−i −1)¢ + 1; pos∞¡(αs−i , βk−i −1)¢ = pos∞ ((αs−i , α0)) + 1 = pos∞¡(αs−k , βs−i −1)¢ + 1. This completes the proof. Proposition 4.2.11. The Poincaré polynomials of Hilbn,n+1,n+2(0) satisfy Xn≥0 Pq³Hilbn,n+1,n+2(0)´ zn = q + 1 (1 − zq)(1 − z2q 2) Ym≥1 1 1 − zm q m−1 . (4.19) Proof. We proved in Proposition 3.2.14 that Hilbn,n+1,n+2(0) has an affine paving, given by the attracting sets for the torus T1+, and that the dimensions of these affine cells are given by the positive parts of the tangent spaces. In the previous section, Proposition 4.1.38, we proved that the Poincaré polynomial for every Hilbn,n+1,n+2(0) can be com- puted by using the positive parts of the tangent spaces at fixed points with respect to the torus T∞ and given by formula (4.3). Then, the existence of map bn with property (4.17), as proved in Lemma 4.2.10, shows that Pq (Hilbn,n+1,n+2(0)) = (q + 1)Pq (Hilbn+1,n+3(0)t r ) thanks to Proposition 4.2.6. Finally we use equation (4.15) to conclude. 4.2. HILBN ,N +2(C2) 113 Now we see how we get for free a cell decomposition of Hilbn,n+2(0), a combinatorial formula for its Poincaré polynomial and a generating function for all the Poincaré polynomials of Hilbn,n+2(0) for different n ∈ N. Lemma 4.2.12. Let Γ = (Γ1, Γ2, Γ3) ⊢ [n, n + 1, n + 2] be a fixed point of Hilbn,n+1,n+2(0)N t r . Call AΓ ⊂ Hilbn,n+1,n+2(0) the attracting affine cell with respect to the torus action T1+. If I = (I1, I2, I3) ∈ AΓ, then I ∈ Hilbn,n+1,n+2(0)N t r . l are not in I Γ3, suppose xα′ Proof. By hypothesis Γ = (Γ1, Γ2, Γ3) ∈ Hilbn,n+1,n+2(0)N t r . This means that either xα′ l or yα′ ∉ I Γ3. Since limt →0 t · I1 = I Γ1 there exists f ∈ I1 such that limt →0 t · f = α′ l , if x f ∈ I3 we would have an absurd. Thus I = (I1, I2, I3) is not trivially related. l . Since limt →0 t · x f = xα′ l Observation 4.2.13. Notice that a similar statement for Γ = (Γ1, Γ2, Γ3) fixed point of Hilbn,n+1,n+2(0)t r is false. Consider for example the fixed point given by the nested monomial ideals (x, y 2) ⊃ (x2, x y, y 2) ⊃ (x2, x y, y 3). The attracting cell is parametrized as ∼= A3 ∼=((x + ωy, y 2) ⊃ ⊃ (x2 + αy 2, x y + βy 2, y 3)) (x2, x y, y 2) AΓ 2 Γ = 1 One can see that all of the points in AΓ with α 6= −ω2 or β 6= ω are in Hilbn,n+1,n+2(0)N t r . Lemma 4.2.14. Consider the T1+ action on R = C[x, y]. Let (Γ1, Γ3) with Γ1 ⊢ n and Γ3 ⊢ n + 2 be a fixed point of Hilbn,n+2(0)t r . Call {α j , αl } = Γ3 \ Γ1 with α j = limt →0 t · (α j + αl ). We define Γ := (Γ1, Γ2, Γ3) with Γ2 := Γ1 ∪ {α j } and denote AΓ ⊂ Hilbn,n+1,n+2(0) the attracting affine cell of Γ. Then if AΓ1,Γ3 ⊂ Hilbn,n+2(0)t r is the attracting set of the fixed point Γ1, Γ3 we have that: p1,3 : AΓ ∼= AΓ1,Γ3 where p1,3 : Hilbn,n+1,n+2(0) → Hilbn,n+2(0) is the projection on the first and third factor. Proof. By Lemma 4.2.3 we have that p −1 1,3(I Γ1, I Γ3) =©I Γ1 ⊃ I Γ1 +¡ω1α j + ω2αl¢ ⊃ I Γ3 ¯¯ [ω1 : ω2] ∈ P1ª ⊂ Hilbn,n+1,n+2(0). Since p1,3 : AΓ → AΓ1,Γ3 is clearly surjective we only need to prove that it is injective. Let I = (I1, I2, I3), J = (J1, J2, J3) ∈ AΓ, with p1,3(I ) = p1,3(J ). Then I = J. In fact let f j , fl ∈ I1 = I3 be such that f j , fl ∉ I3 = J3 and limt →0 t · fi = αi for i = j , l . Then ω1 f j +ω2 fl ∈ I2 if and only if [ω1 : ω2] = [0 : 1]: otherwise the hypothesis α j = limt →0 t · (α j + αl ) implies α j ∈ I Γ2 that is an absurd by definition. The same is true for J2 proving J2 = I2 and thus I = J. 114 CHAPTER4. GENERATINGFUNCTIONANDHILBN ,N +2(C2). Definition 4.2.15. We write Γ = (Γ1, Γ3) ⊢ [n, n + 2] for a couple of nested Young dia- grams of size n and n + 2 respectively. We define a map sn : {Γ ⊢ [n, n + 2]} → {Γ ⊢ [n, n + 1, n + 2]} as sn(Γ1, Γ3) = (Γ1, Γ2, Γ3) where Γ2 := Γ1 ∪©α ∈ Γ3 \ Γ1 ¯¯ α has minimal degree and minimal y degree in Γ3 \ Γ1ª . Proposition 4.2.16. Let n ∈ N. The space Hilbn,n+2(0) has an affine cell decomposition indexed by Γ ⊢ [n, n + 2]. Its Poincaré polynomial is given by: Pq³Hilbn,n+2(0)´ = XΓ⊢[n,n+2] q pos1+ (sn (Γ)) (4.20) where the quantity pos1+ (sn(Γ)) is specified in Formula 3.23 of Proposition 3.2.13. The Poincaré polynomials satisfy: Xn≥0 Pq³Hilbn,n+2(0)´ zn = 1 + q − q z (1 − zq)(1 − z2q 2) Ym≥1 1 1 − zm q m−1 . (4.21) Proof. Consider the action of T1+ on Hilbn,n+2(0): Lemma 4.2.12 and Lemma 4.2.14 show that the attracting sets are affine cells of the affine cell decomposition of Hilbn,n+1,n+2(0). The dimensions of these cells were calculated in Proposition 3.2.13. Then Proposition 1.3.2 proves the first equality for the Poincaré polynomial. To prove the identity (4.21) for the generating function, thank to Lemma 4.2.3, it is enough to use the generating functions (4.15) and (4.19). Bibliography [B+83] Arnaud Beauville et al. Variétés kähleriennes dont la premiere classe de chern est nulle. J. Differential Geom, 18(4):755 -- 782, 1983. [BB73] Andrzej Bialynicki-Birula. Some theorems on actions of algebraic groups. Annals of mathematics, pages 480 -- 497, 1973. [BE13] Michaël Bulois and Laurent Evain. Nested punctual hilbert schemes and commuting varieties of parabolic subalgebras. arXiv:1306.4838, 2013. [Bri77] Joël Briançon. Description de hilb n c {x, y}. 41(1):45 -- 89, 1977. Inventiones mathematicae, [BS16] Dori Bejleri and David Stapleton. The tangent space of the punctual hilbert scheme. arXiv:1604.04915, 2016. [CE12] Pierre-Emmanuel Chaput and Laurent Evain. On the equivariant cohomol- ogy of hilbert schemes of points in the plane. arXiv:1205.5470, 2012. [Che98] Jan Cheah. Cellular decompositions for nested hilbert schemes of points. pacific journal of mathematics, 183(1):39 -- 90, 1998. [dCM00] Mark Andrea A de Cataldo and Luca Migliorini. The douady space of a com- plex surface. Advances in Mathematics, 151(2):283 -- 312, 2000. [ES87] [ES88] Geir Ellingsrud and Stein Arild Strømme. On the homology of the hilbert scheme of points in the plane. Inventiones mathematicae, 87(2):343 -- 352, 1987. Geir Ellingsrud and Stein Arild Strømme. On a cell decomposition of the hilbert scheme of points in the plane. Inventiones mathematicae, 91(2):365 -- 370, 1988. [Fog68] John Fogarty. Algebraic families on an algebraic surface. American Journal of Mathematics, 90(2):511 -- 521, 1968. [Ful13] William Fulton. Intersection theory, volume 2. Springer Science & Business Media, 2013. 115 116 BIBLIOGRAPHY [Göt90] Lothar Göttsche. The betti numbers of the hilbert scheme of points on a smooth projective surface. Mathematische Annalen, 286(1):193 -- 207, 1990. [Göt94] Lothar Göttsche. Hilbert schemes of zero-dimensional subschemes of smooth varieties. 1994. [Gro60] Alexander Grothendieck. Techniques de construction et théorèmes d'existence en géométrie algébrique iv: Les schémas de hilbert. Séminaire Bourbaki, 6:249 -- 276, 1960. [GS93] Lothar Göttsche and Wolfgang Soergel. Perverse sheaves and the cohomol- ogy of hilbert schemes of smooth algebraic surfaces. Mathematische An- nalen, 296(1):235 -- 245, 1993. [Hai98] Mark Haiman. t, q-catalan numbers and the hilbert scheme. Discrete Math- ematics, 193(1):201 -- 224, 1998. [Hai01] Mark Haiman. Hilbert schemes, polygraphs and the macdonald positivity conjecture. Journal of the American Mathematical Society, 14(4):941 -- 1006, 2001. [Iar72] Anthony Iarrobino. Punctual hilbert schemes. Bulletin of the American Mathematical Society, 78(5):819 -- 823, 1972. [Iar77] Anthony A Iarrobino. Punctual Hilbert schemes, volume 188. American Mathematical Soc., 1977. [IY03] Anthony Iarrobino and Joachim Yaméogo. The family g t of graded artinian quotients of k [x, y] of given hilbert function. Communications in Algebra, 31(8):3863 -- 3916, 2003. [Leh99] Manfred Lehn. Chern classes of tautological sheaves on hilbert schemes of points on surfaces. Inventiones mathematicae, 136(1):157 -- 207, 1999. [Mac62] IG Macdonald. The poincaré polynomial of a symmetric product. In Math- ematical Proceedings of the Cambridge Philosophical Society, volume 58, pages 563 -- 568. Cambridge Univ Press, 1962. [MFK94] David Mumford, John Fogarty, and Frances Clare Kirwan. Geometric invari- ant theory, volume 34. Springer Science & Business Media, 1994. [Nak97] Hiraku Nakajima. Heisenberg algebra and hilbert schemes of points on pro- jective surfaces. Annals of mathematics, 145(2):379 -- 388, 1997. [Nak99] Hiraku Nakajima. Lectures on Hilbert schemes of points on surfaces, vol- ume 18. American Mathematical Society Providence, RI, 1999. BIBLIOGRAPHY 117 [NY08] Hiraku Nakajima and Kota Yoshioka. Perverse coherent sheaves on blow- up. ii. wall-crossing and betti numbers formula. arXiv: 0806.0463, 2008. [VW94] Cumrun Vafa and Edward Witten. A strong coupling test of s-duality. Nu- clear Physics B, 431(1):3 -- 77, 1994.
1904.03341	1	1904	2019-04-06T02:29:11	One dimensional topological Galois theory	[ "math.AG" ]	In the preprint we present an outline of the one dimensional version of topological Galois theory. The theory studies topological obstruction to solvability of equations "in finite terms" (i.e. to their solvabilty by radicals, by elementary functions, by quadratures and so on). The preprint is based on the author's book on topological Galois theory. It contains definitions, statements of results and comments to them. Basically no proofs are presented. The preprint was written as a part of the comments to a new edition (in preparation) of the classical book "Integration in finite terms'' by J.F.~Ritt. }	math.AG	math	ONE DIMENSIONAL TOPOLOGICAL GALOIS THEORY Askold Khovanskii Department of Mathematics, University of Toronto, Toronto, Canada Abstract. In the preprint we present an outline of the one dimensional version of topological Galois theory. The theory studies topological obstruction to solvability of equations "in finite terms" (i.e. to their solvabilty by radicals, by elementary functions, by quadratures and so on). The preprint is based on the author's book on topological Galois theory. It contains definitions, statements of results and comments to them. Basically no proofs are presented. The preprint was written as a part of the comments to a new edition (in prepa- ration) of the classical book "Integration in finite terms" by J.F. Ritt. 1. Introduction. As was discovered by Camille Jordan the monodromy group of an algebraic function is isomorphic to the Galois group of the associated extension of the field of rational functions. Therefore the monodromy group is responsible for the representability of an algebraic function by radicals (see [1]). However, not only algebraic functions have the monodromy group. It is defined for any solution of a linear differential equation whose coefficients are rational func- tions and for many more functions, for which the Galois group does not make sense. It is thus natural to try using the monodromy group for these functions instead of the Galois group to prove that they do not belong to a certain Liouvillian class. This particular approach is implemented in topological Galois theory (see [2]), which has a one-dimensional version and a multidimensional version. In the one-dimensional version we consider functions from Liouvillian classes as multi-valued analytic functions of one complex variable. It turns out that there exist topological restrictions on the way the Riemann surface of a function from a certain Liouvillian class can be positioned over the complex plane. If a function does not satisfy these restrictions, then it cannot belong to the corresponding Liouvillian class. Besides a geometric appeal, this approach has the following advantage. Topolog- ical obstructions relate to branching. It turns out that if a function does not belong to a certain Liouvillian class by topological reasons then it automatically does not belong to a much wider class of functions. This wider class can be obtained if we add to the Liouvillian class all single valued functions having at most countable set of singularities and allow them to enter all formulas. Key words and phrases. solvability by radicals, by elementary functions, by quadratures, by generalized quadratures. This work was partially supported by the Canadian Grant No. 156833-17. 1 Typeset by AMS-TEX 2 ASKOLD KHOVANSKII The composition of functions is not an algebraic operation. In differential alge- bra, this operation is replaced with a differential equation describing it. However, for example, the Euler Γ-function does not satisfy any algebraic differential equa- tion. Hence it is pointless to look for an equation satisfied by, say, the function Γ(exp x) and one can not describe it algebraically (but the function y = exp(Γ(x)) satisfies the equation y ′ = Γ′y over a differential field containing Γ and it makes sense in the differential algebra). The only known results on non-representability of functions by quadratures and, say, the Euler Γ-functions are obtained by our method. On the other hand, our method cannot be used to prove that a particular single valued meromorphic function does not belong to a certain Liouvillian class. There are the following topological obstructions to representability of functions by generalized quadratures, k-quadratures and quadratures. Firstly, the functions representable by generalized quadratures and, in particular, the functions representable by k-quadratures and quadratures may have no more than countably many singular points in the complex plane (see section 6). Secondly, the monodromy group of a function representable by quadratures is necessarily solvable (see section 8). There are similar restrictions for for a func- tion representable by generalized quadratures and k-quadratures. However, these restrictions are more involved. To state them, the monodromy group should be regarded not as an abstract group but rather as a transitive subgroup in the per- mutation group. In other terms, these restrictions make use not only of the mon- odromy group but rather of the monodromy pair of the function consisting of the monodromy group and the stabilizer of some germ of the function (see section 9). One can prove that the only reasons for unsolvability in finite terms of Fuchsian linear differential equations are topological (see section 14). In other words, if there are no topological obstructions to solvability of a Fuchsian equation by generalized quadratures (by k-quadratures, by quadratures), then this equation is solvable by generalized quadratures (by k-quadratures or by quadratures respectively). The proof is based on a linear-algebraic part of differential Galois theory (dealing with linear algebraic groups and their differential invariants), 2. Solvability of equations in finite terms. An equation is solvable "in finite terms" (or is solvable "explicitly") if its solutions belong to a certain class of func- tions. Different classes of functions correspond to different notions of solvability in finite terms. A class of functions can be introduced by specifying a list of basic functions and a list of admissible operations. Given the two lists, the class of functions is defined as the set of all functions that can be obtained from the basic functions by repeated application of admissible operations. Below, we define Liouvillian classes of functions in exactly this way. Classes of functions, which appear in the problems of integrability in finite terms, contain multivalued functions. Thus the basic terminology should be made clear. We work with multivalued functions "globally", which leads to a more general un- derstanding of classes of functions defined by lists of basic functions and of admis- sible operations. A multivalued function is regarded as a single entity. Operations on multivalued functions can be defined. The result of such an operation is a set of multivalued functions; every element of this set is called a function obtained from the given functions by the given operation. A class of functions is defined as the ONE DIMENSIONAL TOPOLOGICAL GALOIS THEORY 3 set of all (multivalued) functions that can be obtained from the basic functions by repeated application of admissible operations. 3. Operations on multivalued functions. Let us define, for example, the sum of two multivalued functions on a connected Riemann surface U . Definition 7. Take an arbitrary point a in U , any germ fa of an analytic function f at the point a and any germ ga of an analytic function g at the same point a. We say that the multivalued function ϕ on U generated by the germ ϕa = fa + ga is representable as the sum of the functions f and g. For example, it is easy to see that exactly two functions of one variable are rep- resentable in the form √x + √x, namely, f1 = 2√x and f2 ≡ 0. Other operations on multivalued functions are defined in exactly the same way. For a class of mul- tivalued functions, being stable under addition means that, together with any pair of its functions, this class contains all functions representable as their sum. The same applies to all other operations on multivalued functions understood in the same sense as above. In the definition given above, not only the operation of addition plays a key role but also the operation of analytic continuation hidden in the notion of multivalued function. Indeed, consider the following example. Let f1 be an analytic function defined on an open subset V of the complex line C1 and admitting no analytic continuation outside of V , and let f2 be an analytic function on V given by the formula f2 = −f1. According to our definition, the zero function is representable in the form f1 + f2 on the entire complex line. By the commonly accepted viewpoint, the equality f1 + f2 = 0 holds inside the region V but not outside. Working with multivalued functions globally, we do not insist on the existence of a common region, were all necessary operations would be performed on single- valued branches of multivalued functions. A first operation can be performed in a first region, then a second operation can be performed in a second, different region on analytic continuations of functions obtained on the first step. In essence, this more general understanding of operations is equivalent to including analytic continuation to the list of admissible operations on the analytic germs. 4. Liouvillian classes of single variable functions. In this section, we define Liouvillian classes of single variable functions (for many variables, the corresponding definitions can be found in [ 2]). We will describe these classes by lists of basic functions and admissible operations. We will need the list of basic elementary functions. In essence, this list contains functions that are studied in high-school and which are frequently used in pocket calculators. List of basic elementary functions. 1. All complex constants and an independent variable x. 2. The exponential, the logarithm and the power xα, where α is any complex constant. 3. Trigonometric functions: sine, cosine, tangent, cotangent. 4. Inverse trigonometric functions: arcsine, arccosine, arctangent, arccotangent. 4 ASKOLD KHOVANSKII Lemma 1. Basic elementary functions can be expressed through the exponentials and the logarithms with the help of complex constants, arithmetic operations and compositions. Lemma 1 can be considered as a simple exercise. Its proof can be found in [3]. Let us now proceed with the list of classical operations on functions. List of classical operations. f g, and f /g. 1. The operation of composition takes functions f ,g to the function f ◦ g. 2. The arithmetic operations take functions f , g to the functions f + g, f − g, 3. The operation of differentiation takes function f to the function f ′. 4. The meromorphic operation takes functions f1, . . . , fn to the function F (f1, . . . , fn) where F is a fixed meromorphic function of n complex variables. 5. The operation of integration takes function f to a solution of equation y ′ = f (the function y is defined up to an additive constant). 6. The operation of solving algebraic equations takes functions f1, . . . , fn to the function y such that yn + f1yn−1 +··· + fn = 0 (the function y is not quite uniquely determined by functions f1, . . . , fn since an algebraic equation of degree n can have n solutions). 7. The operation of solving linear differential equations takes functions f1, . . . , fn to the function y such that y(n) + f1y(n−1) + ··· + fn = 0 (the function y is not uniquely determined by functions f1, . . . , fn since an differential equation of order n has an n dimensional space of solutions). We can now return to the definition of Liouvillian classes of single variable func- tions. Functions representable by radicals. List of basic functions: all complex con- stants, an independent variable x. List of admissible operations: arithmetic op- n , n = 2, 3, . . . , of a given erations and the operation of taking the n-th root f function f . 1 Functions representable by k-radicals. List of basic functions: all complex constants, an independent variable x. List of admissible operations: arithmetic operations and the operation of taking the n-th root f n , n = 2, 3, . . . , of a given function f , the operation of solving algebraic equations of degree ≤ k. Elementary functions. List of basic functions: basic elementary functions. List of admissible operations: compositions, arithmetic operations, differentiation. 1 Generalized elementary functions. This class of functions is defined in the same way as the class of elementary functions. We only need to add the operation of solving algebraic equations to the list of admissible operations. Functions representable by quadratures. List of basic functions: basic ele- mentary functions. List of admissible operations: compositions, arithmetic opera- tions, differentiation, integration. Functions representable by k-quadratures. This class of functions is defined in the same way as the class of functions representable by quadratures. We only need to add the operation of solving algebraic equations of degree at most k to the list of admissible operations. ONE DIMENSIONAL TOPOLOGICAL GALOIS THEORY 5 Functions representable by generalized quadratures. This class of functions is defined in the same way as the class of functions representable by quadratures. We only need to add the operation of solving algebraic equations to the list of admissible operations. 5. Simple formulas with complicated topology. Developing topological Ga- lois theory I followed the following plan: I. To find a wide class of multivalued functions such that: a) it is closed under all classical operations; b) it contains all entire functions and all functions from each Liouvillian class; c) for functions from the class the monodromy group is well defined. II. To use the monodromy group instead of the Galois group inside the class. Let us discuss some difficulties that one need to overcome on this way. Example. Consider an elementary function f defined by the following formula: f (z) = ln n Xj=1 λj ln(z − aj) where aj are different points in the complex line, and λj ∈ C are constants. Let Λ denote the additive subgroup of complex numbers generated by the con- It is clear that if n > 2, then for almost every collection of stants λ1, . . . , λn. constants λ1, . . . , λn, the group Λ is everywhere dense in the complex line. Lemma 2. If the group Λ is dense in the complex line, then the elementary function f has a dense set of logarithmic ramification points. Proof. Let g be the multivalued function defined by the formula g(z) = n Xj=1 λj ln(z − aj). Take a point a 6= aj, j = 1, . . . , n and let ga be one of the germs of g at a. A loop around the points a1, . . . , an adds the number 2πiλ to the germ ga, where λ is an element of the group Λ. Conversely, every germ ga + 2πiλ, where λ ∈ Λ, can be obtained from the germ ga by the analytic continuation along some loop. Let U be a small neighborhood of the point a, such that the germ ga has a singe- valued analytic continuation G on U . The image V of the domain U under the map G : U → C is open. Therefore, in the domain V , there is a point of the form 2πiλ, where λ ∈ Λ. The function G − 2πiλ is one of the branches of the function g over the domain U , and the zero set of this branch in the domain U is nonempty. Hence, one of the branches of the function f = ln g has a logarithmic ramification point in U . The set Σ of singular points of the function f is a countable set (see section 6). Under assumptions of Lemma 2 the set Σ is everywhere dense. It is not hard to verify that the monodromy group (see section 7) of the function f has the cardinality of the continuum. This is not surprising: the fundamental 6 ASKOLD KHOVANSKII group π1(C \ Σ) has obviously the cardinality of the continuum provided that Σ is a countable dense set. One can also prove that the image of the fundamental group π1(C \ {Σ ∪ b}) of the complement of the set Σ∪ b, where b 6∈ Σ, in the permutation group is a proper subgroup of the monodromy group of f . The fact that the removal of one extra point can change the monodromy group, makes all proofs more complicated. Thus even simplest elementary functions can have dense singular sets and mon- odromy groups of cardinality of the continuum. In addition the removal of an extra point can change their monodromy groups. 6. Class of S-functions. In this section, we define a broad class of functions of one complex variable needed in the construction of topological Galois theory. Definition. A multivalued analytic function of one complex variable is called a S-function, if the set of its singular points is at most countable. Let us make this definition more precise. Two regular germs fa and gb defined at points a and b of the Riemann sphere S2 are called equivalent if the germ gb is obtained from the germ fa by the analytic continuation along some path. Each germ gb equivalent to the germ fa is also called a regular germ of the multivalued analytic function f generated by the germ fa. A point b ∈ S2 is said to be a singular point for the germ fa if there exists a path γ : [0, 1] → S2, γ(0) = a, γ(1) = b such that the germ has no analytic continuation along this path, but for any τ , 0 ≤ τ < 1, it admits an analytic continuation along the truncated path γ : [0, τ ] → S2. It is easy to see that equivalent germs have the same set of singular points. A regular germ is called a S-germ, if the set of its singular points is at most countable. A multivalued analytic function is called a S-function if each its regular germ is a S-germ. Theorem 3 (on stability of the class of S-functions). The class S of all S-functions is stable under the following operations: 1) differentiation, i. e. if f ∈ S, then f ′ ∈ S; 2) integration, i. e. if f ∈ S and g ′ = f , then g ∈ S; 3) composition, i. e. if g, f ∈ S, then g ◦ f ∈ S; 4) meromorphic operations, i. e. if fi ∈ S, i = 1, . . . , n, the function F (x1, . . . , xn) 5) solving algebraic equations, i. e. if fi ∈ S, i = 1, . . . , n, and f n + f1f n−1 + 6) solving linear differential equations, i.e. if fi ∈ S, i = 1, . . . , n, and f (n) + is a meromorphic function of n variables, and f = F (f1, . . . , fn), then f ∈ S; ··· + fn = 0, then f ∈ S; f1f (n−1) + ··· + fnf = 0, then f ∈ S. Remark. Arithmetic operations and the exponentiation are examples of meromor- phic operations, hence the class of S-functions is stable under the arithmetic oper- ations and the exponentiation. Corollary 4 (see [2]). If a multivalued function f can be obtained from single valued S-functions by integration, differentiation, meromorphic operations, compo- sitions, solutions of algebraic equations and linear differential equations, then the function f has at most countable number of singular points. ONE DIMENSIONAL TOPOLOGICAL GALOIS THEORY 7 Corollary 5. A function having uncountably many singular points cannot be repre- sent by generalized quadratures. In particular it cannot be a generalized elementary function and it cannot be represented by k-quadratures or by quadratures. Example. Consider a discrete group Γ of fractional linear transformations of the open unit ball U having a compact fundamental domain. Let f be a nonconstant meromorphic function on U invariant under the action of Γ. Each point on the boundary ∂U belongs to the closure of the set of poles of f , thus the set Σ of singular points of f contains ∂U . So Σ has the cardinality of the continuum and f cannot be expressed by generalized quadratures. 7. Monodromy group of a S-function. The monodromy group of a S-function f is the group of all permutations of the sheets of the Riemann surface of f which are induced by motions around the singular set Σ of the function f . Below we discuss this definition more precisely. Let Fx0 be the set of all germs of the S-function f at point x0 /∈ Σ. Consider a closed curve γ in S2 \ Σ beginning and ending at the point x0. Given a germ y ∈ Fx0 we can continue it along the loop γ to obtain another germ yγ ∈ Yx0. Thus each such loop γ corresponds to a permutation Sγ : Fx0 → Fx0 of the set Fx0 that maps a germ y ∈ Fx0 to the germ yγ ∈ Fx0. It is easy to see that the map γ → Sγ defines a homomorphism from the fun- damental group π1(S2 \ Σ, x0) of the domain S2 \ Σ with the base point x0 to the group S(Fx0) of permutations. The monodromy group of the S-function f is the image of the fundamental group in the group S(Fx0) under this homomorphism. Remark. Instead of the point x0 one can choose any other point x1 ∈ S2 \ Σ. Such a choice will not change the monodromy group up to an isomorphism. To fix this isomorphism one can choose any curve γ : I → CN \ Σ where I is the segment 0 ≤ t ≤ 1 and γ(0) = x0, γ(1) = x1 and identify each germ fx0 of f with its continuation fx1 along γ. 8. Strong non representability by quadratures. One can prove the following theorem. Theorem 6 (see [2]). The class of all S-functions, having a solvable monodromy group, is stable under composition, meromorphic operations, integration and differ- entiation. Definition. A function f is strongly non representable by quadratures if it does not belong to a class of functions defined by the following data. List of basic functions: basic elementary functions and all single valued S-function. List of admissible operations: compositions, meromorphic operations, differentiation and integration. Theorem 6 implies the following corollary. Result on quadratures. If the monodromy group of an S-function f is not solv- able, then f is strongly non representable by quadratures. Example. The monodromy group of an algebraic function y(x) defined by an equation y5 + y − x = 0 is the unsolvable group S5. Thus y(x) provides an example of a function with finite set of singular points, which is strongly non representable by quadratures. The following corollary 7 contains a stronger result on non representability of algebraic functions by quadratures. 8 ASKOLD KHOVANSKII Corollary 7. If an algebraic function of one complex variable has unsolvable mon- odromy group then it is strongly non representable by quadratures. For algebraic functions of several complex variables there is a result similar to Corollary 7. 9. The monodromy pair. The monodromy group of a function f is not only an abstract group but is also a transitive group of permutations of germs of f at a non singular point x0. Definition. The monodromy pair of an S-function f is a pair of groups, consisting of the monodromy group of f at x0 and the stationary subgroup of a certain germ of f at x0. The monodromy pair is well defined, i.e. this pair of groups, up to isomorphisms, does not depend on the choice of the non singular point and on the choice of the germ of f at this point. Definition. A pair of groups [Γ, Γ0] is an almost normal pair if there is a normal subgroup H of Γ such that H ⊂ Γ0 and the coset Γ0/H is finite. Definition. The pair of groups [Γ, Γ0] is called an almost solvable pair of groups if there exists a sequence of subgroups Γ = Γ1 ⊇ ··· ⊇ Γm, Γm ⊂ Γ0, such that for every i, 1 ≤ i ≤ m − 1 group Γi+1 is a normal divisor of group Γi and the factor group Γi/Γi+1 is either a commutative group, or a finite group. Definition. The pair of groups [Γ, Γ0] is called a k-solvable pair of groups if there exists a sequence of subgroups Γ = Γ1 ⊇ ··· ⊇ Γm, Γm ⊂ Γ0, such that for every i, 1 ≤ i ≤ m − 1 group Γi+1 is a normal divisor of group Γi and the factor group Γi/Γi+1 is either a commutative group, or a subgroup of the group Sk of permutations of k elements. We say that group Γ is almost solvable or k-solvable if pair [Γ, e], where e is the group containing only the unit element, is almost solvable or k-solvable respectively. It is easy to see that an almost normal pair of groups [Γ, Γ0] is almost solvable or k-solvable if and only if the group Γ is almost solvable or k-solvable respectively. 10. Strong non representability by k-quadratures. One can prove the fol- lowing theorem. Theorem 8 (see [2]). The class of all S-functions, having a k-solvable monodromy pair, is stable under composition, meromorphic operations, integration, differentia- tion and solutions of algebraic equations of degree ≤ k. Definition. A function f is strongly non representable by k-quadratures if it does not belong to a class of functions defined by the following data. List of basic func- tions: basic elementary functions and all single valued S-function. List of admissible operations: compositions, meromorphic operations, differentiation,integration and solutions of algebraic equations of degree ≤ k. Theorem 8 implies the following corollary. ONE DIMENSIONAL TOPOLOGICAL GALOIS THEORY 9 Result on k-quadratures. If the monodromy pair of an S-function f is not k- solvable, then f is strongly non representable by k-quadratures. Example. The monodromy group of an algebraic function y(x) defined by an equation yn + y − x = 0 is the permutation group group Sn. For n ≥ 5 the group Sn is not an (n − 1)-solvable group. Thus y(x) provides an example of a function with finite set of singular points which is strongly non representable by (n − 1)-quadratures. This example can be generalized. Corollary 9 (see [2]). If an algebraic function of one complex variable has non k-solvable monodromy group then it is strongly non representable by k-quadratures. Theorem 10 (see [2]). An algebraic function of one variable whose monodromy group is k-solvable, can be represented by k-radicals. Results similar to Corollary 9 and Theorem 10 hold also for algebraic functions of several complex variables. 11. Strong non representability by generalized quadratures. One can prove the following theorem. Theorem 11 (see [2]). The class of all S-functions, having an almost solvable monodromy pair, is stable under composition, meromorphic operations, integration, differentiation and solutions of algebraic equations. Definition. A function f is strongly non representable by generalized quadratures if it does not belong to a class of functions defined by the following data. List of basic functions: basic elementary functions and all single valued S-function. List of admissible operations: compositions, meromorphic operations, differentiation, integration and solutions of algebraic equations. Theorem 11 implies the following corollary. Result on generalized quadratures. If the monodromy pair of an S-function f is not almost solvable, then f is strongly non representable by generalized quadra- tures. Suppose that the Riemann surface of a function f is a universal covering space If n ≥ 3 then the function f over the Riemann sphere with n punched points. is strongly non representable by generalized quadratures. Indeed, the monodromy pair of f consists of the free group with n− 1 generators, and its unit subgroup. It is easy to see that such a pair of groups is not almost solvable. Example. Consider the function z(x), which maps the upper half-plane onto a triangle with vanishing angles, bounded by three circular arcs. The Riemann surface of z(x) is a universal covering space over the sphere with three punched points.1 Thus z(x) is strongly non representable by generalized quadratures. 1it is easy to see that the function z(x) maps its Riemann surface to the open ball whose boundary contains the vertices of the triangle. These properties of the function z(x) play the crucial role in Picard's beautiful proof of his Little Picard Theorem. 10 ASKOLD KHOVANSKII Example. Let K1 and K2 be the following elliptic integrals, considered as the functions of the parameter x: K1(x) = Z 1 0 dt p(1 − t2)(1 − t2x2) and K2(x) = Z 0 1 x dx p(1 − t2)(1 − t2x2) . The functions z(x) can be obtained from K1(x) and from K2(x) by quadratures. Thus both functions K1(x) and K2(x) are strongly non representable by generalized quadratures. In the next section we will list all polygons G bounded by circular arcs for which the Riemann map of the upper half-plan onto G is representable by generalized quadratures. 12. Maps of the upper half-plane onto a curved polygon. Consider a polygons G on the complex plane bounded by circle arcs, and the function fG establishing the Riemann mapping of the upper half-plane onto the polygon G. The Riemann -- Schwarz reflection principle allows to describe the monodromy group LG of the function fG and to show that all singularities of fG are simple enough. This information together with Theorem 11 provide a complete classification of all polygons G for which the function fG is representable in explicit form (see [2]). If a polygon G is obtained from a polygon G by a ]linear transformation w : C → C then f G = w(fG). Thus it is enough to classify G up to a ]linear transformation. 1) The first case of integrability: the continuations of all sides of the polygon G intersect at one point. Mapping this point to infinity by a fractional linear transformation, we obtain a polygon G bounded by straight line segments. All transformations in the group L(G) have the form z → az + b. All germs of the function f = fG at a non-singular point c are obtained from a fixed germ fc by the action of the group L(G) consisting of the affine transformations z → az + b. The germ Rc = (f ′′/f )c is invariant under the action of the group L(G). Therefore, the germ Rc is a germ of a single valued function R. The singular points of R can only be poles (see ). Hence the function R is rational. The equation f ′′/f = R is integrable by quadratures. This integrability case is well known. The function f in this case is called the Christoffel -- Schwarz integral. 2) The second case of integrability: there is a pair of points such that, for every side of the polygon G, these points are either symmetric with respect to this side or belong to the continuation of the side. We can map these two points to zero and infinity by a fractional linear trans- formation. We obtain a polygon G bounded by circle arcs centered at point 0 and intervals of straight rays emanating from 0 (see Figure 2). All transformations in the group L(G) have the form z → az, z → b/z. All germs of the function f = fG at a non-singular point c are obtained from a fixed germ fc by the action of the group L(G) : fc → afc, fc → b/fc. c/fc)2 is invariant under the action of the group L(G). Therefore, The germ Rc = (f ′ the germ Rc is a germ of a single valued function R. The singular points of R can only be poles (see ). Hence the function R is rational. The equation R = (f ′/f )2 is integrable by quadratures ONE DIMENSIONAL TOPOLOGICAL GALOIS THEORY 11 3) The finite nets of circles. To describe the third case of integrability we need to define first the finite net of circles on the complex plane. The classification of finite groups, generated by reflections in the Euclidian space R3 is well known. Each such group is the symmetry group of the following bodies: 1. a regular n-gonal pyramid; 2. a regular n-gonal diheron, or the body formed by two equal regular n-gonal pyramids sharing the base; 3. a regular tetrahedron; 4. a regular cube or icosahedron; 5. a regular dodecahedron or icosahedron. All these groups of isometries, except for the group of dodecahedron or icosahe- dron, are solvable. The intersections of the unit sphere, whose center coincides with the barycenter of the body, with the mirrors, in which the body is symmetric, is a certain net of great circles. Stereographic projections of each of them is a net net of circles on complex plane defined up to a fractional linear transformation. The nets corresponding to the bodies listed above will be called the finite nets of circles. 4) The third case of integrability: every side side of a polygon G belongs to some finite net of circles. In this case the function fG has finitely many branches. Since all singularities of the function fG are algebraic (see [2]), the function fG is an algebraic function. For all finite nets but the net of dodecahedron or icosahedron, the algebraic function fG is representable by radicals. For the net of dodecahedron or icosahedron the function fG is representable by radicals and solutions of degree five algebraic equations (in other words fG is representable by k-radicals). 5) The strong non representability. Our results imply the following: Theorem 12 (see [2]). If a polygon G bounded by circles arcs does not belong to one of the three cases described above, then the function fG is strongly non representable by generalized quadratures. 13. Non solvability of linear differential equations. Consider a homogeneous linear differential equation (3) y(n) + r1y(n−1) + ··· + rny = 0, whose coefficients ri's are rational functions of the complex variable x. The set Σ ⊂ C of poles of ri's is called the set of singular points of the equation (3). At a point x0 ∈ C \ Σ the germs of solutions of (3) form a C-linear space Vx0 of dimension n. The monodromy group M of the equation (3) is the group of all linear transformations of the space Vx0 which are induced by motions around the set Σ. Below we discuss this definition more precisely. Consider a closed curve γ in C\ Σ beginning and ending at the point x0. Given a germ y ∈ Vx0 we can continue it along the loop γ to obtain another germ yγ ∈ Vx0 . Thus each such loop γ corresponds to a map Mγ : Vx0 → Vx0 of the space Vx0 to itself that maps a germ y ∈ Vx0 to the germ yγ ∈ Vx0 . The map Mγ is linear since an analytic continuation respects the arithmetic operations. It is easy to see that the map γ → Mγ defines a homomorphism of the fundamental group π1(C \ Σ, x0) of the domain C \ Σ with the base point x0 to the group GL(n) of invertible linear transformations of the space Vx0 . 12 ASKOLD KHOVANSKII The monodromy group M of the equation (3) is the image of the fundamental group in the group GL(n)) under this homomorphism. Remark. Instead of the point x0 one can choose any other point x1 ∈ C \ Σ. Such a choice will not change the monodromy group up to an isomorphism. To fix this isomorphism one can choose any curve γ : I → CN \ Σ where I is the segment 0 ≤ t ≤ 1 and γ(0) = x0, γ(1) = x1 and identify each germ yx0 of solution of (*) with its continuation yx1 along γ. Lemma 13. The stationary subgroup in the monodromy group M of the germ y ∈ Vx0 of almost every solution of the equation (3) is trivial (i.e. contains only the unite element e ∈ M ). Proof. The monodromy group M contains countable many linear transformations Mi. The space Li ⊂ Vx0 of fixed points of a non identity transformation Mi, is a proper subspace of Vx0 . The union L of all subspaces Li is a measure zero subset of Vx0. The stationary subgroup in M of y ∈ Vx0 \ L is trivial. Theorem 14 (see [2]). If the monodromy group of the equation (3) is not almost solvable (is not k-solvable, or is not solvable) then its almost every solution is strongly non representable by generalized quadratures (correspondingly, is strongly non representable by k-quadratures, or is strongly non representable by quadratures). Consider a homogeneous system of linear differential equations (4) y ′ = Ay where y = (y1, . . . , yn) is the unknown vector valued function and A = {ai,j(x)} is a n× n matrix, whose entries are rational functions of the complex variable x. One can define the monodromy group of the equation (4) exactly in the same way as it was defined for the equation (3). We will say that a vector valued function y = (y1, . . . , yn) belongs to a certain class of functions if all its components yi belong to this class. For example the statement "a vector valued function y = (y1, . . . , yn) is strongly non representable by generalized quadratures" means that at least one component yi of y is strongly non representable by generalized quadratures. Theorem 15. If the monodromy group of the system (4) is not almost solvable (is not k-solvable, or is not solvable) then its almost every solution is strongly non rep- resentable by generalized quadratures (correspondingly, is strongly non representable by k-quadratures, or is strongly non representable by quadratures). 14. Solvability of Fuchsian equations. The differential field of rational func- tions of x is isomorphic to the differential field R of germs of rational functions at the point x0 ∈ C \ Σ. Consider the differential field extension R{y1, . . . , yn} of R where the germs y1, . . . yn form a basis in the space Vx0 of solutions of the equation (3) at x0. Lemma 16. Every linear map Mγ from the monodromy group of equation (3), can be uniquely extended to a differential automorphism of the differential field R{y1, . . . , yn} over the field R. ONE DIMENSIONAL TOPOLOGICAL GALOIS THEORY 13 Proof. Every element f ∈ R{y1, . . . , yn} is a rational function of the independent variable x, the germs of solutions y1, . . . , yn and their derivatives. It can be con- tinued meromorphically along the curve γ ∈ π1(C \ Σ, x0) together with y1, . . . , yn. This continuation gives the required differential automorphism, since the continu- ation preserves the arithmetical operations and differentiation, and every rational function of x returns back to its original values (since it is a single-valued val- ued function). The differential automorphism is unique because the extension is generated by y1, . . . , yn. The differential Galois group (see [2], [3]) of the equation (3) over R is the group of all differential automorphisms of the differential field R{y1, . . . , yn} over the differential field R. According to Lemma 32 the monodromy group of the equation (3) can be considered as a subgroup of its differential Galois group over R. The differential field of invariants of the monodromy group action is a subfield of R{y1, . . . , yn}, consisting of the single-valued functions. Differently from the algebraic case, for differential equations the field of invariants under the action of the monodromy group can be bigger than the field of rational functions. The reason is that the solutions of differential equations may grow exponentially in approaching the singular points or infinity. Example. All solutions of the simplest differential equation y ′ = y are single- valued exponential functions y = C exp x, which are not rational. For a wide class of Fuchsian linear differential equations all the solutions, while approaching the singular points and the point infinity, grow polynomially. The following Frobenius theorem is an analog for Fuchsian equations of C.Jordan theorem (see [ ]) for algebraic equations. Theorem (Frobenius). For Fuchsian differential equations the subfield of the differential field R{y1, . . . , yn}, consisting of single-valued functions, coincides with the field of rational functions. A system of linear differential equations (4) is called a Fuchsian system if the matrix A has the following form: (5) A(x) = k Xi=1 Ai x − ai , where the Ai's are constant matrices. Linear Fuchsian system of differential equations in many ways are similar to linear Fuchsian differential equations In construction of explicit solutions of linear differential equations the following theorem is needed. Theorem (Lie -- Kolchin). Any connected solvable algebraic group acting by linear transformations on a finite-dimensional vector space over C is triangularizable in a suitable basis. Using Frobenius Therem and Lie -- Kolchin Theorem one can prove that the only reasons for unsolvability of Fuchsian linear differential equations and systems of linear differential equations are topological. In other words, if there are no topo- logical obstructions to solvability then such equations and systems of equations are solvable. Indeed, the following theorems hold: 14 ASKOLD KHOVANSKII Theorem 17 (see [2]). If the monodromy group of the linear Fuchsian differential equation (3) is almost solvable (is k-solvable, or is solvable) then its every solution is representable by generalized quadratures (correspondingly, is representable by k- quadratures, or is representable by quadratures). Theorem 18 (see [2]). If the monodromy group of the linear Fuchsian system differential equations (4) is almost solvable (is k-solvable, or is solvable) then its every solution is representable by generalized quadratures (correspondingly, is rep- resentable by k-quadratures, or is representable by quadratures). 15. Fuchsian systems with small coefficients. In general the monodromy group of a given Fuchsian equation is very hard to compute. It is known only for very special equations, including the famous hypergemetric equations. Thus Theorems 17 and 18 are not explicat. If the matrix A(x) in the system (4) is triangular then one can easily solve the system by quadratures. It turns out that if the matrix A(x) has the form (5), where the matrices Ai's are sufficiently small, then the system (4) with a non triangular matrix A(x) is unsolvable by generalized quadratures for a topological reason. Theorem 19 (see [2]). If the matrices Ai's are sufficiently small, kAik < ε(a1, . . . , ak, n), then the monodromy group of the system (6) k y ′ = ( Xi=1 Ai x − ai )y is almost solvable if and only if the matrices Ai's are triangularizable in a suitable basis. Corollary 20. If in the assumptions of Theorem 19 the matrices Ai's are not triangularizable in a suitable basis then almost every solution of the system (6) is strongly non representable by generalized quadratures. 16. Polynomials invertible by radicals. In 1922 J.F.Ritt published (see [5]) the following beautiful theorem which fits nicely into topological Galois theory. Theorem (J.F. Ritt). The inverse function of a polynomial with complex coeffi- cients can be represented by radicals if and only if the polynomial is a composition of linear polynomials, the power polynomials z → zn, Chebyshev polynomials and polynomials of degree at most 4. Outline of proof (following [6]). 1) Every polynomial is a composition of primitive ones: Every polynomial is a composition of polynomials that are not themselves compositions of polynomials of degree > 1. Such polynomials are called primitive. Recall that a permutation group G acting on a non-empty set X is called primitive if G acts transitively on X and G preserves no nontrivial partition of X. A polynomial is primitive if and only if the monodromy group of inverse of the polynomial acts primitively on its branches. 2) Reduction to the case of primitive polynomials: A composition of polynomials is invertible by radicals if and only if each polynomial in the composition is invertible by radicals. Indeed, if each of the polynomials in composition is invertible by radicals, then their composition also is. Conversely, if a polynomial R appears in ONE DIMENSIONAL TOPOLOGICAL GALOIS THEORY 15 the presentation of a polynomial P as a composition P = Q ◦ R ◦ S and P −1 is representable by radicals, then R−1 = Q◦ P −1 ◦ S is also representable by radicals. Thus it is enough to classify only the primitive polynomials invertible by radicals. 3) A result on solvable primitive permutation groups containing a full cycle: A primitive polynomial is invertible by radicals if and only if the monodromy group of inverse of the polynomial is solvable. Since it acts primitively on its branches and contains a full cycle (corresponding to a loop around the point at infinity on the Riemann sphere), the following group-theoretical result of Ritt is useful for the classification of polynomials invertible by radicals: Theorem (on primitive solvable groups with a cycle). Let G be a primitive solvable group of permutations of a finite set X which contains a full cycle. Then either X = 4, or X is a prime number p and X can be identified with the elements of the field Fp so that the action of G gets identified with the action of the subgroup of the affine group AGL1(p) = {x → ax + ba ∈ (Fp)∗, b ∈ Fp} that contains all the shifts x → x + b. 4) Solvable monodromy groups of inverse of primitive polynomials: It can be shown by applying Riemann -- Hurwitz formula that among the groups in Theorem on primitive solvable groups with a cycle, only the following groups can be realized as monodromy groups of inverse of primitive polynomials: 1. G ⊂ S4, 2. Cyclic group G = {x → x + b} ⊂ AGL1(p), 3. Dihedral group G = {x → ±x + b} ⊂ AGL1(p). 5) Description of primitive polynomials invertible by radicals: It can be easily shown (see for instance [see Ritt 22], [Khovanskii 07 Variations], [Burda Khovan- skii11 Branching]) that the following result holds: Theorem 21. If the monodromy group of inverse of a primitive polynomial is a subgroup of the group {x → ±x + b} ⊂ AGL1(p), then up to a linear change of variables the polynomial is either a power polynomial or a Chebyshev polynomial. Thus the polynomials whose inverse have monodromy groups 1-3 are respectively 1. Polynomials of degree four. 2. Power polynomials up to a linear change of variables. 3. Chebyshev polynomials up to a linear change of variables. In each of these cases the fact that the polynomial is invertible by radicals follows from solvability of the corresponding monodromy group or from explicit formulas for its inverse (see for instance [BurdaKhovanskii11Branching]). 17. Polynomials invertible by k-radicals. In this section we discuss the fol- lowing generalization of J.F.Ritt's Theorem. Theorem 22 (see [6]). A polynomial invertible by radicals and solutions of equa- tions of degree at most k is a composition of power polynomials, Chebyshev polyno- mials, polynomials of degree at most k and, if k ≤ 14, certain primitive polynomials whose inverse have exceptional monodromy groups. A description of these excep- tional polynomials can be given explicitly. The proofs rely on classification of monodromy groups of inverse of primitive polynomials obtained by Muller based on group-theoretical results of Feit and on previous work on primitive polynomials whose inverse have exceptional monodromy groups by many authors. Besides the references to these highly involved and techni- cal results an outline of the proof of Theorem 22 is not complicated and it resembles the outline of the proof of Ritt's Theorem. 16 ASKOLD KHOVANSKII Let us start with some background on representability by k-radicals. Definition. Let k be a natural number. A field extension L/K is k-radical if there exists a tower of extensions K = K0 ⊂ K1 ⊂ . . . ⊂ Kn such that L ⊂ Kn and for each i, Ki+1 is obtained from Ki by adjoining an element ai, which is either a solution of an algebraic equation of degree at most k over Ki, or satisfies am i = b for some natural number m and b ∈ Ki. Theorem 23 (see [2]). A Galois extension L/K of fields of characteristic zero is k-radical if and only if its Galois group is k-solvable. An algebraic function z = z(x) of one or several complex variables is said to be representable by k-radicals if the corresponding extension of the field of rational functions is a k-radical extension. Theorem 23 and C. Jordan's Theorem (see [ 1]) imply the following corollary. Corollary 24. An algebraic function is representable by k-radicals if and only if its monodromy group is k-solvable. (Note that Theorem 10 above coincides with a part of Corollary 23). Let us outline briefly the main steps in the proof of Theorem 22: Outline of proof of Theorem 22. : 1) Exactly as in Ritt's theorem one can show that a composition of polynomials is invertible by k-radicals if and only if each polynomial in the composition is invertible by k-radicals. Thus one can reduce Theorem 22 to the case of primitive polynomials. 2) Feit and Jones totally classified all primitive permutation groups of n elements containing a full cycle. 3) Using the this classification and Riemann-Hurwitz formula, Muller listed all groups of permutations of n elements which are monodromy groups of inverses of degree n primitive polynomials. 4) For each group from Muller's list of groups of permutations of n elements one can determine the smallest k for which it is k-solvable and choose the exceptional groups for which k is smaller than n. 5) For each such exceptional group one can explicitly describe polynomials whose inverse has the exceptional monodromy group. 18. Acknowledgement. I would like to thank Michael Singer who invited me to write comments for a new edition of the classical J.F. Ritt's book "Integration in finite terms" [4]. This preprint was written as a part of these comments. I also am grateful to Fedor Kogan who edited my English. REFERENCES [1] A.G. Khovanskii, On representability of algebraic functions by radicals, arXiv:1903.08632 [math.AG] [2] A. Khovanskii, Topological Galois theory. Solvability and unsolvability of equations in finite terms. Translated by Valentina Kiritchenko and Vladlen Tim- orin. Series: Springer Monographs in Mathematics. Springer Berlin Heidelberg. 2014, XVIII, 305 pp. 6 illus. ONE DIMENSIONAL TOPOLOGICAL GALOIS THEORY 17 [3] M. van der Put, M. Singer, Galois theory of linear differential equations (Springer, Berlin/New York, 2003). [4] J. Ritt, Integration in finite terms. Liouvilles theory of elementary methods, N. Y. Columbia Univ. Press. 1948. [5] J. Ritt, On algebraic functions which can be expressed in terms of radicals, Trans. Am. Math. Soc. 24, 21 -- 30 (1922). [6] Yu.Burda, A.Khovanskii, Polynomials invertible in k-radicals. Arnold Math- ematical Journal, V. 2, No 1, 2016, 121 -- 138.
1611.00553	2	1611	2017-06-03T19:21:32	Rational curves on smooth hypersurfaces of low degree	[ "math.AG", "math.NT" ]	We study the family of rational curves on arbitrary smooth hypersurfaces of low degree using tools from analytic number theory.	math.AG	math	RATIONAL CURVES ON SMOOTH HYPERSURFACES OF LOW DEGREE T.D. BROWNING AND P. VISHE Abstract. We study the family of rational curves on arbitrary smooth hypersurfaces of low degree using tools from analytic number theory. Contents Introduction 1. 2. Spreading out 3. The Hardy–Littlewood circle method 4. Geometry of numbers in function fields 5. Weyl differencing 6. The contribution from the minor arcs References 1 3 5 7 9 15 18 1. Introduction The geometry of a variety is intimately linked to the geometry of the space of rational curves on it. Given a projective variety X defined over C, a natural object to study is the moduli space of rational curves on X. There are many results in the literature establishing the irreducibility of such mapping spaces, but most statements are only proved for generic X. Following a strategy of Ellenberg and Venkatesh, we shall use tools from analytic number theory to prove such a result for all smooth hypersurfaces of sufficiently low degree. Let X ⊂ Pn be a smooth Fano hypersurface of degree d defined over C, with n > 3. For each positive integer e, the Kontsevich moduli space M 0,0(X, e) is a compactification of the space M0,0(X, e) of morphisms of degree e from P1 to X, up to isomorphism. According to Koll´ar [13, Thm. II.1.2/3], any irreducible component of M 0,0(X, e) has dimension at least µ = (n + 1 − d)e + n − 4. (1.1) Date: June 6, 2017. 2010 Mathematics Subject Classification. 14H10 (11P55, 14G05). 2 T.D. BROWNING AND P. VISHE Work of Harris, Roth and Starr [9] shows that M 0,0(X, e) is an irreducible, local complete intersection scheme of dimension µ, provided that X is general and d < 1 3 (n+1) by Beheshti and Kumar [6] (assuming that n > 23), and then to d 6 n − 2 by Riedl and Yang [15]. 2(n+1). The restriction on d has since been weakened to d < 2 In the setting d = 3 of cubic hypersurfaces it is possible to obtain results for all smooth hypersurfaces in the family. Thus Coskun and Starr [7] have shown that M 0,0(X, e) is irreducible and of dimension µ for any smooth cubic hypersurface X ⊂ Pn over C, provided that n > 4. (If n = 4 then M 0,0(X, e) has two irreducible components of the expected dimension µ = 2e.) At the expense of a much stronger condition on the degree, our main result establishes the irreducibility and dimension of the space M0,0(X, e), for an arbitrary smooth hypersurface X ⊂ Pn over C. Let n0(d) = 2d−1(5d − 4). (1.2) We shall prove the following statement. Theorem 1.1. Let X ⊂ Pn be a smooth hypersurface of degree d > 3 defined over C, with n > n0(d). Then for each e > 1 the space M0,0(X, e) is irreducible and of the expected dimension. The example of Fermat hypersurfaces, discussed in [7, §1], shows that the analogous result for M 0,0(X, e) is false when d > 3 and e is large enough. When e = 1 we have M 0,0(X, 1) = M0,0(X, 1) = F1(X), where F1(X) is the Fano scheme of lines on X. It has been conjectured, independently by Debarre and de Jong, that dim F1(X) = 2n − d − 3 for any smooth Fano hypersurface X ⊂ Pn of degree d. Beheshti [5] has confirmed this for d 6 8. Taking e = 1 in Theorem 1.1, we conclude that dim F1(X) = 2n − d − 3 for any d > 3, provided that n > n0(d). Our proof of Theorem 1.1 ultimately relies on techniques from analytic num- ber theory. The first step is "spreading out", in the sense of Grothendieck [8, §10.4.11] (cf. Serre [16]), which will take us to the analogous problem for smooth hypersurfaces defined over the algebraic closure of a finite field. Pass- ing to a finite field Fq of sufficiently large cardinality, for a smooth degree d hy- persurface X ⊂ Pn Fq defined over Fq, the cardinality of Fq-points on M0,0(X, e) can be related to the number of Fq(t)-points on X of degree e. We shall access the latter quantity through a function field version of the Hardy–Littlewood circle method. A comparison with the Lang–Weil estimate [10] then allows us to make deductions about the irreducibility and dimension of M0,0(X, e). The idea of using the circle method to study the moduli space of rational curves on varieties is due to Ellenberg and Venkatesh. The traditional setting for the circle method is a fixed finite field Fq, with the goal being to understand RATIONAL CURVES ON SMOOTH HYPERSURFACES 3 the Fq(t)-points on X of degree e, as e → ∞. This is the point of view taken in work of Lee [11, 12] on a Fq(t)-version of Birch's work on systems of forms in many variables. In contrast to this, we will be required to handle any fixed e > 1, as q → ∞. Pugin developed an "algebraic circle method" in his 2011 Ph.D. thesis [14] to study the spaces M0,0(X, e), when X ⊂ Pn Fq is the diagonal cubic hypersurface a0x3 0 + · · · + anx3 n = 0, (for a0, . . . , an ∈ F∗ q). Assuming that n > 12 and char(Fq) > 3, he succeeds in showing that the space M0,0(X, e) is irreducible and of the expected dimension. Our work, on the other hand, applies to arbitrary smooth hypersurfaces of sufficiently low degree, which are defined over the complex numbers. Finally, our investigation bears comparison with work of Bourqui [1, 2], who has also investigated the moduli space of curves on varieties using counting arguments. In place of the circle method, however, Bourqui draws on the theory of universal torsors. Acknowledgements. The authors are grateful to Hamid Ahmadinezhad, Lior Bary Soroker, Roya Beheshti, Emmanuel Peyre and the anonymous ref- eree for useful conversations. While working on this paper the first author was supported by ERC grant 306457. Let X ⊂ Pn be a smooth hypersurface of degree d, defined by a homogeneous 2. Spreading out polynomial F (x0, . . . , xn) = Xi∈Zn+1 >0 i0+···+in=d cixi0 0 . . . xin n , with coefficients ci ∈ C. Rather than working with M0,0(X, e), it will suffice to study the naive space More(P1, X) of actual maps P1 → X of degree e. The expected dimension of More(P1, X) is µ = µ + 3, where µ is given by (1.1), since P1 has automorphism group of dimension 3. We proceed to recall the construction of More(P1, X). Let Ge be the set of all homogeneous polynomials in u, v of degree e > 1, with coefficients in C. A rational curve of degree e on X is a non-constant morphism f : P1 C → X of degree e. It is given by f = (f0(u, v), . . . , fn(u, v)), with f0, . . . , fn ∈ Ge, with no non-constant common factor in C[u, v], such that F (f0(u, v), . . . , fn(u, v)) vanishes identically. We may regard f as a point in P(n+1)(e+1)−1 and the morphisms of degree e on X are parameterised by C 4 T.D. BROWNING AND P. VISHE C, X), which is an open subvariety of P(n+1)(e+1)−1 More(P1 cut out by a system of de + 1 equations of degree d. In this way we obtain the expected dimension C (n + 1)(e + 1) − 1 − (de + 1) = (n + 1 − d)e + n − 1 = µ, of More(P1 components of More(P1 Theorem 1.1 it will therefore suffice to show that More(P1 with dim More(P1 C, X). It follows from Koll´ar [13, Thm. II.1.2] that all irreducible C, X) have dimension at least µ. In order to establish C, X) is irreducible, C, X) 6 µ, provided that n > n0(d). The complement to More(P1 C, X) in its closure is the set of (f0, . . . , fn) with a C, X) by noting that common zero. We can obtain explicit equations for More(P1 f0, . . . , fn have a common zero if and only if the resultant Res(Pi λifi,Pj µjfj) is identically zero as a polynomial in λi, µj. It is clear that both X and More(P1 C, X) are defined by equations with coefficients belonging to the finitely generated Z-algebra Λ = Z[ci], obtained by adjoining the coefficients of F to Z. C, X) as schemes over Λ, with structure morphisms X → Spec Λ and In this way we may view X and More(P1 More(P1 C, X) → Spec Λ. By Chevalley's upper semicontinuity theorem (see [8, Thm. 13.1.3]), there exists a non-empty open set U of Spec Λ such that C, X) 6 dim More(P1 dim More(P1 C, X)m for any closed point m ∈ U. Here More(P1 C, X)m denotes the fibre above m, which is obtained via the base change Spec Λ/m → Spec Λ. Likewise, since integrality is an open condition, the space More(P1 C, X) will be irreducible if More(P1 C, X)m is. Choose a maximal ideal m in U. The quotient Λ/m is a finite field by arith- metic weak Nullstellensatz. By enlarging Λ, we may assume that it contains 1/d!. In particular, it follows that char(Λ/m) = p, say, with p > d, since any prime less than or equal to d is invertible in Λ. The quasi-projective varieties Xm and More(P1 C, X)m are defined over Fp, being given explicitly by reducing modulo m the coefficients of the original system of defining equations. By further enlarging Λ, if necessary, we may assume that Xm is smooth. There exists a finite field Fq0 such that Xm and More(P1 C, XC)m are both defined over Fq0. In view of the Lang–Weil estimate, Theorem 1.1 is a direct consequence of the following result, together with the fact that More(P1 C, XC)m is non empty in the cases under consideration. Theorem 2.1. Let n > n0(d) and let X ⊂ Pn Fq be a smooth hypersurface of degree d > 3 defined over a finite field Fq, with char(Fq) > d. Then for each e > 1 we have lim ℓ→∞ q−ℓµ# More(P1 Fq, X)(Fqℓ) 6 1. RATIONAL CURVES ON SMOOTH HYPERSURFACES 5 3. The Hardy–Littlewood circle method We now initiate the proof of Theorem 2.1. We henceforth redefine qℓ to be q and we replace n by n − 1 in the statement of the theorem. In particular the expected dimension is now µ = (n − d)e + n − 2. Our proof of Theorem 2.1 is based on a version of the Hardy–Littlewood circle method for the function field K = Fq(t), always under the assumption that char(Fq) > d. The main input for this comes from work of Lee [11, 12], combined with our own recent contribution to the subject, in the setting of cubic forms [4]. We begin by laying down some basic notation and terminology. To begin with, for any real number R we set bR = qR. Let O = Fq[t] be the ring of integers of K and let Ω be the set of places of K. These correspond to either monic irreducible polynomials in O, which we call the finite primes, or the prime at infinity t−1 which we usually denote by ∞. The associated absolute value · v is either · for some prime ∈ O or · , according to whether v is a finite or infinite place, respectively. These are given by a/b =(cid:18) 1 qdeg (cid:19)ord(a/b) and a/b = qdeg a−deg b, for any a/b ∈ K ∗. We extend these definitions to K by taking 0 = 0 = 0. For v ∈ Ω we let Kv denote the completion of K at v with respect to · v. We may identify K∞ with the set Fq((1/t)) =(Xi6N aiti : for ai ∈ Fq and some N ∈ Z) . We can extend the absolute value at the infinite place to K∞ to get a non- archimedean absolute value · : K∞ → R>0 given by α = qord α, where ord α is the largest i ∈ Z such that ai 6= 0 in the representation α =Pi6N aiti. In this context we adopt the convention ord 0 = −∞ and 0 = 0. We extend this to vectors by setting x = max16i6n xi, for any x ∈ K n ∞. Next, we put T = {α ∈ K∞ : α < 1} =(Xi6−1 aiti : for ai ∈ Fq) . Since T is a locally compact additive subgroup of K∞ it possesses a unique Haar measure dα, which is normalised so thatRT dα = 1. We can extend dα to a (unique) translation-invariant measure on K∞, in such a way that Z{α∈K∞:α< bN} dα = bN , 6 T.D. BROWNING AND P. VISHE for any N ∈ Z>0. These measures also extend to Tn and K n ∞, for any n ∈ Z>0. There is a non-trivial additive character eq : Fq → C∗ defined for each a ∈ Fq by taking eq(a) = exp(2πi TrFq/Fp(a)/p). This character yields a non-trivial (unitary) additive character ψ : K∞ → C∗ by defining ψ(α) = eq(a−1) for any α =Pi6N aiti in K∞. We may express this polynomial as Let F ∈ Fq[x] be a non-singular form of degree d > 3, with x = (x1, . . . , xn). F (x) = nXi1,...,id=1 ci1,...,idxi1 . . . xid, with coefficients ci1,...,id ∈ Fq. In particular F and the discriminant ∆F are non-zero, or equivalently, maxi ci = 1 and ∆F = 1. We will make frequent use of these facts in what follows. Associated to F are the multilinear forms Ψi(x(1), . . . , x(d−1)) = nXi1,...,id−1=1 ci1,...,id−1,ix(1) i1 . . . x(d−1) id−1 , (3.1) for 1 6 i 6 n. To establish Theorem 2.1 we work with the naive space Me = {x = (x1, . . . , xn) ∈ Ge(Fq)n \ {0} : F (x) = 0} , where Ge(Fq) is the set of binary forms of degree e with coefficients in Fq. Thus Me corresponds to the Fq-points on the affine cone of More(P1 Fq, X), where we drop the condition that x1, . . . , xn share no common factor. Let us set It will clearly suffice to show that bµ = µ + 1 = (n − d)e + n − 1 = (e + 1)n − de − 1. lim q→∞ q−bµ#Me 6 1, (3.2) (3.3) for n > n0(d), where n0(d) is given by (1.2). We proceed by relating #Me to the counting function that lies at the heart of our earlier investigation [4]. Let w : K n ∞ → {0, 1} be given by w(x) =Q16i6n w∞(xi), where w∞(x) =(1, if x < 1, 0, otherwise. Putting P = te+1, we then have #Me 6 N(P ), where N(P ) = Xx∈O n F (x)=0 w(x/P ). RATIONAL CURVES ON SMOOTH HYPERSURFACES 7 It follows from [4, Eq. (4.1)] that for any Q > 1 we have N(P ) = Xr∈O r6 bQ r monic a<rZθ<r−1 bQ−1 X∗ S(cid:16)a r + θ(cid:17) dθ, (3.4) whereP∗ means that the sum is taken over residue classes a < r for which gcd(a, r) = 1, and where S(α) = Xx∈O n ψ(αF (x))w(x/P ), (3.5) for any α ∈ T. We will work with the choice Q = d(e+1)/2, so that bQ = P d/2. The major arcs for our problem are given by r = 1 and θ < P −dqd−1. We let the minor arcs be everything else: i.e. those α = a/r + θ appearing in (3.4) for which either r > q, or else r = 1 and θ > P −dqd−1. The contribution Nmajor(P ) from the major arcs is easy to deal with. Indeed, for θ < P −dqd−1 and x < P we have θF (x) < P −dqd−1qde = q−1, whence ψ(θF (x)) = 1. Thus S(α) = P n, for α = θ belonging to the major arcs, whence Nmajor(P ) = P nZθ<P −dqd−1 dθ = P n−dqd−1 = q bµ. In order to prove (3.3), it therefore remains to show that lim q→∞ q−bµNminor(P ) = 0, (3.6) for n > n0(d), where Nminor(P ) is the overall contribution to (3.4) from the minor arcs. This will complete the proof of Theorem 2.1. 4. Geometry of numbers in function fields The purpose of this section is to record a technical result about lattice point counting over K∞. A lattice in K N ∞ is a set of points of the form x = Λu, where Λ is a N × N matrix over K∞ and u runs over elements of O N . By an abuse of notation we will also denote the set of such points by Λ. Given a lattice M, the adjoint lattice Λ is defined to satisfy ΛT M = IN , where IN is the N × N identity matrix. Let γ = (γij) be a symmetric n × n matrix with entries in K∞. Given any positive integer m, we define the special lattice Mm = t−mIn tmγ 0 tmIn! , 8 T.D. BROWNING AND P. VISHE with corresponding adjoint lattice t−mIn! . Λm = tmIn −tmγ 0 Let bR1, ...,bR2n denote the successive minima of the lattice corresponding to Mm. For any vector x ∈ K 2n ∞ let x1 = (x1, . . . , xn) and x2 = (xn+1, . . . , x2n). We claim that Mm and Λm can be identified with one another. Now Mm is the set of points x = Mmu where u = (u1, u2) runs over elements of O 2n. Likewise, Λm is the set of points y = Λmv where v = (v1, v2) runs over elements of O 2n. We can therefore identify Mm with Λm through the process of changing the sign of v2, then the sign of y2, then switching v1 with v2, and finally interchanging y1 and y2. It now follows from [11, Lemma 3.3.6] (cf. [12, Lemma B.6]) that Rν + R2n−ν+1 = 0, (4.1) for 1 6 ν 6 n. An important step in the proof of [11, Lemma 3.3.6] (cf. [12, Lemma B.6]) is a non-archimedean version of Gram–Schmidt orthogonalisa- tion, which is used without reference in the proof of [11, Lemma 3.3.3] (cf. [12, Lemma B.3]). This deficit is remedied by appealing to recent work of Usher and Zhang [17, Theorem 2.16]. For any Z ∈ R and any lattice Γ we define the counting function Note that Γ(Z) = Γ(⌈Z⌉) for any Z ∈ R. We proceed to establish the following inequality. Lemma 4.1. Let m, Z1, Z2 ∈ Z such that Z1 6 Z2 6 0. Then we have Γ(Z) = #{x ∈ Γ : x < bZ}. > bZ1 bZ2!n Mm(Z1) Mm(Z2) . Proof. Let 1 6 µ, ν 6 2n be such that Rµ < Z1 6 Rµ+1 and Rν < Z2 6 Rν+1. Since Rj is a non-decreasing sequence which satisfies Rj + R2n−j+1 = 0, by (4.1), we must have 0 6 Rn+1, whence in fact µ 6 ν 6 n. It follows from [11, Lemma 3.3.5] (cf. [12, Lemma B.5]) that Mm(Z1) Mm(Z2) 1 = j=1 bRj/cZ1(cid:17) (cZ1/cZ2)ν (cid:16)Qν (cid:16)Qν j=µ+1 bRj/cZ1(cid:17) (cZ1/cZ2)ν if Z1, Z2 < R1, if Z1 < R1 6 Z2, if R1 6 Z1 6 Z2, (cid:3) The statement of the lemma is now obvious. RATIONAL CURVES ON SMOOTH HYPERSURFACES 9 As above, let γ = (γij) be a symmetric n × n matrix with entries in K∞. For 1 6 i 6 n we introduce the linear forms Li(u1, . . . , un) = γijuj. nXj=1 Next, for given real numbers a, Z, we let N(a, Z) denote the number of vectors (u1, . . . , u2n) ∈ O 2n such that and uj <babZ Lj(u1, . . . , un) + uj+n < bZ ba If we put m = ⌊a⌋, then it is clear that for 1 6 j 6 n. Mm(Z − {a}) 6 N(a, Z) 6 Mm(Z + {a}), where {a} denotes the fractional part of a. The following result is a direct consequence of Lemma 4.1. Lemma 4.2. Let a, Z1, Z2 ∈ R with Z1 6 Z2 6 0. Then we have N(a, Z1) N(a, Z2) where K = ⌈Z1 − {a}⌉ − ⌈Z2 + {a}⌉. > bK n, 5. Weyl differencing In everything that follows we shall assume that char(Fq) > d and we will allow all our implied constants to depend at most on d and n. This section is concerned with a careful analysis of the exponential sum (3.5), using the function field version of Weyl differencing that was worked out by Lee [11, 12]. Our task is to make the dependence on q completely explicit and it turns out that gaining satisfactory control requires considerable care. Since we are concerned with hypersurfaces one needs to take R = 1 in [11, 12]. For any β = Pi6N biti ∈ K∞, we let kβk = Pi6−1 biti. Recalling the definition (3.1) of the multilinear forms associated to F , we let N(α) = #(u ∈ O (d−1)n : u1, . . . , ud−1 < P kαΨi(u)k < P −1 (∀i 6 n) ) , (5.1) where u = (u1, . . . , ud−1). We begin with an application of [11, Cor. 4.3.2] (cf. [12, Cor. 3.3]), which leads to the inequality S(α)2d−1 6 P (2d−1−d+1)nN(α), (5.2) for any α ∈ T. The next stage in the analysis of S(α) is a multiple application of the function field analogue of Davenport's "shrinking lemma", as proved in [11, 10 T.D. BROWNING AND P. VISHE Lemma 4.3.3] (cf. [12, Lemma 3.4]), ultimately leading to [11, Lemma 4.3.4] (cf. [12, Lemma 3.5]). Unfortunately the implied constant in these estimates is allowed to depend on q and so we must work harder to control it. Let Nη(α) = #(u ∈ O (d−1)n : u1, . . . , ud−1 < P η kαΨi(u)k < P −d+(d−1)η (∀i 6 n) ) , for any parameter η ∈ [0, 1]. Recalling that P = te+1, we shall prove the following uniform version of [11, Lemma 4.3.4] (cf. [12, Lemma 3.5]). Lemma 5.1. Let α ∈ T and suppose that η ∈ [0, 1) is chosen so that Then we have N(α) 6 P (n−ηn)(d−1)Nη(α). In particular, we have (e + 1)(η + 1) 2 ∈ Z. (5.3) S(α)2d−1 6 P 2d−1n P η(d−1)n Nη(α). Proof. In view of (5.1) and (5.2), the final part follows from the first part. For each v ∈ {0, . . . , d − 1}, define N (v)(α) to be the number of u ∈ O (d−1)n such that u1, . . . , uv < P η, (5.4) and kαΨi(u)k < P −v−1+vη, for 1 6 i 6 n. Thus we have N (0)(α) = N(α) and N (d−1)(α) = Nη(α). It will suffice to show that uv+1, . . . , ud−1 < P N (v)(α) > P −n+ηnN (v−1)(α), (5.5) for each v ∈ {1, . . . , d − 1}. Fix a choice of v, together with u1, . . . , uv−1, uv+1, . . . , ud−1 ∈ O n such that (5.4) holds. For each 1 6 i 6 n we consider the linear form Li(u) = αΨi(u1, . . . , uv−1, u, uv+1, . . . , ud−1) = γijuj, nXj=1 say, for a suitable symmetric n×n matrix γ = (γij), with entries in K∞. Given real numbers a and Z, define N(a, Z) to be the number of vectors (u1, . . . , u2n) in O 2n satisfying uj < \Z + a We are interested in estimating the number of u ∈ O n such that u < P η and kLi(u)k < P −v−1+vη, for 1 6 i 6 n, in terms of the number of u ∈ O n such that u < P and kLi(u)k < P −v+(v−1)η, for 1 6 i 6 n. That is, we wish to compare N(a, Z1) with N(a, Z2), where Lj(u1, . . . , un) − uj+n < \Z − a, for 1 6 j 6 n. and ba = P (v+1−(v−1)η)/2, cZ1 = P (v+1)(η−1)/2, cZ2 = P (v−1)(η−1)/2. RATIONAL CURVES ON SMOOTH HYPERSURFACES 11 that Note thatbabZ1 = P η andbabZ2 = P . Moreover, our hypothesis (5.3) implies (v − 1)(e + 1)(η + 1) (e + 1)(v + 1) (v − 1)(e + 1)η 2 ∈ Z. a = − 2 2 = v(e + 1) − Similarly, (5.3) implies that Z1, Z2 ∈ Z. It now follows from Lemma 4.2 that N(a, Z1) N(a, Z2) >(cid:16) \Z1 − Z2(cid:17)n which thereby completes the proof of (5.5). = P −n+ηn, (cid:3) Lemma 5.1 doesn't allow us to handle the case e = 1 of lines. To circumvent this difficulty we shall invoke a simpler version of the shrinking lemma, as follows. Lemma 5.2. Let α ∈ T and let v ∈ {1, . . . , d}. Then we have S(α)2d−1 6 P (2d−1−d+1)nqe(v−1)nM (v)(α), where M (v)(α) is the number of u ∈ O (d−1)n such that u1, . . . , uv−1 < q, and kαΨi(u)k < P −1 for 1 6 i 6 n. uv, . . . , ud−1 < P Proof. Noting that N(α) = M (1)(α), it follows from (5.2) that it will be enough to prove that M (v−1)(α) 6 qenM (v)(α) for 2 6 v 6 d. The proof follows that of Lemma 5.1 and so we shall be brief. Let u1, . . . , uv−1, uv+1, . . . , ud−1 ∈ O n be vectors satisfying u1, . . . , uv−1 < q, uv+1, . . . , ud−1 < P . Let γ and N(a, Z) be as in the proof of Lemma 5.1, corresponding to this choice. Lemma 4.2 clearly implies that N(e + 1, −e) N(e + 1, 0) > q−en. However, N(e + 1, −e) denotes the number of u ∈ O n such that u < q and kLi(u)k < q−2e−1, for 1 6 i 6 n. The lemma follows on noting that q−2e−1 < q−e−1 = P −1. (cid:3) The next step is an application of the function field analogue of Heath- Brown's Diophantine approximation lemma, as worked out in [11, Lemma 4.3.5] (cf. [12, Lemma 3.6]). Let α = a/r + θ, where a/r ∈ K and θ ∈ T. Note that the maximum absolute value of the coefficients of each multilinear form Ψj is 1. We shall apply [11, Lemma 4.3.5] (cf. [12, Lemma 3.6]) with 12 T.D. BROWNING AND P. VISHE cM = P (d−1)η and bY = P d−(d−1)η. We want a maximal choice of η > 0 such that 1 P d P (d−1)η < min(cid:26)P d−1, r (cid:27) P (d−1)η 6 r max(cid:8)1, P dθ(cid:9) . rθ , This leads to the constraint (e + 1)η 6 Γ, where and Γ = 1 d − 1 ord(cid:18)min(cid:26) P d−1 q , 1 qrθ , P d qr , r max(cid:8)1, P dθ(cid:9)(cid:27)(cid:19) , in which we abuse notation and denote by ord the integer exponent of q that appears. For i ∈ {0, 1}, let [Γ]i denote the largest non-negative integer not exceeding Γ, which is congruent to i modulo 2. We then choose η via (5.6) (e + 1)η =([Γ]0 [Γ]1 if 2 ∤ e, if 2 e. (5.7) One notes that (e + 1)η 6 Γ and (5.3) is satisfied. It now follows from [11, Lemma 4.3.5] (cf. [12, Lemma 3.6]) that Nη(α) 6 Uη, where Uη denotes the number of u ∈ O (d−1)n such that u1, . . . , ud−1 < P η and Ψi(u) = 0, for 1 6 i 6 n. A standard calculation, which we recall here for completeness, now shows that the latter system of equations defines an affine variety V ⊂ A(d−1)n of dimension at most (d − 2)n. To see this, we note that the intersection of V with the diagonal ∆ = {u ∈ A(d−1)n : u1 = · · · = ud−1} is contained in the singular locus of F and so has affine dimension 0. The claim follows on noting that 0 = dim(V ∩ ∆) > dim V + dim ∆ − (d − 1)n = dim V − (d − 2)n . We now apply [4, Lemma 2.8]. Since P η = q(e+1)η, with (e + 1)η ∈ Z, this directly yields the existence of a positive constant cd,n, independent of q, such that Uη 6 cd,nP η(d−2)n. Inserting this into Lemma 5.1, we therefore arrive at the following conclusion. Lemma 5.3. Let L = 2−d+1n, let a/r ∈ K and let θ ∈ T. Let η be given by (5.7). Then there exists a constant cd,n > 0, independent of q, such that S(a/r + θ) 6 cd,nP n−Lη. It turns out that this estimate is inefficient when r is small. Let κ =(1 0 if 2 ∤ e, if 2 e. (5.8) RATIONAL CURVES ON SMOOTH HYPERSURFACES 13 It will also be advantageous to consider the effect of taking (e + 1)η = 1 + κ, instead of (5.7). Since (e + 1)(η + 1) 2 it follows from Lemma 5.1 that = 1 + e + κ 2 ∈ Z, S(α) 6 P nN 2−d+1 q(1+κ)(d−1)L , where N = #(u ∈ O (d−1)n : u1, . . . , ud−1 6 qκ kαΨi(u)k < qκ(d−1)−de−1 (∀i 6 n) ) , (5.9) (5.10) Supposing that α = a/r + θ for a/r ∈ K and θ ∈ T, our argument now bifurcates according to the degree of r. Lemma 5.4 (deg(r) > 1). Let L = 2−d+1n, let a/r ∈ K and let θ ∈ T. Assume that (i) e > 1, q 6 r < qde+1−κ(d−1) and rθ < q−κ(d−1); or (ii) e = 1, q2 6 r 6 qd and rθ 6 q−d. Then there exists a constant c′ d,n > 0, independent of q, such that S(a/r + θ) 6 c′ d,nP nq−L. Proof. To deal with case (i) we apply [11, Lemma 4.3.5] (cf. [12, Lemma 3.6]) with Y = de + 1 − κ(d − 1) and M = κ(d − 1) + 1 2. Our hypotheses ensure that all i 6 n. In particular we have N = 0 unless κ = 1, which we now assume. Pick a prime r with > q. If 6 q2 we may break into residue r < bY and rθ < cM −1. Thus it follows that Ψi(u) ≡ 0 mod r in (5.10), for N 6 Xv1,...,vd−1 # {u1, . . . , ud−1 6 q : ui ≡ vi mod , for 1 6 i 6 d − 1} , classes modulo , finding that where the sum is over all v = (v1, . . . , vd−1) ∈ F(d−1)n such that Ψi(v) = 0, for all i 6 n, over F. The inner cardinality is O((q2/)(d−1)n), with an implied constant that is independent of q. We may use the Lang–Weil estimate to deduce that the outer sum is O((d−2)n), again with an implied constant that depends at most on d and n. Hence we get the overall contribution N ≪ q2(d−1)n n 6 q2(d−1)n−n. Alternatively, if > q2, we may assume that the system of equations Ψi = 0, for i 6 n, has dimension (d − 2)n over F. We now appeal to an argument of 14 T.D. BROWNING AND P. VISHE Browning and Heath-Brown [3, Lemma 4]. Using induction on the dimension, as in the proof of [3, Eq. (3.7)], we easily conclude that N ≪ (q2)(d−2)n 6 q2(d−1)n−2n, for an implied constant that only depends on d and n. Recalling that κ = 1, the first part of the lemma now follows on substituting these bounds into (5.9). We now consider case (ii), in which e = 1, q2 6 r 6 qd and rθ 6 q−d. Let a/r = q−α for 1 6 α 6 d. Let v ∈ {1, . . . , d} be such that d − v − α = −1. Then an application of Lemma 5.2 yields S(α)2d−1 6 P (2d−1−d+1)nq(v−1)nM (v)(α) = P 2d−1n · q(−2d+1+v)nM (v)(α). r Ψi(u) 6 q−α · qd−v = q−1. If we write uj = u′ j, u′′ Let u ∈ O n(d−1) be counted by M (v)(α). Since θ 6 q−d−2, it follows that θΨi(u) 6 q−d−2 · qd−v = q−2−v 6 q−3, for 1 6 i 6 n. Similarly, for 1 6 i 6 n, we have a j , for v 6 j 6 d, where u′ r Ψi(u) is equal to Ψi(u1, . . . , uv−1, u′′ d−1). The condition kαΨi(u)k < P −1 in M (v)(α) implies that this coefficient must necessarily vanish, whence M (v)(α) is at most the number of u1, . . . uv−1, u′ q for which Ψi(u1, . . . , uv−1, u′′ q , then the coefficient of t−1 in the t-expansion of a d−1 ∈ Fn v, . . . , u′′ v, . . . , u′ d−1, u′′ v, . . . , u′′ j ∈ Fn j + tu′′ d−1) = 0, for 1 6 i 6 n. Thus v, . . . , u′′ M (v)(α) ≪ q(d−v)n · q(d−2)n = q(2d−v−2)n, by the Lang–Weil estimate, which implies the statement of the lemma. (cid:3) Lemma 5.5 (deg(r) = 0). Let L = 2−d+1n and let θ ∈ T. Assume that Then there exists a constant c′′ d,n > 0, independent of q, such that q−de−1 6 θ 6 q−1−κ(d−1). S(θ) 6 c′′ d,nP nq−L. Proof. The upper bound assumed of θ implies that θΨi(u) 6 q−1 in (5.10), for 1 6 i 6 n. Hence kθΨi(u)k = θΨi(u) for 1 6 i 6 n. Since α = θ and θ > q−de−1, it follows that the condition kαΨi(u)k < qκ(d−1)−de−1 is equivalent to Ψi(u) < qκ(d−1). If κ = 0 then it follows from (5.10) that N = #(cid:8)u ∈ F(d−1)n q : Ψi(u) = 0 (∀i 6 n)(cid:9) ≪ q(d−2)n, If, on the other hand, κ = 1 then we write by the Lang–Weil estimate. u = u′ + tu′′ in N , under which transformation Ψi(u) < qd−1 is equivalent to Ψi(u′′) = 0, for i 6 n. Applying the Lang–Weil estimate to this system of equations, we therefore deduce that N = O(q(1+κ)(d−1)n−n) for κ ∈ {0, 1}. An application of (5.9) now completes the proof of the lemma. (cid:3) RATIONAL CURVES ON SMOOTH HYPERSURFACES 15 6. The contribution from the minor arcs We assume that d > 3 throughout this section. Our goal is to prove (3.6) for all e > 1, provided that n > n0(d), where n0(d) is given by (1.2). The overall contribution to (3.4) from θ < q−3de is easily seen to be negligible. Hence we may redefine the minor arcs to incorporate the condition θ > q−3de. For α, β ∈ Z>0, let E(α, β) denote the overall contribution to Nminor(P ), from values of a, r, θ for which r = qα and θ = q−β. The contribution is empty unless 0 6 α 6 and α + 6 β 6 3de, (6.1) d(e + 1) 2 d(e + 1) 2 with β 6 de + 1 if α = 0. Since there are only finitely many choices of α, β, in order to prove (3.6), it will suffice to show that q−bµE(α, β) = 0, lim q→∞ for each pair (α, β) under consideration, assuming that n > n0(d). To begin with, summing trivially over a, we have E(α, β) 6 q2α−β+1 max a,r,θ a<r=qα θ=q−β S(a/r + θ). (6.2) We start by dealing with generic values of α and β. Lemma 5.3 implies that E(α, β) 6 cd,nq2α−β+1+(e+1)n−L(e+1)η, where L = 2−d+1n. Recalling the definition (3.2) of bµ, the exponent of q is bµ −bν, with For the choice of η in (5.7), and n > n0(d), we want to determine whenbν > 0. bν = {(n − d)e + n − 1} − {2α − β + 1 + (e + 1)n − L(e + 1)η} Returning to (5.6), we now see that = L(e + 1)η + β − de − 2α − 2. (6.3) Γ = min {(e + 1)(d − 1) − 1, β − α − 1, (e + 1)d − α − 1, α + M} , 1 d − 1 where M = max{0, (e + 1)d − β}. The remainder of the argument is a case by case analysis. When [Γ] 6 1 we shall return to (6.2), and argue differently based instead on Lemmas 5.4 and 5.5. Case 1: α > 2(d − 1) and β > (e + 1)d + 1. In this case M = 0. Using (6.1), one finds that Γ = 1 d − 1 ×(α α − 1 if α < d(e+1) if α = d(e+1) 2 2 , . 16 T.D. BROWNING AND P. VISHE Let ι ∈ {0, 1}. We write α−ι = k(d−1)+ℓ, for k ∈ Z>0 and ℓ ∈ {0, . . . , d−2}. Then (5.7) implies that (e + 1)η = k − δ, where δ =(0 1 if k 6≡ e mod 2, if k ≡ e mod 2. (6.4) We claim that the assumption α > 2(d − 1) implies that k > 2, or else k = 1 and δ = 0. This is obvious when α < d(e+1) . Suppose that k = 1 and α = d(e+1) . Since 2 d > 3, this equation has no solutions in odd integers e. Thus δ = 0. . Then ι = 0 and ℓ = d − 2, whence α = 2(d − 1) = d(e+1) 2 2 Recalling (6.3) and substituting for α, we find that bν = L(k − δ) + β − de − 2 − 2ι − 2k(d − 1) − 2ℓ = (L − 2(d − 1))k − δL + β − de − 2 − 2ι − 2ℓ > (L − 2(d − 1))k − δL − d + 3 − 2ι, since β > (e + 1)d + 1 and ℓ 6 d − 2. Taking 3 − 2ι > 0, we have therefore shown thatbν >bν0, with If k > 2, then we take δ 6 1 to conclude that bν0 = (L − 2(d − 1))k − δL − d. follows that bν0 > (2 − δ)L − 4(d − 1) − d > L − 5d + 4. bν0 = L − 2(d − 1) − d = L − 3d + 2, Thusbν0 > 0 if n > n0(d). Alternatively, if k = 1 then we must have δ = 0. It whencebν0 > 0 if n > n0(d), since n0(d) > 2d−1 · (3d − 2) in (1.2). Case 2: α + de − d + 2 > β and β 6 (e + 1)d. In this case M = (e + 1)d − β. It follows from (6.1) that Γ = 1 d − 1 ×(α + (e + 1)d − β α + (e + 1)d − β − 1 if β > 2α, if β 6 2α. We proceed as before. Thus for ι ∈ {0, 1}, we write α + (e + 1)d − β − ι = k(d − 1) + ℓ, (6.5) with k ∈ Z>0 and ℓ ∈ {0, . . . , d − 2}. Then (5.7) implies that (e + 1)η = k − δ, where δ is given by (6.4). If k > 2 then (6.3) yields bν = L(k − δ) − β + de − 2 − 2ι − 2k(d − 1) − 2ℓ + 2d = (L − 2(d − 1))k − δL − β + de − 2 − 2ι − 2ℓ + 2d > L − 4d + 4 − β + de, RATIONAL CURVES ON SMOOTH HYPERSURFACES 17 since δ, ι 6 1 and ℓ 6 d − 2. But β 6 (e + 1)d, and so it follows that taking ι 6 1 and ℓ 6 d − 2 in (6.5), we must have that bν > L − 5d + 4, which is positive if n > n0(d). Suppose that k 6 1. Then, on α + de − d + 2 6 β, which contradicts the hypothesis. Case 3: α 6 2(d − 1) and β > (e + 1)d + 1. In this case we return to (6.2), and we recall the definition (5.8) of κ. Suppose first that α = 0. It follows from Lemma 5.5 that S(a/r + θ) ≪ P nq−L if 1 + κ(d − 1) 6 β 6 de + 1. The upper bound β 6 de + 1 follows from the definition of the minor arcs when α = 0. Moreover, the lower bound holds, since for e > 1 it follows from (6.1) that β > d > 1 + κ(d − 1). Recalling (3.2), we conclude that E(α, β) ≪ q−β+1+(e+1)n−L = q bµ−bν, Suppose next that α > 1. Then S(a/r + θ) ≪ P nq−L, by Lemma 5.4, withbν = L + β − de − 2 > L > 0, which is satisfactory. provided that e > 1, 1 6 α < de + 1 − κ(d − 1) and α − β < −κ(d − 1), (6.6) or e = 1, 2 6 α 6 d and α − β 6 −d. (6.7) In view of (6.1), it is easily seen that α−β < −(d−1) 6 −κ(d−1). Next, we claim that 2d − 2 < de + 1 − κ(d − 1) for any e > 2. This is enough to confirm (6.6), since α 6 2(d − 1). The claim is obvious when κ = 1 and e > 3. On the other hand, if κ = 0 then e > 2 and it is clear that 2d − 2 6 2d + 1 6 de + 1. Next, suppose that e = 1, so that κ = 1. If α = 1 then we are plainly in the situation covered by (6.6). If α > 2, on the other hand, then (6.1) implies that α 6 d and α − β 6 −d, so that we are in the case covered by (6.7). It follows that E(α, β) ≪ q2α−β+1+(e+1)n−L = q bµ−bν, with bν = L + β − de − 2 − 2α > L + d − 1 − 2α > L − 3d + 3, since α 6 2(d − 1) and β > (e + 1)d + 1. This is positive for n > n0(d). 18 T.D. BROWNING AND P. VISHE Case 4: α+de−d+2 6 β and β 6 (e+1)d. We begin as in the previous case. If α = 0, the same argument goes through, leading to E(α, β) ≪ q bµ−bν, with bν = L + β − de − 2 > L − d. This is certainly positive for n > n0(d). Suppose next that α > 1. Then S(a/r + θ) ≪ P nq−L, by Lemma 5.4, provided that (6.6) or (6.7) hold. Note that α 6 β − de + d − 2 6 2d − 2 < de + 1 − κ(d − 1), for any e > 2, by the calculation in the previous case. Likewise, the previous argument shows that we are covered by (6.6) or (6.7) when e = 1. Thus we find that E(α, β) ≪ q bµ−bν, with bν = L + β − de − 2 − 2α > L − d − α > L − 3d + 2, since α 6 2(d − 1). This is also positive for n > n0(d). References [1] D. Bourqui, Moduli spaces of curves and Cox rings. Michigan Math. J. 61 (2012), 593–613. [2] D. Bourqui, Exemples de comptages de courbes sur les surfaces. Math. Annalen 357 (2013), 1291–1327. [3] T.D. Browning and D.R. Heath-Brown, Rational points on quartic hypersurfaces. J. reine angew. Math. 629 (2009), 37–88. [4] T.D. Browning and P. Vishe, Cubic hypersurfaces over Fq(t). Geom. Funct. Anal. 25 (2015), 671–732. [5] R. Beheshti, Hypersurfaces with too many rational curves. Math. Annalen 360 (2014), 753-768. [6] R. Beheshti and N.M. Kumar, Spaces of rational curves on complete intersections. Compositio Math. 149 (2013), 1041–1060. [7] I. Coskun and J. Starr, Rational curves on smooth cubic hypersurfaces. Int. Math. Res. Not. 24 (2009), 4626–4641. [8] A. Grothendieck, ´El´ements de g´eom´etrie alg´ebrique: IV3. ´Etude locale des sch´emas et des morphismes de sch´emas. Pub. Math. I.H. ´E.S. 28 (1966), 5–255. [9] J. Harris, M. Roth and J. Starr, Rational curves on hypersurfaces of low degree. J. reine angew. Math. 571 (2004), 73–106. [10] S. Lang and A. Weil, Number of points of varieties in finite fields. Amer. J. Math. 76 (1954), 819–827. [11] S.A. Lee, On the applications of the circle method to function fields, and related topics. Ph.D. thesis, University of Bristol, 2013. [12] S.A. Lee, Birch's theorem in function fields. Preprint, 2011. (arXiv:1109.4953) [13] J. Koll´ar, Rational curves on algebraic varieties. Springer-Verlag, 1996. [14] T. Pugin, An algebraic circle method. PhD thesis, Columbia University, 2011. [15] E. Riedl and D. Yang, Kontsevich spaces of rational curves on Fano hypersurfaces. J. reine angew. Math., to appear. RATIONAL CURVES ON SMOOTH HYPERSURFACES 19 [16] J.-P. Serre, How to use finite fields for problems concerning infinite fields. Arithmetic, geometry, cryptography and coding theory, 183–193, Contemporary Math. 14, Ameri- can Math. Soc., 2009. [17] M. Usher and J. Zhang, Persistent homology and Floer–Novikov theory. Geometry and Topology, to appear. School of Mathematics, University of Bristol, Bristol, BS8 1TW, UK E-mail address: [email protected] Department of Mathematical Sciences, Durham University, Durham, DH1 3LE, UK E-mail address: [email protected]
1703.05720	4	1703	2018-06-02T00:54:29	On the bad reduction of certain U(2, 1) Shimura varieties	[ "math.AG", "math.NT" ]	Let $E$ be a quadratic imaginary field and let $p$ be a prime which is inert in $E.$ We study three types of Picard modular surfaces in positive characteristic $p$ and the morphisms between them. The first Picard surface, denoted $S$, parametrizes triples $(A,\phi,\iota)$ comprised of an abelian threefold $A$ with an action $\iota$ of the ring of integers $\mathcal{O}_{E}$, and a principal polarization $\phi$. The second surface, $S_{0}(p)$, parametrizes, in addition, a suitably restricted choice of a subgroup $H\subset A[p]$ of rank $p^{2}$. The third Picard surface, $\widetilde{S}$, parametrizes triples $(A,\psi,\iota)$ similar to those parametrized by $S$, but where $\psi$ is a polarization of degree $p^{2}$. We study the components, singularities and naturally defined stratifications of these surfaces, and their behaviour under the morphisms. A particular role is played by a foliation we define on the blow-up of $S$ at its superspecial points.	math.AG	math	ON THE BAD REDUCTION OF CERTAIN U (2, 1) SHIMURA VARIETIES EHUD DE SHALIT AND EYAL Z. GOREN Dedicated to V. Kumar Murty on the occasion of his 60th birthday Abstract. Let E be a quadratic imaginary field and let p be a prime which is inert in E. We study three types of Picard modular surfaces in positive characteristic p and the morphisms between them. The first Picard surface, denoted S, parametrizes triples (A, φ, ι) comprised of an abelian threefold A with an action ι of the ring of integers OE, and a principal polarization φ. The second surface, S0(p), parametrizes, in addition, a suitably restricted choice of a subgroup H ⊂ A[p] of rank p2. The third Picard surface, (cid:101)S, parametrizes triples (A, ψ, ι) similar to those parametrized by S, but where ψ is a polarization of degree p2. We study the components, singularities and naturally defined stratifications of these surfaces, and their behaviour under the morphisms. A particular role is played by a foliation we define on the blow-up of S at its superspecial points. Contents Introduction A summary of the results 1. Three integral models with Parahoric level structure 1.1. Shimura varieties 1.2. The moduli problems 1.3. Translation into the language of Rapoport and Zink 1.4. Modular curves on the Picard modular surface 2. The structure of the special fiber of S 2.1. Stratification 2.2. The tangent bundle of S 2.3. The blow up of S at the superspecial points 3. Local structure of the three integral models 3.1. Raynaud's classification 3.2. The completed local rings 3.3. Computations 4. The global structure of S0(p) 4.1. The µ-ordinary strata 4.2. The gss strata 4.3. The ssp strata 5. The structure of (cid:101)S 5.1. The global structure of (cid:101)S 2000 Mathematics Subject Classification. 11G18, 14G35. Key words and phrases. Picard surfaces, Shimura varieties, supersingular strata. 1 2 3 6 6 7 12 13 15 15 16 20 20 20 22 26 28 28 35 41 46 46 ON THE BAD REDUCTION OF CERTAIN U (2, 1) SHIMURA VARIETIES 5.2. Analysis of(cid:101)π 6. Appendix 6.1. The classification of the gss Dieudonné modules 6.2. The quotient A/H for gss A References 2 51 55 55 57 58 Introduction Let E be a quadratic imaginary field and let p be a prime which is inert in E. This paper is concerned with the detailed study of three types of Picard modular surfaces in positive characteristic p and the morphisms between them. Deferring precise definitions to the body of the paper, the first Picard surface, denoted S, parametrizes triples (A, φ, ι) comprised of a certain abelian threefold A with an action ι of the ring of integers OE, and a principal polarization φ. Unlike the other two, S is smooth. The second surface, S0(p), parametrizes, in addition, a suitably restricted choice of a subgroup H ⊂ A[p] of rank p2. The third Picard surface, (cid:101)S, parametrizes triples (A, ψ, ι) similar to those parametrized by S, but where ψ is a polarization of degree p2. There are natural morphisms providing us with a diagram S0(p) π π S S. From another perspective, there are three Shimura varieties associated with the unitary group of E of signature (2,1), having parahoric level structure at p. The above mentioned moduli spaces are the special fibers at p of the integral models of these Shimura varieties, studied by Rapoport and Zink in [Ra-Zi]. Before describing the main results of this article, we provide some background, context and motivation. Picard modular surfaces appear in many places in the literature; the book by Langlands and Ramakrishnan [La-Ra] provides a strong motivation for their study as a test case for the Langlands conjectures on modularity of L-functions, as well as a guide to the literature at the time. The local structure at p of S0(p) and related moduli spaces was studied in Bellaïche's thesis [Bel], and later in the work of Bültel-Wedhorn [Bu-We] and Koskivirta [Kos], where the authors applied it to lifting problems of Picard modular forms, Galois representations, and congruence relations for Hecke operators. However, the global structure of S0(p) and of the map S0(p) → S remained opaque. Thus, one of our original motivations was to make this global structure precise. Unlike S0(p), there is little information in the literature on (cid:101)S, or in general indeed (cid:101)S, in contrast to loc. cit., has proven to be extremely well-behaved. on moduli spaces of abelian varieties with non-separable polarizations. The main examples we are aware of are [Cri, dJ1, Nor, N-O], and they tend to exhibit rather pathological phenomena. It is desirable to have additional examples available, and Our main reason for studying the three Picard modular surfaces, was however different. Motivated by questions on the canonical subgroup, or by the search for a geometric proof of the congruence relation (as in [Bu-We, Kos]), it is desirable to ! ! } } ON THE BAD REDUCTION OF CERTAIN U (2, 1) SHIMURA VARIETIES 3 have a surface parametrizing tuples (A, φ, ι, H), where H is a finite flat subgroup scheme which may reduce mod p to the kernel of Frobenius. As this kernel has rank p3 and in characteristic 0 the rank of a p-primary OE-subgroup must be an even power of p, such a surface does not exist. To remedy the situation, one is forced to consider a moduli space as above, but where H is now of rank p6. In the context of modular curves this is akin to passing from X0(p) to X0(p2); a process which is, of course, unnecessary for modular curves, but would be required for many Shimura varieties. It turns out that it is beneficial to modify the moduli problem somewhat and following [dJ2] to consider a filtration of H as part of the datum. That is, (roughly) the following data: (A, φ, ι, H0 ⊆ H), where (A, φ, ι, H0) is an object parametrized by S0(p) and H is a suitable rank p6 finite flat subgroup scheme. We call this moduli problem T , and one of our initial observations is that T ∼= S0(p) ×(cid:101)S S0(p). In characteristic 0, this surface is finite flat of degree (p + 1)(p3 + 1) over S, and represents the Hecke operator T (p). This, therefore, motivated both the introduc- tion of (cid:101)S and the study of the morphism(cid:101)π. The study of the moduli space T will While studying the three moduli spaces S, S0(p) and (cid:101)S, we discovered a new be carried out in a subsequent paper. Nonetheless, the foundations are laid down here. interesting phenomenon. The generic stratum of S in characteristic p parametrizes µ-ordinary abelian threefolds. Although their p-divisible groups are all isomorphic, studying their cotangent spaces we were able to distinguish in the tangent space of S a certain "foliation", amounting in this very simple example to a line sub-bundle closed under the operation of raising to power p (see §2.2). The link between the cotangent space of the universal abelian variety and that of S is supplied by the Kodaira-Spencer map. This foliation extends to the general supersingular locus of S, but fails to extend, in a way made precise, to the superspecial points there. Moreover, we found two other ways to characterize it: the first, as the foliation of "unramified directions" (in the sense of [Ru-Sh]) for a map π : S0(p)(p) → S derived from the map π (Theorem 4.4). The second, in terms of Moonen's generalized Serre- Tate coordinates [Mo] (Proposition 2.4). Shimura curves embedded in S, as well as the supersingular curves in S, are integral curves of this foliation (Theorem 2.3). Does it have any other global integral curves? We expect this new phenomenon to generalize to other Shimura varieties of PEL type whose generic stratum is µ- ordinary but not ordinary; cf. our forthcoming paper [dS-G3] where such a foliation is studied for unitary Shimura varieties of arbitrary signatures. A summary of the results. We now describe briefly the content of this paper. Chapter 1 reviews the three Shimura varieties and their integral models. We explain the precise relation between the moduli problem with parahoric level structure as in [Ra-Zi] and the Raynaud condition appearing in [Bel]. The last section reviews the embeddings of modular curves and Shimura curves in the Picard modular surface. Chapter 2 deals with the Picard modular surface S, where the level at p is a hy- perspecial maximal compact. The mod p fiber is smooth, and its stratification was studied by Vollaard in [Vo]. It consists of three strata. The dense open stratum Sµ parametrizes µ-ordinary abelian threefolds. Its complement Sss parametrizes su- persingular ones, and consists (at least when the tame level N is large, depending ON THE BAD REDUCTION OF CERTAIN U (2, 1) SHIMURA VARIETIES 4 on p) of Fermat curves of degree p + 1, intersecting transversally at their Fp2- rational points. These intersection points support superspecial abelian threefolds (isomorphic, not only isogenous, to a product of supersingular elliptic curves), and constitute the third stratum Sssp. The non-singular locus of the curve Sss supports supersingular, but not superspecial, abelian threefolds, and is denoted Sgss. This is the intermediate stratum. The number of its irreducible components was deter- mined in [dS-G1] using intersection theory on S and a secondary Hasse invariant constructed there. It turns out to be related to the second Chern number of S, and via a result of Holzapfel, expressible as an L-value. Our contribution to the study of S in the present paper is: (a) We introduce the foliation T S+ in the tangent bundle of S, outside Sssp, and prove the results to which we alluded above; (b) We introduce the blow-up S# of S at Sssp and give it a modular interpretation. It has the advantage that the irreducible components of Sss become, after blowing up, disjoint non-singular Fermat curves (even when N is small), i.e. all their inter- sections, including self-intersections, are resolved. The exceptional divisor at every blown-up point x is a projective line Ex defined over Fp2 . The components of Sss intersect Ex at points ζ satisfying ζ p+1 = −1. Embedded Shimura curves, on the other hand, intersect Ex at Fp2-rational points satisfying ζ p+1 (cid:54)= −1. The proofs of these results will have to wait until Theorem 4.11 and §4.3.3. Chapter 3 is based on chapter III of Bellaïche's thesis [Bel] and describes the local models for the completed local rings of the three Shimura varieties, at any point of the special fiber. We are nevertheless interested not only in the completed local rings per se, but in the maps between them. The theory of local models yields these maps only modulo pth powers of the maximal ideal. This is evident already in the case of the germ of the map X0(p) → X between two modular curves, with and without Γ0(p)-level structure, at a supersingular point. In this "baby case" the map between the local models is k[[x]] (cid:44)→ k[[x, y]]/(xy), which is not even flat. The correct map, however, is known ever since Kronecker to be k[[x]] (cid:44)→ k[[x, y]]/((xp − y)(x − yp)), which is finite flat of degree p + 1. Similar but more serious problems arise when we study the maps between the completed local rings of our three Picard surfaces. Luckily, a general theorem of Rudakov and Shafarevich [Ru-Sh] on the local struc- ture of inseparable maps between smooth surfaces, allows us to give a partial answer to our question. In essence, it allows us to determine the maps between the com- pleted local rings of the analytic branches through any given point. Once again, results of this type have to await the study of S0(p) and (cid:101)S in subsequent chapters, where we relate them also to the foliation T S+ mentioned above. Chapter 4 is the longest, and deals with the Picard surface S0(p) of Iwahori level structure, and the map π from S0(p) to S. We caution that π is neither finite nor flat. The special fiber of S0(p) consists of vertical and horizontal components intersecting transversally. There are two horizontal components, multiplicative and étale. The multiplicative component maps under π isomorphically onto S#. The map from the étale component is purely inseparable of degree p3 and factors through Frobenius. The factored map πet is inseparable of degree p, and we show that its "field of unramified directions" is just the foliation T S+, which was defined before ON THE BAD REDUCTION OF CERTAIN U (2, 1) SHIMURA VARIETIES 5 intrinsically on S. The vertical components of π are P1-bundles over Fermat curves, which we call the "supersingular screens". Above each superspecial point x ∈ Sssp lies in S0(p) a "comb", whose base Fx is a P1 along which the two horizontal sheets of S0(p) meet, and whose "teeth" Gx[ζ] belong to the supersingular screens. For a more precise description we refer to Theorems 4.1, 4.5 and 4.11 and their corollaries. Chapter 5 deals with (cid:101)S and the map (cid:101)π. Unlike π, this map is finite flat of (cid:101)π is an isomorphism on the étale component of S0(p) and purely inseparable of degree p on the multiplicative component. The maps π and(cid:101)π allow us to go back and forth between S and (cid:101)S and produce maps that we are able to analyze easily supersingular screens) the map (cid:101)π is pretty intricate. We collect some results on it in light of the modular interpretation. On the vertical components of S0(p) (the degree p + 1. Here again there are horizontal and vertical components. This time in the last section of Chapter 5, but leave some other questions unanswered. The appendix contains some ugly but unavoidable computations with Dieudonné modules, that would have interrupted the presentation, had they been left where needed. Deformation theory of p-divisible groups clearly is a central tool in this work. Un- fortunately, there are at least three traditional approaches to it: Grothendieck's the- ory of crystals, contravariant Dieudonné theory, and covariant Dieudonné-Cartier theory (not counting displays, p-typical curves etc.). We made every effort to re- main faithful to the language and notation used by the various references cited by us. This resulted, however, in a mixture of the three approaches. A very useful guide, and a dictionary between the various languages, can be found in the appendix to [C-C-O]. Notation • If A is an abelian scheme over a base S, At denotes its dual abelian scheme. • If H is a finite flat group scheme over a base S, H D denotes its Cartier dual. • If S is a scheme over Fp we denote by ΦS : S → S the absolute Frobenius morphism of S. If X → S is any scheme, we denote by X (p)/S, or simply by X (p), if no confusion may arise, the fiber product and by F rX/S : X → X (p) the unique morphism over S such that X (p) = S ×ΦS ,S X (ΦS × 1) ◦ F rX/S = ΦX . • If A is an abelian scheme over S then Fr = F rA/S : A → A(p) is an isogeny (the Frobenius of A). The Verschiebung Ver : A(p) → A is the isogeny dual to the Frobenius of At. • If λ : A → At is a polarization of an abelian scheme A and K = ker λ, we denote by eλ : K × K → Gm the Mumford pairing on K. If λ = nφ where φ is a principal polarization, then eλ is Weil's en-pairing associated with φ. • E is a quadratic imaginary field, OE its ring of integers, p a prime that remains inert in E, κ = OE/pOE and Op is the completion of OE at p. We write σ for the non-trivial automorphism of E, extended to Op. • If R is an Op-algebra we denote by Σ the given homomorphism Op → R and Σ = Σ ◦ σ. ON THE BAD REDUCTION OF CERTAIN U (2, 1) SHIMURA VARIETIES 6 • If G is a commutative group scheme over a base S we denote by OE ⊗ G the S-group scheme representing the functor S(cid:48) (cid:55)→ OE ⊗Z G(S(cid:48)). It has an obvious OE action. • If X is a non-singular algebraic variety over a field k we denote its tangent bundle by T X. The fiber of T X at x ∈ X(k) (the tangent space at x) will be denoted by TxX = T Xx. • If X is any scheme we denote by X red the same underlying space, equipped with the reduced induced subscheme structure. Acknowledgments: We thank Ben Moonen and George Pappas for helpful dis- cussions. We are grateful to the research institutes IHES, Bures-sur-Yvette and MFO, Oberwolfach, where part of the research on this paper has been done, for their hospitality. 1. Three integral models with Parahoric level structure 1.1. Shimura varieties. Let E be a quadratic imaginary field. Let Λ = O3 E, equipped with the hermitian form  1 1 1  v, (u, v) = t ¯u which is of signature (2, 1) over R. We denote by e0, e1, e2 the three standard basis vectors. Let G be the group of unitary similtudes GU (Λ, (, )), regarded as a linear algebraic group over Z. The Shimura varieties in the title will be associated with G. More precisely, G∞ = G(R) acts by projective linear transformations on P2(C). The bounded symmetric domain D = {(z0 : z1 : z2) z0z2 + z1z1 + z2z0 < 0} , biholomorphic to the unit ball in C2, is preserved by G∞, which acts on it transi- tively. Denote by K∞ the stabilizer of the "center" (−1 : 0 : 1). For any compact open subgroup Kf ⊂ G(Af ) we put K = K∞Kf ⊂ G(A). The Shimura variety SK is a quasi-projective variety over E whose complex points are identified, as a complex manifold, with SK(C) (cid:39) G(Q) \ G(A)/K (cid:39) G(Q) \ [D × G(Af )/Kf ]. Fix an odd prime p which is inert in E, and let N ≥ 3 be an integer such that p (cid:45) N. Let κ = OE/pOE and denote by Op the ring of integers in the completion Ep. Assume that Kf = KpK p where K p ⊂ G(Ap f ) is the principal level subgroup of level N, and Kp ⊂ Gp = G(Qp). In this paper we are interested in three choices of Kp. As p is inert in E, Gp is non-split, and its semi-simple rank is 1. Its Bruhat-Tits building is a biregular tree of bi-degree (p3+1, p+1). The vertices of degree p3+1 are stabilized by hyperspecial p = G(Zp). This maximal compact subgroups of Gp, which are all conjugate to K 0 subgroup is the stabilizer of the standard self-dual lattice (1.1) Λ0 = Λ ⊗ Zp = (cid:104)e0, e1, e2(cid:105)Op . ON THE BAD REDUCTION OF CERTAIN U (2, 1) SHIMURA VARIETIES 7 The vertices of degree p + 1 are stabilized by special, but not hyperspecial, maximal compact subgroups, which are all conjugate to the stabilizer (cid:101)K 0 p of the lattice Λ1 = (cid:104)pe0, e1, e2(cid:105)Op . Λ2 = (cid:104)pe0, pe1, e2(cid:105)Op . Note that this is also the stabilizer of p−1Λ2, the dual lattice with respect to the hermitian pairing, where p and (cid:101)K 0 stabilized by the standard Iwahori subgroup We call the vertices of degree p3 + 1 vertices of type (hs) and the ones of degree p are called p + 1 of type (s). The vertices v0 and (cid:101)v0 corresponding to K 0 the standard vertices of the respective types. The oriented edge (v0,(cid:101)v0) is then We denote by S (resp. (cid:101)S, resp. S0(p)) the Shimura variety over E of level p (resp. (cid:101)K 0 Proposition 1.1. The Shimura varieties S, (cid:101)S and S0(p) are non-singular quasi- Kf = KpK p, where K p is as above (of full tame level N) and Kp = K 0 resp. K 1 p). The following result is well-known. p ∩ (cid:101)K 0 p , K 1 p = K 0 p . projective surfaces over E and the natural maps π : S0(p) → S, (cid:101)π : S0(p) → (cid:101)S are finite étale of degrees p3 + 1 and p + 1 respectively. We denote by S (resp. (cid:102)S , resp. S0(p)) the integral models of these varieties over Op constructed in chapter 6 of [Ra-Zi]. They are of relative dimension 2, S is smooth over Op, but the other two are not. The relative surface S is the integral model of the Picard modular surface which has been studied in detail by Vollaard [Vo] §§4-6. See [dS-G1] for related results. The surface S0(p) has been studied to some extent in Bellaïche's thesis [Bel]. Previous to this paper, little was known about (cid:102)S , apart from the general facts that follow from [Ra-Zi]. We review these three integral models in the next section. From a general theorem of Görtz [Gö], or from the computations of the local models cited in §3.2, it follows that all three integral models are flat over Op, and their special fibers are reduced. As we shall later show, they are also regular. 1.2. The moduli problems. 1.2.1. The Raynaud condition. Let R be a commutative Op-algebra and H a finite flat group scheme over R of rank p2. Assume that we are given a ring homomorphism ι : OE → EndR(H), and that H is killed by p, or, equivalently, ι factors through the field κ = OE/pOE. Locally on Spec(R), O(H) = A is free of rank p2; the zero section of H is given by an R-homomorphism : A → R whose kernel I, the augmentation ideal, is free of rank p2 − 1. Letting a ∈ κ× act on A via ι(a)∗, this becomes a group action, which preserves I. Let ω : κ× → O× p → R× be the Teichmüller character, and for 1 ≤ i ≤ p2 − 1 let I (i) = {f ∈ I∀a ∈ κ×, ι(a)∗(f ) = ωi(a)f}. Thanks to the fact that p2 − 1 is invertible in R, these are distinct R-submodules, and I is their direct sum. Following [Bel, Ray], we call H Raynaud if each I (i) is free of rank 1 over R. The following facts are easily checked. F : M (p) → M, V : M → M (p), where M (p) = k ⊗σ,k M and σ(x) = xp is the Frobenius on k. The action of κ on H induces an action of κ on M; we let M (Σ) be the subspace on which κ acts through the natural embedding Σ : κ (cid:44)→ k, and M (Σ) the subspace on which it acts via Σ = σ ◦ Σ. Then M = M (Σ) ⊕ M (Σ). Note that (M (p))(Σ) = (M (Σ))(p) and vice versa. We call M balanced if both M (Σ) and M (Σ) are 1-dimensional. Lemma 1.2. H is Raynaud if and only if M (H) is balanced. Proof. We may assume that k is algebraically closed, as both conditions are invari- ant under passage to an algebraic closure. If H is étale, it is constant, and must then be isomorphic, with the OE action, to OE ⊗ Z/pZ, whose Dieudonné module is evidently balanced. Similarly, if H is multiplicative. There remains the local–local case. As a scheme, stripped of the group structure, H is then either (i) Spec(k[X]/(X p2 )) or (ii) Spec(k[X, Y ]/(X p, Y p)), where the second case occurs if and only if H is killed by the Frobenius morphism Fr : H → H (p). Since I is of codimension 1 and, in the local case, also nilpotent, it coincides with the maximal ideal of A = O(H). The cotangent space at the origin, I/I 2, is then kX in case (i) and kX ⊕ kY in case (ii). In case (i) κ may act on the one-dimesional I/I 2 by Σ or Σ, and so does the group κ× act. Either way, κ× acts on I i/I i+1 (1 ≤ i ≤ p2 − 1) via Σi (or Σ ), so every character ωi : κ× → k× must occur in I with multiplicity 1, and H is automatically Raynaud. But in case (i) we also have an exact sequence of finite flat OE-group schemes i 0 → H1 → H Fr→ H (p) 1 → 0. ON THE BAD REDUCTION OF CERTAIN U (2, 1) SHIMURA VARIETIES 8 Spec(R(cid:48)). • Let R → R(cid:48) be any base change. Then if H is Raynaud, so is H ×Spec(R) • The converse holds if Spec(R) is connected. In particular, it is enough to • The constant group scheme OE ⊗ Z/pZ and its dual OE ⊗ µp are Raynaud. It follows from the three properties that étale and multiplicative (dual to étale) group schemes are automatically Raynaud. check then the Raynaud condition at one geometric point. Assume now that R = k is a perfect field containing κ. Let M = M (H) be the covariant Dieudonné module1 of H. Since H is killed by p, M is a 2-dimensional vector space over k, equipped with linear maps Here H1 = ker(Fr : H → H (p)) is a subgroup scheme of rank p, and H (p) is its image. It follows that in case (i) M (H) is an extension of M (H1)(p) by M (H1), so is automatically balanced. Case (ii) is the only case where the "balanced" condition may fail. In this case 1 Fr annihilates H so V = 0 on M = M (H) and Lie(H) = M [V ] = M (see [C-C-O], B.3.5.6-3.5.7). We find that M is balanced if and only if Lie(H), equivalently its dual I/I 2, is balanced. If this is the case, i.e. both Σ and Σ occur in I/I 2, we may choose the variables X and Y so that κ× acts on X via ω and on Y 1We adhere to the conventions of [C-C-O], Appendix B.3. Our M (H) is denoted there M∗(H). F and V can be regarded also as σ or σ−1-linear maps on M. Recall that V is induced by Fr : H → H (p) and F is induced by Ver : H (p) → H. ON THE BAD REDUCTION OF CERTAIN U (2, 1) SHIMURA VARIETIES 9 via ωp, so on X iY j (i, j < p, not both 0) it acts via ωi+jp and every character occurs with multiplicity 1 in I. Thus H is Raynaud in this case. If, on the contrary, I/I 2 is κ×-isotypical, we can not have dim I (i) = 1 for every i, and H is not Raynaud. (cid:3) Let H D denote the Cartier dual of H, which is also finite flat of rank p2, and endow it by an OE-action ιD : OE → EndR(H D) via the formula ιD(a) = ι(a)t, i.e. for any R-algebra R(cid:48) and any x ∈ H(R(cid:48)), y ∈ H D(R(cid:48)), (cid:10)x, ιD(a)y(cid:11) = (cid:104)ι(a)x, y(cid:105) ∈ (R(cid:48))×. Corollary 1.3. H is Raynaud if and only if H D is Raynaud. Proof. M (H D) (identified with the contravariant Dieudonné module of H) is the (cid:3) k-linear dual of M (H), so one is balanced if and only if the other is. 1.2.2. The moduli problem (S). We now define the three integral models for the Shimura varieties with parahoric level structure at p as moduli schemes for moduli problems of PEL type. It is well known and easy to check that in the generic fiber these moduli problems yield the given Shimura varieties. For the relation with the models defined by Rapoport and Zink, and the representability of the moduli problems, see 1.3 below. The Picard modular surface S has a smooth integral model S over Op. It is a fine moduli scheme for the moduli problem which assigns to each Op-algebra R isomorphism classes of tuples A = (A, φ, ι, η), where ∼→ At is a principal polarization. • A is an abelian 3-fold over R. • φ : A • ι : OE → EndR(A) is a ring homomorphism, such that the Rosati involution induced by φ on EndR(A) preserves its image, and is given on it by ι(a) (cid:55)→ ι(a). We furthermore require that Lie(A) becomes an OE-module of type (2, 1) in the sense that it is the direct sum of a locally free R-module of rank 2 on which ι(a)∗ acts like the image of a in R, and a locally free rank 1 module on which it acts like a. • η:Λ/N Λ (cid:39) A[N ] is a full level-N OE-structure (recall p (cid:45) N ≥ 3). Our reference to moduli problems and representability is the comprehensive volume by Lan. In particular, we refer the reader to the precise definition of level structure given there ([Lan] 1.3.6.2), and to the condition of étale liftability. In addition to being compatible with the OE-action, η should carry the polarization pairing (cid:104),(cid:105) : Λ/N Λ × Λ/N Λ → Z/NZ derived from (, ) to the Weil eN-pairing induced by φ on A[N ] × A[N ]. Part of the data involved in η is an isomorphism between the (étale) target groups of the two pairings: νN : Z/NZ (cid:39) µN, making the last condition meaningful. These isomorphisms form a torsor Isom(Z/NZ, µN ) under (Z/NZ)×, and in this way νN becomes a morphism form S to Isom(Z/NZ, µN ), regarded as a scheme over Op of relative dimension 0. We call νN the multiplier morphism. ON THE BAD REDUCTION OF CERTAIN U (2, 1) SHIMURA VARIETIES 1.2.3. The moduli problem ((cid:101)S). The Shimura variety (cid:101)S has an integral model (cid:102)S over Op. It is a fine moduli scheme for the moduli problem which assigns to each Op-algebra R isomorphism classes of tuples A(cid:48) = (A(cid:48), ψ, ι(cid:48), η(cid:48)), where 10 • A(cid:48) is an abelian 3-fold over R. • ψ : A(cid:48) → A(cid:48)t is a polarization of degree p2. • ι(cid:48) : OE → EndR(A(cid:48)) is a ring homomorphism, satisfying the same require- ments as for (S). In addition, we require that ker(ψ) is preserved by ι(cid:48)(OE) and is Raynaud. • η(cid:48) is a full level-N OE-structure. 1.2.4. The moduli problem (S0(p)). The Shimura variety S0(p) has an integral model S0(p) over Op. It is a fine moduli scheme for the moduli problem which assigns to each Op-algebra R isomorphism classes of tuples (A, H) = (A, φ, ι, η, H), where • A is as in (S) • H ⊂ A[p] is a Raynaud OE-subgroup scheme of rank p2, which is isotropic for the Weil pairing ep (the Mumford pairing epφ attached to the polariza- tion pφ). 1.2.5. The maps between the integral models. There are projection maps π : S0(p) → S , (cid:101)π : S0(p) → (cid:102)S extending the maps of Proposition 1.1. The map π is neither finite, nor flat anymore. On the moduli problem, it is simply "forget H". The second map(cid:101)π is defined as follows. Pick (A, H) ∈ S0(p)(R). Let A(cid:48) = A/H. Since H is isotropic for epφ, its annihilator in this pairing is a finite flat subgroup scheme H ⊂ H⊥ ⊂ A[p] and A[p]/H⊥ (cid:39) H D. We claim that H⊥/H is Raynaud. We may assume that R = k is an algebraically closed field of characteristic p. As both H and H D are Raynaud, M (H) and M (A[p]/H⊥) are balanced. It follows that M (H⊥/H) is also balanced, so H⊥/H is Raynaud. The polarization pφ descends canonically to a polarization ψ : A(cid:48) → (A(cid:48))t whose kernel is ker(ψ) = H⊥/H. Its degree is p2. Finally ι(cid:48) and η(cid:48) are defined naturally from ι and η. To check that we obtained a point of (cid:102)S , we need only check one non-trivial2 point, that Lie(A(cid:48)) is indeed of type (2, 1). This can be seen, using the Raynaud condition, as follows. We may assume again that R = k is an algebraically closed field containing κ. The exact sequence 0 → H → A → A(cid:48) → 0 yields, in covariant Dieudonné theory, exact sequences3 and a commutative diagram 0 → M (A) → M (A(cid:48)) → M (H) → 0 0 → M (A)(p) → M (A(cid:48))(p) → M (H)(p) → 0 ↓ V ↓ V ↓ V 2In characteristic 0, or if H is étale, this is obvious, because the Lie algebra is not changed, but in characteristic p the type of the Lie algebra may well change under an isogeny. 3A guide for the perplexed: the covariant Dieudonné modules of a finite flat group scheme (resp. p-divisible group) is defined as the contravariant Dieudonné module of its Cartier (resp. Serre) dual. From the exact sequence 0 → H D → A(cid:48)t → At → 0 we get the top row of the diagram. ON THE BAD REDUCTION OF CERTAIN U (2, 1) SHIMURA VARIETIES 11 where we have abbreviated M (A) = M (A[p∞]) etc. The snake lemma yields 0 → M (H)[V ] → M (A)(p)/V M (A) → 0 → Lie(H) → → Lie(A) (cid:107) (cid:107) → M (A(cid:48))(p)/V M (A(cid:48)) → M (H)(p)/V M (H) → 0 → → M (H)(p)/V M (H) → 0 Lie(A(cid:48)) (cid:107) (cid:107) Thus the type of Lie(A(cid:48)) is also (2, 1) if and only if M (H)[V ] and M (H)(p)/V M (H) have the same type. But from the exact sequence 0 → M (H)[V ] → M (H) V→ M (H)(p) → M (H)(p)/V M (H) → 0 we see that this is the case if and only if M (H) is balanced. We conclude that H being Raynaud is in fact a necessary and sufficient condition for A(cid:48) = A/H to be of type (2, 1) as well. We shall see later that in contrast to π, the map(cid:101)π is finite flat of degree p + 1. If we denote by f : A → A(cid:48) the canonical homomorphism with kernel H, and identify A(cid:48)t with A/H⊥, then f(cid:48)t : A(cid:48)t → At has kernel A[p]/H⊥ and pφ = f t ◦ ψ ◦ f. 1.2.6. The moduli problem ((cid:101)S0(p)). There is a fourth moduli problem that one can The moduli problem ((cid:101)S0(p)) assigns to every Op-algebra R isomorphism classes • A(cid:48) is as in ((cid:101)S) define. It turns out to be equivalent to (S0(p)), yet useful for later calculations and for the study of the moduli problem T mentioned in the introduction. of tuples (A(cid:48), J) where • J ⊂ A(cid:48)[p] is a finite flat OE-subgroup scheme of rank p4, containing ker(ψ), such that J/ ker(ψ) is Raynaud, and which is maximal isotropic for the Mumford pairing epψ. Note that deg(pψ) = p8. Proposition 1.4. The moduli problems (S0(p)) and ((cid:101)S0(p)) are equivalent, hence ((cid:101)S0(p)) is also represented by S0(p). Proof. To pass from (A, H) to (A(cid:48), J) define A(cid:48) = A/H, J = A[p]/H, and observe that J/ ker(ψ) = A[p]/H⊥ is Raynaud, and that J is isotropic (hence, from degree considerations, maximal isotropic) for epψ. To pass from (A(cid:48), J) to (A, H) define A = A(cid:48)/J, descend pψ to obtain a principal polarization φ on A, and let H = A(cid:48)[p]/J. We leave to the reader the verification that we obtain a point of (cid:3) ((cid:101)S0(p)), as well as that these two constructions are inverse to each other. In terms of this new interpretation of S0(p) the map(cid:101)π is simply "forget J". Proposition 1.5. The schemes S , S0(p) and (cid:102)S are regular. They are flat over Op and their special fibers are reduced. The maps π and(cid:101)π are surjective and proper. ON THE BAD REDUCTION OF CERTAIN U (2, 1) SHIMURA VARIETIES 12 Proof. The "flat" and "reduced" assertions follow from the Main Result of [Gö], and from the fact that locally for the étale topology, a neighborhood of a point in the special fiber of S0(p) or (cid:102)S is isomorphic to an open neighborhood in the local condition on J/ ker(ψ). One proves, following de Jong, that this modified moduli criterion. The Raynaud condition is a closed condition, a fact which secures the of the representability of S0(p). For the map π it is done in [Bel] III.3.2.3. For the model. Similarly, regularity follows from the determination of the completed local rings of the three schemes in [Bel] III.3.4.8. Although Bellaïche does not use the formalism of [Ra-Zi], he builds upon the earlier work of de Jong [dJ2], which except for the notation, yields identical results for the completed local rings as what one would get from the more general theory developed by Rapoport and Zink. Properness and surjectivity of π and (cid:101)π are usually proved along with the proof map(cid:101)π the proof is similar, and we only sketch it. It is best described with the new interpretation of S0(p) as representing the moduli problem ((cid:101)S0(p)). Consider first a larger moduli problem ((cid:101)S0(p)(cid:48)) obtained from ((cid:101)S0(p)) by relaxing the Raynaud problem is proper and surjective over ((cid:101)S). Properness follows from the valuative properness of(cid:101)π. Surjectivity clearly holds in the generic fiber. By [Gö], the generic fiber of (cid:102)S is dense. Since (cid:101)π is already known to be proper, its image must be (cid:3) closed, hence is everything. 1.2.7. Diamond operators. If a ∈ (OE/NOE)× we denote by (cid:104)a(cid:105) the automorphism of S , defined on the moduli problem by (cid:104)a(cid:105) (A, φ, ι, η) = (A, φ, ι, η ◦ a) = (A, φ, ι, ι(a) ◦ η). The same notation will be applied to the other moduli schemes. 1.3. Translation into the language of Rapoport and Zink. The moduli prob- lems that we defined in the preceding sections are examples of the moduli problems defined in chapter 6 of [Ra-Zi], although the Raynaud condition is implicit there, as we shall now explain. It follows (from general results of Kottwitz) that, as has been claimed above, they are indeed representable by fine moduli schemes when N ≥ 3. We remark that [Bel] gives an independent proof of the representability of (S0(p)) by proving that it is relatively representable over (S). Using the notation of [Ra-Zi] we take B = E, OB = OE, V = E3 as before and b∗ = b. Let L,(cid:101)L and L0(p) be the following self-dual lattice chains in Vp (see (1.1)): L = {··· ⊂ pΛ0 ⊂ Λ0 ⊂ p−1Λ0 ⊂ ···}, (cid:101)L = {··· ⊂ pΛ1 ⊂ Λ2 ⊂ Λ1 ⊂ p−1Λ2 ⊂ ···}, L0(p) = {··· ⊂ pΛ0 ⊂ Λ2 ⊂ Λ1 ⊂ Λ0 ⊂ p−1Λ2 ⊂ ···}. View the three lattice chains as categories, inclusions as morphisms. The moduli problem of type (L), as defined in [Ra-Zi] Definition 6.9, is clearly our (S); just set A = AΛ0. The moduli problem of type ((cid:101)L) is our ((cid:101)S). Recall the definition of a "principally polarized (cid:101)L-set of abelian varieties of type (2, 1)" over a base ring R as above ([Ra-Zi], Definition 6.6). First, one is given the (cid:101)L-set of abelian schemes AΛ• of ON THE BAD REDUCTION OF CERTAIN U (2, 1) SHIMURA VARIETIES type (2, 1). Then one gives the "principal polarization"4 λ :AΛ• (cid:39) (cid:101)AΛ•. Note that the (cid:101)L-set (cid:101)AΛ• is of type (1, 2) because λ induces the Rosati involution on the 13 endomorphism ring, hence switches types. We set A(cid:48) = AΛ2 , A(cid:48)t = At Λ2 (cid:39) At p−1Λ2 = (cid:101)AΛ1, ψ = λ ◦ ρΛ1,Λ2. Then ψ is a polarization in the ordinary sense, of degree p2 = [Λ1 : Λ2]. If R = k is an algebraically closed field in characteristic p, M (ker(ψ)) = M (AΛ1)/M (AΛ2) = Λ1/Λ2 ⊗ k ([Ra-Zi] 6.10) is balanced, so ker(ψ) is Raynaud. Conversely, if we are given data as in (S), thanks to the fact that ker(ψ) is Raynaud the signature of A(cid:48)(cid:48) = A(cid:48)/ ker(ψ) (with OE-action induced by ι(cid:48)) is (2, 1) (as explained at the end of 1.2.4), so we can define AΛ2 = A(cid:48), AΛ1 = A(cid:48)(cid:48), ρΛ1,Λ2 = the canonical homomorphism, and "polarize" the resulting (cid:101)L-set by letting λ be the unique type-reversing isomor- phism of (cid:101)L-sets satisfying ψ = λ ◦ ρΛ1,Λ2. The proof that the moduli problem of type (L0(p)) is our (S0(p)) is in principle identical, and we only sketch it. Once again, given the data (S0(p)) we construct an L0(p)-set of abelian varieties by interlacing the previous two constructions. First, letting A = AΛ0 (cid:39) ApΛ0 we use the Raynaud condition on H to ensure that A(cid:48) = A/H = AΛ2 is of type (2, 1). Then we continue and define AΛ1 = A/H⊥ and the polarization λ as before. 1.4. Modular curves on the Picard modular surface. 1.4.1. Embedding the modular curve. Maps between Shimura data induce maps between Shimura varieties. Here we have unitary groups of signature (1, 1) at infinity mapping (in many ways) to our G. These group homomorphisms give rise to morphisms of modular curves and Shimura curves to our Picard modular surface. Rather than go through the familiar yoga of Shimura data, we jump straight ahead to the moduli interpretation, thereby giving the morphism on the level of integral structures. We give only one example, which will be explored in connection with the geometry of the special fiber at p later on. Let B0 be a fixed elliptic curve defined over Op with complex multiplication by OE and CM type Σ. Such a curve exists because (p) splits completely in the Hilbert class field H of E, and if P is a prime divisor of (p) in H, B0 may be defined over OH,P = Op. The reduction of B0 modulo p is a supersingular elliptic curve defined over κ. Let φ0 : B0 (cid:39) Bt 0 be the canonical principal polarization of B0, and ι0 : OE (cid:39) End(B0). √−D a fixed square root of it in E. Assume for simplicity that D is odd and (N, D) = 1 (oth- erwise the construction below has to be modified slightly). Let Z0 be the scheme parametrizing OE-isomorphisms η0 : OE/NOE (cid:39) B0[N ]. It is étale of relative di- mension 0 over Op and comes with a "multiplier morphism" νN to Isom(Z/NZ, µN ). the moduli problem ((cid:101)S) with a tilde, hence it made sense to denote the corresponding lattice chain also (cid:101)L. In [Ra-Zi], passing to the dual (cid:101)L-set is also denoted by a tilde, hence the tilde on (cid:101)AΛ•. Recall that p (cid:45) N ≥ 3. Let −D be the discriminant of E and δ = 4We apologize for the unintentional double meaning attributed to tilde. We chose to denote ON THE BAD REDUCTION OF CERTAIN U (2, 1) SHIMURA VARIETIES 14 Write B0 = (B0, φ0, ι0, η0) ∈ Z0(R) for an R-valued point of Z0. Let X = X0(D; N ) be the modular curve parametrizing elliptic curves B with a full level N structure ν : (Z/NZ)2 (cid:39) B[N ] and a cyclic subgroup scheme M of order D. We view X as a scheme over Op. It too comes equipped with a "multiplier morpism" νN to Isom(Z/NZ, µN ). If we identify det(B[N ]) with µN via the Weil pairing, then νN = det ν. We remark that X is neither complete (the cusps are missing) nor connected (det ν is not fixed), and that every subgroup scheme M as above is étale, since D is invertible. Let R be an Op-algebra and B = (B, ν, M ) ∈ X (R). Let A1(B) be the abelian surface OE ⊗ B/(δ ⊗ M ). As D is odd, hence square-free, every class in OE/δOE is represented by a rational integer. As δ kills δ ⊗ M, this subgroup is OE-stable. It is also maximal isotropic for the Mumford pairing induced by the canonical degree D2 polarization 1 : OE ⊗ B → δ−1OE ⊗ B = (OE ⊗ B)t. φ(cid:48) The identification δ−1OE⊗B = (OE⊗B)t is such that the resulting Weil en-pairing between OE ⊗ B[n] and δ−1OE ⊗ B[n] is en(α ⊗ u, β ⊗ v) = eB n (u, v)T rE/Q(αβ), n is Weil's en-pairing on B[n]. We may therefore descend φ(cid:48) where eB 1 to obtain a principal polarization φ1 of A1(B). We let ι1 be the natural action of OE as endomorphisms of A1(B). It is of type (Σ, Σ). Let η1 = id ⊗ ν : (OE/NOE)2 (cid:39) A1[N ], a full level-N OE-structure. Let B0 ∈ Z0(R) and B ∈ X (R) be such that νN (B0) = νN (B). Define A(B0, B) = B0 × A1, φ = φ0 × φ1, ι = ι0 × ι1, η = η0 × η1. The structure A(B0, B) = (A, φ, ι, η) ∈ S (R). Indeed, the assumption νN (B0) = νN (B) allows us to define a multiplier for η so that it becomes compatible with φ, and the rest is obvious. This construction depends functorially on the input. In this way we have defined a morphism Z0 ×Isom(Z/NZ,µN ) X → S . A minor modification of this construction yields a morphism Z0 ×Isom(Z/NZ,µN ) X0(p) → S0(p), when we add a cyclic subgroup of order p to the level. 1.4.2. Endomorphism rings of Fp points of S . Let D be an indefinite quaternion algebra over Q equipped with a positive involution † and assume that E embeds in D as a †-stable subfield. Then D = E ⊕ Eξ where ξ2 > 0 is rational, ξaξ−1 = a for a ∈ E, a† = a and ξ† = ξ. Furthermore E is the unique quadratic imaginary †-stable subfield of D. Let OD be a maximal order in D such that OD ∩ E = OE. In this situation we may define the Shimura curve XD parametrizing abelian surfaces A1 with endomorphisms by OD, a principal polarization inducing † as the Rosati involution on D, and a full level N structure. ON THE BAD REDUCTION OF CERTAIN U (2, 1) SHIMURA VARIETIES 15 Precisely as for the modular curve, we get a morphism from Z0 ×Isom(Z/NZ,µN ) XD to S . Its image in S is called an embedded Shimura curve. The points of S (Fp) lying on the embedded Shimura curves all represent non- simple abelian varieties. There are, however, points A ∈ S (Fp) for which A is simple. We use the Honda-Tate theorem to construct them. More precisely, we construct A's with End0(A) = M a CM field of degree 6. Let L be a totally real non-Galois cubic field, in which p decomposes as pq, where f (p/p) = 2 and f (q/p) = 1. Then M = LE is a degree 6 CM field and p = P P splits in M, while q = Q remains inert. Let π be an element of M such that (π) = P 2hQh, where h kills the class of P 2Q in the class group of M. Then ππ = p2h for a unit of L. Replacing π with −1π2 and h with 2h we may assume that = 1. Let q = p2h. Then π is a Weil q-number, and the Honda-Tate theorem implies that there exists a simple 3-dimensional abelian variety over Fq with End(A) equal to an order of M, and whose Frobenius of degree q is π. It is easily seen that A is absolutely simple. Changing A by an isogeny if necessary we may assume that End(A)⊃ OE, and that A carries a principal polarization. Of course, End0(A) = M. Since End0(A), for any A ∈ S (Fp), must contain a 6-dimensional semi-simple Q-algebra, we see that the "most general" Fp-point of S carries an abelian variety with CM by a field of degree 6. Generic points of the special fiber of S , by contrast, have no endomorphisms except for ι(OE). 2. The structure of the special fiber of S by S, (cid:101)S and S0(p) their geometric special fibers, which are schemes defined over k. 2.1. Stratification. Let k be a fixed algebraic closure of κ. Since we shall have no use for the generic fibers of our integral models any more, we denote from now on We denote by A the universal abelian scheme over S , and by Ax its fiber over a geometric point x ∈ S(k). Let G be the unique (up to isomorphism) connected 1-dimensional p-divisible group over k of height 2. It is self-dual of slope 1/2, and isomorphic to the p- divisible group of any supersingular elliptic curve over k. Fix an embedding λ : Op (cid:44)→ Endk(G) in which a ∈ Op acts on Lie(G) via the natural homomorphism Σ : Op (cid:16) κ (cid:44)→ k, and denote the pair (G, λ) by GΣ. Let GΣ be the same p-divisible group with the embedding λ◦ σ, under which the action of a ∈ Op on Lie(G) is via Σ = Σ ◦ σ. The following theorem is due to Vollaard [Vo], in particular §6. See also [dS-G1] Theorem 2.1. Theorem 2.1. (i) The special fiber S of S is the union of 3 locally closed strata defined over κ. The µ-ordinary stratum Sµ is open and dense, and x ∈ Sµ(k) if and only if Ax[p∞] (cid:39) (OE ⊗ µp∞ ) × GΣ × (OE ⊗ Qp/Zp) as p-divisible groups with OE-action. Its complement, S − Sµ = Sss is called the supersingular locus. It is a reduced (but reducible) complete curve, and if x ∈ Sss(k) then Ax[p∞] is supersingular, i.e. its Newton polygon is of constant slope 1/2. The superspecial locus Sssp ⊂ Sss is 0-dimensional and a point x ∈ Sssp(k) if and only ON THE BAD REDUCTION OF CERTAIN U (2, 1) SHIMURA VARIETIES 16 if Ax[p∞] (cid:39) G2 Σ × GΣ as p-divisible groups with OE-action. We let Sgss = Sss − Sssp and call it the general supersingular locus. Oort's a-number a(Ax) = dimk Hom(αp,Ax[p]) is 1 if x ∈ Sµ(k) or x ∈ Sgss(k) and 3 if x ∈ Sssp(k). Let αp(Ax) be the maximal αp-subgroup of Ax[p]. The action of κ on Lie(αp(Ax)) has signature Σ in the first two cases, and (Σ, Σ, Σ) in the third case. (ii) If S(cid:48) is a connected component of S then S(cid:48) ∩ Sss is a connected component of Sss. The non-singular locus of Sss is precisely Sgss. The irreducible components of Sss are Fermat curves, whose normalizations are isomorphic to the curve C : xp+1 + yp+1 + zp+1 = 0. (iii) If N ≥ N0(p) (an integer depending on p) the following also holds. The irreducible components of Sss are already non-singular, and isomorphic to C . Any two of them intersect at most at one point, and if they intersect, this point belongs to Sssp(k) and the intersection is transversal. There are p3 + 1 superspecial points on each irreducible component of Sss, and there are p + 1 irreducible components of Sss intersecting transversally at each x ∈ Sssp(k). Let X be the geometric special fiber of the modular curve X which was con- structed5 in §1.4. It is a non-singular curve in S. The following corollary is clear from the description of the strata of S. Corollary 2.2. The curve X does not intersect Sgss. If B ∈ X(k) is such that A(B) ∈ Sssp(k) then B is supersingular, and vice versa. 2.2. The tangent bundle of S. 2.2.1. The special line sub-bundle T S+. Outside Sssp, one may define a natural line sub-bundle T S+ of the tangent bundle T S of S. For this recall the following facts from [dS-G1]. Let ΩA/S be the sheaf of relative differentials of the universal abelian variety A, and ωA = f∗ΩA/S where f : A → S is the structure morphism. Then ωA is a rank 3 vector bundle on S, can be identified with the cotangent space of A at the origin, and admits a decomposition ωA = P ⊕ L into a plane bundle P on which OE acts via Σ and a line bundle L on which it acts via Σ. Let Φ : S → S be the absolute Frobenius morphism of degree p, and A(p) = S ×Φ,S A the base change of A. Similar notation will be employed for the base change of the vector bundles P or L. The Verschiebung homomorphism VerA/S : A(p) → A induces maps VP : P → L(p), VL : L → P (p), which, outside Sssp, are both of rank 1. At the superspecial points these maps vanish. Let P0 = ker(VP ). 5We abuse notation and call the curve Z0 ×Isom(Z/NZ,µN ) X simply X . ON THE BAD REDUCTION OF CERTAIN U (2, 1) SHIMURA VARIETIES 17 Outside the superspecial points, P0 is a line sub-bundle of P. Outside Sss, the lines P (p) and VL(L) are distinct, but along Sgss they coincide. In fact, 0 V (p)P ◦ VL, which is a global section of Lp2−1, is the Hasse invariant (cf. [G-N, Appendix B]; one of the main contributions of [G-N] is the construction of the Hasse invariant for unitary Shimura varieties over totally real fields, which is substantially more difficult), and V (p)P ◦ VL = 0 is the equation defining Sss as a subscheme of S. The Kodaira-Spencer isomorphism is an isomorphism KS : P ⊗ L (cid:39) ΩS/k = T S∨. Definition. Outside Sssp, we define T S+ to be the annihilator of the line bundle KS(P0 ⊗ L). We call T S+ the special sub-bundle of T S. By an integral curve of T S+ we mean a nonsingular curve C ⊂ S − Sssp for which T S+C = T C, i.e. T S+ is tangent to C. Theorem 2.3. (i) Sgss is an integral curve of T S+. (ii) The modular curve Xord = X ∩ Sµ is an integral curve of T S+. Proof. Part (i), although not stated there in this form, was proved in [dS-G2] Proposition 3.11. For (ii) observe that if x ∈ Xord(k) ⊂ Sµ(k) then we have the decomposition Ax = B0 × A1,x where A1 is the abelian surface constructed along X from the universal elliptic curve B (and the universal cyclic subgroup of rank D) as in §1.4. For the cotangent space we have accordingly ωAx = ωAx = ωB0 ⊕ ωA1,x , where the first summand is of type Σ and the second of type (Σ, Σ). Thus Px = ωB0 ⊕ ωA1,x (Σ). As A1,x is ordinary, V is injective on ωA1,x (Σ) and P0x = ker(V : Px → L(p)x) = ωB0. As B0 is constant along X, KS(P0⊗Lx) ⊂ ΩS/kx annihilates the line TxX ⊂ TxS. This proves that TxX = T S+x as claimed. (cid:3) There are many modular curves and Shimura curves like X on S, and by sim- ilar arguments they are all integral curves of the special sub-bundle. It would be interesting to know if these are the only integral curves of T S+ in Sµ. This is an "André-Oort type" question. It would imply, in particular, that there are no in- tegral curves passing through the CM points constructed in §1.4.2. Note that in characteristic p there could be many integral curves tangent to a perfectly nice vector field. The curves x − c + λyp = 0, for varying c and λ, are all tangent to the vector field ∂/∂y in A2, and infinitely many of them pass through any given point. The correct formulation of the problem should probably ask for curves annihilated by a larger class of differential operators. Such a class should contain, besides the differential operators generated by T S+, also "divided powers". ON THE BAD REDUCTION OF CERTAIN U (2, 1) SHIMURA VARIETIES 18 2.2.2. A characterization in terms of generalized Serre-Tate coordinates. We shall now give a second characterization of T S+, which relates it to Moonen's work on generalized Serre-Tate coordinates in Sµ. For the following proposition see [Mo], Example 3.3.2 and 3.3.3(d) (case AU, r = 3, m = 1). Theorem 2.5. Let x ∈ Sµ. Then the line T S+x is tangent to the canonical copy Proposition 2.4. Let x ∈ Sµ. Let (cid:98)G be the formal group over k associated with the p-divisible group G and let (cid:98)Gm be the formal multiplicative group over k. Then the formal neighborhood Spf ((cid:98)OS,x) of x has a natural structure of a (cid:98)Gm-torsor over (cid:98)G. In particular, it contains a canonical copy of (cid:98)Gm sitting over the origin of (cid:98)G. of (cid:98)Gm in Spf ((cid:98)OS,x). At a point x lying on a modular curve X as above, the canonical copy of (cid:98)Gm is of X at x coincides with i((cid:98)Gm) as a closed formal subscheme of Spf ((cid:98)OS,x). In this the formal curve i((cid:98)Gm) may no longer be "integrated". Proof. Write (cid:98)Gm = Spf (k[[T − 1]]) with comultiplication T (cid:55)→ T ⊗ T , and let i : (cid:98)Gm (cid:44)→ Spf ((cid:98)OS,x) be the embedding of formal schemes given by Proposition 2.4. It sends the closed point 1 of (cid:98)Gm to x. Let i∗ be the induced map on tangent spaces case the theorem is a consequence of Theorem 2.3(ii). Our claim can therefore be viewed as an extension of Theorem 2.3(ii) to a general µ-ordinary point, at which identified with the classical Serre-Tate coordinate on X, i.e. the formal completion i∗ : T(cid:98)Gm1 (cid:44)→ T Sx. We have to show that i∗(∂/∂T ) annihilates KS(P0 ⊗ L)x. This is equivalent to saying that when we consider the pull back i∗A of the universal abelian scheme to (cid:98)Gm, its Kodaira-Spencer map kills P0 ⊗ L1. For this recall the definition of Let S = (cid:98)Gm and write for simplicity A for i∗A. We then have the following KS = KS(Σ) from [dS-G1], §1.4.2. commutative diagram (2.1) P = ωA/S(Σ) ↓ KS S (cid:39) ω∨ At/S(Σ) ⊗ Ω1 L∨ ⊗ Ω1 (cid:44)→ dR(A/S)(Σ) H 1 dR(A/S)(Σ) ⊗ Ω1 ↓ ∇ S S ←− H 1 in which we identified H 1(A,O) with H 0(At, Ω1)∨ and used the polarization to tion, and the tensor product is over (cid:98)OS = k[[T − 1]]. Although ∇ is a derivation, A/S, reversing types. Here ∇ is the Gauss-Manin connec- identify the latter with ω∨ KS is a homomorphism of vector bundles over (cid:98)OS. We shall show that KS(P0) = 0, where P0 = ker(V : ωA/S → ω(p)A/S) ∩ P. At this point recall the filtration 0 ⊂ F il2 = A[p∞]m ⊂ F il1 = A[p∞]0 ⊂ F il0 = A[p∞] of the p-divisible group of A over S. The graded pieces are of height 2 and OE- stable. They are rigid (do not admit non-trivial deformations as p-divisible groups with OE action) and given by gr2 = OE ⊗ µp∞, gr1 = G, gr0 = OE ⊗ Qp/Zp. ON THE BAD REDUCTION OF CERTAIN U (2, 1) SHIMURA VARIETIES 19 In our case, we can identify D(A[p∞]) with H 1 For any p-divisible group G over S denote by D(G) the Dieudonné crystal asso- ciated to G, and let D(G) = D(G)S, cf. [Gro]. The (cid:98)OS-module D(G) is endowed with an integrable connection ∇ and the pair (D(G),∇) determines D(G). dR(A/S), and the connection with the Gauss-Manin connection. The above filtration on A[p∞] induces therefore dR(A/S) which is preserved by ∇. Since the functor D is a filtration F il• on H 1 contravariant, we write the filtration as dR(A/S) ⊂ F il2H 1 dR(A/S) ⊂ F il3 = H 1 0 ⊂ F il1H 1 dR(A/S) where dR(A/S) = D(A[p∞]/F iliA[p∞]). For example, F il1H 1 As F il2A[p∞] is of multiplicative type, ker(V : H 1 contained in F il2H 1 F iliH 1 dR(A/S) is sometimes referred to as the "unit root subspace". dR(A/S)(p)) is dR(A/S). In particular, P0 ⊂ F il2H 1 Let G = A[p∞]/A[p∞]m, so that F il2H 1 puting KS on P0 we may use the following diagram instead of (2.1): dR(A/S). dR(A/S) = D(G). It follows that in com- dR(A/S) → H 1 (2.2) (cid:44)→ P0 ↓ KS L∨ ⊗ Ω1 D(G)(Σ) ↓ ∇ S ←− D(G)(Σ) ⊗ Ω1 S deformation over Spf ((cid:98)OS,x), the p-divisible groups F il1A[p∞], and dually G = Finally, we have to use the description of the formal neighborhood of x as given in [Mo]. Since we are considering the pull-back of A to S only, and not the full A[p∞]/F il2, are constant over S. Thus over S G (cid:39) G × (OE ⊗ Qp/Zp), and ∇ maps D(G) to D(G) ⊗ Ω1 S. Since P0 = ωG = D(G)(Σ) as subspaces of H 1 dR(A/S), ∇(P0) ⊂ P0 ⊗ Ω1 S. The bottom arrow in (2.2) comes from the homomorphism D(G)(Σ) (cid:44)→ H 1 dR(A/S)(Σ) pr→ H 1(A,O)(Σ) φ(cid:39) H 1(At,O)(Σ) = L∨. But the projection pr kills P0 ⊂ ωA/S. This concludes the proof. (cid:3) We shall later show that the line sub-bundle T S+ has a third characterization, in connection with the ramification in the covering π : S0(p) → S. The definitions and the discussion of this section have obvious generalizations to higher dimensional unitary Shimura varieties. We intend to address them in a future work. ON THE BAD REDUCTION OF CERTAIN U (2, 1) SHIMURA VARIETIES 20 2.3. The blow up of S at the superspecial points. We denote by S# the surface over k which is obtained by blowing up the superspecial points on S. The fiber of S# → S above a superspecial point x is a projective line which we denote by Ex. It is canonically identified with P(TxS). Since S has a canonical model over κ and the stratum Sssp is defined over κ, S# too has a canonical model over κ. In fact, it is the fine moduli space of a moduli problem (S#) which is unique to characteristic p. For any κ-algebra R, S#(R) classifies isomorphism classes of pairs (A,P0) where • A ∈ S(R) • P0 ⊂ ker(V : ωA/R(Σ) → ω(p) which is annihilated by V . A/R(Σ)) is a line sub-bundle of P = ωA/R(Σ) If no geometric fiber of A/R is superspecial then P0 is unique. At superspecial points, however, V kills P, so the additional data amounts to a choice of a line in the plane P. rational. It follows that P is defined over κ too and we can equip each If N = 1 then S is a stack defined over κ and the superspecial points are κ- Ex (cid:39) P(Px) = P(P ⊗ Lx) (cid:39) P(TxS) with a canonical κ-rational structure. If N > 1 then level structure at N forces superspecial points to be defined over larger finite fields, but since P is independent of this extra level structure, the tangent space and the exceptional divisor Ex still carry a canonical κ structure. In practice we use a coordinate ζ on Ex which is derived from a particular choice of a basis for the Dieudonné module of Ax at x ∈ Sssp. This will be explained in Theorem 4.11 below. 3. Local structure of the three integral models 3.1. Raynaud's classification. Recall that k is our fixed algebraically closed field containing κ. In [Ray] Raynaud classifies the finite flat group schemes of rank p2 over k, which admit an action of κ and satisfy the Raynaud condition discussed in 1.2.1. See also [Bel], III.2.3. They are given in the following table. H κ ⊗ Z/pZ κ ⊗ µp κ⊗αp G[p]Σ G[p]Σ αp2,Σ αp2,Σ α∗ p2,Σ α∗ p2,Σ ∅ ∅ Σ, Σ ∅ Σ, Σ Σ, Σ (a0, b0; a1, b1) Lie(H) Lie(αp(H)) α β 2 0 2 1 1 2 2 1 1 (0, 1; 0, 1) (1, 0; 1, 0) (0, 0; 0, 0) (0, 1; 1, 0) (1, 0; 0, 1) (0, 1; 0, 0) (0, 0; 0, 1) (0, 0; 1, 0) (1, 0; 0, 0) 0 2 2 1 1 1 1 2 2 Σ Σ Σ Σ Σ Σ Σ, Σ Σ, Σ Σ Σ Σ Σ γ 1 1 1, 2 1 - 2 - 2 - strata µ µ ssp gss/ssp - gss - gss - Explanations ON THE BAD REDUCTION OF CERTAIN U (2, 1) SHIMURA VARIETIES 21 • Each group scheme is designated by a vector (a0, b0; a1, b1) with entries from {0, 1} where a0b0 = a1b1 = 0. There are 9 possibilities. As a scheme H = Spec(A) where A = k[X, Y ]/(X p−b0Y, Y p−b1X). The group structure (Hopf algebra structure on A) involves the ai. It is completely determined by the condition that the Cartier dual H D is obtained by interchanging a0 with b0, a1 with b1. The twist κ ⊗σ,κ H of H is obtained by interchanging a0 with a1, and likewise b0 with b1. • The column Lie(H) gives the signature of κ on Lie(H), with multiplicities. • The column Lie(αp(H)) gives the signature of κ, with multiplicities, on the Lie algebra of the maximal αp-subgroup of H (whose dimension is Oort's a-number). • The invariants α, β are defined by They satisfy α = dimk Lie(H), β = dimk Lie(H D). α = 2 − b0 − b1, β = 2 − a0 − a1. The third invariant, γ, is not an intrinsic invariant of H, but rather of the way it sits as an isotropic subgroup of A[p]. Recall that if (A, H) is a point of S0(p)(k), we have a filtration 0 ⊂ H ⊂ H⊥ ⊂ A[p] with graded pieces A[p]/H⊥ (cid:39) H D and H⊥/H (cid:39) ker ψ (see §1.2.5). We then set γ = dimk Lie(H⊥/H). • Finally, the last column indicates over which of the strata of S such points (A, H) lie. A hyphen indicates that an H of the given type does not occur as an isotropic subgroup of A[p] for A as in (S). This is the contents of the next lemma. do not occur as isotropic sub- Lemma 3.1. The subgroups G[p]Σ, αp2,Σ and α∗ groups of A[p] for any A as in (S). Proof. We do the first example first. Let M = M (A[p]) be the covariant Dieudonné module of A[p]. It is a 6-dimensional vector space over k, with a κ action of signature (3, 3), and maps F : M (p) → M and V : M → M (p). The principal polarization φ induces a non-degenerate alternating bilinear pairing (cid:104),(cid:105) = (cid:104),(cid:105)M : M × M → k satisfying, for a ∈ κ, x ∈ M (p), y, u, v ∈ M p2,Σ (cid:104)ι(a)u, v(cid:105) = (cid:104)u, ι(a)v(cid:105) (cid:104)F x, y(cid:105)M = (cid:104)x, V y(cid:105)M (p) . By (cid:104),(cid:105)M (p) we denote the base change of (cid:104),(cid:105)M to M (p) = k ⊗σ,k M. The first property shows that M0 and M1, the Σ- and Σ-eigenspaces of κ, are maximal isotropic spaces for the pairing. The second property shows that Lie(A) = Lie(A[p]) = M [V ] = F (M (p)) is another maximal isotropic subspace, which, according to our assumption on the signature of A, intersects M0 in a 2-dimensional space, and M1 in a line. Now let N = M (H) ⊂ M where H is assumed to be of type G[p]Σ and isotropic. Decompose N = N0 ⊕ N1 according to κ-type. Then Lie(H) = N1 is orthogonal to N0 (because N is isotropic) but also to Lie(A)0 = M0[V ] (because Lie(H) ⊂ ON THE BAD REDUCTION OF CERTAIN U (2, 1) SHIMURA VARIETIES 22 Lie(A) and Lie(A) is isotropic). Since N0 is a line lying outside the two-dimensional Lie(A)0, we deduce that N1 is orthogonal to all of M0, contradicting the non- degeneracy of the pairing. The argument for H (cid:39) αp2,Σ is the same. To rule out H (cid:39) α∗ we need another argument, on the αp-subgroup. Lie(H) alone does not distinguish it from α∗ p2,Σ, which, as we shall see later, does occur as a possible isotropic subgroup. If A is either µ-ordinary or general supersingular, then the αp-subgroup of A is of rank p and type Σ, while the αp-subgroup of α∗ is of rank p and type Σ. Hence, α∗ is not isomorphic to a subgroup scheme of A[p]. If A is superspecial, then its p-divisible group is G3, and does not admit a subgroup scheme of type α∗ p2 at all, because the kernels of Verschiebung and Frobenius on A(p) coincide, while α∗ p2 is (cid:3) killed by Frobenius but not by Verschiebung. p2,Σ p2,Σ p2,Σ 3.2. The completed local rings. 3.2.1. Generalities on local models. The method of "local models" was introduced by de Jong [dJ2] and Deligne and Pappas [De-Pa], and developed further by Rapoport and Zink in [Ra-Zi]. See also [P-R-S] and [C-N]. For a point x in the special fiber of a given Shimura variety these authors construct a generalized flag variety, and a point x(cid:48) on it, so that suitable étale neighborhoods of x and x(cid:48) become isomorphic. This allows them to compute the isomorphism type of the completed local rings of the original Shimura variety in terms of linear-algebra data. For the arithmetic schemes S , S0(p) and (cid:102)S these computations were done in [Bel] III.4.3, and in this section we shall quote results from there, adhering as much as possible to the notation used by Bellaïche. The method of local models is flawed when it comes to functoriality with respect to change of level at p. This is because Grothendieck's theory of the Dieudonné crystal, on which it is based, is functorial in divided power neighborhoods, but not beyond. This flaw appears already in the case of the modular curve X0(p) mapping to the j-line X. At a supersingular point y ∈ X0(p)(k) mapping to x ∈ X(k) we get, for the relation between local models in characteristic p while the correct model for the pair (cid:98)Ox (cid:44)→ (cid:98)Oy is known to be, ever since Kronecker, k[[u]] (cid:44)→ k[[u, v]]/(uv), k[[u]] (cid:44)→ k[[u, v]]/((up − v)(vp − u)). Observe that modulo pth powers of the maximal ideal (where there is a canonical divided power structure) the two models are isomorphic, but over the whole formal neighborhood they are not. The second homomorphism is finite flat of degree p + 1 while the first is neither finite nor flat. Despite this flaw, relations between local models of Shimura varieties of PEL type with parahoric level structure suffice to tell us the relations between cotangent spaces, as well as the relations between the infinitesimal deformation theories when we vary the level. 3.2.2. The standard model. Fix y = [A, H] ∈ S0(p)(k). Let x = π(y) ∈ S (k) and (cid:101)x = (cid:101)π(y) ∈ (cid:102)S (k). Then x is represented by the tuple A = (A, φ, ι, η) and (cid:101)x by ON THE BAD REDUCTION OF CERTAIN U (2, 1) SHIMURA VARIETIES 23 A(cid:48) = (A(cid:48), ψ, ι(cid:48), η(cid:48)) where A(cid:48) = A/H and ψ is descended from pφ, i.e. if h : A → A(cid:48) is the canonical isogeny with ker(h) = H then Similarly pφ = ht ◦ ψ ◦ h. ι(cid:48)(a) ◦ h = h ◦ ι(a), η(cid:48) = h ◦ η. Associated with the data (A, φ, ι, A(cid:48), ψ, ι(cid:48), h) is the following linear-algebra data. Let M1 = D(A)W (k), M2 = D(A(cid:48))W (k) be the crystalline Dieudonné modules of the two abelian varieties. Here D(A) is the (contravariant) Dieudonné crystal associated to A, cf. [Gro]. In this section we use crystalline deformation theory as in [Bel]. The translation to covariant Cartier- Dieudonné theory, which will be employed in later sections, is standard (if painful), see the appendix to [C-C-O]. The modules Mi are free W (k)-modules of rank 6, and decompose under the action of OE as a direct sum of two rank-3 submodules, denoted Mi(Σ) and Mi(Σ). The isogeny h induces an injective homomorphism D(h) : M2 → M1 respecting the OE-action, whose cokernel is a two-dimensional vector space over k of type (1, 1), as H is Raynaud. The polarizations result in type-reversing homo- morphism 2 → M2 where we have used the canonical identifications of M∗ crystalline Dieudonné modules of the dual abelian varieties. Clearly 1 (cid:39) M1, B(cid:48) : M∗ B : M∗ i = Hom(Mi, W (k)) with the D(h) ◦ B(cid:48) ◦ D(h)∗ = pB. Denote by M1 the coherent sheaf on S which associates to a Zariski open U the module (A being the universal abelian variety over S ) and define M2 similarly on (cid:102)S . M1(U ) = D(A)U Denote by the same letters their pull-backs to S0(p). Then the same sort of linear- algebra structure is induced on the sheaves Mi, the map D(h) resulting from the canonical isogeny h : A → A/H where H is the universal subgroup scheme of A over S0(p). The following is Théorème III.4.2.5.3 of [Bel]. Theorem 3.2. (i) There exist W (k)-bases {e1, . . . , e6} of M1 and {f1, . . . , f6} of M2 such that, if we denote by {e∗ i } the dual bases, the following properties hold. (a) M1(Σ) is spanned by {e1, e2, e3}, M1(Σ) is spanned by {e4, e5, e6}, and sim- i } and {f∗ ilarly for M2. ON THE BAD REDUCTION OF CERTAIN U (2, 1) SHIMURA VARIETIES 24 (b) The matrices of the homomorphisms B, B(cid:48) in these bases are given by  B = 1 1 1−1 −1 −1  , B(cid:48) =   , p 1 1−1 −1 −p i.e. B(e∗ 1) = −e6, B(cid:48)(f∗ 1 ) = −pf6 etc. (c) The matrix of D(h) is given by  D(h) =  , 1 p 1 p 1 1 i.e. D(h)(f1) = e1, D(h)(f2) = pe2 etc. (ii) The structure (M1,M2, B, B(cid:48), D(h)) is locally Zariski isomorphic to (M1, M2, B, B(cid:48), D(h)) ⊗W (k) O. 3.2.3. The Hodge filtration. Fix y = [A, H] ∈ S0(p)(k) as above. The canonical isomorphism M1 ⊗W (k) k = D(A)k (cid:39) H 1 dR(A/k) defines a 3-dimensional subspace ω0 ⊂ M1 ⊗W (k) k 0 ⊂ which maps isomorphically to ωA/k, and similarly a 3-dimensional subspace ω(cid:48) M2 ⊗W (k) k which maps to ωA(cid:48)/k. These subspaces are OE-invariant of type (2, 1). Furthermore, they are isotropic in the sense that if we denote by ω⊥ 0 the annihilator of ω0 in M∗ 1 ⊗W (k) k, and similarly for ω(cid:48) 0, then 0 ) = ω0, B(cid:48)(ω(cid:48)⊥ B(ω⊥ 0 ) ⊂ ω(cid:48) 0. Equality (rather than inclusion) holds with B because φ, unlike ψ, is principal. Finally, the map D(h) maps ω(cid:48) Lemma 3.3. (i) The invariants (α, β, γ) at the point y are given by the formulae 0 to ω0. β = dimk M1 ⊗W (k) k/(cid:0)ω0 + D(h)(M2 ⊗W (k) k)(cid:1) α = dimk ω0/D(h)(ω(cid:48) 0) γ = dimk ω(cid:48) 0/B(cid:48)(ω(cid:48)⊥ 0 ). (ii) (α, β, γ) form a complete set of invariants of the structure (M1 ⊗W (k) k, M2 ⊗W (k) k, B, B(cid:48), D(h), ω0, ω(cid:48) 0). Namely, any two structures (over k) of this form having the same set of invariants (α, β, γ) are isomorphic. ON THE BAD REDUCTION OF CERTAIN U (2, 1) SHIMURA VARIETIES 25 Proof. Part (ii) is an exercise in linear algebra which we leave out to the reader. In checking it observe that α determines the relative position of ω(cid:48) 0 and ker D(h), β determines the relative position of ω0 and ImD(h), while γ is responsible for the relative position of ker B(cid:48) and ω(cid:48)⊥ 0 . To prove (i) consider the diagram ω∨ HD∩ A(cid:48)t → 0 dR(A(cid:48)/k) → ω∨ ↓ dR(A/k) → ω∨ At → 0 (ht)∗∨ ↓ 0 → ωA(cid:48) → H 1 0 → ωA → H 1 h∗ ↓ ↓ ωH↓ 0 This gives the formulae for α = dimk ωH and β = dimk ωHD = dimk coker(ht)∗∨. The formula for γ comes from the fact that if K = H⊥/H = ker ψ then ωK = (cid:3) ωA(cid:48)/B(cid:48)(ωA(cid:48)t). 3.2.4. Deformations. The following is a consequence of the main theorem of [Gro], characterizing deformations of an abelian variety A (with extra structure) by means of linear-algebra data. See also [dJ2] and [Bel], Proposition III.4.3.6. Let Ck be the category of local Artinian rings (R, mR) of residue field isomorphic to k, equipped with an isomorphism R/mR (cid:39) k. Observe that every object of Ck comes with a canonical homomorphism W (k) → R. The local deformation problem D of the structure (A, φ, ι, A(cid:48), ψ, ι(cid:48), h)/k associates to R ∈ Ck the set D(R) of isomorphism classes of similar structures over R, equipped with an isomorphism between their reduction modulo mR and the given structure over k. theorem is the following. Theorem 3.4. The local deformation problem D is equivalent to the deformation It is represented by the formal scheme Spf((cid:98)OS0(p),y). The local model problem (cid:101)D which associates to every (R, mR) as above the set of structures (ω ⊂ M1 ⊗W (k) R, ω(cid:48) ⊂ M2 ⊗W (k) R) satisfying modulo mR to ω0 and ω(cid:48) 0. (a) ω and ω(cid:48) are rank-3 direct summands, OE-invariant of type (2, 1), reducing (b) B(ω⊥) = ω, B(cid:48)(ω(cid:48)⊥) ⊂ ω(cid:48). (c) D(h)(ω(cid:48)) ⊂ ω. Similar results hold for the moduli problems represented by Spf((cid:98)OS ,x) and Spf((cid:98)O (cid:102)S ,(cid:101)x), obtained by forgetting part of the data. and L(cid:101)x representing the deformation problem (cid:101)D, and deduce isomorphisms Since the local deformation problems (cid:101)D at x and (cid:101)x are obtained from the same (cid:98)OS0(p),y (cid:39) Ly, (cid:98)OS ,x (cid:39) Lx, (cid:98)O (cid:102)S ,(cid:101)x (cid:39) L(cid:101)x. The theorem allows us to compute, quite easily, the complete local rings Ly, Lx problem at y by forgetting part of the data, we get canonical homomorphisms (3.1) L(cid:101)x → Ly ← Lx ON THE BAD REDUCTION OF CERTAIN U (2, 1) SHIMURA VARIETIES 26 between the local models. However, as remarked above, this diagram is not iso- morphic to the corresponding diagram of homomorphisms between the completed local rings of the Picard modular schemes. The best one can get from the general theory is the following. Theorem 3.5. In the above situation the diagrams L(cid:101)x → Ly ← Lx and (cid:98)O (cid:102)S ,(cid:101)x → (cid:98)OS0(p),y ← (cid:98)OS ,x become canonically isomorphic after one divides all the local rings by the pth powers of their maximal ideals. In particular, they induce isomorphic diagrams on cotangent spaces. 3.3. Computations. 3.3.1. Local model diagrams. Let W = W (k) be the ring of Witt vectors of k. The scheme S is smooth over W, so all its completed local rings are isomorphic to Lx = W [[r, s]]. In the following table we catalog the diagrams (3.1) giving the local models at x,(cid:101)x and y, and the maps between them. Proposition 3.6. For a suitable choice of local parameters the local model diagram is given by the following table (where Lx = W [[r, s]]) H at y = [A, H] Ly maps in [Bel] L(cid:101)x W [[r, s]] W [[a, b]] W [[a, b]] W [[a, b]] W [[a, c]] W [[a, c]] W [[r, s, c]]/(cs + p) W [[a, b, c]]/(bc + p) W [[a, b, c]]/(bc + p) W [[a, b, c]]/(bc + p) a (cid:55)→ r, b (cid:55)→ ps r (cid:55)→ pa, s (cid:55)→ pb r (cid:55)→ a, s (cid:55)→ pc a (cid:55)→ cr, b (cid:55)→ s r (cid:55)→ pa, s (cid:55)→ b µ-ord: κ ⊗ µp κ ⊗ Z/pZ gss: G[p] α∗ p2 αp2 ssp: G[p] II.1.c II.3 II.2 I.1.b I.2 II.2 II.1.a I.1.a κ ⊗ αp (generic) W [[a, b, r]]/(ar + p) κ ⊗ αp ( p+1√−1) W [[a, b, r]]/(abr + p) W [[a, b, c]]/(bc + p) W [[a, c]] W [[a, b]] W [[a, c]] r (cid:55)→ a, s (cid:55)→ pc s (cid:55)→ br s (cid:55)→ br, c (cid:55)→ ar Explanations • The first column indicates the stratum to which x belongs and the possible Raynaud types of the subgroup H in the fiber of π above x. The paren- theses distinguishing the two cases where H (cid:39) κ ⊗ αp refer to the value of the coordinate ζ on the projective line Fx ⊂ π−1(x). This line maps isomorphically to Ex ⊂ S# and we endow it with the coordinate ζ as in Section 2.3 and Theorem 4.11 below. The last entry in the table refers to points where ζ p+1 = −1, "generic" refers to all the rest. • The last column refers to the enumeration of the various cases in Bellaïche's thesis [Bel] III.4.3.8 (cas.sous-cas.sous-sous-cas). The table implies that the special fiber S0(p) of S0(p) is equidimensional of dimen- sion 2. As we shall see in Theorems 4.1 and 4.5, it is the union of three smooth surfaces intersecting transversally. These surfaces are the closures of the strata denoted below by Ym, Yet and Ygss. The first two are irreducible, but the third has several connected components. The non-singular points of S0(p), lying on only one of these surfaces, support an H of type κ ⊗ µp, κ ⊗ Z/pZ or G[p]. The points lying p2, αp2 or κ⊗αp (generic). on the intersection of two of them support an H of type α∗ ON THE BAD REDUCTION OF CERTAIN U (2, 1) SHIMURA VARIETIES 27 The remaining points, represented by the last row in the table, are those where all three surfaces meet. The special fiber(cid:101)S of (cid:102)S is the union of two smooth surfaces intersecting transver- sally. One of them, which is the closure of(cid:101)π(Ym) =(cid:101)π(Yet), is irreducible. The other one, which is the closure of (cid:101)π(Ygss), has several connected components. A point (cid:101)x =(cid:101)π(y) lies on the intersection of these two surfaces if and only if y supports an H of type κ ⊗ αp ( p+1 √−1), α∗ p2 or αp2. In the next subsections we work out two sample cases from the table, explaining how one arrives at the given description of the local model diagram. 3.3.2. First example. Assume that x = π(y) is a gss point and y ∈ S0(p)(k) is such that H (cid:39) αp2,Σ (case I.2 in [Bel]). Here the invariants (α, β, γ) = (1, 2, 2). Using Lemma 3.3 one deduces that we may take, without loss of generality, ω0 = (cid:104)e1, e3, e5(cid:105)k , ω(cid:48) 0 = (cid:104)f2, f3, f5(cid:105)k . A little computation yields that the most general deformation satisfying (a) (b) and (c) of Theorem 3.4 is given by ω = (cid:104)e1 − se2, e3 − re2, e5 + re4 + se6(cid:105)R ω(cid:48) = (cid:104)f2 + cf1, f3 + acf1, f5 + af4 + bf6(cid:105)R , where r, s, a, b, c ∈ mR satisfy the relations bc + p = 0, b = s, pa = r. It follows that L(cid:101)x = W (k)[[a, b, c]]/(bc + p) = Ly ⊃ Lx = W (k)[[r, s]]. In the special fiber we get L(cid:101)x ⊗W (k) k = k[[a, b, c]]/(bc) = Ly ⊗W (k) k ← Lx ⊗W (k) k = k[[r, s]] where s (cid:55)→ b and r (cid:55)→ 0. Corollary 3.7. The map (cid:98)O(cid:101)S,(cid:101)x → (cid:98)OS0(p),y is an isomorphism. Identify (cid:98)OS0(p),y with Ly ⊗W (k) k. There are two analytic branches of S0(p) through y, given by c = 0 and b = 0, namely the closed embeddings of formal schemes W = Spf(k[[a, b]]) (cid:44)→ Y = Spf ((cid:98)OS0(p),y) ←(cid:45) Spf(k[[a, c]]) = Z. The map ΩS/kx → ΩW/ky maps ds (cid:55)→ db, dr (cid:55)→ 0. The map ΩS/kx → ΩZ/ky is identically 0. Proof. The map (cid:98)O (cid:102)S ,(cid:101)x → (cid:98)OS0(p),y is an isomorphism even before we reduce these rings modulo p. Indeed, both are 3-dimensional complete regular local rings, and the map between them induces an isomorphism on the cotangent spaces m/m2, hence is an isomorphism. Here we use the fact that the map between cotangent spaces coincides with the corresponding map on the local models, which happens to be an isomorphism. The two branches of Spf((cid:98)OS0(p),y) can be read off the reduction modulo p of the local model Ly. As both branches are smooth over k, and so is the base S at x, the (cid:3) maps on cotangent spaces are easily calculated from the local models. ON THE BAD REDUCTION OF CERTAIN U (2, 1) SHIMURA VARIETIES 28 3.3.3. Second example. For our second example assume that x is an ssp point and y is such that H (cid:39) κ ⊗ αp and ζ p+1 = −1 (case I.1.a in [Bel]). In this case (α, β, γ) = (2, 2, 2) and we may assume that ω0 = (cid:104)e1, e3, e5(cid:105)k , ω(cid:48) 0 = (cid:104)f2, f3, f4(cid:105)k . The most general deformation satisfying (a) (b) and (c) of Theorem 3.4 is given by ω = (cid:104)e1 − re2, e3 − se2, e5 + se4 + re6(cid:105)R ω(cid:48) = (cid:104)f2 + abf1, f3 + bf1, f4 + af5 + cf6(cid:105)R , where r, s, a, b, c ∈ mR satisfy the relations bc + p = 0, s = −rb, c = ra. The local models are therefore L(cid:101)x = W (k)[[a, b, c]]/(bc + p) → Ly = W (k)[[r, a, b]]/(rab + p) ← Lx = W (k)[[r, s]] and the maps between them are given by c (cid:55)→ ra, s (cid:55)→ −rb. Modulo pth powers of the maximal ideals these are also the maps between the completed local rings of the Picard modular surfaces at the corresponding points. 4. The global structure of S0(p) As before, fix an algebraic closure k of κ. In this section we concentrate on the structure of the geometric special fiber S0(p) over k. 4.1. The µ-ordinary strata. 4.1.1. Lots of Frobenii. Let Y = S0(p), and let Y σ = Φ∗ kY This Y σ carries the universal abelian variety A1 = Aσ = Φ∗ be its base change under the Frobenius of k. This is a fine moduli space for tuples (A1, H1) as in the moduli problem (S0(p)) except that the signature of the OE- action on the Lie algebra of A1 is now (1, 2) rather than (2, 1). kA. It should be Y A, which lies over Y . The same remark and notation distinguished from A(p) = Φ∗ applies to the universal subgroup scheme H. The following diagram illustrates the situation. A F rA/Y−→ A(p) −→ (cid:38) ↓ Y (cid:64) F rY /k−→ (cid:38) Aσ ↓ Y σ ↓ Spec(k) −→ −→ (cid:64) (cid:64) A ↓ Y ↓ Φk−→ Spec(k) (cid:38) Spec(Fp) ↓ The three squares are Cartesian. The composition of the arrows in the three top rows are the maps ΦA, ΦY and Φk. ON THE BAD REDUCTION OF CERTAIN U (2, 1) SHIMURA VARIETIES 29 Figure 4.1. The structure of S0(p) Consider now an R-valued point ξ : Spec(R) → Y and let A = ξ∗A be the abelian scheme over Spec(R) represented by ξ (we suppress the role of H and the PEL structure). Consider F rY /k(ξ) = F rY /k ◦ ξ : Spec(R) → Y σ. Then A1 =F rY /k(ξ)∗A1 = ξ∗F r∗ RA = A(p). In the moduli-problem language this means that for (A, H) ∈ Y (R) kA = ξ∗Φ∗ Y /kΦ∗ Y A = Φ∗ F rY /k((A, H)) = (A(p), H (p)). The Frobenius F rA/R is an isogeny F rA/R : A → A(p). All of the above holds (forgetting the group H) also for S instead of S0(p). 4.1.2. The µ-ordinary strata. We study the part of S0(p) lying over Sµ, together with the map π. Recall that we work over the algebraically closed field k. We are motivated by the familiar diagram of maps of modular curves (which takes advantage of the fact that X0(p) is defined over Fp) F rX/k→ X0(p)et ρ (cid:37) (cid:111) ↓ π X0(1) = π ↓ X0(1) X0(p)et where π(A, H) = A, π(A1, H1) = A1/H1 and ρ(A) = (A(p), A(p)[Ver]). Theorem 4.1. (i) Let Yµ = π−1(Sµ) ⊂ S0(p). Then Yµ is the disjoint union of two open sets Ym and Yet. A point (A, H) ∈ S0(p)(k) lies on Ym if and only if H (cid:39) κ ⊗ µp, and on Yet if and only if H (cid:39) κ ⊗ Z/pZ. (ii) The map π : Yµ → Sµ is finite flat of degree p3 + 1. Restricted to Ym it yields an isomorphism πm : Ym (cid:39) Sµ. 𝑆"(𝑝)&'( 𝑆"(𝑝)&) 𝕲[p] 𝛼01 ζ 𝕲[p] 𝛼01∗ 𝜅Ä𝛼0 𝑆"(𝑝)566 ON THE BAD REDUCTION OF CERTAIN U (2, 1) SHIMURA VARIETIES 30 Its inverse is the section σm : Sµ → Ym, σm(A) = (A, A[p]m), cf. the proof below for the notation. (A1, H1) ∈Y σ that A1 (cid:39) A(p) = Φ∗ RA. In fact, let (iii) Consider next Yet and its base change Y σ et under the Frobenius of k. Let et(R) for some k-algebra R. Then there exists a point A ∈ Sµ(R) such K1 = H1 + H⊥ 1 [Fr], where H⊥ 1 is the annihilator of H1 under the pairing epφ1 on A1[p]. Then K1 is a finite flat, maximal isotropic, OE-stable subgroup scheme of A1[p]. Let B = A1/K1, and descend the polarization, endomorphisms, and level-N structure from A1 to B. Then so we may take A = (cid:104)p(cid:105)−1 B. Moreover, under the isomorphism A1 (cid:39) A(p) B(p) (cid:39) (cid:104)p(cid:105) A1 K1 (cid:39) A(p)[Ver]. (iv) Restricted to Yet, π yields a map πet, which is of degree p3 and totally ramified, i.e. 1 − 1 on k-points. It factors as πet = πet ◦ F rY /k where F rY /k : Yet → Y σ totally ramified of degree p. In fact, identify Y σ et is the relative Frobenius morphism, and πet : Y σ et → Sµ is et with the moduli space for tuples (A1, H1) as before. Let K1 and A be as in part (iii). Then the following holds: (4.1) In addition, if (A1, H1) = F rY /k((A, H)) = (A(p), H (p)) for some (A, H) ∈ Yµ(R), then K1 = A(p)[Ver]. πet((A1, H1)) = (cid:104)p(cid:105)−1 (A1/K1) = A. (v) For any R-valued point A of Sµ, H = Fr(A(p)[Ver]) is a finite flat, rank p2, isotropic, Raynaud subgroup scheme of A(p2)[p]. Furthermore, it is étale. Define a map ρet : Sµ → Y σ2 et = Yet by setting Then ρet is finite flat and totally ramified of degree p. We have ρet(A) = (A(p2), Fr(A(p)[Ver])). ρet ◦ πet = F r2 Y /k : Yet → Y σ2 et = Yet, ρet ◦ πet = F rY σ/k. The following diagram summarizes what was said about the maps πet,πet, ρet. Yet = Yet F rY /k−→ πet (cid:38) Y σ et ↓ πet Sµ F rY σ /k−→ (cid:37) ρet F rS/k−→ Yet ↓ πσ Sσ µ et (cid:38) πet F rSσ /k−→ Sµ. ON THE BAD REDUCTION OF CERTAIN U (2, 1) SHIMURA VARIETIES 31 Proof. (i) Let Yµ = π−1(Sµ). This is an open subset of S0(p). If R is any k-algebra and A ∈ Sµ(R), then the group scheme A[p]/R admits a canonical filtration by finite flat OE-subgroup schemes F il3A[p] = 0 ⊂ F il2A[p] = A[p]m ⊂ F il1A[p] = A[p]0 ⊂ F il0A[p] = A[p]. Here F il1 is the maximal connected subgroup-scheme and is of rank p4, while F il2 is the maximal subgroup scheme of multiplicative type (connected, with étale Cartier dual), and is of rank p2. It is also equal to the annihilator of F il1 under the pairing epφ. Moreover, the graded pieces are rigid in formal neighborhoods. This means that over any Artinian neighborhood Spec(R) of a point, we have isomorphisms (gri = F ili/F ili+1) gr2A[p] (cid:39) κ ⊗ µp, gr1A[p] (cid:39) G[p]Σ, gr0A[p] (cid:39) κ ⊗ Z/pZ, as R-group schemes with OE-action. We remark that the filtration and the rigidity of its graded pieces hold for the whole p-divisible group. If R = k (or any other perfect field), A[p] splits canonically as the product of the three graded pieces. As these are pairwise non-isomorphic, the only rank-p2 OE-subgroup schemes of A[p] are then the unique copies of κ ⊗ µp, G[p]Σ or κ ⊗ Z/pZ in it. They are all Raynaud. Only the first and the last are isotropic for the Weil pairing. Thus, if x ∈ Sµ(k), there are only two points of Yµ(k) above x. We call Ym the component of Yµ containing the k-points (A, H) where H (cid:39) κ ⊗ µp, and Yet the component containing the k-points where H (cid:39) κ ⊗ Z/pZ. That these are indeed connected components follows from the above mentioned rigidity. (ii) Let σm : Sµ → Ym be the morphism defined on R-points (R any k-algebra) by A (cid:55)→ (A, A[p]m). It is a section of the map π, both π ◦ σm and σm ◦ π are the identity maps, hence π induces an isomorphism on Ym. This is not the case on Yet, as we can not split the filtration of A[p] functorially over arbitrary k-algebra, only over perfect fields. Let us prove that πet : Yet → Sµ is finite flat and totally ramified of degree p3. It follows from the computations of the completed local rings in §3.2 that Yet is non-singular. The map πet is quasi- finite and proper (see Proposition 1.5), hence finite. Any finite surjective morphism between non-singular varieties is automatically flat ([Eis] 18.17). In fact, the same argument, using regularity of the arithmetic schemes, proves that on the scheme S0(p)(cid:48) obtained by removing Yss = π−1(Sss) from the special fiber of S0(p), the map π is finite flat to S (cid:48) = S − Sss. Since the degree in the generic fiber is p3 + 1, so must be the degree in the special fiber. Since π was shown to be an isomorphism on Ym, on Yet it is finite flat of degree p3, and of course, totally ramified (1 − 1 on geometric points). For another proof see [Bel] III.3.5.12. et is a base-change of an R(cid:48)-point under a homomorphism R(cid:48) → R, where R(cid:48) is reduced. We may therefore assume in the proof of (iii) and (iv) that R is reduced. et is reduced, every R-point of Y σ (iii,iv) Since Y σ We begin by showing that if (A1, H1) is an R-point of Y σ 1 [Fr] is a finite flat subgroup-scheme of rank p3 contained in A1[p]. It is enough to prove this for the universal abelian scheme A1 over Y σ et, and its universal subgroup H1. We use the criterion for flatness, saying that if f : X(cid:48) → X is a finite morphism of schemes, X is reduced, and all the fibers of f have the same rank, then f is also flat ([Mu], p.432). By the open-ness of the flat locus of a morphism, if X is a et, then K1 = H1+H⊥ ON THE BAD REDUCTION OF CERTAIN U (2, 1) SHIMURA VARIETIES 32 variety over a field k, it is enough to check the constancy of the fiber rank at closed points of X. We shall use this criterion here for group schemes over Y σ et, noting that the base is a non-singular variety. First, H⊥ 1 is clearly finite flat of rank p4 1 ∩ A1[Fr] is a closed, hence finite, subgroup scheme. Its over Y σ fiber rank (over the closed points of Y σ et!) is constantly p, so it is also flat. Next, H1 ∩ H⊥ et and H⊥ 1 [Fr] = H1[Fr] = 0. Thus, as a subgroup functor of A1[p], 1 [Fr] = H⊥ H1 + H⊥ 1 [Fr] (cid:39) (H1 × H⊥ 1 [Fr])/(H1 ∩ H⊥ 1 [Fr]) (cid:39) H1 × H⊥ 1 [Fr] is a finite flat group scheme of rank p3. Define πet to be the morphism sending (A1, H1) ∈ Y σ et(R) to (cid:104)p(cid:105)−1 B where B = A1/K1. The type of Lie(B) will now be (2, 1), as can be easily checked. Since K1 is a maximal isotropic subgroup scheme for the Weil pairing on A1[p], the polarization pφ1 on A1 descends to a principal polarization of B. The tame level-N structure on A1 gives rise to a tame level-N structure on B. This completes the definition of πet. If (A1, H1) = (A(p), H (p)) for (A, H) ∈ Yet(R), and R is reduced, then K1 is of rank p3 and killed by Ver, as can be checked fiber-by-fiber. This shows that hence A1/K1 (cid:39) A via Ver : A(p) → A. The polarization pφ1 descends back to φ because φ1 = φ(p). Finally, if K1 = A(p)[Ver], η : Λ/N Λ (cid:39) A[N ] is the level-N structure on A and η1 = η(p), then Ver ◦ η(p) = (cid:104)p(cid:105) ◦ η, concluding the proof that πet(A1, H1) = A. This holds in particular when R = k, which is enough to prove πet = πet ◦ F rY /k. We remark that for a reduced R, to conclude that K1 = A(p)[Ver] we did not have to know that H1 was of the form H (p), only that A1 = A(p). Caution must be exercised when R is non-reduced though, because it is then possible to have A(p) (cid:39) B(p) without A (cid:39) B. The isogeny Ver should be labeled by A or B, and the given isomorphism between A(p) and B(p) may not carry ker(VerA) to ker(VerB). In general, applying the same argument to (A(p) [Fr] = A(p) 1 , H (p) 1 [Ver] 1 ) implies that so B(p) = A(p) 1 /(H (p) [Fr]) = A(p) 1 /A(p) 1 [Ver] (cid:39) A1. By the remark above, K1 = B(p)[Ver]. We emphasize, however, that the group H1 need not be a Frobenius base change of a similar subgroup of B. To guarantee that the level-N structures also match we have to twist B by the diamond operator (cid:104)p(cid:105)−1 and set A = (cid:104)p(cid:105)−1 B. Then A1 (cid:39) A(p). (v) The finite subgroup scheme A(p)[Fr] ∩ A(p)[Ver] is flat over Sµ, as it has constant fiber rank p and the base is reduced. The image Fr(A(p)[Ver]) ⊂ A(p2)[p], is isomorphic to the quotient of A(p)[Ver] by A(p)[Fr]∩A(p)[Ver], hence is also finite and flat of rank p2. It is isotropic, OE-stable and Raynaud. By base change from 1 1 + H (p)⊥ H (p) 1 + H (p)⊥ 1 ON THE BAD REDUCTION OF CERTAIN U (2, 1) SHIMURA VARIETIES 33 the universal case, for any R-valued point A of Sµ, H = Fr(A(p)[Ver]) is a finite flat, rank p2, isotropic, Raynaud subgroup scheme of A(p2)[p]. It is easily seen to be étale. Since ρet is defined functorially in terms of the moduli problem, it is a well-defined morphism. Y /k on k-valued points (A, H) ∈ It is enough to verify the equality ρet ◦ πet = F r2 Yet(k), namely that Fr(A(p)[Ver]) = H (p2), but if A is µ-ordinary this is clear. The relation ρet ◦ πet = F rY σ/k follows from ρet◦πet = F r2 Y /k since πet = πet◦F rY /k and F rY /k is faithfully flat. The remaining (cid:3) assertions on ρet also follow from this relation. Corollary 4.2. Over Yet the universal abelian scheme A (cid:39) A(p) 1 = Yet ×ΦY ,Yet A1 for another abelian scheme A1 of type (1, 2). Proof. In part (iii) of the theorem we showed the same for the universal abelian variety A1 over Y σ et. The corollary follows by base-changing back to Yet, or by (cid:3) repeating the arguments throughout with type (1, 2) replacing type (2, 1). 4.1.3. A lemma on ramification. Before we continue our study of Yµ we need the following result. Lemma 4.3. Let π : Y → X be a finite flat totally ramified morphism of degree p between non-singular surfaces over k, an algebraically closed field of characteristic p. Let π(y) = x. Then there exist local parameters u, v at y ∈ Y so that π∗ : (cid:98)OX,x (cid:44)→ (cid:98)OY,y is The class of up modulo (cid:98)m2 k[[up, v]] (cid:44)→ k[[u, v]]. X,x spans ker(π∗ : ΩX/kx → ΩY /ky), and is therefore independent of any choice. Proof. See [Ru-Sh] Theorem 4, and the Corollary at the bottom of p. 1215 there. (cid:3) Definition. We call the line in TxX which is the annihilator of ker(π∗ : ΩX/kx → ΩY /ky) the unramified direction at x, and denote it by TxX ur. Then T X ur is a line sub-bundle of T X. If C ⊂ X is a non-singular curve such that for every x ∈ C TxC = TxX ur ⊂ TxX (an integral curve for T X ur), then π : π−1(C)red → C is indeed unramified, hence an isomorphism, because π∗ is injective on ΩC/k = ΩX/k/T C⊥ = ΩX/k/ ker(π∗). 4.1.4. The unramified direction of πet. The morphism πet is "too ramified", and we study it via the factorization πet = πet ◦ F rY /k. Since πet is of degree p, it admits, as we have just seen, an "unramified direction". In §2.2 we have defined the special sub-bundle T S+ in T S outside the superspecial locus. We shall now show that over Sµ it coincides with the sub-bundle of unramified directions for πet. Thus the latter can be defined intrinsically in terms of the automorphic vector bundles on S, without any reference to the covering π. ON THE BAD REDUCTION OF CERTAIN U (2, 1) SHIMURA VARIETIES 34 Theorem 4.4. Let x = πet(y) = πet(y(p)) ∈ Sµ. The unramified direction at x for the map πet is TxS+. Equivalently, under the Kodaira-Spencer isomorphism ker(ΩSµ/k → ΩY σ et/k) = KS(P0 ⊗ L). Proof. More precisely, we need to prove that over Y σ et ker(π∗ etΩSµ/k → ΩY σ et/k) = π∗ et(KS(P0 ⊗ L)). In parts (iii) and (iv) of Theorem 4.1 we have seen that if we denote by A1 the etA, and the universal abelian scheme over Y σ morphism πet is induced from Ver : A1 → B, followed by (cid:104)p(cid:105)−1 on the level-N structure. et) where H1 is the universal étale subgroup scheme of A1. The isogeny Ver : A1 → B factors as Consider the abelian scheme C = A1/H1 (over Y σ et then A1 = B(p), where B = π∗ Ver : A1 ψ→ C ϕ→ B where ψ is the isogeny with kernel H1 and ϕ the isogeny with kernel A1[Ver]/H1. Notice that although Ver : A1 → B is pulled back from a similar isogeny over Sµ, et does it factor through C because H1 is not the pull-back of a group only over Y σ scheme on Sµ. Consider now the diagram etP = ωB(Σ) KSB→ ΩY σ π∗ KSC→ ΩY σ ↓ ϕ∗ ωC(Σ) Bt(Σ) ↓ 1 ⊗ (ϕt)∗∨ Ct(Σ) ⊗ ω∨ ⊗ ω∨ et et resulting from the functoriality of the Kodaira-Spencer maps with regard to the isogeny ϕ. Here KSB is the Kodaira-Spencer map for the family B → Y σ et and likewise for C. Note that as B = π∗ etA, KSB is the composition of the isomorphism et(P) π∗ et(KS) : π∗ et(ΩSµ) ⊗ π∗ et(L)∨ ∼→ π∗ (we identify L = ωA(Σ) with ωAt(Σ) via the polarization as usual) and the map induced by π∗ et : π∗ et(ΩSµ ) → ΩY σ et . The kernel of the left vertical arrow ϕ∗ is precisely π∗ et(P0). On the right hand side, however, 1 ⊗ (ϕt)∗∨ is injective. This stems from the fact that the type of C (an étale quotient of A1) is (1, 2) while the type of B is (2, 1). Thus the type of Ct is (2, 1) and that of Bt (1,2). The map (ϕt)∗ being surjective on the Σ-part of the cotangent spaces, its dual is injective. We conclude that KSB(π∗ et(P0)) = 0, hence et(KS(P0 ⊗ L)) ⊂ ker(π∗ π∗ et : π∗ etΩSµ → ΩY σ et ). As both sides are line bundles which are direct summands of the locally free rank 2 sheaf π∗ etΩSµ, the inclusion is an equality between line sub-bundles, as desired. Their annihilators in T S are the "special sub-bundle" T S+ and the "line-bundle of (cid:3) unramified directions" T Sur, hence these two are also equal. In the next section we shall see that the theorem extends to the gss locus. In fact, the same proof applies, once we extend the morphism πet and the factorization πet = πet ◦ F rY /k. See the proof of Theorem 4.5 (iii). ON THE BAD REDUCTION OF CERTAIN U (2, 1) SHIMURA VARIETIES 35 4.2. The gss strata. Recall that the supersingular locus Sss ⊂ S is the union of Fermat curves crossing transversally at the superspecial locus Sssp. The complement of these crossing points was denoted Sgss and is therefore a disjoint union of open Fermat curves. In this section we study its pre-image under the morphism S0(p) → S and show that it is a P1-bundle, intersecting transversally with the horizontal components of S0(p). Understanding the pre-image of Sssp will be taken up in the next section. 4.2.1. The P1-bundles. Theorem 4.5. (i) Let Ygss = π−1(Sgss)red. Then Ygss has the structure of a P1- bundle over the non-singular curve Sgss, with two distinguished non-intersecting non-singular curves Zet and Zm. (ii) The closure Y m of Ym intersects Ygss transversally in Zm. Let Y † A point y = (A, H) ∈ S0(p)(k) lies on Zet if and only if H (cid:39) αp2,Σ and on Zm if and only if H (cid:39) α∗ p2,Σ. The fiber π−1(x) (x ∈ Sgss(k)) intersects each of the curves Zet or Zm at a unique point. At all other k-points (A, H) of Ygss, the group H (cid:39) G[p]Σ. m = Ym∪Zm, a locally closed subscheme of S0(p), and S† m is a non-singular µ is an isomorphism, and the section σm : Sµ → Ym surface. The map πm : Y † extends to a section of πm over S† µ. † et = Yet ∪ † et is a non-singular surface. The (iii) The closure Y et of Yet intersects Ygss transversally in Zet. Let Y Zet, a locally closed subscheme of S0(p). Then Y morphism πet of Theorem 4.1 extends to a morphism µ = Sµ∪ Sgss. Then Y † ∼→ S† m πet : Y † et σ → S† µ, which is finite flat totally ramified of degree p. The factorization πet = πet ◦ F rY /k extends to Y Restricted to Zet the map πet is totally ramified of degree p and πZ = πetZσ † et. is et an isomorphism from Z σ et onto Sgss. (iv) Setting ρet(A) = (A(p2), Fr(A(p)[Ver])) extends the map ρet to a finite flat totally ramified map of degree p from S† We have µ to Y † et. ρet ◦ πet = F r2 Y /k : Y † et → Y †σ2 et = Y † et, ρet ◦ πet = F rY σ/k. The proof of the theorem will be given in the next subsection. We caution the reader that the scheme-theoretic pre-image of Sgss under πet is not reduced. It is †σ et . Similarly rather a nilpotent thickening of degree p of the reduced curve Z σ the scheme-theoretic pre-image π−1(Sgss) is non-reduced along Zet, and only there. et is no longer valid for its continuous extension to Z σ 1 [Fr] is represented by a finite flat group scheme on each of Y σ et separately, but even though the ranks of these group schemes are the same (p3), they do not glue to give a group We also caution that the formula (4.1) giving πet on Y σ et. The group functor H1 + H⊥ et and Z σ et in Y ON THE BAD REDUCTION OF CERTAIN U (2, 1) SHIMURA VARIETIES 36 scheme over the whole of Y kernel of Ver, but this does not hold at closed points of Z σ †σ et . Indeed, at a closed point of Y σ et.6 et this group is the The following diagram summarizes what the extensions of the maps πet, πet, ρet to the gss strata look like. Zet F rZ/k−→ Z σ πet (cid:38) p (cid:39)↓ πet et Sgss F rZσ /k−→ Z σ2 p (cid:37) ρet et = Zet Corollary 4.6. (i) The maps πet and σm induce an isomorphism σm ◦ πet : Z σ et ∼→ Zm. † m → Y (ii) Setting θ = ρet◦ πm : Y † et gives a commutative diagram of totally ramified finite flat morphisms between surfaces, and similarly between embedded curves (the diagonal arrows are embeddings): Zm (cid:39) ↓ Sgss (cid:38) ··· (cid:38) θ−→ m Zet ... (cid:38) θ−→ Y † ↓ (cid:39) F r2(cid:57)(cid:57)(cid:75) Sgss ↓ S† F r2−→ µ † et Y πet (cid:38) ↓ S† µ If Z(cid:48) m and Z(cid:48) et are two κ-components of Zm and Zet (i.e. defined and irreducible The map θ is of degree p, and so is θZm. In particular, the latter factors through the Frobenius of the curve Zm and yields an isomorphism Z σ m over κ) which map to the same κ-component S(cid:48) Proof. The commutativity is easily checked in terms of the moduli problem. The degrees are calculated from the fact that πm is an isomorphism, πet has degree p3 µ and degree p2 on on Y Sgss. To summarize, in the front square we have p3 × p = p4 × 1, and in the back square we have p × p = p2 × 1. The assertion about κ-components follows from the (cid:3) fact that F r2 ∼→ Zet. gss of Sgss then θ(Z(cid:48) † et and degree p on Zet, while F r2 S/K has degree p4 on S† S/k preserves these components. m) = Z(cid:48) et. Remark. We believe that if N = 1 (working with stacks) the geometrically ir- reducible components of Sgss are already defined over κ, hence θ exchanges the irreducible components of Zm and Zet within the same irreducible component of Ygss. This is clearly not the case when N > 1. Compare with supersingular points on the modular curve X(N ). 6If H1 and H2 are finite flat subgroup schemes of a finite flat group scheme G, then H1 ∩ H2 is a finite subgroup scheme, but is not necessarily flat. If it is flat, then the sum H1 + H2, being isomorphic as a group functor to H1 × H2/(H1 ∩ H2), is again represented by a finite flat group scheme. In general, however, the group-functor-quotient of a finite flat group scheme by a closed (hence finite) non-flat subgroup scheme, need not be represented by a group scheme at all, let alone by a finite flat group scheme. Thus the sum of two subgroup schemes need not be a group scheme! ON THE BAD REDUCTION OF CERTAIN U (2, 1) SHIMURA VARIETIES 37 4.2.2. Proof of Theorem 4.5. We first quote [Bu-We], Proposition 3.6. In the nota- tion used there, the Dieudonné module of A[p], for A supersingular but not super- special, is the "Dieudonné space" B(3). Our Dieudonné module M differs from the one appearing in [Bu-We], (3.2)(2) by a "Frobenius twist". This is because we use covariant Dieudonné theory, while [Bu-We] employs Cartier theory. See [C-C-O], Appendix B.3.10, where the first (used here) is denoted M∗, and the second (used in [Bu-We]) is denoted E∗. Proposition 4.7. Let A ∈ Sgss(k), and let M = M (A[p]) be the covariant Dieudonné module of A[p]. Then M has a basis over k denoted {e1, e2, e3, f1, f2, f3} such that (i) OE acts on the ei via Σ and on the fi via Σ. (ii) The antisymmetric pairing induced by the principal polarization φ is given by (cid:104)ei, fj(cid:105) = (−1)jδij, (cid:104)ei, ej(cid:105) = (cid:104)fi, fj(cid:105) = 0. (iii) F and V are given by the following table: e2 0 0 f3 e2 0 e3 0 f1 e1 F −f3 0 V f2 e1 e3 1 = −f3, V e3 = f (p) f1 0 e2 By this we mean that F e(p) M [V ] = (cid:104)e1, e2, f3(cid:105). 1 , etc. In particular, Lie(A) = Let A ∈ S† µ(R), where R is an arbitrary k-algebra. Lemma 4.8. The R-subgroup scheme αp(A(p)) = A(p)[Fr] ∩ A(p)[Ver] is finite flat of rank p, and OE-stable. Proof. We have already encountered the lemma when A was µ-ordinary. The ex- tension to the gss stratum works the same. It is enough to prove the lemma for the universal abelian scheme A over S† µ. In this case αp(A(p)) is clearly finite and OE-stable, and its fibers all have the same rank p, as follows from Proposition 4.7. Let us make this point clear, because the proposition only deals with fibers over µ (not necessarily closed), and {ξ} its closure closed points. Let ξ be any point of S† (a point, a curve, or an irreducible surface). By the open-ness of the flat locus there is a non-empty connected open subset ξ ∈ U ⊂ {ξ} such that αp(A(p))U is finite and flat over U, hence all its fibers, at all the geometric points of U, have the same rank. But U (k) is Zariski dense in U, and at a k-point the proposition tells us that the rank is p. Hence the rank is p at ξ as well. Since S† µ is reduced, by [Mu], Corollary on p. 432, αp(A(p)) is also flat. (cid:3) Proposition 4.9. The finite flat group scheme A[p]/R has a canonical filtration F il3A[p] = 0 ⊂ F il2A[p] ⊂ F il1A[p] ⊂ F il0A[p] = A[p] by finite flat group schemes, which agrees with the canonical filtration over Sµ. The graded pieces are OE-stable, rank p2 and Raynaud. Furthermore, F il1A[p] = F il2A[p]⊥ (with respect to the Weil pairing). Over Sgss every geometric fiber of p2,Σ, gr1A[p] is of type κ⊗ αp, and gr0A[p] is of type αp2,Σ. Let gr2A[p] is of type α∗ R = k and assume that A ∈ Sgss(k). Then, with the notation of Proposition 4.7, F il2M = (cid:104)e2, f3(cid:105) , F il1M = (cid:104)e1, e2, f1, f3(cid:105) . ON THE BAD REDUCTION OF CERTAIN U (2, 1) SHIMURA VARIETIES 38 We remark that unlike µ-ordinary abelian varieties, the above filtration does not split, even if R = k. As we shall see, A[p] does not admit a subgroup scheme of type κ⊗ αp at all, and while it does admit a unique subgroup scheme of type αp2,Σ, this subgroup scheme is contained in F il1A[p], so does not lift gr0A[p]. Proof. Define F il2A[p] = Ver(A(p)[Fr]) (cid:39) A(p)[Fr]/A(p)[Fr] ∩ A(p)[Ver]. This image exists because it is a quotient by a finite flat subgroup scheme. It is a closed subgroup scheme of A[p]. Since A(p)[Fr] is finite flat of rank p3, the Lemma implies that F il2A[p] is finite flat of rank p2. It is furthermore isotropic for the Weil pairing epφ on A[p] associated with the principal polarization φ. By Cartier duality F il1A[p] = F il2A[p]⊥ is finite flat of rank p4. These group schemes are clearly OE-stable. The remaining assertions concern the geometric fibers of A[p], so we assume that R = k. Over the µ-ordinary locus this is the same filtration that we encountered before. Assume that we are over Sgss, and use Proposition 4.7. Let M = M (A[p]). Since F is induced by Ver and V is induced by Fr, we have to compute F (M (p)[V ]). This turns out to be (cid:104)e2, f3(cid:105) . A simple check of the table in §3.1 reveals that (cid:3) gr2 =F il2A[p] is of type α∗ p2,Σ. Similar computations apply to gr1 and gr0. We can now complete the proof of Part (i) of Theorem 4.5. From the analysis of the local models it follows that Ygss is a non-singular surface, mapping under the map π to the non-singular curve Sgss. This is clear at points where H (cid:39) G[p]. At a point y ∈ Ygss where H (cid:39) αp2,Σ or H (cid:39) α∗ p2,Σ the formal neighborhood of y in S0(p) has two non-singular analytic branches which intersect transversally. Since there are at least two irreducible components of S0(p) passing through y, the vertical component Ygss and (at least) one horizontal component, we conclude that there are precisely two such components, and that they are non-singular at y. In particular, Ygss is non-singular at y too. it is enough to prove that for any x ∈ Sgss(k), the scheme-theoretic fiber By the Noether-Enriques Theorem ([Bea] Theorem III.4 and Proposition III.7) Yx ⊂ Ygss of the map π : Ygss → Sgss is isomorphic to P1. We rely on the computation of local models at points y ∈ Yx in [Bel] III.4.3.8. These show that for any y ∈ Yx the map π∗ : ΩSgss,x → ΩYgss,y is injective, and π : Ygss → Sgss is smooth at y. We do not reproduce these com- putations here, but remark that the most problematic points turn out to be the y that lie on Zet (where H (cid:39) αp2,Σ). At such points the claim follows from §3.3.2, as the analytic branch of S0(p) at y determined by Ygss is the one denoted there W, while Sgss ⊂ S is given infinitesimally by the equation r = 0. Yx is therefore a reduced non-singular curve. Let M be the covariant Dieudonné module of A[p], where A = Ax, see Proposi- tion 4.7. The fiber Yx represents the relative moduli problem, sending a k-algebra R to the set of finite flat rank p2 isotropic Raynaud OE-subgroup schemes H ⊂ AR[p]. ON THE BAD REDUCTION OF CERTAIN U (2, 1) SHIMURA VARIETIES 39 Note that since AR is a constant abelian scheme over Spec(R) both Fr and Ver are defined on it, base-changing from k to R the corresponding isogenies of A. Let αp(AR) = AR[Fr] ∩ AR[Ver]. This is a constant (finite flat) subgroup scheme of rank p, and if R = k, Dieudonné submodule is (cid:104)e2(cid:105) . Let βp(AR) = AR[Fr2] ∩ AR[Ver2] ∩ AR[p], its another constant (finite flat) subgroup scheme, of rank p3. If R = k, its Dieudonné submodule is (cid:104)e2, f1, f3(cid:105) . We claim that αp(AR) ⊂ H ⊂ βp(AR), hence classifying H/R is the same as classifying finite flat rank p subgroups of βp(AR)/αp(AR). Since Yx is a reduced non-singular curve, it is enough to check these inclusions when R is reduced and of finite type over k. Since the closed points of Spec(R) are then dense, we may assume R = k. But over k, Ver and Fr are nilpotent on H, which is of rank p2, so both Ver2 and Fr2 must kill it. On the other hand, H must contain an αp-subgroup, because it is local with a local Cartier dual. p (of type (Σ, Σ)) and it is well-known that the moduli problem of classifying its rank-p subgroups is represented by P1 /k. One checks that the isotropy and Raynaud conditions are automatically satisfied for such an H. Now βp(AR)/αp(AR) is nothing but α2 Let R = k. The subgroup scheme H is completely determined by its Dieudonné submodule Nλ = (cid:104)e2, λ1f1 + λ3f3(cid:105) where λ = (λ1 : λ3) ∈ P1(k). Here N0 = N(0:1) = M (H) if H = F il2(A[p]) (cid:39) α∗ p2. Similarly, N∞ = N(1:0) = M (H) where H (cid:39) αp2 because N(1:0) is killed by F and V 2 but not by V. For all other values of λ (cid:54)= 0,∞, Nλ = M (H) where H is of type G[p]Σ, because Nλ is killed by F 2 and V 2 but the kernels of F or V are only 1-dimensional. Part (ii): Let us show that the totality of points (A, H) ∈ Ygss(k) where H (cid:39) α∗ p2, makes up a curve Zm, that π induces an isomorphism of this curve onto Sgss, and that the closure of Ym intersects Ygss transversally in this curve. For this purpose, consider the section σm : S† µ → S0(p) mapping an R-valued point A to (A, H), where H = F il2A[p] = Ver(A(p)[Fr]). The image of the section is a surface isomorphic to the base, intersecting Yµ in its connected component Ym and Ygss in the curve Zm. Finally, the transversality of the intersection of the closure of Ym and Ygss follows from the calculation of the completed local ring of S0(p) at a point y ∈ Zm, see §3.2. Part (iii): We turn our attention to the points (A, H) ∈ Ygss(k) where H (cid:39) αp2. The condition Ver(H (p)) = 0 is a closed condition on the moduli problem S0(p). It is satisfied throughout Yet and on Ygss it holds precisely at the given points where H (cid:39) αp2. We claim that this set forms a curve Zet, which is the intersection of the closure of Yet and Ygss. Indeed, π being proper, the closure of Yet must meet every fiber Yx for x ∈ Sgss(k), and such a fiber has a unique point where H (cid:39) αp2. That the intersection is transversal follows as before from §3.2. ON THE BAD REDUCTION OF CERTAIN U (2, 1) SHIMURA VARIETIES 40 † et)p ⊂ k(Y † et). Write Y †σ et . So is Y We claim that since π : Y † † et = Yet ∪ Zet. The computations in §3.2 show that Y et is non-singular. † † et → S factors through F rY /k : Y et → Y †σ et over the dense open set Yet, it factors through F rY /k everywhere. Indeed, consider the local † et is a closed point. Let y(p) = F rY /k(y) ∈ ring OS,x at x = π(y) ∈ S, where y ∈ Y †σ et . For the function fields we have Y †σ k(S) ⊂ k(Y et ) = k(Y et,y. But the ring on the right is just OY † et the ring R is the intersection of all the OY Thus OS,x ⊂ k(Y et ,y(p), because †σ y is the unique point above y(p) in Y et ,y(p) is normal. For every affine †σ subset U = Spec(R) ⊂ Y et,y for closed † †σ points y ∈ U , and similarly for F rY /k(U ) ⊂ Y et . This proves the claim. †σ et to S† µ. It is a finite morphism, because µ is finite. Both source and target are non-singular surfaces, so by πet : Y [Eis] 18.17 it is also flat, totally ramified of degree p. It therefore defines a line sub- bundle T Sur of unramified directions in the tangent bundle there, as in Lemma 4.3, now over all of S† µ. Recall that the special sub-bundle T S+ was defined on the whole of S† µ as well. The two line sub-bundles T S+ and T Sur coincide over Sµ (Theorem 4.4), hence also over Sgss, by continuity. Thus πet extends to a morphism from Y † et)p ∩ OY † † et and OY † et → S† As T S+ is tangent to Sgss along the general supersingular stratum, we get, from et → Sgss is unramified. As it is the discussion following Lemma 4.3, that πZ : Z σ also totally ramified (bijective on k-points), it is an isomorphism. In retrospect, we can look at the factorization πet = πet ◦ F rY /k also from the etA which is the moduli point of view as follows. Consider the abelian scheme B = π∗ †σ pull-back of the universal abelian scheme over S† et . Consider also the universal et A1 (cid:39) B(p), as was abelian scheme A1 over Y et, shown in the proof of Theorem 4.1. It follows that this relation persists over Z σ et (R) to (cid:104)p(cid:105)−1 Ver(B(p)) ∈ †σ and a-fortiori we may define πet by sending (A1, H1) ∈ Y S† µ(R). †σ et . Over the dense open subset Y σ µ to Y Part (iv): By Lemma 4.8, and the arguments used before, Fr(A(p)[Ver]) is a finite flat rank-p2 isotropic Raynaud subgroup scheme of A(p2)[p], for any A ∈ S† µ(R), for any k-algebra R. Since ρet is now defined functorially in terms of the moduli problems, it is a well defined morphism. The argument is identical to the one used for the proof of Part (v) of Theorem 4.1. Since the equality ρet ◦ πet = F r2 Y /k has already been established on Yet = † et. The relation ρet ◦ πet = F rY σ/k follows Yet(k), it extends by continuity to Y from ρet ◦ πet = F r2 Y /k since πet = πet ◦ F rY /k and F rY /k is faithfully flat. The remaining assertions on ρet also follow from this relation. This concludes the proof of Theorem 4.5. 4.2.3. A closer look at Example 3.3.2. It is instructive to look again at the diagram (cid:98)OS,x (cid:44)→ (cid:98)OS0(p),y ON THE BAD REDUCTION OF CERTAIN U (2, 1) SHIMURA VARIETIES at a point y ∈ Zet(k). We have found the local models (cid:98)OS,x (cid:39) k[[r, s]] and (cid:98)OS0(p),y (cid:39) 41 k[[a, b, c]]/(bc). The map between the local models is r (cid:55)→ 0, s (cid:55)→ b. This is far from the correct map between the completed local rings, which should be injective. Let (cid:98)OW and (cid:98)OZ be the quotients of (cid:98)OY = (cid:98)OS0(p),y which were introduced in §3.3.2. The first is obtained by modding out (c), and is the analytic branch determined by the inclusion Ygss ⊂ S0(p). The second is obtained by modding out (b), and is the analytic branch determined by the inclusion Y † et ⊂ S0(p). Claim 4.10. The diagram (cid:98)OS,x → (cid:98)OW is isomorphic to the diagram and the diagram (cid:98)OS,x → (cid:98)OZ is isomorphic to the diagram k[[r, s]] → k[[a, b]], s (cid:55)→ b + ap, r (cid:55)→ 0, k[[r, s]] (cid:44)→ k[[a, c]], r (cid:55)→ cp2 , s (cid:55)→ ap. This is more than could be deduced from the local models alone. Proof. After a change of variable we may assume that r = 0 is the equation of Sgss in a formal neighborhood of x on S. Therefore r maps to 0 in (cid:98)OW. The local parameter s projects (modulo (r)) to a local parameter of the curve Sgss. We already know that it should map to b modulo pth powers. Since b = c = 0 is the formal equation of the curve Zet (the intersection of the two analytic branches) on W, and since the map Zet → Sgss is purely inseparable of degree p, we see that we may choose a so that s mod (b) = ap. A last change of variables allows us to assume that actually s = b + ap. The second diagram is treated similarly. Here the key point is to recall that the map πet from Z to Spf ((cid:98)OS,x) factors through Fr. The resulting map πet on Zσ was Both diagrams are compatible with (cid:98)OS,x → (cid:98)OY = (cid:98)OS0(p),y being given by shown to be of degree p and unramified in the direction of Sgss. (cid:3) r (cid:55)→ cp2 , s (cid:55)→ b + ap. x 4.3. The ssp strata. 4.3.1. The superspecial combs. We now turn our attention to the superspecial strata of S0(p). Let x ∈ Sssp(k) and Yx = π−1(x). We shall contend ourselves with the determination of the reduced scheme Y red , of finite type over k. The scheme the- oretic pre-image of x will not be reduced along the component denoted below Fx, see the discussion following the theorem. Theorem 4.11. (i) Y red is the union of p + 2 projective lines, arranged as follows. One irreducible component, which we call Fx, intersects the remaining p + 1 pro- jective lines transversally, each at a different point ζ ∈ Fx. With a natural choice of a coordinate on Fx, this ζ can be taken to be7 a p + 1 root of −1. These p + 1 projective lines, which we label as Gx[ζ], are disjoint from each other. A point (A, H) ∈ Yx(k) lies on Fx if and only if H (cid:39) κ ⊗ αp. If this is the case, the invariant γ(A, H) = 1 if (A, H) lies on a non-singular point of Yx, and is equal x 7This is a non-trivial statement, as it has consequences for the cross ratio of the intersection points, which is independent of the chosen coordinate on the basis of the comb. ON THE BAD REDUCTION OF CERTAIN U (2, 1) SHIMURA VARIETIES 42 Figure 4.2. The fiber of S0(p) above a superspecial point P1 P1 P1 G[p] . . . G[p] G[p] κ⊗αp • • • P1 ζ if it is the point ζ). to 2 if it lies at the intersection of Fx and some Gx[ζ] (i.e. Finally, if (A, H) lies on Gx[ζ] but not on Fx, the group H (cid:39) G[p]Σ. (ii) The closure Y µ of Yµ = Ym ∪ Yet in S0(p) intersects Y red in Fx. (iii) Let W be the closure of an irreducible component of Ygss. Then W is a If x ∈ Sssp and P1-bundle over an irreducible component C = π(W ) of Sss. Wx = W ∩ Y red then Wx is one of the Gx[ζ]. Precisely one such W passes through Gx[ζ] for a given x and ζ. Thus the closures of the irreducible components of Ygss do not intersect each other. (iv) The closures of the curves W ∩ Zet and W ∩ Zm intersect Gx[ζ] at the point ζ = Gx[ζ] ∩ Fx. x x See Figures 4.1, 4.2. We refer to the irreducible components W of the closure of Ygss as the supersingular (ss) screens. We refer to the Yx for x superspecial as the superspecial (ssp) combs. The component Fx, which we draw horizontally, is called the base of the comb, and the vertical components Gx[ζ] are called its teeth. The points ζ are called the roots of the teeth. Proof. (i) Let A = Ax. We first analyze what happens on the level of Dieudonné modules. Fix a model of GΣ over k, let GΣ = Gσ Σ and fix the polarization so that the resulting pairing on G[p]Σ, (x, y) (cid:55)→ (cid:104)x, λ(y)(cid:105) is alternating. The group scheme A[p] is isomorphic to so that the polarization induced on it by φx is the product λ2 × λσ of the polariza- tions of the three factors. Consequently [Bu-We], the polarized Dieudonné module M =M (A[p]) is given by M = (cid:104)e1, e2, e3, f1, f2, f3(cid:105)k, where the endomorphisms act on the ei via Σ and on the fi via Σ, where (cid:104)ei, fj(cid:105) = δij, (cid:104)ei, ej(cid:105) = (cid:104)fi, fj(cid:105) = 0 and where the action of F and V is given by the table λ : G[p]Σ ∼→ G[p]D Σ = G[p]Σ G[p]2 Σ × G[p]Σ, e1 0 0 e3 e2 f2 0 −f3 e2 f3 −e1 −e2 0 f1 e1 f3 0 0 F V 7 7 3 3 g g ON THE BAD REDUCTION OF CERTAIN U (2, 1) SHIMURA VARIETIES 43 By this we mean F e(p) Let H ⊂ A[p] be as in (S0(p)). Since M (H) is balanced we may write 3 = −f3, V e3 = f (p) etc. 3 M (H) = (cid:104)α1e1 + α2e2 + α3e3, β1f1 + β2f2 + β3f3(cid:105) . The conditions that have to be satisfied are V (M (H)) ⊂ M (H)(p), F (M (H)(p)) ⊂ M (H), and the isotropy condition α1β1 + α2β2 + α3β3 = 0. 1 e1 + βp 2 e2. If α3 (cid:54)= 0 this forces β1 = β2 = 0, and Observe that M (H) contains βp then the isotropy condition gives also β3 = 0, an absurd. Therefore α3 = 0. We distinguish two cases. Case I (the base of the comb): β1 = β2 = 0. This case is characterized by the fact that M (H) is killed by both V and F, so that H (cid:39) κ⊗ αp. We may take β3 = 1 and H is classified by ζ = (α1 : α2) ∈ P1(k). Consider in this case the group H⊥/H. Its Dieudonné module is given by M (H⊥/H) = (cid:104)e1, e2,−α2f1 + α1f2, f3(cid:105) mod (cid:104)α1e1 + α2e2, f3(cid:105) . An easy check shows that H⊥/H is of type G[p]Σ, unless ζ p+1 = −1, where it is of type κ ⊗ αp. The invariant γ(A, H) = dimk Lie(H⊥/H) is thus 1 in the former case, and 2 in the latter. Case II (the teeth of the comb): (β1 : β2) ∈ P1(k). Then, ζ = (α1 : α2) = (βp 1 : 2 ) and the isotropy condition forces βp αp+1 1 + αp+1 i.e. ζ p+1 = −1. Fix ζ, hence the point (β1 : β2). The H in question are classified by β3 ∈ A1(k). Their M (H) is killed by V 2 and F 2 but neither by V nor by F, so H must be isomorphic to G[p]Σ. We observe that when β3 = ∞, i.e. (β1 : β2 : β3) = (0 : 0 : 1) we are back in Case I. This is the root of the tooth. 2 = 0, This analysis strongly suggests the picture outlined in Part (i), but does not quite prove it. To give a rigorous proof we proceed as follows. The fiber Yx represents the relative moduli problem assigning to any k-algebra R the set of subgroup schemes H ⊂ AR[p] of type (S0(p)). Observe that since AR is constant, both Fr and Ver are defined on it, by base change from A. We let αp(AR) = AR[Fr] ∩ AR[Ver] and αp(H) = H ∩ αp(AR). Case I. Consider first the closed locus Fx ⊂ Y red Fr(H) = 0, Ver(H) = 0. x defined by Over Fx we have αp(H) = H. Indeed, since Fx is a reduced curve it is enough to check the inclusion H ⊂ αp(AR) at geometric points, where it follows from the p,Σ × αp,Σ, analysis of their Dieudonné modules as above. However, αp(AR) = α2 so the problem becomes that of classifying OE-subgroup schemes of type κ ⊗ αp = αp,Σ× αp,Σ in it. As the factor of type αp,Σ is unique, this is the same as classifying subgroup schemes of rank p in α2 /k. This gives us the base of the comb, whose k-points are described in terms of their Dieudonné submodules as before. Gx, the group αp(H) is of rank p. Observe that H ∩ G[p]2 Case II. Let Gx be the open curve which is the complement of Fx in Y red p,Σ, a problem that is represented by P1 . Over Σ is non-zero, because x ON THE BAD REDUCTION OF CERTAIN U (2, 1) SHIMURA VARIETIES 44 Σ were classified before by P1 otherwise, via projection to the third factor, H would be of type G[p]Σ, which is forbidden. It follows that αp(H ∩ G[p]2 Σ) is also non-zero, so must coincide with αp(H). The αp ⊂ G[p]2 /k. Our αp(H) is therefore classified by ζ = (α1 : α2) ∈ P1(R). The Dieudonné module computation above shows that ζ restricts, at every geometric point, to a p + 1 root of −1. However, the equation X p+1 + 1 = 0 is separable, so if R is a local ring in characteristic p and ζ ∈ R satisfies this equation modulo mR, it satisfies it in R. This means that αp(H) is locally constant over Spec(R). There remains the classification of H/αp(H), Σ/αp(H)) × G[p]Σ. The same argument which sits in general "diagonally" in (G[p]2 that was used to show that αp(H) is constant, shows now that the projection Σ/αp(H)) is constant, and in fact is given by the point K of H/αp(H) to (G[p]2 2) ∈ P1(R). The classification of H/αp(H) is therefore the (β1 : β2) = (αp same as the classification of all the R-morphisms of this fixed K to αp(G[p]Σ). This moduli problem, of classifying morphisms from a fixed copy of αp to another, is represented by A1 /k.This gives the tooth of the comb labeled Gx[ζ]. 1 : αp The two cases (I) and (II) cover Y red . It remains to remark that the intersection of the closure of Gx[ζ] with Fx is transversal. This follows, as usual, from §3.2. x (ii) The condition Fr(H) = 0 is a closed condition and holds throughout Ym. It therefore holds also in the intersection of its closure Y m with Yx. As this condition is not satisfied on the teeth of the comb (outside their roots), the closure Y m intersects Yx in Fx. The same argument, applied to the condition Ver(H (p)) = 0 proves that the closure Y et of Yet also intersects Yx in Fx. As we have previously † shown that Y † et are disjoint, Y m and Y et intersect only in the superspecial locus, and their intersection is the union of the Fx for x ∈ Sssp. This intersection is transversal, as follows from the description of the completed local rings in §3.2. m and Y (iii) The classification of the completed local rings of S0(p) shows that through a point ζ ∈ Fx which is not a root of a tooth (i.e. ζ p+1 (cid:54)= −1) pass only 2 analytic branches. As Y et and Y m already account for these two analytic branches, the closure W of a connected component of Ygss can only meet Yx in one of the lines Gx[ζ]. Since the points of Gx[ζ] are generically non-singular on S0(p), exactly one such W passes through every Gx[ζ]. These W are non-singular surfaces projecting to a component C of Sss and the fiber above each geometric point (including now the superspecial points) is P1. By the Noether-Enriques theorem quoted before, they are P1-bundles. (iv) The condition Fr(H) = 0 is a closed condition and holds throughout Zm. It therefore holds also on its closure. It follows that this closure intersects a tooth Gx[ζ] at its root, because points other than the root support an H of type G[p]Σ which is not killed by Fr. A similar argument invoking the condition Ver(H (p)) = 0 proves that the closure of Zet also meets the teeth of the combs in their roots. The two curves Zet and Zm, which are disjoint over the gss locus, intersect over every superspecial point. (cid:3) This concludes the proof of the theorem. 4.3.2. The maps to S#. Recall the construction of the blow-up S# of S at the ssp points, given in §2.3. The exceptional divisor Ex at x = [A] ∈ Sssp(k) classifies lines in P = ωA/k(Σ). ON THE BAD REDUCTION OF CERTAIN U (2, 1) SHIMURA VARIETIES 45 The isomorphism πm : Y † m (cid:39) S† µ extends to an isomorphism m : Y m (cid:39) S#. π# In terms of the moduli problems, it sends (A, H) ∈ Y m(R) to (A, ker(ωA/R(Σ) → ωH/R(Σ)). If R = k, A is µ-ordinary and H = A[p]m then ker(P = ωA/k(Σ) → ωH/k(Σ)) = P0 = P[V ] is uniquely determined by A. The same holds if A is gss and H = F il2A[p]. On the other hand if x = [A] is ssp then P[V ] is the whole of P and H "selects" a line in it. This establishes an isomorphism From the universal property of blow-ups, the projection πet : Y et → S also factors through a map Fx (cid:39) Ex. et : Y et → S# π# mapping Fx to Ex. This map is now proper and quasi-finite, hence finite. The two surfaces are non-singular, so the map is also flat. Its degree is p3. We have seen that on the open dense Y † et it factors through F rY /k, i.e. πet = πet ◦ F rY /k and this forces the map π# whole of Y et. The map π# x . Thus π# shown that it is ramified of degree p along the lines F σ degree p2 along Fx (and of an extra degree p in a normal direction). et ◦ F rY /k over the et to factor in the same way π# et is finite flat totally ramified of degree p, and it can be et is ramified of et = π# We emphasize that π# m and π# et do not agree on Fx. Instead, the following diagram extends the one from Corollary 4.6. θ−→∼ Fx (cid:111) (cid:38) π# m ↓ Ex ··· (cid:38) Fx ... (cid:38) θ−→ ↓ Ex F r2 S/k−→ Y et π# (cid:38) ↓ et S# (cid:57)(cid:57)(cid:75) π# m Y m (cid:111) ↓ S# The degrees of the maps in the front square (on surfaces) are p3 × p = p4 × 1. In the back square (on projective lines) they are p2 × 1 = p2 × 1. 4.3.3. How embedded modular curves meet Fx. Let X be the special fiber of the modular curve X which was constructed on S in §1.4. Consider the modular curve X0(p) parametrizing, in addition to the triple B = (B, ν, M ), also a finite flat subgroup scheme HB ⊂ B[p] of rank p. Enhance the map Z0×Isom(Z/NZ,µN )X → S to a map Z0 ×Isom(Z/NZ,µN ) X0(p) → S0(p), (B0, B, HB) (cid:55)→ (A, H) by setting H to be the image of OE ⊗ HB in A(B0, B). Note that since HB is automatically isotropic, and the polarization on A is induced from the polarizations of B and B0, this H is isotropic. It is also clearly Raynaud. ON THE BAD REDUCTION OF CERTAIN U (2, 1) SHIMURA VARIETIES 46 Proposition 4.12. Let X0(p) be the special fiber of X0(p). Let x ∈ Sssp(k). Then under the above morphism X0(p) meets the component Fx ⊂ Yx in a point ζ satis- fying ζ ∈ κ, ζ p+1 (cid:54)= −1. Thus both the supersingular screens on S0(p) and the modular curves cross the superspecial strata Fx at Fp2-rational points, but while the supersingular screens cross at a ζ satisfying ζ p+1 = −1, the modular curves cross at the remaining ones. Proof. As we shall see in the next chapter, the κ-rational ζ ∈ Fx are characterized by the fact that A(cid:48) = A/H is superspecial. At other points of Fx this A(cid:48) is supersingular of a-number 2, but not superspecial. For the pair (A, H) that is constructed from the "elliptic curve data" on X0(p), it is easily seen that A(cid:48) is either µ-ordinary or superspecial, depending on whether B is ordinary or supersingular. Among these κ-rational points the points with ζ p+1 = −1 are characterized by γ(A, H) = 2, i.e. the group H⊥/H = ker(ψ) being isomorphic to κ ⊗ αp. All the rest have γ = 1. In our case, H = OE ⊗ HB is maximal isotropic in A1(B)[p], so its annihilator in A[p] = A1[p] × B0[p] is H × B0[p]. it follows that H⊥/H (cid:39) B0[p] (cid:39) G[p]Σ 5. The structure of (cid:101)S 1.2.3. Typically, moduli spaces involving parahoric level structure are "compli- cated", and may involve issues such as non-reduced components, complicated sin- 5.1. The global structure of (cid:101)S. The moduli space (cid:102)S was defined in Section gularities etc. It is interesting, and important for our further applications, that (cid:102)S 5.1.1. Flatness of(cid:101)π. The following proposition stands in sharp contrast to the non- turns out to be quite simple. In essence, its special fiber is a collection of smooth surfaces intersecting transversally at a reduced non-singular curve. flatness of π. It is also key to understanding the geometry of the surface and γ = 1. (cid:3) T = S0(p) × (cid:102)S S0(p). This surface, which is generically of degree (p + 1)(p3 + 1) over the Picard modular surface S , "is" the geometrization of the Hecke operator Tp. We intend to study it in a future work. Proposition 5.1. The morphism (cid:101)π : S0(p) → (cid:102)S is finite flat of degree p + 1. Proof. Both arithmetic surfaces are regular. The map(cid:101)π is proper, and, as we shall From now on we concentrate on the structure of the geometric special fiber (cid:101)Sk of (cid:102)S over k, and omit the subscript k. We study (cid:101)S together with the map see below, analyzing its geometric fibers one-by-one, also quasi-finite. It is therefore finite. By [Eis], 18.17, it is flat. The degree can be read off in characteristic 0. (cid:3) (cid:101)π : S0(p) → (cid:101)S and make strong use of the facts that we have already established for S0(p). 47 ON THE BAD REDUCTION OF CERTAIN U (2, 1) SHIMURA VARIETIES given A, the subgroup schemes H ⊂ A[p] for which (A, H) ∈ S0(p)(k). This was achieved by analyzing M (A[p]) and its 2-dimensional, isotropic, balanced OE-stable a given A(cid:48), for all the possible (A, H) yielding A(cid:48) upon the process of dividing by H and descending the polarization. Equivalently, by Proposition 1.4, we have to 5.1.2. The fibers of (cid:101)π. To study the geometric fibers of π we had to study, for a Dieudonné submodules. To study the geometric fibers of (cid:101)π we have to look, for look for all the subgroup schemes J such that (A(cid:48), J) ∈ (cid:101)S0(p)(k). This reduces the computation of the fibers of(cid:101)π to Dieudonné-module computations, as was the case with π. However, starting with one (A, H) mapping under (cid:101)π to A(cid:48), finding all the 5.1.3. The stratification of (cid:101)S. We suppress (ι(cid:48), η(cid:48)) from the notation and refer to R-points of (cid:101)S (R a k-algebra) as (A(cid:48), ψ). Given (A(cid:48), ψ) ∈ (cid:101)S(k) the subgroup scheme others in the fiber above A(cid:48) requires in general the knowledge of M (A[p2]) and not only of M (A[p]). This makes the following sections technically more complicated than the previous ones. ker(ψ) ⊂ A(cid:48)[p] isomorphic to its Cartier dual), stable under ι(cid:48)(OE) is of rank p2, self-dual (i.e. and Raynaud. Its Lie algebra Lie(ker(ψ)) is 1 or 2-dimensional8, and carries an action of κ. We call its type the type (or signature) of ker(ψ) and denote it by τ (ψ). Similarly the maximal αp-subgroup of A(cid:48)[p] is of rank p, p2 or p3, and the κ-type of its Lie algebra is called the a-type of A(cid:48), and denoted a(A(cid:48)). Theorem 5.2. (i) The surface (cid:101)S is the union of 7 disjoint, locally closed, nonsin- gular strata (cid:101)S∗[∗∗], as shown in the table. The name of each stratum indicates the type of A(cid:48)(cid:101)x for (cid:101)x in the stratum (µ-ordinary, gss or ssp), and, in brackets, the type of ker(ψ). The last column indicates what types of (A, H) lie in (cid:101)π−1((cid:101)x). The first entry in the last column refers to the stratum of S in which A lies. The second refers to the type of H (G stands for G[p]). If A is ssp there is a third entry, which we now explain. Recall that the ssp strata of S0(p) are unions of projective lines admitting a natural coordinate ζ. The third entry refers to ζ. Depending on whether ζ ∈ Fp2 or not, and in the case of the components Fx, also on whether it is a p + 1 root of −1, (cid:101)π(A, H) may land in different strata of (cid:101)S. τ (ψ) a(A(cid:48)) #(cid:101)π−1((cid:101)x) Σ (cid:101)π−1((cid:101)x) Stratum of (cid:101)x dim. (cid:101)Sµ (cid:101)Sgss[Σ] (cid:101)Sgss[Σ, Σ] (cid:101)Sgss[Σ] (cid:101)Sssp[Σ] (cid:101)Sssp[Σ, Σ] (cid:101)Sssp[Σ] 2 2 1 1 0 0 0 Σ Σ Σ, Σ p + 1 (µ, et/m) 1 2 3 4 5 6 7 (ii) The closure relations between the various strata are described by the following (gss, G)/(ssp,G,¬Fp2) (ssp, κ ⊗ αp,¬Fp2 ) (ssp, G, Fp2 ) √−1) (ssp, κ ⊗ αp, p+1 √−1) (ssp, κ ⊗ αp, Fp2¬ p+1 Σ, Σ, Σ Σ, Σ Σ, Σ, Σ Σ, Σ, Σ (gss, αp2 /α∗ p2) Σ, Σ Σ, Σ Σ, Σ p + 1 2 2 1 1 1 Σ Σ Σ diagram, where an arrow X → Y indicates specialization, i.e. that Y ⊂ X. 8If it were 0-dimensional, A(cid:48) would be µ-ordinary and ker(ψ) (cid:39) κ ⊗ Z/pZ, but this group is not self-dual. (cid:101)Sgss[Σ] (cid:46) (cid:38) (cid:101)Sµ (cid:101)Sssp[Σ, Σ] (cid:101)Sgss[Σ, Σ] (cid:46) (cid:38) (cid:46) (cid:101)Sgss[Σ] (cid:101)Sssp[Σ] ↓ ON THE BAD REDUCTION OF CERTAIN U (2, 1) SHIMURA VARIETIES 48 (cid:46) (cid:101)Sssp[Σ] The strata (cid:101)Sgss[Σ, Σ] and (cid:101)Sssp[Σ, Σ] are singular on (cid:101)S, and the rest are nonsin- gular. See Figure 5.1 Figure 5.1. The structure of (cid:101)S Proof. The invariants (τ (ψ), a(A(cid:48))) characterize the stratum in (cid:101)S, and the seven cases in the last column are mutually exclusive and exhaustive. It is therefore enough to verify that starting with a point (A, H) ∈ S0(p)(k) in a prescribed stratum of S0(p), we end up with the right pair of invariants (τ (ψ), a(A(cid:48))). For this we use the covariant Dieudonné module M (A[p∞]). (1) If A is µ-ordinary, so is A(cid:48), and vice versa. As in this case A[p∞] (cid:39) (OE ⊗ µp∞) ⊕ GΣ ⊕ (OE ⊗ Qp/Zp) and H is either OE ⊗ µp or OE ⊗ Z/pZ, H⊥/H (cid:39) G[p]Σ so τ (ψ) = Σ. Since upon dividing by H we get A(cid:48)[p∞] (cid:39) A[p∞], a(A(cid:48)) = Σ. The map Ym → (cid:101)Sµ is surjective, purely inseparable of degree p, while Yet → (cid:101)Sµ is an isomorphism. This follows from the following two facts: (a) Yµ → (cid:101)Sµ is finite flat of degree p + 1, (b) If y ∈ Yet(k) then(cid:101)π is étale at y, while if y ∈ Ym(k) it is ramified there (see §3.2). We conclude that if (cid:101)x ∈ (cid:101)Sµ(k) the fiber (cid:101)π−1((cid:101)x) contains precisely 2 points. Alternatively, we could have used the model (cid:101)S0(p) (see §1.2.6) to show that there are precisely two possibilities for J to go with an A(cid:48) ∈ (cid:101)Sµ(k). 𝑆"##[Σ] 𝑆##'[Σ] 𝑆##'[Σ,Σ] 𝑆"##[Σ,Σ] 𝑆"##[Σ] 𝑆) 𝑆##'[Σ] ON THE BAD REDUCTION OF CERTAIN U (2, 1) SHIMURA VARIETIES 49 (2) Assume next that A is gss and H (cid:39) G[p]. The analysis of H⊥/H is easy, since H⊥ ⊂ A[p], so we can use Proposition 4.7. With the notation used there for some (α1 : α2) (cid:54)= 0,∞. It follows that M (H⊥/H) = (cid:10)α2e1 − α1e3, f 1 neither by F nor by V, H⊥/H (cid:39) G[p]. Since M (H⊥/H)[V ] =(cid:10)f 1 (cid:11) where (cid:11), Lie(H⊥/H) is the bar denotes the class modulo M (H). Since this space is killed by V 2 and F 2 but M (H) = (cid:104)e2, α1f1 + α2f3(cid:105) of type Σ. To analyze the αp-subgroup of A(cid:48) and conclude that it is of rank p2 and type (Σ, Σ), we need to know M (A[p2]). This, unlike M (A[p]), depends on the particular A, and not only on it being of type gss. The computations needed to verify this are deferred to the appendix. (3) Assume that A is gss and H (cid:39) αp2,Σ. Using the notation of Proposition 4.7 so M (H⊥/H) =(cid:10)e3, f 3 (cid:11) . This module is killed by both V and F so H⊥/H (cid:39) κ⊗αp, M (H) = (cid:104)e2, f1(cid:105)k and its Lie algebra is of type (Σ, Σ). The computation of a(A(cid:48)) is again deferred to the appendix. The case A gss and H (cid:39) α∗ (4) Assume that A is ssp. Then the covariant Dieudonné module M = M (A[p∞]) is freely spanned over W (k) by a basis e1, e2, e3, f1, f2, f3 satisfying (i) OE acts on the ei via Σ and on the fi via Σ (ii) (cid:104)ei, fj(cid:105) = δij, (cid:104)ei, ej(cid:105) = (cid:104)fi, fj(cid:105) = 0 (iii) the action of F and V is given by the table p2,Σ is treated similarly. e1 e2 e3 f3 F −pf1 −pf2 −f3 pe3 f3 −e1 −e2 −pe3 V f1 e1 f2 e2 pf1 pf2 . See [Bu-We], Lemma (4.1) and [Vo], Lemma 4.2. Note that Vollaard works over W (κ) and uses a slightly different normalization, but over W (k) her model and the one above become isomorphic. Let M = M/pM = M (A[p]) (called in [Bu-We] the Dieudonné space) and denote by ei and f i the images of the basis elements. Using the notation of the proof of Theorem 4.11, we distinguish two cases. Case I (the base of the comb): In this case H is of type κ ⊗ αp and M (H) =(cid:10)α1e1 + α2e2, f 3 (cid:11) ⊂ M . As we have seen in the proof of Theorem 4.11, H⊥/H = ker(ψ) is of type G[p]Σ, unless ζ = (α1 : α2) satisfies ζ p+1 = −1, where it is of type κ ⊗ αp. This gives the entries for τ (ψ) in rows 4,6 and 7 of the table. We proceed to compute the a-number and a-type of A(cid:48). For this observe that M(cid:48) = M (A(cid:48)[p∞]) sits in an exact sequence hence inside the isocrystal MQ 0 → M → M(cid:48) → M (H) → 0, M(cid:48) =(cid:10)ei, p−1((cid:101)α1e1 +(cid:101)α2e2), f1, f2, p−1f3 (cid:11) . Here we let (cid:101)αi denote any element of W (k) mapping to αi modulo p. To compute the Dieudonné module of the αp-subgroup of A(cid:48) we must compute (M(cid:48)/pM(cid:48))[V ] ∩ (M(cid:48)/pM(cid:48))[F ]. ON THE BAD REDUCTION OF CERTAIN U (2, 1) SHIMURA VARIETIES 50 1 f1 +(cid:101)ασ The kernel of V on M(cid:48)/pM(cid:48) {e1, e2, e3,(cid:101)ασ F is spanned by the images of {e1, e2, e3,(cid:101)ασ−1 is spanned over k by the images of the vectors 2 f2} where σ is the Frobenius on W (k). Similarly, the kernel of f2}. The span of {e1, e2, e3} in M(cid:48)/pM(cid:48) is two dimensional and of type Σ, Σ. We see that if ζ = (α1 : α2) /∈ Fp2 then αp(A(cid:48)) is of rank p2, hence A(cid:48) is gss (supersingular but not superspecial), and a(A(cid:48)) = {Σ, Σ}. On the other hand if ζ ∈ Fp2 then αp(A(cid:48)) is of rank p3, so A(cid:48) is superspecial, and a(A(cid:48)) = {Σ, Σ, Σ}. This completes the verification of τ (ψ) and a(A(cid:48)) in rows 4,6 and 7 of the table. f1 +(cid:101)ασ−1 1 2 Case II (the teeth of the comb): In this case H is of type G[p]Σ, M (H) =(cid:10)α1e1 + α2e2, β1f 1 + β2f 2 + β3f 3 (cid:11) ⊂ M . 1 2 3 1 2 1 2 1 : βp M(cid:48) = 1 : βp 2 ). f2, f3}. Note that p−1((cid:101)βσ Likewise M(cid:48)/pM(cid:48)[F ] is spanned over k by the images of We find that M(cid:48)/pM(cid:48)[V ] is spanned over k by the images of 2 ) satisfies ζ p+1 = −1 and β3 ∈ k is arbitrary. Now where ζ = (α1 : α2) = (βp M (H⊥/H) is spanned by the images of −β3e1 + β1e3 and f 3 modulo M (H), so H⊥/H = ker(ψ) is seen to be of type G[p]Σ. This confirms the invariant τ (ψ) in rows 2 and 5 of the table. Regarding a(A(cid:48)) we compute, as in Case I, M(cid:48) = M (A(cid:48)[p∞]) : (cid:68) (cid:69) ei, p−1((cid:101)α1e1 +(cid:101)α2e2), fi, p−1((cid:101)β1f1 +(cid:101)β2f2 +(cid:101)β3f3) {e1, e2, p−1((cid:101)βσ 1 e1 +(cid:101)βσ 2 e2) +(cid:101)βσ 3 e3,(cid:101)ασ 1 f1 +(cid:101)ασ 2 f2, f3}. 1 e1 + (cid:101)βσ 2 e2) ∈ M(cid:48) because of the relation (α1 : α2) = (βp {e1, e2, p−1((cid:101)βσ−1 e1 +(cid:101)βσ−1 e2) +(cid:101)βσ−1 e3,(cid:101)ασ−1 f1 +(cid:101)ασ−1 2 f2 both represent the class of (cid:101)β1f1 +(cid:101)β2f2 in 1 f1 +(cid:101)ασ f2 and (cid:101)ασ Now (cid:101)ασ−1 f1 +(cid:101)ασ−1 e1 +(cid:101)βσ−1 M(cid:48)/pM(cid:48). Similarly p−1((cid:101)βσ 1 e1 +(cid:101)βσ 2 e2) and p−1((cid:101)βσ−1 class of p−1((cid:101)α1e1 +(cid:101)α2e2) in M(cid:48)/pM(cid:48). It follows that the span of f3 and (cid:101)ασ (cid:101)ασ e2) + (cid:101)βσ−1 3 e3 and p−1((cid:101)βσ−1 2 e2) + (cid:101)βσ 1 e1 + (cid:101)βσ If β3 ∈ Fp2 then p−1((cid:101)βσ e2) both represent the 1 f1 + 2 f2 in M(cid:48)/pM(cid:48) is 1-dimensional and of type Σ. Regarding the Σ-component of M(cid:48)/pM(cid:48)[V ] ∩ M(cid:48)/pM(cid:48)[F ], e1 and e2 contribute a 1-dimensional piece there. e3 contribute another 1-dimensional piece, but otherwise they do not agree modulo pM(cid:48). To sum up, if β3 /∈ Fp2 then A(cid:48) is gss and a(A(cid:48)) = {Σ, Σ}. If β3 ∈ Fp2 then A(cid:48) is ssp and a(A(cid:48)) = {Σ, Σ, Σ}. This completes the verification of τ (ψ) and a(A(cid:48)) in rows 2 and 5. follow from the known dimensions of the strata of S0(p). Moreover, each geometric fiber has p + 1 points if one counts multiplicities. We have already noted that the Since the morphism(cid:101)π is finite flat of degree p+1, the dimensions of the strata of (cid:101)S map Ym → (cid:101)Sµ is surjective, purely inseparable of degree p, while Yet → (cid:101)Sµ is an isomorphism. This proves that for (cid:101)x ∈ (cid:101)Sµ(k), #(cid:101)π−1((cid:101)x) = 2, but it also proves that for (cid:101)x ∈ (cid:101)Sgss[Σ, Σ](k) we have #(cid:101)π−1((cid:101)x) = 2. Indeed, such a point must have pre-images both in Zet and in Zm but the morphism (cid:101)π : Ym → (cid:101)S being totally on Zm, since the ramification locus is closed. Thus(cid:101)π is 1:1 on Zm(k). It is clearly ramified and 1:1 on geometric points, must extend to a totally ramified morphism e1 + (cid:101)βσ−1 1 2 1 2 3 1:1 on Zet(k) because it is an isomorphism on Zet. ON THE BAD REDUCTION OF CERTAIN U (2, 1) SHIMURA VARIETIES the comb denoted Fx in Theorem 4.11, where A is ssp and H of type κ ⊗ αp. This Similar arguments show that(cid:101)π is totally ramified of degree p + 1 on the base of shows that #(cid:101)π−1((cid:101)x) = 1 in rows 4,6 and 7 of the table. (i.e. where A is ssp or gss but H is of type G[p]Σ) (cid:101)π induces an isomorphism on Finally, at a generic point y lying on a tooth of a comb or on the gss screens the completed local rings as can be seen from the table in Proposition 3.6, hence is étale. It follows that the image of such a point has p + 1 distinct pre-images. 51 † et) and Ym (or Y † m). This concludes the proof of part (i) of the theorem. Part (ii) follows from the (cid:3) relations between the closures of the pre-images of the seven strata in S0(p). 5.2. Analysis of (cid:101)π. 5.2.1. Analysis of (cid:101)π along the µ-ordinary strata. We denote by (cid:101)πet and (cid:101)πm the restrictions of(cid:101)π to Yet (or even Y Proposition 5.3. (i) The map (cid:101)πet : Yet (cid:101)σet : (cid:101)Sµ If A(cid:48) ∈ (cid:101)Sµ(R) then A(cid:48)[Fr] + ker(ψ) is a finite ∼→ (cid:101)Sµ is an isomorphism. Denote by ∼→ Yet the section which is its inverse. flat subgroup J satisfying the conditions listed in Proposition 1.4, pψ descends to a principal polarization φ on A(cid:48)/J and (cid:101)σet(A(cid:48)) = (A(cid:48)/A(cid:48)[Fr] + ker(ψ), φ, ι(cid:48),(cid:104)p(cid:105)−1 ◦ η(cid:48), A(cid:48)[p]/A(cid:48)[Fr] + ker(ψ)). (ii) The map(cid:101)πm : Ym → (cid:101)Sµ is finite flat totally ramified of degree p. Proof. We have already seen that(cid:101)πet is an isomorphism and that(cid:101)πm is a finite flat totally ramified map of degree p. It remains to check the assertion about (cid:101)σet. Let us first check the claims made about J. As usual, by reduction to the universal object, we may assume that R is reduced. Then A(cid:48)[F r] ∩ ker(ψ) is a finite group scheme over R, all of whose fibers have the same rank p, so is finite flat, and J = A(cid:48)[Fr] + ker(ψ) (cid:39) (A(cid:48)[Fr] × ker(ψ))/(A(cid:48)[F r] ∩ ker(ψ)) descend pψ to a principal polarization of A(cid:48)/J and form the tuple(cid:101)σet(A(cid:48)). It is now is finite flat of rank p4. It is also maximal isotropic for epψ, OE-stable and J/ ker(ψ) is Raynaud. All these statements are checked fiber-by-fiber. We may therefore a simple matter to check that if A(cid:48) = A/H where (A, H) ∈ Yet(k) then A(cid:48)/J = A/A[p] ×p(cid:39) A and A(cid:48)[p]/J = p−1H/A[p] gets mapped back to H. When we add level-N structure twisted by the diamond operator (cid:104)p(cid:105)−1 to the definition of (cid:101)σet(A(cid:48)) we ensure that (cid:101)σet is indeed the inverse of(cid:101)πet. (cid:3) The next corollary follows directly from the definitions of the various maps and we omit its proof. Corollary 5.4. (i) On R-points of the moduli problems the maps jet = πet ◦ (cid:104)p(cid:105) ◦(cid:101)σet : (cid:101)Sµ → Sµ, jm =(cid:101)πm ◦ σm : Sµ → (cid:101)Sµ are given by jet(A(cid:48), ψ, ι(cid:48), η(cid:48)) = (A(cid:48)/A(cid:48)[Fr] + ker(ψ), φ, ι(cid:48), η(cid:48)) jm(A, φ, ι, η) = (A/A[p]m, ψ, ι, η). 52 are given by (ii) The maps ON THE BAD REDUCTION OF CERTAIN U (2, 1) SHIMURA VARIETIES wm(A, H) = (A(p2), Fr(A(p)[Ver])), wet(A, H) = (A, A[p]m). Their compositions are the maps F r2 : Sµ → S(p2) µ = Sµ or F r2 : (cid:101)Sµ → (cid:101)S(p2) µ = (cid:101)Sµ (here we use the fact that S and (cid:101)S are defined over κ). wm = (cid:104)p(cid:105) ◦(cid:101)σet ◦(cid:101)πm : S0(p)m → S0(p)et, wet = σm ◦ πet : S0(p)et → S0(p)m 5.2.2. Analysis of (cid:101)π along the curves Zet and Zm. Proposition 5.5. Let (cid:101)Z be the stratum (cid:101)Sgss[Σ, Σ]. The morphism (cid:101)πet : Zet → (cid:101)Z is an isomorphism. The morphism (cid:101)πm : Zm → (cid:101)Z is totally ramified of degree p. et = Yet ∪ Zet and (cid:101)S† and induces an isomorphism between the open dense subsets Yet (cid:39) (cid:101)Sµ. From the classification of the completed local rings in Proposition 3.6 it follows that (cid:101)S† smooth, hence its local rings are integrally closed and (cid:101)πet is an isomorphism. A similar argument shows that (cid:101)πm : Y † m → (cid:101)S† In principle, the unramified direction (see Lemma 4.3) for (cid:101)πm : Y † a point (cid:101)x ∈ (cid:101)Z could be transversal to (cid:101)Z or tangential to it. We claim that it is everywhere transversal, i.e. the schematic pre-image of (cid:101)Z is Zm (with its reduced structure) but(cid:101)πZm is totally ramified of degree p. This can be seen in a variety of µ = (cid:101)Sµ ∪ (cid:101)Z. The map (cid:101)πet : Y m → (cid:101)S† µ is finite flat totally ramified of degree ways.9 We shall deduce it from Corollary 5.4. Observe first that the maps jet and jm extend to similarly denoted maps et → (cid:101)S† p, where Y † m = Ym ∪ Zm. Proof. Let Y † µ is finite, µ is µ at † jet = πet ◦ (cid:104)p(cid:105) ◦(cid:101)σet : (cid:101)S† jm =(cid:101)πm ◦ σm : S† µ → (cid:101)S† µ, µ → S† µ, and may then be restricted to the gss curves (cid:101)Z and Sgss. The claim follows now from the following established facts: (a) σm : Sgss (cid:39) Zm and (cid:101)σet : (cid:101)Z (cid:39) Zet are isomorphisms, (b) πet : Zet → Sgss is totally ramified of degree p (equivalently, et (cid:39) Sgss is an isomorphism) (c) jet ◦ jm = F r2 hence, restricted to the πet : Z (p) (cid:3) The same argument used to show that (cid:101)πet extends to an isomorphism on Y curve Sgss, it is totally ramified of degree p2. and that(cid:101)πm extends to a totally ramified map on Y † † et, Then (cid:101)πet extends to an isomorphism from Y et to the closure (cid:101)Sµ of (cid:101)Sµ. The map (cid:101)πm extends to a totally ramified map of degree p from Y m to (cid:101)Sµ. Proposition 5.6. Let Y et and Y m denote the closures of Yet and Ym in S0(p). m gives the following. A computation similar to the above, that we leave out, yields the following. m → Y † et be the map θ = ρet ◦ πm (see Corollary 4.6). Corollary 5.7. Let θ : Y † (cid:104)p(cid:105) ◦(cid:101)πet ◦ θ =(cid:101)πm. Then 9Were the unramified direction everywhere tangential to (cid:101)Z, the schematic pre-image of (cid:101)Z would be a nilpotent thickening of order p of Zm, but(cid:101)π would be an isomorphism on the reduced curve. transversal, but tangential to (cid:101)Z at finitely many points. In general, of course, there is also a "mixed option", where the unramified direction is generically ON THE BAD REDUCTION OF CERTAIN U (2, 1) SHIMURA VARIETIES 5.2.3. Analysis of(cid:101)π along the gss screens Ygss. Let W be an irreducible component other. Outside (the closure of) Zet and Zm the restriction of (cid:101)π to W, which we denote from now on (cid:101)πW , is étale. It is also étale at y ∈ Zet(k). This follows from of the closure Y gss of Ygss. As we have seen in Theorem 4.11, these irreducible components are smooth P1-bundles over Fermat curves, and do not intersect each 53 §3.3.2. (1) We have (cid:101)π(Z m ∩ W ) =(cid:101)π(Z et ∩ W ). (5.1) Proof: (cid:101)π(Z et ∩ W ) is an irreducible component of (cid:101)Z, the closure of the stratum (cid:101)Z = (cid:101)Sgss[Σ, Σ]. So is (cid:101)π(Z m ∩ W ). The two intersect at the image of any point ζ irreducible components of (cid:101)Z are disjoint, the two components coincide. which is "a base of a tooth of a comb", points where Z et and Z m meet. Since the (2) We have (cid:101)π(W ) ∩ (cid:101)Z =(cid:101)π(Z et ∩ W ). Proof: this follows from (1) since(cid:101)π−1((cid:101)Z) = Z et ∪ Z m. (3) Let W, W (cid:48) be two components of Y gss. Then(cid:101)π(W ) ∩(cid:101)π(W (cid:48)) = ∅. Proof: Each (cid:101)π(W ) is an irreducible component of (cid:101)π(Y gss). But the irreducible components of (cid:101)π(Y gss) are disjoint from each other and are uniquely determined by their intersection with (cid:101)Sµ, i.e. with (cid:101)Z. The claim follows from (2), since (cid:101)π(Z et ∩ W ) ∩(cid:101)π(Z et ∩ W (cid:48)) = ∅, as(cid:101)πet is an isomorphism. (4) We give another proof of (5.1). It is based on the following lemma, which is of independent interest. Recall that S is defined over κ = Fp2, although we consider it over k = Fp2. It follows that Gal(k/κ) permutes the irreducible components of Sgss. The diamond operators also act on these irreducible components. Lemma 5.8. Let Z be an irreducible component of Sgss. Then F rp2 (Z) = (cid:104)p(cid:105) (Z). Proof. For the proof of the lemma we may increase N. Indeed, if NN(cid:48) and Z, Z(cid:48) are as above for N and N(cid:48), with Z(cid:48) mapping to Z, then the validity of the lemma for Z(cid:48) implies it for Z. Since the closure Z of every irreducible component of Sgss contains at least two superspecial points, and since when N is large enough, through any two superspecial points passes at most one such Z [Vo], it is enough to prove that for x ∈ Sssp(k) F rp2(x) = (cid:104)p(cid:105) (x). Let x = (A, φ, ι, η). Every supersingular elliptic curve B over k has a model B0 over κ, whose Frobenius of degree p2 satisfies F rp2 = p. By the Tate-Honda theorem [Ta], all the endomorphisms of B are already defined over κ. We may therefore assume that A (cid:39) B3 and ι are defined over κ. Since A admits at least one principal polarization defined over κ, and its endomorphisms are all defined over κ, φ is defined over κ. Thus (A, φ, ι) is invariant under F rp2. But the relation F rp2 = p on A[N ] means that F rp2(η) = (cid:104)p(cid:105) ◦ η, which concludes (cid:3) the proof. ON THE BAD REDUCTION OF CERTAIN U (2, 1) SHIMURA VARIETIES 54 Now use the relation (cid:104)p(cid:105)−1 ◦ F r2 p = (cid:104)p(cid:105)−1 ◦ jet ◦ jm = πet ◦(cid:101)σet ◦(cid:101)πm ◦ σm (5.2) (cid:101)πm(y) =(cid:101)πet(y(cid:48)). from Corollary 5.4, and its extension to S† µ from the proof of Proposition 5.5. The left hand side fixes the irreducible components of Sgss, hence also the irreducible components W of Y gss. Let y ∈ Z m ∩ W . Then y(cid:48) =(cid:101)σet ◦(cid:101)πm(y) ∈ Z et ∩ W, or This shows that(cid:101)π(Z m ∩ W ) =(cid:101)π(Z et ∩ W ) as was to be shown. (5) The map(cid:101)πW : W →(cid:101)π(W ) is finite flat of degree p + 1. Proof: This follows from (3) since(cid:101)π in the large is finite flat of degree p + 1. We next want to analyze how (cid:101)π is behaved when restricted to a fiber Wx = Then(cid:101)π(ym) (cid:54)=(cid:101)π(yet). (5.2), which are in the same fiber for (cid:101)π, are distinct. But π(y(cid:48)) = (cid:104)p(cid:105)−1 π(y)(p2). (6) Let ym and yet be the unique points on Zm ∩ Wx and Zet ∩ Wx respectively. Proof: Equivalently, we have to show that the images under π of y and y(cid:48) as in π−1(x) of π above a gss point x. Recall that Wx (cid:39) P1. We claim that if π(y) = (A, φ, ι, η) then already (A, φ, ι) is not defined over κ, so is not isomorphic to (A(p2), φ(p2), ι(p2)). This follows from the fact, established in [Vo], that when N = 1 any irreducible curve Z in the supersingular locus of the coarse moduli space associated with the algebraic stack S is defined over κ, and is birationally isomorphic to the Fermat curve C : xp+1 + yp+1 + zp+1 = 0. Let C → Z be the normalization of Z. This C has p3 + 1 κ-rational points, which are precisely the points mapping to superspecial points on Z. Furthermore, all the self-intersections of Z are at κ-rational points. It follows that no x ∈ Z(k) which is p. Since the diamond operators do not affect (A, φ, ι), a fortiori gss is fixed under F r2 π(y(cid:48)) (cid:54)= π(y). Starting with x = x(1) ∈ Sgss(k) we may now form a sequence of points x(1), . . . , x(r) such that if y(i) m and y(i) et are the respective points on Wx(i) then (cid:101)π(y(i+1) m ) =(cid:101)π(y(i) et ). p (x) = x. of (6), the unique point on Zet ∩ Wx is such a point. Proof: We have to show that the map is generically 1-1. For that it is enough This sequence becomes periodic after d steps, where d is the minimal number so that (cid:104)p(cid:105)−d ◦ F r2d (7) The map(cid:101)π : Wx →(cid:101)π(Wx) is a birational isomorphism. to find a single point y ∈ Wx so that(cid:101)π is étale at y and(cid:101)π−1((cid:101)π(y)) = {y}. In view We do not answer the question whether(cid:101)π is everywhere 1-1. We summarize the Theorem 5.9. The map(cid:101)π induces a bijection between the vertical irreducible com- ponents of (cid:101)S and of S0(p). The map π induces a bijection between the vertical The vertical irreducible components of (cid:101)S are mutually disjoint. Let W be a vertical irreducible component of S0(p). Then (cid:101)πW is finite flat of degree p + 1 and is étale outside W ∩ Z m. The restriction of(cid:101)πW to Wx = π−1(x) for x ∈ Sgss is a birational irreducible components of S0(p) and the irreducible components of the curve Sss. discussion of this section in the following theorem. ON THE BAD REDUCTION OF CERTAIN U (2, 1) SHIMURA VARIETIES 55 isomorphism and maps the unique intersection points of Wx with Zet and Zm to distinct points. 6. Appendix 6.1. The classification of the gss Dieudonné modules. In the appendix we perform some computations on the covariant Dieudonné module of a gss abelian variety. We first recall their classification, following Vollaard [Vo]. Fix δ ∈ µp2−1 ⊂ W (κ) ⊂ W (κ)Q = Ep such that δσ = δp = −δ. Let M be the free W (κ)-module on e1, e2, e3, f1, f2, f3 and let OE act on the ei via Σ (the canonical embedding of E in Ep) and on the fi via Σ. Let F be the σ-linear endomorphism10 of M whose matrix w.r.t. the above basis is  1 p 1 p 1 p  , i.e. F (e1) = pf1,F (e2) = f2, . . . , F (f3) = e3. Let V be the σ−1-linear endomor- phism with the same matrix. Note that τ = V −1F is the identity on M. Let Mk = W (k) ⊗W (κ) M and extend F, V semi-linearly as usual. Then τ becomes σ2-linear. Let (cid:104),(cid:105) be the alternating pairing on Mk satisfying (cid:104)ei, fj(cid:105) = δ · δij, (cid:104)ei, ej(cid:105) = (cid:104)fi, fj(cid:105) = 0. This Mk is the Dieudonné module of Ax[p∞] for any x ∈ Sssp(k). It is isomorphic11 to the module used in part (4) of the proof of Theorem 5.2. The Lie algebra of Ax is identified with Mk/pMk[V ] = V −1pMk/pMk = F Mk/pMk (cid:39) (Mk/V Mk)(p) and is spanned over k by e1, e3, f 2. Following [Vo] we denote M(Σ) = (cid:104)e1, e2, e3(cid:105)W (κ) by M0 and M(Σ) by M1. We introduce on M0 the skew-hermitian form {x, y} = (cid:104)x, F y(cid:105) . We extend it to a bi-additive form on M0,k which is linear in the first variable and σ-linear in the second. It satisfies {x, y} = −{y, τ−1(x)}σ, {τ (x), τ (y)} = {x, y}σ2 . We denote the unitary isocrystal Q⊗ M by N = N0 ⊕ N1 and write also C for N0. When we base-change to the field of fractions of W (k) we shall add, as before, the subscript k. Note that the Qp-group J = GU (C,{,}) is isomorphic, in our case, to G/Qp . (In general, it might be an inner form of it.) If Λ ⊂ C is a W (κ)-lattice we let Λ∨ = {x ∈ C{x, Λ} ⊂ W (κ)}. 10In the appendix we depart from our habit of writing F as a linear map from M(p) to M. 11The change in notation is made to conform with [Vo]. Previously we tried to match [Bu-We]. ON THE BAD REDUCTION OF CERTAIN U (2, 1) SHIMURA VARIETIES 56 If Mk were the Dieudonné module of Ax[p∞] for a superspecial point x, then the components of Sss passing through x are classified, as we have seen before, by the set J = {(1 : ζ) ∈ P1(W (κ)) ζ p+1 + 1 = 0}. The vertices of the Bruhat-Tits tree of J are of two types. The special (s) lattices L (1) are the lattices Λ(cid:48) for which Λ(cid:48) ⊂ Λ(cid:48)∨, lengthW (κ)(Λ(cid:48)∨/Λ(cid:48)) = 2. For example, M0 ∈ L (1). The hyperspecial (hs) lattices L (3) are those satisfying Λ = Λ∨. Finally, the edges of the tree connect a lattice Λ(cid:48) of type (s) to a vertex Λ of type (hs) if Λ(cid:48) ⊂ Λ ⊂ Λ(cid:48)∨. One computes that the p + 1 vertices of type (hs) adjacent to M0 are the lattices where ζ ∈ J and eζ = p−1(e1 + ζe3). skew-hermitian pairing (, ) = {,} mod p is given in this basis by the matrix Fix ζ and let Λ = Λζ, V = Λ/pΛ, a vector space over κ with basis e1, e2, eζ. The Λζ = (cid:104)e1, e2, eζ(cid:105)W (κ)  δ  . 1 1 1 Theorems 2 and 3 of [Vo] imply the following. The k-points of the irreducible component of Sss passing through the superspecial point x and labeled by ζ are in one-to-one correspondence with YΛ(k) = {U ⊂ Vk dim U = 2, U⊥ ⊂ U}. Caution has to be taken as we are over k and not κ : (U⊥)⊥ = τ (U ) and not U. The point x corresponds to U = (cid:104)e1, e2(cid:105) . In general, let a, b ∈ k and U⊥ = {x ∈ Vk (x, U ) = 0}. Ua,b = (cid:104)e1 + aeζ, e2 + beζ(cid:105) . Here Then (a, b) is where and a,b = (cid:104)e1 − bpe2 − apeζ(cid:105) U⊥ is contained in Ua,b if and only if ap + a − bp+1 = 0. It follows ([Vo], Lemma 4.6) that the irreducible components of Sss are isomorphic to the smooth projective curve whose equation is xpz + xzp − yp+1 = 0. This is just the Fermat curve xp+1 + yp+1 + zp+1 = 0 in disguise. Moreover, the Dieudonné module of the abelian variety Aa,b "sitting" at the point Ma,b = M 0 a,b ⊕ M 1 a,b a,b = (cid:104)e1 + [a]eζ, e2 + [b]eζ, peζ(cid:105)W (k) M 0 a,b =(cid:10)f1 − [b]p−1f2 − [a]fζ, f2, pfζ (cid:11) . M 1  F =  V = 1−[bp] γ p −[b] 1 p [b] [ap] + [a] 1 [bp] p 1 −[b1/p] p σ−1(γ) −[b] 1 p [b] 1 [a1/p] + [a] [b1/p] p γ = −p−1([ap] + [a] − [bp+1]) ∈ W (k).   ON THE BAD REDUCTION OF CERTAIN U (2, 1) SHIMURA VARIETIES 57 Here [a] is the Teichmüller representative of a and fζ = p−1(f3 − ζf1). The matrices for F and V can now be computed. To simplify the notation let 1 = e1 + [a]eζ, 2 = e2 + [b]eζ, 3 = peζ, φ1 = f1 − [b]p−1f2 − [a]fζ, φ2 = f2, φ3 = pfζ. Then relative to the basis {1, 2, 3, φ1, φ2, φ3} and with 6.2. The quotient A/H for gss A. Let (a, b) ∈ k2 but not in κ2. This guarantees that Aa,b is gss, and every gss A is of this sort, for an appropriate x ∈ Sssp(k) and an appropriate ζ ∈ J . Let H ⊂ A[p] be an isotropic Raynaud subgroup scheme. Let A(cid:48) = A/H. We know that M (H) ⊂ Ma,b/pMa,b must contain ker V ∩ ker F = (cid:104)3(cid:105)k . In a,b such that F η = 3. We addition, M (H) should contain a vector η from M 1 see that the most general form of such an η is a,b/pM 1 η = u(φ2 + [bp]φ3) + v(φ2 + [b1/p]φ3), (u : v) ∈ P1(W (k)). Thus M (H) = (cid:104)3, η(cid:105)k . Note that by the assumption that (a, b) is not in κ2, neither a nor b lies in κ. Thus H is uniquely classified by (u : v) ∈ P1(k). The point u = 0 corresponds to an H such that M (H) is killed by F , or H is killed by V er. This H will be of type αp2,Σ and (A, H) will lie then on Zet. The point v = 0 will correspond to an H such that M (H) is killed by V, or H is killed by F rob. This H will be of type α∗ p2,Σ and (A, H) will lie then on Zm. type GΣ[p]. Then M (A(cid:48)[p∞]) = M(cid:48) will sit in an exact sequence Assume from now on that we are not in these two special cases, so that H is of and inside Nk, M(cid:48) =(cid:10)1, 2, p−13, φ1, p−1η, φ3 (cid:11) W (k) , provided u (cid:54)= −v. If u = −v the same basis works, if we replace φ3 by φ2. Assume from now on that u (cid:54)= −v. 0 → M → M(cid:48) → M (H) → 0, ON THE BAD REDUCTION OF CERTAIN U (2, 1) SHIMURA VARIETIES 58 We calculate the matrices of F and V in this basis as we did for M = Ma,b before. The matrix of F comes out to be 1−[bp] pγ p p[b](u + v)−1 [ap] + [a] − [b]w p(u + v)−1 [bp] − w 1 uσ + vσ uσ([bp2 ] − [b]) p where we put w = (u[bp] + v[b1/p])(u + v)−1, while the one of V is     . 1 −[b1/p] pσ−1(γ) + vσ−1 uσ−1 ([b1/p2 vσ−1 ] − [b]) p p p[b](u + v)−1 p(u + v)−1 [a1/p] + [a] − [b]w [b1/p] − w 1 We see that M(cid:48)/pM(cid:48)[V ] ∩ M(cid:48)/pM(cid:48)[F ] is spanned by the images modulo pM(cid:48) of φ3 and of x1 + y2 + zp−13 provided x, y, z ∈ W (k) are such that xσ([ap] + [a] − [b]w) + yσ([bp] − w) + zσ ≡ 0 mod p ([a1/p] + [a] − [b]w) + yσ−1 ([b1/p] − w) + zσ−1 ≡ 0 mod p. xσ−1 These two equations are equivalent to x([b1+1/p] − [b1/p]wσ−1 ) + y([b] − wσ−1 ) + z ≡ 0 mod p, x([bp+1] − [bp]wσ) + y([b] − wσ) + z ≡ 0 mod p. p−1/p The solution set (x, y, z) mod p to these two equations is 1-dimensional, unless w ∈ κ and b = 1, where it is 2-dimensional. This last condition however translates into b ∈ κ, which we assumed not to be the case. We conclude that M(cid:48)/pM(cid:48)[V ] ∩ M(cid:48)/pM(cid:48)[F ] is always two-dimensional, of type (Σ, Σ). This settles the a-type of A(cid:48) in the cases that were deferred to the appendix in the proof of Theorem 5.2. References [Bel] [Bea] J. Bellaïche: Congruences endoscopiques et représentations Galoisiennes, Thèse, Paris XI (Orsay), 2002. A. Beauville: Complex algebraic surfaces, second edition, London Mathematical Society Student Texts 34, Cambridge Univ. Press 1996. [Bu-We] O. Bültel, T. Wedhorn: Congruence relations for Shimura varieties associated to some unitary groups, J. Instit. Math. Jussieu 5 (2006), pp. 229-261. [C-C-O] C.-L. Chai, B. Conrad, F. Oort: Complex Multiplication and Lifting Problems, AMS, [C-N] [Cri] [dJ1] [dJ2] Providence, 2013. C.-L. Chai, P. Norman: Singularities of the Γ0(p)-level structure, J. Algebraic Geom. 1 (1992), 251-278. S. E. Crick: Local Moduli of Abelian Varieties, American Journal of Mathematics 97 (1975), pp. 851-861. A.J. de Jong: The moduli spaces of polarized abelian varieties, Mathematische Annalen 295 (1993), pp. 485-503. A.J. de Jong: The moduli spaces of principally polarized abelian varieties with Γ0(p)- level structure, Journal of Algebraic Geometry 2 (1993), pp. 667-688. ON THE BAD REDUCTION OF CERTAIN U (2, 1) SHIMURA VARIETIES 59 [dS-G1] E. de Shalit, E. Z. Goren: Supersingular curves on Picard modular surfaces modulo an inert prime, Journal of Number Theory 171 (2017), pp. 391-421. [dS-G2] E. de Shalit, E.Z. Goren: A theta operator on Picard modular forms modulo an inert [dS-G3] E. De Shalit, E. Z. Goren: Foliations on unitary Shimura varieties in positive character- prime, Res. Math. Sci. 3:28 (2016). istic. 37 pp. arXiv: 1707.08102. [Eis] [Gö] [Gro] [Kos] [Lan] [Mo] [Mu] [Nor] [N-O] [De-Pa] P. Deligne, G. Pappas: Singularités des espaces de modules de Hilbert, en les caracter- istiques divisant le discriminant, Compositio Math. 90 (1994), 59-79. D. Eisenbud, Commutative Algebra with a View Toward Algebraic Geometry, GTM 150, Springer-Verlag, New York, 1995. U. Görtz: On the flatness of models of certain Shimura varieties of PEL type, Math. Ann. 321 (2001), pp. 689-727. [G-N] W. Goldring, M.-H. Nicole: The µ-ordinary Hasse invariant of unitary Shimura varieties. J. Reine Angew. Math. 728 (2017), 137–151. A. Grothendieck: Groupes de Barsotti-Tate et cristaux de Dieudonné, Univ. Montréal, Montréal, 1974. J.-S. Koskivirta: Congruence relations for Shimura varieties associated to unitary groups, Canad. J. Math. 66 (2014), pp. 1305-1326. K.-W. Lan: Arithmetic compactifications of PEL-type Shimura varieties, London Math- ematical Society Monographs 36, Princeton, 2013. [La-Ra] R. Langlands, D. Ramakrishnan: The Zeta functions of Picard modular surfaces, Univ. Montréal, Montréal, 1992. B. Moonen: Serre-Tate theory for moduli spaces of PEL type, Ann. Scient. Éc. Norm. Sup., 4e série, t. 37 (2004), pp. 223-269. D. Mumford: The red book of varieties and schemes, LNM 1358, Springer-Verlag, New- York, 1999. P. Norman: An Algorithm for Computing Local Moduli of Abelian Varieties, Annals of Mathematics 101 (1975), pp. 499-509. P. Norman, F. Oort: Moduli of Abelian Varieties, Annals of Mathematics 112 (1980), pp. 413-439. [P-R-S] G. Pappas, M. Rapoport, B. Smithling: Local models of Shimura varieties, I. Geometry and combinatorics. Handbook of moduli. Vol. III, 135-217, Adv. Lect. Math. (ALM), 26, Int. Press, Somerville, MA, 2013. [Ra-Zi] M. Rapoport, T. Zink: Period Spaces for p-divisible Groups, Annals of Mathematics Studies 141, Princeton University Press, Princeton (1996). [Ray] M. Raynaud: Schémas en groupes de type (p, ..., p), Bulletin de la Société Mathématique de France 102 (1974), pp. 241-280. [Ru-Sh] A. N. Rudakov, I. R. Shafarevich: Inseparable morphisms of algebraic surfaces, Izv. [Ta] [Vo] Akad. Nauk SSSR Ser. Mat. 40 (1976), pp. 1269–1307. J. Tate: Classes d'isogenie des varietes abeliennes sur un corps finis (d'apres T. Honda), Seminaire Bourbaki, Expose 352 (1968/69). I. Vollaard: The supersingular locus of the Shimura variety for GU (1; s), Canad. J. Math. 62 (2010), pp. 668-720. Ehud de Shalit, Hebrew University of Jerusalem, Israel [email protected] Eyal Z. Goren, McGill University, Montréal, Canada [email protected]
1906.12037	1	1906	2019-06-28T04:27:57	A global Torelli theorem for certain Calabi-Yau threefolds	[ "math.AG" ]	We establish a global Torelli theorem for the complete family of Calabi-Yau threefolds arising from cyclic triple covers of $\mathbb P^3$ branched along stable hyperplane arrangements.	math.AG	math	A GLOBAL TORELLI THEOREM FOR CERTAIN CALABI-YAU THREEFOLDS MAO SHENG AND JINXING XU To the memory of Yi Zhang Abstract. We establish a global Torelli theorem for the complete family of Calabi- Yau threefolds arising from cyclic triple covers of P3 branched along stable hyperplane arrangements. 1. Introduction Classical Hodge theory attaches complex algebraic varieties with (mixed) Hodge structures and global Torelli theorem asserts that the attached Hodge structures deter- mine complex algebraic varieties up to isomorphism. There are very rare cases of classes of complex algebraic varieties for which global Torelli theorem holds. The renowned examples include the smooth projective curves [2], Abelian varieties, polarized K3 sur- faces [9, 3, 8] (see also birational global Torelli for hyperkahler manifolds [13]) and cubic fourfolds [14, 6, 7]. In this paper, we add one more example into the above list. Our example stems from our early studies [11, 12] on Calabi-Yau varieties arising from cylic covers of projective spaces branched along hyperplane arrangements. They are cyclic triple covers of P3 branched along six hyperplanes which are stable in the sense of GIT [5] (for hyperplane arrangements in general position (non-unique) crepant resolutions of singular CY varieties exist [11]). We show that the attached weight 3 Hodge structues of these CY varieties are pure (Propositions 2.1, 5.2). The main result of the paper is the following: Theorem. Let X and Y be two Calabi-Yau threefolds which are cyclic triple covers of P3 branched along stable hyperplane arrangements. Then X and Y are isomorphic if and only if H 3(X, Z) and H 3(Y, Z) are isomorphic as polarized Hodge structure. This work is supported by Chinese Universities Scientific Fund (CUSF), Anhui Initiative in Quantum Information Technologies (AHY150200) and National Natural Science Foundation of China (Grant No. 11622109, No. 11721101). 1 2 MAO SHENG AND JINXING XU The theorem is a direct consequence of Theorems 4.4, 5.3 in the text. The method to establish the results is pretty standard, namely using the variation of Hodge structure attached to a universal family. The analysis of the corresponding period map (and its extension to stable locus) is based on the celebrated work of Deligne-Mostow [4]. Inspired by a discussion with Chin-Lung Wang on a possible classification of complete Calabi-Yau families with Yukawa coupling length one, we were led to compare our example with other known families with Yukawa coupling length one. Surprisingly, it turns out our example is essentially not new. Indeed, we show in Appendix that our family over the smooth locus formed by hyperplane arrangements in general position is birationally equivalent to the one constructed by J. Rohde [10] via Borcea-Voisin construction (Propositions A.3, A.4). To the authors' best knowledge, our example (which is essentially the one of Rohde) is the only example in literature of complete families of CY varieties with the global Torelli property. 2. Families from hyperplane arrangements Given a hyperplane arrangement A in Pn in general position, the cyclic cover of Pn branched along A is an interesting algebraic variety. When the hyperplane arrangement A moves in the coarse moduli space of hyperplane arrangements, we get a family of projective variety. In this section, we collect some known facts about Hodge structures of these cyclic covers. We say an ordered arrangement A = (H1,··· , Hm) of hyperplanes in Pn is in general position if no n + 1 of the hyperplanes intersect in a point, or equivalently, if the divisor Pm i=1 Hi has simple normal crossings. Given an odd number n, we set r = n+3 2 . Then for each ordered hyperplane ar- rangement (H1,··· , Hn+3) in Pn in general position, we can define a (unique up to isomorphism) degree r cyclic cover of Pn branched along the divisor Pn+3 i=1 Hi. In this way, if we denote the coarse moduli space of ordered n + 3 hyperplane arrangements in Pn in general position by Mn,n+3, then we obtain a universal family fn : XAR → Mn,n+3 of degree r cyclic covers of Pn branched along n + 3 hyperplane arrangements in general position. In [11], we constructed a simultaneous crepant resolution π : XAR → XAR for the family f without changing the middle cohomology of fibers. Moreover, this simulta- neous crepant resolution gives an n-dimensional projective Calabi-Yau family which is GLOBAL TORELLI THEOREM 3 maximal in the sense that its Kodaira-Spencer map is an isomorphism at each point of Mn,n+3. We denote this smooth projective Calabi-Yau family by fn : XAR → Mn,n+3. Now we recall the relation between a cyclic cover of P1 branched along points and that of Pn branched along hyperplane arrangements. Suppose (p1,··· , pn+3) is a collection of n + 3 distinct points on P1, and put Hi = {pi}× P1 ×···× P1. By the natural identification between Pn and the symmetric power Symn(P1) of P1, we can view each Hi as a hyperplane in Pn. Then it can be shown that (H1,··· , Hn+3) is a hyperplane arrangement in Pn in general position ([11], Lemma 3.4). A direct computation shows that this construction gives an isomorphism between the moduli space M1,n+3 and Mn,n+3 ([11], Lemma 3.5). Moreover, for r = n+3 2 , if we denote C as the r-fold cyclic cover of P1 branched along the n + 3 points, and X as the r-fold cyclic cover of Pn branched along the corresponding hyperplane arrangement (H1,··· , Hn+3), then we have an isomorphism X ≃ C n/N ⋊ Sn. Here N is the kernel of the summation homomorphsim (Z/rZ)n → Z/rZ. The action of Z/rZ on C is induced from the cyclic cover structure, and Sn acts on C n by permutating the n factors. We summarize the properties of the Hodge structures on X and C as the following proposition ([11], Lemma 2.7, Proposition 3.7): Proposition 2.1. Suppose n is an odd number, and (p1,··· , pn+3) is a collection of n + 3 distinct points on P1. For each 1 ≤ i ≤ n + 3, put Hi = γ({pi} × P1 × ··· × P1), viewed as a hyperplane in Pn. Let r = n+3 2 , and C be the r-fold cyclic cover of P1 branched along Pn+3 i=1 pi. Suppose X is the r-fold cyclic cover of Pn branched along Pn+3 i=1 Hi. Then we have: (1) The natural Q-mixed Hodge structure on the middle cohomology group H n(X, Q) is pure. (2) H n(X, C)(i) ≃ ∧nH 1(C, C)(i), for each 1 ≤ i ≤ r − 1, Here and from now on, we fix a primitive r-th root of unity ζ, a generator σ of the cyclic group Z/rZ, and we use H n(X, C)(i) to denote the i-eigenspace {α ∈ H n(X, C)σα = ζ iα} of H n(X, C). The notation H 1(C, C)(i) has the similar meaning. 4 MAO SHENG AND JINXING XU f Remark 2.2. Since the simultaneous crepant resolution XAR → XAR of the universal family XAR −→ Mn,n+3 does not change the middle cohomologies of the fibers, the two Q-PVHS (rational polarized variation of Hodge structures) Rn f∗CXAR and Rnf∗CXAR are isomorphic. 3. The monodromy group and period map: curve case In this section, we first determine the monodromy group of the universal family of cyclic triple covers of P1 branched along six distinct points. Then we recall Deligne- Mostow' result about period maps of this family. Take five distinct points a1,··· , a5 ∈ C. Let C be the smooth projective curve whose affine model is defined by the equation {(x, y) ∈ C2y3 = 5Y (x − ai)}. i=1 The cyclic triple covering structure induces a natural automorphism of C: σ : C → C (x, y) 7→ (x, ωy) ) is a primitive cubic root of unity. where ω = exp( 2π√−1 3 We have the decomposition of Z[ω]-modules: H 1(C, Z) ⊗Z Z[ω] = H 1(C, Z[ω])ω ⊕ H 1(C, Z[ω])¯ω where H 1(C, Z[ω])ωi := {α ∈ H 1(C, Z[ω])σ∗α = ωiα}; The formula h(α, β) := −iQ(α, ¯β) defines an Hermitian form h : H 1(C, Z[ω])ω × H 1(C, Z[ω])ω → Z[ω], i = 1, 2. where Q means the intersection pairing on H 1(C, Z). We can verify that H 1(C, Z[ω])ω is a rank four free Z[ω]-module, and the Hermitian form h is unimodular with signature (3, 1). Now we can describe the monodromy group in the curve case. Let (Λ, h) be a fixed Z[ω]-lattice of signature (3, 1), and let M1,6 := {(z1,··· , z6) ∈ (P1)6zi 6= zi, ∀i 6= j}/P GL(2, C) be the moduli space of ordered six distinct points on P1. Let f : C → M1,6 be the universal family of cyclic triple covers of P1 branched along six distinct points. Fix a base point s ∈ M1,6, and let C := f−1(s) be the fiber over s. Then we GLOBAL TORELLI THEOREM 5 have the monodromy representation ρ : π1(M1,6, s) → Aut(H 1(C, Z[ω])ω, h). Since (H 1(C, Z[ω])ω, h) ≃ (Λ, h), we can view the monodromy group Γ := ρ(π1(M1,6, s)) as a subgroup of Aut(Λ, h). In order to describe Γ, we first introduce some notations. Let θ := ω− ¯ω = √3i. Then V := Λ/θΛ is a four dimensional vector space over the finite field F3 ≃ Z[ω]/θZ[ω], and h reduces to a quadratic form q on V . Let ψ : Aut(Λ, h) → Aut(V, q) be the natural reduction map, and let v : Aut(V, q) → F∗3/F∗2 3 ≃ Z/2Z be the spinor norm. Define the following notations: Aut+(V, q) := kerv Aut+(Λ, h) := ker(v ◦ ψ) Γθ := kerψ S := {±1,±ω,±ω2} ⊂ Aut(Λ, h) S0 := {1, ω, ω2} ⊂ S P Aut(Λ, h) := Aut(Λ, h)/S P Γθ := Γθ/S0 Proposition 3.1. Γ is a subgroup of Γθ, and Γθ is generated by Γ and ω. As a direct consequence, we have the following corollary. Corollary 3.2. The projectified monodromy representation P ρ : π1(M1,6, s) → P Aut(Λ, h) has image P Γθ. and families of curves over them. Define M Before starting the proof of Proposition 3.1, we first introduce some auxiliary spaces := {(z1,··· , z6) ∈ C6 ∀i 6= j, zi 6= zj.}. /S6 be the quotient → M , s is the cyclic triple cover of P1 descends to Obviously the permutation group S6 acts on M space. Similarly as the universal family C such that for each s = (z1,··· , z6) ∈ M branched along the six points z1,··· , z6. Obviously this family C a family ¯C monodromy representations ρ : π1(M and ¯s ∈ ¯M ′, we also have the , s) → Aut(Λ, h) and ρ : π1( ¯M ′, ¯s) → Aut(Λ, h). −→ M1,6, we have a universal family C , the fiber C ′ → ¯M ′ over ¯M ′. Fixing base points s ∈ M ′ . Let ¯M ′ := M ′ ′ ′ ′ ′ → M ′ ′ f ′ ′ ′ Lemma 3.3. The monodromy groups can be determined as follows: 6 MAO SHENG AND JINXING XU (1) The monodromy group of the family ¯C ′ Aut+(Λ, h). → ¯M ′ over ¯M ′ is ρ(π1( ¯M ′, ¯s)) = (2) The monodromy group of the family C ′ → M ′ over M ′ is ρ(π1(M ′ , s)) = Γθ. Proof. (1) It is well known that the fundamental group π1( ¯M ′, ¯s) is isomorphic to the braid group B6. In order to describe it, we suppose the base point ¯s represents the unordered subset B = {1, 2,··· , 6} of C. phic to M odc(C, B), the compactly-supported mapping-class group of the pair (C, B). Moreover, B6 admits standard generators τ1,··· , τ5, where τi denotes the right Dehn half-twist around a loop enclosing the interval [i, i + 1] in C. These five generators It is a standard fact that B6 is isomor- satisfy the braid relation τiτi+1τi = τi+1τiτi+1 for i = 1,··· , 5, as well as the commutation relation τiτj = τjτi for i − j > 1; and these generators and relations give a presentation for B6. As for the monodromy action of B6 on Λ, we can see τi acts on Λ as a −ω-reflection. Then an application of Lemma (7.12) in [1] shows that ρ(π1( ¯M ′, ¯s)) = Aut+(Λ, h). (2) We know that π1(M 2 ,··· , τ 2 generated by τ 2 1 , τ 2 ′ , s) ≃ P B6, the pure braid group of six points, and P B6 is 5 . Moreover, P B6 is normal subgroup of B6, and the quotient group B6/P B6 is isomorphic to S6, the permutation group of six elements. ρ(τ 2 By the arguments in (1), we see ρ(τi) is a −ω-reflection in Λ, for i = 1, 2,··· , 5. So i ) is a ω-reflection in Λ, for i = 1, 2,··· , 5, and hence by the definition of Γθ, the monodromy group ρ(π1(M , s)) is contained in Γθ. By Lemma (4.5) in [1], we have a short exact sequence of group 1 → Γθ → Aut+(Λ, h) → Aut+(V, q) → 1. Then we get ′ [Aut+(Λ, h) : ρ(π1(M ′ , s))] ≥ [Aut+(Λ, h) : Γθ] = Aut+(V, q) = 720. On the other hand, [Aut+(Λ, h) : ρ(π1(M ′ , s))] = [ρ(π1( ¯M ′, ¯s)) : ρ(π1(M ′ , s))] ≤ [B6 : P B6] = S6 = 720. So we obtain ρ(π1(M ′ , s)) = Γθ. (cid:3) Remark 3.4. As a byproduct of the proof of Lemma 3.3, we get the commutative GLOBAL TORELLI THEOREM 7 diagram: (3.4.1) 1 1 ✲ P B6 ✲ B6 ✲ S6 ✲ 1 ❄ ≀ ❄ ✲ Aut+(Λ, h) ✲ Aut+(V, q) ✲ 1 ❄ ✲ Γθ where the rows are short exact sequences and the homomorphism S6 → Aut+(V, q) is an isomorphism. Proof of Proposition 3.1: ′′ Let M := {(z1,··· , z5) ∈ C5zi 6= zj, ∀i 6= j} be the moduli space of five distinct be the family of smooth projective curves whose ′′ ′′ ordered points in C. Let C → M affine model is defined by y3 = Q5 i=1(x − zi). We have the following inclusion of moduli spaces: M1,6 ⊂ i ′′ ✲ M Here we identify M1,6 as the space {(z1, z2, z3) ∈ (C\{0, 1})3zi 6= zj, ∀i 6= j}, and i maps a point (z1, z2, z3) to the point (0, 1, z1, z2, z3) in M . ′′ ′′ ′ ′ ′′ ′′ ′′ ′′ ′′ is Γθ. is homeomorphism to the It is easy to see that, through the map i, the space M product space M1,6 × C × C∗. Moreover, the restriction C M1,6 is isomorphic to the M1,6 → M1,6 is also Γ. On the family C over M1,6. So the monodromy group of C other hand, it can be seen directly from the construction that, the monodromy groups of the two families C two monodromy groups, and by Lemma 3.3 (2), we know the monodromy group of C are isomorphic. We then identify these and C → M → M → M Now we compare the monodromy groups of C → M M1,6. Since M and its restriction C M1,6 → ≃ M1,6 × C × C∗, by fixing a base point s = (z1,··· , z5) ∈ M1,6, , s) is generated by π1(M1,6, s) and a loop µ : θ 7→ the fundamental group π1(M (eiθz1,··· , eiθz5), 0 ≤ θ ≤ 2π. It can be verified directly that the monodromy action induced by µ is ρ(µ) = ω ∈ Γθ. So we obtain that the monodromy group of the family M1,6 → M1,6. This in turn C implies that Γθ is generated by Γ and ω. is generated by ω and the monodromy group of C → M ′′ ′′ ′′ ′′ ′′ ′′ ′′ ′′ ′′ 8 MAO SHENG AND JINXING XU Now we describe the period map of the family f : C → M1,6. Recall (Λ, h) is a Z[ω]-lattice of signature (3, 1). Let B3 := {v ∈ P(Λ ⊗Z[ω] C)h(v, v) < 0)}, which is isomorphic to the three dimensional unit ball B3. For any point s ∈ M1,6, the space H 1,0(Cs, C)¯ω := {α ∈ H 1(C, C)σ∗α = ¯ωα} is a one-dimensional linear space over C. By associating to s the line in H 1(Cs, C)¯ω generated by H 1,0(Cs, C)¯ω, we get a well defined holomorphic map (called the period map): where Muni 1,6 is the universal cover of M1,6. PC : Muni 1,6 → B3 By definition, PC is equivariant under the projectified monodromy representation: P ρ : π1(M1,6, s) → P Aut(Λ, h). Let K be the kernel of P ρ, then PC descends to the following holomorphic map, still denoted by PC fM1,6 := Muni 1,6 /K → B3. the moduli space of stable six ordered points on P1. ThenMs 1,6 := {f : {1, 2,··· , 6} → (P1)6∀a ∈ P1, ♯f−1(a, a,··· , a) ≤ 2}/P GL(2) be 1,6 is a smooth complex 1,6 as the 1,6\M1,6 is a normal crossing divisor. If we denote fMs Fox completion of fM1,6 → M1,6, then By [4], the period map PC : fM1,6 → B3 extends 1,6 ∼−→ B3. The following proposition is the starting point of our to an isomorphism fMs 1,6 → Ms Let Ms manifold and Ms global Torelli theorem. Proposition 3.5. The period mapping induces a bijective map: Ms 1,6/S6 ∼−→ B3/Aut(Λ, h). Proof. By Corollary 3.2, the covering fM1,6 → M1,6 is a Galois cover with deck trans- formation group P Γθ, so we have Ms 1,6/P Γθ ≃ B3/P Γθ. We have seen from the diagram (3.4.1) that P Aut(Λ, h)/P Γθ ≃ Aut+(V, q) ≃ S6, so we get Ms 1,6/S6 ≃ B3/P Aut(Λ, h) = B3/Aut(Λ, h). 1,6 ≃ fMs (cid:3) 4. The monodromy group and period map: Calabi-Yau threefold case In this section, we analyze the monodromy group and period map of the universal family f3 : XAR → M3,6, which is the family of cyclic triple covers of P3 branched along six hyperplane arrangements in general position. Our strategy is to use the correspondence between this family and the family of curves considered in the previous section. GLOBAL TORELLI THEOREM 9 Let H1,··· , H6 be six hyperplanes in general position in P3. Let X be the cyclic triple cover of P3 branched along the divisor P6 i=1 Hi. Similarly with the curve case, we have a natural Z/3Z =< σ > action on X, and we have the eigen-subspace decomposition H 3(X, Z[ω]) = H 3(X, Z[ω])ω ⊕ H 3(X, Z[ω])¯ω where H 3(X, Z[ω])ωi := {α ∈ H 3(X, Z[ω])σ∗α = ωiα}; i = 1, 2. By results in Section 2, we know that there exist six distinct points p1,··· , p6 on P1, such that Hi can be identified with {pi} × P1 × ··· × P1. Moreover, let C be the cyclic triple cover of P1 branched along p1,··· , p6, then the correspondence between C and X shows that the lattice (H 3(X, Z[ω])ω, h) and (H 1(C, Z[ω])ω, h) are isomorphic, both of which are rank four Z[ω]-lattice with signature (3, 1). Here the Hermitian form h on H 3(X, Z[ω]) is defined by h(α, β) = −i Q(α, ¯β), in the same way as the curve case. Let (Λ, h) be a Z[ω]-lattice of signature (3, 1), and let ρ : π1(M3,6, s) → Aut(Λ, h) be the monodromy representation of the family f3. A family version of the correspondence in Section 2 shows that under the association isomorphism φ : M1,6 → M3,6, if s2 = φ(s1), then XAR,s2 ≃ Cs1 × Cs1 × Cs1 N ⋊ S3 . The isometry φΩ : (ΛXs2 , h) ∼−→ (¯ΛCs1 , h) implies the following commutative diagram: π1(M1,6, s) ∼ φ∗ ✲ π1(M3,6, φ(s)) (4.0.1) P ρ P ρ ✲ ✛ P Aut(Λ, h) Keeping the same notations as in Section 3, the commutative diagram (4.0.1) and Corollary (3.2) give the following proposition. Proposition 4.1. The projectified monodromy representation P ρ : π1(M3,6, s) → P Aut(Λ, h) has image P Γθ. (cid:3) 10 MAO SHENG AND JINXING XU Moreover, by Proposition (2.1), we have the decompositions (4.1.1) H 3(X, C)ω = H 3,0(X, C) ⊕ H 2,1(X, C) H 3(X, C)¯ω = H 3,0(X, C) ⊕ H 2,1(X, C). By associating the isomorphic class of X with the point [H 3,0(X, C)] in B3 ≃ {v ∈ P(Λ ⊗Z[ω] Ch(v, v) < 0)}, we get the period map PX : Muni 3,6 is the universal cover of M3,6. As in the curve case, the period map PX is equivariant under the projectified monodromy representation P ρ : π1(M3,6, s) → P Aut(Λ, h). Let K be the kernel of P ρ, then PX descends to the holomorphic map fM3,6 := Muni 3,6 /K → B3, The association isomorphism φ : M1,6 ∼−→ M3,6 gives the commutative diagram 3,6 → B3, where Muni still denoted by PX . relating period maps: Muni 1,6 ∼ φ ✲ Muni 3,6 P C X P ✲ ✛ B3 This diagram descends to the following commutative diagram: (4.1.2) fM1,6 ∼ φ ✲ fM3,6 P C X P ✲ ✛ B3 Recall Ms 3,6 be the moduli space of stable six ordered hyperplanes in P3, which consists six ordered 1,6 is the moduli space of stable six ordered points on P1. Let Ms hyperplanes in P3 with at worst four-fold intersection point. Then the association isomorphism M1,6 ∼−→ M3,6 extends to an isomorphism Ms 1,6 ∼−→ Ms 3,6, and further extends to an isomorphism between Fox completions: 1,6 fMs ❄ Ms 1,6 We have the following proposition. ∼ φass ✲ fMs 3,6 ∼ ✲ Ms ❄ 3,6 GLOBAL TORELLI THEOREM 11 Proposition 4.2. There exists a unique isomorphism fMs map fM3,6 PX−−→ B3. Moreover, the following diagram is commutative: 3,6 PX−−→ B3 extending the period 1,6 fMs φass ≀ ❄ ✲ B3 ✲ PC ∼ X P ∼ 3,6 fMs Proof. By [4], the period map PC : M1,6 → B3 extends to an isomorphism fMs B3. Since fMs 1,6 ∼−→ 3,6 is a normal crossing divisor, the extendability and uniqueness follow from Riemann's extension theorem. The commutative diagram follows from the 3,6 \ Ms diagram (4.1.2). (cid:3) Similarly as Proposition 3.5, we have the following proposition. Proposition 4.3. The period mapping induces a bijective map: Ms 3,6/S6 ∼−→ B3/Aut(Λ, h). (cid:3) As a corollary, we have the following global Torelli type theorem. ′ ′ 1,··· , H Theorem 4.4. Suppose a = (H1,··· , H6) and b = (H 6) are two hyperplane arrangements in general position in P3. Let Xa(resp. Xb) be the cyclic triple cover of P3 branched along a (resp. b). Then the polarized Z-Hodge structures H 3(Xa, Z) and H 3(Xb, Z) are isomorphic if and only if after a permutation, the hyperplane arrange- ments {H1,··· , H6} and {H Proof. Let φ : H 3(Xa, Z) ∼−→ H 3(Xb, Z) be an isomorphism of polarized Z-Hodge 6} are projectively equivalent. 1,··· , H ′ ′ structures. By the decomposition (4.1.1), we see that φ is compatible with the Z/3Z- actions. Then φ induces an isomorphism H 3(Xa, Z[ω])ω ∼−→ H 3(Xb, Z[ω])ω. From this we see a and b have the same image under the period map Ms 3,6 −→ B3/Aut(Λ, h). Then 3,6/S6, which means exactly that after a permutation, the hyperplane arrangements a and b are projectively Proposition 4.3 implies a and b represent the same point in Ms equivalent. (cid:3) 5. Analysis of stable degenerations In this section, we want to extend the global Torelli type Theorem 4.4 to the stable hyperplane arrangement case. 12 MAO SHENG AND JINXING XU We first analyze boundary correspondence under the period mapping. Recall (Λ, h) is a rank four Z[ω] lattice with signature (3, 1), and we realize B3 as the open subset of P(Λ ⊗Z[ω] C) consisting of negative lines. We call a vector r ∈ Λ a short root, if h(r, r) = 1. Denote R for the set of short roots in Λ. For any r ∈ R, define the hyperplane orthogonal to r: Hr := {[v] ∈ B3 ⊂ P(Λ ⊗Z[ω] C)v ∈ Λ, h(v, r) = 0}. we can choose an ω-reflection αr along r in P Γθ. In particular, r is a fixed point of 3,6 ∼−→ B3. We write H := ∪r∈RHr. Proposition 4.2 gives an isomorphism PX : fMs 3,6 \fM3,6 ∼−→ H. Proposition 5.1. PX induces isomorphisms fM3,6 ∼−→ B3 \ H and fMs 3,6 \fM3,6 (resp. H ) is a union of 15 irreducible hypersurfaces Proof. Note first that fMs 3,6 ∼−→ B3 is an isomorphism, it in the complex manifold fMs 3,6 (resp. B3). Since PX : fMs suffices to show PX (fM3,6) ⊂ B3 \H. If x ∈ fM3,6 and PX (x) ∈ Hr for some r ∈ R, then αr. Next we choose γr ∈ π1(fM3,6, s) satisfying P ρ(γr) = αr. Since the period mapping PX−−→ B3 is π1(fM3,6, s)-equivariant, we see PX (γr(x)) = αr(PX (x)) = PX (x). fM3,6 Since PX is injective, we see γr(x) = x. Note the cover fM3,6 → M3,6 is Galois, to the kernel of the monodromy representation P ρ : π1(fM3,6, s) → P Γθ, and hence αr. So we get PX(fMs αr = P ρ(γr) = id is the identity element in P Γθ. This contradicts with the chosen of so any nontrivial deck transformation has no fixed points. This implies γr belongs 3,6 \fM3,6) ⊂ H. (cid:3) In order to give a geometric interpretation of the period mapping PX on the boundary 3,6 \fM3,6, we study the Hodge structure on cyclic triple covers of P3 branched along fMs stable six hyperplanes. For a = (H1,··· , H6) ∈ Ms 3,6\M3,6, we can still define Xa to be the cyclic triple cover of P3 branched along the divisor P6 Proposition 5.2. For each a ∈ Ms is pure. 3,6, Deligne's mixed Hodge structure on H 3(Xa, Q) i=1 Hi. Proof. Denote the homogeneous coordinates on P3 by [X0 : ··· : X3], and for 1 ≤ i ≤ 6, let ℓi be the defining homogeneous linear equation of the hyperplane Hi. Over the rational function field K(P3) of P3, there exists the finite Galois extension L := K(P3)( 3q ℓ2 ). Define Ya to be the normalization of P3 in the Galois extension ,··· , 3q ℓ6 ℓ1 ℓ1 field L. It is not hard to see that Ya is a complete intersection of two degree three GLOBAL TORELLI THEOREM 13 hypersurfaces in P5, and Ya is smooth if a ∈ M3,6. Moreover, the finite abelian group N1, which is defined as the kernel of the summation homomorphism P5 i=0 Z/3Z → Z/3Z, acts on Ya, and Xa is isomorphic to the quotient variety Ya/N1. For details of these claims, one can see [12], section 2.2. Since Xa ≃ Ya/N1, we obtain that the (mixed) Hodge structure H 3(Xa, Q) is isomorphic to H 3(Ya, Q)N1, the N1-invariant part of H 3(Ya, Q). If a ∈ M3,6, then the mixed Hodge structure on H 3(Ya, Q) is pure, since in this case Ya is a smooth projective variety. So the mixed Hodge structure on H 3(Xa, Q) ≃ H 3(Ya, Q)N1 is also pure. If a = (H1,··· , H6) ∈ Ms 3,6\M3,6, we can see that Ya has only isolated singularities. In order to show the mixed Hodge structure on H 3(Xa, Q) ≃ H 3(Ya, Q)N1 is pure, we can assume, without loss of generality, that the first four hyperplanes H1, H2, H3, H4 pass through a common point p = [1 : 0 : 0 : 0] ∈ P3, and P6 i=1 Hi is a normal crossing divisor on P3\{p}. By an automorphism of P3, we can assume the defining equations of Hi (1 ≤ i ≤ 6) are the following: ℓ1 = X0; ℓ2 = X1; ℓ3 = X2; ℓ4 = X3; ℓ5 = X1+X2+X3; ℓ6 = X0+b1X1+b2X2+b3X3. Here bi (1 ≤ i ≤ 3) are complex numbers. Then we can see Ya is the complete intersection in P5 defined by the following two homogeneous equations: Y 3 4 = Y 3 1 + Y 3 2 + Y 3 3 ; Y 3 5 = Y 3 0 + b1Y 3 1 + b2Y 3 2 + b3Y 3 3 where [Y0 : ··· : Y5] are the homogeneous coordinates on P5. The quotient morphism from Ya to Xa is π : Ya → Xa [Y0 : ··· : Y5] 7→ [Y 3 0 : Y 3 1 : Y 3 2 : Y 3 3 ]. The finite abelian group N1 = ker(P5 i=0 Z/3Z way: P −→ Z/3Z) acts on Ya by the following N1 × Ya → Ya ((a0,··· , a5), [Y0 : ··· : Y5]) 7→ [ωa0Y0 : ··· : ωa5Y5] where ω is a primitive cubic root of unity. 14 MAO SHENG AND JINXING XU It is easy to see the singular subset of Ya is π−1([1 : 0 : 0 : 0]) = {[1 : 0 : 0 : 0 : 0 : ωi]i = 0, 1, 2}. Blowing up Ya along these singular points, we get a smooth projective variety Ya, and we can see the exceptional divisor E on Ya is a disjoint union of smooth cubic surfaces. Now we take a sufficient small open neighborhood V of [1 : 0 : 0 : 0] in P3, such that V is biholomorphic to an open ball. Let U1 be the inverse image of V in Ya, and let U2 = Ya\E. Then Ya = U1 ∪ U2 and we get the following exact sequence from the Meyer-Vietoris sequence: H 2(U1∩U2, Q)N1 → H 3( Ya, Q)N1 → H 3(U1, Q)N1⊕H 3(U2, Q)N1 → H 3(U1∩U2, Q)N1. It is not difficult to see that the exceptional divisor E is a deformation retract of U1. Since E is a disjoint union of smooth cubic surfaces, we get H 3(U1, Q) ≃ H 3(E, Q) = 0. On the other hand, we can see U2/N1 ≃ Xa\{p}, where p is the inverse image of [1 : 0 : 0 : 0] ∈ P3. This implies the isomorphism H 3(U2, Q)N1 ≃ H 3(Xa\{p}, Q). Next we consider H 2(U1∩U2, Q)N1 and H 3(U1∩U2, Q)N1. Let Z by the hypersurface in C4 defined by the equation x3 4 = x1x2x3(x1 + x2 + x3), then we can see the quotient space U1 ∩ U2/N1 is homotopic to Z\{(0, 0, 0, 0)}. Since a direct computation shows that both the cohomology groups H 2(Z\{(0, 0, 0, 0)}, Q) and H 3(Z\{(0, 0, 0, 0)}, Q) are vanishing, we see H 2(U1 ∩ U2, Q)N1 = H 3(U1 ∩ U2, Q)N1 = 0. Moreover, a similar Meyer-Vietoris sequence analysis on Xa shows that H 3(Xa\{p}, Q) ≃ H 3(Xa, Q). Putting the above analysis together, we conclude H 3(Xa, Q) ≃ H 3( Ya, Q)N1. Since Ya is a smooth projective variety, we see the mixed Hodge structure on H 3(Xa, Q) is pure. (cid:3) For a stable hyperplane arrangement a = (H1,··· , H6) ∈ Ms i=1 Hi has at worst four-fold intersection point, and the number of four-fold intersection points 3,6, the divisor P6 is less than or equal to 3. Furthermore, if the number of four-fold intersection points is k, then dimQH 3(Xa, Q) = 8 − 2k, and the (H 3(Xa, Z[ω])Z[ω], h) is a rank 4 − k lattice with signature (3 − k, 1). All these facts can be checked by a careful degeneration analysis as that in the proof of Proposition 5.2. Since it is routine and a little tedious, we omit it here. Now we can extend Theorem 4.4 to stable hyperplane arrangements. GLOBAL TORELLI THEOREM 15 ′ ′ Theorem 5.3. Suppose a = (H1,··· , H6) and b = (H 6) are two stable hyper- plane arrangements in P3. Let Xa(resp. Xb) be the cyclic triple cover of P3 branched along a (resp. b). Then the polarized Z-Hodge structures H 3(Xa, Z) and H 3(Xb, Z) are isomorphic if and only if after a permutation, the hyperplane arrangements {H1,··· , H6} and {H 6} are projectively equivalent. 1,··· , H ′ ′ 1,··· , H Proof. Suppose dimQH 3(Xa, Q) = dimQH 3(Xb, Q) = 8 − 2k, with k = 0, 1, 2, 3. Then we discuss the four cases according to k. (1) k = 0: In this case, both a and b are in general position, so it follows from Theorem 4.4. (2) k = 1: Since Ms 3,6 ∼−→ B3 maps H to Br 3,6\M3,6 ≃ H/Aut(Λ, h) is irreducible, we can choose an irre- ducible component H of fMs 3,6 \ fM3,6 and a, b in H over a and b respectively. Suppose the isomorphism PX : fMs 2 := {[v] ∈ P(Λr ⊗Z[ω] C)h(v, v) < 0}. Here r ∈ R is a short root, and Λr := {α ∈ Λh(α, r) = 0} is a free Z[ω]-module of rank three. we can see the isomorphism PX : H ∼−→ Br 2 coincides with the period map H → Br 2 which associates with a point c in the smooth locus of H the line [H 3,0(Xc, C)] in P(H 3(Xc, Z[ω])ω ⊗Z[ω] C), where Xc is the cyclic triple cover of P3 branched along c. Then since the polar- ized Z-Hodge structures H 3(Xa, Z) and H 3(Xb, Z) are isomorphic, we get PX (a) = PX (b). By Proposition 4.3, this in turn implies a and b has the same image in Ms 3,6/S6, so after a permutation, the hyperplane arrangements a and b are projectively equivalent. (3) k = 2: In this case, we can choose two irreducible components H1, H2 of fMs 3,6 \ fM3,6 and points a, b in H1 ∩ H2 over a and b respectively. Moreover, if the isomorphism PX : fMs 2 , i = 1, 2, then r1 and r2 are two perpendicular short roots. Similar as the k = 1 case, the restriction PX : H1 ∩ H2 ∼−→ Br1 2 ∩ Br2 2 coincides with the period map by associating a point c in the smooth locus of H1∩ H2 the line [H 3,0(Xc, C)] in P(H 3(Xc, Z[ω])ω ⊗Z[ω] C). Then this case follows from the same argument as in the k = 1 case. 3,6 ∼−→ B3 maps Hi to Bri (4) k = 3: In this case, both a and b have three fourfold intersection points, then an elementary analysis on the configuration of hyperplane arrangements shows that after a permutation, the hyperplane arrangements a and b are projectively equivalent. 16 MAO SHENG AND JINXING XU (cid:3) Appendix A. Comparison with Rohde's example Given distinct a, b, c ∈ C\{0, 1}, Rohde [10] constructed a singular Calabi-Yau three- ′ fold X in the following way: 1 + y3 equation y3 Let W be the surface in the weighted projective space P(2, 2, 1, 1) defined by the 2 + x0x1(x1 − x0)(x1 − ax0)(x1 − bx0)(x1 − cx0) = 0. Let F be the 2 = 0. Then the cyclic group G = Z/3Z acts on W and F . Fixing a generator σ of G and let ω be a Fermat curve in P2 defined by the homogeneous equation z3 0 + z3 1 + z2 fixed primitive cubic root of unity. We define these actions explicitly: σ : W → W [x0 : x1 : y1 : y2] 7→ [x0 : x1 : ωy1 : ωy2] σ :F → F [z0 : z1 : z2] 7→ [ωz0 : z1 : z2] ′ Rohde [10] constructs the Calabi-Yau threefold as a crepant resolution of the quo- := W × F/G, where G acts on W × F in the diagonal way. Moreover, tient threefold X varying the parameters a, b, c in C\{0, 1}, Rohde obtains a family of Calabi-Yau three- folds X → M1,6, where we recall that M1,6 is the moduli space of ordered six distinct points in P1. The main goal of this section is to show Rohde's family is birationally equivalent to the family XAR → M3,6, which is the universal family of cyclic triple covers of P3 branched along six hyperplanes in general position. ′ We first analyze the structure of the singular surface W . In general, if X1 and X2 are two varieties with G-action, we say X1 and X2 are G-birationally equivalent if there exists a birational map from X1 to X2 compatible with the G-action. Let C be the cyclic triple cover of P1 branched along the six points {0, 1,∞, a, b, c} whose affine model is the curve in C2 defined by the equation y3 − x(x − 1)(x − a)(x − b)(x − c) = 0. Then G acts on C in the following way: σ : C → C (x, y) 7→ (x, ωy) Let G act on the product C × F diagonally, then G acts on the quotient C × F/G through the G-action on the first factor C. GLOBAL TORELLI THEOREM 17 Lemma A.1. W is G-birationally equivalent to the quotient C × F/G. Proof. Let F0 be the affine surface {(z1, z2) ∈ C21+ z3 2 = 0}, and define a G-action on F0 by σ(z1, z2) = (ωz1, ωz2). Then F0 is G-birational to the Fermat curve F . Define 1 + z3 the following morphism: C × F0 → W (x, y, z1, z2) 7→ (x, z1y, z2y) It is easy to see this morphism induces a G-birational equivalent between C × F0/G and W . So W is G-birationally equivalent to C × F/G. (cid:3) Now we consider the following six hyperplanes in P3 which are in general position Hi : Xi = 0 (0 ≤ i ≤ 3), H4 : 3X i=0 Xi = 0, H5 : X0 + aX1 + bX2 + cX3 = 0, where [X0 : ··· : X3] is the homogeneous coordinates on P3. Let X be the cyclic triple cover of P3 branched along P5 i=0 Hi. In order to analyze the structure of X, we define some auxiliary varieties. Let u1, v1 be linear functions of u, v defined by the following relations: We define Y as the following affine variety: Y = {(t1, u, v, y1) ∈ C4y3 1 = uvt1(t1 + 1) u1v1(v1 − u1)}. Let S be the following affine surface: S = {(w, u, v) ∈ C3w3 = uv u1v1(v1 − u1)}. Let G acts on S by σ(w, u, v) = (ω2w, u, v). Lemma A.2. We have the following birational isomorphsims: (1) X is birationally equivalent to Y . (2) Y is birationally equivalent to S × F/G, where G acts on S × F diagonally. (3) S is G-birationally equivalent to C × F/G. Proof. (1) We take the following affine model of X: X1 = {(x1, x2, x3, y) ∈ C4y3 = x1x2x3(1 + x1 + x2 + x3)(1 + ax1 + bx2 + cx3)}. (A.1.1)   u1 = 1 + u + v v1 = a + bu + cv. 18 MAO SHENG AND JINXING XU Under the coordinate transformation   x1 = t x2 = tu x3 = tv y = y the hypersurface X1 is birationally equivalent to the following hypersurface in C4: X2 = {(t, u, v, y) ∈ C4y3 = t3uv(1 + u1t)(1 + v1t)}, where u1, v1 are defined by the equations (A.1.1). Then we can see X2 is birational to Y under the following coordinate transformation:   u = u v = v t = ( 1 v1 − 1 u1 y = tu1v1( 1 )t1 − 1 u1 u1 − 1 )y1. v1 (2) It is direct to see the following affine curve F1 is G-birationally equivalent to the Fermat curve F : F1 = {(t1, x) ∈ C2x3 = t1(t1 + 1)}, where G acts on F1 by σ(t1, x) = (t1, ωx). Then the following rational map induces the desired G-birationally equivalence be- tween S × F/G and Y : S × F1 → Y (w, u, v, t1, x) 7→ (t1, u, v, wx). (3) From the equations (A.1.1), we can view (w, u1, v1) as coordinate system on C3, and we make the following coordinate transformation:   w = w u1 = t2 v1 = t2z. Under this coordinate transformation, we see S is G-birationally equivalent to the following surface: S1 = {(w, t2, z) ∈ C3w3 = (A1t2 + a1)(B1t2 + b1) t3 2z(z − 1) }, where GLOBAL TORELLI THEOREM 19   A1 = b−z b−c B1 = c−z c−b a1 = a−b b−c b1 = a−c c−b , and G acts on S1 by σ(w, t2, z) = (ω2w, t2, z). A further coordinate transformation:   z = z t3 = w1 = B1 − t2+ a1 A1 b1 a1 A1 wt2 A1B1( b1 B1 − a1 A1 ) shows that S1 is G-birational to the surface: S2 = {(z, t3, w1) ∈ C3w3 1 = t3(t3 + 1) z(z − 1)A1B1(A1b1 − B1a1)}, where G acts on S2 by σ(z, t3, w1) = (z, t3, ω2w1). Now let C1 be the following affine curve: C1 := {(z, w2) ∈ C2w3 2 = z(z − 1)A1B1(A1b1 − B1a1)}, and let G act on C1 by σ(z, w2) = (z, ω2w2). We consider the affine model F1 = {(t3, x) ∈ C2x3 = t3(t3 + 1)} of the Fermat curve F as before. The rational map C1 × F1 → S2 (z, w2, t3, x) 7→ (z, t3, x w2 ). gives a G-birationally equivalence between C1 × F1/G and S2. Moreover, we see the smooth projective model of C1 is isomorphic to C, the cyclic triple cover of P1 branched along the six points {0, 1,∞, a, b, c}. By combining all of the birational equivalences above, we obtain S is G-birationally equivalent to C × F/G. (cid:3) Now it is direct to see that, by combining Lemma A.1 and Lemma A.2, we obtain the following birational equivalence. 20 MAO SHENG AND JINXING XU ′ Proposition A.3. Given distinct a, b, c ∈ C\{0, 1}, Rohde's singular Calabi-Yau three- = W × F/G is birational to X, which is the cyclic triple cover of P3 branched fold X along P5 i=0 Hi, with Hi defined by Hi : Xi = 0 (0 ≤ i ≤ 3), H4 : 3X i=0 Xi = 0, H5 : X0 + aX1 + bX2 + cX3 = 0. Note that to give distinct a, b, c ∈ C\{0, 1} is equivalent to give six distinct points {0, 1,∞, a, b, c} in P1. Since the moduli space M1,6 of ordered six distinct points in P1 is isomorphic to the moduli space M3,6 of ordered six hyperplane arrangements in general in P3, the following birational equivalence is a direct consequence of Proposition A.3. Proposition A.4. Rohde's Calabi-Yau family X → M1,6 is birationally equivalent to the universal family XAR → M3,6 of cyclic triple covers of P3 branched along six hyperplanes in general position. ′ References [1] Allcock, D., J.A. Carlson and D. Toledo: The complex hyperbolic geometry of the moduli space of cubic surfaces, J. Algebraic Geom. 11(4), 659-724 (2002). [2] Andreotti, A.: On a theorem of Torelli, Am. J. Math. 80, 801-828 (1958). [3] Burns, D. and M. Rapoport: On the Torelli problem for Kahlerian K3 surfaces, Ann. Sc. ENS. 8, 235-274 (1975). [4] Deligne, P. and G.D. Mostow: Monodromy of hypergeometric functions and non-lattice integral monodromy, Publ. Math. Inst. Hautes ´Etudes Sci. 63, 5-89 (1986). [5] Dolgachev, I. and D. Ortland, Point sets in projective spaces and theta functions, Ast´erisque 165 (1988). [6] Laza, R.: The moduli space of cubic fourfolds via the period map, Ann. of Math. 172(2), 673-711 (2010). [7] Looijenga, E.: The period map for cubic fourfolds, Invent. Math. 177, 213-233 (2009). [8] Looijenga, E. and C. Peters: Torelli theorems for Kahler K3-surfaces, Compos. Math. 42, 145-186 (1981). [9] Piateckii-Shapiro, I. I. and I. R. Safarevic: A Torelli theorem for algebraic surfaces of type K3, Math. USSR, Izv. 5, 547-588 (1971). [10] Rohde, J.: Cyclic coverings, Calabi-Yau manifolds and complex multiplication, Lecture Notes in Math., vol. 1975, Springer, Berlin (2009). [11] Sheng, M., J. Xu and K. Zuo: Maximal families of Calabi-Yau manifolds with minimal length Yukawa coupling, Comm. Math. Statist. 1(1), 73-92 (2013). GLOBAL TORELLI THEOREM 21 [12] Sheng, M., J. Xu and K. Zuo: The monodromy groups of Dolgachev's CY moduli spaces are Zariski dense, Adv. Math. 272, 699-742 (2015). [13] Verbitsky, M.: A global Torelli theorem for hyperkahler manifolds, Duke Math. J. 162, 2929- 2986 (2013). [14] Voisin, C.: Th´eor`eme de Torelli pour les cubiques de P5, Invent. Math. 86, 577-601 (1986). Erratum in Invent. Math. 172, 455-458 (2008). E-mail address: [email protected] E-mail address: [email protected] School of Mathematical Sciences, University of Science and Technology of China, Hefei, 230026, China
1102.3623	1	1102	2011-02-17T16:18:30	Fano manifolds of Calabi-Yau type	[ "math.AG" ]	We introduce and we study a class of odd dimensional compact complex manifolds whose Hodge structure in middle dimension looks like that of a Calabi-Yau threefold. We construct several series of interesting examples from rational homogeneous spaces with special properties.	math.AG	math	FANO MANIFOLDS OF CALABI-YAU TYPE ATANAS ILIEV AND LAURENT MANIVEL Abstract. We introduce and we study a class of odd dimensional com- pact complex manifolds whose Hodge structure in middle dimension looks like that of a Calabi-Yau threefold. We construct several series of interesting examples from rational homogeneous spaces with special properties. 1. Introduction If we think, quite naively, to mirror symmetry as a kind of mysterious process exchanging families of Calabi-Yau threefolds in such a way that the Hodge diamond rotates by a quarter-turn, an obvious difficulty arises for rigid Calabi-Yau's: their mirrors could not be Kahler! The solution given to that issue in [CDP, Sch] was to describe the mirror of certain rigid Calabi-Yau threefolds as families of higher dimensional manifolds of a very special kind. In particular, these mirror manifolds have odd dimension 2n + 1 and their Hodge structure in middle dimension looks like that of a Calabi-Yau threefold, in the sense that the only non-zero Hodge numbers are hn−1,n+2 = hn+2,n−1 = 1 and hn,n+1 = hn+1,n. This is precisely the kind of manifolds that we study in this paper, with the idea that, under some conditions that will be made more precise below, these manifolds of Calabi-Yau type should share some of the very nice properties of Calabi-Yau threefolds. The definition that we will use may not exactly be the correct one and we consider it as provisional. Our main general result will be that for a manifold of Calabi-Yau type with non-obstructed deformations, it implies that the relative intermediate Jacobian forms an integrable system over the gauged moduli space -- a result due to Donagi and Markman for Calabi-Yau threefolds. The period map also has a very similar behavior to the Calabi-Yau setting. Related homological properties are explored in [IK], where it is proved that one can extend at least to certain Fano manifolds of Calabi-Yau type, the celebrated result of Voisin according to which the Griffiths group of a Calabi-Yau threefold cannot be finitely generated. But the main goal of the present paper is to construct examples of man- ifolds of Calabi-Yau type, which will all be Fano's. Our main source of examples will be complete intersections in homogeneous spaces: we will show that, for such a construction to work, one needs strong numerical co- incidences, notably between the dimension and the index of the ambient 1 2 A. ILIEV AND L. MANIVEL homogeneous space. These coincidences are observed in some cases. No- tably, we construct interesting examples as quadratic sections of homoge- neous spaces that are Mukai varieties of even dimension (Proposition 3.3), while for Mukai varieties of odd dimension we need to take double covers branched over quadratic sections (Proposition 3.6). But our most intriguing series of examples is constructed from homoge- neous spaces with the property that their projective dual is a hypersurface of degree equal to the coindex minus one (the coindex of a Fano manifold being defined as the dimension minus the index). We are aware of four cases for which this strange coincidence can be observed, and we show that suit- able linear sections provide examples of Fano manifolds with the required Hodge numbers. In fact we can really conclude that they are of Calabi-Yau type only in half of the cases; for the two others there remains a tedious computation to be done, but we have no doubt about the fact that the final conclusion should be positive. In each case, we would conclude that a cer- tain type of hypersurfaces can be, generically, represented as a linear section of the dual hypersurface to our homogeneous space, in a finite number of ways. For example, a generic cubic sevenfold can be represented as a linear section of the famous Cartan cubic, the E6-invariant hypersurface in P26, in a finite number of different ways. This has interesting consequences for its derived category, as we show in [IM2] where this example in studied in more details. It would be interesting to find out more examples of Fano manifolds of Calabi-Yau type. A possible strategy would be to start from rigid Calabi- Yau threefolds and try to identify their mirrors systematically. We expect that these varieties will exhibit a rich and interesting geometry. 2. Definition and first properties We start with our main definition. Definition 2.1. Let X be a smooth complex compact variety of odd dimen- sion 2n + 1, n ≥ 1. We call X a manifold of Calabi-Yau type if (1) The middle dimensional Hodge structure is numerically similar to that of a Calabi-Yau threefold, that is hn+2,n−1(X) = 1, and hn+p+1,n−p(X) = 0 f or p ≥ 2. (2) For any generator ω ∈ H n+2,n−1(X), the contraction map H 1(X, T X) ω −→ H n−1(X, Ωn+1 X ) is an isomorphism. (3) The Hodge numbers hk,0(X) = 0 for 1 ≤ k ≤ 2n. A Calabi-Yau threefold is of course a manifold of Calabi-Yau type. The definition may not be the optimal one, since in particular, even in dimension three, there exist manifolds of Calabi-Yau type which are not Calabi-Yau stricto sensu. Also, the condition of being a manifold of Calabi-Yau type is FANO MANIFOLDS OF CALABI-YAU TYPE 3 clearly invariant under small deformations but probably not under arbitrary deformations, contrary to being Calabi-Yau. Rather, our main motivation is to find interesting examples of manifolds of Calabi-Yau type, and to investigate their geometry. In many respects they will behave like Calabi-Yau threefolds, and sometimes the examples we will meet will be related, in a rather non trivial way, to Calabi-Yau threefolds. Sometimes we will observe interesting differences of behavior. Deformations of Calabi-Yau manifolds are always unobstructed, as follows from the celebrated Tian-Todorov theorem. We don't know whether this always remains true for manifolds of Calabi-Yau type. (Actually all the concrete examples that we deal with are Fano, hence have unobstructed deformations.) Nevertheless, our first observation is that versal deformations of manifolds of Calabi-Yau type give rise, exactly as families of Calabi-Yau threefolds do, to some beautiful integrable systems. Consider a versal family π : X → B of manifolds of Calabi-Yau type, with special fiber X. We suppose that the base B is identified with an open subset of H 1(X, T X). Since hn+p+1,n−p(X) = 0 for p ≥ 2, the line bundle Rn−1π∗Ωn+2 X /B over B is holomorphic, and the complement B∗ of the zero section, with the pull-back X∗ of X , defines the family of gauged manifolds of Calabi-Yau type. Recall that the intermediate Jacobian of X is the complex torus J(X) = F n+1H 2n+1(X, C)∨/H2n+1(X, Z). Since X is of Calabi-Yau type, the dimension of J(X) is hn+1,n(X) + 1 = h1(T X) + 1 = dim B∗. Globalizing the construction, we get a torus bundle, the relative intermediate Jacobian J (X /B) −→ B∗. The theorem proved by Donagi and Markman for Calabi-Yau threefolds [DM] can be extended to our generalized setting: Theorem 2.2. The torus fibration J (X /B) −→ B∗ defines a completely integrable Hamiltonian system. Proof. The proof of the Theorem for Calabi-Yau threefolds is easy to extend to our setting. The main observation is that the tangent exact sequence of the C∗-bundle B∗ → B can be identified with the Hodge filtration of H 2n+1(X∗/B∗, C). More precisely there is a commutative diagram 0 −→ T(X,ω)(X∗/X ) −→ T(X,ω)X∗ −→ ↓ ↓ TXX ↓ −→ 0 0 −→ H n+2,n−1(X) −→ F n+1H 2n+1(X, C) −→ H n+1,n(X) −→ 0, where the vertical maps are all isomorphisms. In particular the middle one is given by the differential of ω (considered as a section on X∗ of the holo- morphic bundle F n+2H 2n+1(X∗/B∗, C)), by the Gauss-Manin connection: 4 A. ILIEV AND L. MANIVEL this yields, by Griffiths' transversality, a map ∇ω : T(X,ω)X∗ −→ F n+1H 2n+1(X, C), which must be an isomorphism since its two graded parts are. This observation yields a global identification between the cotangent bun- dle of X∗ and the dual of the Hodge bundle F n+1H 2n+1(X∗/B∗, C). Now ΩX∗ has a canonical symplectic structure. We need to check that it descends to the torus fibration J (X /B), which is just a quotient by the relative lat- tice defined by locally constant (n + 1)-cycles. So there remains to verify that such a locally constant cycle, say γ, when considered as a one-form on X∗, is closed. But the argument of Donagi-Markman, which consists in constructing a local primitive by integrating ω over γ, adapts mutatis mutandis. (cid:3) Remark. This implies the existence of a special Kahler structure on the base, with an extremely rich geometry [Fr]. The existence of that integrable system is related to that of a certain cubic form on the base manifold, which in the usual Calabi-Yau setting is the famous Yukawa cubic. The Yukawa cubic is defined, for a gauged Calabi-Yau threefold (X, ω), by the composition Sym3H 1(X, T X) −→ H 3(X, ω−1 X ) SD−→ H 0(X, ω2 X )∨ ω2 −→ C, where SD denotes Serre duality. In our generalized setting, this construction can be extended as follows. First consider the natural map ∧n−1T X ⊗ ∧n−1T X → ∧2n−2T X, which is symmetric if n is odd and skew-symmetric if n is even. Twisting by ω2 X ⊗ ωX with the same parity. We get an induced map (always symmetric, since the cup-product has the same parity as the degree, here n − 1) X we get a map Ωn+2 X ⊗ Ωn+2 X → Ω3 Sym2H n−1(X, Ωn+2 X ) −→ H 2n−2(X, Ω3 sending the square of a generator ω of H n−1(X, Ωn+2 of H 2n−2(X, Ω3 X ⊗ ωX). We can then define a map X ⊗ ωX), X ) to an element s(ω) Sym3H 1(X, T X) −→ H 3(X, ∧3T X) s(ω) −→ H 2n+1(X, ωX ) = C, which is our generalized cubic at the point of B∗ defined by the gauged manifold of Calabi-Yau type (X, ω). For a slightly different view-point, consider the period map P : B∗ → H 2n+1(X0, C), (X, ω) 7→ ω, defined once the family X → B has been trivialized (in the differentiable category). The same proof as in the usual Calabi-Yau setting leads to the following statement: FANO MANIFOLDS OF CALABI-YAU TYPE 5 Proposition 2.3. The period map of a versal family of manifolds of Calabi- Yau type is a local immersion. Moreover, its image is a pointed cone in H 2n+1(X0, C), which is La- grangian relatively to the symplectic structure defined by the intersection product. It would be interesting to investigate under which conditions a Frobenius structure can be defined from our generalized Yukawa cubic. We know from [CDP] that everything works fine for one of the most interesting examples of manifolds of Calabi-Yau type, the cubic sevenfold. Is there any chance to extend this to other examples? One can make a first step in this direction by mimicking the proof of Proposition 3.3 in [Vo], which allows one to prove that: Proposition 2.4. One can choose a local section ω of F n+2H 2n+1(X∗/B∗, C) in such a way that the Yukawa cubic Y ω : Sym3H 1(X, T X) −→ H 3(X, ∧3T X) s(ω) −→ C is defined by the derivatives of some potential F on B. Understanding under which conditions this potential would satisfy the WDVV equation seems to be a very difficult problem, but the question is posed: could there be a version of mirror symmetry for (some) manifolds of Calabi-Yau type? 3. Complete intersections of Calabi-Yau type 3.1. Hodge numbers of complete intersections. In this section we re- view well-known results about the cohomology of complete intersections in weighted projective spaces. Let us begin with a degree d hypersur- face Xd ⊂ wPn, where w = (w0, . . . , wn) is the set of weights defining the weighted projective space wPn. The usual projective space Pn corresponds to w = (1, . . . , 1). In that case, it is well known that the primitive cohomology of Xd can be expressed in terms of the Jacobian ring R = C[x0, . . . , xn]/( ∂F ∂x0 , . . . ∂F ∂xn ), where F denotes an equation of Xd. In fact there exists a general relation of that kind for any quasi-smooth hypersurface Xd ⊂ wPn, which is due to Steenbrink [St]: H n−p−1,p(Xd)0 ≃ R(p+1)d−w, where w = w0 + · · · + wn and Rk denotes the degree k component of R, with respect to the natural grading defined by deg(xi) = wi. 6 A. ILIEV AND L. MANIVEL An obvious consequence is that if d divides w, that is, w = (p + 1)d form some p, then H n−q−1,q(Xd)0 = 0 for q ≤ p − 1, H n−p−1,p(Xd)0 ≃ R0 ≃ C, H n−p−2,p+1(Xd)0 ≃ Rd ≃ H 1(Xd, T Xd). In particular, if n = 2m is even and w = (m − 1)d, the primitive co- homology of Xd defines a Hodge structure of the same type as that of a Calabi-Yau threefold. We find a short list of smooth examples in dimension n − 1 ≥ 4: n − 1 7 5 5 w 19 17 16, 2 d moduli X 3 2, 3 4 84 83 90 cubic sevenfold cubic section of Q6 double quartic fivefold We will meet again the cubic sevenfolds later on. They will be investigated in a more systematic way in [IM2]. 3.2. A technical lemma. We would like to generalize the previous exam- ples and show that certain low degree hypersurfaces in homogeneous spaces are of Calabi-Yau type. The general set-up would be the following. Suppose that Σ is a Fano variety of dimension dΣ, with −KΣ = ιL for some ample line bundle L. Suppose moreover that L is generated by global sections and consider the zero-locus X of a general section of L. Its dimension is of course dX = dΣ − 1, and −KX = (ι − 1)LX . The proof of our next technical lemma is a rather standard play with long exact sequences. It is essentially an application of Griffiths' techniques, for the computation of Hodge structures of hypersurfaces [Gr, Vo]. Lemma 3.1. Let Σ be such that dΣ = 2ι + 2. (1) Suppose that for 0 ≤ k < p < ι, H dΣ−p+k−1(Σ, Ωp−k Σ ⊗ L−k) = H dΣ−p+k(Σ, Ωp−k Σ ⊗ L−k−1) = 0. Then hp,dX −p(X) = 0 for p < ι. (2) Suppose that for 1 ≤ k < ι and ǫ, ǫ′ ∈ {0, 1}, H ι+k+1+ǫ+ǫ′ (Σ, Ωi−k Σ ⊗ L−k+1−ǫ) = 0. Then hι+2,ι−1(X) = 1. (3) Suppose moreover that for 2 ≤ k ≤ ι and ǫ, ǫ′ ∈ {0, 1}, H ι+k+1+ǫ+ǫ′ (Σ, Ωi−k Σ ⊗ L−k−ǫ) = 0. Then X is a manifold of Calabi-Yau type. FANO MANIFOLDS OF CALABI-YAU TYPE 7 Proof. Use the conormal sequence of the pair (X, Σ). Its p-th wedge power is the complex 0 → L−p X → · · · → Ωp−1 Under hypothesis (1), for p ≤ ι − 2, we have Σ ⊗ L−p+1 X → Ω1 Σ ⊗ LX → Ωp ΣX → Ωp X → 0. H dX −p+k(X, Ωp−k Σ ⊗ L−k X ) = 0 for all k from 0 to p. Our first claim follows. For p = ι − 1, our hypothesis imply that H dX −p+k+ǫ(X, Ωp−k Σ ⊗ L−k X ) = 0 for ǫ ∈ {0, 1}. But then, it follows that H ι+2(X, Ωι−1 X ) = H dX (X, L−ι+1 X ) = H dX (X, KX ) = C, which was our second claim. Finally, for p = ι the hypothesis we made are such that we get a complex 0 → H ι+1(X, Ωι X ) → H dX (X, L−ι X ) → H dX (X, Ω1 Σ ⊗ L−ι+1 X ). Using Serre duality, we get the dual complex H 0(X, T ΣX ) → H 0(X, L) → H ι(X, Ωι+1 X ) → 0. But the cokernel of the first arrow is precisely H 1(X, T X), at least if we have H 1(X, T ΣX ) = 0 or, by Serre duality again, H dX −1(X, ΩdX X) = 0. This follows from (3). (cid:3) Σ ⊗ L−ι Of course we can easily imagine variants if this lemma for complete in- tersections, or even more generally, for zero-loci of global sections of vector bundles. For the construction to work, we need a huge number of vanishing conditions on the ambient variety Σ, and it is not so clear a priori that we can find any example that way. We have been able to find some under the hypothesis that Σ is a homogeneous space. 3.3. Linear sections of homogeneous spaces. Suppose that Σ is a ratio- nal homogeneous variety, and that ι = ιΣ is equal to its index. In particular we require that the dimension of Σ is dΣ = 2ιΣ + 2. There is a short list of suitable examples: Σ (P1)6 (P1)3 × P3 (P2)4 (P4)3 G(2, 5) × G(2, 5) G(4, 9) G(3, 11) dim index moduli 6 6 8 12 12 20 24 2 2 3 5 5 9 11 45 55 48 52 51 46 45 8 A. ILIEV AND L. MANIVEL Proposition 3.2. Let Σ belong to the list above, and X = Σ∩H be a smooth hyperplane section of Σ in its minimal complete embedding. Then X is of Calabi-Yau type. Proof. We check case by case that the vanishing conditions of Lemma 3.1 do hold. For Grassmannians this requires the diagrammatic methods first developed in [Sn]. (cid:3) 3.4. Quadratic sections of homogeneous spaces. Now suppose that Σ is again a rational homogeneous variety of even index ιΣ, and let ι = ιΣ/2. The relation dΣ = 2ι + 2 becomes ιΣ = dΣ − 2, which is the definition of the Mukai varieties. It has been proved that apart from complete intersections, Mukai varieties are linear sections of homogeneous varieties (note that a linear section of a Mukai variety, if Fano, is again a Mukai variety), or quadric sections of the cone over G(2, 5). The homogeneous Mukai varieties are the following: Σ dim index moduli (P1)4 Gad 2 P3 × P3 LG(3, 6) IG(2, 6) G(2, 6) S10 4 5 6 6 7 8 10 2 3 4 4 5 6 8 68 62 69 62 68 69 80 In this table we have denoted by IG(k, 2n) the symplectic Grassman- nian, parametrizing isotropic k-dimensional subspaces of a 2n-dimensional vector space endowed with a non-degenerate skew-symmetric form. When k = n we rather use the notation LG(n, 2n) and call this variety the La- grangian Grassmannian. Similarly, S2n denotes the spinor variety, which parametrizes one of the two families of isotropic n-dimensional subspaces of a 2n-dimensional vector space endowed with a non-degenerate quadratic form. The terminology comes from the fact that the minimal equivariant embedding of this variety is inside a projectivized half-spin representation -- a quadratic Veronese is required in order to recover the Plucker embed- ding. Finally, the adjoint variety Gad 2 denotes the closed G2-orbit inside the projectivized adjoint representation. Proposition 3.3. Let X = Σ ∩ Q be a smooth quadric section of a homo- geneous Mukai variety of even dimension. Then X is of Calabi-Yau type. Proof. Again we need to check, case by case, the conditions of Lemma 3.1. Let us treat in some detail the case of S10, which is the most complicated one. Here X is nine dimensional, and the claim is that its middle dimensional Hodge numbers are h9,0 = h8,1 = h7,2 = 0, h6,3 = 1, h5,4 = 80. FANO MANIFOLDS OF CALABI-YAU TYPE 9 Let E denote the tautological rank five vector bundle on Σ = S10. The cotangent bundle of Σ is Ω1 Σ of p-forms can have several components, which are Schur powers of E obtained as follows. Let a = (a0 > a1 > · · · > am) be a decreasing sequence of integers, whose sum is equal to p. Define the sequence Σ ≃ ∧2E. In higher degree, the bundle Ωp λ(a) = (a0, a1 + 1, . . . , am + m, mam−1−am−1, . . . , 1a0−a1−1). Then by the results of [Ko], Ωp Σ = ∧p(∧2E) is the direct sum of the Schur powers Sλ(a)E. Note that since E has rank five, the number of non zero components of λ(a) must not exceed five, which means that a0 − am ≤ 4. Bott's theorem allows to compute the cohomology groups of a Schur power SλE of the tautological bundle, for λ = (λ1, . . . , λ5) a non increasing se- quence of non negative numbers. The rule is the following. Suppose that there exists a pair (i, j) of distinct indices such that λi + λj = i + j − 2. Then SλE is acyclic. Otherwise, let q be the number of pairs such that λi + λj > i + j − 2. Then SλE has a unique non-zero cohomology group, that of degree q. Applying these rules, it is easy to check the following statement. Lemma 3.4. The following bundles of twisted forms on Σ are acyclic: (1) Ω1 (2) Ω2 (3) Ω3 (4) Ω4 Σ(−k) for k = 1, 2, 3. Σ(−k) for k = 1, 2, 3. Σ(−k) for k = 1, 2. Σ(−k) for k = 1, 2. Moreover, for k ≥ 4, Ω1 only in degree dim Σ. And the same is true for Ω3 k ≥ 3. Σ(−k) and Ω2 Σ(−k) have non-zero cohomology groups Σ(−k) when Σ(−k) and Ω4 This is precisely what we need in order to apply Lemma 3.1. (cid:3) 3.5. A quasi-homogeneous Mukai variety. Consider a hyperplane sec- tion Θ of the Lagrangian Grassmannian LG(3, 6). This is a five dimensional variety of index three. Proposition 3.5. The variety Θ has the following properties: (1) it is a quasi-homogeneous P SL3-variety; (2) its non-zero Hodge numbers are hp,p(Θ) = 1, 1 ≤ p ≤ 5; (3) it is a minimal compactification of C5; (4) it is rigid. Proof. Consider the Lagrangian Grassmannian LG(3, 6) in its minimal ho- mogeneous embedding PV = P13. The orbit structure of Sp6 in PV is well-known, in particular it is prehomogeneous: there exists an open orbit, and the generic stabilizer is P SL3 (up to a finite group). Since V is self- dual, we conclude that the general hyperplane section Θ is rigid and admits a P SL3-action. 10 A. ILIEV AND L. MANIVEL An explicit computation shows that P SL3 has four orbits in Θ. This can be seen as follows. The Lagrangian Grassmannian LG(3, 6) is known to be one of the few homogeneous varieties with one apparent double point. This means that a general point of PV belongs to a unique secant xy, where x and y represent two generic isotropic planes. The orbits in Θ are then determined by the relative position with respect to x and y, that is, by the dimensions of these intersections. These can be (0, 0) (the generic case, giving the open orbit), (1, 1) (a hyperplane section), (2, 1) or (1, 2) (two Veronese surfaces). The computation of the Hodge numbers is straightforward. Finally, it is easy to check that a maximal torus in P SL3 has a finite number of fixed points in Θ. The Byalinicki-Birula decomposition [BB] then ensures that Θ is a compactification of C5, and it is clearly minimal. (cid:3) 3.6. Double coverings. Consider a double covering Y → Σ, branched over a smooth hypersurface X of Σ, of degree 2d. The Hodge numbers of Y can be computed in terms of the pair (Σ, X) as follows [Cy]: hp,q(Y ) = hp,q(Σ) + hq(Ωp Σ(log X)(−d)). Recall that Ωp Σ with simple poles at most on X. There is an exact sequence Σ(log X) denotes the vector bundle of logarithmic p-forms on Σ → Ωp defined by taking residues on X. 0 → Ωp Σ(log X) → Ωp−1 X → 0 Σ(−d)) = H q+1(Σ, Ωp Suppose that Σ is homogeneous of odd dimension n = 2m + 1. In partic- ular the Hodge structure of Σ is pure and we get hp,q(Σ) = 0 for p + q = n. Suppose moreover that H q(Σ, Ωp Σ(−d)) = 0. Then hp,q(Y ) = hq(X, Ωp−1 X (−d)), which can be computed by considering a long exact sequence like in the proof of Lemma 3.1. Under favorable circum- stances, we expect to derive that hm+2,m−1(Y ) = hm+2(X, Ωm−2 X (−d)) = h2m(X, OX (−d − (m − 2)2d)) = C if moreover ωX = OX (−(2m − 3)d), that is, ωΣ = OΣ(−(2m − 1)d). If we let d = 1, this means that Σ is a Mukai variety. If we let Σ be homogeneous, or the quasi-homogeneous Mukai variety of the previous section, we check that the expected vanishing theorems do hold and we arrive at the following conclusion. Proposition 3.6. Let Y → Σ be a double covering of a homogeneous or quasi-homogeneous Mukai variety of odd dimension, branched over a smooth quadric section. Then Y is a Fano manifold of Calabi-Yau type. 3.7. The Cayley trick. Let X ⊂ PN be a complete intersection of multide- gree (d1, . . . , dm), where none of these degrees equals one. Let Li = OPN (di) and P = P(E∨), where E = L1 ⊕ · · · ⊕ Lm. The equations of X can be con- sidered as a section σ in H 0(PN , E) ≃ H 0(P, OE(1)). Hence an associated divisor X ⊂ P , which is smooth when X is smooth. By [Na, Lemma 2.7], FANO MANIFOLDS OF CALABI-YAU TYPE 11 the cohomology of X is essentially the same as that of X, with a shift by 2m − 2 = dim X − dim X in the degrees. Moreover the deformations of X and X are both unobstructed, and (1) H 1(X, T X) ≃ H 1(X , T X ). Indeed, the left hand side of (1) is given by the exact sequence X ) → H 0(X, EX ) → H 1(X, T X) → 0. H 0(X, T PN It is easy to check that H 0(X, T PN X ) ≃ H 0(PN , T PN ) = slN . Moreover, the Koszul complex of the section σ of E defining X shows that H 0(X, EX ) is just the cokernel of the natural map H 0(PN , End(E)) → H 0(PN , E) defined by applying an endomorphism of E to the section σ. Similarly, the right hand side of (1) is given by an exact sequence H 0(X , T PX ) → H 0(X , OX (1)) → H 1(X , T X ) → 0. We have H 0(X , OX (1)) ≃ H 0(P, OE (1))/Cσ = H 0(PN , E)/Cσ. Moreover, H 0(X , T PX ) = H 0(P, T P ) can be computed through the differential of the projection π from P to X and the description of its kernel, the vertical tangent space T vP , by the relative Euler sequence 0 → OP → π∗E∨ ⊗ OP (1) → T vP → 0. We get that H 0(P, T P ) is the direct sum of H 0(P, T vP ) and slN , while H 0(P, T vP ) ≃ H 0(PN , End0(E)). Since the maps H 0(PN , End(E)) → H 0(PN , E) and H 0(PN , End0(E)) → H 0(PN , E)/Cσ have the same cok- ernels, the identification (1) follows. We conclude: Proposition 3.7. If X ⊂ PN is a Calabi-Yau manifold, or a complete intersection of Calabi-Yau type, then X ⊂ P is also of Calabi-Yau type. 4. A special series of manifolds of Calabi-Yau type 4.1. A special non-vanishing. Consider the following series of homoge- neous varieties Σ: Σ OP2 S12 G(2, 10) S14 dim index coindex deg 3 16 4 15 5 16 21 8 12 10 10 12 4 5 6 9 Here we denoted by OP2 the so-called Cayley plane, which is the unique Hermitian symmetric space of type E6. Recall that the minimal represen- tation of E6 has dimension 27. They Cayley plane can be obtained as the minimal E6-orbit in the projectivization of this representation (or its dual). 12 A. ILIEV AND L. MANIVEL For more on the extraordinary properties of this variety, see [IM1] and the references therein. Recall that the index r is defined by the identity KΣ = −rH, where H is the ample generator of the Picard group of Σ. The coindex is the difference between the dimension and the index. The last column gives the degree of the dual variety of Σ. Note that this degree is always one less than the coindex, a coincidence that will play a major role below. Lemma 4.1. Let Σ be as above, with index r and coindex c. Then H r+2(Σ, Ωc−2 Σ (c − r − 1)) = C. Proof. This is a straightforward application of the Borel-Weil-Bott theorem. The details are provided in the Appendix. (cid:3) 4.2. Relation with projective duality. The explanation for the non- vanishing of Lemma 4.1 is the following. We start by recalling the usual setting of projective duality, which we summarize in the diagram p I q ???????? Σ∨ PVΣ ⊃ Σ ⊂ PV ∨ Σ where we denoted by I ⊂ PVΣ × PV ∨ Σ the incidence variety, defined as the set of pairs (x, h) with x ∈ Σ and h a hyperplane containing the affine tangent space to Σ at h. The variety Σ∨ parametrizing tangent hyperplanes to Σ is its projective dual variety. For all the cases we are interested in, the dual variety is non-degenerate, which means that it is a hypersurface whose degree has been computed in the framework of prehomogeneous vector spaces. Indeed, the variety Σ is the closed orbit of the simple group Aut(Σ) in a (projectivized) prehomogeneous space P(VΣ) whose orbit structure is known explicitly. Part of the structure is common to all cases: there is of course an open orbit, its complement is an irreducible hypersurface, and this hypersurface contains an orbit closure WΣ whose complement is again a single orbit. This can be observed case by case. If Σ is the Cayley plane, there are only three orbits in P(VΣ), the variety Σ itself, its complement in its secant variety, which is a cubic hypersurface, and the complement of this hypersurface. In particular WΣ = Σ in that case. An equation of the E6-invariant cubic was first written down by Elie Cartan (in terms of the geometry of the 27 lines on a smooth cubic surface) and we therefore call it the Cartan cubic. If Σ = S12, there are four orbits [Ig], whose closures are the whole projec- tive space, a quartic hypersurface, Σ itself, and an intermediate codimension seven variety which is our WΣ. If Σ = G(2, 10), embedded in the projectivization of the space of skew- symmetric forms, there are five orbits determined by the ranks of the forms; FANO MANIFOLDS OF CALABI-YAU TYPE 13 the invariant hypersurface is defined by the Pfaffian; WΣ is the space of forms of rank six at most, its codimension is six. Finally, if Σ = S14, there are exactly nine orbits which were first described by Popov [Po]. His results show that WΣ has codimension five. In each case we expect WΣ to coincide with the singular locus of the invariant hypersurface. At least we know that it contains this singular locus, just because of the orbit structure. The orbit structure is the same in the dual projective space P(V ∨ Σ ), in particular the invariant hypersurface is the projective dual variety of Σ. We summarize the relevant data in the following statement. We have no explanation, even conjectural, of these curious numerical coincidences. Lemma 4.2. The projective dual Σ∨ to Σ is a hypersurface of degree c − 1, whose singular locus has codimension at least r − c + 2. The following table gives VΣ for each case and the dual hypersurface Σ∨. Σ∨ VΣ H3(O) Cartan cubic ∆+ Igusa quartic Σ OP2 S12 G(2, 10) ∧2C10 Pfaffian quintic S14 ∆+ Popov octic Since the quotient of VΣ by the affine tangent bundle is the twisted normal bundle N (−1), we have a natural identification I ≃ P(N (−1)∨). Moreover this identification is such that the relative tautological bundle ON (−1)(1) coincides with q∗O(1). Now, the duality theorem (see e.g. [GKZ]) asserts that (Σ∨)∨ = Σ, and moreover, that the diagram above is symmetric in the sense that I can be defined as parametrizing the tangent hyperplanes to the hypersurface Σ∨ (at least at smooth points, and then one needs to take the Zariski closure). If we denote by FΣ ∈ Symc−1VΣ an equation of Σ∨, this means that I can be obtained as the closure of the set of points ([dFΣ(ξ)], [ξ]), where ξ belongs to the cone over Σ∨ reg. This means in particular that we have a non-zero section δFΣ ∈ H 0(I, Hom(q∗O(−c + 2), p∗O(−1))), vanishing precisely over the singular locus of Σ∨. Pushing this section for- ward to Σ, we get a section ¯δFΣ ∈ H 0(Σ, Symc−2(N (−1))(−1)) = H 0(Σ, Symc−2N (−c + 1))). Now we can use the normal exact sequence 0 → T Σ → T P(VΣ) ⊗ OΣ → N → 0. Taking a suitable wedge power and twist, we get a long exact sequence 0 → ∧c−2T Σ(−c + 1) → · · · → Symc−2N (−c + 1)) → 0. 14 A. ILIEV AND L. MANIVEL A careful check would then allow to conclude that this long exact sequence induces an isomorphism H 0(Σ, Symc−2N (−c + 1))) ≃ H c−2(Σ, ∧c−2T Σ(−c + 1)), and the latter is Serre dual to H r+2(Σ, Ωc−2 Σ (c − r − 1)). In particular, we see that that our non-zero section ¯δFΣ induces a non-zero linear form on this one dimensional cohomology group. 4.3. Linear sections. Let X be a general linear section of Σ, of codimen- sion s = r − c + 1. Then X has dimension n = 2c − 1 and index c − 1. The following statement is a special case of a result of Borcea [Bo]: Lemma 4.3. Any small deformation of X is a linear section of Σ. In particular, the number of moduli for X is the dimension m of the quotient of the Grassmannian G(s, VΣ) by the simple group Aut(Σ), of di- mension δ. The relevant data appear in the table below. Σ OP2 S12 s N Aut(Σ) 9 26 6 31 Spin12 G(2, 10) 5 44 P SL10 4 63 Spin14 S14 E6 δ m 84 78 66 90 99 101 91 149 Theorem 4.4. A generic linear section X of Σ has primitive cohomology only in middle degree. Its non-zero Hodge numbers hp,n−p, for p ≥ c, are the following: hc+1,c−2 = 1, hc,c−1 = m. Proof. Let us denote by L ⊂ V ∨ Its dimension is s = r − c + 1. We use the conormal sequence Σ the space of linear forms that defines X. 0 → OX (−1) ⊗ L → Ω1 ΣX → Ω1 X → 0, and its wedge powers, for positive integers p, 0 → OX (−p) ⊗ SympL → · · · → Ωp−1 ΣX(−1) ⊗ L → Ωp We claim that for p ≤ c − 1, this induces an isomorphism ΣX → Ωp X → 0. H q(X, Ωp ΣX ) ≃ H q(X, Ωp X ). To check this, we will prove that the other vector bundles involved in the long exact sequence above, are all acyclic. These bundles are the bundles of twisted forms Ωp−k ΣX(−k), for 1 ≤ k ≤ p. For example, for k = p we observe that H q(X, OX (−p)) = 0 ∀q ≤ n − 1 by Kodaira's vanishing theorem, while by Serre duality H q(X, OX (−p)) = H n−q(X, OX (p − c + 1))∨ = 0 ∀q ≥ 1 FANO MANIFOLDS OF CALABI-YAU TYPE 15 when p ≤ c − 2. For the remaining terms, we use the Koszul resolution of the structure sheaf of X, 0 → OΣ(c − r − 1) ⊗ ∧r−c+1L → · · · → OΣ(−1) ⊗ L → OΣ → OX → 0. Twisting this long exact sequence by Ωp−k prove the acyclicity of Ωp−k Ωp−k Σ (−k), we see that in order to ΣX(−k), it is enough to check the acyclicity of Σ (−k − ℓ) for 0 ≤ ℓ ≤ r − c + 1. But this follows from Lemma 5.1. By the same Lemma, Ωp Σ(−ℓ) is acyclic for 1 ≤ ℓ ≤ r − c + 1 and p ≤ c − 3. This implies that H q(Σ, Ωp Σ) ≃ H q(X, Ωp ΣX ) ≃ H q(X, Ωp X ) for all q and for all p ≤ c − 2. For p = c−2 there is a unique non acyclic bundle involved, Ωc−2 Σ (c−r−1): by Lemma 4.1, this bundle has a one-dimensional cohomology group in degree r + 2, but the other cohomology groups vanish. Then the Koszul resolution of OX implies that H q(X, Ωc−2 ΣX)0 = H q+r−c+1(Σ, Ωc−2 Σ (c − r − 1)) = δq,c+1C. Finally, for p = c − 1 we find two non-zero cohomology groups for the bundles involved in the complex 0 → OX (1 − c) ⊗ Symc−1L → · · · → Ωc−2 ΣX(−1) ⊗ L → Ωc−1 ΣX → Ωc−1 X → 0. Indeed, by Lemma 5.1 Ωc−2 as above, the cohomology of the bundle Ωc−2 Σ (c − r) is acyclic. Then, by the same argument ΣX(−1) is H q(X, Ωc−2 ΣX (−1)) = H q+r−c(Σ, Ωc−2 Σ (c − r − 1) ⊗ ∧r−cL) = δq,c+2L∨. On the other hand, the line bundle OX(1 − c) is the canonical bundle of X, and therefore it has non trivial cohomology in degree n = 2c − 1, and only in that degree. Note that Ωc−1 ΣX also has non-zero cohomology, but only in degree c − 1, fo which we get the same cohomology group as for Ωc−1 Σ . We deduce a long exact sequence · · · → H q(X, Ωc−1 X )0 → H q+c−1(X, ωX ) ⊗ Symc−1L → → H q+2(X, Ωc−2 ΣX (−1)) ⊗ L → H q+1(X, Ωc−1 X )0 → · · · But note that this long exact sequence has very few non-zero terms, appear- ing precisely for q = c. Note also that by Hodge symmetry, we already know that H c+1(X, Ωc−1 X ) = 0. What remains is the short exact sequence X ) = H c(X, Ωc−2 0 → H c(X, Ωc−1 (2) from which we can easily compute hc,c−1. And we conclude, as claimed, that hc,c−1 = m. (cid:3) X ) → Symc−1L → L∨ ⊗ L → 0, Remark. In general, if Σ has a non degenerate dual of degree d ≤ n, we could expect H n+1−d(Σ, Ωd−1 Σ (d − r)) to be non-zero, and even one-dimensional. 16 A. ILIEV AND L. MANIVEL Under favorable circumstances, this should imply that if X is a general linear section of codimension r − d, hence dimension c + d, then hc+1(Ωd−1 X ) = 1. 4.4. Deformations. Observe that the conclusion of Theorem 4.4 has been obtained by a case by case computation, as a coincidence between two num- bers that were computed from quite different, and at first sight, unrelated data. On one hand, the number m of moduli has been computed as the dif- ference between the dimension of a suitable Grassmannian and that of the Indeed, H 1(X, T X) is given by the exact se- automorphism group of Σ. quence 0 → H 0(X, T ΣX ) → H 0(X, L∨(1)) → H 1(X, T X) → 0. It is straightforward to check that H 0(X, T ΣX ) = H 0(Σ, T Σ) = aut(Σ), the Lie algebra of Aut(Σ), while H 0(X, L∨(1)) = L∨ ⊗ V ∨ Σ /L is the tangent space to the Grassmannian of subspaces of VΣ, at its point defined by L. This tangent space is also the quotient of gl(VΣ) by the stabilizer of L, hence the identification H 1(X, T X) ≃ gl(VΣ)/(aut(Σ) + stab(L)). On the other hand, the exact sequence (2) shows that X )∨ ≃ Symc−1L∨/gl(L). H c(X, Ωc−1 As we have seen, the connection between these two descriptions is provided by the equation FΣ, of degree c − 1, of the dual hypersurface Σ∨. We have a natural map from gl(VΣ) to Symc−1VΣ, sending u to FΣ ◦ u, then to Symc−1L∨ by restriction to L. Since aut(Σ) kills FΣ, this induces a map (3) gl(VΣ)/(aut(Σ) + stab(L)) → Symc−1L∨/gl(L), which is an isomorphism if and only if X is of Calabi-Yau type. This has a modular interpretation: we can associate to X the hypersur- face X ∨ = PL ∩ Σ∨ (which is of course not the projective dual variety of X). By Lemma 4.2, this hypersurface will be smooth of degree c − 1 for a generic L. Moreover, if X ∨ has discrete automorphism group, the quotient Symc−1L∨/gl(L) can be interpreted as H 1(X ∨, T X ∨), Otherwise said, the correspondence between X and X ∨ defines a map from the moduli space of linear sections of Σ to the moduli space of degree c − 1 hypersurfaces in Pr−c, and the morphism (3) is nothing else than the differential of that map. We deduce: Proposition 4.5. The variety X is a Fano manifold of Calabi-Yau type if and only if its small deformations induce a versal family of degree c − 1 hypersurfaces in Pr−c. FANO MANIFOLDS OF CALABI-YAU TYPE 17 Note that if this is true, we can conclude that a generic hypersurface of degree c − 1 in Pr−c can be defined, up to isomorphism, as a linear section of Σ∨ in a finite (non-zero) number of different ways. For the quintic threefold, it has already been noticed by Beauville that there exists finitely many Pfaffian representations (see Schreyer's Appendix to [Be]). Beauville asked how many such representations do exist : in prin- ciple the answer is given by some Donaldson-Thomas invariant. We can ask the same question in our three other cases. The case of cubic sevenfolds is treated in [IM2] where we prove that the very same phenomena hold. That's also what we expect for the two remaining cases, which will deserve further investigations. It is also quite remarkable that we can associate to X either a Calabi- Yau threefold, or another Fano manifold of Calabi-Yau type, by considering either the dual X ∨, or a double covering of Pr−c, branched over X ∨. We get the following varieties X ∗: X ∗ cubic sevenfold double quartic fivefold X OP2 ∩ P17 S12 ∩ P25 G(2, 10) ∩ P40 quintic threefold S14 ∩ P59 double octic threefold The last two varieties are famous examples of Calabi-Yau threefolds. The first two were among our very first examples of Fano manifolds of Calabi-Yau type. They are not Calabi-Yau's stricto sensu but, considered as (compact- ified) Landau-Ginzburg models, we have seen that they appeared in the litterature as mirrors to certain rigid Calabi-Yau threefolds [CDP, Sch]. Note that if the resulting map KX → KX ∗ from the base of the Kuranishi family of X, to that of X ∗, is etale, then we have the same property for the gauged families and we can pull-back the integrable system in intermediate Jacobians of X ∗ and its deformations (see Theorem 2.2). Although we did not check this, it is quite likely that we should recover the integrable system in intermediate Jacobians of X itself and its deformations. In particular the intermediate Jacobians of X and X ∗ should be isomorphic. 4.5. Homological projective duality. We expect that our series of ex- amples of Fano manifolds of Calabi-Yau type should give rise to interest- ing instances of homological projective duality [Ku2]. More precisely, the derived categories of X and X ∗ should both contain a three-dimensional Calabi-Yau category, say AX ⊂ Db(Coh(X)) and AX ∗ ⊂ Db(Coh(X ∗)) and there should exist a natural equivalence between AX and AX ∗. For X ∗ this was observed by Kuznetsov: Proposition 4.6. Let Y be among a the following types of Fano manifolds of Calabi-Yau type: (1) a cubic sevenfold, 18 A. ILIEV AND L. MANIVEL (2) a cubic hypersurface in a six-dimensional quadric, (3) a double-cover of P5 branched over a smmoth quartic. Then the derived category Db(Y ) contains a natural subcategory AY which is a three-dimensional Calabi-Yau category. Proof. For any Fano hypersurface Y ⊂ Pn+1 of degree d, and index ιY = n + 2 − d, the collection hOY , . . . , OY (ιY − 1)i is obviously exceptional. Kuznetsov proved in [Ku1], Corollary 4.3, that if d divides n + 2, its left orthogonal AY is a Calabi-Yau category of dimension n − 2ιY /d. In particular, for d = 3 and n = 7, we can conclude that AY is a three-dimensional Calabi-Yau category. A very similar statement holds for double covers. Suppose that Y is a double-cover of Pn branched over a general hypersurface of degree 2d. Then Y is Fano for d ≤ n, and ω−1 Y = f ∗OPn(n + 1 − d), so that ιY = n + 1 − d. Here, for the same reason as before, the collection hOY , f ∗OPn(1), . . . , f ∗OPn(ιY − 1)i is exceptional. Kuznetsov proved that its orthogonal is, when d divides n+1, a Calabi-Yau category of dimension n − ιY /d. For d = 2 and n = 5, we can conclude that AY is a three-dimensional Calabi-Yau category. Finally the case of hypersurfaces in quadrics in very similar. In dimension 6, the derived category has a Lefschetz decomposition Db(Q6) = hA, A(3)i, where A = hS, OQ6 , OQ6(1), OQ6 (2)i, If i : Y ֒→ Q6 is a cubic hy- and S denotes the rank four spin bundle. persurface, the functor i∗ : Db(Q6) → Db(Y ) is fully faithful on A and the orthogonal AY to i∗(A) in Db(Y ) is a three-dimensional Calabi-Yau cate- gory. (cid:3) The appearance of these three dimensional non-commutative Calabi-Yau's in connection with our manifolds of Calabi-Yau type (but not three dimen- sional!) is a quite remarkable phenomenon, which would deserve to be better understood. 5. Appendix: Proof of Lemma 4.1 First observe that, Σ being a Hermitian symmetric space, its cotangent bundle is an irreducible homogeneous bundle. In particular its cohomology groups, and that of its twists as well, can be computed using Bott's theorem. Moreover the same is true for any bundle of p-forms on Σ, since it must split into a direct sum of irreducible homogeneous bundles. Recall that an irreducible homogeneous vector bundle on a homogeneous space Σ = G/P is defined by an irreducible representation of P , which is in turn defined by its highest weight: some weight λ of G which is P -dominant. We denote the corresponding bundle by Eλ. FANO MANIFOLDS OF CALABI-YAU TYPE 19 When Σ is a Hermitian symmetric space with Picard number one, P is a maximal parabolic subgroup defined by some simple root α, and the cotangent bundle is simply Ω1 Σ = E−α. More generally, a combinatorial formula has been obtained by Kostant for the irreducible components of each Ωp Σ [Ko]. Bott's theorem for an irreducible homogeneous vector bundle Eλ can be stated as the following recipe. Add ρ, the half-sum of the positive roots of G, to λ. If hλ + ρ, α∨i = 0 for some coroot α∨, then Eλ is acyclic. Otherwise, there is a unique w in the Weyl group W of G such that w(λ + ρ) = µ + ρ for some dominant weight µ. Then Eλ has a unique non-zero cohomology group, in degree ℓ(w), and it is an irreducible G-module of lowest weight −µ. In particular, it is one dimensional if and only if µ = 0. First case : Σ = G(2, 10). Here Ω1 Σ = Q∗ ⊗ E, where E denotes the tautological rank two bundle and Q the quotient rank eight bundle. The claim is that H 12(Σ, Ω4 Σ(−5)) = C. The decomposition of Ω4 Σ into irreducible components is Ω4 Σ = ∧4Q∗ ⊗ S4E ⊕ S211Q∗ ⊗ S31E ⊕ S22Q∗ ⊗ S22E. We apply Bott's theorem to each factor. The weights of the three factors are (0, 0, 0, 0, −1, −1, −1, −1, 4, 0), (0, 0, 0, 0, 0, −1, −1, −2, 3, 1), (0, 0, 0, 0, 0, 0, −2, −2, 2, 2). Twisting by OΣ(−5) and adding ρ we get (4, 3, 2, 1, −1, −2, −3, −4, 5, 0), (4, 3, 2, 1, 0, −2, −3, −5, 4, 1), (4, 3, 2, 1, 0, −1, −3, −4, 3, 2). The last two sequences have repeated entries, indicating that the corre- sponding bundles are acyclic. But the first one has pairwise distinct entries, forming, once ordered, a consecutive sequence. This indicates that the cor- responding bundle has one non-zero cohomology group, of dimension one, appearing in degree equal to the number of inversions of the sequence, which is twelve. This proves the claim. Second case : Σ = S12. Here Ω1 rank six bundle. The claim is that Σ = ∧2E, where E denotes the tautological H 12(Σ, Ω3 Σ(−6)) = C. The decomposition of Ω3 Σ into irreducible components is Ω3 Σ = S411100E ⊕ S222000E. Twisting by OΣ(−6) and adding ρ = (5, 4, 3, 2, 1, 0) we get the weights (2, 1, 0, −3, −4, −5) and (2, 0, −1, −2, −3, −7). The last sequence has two 20 A. ILIEV AND L. MANIVEL opposite entries, indicating that the corresponding bundle is acyclic. But this is not the case of the first sequence, whose absolute values form a consecutive sequence starting from zero. This implies that the corresponding bundle has a unique non-zero cohomology group, which is one dimensional. It appears in degree equal to the number of pairs of entries of the sequence whose sum is negative; this is twelve and the claim follows. Third case : Σ = S14. Here again Ω1 logical rank seven bundle. The claim is that Σ = ∧2E, where E denotes the tauto- H 14(Σ, Ω7 Σ(−4)) = C. The decomposition of Ω7 Σ into irreducible components is Ω7 Σ = S6221111E ⊕ S5322110E ⊕ S4333100E ⊕ S4422200E. Twisting by OΣ(−7) and adding ρ = (6, 5, 4, 3, 2, 1, 0) we get the weights (3, 2, 1, 0, −2, −3, −8), (4, 3, 1, −2, −3, −4, −6), (4, 2, 1, −1, −2, −4, −7)), (4, 3, 0, −1, −2, −5, −6). The first three have pairs of opposite entries, which implies that the cor- responding bundles are acyclic. This is not the case of the last sequence, whose absolute values form a consecutive sequence starting from zero. As in the previous case, this implies the claim. Fourth case : Σ = OP2. For convenience we encode each irreducible vec- tor bundle Eλ on the Cayley plane OP2 by a weighted Dynkin diagram of type E6, whose weights are the coordinates of λ on the basis of fundamen- tal weights (themselves in natural bijection with the nodes of the Dynkin diagram). For example we have Ω1 Σ ≃ −21000 0 , Ω2 Σ ≃ −30100 0 . The claim to be checked is that H 14(Σ, Ω2 Σ(−9)) = C. The root system, and the Weyl group W of E6 being rather complicated, we will use a different strategy than in the other case. If λ = −3ω1 + ω4 denotes the highest weight of Ω2 Σ, what we need to check is that µ = λ − 9ω1 + ρ belongs to the W -orbit of ρ. Recall that ρ, the half sum of the positive roots, is also the sum of the fundamental weights. We can thus write ρ = 11111 1 , µ = −B1211 1 , where B stands for eleven (and A will stand for ten, below). Now we can let W act through the simple reflections si associated to the simple root αi. The recipe for the action of si on a weight ω, written in the basis of fundamental weights, is as follows: add the i-th coordinate to the j-th if j is connected to i in the Dynkin diagram; then change the ith coordinate into FANO MANIFOLDS OF CALABI-YAU TYPE 21 its opposite. Here we start with a weight having a negative coordinate. Ap- plying successively simple reflections corresponding to nodes with negative coordinates, we will end up with a dominant weight. Starting from µ, the game goes as follows. −B1211 1 → B−A211 1 → 1A−811 1 → 128−71 −7 → 121−71 7 → → → 12−67−6 7 → 12−616 7 → 1−46−56 1 → 1−4151 1 → −34−351 1 → 31−351 1 → 3−2321 −2 → 3−2121 2 → 12−121 2 → 11111 1 So we end up with ρ after 14 operations. This indicates that we get a one dimensional cohomology group in degree fourteen. The claim is proved. (cid:3) With the very same method we can prove that the one dimensional coho- mology group whose existence is asserted by Lemma 4.1, is the only non-zero cohomology group of twisted form in a wide range. The precise statement is the following: Lemma 5.1. Suppose that p ≤ c − 1 and 1 ≤ k ≤ r − p. Then H q(Σ, Ωp Σ(−k)) = δq,r+2δp,c−2δk,r−c+1C. References [Be] Beauville A.: Determinantal hypersurfaces, Michigan Math. J. 48 (2000), 39 -- 64. [BB] Bialynicki-Birula A.: Some theorems on actions of algebraic groups, Annals of Math. 98 (1973), 480-497. [Bo] Borcea C.: Smooth global complete intersections in certain compact homogeneous complex manifolds, J. Reine Angew. Math. 344 (1983), 65 -- 70. [CDP] Candelas P., Derrick E., Parkes L., Generalized Calabi-Yau Manifolds and the Mirror of a Rigid Manifold, Nucl.Phys. B 407 (1993), 115-154. [Cy] Cynk S. Cohomologies of a double covering of a non-singular algebraic 3-fold, Math. Z. 240 (2002), 731 -- 743. [DM] Donagi R., Markman E.: Cubics, integrable systems, and Calabi-Yau threefolds, in Proceedings of the Hirzebruch 65 Conference on Algebraic Geometry, 199 -- 221, Israel Math. Conf. Proc. 9, 1996. [Fr] Freed D., Special Kahler manifolds, Comm. Math. Phys. 203 (1999), 31-52. [GKZ] Gelfand I.M., Kapranov M.M., Zelevinsky A.V., Discriminants, resultants, and multidimensional determinants. Birkhauser 1994. [Gr] Griffiths, P.A., On the periods of certain rational integrals I, II, Annals of Math. 90 [Ig] [IK] (1969), 460-495, 496-541. Igusa J., A classification of spinors up to dimension twelve, Amer. J. Math. 92 (1970), 997-1028. Iliev A., Katzarkov L., On the Griffiths group of Fano manifolds of Calabi-Yau Hodge type, in preparation. [IM1] Iliev A., Manivel L., The Chow ring of the Cayley plane, Compos. Math. 141 (2005), 146-160. [IM2] Iliev A., Manivel L., On cubic hypersurfaces of dimension seven and eight, preprint 2011. 22 A. ILIEV AND L. MANIVEL [Ko] Kostant B., Lie algebra cohomology and generalized Schubert cells, Annals of Math. 77 (1963), 72-144. [Ku1] Kuznetsov A., Derived categories of cubic and V14 threefolds, Proc. Steklov Inst. Math. 246 (2004), 171 -- 194. [Ku2] Kuznetsov A., Homological projective duality, Publ. Math. Inst. Hautes tudes Sci. 105 (2007), 157 -- 220. [Ma] Manivel L., On the derived category of the Cayley plane, arXiv:math.AG/0907.2784, to appear in the Journal of Algebra. [Na] Nagel J., The Abel-Jacobi map for complete intersections. Indag. Math. (N.S.) 8 (1997), no. 1, 95 -- 113. [Po] Popov V.L., Classification of the spinors of dimension fourteen, Uspehi Mat. Nauk 32 (1977), 199-200. [Sch] Schimmrigk R.: Mirror Symmetry and String Vacua from a Special Class of Fano Varieties, Int. J. Mod. Phys. A 11 (1996), 3049-3096. [Sn] Snow D.: Cohomology of twisted holomorphic forms on Grassmann manifolds and quadric hypersurfaces, Math. Ann. 276 (1986), 159 -- 176. [St] Steenbrink J.: Intersection form for quasi-homogeneous singularities, Compositio Math. 34 (1977), 211 -- 223. [Vo] Voisin C.: Sym´etrie Miroir, Panoramas et Synth`eses 2, SMF 1996. Department of Mathematics, Seoul National University, Seoul 151-747, Ko- rea E-mail address: [email protected] Institut Fourier, Universit´e de Grenoble et CNRS, BP 74, 38402 Saint- Martin d'H`eres, France E-mail address: [email protected]
0704.2744	2	0704	2011-09-02T15:32:04	Nahm transform and parabolic minimal Laplace transform	[ "math.AG" ]	We prove that Nahm transform for integrable connections with a finite number of regular singularities and an irregular singularity of rank 1 on the Riemann sphere is equivalent -- up to considering integrable connections as holonomic $\D$-modules -- to minimal Laplace transform. We assume semi-simplicity and resonance-freeness conditions, and we work in the framework of objects with a parabolic structure. In particular, we describe the definition of the parabolic version of Laplace transform due to C. Sabbah. The proof of the main result relies on the study of a twisted de Rham complex.	math.AG	math	Nahm transform and parabolic minimal Laplace transform Szil´ard Szab´o ∗ November 16, 2018 Contents 1 Parabolic connections and Nahm transform 2 Laplace transform without parabolic structure 2.1 Interpretation as a cokernel . . . . . . . . . . . . . . . . . . . . . 3 Meromorphic connections and D-modules 3.1 Regular singularities . . . . . . . . . . . . . . . . . . . . . . . . . 3.2 Irregular singularity . . . . . . . . . . . . . . . . . . . . . . . . . 3.3 Stability . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4 Minimal parabolic Laplace transform 4.1 Parabolic transform . . . . . . . . . . . . . . . . . . . . . . . . . 5 Outline of the proof 6 Proof of the propositions 6.1 Proposition 5.1 . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6.2 Proposition 5.2 . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6.3 Proposition 5.3 . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6.4 Proposition 5.4 . . . . . . . . . . . . . . . . . . . . . . . . . . . . Abstract We prove that Nahm transform for integrable connections with a finite number of regular singularities and an irregular singularity of rank 1 on the Riemann sphere is equivalent -- up to considering integrable connections as holonomic D-modules -- to minimal Laplace transform. We assume semi-simplicity and resonance-freeness conditions, and we work in the 3 9 10 11 11 13 14 14 16 19 21 21 22 27 28 ∗Department of Geometry, Mathematical Institute, Faculty of Science, Budapest Uni- versity of Technology and Economics, Egry J. u. 1, H ´ep., H-1111 Budapest, Hungary, [email protected] 1 framework of objects with a parabolic structure. In particular, we describe the definition of the parabolic version of Laplace transform due to C. Sabbah. The proof of the main result relies on the study of a twisted de Rham complex. Introduction Nahm transform is a correspondence in the theory of 4-manifolds, between so- lutions of the anti-self-dual (ASD) Yang-Mills equations on R4, invariant by a closed additive subgroup and with finite energy on the quotient on the one hand, and solutions of the ASD equations on the dual vector space (R4)∗, in- variant with respect to the dual subgroup and with finite energy on the quotient on the other hand. It can be considered as a differential geometric variant of Fourier-Mukai transform [1]. When the first subgroup is R2, and after identification of the quotient R2 with C, invariant solutions of the ASD equations are holomorphic connections together with a harmonic metric (see [5]). On the other hand, the dual quotient sion. Hence, Nahm transform maps holomorphic connections with a harmonic is then again a copy of C, that we shall denote by bC in this text to avoid confu- metric on C into holomorphic connections with a harmonic metric on bC. In [15], we describe this transform for holomorphic connections on the Riemann sphere CP1 with a finite number of regular singularities in C, and an irregular singu- larity of Poincar´e rank 1 at infinity, and endowed with a parabolic structure in all singular points. It is well-known that holomorphic connections can be viewed as a special case of holonomic D-modules of finite rank. Furthermore, this cor- respondence is compatible with parabolic structures. On the other hand, there exists a so-called minimal (Fourier-)Laplace transform for D-modules without a parabolic structure (see e.g. [10]). Claude Sabbah extends in [11] this construc- tion to the case with parabolic structure in the singularities. We shall review this extension in Section 4. He also asked whether Nahm transform and parabolic minimal Laplace transform agree under the above-mentioned equivalence of cat- egories. Our main result is an affirmative answer to this question (see Theorem 4.1). This result allows us to give a complex algebraic definition of Nahm trans- form, which is initially defined using L2-theory. An immediate consequence is that Nahm transform is a holomorphic map between moduli spaces of stable holomorphic connections. This is one step in showing that Nahm transform is actually a hyper-Kahler isomorphism between moduli spaces. A few words are due here to explain our assumptions on the singularities (regular singularities except for one irregular singularity of Poincar´e rank 1 at infinity). At first sight this might look an artificial choice; however, it turns out that this class of connections is preserved by our transformation, hence it is natural to study the properties of this correpondence. More importantly, this class of connections appears naturally in the theory of Frobenius manifolds [12] as well as in Mirror Symmetry [8]. Notice however that Laplace transform is defined not only under these assumptions; indeed, in a forthcoming paper [3] 2 we plan to extend the definition of Nahm transform for integrable connections with harmonic metric on the Riemann sphere to a more general setup, and we hope that the present results can be extended to that situation too. Acknowledgments The author would like to express his thanks to Claude Sabbah for drawing his attention on this question, and for the useful remarks during the preparation of this text. Various parts of this text were written while the author stayed at several institutes. It was started during his scholarship at the Max Planck Institute for Mathematics in Bonn: he would like to thank for the hospitality and excellent working conditions. Later parts were written while the author was a junior researcher fellow at the Alfr´ed R´enyi Institute of Mathematics in Budapest; he wishes to thank for the grant provided. Thanks equally to the Budapest University of Technology and Economics, where the text was finished. 1 Parabolic connections and Nahm transform We briefly explain the first correspondence. Let C denote the complex line with its canonical holomorphic coordinate x = x1 + ix2 (denoted z in [15]), and CP1 stand for its one-point compactification, the Riemann sphere. The point at infinity will be denoted ∞. Let E be a holomorphic vector bundle of rank r on CP1, whose underlying C∞ vector bundle and holomorphic structure will be denoted by C∞(E) and ¯∂E respectively. Let P = {p1, . . . , pn} be a fixed finite set in C, the singular locus at finite distance. Let O and Ω1 stand for the sheaf of holomorphic functions and holomorphic 1-forms on CP1 respectively. We will denote by Ω1(∗P ∪ {∞}) the sheaf of meromorphic 1-forms with poles of arbitrary order in the points of P and at infinity, and no other poles. Let ∇ be a meromorphic connection on E with poles in P ∪ {∞}: ∇ : E −→ Ω1(∗P ∪ {∞}) ⊗O E. We suppose that ∇ is logarithmic in the points of P , i.e. in a local holomorphic trivialization of E near a point pj ∈ P , it can be written d + M jdx with M j a matrix whose coefficients have at most a simple pole. We also suppose that the residue of ∇ in all pj is semi-simple, and hence in a suitable trivialization it can be written d + Aj dx x − pj + holomorphic terms, 3 where Aj is a diagonal matrix . . . 0 µj rj +1 Aj = 0     , . . . µj r (1) with µj k 6= 0 for rj < k ≤ r. Let E be endowed with a parabolic structure in P : this means that there exists a decreasing exhaustive filtration E• pj of Epj indexed by a finite set of [0, 1[, compatible with the residue in the sense that all Eβ pj is generated by the last k elements of the above diagonalizing trivialization of Aj, for a suitable k = k(β) between 0 and r. The numbers βj k ∈ [0, 1[ such that the graded pieces grβj pj of the filtration are non-trivial are called parabolic weights. k E• Remark 1. The parabolic structure usually comes from a harmonic Hermitian metric [14]. Namely, assume given a tame harmonic metric h; then the filtration is characterised by the requirement that for any β the vectors e ∈ Eβ pj \ 0 are exactly those vectors whose holomorphic extensions σ satisfy h(σ, σ) < cx − pj2β−ε for all ε > 0 and for a suitable c > 0 (depending on e and ε). Under our assumption of semi-simplicity, for any e ∈ grβE• pj \0 and holomorphic extension σ the metric actually satisfies the stronger asymptotic equality h(σ, σ) ≈ e2x − pj2β for some norm . on Epj coming from a scalar product, where ≈ means that the quotient of the two sides converges to 1 as x → pj. In fact, one can de- fine the notion of parabolic structure for connections with arbitrary logarithmic singularities at finite distance (not necessarily semi-simple ones), and Nahm transform probably carries through to this setup too. In this case, we also need to consider the weight filtration of the nilpotent part of the residue on Epj , and the harmonic metrics we are interested in have a more refined behavior h(σ, σ) ≈ e2x − pj2β log(x − pj)w, where w is the weight of e in the weight filtration. However, in this paper we will stick to the semi-simple case. The behavior of the singularity at infinity is supposed to be of Poincar´e rank in a suitable trivialization of E near ∞, it 1, with diagonalizable polar part: can be written d + M ∞dx, 4 where M ∞ is a holomorphic matrix, and such that up to lower-order terms M ∞ = A + C x . (2) Here and A = ξ1   . . . ξ1 . . .   ξn′ . . . ξn′ µ∞ 1 C =  . . .   µ∞ r are constant diagonal matrices, the eigenvalue ξl of A being displayed on the successive places 1 + al, . . . , a1+l of the diagonal. Finally, we suppose given a parabolic structure in this singularity too: this is again a decreasing exhaustive filtration E• ∞ of E∞ indexed by a finite set of [0, 1[ (the parabolic weights β∞ k at infinity), compatible with the polar part of the connection. Again, the parabolic structure is related to the behaviour of a suitable harmonic metric h near infinity: a holomorphic section σ extending a non-zero vector in grβ∞ k E• ∞ is required to have the asymptotic h(σ, σ) ≈ x−2β∞ k (3) up to a constant multiple. Remark 2. The behavior (1) means in the terminology of Simpson [14] that (E, D) is a regular filtered flat bundle with semi-simple residue near the singular points at finite distance. The behaviour at infinity is irregular of quite simple type; such a connection (with only 0 as regular singular point, but together with some extra data) is called non-commutative Hodge structure of exponential type by Katzarkov, Kontsevich and Pantev [8], where one can also find an account of other related definitions existing in the literature. The following notions will be crucial for several statements of our paper. Definition 1.1. We say that the connection is resonance-free if the following conditions hold: k) /∈ Z, µj k 6= βj 1. for all pj ∈ P and rj < k, m ≤ r one has ℜ(µj k, furthermore µj k − µj m /∈ Z \ {0}; 5 2. for all 1 ≤ k ≤ r, one has ℜ(µ∞ Z \ {0} for any al < k, m ≤ a1+l. k ) /∈ Z, µ∞ k 6= β∞ k , furthermore µ∞ k − µ∞ m /∈ 1 = · · · = µj Remark 3. Clearly, these assumptions are fulfilled for generic choices of sin- gularity parameters. The first and third of each set of conditions simply express that eigenvalues of the residue do not differ by non-zero integers (recall that if rj > 0 then µj rj = 0 is also an eigenvalue); we included them in order to simplify the treatment of parabolic structures in the D-module setup, but we expect that they are not necessary for the results to hold. The second one states that the parabolic weights of the associated Higgs bundle are non-zero on the image of the residue of the Higgs field (c.f. the Table on page 720 of [14]); we included it because we will rely on results of [15] where this is assumed. We will also impose a compatibility condition between the polar part of the connection and the parabolic structure. Definition 1.2. We say that the parabolic structure is admissible if gr0E• pj contains the eigenspace ψ0 pj (E, ∇) of the residue for the 0-eigenvalue; in different terms, the natural map is injective (including the possibility ψ0 = 0). pj (E, ∇) → gr0E• ψ0 pj Remark 4. This condition will be needed for the L2-resolution Lemma 6.2. Notice that it is weaker than its counterpart in [15] in that it also allows zero parabolic weight correspond to non-zero eigenvalues of the residue. In a forth- coming paper [3] we plan on extending Nahm transform to a more general setup; for example, we will not require there that the residue of the Higgs field be semi- simple and that the parabolic weights corresponding to the non-zero eigenvalues of the Higgs bundle be all positive. The hope is that the results of this paper will carry through to the more general situation as well. As usual, we call parabolic degree of (E, ∇) the real number par-deg(E) = deg(E) + Xj∈P ∪{∞},k∈{1,...,r} βj k, and parabolic slope the number par- µ(E) = par-deg(E) rk(E) . (4) (5) Moreover, we say that (E, ∇) is (parabolically) stable if any ∇-invariant holo- morphic sub-bundle F , endowed with the induced parabolic structure, satisfies par- µ(F ) < par- µ(E). (6) The following result is important for our arguments: 6 Theorem 1.1 (C. Simpson [14], C.Sabbah [13], O.Biquard-Ph.Boalch [2]). If (E, ∇) is a parabolically stable holomorphic connection of zero parabolic degree on a punctured compact curve with arbitrary diagonalizable polar parts, then there exists a unique harmonic metric (up to multiplication by a constant) be- having as required in Remark 1 and in (3). A harmonic metric is one that satisfies a certain second-order non-linear partial differential equation that we do not spell out here, see e.g. [5] or [14]. Roughly speaking, it is the differential equation governing the fact that the metric, seen as a map from the universal cover of the punctured Riemann surface to the symmetric space of Hermitian metrics on a fixed vector space, be harmonic as a map between Riemannian manifolds. From now on, we will suppose that (E, ∇) is a stable, resonance-free mero- morphic connection of zero parabolic degree on CP1 having regular singularities with semi-simple residues in P , an irregular singularity of Poincar´e rank 1 at in- finity with diagonalizable polar part, and endowed with an admissible parabolic structure in all singularities (at infinity, this simply means a parabolic structure without any further assumption). 1 1 the one-point compact- Denote by bC the dual complex line of C, and by dCP ification of bC. Given (E, ∇) as above, we define in [15] a holomorphic bundle bE on dCP and a meromorphic connection b∇ with semi-simple regular singu- larities in the points bP = {ξ1, . . . , ξn′ }, and an irregular singularity of Poincar´e rank 1 at infinity with diagonalizable polar part, called the Nahm transform of (E, ∇), with an admissible parabolic structure in all singularities. In what follows, we outline its construction and main properties. First, consider the differential-geometric flat connection defined on C \ P D = ¯∂E + ∇ : C∞(E) −→ Ω1 ⊗C∞ C∞(E), where Ωp stands for smooth p-forms. Let now ξ ∈ bC − bP be a parameter and twist D by the formula Dξ = D − ξdx. Consider the elliptic differential complex Ω0 ⊗ E Dξ−−→ Ω1 ⊗ E Dξ−−→ Ω2 ⊗ E. (7) Endow C with the standard Euclidean metric; it is singular at infinity. It is then possible to show (Theorems 2.6 and 2.16 of [15]) that the L2-cohomologies of degrees 0 and 2 of this complex vanish for all ξ, and that its first L2-cohomology L2H 1(Dξ) is a finite-dimensional vector space, whose dimension is independent implicit function theorem in Hilbert space, these finite-dimensional subspaces of ξ. We let the fiber of bE over the point ξ be L2H 1(Dξ). As ξ varies, by the of L2(Ω1 ⊗ E) define a smooth vector bundle bE bC− bP over bC − bP . It inherits a holomorphic structure from the trivial holomorphic structurebd Hilbert bundle L2(Ω1 ⊗ E), and a holomorphic connection b∇ from bd on the trivial − xdξ, 1,0 0,1 7 1,0 wherebd has an interpretation as the L2-kernel of the Laplace operator stands for the (1, 0)-part of the trivial connection. The fiber bEξ also ∆ξ = DξD∗ ξ + D∗ ξ Dξ, (8) where D∗ ξ is the adjoint of the connection operator with respect to the harmonic metric; in other words as the space of L2 harmonic 1-forms (Theorem 2.21 of 1 defined as the sheaf of local holomorphic sections of bounded an adapted harmonic metric is equivalent to poly-stability [2], it follows that called the transformed metric, and one can prove that if the original parabolic respect to this extension is of the same type as that of ∇ described in (1) and [15]). The L2-norm of such a 1-form defines a Hermitian metricbhξ on the fiber bEξ, and these fiber metrics vary smoothly with ξ over bC − bP . This metric is structure is admissible, then bh induces an extension of bE as a holomorphic bundle to dCP norm (Section 4.4 of [15]). The behaviour of b∇ near the singularities with (2) (with different parameters). In addition, this extension of bE also induces an admissible parabolic structure on bE at the points of bP ∪c∞, i.e. the transformed same thing holds for bh (Theorem 4.9 of [15]). In particular, as the existence of if the original connection (E, ∇) is poly-stable, then so is (bE,b∇). Finally, we obtain an explicit description of the singularity behavior of b∇ in the singular set; in particular, we can show that if ∇ is resonance-free, then so is b∇ (Theorems Let us now recall in more details how the singularity behavior of b∇ could be metric is adapted to the transformed parabolic structure (Section 4.6 of [15]). Furthermore, one can prove that if the original metric h is harmonic, then the obtained. We will only work with the singularity at infinity, because the case of the finitely located singular points is similar (and in fact, even simpler). The computation uses Higgs bundles instead of integrable connections, and relies on the formulae of [14] for the link between the singularity data of the two. Hence, it is sufficient to show how the singularity data of the corresponding Higgs bundle transforms. For this purpose, in Section 4.6 of [15] we constructed explicitly a family of asymptotically harmonic 1-forms σ(ξ) in ξ, i.e. such that the L2-norm of ∆ξ σ(ξ) is negligible compared to the L2-norm of σ(ξ) (the quotient k∆ξ σ(ξ)kL2 /kσ(ξ)kL2 converges to 0 as ξ → ξl), and such that in a 4.30 and 4.31 of [15]). neighborhood of c∞ they match up to give a holomorphic local section for the holomorphic structure of the Higgs bundle associated to bE. We give an outline of that construction. Call (E, θ) the Higgs bundle associated to (E, ∇), and set D′′ ξ = ¯∂E + (θ − ξdx/2). (9) dx x − pj We call the operator (D′′)j = ¯∂E + Aj − diag(βj k)r k=1 2 the model D′′-operator, and we deform it as (D′′ ξ )j = (D′′)j − ξ 2 dx. 8 We introduce the model Hermitian metric hj = diag(x − pj2αj k )r k=1, with αj k = {ℜ(µj k)}, where {.} stands for the fractional part in [0, 1[ of the number. Finally, denote by (D′′ ξ )j for the model metric hj, and set ξ )j∗ the formal adjoint of (D′′ (∆′′ ξ )j = (D′′ ξ )j∗(D′′ ξ )j + (D′′ ξ )j (D′′ ξ )j∗ for the model Laplace operator of (D′′ ξ )j. Now, define the spectral points corre- sponding to ξ as the set of all x ∈ C such that det(θ(x)− ξdx/2) = 0. Of course, these points vary analytically with ξ, and as ξ →c∞, they converge to some pj asymptotically to constant multiples of ξ−1. Denote by χ(r) the standard Gaus- sian function on the plane as a function of the radius, cut-off smoothly outside the disk of radius r > 1 so that it is supported on the disk of radius 2. Let σj k be a holomorphic section of E(pj) near pj extending an eigenvector in the image of the residue of the Higgs field, such that for each ξ the vector σj k(qk(ξ)) represent a non-vanishing class in coker(θ(qk(ξ)) − ξdx/2), and choose a sufficiently small ε0 > 0. Define the E-valued smooth (1, 0)-form on CP1 for every ξ close to c∞ by the formula vj k(ξ, x)dx = χ(ε−1 0 ξ(x − qk(ξ)))σj k(x)dx; (10) it is supported on a disk of radius 2ε0ξ−1 centered at the spectral point qk(ξ). Then there exists an E-valued (0, 1)-form tj k(ξ, x)d¯x, necessarily supported in the same disk, such that (D′′ ξ )j (vj k(ξ, x)dx + tj k(ξ, x)d¯x) = 0 for all ξ. We define σ∞ k (ξ, x) as the harmonic representative of this cohomology class with respect to the model Laplace operator, cut off by χ in a neighborhood of qk(ξ) of diameter O(ξ−1). It can then be shown (see Lemma 4.34 of [15]) that the family in ξ of the sections σ∞ k (ξ, z) is asymptotically harmonic. On k (ξ, z) is approximately equal to ξ−αj the other hand, the L2-norm on C of σ∞ multiplied by the L2-norm of a Gaussian function normalized polynomially and of at most polynomial height with respect to ξ, and is therefore bounded from above by a polynomial in ξ. k 2 Laplace transform without parabolic structure In this section, we recall the classical notions of a module over the Weyl-algebra and its Laplace-transform. Let C[x]h∂xi denote the Weyl algebra in the variable x: this is the algebra of polynomial differential operators over C, generated by the formal variables 9 x and ∂x and the relation [x, ∂x] = −1. Let M be a left C[x]h∂xi-module; this means that we have an action of x and ∂x on the elements of M , satisfying x · (∂x · m) − ∂x · (x · m) = −m. In what follows, when we speak of module, we will always mean left-module. Let ξ be another variable. Then there exists an isomorphism between the Weyl algebras C[x]h∂xi and C[ξ]h∂ξi: C[x]h∂xi −→ C[ξ]h∂ξi x 7→ −∂ξ ∂x 7→ ξ. (11) (12) A C[x]h∂xi-module M automatically becomes a C[ξ]h∂ξi-module via this iso- morphism; we denote this resulting module cM , and call it the Laplace-transform of M . In the same way, a C[ξ]h∂ξi-module N becomes automatically a C[x]h∂xi- module, denoted N and called the inverse Laplace-transform of N . These op- erations are clearly inverses of each other. 2.1 Interpretation as a cokernel It is well-known that Laplace transform has an interpretation as a cokernel (see [7] Lemma 7.1.4). We recall this interpretation here. Denote by C[x, ξ]h∂x, ∂ξi the Weyl-algebra in the variables x and ξ, i.e. the algebra over C generated by the variables x, ξ, ∂x, ∂ξ and relations [x, ∂x] = −1 [ξ, ∂ξ] = −1, all other generators commuting. Let the variables ξ, ∂ξ act trivially on M ; this induces a C[x, ξ]h∂x, ∂ξi-module M = M ⊗C C[ξ] with x, ∂x acting on M and ξ, ∂ξ acting on C[ξ] in the natural way. Then the kernel of the map M ∂x−ξ −−−→ M vanishes, and its cokernel is bijective with M as a set. Moreover, the cokernel inherits the structure of a C[ξ]h∂ξi-module from M: the action of ∂ξ is induced by the operator ∂x + ∂ξ − x − ξ on M. Notice that the action of ∂ξ on the cokernel is then induced by the action of ∂ξ − x on M too, for ∂x − ξ clearly acts trivially on its own cokernel. Therefore, for any m ∈ M , ∂ξ applied to the class of the element m ⊗ 1 in coker(∂x − ξ) is (∂ξ − x) · (m ⊗ 1) = −x · m, so 10 the action of ∂ξ on the cokernel is equal to the action of −x on M . Similarly, the class of m ⊗ 1 in coker(∂x − ξ) is annihilated by ∂x − ξ, so the action of ξ on the cokernel is by ∂x. Hence, we get the transformation given by (11-12). In this picture, we also obtain an interpretation of the individual fibers of let (ξ − ξ0) stand for the left ideal of C[ξ]h∂ξi generated by the monomial ξ − ξ0. the meromorphic connection underlying the transform. Indeed, fix ξ0 ∈ bC, and Define the fiber of cM over ξ0 by Then, we have the isomorphism cMξ0 = cM /(ξ − ξ0)cM . cMξ0 ∼= coker(M ∂x−ξ0−−−−→ M ). 3 Meromorphic connections and D-modules In this section we give the link between the objects studied in the previous sections. Let (E, ∇) be a holomorphic bundle endowed with a meromorphic integrable connection on CP1, with poles in P ∪ {∞} as described in Section 1. Denote by E(∗P ∪ {∞}) the meromorphic bundle E ⊗O O(∗P ∪ {∞}) with poles in P ∪ {∞}. By definition, ∇ acts on E(∗P ∪ {∞}). Denote by DCP1 the sheaf of holomorphic differential operators on CP1: it is the sheaf whose local t, with t a local holomorphic coordinate and ai holomorphic functions. E(∗P ∪ {∞}) then becomes a DCP1-module in the usual way: a holomorphic function a(t) acting by multiplication, and ∂t acting by ∇∂/∂t. This DCP1-module is called the meromorphic bundle associated to (E, ∇), denoted M. It has the following properties: sections are of the form Pi ai(t)∂i 1. It is a holonomic module, i.e. any local section is annihilated by a local differential operator. 2. It has only regular singularities at finite distance, and an irregular singu- larity of Poincar´e rank 1 at infinity. Here is a fundamental notion that will underlie our construction: Definition 3.1. The DCP1 -module N is said to be an extension of M if the meromorphic bundle OCP1 (∗P ∪ {∞}) ⊗OCP1 N induced by N is equal to M. An extension N is called minimal if it admits no submodule or quotient module supported in a point. It is known (see e.g. Section 1 of [13]) that a minimal extension always exists and is unique up to isomorphism; it will usually be denoted by Mmin. 3.1 Regular singularities We will now recall the local structure of meromorphic flat bundles and introduce the notion of parabolic structure for these objects, in the presence of a regular 11 singularity. In this section, we fix a p ∈ P , a small disk ∆ ⊂ C centered at p and containing no other element of P , and we suppose that the meromorphic flat bundle M has a regular singularity at p. O and D will stand for O∆ and D∆ respectively. Definition 3.2. A lattice of the meromorphic flat bundle M∆ is a coherent O-module F such that F ⊗O O(∗{p}) = M∆. A lattice F is called logarithmic if the connection 1-form of ∇ written in any local trivialization of F has a first-order pole. If a logarithmic lattice F is chosen for M, then there is a well-defined residue map of ∇: Res(∇) : F p −→ F p induced by the action of x∂x. Modifying F by a ⊗O(−m{p}) amounts to adding m to all eigenvalues of Res(∇). By a theorem of Deligne (see Corollary 2.21 [12]), it is possible to choose a logarithmic lattice F such that all eigenvalues of Res(∇) lie in the interval ] − 1, 0]. Hence, we obtain a finite set Aj of complex numbers with real part in ]−1, 0], such that Aj +Z is the set of eigenvalues of the residue of ∇ on the various logarithmic lattices of M. For a ∈ Aj and m ∈ Z, the generalized eigenspaces of Res(∇) corresponding to the eigenvalues a and α = a + m on the corresponding lattices are isomorphic through the (possibly meromorphic) gauge transformation (x − p)m. For any α ∈ Aj + Z, let ψα p M stand for the generalized eigenspace of Res(∇) on the lattice F ⊗ O(−m{p}) for the eigenvalue α, where F is Deligne's lattice and m = ⌈ℜα⌉ is the smallest integer greater than or equal to the real part of α. In other words, there exists an isomorphism (x − p)m : ψα−m p M −→ ψα p M, and a = α − m ∈ Aj. The assumption that the parabolic structure in p is compatible with the connection means precisely that for all eigenvalues µj k of µj k,• Aj it defines a decreasing exhaustive filtration ψ pj M indexed by β ∈ [0, 1[. By resonance-freeness, for all α ∈ Aj + Z there exists a unique µj k and N ∈ Z such that k + N. It follows that the parabolic filtrations on the µj ψα,• declaring that the above isomorphisms restrict to filtered isomorphisms k-eigenspaces extend to filtrations indexed by real numbers in the interval [N, N + 1[ for any α ∈ Aj + Z by (13) p α = µj (x − p)N : ψ µj k,β p M −→ ψα,β+N p M, (14) where N is the integer appearing in (13). Definition 3.3. A parabolic structure at a regular singular point p of the mero- morphic bundle M is a collection for all α ∈ C of a decreasing exhaustive fil- tration ψα,• pj M of the residue indexed by pj M of the generalised eigenspace ψα 12 elements β of an interval of the form [k, k + 1[ where k is an integer, such that for all α ∈ C, β ∈ R and N ∈ Z the map (x − p)N : ψα+N,β p M −→ ψα,β+N p M be an isomorphism. Hence, we extended the parabolic structure of the holomorphic bundle E as a parabolic structure of M. Remark 5. By admissibility of the parabolic structure of E and because ∇ is resonance-free, the filtration ψ0,• is the trivial filtration (i.e. gr0ψ0,• p p = ψ0 p). 3.2 Irregular singularity Here -- although it would be possible to work in a more general framework -- , we restrict to the case of a Poincar´e rank 1 irregular singularity with diagonal- izable polar part, as described in Section 1. It is known that after taking the formal completion of the meromorphic bun- dle with respect to the x−1-adic valuation (and possibly after a finite ramified cover, that we shall omit), the irregular connection can be put in the form (2), without extra holomorphic terms. Let us choose such a trivialization of the completion of M. Then, in this basis the germ of connection ∇ − Adz has reg- ∞ the generalized eigenspace corresponding to the eigenvalue µ of this connection. The formal connection ∇ − Adz decomposes as direct sum of the n′ germs of connections with regular singularity d + Cldz/z for l = 1, . . . , n′ where Cl is the diagonal matrix diag(µ∞ a1+l) of size a1+l − al. Since ∇ is supposed to be resonance-free at infinity, none of the µ∞ k can be an integer. ular singularity. Denote by bψµ 1+al , . . . , µ∞ Definition 3.4. A parabolic structure at infinity on M is the data of parabolic structures for the meromorphic bundles associated to all of the regular singular connections d + Cldz/z for l = 1, . . . , n′. Remark 6. Notice that the parabolic structures of the regular singular con- nections d + Cldz/z are then all admissibile, because by resonance-freeness at infinity the 0-eigenvalues of their residues are zero, hence they clearly inject into the graded of the parabolic structure for the weight 0. Lemma 3.1. Near ∞ the minimal extension Mmin agrees with the meromorphic bundle M. Proof. As none of the eigenvalues of the residue of ∇ at ∞ are integers, the successive applications of the residue on any eigenvector yield sections with poles of arbitrarily high order. Hence, any DCP1-module containing E necessarily contains M as well. An obvious consequence of the Lemma is that near ∞ the parabolic struc- ture on the meromorphic bundle induces a parabolic structure on the minimal extension. 13 3.3 Stability Once a parabolic structure of a meromorphic integrable bundle M in all singu- larities (including infinity) is given, it is possible to define its degree: deg(M) = − Xp∈P ∪{∞} Xβ∈[0,1[ Xα∈Ap+Z α dim(grβψα p M) where grβψα filtration ψα,• parabolic degree of M: p M = ψα,β p p M/ψα,>β p M is the graded vector space of the parabolic corresponding to the weight β. Similarly, one can define the par-deg(M) = deg(M) + Xp∈P ∪{∞} Xβ∈[0,1[ β dim(grβ pM). Finally, as usual, we define the (parabolic) slope of M as the quotient of its parabolic degree and its rank. This then allows us to define the notion of stability: Definition 3.5. A meromorphic integrable bundle M endowed with a parabolic structure is said to be parabolically stable if all nontrivial integrable subbundles N have slope smaller than M. Here N is endowed with the induced parabolic filtration: ψα,β p N = ψα p N ∩ ψα,β p M. 4 Minimal parabolic Laplace transform Since M is the meromorphic bundle associated to a meromorphic connection on a holomorphic bundle, and Mmin its minimal extension, they are both modules over the sheaf DCP1(∗∞) of meromorphic differential operators with a pole at infinity. The global sections of DCP1(∗∞) form the Weyl algebra C[x]h∂xi, therefore the sets M and Mmin respectively of global sections on CP1 of M and Mmin admit a C[x]h∂xi-module structure. Definition 4.1. The minimal Laplace transform of a meromorphic flat bundle transform N ) for a C[x]h∂xi-module M (respectively, a C[ξ]h∂ξi-module N ). In Section 2 we defined Laplace transform cM (respectively, inverse Laplace M is the meromorphic flat bundle cM associated to the C[ξ]h∂ξi-module dMmin. N on dCP Similarly, the inverse minimal Laplace transform of a meromorphic flat bundle is the meromorphic flat bundle N associated to the C[x]h∂xi-module Nmin. 1 This is similar to N. Katz' Fourier transform in the ℓ-adic case [6]. Let us now fix the parameters appearing in (1) and in the matrices A, C of (2) and introduce two categories: 14 1. HP(C), whose objects are holomorphic bundles with meromorphic connec- tion (E, ∇) over CP1, with resonance-free semi-simple logarithmic singu- larities in P given by (1), and a resonance-free irregular singularity of rank 1 at infinity with diagonalizable polar part given by (2), endowed with an admissible parabolic structure at all singular points 2. DMP(C), whose objects are meromorphic integrable bundles M admitting a lattice with the type of singularities as in HP(C), endowed with an admissible parabolic structure in all singular points. Morphisms in HP(C) are holomorphic bundle maps compatible with connec- tions and parabolic structure, while in DMP(C), maps of D-modules compatible with parabolic structure. Let us also introduce the full sub-categories HPst 0 (C), DMPst 0 (C) consisting of stable objects of parabolic degree 0. Proposition 4.1. The categories HP(C) and DMP(C) are equivalent. The equivalence restricts to an equivalence between HPst 0 (C) and DMPst 0 (C). Proof. Notice first that there exists an obvious functor HP(C) → DMP(C) (15) mapping a holomorphic bundle endowed with a meromorphic connection and an admissible parabolic structure to the associated meromorphic flat bundle M with parabolic structure as described in Section 3. On the other hand, given an object M of DMP(C), associate to it the integrable connection on a lattice as in its definition 2. To see that this is a quasi-inverse of (15) we need to show that such a lattice is uniquely determined. Near a logarithmic point pj this follows from the fact that because of the non-resonance condi- tion, the logarithm of the local monodromy exp(2πiAj) with given eigenvalues (0, . . . , 0, 2πiµj r) is uniquely determined. Near the irregular sin- gularity at infinity we use a method originally due to Levelt (c.f. Section 2 of [9]). Namely, suppose we are given lattices F1 and F2 of M in a neighborhood of ∞ with the property that in a basis of both of them the connection D reads as in (2), with the same matrices A and C. Since both F1 and F2 are lattices of M, these two bases are then linked by a meromorphic gauge transformation g that can be expanded into Laurent series rj+1, . . . , 2πiµj g(t) = ∞Xb=N gbtb for sufficiently small values of t = x−1 and some N ∈ Z. By assumption, we have g−1(At−2dt + Ct−1dt + O(1)dt)g − g−1dg = At−2dt + Ct−1dt + O(1)dt, or equivalently dg = (At−2dt + Ct−1dt + O(1)dt)g − g(At−2dt + Ct−1dt + O(1)dt). 15 Inserting the Laurent series of g into the latter equation yields the equations AgN − gN A = 0 CgN − gN C + AgN +1 − gN +1A = N gN , and so on. The first of these equations implies that gN is in the Levi subgroup L = Gl(a2 − a1, C) × · · · × Gl(a1+n′ − an′ , C) corresponding to the eigenspace-decomposition of A. In the second expression, the term AgN +1−gN +1A lies in the complementary of L, therefore the projection of CgN − gN C to L must be equal to the right-hand side. Since the restriction of C to any eigen-subspace of A has by assumption distinct eigenvalues modulo the integers (except for multiple eigenvalues), it follows that the only integer eigenvalue of the adjoint action of C restricted to L is 0. Therefore, the second equation implies N = 0; in particular, g is holomorphic. The same argument applied to g−1 shows that F1 = F2. It is clear that this equivalence preserves parabolic degree and stability. 4.1 Parabolic transform The transform defined above is compatible with admissible parabolic structures: if an admissible parabolic structure is given on M, then we can define one on its (inverse) minimal Laplace transform cM by the following construction. 1. In regular singularities: these are the eigenvalues {ξ1, . . . ξn′ } of A. By formal stationary phase (Prop. V.3.6, [12]) and the isomorphism ψαM ∼= ψαMmin for all α /∈ Z, there exists a formal isomorphism ∞M ∼= bψα n′Ml=1 ψα ξl cM, where we recall that bψα ∞ is the generalized eigenspace corresponding to the eigenvalue α of the "regular part" ∇ − Adz of the connection at infinity. The parabolic filtration on the left hand side therefore defines one on the right-hand side. Finally, on the 0-eigenspaces of the residue in ξl, we put the trivial parabolic structure. 2. At infinity: similarly to the regular case, formal stationary phase applied to the inverse Laplace transform, for all α /∈ Z there exists a formal isomorphism nMj=1 ∞cM ∼= bψα ψα pj M, and the parabolic filtrations on the direct summands on the right-hand side induce a parabolic filtration on the space on the left-hand side. 16 The parabolic degree and stability condition corresponding to the structure of this definition behave well with respect to Laplace transform. The following statements are due to C. Sabbah, Proposition 4.1 and Theorem 5.1 of [11]; for sake of completeness we reproduce them here together with their proof. In what follows, we let M denote a parabolically stable meromorphic integrable bundle of parabolic degree 0. regular singular points ξl. In particular, parabolic minimal Laplace transform and inverse parabolic minimal Laplace transform are inverses of each other. Proposition 4.2. The Laplace transform dMmin is a minimal extension at all Proof. Suppose to the contrary that dMmin is not a minimal extension at some ξl, so that it admits a submodule (or a quotient module) supported at ξl. By inverse Laplace transform, it follows that Mmin admits a submodule (or a quotient) N of the form d + ξldx on some trivial bundle. In particular, N has a pole only at ∞, and the residue of the connection is 0. However, by resonance-freeness at infinity M cannot admit such a sub-module. Since dMmin is a minimal extension, it is clearly the minimal extension of cM, whose inverse parabolic minimal Laplace transform is therefore Mmin endowed with its original parabolic structure. This implies the second claim. Remark 7. The statement also holds without the resonance-freeness assump- tion, but the proof then makes use of stability: consider the meromorphic in- tegrable bundle N associated to N . Since the parabolic structure of Mmin is admissible, the induced parabolic structure on N only has 0 weight. Therefore, N is a subbundle of parabolic weight 0 of the meromorphic bundle M associated to Mmin, contradicting stability. Proposition 4.3. If M is a parabolically stable meromorphic integrable bundle of parabolic degree 0, then so is its parabolic minimal Laplace transform cM. Proof. By the stationary phase formula for ordinary Laplace transform [10], the non-zero eigenvalues of the residue of the connection at the regular singularities are switched with the eigenvalues of the residue at infinity. In particular, the sum of the eigenvalues (counted with multiplicity) at all singularities is the same number for M and cM. By our definition of parabolic structure for cM and by admissibility, the same holds for the sum of all parabolic weights. Indeed, as the only weight on ψ0 pj is 0, the two total sums only differ in terms equal to 0. In view of the formulae of Subsection 3.3, this proves equality of the parabolic degree. It follows from Proposition 4.2 that dMmin is the minimal extension of cM. Let bN be a non-trivial meromorphic integrable subbundle of cM. Then, its minimal extension bNmin at all regular singular points is a non-trivial submodule of cMmin. Hence, the inverse Laplace transform Nmin is a non-trivial submodule transform of bN . of Mmin. In particular, the corresponding meromorphic integrable bundle N is a non-trivial subbundle of M. Clearly, N is the inverse minimal Laplace 17 Lemma 4.1. The parabolic structure induced on N from M agrees with the inverse minimal Laplace transformed parabolic structure coming from bN . Proof. We only give a proof of this fact under the extra assumption that for any j ∈ {1, . . . , n} all the non-zero eigenvalues µj k are distinct (not only modulo Z); the general case is a consequence of functoriality of the stationary phase formula, c.f. Subsection 6.4. For any fixed pj ∈ P , the set Aj(N ) of eigenvalues of the residue of N is contained in Aj(M), and for each α ∈ Aj(N )+ Z, we have ψα j M. Since by assumption all of these eigenspaces are either 0 or 1-dimensional, it is clear that the only parabolic weight for the induced parabolic structure of a non-zero ψα j N is equal to the parabolic weight of ψα j M, which in turn is equal to the parabolic weight of ψα the residue. This latter, however, is the parabolic weight of ψα the pj-eigenspace of the leading term of the residue (notice that this latter space is non-zero if and only if ψα c∞cM restricted to the pj-eigenspace of the leading term of c∞ bN restricted to j N 6= 0 in view of the stationary phase formula). j N ⊂ ψα As above for M, the Lemma then implies equality between the parabolic degrees par-deg(N ) and par-deg(bN ). By stability of M, we have par-deg(N ) < 0, hence also par-deg(bN ) < 0, whence stability of cM. We are ready to state our main result. Consider the diagram of functors HPst 0 (C) Nahm inverse Nahm - HPst 0 (bC) (16) DMPst 0 (C) minimal Laplace inverse minimal Laplace DMPst 0 (bC) where the vertical arrows are the equivalences defined in Proposition 4.1. Theorem 4.1. Diagram (16) commutes. Remark 8. The coincidence at the level of ranks follows by comparing the rank formulae for Laplace (Proposition V.1.5. of [10]) and Nahm transforms (formula (1.10) of [15]). 18 - m m - - R R m m R R 5 Outline of the proof The main idea of the proof consists in comparing both constructions to the first hypercohomology of the twisted de Rham complex of Mmin. Set DR(Mmin, ∂x − ξ) = (Mmin ∂x−ξ −−−→ Mmin) for the de Rham complex of Mmin twisted by −ξ, where the two non-zero sheaves are placed in degrees 0 and 1. For any sheaf complex C on CP1, write Hm(C) set for its hypercohomology of degree m. On the other hand, for any ξ ∈ dCP 1 ∇ξ = ∇ − ξdx, 1,0 1 1 . and call this operator the twisted integrable connection. Recall from Section 1 cohomology for the Euclidean metric of the elliptic complex (7), holomorphic formed object. Recall also that Nahm transform of a holomorphic bundle (E, ∇) is induced by d − xdξ, and the transformed metric is defined by the L2- norm of the harmonic representative of an L2-cohomology class. Recall finally that we denote by bP the set {ξ1, . . . , ξn′} of regular singularities of the trans- with meromorphic connection on CP1 is a holomorphic bundle (bE,b∇) with meromorphic connection on dCP : the fiber over a fixed ξ ∈ bC is the first L2- structure is induced by the trivial holomorphic structure with respect to ξ, b∇ that (bE, b∇) admits a holomorphic extension induced by the harmonic metric h: holomorphic sections of bE at a singular point are holomorphic sections in a sections of bE at a singular point are holomorphic sections in a neighborhood of factor in the product CP1 ×dCP Our first aim is to define a natural holomorphic extension to dCP of the bundle of first hypercohomologies of DR(Mmin, ∂x − ξ). Denote the sheaf of holomorphic sections of the vector bundle E also by E, and let F be the sheaf over CP1 whose local sections are meromorphic sections of E ⊗ Ω1 with at most a double pole at infinity and at most a simple pole in the points of P , such that the residue at a point pj ∈ P be in Im(Aj). We then have a sheaf map that point whose norm grows at most polynomially with respect to the inverse of the distance to the point. For j = 1, 2, denote by πj projection to the j-th neighborhood of that point whose norm is bounded. In particular, meromorphic 1 ∇ξ : E −→ F. (17) The bundle of first hypercohomologies of DR(Mmin, ∂x − ξ) carries a natural holomorphic structure on bC \ bP induced by the trivial holomorphic structure of Mmin with respect to ξ. The same holds for the bundle of first hypercohomolo- gies H1(E ∇ξ−−→ F ). ∇ξ−−→ F of sheaves on Proposition 5.1. The complexes DR(Mmin, ∂x −ξ) and E CP1 are quasi-isomorphic. Furthermore, the natural holomorphic structures of the bundles of first hypercohomologies over bC \ bP agree under this isomorphism. 19 Via this proposition the holomorphic bundle of first hypercohomologies H1(DR(Mmin, ∂x − ξ)) on bC \ bP admits a natural holomorphic extension to bC. In particular, this holo- morphic extension will also induce a meromorphic extension over these points. At a regular singular point ξl the holomorphic extension is defined as the sec- ∇ξl−−→ F ). tions whose value at ξl is in the finite-dimensional vector space H1(E 1 It remains to define a holomorphic extension at c∞. We follow Subsection 4.4.2 of [15] and slightly modify the map (17) near infinity as follows: let U0 = bC and Uc∞ = dCP with coordinates ξ and ζ = ξ−1, and let s0 and sc∞ stand for the canonical global sections of the sheaf OdCP \ {0} be the standard affine open charts of dCP 1 (c∞) satisfying 1 s0(ξ) = ξ s0(ζ) = 1 sc∞(ξ) = 1 sc∞(ζ) = ζ on U0 on Uc∞. 1(c∞). On the pullback bundle π∗ , introduce the sheaf map 1 1 E over CP1 ×dCP ∇• = ∇ ⊗ sc∞ − dz ⊗ s0 : π∗ 1 E −→ π∗ 1F ⊗ OdCP (18) 1 H1(E vector in dz−−→ F ) ∼= ⊕j Im Res(∇, pj), sion is clearly induced by the holomorphic extension of the original bundle E at the points pj. the map ∇• is a connection whose restriction to the fiber CP1 × {ξ} is ∇ − ξdx. 1 We may then define the holomorphic extension of the bundle of hypercohomologies H1(DR(Mmin, ∂x − Over the affine U0 and relative to dCP ξ)) to c∞ to consist of those holomorphic sections near c∞ that converge to a the first hypercohomology of the fiber of (18) overc∞. This holomorphic exten- bEξ = L2H 1(Dξ) of (7) for the Euclidean metric with the first hypercohomology Proposition 5.2. For ξ ∈ bC \ bP , the vector spaces bEξ = L2H 1(Dξ) (for the ing holomorphic bundles over bC \ bP are isomorphic. Finally, this isomorphism respects the natural meromorphic extensions to dCP Euclidean metric) and H1(DR(Mmin, ∂x − ξ)) are isomorphic. The correspond- The next ingredient of the proof is an identification of the first L2-cohomology The last step in identifying the meromorphic bundles underlying the two of the twisted de Rham complex of Mmin. 1 . transforms is: 1It might look disturbing that over the chart Uc∞ this is no longer a connection, but instead a ζ-connection. Notice however that we are mainly interested in the first hypercohomology of this complex, and the hypercohomology of a map is independent of scalar rescaling. 20 Proposition 5.3. For all ξ /∈ bP , the first hypercohomology H1(DR(Mmin, ∂x − ξ)) is isomorphic to the fiber ( dMmin)ξ of the minimal Laplace transform in ξ. The corresponding holomorphic bundles over bC \ bP are isomorphic. Finally, the natural meromorphic extensions over dCP of the underlying meromorphic bundles are also isomorphic. 1 Having thus obtained isomorphism of the meromorphic bundles, we move on to match the transformed connection with the ∂ξ-action. Proposition 3.5 of [15] states that the transformed connection b∇ is induced on L2-cohomology by taking the quotient of bd − xdξ. The isomorphism of Proposition 5.2 maps it to the connection induced by taking the quotient of bd − xdξ in hypercohomology. Subsection 2.1, shows that over the open set bC \ bP the connection of the Nahm This, coupled with the interpretation of Laplace transform as a cokernel of transform is mapped to the ∂ξ-action of the minimal Laplace transform. To see that the square of Theorem 4.1 commutes, the last thing to show is that the minimal extension over the singularities of the meromorphic bundle with connection associated to (bE, b∇) is isomorphic to dMmin. Since the under- lying meromorphic connections agree and by uniqueness up to isomorphism of the minimal extension of a meromorphic bundle with connection, Proposition 4.2 implies the claim. Finally, it remains to compare the parabolic structures of the transformed objects: Proposition 5.4. The parabolic structure of (bE,b∇) is mapped under this cor- respondence into the one defined in Subsection 4.1. 6 Proof of the propositions We will carry out all proofs for the direct transforms; the inverse transforms are known to agree with the direct ones up to a sign ξ ↔ −ξ, therefore the same arguments imply the statements for the inverse transforms. 6.1 Proposition 5.1 First, let us define a sequence of sub-sheaves of Mmin: for m ≥ 0 set Fm for the sheaf of sections of Mmin with pole of order at most m at all singular points. These sheaves define an exhaustive filtration of Mmin : E = F0 ⊂ F1 ⊂ F2 ⊂ . . . ⊂ Mmin such that ∂x − ξ maps Fm into Fm+1 for all ξ /∈ bP . Notice that by definition F = F1 ⊗Ω1. Moreover, the quotient sheaves Fm+1/Fm are supported in the set of regular singular points P , and the stalk at pj is of dimension r − rj. Denote by [∂x − ξ] the map induced by ∂x − ξ on these quotients. 21 Lemma 6.1. For all m > 0 the map [∂x − ξ] : Fm/Fm−1 −→ Fm+1/Fm is an isomorphism of complex vector spaces. Proof. Since twisting by ξ does not modify the principal term of ∂x, it is suffi- cient to show the result for ξ = 0. The stalk of Fm/Fm−1 at p ∈ P is isomorphic to the image of the residue of ∇ in p (up to the 1-form dz), and the same thing holds for Fm+1/Fm. The action of [∂x] after this identification is the matrix Res(∇, p) − m · Id. Since ∇ is resonance-free, this matrix admits no 0 eigen- value. Since the source and target spaces are of the same dimension, it is an isomorphism. Clearly, the complexes E ∇ξ−−→ F and E ∂x−ξ −−−→ F1 are quasi-isomorphic. By ∂x−ξ ∂x−ξ −−−→ Fm −−−→ Fm+1 and Fm−1 the lemma, for all m > 0 the complexes Fm ∂x−ξ −−−→ Mmin is are quasi-isomorphic. Since Mmin = ∪mFm, the complex Mmin ∂x−ξ −−−→ Fm+1. The sequence of these the inductive limit of its sub-complexes Fm latter being stationary (up to quasi-isomorphism) for m ≥ 0, we conclude that ∇ξ−−→ F . This Mmin shows the fiber-wise statement. The statement about holomorphic structures is clear. ∂x−ξ −−−→ Mmin is quasi-isomorphic to F0 ∂x−ξ −−−→ F1, hence to E 6.2 Proposition 5.2 Consider the double complex L2(Ω0,1 ⊗ E) ∇ξ / L2(Ω1,1 ⊗ E) (19) ¯∂E L2(E) ¯∂E ∇ξ L2(Ω1,0 ⊗ E) E ∇ξ F? Here we use the convention that whenever L2 is written at a particular place of a diagram, it is meant to be the intersection of the domains of the operators corresponding to all arrows that point out from that place. To prove equality between L2-cohomology and hypercohomology it is suffi- cient to show two statements: Lemma 6.2. For all ξ /∈ bP the vertical arrows in (19) are sheaf resolutions. Remark 9. A similar result is shown in Theorem 6.2 of [16] for the Poincar´e- metric near the punctures and with respect to a Hermitian metric with zero parabolic weight but non-trivial weight-structure, see also Theorem 1.2 of [13]. 22 / / / O O O O / / ? O O O O However, we included a proof because we see no direct way of applying those results here. Lemma 6.3. The first cohomology of the simple complex associated to (19) is equal to L2H 1(Dξ). Proof of Lemma 6.2. This is analogous to Lemma 4.12 of [15], where the same statement is proved for Higgs bundles. As the problem is local, and away from singularities all spaces involved are usual L2-spaces (so exactness is guaranteed by usual L2-theory), we focus on a neighborhood of a singular point that we may suppose to be a disk. Let us first treat the column on the left, near a regular singular point p = pj. Fix a holomorphic local trivialization e1, . . . er of E in which the connection can be written ∇ = d + Aj dx/(x − pj) up to holomorphic terms, where Aj is the matrix in (1). In all that follows, we will omit the index j of the singularity. Let e =Pk ϕkek be a local L2 section of E in the kernel of ¯∂E: this means that the ϕk are meromorphic functions. Recall that the L2 condition is measured with respect to the metric h, and this latter has asymptotic behavior bounded with diag(x − p2βk ), where by admissibility 0 ≤ βk < 1 for rj < k and βk = 0 otherwise. Hence the condition e ∈ L2 is equivalent to the conditions • for k ≤ rj the coefficients ϕk are holomorphic at p; • for rj < k such that βk = 0 the coefficients ϕk are holomorphic at p; and • for rj < k such that βk > 0 the coefficients ϕk have at most a simple pole at p. Let us show that the requirement ∇ξe ∈ L2 implies even in the last case that ϕk must be holomorphic; for this purpose consider their Laurent-series: ϕk(z) = ϕk,−1 1 x − pj + ϕk,0 + ϕk,1(x − pj) + . . . By the local form of ∇ in the trivialization e1, . . . er, the coefficient of ekdz for rj < k in ∇ξe is (cid:18) ∂ ∂x + µk x − p(cid:19) ϕk(x) + rXl=1 hl(x)ϕl(x), where the hl are holomorphic functions. Plugging the Laurent-expansion into this yields (µk − 1) ϕk,−1 (x − p)2 + O(cid:18) 1 x − p(cid:19) , where O(·) denotes terms with at most a simple pole, depending on ϕl,s, µl and hl. Such a 1-form is in L2 if and only if (µk − 1)ϕk,−1 = 0. Since ∇ is resonance- free, this is equivalent to ϕk,−1 = 0, in different terms holomorphicity of ϕk. On the other hand, for k ≤ rj, taking into account that all ϕl are holomorphic and 23 that µk = 0, the coefficient of ekdx in ∇ξe is clearly holomorphic. This proves that e is in the L2-domain of ∇ if and only if it is holomorphic, hence the left column is exact in the term L2(E). Let us now study the second term of this column: given a section f = Pk φkekd¯x of L2(Ω0,1 ⊗ E), we would like to show that there exists an L2 section e =Pk ϕkek of E such that ∇ξe ∈ L2 and ¯∂Ee = f . As e1, . . . , er is a holomorphic trivialization (in particular, ¯∂E is diagonal), we may suppose that f = φekd¯x for some k. For k ≤ rj , by admissibility both ∇ and the Hermitian metric are equivalent to the usual metric, so the usual L2 Dolbeault lemma gives the result. For rj < k, the section φek is in L2 with respect to h if and only if φx − pβk is an L2 function. In the case βk > 0 the parabolic L2 Dolbeault lemma (Claim 4.13 of [15]) implies that there exists ϕ such that ¯∂ϕ = φ and ϕx − pβk−1 ∈ L2. Finally, in the case rj < k and βk = 0 the result again follows from the usual L2 Dolbeault lemma: for f = φekd¯x ∈ L2 we find ϕ ∈ L2 such that ¯∂ϕ = φ; now as by assumption ∇(f ) is also in L2 and the diagram (19) commutes we have that ∇ϕ is (up to subtracting a holomorphic 1-form) also in L2; whence surjectivity of ¯∂E. This finishes the proof of exactness of the left column. All the other statements contained in the lemma (i.e. that the right column is also exact in the regular singularities and that both columns are exact near infinity) are immediate modifications of the same arguments. We leave details to the reader. Proof of Lemma 6.3. The corresponding statement for Higgs bundles is Propo- sition 4.7 of [15]. As the differential of the simple complex associated to (19) is Dξ, this is purely a question of domains. Denote by Cξ the simple complex associated to (19). An element of the degree 1 term C1 ξ of Cξ is represented by a sum f 0,1d¯x + f 1,0dx ∈ L2(Ω0,1 ⊗ E) ⊕ L2(Ω1,0 ⊗ E) such that ∇ξf 0,1d¯x ∈ L2(Ω1,1 ⊗ E) and ¯∂Ef 1,0dx ∈ L2(Ω1,1 ⊗ E). Obviously, one then also has Dξ(f 0,1d¯x + f 1,0dx) ∈ L2(Ω1,1 ⊗ E); in other words, there exists a tautological injection ξ −→ L2(Dξ)1, C1 the latter space being the L2 domain of Dξ on 1-forms. Since terms of degree 0 of the complexes Cξ and L2(Dξ) are the same together with differentials, this injection induces an injection of degree 1 cohomology spaces H 1(Cξ) −→ L2H 1(Dξ). Let us show that this map is surjective too: for this, suppose g = g0,1d¯x+g1,0dx represents a cohomology class in L2H 1(Dξ); by definition, this means that g is in L2 and in the kernel of Dξ. We would like to find a section f = f 0,1d¯x + f 1,0dx representing the same class, such that we have in addition ∇ξf 0,1d¯x ∈ 24 L2(Ω1,1 ⊗ E) and ¯∂Ef 1,0dx ∈ L2(Ω1,1 ⊗ E). Since this is guaranteed away from singularities by usual L2-theory and because ∇ξ is an isomorphism near infinity, it is sufficient to focus on a neighborhood of a regular singularity. In such a neighborhood, Lemma 6.2 allows us to find ϕ ∈ L2(E) such that ¯∂Eϕ = g0,1d¯x and ∇ξϕ ∈ L2(Ω1,0 ⊗ E). The local L2 section f = (g0,1d¯x − ¯∂Eϕ) + (g1,0dx − ∇ξϕ) obviously satisfies the required conditions. 1 Let us come to the second statement of the proposition. Proposition 3.5 of π∗ the quotient in L2-cohomology of the trivial holomorphic structure relative to 1 E. On the other hand, we defined the holomorphic structure on ∇ξ−−→ F ) as the quotient of the . This yields equality of the of [15] gives the holomorphic structure of bE: on the open set bC \ bP , it is dCP trivial holomorphic structure on F relative to dCP the bundle of first hypercohomologies H1(E holomorphic structures away from singularities. Finally, in order to prove that the meromorphic extensions agree, it is suffi- cient to show that the L2-norms of harmonic 1-forms whose limit as ξ → ξl (re- ∇ξl−−→ F ), is of moderate growth. Indeed, in Section 1 the holomorphic extension of the Nahm transformed bundle spectively ξ →c∞) is a vector of the space H1(E bE to the singularities is defined by the condition that the L2-norms of harmonic representatives be bounded; hence, meromorphic sections near a singular point are exactly the ones having at most polynomial growth. We only treat the case of the singularity at infinity, the case of the regular singularities being similar. The crucial point is that since the Laplace operator (8) is (up to a constant 1 2) well-known to be equal to ∆′′ ξ = D′′ ξ (D′′ ξ )∗ + (D′′ ξ )∗D′′ ξ (20) where D′′ the two operators are the same. In particular, for fixed ξ the 1-form σ∞ ξ is the operator defined in (9), the space of L2 harmonic 1-forms for k (ξ, x) are holomorphic for the Dolbeault holomorphic structure, and not the de Rham holomorphic structure, see Section 4.4 of [15]. Recall that σj k was chosen to be constructed in Section 1 is also a vector of bEξ. However, as ξ varies, these vectors a holomorphic section of E(pj) near pj such that for each ξ ∈ bC \ bP the element k (ξ, x) of bE in a punctured disk near c∞ in the σj k(qk(ξ)) represent a non-vanishing class in coker(θ(qk(ξ)) − ξdx/2), and qk(ξ) depends analytically on ξ. Let us now modify the construction of σ∞ k (ξ, x) to obtain a smooth local section τ ∞ following way: instead of σj k , holomorphic with respect to the de Rham holomorphic structure ¯∂E near pj, extending an eigenvector of the image of the residue of ∇ at pj. Notice that by resonance-freeness the images of Res(∇, pj) and of Res(θ, pj) are isomorphic, and formula (1.8) of [2] means that a possible choice for τ j k, pick a local section τ j k is k (x) ≈ x − pjβj τ j k−αj k σj k(x) (21) 25 k is the fractional part of ℜµj where αj k. In particular, for each ξ we still have that τ j k (qk(ξ)) represents a non-trivial class in coker(θ(qk(ξ))−ξdx/2). Therefore, the definitions of vj k (ξ, x) in Section 1 carry over to this situation to define wj k(ξ) satisfying the same equation k(ξ) and σ∞ k(ξ), sj k(ξ), tj D′′ ξ (wj k(ξ, x)dx + sj k(ξ, x)d¯x) = 0, and therefore a local smooth section τ ∞ k (ξ, x) of bE near c∞. k (ξ, x) of bE is holomorphic in ξ with respect to the de Rham holomorphic structure. In particular, the set of local sections τ ∞ k of Im(Res(∇, pj)) defines k (ξ, x) for all pj ∈ P and a basis τ j Lemma 6.4. On bC \ bP , the local smooth section τ ∞ a holomorphic extension bF of bE bC\ bP at c∞. Proof. In the construction of τ ∞ k (ξ, x), the dependence on ξ lies in the spectral point qk(ξ). 2 As this latter varies holomorphically with ξ, the monad construc- tion of [4], Subsection 3.1.3, shows that τ ∞ k (ξ, x) is holomorphic with respect to By Theorem 1.6 of [15], the rank of the transformed integrable bundle is j=1 dim(Im(Res(∇, pj))), which is the number of the local sections k (ξ, x) for all pj ∈ P and τ j τ ∞ k ∈ Im(Res(∇, pj )). To see that these latter define a holomorphic extension, it is therefore sufficient to prove that they are linearly the holomorphic structure on C∞( E) induced by the pull-back to (C\P )×(bC\bP ) of E, which is by definition the de Rham holomorphic structure of bE. equal to Pn independent over C(ξ). This will follow from the fact that as ξ →c∞ they are linked to the linearly independent sections σ∞ k (ξ, x) via (22). Lemma 6.5. The sections τ ∞ k (ξ, x) are asymptotically harmonic. Proof. Identical to the proof of Lemma 4.34 of [15], for this latter does not rely at all on the choice of vectors representing classes in coker(θ(qk(ξ)) − ξdx/2) as in (21). This now implies that the growth of the L2-norm of these sections is at most polynomial. Namely, we have: Lemma 6.6. The L2-norm of τ ∞ metric h satisfies k (ξ, x) over C with respect to the harmonic cξ2−2βj k ≤ kτ ∞ k (ξ, x)k2 L2 ≤ Cξ2−2βj k for some constants 0 < c < C. 2A subtle point here is that in (10) the argument of the Gaussian function χ contains ξ, which is not holomorphic in ξ. However, locally near any ξ0 /∈ bP this factor could be kept constant ξ0 independently of ξ, and the sections so obtained would be cohomologous to the sections (10), so the non-holomorphic dependence is actually only apparent. The point in the choice of the argument of χ is just that its support be sufficiently separated from the other spectral points. Of course, the sections with ξ replaced by ξ0 do not extend to a neighborhood of the singular points bP , so for the construction of asymptotically harmonic sections only the choice in (10) is suitable. 26 Proof. Similar to the proof of Theorem 4.32 of [15]. The difference with that case lies only in the behavior of the metric h on the sections τ j k. In view of (21) and the asymptotic qk(ξ) − pj ≈ 2λj kξ−1 (Claim 4.27, loc. cit.) and because the construction of wj k depends C-linearly with the chosen cokernel vector τ j k and σj k , the sections τ ∞ k (ξ) satisfy k, sj k (ξ) ≈ c0ξ−βj τ ∞ k+αj k σ∞ k (ξ) (22) as ξ →c∞ for some constant c0 6= 0 depending on λj of loc. cit., we have k, αj k, βj k. By Theorem 4.32 c1ξ2−2αj k ≤ kσ∞ k (ξ, x)k2 L2 ≤ C1ξ2−2αj k . Comparing these estimates yields the desired asymptotic. This finishes the proof of the proposition. 6.3 Proposition 5.3 The hypercohomology long exact sequence gives a decomposition H1(Mmin ∂x−ξ −−−→ Mmin) ≃ coker(H 0(Mmin) ∂x−ξ −−−→ H 0(Mmin))⊕ ker(H 1(Mmin) ∂x−ξ −−−→ H 1(Mmin)). By definition, H 0(Mmin) is Mmin, and resonance-freeness at infinity implies that in a neighborhood of infinity the equality Mmin = Mmin(∗{∞}) holds, i.e. locally at infinity Mmin coincides with the meromorphic bundle. In particular, the first cohomology H 1(Mmin) vanishes, and it follows that H1(Mmin ∂x−ξ −−−→ Mmin) ≃ coker(Mmin ∂x−ξ −−−→ Mmin). This gives the fiber-wise statement of the proposition. The holomorphic structure induced on the bundle H1(E ∇•−−→ F ) by taking 1F relative to CP1 is the quotient of the trivial holomorphic structure of π∗ mapped by the above isomorphism to the one induced by the trivial C[ξ]-module structure of Mmin = Mmin ⊗C C[ξ] on the cokernel of ∂x − ξ. This latter is by It remains to match the meromorphic extensions over the singularities. For this purpose, we need to study what happens to the isomorphism of Proposition definition the holomorphic structure of dMmin on bC \ bP . 5.1 as ξ → ξl for some ξl ∈ bP or as ξ → c∞. Near the points ξl, it is known that cM admits a regular singularity. On the other hand, by Theorem 1.6 of [15] the connection b∇ on the transformed extension of bE over ξl is logarithmic; in particular, the meromorphic flat bundle bE(∗bP ) has regular singularities at all points of bP . By uniqueness of the meromorphic extension with regular singularity, the meromorphic flat bundles are naturally isomorphic over bC. 27 Therefore, we only need to treat the case of the irregular singular point described in the cokernel-interpretation of Subsection 2.1 as the lattice generated by the classes of F ⊂ Mmin, i.e. of sections admitting a pole of order 1 at the points of P with respect to the lattice E. By resonance-freeness, Mmin is spanned c∞. For this purpose, consider the holomorphic bundle E over CP1 as a C[x]- submodule of Mmin. The lattice bF near c∞ introduced in Lemma 6.4 can be as a C[x]h∂xi-module by F . Therefore, by Theorem V.2.7 of [12] bF is a lattice of the Laplace transform dMmin of Mmin near infinity; this concludes equality of the meromorphic extensions. 6.4 Proposition 5.4 This follows from comparing the construction of Subsection 4.1 with Subsection 6.2. Notice first that it is sufficient to treat the case of the singularity at c∞. Indeed, the result then also follows for the finitely located singularities using inverse transforms. First, let us prove the stationary phase formulae used in the construction of Subsection 4.1. As we have already obtained the isomorphism of Nahm and minimal Laplace transforms (without parabolic structure), it is sufficient to show the results for Nahm transform. These latter, as we will see, are a simple consequence of the stationary phase formulae for Nahm transform of parabolic Higgs bundles, in view of the non-Abelian Hodge theory of [2]. Hence, we need first to describe how one obtains the stationary phase formulae for Nahm transform of parabolic Higgs bundles. For this purpose, notice that by Claim 4.27 of [15] all of the spectral points qm(ξ) converge to one of the pj's as for which the corresponding qm(ξ) converges to pj. It follows that the sections k (ξ) introduced in Section 1, for all pj ∈ P and a diagonalizing basis σj σ∞ ξ → c∞, and for fixed 1 ≤ j ≤ n there are exactly rk(Aj) = r − rj indices m of Im(Res(θ, pj)) define a lattice bF at c∞ of the transformed Higgs bundle.  Furthermore, by Theorem 4.31 of [15] with respect to this lattice the polar part at infinity of the transformed Higgs field reads  p1 λ1 r1+1 k . . .   dξ ξ . λ1 r . . . λn rn+1 . . . (23) λn r . . . − 1 2    dξ−  p1 . . . pn . . . pn Here, as in (1.11) of [2], the constants λj k = µj k − βj k 2 28 are the eigenvalues of the residue at pj of the original Higgs field corresponding to (E, ∇, h), which are non-zero if and only if µj k 6= 0 because we assume that the parabolic connection is admissible and resonance-free. Recall from Lemma 6.4 that the sections τ ∞ and that we set ζ = ξ−1 for a local coordinate of dCP m (ξ, z) define a lattice bF nearc∞, near c∞. m (ξ, z) of the lattice bF ⊗ 1 (c∞) the polar part of the transformed connection b∇ reads Lemma 6.7. With respect to the trivialization ζ τ ∞ OdCP 1 µ1 r1+1 . . . p1 . . . pn . . . pn   dξ−   − p1   . . . µ1 r . . . µn rn+1 . . . µn r   dξ ξ . k (ξ) is −1 + αj pj ∈ P and a diagonalizing basis τ j k (ξ) for all k of Im(Res(∇, pj)). Now, (22) states that the formula (1.8) of [2]. By Theorem 4.32 of [15], the parabolic weight corresponding to the section σ∞ k; in particular, all the weights corresponding Proof. Recall that the lattice bF had as generating local sections τ ∞ trivializations bF and bF are linked (up to some multiplicative constants) as in to a trivialization of the lattice bF lie in [−1, 0[. Hence, the parabolic weights corresponding to bF ⊗OdCP m (ξ, z) of bF ⊗ OdCP 1(c∞) respectively are then also integrable connection b∇ and of the transformed Higgs field bθ with respect to the lattices bF ⊗ OdCP 1(c∞) are related as in formula (1.10) of loc. 1 (c∞), so the polar part of the transformed Higgs field with respect to the lattice bF ⊗ OdCP 1(c∞) is also equal to (23). This 1 (c∞) lie in [0, 1[. Clearly, the trivializations ζ σ∞ 1 (c∞) and bF ⊗ OdCP cit. On the other hand, in the Dolbeault theory the polar part does not change under tensoring a lattice by OdCP and ζ τ ∞ linked as in formula (1.8) of [2]. Therefore, the polar parts of the transformed 1(c∞) and bF ⊗ OdCP shows the lemma. m (ξ, z) On the other hand, Lemma 6.6 implies that the parabolic weight correspond- m. This, however, means that the parabolic structure at m (ξ, z) is βj ing to ζ τ ∞ parabolic filtrations induced on Im(Aj ). As the filtration on coker(Aj) is by admissibility the canonical filtration, it follows that the parabolic structure at c∞ is the direct sum of the parabolic structures at all the points pj ∈ P of the c∞ is the direct sum of the parabolic structures at all the points pj ∈ P , ignoring the graded subspaces corresponding to the weight 0. 29 References [1] Bartocci C., Bruzzo U., Hern´andez Ruip´erez D. Fourier-Mukai and Nahm Transforms in Geometry and Mathematical Physics, Progress in Mathe- matics 276, Birkhauser (2009) [2] Biquard O., Boalch Ph. Wild nonabelian Hodge theory on curves, Compo- sitio Mathematica (1) 140 (2004), 179-207 [3] Biquard O., Szab´o Sz. Nahm transform and representations of the funda- mental group of a Riemann surface, in preparation. [4] Donaldson S., Kronheimer P. The Geometry of Four-Manifolds, Oxford Mathematical Monographs, Clarendon Press (1990) [5] Hitchin N. J. The self-duality equations on a Riemann surface, Proc. Lon- don Math. Soc. (3) 55 (1987), 59-126 [6] Katz N. M. Rigid local systems, Annals of Mathematics Studies 139, Princeton University Press (1996) [7] Katz N. M., Laumon G. Transformation de Fourier et majoration de sommes exponentielles, Publ. Math. IHES 62 (1985) [8] Katzarkov L., Kontsevich M., Pantev T. Hodge theoretic aspects of mirror symmetry, in: Proceedings of Symposia in Pure Mathematics 78, From Hodge Theory to Integrability and TQFT, Donagi R. Y., Wendland K. (eds.) [9] Levelt A. H. M., Jordan decomposition for a class of singular differential operators, Ark. Mat. 13 (1975), 1-27 [10] Malgrange B. ´Equations diff´erentielles `a coefficients polynomiaux, Progress in Mathematics 96, Birkhauser (1991) [11] Sabbah C. simplement ules l'alg`ebre http://math.polytechnique.fr/∼sabbah/articles.html de Weyl, Transformation exponentiels to letter de avec the Laplace pour les structure author, parabolique downloadable mod- sur from [12] Sabbah C. Isomonodromic Deformations and Frobenius Manifolds, Univer- sitext, Springer (2007) [13] Sabbah C. Harmonic metrics and connections with irregular singularities, Ann. Inst. Fourier Grenoble 49 (1999), 1265-1291 [14] Simpson C. Harmonic bundles on noncompact curves, J. Am. Math. Soc. (3) 3 (1990), 713-770 [15] Szab´o Sz. Nahm transform for integrable connections on the Riemann sphere, M´emoires de la Soci´et´e Math´ematique de France 110 (2007) 30 [16] Zucker, S. Hodge theory with degenerating coefficients: L2 cohomology in the Poincar´e metric, Ann. Math. 109 (1979), 415-476 31
0912.5461	2	0912	2011-02-09T18:42:05	Wonderful models for toric arrangements	[ "math.AG", "math.CO" ]	We build a wonderful model for toric arrangements. We develop the "toric analog" of the combinatorics of nested sets, which allows to define a family of smooth open sets covering the model. In this way we prove that the model is smooth, and we give a precise geometric and combinatorial description of the normal crossing divisor.	math.AG	math	Wonderful models for toric arrangements Luca Moci∗ October 24, 2018 Abstract We build a wonderful model for toric arrangements. We develop the "toric analogue" of the combinatorics of nested sets, which allows to define a family of smooth open sets covering the model. In this way we prove that the model is smooth, and we give a precise geometric and combinatorial description of the normal crossing divisor. 1 Introduction In the spirit of the much studied, ingenious construction of wonderful models for arrangements of linear subspaces by De Concini and Procesi, we here present a construction of wonderful models for toric arrangements. The latter have attracted much interest in recent years, since they proved to be deeply related with a wide number of topics, including vector partition functions and integral points in polytopes ([3], [4]), matroids and zonotopes ([14]), Lie theory ([9], [11], [13]), and index theory ([5]). Toric arrangements can be viewed as periodic counterparts of linear ar- rangements. Accordingly, their theory is much inspired by the theory of hyperplane arrangements as it evolved over the last decades (see in partic- ular [6], [3], [15]). The quest for a wonderful model of toric arrangements is then a natural next step. Let T be a complex torus and Λ its group of characters. Let (cid:101)X be a finite subset of Λ× C∗. For every pair (λ, a) ∈ (cid:101)X we define the hypersurface of T : Hλ,a The collection = {t ∈ Tλ(t) − a = 0} . . (cid:111) (cid:110) Hλ,a, (λ, a) ∈ (cid:101)X is called the toric arrangement defined by (cid:101)X on T . T(cid:101)X . = ∗Institut fuer Mathematik, Technische Universitaet Berlin, Strasse des 17. Juni, 136, Berlin. [email protected] 1 Let R(cid:101)X be the complement of the arrangement: = T \ (cid:91) (λ,a)∈(cid:101)X . R(cid:101)X Hλ,a. open set with complement a normal crossing divisor D, and a proper map In this paper we build a smooth minimal model Z(cid:101)X containing R(cid:101)X as an π : Z(cid:101)X → T extending the identity of R(cid:101)X . We call Z(cid:101)X the wonderful model of T(cid:101)X , in analogy to the wonderful model built by De Concini and Procesi The model Z(cid:101)X has several applications. In particular, it potentially is a powerful tool to describe the cohomology ring H∗(R(cid:101)X , Q). This computa- [2] for arrangements of subspaces in a vector (or projective) space. tion is one of the central, most outstanding questions in the theory of toric arrangements. In the special case of arrangements defined by unimodular lists of vectors, it has been answered in [10], [3]. We plan to face this prob- lem in a future paper, by applying to the model Z(cid:101)X the general method described in [16]. The paper is organized as follows. In Section 2 we give the first def- initions, we make some basic remarks and we build the wonderful model. In Section 3 we develop the necessary combinatorial tools, i.e the "toric analogues" of the notions of irreducible set, building set, nested set, and adapted basis. In Section 4 we define some smooth open sets of the model and we prove that they cover Z(cid:101)X . In Section 5 the open sets are used to prove that the complement of R(cid:101)X in Z(cid:101)X is a normal crossing divisor, and to describe its irreducible components and their intersections (see Theorem 5.3). Acknowledgements I wish to thank my supervisor Corrado De Concini for many illuminating suggestions and helpful remarks. I am also grateful to Maria Angelica Cueto, Jacopo Gandini and Bernd Sturmfels for stimulating discussions. 2 First definitions and remarks 2.1 Toric arrangements Let Λ be a lattice of rank n and U = Λ ⊗Z C the complex vector space obtained by extending the scalars of Λ. Let (cid:101)X be a finite set in Λ × C∗, and set = {λ(λ, a) ∈ (cid:101)X}. . X Given A ⊆ X, we denote by (cid:104)A(cid:105)Z and (cid:104)A(cid:105)C respectively the sublattice of Λ and the subspace of U spanned by A. We will always assume the 2 sublattice (cid:104)X(cid:105)Z to have finite index in Λ; otherwise we can replace Λ with Λ ∩ (cid:104)X(cid:105)C. Then we define = Hom(Λ, C∗). . T The group T is isomorphic to (C∗)n, and its group of characters Hom(T, C∗) Indeed given λ ∈ Λ and t ∈ T , we can take any is identified with Λ. representative ϕt ∈ Hom(Λ, C) of t and set For every pair (λ, a) ∈ (cid:101)X we define: λ(t) . = e2πiϕt(λ). = {t ∈ Tλ(t) − a = 0} . . Hλ,a We remark that in general the hypersurfaces Hλ,a are not connected; and even if they are, their intersections are not (see Remark 2.1 and Example 2.2 all the intersections of the hypersurfaces Hλ,a. This is a poset (with respect to inclusion) which plays a major role in the study of toric arrangements, for many aspects analogous to that of the intersection poset for hyperplane below). Then we consider the set C((cid:101)X) of all the connected components of arrangements (see [14, Section 5]). We call the elements of C((cid:101)X) the layers of the arrangement. Under our assumptions, the minimal elements of C((cid:101)X) are 0-dimensional, hence they are points. We denote by C0((cid:101)X) the set of such layers, which we call the points of the arrangement. For every layer C we define (cid:101)XC . = (cid:111) . (cid:110) (λ, a) ∈ (cid:101)XHλ,a ⊇ C = {λ(λ, a) ∈ (cid:101)XC}. . and The natural surjection (cid:101)XC −→ XC is indeed a bijection, since the condition (λ, a), (λ, b) ∈ XC implies that λ is identically equal to a = b on C. XC 2.2 Primitive vectors Given a system of coordinates (t1, . . . , tn) on T , for every we have a map ν = (ν1, . . . , νn) ∈ Zn e(ν) : T → C∗ (t1, . . . , tn) (cid:55)→ t1 ν1 · . . . · tn νn. It is well known that e is an isomorphism between Zn and Λ = Hom(T, C∗). 3 We will assume every λ ∈ X to be primitive, i.e. Λ ∩ (cid:104)λ(cid:105)C = (cid:104)λ(cid:105)Z. This amounts to require that under the previous isomorphism λ is identified with a vector ν = (ν1, . . . , νn) ∈ Zn such that GCD({νi}) = 1. Remark 2.1. This is not a restrictive assumption; indeed, suppose GCD({νi}) = d > 1 and write ν(cid:48) i . = νi/d. Then (cid:16) ν1 · . . . · tn νn − a = t1 1 · . . . · tn ν(cid:48) ν(cid:48) n t1 (cid:17)d − a = d(cid:89) (cid:16) i=1 1 · . . . · tn ν(cid:48) t1 √ n − ζ i d ν(cid:48) a (cid:17) d(cid:71) where ζ is a primitive d−th root of 1. Then there is a primitive element λ(cid:48) of Λ such that λ = dλ(cid:48), and we can write Hλ,a as the union of its connected components: Hλ,a = Hλ(cid:48),ζi d√ a. Then we can replace every pair (λ, a) ∈ (cid:101)X with all the pairs (λ(cid:48), ζ ia). In this way we get a new set (cid:101)X(cid:48) which defines the same toric arrangement as (cid:101)X. i=1 Example 2.2. Take T = (C∗)2 with coordinates (t, s) and (cid:101)X = {(t2, 1), (s2, 1), (ts, 1), (ts−1, 1)}. Since t2−1 = (t+1)(t−1), the hypersurfaces Ht2 and Hs2 have two connected components each; Hts and Hts−1 are connected, but their intersection is not. The points of the arrangement are: p1 = (1, 1), p2 = (−1,−1), p3 = (1,−1), p4 = (−1, 1). Notice that (cid:101)Xp1 = (cid:101)Xp2 = (cid:101)X, whereas Following Remark 2.1, we can replace (cid:101)X by (cid:101)Xp3 = (cid:101)Xp4 = {(t2, 1), (s2, 1)}. (cid:101)X(cid:48) = {(t, 1), (t,−1), (s, 1), (s,−1), (ts, 1), (ts−1, 1)}. 4 2.3 Construction of the model Given a sublattice ∆ ⊂ Λ, we define its completion For every layer C ∈ C((cid:101)X), we consider the lattice ΛC ∆ = (cid:104)∆(cid:105)C ∩ Λ. . = (cid:104)XC(cid:105)Z and its . completion ΛC. Remark 2.3. The elements of ΛC are the characters taking a constant value on C. Indeed, for every λ ∈ ΛC, we have that dλ ∈ ΛC for some d > 0. Then by definition dλ takes a constant value a on C; hence λ(t)d = a ∀ t ∈ C. Since C is connected and the set of dth roots of unity is discrete, the con- tinuous map λ must be constant. Now let λ1, . . . , λk be an integral basis of ΛC (i.e., a basis spanning over Z the lattice ΛC), and let ai be the constant value assumed by λi on C: then the ideal IC of the regular functions on T that vanish on C is generated by {λ1 − a1, . . . , λk − ak} and the normal space to C in T is NT (C) (cid:39) (cid:18) IC (cid:19)∗ I2 C . We denote by PC its projectified P(NT (C)) and by ϕC the natural map ϕC : T \ C → PC t (cid:55)→ [λ1(t) − a1, . . . , λk(t) − ak]. Now let us fix a subset G ⊆ C((cid:101)X). By collecting the maps {ϕC, C ∈ G} and the inclusion j : R(cid:101)X (cid:44)→ T , we get a map iG = j × (cid:89) ϕC : R(cid:101)X → T × (cid:89) C∈G PC C∈G We define Z(cid:101)X,G as the closure iG(R(cid:101)X ) of the image of R(cid:101)X . In the next section we will describe the subsets G that give arise to models with good geometric properties. Remark 2.4. 5 1. If we choose another basis λ(cid:48) k, we get other generators 1, . . . , λ(cid:48) 1 − a(cid:48) 1, . . . , λ(cid:48) {λ(cid:48) k − a(cid:48) k} of the same ideal IC, hence another basis of IC/I2 C and then another system of projective coordinates for PC. Then our construction does not depend on such choice. 2. In fact Z(cid:101)X,G can be obtained by a sequence of blow-ups along the 3. Since (cid:81) elements of G, listed in any dimension-increasing order (see [12]). C∈G PC is a projective variety, the restriction π : Z(cid:101)X,G → T image is closed in R(cid:101)X ×(cid:81) 4. Since iG is injective, we identify R(cid:101)X with its image iG(R(cid:101)X ). Such C∈G PC; therefore Z(cid:101)X,G contains R(cid:101)X as a dense open set, and the restriction of π to R(cid:101)X is j. of the projection on the first factor T is a projective and thus proper map. C∈G PC, which is open in T ×(cid:81) 2.4 Hyperplane arrangements and complete sets Given a finite set A ⊆ U , a hyperplane arrangement H(A) is defined in the dual space V = U∗ by taking the orthogonal hyperplane to each element of A. To every subset B ⊆ A is associated the subspace B⊥ of V that is the intersection of the corresponding hyperplanes of H(A); in other words, B⊥ is the subspace of vectors that are orthogonal to every element of B. Then we set L(A) = {B⊥, B ⊆ A}. L(A) is a geometric lattice, called the intersection poset of H(A). Its ele- ments are called the flats of the arrangement. Given a subset B ⊂ A, we define its completion = (cid:104)B(cid:105)C ∩ A. . B We say that B is complete in A if B = B. In the language of matroid theory, the complete subsets of A are the flats of the associated matroid, and the flat B is the closure of the subset B. For every Q ∈ L(A), let α(Q) be the set of elements of A which are identically equal to 0 on Q; clearly α(Q)⊥ = Q and α(B⊥) = B. Hence we have a bijection between L(A) and the family of complete subsets of A. 6 Fix p ∈ C0((cid:101)X). For every pair (λ, a) ∈ (cid:102)Xp, λ − a ∈ Ip defines a vector in Ip/I2 p and hence a hyperplane in its dual, which is the normal space to the point, i.e. the tangent space T (p) to p in T . This hyperplane of T (p) is simply the tangent space to the hypersurface H(λ,a) in p. In this way Xp defines in T (p) a hyperplane arrangement Hp, which is locally isomorphic (in 0) to our toric arrangement (in p). Then the map C (cid:55)→ (XC)⊥ is an inclusion-preserving bijection between layers C ∈ C((cid:101)X) containing p XC = α(cid:0)(XC)⊥(cid:1) is a complete subset of Xp. Conversely, for every complete and flats of Hp. Remark 2.5. In particular we see that, for every layer C containing p, subset A of Xp there is a unique layer C(A) such that XC(A) = A and p ∈ C(A). Namely, C(A) is the connected component containing p of the subvariety of T = {t ∈ T λ(t) − λ(p) = 0 ∀ λ ∈ A} . . HA 3 Combinatorial definitions Let B be a finite subset of Λ. An integral decomposition of B is a partition 3.1 Irreducible sets B =(cid:83) A complex decomposition of B is a partition B =(cid:83) i Bi such that (cid:104)B(cid:105)Z = (cid:104)Bi(cid:105)Z. (cid:77) (cid:77) i i Bi such that (cid:104)B(cid:105)C = (cid:104)Bi(cid:105)C. i We say that B is Z−irreducible (resp. C−irreducible) if it does not have We say that a layer C ∈ C((cid:101)X) is Z−irreducible (resp. C−irreducible) a nontrivial integral (resp. complex) decomposition. if XC is. We denote by I (resp. by IC) the set of Z−irreducible (resp. C−irreducible) layers. Remark 3.1. Clearly every integral decomposition is also a complex de- composition, but not conversely: see the example below. Then in general IC (cid:40) I. In the language of [12], C((cid:101)X) is a conical stratification on T , and IC is the set of the irreducible strata. Then a minimal wonderful model can be obtained by blowing up (in any dimension-increasing order) the elements of IC. However, in this model the intersections of irreducible components 7 of the normal crossing divisor fail to be connected (see example below). In order to obtain such property (i.e. the last point of Theorem 5.3), we will blow up all the elements of I. Example 3.2. Take T = (C∗)2 with coordinates (t, s) and (cid:101)X =(cid:8)(ts, 1), (ts−1, 1)(cid:9). Then X is identified with the subset {(1, 1), (1,−1)} of Z2. Thus X is not C−irreducible, but it is Z−irreducible: indeed Z(1, 1) ⊕ Z(1,−1) is a sublattice of index 2 in Z2. The hypersurfaces Hts and Hts−1 are the irreducible components of a normal crossing divisor; however their intersection consists of two points. By blowing them up we optain a model whose normal crossing divisor has four irreducible components, pairwise intersecting in a single point (as in the picture below): We now prove some properties of integral decompositions, which are known (and easier to prove) for complex decompositions (see for instance [4, Chapter 20.1]). From now on we will simply call decompositions the integral decom- layers) the Z−irreducible sets (resp. positions, and irreducible sets (resp. layers). Lemma 3.3. Let B = B1 ∪ B2 be a decomposition and D ⊂ B be an irreducible subset. Then D ⊆ B1 or D ⊆ B2. = D∩B2. We must prove that D = D1∪D2 . Proof. Set D1 is a decomposition; then the irreducibility of D implies that D1 or D2 is empty. We first notice that = D∩B1 and D2 . (cid:104)D(cid:105)Z = (cid:104)D1(cid:105)Z ⊕ (cid:104)D2(cid:105)Z since (cid:104)D1(cid:105)Z ∩ (cid:104)D2(cid:105)Z ⊆ (cid:104)B1(cid:105)Z ∩ (cid:104)B2(cid:105)Z ⊆ (cid:104)B1(cid:105)Z ∩ (cid:104)B2(cid:105)Z = {0}. Then take any λ ∈ (cid:104)D(cid:105)Z. For some positive integer m we have that mλ ∈ (cid:104)D(cid:105)Z and then can be written uniquely as mλ = µ1 + µ2, with µ1 ∈ (cid:104)D1(cid:105)Z and µ2 ∈ (cid:104)D2(cid:105)Z. Moreover, since λ ∈ (cid:104)B(cid:105)Z = (cid:104)B1(cid:105)Z ⊕ (cid:104)B2(cid:105)Z 8 λ can be expressed uniquely as λ = γ1 +γ2, with γ1 ∈ (cid:104)B1(cid:105)Z and γ2 ∈ (cid:104)B2(cid:105)Z. Then mλ = mγ1 + mγ2 = µ1 + µ2 implies µ1 = mγ1 and µ2 = mγ2, hence γ1 ∈ (cid:104)D1(cid:105)Z and γ2 ∈ (cid:104)D2(cid:105)Z. Thus (cid:104)D(cid:105)Z = (cid:104)D1(cid:105)Z ⊕ (cid:104)D2(cid:105)Z. subsets Bi. This decomposition is unique up to the order. Lemma 3.4. Every subset B has a decomposition B =(cid:83) Bi into irreducible Proof. The existence is clear by induction. Now let B =(cid:83) B(cid:48) j be another decomposition into irreducible subsets. By the previous lemma every Bi is contained in some B(cid:48) j and vice versa. Then these factors are the same up to the order. 3.2 Building sets and nested sets of layers Bi of G∗ contained in B. Then we say that B =(cid:83) We now recall some general definitions given in [2] and [4, Chapter 20.1], adapting them to our situation. Let A be a finite subset of Λ. A family G∗ of subsets of A is a building set if every complete subset B of A is decomposed by the maximal elements i Bi is the decomposition of B in G∗ or that the Bis are the G∗−factors of B. A subset S∗ of G∗ is a G∗−nested set if given any B1, . . . , Br ∈ S∗ mutually incomparable, = B1 ∪ . . . ∪ Br . B is a complete set in A with its decomposition in G∗. By [12], an equivalent definition is the following. A flag F∗ is a sequence A1 ⊂ ··· ⊂ Ak of subsets of A. A set S∗ = {B1, . . . , Bs} is G∗−nested if there is a flag F∗ such that all the elements of S∗ are G∗−factors of elements of F∗. The family I∗ of all irreducible subsets of A is clearly a building set. In particular, we call nested sets the I∗−nested sets. Then a nested set is a family S∗ of irreducible subsets such that for every B1, . . . , Br ∈ S∗ mutually incomparable, = B1 ∪ . . . ∪ Br . B is a complete set in A with its decomposition into irreducible subsets. Now let p ∈ C0((cid:101)X) be a point of the arrangement, and let C be any layer containing p. Let G∗ be a building set in Xp, and let XC = (cid:83) i Xi be the decomposition of XC in G∗. By Remark 2.5, there is a unique layer . Ci = C(Xi) containing C and such that XCi = Xi. We call the Cis the G−factors of C; clearly C = ∩Ci. 9 layers G defined as the set of all the G−factors of all the elements of C((cid:101)X). Then we can associate to every building set G∗ in A a building set of In particular for G∗ = I∗ we get that the set I of all irreducible layers is a building set. A flag F of layers is a sequence D1 ⊂ ··· ⊂ Dk. A set of layers S = {C1, . . . , Cs} is G−nested if there is a flag F such that all the elements of S are G−factors of elements of F. We say that S is a nested set of layers if it is I−nested, i.e. if there is a flag F such that all the elements of S are irreducible factors of elements of F. center of S. This is a well defined layer by the following Lemma: Lemma 3.5. Let S be a G−nested set. Then We call the minimal element of the flag (with respect to inclusion) the C(S) . = C (cid:92) C∈S (cid:92) is connected (and then is a layer). Proof. Let M (S) be the set of minimal elements of S, with respect to inclu- sion. Clearly The elements of M (S) are pairwise incomparable, hence C(S) = C∈M (S) C. (cid:77) (cid:88) C∈S ΛC(S) = ΛC = ΛC. C∈M (S) Let us choose an integral basis bC for each of the lattices ΛC, C ∈ M (S). Then (cid:91) b = bC C∈M (S) is an integral basis for ΛC(S). For any λ ∈ ΛC, λ takes a constant value aλ on C by Remark 2.3. It follows that the elements λ− aλ, λ ∈ b generate the ideal of definition of C(S), which is clearly irreducible since b is a basis of a split direct summand in Λ. Remark 3.6. Notice that our proof clearly implies that the intersection C(S) = is transversal. (cid:92) C C∈M (S) 10 A G−nested set of layers is maximal if it is not contained in a larger one; this happens if and only if S contains all the irreducible G−factors of a maximal flag. In this case the center of S is a point p = p(S). We denote by M the set of all maximal G−nested set of layers of C((cid:101)X) and by Mp the set of those having center p. Then we have the partition (cid:71) p∈C0((cid:101)X) M = Mp. The following fact is clear from the definitions (and from Remark 2.5): Lemma 3.7. If S = {C1, . . . , Cs} ∈ Mp is a maximal G−nested set of layers of center p, then S∗ . = {XC1, . . . , XCs} is a maximal G∗−nested set in Xp. Conversely, given a maximal G∗−nested set (cid:98)S in Xp, there is a unique S ∈ Mp such that S∗ = (cid:98)S; namely (cid:110) C(Ai), Ai ∈ (cid:98)S(cid:111) S . = . In particular S = S∗ = n, the rank of X (see [4, Theor 20.9]). Finally we prove an elementary result that we will use frequently in the next sections. Take S ∈ Mp. Lemma 3.8. 1. Let C ∈ I and p ∈ C. Then there is an element C ∈ S which is the maximum among all the elements of S contained in C; we call it the S−core of C. 2. Let C be an element of S which is not minimal in it. Then there is an element s(C) ∈ S which is the maximum among all the elements of S properly contained in C; we call it the successor of C. Proof. The proof is the same for both statements. Let C(cid:48) and C(cid:48)(cid:48) be two elements of S which are contained (or, for the second statement, properly contained) in C. Then XC ⊂ XC(cid:48) ∩ XC(cid:48)(cid:48); hence XC(cid:48) ∪ XC(cid:48)(cid:48) is not a decom- position. Since XC(cid:48) and XC(cid:48)(cid:48) are in the G∗−nested set S∗, they must be comparable; then also C(cid:48) and C(cid:48)(cid:48) are. 3.3 Adapted bases Given a G−nested set S, we say that an integral basis b = {λ1 . . . , λn} for . the lattice Λ is adapted to S if for every C ∈ S, b ∩ ΛC is an integral basis for ΛC. 11 Lemma 3.9. There exists an integral basis bS for Λ adapted to S. (cid:88) (cid:77) D∈S ΛD C∈M (S) Proof. Let us define ΛS . = ΛD. Notice that ΛS = where M (S) is the set of minimal (and hence pairwise incomparable) ele- ments of S. Then by definition ΛS = ΛS. We will prove, by induction on the cardinality of S, that there is a basis of ΛS adapted to S. Then our claim follows: indeed, since the lattice ΛS either coincide with Λ or is a split direct summand of it, the basis of ΛS can be completed to a basis of Λ. If S contains only one element C, the statement is trivial since ΛS = ΛC and every basis of this lattice is adapted to S. Otherwise, take a minimal C ∈ S, and set S(cid:48) = S \ {C}. Since S(cid:48) is G−nested, by inductive hypothesis the lattice (cid:88) D∈S(cid:48) ΛS(cid:48) = ΛD has an integral basis adapted to S(cid:48). Since ΛS(cid:48) = ΛS(cid:48) we can complete the chosen basis of ΛS(cid:48) to an integral basis b of ΛS using elements of ΛC. We claim that this basis is adapted to S. Let us take D in S. If D (cid:54)= C there is nothing to prove. Then assume D = C. In this case we know that ΛS = ΛC ⊕ (cid:77) ΛD. D∈M (S)\{C} By construction, every element in b either lies in ΛC or in (cid:77) ΛD. D∈M (S)\{C} Then every λ ∈ ΛC is in the span of b ∩ ΛC, proving our claim. To every maximal G−nested set of layers S ∈ Mp we associate a function pS : Λ −→ S in the following way. For every λ ∈ Λ we set a . = λ(p), and we define pS(λ) as the maximum element of S on which λ is identically equal to a. This is well defined by Lemma 3.8: indeed pS(λ) = H(λ,a). This function has the following properties: Lemma 3.10. 12 1. For every C ∈ G there exists λ ∈ XC such that pS(λ) = C. 2. The restriction of pS to an adapted basis b is a bijection. Proof. For every C ∈ G, let M (C) be the (possibly empty) set of the el- ements of S properly containing C and minimal with this property. Such is a decomposition. Since XC ⊃ XD for every D ∈ M (C), XD D∈M (C) elements are pairwise incomparable, hence(cid:91) XC ⊃ (cid:91) λ ∈ XC \ (cid:91) D∈M (C) XD XD. D∈M (C) and this inclusion is proper, because XC ∈ G∗. Then there exists Now assume C ∈ S, and let b be an adapted basis to S: then by definition By definition pS(λ) = C, then the first statement is proved. b ∩ ΛC is a basis for ΛC and (cid:77) ΛD. D∈M (C) (cid:71) D∈M (C) (cid:0)b ∩ ΛD (cid:1) is a basis for ΛC (cid:41) (cid:77) (cid:1) \ (cid:71) λ ∈(cid:0)b ∩ ΛC D∈M (C) ΛD. D∈M (C) (cid:0)b ∩ ΛD (cid:1) . Since C ∈ G, we have that Then there exists Clearly pS(λ) = C. Then we proved that the restriction of pS to b is surjec- tive; therefore it is bijective, since b = n = S. Remark 3.11. From now on we will assume for simplicity G = I, and then = Z(cid:101)X,I defined as the closure of the image we will focus on the model Z(cid:101)X iI = j × (cid:89) ϕC : R(cid:101)X → T × (cid:89) of the map PC. . C∈I C∈I However, all the results in this paper may be extended to the case of an arbitrary building set G. 13 4 Open sets and smoothness 4.1 Definition of the open sets To every S ∈ Mp we associate a nonlinear change of coordinates fS and an open set VS defined as follows. Let us take a basis of Λ adapted to S, and denote it by bS = (λC)C∈S = λC(p). Since bS is integral, (λC − aC)C∈S is . . = p−1S (C). Set aC where λC a system of coordinates on T . = D⊆C (cid:102)US . Consider Cn with coordinates zS = (zC)C∈S, and its open set zD (cid:54)= −aC ∀C ∈ S  . Define a map fS : (cid:102)US → T in the given coordinates as  + aC (zC) ∈ Cn (cid:89) (cid:89) (cid:0)fS(zS)(cid:1) = (cid:89) D⊆C λC zD or equivalently as the nonlinear change of coordinates λC − aC = D⊆C zD. (1) Then fS(0) = p. Notice that on the open set of T where λC − aC (cid:54)= 0∀C ∈ S, the map fS can be inverted by the following formula: zC = λC − aC λC −aC λs(C)−as(C) , if C is minimal in S , otherwise (2) (cid:40) where s(C) is the successor defined in Lemma 3.8. Let us define the open set of T = T \ (cid:91) . Tp C p /∈C and set US . = fS−1(Tp). We denote again by fS the restriction US → Tp. Now take any λ ∈ Λ; set a Since bS is adapted to S, an integral basis for ΛC is given by . = λ(p) and C . = pS(λ). bS ∩ ΛC = {λD, D ⊇ C} . 14 In particular λ can be expressed in this basis, and since pS(λ) = C, λ does not lie in the span of {λD, D (cid:41) C}: then λ = mCλC + mDλD for some integers mD and a nonzero integer mC. The previous identity, considered as an equality of regular functions on T , can be written as (cid:88) D(cid:41)C (cid:89) D(cid:41)C λ = λmC C λmD D . C (cid:89) λ − a = Then we have: D − amC λmD C (cid:89) (cid:89) λmC amC  + (cid:89) (cid:1) = βC(λC − αC) and we can write the first summand as C − amC (cid:89) (cid:0)λmC (cid:89) λmD D λmD D where D(cid:41)C D(cid:41)C D(cid:41)C C C D(cid:41)C (λC − ζaC) . = βC λmD D D(cid:41)C ζmC =1,ζ(cid:54)=1  (3) D − a λmD is a regular function on T which is invertible on C. Working in the same way on the second summand of Formula (3) we see that, for some regular functions {βD, D ∈ S}, λ − a = βC (λC − aC) + βD(λD − aD). (cid:88) D(cid:41)C By operating the change of coordinates (1), we get: βC (cid:89) E⊆C λ − a = where we set zE + βD zE  · pλ(zS) zE (4) (cid:89) (cid:88) E⊆D (cid:89) E⊆C  = (cid:89) pλ(zS) . = βC + βD zE. D(cid:41)C D⊇E(cid:41)C We define VS as the open set of US where pλ(zS) (cid:54)= 0. Let us remark that 0 ∈ VS, since for every λ ∈ Xp we have that pλ(0) = βC(p) (cid:54)= 0. Furthermore in VS, for every λ ∈ Xp, we have the equality of regular functions . (5) (cid:88) D(cid:41)C (cid:89) λ∈Xp (cid:89) λ − a pλ(zS) E⊆pS (λ) zE = 15 4.2 Properties of the open sets Let us define the open set of VS VS 0 . = {z ∈ VS zC (cid:54)= 0∀C ∈ S}. We denote by AS the open set of T given by fS(VS) ∩ R(cid:101)X . We remark that by Formula (5) f−1S (AS) = VS 0 and the restriction of fS to VS 0 maps it into AS. By composing this map with the inclusion AS (cid:44)→ R(cid:101)X and with the application φC : R(cid:101)X → PC defined in Section 2.3, we get a map ψC : VS 0 −→ PC. (cid:102)ψC : VS → PC. Lemma 4.1. For every C ∈ I and S ∈ Mp, the map ψC extends uniquely to a map Proof. Let p be the center of S. If C does not contain p the statement is clear: indeed since VS ⊂ US, for every u ∈ VS we have that t = fS(u) /∈ C . so that for at least one index j, λj(t) (cid:54)= aj. Then the projective coordinate λj(t) − aj of PC is nonzero. Then assume p ∈ C, and let C be its S−core (see Lemma 3.8). By the first part of Lemma 3.10, there exists λ1 ∈ XC such that pS(λ1) = C. Since we assumed (Remark 2.1) every element of XC to be primitive, we can complete {λ1} to an integral basis {λ1, . . . , λk} of ΛC. Then if we set ai . = λi(p), we have that [λ1 − a1, . . . , λk − ak] is a system of projective coordinates for PC. Since bS is adapted to S, an integral basis for ΛC is given by bS ∩ ΛC =(cid:8)λD, D ⊇ C(cid:9) . pS(λ1) = C, λ1 does not lie in the span of(cid:8)λD, D (cid:41) C(cid:9) . we can divide every projective coordinate by(cid:81) After making the nonlinear change of coordinates (1) as in Formula (4), E⊆C zE; in this way we get In particular every λi ∈ ΛC ⊆ ΛC can be expressed in this basis, and since that the map ψC : VS 0 −→ PC is given by pλ1(z), pλ2(z) z (cid:55)→ (cid:89) (cid:89)  C(cid:40)E⊆D2 zE, . . . , pλk (z) zE C(cid:40)E⊆Dk = pS(λi). Since by definition pλ1(z) (cid:54)= 0 for z ∈ VS, this . where we set Di map extends to VS. Moreover its image is contained in an affine open set of PC. VS 0 is dense in VS. Finally the uniqueness of the extension is clear since by its very definition 16 By applying the lemma above to all the layers C ∈ I, we get that for every S ∈ Mp the inclusion VS 0 (cid:44)→ Z(cid:101)X extends uniquely to a map jS : VS → Z(cid:101)X . Lemma 4.2. The map jS is an embedding into a smooth open set. Proof. In order to prove that jS is an embedding, it suffices to see that every coordinate zC on VS can be written as the composition of jS and a function on jS(VS). Then take C ∈ S. If C is not minimal, let D = s(C) be the successor of C. Since bS is adapted to S, on PD we have the projective coordinates [λE − aE]E∈S,E⊇D and by the proof of the previous lemma VS maps into the affine subset where λD − aD (cid:54)= 0. Then we can read the coordinate zC in PD by Formula (2): zC = λC − aC λD − aD . If on the other hand C is minimal in S, then zC = λC − aC. of jS(VS) ⊂ Z(cid:101)X on T or on some PD; hence our map is an embedding. In this way all the coordinates zC can be recovered by the projection Moreover, since (zC)C∈S is a system of coordinates on jS(VS), in every point the differential of jS has rank S = n. Then jS(VS) is smooth. Remark 4.3. By abuse of notation, from now on we will write VS for jS(VS), identifying this set with its isomorphic image in Z(cid:101)X . 4.3 Smoothness of the model Let us define (cid:91) S∈M Y(cid:101)X . = VS. In this section we prove that Y(cid:101)X = Z(cid:101)X , and hence Z(cid:101)X is smooth. The main step is the following lemma, which tells that every curve in R(cid:101)X that "has limit" in T , "has limit" in Y(cid:101)X . Let Dε Lemma 4.4. Let f : Dε → T be a curve such that f (Dε \ {0}) ⊆ R(cid:101)X . Then f lifts to a curve in Y(cid:101)X . Proof. Given such a f , let Cf ∈ C((cid:101)X) be the smallest layer containing f (0), and let p ∈ C0((cid:101)X) be a point contained in Cf . For every λ ∈ Xp, we have = {s ∈ C s < ε}. . that locally, near s = 0, we can write λ(f (s)) − a = snλqλ(s) with a = λ(p), nλ ≥ 0 and qλ(0) (cid:54)= 0. 17 For every integer h ≥ 0, let us define = {λ ∈ Xpnλ ≥ h}. . Ah Notice that A0 = Xp and Ah+1 ⊆ Ah; by taking all the irreducible factors of the elements of this flag we get a nested set in Xp. Let us complete it to a maximal nested set S∗; by Lemma 3.7, to S∗ is naturally associated a maximal nested set of layers S ∈ Mp. We claim that for a such S, the curve f : Dε \ {0} → R(cid:101)X extends to a map f : Dε → VS. First notice that f (0) ∈ Tp: indeed for every layer D containing f (0) we have that Cf ⊆ D by minimality and then p ∈ D. Then we have to prove that: (cid:0)f (s)(cid:1) is defined in 0 for every C ∈ S; (cid:0)f (0)(cid:1) (cid:54)= 0 for every λ ∈ Xp. 1. zC 2. pλ Take C ∈ S; if C is minimal in S then zC(f (s)) = λC(f (s)) − aC and there is nothing to prove. Otherwise, let D = s(C) be the successor of C. Then by Formula (2) zC(f (s)) = λC(f (s)) − aC λD(f (s)) − aD = snλC −nλD qλC (s) qλD (s) and nλC ≥ nλD by the definition of S, so zC is well defined in 0. vector λC of the adapted basis bS. As for the second claim, given any λ ∈ Xp set C . = pS(λ) and take the Then by definition of S, nλ = nλC , and by Formulae (1) and (4) we have Therefore pλ(f (0)) = λ − a λC − aC pλ = λ(cid:0)f (0)(cid:1) − a (cid:0)f (0)(cid:1) − aC λC . = qλ(0) qλC (0) (cid:54)= 0. Theorem 4.5. Y(cid:101)X = Z(cid:101)X . In particular Z(cid:101)X is smooth. Proof. By the well known valuative criterion for properness (see for instance [8]), the previous lemma amounts to say that the map is proper. Since also the projection : Y(cid:101)X → T PC → T πY(cid:101)X T × (cid:89) C∈I 18 is proper, the embedding Y(cid:101)X → T × (cid:89) C∈I PC Z(cid:101)X . is proper as well; therefore its image is closed, and thus it coincides with Therefore Z(cid:101)X is smooth, since it is the union of smooth open sets. 5 The normal crossing divisor 5.1 Technical lemmas C, C ∈ S} and {zQ For every C ∈ I, let us define a divisor DC ⊂ Z(cid:101)X as follows. Take a S ∈ M such that C ∈ S. In the open set VS take the divisor of equation zC = 0; let DC be the closure of this divisor in Z(cid:101)X . The following lemma implies that DC does not depend on the choice of S, and yields the theorem below, which describes the geometry of Z(cid:101)X \ R(cid:101)X . Lemma 5.1. Take any two maximal nested sets of layers S ∈ Mp and C , C ∈ Q} be the corresponding sets of Q ∈ Mq. Let {zS coordinates on VS and VQ. Then for every C ∈ S: 1. if C ∈ S \ Q, zS C is regular and invertible as a function on VS ∩VQ. 2. if C ∈ S∩Q, zS Proof. If q /∈ C, then C ∈ S \ Q, and the (first) statement is proved as follows. Take x ∈ Z(cid:101)X such that zS C(x) = 0: then by Formula (1) π(x) ∈ C, where π : Z(cid:101)X → T is the projection defined in Remark 2.4. Therefore π(x) /∈ Tq, hence x /∈ VQ, proving the claim. Therefore we can assume q ∈ C and proceed by induction as in the proof C is invertible as a function on VS ∩ VQ; C/zQ of [4, Lemma 20.39]: • First let us assume C to be a minimal element in I; then necessarily C ∈ S ∩ Q. We recall that zS C = λS = pQ(λS . C − aS C; set C) ⊇ C. D Then for some function a zS C = a zQ E = azQ C zQ E . E∈Q,D⊇E,E(cid:54)=C (cid:89) (cid:89) E∈Q,D⊇E C = λQ In the same way zQ C − aQ C , and if we set C ) ⊇ C = pS(λQ D(cid:48) . 19 zQ F = a(cid:48)zS C F∈S,D(cid:48)⊇F,F(cid:54)=C for some function a(cid:48). Since both D and D(cid:48) contain C, by substituting we get: (cid:89) (cid:89) zS F . zS F . zS F = 1 zQ F∈S,D(cid:48)⊇F C = a(cid:48) (cid:89) C a a(cid:48) (cid:89) aa(cid:48) (cid:89) zS C = zS we get Therefore and hence E∈Q,D⊇E,E(cid:54)=C F∈S,D(cid:48)⊇F,F(cid:54)=C zQ E (cid:89) zQ E (cid:89) E∈Q,D⊇E,E(cid:54)=C F∈S,D(cid:48)⊇F,F(cid:54)=C zS zQ C C = a zQ E E∈Q,D⊇E,E(cid:54)=C is invertible, as claimed. • Now let us take any C ∈ S. By induction, we can assume that our claims are true for every D (cid:40) C, D ∈ S∪Q (if D ∈ Q\S, by symmetry D is assumed to be invertible on VS ∩ VQ). zQ Let D = C ∈ Q be the Q−core of C. Take λ ∈ XC such that = pS(λ). Then G ⊇ C and λ takes on D . pQ(λ) = D, and set G . = λ(p). Notice that D is the and on G the same constant value a Q−core of G. Then for some invertible b, b(cid:48) λ − a = b Hence −1 1 = b b(cid:48) (cid:89) F∈S\Q,G⊇F zS F E∈Q,D⊇E (cid:89) (cid:89) zQ E = b(cid:48) (cid:89) −1 (cid:89) F∈S,G⊇F zQ E zS F . E∈Q\S,D⊇E F∈S∩Q,D⊇F F zQ zS F −1 . (6) We can now prove the first claim. If C /∈ Q then D (cid:40) C. Then all the factors in equation (6) are regular: those of type F , F ∈ S \ Q, G ⊇ F zS obviously, the others by inductive assumption, since they involve ele- ments properly contained in C. Since zS C appears as one of the factors in (6) it is invertible. In the same way if C ∈ Q, and then D = C, all the factors in (6) but (eventually) zS must be regular and invertible. are regular; then also zS C zQ C zQ −1 −1 C C 20 Lemma 5.2. Let be C ∈ I. 1. The divisor DC is well defined. 2. If C /∈ S, then DC ∩ VS = ∅. Proof. 1. Let S,Q be two maximal nested set of layers containing C. Then by the second point of Lemma 5.1, zS C have the same zeros in VS ∩ VQ, which is an open dense set in VS and in VQ. Then the closures of the two divisors coincide. C and zQ 2. Let Q be a maximal nested set of layers containing C. Then by the C is invertible as a function on VS ∩ VQ. C = 0 is contained in Z(cid:101)X \VS. first point of Lemma 5.1, zQ Therefore the divisor of VQ defined by zQ Since this set is closed, it also contains DC which is the closure of the divisor. 5.2 The main theorem Now let us define (cid:91) C∈I DC. D = The geometry of the divisor D is described by the following theorem. Theorem 5.3. 2. D is a normal crossing divisor whose irreducible components are the 1. Z(cid:101)X \ D = R(cid:101)X . divisors DC, C ∈ I. 3. Let be N ⊆ I, and (cid:92) C∈N DN . = DC. Then DN (cid:54)= ∅ if and only if N is nested. 4. If N is nested, DN is smooth and irreducible. Proof. By Theorem 4.5, we can check each statement on every open set VS,S ∈ M. Then the first claim, by the second part of Lemma 5.2, amounts to note that (cid:0)Z(cid:101)X \ D(cid:1) ∩ VS = VS \ (cid:91) C∈S (DC ∩ VS) = VS 0 = R(cid:101)X ∩ VS. 21 (for the definition of VS 0 see the beginning of Section 4.2). The second statement is obvious since D ∩ VS = (DC ∩ VS) = {z ∈ VSzC = 0 for some C ∈ S} (cid:91) C∈S is by definition a normal crossing divisor in VS. For the third statement, note that if N is not nested it is not contained in any maximal nested set of layers; then for every S ∈ M, DN ∩ VS = ∅ by the second part of Lemma 5.2. On the other hand, if N is nested it can be completed to some S ∈ M, and DN ∩ VS = {z ∈ VSzC = 0∀C ∈ N} which is clearly nonempty, smooth and irreducible. Since (cid:91) S⊇N (DN ∩ VS) DN = also the last statement follows. References [1] M. A. Cueto, S. Lin, Tropical secant graphs of monomial curves, arXiv:1005.3364v1 [math.AG]. [2] C. De Concini, C. Procesi, Wonderful models of subspace arrange- ments, Selecta Mathematica 1 (1995), 459-494. [3] C. De Concini, C. Procesi, On the geometry of toric arrangements, Transformations Groups 10, N. 3-4, 2005. [4] C. De Concini, C. Procesi, Topics in hyperplane arrangements, polytopes and box-splines, Springer 2010. [5] C. De Concini, C. Procesi, M. Vergne, Vector partition functions and index of transversally elliptic operators, arXiv: math 0808.2545. [6] Douglass, J. M., Toral arrangements and hyperplane arrangements., Rocky Mountain J. Math. 28 (1998), no. 3, 939-956. [7] E. M. Feichtner, De Concini-Procesi arrangement models - a dis- crete geometer's point of view, in: Combinatorial and Computational Geometry, MSRI Publications 52, Cambridge University Press, 2005, pp. 333-360 22 [8] R. Hartshorne, Algebraic Geometry, Graduate Texts in Mathemat- ics, No. 52, Springer-Verlag, New York, 1977. [9] G. I. Lehrer, The cohomology of the regular semisimple variety, J. Algebra 199 (1998), no. 2, 666-689. [10] E. Looijenga, Cohomology of M3 and M1 3. Mapping class groups and moduli spaces of Riemann surfaces, (Gottingen, 1991/Seattle, WA, 1991), Contemp. Math., 150, Amer. Math. Soc., 1993, 205228. [11] Macmeikan, C., Modules of derivations for toral arrangements., Indag. Math. (N.S.) 15 (2004), no. 2, 25767. [12] R. MacPherson, C. Procesi, Making conical compactifications won- derful, Selecta Mathematica 4 (1998), 125-139. [13] L. Moci, Combinatorics and topology of toric arrangements defined by root systems, Rend. Lincei Mat. Appl. 19 (2008), 293-308. [14] L. Moci, A Tutte polynomial for toric arrangements, arXiv:0911.4823 [math.CO], to appear on Trans. AMS. [15] L. Moci, S. Settepanella, The homotopy type of toric arrange- ments, arXiv:1009.3622v2 [math.AT], to appear on J. of Pure and Appl. Algebra [16] J. Morgan, The algebraic topology of smooth algebraic varieties, Publ. Math. IHES, 48 (1978), 137-204. 23
1001.0607	1	1001	2010-01-05T16:54:26	Regularity of quotients by an automorphism of order $p$	[ "math.AG", "math.AC" ]	Let $B$ be a regular local ring and $G\subset\Aut(B)$ a finite group of local automorphisms. Assume that $G$ is cyclic of prime order $p$, where $p$ is equal to the residue characteristic of $B$. We give conditions under which the ring of invariants $A=B^G$ is again regular.	math.AG	math	Regularity of quotients by an automorphism of order p Stefan Wewers IAZD, Leibniz-Universitat Hannover Abstract Let B be a regular local ring and G ⊂ Aut(B) a finite group of local automorphisms. Assume that G is cyclic of prime order p, where p is equal to the residue characteristic of B. We give conditions under which the ring of invariants A = BG is again regular. Introduction Let X be a regular integral scheme and G ⊂ Aut(X) a finite group of auto- morphisms of X. The quotient scheme Y := X/G may not be regular; its singularities are, by definition, quotient singularities. To study the singularities of Y , we may localize and assume that X = Spec B and Y = Spec A, where B is a local domain and A = BG is the ring of invariants. Quotient singularities have been intensively studied, in connection with reso- lution of singularities and as objects in their own right. However, most results concern tame quotient singularities. In the above situation this means that the order of G is prime to the characteristic of the residue field of B. In a recent preprint [8], D. Lorenzini has studied the resolution graphs of wild quotient singularities in dimension 2, exhibiting many interesting features that do not occur for tame quotient singularities. His results rely heavily on a detailed combinatorial study of the possible intersection matrices that can occur. The present note aims at complementing the methods used in [8]. The basic idea is the following. Let Y = X/G be as above, and let Y ′ → Y be a resolution of Y . Then the fiber product X ′ := X ×Y Y ′ is a G-equivariant modification of X whose quotient Y ′ = X ′/G is regular. Conversely, given a G-equivariant modification X ′ → X, the quotient scheme Y ′ := X ′/G is a modification of Y -- which may or may not be regular. From this point of view it is natural to look for conditions on X ′ under which Y ′ is regular. However, it is difficult to find such conditions in the literature which apply to the case of wild group actions. Our main result (Theorem 1.3) gives a sufficient condition for the regularity of the quotient scheme Y = X/G when G is a cyclic group of order p. This 1 is admittedly a very modest contribution to our motivating problem. Still, our criterion seems to be new, and we hope that it will be useful in the future. The motivation for writing this paper grew out of discussions with F. Kir´aly, who discovered a special case of our main result, see Example 1.10. The author thanks him, W. Lutkebohmert and M. Raynaud for helpful discussions and comments on earlier versions of this paper. 1 The main result 1.1 Let (B, m, k) be a noetherian regular local domain, and let G ⊂ Aut(B) be a finite group of local automorphisms. We are interested in the following question. Problem 1.1 Under which condition is the ring of invariants A := BG = { x ∈ B σ(x) = x for all σ ∈ G} again regular? For the applications we have in mind, the following additional assumptions seem reasonable and useful: (a) The residue field k is algebraically closed. (b) The induced G-action on k is trivial. (c) A is noetherian and B is a finite A-algebra. Assumptions (a) and (b) can always be achieved by passing to the strict henseliza- tion and are therefore harmless. They imply that A is a local domain with residue field k. Condition (c) is satisfied in most situations arising from a geo- metric context. Suppose, for instance, that B is the localization of a finitely generated algebra over an excellent domain R, and that the action of G on B is R-linear. Then (c) holds. See [2] for a general discussion of Condition (c). If the order of the group G is prime to the characteristic of the residue field k, a definitive answer to Problem 1.1 is known. Theorem 1.2 Suppose that (a)-(c) holds and that the order of G is prime to the characteristic of the residue field k. Then the ring A is regular if and only if the image of G in GL(m/m2) is generated by pseudo-reflections1. Proof: See [12] or [14] ✷ The main result of the present paper (Theorem 1.3 below) gives a sufficient criterion for A to be regular in the case where G is cyclic of prime order p. Here p may be equal to the characteristic of k, and hence our result is not covered by Theorem 1.2 1An element σ ∈ GL(V ) is called a pseudo-reflection if rank(σ − 1) ≤ 1 2 Let us now assume that G is cyclic of order p, where p is prime. We 1.2 choose a generator σ ∈ G and consider the ideal Iσ := hσ(x) − x x ∈ BiB ✁ B. By definition, Iσ is the smallest G-invariant ideal such that G acts trivially on B/Iσ. In particular, this shows that Iσ does not depend on the chosen generator σ. Condition (b) says that Iσ is contained in m. An element x ∈ B is called a regular parameter if x ∈ m\m2. Since B is regular, this means that x is part of a regular system of parameters. It follows that B/Bx is a regular local ring and that (x) ✁ B is a prime ideal. Here is our main result: Theorem 1.3 Suppose that there exists a regular parameter π ∈ m\m2 such that (i) Iσ = (πδ), with δ ∈ N, (ii) we either have Iσ = (σ(π) − π), or else σ(π) − π ∈ (πδ+1). Then A = BG is regular. After some preliminary work done in Section 2 and 3, the proof of Theorem 1.3 will be given in Section 4. For the rest of this section, we discuss the scope and the limitations of Theorem 1.3 and some open problems. Remark 1.4 Let ¯σ ∈ GL(m/m2) denote the image of σ. By definition of Iσ, m/(Iσ + m2) is the largest quotient of m/m2 on which ¯σ acts trivially. If the hypothesis of Theorem 1.3 is verified, then it follows that ¯σ is a pseudo- reflection. So for char(k) 6= p, Theorem 1.3 is a direct consequence of Theorem 1.2. Moreover, in this case the natural homomorphism G → GL(m/m2) is known to be injective (see e.g. [14]). This implies that we must have δ = 1 in Condition (i) of Theorem 1.3 and that Iσ = (σ(π) − π) in Condition (ii). In On the other hand, if char(k) = p, δ may be strictly larger than 1. particular, the hypothesis of Theorem 1.3 is stronger then the condition '¯σ is a pseudo-reflection': it is easy to give examples where ¯σ = 1 and A is not regular. See e.g. [10]. Remark 1.5 The author suspects that Condition (ii) in Theorem 1.3 is implied by Condition (i) and could therefore be omitted. However, he did not succeed in proving this. Remark 1.6 The sufficient condition given by Theorem 1.3 is not necessary for A to be regular. See e.g. Example 1.8 below. 3 The Purity Theorem of Zariski-Nagata gives us a necessary condition for regularity. Namely, if A is regular, then all minimal prime ideals of B containing Iσ have height one, and are therefore principal. This prompts the following question: Question 1.7 Is it true that A is regular if and only if Iσ is a principal ideal? 1.3 We shall give three examples that illustrate various points. In all three examples, the ring B is a ring of power series over a complete discrete valuation ring R. In particular, B has dimension 2. We assume moreover that the action of the group G fixes the subring R ⊂ B. condition for regularity of the ring A. The first example shows that the hypothesis of Theorem 1.3 is not a necessary Example 1.8 Let R be a complete discrete valuation ring of mixed character- istic (0, p), containing a pth root of unity ζp. Set B := R[[x]], and let σ : B ∼→ B be the R-automorphism of order p given by Then is regular. However, the ideal σ(x) = ζp · x. A := Bhσi = R[[xp]] is contained in two distinct principal prime ideals, so Condition (i) of Theorem 1.3 fails. Iσ =(cid:0) (ζp − 1)x(cid:1) ✁ B Remark 1.9 Example 1.8 is a special case of the following situation. Assume that the ring B has a regular system of parameters (π1, . . . , πd−1, x) such that σ(πi) = πi for i = 1, . . . , d − 1 and σ(x) ≡ u · x mod (π1, . . . , πd−1) for a unit u ∈ B×. Then A = BG is regular, see [5], Proposition 7.5.2. In the situation of Theorem 1.3 a system of parameters (π1, . . . , πd−1, x) as above exists, but this is a consequence of the proof of Theorem 1.3, and is not obvious beforehand. The next example, which was first studied by F. Kir´aly [6], describes a special situation where the answer to Question 1.7 is affirmative. 4 Example 1.10 Let R and B = R[[x]] be as in Example 1.8. Let σ : B ∼→ B be an automorphism of order p which induces the identity on the residue field k. Contrary to Example 1.8, we assume that σ restricts to a nontrivial automorphism of the subring R ⊂ B. uniformizer, i.e. v(π) = 1. Our assumptions imply that Let v : R → Z ∪ {∞} denote the discrete valuation on R and let π ∈ R be a σ(π) = π + uπµ, with µ ≥ 2 and u ∈ R×. Write fσ := σ(x) − x = a0 + a1 x + a2 x2 + . . . , with ai ∈ R and set ν := min{ v(ai) i = 0, 1, . . .}. Since (π, x) is a system of parameters for B, we have Iσ = (πµ, fσ) ✁ B. It follows that Iσ is a principal ideal if and only if one of the following cases occurs: (I) µ ≤ ν, or (II) v(a0) = ν < µ. In Case (I) we have Iσ = (πµ) and in Case (II) we have Iσ = (πν ). In both cases, the hypothesis of Theorem 1.3 holds, and hence A = BG is regular. In this special case, the statement of Theorem 1.3 has been proved earlier by F. Kir´aly, and his results yield somewhat more. In Case (I), the ring of invariants A is actually a power series ring over RG. Moreover, if neither Case (I) nor Case (II) holds, then Iσ is not a principal ideal and the ring A is not regular. We refer to [6] for more details. The distinction of Case (I) and (II) in Example 1.10 illustrates a dichotomy which is crucial for the proof of Theorem 1.3. Let L := Frac(B) and K := Frac(A) denote the fraction fields. Then L/K is a Galois extension with Galois group G. Assume that the hypothesis of Theorem 1.3 holds. Let v denote the discrete valuation of L which corresponds to the localization of B at the prime ideal (π). It follows from Condition (i) that v is ramified in L/K. The two cases of Condition (ii) correspond to Case (I) and Case (II) in Example 1.10. We shall see in Section 3 that in the first case the extension L/K is totally ramified along v, whereas it is fiercely ramified in the second case (see Definition 3.1 and Proposition 3.2). 5 1.4 If we drop the assumption that the group G acting on B is cyclic of order p, the study of the ramification behavior of the extension L/K becomes much more complicated. The problem is that it is in general difficult to 'separate' a wildly ramified extension in a canonical way into subextensions which are either totally or fiercely ramified. This seems to be the main reason why the statement of Theorem 1.3 does not easily generalize to more general groups. Our last example illustrates this point. Example 1.11 Let R and B = R[[x]] be as in Example 1.8 and 1.10. Let G ⊂ Aut(B) be a finite group of automorphisms which fixes the subring R ⊂ B and acts trivially on the residue field k. We assume that G is an elementary abelian group of order p2, i.e. G ∼= Z/p × Z/p, with generators σ1, σ2. We also assume that the induced action of G on R is faithful. Let Ai := Bhσii, i = 1, 2. The restriction of σ1 to A2 is an automorphism of order p such that Aσ1 2 = A. Likewise, Aσ2 1 = A. Suppose that A1 and A2 are regular. Does it follow that A is regular? The statement of Theorem 1.2 may lead one to believe that the answer to this question is yes. However, this is not the case; in the following we shall give an explicit counterexample, where p = 2. Let K := Qnr 2 denote the maximal unramified extension of Q2. Let L/K be the Galois extension generated by the roots of the polynomial (x2 − 2)(x2 − 3). Hence L/K is generated by elements √2,√3 which are square roots of 2 and 3, respectively. Let G = Gal(L/K) denote the Galois group. Then G ∼= Z/2× Z/2 is generated by the two automorphisms σ1, σ2 determined by σ1(√2) = √2, σ1(√3) = − √3, σ2(√2) = − √2, σ2(√3) = √3. It follows that K1 := Q2(√2) = Lσ1, K2 := Q2(√3) = Lσ2. Let v denote the unique extension to L of the 2-adic valuation, normalized by v(2) = 1. Set π1 := √2, π2 := √3 − 1, π := π2 π1 − 1. One checks that It follows that v(π1) = v(π2) = 1 2 , v(π) = 1 4 . OK1 = Znr 2 [π1], OK2 = Znr 2 [π2], OK = Znr 2 [π]. Let B := OL[[x]] be the ring of power series in OL. We extend the action of G on OL to B by setting σ1(x) := x, σ2(x) = −x. 6 This is the example announced above. Indeed, it is easy to see that A1 = Bσ1 = OK1 [[x]] is a power series ring over OK1 and hence regular. Similarly, if we set y := (1 + π)x, then we find that B = OL[[y]], and σ2(y) = y. It follows that A2 = Bσ2 = OK2 [[y]] is regular, too. However, A = BG = Aσ2 is not regular. This can be checked 1 using the if-and-only-if criterion from Example 1.10. Note that (π1, x) is a regular sequence of parameters for A1. Now and so σ2(π1) − π1 = −2π1, σ2(x) − x = −2x, Iσ2 = (−2π1,−2x) = (2) · mA1 ✁ A1 is not a principal ideal. Using the criterion of [6] mentioned in Example 1.10, we conclude that A is not regular. The reader is invited to check this by a direct calculation of the dimension of mA/m2 A. 2 Derivations and p-cyclic inseparable descent In this section we prove an auxiliary result (Corollary 2.2) which is a crucial step in the proof of Theorem 1.3. The setup is similar as in the previous section; however, we work in equal characteristic p, and instead of an automorphism of order p we consider a derivation of the ring B. Let (B, m, k) denote a noetherian regular local ring of dimension d ≥ 1. We assume moreover that B is complete and has characteristic p. Then it follows from [9], Corollary 2 of Theorem 60, that B is isomorphic to a formal power series ring over k. More precisely, if (x1, . . . , xd) is a regular system of parameters for B, then we get an isomorphism B ∼= k[[x1, . . . , xd]]. We let L denote the fraction field of B. We write Derk(B) for the p-Lie-algebra of k-derivations θ : B → B. Note that any such derivation extends uniquely to a (continuous) derivation of L. Lemma 2.1 Let θ ∈ Derk(B) be a k-derivation of B such that the following holds: (i) the k-linear map induced by θ is not zero, ¯θ : m/m2 → k 7 (ii) we have θp = aθ, for some a ∈ Frac(B). Then there exists a regular system of parameters (x1, . . . , xd) for B such that and θ(x1) ≡ 1 (mod m) θ(x2) = . . . = θ(xd) = 0. Proof: Let θ be as in the statement of the lemma. It follows immediately from (i) that there exists a system of parameters x1, . . . , xd such that θ(x1) ≡ 1 (mod m), θ(xi) ≡ 0 (mod m) for i > 1. (1) Our strategy is to change the xi for i > 1 step by step in order to improve the last congruence modulo arbitrary powers of m. Since B is complete, we can take the limit and find parameters xi such that θ(xi) = 0 for i > 1. Suppose that (mod mn) for i > 1, θ(xi) ≡ 0 (2) for some n ≥ 1. We claim that there exist elements ∆2, . . . , ∆d ∈ mn+1 such that (3) (mod mn+1), i = 2, . . . , d. θ(∆i) ≡ −θ(xi) Assuming this claim, we can set xi := xi + ∆i, i = 2, . . . , d, and obtain θ(xi) ≡ 0 (mod mn+1), i = 2, . . . , d. The lemma then follows by induction and a limit argument. To prove the claim we consider the k-linear map ¯θn : mn+1/mn+2 → mn/mn+1 induced by θ. We have to show that for i > 1 the class of θ(xi) in mn/mn+1 lies in the image of ¯θn. A short calculation, using (1) and (2), shows that the image of ¯θn is spanned by the images of the monomials of degree n xl1 1 xl2 2 ··· xld d , l1 + . . . + ld = n, such that l1 6≡ −1 (mod p). In particular, for n < p − 1 the map ¯θn is surjective, and the claim is true. Suppose now that n ≥ p − 1 and fix an index i > 1. By (2) we can write θ(xi) ≡ n Xl=0 al xl 1yl (mod mn+1), 8 where al ∈ k and yl is a homogenous polynomial of degree n− l in the variables x2, . . . , xd. A straightforward calculation, using (1) and (2), shows that θp(xi) ≡ n Xl=p−1 l(l − 1)··· (l − p + 2) al xl−p+1 1 yl (mod mn−p+2). (4) On the other hand, it follows from (ii) and (2) that θp(xi) = aθ(xi) ≡ 0 (mod mn). (5) (Here we have used that the element a ∈ Frac(B) occuring in (ii) is of the form a = θ(x1)−1θp(x1) and therefore lies in B.) Combining (4) and (5) we find that al = 0 if l ≡ −1 (mod p). This shows that the class of θ(xi) in mn/mn+1 lies in the image of ¯θn and proves the claim. The proof of Lemma 2.1 is now complete. ✷ Corollary 2.2 Let θ ∈ Derk(B) be as in Lemma 2.1. Then: (i) The subring A := Ker(θ) ⊂ B is regular, and B/A is finite and flat of rank p. (ii) Suppose that θ(f ) ∈ B, for some element f ∈ L. Then there exists f0 ∈ B such that θ(f0) = θ(f ). (iii) Suppose that for some element f ∈ L×. Then there exists u ∈ B× such that θ(f )/f ∈ B, θ(u)/u = θ(f )/f. Proof: Let (x1, . . . , xd) be a regular system of parameters as in Lemma 2.1, and set h := θ(x1) ∈ B×. Then θ = hθ0, where θ0(x1) = 1, θ0(x2) = . . . = θ0(xd) = 0. Clearly, we have A = Ker(θ) = Ker(θ0). It is now easy to see that A = k[[xp 1, x2, . . . , xd]] is regular and that B = A ⊕ A · x1 ⊕ . . . ⊕ A · xp−1 1 (6) is finite and flat over A of rank p. So (i) is proved. 9 For the proof of (ii) we write f = a0 + a1x1 + . . . + ap−1xp−1 1 , with ai ∈ K := Frac(A) = L ∩ A. Then the hypothesis θ(f ) ∈ B implies θ0(f ) = a1 + 2a2x1 + . . . + (p − 1)ap−1xp−2 1 ∈ B. Now (6) shows that a1, . . . , ap−1 ∈ A. Therefore, θ(f0) = θ(f ), with f0 := a1x1 + . . . + ap−1xp−1 1 ∈ B. This proves (ii). Assertion (iii) follows from [11], Theorem 2. For the convenience of the reader, we sketch the argument. Let f ∈ L× be given such that θ(f )/f ∈ B. We claim that there exists a Weil divisor d on Spec (A) such that B · d = (f ). (This is obviously a 'descent argument' and explains the title of this section.) Assuming the claim we can prove (ii), as follows. By (i), A is regular and in particular factorial. Therefore, d = (f0), for some f0 ∈ K = Frac(A). So f = uf0 for a unit u ∈ B×, and we get θ(f )/f = θ(u)/u, as desired. To prove the claim, it suffices to show the following: for every prime ideal p ✁ B of height one we have ep vp(f ). Here vp is the normalized valuation on L associated to p and ep denotes the ramification index of p in the field extension L/K. Since ep ∈ {1, p}, we may assume that n := vp(f ) is prime to p, and we have to show that ep = 1. Let t ∈ Bp be a uniformizer for vp. Then we can write f = utn, with u ∈ B× p . From θ(f )/f = θ(u)/u + n · θ(t)/t ∈ B we conclude θ(t)/t ∈ Bp. This means that θ induces a derivation ¯θ on the residue field kp. From assumption (i) of Lemma 2.1 it follows that ¯θ 6= 0. From we conclude kp∩A ⊂ k ¯θ=0 p ( kp fp := [kp : kp∩A] = p. Now the inequality p = [L : K] ≥ ep · fp implies ep = 1, and the claim is proved. ✷ 10 3 Ramification of p-cyclic Galois extensions In the situation of Theorem 1.3, we obtain a cyclic Galois extension L/K of de- gree p by taking fraction fields: L := Frac(B), K := Frac(A) = LG. The present section contains some preliminary investigation of the ramification of this ex- tension with respect to the discrete valuation corresponding to the parameter π. The main point here is that for p-cyclic Galois extensions it is possible to distinguish two types of wild ramification: total ramification and fierce ramifi- cation. Accordingly, the proof of Theorem 1.3 will be divided into these two cases. 3.1 a prime number). Let v be a discrete valuation on L which is fixed by G. Let L/K be a cyclic Galois extension of degree p (where p is, as always, We choose a generator σ of G. The valuation rings of K and L with respect to v are denoted by OK and OL, the residue fields by ¯K and ¯L. The letter π will always denote a uniformizer for v on L (i.e. πOL = pL is the maximal ideal of OL). We normalize the valuation v such that v(π) = 1. We set eL/K := [v(L×) : v(K ×)], fL/K := [ ¯L : ¯K]. These invariants are related by the fundamental equality See e.g. [7], Corollary XII.6.3. eL/K · fL/K = [L : K] = p. Definition 3.1 We classify the ramification behavior of L/K at v as follows: I: If eL/K = p and fL/K = 1, we call L/K totally ramified. There are two subcases: (a) If char( ¯L) 6= p, L/K is called tamely ramified. (b) If char( ¯L) = p, L/K is called totally wildly ramified. II: Suppose that eL/K = 1 and fL/K = p. There are again two subcases: (a) If ¯L/ ¯K is Galois, then L/K is called unramified. (b) If ¯L/ ¯K is inseparable, then L/K is called fiercely ramified. The following invariant is useful to distinguish these cases: Proposition 3.2 δ := min{ v(σ(x) − x) x ∈ OL } ≥ 0. (i) L/K is unramified if and only if δ = 0. (7) (ii) If L/K is tamely ramified, then δ = 1. (Note: the converse of this impli- cation is actually false: L/K may be fiercely ramified.) 11 (iii) The extension L/K is totally ramified if and only if for some uniformizer π. In this case (8) holds for every uniformizer π. δ = v(σ(π) − π), (8) Proof: Assertions (i) and (ii) are classical. See e.g. [13], IV, §1. Assertion (iii) is certainly well known as well. We will nevertheless give a proof because it yields some useful insight. We may assume that δ > 0. Let π be an arbitrary uniformizer. By the definition of δ, the map θπ : OL → ¯L, x 7→ (cid:16) σ(x) − x πδ mod π(cid:17), is well defined and does not vanish. A short calculation shows that θπ is a derivation (i.e. θπ(x + y) = θπ(x) + θπ(y) and θπ(xy) = ¯x · θπ(y) + ¯y · θπ(x)), and hence we have (9) θπ(aπ) = ¯a · θπ(π), for all a ∈ OL. Suppose that θπ(π) = 0. This is equivalent to the condition Then (9) shows that both the equation θπ(π) = 0 and (10) hold in fact for all uniformizers π. Moreover, θπ induces a derivation δ < v(σ(π) − π). (10) θπ : ¯L → ¯L, ¯x 7→ θπ(x) mod π. We have θπ 6= 0 by definition of δ. In particular, there exists a unit x ∈ O× with δ = v(σ(x) − x). It is also clear that L ¯K ⊂ Ker(θπ) ( ¯L. Since [ ¯L : ¯K] ≤ p, it follows that ¯K = Ker(θπ) and [ ¯L : ¯K] = p. In particular, L/K is fiercely ramified (Case II (b)). On the other hand, if L/K is fiercely ramified, then eL/K = 1, which means L . Then we obviously have θπ(π) = that there exists a uniformizer π in OK = OG 0. All claims made in Proposition 3.2 have now been proved. ✷ 3.2 proof of Proposition 3.2 yields the following. Suppose that L/K is fiercely ramified (Case II (b)). In this case the Corollary 3.3 Suppose that L/K is fiercely ramified. Then for every uni- formizer π the map θπ : ¯L → ¯L, ¯x 7→ (cid:16) σ(x) − x πδ mod π(cid:17), is a derivation with the following properties. 12 (i) θπ 6= 0, (ii) Ker(θπ) = ¯K, (iii) θp π = a θπ, for some a ∈ ¯K. Proof: (i) and (ii) have been proved above. Property (iii) is true for all derivations of a field ¯L of characteristic p for which [ ¯L : Ker(θπ)] = p. See e.g. [3], Chapter IV.8, Exercise 3. ✷ The following lemma will be used in Section 4.5. Lemma 3.4 Let x ∈ OL be such that v(σ(x) − x) ≥ δ + n, for some n ≥ 1. Then there exists y ∈ OK with y ≡ x (mod pn L). Proof: We proceed by induction on n, starting with n = 1. Let π ∈ OK be an invariant uniformizer. The assumption v(σ(x)− x) ≥ δ + 1 means θπ(¯x) = 0. It follows from Corollary 3.3 (ii) that ¯x ∈ ¯K. So we can take for y ∈ OK any representative of ¯x. We may hence suppose n ≥ 2. By induction, there exists an element z ∈ OK such that z ≡ x (mod πn−1). Write x = z + aπn−1, with a ∈ OL. Then θπ(¯a) ≡ σ(x) − x πδ+n−1 ≡ 0 (mod π). As before, it follows that ¯a ∈ ¯K. Let b ∈ OK be any lift of ¯a and set y := z + bπn−1 ∈ OK. ✷ 3.3 It is well known that the trace map TrL/K : L → K contains interesting information about the ramification of L/K. We will use this only in the case where L/K is totally wildly ramified. As before, we fix a uniformizer π ∈ OL, v(π) = 1. For any element a ∈ L× we set [a]π := a · π−v(a) mod π ∈ ¯L×. Remark 3.5 The map L× → ¯L×, a 7→ [a]π, obeys the following rules: (i) [π]π = 1, (ii) [a · b]π = [a]π · [b]π, (iii) [a + b]π = [a]π + [b]π, if v(a) = v(b) and [a]π + [b]π 6= 0. 13 Lemma 3.6 Suppose that L/K is totally wildly ramified (Case I (b)). Let π ∈ OL be a uniformizer (i.e. v(π) = 1). (i) The trace induces an ¯L-linear isomorphism Tr : p −(p−1)(δ−1) L /p −(p−1)(δ−1)+1 L ∼ → OK /pK ∼= ¯L. (ii) There exists an element s(π) L/K ∈ ¯L×, uniquely determined by the condition TrL/K(xπ−(p−1)(δ−1)) ≡ s(π) L/K · x (mod π), for all x ∈ OL. (iii) If π′ = uπ is another uniformizer, with u ∈ O× s(π′) L/K = ¯u−(p−1)(δ−1) · s(π) L/K. L , then (iv) We have s(π) L/K = −[σ(π) − π]p−1 π . Proof: (cf. [4], Prop. 2.4) Set λ := σ(π) − π. By Proposition 3.2 (ii), (iii) and the assumption on L/K we have v(λ) = δ ≥ 2 and σ(λ) ≡ λ (mod πδ+1). Also, for i = 1, . . . , p − 1 we have σi(π) − π = λ + σ(λ) + . . . + σi−1(λ). Using Remark 3.5 (iii) and the above, one shows easily that and v(σi(π) − π) = v(λ) = δ (11) (12) Let P ∈ OL[X] denote the minimal polynomial of π over K. By [13], Lemma [σi(π) − π]π = i · [σ(π) − π]. III.6.2, we have TrL/K(cid:0)πiP ′(π)−1(cid:1) =( 0, 1, i = 0, . . . , p − 2, i = p − 1. (13) Since P ′(π) = p−1 Yi=1 (π − σi(π)), 14 we get v(P ′(π)) = (p − 1)δ from (11). To prove (i), note that an element a ∈ p can be uniquely written in the form −(δ−1)(p−1) L a = p−1 Xi=0 ai πi P ′(π) , with a0, . . . , ap−2 ∈ pK and ap−1 ∈ OK. By (13) we have TrL/K(a) = ap−1; . This proves (i). The furthermore, ap−1 ∈ pK if and only if a ∈ p Assertions (ii) and (iii) follow easily. Using (13), (12) and the definition of s(π) L/K, we get −(δ−1)(p−1)+1 L p−1 s(π) L/K = [P ′(π)]π = [ π − σi(π)]π Yi=1 = (p − 1)! · [σ(π) − π]p−1 π = −[σ(π) − π]p−1 π , ✷ proving (iv). 4 The proof of Theorem 1.3 In this section we prove our main result, Theorem 1.3. The assumptions made in Section 1.1 and 1.2 and Conditions (i) and (ii) of Theorem 1.3 are in force throughout. The ring of invariants A := BG is a local, integrally closed domain. By 4.1 Assumption (c) of Section 1.1, A is noetherian and B is a finite A-algebra. Our first step is to show that in the proof of Theorem 1.3 we may assume that B is complete. This assumption will be used later in the proof of Lemma 4.4 and Lemma 4.5. Let A and B denote the completions of A and B, with respect to their maximal ideals. These are complete local rings with residue field k. The action of G on B extends uniquely to B, and we have a canonical map A → BG. Lemma 4.1 The above map is an isomorphism, A ∼= BG. Proof: By definition of A and Assumption (c) we have an exact sequence of finite A-modules (14) Since B is noetherian we have rad(mAB) = mB, so B is the completion of the A-module B. Since A is noetherian, the exact sequence (14) induces an exact sequence 0 → A → B σ−1−→ B. by [9], Thm. 54. This proves the lemma. ✷ 0 → A → B σ−1−→ B, 15 By [1], Prop. 11.24, A is regular if and only if A is regular. So all we have to show is that the action of G on B satisfies Hypothesis (i) and (ii) of Theorem 1.3 (if it does for B). Let x1, . . . , xd be a regular system of parameters for B; it is also a regular system of parameters for B. Moreover, it is easy to check that Therefore, Iσ = (σ(x1) − x1, . . . , σ(xd) − xd) ✁ B. Iσ := (σ(x) − x x ∈ B) = B · Iσ. Our claim follows immediately. For the rest of this section we may therefore assume that B is a complete local ring. We set ¯B := B/Bπ; this is a complete local ring with residue field k. 4.2 Since π is a regular parameter of B, ¯B is regular. It follows from Condition (i) of Theorem 1.3 that the map x 7→ (cid:16) σ(x) − x πδ (mod π)(cid:17) θπ : B → ¯B, is a well defined derivation. Moreover, the ideal generated by the image of θπ is equal to ¯B. Let L := Frac(B) denote the fraction field of B. Let v : L× → Z denote the discrete valuation corresponding to the prime ideal Bπ ✁ B, normalized such that v(π) = 1. So the residue field of v is the fraction field of ¯B, ¯L = Frac( ¯B). Let K := LG be the fixed field. Then L/K is a Galois extension with Galois group G = hσi ∼= Z/pZ, and v is fixed by G. We are therefore in the situation of Section 3. Clearly, the number δ from Theorem 1.3 is the same as the invariant defined by (7): δ = min{ v(σ(x) − x) x ∈ OL }. Since δ > 0, the extension L/K is ramified, i.e. Case II (a) in Definition 3.1 is excluded. There are three cases left. If θπ(π) 6= 0, then L/K is totally ramified by Proposition 3.2 (iii). Moreover, Condition (ii) of Theorem 1.3 implies that θπ(π) ∈ ¯B× is a unit. If δ = 1 then L/K is tamely ramified (Case I (a)); otherwise, L/K is totally wildly ramified (Case I (b)). On the other hand, if θπ(π) = 0, then L/K is fiercely ramified (Case II (b)). In this case the derivation θπ induces a nontrivial derivation θπ : ¯B → ¯B, and the unique extension of θπ to ¯L is the derivation from Corollary 3.3. After these preliminary remarks, the proof of Theorem 1.3 is done by a case-by-case analysis of the three different types of ramification of v in L/K. 16 4.3 Case I (a): L/K is tamely ramified Suppose that char( ¯L) 6= p. By Proposition 3.2 (ii) we have δ = 1. Therefore, the morphism is injective. It follows that G → Aut(Bπ/Bπ2) σ(π) ≡ ζp · π (mod π2), for some primitive pth root of unity ζp ∈ ¯B. But the assumption Iσ = (π) implies that for a unit u ∈ B×. From the above we get σ(π) − π = uπ u ≡ ζp − 1 (mod m). In particular, ζp − 1 6∈ m, which shows that char(k) 6= p. Now the hypothesis of Theorem 1.3 implies that A is regular, see [12] and Remark 1.4. 4.4 Case I (b): L/K is totally wildly ramified In this case Condition (ii) of Theorem 1.3 implies (15) where u ∈ B× is a unit and δ ≥ 2. Let p := A ∩ Bπ ✁ A. This is a prime ideal of height one. σ(π) − π = uπδ, Lemma 4.2 The element λ := NL/K(π) = p−1 Yi=0 σi(π) ∈ A generates p. In particular, p is a principal ideal. Proof: It is obvious that λ ∈ p. Let α ∈ p be an arbitrary element. To prove that p = (λ), it suffices to show that λ α in B. We have π α by definition of p. This means that v(α) > 0. Since α ∈ A ⊂ K and eL/K = p by assumption, we actually have v(α) ≥ p. This implies πp α. By (15) we have σ(π) = π (1 + uπδ−1) ∼ π. Here a ∼ b means: b = va for a unit v ∈ B×. It follows that σi(π) ∼ π for all i and hence λ ∼ πp. Therefore, λα, as desired, and the lemma is proved. ✷ Let ¯A := A/p. We consider ¯A as a subring of ¯B via the natural embedding ¯A ֒→ ¯B = B/Bπ. Lemma 4.3 We actually have ¯A = ¯B. In particular, ¯A is regular. 17 Proof: Clearly, the rings ¯A and ¯B have the same fraction field ¯L. Consider the ¯L-linear map Tr : ¯L → ¯L, ¯x 7→ (cid:0) TrL/K(xπ−(p−1)(δ−1)) mod π(cid:1), studied in Section 3.3. We claim that Tr( ¯B) ⊂ ¯A. Indeed, if x is an element of B, then the proof of Lemma 3.6 shows that v(TrL/K(xπ−(p−1)(δ−1))) ≥ 0. Therefore, TrL/K(xπ−(p−1)(δ−1)) ∈ (B[π−1] ∩ OL)G = BG = A by factoriality of B, proving the claim. It follows from (15) and Lemma 3.6 (ii), (iv) that Tr(¯x) = −¯up−1 ¯x, where ¯u ∈ ¯B× is the image of u in ¯B. We conclude −¯up−1 ¯B = ¯B ⊂ ¯A. So ¯A = ¯B, and the lemma is proved. ✷ Using the two lemmas it is now easy to see that A is regular: if ¯x1, . . . , ¯xd−1 is a regular sequence of parameters for ¯A and xi ∈ A is a lift of ¯xi, then (λ, x1, . . . , xd−1) is a regular system of parameters for A. 4.5 Case II (b): L/K is fiercely ramified Let θ = θπ ∈ Der( ¯L) be the derivation from Corollary 3.3. It follows from the definition of θ that θ( ¯B) ⊂ ¯B. Also, Condition (i) of Theorem 1.3 implies that there exists an element x ∈ B such that σ(x) − x πδ ∈ B×. This shows that the k-linear map ¯θ : m ¯B/m2 ¯B → k. induced by θ is not zero, ¯θ 6= 0. Since θp = a θ for some element a ∈ ¯K (Corollary 3.3 (iii)), the derivation θ satisfies the hypotheses of Lemma 2.1 (with respect to the ring ¯B). Therefore, we may (and will) use Corollary 2.2 in the proof of the following two lemmas. Lemma 4.4 There exists an element λ ∈ A of the form λ = uπ, with u ∈ B×. Proof: By induction, we will construct a sequence of elements λ0, λ1, . . . ∈ B of the form λn = unπ, un ∈ B×, such that: (a) λn ≡ λn−1 (mod πn), (b) v(σ(λn) − λn) ≥ δ + n + 1, 18 for all n ≥ 0. The limit λ := limn λn is then an element λ ∈ A = BG of the form λ = uπ, with u ∈ B×. We start the induction at n = 0 by setting λ0 := π. Then Condition (a) is empty and (b) is true by Proposition 3.2 (iii). The case n = 1 is special. The uniformizer λ1 we are looking for is of the form λ1 = u1π, with u1 ∈ B×. Condition (a) is still empty, and Condition (b) is equivalent to σ(λ1) − λ1 πδ+1 0 ≡ ≡ u1w + θπ(¯u1) (mod π), where w := (σ(π) − π)π−δ−1 ∈ B. So we have to find an element ¯u1 ∈ ¯B× with θπ(¯u1) ¯u1 = − ¯w. (16) By Lemma 3.4, there exists an element µ ∈ OK with µ ≡ π (mod p2 write µ = vπ, with v ∈ O× shows that L). If we L , then the same calculation that resulted in (16) θπ(¯v) ¯v = − ¯w ∈ ¯B. Now it follows from Corollary 2.2 (iii) that there exists ¯u1 ∈ ¯B× such that (16) holds. Now the case n = 1 of the lemma is proved. We may hence assume that n ≥ 2 and that λ0, . . . , λn−1 have already been constructed. Since λn−1 ∼ π, we may assume that λn−1 = π. In particular, v(σ(π) − π) ≥ δ + n. The element λn we are looking for is of the form λn = π(1 + aπn), (17) for some a ∈ B. Using (17) we get σ(λn) − λn πδ+n ≡ c + θπ(¯a) (mod π), where c := (σ(π) − π)π−δ−n ∈ B. This means that we have to find an element ¯a ∈ ¯B such that (18) By (17) and Lemma 3.4, there exists an element µ ∈ OK with µ ≡ π (mod pn L). Write µ = π(1 + bπn), with b ∈ OL. Then the same calculation from which we deduced (18) yields θπ(¯a) = −¯c. Using Corollary 2.2 (ii) we find an element ¯a ∈ ¯B satisfying (18). Now the proof of the lemma is complete. ✷ θπ(¯b) = −¯c. 19 Let λ ∈ A be as in the lemma. We get a natural embedding ¯A := A/Aλ ֒→ ¯B = B/Bπ. It is clear that ¯A is a complete local ring with residue field k and quotient field ¯K. Lemma 4.5 We have ¯A = ¯B ∩ ¯K. Proof: The proof is very similar to the proof of the previous lemma. The inclusion ¯A ⊂ ¯B∩ ¯K is clear. To prove the converse, suppose we have an element x ∈ B such that ¯x ∈ ¯K. We will inductively construct a sequence y0, y1, . . . ∈ B such that (a) x ≡ yn (mod π), (b) yn ≡ yn−1 (mod πn), (c) v(σ(yn) − yn) ≥ δ + n + 1, for all n ≥ 0. The limit y := limn yn is then an element y ∈ A = BG with x ≡ y (mod π). For n = 0 we set y0 := x. Then (a) is clear, (b) is empty and (c) follows from ¯x ∈ ¯K. So we may assume n ≥ 1. The element yn we are looking for is of the form yn = yn−1 + aπn, with a ∈ B. The crucial Condition (c) is equivalent to σ(yn) − yn πδ+n ≡ c + θπ(¯a) ≡ 0 (mod π), where c := (σ(yn−1) − yn−1)π−δ−n ∈ B. So we have to find ¯a ∈ ¯B such that θπ(¯a) = −¯c. (19) By Lemma 3.4, there does exist an element z ∈ OL satisfying the properties required for yn. If we write z = yn−1 + bπn, with b ∈ OL, then we get θπ(¯b) = −¯c ∈ ¯B. Now Corollary 2.2 (ii) shows that ¯a ∈ ¯B satisfying (19) exists, and we are done. The lemma shows that ¯A = Ker(θπ ¯B). By Corollary 2.2 (i) it follows that ¯A is regular. With the same argument used at the end of the previous subsection we conclude that A is regular. Now Theorem 1.3 is proved. ✷ ✷ References [1] M.F. Atiyah and I.G. Mcdonald. Introduction to Commutative Algebra. Addison-Wesley, 1969. [2] J. Fogarty. Kahler differentials and Hilbert's fourteenth problem for finite groups. Amer. J. Math., 102(6):1159 -- 1175, 1980. 20 [3] N. Jacobson. Lectures in Abstract Algebra III: Theory of Fields and Galois Theory. Springer, 1964. [4] K. Kato. Swan conductors with differential values. In Galois Representa- tions and Arithmetic Algebraic Geometry, number 12 in Advanced Studies in Pure Math., pages 315 -- 342, 1987. [5] N.M. Katz and B. Mazur. Arithmetic Moduli of Elliptic Curves. Princeton Univ. Press, 1985. [6] F. Kir´aly. Quotient singularities and models of curves. PhD thesis, Uni- versitat Ulm, expected 2010. [7] S. Lang. Algebra. Addison-Wesley, 1993. [8] D. Lorenzini. Wild quotient singularities of surfaces. Preprint, 2009. [9] H. Matsumura. Commutative Algebra. Benjamin, 2nd edition, 1980. [10] B.R. Peskin. Quotient-singularities and wild p-cyclic actions. J. Algebra, 81:72 -- 99, 1983. [11] P. Samuel. Classes de diviseurs et d´eriv´ees logarithmiques. Topology, 3(1):81 -- 96, 1968. [12] J.-P. Serre. Groupes finis d'automorphismes d'anneaux locaux r´eguliers. In Colloque d'Alg`ebre (Paris, 1967), Exp. 8. Secr´etariat math´ematique, 1967. [13] J.-P. Serre. Corps locaux. Hermann, 3. edition, 1968. [14] J. Watanabe. Some remarks on Cohen-Macaulay rings with many zero divisors and an application. J. Algebra, 39:1 -- 14, 1976. 21
1612.01583	1	1612	2016-12-05T22:47:03	Monodromy of the $SL(n)$ and $GL(n)$ Hitchin fibrations	[ "math.AG", "math.DG" ]	We compute the monodromy of the Hitchin fibration for the moduli space of $L$-twisted $SL(n,\mathbb{C})$ and $GL(n,\mathbb{C})$-Higgs bundles for any $n$, on a compact Riemann surface of genus $g>1$. We require the line bundle $L$ to either be the canonical bundle or satisfy $deg(L) > 2g-2$. The monodromy group is generated by Picard-Lefschetz transformations associated to vanishing cycles of singular spectral curves. We construct such vanishing cycles explicitly and use this to show that the $SL(n,\mathbb{C})$ monodromy group is a {\em skew-symmetric vanishing lattice} in the sense of Janssen. Using the classification of vanishing lattices over $\mathbb{Z}$, we completely determine the structure of the monodromy groups of the $SL(n,\mathbb{C})$ and $GL(n,\mathbb{C})$ Hitchin fibrations. As an application we determine the image of the restriction map from the cohomology of the moduli space of Higgs bundles to the cohomology of a non-singular fibre of the Hitchin fibration.	math.AG	math	MONODROMY OF THE SL(n) AND GL(n) HITCHIN FIBRATIONS DAVID BARAGLIA Abstract. We compute the monodromy of the Hitchin fibration for the moduli space of L- twisted SL(n, C) and GL(n, C)-Higgs bundles for any n, on a compact Riemann surface of genus g > 1. We require the line bundle L to either be the canonical bundle or satisfy deg(L) > 2g − 2. The monodromy group is generated by Picard-Lefschetz transformations associated to vanishing cycles of singular spectral curves. We construct such vanishing cycles explicitly and use this to show that the SL(n, C) monodromy group is a skew-symmetric vanishing lattice in the sense of Janssen. Using the classification of vanishing lattices over Z, we completely determine the structure of the monodromy groups of the SL(n, C) and GL(n, C) Hitchin fibrations. As an application we determine the image of the restriction map from the cohomology of the moduli space of Higgs bundles to the cohomology of a non-singular fibre of the Hitchin fibration. 6 1 0 2 c e D 5 ] . G A h t a m [ 1 v 3 8 5 1 0 . 2 1 6 1 : v i X r a 1. Introduction 1.1. Monodromy of the Hitchin fibration. In this paper we determine the monodromy of the SL(n, C)-Hitchin fibration h : M(n, L) → A for all n. Here M(n, L) is the moduli space of L- twisted SL(n, C)-Higgs bundles on a compact Riemann surface Σ of genus g > 1 and L is a line bundle which is either the canonical bundle K, or satisfies l = deg(L) > 2g− 2. Our proof also gives the monodromy of the corresponding GL(n, C)-Hitchin fibration. For SL(2, C), the monodromy was first determined in [10] using a combinatorial approach and later revisited in [2] from a more geometric point of view. It does not seem possible to extend the arguments used in [10], [2] to rank n > 2. In this paper we introduce new techniques that apply for all n. When n > 2, our results are completely new. M(n, L) to the affine space A = Ln Recall that the Hitchin fibration is a proper, surjective holomorphic map h : M(n, L) → A from j=2 H 0(Σ, Lj) [14, 15, 22]. As shown by Hitchin [15], when L = K the Hitchin map gives M(n, L) the structure of an algebraically completely integrable sys- tem with respect to a natural holomorphic symplectic structure on M(n, L). For L 6= K, we do not have a holomorphic symplectic structure, but it remains the case that the non-singular fibres of the Hitchin map are abelian varieties. Finding the monodromy of this fibration is a natural problem which has a number of important applications, some of which are described in §1.5. To explain our results we will mainly focus on the SL(n, C) case. The GL(n, C) case is covered by Theorem 1.5. Let D ⊂ A denote the locus of singular fibres of the SL(n, C)-Hitchin fibration and let Areg = A \ D be the regular locus. Let Mreg be the points of M lying over Areg, so that h : Mreg → Areg is a non-singular bundle of abelian varieties. The monodromy of the SL(n, C)-Hitchin system is the local system Λ on Areg whose fibre over a point a ∈ Areg is the underlying lattice Λa = H1(h−1(a), Z) of the abelian variety h−1(a). Equivalently Λ is the dual 2010 Mathematics Subject Classification. Primary 14H60 53C07; Secondary 14H70, 14D05. 1 2 DAVID BARAGLIA of the Gauss-Manin local system R1h∗Z. Choose a basepoint a0 ∈ Areg and let ΛP be the fibre of Λ over a0 (we use subscript P because ΛP is the lattice of a Prym variety, see §2). The local system Λ is then equivalent to a representation ρSL : π1(Areg, a0) → Aut(ΛP ). We call ρSL the monodromy representation of the Hitchin fibration. The image ΓSL ⊆ Aut(ΛP ) of ρSL will be called the monodromy group of the SL(n, C)-Hitchin fibration. The smooth fibres of the Hitchin fibration are Prym varieties associated to certain branched covers of Σ called spectral curves, recalled in §2. This construction shows that the smooth fibres are equipped with a natural polarization, which defines a non-degenerate skew-symmetric bilinear pairing h , i : ΛP × ΛP → Z invariant under the monodromy representation. Thus the monodromy of the SL(n, C)-Hitchin fibration is given by a triple (ΛP ,h , i, ΓSL), consisting of: (i) a lattice ΛP , (ii) a skew-symmetric Z-valued bilinear form h , i on ΛP , (iii) a subgroup ΓSL ⊂ Aut(ΛP ) preserving h , i. 1.2. Vanishing cycles. The monodromy group is generated by a certain collection of vanishing cycles. To describe these vanishing cycles, we choose a basepoint a0 ∈ Areg of the form a0 = (0, 0, . . . , 0, an), where an ∈ H 0(Σ, Ln) has only simple zeros. Let tot(L) denote the total space of L and π : tot(L) → Σ the projection. The spectral curve S associated to a0 is given by: S = {λ ∈ tot(L) λn + an(π(λ)) = 0} ⊂ tot(L). Spectral curves of this form carry a Galois action of the cyclic group Zn by deck transformations, and will thus be referred to as cyclic spectral curves. One of the key insights of this paper is that the monodromy of the Hitchin fibration can be computed by studying cyclic spectral curves and small perturbations of them. Given a cyclic spectral curve S, we construct vanishing cycles as follows. Let π : S → Σ denote the restriction of π to S. Then π : S → Σ is a degree n branched cover of Σ, branched over the zeros b1, b2, . . . , bk of an, where k = nl. Let u1, . . . , uk ∈ S be the corresponding ramification points in S. The spectral curve construction identifies the lattice ΛP with the kernel of the Gysin homomorphism π∗ : H 1(S, Z) → H 1(Σ, Z). Suppose that γ : [0, 1] → Σ is an embedded path in Σ joining bi to bj, where i 6= j and assume that γ does not meet any other branch point. Then π−1(γ([0, 1])) ⊂ S consists of n paths γ1, γ2, . . . , γn : [0, 1] → S joining ui to uj. Let t : S → S be the generator of the cyclic Galois action given by λ 7→ e2πi/nλ and order the paths so that tγi = γi+1, i = 1, . . . , n − 1. Then (γ1 − γ2) is a 1-cycle in S. Let lγ ∈ H1(S, Z) be its homology class and cγ ∈ H 1(S, Z) the Poincar´e dual class. Similarly the cycles (γ2 − γ3), . . . , (γn−1 − γn), (γn − γ1) correspond to the cohomology classes tcγ, . . . , tn−2cγ, tn−1cγ ∈ H 1(S, Z). Clearly these cycles are in the kernel of π∗ : H 1(S, Z) → H 1(Σ, Z), so they are elements of ΛP . As the paths γ1, . . . , γn are only determined by γ up to cyclic permutation, the cycles cγ, tcγ, . . . , tn−1cγ are likewise only determined by γ up to cyclic permutation. We call cγ, tcγ, . . . , tn−1cγ the vanishing cycles associated to γ. We then have: Theorem 1.1. Let a = ticγ ∈ ΛP be a vanishing cycle associated to γ. Then the SL(n, C) monodromy group ΓSL contains the Picard-Lefschetz transformation Ta : ΛP → ΛP associated to a, given by: (1.1) Ta(x) = x + ha, xia. MONODROMY OF THE SL(n) AND GL(n) HITCHIN FIBRATIONS 3 1.3. Vanishing lattices. To state our main results, we need to recall the notion of a skew- symmetric vanishing lattice [16]. Let R denote the ring Z or Z2 and let V be a free R-module of rank µ equipped with a bilinear form h , i : V × V → R which is alternating, i.e. hx, xi = 0 for all x ∈ V . For any a ∈ V we have an endomorphism Ta : V → V , called a symplectic transvection, given by Equation (1.1). Clearly Ta acts as an automorphism of V which preserves h , i. Let j : V → V ∗ be the map j(x) = hx, i, let V0 be the kernel of j and V ′ the image of j. Let Sp#V denote the automorphisms of (V,h , i) acting trivially on V ∗/V ′ and note that Ta belongs to Sp#V for any a ∈ V . Definition 1.2 ([16]). Let ∆ ⊆ V and let Γ∆ be the subgroup of Sp#V generated by {Tα}α∈∆. We say that (V,h , i, ∆) is a (skew-symmetric)-vanishing lattice over R if: (i) ∆ is a Γ∆-orbit. (ii) ∆ spans V . (iii) If µ > 1, then there exists δ1, δ2 ∈ ∆ for which hδ1, δ2i = 1. We also call Γ∆ the monodromy group of the vanishing lattice. 1.4. Main Results. Let VC ⊂ ΛP be the collection of all vanishing cycles cγ, . . . , tn−1cγ associated to paths γ joining pairs of branch points. Let ΓVC be the subgroup of Sp#(ΛP ) generated by transvections Ta, where a ∈ VC and set ∆P = ΓVC · VC. Theorem 1.3. We have that (ΛP ,h , i, ∆P ) is a vanishing lattice. Let Γ∆P be the subgroup of Sp#(ΛP ) generated by transvections in ∆P . Then ΓVC = Γ∆P = ΓSL. That is, the mon- odromy group of the SL(n, C)-Hitchin fibration is the monodromy group Γ∆P of the vanishing lattice (ΛP ,h , i, ∆P ). Thus to describe the monodromy group ΓSL and its action on ΛP , it suffices to classify the vanishing lattice (ΛP ,h , i, ∆P ). The classification is as follows (here we use the notation for vanishing lattices introduced in [16, 17], which is recalled in §5): Theorem 1.4. The vanishing lattice (ΛP ,h , i, ∆P ) is isomorphic to: (1) A′(1, 1, . . . , 1, 2, 2, . . . , 2; 0), if n = 2, (2) O# (3) O# (4) Sp#(1, 1, . . . , 1, n, n, . . . , n; 0), if n is even, l is odd and n > 2. a (1, 1, . . . , 1, n, n, . . . , n; 0), where a = (m(m − 1)/2)l, if n = 2m + 1 is odd, a (1, 1, . . . , 1, n, n, . . . , n; 0), where a = m(l/2), if n = 2m and l are even and n > 2, In this classification, the number of 1's is (n − 2)(g − 1) + n(n − 1)l/2 − 1 and the number of n's is g. As mentioned previously, our results also yield the monodromy of the GL(n, C)-Hitchin fibration. Fix a basepoint a0 ∈ Areg with spectral curve S and let ΛS = H 1(S, Z). Notice that ΛP is a sublattice of ΛS. The monodromy of the GL(n, C)-Hitchin fibration is given by a representation ρGL : π1(Areg, a0) → Aut(ΛS). Let ΓGL ⊆ Aut(ΛS) be the image of ρGL. We have: Theorem 1.5. The monodromy group ΓGL is the subgroup of Aut(ΛS) generated by transvections Tα : ΛS → ΛS, where α ∈ ∆P . In §7 we give an application of our monodromy computations. Let M(n, d, L) denote the moduli space of rank n L-twisted Higgs bundles with trace-free Higgs field and determinant equal to a fixed line bundle of degree d. The Hitchin fibration h : M(n, d, L) → A may be defined for any value of d. Let Fa = h−1(a) be a non-singular fibre of the Hitchin fibration lying over a ∈ Areg. Then: 4 DAVID BARAGLIA Theorem 1.6. Let ω ∈ H 2(Fa, Q) be the cohomology class of the polarization on Fa. The image Im(H ∗(M(n, d, L), Q)) → H ∗(Fa, Q)) of the restriction map in cohomology is the subspace spanned by 1, ω, ω2, . . . , ωu, where u = dimC(Fa) is the dimension of the fibre. This result generalises to SL(n, C) a result proved by Thaddeus for SL(2, C) in [26] (see also the appendix of [7]). When n and d are coprime, the moduli space M(n, d, L) is smooth and Theorem 1.6 can be proven by a straightforward generalisation of the argument given in [7]. The argument breaks down when n and d are not coprime, in which case Theorem 1.6 is a new result. 1.5. Applications. Knowledge of the monodromy is important for applications of the moduli space of Higgs bundles in which the Hitchin fibration plays a prominent role. We describe some of these applications here. Cohomology of the moduli space and Ngo's support theorem. Let MGL(n, d, L) denote the moduli space of L-twisted Higgs bundles of rank n, degree d. Assume that n and d are coprime, in which case MGL(n, d, L) is non-singular. We have the Hitchin fibration h : MGL(n, d, L) → AGL whose monodromy is isomorphic to Λ for any value of d. Let Aell GL ⊂ AGL be the elliptic locus of AGL, the locus of points for which the corresponding spectral curve is reduced and irreducible. Ngo's support theorem [21] (see also [6]) applied to the restriction hell : Mell GL of h over Aell GL occuring in the decomposition theorem for hell are supported on the whole of Aell GL. The decomposition theorem then reduces to: GL(n, d, L) → Aell GL implies that the perverse sheaves on Aell ∗ Q[dim(MGL(n, d, L))] =Mi Rhell ICAell GL (∧iΛ∗ ⊗Z Q)[f − i], where f is the dimension of the fibres of the Hitchin fibration. This raises the possibility of com- puting the cohomology of Mell GL(n, d, L) through a knowledge of the local system Λ. In turn, this gives us partial information about the cohomology of the full moduli space MGL(n, d, L). In fact for d > 2g − 2, it was shown by Chaudouard and Laumon that there are no new supports when Aell GL is replaced by AGL [9]. A similar result holds in the SL(n, C) case, except there are additional supports related to the endoscopy theory of SL(n, C) [8]. In a related direction, we note that local monodromy calculations were used extensively in [7] in the proof of the P = W conjecture for GL(2, C), SL(2, C), P SL(2, C). We expect our monodromy calculations to be similarly useful in tackling this conjecture for higher rank groups. Wall crossing and the hyperkahler metric. In the work of Gaiotto, Moore and Neitzke on the Kontsevich-Soibelman wall-crossing formula for N = 2 supersymmetric quantum field theories, a relation is established between hyperkahler metrics and counts of BPS states [13]. In particular, this can be applied to the hyperkahler metric on the moduli space of Higgs bundles (in the untwisted case L = K). Through a twistorial construction, the hyperkahler metric on the moduli space of Higgs bundles is encoded in a family of complex symplectic forms ω(ζ) parametrised by ζ ∈ CP1. In [13], the authors consider local Darboux coordinates for the symplectic forms ω(ζ). The Darboux coordinates χγ, are locally defined C∗-valued functions on Mreg depending on a choice of point ζ ∈ C∗ ⊂ CP1 and local section γ of the monodromy local system Λ. The functions χγ depend multiplicatively on γ in the sense that χγ+γ ′ = χγχγ ′ and are Darboux coordinates in the sense MONODROMY OF THE SL(n) AND GL(n) HITCHIN FIBRATIONS 5 that their Poisson brackets are given by the formula: {χγ, χγ ′} = hγ, γ′iχγχγ ′. The locally defined coordinates χγ do not patch together globally, instead they satisfy a wall-crossing formula as one crosses a real codimension 1 subspace of Mreg×C∗. The wall-crossing formula is given by a holomorphic symplectomorphism, which may be expressed in terms of symplectomorphisms Kγ of the form Kγ : χγ ′ 7→ χγ ′(1 − (−1)q(γ)χγ)hγ ′,γi, where q : Λ → Z2 = {0, 1} is a quadratic function on Λ (see §3.5). Interestingly, the existence of a monodromy invariant quadratic function on Λ is a key ingredient in the classification of Λ as a vanishing lattice. The proposal of [13] is that the existence of coordinates χγ satisfying the wall-crossing formula and some further technical conditions completely determines the hyperkahler metric on the moduli space, through twistor theory. Clearly it is important for this proposal to be able to describe the local system Λ and its quadratic function q. Connected components of real character varieties. Let G be a real or complex reductive Lie group. A representation θ : π1(Σ) → G is called reductive if the action of π1(Σ) on the Lie algebra of G obtained by composing θ with the adjoint representation is a direct sum of irreducible representations. Let Homred(π1(Σ), G) be the space of reductive representations given the compact- open topology. The group G acts on Homred(π1(Σ), G) by conjugation and the quotient Rep(G) = Homred(π1(Σ), G)/G is a Hausdorff space [23], called the character variety of representations of π1(Σ) in G. The non- abelian Hodge correspondence gives a homeomorphism between Rep(G) and the moduli space MG of polystable G-Higgs bundles, with L = K. If GC is a complex reductive group and GR the split real form, then as shown by Schaposnik [24, Theorem 4.12], the image of the natural map MGR → MGC meets the smooth fibres of the Hitchin fibration in points of order 2. The points of order 2 in smooth fibres are described by the monodromy with Z2-coefficients Λ[2] = Λ⊗Z Z2. Thus one may hope to determine the number of connected components of MGR and hence of Rep(GR) by counting the number of orbits of the local system Λ[2]. This strategy has been successfully carried out for SL(2, R) in [25] and extended to GL(2, R), P GL(2, R) and Sp(4, R) with maximal Toledo invariant in [2]. With the calculation of the monodromy Λ in this paper one can similarly obtain the number of components of the character varieties for SL(n, R), GL(n, R), P SL(n, R) and Sp(2n, R) with maximal Toledo invariant for any value of n. There is however a caveat that so far for n > 2, we are only able to count the number of connected components which intersect the regular locus. To turn these counts into a complete proof it will be necessary to show that there are no components of the corresponding character varieties lying entirely within the singular locus. We hope to address this problem in future work. The strategy of counting components of Rep(GR) through monodromy may also be applied to non-split real forms. One may define spectral data or cameral data for arbitrary real forms and at least in some cases, this gives rise to families of smooth spectral curves or cameral curves. The generic fibre F of the Hitchin fibration restricted to MGR is generally not a discrete space anymore, but one can consider the monodromy action on π0(F ) and by counting the number of orbits obtain a count of the components of Rep(GR). 6 DAVID BARAGLIA 1.6. Structure of the paper. This paper is organised as follows. In §2 we recall the spectral curve construction of the regular locus of the Hitchin fibration. In §3, we study in detail the structure of the lattices ΛS, ΛP , where ΛS = H 1(S, Z) is the integral cohomology of the spectral curve over the basepoint a0. In particular, we work out the intersection form h , i in Section 3.3. In Section 3.5 we find that under conditions on n and l, there is a natural quadratic function on ΛP and we compute its Arf invariant. This is an important input for the classification of the vanishing lattice (ΛP ,h , i, ∆P ). In §4 we construct vanishing cycles associated to paths γ in Σ joining pairs of branch points. In §5, we use the vanishing cycles of §4, to construct the vanishing lattice (ΛP ,h , i, ∆P ). We then recall the classification of vanishing lattices over Z2 in Section 5.2, over Z in Section 5.3 and use this to completely determine the structure of (ΛP ,h , i, ∆P ). In §6, we prove that the monodromy group ΓSL of the SL(n, C)-Hitchin fibration coincides with the monodromy group Γ∆P of the vanishing lattice. In §7, we apply our monodromy computations to give a proof of Theorem 7.1, describing the image of the restriction map from the cohomology of the moduli space of Higgs bundles to the cohomology of a non-singular fibre of the Hitchin fibration. 1.7. Acknowledgements. We would like to thank Laura Schaposnik, Tam´as Hausel and Nigel Hitchin helpful discussions. The author is supported by the Australian Research Council grant DE160100024. 2. Spectral curves and the regular locus We recall some basic facts about the moduli space of SL(n, C)-Higgs bundles, in particular the spectral curve construction of the regular locus. An SL(n, C) L-twisted Higgs bundle is a pair (E, Φ) where E is a rank n holomorphic vector bundle on Σ with trivial determinant and Φ is a trace-free holomorphic endomorphism Φ : E → E ⊗ L. As shown by Nitsure [22], one may define notions of semistability and S-equivalence for twisted Higgs bundles and construct a moduli space M(n, L) of S-equivalence classes of semistable SL(n, C) L-twisted Higgs bundles. The moduli space M(n, L) is a quasi-projective complex algebraic variety [22]. As the rank n and line bundle L will be fixed throughout, we will omit them from the notation and simply write M for the moduli space. Recall the Hitchin fibration, also known as the Hitchin map or Hitchin system is a surjective holomorphic map h : M → A from M to the affine space A =Ln j=2 H 0(Σ, Lj) [14, 15, 22]. Using the notion of spectral curves recalled below, it can be shown that the non-singular fibres of the Hitchin map are abelian varieties [15, 3]. Let D ⊂ A denote the locus of singular fibres of the Hitchin map and let Areg = A \ D be the regular locus. We let Mreg denote the points of M lying over the regular locus, so that h : Mreg → Areg is a locally trivial torus bundle. Our goal in this paper is to determine the monodromy of this torus bundle. We recall the spectral curve construction of Mreg from Areg. For this, let tot(L) denote the total space of L and π : tot(L) → Σ the projection. We let λ ∈ H 0(tot(L), π∗L) denote the tautological section of π∗(L) on tot(L). Given a = (a2, a3, . . . , an) ∈ A, let sa be the section of π∗Ln on tot(L) given by sa = λn + a2λn−2 + ··· + an. The vanishing locus of sa defines a curve Sa ⊂ tot(L) called the spectral curve associated to a. We use π : Sa → Σ to denote the restriction of π : tot(L) → Σ. As in [15], Bertini's theorem implies that Sa is smooth for generic a ∈ A. Moreover it can be shown that Sa is smooth if and only if the corresponding fibre of the Hitchin system h : M → A is non-singular [18]. When this is the case the corresponding fibre of the Hitchin system is: h−1(a) = {M ∈ Jacln(n−1)/2(Sa) N m(M ) = Ln(n−1)/2}, MONODROMY OF THE SL(n) AND GL(n) HITCHIN FIBRATIONS 7 where Jacd(Sa) denotes the degree d component of P ic(Sa) and N m : P ic(S) → P ic(Σ) is the norm map associated to π : Sa → Σ [3]. Note that h−1(a) is naturally a torsor over the abelian variety P rym(Sa, Σ), which is defined as: The abelian variety P rym(Sa, Σ) is called the Prym variety associated to π : Sa → Σ. P rym(Sa, Σ) = {M ∈ Jac(Sa) N m(M ) = O}. Let q : S → Areg denote the family of smooth spectral curves parametrised by Areg. This may be defined as the set of pairs (a, x) ∈ Areg×tot(L) such that sa(x) = 0. This is a family of non-singular curves parametrised by Areg and so we may construct the relative Jacobian j : Jac(S/Areg) → Areg. Let N m : Jac(S/Areg) → Areg × Jac(Σ) be the fibrewise norm map and let p : P rym(S/Areg) → Areg be the fibrewise kernel of N m. Then P rym(S/Areg) is a family of Prym varieties parametrised by Areg. Remark 2.1. If n is odd or l is even, then there exists on Σ a square root L(n−1)/2 of Ln−1. The map sending a ∈ Areg to π∗(L(n−1)/2) ∈ Jacln(n−1)/2(Sa) is a section of Mreg → Areg and this gives an isomorphism Mreg ≃ P rym(S/Areg). Definition 2.2. We define the following local systems on Areg: (1) Λ = Hom(R1h∗Z, Z), the monodromy of the Hitchin fibration. (2) ΛJ = Hom(R1j∗Z, Z), the monodromy of the family of Jacobians. (3) ΛP = Hom(R1p∗Z, Z), the monodromy of the family of Prym varieties. (4) ΛS = Hom(R1q∗Z, Z), the monodromy of the family of spectral curves. (5) ΛΣ will denote the trivial local system with coefficient group H 1(Σ, Z). Proposition 2.3. We have the following relations between local systems: (1) ΛS ≃ ΛJ (2) Λ ≃ ΛP (3) The fibrewise norm map N m : Jac(S/Areg) → Areg×Jac(Σ) induces a short exact sequence: (2.1) Proof. (1). For any smooth spectral curve S there is a canonical isomorphism H 1(S, Z) ≃ H 1(Jac(S), Z). Clearly this isomorphism extends to a canonical isomorphism of local systems R1q∗Z ≃ R1j∗Z on Areg associated to the families q : S → Areg and j : Jac(S/Areg) → Areg. Thus ΛS ≃ ΛJ . 0 −→ ΛP −→ ΛJ N m−→ ΛΣ −→ 0. (2). Recall that P rym(S/Areg) is the pre-image under the fibrewise norm map N m : Jac(S/Areg) → Areg × Jac(Σ) of Areg × {O}. Similarly Mreg is the pre-image under N m of Areg × {Ln(n−1)/2}. Thus p : P rym(S/Areg) → Areg is a bundle of abelian varieties and h : Mreg → Areg is a bundle of torsors for P rym(S/Areg). In particular this gives a canonical isomorphism R1h∗Z ≃ R1p∗Z of local systems. Hence Λ ≃ ΛP . (3). We have a short exact sequence of bundles of abelian varieties over Areg: P rym(S/Areg) → Jac(S/Areg) → Areg × Jac(Σ). Taking homology of the fibres gives the short exact sequence (2.1). (cid:3) Instead of working with local systems directly we will fix a basepoint a0 ∈ Areg and work with representations of π1(Areg, a0). Let π : S → Σ denote the spectral curve corresponding 8 DAVID BARAGLIA to the basepoint a0. We let ΛS denote H 1(S, Z), let ΛΣ denote H 1(Σ, Z) and let ΛP denote the kernel of the Gysin map π∗ : ΛS → ΛΣ. Then Λ ≃ ΛP corresponds to a representation ρSL : π1(Areg, a0) → Aut(ΛP ) which we call the monodromy representation of the SL(n, C) Hitchin fibration. Similarly ΛS ≃ ΛJ corresponds to a representation ρGL : π1(Areg, a0) → Aut(ΛS), which we refer to as the monodromy representation of the GL(n, C) Hitchin fibration. We note that (2.1) corresponds to a short exact sequence of representations: 0 −→ ΛP −→ ΛS π∗−→ ΛΣ −→ 0. Remark 2.4. Although it will play no part in subsequent calculations, we note that the dual local system Λ∗ and corresponding representation ρ∗ P ) give the monodromy of the Hitchin fibration for the group P GL(n, C). This is an instance of Langlands duality for Hitchin systems, which in general implies that Langlands dual groups give rise to dual monodromy representations. SL : π1(Areg, a0) → Aut(Λ∗ 3. The lattices ΛS, ΛP 3.1. Decomposition of ΛS. In this section we fix a basepoint a0 ∈ Areg and examine the structure of the lattices ΛS, ΛP in detail. For this we choose a0 to be of the form a0 = (0, 0, . . . , 0, an), for some an ∈ H 0(Σ, Ln). The corresponding spectral curve S is given by the equation λn + an = 0 and it is clear that S is smooth if and only if an has only first order zeros. Let b1, b2, . . . , bk ∈ Σ be the zeros of an, where k = nl and let u1, . . . , uk ∈ S be the corresponding ramification points. Definition 3.1. Let D be the open unit disc in C of radius 1 centred at 0 and D the corresponding closed unit disc. We say that an embedding i : D → Σ is a trivialising disc if i(D) contains the branch points of π : S → Σ and the restriction of S to Σ\ D is the trivial covering space (n disjoint copies of Σ \ D). Proposition 3.2. For any an ∈ H 0(Σ, Ln) with simple zeros, the corresponding spectral curve π : S → Σ admits a trivialising disc. Proof. Let Σ′ = Σ\{b1, . . . , bk}. Let C ⊂ H1(Σ′, Z) be the subgroup generated by cycles l1, l2, . . . lk around the points b1, . . . , bk, so that we have an exact sequence: 0 / C / H1(Σ′, Z) / H1(Σ, Z) / 0. The monodromy of the branched cover S → Σ defines a homomorphism φS : H1(Σ′, Z) → Zn such that φS(li) = 1 for each i. Let i : D → Σ be an embedding of D in Σ for which i(D) contains the branch points. Then i determines a splitting si : H1(Σ, Z) → H1(Σ′, Z) given by the composition of the isomorphism H1(Σ, Z) ≃ H1(Σ \ i(D), Z) with the inclusion induced map H1(Σ \ i(D), Z) → H1(Σ′, Z). We have that i gives a trivialising disc if and only if the composition φS ◦ si : H1(Σ, Z) → Zn is the trivial homomorphism. Choose a branch point bi and let l be an embedded loop in Σ′ with underlying homology class of the form [l] = li + si(m), where m ∈ H1(Σ, Z) is the homology class of l as a loop in Σ. Let τl : Σ → Σ be a Dehn twist around l, supported in a neighbourhood of l containing no branch points. So τl preserves the branch points and acts a homeomorphism τl : Σ′ → Σ′. The composition i′ = τl ◦ i : D → Σ gives a new embedding and corresponding splitting si′ : H1(Σ, Z) → H1(Σ′, Z). / / / / MONODROMY OF THE SL(n) AND GL(n) HITCHIN FIBRATIONS 9 From the commutative diagram: si H1(Σ, Z) ∼=o H1(Σ \ i(D), Z) H1(Σ′, Z) τl ∗ τl ∗ τl ∗ H1(Σ, Z) ∼=o H1(Σ \ i′(D), Z) / H1(Σ′, Z) si′ we see that si′ = τl∗◦ si◦ τl ∗ )(a) = si(a) +ha, mili, hence (φS ◦ si′ )(a) = (φS ◦ si)(a) +ha, mi. From this we see that by applying to i a series of Dehn twists around suitably chosen loops, we can obtain an embedding i of D with splitting si such that φS ◦ si is trivial, as required. −1 ∗ . One finds that (τl∗siτl −1 (cid:3) By Proposition 3.2, there exists a trivialising disc i : D → Σ. Let D′ be an open disc obtained from D by shrinking the radius slightly, but for which i(D′) still contains all branch points. Let U0 = i(D) and U1 = Σ\i(D′), where i(D′) is the closure of i(D′) in Σ. The Mayer-Vietoris sequence applied to the cover V0 = π−1(U0), V1 = π−1(U1) of S gives: 0 / H2(S, Z) ∂ / H1(V0 ∩ V1, Z) / H1(V0, Z) ⊕ H1(V1, Z) (i0 ∗,i1 ∗) j0 ∗−j1 ∗/ / H1(S, Z) / 0, where ia is the inclusion V0∩V1 → Va and ja is the inclusion Va → S. Since i : D → Σ is a trivialising i=1 U1 →`n disc we have that V1 consists of n disjoint copies of U1 and that the inclusion`n i=1 Σ i=1H 1(Σ, Z). Similarly, V0 ∩ V1 is homotopy equivalent to induces an isomorphism H1(V1, Z) → ⊕n n copies of the circle. These circles correspond to the n lifts to S of the boundary of the disc i(D), hence the map i1∗ is trivial. Let ΛS,0 be the cokernel of i0∗ : H1(V0 ∩ V1, Z) → H1(V0, Z) and i=1H 1(Σ, Z). Then the above sequence gives an isomorphism H1(S, Z) ≃ ΛS,1 = H1(V1, Z) = ⊕n ΛS,0 ⊕ ΛS,1. By Poincar´e duality we have ΛS = H 1(S, Z) ≃ H1(S, Z), so (3.1) ΛS ≃ ΛS,0 ⊕ ΛS,1. We will make extensive use of this decomposition of ΛS in the following sections. 3.2. Z[t]-module structure. Let t : S → S be the map sending λ to ξλ, where ξ = e2πi/n. Then t generates a Zn-action on S. We can thus view ΛS as a Z[t]-module, where the action of t satisfies tn = 1. Clearly (3.1) is a direct sum of Z[t]-modules. The lattice ΛS,1 was seen to consist of a direct sum of n copies of ΛΣ = H 1(Σ, Z) ≃ H1(Σ, Z). Further, t acts on ΛS,1 by cyclic permutation of these n copies so that as Z[t]-modules we have: Next consider ΛS,0. From its definition ΛS,0 fits into an exact sequence: ΛS,1 = Z[t] htn − 1i ⊗Z ΛΣ. 0 / H2(S, Z) ∂ / H1(V0 ∩ V1, Z) i0 ∗ / / H1(V0, Z) / ΛS,0 / 0. Up to homotopy V0 ∩ V1 can be identified with the boundary of V0 and i0 with the inclusion map. We have that V0 is a degree n branched cover of the unit disc D ⊂ C. Applying a suitable ( ( o / / 6 6 o / / / / / / / / / 10 DAVID BARAGLIA γ1 1 γ2 1 γ1 2 γ2 2 u1 u2 u3 γn−1 1 γn 1 γ1 b2 γn−1 2 γn 2 π γ2 b3 b1 Figure 1. The branched covering π−1([−1/2, 1/2]) → [−1/2, 1/2] homeomorphism to D, we can assume that the branch points b1, b2, . . . , bk lie on the real axis with −1/2 = b1 < b2 < ··· < bk = 1/2. A deformation retraction of D onto the interval [−1/2, 1/2] lifts to a deformation retraction of V0 onto π−1([−1/2, 1/2]). The branched cover π−1([−1/2, 1/2]) → [−1/2, 1/2] is depicted in Figure 1. For 1 ≤ i ≤ k − 1, let γi be the path joining bi to bi+1 along the interval [bi, bi+1]. Let γ1 i be the lifts of γi to paths in π−1([−1/2, 1/2]) joining ui to ui+1, as shown in Figure 1. We may order the lifts in such a way that γj i . Then tj−1ci is the homology class of γj i = −(ci + tci + ··· + tn−2ci), but that i − γ1 the cycles ci, tci, . . . , tn−2ci are independent. In fact, we have: i . Let ci be the homology class of the 1-cycle γ1 i = tj−1γ1 i . Note that tn−1ci = γn i − γj+1 i − γ2 i , γ2 i , . . . , γn Proposition 3.3. The homology group H1(V0, Z) is free as a Z-module, with basis given by the cycles tjci for 1 ≤ i ≤ k − 1, 0 ≤ j ≤ n − 2. Proof. We have shown that V0 admits a deformation retraction to π−1([−1/2, 1/2]). By considering Figure 1, the proposition is easily seen to follow. (cid:3) Next, we will determine the image of i0∗ : H1(V0 ∩ V1, Z) → H1(V0, Z). For this we introduce a convention on the ordering of the paths γj i as follows. For each i, let Di be a small disc in D centered around the point bi. We choose these discs small enough so that they are mutually disjoint. i divide Di into n segments. We will assume that the lifts γj Let Di = π−1(Di). Then γ1 have been ordered in such a way that for i < i < k and 1 ≤ j ≤ n− 1, we have that γj i−1 ∩ π−1(Di) lies in the segment between γj . By induction on i, such an ordering exists. Henceforth we will assume that such an ordering has been chosen. Figure 2 shows the order of paths entering and leaving the point ui under our convention. i and γj+1 i , . . . , γn i , γ2 i i Proposition 3.4. The image of i0∗ : H1(V0 ∩ V1, Z) → H1(V0, Z) is the subgroup spanned by ∂, t∂, . . . , tn−1∂, where MONODROMY OF THE SL(n) AND GL(n) HITCHIN FIBRATIONS 11 γ2 i−1 γ3 i γ2 i γ3 i−1 γ1 i−1 γ1 i ui Figure 2. Ordering convention for the paths γj i ∂ = c1 + (1 + t)c2 + (1 + t + t2)c3 + ··· + (1 + t + ··· + tn−2)ck−1. Proof. Recall that V0 ∩ V1 may be identified with the boundary of V0, which a disjoint union of n circles. Moreover t : S → S cyclically permutes the circles. Let ∂D be a clockwise loop in D enclosing the branch points b1, . . . , bk. Let ∂ be any lift of ∂D to a loop in S. It follows that the image of i0∗ is spanned by the cycles ∂, t∂, . . . , tn−1∂. Thus it remains to determine the cycle ∂. We think of ∂ as consisting of an upper segment ∂+ and a lower segment ∂−. b1 b2 ∂+ ∂− bk Due to our ordering conventions, it is easy to see that ∂+ can be lifted to γ1 1 + γ1 2 + ··· + γ1 k−1 and ∂− to −γ2 1 − γ3 3 − ··· − γn 1) + (γ1 1 − γ2 2 − γ4 ∂ = (γ1 = c1 + (1 + t)c2 + (1 + t + t2)c3 + ··· + (1 + t + ··· + tn−2)ck−1 k−1. We then have: 2 ) + ··· + (γ1 2 − γ3 k−1 − γn k−1) as required. (cid:3) Corollary 3.5. As a Z[t]-module, ΛS,0 is isomorphic to k− 2 copies of Z[t]/h1 + t + t2 +··· + tn−1i. We shall use the same notation tjci for the cycle in H1(V0, Z) and for its image in ΛS,0. This should not cause confusion, since from this point onward we will only be concerned with ΛS,0 as opposed to H1(V0, Z). We have that the tjci span ΛS,0, but are not linearly independent, since by 12 DAVID BARAGLIA γ3 1 tc1 γ2 1 γ3 1 γ2 1 tc1 c1 γ1 1 u1 u2 γ1 1 c1 Figure 3. hc1, tc1i = 1 Proposition 3.4 we have the relation c1 + (1 + t)c2 + (1 + t + t2)c3 + ··· + (1 + t + ··· + tn−2)ck−1 = 0. 3.3. Intersection form. Consider the intersection pairing h , i : ΛS × ΛS → Z defined by the cup product. Clearly h , i is t-invariant and the decomposition ΛS = ΛS,0 ⊕ ΛS,1 is orthogonal. Under the identification ΛS,1 = ⊕n i=1H 1(Σ, Z), we have that the restriction h , iΛS,1 is given by n copies of the usual intersection form on H 1(Σ, Z). The intersection form on ΛS,0 is given by the following: Proposition 3.6. We have the following intersection pairings: hci, tcii = 1, hci, ci+1i = 1, hci, tci+1i = −1, hci, tjcii = 0, hci, tjci+1i = 0, hci, tjci′i = 0, whenever i − i′ > 1. for 2 ≤ j ≤ n − 2, for 2 ≤ j ≤ n − 2, Proof. We choose representatives of the cycles tjci meeting transversally and directly compute their intersection. Figure 3 shows the computation of hc1, tc1i = 1 and Figure 4 shows the computations of hc1, c2i = 1 and hc1, tc2i = −1. The remaining intersections are computed similarly. (cid:3) htn−1i ⊗Z ΛΣ. 1, where π∗ The intersection relations described in Proposition 3.6 may be visualised as a graph, shown in Figure 5. Here the vertices correspond to the elements {tjci}, for 1 ≤ i ≤ k − 1 and 0 ≤ j ≤ n − 2. We draw an oriented edge from from u to v whenever hu, vi = 1. 3.4. Decomposition of ΛP . Recall the decomposition ΛS = ΛS,0 ⊕ ΛS,1. Let iu : ΛS,u → ΛS for u = 0, 1 be the inclusions and ju : ΛS → ΛS,u the projections. Recall also the identification ΛS,1 = Z[t] Proposition 3.7. The map π∗ : ΛΣ → ΛS factors as π∗ = i1 ◦ π∗ given by π∗ (π1)∗ : ΛS,1 → ΛΣ is given by (π1)∗(tja) = a, for a ∈ ΛΣ. Proof. Let i : D → Σ be a trivialising disc and ΛS = ΛS,0 ⊕ ΛS,1 the corresponding decomposition. Any class in ΛΣ can be represented by a cycle lying outside of D and the factorsation π∗ = i1 ◦ π∗ 1 follows. Similarly, any class in ΛS,0 is represented by a cycle whose image under π lies in D. So π∗ is trivial on ΛS,0 and the factorisation π∗ = (π1)∗ ◦ j1 follows. (cid:3) 1 : ΛΣ → ΛS,1 is 1(a) = (1 + t + ··· + tn−1)a. The map π∗ : ΛS → ΛΣ factors as π∗ = (π1)∗ ◦ j1, where MONODROMY OF THE SL(n) AND GL(n) HITCHIN FIBRATIONS 13 γ2 1 c1 γ3 2 tc2 γ2 2 γ1 1 c2 γ1 2 u2 Figure 4. hc1, c2i = 1, hc1, tc2i = −1 ck−2 c3 ck−1 c1 c2 tc1 t2c1 tn−3c1 tn−2c1 tck−1 tn−3ck−1 tn−2ck−1 Figure 5. Intersection graph for ΛS,0 Corollary 3.8. The decomposition ΛS = ΛS,0⊕ ΛS,1 induces a similar decomposition ΛP = ΛP,0 ⊕ ΛP,1, where ΛP,0 = ΛS,0 and ΛP,1 is the kernel of the map (π1)∗ : ΛS,1 → ΛΣ defined in Proposition 3.7. Moreover, we have an isomorphism of Z[t]-modules ΛP,1 ≃ Z[t]/h1 + t + ··· + tn−1i ⊗Z ΛΣ. Corollary 3.9. The polarization type of ΛP is (1, 1, . . . , 1, n, n, . . . , n), where 1 occurs (n − 2)(g − 1) + n(n − 1)l/2 − 1 times and n occurs g times. Proof. Since the decomposition ΛS = ΛS,0 ⊕ ΛS,1 is orthogonal and h , i is unimodular on ΛS, it follows that the restriction of h , i to ΛS,0 or ΛS,1 is unimodular. So the type of ΛP,0 = ΛS,0 is (1, 1, . . . , 1). Choose a symplectic basis a1, b1, . . . , ag, bg of ΛΣ. There is an induced decomposition of ΛP,1 into g orthogonal Z[t]-submodules M1, M2, . . . , Mg, where Mu is spanned by {(1−t)au, t(1− t)au, . . . , tn−2(1−t)au, (1−t)bu, t(1−t)bu, . . . , tn−2(1−t)bu}. The intersection matrix on Mu can be obtained as the tensor product of the Cartan matrix for An−1 with the standard 2 × 2 symplectic matrix (cid:20) 0 It follows that the type of Mu is given by the invariants of the An−1 Cartan matrix, which are (1, 1, . . . , 1, n), where there are n − 1 entries equal to 1. Thus the type of ΛP,1 is of the form (1, 1, . . . , n, . . . , n), where there are g copies of n. −1 0(cid:21). 1 (cid:3) 14 DAVID BARAGLIA In order to apply the classification of vanishing lattices in Section 5, it will be necessary to consider cohomology with Z2-coefficients. Let ΛS[2] = ΛS ⊗Z Z2 be the mod 2 reduction of ΛS and similarly define ΛP [2], ΛΣ[2]. Proposition 3.10. Suppose n is odd. Then π∗ : ΛΣ[2] → ΛS[2] gives a splitting of the sequence 0 / ΛP [2] / ΛS[2] / ΛΣ[2] / 0. π∗ The decomposition induced by this splitting is orthogonal. ΛS[2] = ΛP [2] ⊕ ΛΣ[2] Suppose n is even. Then the image of π∗ : ΛΣ[2] → ΛS[2] is contained in ΛP [2]. We have that π∗ΛΣ[2] is the null space of h , iΛP [2] and hence the quotient H = ΛP [2]/π∗ΛΣ[2] has an induced non-degenerate alternating bilinear form h , iH . The filtration π∗ΛΣ[2] ⊂ ΛP [2] ⊂ ΛS[2] can be split in such a way that under the induced isomorphism ΛS[2] ≃ ΛΣ[2] ⊕ (ΛP [2]/ΛΣ[2]) ⊕ (ΛS[2]/ΛP [2]) ≃ ΛΣ[2] ⊕ H ⊕ ΛΣ[2], we have that π∗(a) = (a, 0, 0), π∗(a, b, c) = c and h(a, b, c), (a′, b′, c′)i = ha, c′i + hb, b′iH + ha′, ci, for all a, a′, c, c′ ∈ ΛΣ[2] and b, b′ ∈ H. Proof. For a ∈ ΛΣ we have π∗π∗(a) = na. Thus when n is odd we have that π∗ : ΛΣ[2] → ΛS[2] is a splitting of π∗ : ΛS[2] → ΛΣ[2]. For a ∈ ΛΣ[2], b ∈ ΛP [2], we have hπ∗(a), bi = ha, π∗(b)i = 0. So for n odd, the splitting provided by π∗ is orthogonal. Now assume that n is even. We have shown that π∗ΛΣ[2] ⊂ ΛP [2]. By non-degeneracy of h , i on ΛS[2] we have that Λ⊥ P [2] = ker(π∗)⊥ = im(π∗), so that π∗ΛΣ[2] is the nullspace of h , iΛP [2] as claimed. The quotient H = ΛP [2]/π∗ΛΣ[2] then has an induced non-degenerate alternating bilinear form h , iH . The last claim concerning the splitting of the filtration is straightforward. (cid:3) 3.5. Quadratic functions on ΛS[2], ΛP [2]. Under conditions on n and l we will construct mon- odromy invariant quadratic functions on ΛS[2], ΛP [2] and compute their Arf invariants. Definition 3.11. Let V be a finite dimensional Z2-vector space and let h , i be a bilinear form on V which is alternating, i.e. hx, xi = 0 for all x ∈ V . A quadratic function associated to (V ,h , i) is a Z2-valued function q on V satisfying: for all x, y ∈ V . q(x + y) = q(x) + q(y) + hx, yi Let V 0 = {x ∈ V hx, i = 0} and let p be the dimension of V 0. Then V /V 0 is even-dimensional as it carries a non-degenerate alternating bilinear form. Let 2r be the dimension of V /V 0, so that V has dimension µ = 2r + p. If q is a quadratic function, observe that qV 0 is linear. If q vanishes on V 0, then q descends to a quadratic function on V /V 0 and we define the Arf invariant Arf(q) ∈ Z2 i=1 q(ei)q(fi), where e1, . . . , er, f1, . . . , fr ∈ V project to a symplectic basis of V /V 0. This can be shown to be independent of the choice of e1, . . . , fr. If qV 0 is not identically zero, of q to bePr / / / / MONODROMY OF THE SL(n) AND GL(n) HITCHIN FIBRATIONS 15 then Arf(q) is left undefined. For fixed r, p there are at most 3 isomorphism classes of quadratic functions, distinguished by whether the Arf invariant is 0, 1 or undefined. Let X be a compact Riemann surface with canonical bundle KX. Spin structures on X may be identified with square roots of KX . A spin structure gives a KO-orientation and hence a Gysin homomorphism ϕX : KO(X) → KO−2(pt) = Z2. If E is a holomorphic vector bundle on X with orthogonal structure, then ϕX ([E]) is the mod 2 index [1]: ϕX ([E]) = dim(cid:16)H 0(X, E ⊗ K 1/2 X )(cid:17) (mod 2). Note that if A is any line bundle on X, then A ⊕ A∗ has an orthogonal structure given by pairing A and A∗. By Riemann-Roch, we have ϕX (A ⊕ A∗) = deg(A) (mod 2). The restriction of ϕX to H 1(X, Z2), the space of Z2-line bundles, is a quadratic function with constant term, in the sense that: ϕX (a + b) = ϕX (a) + ϕX (b) + ha, bi + ϕX (0). In particular, if we let ϕX = ϕX + ϕX (0), then ϕXH 1(X,Z2) is a quadratic function. The spin structure K 1/2 X is called even or odd according to whether ϕX (0) is 0 or 1. If X has genus gX then ϕX has 2gX −1(2gX + 1) zeros [1]. It follow that the Arf invariant of ϕH 1(X,Z2) equals ϕX (0). Now consider the case of a spectral curve π : S → Σ and let KS denote the canonical bundle of S. We will use the mod 2 index to construct a monodromy invariant quadratic function on ΛS[2] provided that either n is odd or n and l are both even. We will later see that when n is even and l is odd, there are no monodromy invariant quadratic functions on ΛS[2]. By the adjunction formula we have KSπ∗(K −1) = π∗(Ln−1). Under the assumption that either n is odd or n and l are both even, there exists a square root L(n−1)/2 of Ln−1 on Σ. This defines a relative KO-orientation for π : S → Σ and hence a Gysin homomorphism π! : KO(S) → KO(Σ). If E is a holomorphic vector bundle on S with orthogonal structure, then π! is related to taking the direct image by: π![E] = [π∗(E ⊗ π∗(L(n−1)/2))]. Note that by relative duality, π∗(E ⊗ π∗(L(n−1)/2)) inherits an orthogonal structure from the or- thogonal structure on E. Now choose a square root K 1/2 of K and set K 1/2 S = π∗(L(n−1)/2K 1/2). As we have chosen our spin structures to be compatible with the relative spin structure, we have that ϕS = ϕΣ ◦ π!. Moreover, since K 1/2 and L(n−1)/2 are defined on Σ, it is clear that ϕSΛS [2] is a monodromy invariant quadratic function on ΛS[2]. Proposition 3.12. Suppose that either n is odd or n and l are even. Choose square roots of K and L(n−1) on Σ so that the function ϕS : KO(S) → Z2 is defined. Then: (1) The restriction of ϕS to ΛP [2] is independent of the choice of square roots and defines a monodromy invariant quadratic function q : ΛP [2] → Z2. (2) If n = 2m + 1 is odd, then Arf (q) = (m(m − 1)/2)l (mod 2). (3) If n = 2m is even, then Arf (q) = m(l/2) (mod 2). Proof. To show independence of choices, suppose we replace K 1/2 and L(n−1)/2 by K 1/2 ⊗ A1 and L(n−1)/2⊗A2, where A1, A2 ∈ ΛΣ[2]. Then K 1/2 S ⊗π∗(A), where A = A1⊗A2. This has the effect of replacing the function x 7→ ϕS(x) with x 7→ ϕS(x + π∗[A]). But if x ∈ ΛP [2] is replaced with K 1/2 S 16 DAVID BARAGLIA then hx, π∗[A]i = 0, so ϕS(x + π∗[A]) + ϕS(π∗[A]) = ϕS(x) + ϕS(π∗[A]) + hx, π∗[A]i + ϕS(0) + ϕS(π∗[A]) = ϕS(x) + ϕS(0) = ϕS(x), which shows independence of ϕS on the choice of square roots. Now suppose that n = 2m + 1 is odd. If A ∈ ΛΣ[2], we find that ϕS(π∗[A]) = ϕΣ(π!π∗[A]) = ϕΣ([A ⊗ (Lm ⊕ Lm−1 ⊕ ··· ⊕ L−m)]) = ml + (m − 1)l + ··· + l + ϕΣ([A]) = (m(m − 1)/2)l + ϕΣ([A]). In this calculation we have used that π∗OS = OΣ ⊕ L−1 ⊕ L−2 ⊕ ··· ⊕ L−(n−1) [3]. It follows that ϕS(0) = (m(m − 1)/2)l + ϕΣ(0) and that ϕS(π∗[A]) = ϕΣ([A]). Thus the restriction of ϕS to ΛΣ[2] has Arf invariant equal to ϕΣ(0). Now since n is odd, we have an orthogonal decomposition ΛS[2] = ΛP [2]⊕ ΛΣ[2] and from the additivity of the Arf invariant we have ϕS(0) = Arf (q)+ ϕΣ(0). Hence Arf (q) = ϕS(0) + ϕΣ(0) = (m(m − 1)/2)l. Lastly, suppose that n and l are even and set n = 2m. If A ∈ ΛΣ[2], we find that ϕS(π∗[A]) = ϕΣ(π!π∗[A]) = ϕΣ([A ⊗ (L(2m−1)/2 ⊕ L(2m−3)/2 ⊕ ··· ⊕ L−(2m−1)/2)]) = (2m − 1)l/2 + (2m − 3)l/2 + ··· + l/2. = m(l/2) (mod 2). In particular, ϕS(0) = m(l/2) and ϕS(π∗[A]) = 0 for all A ∈ ΛΣ[2]. By Proposition 3.10, we have ΛS[2] = ΛΣ[2]⊕H⊕ΛΣ[2], where H is orthogonal to the two factors of ΛΣ[2]. Let J be the subspace of ΛS[2] given by J = ΛΣ[2]⊕ 0⊕ ΛΣ[2]. Then we have an orthogonal decomposition ΛS[2] = H ⊕ J. By the above computation, ϕS has at least 22g zeros on the subspace J, hence the Arf invariant of ϕSJ is 0. This implies that Arf ( ϕSH ) = Arf ( ϕSΛS [2]) = ϕS(0) = m(l/2). Now recall that ΛP [2] = π∗ΛΣ[2] ⊕ H, and that π∗ΛΣ[2] is the null space of h , iΛP [2]. We have just shown that ϕS vanishes on π∗ΛΣ[2], so the Arf invariant of ϕSΛP [2] is defined and equals the Arf invariant of ϕSH , which is m(l/2). (cid:3) 4. Constructing vanishing cycles 4.1. General construction. Our goal is to construct vanishing cycles associated to singular spec- tral curves and to show that the corresponding transvections occur in the monodromy of the Hitchin fibration. These vanishing cycles will occur as a special case of the general construction given in this section. Let X be a Riemann surface which may be non-compact and let f : Y → X be a degree n branched cover satisfying the following conditions: (i) All branch points have ramification index n. (ii) There is an action of the cyclic group Zn on Y by deck transformations. MONODROMY OF THE SL(n) AND GL(n) HITCHIN FIBRATIONS 17 We let t : Y → Y be a generator of the Zn-action. Let b1, b2, . . . , bk denote the branch points of f : Y → X. By assumption each branch point bi ∈ X has a unique ramification point ui ∈ Y lying over it and f −1(bi) = ui. Let i 6= j and suppose that γ : [0, 1] → X is an embedded path in X joining bi to bj for which γ(t) is not a branch point for any t ∈ (0, 1). Then f −1(γ([0, 1])) is the union of n paths γ1, γ2, . . . , γn : [0, 1] → Y , each of which goes from ui to uj. We may order the paths such that tγi = γi+1 for i = 1, . . . , n − 1. We think of γi as 1-chains in Y with boundary uj − ui. Thus γ1 − γ2 is a 1-cycle in Y . Let lγ ∈ H1(Y, Z) be the underlying homology class. Similarly (γ2− γ3), . . . , (γn−1− γn), (γn− γ1) define cycles tlγ, . . . , tn−2lγ, tn−1lγ ∈ H1(Y, Z). Note that while lγ depends on the choice of lift γ1, the collection {lγ, tlγ, t2lγ, . . . , tn−1lγ} depends only on γ. Definition 4.1. We call lγ, tlγ, . . . , tn−1lγ the vanishing cycles associated to γ. Let γ, γ′ be two embedded paths from bi to bj which avoid all other branch points. We say that γ, γ′ are isotopic, if they are homotopic through a path of embedded paths from bi to bj which avoid the other branch points. If γ, γ′ are isotopic then clearly γ and γ′ define the same set of vanishing cycles. 4.2. Vanishing cycles of the An−1 singularity. We now consider a local calculation of vanishing cycles around An−1 singularities. This will subsequently be converted into a global calculation for spectral curves. Let C2 have coordinates (λ, z) and consider the function f : C2 → C given by f (λ, z) = λn + z2. The zero locus of f is a hypersurface in C2 with an isolated singularity at (0, 0). The germ of this hypersurface around (0, 0) is the An−1 plane curve singularity. Let Cn = C2 × Cn−2 have coordinates (λ, z, u1, u2, . . . , un−2) and consider the map f : Cn → Cn−1 given by f (λ, z, u1, u2, . . . , un−2) = (λn + un−2λn−2 + ··· + u2λ2 + u1λ + z2, u1, u2, . . . , un−2). This is a versal deformation of the An−1 singularity. Let B be a sufficiently small open ball around the origin in Cn with closure B and boundary ∂B. For any such B, there exists a sufficiently small open ball B′ around the origin in Cn−1 such that f is a submersion along ∂B ∩ f −1(B′). Let D′ ⊂ B′ be the set of critical values of f restricted to B ∩ f −1(B′). The restriction of f to B ∩ f −1(B′ \ D′) is a smooth fibre bundle over B′ \ D′, the Milnor fibration associated to the germ of f around 0. Let b ∈ B′ \ D′ be a regular value of f in B′ and let X b = f −1(b) ∩ B be the (compact) Milnor fibre. The compact Milnor fibre is a compact manifold with boundary ∂X b = f −1(b) ∩ ∂B. The geometric monodromy [19] of the Milnor fibration is given by a represen- tation ρgeom : π1(B′ \ D′, b) → Iso0(X b, ∂X b), where Iso0(X b, ∂X b) is the group of relative isotopy classes of homeomorphisms of X b which are the identity on ∂X b. Let w0 ∈ C\{0} be small enough that (w0, 0, 0, . . . , 0) ∈ B′. Then we take b = (w0, 0, . . . , 0) ∈ B′ as a basepoint. The fibre of f over b is given by the equation λn + z2 − w0 = 0, which is smooth for any w0 6= 0, so b ∈ B′ \ D′. We have: X b = f −1(b) ∩ B = {(λ, z) λn + z2 − w0 = 0, (λ, z, 0, . . . , 0) ∈ B}. Observe that X b is a branched cover of a closed ball in C via the map (λ, z) 7→ z. This is a degree n cyclic branched cover with branch points ±√w0. As with cyclic spectral curves, we let t : X b → X b We construct loops in B′\D′ representing the generators σ1, . . . , σn−1 as follows. For i = 1, . . . , n−1, we can find a loop in B′ \ D′ exchanging ξi−1q, ξiq along paths joining ξi−1q and ξiq while keeping all other roots fixed. To be specific, let τ + i , τ − i : [0, 1] → C be the paths τ + i (t) = ξi−1(cid:18)1 + i (t) = ξi−1(cid:18)ξ + τ − 1 − ξ 2 ξ − 1 2 (eiπt − 1)(cid:19) q, (eiπt − 1)(cid:19) q. 18 DAVID BARAGLIA be the generator of the cyclic action given by t(λ, z) = (ξλ, z), where ξ = e2πi/n. Let γ : [0, 1] → C be the straight line in C joining −√w0 to √w0 (the choice of which square root of w0 is taken to be √w0 will be unimportant). Associated to γ we have the vanishing cycles lγ, tlγ, . . . , tn−1lγ. Proposition 4.2. The geometric monodromy representation ρgeom : π1(B′\D′, b) → Iso0(X b, ∂X b) is generated by Dehn twists of X b around the loops lγ, tlγ, . . . , tn−2lγ. Remark 4.3. Of course the Dehn twist around tn−1γ is also in the image of the geometric monodromy representation, but can be expressed in terms of lγ, . . . , tn−2lγ. Proof. Recall that D′ is the set of critical values of f restricted to B ∩ f −1(B′). It is easy to see that D′ is given by: D′ = {(w, u1, u2, . . . , un−2) ∈ B′ λn + un−2λn−2 + ··· + u1λ − w has multiple roots }. Then π1(B′ \ D′, b) is the n-th Artin braid group with generators σ1, σ2, . . . , σn−1 exchanging pairs of roots of λn − w0. More precisely, let q ∈ C be an n-th root of w0, so qn = w0. Then λn − qn = (λ − q)(λ − ξq) . . . (λ − ξn−1q). Then we let σi ∈ π1(B′ \ D′, b) be the loop in which ξi−1q moves on τ + i and all other roots fixed. It follows that the geometric monodromy corresponding to σi is given by a Dehn twist around a cycle ci, which we now define. Let τi(t) = ξi−1 (1 + (ξ − 1)t) q be the straight line path joining ξi−1q to ξiq. The cycle ci may be taken to be the pre-image of τi in X b under the branched double cover (λ, z) 7→ λ. We have shown that ρgeom(σi) is a Dehn twist around the loop ci. To conclude we note that it is easy to see that the cycles c1, c2, . . . , cn−1 are homologous to lγ, tlγ, tn−2lγ (possibly after a cyclic re-ordering of lγ, . . . , tn−1lγ). (cid:3) i , ξiq moves on τ − 4.3. Vanishing cycles for spectral curves. In this section we will construct a collection of van- ishing cycles α ∈ H 1(S, Z) such that the corresponding transvections Tα generate the monodromy action of the Hitchin system. We continue to assume that we have chosen a basepoint a0 ∈ Areg of the form a0 = (0, . . . , 0, an), where an has only simple zeros. Let i 6= j and suppose that γ : [0, 1] → Σ is an embedded path in Σ joining bi to bj for which γ(t) is not a branch point for any t ∈ (0, 1). Let lγ, tlγ, . . . , tn−1lγ ∈ H1(S, Z) be the corresponding vanishing cycles. We let cγ, tcγ, . . . , tn−1cγ ∈ H 1(S, Z) be the Poincar´e dual cohomology classes. We note that π∗(tjcγ) = 0 for all j, so tjcγ ∈ ΛP . In particular the transvection Ttj cγ : ΛS → ΛS preserves ΛP . The main result of this section is the following: Theorem 4.4. The transvections Tcγ , Ttcγ , . . . , Ttn−1cγ belong to the SL(n, C) monodromy group. Proof. Let D denote the unit disc in C and choose an embedding e : D → Σ such that bi = e(−1/2), bj = e(1/2) and γ(t) = e(t − 1/2). We can also choose D so that the image e(D) contains no other MONODROMY OF THE SL(n) AND GL(n) HITCHIN FIBRATIONS 19 branch points. For each w ∈ D, let Dw be the degree 2 divisor given by Dw = e(√w) + e(−√w). Then Dw consists of two distinct points for w 6= 0 and D0 = 2e(0). We also have D1/4 = bi +bj. Let D∗ be the divisor D∗ =Pk6=i,j bj. Denote by SmΣ the m-th symmetric product and α : SmΣ → Jacm(Σ) the Abel-Jacobi map, where Jacm(Σ) is the space of degree m line bundles on Σ. We also let eSmΣ denote the subspace of SmΣ consisting of m-tuples of distinct points on Σ. For m > 2g− 2, the restriction eSmΣ → Jacm(Σ) of the Abel-Jacobi map to eSmΣ is known to be a Serre fibration [12]. Now consider the constant map f : D → eSnl−2Σ given by f (w) = D∗. Then α ◦ f is the constant map D → Jacnl−2(Σ) taking the value α(D∗) = Ln ⊗ [D1/4]−1. By the lifting property of Serre fibrations, f is homotopic to a map f ′ : D → eSnl−2Σ for which α(f ′(w)) = Ln ⊗ [Dw]−1. Here we take homotopies relative to the point 1/4 ∈ D. Thus we can assume that f ′(1/4) = D∗. Consider the family of divisors {Dw + f ′(w)}w∈D. We would like it to be true that for all w ∈ D, the divisors Dw and f ′(w) have no points in common. Unfortunately this may not be the case. To avoid this, we will need to change f ′ by a homotopy and also shrink the disc D as we now explain. Define spaces X and Y as follows: let X = {(N, w) ∈ eSnl−2Σ × D [N ] = Ln ⊗ [Dw]−1}. Let Y be the subspace of X consistsing of pairs (N, w) for which some point of N belongs to Dw. Then X and Y naturally fiber over D giving a commutative diagram: i Y ❅ ❅ ❅ ❅ ❅ ❅ ❅ ❅ X p D Clearly X is a complex manifold and Y a subvariety of codimension 1. We also find that the projection p : X → D is a submersion. The map f ′ : D → X is a section of p. We may replace f ′ by a homotopic map f ′′ : D → X, which is a section of p such that f ′′(D) meets Y transversally in smooth points of Y and away from the fibre of Y over 0 ∈ D. We again choose our homotopy relative to the basepoint 1/4 ∈ D, so that f ′′(1/4) = D∗ /∈ Y . So Y meets f ′′(D) in a discrete set of points different from f ′′(1/4) and f ′′(0). Choose an embedded path p(t) : [0, 1] → D in D from 1/4 to 0 with f ′′(p(t)) /∈ Y for all t. Let D′ be a tubular neighborhood of the path p([0, 1]) not meeting any points of Y . We can assume D′ is homeomorphic to a disc. Replacing D with D′, we have obtained a map f ′′ : D′ → X such that f ′′(1/4) = D′ and such that the divisor f ′′(w) does not intersect with Dw for any value of w ∈ D′. Letpp(t) denote the square root of p(t) that goes from −1/2 to 0. We obtain a path p : [0, 1] → D from −1/2 to 1/2 by taking pp(t) from −1/2 to 0 and then taking −pp(t) in reverse from 0 to 1/2. Let γ′ : [0, 1] → Σ be e ◦ p. Then γ′ is isotopic to γ in the space of embedded paths from bi to bj avoiding all other branch points. In particular, the vanishing cycles cγ, tcγ, . . . , tn−1cγ coincide with the cycles cγ ′, tcγ ′, . . . , tn−1cγ ′. Consider now the family of effective divisors {Dw + f ′′(w)}w∈D′ parametrised by w ∈ D′. By construction, this family has the following properties: (i) For every w, the line bundle associate to Dw + f ′′(w) is Ln. (ii) When w = 1/4, the divisor D1/4 + f ′′(1/4) is the divisor of zeros of an. (iii) For w 6= 0, the divisor Dw + f ′′(w) has no multiple points. / / 20 DAVID BARAGLIA (iv) For w = 0, the divisor D0 + f ′′(0) has one point of multiplicity 2 and no other multiple points. (v) If w follows a loop based at 1/4 going once around 0, this has the effect of swapping bi, bj along paths isotopic to γ while all other branch points move on contractible loops. By (i), we have constructed a map d : D′ → P(H 0(Σ, Ln)). Since D′ is contractible this can be lifted to a map a : D′ → H 0(Σ, Ln), where the divisor of a(w) is d(w). We may further choose a so n = a(0) ∈ H 0(Σ, Ln). Then a′ that a(1/4) = an. Let a′ n has a single zero of order 2 and no other multiple zeros. Let x = e(0) ∈ Σ be the double zero of a′ n. Let u ∈ tot(L) be the origin of the fibre n given by λn + a′ of L lying over x. The spectral curve Sa′ n = 0 has exactly one singularity, located n(z) = z2(dz)n and S′ at the point u. In a suitable local coordinate z centered at x we have that a′ a′ n is locally given by λn + z2 = 0, a plane curve singularity of type An−1. In what follows, we aim to show that the monodromy of this An−1 singularity occurs as monodromy of the Hitchin fibration and that this monodromy is generated by Picard-Lefschetz transformations in the vanishing cycles cγ, tcγ, . . . , tn−1cγ. Let A′ be the affine subspace of A consisting of points a = (a2, a3, . . . , an−1, a′ n) whose H 0(Σ, Ln)- n. For each a ∈ A′, the spectral curve Sa passes through the point u ∈ tot(L) term is given by a′ and we may consider the germ of the hypersurface Sa ⊂ tot(L) around the point u. Thus A′ gives a family of deformations of the An−1 singularity of Sa′ n located at u. We claim this is a versal family of deformations. Under the family of deformations of the An−1 singularity defined by the space A′, we have that the germ of λn + z2 is deformed to λn + an−2(z)λn−2 + ··· + a2(z)λ2 + a1(z)λ + z2. The Kodaira-Spencer map for this family (see [19]) is to the map A′ → Cn−2 sending (a2, a3, . . . , an−1, a′ n) to (a2(x), a3(x), . . . , an−1(x)). It is clear that for any point x ∈ Σ, this map surjects to Cn−2, which shows that A′ provides a versal deformation of the singularity. It follows that the geometric monodromy of Section 4.2 is realised as the monodromy of the Hitchin fibration around certain loops in Areg, defined in the vicinity of a′ n. These loops act on S by Dehn twists around the vanishing cycles of this An−1 singularity. Clearly a Dehn twist around a vanishing cycle v ∈ H 1(S, Z) acts on cohomology by the corresponding Picard-Lesfschetz transformation Tv. To complete the proof of the theorem, it remains to identify the vanishing cycles associated to the An−1 singularity of Sa′ n with specific cohomology classes in H 1(S, Z). Let B be a small open ball around u ∈ tot(L). Let U ⊂ A be an open neighbourhood of (0, 0, . . . , 0, a′ n) in A sufficiently small so that for any a ∈ U , the spectral curve Sa is smooth outside of B. Then after possibly further shrinking U , we know that the monodromy of the Hitchin fibration associated to loops in U \ D ∩ U is generated by transvections of the vanishing cycles associated to the An−1 singularity of Sa′ 0 be an embedded path in e(D) joining e(pp(t0)) to e(−pp(t0)), in fact we may take γ′ 0(t) to be the restriction of γ′ to [t0/2, 1− t0/2]. Then by Proposition 4.2, the monodromy of the Hitchin fibration over U \ D ∩ U is generated by transvections in the vanishing cycles cγ ′ 0, Z) associated to γ′ 0 given by restricting a(p(t)) to [0, t0]. The Gauss-Manin connection over this path defines an isomorphism H 1(Sa′ 0 , Z) ≃ H 1(S, Z) and 0 are it is easy to see that the vanishing cycles cγ ′ mapped to the vanishing cycles cγ ′, tcγ ′, . . . , tn−1cγ ′ ∈ H 1(S, Z) associated to γ′. This proves the theorem. 0. Consider the path in Areg joining an to a′ 0, Z) associated to γ′ n . Let t0 ∈ [0, 1) be such that a′ 0 = a(p(t0)) ∈ U . Let γ′ 0, tcγ ′ 0 , . . . , tn−1cγ ′ 0, . . . , tn−1cγ ′ 0 ∈ H 1(Sa′ 0 ∈ H 1(Sa′ (cid:3) MONODROMY OF THE SL(n) AND GL(n) HITCHIN FIBRATIONS 21 Remark 4.5. In the process of proving Theorem 4.4 we constructed a family a : D′ → H 0(Σ, Ln) of sections of Ln parametrised by a space D′ homeomorphic to a disc. Moreover, D′ contained exactly one point 0 ∈ D′ for which the corresponding spectral curve was singular. If we take a loop in D′ that winds once anti-clockwise around 0, the corresponding monodromy transformation is seen to be Tcγ Ttcγ Tt2cγ . . . Ttn−2cγ . Next, we will apply Theorem 4.4 to construct some specific examples of vanishing cycles. In fact we will later see that the vanishing cycles constructed below already are sufficient to generate the monodromy group. Recall from Section 3.1 that on choosing a trivialising disc i : D → Σ, we obtain a decomposition ΛS = ΛS,0 ⊕ ΛS,1, where ΛS,0 is spanned by the cycles {tjci}i,j and ΛS,1 can be identified with Z[t]/htn − 1i ⊗Z ΛΣ. We note that once a choice of trivialising disc is given, the identification (4.1) ΛS,1 ≃ Z[t]/htn − 1i ⊗Z ΛΣ is determined only up to composition with a power of t. Proposition 4.6. For i = 1, 2, . . . , k − 1, there exists a path from bi to bi+1 whose associated vanishing cycles are {ci, tci, . . . , tn−1ci}. Let a1, b1, . . . , ag, bg be a symplectic basis for ΛΣ. For u = 1, 2, . . . , g, there exist paths γau and γbu from b1 to b2 such that under a suitable choice of identification (4.1), the associated vanishing cycles are {c1 +(1−t)au, t(c1 +(1−t)au), . . . , tn−1(c1 + (1 − t)au)} and {c1 + (1 − t)bu, t(c1 + (1 − t)bu), . . . , tn−1(c1 + (1 − t)bu)}. Proof. Let γi be the path from bi to bi+1 constructed in Section 3.2 (γi is depicted in Figure 1). Then by the definition of ci, tci, . . . , tn−1ci, it is clear that these are the vanishing cycles associated to γi. Consider a path γa1 in Σ starting at b1 and moving left to the boundary of the trivialising disc, going around the loop a1, then returning back to the trivialising disc and terminating at b2. The corresponding vanishing cycles will clearly have the form tj(c1 + (1− t)twa1), for some value of w. Similarly define paths γb1 , . . . , γag , γbg . The corresponding vanishing cycles will have the form tj(c1 + (1− t)twau) or tj(c1 + (1− t)twbu) for the same value of w. Choosing a different identification in (4.1) if necessary, we may assume w = 0. (cid:3) 5. The vanishing lattice 5.1. Construction of vanishing lattice on ΛP . In this section we give ΛP the structure of a vanishing lattice (ΛP ,h , i, ∆P ), made out of vanishing cycles constructed in Section 4. We then recall the classification of vanishing lattices in [16, 17] and use these results to classify the vanish- ing lattice on ΛP . In Section 6 we will show that the group generated by transvections in ∆P is precisely the monodromy group ΓSL of the Hitchin fibration. Vanishing lattices were classified for R = Z2 in [16] and R = Z in [17]. If (V,h , i, ∆) is a vanishing lattice over Z then we obtain a vanishing lattice (V ,h , i, ∆) over Z2, where V = V ⊗Z Z2 and ∆ is the image of ∆ under mod 2 reduction. Proposition 5.1. Let V be a free R-module of finite rank with alternating bilinear form h , i. Let S ⊆ V be a subset of V satisfying the following conditions: (i) S spans V . (ii) For any two distinct elements u, v ∈ S, there exists a sequence u = s1, s2, . . . , sm = v of elements of S such that hsi, si+1i = ±1 for, 1 ≤ i ≤ m − 1. 22 DAVID BARAGLIA Let ΓS be the subgroup of Sp#V generated by {Ts}s∈S and set ∆ = ΓS · S. Then (V,h , i, ∆) is a vanishing lattice and Γ∆ = ΓS. Proof. Clearly ∆ spans V , as S spans V . If V has rank µ > 1 then by (i), S has more than one element. Hence by (ii) there exists δ1, δ2 ∈ S ⊆ ∆ with hδ1, δ2i = 1. Next, we claim that for any two elements u, v ∈ S, we have u = g(v) for some g ∈ ΓS. By (ii) it suffices to show this in the case that hu, vi = ±1. The claim follows, as u = TvTu(v), if hu, vi = 1 and u = T −1 u (v), if hu, vi = −1. Our claim shows that ∆ is a ΓS-orbit, hence also a Γ∆-orbit. Thus (V,h , i, ∆) is a vanishing lattice. Moreover, since ∆ is a ΓS-orbit, we have that for any α ∈ ∆ there exists u ∈ S and g ∈ ΓS with α = g(u). Then Tα = g ◦ Tu ◦ g−1 ∈ ΓS, hence Γ∆ = ΓS. v T −1 (cid:3) Consider ΛP equipped with the intersection form h , i. We will give (ΛP ,h , i) the structure u=1 be a symplectic basis for ΛΣ. Let SP ⊂ ΛP be the of a vanishing lattice. For this let {au, bu}g subset: SP = {tjci} 0≤j≤n−2 1≤i≤nl−1 ∪ {tj(c1 + (1 − t)au)}0≤j≤n−2 1≤u≤g ∪ {tj(c1 + (1 − t)bu)}0≤j≤n−2 1≤u≤g By Proposition 4.6, we see that elements of SP are vanishing cycles associated to certain paths in Σ. By Theorem 4.4, it follows that the transvections {Tv}v∈SP belong to the SL(n, C)-monodromy group ΓSL. Clearly SP satisfies the conditions of Proposition 5.1, so we obtain a vanishing lattice (ΛP ,h , i, ∆P ), where Γ∆P is the group generated by transvections in SP and ∆P = Γ∆P · SP . So Γ∆P is a subgroup of ΓSL. We will eventually show that Γ∆P = ΓSL. Remark 5.2. Let VC ⊂ ΛP be the set of all vanishing cycles associated to paths γ joining pairs of branch points in Σ and let ΓVC be the group generated by transvections in VC. We will eventually be able to show that ∆P = ΓVC · VC and ΓVC = Γ∆P , so that (ΛP ,h , i, ∆P ) is the same as the vanishing lattice described in the introduction. 5.2. Classification over Z2. We recall the classification of vanishing lattices in [16, 17]. Since the classification over Z depends on the classification over Z2, we begin with the Z2 case. In this section V will be a finite dimensional Z2 vector space of dimension µ, equipped with an alternating bilinear form h , i. Let V 0 be the null space of h , i and let p be the dimension of V 0. Then µ = 2r + p for some integer r. If q is a quadratic function on V , we see that the transvection Tv associated to v ∈ V preserves q if and only if q(v) = 1. We observe that to any basis B of V , we can associate a unique quadratic function qB with the property that qB(v) = 1 for all v ∈ B. Clearly the group generated by transvections by elements of B preserves qB. Definition 5.3. Let (V ,h , i, ∆) be a vanishing lattice. A basis B of V is called weakly distinguished if Γ∆ is generated by {Tv}v∈B. In this case, Γ∆ preserves qB, hence qB(α) = 1 for all α ∈ ∆. Note that a vanishing lattice does not necessarily admit a weakly distinguished basis. To any basis B of V , we construct a graph Gr(B) as follows. The vertices of Gr(B) are elements of B and for every distinct pair of element u, v ∈ B, there is a single edge joining u and v if and only if hu, vi = 1. Remark 5.4. If (V ,h , i, ∆) is a vanishing lattice and B a weakly distinguished basis consisting of elements of ∆, then (V ,h , i, ∆) can be completely recovered from Gr(B). Indeed, Γ∆ is the group MONODROMY OF THE SL(n) AND GL(n) HITCHIN FIBRATIONS 23 generated by {Tv}v∈B and ∆ = Γ∆ · B. Note however that different graphs can give rise to the same underlying vanishing lattice. We now state the classification in [16] of vanishing lattices over Z2. If (V ,h , i, ∆) is a van- ishing lattice which admits a weakly distinguished basis B whose elements belong to ∆, then by Remark 5.4 it is enough to simply give the graph Gr(B). The remaining cases will be described individually. Vanishing lattices over Z2 can be broadly classified into three main types: symplectic, orthogonal and special. In both the orthogonal and special cases, the are three sub-cases to consider. Case 1: Symplectic. In this case ∆ = V \ V 0 and Γ∆ = Sp#V . Symplectic vanishing lattices do not admit weakly distinguished bases and the group Sp#V does not preserve any quadratic functions. For fixed values of (r, p) there is only one symplectic vanishing lattice, which is denoted by Sp#(2r, p). Case 2: Orthogonal. In this case, a weakly distinguished basis B exists, ∆ = {v ∈ V \ V 0 qB(v) = 1} and Γ∆ = O#(qB) is the subgroup of Sp#V preserving qB. There are three sub-cases which are distinguished according to whether the Arf invariant of qB is 0, 1 or undefined. In all cases, a weakly distinguished basis can be chosen so that Gr(B) is one of the following: 2 3 4 5 for which Arf(qB) =(1 if r = 2, 3 (mod 4), 0 if r = 0, 1 (mod 4). 1 3 4 5 6 7 2 1 for which Arf(qB) =(1 if r = 0, 1 (mod 4), 0 if r = 2, 3 (mod 4). 2 3 4 5 6 1 for which Arf(qB) is undefined. 2r 2r + 1 2r + p 2r 2r + 1 2r + p 2r + 1 2r + 2 2r + p For fixed (r, p) there are at most three orthogonal vanishing lattices, according to whether the 1 (2r, p) and O#(2r, p). Note Arf invariant is 0, 1 or undefined. We denote these by O# 0 (2r, p), O# 24 DAVID BARAGLIA that O#(2r, p) only exists for p > 0. Case 3: Special. There are three sub-cases, of which two admit weakly distinguished bases. We describe these cases first. In the first subcase B can be chosen so that Gr(B) is: 1 2 2r − 1 2r 2r + 1 2r + p We have Arf(qB) =(1 if r = 1, 2 (mod 4), 0 if r = 0, 3 (mod 4). In the second subcase, Gr(B) is: This vanishing lattice is denoted Aev(2r, p). 1 2 2r 2r + 1 2r + 2 2r + p This vanishing lattice is denoted Aodd(2r, p). We have Arf(qB) = 1 if r = 1 (mod 4), 0 if r = 3 (mod 4), undefined if r = 2, 4 (mod 4). The third subcase is obtained by taking the vanishing lattice Aodd(2r, p + 1) and taking the quotient by the subspace spanned by e1 + e3 + e5 + ··· + e2r+1, where ei denotes the basis vector corresponding to the i-th vertex. The resulting vanishing lattice is denoted by A′(2r, p). Note that the quadratic form of the Aodd(2r, p + 1) vanishing lattice descends to the quotient if and only if r is odd. The following is a very useful criterion for determining whether or not a vanishing lattice is special: Proposition 5.5 ([16], Proposition 4.13). Let (V ,h , i, ∆) be a vanishing lattice over Z2. The following are equivalent: (1) (V ,h , i, ∆) is not of special type. (2) There exists V 1 ⊆ V , ∆1 ⊆ ∆ such that (V 1,h , i, ∆1) is a vanishing lattice of type O# 1 (6, 0). The condition that (V ,h , i, ∆) contains O# 1 (6, 0) can be re-stated more simply as the condition that there exists e1, e2, e3, e4, e5, e6 ∈ ∆ such that the graph of {e1, e2, e3, e4, e5, e6} is the E6 Dynkin diagram: 2 3 4 5 6 1 MONODROMY OF THE SL(n) AND GL(n) HITCHIN FIBRATIONS 25 Let (ΛP ,h ,i, ∆P ) be the vanishing lattice constructed in Section 5.1 and let (ΛP [2],h , i, ∆P ) be its mod 2 reduction. Theorem 5.6. Let µ = (n − 1)(nl + 2g − 2), let p = 2g if n is even and p = 0 if n is odd. Further define r such that µ = 2r + p. The vanishing lattice (ΛP [2],h , i, ∆P ) is isomorphic to: (1) A′(2l − 2, 2g), if n = 2, (2) O# (3) O# (4) Sp#(2r, 2g), if n is even, l is odd and n > 2. a (2r, 0), where a = (m(m − 1)/2)l, if n = 2m + 1 is odd, a (2r, 2g), where a = m(l/2), if n = 2m and l are even and n > 2, Proof. Consider first the case that n = 2. Then SP = {c1, c2, . . . , c2l−1, c1 + (1 − t)a1, c1 + (1 − t)b1, . . . , c1 + (1 − t)ag, c1 + (1 − t)bg}. The intersection graph of SP is given by: c2l−1 c2l−2 c2 c1 c1 + (1 − t)a1 c1 + (1 − t)bg So (ΛP [2],h , i, ∆P ) must either be of type Aodd(2l − 2, 2g + 1) or A′(2l − 2, 2g). However, we have the relation c1 + c3 + c5 + ··· + c2l−1 = 0, so the vanishing lattice is of type A′(2l − 2, 2g). Consider the case n > 2. We will show that (ΛP [2],h , i, ∆P ) contains a copy of O# 1 (6, 0). In fact, consider the elements tc3, tc1, c2 + tc2, c3, c4, c5 ∈ ∆P . Note that c2 + tc2 ∈ ∆P because c2 + tc2 = Tc2(tc2). The intersection graph of these elements is the E6 Dynkin diagram: tc1 c2 + tc2 c3 c4 c5 tc3 Hence by Proposition 5.5, the vanishing lattice (ΛP [2],h , i, ∆P ) is not special for n > 2. If n is odd or n and l are even, we saw in Section 3.5, that ΛP [2] has a monodromy invariant qua- dratic function q and we calculated the Arf invariant of q in Proposition 3.12. Thus if n is odd or n and l are both even, then (ΛP [2],h , i, ∆P ) is of orthogonal type with the specified Arf invariant. Lastly, in the case that n > 2 is even and l is odd we will show that (ΛP [2],h , i, ∆P ) is not of orthogonal type, hence it must be symplectic. Suppose on the contrary that there is a quadratic . Then we must have q(tj ci) = 1 for all i and j. From function q : ΛP [2] → Z2 invariant under Γ∆P this it follows that q(c1 + (1 + t)c2 +··· + (1 + t +··· + tn−2)ck−1) = 1. But this is impossible, since c1 + (1 + t)c2 + ··· + (1 + t + ··· + tn−2)ck−1 = 0. (cid:3) 26 DAVID BARAGLIA 5.3. Classification over Z. Let (V,h , i, ∆) be a vanishing lattice over Z. Recall [4] that for an alternating bilinear form h , i on V there exists a basis {e1, f1, . . . , er, fr, g1, . . . , gp} of V for which the matrix of h , i has the form d1 0 0 −d2 d2 0 . . . 0 −d1  0 −dr dr 0 0 . . .  0 where the di are positive integers and di divides di+1 (i = 1, . . . , r − 1). The di are called the elementary divisors of h , i and are uniquely determined. Let (V ,h , i, ∆) be the mod 2 reduction, let V 0 be the null space of h , i on V . Choose an element δ ∈ ∆ and let V 00 = {v ∈ V 0 v + δ ∈ ∆}. Then V 00 is a subspace of V 0 and does not depend on the choice of δ [16, Lemma 2.11]. Let j : V → V ∗ be the homomorphism j(x) = hx, i and consider the homomorphism j−1(2V ∗) → j−1(2V ∗)/2V ≃ V 0 → V 0/V 00. In [16] it is shown that V 0/V 00 is either 0 or Z2. Hence we obtain a homomorphism φ : j−1(2V ∗) → Z2, where in the case V 0/V 00 = 0, we take φ = 0. Define k0(V ) = max{k φ(j−1(2kV ∗)) 6= 0}, with the conventions that k0(V ) = ∞ if no such maximum value of k exists and k0(V ) = 0 if φ = 0. Then we have: Theorem 5.7 ([17], Theorem 7.5). Let (V1,h , i, ∆1) and (V2,h , i, ∆2) be vanishing lattices over Z. They are isomorphic if and only if (i) they are isomorphic as lattices with bilinear form, (ii) their mod 2 reductions are isomorphic as vanishing lattices and (iii) k0(V1) = k0(V2). Thus integral vanishing lattices are classified by their mod 2 reduction, the invariants d1, d2, . . . , dr, p of the bilinear form and the invariant k0. Moreover, we have that V 00 = V 0 except in the O# and Aodd cases. Following Janssen, we use notation like O#(d1, . . . , dr; p; k0) to denote a vanishing lat- tice with mod 2 reduction of type O# and invariants d1, . . . , dr, p, k0. In all cases other than O# and Aodd we may omit k0 from the notation. We will denote by u the number of di's which are odd. MONODROMY OF THE SL(n) AND GL(n) HITCHIN FIBRATIONS 27 Theorem 5.8 ([17], Theorem 7.8). Up to isomorphism integral vanishing lattices are given by the following list: O# O# 1 (d1, . . . , dr; p) 0 (d1, . . . , dr; p) O#(d1, . . . , dr; p; k0) Sp#(d1, . . . , dr; p) Aev(d1, . . . , dr; p) Aodd(d1, . . . , dr; p; k0) A′(d1, . . . , dr; p) (u ≥ 3) (u ≥ 2, r > u or p > 0, k0 > 0) (if u = 1, then r = 1 and p = 0) (r > u or p > 0; k0 = 0 iff u = 1) (u ≥ 2). Now we can determine the isomorphism class of the integral vanishing lattice (ΛP ,h , i, ∆P ): Theorem 5.9. The vanishing lattice (ΛP ,h , i, ∆P ) is isomorphic to: (1) A′(1, 1, . . . , 1, 2, 2, . . . , 2; 0), if n = 2, (2) O# (3) O# (4) Sp#(1, 1, . . . , 1, n, n, . . . , n; 0), if n is even, l is odd and n > 2. a (1, 1, . . . , 1, n, n, . . . , n; 0), where a = (m(m − 1)/2)l, if n = 2m + 1 is odd, a (1, 1, . . . , 1, n, n, . . . , n; 0), where a = m(l/2), if n = 2m and l are even and n > 2, In this classification, the number of 1's is (n − 2)(g − 1) + n(n − 1)l/2 − 1 and the number of n's is g. Proof. The mod 2 classification of (ΛP ,h , i, ∆P ) was given in Theorem 5.6. In particular we saw that (ΛP ,h , i, ∆P ) does not have type O# or Aodd, so we do not need to consider the invariant k0. The intersection form h , i is non-degenerate on ΛP , so p = 0 and the invariants (d1, . . . , dr) are precisely the polarization type of the Prym variety P rym(S, Σ). This was calculated in Corollary 3.9. (cid:3) 6. Proof of the main theorem Let δ be a smooth point of the discriminant locus D. Let Dδ be a copy of the unit disc in C, embedded in A such that Dδ intersects D transversally at the point δ and meets no other point of D. Consider a loop γδ in A \ D, which starts at the basepoint a0, follows a path p from a0 to the boundary of Dδ, goes once around the boundary of Dδ and goes back to a0 along p−1. Such a loop will be called a meridian. Kouvidakis and Pantev showed in the case L = K, that the discriminant locus D ⊆ A is an irre- ducible hypersurface [18]. Their proof easily extends to the case where L 6= K and deg(L) > deg(K). It follows that π1(A\ D, a0) is generated by meridians and since D is irreducible, any two meridians are conjugate in π1(A\D, a0). Let D0 ⊂ D be the locus of points a ∈ D for which the corresponding spectral curve Sa is irreducible and has an ordinary double point as its only singularity. Kouvidakis and Pantev also showed that D0 is a non-empty Zariski open subset of D, in the case L = K [18]. This is clearly also true in the case L 6= K and deg(L) > deg(K). Since D0 is Zariski dense in D, we see that π1(A \ D, a0) is generated by meridians around points in D0. Let δ ∈ D0 and suppose that Dδ is an embedded disc meeting D transversally in δ, as above. Consider the family X → Dδ of spectral curves over Dδ defined by pullback under the incusion 28 DAVID BARAGLIA Dδ ⊂ A. The fibre Xδ over δ of this family has an isolated non-degenerate singularity. It follows that the monodromy ρ(γδ) ∈ Aut(ΛS) of a meridian γδ around δ is a Picard-Lefschetz transformation: ρ(γδ)(x) = Tα(x) = x + hα, xiα, where α ∈ ΛS is the vanishing cycle associated to γδ. In fact, since π∗ ◦ ρ(γδ) = π∗, we see that α ∈ ΛP . We then have: Lemma 6.1. Let ∆ ⊆ ΛP be the set of vanishing cycles associated to meridians around points in D0. The monodromy group ΓSL is generated by the transvections {Tα}α∈∆. For any two vanishing cycles α, β ∈ ∆, we have α = gβ or α = −gβ for some g ∈ ΓSL. Proof. Only the last statement of the lemma requires proof. Let α, β ∈ ∆. Then Tα = ρ(γ1), Tβ = ρ(γ2) for some meridians γ1, γ2. But we have seen that all meridians are conjugate, so there exists x ∈ π1(A \ D, a0) for which γ1 = xγ2x−1. Applying ρ, we get Tα = gTβg−1 = Tgβ, where g = ρ(x) ∈ ΓSL. As h , i is non-degenerate it is easy to see that Tα = Tgβ implies that α = gβ or α = −gβ. Lemma 6.2 ([16], Theorem 2.9). Let (V,h , i, ∆) be an integral vanishing lattice and x ∈ V . Then x ∈ ∆ if and only of there exists y ∈ V and δ ∈ ∆ such that hx, yi = 1 and x − δ ∈ 2V . Theorem 6.3. Let (ΛP ,h , i, ∆P ) be the integral vanishing lattice constructed in Section 5.1 and Γ∆P ⊆ Aut(ΛP ) the group generated by transvections by elements of ∆P . We have an equality ΓSL = Γ∆P . Proof. We have already established the inclusion Γ∆P ⊆ ΓSL, which follows by Theorem 4.4 and Proposition 4.6. It remains to prove the reverse inclusion. (cid:3) Consider first the case n > 2. By Lemma 6.1, it is sufficient to show that Tα ∈ Γ∆P , where α is the vanishing cycle of a meridian around a point in D0. It is easy to see that the vanishing cycles {tjci} constructed in Section 4 are vanishing cycles associated to meridians. Applying Lemma 6.1, we have that there exists a g ∈ ΓSL for which α = gc1 or α = −gc1. As (ΛP ,h , i, ∆P ) is a vanishing lattice, it can be shown that there exists h ∈ Γ∆P such that hc1 = −c1 [16]. So replacing g by gh if necessary, we can assume α = gc1. If n is odd or n and l are both even, then (ΛP ,h , i, ∆P ) is of orthogonal type. Let ∆P ⊂ ΛP [2] be the mod 2 reduction of ∆P . Then as explained in Section 5.2, ∆P = {v ∈ V \ V 0 q(v) = 1}, where V = ΛP [2] and q is the monodromy invariant quadratic function. It follows that the mod 2 reduction of α belongs to ∆P , since q(gc1) = q(c1) = 1. More- over hα, yi = 1, where y = g(tc1), so by Lemma 6.2 we have α ∈ ∆P . If n is even and l is odd, then (ΛP ,h , i, ∆P ) is of symplectic type. So ∆P = V \ V 0 and by a similar argument we have α ∈ ∆P . Lastly, consider the case n = 2. It was shown in [2] that ΓSL is generated by transvections Tcγ , where cγ is the cycle associated to a path γ joining two branch points, as in Theorem 4.4. It is also clear that cγ is the vanishing cycle of a meridian. Suppose that γ joins branch points bi, bj and assume i < j (the case i > j is similar). Let c ∈ ΛP [2] be the mod 2 reduction of cγ. Then it is easy to see that c must have the form c = ci + ci+1 + ··· + cj−1 + (1 + t)a for some a ∈ ΛΣ[2]. In the case n = 2, we have that ∆P ⊂ ΛP [2] is of special type. Then using [16, Lemma 3.11], we see that c ∈ ∆P . By the same argument as in the n > 2 case, we have cγ = g(c1) for some g ∈ ΓSL. So hcγ, yi = 1 for y = g(c2). Then by Lemma 6.2, we have cγ ∈ ∆P . (cid:3) The next theorem shows that the monodromy of the GL(n, C) Hitchin fibration is determined by the vanishing lattice (ΛP ,h , i, ∆P ) together with the extension ΛP → ΛS → ΛΣ. MONODROMY OF THE SL(n) AND GL(n) HITCHIN FIBRATIONS 29 Theorem 6.4. The monodromy group ΓGL is the subgroup of Aut(ΛS) generated by transvections Tα : ΛS → ΛS, where α ∈ ∆P . Proof. By the same argument used for ΓSL, we have that ΓGL is the subgroup of Aut(ΛS) generated by transvections in vanishing cycles associated to meridians around points in D0. Let ΓGL,∆P be the subgroup of Aut(ΛS) generated by transvections in ∆P . Recall from Section 5.1 the subset SP ⊂ ∆P . We let ΓGL,SP be the subgroup of Aut(ΛS) generated by transvections in SP . In the proof of Theorem 6.3, we established that if α is such a vanishing cycle, then α ∈ ∆P . Hence ΓGL ⊆ ΓGL,∆P . On the other hand we have that the elements of SP are vanishing cycles associated to meridians, so ΓGL,SP ⊆ ΓGL. Arguing as in the proof of Proposition 5.1, we see that the subgroup of Aut(ΛS) generated by transvections in ∆P is also generated by transvections in SP . Hence ΓGL,SP = ΓGL = ΓGL,∆P . In particular, ΓGL is the subgroup of Aut(ΛS) generated by transvections in ∆P . (cid:3) Proposition 6.5. Let VC ⊂ ΛP be the set of vanishing cycles associated to paths in Σ joining pairs of branch points and let ΓVC be the group generated by transvections by cycles in VC. Then ∆P = ΓVC · VC and ΓVC = Γ∆P = ΓSL. Proof. The elements of SP are all vanishing cycles associated to certain paths joining branch points, so SP ⊆ VC and ΓSP ⊆ ΓVC, where ΓSP is the group generated by transvections in SP . By Proposition 5.1, we have ΓSP = Γ∆P and Γ∆P = ΓSL by Theorem 6.3, so ΓSL ⊆ ΓVC. On the other hand ΓVC ⊆ ΓSL, by Theorem 4.4. So ΓVC = ΓSL. It remains to show that VC ⊆ ∆P . Let α ∈ VC. Then as Tα is the monodromy around a meridian, arguing along the same lines as in the proofs of Lemma 6.1 and Theorem 6.3 gives α = gc1 for some g ∈ ΓSL. But gc1 ∈ ∆P , so we have shown that VC ⊆ ∆P and hence ΓVC · VC ⊆ ΓSL · ∆P = ∆P . However, ∆P is an orbit of ΓSL = ΓVC, so we must have ΓVC · VC = ∆P . (cid:3) 7. Application to the topology of Higgs bundle moduli spaces Let M(n, d, L) denote the moduli space of rank n L-twisted Higgs bundles with trace-free Higgs field and determinant equal to a fixed line bundle D of degree d. Up to isomorphism, M(n, d, L) depends on D only through the degree d. We can define the Hitchin fibration h : M(n, d, L) → A and one finds that for any value of d, the moduli space M(n, d, L) is a torsor for the family of Prym varieties p : P rym(S/Areg) → Areg as defined in Section 2. In particular, this implies that the monodromy local system of M(n, d, L) is Λ, independent of d. Let a ∈ Areg and let Fa = h−1(a) be the corresponding non-singular fibre of the Hitchin fibration. Our goal this section will be to prove the following: Theorem 7.1. Let ω ∈ H 2(Fa, Q) be the cohomology class of the polarization on Fa. The image Im(H ∗(M(n, d, L), Q)) → H ∗(Fa, Q)) of the restriction map in cohomology is the subspace spanned by 1, ω, ω2, . . . , ωu, where u = dimC(Fa) is the dimension of the fibre. To prove this theorem we need to use a result concerning symplectic vector spaces over finite fields of prime order. Let p be an odd prime, V a vector space over Zp of dimension 2v and h , i a non-degenerate alternating bilinear form over V . Given a symplectic basis {e1, f1, e2, f2, . . . , ev, fv} 30 DAVID BARAGLIA and m ∈ {0, 1, 2, . . . , v}, let α2m ∈ ∧2mV be given by (7.1) α2m = Xi1<i2<···<im (ei1 ∧ fi1) ∧ ··· ∧ (eim ∧ fim). It can be shown that α2m is independent of the choice of symplectic basis and is invariant under the group Sp(V, Zp) of symplectic transformations of V [11]. Lemma 7.2. The subspace of ∧∗V invariant under Sp(V, Zp) is spanned by 1, α2, α4, . . . , α2v. Proof. We use induction on the dimension 2v of V . Assume the result holds in dimension 2v − 2. Choose a symplectic basis B = {e1, f1, . . . , ev, fv} for V . For i = 1, 2, . . . , v, let Vi be the subspace spanned by B \ {ei, fi}, let ιi : Vi → V be the inclusion and πi : V → Vi the projection with kernel spanned by ei, fi. Let λ ∈ ∧kV be invariant. Consider first the case where k is odd. For any i we have that πi(λ) is invariant under Sp(Vi, Zp), so πi(λ) = 0 by induction. It follows that λ = ei ∧ α + fi ∧ β + ei∧ fi ∧ γ, where α, β, γ are in the image of ιi : ∧∗Vi → ∧∗V . But λ is invariant under the transvections Tei , Tfi and it follows easily that α = β = 0, so that λ is a multiple of ei ∧ fi. Since i was arbitrary, we have that λ is a multiple of e1∧f1∧···∧ev∧fv. But λ has odd degree so this can only happen if λ = 0. Consider the case where k = 2m is even. Then for each i, we have that πi(λ) is invariant under Sp(Vi, Zp), so by induction we have πi(λ) = ciπi(α2m) for some ci ∈ Zp. Clearly this can only happen if c1 = c2 = ··· = cm and thus πi(λ − c1α2m) = 0 for all i. By the same argument as used in the case where k is odd, we see that either λ − c1α2m = 0, or m = v and λ − c1α2v is a multiple of e1 ∧ f1 ∧ ··· ∧ ev ∧ fv = α2v. In either case λ is a multiple of α2m. Lemma 7.3. The space H ∗(Fa, Q)ρSL of monodromy invariant rational cohomology classes on Fa is spanned by 1, ω, ω2, . . . , ωu. Proof. Suppose that µ ∈ H k(Fa, Q)ρSL is a monodromy invariant cohomology class on Fa. Multi- plying µ by a sufficiently large positive integer, it suffices to assume that µ ∈ H k(Fa, Z)ρSL . Let p be an odd prime not dividing n and let V = Λ∗ P , we have H k(Fa, Zp) = ∧kV . By Corollary 3.9, the polarization type of h , i on ΛP is (1, 1, . . . , 1, n, n, . . . , n), where there are g copies of n. It follows that the mod p reduction of h , i is a non-degenerate alter- nating bilinear form on ΛP ⊗Z Zp and so by duality induces a non-degenerate alternating bilinear form h , i on V preserved by the monodromy action. Reduction mod p thus induces a homomor- phism φ : ΓSL → Sp(V, Zp) and the reduction of µ mod p gives an element µp ∈ ∧kV invariant under φ(ΓSL). Since p is odd, it follows from [5, Theorem 2.7 and Theorem 6.5] and the fact that ΓSL is generated by transvections in the set SP given in Section 5.1, that φ is actually surjective. Then by Lemma 7.2, we have that µp = 0 if k is odd and µp is a multiple of α2m if k = 2m is even. P ⊗Z Zp. Then since H k(Fa, Z) = ∧kΛ∗ (cid:3) Suppose first that k is odd. Then µ is divisible by infinitely many primes, hence µ = 0. Now suppose that k = 2m is even. Then for every odd prime p not dividing n we have that the mod p reduction of µ is a multiple of α(p) 2m (here we use a superscript p to remind us that α2m = α(p) 2m depends on p). Suppose also that p > m. Then from Equation (7.1), we have that α(p) 2m is the mod p reduction of ωm/m!, where ω ∈ ∧2Λ∗ P is the alternating form h , i thought of as a 2-form. P be a primitive vector such that ωm/m! is an integral multiple of τ . The mod p Let τ ∈ ∧2mΛ∗ reduction of µ is a multiple of the mod p reduction of τ for infinitely many primes p. It follows that µ is in the Z-span of τ and thus µ is a rational multiple of ωm in ∧2mΛ∗ (cid:3) P ⊗Z Q. MONODROMY OF THE SL(n) AND GL(n) HITCHIN FIBRATIONS 31 Proof of Theorem 7.1. As the restriction map r : H ∗(M(n, d, L), Q) → H ∗(Fa, Q) factors through Mreg(n, d, L), it is clear that the image of r is contained in H ∗(Fa, Q)ρSL , the subgroup of mon- odromy invariants. By Lemma 7.3, H ∗(Fa, Q)ρSL is spanned by 1, ω, . . . , ωu. On the other hand since M(n, d, L) is quasi-projective, there exists a class α ∈ H 2(M(n, d, L), Q) whose restriction r(α) to Fa is a Kahler class, hence non-zero. Thus r(α) must be some non-zero rational multiple of ω. It follows that the image of r is precisely the span of 1, ω, ω2, . . . , ωu. (cid:3) Remark 7.4. When n and d are coprime and L = K, Theorem 7.1 may also be proved as follows. By Theorem 7 of [20], the cohomology ring of M(n, d, K) is generated by the Kunneth factors of the Chern classes of the universal P GL(n, C)-Higgs bundle on M(n, d, K)×Σ. Then by a generalisation of the proof of Proposition 5.1.2 in [7], one can determine the image of the generators on restriction to a non-singular fibre. The advantage of our proof is that it applies without restriction on the values of n and d and does not require knowing a set of generators for the cohomology of M(n, d, L). References 1. M. F. Atiyah, Riemann surfaces and spin structures. Ann. Sci. ´Ecole Norm. Sup. Vol. 4, no. 4 (1971) 47-62. 2. D. Baraglia, L. Schaposnik, Monodromy of rank 2 twisted Hitchin systems and real character varieties, to appear in Trans. Amer. Math. Soc., arXiv:1506.00372 (2015). 3. A. Beauville, M. S. Narasimhan, S. Ramanan, Spectral curves and the generalised theta divisor. J. Reine Angew. Math. 398 (1989), 169-179. 4. N. Bourbaki, ´El´ements de math´ematique. Alg`ebre. Chapitre 9. Reprint of the 1959 original. Springer-Verlag, Berlin, 2007. 211 pp. 5. R. Brown, S. P. Humphries, Orbits under symplectic transvections. I. Proc. London Math. Soc. (3) 52 (1986), no. 3, 517-531. 6. M. A. A. de Cataldo, L. Migliorini, The decomposition theorem, perverse sheaves and the topology of algebraic maps. Bull. Amer. Math. Soc. (N.S.) 46 (2009), no. 4, 535-633. 7. M. A. A. de Cataldo, T. Hausel, L. Migliorini, Topology of Hitchin systems and Hodge theory of character varieties: the case A1. Ann. of Math. (2) 175 (2012), no. 3, 1329-1407. 8. M. A. A. de Cataldo, A support theorem for the Hitchin fibration: the case of SLn, arXiv:1601.02589 (2016). 9. P.-H. Chaudouard, G. Laumon, Un th´eor`eme du support pour la fibration de Hitchin. Ann. Inst. Fourier (Greno- ble) 66 (2016), no. 2, 711-727. 10. D. J. Copeland, Monodromy of the Hitchin map over hyperelliptic curves. Int. Math. Res. Not. (2005), no. 29, 1743-1785. 11. B. De Bruyn, On the Grassmann modules for the symplectic groups. J. Algebra 324 (2010), no. 2, 218-230. 12. I. Dolgachev, A. Libgober, On the fundamental group of the complement to a discriminant variety. In: Algebraic geometry, Lecture Notes in Math. 862, 1-25, Springer, Berlin-New York, 1981. 13. D. Gaiotto, G. W. Moore, A. Neitzke, Four-dimensional wall-crossing via three-dimensional field theory. Comm. Math. Phys. 299 (2010), no. 1, 163-224. 14. N. J. Hitchin, The self-duality equations on a Riemann surface. Proc. London Math. Soc. (3) 55 (1987), no. 1, 59-126. 15. N. J. Hitchin, Stable bundles and integrable systems. Duke Math. J. 54 (1987), no. 1, 91-114. 16. W. A. M. Janssen, Skew-symmetric vanishing lattices and their monodromy groups. Math. Ann. 266 (1983), no. 1, 115-133. 17. W. A. M. Janssen, Skew-symmetric vanishing lattices and their monodromy groups. II. Math. Ann. 272 (1985), no. 1, 17-22. 18. A. Kouvidakis, T. Pantev, The automorphism group of the moduli space of semistable vector bundles. Math. Ann. 302 (1995), no. 2, 225-268. 19. E. J. N. Looijenga, Isolated singular points on complete intersections. London Mathematical Society Lecture Note Series, 77. Cambridge University Press, Cambridge, 1984. xi+200 pp. 20. E. Markman, Generators of the cohomology ring of moduli spaces of sheaves on symplectic surfaces. J. Reine Angew. Math. 544 (2002), 61-82. 21. B. C. Ngo Le lemme fondamental pour les alg´ebres de Lie. Publ. Math. Inst. Hautes ´Etudes Sci. No. 111 (2010), 1-169. 32 DAVID BARAGLIA 22. N. Nitsure, Moduli space of semistable pairs on a curve. Proc. London Math. Soc. (3) 62 (1991), no. 2, 275-300. 23. R. W. Richardson, Conjugacy classes of n-tuples in Lie algebras and algebraic groups. Duke Math. J. 57 (1988), no. 1, 1-35. 24. L. P. Schaposnik, Spectral data for G-Higgs bundles. DPhil thesis. arXiv:1301.1981 (2013). 25. L. P. Schaposnik, Monodromy of the SL2 Hitchin fibration. Internat. J. Math. 24 (2013), no. 2, 1350013, 21 pp. 26. M. Thaddeus, Topology of the moduli space of stable vector bundles over a compact Riemann surface, Masters thesis, University of Oxford (1989). School of Mathematical Sciences, The University of Adelaide, Adelaide SA 5005, Australia E-mail address: [email protected]
1904.12917	1	1904	2019-04-29T19:28:31	Branch Stabilisation for the Components of Hurwitz Moduli Spaces of Galois Covers	[ "math.AG", "math.GT" ]	We consider components of Hurwitz moduli space of G-Galois covers and set up a powerful algebraic framework to study the set of corresponding equivalence classes of monodromy maps. Within that we study geometric stabilisation by various G-covers branched over the disc. Our results addresses the problem to decide equivalence and stable equivalence algebraically. We recover a homological invariant, which we show to distinguish the equivalence classes of given boundary monodromy and Nielsen type, if the latter is sufficiently large in the appropriate sense.	math.AG	math	BRANCH STABILISATION FOR THE COMPONENTS OF HURWITZ MODULI SPACES OF GALOIS COVERS MICHAEL L ONNE Abstract. We consider components of Hurwitz moduli space of G-Galois covers and set up a powerful algebraic framework to study the set of corresponding equivalence classes of monodromy maps. Within that we study geometric stabilisation by various G-covers branched over the disc. Our results addresses the problem to decide equivalence and stable equivalence algebraically. We recover a homological invariant, which we show to distinguish the equivalence classes of given boundary monodromy and Nielsen type, if the latter is sufficiently large in the appropriate sense. 1. Introduction In this article we want to discuss at length some group theoretical aspects crucial to the study of connected components of Hurwitz spaces of coverings, which can be traced back to [Cleb72, Hur91]. In the narrow sense, the objective is the study of morphisms p : C Ñ C{G ": C1 induced by an effective action of a finite group G. The existence of such an effective action provides information about the complex structure of the curve and also about the group. In the broader sense, we include branched covers over Riemann surfaces, usually with at least one boundary component. Let us recall first that the geometry of the covering p encodes several numerical invariants that are constant under deformation: the genus g1 of the base C1 and its number of boundary components, the number d of branch points y1, . . . , yd P C1 and the orders m1, . . . , md of the local monodromies - strictly speaking as an unordered multi-set. These invariants form the primary numerical type and have been the object of intensive studies since long. A refined invariant is obtained from the monodromy µ : π1pC1zty1, . . . , yduq Ñ G of the regular unramified cover given by restriction of p to the complement of the branch points, p´1pC1zty1, . . . , yduq. Instead of keeping track only of the orders mi of the elements in G associated to the local monodromies at the punctures yi, we consider the multi-set of their conjugacy classes. This is given by the Nielsen type, cf. [Niel37], the class function ν which on each conjugacy class C in G takes the cardinality of local monodromies in C as its value. The function ν can be characterised also without taking recourse to the quotient C1. For each conjugacy class C just count the number of G- orbits on C such that an element is in C which rotates a disc around some point of the orbit by the smallest possible angle. While many groups have been shown to allow classification by primary numerical type and Nielsen type, we are interested into arbitrary groups and get the motivation by the progress in the case of genus stabilisation. Date: May 1, 2019. The present work was done in the framework of the ERC Advanced grant n. 340258, 'TADMICAMT' . 1 2 MICHAEL L ONNE Let us briefly recall the main results in the case of free actions. There the second homology group H2pGq was shown ‚ to classify equivalence classes of unbranched G-coverings for abelian and metabelian groups [Edm82, Edm83]. ‚ to classify stable equivalence classes for every group [Liv85]. ‚ to classify equivalence classes for every group if the genus g1 is sufficiently large [DT06]. An analogous result to the last in the case of non-free action was proved with F. Catanese and F. Perroni [CLP16], where the second homology group had to be replaced by a quotient H2,Γ. The case of branch stabilisation is more involved. While genus stabilisation corre- sponds to connected sum with a trivial G-cover over the torus, summing a branched cover does involve making non-trivial choices for the monodromies at the branch points. In fact we will stabilise by boundary connected sum with a punctured disc, but consider very general choices for the branch monodromies. We will succeed following the above program in the following sense ‚ we classify various sets of equivalence classes of G-covers by elements of a monoid or a set with a monoid action. ‚ we classify certain sets of stable equivalence classes of G-covers by elements of a set, which can be distinguished by the primary numerical type, the Nielsen type and using H2,Γ. ‚ we classify equivalence classes of G-covers by their stable equivalence classes, if the Nielsen type is sufficiently large. For example we prove the algebraic version, Thm.4.9 of the following geometric result. Theorem 1.1. Suppose Γ is a union of conjugacy classes generating G, then there exists an integer N such that the number of equivalence classes of G-covers of the disc with local mono- dromies in Γ and fixed Nielsen type ν is independent of ν, if ν takes a value at least N on each conjugacy class. Let us give a short overview on the content of this article: In the next section we recall Hopf formula and explain its validity in the setting of crossed modules. We then take some care to motivate the definition of a G-crossed module taking a union Γ of conjugacy classes into account, which gives rise to a finite abelian group which is later shown to be equal to H2,Γ. In section 3 we equip the set of isomorphism classes of G-covers with an algebraic structure of monoid. This provides the necessary tool to investigate the geometric notion of stable equivalence by algebraic means. In the following section we address the stabilisation of G-covers branched over the disc, and determine conditions on the Nielsen types, such that stabilisation is surjective, resp. bijective. Section 5 considers various geometric generalisations of G-covers and their equivalence classes. The corresponding algebraic setting is then presented and explored. In the final section 6 we revise the definition of tautological central extension and recall the definition of quotient H2,Γ of H2pGq from [CLP15]. We will then see how it fits very well with the set-up of the previous sections and deduce a classification result for G-covers with sufficiently large Nielsen type. BRANCH STABILISATION 3 2. algebraic setting The formula of Hopf describes the second homology group H2pG, Zq as the kernel of a natural map associated to a given group G. To avoid unnecessary generality we restrict to the case of a finite group of order n :" ordpGq. In that case there is a natural finite presentation of G for any map S Ñ G, such that the image generates G as a group. It is expressed in a short exact sequence 1 Ñ R Ñ F Ñ G Ñ 1 , where F " FS is the free group freely generated by elements of S and R " RS Ă F is the free subgroup of relations. The Hopf formula [Hopf42] then states H2pG, Zq " R X rF, Fs rF, Rs . The same information is conveyed in the following exact sequences 1 Ñ H2pG, Zq Ñ F 1 Ñ H2pG, Zq Ñ R rF, Fs G Ñ 1 rF, Rs Ñ F rF, Fs Ñ rF, Rs Ñ F F RrF, Fs Ñ 1 However, there is another setting which also provides an approach to the second homology of G using a free object, the category of G-crossed modules. Let us quickly recall the basic definition. Definition 2.1 A group homomorphism B : C Ñ G with an action C G Ñ C ÞÑ cg pc, gq is called a G-crossed module, if XM1: B is G-equivariant for the conjugation action of G on itself, @ c P C, g P G. Bpcgq " pBcqg " g´1pBcqg, XM2: the Peiffer identities hold ca " acBa, Example 2.2 If S Ă G generates G then FS @ a, c P C. BS : rFS, RSs ÝÑ G ÞÑ {g´1ag G FS ÝÑ FS pg, aq is a G-crossed module which is a free G-crossed module on the inclusion S Ñ G, cf. [Rat80]. If S is invariant under conjugation the G-action is simply induced by which readily gives the G-equivariance of BS: BSpagq " BSp{g´1agq " g´1ag " g´1pBSaqg. We note that the denominator of the Hopf formula has been incorporated into the crossed module. So it remains to perform the intersection of its numerator, between a 4 MICHAEL L ONNE derived subgroup and a kernel. This can be done with lots of G-crossed modules with still the second homology of G dropping out. Theorem 2.3 ([ElPo86]). If B : C Ñ G is a projective G-crossed module ( in particular if it is a free G-crossed module) with B surjective, then H2pG, Zq " kerB X rC, Cs. We will now define a G-crossed module depending on a union Γ Ă G of conjugacy classes. To this end, we first relax the notion of G-crossed module, dispensing with the group structure of the domain. Definition 2.4 A map ε : Q Ñ G from a G-set Q to the group G is called an augmentation and Q an augmented quandle if AQ1: ε is G-equivariant for the conjugation of G on itself, (Q is a G-crossed set) εppgq " g´1εppqg, @ p P Q, g P G AQ2: idempotency holds in the sense pεppq " p, @ p P Q. G-crossed modules are augmented quandles, but also, more importantly, the set Γ. Example 2.5 The union of conjugacy classes Γ Ă G is a G set for the action by conjugation. Therefore the injection ε : Γ Ñ G is G-equivariant in the sense of AQ1. Property AQ2 also holds since any element in G is unchanged under conjugation with itself. We do not want to recall the story of the notion of quandle here, but refer for this and more on augmented quandles to the source [Joy82]. Instead we take the quickest path back to G-crossed modules. It leads via the following definition. Definition 2.6 Let Q be an augmented quandle then the adjoint group Adj Q is the group presented by Adj Q " xeq, q P Q epeq " eqepεpqqy. It has a unique group homomorphism to G compatible with the augmentation ε BQ : Adj Q Ñ G, eq ÞÑ εpqq since Adj Q has the universal property for quandle maps to groups. And it serves our purpose, thanks to Proposition 2.7. Suppose ε : Q Ñ G is an augmented quandle, then BQ : Adj Q Ñ G is a crossed module over G with respect to the tautological action of G on Adj Q induced by the action of G on Q Adj Q G Ñ Adj Q peq, gq ÞÑ eq g " eqg BRANCH STABILISATION Proof. First we check that BQ is G-equivariant: BQpeq gq " BQpeqgq " BQpεQpqgqq " εpqgq " g´1εpqqg " BQpεQpqqqg " BQpeqqg. Second we have to check the Peiffer identities: eBQpeqq p " eBQpεQpqqq p " eεpqq p " epεpqq " epq " e´1 q epeq 5 (cid:3) In analogy to the Hopf formula we get an abelian group HpΓ, Gq :" kerBΓ X rAdj Γ, Adj Γs. (1) As we will see, that intersection is a proper quotient of the second homology in general. 3. the Hurwitz monoid To associate to a G-cover of the punctured disc an algebraic object, we recall the Let D be the closed unit disc, ppnq Ă D a sequence of distinct points in the interior, following geometric set-up: p0 P BD a base point and let pγnq be a geometric basis for π1pDzppnq, p0q. Any G-cover of D -- connected or not -- which is unbranched outside p1, . . . , pd has a well-defined monodromy map relative p0. To the homotopy class of a closed path it assigns the unique bijection acting from the right on the fibre at p0 which gives the same map as path-lifting. Only after choosing a point p in the fibre, i.e. for a pointed G-cover, a monodromy map µ with values in G is well-defined in general1: Via the (left) G-action the fibre at p0 is identified with tg pg P Gu and µ takes the value g if path-lifting maps p to g p. From now we tacitly assume all G-covers and their isomorphisms to be pointed and only put an occasional (pointed) to remind the reader of this fact. In these circumstances, the chosen datum associates to a G-cover a monodromy tuple of elements in G. µpγ1q, . . . , µpγdq. As in previous papers, see [CLP15, CLP16], we want to call such a tuple a Hurwitz vector of genus 0 according to Definition 3.1 For any d ą 1 the set Gd is a Brd ´set by the well-known Hurwitz action of the braid group. An element of this Brd ´set is called a p0, dq-Hurwitz vector, while an element of Two Hurwitz vectors v, w are braid equivalent if they belong to the same braid group orbit, Gd is simply called a Hurwitz vector of genus 0. š v « w. Since on the geometric side we only consider isomorphism classes of G-covers, i.e. up to G-equivariant (pointed) maps covering a map preserving ppnq as a set, we get a one to one correspondence tG-covers of D, branched outside p1, . . . , pdu{iso 1:1"" Gd{ Brd . 1points in the same orbit under the center of G determine G-equivariantly isomorphic covers, in particular the choice is superfluous in case of abelian G. 6 MICHAEL L ONNE Now on the algebraic side we have a natural composition Gd Ge Ñ Gd`e given by juxtaposition which obviously is associative. It is equivariant under the inclusions Brd Ñ Brd`e and Bre Ñ Brd`e, where in the second case a braid on strands one to e is moved to the corresponding braid on strands d ` 1 to d ` e. The resulting monoid is called the Hurwitz class monoid of G: ž dě0 HG :" Gd{ Brd where the unit is understood to be represented by the empty tuple. This composition can also be constructed on the geometric side. Given a pair of G-covers unbranched outside p1, . . . , pd, respectively p1, . . . , pe, there is a unique class of G-covers unbranched outside p1, . . . , pd`e, which is isomorphic to the first over a regular neighbourhood of γ1 YY γd and to the second over a regular neighbourhood of γd`1 Y Y γd`e, pointed by the same point p over p0. On a moments thought, we may restrict on the geometric side to G-covers with local monodromies in a union of conjugacy classes Γ Ă G and get on the algebraic side a submonoid ž HG,Γ :" Γd{ Brd . dě0 š We next define equivalence of G-covers with respect to stabilisation by a G-cover Cu corresponding to some u P Ge. Definition 3.2 Two G-covers are called Cu´stably equivalent if composition of either with the same number (cid:96) of copies of Cu yield isomorphic G-covers. Gd are called u´stably equivalent if vu(cid:96) « wu(cid:96), The corresponding elements v, w P for some (cid:96). So u-stable equivalence of v, w is equality in the Hurwitz class monoid HG of elements represented by vu(cid:96), wu(cid:96) for some (cid:96). In fact we can push this further to an equality in some monoid of fractions, but we need to recall the evaluation map on HG first. Lemma 3.3. The evaluation map defined on Gd with values in G š v " pv1, . . . , vdq ÞÑ evpvq :" v1 vd P G has the following properties: i) ev is a monoid homomorphism. ii) ev is constant on braid group orbits. iii) ev induces an monoid homomorphism on the monoid HG, also called evaluation. Now evpuq P G has finite order -- say n -- so we deduce that the elements u(cid:96)n are The proof is easy and left to the reader. central in HG thanks to the following lemma proved in [CLP15]. Lemma 3.4. If v, w are Hurwitz vectors of genus 0 and evpvq " 1 P G, then vw and wv are braid equivalent. In particular these elements form a central submonoid of HG, which is the same as a central multiplicative set. The following technical result will provide the existence of corresponding monoids of fractions and expose their relation with the enveloping group of HG, defined by the universal property, that every monoid homomorphism to a group factors uniquely through the monoid homomorphism to the enveloping group. Proposition 3.5. Suppose S Ă HG is a submonoid, equal to HG or central and gener- ated by an element u(cid:96) with evpuq of order dividing (cid:96), then BRANCH STABILISATION 7 i) there is an equivalence relation on HG S given by ii) there is a well-defined monoid structure on equivalence classes induced by mul- ðñ pv, sq „ pv1, s1q Equivalence classes in case S generated by u(cid:96) are written rv{ssu. tiplication in HG, which for S central is u " This monoid is denoted by HGS´1. iii) the total monoid of fractions HGH´1 phic to the enveloping group of HG Dv P HG, s P S : vr " v1s, sr " s1s " w um(cid:96) ‰ vw upn`mq(cid:96) u . v un(cid:96) u " " ‰ ‰ G together with the map is uniquely isomor- iv) the monoid of fractions associated to u is isomorphic to the enveloping group if g P G ùñ Dv P HG, n ą 0 : gv " un(cid:96) P HG Proof. The first two claims follow from the following two properties, well-know from localisation of rings, see [Sko06]. ‚ s P S, v P HG ùñ sHG X vS ‰ H ‚ s P S, v, w P HG, sv " sw ùñ Ds1 P S : vs1 " ws1 Both are immediate for S central. The argument in the case S " HG is only slightly more difficult. In fact we note that every element has a power that is central, since its evaluation in the finite group G has finite order and by Lemma 3.4. s P S, v P HG ùñ svord evpvq " vord evpvqs P sHG X vS s P S, v, w P HG, sv " sw ùñ sord evpsqv " sord evpsqw ùñ vsord evpsq " wsord evpsq To get the last two claims we first notice that for the monoids of fractions HGS´1 by construction any monoid homomorphism to a group factors. Thus we only need to show that every element has an inverse. This is trivially true for the the total monoid of fraction. In the case of claim ivq the elements of HG given by a single letter g P G have an inverse rv{un(cid:96)s by hypothesis. Since these elements generate HG every element (cid:3) in the monoid has an inverse. We followed this route of progressive abstraction to have a criterion for stably equiv- alence in terms of an equality in a monoid of fraction which even is a group in the cases we are mostly interested in: Corollary 3.6. Suppose u P HG has evpuq P G of order (cid:96), then the following are equivalent i) v, w P HG are u´stably equivalent ii) rv{1su " rw{1su P HGS´1 Obviously, the proof of the proposition works also to proof analogous claims for submonoids of the monoids HG,Γ. However we skip the details and only spell out the corresponding corollary. Corollary 3.7. Suppose u P HG,Γ has evpuq P G of order (cid:96) and S " tun(cid:96)u, then the following are equivalent i) v, w P HG,Γ are u´stably equivalent ii) rv{1su " rw{1su P HG,ΓS´1 8 MICHAEL L ONNE In particular, if we want to decide u-stable equivalence in HG,Γ under the hypothesis that the monoid of fractions is a group, then we can do so in the finitely presented group Adj Γ. Proposition 3.8. Suppose u P HG,Γ has evpuq P Γ of order (cid:96) and S " tun(cid:96)u is a denominator set such that HG,ΓS´1 is a group, then v, w P Γd are u ´ stably equivalent ðñ ev1 evd " ew1 ewd P Adj Γ Proof. By the corollary above it suffices to show that the second claim of the proposi- tion is equivalent to the second claim of the corollary. This follows since both groups involved are the enveloping group of the monoid HG, and the element involved are mapped to each other by the unique map provided by the universal property. For the monoid of fractions this was shown above, for the adjoint group it is proved by Kamada, Matsumoto [KM05]. It uses the fact that the universal properties of the en- veloping group and the adjoint group provide mutually inverse group homomorphisms (cid:3) between the two. 4. the genus zero case We investigate the relation between equivalence and stable equivalence of (pointed) G-covers branched over the disc, i.e. with g " 0. They will be shown to be equal for the G-covers with sufficiently 'rich' branching and detectable in monoids of fractions from the last section. This involves studying conditions on the Hurwitz vector u such that stability holds for geometric stabilisation by the corresponding G-cover Cu according to Definition 4.1 We say that stability holds for geometric stabilisation by Cu, if there exists a positive integer m, such that Cu ´ stable equivalence " G ´ cover equivalence on the G-covers equivalent to covers obtained by m iterations of the Cu-stabilisation. The condition of the definition easily translates into a condition on the algebraic side. Stability holds if u ´ stable equivalence " braid equivalence on the set of tuples braid equivalent to some vum. We are thus bound to study properties of Hurwitz vectors of the last kind in more detail. Before entering into the discussion of stability, let us review some important tools, giving the definitions and providing their basis properties. š Gd of Hurwitz vectors for the finite group G Definition 4.2 On the set i) the relation v ď w : ðñ Du : vu " w is called the prefix order on Gd. ii) the relation v AE w : ðñ Du : vu « w š š is called the weak prefix order on iii) an element v P Gd is said to generate the subgroup H Ă G if Gd, and so is the induced order on HG. xvy :" xv1, . . . , vdy " H. BRANCH STABILISATION 9 Now let tCiu denote the set of equivalence classes for conjugacy on G, then i, @i " 'i ZCi with : ðñ νi ď ν1 ν ď ν1 Z G ` L „ is a free abelian group with a partial order. Note that more generally the free abelian group on a union Γ of conjugacy classes of G is naturally isomorphic to the abelianisation of the adjoint group for the conjugation quandle Γ Next we recall the basic properties of the Nielsen type and the subgroup generated by a Hurwitz vector, but leave the proofs to the reader, see also [CLP15, CLP16]. Gd with values in 'Ci ZCi " ZpG{„q, Lemma 4.3. The Nielsen map defined on v " pv1, . . . , vdq š 'CiĂΓZCi " H1pAdjpΓq, Zq. ÿ νiCi νi " #tj vj P Ciu ÞÑ νpvq :" has the following properties: i) ν is a monoid homomorphism. ii) ν is constant on braid group orbits. iii) ν induces a monoid homomorphism on HG also called Nielsen map. iv) the Nielsen map is order preserving for both prefix orders ùñ νpvq ď νpwq, ùñ v AE w v ď w š Lemma 4.4. The map x y : ordered by inclusion has the following properties: Gd Ñ tH H Ă Gu to the subgroups of G partially i) x y is order preserving for both prefix orders on ii) x y is constant on braid group orbits. iii) x y induces an order preserving map on HG with respect to the weak prefix order. A special role will be given to the Hurwitz vector defined in terms of Γ Ă G Gd. loooomoooon uΓ " pg1, . . . , g1 loooomoooon q, , . . . , gr, . . . , gr š ord g1 ord gr where g1, . . . , gr is an enumeration of the elements of Γ. Its braid equivalence class is independent of choices thanks to Lemma 3.4, since evpuΓq " 1G. The corresponding denominator set generated by uΓ will be denoted by SΓ. In the sequel, our argument is motivated by the approach of Conway and Parker, see [FV91]. The first step is the first part of [MM99, Lemma 6.9]. Lemma 4.5. Suppose g1, g2 are conjugate elements of order n in a finite group G and v is a Hurwitz vector which generates G, then 1 « vgn vgn 2 . The second part of that lemma inspired our next result: Lemma 4.6. Let w be a Hurwitz vectors generating G and u a Hurwitz vector with entries in Γ, then νpwq ě νpuΓuq ùñ Dv : w « vu, v generates G 10 MICHAEL L ONNE Proof. Thanks to induction, it suffices to prove the claim in case u consists of a single entry g. By hypothesis the conjugacy class of g occurs so often that by the pigeonhole principle there is a g1 conjugate to g which occurs at least n " ord g ` 1 times. By Hurwitz moves we can obtain a Hurwitz vector w1gn`1 braid equivalent to w, hence w1g1 generates G. We apply Lemma 4.5 and obtain w « w1g1gn. Thus our claim follows for u " g and v " w1g1gn´1. (cid:3) 1 Now we are ready to prove a bunch of bijectivity results for certain map induced by * stabilisation maps. Lemma 4.7. Let u be a Hurwitz vector and suppose ν0 ě νpuΓpuqq, then for all n ě 1 * " " rws« νpwq " pn ` 1qνpuq ` ν0 νpvq " nνpuq ` ν0 xuy Ă xvy rvs« u : ÝÑ ÞÑ xuy Ă xwy rvus« rvs« is surjective and there exists m " mpu, ν0q, such that it is bijective for n ě m. Indeed, for w with rws Proof. To prove surjectivity we want to apply Lemma 4.6. in the range, we let G " xwy and Γ " Γpuq X G. Then u has entries in Γ and νpwq ě νpuq ` ν0 ě νpuΓuq, since uΓ AE uΓpnq. Hence we get the conclusion, that rws " rw1us with xw1y " G Ą xuy. Bijectivity follows from the fact, that an infinite chain of surjective maps between (cid:3) Γy " Gq, Γq " νpuΓq ñ xu1 finite sets must eventually stabilise. Proposition 4.8. Suppose uΓ generates G invariably pi.e. νpu1 then there exists an integer m " mpGq such that for Γdq š ‚ any ν1 ě ν0 ě mνpuΓq with ν0, ν1 P νp ‚ any Hurwitz vector u with νpuq ` ν0 " ν1 * juxtaposition of u induces a bijective map * " rvs« νpvq " ν0 rws« " rvs« νpvq " ν1 rwus« ÝÑ ÞÑ Proof. Apply Lemma 4.7 and get m " mpuΓ, νpuΓqq such that uΓ is bijective for n ě m. There are now Hurwitz vectors u0, u1 such that u : νpu0q ` mνpuΓq " ν0, u0uu1 « uk Γ for some k ą 0. (cid:32) Then we get a factorisation of the bijective map uk Γ: (cid:32) (cid:32) νpvq " mνΓ νpvq " ν0 νpvq " ν1 νpvq " kνΓ (cid:32) ( u0ÝÑ ( u1ÝÑ ( uÝÑ ( Since uΓ generates G invariantly, every element in the given sets generates G, hence Lemma 4.7 gives surjectivity of all three maps. Thus we conclude that each of the surjective maps, in particular u is bijective. (cid:3) Theorem 4.9. Suppose uΓ generates G, then there exists an integer m " mpGq such that š i) for‚ any ν1 ě ν0 ě mνpuΓq with ν0, ν1 P νp Γdq " NpΓ{„q ‚ any subgroup H of G, ‚ any Hurwitz vector u P H d with νpuq ` ν0 " ν1 š BRANCH STABILISATION " rvs« u : juxtaposition of u induces a bijective map ÝÑ š ÞÑ ii) for any ν1, ν0 ě mνpuΓq with ν0, ν1 P νp xvy " H rws " * νpvq " ν0 * νpvq " ν0 rvs« # " # " rvs« " Γdq rvs« νpvq " ν1 * νpvq " ν1 xvy " H rwus 11 * Γq " νpuΓq. Then we can handle the claim of iq. š In case ν0 R νp Γq " νpuΓq a multiple of uΓXH. Claim iq then follows as in the previous proofs. Proof. Let m be a common stability bound for the finitely many Hurwitz vectors u1 with νpu1 H dq, also ν1 is not contained, so both sides are empty sets and the claim is trivially true. Otherwise pick u0 P H d with νpu1 H d with νpu0q " ν0 and u1 š š Γ P The second claim is an immediate corollary, since the sets involved decompose over all possible subgroups H of G into the sets of the first claim and shown there to be (cid:3) bijective. Γ Definition 4.10 Call m as in the theorem a Γ-stability bound. Proposition 4.11. Suppose v, w are Hurwitz vectors generating G with entries in Γ, and νpvq ě νpum Γ q, where m is a Γ-stability bound, then v « w ðñ rv{1s " rw{1s P HG,ΓS´1 Γ , Proof. We have only the prove the reverse implication. The first hypotheses on the right implies v «u w Γ « wu(cid:96) v « w ùñ vu(cid:96) ùñ uΓ-stably Γ for some (cid:96) by theorem 4.9 (cid:3) In Corollary 3.6 we showed that elements with the same invariant are stably equiva- lent. In this proposition we show that elements with sufficiently high Nielsen class belong to the stable range, ie. where stable equivalence implies equivalence. The existence of a stability range expresses the fact that stability holds. 5. generalisations In this section we will aim for some generalisations going beyond the case of G-covers of the disc branched at a finite set of points up to G-equivariant covering isomorphisms. Though we still want to understand stabilisation under 'geometric' composition with a G-cover Cu over the disc corresponding to some u P Ge, there are several different direction open to generalisation: ‚ include G-covers over surfaces of higher genus or higher number of boundary components, one boundary component is needed at least to perform geometric composition. ‚ modify the notion of isomorphism on the base: restrict the induced maps to preserve isotopy classes of appropriate geometric objects. 12 MICHAEL L ONNE ‚ modify the notion of isomorphism on the fibres: restrict to maps preserving G-markings of a set of fibres. To get the flavour of these generalisations, let us look at some examples and the corresponding algebraic structure. Example 5.1 On the geometric side consider a Riemann surface Σ " Σ1 g of genus g with one boundary component. Let p0 be a point on the boundary, ppnq Ă Σ be a sequence of distinct points in the interior. Finally let αi, βi corresponding to handles, and a sequence pγnq corresponding to the interior points be elements of a geometric basis for π1pΣzppnq, p0q As before any (pointed) G-cover of Σ, which is unbranched outside p1, . . . , pd has a monodromy map that gives rise to the monodromy tuple of elements in G µpα1q, µpβ1q, . . . , µpαgq, µpβgq, µpγ1q, . . . , µpγdq, g,d :" MappΣ1 In the present case, such tuples are naturally acted on by the mapping class group gq, and we recall that elements of G2g`d are called pg, dq-Hurwitz Map1 vectors, since the Cartesian product is considered as a Map1 g,d-set. We get the one to one correspondence tG-covers of Σ, branched outside p1, . . . , pdu{iso 1:1"" G2g`d{Map1 g,d base with geometric basis Example 5.2 Let us consider the G-covers of Σ as in the previous example. However we define two G-covers to be isomorphic preserving chains if they are isomorphic via some G-equivariant covering map such that the induced map preserves the free isotopy classes of the αi, βi. Accordingly on the algebraic side we no longer have the action of the full mapping class group but rather that of the subgroup preserving the said isotopy classes. This subgroup can be identified with the mapping class group Mapg`1 0,d of a subsurface Σ1 of genus 0 with d punctures and g ` 1 boundary components obtained from Σ by cutting a regular neighbourhood of simple curves representing the given isotopy classes, thus tG-covers of Σ, branched outside p1, . . . , pdu{isoα,β 1:1"" G2g`d{Mapg`1 0,d BRANCH STABILISATION 13 base with curves representing two chains Example 5.3 For the third kind of generalisation we look again at G-covers of the disc D equipped with distinct interior points ppnq, a geometric basis pγnq with respect to p0 on the boundary. Let q0, . . . , qk be distinct points on the boundary with q0 " p0. We introduce a q-marking to be given by G-equivariant maps (cid:96)i from G to the fibres over the qi. Then we get an extended monodromy map defined on the fundamental groupoid of the punctured base relative to the finite set of points qi. µ : πgr´oid 1 pDztp1, . . . , pdu,tq0, . . . , qkuq ÝÑ G To the homotopy class of a path from qi to qi1 it associates the unique element g in G such that g(cid:96)i1p1q gives the same point in the fibre at qi1 as path lifting to (cid:96)ip1q. The domain is a free groupoid on d ` k generators, and we end up with the one to one correspondence tmarked G-covers of D, branched outside p1, . . . , pdu{isomarked 1:1"" Gk`d{ Brd The set Gk`d in this case is a Brd-set as the Cartesian product of the trivial Brd-set Gk and the Brd-set Gd of Hurwitz vectors of genus 0. base with geometric base of groupoid The common features of these generalisations -- and others -- on the algebraic side should be well noted: ‚ we classify orbits of G-tuples, ‚ the action is by some geometrically distinguished mapping class group, ‚ Brd acts as the mapping class group of a suitable neighbourhood of γ1, . . . , γd. 14 MICHAEL L ONNE They will allow to get an algebraic model for composition with G-covers of genus 0 on the geometric side, which relies on the notion in the following definition: Definition 5.4 Let M be a monoid and S be a set together with a map such that ρ : S M Ñ S, spm1m2q " psm1qm2, ps, mq ÞÑ sm s1M " s Then we will call ρ a M -action and S a M -set (instead of the more common M -act). Let us formulate the hypothesis in more abstract terms Proposition 5.5. Suppose there is an array of nested groups Md Y Brd M0 Ă M1 Ă M2 Y Y Br1 Ă Br2 š d Gr`d is a š with Md acting on Gr`d such that as a Brd-set Gr`d is the Cartesian product of a trivial factor Gr and the Brd-set Gd of Hurwitz vectors of genus 0, then i) ii) there is an induced action of the Hurwitz class monoid HG on the M-orbits d Gd-set for concatenation, Gr`d{Md š Gd-set and the action is by concatenation. d Gr`d is an invariant subset. For the second claim we have to show that the action on equivalence classes does not Proof. Of course, the monoid Thus the first claim follows from the observation that š QG,M :" š Gd is a ž d depend on the representatives. In fact xMe " yMe, v Brd " w Brd ùñ xvMe`d " ywMe`d since the action of Brd on Gd is the same as that on the second factor of Gr`e`d " Gr`e Gd via the action of Bre`d Ă Me`d provided Brd is considered as the subgroup (cid:3) of Bre`d braiding only the last d strands. There is still one more essential feature: ‚ the number of local monodromies in each conjugacy class is invariant under the equivalence. In fact a choice of a geometric free basis can be made in such a way, that the local monodromies correspond to the entries of a tuple in Gr`d except for the first s entries, with s ď r. Hence in extension of the Nielsen type we get Definition 5.6 The Nielsen map ν is defined on Gr`d with values in 'Ci ZCi " ZpG{„q, ÿ v " pv1, . . . , vr`dq ÞÑ νpvq :" νiCi νi " #tj ą s vj P Ciu At this point we have collected enough information to turn back to the previous sections and see that their results generalise all along the way we have gone. Remark 5.7 The action of HG on QG,M can be 'localised' at a denominator set S to yield a group action of the group HGS´1 on the set QG,MS´1. BRANCH STABILISATION 15 Moreover if u P HG has evpuq P G of order (cid:96) and S " tun(cid:96)u, then the following are equivalent i) v, w P QG,M are u´stably equivalent ii) rv{1su " rw{1su P QG,MS´1 We will say more about stabilisation. However, instead of pushing the generalisation to the limits, we fix Γ " ΓG " Gzt1Gu which contains all non-trivial elements of G and * generates G invariably. Lemma 5.8. Let u be a Hurwitz vector and suppose ν0 ě pr ´ s ` 1qνΓ, then " " * νpwq " pn ` 1qνpuq ` ν0 νpvq " nνpuq ` ν0 vM vM wM uÝÑ ÞÑ vuM is surjective for n ą 0 and there exists m " mpu, ν0q, such that it is bijective for n ě m. Proof. The larger bound for ν0 is needed to make sure that for w P Gr`d its tail τw P Gd has νpτwq ě νpuq ` νΓ. Then we proceed as in the proof of Lemma 4.7 using Lemma 4.6 and that every Hurwitz vector with ν ě νΓ generates G by our general assumption Γ " ΓG. (cid:3) Theorem 5.9. Let Γ " ΓG, then there exists an integer m " mpGq such that for juxtaposition of u induces a bijective map * " rvsM ‚ any ν1 ě ν0 ě mνpuΓq " ‚ any Hurwitz vector u with νpuq ` ν0 " ν1 rvsM * νpvq " ν0 " uÝÑ ÞÑ vM νpvq " nνΓ vM * νpvq " ν1 vuM Proof. Apply Lemma 5.8 with u " uΓ and ν0 " pr ` 1qνΓ to get m and let m " mpuΓ, νpuΓqq ` r ´ s ` 1 such that uΓ is bijective for n ě m on There are now Hurwitz vectors u0, u1 such that u0uu1 « uk Then we can conclude as in the proof of Prop. 4.8. νpu0q ` mνpuΓq " ν0, Γ for some k ą 0. (cid:3) This is the important task, since the analogous result for genus stabilisation will be applied in the argument for homological stability. 6. The tautological lift In the final section we revise the definition of tautological central extension and recall some information, see [CLP15], in particular the definition and properties of a quotient H2,Γ of H2pGq which proved to be crucial in our classification of curves with dihedral group of automorphisms. We will then see how to recover that group in the set-up of the previous sections and deduce a classification result for (pointed) G-covers in the stable range. Definition 6.1 Let G be a finite group and let F " FG, R " RG be as before. For any union of conjugacy classes Γ Ă G, define 16 MICHAEL L ONNE i) the tautological lift G Ñ FG, which maps g ÞÑ g, ii) the tautological map on iii) the normal subgroup RΓ normally generated by commutators in rF, Rs and tau- Gd : v " pv1, . . . , vdq ÞÑ v1 vd tological lifts of conjugacy relations for elements in Γ: š @@ RΓ " rF, Rs, abc´1b´1 ( Note that the given elements generate this normal subgroup as subgroup. ) iv) the quotient group of F by RΓ DD ´1ab @a P Γ, b P G, c " b L GΓ " F RΓ . L RΓ. In particular v) the boundary homomorphism α : GΓ Ñ G, induced by a ÞÑ a with kernel KΓ. Lemma 6.2. With the notation just introduced, RΓ Ă R and KΓ " R KΓ is contained in the centre of GΓ and the short exact sequence 1 Ñ R RΓ Ñ GΓ Ñ G Ñ 1 RΓ is a central extension. Proof. rF, Rs Ă R because R is normal in F . Moreover abc´1b´1 P R for any a, b, c P G with ab " bc, therefore RΓ Ă R. By the definition of α we have that KΓ " R . Finally, KΓ is in the centre of GΓ because rF, Rs Ă RΓ. (cid:3) The tautological lift G Ñ GΓ, a ÞÑ a is not a group homomorphism in general, but every element in GΓ with image g P G can be written as gz " zg, with z P KΓ. Here, by abuse of notation, g denotes also the class of g P F in GΓ " F Remark 6.3 Let Γ Ă G be the union of distinct conjugacy classes C1, . . . , Ct and let g1, . . . , gr be the elements of GzΓ, then the abelianisation Gab Γ of GΓ is the free abelian group on t ` r generators Gab Γ š -- ZC1 ' ' ZCt ' Zg1 ' ' Zgr. š Γd factors through the tautological map and the abeliani- dΓd ÝÑ GΓ ÝÑ Gab The Nielsen map ν on sation as L RΓ. Γ ÝÑ 'iZCi ` H2,ΓpGq " ker GΓ Ñ G Gab Γ , Definition 6.4 Let Γ Ă G be a union of non-trivial conjugacy classes of G. We define where GΓ Ñ G Gab second component the abelianisation. Γ is the morphism with first component the boundary map α and Let us recall from [CLP15] the precise relation between H2pG, Zq and H2,ΓpGq. Lemma 6.5. Let G be a finite group and let Γ Ă G be a union of nontrivial conjugacy classes. Write G " F . Then, there is a short exact sequence R and GΓ " F 1 Ñ RΓ X rF, Fs RΓ rF, Rs Ñ H2pG, Zq Ñ H2,ΓpGq Ñ 1 . In particular H2,ΓpGq is abelian. BRANCH STABILISATION Remark 6.6 The Schur multiplier is often interpreted as a cohomology group L L which is algebraically dual to H2pG, Zq. In case Γ " G Moravec [Mo12] identified the Zq introduced by Bogomolov and justly group H2,GpGq with a subgroup of H 2pG, Q called Bogomolov multiplier by Moravec. H 2pG, Q Zq 17 The algebraic object from the previous to enter the stage is the adjoint group asso- ciated a quandle from Def.2. To emphasise its importance for the pair G, Γ we give a more specific name: Definition 6.7 Suppose G is a finite group and Γ Ă G a union of conjugacy classes, then the adjoint group for the quandle Γ Adj Γ :" xeg, g P Γ eaeb " ebeab, a, b P Γy is called the tautological crossed module of the pair G, Γ. The structural maps are BΓ : eg ÞÑ g, Adj Γ G Ñ Adj Γ : pea, gq ÞÑ eag . (It is a crossed module thanks to Prop.2.7.) The two following results establish the close relation between the 'old' tautological central extension and the 'new' tautological crossed module. Proposition 6.8. The map ea ÞÑ a for all a P Γ extends to an injective group homo- morphism Adj Γ ÝÑ F{RΓ with left inverse. Proof. For the extension we only need to check that the relations of the domain map to RΓ: eaebe´1 ab eb´1 ÞÑ abc´1b´1, with c " ab " b´1ab. For the left inverse we first define a map F Ñ Adj Γ by h ÞÑ eg1 egr where for each h P G a unique factorisation h " g1 gr, r " rphq has been chosen, with r " 1, g1 " h if h P Γ. Since the composition of the two maps is the identity on Γ, it induces the identity map on Adj Γ if RΓ is in the kernel of this second map. Let us note first, that by construction the map F Ñ G factors through BΓ. Accord- ingly R maps to kerBΓ. Now recall that for any crossed module B : C Ñ G the kernel is central in C. In fact, conjugation by any element in the kernel is trivial by the Peiffer identity. We infer, that R maps to the centre of Adj Γ and hence rF, Rs maps to the identity. We complete the proof by showing that this is true also for the remaining elements generating RΓ: h{h´1gh gh ÞÑ egeg1 egr " eg1egg1 eg2 egr " eg1 egregh ÞÑ eg1 egregh Proposition 6.9. If Γ generates G, then H2,Γ " HpΓ, Gq :" kerBΓ X rAdj Γ, Adj Γs (cid:3) 18 MICHAEL L ONNE Proof. Since the maps from the previous proposition induce the identity on the quotient G and maps on the respective abelianisations, they also induce maps on the given groups. Therefore the group of the right is a subgroup of the other and it remains to show, that the kernel N of GΓ Ñ Adj Γ intersects H2,Γ trivially. By the construction of the map, N is normally generated by elements g1 . . . gr h´1, for all h R Γ They all map to 1 P G, so they are central and generate a free abelian subgroup in GΓ of rank GzΓ. The image in ZpGzΓq via Gab Γ is of the same rank, since each generator is mapped to a standard generator. Thus N maps injectively to Gab Γ and thus intersects (cid:3) trivially with H2,Γ. To pursue the proof of the following classification result, we first need to introduce another Nielsen map. Definition 6.10 The canonical extension of the Nielsen map HG Ñ 'iZCi to the localisation Γ ÝÑ 'iZCi ν : HGS´1 Γs rv{u(cid:96) š is also called Nielsen map. νpvq " ν0 Theorem 6.11. Suppose uΓ generates G, then there exists an integer m " mpGq such that for any ν0 ě mνpuΓq with ν0 P νp ÞÑ νpvq ´ (cid:96)νpuΓq * " #H2,Γ rG, Gs " rvs« Γdq # xvy " G Proof. First we note that there is an exact sequence induced from the central extension in Lemma 6.2 Moreover, any rv{u(cid:96) uΓ,v : * " Γs P ker ν gives two maps ÝÑ νpvq " ν0 ` (cid:96)νΓ H2,Γ ÝÑ ker ν ÝÑ rG, Gs νpvq " ν0 " rvs« rvs« * xvy " G xvy " G Both are bijective since we are above the stability bound, hence we can define a well- defined action rws«rv{u(cid:96) Γs " rw1s, such that wv « w1u(cid:96) Γ. Again using the fact, that we are in the stable range, the set is mapped injectively to the enveloping group Adj Γ. This can be exploited to show that the action is free and transitive, because the action is now identified with multiplication inside Adj Γ by the subgroup ker ν. While freeness is immediate, we are left to check, that any two elements w, w1 are in one orbit: There is a unique element in the group Adj Γ which maps one to the other, Γs. Under the Nielsen map, it must map to 0, since which again we can write as rv{u(cid:96) ν is a homomorphism on Adj Γ. Thus this element is in ker ν and we have also proved (cid:3) transitivity. The claim on the cardinality of the set is then obvious. Rephrased in more geometrical terms the statement of the theorem tells us: In the stable range connected (pointed) G-covers are classified up to equivalence by the Nielsen type, the evaluation, and an element in H2,Γ. BRANCH STABILISATION 19 Do not miss the caveat: the homological information is not canonical, but depends on the choice of an element in each fibre of Adj Γ Ñ G 'iZCi. [CLP11] [CLP15] [CLP16] [Cleb72] [DT06] [Edm82] [Edm83] [ElPo86] [FV91] [Hopf42] [Hur91] [Joy82] [KM05] [Liv85] [MM99] [Mo12] [Niel37] [Rat80] [Sko06] References Catanese, F., Lonne, M., Perroni, F. Irreducibility of the space of dihedral covers of algebraic curves of fixed numerical type. Atti Accad. Naz. Lincei Cl. Sci. Fis. Mat. Natur. Rend. Lincei (9) Mat. Appl. 22 (2011), 1 -- 19. Catanese, F., Lonne, M., Perroni, F. The irreducible components of the moduli space of dihedral covers of algebraic curves, Groups Geom. Dyn. 9 (2015), 1185 -- 1229. Catanese, F., Lonne, M., Perroni, F. Genus stabilization for the components of moduli spaces of curves with symmetries, Algebr. Geom. 3 (2016), 23 -- 49. Clebsch A., Zur Theorie der Riemann'schen Flachen. Math. Ann. 6, (1872), 216-230. Dunfield, N.M., Thurston, W.P. Finite covers of random 3-manifolds. Invent. Math. 166, (2006), 457 -- 521. Edmonds, A.L., Surface symmetry I, Michigan Math. J., 29 (1982), 171 -- 183. Edmonds, A.L., Surface symmetry II, Michigan Math. J., 30 (1983), 143 -- 154. Ellis, G.J., Porter, T. Free and Projective Crossed Modules and the Second Homology Group of a Group. J. of Pure and Appl. Alg. 40 (1986), 27 -- 31. Fried, M.D., Volklein, H. The inverse Galois problem and rational points on moduli spaces. Math. Ann. 290 (1991), 771 -- 800. Hopf, H., Fundamentalgruppe und zweite Bettische Gruppe, Comment. Math. Helv. 14 (1942). 257 -- 309. Hurwitz, A.: Ueber Riemann'schen Flachen mit gegebenen Verzweigungspunkten. Math. Ann. 39, (1891), 1 -- 61. Joyce, D. A classifying invariant of knots, the knot quandle, J. Pure Appl. Algebra 23 (1982), 37 -- 65. Kamada, S., Matsumoto, Y. Enveloping monoidal quandles, Top. and Appl. 146/147 (2005), 133 -- 148. Livingston, C. Stabilizing surface symmetries. Michigan Math. J. 32 (1985), 249 -- 255. Malle, G.,Matzat B.H. Inverse Galois Theory, Monographs in Mathematics, Springer Verlag Berlin, 1999. P. Moravec, Unramified Brauer groups of finite and infinite groups, Amer. J. Math. 134 (2012), 1679 -- 1704. Nielsen, J. Die Struktur periodischer Transformationen von Flachen. Danske Vid. Selsk. Math.-Fys. Medd. 15, (1937), 1 -- 77. Ratcliffe, J. Free and projective crossed modules. J. LMS (2) 22, (1980), no.1, 66 -- 74. Skoda, Z. Noncommutative localization in noncommutative geometry. Noncommutative Localization in Algebra and Topology, LMS Lecture Note Series, CUP, Cambridge, (2006), 220 -- 310.
1511.08288	1	1511	2015-11-26T04:44:49	Rectification of Deligne's mixed Hodge structures	[ "math.AG" ]	We promote Beilinson's triangulated equivalence between the bounded derived category of rational polarizable mixed Hodge structures and the derived category of rational polarizable mixed Hodge complexes to an equivalence of symmetric monoidal quasi-categories. We use this equivalence to construct a presheaf of commutative differential graded algebras in the ind-completion of the category of rational mixed Hodge structures which computes Deligne's mixed Hodge structure on the rational Betti cohomology of $\mathbf{C}$-schemes of finite type. This leads to a preshea--in the quasi-categorical sense--of $\mathbf{E}_{\infty}$-algebras computing integral mixed Hodge structures.	math.AG	math	Rectification of Deligne’s mixed Hodge structures Brad Drew Abstract We promote Beilinson’s triangulated equivalence between the bounded derived cat- egory of rational polarizable mixed Hodge structures and the derived category of rational polarizable mixed Hodge complexes to an equivalence of symmetric monoidal quasi-categories. We use this equivalence to construct a presheaf of commutative differ- ential graded algebras in the ind-completion of the category of rational mixed Hodge structures which computes Deligne’s mixed Hodge structure on the rational Betti coho- mology of C-schemes of finite type. This leads to a presheaf—in the quasi-categorical sense—of E∞-algebras computing integral mixed Hodge structures. Contents 1 Deriving Tannakian categories 2 Mixed Hodge coefficients 3 Rectification 4 Integral coefficients Introduction 7 13 17 23 /C denote the category of C-schemes of finite type and let X ∈ Schft Let Schft /C. The de- rived category D(X(C)an,Q) of analytic sheaves of Q-modules, M. Saito’s derived category Db MHM(X) of mixed Hodge modules ([Sai88, Sai90]) and the P1-stable A1-homotopy cat- egory SH(X) of F. Morel and V. Voevodsky, as developed by J. Ayoub ([Ayo07a, Ayo07b]), admit Grothendieck six-functor formalisms. They are related by symmetric monoidal Betti,X : SH(X) → ∗ triangulated functors ω D(X(C)an,Q) ([Ayo10]) compatible with Grothendieck’s six functors. One expects that ⊆ SH(X) spanned by the ℵ0- ∗ Betti,X to the full subcategory SH(X)ℵ0 the restriction of presentable objects actually factors as the composite of a symmetric monoidal triangulated X : Db MHM(X) → D(X(C)an,Q) ([Sai90]) and ∗ 1 Hodge realization functor ∗ Hdg,X is itself compatible with Grothendieck’s six functors. ∗ Hdg,X : SH(X)ℵ0 → Db MHM(X) and ω ∗ X, and that this functor Q). Higher algebra, as developed in [Lur14], offers an elegant approach to this open prob- ∗ lem, consisting of two main ingredients: (i) a sufficiently refined construction of Hdg,X for the base case X = Spec(C), and (ii) some general results concerning Grothendieck’s six- functor formalism in the context of stable symmetric monoidal quasi-categories. Our goal here is to address (i) as follows, deferring discussion of (ii) and the question of constructing Hodge realization functors over more general bases X to a forthcoming preprint. For Λ ∈ {Z,Q}, let MHSp Λ denote the category of polarizable mixed Hodge Λ-structures and gr MHSp Λ the category of Z-graded objects thereof. We rectify P. Deligne’s functor Betti(−,Z) : (Schft • Z, assigning to X the graded polarizable mixed Hodge H Z-structure on its Betti cohomology ([Del74]), to a presheaf of with values in a quasi- category of E∞-algebras in the derived quasi-category of MHSp Λ. With rational coefficients, we further rectify this to a strict presheaf of commutative differential graded algebras in the ind-completion Ind(MHSp /C)op → gr MHSp Z) (cid:39) Db Q) (cid:39) Db Hp,Q can be promoted to an equivalence Hp,Z from the bounded derived category of MHSp The issue of rectification of mixed Hodge structures considered here is of interest independent from the aforementioned motivic questions. Indeed, the question has a long history: cf. [NA87, 8.15], [GNA02, 2.3.6], [CG14, 4.4]. The fundamental component in such rectification results is always A. Beilinson’s equivalence of triangulated categories Db(MHSp Z to the derived category of polarizable mixed Hodge Z-complexes ([Bei86, 3.11]). Below, we establish the following refinement of this equivalence with rational coefficients. Theorem 2.7. Beilinson’s equivalence Db(MHSp of symmetric monoidal quasi-categories. Combined with technical results about model structures on complexes in ind-completions of Q-linear Tannakian categories established in §1, this allows us to deduce the following rather strong rectification result. Theorem 3.6. There is functor ΓHdg(−,Q) from (Schft /C)op the category of commutative differ- ential graded algebras in Ind(MHSp Q) such that, for each X, the cohomology of ΓHdg(X,Q) is • Betti(X,Q). naturally isomorphic to Deligne’s mixed Hodge structure on H Gluing the functor ΓHdg(−,Q) of 3.6 with the “singular cochain complex” functor, we obtain the following variant with integral coefficients. Corollary 4.9. There is a functor ΓHdg(−,Z) from (Schft /C)op to the quasi-category of E∞- algebras in the symmetric monoidal derived quasi-category D(Ind(MHSp Z)) such that, for each X, the cohomology of ΓHdg(X,Z) is naturally isomorphic to Deligne’s mixed Hodge structure on • Betti(X,Z). H As mentioned above, our intended application of these results is the construction of ∗ Hdg,X. It is straightforward to check, using 4.6 and M. Robalo’s Hodge realization functors 2 universal property of the P1-stable A1-homotopy quasi-category ([Rob15, Corollary 1.2]), that these results do indeed lead to the desired Hodge realization functor over X = Spec(C). This construction actually provides a symmetric monoidal functor between symmetric monoidal quasi-categories, rather than a mere symmetric monoidal triangulated functor. In this sense, we obtain a refined version of previous constructions due to A. Huber ([Hub95, Hub00, Hub04]), M. Levine ([Lev98, 2.3.10]), and F. Lecomte and N. Wach ([LW13]). Aside from playing a key role in the larger project of constructing ∗ Hdg,X for more general X, let us mention another application of our results. Using 3.6 (resp. 4.9), one can construct a motivic E∞-ring spectrum EHdg in SH(C) representing rational (resp. integral) absolute Hodge cohomology ([Bei86, §5]) and a morphism of E∞-spectra HZ → EHdg from the motivic Eilenberg-Mac Lane spectrum to EHdg inducing regulator (resp. cycle-class) morphisms from the rational higher K-theory (resp. higher Chow groups) of the smooth C-scheme of finite type X to its rational (resp. integral) absolute Hodge cohomology. This motivic E∞-spectrum EHdg will be crucial to our approach to the construction ∗ of Hdg,X for more general X: higher algebra allows us to make sense of a well-behaved symmetric monoidal quasi-category of modules over EHdg in SH(X) and we show in a forthcoming work that this quasi-category of modules is naturally a full sub-quasi-category of the derived quasi-category D(Ind(MHSp Z)). Our strategy for bases X of higher dimension is to generalize this result. Relation to other work In some form or other, the essential results of the first three sections below are contained in the author’s 2013 PhD thesis. Interesting related work has appeared since then. In [Pri, Appendix A.2], J.P. Pridham discusses an construction related to 3.6, applying different techniques and treating only the case of smooth C-schemes. Working with real coefficients, in [BNT15], U. Bunke, T. Nikolaus and G. Tamme have used similar techniques for lifting regulator morphisms to morphisms of motivic E∞-ring spectra, further analyzing the structure of the motivic E∞-ring spectrum representing real absolute Hodge cohomology and its relation to differential algebraic K-theory. An alternative construction of a motivic commutative ring spectrum representing absolute Hodge cohomology with real coefficients will be presented in the PhD thesis of A. Navarro Garamendia ([NG16]). Using our rectification result, W. Soergel and M. Wendt have studied a motivic E∞-ring spectrum, denoted by EGrH in [SW15], which represents an interesting variant of absolute Hodge cohomology. Organization We begin in §1 by establishing some technical results on model structures and Tannakian categories, specifically showing that the injective model structure on the category of 3 complexes of ind-objects in a Q-linear Tannakian category is symmetric monoidal (1.7); that the category of commutative algebras in this symmetric monoidal category admits a model structure (1.8); and that this model structure on commutative algebras allows for a useful rectification result (1.10). In §2, working with rational coefficients, we construct a symmetric monoidal quasi- category of mixed Hodge complexes and use it to promote Beilinson’s triangulated equiv- alence Db(MHSp Hp,Q to an equivalence of symmetric monoidal quasi-categories (2.7). Q) (cid:39) Db In §3, we combine the results from the previous sections to construct the functor of 3.6 in two steps. First, we restrict construct the functor after restricting the domain to separated, smooth C-schemes (3.2). Then, using a result of V. Voevodsky, we extend the functor to all C-schemes of finite type (3.6). In §4, we establish some general results on fiber products of stable quasi-categories and t-structures (4.2, 4.3) and use them to show that the derived quasi-category of mixed Hodge Z-structures is a fiber product of the derived quasi-categories of mixed Hodge Q-structures and Abelian groups over the derived quasi-category of Q-modules (4.5). We then establish the required functoriality of singular cochain complexes of associated analytic spaces (4.8) and deduce 4.9. Notation and conventions Grothendieck universes: We assume that each set is an element of a Grothendieck universe. Fix uncountable Grothendieck universes U ∈ V such that the categories Set, Ab and Cat of U-sets, U-small Abelian groups and U-small categories are V-small. Unless context dictates otherwise, all commutative rings and schemes will be U-small. We shall consider variations on such monstrosities as the category CAT of V-small categories, which is not V-small, but ambiguity is unlikely to result from our refusal to name a sufficiently large third Grothendieck universe. Quasi-categories: We freely employ the language of quasi-categories and higher algebra as developed in [Lur09, Lur14]. For brevity, we contract the word “quasi-category” to “qcategory”. Categories as qcategories: We regard all categories as qcategories by tacitly taking their nerves. As justification for this convention, observe that the nerve functor N : Cat → Set∆ is right Quillen with respect to the model structure on Cat whose weak equivalences and fibrations are the equivalences of categories and the isofibrations, respectively, and the Joyal model structure on Set∆, and the functor induced between the qcategories underlying these model structures is fully faithful ([Joy08, 2.8]). Functors and limits: We say that a functor F : C → D between qcategories is U- continuous (resp. U-cocontinuous) if it preserves U-limits (resp. U-colimits), i.e., limits (resp. colimits) of U-small diagrams. We also write F (cid:97) G to indicate that the functor 4 F : C → D is left adjoint to the functor G : D → C ([Lur09, 5.2.2.1]). Presentability: Let κ denote an infinite regular U-cardinal. We preserve the terminology from the theory of 1-categories ([AR94]) and refer to an object X of a qcategory C as κ- presentable if mapC(X, −) : C → Spc preserves κ-filtered colimits, i.e., if it is “κ-compact” in the sense of [Lur09, 5.3.4.5]. We say that the qcategory C is locally U-presentable (resp. locally κ-presentable) if it is “presentable” (resp. “κ-compactly generated”) in the sense of [Lur09, 5.5.0.18, 5.5.7.1]. Localization: If C is a U-small qcategory and W a class of morphisms of C, then there exists a functor λ : C → C[W −1] with the universal property that, for each U-small qcategory −1], D) (cid:44)→ Fun(C, D) whose D, composition with λ induces a fully faithful functor Fun(C[W essential image is spanned by those functors that send each element of W to an equivalence −1], as a localization of C with respect in D ([Lur14, 1.3.4.2]). We refer to λ or, abusively, C[W to W. If λ admits a fully faithful right adjoint ι, then we say that λ is a reflective localization of C. This applies in particular to the locally presentable setting: if C is a locally U-presentable category and S is a U-set of morphisms of C, then the localization λ : C → C[S −1] is reflective, the essential image of its right adjoint is the full subqcategory spanned by the S-local objects, i.e., the objects X ∈ C such that, for each f ∈ S, the morphism mapC(f , X) is a weak −1] is locally U-presentable ([Lur09, 5.5.4.15, 5.5.4.20]). homotopy equivalence, and C[S Symmetric monoidal qcategories: A symmetric monoidal qcategory is, by definition ([Lur14, 2.0.0.7]), a coCartesian fibration p : C⊗ → Fin∗ such that the morphisms ρi : (cid:104)n(cid:105) → (cid:104)1(cid:105) given by ρi(j) := 1 if i = j and ρi(j) = ∗ if i (cid:44) j induce functors ρi (cid:104)1(cid:105) which in turn induce equivalences C⊗ (cid:104)1(cid:105))n. We systematically suppress the fibration p from ”. We also refer to C := C⊗ the notation, referring to “the symmetric monoidal qcategory C⊗ (cid:104)1(cid:105) : C⊗ → D⊗ as the qcategory underlying C⊗ ⊗ for a possibly . Similarly, we use the notation F lax symmetric monoidal functor and F : C → D for the underlying functor. (cid:104)n(cid:105) (cid:39) (C⊗ ! : C⊗ (cid:104)n(cid:105) → C⊗ × × Appealing to [Lur14, 2.4.2.6], the category CAlg(QCat ) of commutative algebra ob- jects of QCat is a convenient model for the qcategory of U-small symmetric monoidal qcategories: its objects correspond to U-small symmetric monoidal qcategories and its morphisms to symmetric monoidal functors. Qcategories underlying model categories: The model categories appearing in the sequel will prove to be U-combinatorial, i.e., cofibrantly generated model categories whose underlying categories are locally U-presentable ([Bek00, 1.8], [Lur09, §A.2.6], [Bar10, 1.21]). Many will even prove to be U-tractable model categories, i.e., U-combinatorial model categories whose generating cofibrations and trivial cofibrations have cofibrant domains ([Bar10, 1.21]). If M is a U-combinatorial model category and W is its class of weak equivalences, then −1] is locally U-presentable ([Lur14, 1.3.4.15, 1.3.4.16]). If is a symmetric monoidal U-combinatorial model category, then it admits an underly- ([Lur14, 4.1.3.6, 4.1.4.8]). its underlying qcategory M[W M⊗ − ing locally U-presentable symmetric monoidal qcategory M[W ⊗ ] 5 ⊗ −1] Strictly speaking, the qcategory underlying the symmetric monoidal qcategory M[W −1] spanned by the cofibrant objects of M, is defined to be the full subqcategory of M[W but the inclusion of this full subqcategory is an equivalence as each object of M is weakly equivalent to a cofibrant object. ([Lur14, 2.1.3.1]) ) Cℵ0 C(cid:113) C× −1] C[W ⊗ Fun hr Ab CAlg(C⊗ ) (C⊗ , D⊗ Notation: While we maintain most of the notations of [Lur09, Lur14], the following list specifies the notable deviations therefrom and other frequently recurring symbols. the full subqcategory of C spanned by the ℵ0-presentable objects ([Lur09, 5.3.4.5]) the coCartesian symmetric monoidal structure on the qcategory C ([Lur14, §2.4.3]) the Cartesian symmetric monoidal structure on the qcategory C ([Lur14, 2.4.1.1]) a localization of the qcategory C with respect to the class of morphisms W ([Lur14, 1.3.4.1]) the category of U-small Abelian groups the qcategory of commutative algebras in the symmetric monoidal qcategory C⊗ the qcategory of symmetric monoidal functors F ([Lur14, 2.1.3.7]) ≥r : C → C♥ ≤rt the degree r cohomology functor t the stable qcategory C ([Lur14, 1.2.1.4]) the homotopy category of the qcategory C ([Lur09, 1.2.3]) the ind-completion of the qcategory C ([Lur09, 5.3.5.1]) the mapping space between two objects X and Y of the qcategory C ([Lur09, 1.2.2]) the qcategory of modules over A ∈ CAlg(C⊗ the nerve of the category C two constructions of the differential graded nerve of the differential graded category C ([Lur14, 1.3.1.6, 1.3.1.16]) the simplicial nerve of the simplicial category C ([Lur09, 1.1.5.5]) Fun(Cop, D) the qcategory of U-small (resp. V-small) qcategories ([Lur09, 3.0.0.1]) the qcategory of U-small (resp. V-small) stable qcategories and exact functors ([Lur14, §1.1.4]) ModA(C) N(C) Ndg(C), Ndg(C) ho(C) Ind(C) mapC(X, Y) N∆(C) PSh(C, D) QCat ⊗ : C⊗ → D⊗ of a t-structure on ) ([Lur14, 4.5.1.1]) QCatEx, QCATEx 6 × QCat Schft /S /S, Smsft Smft /S Set∆ Spc Spc∗ ≤r, t ≥r t U ∈ V y /S spanned by the smooth (resp. smooth the Cartesian symmetric monoidal qcategories of U-small (resp. V-small) qcategories ([Lur14, 2.4.1.5]) the essentially U-small category of S-schemes of finite type the full subcategory of Schft and separated) S-schemes the category of simplicial U-sets the qcategory of U-small spaces, i.e., the qcategory underlying the model structure on Set∆ whose weak equivalences and fibrations are the weak homotopy equivalences and the Kan fibrations, respectively ([Lur09, 1.2.16.1]) the qcategory of pointed objects in Spc ([Lur14, 4.8.1.20]) the truncations of a cohomological t-structure on a stable qcategory C fixed Grothendieck universes the Yoneda embedding ([Lur09, §5.1.3]) 1 Deriving Tannakian categories −1] can be rectified to a functor F (cid:48) Motivation. One nice property of combinatorial model structures—among many oth- ers—is that they allow for convenient rectification results. For instance, by [Lur14, 1.3.4.25], if M is a U-combinatorial model category, W its class of weak equivalences and C a U-small category, then any functor F : C → M[W : C → M, i.e., : C → M whose composite with the localization M → M[W −1] (cid:48) there exists a functor F is equivalent to F. In a similar vein but under more restrictive hypotheses, if M⊗ is a U-combinatorial symmetric monoidal model category, then commutative algebras in the underlying symmetric monoidal qcategory M[W can be rectified to commutative algebras in M⊗ ([Lur14, 4.5.4.7]). Our goal in this section is to show that the category of complexes of ind-objects in a Tannakian category admits a U-combinatorial model structure allowing for both of these types of rectifications. Summary. After a brief review of the theory of Tannakian categories, we show that, for T a U-small Tannakian category, the categories Cpx(Ind(T)) and CAlg(Cpx(Ind(T)) ) admit U- combinatorial model structures (1.7, 1.8). Theorem 1.10 shows that commutative algebras ⊗ in D(Ind(T)) Definition 1.1. Let A be an Abelian category. can be rectified to commutative algebras in Cpx(Ind(T)) ⊗ −1] ⊗ ⊗ . (i) As in [Dre15, 3.1], we construct its bounded derived qcategory Db(A) as follows: take the differential graded nerve Kb(A) := Ndg(Cpxb(A)) of the differential graded category of bounded cochain complexes in A ([Lur14, 1.3.1.6]) and then take the Verdier quotient 7 Db(A) := Kb(A)/ Ac(A) with respect to the full subqcategory Ac(A) spanned by the acyclic complexes. The analogous construction for unbounded complexes results in the unbounded derived qcategory D(A) of A. (ii) The category A is U-Grothendieck Abelian if it is an Abelian, locally U-presentable category in which ℵ0-filtered U-colimits preserve finite limits. By [Bek00, 3.10], this is equivalent to the classical definition as an AB5 category with a generator. If A is es- sentially U-small, then its ind-completion Ind(A) is U-Grothendieck Abelian. If A is U-Grothendieck Abelian, then Cpx(A) admits a U-combinatorial model structure ([Lur09, A.2.6.1]) whose cofibrations and weak equivalences are the monomorphisms and quasi- isomorphisms, respectively ([Bek00, 3.13]), called the injective model structure and denoted by Cpx(A)inj. Its homotopy category is the unbounded derived category of A. Let D(A) denote the stable, locally U-presentable qcategory underlying Cpx(A)inj ([Lur14, 1.3.4.22]). is a symmetric monoidal structure on A, then Cpxb(A) and Cpx(A) inherit (iii) If A⊗ symmetric monoidal structures given informally by (cid:77) (Kr ⊗ Ln−r), r∈Z (1.1.1) (K⊗ L)n := d(x⊗ y) := d(x)⊗ y + (−1)deg(x)x⊗ d(y). ⊗ and Cpx(A) ⊗ , respectively. (iv) If A⊗ By [Lur14, 1.3.4.5, 4.1.3.4], Kb(A) and K(A) inherit symmetric monoidal structures from Cpxb(A) is a symmetric monoidal structure on A such that (−)⊗ (−) is exact separately in each variable, then, by [Dre15, 3.2], the Verdier quotient functor q : Kb(A) → Db(A) as a symmetric monoidal underlies a symmetric monoidal functor q ⊗ Verdier quotient of Kb(A) by Ac(A) in the sense of [Dre15, 1.5]. This means that, for each ⊗ stable symmetric monoidal qcategory C⊗ , composition with q induces a fully faithful ) (cid:44)→ Fun ⊗ ⊗ ⊗ , C⊗ , C⊗ (Kb(A) functor Fun ) whose essential image is spanned by the symmetric monoidal functors sending each object of Ac(A) to a zero object. Definition 1.2. An essentially U-small closed symmetric monoidal category T⊗ nakian if it satisfies the following conditions: ⊗ realizing Db(A) ⊗ (Db(A) is Tan- ⊗ (i) T is Abelian; (ii) homT(1T, 1T) is a field of characteristic zero; (iii) each object of T is ⊗-dualizable ([Lur14, 4.6.1.12]); ∨ (cid:39) V (iv) for each V ∈ T, the composite 1T η−→ V⊗ V ∨ ⊗ V ε−→ 1T is a nonnegative integer under the identification of Z with its image in the field homT(1T, 1T) of characteristic zero, where η and ε are the coevaluation and evaluation morphisms, respectively, and the equivalence V⊗ V Remark 1.3. This definition is more restrictive than the original one of [SR72, III, 3.2.1], but the two are equivalent once we require homT(1T, 1T) to be a field of characteristic zero by [Del90, 7.1]. ∨ ⊗ V is the symmetry isomorphism. ∨ (cid:39) V 8 Example 1.4. Let K be a field of characteristic zero. (i) If G is an affine gerbe on the fpqc-site (Schft /K)fpqc of K-schemes of finite type, then ∨ the essentially U-small category QCoh(G) of locally free quasi-coherent sheaves of finite rank on G is Tannakian when equipped with the usual tensor product of quasi-coherent sheaves. By [Del90, 1.12], every essentially U-small Tannakian category T arises in this way. ⊗ (ii) More concretely, the category MHSK of mixed Hodge K-structures ([Del71, 2.3.8]) ⊆ MHSK spanned by the objects (H, W, F) is Tannakian, as is the full subcategory MHSp K such that the pure Hodge K-structure on grW n (H) induced by the filtration F admits a polarization ([Del71, 2.1.15]) for each n ∈ Z. Remark 1.5. Let T⊗ be an essentially U-small Tannakian category. : T⊗ → ModK(cid:48)(Ab) ⊗ (i) By [Del90, 7.1], there exists a field extension K := homT(1T, 1T) (cid:44)→ K(cid:48) and a K- . By [DM89, 1.19], ω is linear, exact symmetric monoidal functor ω faithful. Using the exactness and faithfulness of ω, one finds that T is Noetherian, since ModK(cid:48)(Ab) is. By Noetherian, we mean that each family of subobjects of each fixed object V ∈ T contains a maximal element. (ii) If A is an Abelian category, we define homological dimension of A ∈ A to be hdim(A) := sup{n ∈ Z≥0 ∃B ∈ A[extn A(A, B)]} ∈ Z≥0 ∪{∞} and we define the homological dimension of A to be hdim(A) := sup{hdim(A) A ∈ A} ∈ Z≥0 ∪ {∞}. Since each V ∈ T is ⊗-dualizable, the adjunction (−) ⊗ V (cid:97) V ∨ ⊗ (−) shows that hdim(T) = hdim(1T). As a consequence, if hdim(1T) < ∞, then T satisfies the hypotheses of [Dre15, 4.7] and it follows that the ℵ0- presentable objects of D(Ind(T)) are precisely the ⊗-dualizable objects and the natural symmetric monoidal functor Ind(Db(T)) is an equivalence. This applies in particular to T⊗ ⊗ Lemma 1.6. Let K be a field of characteristic zero, F monoidal functor between two essentially U-small K-linear Tannakian categories. Then: → Cpx(Ind(T(cid:48) ⊗ (i) Cpx(Ind(F)) : Cpx(Ind(T)) inj ⊗ inj is a K-linear, exact, faithful symmetric )) K , since hdim(MHSp K) = 1 by [PS08, 3.35]. : T⊗ → T(cid:48)⊗ a K-linear, exact symmetric = MHSp,⊗ ⊗ → D(Ind(T)) ⊗ monoidal left Quillen functor; (ii) there is an essentially commutative square T F T(cid:48) Cpx(Ind(T)) Cpx(Ind(F)) / / Cpx(Ind(T(cid:48) )) in which vertical arrows are the evident inclusions in degree zero; and (iii) the functor D(Ind(F)) : D(Ind(T)) → D(Ind(T(cid:48) )) is conservative and t-exact with respect to the natural t-structures. Proof. By [DM89, 1.19], F is faithful. The existence of Cpx(Ind(F)), as well as its K-linearity, 9 / / _ _ exactness, faithfulness, U-cocontinuity and compatibility with the symmetric monoidal structures, follows from [SGA72a, Exposé I, 8.9.8, 8.6.4] and the obvious functorial prop- erties of complexes and categories of ind-objects. The Adjoint Functor Theorem ([AR94, 1.66]) implies that the U-cocontinuous functor Cpx(Ind(F)) is a left adjoint. As Cpx(Ind(F)) preserves quasi-isomorphisms and monomorphisms, it is left Quillen with respect to the injective model structures. This proves (i), and (ii) is obvious. As Cpx(Ind(F)) preserves quasi-isomorphisms, it induces D(Ind(F)) : D(Ind(T)) → D(Ind(T(cid:48) )) by the universal property of the localization. The derived functor of an exact functor between Abelian categories is t-exact with respect to the natural t-structures, so D(Ind(F)) is t-exact. Let us check that D(Ind(F)) is conservative. As D(Ind(F)) is an exact functor between stable qcategories, it suffices to check that it reflects zero objects. The natural t-structure on the derived qcategory of an Abelian category is nondegenerate, so it suffices to show that, for each object K of the heart D(Ind(T)) , D(Ind(F))(K) = 0 implies K = 0. In this case, we may assume K is concentrated in degree zero, given by an object V of T. As Ind(F) is faithful, we have a commutative square ♥ π0 mapD(Ind(T))(K, K) D(Ind(F)) π0 mapD(Ind(T(cid:48)))(D(Ind(F))(K), D(Ind(F))(K)) ∼ ∼ / homInd(T)(V, V) Ind(F) / homInd(T(cid:48))(Ind(F)(V), Ind(F)(V)) in which the vertical arrow on the right is injective and (iii) follows. Proposition 1.7. Let T⊗ proper, stable, U-tractable symmetric monoidal model category satisfying the monoid axiom. be a U-small Tannakian category. Then Cpx(Ind(T)) ⊗ inj is a left ⊗ ⊗ ⊗ ⊗ → ModK(cid:48)(Ab) ⊗ ) : T⊗ → ModK(cid:48)(Ab) ⊗ ⊗ → Cpx(ModK(cid:48)(Ab)) Proof. The U-tractability follows from the remark that all objects is cofibrant. The stability ⊗ follows from [Lur14, 1.3.4.24, 1.4.2.27]. Let us show that Cpx(Ind(T)) inj is a symmetric as in 1.5(i). As the canonical monoidal model category. Choose ω U-cocontinuous symmetric monoidal functor Ind(ModK(cid:48)(Ab)ℵ0 is an induces a K-linear, exact, faithful, U-cocontinuous symmetric monoidal equivalence, ω functor Cpx(Ind(T)) As 1T is cofibrant, it remains to establish the pushout-product axiom. Let f : K → K (cid:48) and g : L → L (cid:48) be two cofibrations of Cpx(Ind(T))inj, i.e., two monomorphisms. We claim that the canonical morphism f (cid:3) g : (K⊗ L (cid:48) is a monomorphism, and that it is moreover a quasi-isomorphism if f or g is. The image of f (cid:3) g under ω is the morphism ω(f ) (cid:3) ω(g). Faithful, exact functors preserve and reflect monomorphisms, ⊗ so the pushout-product axiom in Cpx(ModK(cid:48)(Ab)) inj, which holds by [Dre15, 2.3], implies that f (cid:3) g is a monomorphism. Similarly, the faithful, exact functor ω preserves and reflects quasi-isomorphisms, so if f or g is a quasi-isomorphism, then ω(f ) or ω(g) is. This implies that ω(f ) (cid:3) ω(g), and hence also f (cid:3) g, is a quasi-isomorphism. Hence, Cpx(Ind(T))inj is a by 1.6, which we abusively denote by ω )(cid:113)K⊗L (K (cid:48) ⊗ L) → K (cid:48) ⊗ L (cid:48) ⊗ . 10 / / _ symmetric monoidal model category. Since each object is cofibrant, it is left proper and satisfies the monoid axiom ([SS00, 3.4]). Lemma 1.8. Let T⊗ ) admits a U-combinatorial model structure whose weak equivalences (resp. fibrations) are the morphisms inducing quasi-isomorphisms (resp. fibrations) between the underlying objects of Cpx(Ind(T))inj. be a U-small Tannakian category. The category CAlg(Cpx(Ind(T)) ⊗ ⊗ Proof. By [Lur14, 3.2.3.5] and 1.7, CAlg(Cpx(Ind(T)) ) is locally U-presentable. It therefore suffices to construct a cofibrantly generated model structure with the prescribed weak equivalences and fibrations. Let f : K → L be a morphism of Cpx(Ind(T)). We have the pushout-product morphism f (cid:3)2 := f (cid:3) f : (K ⊗ L) (cid:113)K⊗K (L ⊗ K) → L ⊗ L. Iterating, we obtain morphisms f (cid:3)n for n ∈ Z≥0, which are Sn-equivariant with respect to the Sn-actions permuting factors of the tensor products appearing in the domain and codomain. We thus regard f (cid:3)n as a morphism of Cpx(Ind(T))Sn, the category of functors from the groupoid Sn into Cpx(Ind(T)). The functor Cpx(Ind(T)) → Cpx(Ind(T))Sn sending K to itself with the trivial Sn-action admits a left adjoint, the Sn-coinvariants functor (−)/Sn. ⊗ egory CAlg(Cpx(Ind(T)) ⊗ that Cpx(Ind(T)) cofibration f : K → L of Cpx(Ind(T)) ⊗ cofibration in Cpx(Ind(T))inj. By [Whi14, 3.2], the existence of a cofibrantly generated model structure on the cat- ) with the prescribed weak equivalences will follow if we show satisfies the commutative monoid axiom ([Whi14, 3.1]): for each trivial and each n ∈ Z>0, the morphism f (cid:3)n/Sn is a trivial : T⊗ → ⊗ As in 1.5, choose a K-linear, faithful, exact symmetric monoidal functor ω ⊗ ⊗ ℵ0 ModK(cid:48)(Ab) := Cpx(Ind(ω)) , and let f be a trivial cofibration of Cpx(Ind(T)) and n ∈ Z>0. By 1.6, ω reflects trivial cofibrations, so it suffices to show that ω(f (cid:3)n/Sn) is a trivial cofibration. On the other hand, ω also preserves trivial cofibrations by 1.6, so ωf is a trivial cofibration. Since ω is symmetric monoidal and U-cocontinuous we have ω(f (cid:3)n/Sn) (cid:39) (ωf )(cid:3)n/Sn. The claim therefore follows from the is freely powered ([Lur14, 7.1.4.7]), hence (ωf )(cid:3)n is a projective fact that Cpx(ModK(cid:48)(Ab)) trivial cofibration in Cpx(ModK(cid:48)(Ab))Sn. Indeed, as (−)/Sn is left Quillen with respect to the projective model structure on its domain, (ωf )(cid:3)n/Sn is a trivial cofibration, as desired. Lemma 1.9. Consider the following data: (i) T⊗ , a U-small Tannakian category; (ii) K := homT(1T, 1T) (cid:44)→ K(cid:48) , a field extension; : T⊗ → ModK(cid:48)(Ab) ⊗ ℵ0 , a K-linear exact symmetric monoidal functor; (iii) ω (cid:48) (iv) W (resp. W ⊗ CAlg(Cpx(Ind(T)) ), the class of weak equivalences in the model structure of 1.8 on the category ) (resp. CAlg(Cpx(ModK(cid:48)(Ab)) for some field extension homT(1T, 1T) (cid:44)→ K(cid:48) )); and ⊗ , let ω ⊗ ⊗ ⊗ ⊗ (v) C, a U-small category. 11 (cid:48) ⊗ ⊗ ⊗ )[W )[W )[W ⊗ : CAlg(Cpx(ModK(cid:48)(Ab)) If the forgetful functor ψ indexed colimits, then so does the forgetful functor ψ : CAlg(Cpx(Ind(T)) (cid:48)−1] → D(ModK(cid:48)(Ab)) preserves C- −1] → D(Ind(T)). ⊗ (cid:48) Proof. Suppose ψ and let CAlg(ω := Cpx(Ind(ω)) preserves C-indexed colimits. Let ω ) : −1] → CAlg(Cpx(ModK(cid:48)(Ab)) ⊗ ⊗ (cid:48)−1] denote the induced functor. CAlg(Cpx(Ind(T)) )[W ) (cid:39) ⊗ (cid:48) CAlg(ω (cid:48) It sends W to W by 1.6. We claim that there is a homotopy equivalence ψ ωψ. Indeed, the corresponding square of model categories is essentially commutative by inspection, and each functor involved preserves weak equivalences. Passing to underlying qcategories, we obtain the desired homotopy commutative square. ) be a functor. We claim that the canonical mor- phism colimγ∈C ψAγ → ψ colimγ∈C Aγ is an equivalence. As ω reflects weak equivalences by 1.6, it suffices to show that ω colimγ∈C ψAγ (cid:39) ωψ colimγ∈C Aγ is an equivalence. Note that ⊗ ⊗ is U-cocontinuous. In particular, it admits a lax symmetric monoidal right adjoint υ ), which is thus U-cocontinuous. preserve C-indexed colimits, provide a ω ([Lur14, 7.3.2.7]), and CAlg(υ These remarks, along with the hypothesis that ψ homotopy commutative diagram Let γ (cid:55)→ Aγ : C → CAlg(Cpx(Ind(T)) ) is right adjoint to CAlg(ω ⊗ ⊗ ⊗ (cid:48) colimγ∈C ψ (cid:48) CAlg(ω ∼ (cid:48) ψ colimγ∈C CAlg(ω ⊗ ⊗ )Aγ )Aγ ∼ colimγ∈C ωψAγ ∼ / ω colimγ∈C ψAγ ∼ / (cid:48) CAlg(ω ⊗ / ψ )colimγ∈C Aγ ∼ / / ωψ colimγ∈C Aγ ⊗ and the claim follows. Theorem 1.10. Let T⊗ phisms of CAlg(Cpx(Ind(T)) Cpx(Ind(T)), then the canonical functor φ : CAlg(Cpx(Ind(T)) is an equivalence. be a U-small Tannakian category. If W denotes the class of mor- ) inducing quasi-isomorphisms between the underlying objects of −1] → CAlg(D(Ind(T)) ⊗ ) for some field extension homT(1T, 1T) (cid:44)→ K(cid:48) : T⊗ → Proof. As in 1.5, choose a K-linear, faithful, exact symmetric monoidal functor ω ⊗ ⊗ ℵ0 ModK(cid:48)(Ab) := Cpx(Ind(ω)) . (cid:48)−1] → CAlg(D(ModK(cid:48)(Ab)) ⊗ By [Lur14, 7.1.4.7, 4.5.4.7], CAlg(Cpx(ModK(cid:48)(Ab) ) is an denotes the class of morphisms inducing quasi-isomorphisms equivalence, where W between the underlying complexes of K(cid:48) -modules. To prove the claim, we use this special case and the properties of ω to show that the conditions of [Lur14, 4.7.4.16] are satisfied. Consider the diagram and set ω ))[W )[W ⊗ ⊗ ⊗ ⊗ ⊗ (cid:48) CAlg(Cpx(Ind(T)) ⊗ −1] )[W CAlg(D(Ind(T)) ⊗ ) φ G D(Ind(T)) (cid:48) G 12 / / / / / ) ) v v (1) The qcategories D(Ind(T)) and CAlg(Cpx(Ind(T)) −1] are locally U-presentable by 1.1(ii) and 1.8: the qcategory underlying a U-combinatorial model category is lo- cally U-presentable ([Lur14, 1.3.4.22]). The qcategory CAlg(D(Ind(T)) ) is also locally U-presentable by [Lur14, 4.1.4.8, 3.2.3.5] and 1.7. The forgetful functor )[W ⊗ ⊗ ⊗ CAlg(Cpx(Ind(T)) ) → Cpx(Ind(T)) admits a left adjoint given by the free commutative algebra functor, denoted by sym, and ⊗ these form a Quillen adjunction by definition of the model structure on CAlg(Cpx(Ind(T)) ). In particular, by [Lur14, 1.3.4.27], as G is obtained from a right Quillen functor by passing (cid:48) to underlying qcategories, it admits a left adjoint F. By [Lur14, 3.1.3.5], G also admits a left adjoint F . (cid:48) is conservative by [Lur14, 3.2.2.6]. (2) The functor G preserves geometric realizations of simplicial objects by 1.9 and (cid:48) [Lur14, 7.1.4.7, 4.5.4.12]. The functor G preserves geometric realizations of simplicial objects by [Lur14, 3.2.3.2]. The functor G is conservative as it tautologically preserves weak (cid:48) equivalences, and G (K) → (3) We now claim that, for each K ∈ Cpx(Ind(T)), the canonical morphism G (cid:48) GF(K) is an equivalence. As explained in step (e) of the proof of [Lur14, 4.5.4.7], it suf- fices to prove that, for each K, the colimit defining the total symmetric power sym(K) := n∈Z≥0 symn(K) is a homotopy colimit. Let Lsym denote the corresponding homotopy colimit functor. As ω is symmetric monoidal, U-cocontinuous and homotopically U- cocontinuous, we have a homotopy commutative square (cid:96) (cid:48) F ⊗ Lsym(ωK) ∼ / ω Lsym(K) sym(ωK) ∼ / ω sym(K) and the left vertical arrow is an equivalence by [Lur14, 7.1.4.7] and step (e) of the proof of [Lur09, 4.5.4.7] applied to A⊗ . Since ω is conservative (1.6), the claim follows. Thus, the conditions of [Lur14, 4.7.4.16] are satisfied. ⊗ := Cpx(ModK(cid:48)(Ab)) 2 Mixed Hodge coefficients Notation 2.0. Throughout this section, we fix K (cid:44)→ R, a subfield of the real numbers. Motivation. While mixed Hodge structures arise very naturally in algebraic geometry, they tend to do so as the cohomology of much larger objects, to wit, mixed Hodge complexes. There is thus a dichotomy between complexes of mixed Hodge structures, which are hard to construct but form a very well-behaved category, and mixed Hodge complexes, which are much easier to construct but, as a 1-category, leave much to be desired. By a result of A. Beilinson ([Bei86, 3.11]), their derived categories are nevertheless equivalent. In this section, we translate this into an equivalence of symmetric monoidal qcategories. 13 / / Summary. We begin by reviewing the construction of the symmetric monoidal differ- ential graded category of mixed Hodge complexes (2.1, 2.2, 2.3). Theorems 2.6 and 2.7 lift A. Beilinson’s equivalence ([Bei86, 3.11]) to an equivalence of symmetric monoidal qcategories. Definition 2.1 ([Bei86, 3.9]). We recall the following constructions: (i) A mixed Hodge K-complex (K·, F, W, α, β) is a diagram β←− (K3, F, W) (K1, W) α−→ (K2, W) (2.1.1) in which (K1, W) (resp. (K2, W), resp. (K3, F, W)) is a filtered (resp. filtered, resp. bifil- tered) complex of K-modules (resp. C-modules, resp. C-modules), α is a filtered quasi- isomorphism (K1 ⊗K C, W⊗K C) ∼→ (K2, W), β is a filtered quasi-isomorphism (K3, W) ∼→ (K2, W) and the following conditions are satisfied: (cid:76) (a) (b) for each k ∈ Z, the differentials of the complex grW with the filtration induced by F; and (c) for each (k, n) ∈ Z2, the isomorphism β −1α : hn grW k (K3) and k (K1) with a pure Hodge structure of k (K3) are strictly compatible k (K1)⊗K C ∼→ hn grW n∈Z hnK1 is of finite rank; the filtration induced by F endow hn grW weight k + n. We refer to the filtrations denoted by “W” as weight filtrations and those by “F” as Hodge filtrations. By convention, W will always be increasing and F will always be decreasing. We will also frequently suppress the morphisms α and β and refer abusively to the mixed Hodge K-complex (K·, F, W). Note that this definition is not equivalent to [Bei86, 3.2], but rather to [Bei86, 3.9], the difference being the shift in the weights by the cohomological degree appearing in condition (c) above. It seems likely that the following techniques apply to both settings with slight modification. (k, n) ∈ Z2, the pure Hodge K-structure (hn grW (ii) We say that a mixed Hodge K-complex (K·, F, W, α, β) is polarizable if, for each (iii) We define a morphism of polarizable mixed Hodge K-complexes f : (K·, F, W, α, β) → (cid:48) (cid:48)·, F, W, α ) to be a morphism of diagrams, consisting of morphisms of (bi)filtered (K complexes f1 : (K1, W) → (K 1, W), f2 : (K2, W) → (K 2, W) and f3 : (K3, F, W) → (K (cid:48) (cid:48) (cid:48) 3, F, W) (f1⊗KC) = (f2⊗KC)α and β K denote the category polarizable such that α mixed Hodge K-complexes and morphisms of such. K inherits a K-linear-differential-graded-category structure from the differential graded categories of filtered (resp. filtered, resp. bifiltered) complexes of K-modules (resp. C-modules, resp. C-modules) as explained in [Ivo15, 1.2.1]. k (K1), F) is polarizable. f3 = f3β. Let MHCp The category MHCp With translations and cones defined in the evident way ([Ivo15, 1.2.6]), MHCp K is pretriangulated. Note that MHCp K is locally U-small, but neither essentially U-small nor locally U-presentable, so we will view it as a V-small category for some suitably large Grothendieck universe V containing U. (cid:48) , β (cid:48) (cid:48) 14 Definition 2.2. Let K and L be two complexes of K-modules. (i) In (1.1.1), we defined the tensor product K⊗ L. If K and L are given filtrations F and G, respectively, then we define the tensor product (K, F)⊗K (L, G) to be K⊗K L equipped with the filtration given by (cid:77) (cid:88) r∈Z s∈Z (F⊗K G)k(K⊗K L)n := (FsKr ⊗K Gk−sLn−r) for each (k, n) ∈ Z2. Defining the tensor product of the additional filtrations analogously, we have a reasonable construction of the tensor product of bifiltered complexes. With this definition, the category of increasingly filtered complexes of K-modules is a symmetric monoidal category with unit given by K as a complex concentrated in degree 0 equipped with the trivial filtration FkK := 0 for k ∈ Z<0 and FkK = K for k ∈ Z≥0. This too extends easily to the setting of bifiltered complexes. , W ⊗ W (cid:48) , β , α ⊗ α ) is defined by complexes (K·, F, W, α, β) and (K (ii) The tensor product (K· ⊗ K (cid:48)·, F ⊗ F (cid:48) (cid:48) (cid:48) (cid:48)·, F (cid:48) , W , α α⊗Cα 2, W⊗C W −−−−−→ (K2 ⊗C K (cid:48) (cid:48) β⊗Cβ (cid:48)←−−−−− (K3 ⊗C K ) 3, F⊗C F (cid:48) (cid:48) , W⊗C W (cid:48) ). ) of two mixed Hodge K- 1, W⊗K W (cid:48) (K1 ⊗K K (cid:48) , β ⊗ β (cid:48) (cid:48) (cid:48) ) (cid:48) K of V-small K-linear differential graded categories ([Dre15, 2.1(vi)]). K , we construct a V-small stable symmetric monoidal qcategory Ndg(MHCp K) A filtered variant of the Künneth formula, along with the observation that the tensor product of two polarizable pure Hodge K-structures is another such, show that this is another object of MHCp K ([PS08, 3.20]), so this tensor product makes MHCp K a symmetric monoidal category MHCp,⊗ K . The tensor product bifunctor is compatible with the differential graded structures on filtered and bifiltered complexes, so MHCp,⊗ K is in fact a symmetric monoidal K-linear dif- ferential graded category, i.e., a commutative monoid in the symmetric monoidal category DGCAT⊗ (iii) From the pretriangulated K-linear symmetric monoidal differential graded cate- gory MHCp,⊗ ⊗ ∈ CAlg(QCATEx,⊗ denotes the qcategory of V-small sta- ble qcategories equipped with the symmetric monoidal structure of [Lur11, 5.4.7]. The homotopy category ho(Ndg(MHCp Definition 2.3. To each K ∈ Cpxb(MHSp K) we assign a diagram (K·, F, W) as in (2.1.1) satisfying conditions (a) and (b) of 2.1(i) in the evident way: K1 and K2 = K3 are the complexes of K-modules and C-modules underlying K, respectively, and F and W are the Hodge and weight filtrations, respectively. We must, however, shift the weight filtration in order to obtain a diagram satisfying 2.1(i)(c), setting Wk(Kn) := Wk+n(Kn) for each (k, n) ∈ Z2. This assignment K (cid:55)→ (K·, F, W) extends to a K-linear differential graded functor χ : Cpxb(MHSp ), using [Dre15, 2.5], where QCATEx,⊗ K)) is equivalent to H0(MHCp K) ([Dre15, 2.1(viii)]). K) → MHCp K. 15 (cid:18)(cid:77) (cid:19) (Kr ⊗K Ln−r) = ( W⊗K W)k(χ(K)⊗K χ(L))n = r∈Z (cid:77) (cid:77) r∈Z (cid:88) (cid:88) s∈Z r∈Z t∈Z (WsKr ⊗K Wk+n−sLn−r), (Wr+tKr ⊗K Wk+n−(r+t)Ln−r). ⊗ K) → MHCp Proposition 2.4. The K-linear differential graded functor χ : Cpxb(MHSp K is symmetric monoidal with respect to the symmetric monoidal structures of 1.1(iii) and 2.2(ii) and induces an exact symmetric monoidal functor χ Proof. Once we have shown that χ underlies a K-linear symmetric monoidal differential graded functor, the last assertion will follow from [Dre15, 2.5]. Let K and L be objects K). By definition, then complex of K-modules underlying χ(K ⊗ L) is the of Cpxb(MHSp complex of K-modules underlying K⊗ L. To see that χ is symmetric monoidal, it therefore suffices to check that the Hodge and weight filtrations on χ(K⊗ L) are equal to the tensor products of the Hodge and weight filtrations, respectively, on χ(K) and χ(L). Fix (k, n) ∈ Z2. We have the following computations: ⊗ → Ndg(MHCp K) : Kb(MHSp K) ⊗ . Wk(χ(K⊗K L))n = Wk+n ⊗ Reindexing the last expression by t := s − r, we find that the two filtrations are equal. Essentially the same argument applies for the Hodge filtrations. Sparing the reader the predictably tedious verification that χ satisfies the coherence properties required of a symmetric monoidal differential graded functor, the claim follows. Definition 2.5. Let (K·, F, W) ∈ MHCp K. We say that (K·, F, W) is acyclic if the underly- ing complex of K-modules K1 is acyclic. Let ι : Ac (cid:44)→ Ndg(MHCp K) denote the full sub- qcategory spanned by the acyclic objects. Since the forgetful differential graded functor (K·, F, W) (cid:55)→ K1 : MHCp → Cpx(ModK(Ab)) preserves cones and translations, Ac is a stable subqcategory and ι is an exact functor. We define MHCp K)/ Ac to be the cofiber of ι in QCATEx, i.e., the Verdier quotient of Ndg(MHCp Theorem 2.6 ([Bei86, 3.11]). The exact functor χ : Kb(MHSp an equivalence χ : Db(MHSp K := Ndg(MHCp K) → Ndg(MHCp K) of 2.4 induces K) ∼→ MHCp K. K) by Ac. K K by the universal property of the cofiber defining D(MHSp Proof. A complex of polarizable mixed Hodge K-structures is acyclic if and only if the com- K) → plex of underlying K-modules is acyclic, so χ induces an exact functor χ : Db(MHSp MHCp K) (1.1(i)). At the level of homotopy categories, χ induces the functor of [Bei86, 3.11], so ho(χ) is an equivalence. This proves the first assertion: the exact functor χ, whose domain and codomain are stable qcategories, is an equivalence if and only if ho(χ) is an equivalence. Theorem 2.7. The canonical functor π : Ndg(MHCp monoidal Verdier quotient π K underlies a symmetric by Ac 1.1(iv) K) → MHCp ⊗ K of Ndg(MHCp K) ⊗ → MHCp,⊗ : Ndg(MHCp K) ⊗ 16 ⊗ and the equivalence χ of 2.6 underlies a symmetric monoidal equivalence χ MHCp,⊗ K . Proof. The functor (K·, F, W) (cid:55)→ K1 : MHCp,⊗ and reflects acyclicity, so it follows from [Dre15, 3.4] applied to A⊗ tensor product in MHCp assertion follows by [Dre15, 2.6]. → Cpx(ModK(Ab)) is symmetric monoidal that the K preserves acyclic objects separately in each variable. The first : Db(MHSp K) ⊗ → ⊗ = ModK(Ab) K By the universal property of the symmetric monoidal Verdier quotient π , the com- ⊗ → ⊗ ⊗ factors through the symmetric monoidal Verdier quotient Kb(MHSp posite π K) χ ⊗ Db(MHSp of 1.1(iv). The exact functor underlying the resulting symmetric monoidal K) ⊗ → MHCp,⊗ ⊗ : Db(MHSp K must be equivalent to χ by the universal property functor χ K) of the cofiber D(MHSp K) := Kb(MHSp K) de- notes the full subqcategory spanned by acyclic complexes. By [Lur14, 2.1.3.8], the fact (2.6) that χ is an equivalence implies that the symmetric monoidal functor χ is an equiva- lence. K) (cid:44)→ Kb(MHSp K), where Ac(MHSp K)/ Ac(MHSp ⊗ ⊗ 3 Rectification Notation 3.0. Throughout this section, we fix the following notation: (i) K (cid:44)→ R, a subfield of the real numbers; and (ii) κ (cid:44)→ C, a subfield of the complex numbers. Motivation. We arrive now at our intended applications. Having in the previous two sections constructed the requisite equivalence between the symmetric monoidal qcate- gories of complexes of mixed Hodge structures and mixed Hodge complexes and the necessary ingredients for the rectification of presheaves of commutative algebras in the symmetric monoidal derived category of mixed Hodge structures, we now perform the desired rectifications. Specifically, we show that the functor assigning to each X ∈ Schft • /κ Betti(X,K) of [Del74, 8.2.1], equipped with the graded polarizable mixed Hodge structure H the ring structure given by the cup product, can be obtained by taking the cohomology of a presheaf ΓHdg : (Schft /κ)op → CAlg(Cpx(Ind(MHSp K)) ⊗ ) of commutative algebras in the symmetric monoidal category Cpx(Ind(MHSp Summary. We begin by constructing the presheaf ΓHdg on the category of separated κ- schemes of finite type (3.2). In order to extend ΓHdg to singular κ-schemes, we appeal to a result of V. Voevodsky that requires some terminology from A1-homotopy theory, which we recall in 3.4. Proposition 3.5 is a general result providing sufficient conditions for a /κ. The desired presheaf on Schft presheaf on Smsft /κ is then constructed in 3.6. /κ to extend naturally to a functor on Schft K)) . ⊗ 17 Definition 3.1. We denote by Cpt the category of smooth compactifications, whose objects are the dense open immersions j : X (cid:44)→ X in Smsft /C such that X is smooth and proper over Spec(C) and X− X is a normal crossings divisor, and whose morphisms are commutative squares in Smsft /C. We abusively denote objects of Cpt by ordered pairs (X, X), suppressing the morphism j. Theorem 3.2. There exists a functor ΓHdg : (Smsft ) such that, /κ and each r ∈ Z, hr ΓHdg(X) is naturally isomorphic to the K-linear Betti for each X ∈ Smsft Betti(X⊗κ C,K) equipped with the mixed Hodge K-structure of [Del71, 3.2.5]. cohomology Hr Proof. It suffices to treat the case in which κ = C and then compose with the functor /C)op. Making the obvious modification from the Q-linear to the K-linear (Smsft setting and forgetting the Z-linear component, [NA87, 8.15] provides us with a functor Γ 0 : Cptop → CAlg(Z0(MHCp,⊗ K )). By construction, for each r ∈ Z, the image of an object /κ )op → CAlg(Cpx(Ind(MHSp K)) /κ )op → (Smsft ⊗ (X, X) under the composite (3.2.1) Cptop Γ 0−−→ CAlg(Z0(MHCp,⊗ K )) → Z0(MHCp K) → MHCp K χ −1−−−→ Db(MHSp K) hr−−→ MHSp K is naturally isomorphic to Deligne’s mixed Hodge structure on Hr in [Del71, 3.2.5(iii)]. Betti(X,K) as constructed Let W be the class of weak equivalences of the model structure of 1.8 on the category −1] as the ). We construct Γ : Cptop → CAlg(Cpx(Ind(MHSp ⊗ K)) ⊗ K)) )[W CAlg(Cpx(Ind(MHSp following composite: / CAlg(Z0(MHCp,⊗ K )) ⊗ / CAlg(Ndg(MHCp K) ) q Γ 0 π Cptop Γ CAlg(MHCp,⊗ K ) ∼ −1 χ CAlg(Cpx(Ind(MHSp K)) ⊗ −1] −1 φ ∼o )[W ⊗ K)) CAlg(D(Ind(MHSp ι ) CAlg(Db(MHSp K) ). −1 are those of 2.7, ι is that of [Dre15, −1 do not affect cohomology objects, so Γ also ⊗ −1 is that of 1.10. Note that ι and φ Here, q is the functor given by [Dre15, 2.7], π and χ 4.7] and φ recovers Deligne’s mixed Hodge structures by (3.2.1). As Γ ∈ Fun(Cptop, CAlg(Cpx(Ind(MHSp K)) ⊗ −1]) )[W and its codomain is the qcategory underlying a U-combinatorial model category (1.8), [Lur14, 1.3.4.25] implies that Γ can be rectified to Γ : Cptop → CAlg(Cpx(Ind(MHSp K)) ⊗ ). 18 / / o ? _ o o We now have a functor Γ between 1-categories. Let CptX ⊆ Cpt denote the subcategory of smooth compactifications of a fixed object X ∈ Smsft /C, i.e., the subcategory spanned by the morphisms (f , f ) such that f = idX. Then CptX is nonempty by theorems of M. Nagata ([Con07, 4.1]) and H. Hironaka ([Hir64]) and ℵ0-filtered by a standard argument ([Del71, 3.2.11]). Also, if (f , f ) : (X, X) → (X, X ) is a morphism of Cpt, then Γ (f , f ) ∈ W. Indeed, the morphism of complexes of K-modules underlying Γ 0(f , f ) is the identity on the singular /C)op → cochain complex of X. We may therefore construct the desired functor ΓHdg : (Smsft CAlg(Cpx(Ind(MHSp ) by defining (cid:48) ⊗ K)) ΓHdg(X) := colim X∈CptX Γ (X, X) and having ΓHdg act in the evident way on morphisms. Corollary 3.3. The functor ΓHdg of 3.2 induces a functor ΓHdg : (Smsft underlying a symmetric monoidal functor Γ ⊗ Hdg : (Smsft /κ )op,(cid:113) → D(Ind(MHSp ⊗ K)) . /κ )op → D(Ind(MHSp K)) ⊗ K)) )[W −1] /κ )op and Γ λ : CAlg(Cpx(Ind(MHSp Proof. Recall that (Smsft ([Lur14, §2.4.3]). Let Γ denotes the coCartesian symmetric monoidal structure /κ )op,(cid:113) (cid:48) Hdg be the composite of ΓHdg, the localization ) → CAlg(Cpx(Ind(MHSp ⊗ K)) (cid:48) (1.8) and the equivalence φ of 1.10. Let ΓHdg denote the composite of Γ Hdg with the forgetful ) → D(Ind(MHSp ⊗ functor ψ : CAlg(D(Ind(MHSp K)) K)). By [Lur14, 3.2.4.9], ΓHdg underlies a ⊗ lax symmetric monoidal functor Γ Hdg with respect to the coCartesian symmetric monoidal ⊗ (cid:48) Hdg is symmetric monoidal if Γ structure on (Smsft Hdg preserves finite coprod- (cid:48) ucts. The object of D(Ind(MHSp Hdg(Y) in ), i.e., its image under ψ, is the tensor product ΓHdg(X) ⊗ ΓHdg(Y) CAlg(D(Ind(MHSp ([Lur14, 3.2.4.8]). By the Künneth formula, ΓHdg(X)⊗ ΓHdg(Y) (cid:39) ΓHdg(X×κ Y) and X×κ Y is /κ )op. Indeed, composing with the conservative symmetric the coproduct of X and Y in (Smsft ⊗ → D(ModK(Ab)) ⊗ monoidal functor D(Ind(MHSp K)) reduces the problem to the Künneth formula for Betti cohomology. As ψ is conservative, the claim follows. Definition 3.4. Let S be a quasi-compact quasi-separated scheme. Let S ⊆ Sch/S be a full subcategory stable under fiber products and containing ∅, S and A1 S, C a qcategory, F : Sop → C a functor, and Q a class of Cartesian squares in S of the form (3.4.1) K)) underlying the coproduct of Γ (cid:48) Hdg(X) and Γ ⊗ K)) (cid:48) Y (cid:48) (cid:48) X g Y g / X. (cid:48) f Q f 19 / / / (i) We say that F is: (a) excisive with respect to Q if F(∅) is a final object of C and, for each Q ∈ Q as in (3.4.1), the square F(Q) is Cartesian in C; (b) Nisnevich excisive if it is excisive with respect to the class QNis(S) of squares of the form (3.4.1) such that f is an open immersion, g an étale morphism and the induced morphism g → (X− X −1(X− X )red is an isomorphism; )red (c) cdh-excisive if it is Nisnevich excisive and also excisive with respect to the class Qcdh(S) of squares of the form (3.4.1) such that f is a closed immersion, g is proper and the induced morphism g is an isomorphism; ) → X− X −1(X− X (cid:48) (cid:48) (cid:48) (cid:48) (d) scdh-excisive if it is Nisnevich excisive and also excisive with respect to the class are smooth Qscdh(S) of squares of the form (3.4.1) such that X, X S-schemes, f is a closed immersion and g is the blow-up of X along X (e) A1-invariant if, for each X ∈ S, the morphism F(X) → F(A1 , Y and Y X) induced by the ; (cid:48) (cid:48) (cid:48) canonical projection is an equivalence in C. (ii) Let C = Spc be the qcategory of spaces. If τ = Nis (resp. τ = cdh, resp. τ = scdh), then we let Shτ(S, Spc) ⊆ PSh(S, C) denote the full subqcategory spanned by the Nisnevich- excisive (resp. cdh-excisive, resp. scdh-excisive) functors, and we let Hτ(S) ⊆ Shτ(S, Spc) denote the full subqcategory spanned by the functors which are moreover A1-invariant. The inclusions Shτ(S, Spc) (cid:44)→ PSh(S, Spc) and Hτ(S) (cid:44)→ Shτ(S, Spc) are reflective subqcat- egories with respective left adjoints λτ and λA1. Indeed, by the Yoneda lemma ([Lur09, 5.1.3.1]), the excision property for F ∈ PSh(S, Spc) is equivalent to requiring that F be Wτ-local ([Lur09, 5.5.4.1]), where Wτ is the class of morphisms of the form ζQ : y (X (cid:48) )(cid:113)y (Y(cid:48)) y (Y) → y (X) induced by the universal property of the pushout with Q ∈ Qτ(S), where y : S (cid:44)→ PSh(S, Spc) denotes the Yoneda embedding. Similarly, A1-invariance is equivalent to re- X) → y (X) quiring that F be WA1-local, where WA1 is the class of morphisms of the form y (A1 with X ∈ S. The subqcategories Shτ(S, Spc) and Hτ(S) are therefore reflective by [Lur09, 5.5.4.15]. Proposition 3.5. Let C⊗ /κ )op → CAlg(C⊗ F : (Smsft /κ (cid:44)→ Schft /κ)op → CAlg(C⊗ ) a Nisnevich-excisive, A1-invariant functor. /κ denotes the inclusion, there exists a cdh-excisive, A1-invariant be a stable locally U-presentable symmetric monoidal qcategory and (i) If ι : Smsft functor F : (Schft ) such that Fιop (cid:39) F. ⊗ (ii) If F corresponds to a symmetric monoidal functor F via [Lur14, 2.4.3.18], then F corresponds to a symmetric monoidal functor F Proof. By [Lur14, 3.2.3.5], CAlg(C⊗ excisive. The forgetful functor φ : CAlg(C⊗ ) is locally U-presentable. We claim that F is scdh- ) → C reflects limits ([Lur14, 3.2.2.5]), so it . : (Smsft ⊗ /κ )op,(cid:113) → C⊗ : (Schft /κ)op,(cid:113) → C⊗ 20 suffices to show that the composite F := φF is scdh-excisive. For each Q ∈ Qscdh(Smft /κ), the square F(Q) is Cartesian in C if and only if, for each C ∈ C, the square mapC(C, F(Q)) is Cartesian in Spc∗. Indeed, this follows from the Yoneda lemma ([Lur09, 5.1.3.1]). Thus, F is scdh-excisive if and only if the functor mapC(C, F(−)) : (Smft /κ)op → Spc∗ is scdh-excisive for each C ∈ C. Since C is stable, we have an equivalence mapC(Σ1C, F(−)) (cid:39) Ω1 mapC(C, F(−)) for each C ∈ C, where Σ1 denotes the suspension endofunctor of C and Ω1 denotes the loop functor of Spc∗ ([Lur14, 1.1.2.6]). The claim that F, and hence also F, is scdh-excisive therefore follows from [Bla01, Lemma 5.1]. By [Lur09, 5.1.5.6], there is a canonical equivalence Funcont(PSh(Smsft /κ , Spc)op, CAlg(C⊗ )) ∼→ PSh(Smsft /κ , CAlg(C⊗ )) under which F corresponds to an object in the essential image of the fully faithful functor Funcont(Hscdh(Smsft F(X)⊗ F(Y) ∼ αXY F(X×κ Y) ∼ F(Xred)⊗ F(Y) αXredY/ / F(Xred ×κ Y) 21 /κ )op, CAlg(C⊗ /κ )op → CAlg(C⊗ )) (cid:44)→ Funcont(PSh(Smsft op scdh : PSh(Smsft op A1λ given by composition with the localization λ (cid:48) of 3.4(ii). Let F ∗ 4.7], composition with ιop induces an equivalence ι denote the composite : Hscdh(Smsft /κ )op ) denote the corresponding functor. By [Voe10, /κ ). Let F : Hcdh(Schft )) /κ , Spc)op, CAlg(C⊗ /κ , Spc)op → Hscdh(Smsft /κ) ∼→ Hscdh(Smsft (cid:48)−−→ CAlg(C⊗ ). (Schft /κ)op → Hcdh(Schft /κ)op ι ∗,op−−−→ Hscdh(Smsft /κ )op F ⊗ Suppose the lax symmetric monoidal functor F By construction, F is cdh-excisive and A1-invariant, and Fιop (cid:39) F, which proves (i). Note that F inherits A1-invariance and cdh-excisiveness from F. : (Smsft associated with F by [Lur14, 2.4.3.18] is in fact a symmetric monoidal functor. We must show that the lax ⊗ associated with F is also symmetric monoidal. By [Lur14, symmetric monoidal functor F 3.2.4.9], this amounts to showing that, for all X and Y in Schsft /κ , the canonical morphism αXY : F(X)⊗ F(Y) → F(X×κ Y) is an equivalence. /κ )op,(cid:113) → C⊗ Fix objects X and Y of Schft /κ. Since F restricts to a functor equivalent to F on (Smsft /κ )op ⊗ and F is symmetric monoidal, αXY is an equivalence if X and Y are smooth, separated κ-schemes. Also, if X = ∅, then X ×κ Y = ∅ and F(X), F(X)⊗ F(Y) and F(X ×κ Y) are zero objects in C, so we may assume X and Y are nonempty and X is singular. Suppose X and Y are separated over κ and Y is smooth over κ. If dim(X) = 0, then Xred is also smooth and Xred×κ Y = (X×κ Y)red. By cdh-excision, the inclusion Xred (cid:44)→ X induces an equivalence F(X) ∼→ F(Xred), so αXY is an equivalence, because we have a homotopy commutative square / / If dim(X) > 0, suppose αSY is an equivalence for all separated κ-schemes S of dimension < dim(X). By [Hir64], there exists an element (3.5.1) Z Z Q X / X of Qcdh(Schft dim(X). Tensoring F(Q) with F(Y), we have a Cartesian square /κ) such that X is a smooth κ-scheme, dim(Z) < dim(X) and dim(Z) < dim(X) = F(X)⊗ F(Y) F(X)⊗ F(Y) F(Z)⊗ F(Y) / F(Z)⊗ F(Y). is a stable symmetric monoidal qcategory, so the endofunctor (−)⊗ C is exact Indeed, C⊗ for each C ∈ C and, in particular, it preserves Cartesian squares. The morphisms αZY and αZY are equivalences by the inductive hypothesis and αXY is an equivalence since X and Y are both smooth, separated κ-schemes. It follows that αXY is an equivalence. Suppose both X and Y are separated κ-schemes. If dim(X) = 0, then Xred is smooth over κ, so αXredY is an equivalence by the previous case and we have a homotopy commutative square F(X×κ Y) ∼ F(X)⊗ F(Y) ∼ F(Xred)⊗ F(Y) ∼ / / F(Xred ×κ Y) in which the vertical arrows are equivalences by cdh-excision, since Xred (cid:44)→ X is a universal homeomorphism. In general, suppose αSY is an equivalence for each separated κ-scheme S of dimension < dim(X) and consider the square Q of (3.5.1). Tensoring F(Q) with F(Y) and again using the fact that C⊗ is a stable symmetric monoidal qcategory, it suffices to show that αZY, αZY and αXY are equivalences. However, αZY and αZY are equivalences by the inductive hypothesis, and αXY is also an equivalence: permuting the tensor factors, it becomes αYX and X is smooth, so we are in the previous case. If Y is separated and X is arbitrary, choose a finite Zariski cover {jβ : Xβ (cid:44)→ X}1≤β≤n such that Xβ is separated for each 1 ≤ β ≤ n and n > 1. Let X 1≤β<n Xβ. We have an element :=(cid:83) (cid:48) X (cid:48) ∩ Xn (cid:48) j n X (cid:48) (cid:48)(cid:48) j Q (cid:48) j Xn jn / X 22 / / / / / / / / / / / (cid:48) /κ) and X (cid:48) ∩ Xn and X are both unions of n − 1 separated open subschemes. of QNis(Schft Applying the Nisnevich-excisive functor F to the square Q and tensoring F(Q) with F(Y), we find by induction on n that αXY is an equivalence. Inducting now on the number of elements in a Zariski cover of Y by separated S-subschemes, we find that αXY is an equivalence for arbitrary X and Y. Theorem 3.6. that, for each X ∈ Schft equipped with the mixed Hodge K-structure of [Del74, 8.2.1]. /κ and each r ∈ Z, hrΓHdg(X) is naturally isomorphic to Hr /κ)op,(cid:113) → D(Ind(MHSp K)) (i) There exists a symmetric monoidal functor Γ such Betti(X⊗κ C,K) ⊗ Hdg : (Schft ⊗ (ii) The underlying functor ΓHdg factors up to equivalence as (Schft /κ)op ΓHdg−−−→ CAlg(Cpx(Ind(MHSp K)) ⊗ ) φ (cid:48)−−→ D(Ind(MHSp K)), ⊗ K)) is induced by the functor φ of 1.10. := D(Ind(MHSp (cid:48) where φ Proof. For brevity, let C⊗ denote the symmet- ⊗ ric monoidal functor denoted by Γ Hdg in 3.3. Composing F with the “underlying complex of K-modules” functor ω : C → D(ModK(Ab)) results in a functor assigning to each X ∈ Smsft /κ the K-linear singular cochain complex of X(C)an or, in any case, a complex of K-modules quasi-isomorphic to it. /κ )op,(cid:113) → C⊗ ⊗ . Let F : (Smsft Betti cohomology is A1-homotopy invariant and Nisnevich excisive: the A1-invariance C(C)an and homotopy invariance of singular coho- follows from the contractibility of A1 mology; one may check that it is Nisnevich excisive by standard cohomological descent arguments ([SGA72b, Exposé V bis, 4.1.8, 5.2.3]), or use the fact ([Ayo10, 3.3]) that the derived categories D(X(C)an, K) of analytic sheaves of complexes of K-modules form a stable homotopy 2-functor ([Ayo07a, 1.4.1]) and remark that the existence of localization sequences ([Ayo07a, 1.4.9]) implies Nisnevich excision. As ω is conservative by 1.6, F is also A1-invariant and Nisnevich excisive. Claim (i) now follows from 3.5. ⊗ Hdg of (i) is classified by an essentially unique functor By [Lur14, 2.4.3.18], the functor Γ (cid:48) Γ Hdg : (Schft /κ)op → CAlg(C⊗ ). By 1.10, we therefore have a functor /κ)op → CAlg(Cpx(Ind(MHSp ⊗ K)) (cid:48) Hdg : (Schft −1 Γ φ −1]. )[W By 1.8 and [Lur14, 1.3.4.25], we can rectify φ CAlg(Cpx(Ind(MHSp ), which proves (ii). ⊗ K)) −1 Γ (cid:48) Hdg to a functor ΓHdg : (Schft /κ)op → 4 Integral coefficients Notation 4.0. Throughout this section, we fix the following: 23 Λ⊗Z Q is a field, e.g., Λ ∈ {Z,Q,R}; (i) Λ (cid:44)→ R, a Noetherian subring of the real numbers of global dimension ≤ 1 such that (ii) K := Λ⊗Z Q; and (iii) κ (cid:44)→ C, a subfield of the complex numbers. Motivation. Working in the K-linear setting has simplified things in several ways: it allowed us to apply the results of [Dre15, §2] to pass from symmetric monoidal differen- tial graded categories to symmetric monoidal qcategories without being forced to deal with cofibrant resolutions of our differential graded categories; it allowed us to equip the bounded derived category of MHSp K with a symmetric monoidal structure without constructing flat resolutions; and it allowed us to apply the rectification of 1.10. In this section, we show that it is possible to work with integral rather than rational coefficients. However, whereas in 3.6 we constructed a presheaf of strictly commutative differential graded algebras, in the integral setting, one may at best hope for a presheaf of E∞-algebras. In fact, we content ourselves to ask for a presheaf of E∞-algebras at the level of symmetric monoidal qcategories and set aside the question of establishing an analogue of the rectification result 1.10 for E∞-algebras with integral coefficients. Summary. We begin by constructing a t-structure on the fiber product of two stable qcategories equipped with t-structures over a third (4.2) and studying the heart of this t-structure (4.3). We then apply this to show that the derived qcategory of MHSp Λ is the fiber product of Db(MHSp ) (4.5). After constructing a symmetric monoidal functor computing Λ-linear Betti cohomology (4.8), this allows us to establish in 4.9 the Λ-linear analogue of 3.6(i). Lemma 4.1. Consider a commutative diagram / A K) and Db(ModΛ(Ab)ℵ0 ) over D(ModK(Ab)ℵ0 B C (4.1.1) f g b B(cid:48) (cid:48) f a / A(cid:48) (cid:48) g c C(cid:48) in QCat. If a, b and c are fully faithful, then so is the induced functor φ : B×A C → B(cid:48) ×A(cid:48) C(cid:48) . Proof. By [Lur09, 3.3.3.2], an object of the fiber product B ×A C in QCat is determined by objects A ∈ A, B ∈ B, C ∈ C and equivalences f B (cid:39) A (cid:39) gC. Moreover, if D and D (cid:48) are objects of B×A C corresponding to such (A, B, C) and (A (cid:48) ), respectively, then we have (4.1.2) This also applies to B(cid:48) ×A(cid:48) C(cid:48) . If a, b and c are fully faithful, then the induced morphism )×mapA(A, A(cid:48)) mapC(C, C (cid:48) (cid:48) mapB×AC(D, D ) (cid:39) mapB(B, B (cid:48) (cid:48) , C (cid:48) , B ). (cid:48) mapB(B, B )×mapA(A, A(cid:48)) mapC(C, C (cid:48) ) → mapB(cid:48)(bB, bB (cid:48) )×mapA(cid:48) (aA, aA(cid:48)) mapC(cid:48)(cC, cC (cid:48) ) is the fiber product of three equivalences in Spc, hence an equivalence itself. 24 / o o / o o (cid:48) Proposition 4.2. If f : B → A and g : C → A are t-exact functors between U-small stable qcategories equipped with t-structures, then D := B×A C admits a t-structure with respect to which the canonical functors g (cid:48) Proof. By [Lur14, 1.1.4.2], f full subqcategory spanned by the objects D such that g D>0 ⊆ D analogously. If these define a t-structure, then f Since f functor Ω preserves D>0. are morphisms of QCatEx. Define D≤0 ⊆ D to be the D ∈ C≤0, and define are necessarily t-exact. are exact, the suspension functor Σ preserves D≤0 and the loop space We claim that, if X ∈ D≤0 and Y ∈ D>0, then π0 mapD(X, Y) = 0. As in (4.1.2), we have : D → B and f and g : D → C are t-exact. D ∈ B≤0 and f and g and g (cid:48) (cid:48) (cid:48) (cid:48) (cid:48) (cid:48) (cid:48) (cid:48) mapD(X, Y) (cid:39) mapB(g (cid:48) (cid:48) X, g Y)×mapA(f g (cid:48)X, f g (cid:48)Y) mapC(f (cid:48) X, f (cid:48) Y). This homotopy fiber product in Set∆ induces an exact sequence of homotopy groups (cid:48) (cid:48) (4.2.1) π1 mapA(f g Y)⊕π0 mapC(f with base points given by the zero morphisms. The last term vanishes since g Y ∈ B>0, f (cid:48) g Y) → π0 mapD(X, Y) → π0 mapB(g X ∈ C≤0 and f X, f g X, g (cid:48) (cid:48) (cid:48) (cid:48) (cid:48) (cid:48) X, f Y) X ∈ B≤0, (cid:48) Y ∈ C>0; the first is isomorphic to (cid:48) Y) (cid:39) π0 mapA(Σf g X, f g (cid:48) (cid:48) Y) = 0, X, f g π0ΩmapA(f g (cid:48) Y ∈ A>0. (cid:48) X ∈ A≤0 and f g (cid:48) (cid:48) (cid:48) (cid:48) (cid:48) : f (cid:48) X and η X → t>0g ≤0 of η to t X, respectively, so X (cid:48) and g X and t (cid:48) (cid:48) are exact and send η to η ≤0 ∈ D≤0, as required. ≤0g since Σf g ≤0 → X → X>0 in D with Let X ∈ D. We claim that there exists a fiber sequence X ≤0 ∈ D≤0 and X>0 ∈ D>0. Consider the functors δ : ∆1 → B and δ : ∆1 → C corresponding (cid:48) X X → t>0f (cid:48) X, respectively. We have to the canonical morphisms η : g a homotopy f δ (cid:39) gδ (cid:48) since f and g are t-exact. By the universal property of the fiber induce a functor ∆1 → D corresponding to a morphism η : X → X>0. By (cid:48) product D, δ and δ construction, X>0 ∈ D>0. Since f and η, respectively, they (cid:48) ≤0f send the fiber X Proposition 4.3. Consider a commutative diagram (4.1.1) in which f are t-exact functors between U-small stable qcategories equipped with t-structures and a, b and c are the fully faithful inclusions of the hearts of these t-structures. If f is an isofibration, then the heart is equivalent to the fiber product B×A C, formed in the of the t-structure of 4.2 on B(cid:48) ×A(cid:48) C(cid:48) 1-category Cat of U-small categories. Proof. Recall that an isofibration is a functor of categories F : D → D(cid:48) such that, for each D ∈ D and each isomorphism α : FD ∼→ D (cid:48) such that Fβ = α. By [Joy08, 2.8], the nerve functor N : Cat → Set∆ is right Quillen with respect to the model structure on Cat whose weak equivalences and fibrations are the equivalences and the isofibrations, respectively, and the Joyal model structure on Set∆ ([Lur09, 2.2.5.1]). Each object is fibrant in this model structure on Cat, so the model structure is right proper. , there exists an isomorphism β : D ∼→ D (cid:48) in D(cid:48) and g (cid:48) (cid:48) 25 ♥ ) and χ1 : ∆0 → C(cid:48) In particular, pullbacks along isofibrations are homotopy pullbacks ([Bar10, 1.19]). Right Quillen functors between U-combinatorial model categories induce U-continuous functors between the underlying qcategories ([Lur14, 1.3.4.26]), so N sends the fiber product of B and C over A in Cat to their fiber product in QCat. By 4.1, it follows that the induced functor B×A C → B(cid:48) ×A(cid:48) C(cid:48) is fully faithful. Its essential image is contained in (B(cid:48) ×A(cid:48) C(cid:48) ♥ ) , so it remains to show that the essential image contains each X ∈ (B(cid:48) ×A(cid:48) C(cid:48) . Such an object X is classified by a functor χ : ∆0 → B(cid:48) ×A(cid:48) C(cid:48) , which is in turn classified by functors χ0 : ∆0 → B(cid:48) χ0 (cid:39) g (cid:48) (cid:48) χ1. By construction of the t- and a homotopy f structure on B(cid:48) ×A(cid:48) C(cid:48) , χ0 and χ1 must factor through B = B(cid:48)♥ and C = C(cid:48)♥ , respectively, so χ factors through B×A C, as desired. ×ModK(Ab)ℵ0 Lemma 4.4. The natural functor MHSp Λ equivalence. Proof. Following the discussion in the proof of 4.3, it suffices to show that we have the → ModK(Ab)ℵ0 corresponding equivalence in Cat and that the fiber functor ω : MHSp K sending polarizable mixed Hodge K-structure to its underlying K-module is an isofibration. The former is true by definition of MHSp Λ: its objects are given by pairs (M, V) such that M is a Λ-module of finite type and V is an object of MHSp K whose underlying K-module is M ⊗Λ K, and similarly for morphisms. Also, we can push polarizable mixed Hodge K-structures forward along isomorphisms of K-modules to isomorphic mixed Hodge K-structures, so ω is an isofibration. Λ) ∼→(cid:101)Db(MHSp Theorem 4.5. There are natural equivalences Λ)) ∼→ D(Ind(MHSp K))×D(ModK(Ab)) D(ModΛ(Ab)). (i) φ : Db(MHSp (ii) Φ : D(Ind(MHSp K)×Db(ModK(Ab)ℵ0 ) ModΛ(Ab)ℵ0 in QCat is an Db(ModΛ(Ab)ℵ0 ), Λ) := Db(MHSp → MHSp K Proof. The functors in question are those induced by the universal properties of the fiber products defining the codomains. They are thus exact by [Lur14, 1.1.4.2]. The functor φ is t-exact with respect to the natural t-structure on its domain and the one on its codomain induced by 4.2 and the natural t-structures on Db(MHSp K), Db(ModK(Ab)ℵ0 ) and Db(ModΛ(Ab)ℵ0 ). By 4.3 and 4.4, φ induces an equivalence between the hearts of these t-structures. As Λ is of global dimension ≤ 1 (4.0), hdim(ModΛ(Ab)ℵ0 ) = 1 (1.5(ii)); hdim(MHSp ) = 0. Using the exact sequence of (4.2.1), these homological-dimension computations imply that K) = 1 by [PS08, 3.35]; and hdim(ModK(Ab)ℵ0 π0 map(cid:101)Db(MHSp for each r ∈ Z≥2 and each pair of objects M and N of(cid:101)Db(MHSp . By the argument of [Wil12, 2], which develops a remark of [DG05, p.3] on the proof of [Buc60, 4.2], φ is fully ♥ Λ) Λ) (M, N[r]) = 0 26 faithful. Since the t-structures on Db(MHSp hence also(cid:101)Db(MHSp containing the heart is(cid:101)Db(MHSp K), Db(ModΛ(Ab)ℵ0 Λ), are all bounded, the smallest stable subqcategory of(cid:101)Db(MHSp (cid:16) Λ) itself, and it follows that φ is essentially surjective. ) and Db(ModK(Ab)ℵ0 ), and Λ) By [Lur14, 5.3.2.11(3)], the canonical functor (cid:17) Ind Db(MHSp K)×Db(ModK(Ab)ℵ0 ) Db(ModΛ(Ab)ℵ0 ) Ind(Db(MHSp K))×Ind(Db(ModK(Ab)ℵ0 )) Ind(Db(ModΛ(Ab)ℵ0 )) ψ is an equivalence. By [Dre15, 4.6], there is a canonical equivalence Ind(Db(A)) (cid:39) D(Ind(A)) for each U-small Noetherian Abelian category A such that hdim(A) < ∞. Under these canonical equivalences, Φ is equivalent to ψ Ind(φ). Lemma 4.6. If C⊗ ⊗-dualizable, then there is a symmetric monoidal equivalence Cop,⊗ ∼→ C⊗ is a U-small stable symmetric monoidal qcategory such that each V ∈ C is is the , where Cop,⊗ symmetric monoidal structure of [Lur14, 2.4.2.7]. Proof. Let C⊗rev denote the “reverse” symmetric monoidal structure on C given informally by V ⊗rev W := W ⊗ V. Specifically, we define C⊗rev as the pullback of the coCartesian fibration C⊗ → E⊗∞ along the reversal involution E⊗∞ → E⊗∞ of [Lur14, 5.2.5.25]. This reversal involution is homotopic to the identity, since E⊗∞ (cid:39) Fin∗ is the final object of the qcategory of (∞,1)-operads, from which we deduce that C⊗rev is equivalent to C⊗ Applying [Lur14, 5.2.5.27] in the case k = ∞, i.e., taking the limit of the functors (CPair) constructed therein as k approaches ∞, we obtain a pairing × (QCat ⊗ → C⊗×Fin∗ C⊗rev. By [Lur14, : Cop,⊗ → C⊗rev, which is given in- := morC(V, 1C), where morC(−, −) is the internal hom-object bifunctor. is in fact a symmetric monoidal AlgEk of symmetric monoidal qcategories ([Lur14, 5.2.2.21]) Dual(C) 5.2.2.25], this pairing induces a morphism of (∞,1)-operads (−) we refer to as the duality functor. The functor (−) ∨ formally by V (cid:55)→ V Since each object of C is ⊗-dualizable by hypothesis, (−) ∨,⊗ ∨ → (W⊗ V) ∨ ⊗rev W ∨ functor: the natural morphism V is an equivalence. The functor underlying (−) ∨,⊗ is an equivalence. Indeed, essential surjectivity follows from the observation that V (cid:39) (V for each V ∈ C and full faithfulness from the fact that ∨ ∨ ) : Cop → C underlying (−) ) → AlgEk ∨,⊗ ∨,⊗ ∨ . mapCop(V, W) → mapC(V ∨ ∨ , W ) (cid:39) mapC(W, V) is an equivalence for each (V, W) ∈ C2. By [Lur14, 2.1.3.8], as the underlying functor (−) ∨ with the above equivalence C⊗rev (cid:39) C⊗ is an equivalence, so is (−) completes the proof. . Composing (−) ∨,⊗ ∨,⊗ 27 ⊗ ⊗ Betti : (Smft such that the induced functor ΓBetti : (Smft /κ)op,(cid:113) → Theorem 4.7 ([Ayo10, §1]). There is a symmetric monoidal functor Γ /κ)op → CAlg(D(ModΛ(Ab)) ⊗ D(ModΛ(Ab)) ) ([Lur14, 2.4.3.18]) sends each X ∈ Smft /κ to an E∞-algebra ΓBetti(X) naturally equivalent to the Λ-linear singular cochain complex of X(C)an equipped with the E∞-algebra structure corre- sponding to the cup product. ⊗ Proof. Here, D(ModΛ(Ab)) denotes the stable locally U-presentable symmetric monoidal qcategory underlying the projective model structure on Cpx(ModΛ(Ab)) ([Hov99, 4.2.13]). ⊗ Let DA(κ, Λ) denote the stable locally U-presentable symmetric monoidal qcategory un- derlying the U-combinatorial symmetric monoidal model category used to define DA(S, Λ) in [Ayo11, 3.3] for S = Spec(κ). This is a Λ-linear variant of the construction of the P1- stable A1-homotopy category SH(κ) of [Voe98, 5.7]. This category is equipped with a canonical symmetric monoidal functor Σ whose essential im- age is spanned by ℵ0-presentable objects ([Rio05, 1.3]). From the symmetric monoidal Quillen equivalences constructed in [Ayo10, §1], we deduce a U-cocontinuous symmet- ric monoidal functor . By the results of [Ayo10, §3], preserves ℵ0-presentable objects and, at the level of underlying homotopy categories, ∗ ∗ is compatible with Grothendieck’s six operations. By [Rio05, 1.4, 2.2], since κ admits ⊆ DA(κ, Λ) is equal resolution of singularities ([Hir64]), the full subqcategory DA(κ, Λ)ℵ0 to the full subqcategory spanned by the ⊗-dualizable objects. Combining these results with 4.6, we obtain a symmetric monoidal functor Γ ⊗ → D(ModΛ(Ab)) × → DA(κ, Λ) ∞ T : (Smft /κ) : DA(κ, Λ) ∗,⊗ ⊗ ⊗ ⊗ ⊗ Betti given by the composite ∗,⊗−−−→ D(ModΛ(Ab)) ⊗ ⊗ . ∞,op ∗ (Smft /κ)op,(cid:113) Σ )op,⊗ ∼→ DA(κ, Λ) T−−−−−→ (DA(κ, Λ)ℵ0 ∗ Unwinding the constructions and using the compatibility of with Grothendieck’s six operations, one finds that ΓBetti(X) (cid:39) f∗f ΛX(C)an, where f : X → Spec(κ) is the structural ∗ morphism and ΛX(C)an is the constant sheaf associated to Λ on X(C)an. Since f∗f ΛX(C)an and the Λ-linear singular cochain complex of X(C)an are naturally equivalent as E∞-algebras when the latter is equipped with the cup product, the claim follows. ⊗ Definition 4.8. Let D(Ind(MHSp Λ)) ⊗ ×D(ModK(Ab)) D(Ind(MHSp K)) (4.8.1) in CAlg(PrL,⊗ ), the qcategory of commutative algebras in the symmetric monoidal qcate- gory of locally U-presentable qcategories and U-cocontinuous functors ([Lur14, 4.8.1.14]). The functor C⊗ (cid:55)→ CAlg(C⊗ ) preserves fiber products, so, by 4.5, the qcategory underly- ing D(Ind(MHSp Λ)). This circuitous construction of the symmetric monoidal structure on the derived category of Ind(MHSp Λ) facilitates the proof of 4.9. Given any other more direct construction of this symmetric monoidal structure, it should not be difficult to show that it is equivalent to the above using the universal property of the fiber product (4.8.1). is equivalent to D(Ind(MHSp denote the fiber product ⊗ D(ModΛ(Ab)) Λ)) ⊗ ⊗ 28 Corollary 4.9. There is a symmetric monoidal functor ⊗ Hdg,Λ : (Schft /κ)op,(cid:113) → D(Ind(MHSp Λ)) ⊗ Γ such that, for each X ∈ Schft /κ: (i) hrΓHdg,Λ(X) is naturally isomorphic to Hr Λ-structure of [Del74, 8.2.1] for each r ∈ Z; and Betti(X, Λ) equipped with the mixed Hodge (ii) when equipped with the E∞-algebra structure induced by [Lur14, 2.4.3.18], the complex of Λ-modules underlying ΓHdg,Λ(X) is naturally equivalent to the Λ-linear singular cochain complex of X(C)an with the E∞-algebra structure given by the cup product. K)) Proof. By 3.5 and the basic properties of Betti cohomology cited in the proof of 3.6, ⊗ the symmetric monoidal functor Γ Betti of 4.7 extends to a symmetric monoidal functor /κ)op,(cid:113) → D(ModΛ(Ab)) ⊗ ⊗ (Schft Betti, computing the Λ-linear Betti , abusively denoted by Γ cohomology of each X ∈ Schft /κ. After composition with the evident symmetric monoidal ⊗ → D(ModK(Ab)) functors D(Ind(MHSp , re- ⊗ spectively, the symmetric monoidal functors Γ Betti become equivalent: both compute K-linear Betti cohomology. The claim now follows from the universal property of the fiber product (4.8.1). Remark 4.10. As mentioned in the introduction, from the symmetric monoidal functor ⊗ Hdg,Λ, one may construct a motivic E∞-ring spectrum representing Λ-linear absolute Γ Hodge cohomology ([Bei86, §5]). This observation will be exploited and explored further in a forthcoming preprint. and D(ModΛ(Ab)) ⊗ Hdg of 3.6 and Γ ⊗ → D(ModK(Ab)) ⊗ ⊗ References [AR94] [Ayo07a] [Ayo07b] [Ayo10] [Ayo11] Jiří Adámek and Jiří Rosický. Locally presentable and accessible categories, volume 189 of London Mathematical Society Lecture Note Series. Cambridge University Press, Cambridge, 1994. Joseph Ayoub. Les six opérations de Grothendieck et le formalisme des cycles évanescents dans le monde motivique. I. Astérisque, (314):x+466 pp. (2008), 2007. Joseph Ayoub. Les six opérations de Grothendieck et le formalisme des cycles évanescents dans le monde motivique. II. Astérisque, (315):vi+364 pp. (2008), 2007. Joseph Ayoub. Note sur les opérations de Grothendieck et la réalisation de Betti. J. Inst. Math. Jussieu, 9(2):225–263, 2010. Joseph Ayoub. La réalisation étale et les opérations de Grothendieck. http://www.math.uiuc.edu/K-theory/0989/, January 2011. 29 [Bar10] [Bei86] [Bek00] [Bla01] Clark Barwick. On left and right model categories and left and right Bousfield localizations. Homology, Homotopy and Applications, 12(2):245–320, 2010. Alexander Beilinson. Notes on absolute Hodge cohomology. In Applications of algebraic K-theory to algebraic geometry and number theory, Part I, II (Boulder, Colorado, 1983), volume 55 of Contemporary Mathematics, pages 35–68. American Mathematical Society, Providence, RI, 1986. Tibor Beke. Sheafifiable homotopy model categories. Math. Proc. Cambridge Philos. Soc., 129(3):447–475, 2000. Benjamin A. Blander. Local projective model structures on simplicial presheaves. K-Theory, 24(3):283–301, 2001. [BNT15] Ulrich Bunke, Thomas Nikolaus, and Georg Tamme. The Beilinson regulator is a map of ring spectra. http://arxiv.org/abs/1509.05667, September 2015. [Buc60] David A. Buchsbaum. Satellites and universal functors. Ann. of Math. (2), [CG14] [Con07] [Del71] [Del74] [Del90] [DG05] [DM89] [Dre15] 71:199–209, 1960. Joana Cirici and Francisco Guillén. E1-formality of complex algebraic varieties. Algebr. Geom. Topol., 14(5):3049–3079, 2014. Brian Conrad. Deligne’s notes on Nagata compactifications. J. Ramanujan Math. Soc., 22(3):205–257, 2007. Pierre Deligne. Théorie de Hodge. II. Inst. Hautes Études Sci. Publ. Math., (40):5–57, 1971. Pierre Deligne. Théorie de Hodge. III. Inst. Hautes Études Sci. Publ. Math., (44):5–77, 1974. Pierre Deligne. Catégories tannakiennes. In The Grothendieck Festschrift, Vol. II, volume 87 of Progr. Math., pages 111–195. Birkhäuser Boston, Boston, MA, 1990. Pierre Deligne and Alexander B. Goncharov. Groupes fondamentaux motiviques de Tate mixte. Ann. Sci. École Norm. Sup. (4), 38(1):1–56, 2005. Pierre Deligne and James S. Milne. Tannakian Categories. In Hodge Cycles, Motives, and Shimura Varieties, number n Springer, 1989. Brad Drew. Verdier quotients of stable quasi-categories are localizations. Preprint, available at http://www-bcf.usc.edu/~bdrew/, November 2015. 900 in Lecture Notes in Mathematics. ◦ [GNA02] Francisco Guillén and Vicente Navarro Aznar. Un critère d’extension des foncteurs définis sur les schémas lisses. Publ. Math. Inst. Hautes Études Sci., (95):1–91, 2002. [Hir64] Heisuke Hironaka. Resolution of singularities of an algebraic variety over a field of characteristic zero. I, II. Ann. of Math. (2) 79 (1964), 109–203; ibid. (2), 79:205–326, 1964. 30 [Hov99] Mark Hovey. Model categories, volume 63 of Mathematical Surveys and Monographs. American Mathematical Society, Providence, RI, 1999. [Hub95] Annette Huber. Mixed motives and their realization in derived categories, volume 1604 of Lecture Notes in Mathematics. Springer-Verlag, Berlin, 1995. [Hub00] Annette Huber. Realization of Voevodsky’s motives. J. Algebraic Geom., 9(4):755–799, 2000. [Hub04] Annette Huber. Corrigendum to: “Realization of Voevodsky’s motives” [J. Algebraic Geom. 9 (2000), no. 4, 755–799; mr1775312]. J. Algebraic Geom., 13(1):195–207, 2004. Florian Ivorra. Mixed Hodge complexes on algebraic varieties and t-structures. J. Algebra, 433:107–167, 2015. André Joyal. Notes on quasi-categories. Preprint, available at http://www.math.uchicago.edu/~may/IMA/Joyal.pdf, 2008. [Ivo15] [Joy08] [Lur09] [Lur14] [Lur11] [Lev98] Marc Levine. Mixed motives, volume 57 of Mathematical Surveys and Monographs. American Mathematical Society, Providence, RI, 1998. Jacob Lurie. Higher topos theory, volume 170 of Annals of Mathematics Studies. Princeton University Press, Princeton, NJ, 2009. Jacob Lurie. Derived Algebraic Geometry VIII: Quasi-Coherent Sheaves and Tannaka Duality Theorems. http://www.math.harvard.edu/~lurie/papers/DAG-VIII.pdf, 2011. Jacob Lurie. Higher Algebra. Preprint, available at http://www.math.harvard.edu/~lurie/papers/higheralgebra.pdf, September 2014. Florence Lecomte and Nathalie Wach. Réalisation de Hodge des motifs de Voevodsky. Manuscripta Math., 141(3-4):663–697, 2013. Vicente Navarro Aznar. Sur la théorie de Hodge-Deligne. Invent. Math., 90(1):11–76, 1987. Alberto Navarro Garmendia. The Riemann-Roch theorem and Gysin morphism in arithmetic geometry. PhD thesis, 2016. In preparation. Jonathan P. Pridham. Chris A. M. Peters and Joseph H. M. Steenbrink. Mixed Hodge structures, volume 52 of Ergebnisse der Mathematik und ihrer Grenzgebiete. 3. Folge. A Series of Modern Surveys in Mathematics [Results in Mathematics and Related Areas. 3rd Series. A Series of Modern Surveys in Mathematics]. Springer-Verlag, Berlin, 2008. Joël Riou. Dualité de Spanier-Whitehead en géométrie algébrique. C. R. Math. Acad. Sci. Paris, 340(6):431–436, 2005. [Pri] [PS08] [LW13] [NA87] [NG16] [Rio05] 31 [Rob15] Marco Robalo. K-theory and the bridge from motives to noncommutative motives. Adv. Math., 269:399–550, 2015. [Sai88] Morihiko Saito. Modules de Hodge polarisables. Publ. Res. Inst. Math. Sci., 24(6):849–995 (1989), 1988. [Sai90] Morihiko Saito. Mixed Hodge modules. Publ. Res. Inst. Math. Sci., 26(2):221–333, 1990. [SGA72a] Théorie des topos et cohomologie étale des schémas. Tome 1: Théorie des topos. Lecture Notes in Mathematics, Vol. 269. Springer-Verlag, Berlin, 1972. Séminaire de Géométrie Algébrique du Bois-Marie 1963–1964 (SGA 4), Dirigé par M. Artin, A. Grothendieck, et J. L. Verdier. Avec la collaboration de N. Bourbaki, P. Deligne et B. Saint-Donat. [SGA72b] Théorie des topos et cohomologie étale des schémas. Tome 2. Lecture Notes in Mathematics, Vol. 270. Springer-Verlag, Berlin, 1972. Séminaire de Géométrie Algébrique du Bois-Marie 1963–1964 (SGA 4), Dirigé par M. Artin, A. Grothendieck et J. L. Verdier. Avec la collaboration de N. Bourbaki, P. Deligne et B. Saint-Donat. Neantro Saavedra Rivano. Catégories Tannakiennes. Lecture Notes in Mathematics, Vol. 265. Springer-Verlag, Berlin, 1972. Stefan Schwede and Brooke E. Shipley. Algebras and modules in monoidal model categories. Proc. London Math. Soc. (3), 80(2):491–511, 2000. [SR72] [SS00] [SW15] Wolfgang Soergel and Matthias Wendt. Perverse motives and graded derived [Voe98] [Voe10] category O. http://arxiv.org/abs/1404.6333, April 2015. Vladimir Voevodsky. A1-homotopy theory. In Proceedings of the International Congress of Mathematicians, Vol. I (Berlin, 1998), number Extra Vol. I, pages 579–604 (electronic), 1998. Vladimir Voevodsky. Unstable motivic homotopy categories in Nisnevich and cdh-topologies. J. Pure Appl. Algebra, 214(8):1399–1406, 2010. [Whi14] David White. Model structures on commutative monoids in general model [Wil12] categories. http://arxiv.org/abs/1403.6759/, March 2014. Jörg Wildeshaus. Erratum à la note “f -catégories, tours et motifs de Tate” [C. R. Acad. Sci. Paris, Ser. I 347 (2009) 1137–1342], avec un appendice de Q. Liu et de J. Vitória [mr2588777]. C. R. Math. Acad. Sci. Paris, 350(3-4):129–131, 2012. 32
1701.06864	1	1701	2017-01-24T13:36:03	3x3 Singular Matrices of Linear Forms	[ "math.AG" ]	We determine the irreducible components of the space of 3x3 matrices of linear forms with vanishing determinant. We show that there are four irreducible components and we identify them concretely. In particular, under elementary row and column operations with constant coefficients, a 3x3 matrix with vanishing determinant is equivalent to one of the following four: a matrix with a zero row, a zero column, a zero 2x2 square or an antisymmetric matrix.	math.AG	math	3 × 3 SINGULAR MATRICES OF LINEAR FORMS DAMIANO TESTA Abstract. We determine the irreducible components of the space of 3×3 matrices of linear forms with vanishing determinant. We show that there are four irreducible components and we identify them concretely. In particular, under elementary row and column operations with constant coefficients, a 3 × 3 matrix with vanishing determinant is equivalent to one of the following four: a matrix with a zero row, a zero column, a zero 2 × 2 square or an antisymmetric matrix. Introduction If X is a square matrix over a field and the determinant of X vanishes, it follows that there is a non-zero linear combination of the rows of X that vanishes. This is no longer true if the matrix is a matrix of linear forms and we only allow linear combinations by constants. For instance, letting x1, . . . , x5 be independent variables, no non-trivial constant linear combination of the rows of the matrices (1) (cid:18)x1 x1 x2 x2(cid:19) x1 x2 x3 0 x4 0 x5 0 0     vanishes, although both matrices have zero determinant. In the case of the first matrix in (1), the difference of the columns vanishes, while in the case of the second matrix there is also no non-trivial linear combination of the columns that vanishes. Thus, we have identified three types of 3 × 3 matrices with vanishing determinants: matrices with a zero row, matrices with a zero column and matrices with a zero 2 × 2 square. Up to multiplication on the left and on the right by an invertible 3 × 3 matrix of constants, there is one more kind of matrix that has vanishing determinant (Theorem 7) and is not of one of the three types above: antisymmetric matrices (see Table 1). ⋆ ⋆ ⋆ ⋆ ⋆ ⋆ 0 0 0     ⋆ ⋆ ⋆ 0 ⋆ 0 0 ⋆ 0     Zero row Zero square Antisymmetric 0 x1 −x2 0 x3 0 −x1 x2 −x3     0 ⋆ ⋆ 0 ⋆ ⋆ 0 ⋆ ⋆     Zero column Table 1. Matrices with vanishing determinant For our purposes, the number of variables for the linear forms appearing in the matrices is not especially relevant. Nevertheless, the effective number of variables is at most 9, as we Date: March 13, 2018. 2010 Mathematics Subject Classification. 14M12, 15A54, 13C40. Key words and phrases. Determinantal varieties, determinantal ideals, matrices of linear forms. 1 can always replace the initial variables with a basis for the span of the entries of the 3 × 3 matrix in question. As a consequence of Theorem 7, if the dimension of the span of the entries of the matrix is at least 7, then the determinant cannot vanish identically. Thus, in the interesting cases, the number of variables could be limited to 6. In a broader context, for m, n, r positive integers, the loci of m×n matrices with entries in a field and rank at most r are called determinantal varieties, see for instance [Har95, Lecture 9]. The extension of the definition of determinantal varieties to the case in which the entries of the matrix are linear forms is especially relevant in the context of rational normal scrolls and minimal varieties: all of this is very classical and the paper [EH87] is a good introduction to the topic, its history and relevant results. Eisenbud introduces in [Eis88] the property of 1- genericity for spaces of matrices: a sort of non-degeneracy for the entries of a matrix of linear forms. One of his aims is to obtain conditions under which sections of small codimension of certain spaces of matrices inherit the irreducibility property. His ideas have later been applied, developed and extended, see for instance [Eis07, Kat08, Rai12, Rai13, KPU17]. Our work starts with similar questions, but we do not limit the entries of our matrices. First, we do not impose any genericity condition: we study 3 × 3 matrices of linear forms with no restrictions. Second, we find several irreducible components for our schemes and, in fact, we are able to determine all of them. Finally, this leads to questions regarding a deeper understanding of the irreducible components of the schemes we introduce and their ideals. 1. Notation and preliminary results Let k be a field and let n be a positive integer. We denote • the algebra of polynomials in x1, . . . , xn with coefficients in k by R = k[x1, . . . , xn], • the field of fractions of R by K = k(x1, . . . , xn). Let M be a matrix with entries in R. The matrix M is homogeneous if the entries of M are homogeneous forms of the same degree; the matrix M is primitive if the only polynomials in R dividing all the entries of M are constants; the degree of M is d = deg M if the maximum of the degrees of the entries of M is d. In the cases of interest below, the entries of M will in fact be homogeneous of the same degree; we state our results for not necessarily homogeneous matrices: the statements in the homogeneous case follow at once observing that if the product of two polynomials is homogeneous and non-zero, then the two polynomials are themselves homogeneous. A vector is a matrix consisting of a single column; a row vector is a matrix consisting of a single row. We begin with an easy lemma on matrices of rank 1 and entries in R. Lemma 1. Let a, b be positive integers and let X be an a × b matrix with entries in R. If the rank of X is 1, then there exist two non-zero column vectors u ∈ Ra and v ∈ Rb such that the identity X = uvt holds. The vectors u and v satisfy deg u + deg v = deg X and if the matrix X is homogeneous, then they are also homogeneous. Proof. Denote by u′ a non-zero column of X; let u ∈ R denote the greatest common divisor of all the entries of the vector u′ and let u be the primitive column vector in Ra such that u′ = uu. Since the rank of the matrix X is one, all the columns of X are proportional over K to the vector u; since the entries of the matrix X are contained in R and the vector u is primitive, all the columns of X are obtained from u by multiplication by an element of R. Denote by v = (v1 · · · vb)t ∈ Rb the column vector such that, for i ∈ {1, . . . , b}, the 2 i-th entry of v is the factor vi ∈ R with the property that viu is the i-th column of X. By construction the identity X = uvt holds. The assertion about homogeneous matrices is immediate. (cid:3) The next result is a direct consequence of Cramer's rule. Recall that for a square matrix M, the adjoint of M is the matrix adj(M) whose entry in position (i, j) is (−1)i+j times the determinant of the matrix obtained from M by removing the i-th column and the j-th row. Lemma 2. Let d, r be non-negative integers and let X be an (r + 1) × (r + 1) matrix of rank r and degree d with entries in R. There exist two non-zero vectors u and v with entries in R such that the identities Xu = (X t)v = 0 and deg u + deg v ≤ dr hold. If X is homogeneous, then the vectors u and v can be chosen to be homogeneous and satisfy the equality deg u + deg v = dr. Proof. Let adj(X) denote the adjoint of the matrix X so that the identities X adj(X) = adj(X)X = det(X) Id = 0 hold. The matrix adj(X) has degree at most dr; if X is homogeneous of degree d, then adj(X) is homogeneous of degree dr. Since the rank of X is r, it follows that the matrix adj(X) is non-zero; moreover the columns of adj(X) lie in the kernel of X and hence are proportional over the fraction field K of R, because the kernel of X is one-dimensional. Thus the rank of the matrix adj(X) is one and we can apply Lemma 1 to the matrix let u ∈ Rm and v ∈ Rn be vectors such that the equalities adj(X) = uvt and adj(X): deg u + deg v = deg adj(X) ≤ dr hold. Finally, the identity Xuvt = 0 shows that the vector Xu is the zero vector, and similarly the row vector vtX is the zero vector, completing the proof. (cid:3) To state the next lemma, we introduce a little notation. Denote by P the projective space of dimension n − 1 whose homogeneous coordinate ring is R. Let r be a non-negative integer and let ℓ = (ℓ1, . . . , ℓr) ∈ Rr be a sequence of r linear forms in R. Denote by L ⊂ P the linear subspace of P defined by the vanishing of the forms in ℓ and by c the codimension of L in P. Denote by IL the ideal sheaf of L ; the sequence ℓ defines a surjective homomorphism −→ IL , sending (f1, . . . , fr) to P fiℓi. Twisting by OP(2) and r,c as the kernel of the homomorphism r,c defined above is (r − c)n +(cid:0)c 2(cid:1). taking global sections, we define the vector space V n of sheaves (cid:0)OP(−1)(cid:1)⊕r H0(cid:0)P, OP(1)⊕r(cid:1) −→ H0(cid:0)P, IL (2)(cid:1). Lemma 3. The dimension of the vector space V n Proof. We will use the diagram 0 ↑ (2) 0 −→ IL (2) −→ OP(2) −→ OL (2) −→ 0 ↑ ℓ OP(1)⊕r of sheaves on P. Observe that diagram (2) is exact and stays exact if we take global sections. By definition, the vector space V n r,c is the kernel of the homomorphism H0(cid:0)P, OP(1)⊕r(cid:1) −→ H0(cid:0)P, IL (2)(cid:1). 3 Standard computations yield the equalities 2(cid:19) dim H0(cid:0)P, IL (2)(cid:1) = cn −(cid:18)c dim H0(cid:0)P, OP(1)⊕r(cid:1) = rn r,c = (r − c)n +(cid:0)c 2(cid:1). and, taking differences, we find dim V n (cid:3) Lemma 4. Let ℓ = (ℓ1, ℓ2, ℓ3) be a triple of linear forms and let Vℓ be the k-vector space of triples of linear forms (f1, f2, f3) such that the identity f1ℓ1 + f2ℓ2 + f3ℓ3 = 0 holds. (1) If the linear forms ℓ1, ℓ2, ℓ3 are k-linearly independent, then Vℓ is spanned by the triples (0, ℓ3, −ℓ2), (−ℓ3, 0, ℓ1), (ℓ2, −ℓ1, 0). (2) If the linear forms ℓ1, ℓ2 are k-linearly independent and ϕ1ℓ1 + ϕ2ℓ2 + ℓ3 = 0 is a k-linear relation, then Vℓ is spanned by the triples (ℓ2, −ℓ1, 0) and ℓ(ϕ1, ϕ2, 1), as ℓ varies among all linear forms in R. Proof. We begin with some easy checks: • the stated triples satisfy the identity f1ℓ1 + f2ℓ2 + f3ℓ3 = 0; • if item (1) holds, the dimension of the span of the three vectors (0, ℓ3, −ℓ2), (−ℓ3, 0, ℓ1), (ℓ2, −ℓ1, 0) is 3; • if item (2) holds, the dimension of the span of the n + 1 vectors (ℓ2, −ℓ1, 0) and ℓ(ϕ1, ϕ2, 1), as ℓ ranges among the n variables of the ring R, is n + 1. To conclude it suffices to show that the dimension of the vector space Vℓ is 3 in case (1) and n + 1 in case (2). A moment's thought makes it clear that the vector space Vℓ mentioned in this lemma is the same vector space V n r,c defined in Lemma 3 in the case (r, c) = (3, 3), if the forms ℓ1, ℓ2, ℓ3 are independent (item (1)), and in the case (r, c) = (3, 2), if the forms ℓ1, ℓ2 are independent and ℓ1, ℓ2, ℓ3 are not (item (2)). Specializing the results of Lemma 3 we finally obtain dim V n (cid:3) 3,2 = n + 1, as required. 3,3 = 3 and dim V n 2. Main result Let Mn denote the k-vector space of 3 × 3 matrices whose entries are linear forms in R: the dimension of Mn is 9n. We denote the projective space associated to Mn by PMn, a projective space of dimension 9n − 1. By construction, the determinant induces a function det : Mn → R whose image is contained in the vector space R3 of forms of degree 3 in R. The dimension of the vector space R3 is r3 = (cid:0)n+2 det corresponds to the simultaneous vanishing of r3 cubic forms. 3 (cid:1), so that the vanishing of the function Definition 5. The space of 3 × 3 singular matrices of linear forms is the subscheme Fn of PMn defined by the vanishing of the function det. When the index n is clear from the context, we will omit it; for instance, we may denote the scheme Fn simply by F . The space F is our main interest. An immediate consequence of the definition of the space F is that multiplying a 3 × 3 matrix of linear forms by an invertible matrix with entries in k on the left or on the right stabilizes F and its complement. 4 Definition 6. Two f × g matrices X, Y with entries in R are f -equivalent if there are an invertible f × f matrix F and an invertible g × g matrix G both with entries in k such that Y = F XG. There is a straightforward relationship between Fn and linear subspaces of determinantal varieties. Let P8 denote the 8-dimensional projective space with coordinates {xij}i,j∈{1,2,3} and let X be the determinantal cubic with equation X : det  x11 x12 x13 x21 x22 x23 x31 x32 x33   = 0. A point in the variety Fn is a linear map of a vector space to the variety X. Equivalently, points of Fn correspond to (not necessarily injective) parameterizations of linear subspaces contained in X. Theorem 7. Let X be a 3 × 3 matrix of linear forms. If the determinant det(X) vanishes, then X is f -equivalent to one of the following: (1) a matrix with a zero row; (2) a matrix with a zero column; (3) a matrix with a zero square; (4) an antisymmetric matrix. Proof. If X is the zero matrix, then there is nothing to prove. If the rank of X is 1, then we can apply Lemma 1 to conclude that either the rows or the columns of X are k-proportional and we are in case (1) or (2). Thus, we reduce to the case in which the rank of the matrix X is 2 and we apply Lemma 2: let u, v be non-zero column vectors such that u is in the kernel of X, v is in the kernel of X t and deg u + deg v = 2. If one of the vectors u or v has degree 0, i.e. is constant, then the columns or the rows of X are k-proportional and we are in case (1) or (2). Therefore we further reduce to the case in which the vectors u and v have degree 1 and are not R- proportional to constant vectors: the entries of u span a k-vector space of dimension at least 2 of linear forms and similarly for the entries of v. The rows of the matrix X therefore are triples contained in the vector space Vu of Lemma 4 and, by our reductions, they span a three-dimensional linear subspace of Vu. If the entries u1, u2, u3 of u are k-linearly independent, then after multiplying the matrix X on the left by an invertible matrix with entries in k, we may assume that the rows of X are (0 u3 −u2), (−u3 0 u1), (u2 −u1 0): thus X is f -equivalent to the antisymmetric matrix 0   u3 −u2 0 u1 0  . −u3 u2 −u1 If the entries u1, u2, u3 of u are k-linearly dependent and ϕ1u1 + ϕ2u2 + ϕ3u3 = 0 is a non-trivial linear relation among them, then after multiplying the matrix X on the left by an invertible matrix with entries in k, we may assume that the last two of the rows of X are (ℓϕ1 ℓϕ2 ℓϕ3) and (mϕ1 mϕ2 mϕ3), where ℓ, m are linear forms. The matrix X 5   and hence, after a sequence of column (cid:3)   with a zero square.  ∈ PMn    ∈ PMn    ∈ PMn   dim Rn = 6n − 1; dim Sn = 5n − 1; dim An = 3n − 1; dim Cn = 6n − 1. is f -equivalent to the matrix   operations, also to the matrix   ⋆ ℓϕ2 ⋆ ⋆ ℓϕ1 ℓϕ3 mϕ1 mϕ2 mϕ3 ⋆ ⋆ ⋆ 0 0 ℓ m 0 0 We define four linear subspaces of PMn: ⋆ ⋆ ⋆ ⋆ ⋆ ⋆ 0 0 0 ⋆ ⋆ ⋆ ⋆ 0 0 ⋆ 0 0 • (zero row) • (zero square) Rn =     Sn =     • (antisymmetric) An =     Cn =     • (zero column) ϕ −χ 0 0 −ϕ ψ 0 χ −ψ ⋆ ⋆ 0 ⋆ ⋆ 0 ⋆ ⋆ 0  ∈ PMn   Let G = GL3 × GL3 be the group of ordered pairs of invertible 3 × 3 matrices. The group G acts on M by the rule (f, g) · m = f mgt and this action induces an analogous action on PM . The orbits under the G-action are precisely the f -equivalence classes of matrices. Denote • by Rn the closure of the G-orbit of Rn, • by Sn the closure of the G-orbit of Sn, • by An the closure of the G-orbit of An, • by Cn the closure of the G-orbit of Cn. Each of these schemes is clearly irreducible, being the orbit of a linear subspace under an irreducible group. As a consequence of Theorem 7, we know that each G-orbit of a point of F intersects at least one of the linear subspaces R, S, A, C: the set of k-rational points of F is the union of the four subvarieties R, S , A , C . In Section 3 we examine these four subvarieties of F . We shall see that the irreducible components of the scheme F are the varieties R, S , A , C . 3. The four components We analyze here the four components of the scheme Fn that we determined in Theorem 7. Throughout this section, we assume that the integer n satisfies the inequality n ≥ 2: the case n = 1 is essentially the case of 3 × 3 matrices over a field. We begin by determining the stabilizers in G of each of the four linear subspaces Rn, Sn, An, Cn. 6 The stabilizers of matrices with a zero row or column. Let (f, g) in G be a pair preserving the set of matrices R with last row equal to zero. Observe that we can multiply the second element of the pair, g, by any invertible matrix and the pair still preserves the set R. It is then immediate to check that the stabilizers of R and C are the pairs of invertible matrices of the form Stabilizer of R Stabilizer of C ⋆ ⋆ ⋆ ⋆ ⋆ ⋆ 0 0 ⋆       , g  ,  f,  ⋆ ⋆ ⋆ ⋆ ⋆ ⋆ 0 0 ⋆     for any f, g in GL3. The stabilizer of matrices with a zero square. Let (f, g) in G be a pair preserving the set of matrices S with a square of zeros. For i, j in {1, 2, 3}, let fij and gij denote the (i, j)-th entry of f and of g and let eij denote the matrix with (i, j)-th entry equal to 1 and remaining entries equal to 0. The product f eijgt is the rank one matrix that is the product of the i-th column of f times the j-th row of gt. Thus the condition that the two matrices f e11gt, f e12gt lie in S translates to the equations f21g21 = f21g31 = f21g22 = f21g32 = 0 f31g21 = f31g31 = f31g22 = f31g32 = 0. Since the matrix g is invertible, the square on the four entries g21, g31, g22, g32 of g cannot vanish and we deduce that the entries f21 and f31 of f vanish. Similarly, the entries g21 and g31 of g vanish as well and finally the stabilizer of S consists of the matrices of the form ⋆ ⋆ ⋆ 0 ⋆ ⋆ 0 ⋆ ⋆       ,  ⋆ ⋆ ⋆ 0 ⋆ ⋆ 0 ⋆ ⋆    .  The stabilizer of antisymmetric matrices. Let (f, g) in G be a pair preserving the set of antisymmetric matrices A; for s, t, u in {0, 1} define the antisymmetric matrix estu by estu =   s −t 0 −s 0 u t −u 0   . The matrices f e100gt, f e010gt, f e001gt are antisymmetric; in particular, their diagonal entries vanish. The implied equations force each row of the matrix g to be proportional to the corresponding row of f . Imposing now the further constraints arising from the matrices f e100gt, f e010gt, f e001gt being antisymmetric shows that the matrices f and g themselves are proportional; clearly, such matrices stabilize A and we find that the stabilizer of A consists of matrices of the form (f, λf ) for f ∈ GL3 and λ ∈ k∗. Having determined the stabilizers of the linear subspaces Rn, Sn, An, Cn we now compute the dimensions of the irreducible components of the space Fn of 3 × 3 singular matrices of 7 linear forms. Indeed, the dimension of the orbit under G of each linear subspace is the sum of the dimension of the linear subspace plus the codimension of their stabilizer: dim(Rn) = 6n + 1, dim(Sn) = 5n + 3, dim(An) = 3n + 7, dim(Cn) = 6n + 3. The special case n = 2. In the special case in which the number of variables of the ring R equals 2, the dimensions of the four subvarieties R2, A2, S2, C2 coincides and equals 13. An easy check shows that the subvarieties A2 and S2 coincide in this case, so that F2 is the union of three subvarieties: R2, A2 = S2, C2. A quick computation using the computer algebra package Magma shows that, as subvarieties of the projective space PM2 ≃ P17, the degree of R2 and of C2 is 15 and the degree of S2 = A2 is 51. We introduce 9n coordinate functions on Mn: for i, j ∈ {1, 2, 3} and ℓ ∈ {1, . . . , n}, let aℓ ij denote the coefficient of xℓ in the (i, j)-th entry of a matrix in Mn: a1 11x1 + a2 21x1 + a2 a1 31x1 + a2 a1 11x2 + · · · + an 21x2 + · · · + an 31x2 + · · · + an 11xn 21xn 31xn   12x1 + a2 a1 22x1 + a2 a1 32x1 + a2 a1 12x2 + · · · + an 22x2 + · · · + an 32x2 + · · · + an 22xn 32xn 12xn 13x1 + a2 a1 23x1 + a2 a1 33x1 + a2 a1 13x2 + · · · + an 23x2 + · · · + an 33x2 + · · · + an 23xn 33xn 13xn .   The orbits of matrices with a vanishing row or column. Let M be a matrix in Mn. If M lies in Rn, then it follows that the rows of M are linearly independent. Making use of the coordinates {aℓ ij} defined above, we form the 3n × 3 matrix A =   11 a1 11 a2 a1 21 a2 31 a2 a1 21 31 11 · · · an · · · an · · · an 21 31 a1 12 a1 22 a1 32 12 · · · an · · · an · · · an 22 32 a1 13 a1 23 a1 33 13 · · · an · · · an · · · an 23 33 .   Therefore, in terms of the associated matrix A, the matrix M lies in Rn if and only if the matrix A has rank at most 2. Hence, the (cid:0)3n 3(cid:1) determinants of the 3 × 3 submatrices of the matrix A vanish exactly on the subvariety Rn. Thus, the variety Rn is actually the determinantal variety of 3n × 3 matrices of rank at most 2. Of course, transposed assertions apply to the variety Cn. The orbits of antisymmetric matrices or of matrices with a zero square. We do not know a similar explicit description of the varieties An and Sn. We checked that the ideals of An and Sn agree with the ideal of Fn in all degrees up to 4 in a few examples. Concluding questions An immediate question that arises from our work is to study further the orbit Sn of the matrices with a square of zeros and the orbit An of the antisymmetric matrices. We would also like to present the following personal point of view on the initial question. Computing a determinant involves usually a fair number of operations. Moreover, if the calculation shows that the determinant vanishes, I usually try to find a "justification" of this fact that is independent of computing the determinant. Possible such justifications are 8 • some evident linear combination of the rows or columns of the matrix that vanishes (e.g., one row is the sum of two other rows); • the matrix is of odd order and is antisymmetric; • the matrix has "lots" of zeros, (e.g., a 3 × 3 matrix with a zero square). These are precisely the irreducible components of the schemes Fn! As the results in this paper handle the case of matrices of order 3, we are naturally lead to wonder what happens next. Question. Let n be a positive integer. What are the irreducible components of the spaces of n × n matrices of linear forms and identically vanishing determinant? References [Eis88] David Eisenbud, Linear sections of determinantal varieties, Amer. J. Math. 110 (1988), no. 3, 541–575. ↑2 [Eis07] , Syzygies, degrees, and choices from a life in mathematics. Retiring presidential address, Bull. Amer. Math. Soc. (N.S.) 44 (2007), no. 3, 331–359. ↑2 [EH87] David Eisenbud and Joe Harris, On varieties of minimal degree (a centennial account), Algebraic geometry, Bowdoin, 1985 (Brunswick, Maine, 1985), Proc. Sympos. Pure Math., vol. 46, Amer. Math. Soc., Providence, RI, 1987, pp. 3–13. ↑2 [Har95] Joe Harris, Algebraic geometry, Graduate Texts in Mathematics, vol. 133, Springer-Verlag, New York, 1995. A first course; Corrected reprint of the 1992 original. ↑2 [Kat08] Mordechai Katzman, On ideals of minors of matrices with indeterminate entries, Comm. Algebra 36 (2008), no. 1, 104–111. ↑2 [KPU17] Andrew R. Kustin, Claudia Polini, and Bernd Ulrich, A matrix of linear forms which is annihilated by a vector of indeterminates, J. Algebra 469 (2017), 120–187. ↑2 [Rai13] Claudiu Raicu, 3 × 3 minors of catalecticants, Math. Res. Lett. 20 (2013), no. 4, 745–756. ↑2 [Rai12] , Secant varieties of Segre-Veronese varieties, Algebra Number Theory 6 (2012), no. 8, 1817–1868. ↑2 Mathematical Institute Zeeman Building University of Warwick Coventry CV4 7AL UK E-mail address: [email protected] 9
1504.05395	1	1504	2015-04-21T11:50:02	The dual boundary complex of the $SL_2$ character variety of a punctured sphere	[ "math.AG" ]	Suppose $C_1,\ldots , C_k$ are generic conjugacy classes in $SL_2({\mathbb C})$. Consider the character variety of local systems on ${\mathbb P}^1-\{ y_1,\ldots , y_k\}$ whose monodromy transformations around the punctures $y_i$ are in the respective conjugacy classes $C_i$. We show that the dual boundary complex of this character variety is homotopy equivalent to a sphere of dimension $2(k-3)-1$.	math.AG	math	THE DUAL BOUNDARY COMPLEX OF THE SL2 CHARACTER VARIETY OF A PUNCTURED SPHERE CARLOS SIMPSON Abstract. Suppose C1, . . . , Ck are generic conjugacy classes in SL2(C). Consider the character variety of local systems on P1 − {y1, . . . , yk} whose monodromy transformations around the punc- tures yi are in the respective conjugacy classes Ci. We show that the dual boundary complex of this character variety is homotopy equivalent to a sphere of dimension 2(k − 3) − 1. 1. Introduction Given a smooth projective variety X, choose a normal crossings compactification X = X ∪ D and define a simplicial set called the dual boundary complex D∂X, containing the combinatorial informa- tion about multiple intersections of divisor components of D. Danilov, Stepanov and Thuillier have shown that the homotopy type of D∂X is independent of the choice of compactification, and this structure has been the subject of much study. We consider the case when X = MB(S; C1, . . . , Ck) is the character variety, of local systems on a punctured sphere S ∼ P1 − {y1, . . . , yk} such that the conjugacy classes of the monodromies around the punc- tures are given by C1, . . . , Ck respectively [36]. If these conjugacy classes satisfy a natural genericity condition then the character variety is a smooth affine variety. We prove that its dual boundary complex is a sphere of the appropriate dimension (see Conjecture 11.1), for local systems of rank 2. Theorem 1.1. Suppose C1, . . . , Ck are conjugacy classes in SL2(C) satisfying the genericity Condition 7.1. Then the dual boundary com- plex of the character variety is homotopy equivalent to a sphere: D∂MB(S; C1, . . . , Ck) ∼ S2(k−3)−1. 2010 Mathematics Subject Classification. Primary 14D20; Secondary 14T05, 57M50. Key words and phrases. Character variety, Dual complex, Fenchel-Nielsen coor- dinates, Moduli stack, P=W conjecture. 1 2 C. SIMPSON This statement is a part of a general conjecture about the bound- aries of moduli spaces of local systems [28]. The conjecture says that the dual boundary complex of the character variety or “Betti moduli space” should be a sphere, and that it should furthermore be naturally identified with the sphere at infinity in the “Dolbeault” or Hitchin mod- uli space of Higgs bundles. We will discuss this topic in further detail in Section 11 at the end of the paper. The case k = 4 of our theorem is a consequence of the Fricke-Klein expression for the character variety, which was indeed the motivation for the conjecture. The case k = 5 of Theorem 1.1 has been proven by Komyo [31]. 1.1. Strategy of the proof. Here is the strategy of our proof. We first notice that it is possible to make some reductions, based on the follow- ing observation (Lemma 2.3): if Z ⊂ X is a smooth closed subvariety of a smooth quasiprojective variety, such that the boundary dual complex is contractible D∂Z ∼ ∗, then the natural map D∂X → D∂(X − Z) is a homotopy equivalence. This allows us to remove some subvarieties which will be “negligeable” for the dual boundary complex. The main criterion is that if Z = A1 ×Y then D∂Z ∼ ∗ (Corollary 2.5). Together, these two statements allow us successively to remove a whole sequence of subvarieties (Proposition 2.6). The main technique is to express the moduli space MB(S; C1, . . . , Ck) in terms of a decomposition of S into a sequence of “pairs of pants” Si which are three-holed spheres. The decomposition is obtained by cutting S along (k − 3) circles denoted ρi. In each Si, there is one boundary circle corresponding to a loop ξi around the puncture yi, and two other boundary circles ρi−1 and ρi along which S was cut. At the start and the end of the sequence, two of the circles correspond to ξ1, ξ2 or ξk−1, ξk and only one to a cut. One may say that ρ1 and ρk−1 are confused with the original boundary circles ξ1 and ξk respectively. We would like to use this decomposition to express a local system V on S as being the result of “glueing” together local systems V Si on each of the pieces, glueing across the circles ρi. A basic intuition, which one learns from the elementary theory of classical hypergeomet- ric functions, is that a local system of rank 2 on a three-holed sphere is determined by the conjugacy classes of its three monodromy trans- formations. This is true generically, but one needs to take some care in degenerate cases involving potentially irreducible local systems, as will be discussed below. DUAL BOUNDARY COMPLEX 3 The conjugacy classes of the monodromy transformations around ρi are determined, except in some special cases, by their traces. The special cases are when the traces are 2 or −2. If we assume for the moment the uniqueness of V Si as a function of Ci and the traces ti−1 and ti of the monodromies around ρi−1 and ρi respectively, then the local system V is roughly speaking determined by specifying the values of these traces t2, . . . , tk−2, plus the glueing parameters. The glueing parameters should respect the monodromy transformations, and are defined modulo central scalars, so each pa- rameter is an element of Gm. In this rough picture then, the moduli space could be viewed as fibering over (A1)k−3 with fibers Gk−3 m . The resulting coordinates are classically known as Fenchel-Nielsen coordinates. Originally introduced to parametrize P GL2(R) local sys- tems corresponding to points in Teichmuller space, they have been extended to the complex character variety by Tan [48]. In the above discussion we have taken several shortcuts. We as- sumed that the traces ti determined the monodromy representations, and in saying that the glueing parameters would be in Gm we implicitly assumed that these monodromy transformations were diagonal with distinct eigenvalues. These conditions correspond to saying ti 6= 2, −2. We also assumed that the local system V Si was determined by Ci, ti−1 and ti. This is not in general true if it can be reducible, which is to say if there is a non-genericity relation between the conjugacy classes. The locus where that happens is somewhat difficult to specify explicity since there are several possible choices of non-genericity relation (the different choices of ǫi in Condition 4.3). We would therefore like a good way of obtaining such a rigidity even over the non-generic cases. Such a property is provided by the notion of stability. One may envision assigning parabolic weights to the two eigenvalues of Ci and assigning parabolic weights zero over ρj. The parabolic weights induce a notion of stable local system over Si. But in fact we don’t need to discuss parabolic weights themselves since the notion of stability can also be defined directly: a local system Vi on Si is unstable if it admits a rank 1 subsystem L such that the monodromy matrix in Ci acts on L by c−1 (a previously chosen one of the two eigenvalues of Ci). It is stable otherwise. Now, it becomes true that a stable local system is uniquely determined by Ci, ti−1 and ti. This will be the basis of our calculations in Section 10, see Corollary 10.3. i The first phase of our proof is to use the possibility for reductions given by Proposition 2.6 to reduce to the case of the open subset M ′ ⊂ MB(S; C1, . . . , Ck) 4 C. SIMPSON consisting of local systems V such that ti ∈ A1 − {2, −2} and such that V Si is stable. In order to make these reductions, we show in Sections 7 and 8 that the strata where some ti is 2 or −2, or where some V Si is unstable, have a structure of product with A1, hence by Lemma 2.5 these strata are negligeable in the sense that Lemma 2.3 applies. For the open set M ′, there is still one more difficulty. The glueing parameters depend a priori on all of the traces, so we don’t immediately get a decomposition of M ′ as a product. A calculation with matrices and a change of coordinates allow us to remedy this and we show in Theorem 10.6 that M ′ ∼= Qk−3 where Q is a space of choices of a trace t together with a point [p, q] in a copy of Gm. It turns out that this family of multiplicative groups over A1−{2, −2} is twisted: the two endpoints of the fibers Gm get permuted as t goes around 2 and −2. This twisting property is what makes it so that D∂Q ∼ S1, and therefore by [43, Lemma 6.2], D∂(Qk−3) ∼ S2(k−3)−1. This cal- culates D∂M ′ and hence also D∂MB(S; C1, . . . , Ck) to prove Theorem 1.1. We should consider the open subset M ′ as the natural domain of def- inition of the Fenchel-Nielsen coordinate system, and the components in the expression M ′ ∼= Qk−3 are the Fenchel-Nielsen coordinates. 1.2. Relation with other work. What we are doing here is closely related to a number of things. Firstly, as pointed out above, our calcu- lation relies on the Fenchel-Nielsen coordinate system coming from a pair of pants decomposition, and this is a well-known technique. Our only contribution is to keep track of the things which must be removed from the domain of definition, and of the precise form of the coordinate system, so as to be able to conclude the structure up to homotopy of the dual boundary complex. A few references about Fenchel-Nielsen coordinates include [8] [13] [42] [52], and for the complex case Tan’s paper [48]. Nekrasov, Rosly and Shatashvili’s work on bend parameters [39] involves similar co- ordinates and is related to the context of polygon spaces [12]. The work of Hollands and Neitzke [25] gives a comparison between Fenchel- Nielsen and Fock-Goncharov coordinates within the theory of spectral networks [11]. Jeffrey and Weitsman [26] consider what is the effect of a decomposition, in arbitrary genus, on the space of representations into a compact group. Recently, Kabaya uses these decompositions to give algebraic coordinate systems and furthermore goes on to study the DUAL BOUNDARY COMPLEX 5 mapping class group action [27]. These are only a few elements of a vast literature. Conjecture 11.1 relating the dual boundary complex of the character variety and the sphere at infinity of the Hitchin moduli space, should be viewed as a geometric statement reflecting the first weight-graded piece of the P = W conjecture of de Cataldo, Hausel and Migliorini [5] [16]. This will be discussed a little bit more in Section 11 but should also be the subject of further study. Komyo gave the first proof of the theorem that the dual boundary complex was a sphere, for rank 2 systems on the projective line minus 5 points [31]. He did this by constructing an explicit compactification and writing down the dual complex. This provides more information than what we get in our proof of Theorem 1.1, because we use a large number of reduction steps iteratively replacing the character variety by smaller open subsets. I first heard from Mark Gross in Miami in 2012 about a statement, which he attributed to Kontsevich, that if X is a log-Calabi-Yau variety (meaning that it has a compactification X = X ∪ D such that KX + D is trivial), then D∂X should be a sphere. Sam Payne points out that this idea may be traced back at least to [33, Remark 4] in the situation of a degeneration. Gross also stated that this property should apply to character vari- eties, that is to say that some or all character varieties should be log- CY. That has apparently been known folklorically in many instances cf [15]. Recently, much progress has been made. Notably, Koll´ar and Xu have proven that the dual boundary of a log-CY variety is a sphere in dimension 4, and they go a long way towards the proof in general [30]. They note that the correct statement, for general log-CY varieties, seems to be that D∂X should be a quotient of a sphere by a finite group. In our situation of character varieties, part of the statement of Conjecture 11.1 posits that this finite quotienting doesn’t happen. This is supported by our theorem, but it is hard to say what should be expected in general. De Fernex, Koll´ar and Xu have introduced a refined dual bound- ary complex [6], which is expected to be a sphere in the category of PL manifolds. That is much stronger than just the statement about homo- topy equivalence. See also Nicaise and Xu [40]. For character varieties, as well as for more general cluster varieties and quiver moduli spaces, the Kontsevich-Soibelman wallcrossing picture could be expected to 6 C. SIMPSON be closely related to this PL sphere, more precisely the Kontsevich- Soibelman chambers in the base of the Hitchin fibration should to cor- respond to cells in the PL sphere. One may witness this phenomenon by explicit calculation for SL2 character varieties of the projective line minus 4 points, under certain special choices of conjugacy classes where the character variety is the Cayley cubic. Recently, Gross, Hacking, Keel and Kontsevich [15] building on work of Gross, Hacking and Keel [14], have given an explicit combinato- rial description of a boundary divisor for log-Calabi-Yau cluster va- rieties. Their description depends on a choice of toroidal cluster co- ordinate patches, and the combinatorics involve toric geometry. It should in principle be possible to conclude from their construction that D∂MB(S; C1, . . . , Ck) is a sphere, as is mentioned in [15, Remark 9.12]. Their technique, based in essence on the Fock-Goncharov coordinate systems, should probably lead to a proof in much greater generality than our Theorem 1.1. 1.3. Varying the conjugacy classes. In the present paper, we have been considering the conjugacy classes C1, . . . , Ck as fixed. As Deligne pointed out, it is certainly an interesting next question to ask what happens as they vary. Nakajima discussed it long ago [38]. This has many different aspects and it would go beyond our current scope to enter into a detailed discussion. I would just like to point out that the natural domain on which ev- erything is defined is the space of choices of C1, . . . , Ck which satisfy the Kostov genericity Condition 4.3. This is an open subset of Gk m, the complement of a divisor K whose components are defined by multiplica- tive monomial equalities. It therefore looks like a natural multiplica- tive analogue of the hyperplane arrangement complements which enter into the theory of higher dimensional hypergeometric functions [50]. The variation with parameters of the moduli spaces MB(S; C1, . . . , Ck) leads, at the very least, to some variations of mixed Hodge structure over Gk m − K which undoubtedly have interesting properties. 1.4. Acknowledgements. I would like to thank the Fund for Mathe- matics at the Institute for Advanced Study for support. This work was also supported in part by the ANR grant 933R03/13ANR002SRAR (Tofigrou). It is a great pleasure to thank L. Katzarkov, A. Noll and P. Pandit for all of the discussions surrounding our recent projects, which have provided a major motivation for the present work. I would specially like to thank D. Halpern-Leistner, L. Migliorini and S. Payne for some DUAL BOUNDARY COMPLEX 7 very helpful and productive discussions about this work at the Insti- tute for Advanced Study. They have notably been suggesting several approaches to making the reasoning more canonical, and we hope to be able to say more about that in the future. I would also like to thank G. Brunerie, G. Gousin, A. Ducros, M. Gross, J. Koll´ar, A. Komyo, F. Loray, N. Nekrasov, M.-H. Saito, J. Weitsman, and R. Wentworth for interesting and informative discussions and suggestions. 1.5. Dedication. It is a great honor to dedicate this work to Vadim Schechtman. Vadim’s interests and work have illuminated many as- pects of the intricate interplay between topology and geometry in the de Rham theory of algebraic varieties. His work on hypergeometric functions [50] motivates our consideration of moduli spaces of local sys- tems on higher-dimensional varieties. His work with Hinich on dga’s in several papers such as [22] was one of the first instances of homo- topy methods for algebraic varieties. His many works on the chiral de Rham complex have motivated wide developments in the theory of D-modules and local systems. The ideas generated by these threads have been suffused throughout my own research for a long time. 2. Dual boundary complexes Suppose X is a smooth quasiprojective variety over C. By resolu- tion of singularities we may choose a normal crossings compactification X ⊂ X whose complementary divisor D := X − X has simple normal crossings. In fact, we may assume that it satisfies a condition which might be called very simple normal crossings: i=1 Di is the decomposition into irreducible components, then we can ask that any multiple intersection Di1 ∩ · · · ∩ Dik be either empty or connected. If the compactification satisfies this condition, then we obtain a simpli- cial complex denoted D∂X, the dual complex D(D) of the divisor D, defined as follows: there are m vertices e1, . . . , em of D∂X, in one-to- one correspondence with the irreducible components D1, . . . , Dm of D; and a simplex spanned by ei1, . . . , eik is contained in D∂X if and only if Di1 ∩ · · · ∩ Dik is nonempty. if D = Sm This defines a simplicial complex, which could be considered as a simplicial set, but which for the present purposes we shall identify with its topological realization which is the union of the span of those simplicies in Rm with ei being the standard basis vectors. The simplicial complex D∂X goes under several different terminolo- gies and notations. We shall call it the dual boundary complex of X. It contains the purely combinatorial information about the divisor com- pactifying X. The main theorem about it is due to Danilov [2]: 8 C. SIMPSON Theorem 2.1 (Danilov). The homotopy type of D∂X is independent of the choice of compactification. The papers of Stepanov [46] [47], concerning the analogous question for singularities, started a lot of renewed activity. Following these, a very instructive proof, which I first learned about from A. Ducros, was given by Thuillier [49]. He interpreted the homotopy type of D∂X as being equivalent to the homotopy type of the Berkovich boundary of X, namely the set of points in the Berkovich analytic space [1] associated to X (over the trivially valued ground field), which are valuations centered at points outside of X itself. Further refinements were given by Payne [43] and de Fernex, Koll´ar and Xu [6]. Payne showed that the simple homotopy type of D∂X was invariant, and proved several properties crucial to our arguments below. De Fernex, Koll´ar and Xu defined in some cases a special choice of compactification leading to a boundary complex D∂X whose PL homeomorphism type is invariant. Nicaise and Xu show in parallel, in the case of a degeneration at least, that the essential skeleton of the Berkovich space is a pseudo-manifold [40]. Manon considers an embedding of “outer space” for character varieties, into the Berkovich boundary [37]. These refined versions provide very interesting objects of study but for the present paper we just use the homotopy type of D∂X. Our goal will be to calculate the homotopy type of the dual boundary complex of some character varieties. To this end, we describe here a few important reduction steps allowing us to modify a variety while keeping its dual boundary complex the same. One should be fairly careful when manipulating these objects, as some seemingly insignificant changes in a variety could result in quite different boundary complexes. For example, in the case of an isotrivial fibration it isn’t enough to know the homotopy types of the base and the fiber—essentially, the fibration should be locally trivial in the Zariski rather than etale topology in order for that kind of reasoning to work. The space Q to be considered at the end of the paper provides an example of this phenomenon. In a similar vein, I don’t know of a good notion of dual boundary complex for an Artin stack. It is possible that a theory of Berkovich stacks could remedy this problem, but that seems difficult. Payne has suggested, in answer to the problem of etale isotrivial fibrations, to look at an equivariant notion of isotrivial fibration which could give a natural group action on a dual boundary complex such that the quotient would DUAL BOUNDARY COMPLEX 9 be meaningful. This type of theory might give an alternate approach to some of our problems. Let us get now to the basic properties of dual boundary complexes. The first step is to note that if U ⊂ X is an open subset of a smooth quasiprojective variety, then we obtain a map D∂X → D∂U. Lemma 2.2 (Payne). If X is an irreducible smooth projective variety and Z → X is obtained by blowing up a smooth center, then it induces a homotopy equivalence on dual boundary complexes D∂Z ∼ D∂X. Proof. See [43], where more generally boundary complexes of singular varieties are considered but we only need the smooth case. (cid:3) It follows from this lemma that if U ⊂ X is an open subset of a smooth quasiprojective variety, it induces a natural map of boundary complexes ∂X → ∂U. Indeed, for that we may assume by resolving sin- gularities and applying the previous lemma, that U is the complement of a divisor B ⊂ X, and furthermore there is a very simple normal crossings compactification X = X ∪ D such that B ∪ D also has very simple normal crossings. Then D∂U is the dual complex of the divisor B ∪ D, which contains D∂X, the dual complex of D, as a subcomplex. Following up on this idea, here is our main reduction lemma: Lemma 2.3. Suppose U ⊂ X is an open subset of an irreducible smooth quasiprojective variety, obtained by removing a smooth irre- ducible closed subvariety of smaller dimension Y = X − U ⊂ X. Sup- pose that D∂Y ∼ ∗ is contractible. Then the map D∂X → D∂U is a homotopy equivalence. Proof. Let X BlY be obtained by blowing up Y . From the previous lemma, D∂X BlY ∼ D∂X. Let Bl(Y ) ⊂ X BlY be the inverse image of Y . It is an irreducible smooth divisor, and U is also the complement of this divisor in X BlY . By resolution of singularities we may choose a compactification X BlY such that the boundary divisor D, plus the closure B := Bl(Y ), form a very simple normal crossings divisor. This combined divisor is therefore a boundary divisor for U, so D∂U ∼ D(D ∪ B). Now this bigger dual complex D(D ∪ B) has one more vertex than D(D), corresponding to the irreducible component B. The star of this vertex is the cone over D∂Bl(Y ) = D(B ∩ D). The cone is attached to D(D) via its base D(B ∩ D), to give D(B ∪ D). We would like to show that D∂Bl(Y ) ∼ ∗. The first step is to notice that Bl(Y ) → Y is the projective space bundle associated to the vector bundle NY /X over Y . 10 C. SIMPSON We claim in general that if V is a vector bundle over a smooth quasiprojective variety Y , then D∂(P(V )) ∼ D∂(Y ). The proof of this claim is that there exists a normal crossings compactification Y of Y such that the vector bundle V extends to a vector bundle on Y . That may be seen by choosing a surjection from the dual of a direct sum of very ample line bundles to V , getting V as the pullback of a tautological bundle under a map from Y to a Grassmanian. The compactification may be chosen so that the map to the Grassmanian extends. We obtain a compactification of P(V ) wherein the boundary divisor is a projective space bundle over the boundary divisor of Y , and with these choices D∂P(V ) = D∂Y . It follows from Danilov’s theorem that for any other choice, there is a homotopy equivalence. Back to our situation where Bl(Y ) = P(NY /X), and assuming that D∂Y ∼ ∗, we conclude that D∂Bl(Y ) ∼ ∗ too. Therfore the dual complex D(B ∩ D) is contractible. Now D∂U = D(B ∪ D) is obtained by attaching to D(D) the cone over D(B ∩ D). As we have seen above D(B ∩ D) is contractible, so coning it off doesn’t change the homotopy type. This shows that the map D∂X = D(D) → D(B ∪ D) = D∂U is a homotopy equivalence. (cid:3) In order to use this reduction, we need a criterion for the condition D∂Y ∼ ∗. Note first the following general property of compatibility with products. Lemma 2.4 (Payne). Suppose X and Y are smooth quasiprojective varieties. Then D∂(X × Y ) is the join of D∂(X) and D∂(Y ), in other words we have a homotopy cocartesian diagram of spaces D∂(X) × D∂(Y ) → D∂(Y ) ↓ D∂(X) ↓ → D∂(X × Y ) . Proof. This is [43, Lemma 6.2]. (cid:3) Corollary 2.5. Suppose Y is a smooth quasiprojective variety. Then D∂(A1 × Y ) ∼ ∗. Proof. Setting X := A1 in the previous lemma, we have D∂(X) ∼ ∗, so in the homotopy cocartesian diagram the top arrow is an equivalence and the left vertical arrow is the projection to ∗; therefore the homotopy pushout is also ∗. (cid:3) DUAL BOUNDARY COMPLEX 11 Proposition 2.6. Suppose U ⊂ X is a nonempty open subset of a smooth irreducible quasiprojective variety, and suppose the complement Z := X − U has a decomposition into locally closed subsets Zj such ∼= A1 × Yj. Suppose that this decomposition can be ordered into that Zj a stratification, that is to say there is a total order on the indices such thatSj≤a Zj is closed for any a. Then D∂(X) ∼ D∂(U). Proof. We first prove the proposition under the additional hypothesis that the Yj are smooth. Proceed by induction on the number of pieces in the decomposition. Let Z0 be the lowest piece in the ordering. The ordering hypothesis says that Z0 is closed in X. Let X ′ := X − Z0. Now U ⊂ X ′ is the complement of a subset Z ′ =Sj>0 Zj decomposing in the same way, with a smaller number of pieces, so by induction we know that D∂(X ′) ∼ D∂(U). By hypothesis Z0 ∼= A1×Y0. Lemma 2.5 tells us that D∂(Z0) ∼ ∗ and now Lemma 2.3 tells us that D∂(X) ∼ D∂(X ′), so D∂(X) ∼ D∂(U). This completes the proof of the proposition under the hypothesis that Yj are smooth. Now we prove the proposition in general. Proceed as in the first paragraph of the proof with the same notations: by induction we may assume that D∂(X ′) ∼ D∂(U) where X ′ = X −Z0 such that Z0 is closed and isomorphic to A1 × Y0. Choose a totally ordered stratification of Y0 by smooth locally closed subvarieties Y0,i. Set Z0,i := A1 × Y0,i. This collection of subvarieties of X now satisfies the hypotheses of the proposition and the pieces are smooth. Their union is Z0 and its complement in X is the open subset X ′. Thus, the first case of the proposition treated above tells us that D∂(X) ∼ D∂(X ′). It follows that D∂(X) ∼ D∂(U), completing the proof. (cid:3) Caution: A simple example shows that the condition of ordering, in the statement of the propostion, is necessary. Suppose X is a smooth projective surface containing two projective lines D1, D2 ⊂ X such that their intersection D1 ∩ D2 = {p1, p2} consists of two distinct points. Then we could look at Z1 = D1 − {p1} and Z2 = D2 − {p2}. Both Z1 and Z2 are affine lines. Setting U := X − (D1 ∪ D2) = X − (Z1 ∪ Z2) we get an open set which is the complement of a subset Z = Z1 ⊔ Z2 decomposing into two affine lines; but D∂X = ∅ whereas D∂U ∼ S1. 3. Hybrid moduli stacks of local systems The moduli space of local systems is different from the moduli stack, even at the points corresponding to irreducible local systems. Indeed, the open substack of the moduli stack parametrizing irreducible GLr- local systems is a Gm-gerbe over the corresponding open subset of the 12 C. SIMPSON moduli space. Even by considering SLr-local systems we can only reduce this to being a µr-gerbe. However, it is usual and convenient to consider the moduli space instead. In this section, we mention a construction allowing to define what we call1 a hybrid moduli stack in which the central action is divided out, making it so that for irreducible points it is the same as the moduli space. Our initial discussion will use some simple 2-stacks, however the reader wishing to avoid these is may refer to Proposition 3.1 which gives an equivalent definition in more concrete terms. Consider a reductive group G with center Z. The fibration sequence of 1-stacks may be transformed into the cartesian diagram BZ → BG → B(G/Z) (3.1) BG → B(G/Z) ↓ ∗ → K(Z, 2) ↓ of Artin 2-stacks on the site Aff f t,et C with the etale topology. C of affine schemes of finite type over Suppose now that S is a space or higher stack. Then we may consider the relative mapping stack M(S, G) := Hom(S, B(G/Z)/K(Z, 2)) → K(Z, 2). It may be defined as the fiber product forming the middle arrow in the following diagram where both squares are cartesian: Hom(S, BG) → M(S, G) → Hom(S, B(G/Z)) ↓ ∗ ↓ ↓ . → K(Z, 2) → Hom(S, K(Z, 2)) Here the bottom right map is the “constant along S” construction induced by pullback along S → ∗. The bottom left arrow ∗ → K(Z, 2) is the universal Z-gerbe, so its pullback on the upper right is again a Z-gerbe. We have thus con- structed a stack M(S, G) over which Hom(S, BG) is a Z-gerbe. From the definition it is a priori a 2-stack, and indeed M(∅, G) = K(Z, 2), but the following alternate characterization tells us that M(S, G) is usually a usual 1-stack. 1This is not new but I don’t remember where it is from, so no claim is made to originality. DUAL BOUNDARY COMPLEX 13 Proposition 3.1. Suppose S is a nonempty connected CW-complex with basepoint x. Then the hybrid moduli stack may be expressed as the stack-theoretical quotient M(S, G) = Rep(π1(S, x), G)//(G/Z). In particular, it is an Artin 1-stack. Proof. The representation space may be viewed as a mapping stack Rep(π1(S, x), G) = Hom((S, x), (BG, o)). Consider the big diagram Hom((S, x), (BG, o)) → Hom((S, x), (B(G/Z), o)) ↓ ∗ ↓ → Hom((S, x), (K(Z, 2), o)) → ↓ ↓ ∗ ↓ K(Z, 2) → Hom(S, K(Z, 2)) → K(Z, 2) where the bottom right map is evaluation at x. The pointed mapping 2-stack in the middle is defined by the condition that the bottom right square is homotopy cartesian. The composition along the bottom is the identity, so if we take the homotopy fiber product on the bottom left, the full bottom rectangle is a pullback too so that homotopy fiber product would be ∗ as is written in the diagram. In other words, the bottom left square is also homotopy cartesian. The middle horizontal map on the left sends the point to the map S → o ֒→ K(Z, 2), indeed it is constant along S because it comes from pullback of the bottom left map, and its value at x is o because of the right vertical map. Now, the upper left square is homotopy cartesian, just the result of applying the pointed mapping stack to the diagram (3.1). It follows that the whole left rectangle is homotopy cartesian. Consider, on the other hand, the diagram Hom((S, x), (BG, o)) → Hom((S, x), (B(G/Z), o)) → ↓ M(S, G) ↓ K(Z, 2) ↓ ∗ ↓ → → Hom(S, B(G/Z)) → B(G/Z) ↓ Hom(S, K(Z, 2)) . The bottom square is homotopy-cartesian by the definition of M(S, G). We proved in the previous paragraph that the full left rectangle is homotopy cartesian. In this 2-stack situation note that a commu- tative rectangle constitutes a piece of data rather than just a prop- erty. In this case, these data for the left squares are obtained by just considering the equivalence found in the previous paragraph, from 14 C. SIMPSON Hom((S, x), (BG, o)) to the homotopy pullback in the full left rec- tangle which is the same as the composition of the homotopy pull- backs in the two left squares. In particular, the upper left square is homotopy-cartesian. It now follows that the upper full rectangle is homotopy-cartesian. That exactly says that we have an action of G/Z on Hom((S, x), (BG, o)) = Rep(π1(S, x), G) and M(S, G) is the quo- tient. (cid:3) The hybrid moduli stacks also satisfy the same glueing or factoriza- tion property as the usual ones. Lemma 3.2. Suppose S = S1 ∪ S2 with S12 := S1 ∩ S2 excisive. Then M(S, G) ∼= M(S1, G) ×M (S12,G) M(S2, G). Proof. The mapping stacks entering into the definition of M(S, G) as a homotopy pullback, satisfy this glueing property. Notice that this is true even for the constant functor which associates to any S the stack K(Z, 2). The homotopy pullback therefore also satisfies the glueing property since fiber products commute with other fiber products. (cid:3) Suppose G = GLr so Z = Gm and G/Z = P GLr, and suppose S is a connected CW-complex. Let Hom(S, BGLr)irr ⊂ Hom(S, BGLr) denote the open substack of irreducible local systems. It is classical that the stack Hom(S, BGLr) has a coarse moduli space MB(S, GLr), and that the open substack Hom(S, BGLr)irr is a Gm-gerbe over the corresponding open subset of the coarse moduli space MB(S, G)irr. Proposition 3.3. In the situation of the previous paragraph, we have a map which restricts to an isomorphism M(S, GLr) → MB(S, GLr) M(S, GLr)irr ∼= MB(S, GLr)irr between the open subsets parametrizing irreducible local systems. The same holds for G = SLr. Lemma 3.4. The determinant map GLr diagram det→ Gm induces a cartesian M(S, SLr) → M(S, GLr) ↓ ∗ ↓ → M(S, Gm) DUAL BOUNDARY COMPLEX 15 which essentially says that M(S, SLr) is the substack of M(S, GLr) parametrizing local systems of trivial determinant. Note that M(S, Gm) is isomorphic to the quasiprojective variety Hom(H1(S), Gm). In what follows, we shall use these stacks M(S, GLr) which we call hybrid moduli stacks as good replacements intermediary between the moduli stacks of local systems and their coarse moduli spaces. 4. Boundary conditions Let S denote a 2-sphere with k open disks removed. It has k bound- ary circles denoted ξ1, . . . , ξk ⊂ S and From now on we consider rank 2 local systems on this surface S. ∂S = ξ1 ⊔ · · · ⊔ ξk. Fix complex numbers c1, . . . , ck all different from 0, 1 or −1. Let Ci :=(cid:26)P(cid:18) ci 0 0 c−1 i (cid:19) P −1(cid:27) denote the conjugacy class of matrices with eigenvalues ci, c−1 . i Consider the hybrid moduli stack M(S, GL2) constructed above, and let M(S; C·) ⊂ M(S, GL2) denote the closed substack consisting of local systems such that the monodromy transformation around ξi is in the conjugacy class Ci. See [36]. If we choose a basepoint x ∈ S and paths γi going from x out by straight paths to the boundary circles, around once and then back to x, then π1(S, x) is generated by the γi subject to the relation that their product is the identity. Therefore, the moduli stack of framed local systems is the affine variety Hom((S, x), (BGL2, o)) = Rep(π1(S, x), GL2) = {(A1, . . . , Ak) ∈ (GL2)k s.t. A1 · · · Ak = 1}. The unframed moduli stack is the stack-theoretical quotient Hom(S, BGL2) = Rep(π1(S, x), GL2)//GL2 by the action of simultaneous conjugation. The center Gm ⊂ GL2 acts trivially on Rep(π1(S, x), GL2) so the action of GL2 there factors through an action of P GL2. Proposition 3.1 may be restated as 16 C. SIMPSON Lemma 4.1. The hybrid moduli stack M(S, GL2) may be described as the stack-theoretical quotient M(S, GL2) = Rep(π1(S, x), GL2)//P GL2. Let Rep(π1(S, x), GL2; C·) ⊂ Rep(π1(S, x), GL2) denote the closed subscheme of representations which send γi to the conjugacy class Ci. These conditions are equivalent to the equations Tr(ρ(γi)) = ci + c−1 . We have i Rep(π1(S, x), GL2; C·) = {(A1, . . . , Ak) s.t. Ai ∈ Ci and A1 · · · Ak = 1}. Corollary 4.2. The hybrid moduli stack with fixed conjugacy classes is given by M(S; C·) = Rep(π1(S, x), GL2; C·)//P GL2 = {(A1, . . . , Ak) s.t. Ai ∈ Ci and A1 · · · Ak = 1}//P GL2. It is also isomorphic to the stack one would have gotten by using the group SL2 rather than GL2. Proof. Our conjugacy classes have been defined as having determinant one. Since the γi generate the fundamental group, if the ρ(γi) have determinant one then the representation ρ goes into SL2. As P GL2 = P SL2, the hybrid moduli stack for GL2 is the same as for SL2. (cid:3) Recall the following Kostov-genericity condition [35] on the choice of the numbers ci. Condition 4.3. For any choice of ǫ1, . . . , ǫk ∈ {1, −1} the product is not equal to 1. 1 · · · cǫk cǫ1 k The following basic lemma has been observed by Kostov and others. Lemma 4.4. If Condition 4.3 is satisfied then any representation in Rep(π1(S, x), GL2; C·) is irreducible. In particular, the automorphism group of the corresponding GL2 local system is the central Gm. The set of (c1, . . . , ck) satisfying this condition is a nonempty open subset of (Gm − {1, −1})k. We also speak of the same condition for the sequence of conjugacy classes C·. Proposition 4.5. Suppose C· satisfy Condition 4.3. The hybrid moduli stack M(S; C·) is an irreducible smooth affine variety. It is equal to the coarse, which is indeed fine, moduli space MB(S; C1, . . . , Ck) of local systems with our given conjugacy classes. DUAL BOUNDARY COMPLEX 17 Proof. The representation space Rep(π1(S, x), GL2; C·) is an affine va- riety, call it Spec(A), on which the group P GL2 acts. The moduli space is by definition MB(S; C1, . . . , Ck) := Spec(AP GL2). By Lemma 4.4 and using the hypothesis 4.3 it follows that the stabiliz- ers of the action are trivial. Luna’s etale slice theorem (see [7]) implies that the quotient map Spec(A) → Spec(AP GL2) is an etale fiber bundle with fiber P GL2. Therefore this quotient is also the stack-theoretical quotient: Spec(AP GL2) = Spec(A)//P GL2. By Corollary 4.2 that stack-theoretical quotient is M(S; C·), complet- ing the identification between the hybrid moduli stack and the moduli space required for the proposition. Smoothness of the moduli space has been noted in many places, see for example [17] [36]. Irreducibility is proven in a general context in [18] [36] as a consequence of computations of E-polynomials, and a different proof is given in [45] using moduli stacks of parabolic bundles. In our case irreducibility could also be obtained by including dimension estimates for the subvarieties which will be removed in the course of our overall discussion. (cid:3) This proposition says that our hybrid moduli stack M(S; C·) is the same as the usual moduli space. A word of caution is necessary: we shall also be using M(S ′, C·) for subsets S ′ ⊂ S, and those are in general stacks rather than schemes, for example when Condition 4.3 doesn’t hold over S ′. 5. Interior conditions and factorization We now define some conditions concerning what happens in the in- terior of the surface S. These conditions will serve to define a stratifi- cation of M(S; C·). The biggest open stratum denoted M ′, treated in detail in Section 10, turns out to be the main piece, contributing the essential structure of the dual boundary complex. The smaller strata will be negligeable for the dual boundary complex, in view of Lemmas 2.3 and 2.5 as combined in Proposition 2.6. Divide S into closed regions denoted S2, . . . , Sk−1 such that Si ∩ Si+1 = ρi is a circle for 2 ≤ i ≤ k − 2, and the regions are otherwise disjoint. We assume that Si encloses the boundary circle ξi, so it is a 3-holed sphere with boundary circles ρi−1, ξi and ρi. The orientation of 18 C. SIMPSON ρi−1 is reversed when it is viewed from inside Si. The end piece S2 has boundary circles ξ1, ξ2 and ρ2 while the end piece Sk−1 has boundary circles ρk−2, ξk−1 and ξi. This is a “pair of pants” decomposition. Factorization properties, related to chiral algebra cf [9] [10], are a kind of descent. We will be applying the factorization property of Theorem 5.4 to the decomposition of our surface into pieces Si. This classical technique in geometric topology was also used extensively in the study of the Verlinde formula. The factorization is often viewed as coming from a degeneration of the curve into a union of rational lines with three marked points. For our argument it will be important to consider strata of the mod- uli space defined by fixing additional combinatorial data with respect to our decomposition. To this end, let us consider some nonempty sub- sets σi ⊂ {0, 1} for i = 2, . . . , k − 1, and conjugacy-invariant subsets G2, . . . , Gk−2 ⊂ SL2. We denote by α = (σ1, . . . , σk−1; G2, . . . , Gk−2) this collection of data. The subsets Gi will impose conditions on the monodromy around the circles ρi, while the σi will correspond to the following stability condition on the restrictions of our local system to Si. Recall that a local system V ∈ M(S; C·) is required to have mon- odromy around ξi with eigenvalues ci and c−1 . We are making a choice of orientation of these boundary circles, and ci 6= c−1 i by hypothesis, so the c−1 eigenspace is a well-defined rank 1 subspace of V ξi. i i Definition 5.1. We say that a local system V Si on Si, satisfying the conjugacy class condition, is unstable if there exists a rank 1 subsystem L ⊂ V Si such that the monodromy of L around ξi is c−1 . Say that V Si is stable otherwise. i An irreducible local system V Si is automatically stable; one which If V Si is a decomposes as a direct sum is automatically unstable. nontrivial extension with a unique rank 1 subsystem L, then V Si is unstable if Lξi -eigenspace of the monodromy, whereas it is stable if Lξi is the ci-eigenspace. We will later express these condi- tions more concretely in terms of vanishing or nonvanishing of a certain matrix coefficient. is the c−1 i Definition 5.2. Let M α(S; C·) ⊂ M(S; C·) denote the locally closed substack of local systems V satisfying the following conditions: • if σi = {0} then V Si is required to be unstable; if σi = {1} then it is required to be stable; and if σi = {0, 1} then there is no condition; and • the monodromy of V around ρi should lie in Gi. DUAL BOUNDARY COMPLEX 19 Consider a subset S ′ ⊂ S made up of some or all of the Si or the circles. Let M α(S ′; C·) denote the moduli stack of local systems on S ′ satisfying the above conditions where they make sense (that is, for the restrictions to those subsets which are in S ′). In the case of the inner boundary circles we may just use the notation M α(ρi) since the choices of conjugacy classes Ci, corresponding circles ξi, don’t intervene. In the case of Si, only the conjugacy class Ci matters so we may use the notation M α(Si; Ci). Suppose S ′ ⊂ S is connected and x ∈ S ′. Let Repα(π1(S ′, x), GL2; C·) ⊂ Rep(π1(S ′, x), GL2) denote the locally closed subscheme of representations which satisfy conjugacy class conditions corresponding to C· and the conditions cor- responding to α, that is to say whose corresponding local systems are in M α(S ′; C·). Proposition 3.1 says: Lemma 5.3. The simultaneous conjugation action of GL2 on the space of representations Repα(π1(S ′, x), GL2; C·) factors through an action of P GL2 and M α(S ′; C·) = Repα(π1(S ′, x), GL2; C·)//P GL2 is the stack-theoretical quotient. The hybrid moduli stacks allow us to state a glueing or factorization property, expressing the fact that a local system L on S may be viewed as being obtained by glueing together its pieces LSi along the circles ρi. Theorem 5.4. We have the following expression using homotopy fiber products of stacks: M α(S; C·) = M α(S2; C·) ×M α(ρi) M α(S3; C·) · · ·×M α(ρk−2) M α(Sk−1; C·). Proof. Apply Lemma 3.2. (cid:3) Corollary 5.5. Suppose the requirements given for the boundary pieces of ∂S ′ (which are circles either of the form ξi or ρi) satisfy Condition 4.3 for S ′. Then the moduli stack M α(S ′; C·) is in fact a quasiprojective variety. Proof. This follows from Proposition 4.5 applied to S ′. (cid:3) 20 C. SIMPSON 6. Universal objects Let us return for the moment to the general situation of Section 3, of a space S and a group G. If x ∈ S is a basepoint, then we obtain a principal (G/Z)-bundle over Hom(S, B(G/Z)), and this pulls back to a principal (G/Z)-bundle denoted F (S, x) → M(S, G). It may be viewed as the bundle of frames for the local systems, up to action of the center Z. If y ∈ S is another point, and γ is a path from x to y then it gives an isomorphism of principal bundles F (S, x) ∼= F (S, y) over M(S, G). In particular, π1(S, x) acts on F (S, x) in a tautological representation. Suppose S = S1 ∪ S2 such that the intersection S12 = S1 ∩ S2 is connected. Choose a basepoint x ∈ S12. This yields principal (G/V )- bundles F (S1, x) and F (S2, x) over M(S1, G) and M(S2, G) respec- tively. The fundamental group π1(S12, x) acts on both of these. We may restate the glueing property of Lemma 3.2 in the following way. Proposition 6.1. We have an isomorphism of stacks lying over the product M(S1, G) × M(S2, G), M(S, G) ∼= Isoπ1(S12,x)−G/V (p∗ 1F (S1, x), p∗ 2F (S2, x)) where on the right is the stack of isomorphisms, relative to M(S1, G) × M(S2, G), of principal G/V -bundles provided with actions of π1(S12, x). Return now to the notation from the immediately preceding sections. There are several ways of dividing our surface S into two or more pieces, various of which shall be used in the next section. Choose basepoints xi in the interior of Si, and si on the boundary circles ρi. Connect them by paths, nicely arranged with respect to the other paths γi. Then, over any subset S ′ containing a basepoint xi, we obtain a principal P GL2-bundle F (S ′, xi) → M(S ′, C·), and the same for si. Our paths, when in S ′, give isomorphisms between these principal bundles. It will be helpful to think of the description of glueing given by Proposition 6.1, using these basepoints and paths. The following local triviality property is useful. Lemma 6.2. Suppose S ′ has at most one boundary circle of the form ρi, and suppose that the conjugacy classes determining the moduli prob- lem on M α(S ′, C·) satisfy Condition 4.3, and suppose that x ∈ S ′ is one of our basepoints. Then the principal P GL2-bundle F (S ′, x) → M α(S ′, C·) is locally trivial in the Zariski topology of the moduli space M α(S ′, C·), and Zariski locally F (S ′, x) may be viewed as the projective frame bundle of a rank 2 vector bundle. DUAL BOUNDARY COMPLEX 21 Proof. Consider a choice of three loops (γj1, γj2, γj3) and a choice of one of the two eigenvalues of the conjugacy class Cj1, Cj1, or Cj1 for each of them. This gives three rank 1 eigenspaces in Vx for any local system V . Over the Zariski open subset of the moduli space where these three subspaces are distinct, they provide the required projective frame. Notice that the eigenspaces of the γj cannot all be aligned since these loops generate the fundamental group of S ′, by the hypothesis that there is at most one other boundary circle ρi. Therefore, as the our choices of triple of loops and triple of eigenvalues range over the possible ones, these Zariski open subsets cover the moduli space. We get the required frames. A framed P GL2-bundle comes from a vector bundle so F (S ′, x) locally comes from a GL2-bundle. (cid:3) 7. Splitting along the circle ρi In this section we consider one of the circles ρi which divides S into two pieces. Let S<i :=[j<i Sj , S>i :=[j>i Sj, and similarly define S≤i and S≥i. We have the decomposition S = S≤i ∪ S>i into two pieces intersecting along the circle ρi. Thus, M α(S; C·) = M α(S≤i; C·) ×M α(ρi) M α(S>i; C·). This factorization will allow us to analyze strata where Gi is a unipotent or trivial conjugacy class. The following condition will be in effect: Condition 7.1. We assume that the sequence of conjugacy classes C1, . . . , Ck is very generic, meaning that for any i the partial sequences C1, . . . , Ci and Ci, . . . , Ck satisfy Condition 4.3, and they also satisfy that condition if we add the scalar matrix −1. That is to say, no product of eigenvalues or their inverses should be equal to either 1 or −1. Suppose that Gi = {1}. Then M α(ρi) = B(P GL2). On the other hand, Condition 7.1 means that the sequences of conjugacy classes defining the moduli problems on S≤i and S>i themselves satisfy Con- dition 4.3. Therefore Proposition 4.5 applies saying that the moduli stacks M α(S≤i; C·) and M α(S>i; C·) exist as quasiprojective varieties. The projective frame bundles over a basepoint of ρi are principal P GL2-bundles denoted F≤i → M α(S≤i; C·) 22 and C. SIMPSON F>i → M α(S>i; C·). These principal bundles may be viewed as given by the maps M α(S≤i; C·) → M α(ρi) = B(P GL2) ← M α(S>i; C·). These principal bundles are locally trivial in the Zariski topology by Lemma 6.2. The principal bundle description of the moduli space in Proposition 6.1 now says M α(S; C·) = Iso(p∗ 1(F≤i), p∗ 1(F≤i))/M α(S≤i; C·) × M α(S>i; C·). The bundle of isomorphisms between our two principal bundles, is a fiber bundle with fiber P GL2, locally trivial in the Zariski topology be- cause the two principal bundles are Zariski-locally trivial. We may sum up this conclusion with the following lemma, noting that the argument also works the same way if Gi = {−1}. Lemma 7.2. Under the assumption that Gi = {1}, the moduli space M α(S; C·) is a fiber bundle over M(S≤i; C·)×M(S>i; C·), locally trivial in the Zariski topology, with fiber P GL2. The same holds true if Gi = {−1}. Consider the next case: suppose that Gi is the conjugacy class of matrices conjugate to a nontrivial unipotent matrix U =(cid:18) 1 1 0 1 (cid:19) . In that case, M α(ρi) = BGa. The situation is the same as before: the moduli spaces M α(S≤i; C·) and M α(S>i; C·) are quasiprojective vari- eties, and we have principal bundles F≤i and F>i. This time, these principal bundles have unipotent automorphisms denoted R′ and R respectively, in the conjugacy class of U. We have M α(S; C·) = IsoM α(S≤i;C·)×M α(S>i;C·)(p∗ 1(F≤i, R′), p∗ 1(F≤i, R)). This means the relative isomorphism bundle of the principal bundles together with their automorphisms. We claim that these principal bundles together with their automor- phisms may be trivialized locally in the Zariski topology. For the prin- cipal bundles themselves this is Lemma 6.2. The unipotent endomor- phisms then correspond, with respect to these local trivializations, to maps into P GL2/Ga. One can write down explicit sections of the pro- jection P GL2 → P GL2/Ga locally in the Zariski topology of the base, and these give the claimed local trivializations. One might alternatively DUAL BOUNDARY COMPLEX 23 notice here that a Ga-torsor for the etale topology is automatically lo- cally trivial in the Zariski topology by “Hibert’s theorem 90”. From the result of the previous paragraph, M α(S; C·) is a fiber bun- dle over M α(S≤i; C·) × M α(S>i; C·), locally trivial in the Zariski topol- ogy, with fiber the centralizer Z(R) ⊂ P GL2 of a unipotent element ∼= A1. We obtain the following R ∈ P GL2. This centralizer is Ga statement. Lemma 7.3. Under the assumption that Gi is the unipotent conjugacy class, the moduli space M α(S; C·) is a fiber bundle over M α(S≤i; C·) × M α(S>i; C·), locally trivial in the Zariski topology, with fiber A1. The same holds true if Gi is the conjugacy class of matrices conjugate to −U. We may sum up the conclusion of this section as follows. Proposition 7.4. With the hypothesis of Condition 7.1 in effect, sup- pose that the datum α is chosen such that for some i, Gi is one of the following four conjugacy classes {1}, {−1}, {P UP −1}, or {−P UP −1}, that is to say the conjugacy classes whose traces are 2 or −2. Then the dual boundary complex of the α-stratum is contractible: D∂M α(S, C·) ∼ ∗. Proof. In all four cases, covered by Lemmas 7.2 and 7.3 above, the space M α(S, C·) admits a further decomposition into locally closed pieces all of which have the form A1 × Y . Therefore, Lemmas 7.2 and 7.3 apply to show that the dual boundary complex is contractible. (cid:3) 8. Decomposition at Si in the unstable case Define the function ti : M(S, C·) → A1 sending a local system to the trace of its monodromy around the circle ρi. In the previous section, we have treated any strata which might be defined in such a way that at least one of the Gi is a conjugacy class with ti equal to 2 or −2. Therefore, we may now assume that all of our subsets Gi consist entirely of matrices with trace different from 2, −2. In particular, these matrices are semisimple with distinct eigenvalues. If Gi consists of a single conjugacy class, it is possible to choose one of the two eigenvalues. But in general, this is not possible. However, in the situation considered in the present section, where one of the σi indicates an unstable local system, then the destabilizing subsystem serves to pick out a choice of eigenvalue. 24 C. SIMPSON In the case where one of the σi is {0} stating that V Si should be un- stable, we will again obtain a structure of decomposition into a product with A1 locally over a stratification, essentially by considering the ex- tension class of the unstable local system. Some arguments are needed in order to show that this leads to direct product decompositions. 8.1. Some cases with Gi−1 and Gi fixed. We suppose in this sub- section that Gi−1 and Gi are single conjugacy classes, with traces differ- ent from 2, −2, and furthermore chosen so that the moduli problem for M α(S>i; C·) on one side is Kostov-generic. Hence, that moduli stack is a quasiprojective variety. Furthermore we assume that σi = {0}. Therefore, M α(Si; Ci) is the moduli stack of unstable local systems on Si. The elements here are local systems V fitting into an exact sequence 0 → L → V → L′ → 0 such that the monodromy of L on ξi has eigenvalue c−1 that M α(Si; Ci) is nonempty. i . We assume Remark 8.1. If we are given the conjugacy classes Gi−1 and Gi such that there exists an unstable local system V on Si, then the eigenvalues bi−1 of L on ρi−1, and bi of L on ρi, are uniquely determined. i along ξi. Suppose that bi−1c−1 Proof. The conjugacy classes Gi−1, Gi determine the pairs (bi−1, b−1 i−1) and (bi, b−1 i ) respectively. The instability condition says that L has eigenvalue c−1 i bi = 1 so there exists a local system L with eigenvalues bi−1 and bi. We show that the other products with either b−1 or both, are different from 1. For example, bi−1c−1 6= 1 since we are assuming that Gi is a conjugacy class with distinct eigenvalues. Thus bi−1c−1 6= 1. Similarly, b−1 6= 1. This shows that if there is one possible combination of eigenvalues for a sub-local system, then it is unique. (cid:3) i−1 or b−1 i = b−2 , but b2 i i b−1 i−1c−1 i bi 6= 1. Also, b−1 i−1c−1 i b−1 i = c−2 i b−1 i i i i From the assumption that M α(Si; Ci) is nonempty and the previous remark, we may denote by bi−1 and bi the eigenvalues of L on ρi−1 and ρi respectively. We are assuming a genericity condition implying that M α(S>i; C·) is a quasiprojective variety. It has a universal principal bundle F>i over it, and this has an automorphism R corresponding to the monodromy transformation around ρi. The eigenvalues of R are bi and b−1 Restrict to a finer stratification of M α(S>i; C·) into strata denoted M α(S>i)a on which (F>i, R) is trivial. Let M α(S; C·)a be the inverse image of M α(S>i; C·)a under the map M α(S; C·) → M α(S>i; C·). . i DUAL BOUNDARY COMPLEX 25 Proposition 8.2. We have M α(S; C·)a = M α(S>i; C·)a × M α(S≤i; C·)fr,R where M α(S≤i; C·)fr,R is the moduli space of framed local systems, that is to say local systems with a projective framing along ρi compatible with the monodromy and having the specified eigenvalues (bi, b−1 i ). Proof. Use Proposition 6.1. (cid:3) Without the conditions α = (σ·, G·), the framed moduli space is just the space of sequences of group elements A1, . . . , Ai, in conjugacy classes C1, . . . , Ci respectively, such that A1 · · · AiR = 1. Denote this space by Rep(C1, . . . , Ci; R). The moduli space M α(S≤i, C·)fr,R is the subspace of Rep(C1, . . . , Ci; R) given by the conditions σ· and G·. Notice here that, since we don’t know a genericity condition for (C1, . . . , Ci, Gi) the moduli space might not be smooth. Even though we are considering framed representations, at a reducible representa- tion the space will in general have a singularity. Furthermore, the conditions Gj might, in principle, introduce other singularities. Theorem 8.3. With the above notations, let R′ be an element in the conjugacy class Gi−1. We have M α(S≤i; C·)fr,R ∼= A1 × M α(S≤i−1; C·)fr,R′ . Proof. It isn’t too hard to see that the moduli space is an A1-bundle over the second term on the right hand side, where the A1-coordinate is the extension class. The statement that we would like to show, saying that there is a natural decomposition as a direct product, is a sort of commutativity property. Let Rep(C1, . . . , Ci; R)u denote the subspace of Rep(C1, . . . , Ci; R) consisting of representations which are unstable on Si. This is equiva- lent to saying that Ai fixes, and acts by c−1 on the eigenvector of R of eigenvalue bi. We will show an isomorphism i Rep(C1, . . . , Ci; R)u ∼= A1 × Rep(C1, . . . , Ci−1; R′), and this isomorphism will preserve the conditions (σ·, G·) over Si−1 so it restricts to an isomorphism between the moduli spaces as claimed in the theorem. Write R =(cid:18) b−1 i 0 0 bi (cid:19) . 26 C. SIMPSON Then Rep(C1, . . . , Ci; R)u is the space of sequences (A1, . . . , Ai) such that A1 · · · AiR = 1 and (8.1) for some y ∈ A1. Similarly, write y Ai =(cid:18) ci R′ =(cid:18) b−1 i−1 0 0 c−1 i (cid:19) bi−1 (cid:19) , 0 A′ 1 · · · A′ i−1R′ = 1. and Rep(C1, . . . , Ci−1; R′) is the space of sequences (A′ that 1, . . . , A′ i−1) such Suppose (A1, . . . , Ai) is a point in Rep(C1, . . . , Ci; R)u and let y ∈ A1 i bi = bi−1 so be the lower left coefficient of Ai from (8.1). Note that c−1 Let AiR =(cid:18) b−1 i ci b−1 i y i y bi−1 (cid:19) . i−1 b−1 0 0 c−1 i bi (cid:19) =(cid:18) b−1 U :=(cid:18) 1 0 u 1 (cid:19) be chosen so that UAiRU −1 = R′, which happens if and only if in other words b−1 i−1u + b−1 i y − bi−1u = 0, u := −b−1 i y b−1 i−1 − bi−1 . The denominator is nonzero because we are assuming the trace of Gi−1 is different from 2 or −2, which is equivalent to asking bi−1 6= b−1 i−1. Then put A′ j := UAjU −1. From the equation UAiRU −1 = R′ we get A′ 1 · · · A′ i−1R′ = U(A1 · · · Ai−1)U −1(UAiRU −1) = 1. Hence, (y, (A′ defines the map 1, . . . , A′ i−1)) is a point in A1 ×Rep(C1, . . . , Ci−1; R′). This Rep(C1, . . . , Ci; R)u → A1 × Rep(C1, . . . , Ci−1; R′), 1, . . . , A′ Its inverse is obtained by mapping (y, (A′ i−1)) to (A1, . . . , Ai) where for 1 ≤ j ≤ i − 1 we put Aj = U −1A′ jU with U defined as above using y, and Ai is the upper triangular matrix (8.1). We obtain the claimed isomorphism. (cid:3) DUAL BOUNDARY COMPLEX 27 By symmetry the same holds in case of Kostov-genericity on the other side, giving a statement written as M α(S≥i−1; C·)fr,R′ ∼= A1 × M α(S≥i; C·)fr,R. 8.2. Open Gi−1 and Gi. If σi = 0 and the moduli space is nonempty, then we cannot have both sides being Kostov-nongeneric at once. Therefore, the remaining case is when Gi−1 and Gi are open sets which are unions of all but finitely many conjugacy classes (that is to say, allowing all traces but a finite number), such that the moduli problems on both S<i and S>i are Kostov-generic. In this situation, which we now assume, the moduli spaces M α(S<i; C·) and M α(S>i; C·) exist and have principal bundles F<i and F>i respectively. We have a map M α(S<i; C·) × M α(S>i; C·) → Gi−1 × Gi. Consider the etale covering spacefGi which parametrizes matrices with a choice of one of the two eigenspaces. This was considered extensively by Kabaya [27]. Let fM α(S>i; C·) := M α(S>i; C·) ×GifGi and similarly for fM α(S<i; C·). Our hypothesis that σi = {0}, in other words that for any local sys- tem V in M α(S; C·) the restriction is unstable, provides a factorization of the projection map through M α(S; C·) → fM α(S<i; C·) ×fM α(S>i; C·). Indeed the destabilizing rank one subsystem is uniquely determined by the condition that the monodromy around ξi have eigenvalue c−1 , and this rank one subsystem serves to pick out the eigenvalues of the matrices for ρi−1 and ρi. i Now the same argument as before goes through. We may choose a stratification such that on each stratum the principal bundles have framings such that the automorphisms R′ and R are diagonal (note, however, that the eigenvalues are now variable). We reduce to the following situation: Z is a quasiprojective variety with invertible functions bi−1 and bi such that b−1 i bi = 1, and we look at the moduli space of quadruples (z, Vi, β ′, β) such that z ∈ Z, Vi is an unstable local system on Si, and i1 c−1 β ′ : V ρi−1 β : V ρi ∼= (V, R′(z)), ∼= (V, R(z)) 28 C. SIMPSON where R′(z) =(cid:18) bi−1(z)−1 0 0 bi−1(z) (cid:19) R(z) =(cid:18) bi(z)−1 0 0 bi(z) (cid:19) . The map Y = β ′β −1 is an automorphism of V (defined up to scalars, so it is a group element in P GL2) and it preserves the marked subspace, It uniquely determines the data so it is a lower-triangular matrix. ∼= V using (Vi, β, β ′) up to isomorphism. Indeed we may consider Vi for example β ′, then our local system is (R′, Ai, Y RY −1) where Ai is specified by the condition (R′)−1AiY RY −1 = 1. As the group of lower triangular matrices in P GL2 is isomorphic to Gm × Ga we obtain an isomorphism between our stratum and Z × Gm × Ga. Alternatively, one could just do a parametrized version of the proof of Theorem 8.3. 8.3. Synthesis. We may gather together the various cases that have been treated in this section so far. Theorem 8.4. Suppose α is any datum such that for some i we have σi = {0}. If M α(S; C·) is nonempty, then D∂M α(S; C·) ∼ ∗. Proof. Let Gv be the set of matrices R with Tr(R) 6= 2, −2. In the previous section we have treated the cases where any Gi is one of the four conjugacy classes of trace 2 or −2. Therefore we may assume that Gi−1, Gi ⊂ Gv. Suppose that Gi−1 and Gi are conjugacy classes chosen so that the sequences (C1, . . . , Ci−1, Gi−1) and (Gi, Ci+1, . . . , Ck) are both Kostov- nongeneric. Under the hypothesis σi = {0} and supposing M α(S; C·) nonempty, containing say a local system V , then an eigenvalue of Gi−1 is the product of an eigenvalue of Ci and an eigenvalue of Gi, since there exists a rank one subsystem of V Si. The same holds for the other eigenvalue of Gi−1. Combining with the nongenericity relations among eigenvalues of (C1, . . . , Ci−1, Gi−1) and (Gi, Ci+1, . . . , Ck), we obtain a nongenericity relation for (C1, . . . , Ck). This contradicts the hypothesis of Condition 4.3 for C·. Therefore, we conclude that if M α(S; C·) is nonempty, then for any specific choice of conjugacy classes Gi−1 and Gi, at least one of the moduli problems over S<i or S>i has to satisfy Condition 4.3. These cases are then covered by Theorem 8.3 above. There are finitely many choices of single conjugacy classes Gi−1 (resp. Gi) such that (C1, . . . , Ci−1, Gi−1) (resp. (Gi, Ci+1, . . . , Ck)) is Kostov non-generic. We may therefore isolate these choices and treat them by Theorem 8.3 according to the previous paragraph. Let now Gi−1 and Gi be the complement in Gv of these nongeneric conjugacy classes. These are open subsets such that for any conjugacy classes therein, the DUAL BOUNDARY COMPLEX 29 moduli problems on S<i and S>i satisfy Condition 4.3. The discussion of subsection 8.2 now applies to give the conclusion that this part of M α(S, C·) has contractible dual boundary complex. (cid:3) 9. Reduction to M ′ In this section, we put together the results of the previous sections to obtain a reduction to the main biggest open stratum. Recall from Condition 7.1 that we are assuming that C· is very generic. Let the datum α′ consist of the following choices: for all i, σ′ and Gi is the set Gv of matrices with trace 6= 2, −2. Then we put i = {1} M ′ := M α′ (S, C·). It is an open subset of M(S, C·) since stability, and the conditions on the traces, are open conditions. Theorem 9.1. There exists a collection of data denoted αj such that M(S, C·) = M ′ ⊔aj M αj (S, C·) is a stratification, i.e. a decomposition into locally closed subsets ad- mitting a total order satisfying the closedness condition of 2.6. Fur- thermore, this admits a further refinement into a stratification with M ′ together with pieces denoted Z j,a ⊂ M αj (S, C·), such that all of the pieces Z j,a have the form Z j,a = Y j,a × A1. Proof. Let Gv be the set of matrices of trace 6= 2, −2 and let Gu be the set of matrices of trace 2 or −2. Let αj run over the 22k−3 choices of (σ2, . . . , σk−1; G2, . . . , Gk−2) with σi either {0} or {1}, and Gi either Gu or Gv. The locally closed pieces M αj (S, C·) are disjoint and their union is M(S, C·). Furthermore, the set of indices is partially ordered with the product order induced by saying that {0} < {1} and Gu < Gv and j1 ≤ j2 if each component of αj1 is ≤ the corresponding component of αj2. If J is a downward cone in this partial ordering thenSj∈J M αj (S, C·) is closed, because specialization decreases the indices (stable specializes to unstable and Gv specializes to Gu). Choosing a compatible total ordering we obtain the required closedness property. The highest element in the partial ordering is the datum α′ consid- ered above, so M ′ is the open stratum of the stratification. The discussion of the previous two sections allows us to further decompose all of the other strata M αj (S, C·), in a way which again preserves the ordered closedness condition, into pieces of the form Z j,a = Y j,a × A1. (cid:3) 30 C. SIMPSON Corollary 9.2. The natural map D∂M(S, C·) → D∂M ′ is a homotopy equivalence. Proof. Apply Proposition 2.6 to the stratification given by the theorem. Note that M ′ is nonempty and the full moduli space is irreducible so the other strata are subvarieties of strictly smaller dimension. (cid:3) 10. Fenchel-Nielsen coordinates We are now reduced to the main case M ′ = M α(S; C·) for α′ such that all σi = {1} and all Gi = Gv. We would like to get an expression for M ′ allowing us to understand its dual boundary complex by in- spection. We will show M ′ ∼= Qk−3 where Q is defined near the end of this section, such that D∂(Q) ∼ S1. The conclusion D∂M ′ ∼ S2(k−3)−1 then follows from Lemma 2.4. This product decomposition is a system of Fenchel-Nielsen coordi- nates for the open subset M ′ of the moduli space. One of the main things we learn from the basic theory of the clas- sical hypergeometric function is that a rank two local system on P1 − {0, 1, ∞} is heuristically determined by the three conjugacy classes of the monodromy transformations at the punctures. This general princi- ple is not actually true, in cases where there might be a reducible local system. But, imposing the condition of stability provides a context in which this rigidity holds precisely. This is the statement of Corollary 10.3 below. Let ti−1 and ti be points in A1 − {2, −2}. We will write down a stable local system Vi(ti−1, ti) on Si, whose monodromy traces around ρi−1 and ρi are ti−1 and ti respectively, and whose monodromy around ξi is in the conjugacy class Ci. Furthermore, any stable local system with these traces is isomorphic to Vi(ti−1, ti) in a unique way up to scalars. Construct Vi(ti−1, ti) together with a basis at the basepoint xi, by exhibiting monodromy matrices R′ i−1, Ri and Ai in SL2. Set Ai :=(cid:18) ci 0 0 c−1 i (cid:19) and Ri :=(cid:18) ui wi 1 (ti − ui) (cid:19) with ui given by the formula (10.1) to be determined below, and wi := ui(ti − ui) − 1 because of the determinant one condition. We could just write down the formula for ui but in order to motivate it let us first calculate R′ i−1 = AiRi =(cid:18) ciui c−1 i wi ci i (ti − ui) (cid:19) . c−1 DUAL BOUNDARY COMPLEX 31 We need to choose ui such that ti−1 = Tr(R′ i−1) = ciui + c−1 i (ti − ui). This gives the formula (10.1) ui = i ti ti−1 − c−1 ci − c−1 i . The denominator is nonzero since by hypothesis ci 6= c−1 i . Lemma 10.1. Suppose Vi is an SL2 local system with traces ti−1 and ti. Suppose Vi is given a frame at the base point xi, such that the monodromy matrix around the loop γi is diagonal with ci in the upper left, and such that the monodromy matrix around ρi (via the path going from xi to si ∈ ρi) has a 1 in the upper right corner. Then the three monodromy matrices of Vi are the matrices R′ i−1, Ri and Ai defined above. Proof. The matrix Ai is as given, by hypothesis. The matrix Ri has trace ti and upper right entry 1 by hypothesis, so it too has to look as given. Now the calculation of the trace ti−1 as a function of ui has a unique inversion: the value of ui must be given by (10.1) as a function of ti−1, ti and ci. This determines the matrices. (cid:3) Lemma 10.2. Suppose Vi is an SL2 local system with traces ti−1 and ti different from 2 or −2, and suppose Vi is stable. Then, up to a scalar multiple, there is a unique frame for Vi over the basepoint xi satisfying the conditions of the previous lemma. i Proof. Let e1 and e2 be eigenvectors for the monodromy around γi, with eigenvalues ci and c−1 respectively. They are uniquely determined up to a separate scalar for each one. We claim that the upper right entry of the monodromy around ρi is nonzero. If it were zero, then the subspace generated by e2 would be fixed, with the monodromy around ξi being c−1 i ; that would contradict the assumption of stability. Now since the upper right entry of the monodromy around ρi is nonzero, we may adjust the vectors e1 and e2 by scalars such that this entry is equal to 1. Once that condition is imposed, the only further allowable change of basis vectors is by multiplying e1 and e2 by the same scalar. (cid:3) Corollary 10.3. Suppose Vi is a local system on Si, with conjugacy class Ci around ξi, stable, and whose traces around ρi−1 and ρi are ti−1 and ti respectively. Then there is up to a scalar an unique isomorphism Vi ∼= Vi(ti−1, ti) with the system constructed above. 32 C. SIMPSON Suppose V is a point in M ′, and let ti denote the traces of the monodromies of V around the loops ρi. Then by the definition of the datum α′, ti 6= 2, −2 and the restriction to each Si is stable, so by ∼= the corollary there is up to scalars a unique isomorphism hi : V Si Vi(ti−1, ti). Recall that xi is a basepoint in Si, and that we have chosen a path in Si from xi to a basepoint si in ρi, and then a path in Si+1 from si to xi+1. Let ψi denote composed the path from xi to xi+1, and use the same symbol to denote the transport along this path which is an ∼= Vxi+1. The stalk of the local system Vi(ti−1, ti) isomorphism ψi : Vxi at xi is by construction C2, and the same at xi+1, so the map Pi := hi+1ψih−1 i : Vi(ti−1, ti)xi → Vi+1(ti, ti+1)xi+1 is a matrix Pi : C2 → C2 well-defined up to scalars, that is Pi ∈ P GL2. By the factorization property of M ′, the local system V is determined by these glueing isomorphisms Pi, subject to the constraint that they should intertwine the monodromies around the circle ρi for Vi and Vi+1. We have used the notation R′ i for the monodromy of the local system Vi+1 around the circle ρi, whereas Ri denotes the monodromy of Vi around here. We will have made sure to use the same paths from xi or xi+1 to the basepoint si ∈ ρi in order to define these monodromy matrices as were combined together to make the path ψi. Therefore, the compatibility condition for Pi says (10.2) R′ i ◦ Pi = Pi ◦ Ri. The frames for Vxi are only well-defined up to scalars, so the matri- ces Pi are only well-defined up to scalars and conversely if we change them by scalars then it doesn’t change the isomorphism class of the local system. Putting together all of these discussions, we obtain the following preliminary description of M ′. Lemma 10.4. The moduli space M ′ is isomorphic to the space of (t2, . . . , tk−2) ∈ (A1 −{2, −2})k−3 and (P2, . . . , Pk−2) ∈ (P GL2)k−3 sub- ject to the equations (10.2), where R′ i and Ri are given by the previous formulas in terms of the tj. For the end pieces, one should formally set t1 := c1 + c−1 1 and tk−1 := ck + c−1 k . At this point, we have not yet obtained a good “Fenchel-Nielsen” style coordinate system, because the equation (10.2) for Pi contains R′ i which depends on ti+1 as well as ti, and Ri which depends on ti−1 as well as ti. DUAL BOUNDARY COMPLEX 33 We now proceed to decouple the equations. The strategy is to intro- duce the matrices Ti :=(cid:18) 0 −1 ti (cid:19) 1 0 Ui :=(cid:18) 1 ui 1 (cid:19) −1 ti (cid:19)(cid:18) 1 1 which serve as a canonical normal form for matrices with given traces ti, not requiring the marking of one of the two eigenvalues. Notice that if we set then U −1 i TiUi =(cid:18) 1 −ui 1 (cid:19)(cid:18) 0 0 0 ui 1 (cid:19) =(cid:18) ui wi 1 ti − ui (cid:19) with wi as before. Therefore, using the formula (10.1) for ui we may write Now Ri(ti−1, ti) = U −1 i TiUi. R′ i−1 = AiRi = AiU −1 i TiUi = U −1 i Ti)Ui. is lower triangular with ci and c−1 (UiAiU −1 i Furthermore, UiAiU −1 along the diagonal, and when we multiply with Ti it gives a matrix of the form i i U −1 i AiUiTi =(cid:18) ci 0 ∗ c−1 i (cid:19)(cid:18) 0 −1 ti (cid:19) =(cid:18) 0 1 −c−1 i ci ∗ (cid:19) . However, we know that ui was chosen so that this matrix has trace ti−1 (it is conjugate to R′ i−1), therefore in fact U −1 i AiUiTi =(cid:18) 0 −c−1 i ci ti−1 (cid:19) as could alternately be seen by direct computation. By inspection this matrix is conjugate to Ti−1 as it should be from its trace. Interestingly enough, the conjugation is by the matrix 1 A 2 i with := c 1 2 i 0 i ! , 0 − 1 2 c U −1 i AiUiTi = A 1 2 − 1 i Ti−1A 2 i . This half-power seems also to occur somewhere in the classical treat- ments of the Fenchel-Nielsen coordinates, We obtain R′ i−1 = U −1 i (UiAiU −1 i Ti)Ui = U −1 i A i Ti−1A 1 2 − 1 i Ui. 2 34 C. SIMPSON Recall that the equation (10.2) for Pi−1 reads R′ i−1 ◦ Pi−1 = Pi−1 ◦ Ri−1, and using the above formula for R′ this equation reads i−1 as well as Ri−1 = U −1 i−1Ti−1Ui−1, 1 2 U −1 i A i Ti−1A − 1 i Ui ◦ Pi−1 = Pi−1 ◦ U −1 2 i−1Ti−1Ui−1. (10.3) Set Qi−1 := A − 1 i UiPi−1U −1 i−1. 2 This is a simple change of variables of the matrix Pi−1, with the ma- trices entering into the change of variables depending however on ti−2, ti−1 and ti. Notice that the coefficients of Qi−1 are linear functions of the coefficients of Pi−1, in particular the action of scalars is the same on both. Our equation which was previously (10.2) (but for i − 1 instead of i), has become (10.3) which, after multiplying on the left by Ui then − 1 by A 2 i i−1 and substituting Qi−1, becomes: and on the right by U −1 (10.4) Ti−1 ◦ Qi−1 = Qi−1Ti−1. A sequence of matrices Qi satisfying these equations leads back to a sequence of matrices Pi satisfying (10.2) and vice-versa. Recall that the glueing for the local system depended on these matrices modulo scalars, that is to say in P GL2. We may sum up with the following proposition: Proposition 10.5. The moduli space M ′ is isomorphic to the space of choices of (t2, . . . , tk−2) ∈ (A1 − {2, −2})k−3 and (Q2, . . . , Qk−2) ∈ (P SL2)k−3 subject to the equations TiQi = QiTi. This expression for the moduli space is now decoupled, furthermore the equations are in a nice and simple form. Theorem 10.6. Let Q be the space of pairs (t, [p : q]) ∈ A1 × P1 such that t 6= 2, −2 and (10.5) Then we have p2 + tpq + q2 6= 0. M ′ ∼= Qk−3. DUAL BOUNDARY COMPLEX 35 Proof. This will follow from the previous proposition, once we calculate that the space of matrices Qi in P GL2 commuting with Ti, is equal to the space of points [p, q] ∈ P1 such that p2 + tipq + q2 6= 0. Write then whereas q p′ q′ (cid:19) Qi =(cid:18) p q′ (cid:19)(cid:18) 0 −1 ti (cid:19) =(cid:18) −q q′ (cid:19) =(cid:18) −1 ti (cid:19)(cid:18) p p′ p′ 1 q QiTi =(cid:18) p TiQi =(cid:18) 0 q 1 p + tiq −q′ p′ + tiq′ (cid:19) tip′ − p tiq′ − q (cid:19) . p′ q′ The equation QiTi = TiQi thus gives from the top row and then, those actually make the other two equations hold automati- cally. Therefore a solution Qi may be written p′ = −q, q′ = p + tiq Qi =(cid:18) p −q p + tiq (cid:19) . q The statement Qi ∈ P GL2 means that Qi is taken up to multiplica- tion by scalars, in other words [p : q] is a point in P1 (clearly those coordinates are not both zero); and det(Qi) = p2 + tipq + q2 6= 0. We conclude that the space of (ti, Qi) ∈ (A1 − {2, −2}) × P GL2 such that TiQi = QiTi is isomorphic to Q. Therefore Proposition 10.5 now says M ′ ∼= Qk−3. (cid:3) Lemma 10.7. The dual boundary complex of Q is Therefore D∂Q ∼ S1. D∂Qk−3 ∼ S2(k−3)−1. Proof. Let Φ ⊂ P1 × A1 be the open subset defined by the same inequa- tion (10.5). Then Q ⊂ Φ is an open subset, whose complement is the disjoint union of two affine lines. Furthermore, Φ := P1 × P1 is a (non simple) normal crossings compactification of Φ. The divisor at infinity is the union of two copies of P1, namely the fiber over t = ∞ and the conic defined by p2 + tpq + q2 = 0. These intersect transversally in two points. Therefore, the incidence complex of Φ ⊂ Φ at infinity is a graph with two vertices and two edges joining them. 36 C. SIMPSON It follows that the incidence complex at infinity for Q is a circle. That may also be seen directly by blowing up two times over each ramification point of the conic lying over t = ±2. Now applying Lemma 2.4 successively, and noting that the successive join of k−3 times the circle is S2(k−3)−1, we obtain the second statement. (cid:3) Corollary 10.8. Let C· be a collection of conjugacy classes satisfying Condition 7.1. Then the moduli space MB(S; C1, . . . , Ck) of rank 2 local systems with those prescribed conjugacy classes, has dual boundary complex homotopy equivalent to a sphere D∂MB(S; C1, . . . , Ck) ∼ S2(k−3)−1. Proof. We have been working with the hybrid moduli stack M(S; C·) above, but Proposition 4.5 says that this is the same as the moduli space MB(S; C1, . . . , Ck). By Corollary 9.2, D∂M(S; C·) ∼ D∂M ′. By Theorem 10.6, M ′ ∼= Qk−3, and by Lemma 10.7 D∂Qk−3 ∼ S2(k−3)−1. Putting these all together we obtain the desired conclusion. (cid:3) This completes the proof of Theorem 1.1. Remark 10.9. The space Φk−3 itself has a modular interpretation: it is M α(S, C·) for α given by setting all σi to {1} (requiring stability of each V Si), but having Gi = GL2 for all i, that is no longer constraining the traces. 11. A geometric P = W conjecture In this section we discuss briefly the relationship between the theo- rem proven above, and the Hitchin fibration. For this discussion, let us suppose that the eigenvalues ci are ni-th roots of unity, so the con- jugacy classes Ci have finite order ni. Fix points y1, . . . , yk ∈ P1 and let X := P1[ 1 n1 y1, . . . , yk] 1 nk be the root stack with denominators ni at the points yi respectively. It is a smooth proper Deligne-Mumford stack. The fundamental group of its topological realization [41] is generated by the paths γ1, . . . , γk subject to the relations that γ1 · · · γk = 1 and γni i = 1. We may also let S be a punctured sphere such as considered above, the complement of a collection of small discs in P1 centered at the points yi. Therefore, a local system on X top is the same thing as a local system on S such that the monodromies around the boundary loops ξi have order ni DUAL BOUNDARY COMPLEX 37 respectively. We have MB(X top, GLr) = a(C1,...,Ck) MB(S, C·) where the disjoint union runs over the sequences of conjugacy classes such that Ci has order ni. Recall that if we assume that C· satisfies the Kostov-genericity condition then the character variety with fixed conjugacy classes MB(S, C·) is the same as the hybrid moduli stack M(S, C·). It may be seen as a connected component of the character variety MB(X top, GLr). Now we recall that there is a homeomorphism between the char- acter variety MB(X top, GLr) and the Hitchin-Nitsure moduli space MDol(X top, GLr) of Higgs bundles. One may consult for example [44], [32], [38] for the general theory in the open or orbifold setting. We denote by MDol(S, C·) the connected component of MDol(X top, GLr) corresponding to the choice of conjugacy classes, which it may be re- called corresponds to fixing appropriate parabolic weights for the par- abolic Higgs bundles. Hitchin’s equations give a homeomorphism , the “nonabelian Hodge correspondence” (11.1) MDol(S, C·)top ∼= MB(S, C·)top. Recall that the resulting two complex structures on the same underly- ing moduli space, form a part of a hyperkahler triple [24]. In the smooth proper orbifold setting we have the same theory of the Hitchin map MDol(S, C·) → An which is a Lagrangian fibration to the space of integrals of Hitchin’s Hamiltonian system [23]. In particular, n is one-half of the complex dimension of the moduli space, that dimension being even because of the hyperkahler structure. Fix a neighborhood of infinity in the Hitchin base B∗ ⊂ An, and let N ∗ Dol denote its preimage in MDol(S, C·). Similarly, let N ∗ B denote a neighborhood of infinity in MB(S, C·). The homeomorphism 11.1 gives a natural homotopy equivalence N ∗ Dol ∼ N ∗ B. The neighborhood at infinity B∗ ⊂ An has the homotopy type of the sphere S2n−1, and indeed we may view S2n−1 as the quotient of B∗ by radial scaling, so the Hitchin map provides a natural map N ∗ Dol → S2n−1. On the other hand, there is a natural projection N ∗ B → D∂MB(S, C·). This is a general phenomenon, indeed if we have chosen a very simple normal crossings compactification with divisor components D1, . . . , Dm 38 C. SIMPSON then we may choose an open covering of N ∗ B by open subsets U1, . . . , Um punctured neighborhoods of the Di, such that Ui1 ∩· · ·∩Uir is nonempty if and only if Di1 ∩ · · · ∩ Dir is nonempty. Then, any partition of unity B → Rm which just goes into the for this covering provides a map N ∗ subspace D∂MB(S, C·). Recall the following conjecture [28], which was motivated by consid- eration of the case P1 − {y1, y2, y3, y4}. Conjecture 11.1. There is a homotopy-commutative square ∼→ N ∗ Dol ↓ N ∗ B ↓ S2n−1 ∼→ D∂MB(S, C·) where the top and side maps are those described above, such that the bottom map is a homotopy equivalence. Our main theorem provides a homotopy equivalence such as the one which is conjectured to exist on the bottom of the square, for the group GL2 on P1 − {y1, . . . , yk}. This was our motivation, and it was also the motivation for Komyo’s proof in the case k = 5 [31]. We haven’t shown anything about commutativity of the diagram. This is one of the motivations for looking at the geometric theory of harmonic maps to buildings developed in [28] [29]. A result in this direction is shown by Daskalopoulos, Dostoglou and Wentworth [3]. The Kontsevich-Soibelman wallcrossing picture [34] should provide a global framework for this question. Conjecture 11.1 may be viewed as a geometrical analogue of the first weight-graded piece of the P = W conjecture [5] [16]. That conjecture states that weight filtration W of the mixed Hodge structure on the cohomology of the character variety MB should be naturally identified with the perverse Leray filtration P induced by the Hitchin fibration. For the case of rank two character varieties on a compact Riemann surface, it was in fact proved by de Cataldo, Hausel and Migliorini [5]. Davison treats a twisted version [4]. It is known [43] that the cohomology of the dual boundary complex is the first weight-graded piece of the cohomology of MB. Conjecture 11.1 states that this should come from the cohomology of the sphere at infinity in the Hitchin fibration, which looks very much like a Leray piece. Furthermore, indeed from discussions with L. Migliorini and S. Payne it seems to be the case that the characterization of the cohomology of the dual boundary complex in [43], and the computations [19] [20] [17] [18] of the cohomology ring of MDol used to prove the P = W conjecture DUAL BOUNDARY COMPLEX 39 for SL2 in [5], should serve to show commutativity of the diagram in rational cohomology. The question of proving the analogue of our Theorem 1.1 for a com- pact Riemann surface, even in the rank 2 case, is an interesting problem for further study. One may also envision the case of a punctured curve of higher genus. The techniques used here involved a choice of sta- bility condition on each of the pieces of the decomposition, which in the higher genus case would require having at least a certain number of punctures. Weitsman suggests, following [51] and [26], that it might be possible to obtain a similar argument with only at least one puncture. The compact case would seem to be more difficult to handle. Let us note that Kabaya [27] gives a general discussion of coordinate systems which can be obtained using decompositions, and he treats the problems of indeterminacy of choices of eigenspaces up to permutations. The other direction which needs to be considered is local systems of higher rank. Here, the first essential case is P1 − {0, 1, ∞}, where there is no useful decomposition of the surface into simpler pieces. We could hope that if this basic case could be treated in all ranks, then the reduction techniques we have used above could allow for an extension to the case of many punctures. References 1. V. G. Berkovich. Spectral Theory and Analytic Geometry over non-Archimedean fields, Mathematical Surveys and Monographs 33, AMS, Providence (1990). 2. V. Danilov. Polyhedra of schemes and algebraic varieties. Mathematics of the USSR-Sbornik 26 (1975), 137-149. 3. G. Daskalopoulos, S. Dostoglou, R. Wentworth. On the Morgan-Shalen com- pactification of the SL(2, C) character varieties of surface groups. Duke Math. J. 101 (2000), 189-207. 4. B. Davison. Cohomological Hall algebras and character varieties. Preprint arXiv:1504.00352 (2015). 5. M. A. de Cataldo, T. Hausel, L. Migliorini. Topology of Hitchin systems and Hodge theory of character varieties: the case A1. Ann. of Math. 175 (2012), 1329-1407. 6. T. de Fernex, J. Koll´ar, C. Xu. The dual complex of singularities. Preprint arXiv:1212.1675 (2012). 7. J.M. Dr´ezet. Luna’s slice theorem and applications. Algebraic group actions and quotients, J. A. Wisniewski, ed., Hindawi (2004), 39-90. 8. W. Fenchel, J. Nielsen. Discontinuous groups of non-Euclidean motions. Un- published manuscript. 9. J. Francis, D. Gaitsgory. Chiral Koszul duality. Selecta Mathematica 18 (2012), 27-87. 10. E. Frenkel, D. Ben-Zvi. Vertex algebras and algebraic curves. Mathematical Surveys and Monographs 88, AMS, Providence (2001). 40 C. SIMPSON 11. D. Gaiotto, G.W. Moore, A. Neitzke. Spectral networks. Annales Henri Poincar´e 14 (2013), 1643-1731. 12. L. Godinho, A. Mandini. Hyperpolygon spaces and moduli spaces of parabolic Higgs bundles. Advances in Mathematics 244 (2013), 465-532. 13. W. Goldman. The complex-symplectic geometry of SL(2, C)-characters over surfaces. Preprint arXiv math/0304307 (2003). 14. M. Gross, P. Hacking, S. Keel. Mirror symmetry for log Calabi-Yau surfaces I. Preprint arXiv:1106.4977 (2011). 15. M. Gross, P. Hacking, S. Keel, M. Kontsevich. Canonical bases for cluster algebras. Preprint arXiv:1411.1394 (2014). 16. T. Hausel. Global topology of the Hitchin system. Preprint arXiv:1102.1717 (2011). 17. T. Hausel, E. Letellier, F. Rodriguez-Villegas. Arithmetic harmonic analysis on character and quiver varieties, Duke Math. J. 160 (2011) 323-400. 18. T. Hausel, E. Letellier, F. Rodriguez-Villegas. Arithmetic harmonic analysis on character and quiver varieties II, Adv. Math. 234 (2013), 85-128. 19. T. Hausel, M. Thaddeus. Relations in the cohomology ring of the moduli space of rank 2 Higgs bundles. J.A.M.S. 16 (2003), 303-329. 20. T. Hausel, M. Thaddeus. Generators for the cohomology ring of the moduli space of rank 2 Higgs bundles. Proc. London Math. Soc. 88 (2004), 632-658. 21. T. Hausel, F. Rodriguez-Villegas. Mixed Hodge polynomials of character vari- eties. Invent. Math. 174 (2008), 555-624. 22. V. Hinich, V. Schechtman. On homotopy limit of homotopy algebras. K-theory, Arithmetic and Geometry, Springer (1987), 240-264. 23. N. Hitchin. Stable bundles and integrable systems. Duke Math. J. 54 (1987), 91-114. 24. N. Hitchin. The self-duality equations on a Riemann surface. Proc. London Math. Soc. 55 (1987), 59-126. 25. L. Hollands, A. Neitzke. Spectral networks and Fenchel-Nielsen coordinates. Preprint arXiv:1312.2979 (2013). 26. L. Jeffrey, J. Weitsman. Toric structures on the moduli space of flat connections on a Riemann surface II: Inductive decomposition of the moduli space. Math. Annalen 307 (1997), 93-108. 27. Y. Kabaya. Parametrization of PSL(2, C)-representations of surface groups. Geometriae Dedicata 170 (2014), 9-62. 28. L. Katzarkov, A. Noll, P. Pandit, C. Simpson. Harmonic maps to buildings and singular perturbation theory. Comm. Math. Physics 336 (2015), 853-903. 29. L. Katzarkov, A. Noll, P. Pandit, C. Simpson. Constructing buildings and har- monic maps. Preprint arXiv:1503.00989 (2015). 30. J. Koll´ar, Chenyang Xu. The dual complex of Calabi-Yau pairs. Preprint arXiv:1503.08320 (2015). 31. A. Komyo. On compactifications of character varieties of n-punctured projective line. Preprint arXiv:1307.7880 (2013). 32. H. Konno. Construction of the moduli space of stable parabolic Higgs bundles on a Riemann surface. J. Math. Soc. Japan 45 (1993), 253-276. 33. M. Kontsevich, Y. Soibelman. Homological mirror symmetry and torus fibra- tions. Symplectic geometry and mirror symmetry (Seoul, 2000), World Sci. Pub- lishing (2001), 203-263. DUAL BOUNDARY COMPLEX 41 34. M. Kontsevich, Y. Soibelman. Wall-crossing structures in Donaldson-Thomas invariants, integrable systems and Mirror Symmetry. Homological Mirror Sym- metry and Tropical Geometry, R. Castano-Bernard et al eds., Springer (2014), 197-308. 35. V. Kostov. On the Deligne-Simpson problem. Preprint arXiv math/0011013 (2000). 36. E. Letellier. Character varieties with Zariski closures of GLn-conjugacy classes at punctures. Selecta 21 (2015), 293-344. 37. C. Manon. Toric geometry of SL2(C) free group character varieties from outer space. Preprint arXiv:1410.0072 (2014). 38. H. Nakajima. Hyper-Kahler structures on moduli spaces of parabolic Higgs bundles on Riemann surfaces. Moduli of vector bundles (Sanda, Kyoto 1994), M. Maruyama ed., Lecture notes in pure and applied math. (1996), 199- 208. 39. N. Nekrasov, A. Rosly, S. Shatashvili. (2011). Darboux coordinates, Yang-Yang functional, and gauge theory. Nuclear Physics B-Proceedings Supplements 216 (2011), 69-93. 40. J. Nicaise, Chenyang Xu. The essential skeleton of a degeneration of algebraic varieties. Preprint arXiv:1307.4041 (2013). 41. B. Noohi. Fundamental groups of algebraic stacks. J. Inst. Math. Jussieu 3 (2004), 69-103. 42. J. Parker, I. Platis. Complex hyperbolic FenchelNielsen coordinates. Topology 47 (2008), 101-135. 43. S. Payne. Boundary complexes and weight filtrations. Michigan Math. J. 62 (2013), 293-322. 44. C. Simpson. Local systems on proper algebraic V -manifolds. Pure and Appl. Math. Quarterly (Eckart Viehweg’s volume), 7 (2011), 1675-1760. 45. A. Soibelman. The moduli stack of parabolic bundles over the projec- tive line, quiver representations, and the Deligne-Simpson problem. Preprint arXiv:1310.1144 (2013). 46. D.A. Stepanov. A remark on the dual complex of a resolution of singularities. Uspekhi Mat. Nauk 61 (367), (2006), 185-186. 47. D.A. Stepanov. A note on resolution of rational and hypersurface singularities. Proc. Amer. Math. Soc. 136 (2008), 2647-2654. 48. S.P. Tan. Complex Fenchel-Nielsen coordinates for quasi-Fuchsian structures. International J. Math. 5 (1994), 239-251. 49. A. Thuillier, G´eom´etrie toroıdale et g´eom´etrie analytique non archim´edienne. Application du type d’homotopie de certains sch´emas formels. manuscripta math. 123 (2007), 381-451. 50. V. Schechtman, A. Varchenko. Hypergeometric solutions of Knizhnik- Zamolodchikov equations. Lett. Math. Phys. 20 (1990), 279-283. 51. J. Weitsman. Geometry of the intersection ring of the moduli space of flat connections and the conjectures of Newstead and Witten. Topology 37 (1998), 115-132. 52. S. Wolpert. The Fenchel-Nielsen deformation. Annals of Math. (1982), 501-528. 42 C. SIMPSON CNRS, Laboratoire JAD, UMR 7351 Universit´e Nice Sophia Antipolis 06108 Nice Cedex 2 France
1003.5200	2	1003	2011-07-12T13:57:21	Hori--Vafa mirror models for complete intersections in weighted projective spaces and weak Landau--Ginzburg models	[ "math.AG" ]	We prove that Hori--Vafa mirror models for smooth Fano complete intersections in weighted projective spaces admit an interpretation as Laurent polynomials.	math.AG	math	Hori -- Vafa mirror models for complete intersections in weighted projective spaces and weak Landau -- Ginzburg models VICTOR PRZYJALKOWSKI Abstract. We prove that Hori -- Vafa mirror models for smooth Fano complete intersec- tions in weighted projective spaces admit an interpretation as Laurent polynomials. Mirror symmetry of variations of Hodge structures states that for any smooth Fano va- riety X there exists a dual Landau -- Ginzburg model f : Y → C such that an essential part of the regularized quantum differential equation for X is of Picard -- Fuchs type. In other words, the solutions of a certain differential equation (constructed via genus 0 Gromov -- Witten invariants for X -- the numbers which count rational curves lying on X) are the periods of the dual family (for more details and references see [Prz09]). By definition, the relevant Picard -- Fuchs differential equation depends only on relative birational type of Y . If one assumes Y = (C∗)N one can translate mirror correspondence to the quantitative level, that is, to combinatorics of Laurent polynomials. Then f may be represented by a Laurent polynomial, which is called a (very) weak Landau -- Ginzburg model. The follow- ing conjecture states that this hypothesis is not very restrictive, particulary for the case of Pic X = Z. Conjecture 1 ([Prz09]). Any smooth Fano variety of dimension N with Picard rank 1 has a weak Landau -- Ginzburg model f ∈ C[x±1 1 , . . . , x±1 N ]. This conjecture holds for threefolds, complete intersections in projective spaces, Grass- mannians and some complete intersections therein, varieties admitting degenerations to"good" toric varieties (for more details see [Prz09]). In the paper we prove that it also holds for smooth complete intersections of Cartier divisors in weighted projective spaces. That is, we prove that Hori -- Vafa suggestions for Landau -- Ginzburg models for such varieties may be interpreted as Laurent polynomials. The similar problems (for weighted projective spaces in an orbifold setup) are studied in [Dou06] and [DM09]. 1. Hori-Vafa models We give some basic definitions and notions about weighted projective spaces and com- plete intersections therein mostly following [Dol82]. We consider weighted projective spaces as projective varieties (not as smooth stacks). We denote by (a1, . . . , ar) the greatest common divisor of a1, . . . , ar ∈ N. Definition 2. A weighted projective space P(w0, . . . , wn) is called normalized if for any i we have (w0, . . . , wi, . . . , wn) = 1 and w0 ≤ w1 ≤ . . . ≤ wn. Remark 3. It is easy to see that P(w0, dw1 . . . , dwn) ∼= P(w0, w1 . . . , wn), so any weighted projective space is isomorphic to a unique normalized one. The work was partially supported by FWF grant P20778, RFFI grants 11-01-00336-a and 11-01-00185- a, grants NSh−4713.2010.1, MK-503.2010.1, and AG Laboratory GU-HSE, RF government grant, ag. 11 11.G34.31.0023. 1 Definition 4. The zero set of (weighted) homogenous polynomial f ∈ C[x0, . . . , xn], wt(xi) = wi, of weighted degree d is called a hypersurface of degree d in P = P(w0, . . . , wn). As the rank of the Weil group of a weighted projective space is 1, any effective Weil divisor is proportional to the zero locus of some weighted homogenous polynomial. Its degree is called the degree of the divisor. It is easy to see that a Weil divisor of degree d is Cartier if and only if d is integral and all wi's divide d. The singular locus of P is the union of subvarieties of form {xi1 = . . . = xir = 0}, where wi1, . . . , wir is a minimal collection of weights such that the rest of the weights have common prime divisor. Consider a complete intersection X = X1 ∩ . . . ∩ Xk, where X1, . . . , Xk are Cartier divisors. It is quasismooth as a complete intersection of weighted Fermat hypersurfaces is quasismooth. By Proposition 8 in [Dim86] together with Propo- sition 2 in loc. cit. the singularities of X are the intersection of X with the singularities of P. In particular X is smooth if and only if the maximal dimension of the strata of singularities of P is less then k. This means that (wi1, . . . , wik+1) = 1 for any collection of weights wi1, . . . , wik+1 (cf. [Dim86]). Let deg Xi = di. The canonical sheaf of X is O(d1 + . . . + dk − w0 − . . . − wn)X. So X is Fano if and only if P di < P wj. Definition 5 ([HV00], see also [Giv96]). Consider a smooth complete intersection X = X1 ∩ . . . ∩ Xk in P(w0, . . . , wn) such that Xi is a Cartier divisor of degree di and there are k non-intersecting subsets Ii ⊂ {0, . . . , n} such that Pj∈Ii wj = di (we call this splitting a Q-nef-partition1). Then a Hori -- Vafa model for X is an affine variety n = 1, 0 · . . . · xwn (cid:26) xw0 Pj∈Ii xj = 1 with function (potential) f = x0 + . . . + xn. Up to a shift f 7→ f − k we can define the potential as the sum of variables whose indices do not lie in any Ii's. 2. Weak Landau -- Ginzburg models We define a (very) weak Landau -- Ginzburg models for Fano variety following [Prz08]. Let X be a smooth Fano variety. Given its Gromov -- Witten invariants (the numbers which count rational curves lying on X) one can construct the so called regularized quan- tum differential equation for X. This equation for a complete intersection in weighted projective spaces is of type DN and has a unique normalized analytic solution I X H 0(t) called the constant term of regularized I-series. Consider a Laurent polynomial f ∈ C[x±1 n ]. Let bi be the constant term of f i. The series Φf (t) = 1 + b1t + b2t2 + . . . is called the constant terms series for f . It is an analytic solution of the Picard -- Fuchs differential equation for a pencil of hypersurfaces in the torus given by f . 1 , . . . , x±1 Definition 6. The polynomial f (or the pencil of hypersurfaces associated to it) is called a very weak Landau -- Ginzburg model for X if the regularized quantum differential equation for X coincides with the Picard -- Fuchs equation for f , or, equivalently, if I X H 0 = Φf . It is called a weak Landau -- Ginzburg model for X if, in addition, the general element of the pencil given by f is birational to a Calabi -- Yau variety. 1It is called a nef-partition in Gorenstein case. 2 Proposition 7. Let X be a smooth complete intersection of Cartier divisors of degrees d1, . . . , dk in normalized weighted projective space P(w0, . . . , wn). Assume that X is a Fano variety. Then there are k non-intersecting subsets I1, . . . , Ik ⊂ {0, . . . , n} such that di = Pj∈Ii wj for any i and wj = 1 for all j /∈ I1 ∪ . . . ∪ Ik. Proof. We have the following numerical conditions: w0 ≤ w1 ≤ . . . ≤ wn, wjdi, j = 0 . . . n, i = 1, . . . , k, (wi1, . . . , wik+1) = 1, {i1, . . . , ik+1} ⊂ {0, . . . , n}. X di < X wi, Apply the following "reduction process". Let p be a divisor of one of the weights. Divide by p all the degrees and those of the weights which are divisible by p. Up to renumbering of the weights we get a collection of weights and degrees satisfying the conditions above. Repeat the procedure until all weights become equal to 1. Consider k non-intersecting subsets I1, . . . , Ik of {0, . . . , n} whose orders equal the degrees we got on the last step. Let I = {0, . . . , n} \ {I1 ∪ . . . ∪ Ik}. Denote P wj, wj ∈ Ii, by Ii. Start the reduction process in the reverse direction. The weights and the degrees change on each step. Change the elements of I and Ii's on the first step in such a way that each of Ii's contains at most one index of an increasing weight and I contains none of them. Change I and Ii's on each step in the following way: if Ii = di (where di's are the degrees on this step) do nothing. If Ii < di, add indices from I whose weights increase to Ii (it is easy to check that di − Ii is not less then the prime divisor p we increase). As the number of increasing weights is not greater than k we get I containing only indices corresponding to the weight 1. We get Ii ≤ di. If Ii < di, add di − Ii indices from I. Doing such changes of I and Ii's on each step we get Ii = di, and all the weights whose indices lie in I equal 1. Finally we get the partition we need. (cid:3) Remark 8. Let d0 = P wi −P dj be the Fano index of X. It is easy to see from the proof of Proposition 7 that there are actually at least d0 + 1 weights that are equal to 1. This bound is strict. The example is hypersurface of degree 6 in P(1, 1, 2, 3). Theorem 9. Let X be a smooth Fano complete intersection of Cartier divisors in normal- ized weighted projective space P(w0, . . . , wn). Then X has a very weak Landau -- Ginzburg model. Proof. In order to keep the notation simple we prove this theorem for the case when X is a hypersurface; the general case can be proved identically. Prove that the pencil given by the Hori -- Vafa model for X is relative birationally isomorphic to a pencil of hypersurfaces in (C∗)n−1. By Proposition 7, d1 is the sum of some weights such that the rest of the weights equal 1. Renumber weights for convenience such that d1 = wr+1 + . . . + wn. Then do the well-known trick with a projective change of coordinates for Hori -- Vafa model for X. That is, the Hori -- Vafa model is with potential f = x0 + . . . + xr. Consider (projective) change of coordinates (cid:26) xw0 0 · . . . · xwn n = 1, xr+1 + . . . + xn = 1 xi = yi yr+1 + . . . + yn 3 , i = r + 1, . . . , n. The second equation of the system disappears. As w0 = 1 we may rewrite the first variable as (yr+1 + . . . + yn)d1 · . . . · xwr r ywr+1 r+1 · . . . · ywn n x0 = xw1 1 . In the local chart, say, yn = 1 we finally get a very weak Landau -- Ginzburg model fX = xw1 1 (yr+1 + . . . + yn−1 + 1)d1 · . . . · xwr r r+1 · . . . · ywn−1 · ywr+1 n−1 + x1 + . . . + xr. The constant term of regularized I-series for X is given by I X H 0 = ∞ Xm=0 (d0m)!(d1m)! (w0m)! · . . . · (wnm)! td0m (see [Prz07]). One can see that this series coincides with the constant terms series for fX . (cid:3) Remark 10. Let Xd be a smooth hypersurface in a weighted projective space such that its Fano index iX = w0 + . . . + wn − d is one. Let fX be a very weak Landau -- Ginzburg model fX given by Theorem 9. One can prove (see the proof of Theorem 14 in [Prz09]) that a general element of the pencil defined by fX birational to a Calabi -- Yau variety. In other words, fX is actually a weak Landau -- Ginzburg model for X. Problem 11. Prove that this is true for all smooth Fano complete intersections of Cartier divisors in weighted projective spaces. Remark 12. In [?] N. Ilten and the author prove that very weak Landau -- Ginzburg models of Hori -- Vafa type for complete intersections are toric. This means that their Newton polytopes are fan polytopes of toric degenerations of these complete intersections. It seems that the assumptions on varieties we need for Hori -- Vafa mirror models can be weakened. One can consider a complete intersection X which does not intersect the singular locus of P (in a classical setup this condition is necessary, because otherwise X should be considered as an orbifold). On the numerical level this means that for any q > 1 the number of weights divisible by q is not greater then the number of degrees divisible by q. In all examples we consider we still get an appropriate Q-nef-partition. Question 13. Is it always true? If not, what conditions should we put to have an appro- priate Q-nef partition? Remark 14. If this is true then this statement can be strengthened in the following way. There is a Q-nef-partition I1, . . . , Ik such that wi = 1 for i /∈ I1 ∩ . . . ∩ Ik. Indeed, consider the given Q-nef-partition. Delete w0. All numerical conditions still hold for a collection of weights and degrees we get. Thus there is another appropriate Q-nef-partition. Deleting the smallest weights step by step we find the partition we need. It is also natural to consider Hori -- Vafa models for quasismooth Fano complete inter- sections. But even in the case of a Cartier hypersurface it is not always possible to write down a Hori -- Vafa model. An example is a hypersurface of degree 30 in P(1, 6, 10, 15): it has no Q-nef partition. Another example, suggested to the author by S. Galkin, is a hypersurface of degree 30 in P (1, 6, 6, 6, 6, 10, 10, 15). It shows that even nef-partition (that is Q-nef-partition in Gorenstein variety) does not necessarily exist. The reason of this phenomenon seems to be the following: such complete intersections should be considered as smooth stacks instead of considering them as singular varieties. 4 Question 15. Is there a stacky version of Hori -- Vafa procedure? If yes, can it be refor- mulated in Laurent polynomials terms? Even if a hypersurface had a Hori -- Vafa model, it can have no very weak Landau -- Ginzburg model of type discussed in the paper. An example is a hypersurface of degree 30 in P(1, 1, 1, 1, 1, 6, 10, 15). Question 16. Does this hypersurface (or all complete intersections having Hori -- Vafa models) have another weak Landau -- Ginzburg models, not of Hori -- Vafa type? In other words, is it rational? The author is grateful to I. Cheltsov, S. Galkin, V. Golyshev, L. Katzarkov, D. Orlov, K. Shramov, D. Stepanov, and D. van Straten for helpful comments and T. Logvinenko for detailed proofreading and English checking. References [Dim86] A. Dimca, Singularities and coverings of weighted complete intersections, J. Reine Angew. Math. 366 (1986), 184 -- 193. [DM09] A. Douai, E. Mann, The small quantum cohomology of a weighted projective space, a mirror D- module and their classical limits, 2009, arXiv:0909.4063. [Dol82] I. Dolgachev, Weighted projective varieties, Lecture Notes in Mathematics 956 (1982), 34 -- 71. [Dou06] A. Douai, Construction of Frobenius manifolds via Laurent polynomials: a different approach. (Construction de vari´et´es de Frobenius via les polynomes de Laurent: une autre approche.). [Giv96] A. Givental, Equivariant Gromov-Witten invariants., Int. Math. Res. Not. 1996 (1996), no. 13, 613 -- 663. [HV00] K. Hori, C. Vafa, Mirror symmetry, 2000, arXiv:hep-th/0002222. [Prz07] V. Przyjalkowski, Quantum cohomology of smooth complete intersections in weighted projective spaces and in singular toric varieties., Sb. Math. 198 (2007), no. 9, 1325 -- 1340. [Prz08] V. Przyjalkowski, On Landau -- Ginzburg models for Fano varieties, Com. Num. Th. Phys. 1, No. 4 (2008), 713 -- 728, arXiv:0707.3758. [Prz09] V. Przyjalkowski, Weak Landau -- Ginzburg models for smooth Fano threefolds, 2009, arXiv:0902.4668. Steklov Mathematical Institute, 8 Gubkina street, Moscow 119991, Russia E-mail address: [email protected], [email protected] 5
1705.03171	2	1705	2017-07-16T15:01:25	Lazarsfeld-Mukai Reflexive Sheaves and their Stability	[ "math.AG" ]	Consider an ample and globally generated line bundle $L$ on a smooth projective variety $X$ of dimension $N\geq 2$ over $\mathbb{C}$. Let $D$ be a smooth divisor in the complete linear system of $L$. We construct reflexive sheaves on $X$ by an elementary transformation of a trivial bundle on $X$ along certain globally generated torsion-free sheaves on $D$. The dual reflexive sheaves are called the Lazarsfeld-Mukai reflexive sheaves. We prove the $\mu_L$-(semi)stability of such reflexive sheaves under certain conditions.	math.AG	math	LAZARSFELD-MUKAI REFLEXIVE SHEAVES AND THEIR STABILITY POORNAPUSHKALA NARAYANAN Abstract. Consider an ample and globally generated line bundle L on a smooth pro- jective variety X of dimension N ≥ 2 over C. Let D be a smooth divisor in the complete linear system of L. We construct reflexive sheaves on X by an elementary transformation of a trivial bundle on X along certain globally generated torsion-free sheaves on D. The dual reflexive sheaves are called the Lazarsfeld-Mukai reflexive sheaves. We prove the µL-(semi)stability of such reflexive sheaves under certain conditions. 1. Introduction Lazarsfeld-Mukai bundles were introduced by Lazarsfeld [15] and Mukai [21] in the 1980s. They are an important class of vector bundles obtained from certain elementary transformations and have found applications in studying syzygies and Brill-Noether the- ory. These bundles play a crucial role in Lazarsfeld's proof of the Gieseker-Petri theorem [15] and Voisin's proof of the generic Green's conjecture [25, 26]. Suppose X is a smooth projective surface over C, and C is a smooth, irreducible curve on X. Consider a globally generated line bundle A on C. Denote by i∗A, the direct image of A on X where i : C ֒→ X is the inclusion. Then i∗A is a globally generated coherent sheaf on X. We thus have the following exact sequence on X where the kernel F is a vector bundle: 0 −→ F −→ H 0(A) ⊗ OX ev−→ i∗A −→ 0 . The dual of F is called the Lazarsfeld-Mukai bundle on X associated to the pair (C, A). Lelli-Chiesa [17] has studied the (semi)stability of the Lazarsfeld-Mukai bundles on K3- surfaces, and similar results have been obtained by us on abelian surfaces [22]. Also, [2] and references therein give a general survey of Lazarsfeld-Mukai bundles with other applications. In this article we generalize the above construction to higher dimensional varieties. We in fact obtain reflexive sheaves as kernels. We study their µ-(semi)stability properties in various cases. This construction also enables us to obtain on any smooth projective variety X, semistable vector bundles E with rank E = dim X. Suppose X is a smooth projective variety of dimension N ≥ 2 over C. Let D i ֒−→ X be a smooth, irreducible divisor on X and A be an ample and globally generated line bundle on D. Consider a general subspace V ⊂ H 0(D, A) of dimension r ≥ 2. Let Z(V ) ֒→ D be the closed subscheme defined by the vanishing of sections of V and IZ(V ) ⊂ OD be its ideal sheaf. Then A ⊗ IZ(V ) is a globally generated torsion-free sheaf on D. We have Key words and phrases. (Semi)stability, Reflexive sheaves. Mathematics Classification numbers: 14C20, 14J60, 14M10. 1 2 P. NARAYANAN the following short exact sequence, which defines the sheaf FD,A,V associated to the triple (D, A, V ) on X: 0 −→ FD,A,V −→ V ⊗ OX −→ i∗(A ⊗ IZ(V )) −→ 0 . The kernel FD,A,V is a reflexive sheaf of rank r on X, whose dual is called the Lazarsfeld- Mukai reflexive sheaf (see § 3). We remark that the same construction can be carried out under the weaker assumption that D is just reduced and irreducible but not necessarily smooth. But for the purpose of this paper, we confine ourselves mainly with smooth and irreducible divisors D. The µ-(semi)stability properties of the sheaves FD,A,V are studied in § 4. The first case we consider is that of a variety X whose Picard group is cyclic. Theorem 1.1. Suppose X is an irreducible smooth projective variety over C of dimension N ≥ 2, such that Pic X = Z · [H], where [H] is the class of an ample divisor. Let D ∈ OX(H) and A be an ample, globally generated line bundle on D. Consider V ⊂ H 0(D, A), an r-dimensional subspace where r ≥ 2. Then the reflexive sheaf FD,A,V is µH-stable. Any D ∈ OX(H) is reduced, irreducible and Cohen-Macaulay. Hence we can consider In the specific case when X is the projective reflexive sheaves FD,A,V for all such D. space, we have the following theorem. Theorem 1.2. Suppose X = PN C for N ≥ 2. Consider L = O(d) with d > 0 on X. Let D ∈ L be a general smooth, irreducible hypersurface and A be an ample, globally gener- ated line bundle on D. Suppose that V is a general r-dimensional subspace of H 0(D, A), where r ≥ 2. Then the following table summarizes the conditions on L, A and r under which the sheaves FD,A,V are µO(1)-(semi)stable. L (a) L = O(1) (b) L = O(2) (c) L = O(2l) which A all A all A (d) L = O(d) A = O(md)D 2 ≤ r ≤ (cid:0)N −1+m A = O(l)D when d > 1 r r ≥ 2 r = 2 r = 2 stability µO(1)-stable µO(1)-semistable µO(1)-semistable m (cid:1) µO(1)-semistable See § 4.3 for a proof. Part (a) of the above theorem follows from Theorem 1.1. Part (b) is proved by applying a lemma from [23]. In case of part (c), we prove that FD,A,V D is µO(1)D -semistable, which implies our result. We prove part (d) of the theorem by proving it in general for any smooth, irreducible projective variety X, cf. Theorem 1.3. We remark that, in case (d) of the above theorem, if the condition A = O(md)D is weakened, the assertion is not necessarily true. Lemma 4.11 in § 4.3 gives a class of such examples. Note that parts (a) and (b) of the theorem hold for all reduced and irreducible D in the linear system and all r-dimensional subspaces V . We prove the following theorem on the µ-(semi)stability of the reflexive sheaves FD,A,V on arbitrary smooth projective varieties. Theorem 1.3. Suppose X is an irreducible, smooth projective variety of dimension N ≥ 2 over C. Consider an ample, globally generated line bundle L on X and an irreducible, LAZARSFELD-MUKAI REFLEXIVE SHEAVES AND THEIR STABILITY 3 smooth D ∈ L. For m > 0, let V ⊂ H 0(D, L⊗m 2 ≤ r ≤ (cid:0)N −1+m µL-semistable. m (cid:1). Then, for a general pair (D, V ), the reflexive sheaf FD,L⊗m D ) be an r-dimensional subspace, where D ,V is See § 4.2. The method of proof employed is the following. Consider an appropriate finite morphism X → PN , where N = dim X. We prove the corresponding (semi)stability statement for the projective space. We know by [19, Lemma 1.17], that the pullback of a semistable torsion-free sheaf under a finite morphism is semistable; and that semistability is an open condition in flat families [13, Proposition 2.3.1]. We thus get the required result. The same technique is applied to study the (semi)stability of some kernel bundles. A kernel bundle ML,W is defined as follows. Consider an ample and globally generated line bundle L on a smooth, irreducible projective variety X of dimension n. Let W ⊂ H 0(X, L) be a subspace such that the linear system PW is base-point free. Hence we have the following short exact sequence where ML,W is the kernel vector bundle: 0 −→ ML,W −→ W ⊗ OX −→ L −→ 0 . Let W ⊂ H 0(X, L) be a general (n + 1)-dimensional subspace. We prove that the kernel bundle ML,W associated to (L, W ) is µL-polystable, cf. Proposition 5.2. We mention that for curves, certain surfaces and projective spaces, the µL-(semi)stability of ML,W has been obtained in [4, 5, 6, 7, 8, 9, 10, 20, 24], for W = H 0(X, L) with certain assumptions on L. Acknowledgements. I thank Dr. Jaya NN Iyer for her guidance during the course of this project. I also thank Dr. T. E. Venkata Balaji and Prof. D. S. Nagaraj for helpful discussions and their support. I also thank the referee and the editor for useful comments which helped make the exposition better. 2. Preliminaries 2.1. Definitions and Notations. Let X be a smooth projective variety of dimension N ≥ 2 over C. (1) Let L be an ample and globally generated line bundle on X. By Bertini's theorem, the set smL as given below is a dense open set of L, smL = {D ∈ L : D is smooth and irreducible} . (2) Let W be a vector space over C. Then G(m, W ) (1 ≤ m ≤ dim W ) denotes the Grassmannian of m-dimensional subspaces of W . 2.2. Mumford-Takemoto (semi)stability. Let L be an ample line bundle on X (as above). Consider a torsion-free coherent sheaf F of rank r on X. (1) The slope of F with respect to L is defined as: µL(F ) = c1(F ) · (LN −1) r . (2) The sheaf F is said to be µL-semistable (resp. µL-stable), if for any coherent subsheaf E ⊂ F of rank s where 0 < s < r, one has µL(E) ≤ µL(F ) (resp. µL(E) < µL(F )). 4 P. NARAYANAN (3) The torsion-free coherent sheaf F is µL-polystable if it is a direct sum of µL-stable sheaves of the same slope. 3. Construction of Reflexive sheaves Consider a smooth projective variety X of dimension N ≥ 2 over C, and an ample, globally generated line bundle L on X. For a divisor D ∈ smL, let i : D ֒→ X denote the inclusion. Let A be an ample, globally generated line bundle on D. By the Noether-Lefschetz Theorem, if N ≥ 4, then Pic X −→ Pic D is an isomorphism, cf. [16, Example 3.1.25]. Hence, the line bundle A is the restriction of a line bundle from X. Consider G(r, H 0(D, A)), the Grassmannian of r-dimensional subspaces of the space of global sections H 0(D, A), where 2 ≤ r ≤ h0(D, A). For V ∈ G(r, H 0(D, A)), let Z(V ) denote the closed subscheme of D defined by the vanishing of sections of V . Recall that Z(V ) has codimension at most r in D. In fact, Z(V ) has codimension exactly r in D for a general V ∈ G(r, H 0(D, A)). The base locus of the linear system PV corresponding to (A, V ) on D is Z(V ). The ideal sheaf IZ(V ) of Z(V ) is the image of the morphism V ⊗ A∨ −→ OD on D. This gives the surjective evaluation map V ⊗ OD ։ A ⊗ IZ(V ) on D. Push-forward this morphism by the closed immersion i to X, and consider the following composition: V ⊗ OX ։ V ⊗ i∗OD −→ i∗(A ⊗ IZ(V )) −→ 0 . Let FD,A,V denote the kernel of the composition. Thus, we get: (1) 0 −→ FD,A,V −→ V ⊗ OX −→ i∗(A ⊗ IZ(V )) −→ 0 . Proposition 3.1. The kernel sheaf FD,A,V associated to (D, A, V ) has the following initial properties: (a) The sheaf FD,A,V is reflexive of rank r. (b) For a general V ∈ G(r, H 0(D, A)), the sheaf FD,A,V is locally free when r ≥ N. (c) The determinant of FD,A,V is OX (−D) ≃ L∨. (d) The sheaf FD,A,V has no non-zero global sections, i.e. H 0(X, FD,A,V ) = 0. Proof. For part (a), we note that in the exact sequence (1), FD,A,V is the elementary transformation of a locally free sheaf by a torsion-free sheaf supported on the smooth divisor D. By [1, Lemma 2.4], FD,A,V is a reflexive sheaf of rank r. When r ≥ N, a general V ∈ G(r, H 0(D, A)) has codimDZ(V ) = r ≥ N. Hence, Z(V ) is empty as D is of dimension N − 1. In this case, A ⊗ IZ(V ) ≃ A and FD,A,V is the usual elementary transformation, and is locally free. From (1), det FD,A,V ≃ det i∗(A⊗IZ(V ))∨. Since Z(V ) is of codimension at least 2 in X, we have det i∗(A ⊗ IZ(V )) ≃ det i∗A ≃ OX (D). Part (d) can be proved by consider the long exact sequence of cohomology associated to (1). (cid:3) This construction gives us reflexive sheaves of rank r ≥ 2 on smooth projective varieties D,A,V , the Lazarsfeld-Mukai of dimension N ≥ 2. We call the dual sheaves ED,A,V = F ∨ reflexive sheaves. Dualizing the exact sequence (1) defining FD,A,V , we get: 0 −→ V ∨ ⊗ OX −→ ED,A,V −→ E xt1(i∗(A ⊗ IZ(V )), OX) −→ 0 . (2) LAZARSFELD-MUKAI REFLEXIVE SHEAVES AND THEIR STABILITY 5 As with Lazarsfeld-Mukai bundles, the Lazarsfeld-Mukai reflexive sheaves are naturally equipped with an r-dimensional space of global sections V ∨ ⊂ H 0(X, ED,A,V ). Remark 3.2. Suppose X is a smooth, irreducible projective variety of dimension N ≥ 2 over C. Let D be a smooth and irreducible divisor on X and A be an ample and globally generated line bundle on D. Let r be such that 2 ≤ r ≤ h0(D, A). Consider the dense open subscheme Ur ⊂ G(r, H 0(D, A)) given by: Ur = {V ∈ G(r, H 0(D, A)) Z(V ) has codimension r in D} . If V ∈ Ur, it is well-known that the ideal sheaves IZ(V ) form a flat family of sheaves parametrized by Ur. Consequently, the torsion-free sheaves {A ⊗ IZ(V )}V ∈Ur and the dual Lazarsfeld-Mukai reflexive sheaves {FD,A,V }V ∈Ur form flat families of sheaves parametrized by Ur. 4. (Semi)stability of the sheaves FD,A,V In this section we study the (semi)stability properties of the reflexive sheaves FD,A,V . 4.1. Varieties with cyclic Picard group. Suppose that X is a smooth projective variety of dimension N ≥ 2 such that Pic X ≃ Z · [H]. Here [H] is the class of the ample generator the Picard group. Projective spaces, smooth hypersurfaces in Pn for n ≥ 4, general complete intersections in higher dimensional projective spaces are some examples of such varieties. We now prove Theorem 1.1. When X is a K3-surface with cyclic Picard group and FD,A,V is a vector bundle, the µH-stability and the simplicity of FD,A,V is known, cf. [15, Lemma 1.3]. We generalize this to higher dimensional varieties when FD,A,V is a reflexive sheaf. Proof of Theorem 1.1. Denote E := ED,A,V = F ∨ D,A,V . Recall the exact sequence (2) defin- ing the Lazarsfeld-Mukai reflexive sheaf. Since the cokernel sheaf E xt1(i∗(A ⊗ IZ(V )), OX) is supported only on D, there is a generically surjective morphism Or X −→ E. Assume the contrary, i.e. E (equivalently FD,A,V ) is not µH-stable. Then, there is a torsion-free quotient Q of E, i.e. E ։ Q of rank s < r, such that µH(E) ≥ µH(Q). From Proposition 3.1, det (E) ≃ OX (H), thus c1(Q) · (H N −1) s ≤ H N r . Since Pic X ≃ Z, we get c1(Q) ≤ 0. Let eQ = Q∨∨, the double dual of the torsion-free sheaf Q. The sheaf eQ on X is reflexive with c1(eQ) = c1(Q), and there is a natural inclusion Q ֒→ eQ which is generically X −→ E −→ Q ֒→ eQ. As each morphism is an isomorphism. We have morphisms Or X −→ eQ . The s-th generically surjective, we have a generically surjective morphism Or wedge product of the above morphism gives a generically surjective morphism: s^ Or X −→ s^ eQ ≃ det eQ . By [13, Proposition 1.2.7], we have 0 = µH(Vs Or implying that c1(eQ) ≥ 0. Hence, c1(eQ) = 0, i.e. det Q ≃ det eQ ≃ OX . X) ≤ µH(det eQ) = c1(det eQ) · (H N −1) , 6 P. NARAYANAN Let ξ be the generic point of X. The generically surjective morphism Or X −→ eQ, X )ξ ։ eQξ. Suppose the morphism gives the surjective morphism of Oξ-vector spaces (Or X −→ eQ is given by global sections t1, t2, · · · , tr of eQ, then the stalks of these at ξ Or generate eQξ as an Oξ-vector space. Thereby, there are s among these, say t1, t2, · · · , ts which form a basis of eQξ. These sections give a generically surjective morphism: f : Os X −→ eQ . The map f is in fact an isomorphism. Indeed, if K and C denote the kernel and cokernel of f , we get: 0 −→ K −→ Os X f −→ eQ −→ C −→ 0 . Since f is generically surjective, rank C = 0. Then, rank K = 0, and thus K = 0. Also, det C ≃ OX. Therefore, C is supported on a closed set of codimension ≥ 2 on X. Hence, X and eQ are isomorphic on an open set whose complement has codimension ≥ 2. By Os [12, Proposition 1.6 (iii)], eQ ≃ Os X . X ≃ eQ∨ ≃ Q∨∨∨ ≃ Q∨. From By [12, Corollary 1.2], Q∨ is reflexive. This gives Os the surjection E ։ Q, we get Q∨ ≃ Os X ֒→ FD,A,V . This contradicts H 0(X, FD,A,V ) = 0 (Proposition 3.1 (d)). Hence, the kernel reflexive sheaf is µH-stable. (cid:3) Remark 4.1. From the proof of theorem, the following commutative diagram gives Q ≃ Os X . Os X ≃ / eQ ≃ Os :✈ ✈ X ✈ ✈ ✈ ✈ ✈ ✈ ✈ ✈ Q The following result can be inferred from the above. Consider a smooth projective variety X. Let Q be a torsion-free sheaf on X with trivial determinant over an open set whose complement has codimension ≥ 2. Suppose Q admits a generically surjective morphism Or X −→ Q, then the torsion-free sheaf Q is itself trivial. We thank the referee for pointing this out. Remark 4.2. The proof of Theorem 1.1 proves essentially the following statement: Sup- pose a smooth projective variety X has Pic X = Z · [H], where H is ample. Let F be a reflexive sheaf on X such that a. the determinant det F = OX(−H), b. there is a generically surjective morphism from a trivial bundle on X to F ∨, c. the space of global sections of F , i.e. H 0(X, F ) = 0. Then the reflexive sheaf F is µH-stable. 4.2. (Semi)stability in case of arbitrary smooth projective varieties. Suppose that X is an irreducible, smooth projective variety of dimension N ≥ 2 over C. Let L be an ample and globally generated line bundle on X. Lemma 4.3. For a general D ∈ smL, there is a finite, flat morphism φ : X −→ PN such that D maps to the hyperplane H = Z(x0) in PN , where x0, x1, · · · , xN ∈ H 0(PN , O(1)) are the homogeneous coordinates. / , : LAZARSFELD-MUKAI REFLEXIVE SHEAVES AND THEIR STABILITY 7 Proof. Since L is globally generated and dim X = N, any general collection of N + 1 sections in H 0(X, L) generate L. If D ∈ smL is general, then D = Z(s0) for some s0 ∈ H 0(X, L), and we can find s1, s2, · · · , sN ∈ H 0(X, L) such that {s0, s1, s2, · · · , sN } generate L. These sections give a morphism φ : X −→ PN such that φ∗(O(1)) = L and si = φ∗xi. If H is the hyperplane Z(x0) ⊂ PN , we get the following commutative diagram. D φD H X φ / PN i j Since X is a smooth projective variety over C and φ is defined by the sections of an ample line bundle L, the morphism φ is finite [16, Corollary 1.2.15]. This also implies that φ is surjective. Thereby, φ is a flat morphism [11, Exercise III.9.3 (a)]. (cid:3) We remark that, over positive characteristic we will obtain a finite map from X to the projective space by choosing a very ample line bundle L. Any ample, globally generated line bundle on the hyperplane H is of the form OH(m) ≃ OPN (m)H for some m > 0. Consider V ′ ∈ G(r, H 0(H, OH(m)) where 2 ≤ r ≤ h0(H, OH(m)), such that codimHZ(V ′) = r. Then we have: (3) 0 −→ FH,OH (m),V ′ −→ V ′ ⊗ OPN −→ j∗(OH (m) ⊗ IZ(V ′)) −→ 0 . By Theorem 1.1, FH,OH (m),V ′ is µO(1)-stable. We now have the following lemma. Lemma 4.4. Let A = (LD)⊗m on D. Then (a) the subspace V = φ∗ (b) the sheaf FD,A,V ≃ φ∗FH,OH (m),V ′ and is µL-semistable. DV ′ ⊂ H 0(D, A) and codimDZ(V ) = r, Proof. Note that, A = (LD)⊗m ≃ φ∗ H 0(D, A) . Since V = φ∗ get the following commutative diagram. DOH(m) . This gives V = φ∗ DV ′ ⊂ φ∗ DV ′, the closed subscheme Z(V ) maps to Z(V ′) under φ, and we DH 0(H, OH(m)) ⊂ i′ Z(V ) D φZ(V ) φD j ′ Z(V ′) / H i j / X φ / PN Since φ is a finite, surjective and flat morphism, so are φD and φZ(V ). This implies that Z(V ) and Z(V ′) have the same dimension. Thus, codimD(Z(V )) = codimH(Z(V ′)) = r. This proves (a). Consider the pullback of the exact sequence (3) by φ: (4) 0 −→ φ∗FH,OH (m),V ′ −→ V ⊗ OX −→ φ∗j∗(OH (m) ⊗ IZ(V ′)) −→ 0 . By [11, Proposition III.9.3], φ∗j∗(OH (m) ⊗ IZ(V ′)) ≃ i∗(φD)∗(OH (m) ⊗ IZ(V ′)) ≃ i∗(A ⊗ φ∗ DIZ(V ′)) . D IZ(V ′) ≃ IZ(V ) ⊂ OD. Hence, the short exact sequence (4) becomes Note that φ∗ 0 −→ φ∗FH,OH (m),V ′ −→ V ⊗ OX −→ i∗(A ⊗ IZ(V )) −→ 0 . / / / / / / / / 8 P. NARAYANAN Therefore, φ∗FH,OH (m),V ′ ≃ FD,A,V . Further, since FH,OH (m),V ′ is µO(1)-stable, and φ is a finite morphism, by [19, Lemma 1.17], FD,A,V is µL-semistable. (cid:3) We prove Theorem 1.3. Proof of Theorem 1.3. Let D ∈ smL be a divisor which is general in the linear system. By Lemma 4.3, there is a finite morphism φ : X −→ PN such that φ(D) is the hyperplane H = Z(x0). Consider the line bundle A = (LD)⊗m on D for any m > 0. By Remark 3.2, there is a flat family of rank r reflexive sheaves (where 2 ≤ r ≤ h0(D, A)) parametrized by V ∈ Ur = {V ∈ G(r, H 0(D, A)) codimDZ(V ) = r}. Each V ∈ Ur corresponds to the reflexive sheaf FD,A,V ≃ FD,L⊗m D ,V . Let V = φ∗V ′, where V ′ ∈ G(s, H 0(H, OH(m)) is such that codimH Z(V ′) = s. Then by Lemma 4.4, V ∈ Us and FD,A,V is µL-semistable. Note that s can only vary in the range 2 ≤ s ≤ h0(H, OH(m)) = (cid:0)N −1+m m (cid:1), we m (cid:1). Hence, for any r in the range 2 ≤ r ≤ (cid:0)N −1+m have a V ∈ Ur with the corresponding FD,A,V being µL-semistable. By [13, Proposition 2.3.1], semistability is an open condition in flat families. Therefore, for a general V ∈ (cid:3) G(r, H 0(D, A)) where 2 ≤ r ≤ (cid:0)N −1+m 4.3. (Semi)stability in case of projective space. Suppose X = PN for N ≥ 2. m (cid:1), FD,A,V is µL-semistable. 4.3.1. We first consider the case L = O(2) and r = 2. Let D ∈ smO(2) be a degree two hypersurface. Let A be an ample, globally generated line bundle on D. We now discuss the µO(1)-stability of the rank 2 reflexive sheaf FD,A,V where V ∈ G(2, H 0(D, A)). We recall the concept of normalization of a torsion-free sheaf [23, Chapter II, § 1.2]. A torsion-free sheaf E of rank 2 over X = PN has a uniquely determined integer kE associated to it, namely, c1(E) c1(E) + 1 for c1(E) odd . 2 (5) kE = − if c1(E) even, and kE = − 2 Note that, c1(E(kE)) ∈ {0, −1}. We set Enorm := E(kE), and call E normalized if E = Enorm. Lemma 4.5. [23, Lemma II.1.2.5] A reflexive sheaf E of rank 2 over X = PN is stable if and only if Enorm has no sections : H 0(X, Enorm) = 0. If c1(E) is even, then E is semistable if and only if H 0(X, Enorm(−1)) = 0. Corollary 4.6. Consider L = O(2) on X = PN for N ≥ 2. Let D ∈ smL, A be an ample, globally generated line bundle on D and V ∈ G(2, H 0(D, A)) be a 2-dimensional subspace. Then the associated rank two FD,A,V is µO(1)-semistable. Proof. If L = O(2), then c1(FD,A,V ) = −2. Then, kFD,A,V = 1. Thus, FD,A,V norm = FD,A,V (1). Now, the result follows from the fact that H 0(X, FD,A,V ) = 0 (Proposition 3.1 (d)). (cid:3) 4.3.2. We consider the case L = O(2l) and r = 2. Start with L = O(d) for any d > 0. Consider D ∈ smL where i : D ֒→ X denotes the inclusion. Suppose A is an ample, globally generated line bundle on D. For r ≥ 2, let V ∈ G(r, H 0(D, A)) be such that codimDZ(V ) = r. LAZARSFELD-MUKAI REFLEXIVE SHEAVES AND THEIR STABILITY 9 We restrict the exact sequence (1) to the open set eU = X \ Z(V ) to get: 0 −→ FD,A,V eU −→ V ⊗ O eU −→ i∗(A ⊗ IZ(V )) eU −→ 0. (6) Let U = D \ Z(V ), an open subset of D. By [11, Proposition II.6.5], U is a divisor in eU and we have the following commutative diagram. U iU eU D i / X By [11, Proposition III.9.3], i∗(A ⊗ IZ(V )) eU ≃ (iU )∗(AU ⊗ IZ(V )U ) ≃ (iU )∗(AU ). The short exact sequence (6) then becomes: 0 −→ FD,A,V eU −→ V ⊗ O eU −→ (iU )∗(AU ) −→ 0 . Thus by [1, Lemma 1.1], FD,A,V eU is an elementary transformation, and hence a locally free sheaf of rank r on eU . Then, (iU )∗FD,A,V eU = FD,A,V U on U ⊂ D is also locally free of rank r. Now, U is an open subset of D whose complement Z(V ) is of codimension r ≥ 2 in D. This implies that the stalks of FD,A,V D are torsion-free. Remark 4.7. The sheaf FD,A,V D is torsion-free of rank r. Assume now that r = 2. Thus, FD,A,V and FD,A,V D have rank 2. Restrict the exact sequence (1) to D to get: 0 −→ K −→ FD,A,V D −→ V ⊗ OD −→ A ⊗ IZ(V ) −→ 0 on D. Here K denotes the kernel, which is torsion-free by Remark 4.7. Note that K is a sheaf of rank 1 on D. Consider the kernel M of the surjective evaluation map V ⊗OD ։ A⊗IZ(V ). The sheaf M is reflexive, cf. [12, Proposition 1.1]. Since M is of rank 1, by [12, Proposition 1.9], M is a line bundle. By comparing determinants, M ≃ A∨, and we get: 0 −→ A∨ −→ V ⊗ OD −→ A ⊗ IZ(V ) −→ 0 . From the exact sequences above, we get the following exact sequence on D: (7) 0 −→ K −→ FD,A,V D −→ A∨ −→ 0. Remark 4.8. As FD,A,V D is torsion-free, det FD,A,V D ≃ (detFD,A,V )D ≃ (LD)∨. Thus, from the exact sequence (7), the determinant of the rank 1 torsion-free sheaf K is: det K ≃ A ⊗ det FD,A,V D ≃ A ⊗ (LD)∨ . Proposition 4.9. Consider L = O(2l) for l > 0. Let D ∈ smL, A = O(l)D and V ∈ G(2, H 0(D, A)) such that codimDZ(V ) = 2. Then the torsion-free sheaf FD,A,V D on D is µO(1)D -semistable. Thus FD,A,V is µO(1)-semistable. Proof. Both K and A∨ are rank one torsion-free sheaves on D and hence are µO(1)D -stable. Further they have the same determinant. Indeed, A∨ ≃ O(−l)D and det K ≃ A ⊗ (LD)∨ ≃ O(l)D ⊗ O(−2l)D ≃ O(−l)D . / / _ _ / 10 P. NARAYANAN Hence FD,A,V D is a torsion-free sheaf which is an extension of µO(1)D -stable sheaves with the same slope. By [19, Lemma 1.10 (3)], FD,A,V D is µO(1)D -semistable. This implies that FD,A,V is µO(1)-semistable, cf. [18, Chapter 11]. (cid:3) We collect our observations to prove Theorem 1.2. Proof of Theorem 1.2. Part (a) of the theorem follows from Theorem 1.1. Part (b) of the theorem follows from Corollary 4.6. Proposition 4.9 proves part (c) of the theorem. Finally, part (d) follows from Theorem 1.3. (cid:3) Remark 4.10. If L = O(d), then for all r such that (r, d) = 1, a µO(1)-semistable FD,A,V is in fact stable. Indeed, if D ∈ O(d), then deg FD,A,V = c1(FD,A,V ) · (O(1)N −1) = −d and rank FD,A,V = r. Since the rank and degree are coprime, by [13, Lemma 1.2.14], semistability and stability coincide. Theorem 1.2 shows that there are families of µO(1)-(semi)stable rank r reflexive sheaves on the projective space with any prescribed c1. The following lemma shows that if we weaken our condition on A in part (d) of Theorem 1.2, then we may not get the required (semi)stability. Lemma 4.11. Consider L = OX (d) (d > 0) on X = PN for N ≥ 2. Let D ∈ L be a general smooth and irreducible hypersurface. Let A = OX(l)D and V ⊂ H 0(D, A) be a general r-dimensional subspace (r ≥ 2) such that r and l satisfy the following condition: Then FD,A,V is not µO(1)-semistable. 0 < l < d(r − 1) r . Note that if l is an integer such that 0 < l < d 2, then for any r ≥ 2, the above inequality is satisfied. Proof. Let l and r be integers that satisfy the given condition, i.e. 0 < l < d(r − 1) r . Consider the line bundle OX (l) on the projective space X. Since D ∈ OX(d) and A = OX (l)D, we have H 0(X, OX (l)) ≃ H 0(D, A). Choose a general V ∈ G(r, H 0(X, OX (l))) ≃ G(r, H 0(D, A)). Let ZX(V ) (resp. ZD(V )) denote the closed subscheme of X (resp. D), defined by the vanishing of sections of V ⊂ H 0(X, OX(l)) (resp. V ⊂ H 0(D, A)). In fact ZD(V ) = ZX(V ) ∩ D. Note that for a general D and V , codimX ZX(V ) = codimDZD(V ) = r. We get the following exact sequence on X: 0 −→ M −→ V ⊗ OX −→ OX (l) ⊗ IZX (V ) −→ 0 . Here M is a torsion-free (in fact reflexive) sheaf of rank r − 1 and determinant OX (−l). Thus µOX (1)(M) = −l r−1. Also, corresponding to the triple (D, A, V ), we get the following exact sequence on X (where i : D ֒→ X denotes the inclusion): 0 −→ FD,A,V −→ V ⊗ OX −→ i∗(A ⊗ IZD(V )) −→ 0 . LAZARSFELD-MUKAI REFLEXIVE SHEAVES AND THEIR STABILITY 11 Note that FD,A,V is of rank r and det FD,A,V = OX(−d). We get µOX (1)(FD,A,V ) = −d r . Comparing the above exact sequences, there is an inclusion M ֒→ FD,A,V . But µOX (1)(M) = −l r − 1 > −d(r − 1) r(r − 1) = −d r = µOX (1)(FD,A,V ) . Hence FD,A,V is not µO(1)-semistable. (cid:3) 5. An application to Kernel bundles In this section we see an application of the techniques used so far. As before, we work over the field of complex numbers. Consider the line bundle O(d) (d > 0) on Pn, n ≥ 2. We have the following exact sequence corresponding to the line bundle O(d) where MO(d) is the kernel vector bundle. (8) 0 −→ MO(d) −→ H 0(O(d)) ⊗ OPn −→ O(d) −→ 0 . Flenner [10, Corollary 2.2] proved that the kernel bundles MO(d) are µO(1)-semistable. Remark 5.1. In fact, in case of the line bundle O(1) on Pn, the kernel bundle MO(1) is stable. Indeed, note that MO(1) is a locally free sheaf on Pn with det MO(1) = O(−1). Fur- ther, there is a surjective morphism from a trivial bundle on Pn to M ∨ O(1). For, dualizing (8) when d = 1, we get the exact sequence: O(1) −→ 0 . Finally, since H 0(X, MO(1)) = 0, by Remark 4.2, MO(1) is µO(1)-stable. 0 −→ O(−1) −→ H 0(O(1))∨ ⊗ OPn −→ M ∨ Proposition 5.2. Let X be an irreducible, smooth projective variety of dimension n over C. Consider an ample and globally generated line bundle L on X. For a general subspace W ⊂ H 0(X, L) of dimension n + 1, the kernel bundle ML,W associated to (L, W ) is µL-polystable. Proof. For a general W ∈ G(n + 1, H 0(X, L)), the corresponding linear system PW is basepoint-free and gives a finite surjective morphism ψW : X −→ Pn . Note that ψ∗ W H 0(Pn, O(1)) ≃ W . We have the kernel bundle MO(1) on Pn defined by the exact sequence of the form (8) for d = 1. Pullback this sequence to X, we get: W O(1) ≃ L and that ψ∗ 0 −→ ψ∗ W MO(1) −→ W ⊗ OX −→ L −→ 0 . W MO(1) is the kernel bundle on X associated to (L, W ), i.e. ψ∗ Hence, ψ∗ W MO(1) ≃ ML,W . Again, since MO(1) is µO(1)-stable, by [19, Lemma 1.17], ML,W is µL-semistable. In fact, by a result of Kempf [14, Theorem 1], ML,W is µL-polystable on X. (cid:3) References [1] Abe, T. (2007), The elementary transformation of vector bundles on regular schemes, Trans. Amer. Math. Soc. 359 (9), 4285-4295. [2] Aprodu, M. (2013), Lazarsfeld-Mukai bundles and applications, Commutative Algebra, 1-23, Springer, New York. [3] Aprodu, M., Nagel, J., (2010), Koszul cohomology and algebraic geometry, University Lecture Series, Vol. 52, Amer. Math. Soc., Providence. [4] Beauville, A., (2003), Some stable vector bundles with reducible theta divisor, Manuscripta Math. 110, 343-349. 12 P. NARAYANAN [5] Butler, D., (1994), Normal generation of vector bundles over a curve, J. Diff. Geom. 39, 1-34. [6] Camere, C., (2008), About the stability of the tangent bundle of Pn restricted to a curve, C. R. Acad. Sci. Paris, Ser. I 346, 421-426. [7] Camere, C., (2012), About the stability of the tangent bundle of Pn restricted to a surface, Math. Z. 271, no. 1-2, 499-507. [8] Ein, L., Lazarsfeld, R., (1992), Stability and restrictions of Picard bundles, with an application to the normal bundles of elliptic curves, Complex Projective Geometry, Lond. Math. Soc. Lect. Note Ser. 179, 149-156. [9] Ein, L., Lazarsfeld, R., Mustopa, Y., (2013), Stability of syzygy bundles on an algebraic surface, Math. Res. Lett. 20 (1), 73-80. [10] Flenner, H., (1984), Restrictions of semistable bundles on projective varieties, Comment. Math. Helv. 59, 635-650. [11] Hartshorne, R., (1977), Algebraic geometry, Graduate Texts in Mathematics, No. 52, Springer- Verlag, New York. [12] Hartshorne, R., (1980), Stable Reflexive Sheaves, Math. Ann. 254 (2), 121-176. [13] Huybrechts, D., Lehn, M., (2010), The geometry of moduli spaces of sheaves. 2nd ed., Cambridge University Press, Cambridge. [14] Kempf, G., (1992), Pulling back bundles, Pacific J. Math. 152 (2), 319-322. [15] Lazarsfeld, R., (1986), Brill-Noether-Petri without degenerations, J. Diff. Geom. 23, 299-307. [16] Lazarsfeld, R., (2004), Positivity in algebraic geometry I & II, Ergebnisse der Mathematik und ihrer Grenzgebiete, vol. 48 & 49, Springer-Verlag, Berlin. [17] Lelli-Chiesa, M., (2013), Stability of rank-3 Lazarsfeld-Mukai bundles on K3 surfaces, Proc. Lond. Math. Soc. 107 (3), 451-479. [18] Le Potier, J., (1997), Lectures on vector bundles, Cambridge Studies in Advanced Mathematics 54, Cambridge University Press. [19] Maruyama, M., (1981), The theorem of Grauert-Mulich-Spindler, Math. Ann. 255 (3), 317-333. [20] Mistretta, E., (2008), Stability of line bundle transforms on curves with respect to low codimensional subspaces, J. Lond. Math. Soc. 78 (2), 172-182. [21] Mukai, S., (1989), Biregular classification of Fano threefolds and Fano manifolds of coindex 3, Proc. Nat. Acad. Sci. USA 86, 3000-3002. [22] Narayanan, P., (2017), On the semistability of certain Lazarsfeld–Mukai bundles on abelian surfaces, Annali dell'Universita di Ferrara. doi:10.1007/s11565-017-0277-z. [23] Okonek, C., Schneider, M., Spindler, H., (1980), Vector Bundles on Complex Projective Spaces, Progress in Mathematics, vol. 3, Birkhauser, Boston, MA. [24] Paranjape, K., and Ramanan, S., (1988), On the canonical ring of a curve, Algebraic geometry and commutative algebra, Volume II, Kinokuniya, Tokyo, 503-516. [25] Voisin, C., (2002) Greens generic syzygy conjecture for curves of even genus lying on a K3 surface, J. European Math. Soc. 4, 363-404 . [26] Voisin, C., (2005) Greens canonical syzygy conjecture for generic curves of odd genus. Compositio Math. 141, 1163-1190. Department of Mathematics, Indian Institute of Technology Madras, Chennai - 600036. E-mail address: [email protected]
1911.00780	1	1911	2019-11-02T20:07:32	From non Defectivity to Identifiability	[ "math.AG" ]	A projective variety $X\subset\mathbb{P}^N$ is $h$-identifiable if the generic element in its $h$-secant variety uniquely determines $h$ points on $X$. In this paper we propose an entirely new approach to study identifiability, connecting it to the notion of secant defect. In this way we are able to improve all known bounds on identifiability. In particular we give optimal bounds for some Segre and Segre-Veronese varieties and provide the first identifiability statements for Grassmann varieties.	math.AG	math	FROM NON DEFECTIVITY TO IDENTIFIABILITY ALEX CASAROTTI AND MASSIMILIANO MELLA Abstract. A projective variety X ⊂ PN is h-identifiable if the generic element in its h-secant variety uniquely determines h points on X. In this paper we propose an entirely new approach to study identifiability, connecting it to the notion of secant defect. In this way we are able to improve all known bounds on identifiability. In particular we give optimal bounds for some Segre and Segre-Veronese varieties and provide the first identifiability statements for Grassmann varieties. Introduction The notion of identifiability is ubiquitous both in applied and classical algebraic In general we say that an element p of a projective space PN is h- geometry. identifiable via a variety X if there is a unique way to write p as linear combination of h elements of X. In the classical setting this very often translates into rationality problems and it is linked to the existence of birational parameterizations. In the applied set up one usually considers a tensor space and the identifiability allows to reconstruct a tensor via a subset of special tensors defined by rank conditions or other special requirements. For applications ranging from biology to Blind Signal Separation, data compression algorithms, and analysis of mixture models, [DDL1] [DDL2] [DDL3] [KADL] [Si], uniqueness of decompositions allows to solve the prob- lem once a solution is determined. For all these reasons it is interesting and often crucial to understand identifiability. Over a decade ago the notion of h-weakly defective varieties has been connected to identifiability of polynomials, [Me]. This provided the first systematic study of identifiability for Veronese varieties. More recently with the work of Luca Chiantini and Giorgio Ottaviani, [CO], weakly defective varieties have been substituted by h-tangentially weakly defective varieties to study identifiability problems. In both approaches to provide identifiability one has to check the behavior of special linear systems and quite often this is done by an ad hoc degeneration argument. As a consequence identifiability has been proved in very few cases and quite often the result obtained are not expected to be sharp, [CO] [BDdG] [BC] [BCO] [Kr]. In this paper we want to develop an entirely new approach to study generic identifiability, see Definition 6 for the precise set up. Starting from the seminal paper [CC10], where the geometry of contact loci has been carefully studied, and the improvement presented in [BBC], we derive identifiability statements for non secant defective varieties. Even if new this is not really surprising since weakly de- fectiveness and tangentially weakly defectiveness, thanks to Terracini Lemma, have secant defectiveness as a common ancestor. With this new approach we are able to translate all the literature on defective varieties into identifiability statements, providing in many cases sharp classification of h-identifiability. One of the results Date: November 5, 2019. 2010 Mathematics Subject Classification. Primary 15A69, 15A72, 11P05; Secondary 14N05, 15A69. Key words and phrases. Tensor decomposition, Waring decomposition, identifiability, defective varieties. 1 2 ALEX CASAROTTI AND MASSIMILIANO MELLA we prove in this direction is the following conditional relation between identifiability and defectivity, we refer to section 1 for the necessary definitions. Theorem. Let X ⊂ PN be an irreducible reduced variety. Assume that h > dim X, X is not (h − 1)-tangentially weakly defective and it is not h-identifiable. Then X is (h + 1)-defective. As already mentioned identifiability issues are particularly interesting for tensor spaces. As a corollary we get the best asymptotic identifiability result so far for Segre, Segre-Veronese, and Grassmann varieties, that is, tensors and structured tensors, see section 3. As a sample we state the application to binary tensors. Theorem. The Segre embedding of n copies of P1, with n ≥ 5 is h-identifiable for any h ≤ ⌊ 2n n+1 ⌋ − 1. Recall that the generic rank of the Segre embedding of (P1)n is ⌈ 2n n+1 ⌉, therefore our result shows generic identifiability of all sub-generic binary tensors, qbits if you like the quantum computing dictionary, in the perfect case, that is when 2n n+1 is an integer, and all but the last one in general, as predicted by the conjecture posed in [BC]. The starting point of our analysis was the observation that, in all known ex- amples, when a variety X is not h-identifiable then any element in Sech+1(X) has infinitely many decompositions, [CMO] [BBC] [BCO] [BV] [COV1]. Going back to the ideas in [Me] we realized that the best way to use this observation is to set a connection between the abstract secant map and the tangential projection. Via this we reduce the problem to study fiber type tangential projections. The latter is accomplished via the construction of a map from the Hilbert scheme of points of the contact loci of h-tangentially weakly defective varieties to a suitable Grassmannian. Under the right assumptions this is proved to be of fiber type and it allows us to connect defectivity and non identifiability. This, together with an improvement of the contact loci geometry studied in [CC10] and [BBC], lead us to derive identifiability from non defectivity under very mild hypothesis. Theorem. Let X ⊂ PN be a smooth variety. Assume that πX generically finite and k > 2 dim X. Then X is (k − 1)-identifiable. k : Seck(X) → PN is The paper is structured as follows. In section 2 we study the geometry of contact locus and prove the main general results about the connections between defectivity and non identifiability. In the final section we apply our techniques to varieties that are meaningful for tensor decomposition and streamline our argument for those. We are much indebted with Luca Chiantini for many conversations on the subject and for explaining us the connection between tangentially weakly defective varieties and identifiability when we started to work on the subject. 1. Notation We work over the complex field. A projective variety X ⊂ PN is non degenerate if it is not contained in any hyperplane. Let X ⊂ PN be an irreducible and reduced non degenerate variety. Let X (h) be the h-th symmetric product of X, that is the variety parameterizing unordered sets h ⊂ X (h) be the smooth locus, given by sets of h distinct of h points of X. Let U X points. Definition 1. A point z ∈ U X h represents a set of h distinct point, say {z1, . . . , zh}. We say that a point p ∈ PN is in the span of z, p ∈ hzi, if it is a linear combination of the zi. FROM NON DEFECTIVITY TO IDENTIFIABILITY 3 With this in mind we define Definition 2. The abstract h-Secant variety is the irreducible and reduced variety sech(X) := {(z, p) ∈ U X h × PN p ∈ z} ⊂ X (h) × PN . Let π : X (h) × PN → PN be the projection onto the second factor. The h-Secant variety is Sech(X) := π(sech(X)) ⊂ PN , and πX h := πsech(X) : sech(X) → PN is the h-secant map of X. The irreducible variety sech(X) has dimension (hn + h − 1). One says that X is h-defective if dim Sech(X) < min{dim sech(X), N }. Remark 3. If X is h-defective then the h-secant map is of fiber type. Note that a general point in sech(X) is a linear combination of h points of X. Thanks to the non degeneracy assumption a general point in sech(X) ( PN is not a linear combination of less points on X. The tricky part in studying secant varieties is the closure. Many different things can happen, the h points can group in infinitely near cluster or positive dimensional intersection can appear. As a matter of facts these special loci are really difficult to control and the main advantage to use birational geometry is the opportunity to get rid of them. Definition 4. Let X ⊂ PN be a non degenerate subvariety. We say that a point h and p 6∈ hz′i for p ∈ PN has rank h with respect to X if p ∈ hzi, for some z ∈ U X any z′ ∈ U X h′ , with h′ < h. Remark 5. With this in mind it is easy to produce examples of limits of rank h points with different rank. If we let one of the point degenerate to the span of the others we lower the rank. If we let two points collapse the rank, generically, increases. Definition 6. A point p ∈ PN is h-identifiable with respect to X ⊂ PN if p is of h )−1(p) is a single point. The variety X is said to be h-identifiable rank h and (πX if πX h is a birational map, that is the general point of Sech(X) is h-identifiable It is clear, by the above remark, that when X is h-defective, or more generally when πX h is of fiber type X is not h-identifiable. The next ingredient we need to introduce is Terracini Lemma. Theorem 7 (Terracini Lemma). [CC02] Let X ⊂ PN be an irreducible variety. Then we have • for any x1, . . . , xk ∈ X and z ∈ hx1, . . . , xki hTx1 X, . . . , Txk Xi ⊆ Tz Seck(X), • there is a dense open set U ⊂ X (k) such that hTx1 X, . . . , Txk Xi = Tz Seck(X), for a general point z ∈ hx1, . . . , xki with (x1, . . . , xk) ∈ U . Terracini Lemma yields a direct consequence of h-defectiveness. If X is h- defective then the general fiber of πX h has positive dimension. Therefore by Ter- racini the general hyperplane tangent at h points of X is singular along a positive dimensional subvariety. This property does not characterize defective varieties. Definition 8. Let X ⊂ PN be a non degenerate variety. The variety X is said h- weakly defective if the general hyperplane singular along h general points is singular along a positive dimensional subvariety. 4 ALEX CASAROTTI AND MASSIMILIANO MELLA There is a direct connection, proven in [CC02], between h-weakly defectiveness and identifiability. Theorem 9. If X is not h-weakly defective then it is h-identifiable. The main problem is that it is quite hard in general to verify if a variety is h-weakly defective. To overcome this problem the notion of tangentially weakly defective varieties has been introduced, [CO]. Here we follow the notations of [BBC]. For a subset A = {x1, . . . , xh} ⊂ X of general points we set MA := h[i Txi Xi. By Terracini Lemma the space MA is the tangent space to Sech(X) at a general point in hAi. Definition 10. The tangential h-contact locus Γh :=: Γ(A) is the closure in X of the union of all the irreducible components which contain at least one point of A, of the locus of points of X where MA is tangent to X. We will write γh := dim Γ(A). We say that X is h-twd (tangentially weakly defective) if γh > 0. Remark 11. It is clear that if X is h-twd then it is h weakly defective, it is (h+ 1)- twd and Γh ⊆ Γh+1. Using scrolls it is not too difficult to produce explicit examples of varieties that are h-weakly defective but are not h-twd, see also Remark 19. For what follows it is useful to introduce also the notion of tangential projection. Definition 12. Let X ⊂ PN be a variety and A = {x1, . . . , xh} ⊂ X a set of general points. The h-tangential projection (from A) of X is τh : X 99K PM the linear projection from MA. That is, by Terracini Lemma, the projection from the tangent space of a general point z ∈ hAi of Sech(X) restricted to X. 2. Relation between twd and defectivity We start collecting properties of the tangential contact loci that will be useful for our purpose. Theorem 13. Let X ⊂ PN be an irreducible, reduced, and non degenerate variety. Let A ⊂ X be a set of h general points and Γ the associated contact locus. Assume that Sech−1(X) ( PN . Then we have: a) Γ is equidimensional and it is either irreducible (type I) or reduced (type II) with exactly h irreducible component, each of them containing a single point of A [CC10, Proposition 3.9], b) hΓi = Sech(Γ) and Seci(Γ) 6= hΓi for i < h [CC10, Proposition 3.9], h )−1(z)) ⊂ hΓi, [CC10, Proposition 3.9], c) for z ∈ hΓi general πX d) if we are in type I γh > γh−1, [BBC, Lemma 3.5] e) if γh = γh+1 and Sech+1(X) is not defective and does not fill up PN we are in type II, the irreducible components of both contact loci are linearly independent linear spaces, [BBC, Lemma 3.5], h ((πX f) if we are in type I and Sech+1(X) is not defective and does not fill up PN then Γh+1 is of type I. Proof. Points a) -- e) are proved in the cited papers under the assumption that Sech(X) ( PN . Points a) -- d) are immediate when Sech(X) = PN and Sech−1(X) ( PN . FROM NON DEFECTIVITY TO IDENTIFIABILITY 5 We have only to prove point f). Let A = {x1, . . . , xh} and B = A ∪ {xh+1} be general sets in X. Assume that Γ(B) is of type II. By definition Γ(A) ⊂ Γ(B), on the other hand by point a) the irreducible component of Γ(B) through x1 does not contain x2 and therefore it cannot contain Γ(A). This contradiction proves the claim. (cid:3) From the point of view of identifiability the notions of weakly defectiveness and twd behave the same. The following proposition is well known to the experts in the field but we were not able to found a written version of it. Proposition 14. [Ch] Let X ⊂ PN be an irreducible, reduced, and non degenerate variety. Assume that X is not h-twd, then X is h-identifiable. Proof. Assume that X is not h-identifiable and let z ∈ Sech(X) be a general point. Let z ∈ hx1, . . . xhi, for xi general in X. The existence of a different decomposition yields a new set {y1, . . . yh} ⊂ X such that z ∈ hy1, . . . , yhi. Moving the point z in the linear space hx1, . . . , xhi yields a positive dimensional contact locus. (cid:3) Remark 15. We want to stress that h-identifiability is not equivalent to non h-twd. In [COV2] and [BV] are described examples of Segre and Grassmannian varieties that are h-identifiable but h-twd. We aim to study the relation between twd and defectivity. The next lemma is a first step in this direction. Lemma 16. Let X ⊂ PN be an irreducible, reduced, and non degenerate variety of dimension n, πX k : seck(X) → PN the k-secant map, τ X Γ(x1, . . . , xk) the k-contact locus associated to the genral points x1, . . . , xk. k−1 : X 99K PM the (k − 1)-tangential projection, and Γ := k is of fiber type if and only if τ X i) The map πX ii) Let {x1, . . . , xk, y1, y2} be general points. Then k−1 is of fiber type. dim(Γ(x1, . . . , xk, y1) ∩ Γ(x1, . . . , xk, y2)) > 0 in a neighborhood of xi only if either X is k-twd or πX mensional fibers. k+2 has positive di- iii) The map (τ X k−1)Γ : Γ 99K Pγk is either of fiber type or dominant. Proof. i) By Terracini Lemma πX k is of fiber type if and only if Tz Seck−1(X) ∩ TyX 6= ∅ for y ∈ X general. This condition is clearly equivalent to have τ X k−1 of fiber type. ii) Assume that X is not k-twd and dim(Γ(x1, . . . , xk, y1)∩Γ(x1, . . . , xk, y2)) > 0 in a neighborhood of xi. Set MAi = hTx1 X, . . . , Txk X, Tyi Xi, the variety X is not k-twd therefore In particular we have and hence MA1 ∩ MA2 ) hTx1 X, . . . , Txk Xi. (MA1 ∩ MA2) ∩ Tyi X 6= ∅, hTx1 X, . . . , Txk X, Ty1 Xi ∩ Ty2 X 6= ∅. This shows, by the generality of the points and point i) that πX k+2 is of fiber type. k−1)Γ is not of fiber type. Then by point b) of Theorem 13 we k−1 and both maps are dominant (cid:3) k−1)Γ = τ Γ iii) Assume that (τ X have dimhΓi = k(γk + 1) − 1. Hence (τ X onto Pγk . 6 ALEX CASAROTTI AND MASSIMILIANO MELLA Next we prove a general statement for type II contact loci. Lemma 17. Let X ⊂ PN be an irreducible, reduced, and non degenerate variety. Assume that: a) X is k-twd, b) X is not (k − 1)-twd c) the k-contact locus is of type II. Then πX k+1 is of fiber type. Proof. By point i) in Lemma 16 it is enough to prove that τ X k is of fiber type. Then by projection it is enough to prove the latter for k = 2. Let {x1, x2, y} ⊂ X be a set of general points and Γ = Γ(x1, x2, y) the contact locus associated to {x1, x2, y}. To conclude the proof it is enough to prove that hTx1 , Tx2 i ∩ TxX 6= ∅, for x ∈ Γ a general point. For a general point p ∈ Γ we set Γi p ⊂ Γ(xi, p) the irreducible component of the contact locus Γ(xi, p) through p. The contact locus is of type II, therefore Γi p we have TxX ⊂ hTx1 X, TpXi. Then by semicontinuity for any point w ∈ Γ1 p there is a linear space of dimension n, say Aw ⊆ TwX, contained in the span. p 6∋ x1, x2. Note that for a general point x ∈ Γ1 Set T(Γ1 p) = hAwiw∈Γ1 . p We may assume that X is not 2-defective, otherwise there is nothing to prove, that is and, since y is general, Tx1 X ∩ Tx2 X = ∅, codimT(Γ1 y)(T(Γ1 y) ∩ Tx1 X) = n + 1. (1) (2) The variety X is not 1-twd, then there are points z ∈ Γ1 z ∈ Γ1 y be a point with y with Az ∩ Tx1 X 6= ∅. Let (3) The contact locus is of type II, therefore z 6= x1. We want to stress that this is the only point in the proof where we use the assumption that Γ is of type II. Az ∩ Tx1 X 6= ∅, If Az ∩ Tx2 X 6= ∅, by Equations (1) and (2) we have codimT(Γ1 y)(T(Γ1 y) ∩ hTx1 , Tx2i) ≤ n and we conclude TyX ∩ hTx1 , Tx2i 6= ∅, that is τ X 2 is of fiber type. Assume that Az ∩ Tx2 X = ∅. Then we consider the span hAz, Tx2 i. By semicon- z its irreducible tinuity to this linear space is associated a contact locus and we set Γ2 component passing through z. As before we have and by Equations (1) and (3) we conclude that codimT(Γ2 z)(T(Γ2 z) ∩ Tx2 ) = n + 1, codimT(Γ2 z)(T(Γ2 z) ∩ hTx1 , Tx2i) ≤ n. This yields for any point w ∈ Γ2 and the assumption that X is not 2-defective ensure that Aw ∩ hTx1 , Tx2 i 6= ∅, (4) z. We have z 6= y1, then the general choice of the points xi, for general w ∈ Γ2 z. Aw ∩ Tx1 X = ∅ (5) FROM NON DEFECTIVITY TO IDENTIFIABILITY 7 We set Γ1 w the irreducible component through w of the contact locus associated to hAw, Tx1 Xi. Again z 6= x1 and the general choice of the xi ensure that z 6∈ Γ1 w. In particular Set w 6= Γ2 Γ1 z. S2 := [v∈Γ1 ygeneral Γ2 v. Then Γ2 z is in the closure of S2 and, for p ∈ Γ1 y general, Γ1 y is in the closure of Hence Γ1 y is in the closure of [w∈Γ2 pgeneral Γ1 w. S1 := [w∈Γ2 zgeneral Γ1 w. In particular the general point of S1 is a general point of X. By construction we have codimT(Γ1 Equations (4) and (5) then give w)(T(Γ1 w) ∩ Tx1 X) ≤ n + 1. codimT(Γ1 w)(T(Γ1 w) ∩ hTx1 , Tx2)i ≤ n and this concludes the proof. (cid:3) We are ready to prove our main result that connects twd and defectivity. Theorem 18. Let X ⊂ PN be an irreducible, reduced, and non degenerate variety of dimension n. Assume that: a) X is k-twd, b) X is not (k − 1)-twd c) k > n and N ≥ (k + 1)(n + 1) − 1. Then πX k+1 is of fiber type. Proof. Thanks to Lemma 17 we may assume that the contact locus is of type I. By hypothesis the variety X is k-twd. Let A = {x1, . . . , xk} ⊂ X be a set of general points and Γ := Γ(A) the associated contact locus of dimension γ > 0. Let z ∈ hAi be a general point, τk := τ X : X 99K PM the associated k-tangential projection, k and y ∈ X a general point. For a general set Y := {y1, . . . , yk−1} ⊂ Γ let Γ(Y ∪{y}) be the contact locus associated to {y1, . . . , yk, y}. Assume that πX k+1 is not of fiber type. then, by i) in Lemma 16 τk is not of fiber type and by point iii) in Lemma 16, τk(Γ(Y ∪ {y}) is a linear space of dimension γ through z := τk(y). This gives a map χ : Hilbk−1(Γ) 99K G(γ − 1, M − 1). The point z is smooth hence all these linear spaces sit in Tzτk(X) ∼= Pn. In other words we have a map χ : Hilbk−1(Γ) 99K G(γ − 1, n − 1) ⊂ G(γ − 1, M − 1). Note that dim G(γ − 1, n − 1) = γ(n − γ) and dim Hilbk−1(Γ) = (k − 1)γ. By hypothesis k > n and γ > 0 hence we have (k − 1)γ > γ(n − γ). Then the map χ is of fiber type and fibers have, at least, dimension γ(k − n+ γ − 1). 8 ALEX CASAROTTI AND MASSIMILIANO MELLA Set [Y1], [Y2] ∈ χ−1([Λ]) general points, for [Λ] ∈ χ(Hilbk−1(Γ)) ⊂ G(γ − 1, n− 1) a general point. The variety X is not (k − 1)-twd and we are assuming that πX is not of fiber type therefore, by ii) in Lemma 16, k+1 dim(Γ ∩ Γ(Yi ∪ {y})) = 0, in a neighborhood of yi. Since the fiber of χ is positive dimensional we have Γ(Y1 ∪ {y}) 6⊃ Y2. (6) The contact loci are irreducible then, by Equation (6), we conclude that Γ(Y1 ∪ {y}) 6= Γ(Y2 ∪ {y}). Therefore, by point iii) in Lemma 16, the positive dimensional fiber of χ induces a positive dimensional fiber of τk and we derive, by point i) Lemma 16, the contra- diction that that πX (cid:3) k+1 is of fiber type. Remark 19. Both assumption b) and c) alone are reasonable and not over- demanding. Unfortunately the combination of them is quite restrictive and narrows the range of application we are aiming at. We believe the statement is not optimal with respect to assumption c). But we are not sure it is true, in full generality, without any assumption of this kind. On the other hand we strongly believe that for many interesting varieties, like Segre, Grassmannian, Veronese and their combinations, twd can occur only one step before the secant map becomes of fiber type. This is not the case for weakly defectiveness as is shown in [BV]. In [BV, Theorem 1.1 a)] it is proven that G(2, 7) is 2 and 3 weakly defective without being 3-defective. Note that this variety is 3-twd and it is not 2-twd. The next result generalizes the main result in [BBC] and it allows to avoid the bottleneck introduced by conditions b) and c) of Theorem 18 in many interesting situations. Lemma 20. Let X ⊂ PN be an irreducible, reduced, and non degenerate variety. Assume that X is not 1-twd and πX k+1 is generically finite, in particular X is not (k + 1)-defective. If X is k-twd then γk < γk+1. Proof. The variety X is not 1-twd, then we may assume, without loss of generality, that γk−1 < γk = γk+1. Then γk+1 < n and Seck+1(X) ( PN , hence by e) in Theorem 13, the contact loci are of type II and linearly independent linear spaces. Fix {x1, . . . , xk, y} ⊂ X a set of general points and let Γ(x1, . . . , xk, y) = ∪k 1 Pi ∪ Py the contact locus. Moreover the assumption γk = γk+1 and point a) in Theorem 13 force Γ(x1, . . . , xk−1, y) = ∪k−1 1 Pi ∪ Py, with the same Pi's. Then hTx1 X, . . . , Txk−1 X, TyXi ⊃ hTzXiz∈Pi,i=1,. . . , k-1 \y∈X We are assuming that γk−1 < γk therefore and we have a proper inclusion Pi 6⊂ Γ(x1, . . . , xk−1), hTzXiz∈Pi,i=1,. . . , k-1 ) hTx1 X, . . . , Txk−1 Xi. FROM NON DEFECTIVITY TO IDENTIFIABILITY 9 Set for general points y1, y2 ∈ X. Then we have MAi = hTx1 X, . . . , Txk−1 X, Tyi Xi, MA1 ∩ MA2 ⊃ hTzXiz∈Pi,i=1,. . . , k-1 ) hTx1 X, . . . , Txk−1 Xi. and we conclude that This shows that (MA1 ∩ MA2) ∩ Tyi 6= ∅. hTx1 X, . . . , Txk−1 X, Ty1 Xi ∩ Ty2 X 6= ∅. hence the k-tangential projection τ X the contradiction that πX k+1 is of fiber type. k is of fiber type and by Lemma 16 we derive (cid:3) Remark 21. Let us recall that 1-twd varieties are classified in [GH] and are es- sentially generalized developable varieties. In particular they are ruled by linear spaces and, with the unique exception of linear spaces, they are singular. We are ready to apply the above results to get not tangentially weakly defec- tiveness and hence identifiability statements. Corollary 22. Let X ⊂ PN be an irreducible, reduced, and non degenerate variety that is not 1-twd, for instance a smooth variety or a variety that is not covered by linear spaces. Assume that πX k is generically finite and k ≥ dim X. Then X is not (k − dim X)-twd and it is not (k − dim X + 1)-twd if πX dominant. If moreover either k > 2 dim X or πX then X is not (k − 1)-twd. k is not k is not dominant and k ≥ 2 dim X In all the above cases X is h-identifiable. Proof. By hypothesis πh is generically finite for any h ≤ k. Then by Theorem 20 if it is j-twd γj < γj+1. The contact locus is a subvariety of X, hence γk−dim X = 0. This proves the first statement. If πX k is not dominant then the contact locus is a proper subvariety and we have γk−dim X+1 = 0. Assume that k ≥ 2 dim X then by the first part X is not j-twd for some j > dim X. Then we apply Theorem 18 recursively to conclude. We derive identifiability by Proposition 14. (cid:3) Remark 23. The first part of Corollary 22 extends the bounds in [BBC] to non 1-twd varieties. The main novelty is the second part that allows to derive identifi- ability from non defectivity for large enough secant varieties. 3. Application to tensor and structured tensor spaces As we already mentioned identifiability is particularly interesting for tensor spaces. In this section we use our main result to explicitly state identifiability of a variety of tensor spaces. For this we will consider Segre, Segre-Veronese and Grassmannian varieties and their h-twd properties. We start with some notation Notation 24. The variety Σ(d1, ..., dr; n1, ..., nr) is the Segre-Veronese embedding of Pn1 × ... × Pnr in PQ(ni+di ni )−1 via the complete linear system O(d1, . . . , dr). 10 ALEX CASAROTTI AND MASSIMILIANO MELLA When all di's are one we have the Segre embedding and we let Xn1,...,nr := n := Σ(1, . . . , 1; n, . . . , n) ∼= (Pn)r. The expected Σ(1, . . . , 1; n1, . . . , nr) and X r generic rank is Using the notations in [AOP] we define ⌉ gr(Σ(d1, ..., dr; n1, ..., nr)) = ⌈ Q(cid:0)ni+di ni (cid:1) (P ni) + 1 s(Σ(d1, ..., dr; n1, ..., nr) := ⌊ Q(cid:0)ni+di ni (cid:1) (P ni) + 1 ⌋. For simplicity in the case n1 = ... = nr = n and d1 = ... = dr = 1 we set sr n := s(Σ(d1, ..., dr; n1, ..., nr)) The variety G(k, n) is the Grassmannian parameterizing k−planes in Pn embed- k+1 ded in P( V V ) via the Plucker embedding. The expected generic rank is gr(G(k, n)) = ⌈ (n − k)(k + 1) + 1 ⌉ (cid:0)n+1 k+1(cid:1) Remark 25. Note that we always have s(Σ(d1, ..., dr; n1, ..., nr) ≥ gr(Σ(d1, ..., dr; n1, ..., nr)) − 1 and equality occurs only when h < s(X) we have Sech(X) ( PN . Q(ni+di ni ) (P ni)+1 is not an integer. In particular for any The defectivity of Segre and Segre-Veronese varieties is in general very far from being completely understood, [AOP] [AB] [AMR], but it is in better shape than their identifiability. For the latter the best asymptotic bounds we are aware of is in [BBC]. We start proving the theorem in the introduction. Theorem 26. Let X = X k 1 in the following range: ∼= (P1)k. Then X is not h-twd and hence h-identifiable - (k, h) = (2, 1), (3, 2), (4, 2), (5, 4), (6, 9), - k ≥ 7 h < s(X) Proof. For k ≤ 5 this is well known, and can be easily checked also via a direct computation with commutative algebra software. For k = 6 this has been checked in [BC] by a computer aided computation. Let us fix k ≥ 7. By [CGG, Theorem 4.1] X is never defective. In particular the morphism πX h is generically finite for h ≤ sk 1. When k ≥ 7 we have 2 dim X = 2k < 2k k + 1 − 1 < sk 1 , then we can apply Corollary 22. (cid:3) Remark 27. The Theorem answers positively Conjecture 1.2 in [BC] when the 2k k+1 ∈ N. For k ≤ 6 the one listed are the only generic rank is an integer, that is identifiable cases. For 3-factors Segre we plug [CO] directly in Theorem 18 to get the following. Theorem 28. Let X = X 3 n. Then X is h-identifiable for h < s(X). FROM NON DEFECTIVITY TO IDENTIFIABILITY 11 Proof. For n ≤ 7 the statement is proved in [CO, Theorem 1.2]. For n > 7, by [Li], the variety X is not h-defective for h ≤ s3 n and by the results in [CO] X is not h-twd for h = 3n, confront the table in [CO, Theorem 1.2] . Then we are in the condition to apply Theorem 18 recursively to prove that X is not h-twd, and hence identifiable, for h < s3 n. (cid:3) For general diagonal Segre we have a similar statement using [AOP]. Theorem 29. Let X = X k n, with n ≥ 2 and k ≥ 4. Let n ≥ δ(X) ≡ sk n mod(n + 1) Then X is not h-twd and hence h−identifiable for h < s(X) − δ(X). In particular when δ(X) = 0 X is h-identifiable for all h < s(X). Proof. Using the notations in [CO, Theorem 6.7] let α be the greatest integer such that n + 1 ≥ 2α. First we prove the statement for all but finitely many cases. Claim 1. If (k, n) 6∈(cid:26) (k,6) with k ≤ 6, (k,4) with k ≤ 5, then X is h−identifiable for h < s(X) − δ(X) (k,5) with k ≤ 5, (k,3) with k ≤ 4 (cid:27) Proof. By [AOP, Theorem 5.2] we know that X is not h-defective as long as h ≤ s(X) − δ(X). The variety X k n is not h−twd for h ≤ 2(k−1)α−(k−1) = 2(k−1)(α−1) by [CO, Theorem 6.7]. Let us assume that n 6= 2. A short hand computation shows that 2(k−1)(α−1) > dim(X) = kn is satisfied for every (k, n) in the list. Then, using recursively Theorem 18, we conclude. (cid:3) For the case n = 2 it is easy to check that the inequality sk 2 = ⌊ 3k 2k + 1 ⌋ − δ(X k 2 ) > 4k = 2dim(X k 2 ) is satisfied for every k ≥ 5 and so we can conclude using Corollary 22. When (k, n) = (6, 6), (5, 6) we have the inequalities s6 6 − δ(X 6 6 ) > 2 · 36 = 2 dim X 6 6 and s5 6 − δ(X 5 Then we conclude by Corollary 22. 6 ) > 2 · 30 = 2 dim X 5 6 . For all the remaining cases we have that (n + 1)k ≤ 15000 and we may use the computation in [COV1, Theorem 1.1] to conclude the required identifiability. (cid:3) The next class of Segre varieties we treat in details is given by X[k, n] := Pk × (Pn)k+1. For these varieties we have gr(X[k, n]) = (k + 1)(n + 1)k+1 (k + 1)n + k + 1 = (n + 1)k. In particular gr(X[k, n]) = s(X[k, n]) is always an integer, that is X[k, n] is always perfect. Thanks to this special condition we have the following. Theorem 30. Let X = X[k, n] with n odd and k > 1. Then X is h−identifiable for h < gr(X). 12 ALEX CASAROTTI AND MASSIMILIANO MELLA Proof. The proof is entirely similar to that of Theorem 29. Indeed by [AOP, The- orem 5.11] we know that all these Segre are non defective. If the inequality (k, n) 6= (4, 1), (3, 1), (2, 1), (2, 3), (2, 5) (n + 1)k > 2(k + kn + n) = 2dim(X) is satisfied and we conclude using Corollary 22. For all the exceptional cases we have (k + 1)(n + 1)k+1 ≤ 15000 hence we may apply [COV1, Theorem 1.1]. (cid:3) Remark 31. Defective Segre are expected to be quite rare, beside the unbalanced ones, see the conjecture in [AOP]. This conjecture has been checked via a com- puter in many cases, [Va] [COV1]. For all these special values our argument gives identifiability confirming the numerical computation in [COV1]. Next we apply the same strategy to Segre -- Veronese varieties. For this class of varieties the defectivity results are much weaker and so are our bounds. Again the special case of binary forms is in better shape. We start recalling the notation of [LP]. Definition 32. We say that (d1, ..., dr; n) is special if (d1, ..., dr; n) = (2, 2a; 2a + 1), (1, 1, 2a; 2a + 1), (2, 2, 2; 7), (1, 1, 1, 1; 3) for a ≥ 1. Otherwise (d1, ..., dr; n) is called not special. Theorem 33. Let X = Σ(d1, ..., dr; 1, .., 1) with r = dimX. Assume that (d1, ..., dr; n) is not special and r ≥ 6. Then X is h−identifiable for h < s(X). Proof. We are assuming that (d1, ..., dr; n) is not special. Then, by [LP, Theorem 2.1], the variety X is not h-defective for h ≤ gr(X). Thanks to Theorem 26 we may assume, without loss of generality that d1 > 1 and we have s(X) = ⌊ (d1 + 1) · · · (dr + 1) r + 1 ⌋ ≥ 3 · 2r−1 r + 1 − 1. In particular holds for every r ≥ 6. 3 · 2r−1 r + 1 − 1 > 2r = 2dimX The variety X is not 1-twd and so we conclude by Corollary 22. (cid:3) For general Segre-Veronese we have the following. Theorem 34. Let X := Σ(d1, ..., dr, n1, ..., nr) be the Segre -- Veronese variety. As- sume r ≥ 2, and set d = d1 + . . . + dr. Then X is h−identifiable for h ≤ n⌊log2(d−1)⌋ 1 − 1. n⌊log2(d−1)⌋ 1 ≥ 2(n1 + . . . + nr), Proof. By [AMR, Theorem 1.1] X is not h-defective for In our numerical assumptions Sech(X) ( PN and we may assume h ≤ n⌊log2(d−1)⌋ 1 − (n1 + ... + nr) + 1. h ≥ 2 dim X. Then we conclude by Corollary 22. (cid:3) Remark 35. For the Veronese variety of Pn, that is Σ(d1, n1) it is easy, via Corol- lary 22 and [AH], to reprove the identifiability results in [Me] and [COV2]. Let and N = (m + 1)(cid:0)n+2 2 (cid:1) − 1 r(m, n) =(m3 − 2m (m−2)(m+1)2 2 if m even and n odd otherwise s(X) = ⌊ (m + 1)(cid:0)n+2 2 (cid:1) m + n + 1 ⌋ FROM NON DEFECTIVITY TO IDENTIFIABILITY 13 As in the Segre case, for special classes of Segre -- Veronese there are better non defectivity results. Here we recall the notation in [AB]. Let X :=P(1, 2; m, n) be the Segre-Veronese variety Pm × Pn embedded by O(1, 2) in PN where the meaningful numbers of X. With this in mind we have the following. Corollary 36. Let X =P(1, 2; m, n). If n > r(m, n) and ⌋ ≥ 2(m + n) ⌊ (m + 1)(cid:0)n+2 2 (cid:1) m + n + 1 then X is not h−twd and hence h−identifiable for h < s(X). Proof. In our range X is not h-defective by [AB, Theorem 1.1] and Sech(X) ( PN . Moreover s(X) = ⌊ ⌋ ≥ 2(m + n) = 2 dim X (m + 1)(cid:0)n+2 2 (cid:1) m + n + 1 and we may apply Corollary 22 to conclude. (cid:3) Let us consider now the case of Pm × Pn embedded with O(1, d) for d ≥ 3. Corollary 37. Let X =P(1, d; m, n) with d ≥ 3 and m, n ≥ 1. Let s(X) = max(s ∈ Ns is a multiple of (m + 1) and s ≤ ⌊ (m + 1)(cid:0)n+d d (cid:1) m + n + 1 ⌋) If s(X) > 2(m + n) then X is not h-twd and hence h-identifiable for h < s(X). Proof. By [BCC, Theorem 2.3] X is not h-defective for h ≤ s(X) and Sech(X) ( P(m+1)(n+d d )−1. X is smooth, in particular it is not 1-twd. Since s(X) > 2(m + n) = 2dim(X) we can apply Corollary 22 to conclude. (cid:3) Remark 38. Similar statements about subgeneric identifiability of Pn × P1 em- bedded with O(a, b) can be derived applying Corollary 22 using the non defectivity results in [BBC1]. Finally we consider Grassmannian varieties. For this class of tensor spaces very few is known about identifiability. To the best of our knowledge the following is the first non computer aided result for them. Theorem 39. Let X = G(k, n) such that 2k + 1 ≤ n. Assume that k + 1(cid:19)⌊log2(k)⌋ ⌊(cid:18) n + 1 . Then X is h−identifiable for ⌋ ≥ 2(n − k)(k + 1) h ≤(cid:18) n + 1 k + 1(cid:19)⌊log2(k)⌋ − 1 14 ALEX CASAROTTI AND MASSIMILIANO MELLA Proof. By [MR, Theorem 5.4] in our numerical range X is not h−defective and Sech(X) ( PN . Then we conclude by Corollary 22. (cid:3) The technique we developed can be applied to many other classes of varieties, once it is known their defectivity behavior. As a sample we conclude with the following example. Example 40. C. Am´endola, J.-C. Faugre, K. Ranestad and B. Sturmfels in [AFS] and [ARS] studied the Gaussian moment variety G1,d ⊂ Pd whose points are the vectors of all moments of degree ≤ d of a 1−dimensional Gaussian distribution. They proved that G1,d is a surface for every d and Sech(G1,d) has always the expected dimension. In [BBC, Example 5.8] it is shown that G1,d is not uniruled by lines, in particular it is not 1-twd. As usual let s(G1,d) = ⌊ d + 1 3 ⌋ ≥ gr(G1,d) − 1 Then by Corollary 22 G1,d is h-identifiable, for h < s(G1,d) when d ≥ 14. [AB] [AOP] [AH] [AFS] [ARS] [AMR] [BBC] [BBC1] [BDdG] [BCC] [BV] [BC] [BCO] [BO] [CGG] [Ch] [CC02] [CC10] [CMO] [CO] References Abo, H; Brambilla, M.C. Secant Varieties of Segre-Veronese Varieties Pm × Pn Em- bedded by O(1, 2) Experiment. Math. 18 (2009), no. 3, 369-384. Abo, H; Ottaviani, G; Peterson, C Induction for secant varieties of Segre varieties Transactions of the AMS 361 (2009), no. 2, 767-792. Alexander, J.; Hirschowitz, A. The blown-up Horace method: application to fourth- order interpolation Invent. Math. 107 (1992), no. 3, 585602. Am´endola, C; Faugre J.-C.; Sturmfels, B Moment varieties of Gaussian mixtures J. Algebr. Stat. 7 (2016), 1428. Am´endola, C; Ranestad, K; Sturmfels, B Algebraic Identifiability of Gaussian Mixtures Int. Math. Res. Not. IMRN, DOI: 10.1093/imrn/rnx090 (2016,to appear) Preprint Araujo, C; Massarenti, A; Rischter, R On non-secant defectivity of Segre-Veronese varieties Transactions of the AMS 371 (2019), no.4, 2255-2278. Ballico, E; Bernardi, A; Chiantini, L On the dimension of contact loci and the iden- tifiability of tensors Ark. Mat. 56 (2018), no. 2, 265-283. Ballico, E; Bernardi, A; Catalisano, M.V. Higher secant varieties of Pn × P1 embedded in bi-degree (a, b) Comm. Algebra 40 (2012), 3822-3840 Baur, K.; Draisma, J. ; de Graaf, W., Secant dimensions of minimal orbits: compu- tations and conjectures, Exp. Math. 16 (2007), 239250. Bernardi, A; Carlini, E; Catalisano, M.V. Higher secant varieties of Pn ×Pm embedded in bi-degree (1, d) Journal of Pure and Applied Algebra, 215 (2011), no. 12, 2853-2858 Bernardi, A.; Vanzo, D. A new class of non-identifiable skew-symmetric tensors Ann. Mat. Pura Appl. (4) 197 (2018), no. 5, 1499-1510. Bocci, C.; Chiantini, L. On the identifiability of binary Segre products J. Algebraic Geom. 22 (2013), no. 1, 1-11. Bocci, C.; Chiantini, L.; Ottaviani, G. Refined methods for the identifiability of tensors Ann. Mat. Pura Appl. (4) 193 (2014), no. 6, 1691-1702. Boralevi, A; A note on secants of Grassmannians Rendiconti dellIstituto di Matem- atica dellUniversit di Trieste, Volume 1 (2013) Catalisano, M.V.; Geramita, A.; Gimigliano, A. Secant varieties of P1 × · · · × P1 are not defective for n ≥ 5. J. Algebraic Geom. 20 (2011), no. 2, 295-327. Chiantini, L. private comunication. Chiantini, L., Ciliberto, C.: Weakly defective varieties Trans. Am. Math. Soc. 354, 151-178 (2002) Chiantini,L Ciliberto C.: On the dimension of secant varieties. J. Eur. Math. Soc. (JEMS) 12 (2010), no. 5, 1267-1291. Chiantini, L.; Mella, M.; Ottaviani, G. One example of general unidentifiable tensors J. Algebr. Stat. 5 (2014), no. 1, 64-71. Chiantini, L., Ottaviani, G.: On generic identifiability of 3-tensors of small rank SIAM. J. Matrix Anal. Appl. 33, 1018-1037 (2012) FROM NON DEFECTIVITY TO IDENTIFIABILITY 15 [COV1] [COV2] [DDL1] [DDL2] [DDL3] [GH] [KADL] [Kr] [LP] [Li] [MR] [Me] [Si] [Va] Chiantini, L., Ottaviani, G., Vannieuwenhoven, N.: An algorithm for generic and low-rank specific identifiability of complex tensors SIAM. J. Matrix Anal. Appl. 35, 1265-1287 (2014) Chiantini, L.; Ottaviani, G.; Vannieuwenhoven, N. On generic identifiability of sym- metric tensors of subgeneric rank Trans. Amer. Math. Soc. 369 (2017), no. 6, 4021- 4042. Domanov, I.; De Lathauwer, L., On the uniqueness of the canonical polyadic decompo- sition of third-order tensorsPart I: Basic results and uniqueness of one factor matrix, SIAM J. Matrix Anal. Appl. 34 (2013), 855875. Domanov, I.; De Lathauwer, L., On the uniqueness of the canonical polyadic decom- position of third-order tensorsPart II: Uniqueness of the overall decomposition, SIAM J. Matrix Anal. Appl. 34 (2013), 876903. Domanov, I.; De Lathauwer, L., Generic uniqueness conditions for the canoni- cal polyadic decomposition and INDSCAL, SIAM J. Matrix Anal. Appl. 36 (2015), 15671589. Griffiths, P.; Harris, J. Algebraic geometry and local differential geometry Ann. Sci. cole Norm. Sup. (4) 12 (1979), no. 3, 355-452. Karfoul, A; Albera, L.; De Lathauwer, L. Iterative methods for the canonical decom- position of multi-way arrays: Application to blind underdetermined mixture identifi- cation Signal Processing 91, Issue 8 2011, 1789-1802 Kruskal, J. B., Three-way arrays: rank and uniqueness of trilinear decompositions, with application to arithmetic complexity and statistics, Linear Algebra Appl. 18 (1977), 95138. Laface, A.; Postinghel, E. Secant varieties of Segre-Veronese embeddings of (P1)r eprint arXiv:1105.2136, 2011 Lickteig, T. Typical tensorial rank Linear Algebra Appl. 69 (1985) 95-120. Massarenti, A; Rischter, R Non-secant defectivity via osculating projections Annali della Scuola Normale Superiore di Pisa, Vol XIX, Issue 1 (2019), 1-34. Mella, M. Singularities of linear systems and the Waring problem Trans. Amer. Math. Soc. 358 (2006), no. 12, 5523-5538. Sidiropoulos, N. D; Bro, R., On the uniqueness of multilinear decomposition of N-way arrays, J. Chemom. 14 (2000), 229239. Vannieuwenhoven, N. A randomized algorithm for testing nonsingularity of structured matrices with an application to asserting nondefectivity of Segre varieties, IMA J Numer. Anal. (2014), 1-34 Dipartimento di Matematica e Informatica, Universit`a di Ferrara, Via Machiavelli 35, 44121 Ferrara, Italy E-mail address: [email protected] Dipartimento di Matematica e Informatica, Universit`a di Ferrara, Via Machiavelli 35, 44121 Ferrara, Italy E-mail address: [email protected]
1204.4740	1	1204	2012-04-20T20:21:57	SYZ duality for parabolic Higgs moduli spaces	[ "math.AG" ]	We prove the SYZ (Strominger-Yau-Zaslow) duality for the moduli space of full flag parabolic Higgs bundles over a compact Riemann surface.The SYZ duality was proved for moduli spaces of Higgs vector bundles over a compact Riemann surface by Hausel and Thaddeus.	math.AG	math	SYZ DUALITY FOR PARABOLIC HIGGS MODULI SPACES INDRANIL BISWAS AND A. DEY Abstract. We prove the SYZ (Strominger-Yau-Zaslow) duality for the moduli space of full flag parabolic Higgs bundles over a compact Riemann surface. In [HT2], the SYZ duality was proved for moduli spaces of Higgs vector bundles over a compact Riemann surface. 1. Introduction 1.1. Mirror symmetry and SYZ duality. Mirror symmetry was discovered in the late 1980's by physicists studying superconformal field theories. Let X be an n dimen- sional complex Calabi-Yau manifold with a Ricci-flat Kahler form ω and a nowhere van- ishing holomorphic n-form Ω on X. A submanifold Z ⊂ X of real dimension n is called Lagrangian if ωZ = 0; further a Lagrangian submanifold is said to be special if (Im Ω)Z = 0. After simplifying a great deal, mirror symmetry is an one-to-one duality (in an appropriate sense) between two class of objects: (1) Pairs of the form (Z , L), where Z is a holomorphic submanifold of X and L is a holomorphic line subbundle on Z (such a pair is called a holomorphic D-brane). (2) Special Lagrangian D-branes, which is a pair (bZ , bL) where bZ is a special La- grangian submanifold of a certain Calabi-Yau manifold bX (mirror partner), and bL is a flat U(1) line bundle on bZ. Since any point x ∈ X is a submanifold, it should correspond to a pair (bZ , Λ), where bZ is a special Lagrangian submanifold of a fixed Calabi-Yau manifold bX and Λ is a flat U(1)-line bundle on bZ. By a theorem of McLean, the deformation space for a spe- cial Lagrangian manifold bZ is unobstructed and is parametrized by H 1(bZ , R), hence dim H 1(bZ, R) = dimC(X) = n. The moduli space of flat U(1) line bundles is given by the torus H 1(bZ, R)/H 1(bZ, Z). This gives a hint that a moduli of special Lagrangian submanifolds of a fixed Calabi-Yau manifold should have a n-torus fibration over an affine base of real dimension n. Motivated by this Strominger-Yau-Zaslow made a conjecture [SYZ]. SYZ Conjecture: If X and bX are mirror pair of Calabi-Yau n-folds, then there exist fibrations f : X −→ B and bf : bX −→ B whose fibers are special Lagrangian such that canonically Xb = H 1(cXb, S1) and cXb = H 1(Xb, S1), whenever the fibers Xb and cXb are the general fiber is an n-torus. Furthermore, these fibrations are dual, in the sense that non-singular tori. 2000 Mathematics Subject Classification. 14D20, 14D21. Key words and phrases. Parabolic bundle, Higgs field, SYZ duality, gerbe. 1 2 I. BISWAS AND A. DEY 1.2. Reformulation of the SYZ conjecture in terms of unitary gerbes. Hitchin introduced the notion of a flat unitary gerbe (known as B-fields to physicists) and re- formulated the SYZ conjecture in terms of this B-fields [Hi1]. To make sense one needs a further assumption that the special Lagrangian fibers are linearly equivalent for both X and bX. Then by [Hi1, Theorem 3.3], the restriction map for the second cohomology H 2(X, R) −→ H 2(Xb, R) is zero. This means that the restriction map H 2(X, S1) −→ H 2(Xb, S1) is trivial. So the flat unitary gerbe B has trivial holonomy on each torus fiber (hence trivial). Therefore, one should work with pairs (Xb, T ), where Xb is a special Lagrangian submanifold and T is a flat trivialization of the gerbe B on Z. The modified mirror conjecture as proposed by Hitchin, [Hi1], is the following: Conjecture: The mirror of a Calabi-Yau manifold X with a B-field is the moduli space of pairs (Z , T ), where Z is a special Lagrangian submanifold of X and T is a flat trivialization of the gerbe B on Z. Two Calabi-Yau n-orbifolds M and cM, equipped with flat unitary gerbes B and bB respectively, are said to be SYZ mirror partners if there is an orbifold N of real dimension n and there are smooth surjections µ : M −→ N , bµ : cM −→ N such that for every x ∈ N which is a regular value of µ and bµ, the fibers Lx := µ−1(x) ⊂ M and bLx := bµ−1(x) ⊂ cM are special Lagrangian tori which are dual to each other in the sense that there are smooth identifications Lx = TrivU 1( Lx, B) and bLx = TrivU 1(Lx, B) that depend smoothly on x. 1.3. The result of Hausel and Thaddeus. The moduli spaces of Higgs bundles admit natural dual pairs of hyper-Kahler integrable systems [Hi2], [Hi3]. The hyper-Kahler met- ric and the collection of Poisson-commuting functions determining the integrable system produce a family of special Lagrangian tori on the moduli spaces, which is a key require- ment of SYZ conjecture. Moreover, the families of tori on the SL(r, C) and PGL(r, C) moduli spaces are dual in the appropriate sense, which is the other requirement of SYZ conjecture. This work of Hausel and Thaddeus was extended to principal G2 in [Hi4]. In [DP], this was extended to all semisimple groups. (See related works [FW], [GW] and [Wi].) In [HT1], Hausel-Thaddeus made an announcement that moduli spaces of SL(r, C) and PGL(r, C) parabolic Higgs bundles are mirror partner to each other (in the sense of SYZ) and their stingy E polynomials are same. In [HT2], they gave a proof of this conjecture in non-parabolic case. Our aim here is to address the case of parabolic vector bundles with complete quasi- parabolic flags. We follow the proof of Hausel-Thaddeus; the key ingredient here is the identification of the parabolic Hitchin fiber as the Prym variety of a certain spectral cover (which was done in [BM], [GL]). It should be clarified that the Higgs fields that we consider have nilpotent residue. The corresponding moduli space forms a symplectic leaf of the moduli space of Higgs bundles SYZ DUALITY FOR PARABOLIC HIGGS MODULI 3 for which the residue of the Higgs field satisfies the weaker condition that it is only flag preserving. Acknowledgements. We thank the referee for pointing out references. The second author would like to thank J. Martens for an useful discussion. He also thanks TIFR for hospitality while some part of this work was done. Both authors thank The Institute of Mathematical Sciences at Chennai for hospitality. 2. Preliminaries 2.1. Parabolic Higgs bundles. Let X be an irreducible smooth projective curve over C of genus g, with g ≥ 2. Let D ⊂ X be a nonempty finite subset of n points. A quasi-parabolic structure, over D, on a holomorphic vector bundle E −→ X is a filtration Ex = Ex,0 ⊃ Ex,1 ⊃ · · · ⊃ Ex,rx ⊃ Ex,rx+1 = {0} for each x ∈ D. A parabolic structure on E is a quasi-parabolic structure as above together with rational numbers 0 ≤ αx,0 < αx,1 < · · · < αx,rx < 1 , which are called parabolic weights. A parabolic vector bundle over X of rank r is a holo- morphic vector bundle of rank r on X equipped with a quasi-parabolic structure over D together with parabolic weights. The system of parabolic weights {(αx,0 , · · · , αx,rx)}x∈D will be denoted by α. For a parabolic vector bundle E∗ = (E, {Ex,i}, α∗), the parabolic degree is defined to be par-deg (E∗) = deg(E) + X rxX αx,i , x∈D i=0 and the parabolic slope is defined to be par-µ (E∗) := par-deg (E∗)/rk(E). Any holomor- phic subbundle F of E has a parabolic structure induced by the parabolic structure on E; the resulting parabolic vector bundle will be denoted by F∗. A parabolic bundle is said to be stable (respectively, semistable) if for all holomorphic subbundles F ⊂ E with 0 < rk(F ) < rk(E), (2.1) par-µ (F ) < par-µ (E) (respectively, par-µ (F ) ≤ par-µ (E)) . The moduli space of semistable parabolic vector bundles of rank r and degree d with fixed parabolic data was constructed by Mehta and Seshadri, [MS], using Mumford's Geometric Invariant Theory. This moduli space, which we will denote by PMd α, is smooth for a generic choice of weights α. We recall that given rank and degree, a system of parabolic weights α with multiplication is called generic if the semistability condition implies the stability condition. Consider the determinant morphism (2.2) det : PMd α −→ Jacd(X) that sends a parabolic vector bundle E∗ to Vtop E. Choose Λ ∈ Jacd(X), and define (2.3) PMΛ := det−1(Λ) . 4 I. BISWAS AND A. DEY So PMΛ is a moduli space of twisted SL(r, C)–bundles with parabolic structure (see [BLS, Section 2] for twisted SL(r, C)–bundles). For any other line bundle Λ1 ∈ Jacd(X), the morphism det−1(Λ) −→ det−1(Λ1) that sends any parabolic vector bundle E∗ to E∗ ⊗ ζ, where ζ is a fixed r-th root of Λ1 ⊗ Λ−1, is an isomorphism. Thus the isomorphism class of the moduli space PMΛ does not depend on the choice of Λ ∈ Jacd(X). The abelian variety Pic0(X) = Jac0(X) acts on PMd α via (L, E∗) 7−→ L⊗E∗ . The quotient ^ PMd α := PMd α/ Pic0(X) , which exist as a projective variety by [Se], is the component of the moduli space of parabolic PGL(r, C)–bundles corresponding to degree d (the connected components of the moduli space of parabolic PGL(r, C)–bundles are irreducible). Let (2.4) be the group of r-torsion points of the Jacobian; it is isomorphic to (Z/rZ)2g, in particular, α preserves the subvariety PMΛ defined in (2.3). its order is r2g. The action of Γ on PMd Let Γ := Pic0(X)[r] ^ PMΛ := PMΛ/Γ be the quotient; it is a projective variety by [Se]. Let K be the holomorphic cotangent bundle of X. The line bundle K ⊗ OX (D) will be denoted by K(D). A parabolic Higgs bundle is a pair (E∗ , Φ), where E∗ is a parabolic vector bundle and Φ : E −→ E ⊗ K(D) is a homomorphism which is strongly parabolic, meaning Φ(Ex,i) ⊂ Ex,i+1 ⊗ K(D)x for each point x ∈ D and i ∈ [0 , rx]. A parabolic Higgs bundle (E∗ , Φ) is called (semi)-stable if the slope condition in (2.1) Higgs denote the moduli space of semistable para- holds whenever Φ preserves F . Let PMd bolic Higgs bundles of rank r and degree d with the given parabolic data. Let PMΛ Higgs ⊂ PMd Higgs be the subvariety consisting of all (E∗ Φ) such that det(E) = Λ and trace(Φ) = 0. For E∗ ∈ PMd α, TE∗PMd α = H 1(X, End(E∗)) [Yo1], [Yo2], where End(E∗) is the sheaf of endomorphisms of the underlying vector bun- dle E preserving the quasi-parabolic filtrations. Applying the parabolic analog of Serre duality, [Yo1, Section 3], TE∗PMd α = H 0(X, SEnd(E∗) ⊗ K(D))∗ , SYZ DUALITY FOR PARABOLIC HIGGS MODULI 5 where SEnd(E∗) ⊂ End(E∗) is the strongly parabolic endomorphisms. Hence the total space T ∗ Higgs. This map is an open embed- ding. α of the cotangent bundle maps to PMd E∗PMd The group Γ acts on PMΛ Higgs via tensor product (the Higgs field does not change). The quotient PMΛ Higgs/Γ will be denoted by ^ PMΛ Higgs. 2.2. Parabolic Hitchin system. In this subsection we recall the Hitchin map and the spectral curve for a parabolic Higgs bundle (see [BM], [GL], [LM] for details). For notational convenience, the line bundle K ⊗a ⊗ OX(bD) will be denoted by K aDb. The parabolic Hitchin space is defined as H = H 0(K 2D) ⊕ · · · ⊕ H 0(K rDr−1) . The characteristic polynomial or trace map of a Higgs field defines the parabolic Hitchin map (2.5) Higgs −→ H It is known that this morphism h is proper (see [Yo2]). h : PMΛ Let Z := Spec Sym•(K −1 ⊗ OX(D)−1) be the total space of the line bundle K(D) which is a quasi-projective surface. Let (2.6) p : Z −→ X be the natural projection. For s ∈ H, there exists an algebraic curve, denoted as Xs, in Z which is known as the spectral curve. We will very briefly recall it (for details see [BM], [GL]). For any (s1, · · · , sr−1) ∈ H, consider the map S from Z to the total space of the line bundle K rDr given by z 7−→ z⊗r + z⊗(r−2) ⊗ s1(p(z)) + · · · + sr−1(p(z)) ∈ (K rDr)p(z) where z ∈ Z, and p is the projection in (2.6). The inverse image S−1(0X) ⊂ Z, where OX ⊂ K rDr is the zero section, is the parabolic spectral curve associated to s := (s1, · · · , sr−1). This parabolic spectral curve will be denoted by Xs. The restriction of p to Xs will again be denoted as p. Henceforth, we assume that for each point x ∈ D, the quasi-parabolic flag is complete. In other words, rx = r − 1. There is a Zariski open dense subset U ⊂ H such that for any s ∈ U the spectral curve Xs is smooth and connected [GL, Lemma 3.1]; the assumption that the quasi-parabolic flags are complete is needed for this. The fiber over s is isomorphic to Prymd′ (Ys) = {L ∈ Picd′ (Ys) det(π∗(L)) = ξ} , where d′ = d + r(r − 1)(n + 2g − 2)/2 [GL, Lemma 3.2]. This is a Γ-invariant closed subvariety of PMΛ Higgs. Note that tensoring by a line bundle does not change the characteristic polynomial, hence the Hitchin map h in (2.5) descends down to (2.7) ^ PMΛ Higgs −→ H . eh : 6 I. BISWAS AND A. DEY So h is the composition of eh with the quotient map PMΛ Higgs −→ We will describe the fibers of the Hitchin maps in (2.5) and (2.7). For any s ∈ U, ^ PMΛ Higgs. (1) P d′ := h−1(s) = Prymd′ (Xs) = Nm−1(OX(d′x)), where is the norm map defined by OXs(Pi dixi) 7−→ OX(Pi diπ(xi)). Nm : Picd(Xs) −→ Picd(X) (2) bP d′ The fiber P d′ := bh−1(s) = Prymd′ is a torsor for (Xs)/Γ (see (2.4) for Γ). P 0 := Nm−1(OX) , is a torsor for bP 0 = Prym0(Xs)/Γ. Hence hU and hU can be thought of as P 0 and bP d′ and bP 0 torsors respectively over U. Let G be an abelian group-scheme over X, and let A be a G–torsor. For any integer n, the G–torsor on X obtained by extending the structure group of A using the endomor- phism of G defined by z 7−→ zn will be denoted by (A)n. Lemma 2.1. For any integer d, we have (1) P d ∼= (P 1)d as P 0 torsors over U, and (2) bP d ∼= (bP 1)d as bP 0 torsors over U. In section 4 we will show that P d and bP d′ "B" field. are mirror partners for a certain choice of a 3. Picard category and Gerbes We briefly recall definition of sheaf of categories over a scheme (for details see [DG], [DM], [Gi]). Let Schet(X) denote the category of all ´etale neighborhoods of a scheme X. A presheaf of categories on Schet(X) is a contravariant functor Q which assigns to every object U −→ X in Schet(X) a category Q(U) and to every morphism f : U1 −→ U2 in Schet(X) a functor f ∗ Q : Q(U2) −→ Q(U1). Moreover, for every composition there is a transformation f ∗ for three-fold compositions. Q ◦ g∗ f U1 −→ U2 Q −→ (g ◦ f )∗ g −→ U3 , Q satisfying an obvious compatibility relation A presheaf Q of categories on Schet(X) is said to be a sheaf of categories if the following two axioms hold: (1) For U −→ X in Schet(X) and a pair of objects C1, C2 ∈ Q(U), the presheaf of sets on Schet(U) that assigns to f : U ′ −→ U the set HomQ(U ′)(f ∗ Q(C1), f ∗ Q(C2)) is a sheaf. (2) If f : U ′ −→ U is a covering, then the category Q(U) is equivalent to the category of descent data on Q(U ′) with respect to f , meaning every descent data on Q(U ′) with respect to f is. SYZ DUALITY FOR PARABOLIC HIGGS MODULI 7 A Picard category is a tensor category, in which every object is invertible. A basic example is the category of line bundles over a scheme. A sheaf of categories P is said to be a sheaf of Picard categories if for every (U → X) ∈ Schet(X) , P(U) is endowed with a structure of a Picard category such that the pull-back functors f ∗ P are compatible with the tensor product in an appropriate sense. If P1 and P2 are two sheaves of Picard categories, one defines (in a straightforward fashion) a tensor functor between them. A category Q is said to be a gerbe over the Picard category P , if P acts on Q as a tensor category, and for any object C ∈ Q the functor P −→ Q given by is an equivalence. B ∈ P =⇒ Action(P, C) ∈ Q Now, if P is a sheaf of Picard categories and Q is another sheaf of categories we say that Q is a gerbe over P, if the following two conditions hold: • For every (U → X) ∈ Schet(X), Q(U) has the structure of a gerbe over P(U). This structure is compatible with the pull-back functors f ∗ P and f ∗ Q. • There exists a covering U −→ X such that Q(U) is non-empty. The basic example of a gerbe over an arbitrary sheaf of Picard categories P is P itself; it is called the trivial P–gerbe. Let A be a sheaf of abelian groups over Schet(X) which takes values in an abelian group A. For an object f : U −→ X of Schet(X), let TorsA(U) denote the category of AU –torsors on U. This is a Picard category, and the assignment U −→ TorsA(U) defines a sheaf of Picard categories on Schet(X) which we will call TorsA or A-torsor. A gerbe over the sheaf of Picard categories A-torsor will be called an A–gerbe. Hence an A–gerbe will be thought of as a torsor over the sheaf of Picard categories A–torsors. An isomorphism between A–gerbes is an equivalence of sheaves of categories as torsors over the sheaf of A–torsors. The isomorphism classes of A–gerbes are in one-to-one correspondence with H 2(X, A) (see [Gi]). A trivialization of an A-gerbe is an isomorphism with the trivial gerbe A–torsor. Two trivializations z , z′ are equivalent if the automorphism z′ ◦ z−1 is given by tensorization with a trivial A-torsor. The space of equivalence classes of trivializations of a trivial A–gerbe B denoted TrivA(X, B) is an H 1(X, A)-torsor over a point [Gi]. Remark 3.1. Fix a short exact sequence 0 −→ A −→ A′′ −→ A′ −→ 0 of sheaves of groups (they need not be abelian) on X, and let τA′ be an A′–torsor over X. We introduce a sheaf of categories Q = QτA′ as follows: For U ∈ Schet(X), let Q(U) be the category of all "liftings" of τA′U to an A′′U –torsor. It is easy to check that Q is a A-gerbe over X. Remark 3.2. Let Q1 be a P1-gerbe over X, and let a : P1 −→ P2 be a tensor functor of Picard category over X. Then one can construct a canonical induced P2–gerbe Q2 over 8 I. BISWAS AND A. DEY X with the property that there exists a functor Q1 −→ Q2, compatible with the actions of P1 and P2 via a. Let Q1 and Q2 be two A–gerbes over X. Then Q1 ×X Q2 is an A ×X A–gerbe over X. Q2 be the A–gerbe Consider the multiplication homomorphism A ×X A −→ A. Let Q1 ⊗ P over X given by Q1 ×X Q2 using this homomorphism (see Remark 3.2). This A–gerbe Q1 ⊗ P Q2 is called the tensor product of Q1 and Q2. Now consider the inversion homomorphism A −→ A. The A–gerbe over X given by Q1 using this homomorphism (see Remark 3.2) will be denoted by (Q1)−1. Notation. Let Q be a A–gerbe over X. For any positive integer n, the n-fold tensor Q will be denoted by (Q)d. For any negative integer n, the n-fold tensor product Q⊗ P Q−1 will be denoted by (Q)−d. product Q−1 ⊗ P · · ·⊗ P · · · ⊗ P Remark 3.3. If B is a trivial A–gerbe over X, then the tensor power Be for any e ∈ Z is also trivial, and, moreover, the set of all trivializations, which is a H 1(X, A)-torsor, has the following identification (it is an identification of H 1(X, A)-torsors). TrivA(X, Be) = TrivA(X, B)e 4. Trivializations and B fields Let U(1) (respectively, Zr) be the sheaf of abelian groups over Schet(PMd Higgs) which takes values in U(1) (respectively, Zr), and let T orsU (1) (respectively, T orsZr ) be the sheaf of Picard categories over PMd Higgs (see Section 3 for definition). There is a universal parabolic Higgs vector bundle over PMd Higgs × X, because the parabolic flags are complete [BY, p. 465, Proposition 3.2]. Let (E, Φ) be a Universal parabolic Higgs bundle on PMd Higgs ×{c}, where c ∈ X \D is a fixed point, we get a vector bundle E on PMd Higgs. Let P := P (E) be the associated Higgs parametrizing line in the fibers of E. From the exact projective bundle on PMd sequence Higgs ×X. Restricting E to PMd e −→ Zr −→ SL(r, C) −→ PGL(r, C) −→ e it follows that the obstruction to lift the PGL(r, C)–bundle P to a SL(r, C)–bundle gives a class B ∈ H 2(PMd Higgs, Zr). This cohomology class B corresponds to the Zr–gerbe on PMd Higgs defined by the liftings of P to a SL(r, C) bundle (see Remark 3.1). Lemma 4.1. The restriction of B to each regular fiber P d of the Hitchin map h (see (2.5)) is trivial as a Zr–gerbe. Proof. Let L be a universal line bundle on P d × Xs (see Section 2 for P d). The projection of the spectral curve Xs to X will be denoted by π. The push-forward (Id × π)∗L is a vector bundle which admits a family of parabolic Higgs field inducing the inclusion Higgs. Hence we have P((Id×π)∗L)P d×{c} = P. Note that det((Id×π)∗L)P d×{c} P d ⊂ PMd is isomorphic to ξ = ⊗y∈π−1(c)LP d×{y} (the points of π−1(c) are taken with multiplicities). SYZ DUALITY FOR PARABOLIC HIGGS MODULI 9 Consider the above line bundle ξ. Let η be the r-th root of the line bundle ξ∗ on P d, meaning ηr = ξ∗. Note that since the N´eron–Severi class of ξ is divisible by r, such a line bundle η exists. It is easy to see that (Id × π)∗(L ⊗ p∗η), where p is the projection of P d × Xs to P d, is a SL(r, C) bundle on P d × X such that P((Id × π)∗(L ⊗ p∗η)) = P. Hence B is a trivial Zr–gerbe when restricted to P d. (cid:3) As seen in the proof of Lemma 4.1, a trivialization of B on P d is equivalent of giving a universal line bundle L −→ P d × Xs such that det(Id × π)∗LP d×{c} is trivial on P d × {c}. Hence we have a natural identification of the set of trivializations of B denoted as TrivZr(P d, B) with the set of isomorphism classes of such line bundles on P d × Xs; define T := {L → P d × Xs L is universal bundle with det(Id × π)∗(L)P d×{c} = OP d} . Note that this T is naturally a bP 0[r]–torsor since det(Id × π)∗(L ⊗ p∗L) = det(Id × π)∗(L) ⊗ Lr . We have the following natural isomorphism H 1(P d, Zr) = H 1(P 0, Zr) = bP 0[r] . In terms of this isomorphism, the H 1(P d, Zr)–torsor TrivZr (P d, B) gets identified with the bP 0[r]–torsor T . Using this identification, the set of trivialization of B on P d will be considered as a H 1(P d, Zr)-torsor. By Remark 3.2, any Zr–gerbe extends to a U(1)–gerbe. Let B denote the U(1)-gerbe given by the Zr–gerbe B. Since the Zr gerbe B on P d is trivial, the extended U(1)– gerbe B on P d is also trivial. The set of all trivializations of B on P d as is denoted by TrivU (1)(P d, B). This TrivU (1)(P d, B) is a H 1(P d, U(1))–torsor. Theorem 4.2. For any d , e ∈ Z, there is an isomorphism of bP 0–torsors TrivU (1)(P d, Be) ∼−→ bP e . Proof. From Lemma 2.1 and Remark 3.3, TrivU (1)(P d, Be) ∼= (TrivU (1)(P d, B1))e and bP e ∼= (bP 1)e . Therefore, it is enough to prove the theorem under the assumption that e = 1. So set e = 1. We have a natural identification as extension of scalers, TrivU (1)(P d, B) = TrivZr (P d, B) × H 1(P d, U(1)) H 1(P d, Zr) . Under this identification, the above torsor TrivU (1)(P d, Be) can be identified set theoreti- cally with T1 defined as follows: {L −→ P d × Xs L is a universal line bundle and LP d×{y} ∈ Pic0(P d) ∀ y ∈ Xs} . We have a natural identification Pic0(P d) ∼= Pic0(f J 0) Pic0(J 0) = f J 0 J 0 [HT2, Lemma 2.2, Lemma 2.3], hence T1 can be identified with T J 0 where T is defined as follows: {L −→ fJ d × Xs L is a universal line bundle and Lf J d×{y} ∈ Pic0fJ d ∀ y ∈ Xs} . 10 I. BISWAS AND A. DEY There is a natural isomorphism of J 0 with Pic0(J 0) given by the natural theta polar- ization on J 0. In terms of this identification, the action of J 0 on T corresponds to the f action of Pic0(J 0) defined by pull-back. Note that bP d can be identified with J d J d . So it is enough to show that T and fJ 1 are isomorphic as fJ 0-torsors. The idea is to give two surjective set theoretic maps f1 and f2 from Xs to these two torsors: such that, (4.1) Xs f1 ~⑦⑦⑦⑦⑦⑦⑦⑦ f2 ❄❄❄❄❄❄❄❄ T fJ 1 f1(y′) − f1(y) = f2(y) − f2(y′) . Now we will construct f1 and f2. The map f1 is the Abel-Jacobi map which takes In view of this equality and the fact that both are fJ 0-torsors, the identification of f1(y) with f2(y) gives the required isomorphism between fJ 1 and T as fJ 0-torsors. any y ∈ Xs to the line bundle OXs(y) (which is an element of fJ 1). The map f2 sends any y to the unique universal line bundle L on P d × Xs satisfying the condition that LP d×{y} = OP d. To show that (4.1) holds, we need the following: (4.2) (L ⊗ p∗(OXs(−y′))P d×{y} = f1(y) − f1(y′) for any y′ ∈ Xs . Now, (4.2) follows from two facts: Firstly, any universal bundle on P d × Xs is of the 2(L0) ⊗ F ∗P, where p2 is the projection to Xs, L0 ∈ fJ d is a fixed line bundle, P is form p∗ the universal line bundle on Pic0(eJ 0) × eJ 0 (= eJ 0 × eJ 0), and F : fJ d × Xs −→ fJ 0 × fJ 0 0 , f1(y) − f1(y′)). is defined by (L, y) 7−→ (L ⊗ L−1 Secondly, the involution of fJ 0 × fJ 0 exchanging the two factors takes the universal line bundle on fJ 0 × fJ 0 = Pic0(eJ 0) × eJ 0 to its dual. (cid:3) Let (4.3) eΓ = G γ∈Γ Lγ − {0} be the disjoint union of the total spaces of the nonzero vectors of the line bundles Lγ. This has the structure of a group scheme over X whose fiber at x ∈ X is an abelian extension (4.4) The group Γ acts on PMd 1 −→ C∗ −→ eΓc −→ Γ −→ 0 . Higgs; the action of any L ∈ Γ sends any (E∗, φ) to (E∗ ⊗L, φ⊗ IdL); since the parabolic structure of L is trivial, we may use the notation of the usual tensor product (note that φ ⊗ IdL is a Higgs field on E∗ ⊗ L in a natural way). Since the quasi-parabolic flags are complete, there exists an universal parabolic Higgs bundle (E, Φ) on PMd Higgs × X and Higgs × X [BY]. In particular, E is a universal vector bundle on PMd ~ SYZ DUALITY FOR PARABOLIC HIGGS MODULI 11 XK(D)), where pX is the projection of PMd Φ ∈ H 0(End(E) ⊗ p∗ Higgs × X to X. Consider the projective bundle P (E) parametrizing the lines in the fibers of E. The group Γ acts on P (E); the action of any L ∈ Γ sends any ((E∗ , φ) , ξ) to ((E∗ ⊗ L , φ ⊗ IdL) , ξ ⊗ Ly), where ξ ∈ (E∗)y. Fix a point c ∈ X. Restricting P (E) to PMd Higgs × {c} we get a Γ- equivariant projective bundle P on PMd Higgs. The obstruction class to lift the Γ-equivariant PGL(r, C)-bundle P into a Γ-equivariant SL(r, C)-bundle gives a nontrivial Γ-equivariant gerbe B ∈ H 2 structure on B. Higgs, Zr). Let eB be the Zr-gerbe Higgs forgetting the Γ-equivariant Γ(PMd ^ PMd The following technical lemma which will be used in proving Lemma 4.4. Lemma 4.3 ([HT2, Lemma 3.3]). Let L −→ J 0 × X be the universal line bundle which is trivial on J 0 × {c}. Then there is an action over X of eΓ (constructed in (4.3)) on the total space of L, lifting the action of Γ on J 0 by translation, so that the scalars C∗ act with weight one on the fibers. Proposition 4.4. The restriction of eB to each regular fiber eP d of the Hitchin map is trivial as a Zr-gerbe. Proof. The statement that the restriction of eB to each regular fiber eP d of the Hitchin map is trivial as a Zr-gerbe is equivalent to the statement that the projective bundle PP d is the projectivization of a Γ-equivariant vector bundle on P d. We have already seen in Lemma 4.1 that B is a trivial gerbe on P d. So PP d is the projectivization of a vector bundle V on P d. Therefore, the only thing to check is that the vector bundle V can be chosen to be Γ equivariant. Recall how we got hold of a vector bundle V on P d; it came from a universal line bundle L on P d × Xs. Hence by Lemma 4.3, giving a Γ-equivariant vector bundle V on P d is equivalent of giving a universal line bundle L over P d × Xs on which eΓ acts such that the scalers C∗ ⊂ eΓc (see (4.4)) act with weight one and det(π∗(eL P d×{c})) is in Pic0 Γ(P d). The rest of the proof will be devoted in showing the existence of such a universal line bundle L on P d × Xs. Let eL be any universal bundle on eJ d × Xs and L be the Poincar´e universal bundle on J 0 × X rigidified using c. We have following natural projection maps, (4.5) P d × J 0 p12 w♣♣♣♣♣♣♣♣♣♣♣♣ p23 'PPPPPPPPPPPPP p13 P d × J 0 × Xs fp12 P d×J 0 Γ × Xs J 0 × Xs P d × Xs Id×π J 0 × X. The action of the group Γ on P d × J 0 is given by γ · (L , M) 7−→ (L ⊗ Lγ , M ⊗ Lγ) . By [HT2, Lemma 2.2], we have an identification P d×J 0 fp12 ∗(eL) has a natural Γ action; this action can be extended to an action of eΓ on fp12 Γ = fJ d. The pulled back line bundle ∗(eL) / / w ' 12 I. BISWAS AND A. DEY by making C∗ act trivially. Define M := fp12 ∗(eL) ⊗ p∗ 12(eL−1) ⊗ p∗ 23(Id × π)∗L . Note that for any x, y and z in P d, J 0 and Xs respectively, we have M{x}×J 0×Xs = MP d×{y}×Xs = MP d×J 0×{z} = O . Hence by the theorem of the cube (see [Mu, p. 87, Theorem]), the line bundle M is trivial, equivalently, there is a eΓ–equivariant isomorphism ∗eL = p∗ 12(eL) ⊗ p∗ fp12 23(Id × π)∗(L−1) . (4.6) Since the universal Poincar´e line bundle L is trivial on J 0 × {c}, by Lemma 4.3, the group scheme eΓ acts on p∗ 23(id × π)∗(L−1) with the scalers C∗ acting by weight −1. From the 12(eL) with the scalers C∗ acting by weight isomorphism in (4.6) we get that eΓ acts on p∗ 12(eL) to any base point of J 0 we get a eΓ-action on eL; this produces a one. Restricting p∗ Γ-action on eL which we wanted. So we can make eL in the proof of Lemma 4.1 to be a Γ-equivariant line bundle. (cid:3) The set of all Γ-equivariant trivializations TrivZr (bP d, eB) can be identified with the set eT defined by {eL −→ P d × Xs eL is a eΓ-equivariant such that det(Id × π)∗(eL)P d×{c} = OP d} . Note that in eT we are interested only those eΓ-actions on eL such that the action of C∗ is of weight one, because by Lemma 4.3 this will ensure that eLP d×{y} is a Γ-equivariant line bundle on P d. It is easy to see that eT is a Pic0 Γ(P d)[r]-torsor. Note that we have the following identifications Pic0 Γ(P d)[r] = Pic0(P d/Γ)[r] = Pic0(P 0/Γ)[r] = P 0[r] , where the last equality follows from the fact that the dual of P 0, namely H 1(P 0, U(1)), is P 0/Γ [HT2, Lemma 2.3]. Let eB denote the extended U(1)-gerbe given by the Zr-gerbe (see Remark 3.2) which is also trivial over bP d. Theorem 4.5. For any d, e ∈ Z, there is a smooth isomorphism of P 0-torsors TrivU (1)(bP d, bBe) ∼= P e . Proof. As in the proof of Theorem 4.2 we can assume e = 1. The set of all trivializa- tion of eB, which is denoted as TrivU (1)(bP d , eB) is a H 1(P d/Γ, U(1)) (= P 0)-torsor. This TrivU (1)(bP d , eB) can be identified with eT defined by {eL −→ P d × Xs eL is a eΓ-equivariant such that det(Id × π)∗(eL)P d×{c} ∈ Pic0 Γ(P d)} . Note that in eT we are interested in only those eΓ action on eL such that the action of C∗ is of weight one so that eL P d×{y} is a Γ-equivariant line bundle on P d. Note that the difference between eT and eT is that in eT we require the determinant of eL is trivial, while in eT we require the determinant to be a Γ-equivariant line bundle on P d of degree zero. In the rest of the proof we will show that the P 0-torsor eT is isomorphic to the P 0-torsor P 1. SYZ DUALITY FOR PARABOLIC HIGGS MODULI 13 First note that P 1 sits naturally inside fJ 1 as a P 0-torsor. In the proof of Theorem 4.2 we have seen that the set of all universal line bundles eL −→ fJ d × Xs with eLf J d×{c} ∈ Pic0(fJ d) is isomorphic to fJ 1 as a fJ 0-torsor. We will use this fact to give an inclusion of eT into fJ 1 compatible with respect to P 0 ⊂ fJ 0. Send any eL ∈ eT to 13(eL) ⊗ p∗ 23(id × π)∗(L)−1 by weights +1 p∗ and -1 respectively, the above tensor product is a eΓ-equivariant line bundle with scalers acting trivially, or in other words, it is a Γ-equivariant line bundle on P d ×J 0 ×Xs. Hence 13(eL) ⊗ p∗ p∗ 23(Id × π)∗(L)−1. Since C∗ acts on p∗ 23(Id × π)∗(L)−1 descends down to 13(eL) and p∗ P d × J 0 Γ × Xs = fJ d × Xs . One can check easily that p∗ map 13(eL) ⊗ p∗ 23(Id × π)∗(L)−1f J d×{c} ∈ Pic0(fJ d), and the resulting 23(Id × π)∗(L)−1 is injective. eT −→ fJ 1 , eL 7−→ p∗ 13(eL) ⊗ p∗ So eT and P 1 are now both P 0-subtorsors of fJ 1. The quotient by either is the constant torsor J 0 (the Jacobian of C). Therefore, the image of one in the quotient by the other gives a morphism from the base U to J 0. But U is a Zariski open set in an affine space, so its only morphisms to an abelian variety are the constant ones. Indeed, any nonconstant morphism from a Zariski open subset of A1 to an abelian variety extends to a nonconstant morphism from P1 (cid:3) C to the abelian variety. But there is no such map. From Theorem 4.2 and Theorem 4.5 we conclude that over a generic open subset U ⊂ H the Hitchin fibers are SYZ mirror partners in the sense of Hitchin. References [BLS] [BM] [BY] [DM] [DG] [DP] [FW] [Gi] [GL] [GW] [HT1] A. Beauville, Y. Laszlo and C. Sorger: The Picard group of the moduli of G-bundles on a curve, Compos. Math. 112 (1998), 183–216. I. Biswas and A. Mukherjee: Symplectic structures on moduli spaces of parabolic Higgs bundles and Hilbert scheme, Comm. Math. Phys. 240 (2003), 149–159. H. U. Boden and K. Yokogawa: Rationality of moduli spaces of parabolic bundles, Jour. London Math. Soc. 59 (1999), 461–478. P. Deligne and J. S. Milne: Tannakian categories, in: Hodge Cycles, Motives, and Shimura varieties, 101–229, Lect. Notes Math. 900, Springer-Verlag, Berlin-New York, 1982. R. Y. Donagi and D. Gaitsgory: The gerbe of Higgs bundles, Transform. Groups 7 (2002), 109–153. R. Y. Donagi and T. Pantev: Langlands duality for Hitchin systems, Inv. Math. (2012), DOI: 10.1007/s00222-012-0373-8. E. Frenkel and E. Witten: Geometric endoscopy and mirror symmetry, Commun. Number The- ory Phys. 2 (2008), 113–283. J. Giraud: Cohomologie non ab´elienne, Springer-Verlag, Berlin-New York, 1971. T. L. G´omez and M. Logares: A Torelli theorem for the moduli space of parabolic Higgs bundles, Adv. Geom. 11 (2011), 429–444. S. Gukov and E. Witten: Gauge theory, ramification, and the geometric Langlands program, in: Current developments in mathematics, 2006, 35–180, Int. Press, Somerville, MA, 2008. T. Hausel and M. Thaddeus: Examples of mirror partners arising from integrable systems, C. R. Acad. Sci. Paris Math. 333 (2001), 313–318. 14 [HT2] [Hi1] [Hi2] [Hi3] [Hi4] [LM] [MS] [Mu] [Se] [SYZ] [Wi] [Yo1] [Yo2] I. BISWAS AND A. DEY T. Hausel and M. Thaddeus: Mirror symmetry, Langlands duality, and the Hitchin system, Invent. Math. 153 (2003), 197–229. N. J. Hitchin: Lectures on special Lagrangian submanifolds, in: Winter School on Mirror Symmetry, Vector Bundles and Lagrangian Submanifolds, Cambridge, MA (1999), 151–182, AMS/IP Stud. Adv. Math., 23, Amer. Math. Soc., Providence, RI, 2001. N. J. Hitchin: The self-duality equations on a Riemann surface, Proc. London Math. Soc. 55 (1987), 59–126. N. J. Hitchin: Stable bundles and integrable systems, Duke Math. Jour. 54 (1987), 9–114. N. J. Hitchin: Langlands duality and G2 spectral curves, Quart. Jour. Math. 58 (2007), 319–344. M. Logares and J. Martens: Moduli of parabolic Higgs bundles and Atiyah algebroids, Jour. Reine Angew. Math. 649 (2010), 89–116. V. Mehta and C. Seshadri: Moduli of vector bundles on curves with parabolic structures, Math. Ann. 248 (1980), 205–239. D. Mumford: Abelian varieties, With appendices by C. P. Ramanujam and Yuri Manin, Cor- rected reprint of the second (1974) edition, Tata Institute of Fundamental Research Studies in Mathematics, 5, Hindustan Book Agency, New Delhi, 2008. C. S. Seshadri: Quotient space by an abelian variety, Math. Ann. 152 (1963), 185–194. A. Strominger, S.-T. Yau and E. Zaslow: Mirror symmetry is T-duality, Nuclear Phy. B 479 (1996), 243–259. E. Witten: Mirror symmetry, Hitchin's equations, and Langlands duality, in: The many facets of geometry, 113–128, Oxford Univ. Press, Oxford, 2010. K. Yokogawa: Infinitesimal deformation of parabolic Higgs sheaves, Inter. Jour. Math. 6 (1995), 125–148. K. Yokogawa: Compactification of moduli of parabolic sheaves and moduli of parabolic Higgs sheaves, Jour. Math. Kyoto Univ. 33 (1993), 451–504. School of Mathematics, Tata Institute of Fundamental Research, Homi Bhabha Road, Bombay 400005, India E-mail address: [email protected] Department of Mathematics, Indian Institute of Technology, Madras, Chennai 600036, India E-mail address: [email protected]
1808.04783	3	1808	2019-01-05T00:49:16	Tate Conjecture And Finiteness of Abelian Varieties over Finite Field	[ "math.AG" ]	In this paper we will prove that Tate conjecture of abelian varieties over finite field is equivalent to the finiteness of isomorphism classes of abelian varieties with a fixed dimension. We give a different approach with Zarhin's result.	math.AG	math	TATE CONJECTURE AND FINITENESS OF ABELIAN VARIETIES OVER FINITE FIELD ANNINGZHE GAO Abstract. In this paper we will prove that Tate conjecture of abelian varieties over finite field is equivalent to the finiteness of isomorphism classes of abelian varieties with a fixed dimension. We give a different approach with Zarhin's result [6]. Contents Introduction 1. 2. Some Basic Facts About Abelian Varieties 3. The Finiteness of Isogenous Classes 4. Some Calculus of the Tate Module 5. Finish the Proof References 1 2 3 4 8 10 Through the paper, k will denote a finite field of char k = p and ¯k means the algebraic closure of k. In [3], Tate proved the following famous theorem 1. Introduction Theorem 1.1 (Tate). Let k = Fq be a finite field where q is a power of a prime p. Let A, B be two abelian varieties over k. Let G = Gal(¯k/k) be the absolute Galois group of k. If l is a prime and l 6= p, then we have the isomorphism HomAV (A, B) ⊗ Zl ∼= HomZl[G](Tl(A), Tl(B)) Here Tl is the Tate module of an abelian variety. In the proof of the theorem, Tate used the following fact: Fix an integer d, a prime l and an abelian variety A over k, the isomorphism classes of abelian varieties which admits a polarization of degree d2 and an isogeny B → A of degree ln for some n is finite. If we denote g = dimA, then since there is an isogeny between B and A, so dimB = g. It is well known that over k (a finite field), the isomorphism classes of abelian variety of dimension g is finite, so we have F initeness of isomorphism classes of abelian varieties over k with a f ixed dimension 1 2 ANNINGZHE GAO =⇒ T ate conjecture of abelian varieties over k In this paper we will prove the converse direction, i.e. Theorem 1.2 (Main Theorem). Tate conjecture of abelian varieties over k implies that thaere are only finitely many abelian varieties of dimension g over k. This result is first proved by Zarhin [6]. We will give a different approach to this result. The organization of this paper is as follows: In Section 2 we will recall some basic properties of abelian varieties over k, in Section 3 we will use Tate isogeny theorem to prove there are only finitely many isogeny classes of abelian varieties, in Section 4 and Section 5 we will show there are finitely many abelian varieties isogenous to a fixed abelian variety. Notations: We will use k to represent a finite field of characteristic p, ¯k its algebraic closure, G = Gal(¯k/k) the absolute Galois group. σ is the Frobenius element. For a projective variety X over k, we use πX to denote the Frobenius morphism of X. I would like to thank my advisor, Martin Olsson, introduced me this interesting topic and lots of useful discussions. And Daniel Bragg and Thomas Preu for helpful discussion on some details. 2. Some Basic Facts About Abelian Varieties group. In this section we recall the Tate module of an abelian variety and the p−divisible Let A be an abelian variety over k with dimA = g. Choose l a prime number with l 6= p. We know that if (p, n) = 1, the morphism n : A → A is a separable isogeny of degree n2g, denote A[n] = Ker(n : A → A), then A[n](¯k) ∼= (Z/nZ)2g. By definition the Tate module Tl(A) = lim ←−n A[ln](¯k) We know Tl(A) ∼= Z2g l non-canonically. The Galois group G acts on Tl(A) in a natural way. This action is continuous, since the Frobenius is an topological generator of G so the action of σ determines the action of G. The Frobenius morphism πA : A → A is a morphism in HomAV (A, A), by definition the image of πA under the isomorphism HomAV (A, A) ⊗ Zl ∼= HomZl[G](Tl(A), Tl(A)) is σ. For πA, we define a function PπA(n) = deg(n − πA), then we know PπA is a polynomial of degree 2g with Z coefficients. It is the same as the characteristic polynomial of σ on Vl(A) = Tl(A)⊗ Ql. In particular the characteristic polynomial of σ on Vl(A) is independent of l. TATE CONJECTURE AND FINITENESS OF ABELIAN VARIETIES OVER FINITE FIELD 3 In the case if l = p, since now p : A → A is not separate, things are a little different. We use the p−divisible group in this case. We define A[p∞] = lim −→n A[pn] and Dieudonne modules, for details, see [5] and [2]. To introduce the Tate p−conjecture, we need to use definition of Dieudonne ring Let Dk be the Dieudonne ring of k, it is a non-commutative associative W (k)−algebra (W (k) is the ring of Witt vectors) with two generators F, V satisfying the following conditions: F V = V F = p F (c) = φ(c)F cV = V φ(c) for any c ∈ W (k). Here φ is the automorphism of W (k) induced by the automor- phism x → xp on k. So if k = Fp, then Dk is commutative. By the standard procedure (see [2]) we can associate A[p∞] with a Dk module M(A). It is a free W (k) module of rank 2g. Its Dk action is uniquely determined by the action of F (or V ), then Tate proved Theorem 2.1 (Tate). For two abelian varieties A, B over k, under the above notation, we have a natural isomorphism HomAV (A, B) ⊗ Zp ∼= HomDk(M(B), M(A)) This theorem can be found in [5]. In particular, we have HomAV (A, A) ∼= HomDk(M(A), M(A)) In this case, if we denote σA is image of πA (The Frobenius of A) under this isomorphism, then σA = F m if k = Fpm. And the character polynomial of σA is just PπA. 3. The Finiteness of Isogenous Classes In this section we will prove that there are finitely many isogeny classes of abelian varieties over k of dimension g. We first recall the isogeny thoerem. Theorem 3.1 (Tate). Given two abelian varieties A, B over k, then A and B are isogenous ⇐⇒ PπA = PπB ⇐⇒ Tl(A) ⊗ Ql ∼= Tl(B) ⊗ Ql as Ql[G] modules So to consider the isogenous classes we just need to consider the characteristic polynomials of the Frobenius. But we have the following big theorem. 4 ANNINGZHE GAO Theorem 3.2 (Weil Conjecture). Let A be an abelian variety over k with dimen- sion g, then PπA is a monic polynomial with coefficients in Z with degree 2g, and if α is a root of PπA, then for any Galois embedding η : ¯Q → ¯Q over Q, we have η(α) = √q, here q is the number of elements in k. With these two theorems, we can state Corollary 3.1. There are only finitely many isogenous classes of abelian varieties over k of dimension g. Proof. By theorem 3.1, it suffices to prove there are only finitely many character- istic polynomials. Suppose k = Fq. If i=0aix2g−i of P (x), then by theorem 3.2, αi ≤ √q, so we have P (x) = Σ2g is the characteristic polynomial of some abelian variety, and α1, α2, ..., α2g are roots as = Σ1≤i1<i2<...<is≤2gαi1...αis ≤ Σ1≤i1<i2<...<is≤2gαi1...αis ≤ Σ1≤i1<i2<...<is≤2g√qs ≤ M√qs for some M. So we know all ai are bounded by some number which only depends on the field k. But we know all ai are integers, so we only have finitely many choices, so there are only finitely many polynomials can be the characteristic polynomial of some abelian variety. So there are only finitely many isogenous classes. (cid:3) So to prove there are finitely many isomorphism classes it suffices to show every isogenous class of abelian varieties only contains finitely many isomorphism classes. 4. Some Calculus of the Tate Module In this section we fix an abelian variety A over k with dimension g. πA will denote the Frobenius morphism of A, PπA is its characteristic polynomial. Let C be the isogenous class containing A. We will also use πA to mean the element in HomZl[G](Tl(A), Tl(A)) under the Tate's isomorphism, which can be regarded as a 2g × 2g matrix with element in Zl. The main proposition of this section is: Proporsition 4.1. With the above data, there exists a positive integer N which only depends on A (we will see from the proof N only depends on C), such that for any B ∈ C and l > N, Tl(B) ∼= Tl(A) as Zl[G] modules, and for l < N, the set {Tl(B)B ∈ C} (consider as Zl[G] modules) is a finite set (we include the case l = chark, in which case we consider Dieudonne modules as in section 2). TATE CONJECTURE AND FINITENESS OF ABELIAN VARIETIES OVER FINITE FIELD 5 Before the proof, we first notice that we must have Tl(A) ⊗ Ql ∼= Tl(B) ⊗ Ql as Ql[G] module. As we discussed, the G action action on the Tate module is uniquely determined by the action of the Frobenius. So we can see Tl(A) ∼= Tl(B) as Zl[G] modules if and only if πA and πB are conjugate by some matrix in GL2g(Zl) (not GL2g(Ql), they already conjugate by some matrix in GL2g(Ql) by Tate's isogeny theorem). We separated the proof into two parts, consists of the following two lemmas. They are all purely linear algebra things. Lemma 4.1. There exists a posotive N such that for any abelian variety B which is isogenous to A, and l > N, we can find basis of Tl(B) and Tl(A) such that the matrices of πA and πB will be the same. Proof. We know A is isogenous to Ab1 s where all Ai are simple, non-isogenous with each other. We know for a simple abelian variety Ai, the char- acteristic polynomial PπAi of the Frobenius is a power of an irreducible polynomial, is also a power of an irreducible polynomial. The characteristic polyno- so Pπ 2 × ... × Abs 1 × Ab2 bi A i mials Pπ theorem, there exists gi ∈ Q[x] such that bi A i should coprime with each other, let Ki = Πj6=iPπ then by Bezout bj j A Σs i=1gi(x)Ki(x) = 1 Then choose N0 be a positive number such that if l > N0, then all gi(x) ∈ Zl[x] (i.e. l doesn't divide any denominators in gi). And denote Mi = Ki(π)Tl(A), then since all gi ∈ Zl[x], so if l > N0, Tl(A) = ⊕Mi. And this N0 only depends on the chosen isogenous class C, and on each Mi, the characteristic polynomial of πA is Pπ , which is a power of an irreducible polynomial. Then we will concentrate on bi A i one M1, i.e. we just assume M1 = Tl(A), and we can see the similar procedure can be applied to all 2 ≤ i ≤ s and prove the lemma in the general case. We know πA is an invertible matrix with coefficients in Zl. Let {α1, ..., αt} be the roots of PπA, then we have PπA = ((x − α1)...(x − αt))e for some e and Q(x) = Πt i=1(x − αi) ∈ Z[x] is irreducible. Then we define Pi(x) = Πj6=i(x − αj) for 1 ≤ i ≤ k. Since they don't have common factors, so we may choose hi(x) ∈ ¯Q[x] such that ΣhiPi = 1 Choose N1 such that if l > N1, then we have hi ∈ ¯Zl[x]. Then we can see for any v ∈ Tl(A), v = Σhi(πA)Pi(πA)v. Also it is easy to check Li = Pi(πA)Tl(A) lies in αi eigenspace (consider this over ¯Zl). Let ¯Li be the ¯Zl linear expansion of Li in Tl(A) ⊗ ¯Zl. Since every eigenspace of different eigenvalues are linearly 6 ANNINGZHE GAO independent, so we have Tl(A) ⊗ ¯Zl = ⊕ ¯Li Define D = Πi6=j(αi − αj)2, then D ∈ Z. We pick u1, ..., ue to be an integral basis of ¯L1 over ¯Zl, such that u1, ..., ue can be represented by ui = P1(πA)(wi) for w1, ..., we in Tl(A) (This is true by linear algebra and the definition of ¯L1). Define vi = ΣPj(πA)(wi). Then we have vi ∈ Tl(A). We prove if l > max(N1,D), then {v1, πA(v1), ..., πt−1 A (v1), v2, ..., πt−1 A (v2), ..., ve, ..., πt−1 A (ve)} is an integral basis of Tl(A). Since v = Σhi(πA)Pi(πA)v, so it suffices to show Pi(πA)(v) can be represented over ¯Zl by these elements. From the Galois theory Pi(x) = φ(P1(x)) for some φ ∈ Gal( ¯Ql/Ql). By definition of wi, P1(πA)(v) = ΣβjP1(πA)(wj) for some βj ∈ ¯Zl, so we have Pi(πA)(v) = Σφ(βj)Pi(πA)(wj) Then we just need to show Pi(πA)(wj) can be represented over ¯Zl by these elements. This is solved by considering the system linear equations: v1 = ΣPi(πA)(w1) πA(v1) = ΣαiPi(πA)(w1) .... πt−1 A (v1) = Σαt−1 i Pi(πA)(w1) then the matrix of this system of linear equations has determinant D, so by defini- tion of l and the Crammer's rule, Pi(πA)(wj) can be represented over ¯Zl by these elements. So {v1, πA(v1), ..., πt−1 A (v1), v2, ..., πt−1 A (v2), ..., ve, ..., πt−1 A (ve)} is an integral basis of Tl(A) (Here we only proved every element can be repre- sented integrally by these elements. But we have te elements here and te = 2g = dimTl(A)⊗ Ql, so they must form a basis). And the matrix of πA under this basis is uniquely determined by PπA, just denote this matrix by C (independent of A). So if we set N = max(N1,D), then for l > N1, we can choose a basis as above such that the Frobenius acts on Tl(A) is represented by the matrix C. But this is independent of A, so we can do the same thing for B, so the matrices of πA and πB are the same. For the general case, we can find Ni for each Mi, they are all only depend on PπA, so just choose N = max(N0, N1, ..., Ns), then from the above procedure, we can choose basis such that πA and πB have the same matrix when l > N. We proved the lemma. (cid:3) TATE CONJECTURE AND FINITENESS OF ABELIAN VARIETIES OVER FINITE FIELD 7 Lemma 4.2. With N defined as above, for l < N, the set {Tl(B)B ∈ C} (consider as Zl[G] modules) is a finite set (we include the case l = chark, in which case we consider Dieudonne modules as in section 2). i=1Pπ Proof. We use the same idea as in Lemma 4.1. We collect them here. Let PπA = Πs Ki = Πj6=iPπ have gi(x) ∈ Q[x] such that is a power of an irreducible polynomial. Set for 1 ≤ i ≤ s, then these Ki(x) don't have common factors. So we where Pπ bi A i bj j A bi A i Σs i=1gi(x)Ki(x) = 1 Fix some l < N. Define Mi = Ki(πA)Tl(A). If we set s1 to be the smallest integer such that ls1gi(x) ∈ Zl[x], then we have ls1Tl(A) ⊆ ⊕Mi ⊆ Tl(A) Be careful this s1 only depends on PπA and l. Then πA acts on Mi, and its characteristic polynomial is just Pπ . Write Pπ = (Πri j=1(x − αij))ei. Define bi A i bi A i for 1 ≤ i ≤ s and 1 ≤ j ≤ ri. Then by Bezout's theorem, we may find hij ∈ ¯Q[x] such that Pij = Πn6=j(x − αin) Define Σri j=1hijPij = 1 Lij = Pij(πA)Mi and ¯Lij be the ¯Zl expansion of Lij in Tl(A)⊗ ¯Zl. Then ¯Lij is a free module of rank ei. Choose {wi1, ..., wiei} such that Pi1(πA)(wij) 1 ≤ j ≤ ei is an integral basis of Li1. Define Define Ni to be the submodule of M1 generated by vij = Σri n=1Pin(πA)(wij), 1 ≤ i ≤ s, 1 ≤ j ≤ e1 {vi1, πA(vi1), ..., πri−1 A (vi1), ....., viei, ..., πri−1 A (viei)} Let D = (Πt number such that i=1Π1≤j,k≤ri(αij − αik))2(e1+e2+...+es), and choose s2 to be the smallest ls D ∈ Z Then similar to the proof in Lemma 4.1 we can see ls2(⊕Mi) ⊆ ⊕Ni ⊆ ⊕Mi ls2hij ∈ Zl[x], Then we have ⊕Ni ⊆ Tl(A) ⊆ l−s1−s2 ⊕ Ni Note that the matrix of πA on ⊕Ni is only determined by PπA in the chosen basis. Also the action of πA on Tl(A) is induced from l−s1−s2 ⊕ Ni. But l−s1−s2 ⊕ Ni/⊕ Ni is a finite set, so we proved the finiteness of {Tl(B)B ∈ C} if l 6= p. 8 ANNINGZHE GAO The l = p case is similar as we can see we can do the similar calculus for W (k) module M(A) with the action πA = Lm if k = Fpm. Then we can see that the set of M(A) with the action of πA is finite, but for fixed πA, there are only finitely many choices of L since they must be semi-simple. So we have the set of M(A) with Dk action is a finite set. (cid:3) By Proposition 4.1, to prove the finiteness of isomorphism classes of abelian varieties, it suffices to show for a fixed abelian variety A of dimension g, the set {B an abelian variety, Tl(B) ∼= Tl(A) as Zl[G] modules f or all prime l} is a finite set. Here we include the case l = p, which we consider the Dk module M(A). We will show this in the next section. 5. Finish the Proof In this section, we will show the set in the previous section {B an abelian variety, Tl(B) ∼= Tl(A) as Zl[G] modules f or all prime l} is finite with the fixed A. Lemma 5.1. For a prime l 6= p, if Tl(A) ∼= Tl(B) as a Zl[G] module, then there exists an isogeny π : B → A with (degπ, l) = 1. Proof. From Tate conjecture, we have the following isomorphism HomAV (B, A) ∼= HomZl[G](Tl(B), Tl(A)) If σ : Tl(B) → Tl(A) the isomorphism, then since Z is dense in Zl, so we may find an isogeny π : B → A such that the image of π is close to σ. So the image of π is also an isomorphism. Then set N = Kerπ, so we have the exact sequence 0 → Tl(B) → Tl(A) → Nl → 0 here Nl means the sylow l group of N, see [4] Prop. 10.6. Since π induces isomor- phism between Tate modules, so we must have Nl = {0}, so (degπ, l) = 1. (cid:3) Lemma 5.2. The same holds for l = p case. Proof. The proof is really similar, the exact sequence is 0 → Np → B[p∞] → A[p∞] → 0 see [4] 10.17 We can conclude the previous lemma into one property: (cid:3) Proporsition 5.1. Fix an abelian variety A. If there exists an abelian variety B such that Tl(B) ∼= Tl(A) for l 6= p and M(B) ∼= M(A) as Dk modules, then for any prime l (maybe l = p), we have an isogeny πl : B → A such that deg(π) is coprime to l. We need a technique lemma. TATE CONJECTURE AND FINITENESS OF ABELIAN VARIETIES OVER FINITE FIELD 9 Lemma 5.3. If we have two abelian varieties A and B and two isogenies π1 : B → A and π2 : B → A. If we have two integers m1, m2 such that (m1, m2) = 1 and (m1, degπ1) = (m2, degπ2) = 1, then we have an isogeny π : B → A such that (degπ, m1m2) = 1. Proof. Set π = m2π1 + m1π2. First we show π is an isogeny. So pick some l m1, then consider the image of π in HomZl[G](Tl(B), Tl(A)) under the Tate isomorphism. We can see by condition m2π1 under this isomorphism induces an isomorphism since (m2degπ1, l) = 1, and π and m2π1 are differ by l times some homomorphism, so we have π is an isomorphisms of Tate modules, so π is an isogeny. If (degπ, m1m2) 6= 1, then there exists some x ∈ Kerπ ∩ B[m1m2]. Then by replacing x by some multiple, we may assume there exists some prime factor l of m1m2, just say a prime factor of m1 (the case of m2 is the same), such that x 6= 0, lx = 0 and x ∈ Kerπ. But then we have 0 = π(x) = m2π1(x) + m1π2(x) = m2π1(x) so x ∈ Ker(m2π1), so x ∈ Ker(m2π1) ∩ B[l] = {0}, which is a contradiction. So this π satisfies our requirements. (cid:3) Now we comes to the last lemma. Lemma 5.4. Fix an abelian variety A. If there exists an abelian variety B such that Tl(B) ∼= Tl(A) for l 6= p and M(B) ∼= M(A) as Dk modules, then B is a direct component of A × A. Here direct component means we have an abelian variety C such that B × C ∼= A × A Proof. Choose any isogeny π1 : B → A, then by Lemma 5.3, Prop 5.1 and the induction procedure, we may find an isogeny π2 such that (degπ1, degπ2) = 1, then we have an embedding Then we may find isogenies φ1 : A → B and φ2 : A → B such that g : B → A × A b → (π1(b), π2(b)) φ1π1 = degπ1 since (degπ1, degπ2) = 1, so there exists s, t ∈ Z such that sdegπ1 + tdegπ2 = 1, define φ2π2 = degπ2 f : A × A → B (a1, a2) → sφ1(a1) + tφ2(a2) Then we have f g = 1B, so B is a direct factor of A × A. Now we can finish our proof. (cid:3) Theorem 5.1. There are only finitely abelian varieties of dimension g over k. 10 ANNINGZHE GAO Proof. By [1] Theorem 18.7, for a fixed abelian variety, there are only finitely many direct components, then the theorem follows from Corollary 3.1, Prop 4.1 and Lemma 5.4. (cid:3) References [1] J. S. Milne. Abelian varieties. In Arithmetic geometry, pages 103 -- 150. Springer, 1986. http://www. math. [2] R. Pink. Finite schemes. Notes group de Cours ethz. ch/pink/FiniteGroupSchemes. html, 2004. [3] J. Tate. Endomorphisms of abelian varieties over finite fields. Inventiones mathematicae, 2(2):134 -- 144, 1966. [4] G. van der Geer and B. Moonen. Abelian varieties. [5] W. C. Waterhouse and J. Milne. Abelian varieties over finite fields. PhD thesis, Harvard University, 1968. [6] Y. Zarhin. Endomorphisms of abelian varieties and points of finite order in characteristic p. Mathematical notes of the Academy of Science of the USSR, 21:415 -- 419, 1977.
1401.6678	3	1401	2015-05-16T10:41:08	Relations between derived Hochschild functors via twisting	[ "math.AG", "math.AC" ]	Let $k$ be a regular ring, and let $A,B$ be essentially finite type $k$-algebras. For any functor $F:{D}(A)\times\dots\times{D}(A)\to{D}(B)$ between their derived categories, we define its twist $F^{!}:{D}(A)\times\dots\times{D}(A)\to{D}(B)$ with respect to dualizing complexes, generalizing Grothendieck's construction of $f^{!}$. We show that relations between functors are preserved between their twists, and deduce that various relations hold between derived Hochschild (co)-homology and the $f^{!}$ functor. We also deduce that the set of isomorphism classes of dualizing complexes over a ring (or a scheme) form a group with respect to derived Hochschild cohomology, and that the twisted inverse image functor is a group homomorphism.	math.AG	math	RELATIONS BETWEEN DERIVED HOCHSCHILD FUNCTORS VIA TWISTING LIRAN SHAUL ABSTRACT. Let k be a regular ring, and let A, B be essentially finite type k-algebras. For any functor F : D(Mod A) × · · · × D(Mod A) → D(Mod B) between their derived categories, we define its twist F ! : D(Mod A) × · · · × D(Mod A) → D(Mod B) with respect to dualizing complexes, generalizing Grothendieck's construction of f !. We show that relations between functors are preserved between their twists, and deduce that various relations hold between derived Hochschild (co)-homology and the f ! functor. We also deduce that the set of isomorphism classes of dualizing complexes over a ring (or a scheme) form a group with respect to derived Hochschild cohomology, and that the twisted inverse image functor is a group homomorphism. 1. INTRODUCTION : D+ f (Mod A) → D+ All rings in this note are commutative. Let k be a regular noetherian ring of finite Krull dimension. Let f : A → B be a map between two essentially finite type k- algebras. Grothendieck duality theory, whose details first appeared in [RD], centers around the twisted inverse image functor f ! f (Mod B) (See [Li] for a modern account). Under the above assumption on k, this functor may be constructed as a twist of the inverse image functor Lf ∗(−) := B ⊗L A −. The twist is given by f !(−) := DB(Lf ∗(DA(−))) where for an essentially finite type k-algebra g : k → C, we have set DC(−) := R HomC (−, RC ), where RC = g!(k) is the canonical (or rigid) dualizing complex over C relative to k. Similarly to this construction, given any func- tor F from Df (Mod A) to Df (Mod B), one may construct the twist of F by declaring F !(−) := DB(F (DA(−)). Under suitable finiteness assumptions, if F, G and H are three functors of this form, and if F ∼= G ◦ H, then it is easy to see that F ! ∼= G! ◦ H !. This means that relations between such functors give rise to relations between their twists. If the ring A is projective over k, then the Hochschild cohomology functor of A over k is defined by Ext∗ A⊗kA(A, −). When dropping the projectivity assumption, an impor- tant variation of this construction is given by Shukla cohomology, also known as derived Hochschild cohomology. This functor, recently studied in great detail in [AILN], is defined by the formula R HomA⊗L k A(A, −) k A is taken in the category of DG-algebras. Taking where the derived tensor product A ⊗L the coefficients complex to be of the form M ⊗L k N where M, N ∈ D(Mod A), it was shown in [AILN, Theorem 4.1], under suitable technical assumptions, that this functor has a particularly nice reduction formula: There is a bifunctorial isomorphism (1) R HomA⊗L k A(A, M ⊗L k N ) ∼= R HomA(R HomA(M, RA), N ). Key words and phrases. derived Hochschild cohomology, dualizing complex, symmetric monoidal structure. Mathematics Subject Classification 2010. 13D03, 18E30, 18D10. 1 2 LIRAN SHAUL A −. This suggests the notation − ⊗! We will see below that the right hand side of this formula is canonically isomorphic to A − for this functor. the twist of the bifunctor − ⊗L A similar result ([AILN, Theorem 4.6]) was given for derived Hochschild homology, and, similarly, we interpret this result by showing that it is canonically isomorphic to the twist of the bifunctor R HomA(−, −), which suggests the notation Hom! A(−, −) for this functor. Thus, having identified the twists of the inverse image functor, the derived tensor func- tor and the derived hom functor, we will immediately deduce various relations that hold between the twisted inverse image, the derived Hochschild homology and the derived Hochschild cohomology functors. 2. TWISTED FUNCTORS Setup 2. Throughout this section, k will be a fixed regular noetherian ring of finite Krull dimension. Let A be an essentially finite type k-algebra. This means that the structure map k → A may be factored as k ֒→ k[x1, . . . , xn] ֒→ U −1k[x1, . . . , xn] ։ A where U ⊆ k[x1, . . . , xn] is some multiplicatively closed set. algebras, the canonical dualizing complex of A is given by In the category of k- RA := R HomU −1 k[x1,...,xn](A, Ωn U −1 k[x1,...,xn]/k[n]), where Ωn is the module of Kähler n-differentials. This construction is independent of the chosen factorization (see for example [AILN, Theorem 1.1]). An intrinsic characterization of the complex RA is given by the important property of f (Mod A), rigidity: recall that according to [YZ1, Definition 2.1], a pair (R, ρ) where R ∈ Db and ρ : R → R HomA⊗L k A(A, R ⊗L k R) is an isomorphism, is called a rigid complex over A relative to k. If moreover the complex R is a dualizing complex then (R, ρ) is called a rigid dualizing complex over A relative to k. This notion originated in [VdB]. It is shown in [YZ1, Theorem 3.6] that a rigid dualizing complex exists, and is unique in a strong sense (see also [AIL1, Theorem 8.5.6] for a stronger existence result). It is given by the complex RA defined above. In the remaining of this note, for any essentially finite type k-algebra A, we will denote by RA the canonical (=rigid) dualizing complex of A, and by DA the functor DA(M ) := R HomA(M, RA). In a somewhat unorthodox manner, we will also set for n > 1, DA(M1, . . . , Mn) := (DA(M1), . . . , DA(Mn)). For a category A, we will denote by An the product category A × · · · × A . } {z n Generalizing Grothendieck's construction of the twisted inverse image functor leads us to the following definition: Definition 3. Let A, B be two essentially finite type k-algebras, and let F : D(Mod A)n → D(Mod B)m be a functor. The twist of F is the functor F !(−) := DB ◦ F ◦ DA(−) : D(Mod A)n → D(Mod B)m. We now give several examples to demonstrate the usefulness of this definition. RELATIONS BETWEEN DERIVED HOCHSCHILD FUNCTORS VIA TWISTING 3 Example 4. Let A, B be two essentially finite type k-algebras, and let f : A → B be a k-algebra map. Consider the functor Lf ∗ : D+ f (Mod B) given by f (Mod A), there is Lf ∗(−) := − ⊗L an isomorphism of functors f (Mod A) → D+ A B. Then by [YZ1, Theorem 4.10], for any M ∈ D+ (Lf ∗)!(M ) ∼= f !(M ). Example 5. Let A be an essentially finite type k-algebra, and let F (M, N ) := M ⊗L and G(M, N ) := R HomA(M, N ). Then by definition F !(M, N ) := R HomA(R HomA(M, RA) ⊗L A R HomA(N, RA), RA), A N and G!(M, N ) := R HomA(R HomA(R HomA(M, RA), R HomA(N, RA)), RA). A N := F !(M, N ), and Hom! We set M ⊗! D+ D− almost immediately from the results of [AILN], we will identify these two functors. A(M, N ) := G!(M, N ). Note that if M, N ∈ f (Mod A) and N ∈ f (Mod A). In Theorem 7 below, which follows f (Mod A), then M ⊗! f (Mod A) then Hom! f (Mod A). Similarly, if M ∈ D+ A N ∈ D+ A(M, N ) ∈ D− Example 6. Let A be an essentially finite type k-algebra, and let a ⊆ A be an ideal. The a-torsion and a-completion functors are defined by Γa(−) := lim−→ HomA(A/an, −) and Λa(−) := lim←− A/an ⊗A − respectively. Their derived functors RΓa and LΛa exist, and are calculated using K-injective and K-flat resolutions respectively (see [AJL, Section 1]). It follows from the Greenlees-May duality (specifically, from [AJL, Corollary 5.2.2]) that for any M ∈ Df (Mod A), there is an isomorphism of functors (RΓa)!(M ) ∼= LΛa(M ), and for any M ∈ D(Mod A) such that RΓa(M ) ∈ Df (Mod A), there is an isomorphism of functors (LΛa)!(M ) ∼= RΓa(M ). Conversely, once one knows that the functors RΓa and LΛa are twists of each other, the Greenlees-May duality R HomA(RΓa(M ), N ) ∼= R HomA(M, LΛa(N )) follows easily, so the language of Definition 3 is suited to describe this fundamental relation between these functors. Theorem 7. Let A be an essentially finite type k-algebra. (1) For any M ∈ Db f (Mod A), and any N ∈ Df (Mod A), there is an isomorphism of functors M ⊗! (2) For any M, N ∈ Db Hom! A N ∼= R HomA⊗L k A(A, M ⊗L k N ). f (Mod A), there is an isomorphism of functors A(M, N ) ∼= A ⊗L k A R Homk(M, N ) A⊗L Proof. For the first claim, by [AILN, Theorem 4.1], there is an isomorphism of functors R HomA⊗L k A(A, M ⊗L k N ) ∼= R HomA(DA(M ), N ). Since N ∈ Df (Mod A), we have that N ∼= DA(DA(N )). Hence, by the derived hom- tensor adjunction R HomA(DA(M ), N ) ∼= R HomA(DA(M ), R HomA(DA(N ), RA)) ∼= A DA(N ), RA), R HomA(DA(M ) ⊗L which proves the result. To show the second claim, by [AILN, Theorem 4.6], there is an isomorphism of functors A ⊗L k A R Homk(M, N ) ∼= DA(M ) ⊗L A N. A⊗L 4 LIRAN SHAUL Since DA(M ) ⊗L A N ∈ Df (Mod A), we have that DA(M ) ⊗L A N ∼= DA(R HomA(DA(M ) ⊗L A N, RA)). So by the derived hom-tensor adjunction DA(R HomA(DA(M )⊗L AN, RA)) ∼= DA(R HomA(DA(M ), DA(N ))) = Hom! A(M, N ). (cid:3) The main reason we introduced the twisting formalism is that relations between functors are preserved between their twists: Proposition 8. Let A, B, C be three essentially finite type k-algebras. Let F : D(Mod A)m → D(Mod C)k, G : D(Mod B)n → D(Mod C)k and H : D(Mod A)m → D(Mod B)n be three functors such that there is an isomorphism of functors F ∼= G ◦ H. Then there is an isomorphism of functors F !(M1, . . . , Mm) ∼= G! ◦ H !(M1, . . . , Mm) for any M1, . . . , Mm ∈ D(Mod A) such that H ◦ DA(M1, . . . , Mm) ∈ Df (Mod B)n. Proof. Since F ∼= G ◦ H, it follows that F ! ∼= (G ◦ H)! := DC ◦ G ◦ H ◦ DA. On the other hand, by definition G! ◦ H ! := DC ◦ G ◦ DB ◦ DB ◦ H ◦ DA. By assumption, H ◦ DA has finitely generated cohomology. Hence, DB ◦ DB ◦ H ◦ DA H ◦ DA which proves the result. ∼= (cid:3) From this proposition, the following relations between twisted functors follow immedi- ately: Corollary 9. Let A be an essentially finite type k-algebra. Then the following holds: (1) Let B be another essentially finite type k-algebra, and let f : A → B be a k- f (Mod A) there is a bifunctorial isomorphism algebra map. For any M, N ∈ D+ f !(M ⊗! A N ) ∼= (f !(M )) ⊗! B (f !(N )). in D(Mod B). (2) For any M, N, K ∈ D+ f (Mod A), there is a trifunctorial isomorphism A (N ⊗! A K) ∼= (M ⊗! A N ) ⊗! A K M ⊗! in D(Mod A). (3) For any M ∈ D+ f (Mod A), N ∈ Db f (Mod A) and K ∈ D− f (Mod A), there is a trifunctorial isomorphism Hom! A(M ⊗! A N, K) ∼= Hom! A(M, Hom! A(N, K)) in D(Mod A). Proof. Each of these statements follows from applying Proposition 8 to the following canonical isomorphisms: A N ) ⊗L A (N ⊗L A B ∼= (M ⊗L A K) ∼= (M ⊗L A B) ⊗L A N ) ⊗L B (N ⊗L A K. A B). (1) (M ⊗L (2) M ⊗L (3) R HomA(M ⊗L A N, K) ∼= R HomA(M, R HomA(N, K)). (cid:3) RELATIONS BETWEEN DERIVED HOCHSCHILD FUNCTORS VIA TWISTING 5 Remark 10. It follows easily from this corollary that the operation − ⊗! A − defines a symmetric monoidal structure on D+ f (Mod A) for any essentially finite type k-algebra A. The canonical dualizing complex RA is a monoidal unit. If B is another essentially finite type k-algebra, and f : A → B is a k-algebra map, then the twisted inverse image functor f ! : D+ f (Mod A) → D+ f (Mod B) is a monoidal functor. Further, restricting attention to the subcategory of D+ f (Mod A) made of complexes with finitely generated cohomology and of finite injective dimension over A, one obtain a closed symmetric monoidal category, with the internal hom being Hom! A(−, −). Remark 11. We first encountered the idea that Hochschild cohomology defines a sym- metric monoidal structure in [Ga]. There, in [Ga, Corollary 5.6.8], assuming k is a field of characteristic zero, it was stated without proof that the operation R HomA⊗kA(A, − ⊗k −) defines a symmetric monoidal structure on the category of indcoherent sheaves on Spec A. Combining Theorem 7 and Corollary 9 we immediately obtain various relations be- tween the derived Hochschild functors and the twisted inverse image functor. Corollary 12. Derived Hochschild cohomology commutes with the twisted inverse image functor: Let k be a regular noetherian ring of finite Krull dimension, and let A, B be two essentially finite type k-algebras. Let f : A → B be a k-algebra map. Let M, N ∈ f (Mod A) and assume that the complexes f !(M ), f !(N ) have bounded cohomology. Db Then there is a bifunctorial isomorphism f !R HomA⊗L k A(A, M ⊗L k N ) ∼= R HomB⊗L k B(B, f !(M ) ⊗L k f !(N )) in D(Mod B). Remark 13. If in the above corollary the map f : A → B has a finite flat dimension, then by [AIL2, Proposition 2.5.4], assuming that M, N have a bounded cohomology implies that f !(M ), f !(N ) have bounded cohomology. Corollary 14. Adjunction between derived Hochschild homology and derived Hochschild cohomology: Let k be a regular noetherian ring of finite Krull dimension, and let A be an essentially finite type k-algebra. Let M, N, K ∈ Db f (Mod A) be three complexes, and assume that the complexes M ⊗! A(N, K) are also bounded. Then there is a trifunctorial isomorphism A N, Hom! A ⊗L A⊗L A ⊗L k A R Homk(R HomA⊗L k A R Homk(M, A ⊗L A⊗L k A(A, M ⊗L k N ), K) ∼= k A R Homk(N, K)). A⊗L in D(Mod A). Corollary 15. Associativity of derived Hochschild cohomology: Let k be a regular noe- therian ring of finite Krull dimension, and let A be an essentially finite type k-algebra. Let M, N, K ∈ Db A N and N ⊗! f (Mod A) be three complexes, and assume that the complexes M ⊗! A K are also bounded. Then there are trifunctorial isomorphisms R HomA⊗L R HomA⊗L k R HomA⊗L k A(A, M ⊗L k A(A, R HomA⊗L k A(A, N ⊗L k N ) ⊗L k K)) ∼= k K) ∼= k A(A, M ⊗L R HomA(R HomA(M, RA) ⊗L A R HomA(N, RA) ⊗L A R HomA(K, RA), RA) in D(Mod A). 6 LIRAN SHAUL Proof. The first isomorphism follows from Theorem 7 and Corollary 9. To get the second k K) with (M ⊗! isomorphism, first replace R HomA⊗L A N ) ⊗! (cid:3) A K, and now use the derived hom-tensor adjunction. k A(A, R HomA⊗L k A(A, M ⊗L k N )⊗L The second isomorphism in the above Corollary can be thought of as a reduction for- mula for derived 3-Hochschild cohomology. One might wonder if this functor is canoni- k K). The answer to this is positive, cally isomorphic to R HomA⊗L and in fact, more generally, for every n > 1, there is an isomorphism of functors A · · · ⊗! k Mn) ∼= M1 ⊗! k A(A, M ⊗L (A, M1 ⊗L k M2 ⊗L k · · · ⊗L k N ⊗L R Hom A Mn k A⊗L A ⊗L k A ⊗L k · · · ⊗L {z n k A } For every complexes of A-modules M1, . . . Mn which satisfy suitable finiteness condi- tions. Proof of this fact will appear elsewhere (the case n = 2 is, as seen above, just a rephrase of [AILN, Theorem 4.1]). In the special case where n = 4, this follows easily, using Corollary 12, under an additional flatness hypothesis: Corollary 16. Let k be a regular noetherian ring of finite Krull dimension, and let A be a flat essentially of finite type k-algebra. Let M1, M2, M3, M4 ∈ Db f (Mod A) be four complexes. Assume that M1 ⊗! A M4 are also bounded. Then there is a quad- functorial isomorphism A M2, M3 ⊗! R HomA⊗kA⊗kA⊗kA(A, M1 ⊗L R HomA(R HomA(M1, RA) ⊗L R HomA(M3, RA) ⊗L k M3 ⊗L k M4) ∼= k M2 ⊗L A R HomA(M2, RA) ⊗L A A R HomA(M4, RA), RA). Hence, under the above hypothesis, the quad-functor R HomA⊗kA⊗kA⊗kA(A, − ⊗L − ⊗L D(Mod A)4 → D(Mod A). k −) is canonically isomorphic to the twisting of the functor − ⊗L A − ⊗L A − ⊗L k − ⊗L A − : k Proof. Let C = A ⊗k A, and let ∆ : C → A be the diagonal map. Then by Corollary 12, there is a natural isomorphism ∆!((M1 ⊗L k M2) ⊗! C (M3 ⊗L k M4)) ∼= ∆!(M1 ⊗L k M2) ⊗! A ∆!(M3 ⊗L k M4). Since ∆ is a finite map, ∆!(−) ∼= R HomC (A, −). By Theorem 7, the left hand side is canonically isomorphic to R HomA⊗kA(A, R HomA⊗kA⊗kA⊗kA(A ⊗k A, (M1 ⊗L k M2) ⊗L k (M3 ⊗L k M4))) and by the derived hom-tensor adjunction this is canonically isomorphic to R HomA⊗kA⊗kA⊗kA(A, (M1 ⊗L k M2) ⊗L k (M3 ⊗L k M4)). Applying Theorem 7 to the right hand side, we obtain: DA(DA(DA(DA(M1) ⊗L A DA(M2))) ⊗L The result now follows from the fact that DA ◦ DA DA(M3) ⊗L A DA(M4). A DA(DA(DA(M3) ⊗L ∼= 1 on DA(M1) ⊗L A DA(M4)))). A DA(M2) and on (cid:3) In the remaining of this note, we analyze how the above twisting operations act on du- alizing complexes. It has been known long ago (see [RD, Theorem V.3.1]) that the set of isomorphism classes of dualizing complexes over a commutative ring (at least if Spec A is connected) is classified by the Picard group of A and the integers. Moreover, this set is isomorphic to a set which naturally carries a group structure, namely, the group of iso- morphism classes of tilting complexes (See [Ye] for details). Let A be a commutative RELATIONS BETWEEN DERIVED HOCHSCHILD FUNCTORS VIA TWISTING 7 noetherian ring. Recall that a complex P ∈ D(Mod A) is called a tilting complex if there A Q ∼= A. If R1, R2 are two dualizing exist a complex Q ∈ D(Mod A) such that P ⊗L complexes over A then P = R HomA(R1, R2) is a tilting complex, and there is an iso- A P . The set of isomorphism classes of tilting complexes under the morphism R2 derived tensor product operation form an abelian group, called the derived Picard group of A and denoted by DPic(A). Given any dualizing complex R, the duality operation R HomA(−, R) defines a bijection between the set of isomorphism classes of dualizing complexes and the set of isomorphism classes of tilting complexes. ∼= R1 ⊗L In the next result, we observe, using the above formalism, that the set of isomorphism classes of dualizing complexes also carries naturaly a group structure, with the operation being the functor underlying derived Hochschild cohomology (that is, Shukla cohomol- ogy). Theorem 17. Let k be a regular noetherian ring of finite Krull dimension, and let A be an essentially finite type k-algebra. Then the set DA of isomorphism classes of dualizing com- k A(A, − ⊗L plexes over A form an abelian group with respect to the operation R HomA⊗L −). The inverse of a dualizing complex R is given by R′ := A ⊗L k A R Homk(R, RA). The rigid dualizing complex RA is the identity of the group. The map R HomA(−, RA) is a group isomorphism between DA and DPic(A). If B is another essentially finite type k algebra, and f : A → B is a k-algebra map then f ! : DA → DB is a group homomor- phism. A⊗L k Proof. First, suppose that R1, R2 are dualizing complexes over A. Then DA(R1) and DA(R2) are tilting complexes, so that P = DA(R1) ⊗L A DA(R2) is also a tilting complex. Hence, DA(P ) = R HomA(P, RA) ∼= R HomA(P, A) ⊗L A RA. But R HomA(P, A) is also tilting, so that DA(P ) ∼= R1 ⊗! by Theorem 7, the complex A R2 is a dualizing complex, so R HomA⊗L k A(A, R1 ⊗L k R2) is also a dualizing complex. Next, let R be a dualizing complex over A. Let R′ = A(R, RA). A similar calculation to the above now A ⊗L shows that R′ is a dualizing complex, and that k A R Homk(R, RA) ∼= Hom! A⊗L R HomA⊗L k A(A, R ⊗L k R′) ∼= R ⊗! A R′ ∼= RA. By Corollary 15, the operation R HomA⊗L k −) is associative. It follows that DA is an abelian group. It is clear that the map DA(−) : DA → DPic(A) is bijective (the inverse map is also DA). To see that it is a group map, simply note that k A(A, − ⊗L DA(R HomA⊗L = DA(DA(DA(R1) ⊗L k A(A, R1 ⊗L k R2)) ∼= DA(R1 ⊗! A DA(R2))) ∼= DA(R1) ⊗L A R2) = A DA(R2). Finally, if f : A → B is a k-algebra map, then it is well known that f ! maps DA to DB, and Corollary 12 shows that it is a homomorphism. (cid:3) Remark 18. Returning to the symmetric monoidal structure on D+ f (Mod A) described in Remark 10 above, one may easily see that the units of this structure are precisely the dualizing complexes. Thus, the fundamental result that f ! carries dualizing complexes to dualizing complexes is expressed by the fact that monoidal functors carry unit elements to unit elements. Similarly, the fact that f ! carries rigid dualizing complexes to rigid dualizing 8 LIRAN SHAUL complexes ([YZ2, Theorem 0.1], [AIL2, Corollary 3.3.5]) can now be expressed simply by saying that monoidal functors preserve the monoidal identity. We end this note with a couple remarks about possible generalizations of the above theory. Remark 19. In [AILN, Corollary 6.5], there is a global version of the reduction formula for derived Hochschild cohomology under the additional assumption that the given scheme is flat over the base. A similar result for derived Hochschild homology is shown in [ILN, Theorem 4.1.8]. Using these results, all results of this note immediately generalize to the global case of schemes, under the additional assumption that they are flat over k. Remark 20. Another possible generalization is relaxing the assumptions on k. As a first step, one can relax the regularity assumption and assume instead that k is Gorenstein. Rigid dualizing complexes still exist (see [AIL1, Theorem 8.5.6]), so most of the above will still make sense and will be true, under the additional assumption that all algebras are of finite flat dimension over k. Going further, we can simply assume that k is a noetherian ring. Then, in the (possible) absence of dualizing complexes, we may use instead the notion of a relative dualizing complex (see [AILN, Section 1]). Again, we will have to assume that all algebras are of finite flat dimension over k, and further, to have the biduality isomorphism of [AILN, Theorem 1.2], we must assume that all complexes involved are also of finite flat dimension over k. Acknowledgments. The author would like to thank Professor Joseph Lipman and Pro- fessor Amnon Yekutieli for some useful suggestions. The author also wish to thank the Department of Mathematics at the Weizmann Institute of Science, where this work was carried out. REFERENCES [AIL1] Avramov, L., Iyengar, S., and Lipman, J. (2010). Reflexivity and rigidity for complexes, I: Commuta- tive rings. Algebra and Number Theory, 4(1), 47-86. [AIL2] Avramov, L., Iyengar, S., and Lipman, J. (2011). Reflexivity and rigidity for complexes, II: Schemes. Algebra and Number Theory, 5(3), 379-429. [AILN] Avramov, L., Iyengar, S., Lipman, J., and Nayak, S. (2010). Reduction of derived Hochschild functors [AJL] [Ga] [ILN] [Li] [RD] [VdB] [Ye] [YZ1] [YZ2] over commutative algebras and schemes. Advances in Mathematics, 223(2), 735-772. Alonso L., Jeremias A. and Lipman, J., Local homology and cohomology of schemes, Ann. Sci. ENS 30 (1997), 1-39. Gaitsgory, D. Ind-coherent sheaves. arXiv preprint arXiv:1105.4857 (2011). Iyengar, S., Lipman, J., and Neeman, A. (2013). Relation between two twisted inverse image pseudo- functors in duality theory. arXiv preprint arXiv:1307.7092. Lipman, J. (2009). Notes on derived functors and Grothendieck duality, Foundations of Grothendieck Duality for Diagrams of Schemes, Lecture Notes in Math. 1960, Springer-Verlag, New York; 1-261. R. Hartshorne, "Residues and Duality," Lecture Notes in Math. 20, Springer-Verlag, Berlin, 1966. Van den Bergh, M. (1997). Existence theorems for dualizing complexes over non-commutative graded and filtered rings. Journal of Algebra, 195(2), 662-679. Yekutieli, A. (1999). Dualizing complexes, Morita equivalence and the derived Picard group of a ring. Journal of the London Mathematical Society, 60(3), 723-746. Yekutieli, A., and Zhang, J. J. (2009). Rigid dualizing complexes over commutative rings. Algebras and representation theory, 12(1), 19-52. Yekutieli, A., and Zhang, J. J. (2004). Rigid dualizing complexes on schemes. arXiv preprint math/0405570. UNIVERSITEIT ANTWERPEN, DEPARTEMENT WISKUNDE-INFORMATICA, MIDDELHEIM CAMPUS, MID- DELHEIMLAAN 1, 2020 ANTWERP, BELGIUM E-mail address: [email protected]
1209.4342	5	1209	2013-07-22T13:34:22	One Cycles on Rationally Connected Varieties	[ "math.AG" ]	All curves on a separably rationally connected variety are rationally equivalent to a (non-effective) integral sum of rational curves, hence the first Chow group is generated by rational curves. Applying the same techniques, we also proved that the first Chow group of all separably rationally connected Fano complete intersections with index at least 2 is generated by lines. As a consequence, a question of Professor Burt Totaro about integral Hodge classess on rationally connected 3-folds is solved, and positive answer to the question for general n-fold due to Professor J\'anos Koll\'ar will follow from the Tate conjecture for surfaces over finite fields.	math.AG	math	One cycles on rationally connected varieties Zhiyu Tian and Hong R. Zong Abstract We prove that every curve on a separably rationally connected variety is rationally equivalent to a (non-effective) integral sum of rational curves. That is, the Chow group of 1-cycles is generated by rational curves. Applying the same technique, we also show that the Chow group of 1-cycles on a separably rationally connected Fano complete intersections of index at least 2 is generated by lines. As a consequence, we give a positive answer to a question of Professor Totaro about integral Hodge classes on rationally connected 3-folds. And by a result of Professor Voisin, the general case is a consequence of the Tate conjecture for surfaces over finite fields. 1. Introduction In [Kol10] and [Voi12], the following question is asked by Professor J´anos Koll´ar and Professor Claire Voisin (and in the case of dimension 3 by Professor Burt Totaro): Question 1.1. Is every integral Hodge (n − 1, n − 1)-class on a smooth projective rationally connected n-dimensional variety over C a Z–linear combination of cohomology classes of rational curves? This question can be separated into two questions, as in [Voi12]. – Is every integral Hodge (n − 1, n − 1)-class on a smooth projective rationally connected n-fold over C a Z-linear combination of the cohomology classes of curves? – Is every curve class on a smooth projective rationally connected n-fold over C a Z-linear combination of the cohomology classes of rational curves? The first question is known in some special cases. Theorem 1.2. Let X be a smooth projective variety over C. Then every integral Hodge (n − 1, n − 1)-class on X is a Z-linear combination of the cohomology classes of curves if X is one of the followings. – X is a uniruled or Calabi-Yau 3-fold.[Voi06] – X is a Fano 4-fold. [HV11] – X is a Fano 5-fold with index 2. [HV11] – X is a Fano n-fold (n > 8) with index n − 3. [Flo] It is shown in [Voi12] that a positive answer to the first question in general is implied by the Tate conjecture for divisor classes on surfaces defined over finite fields. 2010 Mathematics Subject Classification 14M22 (primary), 14C25(secondary). Keywords: Rationally connected varieties; Algebraic cycles Zhiyu Tian and Hong R. Zong Our main theorem gives a positive answer to the second question. Theorem 1.3. Let X be a smooth proper and separably rationally connected variety over an algebraically closed field. Then every 1-cycle is rationally equivalent to a Z-linear combination of the cycle classes of rational curves. That is, the Chow group CH1(X) is generated by rational curves. Recall that rational equivalence implies cohomological equivalence over the complex numbers C. Thus combining Theorem 1.2 and Theorem 1.3, we have: Corollary 1.4. Let X be a smooth projective variety over C. Then every integral Hodge (n − 1, n − 1)-class on X is a Z-linear combination of the cohomology classes of rational curves if X is one of the followings. – X is a rationally connected 3-fold. – X is a Fano 4-fold. – X is a Fano 5-fold with index 2. – X is a Fano n-fold (n > 8) with index n − 3. In particular, the 3-fold case of Question 1.1 is solved. And by [Voi12], we have: Corollary 1.5. Assume that the Tate conjecture for surfaces over finite fields is true. Then every integral Hodge (n − 1, n − 1)-class on a smooth projective rationally connected n-fold is a Z-linear combination of the cohomology classes of rational curves. For the proof of the main theorem, one can first reduce rational equivalence to algebraic equiv- alence (Lemma 3.4). Then in characteristic 0 one can apply the main construction of [GHS03] to the trivial product X × P1 and degenerate any curve to a sum of rational curves (this part is already stated in the unpublished note [Zon12a]). Professor J´anos Koll´ar pointed out that a more direct and simpler proof is possible by looking at the natural forgetful map Φ : M g,0(X, [C]) → M g,0 from the Kontsevich moduli space of stable maps to the moduli space of stable curves. Once we know this map is surjective, a degeneration argument proves the main theorem in all character- istics. We still include a section on the proof by the product trick which started the whole projects with a few remarks and applications. Next we turn to the study of one cycles on low degree complete intersections. First we have: Theorem 1.6. Let X be a (possibly singular) complete intersection of type (d1, . . . , dc) with d1 + . . . + dc 6 n − 1. Then every rational curve on X is algebraically equivalent to an effective sum of lines. Then by the same technique proving our main theorem, we have the following. Theorem 1.7. Let X be a smooth complete intersection of type (d1, . . . , dc) over an algebraically closed field k. Assume d1 + . . . + dc 6 n − 1. If Char(k) = p > 0, also assume that X is separably rationally connected. Then the Chow group of 1-cycles CH1(X) is generated by lines. Remark 1.8. A general Fano hypersurface in characteristic p > 0 is separably rationally con- nected [Zhu11]. And it is plausible that the same is true for a general Fano complete intersection. 2 One cycles on rationally connected varieties Corollary 1.9. Let X be a smooth separably rationally connected complete intersection of degree (d1, . . . , dc) such that c X i=1 di(di + 1) 2 6 n. Then the Chow group of 1-cycles CH1(X) is isomorphic to Z. Proof. Under the assumption, the Chow group of 1-cycles is generated by lines and any two lines are rationally equivalent since the Fano scheme of lines in X is rationally chain connected (c.f. Proof of Theorem 4.2, Chap. V, [Kol96]). Remark 1.10. One can prove that the Chow group of one cycles is isomorphic to Z for all complete intersections with a slightly worse degree bound (Proposition 7.2). The result with Q coefficients is well-known (Theorem 4.2, Chap. V, [Kol96]) even without the smoothness assump- tion. 2.1 Rationally connected varieties and smoothing of curves Recall the following definition. 2. Preliminaries Definition 2.1. Let X be a proper variety over a field k. It is separably rationally connected (SRC) if it is smooth and there is a generic smooth family of 1–cycles with geometrically rational components: u −→ X U g ↓ B , over k such that the double evaluation map: U ×B U (u,u) −→ X × X is generically ´etale and dominates X × X. If we drop the ´etale condition, it is called rationally connected (RC). And if we don't require the generic smoothness of U → B, it is called rationally chain connected (RCC). Definition 2.2. Let X be a smooth proper variety. A morphism f : P1 → X is free (resp. very free) if f ∗TX is non-negative (resp. ample). A variety over an algebraically closed field is separably rationally connected if and only if there is a very free rational curve. Definition 2.3. Let k be an arbitrary field. A comb with n teeth over k is a projective curve with n + 1 irreducible components C0, C1, . . . , Cn over ¯k satisfying the following conditions: (i) The curve C0 is defined over k. (ii) The union C1 ∪ · · · ∪ Cn is defined over k (Each individual curve may not be defined over k). (iii) The curves C1, . . . , Cn are smooth rational curves disjoint from each other, and each of them meets C0 transversely in a single smooth point of C0 (which may not be defined over k). 3 Zhiyu Tian and Hong R. Zong The curve C0 is called the handle of the comb, and C1, . . . , Cn the teeth. A rational comb is a comb whose handle is a smooth rational curve. One of the most important techniques in studying separably rationally connected varieties is the smoothing of a comb. Given a morphism f : C → X from a comb C with handle C0, a smoothing of f is a family Σ → T over a pointed curve (T, 0) together with a morphism F : Σ → X such that Σ0 ∼= C, F Σ0 = f and Σt is a smooth curve, isomorphic to C0, for a general t. The following result is implicit in the book [Kol96] and follows from the work of Koll´ar- Miyaoka-Mori. Proposition 2.4. Given a morphism from a smooth projective curve f0 : C0 → X to a smooth proper and separably rationally connected variety X over a field k, and an integer d, there are p ≫ 0 very free rational curves fi : Ci → X, 1 6 i 6 p, such that (i) C = C0 ∪ C1 ∪ ... ∪ Cp is a comb in the sense of 2.3. Furthermore, there is a morphism f : C → X (defined over k) and a smoothing of the comb Σ → T, F : Σ → X. (ii) H 1(Σt, F ∗ t TX ⊗ M ) = 0 for a general member Ft : Σt → X of the smoothing and any line bundle M of degree d. For the proof, we need the following lemma. Lemma 2.5. Let D be a smooth projective connected curve of genus g and f : D → X be a morphism to a smooth variety. Then given an integer d, there is a number n such that for any line bundle L of degree at least n, H 1(D, f ∗TX ⊗ L ⊗ M ) = 0 for any line bundle M of degree d. Proof. There is a line bundle L0 on D such that H 1(D, f ∗TX ⊗ L0) = 0. Choose a line bundle L1 of degree at least g − d. Then H 0(D, OD(L1 + M )) is non-zero for any line bundle M of degree d since any divisor of degree at least g is rationally equivalent to an effective divisor. Take EM to be an effective divisor in the linear system L1 + M . Then we have the short exact sequence of sheaves 0 → f ∗TX ⊗ O(L0) → f ∗TX ⊗ O(L0 + EM ) → f ∗TX ⊗ O(L0 + L1 + M )EM → 0. Thus H 1(D, f ∗TX ⊗ O(L0 + L1 + M )) = 0 for any line bundle M of degree d. Take n to be deg L0 + deg L1 + g. Then for any line bundle L whose degree is at least n, H 0(D, OD(L − L0 − L1)) is non-zero, again since any divisor of degree at least g is rationally equivalent to an effective divisor. Take a divisor E in the linear system L − L0 − L1. We have the following short exact sequence: 0 → f ∗TX ⊗ O(L0 + L1 + M ) → f ∗TX ⊗ O(E + L0 + L1 + M ) → f ∗TX ⊗ O(L + M )E → 0. Thus H 1(D, f ∗TX ⊗ O(L + M )) = 0. Proof of Proposition 2.4. We can assemble a comb whose handle is C0 and whose teeth are very free curves Ci, 1 6 i 6 m. By Theorem 7.9, Chap II, [Kol96], at least m − h1(C0, f ∗TX ) teeth can be smoothed. We choose m to be large enough so that m − h1(C0, TX C ) > n, where n is the number in Lemma 2.5. Then a general smoothing Σt of the subcomb has H 1(Σt, F ∗ t TX ⊗ M ) = 0 for any line bundle M of degree d by Lemma 2.6 (Take E to be F ∗TX ⊗ N , where N is the line bundle on Σ whose restriction to a general fiber and the handle C0 is isomorphic to M and the restriction to the teeth are trivial line bundles). 4 One cycles on rationally connected varieties Lemma 2.6 [Kol96], Chap. II, Lemma 7.10.1. Let C = C0 ∪ C1 ∪ C2... ∪ Cp be a comb with p teeth as in Definition 2.3. Let q : Σ → T be a smoothing of C. And let E be a vector bundle over Σ such that the restrictions ECi are all ample for 1 6 i 6 p and H 1(C0, L ⊗ EC0 ) = 0 for every line bundle L on C0 of degree at least p. Then H 1(Σt, EΣt ) = 0 for a general t. Remark 2.7. Let f : D → X be a curve in a separable rationally connected variety X. Assume D is embedded. One can add very free rational curves to form a higher genus nodal curve and smooth it, as is done in Section 2.2, [GHS03]. When D is not embedded in X, one can first embed D in X × P3 and then do the above operation and project to X. In other word, after adding suitable rational curves, we can deform the reducible curve to a curve of higher genus in X. 2.2 Moduli space of stable maps to projective varieties Definition 2.8. Let C be a connected nodal curve, and X be a projective variety. We call a morphism f : C → X a stable map if every irreducible component of C which is mapped to a point is one of the followings: – A curve of arithmetic genus at least 2. – A curve of arithmetic genus 1 having at least 1 intersection point with other components of C. – A curve of arithmetic genus 0 having at least 3 intersection points with other components of C. Let β be a curve class of X. We have a proper moduli space of stable maps M g,0(X, β) ([FP97]). When X is a point, we recover the moduli space of stable curves M g,0 as in [DM69]. Remark 2.9. When X is not projective, we might not have a proper moduli stack of stable maps M g,0(X, β): given a family of stable maps over the generic point of a discrete valuation domain to a proper, non-projective variety, it is always possible to extend the family over the closed point (possibly after a base change) to a family of prestable maps (i.e. maps from a family of nodal curves). But it is not clear that one can extend the family over the closed point (even after a base change) to a family of stable maps. Let X and Y be projective varieties with a morphism π : X → Y . Fix a curve class β of X. We have then a natural forgetful map ([BM96]): Φ : M g,0(X, β) → M g,0(Y, π∗β) defined by composing a map f : C → X with π and collapsing components of C as necessary to make the composition π ◦ f stable. Now we have an easy but important observation. Lemma 2.10. For a stable map f : C → X, the components that are contracted under are all rational. Φ : M g,0(X, β) → M g,0(Y, π∗β) Proof. By Definition 2.8, the only possible non-stable components that need to be contracted are smooth rational curve having at most 2 intersection points with other components of C. 5 Zhiyu Tian and Hong R. Zong 3. Reducing rational equivalence to algebraic equivalence The following proposition is essentially Proposition 3.13.3, Chap. IV, [Kol96]. The existence of the number N follows from the argument there but is not explicitly stated. We include the proof here for completeness. Proposition 3.1. Let X be a proper rationally chain connected variety over an algebraically closed field. Then there is a positive integer N together with a family of effective 1-cycles with rational components u −→ X F g ↓ B , where u is the evaluation map, such that for any 1-cycle D in X, there are integers mi's and rational curves Fi's in the fibers of g : F → B, which satisfies the following rational equivalence relation: N · D ∼r.e. X miu∗(Fi). i Proof. By the definition of rational chain connectedness, there is a family of connected effective 1-cycles with rational components g : F → B in X with the evaluation map u : F → X, and such that the map m : F ×B F → X × X is generically finite of degree N and surjective. Now pick a general point x0 of X. Then u0 : u−1(x0) ×B F → x0 × X is generically finite of degree N and surjective. Denote u−1(x0) ×B F by F0. And let B0 = g(u−1(x0)). Then we have a diagram: u0−→ X F0 g ↓ B0 Fix an irreducible curve C in X. Then the class [C] is rationally equivalent to a class of the form [C1] − [C2], where C1 and C2 are curves in X containing a general point. To see this, first find a projective birational morphism X ′ → X which is an isomorphism near the generic point of C. Then there is a unique lifting of C to X ′. Thus it suffices to prove this when X is projective. Then one can find an irreducible curve C ′, which contains a general point, as a residual curve of C such that C ∪ C ′ is a complete intersection of very ample divisors. Then [C] is rationally equivalent to the difference of a general complete intersection and C ′. So we may assume that C contains a general point of X. Let C be the one dimensional component of the inverse image of C under u0. Then u0∗[ C] ∼r.e. N · C. The curve C is a section of the chain of ruled surfaces, intersecting each other at a section: F0 ×B0 C → C. There is another section C0 : u−1(x0) ×B0 C, which is mapped to x0 via the evaluation map u0. Since two sections of a ruled surface are rationally equivalent modulo rational curves in the fiber, we have C − C0 ∼r.e. X niFi i 6 One cycles on rationally connected varieties where Fi's are irreducible components of the fibers of F → B. Push forward this relation to X. Since u0(C0) = x0, we get N · C ∼r.e. X miFi. i Let CH1(X)alg be the subgroup of the Chow group of 1-cycles on X which are algebraically equivalent to 0. Here is a classical lemma. Lemma 3.2 [BO74], Lemma 7.10 . The group CHp(X)alg is a divisible group. Namely, for any integer N > 0 and any element x in CHp(X)alg, there is another element y in CHp(X)alg such that x = N · y. Proof. The subgroup CHp(X)alg is generated by cycles of the form Σp − Σq, where Σ ⊂ C × X is a family of cycles over a smooth projective curve C. So the divisibility of CHp(X)alg follows from the divisibility of CH0(C)alg, which is isomorphic to the Jacobian of the curve C. Denote by ·N the map of multiplication by N in CH1(X)alg. Then CH1(X)alg ·N −→ CH1(X)alg is surjective by the divisibility of CH1(X)alg. Together with Proposition 3.1, one immediately has: Corollary 3.3. Notation as in Proposition 3.1. Then CH1(X)alg is generated by cycles with only rational components. This allows us to reduce Theorem 1.3 to a similar statement for cycles modulo algebraic equivalence. Lemma 3.4. In order to prove the main theorem 1.3, it suffices to prove that rational curves generate the group of one cycles modulo algebraic equivalence. Proof. Assume that for any curve C in X, there are rational curves Ci's and integers n′ that is, such [C] − X ni[Ci] i is algebraically equivalent to 0, namely, an element in CH1(X)alg. Then by Corollary 3.3, we have rational equivalence relation [C] − X ni[Ci] ∼r.e. X mj[Fj] i j where mi's are integers and Fj's are rational curves. So C is rationally equivalent to an integral sum of rational curves. 4. Proof of the main theorem over C with a product trick We can use a product trick to prove the main result of this paper over the field of complex numbers C, by applying the main argument of the celebrated paper "Families of Rationally Connected Varieties" of T. Graber, J. Harris and J. Starr ([GHS03]). 7 Zhiyu Tian and Hong R. Zong Theorem 4.1. Let X be a smooth proper rationally connected variety over C. Then every curve on X is algebraically equivalent to a Z-linear combination of rational curves. Remark 4.2. It is easy to reduce to the projective case as follows: by Chow's lemma and reso- lution of singularities, there is a smooth projective variety X ′, with proper birational morphism f : X ′ → X. One can smooth any curve C ⊂ X (after adding free rational curves) to a general curve C ′ not supported in the exceptional locus of f . Then just take a lift of C ′ in X ′. Assuming X is projective, the idea of proof is to first lift any irreducible curve C in X to X × P1. By [GHS03], there are very free rational curves which are horizontal with respect to the projection to P1, just add enough of these curves to form a comb and then smooth to a curve C which is f lexible in the sense of [GHS03], i.e. the natural map M g,0(X × P1, β) → M g,0(P1, d) is proper and surjective at the corresponding component. Degenerating C to a sum of rational curves and pushing forward this relation to X, we get the desired result. The following subsection is largely a restatement of the unpublished note [Zon12a]. 4.1 Families of rationally connected varieties Let Y be a smooth projective variety with a morphism Y → P1 whose general fibers are rationally connected. For a class β ∈ H2(Y, Z) having intersection number d with a fiber of the map π. We have then a natural morphism: ϕ : M g,0(Y, β) → M g,0(P1, d). Definition 4.3. Let f : C → Y be a stable map from a nodal curve C of genus g to X with class f∗[C] = β. We say that f is flexible relative to π if the map ϕ : M g,0(Y, β) → M g,0(P1, d) is dominant at the point [f ] ∈ M g,0(Y, β) and π : C → P1 is flat. Proposition 4.4. A flexible curve f : C → Y can be degenerated to an effective sum of rational curves in Y . Proof. It is a classical fact that the variety M g,0(P1, d) has a unique irreducible component whose general member corresponds to a flat map f : C → P1, see [Ful69]. Since the map ϕ : M g,0(Y, β) → M g,0(P1, d) is proper and dominant, ϕ is surjective on the component of π : C → Y . By Lemma 2.10 it is enough to find a degeneration of C → P1 in M g,0(P1, d) as a sum of rational curves, which is elementary. Theorem 4.5. (Main Construction of [GHS03])For any multisection B → P1, there are rational curves Ci's such that B ∪ C1 ∪ ... ∪ Cm can be deformed to a flexible curve of Y → P1. Now we can prove Theorem 4.1. Proof of Theorem 4.1. Take Y = X × P1, for any irreducible curve C ⊂ X, lift it to a curve C ′ in X × 0 ⊂ Y . Since Y is rationally connected, we can add enough free curves of Y which are horizontal with respect to the projection Y → P1, such that the comb can be deformed to a multisection Γ of the fibration Y → P1. By Theorem 4.5, we can add some other rational curves to Γ and deform the reducible curve to a flexible curve. Then by Proposition 4.4, it can be degenerated to a sum of rational curves, so C ′ is algebraically equivalent to an integral sum of rational curves in Y . Pushing forward the relation to X finishes the proof. 8 One cycles on rationally connected varieties Remark 4.6. It might be possible to prove the main theorem in all characteristics by applying the argument of [dJS03] to X × P1. 5. Proof of the main theorem in all characteristics We begin with a lemma. Lemma 5.1. Let f : C → X be a morphism from a smooth projective connected curve C of genus g(> 2) to a projective variety X. Assume that the image of f lies in the smooth locus X sm of X and H 1(C, f ∗TX) = 0. Then the moduli space M g,0(X, [C]) is smooth at the point represented by (f : C → X) and the natural forgetful map: Φ : M g,0(X, [C]) → M g,0, restricted to the (unique) irreducible component containing the point represented by (f : C → X), is surjective. Proof. The deformation and obstruction space of the stable map (f : C → X) is where Ω· f is the complex Def (f ) = H1(C, RHomOC (Ω· f , OC )) Obs(f ) = H2(C, RHomOC (Ω· f , OC )) −1 0 f ∗ΩX df † −−−−→ ΩC . We have the long exact sequence 0 → H 0(C, TC ) → H 0(C, f ∗TX ) → H1(RHomOC (Ω· → H 1(C, TC ) → H 1(C, f ∗TX ) → H2(RHomOC (Ω· f , OC )) f , OC )) → 0 f , OC )) → H 1(C, TC ) is the map between tangent spaces of M g,0(X, [C]) Note that H1(RHomOC (Ω. and M g,0. Thus the vanishing of H 1(C, TX C ) implies both the surjectivity of the tangent space map and smoothness of M g,0(X, [C]) at the point represented by (f : C → X). Therefore the restriction of the forgetful morphism Φ to the component containing the point (f : C → X) is dominant. Since the morphism Φ is also proper when restricted to this component, it is also surjective. Now we can finish the proof of the main theorem by the following result and Lemma 3.4. Theorem 5.2. Let X be a smooth proper and separably rationally connected variety over an algebraically closed field of arbitrary characteristic. Then every curve on X is algebraically equiv- alent to a Z-linear combination of rational curves. Proof. Choose an irreducible curve f : C → X. We may assume the genus of C is at least 2 by Remark 2.7. Then by Proposition 2.4, we may also assume that H 1(C, f ∗TX (−p)) = 0 for any point p in C (up to adding very free rational curves and smoothing). Note that H 1(C, f ∗TX) also vanishes in this case. By Chow's lemma, there is a normal projective variety X ′ with a birational morphism π : X ′ → X. The exceptional locus of π has codimension at least 2 in X. Thus there is a open subset U of X, whose compliment has codimension at least 2, and is isomorphic to an open subset V of X ′. Since H 1(C, f ∗TX(−p)) = 0, a general deformation of C lies in U . By 9 Zhiyu Tian and Hong R. Zong abuse of notations, we still write the general deformation as C. Then by upper semi-continuity, H 1(C, f ∗TU ) = H 1(C, f ∗TX ) = H 1(C, f ∗TX ′) = 0. And it suffices to prove the statement for the map f : C → U ∼= V ⊂ X ′. Finally Lemma 5.1 implies that the forgetful map Φ : M g,0(X ′, [C]) → M g,0 is surjective when restricted to the irreducible component of the Kontsevich moduli space of stable map M g,0(X ′, [C]) containing the point represented by (f : C → X ′). The moduli space M g,0 is irreducible [DM69]. So one can specialize the image of C in M g,0 to a singular stable curve whose irreducible components are all rational curves. And then choose any preimage of the point in the irreducible component of M g,0(X ′, [C]) containing (f : C → X ′). By Lemma 2.10, the preimage is represented by a map from a curve with only rational components to X ′. Thus the curve class [C] is algebraically equivalent to the class of a union of rational curves. Remark 5.3. An elliptic curve without any marked point is not stable. So we increase the genus first to get a stable curve. One can also use M g,1 and slightly modify Lemma 5.1 to adapt to this case. 6. One cycles on low degree complete intersections In this section, we prove the following theorem by the same technique as in previous sections. Theorem 6.1. Let X be a smooth complete intersection of type (d1, . . . , dc) over an algebraically closed field k. Assume d1 + . . . + dc 6 n − 1. If Char(k) = p, also assume that X is separably rationally connected. Then the Chow group of 1-cycles CH1(X) is generated by lines. By the assumption, X is rationally chain connected by chains of lines (Lemma 4.8.1, Chap. V of [Kol96]). Let F (X) be the Fano scheme of lines of X. Then (i) CH0(F (X)) ⊗ Q → CH1(X) ⊗ Q is surjective ( Proposition 3.13.3, Chap. IV [Kol96]). (ii) The cokernel of CH0(F (X))alg → CH1(X)alg is annihilated by an integer N (c.f. Proposition 3.1). Thus it suffices to show that every curve is algebraically equivalent to an integral sum of lines by the same argument as in Lemma 3.4. By the main theorem 1.3, it suffices to show that every rational curve is algebraically equivalent to an integral sum of lines. Theorem 6.2. Let X be a (possibly singular) complete intersection of type (d1, . . . , dc) with d1 + . . . + dc 6 n − 1 in P(V ) ∼= Pn. Then every rational curve on X is algebraically equivalent to an effective sum of lines. We will need the following connectedness result of Hartshorne. Lemma 6.3 (Hartshorne, [Har62]). Let X be a subscheme in PN defined by M homogeneous polynomials. And let Y be a closed subset of X with dimension less than N − M − 1. Then X − Y is connected. Proof of Theorem 6.2. We use induction on the degree of the considered rational curve. The degree 1 case is obvious. 10 One cycles on rationally connected varieties Assume the statement is true for all rational curves whose degree is less than e(> 2). Fix a copy C of P1, and consider the parameter space P := P(Hom(V ∗, H 0(C, O(e)))) If we choose homogeneous coordinates x0, . . . , xn on P(V ), then P parametrizes (n + 1)-tuples [u0, ..., un] of homogeneous degree e polynomials on C. Let X be a complete intersection of codimension c of type (d1, . . . , dc) defined by nonzero degree di homogeneous polynomials Fi(1 6 i 6 c) on P(V ). Define PX to be the closed subset of P parameterizing [u0, ..., un] such that Fi(u0, ..., un) equals 0 for all 1 6 i 6 c. The subvariety PX is defined by homogeneous polynomials of degree , . . . , dc, . . . , dc d1, . . . , d1 } } {z ed1+1 {z edc+1 on P. Let B ⊂ P be the closed subvariety parameterizing tuples [u0, . . . , un] where span{u0, ..., un} in H 0(C, O(e)) is 1-dimensional, i.e., every pair ui, uj satisfies a scalar linear relation. The codi- mension of B in P is ne. Let D ⊂ P be the closed subvariety parameterizing (n + 1)-tuples [u0, ..., un] that have a common zero in C. Clearly D contains B. The point outside D parametrizes a degree e map from C to Pn. Now we see that PX ∩ B parameterizes a degree e polynomial in two variables (up to scaling) and a point in X. So its dimension is n − c + e. Then in the situation of Lemma 6.3, take Y to be PX ∩ B and N = ne + n + e, M = e(d1 + ... + dc) + c. The condition n − c + e < N − M − 1 is simply Therefore PX − B is connected for all e > 2. e(n − d1 − d2 − ... − dc) > 1. Now let u = [u0, . . . , un] be any element of PX − D, e.g., a parameterized, degree e morphism from C to X. Let u′ be a degree e multiple cover of a line in X. Then u′ is algebraically equivalent to a union of lines. Since PX − B is connected, there exists a connected curve mapping to PX −B whose image connects these two points. We may assume that it is a chain of irreducible components, with each consecutive pair meeting in a single node. If none of the nodes is contained in D, then after deleting finitely many points from the curve, none of which are the nodes, we may assume the entire connected curve is disjoint from D. Thus this connected curve parameterizes a family of stable maps. Then u is algebraically equivalent to u′, and hence algebraically equivalent to a union of lines. So it remains to consider the case when some of the nodes lie in D but not in B. Consider the first such node, starting from the component containing u. Approaching the node from the component closer to u, we have a family of elements in the complement of D which approaches a point of D, i.e., we have a family of stable maps which approaches a point of D. But this is a point of D which is not in B. So the corresponding rational map from C to X is non-constant of degree strictly less than e. It follows that the limiting stable map has at least one component of positive degree which is strictly less than e, i.e., this is a point in the boundary. But then, each component of this boundary stable map has degree strictly less than e, hence by induction each component is algebraically equivalent to a union of lines. Thus u is also algebraically equivalent to a union of lines. 11 Zhiyu Tian and Hong R. Zong Remark 6.4 (suggested by C. Voisin). The argument presented here proves that for a smooth Fano complete intersection of type (d1, . . . , dc) such that d1 + . . . + dc 6 n − 1, the Griffiths group of one cycles homologous to 0 modulo algebraic equivalence is trivial. Indeed, this follows from the statement that every one cycle is rationally equivalent to a linear combination of lines and the fact that the Fano scheme of lines is connected in all but one case, i.e. the quadric surface in P3 ([Kol96], Theorem 4.3, Chap. V, [Zon12b], Theorem 1.3). But the statement is easily seen to be true for the quadric surface. The product trick gives a weak form of Proposition 3.1. 7. Remarks on the product trick Proposition 7.1. Let X be a proper rationally connected variety over an algebraically closed field, then CH1(X) ⊗Z Q is generated by rational curves. Proof. For any curve C in X, consider the graph Γ ⊂ C × X. Take any point x ∈ X. Let C ′ be the trivial section x × C ⊂ X × C. Since X ×C K(C) is a rationally connected variety over the function field K(C) of C, after a base change S → C there is a rational curve in X ×C K(S) connecting Γ ×C K(S) and C ′ ×C K(S ′). Thus there is a ruled surface with two sections Γ and C ′. Recall again the fact that two sections of a ruled surface are rationally equivalent modulo fibers, which are all rational curves. The result follows from pushing forward everything to X and the fact that the push-forward of C ′ is 0 in the Chow group of X. We give one more application of the product trick. Although, strictly speaking, the trick is not necessary, it does make the proof more transparent. i=1 d2 Proposition 7.2. Let X ⊂ Pn be a closed subscheme defined by equations of degree d1, . . . , dc such that Pc Proof. Let f : C → X be a curve in X. We first show that C is rationally equivalent to a sum of lines in X. i 6 n. Then CH1(X) ∼= Z. Consider the graph Γ of the map in C × X. Choose a trivial section S = C × x of C × X → C, where x is a closed point in X. Now think of the fibration over C as a scheme X defined over K(C), the function field of C, by equations of degree d1, . . . , dc and the graph and the trivial section as two rational points of X . Under the assumptions on di's, X is rationally chain connected by chains of two lines. And furthermore, the scheme parameterizing the chain of two lines containing two points is defined by equations of degree (1, 2, . . . , d1 − 1, 1, 2, . . . , d1 − 1, d1, . . . , 1, 2, . . . , dc − 1, 1, 2, . . . , dc − 1, dc). To see this, just consider the equations for the intersection point of the two lines. In particular, the scheme parameterizing chains of two lines containing the two rational points is a scheme defined over K(C) by equations of the above degrees. Then by Tsen's theorem, there is a K(C)- rational point in the scheme. The chain of lines between the two rational points is equivalent to a chain of ruled surfaces over C containing the two sections Γ and S as sections (of the ruled surface). So in C × X. Then push forward the relations to X. Notice that the push-forward of [S] is 0. Thus we get the desired relation in X. Γ ∼r.e. S + lines 12 One cycles on rationally connected varieties Finally note that any two lines in X are rationally equivalent since the Fano scheme of X is rationally chain connected under the assumptions on di. Acknowledgements We thank Professor Burt Totaro for introducing the question to us by his enlightening lectures, correcting many mistakes in the first draft of this paper, and helping us to form the final ar- gument, Professor Claire Voisin for the argument of reducing rational equivalence to algebraic equivalence and Remark 6.4, Professor J´anos Koll´ar for his constant support for the second named author and enlightening comments on the proof, Professor Jason Starr for helpful discus- sions about rational curves on Fano complete intersections, and Professor Chenyang Xu for very encouraging comments during the project. References AK03 C. Araujo and J. Koll´ar, Rational Curves on Varieties, Higher Dimensional Varieties and rational points (Budapest 2001), Bolyai Soc. Math. Stud., vol. 12, Springer, Berlin, 2003, pp. 13-68. BO74 S.Bloch, O.Ogus, Gersten's conjecture and the homology of schemes, Annales scientifiques de l'´Ecole Normale Suprieure, 4e s´erie, tome 7, no 2 (1974), p. 181-201. BM96 K. Behrend and Yu. Manin. Stacks of stable maps and Gromov-Witten invariants. Duke Math. J. Volume 85, Number 1 (1996), 1-60. dJS03 A.J. de Jong, J. Starr, Every rationally connected variety over the function field of a curve has a rational point, American Journal of Mathematics, 125, 567-580 (2003). DM69 P. Deligne and D. Mumford, The irreducibility of the space of curves of given genus, Publications Mathmatiques de l'IHS (Paris), 1969, 36: 75109. Flo E. Floris. Fundamental divisors on Fano varieties of index n-3. to appear in Geometriae Dedicata. Ful69 W. Fulton, Hurwitz Schemes and Irreducibility of Moduli of Algebraic Curves, Ann. of Math., Vol. 90, No. 3, Nov., 1969, page 542-575. FP97 W. Fulton and R. Pandharipande, Notes on stable maps and quantum cohomology, Notes on stable maps and quantum cohomology. Algebraic geometry Santa Cruz 1995, 45-96, Proc. Sympos. Pure Math., 62, Part 2, Amer. Math. Soc., Providence, RI, 1997. GHS03 T. Graber, J. Harris and J. Starr, Families of rationally connected varieties. J. Amer. Math. Soc., 16(1), 57-67 (2003). Har62 R. Hartshorne, Complete Intersections and Connectedness, American Journal of Mathematics, Vol. 84, No. 3 (Jul., 1962), pp. 497-508 HV11 Andreas Horing and Claire Voisin. Anticanonical divisors and curve classes on Fano manifolds. Pure Appl. Math. Q., 7(4, Special Issue: In memory of Eckart Viehweg):1371-1393, 2011. Kol96 J´anos Koll´ar. Rational curves on algebraic varieties, volume 32 of Ergebnisse der Mathematik und ihrer Grenzgebiete. 3. Folge. A Series of Modern Surveys in Mathematics [Results in Mathematics and Related Areas. 3rd Series. A Series of Modern Surveys in Mathematics]. Springer-Verlag, Berlin, 1996. Kol10 J. Koll´ar. Holomorphic and pseudo-holomorphic curves on rationally connected varieties, Portu- galiae Mathematica, Volume 67, Issue 2, (2010) 155179. Voi06 C. Voisin, On integral Hodge classes on uniruled and Calabi-Yau threefolds, in Moduli Spaces and Arithmetic Geometry, Advanced Studies in Pure Mathematics 45, pp. 43-73, (2006). Voi12 C. Voisin, Remarks on curve classes on rationally connected varieties. arXiv:1201.1072. Zhu11 Y. Zhu, Fano hypersurfaces in positive characteristic. arXiv:1111.2964. 13 One cycles on rationally connected varieties Zon12a Hong R. Zong, Curve classes on rationally connected varieties. arXiv:1207.0575. Zon12b Hong R. Zong, On the Space of Conics on Complete Intersections. arXiv:1211.1946. Zhiyu Tian Department of Mathematics 253-37, California Institute of Technology, Pasadena, CA, 91125 [email protected] Hong R. Zong Department of Mathematics, Princeton University, Princeton, NJ, 08544-1000 [email protected] 14
1212.2859	2	1212	2015-11-25T16:29:34	Cyclic homology of categories of matrix factorizations	[ "math.AG", "math.CV" ]	In this paper, we will show that for a smooth quasi-projective variety over $\C,$ and a regular function $W:X\to \C,$ the periodic cyclic homology of the DG category of matrix factorizations $MF(X,W)$ is identified (unde Riemann-Hilbert correspondence) with vanishing cohomology $H^{\bullet}(X^{an},\phi_W\C_X),$ with monodromy twisted by sign. Also, Hochschild homology is identified respectively with the hypercohomology of $(\Omega_X^{\bullet},dW\wedge).$ One can show that the image of the Chern character is contained in the subspace of Hodge classes. One can formulate the Hodge conjecture stating that it is surjective ($\otimes\Q$) onto Hodge classes. For W=0 and $X$ smooth projective this is precisely the classical Hodge conjecture.	math.AG	math	CYCLIC HOMOLOGY OF CATEGORIES OF MATRIX FACTORIZATIONS ALEXANDER I. EFIMOV Abstract. In this paper, we will show that for a smooth quasi-projective variety over C, and a regular function W : X → C, the periodic cyclic homology of the DG category of matrix factorizations M F (X, W ) is identified (unde Riemann-Hilbert correspondence) with vanishing cohomology H •(X an, φW CX), with monodromy twisted by sign. Also, Hochschild homology is identified respectively with the hypercohomology of (Ω• X , dW ∧). One can show that the image of the Chern character is contained in the subspace of Hodge classes. One can formulate the Hodge conjecture stating that it is surjective ( ⊗Q ) onto Hodge classes. For W = 0 and X smooth projective this is precisely the classical Hodge conjecture. Contents 1. 2. Introduction (Curved) DG categories 2.1. DG categories 2.2. Matrix factorizations 2.3. Curved DG categories 3. Mixed complexes with a connection 3.1. Mixed complexes. 3.2. u -connections on mixed complexes. 3.3. Examples of u -connections. 3.4. Sheaves of mixed complexes with u -connections. 4. Affine case 5. General case 6. Concluding remarks References MSC: 14F05, 32S30. This work is supported by the RSF under a grant 14-50-00005. 1 2 5 5 7 8 11 11 13 14 18 23 24 27 28 2 ALEXANDER I. EFIMOV 1. Introduction It is classically known that Hochschild homology is a non-commutative analog of differen- tial forms, and in characteristic zero Connes differential corresponds to de Rham differential [HKR], [FT], [C]. The de Rham cohomology hence corresponds to periodic cyclic homology. An identification of the de Rham cohomology with periodic cyclic homology for affine smooth schemes was shown by Feigin and Tsygan [FT], and they also proved that for non- smooth affine schemes the periodic cyclic homology is identified with crystalline cohomology (which in the case of complex numbers is just the singular cohomology of the associated topological space). Weibel [W2] (see also Weibel and Geller [WG]) has shown that if one defines the cyclic homology of schemes of finite type by sheafifying the cyclic complex and then taking RΓ, one still gets the crystalline cohomology. Keller [Ke2] proved that this version of cyclic homology actually coincides with the cyclic homology of the DG category of perfect complexes on a scheme. Therefore, the periodic cyclic homology of Perf(X) is identified with the crystalline cohomology. In this paper we compute periodic cyclic homology for matrix factorization categories. More precisely, following [KKP], we consider periodic cyclic homology of a Z/2 -graded DG category as a Z/2 -graded vector bundle with connection on the formal punctured disk. We show that given a quasi-projective smooth complex algebraic variety with a regular function, the periodic cyclic homology is identified under Riemann-Hilbert correspondence with cohomology of the sheaf of vanishing cycles with the monodromy twisted by a sign. This can be considered as a generalization of a computation of the periodic cyclic homology of perfect complexes. For a scheme X of finite type over a field k, together with a regular function W ∈ O(X), we denote by M Fcoh(X, W ) the Z/2 -graded DG category of matrix factorizations, defined in [EP] in a more general context of quasi-coherent sheaves of curved DG algebras. Further, for any small Z/2 -graded DG category T over k we recall the Hochschild homology HH•(T ) := H −•(IdT L ⊗T ⊗T op IdT ), where IdT is the diagonal DG bimodule. It can be computed by the standard Hochschild complex (Hoch(T ), b), obtained from the bar resolution of one of the copies of IdT . There is a well-known Connes differential B on Hoch(T ), satisfying It allows one to define the periodic cyclic homology Bb + bB = 0. HP•(T ) := H •(Hoch((u)), b + uB), CYCLIC HOMOLOGY OF CATEGORIES OF MATRIX FACTORIZATIONS 3 as well as ordinary cyclic homology HC•(T ) and negative cyclic homology HC − we will not consider. • (T ), which Periodic cyclic homology HP•(T ) can be viewed as a vector bundle over the formal introduced in [KKP] and punctured disk Spf k((u)), with a natural connection ∇GM written down explicitly in [Sh1]. u . Given a finite-dimensional complex vector space V with an automorphism T : V → V, choose any M : V → V such that and write exp(−2πiM ) = T, −1 (E, T ) := (V ((u)), d + M du u ). dRH This is a bundle with connection on Spf C((u)). Our main result is the following: Theorem 1.1. Let X be a smooth quasi-projective algebraic variety over C, and W : X → C a regular function. Then we have an identification of Z/2 -graded vector bundles with connection on the formal punctured disk: (HP•(M Fcoh(X, W )), ∇GM u −1 (H •−1 an (W −1(0), φW CX), T · (−1)•), ) ∼= dRH where T is the monodromy automorphism. There is an algebraic formula for the vanishing cohomology with monodromy, conjectured by Kontsevich and proved by Sabbah [Sab2] (see also the paper by Sabbah and M. Saito [SS]). We formulate its special case needed for our purposes. Theorem 1.2. ([Sab2]) Let X be a smooth quasi-projective algebraic variety over C, and W : X → C a regular function. Assume that the critical locus of W is contained in W −1(0). Then we have an isomorphism of Z -graded bundles with connections on the formal punctured disk: (H • Zar(X, (Ω• X ((u)), −dW + ud)), ∇u = d du + W u2 ) ∼= dRH −1 (H •−1 an (W −1(0), φW CX), T ). Our Theorem 1.1 is actually obtained as a combination of Sabbah Theorem and the following result valid over arbitrary field of characteristic zero. Theorem 1.3. Let X be a smooth separated scheme of finite type over a field k of char- acteristic zero, and W ∈ O(X) a regular function. Assume that the critical locus of W is contained in W −1(0). Then we have natural isomorphisms of Z/2 -graded bundles with connections (HP•(M Fcoh(X, W )), ∇GM u ) ∼= (H •(Ω• X((u)), −dW + ud), ∇DR u = d du + Γ u + W u2 ), 4 ALEXANDER I. EFIMOV where ΓΩp := − p 2 · id, and an isomorphism of Z/2 -graded vector spaces HH•(M Fcoh(X, W )) ∼= H •(X, (Ω• X , −dW )). Here by dW one means the operator of multiplication by dW. We remark that for Hochschild cohomology it was shown in [LP] that HH •(M Fcoh(X, W )) ∼= H •(X, (Λ•TX, ιdW )), again under assumption that the critical locus is in the fiber over zero. Theorem 1.3 is implied by a stronger result on mixed complexes with u -connections (see Section 3 for the precise definitions). Namely, a ( 2 -periodic) mixed complex is just a Z/2 - graded vector space C with a pair of anti-commuting differentials (b, B). The morphism of such mixed complexes is defined in the obvious way, and it is said to be a quasi-isomorphism if it induces isomorphisms on b -cohomology. A u -connection on a mixed complex is defined in the natural way so that it induces a connection on H •(C((u)), b + uB). There is a natural notion of weak and strict morphisms of mixed complexes with u - connections. Also, for a sheaf of mixed complexes with u -connections on a scheme X of finite type over a field, there is a functor RΓ(X, −) with values in mixed complexes with u -connections. We have the following result. Theorem 1.4. Assume that the assumptions of Theorem 1.3 are satisfied. Then we have a chain of quasi-isomorphisms between mixed complexes with u -connections: (Hoch(M Fcoh(X, W )), b, B, ∇GM u ), RΓ(X, (Ω• X , −dW, d, ∇DR u )). The special case of this result for an affine space with a polynomial with isolated singu- larity has already been shown by Shklyarov [Sh1]. The paper is organized as follows. In Section 2 we recall the DG categories and curved DG categories over a field, and define the DG categories of matrix factorizations. Section 3 is devoted to mixed complexes with u -connections, already mentioned above. First, in Subsection 3.1 we introduce mixed complexes, morphisms and quasi-isomorphisms between them. Also, we give examples of Hochschild complexes (of the second kind) of small (curved) DG categories and twisted de Rham mixed complex. Then, in Subsection 3.2 we introduce u -connections of mixed complexes, and also the notions of weak and strict morphisms between mixed complexes with u -connections. Next, in Subsection 3.3 we recall the u -connections on the Hochschild complexes (of the second kind) and for CYCLIC HOMOLOGY OF CATEGORIES OF MATRIX FACTORIZATIONS 5 twisted de Rham mixed complex. For Hochschild complexes of the second kind it was not written explicitly before, but it is obtained naturally e.g. by looking at the general formula for A∞ -algebras, and adding a term corresponding to the m0, which corresponds In Subsection 3.4 we study (pre)sheaves of mixed complexes with u - to the curvature. connections on a scheme of finite type over a field. Such (pre)sheaves arise for example from presheaves of (curved) DG categories. Here we construct in particular the derived global section functor with values in mixed complexes with u -connections, which preserves weak quasi-isomorphisms. In Section 4 we prove Theorem 1.4 in the affine case: X = Spec A. It uses the results of Polishchuk and Positselski on Hochschild complexes of the second kind, and the Hochschild- Kostant-Rosenberg map [HKR] from Hochschild complex of the second kind of the CDG algebra (A, W ) to the twisted de Rham mixed complex. The special case when A is the polynomial ring, and W has isolated singularity at the origin, was done by Shklyarov [Sh1]. In Section 5 we prove the general case of Theorem 1.4, using sheafification argument. Actually, the idea of sheafification for Hochschild complexes of DG categories is due to Keller [Ke2], who used this for perfect complexes. Our argument is very similar to Keller’s one, although the chain of quasi-isomorphisms occures to be much longer. We also show here that Theorem 1.4 implies Theorem 1.3, which in turn implies Theorem 1.1. In Section 6 we write down some remarks. Acknowledgements. I am grateful to L. Positselski and V. Vologodsky for useful discus- sions. 2. (Curved) DG categories In this section we recall the notion of (curved) DG categories and (curved) DG functors, and also define DG categories of matrix factorizations. We fix some basic field k. 2.1. DG categories. Our basic reference for DG categories is [Ke1]. The DG quotients were introduced in [Ke4], [Dr], and we give the explicit construction from [Dr]. Definition 2.1. 1) A Z/2 -graded DG category is a category, for which the sets of mor- phisms Hom(X, Y ) are Z/2 -graded complexes of vector spaces, the composition maps Hom(Y, Z) ⊗ Hom(X, Y ) → Hom(X, Z) are morphisms of complexes, and the identity morphisms are closed of degree zero. 2) A DG functor F : T → T ′ between DG categories is a functor such that the maps F (X, Y ) : Hom(X, Y ) → Hom(F (X), F (Y )) are morphisms of complexes. 6 ALEXANDER I. EFIMOV The basic example is the DG category Com k of complexes of k -vector spaces., with Hom’s being Hom-complexes. DG algebra is a DG category with one object. A morphism of DG algebras is a DG functor of the associated DG categories. To any DG category T we can associate a graded k -linear homotopy category Ho•(T ), with HomHo•(T )(X, Y ) := H •(HomT (X, Y )). Definition 2.2. A DG functor F : T → T ′ functor is called a quasi-equivalence if the induced Ho•(F ) : Ho•(T ) → Ho•(T ′) is an equivalence of graded categories. Recall that the right DG module over a small DG category T is a DG functor T op → Com k. One gets the derived category D(T ) of such DG modules, obtained by localizing with respect to quasi-isomorphisms. There is a more general notion of Morita equivalence. Definition 2.3. A DG functor F : T → T ′ between small DG categories is said to be a Morita equivalence if the restriction functor D(T ′) → D(T ) is an equivalence. We will need the notion of a DG quotient for DG categories. Definition 2.4. ([Dr]) Let S ⊂ T be a full DG subcategory of a small DG category T. Define the Drinfeld DG quotient T /S as follows. This is a DG category with Ob(T /S) = Ob(T ), and it is obtained from T by formally adding odd morphisms h(X) : X → X such that That is, we have an identification of graded vector spaces d(h(X)) = idX . HomT /S(X, Y ) = HomT (X, Y )⊕ M n≥1, Z1,...,Zn∈S HomT (X, Z1)⊗HomT (Z1, Z2)⊗· · ·⊗HomT (Zn, Y )[n], and the differential comes from differentials on Hom ’s in T and from the condition d(h(Zi)) = idZi . CYCLIC HOMOLOGY OF CATEGORIES OF MATRIX FACTORIZATIONS 7 2.2. Matrix factorizations. Now we define the DG category of matrix factorizations. Let X be a separated scheme of finite type over k, and W ∈ O(X) a regular function. Define the following Z/2 -graded DG categories. First, the category M F nv(X, W ) has as objects pairs (E, δ) of a Z/2 -graded locally free sheaf E = E0 ⊕ E1 on X and an odd map δ : E → E such that δ2 = W · idE . The complexes of morphisms are defined by taking standard Hom -complexes (the differ- ential squares to zero). Second, the category M F nv coh(X, W ) is defined in the same way but the objects are pairs (E, δ), where E is a Z/2 -graded coherent sheaf on X. In particular, we have a fully faithful DG functor M F nv(X, W ) → M F nv coh(X, W ). coh(X, W ) ⊂ M Fcoh(X, W ) is defined to consist of convolutions of exact triples of coherent matrix factorizations. Following [EP], [Or2], [PV], Further, the full DG subcategory M F ex [LP], we define the category of coherent matrix factorizations as a DG quotient M Fcoh(X, W ) := M F nv coh(X, W )/M F ex coh. In particular, we have a natural DG functor (2.1) M F nv(X, W ) → M Fcoh(X, W ). Proposition 2.5. ([EP]) Let X be affine and smooth over k. Then the DG functor (2.1) is a quasi-equivalence. Theorem 2.6. ([Or1],[EP]) Assume that X is smooth, and W is non-zero on each con- nected component. Then there is an equivalence of triangulated categories Ho(M Fcoh(X, W )) ∼= Dsg(W −1(0)), where Dsg(Y ) := Db coh(Y )/ Perf(Y ). It can be easily seen in a number of ways that for any open subset U ⊂ X containing W −1(0) the restriction functor M Fcoh(X, W ) → M Fcoh(U, W ) is a quasi-equivalence. In particular, we may (and will if it is convenient) assume that the critical locus of W is contained in W −1(0). 8 ALEXANDER I. EFIMOV 2.3. Curved DG categories. Our basic reference for curved DG categories is [PP]. Definition 2.7. A Z/2 -graded curved DG category (CDG category) is a category D, for which the sets of morphisms are Z/2 -graded vector spaces equipped with odd maps d : Hom(X, Y ) → Hom(X, Y ) (the ”differentials”), and for each object there is a distinguished even morphism hX : X → X (a curvature). This data is required to satisfy the following properties: 1) We have for any composable homogeneous morphisms f, g; d(f g) = (df )g + (−1)f f dg 2) We have for any f : X → Y ; 3) We have for any object X; d2(f ) = hY f − f hX dhX = 0 3) The identity morphisms idX are of degree zero (this automatically implies d idX = 0 ). CDG algebra is a CDG category with one object. In particular, any DG category can be treated as a CDG category with hX = 0. Moreover, for any CDG category, its full sub(CDG)category consisting of objects with zero curvature is actually a DG category. There is a natural notion of an opposite CDG category Dop. Namely, Ob(Dop) = OB(D), HomC op(X op, Y op) = HomC(Y, X), dop = d, hX op = −hX, and the composition changes by a sign: f opgop = (−1)f g(gf )op. We have the following basic example: Its objects are Z/2 -graded vector spaces X equipped with an odd map dX : X → X. Further, Hom((X, dX ), (Y, dY )). is the graded space of morphisms. It is equipped with a ”differen- tial” the CDG category Pre(k-mod). d(f ) = dY f − (−1)f f dX. The composition is obvious, and the curvature of (X, dX ) equals d2 X . We have a notion of a CDG functor CYCLIC HOMOLOGY OF CATEGORIES OF MATRIX FACTORIZATIONS 9 Definition 2.8. A curved DG functor (F, α) : D → D′ (CDG functor) between CDG categories is a functor together with distinguished odd morphisms αX : F (X) → F (X) such that 1) The maps F (X, Y ) : Hom(X, Y ) → Hom(F (X), F (Y )) are maps of graded vector spaces; 2) We have F (df ) = d(F (f )) + αY F (f ) − (−1)f F (f )αX. for any homogeneous f ∈ HomD(X, Y ); 3) We have F (hX ) = hF (X) + dαX + α2 X . A morphism of CDG algebras is a CDG functor of the associated CDG categories. It is easy to see that CDG functors from a small CDG category D to any CDG category D′ form a DG category: the space Hom((F, α), (G, β)) is the graded space of morphisms between graded functors, the differential is given by the formula d(f )X = d(fX ) + βX f − (−1)f f αX for a homogeneous morphism f : (F, α) → (G, β), and the composition is obvious. We will call a CDG functor strict if the morphisms αX equal to zero. In particular, any DG functor between DG categories can be treated as a strict CDG functor between associated CDG categories. Definition 2.9. A (left) CDG module over a CDG category D is a strict CDG functor F : D → Pre(k-mod). Left CDG-modules over D form a DG category D − modcdg. We will also need a notion of a QDG functor. Definition 2.10. A QDG functor F : D → D′ between CDG categories is given by the same data as a CDG functor, but we do not require the condition 3) of Definition 2.8 to hold (the relation with curvatures). The QDG functors form a CDG category: the morphisms and the ”differentials” are defined in the same way as for CDG functors, and the curvatures are given by the formula (hF )X := hF (X) + dαX + α2 X − F (hX ). 10 ALEXANDER I. EFIMOV It is clear that the DG category of CDG functors is a full CDG subcategory of the CDG category of QDG functors, consisting of QDG functors with zero curvature. Again, we call a QDG functor strict, if αX = 0. Definition 2.11. A (left) QDG module over a CDG category D is a strict QDG functor F : D → Pre(k-mod). Left QDG-modules over D form a CDG category D − modqdg. In particular, we have Pre(k-mod) ∼= k-modqdg. There is no reasonable notion of a weak equivalence for CDG categories (like e.g. quasi- equivalences or Morita equivalences of DG categories). However, we will need a notion of a pseudo-equivalence [PP], which is quite useful for our purposes. First, define the notion of a twist of an object. Let X ∈ D be an object of a CDG category, and τ : X → X an odd endomorphism. Then an object Y ∈ D is called a twist of an object X by τ if there exist even morphisms i : X → Y, j : Y → X such that ij = idY , ji = idX, jd(i) = τ, jhY i = hX + dτ + τ 2. The shift X[1] of X is defined as an object Y equipped with odd morphisms i : X → Y, j : Y → X such that ij = idY , ji = idX, d(i) = 0, d(j) = 0. Further, given a family of objects Xα ∈ D, there direct sum is an object X, which is equipped with a structure of a direct sum of objects Xα in the graded category Dgr, and the structural even morphisms iα : Xα → X satisfy diα = 0. Definition 2.12. A CDG functor F : D → D′ is called a pseudo-equivalence if it is fully faithful (as a functor between graded categories) and each object of D′ can be obtained from the image of F using direct sums, direct summands, shifts and twists. Take any small CDG category D. Then we have full (C)DG subcategories D-modcdg f gp ⊂ D-modcdg, D-modqdg f gp ⊂ D-modqdg consisting of CDG (resp. QDG) modules which are finitely generated projective as graded D -modules. Note that we have an Yoneda strict CDG functor Dop → D-modqdg f gp. Proposition 2.13. ([PP])The strict CDG functors D-modcdg f gp → D-modqdg f gp ← Dop CYCLIC HOMOLOGY OF CATEGORIES OF MATRIX FACTORIZATIONS 11 are pseudo-equivalences. For any commutative k -algebra A of finite type, and any W ∈ A, we have a CDG algebra (A, W ) concentrated in even degree, with zero ”differential” and curvature W. Then there is a natural identification of DG categories (A, W )-modcdg f gp ∼= M F nv(Spec A, W ). 3. Mixed complexes with a connection We will consider Z/2 -graded complexes. Again, we fix some basic field k. 3.1. Mixed complexes. Definition 3.1. 1) A mixed complex is a triple (C, b, B), where C is a Z/2 -graded vector space, b and B are odd differentials on C satisfying 2) A morphism of mixed complexes bB + Bb = 0. f : (C, b, B) → (C ′, b′, B′) is a morphism f : C → C ′ of graded vector spaces (homogeneous of degree zero), such that b′f = f b, B′f = f B. 3) A morphism f : (C, b, B) → (C ′, b′, B′) is called a quasi-isomorphism if it induces a quasi-isomorphism of complexes (C, b) → (C ′, b′). It is convenient to think of mixed complexes as DG modules over khBi/B2, where B is an odd variable. We will consider the following examples of mixed complexes. Example: Hochschild complexes. Let T be a small Z/2 -graded DG category over a field k. Define the DG category T e by formally adding a (closed) identity morphism eX for each object X ∈ T. That is, Ob(T e) := Ob(T ), HomT e(X, Y ) :=   and the composition is obvious. HomT (X, Y ) if X 6= Y ; HomT (X, X) ⊕ k · eX if X = Y, 12 ALEXANDER I. EFIMOV For a Z/2 -graded vector space V, denote by sV the same space with grading shifted by 1 (reversed grading), and for a homogeneous element v ∈ V denote by sv the corresponding element of sV. Now, put HomT (X, X)⊕ Hoch(T ) := M M X∈Ob(T ) n≥1, HomT e(Xn, X0) ⊗ s HomT (Xn−1, Xn) · · · ⊗ s HomT (X0, X1). X0,...,Xn∈Ob(T ) We will write (fn, fn−1, . . . , f0) for fn ⊗ sfn−1 ⊗ . . . sf0. The Hochschild differential b is defined by the formula b = bm2 + bm1, where bm2(f0 ⊗ f1 ⊗ . . . fn) := n P sfk+1 (−1) k=i+1 n−1X i=0 (fn, . . . , fi+1fi, . . . f0)+ sf0(fn+ (−1) n−1 P k=1 sfk) (f0fn, fn−1, . . . , f1), and bm1(fn, fn−1, . . . , f0) = n P sfk (−1) k=i+1 (fn, . . . , dfi, . . . , f0). nX i=0 Further, the Connes differential B is defined by the formula It is straightforward to check that (Hoch(T ), b, B) is a mixed complex. Example: Hochschild complexes of the second kind. Suppose that D is a Z/2 -graded curved small DG category (for instance, D can be just a DG category, with zero curvatures). Again, we consider the CDG category De defined in the same way as above. Define the graded vector space HochΠ(D) by the formula HochΠ(D) := M X∈Ob(D) ( M Y n≥1 X0,...,Xn∈Ob(D) HomD(X, X)⊕ HomDe(Xn, X0) ⊗ s HomlD(Xn−1, Xn) ⊗ · · · ⊗ s HomD(X0, X1)). The Hochschild differential is defined by the formula b = bm2 + bm1 + bm0, B(fn, fn−1, . . . , f0) := i−1 P ( (−1) k=0 nPi=0 sfk)( n P l=i sfl) 0 (eXi, fi−1, . . . f0, fn, . . . fi) if fn ∈ HomT (Xn, X0); if fn ∈ k · eX0.   CYCLIC HOMOLOGY OF CATEGORIES OF MATRIX FACTORIZATIONS 13 where bm1 and bm2 are as above, and bm0(fn, fn−1, . . . f0) := (−1) nX i=0 n P k=i sfk (fn, . . . , fi, hXi, fi−1, . . . , f0). The Connes differential B : HochΠ(D) → HochΠ(D) is defined by the same formula as above. Remark 3.2. One can show (by an easy spectral sequence argument) that if we would take direct sums instead of direct products in the definition of HochΠ(D), and at least one curvature hX is non-zero, then the resulting complex would be acyclic (see [PP]). Example: Twisted de Rham mixed complex. Let A be a commutative smooth k -algebra of finite type. Then is a mixed complex. (Ω•(A), −dW ∧, dDR) 3.2. u -connections on mixed complexes. Take a formal even variable u. Starting from a mixed complex (C, b, B), we can form a new complex (C((u)), b + uB), where C((u)) = colimn u−n · C[[u]], C[[u]] = Y m≥0 um · C. Proposition 3.3. Any morphism of mixed complexes f : (C, b, B) → (C ′, b′, B′) induces a morphism of complexes f ((u)) : (C((u)), b + uB) → (C ′((u)), b′ + uB′). Moreover, if f is a quasi-isomorphism, then so is f ((u)). Proof. The first statement is obvious, and the second follows from the spectral sequence argument. (cid:3) Definition 3.4. A u -connection on a mixed complex (C, b, B) is a k -linear operator which has the form ∇u : C((u)) → C((u)), ∇u = d du + A(u), where A(u) is a k((u)) -linear operator from C((u)) to itself, and, moreover, [∇u, b + uB] = 1 2u (b + uB). We will consider morphisms compatible with u -connection. 14 ALEXANDER I. EFIMOV Definition 3.5. A morphism f : (C, b, B, ∇u) → (C ′, b′, B′, ∇′ u) of mixed complexes is 1) strictly compatible with u -connections if ∇′ uf ((u)) = f ((u))∇u. In this case we will call f a strict morphism; 2) weakly compatible with u -connections if the k((u)) -linear operator (∇′ uf ((u)) − f ((u))∇u) : C((u)) → C ′((u)) is u -homotopic to zero. That is, there exists a k((u)) -linear operator H : C((u)) → C ′((u)), such that ∇′ uf ((u)) − f ((u))∇u = (b′ + uB′)H + H(b + uB). In this case we will call a pair (f, H) a weak morphism. 3.3. Examples of u -connections. Example: Hochschild complexes. Let T be a Z/2 -graded DG category. Following Shklyarov, we introduce the following u -connection on the mixed complex Hoch(T ) : where ∇u := d du + Um1 u2 + Vm1 + Γ u , Γ(fn, . . . , f0) := −n 2 (fn, . . . , f0); Um1(fn, . . . , f0) := n P k=1 (−1) sfk+f0(fn+ n−1 P k=1 sfk) ((df0)fn, fn−1, . . . , f1); (−1)ǫij (eXj+1 , fj, . . . , dfi, . . . , f0, fn, . . . , fj+1) if fn ∈ HomT (Xn, X0); if fn ∈ k · eX0, 1 2 1 2 P0≤i≤j≤n−1   Vm1(fn, . . . , f0) := where 0 ǫij = nX k=i+1 sfk + ( jX k=0 sfk + 1)( nX l=j+1 sfl) It was checked by Shklyarov ([Sh1], section C.2) that ∇u is indeed a u -connection. Remark 3.6. This u -connection was actually introduced in [KKP] (without explicit for- mula). Namely, consider a deformation Tt of T over Gm : objects, spaces of morphisms and compositions do not change, but the differentials change by the formula dt = t · d CYCLIC HOMOLOGY OF CATEGORIES OF MATRIX FACTORIZATIONS 15 (this deformation is naturally isomorphic to the one in [KKP], the isomorphism map is multiplication by t−1 on morphisms). Then we have a Z/2 -graded bundle over Gm × Spf k((u)) : E = H •(Hoch(Tt)((u)), bt + uBt). It is equipped with Getzler-Gauss-Manin [G] connection ∇GM in the t -direction. Note that we have a natural identification of mixed complexes (Hoch(Tt), bt, Bt) → (C, tb, t−1B), (fn, . . . , f0) 7→ t−n(fn, . . . , f0). This equips E with a Gm -equivariant structure with respect to the action µ · (t, u) = (µt, µ2u). Let Λ : E → E be an operator of differentiation by d idE . Therefore, the differential operator dµ . Then Λ has symbol (t ∂ ∂t + 2u ∂ ∂u ) · Λ 2u − ∇GM t 2u ∂ ∂t has symbol ∂ ∂u . Its restriction to the fiber t = 1 is precisely the u -connection described above. Below we will always write Hoch(T ) for the mixed Hochschild complex with u - connection defined above. Proposition 3.7. A DG functor F : T → T ′ induces a morphism of mixed complexes F∗ : Hoch(T ) → Hoch(T ′), strictly compatible with a u -connections. Moreover, if F is a Morita equivalence, then F∗ is a quasi-isomorphism. Proof. The first statement is obvious, and the second one is due to Keller [Ke3]. (cid:3) We will need one more result here (due to Keller). Definition 3.8. A sequence of small DG categories and DG functors T ′ → T → T ′′ said to be exact up to Morita equivalence, if the image of the composition consists of objects is homotopic to zero. and the choice of contracting homotopies defines a Morita equivalence T /T ′ → T ′′ Theorem 3.9. ([Ke3]) Given an exact sequence of small DG categories T ′ → T → T ′′, one has a natural exact triangle of Hochschild complexes Hoch(T ′) → Hoch(T ) → Hoch(T ′′) → Hoch(T ′)[1]. 16 ALEXANDER I. EFIMOV Example: Hochschild complexes of the second kind. Let D be a Z/2 -graded CDG category. Then we define a u -connection on the mixed complex HochΠ(D) by the formula u2 where Γ, Um1 and Um2 are as above, and 2Um0 + Um1 + 2Vm0 + Vm1 + Γ u , ∇u := d du +   Vm0(fn, . . . , f0) := where Um0(fn, . . . , f0) := (hX0 fn, fn−1, . . . , f0); P0≤i≤j≤n (−1)ηij (eXj , fj−1, . . . , hXi , . . . , f0, fn, . . . , fj) 0 ηij = nX k=i sfk + ( j−1X k=0 sfk + 1)( nX l=j sfl) + 1 if fn ∈ HomT (Xn, X0); if fn ∈ k · eX0 , One can check that this indeed defines a u -connection (just analogously to the case of DG categories). A special case of polynomial algebra with zero differential and some curvature was verified in [Sh1], section D.3. Below we will write HochΠ(D) for the mixed Hochschild complex of the second kind with u -connection defined above. Proposition 3.10. ([PP]) A strict CDG functor F : D → D′ induces a morphism of mixed complexes F∗ : HochΠ(D) → HochΠ(D′), strictly compatible with u -connections. Moreover, if F is a pseudo-equivalence, then F∗ is a quasi-isomorphism. One can also show that non-strict CDG functor induces a morphism of mixed complexes weakly compatible with u -connections, but we do not need it. Proposition 3.11. Let T be a Z/2 -graded DG category. Then the natural map of mixed complexes is strictly compatible with u -connections. Hoch(T ) → HochΠ(T ) Proof. This follows immediately from the definitions. (cid:3) We need the following result, due to Polishchuk and Positselski (they consider just Hochschild complexes without additional data). Proposition 3.12. ([PP]) Let X = Spec A be affine smooth algebraic variety over k, where chark = 0, and W ∈ O(X) Assume that the critical locus of W is in the fiber W −1(0). Then the morphism of mixed complexes with u -connections Hoch(M F nv(X, W )) → HochΠ(M F nv(X, W )) is a quasi-isomorphism. CYCLIC HOMOLOGY OF CATEGORIES OF MATRIX FACTORIZATIONS 17 We will need one more fact. Proposition 3.13. For any small curved CDG category D, we have a natural weak quasi- isomorphism of mixed complexes with u -connections: (Φ, H) : HochΠ(D) → HochΠ(Dop), bop, Bop). Proof. The proof is a generalization of the proof of Proposition 3.8 of [Sh1]. First, the morphism Φ is given by the formula Φ(fn, fn−1, . . . , f0) := (−1) 0≤i<j≤n−1 n+ P sfisfj (f op n , f op 0 , . . . , f op n−1). It is easily seen to commute with Hochschild and Connes differentials. Moreover, it is obviously a quasi-isomorphism. Further, the odd morphism H = Hm1 + Hm0 : HochΠ(D)((u)) → HochΠ(Dop)((u)) (−1)ǫi 2u2 n−1Pi=0 (df op i , f op i+1, . . . , f op n , f op 0 , . . . , f op i−1) if fn ∈ HomD(Xn, X0); if fn ∈ k · eX0; if fn ∈ HomD(Xn, X0); if fn ∈ k · eX0, is given by the formulas Hm1(fn, . . . , f0) := Hm0(fn, . . . , f0) := where 0     nPi=0 (hX op i (−1)ηi u2 0 , f op i , . . . , f op n , f op 0 , . . . , f op i−1) ηi = X 0≤k<l≤i−1 sfksfl + X i≤k<l≤n sfksfl, ǫi = ηi + 1 for 0 ≤ i ≤ n − 1. It is straightforward to check that the pair (f, H) defines a weak morphism. (cid:3) Example: Twisted de Rham mixed complex. If A is a smooth commutative k -algebra of finite type, we define a u -connection on the mixed complex (Ω•(A), −dW ∧, d) by the formula ∇DR u := W u2 , d du + + Γ u where ΓΩp(A) := − p 2 · id . 18 ALEXANDER I. EFIMOV Proposition 3.14. Let A be as above. Then the Hochschild-Kostant-Rosenberg map defined by the formula ε : HochΠ(A, W ) → (Ω•(A), −dW ∧, d), ε(a0, . . . , an) := 1 n! a0da1 ∧ · · · ∧ dan, is a quasi-isomorphism of mixed complexes, weakly compatible with connections. Moreover, the contracting homotopy is given by the formula H = − W d 2u ε. Proof. It was checked in [Sh1], section D.5, that the pair (ε, H) defines a weak morphism for the special case of a polynomial algebra A with W having isolated singularity, but the checking is the same in the general case. Further, the fact that ε is a quasi-isomorphism follows from the classical Hochschild-Kostant-Rosenberg theorem [HKR] and an obvious spectral sequence argument, see [Seg]. (cid:3) 3.4. Sheaves of mixed complexes with u -connections. Let X be a scheme of finite type over a field k. For any presheaf of k -vector spaces F on X we define a presheaf F((u)) of k((u)) -vector spaces by the formula F((u))(U ) := F(U )((u)). Lemma 3.15. If F is a sheaf on X, then so is F((u)). The functor F 7→ F((u)) is exact on the category of sheaves of k -vector spaces. Proof. The first statement is an easy consequence of the fact that any open subset of X is quasi-compact. Indeed, because of this fact we have to check that for any finite collection of open subsets U1, . . . , Un ⊂ X there is an exact sequence 0 → F( n[ i=1 Ui)((u)) → nY i=1 F(Ui)((u)) → Y 1≤i<j≤n F(Ui ∩ Uj)((u)). But this follows from the corresponding exact sequence for the sheaf F. The second statement follows from the first one. (cid:3) Definition 3.16. 1) A (pre)sheaf of mixed complexes with a u -connection on X is a (pre)sheaf on X (with Zariski topology) with values in the category of mixed complexes with u -connections and morphisms strictly compatible with u -connections. That is, to define such a (pre)sheaf (C, b, B, ∇u) we need to define for each open sub- set U ⊂ X a mixed complex with u -connection (C(U ), b(U ), B(U ), ∇u(U )), and for any CYCLIC HOMOLOGY OF CATEGORIES OF MATRIX FACTORIZATIONS 19 inclusion of open subsets V ⊂ U a map of mixed complexes strictly compatible with u - connections fU V : (C(U ), b(U ), B(U ), ∇u(U )) → (C(V ), b(V ), B(V ), ∇u(V )), which satisfy the axioms of a (pre)sheaf. 2) A strict morphism of (pre)sheaves of mixed complexes with u -connections f : (C, b, B, ∇u) → (C ′, b′, B′, ∇u ′) is an assignment to each open subset U ⊂ X a morphism f (U ) : (C(U ), b(U ), B(U ), ∇u(U )) → (C ′(U ), b′(U ), B′(U ), ∇u ′(U )) of mixed complexes strictly compatible with u -connections, such that for any V ⊂ U we get a strictly commutative square. 3) A weak morphism of (pre)sheaves of mixed complexes with u -connections (f, H) : (C, b, B, ∇u) → (C′, b′, B′, ∇u ′) is a pair of a morphism f : (C, b, B) → (C ′, b′, B′) of (pre)sheaves of mixed complexes and an odd k((u)) -linear morphism H : C((u)) → C ′((u)) of Z/2 -graded (pre)sheaves of k((u)) -vector spaces, such that for any open subset U ⊂ X we have ∇u ′(U )f (U )((u)) − f (U )((u))∇u(U ) = (b′(U ) + uB′(U ))H + H(b(U ) + uB(U )). In the case of sheaves, such a morphism is called a quasi-isomorphism if the morphism is a quasi-isomorphism of complexes of (pre)sheaves of vector spaces. f : (C, b) → (C ′, b′) Note that a strict morphism f can be treated as a weak morphism (f, 0). Clearly, the standard sheafification construction works for mixed complexes with u - connections. We will denote by Csh the sheafification of C. We have an obvious sufficient condition for a morphism of mixed complexes with u - connections to be a quasi-isomorphism. 20 ALEXANDER I. EFIMOV Proposition 3.17. If a weak morphism of presheaves (f, H) : (C, b, B, ∇u) → (C′, b′, B′, ∇u ′) induces quasi-isomorphisms f (U ) : (C(U ), b(U )) → (C ′(U ), b′(U )) for each open affine subset U ⊂ X, then (f sh, H sh) : (C sh, b, B, ∇u) → (C ′sh, b′, B′, ∇u ′) is a quasi-isomorphism. Proof. Indeed, open affine subsets of X form a base of Zariski topology, which implies the result. (cid:3) We will be mostly interested in sheaves of mixed complexes coming from presheaves of (curved) DG categories. Definition 3.18. 1) A presheaf T of (curved) DG categories on X is an assignment to each open subset U ⊂ X of a small Z/2 -graded (curved) DG category T (U ), and to each inclusion V ⊂ U a DG (resp. strict CDG) functor FU V : T (U ) → T (V ) such that for U1 ⊃ U2 ⊃ U3 we have an equality of (C)DG functors FU1U3 = FU2U3FU1U2. 2) A morphism of presheaves of (C)DG categories F : T → T ′ is a collection of DG (resp. strict CDG) functors F (U ) : T (U ) → T ′(U ) such that for any inclusion V ⊂ U we get a strictly commutative square. Starting from a presheaf of (C)DG categories T on X, we obtain a sheaf of mixed complexes with u -connections Hoch(T ) (resp. HochΠ(T ) ) associated with a presheaf U 7→ Hoch(T (U )) (resp. U 7→ HochΠ(T (U ))). Proposition 3.19. 1) Let F : T → T ′ be a morphism of presheaves of DG categories on X. Then it induces a strict morphism of sheaves of mixed complexes with u -connections F∗ : Hoch(T ) → Hoch(T ′). Moreover, if for any open affine subset U ⊂ X the DG functor F (U ) is a quasi-equivalence, then F∗ is a quasi-isomorphism. 2) Let G : D → D′ be a morphism of presheaves of CDG categories on X. Then it induces a strict morphism of sheaves of mixed complexes with u -connections G∗ : HochΠ(D) → HochΠ(D′). CYCLIC HOMOLOGY OF CATEGORIES OF MATRIX FACTORIZATIONS 21 Moreover, if for any open affine subset U ⊂ X the CDG functor G(U ) is a pseudo- equivalence, then G∗ is a quasi-isomorphism. 3) For any presheaf of DG categories T on X we have a strict morphism of sheaves of mixed complexes with u -connections Hoch(T ) → HochΠ(T ). Proof. The strict morphisms are defined in the obvious way, and the quasi-isomorphisms follow from Proposition 3.17, Proposition 3.7 and Proposition 3.10. (cid:3) We will need also a sheaf of de Rham mixed complexes with u -connections. Definition 3.20. Assume that X is smooth and chark = 0, and W ∈ O(X) is a regular function. Then we have a (twisted de Rham) sheaf Γ u of mixed complexes with u -connections, where ΓΩp = −p 2 · id is as above. X, −dW ∧, d, ∇DR W u2 ) d du (Ω• u = + + Proposition 3.21. Let X be smooth and chark = 0. Suppose that we have a function W ∈ O(X), and consider the presheaf (OX , W ) of CDG algebras. Then the maps ε and H from Proposition 3.14 define a weak morphism of mixed com- plexes with u -connections (ε, H) : HochΠ(OX , W ), b, B, ∇u) → (Ω• X, −dW ∧, d, ∇u = d du + Γ u + W u2 ), which is a quasi-isomorphism. Proof. Indeed, this follows from Proposition 3.17 and Proposition 3.14. (cid:3) Finally we would like to define the derived functor of global sections for sheaves of mixed complexes with u -connections. For any sheaf F of k -vector spaces on X we have a flasque Godement resolution where and F → G(F), G0(F)(U ) = Gode(F)(U ) := Y x∈U Fx, Gi := Gode(Coker(Gi−2(F) → Gi−1(F))), i > 0, where we put G−1(F) := F. Suppose that the scheme X has dimension d. Then we can apply the canonical trunca- tion and get the truncated Godement resolution by acyclic sheaves G≤d(F) := τ≤dG(F). 22 ALEXANDER I. EFIMOV Note that both Godement and truncated Godement resolutions are functorial. This allows us to apply them for mixed complexes with u -connections. Definition 3.22. Let X be a separated scheme of finite type over k, and dim X = d. Take any sheaf of mixed complexes with connections C = (C, b, B, ∇u). Applying to it truncated Godement resolution and taking the total complex, we get a new sheaf of mixed complex with u -connections G≤d(C). We put This is a mixed complex with u -connection. RΓ(X, C) := Γ(X, G≤d(C)). Clearly, RΓ is functorial with respect to weak and strict morphisms. We need the following basic result. Proposition 3.23. The functor RΓ preserves weak quasi-isomorphisms. Proof. It suffices to prove that for any acyclic Z/2 -graded complex of sheaves of vector spaces C the complex Γ(X, G≤d(C)) is acyclic. To show this, note that by induction the complex Gn(C) is acyclic for all n ≥ 0. Hence, the complex G≤d(C) is acyclic. But it consists of acyclic sheaves. This implies that Γ(X, G≤d(C)) is also acyclic. (cid:3) Proposition 3.24. Let (C, b, B, ∇u) be a sheaf of mixed complexes with u -connections. Then we have an isomorphism of bundles with connection on the formal punctured disk: (H •(RΓ(X, (C )((u)), b + uB), ∇u)) ∼= (H •(X, (C((u)), b + uB)), ∇u). In particular, if we have a chain of quasi-isomorphisms between mixed complexes (C ′, b′, B′, ∇′ u), RΓ(X, (C, b, B, ∇u)), then we get an isomorphism (H •(C ′((u)), b′ + uB′), ∇′ u) ∼= (H •(X, (C((u)), b + uB)), ∇u). Proof. For any sheaf F of k -vector spaces on X, we claim that G≤d(F)((u)) is an acyclic resolution of the sheaf F((u)). Indeed, the fact that it is a resolution follows from the exactness of −((u)) (Lemma 3.15), so we have to check that it consists of acyclic sheaves. This reduces to the following lemma. Lemma 3.25. For any acyclic sheaf G on X, the sheaf G((u)) is also acyclic. CYCLIC HOMOLOGY OF CATEGORIES OF MATRIX FACTORIZATIONS 23 Proof. In the special case when G is flasque, we have that G((u)) is also flasque, hence acyclic. Indeed, the restriction maps F(U )((u)) → F(V )((u)) are surjective since the maps F(U ) → F(V ) are surjective. For arbitrary acyclic G, take any of its flasque resolution K•(G). Then K•(G)((u)) is a flasque resolution of G((u)). Applying Γ(X, −) to this resolution, we see that G((u)) is acyclic. (cid:3) Now, we have a natural morphism of acyclic resolutions (which is of course a quasi- isomorphism): G≤d(F((u))) → G≤d(F)((u)). By functoriality, it follows that we have a quasi-isomorphism of Z/2 -graded complexes of acyclic sheaves of k((u)) -modules, with u -connections: u -connections: (G≤d(C((u))), G≤d(b)+uG≤d(B), G≤d(∇u)) → (G≤d(C)((u)), G≤d(b)+uG≤d(B), G≤d(∇u)). Applying to it the functor Γ(X, −) and passing to cohomology, we obtain the desired isomorphism. This proves the proposition. (cid:3) 4. Affine case Fix some field k of characteristic zero. In this section we will prove Theorem 1.4 in the affine case. Let A be a smooth commutative algebra of finite type over k, and take some W ∈ A. Put X := Spec A We assume that the critical locus of W is contained in W −1(0) ⊂ X. Note that the strict morphism of complexes with u -connections (4.1) (Ω• A, −dW, d, ∇DR u ) → RΓ(X, (Ω• X , −dW, d, ∇DR u )) is a quasi-isomorphism. This follows from the fact that the sheaves Ωp X is affine. X are acyclic since So our goal is to construct a chain of weak quasi-isomorphisms between Hoch(M Fcoh(X, W )) and (Ω• A, −dW, d, ∇DR u ). First, by Proposition 2.5 and Proposition 3.7 we have a strict quasi-isomorphism (4.2) Hoch((A, W )-modcdg f gp) ∼→ Hoch(M Fcoh(X, W )). Further, by Proposition 3.12 we have a strict quasi-isomorphism (4.3) Hoch((A, W )-modcdg f gp) ∼→ HochΠ((A, W )-modcdg f gp). 24 ALEXANDER I. EFIMOV Applying Proposition 3.10 to pseudo-equivalences from Proposition 2.13, we obtain strict quasi-isomorphisms (4.4) HochΠ((A, W )-modcdg f gp) ∼→ HochΠ((A, W )-modqdg f gp) ∼← HochΠ(A, −W ). Next, we apply Proposition 3.13 to the curved DG algebra (A, −W ), and get a weak quasi-isomorphism (4.5) HochΠ(A, −W ) ∼ → HochΠ(A, W ). Finally, by Proposition 3.14 we have a weak quasi-isomorphism (4.6) HochΠ(A, W ) ∼→ (Ω• A, −dW, d, ∇DR u ). Combining quasi-isomorphisms (4.1)-(4.6), we obtain the following result. Theorem 4.1. Let X be a smooth affine scheme of finite type over a field k of character- istic zero. Then there is a chain of weak quasi-isomorphisms between mixed complexes with u -connections Hoch(M Fcoh(X, W )) and (Γ(X, Ω• X ), −dW, d, ∇DR u ). 5. General case Here we will prove Theorem 1.4 in the general case. Let X be a smooth separated scheme of finite type over a field k of characteristic zero, and W ∈ O(X). Again, we assume that the critical locus of W is contained in W −1(0). Consider the following presheaves of CDG categories: M Fcoh(X, W )(U ) := M Fcoh(U, W ); M F nv(X, W )(U ) := M F nv(U, W ); (OX , W )-modqdg l.f. (U ) := (OU , W )-modqdg l.f. ; (OX , ±W )(U ) := (OX (U ), ±W ). Here the CDG category (OU , W )-modqdg f gp, but we l.f. take locally free sheaves of finite rank on U instead of finitely generated projective modules. is defined in the same way as D-modqdg Again, we have natural morphisms of presheaves of CDG categories M F nv(X, W ) → (OX , W )-modqdg l.f. ← (OX , −W ). Now our goal is to construct a chain of quasi-isomorphisms between Hoch(M Fcoh(X, W )) and RΓ(X, (Ω• X , −dW, d, ∇DR u )). Combining Proposition 3.17 with quasi-isomorphisms (4.1)-(4.6) from the previous sec- tion, we obtain strict quasi-isomorphisms of sheaves of mixed complexes with u -connections: (5.1) Hoch(M F nv(X, W )) → Hoch(M Fcoh(X, W )); CYCLIC HOMOLOGY OF CATEGORIES OF MATRIX FACTORIZATIONS 25 Hoch(M F nv(X, W )) → HochΠ(M F nv(X, W )); HochΠ(M F nv(X, W )) → HochΠ((OX , W )-modqdg l.f. ) ← HochΠ(OX , −W ); HochΠ(OX , −W ) → HochΠ(OX , W ), HochΠ(OX , W ) → (Ω• X, −dW, d, ∇DR u ). (5.2) (5.3) (5.4) (5.5) Applying the functor RΓ(X, −) to the quasi-isomorphisms (5.1)-(5.5), and using Propo- sition 3.23 we get a chain of quasi-isomorphisms between RΓ(X, Hoch(M Fcoh(X, W ))) and RΓ(X, (Ω• X , −dW, d, ∇DR u )). Thus, we are left to prove the following result Proposition 5.1. The morphism (5.6) Hoch(M Fcoh(X, W )) → RΓ(X, Hoch(M Fcoh(X, W ))) is a quasi-isomorphism. Proof. The statement is just about Hochschild complexes without additional structures, so we can forget about Connes differential and the connection. In [EP], the support of a matrix factorization (E, δ) is defined as the smallest closed subset Supp(E, δ) ⊂ X such that the restriction of (E, δ) onto X\Supp(E, δ) is isomorphic to zero (in the homotopy category of matrix factorizations). Further, for any closed subset Z ⊂ X denote by M Fcoh,Z(X, W ) ⊂ M Fcoh(X, W ) the full DG subcategory consisting of matrix factorizations with support contained in Z. By [EP], for any closed subset Z ⊂ X we have a short exact sequence of DG categories M Fcoh,Z(X, W ) → M Fcoh(X, W ) → M Fcoh(X \ Z, W ). It follows from Theorem 3.9 that we have an exact triangle of Hochschild complexes: (5.7) Hoch(M Fcoh,Z(X, W )) → Hoch(M Fcoh(X, W )) → Hoch(M Fcoh(X \ Z, W )) → Hoch(M Fcoh,Z(X, W ))[1]. Further, if X = U1 ∪ U2 is an open covering, then putting Z = X \ U1 and applying triangle (5.7) to X and U2 we get a pair of exact triangles Hoch(M Fcoh,Z(X, W )) → Hoch(M Fcoh(X, W )) → Hoch(M Fcoh(U1, W )) → Hoch(M Fcoh,Z(X, W ))[1]. 26 ALEXANDER I. EFIMOV Hoch(M Fcoh,Z(U2, W )) → Hoch(M Fcoh(U2, W )) → Hoch(M Fcoh(U1 ∩ U2, W )) → Hoch(M Fcoh,Z(U2, W ))[1]. Since the natural restriction DG functor is a quasi-equivalence, we get a Mayer-Vietoris exact triangle M FZ (X, W ) → M FZ (U2, W ) Hoch(M Fcoh(X, W )) → Hoch(M Fcoh(U1, W )) ⊕ Hoch(M Fcoh(U2, W )) → Hoch(M Fcoh(U1 ∩ U2, W )) → Hoch(M Fcoh(X, W ))[1]. It naturally maps to the usual Mayer-Vietoris exact triangle for the complex of sheaves Hoch(M Fcoh(X, W )) : RΓ(X, Hoch(M Fcoh(X, W ))) → RΓ(U1, Hoch(M Fcoh(X, W ))) ⊕ RΓ(U2, Hoch(M Fcoh(X, W ))) → RΓ(U1 ∩ U2, Hoch(M Fcoh(X, W ))) → RΓ(X, Hoch(M Fcoh(X, W )))[1]. Therefore, to show that (5.6) is a quasi-isomorphism for the pair (X, W ), it suffices to show it for the pairs (U1, W ), (U2, W ) and (U1 ∩ U2, W ). But the scheme X is separated of finite type, so the problem reduces to affine varieties. And in the affine case this follows from Theorem 4.1. This proves the proposition. (cid:3) We have proved the following theorem (Theorem 1.4): Theorem 5.2. Under the above assumptions, we have a chain of quasi-isomorphisms be- tween mixed complexes with u -connections: (Hoch(M Fcoh(X, W )), b, B, ∇GM u ), RΓ(X, (Ω• X , −dW, d, ∇DR u )). This Theorem implies implies the following (Theorem 1.3) Theorem 5.3. Under the above assumptions, we have natural isomorphisms (HP•(M Fcoh(X, W )), ∇GM u and ) ∼= (H •(X, (Ω• X ((u)), −dW + ud)), ∇DR u ), HH•(M Fcoh(X, W )) ∼= H •(X, (Ω• X , −dW )). Proof. Indeed, the second isomorphism directly follows from Theorem 5.2, and the first one is a combination of Theorem 5.2 and Proposition 3.24. (cid:3) Further, we have the following corollary over complex numbers (Theorem 1.1). CYCLIC HOMOLOGY OF CATEGORIES OF MATRIX FACTORIZATIONS 27 Theorem 5.4. Let X be a smooth quasi-projective algebraic variety over C. Then we have an identification of Z/2 -graded vector bundles with connection on the formal punctured disk: (HP•(M Fcoh(X, W )), ∇GM u −1 ) ∼= dRH (H •−1 an (W −1(0), φW CX), T · (−1)•). Proof. This follows from the first isomorphism of Theorem 5.3 and from Theorem 1.2 (result of Sabbah). Indeed, adding the term Γ u to the connection on the twisted de Rham complex corresponds to multiplying the monodromy by (−1)•. This proves the theorem. (cid:3) 6. Concluding remarks Here we write down some remarks. First, we know by Weibel [W2] and Keller [Ke2] that for any complex quasi-projective va- riety X (not necessarily smooth) the periodic cyclic homology HP•(Perf(X)) is identified with H • top(X, C). The naive hope would be that the similar result should hold for locally free matrix factorizations. More precisely, one can define the full DG subcategory M Fl.f.(X, W ) ⊂ M Fcoh(X, W ) to consist of matrix factorizations locally isomorphic (in the derived sense) to the free matrix factorizations of finite rank. For smooth X, these two categories coincide. However, it is shown in [EP] that the categories M Fl.f.(X, W ) in general behave very bad. In particular, the example discussed in [EP], Section 3.3 shows that: Thomason localization theory fails; there may not exist a single generating object; periodic cyclic homology can be infinite- dimensional. Thus, we cannot expect any kind of analogue of Theorem 1.1 for the categories M Fl.f.(X, W ) for singular X. On the other hand, the categories M Fcoh(X, W ) behave very well for arbitrary separated X of finite type over a field of characteristic zero. We expect an isomorphism (HP•(M Fcoh(X, W )), ∇u) ∼= dRH −1 ((H •−1 c (W −1(0), φW CX))∨, T ∨ · (−1)•). Of course, these isomorphisms should admit a version with support: for any closed sub- scheme Z ⊂ X, we can take matrix factorizations with support on Z in the LHS, and the cohomology supported on Z in the RHS. The Z -grading on the vanishing cohomology should come from lambda-decomposition for the sheaf of Hoschschild complexes of the sheaf of curved DG algebras (OX , W ), as in [W2] for W = 0. In our case of smooth varieties this is straightforward. We also note that Thom-Sebastiani theorem for vanishing cohomology [M] over com- plex numbers can be viewed as a combination of the Kunneth isomorphism for peridoic 28 ALEXANDER I. EFIMOV cyclic homology ([HS], [Sh2]) for matrix factorizations categories, and the Thom-Sebastiani Theorem for matrix factorizations [Pr]. Also, in the case of compact critical locus and smooth X, the Kontsevich-Soibelman degeneration conjecture for the DG category M Fcoh(X, W ) is a combination of our Theo- rem 1.4 and the theorem of Barannikob-Kontsevich (unpublished) on degeneration for the twisted de Rham complex, proved also in [Sab1] and [OV]. Finally, for a smooth algebraic variety over C, and a regular function W, the categorical Chern character gives a map K0(M Fcoh(X, W )c) → H odd an (W −1(0), φW CX)T (here we get odd cohomology because in the case of W = 0 the standard degree on coho- mology is shifted by 1. ). The superscript c means that we take the Karoubi completion. This map can be shown to map to Hodge classes. The analogue of Hodge conjecture (see also [KKP]) states that this Chern character ⊗Q is surjective onto rational Hodge classes (at least in the case of compact critical locus). References [C] [Dr] [FT] A. Connes. Noncommutative differential geometry. Inst. Hautes ´Etudes Sci. Publ. Math. No. 62 (1985), 257-360. V. Drinfeld, DG quotients of DG categories. J. Algebra 272 (2004), no. 2, 643-691. B. Feigin, B. Tsygan, Additive K-theory, Lecture Notes in Mathematics Volume 1289, 1987, pp 67-209. [G] E. Getzler, Cartan homotopy formulas and the Gauss-Manin connection in cyclic homology. (English summary) Quantum deformations of algebras and their representations (Ramat-Gan, 1991/1992; Rehovot, 1991/1992), 6578, Israel Math. Conf. Proc., 7, Bar-Ilan Univ., Ramat Gan, 1993. [HKR] G. Hochschild, B. Kostant, A. Rosenberg, Differential forms on regular affine algebras, Transactions AMS 102 (1962), No.3, 383-408 (reprinted in G. P. Hochschild, Collected Papers. Volume I 1955-1966, Springer 2009, 265-290) [HS] Ch. E. Hood, J. D. S. Jones, Some algebraic properties of cyclic homology groups. K-Theory 1 (1987), no. 4, 361-384. [KKP] L. Katzarkov, M. Kontsevich, and T. Pantev, Hodge theoretic aspects of mirror symmetry, in From Hodge theory to integrability and TQFT: tt*-geometry, R. Donagi et K. Wendland ed., Proc. Sym- posia in Pure Math. vol. 78, Amer. Math. Soc. 2008, p. 87-174. [Ke1] B. Keller, Deriving DG categories, Ann. scient. Ec. Norm. Sup., 4e s´erie 27 (1994), 63-102. [Ke2] B. Keller, On the Cyclic Homology of Ringed Spaces and Schemes, Doc. Math. J. DMV 3 (1998), 231-259. [Ke3] B. Keller, Invariance and Localization for cyclic homology of DG algebras, J. Pure Appl. Alg. 123 (1998), 223-273. [Ke4] B. Keller, On the cyclic homology of exact categories. J. Pure Appl. Algebra 136 (1999), no. 1, 156. CYCLIC HOMOLOGY OF CATEGORIES OF MATRIX FACTORIZATIONS 29 [KS] M. Kontsevich, Y. Soibelman, Notes on A∞ -algebras, A∞ -categories and non-commutative geom- etry. Homological mirror symmetry, 153-219, Lecture Notes in Phys., 757, Springer, Berlin, 2009. [LP] K. Lin, D. Pomerleano, Global matrix factorizations, Math. Res. Lett. 20 (2013), no. 1, 91-106. [M] D. Massey, Sebastiani-thom isomorphism in the derived category, Compositio Mathematica, Volume 125, Number 3, February 2001 , pp. 353-362(10). [OV] A. Ogus, V. Vologodsky, Nonabelian Hodge theory in characteristic p. Publ. Math. Inst. Hautes ´Etudes Sci. No. 106 (2007), 1-138. [Or1] D. Orlov, Triangulated categories of singularities and D-branes in Landau-Ginzburg models. (Rus- sian) Tr. Mat. Inst. Steklova 246 (2004), Algebr. Geom. Metody, Svyazi i Prilozh., 240–262; transla- tion in Proc. Steklov Inst. Math. 2004, no. 3 (246), 227-248 [Or2] D. Orlov, Matrix factorizations for nonaffine LG-models. Math. Ann. 353 (2012), no. 1, 95-108. [EP] A.I. Efimov, L. Positselski, Coherent analogues of matrix factorizations and relative singularity categories, Algebra Number Theory 9 (2015), no. 5, 1159-1292. [PP] A. Polishchuk, L. Positselski, Hochschild (co)homology of the second kind I. Transactions of the Amer. Math. Soc. 364 (2012), #10, p. 5311-5368. [PV] A. Polishchuk, A. Vaintrob, Matrix factorizations and singularity categories for stacks, Ann. Inst. Fourier (Grenoble) 61 (2011), no. 7, 2609-2642. [Pr] A. Preygel. Thom-Sebastiani and Duality for Matrix Factorizations, and Results on the Higher Structures of the Hochschild Invariants. Thesis (Ph.D.)-Massachusetts Institute of Technology. 2012. [Sab1] C. Sabbah, On a twisted de Rham complex. Tohoku Math. J. (2) 51 (1999), no. 1, 125-140. [Sab2] C. Sabbah, On a twisted de Rham complex, II, arXiv:1012.3818 (preprint). [SS] C. Sabbah, M. Saito, Kontsevich’s conjecture on an algebraic formula for vanishing cycles of local systems, Algebr. Geom. 1 (2014), no. 1, 107-130. [Seg] E. Segal, The closed state space of affine Landau-Ginzburg B-models, J. Noncommut. Geom. 7 (2013), no. 3, 857-883. [Sh1] D. Shklyarov, Non-commutative Hodge structures: Towards matching categorical and geometric examples, Trans. Amer. Math. Soc. 366 (2014), no. 6, 2923-2974. [Sh2] D. Shklyarov, On a Hodge theoretic property of the Knneth map in periodic cyclic homology. J. Algebra 446 (2016), 132-153. [W1] C. A. Weibel, An Introduction to Homological Algebra, Cambridge University Press, 1994 [W2] C. A. Weibel, The Hodge Filtration and Cyclic Homology, K-Theory 12 (1997), 145-164. [WG] C. A. Weibel, S. C. Geller, Etale descent for Hochschild and cyclic homology, Comment. Math. Helv., vol. 66, no. 1, pp. 368-388, 1991 Steklov Mathematical Institute of RAS, Gubkin str. 8, Moscow 119991, Russia E-mail address: [email protected]
1208.1813	2	1208	2012-09-05T08:14:42	Determinantal Quintics and Mirror Symmetry of Reye Congruences	[ "math.AG", "hep-th" ]	We study a certain family of determinantal quintic hypersurfaces in $\mathbb{P}^{4}$ whose singularities are similar to the well-studied Barth-Nieto quintic. Smooth Calabi-Yau threefolds with Hodge numbers $(h^{1,1},h^{2,1})=(52,2)$ are obtained by taking crepant resolutions of the singularities. It turns out that these smooth Calabi-Yau threefolds are in a two dimensional mirror family to the complete intersection Calabi-Yau threefolds in $\mathbb{P}^{4}\times\mathbb{P}^{4}$ which have appeared in our previous study of Reye congruences in dimension three. We compactify the two dimensional family over $\mathbb{P}^{2}$ and reproduce the mirror family to the Reye congruences. We also determine the monodromy of the family over $\mathbb{P}^{2}$ completely. Our calculation shows an example of the orbifold mirror construction with a trivial orbifold group.	math.AG	math	DETERMINANTAL QUINTICS AND MIRROR SYMMETRY OF REYE CONGRUENCES SHINOBU HOSONO AND HIROMICHI TAKAGI Abstract. We study a certain family of determinantal quintic hypersurfaces in P4 whose singularities are similar to the well-studied Barth-Nieto quintic. Smooth Calabi-Yau threefolds with Hodge numbers (h1,1, h2,1) = (52, 2) are obtained by taking crepant resolutions of the singularities. It turns out that these smooth Calabi-Yau threefolds are in a two dimensional mirror family to the complete intersection Calabi-Yau threefolds in P4 × P4 which have ap- peared in our previous study of Reye congruences in dimension three. We compactify the two dimensional family over P2 and reproduce the mirror fam- ily to the Reye congruences. We also determine the monodromy of the family over P2 completely. Our calculation shows an example of the orbifold mirror construction with a trivial orbifold group. Contents ) 0 0 0 0 ) 2 0 Introduction 2 ) and e( X sp,♯ 0 2 and X sp,♯ e(Ds) for Z sp e(Ds) for X sp,♯ 1. 2. Backgrounds and summary of main results 2.1. Three dimensional Reye congruences 2.2. Orbifold mirror construction of X0 2.3. Special determinantal quintics Z sp 3. The Euler numbers e(Z sp 3.1. Euler number e(Z sp 2 ) 3.2. Euler number e( X sp,♯ 4. Calculations of the Euler numbers e(Ds) 4.1. 4.2. 5. Crepant resolutions X ∗ 5.1. Singular loci of Z sp 5.2. Singular loci of X sp 5.3. Blowing-ups of X sp 5.4. Hodge numbers 6. Picard-Fuchs equations and monodromy matrices 6.1. Picard-Fuchs differential equations 6.2. Determinantal quintics 6.3. 6.4. Mirror symmetry of Reye congruences 7. Special families of Steinerian and Hessian quintics 7.1. Steinerian and Hessian quintics. 7.2. Singular loci of Hsp 7.3. The ramified covering Ysp → Hsp References Integral, symplectic basis and monodromy matrices 1 and Z sp 2 0 → X sp 0 1 2 3 3 5 7 9 9 11 13 13 14 15 15 16 18 26 30 30 31 32 40 42 42 42 44 46 1. Introduction Quintic hypersurfaces in the projective space P4 have been invaluable testing grounds for the interesting mathematical ideas coming from the string theory. This has been for long since the historical discovery of an exact solution of N=2 su- perconformal field theory/string theory and its profound relations to the quintic hypersurfaces [Gep]. The idea of mirror symmetry of Calabi-Yau manifolds, for example, has been verified first by translating a special involution in the set of N=2 theories into some operation, now-called orbifold mirror construction, in the algebraic geometry of quintic hypersurfaces [GP], [Yau], and also the surprising ap- plications of the mirror symmetry to Gromov-Witten theory have been started from the Hodge theoretical investigations of the mirror quintic hypersurfaces [CdOGP]. In this paper, we will be concerned with certain special types of quintic hypersur- faces in P4 which are called determinantal quintics. The determinantal quintics are interesting not only from the viewpoint of the mirror symmetry but also from the viewpoint of the classical projective geometry. In fact these quintics have appeared in our previous study of the so-called Reye congruences in dimension three [HT], where a beautiful interplay between the mirror symmetry and the classical projec- tive geometry has been observed. Historically, Reye congruences represent certain Enriques surfaces, called nodal Enriques surfaces [Co], [Ty], and their study goes back to the 19th century, where the term 'congruence' arose in relation to the ge- ometry of the Grassmannian G(2, 4). They naturally come with K3 surfaces which admit fixed point free involutions. In dimension three, the corresponding Reye con- gruences turn out to be Calabi-Yau manifolds with non-trivial fundamental groups [Ol], and they also come with Calabi-Yau threefolds equipped with fixed point free involutions which we call covering Calabi-Yau threefolds of the Reye congruences. In our previous work [HT], we have studied the mirror symmetry of the three dimensional Reye congruences through the covering Calabi-Yau threefolds using the methods in the toric geometry [BaBo]. In this paper, we will reconsider the mirror symmetry based on the orbifold mirror construction and will observe that the projective geometries of certain singular determinantal quintics come into play in an interesting way. Also we find that, in our case, the so-called orbifold group is a trivial group, Gorb = {id}. The last property naturally leads us to a problem that how is the mirror involution in the corresponding N=2 string theory realized in such cases, although we will not discuss the problem in this article. The construction of this paper is as follows: In the next section we will summarize the geometries of the Reye congruences following the previous work. There, after setting up the notation and the problems in details, we describe the main results of this paper. In Section 3, we calculate the topological Euler numbers of certain (singular) determinantal quintic hypersurfaces. In Section 4, we describe the details about the calculations of some Euler numbers needed in Section 3. In Section 5, we will obtain the mirror family to the covering Calabi-Yau threefolds of the Reye congruences. In Section 6, we will determine completely the monodromy properties of the mirror family. Taking the fixed point free involution into account, we construct the mirror family to the Reye congruences. In Section 7, we will discuss some geometry of the singular Hessian quintics. 2 Acknowledgements: The authors would like to thank Prof. B. van Geemen for his kind and helpful correspondence to their question about the étale cohomology. They also would like to thank Prof. C. Vafa for his correspondence. This work is supported in part by Grant-in Aid Scientific Research (C 22540041, S.H.) and Grant-in Aid for Young Scientists (B 30322150, H.T.). 2. Backgrounds and summary of main results 2.1. Three dimensional Reye congruences. Let us consider the product P4×P4 of the complex projective spaces with its bi-homogeneous coordinate ([z], [w]). We consider a generic complete intersection X0 of five (1, 1) divisors in the product. In terms of the bi-homogeneous coordinates, X0 may be written by f1 = ··· = f5 = 0 in P4 × P4, where we set fk := tzAkw with 5 × 5 matrices A1, ..., A5 over C. When Ak are generic, X0 defines a smooth Calabi-Yau threefold with its Hodge numbers h11( X0) = 2, h21( X0) = 52. Despite this simple descriptions, X0 has interesting birational geometries which we summarize in the following diagram: (2.1) X0 π2 ❄❄❄❄❄❄❄❄ X2 p1 ⑧⑧⑧⑧⑧⑧⑧⑧ Z2 p2 ❇❇❇❇❇❇❇❇ X ♯ 0 , where Z2 and X ♯ 0 are determinantal quintics defined by and X ♯ 0 =([λ] ∈ P4 Z2 =(cid:8) [w] ∈ P4 det(A1wA2w...A5w) = 0(cid:9) , λkAk(cid:1) = 0) , λ(cid:12)(cid:12)(cid:12)(cid:12)det(cid:0) 5Xk=1 λ Aλw = 0(cid:9) , where Aλ = X2 =(cid:8)[w] × [λ] ∈ P4 × P4 λkAk. 5Xk=1 respectively, and X2 is defined by P4 λ is the projective space defined from the C-vector space spanned by the matrices Ak(k = 1, ..., 5). The maps π2 and pi (i = 1, 2) in the diagram (2.1) are defined by the natural projections; the projection to the second factor P4 × P4 → P4 for π2, and the projections from P4 × P4 λ to the first and the second factors for pi (i = 1, 2), respectively. As we can see in the definitions, both Z2 and X ♯ 0 are quintic hypersurfaces in the respective projective spaces, and X2 is a complete intersection of five (1, 1) divisors in the product P4× P4 λ. When the matrices Ak are generic, the both Z2 and X ♯ 0 determine generic determinantal varieties in P4 and P4 λ, respectively. Generic determinantal varieties are known to be singular along codimension three loci, where the matrices have corank two (see [HT, Lemma 3.2] 3 for example). In our case, the degree of the singular loci is 50. Hence generic Z2 and X ♯ 0 are singular at 50 points, where the rank of the relevant matrices decreases to three, and actually these consist of 50 ordinary double points [ibid. Proposition 3.3]. We also note that X0 and X2 are birational but not isomorphic in general [ibid. Sect.(3-2)]. The geometries of X0 and X ♯ 0 in the above diagram fit well to the classical projective duality, since the projective dual (P4×P4)∗ to the Segre variety P4×P4 ֒→ P24 is naturally given by the determinantal variety in the dual projective space (P24)∗ and X ♯ 0 is given by a linear section of this determinantal variety. Based on this, in [ibid. Sect. (3-1)] we have called the determinantal quintic X ♯ 0 as the Mukai dual of X0. The diagram (2.1) shows further interesting properties if we require the matrices Ak to be symmetric. When we identify these symmetric matrices with quadrics in P4, the projective space P4 λ is nothing but the 4-dimensional linear system of the quadrics spanned by Ak, which we denote by P = A1, A2, ..., A5. In general, an n-dimensional linear system of quadrics in Pn is called regular if it is base point free and satisfies a further condition [Co], [Ty]. In our present case, for a regular linear system of quadrics P = A1, A2, ..., A5, we have a smooth Calabi-Yau threefold X = X0 which admits a fixed point free involution; σ : ([z], [w]) 7→ ([w], [z]). Unlike the 2-dimensional case, this involution preserves the holomorphic three form and we obtain a Calabi-Yau threefold X = X/hσi, which is called a Reye congruence in dimension three [Ol]. X has the Hodge numbers h1,1(X) = 1, h2,1(X) = 26 and degree 35. Corresponding to the diagram (2.1), we have (2.2) X /hσi X "❋❋❋❋❋❋❋❋❋ S U #●●●●●●●●● ②②②②②②②②② Y 2:1 H . Here we have adopted the historical notations S and H for the determinantal va- rieties of symmetric matrices; S will be called the Steinerian quintic and H the Hessian quintic. These belong to the special families of the previous determinan- tal quintics Z2 and X ♯ 0, i.e., the Steinerian quintic is defined by the equations det(Aw) = 0 with Aw := (A1wA2w...A5w) and similarly for the Hessian quintic with Aλ. However, for the generic regular linear system P , Aw is not symmetric while Aλ is. Due to this, S has generically 50 ordinary double points while H is singular along a (smooth) curve of genus 26 and degree 20. In our previous work, guided by the calculations from mirror symmetry, we have found: Theorem ([HT, Theorem 3.14]) There exists a double covering Y of the Hessian quintic H branched along the singular locus, which is a smooth curve of genus 26 and degree 20. Y is a smooth Calabi-Yau threefold with the Hodge numbers h1,1(Y ) = 1, h2,1(Y ) = 26 and degree 10 with respect to OY (1). An explicit description of the covering Y will be given in Definition 7.6 and Remark 7.8. We can observe an interesting projective duality behind the diagram (2.2). This time the projective dual we start with is the dual (Sym2P4)∗ associated to the embedding Sym2P4 ֒→ P14 by the Chow form. This duality has quite similar prop- 2 n(n−1)−1, which appeared erties to that of the Grassmannians under G(2, n) ֒→ P 1 4 " # in [Ro], [BoCa], [Ku]. Observing this similarity, and also from the mirror symmetry, it has been conjectured that the Calabi-Yau threefolds X and Y in the diagram have the equivalent derived categories of coherent sheaves although they are not birational (see [Hor], [JKLMR] and references therein for physical arguments on this). Our main objective in this paper is to construct 'the mirror diagrams' to the two diagrams (2.1) and (2.2). For this, we start with the orbifold mirror construction of X0. 2.2. Orbifold mirror construction of X0. Orbifold mirror constructions in gen- eral consist of the following three main steps: Given a generic complete intersection Calabi-Yau manifold (CICY) in a product of (weighted) projective spaces, we first consider it in its deformation family. Then, secondly we try to find a suitable spe- cial family of the generic deformation family. In general, we encounter singularities in the generic members of the special family. We may seek crepant resolutions of them at this point or defer them to the next step since crepant resolutions may not exist at this point. As the third step, we try to find a suitable finite group Gorb which acts on generic fibers of the family and preserves holomorphic three forms on them. Gorb is required to have the property that we have the mirror relations in the Hodge numbers when we take the quotient (orbifold) of the generic fibers and after making crepant resolutions of the singularities, if any. Apart from the hypersurfaces of Fermat type in the weighted projective spaces [GP], [Ba], the existence of the suitable special family and also Gorb is based on case-by-case studies for general CICY's (see [BeH] for Calabi-Yau hypersurfaces of non-Fermat type). In our case of the complete intersection X0, we first consider the following special (two dimensional) family of the complete intersection: (2.3) z1w1 + a z2w1 + b z1w2 = 0, z3w3 + a z4w3 + b z3w4 = 0, z5w5 + a z1w5 + b z5w1 = 0, z2w2 + a z3w2 + b z2w3 = 0, z4w4 + a z5w4 + b z4w5 = 0, where a and b are the parameters of the family. In what follows in this paper, by fk= tzAkw (k = 1, .., 5) we represent the above defining equations, i.e., we set A1 = 1 b 0 0 0 0 0 0 0 0! , A3 = 0 0 0 0 0 0 0 0 0 0! , 0 0 0 0 0 0 0 1 b 0 0 0 a 0 0 0 1 b 0 0 0 a 0 0 0 0 0 0 0 0 a 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0! , A2 = 0 0 0 0 0 A4 = 0 0 0 0 0 0 0 0 a 0! , A5 = 0 0 0 0 a b 0 0 0 1! . 0 0 0 0 0 0 0 0 0 0 0 0 0 1 b 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 We consider the above family over (C∗)2 by taking (a, b) ∈ (C∗)2, and denote by X sp 0 a general fiber of this family. This special form of the defining equations has been chosen so that period integrals of X sp calculated in terms of fk reproduce 0 the period integrals from the toric mirror construction [BaBo], [HKTY], see Sect.6. The validity of this choice will be confirmed by the mirror symmetry among the Hodge numbers (see Theorem 5.17). We may consider the restriction X sp 0 (C∗)8 of X sp 0 to (C∗)4 × (C∗)4 ⊂ P4 × P4. 5 Proposition 2.1. The restriction X sp and becomes singular when the following discriminant vanishes: 0 (C∗)8 is smooth for generic (a, b) ∈ (C∗)2 (2.4) dis( X sp 0 (C∗)8 ) = 4Yk,l=0 (µk a + µl b + 1), (µ5 = 1, µ 6= 1). Proof. The form of the discriminant follows from the Jacobian ideal by eliminating the homogeneous coordinates of the projective spaces. To implement the restric- tion to (C∗)4 × (C∗)4 ⊂ P4 × P4, we impose additional relations z1z2z3z4z5 = 1 and w1w2w3w4w5 = 1 to the Jacobian ideal. The eliminations may be done by Macauley2 [GS]. (cid:3) In the sections 5.1 and 5.2, we will derive the following property (see Proposition 5.3 for details): Proposition 2.2. For (a, b) ∈ (C∗)2 with non-vanishing discriminant (2.4), the complete intersection X sp 0 is singular along 20 lines of A1-singularity which intersect at 20 points. Local geometries about the intersections are classified into two types, which we call (3A1,U1) and (2A1,U2), with the 20 points being split into 10 points for each. We determine the singular loci above essentially by the Jacobian criterion, how- ever straightforward calculations do not work since the Jacobian ideal turns out to be complicated. We avoid this complication by studying the singular loci of the determinantal quintics which are naturally associated to X sp (see the next subsec- 0 tion). Detailed analysis will be given in Sect.5. There, the type of the singularities and also the configuration of them will be determined (see Fig.5.1). The configu- ration of the singular loci, consisting of 20 lines of A1-singularity, is similar to the Barth-Nieto quintic studied in [BaN] (see also [HSvGvS]). While the local geometry (3A1,U1) has the corresponding geometry in the Barth-Nieto quintic, the geometry (2A1,U2) (and also (∂A1,U3) which will be introduced in Sect. 5) is new in our case. For the resolution of the singularities, as in [BaN] (see also [HSvGvS]), we start with the blowing-up at the 10 points of (3A1,U1) singularity and continue the blowing-up along the strict transforms of the lines in the prescribed way in Sect.5. Then we finally obtain the following result: Main Result 1. vanishing discriminant (2.4), there exists a crepant resolution X∗0 → X sp Hodge numbers: (Theorem 5.11, Theorem 5.17) For (a, b) ∈ (C∗)2 with non- 0 with the h1,1( X∗0 ) = h2,1( X0) = 52, h2,1( X∗0 ) = h1,1( X0) = 2. Namely, the resolution X∗0 is a mirror Calabi-Yau threefold to X0. In particular, we have a trivial finite group Gorb = {id} for the orbifold mirror construction. 6 2.3. Special determinantal quintics Z sp obtain two determinantal quintics Z sp into the following diagram: 2 and X sp,♯ 0 2 and X sp,♯ 0 from X sp . As in the diagram (2.1), we 0 , which can be arranged (2.5) X∗0 / X sp 0 X sp 2 }⑤⑤⑤⑤⑤⑤⑤⑤ !❇❇❇❇❇❇❇❇ Z sp 2 "❊❊❊❊❊❊❊❊ X sp,♯ 0 , where we define X sp is defined by the map π2 : X sp 2 associated with the projection to the second factor P4 × P4 → P4. The defining equation det(A1w A2w...A5w) = 0 is given by the following quintic: 2 as X2 in (2.1). The first determinantal quintic Z sp 0 → Z sp 2 det w1 + bw2 0 aw1 w2 + bw3 0 0 0 0 0 aw2 w3 + bw4 0 0 aw3 0 0 0 0 w4 + bw5 aw5 0 0 0 aw4 w5 + bw1  = a5w1w2w3w4w5 +(cid:0)w1 + bw2(cid:1)(cid:0)w2 + bw3(cid:1)(cid:0)w3 + bw4(cid:1)(cid:0)w4 + bw5(cid:1)(cid:0)w5 + bw1(cid:1). Similarly, the second determinantal quintic X sp,♯ with 0 is defined by det(Pk λkAk) = 0 (2.6) (2.7) aλ5 0 0 bλ4 λ5  det λ1 aλ1 0 0 bλ5 bλ1 λ2 aλ2 0 0 0 bλ2 λ3 aλ3 0 0 0 bλ3 λ4 bλ4 =(cid:0)1 + a5 + b5(cid:1)λ1λ2λ3λ4λ5 + a2b2(cid:0)λ1λ2 − ab(cid:0)λ1λ2λ3λ2 4 + λ2λ2 4 + λ2λ3λ4λ2 2λ2 3λ2 5 + λ3λ2 4λ2 1 + λ4λ2 5 + λ3λ4λ5λ2 5λ2 2 + λ5λ2 1 + λ4λ5λ1λ2 1λ2 2 + λ5λ1λ2λ2 3(cid:1) 3(cid:1). Proposition 2.3. The singular loci of Z sp (C∗)2, i.e., the restriction Z sp 2 (C∗)4 of Z sp Z sp 2 (C∗)4 becomes singular for a, b on the discriminant {dis(Z sp (C∗)2, where 2 are in P4 \ (C∗)4 for generic (a, b) ∈ to the torus (C∗)4 ⊂ P4 is smooth. 2 (C∗)4) = 0} ⊂ 2 4Yk,l=0 (C∗)4 of X sp,♯ 0 (2.8) dis(Z sp 2 (C∗)4 ) = a5 (µk a + µl b + 1) (µ5 = 1, µ 6= 1). Similar restriction X sp,♯ becomes singular for the values on the discriminant {dis( X sp,♯ where is smooth for generic (a, b) ∈ (C∗)2 and (C∗)4) = 0} ⊂ (C∗)2, 0 0 (2.9) dis( X sp,♯ 0 (C∗)4) = 4Yk,l=0 (µk a + µl b + 1) × 7 4Yk=0 (a − µkb)2 (µ5 = 1, µ 6= 1). / ! } " Proof. As in Proposition 2.1, we impose the restrictions to (C∗)4 by adding the equations w1w2...w5 = 1 or λ1λ2...λ5 = 1 to the Jacobian ideals. By the elimina- tions, we obtain the claimed forms of the discriminants. (cid:3) X sp 2 is defined by special forms of five (1, 1)-divisors in P4×P4 λ. We can verify that the restriction to the tori (C∗)8 = (C∗)4 × (C∗)4 is smooth and has the following form of the discriminant: (2.10) dis( X2(C∗)8) = 4Yk,l=0 (µk a + µl b + 1), (µ5 = 1, µ 6= 1). Proposition 2.4. 1) For (a, b) ∈ (C∗)2 with non-vanishing discriminant (2.8), the determinantal quintic Z sp is singular along 5 coordinate lines, each of them is of 2 A2 type, and singular also along 10 lines of A1 singularity. These lines intersect at 15 points. 2) For (a, b) ∈ (C∗)2 with nonvanishing (2.9), the determinantal quintic X sp,♯ is singular along 5 coordinate lines, each of which is of A3 type, and singular also along 5 additional lines of A1 singularity. These lines intersect at 10 points. 0 Proof. We present the details of 1) in Sect. 5 and Fig. 5.1. The property of 2) is obtained in 7.3 (see Fig. 7.1). (cid:3) The complete intersections X sp 0 and X sp 0 Precisely, the crepant resolution X∗0 of X sp , respectively. In fact, all the singularities along the 15 lines in Z sp 2 give partial crepant resolutions of Z sp 2 and X sp,♯ 2 are (partially) resolved to the singularities of A1 type along the 20 lines in Proposition 2.2. The similar property also holds for the projection X sp (cf. Fig. 7.1). (claimed in Main Result 1) is valid for (a, b) ∈ (C∗)2 being away from the zero-loci of the discriminant (2.4) in (C∗)2. It is easy to see that X sp 0 with two different values of (a, b) and (µka, µlb) (µ5 = 1) are isomorphic to each other by a simple coordinate change. Based on this, we introduce the affine variables x = −a5, y = −b5 to have a smooth family over (C∗)2 ∋ (x, y), which will be compactified to a family over P2 (see Sect. 6). Main Result 2. (Propositions 6.6, 6.7, 6.8) Let X∗ be the family of Calabi-Yau manifolds X∗0 over P2, and consider the period integrals of the family. Then, the integral and symplectic basis of the period integrals are generated by the cohomology- valued hypergeometric series defined in [Ho2, Conj. 2.2] and [Ho1, Prop.1]. 2 → X sp,♯ 0 0 0 We remark that our crepant resolution X∗0 → X sp is valid also for a = b ∈ C∗ as far as we have non-vanishing discriminant (2.4). Hence we can consider the restriction of the family X∗ over P2 to a family over {x = y} ∼= P1 and have the following properties: (Proposition 6.9) Over the 'diagonal' {x = y} ∼= P1, except Main Result 3. 32 , the family X∗ → P2 admits a fiberwise fixed point free involution x = y = 1 found in [HT, Prop. 2.9]. By taking the fiberwise unramified quotient under this involution over P1, we obtain the mirror family X∗P1 of the Reye congruence X. In particular, the period integrals and the monodromy matrices from Main Result 2 reproduce the previous results obtained in [HT, Prop. 2.10, 3)]. We summarize the geometries of the generic fiber X∗ of X∗P1 → P1 as follows: 8 (2.11) X∗0 /Z2 X∗ X sp 0 !❈❈❈❈❈❈❈❈ Ssp ~⑤⑤⑤⑤⑤⑤⑤⑤ Usp "❉❉❉❉❉❉❉❉ Hsp , where Ssp and Hsp are the special forms of the Steinerian quintic and the Hessian quintic defined by (2.6) and (2.7) with a = b, respectively. We close this section noting some properties of the Hessian quintic Hsp. When a = b, the discriminant (2.9) of the Hessian Hsp vanishes. In Sect.7, we will explain this (Proposition 7.3) by observing that an elliptic normal quintic appears as a new component of the singular loci of X sp,♯ 0 when a = b. There we will also discuss that the Hessian quintic Hsp admits a double covering Ysp ramified along its singular loci. From the mirror symmetry considerations given in [HT], it is expected that there is a crepant resolution Y ∗sp of Ysp which gives a mirror Calabi-Yau threefold Y ∗(= Y ∗sp) to the Y of the Reye congruence X. Namely we expect that the pair (X, Y ) of Calabi-Yau manifolds associated with the Reye congruence is mirrored to another pair (X∗, Y ∗) of the mirror Calabi-Yau manifolds. Here Y ∗ can be either birational to X∗ or a Fourier-Mukai partner to X∗. Both cases are consistent with the homological mirror symmetry [Ko]. The construction of Y ∗sp is left for future study. 3. The Euler numbers e(Z sp 2 ) and e( X sp,♯ 0 ) In this section, we determine the Euler numbers of the determinantal quintics. Since these quintics are singular, we invoke to a topological method. We assume that (a, b) ∈ (C∗)2 is away from the zero of the discriminant (2.8) and (2.9), re- spectively, for the determinantal quintics Z sp 2 and X sp,♯ 0 . 3.1. Euler number e(Z sp 2 ). We compute the Euler numbers of the singular de- terminantal quintic Z sp in (2.6) with the following affine line l ⊂ P4 such that l ∪{v0} = P1 with v0 = [0 : 0 : 0 : 0 : 1] ∈ Z sp 2 : 2 by considering the intersections of Z sp 2 l : [w1 : w2 : w3 : w4 : w5] = [x1 : x2 : x3 : x4 : t] (t ∈ C). Substituting the coordinates of this line into the defining equation (2.6), we obtain f (t) = c2t2 + c1t + c0, where (3.1) c2 = b(x1 + bx2)(x2 + bx3)(x3 + bx4), c1 = a5x1x2x3x4 + 1 b c2(b2x1 + x4), c0 = x1x4c2 . The equation f (t) = 0 determines the intersection of l with Z sp 2 as the fiber over each point [x1 : x2 : x3 : x4 : 0] associated to the projection P4 \ {v0} → P3. Then, by counting the numbers of the solutions, we can calculate the Euler number 9 / / ! ~ " e(Z sp 2 ). The fiber over each point varies depending on the values of c2, c1,c0, and the followings are two extreme cases: 1) c2 = c1 = c0 = 0, and 2) c2 = c1 = 0 but c0 6= 0. We regard that the fiber over the former loci is P1 = l ∪ {v0}. The fiber over 2) is empty, however we may regard it as the point at infinity {v0}. We see that, in the present case, 2) does not occur since c2 = 0 implies c0 = 0, however, the following arguments are not restricted to such cases. For other cases than 1) and 2), the numbers of solutions of the equation f (t) = 0 (t ∈ C) are either 2 or 1. Depending on the numbers of solutions we define the following subsets in P3: Then the Euler number is evaluated by UP1 =(cid:8)[x] c2 = c1 = c0 = 0(cid:9), U2 =(cid:8)[x] c2 6= 0, c2 U1 =(cid:8)[x] c2 6= 0, c2 1 − 4c2c0 = 0(cid:9) ⊔(cid:8)[x] c2 = 0, c1 6= 0(cid:9). 2 ) = 2 e(U2) + e(U1) +(cid:0)e(P1) − 1(cid:1) e(UP1) + e(v0). We denote the discriminant surface by Ds, i.e., Ds =(cid:8)[x] ∈ P3 c2 e(Z sp The subset U1 consists of those points in an open subset of Ds or where f (t) becomes linear. Consider the inclusions: 1 − 4c2c0 6= 0(cid:9), 1− 4c2c0 = 0(cid:9). (3.2) {c2 = 0} ⊃ {c2 = c1 = 0} ⊃ {c2 = c1 = c0 = 0}, (3.3) and denote these by Vc2 ⊃ Vc2,c1 ⊃ Vc2,c1,c0 with the obvious definitions. Lemma 3.1. (3.4) e(Z sp 2 ) = 2 e(P3) + 1 − e(Ds) − e(Vc2 ) + e(Vc2,c1,c0). Vc2,c1. Since the union is disjoint, we have Proof. By definition, we have U1 =(cid:0)Ds \ (Ds ∩ Vc2 )(cid:1)⊔ (Vc2 \ Vc2,c1) and Ds ∩ Vc2 = e(U1) =(cid:0)e(Ds) − e(Vc2,c1)(cid:1) +(cid:0)e(Vc2 ) − e(Vc2,c1)(cid:1) = e(Ds) + e(Vc2) − 2 e(Vc2,c1). Similarly, we have U2 = P3 \ (Ds ∪ Vc2 ) and hence e(U2) = e(P3) − e(Ds ∪ Vc2 ) = e(P3) − e(Ds) − e(Vc2) + e(Vc2,c1). Also note that UP1 = Vc2,c1,c0 holds. Substituting all these expressions into (3.2), the claimed formula follows. (cid:3) Let us introduce the C-bases e1, ..., e5 of C5 by which we can write [x1 : x2 : ... : x5] = [x1e1 + x2e2 + ... + x5e5] for the projective space P4 = P(C5). We define coordinate (projective) lines Lij and also coordinate (projective) planes Lijk by Lij = hei, eji, Lijk = hei, ej, eki, where hei1 , ei2 , .., eiki represents the projective space spanned by the vectors ei1 , ei2, .., eik . We also define the following projective lines and planes: Li,jk = hei,bej − eki, Lij,mn = hei, ej, bem − eni. Lemma 3.2. We have e(Vc2 ) = 4 and e(Vc2,c1,c0) = 3. Proof. From the form of c2 we have the following decomposition into planes: Vc2 = L12,34 ∪ L41,23 ∪ L34,12, where three components are normal crossing in P3. From this, we have e(Vc2) = 3 e(P2) − 3 e(P1) + 1 = 4. Observe that Vc2,c1 = Vc2 ∩ {x1x2x3x4 = 0}. From this, we deduce that Vc2,c1 = ∂L12,34 ∪ ∂L41,23 ∪ ∂L34,12, 10 where ∂L12,34 represents the union of the boundary 3 lines L12 ∪ L2,34 ∪ L1,34, and similarly for ∂L41,23 and ∂L34,12. Inspecting the intersection points of the 9 lines carefully, we evaluate e(Vc2,c1) = 3. Note that in the present case, we have Vc2,c1,c0 = Vc2,c1, hence e(Vc2,c1,c0) = 3. (cid:3) Proposition 3.3. For (a, b) ∈ (C∗)2 with non-vanishing discriminant (2.8), the discriminant Ds is an irreducible, singular octic surface in P3 with its Euler number e(Ds) = 18. Proof. We defer the detailed calculations to the next section. (cid:3) Using the above Proposition and the preceding two Lemmas, we evaluate the Euler number e(Z sp 2 ) = 9− 18− 4 + 3 = −10. We remark that the arguments above are still valid for non-vanishing a = b as long as a, b are away from the zero of the discriminant (2.8). Proposition 3.4. For (a, b) ∈ (C∗)2 with non-vanishing discriminant (2.8), the determinantal quintic Z sp 2 ) = −10. 2 Also for a = b ∈ C∗ with the same property, we have e(Ssp) = −10. (2.6) has its topological Euler number e(Z sp 3.2. Euler number e( X sp,♯ (2.9), similar calculations apply to the determinantal quintic X sp,♯ Let us first consider the affine line ). For (a, b) ∈ (C∗)2 with non-vanishing discriminant given in (2.7). 0 0 l : [λ1 : λ2 : λ3 : λ4 : λ5] = [x1 : x2 : x3 : x4 : t] such that l ∪ {v0} = P1 with v0 = [0 : 0 : 0 : 0 : 1] ∈ X sp,♯ coordinates into the defining equation of X sp,♯ with 0 (t ∈ C) . Substituting the , we obtain f (t) = c2t2 + c1t + c0 0 3 + x4x2 c2 =a2b2(x2x2 2) − ab x2x3x4, c1 =(1 + a5 + b5)x1x2x3x4 + a2b2x2 1x2 c0 =a2b2(x1x2 4 + x3x2 4x2 2x2 1) − ab x1x2x3x2 4. 3 − ab(x3x4x2 1 + x4x1x2 2 + x1x2x2 3), This time, it turns out that the discriminant surface Ds = {c − 4c2c0 = 0} in P3 consists of two irreducible components D1 s := {x1 = 0}, where the compo- s is an irreducible, singular septic in P3. We may verify these properties by nent D1 Macaulay2. The general formula (3.4) is still valid for the present case of e( X sp,♯ ), since it is topological. However we see some complications in the necessary calcu- lations, which we will sketch briefly below. s and D2 0 We use Macaulay2 for the calculations e(Vc2 )and e(Vc2,c1,c0). For these Euler numbers, we make suitable primary decompositions of the ideals of Vc2 and Vc2,c1,c0, respectively. From the decompositions, we obtain (3.5) Vc2 = L134 ∪ Cone([e1], C0), where C0 is a plane conic defined by C0 := V (x1, ab x2 Cone([e1], C0) is the cone over C0 from the vertex [e1] ∈ P3. Also we have 3 + ab x2x4 − x3x4) in P3 and Vc2,c1,c0 = C0 ∪ L12 ∪ L14 ∪ L34 ∪ {q1,q2}, 11 where the set of two points {q1,q2} is given by the intersection of the plane a2b2 x2− (a5 + b5)x4 = 0 with the (space) conic Q in P3 defined by Q = V(cid:0)a2b2x1 − (a5 + b5)x3, ab x2 3 + abx2x4 − x3x4(cid:1). Proposition 3.5. For (a, b) ∈ (C∗)2 with non-vanishing discriminant (2.9), the topological Euler number of the determinantal quintic X sp,♯ ) = 11 − e(Ds), where e(Ds) is the Euler number of the (reducible) discriminant octic surface. is given by e( X sp,♯ 0 0 Proof. For the numbers e(Vc2) and e(Vc2,c1,c0), it suffices to see the intersections of each component of the respective irreducible decompositions. For the former, we see L134 ∩ Cone([e1], C0)) = L14 ∪ L′1,34, where L′i,jk = hei, ej + ab eki represents the lines generated by the two vectors indicated. Using this, we obtain e(Vc2 ) = e(L134 ∪ Cone([e1], C0)) = e(L134) + e(Cone([e1], C0)) − e(L14 ∪ L′1,34), which we evaluate as 3 + 3− 3 = 3. For the latter e(Vc2,c1,c0), we note that the two points q1 and q2 do not lie on any other components for the values of a, b. We also note C0 ∩ (L12 ∪ L14 ∪ L34) = {[e2], [e4], [e3 + ab e4]}. Looking the configurations of the lines L12 ∪ L14 ∪ L34, we see two intersection points among the lines. Taking into account 3 + 2 = 5 intersection points in total, we finally evaluate the Euler number as e(Vc2,c1,c0) = e(C0 ∪ L12 ∪ L14 ∪ L34 ∪ {q1,q2}) = 2 × 4 + 2 − 5 = 5. The claim follows from the general formula (3.4), i.e., e( X sp,♯ 5. 0 ) = 9 − e(Ds) − 3 + (cid:3) Remark. In Proposition 3.5, we assumed non-vanishing discriminant (2.9) and (a, b) ∈ (C∗)2. However, as we see in the arguments above, Proposition 3.5 holds also for k,l=0(µka + µlb + 1) of the discriminant does not a = b ∈ C∗ as long as the factorQ4 vanish. [] Proposition 3.6. The reducible octic surface Ds = D1 s has the topological number e(Ds) = 21 (resp. 16) for (a, b) ∈ (C∗)2 with non-vanishing discriminant k,l=0(µka + µlb + 1) 6= 0). Hence we have (2.9) (resp. e( X sp,♯ for a = b ∈ C∗ with Q4 ) = −10 and also e(Hsp) = −5. s ∪ D2 0 Proof. We briefly sketch the calculations of e(Ds) in the next section. We evaluate the Euler numbers by e( X sp,♯ (cid:3) 0 ) = 11 − e(Ds). 12 4. Calculations of the Euler numbers e(Ds) This section is devoted to rather technical calculations of the Euler number e(Ds) appeared in Propositions 3.3, 3.5 and 3.6. Our method is essentially based on a similar formula to (3.2) which counts the number of solutions for a given equation. Since the degrees of the relevant polynomial equations become higher, the necessary calculations are more involved than the previous section. For readers' convenience, we briefly summarize the technical details required to do the calculations. 4.1. e(Ds) for Z sp surface Ds is defined as the discriminant of f (t) = c2t2 + c1t + c0 : 2 . Let us consider the determinantal quintic Z sp 2 . The octic Ds : (c2 1 − 4c2c1 = 0) ⊂ P3, with the definitions of ci = ci(x1, x2, x3, x4) as in (3.1). We first note that [0 : 0 : 0 : 1] ∈ P3 is a point on Ds. We then consider an affine line ℓ : [y1; y2 : y3 : t] (t ∈ C) such that ℓ ∪ [0 : 0 : 0 : 1] = P1. As before, we understand that t = ∞ represents [0 : 0 : 0 : 1]. The number of the intersection points ℓ ∩ Ds is determined by the number of solutions of g(t) = 0 with g(t) = d4t4 + d3t3 + d2t2 + d1t + d0, where the coefficients di = di(y1, y2, y3) are read from the defining octic equation of Ds. As in the previous section, we can determine e(Ds) by carefully analyzing the numbers of solutions of the quartic equation g(t) = 0 parametrized by [y1 : y2 : y3] ∈ P2. There may appear several possibilities for the equation g(t) = 0. If d4 6= 0, then g(t) = 0 is an quartic equation which has 4 roots admitting following types of multiple roots: 2 + 1 + 1, 2 + 2, 3 + 1, 4. For each type of the multiple roots, we can determine the corresponding component of the discriminant of g(t) as follows: As an example, consider the case of 2 + 1 + 1, i.e., one double roots and two simple roots. We assume the following forms for g(t): g(t) = d4(t − α)2(t − β)(t − γ) = d4t4 + d3t3 + d2t2 + d1t + d0, and read an ideal in C[α, β, γ, y1, y2, y3] by comparing the coefficients of tk(k = 0, .., 4) in the second equality. Then the elimination ideal in C[y1, y2,y3] determines the Zariski closure of the components where we have multiple roots of type 2+1+1. The loci of the other types of multiple roots can be analyzed in a similar way. Lemma 4.1. For (a, b) ∈ (C∗)2 with non-vanishing (2.8), the equation g(t) = 0 has multiple roots of type 2 + 1 + 1 over the generic points of a plane curve C of degree 9. C is singular at 2 points of A1singularity, 3 points of E6 singularity, and 2 points of E12 singularity (according to Arnold's classification [AGV]). Doing similar calculations, we can stratify the discriminant loci of the equation Incorporating the cases where d4 = 0, we have summarized g(t) = 0 (d4 6= 0). the entire picture of the degeneracies of the solutions for the equation g(t) = 0 in Fig.4.1: For the generic points on the nonic curve C, the equation has the multiplicity 2+1+1 and this changes at special points as shown. Since the equation of C is lengthy, we refrain from writing it here. Over the other components, the multiplicities may be seen from the following forms of the polynomial g(t): Over the coordinate lines he1, e2i,he2, e3i,he3, e1i, respectively, g(t) are given by g(t) = b2y2 3(t−b2y1)2(bt+y2)2. Over the (broken) lines he1, e23i and he3, e12i, respectively, g(t) becomes quadrics of the form a10b2y2 2(y1+by3)2t2(t−b2y1)2, b2y2 3t2 and a10b2y4 2(y2+by3)2t2(y3+bt)2 and b2y2 1y2 2y2 3t2. 1y4 13 2+1+1 0 . e1[ ] E12 2+1+1 E 6 2+2 2 2+1+1 . 0 2+2 x A1 2 + 2 4 * A1 x 2+2 e2[ ] E 6 * 4 2+1+1 2 0 . E 6 2+2 2+1+1 0 . e3[ ] E12 C 2+1+1 Figure 4.1. Several components of the discriminant of g(t) are drawn in P2 = he1, e2, e3i. The singular plane curve C of degree 9 is the main component where we have quartics g(t) = 0 with multiple roots of type 2 + 1 + 1. Over the broken lines, g(t) = 0 becomes quadrics with double roots. Also over the 4 points, indicated by •, the equation g(t) = 0 becomes an identity g(t) ≡ 0. Over the complement of the discriminant in P2, we have quartics g(t) = 0 with only simple roots. See the text for more details. Proposition 4.2. We have e(Ds) = 18. Proof. We first calculate the Euler number of the curve C as e(C) = −10. We can determine this number by representing P2 as the cone from [0 : 0 : 1] over P1 = {z = 0}. Or one can obtain the same number by taking into account the vanishing cycles of the singularities 3× E6, 2× E12 and 2× A1 to the Euler number of smooth plane curve of degree 9: e(C) = 2 − 2g + 3 × 6 + 2 × 12 + 2 × 1 = −10. Now we count the numbers of the solutions g(t) = 0 with forgetting multiplicities from the preceding Lemma and Fig. 4.1. Four solutions are possible only for g(t) being a quartic with only simple roots. This occurs over P2 \(cid:0)C ∪ (5 lines)(cid:1), where 5 lines are those depicted in the figure. The case of three solutions are given over C \ (8 points) as we see in the figure. The case of two solutions occurs over three coordinate lines he1, e2i,he2, e3i,he3, e1i except three points for each, and also two points of A1 singularity on C. The case of one solution occurs over the two broken lines in the figure except two points (•'s ) for each, and also over the two points indicated by ∗ . Over the four points shown by • in the figure, we have g(t) ≡ 0, i.e., the entire P1 as the 'solutions'. For each case above, we evaluate the Euler number of the corresponding loci. Summing up all the cases, we evaluate e(Ds) as e(Ds) = 4(cid:8)e(P2) − e(C) − 5(e(P1) − 3) − 1(cid:9) + 3(cid:8)e(C) − 8(cid:9) + 2(cid:8)3(e(P1) − 3) + 2(cid:9) + 1(cid:8)2 + 2(e(P1) − 3) + 1(cid:9) + 4e(P1) − 3 = 4(3 + 10 + 4) + 3(−18) + 2(−1) + 1 + 8 − 3 = 18. 0 4.2. e(Ds) for X sp,♯ . For this case, we consider again an affine line ℓ : [y1 : y2 : y3 : t](t ∈ C) such that ℓ ∪ [0 : 0 : 0 : 1] = P1. This time we have g(t) = d3t3 + d2t2 + d1t + d0 for the equation g(t) = 0 which determines the intersection ℓ ∩ Ds. Although g(t) looks simpler than the previous section, the stratification of the discriminant of the equation g(t) = 0 turns out to be more complicated. For example, for (a, b) ∈ (C∗)2 with non-vanishing (2.9), we have a singular irreducible 14 (cid:3) curve of degree 9 for the locus of the multiplicity 2 + 1 which intersects with other components at many points in a rather complicated way. For a = b ∈ C∗ with k,l=0(µka + µlb + 1) 6= 0, this irreducible curve split into two smooth cubics and simplifies the stratification slightly. Since the calculations are essentially the same as in the previous subsection, we Q4 omit the details here. After careful analysis, we obtain: Proposition 4.3. We have e(Ds) = 21 (resp.16) for (a, b) ∈ (C∗)2 with non- vanishing (2.9) (resp. for a = b ∈ C∗ withQ4 k,l=0(µka + µlb + 1) 6= 0) . 5. Crepant resolutions X∗0 → X sp 0 To study the resolution of X sp 0 it will be convenient to extend our diagram (2.5) to (5.1) X sp 1 !❈❈❈❈❈❈❈❈ ②②②②②②②② X sp 0 π1 }⑤⑤⑤⑤⑤⑤⑤⑤ Z sp 1 X sp,♯ 0 π2 !❇❇❇❇❇❇❇❇ Z sp 2 X sp 2 }⑤⑤⑤⑤⑤⑤⑤⑤ "❊❊❊❊❊❊❊❊ X sp,♯ 0 , where π1, π2 represent the projections to the first and the second factors of P4× P4, respectively, and Z sp X sp 1 =(cid:8) [z] ∈ P4 det( tzA1 1 =(cid:8) [z] × [λ] ∈ P4 × P4 tzAλ = 0(cid:9). tzA2... tzA5) = 0(cid:9) , 1 and Z sp is singular along 15 lines and so is Z sp 2 by simply exchanging a and wi with b and zi, respectively. 0 = {det(Aλ) = 0} appears twice in 1 may be obtained from Note that the same quintic hypersurface X sp,♯ the diagram. Note also that the defining equation of Z sp Z sp 5.1. Singular loci of Z sp 2 . As introduced in Proposition 2.4, the determi- nantal quintic Z sp 1 . To write down all these 2 lines and their configurations, we denote as before by [ei] the coordinate points of the projective space P4 = he1, e2, ..., e5i, which is the second factor in the product P4× P4. Similarly, we use the notation [ei] for the first factor P4 = he1, ..., e5i of the product P4× P4. For these projective spaces, the coordinate lines are the projective lines spanned by the coordinate points, i.e., hei,eji and hei, eji. More generally, we use the notation hvi,vji, hvi,vj, vki, etc. to describe the projective lines, planes, etc. spanned by the vectors indicated. Using this, we define the following lines: qi = hei, ei+1i, li = hei i+1, ei+2i, qi = hei,ei+1i, li = hei i+1, ei+2i, where we set eij = −a ei + ej, eij = −b ei + ej and the indices i, j = 1, ..., 5 should be read cyclically, i.e., by modulo 5. 15 ! } ! } " Since the quintic Z sp 2 has a rather simple defining equation (2.6), we can derive the following results by using Macaulay2 or Singular [DGPS]: Proposition 5.1. For (a, b) ∈ (C∗)2 with non-vanishing discriminant (2.8), the determinantal quintic Z sp 2 is singular along the lines qi(i = 1, ..., 5) with singularities of type A2, and also singular along li (i = 1, ..., 5) and additional 5 lines (see Remark 5.4) with singularities of type A1. Likewise Z sp 1 is singular along qi of A2 singularity, and singular along li (i = 1, ..., 5) and additional 5 lines of A1-singularities. Proof. These are among the properties described in Proposition 2.4. For the deriva- tions we use the Jacobian criteria and primary decompositions for the corresponding ideals. For each lines, taking local coordinates of the normal bundles, we can de- termine the claimed types of singularities. Since calculations are straightforward, we omit the details. (cid:3) 2 0 1 and Z sp 2 . The map: π2 : X sp 2 0 → Z sp 2 ([w]) of a point [w] ∈ Z sp 0 . As we see in the diagram (5.1), X sp 5.2. Singular loci of X sp is a partial res- olution of both of Z sp is birational since the inverse image π−1 is given by the left kernel of the matrix tz(A1wA2w...A5w) = 0, which is uniquely de- (A1wA2w...A5w), i.e., ([z], [w]) s.t. termined for a generic [w] ∈ Z sp 2 . The birational map π2 has non-trivial fibers over the loci where the matrix has co-rank ≥ 2, and X sp 0 naturally defines a blow-up along these loci introducing the projective spaces spanned by the null spaces. The 0 → Z sp same property holds for the first projection π1 : X sp 1 . Proposition 5.2. The birational map π2 has non-trivial fibers over the 5 coordinate lines qi (i = 1, .., 5), and over the complement of these, this is an isomorphism. The fiber π−1 2 ([ei]) is given by the plane hei i+1, ei+2, ei+3i ≃ P2, and the inverse image π−1 2 (qi) of the line qi, more precisely the closure of the inverse image of qi \ {[ei,], [ei+1]}, is isomorphic to P1 × P1. Similar properties hold also for π1 : X sp 0 → Z sp Proof. By studying the left kernels of matrices (A1wA2wA3wA4wA5w) with [w] ∈ Z sp 2 , it is straightforward to obtain the claimed properties of π2. For the properties of π1, we study the right kernel of matrices ( tzA1 1 , where tzA2 ... tzA5):= t( tA1z tA2z ... tA5z) for simplicity. (cid:3) we use a convention ( tzA1 tzA2 ... tzA5) with [z] ∈ Z sp 1 ([ei]) = hei,i+1, ei+2, ei+3i. 1 with π−1 The Jacobian criterion for the complete intersection X sp 0 is rather involved, since we need to handle large ideal. In our case, however, we can utilize the properties of the partial resolutions π1and π2 efficiently. For example, we can deduce that the singular loci of X sp 1 (resp. Z sp 2 ) under π1(resp. π2) (see Proposition 5.1). Combining this fact with the Jacobian criterion for X sp Proposition 5.3. For (a, b) ∈ (C∗)2 with non-vanishing discriminant (2.4), the complete intersection X sp 0 0 must be in the inverse images of the 15 lines in Z sp is singular along the following 20 lines: 0 , we obtain the following: (5.2) Qi =(cid:8)[a2s ei + (s + bt) ei+1,i+2] × [sei + tei+1] [s, t] ∈ P1(cid:9) , Qi =(cid:8)[sei + tei+1] × [b2s ei + (s + at) ei+1,i+2] [s, t] ∈ P1(cid:9) , Li = [ei] × hei,i+1, ei+2i, Li = hei,i+1, ei+2i × [ei], (i = 1, ..., 5), 16 ~ sp X 0 2π sp Z 2 e~ 23 Q1 e~ 5 e~ 34 e~ 4 Q5 e~ 4 ~ e12 ~ Q 1 ~ L1 e~ 3 L1 e~ 1 e~ 2 e~ 5 q1 e1 e2 1l π ~ 2(Q1) q2 e3 ~ Q 5 e~ 51 e~ 3 e~ 2 e~ 45 e~ 1 q5 e5 e4 Figure 5.1. The blow-up π2 : X sp 2 . Bold lines and broken lines upstairs are lines with A1 singularity. Not all broken lines are drawn on the upstairs. 0 → Z sp 1 (qi). Li and Li are the proper transforms of the where Qi ⊂ π−1 lines li and li under π2 and π1, respectively. The singularities along these 20 lines are of A1 type for all, and these lines intersect at 20 points. 2 (qi), Qi ⊂ π−1 Remark 5.4. Now we may restate Proposition 5.1 as follows: For (a, b) ∈ (C∗)2 with non-vanishing discriminant (2.4), the determinantal quintic Z sp is singular 2 along the lines qi(i = 1, ..., 5) with singularities of type A2, and also singular along li and π2( Qi) (i = 1, ..., 5) with singularities of type A1. These 15 lines intersect at is singular along 15 lines qi, li and π1(Qi) intersecting at 15 points. Similarly, Z sp 1 15 points. We have depicted the schematic picture of the blow-up π2 : X sp in Fig.5.1. The intersection points should be clear in this figure. 0 → Z sp 2 We note that the structure of singularities in X sp 0 is quite similar to that of the Barth-Nieto quintic [BaN], where we see 20 lines of A1 singularity intersecting at 15 points, in addition to 10 isolated ordinary double points (called Segre points). In the case of the Barth-Nieto quintic, the local geometries near 15 intersection points are all isomorphic. In our case of the complete intersection X sp 0 , which is a partial resolution of the determinantal quintics Z sp 2 , the 20 intersection points of the 20 lines (Qi, Qi, Li, Li) split into two isomorphic classes as we see below. Also we see in the next sub-section a new isomorphic local geometry near the infinity points of the 10 lines Li, Li (see Fig. 5.2). 1 and Z sp 17 ~ X sp 0 t ~ Q 1 ~ e1 Q 1 t π1 e~ 12 ~ l 1 U 1 U 2 e~ 2 U 3 e~ 3 e~ 4 e~ 5 Z sp 1 e~ 12 U 1 s Q 5 ~ L1 u U 2 s U 3 π2 e~ 3 e1 u e~ 4 sp 2 Z Figure 5.2. The local affine coordinates U1, U2 and their coor- dinate axes along singular lines with the projections to Z sp 1 and Z sp 2 . 5.3. Blowing-ups of X sp 0 . There are two types of the intersections among the 20 singular lines in X sp 0 : 1) the point where 3 lines of the singularities meet, 2) the point where 2 lines meet. This should be clear from a careful inspection of Fig.5.1 and also from the symmetry of the defining equations. We denote by (3A1,U1) and (2A1,U2), respectively, the local geometries around the points of type 1) and 2). These may be summarized as follows: (3A1,U1) around Qi ∩ Li ∩ Qi−1 and Qi ∩ Li ∩ Qi−1 (i = 1, ..., 5), (2A1,U2) around Li ∩ Qi and Li ∩ Qi (i = 1, ..., 5). Also, from the reasons which will become clear soon (in the proof of Proposition 5.9, 2)), we need to study (isomorphic) local geometries around the points [ei+2] × [ei] on Li and [ei] × [ei+2] on Li (i = 1, ..., 5), which we denote by (∂A1,U3). It will be helpful to list the relevant local geometries on each lines as follows: (5.3) (∂A1,U3), (2A1,U2), (3A1,U1) on each Li, Li, (3A1,U1), (3A1,U1), (2A1,U2) on each Qi, Qi. In the following arguments, we will focus on the point Q1∩ L1∩Q5 = [e12]×[e1] for (3A1,U1), and Q1 ∩ L1 = [−a e12 + e3]× [e1] for (2A1,U2). We will also focus on the point [e3] × [e1] on L1 for (∂A1,U3). See Fig.5.2. Also, in the following arguments in this subsection, we assume non-vanishing discriminant (2.4) and ab 6= 0 for the parameters a and b. 5.3.1. Resolution of (3A1,U1). We choose an affine coordinate (s, t, u, v, ω2,ω3,ω4,ω5) so that s, t and u, respectively, coincide with the local parameters of the curves Q5, Q1 and L1, and also the origin represents the point [e12] × [e1] = [e1 − 1 a e2] × [e1]. For this, we use the parametrization given in (5.2). Explicitly, we write the points 18 e23 e4 e12 4e z y e5 e34 e5 e~ 2 ~ L1 e3 e3 e45 e1 π1 s' e1 U3 2e x s' e~ 1 U3 x e~ 3 e~ 4 51e 2e ~ sp X0 e~ 5 Z sp 1 e~ 23 e~ 4 = ~ e12 ~ L1 s' e~ 3 z U3 e~ 4 x y e~ 1 e~ 3 e~ 2 2π e1 U3 e~ 51 e~ 2 e~ 3 e~ 4 ~ sp X0 e~ 45 e~ 1 y e5 z e3 sp Z 2 e4 e~ 5 e~ 34 e~ 5 e2 Figure 5.3. The local geometry (∂A1,U3) with affine coordinates. on Q5 by In order to see the local geometry about the origin, we work in the local ring C[s, t, u, v, ω2, .., ω5]m0 with respect to the maximal ideal m0 of the origin. Writing the defining equations of X sp in this ring, it is straightforward to see that the three 0 equations (1st, 2nd and 5th equations in (2.3)) may be solved as ω2 = ω5 = 0 and ω3 = − a3tu b2(1−at) . After substituting these into the remaining equation, we obtain b3(1 − at)uω4 + at(u + av)(1 − at − a2u) = 0, ω4(a + s)(as + v) − b2sv = 0. Setting ω4 = w, and focusing on the property near the origin, we have: Proposition 5.5. The local geometry (3A1,U1) near the singular point [e12] × [e1] is represented by the germ ({g1, g2}, C5) near the origin (s, t, u, v, w) = (0, ..., 0) with g1 = a t(u + a v) + b3uw, g2 = aw(v + a s) − b2sv. Remark. The coordinate w = w4 has a special meaning related to the blow-up π1 : X sp 0 → Z sp 1 . In fact, in our affine coordinate (s, t, u, v, ω2,ω3,ω4,ω5), the exceptional 19 ha2s e5 + (s + b)e12i ×hse5 + e1i =he1 − =he1 − 1 a 1 a as s + b e2 − e5i×he1 + se5i e5i, s + a bs e2 + se5i ×he1 − s+b in the middle to s using Aut(P1). Similarly, we can 1 a where we have changed − as parametrize the points on Q1 and L1, respectively, by e2 + te2i ×he1 − he1 − a(cid:1)e2+ue3+ve4+se5i×he1+(cid:0)ω2− he1+(cid:0)t− b2 t e3i , he1 − t(cid:1)e2+(cid:0)ω3+ t e2 + a b a b a 1 Introducing additional parameters v, ω2, ω3, ω4, ω5, we take an affine coordinate of C4 × C4 ⊂ P4 × P4 by 1 a e2 + ue3i ×he1i. a b2 t(cid:1)e3+ω4e4+(cid:0)ω5− bs s + a(cid:1)e5i. he1 − 1 a e2 + te2i ×he1 − a b t e2 + a b2 t e3 + ω4e4i (t, ω4 ∈ C). divisor over the line q1 can be written as Based on this, after eliminating the variable w from the local equations by {g1, g2}, we have a germ (g3,C4) near the origin with g3 = a2 t(u + a v)(v + a s) − b5suv = a3 stu + a4 stv + b5 suv + a2 tuv + a3 tv2, where the polynomial g3 coincides with the lowest oder terms of the defining (quintic) polynomial Z sp represented by the local parameters (1, z2, z3, z4,z5) = 1 a , u, v, s). The A1-singularity along Q1 is the partial resolution of the A2- (1, t − 1 singularity along q1 in Z sp 2 . [] Let us consider the blowing-up C5 → C5 at the origin of the local geometry ({g1, g2}, C5), and denote the exceptional divisor by E1. E1 is the surface {g1 = g2 = 0} considered in P4 with the homogeneous coordinate [S : T : U : V : W ] corresponding to (s, t, u, v, w). Proposition 5.6. E1 is a singular del Pezzo surface of degree four with three nodal points, and has the Euler number e(E1) = 5. Proof. The equations g1 = g2 = 0 in P4 define a del Pezzo surface of degree 4. By evaluating the Jacobian ideal, it is immediate to see that this is singular at [S : T : U : V : W ] = [1 : 0 : 0 : 0 : 0], [0 : 1 : 0 : 0 : 0] and [0 : 0 : 1 : 0 : 0], where the exceptional divisor E1 intersects with the s-, t- and u-axes of A1-singularities. Since E1 is a singular Pezzo surface of degree 4 with three ordinary double points, it can be given as P2 blown-up at 5 points and then contracting three (−2) curves [HW]. Therefore we have e(E1) = 8 − 3 = 5. Proposition 5.7. After the blowing-up (3A1,U1) at the origin, the three singular lines separate from each other and intersect with E1 at the three nodal points. Proof. We have chosen our coordinate of U1 so that s-, t-, u-axes coincide with the lines of A1-singularity. The blowing-up C5 → C5 at the origin introduces the exceptional set P4, which separate the coordinate axes. Hence the blowing-up separates the s-, t-, u-axes of A1-singularity from each other with introducing the exceptional divisor E1. The intersection points of the (proper transforms of the) s-, t-, u-axes with E1 coincides with the three nodal points of E1. (This is similar to the case of the Barth-Nieto quintic [BaN].) (cid:3) (cid:3) 0 → X sp 0 . Also we denote by Q(1) Now, we blow-up all the 10 local geometries of type (3A1,U1) at their origins, and , L(1) denote the blow-ups by ϕ1 : X sp,(1) the proper transforms of the 20 lines Qi, Qi, Li, Li of A1 singularity, respectively. Along these proper transforms of lines, we still have A1 singularities. Also, these lines intersect at the origins of the 10 isomorphic local geometries of type (2A1,U (1) 2 ), 0 . We also denote by (∂A1,U (1) 3 ) ∼= (∂A1,U3) which is isomorphic to (2A1,U1) in X sp , L(1) the 10 isomorphic local geometries near the infinity points of the lines L(1) (see Fig. 5.2). The local geometries on each lines are now summarized as , Q(1) , L(1) i i i i i i (5.4) (∂A1,U (1) 3 ), (2A1,U (2) (2A1,U (1) 2 ) on each L(1) 1 ) on each Q(1) i i i , L(1) , , Q(1) i . 20 and also the line Q(1) 1 by (cid:2)a2e1 + (1 + t)e23(cid:3) ×he1 + Taking these forms into account, we introduce the affine coordinate by 1 b te2i =(cid:2)a2e1 − a(1 + t)e2 + (1 + t)e3(cid:3) ×he1 + t b e2i. 5.3.2. Resolution of (2A1,U (1) 2 ). As in the previous case, we choose an affine coor- dinate (s, t, u, v, ω2, ω3, ω4, ω5) centered at [−a e12 + e3] × [e1] with s, t being along in X sp,(1) the lines L(1) . For this we parametrize the line L(1) 1 , Q(1) 1 1 by 0 (cid:2)−ae12 + (1 + s)e3(cid:3) × [e1] =(cid:2)a2e1 − ae2 + (1 + s)e3(cid:3) × [e1], (cid:2)a2e1 − a(1 + t)e2 + (1 + s + t)e3 + ue4 + ve5(cid:3) b(cid:1)e2 + ω3e3 + ω4e4 + ω5e5i. ×he1 +(cid:0)ω2 + t In the local ring C[s, t, u, v, ω2, .., ω5]m0 , four of the five defining equations of X sp,(1) may be solved with respect to ω2, ω3, ω4, ω5 and one equation leftover determines the germ about the origin. Proposition 5.8. The local geometry (2A1,U (1) 2 ) near the singular point [−a e12 + e3] × [e1] is represented by a germ (h, C4) near the origin (s, t, u, v) = (0, 0, 0, 0) with 0 (5.5) h = b5uv + a3 stu + a4 stv + b5suv + 2b5tuv. 0 1 0 → Z sp (composed with the blow-up X sp,(1) We may derive the same form directly from the quintic equation of Z sp 1 since → X sp the projection π1 : X sp 0 ) defines an isomorphism on the neighborhood U2, see Fig. 5.2. In the figure, the geometric meaning of the parameters u and v should be clear. By our choice of the coordinates, we have A1-singularities along s- and t-axes, i.e., along the lines L(1) 1 and Q(1) 1 , respectively. We will consider the blowing-up along L(1) 1 , which is locally described by the blowing-up C × C3 → C × C3 along the s-axis. Proposition 5.9. 1) The exceptional divisor of the blow-up C × C3 → C × C3 of (2 A1,U (1) 2 ) along the s-axis is a conic bundle over s ∈ C (s ≪ 1), which has a reducible fiber over s = 0. This conic bundle is singular only at an ODP over s = 0. 2) The conic bundle over C extends to a conic bundle E2 → L(1) 1 ∼= P1, which has reducible fibers over s = 0 and s = ∞. This conic bundle is singular only at an ODP over s = 0, and also admits a section. 3) After the blowing-up of (2A1,U (1) 2 ), the singularity leftover near the local ge- ometry is the A1 singularity along the proper transform of the t-axis. The proper transform of Q(1) 1 Proof. 1) We introduce the coordinate (s, [T, : U : V ]) for the exceptional set C× P2 of the blow-up C × C3 → C × C3. Then from the local equation of (2A1,U2), we have the equation of the exceptional divisor as intersects with the conic bundle E2 at the ODP over s = 0. b5U V + a3sT U + a4sT V + b5sU V = 0. This defines a family of conics in P2 over s ∈ C (s ≪ 1), which is reducible at s = 0. Also we see that the conic bundle is singular only at an ODP over s = 0. 2) To see the geometry of the exceptional divisor over C(= P1 \ {s = ∞}), we need to have the equation (5.5) in all order in s but with homogeneous of degree 21 two for t, u, v. It is easy to have the equation from the defining equation of Z sp 1 . After some algebra, we have the equation for the exceptional divisor: (5.6) (s + 1)(cid:0)b5U V + a3sT U + a4sT V(cid:1) = 0, which defines a conic bundle over s ∈ C (s 6= −1) with only one singular fiber over s = 0. We see that s = −1 correspond to the intersection point of the exceptional divisor E1 and the s-axis. Since this intersection point is one of the three nodal points on E1, we see that the conic bundle extends to s = −1 with smooth fiber over it. The point s = ∞ in the s-axis corresponds to the center of the local geometry (∂A1,U (1) s , x, y, z to represent the relevant lines in this geometry, see Fig. 5.3. With other parameters v1, v5 and ω2, ω4, we consider the following affine coordinate centered at [e3]× [e1] of P4 × P4 : 3 ). We introduce the local parameters s′ = 1 [(v1 − as′)e1 + s′e2 + e3 + xe4 + v5e5]× [e1 + ω2e2 + (b2y − z)e3 + (ω4 − by + z)e4 + ye5]. Writing the defining equations (2.3) in this coordinate, we can solve four equations with respect to v1, v5, ω2, ω5 to obtain one equation abxz+s′(bxz+a4yz−a4by2) = 0 which describes the local geometry (∂A1,U (1) 3 ) near the origin. Now we have the following local equation of the exceptional divisor of the blowing up along s′-axis: abXZ + s′(cid:0)bXZ + a4Y Z − a4bY 2(cid:1) = 0 (s′ ≪ 1), where (s′, [X, Y, Z]) represents the coordinates of the exceptional set C × P2 of the blow-up C× C3 → C× C3. From this equation, we see that the exceptional divisor is a conic bundle with reducible fiber over s′ = 0 (s = ∞) but smooth for s′ ≪ 1. Finally, from the equation (5.6), we see that U = V = 0, for example, gives a section. 3) Let (s, t, u, v) = (s, t, U T ) be the one of the affine coordinates of the blow- up. Then we have u = ut, v = vt. Substituting these into the local equation h of (2A1,U2), i.e., for s,t ≪ 1, we obtain T , V h = b5 uv + a3su + a4sv + b5suv + 2b5tuv with h = t2h. If we set t = 0, then we have the equation of the exceptional divisor (s ≪ 1) above. When we set s = 0, then we have h = uv(b5 + 2b5t). This shows that the ODP of the exceptional divisor E2 over s = 0 merges to the A1-singularity along the proper transform of the t-axis ( see Fig. 5.4). Since the singularity along 1 ∩ L(1) the line Q(1) 1 , we 1 now see that, near t = 0, the singularity along the proper transform of Q(1) is of 1 A1-type. (cid:3) is of A1-type except t = 0, i.e., at the intersection Q(1) All the intersections of Q(1) and L(1) ( Q(1) i i i 2 ). We blow-up along all the 10 lines L(1) i and L(1) i ) have the local geometries , and . We denote the proper transforms and L(1) i isomorphic to (2 A1,U (1) denote the blow-ups by ϕ2 : X sp,(2) of the 10 lines Q(1) and Q(1) 0 i → X sp,(1) 0 , respectively, by Q(2) i and Q(2) i . i 22 i i . and Q(2) 0 . We finally construct a crepant resolution. 5.3.3. Crepant Resolution X∗0 → X sp Proposition 5.10. 1) All the singularities of X sp,(2) 10 lines Q(2) 2) The singularities along Q(2) of Q(2) which is a P1-bundle with a section. Proof. 1) Since each line of Qi and Qi intersects with others at the center of the local geometry (3A1,U1), it is clear that Q(2) are separated after the blow-ups (see Proposition 5.7). are of A1-type. Blowing-up along each line resolves the singularity with introducing the exceptional divisor E3 are along the non-intersecting and Q(2) 1 and Q(2) and Q(2) 0 i i i i i 0 in X sp,(2) 2) By symmetry, it suffices to show the properties for a line, say Q(2) 1 . Note that Q(2) is given by the successive proper transform of the line Q1 in X sp 0 1 under the blow-ups of the local geometries, two (3A1,U1)'s and one (2A1,U2) on the line. Therefore, the local geometry around Q(2) is isomorphic to that around 1 the line Q1 except the three centers of the blowing-ups on the line. We further note that the local geometry around the line Q1 except the three centers is pro- jected isomorphically to Z sp 2 . The local geometry around π2( Q1) is easily analyzed by introducing the following affine coordinate: 2 under the partial resolution π2 : X sp 0 → Z sp [w1 : w2 : ... : w5] = [e1 + (u − bt)e2 + te3 + ve4 + we5], where t parametrizes the line π2( Q1). Substituting this into the defining equation (2.6) of Z sp 2 and taking the polynomial of homogeneous degree up to two with respect to u, v, w but all for t, we obtain (5.7) bt(cid:8)a5tvw − (1 − b2t)u(v + bw)(cid:9) = 0, which shows A1-singularity along the t-axis except t = 0, 1 b2 and ∞. These three values exactly correspond to the two local geometries (3A1,U1)'s and one (2A1,U2) on the line Q1, whose blowing-up we studied in Proposition 5.7 and Proposition 5.9. Combined with the results there, we conclude that the singularity along Q(2) 1 is of A1-type, and it is resolved by the blowing-up along the line with introducing an exceptional divisor E3 which is isomorphic to a P1-bundle over the line. Also from the equation (5.7), it is easy to see that E3 has a section (cf. Proposition 5.9 2) ). (cid:3) Let us now denote the blowing-up along the 10 lines by ϕ3 : X sp,(3) → X sp,(2) . , we may summarize the whole process of the blowing-ups 0 0 Defining X∗0 := X sp,(3) by 0 ϕ : X∗0 = X sp,(3) X sp,(2) 0 −→ϕ2 represents the composition. −→ϕ3 0 X sp,(1) 0 X sp 0 , −→ϕ1 0 where ϕ : X∗0 → X sp Theorem 5.11. For (a, b) ∈ (C∗)2 with non-vanishing discriminant (2.4), the blowing-up ϕ : X∗0 → X sp is a crepant resolution and gives a smooth Calabi-Yau manifold with the Euler number e( X∗0 ) = 2(h1,1( X∗0 ) − h2,1( X∗0 )) = 100. Proof. For the proof of K X ∗ , we show the existence of a nowhere vanishing holomorphic 3-form explicitly, although an abstract argument is possible. We first 0 ∼= O X ∗ 0 0 23 consider the blow-up, ϕ1 : X sp,(1) 0 at the origin of the local geometries (3A1,U1). As before we introduce the affine coordinate (s, t, u, v, w2,··· , w5). We start with the standard form of a nowhere vanishing holomorphic 3-form Ω( X sp 0 ) for the complete intersection Calabi-Yau variety X sp 0 given in (6.1). In this affine coordinate, we have → X sp 0 Evaluating the Jacobian ∂(w2,w3,w5) Ω( X sp 0 )U1 = Resf1=···=f5=0(cid:16) ds ∧ dt ∧ du ∧ dv ∧ dw2 ∧ ··· ∧ dw5 (cid:17), Resg1=g2=0(cid:16) ds ∧ dt ∧ du ∧ dv ∧ dw 0 )U1 = −1 ∂(f1,f2,f5) = f1f2f3f4f5 Ω( X sp b2(a+s)(1−at) , we calculate the residue as (cid:17). g1g2 −a a (5.8) where w = w4 and g1,g2 are given in Proposition 5.5 (precisely g1,g2 here contain all higher order terms, but this does not affect the following arguments). Consider the blow-up ϕ1 : C5 → C5 at the origin, and one of the affine coordinate (s, t, u, v, w) = S ) with t = ts, u = us, v = vs, w = ws. Then, pulling back the 3-form, (s, T it is immediate to have S , W S , U S , V (5.9) ϕ∗1Ω( X sp = −1 a Resg1=g2=0(cid:16) ds ∧ dt ∧ du ∧ dv ∧ d w g1g2 (cid:17), 0 )(cid:12)(cid:12)(cid:12)U (1) 1 1 0 1 0 0 0 0 0 0 ) of X sp,(1) = ϕ∗3Ω( X sp,(2) 0 ) coincides with Ω( X sp,(1) )(cid:12)(cid:12)U (2) )(cid:12)(cid:12)U (3) it is straightforward to see that Ω( X sp,(2) where g1 = s2g1, g2 = s2g2 and g1 = g2 = 0 is the defining equation of the blow-up. Up to the non-vanishing constant, the right hand side is the holomor- phic 3-form Ω( X sp,(1) . Calculations are similar for other affine coordi- nates, and we see that the pull-back ϕ∗1Ω( X sp ), i.e., ϕ1 is crepant. The next step ϕ2 : X sp,(2) → X sp,(1) has an effect on (5.9) as the blowing- 0 up along the s-axis. Again, = ϕ∗2Ω( X sp,(1) holds up to a non-vanishing constant on all the affine coor- dinates. Doing similar calculations for the blow-up ϕ3, we finally verify that Ω( X sp,(3) . Thus near the 10 points of the local geome- try (3A1,U1), we see that ϕ : X∗0 → X sp For the local geometry (2A1,U2), since the first blow-up ϕ1 has no effect, we start with Ω( X sp,(1) . As in the previous subsection, we introduce the affine coordinate (s, t, u, v, w2, w3, w4). Evaluating the Jacobian ∂(w2,w3,w4,w5) ∂(f1,f2,f3,f5) , we have )(cid:12)(cid:12)U (3) 0 )(cid:12)(cid:12)U2 Resh=0(cid:16) ds ∧ dt ∧ du ∧ dv where h is given in (5.5) (again, precisely h should be understood with the higher order terms). Then ϕ2 is the blow-up along the s-axis, see Proposition 5.9. Using one of the affine coordinate of the blow-up, (s, t, u, v) = (s, t, U T ) with u = ut, v = vt, we evaluate the pull-back as )(cid:12)(cid:12)U2 0 )(cid:12)(cid:12)U2 )(cid:12)(cid:12)U (2) = −1 a = Ω( X sp is crepant. Ω( X sp (cid:17), T , V h 0 0 0 1 1 ϕ∗2Ω( X sp,(1) 0 with h = t2h. Since h = 0 is the local equation of the blow-up X sp,(2) that Ω( X sp,(2) blow-up ϕ3 is along the t-axis, and this is done locally by (s, t, u′, v′) = (s, t, , we see up to a non-vanishing constant. The next = ϕ∗2Ω( X sp,(1) 0 0 0 U S , V S ) )(cid:12)(cid:12)U (2) 2 )(cid:12)(cid:12)U (2) 2 (cid:17), h = −1 a Resh=0(cid:16) ds ∧ dt ∧ du ∧ dv )(cid:12)(cid:12)U (2) 2 24 E3 E3 (3) E1 ϕ 3 E3 (3) E2 (2) Q 1 ~ (2) Q 1 (2) E1 E 2 (2) Q 5 ϕ 2 ~ (1) Q 1 (1) Q 5 ϕ 1 E1 ~ Q 1 Q 5 (1) Q 1 ~(1) L1 Q 1 ~ L1 Figure 5.4. Exceptional divisors E1, E2, E3 of the blowing-ups ϕ1, ϕ2, ϕ3, respectively. Only the local geometries around the line L1 in X sp 0 are depicted. 2 0 0 2 0 0 is a )(cid:12)(cid:12)U (3) 0 = ϕ∗3Ω( X sp,(2) )(cid:12)(cid:12)U (3) ) = 0. For the property i), we note that K X sp with u = u′s, v = v′s. The local equation of the blow-up is given by h′ = 0 with h = s2h′, and we have Ω( X sp,(3) , up to a non-vanishing constant. From the local equation h′ = 0, we see that X sp,(3) = X∗0 is smooth. The calculations are valid for all the 10 points of the local geometry (2A1,U2). crepant resolution. Combined with the results for (3A1,U1), we conclude that ϕ : X∗0 → X sp 0 ∼= O X ∗ Next we show that X∗0 is a Calabi-Yau manifold, namely, i) K X ∗ and 0 ∼= O X sp ii) h1(O X ∗ ) = h2(O X ∗ since X sp is a complete intersection of 5 divisors of (1, 1)-type in P4 × P4. Then 0 0 ∼= O X ∗ is immediate since ϕ is crepant. For the second ii), we K X ∗ note that all the higher direct images Riϕ∗O X ∗ (i > 0) vanish by the Grauert- Riemenschneider vanishing since ϕ is crepant. Then, by the Leray spectral se- ) ∼= H i(O X sp quence, we have H i(O X ∗ ) (i = 1, 2). Hence we have only to show that the r.h.s vanishes. Note that X sp 0 (1, 1)-type in P4 × P4, and consider the following Koszul resolution of O X sp 0 → OP4×P4(−5,−5) → OP4×P4(−4,−4)⊕4 → OP4×P4(−3,−3)⊕10 → OP4×P4(−2,−2)⊕10 → OP4×P4(−1,−1)⊕5 → OP4×P4 → O X sp 0 → 0. is a complete intersection of 5 divisors of = ϕ∗K X sp : 0 0 0 0 0 0 0 0 0 0 0 ) ∼= H 0(OP4×P4) ∼= C and H i(O X sp For the calculation of the Euler number, let us first note that we have e( X sp , all the cohomology groups As for the sheaves in this exact sequence except O X sp vanish except H 5(OP4×P4(−5,−5)) ∼= C and H 0(OP4×P4) ∼= C by the Kodaira ) ∼= vanishing theorem and the Serre duality. Now it is standard to see that H 2(O X sp H 5(OP4×P4(−5,−5)) ∼= C, H 0(O X sp ) (i = 1, 2) vanish. 2 ) + 5 ×(cid:0)e(P2) − 1(cid:1) = −10 + 10 = 0. This follows form Proposition 3.4 and e(Z sp Proposition 5.2, see also Fig. 5.1. Now we note that, under the blow-up, the origin of (3A1,U1) is replaced by the exceptional divisor E1 with its Euler number e(E1) = 5. Similarly for (2 A1,U (1) 2 ), one line is replaced by a conic bundle E2 over P1 with two reducible fibers, hence e(E2) = 6. Since we have 10 isomorphic geometries for (3A1,U1) and 10 for (2 A1,U (1) 2 ), taking into account the final blow- ups of 10 lines, we evaluate the Euler number e( X∗0 ) as 0 ) = 0 0 e( X∗0 ) = 10 ×(cid:0)e(E1) − 1(cid:1) + 10 ×(cid:0)e(E2) − e(P1)(cid:1) + 10 ×(cid:0)e(E3) − e(P1)(cid:1) = 40 + 10 × (6 − 2) + 10 × 2 = 100. 25 (cid:3) 0 0 0 2 0 0 1 0 0 in X sp,(3) → X sp → X sp,(1) 0 , ϕ2 : X sp,(2) 5.4. Hodge numbers. Recall that the crepant resolution is obtained as the com- posite of the blowing-ups ϕ1 : X sp,(1) , ϕ3 : X sp,(3) → . The first blow-up ϕ1 introduces the exceptional divisors E1(=: E(1) X sp,(2) 1 ) in X sp,(1) which is a del Pezzo surfaces of degree 4 with three lines are contracted to three points. One of the three points is resolved in the proper transform E(2) under ϕ2, and the other two are resolved in the proper transform E(3) 1 under ϕ3. Similarly, the resolution ϕ2 introduces the conic bundle E2(=: E(2) 2 ) over P1 which has an ordinary double point (over s = 0), and ϕ3 resolves this singularity to have smooth ruled surface E(3) . The final blow-up ϕ3 introduces the divisor E3 = E(3) 3 which is a P1-bundle over P1 with a section. Note that all these divisors 1 , E(3) E(3) In this subsection, following [HSvGvS], we apply the Weil conjecture to deter- mine the Hodge numbers of the resolution X∗0 . We set our parameters to a = b = 1 and consider the mod p reduction of X∗0 . We write Fp = Z/pZ. Lemma 5.12. For all but finite primes, the reduction of X∗0 modulo p is smooth over Fp. Proof. The smoothness of X∗0 in the tori (C∗)4×(C∗)4 follows from the discriminant (in Proposition 2.3) dis( X sp 0 (C∗)4) = 3× 113 for a = b = 1. The exceptional divisors E1, E2 of the blowing-ups ϕ1 and ϕ2, respectively, are blown-up to smooth surfaces 1 and E(3) E(3) in X∗0 , hence the resolution X∗0 is smooth over Fp except finite primes p. are smooth in X sp,(3) and E(3) = X∗0 . (cid:3) 2 0 2 3 Let X∗0 (Fp) be the set of points in X∗0 which are rational over Fp. We use the Lefschetz fixed point formula due to Grothendieck, (5.10) with tj = tr (Frob∗p(cid:12)(cid:12)H j # X∗0 (Fp) = 1 − t1 + t2 − t3 + t4 − t5 + t6, ´et( X∗0 , Qℓ)) and Frobp : X∗0 → X∗0 the Frobenius morphism. Since X∗0 is a Calabi-Yau threefold, we have t0 = 1, t1 = t5 = 0, t6 = p3. By the Weil conjecture (see [Har, Appendix C] for example), the eigenvalues of Frobp on H j ´et( X∗0 , Qℓ) are algebraic integers, which do not dependent on ℓ, with absolute values pj/2. Also, by the Weil conjecture again, tj's are (ordinary) integers and satisfy tj ≤ bj( X∗0 ) pj/2. We derive the following property following the arguments in [HSvGvS, Prop. 2.4] made for the Barth-Nieto quintic. Proposition 5.13. For every good prime p, all eigenvalues of Frobp on H 2 are equal to p. ´et( X∗0 , Qℓ) Proof. Due to Lemma 5.14 below, we can use the Lefschetz hyperplane theorem [FK, Corollary I.9.4] and have the claimed property for X sp 0 . Then from the Leray spectral sequence associated to ϕ1 : X sp,(1) 0 , we obtain the claimed property for X sp,(1) , we use the Leray spectral sequence associated to ϕ2 : X sp,(2) (see [HSvGvS, Lemma 2.16]). To go further to X sp,(2) 0 0 0 → X sp → X sp,(1) , , R2−jϕ2∗(Qℓ)) ⇒ H 2 2 = 0 and E0,2 0 0 26 ´et( X sp,(1) 0 Ej,2−j 2 = H j ´et( X sp,(1) 0 ´et( X sp,(2) 0 2 = H 2 , Qℓ), ´et( X sp,(1) where E2,0 0 Due to Lemma 5.15 below, we have the claimed property for E0,2 , Qℓ), E1,1 2 = H 2 , R2ϕ2∗(Qℓ)). as well as E2,0 2 , 2 0 ´et( X sp,(2) , Qℓ), too. To go from X sp,(2) hence for H 2 the argument in [ibid, Lemma 2.16] since the exceptional divisor E3(= E(3) ϕ3 : X sp,(3) 0 for H 2 = X∗0 , we can use 3 ) of is a P1-bundle over P1. Thus we obtain the claimed property (cid:3) → X sp,(2) to X sp,(3) 0 0 0 ´et( X∗0 , Qℓ). Lemma 5.14. Consider X sp 0 as the linear section (P4× P4)∩H1∩...∩H5 in P24 by the Segre embedding with Hk representing the defining equation fk = tzAkw (k = 1, ..., 5). Then for all but finite primes p, there exists a sequence linear forms H′1, H′2, ..., H′5 over Z with the following properties over Fp: 1) Sing(Xi−1)⊂ Xi holds for i = 2, .., 5, where Xi = (P4 × P4) ∩ H′1 ∩ .. ∩ H′i and Sing(Xi−1) is the singular loci of Xi−1. 2) X5 = X sp 0 . Proof. Since fk's are defined over Z, it suffices to have the properties 1) and 2) over C. We can verify explicitly that the sequence H′1, H′2, ..., H′5 corresponding to f1 + f3 + f5, f2 + f4, 3f2 + f5, 5f3 + f4, f5 satisfies the desired properties over C. (cid:3) Lemma 5.15. All the eigenvalues of Frobp on H 0 to p for every good prime p. ´et( X sp,(1) 0 , R2ϕ2∗(Ql)) are equal . First note that H 0 Proof. Set ρ1 := ϕ2E(2) Let ρ2 : E(3) fiber of E(2) ρ2 and ρ1. We have the spectral sequence: ´et(E(2) be the blow-up of the ordinary double point of E(2) 2 → E(2) on the 2 → P1 over s = 0 (see Proposition 5.9). Denote by ρ the composite of , R2ϕ2∗Ql) ≃ H 0 ´et( X sp,(1) 2 , R2ρ1∗Ql). 2 0 2 2 (5.11) Ei,j 2 := Riρ1∗(Rjρ2∗Ql) =⇒ Ri+jρ∗(Ql). By standard calculations, we have • E2,0 2 = R2ρ1∗(Ql). • Since the nontrivial fiber of ρ2 is a P1, we have R1ρ2∗(Ql) = 0. Hence E1,1 2 = 0. • E0,2 2 ≃ H 2 ´et(P1, Ql), where P1 is the nontrivial fiber of ρ2, and we consider ´et(P1, Ql) as a skyscraper sheaf supported on s = 0. H 2 Then, by standard properties of the spectral sequence, we have the following exact sequence: 0 → R2ρ1∗(Ql) → R2ρ∗(Ql) → H 2 ´et(P1, Ql) → 0. Therefore, to show the claimed property for H 0 to show that the claimed property holds for H 0 0 ´et( X sp,(1) ´et(E(2) 2 , R2ρ∗(Ql)). , R2ϕ2∗(Ql)), we have only Let ρ3 : E(3) 2 → E′2 be the contraction of three (−1)-curves on E(3) 2 , two of which are the strict transforms of the components of the fiber of E(2) 2 → P1 over s = 0, and the remaining one of which is one component of the fiber of E(2) 2 → P1 over s = ∞ (see Proposition 5.9). Denote by ρ4 : E′2 → P1 the natural induced morphism, which defines a P1-bundle structure. We have the spectral sequence: (5.12) Riρ4∗(Rjρ3∗(Ql)) =⇒ Ri+jρ∗(Ql). By similar considerations to those for (5.11), we have the following exact sequence: 0 → R2ρ4∗(Ql) → R2ρ∗(Ql) → H 2 ´et(P1, Ql)⊕3 → 0. 27 (2) E 2 ρ 1 (3) E2 ρ 2 ρ 3 E'2 ODP s = 0 oo ρ 4 s = 0 oo s = 0 oo Figure 5.5. Note that all eigenvalues of Frobp on H 2 show that the claimed property holds for H 0 that the claimed property holds for H 0 ´et(P1, Ql)⊕3 are equal to p. Therefore, to 2 , R2ρ∗(Ql)), we have only to show ´et(E(2) ´et(E′2, R2ρ4∗(Ql)). Now we consider the Leray spectral sequence: ´et(P1, Rjρ4∗(Ql)) =⇒ H i+j H i ´et (E′2, Ql). Since ρ4 is a P1-bundle, we have R1ρ4∗(Ql) = 0. Therefore, in a similar way as above, we have the following exact sequence: ´et(P1, Ql) → H 2 ´et(E′2, Ql) → H 0 0 → H 2 Since ρ1 : E(2) 2 → P1 has a section defined over Q, due to 2) in Proposition 5.9 ´et(E′2, Ql) is generated by applied to a, b ∈ Z, so does ρ4 : E′2 → P1. Therefore H 2 the classes of divisors defined over Q, which are a section and a fiber. Hence all ´et(E′2, Ql) are equal to p [vGN], and then the claimed eigenvalues of Frobp on H 2 property holds for H 0 ´et(E′2, R2ρ4∗(Ql)) → 0. ´et(E′2, R2ρ4∗(Ql)). (cid:3) From the Proposition 5.13 and the fixed point formula (5.10), we have 3 2 , (5.13) (cid:12)(cid:12)1 + (50 + h21)(p + p2) + p3 − # X∗0 (Fp)(cid:12)(cid:12) ≤ (2 + 2 h21)p where we have used b2 = b4 by the Poincaré duality and also expressed b2 = h11 = (50 + h21) from e( X∗0 ) = 2(h11 − h21) = 100. Proposition 5.16. The number of rational points # X∗0 (Fp) is given by 2 (Fp) + 10 × #E1(Fp) + 30 p2 + 40 p − 10, # X∗0 (Fp) = #Z sp where #Z sp 2 (Fp) and #E1(Fp) are the numbers of rational points over Fp for the determinantal quintic (2.6) and the singular del Pezzo surface in Proposition 5.5, respectively, with a = b = 1. Proof. The projection π2 : X sp 0 → Z sp is isomorphic outside the coordinate lines qi ( see Fig. 5.1). Since the fibers over the coordinate point [ei] and qi \ {[ei], [ei+1]} are P2 and P1, respectively, we obtain 2 # X sp 0 (Fp) = #Z sp = #Z sp 2 (Fp) + 5 × (NP2 − 1) + 5 × (NP1 − 1)(NP1 − 2) 2 (Fp) + 10p2, where NP2 = p2 + p + 1 and NP1 = p + 1, respectively, count the number of rational points in P2 and P1 over Fp. We count the number of rational points on the conic 28 ● bundle E2 (with two reducible fibers) over P1 as #E2(Fp) = (p + 1)(p − 1) + (2p + 1) × 2 = p2 + 4p + 1. The counting for E3 is given by #E3(Fp) = (p + 1)2. Now summarizing all, we obtain # X∗0 (Fp) = # X sp 0 (Fp) + 10 ×(cid:0)#E1(Fp) − 1(cid:1) + 10 ×(cid:0)#E2(Fp) − (p + 1)(cid:1) + 10 ×(cid:0)#E3(Fp) − (p + 1)(cid:1) 2 (Fp) + 10 × #E1(Fp) + 30 p2 + 40 p − 10. = #Z sp (cid:3) Writing a straightforward computer codes, we have evaluated the number # X∗0 (Fp). After the computations in several minutes, we verify the inequality (5.13) for p ≤ 97 with h2,1 = 2 or 3. For example, we obtain # X∗0 (Fp) = 669880, 1118250 and 1408330 for p = 73, 89 and 97, respectively. We observe that the inequality (5.13) holds only if h2,1 = 2 for p = 59, 61, 71, 73, 89, 97. Also we can verify that these are good primes by analyzing the Jacobian ideals over the field Fp. Since the inequality holds for all good primes, we conclude that: Theorem 5.17. The smooth Calabi-Yau manifold X∗0 has Hodge numbers; h1,1( X∗0 ) = 52, h2,1( X∗0 ) = 2. In particular this is mirror symmetric to the generic complete intersection X0 with h1,1( X0) = 2, h2,1( X0) = 52. 29 (6.1) where Ω( X sp f1f2 ··· f5(cid:19) , 0 ) = Resf1=...=f5=0(cid:18) dµ1 ∧ dµ2 5Xi=1 (−1)izidz1 ∧ ··· ∧ cdzi ∧ ··· ∧ dz5, dµ1 = − 6. Picard-Fuchs equations and monodromy matrices 6.1. Picard-Fuchs differential equations. We consider a family of Calabi-Yau manifolds X∗0 defined over (C∗)2 ∋ (a, b). Here we briefly introduce a natural compactification of (C∗)2 to P2 which follows from the differential equations satisfied by the period integrals, see [HKTY] and [HT] for details. To formulate the set of differential operators, we slightly modify the defining equations (2.3) to fi = ciziwi + aizi+1wi + biziwi+1 (i = 1, ..., 5), where the indices are considered modulo five as before. Clearly, the original forms are recovered by setting ai = a, bi = b, ci = 1. Since we have X∗0(C∗)4×(C∗)4 ≃ X sp 0 (C∗)4×(C∗)4 for (C∗)4 × (C∗)4 ⊂ P4 × P4, a holomorphic 3-form of the crepant resolution X∗0 may be given by the corresponding 3-form of X sp if the 3-cycles of the period integrals are contained in (C∗)4 × (C∗)4. For the complete intersection X sp 0 , the following expression of a holomorphic 3-form is well-known [Gr]: 0 (cid:26) ∂ and similar definition for dµ2 with the coordinates wk's . The period integral ´Γ Ω( X∗0 ) for a 3-cycle Γ ∈ H3( X∗0 , Z) satisfies a system of differential equations, the so-called Picard-Fuchs differential equations, see [Mo], [DGJ] for example. In the present case, assuming that the cycle Γ is contained in (C∗)4 × (C∗)4, we can describe the system by noting rather trivial algebraic relations represented in terms of differential operators, e.g., ∂ ∂a5(cid:27) Ω( X sp ∂ ∂c3 ∂ ∂c4 ∂c1 ∂ ∂a1 ∂ ∂a4 ∂ ∂a3 ∂ ∂a2 0 ) = 0, ∂ ∂ ∂c5 − ∂c2 which represents Π5 i=1ziwi − Π5 i=1zi+1wi = 0. We should also note that the holo- morphic 3-form is invariant under the (C∗)4-action zi 7→ tizi, (t1t2 ··· t5 = 1) and similar (C∗)4-action on the coordinates wi's. We note further that Ω( X sp 0 ) has a simple scaling property under fi 7→ rifi (ri ∈ C∗). All these properties of in- variance (or covariance) may be expressed by the corresponding linear differential operators, and may be used to reduce the enlarged parameters to the original a and b. The system of differential operators which we obtain in this way is an example of the Gel'fand-Kapranov-Zelevinski (GKZ) system [GKZ] for which a natural com- pactification of the parameters is known. In the present case, from the C∗-actions above and the form of the defining equations (2.3), it is rather easy to deduce that (C∗)2 ∋ (a, b) is compactified to P2 ∋ [a5 : b5 : 1]. According to the mirror sym- metry calculations formulated in [HKTY], we actually come to the affine charts {(x, y),A0},{(x1, y1),A1} and {(x2, y2),A2} defined by x = −a5, y = −b5; x1 = − b5 a5 , y1 = − 1 a5 ; x2 = − a5 b5 , y2 = − 1 b5 . Up to signs, these relations are in accord with the standard relations [a5 : b5 : 1] = [1 : b5 b5 ] of the affine coordinates of P2. The extra minus signs follows from the general definition given in [HKTY]. a5 ] = [ a5 b5 : 1 : 1 a5 : 1 30 Proposition 6.1. On the affine chart {(x, y),A0}, the following differential oper- ators determine the period integrals as the solutions: D1(x, y) = 2θ3 D2(x, y) = 2θ2 where θx = x ∂ given by the following gauge transforms of the operators D1(x, y),D2(x, y): y − (θx + θy)2(cid:8)(2θx + 3θy)x − (3θx + 2θy)y(cid:9), x − 3θ2 x − 3θxθy + 2θ2 ∂x , θy = y ∂ ∂y . On the other affine charts the differential operators are y − 2θ3 y − (2θ2 x + 7θxθy + 2θ2 x + 7θxθy + 7θ2 y)x − (7θ2 xθy + 3θxθ2 y)y, D′1(x1, y1) := x1D1(x1, y1)x−1 1 , D′2(x1, y1) := D2(x1, y1) on {(x1, y1),A1} and D′′1 (x2, y2) := x2D1(x2, y2)x−1 2 , D′′2 (x2, y2) := D2(x2, y2) on {(x2, y2),A2}. Proof. For the derivation of the differential operators D1,D2, we refer to [HKTY]. Also see Prop.2.6 in [HT]. Note that the parameters (ai, bi, ci) in [HT, (2.6) ] should be read as (ci, ai, bi) here (see the defining equations fi ). (cid:3) 6.2. Determinantal quintics. For the determinantal quintics Z sp have the following standard forms of holomorphic 3-forms: 2 , and X sp,♯ 0 , we (6.2) Ω(Z sp 2 ) = ResFw=0(cid:18) dµ2 2 = {Fw = 0} and X sp,♯ Fw(cid:19) , Ω( X sp,♯ 0 ) = ResFλ=0(cid:18) dµλ Fλ (cid:19) , where Z sp 0 = {Fλ = 0} (see (2.6)). We may derive these holomorphic 3-forms from (6.1) by evaluating the residue integrals: Let us take an affine coordinate [z1 : z2 : z3 : z4 : 1] of P4, and regard the relations f1 = ··· = f4 = 0 as linear equations for z1,..., z4 with fixed wk's, i.e., B(cid:18) z1 Then, changing the variables to t(ξ1, ..., ξ4) = B t(z1,..., z4) and taking into account the Jacobian factor dz1 ∧ ... ∧ dz4 = 1 detB dξ1 ∧ ... ∧ dξ4, we obtain z4(cid:19) =(cid:18) 0 −aw4(cid:19) . z2 z3 0 0 Ω( X sp 0 ) = Resf5=0(cid:18) dµ2 det B f5(cid:19) . 1 and X sp 0 ) = Ω(Z sp Since we can verify the equality det B f5 = Fw, we see that Ω( X sp 2 ) holds. By changing the roles of zk's with wk's, we have a similar result for Ω(Z sp 1 ). The threefolds X sp in the diagram (5.1) also have the form of complete intersections of five (1, 1)-divisors. The same formal arguments as above apply to the cases of X sp 2 ), respectively. By evaluating the residues, the holomorphic 3-forms Ω( X sp i ) (i = 1, 2) can also be connected to the holomorphic 3-form Ω( X sp,♯ i ). Noting that there are 3-cycles contained in the tori (see the next subsection), we have: starting from Ω( X sp ) as well as Ω( Z sp 1 ) and Ω( X sp 1 and X sp 2 2 0 Proposition 6.2. The period integrals of Z sp holomorphic 3-forms Ω(Z sp satisfy the same Picard-Fuchs differential equations as in Proposition 6.1. 2 ) and Ω( X sp,♯ 2 , and X sp,♯ 1 ), Ω( X sp 2 ), Ω( X sp 1 ), Ω(Z sp 1 , X sp 2 , X sp 1 , Z sp with the ), respectively, 0 0 31 6.3. Integral, symplectic basis and monodromy matrices. As in [CdOGP], we can evaluate the period integral of Ω(Z sp 2 ) over certain torus cycles. Let us first note that Γ0 = {[w1 : w2 : w3 : w4 : 1] ∈ Sspw1 = w2 = w3 = ε} defines a 3-cycle in Z sp 2 . This simply follows by observing that the substitution of wk = εe√−1θk (k = 1, 2, 3) (in the affine coordinate w5 = 1) into the defining equation of Z sp 2 entails a quadratic equation for w4, and one of the two roots goes to zero when ε → 0. Choosing this vanishing root defines a 3-cycle Γ0. Combined with the residues contained in the definition of Ω(Z sp 2 ), one obtain dµ2 Fw (2πi)4 γ0 Γ0 Ω(Z sp 2 ) = (6.3) 1 , where γ0 = {w1 = ··· = w4 = ε, w5 = 1} is a torus cycle in P4. Proposition 6.3. The period integral (6.3) can be evaluated in three different ways depending on the (relative) magnitudes of a and b: 1 1 b5 a5 Ω(Z sp Γ0 a5 ,− a5(cid:1), a5 w0(cid:0)−1 2 ) = w0(cid:0)−a5,−b5(cid:1), where we set w0(x, y) = Pn,m≥0 absolutely for x,y < 1 25 . Proof. Since the cycle γ0 is contained in the affine coordinate w5 = 1 (in fact γ0 ⊂ (C∗)4), we may use dµ2 Fw(w1,··· ,w4,1) for the evaluations. The claimed expansions follow from the three different ways of handling 1 : The first one is Fw obtained by ((n+m)!)5 (n!)5(m!)5 xnym. The series w0(x, y) converges b5(cid:1), b5 , −1 b5 w0(cid:0)− = dw1∧···∧dw4 Fw dw1 ∧ ··· ∧ dw4 Fw =(cid:18)1 +(cid:18) Fw w1w2w3w4 − 1(cid:19)(cid:19)−1 dw1 ∧ ··· ∧ dw4 w1w2w3w4 and taking the residue integrals about wk = ε, see [BaCo] for example. Similarly, the second one follows from dw1 ∧ ··· ∧ dw4 Fw = 1 a5(cid:18)1 +(cid:18) Fw a5w1w2w3w4 − 1(cid:19)(cid:19)−1 dw1 ∧ ··· ∧ dw4 w1w2w3w4 . For the third one, we simply replace the a5's by b5's in the above equation. Since the convergence follows from the standard estimates using the duplication formula of the Γ-functions, our derivation may be brief here. Assume x,y < r, then we have (d + 1)c[ d 2 ],[ d+1 2 ]rd ≤ 1 +Xd≥1 (d + 1)(cid:18) 2d √π(cid:19)5 rd, n,m≥0 cn,mxnym ≤ 1 +Xd≥1 Xd≥0 Xn+m=d where cn,m = (cid:0) (n+m)! n!m! (cid:1)5 inequality. Since the last series converges for 25r < 1, we obtain the claim. , and the duplication formula is used to have the second (cid:3) It should be clear that the three different series expansions of the period integral originate from the symmetry of the defining equations fi of X sp 0 , which we started with. Also, we can observe here the natural compactification of the deformations by (a, b) ∈ (C∗)2 to [a5 : b5 : 1] ∈ P2 discussed above. Moreover, we may observe that the three infinity points [0 : 0 : 1], [0 : 1 : 0] and [1 : 0 : 0] are all isomorphic up to suitable factors or "gauge" transformations as claimed in Proposition 6.1. 32 In the next subsection, we set up a canonical integral, symplectic basis for the solutions which follows from the mirror symmetry. 6.3.1. Canonical integral and symplectic basis. The space of the solutions of the Picard-Fuchs differential equation is endowed with an integral and symplectic struc- ture in their monodromy property which come from those in H3( X∗0 , Z). Using the mirror symmetry of X∗0 to X0, we have a canonical form of the (conjectural) integral and symplectic basis of the solutions [Ho1, Prop.1], [Ho2, Conj.2.2]. O X0 Euler characteristic χ(E,F ) =Pi(−1)i dim Exti Recall that, under the mirror symmetry, the integral and symplectic structure in H3( X∗0 , Z) is conjecturally isomorphic to those in the (numerical) Grothendieck group K( X0) of the mirror Calabi-Yau manifold X0 to X∗0 [Ko]. Note that the (E,F ) of coherent sheaves E,F on X0 defines a skew symmetric form on the Grothendieck group K( X0) due to the fact that X0 is a Calabi-Yau threefold. This skew symmetric form (as well as the integral structure) in K( X0) may be transferred into H even( X0, Q) by the Chern character homomorphism: ch : K( X0) → H even( X0, Q) and the Riemann-Roch formula for χ(E,F ). Explicitly, the skew form on H even( X0, Q) may be written by (α, β) = ´ X0 represents the Todd class and α = α0 + α2 + α4 + α6 represents the decomposition with respect to H even( X0, Q) = ⊕3 The Calabi-Yau manifold X0 is a smooth complete intersection of five generic (1, 1)-divisors in P4× P4. The cohomology H even( X0, Q) is generated by the hyper- plane classes J1, J2 from the respective projective spaces P4 with the ring structure J 3 compatible with their intersection numbers (´ X0 2 ) = (5, 10, 10, 5). Using this ring structure in H even( X0, Q), the mirror symmetry stated above can be summarized into the following cohomology-valued hypergeometric se- ries [Ho2, Sect.2]: i=0H 2i( X0, Q) and similarly for β = β0 + β2 + β4 + β6. (α0− α2 + α4− α6)∪ (β0 + β2 + β4 + β6)∪ T d X0 J 2 1 J2,´ X0 , where T d X0 J 3 1 ,´ X0 1 ,´ X0 J2J 2 (6.4) ω(cid:18)x, y; J1 2πi , J2 2πi(cid:19) = Xn,m≥0 Γ(1 + n + m + J1 2πi + J2 2πi )5 2πi )5Γ(1 + m + J2 2πi )5 Γ(1 + n + J1 xn+ J1 2πi ym+ J2 2πi , where the right hand side is defined by the series expansion with respect to the nilpotent elements J1, J2 in the cohomology ring. By this series expansion in the cohomology ring, we effectively generate the solutions of the Picard-Fuch differential equations formulated in [HLY], [HKTY]. Then the (conjectural) claim made in [Ho1, Prop.1], [Ho2, Conj.2.2] is as follows: In this form of the cohomology-valued hypergeometric series, the integral and symplectic structure in H even( X0, Q) is transformed canonically to that of the hypergeometric series representing the period integrals. The canonical integral, symplectic structure may be read by arranging 2πi(cid:1) as follows: 2πi , J2 ω(cid:0)x, y; J1 w0(x, y)1 +Xk w(1) k (x, y)(cid:0)Jk −Xl +Xk CklKl(cid:1)T d−1 = − 1 JkKl = δkl and ´ X0 are defined so that we have ´ X0 1 + J 3 where T d X0 2 ). Here, Kk and V X0 = −1. Ckl's are constants satisfying Ckl = Clk which must be fixed (by hand) from the ex- plicit monodromy calculations of the hypergeometric series (Proposition 6.6). The k (x, y)Kk + w(3)(x, y)V X , is the Todd class and Kk = 1 = 1 + c2( X0) 10 (J 3 V X0 k , V X0 w(2) 5 J 2 X 12 33 (6.5) Σ0 = 0 0 0 0 0 1 0 0 0 0 1 0 0 0 0 1 0 0 0 0 −1 0 0 0 0 −1 0 0 0 0 0 0 0 0 −1 0  integral structure on H even( X0, Q) can be introduced through the basis {1,(cid:0)Jk − Pl CklKl(cid:1)T d−1 , etc. Then, with respect to this basis, the symplectic form (∗,∗) : H even( X0, Q)× H even( X0, Q) → Z described above takes the following form: , Kk, V X} by noting ch(O X0 ) = 1, ch(Op) = −V X0 X with no dependence on Ckl. From the above calculations of the cohomology-valued hypergeometric series, we read the (conjectural) integral, symplectic basis of the period integrals as Π(x, y) = t(w0(x, y), w(1) 1 (x, y), w(1) 2 (x, y), w(2) 2 (x, y), w(2) 1 (x, y), w(3)(x, y)). For notational simplicity, we will understand by Π(x, y) = t(w0(x, y), w(1) k (x, y), w(2) l (x, y), w(3)(x, y)), (k, l = 1, 2) the period integrals arranged in the above order. We observed in Proposition 6.1 that there appear two other local structures on {(x1, y1),A1} and {(x2, y2),A2}. It has been noted in [HT] that these local struc- tures correspond to X1 and X2, respectively, both of which are smooth complete intersections of (1, 1)-divisors and birational to X0(6≃ Xi, i = 1, 2). By symmetry, up to the gauge transformations, we have the corresponding cohomology valued hypergeometric series x1ω(cid:18)x1, y1; J′1 2πi , J′2 2πi(cid:19) , x2ω(cid:18)x2, y2; J′′1 2πi , J′′2 2πi(cid:19) under the integral, symplectic structures on H even( X1, Q) and H even( X2, Q), re- spectively. The definitions and the calculations of these cohomology valued hyper- geometric series are parallel to (6.4) with the corresponding generators J′k and J′′k . We read the canonical symplectic form Σ0 as above, and the canonical integral, symplectic basis of the period integrals as Π′(x1,y1) = t(x1w0(x1, y1), x1w(1) Π′′(x2, y2) = t(x2w0(x2, y2), x2w(1) k (x1, y1), x1w(2) k (x2, y2), x2w(2) l l (x1, y1), x1w(3)(x1, y1)), (x2, y2), x2w(3)(x2, y2)). Note that Π(x, y), Π′(x1, y1) and Π′′(x2, y2) contain the same unknown constants Ckl in common, which will be fixed later in Proposition 6.6. To make the Taylor expansion of the cohomology valued hypergeometric series (6.4), let us introduce the following notation: ρ2 ρ1 2πi ∂ρk w(x, y) = ∂ ∂ρk ∂2 2πi(cid:1)(cid:12)(cid:12)ρ=0, 2πi(cid:1)(cid:12)(cid:12)ρ=0,··· with formal variables ρ1, ρ2. Using the intersection numbers ´ X0 c2J1 = ´ X0 ´ X0 w(cid:0)x, y; w(cid:0)x, y; 2 = 10, and also the values ´ X0 the explicit form of the period integral Π(x, y): J 2 1 J3 = ´ X0 ∂ρk ∂ρlw(x, y) = ρ1 2πi ∂ρk∂ρl J1J 2 ρ2 , , J 3 2 = 5, c2J2 = 50, we have J 3 1 = ´ X0 34 (6.6) Π(x, y) =  w0(x,y) ∂ρ1 w(x,y) ∂ρ2 w(x,y) 2 ∂2 w+10∂ρ1 ∂ρ2 w+ 5 ρ2 w+10∂ρ1 ∂ρ2 w+5∂2 ρ2 w)−5(∂2 ρ1 5∂2 ρ1 2 ∂2 5 ρ1 w+∂3 ρ2 w+Pb C2b∂ρb w w+Pb C1b∂ρb w − 5 6 (∂3 ρ1 ∂ρ2 w+∂ρ1 ∂2 ρ2 w)− 50 12 (∂ρ1 w+∂ρ2 w) ,  and similar forms for Π′(x1, y1) and Π′′(x2,y2). In the following calculations, we use the powerseries expansions of these period integrals to sufficiently higher orders. 6.3.2. Analytic continuations. Let us consider the analytic continuations of the three isomorphic local structures noted in Proposition 6.1 to the 'center' [1 : 1 : 1] of P2. We introduce a local coordinate s = x + 1, t = y + 1 of A0 which locates the center [1 : 1 : 1] at the origin, and write the Picard-Fuchs differential equations as D1ϕk(s, t) = D2ϕk(s, t) = 0 (k = 0, ..., 5). We arrange the solutions into the column vector ϕ(s, t) = t(ϕ0(s, t), ϕ1(s, t), ϕ2(s, t), ϕ3(s, t), ϕ4(s, t), ϕ5(s, t)). Similarly we consider the local solutions satisfying D′1ϕ′k(s1, t1) = D′2ϕ′k(s1, t1) = 0 with the local coordinates s1 = x1 + 1, t1 = y1 + 1 of A1, and also D′′1 ϕ′′k(s2, t2) = D′′2 ϕ′′k(s2, t2) = 0 with s2 = x2 + 1, t2 = y2 + 1 of A2. Since the center [1 : 1 : 1] is a regular point of the differential equations D1ϕk(s, t) = D2ϕk(s, t) = 0 (see Proposition 6.5), we have 6 power series solutions. After some calculations, we see that the following leading behaviors determine the local solutions uniquely: (6.7) ϕ0(s, t) = 1 + c(0) ϕ2(s, t) = s + c(2) ϕ4(s, t) = t2 + c(4) 11 st + ··· , 11 st + ··· , 11 st + ··· , ϕ1(s, t) = t + c(1) ϕ3(s, t) = s2 + c(3) ϕ5(s, t) = s3 + ··· , 11 st + ··· , 11 st + ··· , where ··· represent higher order terms (degree ≥ 3) which do not contain the s3- term. Since the differential operators D′i and D′′i are related to Di as in Proposition 6.1, the corresponding local solutions are simply given by (6.8) ϕ′i(s1, t1) = (s1 − 1)ϕi(s1, t1) , ϕ′′i (s2, t2) = (s2 − 1)ϕi(s2, t2). Proposition 6.4. The three local solutions are related by ϕ′(s1, t1) = M1 ϕ(s, t) , ϕ′′(s2, t2) = M2 ϕ(s, t) , with M1 = 121 11 0 − 58 −1 0 −1 − 7 0 − 3 0 0 −1 1 0 − 6 1 0 0 1 0 0 0 0 0 1 −1 0 0 0 0 0 0 0 26 11 13 11 11 11 0 −1  ,M2 = 0 3 11 −1 −1 0 0 0 0 0 0 0 − 7 11 0 1 −1 0 0 −3 0 0 1 0 0 0 1 1 − 13 0 0 −1 0 −1 0 0 0 13 11 11  . Proof. Since [−x : −y : 1] = [1 : −y1 : −x1] = [−y2 : 1 : −x2] by definition, we have [1 − s : 1 − t : 1] = [1 : 1 − t1 : 1 − s1] = [1 − t2 : 1 : 1 − s2] and s − t 1 − t ; s2 = −t 1 − t s1 = −s 1 − s t − s 1 − s , t1 = , t2 = 35 . Then we should have t − s 1 − s(cid:1) = M1ϕ(s, t) , ϕ′′(cid:0) −t 1 − t , s − t 1 − t(cid:1) = M2ϕ(s, t), for s,t ≪ 1. Using the relations (6.8) for the left hand sides, we obtain the claimed form of the matrices M1, M2. (cid:3) ϕ′(cid:0) −s 1 − s , Now by Proposition 6.4, the connection problems of the three period integrals Π(x, y), Π′(x1, y1), Π′′(x2, y2) to each other may be solved by the analytic continua- tions of each to the corresponding local solutions around the center. By symmetry, we note that connecting Π(x, y) to ϕ(s, t) is sufficient for our purpose. Proposition 6.5. The singular loci of the Picard-Fuchs differential equations con- sist of the three coordinate lines of P2and an irreducible nodal rational curve of genus 6. The defining equation of the nodal curve in the affine chart {(x, y),A0} has the following form dis0 = (1 − x − y)5 − 54xy(1 − x − y)2 + 55xy(xy − x − y). Proof. This follows from calculating the characteristic variety of the differential operators D1(x, y), D2(x, y), see [HT, Remark 2.7]. (cid:3) Since the irreducible component dis0 = 0 is rational, this can be parametrized globally by P1. In fact, we can verify that the equation dis0 = 0 follows from the discriminant dis( X sp 0 (C∗)4) = 0 determined in Proposition 2.3 eliminating the variables a and b under the relations x = −a5, y = −b5. Hence as a global param- eter of the curve we can adopt an affine line a + b + 1 = 0 in (C∗)2 (which we compactify to P1 with infinity). Using this, we have depicted a schematic picture of the singular loci in Fig.6.1. In the figure, the curve dis0(x, y) = 0 of complex- one dimension is reduced to the corresponding real curve by imposing a condition Im(x) = Im(y). The real plane curves drawn in the figure are the projection of the space curve {(Re(x), Re(y), Im(x)) dis0(x, y) = 0} to the first two coordinates. Also, the three affine coordinates are taken "outward direction" from the standard right-triangular shape of the moment polytope of P2 whose vertices represent the three affine coordinates [a5 : b5 : 1] = [1 : b5 a5 ] = [ a5 b5 : 1 : 1 a5 : 1 b5 ]. 6.3.3. Monodromy transformations shown in Fig.6.1. As explained above, the defin- ing equation dis0(x, y) = 0 can be solved by the line a + b + 1 = 0. We set a = 1 2 − (α + iβ) and solve the additional condition Im(x) = Im(y) (x = −a5, y = −b5) for β to have the space curve 2 + (α + iβ), b = 1 {(Re(x(α)), Re(y(α)), Im(x(α))) − ∞ < α < +∞}. Solving the equation Im(x) = Im(y) for β introduces five branches for the solutions. Each of the solutions determines a partial parametrization of the curve by α. As shown in Fig.6.1, we have two connected components for the real space (plane) curve in this way. One component comes from the obvious solution β = 0, and this is represented by the component that consists of 3 solid-bold (hyperbola-shaped) lines and 3 broken lines. Due to the repetition of the regions in the coordinate planes, each of the 3 broken lines should be identified with the solid-bold line in 36 y1 T 2p' 1Tx l1 x1 y l 0 Tp"3 1, Tp"3 2, 1Ty Tp' ,11 T ,21p' Tz'' [1,1,1] Tz T ,3p 2 1T ,p3 Tx 1T ,1 p Tp2 Ty p1T ,2 Tz' T ,3 2p' 1T ,3p' Ty2 T ,p"1 2 T ,1p"1 Tx2 Tp"2 l2 2x x y2 Figure 6.1. Singular loci of the Picard-Fuchs differential equa- tions. Each loop represents the monodromy transformation with a base point near (x, y) = (0, 0) and a path taken over the lines ℓi (Im(x) = Im(y) > 0). the opposite side. The other component contains the 6 nodes. It is left to readers to draw a picture of real Riemannian surface of genus 6 with 6 nodes whose real hyperplane section is given by the plane curves shown in Fig.6.1. In Fig. 6.1, we have also drawn three lines ℓi : ℓ0 : (x = y), ℓ1 : (x1 = y1), ℓ2 : (x2 = y2) which intersect at the center [1, : 1 : 1]. Each line intersects with the curve at the two nodal points, and transversally at one point, as shown in the figure. We name all these points of the intersection by p1, p2, p3; p′1, p′2, p′3; p′′1 , p′′2 , p′′3 with their explicit coordinates: 32 p1 : [−ρ− : −ρ− : 1], p2 :(cid:2)−1 p′1 : [1 : −ρ− : −ρ−], p′2 :(cid:2)1 : −1 p′′1 : [−ρ− : 1 : −ρ−], p′′2 :(cid:2)−1 : −1 32 : −1 32 : 1 : −1 32 + + : 1], : −ρ : 1(cid:3), p3 : [−ρ 32(cid:3), p′3 : [1 : −ρ+ : −ρ+], 32(cid:3), p′′3 : [−ρ+ : 1 : −ρ+], where ρ∓ = 11∓√5 . 2 For the monodromy calculation of the period integral Π(x, y), we take a base point o near the origin (x, y) = (0, 0). We fix it to be a real point near (0, 0) and 1 ≫ Im(x) > 0 in the figure. Starting this base point, we define the monodromy transformations Tx,Ty around the coordinate axes via the loops shown. Similarly we define monodromy transformations Tp1,1,Tp1,2;Tp2; Tp3,1, Tp3,2;Tz by connecting the small loops shown in the figure with the paths 'over' the line ℓ0 (a line near ℓ0 with Im(x) = Im(y) > 0) from the base point. We define the monodromy representation, ρ : π1(P2 \ DP F , o) → Sp(6, Z) with DP F representing the singular loci of the Picard-Fuchs differential operators and Sp(6, Z) = { P ∈ GL(6, Z) tP Σ0P = Σ0 } 37 with respect to the symplectic form Σ0 in (6.5). We adopt the convention that, for example, Tx.Π(x, y) = ρ(Tx)Π(x, y) represents the analytic continuation Tx.Π(x, y) of the local solution Π(x, y) along the path with the loop Tx in terms of the local solution Π(x, y). Thus under our convention, the monodromy representation ρ is an anti-homomorphism satisfying ρ(T1T2) = ρ(T2)ρ(T1). Proposition 6.6. When we take C11 = C22 = − 1 2 , C12 = C21 = 0 in the canonical integral, symplectic basis Π(x, y) in (6.6), all the monodromy transformations above are represented by the elements ρ(T∗) in Sp(6, Z). Explicitly, the corresponding monodromy matrices acting on the period integral Π(x, y) are given by: Tx : Tp1,1 : 1 0 0 0 0 0 1 1 0 0 0 0 0 0 1 0 0 0 5 10 10 1 0 0 2 5 10 0 1 0 -5 -3 -5 0 -1 1 5 1 5 2 -1 0 -1 -4 0 0 0 0 0 -5 3 -1 0 -1 -10 10 4 -1 0 -2 -25 25 10 -5 1 -5 25 -25 -10 5 0 6 1 0 0 0 0 0 1 1 0 0 0 0 1 0 1 0 0 0 2 10 5 1 0 0 5 10 10 0 1 0 -5 -5 -3 -1 0 1  , Ty :  , Tp1,2 : -2 -5 0 1 1 6 -1 -5 0 1 1 5 0 0 1 0 0 0 25 -10 -25 1 5 5 10 -4 -10 0 3 2 -25 10 25 0 -5 -4  , Tz :  , Tp3,2 : 41 -4 -20 0 12 16 30 -2 -15 0 9 12 0 0 50 -5 -25 1 15 20 10 -1 -5 0 4 4 -100 10 50 0 -30 -39 1 0 0 0 Tp3,1 : 0 -6 6 15 41 -17 -17 6 1 3 -6 -2 4 4 -2 1 0 3 -72 28 23 -13 -9 -28 -72 23 28 -9 -13 -28 -30 18 18 -4 -4 -9  , Tp2 : 1 0 0 0 0 1 0 1 0 0 0 0 0 0 1 0 0 0 0 0 0 1 0 0 0 0 0 0 1 0 0 0 0 0 0 1  ,  , 1 -39 20 4 -12 0 -16 0 0 0 0 0 -9 0 -12 -30 15 4 -10 5 1 -2 0 -4 -50 25 5 -15 1 -20 100 -50 -10 30 0 41  . Proof. Our proof is based on numerical calculations except for Tx and Ty. To have the matrix of Tp1,1, for example, we generate the power series for Π(x, y) in (6.6) up to total degree 60. From the base point to a small loop for Tp1,1, we may take a path over the line ℓ0, i.e., a real line near ℓ0 with Im(x) = Im(y) > 0. This choice of path, however, is not efficient to attain numerically high accuracy due to the 'degeneration' of the period integrals which we see in w(k) 2 (x, y) when x = y. To avoid this degeneration, we deform the path satisfying Im(x) = Im(y) to that satisfying Im(y) = 0 by making use of the homotopy εIm(x) = Im(y), ε ∈ [0, 1]. Thereby, we verify that the path does not intersect the singular loci DP F at any ε ∈ [0, 1]. The path for our actual calculation is a path over the ℓ0 satisfying Im(y) = 0. We divide the deformed line into 200 segments and also the small loop into 100 arcs. Then, for each endpoint of them, we have constructed the local solutions imposing the same leading behavior in (6.7). The monodromy matrix, by definition, follows by relating these solutions at each ends along the path. We obtained the claimed integral, symplectic matrix for Tp1,1 in the accuracy 10−5 ∼ 10−6. Other monodromies are determined in the same way with the same level of accuracy in their numerical calculations. (cid:3) 1 (x, y) = w(k) We now consider the analytic continuation of the local solution Π(x, y) from the base point to the center [1, : 1 : 1] along (over) the line ℓ0, and further continue to a point near (x1, y1) = (0, 0) along (over) the line ℓ1. We express the local solution Π′(x1,y1) in terms of the analytically continued solution Π(x, y) by Π′(x1, y1) = C10Π(x, y). In a similar way, we consider the analytic continuation of Π(x, y) along the line ℓ0 followed by ℓ2, and define the relation Π′′(x2, y2) = C20Π(x, y). Proposition 6.7. The above relations Π′(x1, y1) = C10Π(x, y) and Π′′(x2, y2) = C20Π(x, y) are solved by 38 4 4 -3 -6 -17 0  -4 -2 -1 0 4 6 -8 -4 -2 -4 0 17 -1 0 -1 -2 -4 4 2 1 0 -1 -2 -3 0 1 -2 -4 -8 4  . C10 = -4 -4 3 6 17 0 8 4 2 4 0 -17 4 2 1 0 -4 -6 -2 -1 0 1 2 3  1 0 -1 0 1 2 2 4 4 8 -4 -4  , C20 = Proof. As in the previous proposition, we do numerically the analytic continuation of Π(x, y) to ϕ(s, t) along ℓ0, and Π′(x1, y1) to ϕ′(s1, t1) along ℓ1. Then use Propo- sition 6.4 to relate ϕ(s, t) and ϕ′(s1, t1), and obtain the claimed matrix C10. The matrix C20 follows in the same way. (cid:3) Since the three local forms of the period integral Π(x, y), Π′(x1, y1) and Π′′(x2, y2) are governed by the isomorphic system of differential equations (see Proposition 6.1) and also from the obvious symmetry in Fig.6.1, the entire monodromy properties of the period integral Π(x, y) can be described by the monodromy transformations or the corresponding transformations: {ξm} := {Tx,Ty, Tp1,1,Tp1,2, Tp2, Tp3,1, Tp3,2, Tz}, {ξ′m} := {Tx1, Ty1, Tp′ {ξ′′m} := {Tx2, Ty2, Tp′′ 1,1,Tp′1,2, Tp′ 1 ,1,Tp′′ 1 ,2, Tp′′ 2 , Tp′ , Tp′′ 3,1, Tp′ 3 ,1, Tp′′ 3,2, Tz′}, or 3 ,2, Tz′′}. 2 Proposition 6.8. 1) For the monodromy matrices we have ρ(ξ′m) = C−1 10 ρ(ξm)C10, ρ(ξ′′m) = C−1 20 ρ(ξm)C20. 2) The following relations can be observed: p2 Ty) , ρ(Tp1,2) = ρ(T −1 x Tp2 Tx) ρ(Tp1,1) = ρ(T −1 y T −1 ρ(Tp3,1) = C10 ρ(TxT −1 ρ(Tp3,2) = C20 ρ(TxT −1 y Tp2TyT −1 p2 TyT −1 y T −1 x )C−1 10 , x )C−1 20 . 1,2) = ρ(T −1 1 ,2Tx2Tp′′ p′′ 3) We have ρ(Tz) = ρ(T −1 1,2Tx1Tp′ p′ 4) The image of the monodromy transformations in Sp(6, Z) is given by 1 ,2). hρ(T ±1 x ), ρ(T ±1 y ), ρ(T ±1 p2 ),C±1 10 ,C±1 20 i. Proof. 1) By the symmetry summarized in Proposition 6.1 and the definitions of C10 and C20, the first claim follows. For 2), we verify directly the claimed relations using the monodromy matrices in Proposition 6.6. When doing this, we should note that ρ is defined as an anti-homomorphism, ρ(TαTβ) = ρ(Tβ)ρ(Tα). Using the results 1) and 2), we verify the relation 3). We can also verify 3) by deforming the contours of the analytic continuations (see Fig. 6.1). The property 4) follows from 1) to 3). (cid:3) Remark. In the claim 3) of Proposition 6.8, not all monodromy relations which we read from Fig. 6.1 are written out. By deforming the paths in the figure, it is easy to deduce relations, for example: ρ(Tp1,1TyT −1 p1,1) = ρ(Tz′′ ), ρ(T −1 z Tp3,1Tz) = ρ(T −1 p′ 3 ,2Tz′′ ) = ρ(Tp3,2 ). z′′ Tp′′ 1,1), ρ(T −1 39 We can also observe relations among the generators in the claim 4), for example, ρ(Tp1,1)C10C20ρ(Tp1,2) = C10C20. The determination of the minimal set of relations is left for a future study. Also some simplifications in the matrix expressions, like C10C20 = (-1)⊕ ( 0 -1 -1 0 )⊕ (-1), [] may have some interpretations. -1 0 )⊕ ( 0 -1 1 (x, x) = w(k) 6.4. Mirror symmetry of Reye congruences. Over the line ℓ0 : x = y = 0 the six period integrals contained in Π(x, y) reduce to four independent integrals due to the degeneration w(k) 2 (x, x) (k = 1, 2). This is related to the symmetry under the exchange zi ↔ wi of the defining equations fi = 0 of X sp 0 when a = b. More generally, taking the automorphisms of X sp into account, this 0 symmetry appears when a = µkb with µ5 = 1, i.e., when x = y. 32 , the involution zi ↔ wi(∼= Z2) acting on X sp Proposition 6.9. When x = y 6= 1 0 has no fixed point. This action naturally lifts to a fixed point free Z2 action on the crepant resolution X∗0 . Taking a quotient by this, we obtain a Calabi-Yau threefold X∗ = X∗0 /Z2 with the Hodge numbers h1,1(X∗) = 26, h2,1(X∗) = 1. 0 when a = b (a = µkb in general). Proof. As above, it is clear from the form of the defining equations that the in- volution zi ↔ wi acts on X sp It is also straightforward to see that if a = b 6= − 1 2 , there is no solution for fi = 0 with zi = wi except zi = wi = 0 (i = 1, ..., 5). Clearly the involution acts on the singular loci. Hence it lifts to a fixed point free Z2 action on the crepant res- olution X∗0 when x = y 6= 1 32 . Since h0,1( X∗0 ) = h0,2( X∗0 ) = 0 for the res- olution, we have h0,1(X∗) = h0,2(X∗) = 0 for the quotient. The calculations e(X∗) = e( X∗0 )/2 = 50 and #X∗(Fp) = # X∗0 (Fp)/2 are valid for the free quo- tient. Hence the proof of Proposition 5.13 applies to the present case, and we have h1,1(X∗) = 26, h2,1(X∗) = 1. (cid:3) In [HT, Propositions 2.9,2.10], we observed that, when x = y = 1 32 , one ordinary double point appears in X sp 0 as a fixed point of the involution, and this results in a singular point of X∗ where a lens space (∼= S3/Z2) vanishes. In fact, this property has been predicted by noting a specific form of the Picard-Lefschetz monodromy [EvS] in their study of 4th order differential equations (see also [AEvSZ]). In order to connect the vanishing lens space directly with the Picard-Lefschetz monodromy, let us introduce the following monodromy matrices: (6.9) Rα1 = ρ(T −1 p1,1Tp1,2), R0 = ρ(TxTy), R 1 = ρ(Tp2), 32 Rα2 = ρ(T −1 p3,2Tp3,1), R∞ = ρ(Tz). As we see in Fig.6.1, these represent the monodromy transformations of Π(x, y) around the intersections of ℓ0 ∼= P1 with the discriminant, and satisfy a relation R∞Rα2 Rα1 R0R 1 These correspond to the matrices Mα1, M0, M 1 Table 1]. Explicitly we evaluate the matrices (6.9) as follows: = id. 32 , Mα2, M∞ of Π(x) given in [HT, 32 40 0 1 0 0 0 0 0 1 0 0 0 0 1 1 1 0 0 0 17 20 15 1 0 0 17 15 20 0 1 0 -20 -18 -18 -1 -1 1  , R∞ : w(3)(cid:17) ⊕ t(cid:16) 1 2 1 0 0 0 0 1 0 1 0 0 0 0 0 0 1 0 0 0 0 0 0 1 0 0 0 0 0 0 1 0 0 0 0 0 0 1  ,  , R 1 32 : 0 -6 41 -17 -17 6 6 15 -6 -2 4 1 3 4 0 3 -2 1 -72 28 23 -13 -9 -28 -72 23 28 -9 -13 -28 -30 18 18 -4 -4 -9  , (w(1) 1 − w(1) 1 ), 1 2 (w(2) 2 )(cid:17) 1 − w(2) ∞ instead of Rα2 since the matrices in [HT, Table 1] = id. Now we define Π(x, y) by Rα1 : 11 -7 -7 1 1 2 -1 -5 0 1 1 5 5 -5 -1 1 0 1 35 -20 -29 3 5 7 35 -29 -20 5 3 7 -50 35 35 -5 -5 -9  , R0 : 1 5 5 0 0 2 0 2 -1 -2 2 0 0 -1 2 2 -2 0 0 -12 -13 0 1 -5 0 -13 -12 1 0 -5 0 0 0 0 0 1 R∞Rα2R−1 ∞ : where we consider R∞Rα2R−1 satisfy Mα2M∞Mα1M0M 1 (w(2) 1 + w(1) 2 ), (w(1) 32 1 2 t(cid:16)w0, 1 2 1 + w(2) 2 ), 1 2 so that the second summand becomes t(0, 0) on the line ℓ0(x = y). It is straight- forward to see the following property: Proposition 6.10. In terms of the period integral Π(x, y) = P Π(x, y), we have the decomposition where Nk's are 2× 2 matrices and we set Rk = PRkP−1 with {Rk}4 , Mα2, M∞}. R 1 32 , R∞Rα2 R−1 From the explicit forms of (6.9), it is clear that R 1 ∞ , R∞} and {Mk}4 Rk = Mk ⊕ Nk (k = 1, .., 5), k=1 = {Mα1, M0, M 1 32 represents the Picard- Lefschetz monodromy of the vanishing cycle which appears in the fiber over x = y = 1 32 , from which we identify w(3)(x, y) as the period integral of the vanishing cycle. We note that w(3)(x, y) is contained in Π(x, y) with the prefactor 1 2 . (If this prefactor were taken to be 1, P should be symplectic with respect to Σ0.) From the with the Picard- above proposition and Proposition 6.9, we can now identify M 1 Lefschetz monodromy of the vanishing lens space (S3/Z2) in X∗ for x = y = 1 32 which we described above. 32 32 k=1 = {Rα1, R0, Finally we remark that both R0 and R∞ have the same Jordan normal form J(1, 4) ⊕ J(1, 2) with eigenvalues 1. Proposition 6.10 implies that the period inte- gral Π(x, y) is compatible with the Jordan decomposition and the first summand of Π(x, y) shows the maximally unipotent monodromies both at x = 0 and ∞. The mirror geometry which arises from x = 0 has been identified with the Reye congruence Calabi-Yau threefold, and that from x = ∞ has been identified with a new Calabi-Yau manifold which doubly covers the generic Hessian quintic with ramification locus being a smooth curve of genus 26 and degree 20. 41 7. Special families of Steinerian and Hessian quintics 7.1. Steinerian and Hessian quintics. Here we discuss the special family of the Steinerian and Hessian quintics defined by (2.6) and (2.7), respectively, for a = b. Together with the mirror family X∗P1, we summarize the related families over P1 by writing the generic fibers: X∗0 X sp 0 ϕ (7.1) /Z2 X∗ Usp !❈❈❈❈❈❈❈❈ Ssp ~⑤⑤⑤⑤⑤⑤⑤⑤ Ysp Y ∗ ρ 2:1 . !❈❈❈❈❈❈❈❈ Hsp The Steinerian quintic Ssp is defined as the determinantal quintics Z sp for a = b. Usp in the diagram provides a partial resolutions of Ssp, and is given as X sp 2 for a = b. There is a natural projection from Usp to the Hessian quintic Hsp, i.e., the determinantal quintic X sp,♯ 1 = Z sp for a = b. 2 0 a = b, i.e., a = b ∈ C∗ withQ4 In what follows, we describe the singularity of the Hessian quintic Hsp for generic k,l=0(µk a + µl b + 1) 6= 0. Then we define the double covering Ysp → Hsp branched along the singular loci of Hsp. It is expected that there is a crepant resolution ρ : Y ∗sp → Ysp. 7.2. Singular loci of Hsp. The Hessian quintic Hsp is defined in P4 λ by the equa- tion (2.7) with a = b ∈ C∗, which may be written det Aλ = 0. We note that this is actually defined for x = −a5 ∈ P1 due to the automorphism Hsp ⊂ P4 λ. Since Hsp is a hypersurface, it is rather easy to determine the singular loci by the Ja- cobian criterion. To describe the singular loci, let us denote the projective space λ = he∗1, e∗2, ..., e∗5i choosing a C-bases e∗i (i = 1, ..., 5). As before, we denote by P4 the coordinate points and lines, respectively, by [e∗i ] and Lij = he∗i , e∗ji. We also introduce the lines: where the indices are considered cyclic or modulo 5. Mi = he∗i−2 + ab e∗i−1, e∗i i (i = 1, 2, ..., 5), Proposition 7.1. For generic a = b, we have: 1) the singular loci of Hsp contain a component of a curve CE given by the 4 × 4 Pfaffians of 2) CE is a smooth genus one curve of degree 5 in P5 quintic. λ , i.e., an elliptic normal Proof. Our proof of 1) is based on the calculations of the primary decomposition of the Jacobian ideal. As we describe below, there appear 10 lines in the irreducible components of the singular loci. It is efficient to take the saturations repeatedly with respect to the ideals representing these lines. The claimed matrix form of the ideal may be deduced from the minimal resolution of the ideal, and has been determined by taking suitable linear combinations of the generators. 42 (7.2)  0 λ2 −aλ5 −aλ1 λ5 λ4 −aλ2 −aλ3 −λ2 0 λ1 0 aλ5 −λ4 −aλ4 λ3 0 aλ2 −λ1 aλ1 −λ3 aλ4 aλ3 −λ5 0 .  / / ! ~ ! o o e4 e4 e1 e4 e5 e3 e5 e5 e3 2e e3 Ysp 2:1 e1 2e e1 Usp 2e p2 e1* CE * e2 M2 M3 M4 M1 M5 Hsp * e5 e3 * * e4 Figure 7.1. The partial resolution Usp → Hsp of the Hessian quintic and the double covering Ysp → Hsp for a generic a = b ∈ C∗. CE is an elliptic normal quintic. By the partial resolution, the A1 singularities along the lines Mi(i = 1, .., 5) and CE are resolved in Usp. The A3 singularities along the coordinate lines Li i+1 are In blown up to A1 singularities along two lines for each Li i+1. 2 ([e∗i ]) ∼= P2, there exist A1-singularities along the broken each p−1 lines and the coordinate line hei+2, ei+3i. The claim 2) is a consequence of 1), since CE is a Pfaffian variety of 5 × 5 anti- symmetric matrices which may be identified as a linear section of the Grassmannian G(2, 5). The smootheness is verified by the Jacobian criterion. (cid:3) Remark 7.2. When a10 + 11a5 − 1 = 0, CE becomes nodal at one point (at [1 : 1 : 1 : 1 : 1] up to automorphisms of Hsp ⊂ P4 λ for every solution a). When a = 0 (resp. ∞), CE becomes reducible: CE = L13 ∪ L35 ∪ L52 ∪ L24 ∪ L41 (reps. CE = L12 ∪ L23 ∪ L34 ∪ L45 ∪ L51). See [Hu] for the geometry of elliptic normal [] quintics. By studying the Jacobian ideal of Hsp in details, we obtain the structure of the singularities in the special Hessian quintic for generic a = b as follows: Proposition 7.3. 1) For generic a = b ∈ C∗, the special Hessian quintic Hsp is singular along the 5 coordinate lines Li i+1 and 5 lines Mi (i=1,..,5) and also the smooth elliptic normal quintic CE. The type of singularities are of A3-type along the lines Li i+1 and of A1-type along the lines Mi and CE. These irreducible components intersect at 10 points as shown schematically in Fig. 7.1. 2) The singular loci of Hsp coincide set-theoretically with {[λ] ∈ P4 λ rk(Aλ) ≤ 3}. Proof. 1) Singular loci are determined by studying the Jacobian ideal as described in the proof of the previous proposition. The type of the singularities are deter- mined by taking the local coordinates of the normal bundle at generic points of the irreducible components. 2) The loci of rk(Aλ) ≤ 3 are determined by the ideal generated by the 4 × 4 minors of Aλ. We compare the primary decompositions of this ideal with that of the Jacobian ideal. The claim follows since we verify that the radicals of each component coincide. (cid:3) 43 Remark 7.4. The Hessian quintic Hsp is the special quintic hypersurface X sp,♯ 0 with a = b. If a 6= b (more generally a 6= µkb with µ5 = 1), it is easy to observe that the irreducible component CE disappears from the singular loci. This explains the additional factor Q4 k=0(a − µkb) in the discriminant (2.9). We note that the ) are resolved by the partial resolution p2 : A1-singularities in Hsp (resp. Usp → Hsp (resp. p2 : X sp ). The configuration of the singularities in Usp is depicted in Fig. 7.1. As depicted in the figure, there appear the following 25 lines along which Usp has A1-singularities: Three lines in each fiber p−1 2 ([e∗i ]) = he∗i+2, e∗i+3, e∗i+4i given by X sp,♯ 2 → X sp,♯ 0 0 hei+2,−bei+3 + ei+4i, h−bei+2 + ei+3, ei+4i, hei+2, ei+3i (i = 1, ..., 5), and two lines in each inverse image p−1 2 (Li i+1) (i = 1, ..., 5) given by [ei+3] × Li i+1, [−b ei+3 + ei+4] × Li i+1 (i = 1, ..., 5). Since determining these singular loci is essentially the same as we did for X sp in 0 Proposition 5.3, we omit the details. Inspecting the configuration shown in Fig. 7.1, it is immediate to have the Euler number of Usp by using e(Hsp). [] Proposition 7.5. The Euler number e(Usp) = 0. Proof. We evaluate e(Usp) from the projection p2 : Usp → Hsp shown in Fig. 7.1. We note that e(p−1 2 ([e∗i ]) = e(P2) = 3. Also we note that for the generic point z of the coordinate line Li i+1 (resp. the line Mi), we have e(p−1 2 (z)) = 3e(P1) − 2 = 4 (resp. e(p−1 2 (z)) = e(P1) = 2 ). We also note that Mi ∩ Mj = φ. Now inspecting the Fig. 7.1 carefully, we can evaluate the Euler number as e(Usp) = e(Hsp) + 5{e(P2) − 1} + {e(CE) − 5}{e(P1) − 1} = −5 + 10 − 5 = 0, (cid:3) where we use the result e(Hsp) = −5 obtained in Proposition 3.6. 7.3. The ramified covering Ysp → Hsp. For the generic Hessian quintic H, we have found a ramified double covering Y → H which gives us a smooth Calabi- Yau threefold [HT, Theorem 3.14] with h2,1(Y ) = h2,1(X) = 26. The diagram (2.2) shows relations among the generic fibers of the related families over the 26- dimensional deformation space. We obtain the diagram (7.1) by restricting (2.2) to the special families over P1. k=1 λkAk for generic a = b ∈ C∗. We recall that our special family of the Hessian is defined by Hsp = (cid:8)[λ] ∈ λ detAλ = 0(cid:9) with Aλ =P5 Definition 7.6. Consider the weighted projective space P9(25, 15) with its (weighted) homogeneous coordinate [ξ : λ] = [ξ1 : ... : ξ5 : λ1 : ... : λ5]. We denote by ¯ϕλ : P9(25, 15) 99K P4 λ the natural projection to the second half of the components. We define Ysp ⊂ P9(25, 15) by the following (weighted) homogeneous equations: (7.3) ξiξj = ∆(Aλ)ij (1 ≤ i, j ≤ 5), Aλξ = 0, P4 where ∆(Aλ)ij (=: ∆ij) represents the ij-cofactor of the symmetric matrix Aλ. [] We read the above definition from [Ca, Sect.3]. We note that if det(Aλ) 6= 0, then we have ξ = 0 hence ∆(Aλ)ij = 0 (1 ≤ i, j ≤ 5), which is a contradiction. Therefore we have a map ϕλ : Ysp → Hsp which follows from ¯ϕλ : P9(25, 15) 99K P4 λ. 44 : ... : ± ∆15√∆11 λ ([λ]) =(cid:8)[± ∆11√∆11 : λ1 : ... : λ5](cid:9) . This completes the proof. Proposition 7.7. 1) The map ϕλ : Ysp → Hsp is surjective. Moreover it is a double covering ramified along the singular loci of Hsp (see Proposition 7.3). 2) The singular loci of Ysp consist of 5 lines Li i+1(i = 1, ..., 5) of A1-singularities, where Li i+1 represents the coordinate line Li i+1 ⊂ Hsp considered in Ysp. 3) The Euler number of Ysp is given by e(Ysp) = −10. Proof. 1) The singular loci Sing Hsp of Hsp coincides with the loci with rk(Aλ) ≤ 3 due to 2) of Proposition 7.3. If rk(Aλ) ≤ 3, then ∆ij = 0 (1 ≤ i, j ≤ 5). Hence, from (7.3), we have ξi = 0, which implies that ϕλ is bijective over Sing Hsp. Now take a point [λ] ∈ Hsp such that rk(Aλ) = 4, then we have rk(∆ij ) = 1 for the matrix of the cofactors. Then there exists ξ such that (∆ij ) = (ξiξj). We may assume, without loss of generality, that ∆11 6= 0. Solving ξ2 1 = ∆11, we obtain ϕ−1 2) Let us denote by Si(i = 1, ..., 5) the affine subsets {[ξ, λ] λi 6= 0} in the weighted projective space P9(25, 15). We note that Si ≃ C9(i = 1, ..., 5) are in the smooth loci of the weighted projective space, and Ysp is contained in the union ∪iSi of these affine subsets. Here we study the singular loci of Ysp on the affine subset S5 with the affine coordinates [ξ1 : ... : ξ5 : λ1 : ... : λ4 : 1], but the calculations on the other Si's are quite parallel due to the symmetry of the defining equations Ysp and Hsp. We obtain 20 equation from the defining equations (7.6) expressed by the affine coordinates of S5. We observe that one of the 5 equations from Aλξ = 0 can be solved by ξ5 = −ξ1 − λ4ξ4. Eliminating ξ5 by this, we have 19 equations for (ξ1, ξ2, ξ3, ξ4, λ1, λ2, λ3, λ4) and make the Jacobian matrix of size 8 × 19. Since dim Ysp=3, the Jacobian ideal of Ysp is generated by 5× 5 minors of the Jacobi ma- trix. By studying this Jacobian ideal in a straightforward way, we obtain L45 ∪ L51 for the singular loci of Ysp restricted on S5. Combined with the similar calculations for the other Si's, we determine the singular loci of Ysp as claimed. To determine the type of singularity, we work with the affine coordinates [ξ1 : ... : ξ5 : λ1 : ... : λ4 : 1] and focus on the line L51. Note that the corresponding line L51 ⊂ Hsp intersects with the line M2 at [λ] = [a2 : 0 : 0 : 0 : 1]. We describe the local geometry around the point [ξ : λ] = [05 : a2 : 0 : 0 : 0 : 1] by introducing the affine coordinates by [x1 : ... : x5 : a2 + y1 : y2 : y3 : y4 : 1]. We work on the defining equations (7.6) in the local ring C[x1, x2, x3, x4, x5,y1, y2, y3, y4]m0 at the origin. By inspecting the defining equations of Ysp, we see that x1, x2, y3 as well as x5 can be solved by other variables in the local ring. After eliminating these variables, we study the (eliminated) ideal of Ysp in the local ring C[x3, x4, y1, y2, y4]m′ at the origin. It turns out that the local geometry of Ysp is described by the two equations and g2 is quadric with respect to y4. g1 = g2 = 0 with g1, g2 ∈ C[x3, x4, y1, y2, y4]m′ Solving the quadric equation g2 = 0, we eliminate y4 from g1. Choosing suitable branch of the solutions, we obtain g1 = x3x4 + x2 0 4 + y1y2 2 + y2x2 3 − y1y2x2 3 + ··· , 0 where ··· represents the higher order terms of the total degree greater than four. From this local equation, we read the singularities along the line L51 are of A1- type generically. (We may also observe that, as before, the blowing-up along L51 introduces a smooth conic bundle with reducible fiber at the origin, i.e., over the intersection point of L51 and M2. ) 3) The singular loci Sing Hsp, i.e., the ramification loci of Ysp → Hsp, consists of 10 lines Li i+1, Mi and the curve CE which intersect as shown in Fig. 7.1. Combined 45 with e(Hsp) = −5 in Proposition 3.6, we have e(Ysp) = 2 {e(Hsp) − e(Sing Hsp)} + e(Sing Hsp) = 2 × (−5 − 0) + 0 = −10. (cid:3) P = A1, A2, ..., Ak and Aλ =P5 Remark 7.8. The A1-singularities along the ramification loci of Hsp are resolved by the covering Ysp → Hsp. This is, in fact, a general property valid for a normal variety Ysp. If we consider the generic Hessian quintic H with a regular linear system k=1 λkAk, then the singular loci of H consist of a curve of A1-singularity. Hence, Definition 7.6 applied for generic Hessian quintic H gives us a smooth Calabi-Yau threefold ramified along the curve, which is Y in the theorem of the subsection 2.1, see also the diagram (2.5). This explicit realization of the geometry Y in P4(25, 15) should have corresponding descriptions in physics [] [Hor], [JKLMR]. As is shown in the diagram (7.1), we expect a crepant resolution ρ : Y ∗sp → Ysp with a Calabi-Yau manifold Y ∗sp which satisfies (h1,1(Y ∗sp), h2,1(Y ∗sp)) = (h1,1(X∗), h2,1(X∗)) = (26, 1) for the Hodge numbers. The existence of Y ∗sp with these ex- pected properties is left for future study. References [AEvSZ] G. Almkvist, C. van Enckevort, D. van Straten, W. Zudilin, Tables of Calabi -- Yau equa- tions, arXiv:math/0507430. [AGV] V.I. Arnold, S.M. Gusein-Zade and A.N. Varchenko, Singularities of Differential Maps, Volume I. Birkhäuser (1985). [BaN] W. Barth and I. Nieto, Abelian Surfaces of type (1,3) and Quartic Surfaces with 16 Skew Lines, J. Alg. Geom. 3 (1994) 173-222. [Ba] V. Batyrev, Dual polyhedra and mirror symmetry for Calabi-Yau hypersurfaces in toric varieties, J. Alg. Geom. 3(1994), 493-535. [BaBo] V. Batyrev and L. Borisov, On Calabi-Yau Complete Intersections in Toric Varieties, Higher-dimensional complex varieties (Trento, 1994), 39 -- 65, de Gruyter, Berlin, 1996. [BaCo] V. Batyrev and D.A. Cox, On the Hodge structure of projective hypersurfaces in toric varieties, Duke Math. J. 75 (1994) 293 -- 338. [BeH] P. Berglund and T. Hübsch, A generalized construction of mirror manifolds, Nuclear Phys. B 393 (1993), no. 1-2, 377 -- 391. [BoCa] L. Borisov and A. Caldararu, The Pfaffian-Grassmannian derived equivalence, J. Alge- braic Geom. 18 (2009), no. 2, 201 -- 222.math/0608404. [CdOGP] P. Candelas, X.C. de la Ossa, P.S. Green, and L.Parkes, A pair of Calabi-Yau manifolds as an exactly soluble superconformal theory, Nucl.Phys. B356(1991), 21-74. [Ca] F. Catanese, Commutative algebra methods and equations of regular surfaces. Algebraic ge- ometry, Bucharest 1982 (Bucharest, 1982), 68 -- 111, Lecture Notes in Math., 1056, Springer, Berlin, 1984. [Co] F. Cossec, Reye congruence, Transactions of the A.M.S. 280 (1983), 737 -- 751. [DGPS] W. Decker, G.-M. Greuel, G. Pfister and H. Schönemann, Singular 3-1-3 -- A computer algebra system for polynomial computations, http://www.singular.uni-kl.de (2011). [DGJ] C. Doran, B. Greene and S. Judes, Families of quintic Calabi-Yau 3-folds with discrete symmetries, Comm. Math. Phys. 280 (2008), no. 3, 675 -- 725 46 [EvS] C. Enckevort and D. van Straten, Monodromy calculations of fourth order equations of Calabi-Yau type,in Mirror symmetry. V, 539 -- 559, AMS/IP Stud. Adv. Math., 38, Amer. Math. Soc., Providence, RI, 2006, math.AG/0412539. [FK] E. Freitag and R. Kiehl, Etale Cohomology and The Weil Conjecture, Springer-Verlag, Berlin and New York (1988). [GS] D. R. Grayson and M. E. Stillman, Macaulay2, a software system for research in algebraic geometry, Available at http://www.math.uiuc.edu/Macaulay2/. [vGN] B. van Geemen and N. O. Nygaard, On the Geometry and Arithmetics of Some Siegel Modular Threefolds, Jour. of Number Theory 53(1995), 45 -- 87. [Gep] D.Gepner, Exactly solvable string compactifications on manifolds of SU(n) holonomy, Phys.Lett.199B(1987)380. [GP] B.R.Greene and M.R.Plesser, Duality in Calabi-Yau moduli space, Nucl.Phys. B338 (1990) 15-37. [GKZ] I.M. Gel'fand, A. V. Zelevinski, and M.M. Kapranov, Equations of hypergeometric type and toric varieties, Funktsional Anal. i. Prilozhen. 23 (1989), 12 -- 26; English transl. Functional Anal. Appl. 23(1989), 94 -- 106. [Gr] P. A. Griffiths, On the periods of certain rational integrals. I, II. Ann. of Math. (2) 90 (1969), 460-495; ibid. (2) 90 1969 496 -- 541. [Har] R. Hartshorne, Algebraic geometry. Graduate Texts in Mathematics 52, Springer-Verlag, New York, Heidelberg, Berlin, 1977. [Ho1] S. Hosono, Local Mirror Symmetry and Type IIA Monodromy of Calabi-Yau Manifolds, Adv. Theor. Math. Phys. 4 (2000), 335 -- 376. [Ho2] S. Hosono, Central charges, symplectic forms, and hypergeometric series in local mirror symmetry, in "Mirror Symmetry V", S.-T.Yau, N. Yui and J. Lewis (eds), IP/AMS (2006), 405 -- 439. [HKTY] S. Hosono, A. Klemm, S. Theisen and S.-T. Yau, Mirror Symmetry, Mirror Map and Applications to complete Intersection Calabi-Yau Spaces, Nucl. Phys. B433(1995)501 -- 554. [HLY] S. Hosono, B.H. Lian, and S.-T. Yau, GKZ-Generalized hypergeometric systems in mirror symmetry of Calabi-Yau hypersurfaces, Commun. Math. Phys. 182 (1996) 535 -- 577. [HT] S. Hosono and . H. Takagi, Mirror Symmetry and Projective Geometry of Reye congruences I, preprint arXiv:1101.2746(2011) to appear in J. Alg. Geom. [HW] F. Hidaka, K. Watanabe, Normal Gorenstein surfaces with ample anti-canonical divisor, Tokyo J. Math. 4 (1981), no. 2, 319 -- 330. [Hor] K. Hori, Duality In Two-Dimensional (2,2) Supersymmetric Non-Abelian Gauge Theories, arXiv:1104.2853, hep-th.(2011). [Hu] K. Hulek, Projective geometry of elliptic curves, Astérisque No. 137 (1986), 143 pp. [HSvGvS] K. Hulek, J. Spandaw, B. van Geemen, and D. van Straten, The modularity of the Barth-Nieto quintic and its relatives, Adv. Geom. 1 (2001) 263 -- 289. [JKLMR] H. Jockers, V. Kumar, J. M. Lapan, D. R. Morrison, M. Romo, Nonabelian 2D Gauge Theories for Determinantal Calabi-Yau Varieties, arXiv:hep-th/1205.3192. [Ko] M. Kontsevich, Homological algebra of mirror symmetry, Proceedings of the Interna- tional Congress of Mathematicians (Zürich, 1994) Birkhäuser (1995) pp. 120 -- 139. [Ku] A. Kuznetsov, Homological projective duality for Grassmannians of lines, arXiv:math/0610957. [Mo] D. R. Morrison, Picard-Fuchs equations and mirror maps for hypersurfaces, in "Essays on Mirror Manifolds", Ed. S.-T.Yau, International Press, Hong Kong (1992) 241-264. [Ol] C. Oliva, Algebraic cycles and Hodge theory on generalized Reye congruences, Compositio Math. 92 (1994), 1 -- 22. [Ro] E.A. Rødland, The Pfaffian Calabi-Yau, its Mirror and their link to the Grassmannian G(2, 7), Compositio Math. 122 (2000), no. 2, 135 -- 149, math.AG/9801092. [Ty] A.N. Tyurin, On intersections of quadrics, Russian Math. Surveys 30 (1975), 51 -- 105. [Yau] "Essays on Mirror Manifolds", Ed. S.-T.Yau, International Press, Hong Kong (1992). Graduate School of Mathematical Sciences, University of Tokyo, Meguro-ku, Tokyo 153-8914, Japan e-mail addresses: [email protected], [email protected] 47
1606.05810	4	1606	2018-06-01T08:55:06	The slice spectral sequence for singular schemes and applications	[ "math.AG", "math.KT" ]	We examine the slice spectral sequence for the cohomology of singular schemes with respect to various motivic T-spectra, especially the motivic cobordism spectrum. When the base field k admits resolution of singularities and X is a scheme of finite type over k, we show that Voevodsky's slice filtration leads to a spectral sequence for MGL(X) whose terms are the motivic cohomology groups of X defined using the cdh-hypercohomology. As a consequence, we establish an isomorphism between certain geometric parts of the motivic cobordism and motivic cohomology of X. A similar spectral sequence for the connective K-theory leads to a cycle class map from the motivic cohomology to the homotopy invariant K-theory of X. We show that this cycle class map is injective for projective schemes. We also deduce applications to the torsion in the motivic cohomology of singular schemes.	math.AG	math	THE SLICE SPECTRAL SEQUENCE FOR SINGULAR SCHEMES AND APPLICATIONS AMALENDU KRISHNA AND PABLO PELAEZ Abstract. We examine the slice spectral sequence for the cohomology of singular schemes with respect to various motivic T -spectra, especially the motivic cobordism spectrum. When the base field k admits resolution of singularities and X is a scheme of finite type over k, we show that Voevodsky's slice filtration leads to a spectral sequence for MGLX whose terms are the motivic cohomology groups of X defined using the cdh-hypercohomology. As a consequence, we establish an isomorphism between certain geometric parts of the motivic cobordism and motivic cohomology of X. A similar spectral sequence for the connective K-theory leads to a cycle class map from the motivic cohomology to the homotopy invariant K-theory of X. We show that this cycle class map is injective for a large class of projective schemes. We also deduce applications to the torsion in the motivic cohomology of singular schemes. Contents Introduction 1 1. 3 2. A descent theorem for motivic spectra 9 3. Motivic cohomology of singular schemes 11 4. Slice spectral sequence for singular schemes 18 5. Applications I: Comparing cobordism, K-theory and cohomology 22 6. The Chern classes on KH-theory 7. Applications II: Intermediate Jacobian and Abel-Jacobi map for singular schemes 28 32 8. Applications III: Roitman torsion and cycle class map References 34 1. Introduction The motivic homotopy theory of schemes was put on a firm foundation by Voevodsky and his coauthors beginning with the work of Morel and Voevodsky [67] and its stable counterpart [84]. It was observed by Voevodsky [86] that the motivic T -spectra in the stable homotopy category SHX over a noetherian scheme X of finite Krull dimension can be understood via their slice filtration. This slice filtration leads to spectral sequences which then become a very powerful tool in computing various cohomology theories for smooth schemes over X. The main problem in the study of the slice filtration for a given motivic T -spectrum is twofold: the identification of its slices and the analysis of the convergence properties for the corresponding slice spectral sequence. When k is a field which admits resolution of singulari- ties, the slices for many of these motivic T -spectra in SHk are now known. In particular, we can compute these generalized cohomology groups of smooth schemes over k using the slice spectral sequence. 2010 Mathematics Subject Classification. Primary 14C25, 14C35, 14F42, 19E08, 19E15. Key words and phrases. Algebraic Cobordism, Milnor K-theory, Motivic Homotopy Theory, Motivic Spec- tral Sequence, K-theory, Slice Filtration, Singular Schemes. 1 2 AMALENDU KRISHNA AND PABLO PELAEZ In this paper, we study some descent property of the motivic T -spectra in SHX when X is a possibly singular scheme of finite type over k. This descent property tells us that the cohomology groups of a scheme Y ∈ SmX, associated to an absolute motivic T -spectra in SHX [22, §1.2], can be computed using only SHk. Even though our methods apply to any of these absolute T -spectra, we restrict our study to the motivic cobordism spectrum MGLX. We show using the above descent property of motivic spectra that MGLX can be computed using the motivic cohomology groups of X. Recall from [27, Defs. 4.3, 9.2] that the motivic cohomology groups of X are defined to be the cdh-hypercohomology groups H p(X, Z(q)) = Hp−2q k, 0)cdh). Using these motivic cohomology groups, we show: cdh (X, C∗zequi(Aq Theorem 1.1. Let k be a field which admits resolution of singularities and let X be a separated scheme of finite type over k. Then for any integer n ∈ Z, there is a strongly convergent spectral sequence (1.2) and the differentials of this spectral sequence are given by dr : Ep,q more, this spectral sequence degenerates with rational coefficients. Ep,q 2 = H p−q(X, Z(n − q)) ⊗Z Lq ⇒ MGLp+q,n(X) r → Ep+r,q−r+1 r . Further- If k is a perfect field of positive characteristic p, we obtain a similar spectral sequence after inverting p except that we can not guarantee strong convergence unless X is smooth over k (see 4.25). As a consequence of 1.1 and its positive characteristic version, we get the following relation between the motivic cobordism and cohomology of singular schemes. Theorem 1.3. Let k be a field which admits resolution of singularities (resp. a perfect field of positive characteristic p). Then for any separated (resp. smooth) scheme X of finite type over k and dimension d and every i ≥ 0, the edge map in the spectral sequence (1.2): resp. νX : MGL2d+i,d+i(X)⊗Z νX : MGL2d+i,d+i(X) → H 2d+i(X, Z(d + i)) p ] → H 2d+i(X, Z(d + i))⊗Z Z[ 1 Z[ 1 p ] is an isomorphism. We apply our descent result to obtain a similar spectral sequence for the connective KH- theory, KGL0 (see §5). We use this spectral sequence and the canonical map CKH(−) → KH(−) from the connective KH-theory to obtain the following cycle class map from the motivic cohomology of a singular scheme to its homotopy invariant K-theory. Theorem 1.4. Let k be a field of exponential characteristic p and let X be a separated scheme of dimension d which is of finite type over k. Then the map KGL0 for every integer i ≥ 0, an isomorphism CKH 2d+i,d+i(X)⊗Z X → s0 KGLX ∼= HZ induces ∼=−→ H 2d+i(X, Z(d + i))⊗Z Z[ 1 p ] Z[ 1 p ]. In particular, there is a natural cycle class map cyci : H 2d+i(X, Z(d + i))⊗Z Z[ 1 p ] → KHi(X)⊗Z Z[ 1 p ]. We use this cycle class map and the Chern class maps from the homotopy invariant K- theory to the Deligne cohomology of schemes over C to construct intermediate Jacobians and Abel-Jacobi maps for the motivic cohomology of singular schemes over C. More precisely, we prove the following. This generalizes intermediate Jacobians and Abel-Jacobi maps of Griffiths and the torsion theorem of Roitman for smooth schemes. THE SLICE SPECTRAL SEQUENCE FOR SINGULAR SCHEMES AND APPLICATIONS 3 Theorem 1.5. Let X be a projective scheme over C of dimension d. Assume that either d ≤ 2 or X is regular in codimension one. Then, there is a semi-abelian variety J d(X) and an Abel- Jacobi map AJd X : H 2d(X, Z(d))deg 0 → J d(X) which is surjective and whose restriction to the torsion subgroups is an isomorphism. In a related work, Kohrita (see [50, Corollary 6.5]) has constructed an Abel-Jacobi map for the Lichtenbaum motivic cohomology H 2d L (X, Z(d)) of singular schemes over C using a differ- ent technique. He has also proven a version of Roitman torsion theorem for the Lichtenbaum motivic cohomology. The natural map H 2d(X, Z(d)) → H 2d L (X, Z(d)) is not an isomorphism in general if d ≥ 3. Note also that the Roitman torsion theorem for H 2d(X, Z(d)) is a priori a finer statement than that for the analogous Lichtenbaum cohomology. Using Theorem 1.5, we prove the following property of the cycle class map of 1.4 which is our final result. The analogous result for smooth projective schemes was proven by Marc Levine [59, Thm. 3.2]. More generally, Levine shows that a relative Chow group of 0-cycles on a normal projective scheme over C injects inside K0(X). Theorem 1.6. Let X be a projective scheme of dimension d over C. Assume that either d ≤ 2 or X is regular in codimension one. Then the cycle class map cyc0 : H 2d(X, Z(d)) → KH0(X) is injective. We end this section with the comment that our motivation behind this work was to exploit powerful tools of the motivic homotopy theory to study several questions about the motivic cohomology and K-theory of singular schemes, which were previously known only for smooth schemes. We hope that the methods and the techniques of our proofs can be advanced further to answer many other cohomological questions about singular schemes. We refer to [55] for more results based on the techniques of this text. 2. A descent theorem for motivic spectra In this section, we set up our notations, discuss various model structures we use in our proofs and show the Quillen adjunction property of many functors among these model structures. The main objective of this section is to prove a cdh-descent property of the motivic T -spectra, see 2.14. 2.1. Notations and preliminary results. Let k be a perfect field of exponential charac- teristic p (in some instances we will require that the field k admits resolution of singularities [87, Def. 4.1]). We will write Schk for the category of separated schemes of finite type over k and Smk for the full subcategory of Schk consisting of smooth schemes over k. If X ∈ Schk, let SmX denote the full subcategory of Schk consisting of smooth schemes over X. We will write (Smk)N is (resp. (SmX )N is, (Schk)cdh, (Schk)N is) for Smk equipped with the Nis- nevich topology (resp. SmX equipped with the Nisnevich topology, Schk equipped with the cdh-topology, Schk equipped with the Nisnevich topology). The product X ×Spec k Y will be denoted by X × Y . Let M (resp. MX , Mcdh) be the category of pointed simplicial presheaves on Smk (resp. SmX, Schk) equipped with the motivic model structure described in [39] considering the Nisnevich topology on Smk (resp. Nisnevich topology on SmX , cdh-topology on Schk) and the affine line A1 k as an interval. A simplicial presheaf will often be called a motivic space. s = S0 We define T in M (resp. MX , Mcdh) as the pointed simplicial presheaf represented by s ∧ S1 S1 s denotes the simplicial circle. Given an arbitrary integer r ≥ 1, let Sr t ) denote the iterated smash t ∧ ··· ∧ S1 s ∧ ··· ∧ S1 product of S1 t will be k \ {0}) pointed by 1, and S1 t ); S0 k \ {0} (resp. A1 t ) with r-factors: S1 s (resp. Sr s (resp. S1 X \ {0}, A1 t , where S1 t is A1 s (resp. S1 4 AMALENDU KRISHNA AND PABLO PELAEZ by definition equal to the pointed simplicial presheaf represented by the base scheme Spec k (resp. X, Spec k). s ◦ Σn Since T is cofibrant in M (resp. MX, Mcdh) we can apply freely the results in [36, §8]. Let Spt(M) (resp. Spt(MX ), Spt(Mcdh)) denote the category of symmetric T -spectra on M (resp. MX, Mcdh) equipped with the motivic model structure defined in [36, 8.7]. We will write SH (resp. SHX , SHcdh) for the homotopy category of Spt(M) (resp. Spt(MX ), Spt(Mcdh)) which is a tensor triangulated category. For any two integers m, n ∈ Z, let Σm,n denote the automorphism Σm−n t : SH → SH (this also makes sense in SHX and SHcdh). T for Σ2n,n, and E ∧ F for the smash product of E, F ∈ SH (resp. SHX, We will write Σn SHcdh). Given a simplicial presheaf A, we will write A+ for the pointed simplicial presheaf obtained by adding a disjoint base point (isomorphic to the base scheme) to A. For any B ∈ M, let Σ∞ T (B) denote the object (B, T ∧ B,··· ) ∈ Spt(M). This functor makes sense for objects in Mcdh and MX as well. If F : A → B is a functor with right adjoint G : B → A, we shall say that (F, G) : A → B is an adjunction. We shall freely use the language of model and triangulated categories. We will write Σ1 for the suspension functor in a triangulated category, and for n ≥ 0 (resp. n < 0), Σn will be the suspension (resp. desuspension) functor iterated n (resp. −n) times. We will use the following notation in all the categories under consideration: ∗ will denote the terminal object, and ∼= will denote that a map (resp. functor) is an isomorphism (resp. equivalence of categories). 2.2. Change of site. Let X ∈ Schk and let v : X → Spec k denote the structure map. We will write P reX (resp. P rek) for the category of pointed simplicial presheaves on SmX (resp. Schk). If X = Spec k where k is the base field, we will write P rek instead of P reX . These categories are equipped with the objectwise flasque model structure [39, §3]. To recall this model structure, we consider a finite set I of monomorphisms {Vi → U}i∈I for any U ∈ SmX. The categorical union ∪i∈I Vi is the coequalizer of the diagram `i,j∈I Vi×U Vj / `i∈I Vi formed in P reX. We denote by iI the induced monomorphism ∪i∈I Vi → U . Note that ∅ → U arises in this way. The push-out product of maps of iI and a map between simplicial sets exists in P reX. In particular, we are entitled to form the sets I sch clo (SmX ) = {iI (cid:3) (∂∆n ⊂ ∆n)+}I,n≥0 clo (SmX) = {iI (cid:3) (Λn J sch i ⊂ ∆n)+}I,n≥0,0≤i≤n, and where I is a finite set of monomorphisms {Vi → U}i∈I , U ∈ SmX; and iI : ∪i∈I induced monomorphism defined above. Vi → U is the A map between simplicial presheaves is called a closed objectwise fibration if it has the right lifting property with respect to J sch clo (SmX). A map u : E → F between simplicial presheaves is called a weak equivalence if E(U ) → F (U ) is a weak equivalence of simplicial sets for each U ∈ SmX. A closed objectwise cofibration is a map having the left lifting property with respect to every trivial closed objectwise fibration. Note that this notion of weak equivalence, cofibrations and fibrations makes sense for simplicial presheaves in= any category with finite products (e.g., Smk, Schk). It follows from [39, Thm. 3.7] that the above notion of weak equivalence, cofibrations and fibration forms a proper, simplicial and cellular model category structure on P rek, P reX and P rek. We call this the objectwise flasque model structure. Our reason for choosing this model structure is the following result. / / / THE SLICE SPECTRAL SEQUENCE FOR SINGULAR SCHEMES AND APPLICATIONS 5 Lemma 2.3. ([39, Lem. 6.2]) If V → U is a monomorphism in Smk (resp. SmX, Schk), then U+/V+ is cofibrant in the flasque model structure on P rek (resp. P reX , P rek). In particular, T n ∧ U+ is cofibrant for any n ≥ 0. It is clear that P reX (resp. P rek) is a cofibrantly generated model category with generat- clo (SmX) clo (Schk)) and generating trivial cofibrations J sch clo (SmX) (resp. I sch ing cofibrations I sch (resp. J sch clo (Schk)). Let π : (Schk)cdh → (Smk)N is be the continuous map of sites considered by Voevodsky in [87, §4]. We will write (π∗, π∗) : P rek → P rek, (v∗, v∗) : P rek → P reX for the adjunctions induced by π, v respectively. We will also consider the morphism of sites πX : (Schk)cdh → (SmX)N is and the cor- X, πX∗) : P reX → P rek. These adjunctions are related by the responding adjunction (π∗ following. Lemma 2.4. The following diagram commutes: π∗ P rek P rek ❈ ❈ ❈ ❈ ❈ ❈ ❈ v∗ πX ∗ !❈ P reX. Proof. We first notice that for every simplicial set K, Y ∈ Smk and Z ∈ SmX , one has v∗(K ⊗ Y+) = K ⊗ (Y × X)+ ∈ P reX and (2.5) π∗(K ⊗ Y+) = K ⊗ Y+ ∈ P rek; π∗ X(K ⊗ Z+) = K ⊗ Z+ ∈ P rek. We observe that π∗, v∗ commute with colimits since they are left adjoint; and that πX∗ also commutes with colimits since it is a restriction functor. Hence, it suffices to show that for every simplicial set K and every Y ∈ Smk, πX∗(π∗(K ⊗ Y+)) = v∗(K ⊗ Y+). Finally, a direct computation shows that πX∗(K ⊗ Y+) = K ⊗ (Y × X)+ ∈ P reX and we conclude by (2.5). Lemma 2.6. The adjunctions (π∗, π∗) : P rek → P rek, (v∗, v∗) : P rek → P reX and (π∗ X, πX∗) : P reX → P rek are all Quillen adjunctions. Moreover, πX∗ and π∗ preserve weak equivalences. (cid:3) clo (Schk), π∗(J sch Proof. We have seen above that all the three model categories (with the objectwise flasque model structure) are cofibrantly generated. Moreover, it follows from (2.5) that π∗(I sch clo (Smk)) ⊆ I sch clo (Smk)) ⊆ J sch clo (SmX) and π∗ clo (Schk). Hence, it follows from [35, Lem. 2.1.20] that (π∗, π∗), (v∗, v∗) and (π∗ X , πX∗) are Quillen adjunctions. The second part of the lemma is an immediate consequence of the observation that πX∗ and π∗ are restriction functors and the weak equivalences in the objectwise flasque model structure are defined schemewise. (cid:3) clo (Smk)) ⊆ J sch X(I sch clo (Schk); v∗(I sch clo (Schk), π∗ clo (Smk)) ⊆ J sch X(J sch clo (SmX)) ⊆ J sch clo (SmX)) ⊆ I sch clo (SmX ), v∗(J sch To show that the Quillen adjunction of 2.6 extends to the level of motivic model structures, we consider a distinguished square α (see [87, §2]): Y ′ Z ′ (2.7) Z / Y in (Smk)N is (resp. (SmX )N is, (Schk)cdh). We will write P (α) for the pushout of Z ← Z ′ → Y ′ in P rek (resp. P reX, P rek). / / ! / / / 6 AMALENDU KRISHNA AND PABLO PELAEZ The motivic model category M (resp. MX , Mcdh, Mf t) is the left Bousfield localization of P rek (resp. P reX , P rek, P rek) with respect to the following two sets of maps: (Schk)cdh, (Schk)N is), • P (α) → Y indexed by the distinguished squares in (Smk)N is (resp. • pY : Y × A1 k → Y for Y ∈ Smk (resp. Y ∈ SmX, Y ∈ Schk, Y ∈ Schk). (SmX )N is, Notice that since we are working with the flasque model structures, by [39, Thms. 4.8-4.9] it is possible to consider maps from the ordinary pushout P (α) instead of maps from the homotopy pushout of the diagram Z ← Z ′ → Y ′ (2.7). Remark 2.8. We will also consider the Nisnevich (resp. cdh) local model structure, i.e., the left Bousfield localization of P rek (resp. P rek) with respect to the set of maps: P (α) → Y indexed by the distinguished squares in (Smk)N is (resp. (Schk)cdh). We will abuse notation and write (π∗, π∗) : M → Mcdh, (v∗, v∗) : M → MX , (π∗ MX → Mcdh for the adjunctions induced by π, v and πX, respectively. Proposition 2.9. The adjunctions (π∗, π∗) : M → Mcdh, (v∗, v∗) : M → MX, (π∗ MX → Mcdh are Quillen adjunctions. Proof. We will give the argument for (π∗, π∗), since the other cases are parallel. Consider the commutative diagram X , πX∗) : X, πX∗) : π∗ P rek P rek id id M π∗ /❴❴ Mcdh, k → Y ) are weak equivalences in Mcdh. On the one hand, it is immediate that π∗(Y × A1 where the solid arrows are left Quillen functors by [34, Lem. 3.3.4(1)] and 2.6. Thus, it follows from [34, Def. 3.1.1(1)(b), Thm. 3.3.19] that it suffices to check that π∗(P (α) → Y ) and π∗(Y × A1 k → Y ) ∈ Mcdh; hence a weak equivalence in Mcdh. On the other hand, π∗ commutes with pushouts since it is a left adjoint functor. It follows therefore from (2.5) that π∗(P (α) → Y ) = (P (α) → Y ) ∈ Mcdh, hence a weak equivalence in Mcdh. k → Y ) = (Y × A1 (cid:3) We will write H (resp. HX, Hcdh) for the homotopy category of M (resp. MX, Mcdh) and X, RπX∗) : HX → Hcdh for the derived (Lπ∗, Rπ∗) : H → Hcdh, (Lv∗, Rv∗) : H → HX , (Lπ∗ adjunctions of the Quillen adjunctions in 2.9 (see [34, Thm. 3.3.20]). 2.10. A cdh-descent for motivic spectra. It follows from (2.5) that the adjunctions be- tween the categories of motivic spaces induce levelwise adjunctions (π∗, π∗) : Spt(M) → Spt(Mcdh), (v∗, v∗) : Spt(M) → Spt(MX), (π∗ X, πX∗) : Spt(MX) → Spt(Mcdh) between the corresponding categories of symmetric T -spectra such that the following diagram commutes (see (2.4)): (2.11) Spt(M) ▼ π∗ ▼ ▼ ▼ ▼ v∗ Spt(Mcdh) πX ∗ ▼ ▼ ▼ ▼ ▼ &▼ We further conclude from 2.9 and [36, Thm. 9.3] the following: Spt(MX ). Proposition 2.12. The pairs (1) (π∗, π∗) : Spt(M) → Spt(Mcdh), (2) (v∗, v∗) : Spt(M) → Spt(MX ) and / / / / / & THE SLICE SPECTRAL SEQUENCE FOR SINGULAR SCHEMES AND APPLICATIONS 7 (3) (π∗ X , πX∗) : Spt(MX) → Spt(Mcdh) are Quillen adjunctions between stable model categories. X(−). s ∧ Sb X ◦ Σm,n(−) ∼= Σm,n ◦ Lπ∗ We deduce from 2.12 that there are pairs of adjoint functors (Lπ∗, Rπ∗) : SH → SHcdh, (Lv∗, Rv∗) : SH → SHX and (Lπ∗ X , RπX∗) : SHX → SHcdh between the various stable homotopy categories of motivic T -spectra. We observe that for a ≥ b ≥ 0, the suspension functor Σa,b in SH (resp. SHX, SHcdh) is the derived functor of the left Quillen functor t ∧ E in Spt(M) (resp. Spt(MX), Spt(Mcdh)). Since the functors π∗, v∗, π∗ E 7→ Sa−b X are simplicial and symmetric monoidal, we deduce that they commute with the suspension functors Σm,n, i.e., for every m, n ∈ Z: Lπ∗ ◦ Σm,n(−) ∼= Σm,n ◦ Lπ∗(−), Lv∗ ◦ Σm,n(−) ∼= Σm,n ◦ Lv∗(−) and Lπ∗ Recall that Mf t is the motivic category for the Nisnevich topology in Schk. We will write Spt(Mf t) for the category of symmetric T -spectra on Mf t equipped with the stable model structure considered in [36, 8.7]. It is well known [44, p. 198] that Spt(Mf t) (resp. Spt(MX ), X ∈ Schk) is a simplicial model category [34, 9.1.6]. For E, E′ ∈ Spt(Mf t) (resp. Spt(MX)), we will write M ap(E, E′) (resp. M apX(E, E′)) for the simplicial set of maps from E to E′, i.e., the simplicial set with n-simplices of the form HomSpt(Mf t)(E ⊗ ∆n, E′) (resp. HomSpt(MX )(E ⊗ ∆n, E′)). Given f : X → X ′, we observe that the Quillen adjunction (f ∗, f∗) : Spt(MX ′) → Spt(MX) [5, Thm. 4.5.14] is enriched on simplicial sets, i.e. M apX(f ∗E′, E) ∼= M apX ′(E′, f∗E) for E ∈ Spt(MX ), E′ ∈ Spt(MX ′). homotopy theory [4, Cor. 1.7.18], [16, 2.3.11(2)], [15, Prop. 3.7]. Proposition 2.13. The functor Lv∗ is naturally equivalent to the composition RπX∗ ◦ Lπ∗. Proof. We observe that the following diagram of left Quillen functors commutes: The following result is a direct consequence of the proper base change theorem in motivic Spt(M) ▼ π∗ Spt(Mcdh) ▼ ▼ ▼ ▼ ▼ π∗ f t ▼ ▼ ▼ ▼ &▼ id Spt(Mf t). f tE → E′ be a functorial fibrant replacement of π∗ Let E be a motivic T -spectrum in Spt(M). Without any loss of generality, we can assume that E is cofibrant in Spt(M). Let ν : π∗ f tE in Spt(Mf t). The argument of Jardine in [44, pp. 198-199] shows that the restriction functor πX∗ maps weak equivalences in Spt(Mf t) into weak equivalences in Spt(MX ). Combining this f tE) = πX∗(π∗E) = v∗E → πX∗E′ is a weak with (2.11), we deduce that πX∗(ν) : πX∗(π∗ equivalence in Spt(MX ). Since E is cofibrant in Spt(M), Lv∗E ∼= v∗E. Hence, in order to conclude it suffices to show that E′ is fibrant in Spt(Mcdh). We shall use the following notation in the rest of the proof: for Y ∈ Schk, we will write Y E ∼= vY : Y → Spec (k) for the structure map. Notice that we have proved that Lv∗ πY ∗E′ in SHY . Consider a distinguished abstract blow-up square in Schk, i.e., a distinguished square in the lower cd-structure defined in [87, §2]: Y ′ Y E ∼= v∗ Z ′ i′ f ′ f Z / Y i / / & O O / / / 8 AMALENDU KRISHNA AND PABLO PELAEZ Y E) ∼= Rf∗L(vY ◦ f )∗E ∼= Rf∗πY ′∗E′ ∼= f∗πY ′∗E′ in Let j = i ◦ f ′. Then Rf∗Lf ∗(Lv∗ SHY ; where the last isomorphism follows from the fact that πY ′∗E′ is fibrant in Spt(MY ′), since E′ is fibrant in Spt(Mf t) and the restriction functor πY ′ : Spt(Mf t) → Spt(MY ′) is a right Quillen functor (using the same argument as in 2.12). Similarly, we conclude that Ri∗Li∗(Lv∗ Y E) ∼= i∗πZ∗E′ and Rj∗Lj∗(Lv∗ Y E) ∼= j∗πZ ′∗E′ in SHY . Thus, by [15, Prop. 3.7] we conclude that the commutative diagram: πY ∗E′ f∗πY ′∗E′ i∗πZ∗E′ / j∗πZ ′∗E′ is a homotopy cofiber square in Spt(MY ) [34, 13.5.8], thus also a homotopy fibre square since Spt(MY ) is a stable model category, i.e. its homotopy category is triangulated. Since Σ∞ T Y+ is cofibrant in Spt(MY ) and πY ∗E′, f∗πY ′∗E′, i∗πZ∗E′, j∗πZ ′∗E′ are fibrant; combining [34, 9.1.6 M7] and [34, 9.7.5(1)] we conclude that the induced commutative diagram is a homotopy fibre square of simplicial sets: M apY (Σ∞ T Y+, πY ∗E′) M apY (Σ∞ T Y+, f∗πY ′∗E′) M apY (Σ∞ T Y+, i∗πZ∗E′) / M apY (Σ∞ T Y+, j∗πZ ′∗E′) Since the adjunction (f ∗, f∗) is enriched in simplicial sets, we conclude that: T Y ′ T Y+, f∗πY ′∗E′) ∼= M apY ′(f ∗Σ∞ T Y+, πY ′∗E′) ∼= M apY ′(Σ∞ M apY (Σ∞ +, πY ′∗E′) and by definition M apY ′(Σ∞ M apY (Σ∞ M apY (Σ∞ square of simplicial sets: T Y+, πY ∗E′) ∼= M ap(Σ∞ T Y+, j∗πZ ′∗E′) ∼= M ap(Σ∞ T Y ′ +, πY ′∗E′) ∼= M ap(Σ∞ T Y+, E′), M apY (Σ∞ +, E′). Similarly, we conclude that T Z+, E′) and +, E′). Therefore, the following is a homotopy fibre T Y ′ T Y+, i∗πZ∗E′) ∼= M ap(Σ∞ T Z ′ M ap(Σ∞ T Y+, E′) M ap(Σ∞ T Z+, E′) M ap(Σ∞ T Y ′ +, E′) / M ap(Σ∞ T Z ′ +, E′) T Z ′ + → Σ∞ T Y ′ Since Σ∞ + is a cofibration in Spt(Mf t) and E′ is fibrant in Spt(Mf t), we deduce +, E′) is a fibration of simplicial sets [34, 9.1.6 M7]. We that M ap(Σ∞ observe that the functor M ap(−, E′) maps pushout squares in Spt(Mf t) into pullback squares of simplicial sets [34, 9.1.8]; thus by [34, 13.3.8] we conclude that the map: T Y ′ +, E′) → M ap(Σ∞ T Z ′ M ap(Σ∞ T Y+, E′) → M ap(Σ∞ T P (α), E′) induced by P (α) → Y is a weak equivalence of simplicial sets, where P (α) is the pushout of Z ← Z ′ → Y ′ in P rek. Finally, by [34, 4.1.1(2)] we conclude that E′ is fibrant in Spt(Mcdh) since by construction Spt(Mcdh) is the left Bousfield localization of Spt(Mf t) with respect to the maps of the form Σ∞ T (P (α) → Y+) indexed by the abstract blow-up squares in Schk. (cid:3) The following result should be compared with [15, Prop. 3.7]. Theorem 2.14. Let v : X → Spec (k) be in Schk. Given a motivic T -spectrum E ∈ SH, Y ∈ SmX and integers m, n ∈ Z, there is a natural isomorphism HomSHX (Σ∞ T Y+, Σm,nLv∗E) ∼= HomSHcdh(Σ∞ T Y+, Σm,nLπ∗E). / / / / / / / / / THE SLICE SPECTRAL SEQUENCE FOR SINGULAR SCHEMES AND APPLICATIONS 9 Proof. By 2.13, Lv∗(−) ∼= (RπX∗ ◦ Lπ∗)(−) in SHX . Thus, by adjointness: T Y+, Lv∗(Σm,nE)) HomSHX (Σ∞ T Y+, Σm,nLv∗E) ∼= HomSHX (Σ∞ ∼= HomSHcdh(Lπ∗ ∼= HomSHcdh(Lπ∗ X Σ∞ X Σ∞ T Y+, Lπ∗(Σm,nE)) T Y+, Σm,nLπ∗E). Finally, it follows from 2.3 that Σ∞ T Y+ is cofibrant in the levelwise flasque model structure and hence in any of its localizations. In particular, it is cofibrant in the stable model structure of motivic T -spectra. We conclude that Lπ∗ T Y+. The corollary now follows. (cid:3) T Y+ ∼= Σ∞ T Y+ ∼= π∗ X Σ∞ XΣ∞ Remark 2.15. The above result could be called a cdh-descent theorem because it implies cdh- descent for many motivic spectra (see [15, Prop. 3.7]). In particular, it implies cdh-descent for absolute motivic spectra (e.g., KGL and MGL). Recall from [22, §1.2] that an absolute motivic spectrum E is a section of a 2-functor from Schk to triangulated categories such that for any f : X ′ → X in Schk, the canonical map f ∗EX → EX ′ is an isomorphism. Lemma 2.16. Let f : Y → X be a smooth morphism in Schk. Let v : X → Spec (k) be the structure map and u = v ◦ f . Given any E ∈ SH, the map HomSHX (Σ∞ T Y+, Lv∗E) → HomSHY (Σ∞ Proof. The functor Lf ∗ : SHX → SHY admits a left adjoint Lf♯ : SHY → SHX by [5, Prop. 4.5.19] (see also [4, Scholium 1.4.2]). Since f : Y → X is smooth, we have Lf♯(Σ∞ T Y+) = Σ∞ T Y+ by [67, Prop. 3.1.23(1)] and we get T Y+, Lu∗E) is an isomorphism. HomSHX (Σ∞ T Y+, Lv∗E) ∼= HomSHX (Lf♯(Σ∞ ∼= HomSHY (Σ∞ ∼= HomSHY (Σ∞ T Y+), Lv∗E) T Y+, Lf ∗ ◦ Lv∗E) T Y+, Lu∗E) and the lemma follows. A combination of 2.16 and 2.14 yields: (cid:3) Corollary 2.17. Under the same hypotheses and notation of 2.14, assume in addition that X ∈ Smk. Then there are natural isomorphisms: HomSH(Σ∞ T Y+, Σm,nLv∗E) ∼= HomSHcdh(Σ∞ T Y+, Σm,nE) ∼= HomSHX (Σ∞ T Y+, Σm,nLπ∗E). 3. Motivic cohomology of singular schemes We continue to assume that k is a perfect field of exponential characteristic p. In this section, we show that the motivic cohomology of a scheme X ∈ Schk, defined in terms of a cdh-hypercohomology (see 3.1), is representable in the stable homotopy category SHcdh. Recall from [64, Lecture 16] that given T ∈ Schk and an integer r ≥ 0, the presheaf zequi(T, r) on Smk is defined by letting zequi(T, r)(U ) be the free abelian group generated by the closed and irreducible subschemes Z ( U × T which are dominant and equidimensional of relative dimension r (any fiber is either empty or all its components have dimension r) over a component of U . It is known that zequi(T, r) is a sheaf on the big ´etale site of Smk. Let C ∗zequi(T, r) denote the chain complex of presheaves of abelian groups associated via the Dold-Kan correspondence to the simplicial presheaf on Smk given by Cnzequi(T, r)(U ) = zequi(T, r)(U × ∆n k ). The simplicial structure on C∗zequi(T, r) is induced by the cosimplicial scheme ∆• k. Recall the following definition of motivic cohomology of singular schemes from [27, Def. 9.2]. Definition 3.1. The motivic cohomology groups of X ∈ Schk are defined as the hypercoho- mology H m(X, Z(n)) = Hm−2n cdh (X, C ∗zequi(An k , 0)cdh) = A0,2n−m(X, An). 10 AMALENDU KRISHNA AND PABLO PELAEZ We will also need to consider Z[ 1 p ]-coefficients. In this case, we will write: p ](n)) = Hm−2n For n < 0, we set H m(X, Z(n)) = H m(X, Z[ 1 H m(X, Z[ 1 cdh p ](n)) = 0. (X, C ∗zequi(An k , 0)[ 1 p ]). 3.2. The motivic cohomology spectrum. In order to represent the motivic cohomology of a singular scheme X in SHX, let us recall the Eilenberg-MacLane spectrum HZ = (K(0, 0), K(1, 2),··· , K(n, 2n),··· ) in Spt(M), where K(n, 2n) is the presheaf of simplicial abelian groups on Smk associated to the presheaf of chain complexes C ∗zequi(An k , 0) via the Dold-Kan correspondence. The external product of cycles induces product maps K(m, 2m)∧ K(n, 2n) → K(m + n, 2(m + n)). Notice that K(1, 2) ∼= C ∗(zequi(P1 k, 0)) [64, 16.8], so composing the product maps with k, 0)) ∼= K(1, 2) (where the first the canonical map g : T ∼= P1 map assigns to any morphism U → P1 k), we obtain the bonding maps. HZ is a symmetric spectrum whose symmetric structure is obtained by permuting the coordinates in An k . We shall not distinguish between a simplicial abelian group and the associated chain complex of abelian groups from now on in this text and will use them interchangeably. k its graph in U × P1 k → C∗(zequi(P1 k, 0)/zequi(P0 k, 0)/zequi(P0 k/P0 3.3. Motivic cohomology via SHcdh. Let 1 = Σ∞ let 1[ 1 p ] ∈ SH be the homotopy colimit [71, 1.6.4] of the filtering diagram in SH: s ) be the sphere spectrum in SH, and T (S0 1 p / 1 p / 1 p / ··· r→ 1 is the composition of the sum map with the diagonal: 1 where 1 E ∈ SH, we define E[ 1 singularities, and the main result in [45] when k has positive characteristic. p ] ∈ SH to be E ∧ 1[ 1 p ]. This also makes sense in SHX and SHcdh. The following is a reformulation of the main result in [27] when k admits resolution of Σ→ 1. For ∆→ ⊕r i=11 Theorem 3.4 ([17]). Let k be a perfect field of exponential characteristic p, and let v : X → Spec (k) be a separated scheme of finite type. Then for any m, n ∈ Z, there is a natural isomorphism: (3.5) θX : H m(X, Z[ 1 p ](n)) ∼=−→ HomSHX (Σ∞ T X+, Σm,nLv∗HZ[ 1 p ]). Proof. Recall that H m(X, Z[ 1 k , 0) is the motive with compact supports M c(An k [88, §4.1], [64, 16.13]. Combining [88, Cor. 4.1.8] (or [64, 16.7, 16.14]) with [17, 4.2, Prop. 4.3, Thm. 5.1 and Cor. 8.6], we conclude that p ](n)) = A0,2n−m(X, An) (3.1). We observe that C ∗zequi(An k ) of An H m(X, Z[ 1 p ](n)) ∼= HomSHX (Σ2n−m,0(Σ∞ T X+), Σ2n,nLv∗HZ[ which finishes the proof. As a combination of 2.14 and 3.4, we get 1 p ]) (cid:3) Corollary 3.6. Under the hypothesis and with the notation of 3.4, there are natural isomor- phisms H m(X, Z[ 1 p ](n)) ∼= HomSHcdh(Σ∞ ∼= HomSHX (Σ∞ T X+, Σm,nLπ∗HZ[ 1 p ]) T X+, Σm,nLv∗HZ[ 1 p ]). / / / THE SLICE SPECTRAL SEQUENCE FOR SINGULAR SCHEMES AND APPLICATIONS 11 4. Slice spectral sequence for singular schemes Let k be a perfect field of exponential characteristic p. Given X ∈ Schk, recall that TSHeff X eff and is closed Voevodsky's slice filtration of SHX is given as follows. For an integer q ∈ Z, let Σq denote the smallest full triangulated subcategory of SHX which contains C q under arbitrary coproducts, where s ∧ Ss t ∧ Y+) : n, r, s ≥ 0, s − n ≥ q, Y ∈ SmX}. C q eff = {Fn(Sr (4.1) In particular, SHeff under infinite direct sums and contains all spectra of the type Σ∞ slice filtration of SHX (see [86]) is the sequence of full triangulated subcategories X is the smallest full triangulated subcategory of SHX which is closed T Y+ with Y ∈ SmX. The ··· ⊆ Σq+1 T SHeff X ⊆ Σq TSHeff X ⊆ Σq−1 T SHeff X ⊆ ··· It follows from the work of Neeman [70], [71] that the inclusion iq : Σq X → SHX admits a right adjoint rq : SHX → Σq X and that the functors fq, s<q, sq : SHX → SHX are triangulated; where rq ◦ iq is the identity, fq = iq ◦ rq and s<q, sq are characterized by the existence of the following distinguished triangles in SHX: (4.2) TSHeff TSHeff / E / s<qE fqE fq+1E / fqE / sqE for every E ∈ SHX. Definition 4.3. Let a, b, n ∈ Z and Y ∈ SmX . Let F nEa,b(Y ) be the image of the map induced by fnE → E (4.2): HomSHX (Σ∞ T Y+, Σa,bE). This determines a decreasing filtration F • on Ea,b(Y ) = HomSHX (Σ∞ T Y+, Σa,bE), and we will write grnF • for the associated graded F nEa,b(Y )/F n+1Ea,b(Y ). T Y+, Σa,bfnE) → HomSHX (Σ∞ The following result is well known [86, §2]. Proposition 4.4. The filtration F • on Ea,b(Y ) is exhaustive in the sense of [13, Def. 2.1]. Proof. Recall that SHX is a compactly generated triangulated category in the sense of Neeman [70, Def. 1.7], with set of compact generators [5, Thm. 4.5.67]: ∪q∈ZC q eff (4.1). Therefore a map f : E1 → E2 in SHX is an isomorphism if and only if for every Y ∈ SmX and every m, n ∈ Z the induced map of abelian groups HomSHX (Σm,nY+, E1) → HomSHX (Σm,nY+, E2) is an isomorphism. Thus, we conclude that E ∼= hocolimfqE in SHX. Therefore, we deduce that for every a, b ∈ Z and every Y ∈ SmX ; there exist the following isomorphisms [70, Lem. 2.8], [39, Thm. 6.8]: colim n→−∞ F nEa,b(Y ) ∼= colim n→−∞ HomSHX (Σ∞ T Y+, Σa,bfnE) ∼= HomSHX (Σ∞ T Y+, Σa,bhocolimfqE) ∼= Ea,b(Y ), so the filtration F • is exhaustive. (cid:3) 4.5. The slice spectral sequence. Let Y ∈ SmX be a smooth X-scheme and G ∈ SHX. Since SHX is a triangulated category, the collection of distinguished triangles {fq+1G → fqG → sqG}q∈Z determines a (slice) spectral sequence: Ep,q 1 = HomSHX (Σ∞ T Y+, Σp+q s spG) with G∗,∗(Y ) as its abutment and differentials dr : Ep,q r → Ep+r,q−r+1 r . / / / / 12 AMALENDU KRISHNA AND PABLO PELAEZ In order to study the convergence of this spectral sequence, recall from [86, p. 22] that G ∈ SHX is called bounded with respect to the slice filtration if for every m, n ∈ Z and every Y ∈ SmX, there exists q ∈ Z such that (4.6) HomSHX (Σm,nΣ∞ T Y+, fq+iG) = 0 for every i > 0. Clearly the slice spectral sequence is strongly convergent when G is bounded. Proposition 4.7. Let k be a field with resolution of singularities. Let F ∈ SH be bounded with respect to the slice filtration and let G = Lv∗F ∈ SHX, where v : X → Spec k. Then G is bounded with respect to the slice filtration. Proof. Since the base field k admits resolution of singularities, we deduce by [74, Thm. 3.7] that fqG ∼= Lv∗fqF in SHX for every q ∈ Z. It follows from 2.14 that for every m, n ∈ Z; and every Y ∈ SmX, we have HomSHX (Σm,nΣ∞ T Y+, fq+iG) ∼= HomSHcdh(Σm,nΣ∞ T Y+, Lπ∗(fq+iF )) for every i > 0. If X ∈ Smk, then Y ∈ Smk and we have HomSHcdh(Σm,nΣ∞ T Y+, Lπ∗(fq+iF )) ∼= HomSH(Σm,nΣ∞ T Y+, fq+iF ) for every i > 0 by 2.17. Since F is bounded with respect to the slice filtration, we deduce from (4.6) that G is also bounded in SHX in this case. n ∈ Z and every Y ′ ∈ Schk with dim(Y ′) < dim(Y ); there exists q ∈ Z such that Finally, we proceed by induction on the dimension of Y , and assume that for every m, HomSHcdh(Σm,nΣ∞ T Y ′ +, Lπ∗(fq+iF )) = 0 for every i > 0. Since the base field k admits resolution of singularities, there exists a cdh- cover {X ′ ∐ Z → Y } of Y such that X ′ ∈ Smk, dim(Z) < dim(Y ) and dim(W ) < dim(Y ), where we set W = X ′ ×Y Z. Let q1 (resp. q2, q3) be the integers such that the vanishing condition (4.6) holds for (X ′, m, n) (resp. (Z, m, n), (W, m + 1, n)). Let q be the maximum of q1, q2 and q3. Then by cdh-excision, for every i > 0, the following diagram is exact: HomSHcdh(Σm+1,nΣ∞ HomSHcdh(Σm,nΣ∞ T W, Lπ∗(fq+iF )) → HomSHcdh(Σm,nΣ∞ T X ′ +, Lπ∗(fq+iF )) ⊕ HomSHcdh(Σm,nΣ∞ T Y+, Lπ∗(fq+iF )) → T Z+, Lπ∗(fq+iF )). By the choice of q, both ends in the diagram vanish, hence the group in the middle also (cid:3) vanishes as we wanted. In order to get convergence results in positive characteristic, we need to restrict to spectra E ∈ SH which admit a structure of traces [45, 4.2.27 and 4.3.1]. Lemma 4.8. With the notation of 2.11, let X ∈ Schk. Then: p ]) ∼= (Lπ∗E)[ 1 (1) For every E ∈ SH, Lπ∗(E[ 1 (2) For every E ∈ SHcdh and every a, b ∈ Z: HomSHcdh(Σa,bΣ∞ T (X+), E[ 1 p ] and Lv∗(E[ 1 p ]) ∼= (Lv∗E)[ 1 p ]. p ]) ∼= HomSHcdh(Σa,bΣ∞ T (X+), E) ⊗ Z[ 1 p ]. Proof. (1): It follows from the definition of homotopy colimit [71, 1.6.4] that Lπ∗ and Lv∗ commute with homotopy colimits since they are left adjoint. This implies the result since E[ 1 p ] is given in terms of homotopy colimits. (2): Since Σa,bΣ∞ T (X+) is compact in SHcdh [5, Thm. 4.5.67], the result follows from [70, (cid:3) Lem. 2.8]. Lemma 4.9. Let X ∈ Schk and E ∈ SHX . Then for every r ∈ Z, fr(E[ 1 sr(E[ 1 p ]) ∼= (srE)[ 1 p ]. p ]) ∼= (frE)[ 1 p ] and THE SLICE SPECTRAL SEQUENCE FOR SINGULAR SCHEMES AND APPLICATIONS 13 TSHeff Proof. Since the effective categories Σq that the functors fr, sr commute with homotopy colimits. Proposition 4.10. Let F ∈ SH and G = Lv∗F ∈ SHX, where v : X → Spec k. Assume that for every r ∈ Z, sr(F [ 1 p ]) has a weak structure of smooth traces in the sense of [45, 4.2.27]; and that F [ 1 p ] is bounded with respect to the slice filtration, then G[ 1 X are closed under infinite direct sums, we conclude (cid:3) p ] has a structure of traces in the sense of [45, 4.3.1]. If F [ 1 p ] is bounded as well. Proof. Since the base field k is perfect and F [ 1 and 4.9, we conclude that fqG[ 1 p ] ∼= Lv∗fqF [ 1 p ] in SHX for every q ∈ Z. p ] is clearly Z[ 1 p ]-local, combining [45, 4.2.29] It follows from 2.14 that for every m, n ∈ Z; and every Y ∈ SmX, we have HomSHX (Σm,nΣ∞ T (Y+), fq+iG[ 1 p ]) ∼= HomSHcdh(Σm,nΣ∞ T (Y+), Lπ∗(fq+iF [ 1 p ]) for every i > 0. If X ∈ Smk, then Y ∈ Smk and we have HomSHcdh(Σm,nΣ∞ T (Y+), Lπ∗(fq+iF [ 1 p ])) ∼= HomSH(Σm,nΣ∞ T (Y+), fq+iF [ 1 p ]) for every i > 0 by 2.17. Since F [ 1 from (4.6) that G[ 1 p ] is bounded with respect to the slice filtration, we deduce p ] is also bounded with respect to the slice filtration in SHX in this case. Finally, we proceed by induction on the dimension of Y , and assume that for every m, n ∈ Z and every Z ∈ Schk with dimk(Z) < dimk(Y ); there exists q ∈ Z such that HomSHcdh(Σm,nΣ∞ T (Z+), Lπ∗(fq+iF [ 1 p ])) = 0 for every i > 0. Since k is perfect, by a theorem of Gabber [38, Introduction Thm. 3(1)] and Temkin's strengthening [82, Thm. 1.2.9] of Gabber's result, there exists W ∈ Smk and a surjective proper map h : W → Y , which is generically ´etale of degree pr, r ≥ 1. In particular h is generically flat, thus by a theorem of Raynaud-Gruson [32, Thm. 5.2.2], there exists a blow-up g : Y ′ → Y with center Z such that the following diagram commutes, where h′ is finite flat surjective of degree pr and g′ : W ′ → W is the blow-up of W with center h−1(Z): (4.11) W ′ h′ / Y ′ g′ W g / Y. h Thus we have a cdh-cover {Y ′∐Z → Y } of Y , such that dimk(Z) < dimk(Y ) and dimk(E) < dimk(Y ), where we set E = Y ′ ×Y Z. Let q1 (resp. q2, q3) be the integers such that the vanishing condition (4.6) holds for (W, m, n) (resp. (Z, m, n), (E, m + 1, n)). Let q be the maximum of q1, q2 and q3. Then by cdh-excision, for every i > 0, the following diagram is exact: HomSHcdh(Σm+1,nΣ∞ T (E+), Lπ∗(fq+iF [ 1 HomSHcdh(Σm,nΣ∞ T (Y ′ +), Lπ∗(fq+iF [ 1 p ])) → HomSHcdh(Σm,nΣ∞ p ])) ⊕ HomSHcdh(Σm,nΣ∞ T (Y+), Lπ∗(fq+iF [ 1 T (Z+), Lπ∗(fq+iF [ 1 p ])) → p ])). By the choice of q, this reduces to the following exact diagram: g∗ 0 / HomSHcdh(Σm,nΣ∞ T (Y+), Lπ∗(fq+iF [ 1 +), Lπ∗(fq+iF [ 1 So it suffices to show that g∗ = 0. In order to prove this, we observe that the diagram 4.11 commutes, therefore by the choice of q: HomSHcdh(Σm,nΣ∞ p ])) = 0; and we conclude that h′∗ ◦ g∗ = g′∗ ◦ h∗ = 0. Thus, it is enough to see that p ])) → HomSHcdh(Σm,nΣ∞ h′∗ : HomSHcdh(Σm,nΣ∞ T (W+), Lπ∗(fq+iF [ 1 / HomSHcdh(Σm,nΣ∞ +), Lπ∗(fq+iF [ 1 +), Lπ∗(fq+iF [ 1 T (W ′ T (Y ′ T (Y ′ p ])) p ])) p ])). / / / / 14 AMALENDU KRISHNA AND PABLO PELAEZ is injective. Let v′ : Y ′ → Spec k, and ǫ : Lv′∗(fq+iF [ 1 map given by the unit of the adjunction (Lh′∗, Rh′ in 2.13 we deduce that h′∗ gets identified with the map induced by ǫ: p ]) → Rh′ p ]) be the ∗). By the naturality of the isomorphism ∗Lh′∗Lv′∗(fq+iF [ 1 ǫ∗ : HomSHY ′ (Σm,nΣ∞ T Y ′ +, Lv′∗fq+iF [ 1 p ]) → HomSHY ′ (Σm,nΣ∞ T Y ′ +, Rh′ ∗Lh′∗Lv′∗fq+iF [ 1 p ]) p ] has a structure of traces and sr(F [ 1 Since F [ 1 p ]) has a weak structure of smooth traces for every r ∈ Z, it follows from [45, 4.3.7] that fq+i(F [ 1 p ]) has a structure of traces in the sense of [45, 4.3.1]. Thus, we deduce from [45, 4.3.1(Deg) p. 101] that ǫ∗ is injective since h′ is finite flat surjective of degree pr. This finishes the proof. (cid:3) If we only assume that the slices srE have a structure of traces, then we get the weaker conditions 4.15. Corollary 4.12. Let F ∈ SH and G = Lv∗F ∈ SHX, where v : X → Spec k is the structure map. Assume that the following hold. (1) For every r ∈ Z, sr(F [ 1 (2) F [ 1 p ] is bounded with respect to the slice filtration. p ]) has a structure of traces in the sense of [45, 4.3.1]. Then for every m, n in Z, there exists q ∈ Z such that HomSHX (Σm,nΣ∞ for every i > 0 (see 4.6). Proof. Since sr(F [ 1 p ]) has a weak structure of smooth traces [45, 4.2.27]. Thus combining 4.8, 4.9 and [45, 4.2.29] we conclude that for every r ∈ Z: srG[ 1 p ]) has a structure of traces [45, 4.3.1], we observe that in particular sr(F [ 1 p ] ∼= Lv∗frF [ 1 p ]. p ] ∼= Lv∗srF [ 1 p ] and frG[ 1 T X+, sq+iG[ 1 p ]) = 0 If X ∈ Smk, we have HomSHX (Σm,nΣ∞ T X+, Lv∗(sq+iF [ 1 p ])) ∼= HomSH(Σm,nΣ∞ T X+, sq+iF [ 1 p ]) for every i > 0 by 2.17. Since F [ 1 p ] is bounded with respect to the slice filtration, there exist q1, q2 ∈ Z such that the vanishing condition (4.6) holds for (X, m, n), (X, m − 1, n) respec- tively. Let q be the maximum of q1 and q2, then using the distinguished triangle fq+iF [ 1 p ] → sq+iF [ 1 sfq+i+1F [ 1 p ]) = 0 for every i > 0 as we wanted. p ] in SH we conclude that HomSH(Σm,nΣ∞ T (X+), sq+iF [ 1 p ] → Σ1 fq+iF [ 1 p ] with sq+iF [ 1 When X ∈ Schk, the argument in the proof of 4.10 works mutatis-mutandis replacing p ] has a structure of traces [45, 4.3.1]. (cid:3) Corollary 4.13. Assume that the conditions (1) and (2) of 4.12 hold. Then for every m, n in Z, there exists q ∈ Z such that the map fq+i+1G[ 1 p ]; since for every j ∈ Z, sjF [ 1 p ] induces an isomorphism p ] → fq+iG[ 1 HomSHX (Σm,nΣ∞ T X+, fq+i+1G[ 1 p ]) ∼= HomSHX (Σm,nΣ∞ T X+, fq+iG[ 1 p ]) for every i > 0. Proof. Let q1, q2 ∈ Z be the integers corresponding to (m, n), (m + 1, n) in 4.12, respectively. Let q be the maximum of q1 and q2. Then the result follows by combining the vanishing in 4.12 with the distinguished triangle Σ−1 p ] in SHX . (cid:3) Remark 4.14. Combining 4.3 and 4.13, we deduce that for every a, b ∈ Z, there exists m ∈ Z such that F nG[ 1 p ] → fq+i+1G[ 1 p ] → fq+iG[ 1 p ] → sq+iG[ 1 s sq+i[ 1 p ]a,b(X) = F mG[ 1 p ]a,b(X) for every n ≥ m. Proposition 4.15. Assume that the conditions (1) and (2) of 4.12 hold. Then for every n ∈ Z, the slice spectral sequence (4.5): 1 (X, n) = HomSHX (Σ∞ T X+, Σa+b+n,nsaG[ 1 p ]a+b+n,n(X) Ea,b p ]) ⇒ G[ 1 THE SLICE SPECTRAL SEQUENCE FOR SINGULAR SCHEMES AND APPLICATIONS 15 satisfies the following. (1) For every a, b ∈ Z, there exists N > 0 such that Ea,b ∞ for r ≥ N , where ∞ is the associated graded graF • with respect to the descending filtration F • on p ]a+b+n,n(X), see 4.3. r = Ea,b Ea,b G[ 1 (2) For every m, n ∈ Z, the descending filtration F • on G[ 1 and complete [13, Def. 2.1]. p ]m,n(X) (4.3) is exhaustive Proof. (1): It suffices to show that for every a, b ∈ Z only finitely many of the differentials dr : Ea,b p ]m,n(X) is exhaustive. Finally, the completeness of F • (2): By 4.4 the filtration F • on G[ 1 are nonzero. But this follows from 4.12. r → Ea+r,b−r+1 r follows by combining 4.14 with [13, Prop. 1.8 and Prop. 2.2(c)]. (cid:3) 4.16. The slice spectral sequence for MGL(X). Our aim here is to apply the results of the previous sections to obtain a Hopkins-Morel type spectral sequence for MGL∗,∗(X) when X is a singular scheme. For smooth schemes, the Hopkins-Morel spectral sequence has been studied in [58], [37], and over Dedekind domains in [78]. Recall from [84, § 6.3] that for any noetherian scheme S of finite Krull dimension, the scheme GrS(N, n) parametrizes n-dimensional linear subspaces of AN S and one writes BGLS,n = colimN GrS(N, n). There is a universal rank n bundle US,n → BGLS,n and one denotes the Thom space T h(US,n) of this bundle by MGLS,n. Using the fact that the Thom space of a direct sum is the smash product of the corresponding Thom spaces and T = T h(OS ), one gets a T -spectrum MGLS = (MGLS,0, MGLS,1,··· ) ∈ Spt(MS). There is a structure of symmetric spectrum on MGLS for which we refer to [72, §2.1]. We now let k be a field of characteristic zero and let X ∈ Schk. We shall use MGL as a short hand for MGLk throughout this text. It follows from the above definition of MGLX (which shows that MGLX is constructed from presheaves represented by smooth schemes) and 2.12 that the canonical map Lv∗(MGL) → MGLX is an isomorphism. Definition 4.17. We define MGL∗,∗(X) to be the generalized cohomology groups MGLp,q(X) := HomSHX (Σ∞ ∼= HomSHX (Σ∞ T X+, Σp,q MGLX ) T X+, Σp,qLv∗ MGL). It follows from 2.14 that: (4.18) MGLp,q(X) ∼= HomSHcdh(Σ∞ T X+, Σp,qLπ∗ MGL). We now construct the spectral sequence for MGL∗,∗(X) using the exact couple technique as follows. For p, q, n ∈ Z, define Ap,q(X, n) := [Σ∞ Ep,q(X, n) := [Σ∞ s T X+, Σp+q−n T X+, Σp+q−n Σn t (fp MGLX )]. Σn t sp MGLX ]. s It follows from (4.2) that there is an exact Here, [−,−] denotes the morphisms in SHX. sequence (4.19) Ap+1,q−1(X, n) ap,q bp,q n−−→ Ap,q(X, n) n−−→ Ep,q(X, n) n−−→ Ap+1,q(X, n). cp,q n : E1 n := ⊕cp,q n → E1 n . This gives an exact couple {D1 n := n = n shows that (E1, d1) is a complex. Thus, by repeatedly taking the homology Set D1(X, n) := ⊕p,qAp,q(X, n) and E1(X, n) := ⊕p,qEp,q(X, n). Write a1 ⊕bp,q n and c1 n ◦ c1 b1 functors, we obtain a spectral sequence. For the target of the spectral sequence, let Am(X, n) := colimq→∞ Am−q,q(X, n). Since X is a compact object of SHX (see [84, Prop. 5.5], [5, Thm. 4.5.67]), the colimit enters into n , b1 n} and the map d1 n := ⊕ap,q n, E1 n, a1 n, b1 n, c1 16 AMALENDU KRISHNA AND PABLO PELAEZ [−,−] so that Am(X, n) = [Σ∞ T X+, Σm−n couples then yields us a spectral sequence s t MGLX ] = MGLm,n Σn X (X). The formalism of exact (4.20) We now have (4.21) Ep,q 1 (X, n) = Ep,q 1 ⇒ MGLm,n X (X). Ep,q s 1 (X, n) = [Σ∞ ∼=1 [Σ∞ ∼=2 [Σ∞ ∼=3 [Σ∞ ∼= [Σ∞ T X+, Σp+q−n T X+, Σp+q−n T X+, Σp+q−n T X+, Σp+q−n T X+, Σp+q−n s s s s Σn t sp MGLX] t spLv∗ MGL] Σn t Lv∗(sp MGL)] Σn t Lv∗(Σp Σn t Σp Σn T H(L−p))] T Lv∗(H(L−p))]. s 1 Σn Ep+1,q ∼=−→ Σp Since L is a torsion-free abelian group, it follows from 3.6 that the last term of (4.21) is In this sequence of isomorphisms, ∼=1 is shown above, ∼=2 follows from [74, Thm. 3.7] and ∼=3 follows from the isomorphism sp MGL T H(L−p), as shown, for example, in [37, 8.6, p. 46], where L = ⊕i≤0Li ∼= ⊕i≥0M U2i is the Lazard ring. same as H 3p+q(X, Z(n + p)) ⊗Z L−p. The spectral sequence (4.20) is actually identical to an E2-spectral sequence after rein- dexing. Indeed, letting eEp′,q′ 2 = H p′−q′ and using (4.21), an elementary calculation shows that the invertible transformation (3p + q, n + p) 7→ (p′ − q′, n − q′) yields (4.22) (X, Z(n − q′)) ⊗Z Lq′ It is clear from (4.19) that the E1-differential of the above spectral sequence is dp.q 1 and (4.22) shows that this differential is identified with the differential dp′,q′ t sp+1 MGLX ] T X+, Σp+q+1−n ∼= [Σ∞ ∼= H (p′+2)−(q′−1)(X, Z(n − (q′ − 1)) ⊗Z Lq′−1 = eEp′+2,q′−1 . : Ep,q 1 → 2 = dp,q : dr−→ Ep+r,q−r+1 } is }. Combining this with (4.18), we conclude the following. dr−→ eEp′+r+1,q′−r . Inductively, it follows that the chain complex {Ep,q eEp′,q′ 2 → eEp′+2,q′−1 transformed to the chain complex {eEp′,q′ Theorem 4.23. Let k be a field which has characteristic zero and let X ∈ Schk. Then for any integer n ∈ Z, there is a strongly convergent spectral sequence: (4.24) The differentials of this spectral sequence are given by dr : Ep,q ; and for every p, r q ∈ Z, there exists N > 0 such that Ep,q ∞ is the associated graded gr−qF • with respect to the descending filtration on MGLp+q,n(X) (see 4.3). Furthermore, this spectral sequence degenerates with rational coefficients. Ep,q 2 = H p−q(X, Z(n − q)) ⊗Z Lq ⇒ MGLp+q,n(X). ∞ for r ≥ N , where Ep,q r → Ep+r,q−r+1 r = Ep,q Ep+1,q r+1 r+1 1 1 2 2 r r Proof. The construction of the spectral sequence is shown above. Since MGL is bounded by [37, Thm. 8.12], it follows from 4.7 that the spectral sequence (4.24) is strongly convergent. Thus, we deduce the existence of N > 0 such that Ep,q For the degeneration with rational coefficients, we note that the maps fp MGL → sp MGL ∼= Σp T H(L−p) rationally split to yield an isomorphism of spectra MGLQ Q ) in SH [68, Thm. 10.5 and Cor. 10.6(i)]. The desired degeneration of the spectral sequence now follows immediately from its construction above. ∼=−→ ⊕p≥0Σp T H(L−p ∞ for r ≥ N . r = Ep,q (cid:3) Remark 4.25. If k is a perfect field of positive characteristic p, we observe that sr(MGL[ 1 Σr smooth traces [45, 5.2.4]. Thus, we can apply [45, 4.2.29] to conclude that Lv∗sr(MGL[ 1 p ] for every r ∈ Z [37, 8.6, p. 46], so sr(MGL[ 1 p ]) ∼= p ]) has a weak structure of p ]) ∼= T H(L−r)[ 1 THE SLICE SPECTRAL SEQUENCE FOR SINGULAR SCHEMES AND APPLICATIONS 17 sr(Lv∗ MGL[ 1 on the characteristic of k. We thus obtain a spectral sequence as in (4.24): p ]). Except for this identification, the proof of Theorem 4.23 does not depend Ea,b 2 = H a−b(X, Z(n − b)) ⊗Z Lb[ 1 p ] ⇒ MGLa+b,n(X)[ 1 p ]. But we can only guarantee strong convergence when X ∈ Smk [37, Thm. 8.12]. In general, for X ∈ Schk, the spectral sequence satisfies the weaker convergence of 4.15(1)-(2). In this case, the strong convergence will follow if one knew that MGL has a structure of traces. 4.26. The slice spectral sequence for KGL. For any noetherian scheme X of finite Krull dimension, the motivic T -spectrum KGLX ∈ Spt(MX) was defined by Voevodsky (see [84, § 6.2]) which has the property that it represents algebraic K-theory of objects in SmX if X is regular. It was later shown by Cisinski [15] that for X not necessarily regular, KGLX represents Weibel's homotopy invariant K-theory KH∗(Y ) for Y ∈ SmX . Like MGLX, there is a structure of symmetric spectrum on KGLX for which we refer to [43, p. 157, p. 176]. Let k be a field of exponential characteristic p. The map Lv∗(KGLk) → KGLX is an It is known that sr KGLk ∼= Σr T HZ, for r ∈ Z (see [61, isomorphism by [15, Prop. 3.8]. Thm. 6.4.2] if k is perfect and [76, § 1, p. 1158] in general). It follows from [74, Thm. 3.7] (in p ]k) ∼= sr(Lv∗ KGL[ 1 p ]k) ∼= positive characteristic we use instead [45, 4.2.29]) that Lv∗(sr KGL[ 1 sr KGL[ 1 T HQ in SH [75, 5.3.17 and 5.3.10]. We can thus use the Bott periodicity of KGLX and repeat the construction of § 4.16 mutatis mutandis (with n = 0) to conclude the following. p ]X . One also knows that (KGLk)Q ∼= ⊕p∈ZΣp Theorem 4.27. Let k be a field that admits resolution of singularities (resp. a field of exponential characteristic p > 1); and let X ∈ Schk. Then there is a strongly convergent spectral sequence (4.28) (4.29) Ea,b 2 = H a−b(X, Z(−b)) ⇒ KH−a−b(X) resp. Ea,b 2 = H a−b(X, Z(−b)) ⊗Z Z[ 1 p ] ⇒ KH−a−b(X) ⊗Z Z[ 1 p ]. r → Ea+r,b−r+1 The differentials of this spectral sequence are given by dr : Ea,b b ∈ Z, there exists N > 0 such that Ea,b gr−bF • with respect to the descending filtration on KH−a−b(X) (resp. KH[ 1 4.3. Furthermore, this spectral sequence degenerates with rational coefficients. ∞ for r ≥ N , where Ea,b ; and for every a, r ∞ is the associated graded p ]−a−b(X)), see r = Ea,b Proof. If k admits resolution of singularities, we just need to show that the spectral sequence is convergent. For this, we observe that KGLk is the spectrum associated to the Landweber exact L-algebra Z[β, β−1] that classifies the multiplicative formal group law [79, Thm. 1.2]. Thus [37, 8.12] implies that KGLk is bounded with respect to the slice filtration (this argument also applies in positive characteristic). Hence, the convergence follows from 4.7. p ]k satisfies the conditions in 4.10. In the case of positive characteristic, the existence of the spectral sequence follows by combining the argument of § 4.16 with 4.8 and 4.9. To establish the convergence, it suffices to check that KGL[ 1 We have already seen that KGLk is bounded with respect to the slice filtration. Thus, by p ]k is bounded with respect to the slice filtration as well. On p ]k has a structure of traces in the sense T HZ for r ∈ Z, combining [45, 5.2.4] and 4.9, we p ])k has a weak structure of smooth traces in the sense of [45, 4.2.27]. (cid:3) 4.8(2) we conclude that KGL[ 1 the other hand, it follows from [45, 5.2.3] that KGL[ 1 of [45, 4.3.1]. Finally, since sr KGLk ∼= Σr deduce that sr(KGL[ 1 This finishes the proof. Remark 4.30. When k has characteristic zero, the spectral sequence of Theorem 4.27 is not new and was constructed by Haesemeyer (see [33, Thm. 7.3]) using a different approach. 18 AMALENDU KRISHNA AND PABLO PELAEZ However, the expected degeneration (rationally) of this spectral sequence and its positive characteristic analogue are new. As a combination of 4.27 and [83, Thm. 9.5, 9.6], we obtain the following spectral sequence for the algebraic K-theory K B(−) of singular schemes [83]. Corollary 4.31. Let k be a field of exponential characteristic p > 1. Let ℓ 6= p be a prime and m ≥ 0 any integer. Given any X ∈ Schk, there exist strongly convergent spectral sequences (4.32) Ea,b 2 = H a−b(X, Z(−b)) ⊗Z Z[ 1 p ] ⇒ K B Ea,b 2 = H a−b(X, Z/ℓm(−b)) ⇒ K B/ℓm −a−b(X) ⊗Z Z[ 1 p ]; −a−b(X). (4.33) 5. Applications I: Comparing cobordism, K-theory and cohomology In this section, we deduce some geometric applications of the slice spectral sequences for singular schemes. More applications will appear in the subsequent sections. Consider the edge map MGL = f0 MGL → s0 MGL ∼= HZ in the spectral sequence (4.24). This induces a natural map νX : MGLi,j(X) → H i(X, Z(j)) for every X ∈ Schk and i, j ∈ Z. The following result shows that there is no distinction between algebraic cycles and cobor- dism cycles at the level of 0-cycles. Theorem 5.1. Let k be a field which admits resolution of singularities (resp. a perfect field of positive characteristic p). Then for any X ∈ Schk of dimension d, we have H 2a−b(X, Z(a)) = 0 (resp. H 2a−b(X, Z(a)) ⊗Z Z[ 1 p ] = 0) whenever a > d + b. In particular, for every X ∈ Schk (resp. X ∈ Smk), the map (5.2) resp. νX : MGL2d+i,d+i(X) ⊗Z Z[ 1 νX : MGL2d+i,d+i(X) → H 2d+i(X, Z(d + i)) p ] → H 2d+i(X, Z(d + i)) ⊗Z Z[ 1 p ] (5.3) is an isomorphism for all i ≥ 0. Proof. Using the spectral sequence (4.24) (resp. 4.25) and the fact that L>0 = 0, the isomor- phism of (5.2) (resp. 5.3) follows immediately from the vanishing assertion for the motivic cohomology. To prove the vanishing result, we note that for X ∈ Smk, there is an isomorphism H 2a−b(X, Z(a)) ∼= CHa(X, b) by [85], and the latter group is clearly zero if a > d + b by definition of Bloch's higher Chow groups. If X is not smooth and k admits resolution of singularities, our assumption on k implies that there exists a cdh-cover {X ′∐Z → X} of X, such that X ′ ∈ Smk, dim(Z) < dim(X) and dim(W ) < dim(X), where we set W = X ′ ×X Z. The cdh-descent for the motivic cohomology yields an exact sequence H 2a−b−1(W, Z(a)) ∂−→ H 2a−b(X, Z(a)) → H 2a−b(X ′, Z(a)) ⊕ H 2a−b(Z, Z(a)). The smooth case of our vanishing result shown above and an induction on the dimension together imply that the two end terms of this exact sequence vanish. Hence, the middle term vanishes too. If X is not smooth and k is perfect of positive characteristic, we argue as in 4.10. Namely, by a theorem of Gabber [38, Introduction Thm. 3(1)] and Temkin's strengthening [82, Thm. 1.2.9] of Gabber's result, there exists W ∈ Smk and a surjective proper map h : W → X, which is generically ´etale of degree pr, r ≥ 1. Then by a theorem of Raynaud-Gruson [32, Thm. 5.2.2], there exists a blow-up g : X ′ → X with center Z such that the following diagram THE SLICE SPECTRAL SEQUENCE FOR SINGULAR SCHEMES AND APPLICATIONS 19 commutes, where h′ is finite flat surjective of degree pr and g′ : W ′ → W is the blow-up of W with center h−1(Z): (5.4) W ′ h′ / X ′ g′ W g / X. h Thus we have a cdh-cover {X ′ ∐ Z → X} of X, such that dimk(Z) < dimk(X) and dimk(E) < dimk(X), where we set E = X ′ ×X Z. Then by cdh-excision, the following diagram is exact: H 2a−b−1(E, Z(a)) ⊗Z Z[ 1 H 2a−b(X ′, Z(a)) ⊗Z Z[ 1 p ] → H 2a−b(X, Z(a)) ⊗Z Z[ 1 p ] → p ] ⊕ H 2a−b(Z, Z(a)) ⊗Z Z[ 1 p ]. By induction on the dimension, this reduces to the following exact sequence: 0 / H 2a−b(X, Z(a)) ⊗Z Z[ 1 p ] g∗ / H 2a−b(X ′, Z(a)) ⊗Z Z[ 1 p ] . So it suffices to show that g∗ = 0. In order to prove this, we observe that (5.4) commutes, therefore since W ∈ Smk: H 2a−b(W, Z(a)) ⊗Z Z[ 1 p ] = 0; and we conclude that h′∗ ◦ g∗ = g′∗ ◦ h∗ = 0. Thus, it is enough to see that h′∗ : H 2a−b(X ′, Z(a)) ⊗Z Z[ 1 p ] → H 2a−b(W ′, Z(a)) ⊗Z Z[ 1 p ] ∗Lh′∗Lv′∗HZ[ 1 p ] → Rh′ p ] be the map given ∗). By the naturality of the isomorphism in 2.13, we is injective. Let v′ : X ′ → Spec k, and ǫ : Lv′∗HZ[ 1 by the unit of the adjunction (Lh′∗, Rh′ deduce that h′∗ gets identified with the map induced by ǫ (see 3.6): T (X ′ ǫ∗ : HomSHX ′ (Σm,nΣ∞ By [45, 5.2.4], HZ[ 1 p ] has a structure of traces in the sense of [45, 4.3.1]. Thus, we deduce from [45, 4.3.1(Deg) p. 101] that ǫ∗ is injective since h′ is finite flat surjective of degree pr. This finishes the proof. (cid:3) p ]) → HomSHX ′ (Σm,nΣ∞ ∗Lh′∗Lv′∗HZ[ 1 +), Lv′∗HZ[ 1 +), Rh′ T (X ′ p ]). Remark 5.5. For X ∈ Smk and i = 0, the isomorphism of (5.2) was proved by D´eglise [21, Cor. 4.3.4]. When A is a field, the following result was proven by Morel using methods of unstable motivic homotopy theory [66, p. 9 Cor. 1.25]. Taking for granted the result for fields, D´eglise [21] proved 5.6 using homotopy modules. Spitzweck proved 5.6 for localizations of a Dedekind domain in [78, Cor. 7.3]. Theorem 5.6. Let k be a perfect field of exponential characteristic p. Then for any regular semi-local ring A which is essentially of finite type over k, and for any integer n ≥ 0, the map (5.7) MGLn,n(A) ⊗Z Z[ 1 p ] → H n(A, Z(n)) ⊗Z Z[ 1 p ]. p ] if k is also infinite. In particular, there is a natural isomorphism MGLn,n(A) ⊗Z Z[ 1 is an isomorphism. n (A) ⊗Z Z[ 1 K M Proof. Using the spectral sequence (4.24) and the fact that L>0 = 0, it suffices to prove that En+i+j,−i (A) = 0 for every j ≥ 0 and i ≥ 1. In positive characteristic, we can use 4.25 since A is regular. Notice that (4.24)-(4.25) are strongly convergent for A by [37, 8.9-8.10]. 2 p ] ∼= / / / / 20 AMALENDU KRISHNA AND PABLO PELAEZ On the one hand, we have isomorphisms En+i+j,−i 2 (A) = H n+2i+j(A, Z(n + i)) ⊗Z Z[ 1 p ] ∼= CHn+i(A, 2n + 2i − n − 2i − j) ⊗Z Z[ 1 p ] = CHn+i(A, n − j) ⊗Z Z[ 1 p ]. On the other hand, letting F denote the fraction field of A, the Gersten resolution for the higher Chow groups (see [11, Thm. 10.1]) shows that the restriction map CHn+i(A, n − j) → CHn+i(F, n − j) is injective. But the term CHn+i(F, n − j) is zero whenever j ≥ 0, i ≥ 1 for dimensional reasons. We conclude that En+i+j,−i (A) = 0. The last assertion of the theorem now follows from the isomorphism CHn(A, n) ∼= K M n (A) by [47, Thm. 1.1]. (cid:3) 2 5.8. Connective K-theory. Let k be a field of exponential characteristic p and let X ∈ Schk. Recall that the connective K-theory spectrum KGL0 X is defined to be the motivic T -spectrum f0 KGLX in SHX (see (4.2)). Strictly speaking, KGL0 X should be called effective K-theory, nevertheless we will follow the terminology of Dai-Levine [20]. In particular, there is a canonical map uX : KGL0 morphisms from objects of SHeff X). Using an analogue of 4.23 for KGL0 HomSHX (Σ∞ tence of the cycle class map for the higher Chow groups as follows. X → KGLX which is universal for X to KGLX. For any Y ∈ SmX, we let CKH p,q(Y ) = X , one can prove the exis- T Y+, Σp,q KGL0 Theorem 5.9. Let k be a field of exponential characteristic p and let X ∈ Schk have dimen- sion d. Then the map KGL0 p ] induces for every integer i ≥ 0, an isomorphism p ] → s0 KGLX [ 1 p ] ∼= HZ[ 1 X [ 1 (5.10) CKH 2d+i,d+i(X) ⊗Z Z[ 1 p ] ∼=−→ H 2d+i(X, Z(d + i)) ⊗Z Z[ 1 p ]. In particular, the canonical map KGL0 X → KGLX induces a natural cycle class map (5.11) cyci : H 2d+i(X, Z(d + i)) ⊗Z Z[ 1 p ] → KHi(X) ⊗Z Z[ 1 p ]. X is a connective T -spectrum and we have Lv∗(KGL0 k) Proof. First we assume that k admits resolution of singularities. It follows from the definition that KGL0 X by [74, Thm. 3.7]. One also knows that sr KGL0 T HZ for r ≥ 0 [61, Thm. 6.4.2] and zero otherwise. The proof of 4.23 can now be repeated verbatim to conclude that for each n ∈ Z, there is a strongly convergent spectral sequence ∼=−→ KGL0 k ∼= Σr (5.12) Ea,b 2 = H a−b(X, Z(n − b)) ⊗Z Zb≤0 ⇒ CKH a+b,n(X), If the characteristic of k is positive, then sr(KGL0 where Zb≤0 = Z if b ≤ 0 and is zero otherwise. Furthermore, this spectral sequence degenerates with rational coefficients. One now repeats the proof of 5.1 to conclude that the edge map CKH 2d+i,d+i(X) → H 2d+i(X, Z(d + i)) is an isomorphism for every i ≥ 0. Finally, we compose the inverse of this isomorphism with canonical map CKH 2d+i,d+i(X) → KHi(X) to get the desired cycle class map. T HZ for every r ≥ 0 and zero k[ 1 otherwise [61, Thm. 6.4.2]. So sr(KGL0 p ]) has a weak structure of traces [45, 5.2.4]. By 4.9, we deduce that sr(KGL0 T HZ[ 1 p ] for every r ≥ 0 and zero otherwise. Thus, we can apply [45, 4.2.29] to conclude that Lv∗(KGL0 p ]. Then the argument of 4.27 applies, and we conclude that for each n ∈ Z, there is a strongly convergent spectral sequence (5.13) p ]) ∼= KGL0 k[ 1 p ]) ∼= Σr k[ 1 k) ∼= Σr X[ 1 Ea,b 2 = H a−b(X, Z(n − b)) ⊗Z Z[ 1 p ]b≤0 ⇒ CKH a+b,n(X) ⊗Z Z[ 1 p ], THE SLICE SPECTRAL SEQUENCE FOR SINGULAR SCHEMES AND APPLICATIONS 21 By 5.1, H 2a−b(X, Z(a)) ⊗Z Z[ 1 sequence (5.13) and the fact that L>0 = 0, we deduce the isomorphism of (5.10) with Z[ 1 coefficients: p ] = 0 whenever a > d + b. Thus, combining the spectral p ]- CKH 2d+i,d+i(X) ⊗Z Z[ 1 p ] ∼=−→ H 2d+i(X, Z(d + i)) ⊗Z Z[ 1 p ]. (cid:3) An argument identical to the proof of 5.6 shows that for any regular semi-local ring A which is essentially of finite type over an infinite field k and any integer n ≥ 0, there is a natural isomorphism (notice that in positive characteristic, the spectral sequence is also strongly convergent integrally since A is regular): (5.14) CKH n,n(A) ∼=−→ K M n (A). Moreover, the canonical map CKH n,n(A) → Kn(A) respects products [73, Thm. 3.6.9], hence it coincides with the known map K M n (A) → Kn(A). This shows that the Milnor K-theory is represented by the connective K-theory and one gets a lifting of the relation between the Milnor and Quillen K-theory of smooth semi-local schemes to the level of SH. In particular, it is possible to recover Milnor K-theory and its map into Quillen K-theory from the T -spectrum KGL (which represents Quillen K-theory in SH for smooth k-schemes) by passing to its (−1)-effective cover f0KGLk → KGLk. As another consequence of the slice spectral sequence, one gets the following comparison result between the connective and non-connective versions of the homotopy K-theory. The homological analogue of this results was shown in [20, Cor. 5.5]. Theorem 5.15. Let k be a field of exponential characteristic p and let X ∈ Schk have dimen- sion d. Then the canonical map CKH 2n,n(X)⊗Z Z[ 1 p ] is an isomorphism for every integer n ≤ 0. Proof. If k admits resolution of singularities, we observe that the slice spectral sequence is functorial for morphisms of motivic T -spectra. Since H 2q(X, Z(q)) = 0 for q < 0, a comparison of the spectral sequences (4.28) and (5.12) shows that it is enough to prove that for every r ≥ 2 and q ≤ 0, either q +r−1 ≤ 0 or H −q−r−(q+r−1)(X, Z(1−r−q)) = H 1−2r−2q(X, Z(1−r−q)) = 0. But this is true because H 1−2r−2q(X, Z(s)) = 0 if s < 0. In positive characteristic, we use the same argument as above for the spectral sequences p ] → KH0(X)⊗Z Z[ 1 (4.29) and (5.13) to deduce that is an isomorphism for every prime ℓ 6= p. Then, the result follows from 4.8(2). CKH 2n,n(X) ⊗Z Z(ℓ) → KH0(X) ⊗Z Z(ℓ) (cid:3) Yet another consequence of the above spectral sequences is the following direct verifica- tion of Weibel's vanishing conjecture for negative KH-theory and negative CKH-theory of singular schemes. For KH-theory, there are other proofs of this conjecture by Haesemeyer [33, Thm. 7.1] in characteristic zero and by Kelly [46, Thm. 3.5] and Kerz-Strunk [48] in positive characteristic using different methods. We refer the reader to [15], [18], [29], [49], [51] and [90] for more results associated to Weibel's conjecture. The vanishing result below for CKH-theory is new in any characteristic. Theorem 5.16. Let k be a field of exponential characteristic p and let X ∈ Schk have dimension d. Then CKH m,n(X) ⊗Z Z[ 1 p ] = 0 whenever 2n− m < −d and KH−d(X) ⊗Z Z[ 1 Proof. When k admits resolution of singularities, using the spectral sequences (4.28) and (5.10), it suffices to show that H p−q(X, Z(n − q)) = 0 whenever 2n − p − q + d < 0. p ] = KH2n−m(X)⊗Z Z[ 1 cdh(X, Z) ⊗Z Z[ 1 p ]. p ] ∼= H d 22 AMALENDU KRISHNA AND PABLO PELAEZ If n − q < 0, then we already know that this motivic cohomology group is zero. So we can assume n− q ≥ 0. We set a = n− q, b = 2n− p− q so that 2a− b = 2n− 2q− 2n + p + q = p− q. Since 2n − p − q + d < 0 and n − q ≥ 0 by our assumption, we get b + d − a = 2n − p − q + d − n + q = n − p + d = (2n − p − q + d) − (n − q) < 0. The theorem now follows because we have shown in the proof of 5.1 that H p−q(X, Z(n − q)) = H 2a−b(X, Z(a)) = 0 as a > b + d. This argument also shows that KH−d(X) ∼= H d(X, Z(0)) ∼= H d In positive characteristic, we use the same argument with the spectral sequences (4.29) and (5.13) to deduce that CKH m,n(X)⊗Z Z[ 1 p ] = 0 whenever 2n−m < −d and KH−d(X) ⊗Z Z[ 1 p ] = KH2n−m(X)⊗Z Z[ 1 cdh(X, Z) ⊗Z Z[ 1 p ]. p ] ∼= H d cdh(X, Z). (cid:3) Weibel's conjecture on the vanishing of certain negative K-theory was proven (after invert- ing the characteristic) by Kelly [46]. Using our spectral sequence (which uses the methods of [45]), we can obtain the following result (which follows as well from [46] via the cdh-descent spectral sequence). The characteristic zero version of this computation was proven in [19, Thm. 02(1)], and for arbitrary noetherian schemes, we refer the reader to [49, Cor. D]. Corollary 5.17. Let k be a field of exponential characteristic p and let X ∈ Schk have dimension d. Then K B −d(X) ⊗Z Z[ 1 p ] ∼= H d cdh(X, Z) ⊗Z Z[ 1 p ]. 6. The Chern classes on KH-theory In order to obtain more applications of the slice spectral sequence for KH-theory and the cycle class map (see 5.9), we need to have a theory of Chern classes on the KH-theory of singular schemes. Gillet [31] showed that any cohomology theory satisfying the projective bundle formula and some other standard admissibility axioms admits a theory of Chern classes from algebraic K- theory of schemes over a field. These Chern classes are very powerful tools for understanding algebraic K-theory groups in terms of various cohomology theories such as motivic cohomol- ogy and Hodge theory. The Chern classes in Deligne cohomology are used to define various regulator maps on K-theory and they also give rise to the construction of intermediate Jaco- bians of smooth projective varieties over C. If k is a perfect field of exponential characteristic p ≥ 1, then Kelly [45, Th. 5.5.10] has shown that the motivic cohomology functor X 7→ {H i(X, Z(j))[ 1 p ]}i,j∈Z satisfies the projective bundle formula in Schk. This implies in particular by Gillet's theory that there are functorial Chern class maps (6.1) ci,j : Kj(X) → H 2i−j(X, Z(i))[ 1 p ]. In this section, we show that in characteristic zero, Gillet's technique can be used to con- struct the above Chern classes on the homotopy invariant K-theory of singular schemes. Applications of these Chern classes to the understanding of the motivic cohomology and KH-theory of singular schemes will be given in the following two sections. Let k be a field of characteristic zero and let SchZar/k denote the category of separated schemes of finite type over k equipped with the Zariski topology. Let SmZar/k denote the full subcategory of smooth schemes over k equipped with the Zariski topology. For any X ∈ Schk, let XZar denote small Zariski site of X. A presheaf of spectra on Schk or Smk will mean a presheaf of S1-spectra. Let P re(SchZar/k) be the category of presheaves of simplicial sets on SchZar/k equipped with the injective Zariski local model structure, i.e., the weak equivalences are the maps that induce a weak equivalence of simplicial sets at every Zariski stalk and the cofibrations are THE SLICE SPECTRAL SEQUENCE FOR SINGULAR SCHEMES AND APPLICATIONS 23 given by monomorphisms. This model structure restricts to a similar model structure on the category P re(XZar) of presheaves of simplicial sets on XZar for every X ∈ Schk. We will write Hbig Zar(k), Hsml Zar(X) for the homotopy categories of P re(SchZar/k) and P re(XZar), respectively. 6.2. Chern classes from KH-theory to motivic cohomology. For any X ∈ Schk, let ΩBQP (X) denote the simplicial set obtained by taking the loop space of the nerve of the category QP (X) obtained by applying Quillen's Q-construction to the exact category of locally free sheaves on XZar. Let K denote the presheaf of simplicial sets on SchZar/k given by X 7→ ΩBQP (X). One knows that K is a presheaf of infinite loop spaces so that there is a presheaf of spectra eK on Schk such that K = (eK)0. Let eKB denote the Thomason-Trobaugh presheaf of spectra on Schk such that eKB(X) = K B(X) for every X ∈ Schk. There is a natural map of presheaves of spectra eK → eKB which induces isomorphism between the Recall from [40, Thm. 2.34] that the category of presheaves of spectra on SchZar/k has a closed model structure where the weak equivalences are given by the stalk-wise stable equivalence of spectra and a map f : E → F is a cofibration if f0 is a monomorphism and En+1 ∐S1∧En S1 ∧ Fn → Fn+1 is a monomorphism for each n ≥ 0. Let Hs Zar(k) denote the : Hbig Zar(k) → Hs associated homotopy category. There is a functor Σ∞ Zar(k) which has a s right adjoint. We can consider the above model structure and the corresponding homotopy categories with respect to the Nisnevich and cdh-sites as well. non-negative homotopy group presheaves. Let eKcdh → eKB cdh denote the map between the functorial fibrant replacements in the above model structure on presheaves of spectra on Schk with respect to the cdh-topology. Let KH denote the presheaf of spectra on Schk such that KH(X) is Weibel's homotopy invariant K-theory of X (see [89]). The following is a direct consequence of the main result of [33]. Lemma 6.3. Let k be a field of characteristic zero. For every X ∈ Schk and integer p ∈ Z, there is a natural isomorphism KHp(X) cdh(X,Kcdh). ∼=−→ H−p πp(eKcdh(X)) = HomHs cdh(k)(Σ∞ ∼= HomHcdh(k)(Sp ∼= H−p cdh(X,Kcdh). s (Sp s ∧ X,K) s ∧ X), eK) Proof. We have a natural isomorphism (6.4) It is well known that the natural maps Kp(X) → πp(eKcdh(X)) → πp(eKB cdh(X)) are isomor- phisms for all p ∈ Z when X is smooth over k. In general, let X ∈ Schk. We can find a Cartesian square (6.5) Z ′ X ′ f Z / X, cdh now show that the map πp(eKcdh(X)) → πp(eKB where X ′ ∈ Smk and f is a proper birational morphism which is an isomorphism outside the closed immersion Z ֒→ X. An induction on dimension of X and cdh-descent for eKcdh as well as eKB cdh(X)) is an isomorphism for all p ∈ Z. Composing the inverse of this isomorphism with the map in (6.4), we get a natural isomorphism πp(eKB On the other hand, it follows from [33, Thm. 6.4] that the natural map KH(X) → eKB is a homotopy equivalence. We conclude that there is a natural isomorphism νX : KHp(X) H−p cdh(X,Kcdh) for every X ∈ Schk and p ∈ Z. cdh(X,Kcdh). cdh(X) ∼=−→ ∼=−→ H−p cdh(X)) (cid:3) / / / 24 AMALENDU KRISHNA AND PABLO PELAEZ Let BGL denote the simplicial presheaf on Schk given by BGL(X) = colimn BGLn(O(X)). ∼=−→ It is known (see [31, Prop. 2.15]) that there is a natural section-wise weak equivalence KX Z × Z∞BGLX in P re(SchZar/k) (see § 6.2), where Z∞(−) is the Z-completion functor of Bousfield-Kan. To simplify the notation, for any integer q ∈ Z, we will write Γ(q) for the presheaf on SchZar/k given by (see § 3): Γ(q)(U ) =(cid:26) C ∗zequi(Aq 0 k, 0)(U )[−2q] if q ≥ 0 if q < 0. It is known that the restriction of Γ(q) on SmZar/k is a sheaf (see, for instance, [64, Def. 16.1]). We let Γ(q)[2q] → K(Γ(q), 2q) denote a functorial fibrant replacement of Γ(q)[2q] with respect to the injective Zariski local model structure. It follows from [3, Ex. 3.1] that K(Γ(q), 2q) is a cohomology theory on SmZar/k satisfying all conditions of [31, Defs. 1.1-1.2]. We conclude from Gillet's construction (see [31, §2, p. 225]) that for any X ∈ SmZar/k, there is a morphism of simplicial presheaves Cq : BGLX → ∼=−→ Z× Z∞BGLX K(Γ(q), 2q)X in Hsml Zar(X) which is natural in X. Composing this with KX and using the isomorphism Z∞K(Γ(q), 2q) ∼= K(Γ(q), 2q), we obtain a map Cq : KX ∼=−→ Z × Z∞BGLX → Z × K(Γ(q), 2q)X → K(Γ(q), 2q)X Zar(X), where the last arrow is the projection. in Hsml Since K(Γ(q), 2q) is fibrant in P re(SchZar/k), it follows from [41, Cor. 5.26] that the restrictionK(Γ(q), 2q)X is fibrant in P re(XZar). Since KX is cofibrant (in our local injective model structure), Gillet's construction yields a map of simplicial presheaves (see [31, p. 225]) Cq : KX → K(Γ(q), 2q)X in P re(XZar). In particular, a map K(X) → K(Γ(q), 2q)(X). Furthermore, the naturality of the construction gives for any morphism f : Y → X in Smk, a diagram that commutes up to homotopy (see, for instance, [3, 5.6.1]) (6.6) Cq Cq K(X) f ∗ K(Y ) K(Γ(q), 2q)(X) f ∗ / K(Γ(q), 2q)(Y ). Equivalently, there is a morphism of simplicial presheaves Cq : K → K(Γ(q), 2q) in Hbig Zar(k) and hence a morphism in (Smk)N is (see § 2.1). Pulling back Cq via the morphism of sites π : (Schk)cdh → (Smk)N is [41, p. 111], and considering the cohomologies of the associated cdh-sheaves, we obtain for any X ∈ Schk, a closed subscheme Z ⊆ X and p, q ≥ 0, the Chern class maps X,p,q : H−p cZ Z,cdh(X,Kcdh) (6.7) Z,cdh(X, Lπ∗(K)) Z,cdh(X, Lπ∗(K(Γ(q), 2q))) Z,cdh(X, C∗zequi(Aq k, 0)cdh) := H−p → H−p = H−p := H 2q−p Z (X, Z(q)). It follows from 6.3 that H−p p (X), where the KH Z (X) is the homotopy fiber of the map KH(X) → KH(X \ Z). Let (X, Z) denote the pair consisting of a scheme X ∈ Schk and a closed subscheme Z ⊆ X. A map of pairs f : (Y, W ) → (X, Z) is a morphism f : Y → X such that f −1(Z) ⊆ W . We have then shown the following. Z,cdh(X,Kcdh) = KH Z / / / THE SLICE SPECTRAL SEQUENCE FOR SINGULAR SCHEMES AND APPLICATIONS 25 Theorem 6.8. Let k be a field of characteristic zero. Then for any pair (X, Z) in Schk and for any p ≥ 0, q ∈ Z, there are Chern class homomorphisms p (X) → H 2q−p cZ X,p,q : KH Z X,0,0 with K0(X) → KH0(X) is the rank map. For any map of (X, Z(q)) Z such that the composition of cX pairs f : (Y, W ) → (X, Z), there is a commutative diagram (6.9) cZ X,p,q KH Z p (X) H 2q−p Z (X, Z(q)) f ∗ f ∗ cW Y,p,q KH W p (Y ) / H 2q−p W (Y, Z(q)). 6.10. Chern classes from KH-theory to Deligne cohomology. Let CZar denote the category of schemes which are separated and of finite type over C with the Zariski topology. We denote by CNis the same category but with the Nisnevich topology. Let Can denote the category of complex analytic spaces with the analytic topology. There is a morphism of sites ǫ : Can → CZar. For any q ∈ Z, let Γ(q) denote the complex of sheaves on CZar defined as follows: (6.11) Γ(q) =(cid:26) ΓD(q) Rǫ∗((2π√−1)Z) if q ≥ 0 if q < 0, where ΓD(q) is the Deligne-Beilinson complex on CZar in the sense of [26]. Then Γ(q) is a cohomology theory on SmC satisfying Gillet's conditions for a theory of Chern classes (see, for instance, [3, Ex. 3.4]). Applying the argument of 6.8 in verbatim, we obtain the Chern class homomorphisms (6.12) cZ X,p,q : KH Z p (X) → H2q−p Z,cdh(X, (ΓD(q))cdh) for a pair of schemes (X, Z) in SchC which is natural in (X, Z). Let us now fix a scheme X ∈ SchC. Recall from [24, § 6.2.5-6.2.8] that a smooth proper hypercovering of X is a smooth simplicial scheme X• with a map of simplicial schemes pX : X• → X such each map Xi → X is proper and pX satisfies the universal cohomological descent in the sense of [24]. The resolution of singularities implies that such a hypercovering exists. The Deligne cohomology of X is defined to be [24, 5.1.11] (6.13) H p D(X, Z(q)) := Hp Zar(X, RpX∗ΓD(q)) = Hp Zar(X•, ΓD(q)). Gillet's theory of Chern classes gives rise to the Chern class homomorphisms D (X, Z(q)) cQ X,p,q : Kp(X) → H 2q−p (6.14) for any X ∈ SchC which is contravariant functorial, where Ki(X) = πi(ΩBQP (X)) is the Quillen K-theory (see, for instance, [6, § 2.4]). Our objective is to show that these Chern classes actually factor through the natural map K∗(X) → KH∗(X). The construction of the Chern classes from KH-theory to the Deligne cohomology (see 6.20 below) will be achieved by the cdh-sheafification of Gillet's Chern classes at the level of presheaves of simplicial sets, followed by considering the induced maps on the hypercoho- mologies. Therefore, in order to factor the classical Chern classes cQ X,p,q on Quillen K-theory through KH-theory, we only need to identify the target of the Chern class maps in (6.12) with the Deligne cohomology. To do this, we let for any X ∈ SchC, H ∗ an(X,F) denote the cohomology of the analytic space Xan with coefficients in the sheaf F on Can. Let Z → Sing∗ denote a fibrant replacement ∼=−→ ǫ∗(Sing∗). Set Z(q) = (2π√−1)qǫ∗(Sing∗) ∼= Rǫ∗(Z). of the sheaf Z on Can so that Rǫ∗(Z) / / / 26 AMALENDU KRISHNA AND PABLO PELAEZ an(X, Z) ∼= Hp an(X, Z) → Hp Zar(X, Z(q)), it suffices to show that the map Hp Lemma 6.15. For any X ∈ SmC, the map H p Proof. Since H p Hp cdh(X, Z(q)cdh) is an isomorphism. Zar(X, Z(q)) → cdh(X, Z(q)cdh) is an isomorphism. Let Cloc denote the category of schemes which are separated and of finite type over C. We will consider Cloc as a Grothendieck site with coverings given by maps Y ′ → Y where the associated map of the analytic spaces is a local isomorphism of the corresponding topological spaces [2, Expos´e XI p. 9]. Since a Nisnevich cover of schemes is a local isomorphism of the associated analytic spaces, there is a commutative diagram of morphisms of sites: (6.16) Cloc ν CNis δ τ Can ǫ / CZar. an(X,F ∗) → H p Since every local isomorphism of analytic spaces is refined by open coverings, it is well known that the map Hp loc(X,F ∗) is an isomorphism for any complex of sheaves on Can (see, for instance, [65, Prop. 3.3, Thm. 3.12]). We set (Z(q))N is = τ ∗(Z(q)) = ν∗ ◦ δ∗(Sing∗). We observe that for every i ∈ Z, the cohomology sheaf Hi associated to the complex Z(q) is isomorphic to the Zariski (or Nis- nevich) sheaf on SchC associated to the presheaf U 7→ H i an(U, Z). But this latter presheaf on SmC is homotopy invariant with transfers. It follows from [81, Cor. 1.1.1] that the map Hp Zar(X, Z(q)) → Hp N is(X, (Z(q))N is) is an isomorphism. We are thus reduced to showing that the map Hp But this follows again from [81, Cor. 1.1.1, 5.12.3, Thm. 5.13] because each Hi ∼= Riν∗(Z) is a Nisnevich sheaf on SmC associated to the homotopy invariant presheaf with transfers U 7→ H i (cid:3) cdh(X, (Z(q))cdh) is an isomorphism for X ∈ SmC. N is(X, (Z(q))N is) → Hp an(U, Z). The proof is complete. For any X ∈ SchC, there are natural maps Zar(X•, ΓD(q)) → Hp (6.17) H p Lemma 6.18. For a projective scheme X over C, the map H p is an isomorphism. D(X, Z(q)) ∼= Hp N is(X•, (ΓD(q))N is) → Hp D(X, Z(q)) → Hp cdh(X•, (ΓD(q))cdh). cdh(X•, (ΓD(q))cdh) Proof. Our assumption implies that each component Xp of the simplicial scheme X• is smooth and projective. Given a complex of sheaves F ∗ • (in Zariski or cdh-topology), there is a spectral sequence p )Zar/cdh) ⇒ Hp+q Zar/cdh(X•, (F ∗ • )Zar/cdh), Ep,q 1 = Hq Zar/cdh(Xp, (F ∗ Zar(X, ΓD(q)) → Hp (see, for instance, [3, Appendix]). Using this spectral sequence and (6.17), it suffices to show that the map H p cdh(X, (ΓD(q))cdh) is an isomorphism for any smooth projective scheme X over C. For q ≤ 0, this follows from 6.15. So we assume q > 0. analytic sheaves Z(q) → OXan → Ω1 triangle Since X is smooth and projective, the analytic Deligne complex Z(q)D is the complex of . In particular, there is a distinguished Xan → ··· → Ωq−1 Xan Rǫ∗(Ω<q Xan [−1]) → ΓD(q) → Z(q) → Rǫ∗(Ω<q Xan ) in the derived category of sheaves on XZar. Since X is projective, it follows from GAGA that the natural map Ω<q ) is an isomorphism in the derived category of sheaves on XZar. In particular, we get a distinguished triangle in the derived category of sheaves on XZar: X/C → Rǫ∗(Ω<q Xan (6.19) X/C[−1] → ΓD(q) → Z(q) → Ω<q Ω<q X/C. / / / THE SLICE SPECTRAL SEQUENCE FOR SINGULAR SCHEMES AND APPLICATIONS 27 We thus have a commutative diagram of exact sequences: Hp−1 Zar (X, Z(q)) Hp−1 cdh (X, (Z(q))cdh) Hp−1 Zar (X, Ω<q X/C) / Hp−1 cdh (X, (Ω<q X/C)cdh) H p Zar(X, ΓD(q)) / H p cdh(X, (ΓD(q))cdh) / Hp Zar(X, Z(q)) / Hp cdh(X, (Z(q))cdh) Hp Zar(X, Ω<q X/C) / Hp cdh(X, (Ω<q X/C)cdh). It follows from 6.15 that the first and the fourth vertical arrows from the left are iso- morphisms. The second and the fifth vertical arrows are isomorphisms by [19, Cor. 2.5]. We conclude that the middle vertical arrow is also an isomorphism and this completes the proof. (cid:3) As a combination of 6.3, (6.14) and 6.18, we obtain a theory of Chern classes from KH- theory to Deligne cohomology as follows. Theorem 6.20. For every projective scheme X over C, there are Chern class homomorphisms cX,p,q : KHp(X) → H 2q−p D (X, Z(q)) such that for any morphism of projective C-schemes f : Y → X, one has f ∗◦cX,p,q = cY,p,q◦f ∗. ∼=−→ Proof. We only need to show that there is a natural isomorphism αX : Hp H p D(X, Z(q)). Given a morphism of projective C-schemes f : Y → X, there exists a commutative diagram cdh(X, (ΓD(q))cdh) Y• f• X• pY pX Y f / X, where the vertical arrows are the simplicial hypercoverings maps. In particular, there is a commutative diagram Hp Zar(X, ΓD(q)) ◗◗◗◗ ◗◗◗ ◗◗◗◗ Hp (◗◗ cdh(X, (ΓD(q))cdh) Hp Zar(Y, ΓD(q)) ◗◗◗◗ ◗◗◗◗ ◗◗◗◗ (◗ Hp cdh(Y, (ΓD(q))cdh) H p D(X, Z(q)) H p D(Y, Z(q)) ◗◗◗◗ ◗◗◗ (◗◗ ◗◗◗◗ Hp cdh(X•, (ΓD(q))cdh) ◗◗◗◗ ◗◗◗◗ (◗◗ ◗◗◗ cdh(Y•, (ΓD(q))cdh). / Hp Using 6.18, we get a map αX : Hp D(X, Z(q)) such that f ∗ ◦ αX = αY ◦ f ∗ for any f : Y → X as above. Moreover, we have shown in the proof of 6.18 that this map is an isomorphism if X ∈ SmC. Since the source as well as the target of αX satisfy cdh(X, (ΓD(q))cdh) → H p / / / / / / / / / / / / / / / / / / / / ( ( / / / / ( ( / 28 AMALENDU KRISHNA AND PABLO PELAEZ cdh-descent by 6.18 (see [81, Lem. 12.1]), we conclude as in the proof of 6.3 that αX is an isomorphism for every projective C-scheme X. (cid:3) 7. Applications II: Intermediate Jacobian and Abel-Jacobi map for singular schemes Recall that a very important object in the study of the geometric part of motivic cohomology of smooth projective varieties is an intermediate Jacobian. The intermediate Jacobians were defined by Griffiths and they receive the Abel-Jacobi maps from certain subgroups of the geometric part H 2∗(X, Z(∗)) of the motivic cohomology groups. A special case of these intermediate Jacobians is the Albanese variety of a smooth projective variety. The most celebrated result about the Albanese variety in the context of algebraic cycles is that the Abel-Jacobi map from the group of 0-cycles of degree zero to the Albanese variety is an isomorphism on the torsion subgroups. This theorem of Roitman tells us that the torsion part of the Chow group of 0-cycles on a smooth projective variety over C can be identified with the torsion subgroup of an abelian variety. Roitman's torsion theorem has had enormous applications in the theory of algebraic cycles and algebraic K-theory. For example, it was predicted as part of the conjectures of Bloch and Beilinson that the Chow group of 0-cycles on smooth affine varieties of dimension at least two should be torsion-free. This is now a consequence of Roitman's torsion theorem. We hope to use the Roitman's torsion theorem of this paper to answer the analogous question about the motivic cohomology H 2d(X, Z(d)) of a d-dimensional singular affine variety in a future project. It was predicted as part of the relation between algebraic K-theory and motivic cohomology that the Chow group of 0-cycles should be (integrally) a subgroup of the Grothendieck group. This is also now a consequence of Roitman's theorem. We shall prove the analogue of this for singular schemes in the next section. Recall that the Riemann-Roch theorem says that this inclusion of the Chow group inside the Grothendieck group is always true rationally. For applications concerning the relation between Chow groups and ´etale cohomology, see [12]. In this section, we shall apply the theory of Chern classes from KH-theory to Deligne cohomology from § 6 to construct the intermediate Jacobian and Abel-Jacobi map from the geometric part of the motivic cohomology of any singular projective variety over C. In the next section, we shall use the Abel-Jacobi map to prove a Roitman torsion theorem for singular schemes. As another application of our Chern classes and the Roitman torsion theorem, we shall show that the cycle map from the geometric part of motivic cohomology to the KH groups, constructed in 5.9, is injective for a large class of schemes. 7.1. The Abel-Jacobi map. In the rest of this section, we shall consider all schemes over C and mostly deal with projective schemes. Let X be a projective scheme over C of dimension d. Let Xsing and Xreg denote the singular (with the reduced induced subscheme structure) and the smooth loci of X, respectively. Let r denote the number of d-dimensional irreducible components of X. We shall fix a resolution of singularities f : eX → X and let E = f −1(Xsing) throughout this section. The following is an immediate consequence of the cdh-descent for Deligne cohomology. Lemma 7.2. For any integer q ≥ d + 1, one has H q+d+i Proof. If X is smooth, it follows immediately from (6.19). Deligne cohomology (see 6.18 or [6, Variant 3.2]) implies that there is an exact sequence (X, Z(q)) = 0 for i ≥ 1. In general, the cdh-descent for D H q+d+i−1 D (E, Z(q)) → H q+d+i D (X, Z(q)) → H q+d+i D We conclude the proof by using this exact sequence and induction on dim(X). (cid:3) (eX, Z(q)) ⊕ H q+d+i D (Xsing, Z(q)). THE SLICE SPECTRAL SEQUENCE FOR SINGULAR SCHEMES AND APPLICATIONS 29 It follows from the definition of the Deligne cohomology that there is a natural map of complexes ΓD(q)X → Z(q)X (see (6.19)) and in particular, there is a natural map H p an(X, Z(q)). For any integer 0 ≤ q ≤ d, the intermediate Jacobian J q(X) D(X, Z(q)) is defined so that we have an exact sequence κX−−→ H p It follows from 6.20 that there is a commutative diagram 0 → J q(X) → H 2q D (X, Z(q)) κX−−→ H 2q an(X, Z(q)). cX,d,0 KH0(X) H 2d D (X, Z(d)) (7.3) f ∗ f ∗ cX,d,0 KH0(eX) / H 2d D (eX, Z(d)) κX κ eX H 2d an(X, Z(d)) f ∗ / H 2d an(eX, Z(d)). an(eX, Z(d)) ⊕ H 2d κX−−→ H 2d It follows from (6.19) that κ eX is surjective. The cdh-descent for the Deligne cohomology and 7.2 together imply that the middle vertical arrow in (7.3) is surjective. The cdh-excision property of singular cohomology (see [24, 8.3.10]) yields an exact sequence H 2d−1 an (E, Z(d)) → H 2d an(X, Z(d)) → H 2d an(Xsing, Z(d)) → H 2d+1 an (E, Z(d)). Since Xsing and E are projective schemes of dimension at most d− 1, it follows that the right vertical arrow in (7.3) is an isomorphism. We conclude that there is a short exact sequence (7.4) 0 → J d(X) → H 2d D (X, Z(d)) an(X, Z(d)) → 0. A similar Mayer-Vietoris property of the motivic cohomology yields an exact sequence H 2d−1(E, Z(d)) → H 2d(X, Z(d)) → H 2d(eX, Z(d)) ⊕ H 2d(Xsing, Z(d)) → H 2d+1(E, Z(d)). It follows from 5.1 that H 2d(Xsing, Z(d)) = H 2d+1(E, Z(d)) = 0. In particular, there exists a short exact sequence H 2d−1(E, Z(d)) (7.5) 0 → H 2d−1(eX, Z(d)) + H 2d−1(Xsing, Z(d)) → H 2d(X, Z(d)) → H 2d(eX, Z(d)) → 0. Since the map H 2d(eX, Z(d)) ∼= CHd(eX) → H 2d an(eX, Z(d)) is the degree map which is surjec- tive, we conclude that the 'degree' map H 2d(X, Z(d)) → H 2d let Ad(X) denote the kernel of this degree map. It follows from 6.20 that there is a Chern class map (take p = 0) cX,q : KH0(X) → H 2q D (X, Z(q)). Theorem 5.9 says that the spectral sequence (4.28) induces a cycle class map cycX,0 : H 2d(X, Z(d)) → KH0(X). Composing the two maps, we get a cycle class map from motivic to Deligne cohomology an(X, Z(d)) is also surjective. We (7.6) X : H 2d(X, Z(d)) → H 2q ecd and a commutative diagram of short exact sequences: D (X, Z(q)) 0 / Ad(X) / H 2d(X, Z(d)) H 2d an(X, Z(d)) / 0 (7.7) AJd X ecd X 0 / J d(X) / H 2q D (X, Z(q)) / H 2d an(X, Z(d)) / 0. It is known that J d(X) is a semi-abelian variety whose abelian variety quotient is the classical Albanese variety of eX (see [10, Thm. 1.1] or [7]). The induced map AJd X : Ad(X) → J d(X) will be called the Abel-Jacobi map for the singular scheme X. We shall prove our main result about this Abel-Jacobi map in the next section. Here, we recall the following description of J d(X) in terms of 1-motives. Recall from [8, § 12.12] that every projective / / / / / / / / / / / / / / / 30 AMALENDU KRISHNA AND PABLO PELAEZ scheme X of dimension d over C has an 1-motive Alb+(X) associated to it. This is called the cohomological Albanese 1-motive of X. This is a generalization of the Albanese variety of smooth projective schemes. Theorem 7.8. ([7, Cor. 3.3.2]) For a projective scheme X of dimension d over C, there is a canonical isomorphism J d(X) ∼= Alb+(X). 7.9. Levine-Weibel Chow group and motivic cohomology. In order to prove our main theorem of this section, we need to compare the motivic cohomology of singular schemes with another 'motivic cohomology', called the (cohomological) Chow-group of 0-cycles, introduced by Levine and Weibel [62]. We shall assume throughout our discussion that X is a reduced projective scheme of dimension d over C. However, we remark that the following definition of the Chow group of 0-cycles makes sense over any ground field. Let Z0(X) denote the free abelian group on the closed points of Xreg. Definition 7.10. Let C be a pure dimension one reduced scheme in SchC. We shall say that a pair (C, Z) is a good curve relative to X if there exists a finite morphism ν : C → X and a closed proper subscheme Z ( C such that the following hold. (1) No component of C is contained in Z. (2) ν−1(Xsing) ∪ Csing ⊆ Z. (3) ν is local complete intersection morphism at every point x ∈ C such that ν(x) ∈ Xsing. Let (C, Z) be a good curve relative to X and let {η1,··· , ηr} be the set of generic points of C. Let OC,Z denote the semilocal ring of C at S = Z ∪ {η1,··· , ηr}. Let C(C) denote the ring of total quotients of C and write O× C,Z for the group of units in OC,Z. Notice that OC,Z coincides with k(C) if Z = ∅. As C is Cohen-Macaulay, O× C,Z is the subgroup of k(C)× consisting of those f which are regular and invertible in the local rings OC,x for every x ∈ Z. Given any f ∈ O× C,Z ֒→ C(C)×, we denote by div(f ) the divisor of zeros and poles of f on C, that is defined as follows. If C1, . . . , Cr are the irreducible components of C, we set div(f ) to be the 0-cycle Pr i=1 div(fi), where (f1,··· , fr) = θ(C,Z)(f ) and div(fi) is the usual divisor of a rational function on an integral curve in the sense of [28]. Let Z0(C, Z) denote the free abelian group on the closed points of C \ Z. As f is an invertible regular function on C along Z, div(f ) ∈ Z0(C, Z). By definition, given any good curve (C, Z) relative to X, we have a pushforward map Z0(C, Z) ν∗−→ Z0(X). We shall write R0(C, Z, X) for the subgroup of Z0(X) generated by the set {ν∗(div(f ))f ∈ O× (X) denote the subgroup of Z0(X) generated by the image of the map Z0(C, Z, X) → Z0(X), where Z0(C, Z, X) runs through all good curves. (X) = Z0(X) We let CHBK (X) C,Z}. Let RBK RBK 0 0 . 0 If we let RLW (X) denote the subgroup of Z0(X) generated by the divisors of rational functions on good curves as above, where we further assume that the map ν : C → X is a closed immersion, then the resulting quotient group Z0(X)/RLW (X). There is a canonical surjection CHLW (X). However, we can say more about this map in the present context. This comparison will be an essential ingredient in the proof of 8.4. (X) is denoted by CHLW (X) ։ CHBK 0 0 0 0 0 Theorem 7.11. For a projective scheme X over C, the map CHLW isomorphism. 0 (X) ։ CHBK 0 (X) is an Proof. By [9, Lem. 3.13], there are cycle class maps CHLW (X) → K0(X) and one knows from [59, Cor. 2.7] that the kernel of the composite map is (d − 1)!-torsion. In follows that Ker(CHLW (X)deg 0. (X)) is torsion. In particular, it lies in CHLW (X) ։ CHBK 0 0 0 0 (X) → CHBK 0 THE SLICE SPECTRAL SEQUENCE FOR SINGULAR SCHEMES AND APPLICATIONS 31 On the other hand, it follows from [9, Prop. 9.7] that the Abel-Jacobi map CHLW (X)deg 0 → J d(X) (see [10, Th. 1.1]) factors through CHLW (X)deg 0 → J d(X). More- over, it follows from [10, Th. 1.1] that the composite map is isomorphism on the torsion It must subgroups. therefore be zero. (cid:3) In particular, Ker(CHLW (X)deg 0) is torsion-free. (X)deg 0 ։ CHBK (X)deg 0 ։ CHBK 0 0 0 0 0 In the rest of this text, we shall identify the above two Chow groups for projective schemes an(X, Z(d)) ∼= over C and write them as CHd(X). There is a degree map degX : CHd(X) → H 2d Zr. Let CHd(X)deg 0 denote the kernel of this degree map. In order to obtain applications of the above Abel-Jacobi map, we connect CHd(X) with the motivic cohomology as follows. C, 0)cdh) → H0 {x},cdh(U, C∗zequi(Ad Lemma 7.12. There is a canonical map γX : CHd(X) → H 2d(X, Z(d)) which restricts to a map γX : CHd(X)deg 0 → Ad(X). Proof. We let U denote the smooth locus of X and let x ∈ U be a closed point. The ex- cision for the local cohomology with support in a closed subscheme tells us that the map H0 {x},cdh(X, C∗zequi(Ad C, 0)cdh) is an isomorphism. On the other hand, the purity theorem for the motivic cohomology of smooth schemes and the iso- morphism between the motivic cohomology and higher Chow groups [85] imply that the map {x},cdh(U, C∗zequi(Ad H0 C, 0)cdh) is same as the map of the Chow groups Z ∼= CH0({x}) → CH0(U ). In particular, we obtain a map cdh(X, C∗zequi(Ad γx : Z → H0 We let γX ([x]) be the image of 1 ∈ Z under this map. This yields a homomorphism We first assume that X is a reduced curve. In this case, an easy application of the spectral γX : Z0(X) → H 2d(X, Z(d)). We now show that this map kills R0(X). sequence of 4.27 and the vanishing result of 5.1 shows that there is a short exact sequence C, 0)cdh) = H 2d(X, Z(d)). {x},cdh(X, C∗zequi(Ad C, 0)cdh) → H0 C, 0)cdh) → H0 cdh(U, C∗zequi(Ad (7.13) 0 → H 2(X, Z(1)) → KH0(X) → H 0(X, Z(0)) → 0. Using the isomorphism H 0(X, Z(0)) ∼=−→ H 0 an(X, Z) and the natural map K∗(X) → KH∗(X), we have a commutative diagram of the short exact sequences (7.14) 0 0 / Pic(X) / K0(X) H 0 an(X, Z) / 0 / H 2(X, Z(1)) / KH0(X) / H 0 an(X, Z) / 0. It follows from [9, Lem. 3.11] that the map Z0(X) → K0(X) given by cycX ([x]) = [O{x}] ∈ ∼=−→ Pic(X). Note that x ∈ U and hence the class K0(X) defines an isomorphism CH1(X) [O{x}] in K0(X) makes sense. We conclude from this isomorphism and (7.14) that the com- posite map Z0(X) → K0(X) → KH0(X) has image in H 2(X, Z(1)) and it factors through CH1(X). C,Z . We need to show that γX (ν∗(div(f ))) = 0. By [9, Lem. 3.4], we can assume that ν is an lci morphism. In particular, there is a functorial push-forward map ν∗ : H 2(C, Z(1)) → H 2d(X, Z(d)) by 3.6 We now assume d ≥ 2 and let ν : (C, Z) → X be a good curve and let f ∈ O× / / / / / / / / / 32 AMALENDU KRISHNA AND PABLO PELAEZ and [69, Def. 2.32, Thm. 2.33]. We thus have a commutative diagram: (7.15) γC Z0(C, Z) ∼= ⊕x /∈Z H 0({x}, Z(0)) H 2(C, Z(1)) ν∗ ∼= Z0(X) ν∗ ν∗ ⊕x /∈XsingH 0({x}, Z(0)) γX / H 2d(X, Z(d)). The two horizontal arrows on the right are the push-forward maps on the motivic coho- mology since the inclusion {x} ֒→ X is an lci morphism for every x /∈ Xsing. We have shown that γC(div(f )) = 0 and hence γX(ν∗(div(f ))) = ν∗(γC((div(f )))) = 0. Furthermore, the an(eX, Z(d)) ∼= Zr is the degree map. composite Z0(X) → H 2d(X, Z(d)) → H 2d(eX, Z(d)) → H 2d This shows that γX (Z0(X)deg 0) ⊆ Ad(X). (cid:3) 8. Applications III: Roitman torsion and cycle class map Let X be a projective scheme of dimension d over C. Using the map γX : CHd(X) → H 2d(X, Z(d)) and the Abel-Jacobi map AJd X of (7.7), we shall now prove our main result on the Abel-Jacobi map and Roitman torsion for singular schemes. We shall use the following lemma in the proof. Lemma 8.1. Let X be a reduced projective scheme of dimension d over C. There is a cycle class map cycQ X,0 : CHd(X) → K0(X) and a commutative diagram (8.2) cycQ X,0 CHd(X) K0(X) γX H 2d(X, Z(d)) / KH0(X). cycX,0 Proof. Every closed point x ∈ U defines the natural map Z = K0({x}) = K {x} (X) → K0(X) and hence a class [O{x}] ∈ K0(X). This defines a map cycQ X,0 : Z0(X) → K0(X) and it factors through CHd(X) by [62, Prop. 2.1]. Since CHd(X) is generated by the closed points in U , it suffices to show that for every closed point x ∈ U , the diagram 0 (8.3) K {x} 0 (X) / K0(X) KH {x} 0 (X) / KH0(X) commutes. But this is clear from the functorial properties of the map of presheaves K(−) → KH(−) on SchC. (cid:3) We can now prove: Theorem 8.4. Let X be a projective scheme over C of dimension d. Assume that either d ≤ 2 or X is regular in codimension one. Then, there is a semi-abelian variety J d(X) and an Abel-Jacobi map AJd X : Ad(X) → J d(X) which is surjective and whose restriction to the torsion subgroups AJd X : Ad(X)tors → J d(X)tors is an isomorphism. , , / / / / / / 1 1 / / / / / / THE SLICE SPECTRAL SEQUENCE FOR SINGULAR SCHEMES AND APPLICATIONS 33 Proof. We can assume that X is reduced. We first consider the case when X has dimension at most two but has arbitrary singularity. In this case, we only need to prove that AJd X is surjective and its restriction to the torsion subgroups is an isomorphism. The map AJd X is induced by the Chern class map cX,d,0 : KH0(X) → H 2d D (X, Z(d)) and the composite map K0(X) → KH0(X) → H 2d D (X, Z(d)) is the Gillet's Chern class map C Q X,d,0 of (6.14). Composing these maps with the cycle class maps and using 8.1, we get a commutative diagram CHd(X)deg 0 γX Ad(X) (8.5) ❏ ❏ ❏ ❏ ❏ ❏ AJd,Q X ❏ ❏ $❏ AJd X J d(X). The map AJd,Q X : J d(X)tor X is surjective and is an isomorphism on the torsion subgroups by [6, Main Theorem]. It follows that AJd X is also surjective. To prove that it is an isomorphism on the torsion subgroups, we apply 7.8 and [8, Cor. 13.7.5]. It follows from these results that there ∼=−→ Ad(X)tor. Since J d(X) is a semi-abelian variety, is indeed an isomorphism φd we know that for any given integer n ≥ 1, the n-torsion subgroup nJ d(X) is finite. It follows that nAd(X) and nJ d(X) are finite abelian groups of same order. We conclude that the Abel-Jacobi map AJd X : nAd(X) → nJ d(X) between finite abelian groups which have same order. Therefore, this map will be an isomorphism if and only if it is a surjection. But this is true by (8.5) because we have seen above that the composite map AJd,Q X is isomorphism between the n-torsion subgroups. Since n ≥ 1 is arbitrary in this argument, we conclude the proof of the theorem. We now consider the case when X has arbitrary dimension but it is regular in codimension X : Ad(X) → J d(X) induces the map AJd one. Let f : eX → X be a resolution of singularities of X. It is then known that J d(X) ∼= J d(eX) = Alb(eX) (see [63, Rem. 2, pp. 505]). We have a commutative diagram Since the lower horizontal arrow in this diagram is an isomorphism, it uniquely defines the X . The map f ∗ ◦ γX is known to be surjective by the moving lemma for is also known to Abel-Jacobi map AJd 0-cycles on smooth schemes. In particular, f ∗ is surjective. The map AJd eX be surjective. It follows that AJd X is surjective. To prove that this is an isomorphism on the torsion subgroups, we can argue exactly as X is surjective on the X (and also ) is isomorphism on the n-torsion subgroups by [10, Th. 1.1]. This finishes the proof of (cid:3) in the first case of the theorem. This reduces us to showing that AJd n-torsion subgroups for every given integer n ≥ 1. But this follows because AJLW AJd eX the theorem. Remark 8.7. For arbitrary d ≥ 1, the map AJd,Q X in (8.5) is known to be an isomorphism only up to multiplication by (d−1)!. This prevents us from extending 8.4 to higher dimensions if X has singularities in codimension one. We also warn the reader that unlike AJd,Q X in (8.5), the map AJLW X in (8.6) is not defined via the Chern class map on K0(X). These maps coincide only up to multiplication by (d − 1)!. (8.6) CHd(X)deg 0 γX / Ad(X) f ∗ ◆ ◆ ◆ ◆ ◆ ◆ ◆ AJLW X ◆ ◆ ◆ &◆ AJd X J d(X) ∼= AJd eX Ad(eX) / J d(eX). / / $ / / & / / / / / / 34 AMALENDU KRISHNA AND PABLO PELAEZ 8.8. Injectivity of the cycle class map. Like the case of smooth schemes, the Roitman torsion theorem for singular schemes has many potential applications. Here, we use this to prove our next main result of this section. It was shown by Levine in [59, Thm. 3.2] that for a smooth projective scheme X of dimension d over C, the cycle class map H 2d(X, Z(d)) → K0(X) (see (5.11)) is injective. We generalize this to singular schemes as follows. Theorem 8.9. Let X be a projective scheme of dimension d over C. Assume that either d ≤ 2 or X is regular in codimension one. Then the cycle class map cyc0 : H 2d(X, Z(d)) → KH0(X) is injective. Proof. We note that cyc0 : H 2d(X, Z(d)) → KH0(X) is induced by the spectral sequences (4.28) and (5.11), both of which degenerate with rational coefficients. In particular, Ker(cyc0) is a torsion group. H 2d(X, Z(d)) We must therefore have Ker(cyc0) = 0. cyc0−−→ KH0(X) On the other hand, if dim(X) ≤ 2, (7.7) and 8.4 tell us that the composite map ecd If X is regular in codimension one, we let eX → X be a resolution of singularities and X : D (X, Z(d)) is isomorphism on the torsion subgroups. cX,0,d−−−→ H 2d consider the commutative diagram H 2d(X, Z(d)) cycX ,0 KH0(X) f ∗ f ∗ cyc eX,0 H 2d(eX, Z(d)) / K0(eX). We have shown in the proof of 8.4 that the left vertical arrow is isomorphism on the torsion subgroups. The bottom horizontal arrow is injective by [59, Th. 3.2]. It follows that cycX,0 is injective on the torsion subgroup. We must therefore have Ker(cycX,0) = 0. This finishes the proof. (cid:3) Acknowledgements. The authors would like to thank Shane Kelly for having read an earlier version of this paper and pointing out some missing arguments. The authors would like to thank the referee for very carefully reading the paper and suggesting many improvements. References [1] A. Altman, S. Kleiman, Bertini theorems for hypersurface sections containing a subscheme, Comm. Alg., 7, (1979), no. 8, 775 -- 790. [2] M. Artin, A. Grothendieck, J. L. Verdier, Th´eorie des topos et cohomologie ´etale des sch´emas. Tome 3. (French) S´eminaire de G´eom´etrie Alg´ebrique du Bois-Marie 1963-1964 (SGA 4), Lecture Notes in Mathematics, 305, (1973), vi+640. [3] M. Asakura, K. Sato, Chern class and Riemann-Roch theorem for cohomology theory without homotopy invariance, arXiv:1301.58297v1[math.AG], (2015). [4] J. Ayoub, Les six op´erations de Grothendieck et le formalisme des cycles ´evanescents dans le monde motivique I, Ast´erisque, 314, (2007), X+466 pp. [5] J. Ayoub, Les six op´erations de Grothendieck et le formalisme des cycles ´evanescents dans le monde motivique II, Ast´erisque, 315, (2008), vi+364 pp. [6] L. Barbieri-Viale, C. Pedrini, C. Weibel, Roitman's Theorem for singular Varieties, Duke Math. J., 84, (1996), 155 -- 190. [7] L. Barbieri-Viale, V. Srinivas, Albanese and Picard 1-motives , M´emoires de la SMF, 87, (2001). [8] L. Barbieri-Viale, B. Kahn, On the derived category of 1-motives, arXiv:1009.1900[math.AG], (2010). [9] F. Binda, A. Krishna, Zero cycles with modulus and zero cycles on singular varieties, Comp. Math., 154, (2018), 120 -- 187. [10] J. Biswas, V. Srinivas, Roitman's theorem for singular projective varieties, Comp. Math., 119, (1999), 213 -- 237. [11] S. Bloch, Algebraic cycles and higher K-theory, Adv. Math., 61, (1986), 267 -- 304. [12] S. Bloch, Torsion algebraic cycles and a theorem of Roitman, Comp. Math., 39, (1979), 107 -- 127. / / / THE SLICE SPECTRAL SEQUENCE FOR SINGULAR SCHEMES AND APPLICATIONS 35 [13] J. M. Boardman, Conditionally convergent spectral sequences, Homotopy invariant algebraic structures, Contemporary Mathematics, 239, (1999), 49 -- 84. [14] K. S. Brown, S. M. Gersten, Algebraic K-theory and generalized sheaf cohomology, Lecture Notes in Mathematics, 341, Springer-Verlag, (1973). [15] D. Cisinski, Descente par ´eclatements en K-th´eorie invariante par homotopie, Ann. of Math., 177, no. 2, (2013), 425 -- 448. [16] D.-C. Cisinski, F. D´eglise Triangulated categories of mixed motives, arXiv:0912.2110v3[math.AG], (2009). [17] D.-C. Cisinski, F. D´eglise Integral mixed motives in equal characteristic, Doc. Math., Extra Volume: Alexander S. Merkurjev's Sixtieth Birthday, (2015), 145 -- 194. [18] G. Cortinas, C. Haesemeyer, M. Schlichting, C. Weibel, Cyclic homology, cdh-cohomology and negative K-theory, Ann. of Math., (2), 167, (2008), no. 2, 549 -- 573. [19] G. Cortinas, C. Haesemeyer, C. Weibel, K-regularity, cdh-fibrant Hochschild homology, and a conjecture of Vorst, J. Amer. Math. Soc., 21, (2008), no. 2, 547 -- 561. [20] S. Dai, M. Levine, Connective algebraic K-theory, J. K-Theory, 13, (2014), no. 1, 9 -- 56. [21] F. D´eglise, Orientable homotopy modules, Amer. J. Math., 135, no. 2, (2013), 519 -- 560. [22] F. D´eglise, Orientation theory in arithmetic geometry, arXiv:1111.4203v2[math.AG], (2014). [23] P. Deligne, Voevodsky's lectures on motivic cohomology 2000/2001, In Algebraic topology, Abel Symp., 4, (2009), 355 -- 409. Springer, Berlin. [24] P. Deligne, Th´eorie de Hodge: III, Publ. Math. IHES, 44, (1974), 5 -- 77. [25] H. Esnault, V. Srinivas, E. Viehweg, The universal regular quotient of the Chow group of points on projective varieties, Invent. math., 135, (1999), 595 -- 664. [26] H. Esnault, E. Viehweg, Deligne-Beilinson cohomology, In, "Beilinson's conjectures on special values of L-functions", Persp. Math., Academic Press, Boston, 4, (1998), 305 -- 372. [27] E. Friedlander, V. Voevodsky, Bivariant cycle cohomology, In "Cycles, Transfers and Motivic Homology Theories", Annals of Math. Stud., 143, Princeton Univ. Press, Princeton, NJ, (2000), 138 -- 187. [28] W. Fulton, Intersection theory, Second Edition, Ergebnisse der Mathematik und ihrer Grenzgebiete 3, Folge. A Series of Modern Surveys in Mathematics, 2, Springer-Verlag, Berlin, (1998). [29] T. Geisser, L. Hesselholt, On the vanishing of negative K-groups, Math. Annalen, 348, (2010), no. 3, 707 -- 736. [30] S. Geller, C. Weibel, K1(A, B, I), J. reine angew. Math., 342, (1983), 12 -- 34. [31] H. Gillet, Riemann-Roch theorems for higher algebraic K-theory, Adv. Math., 40, (1981), 203 -- 289. [32] L. Gruson, M. Raynaud, Crit`eres de platitude et de projectivit´e, Invent. Math., 13, (1971), 1 -- 89. [33] C. Haesemeyer, Descent property of homotopy K-theory, Duke J. Math., 125, no. 3, (2004), 589 -- 619. [34] P. Hirschhorn, Model categories and their localizations, Mathematical surveys and monograph series, 99, Amer. Math. Soc., Providence, (2002). [35] M. Hovey, Model categories, Mathematical surveys and monograph series, 63, Amer. Math. Soc., Provi- dence, (1999). [36] M. Hovey, Spectra and symmetric spectra in general model categories, J. Pure Appl. Algebra, 165, (2001), no. 1, 63-127. [37] M. Hoyois, From algebraic cobordism to motivic cohomology, J. reine angew. Math., 702, (2015), 173 -- 226. [38] L. Illusie, Y. Laszlo, F. Orgogozo, Travaux de Gabber sur l'uniformisation locale et la cohomologie ´etale des sch´emas quasi-excellents, Ast´erisque, 363-364, (2014). [39] D. Isaksen, Flasque model structures for simplicial presheaves, K-Theory, 36, (2005), 371-395. [40] J. Jardine, Generalized ´etale cohomology theories, Progress in Math., 146, Birkhauser, (1997). [41] J. Jardine, Local homotopy theory, Springer Monographs in Mathematics, Springer-Verlag, (2015). [42] J. Jardine, Motivic symmetric spectra, Doc. Math., 5, (2000), 445 -- 552. [43] J. Jardine, The K-theory presheaf of spectra, Geometry and topology monographs, 16, (2009), 151 -- 178. [44] J. Jardine, The separable transfer, J. Pure Appl. Algebra, 177, (2003), no. 2, 177 -- 201. [45] S. Kelly, Voevodsky motives and ldh-descent, Ast´erisque, 391, (2017), iv+125. [46] S. Kelly, Vanishing of negative K-theory in positive characteristic, Comp. Math., 150, (2014), 1425 -- 1434. [47] M. Kerz, The Gersten conjecture for Milnor K-theory, Invent. Math., 175, (2009), 1 -- 33. [48] M. Kerz, F. Strunk, On the vanishing of negative homotopy K-theory, J. Pure Appl. Algebra, 221, (2017), 1641 -- 1644. [49] M. Kerz, F. Strunk, G. Tamme, Algebraic K-theory and descent for blow-ups, Invent. Math., 211, (2018), no. 2, 523 -- 577 [50] T. Kohrita, Deligne-Beilinson cycle maps for Lichtenbaum cohomology, Homology, Homotopy and Appli- cations, 16 (2), 2014, 1 -- 26. 36 AMALENDU KRISHNA AND PABLO PELAEZ [51] A. Krishna, On the negative K-theory of schemes in finite characteristic, J. Algebra, 322, (2009), no. 6, 2118 -- 2130. [52] A. Krishna, Zero cycles on singular surfaces, J. K-Theory, 4, (2009), no. 1, 101 -- 143. [53] A. Krishna, 0-cycles on singular schemes and class field theory, arXiv:1502.01515[math.AG], (2015). [54] A. Krishna, Zero cycles on affine varieties, arXiv:1511.04221v1[math.AG], (2015). [55] A. Krishna, P. Pelaez, Motivic spectral sequence for relative homotopy K-theory, arXiv:1801.00922, (2018). [56] A. Krishna, V. Srinivas, Zero cycles and K-theory on normal surfaces, Ann. of Math., 156, no. 2, (2002), 155 -- 195. [57] S. Lang, Abelian Varieties, Interscience Tracts in Pure and Applied Math., 7, Interscience Publishers, (1959). [58] M. Levine, Comparison of cobordism theories, J. Algebra, 322, (2009), no. 9, 3291 -- 3317. [59] M. Levine, Zero-cycles and K-theory on singular varieties, Proc. Symp. Pure Math., 46, Amer. Math. Soc., Providence, (1987), 451 -- 462. [60] M. Levine, Lambda-operations, K-theory and motivic cohomology, in Algebraic K-theory, (volume in the memory of R. Thomason), Fields Inst. Communications, 16, (1997), 13 -- 184. [61] M. Levine, The homotopy coniveau tower, J. Top., 1, (2008). 217 -- 267. [62] M. Levine, C. Weibel, Zero cycles and complete intersections on singular varieties, J. Reine Angew. Math., 359, (1985), 106 -- 120. [63] V. Mallick, Roitman's theorem for singular projective varieties in arbitrary characteristic, J. K-Theory 3, (2009), 501 -- 531. [64] C. Mazza, V. Voevodsky, C. Weibel, Lecture notes on motivic cohomology, Clay Mathematics Monographs, 2, American Mathematical Society, Providence, (2006) . [65] J. Milne, ´Etale Cohomology, Princeton Math. Ser., 33, Princeton University Press, Princeton, (1980). [66] F. Morel, A1 algebraic topology over a field, Lecture Notes in Mathematics, 2052, Springer-Verlag, (2012). [67] F. Morel, V. Voevodsky, A1-homotopy theory of schemes, Publ. Math. IHES, 90, (1999), 45 -- 143. [68] N. Naumann, M. Spitzweck, P. A. Østvaer, Motivic Landweber exactness, Doc. Math., 14, (2009), 551 -- 593. [69] A. Navarro, Riemann-Roch for homotopy invariant K-theory and Gysin morphisms, arXiv:1605.00980[math.AG], (2016). [70] A. Neeman, The Grothendieck duality theorem via Bousfield's techniques and Brown representability, J. Amer. Math. Soc., 9, (1996), no. 1, 205 -- 236. [71] A. Neeman, Triangulated categories, Annals of Math. Stud., 148, Princeton Univ. Press, Princeton, NJ, (2001). [72] I. Panin, K. Pimenov, O. Rondigs, A universality theorem for Voevodsky's algebraic cobordism spectrum, Homology, Homotopy and Applications, 10, 2008, no. 2, 211 -- 226. [73] P. Pelaez, Multiplicative Properties of the Slice Filtration, Ast´erisque, 335, (2011), xvi+289. [74] P. Pelaez, On the functoriality of the slice filtration, J. K-Theory, 11, (2013), no. 1, 55 -- 71. [75] J. Riou, Algebraic K-theory, A1-homotopy and Riemann-Roch theorems, J. Top., 3, (2010), 229 -- 264. [76] O. Rondings, P. Østvaer, Slices of hermitian K-theory and Milnor's conjecture on quadratic forms, Ge- ometry and Topology, 20, (2016), 1157 -- 1212. [77] J.-P. Serre, Morphismes universels et vari´et´e d' Albanese, S´eminaire Chevalley, expos´e 11, ann´ee, (1958/59). [78] M. Spitzweck, Algebraic cobordism in mixed characteristic, arXiv:1404.2542[math.AT], (2014). [79] M. Spitzweck, P. A. Østvaer, The Bott inverted infinite projective space is homotopy algebraic K-theory, Bull. London Math. Soc., 41, (2009), 281 -- 292. [80] A. Suslin, V. Voevodsky, Singular homology of abstract algebraic varieties, Invent. Math., 123, (1996), 61 -- 94. [81] A. Suslin, V. Voevodsky, Bloch-Kato conjecture and motivic cohomology with finite coefficients, in "The Arithmetic and Geometry of Algebraic cycles", Banff, Canada, NATO Sci. Ser. C Math. Phys. Sci., 548, Kluwer, Dordrecht, (2000), 117 -- 189. [82] M. Temkin, Tame destillation and desingularization by p-alterations, Ann. of Math., 186, (2017), 1 -- 30. [83] R. Thomason, T. Trobaugh, Higher algebraic K-theory of schemes and of derived categories, in 'The Grothendieck Festschrift III', Birkhauser Boston, (1990), 247 -- 435. [84] V. Voevodsky, A1-homotopy theory, Proceedings of the International Congress of Mathematicians, I, (Berlin, 1998). Doc. Math., (1998), Extra Vol. I, 579 -- 604. [85] V. Voevodsky, Motivic cohomology groups are isomorphic to higher Chow groups in any characteristic, Int. Math. Res. Not., 7, (2002), 351 -- 355. THE SLICE SPECTRAL SEQUENCE FOR SINGULAR SCHEMES AND APPLICATIONS 37 [86] V. Voevodsky, Open problems in the motivic stable homotopy theory. I, In, "Motives, polylogarithms and Hodge theory, Part I (Irvine, CA, 1998)", Int. Press Lect., 3, (2002), 3 -- 34. [87] V. Voevodsky, Unstable motivic homotopy categories in Nisnevich and cdh-topologies, J. Pure Appl. Algebra, 214, (2010), no. 8, 1399 -- 1406. [88] V. Voevodsky, Triangulated categories of motives over a field, Cycles, transfers, and motivic homology theories, Ann. of Math. Stud., 143, (2000). [89] C. Weibel, Homotopy algebraic K-theory, AMS Contem. Math., 83, (1988), 461 -- 488. [90] C. Weibel, Negative K-theory of Normal Surfaces, Duke J. Math., 108, (2001), 1 -- 35. School of Mathematics, Tata Institute of Fundamental Research, 1 Homi Bhabha Road, Co- laba, Mumbai, India. E-mail address: [email protected] Instituto de Matem´aticas, Ciudad Universitaria, UNAM, DF 04510, M´exico. E-mail address: [email protected]
0909.4690	4	0909	2010-07-19T15:34:40	Deformations of Kahler manifolds with non vanishing holomorphic vector fields	[ "math.AG", "math.DG" ]	In this article we study compact K\"ahler manifolds $X$ admitting non-singular holomorphic vector fields with the aim of extending to this setting the classical birational classification of projective varieties with tangent vector fields. We prove that any such a K\"ahler manifold $X$ admits an arbitrarily small deformation of a particular type which is a suspension over a torus; that is, a quotient of $F\times \mbb C^s$ fibering over a torus $T=\mbb C^s/\Lambda$. We derive some results dealing with the structure of such manifolds. In particular, we prove an extension of Calabi's theorem describing the structure of compact K\"ahler manifolds with $c_1(X)=0$ to general K\"ahler manifolds with non-vanishing vector fields. A complete classification when $X$ is a projective manifold or when $\dim X\leq s+2$ is also given. As an application, it is shown that the study of the dynamics of holomorphic tangent fields on compact K\"ahler manifolds reduces to the case of rational manifolds.	math.AG	math	DEFORMATIONS OF K AHLER MANIFOLDS WITH NON VANISHING HOLOMORPHIC VECTOR FIELDS JAUME AMOR ´OS, M `ONICA MANJAR´IN, MARCEL NICOLAU Abstract. We study compact Kahler manifolds X admitting nonvanishing holomorphic vector fields, extending the classical birational classification of projective varieties with tangent vector fields to a classification modulo de- formation in the Kahler case, and biholomorphic in the projective case. We introduce and analyze a new class of tangential deformations, and show that they form a smooth subspace in the Kuranishi space of deformations of the complex structure of X. We extend Calabi's theorem on the structure of com- pact Kahler manifolds X with c1(X) = 0 to compact Kahler manifolds with nonvanishing tangent fields, proving that any such manifold X admits an arbi- trarily small tangential deformation which is a suspension over a torus; that is, a quotient of F × Cs fibering over a torus T = Cs/Λ. We further show that ei- ther X is uniruled or, up to a finite Abelian covering, it is a small deformation of a product F × T where F is a Kahler manifold without tangent vector fields and T is a torus. A complete classification when X is a projective manifold, in which case the deformations may be omitted, or when dim X ≤ s + 2 is also given. As an application, it is shown that the study of the dynamics of holomorphic tangent fields on compact Kahler manifolds reduces to the case of rational varieties. Contents Introduction 1. Groups of automorphisms of Kahler manifolds 2. Locally free actions on Kahler and Fujiki manifolds 3. Projective manifolds with nonvanishing tangent fields 4. Tangential deformations of locally free holomorphic actions 5. Tangential deformations of locally free Cs-actions on Kahler and Fujiki manifolds 6. The approximation theorem 7. Structure of Kahler manifolds with non vanishing vector fields 8. Locally free Cs-actions on Kahler manifolds with small codimension 9. Dynamics of holomorphic vector fields References 1 5 7 10 13 15 21 25 31 34 36 Introduction This article is devoted to the study of compact Kahler manifolds admitting non- vanishing holomorphic vector fields or, what is equivalent, endowed with a locally 1991 Mathematics Subject Classification. Primary 32G07, 32M; Secondary 14M17, 14L27, 32J27, 37F75. Key words and phrases. Kahler manifold, deformation, vector field, Fujiki manifold. This work was partially supported by the Ministerio de Educaci´on y Ciencia of Spain, grants MTM2008-02294 and MTM2009-14163-C02-02, and by the Generalitat de Catalunya grants 2009SGR 1207 and 2009SGR 1284. The second author was supported by the Programa de Movil- idad de Recursos Humanos del Plan Nacional de I-D+I 2008/2011. 1 2 JAUME AMOR ´OS, M `ONICA MANJAR´IN, MARCEL NICOLAU free holomorphic Cs-action for a certain s > 0. We introduce a particular type of deformations of such manifolds and, as an application, we describe quite precisely their structure. Our study relies on the work carried out simultaneously and in- dependently by A. Fujiki [17] and D. Lieberman [29]. We use specially their deep result on the structure of the group of holomorphic automorphisms of a compact Kahler manifold (cf. Theorem 1.2). Our work is a continuation of the classical study of complex projective manifolds with tangent vector fields, which led to their birational classification up to finite covering, successively developed by F. Severi, R. Hall and D. Lieberman ([32], [21], [28]). The conclusions were summarized in Theorem 1 of [28] as Theorem 0.1. Let X be a complex projective manifold with a holomorphic tangent vector field v. Then: a) X has a finite ´etale Abelian covering X ′ which is birational to a product M × T × CP r, where M is another projective manifold, T an Abelian variety, and CP r the projective space of dimension r ≥ 0, such that the lift of the tangent vector field v to X ′ has trivial component in T M . b) If the tangent vector field v on X vanishes at some point then X is ruled. In Theorem 3.2 below we continue this classification in the case of nonvanishing vector fields, obtaining a classification up to biholomorphism by using suspensions over Abelian varieties (see Example 2.4) and repeated ´etale Abelian coverings in order to exhaust all tangent vector fields. Tangent vector fields play a role in the classification of algebraic varieties, spe- cially in the case of Kodaira dimension 0. The earliest indication of this fact known to the authors is the result stated by Eugenio Calabi in [9]. In that article, E. Calabi formulated his celebrated conjecture about Kahler manifolds admitting Ricci -- flat metrics and, under the assumption that the conjecture was true, he proved the result stated as Theorem 0.2 below. This theorem describes the structure of com- pact Kahler manifolds with trivial first Chern class, i.e. with c1(X) = 0. It was proved for projective manifolds by Y. Matsushima in [31] using previous results by A. Lichnerowicz, but without the assumption of Calabi's conjecture, and extended to the general Kahler case by F.A. Bogomolov [7] and independently by Lieberman [29]. Theorem 0.2 (Calabi, [9], [31], [7], [29]). Let X be a compact Kahler manifold with c1(X) = 0 and b1(X) > 0. Then X admits, as a finite ´etale covering, a product X ′ = F × T , with T a complex torus of real dimension b1(X ′) and F a Kahler manifold with c1(F ) = 0 and b1(F ) = 0. Moreover, X ′ is a regular covering space of X and the group of covering transformations is solvable. This Theorem was improved by Beauville in [3], where he completely describes the topology and the geometry of the fiber F . The proof of Calabi's theorem may be succintly described as follows: Use the triviality of the canonical bundle to show that X (or a finite covering of it) has q = 1 2 b1(X) linearly independent nonvanishing tangent vector fields; check that the Albanese morphism is onto, the tangent vector fields are lifts from the tangent bundle on the Albanese torus, and they make X (or a finite covering of it) a suspension with trivial monodromy over the Albanese torus. In this work we develop a refinement of the above argument to prove an extension to general Kahler manifolds of Calabi's theorem, namely our Theorem 3.2 and Proposition 7.3 may be condensed as: DEFORMATIONS OF K AHLER MANIFOLDS WITH VECTOR FIELDS 3 Theorem 0.3. Let X be a compact Kahler manifold with s non vanishing tangent vector fields v1, . . . , vs, defining a locally free Cs-action. Then X admits a small deformation Xǫ, which is a suspension over an s -- dimensional torus T . Furthermore, there is a finite Abelian covering X ′ ǫ of a small deformation Xǫ of X, a torus T ′ of dimension s′ ≥ s and a compact Kahler manifold F without non vanishing vector fields in such a way that, (i) if kod(X) ≥ 0 then X ′ (ii) if kod(X) = −∞ then X ′ ǫ = F × T ′ and F has no vector fields, ǫ is a topologically trivial suspension over T ′ and fiber F . In both cases kod(F ) = kod(X). If X is a complex projective manifold one may omit the small deformation of the complex structure, and the above decomposition holds for Xǫ = X. Here, kod(X) denotes the Kodaira dimension of X. The definition of a suspen- sion manifold is recalled in Example 2.4. We notice that a locally free holomorphic Cs-action, i.e. a holomorphic Cs-action for which every isotropy group is discrete, is determined by s holomorphic vector fields v1, . . . , vs generating an s-dimensional Abelian Lie algebra a such that each vector field v in a different from zero is a non vanishing vector field. On a compact Kahler manifold X, the set h1 of vector fields having zeros is an ideal of the Lie algebra h of holomorphic vector fileds on X, there is a direct sum decomposition h = h1 ⊕ a, where a is an Abelian subalgebra of h, and X admits a locally free holomorphic Cs-action if and only if dim a > 0 (cf. Propositions 1.4 and 1.6). Hence the hypothesis in the above Theorem is fulfilled whenever s = dim(h/h1) > 0. This is the case, in particular, if X admits a non vanishing vector field. The small deformation of the complex structure of X cannot be avoided in the general Kahler case if c1(X) 6= 0, as it is shown by the examples discussed in Section 7. On the other hand, for X compact Kahler with c1(X) = 0, the dimension of the space of holomorphic vector fields on X is b1(X)/2 and this fact also renders innecessary the small deformation of complex structure, thus Corollary 6.8 to our Theorem 0.3 refines Calabi's theorem, as we show that the group of covering trans- formations is in fact Abelian. This refinement had been previously established in [7] in the case when X is a complex projective manifold. Besides this extension of Calabi's theorem from Calabi -- Yau manifolds to any Kahler manifolds with nonvanishing tangent vector fields, the authors hope that the results of this work may have further applications in the case of manifolds of Kodaira dimension zero. K. Ueno (cf. [33, §11]) made a conjecture, which is the extension of Calabi's one to this setting, and was refined by J´anos Koll´ar (cf. [25] 4.16) as: Conjecture 0.4 (Ueno, Koll´ar). Let X be a smooth, projective variety with kod(X) = 0. Then X has a finite, ´etale covering X such that X is birational to the product of an Abelian variety and a simply connected variety F with kod(F ) = 0. Corollary 8.10 is a modest contribution to this problem, establishing the conjec- ture if one has dim X − 2 linearly independent nonvanishing tangent vector fields, a particular case being threefolds with non vanishing vector fields. This Corollary is a particular case of the classification of compact Kahler and projective manifolds with dim X − 2 linearly independent nonvanishing tangent vector fields carried out in Section 8 based on Theorem 0.3 and the classification of curves and surfaces. 4 JAUME AMOR ´OS, M `ONICA MANJAR´IN, MARCEL NICOLAU In Section 9 we extend the results in [29] concerning the dynamical properties of a holomorphic tangent vector field v on a compact Kahler manifold X. Applying Theorem 0.3, we show that the dynamical system (X, v) becomes trivial after taking a finite unramified covering space of X if kod(X) ≥ 0 and that, in all cases, the dynamics of the system reduce to the dynamics of an Abelian group Cp × (C∗)q acting on a rational variety (Theorem 9.2). In order to achieve all the results described above, we introduce the notion of tangential deformation of a holomorphic group action. A locally free holomorphic action of a (connected) complex Lie group G on a complex manifold X defines a holomorphic foliation F on X whose leaves are the orbits of the action. G- equivariant deformations of X change the complex structure of X as well as the holomorphic foliation F . The existence of a versal space of deformations for equi- variant deformations of compact complex manifolds endowed with a holomorphic action was proved by Cathelineau in [14]. In this article we consider equivariant deformations of the locally free G-action for which the foliation F keeps fixed its holomorphic transverse structure. This type of deformations, called tangential de- formations of the action, also admits a versal space of deformations whose tangent space at the origin is naturally identified to the space of infinitesimal tangential deformations (Theorem 4.3). We focus on the case in which X is a Kahler manifold. Under that hypothesis and as a consequence of the mentioned theorem by Fujiki and Lieberman, the group G is necessarily Abelian. For a given locally free holomorphic Cs-action on a compact Kahler manifold X we construct a family XR of tangential deformations of the action parametrized by smooth space R with the following properties: (1) R is an open subset of an Euclidean space CN , complementary to an affine real algebraic variety. (2) The family XR is versal at each point r ∈ R (Theorem 5.5 and Proposition 5.19). (3) Each element Xr, with r ∈ R, is a Kahler manifold (Theorem 5.17). We emphasize that R is a (real) Zariski open set and not merely an small open ball of CN . Looking at general properties of tangential deformations of the action, we prove that the Kodaira dimension kod(X) of X is constant under tangential deformations. We also relate R to the Kuranishi space KX of X showing that the forgetful map R −→ KX has a smooth image of dimension s · b1(X)/2. In particular this shows the following (cf. Theorem 5.6) Theorem 0.5. The space H 1(X, ΘX ) of infinitesimal deformations of a compact Kahler manifold X endowed with a locally free Cs-action contains a subspace of dimension ℓ = s·b1(X)/2 of unobstructed infinitesimal deformations. Consequently the Kuranishi space KX contains a smooth subspace of dimension ℓ. Many of the results of this article are also valid for manifolds X belonging to the class C, the class introduced by A. Fujiki and whose elements are those manifolds that are bimeromorphic to Kahler manifolds. Hence some of the results are stated in that context. However, we do not dispose of a characterization of those suspension manifolds that are in the class C (cf. Remark 2.5). Moreover, the class C is not stable under small deformations and by these reasons we do not know if the structure theorems of Section 7 are valid for Fujiki manifolds and we are able to prove them only in the case of Kahler manifolds. Even if it is not said explicitly, all the group actions considered along this article are holomorphic and the covering spaces are always unramified (´etale). DEFORMATIONS OF K AHLER MANIFOLDS WITH VECTOR FIELDS 5 We are very indebted to M. Brunella for inspiring discussions and to F. Touzet who communicated the proof of Proposition 1.6 to us. 1. Groups of automorphisms of Kahler manifolds Let X be a compact complex manifold of dimension n. All along this article, with the only exception of Section 4, the manifold X will be assumed to be Kahler or, more generally, to belong to the class C introduced by A. Fujiki in [16]. We recall that a manifold belongs to the class C (or Fujiki class) if it is bimeromorphic to a Kahler manifold. Notice that Moishezon manifolds are in C and that a submanifold of a Fujiki manifold is again in the class C. Recently, Demailly and Paun have characterized compact manifolds in C as those manifolds carrying a Kahler current (cf. [15]). Compact manifolds X in the class C fulfill many of the properties of compact Kahler manifolds. In this article we will make use of the following ones: (i) holomor- phic forms on X are closed, (ii) there are C-antilinear isomorphisms H q(X, Ωp X ) ≡ H p(X, Ωq X ), (iii) the Hodge theorem holds, i.e. H m(X, C) ≡ ⊕p+q=mH q(X, Ωp X ) and (iv) H 1(X, C) is isomorphic to H 1(Alb(X), C), where Alb(X) is the Albanese torus of X. For the proof of the above statements and a general account on the properties of Fujiki manifolds we refer to [17] and [18]. In a sharp contrast, the class C is not stable under small deformations (cf. [10] and [26]). Let us denote by hX the Lie algebra of holomorphic vector fields on X and let h1 X be the subspace of hX whose elements are the vector fields on X annihilated by all the holomorphic 1-forms on X. Since global holomorphic 1-forms are closed, h1 X is an ideal of hX containing the derived subalgebra [hX , hX ] of hX . The following characterization of h1 X obtained by Carrell and Lieberman in [12] was extended by Fujiki to the class C in [16] . Theorem 1.1 (Carrell-Lieberman [12]). Let X be a compact Kahler manifold. Then h1 X is the vector space of holomorhic vector fields with zeros. In particular, if X admits a non vanishing holomorphic vector field v, then there is a holomorphic 1-form α on X such that α(v) = 1. Let φ : X → Alb(X) be the natural map from X to its Albanese torus Alb(X). Each element of the group AutC(X) of holomorphic automorphisms of X induces an automorphism of Alb(X) and the correspondence Φ : AutC(X) −→ AutC(Alb(X)) is a group morphism. The map φ is equivariant with respect to the action of AutC(X) on Alb(X) induced by Φ. The connected component of the identity C(Alb(X)) ∼= Alb(X) and its image Aut0 is a sub-torus TX of Alb(X). Notice that TX is contained in the image φ(X) of X into Alb(X). C(X) of AutC(X) is mapped by Φ into Aut0 The following result was obtained independently and at the same time by A. Fu- jiki and by D. Lieberman. The latter proved it for Kahler manifolds and the former for manifolds in the class C. Theorem 1.2 (Fujiki [17], Lieberman [29]). Let X be a compact complex manifold that is Kahler or, more generally, that belongs to the Fujiki class. Then there is an exact sequence of groups (1) 1 → L → Aut0 C(X) Φ→ TX → 1 where L is a linear algebraic group (so with finitely many connected components) with Lie algebra h1 C(X) such that the composition Φ ◦ ι is the universal covering of TX . X . Moreover, there is a group morphism ι : Cr → Aut0 6 JAUME AMOR ´OS, M `ONICA MANJAR´IN, MARCEL NICOLAU Corollary 1.3. If h1 X = 0 then Aut0 C(X) is a complex torus. Notice that Theorem 1.1 is also a corollary of Theorem 1.2. Another consequence of the theorem is the following: Proposition 1.4. There is an Abelian Lie subalgebra a of hX such that (2) hX = h1 X ⊕ a, where the direct sum is in the sense of vector spaces. Remarks 1.5. (a) We denote s = dimC TX = dimC a = dimC hX /h1 s ≤ dimC Alb(X) = b1(X)/2 and s ≤ dimC(X). X . Then we have (b) The subalgebra a is not unique. Such a Lie subalgebra can be character- ized as a vector subspace of hX of maximal dimension generated by non vanishing commuting vector fields. (c) Let v1, . . . , vs be a basis of a. Then the vector fields v1, . . . , vs are linearly independent at each point of X. (d) The choice of a basis v1, . . . , vs of a determines an action of Cs on X which is locally free, i.e. an action which is injective at the Lie algebra level or, what is equivalent, which has discrete isotropy groups. Conversely, given such a locally free action : Cm × X → X, of maximal rank (i.e. m is maximal), one has m = s and the fundamental vector fields of the action generate a subalgebra of hX complementary to h1 if the action factorizes through a torus T = Cs/Λ, then a is contained in the center of hX . X . If the orbits are given by a torus action, i.e. The following result is stated without proof in [29]. The proof we give here is due to F. Touzet. Proposition 1.6. The Abelian subalgebra a of hX fulfilling (2) can be chosen in the center of hX . Proof. First we notice that L can be assumed to be connected just by replacing TX by a finite covering, which is again a torus. Let Aut0(h1 X ) be the identity component of the automorphism group of the complex Lie algebra h1 X = Lie(L). Since the group L is normal in G = Aut0 C(X), the adjoint representation ψ : G → Aut0(h1 Adg 7→ g X ) is well defined and induces a group morphism ψ : TX −→ H = Aut0(h1 Ad(L) X ) . X ), as Lie(Ad(L)) is an ideal of Der0(h1 Notice that Ad(L) is closed in Aut0(h1 X ), as L is an algebraic group, and that Ad(L) is a normal subgroup of Aut0(h1 X ). Therefore the connected group Ad(L) is a normal and closed subgroup of the algebraic group Aut0(h1 [23]) and the image by ψ of the compact group TX is necessarily constant, thus equal to the identity. As a consequence, ψ has image contained in Ad(L). We deduce that, for each v ∈ hX , there is v1 = v1(v) ∈ h1 X ). Hence the quotient H is also a linear algebraic group (cf. X such that [v, w] = [v1, w] ∀w ∈ h1 X , that is, v − v1 commutes with all the elements in h1 X . We consider now v ∈ hX with the property that Φ({exp tv}) is dense in the torus TX . Then v − v1 has the same property. Hence the closure Av−v1 of the group {exp t(v − v1)} is an Abelian subgroup of G that is mapped onto TX by Φ. This implies that v − v1 is in the center of hX and therefore that Av−v1 is in the DEFORMATIONS OF K AHLER MANIFOLDS WITH VECTOR FIELDS 7 center of G. Now we take a group morphism ι : Cr → Av−v1 with the property that the composition Φ ◦ ι is the universal covering of TX. Then the morphism ι determines an Abelian subalgebra a of hX with the requiered properties. (cid:3) Remarks 1.7. (a) It follows from the above proposition that each non vanishing holomorphic vector field v on X is included in an Abelian Lie algebra a fulfilling (2). (b) Assume that the subalgebra a of h fulfilling the properties of Proposition (1.4) has been fixed and set a = hv1, . . . , vsi. The space of holomorphic 1-forms on X can be decomposed as H 0(X, Ω1 X ) = hα1, . . . , αsi ⊕ hβ1, . . . , βp−si, where p = b1(X)/2 = dimC Alb(X), and the forms αi and βk fulfill αi(vj ) = δi βk(vj ) = 0 for each j = 1, . . . , s. j and Although the forms αi are not uniquely determined unless p = s, the subspace hβ1, . . . , βp−si is canonically associated to the complex manifold X. In particular, it does not depend on the choice of the subalgebra a. In fact, hβ1, . . . , βp−si is just the kernel of the natural morphism H 0(X, Ω1 X ) −→ H 0(X, ΘX )∨, where ΘX denotes the sheaf of germs of holomorphic vector fields on X. A manifold X will be called ruled if it is bimeromorphic to a geometrically ruled manifold, that is if X is bimeromorphic to the total space of a fiber bundle Y → B, which is analytically locally trivial over a complex manifold B, has fiber a complex projective space CP k and has changes of trivialization in the sheaf PGL(k + 1, OB). In the case in which X is algebraic this is equivalent, by the GAGA theorem, to say that X is bimeromorphic to the trivial bundle CP k × B but this equivalence is no longer true in the setting of Kahler manifolds (cf. [17]). A manifold X will be called uniruled if every point x ∈ X lies in a rational curve C ⊂ X. Theorem 1.8 (Fujiki, [17]). If a compact manifold X in the class C admits a non- trivial holomorphic vector field v with vx = 0 for some point x ∈ X, then X is uniruled. Remark 1.9. If the manifold X in Theorem 1.8 is complex projective then it follows from the birational classification theorem (Theorem 0.1) that X is ruled. In [29], Lieberman states that if X is a compact Kahler manifold then it is also ruled. In the particular case when the automorphism group of X includes a C∗-action this was proved by Carrell and Sommese in [13]. The authors have not been able to complete Lieberman's argument from [29] in the general compact Kahler case. We end this section by recalling the following result of Lieberman concerning the Kodaira dimension, kod(X), of X and the dimension of the Lie algebra hX . Theorem 1.10 (Lieberman [29]). Let X be a compact Kahler manifold. Then (3) dimC hX + kod(X) ≤ dimC X. 2. Locally free actions on Kahler and Fujiki manifolds Let X be a compact manifold endowed with a locally free G-action. In case X is a Kahler or Fujiki manifold there are very few possibilities for the Lie group G. F. Bosio has remarked the following corollary of the mentioned theorem of Carrell and Lieberman. Theorem 2.1 (Bosio [8]). Let G be a connected complex Lie group acting locally freely on a compact manifold X belonging to the class C. Then G is Abelian. 8 JAUME AMOR ´OS, M `ONICA MANJAR´IN, MARCEL NICOLAU Proof. Theorem 1.1 implies that the induced action of G on the Albanese torus Alb(X) is also locally free. (cid:3) Here are some examples of compact complex manifolds naturally endowed with locally free Cs-actions. Example 2.2 (Tori). Complex tori, that is Ts = Cs/Λ where Λ ∼= Zs is a lattice, are the only compact manifolds X with a locally free Cs-action such that s = dimC X. Example 2.3 (Principal torus fibrations). The total space X of a principal torus bundle Ts −→ X ↓ B where B is a compact complex manifold, is endowed with a natural action of Ts and therefore of Cs. If X is a Kahler manifold then the Kahler metric can be made invariant by averaging it with respect to the action by the compact group Ts. Hence the base space B is necessarily a Kahler manifold. Furthermore, it will be seen below (cf. Remark 2.7 (a)) that the principal bundle is flat. Conversely, if B is a Kahler manifold and the bundle is flat then the total space is a Kahler manifold. This follows from a theorem of Blanchard (cf. [6], Th´eor`eme principal II). Example 2.4 (Suspensions over complex tori). Let ρ : Λ → AutC(F ) be a group representation, where Λ = hτ1, . . . , τ2si ∼= Z2s is a lattice of Cs and F is a compact complex manifold, and denote fi = ρ(τi) ∈ AutC(F ). The suspension of the repre- sentation ρ is the manifold X = F ×Λ Cs obtained as the quotient of the product F × Cs by the equivalence relation defined by the diagonal action of Λ, that is (x, z) ∼ (fi(x), z + τi), for i = 1, . . . , 2s. The suspension X = F ×Λ Cs fibers over the torus Ts = Cs/Λ with fiber F and the natural Cs-action on F × Cs commutes with Λ inducing a locally free action on X. We say that X is a suspension over Ts. It is proved in [30, Theorem 3.19] that X is a Kahler manifold if and only if F is Kahler and there are integers ni, for i = 1, . . . , s, such that f ni C(F ). i ∈ Aut0 Remark 2.5. The above characterizations of Kahler manifolds that are principal torus bundles or suspensions over tori stated in the last two examples rely on a criterium, due to Blanchard [6], for deciding if the total space of a fibration is a Kahler manifold. It is not know if Blanchard's criterium can be extended to the class C. Therefore we are not able to formulate such kind of characterizations for Fujiki manifolds. This is the main difficulty to extend the structure theorems in Section 7 to manifolds in the class C. From now on X will denote a given compact manifold in the class C and we will use the notation introduced in Section 1. We denote b1(X) = 2p. We also assume from now on that s = dimC h/h1 is strictly positive and we fix a subalgebra a ⊂ hX with h = h1 X ⊕ a as well as a basis v1, . . . , vs of a. We denote by the locally free Cs-action defined by the choice of the basis and by F the induced foliation. It follows from Theorem 1.2 that there are holomorphic 1-forms α1, . . . , αs such that αi(vi) = 1 and αj (vi) = 0 if j 6= i. The forms αi generate the cotangent bundle T ∗F of F at each point and we assume that they have also been fixed. The distribution ker α1 ∩ · · · ∩ ker αs is integrable and defines a holomorphic foli- ation G that depends on the choice of α1, . . . , αs. This foliation is transverse and complementary to F , and also invariant by the action of Cs. Therefore, we have DEFORMATIONS OF K AHLER MANIFOLDS WITH VECTOR FIELDS 9 Proposition 2.6. A foliation F on a compact manifold X in the class C defined by a locally free Cs-action admits a holomorphic foliation G, which is transverse and complementary to F , and invariant by the action . Remarks 2.7. (a) This proposition implies in particular that a principal torus bundle whose total space is a manifold in the class C is necessarily flat (cf. Example 2.3). In the case in which the leaves of G are compact they define a fibration of X over a complex torus of dimension s and X is a suspension. (b) Notice that all the leaves of G are biholomorphic. We denote by νF the normal bundle of the foliation F , i.e. νF = T X/T F . The existence of the Cs-invariant foliation G implies that, at each point of X, we can find local coordinates (t1, . . . , ts, z1, . . . , zn−s) with the properties (4) vi = ∂ ∂ti and αi = dti. In these coordinates F is defined by zk = const. and G by ti = const. As a consequence, we obtain: Proposition 2.8. There is a natural isomorphism between the canonical bundle KX of X and the determinant det ν∗F of the conormal bundle ν∗F of F . Proof. If (t1 i , z1 ηi = α1 ∧ · · · ∧ αs ∧ dz1 i , . . . , ts i , . . . , zn−s i ∧ · · · ∧ dzn−s i i ) are local coordinates fulfilling (4) and we define , then ηj = gjiηi with gji = det(cid:18) ∂zk i (cid:19). j ∂zl Hence the cocycle (gji) defines the line bundle KX as well as det ν∗F . (cid:3) We then obtain: Corollary 2.9. Assume that X is a suspension F ×Λ Cs over a torus T = Cs/Λ. Then one has kod(X) = kod(F ) and c1(F ) = ι∗c1(X) where ι : F ֒→ X is the identification of F with a fiber of the projection X → T . Proof. Notice that the tangent bundle T F of F is canonically isomorphic to the restriction of νF to F . As the Kodaira dimension is invariant by finite coverings we can assume, using [30, Theorem 3.19], that the monodromy of the suspension has ∼= (det ν∗F )⊗m values in Aut0 for every m ∈ Z. Consequently the spaces H 0(X, K ⊗m X ) are naturally identified with the subspaces of H 0(F, (det ν∗F )⊗m) = H 0(F, K ⊗m F ) invariant by the monodromy. Ueno proved [33, Corollary 14.8] that Aut0 C(F ) acts trivially on the pluricanonical section spaces H 0(F, K ⊗m C(F ). By Proposition 2.8 there are isomorphisms K ⊗m F ). Therefore h0(X, K ⊗m X ) = h0(F, K ⊗m F ) for every m. X The relation between the first Chern classes follows from the adjunction formula. (cid:3) Let Ωk X be the sheaf of germs of holomorphic k-forms on X. As remarked above p = b1(X)/2 ≥ s. We can choose holomorphic 1-forms β1, ..., βp−s on X such that {α1, ..., αs, β1, ..., βp−s} (5) is a basis of H 0(X, Ω1 X ) and that βj (vi) = 0 for i = 1, . . . , s and j = 1, . . . , p − s. Since βj are closed they are in fact basic forms with respect to F . We recall that a differential form γ is said to be basic with respect to a foliation if it fulfills iwγ = 0 and iwdγ = 0 for each local vector field w tangent to the foliation. We shall denote by Ωk X of those k-forms that are basic with respect to F . The following decomposition of global holomorphic forms on X will be useful in the sequel. X/F the subsheaf of Ωk 10 JAUME AMOR ´OS, M `ONICA MANJAR´IN, MARCEL NICOLAU Lemma 2.10. Each global holomorphic k-form γ on X can be written in a unique way as a sum γ = X0≤I≤min(k,s) αI ∧ γI with I = (i1, . . . , im), 1 ≤ i1 < · · · < im ≤ s and I = m, and where αI = αi1 ∧ · · · ∧ αim and γI is a holomorphic (k − m)-form basic with respect to F . In particular we have min(k,s) i H 0(X, Ωk X ) ∼= Mi=0 ^hα1, ..., αsiC ⊗ H 0(X, Ωk−i X/F ). Proof. Let γ be a global holomorphic k-form on X and set ℓ = min(k, s). For each I = (i1, . . . , im), with m ≤ k and 1 ≤ i1 < · · · < im ≤ s, and a given k-form ω we denote ωI = ivim ◦ · · · ◦ ivi1 ω. We put γ0 = γ and we define γ1, . . . , γℓ recursively by γm+1 = γm − XI=ℓ−m αI ∧ γI m. Notice here that the (k − ℓ + m)-forms γI they are closed. Moreover, by construction they fulfill ivγI tangent to F . Therefore they are basic with respect to F and m are holomorphic global forms and thus m = 0 for each vector v γ = ℓ Xm=1 XI=m αI ∧ γI ℓ−m + γℓ is the required decomposition. (cid:3) 3. Projective manifolds with nonvanishing tangent fields In this section we will assume that X is a complex projective manifold. Our study in this algebraic context becomes a continuation of the classical study of complex projective manifolds with holomorphic tangent vector fields, which led to their birational classification summarized in Theorem 0.1. The Albanese mapping and Poincar´e's reducibility theorem lead to a biholomorphic classification of these manifolds which is simpler than in the Kahler case. Following Theorem 1.2 and Remark 1.5 about the structure of the Lie algebra of holomorphic tangent vector fields, the existence of nonvanishing tangent vector fields is equivalent to the existence of a holomorphic locally free Cr -- action on the manifold. We will work directly with these locally free actions. The basic fact beyond the classical theory is: Proposition 3.1. Let X be a complex projective manifold, admitting a locally free holomorphic Cr-action. Then X is a suspension F ×Λ Cs over an Abelian variety of dimension s ≥ r, with fiber a connected projective manifold F . Proof. As X is closed Kahler, the Cr-action can be extended to a locally free holomorphic action of maximal rank s = dimC TX = dim hX /h1 X as we have recalled in Section 1. The Albanese torus Alb(X) of X is an Abelian variety, therefore the subtorus TX ⊂ Alb (X) in the Fujiki-Lieberman structure Theorem 1.2 is also an Abelian variety, and by Poincar´e's Reducibility Theorem (see [5, §5.3]) it has a complemen- tary Abelian subvariety Z such that TX ∩ Z is finite, TX + Z = Alb(X), and the addition law induces an isogeny c : Alb(X) → TX × Z. The composition of the Albanese morphism of X with this isogeny and the natural projection X φ −→ Alb(X) c−→ TX × Z p1−→ TX DEFORMATIONS OF K AHLER MANIFOLDS WITH VECTOR FIELDS 11 maps the tangent subspace of X generated by the Cs-action isomorphically into the tangent space of TX. Consequently X is a suspension over TX, with parallel transport given by the Cs-action and possibly disconnected fibers. The Stein fac- torization of X → TX is thus unramified, and yields a suspension X → NX with connected fiber F , over an Abelian variety NX isogenous to TX . (cid:3) If X is not ruled, it follows from Theorems 0.1 and 1.2 that Aut0 C(X) is an Abelian variety of dimension s = dim hX , isogenous to NX . The suspension morphism X → NX shows that the natural action of Aut0 C(X) on X is holomorphically injective, and the above quoted theorems make Proposition 3.1 essentially equivalent to Carrell's classification theorem for such actions of Abelian varieties (cf. [11, Thm. 8]). The classification of projective manifolds with locally free holomorphic Cs-actions follows from Proposition 3.1 and Theorem 1.8. Putting them together yields: Theorem 3.2. Let X be a complex projective manifold admitting a locally free holomorphic Cr-action. (a) If X is not uniruled, it is a quotient (F × T )/Γ, with T an Abelian variety of dimension s ≥ r, F a projective manifold with no holomorphic tangent vector fields and kod(F ) = kod(X), and Γ a finite Abelian group of biholomorphisms operating freely on F × T . (b) If X is uniruled, it admits a finite, Abelian, ´etale cover X ′ which is a suspension over an Abelian variety T of dimension s ≥ r, with uniruled fiber F such that any holomorphic tangent vector field on F has zeros. The monodromy of the suspension X ′ → T has values in the group Aut0 C(F ). Remark 3.3. This Theorem proves the algebraic part of Theorem 0.3. Its proof mirrors that of Proposition 6.5 which deals of the general case of Kahler manifolds. As examples in Section 7 show, the fact that Poincar´e's reducibility theorem is no longer valid in the Kahler setting forces the introduction of deformations of the complex structure. Proof. We may assume that the locally free action on X is of maximal rank s. By Proposition 3.1, X is a suspension over an s -- dimensional Abelian variety NX , isogenous to TX , and with fiber a connected projective manifold F1. Let the suspension structure on X be given by NX = Cs/Λ, with Λ = hτ1, . . . , τ2si a cocompact lattice defining the torus NX , ρ : Λ → AutC(F ) the monodromy of the suspension, and fi = ρ(τi) the generators of the monodromy. By 2.4, the fact that the total space X of the suspension is Kahler implies that all the monodromy automorphisms have a finite power f ni C(F ), i.e. there is a finite ´etale Abelian covering T1 → NX such that the pullback of X over T1 is a finite ´etale Abelian covering X1 → X, and also a suspension over T1 with fiber F1 and monodromy automorphisms in Aut0 i ∈ Aut0 C(F1). If F1 has no nonvanishing holomorphic tangent vector fields it can be checked that we have reached the sought cover: (i) kod(F1) = kod(X) by Corollary 2.9. (ii) If X is not uniruled neither is X1 nor F1. Indeed, as all the fibers of X1 → T1 are isomorphic, if F1 were uniruled there would be a rational curve passing by each given point of X1. Now Fujiki's Theorem 1.8 implies that h1 F1 = 0. Therefore the suspension monodromy has values in Aut0 C(F1) = {Id}, i.e. X1 is biholomorphic to F1 × T1. (iii) If X is uniruled so is its ´etale cover X1. Since rational curves map to points in the suspension basis T1, the fiber F1 is uniruled as well. If F1 has nonvanishing holomorphic tangent vector fields, we apply an iterative let hF1 be the Lie algebra of holomorphic vector fields in F1, and apply process: 12 JAUME AMOR ´OS, M `ONICA MANJAR´IN, MARCEL NICOLAU Proposition 1.6 to choose an Abelian subalgebra aF1 corresponding to a Cr -- action of maximal rank in F1, such that aF1 is central in hF1 . As the monodromy of the suspension X1 → T1 lies in Aut0 C(F1) it must com- mute with the central algebra aF1. Therefore its tangent vector fields may be extended by parallel transport from the fiber F1 to the suspension total space X1 as nonvanishing tangent vector fields. These extensions, to be denoted hv1, . . . , vri, commute with the horizontal vector fields hh1, . . . , hsi which are the lifts to X1 of hT1 = H 0(T1, ΘT1) defining the suspension's parallel transport. Thus aX1 = hv1, . . . , vr, h1 . . . hsi is an Abelian subalgebra of hX1 , spanned by r + s linearly independent nonvanishing vector fields. It has maximal rank among such algebras: the algebra hX1 is the direct sum hX1 = hh1, . . . , hsi ⊕ H 0(X1, Vert) where Vert is the sheaf spanned by the vector fields vertical with respect to the projection X1 → T1. As the horizontal fields h1, . . . , hs are central in hX1 , the morphism H 0(X1, Vert) → hF1 given by restriction to a fiber is injective. This induces a natural inclusion of Lie algebras hX1 = hh1, . . . , hsi ⊕ H 0(X1, Vert) ⊂ hh1, . . . , hsi ⊕ hF1 . If there existed an Abelian Lie subalgebra a ⊂ hX1 formed by nonvanishing tangent fields, with dim a > r + s, then its inclusion in hh1, . . . , hsi ⊕ hF1 would have an intersection with hF1 of dimension > r. This contradicts the maximality of aF1. By Proposition 3.1, the algebra aX1 determines a suspension X1 → N1, over an (r + s) -- dimensional Abelian variety, with a connected fiber F2. Again by 2.4, the monodromy automorphisms of this suspension have a finite power in Aut0 C(F2). Consequently, there is a finite ´etale Abelian covering T2 → N1 such that the pull- back X2 of X1 over T2 is a suspension with fiber F2 and monodromy in Aut0 C(F2). The induced map X2 → X1 is also a finite ´etale Abelian cover. Moreover, the composition of coverings X2 → X1 → X is still regular and Abelian. The reason for its regularity is that the automorphisms of the cover X1 → X are realized by integration up to time 1 in X1 of suitable linear combinations of the horizontal fields h1, . . . , hs. These tangent fields have become parallel transport vector fields of the suspension X1 → T1 by our choice of algebra hX1 , and the isogenous base change T2 → T1 lifts canonically to X2 the vector fields h1, . . . , hs, thus also the integration up to time 1 of their linear combinations. The Abelianity of the cover follows from the fact that the vector fields v1, . . . , vr, h1 . . . hs commute in X1, hence also in X2, and the automorphisms of both covers X1 → X, X2 → X1 are defined by integration up to time 1 of suitable linear combinations of them. The iteration step concludes here. We have obtained a suspension X2 → T2, with X2 a finite, Abelian ´etale cover of X, fiber F2 with kod(F2) = kod(X2) = kod(X), and monodromy in Aut0 C(F2). Repetition of the above procedure yields a sequence of suspensions Xi → Ti with fiber Fi, again with Xi Abelian ´etale covers of X, kod(Fi) = kod(X) and monodromy in Aut0 C(Fi). As the dimension of the basis tori T2, T3, . . . grows strictly the iterative process must reach a final suspension Xf → Tf such that its fiber Ff does not have any nonvanishing vector field, i.e. no holomorphic locally free Cr -- action. If X is not uniruled neither is Xf nor Ff . By Fujiki's Theorem 1.8 we have C(Ff ) = = 0. Hence the fiber Ff cannot have any tangent vector field so Aut0 h1 Ff {Id} and the trivial suspension Xf = Ff × Tf has been reached. If X is uniruled, so are all its ´etale covers up to Xf , and the fiber Ff as well (cid:3) because rational curves in Xf must map to a point in Tf . DEFORMATIONS OF K AHLER MANIFOLDS WITH VECTOR FIELDS 13 We can use the structure Theorem 3.2 to classify algebraic manifolds with suf- ficiently many vector fields and to establish for them Ueno's conjecture 0.4. This is done in Corollaries 8.9 and 8.10 below as specialization of the arguments in the Kahler case. 4. Tangential deformations of locally free holomorphic actions Through this section X will be an arbitrary compact complex manifold, i.e. not necessarily Kahler or belonging to the Fujiki class. Let us assume that X is endowed with a holomorphic action : G × X → X, where G is a connected complex Lie group. Such an action is defined by a representation (6) ρ : G −→ AutC(X) from G into the group AutC(X) of holomorphic automorphisms of X. When is locally free, i.e. it has discrete isotropy groups, then it induces a holomorphic foliation F on X whose leaves are the orbits of the action. Let F tr denote the transversely holomorphic foliation obtained from F by for- getting the complex structure along the leaves. One can consider deformations of the holomorphic foliation F that keep fixed its transversal type; that is, holomor- phic deformations Fr of F such that the transversely holomorphic foliation F tr r coincides with F tr. This type of deformations were studied in [20] and [19], where they are called f -deformations. In [20], G´omez-Mont studies the space of infini- tesimal f -deformations of holomorphic foliations, which is naturally identified to H 1(X, T F ), i.e. the first cohomology group of X with values in the tangent bundle T F of F . The existence of a versal or Kuranishi space for f -deformations, i.e. a germ of analytic space (Kf , 0) parametrizing a versal family of f -deformations of F , is proved in [19]. The tangent space of (Kf , 0) at 0 is H 1(X, T F ). Families of f -deformations of F can be viewed as unfoldings of F , that is families of complex structures on X for which F tr becomes holomorphic. More precisely, a family of f -deformations of F parametrized by a germ of analytic space (R, r0) is given by a proper and flat morphism π : XR → R, an identification ι : X ֒→ XR of X with the fiber Xr0 = π−1(r0) and a holomorphic foliation FR on XR of the same codimension than F , which is transverse to the projection π and such that the restriction Fr0 of FR to X ≡ Xr0 coincides with F . Such a family of f -deformations will be denoted by (XR, ι, π, FR). Notice however that f -deformations Fr of F need not to be equivariant, in the sense that Fr are not necessarily defined by an action. By this reason we give the following definition, which is inspired in the notion of equivariant deformations of complex manifolds introduced by Cathelineau in [14]. Definition 4.1. Let X be a compact complex manifold endowed with a locally free holomorphic action : G × X → X and let F be the foliation defined by . By a family of tangential deformations of we mean a family of f -deformations (XR, ι, π, FR) of F and a holomorphic action R of G on XR fulfilling the following properties: (i) R extends the action of G on X ≡ ι(X) = Xr0, (ii) R preserves the fibers of π and (iii) the orbits of the action are tangent to FR. Then R induces a locally free G-action r on Xr for each r ∈ R and the leaves of Fr = F Xr are the orbits of r. Such a family will be denoted by (XR, ι, π, R) and each pair (Xr, r), with r ∈ R, will be called a tangential deformation of (X0, 0) ∼= (X, ). We will just write (XR, R) instead of (XR, ι, π, R) when there is no danger of confusion. 14 JAUME AMOR ´OS, M `ONICA MANJAR´IN, MARCEL NICOLAU Definition 4.2. Two families (XR, ι, π, R) and (X ′ R) of tangential de- formations of (X, ), parameterized by the same germ of analytic space R, are said to be isomorphic if there is a G-equivariant biholomorphism φ : XR → X ′ R fulfilling φ ◦ ι = ι′ and π′ ◦ φ = π. R, ι′, π′, ′ A family of tangential deformations (XR, R) of (X, ) is called versal if it fulfills the following property: for any other family (XS, S) of tangential deformations of (X, ) there is an analytic morphism of germs of analytic spaces ϕ : S → R such that (i) the pull-back family (ϕ∗(XR), ϕ∗(R)) parameterized by S is isomorphic to (XS, S), and (ii) the tangent map ds0 ϕ of ϕ at the distinguished point s0 of S is unique. If such a versal family exists then it is unique up to isomorphism. Notice now that the sheaf of germs of holomorphic vector fields tangent to F is a G-sheaf and therefore we can consider the equivariant cohomology of X with values in the tangent bundle T F of F , that we denote by H ∗ G(X, T F ). Let Ω0,q(X, T F ) be the Frechet space of (0, q)-forms on X with values in T F and de- note by Cp h(G, Ω0,q(X, T F )) the space of holomorphic p-cochains with values in the G-module Ω0,q(X, T F ). Then H ∗ G(X, T F ) is the cohomology of the double complex (7) ((Cp h(G, Ω0,q(X, T F )))p,q, δ, ∂) where δ denotes the Eilenberg-Maclane coboundary operator and ∂ is the usual delta-bar operator. As usual, infinitesimal tangential deformations of (X, ) can be defined as isomorphism classes of families parameterized by the double point D = {∗, C(t)/(t2)}. The space of infinitesimal deformations is naturally identified to H 1 G(X, T F ). Moreover, to each family of tangential deformations (XR, R) of (X, ) there is associated in a natural way a linear map κ : Tr0R −→ H 1 G(X, T F ) which is called the Kodaira-Spencer map of the family (cf. [14]). The following theorem states the existence of a versal family and it can be proved in the same lines as the corresponding statement for equivariant deformations of complex manifolds (cf. [14, Th´eor`eme 1 and Proposition 1]). Theorem 4.3. Let X be a compact complex manifold endowed with a locally free holomorphic action : G × X → X and let F be the induced holomorphic foliation. Then there is a germ of analytic space (Kt G, 0) parametrizing a versal family of tangential deformations of the action . The tangent space to Kt G at 0 is isomorphic to H 1 G(X, T F ) and it is naturally identified to the space of infinitesimal tangential deformations of . The following proposition is a general fact in deformation theory and can be deduced easily from the existence of the versal space of tangential deformations. Proposition 4.4. Let (XR, R) be a family of tangential deformations of (X, ) whose parameter space R is smooth and such that the corresponding Kodaira-Spencer map κ is an isomorphism. Then the family of deformations (XR, R) is versal. G(X, T F ) and whose second term is Ep,q Associated to the double complex (7) there is an spectral sequence converging to H ∗ h(G, H q(X, T F )). Here the index h denotes the cohomology of holomorphic cochains. In particular one has the following exact sequence that is useful in the computation of H 1 2 = H p G(X, T F ) (8) 0 → H 1 h(G, H 0(X, T F )) → H 1 → H 2 χ → H 1(X, T F )G → G(X, T F ) h(G, H 0(X, T F )) → H 2 G(X, T F ). DEFORMATIONS OF K AHLER MANIFOLDS WITH VECTOR FIELDS 15 5. Tangential deformations of locally free Cs-actions on Kahler and Fujiki manifolds The aim of this section is to construct the versal family of tangential deformations of a locally free Cs-action on a given compact Kahler or Fujiki manifold X and to study the properties of that family. With this purpose we consider the exact sequence (8) when G is the Abelian group Cs. In that case T F is trivial as a G-bundle. So G = Cs acts trivially on h(G, H 0(X, T F )) is just the space of holomorphic group ho- H 0(X, T F ) ≡ Cs and H 1 momorphisms from G into Cs, that is the space End Cs of C-linear endomorphisms of Cs. Moreover, G = Cs acts also trivially on H 1(X, T F ) = H 1(X, OX )s ∼= H 0(X, Ω1 X )s, where the last isomorphism is C-antilinear. Therefore, the first terms of sequence (8) can be written (9) 0 → End Cs τ → H 1 G(X, T F ) χ → H 0(X, Ω1 X )s → . . . We will see below that the morphism χ is surjective and therefore that H 1 isomorphic to the direct sum End Cs⊕H 0(X, Ω1 of the family of tangential deformations of (X, ) that follows. G(X, T F ) is X )s. This motivates the construction As above v1, . . . , vs will denote the fundamental vector fields of the action corresponding to the canonical basis of Cs, α1, . . . , αs will be holomorphic 1-forms fulfilling αi(vj ) = δi X ) where βk are basic forms. Although the vector fields vi are naturally associated to the action , the 1-forms αi are not uniquely determined. If αi is another choice then αi − αi are basic forms, hence vanishing on T F . j and {α1, ..., αs, β1, ..., βp−s} a basis of H 0(X, Ω1 We denote Ξ = End Cs ⊕ H 0(X, Ω1 X )s. Given an element r = (C, θ) ∈ Ξ, where θ = (θ1, . . . , θs) and θi = Ps j=1 ai (10) αi r = αi + θ i = αi + s jαj +Pp−s k=1 bi Xk=1 Xj=1 ai jαj + p−s kβk with ai j, bi k ∈ C, we set k i b kβ for i = 1, . . . , s. We denote A = (ai j ) and we define the matrix MA by MA = (cid:18) I −A I (cid:19) −A Definition 5.1. Let R0 be the open subset of H 0(X, Ω1 X )s of those elements θ = (θ1, . . . , θs) for which det MA 6= 0 and denote by R the open subset of Ξ whose elements are the pairs r = (C, θ) fulfilling (i) det C 6= 0 and (ii) det MA 6= 0, that is, R = GL(s, C) × R0. The set R is the complementary of an affine real algebraic variety and its tangent space at (id, 0) is naturally identified to Ξ = End Cs ⊕ H 0(X, Ω1 X )s. The space R will be the parameter space of the versal family of tangential de- formations of (X, ). Moreover, in the case in which X is a Kahler manifold, that family will be versal at each point of R. Condition (ii) in the above definition implies that α1 r are linearly inde- pendent at each point. Hence the forms αi r generate a real subbundle Qr of T ∗X C. One has Qr ∩ ν∗F = 0, so there are well-defined complex valued smooth vector fields w1, . . . ws that are tangent to F and are determined by the conditions r, . . . , αs (11) r(wj ) = δi αi j and αi r(wj) = 0. 16 JAUME AMOR ´OS, M `ONICA MANJAR´IN, MARCEL NICOLAU We set (12) and we notice that (13) Then we define (14) wi = vi − s Xj=1 aj i vj for i = 1, . . . , s, hw1, . . . , wsiC = (cid:10) w1, . . . , ws(cid:11)C ∼= Cs. vr i = Cwi. In this situation one has Proposition 5.2. Let an element r = (C, θ) of R be given and set N 1,0 r = Qr ⊕ ν∗F 1,0. Then T ∗X C = N 1,0 and the almost complex structure so defined is integrable defining a complex structure Xr. Moreover, the vector fields vi r are holomorphic and induce a locally free Cs-action r on Xr, which is a tangen- tial deformation of . This construction defines a family (XR, R) of tangential deformations of (X, ) parametrized by R. r ⊕ N 1,0 r Proof. Condition T ∗X C = N 1,0 αi + θ and induces a complex structure Xr by Newlander-Nirenberg's theorem. are closed, the almost complex structure on X defined by N 1,0 is equivalent to det MA 6= 0. Since αi r = is integrable r ⊕ N 1,0 r r i Since v1, . . . , vs are holomorphic on X we see, using (13), that the vector fields It is also clear that wi are of type (1, 0) on It only remains to show that wi are in fact w1, . . . , ws commute to each other. Xr and so are the vector fields wi. holomorphic on Xr. Let (t1, . . . , ts, z1, . . . , zn−s) be local coordinates fulfilling (4). A local basis of vector fields of type (1, 0) on Xr is given by ∂ s s ∂ ∂z1 − (15) Xj=1 cj 1vj, . . . , ∂zn−s − (cid:8) w1, . . . , ws, l = Pn−s k=1 bj cj Xj=1 n−svj(cid:9), where cj kβk(∂/∂zl). Notice that the functions cj l are holomorphic on Xr and basic with respect to the foliation F , that is cj l only depend on the coordinates z1, . . . , zn−s. The vector fields wi are holomorphic if and only if the Lie brackets [wi, w] are of type (0, 1) on Xr for each local vector field w of type (1, 0) on Xr. But this is a straightforward computation using the local basis of T X 1,0 given above. (cid:3) r Remarks 5.3. (a) Notice that the almost complex structure on X by the decompo- sition T ∗X C = N 1,0 is real-analytic. r ⊕ N 1,0 r (b) The matrix C in r = (C, θ) corresponds to a choice of a basis of fundamental vector fields of the action. Hence, if r = (C, θ) and r′ = (C′, θ) with C′ 6= C then the complex manifolds Xr and Xr′ are identical, as well as the holomorphic foliations Fr and Fr′, although the actions r and r′ are different. (c) One can consider what, in principle, could be a more general class of tangen- tial deformations of the action, namely those determined by the global 1-forms αℓ on X defined as (16) αℓ = αℓ +X eℓ k, gℓ j, f ℓ jαj +X f j, hℓ k ℓ kβ +X gℓ jαj +X hℓ kβk, where the coefficients eℓ In fact, if these coeffi- cients are small enough then there are smooth vector fields uniquely defined by (vj) = 0. Moreover, if we denote by Q the real the conditions αℓ(vj) = δℓ subbundle of T ∗X C generated by the 1-forms αℓ, then N 1,0 = Q ⊕ ν∗F 1,0 defines k are complex numbers. j and α ℓ DEFORMATIONS OF K AHLER MANIFOLDS WITH VECTOR FIELDS 17 an integrable almost complex structure and the vector fields vj are holomorphic on the new complex manifold X. Notice however that this complex structure and the vector fields vj are the same as the ones determined by the 1-forms αℓ +X eℓ jαj +X f k ℓ kβ +X gℓ jαj. Furthermore, X endowed with the Cs-action defined by the vector fields vj can be identified with a suitable (Xr, r) in the family constructed above. More precisely, j) then X coincides with Xr, where r = if we denote by D the invertible matrix (δℓ (C, (θ1, . . . , θs)) is determined by C = (ci with m ai j . Therefore this construction does not provide more general deformations of (X, ). In fact it is proved in the following theorem that (XR, R) is the versal family of tangential deformations of (X, ). In particular the family is complete and it contains all the small tangential deformations of (X, ) up to isomorphism. j+gℓ m) = D−1 and θ jαj +P b j = P ci j = P ci = P ai mem i j and b k i kβ mf i (d) It follows from the above remark that the family (XR, R) does not depend j. Therefore that family is on the choice of the 1-forms αi fulfilling αi(vj ) = δi naturally associated to (X, ). Remark 5.4. Let ′ be a locally free Cs-action on a compact complex mani- fold X ′ and assume that (X ′, ′) coincides with a pair (Xr, r) in the family (XR, R). Then (X, ) coincides with an element of the family (X ′ R) associ- ated to (X ′, ′). Therefore the original action : G × X → X is also a tangential deformation of r : G × Xr → Xr for each r ∈ R. R, ′ Theorem 5.5. The family (XR, R) defined above is the versal family of tangential deformations of (X, ). Proof. Let κ : T0R → h1(X, T X) denote the Kodaira-Spencer map associated to the family (XR, R) at 0. Let (C = id + C′, θ) be a given element in Ξ close to (id, 0). Then (C′, θ) is an element of T0R = End Cs ⊕ H 0(X, Ω1 X )s, close to zero, h(Cs, Ω0,1(X, T F )). An that can be seen as a cocycle in C1 easy computation shows that κ(C′, θ) is just the cohomology class in H 1 Cs (X, T F ) of that cocycle. This implies that: (i) the restriction of κ to End Cs is just the X )s is a section of map τ in (9), and (ii) the restriction of κ to T0R0 ≡ H 0(X, Ω1 the morphism χ in (9). Hence κ is an isomorphism and the statement follows from Proposition 4.4. (cid:3) h(Cs, Ω0,0(X, T F )) ⊕ C0 Let KX denote the Kuranishi space of X, that is the parameter space of the versal family of deformations of the complex manifold X. Recall that KX is a (possibly singular) analytic space whose tangent space at the distinguished point 0 ∈ KX is naturally identified to H 1(X, ΘX ). The restriction XR0 to R0 of the family XR can be seen as a family of deformations of the complex structure of X. Hence there is a well defined forgetful map of germs of analytic spaces φ : (R0, 0) −→ (KX , 0). Moreover its tangent map d0φ at 0 coincides with the morphism H 1(X, T F ) → H 1(X, ΘX ) induced by the decomposition T X = T F ⊕ T G. More precisely, given an element kβk, its image d0φ(θ) is the vector θ = (θ1, . . . , θs) in T0R0 with θi = P ai valued (0, 1)-form jαj +P bi θ = X ¯θi ⊗ vi. 18 JAUME AMOR ´OS, M `ONICA MANJAR´IN, MARCEL NICOLAU Since holomorphic forms are closed and non exact, one can see using local coor- dinates fufilling (4) that θ defines a cohomology class in H 1(X, ΘX ) which is not trivial unless θi = 0 for all i = 1, . . . s. This implies that φ is an embedding proving the following statement. Theorem 5.6. Let X be a compact manifold in the class C such that s = dimC TX = dimC hX /h1 X 6= 0. Then the Kuranishi space KX of X contains a smooth subspace of dimension s · b1(X)/2. More precisely, the canonical morphism H 1(X, T F ) ≡ H 1(X, OX )s → H 1(X, ΘX) is injective and its image are unobstructed infinitesimal deformations of the complex structure of X. We shall make distinction of two particular types of tangential deformations defined as follows Definition 5.7. Let r = (C, θ) ∈ R be given and set θ = (θ1, . . . , θs). We say that the associated tangential deformation (Xr, r) is (i) a purely tangential deformation, or t-deformation, of (X, ) if θi ∈ hα1, . . . , αsi, i.e. if the coefficients bi k in (10) are all zero, (ii) a basic deformation, or b-deformation, of (X, ) if θi ∈ hβ1, . . . , βp−si, i.e. if the coefficients ai j in (10) are all zero. Remarks 5.8. (a) The holomorphic and invariant foliation G, which is transversal and complementary to F , remains holomorphic and invariant for each t-deformation of (X, ). On the other hand the foliation F remains holomorphic for each b- deformation of (X, ). (b) If (Xr, r) is a t-deformation (resp. b-deformation) of (X, ) then (X, ) is a t-deformation (resp. b-deformation) of (Xr, r). Remark also that each small tangential deformation (Xr′ , r′) of (X, ) can be seen as a b-deformation of a t-deformation (Xr, r) of (X, ). More precisely, if then the manifold Xr jαj is a t-deformation of X. Moreover, jαj + P b Xr′ is determined by 1-forms αi = αi + P ai determined by the 1-forms αi = αi + P ai since αi = αi +P b Notice that, as X is a tangential deformation of Xr, the previous argument also proves that (Xr′, r′) can be seen as a t-deformation of a b-deformation of (X, ). the manifold Xr′ is a b-deformation of Xr. i kβ i kβ k k As it is remarked in the proof of Proposition 5.2, local holomorphic functions on X that are basic with respect to F are also holomorphic on Xr and basic with respect to Fr. By definition, the global 1-forms αi r are of type (1, 0) on Xr and they are also closed by construction. Hence αi r are holomorphic forms on Xr. It follows that, for each r ∈ R, one has h1,0(Xr) = h1,0(X), h0,1(Xr) = h0,1(X) and b1(X) = h1,0(Xr) + h0,1(Xr). In particular, the space of holomorphic 1-forms on Xr can be decomposed as H 0(Xr, Ω1 (17) Proposition 5.9. Let aX be an Abelian algebra fulfilling hX = aX ⊕ h1 Xr be a tangential deformation of X. ri ⊕ hβ1, . . . , βp−si. Xr ) = hα1 r, . . . , αs X and let (i) If aX is in the center of hX then h1 (ii) If Xr is a b-deformation of X then aX ⊂ hXr and hXr = aX ⊕ h1 If both conditions are fulfilled then hXr = hX . Proof. (i) Let w ∈ h1 Since [aX, h1 X be given. Notice that αi(w) = αi X ] = 0, w is written locally as Xr = h1 X . Xr . r(w) = 0 for i = 1, . . . , s. w = n−s Xk=1 f k(z) ∂ ∂zk , DEFORMATIONS OF K AHLER MANIFOLDS WITH VECTOR FIELDS 19 where f k are holomorphic functions on z1, . . . , zn−s. It follows that w is of type (1, 0) on Xr. Now the same argument used in the proof of Proposition 5.2 shows that w is in fact holomorphic in Xr. The original manifold X is a tangential deformation of Xr (cf. Remark 5.4), so h1 Xr does not contain any additional tangent fields besides those arising in X. (ii) If Xr is a b-deformation of X then the elements v ∈ aX are still holomorphic (cid:3) on Xr, as it follows from the identities (11) to (14). We discuss now the tangential deformations of the examples of Kahler manifolds introduced in Section 2. Example 5.10 (Deformations of tori). If X is a complex torus Ts = Cs/Λ then H 0(Ts, Ω1 Ts)s can be identified with the dual space of H 1(Ts, ΘTs). In that case the space R is just the product of GL(s, C) with the versal space of deformations of the complex manifold Ts as it is described by Kodaira and Spencer in [24]. Example 5.11 (Deformations of principal torus bundles). Assume that X is the total space of a principal torus bundle Ts → X → B where Ts = Cs/Λ. As it was already noticed (cf. Example 2.3), when X is Kahler then B is also a Kahler manifold and the bundle is flat. The possible choices of the s-tuple (α1, . . . , αs) correspond exactly to the different flat connections on that bundle. In this situation the space of basic deformations of the action is naturally identi- B → B). The exact sequence of sheaves 0 → Λ → Os B)s ∼= H 1(B, Os fied to H 0(B, Ω1 Os B/Λ → 0 over B induces the exact cohomology sequence . . . −→ H 1(B, Λ) −→ H 1(B, Os B) −→ H 1(B, Os B/Λ) τ−→ H 2(B, Λ) −→ . . . It follows that H 1(B, Os B) is maped onto ker τ , which is the space of (isomorphism classes of) topologically trivial Cs/Λ-principal bundles over B. Replacing B by a finite (unramified) covering we can suppose that H 2(B, Λ) has no torsion. In that case each topologically trivial principal torus bundle over B can be obtained from X by a suitable basic deformation of the action and, in particular, the trivial bundle is a tangential deformation of (X, ). In general, each tangential deformation of (X, ) is a principal torus bundle over B. Purely tangential deformations r ∈ R just deform the fiber Cs/Λr. Example 5.12 (Deformations of suspensions). Assume that X is a suspension over a torus Ts = Cs/Λ, that is X = F ×Λ Cs. It follows from Remark 5.8 (a) that each purely tangential deformation of X is still a suspension, although this is not true for general tangential deformations of X. If X is a suspension one has H 0(X, Ω1 F ). Therefore if b1(F ) = 0, or what is equivalent b1(X) = 2s, then X has no basic deformations and each tangential deformation of X is again a suspension. X/F ) = H 0(F, Ω1 We end this section stating several general properties of tangential deformations of compact manifolds of the class C endowed with a locally free Cs-action. If X → X is an unramified covering then X is also endowed with a locally free Cs-action in a natural way. Notice however that the maximal rank of such an action on X can increase, that is dimC h X /h1 ≥ s. The following statement follows X straightforward from the definitions. Proposition 5.13. A finite covering X ′ r of a tangential deformation Xr of X is a tangential deformation of a finite covering X ′ of X. Conversely, a tangential deformation X ′ r of a finite covering X ′ of X is a finite covering of a tangential deformation Xr of X. 20 JAUME AMOR ´OS, M `ONICA MANJAR´IN, MARCEL NICOLAU Proposition 5.14. Let X be a geometrically ruled manifold in the class C. Then Xr is a geometrically ruled compact complex manifold for each r ∈ R. Proof. Let X be a compact manifold in the class C endowed with a locally free Cs-action , and assume that X is the total space of a locally trivial fiber bundle, ϕ : X → B, with fiber CP k and defined by a cocyle {gij} in H 1(B, PGL(k+1, OB)). Then B is also in the class C. Since automorphisms of X close to the identity must preserve the fibration (cf. [6]) each fundamental vector field v of the action projects into a non vanishing vector field v on B. These vector fields define a locally free Cs-action on B and the projection ϕ is equivariant with respect to the Cs-actions. Notice that every global holomorphic 1-form on X vanishes on the fibers of ϕ. Hence ϕ∗ : H 0(B, Ω1 X ) is an isomorphism. Using this identification, one can associate, to each tangential deformation Xr given by a pair r = (C, θ), a tangential deformation Br defined by the same expression (10). Then the projection ϕ : Xr → Br is holomorphic. B) → H 0(X, Ω1 Now, as the Cs-action is transverse and preserves the fibration, the cocycle {gij} can be chosen to be constant along the orbits of the Cs-action on B. This implies that the matrix-valued functions gij are still holomorphic on Br and therefore they define Xr as a geometrically ruled manifold over Br. (cid:3) Proposition 5.15. A tangential deformation Xr of X uniquely determines a tan- gential deformation Alb(X)r of the Albanese torus Alb(X) of X such that Alb(X)r = Alb(Xr), i.e. Alb(X)r is the Albanese torus of Xr. Proof. The locally free Cs-action on X induces a locally free Cs-action on the corresponding Albanese torus Alb(X) and the natural map φ : X → Alb(X) is equivariant with respect to these actions. Since the map φ induces an isomorphism φ∗ : H 0(Alb(X), Ω1 X ), there is a natural correspondence, as the one considered in the proof of the above proposition, associating to each tangential deformation Xr of X a tangential deformation Alb(X)r of Alb(X). It follows from the definitions that Alb(X)r = Alb(Xr). (cid:3) Alb(X)) → H 0(X, Ω1 Proposition 5.16. The manifolds X and Xr have the same plurigenera for all r ∈ R. In particular the Kodaira dimension kod(Xr) of Xr is equal to kod(X) for each r ∈ R. Proof. Let (t1, . . . , ts, z1, . . . , zn−s) be local coordinates of X fulfilling (4) and let r be the global holomorphic 1-forms on X defined by (10). Notice that zj are αi also holomorphic functions on Xr. A local section of K ⊗m is of the form f (t, z)ηm Xr where η = α1 ∧ · · · ∧ αs ∧ dz1 and f = f (t, z) is a holomorphic function. Nakamura and Ueno have proved that, for each compact complex manifold Y , the group Aut0 Y ) (cf. [33, Corollary 14.8]). Applying this result to the Cs-action we see that f is necessarily a basic function, i.e. f = f (z) only depends on the coordinates (z1, . . . , zn−s) transverse to F . Therefore H 0(Xr, K ⊗m ) is naturally identified to the space of basic global sections of the line Xr bundle (det ν∗Fr)m and this space is independent of r ∈ R. (cid:3) C(Y ) acts trivially on H 0(Y, K m i ∧ · · · ∧ dzn−s i If X is a Kahler manifold then Kodaira's stability theorem implies that Xr is also Kahler for every r in a neighborhood of (id, 0) in R, and we will prove the following result, stating that the class of Kahler manifolds is stable under tangential deformations. In sharp contrast the class C is not stable under small deformations. We do not know if tangential deformations of Fujiki manifolds still belong to the class C. Theorem 5.17. Assume that X is a Kahler manifold, then Xr are also Kahler manifolds for each r ∈ R. DEFORMATIONS OF K AHLER MANIFOLDS WITH VECTOR FIELDS 21 We postpone the proof of this theorem to Section 7, where the assertion will be a corollary of general results on the structure of such manifolds. Nevertheless, accepting that the theorem has already been proved we can state the following two propositions: Proposition 5.18. Assume that X and Xr belong to the class C. Then (i) hk,0(X) = hk,0(Xr) and h0,k(X) = h0,k(Xr) for each k, (ii) h1,1(X) = h1,1(Xr). If X is a Kahler manifold the above identities hold for each r ∈ R. Proof. Lemma 2.10 implies that there is a well defined C-linear map from H 0(X, Ωk to H 0(Xr, Ωk Xr ) determined by the condition X ) γ = X0≤I≤ℓ r = αi1 αI ∧ γI 7−→ γr = X0≤I≤ℓ r = αi + θ iI r ∧ · · · ∧ α r αI r ∧ γI. i where ℓ = min(k, s), αI and γI are global basic forms. Clearly this map is in fact an isomorphism. Preservation of b2, h2,0, h0,2 implies that of h1,1. (cid:3) , αi Proposition 5.19. Assume that X is a Kahler manifold. Then the family (XR, R) is the versal family of (Xr, r) at each point r ∈ R. Proof. In that case Xr is a Kahler manifold for each r ∈ R and H 1(Xr, OXr ) is isomorphic to H 0(Xr, Ω1 Xr ). Moreover, the dimension of these vector spaces does not change with r ∈ R. Hence the argument used to prove Theorem 5.5 also applies to each Xr with r ∈ R. (cid:3) 6. The approximation theorem In this section we prove a key result for describing the structure of compact complex manifolds X in the class C endowed with a locally free holomorphic Cs- action. Namely, we prove that such a manifold can be approximated by tangential deformations Xǫ of X that are suspensions over a torus of dimension s, i.e. Xǫ belonging to the class of manifolds described in Example 2.4. We shall make use of the following criterium, which is valid for arbitrary compact complex manifolds. Lemma 6.1. Let X be a compact complex manifold endowed with a locally free action of Cs and let v1, . . . , vs be a basis of fundamental vector fields of the action. Then X is a suspension over a torus T = Cs/Λ, i.e. X = F ×Λ Cs for a suitable representation Λ → AutC(F ), if and only if one can choose holomorphic 1-forms αj on X fulfilling αj (vi) = δj Λ = {(Zγ i and such that the set of periods of α1, . . . , αs α1, . . . ,Zγ αs) γ ∈ H1(X, Z)} is a lattice of Cs. Proof. For each i = 1, . . . , s let f i be a multivalued holomorphic function such that αi = df i. If the set of periods Λ is a lattice of Cs then f = (f 1, . . . , f s) is a well-defined holomorphic submersion from X onto Cs/Λ. The converse is clear. (cid:3) Remark 6.2. The fiber space F in the above theorem can always be taken connected. Indeed, if the fibers of the submersion f : X → T are not connected, one can 22 JAUME AMOR ´OS, M `ONICA MANJAR´IN, MARCEL NICOLAU consider the Stein factorization of f X ✲p N ❅ ❅ ❅❘f π ❄ T Since the map p is finite and f is a submersion, the map π is necessarily an un- branched cover, thus N is also a torus that covers T . The map p is then the sought suspension. The following result is an improvement of Theorem 2.16 in [30]. For our purposes here, an essential point of the theorem is the fact that the manifold Xǫ, which approximates X and is a suspension over a torus, can be taken to be a basic deformation of X. Theorem 6.3. Let X be a compact complex manifold in the class C endowed with a locally free Cs-action. There is an arbitrarily small b-deformation Xǫ of X which is a suspension over a complex torus T = Cs/Λ. More precisely, there is a connected compact complex manifold F and a group representation ρ : π1(T ) = Λ → AutC(F ) such that Xǫ is the suspension of ρ. Proof. First, we claim that there exist arbitrarily small holomorphic 1-forms θi, ηi on X such that, if we set αi = αi + θ + ηi, then i Λ = (cid:8)(cid:0)Zγ α1, . . . ,Zγ αs(cid:1) γ ∈ H1(X, Z)(cid:9) is a lattice of Cs. The proof is essentially the same as the one given in [30, p. 490]. Let {α1, ..., αs, β1, ..., βp−s} be the basis of H 0(X, Ω1 X ) considered in (5). We decompose θi = θi t + θi b = X ai jαj +X bi kβk, ηi = ηi t + ηi b = X ci jαj +X di kβk. i It follows from Remark 5.3 (c) that the 1-forms αi + θ t determine a complex manifold endowed with a Cs-action which is isomorphic to a suitable (Xr′ , r′) close to (X, ). Since αi are holomorphic 1-forms on Xr′, Lemma 6.1 implies that the manifold Xr′ is a suspension over a torus. + ηi As it has been noticed in Remark 5.8 (b), we can see (Xr′, r′) as a t-deformation of a suitable (Xr, r) close to (X, ) with the property that (Xr, r) is a b- deformation of (X, ). But Xr is a suspension over a torus as t-deformations of suspensions are again suspensions (cf. Example 5.12). Hence (Xr, r) fulfills the required conditions as Remark 6.2 assures that the fibre F of the suspension can always be taken connected. (cid:3) It follows from the above theorem that there is an exact sequence of groups 1 → π1(F ) → π1(X) → Z2s → 1. Remarks 6.4. (a) It follows from Remark 5.3 (a) that Xǫ is real-analitically isomor- phic to the manifold X. More precisely, as the two complex manifolds X and Xǫ have the same underlying real-analytic structure, the identity map id : X → Xǫ is real-analytic. (b) Assume that X is a suspension over a torus T = Cs/Λ, i.e. X = F ×Λ Cs. If T ′ is a torus isogenous to T then T ′ = Cs/Λ′, where Λ′ is a sublattice of Λ, and there is an Abelian finite covering X ′ of X which is a suspension over T ′. From this fact one deduces easily that if X is a suspension over an s-dimensional torus, DEFORMATIONS OF K AHLER MANIFOLDS WITH VECTOR FIELDS 23 where s = dimC TX = dimC hX /h1 is a suspension over TX = Φ(Aut0 X , then X has an Abelian finite covering X ′ that C(X)) ⊂ Alb(X). (c) Let X be a Kahler manifold in the hypothesis of the above theorem. Then Xǫ is the suspension F ×Λ Cs of a representation ρ : Λ → AutC(F ). Since Xǫ is a Kahler manifold, there is a sublattice Λ′ of Λ such that ρ(Λ′) ⊂ Aut0 C(F ) (cf. 2.4). ǫ = F ×Λ′ Cs is an Abelian finite covering of Xǫ. Notice Then the suspension X ′ ǫ is topologically isomorphic to the product F × T ′, where T ′ = Cs/Λ′, and that X ′ that X ′ ǫ can be seen as a b-deformation of a finite covering X ′ of X (cf. 5.13 and 5.4). (d) Any (unramified) finite covering X ′ of X is naturally endowed with a Cs- action but the maximal rank can increase, that is dimC hX ′ /h1 X ′ ≥ dimC hX /h1 X . Examples where this inequality is strict can be constructed by considering suspen- sions of bielliptic surfaces. A more precise description of the fiber manifold F and of the representation ρ : Λ → AutC(F ) in the above theorem can be obtained by considering (unram- ified) finite coverings X ′ of the manifold X. This is the content of the following proposition. It can be seen as an improvement of the approximation theorem and will play a key role in the structure theorems of the next Section. X > 0. There is a finite Abelian covering X ′ Proposition 6.5. Let X be a compact Kahler manifold such that s = dimC TX = dimC hX /h1 ǫ of an arbitrarily small b- deformation Xǫ of X, which is the suspension over a torus T = Cs′ /Λ, of dimension s′ ≥ s, associated to a representation ρ : Λ → AutC(F ) with the following properties: F = 0, i.e. F has no (i) F is a connected compact Kahler manifold with hF /h1 non-singular holomorphic vector fields, (ii) ρ(Λ) ⊂ Aut0 C(F ). Remark 6.6. The manifold X ′ Abelian covering X ′ of X. This follows from Proposition 5.13. ǫ can also be seen as a (small) b-deformation of a finite Proof. Let aX be an central Abelian subalgebra of hX fulfilling hX = aX ⊕ h1 X . It defines a locally free Cs-action on X and, by the above Theorem 6.3 there is an Cs associated arbitrarily small b-deformation Xǫ of X which is the suspension F1×Λ1 to a representation ρ1 : Λ1 → AutC(F1). Notice here that, applying Proposition 5.9, we have aX ⊂ hXǫ . Therefore we can chose aXǫ = aX as the complementary of h1 Xǫ in hXǫ . 1 ǫ = F1 ×Λ′ ǫ /Γ1 where Γ1 = Λ1/Λ′ As in Remark 6.4 (c) above, we consider the sublattice Λ′ C(F1)) of Λ1. Then the suspension X 1 Cs is an Abelian finite covering of Xǫ = X 1 1. More precisely, each vector field on Xǫ lifts to a unique vector field on X 1 ǫ and therefore we can identify aXǫ = aX to an Abelian subalgebra of hX 1 ǫ that we still denote by aX . Then there are vi ∈ aX such that the monodromies ρ1(γ) for the set of elements γ ∈ Λ1 generating Γ1 lie in hγ1, . . . , γki, where γi = exp1(vi). 1 = ρ−1 1 (Aut0 If F1 has no vector fields without zeros we are done. So, assume that hF1 /h1 F1 6= 0. In that case we choose a subalgebra aF1 of the center of hF1 such that hF1 = aF1 ⊕h1 F1 (cf. Proposition 1.6). Since ρ1(Λ′ C(F1) each element w of aF1 is ρ1(Λ′ 1)-invariant. This implies that w is the restriction of a globally defined holomorphic vector field w on X 1 ǫ without zeros and commuting to aX . More Cs of the vector field (w, 0) on the product precisely, w is the projection over F1 ×Λ1 manifold F1 × Cs. Remark that, by construction, aX 1 ǫ = aX ⊕ aF1 is Abelian and 1) is contained in Aut0 24 JAUME AMOR ´OS, M `ONICA MANJAR´IN, MARCEL NICOLAU ǫ ⊕ h1 that hX 1 X 1 ǫ beginning with X 1 ǫ = aX 1 ǫ instead of X. . Denote r = dim aX 1 ǫ > s, and repeat the above construction ǫ ǫ There is a small b-deformation X 1 Cr over a torus Cr/Λ2, associated to a representation ρ2 : Λ2 → AutC(F2). We see, using Proposition 5.13, that X 1 δ is a finite Abelian covering of a b-deformation Xδ of X. ⊕ h1 And, as above, aX 1 . In X 1 δ particular, γi = exp1(vi) can be seen as holomorphic transformations of X 1 δ . ǫ which is the suspension F2 ×Λ2 is an Abelian subalgebra of hX 1 with hX 1 δ of X 1 = aX 1 2 (Aut0 2 = ρ−1 We set Λ′ C(F2)) and we define X 2 Cr. Then X 2 δ is a finite Abelian covering of X 1 2. Also as above, It follows in particular that the aX 1 automorphisms γi = exp1(vi) for vi ∈ aX can be lifted as biholomorphisms γi of X 2 δ = F2 ×Λ′ δ /Γ2 where Γ2 = Λ2/Λ′ ǫ = aX ⊕ aF1 is naturally included in hX 2 δ . This already implies that the composition δ = X 2 δ . δ δ 2 X 2 δ −→ X 1 δ −→ Xδ is a regular covering. Moreover, we can realize the monodromies ρ2(γ) for a set of elements in Λ2 generating Γ2 within a subgroup of automorphisms hγ1, . . . , γmi, where γj = exp1(wj ) and wj ∈ aX 1 ǫ . Since all the biholomorphisms γi and γj are exponentials of vector fields belonging to the same Abelian algebra, they commute to each other proving that the regular covering X 2 δ → Xδ is in fact Abelian. If F2 does not have non-singular vector fields we have done. If this is not the case we repeat again the above construction starting with X 2 δ . Since the rank of locally free actions is bounded by the dimension n of X this procedure ends after a finite number of steps. (cid:3) X , with aX central, and X ′ Remark 6.7. Let hX = aX ⊕ h1 be as in the above Proposition. Along the proof it is shown that the Lie algebra decomposition ǫ ⊕ h1 ǫ is generated by the can be taken with the property that aX ′ hX ′ X ′ ǫ ǫ of linear vector fields on Cs′ projection over X ′ ǫ of aX = aXǫ ǫ . is included in aX ′ and that the lift to X ′ ǫ = aX ′ ǫ = F ×Λ Cs′ We are ready now to establish our refinement of Calabi's Theorem 0.2, which was originally stated by Calabi in [9] with solvable Galois group Γ,. It was proved by Bogomolov in [7] for complex projective manifold X with Abelian Γ and extended by Beauville [3] who showed that for a suitable finite Γ the factor F is a simply- connected manifold with special holonomy. Corollary 6.8. Let X be a compact Kahler manifold with c1(X) = 0. Then X admits a finite, Abelian ´etale covering X ′ = F × T with Galois group Γ, which is the product of a complex torus T of dimension b1(X ′)/2 and a compact Kahler manifold F with c1(F ) = 0 and b1(F ) = 0. Proof. If a compact Kahler manifold has a trivial first Chern class c1(X) = 0 then the natural coupling between hX and H 0(M, Ω1 X ) is non degenerate (cf. [27]). This implies that Aut0 C(X) is a complex torus, and in particular hX = aX . Therefore, in that case there is a unique locally free Cs action of maximal rank s = dim aX = b1(X)/2. The non degeneracy of the above coupling also implies that the foliation F determined by the action does not have basic 1-forms. Hence (X, ) does not have non-trivial b-deformations and the approximation theorem says in that case that X is already a suspension F ×Λ Cs. Now the statement can be proved as the above theorem just remarking that, at each stage, the fiber manifold F also fulfills c1(F ) = 0 (cf. 2.9) and therefore there is no need to consider tangential deformations. (cid:3) DEFORMATIONS OF K AHLER MANIFOLDS WITH VECTOR FIELDS 25 7. Structure of Kahler manifolds with non vanishing vector fields With the exception of Proposition 7.1, we assume from now on that the compact complex manifold X endowed with a locally free Cs-action is of Kahler type. As we pointed out, examples of such kind of manifolds are given by (i) complex tori, (ii) flat principal torus bundles over a Kahler manifold and (iii) suspensions over a torus T = Cs/Λ with fiber a Kahler manifold F and monodromy ρ : Λ → Aut0 C(F ). A natural question is whether that list of examples covers all the possible Kahler manifolds X with locally free Abelian actions. The answer is positive, up to a finite covering, if the action has codimension one, i.e. if s = n − 1 (as it was proved by F. Bosio in [8]), if c1(M ) = 0 (Corollary 6.8) or if the manifold X is projective (as it is proved in Section 3). In this section we show that, for a general Kahler manifold, the answer is also positive but up to a finite covering and up to a (small) tangential deformation of the manifold. The first two Propositions are precise statements of this fact. We also show (Examples (7.6) and (7.7)) that tangential deformations cannot be avoided. Afterwards, combining tangential deformations with deformations of represen- tations, we are able to give a general structure theorem for Kahler manifolds with locally free Abelian actions (Theorem (7.11)). From it we deduce that such a man- ifold X has a finite covering which is a (non necessarily small) deformation of a product F ×T . Also as a corollary of that construction, we prove that the manifolds Xr in the versal family XR of tangential deformations of X are Kahler manifolds for each r ∈ R. Assume that the Abelian group defining the action on X is a complex torus. It was proved by Fujiki [17] and Lieberman [29] that, under this hypothesis, the manifold X is a Seifert fibration with the torus as the typical fiber. The following statement is an slight improvement of that result. Proposition 7.1. Let X be a compact manifold in the class C endowed with a locally free Cs-action. Assume that all the orbits of the action are closed. Then X is the quotient by an Abelian finite group of a flat torus bundle X ′ T → X ′ ↓ F where T is a complex torus of dimension s and F is a compact manifold in the class C. Furthermore, there is an arbitrarily small b-deformation Xǫ of X such that Xǫ is a finite quotient of the product manifold F × T . Proof. By Theorem 6.3, there is a small b-deformation Xǫ of X that is the suspen- sion F ×Λ Cs over a torus T = Cs/Λ defined by a representation ρ : Λ = π1(T ) → AutC(F ). The orbits of the action on Xǫ are still closed and all orbits are finite coverings of T . In particular all the elements of H = ρ(Λ) are of finite order and [22]). Hence, the Cs-action on this implies that the group H itself is finite (cf. Xǫ defines a Seifert fibration over the complex orbifold F = F/H. Notice that the natural projection F → F is a ramified covering. Set T0 = Cs/Λ0, where Λ0 = ker ρ. The pull-back of the fibration Xǫ → T by the covering map T0 → T has total space an Abelian finite covering X ′ ǫ of Xǫ naturally identified to the product F × T0. Indeed, since H = Λ/Λ0 we have Xǫ = F ×H T0. ǫ is a b-deformation of X ′ (cf. Proposition 5.13). Then the covering space X ′ has the required properties. (cid:3) Finally, notice that there is an Abelian finite covering X ′ of X such that X ′ 26 JAUME AMOR ´OS, M `ONICA MANJAR´IN, MARCEL NICOLAU Remark 7.2. It follows from the above proof that, under the hypothesis of the theorem, X is a Seifert fibration over a good orbifold, i.e. an orbifold which is the quotient of a manifold by a finite group. Combining previous results we can state Proposition 7.3. Let X be a compact Kahler manifold endowed with a locally free Cs-action, where s = dimC hX /h1 ǫ of a small b-deformation Xǫ of X, a torus T of dimension s′ ≥ s and a compact Kahler manifold F without non vanishing vector fields in such a way that X . There is a finite Abelian covering X ′ (i) if h1 (ii) if h1 X = 0 then X ′ X 6= 0 then X ′ ǫ = F × T , ǫ is a suspension over T with fiber F , monodromy in Aut0 C(F ), and hF = h1 F 6= 0. In particular F and X are uniruled manifolds. In both cases kod(F ) = kod(X). Proof. Recall that h1 C(X) is a torus and in that case the orbits of the action are closed. Hence case (i) follows from Proposition 6.5 and Proposition 7.1. X = 0 implies that Aut0 Without loss of generality we can assume that the subalgebra a has been chosen in the center of hX . Then the space of vector fields with zeros is the same for every tangential deformation of the manifold (cf. Proposition 5.9). Therefore, we can apply again Proposition 6.5 to case (ii), and we have h1 F 6= 0, as vector fields with zeros are always tangent to fibers in the case of suspensions. In particular the manifolds X and F are both uniruled by virtue of Theorem 1.8. Xǫ 6= 0 and also h1 Finally, the last statement follows directly from Proposition 5.16, Corollary 2.9 and from the fact that taking an (unramified) finite covering does not change the Kodaira dimension. (cid:3) Remark 7.4. If h1 So there are the following possibilities according to the Kodaira dimension of X: X 6= 0 then X is a uniruled manifold and therefore kod(X) = −∞. (a) kod(X) ≥ 0. Then we are in case (i) in the above proposition and the manifold F fulfills the stronger condition hF = 0, because kod(F ) ≥ 0. (b) kod(X) = −∞. Since the Cs-action can have closed orbits both possibilities, (i) and (ii) in the above proposition, can occur. We are now ready to complete the proof of Theorem 0.3: Proof of Theorem 0.3. The algebraic case follows from Theorem 3.2. The general statement follows directly from Proposition 7.3, Remark 7.4 and from the observa- tion that, when kod(X) ≥ 0, the connected component of the identity Aut0 C(F ) of the group of automorphisms of the fiber manifold F reduces to the identity. (cid:3) We give here two examples showing that, in order to obtain a complete classifica- tion up to finite coverings of Kahler manifolds with non-singular vector fields, one cannot avoid the use of tangential deformations. More precisely, we exhibit exam- ples of Kahler manifolds X endowed with a locally free C-action with the property that neither X nor any finite covering X ′ of X are suspensions or torus bundles. We will use the following characterization of manifolds that are suspensions over a complex torus in terms of their Albanese torus. Proposition 7.5. Let X be a compact Kahler manifold with s = dimC TX = dimC hX /h1 X > 0. The manifold X is a suspension over an s-dimensional torus Ts with a connected fiber if and only if the Albanese torus Alb(X) splits, up to isogeny, as a product TX × T ′ DEFORMATIONS OF K AHLER MANIFOLDS WITH VECTOR FIELDS 27 Proof. Assume that p : X → Ts is a suspension. The universal property of Alb(X) gives a commutative diagram X ✲φ Alb(X) ❅ ❅ p ❅ ❅❅❘ ψ ❄ Ts Then the restriction of ψ to TX is a surjective group homomorphism with finite kernel and Alb(X) is isogenous to TX × ker ψ. Conversely, if Alb(X) is isogenous to TX × T ′ there is a surjective group homo- morphism ψ : Alb(X) → TX . The composition ψ ◦ φ : X → TX is an immersion when restricted to the orbits of the Cs-action, hence it defines X as a suspension over TX . Finally, by replacing TX by a finite covering of it, we can assume that the projection ψ ◦ φ has connected fibers (cf. Remark 6.2). (cid:3) Example 7.6 (Examples that are not suspensions). Let Cg be a compact Riemann surface of genus g > 1, and Π its g × 2g period matrix for fixed basis of H1(Cg, Z) and H 0(Cg, Ω1 Cg ). If Λ = Π · Z2g is the lattice of Cg generated by the columns of the matrix Π then Alb(Cg) = Cg/Λ. Let E = C/(Z ⊕ Zτ ) be a fixed elliptic curve. The complex tori T that are extensions of the type (18) 0 → E → T → Alb(Cg) → 0 are parametrized by the (g + 1) × (2g + 2) complex matrices of the form 1 τ 0 0 ... ... 0 0   e1 . . . e2g Π .   τ(cid:1) and Π, Denote by Te the extension defined by the fixed period matrices (cid:0)1 and a given vector e = (e1, . . . , e2g) ∈ C2g. Then Te is the Albanese torus of an elliptic surface Se, which is the analitically locally trivial fibration E → Se ↓ Cg having as monodromy automorphisms the translations by e1, . . . e2g ∈ E = C/(Z ⊕ Zτ ) along the selected basis of H1(Cg, Z). That is, Se is the suspension over Cg associated to the representation pe : π1(Cg) → E induced by the above translations. As the monodromy is isotopic to the identity, Se has even first Betti number and therefore it is a Kahler surface. Notice also that the Albanese morphism φSe : Se → Te = Alb(Se) maps Se isomorphically onto its image φSe (Se) ⊂ Te and that the map φSe identifies E with the subtorus TSe = Φ(Aut0 C(Se)) of Alb(Se) canonically associated to Se. The elliptic surfaces Se are endowed with a locally free holomorphic C-action given by the constant tangent vector fields of the fiber E, as they are preserved by the monodromy automorphisms of the suspension Se → Cg. By Proposition 7.5, a necessary condition for Se to be a suspension over a (one-dimensional) torus is that its Albanese torus Te splits, up to isogeny, as a product E × Alb(Cg). Yet it is known that the exact sequence (18) splits up to isogeny for only countably many choices modulo isomorphism equivalence of extension data, while the set of 28 JAUME AMOR ´OS, M `ONICA MANJAR´IN, MARCEL NICOLAU all extensions modulo isomorphism equivalence is a g-dimensional complex analytic variety (cf. [4]). Therefore, a generic choice of e ∈ C2g yields a Kahler elliptic surface Se, which has a locally free C-action along its fibers, but is not a suspension over a torus (although it is a suspension over Cg by construction). Suspensions over tori that are projective manifolds can be characterized in terms of their Albanese torus as it is stated in Proposition 8.6. In particular and by virtue of Poincar´e's reducibility theorem, the manifolds Se so constructed are not projective and their Albanese tori Alb(Se) are not Abelian varieties. Finally, consider a finite covering Se of Se. Necessarily, Se is also a E-principal fi- bration over a Riemann surface Cg, where E and Cg are finite coverings of E and Cg respectively. If Se were a suspension over a torus then it would be a finite quotient of a product of two Riemann surfaces and hence a projective manifold. In that case Alb( Se) would be an Abelian variety and, as the natural map Alb( Se) → Alb(Se) is surjective, Alb(Se) would also be an Abelian variety leading to a contradiction. Example 7.7 (Examples that are neither suspensions nor torus bundles). Using the above example we construct now a compact Kahler 3-fold X endowed with a locally free C-action in such a way that neither X nor any of its finite coverings are a suspension over a torus or a torus bundle. Choose a closed Riemann surface Cg, an elliptic curve E and a Kahler elliptic surface Se which is not a suspension over a torus, as in Example 7.6. Let L ∈ Pic0(E) be a flat line bundle over E which is non-torsion. Sum L with the trivial local system to get a rank 2 one, V = C ⊕ L, and let Y = PE(V ) be the ruled surface over E obtained by projectivizing the holomorphic rank 2 vector bundle defined by V ⊗ OE. The ruled surface Y admits a locally free holomorphic C-action, given by parallel transport along the flat connection on V , with the property that the projection Y → E is equivariant when we consider on E the natural C-action. It induces a group morphism ν : C → Aut0 C(Y ). The monodromy morphism pe : π1(Cg) → E defining the elliptic surface Se of Example 7.6 can be lifted to a morphism pe : π1(Cg) → C from π1(Cg) to the universal covering C of E. Then the representation π1(Cg) pe−→ C ν−→ Aut0 C(Y ) defines an analytically locally trivial fibration X which is a suspension over Cg with fiber Y Y −→ X ↓ Cg and it follows from Blanchard's criterium [6, Th´eor`eme principal II] that X is a Kahler manifold. Notice that the 3-fold X is naturally endowed with a locally free holomorphic C-action, coming from the flat parallel transport of Y over E, which is invariant under translations in E and may therefore be glued under the monodromy ν ◦ pe. Each holomorphic vector field on Y projects over E, hence there is a natural surjective morphism η : Aut0 C(Y ) → Aut0 C(E) = E fulfilling pe = η ◦ ν ◦ pe DEFORMATIONS OF K AHLER MANIFOLDS WITH VECTOR FIELDS 29 so we may glue the projections Y → E to get a commutative diagram of analytically locally trivial fibrations X −→ Se ց ↓ Cg The fibration X → Se has fiber CP 1, so it induces an isomorphism Alb(X) ∼= Alb(Se) and an identification of the Albanese images φX (X) = φ(Se). Therefore the subtorus E ⊂ Alb(X) does not have a complementary subtorus, even up to isogeny, and X is not a suspension. In particular the manifolds X and Alb(X) are not algebraic. Moreover, two choices of local systems V = C ⊕ L, V ′ = C ⊕ L′ yield isomor- phic ruled surfaces Y and Y ′ = PE(V ′) if and only if L′ ∼= L⊗±1 (cf. [2, III.7]). Therefore, the choice of a non-torsion line bundle L assures us that the orbits of the C-action on Y do not cover E with finite degree, and this implies that neither Y nor X can have the structure of a torus bundle. We remark finally that each unramified finite covering of X has the same prop- erties. Indeed, a finite covering X of X is necessarily an analytically locally trivial fibration over a Riemann surface Cg with fiber a ruled surface Y such that Cg and Y are finite coverings of Cg and Y respectively. Moreover, there is a finite cov- ering q : E → E such that Y is the projectivization P E(V ′) of the rank 2 bundle V ′ = q∗V = C ⊕ q∗L. Since the line bundle q∗L over E is also non-torsion the man- ifolds Y , X cannot have the structure of a torus bundle. The fact that X ′ cannot be a suspension over a torus follows from the same argument used to discuss the previous example. In order to describe more accurately the structure of Kahler manifolds X hav- ing non-singular vector fields, specially of those fulfilling h1 X > 0, we introduce here a more general class of deformations that combines the notion of tangential deformation introduced above with that of deformation of representations. In the sequel we assume that a lattice Λ of Cs, as well as a set of generators of it, γ1, . . . , γ2s, have been fixed. We write γj = exp1(vj ) with vj a linear combination (19) vj = s Xi=1 Bi jvi of the coordinate vector fields v1, . . . , vs of Cs and where expt(w) stands for the exponential of a vector field w, at the time t . Let us consider the product Z = F × Cs where F is a fixed compact Kahler manifold. We assume that F has no vector fields without zeros and we fix an Abelian subalgebra b of h1 F = hF (we do not exclude the case b = 0). We think of the coordinate vector fields v1, . . . , vs of Cs as defined on the product Z = F × Cs and, in the same way, we identify each w ∈ b with the vector field (w, 0) on Z. Then b and the vector fields vi generate an Abelian Lie algebra z of dimension dimC z = s + dimC b. We fix vector fields w1, . . . , w2s in b and define ϕj = exp1(vj + wj), for j = 1, . . . , 2s. Then ϕ1, . . . , ϕ2s are commuting biholomorphisms of Z defining a free Z2s-action on Z. The quotient manifold Z is naturally endowed with the locally free Cs-action induced by the projection of the vector fields vi. Notice that Z is a suspension over the torus Cs/Λ with fiber F . If the vector fields wj are all zero, then the manifold Z is just the product F × Cs/Λ. We generalize now the above construction of the manifold Z in the following way. Here, we denote by α1, . . . , αs the dual basis of v1, . . . , vs, also thought as 30 JAUME AMOR ´OS, M `ONICA MANJAR´IN, MARCEL NICOLAU holomorphic 1-forms on Z. We consider the tangential deformations of the Cs- action on Z as defined in Section 5. That is, for a given r = (C, θ) ∈ Ξ = End Cs ⊕ hα1, . . . , αsi, we define αi r as in (10), we denote by R the subset of Ξ fulfilling the conditions stated in Definition 5.1 and we define new vector fields vr i by the i are the vector fields on Z that are tangent to the condition vr Cs-action and are determined by the conditions αi i, where v′ i = Cv′ j ) = δi j and αi j) = 0. r(v′ r(v′ 1, . . . , vr We denote by Zr the product F × Cs endowed with this new complex structure i . Notice that s generate an Abelian Lie algebra zr of holomorphic vector fields on and with the holomorphic Cs-action determined by the vector fieds vr b and vr Zr. Finally, we consider the vector fields vr Xi=1 2s on Zr defined as 1, . . . , vr vr j = jvr i (20) Bi s Definition 7.8. Let Λ be a fixed lattice of Cs. Let F be a compact Kahler manifold without nonsingular vector fields and let b be an Abelian subalgebra of h1 F = hF . Fix vector fields w1, . . . , w2s in b and, for a given u = (a1, . . . , a2s) in C2s and r ∈ R, define the commuting automorphisms ϕj u,r of Zr by (21) ϕj u,r = exp1(vr j + ajwj ), j are the vector fields defined in (20). Let S be the subset of C2s × R of where vr those pairs (u, r) such that the Z2s-action on Zr defined by the automorphisms ϕj is free. In that case we denote by Zu,r the quotient manifold. u,r The family ZS = {Zu,r} is a holomorphic family of deformations of compact complex manifolds endowed with locally free Cs-actions parametrized by S. Remarks 7.9. (a) If (u, r) is an element of S then (t · u, r) also belongs to S for each t ∈ [0, 1], hence S is a connected set containing a neighborhood of R ∼= {0} × R. The restriction of the family ZS to {0} × R is just the product of F with the family of complex tori. (b) The manifolds Zu,r are not necessarily suspensions over a torus. In fact each (small) tangential deformation of Zu,r is an element of ZS. Proposition 7.10. The complex manifolds Zu,r of the family ZS defined above are Kahler manifolds for each (u, r) ∈ S. Proof. The manifolds Z0,r, being products of F with complex tori, are all Kahler manifolds. The stability theorem of Kodaira implies that Zu,r are also Kahler manifolds for u small enough. Now let an element Zu,r of the family ZS be given. We notice that Zu,r is a finite covering of the complex manifold Ym defined as the quotient of Zr by the automorphisms ψj = exp1/m(vr j + ajwj) = exp1( 1 m vr j + aj m wj) where m is a positive integer. Hence it suffices to prove that Ym is a Kahler manifold. But the map h : F × Cs → F × Cs defined by h(z, t) = (z, m t) is a holomorphic automorphism of Zr that induces a biholomorphism from Ym onto Zu′,r, where u′ = u/m. Since m can be taken arbitrarily big, Zu′,r is arbitrarily close to Z0,r and this ends the proof. (cid:3) Now, we are able to prove now the main result of this section: Theorem 7.11. Let X be a compact Kahler manifold with non-singular vector fields. There is a finite Abelian covering X ′ of X, a lattice Λ of Cs, a compact Kahler manifold F with h1 F = hF and vector fields w1, . . . , w2s in an Abelian sub- algebra b of hF such that X ′ is biholomorphic to a manifold Zu,r in the family ZS constructed above for a suitable pair (u, r) ∈ S. DEFORMATIONS OF K AHLER MANIFOLDS WITH VECTOR FIELDS 31 Proof. We know the existence of the covering X ′ of X, a small tangential deforma- ǫ of X ′ and a compact Kahler manifold F without non vanishing vector fields tion X ′ ǫ is the suspension over a torus T = Cs/Λ associated to a representation such that X ′ ρ : Λ → Aut0 C(F ) (cf. Proposition 6.5). Remark that it is sufficient to prove that X ′ ǫ belongs to the family ZS. If ρ reduces to the identity then we are in case (i) of Proposition 7.3 and the assertion is clear as X ′ ǫ = F ×T . So, assume that ρ is not constant and let γ1, . . . , γ2s be a set of generators of Λ. Notice that Aut0 C(F ) is an algebraic group. This is a consequence of Fujiki-Lieberman theorem because hF = h1 F . Therefore the Zariski closure B of ρ(Λ) in Aut0 C(F ) is an Abelian algebraic group and in particular it has a finite number of connected components. So, replacing X ′ by an Abelian finite covering if it is necessary, we can assume that the group B is connected. Let b be the Lie algebra of B. As connected Abelian Lie groups are exponential, we can write ρ(γj) = exp1(wj ), for j = 1, . . . , 2s and suitable vector fields wj ∈ b. Then the lattice Λ, the Kahler manifold F , the Abelian Lie algebra b and the vector fields w1, . . . , w2s fulfill the required conditions. (cid:3) As a corollary we obtain Corollary 7.12. Let X be a compact Kahler manifold endowed with a locally free Cs-action. Then there is a finite Abelian covering X ′ of X, which is a deformation of the product F × T of a Kahler manifold F without non vanishing vector fields and a complex torus T . Remark 7.13. Notice that s = dimC TX = dimC hX /h1 dimension of T . X can be smaller than the We are now in situation of proving Theorem 5.17, i.e. if X is a Kahler manifold then each Xr in the versal family XR of tangential deformations of (X, ) is also a Kahler manifold. Proof of Theorem 5.17. Let r ∈ R be given. An easy computation shows that if r′ is close enough to (id, 0) then Xr can be identified to an element of the versal family (Xr′ )R of Xr′. Hence we can assume without loss of generality that the Kahler manifold X is already a suspension over a torus T = Cs/Λ. Since finite coverings and finite quotients of Kahler manifolds are Kahler manifolds too, we can also assume that X is the suspension over T of a representation ρ : Λ → Aut0 C(F ), where F is a Kahler manifold without non-singular vector fields. F . Then, if we set γj = exp1(vj) = exp1(P Bi As in the proof of Theorem 7.11, we can write ρ(γj) = exp1(wj ), where γ1, . . . , γ2s are set of generators of Γ and wj are suitable vector fields in an Abelian subalgebra b of h1 jvi) as in (19), the manifold X is the quotient of Z = F × Cs by the Abelian group generated by the automorphisms ϕj = exp1(vj + wj). Hence X is an element of the family ZS associated to Λ, F and the commuting vector fields wj on F . We claim that Xr belongs also to the family ZS. Indeed, Xr is the quotient of Z by the group generated by the automorphisms that is Xr coincides with the manifold Zu1,r where u1 = (1, . . . , 1). Now the statement is a consequence of Proposition 7.10. (cid:3) exp1(X Bi jvr j + wj ), 8. Locally free Cs-actions on Kahler manifolds with small codimension In this section we classify compact Kahler manifold X endowed with a locally free holomorphic Cs-action when the codimension of the action is n − s ≤ 2. Recall that, by inequality (3), we have kod(X) ≤ n − s. 32 JAUME AMOR ´OS, M `ONICA MANJAR´IN, MARCEL NICOLAU If s = n the manifold X is a quotient of Cs, hence we have (notice that the hypothesis of being Kahler in not needed here) Proposition 8.1. Assume that s = n = dimC X, then X is a complex torus. The case s = n− 1 was first discussed by Bosio (cf. [8]). The following statement is a slight improvement of Bosio's result. We give here an alternative and simpler proof. Proposition 8.2 (Bosio). Assume that s ≥ n − 1. Then (i) If kod(X) = 1 then X has a finite Abelian covering X ′ which is a (n−1)-torus bundle over a Riemann surface Cg of genus g ≥ 2. Moreover, there is a torus T and a small b-deformation X ′ ǫ of X ′ such that X ′ ǫ = Cg × T . (ii) If kod(X) = 0 then X has a finite Abelian covering X ′ which is a torus. (iii) If kod(X) = −∞ then X is a suspension over a torus T with fiber CP 1; that is, X is a flat ruled manifold over T . In this case h1 X 6= 0. Proof. Notice that a small b-deformation Xǫ of X is a suspension over a torus T with a fiber F that can be a Riemann surface of genus g ≥ 2, an elliptic surface or CP 1. As kod(F ) = kod(Xǫ) = kod(X) these three cases correspond respectively to kod(X) equal to 1, 0 or −∞. The first two statements follow directly from Proposition 7.1 (cf. Remark 7.4). If kod(X) = −∞ then F = CP 1. But, since in this case b1(F ) = 0, the manifold X has no b-deformations and X = Xǫ. Hence X is a flat ruled manifold over T . Notice finally that, as the monodromy ρ : π1(T ) → PGL(2, C) defining the suspension is Abelian, it fixes a vector field of CP 1. This proves that h1 X 6= 0. (cid:3) Remark 8.3. Let X be a compact complex manifold in the class C endowed with a locally free Cs-action. If n − s ≤ 1 then X is a Kahler manifold. Indeed, if s = n Proposition (8.1) applies and X is a torus. If s = n − 1 then a small tangential deformation Xǫ is a suspension over a torus T with fiber a Riemann surface C. Since b2(C) = 1 the suspension Xǫ is a Kahler manifold (cf. [30, Corollary 3.21]) and Theorem (5.17) implies that X is also of Kahler type. Proposition 8.4. Assume that s ≥ n − 2. Then (I) If kod(X) ≥ 0 then X is a (n − 2)-torus bundle over a Kahler surface F . Moreover, there is a torus T , a finite, Abelian covering X ′ of X and a small b-deformation X ′ ǫ = F ×T . If b1(F ) = 0 then X ′ = F ×T . Furthermore ǫ of X ′ such that X ′ (i) If kod(X) = 2 then F is a surface of general type. (ii) If kod(X) = 1 then F is an elliptic surface. (iii) If kod(X) = 0 then the minimal model of F is a K3 surface or a torus. If F is a torus then X is a quotient of a torus. (II) If kod(X) = −∞ then the manifold X is uniruled and it has a finite, Abelian covering X ′ which belongs to one of the following types: (i) A suspension of a group representation ρ : π1(T ) → Aut0 C(F ) where F is a rational surface. (ii) A small b-deformation of the suspension of a group representation ρ : C(F ) where F is a ruled surface over a Riemann surface π1(T ) → Aut0 of genus g ≥ 1. In this case the manifold X ′ is ruled. Proof. Let us consider a small b-deformation Xǫ of X which is a suspension over a torus T with fiber a Kahler surface F . Then kod(F ) = kod(X). Assume first that kod(X) ≥ 0. In that case the statements follows from Propo- sition 7.1 and Remark 7.4. We just notice that, if kod(X) = 0 then F is a Kahler surface whose minimal model F0 can be a K3 surface, a torus, an Enriques surface DEFORMATIONS OF K AHLER MANIFOLDS WITH VECTOR FIELDS 33 or a bielliptic surface. Since Enriques surfaces and bielliptic surfaces are Abelian quotients of K3 surfaces and torus respectively, by considering an appropriate finite covering one can assume that F0 is of one of the first two types. Suppose now that kod(X) = −∞. A small b-deformation X ′ ǫ of an Abelian finite covering X ′ of X is the suspension over a torus T associated to a representation ρ : π1(T ) → Aut0 C(F ), where kod(F ) = −∞. Thus, F is either a rational surface or a ruled surface over a Riemann surface of genus g ≥ 1. If F is rational, b1(F ) = 0 and the manifold Xǫ admits no b-deformations, therefore X = Xǫ. C(F ) → Aut0 If F is a ruled surface then there is a geometrically ruled surface F obtained by recursively blowing down rational (−1)-curves of F . The set of (−1)-curves is discrete, so any element of Aut0 C(F ) fixes them. Therefore there is well defined group C( F ). The composition of the monodromy ρ with that morphism Aut0 morphism defines a suspension manifold X ′ ǫ with the following two properties: (i) it is geometrically ruled and (ii) it is obtained from X ′ ǫ by recursively contracting the families of (−1)-curves. In particular X ′ ǫ and ruled. Moreover the non-trivial fibers of the contraction X ′ ǫ are transverse to the action. This allows us to define a tangential deformation X ′ of X ′ ǫ such that it is dominated ǫ. By Proposition 5.14 the manifold X ′ is by the tangential deformation X ′ of X ′ geometrically ruled and so X ′ is ruled. ǫ is bimerophic to X ′ ǫ → X ′ (cid:3) In the case of kod(X) = −∞ we derive the following consequence: Theorem 8.5. Assume that s ≥ n − 2 and, in the case s = n − 2, let us suppose also that kod(X) = −∞. Then there is an arbitrarily small tangential deformation Xǫ of X which is a projective manifold. In order to prove the theorem we begin by characterizing when a suspension over a torus is a projective manifold. The following proposition is a direct consequence of a result by A. Blanchard [6, p. 198] since a manifold is projective if and only if a given (unramified) covering of it is projective. Proposition 8.6. Let F be a compact Kahler manifold and T = Cs/Λ a complex torus. Let X be the suspension of a group representation ρ : π1(T ) → AutC(F ) and assume that it is a Kahler manifold. The following conditions are equivalent: (i) X is projective. (ii) F and Alb(X) are projective. In that case, the complex torus T is projective. Remark 8.7. An easy computation shows that if ρ(Λ) ⊂ Aut0 C(F ) then Alb(X) is isomorphic to the suspension over the torus T of the group representation ρ′ = Φ ◦ ρ : π1(T ) → Alb(F ), where Φ is the group morphism in the exact sequence (1). Proposition 8.8. Let F be a compact projective manifold with hF /h1 F = 0. A Kahler manifold X obtained by suspension over a torus T of a group representation ρ : π1(T ) → AutC(F ) is projective if and only if T is projective. Proof. We can assume that ρ(Λ) ⊂ Aut0 zeros the map Φ is identically zero, therefore Alb(X) = Alb(F ) × T . C(F ). As F admits no vector fields without (cid:3) Proof of Theorem 8.5. By Proposition 6.5 there is a small tangential deformation τ of a finite covering X ′ of X which is a suspension over a torus T and has a fiber X ′ F fulfilling hF /h1 F = 0. With the hypothesis made, F is a projective manifold and the above Proposition applies. A t-deformation X ′ τ is still a suspension with the same fiber F and over a suitable deformation Tǫ of the torus T (cf. Example 5.12). As Abelian varieties are ǫ of X ′ 34 JAUME AMOR ´OS, M `ONICA MANJAR´IN, MARCEL NICOLAU dense in the space of complex tori we can assume that Tǫ is projective and in that case X ′ ǫ is a finite covering of a tangential deformation Xǫ of X which is also an algebraic manifold. (cid:3) ǫ is projective too. Finally, X ′ In the case of projective manifolds with kod(X) ≥ 0 we can replace the use of tangential deformations by Theorem 3.2 and we obtain a more refined classification: Corollary 8.9. Let X be a complex projective manifold of dimension n, with kod(X) ≥ 0, and such that it admits a locally free holomorphic Cs-action with s ≥ n − 2. Up to a finite, ´etale, Abelian covering, X is either of the following: (i) T n, (ii) a product Cg × T n−1 with Cg a closed Riemann surface of genus g ≥ 2, (iii) a product F ×T n−2 with F a complex projective surface with kod(F ) = kod(X) and no tangent vector fields, with T k denoting an Abelian variety of dimension k. Applying the classification of compact complex surfaces to Corollary 8.9 we de- rive the following Corollary 8.10. Conjecture 0.4 is true for projective manifolds X with a locally free holomorphic Cs-action of rank s ≥ dimC X − 2. Proof. By Corollary 8.9, such X admits as a finite ´etale cover either an Abelian variety, or a product F × T n−2 with T n−2 another Abelian variety and F a surface with kod(F ) = 0. By the Kodaira -- Enriques classification of surfaces F can be an Abelian, hyperelliptic, K3 or Enriques surface. In the two first cases, a finite ´etale cover becomes an Abelian variety. In the two last cases, an ´etale cover of degree 1 or 2 becomes a torus times a simply connected F ′ with kod(F ) = 0. (cid:3) 9. Dynamics of holomorphic vector fields A tangent vector field on a manifold defines a 1 -- parametric flow, consisting of biholomorphisms if both are complex analytic. The flow is complete when the manifold is compact. Consider the continuous dynamical system (X, v), formed by a compact Kahler manifold X and a holomorphic tangent vector field v on X. The classification of compact Kahler manifolds with tangent vector fields provided in Theorem 0.3 and Proposition 7.3 may be applied to study the dynamics of such dynamical systems. This is a refinement of the study carried out by D. Lieberman in [28] and [29]. The conclusion is that their dynamics reduce to the case of an Abelian Lie group action on a rational variety, i.e. on a variety bimeromorphic to CP n. The classification of these dynamical systems is based on the decomposition described in Propositions 1.4 and 1.6 of the Lie algebra hX of holomorphic vector fields on X as a direct sum hX = h1 X ⊕ aX , where h1 X is the subalgebra of tangent vector fields with zeros and aX is a maximal rank Abelian subalgebra of vector fields linearly independent in every point x ∈ X, which has been chosen in the center of hX . Recall that, by Fujiki's Theorem 1.8, if h1 X 6= 0 then X is uniruled. In the case of compact Kahler manifolds X with h1 X = 0, Theorem 4.4 in [29] establishes that the dynamical system (X, v) is integrable, in the sense that X admits a Seifert fibration by tori defined by the closures of the orbits of the tangent vector fields. Proposition 7.3 allows us to make the above result more precise: DEFORMATIONS OF K AHLER MANIFOLDS WITH VECTOR FIELDS 35 Corollary 9.1. Let X be a compact Kahler manifold with h1 X = 0 (e.g., if kod(X) ≥ 0), and let v be a holomorphic tangent vector field on it. There exists a finite, Abelian, unramified cover X → X such that the dynamical system ( X, v), where v is the lift of v to X, is real -- analytically conjugate to (F × T, v′), where F is a compact Kahler manifold, T is a complex torus, and v′ is a linear vector field on T . If X is a complex projective manifold then ( X, v) is biholomorphic to (F × T, v′). X = 0, the group Aut0 Proof. Under the hypothesis h1 C(X) is a torus and the vector field v is non-singular. By Proposition 7.3, X admits a small b-deformation Xǫ and there is a finite Abelian covering Xǫ → Xǫ such that Xǫ = F × T where T is a torus and F does not have vector fields without zeros. Recall that Xǫ can be seen as a b-deformation of a finite covering X of X. Moreover Xǫ is real-analitycally isomorphic to X (cf. 6.4 (a)). Notice also that, by Proposition 5.9, the Lie algebra hXǫ coincides with hX = aX . As we have noted in Remark 6.7, the commutative , the maximal Abelian algebra aX of vector fields on Xǫ lifts to a subalgebra of a Xǫ algebra of non vanishing vector fields in Xǫ formed by the vector fields that are tangent to the factor T . This proves the corollary. Notice that the last statement follows from Theorem 3.2. (cid:3) We consider now the case h1 X 6= 0. Theorem 9.2. Let X be a compact Kahler manifold with h1 X 6= 0 and let v be a nonvanishing holomorphic tangent vector field on it. There exists a finite, Abelian, unramified cover X → X such that, if v is the lift to X of v, the dynamical system ( X, v) is real -- analytically conjugate to ( Xǫ, v′ + w), where: (i) Xǫ is a small b-deformation of X which is a suspension F ×Λ Cs over a compact torus T = Cs/Λ, with fiber an uniruled compact Kahler manifold F without non-singular vector fields. (ii) There is a linear algebraic subgroup G ⊂ Aut0 that the monodromy of the suspension ρ : π1(T ) ∼= Λ → Aut0 in G. C(F ), with G ∼= Cp ⊕ (C∗)q, such C(F ) has values (iii) w ∈ Lie(G) ⊂ h1 F . (iv) The lift of the vector field v′ from F ×Λ Cs to its cover F × Cs is a linear vector field in Cs. (v) [v′, w] = 0. (vi) The topological closures G · x ⊂ F of the G -- orbits are rational varieties. If the original tangent vector field v vanishes at some point in X, the above conjugation holds with v′ = 0. If X is a projective manifold then ( X, v) is biholomorphic to ( Xǫ, v′ + w). Proof. Let aX be a central Abelian subalgebra of hX such that hX = h1 X ⊕ aX . By Proposition 7.3, X admits a small b-deformation Xǫ and there is a finite Abelian covering Xǫ → Xǫ such that Xǫ is the suspension F ×Λ Cs, over a torus T = Cs/Λ and with fiber a Kahler manifold F without non-singular vector fields, associated to a representation ρ : Λ = π1(T ) ∼= Z2s → Aut0 C(F ). As above Xǫ can be seen as a b-deformation of a finite covering X of X. We set Λ = π1(T ) = hγ1, . . . , γ2si. By Proposition 7.3 and Theorem 1.2, hF = h1 C(F ) is a linear algebraic group. Hence the Zariski closure of ρ(Λ) is an Abelian linear algebraic subgroup. In particular it has a finite number of connected components. By replacing T by a suitable Abelian finite covering of it, we can assume that γi = exp1(ξi) with ξi holomorphic vector fields on F generating an Abelian subalgebra of hF = h1 F . F and Aut0 36 JAUME AMOR ´OS, M `ONICA MANJAR´IN, MARCEL NICOLAU X and v′ ∈ aX. = h1 Xǫ ⊕ a Xǫ Xǫ = h1 X , aXǫ = aX and hXǫ = h1 Let us write v = w + v′, where w ∈ h1 If v has zeros then v′ = 0. The algebra aX is central, so [v′, w] = 0. Using Proposition 5.9, we see Xǫ ⊕ aXǫ . Hence v, w, v′ are holomorphic that h1 vector fields on Xǫ and their lifts v, w and v′ to Xǫ, still fulfill v = w + v′ ∈ h Xǫ , and [v′, w] = 0. There is a Lie algebra decomposition h Xǫ where the are projections on Xǫ = F ×Λ Cs of linear vector fields on Cs and elements of a Xǫ h1 F . On the one hand, we have Xǫ v′ ∈ a Xǫ since it is a vector field with zeros. In particular we can think of w as a vector field on F which is invariant by the monodromy. That is, fulfilling w = (ργ)∗ w for each γ ∈ Λ or, what is equivalent, by Remark 6.7. On the other hand w is an element of h1 Xǫ is naturally identified to a subalgebra of hF = h1 expk(ξi)∗ w = w, for 1 = 1, . . . , 2s and k ∈ Z. Let H be the Abelian subgroup of Aut0 C(F ) generated by the elements expk(ξi) and expt( w), where i = 1, . . . , 2s, k ∈ Z and t ∈ C. The Zariski closure G of H is an Abelian linear algebraic subgroup and by replacing again T by a suitable Abelian finite covering we can assume that G is connected. Therefore, the group G is of the form Cp × (C∗)q. The group Aut0 C(F ) not only acts compactifiably on F in the sense of Lieberman but, by [17, Remark 2.3], it also admits a compactification Aut0 C(F )∗ which is a projective variety. The group G acts on F as a linear algebraic subgroup of Aut0 C(F )∗ compactifies the action of G in F . C(F ). The Zariski closure G∗ of G in Aut0 x is an algebraic subgroup of G. Let G0 Let x be a point in F and denote by G · x its G-orbit. Then G · x = G/Gx, x of Gx in G∗ is algebraic where Gx is the stabilizer of x. The Zariski closure G∗ by GAGA. Therefore Gx = G ∩ G∗ x be the connected component of the identity in Gx. It is a connected Abelian linear algebraic subgroup of G, so the quotient G/G0 x is also an Abelian linear algebraic connected group (cf. [23]). It follows that G/G0 x is isomorphic to a group of type Cp × (C∗)q. Since the orbit G · x = G/Gx is a finite quotient of G/G0 x, it is again of type Cp × (C∗)q. As the action of G in F is compactifiable, by [29, Proposition 3.7] the orbit G · x is a dense Zariski open set of its closure G · x, which is therefore a rational variety. The fiber F is covered by such orbits, so it is uniruled. If X is a projective manifold then there is no need of considering tangential (cid:3) deformations by virtue of Theorem 3.2. Remark 9.3. In the case when the tangent field v vanishes at some point the torus T may be trivial, and our theorem says that the closures of the orbits of v are rational varieties. Applied to particular vector fields, this is already the central argument in the proof of Fujiki's Theorem 1.8, and Lieberman's version in [29]. The preceding results mean that, if one forgets linear vector fields on tori, the dynamics of a holomorphic vector field on a compact Kahler manifold reduce to the dynamics of an Abelian connected linear algebraic group (thus of type Cp × (C∗)q) acting on a rational variety. References [1] Barth, Wolf P. and Hulek, Klaus and Peters, Chris A. M. and Van de Ven, Antonius, Compact Complex Surfaces. A Series of Modern Surveys in Mathematics, Vol. 4, Springer, 2004. [2] Beauville, Arnaud, Complex Algebraic Surfaces. 2nd Edition. LMS Student Texts 34. Cam- bridge U.P., 1996. [3] Beauville, Arnaud, Vari´et´es Kahleriennes dont la premi`ere classe de Chern est nulle. J. Differential Geom. 18 (1983), no. 4, 755 -- 782. DEFORMATIONS OF K AHLER MANIFOLDS WITH VECTOR FIELDS 37 [4] Birkenhake, Christina; Lange, Herbert, Complex Abelian Tori. Progress in Mathematics 177. Birkhauser, 1999. [5] Birkenhake, Christina; Lange, Herbert, Complex Abelian Varieties. Second edition. Grundlehren der Mathematischen Wissenschaften, 302. Springer-Verlag, Berlin, 2004 [6] Blanchard, M. Andr´e, Sur les vari´et´es analytiques complexes. Ann. Sci. Ecole Norm. Sup. (3) 73 (1956) 157 -- 202. [7] Bogomolov, Fedor. A. Kahler manifolds with trivial canonical class. Izv. Akad. Nauk SSSR Ser. Mat. 38 (1974), 11 -- 21. [8] Bosio, Fr´ed´eric, Actions holomorphes et localement libres de groupes de Lie ab´eliens. Th`ese de Doctorat, ´Ecole Normale Sup´erieure de Lyon, 1996. [9] Calabi, Eugenio On Kahler manifolds with vanishing canonical class. Algebraic geometry and topology. A symposium in honor of S. Lefschetz, pp. 78 -- 89. Princeton University Press, Princeton, N. J., 1957. [10] Campana, Fr´ed´eric, The class C is not stable by small deformations. Math. Ann. 290 (1991), no. 1, 19 -- 30. [11] Carrell, James B., Holomorphically injective complex toral actions. Proceedings of the Con- ference on Compact Transformation Groups. U. of Massachussets, Amherst 1971. Vol. 2. Lect. Notes in Math. 299. Springer-Verlag, 1972. [12] Carrell, James B.; Lieberman, David I. Holomorphic vector fields and Kahler manifolds. Invent. Math. 21 (1973), 303 -- 309. [13] Carrell, James B.; Sommese, Andrew John C ∗-actions. Math. Scand. 43 (1978/79), no. 1, 49 -- 59. [14] Cathelineau, Jean-Louis, D´eformations ´equivariantes d'espaces analytiques complexes com- pacts. Ann. Sci. ´Ecole Norm. Sup. (4) 11 (1978), no. 3, 391 -- 406. [15] Demailly, Jean-Pierre; Paun, Mihai, Numerical characterization of the Kahler cone of a compact Kahler manifold. Ann. of Math. (2) 159 (2004), no. 3, 1247 -- 1274. [16] Fujiki, Akira, Closedness of the Douady spaces of compact Kahler spaces. Publ. Res. Inst. Math. Sci. 14 (1978), no. 1, 1 -- 52. [17] Fujiki, Akira, On automorphism groups of compact Kahler manifolds. Invent. Math. 44 (1978), no. 3, 225 -- 258. [18] Fujiki, Akira, On the structure of compact complex manifolds in C. Algebraic varieties and analytic varieties (Tokyo, 1981), 231 -- 302, Adv. Stud. Pure Math., 1, North-Holland, Ams- terdam, 1983. [19] Girbau, Joan; Nicolau, Marcel, On deformations of holomorphic foliations. Ann. Inst. Fourier (Grenoble) 39 (1989), no. 2, 417 -- 449. [20] G´omez-Mont, Xavier, Transverse deformations of holomorphic foliations. The Lefschetz cen- tennial conference, Part I (Mexico City, 1984), 127 -- 139, Contemp. Math., 58, Amer. Math. Soc., Providence, RI, 1986. [21] Hall, R., On algebraic varieties which possess finite continuous commutative groups of bira- tional self-transformations. J. London Math. Soc. 30 (1955) 507 -- 511. [22] Holmann, Harald, On the stability of holomorphic foliations with all leaves compact. Vari´et´es analytiques compactes (Colloq., Nice, 1977), pp. 217 -- 248, Lecture Notes in Math., 683. Springer, Berlin, 1978. [23] Humphreys, James E., Linear algebraic groups. Graduate Texts in Mathematics, No. 21. Springer-Verlag, New York-Heidelberg, 1975. [24] Kodaira, K; Spencer, D.C. On deformations of complex analytic structures. I, II. Ann. of Math. (2) 67 (1958) 328 -- 466. [25] Koll´ar, J´anos, Shafarevich Maps and Automorphic Forms. M. B. Porter Lectures. Princeton University Press, Princeton, NJ, 1995. [26] LeBrun, Claude; Poon, Yat Sun, Twistors, Kahler manifolds, and bimeromorphic geometry. II. J. Amer. Math. Soc. 5 (1992), no. 2, 317 -- 325. [27] Lichnerowicz, Andr´e, Vari´et´es kahleriennes et premi`ere classe de Chern. J. Differential Ge- ometry 1 (1967) 195 -- 223. [28] Lieberman, David I., Holomorphic vector fields on projective varieties. Proc. Symp. Pure Math. 30. AMS 1977. [29] Lieberman, David I., Compactness of the Chow scheme: applications to automorphisms and deformations of Kahler manifolds. Fonctions de plusieurs variables complexes, III (S´em. Fran¸cois Norguet, 1975 -- 1977), pp. 140 -- 186, Lecture Notes in Math., 670, Springer, Berlin, 1978. [30] Manjar´ın, M`onica, Normal almost contact structures and non-Kahler compact complex man- ifolds. Indiana Univ. Math. J. 57 (2008), no. 1, 481 -- 507. [31] Matsushima, Yozo , Holomorphic vector fields and the first Chern class of a Hodge manifold. J. Differential Geometry 3 (1969) 477 -- 480. 38 JAUME AMOR ´OS, M `ONICA MANJAR´IN, MARCEL NICOLAU [32] Severi, Francesco, Funzione quasi abeliane. Pontificiae Academiae Scientiarum Scripta Varia, vol 4. Citt`a del Vaticano, 1947. [33] Ueno, Kenji, Classification theory of algebraic varieties and compact complex spaces. Lecture Notes in Math., 439, Springer-Verlag, Berlin, 1975. [34] Wang, Hsien-Chung, Complex parallisable manifolds, Proc. Amer. Math. Soc. 5 (1954), 771 -- 776. Jaume Amor´os. Departament de Matem`atica Aplicada I, Universitat Polit`ecnica de Catalunya, Diagonal 647, E-08028 Barcelona, Spain M`onica Manjar´ın. Institut de Recherche Math´ematique de Rennes (IRMAR), Campus de Beaulieu Bat.22, 35042 Rennes cedex, France Marcel Nicolau. Departament de Matem`atiques, Universitat Aut`onoma de Barcelona, Bellaterra 08193, Spain E-mail address: [email protected], [email protected], [email protected]
1308.4104	2	1308	2015-08-05T16:43:18	Homology of Hilbert schemes of points on a locally planar curve	[ "math.AG" ]	Let C be a proper, integral, locally planar curve, and consider its Hilbert schemes of points C^[n]. We define 4 creation/annihilation operators acting on the rational homology groups of these Hilbert schemes and show that the operators satisfy the relations of a Weyl algebra. The action of this algebra is similar to that defined by Grojnowski and Nakajima for a smooth surface. As a corollary, we compute the cohomology of C^[n] in terms of the cohomology of the compactified Jacobian of C together with an auxiliary grading on the latter. This recovers and slightly strenghtens a formula recently obtained in a different way by Maulik and Yun and independently Migliorini and Shende.	math.AG	math	Homology of Hilbert schemes of points on a locally planar curve Jørgen Vold Rennemo Let C be a proper, integral, locally planar curve, and consider its Hilbert schemes of points C [n]. We define 4 creation/annihilation operators acting on the rational homology groups of these Hilbert schemes and show that the operators satisfy the relations of a Weyl algebra. The action of this algebra is similar to that defined by Grojnowski and Nakajima for a smooth surface. As a corollary, we compute the cohomology of C [n] in terms of the coho- mology of the compactified Jacobian of C together with an auxiliary grading on the latter. This recovers and slightly strenghtens a formula recently ob- tained in a different way by Maulik and Yun and independently Migliorini and Shende. 1 Introduction Let C be a proper, integral, complex curve with planar singularities. Denote by C [n] the Hilbert scheme of length n subschemes of C. Let J be the compactified Jacobian, i.e. the space of torsion free sheaves on C with rank 1 and degree 0. These spaces are related by the Abel -- Jacobi morphism AJ : C [n] → J, which sends a subscheme Z to the sheaf IZ ⊗ O(x)⊗n, where x ∈ C is a chosen nonsingular point. Under our assumptions on C, both C [n] and J are reduced and irreducible with l.c.i. singularities [AIK77, BGS81]. Let g be the arithmetic genus of C. For n ≥ 2g − 1 the map AJ is a Pn−g-bundle (see [AK80]), so the rational homology group H∗(C [n]) is determined up to isomorphism by H∗(J). The formula below extends this by expressing H∗(C [n]) in terms of H∗(J) even for n < 2g − 1. In order to state the result, we will define a new grading on H∗(J), with the m-th graded piece denoted DmH∗(J). This D-grading combines with the homological grading to give a bigrading, and we have DmH∗(J) = 0 unless 0 ≤ m ≤ 2g. We then have the following formula. Proposition 1.1. There is an isomorphism of homologically graded vector spaces H∗(C [n]) ∼= Mm≤n DmH∗(J) ⊗ Symn−m(Q ⊕ Q[2]). 1 Here Q[2] denotes the space Q with homological degree 2. A very similar statement was recently shown by Maulik and Yun [MY14] and Migliorini and Shende [MS13]. See Section 1.5 for a discussion of how these papers relate to this one. 1.1 Algebra action Proposition 1.1 will be obtained as a corollary of our main result, which we now describe. Consider the vector space V (C) :=Mn≥0 H∗(C [n]). We shall define two pairs of creation and annihilation operators acting on V (C). The first pair is denoted µ±[pt] : H∗(C [n]) → H∗−1±1(C [n±1]) and corresponds to adding or removing a fixed nonsingular point x ∈ C. Indeed, any such x induces an inclusion i : C [n] ֒→ C [n+1] by letting Ii(Z) = IZ · Ix for every Z ∈ C [n]. We then take µ+[pt] = i∗ and µ−[pt] = i!, where i! is the intersection pullback map. The second pair is denoted µ±[C] : H∗(C [n]) → H∗+1±1(C [n±1]), and the operators are correspondences induced by the diagram C [n,n+1] p z✈✈✈✈✈✈✈✈✈✈ %❑❑❑❑❑❑❑❑❑❑ q C [n] C [n+1], that is µ+[C] = q∗p! and µ−[C] = p∗q!. Here C [n,n+1] is the flag Hilbert scheme of pairs (Z, Z ′) ∈ C [n] × C [n+1] such that Z ⊂ Z ′. The fact that the Gysin maps p!, q! are well defined is nontrivial, since all three schemes are in general singular. In particular, the definition depends on the assumption that C is locally planar; see Section 2. The main result of this paper is the following. Theorem 1.2. (i) The operators µ±[pt], µ±[C] ∈ End(V (C)) satisfy the commutation relations [µ−[pt], µ+[C]] = [µ−[C], µ+[pt]] = id, and all other pairs of operators commute. (ii) Let W = ker µ−[pt] ∩ ker µ−[C]. Then the natural map W ⊗ Q[µ+[pt], µ+[C]] → V (C) is an isomorphism. (iii) The Abel -- Jacobi pushforward map AJ∗ : V (C) → H∗(J) induces an isomorphism W ∼= H∗(J). 2 z % Point (i) can be rephrased as saying that the subalgebra of End(V (C)) generated by µ±[pt], µ±[C] is isomorphic to the Weyl algebra C[x1, x2, ∂1, ∂2]. Note that V (C) is naturally bigraded by taking the (i, n)-th homogeneous piece to be Hi(C [n]). The four operators are bihomogeneous, so the space W in the theorem inherits a bigrading, and so by part (iii) we get an induced bigrading on H∗(J). We let DnHi(J) denote the (i, n)-th homogeneous part of H∗(J). Restricting the isomorphism of (ii) to a single H∗(C [n]) then gives Proposition 1.1. 1.2 On the proof Assuming the commutation relations of Theorem 1.2 (i), the proof of part (ii) is a matter of elementary algebra. The proof of (iii) is then quite easy, using the fact that for large n the map C [n] → J is a projective space bundle [AK80]. Finally, for checking the commutation relations of (i), the idea is the following. The operators can all be thought of as correspondences. If the C [n] were smooth, we could apply the usual composition formula for correspondences, and so reduce the calculation of each commutator to computing a specific class in H∗(C [n] × C [n′]), with n′ ∈ {n − 2, n, n + 2}. The idea for circumventing the non-smoothness of the C [n] is to embed C in an al- gebraic family C → B over a smooth base B, such that the relative Hilbert schemes C[n] → B are nonsingular for all n. That this is possible follows from the fact that C is locally planar, as was shown by Shende [She12, Cor. 15]. Given such a family, we may compose correspondences in the family, compute the commutators (this is possible by the nonsingularity of C[n]), and finally restrict to the fibre C [n]. 1.3 Variants The main theorem has natural variants in cohomology and Chow homology: 1.3.1 Cohomology Since we are working with Q-coefficients, we may dualise every vector space and consider cohomology instead of homology. Let V c(C) = Mi,n≥0 H i(C [n], Q). We let the operators µc µ∓[pt]∗ and µc ±[C] = µ∓[C]∗. ± acting on cohomology be defined by dualising, i.e. by µc ±[pt] = Then from Theorem 1.2 we easily get the following cohomological version. Theorem 1.3. (i) The operators µc ±[pt], µc ±[C] ∈ End(V c(C)) satisfy the commutation relations −[pt], µc [µc +[pt]] = id, −[C], µc +[C]] = [µc and all other pairs of operators commute. 3 (ii) Let W c = V c(C)/(im µc +[pt] + im µc +[C]). Then the natural maps ker µc −[pt] ∩ ker µc −[C] → W c and (ker µc −[pt] ∩ ker µc −[C]) ⊗ Q[µc +[pt], µc +[C]] → V c(C) are isomorphisms. (iii) The Abel -- Jacobi pullback map AJ ∗ : H ∗(J) → H ∗(C [n]) induces an isomorphism H∗(J) ∼= W c. The natural bigrading on V c(C) induces a bigrading on W c, and hence a bigrading on H ∗(J), which we write as H ∗(J) = ⊕i,nDnH i(J). As in the case of homology, we recover every H ∗(C [n]) from the data of H ∗(J) with this D-grading, i.e. H ∗(C [n]) ∼= Mm≤n DmH ∗(J) ⊗ Symn−m (Q ⊕ Q[−2]) . (1) The following question seems natural. Question 1.4. Is the cup product on H ∗(J) homogeneous with respect to the D-grading? 1.3.2 Chow homology Instead of the homology groups H∗(C [n]) and H∗(J) we may work with Chow homology groups A∗(C [n]) and A∗(J) (with rational coefficients). The operators µ±[pt] and µ±[C] can still be defined in this setting, and Theorem 1.2 then holds. The proof is the same as in the case of singular homology, and we shall only indicate the changes necessary at the few places where these occur. Note that in this setting the operators µ±[pt] will in general depend on the particular point x ∈ C chosen for the definition of C [n] ֒→ C [n+1]. 1.4 Applications to curve counting and BPS numbers The present work is related to considerations in curve counting on Calabi -- Yau 3-folds. See also the introduction to [MS13] or the survey paper [PT14] for background on these curve counting theories. Under our assumptions on the curve C, Pandharipande and Thomas [PT10, App. B] show that there are integers ng such that q1−g(C) ∞Xn=0 χ(C [n])qn = g(C)Xg=g( eC) ng(cid:18) q (1 − q)2(cid:19)1−g(C) . (2) Here g(C) and g(eC) are the arithmetic and geometric genera of C, respectively. If C lies in a Calabi -- Yau 3-fold X, then one may in certain good cases interpret the ng as the contribution of C to the BPS invariant ng,[C] of Gopakumar and Vafa, see [PT10]. 4 In Gopakumar and Vafa's original proposal [GV98a, GV98b] the BPS invariants ng,[C] of a Calabi -- Yau 3-fold X are computed from the cohomology of the space of pure 1- dimensional sheaves on X. For a single curve C, this computation suggests the following alternative way of defining the contribution of C to ng,[C]: The cohomology H ∗(J) should in some sense split as the direct sum of cohomologies H ∗(T 2g) for different g, where T 2g is real 2g-dimensional torus. The contribution of C to ng,[C] should then be the number of copies of H ∗(T 2g) appearing in the decomposition. Formula (1) gives one way of making this precise, as follows. The right hand side of (2) is a rational function invariant under q 7→ q−1, hence the left hand side is as well. Let χ(DnH ∗(J)) = dim DnH even(J) − dim DnH odd(J). Applying (1) one can then check that the Laurent polynomial q−g(C) 2g(C)Xn=0 χ(DnH ∗(J))qn is invariant under q 7→ q−1 as well.1 Thinking of (q−1−2+q)g as the shifted Poincaré polynomial of T 2g, it is then reasonable to define the contribution n′ g of C to ng,[C] by q−g(C) 2g(C)Xn=0 χ(DnH ∗(J))qn = 2g(C)Xg=0 g(q−1 − 2 + q)g. n′ From (1) we then easily get the following proposition. Proposition 1.5. The two definitions of the contribution of C to the BPS number ng,[C] agree, i.e. we have ng = n′ g for all g. 1.5 Relation to existing work The results in this paper are motivated by the recent work of Maulik and Yun [MY14] In these papers H ∗(J) is endowed with a certain and Migliorini and Shende [MS13]. ∗ H ∗(J) then recovers H ∗(C [n]) in the perverse filtration P , and the P -graded space grP same way that our D-graded H ∗(J) recovers H ∗(C [n]). In Section 7, we show that the grading D is in fact a splitting of the filtration P . This filtration P arises in a completely different way to our D-grading. Consider a deformation family C → B such that the relative compactified Jacobian f : J → B is nonsingular. Then Rf∗(QJ ) ∈ Db c(B) has a filtration induced by the perverse t- c(B), which restricts to give the filtration P on H ∗(J). The main result structure on Db 1The can be refined to symmetry of this polynomial an isomorphism DnH k(J) ∼= D2g(C)−nH k+2g(C)−2n(J). This follows from the relation between the D-grading and the perverse fil- tration on H ∗(J) (Prop. 7.1) and the relative hard Lefshetz theorem applied to the perverse filtration, see [MY14, 2.16]. 5 of [MY14, MS13] is a description of the object Rf∗(QJ ), with the formula for H ∗(C [n]) then appearing as a corollary. In contrast, we restrict ourselves to the study of the single curve C. This paper grew out of an attempt to prove Proposition 1.1 without the technology of perverse sheaves and the decomposition theorem. That such a proof should exist was suggested to us by Richard Thomas. The approach we take is inspired by Nakajima's [Nak97] and Grojnowski's [Gro96] construction of an action of an infinite-dimensional Heisenberg algebra on the homologies of the Hilbert schemes of a smooth surface. Both the definition of our operators and the strategy for proving their commutation relations are analogous to the corresponding parts of Nakajima's paper. The main technical contribution of this paper lies in defining the operators and proving the commutation relations in the context of the singular spaces C [n]. For a curve C which is smooth over a quasi-projective smooth base variety S, Moonen and Polischuk [MP10] have computed ⊕n≥0A∗(C [n]) in terms of A∗(J), using a similar strategy to that of this paper. Their computation holds in Chow groups with integral coefficients. Specialising to S = Spec C and tensoring the Chow groups with Q, we recover the Chow version of Proposition 1.1 for a smooth C. 1.6 Outline of the paper The paper is laid out as follows. In Section 2 we give the precise definitions of the 4 operators. In Section 3 we assume the commutation relations of Theorem 1.2 (i) and deduce parts (ii) and (iii). For the proof of the commutation relations, it will be convenient to use the language of bivariant homology theory, as laid out in [FM81]. In Section 4 we give a summary of the relevant parts of this theory, and in Section 5 we prove Theorem 1.2 (i). In Section 6 we collect a few lemmas on the incidence schemes C [n,n+1] which we need elsewhere. Finally, in Section 7 we show that the grading D is a splitting of the perverse filtration of [MY14, MS13]. 1.7 Acknowledgements I would like to thank S. Kleiman, A. MacPherson, A. Oblomkov, R. Pandharipande, J. Rognes, V. Shende, C. Vial and my supervisor R. Thomas for useful discussions and comments related to this work. Special thanks to C. Vial for pointing out and correcting an error in the proof of Thm. 1.2 (iii). 2 Definition of the four operators 2.1 The deformation family of C The following construction is essential for the definition of the operators µ±[C] and for proving the commutation relations. 6 Choose an algebraic family f : C → B, where B is nonsingular, such that f −1(0) ∼= C for some 0 ∈ B. Let C[n] → B be the relative Hilbert scheme, that is the scheme such that the fibre over b ∈ B is (Cb)[n]. By [She12, Cor. 15] we may choose the family so that the scheme C[n] is nonsingular for all n. Possibly after an étale base change, we may assume that the family admits a section s : B → C such that the image of s is disjoint from the discriminant locus of f . Restricting the base further, we may assume that every curve in the family is reduced and irreducible. For the remainder of the paper, we fix the data of the family C → B, the section s : B → C and the nonsingular point x = s(0) ∈ C. 2.2 Definition of µ±[pt] Let i : C [n] → C [n+1] be the morphism defined on the level of points by Ii(Z) = Ix · IZ ∀Z ∈ C [n]. In other words, the map i is defined by adding a point at x. Lemma 2.1. The embedding i : C [n] ֒→ C [n+1] is regular. Proof. The property of being regular is analytic local [ACG11, Lemma 2.6]. Let Z ∈ C [n] be a point such that Z has length k at x. Choose an analytic open U around x such that the only component of Z contained in U is the one at x. Then locally around Z the morphism i is isomorphic to (U )[k] × (C \ U )[n−k] (j,id) ֒→ (U )[k+1] × (C \ U )[n−k], where j : (U )[k] ֒→ (U )[k+1] is the morphism which adds a point at x. Since (U )[k] and (U )[k+1] are smooth, j is a regular embedding, and hence so is i. As a consequence of Lemma 2.1, there is a Gysin map i! : H∗(C [n]) → H∗−2(C [n−1]). We let µ+[pt] = i∗ and µ−[pt] = i!. 2.3 Definition of µ±[C] The operators µ±[C] are defined as correspondences in the following way. Let C [n,n+1] ⊂ C [n] × C [n+1] be the flag Hilbert scheme parametrising pairs (Z, Z ′) such that Z ⊂ Z ′. Let C[n,n+1] be its relative version, that is the scheme over B such that for every b ∈ B, the fibre over b is (Cb)[n,n+1]. We then have the diagram C [n,n+1] %❑❑❑❑❑❑❑❑❑❑ q p z✉✉✉✉✉✉✉✉✉✉ C [n] C [n+1]. We will define maps p! : H∗(C [n]) → H∗+2(C [n,n+1]) and q! : H∗(C [n+1]) → H∗(C [n,n+1]). With these maps defined, we let µ+[C] = q∗p! and µ−[C] = p∗q!. 7 z % Consider the Cartesian square C [n,n+1] / C[n,n+1] p C [n] / C[n] Let d = dim C[n]. By Lemma 6.5, C[n,n+1] is irreducible of dimension d + 1. Since C[n] is nonsingular, we have H∗(C [n]) ∼= H ∗(C[n], C[n] \ C [n]). It then follows from [Ful98, Ex. 19.9.10] that there exists a refined intersection product − × − : Hk(C [n]) ⊗ H BM l (C[n,n+1]) → Hk+l−2d(C [n,n+1]). Now let α ∈ Hk(C [n]), and let [C[n,n+1]] ∈ H BM 2d+2(C[n,n+1]) be the fundamental class. We then define p!(α) = α × [C[n,n+1]] ∈ Hk+2(C [n,n+1]). The definition of q! is similar. 3 Proof of main theorem from commutation relations In this section, we take the commutation relations of Theorem 1.2 (i) for granted and show how parts (ii) and (iii) of the theorem follow from this. Part (ii) is a formal consequence of the commutation relations and the fact that µ−[pt] and µ−[C] are locally nilpotent. Lemma 3.1. Let V be a vector space over a field k with char(k) = 0, and let µ−, µ+ ∈ End(V ) satisfy [µ−, µ+] = id. Assume further that for every v ∈ V there is an integer n ≥ 0 such that µn −v = 0. Then the natural map is an isomorphism. (ker µ−) ⊗ k[µ+] → V Proof. Note first of all that if v ∈ ker µ−, the commutation relation implies that µ−µn nµn−1 + v. Let φ : (ker µ−) ⊗ k[µ+] → V be the natural map. We first show that φ is injective. +v = Suppose not, then there is some relation nXi=0 µi +vi = 0 vi ∈ ker µ− with vn non-zero. Acting on this relation by µn gives n!vn = 0, which is a contradiction. − and using the commutation relation We next show that φ is surjective. For any v ∈ V , we define the nilpotency of v to be the smallest n ≥ 0 such that µn −v = 0. Suppose φ is not surjective, and let v ∈ V be an element of minimal nilpotency among those such that v 6∈ im φ. The nilpotency of µ−v is less than that of v, so we have µ−v ∈ im φ. Hence we have µ−v = nXi=0 µi +vi vi ∈ ker µ−. 8 / / Now write v = nXi=0 1 i + 1 µi+1 + vi + v′ (3) for some v′ ∈ V . Applying µ− to (3) shows that v′ ∈ ker µ−. The right hand side of (3) then clearly belongs to im φ, hence v does. Proof of Theorem 1.2 (ii). Since µ−[pt] commutes with µ−[C] and µ+[pt], the action of µ−[C] and µ+[pt] preserves ker µ−[pt]. Applying Lemma 3.1 with V = ker µ−[pt], µ− = µ−[C], and µ+ = µ+[pt], we see that the natural map W ⊗ Q[µ+[pt]] = (ker µ−[pt] ∩ ker µ−[C]) ⊗ Q[µ+[pt]] → ker µ−[pt] is an isomorphism. Similarly we find that the map ker µ−[pt] ⊗ Q[µ+[C]] → V (C) is an isomorphism. Combining these two isomorphisms and the fact that µ+[C] and µ+[pt] commute gives the result. Let g be the arithmetic genus of C. Lemma 3.2. The map AJ∗ : ker µ−[pt] ∩ H∗(C [n]) → H∗(J) is injective for any n, and is an isomorphism for n ≥ 2g. Proof. Since the map µ+[pt] : ker µ−[pt]∩H ∗(C [n]) → ker µ−[pt]∩H ∗(C [n+1]) is injective by Theorem 1.2 (ii) and AJ∗ = AJ∗ ◦ µ+[pt], it suffices to prove the claim when n ≥ 2g. For n ≥ 2g − 1 the morphism AJ : C [n] → J is a Pn−g-bundle [AK80]. Let ω = [i(C [n−1])] ∈ H 2(C [n]), where i is the inclusion map i : C [n−1] ֒→ C [n], and let r = n−g be the fibre dimension of C [n] → J. The divisor i(C [n−1]) ⊂ C [n] is a projective subbundle, hence we may express every α ∈ H∗(C [n]) uniquely as α = rXi=0 ωi ∩ AJ !(αi) αi ∈ H∗(J), (4) where AJ ! is the Gysin pull-back associated to a projective bundle. (See [Ful98, Thm. 3.3] for a proof of this in the case of Chow groups.) Note that we have AJ∗(α) = αr. We first prove injectivity of AJ∗. By part (ii) of the main theorem, µ+[pt] is injective. Hence ker µ−[pt] = ker(µ+[pt]µ−[pt]). By definition of the operators we have µ+[pt]µ−[pt](α) = i∗i!(α) = ω ∩ α ∀α ∈ H∗(C [n]). Suppose AJ∗(α) = 0 and µ−[pt](α) = 0. Writing α as above this means αr = 0, and further that ω ∩ α = 0. This implies αi = 0 for all i, hence α = 0. To prove surjectivity when n ≥ 2g, we note first that r = n − g ≥ g = dim J. Let 0 6= β ∈ Hk(J), and let α = ωr+1 ∩ AJ !(β). Write α in terms of αi as in (4). Since β 6= 0 9 we have k ≤ 2 dim J ≤ 2r, and then the homological degree of α0 is k − 2 − 2r ≤ −2, so we have α0 = 0. We now take γ = ωr ∩ AJ !(β) − r−1Xi=0 ωi ∩ AJ !(αi+1). We see that AJ∗(γ) = β and µ+[pt]µ−[pt](γ) = ω ∩ γ = 0, hence µ−[pt](γ) = 0. Proof of Theorem 1.2 (iii). The inclusion map i : C [n] → C [n+1] commutes with the Abel -- Jacobi map, in the sense that AJ ◦ i = AJ. It follows that AJ∗ = AJ∗ ◦ µ+[pt]. We first show AJ∗ : W → H∗(J) is surjective. Let α ∈ H∗(J). By Lemma 3.2 there exists some class α ∈ ker µ−[pt] such that AJ∗(α) = α. But by Theorem 1.2 (ii) we may write α =Xi µ+[pt]iαi with αi ∈ W , which implies α = AJ∗(P αi). Using Theorem 1.2 (ii) and the fact that C [n] → J is a Pn−g-bundle, one checks that dim W = dim H∗(J), hence AJ∗ is an isomorphism. If we want to prove the version of Theorem 1.2 (iii) for Chow groups, the dimensions of W and A∗(J) may be infinite. In this case we can prove injectivity directly as follows. Let α ∈ W be such that AJ∗(α) = 0. If α ∈ W ∩ A∗(C [n]) for some n, then Lemma i=m αi, with αi ∈ 3.2 shows α = 0. If this is not the case, then we can write α = Pn A∗(C [i]) ∩ W and αm, αn 6= 0. Let β = P µ+[pt]n−i(αi). Then µ−[pt](β) = 0 and AJ∗(β) = 0, hence by Lemma 3.2 we have β = 0. But µ−[C](β) = P(n − i)αi 6= 0, which gives a contradiction. 4 Bivariant homology formalism In order to be precise about which Gysin pull back maps we are using and what the com- patibilities between them are, we use the formalism of bivariant homology as presented by Fulton and MacPherson in [FM81]. As the scope of the general theory is quite broad, we give here a recap of the parts of the theory we need. See [FM81] for the full story and in particular Section I.3 for details on the topological case. 4.1 Description of the bivariant theory The bivariant Borel -- Moore homology theory assigns to each map f : X → Y of reason- able2 topological spaces a graded abelian group H ∗(X with 3 operations. f → Y ). The theory is equipped • Product: Given maps X f → Y and Y g → Z, there is a product homomorphism H i(X f → Y ) ⊗ H j(Y g → Z) → H i+j(X g◦f → Z). 2We require that X and Y can be written as closed subspaces of Rn for some n; see [FM81, I.3.1.1]. 10 For α ∈ H i(X Z). f → Y ) and β ∈ H j(Y g → Z) we thus get a product α·β ∈ H i+j(X g◦f → • Pushforward: For any proper map X g◦f → Z) → H ∗(Y forward homomorphism f∗ : H ∗(X g → Z). f → Y and any map Y g → Z there is a push- • Pullback: For any Cartesian square X ′ g Y ′ X f / Y there is a pullback homomorphism H ∗(X f → Y ) → H ∗(X ′ g → Y ′). These operations satisfy various compatibility axioms, see [FM81, Sec. I.2.2]. 4.2 Relation to homology For any space X, the groups H i(X → pt) and H i(X −i (X) and H i(X), respectively. Note that the three bivariant operations recover the usual homological operations of cup and cap product, proper pushforwards in homology and arbitrary pullbacks in cohomology. id → X) are identified with H BM 4.3 Nonsingular targets The following observation will be crucial. If Y is a nonsingular variety and f : X → Y is any morphism, the induced homomorphism H ∗(X f → Y ) → H ∗−2 dim Y (X → pt) = H BM 2 dim Y −∗(X) given by taking the product with [Y ] ∈ H −2 dim Y (Y → pt) is an isomorphism. In such a situation we will frequently identify H ∗(X → Y ) with H BM In particular, if X has a fundamental class [X] ∈ H BM 2 dim Y −∗(X). 2 dim X(X), this induces a class [X] ∈ H 2(dim Y −dim X)(X → Y ). 4.4 Gysin maps Any class α ∈ H i(X f → Y ) defines a Gysin pull-back map f ! : H BM ∗ (Y ) → H BM ∗−i (Y ) by f !(β) = α · β, ∀β ∈ H BM ∗ (Y ). This relates to the Gysin maps p! and q! in the definition of µ±[C] (see Sec. 2.3) as follows. Consider the Cartesian square C [n,n+1] / C[n,n+1] p C [n] / C[n] 11 / / / / / as in Sec. 2.3. The fundamental class [C[n,n+1]] ∈ H BM (C[n,n+1]) is identified with an element [C[n,n+1]] ∈ H −2(C[n,n+1] → C[n]), since C[n] is nonsingular. Cartesian pullback defines an element ^[C[n,n+1]] ∈ H −2(C [n,n+1] → C [n]), and the Gysin pullback map asso- ciated with ^[C[n,n+1]] coincides with p!. A similar description can be given for q!. ∗ 4.5 Notation In a commutative diagram a Latin letter next to an arrow denotes the morphism, while a Greek letter denotes a bivariant homology class, so that e.g. the α in X Y denotes α −→ f an element α ∈ H ∗(X f −→ Y ). 4.6 Chow theory There is a bivariant operational Chow theory assigning to every morphism of varieties X → Y an abelian group A∗(X → Y ) [FM81, Sec. I.9]. In this case A∗(X → pt) equals the ordinary Chow group A∗(X) of X. This bivariant theory is equipped with the same operations as the Borel -- Moore theory satisfying the same compatibilities. It also has the ·[Y ] → A∗−dim Y (X → pt) is an isomorphism for nonsingular Y property that A∗(X → Y ) [FM81, I.9.1.3]. Because of this, the proof of the commutation relations goes through verbatim upon replacing every H with an A. 5 Proof of commutation relations We now show that the operators obey the commutation relations of Theorem 1.2 (i). 5.1 Proof of [µ−[pt], µ+[C]] = [µ−[C], µ+[pt]] = id Consider the diagrams θ κ q z✈✈✈✈✈✈✈✈✈✈ p C[n,n+1] C[n+1] eι ei ι i X? eκ eq C[n] C[n] and C[n−1,n] θ′ zttttttttt p′ $❍❍❍❍❍❍❍❍❍❍ κ′ q′ C[n], C[n] ι′ i′ C[n−1] where in the first diagram X = C[n,n+1] ×C[n+1] C[n] and the square containing X is Cartesian. The morphisms i, i′ correspond to adding a point at the section s : B → C, see Sections 2.1, 2.2. The bivariant classes θ, ι, κ and their primed versions are the ones 12 z _ o o ? _ o o z $ ? _ o o pullbacks of ι and κ, respectively. defined by fundamental classes, as in Section 4.3. The classeseι and eκ are the Cartesian Both of these diagrams are defined over the base B of the family C. For any scheme, morphism or bivariant class we denote the result of performing the base change to 0 ∈ B by appending a subscript 0 to the object in question. We first treat the case of [µ−[pt], µ+[C]]. For any α ∈ H∗(C [n]), we have and µ−[pt]µ+[C](α) = ι0 · (q0)∗(θ0 · α) = (eq0)∗(eι0 · θ0 · α) µ+[C]µ−[pt](α) = (q′ 0 · α). 0)∗(θ′ 0 · ι′ Lemma 5.1. Under the identification of H ∗(X p◦ei → C[n]) with H BM ∗+2 dim C[n](X), we have Proof. The classeι is the same as the class induced by X ֒→ C[n,n+1] being the embedding of a Cartier divisor. It follows that eι · θ = [X] eι · θ · [C[n]] =eι · [C[n,n+1]] = [X]. We will now compute [X] by describing the irreducible components of X. In order to do this, we define certain maps f : C[n−1,n] → X and g : C[n] → X. Since X = C[n,n+1] ×C[n+1] C[n], we can describe f and g as products of suitable maps to C[n,n+1] and C[n]. We then let f be the product of the map C[n−1,n] → C[n,n+1] sending (Z, Z ′) to (i(Z), i(Z ′)) with the map q′ : C[n−1,n] → C[n]. We let g be the product of the map C[n] → C[n,n+1] sending Z to (Z, i(Z)) with the identity map on C[n]. Lemma 5.2. In H BM ∗ (X) the equation [X] = f∗[C[n−1,n]] + g∗[C[n]] holds. Proof. It is easy to check on the level of points that X = f (C[n−1,n]) ∪ g(C[n]), and that f and g are both injective. As C[n] and C[n−1,n] are both irreducible by Lemma 6.5, we get that X = f (C[n−1,n]) ∪ g(C[n]) is the decomposition of X into irreducible components. Furthermore, on the complement of f (C[n−1,n]) ∩ g(C[n]) one checks that f and g are local isomorphisms of schemes. The claim follows. It follows that we have eι · θ = [X] = f∗([C[n−1,n]]) + g∗([C[n]]) = f∗(θ′ · ι′) + g∗(1), 13 where 1 is the unit element in H 0(C[n] id→ C[n]). Using this we now compute µ−[p]µ+[C](α) = (eq0)∗(eι0 · θ0 · α) = (eq0)∗(f∗(θ′ · ι′)0 · α) + (eq0)∗(g∗(1)0 · α) 0 · ι′ 0 · α) + (id)∗(α) = µ+[C]µ−[p](α) + α, 0 · α) + (eq0 ◦ g0)∗(α) 0 · ι′ = (eq0 ◦ f0)∗(θ′ 0)∗(θ′ = (q′ (5) which is what we wanted to show. The proof of [µ−[C], µ+[pt]] = id is similar to the above case. Here we have µ−[C]µ+[pt](α) = (p0)∗(κ0 · (i0)∗(α)) = (p0 ◦ei0)∗(eκ0 · α) and µ+[pt]µ−[C](α) = (i′ 0 ◦ p′ 0)∗(κ′ 0 · α). Under the identification of H ∗(X follows from eq → C[n]) with H BM ∗+2 dim C[n](X) we have eκ = [X]. This eκ · [C[n]] =eκ · ι · [C[n+1]] =eι · κ[C[n+1]] =eι · [C[n,n+1]] = [X], where the last equality is obtained as in the proof of Lemma 5.1. Using Lemma 5.2 we get A computation similar to (5) now shows µ−[C]µ+[pt](α) = µ+[pt]µ−[C](α)+α as needed. eκ = [X] = f∗[C[n−1,n]] + g∗[C[n]] = f∗(κ′) + g∗(1). 5.2 Proof of [µ+[C], µ−[C]] = 0 The relevant diagrams are X eκ ysssssssss %❏❏❏❏❏❏❏❏❏ q eq %❑❑❑❑❑❑❑❑❑ yttttttttt κ C[n,n+1] $❍❍❍❍❍❍❍❍❍❍ p C[n,n+1] θ z✈✈✈✈✈✈✈✈✈✈ C[n+1] Y eθ ysssssssss %❏❏❏❏❏❏❏❏❏ p′ ep %❑❑❑❑❑❑❑❑❑ yttttttttt θ′ C[n−1] C[n−1,n] κ′ z✈✈✈✈✈✈✈✈✈✈ C[n−1,n] $■■■■■■■■■ q′ (6) (7) C[n] C[n]. and C[n] C[n] Here X = C[n,n+1] ×C[n+1] C[n,n+1], Y = C[n−1,n] ×C[n−1] C[n−1,n], and the squares con- taining X and Y are Cartesian. The bivariant classes θ, κ, θ′, κ′ are the ones induced by 14 y % z % y $ y % z % y $ fundamental classes, while eθ and eκ are the Cartesian pullbacks of θ′ and κ, respectively. As in Section 5.1, a base change to the central fibre C = C0 is denoted by a subscript 0. Let α ∈ H∗(C [n]). We then have µ−[C]µ+[C](α) = (p0)∗(κ0 · (q0)∗(θ0 · α)) = (p0 ◦eq0)∗(eκ0 · θ0 · α) and µ+[C]µ−[C](α) = (q′ 0)∗(θ′ 0 · (p′ 0)∗(κ′ 0 · α)) = (q′ 0 ◦ep0)∗(eθ0 · κ′ 0 · α). Lemma 5.3. The scheme X is equidimensional, and the scheme Y is irreducible. Both are generically reduced, and dim X = dim Y = dim C[n+1]. Proof. As X = C[n,n+1] ×C[n+1] C[n,n+1], every irreducible component has dimension at least 2 dim C[n,n+1] − dim C[n+1] = dim C[n+1]. Let ∆ ⊂ B be the discriminant locus, i.e. the set of b ∈ B such that Cb is singular. By Lemma 6.3 (iii) we have dim X∆ = dim ∆ + n + 1 ≤ (dim B − 1) + n + 1 < dim C[n+1]. It follows that X \ X∆ is dense in X. Write X = {(Z1, Z2, Z3) ∈ C[n] ×B C[n+1] ×B C[n] Z1, Z3 ⊂ Z2}. Let X1 ⊂ X be the locus where Z1 = Z3, and let X2 = X \ X1. It is then easy to check that X1 ∩ (X \ X∆) and X2 ∩ (X \ X∆) are irreducible, generically nonsingular, and of dimension equal to dim C[n+1]. This proves the claims for X. Arguing similarly for Y , using Lemma 6.3 (iv) we find that Y \ Y∆ is dense in Y . There is a morphism C[n−1,n] → C taking a pair (Z, Z ′) to the point where Z and Z ′ differ. Using this we get a map Y = C[n−1,n] ×C[n−1] C[n−1,n] → C ×B C[n−1] ×B C. One checks that restricting both source and target to the locus of nonsingular curves this map is an isomorphism, hence Y \ Y∆ is isomorphic to (C ×B C[n−1] ×B C) \ (C ×B C[n−1] ×B C)∆. In particular Y \ Y∆ is nonsingular and irreducible of dimension equal to dim C[n+1], and the claims for Y follow. Let π : X → C[n] and π′ : Y → C[n] be the natural maps going down the left hand side of diagrams (6) and (7), respectively. Lemma 5.4. Identifying H(X π→ C[n]) with H BM H(Y π′ → C[n]) with H BM ∗ ∗ (Y ) gives eθ · κ′ = [Y ]. (X) gives eκ · θ = [X]. Identifying 15 Proof. We treat the case of X; the case of Y is similar. We must show that eκ · θ · [C[n]] = [X], and as θ·[C[n]] = [C[n,n+1]], it suffices to showeκ·[C[n,n+1]] = [X]. The classeκ·[C[n,n+1]] can be identified with the refined intersection product [C[n,n+1] × C[n,n+1]] ∩ ∆ ∈ H BM ∗ (X), where we intersect the classes inside C[n+1] × C[n+1], and ∆ denotes the diagonal in this space. As X is generically reduced, the intersection multiplicity at each component is 1, by [Ful98, Prop. 8.2], and so this intersection product equals [X]. Let f : X → C[n] ×B C[n] and g : Y → C[n] ×B C[n] be the maps induced by composing down both sides of diagrams (6) and (7), respectively. Lemma 5.5. In H BM ∗ (C[n] ×B C[n]), the equality f∗[X] = g∗[Y ] holds. Proof. Let X = X1 ∪ X2 be the decomposition of X into irreducible components, where X1 and X2 are as in the proof of Lemma 5.3. By definition of X1 the image f (X1) is contained in the diagonal C[n] ⊂ C[n] ×B C[n]. Hence dim f (X1) < dim C[n+1] = dim X1, and so f∗[X1] = 0. Let U =(cid:0)C[n] ×B C[n](cid:1) \ C[n]. We claim that over U the maps f X2 and g are injective with the same image. To see this, note that if (Z1, Z3) ∈ U , then Z1 6= Z3, and so (Z1, Z3) ∈ f (X2) ⇔ (Z1, Z1 ∪ Z3, Z3) ∈ X2 ⇔ l(Z1 ∪ Z3) = n + 1 ⇔ l(Z1 ∩ Z3) = n − 1 ⇔ (Z1, Z1 ∩ Z3, Z3) ∈ Y ⇔ (Z1, Z3) ∈ g(Y ). As both X2 and Y are generically reduced, it follows that f∗[X] = f∗[X2] = g∗[Y ]. Let π1, π2 : C[n] ×B C[n] → C[n] be the projections. Combining Lemmas 5.4 and 5.5 shows that in H(C[n] ×B C[n] π1→ C[n]) we have f∗(eκ · θ) = g∗(eθ · κ′). Let α ∈ H∗(C [n]), and compute µ−[C]µ+[C](α) = (p0 ◦eq0)∗(eκ0 · θ0 · α) = ((π2)0 ◦ f0)∗(eκ0 · θ0 · α) = ((π2)0)∗((f0)∗(eκ0 · θ0) · α) = ((π2)0)∗((g0)∗(eθ0 · κ′ = ((π2)0 ◦ g0)∗(eθ0 · κ′ 0 ◦ep0)∗(eθ0 · κ′ = µ+[C]µ−[C](α), 0) · α) 0 · α) = (q′ 0 · α) which is what we wanted. 16 5.3 Proof of [µ±[pt], µ±[C]] = 0 Consider the diagram C [n,n+1] i′′ / C [n+1,n+2] p z✉✉✉✉✉✉✉✉✉✉ %❑❑❑❑❑❑❑❑❑❑ q p′ xrrrrrrrrrr q′ &▼▼▼▼▼▼▼▼▼▼ C [n] i / C [n+1] i′ / C [n+2]. We have and µ+[pt]µ+[C] = i′ ∗q∗p! = q′ ∗i′′ ∗p! = q′ ∗(p′)!i∗ = µ+[C]µ+[pt] µ−[C]µ−[pt] = p∗q!(i′)! = p∗(i′′)!(q′)! = i!p′ ∗(q′)! = µ−[pt]µ−[C], where the required compatibilities are easily checked. 5.4 Proof of [µ−[pt], µ+[pt]] = 0 For α ∈ H∗(C [n]), we have µ+[pt]µ−[pt](α) = i∗(i!(α)) = [i(C [n−1])] ∩ α, where [i(C [n−1])] ∈ H 2(C [n]) is the class of the Cartier divisor C [n−1]. On the other hand, µ−[pt]µ+[pt](α) = i!(i∗(α)) = i∗[i(C [n])] ∩ α. It thus suffices to show the equality [i(C [n−1])] = i∗[i(C [n])] in H 2(C [n]). For any nonsingular point y ∈ C, let iy : C [n] → C [n+1] be defined by adding a point at y, so that we have i = ix for our chosen point x. For any y 6= x we have [i(C [n−1])] = i∗ y[i(C [n])], in H 2(C [n]), which follows from the corresponding equality of Cartier divisors. As i∗ the claim follows.3 y = i∗ 6 Flag Hilbert schemes In this section we prove some dimension estimates for the flag Hilbert schemes C [n,n+1] and related schemes. Let Hn ⊂ (A2)[n] be the set of Z ∈ (A2)[n] such that Z is supported at 0 ∈ A2. Similarly, let Hn,n+1 ⊂ (A2)[n,n+1] be the set of pairs (Z, Z ′) ∈ (A2)[n] × (A2)[n+1] such that Z ⊂ Z ′ and both are supported at 0. We follow the convention that dim ∅ = −1. 3For the case of Chow homology we need the equality [i(C [n−1])] = i∗[i(C [n])] in Pic(C [n]). At the level y = i∗, but the relation still holds by noting the of rational equivalence, it is no longer true that i∗ equality of Cartier divisors [i(C [n−1])] = i∗ y[i(C [n])] and then letting y tend to x. 17 z % / x & / / Lemma 6.1. We have dim Hn,n+1 = n for all n ≥ 0. Proof. For any Z ∈ Hn, let d−(Z) = dim{Z ′ ∈ Hn−1 Z ′ ⊂ Z} and let d+(Z) = dim{Z ′ ∈ Hn+1 Z ⊂ Z ′}. We then have d+(Z) = d−(Z) + 1 (8) for all Z, see [ES98, Sec. 3]. Let Vn,k ⊆ Hn be the set of Z ∈ Hn such that d−(Z) = k. Using (8) we find max k {dim Vn,k + k + 1} = dim Hn,n+1 = max k {dim Vn+1,k + k}. (9) From (9) we find Hn+1,n+2 = Hn,n+1 + 1, hence the claim of the lemma follows by induction from dim H0,1 = 0. Lemma 6.2. For all n ≥ 0, we have (i) dim Hn,n+1 ×Hn+1 Hn,n+1 = n. (ii) dim Hn,n+1 ×Hn Hn,n+1 = n + 1, unless n = 0, in which case dim H0,1 ×H0 H0,1 = 0. Proof. For P = i, ii and n ≥ 0, let (P )n denote the claim that equation (P ) holds for the given value of n. We will prove the claims by induction, starting from the trivial cases (i)0 and (ii)0. Let Xn and Yn denote the schemes appearing on the left hand side of (i) and (ii), respectively. (ii)n−1 =⇒ (i)n: The diagonal map defines an inclusion Hn,n+1 ֒→ Xn, whence by Lemma 6.1 we have dim Xn ≥ n, and it suffices to show that dim(Xn \ Hn,n+1) ≤ n. The set of points of Xn is {(Z1, Z2, Z3) ∈ Hn × Hn+1 × Hn Z1, Z3 ⊂ Z2}, and Xn \ Hn,n+1 is the locus of triples (Z1, Z2, Z3) where Z1 6= Z3. For such triples we must have Z2 = Z1 ∪ Z3. Let l(Z) denote the length of Z. Using the relation l(Z1 ∪ Z3) = l(Z1) + l(Z3) − l(Z1 ∩ Z3) we get bijections Xn \ Hn,n+1 = {(Z1, Z3) ∈ Hn × Hn l(Z1 ∪ Z3) = n + 1} = {(Z1, Z3) ∈ Hn × Hn l(Z1 ∩ Z3) = n − 1} = {Z1, Z1 ∩ Z3, Z3} ⊆ Yn−1, hence dim(Xn \ Hn,n+1) ≤ dim Yn−1 ≤ n, by our assumption (ii)n−1. (i)n−1 =⇒ (ii)n: Let d−, d+ : Hn → Z and Vn,k = d−1 − (k) ⊆ Hn be as in the proof of Lemma 6.1. We write as above Xn−1 = {(Z1, Z2, Z3) ∈ Hn−1 × Hn × Hn−1 Z1, Z3 ⊂ Z2}. For any Z ∈ Hn, the fibre over Z under the projection Xn−1 → Hn has dimension 2d−(Z). It follows that the locus in Xn−1 such that Z2 ∈ Vn,k has dimension dim Vn,k + 2k. 18 Similarly Yn = {(Z ′ 1, Z ′ 2, Z ′ 3) ∈ Hn+1 × Hn × Hn+1 Z ′ 2 ⊂ Z ′ 1, Z ′ 3}, and the fibre over Z ∈ Hn under the projection Yn → Hn has dimension 2d+(Z) = 2d−(Z)+2, by (8). Hence the locus in Yn where Z ′ 2 ∈ Vn,k has dimension dim Vn,k +2k+2. We get dim Xn−1 = max k {dim Vn,k + 2k} = max k {dim Vn,k + 2k + 2} − 2 = dim Yn − 2, hence by the induction assumption (i)n−1 we get dim Yn = dim Xn−1 + 2 = n + 1. This concludes the induction procedure. Lemma 6.3. Let C be a locally planar reduced curve, and let Csm ⊆ C be its nonsingular locus. For all n ≥ 0, we have (i) dim C [n,n+1] = n + 1. (ii) dim(C [n,n+1] \ (Csm)[n,n+1]) < n + 1. (iii) dim C [n,n+1] ×C[n+1] C [n,n+1] = n + 1. (iv) dim C [n,n+1] ×C[n] C [n,n+1] = n + 2. Proof. For points (i), (iii) and (iv) the claim LHS ≥ RHS is straightforward to see by replacing C with Csm, hence it suffices to prove the claim LHS ≤ RHS. We shall only prove the claim (iv); the other three claims can be handled by similar arguments. Let X = C [n,n+1] ×C[n] C [n,n+1], and write X = {(Z1, Z2, Z3) ∈ C [n+1] × C [n] × C [n+1] Z2 ⊂ Z1, Z3}. Let {x1, . . . , xk} be the set of singular points of C. We partition X into disjoint subsets X(a0, . . . , ak, r, s), where the ai are non-negative integers whose sum is n, and where r, s are integers such that 0 ≤ r, s ≤ k. The subset X(a0, . . . , ak, r, s) parametrises (Z1, Z2, Z3) satisfying the two conditions 1. Z2 has support of length a0 over the smooth locus of C and of length ai at the point xi for i > 0. 2. The scheme Z1 (resp. Z3) differs from Z2 at point xr if r > 0 (resp. xs if s > 0), and differs at a smooth point of C if r = 0 (resp. s = 0). Let Csm ⊂ C be the nonsingular locus. Using the local planarity of C we see that X(a0, . . . , ak, r, s) is isomorphic to a subset of one of the following schemes, depending on r and s. • r = s = 0 : ((Csm)[a0,a0+1] ×(Csm)[a0] (Csm)[a0,a0+1]) × Ha1 × · · · × Hak . • r = s 6= 0 : (Csm)[a0] × Ha1 × . . . × (Har ,ar+1 ×Har Har ,ar+1) × . . . × Hak . 19 • r 6= s = 0 : (Csm)[a0,a0+1] × Ha1 × . . . × Har ,ar+1 × . . . × Hak . • 0 6= r 6= s 6= 0 : (Csm)[a0] × Ha1 × . . . × Har ,ar+1 × . . . × Has,as+1 × . . . × Hak . Using Lemmas 6.1, 6.2 and the fact that dim Hm = max(0, m − 1) (see [Iar72]), we find that each of the above listed schemes has dimension ≤ a0 + · · · + ak + 2 = n + 1. The claim of the lemma follows. Lemma 6.4. Let C be a locally planar reduced curve. Then (Csm)[n,n+1] is dense in C [n,n+1]. Proof. Kleiman and Altman [KA79] have shown that we may embed C in a nonsingular quasiprojective surface S. By work of Cheah and Tikhomirov [Che98, Tik97], the scheme S[n,n+1] is nonsingular of dimension 2n + 2. Because of the Cartesian diagram C [n,n+1] S[n,n+1] C [n+1] / S[n+1], every irreducible component of C [n,n+1] has dimension at least equal to dim C [n+1] + dim S[n,n+1] − dim S[n+1] = n + 1. By Lemma 6.3 (ii) we have dim(C [n,n+1] \ (Csm)[n,n+1]) < n + 1. It follows that the open subset (Csm)[n,n+1] intersects every irreducible component in C [n,n+1], hence it is dense as claimed. Let C → B be a family of curves satisfying the hypotheses of Section 2.1, that is, C[n] is nonsingular and every curve in the family is irreducible and reduced. Lemma 6.5. The relative flag Hilbert scheme C[n,n+1] is irreducible of dimension equal to dim C[n+1]. Proof. Let U ⊆ C be the locus of q ∈ C such that q ∈ Cb with Cb smooth at q. As every curve in the family is irreducible, we get that U [n,n+1] ⊂ C[n,n+1] is irreducible. Now for every fibre Cb, we have that C[n,n+1] , by Lemma 6.4. Hence U [n,n+1] is dense in C[n,n+1], and the claim follows. ∩ U [n,n+1] is dense in C[n,n+1] b b Using the techniques of [She12] and the fact that for a smooth surface S the variety S[n,n+1] is nonsingular (see [Che98, Tik97]) one can show that C[n,n+1] is nonsingular. As we do not need this stronger statement, we omit the proof. 20 / / / 7 The D-grading splits the perverse filtration In this section we relate the D-grading on H ∗(J) to the perverse filtration appearing in [MY14, MS13]. We use the notation of Section 1.3.1, namely V c(C) = ⊕n≥0H ∗(C [n]), the operators µc +[C]). The space V c(C) is equipped with a grading D, by letting ±[C] act on V c(C), and W c = V c(C)/(im µc +[pt] + im µc ±[pt] and µc DnV c(C) = H ∗(C [n]), and the spaces W c and H ∗(J) inherit this grading using the isomorphisms of Theorem 1.3. Recall also the formula H ∗(C [n]) ∼= Mm≤n DmH ∗(J) ⊗ Symn−m (Q ⊕ Q[−2]) . (10) We fix as usual the versal deformation family C → B and let 0 ∈ B be the point such that C0 = C. Following [MY14], we define the perverse filtration P≤j on H ∗(J) as follows. The versal deformation family f : C → B induces a family of compactified ≤jRfJ ∗QJ Jacobians fJ : J → B. The object RfJ ∗QJ ∈ Db −1(0) = J, and so if g is the induced by the perverse t-structure on Db inclusion 0 ֒→ B, we naturally have g∗(RfJ ∗QJ ) = H ∗(J). We may now define the perverse filtration by c(B) has a filtration τ p c(B). We have fJ P≤jH ∗(J) = Im(cid:16)g∗(cid:16)τ p ≤jRfJ ∗QJ(cid:17) → g∗ (RfJ ∗QJ ) = H ∗(J)(cid:17) . Replacing J and J with C [n] and C[n] in the above construction we get a filtration P≤j on H ∗(C [n]) as well. For X = J or X = C [n] we normalise the indices of the perverse filtration by letting It follows that P≤−1H ∗(X) = 0 and letting 1 ∈ H 0(X) be contained in P≤0H ∗(X). grP i (H ∗(X)) = 0 unless 0 ≤ i ≤ 2 dim X. Comparing the formula of [MY14, Thm. 1.1] with (10) we find an isomorphism grP • H ∗(J) ∼= D•H ∗(J) of bigraded vector spaces. In other words the filtrations P≤n and D≤n on H ∗(J) have isomorphic associated graded objects. The remainder of this section is devoted to showing that the filtrations are in fact equal. Proposition 7.1. D≤nH ∗(J) = P≤nH ∗(J). Let X = C [n] or X = J. We define the filtration Q≤j on H ∗(X) by Q≤jH i(X) = P≤i+jH i(X). Lemma 7.2. The maps µc ±[pt], µc ±[C] and AJ ∗ all preserve the Q-filtration. Proof. The statement follows from the fact that each of the maps is the restriction to 0 of a map of complexes on B. We will give the details for µc +[C]; the remaining cases are similar and left to the reader. +[pt] and µc 21 For µc +[pt], we have the following diagram C[n] !❇❇❇❇❇❇❇❇ f [n] i B C[n+1] ②②②②②②②②② f [n+1] . Since C[n] and C[n+1] are nonsingular, we have i!(QC[n+1]) = QC[n][−2]. By adjunction we get a map i∗QC[n] → QC[n+1][2], which we push down to get a map Rf [n] ∗ QC[n] = R(f [n+1]i)∗QC[n] → Rf [n+1] ∗ QC[n+1][2]. One can check that the restriction of this map to 0 ∈ B agrees with µc +[pt]. Now, the composed map ≤jRf [n] τ p ∗ QC[n] → Rf [n] ∗ QC[n] → Rf [n+1] ∗ QC[n+1][2] factors through τ p P≤jH ∗(C[n]) to P≤j+2H ∗+2(C[n+1]), which is the same as saying µc Q-filtration. QC[n+1][2]) → Rf [n+1] ≤j(Rf [n+1] QC[n+1][2]. It follows that µc +[pt] sends +[pt] preserves the ∗ ∗ For the case of µc +[C], we have the diagram C[n,n+1] C[n] p z✈✈✈✈✈✈✈✈✈✈ $■■■■■■■■■■ f [n] q %❏❏❏❏❏❏❏❏❏ ysssssssssss f [n+1] C[n+1] B . Using the identification of H BM ∗ (C[n,n+1]) with R−∗Γ(Q∨ C[n,n+1]) = HomD(C[n,n+1])(QC[n,n+1], Q∨ C[n,n+1][−∗]), the fundamental class of C[n,n+1] induces a map QC[n,n+1] → Q∨ C[n,n+1][−2 dim C[n,n+1]] = q!QC[n+1]. Hence we get a map Rq∗QC[n,n+1] → QC[n+1], and thus a composed map Rf [n] ∗ QC[n] → R(f [n]p)∗QC[n,n+1] = R(f [n+1]q)∗QC[n,n+1] → Rf [n+1] ∗ QC[n+1]. Again one can check that the restriction of this map to 0 ∈ B equals µc argument as for µc +[pt], we then get that µc +[C] preserves the Q-filtration. +[C]. By the same 22 / / ! z % $ y As a consequence of the above lemma, the operators µc ±[C] act on grQV c(C). Since they still obey the Weyl algebra commutation relations when acting on this space, the proof of the implication (i) ⇒ (ii) in Theorem 1.2 applies to show that ±[pt] and µc grQW c ⊗ Q[µc +[pt], µc +[C]] ∼= grQV c(C). (11) This is an isomorphism of (H, Q, D)-graded spaces, where µc (2, 0, 1) and (0, 0, 1), respectively. +[pt] and µc +[C] have degrees Let DnH iW c be the image of H i(C [n]) under the map V c(C) → W c. Define three generating functions FV (x, y, z) =Xi,j,n FW (x, y, z) =Xi,j,n FJ (x, y, z) =Xi,j dim grQ j H i(C [n])xiyjzn dim grQ j DnH iW cxiyjzn dim grQ j H i(J)xiyjzi+j. The isomorphism (11) implies FW · (1 − z)−1(1 − x2z)−1 = FV . It follows from [MS13, Cor. 2] that FJ · (1 − z)−1(1 − x2z)−1 = FV , and hence FJ = FW . As a consequence, the coefficient of xiyjzn in FW is 0 unless n = i + j, and it follows that D≤i+jH iW c = Q≤jH iW c. (12) From FJ = FW we get that dim grQ j W c for all j. Since AJ ∗ : H ∗(J) → W c is an isomorphism and preserves the Q-filtration, it must then be an isomorphism of Q-filtered spaces. Obviously AJ ∗ preserves the cohomological grading. The D-grading on H ∗(J) is defined via AJ ∗ and so preserved by definition, hence by (12) we have j H ∗(J) = dim grQ D≤i+jH i(J) = Q≤jH i(J) = P≤i+jH i(J), and the proof of Proposition 7.1 is complete. References [ACG11] E. Arbarello, M. Cornalba, and P. A. Griffiths. Geometry of algebraic curves. Volume II, volume 268 of Grundlehren der Mathematischen Wissenschaften. Springer, Heidelberg, 2011. With a contribution by J. D. Harris. 23 [AIK77] A. B. Altman, A. Iarrobino, and S. L. Kleiman. Irreducibility of the compact- ified Jacobian. In Real and complex singularities (Proc. Ninth Nordic Summer School/NAVF Sympos. Math., Oslo, 1976), pages 1 -- 12. Sijthoff and Noordhoff, Alphen aan den Rijn, 1977. [AK80] A. B. Altman and S. L. Kleiman. Compactifying the Picard scheme. Advances in Mathematics, 35(1):50 -- 112, 1980. [BGS81] J. Briançon, M. Granger, and J.-P. Speder. Sur le schéma de Hilbert d'une courbe plane. Ann. Sci. École Norm. Sup. (4), 14(1):1 -- 25, 1981. [Che98] J. Cheah. Cellular decompositions for nested Hilbert schemes of points. Pacific J. Math., 183(1):39 -- 90, 1998. [ES98] G. Ellingsrud and S. A. Strømme. An intersection number for the punctual Hilbert scheme of a surface. Trans. Amer. Math. Soc., 350(6):2547 -- 2552, 1998. [FM81] W. Fulton and R. MacPherson. Categorical framework for the study of singular spaces. Mem. Amer. Math. Soc., 31(243):vi+165, 1981. [Ful98] W. Fulton. Intersection theory, volume 2 of Ergebnisse der Mathematik und ihrer Grenzgebiete. 3. Folge. A Series of Modern Surveys in Mathematics. Springer-Verlag, Berlin, second edition, 1998. [Gro96] I. Grojnowski. Instantons and affine algebras. I. The Hilbert scheme and vertex operators. Math. Res. Lett., 3(2):275 -- 291, 1996. [GV98a] R. Gopakumar and C. Vafa. M-theory and topological strings -- I. arXiv e-print hep-th/9809187, 1998. [GV98b] R. Gopakumar and C. Vafa. M-theory and topological strings -- II. arXiv e-print hep-th/9812127, 1998. [Iar72] A. Iarrobino. Punctual Hilbert schemes. Bull. Amer. Math. Soc., 78(5):819 -- 823, 1972. [KA79] S. L. Kleiman and A. B. Altman. Bertini theorems for hypersurface sections containing a subscheme. Comm. Algebra, 7(8):775 -- 790, 1979. [MP10] B. Moonen and A. Polishchuk. Algebraic cycles on the relative symmetric powers and on the relative Jacobian of a family of curves. II. J. Inst. Math. Jussieu, 9(4):799 -- 846, 2010. [MS13] L. Migliorini and V. Shende. A support theorem for Hilbert schemes of planar curves. J. Eur. Math. Soc., 15(6):2353 -- 2367, 2013. [MY14] D. Maulik and Z. Yun. Macdonald formula for curves with planar singularities. J. Reine Angew. Math., 694:27 -- 48, 2014. 24 [Nak97] H. Nakajima. Heisenberg algebra and Hilbert schemes of points on projective surfaces. Annals of Mathematics, 145(2):pp. 379 -- 388, 1997. [PT10] R. Pandharipande and R. P. Thomas. Stable pairs and BPS invariants. J. Amer. Math. Soc., 23(1):267 -- 297, 2010. [PT14] R. Pandharipande and R. P. Thomas. 13/2 ways of counting curves. In Moduli spaces, volume 411 of London Math. Soc. Lecture Note Ser., pages 282 -- 333. Cambridge Univ. Press, Cambridge, 2014. [She12] V. Shende. Hilbert schemes of points on a locally planar curve and the Severi strata of its versal deformation. Compos. Math., 148(2):531 -- 547, 2012. [Tik97] A. S. Tikhomirov. The variety of complete pairs of zero-dimensional subschemes of an algebraic surface. Izv. Ross. Akad. Nauk Ser. Mat., 61(6):153 -- 180, 1997. 25
1604.08662	1	1604	2016-04-29T01:01:00	Triangulated endofunctors of the derived category of coherent sheaves which do not admit DG liftings	[ "math.AG", "math.KT" ]	Recently, Rizzardo and Van den Bergh constructed an example of a triangulated functor between the derived categories of coherent sheaves on smooth projective varieties over a field $k$ of characteristic $0$ which is not of the Fourier-Mukai type. The purpose of this note is to show that if $char \, k =p$ then there are very simple examples of such functors. Namely, for a smooth projective $Y$ over $\mathbb Z_p$ with the special fiber $i: X\hookrightarrow Y$, we consider the functor $L i^* \circ i_: D^b(X) \to D^b(X)$ from the derived categories of coherent sheaves on $X$ to itself. We show that if $Y$ is a flag variety which is not isomorphic to $\mathbb P^1$ then $L i^ \circ i_$ is not of the Fourier-Mukai type. Note that by a theorem of Toen (\cite{t}, Theorem 8.15) the latter assertion is equivalent to saying that $L i^ \circ i_$ does not admit a lifting to a $\mathbb F_p$-linear DG quasi-functor $D^b_{dg}(X) \to D^b_{dg}(X)$, where $D^b_{dg}(X)$ is a (unique) DG enhancement of $D^b(X)$. However, essentially by definition, $L i^ \circ i_*$ lifts to a $\mathbb Z_p$-linear DG quasi-functor.	math.AG	math	TRIANGULATED ENDOFUNCTORS OF THE DERIVED CATEGORY OF COHERENT SHEAVES WHICH DO NOT ADMIT DG LIFTINGS VADIM VOLOGODSKY Abstract. In [RV], Rizzardo and Van den Bergh constructed an example of a triangulated functor between the derived categories of coherent sheaves on smooth projective varieties over a field k of characteristic 0 which is not of the Fourier-Mukai type. The purpose of this note is to show that if char k = p then there are very simple examples of such functors. Namely, for a smooth projective Y over Zp with the special fiber i : X ֒→ Y , we consider the functor ◦ i∗ : Db(X) → Db(X) from the derived categories of coherent sheaves on Li∗ X to itself. We show that if Y is a flag variety which is not isomorphic to P1 ◦ i∗ is not of the Fourier-Mukai type. Note that by a theorem of Toen then Li∗ ([T], Theorem 8.15) the latter assertion is equivalent to saying that Li∗ ◦ i∗ dg(X) → Db does not admit a lifting to a Fp-linear DG quasi-functor Db dg(X), dg(X) is a (unique) DG enhancement of Db(X). However, essentially by where Db definition, Li∗ ◦ i∗ lifts to a Zp-linear DG quasi-functor. Given smooth proper schemes X1, X2 over a field k and an object E ∈ Db(X1 ×X2) of the bounded derived category of coherent sheaves on X1 × X2 define a triangulated functor (0.1) ΦE : Db(X1) → Db(X2) sending a bounded complex M of coherent sheaves on X1 to Rp2∗(E 1M ), where pi : X1 × X2 → Xi are the projections. Recall that a triangulated functor Db(X1) → Db(X2) is said to be of the Fourier-Mukai type if it is isomorphic to ΦE for some E. Let Y be a smooth projective scheme over Spec Zp and let X be its special fiber, i : X ֒→ Y the closed embedding. Consider the triangulated functor G : Db(X) → Db(X) given by the formula L ⊗ p∗ We shall see that in general G is not of the Fourier-Mukai type. G = Li∗ ◦ i∗ Theorem 1. Let Z a smooth projective scheme over Spec Zp, Y = Z × Z, X = Y ×Spec Zp Spec Fp . Assume that (1) The Frobenius morphism F r : Z → Z, where Z = Z × Spec Fp, does not lift modulo p2. (2) H 1(X, TX ) = 0, where TX is the tangent sheaf on X. Then G = Li∗ ◦ i∗ : Db(X) → Db(X) is not of the Fourier-Mukai type. For example, let GLn be the general linear group over Spec Zp, B ⊂ GLn a Borel subgroup. Then, by Theorem 6 from [BTLM], for any n > 2, the flag variety Z = GLn/B satisfies the first assumption of the Theorem i.e., the Frobenius F r : Z → Z does not lift on Z × Spec Z/p2Z. By ([KLT], Theorem 2), we have that H 1(Z, TZ) = 1 2 VADIM VOLOGODSKY H 1(Z, OZ) = 0. It follows that H 1(X, TX ) = 0. Hence, by the Theorem, for n > 2, G : Db(X) → Db(X) is not of the Fourier-Mukai type. Proof. Assume the contrary and let E ∈ Db(X × X) be the Fourier-Mukai kernel. By definition, for every M ∈ Db(X) we have a functorial isomorphism (0.2) G(M ) ∼ −→ Rp2∗(E L ⊗ p∗ 1M ). By the projection formula ([H], Chapter II, Prop. 5.6) we have that p L ⊗ i∗(OX ) ∼ −→ OY ) ∼ −→ i∗(M ) −→ i∗(M ) ⊗ (OY ∼ i∗ ◦ Li∗ ◦ i∗(M ) −→ i∗(M ) ⊕ i∗(M )[1] In particular, if M is a coherent sheaf then H i(G(M )) ≃ M for i = 0, −1 and H i(G(M )) = 0 otherwise. Applying this observation and formula (0.2) to skyscraper sheaves, M = δx, x ∈ X(Fp), we conclude that the coherent sheaves H i(E) are set theoretically supported on the diagonal ∆X ⊂ X × X. Applying the same formulas to M = OX we see that p2∗(H i(E)) = OX for i = 0, −1 and p2∗(H i(E)) = 0 otherwise. In fact, every coherent sheaf F on X × X which is set theoretically supported on the diagonal and such that p2∗F = OX is isomorphic to O∆X . It follows that H 0(E) = H −1(E) = O∆X . In the other words, E fits into an exact triangle in Db(X × X) −→ E −→ O∆X O∆X [1] α (0.3) for some β ∈ Ext2 (O∆X , O∆X ). We wish to show that the second assumption in the Theorem implies that β = 0, while the first one implies that β 6= 0. For every M ∈ Db(X), (0.3) gives rise to an exact triangle αM−→ G(M ) −→ M −→ O∆X [2] βM−→ M [2] M [1] (0.4) OX×X β Our main tool is the following result. Lemma 0.1. For a coherent sheaf M the following conditions are equivalent. (1) βM = 0. (2) G(M ) (3) There exists a morphism λ : G(M ) → M [1] such that λ ◦ αM is an isomor- ∼ −→ M ⊕ M [1]. phism. (4) M admits a lift modulo p2 i.e., there is a coherent sheaf M on Y flat over Z/p2Z such that i∗ M ≃ M . Proof. The equivalence of (1), (2) and (3) is immediate. Let us check that (3) is equivalent to (4). A morphism λ : G(M ) → M [1] gives rise by adjunction a morphism γ : i∗M → i∗M [1]. Note that M = cone γ[1] is a coherent sheaf on Y which is an extension of i∗M by itself. It suffices to prove that λ ◦ αM : M [1] → M [1] is an isomorphism if and only if M is flat over Z/p2Z. Indeed, from the exact triangle Li∗i∗M → Li∗( M ) → Li∗i∗M → Li∗i∗M [1] we get a long exact sequence of the cohomology sheaves 0 → M → L1i∗( M ) → M λ◦αM−→ M → i∗( M ) → M → 0 Thus λ ◦ αM is an isomorphism if and only if in the exact sequence 0 → i∗M → M → i∗M → 0 the image of second map is the kernel of the multiplication by p on M and also the image of this map. The latter is equivalent to flatness of M over Z/p2Z. (cid:3) TRIANGULATED ENDOFUNCTORS OF THE DERIVED CATEGORY OF COHERENT SHEAVES WHICH DO NOT ADMIT DG LIFTINGS3 We have a spectral sequence converging to Ext∗ OX×X (O∆X , O∆X ) whose second page is H ∗(X, Ext∗ OX×X (O∆X , O∆X )). In particular, we have a homomorphism Ext2 OX×X (O∆X , O∆X ) → H 0(X, Ext2 OX×X (O∆X , O∆X )) ∼ −→ H 0(X, ∧2TX ). Let us check that the image µ of β under this map is 0. To do this we apply the Lemma to skyscraper sheaves δx, where x runs over closed points of X. On the one hand, the evaluation of the bivector field µ at x is equal to the class of βδx in Ext2 −→ ∧2Tx,X. On the other hand, by the Lemma, βδx = 0 since δx is liftable modulo p2. Next, the assumption that H 1(X, TX) = 0 implies that β lies in the image of the map (0.5) v : H 2(X, OX ) OX×X (O∆X , O∆X )) → Ext2 OX×X (O∆X , O∆X ). (δx, δx) ∼ −→ H 2(X, Ext0 OX ∼ (O∆X , O∆X ) → H 2(X, OX ) which takes The map (0.5) has a left inverse u : Ext2 β to βOX . But, by the Lemma, the later class is equal to 0 since OX is liftable modulo p2. It follows that β is 0. OX×X On the other hand, let Γ ⊂ X = Z × Z be the graph of the Frobenius morphism F r : Z → Z and OΓ the structure sheaf of Γ viewed as a coherent sheaf on X. Then, by our first assumption, the sheaf OΓ is not liftable modulo p2. Hence, by the Lemma, βOΓ is not 0. This contradiction completes the proof. (cid:3) Acknowledgements. I would like to thank Alberto Canonaco and Paolo Stellari: their interest prompted writing this note. Also, I am grateful to Alexander Samokhin for stimulating discussions and references. References [BTLM] A. Buch, J. F. Thomsen, N. Lauritzen, and V. Mehta, The Frobenius morphism on a toric variety, Tohoku Math. J. (2), Volume 49, Number 3 (1997), 355-366. [CV] D. Calaque, M. Van den Bergh, Hochschild cohomology and Atiyah classes, Adv. Math. 224 (2010), no. 5, 1839 -- 1889. R. Hartshorne, Resudues and duality, LNM 20 (1966). [H] [KLT] S. Kumar, N. Lauritzen, J.F.Thomsen, Frobenius splitting of cotangent bundles of flag vari- eties, Invent. Math. 136 (1999), pp. 603-62 [RV] A. Rizzardo, M. Van den Bergh, An example of a non-Fourier?Mukai functor between derived [T] categories of coherent sheaves, arXiv:1410.4039. B. Toen, The homotopy theory of dg-categories and derived Morita theory, Invent. Math. 167 (2007), National Research University "Higher School of Economics" and the University of Oregon E-mail address: [email protected]
1507.07528	1	1507	2015-07-27T19:05:23	Dolbeault dga and $L_\infty$-algebroid of the formal neighborhood	[ "math.AG", "math.DG" ]	We continue the study the Dolbeault dga of the formal neighborhood of an arbitary closed embedding of complex manifolds previously defined by the author in \cite{DolbeaultDGA}. The special case of the diagonal embedding has been studied in \cite{Diagonal}. We describe the Dolbeault dga explicitly in terms of the formal differential geometry of the embedding. Moreover, we show that the Dolbeault dga is the completed Chevalley-Eilenberg dga an $L_\infty$-algebroid structure on the shifted normal bundle of the submanifold. This generlizes the result of Kapranov on the diagonal embedding and Atiyah class.	math.AG	math	DOLBEAULT DGA AND L∞-ALGEBROID OF THE FORMAL NEIGHBORHOOD SHILIN YU ABSTRACT. We continue the study the Dolbeault dga of the formal neighborhood of an arbitary closed embedding of complex manifolds previously defined by the author in [Yua]. The special case of the diagonal embedding has been studied in [Yu15]. We describe the Dolbeault dga ex- plicitly in terms of the formal differential geometry of the embedding. Moreover, we show that the Dolbeault dga is the completed Chevalley-Eilenberg dga an L∞-algebroid structure on the shifted normal bundle of the submanifold. This generlizes the result of Kapranov on the diago- nal embedding and Atiyah class. 5 1 0 2 l u J 7 2 ] . G A h t a m [ 1 v 8 2 5 7 0 . 7 0 5 1 : v i X r a CONTENTS 1. Introduction 2. Dolbeault dga of formal neighborhoods 2.1. Definitions and notations 3. Dolbeault dgas of diagonal embeddings 3.1. Diagonal embeddings and jet bundles 3.2. Formal geometry 3.3. Kapranov's result for Kahler manifolds 4. Case of general embeddings 4.1. Differential geometry of complex submanifolds 4.2. Taylor expansions in normal direction 5. L∞-algebroid of the formal neighborhood 5.1. Conventions and notations 5.2. L∞- and L∞[1]-algebroids 5.3. L∞[1]-algebroid of the formal neighborhood Bibliography References 2 3 3 5 6 6 11 13 13 14 22 22 22 26 27 27 2010 Mathematics Subject Classification. Primary 14B20; Secondary 16E45, 58A20. Key words and phrases. Formal neighborhood, Jet Bundle, Differential graded algebra, Formal geometry, Atiyah class, L∞-algebra, L∞-algebroid. This research was partially supported by the grant DMS1101382 from the National Science Foundation. 1 2 SHILIN YU 1. INTRODUCTION This is the continuation of [Yua] and [Yu15]. In [Yua] we introduced the notion of Dolbeaut differential graded algebra (dga) of the formal neighborhood of a closed embedding of complex manifolds, which contains all the formal geometric information of the embedding. Then in [Yu15] we studied the Dolbeault dga of the diagonal embedding and recovered Kapranov's description of the formal neighborhood of the diagonal in terms of the Atiyah class ([Kap99]). In the current paper, we will generalize the results in [Yu15] to arbitrary closed embeddings and show how to describe the Dolbeault dga explicitly in terms of the differential geometry of the submanifold, at least when the ambient manifold has a Kahler metric. The Dolbeault dga A = (A•( Y), ¶ ) for the formal neighborhood Y of any closed embedding i : X ֒→ Y is defined in a canonical way, independent of any auxiliary geometric structures of the manifolds. A certain dg category PA of certain dg-modules over A was built in [Yua] following the work of [Blo10] and was shown to be a dg-enhancement of the derived category of coherent sheaves over Y. However, we are interested in explicit construction of objects in PA, among which the most important one for us is the derived direct image of OX on Y, which will be the main content of our upcoming work [Yub]. There we will study the quantized analytic cycle class defined by Grivaux [Gri14], which specializes to the usual Todd class in the case of diagonal embedding. For this purpose, we need a geometric description of the Dolbeault dga, which reflects how the submanifold 'curls' in the ambient manifold. X×X and the dga (A0,• The paper [Yu15] provides such a description in the case of the diagonal embedding ∆ : X ֒→ X × X. It was shown that there exist isomorphisms between the Dolbeault dga of the formal X ( S(T∗X)), Dσ) of the Dolbeault resolution of the com- neighborhood X(∞) pleted symmetric algebra of the cotangent bundle of X, where the differentials Dσ depend on sections σ of certian jet bundle with infinite dimensional fibers and related to the Atiyah class of X. In the current paper, we generalize this result to the case of a general embedding i : X ֒→ Y, i.e., we show that there are isomorphisms between the dgas (A•( Y), ¶ ) and (A0,• X ( S(N∨)), D), where N∨ is the conormal bundle of the submanifold and the differential D again depends on sections of some jet bundle. The main idea is to consider the graph ei : X ֒→ X × Y of i, which again is an embedding, and the natural map between the pairs (X, X × Y) → (X, Y). The for- mal neighborhood Y of X inside Y can then be studied by studying the formal neighborhood X(∞) X×Y of X inside X × Y. The latter has a similar description as that of the diagonal embedding (Theorem 3.4). Kapranov's original result was formulated as an L∞-algebra structure on the shifted tangent bundle TX[−1], whose binary bracket is given by the Atiyah class. In our language, the com- pleted Chevalley-Eilenberg dga of TX[−1] is the Dolbeault dga A•(X(∞) X×X) or (A0,• X ( S(T∗X)), Dσ). DOLBEAULT DGA AND L∞-ALGEBROID OF THE FORMAL NEIGHBORHOOD 3 In particular, it induces an Lie algebra structure on TX[−1] as an object in the derived cat- egory of X. Our result on general embeddings can also be reformulated as an L∞-structure on the shifted normal bundle N[−1]. The novel discovery here is an extra ∞-anchor map N[−1] → TX which makes N[−1] into an L∞-algebroid. In the case of the diagonal embedding, it recovers Kapranov's L∞-algebra, where the anchor map vanishes. To our knowledge, the notion of L∞-algebroid was first defined in the work of Kjeseth ([Kje01b], [Kje01a]) under the name of strong homotopy Lie-Rinehart algebras. L∞-algebroids also appear in other context, such as string theory ([SSS12]) and the study of foliations ([Vit14]). We want to mention that Calaque, Caldararu and Tu have established similar results in the algebraic setting ([CCT14]). While they built a dg-Lie algebroid on some dg-sheaf which is quasi-isomorphic to i∗ N[−1] in the derived category of Y, our L∞-algebroid has the Dolbeault resolution of the normal bundle as the underlying complex and the higher brackets do not vanish in general. The paper is organized as follows. In § 2 we recall the general definition of the Dolbeault dga A•( Y) for the formal neighborhood Y of an arbitrary closed holomorphic embedding i : X ֒→ Y of complex manifolds. In § 3 we briefly review our reformulation ([Yu15]) of Kapranov's result of the diagonal embedding. We recall various infinite dimensional fiber bundles arising from formal geometry, which we already used heavily in [Yu15] to derive Kapranov's results. We then apply them in 4 to get a description of the Dolbeault dga of an arbitrary embedding i : X ֒→ Y, in which many other differential geometric quantities other than the curvature, such as Kodaira-Spencer class and shape operator, come into the picture. The main result is Theorem 4.5. For convenience we only discuss the Kahler case, yet the formulas make sense in broader context (see the comment at the beginning of § 4.2.2) . We construct an isomorphism from our canonical yet abstractly defined Dolbeault dga A•( Y) to a concrete dga, namely the completed symmetric algebra A0,•( S(N∨)) of the conormal bundle of X in Y, and compute the differential on it. The main result is Theorem 4.5 for the final answer. Finally, § 5 is contributed to the equivalent L∞-description. We will recall the basic definitions of L∞-algebroids from [Vit14]. For convenience, we will mainly use L∞[1]-algebroid, which is a shifted version of L∞-algebroid, since the signs involved in the formula are enormously simplified. Acknowledgements. The author would like to thank Jonathan Block, Damien Calaque, An- drei Caldararu, Nigel Higson and Junwu Tu for many discussions.This research was partially supported under NSF grant DMS-1101382. 2. DOLBEAULT DGA OF FORMAL NEIGHBORHOODS 2.1. Definitions and notations. We review the notations and definitions from [Yua]. Let i : (X, OX) ֒→ (Y, OY) be a closed embedding of complex manifolds where OX and OY are the 4 SHILIN YU structure sheaf of germs of holomorphic functions over the complex manifolds X and Y re- spectively. Let I the ideal sheaf of OY of holomorphic functions vanishing along X. The main objects studied by this paper are the r-th formal neighborhood Y(r) of X in Y, which is defined as the ringed space (X, O Y(r)) whose the structure sheaf is O Y(r) = OY/I r+1, and the (complete) formal neighborhood Y = Y(∞), which is defined to be the ringed space (X, O Y) where O Y = lim←− r O Y(r) = lim←− r OX/I r+1. We will also use the notations X(∞) manifolds. Y = Y and X(r) Y = Y(r) when we need to emphasize the sub- In [Yua] the Dolbeault differential graded algebra (dga) of the embedding i : X ֒→ Y is defined as Y , ¶ ) be the Dolbeault complex of Y, thought of as a dga. For r is set to be the graded ideal of A0,•(Y) consisting of those forms follows. Let (A0,•(Y), ¶ ) = (∧•Ω0,1 each nonnegative integer r, a• ω ∈ A0,•(Y) satisfying i∗(LV1LV2 · · · LVlω) = 0, ∀ 1 ≤ j ≤ l, (2.1) for any collection of smooth (1, 0)-vector fields V1, V2, . . . , Vl over Y, where 0 ≤ l ≤ r. By r is invariant under the action of ¶ and hence is a dg-ideal of (A0,•(Y), ¶ ). Proposition 2.1, [Yua], a• Definition 2.1 (Definition 2.3, [Yua]). The Dolbeault dga of the r-th formal neighborhood Y(r) is the quotient dga The Dolbeault dga of the complete formal neighorhood Y is defined to be the inverse limit A•( Y(r)) := A0,•(Y)/a• r . A•( Y) = A•( Y(∞)) := lim←− r A•( Y(r)). We will write A( Y) = A0( Y) and A( Y(r)) = A0( Y(r)) for the zeroth components of the Dol- beault dgas. The Dolbeault dga A•( Y(r)) can be sheafified to a soft sheaf of dgas A •( Y(r)) over X for r ∈ N or r = ∞ (see [Yua] for details). Moreover, there are natural inclusions of sheaves of algebras O Y(r) ֒→ A ( Y(r)) The following result was proved in [Yua]. Theorem 2.2 (Prop. 2.8., [Yua]). For any nonnegative integer r or r = ∞, the complex of sheaves is exact, where m = dim X. In other words, (A • −→ · · · Y(r) 0 → O Y(r) → A 0 Y(r) −→ A m −→ A 1 Y(r), ¶ ) is a soft resolution of O Y(r). Y(r) → 0 ¶ ¶ ¶ DOLBEAULT DGA AND L∞-ALGEBROID OF THE FORMAL NEIGHBORHOOD 5 As the completion of A0,•(Y) with respect to the filtration a• r , the dga A•( Y) is itself filtered and its associated graded dga is gr A•( Y) ≃ (A0,•(Y)/a• 0) ⊕ ∞M r=0 r /a• a• r+1. r /a• Note that A0,•(X) ≃ A0,•(Y)/a• each r ≥ 0, a 'cosymbol map' of complexes 0 and a• r+1 are dg-modules over (A0,•(X), ¶ ). We define, for τr : (a• r /a• r+1) → A0,• X (Sr+1N∨), (2.2) where Sr+1N∨ is the (r + 1)-fold symmetric tensor of the conormal bundle N∨ of the embed- ding. Given any (r + 1)-tuple of smooth sections µ1, . . . , µr+1 of N, we lift them to smooth sections of TYX and extend to smooth (1, 0)-tangent vector fields µ1, . . . , µr+1 on Y (defined near X). We then define the image of ω+ a• r , thought of as linear functionals on (N)⊗(r+1), by the formula r+1 under τ for any ω ∈ a• r+1 ∈ a• r /a• τr(ω+ a• r+1)(µ1 ⊗ · · · ⊗ µr+1) = i∗L µ1L µ2 · · · L µr+1ω. (2.3) The map τr is well-defined and is independent of the choice of the representative ω and µj's. Moreover, the tensor part of τr(ω+ a• Proposition 2.3. The map τr in (2.2) is an isomorphism of dg-modules over (A0,•(X), ¶ ). r+1) is indeed symmetric. Corollary 2.4. We have a natural isomorphism of dgas gr A•( Y) ≃ ∞M n=0 A0,• X (SnN∨). 3. DOLBEAULT DGAS OF DIAGONAL EMBEDDINGS Among all important and interesting examples is the diagonal embedding ∆ : X ֒→ X × X of a complex manifold X into product of two copies of itself. We then have the formal neigh- X×X), ¶ ) constructed borhood of the diagonal, X(∞) in the previous section. It is isomorphic to the Dolbeault resolution of the infinite holomorphic X . In this case one can write down the ¶ -derivation explicitly (at least when X is jet bundle J ∞ Kahler) due to a theorem below by Kapranov ([Kap99]), of which we will reproduce the Kahler X×X, which is understood as the dga (A0,•(X(∞) case in a slightly different way and discuss the general situation later. Intuitively one would expect that there is an isomorphism A0,•(X(∞) X×X) ≃ A0,•( S• X(T∗X)) (3.1) by taking 'Taylor expansions', where S•(T∗X) is the bundle of complete symmetric tensor al- gebra generated by the cotangent bundle of X (which is natural identified with the conormal bundle of the diagonal embedding). Such an isomorphism does exist, but there is no canonical 6 SHILIN YU way to define it since one need to first choose some local coordinates to get Taylor expansions. Indeed we will see that the isomorphism depends (in a more or less 1-1 manner) on a smooth choice of formal (holomorphic) coordinates on X. 3.1. Diagonal embeddings and jet bundles. We consider the case of diagonal embeddings. Let X be a complex manifold and let ∆ : X ֒→ X × X be the diagonal map. For convenience, we write ∆(r) = X(r) X×X for r ∈ N or r = ∞ throughout this section.Denote by pr1, pr2 : X × X → X the projections onto the first and second component of X × X respectively. The jet bundle J r X of order r (r ∈ N or r = ∞) can be viewed as the sheaf of algebras J r X = pr1∗ O ∆(r), which is a sheaf of OX-modules where the OX-actions are induced from the projection pr1. The sheaf (A 0,•(J r X is a sheaf of dgas and its global sections forms a dga X), ¶ ) of Dolbeault complexes of J r A0,•(J r X) = Γ(X, A 0,•(J r X)). Since J r X is a sheaf of OX-modules, A0,•(J r X) is a dga over the dga (A0,•(X), ¶ ). The Dolbeault dga (A•(∆(r)), ¶ ) of the formal neighborhood (A•(∆(r)), ¶ ) (r ∈ N or r = ∞) is also an (A0,•(X), ¶ )-dga via the compositions of homomorphisms of dgas A0,•(X) pr∗ 1−→ A0,•(X × X) → A•(∆(r)). Proposition 3.1 (Prop. 2.8., [Yu15]). The natural inclusion O ∆(r) ֒→ A ∆(r) determines an (A0,•(X), ¶ )- linear morphism of dgas Ir : A0,•(J r X) ≃−→ A•(∆(r)) either when r ∈ N or r = ∞. Similar results hold for the corresponding sheaves. 3.2. Formal geometry. 3.2.1. Differential geometry on formal discs. Fix a complex vector space V of dimension n. The formal disc bV is the formal neighborhood of 0 in V. Its function algebra is the formal power series algebra F = CJV∗K = S(V∗) = (cid:213) SiV∗. i≥0 It is a complete regular local algebra with a unique maximal ideal m consisting of formal power series with vanishing constant term. The associated graded algebra with respect to the usual m-filtration is the symmetric algebra gr F = S(V∗) = M SiV∗. i≥0 Since we are in the complex analytic situation, we endow F with the canonical Fr´echet topol- ogy . In algebraic setting, one need to use the m-adic topology on F . However, the associated DOLBEAULT DGA AND L∞-ALGEBROID OF THE FORMAL NEIGHBORHOOD 7 groups and spaces in question remain the same, though the topologies on them will be differ- ent. Since our arguments work for both Fr´echet and m-adic settings, the topology will not be mentioned explicitly unless necessary. We also use bV = Spf F to denote the formal polydisc, either as a formal analytic space or a formal scheme. We recall several definitions in §4, [Yu15]: G(∞)(V) = the proalgebraic group of automorphisms of the formal space bV, J(∞)(V) = Ker(d0 : G(∞) → GLn(V)), where d0(φ) is the tangent map of φ ∈ G(∞) at 0, g(∞)(V) = Lie algebra of G(∞)(V) = Lie algebra of formal vector fields on bV vanishing at 0 = (cid:213) = (cid:213) Hom(V∗, SiV∗), V ⊗ SiV∗ i≥1 i≥1 j(∞)(V) = Lie algebra of J(∞)(V) (3.2) = Formal vector fields on bV with vanishing constant and linear terms = (cid:213) = (cid:213) Hom(V∗, SiV∗), V ⊗ SiV∗ i≥2 i≥2 g(∞)(V) and j(∞)(V) act on F = (cid:213) we will write G(∞) = G(∞)(V) and so on. i≥0 SiV∗ as derivations in the obvious way. For convenience The short exact sequence 1 → J(∞) → G(∞) → GLn(V) → 1 (3.3) splits canonically by identifying elements of GLn = GLn(V) as jets of linear transformations on bV. Hence G(∞) is the semidirect product G(∞) = J(∞) ⋊ GLn. The canonical bijection of sets q : J(∞) ≃−→ G(∞)/GLn (3.4) is G(∞)-equivariant if we endow G(∞)/GLn with the usual left G(∞)-action and the left G(∞)-action on J(∞) is given by ψ· ϕ = ψ◦ ϕ◦ (d0ψ)−1, ∀ ψ ∈ G(∞), ∀ ϕ ∈ J(∞). (3.5) ≃−→ In view of (3.5), it is natural to interpret J(∞) as the set of formal exponential map ϕ : dT0V bV, where dT0V is the completion of the tangent space T0V at the origin, which is canonically identified with bV itself. Such ϕ induces the identity map on the tangent spaces of the two 8 SHILIN YU formal spaces at the origins. Moreover, giving a formal exponential map ϕ is equivalent to giving an isomorphism between filtered algebras ϕ∗ : F → FT, where FT := S((T0V)∗) = S(V∗) is the algebra of functions on dT0V, such that the induced isomorphism between the associated graded algebras, which are both S(V∗), is the identity map. In [Yu15] we introduced the set Conn of all flat torsion-free connections on bV, i.e., each ele- ment of Conn is a (nonlinear) map ∇ : TbV → TbV ⊗ObV T∗bV satisfying the Leibniz rule and the flatness and torsion-freeness conditions. By abuse of no- tation, we also use ∇ to denote the induced connection on the cotangent bundle of bV and its associated tensor bundles: There is a canonical bijection ∇ : T∗bV → T∗bV ⊗ObV T∗bV. exp : Conn ≃−→ J(∞), ∇ 7→ exp∇, (3.6) which assigns to each connection ∇ a formal exponential map exp∇ : dT0V → bV, which is completely determined by the way it pulls back functions f ∈ F , exp∗ ∇( f ) = (∇i f 0)i≥0 = ( f (0), ∇ f 0, ∇2 f 0, · · · ) ∈ (cid:213) i≥0 Si(T0V)∗ = FT, (3.7) where ∇ f = df and ∇i f = ∇i−1df for i ≥ 2. The torsionfreeness and flatness of ∇ guanrantee that the terms in the expression are symmetric tensors. Moreover, G(∞) naturally acts on Conn from left by pushing forward connections via auto- morphisms of bV. By Lemma 4.1, [Yu15], the bijection exp : Conn → J(∞) is G(∞)-equivariant. 3.2.2. Bundle of formal coordinates and connections. We introduce the bundle of formal coordinate systems Xcoord of a smooth complex manifold X from §4.4., [Kap99]. At each point x ∈ X the fiber Xcoord,x is the space of infinite jets of biholomorphisms ϕ : V ≃ Cn → X with ϕ(0) = x. Xcoord is naturally a holomorphic principal G(∞)-bundle. We can apply the associated bundle construction to the principal G(∞)-bundle Xcoord to glob- alize various objects defined in §3.2.1. There is a canonical isomorphism between bundles of algebras and hence we have a tautological trivalization of the jet bundle J ∞ X over Xcoord Xcoord ×G(∞) F ≃ J ∞ X Xcoord ×X J ∞ X ≃ Xcoord × F . (3.8) DOLBEAULT DGA AND L∞-ALGEBROID OF THE FORMAL NEIGHBORHOOD 9 Other related jet bundles, such as J ∞TX (J ∞T∗ X, resp.), the jet bundle of the tangent bundle (cotangent bundle, resp.), can be obtained in a similar way by the associated bundle construc- tion. Another related bundle π : Xexp → X is the bundle of formal exponential maps introduced in [Kap99], which we denote by Xexp. At each x ∈ X the fiber Xexp,x is the space of jets of holomorphic maps φ : TxX → X such that φ(0) = x, d0φ = IdTxX. We have a canonical isomorphism Xexp → Xcoord ×G(∞) J(∞) ≃ Xcoord ×G(∞) (G(∞)/GLn) which hence induces a biholomorphism Xcoord/GLn ≃ Xexp (3.9) We also defined in [Yu15] the bundle of jets of flat torsion-free connection Xconn = Xcoord ×G(∞) Conn whose fiber at a each point x ∈ X consists of all flat torsion-free connections on the formal ≃−→ J(∞) induces an identification neighborhood of x. The G(∞)-equivariant bijection exp : Conn between the Xconn and Xexp. We regard them as the same bundle with different descriptions. There is a tautological flat and torsion-free connection over Xconn, ∇tau : π∗J ∞T∗X → π∗J ∞T∗X ⊗π∗J ∞ X π∗J ∞T∗X, which is OXconn-linear and satisfies the Leibniz rule with respect to the differential that is the pullback of ed(∞) : π∗J ∞ X → π∗J ∞T∗X d(∞) : J ∞ X → J ∞T∗X. Here d(∞) is a OX-linear differential obtained by apply the associated bundle construction with Xcoord and the de Rham differential d : ObV → T∗bV on the formal disc bV. On the other hand, since Xconn can also be interpreted as the bundle Xexp of formal exponen- tial maps, we have a tautological isomorphism between sheaves of algebras over Xconn = Xexp, Exp∗ : π∗(Xcoord ×G(∞) F ) → π∗(Xcoord ×G(∞) FT). The domain is identified as π∗J ∞ formal neighborhood of the diagonal in X × X, while for the codomain we have X or π∗O , the pullback via π of the structure sheaf of the X(∞) X×X Xcoord ×G(∞) FT ≃ Xcoord ×G(∞) GLn ×GLn FT ≃ Xcoord/J(∞) ×GLn FT 10 SHILIN YU by our definition of the G(∞)-action on FT. But the principal GLn-bundle Xcoord/J(∞) is exactly the bundle of (0th-order) frames on X, so Xcoord/J(∞) ×GLn V ≃ Xcoord/J(∞) ×GLn T0V ≃ TX. and similarly Since the GLn-action respects the decomposition FT = (cid:213) i≥0 SiV∗, we get Xcoord/J(∞) ×GLn V∗ ≃ T∗X. Xcoord/J(∞) ×GLn FT ≃ (cid:213) SiT∗X = S(T∗X), i≥0 which is the structure sheaf of X(∞) we have a tautological exponential map TX, the formal neighborhood of the zero section of TX. In short, or equivalently, a tautological Taylor expansion map Exp : π∗X(∞) TX → π∗X(∞) X×X Exp∗ : π∗O X (∞) X×X → π∗O . X (∞) TX which is an isomorphism of bundles of topological algebras. The induced map between asso- ciated bundle of graded algebras gr Exp∗ : π∗ gr O X(∞) X×X = π∗ S(T∗X) → π∗ S(T∗X) is the identity map. By (3.7) we can write Exp∗ explicitly in terms of ∇tau, Exp∗( f ) = (∇i tau f 0)i≥0 = ( f (0), ∇tau f 0, ∇2 tau f 0, · · · ) ∈ π∗ S(T∗X), (3.10) where 0 stands for the 'restriction to the origin' map π∗ SiJ ∞T∗X → π∗ SiT∗X. It is the glob- alization of the natural restriction map T∗bV → T∗ 0 bV = V∗ by applying the associated bundle construction with Xcoord and then pulling back onto Xconn via π. Again ∇tau f means d(∞) f and so on. The Taylor expansion map Exp∗ induces a natural bijection between global smooth sections X and S(T∗X) which are the identity of Xconn and all possible smooth isomorphisms between J ∞ map on the level of associated graded algebras. Given any smooth sectionσ of Xconn, we denote by exp∗ σ : J ∞ X → S(T∗X) the corresponding smooth homomorphism of bundles of algebras over X. It is holomorphic if and only if σ is holomorphic. In general, Xconn carries a flat (0, 1)-connection d, such that for any given smooth section σ of Xconn, its anti-holomorphic differential ωσ := dσ ∈ A0,1(j(∞)(TX)) (3.11) DOLBEAULT DGA AND L∞-ALGEBROID OF THE FORMAL NEIGHBORHOOD 11 is well-defined and it satisfies a Maurer-Cartan type equation ¶ ωσ + 1 2 [ωσ, ωσ] = 0. (3.12) We denote by σ ∈ A0,1 αn X (Hom(SnTX, TX)) = A0,1 X (Hom(T∗X, SnT∗X)). the n-th graded component of ω in the decomposition (3.2). By abuse of notation, we also denote the A0,•(X)-linear extension of exp∗ σ by exp∗ σ : A•(X(∞) X×X) → A0,• X ( S(T∗X)) (3.13) between graded algebras. It is in general not a homomorphism of dgas. The deficiency is measured exactly by ω since ωσ = ¶ exp∗ σ ◦(exp∗ σ)−1, (3.14) σ = [¶ where ¶ exp∗ σ is the odd derivation of the graded algebra A0,•( S(T∗X)) induced by αn eαn (3.12) and we have σ]. Define a new differential Dσ = ¶ − (cid:229) n≥2 eαn , exp∗ σ on A0,• X ( S(T∗X), where σ. Then D2 σ = 0 by Proposition 3.2 ([Yu15], Prop. 4.4.). For any given smooth section σ of Xconn, the map exp∗ σ : (A•(X(∞) X×X), ¶ ) → (A0,•( S(T∗X)), Dσ) is an isomorphism of dgas. 3.3. Kapranov's result for Kahler manifolds. Now suppose that X is equipped with a Kahler metric h. Let ∇ be the canonical (1, 0)-connection in TX associated with h, so that [∇, ∇] = 0 in A2,0 X (End(TX)). (3.15) and it is torsion-free, which is equivalent to the condition for h to be Kahler. is the (0, 1)-connection defining the complex structure. The curva- Set e∇ = ∇ + ¶ ture of e∇ is just , where ¶ R = [¶ , ∇] ∈ A1,1 X (End(TX)) = A0,1 X (Hom(TX ⊗ TX, TX)) (3.16) which is a Dolbeault representative of the Atiyah class αTX of the tangent bundle. In particular one has the Bianchi identity: ¶ R = 0 in A0,2 X (Hom(TX ⊗ TX, TX)) Actually, by the torsion-freeness we have R ∈ A0,1 X (Hom(S2TX, TX)) 12 SHILIN YU Now define tensor fields Rn, n ≥ 2, as higher covariant derivatives of the curvature: Rn ∈ A0,1 X (Hom(S2TX ⊗ TX⊗(n−2), TX)), R2 := R, Ri+1 = ∇Ri (3.17) In fact Rn is totally symmetric, i.e., Rn ∈ A0,1 X (Hom(SnTX, TX)) = A0,1 X (Hom(T∗X, SnT∗X)) by the flatness of ∇ (3.15). Note that if we think of ∇ as the induced connection on the cotangent bundle, the same formulas (3.16) and (3.17) give −Rn. The connection ∇ determines a smooth section of Xconn,which we write as σ = [∇]∞, by assigning to each point x ∈ X the holomorphic jets of ∇. This has been done implicitly in the proof of Lemma 2.9.1., [Kap99]. One can check that the induced Taylor expansion map exp∗ σ : A•(X(∞) X×X) ≃−→ A0,• X ( S(T∗X)) by exp∗ σ([η]∞) = (∆∗η, ∆∗∇η, ∆∗∇2η, · · · , ∆∗∇nη, · · · ) ∈ A0,• X ( S(T∗X)) (3.18) for any [η]∞ ∈ A•(X(∞) X×X). Here ∇ is understood as the pullback of ∇ (on the cotangent bundle) via pr2, which is a constant family of connections along fibers of pr1, instead of jets of ∇. The key observation is that the right hand side of the formula only depends on the class [η]∞ ∈ A•(X(∞) X×X) and the holomorphic jets of the ∇. In [Yu15], we reformulated and reproved Theorem 2.8.2., [Kap99] in our language. Theorem 3.3 (Thm 3.2., [Yu15]). Assume X is Kahler. With the notations from §3.2.2 and above, we have αn σ = −Rn, i.e., ωσ = − (cid:229) n≥2 Rn. Thus there is an isomorphism between dgas exp∗ σ : (A•(X(∞) X×X), ¶ ) → (A0,• X ( S(T∗X)), Dσ) The derivation Dσ = ¶ + (cid:229) n≥2 eRn, where eRn is the odd derivation of A0,•( S(T∗X)) induced by Rn. We conclude this section by a slightly more generalized version of Theorem 3.3, which will be used later in § 4.2. Suppose f : X → Y is an arbitrary holomorphic map. We consider the graph of f ef = (Id, f ) : X → X × Y which is a closed embedding. So we can consider the formal neighborhood X(∞) X×Y. All the constructions above can be carried out in almost the same way with only slight adjustment and give us a description of the Dolbeault dga A•(X(∞) X×Y). Namely, consider the pullback bundle f ∗Yconn over X. Each smooth section σ of f ∗Yconn naturally corresponds to an isomorphism ησ : A•(X(∞) X×Y) ֒→ A0,• X ( S( f ∗T∗Y)) DOLBEAULT DGA AND L∞-ALGEBROID OF THE FORMAL NEIGHBORHOOD 13 of graded algebras. One can also view sections of f ∗Yconn as jets of flat torsion-free connections on X(∞) X×Y along Y-fibers. In particular, when Y carries a Kahler metric and the canonical (1, 0)- connection ∇, we can pullback ∇ via the projection X × Y → Y, which determines a smooth section σ of f ∗Yconn and a Taylor expansion map X ( S( f ∗T∗Y)), X×Y) → A0,• [ζ]∞ 7→ (ef ∗ζ, ef ∗∇ζ, ef ∗∇2ζ, · · · ), σ : A•(X(∞) exp∗ for any [ζ]∞ ∈ A•(X(∞) X×Y). By abuse of notations, we still write Rn ∈ A0,1 Z (Hom( f ∗TY, Sn( f ∗TY))) as the pullback of the curvature form of Y and its covariant derivatives via f . Then we have the following theorem, Theorem 3.4. We have an isomorphism between dgas X ( S•( f ∗T∗Y)), Dσ) where Dσ = ¶ + (cid:229) n≥2 eRn and eRn is the derivation of degree +1 induced by Rn. σ : (A•(X(∞) X×Y), ¶ ) ≃−→ (A0,• exp∗ 4. CASE OF GENERAL EMBEDDINGS Let i : X ֒→ Y be an arbitrary embedding and A•( Y) the Dolbeault dga associated to the formal neighborhood of X inside Y as in § 3.2.2. The goal is to build some appropriate isomor- X ( S•(N∨)) and write down the ¶ -derivation explicitly under this identifi- phism A•( Y) ≃ A0,• cation. We will show that this can be derived from the special yet universal case considered in § 3. 4.1. Differential geometry of complex submanifolds. 4.1.1. Splitting of normal exact sequence and Kodaira-Spencer class. Over X we have the short exact sequence of holomorphic vector bundles defining the normal bundle N = N and its dual 0 → TX ι−→ i∗TY p −→ N → 0 0 → N∨ p∨ −→ i∗T∗Y ι∨ −→ T∗X → 0 (4.1) (4.2) We fix a choice of C∞-splitting of the normal exact sequence (4.1), i.e., two smooth homomor- phisms of vector bundles τ : i∗TY → TX and ρ : N → i∗TY satisfying τ ◦ ι = IdTX, p ◦ ρ = IdN, ι ◦ τ + ρ◦ p = Idi∗TY and denote the corresponding dual splitting on the conormal exact sequence (4.2) by τ∨ : T∗X → i∗T∗Y and ρ∨ : i∗T∗Y → N∨. We can choose the splittings as the orthonormal de- composition induced by a Kahler metric on Y if there is one, but again we will never need a metric in our general discussion. 14 SHILIN YU Think of τ∨ as a C∞-section of the holomorphic vector bundle Hom(T∗X, i∗T∗Y), we can form In fact β = βX/Y := ¶ τ∨ ∈ A0,1 X (Hom(T∗X, T∗Y)). β ∈ A0,1 X (Hom(T∗X, N∨)). To see this, just apply ¶ on both sides of equality ι∨ ◦ τ∨ = IdT∗X and note that ι∨ is holo- morphic and hence ¶ ι∨ = 0. By definition ¶ β = 0, thus β defines a cohomology class [β] ∈ Ext1 X(T∗X, N∨), which is the obstruction class for the existence of a holomorphic splitting of the exact sequence (4.2) or (4.1). We call it the Kodaira-Spencer class. Also note that β = −¶ ρ = −¶ ρ∨ ∈ A0,1(Hom(N, TX)) = A0,1(Hom(T∗X, N∨)) (4.3) 4.1.2. Shape operator. Suppose ∇ is an arbitrary (1, 0)-connection on TY without any additional assumption. We use the same notation for the pullback connection on i∗TY or i∗T∗Y over X. The induced connection on the normal bundle via the chosen splitting is denoted by ∇⊥: ∇⊥ Vµ := p(∇Vρ(µ)) ∈ C∞ X (N), ∀ µ ∈ C∞ X (N), V ∈ C∞(TX) (here we identify T1,0X with TX). Analogous to the shape operator in Riemannian geometry, we also define a linear operator SN : TX ⊗ N → TX by Sµ N(V) = −τ(∇Vρ(µ)), ∀ µ ∈ C∞ X (N), V ∈ C∞(TX). (4.4) That is, we first lift a smooth section µ of the normal bundle to a section of i∗TY via the splitting, then take its derivatives with respect to the induced connection on i∗TY and finally project the output into TX. Note that SN is in general not a holomorphic map between vector bundles. 4.2. Taylor expansions in normal direction. 4.2.1. General discussions. Let i : X ֒→ Y be a closed embedding, where Y is not necessarily Kahler. Similar to what has been done in § 3.2.2, we can consider all isomorphisms A•( Y) ≃−→ A0,• X ( S(N∨)) which induces identity on the associated graded bundle S(N∨) and there is a infinite dimen- sional bundle ΨX/Y → X whose smooth sections correspond exactly to such isomorphisms. Indeed, for each x ∈ X, the fiber ΨX/Y,x is the space of jets of holomorphic maps ψ : Nx → Y with ψ(0) = x and p ◦ d0ψ = IdNx, where Nx is the fiber of the normal bundle at point x ∈ X and p : i∗TY → N is the natural projection. DOLBEAULT DGA AND L∞-ALGEBROID OF THE FORMAL NEIGHBORHOOD 15 Similarly, we can define another bundle ΘX/Y over X whose fiber at x is the space of jets of maps θ : Nx → TxY with θ(0) = x and p ◦ d0θ = IdNx. Then ΘX/Y admits a natural left action of J(∞)(T∗Y) (or more precisely, J(∞)(T∗YX)). In fact, we have a canonical isomorphism ΨX/Y ≃ YexpX ×J(∞)(T∗Y) ΘX/Y ≃ YconnX ×J(∞)(T∗Y) ΘX/Y. For our purpose here, however, it is not convenient to deal with ΨX/Y since even we have already understood Yconn in various geometric ways, general sections of the bundle ΘX/Y are difficult to handle. So instead we only look at linear liftings Nx → TYx which form a subbundle (1) X/Y sending jets of maps Θ ψ : Nx ֒→ Y to their linearizations d0ψ. Thus we have a fiberwise surjection of bundles (1) X/Y ⊂ ΘX/Y. Moreover, there is a canonical retraction ΘX/Y → Θ κ : YconnX ×X Θ (1) X/Y → YconnX ×J(∞)(T∗Y) ΘX/Y = ΨX/Y (1) As an affine bundle over the vector bundle N∨ ⊗ TX, Θ X/Y admits smooth sections which are nothing but C∞-liftings ρ : N → TY. Such a lifting ρ and any section σ of YconnX together determine a section Ξ of ΨX/Y via κ, and hence an isomorphism of graded algebras exp∗ X/Y,Ξ : A•( Y) ≃−→ A0,• X ( S•(N∨)) The rest of the job is to determine which differential we should put on the codomain of this isomorphism in terms of ρ and σ to make it into an isomorphism of dgas. As in Theorem 3.4, denote the graph of i by ei := (Id, i) : X → X × Y. We regard X as a submanifold of X × Y viaei, then by Theorem 3.4 a section σ of YconnX induces exp∗ σ : (A•(X(∞) X×Y), ¶ ) ≃−→ (A0,• X ( S•(T∗Y)), Dσ) where T∗Y is understood as the pullback i∗T∗Y (we will omit i∗ as long as there is no confusion). The derivation Dσ = ¶ + (cid:229) n≥2 eRn (4.5) and eRn is induced by (the pullback of) the covariant derivatives of curvature forms of Y. Note that we have a commutative diagram of holomorphic maps X ⊂ ei > X × Y wwwwww ⊂ X π ∨ ∨ > Y i where π : X × Y → Y is the natural projection. Thus by functoriality, π induces an injective homomorphism of dgas π∗ : (A•( Y), ¶ ) → (A0,•(X(∞) X×Y), ¶ ), [η]∞ 7→ [π∗η]∞. (4.6) 16 SHILIN YU We then compose π∗ with the isomorphism exp∗ σ to get σ ◦ π∗ : (A•( Y), ¶ ) → (A0,• exp∗ X ( S•(T∗Y), Dσ) We then extend ρ∨ : T∗Y → N∨ to obtain a homomorphism of graded algebras ρ∨ : A0,• X ( S•(T∗Y)) → A0,• X ( S•(N∨)). Composing with exp∗ σ ◦π∗ in (4.11) we get a homomorphism of graded algebras ρ∨ ◦ exp∗ σ ◦π∗ : A•( Y) → A0,• X ( S•(N∨)). Lemma 4.1. With all the notations above, we have Proof. Follows immediately from the definition of κ. exp∗ X/Y,Ξ = ρ∨ ◦ exp∗ σ ◦π∗ (4.7) (4.8) (cid:3) Via the isomorphism exp∗ X ( S•(N∨)), denoted as D, i.e., A0,• X/Y,Ξ we can transfer the ¶ -derivation on A•( Y) to a derivation on D = exp∗ X/Y,Ξ ◦ ¶ ◦ (exp∗ X/Y,Ξ)−1. Hence exp∗ X/Y,Ξ becomes an isomorphism of dgas exp∗ X/Y,Ξ : (A•( Y), ¶ ) ≃−→ (A0,• X ( S•(N∨)), D). Thus we can also transfer the homomorphism π∗ in (4.6) to a homomorphism eπ∗ : (A0,• X ( S•(N∨)), D) → (A0,• X ( S•(T∗Y), Dσ), by setting We then get the following commutative diagram: eπ∗ = exp∗ σ ◦ π∗ ◦ (exp∗ X/Y,Ξ)−1. (A•( Y), ¶ ) π∗ > (A0,•(X(∞) X×Y), ¶ ) ≃ exp∗ X/Y,Ξ ∨ (A0,• X ( S•(N∨)), D) eπ∗ ≃ exp∗ σ ∨ > (A0,• X ( S•(T∗Y), D) Note that by (4.8) and (4.9) we have ρ∨ ◦ eπ∗ = Id : (A0,• X ( S•(N∨)), D) → (A0,• X ( S•(N∨)), D) even though ρ∨ is not a homomorphism of dgas. (4.9) (4.10) DOLBEAULT DGA AND L∞-ALGEBROID OF THE FORMAL NEIGHBORHOOD 17 4.2.2. Description of eπ∗. From now on, let us assume Y is Kahler and the section σ of YconnX is determined by the associated (1, 0)-connection ∇. However, we want to keep the reader aware that all the arguments and computations below will still work even if the Kahler condition is dropped and the section σ is arbitrary. All the ∇ appearing in the formulae can be intepreted as formal connections determined by σ without any change just as in Proposition 3.2. The Kahler assumption just makes sure that those terms in the final formula (4.20) of D have clearer differential geometric meanings. By the discussion at the end of §3.3, the homomorphism exp∗ σ ◦ π∗ : (A•( Y), ¶ ) → (A0,• X ( S•(T∗Y), Dσ) is given by [η]∞ 7→ (i∗η, i∗∇η, i∗∇2η, · · · ) (4.11) Y while the one on the left hand side is the pullback one along Y-fibers of X × Y. sinceei∗∇nπ∗η = i∗∇nη, where the ∇ on the right hand side is the original (1, 0)-connection on Before we give a description of the homomorphism eπ∗, we make some conventions on no- tations. We abuse the notations and write TY = TX ⊕ N and T∗Y = T∗X ⊕ N∨ induced by the fixed splittings. The decompositions extend to tensor products, i.e., tensor product TY⊗n can be decomposed into direct sum of mixed tensors of TX and N components and similarly for T∗Y⊗n. The same happens for symmetric tensor products: SnT∗Y = M p+q=n SpT∗X · SqN∨, where the dot stands for the commutative multiplication in the symmetric algebras. We define a derivation ∇ : A0,• X ( S•T∗Y) → A0,• X ( S•+1T∗Y) (4.12) (4.13) of degree 0 for the grading from A0,• X as the composition of the operators ∇ := Sym ◦ ∇TX Namely, for any η ∈ A0,• get a T∗X ⊗ SnT∗Y-valued form X (SnT∗Y), first apply the connection ∇TX on i∗ SnT∗Y induced by ∇ to ∇TXη ∈ A0,• X (T∗X ⊗ SnT∗Y), then apply a variation of the usual symmetrization map Sym : T∗X ⊗ SnT∗Y → Sn+1T∗Y, which we define on each component of the decomposition (4.12) as Symm,n : T∗X ⊗ (Sm−1T∗X · Sn−m+1N∨) → SmT∗X · Sn−m+1N∨ 18 by the formula SHILIN YU Symm,n(v0 ⊗ (v1 · v2 · · · vn)) = v0 · v1 · · · vn, 1 m ∀ v0 ⊗ (v1 · v2 · · · vn) ∈ T∗X ⊗ (Sm−1T∗X · Sn−m+1N∨) (4.14) and we finally get When η ∈ A0,• of forms on X. We can inductively apply ∇ and get X (S0(T∗Y)) = A0,• ∇η ∈ A0,• X (Sn+1T∗Y) X , ∇η = τ∨(¶ η) ∈ A0,• X (T∗Y) where ¶ is the (1, 0)-derivation k ∇ : A0,• X ( S•(T∗Y)) → A0,• X ( S•+k(T∗Y)) In particular, since N∨ is naturally identified as a subbundle of T∗Y via p∨ : N∨ → T∗Y, we can form the restriction of ∇ on S•(N∨): k k ∇ : A0,• X ( S•(N∨)) → A0,• X ( S•+k(T∗Y)) Note that here ∇ 0 is the natural inclusion S•(N∨) ֒→ S•(T∗Y). Proposition 4.2. We have eπ∗ = k ∇ ∞(cid:229) k=0 X ( S•(N∨)) → A0,• where eπ∗ : A0,• the n-th component of its image ν = eπ∗(µ) is X ( S•(T∗Y)) is as in (4.9). That is, given µ = (µk)∞ k=0 ∈ A0,• X ( S•(N∨)), νn = ∇ k µn−k ∈ A0,• X (SnN∨). n(cid:229) k=0 Proof. Assume that exp∗ X/Y,Ξ([η]∞) = µ where [η]∞ ∈ A•( Y). By (4.8) and (4.11), this means where ρ∨ is the projection µk = ρ∨(i∗∇kη) ρ∨ : A0,• X ( S•(T∗Y)) → A0,• as in (4.7). Moreover, by the definition of eπ∗ (4.9), we have X/Y,Ξ)−1(µ) = exp∗ σ ◦ π∗ ◦ (exp∗ eπ∗(µ) = exp∗ X ( S•(N∨)) σ ◦ π∗([η]∞) = (i∗∇kη)∞ k=0. Thus all we need to show is that i∗∇nη = n(cid:229) k=0 k ∇ (ρ∨(i∗∇n−kη)) DOLBEAULT DGA AND L∞-ALGEBROID OF THE FORMAL NEIGHBORHOOD 19 for all n ≥ 0. We prove it by induction on n. The n = 0 case is trivial. For n ≥ 1, note that we can write i∗∇nη = ρ∨(i∗∇nη) + component in T∗X · Sn−1T∗Y ⊂ SnT∗Y via the decomposition (4.12). The second term on the right hand side is nothing but ∇(i∗∇n−1η). To see this, consider what happens when we evaluate i∗∇nη at some section s of TY⊗n, which lies in a mixed tensor of m copies of TX (m ≥ 1) and n − m copies of N (of arbitrary order). One can permute any of the TX-factors to the first place and plug it into the first ∇ in i∗∇nη, since i∗∇nη is a symmetric tensor. This implies that i∗∇nη should be a symmetrization of ∇TX(i∗∇n−1η). The value of the latter at s, however, is m times what we need since s contains m TX-factors. This explains the fractional factor 1/m in the formula (4.14). Finally we end the proof by applying the inductive assumption. (cid:3) 4.2.3. Description of the derivation D. To determine the derivation D, note that by (4.10) we have D = ρ∨ ◦ eπ∗ ◦ D = ρ∨ ◦ D ◦ eπ∗ where the last equality is by the definition of D. Thus for given µ = (µk)∞ the n-th component of Dµ is k=0 ∈ A0,• X ( S•(N∨)), (Dµ)n = (cid:229) ρ∨ ◦ ¶ ◦ ∇ s µt + (cid:229) s+t=n r+s+t=n r≥1 ρ∨ ◦ eRr+1 ◦ ∇ s µt (4.15) by Proposition 4.2 and (4.5). To simplify the right hand side further, first observe that all we s µt lying in S•(N∨) and T∗X · S•(N∨) and we can ignore the need are the components of ∇ remaining ones in S2T∗X · S•(T∗Y). The reason is that if we apply the derivations ¶ and eRn on any section from S2T∗X · S•(T∗Y), the outcomes must lie in T∗X · S•(T∗Y), which will then be eliminated by the projection ρ∨. Thus we first denote the projections onto the only two 'effective' components respectively by and P0 = p∨ ◦ ρ∨ : S•(T∗Y) → S•(N∨) ⊂ S•(T∗Y) P1 : S•(T∗Y) → T∗X · S•(N∨) ⊂ S•(T∗Y). Secondly, we define the derivation of degree +1 eβ : A0,• X ( S•(T∗Y)) → A0,•+1 X ( S•(T∗Y)) induced by β ∈ A0,1 X (Hom(T∗X, N∨)) ⊂ A0,1 X (Hom(T∗Y, T∗Y)). 20 SHILIN YU where the last inclusion comes again from the splitting of TY. Note that eβacts on A0,• X ( S•(T∗Y)) as the zero map. So we can also think of eβ as the operator A0,• X ( S•(N∨)) ⊂ eβ : A0,• X (T∗X · S•(N∨)) → A0,•+1 X ( S•+1(N∨)) The following lemma is immediate from (4.3). Lemma 4.3. As derivations A0,• X ( S•(T∗Y)) → A0,•+1 X ( S•(N∨)), Hence the first sum on the right hand side of (4.15) can be rewritten as [ρ∨, ¶ ] = eβ◦ P1 ρ∨ ◦ ¶ ◦ ∇ s µt = (cid:229) ¶ ◦ ρ∨ ◦ ∇ s µt + (cid:229) s+t=n eβX ◦ P1 ◦ ∇ s µt (4.16) s+t=n The last equality is because that ∇ s+t=n = ¶ µn + (cid:229) s µt ∈ A0,• s µt s+t=n eβX ◦ P1 ◦ ∇ X (T∗X · S•(T∗Y)) unless s = 0. Thirdly, we split Rn ∈ A0,1 X (Hom(T∗Y, SnT∗Y)) into two components, n := ρ∨ ◦ Rn ◦ p∨ ∈ A0,1 R⊥ X (Hom(N∨, SnN∨)) and n := ρ∨ ◦ Rn ◦ τ∨ ∈ A0,1 R⊤ X (Hom(T∗X, SnN∨)) and denote the induced operators by eR⊥ n : A0,• X ( S•(N∨)) → A0,•+1 X ( S•+n−1(N∨)) and respectively. To unify notations, we also write R⊤ eR⊤ n : A0,• X (T∗X · S•(N∨)) → A0,•+1 X ( S•+n(N∨)) 1 := eβ. 1 := β and eR⊤ We can then split the second term on the right hand side of (4.15): r+s+t=n r≥1 = (cid:229) r+s+t=n r≥1 = (cid:229) r+s+t=n r≥1 n(cid:229) k=2 eR⊥ = s µt + (cid:229) r+s+t=n r≥1 ρ∨ ◦ eRr+1 ◦ P1 ◦ ∇ s µt eR⊤ r+1 ◦ P1 ◦ ∇ s µt (4.17) ρ∨ ◦ eRr+1 ◦ ∇ s µt ρ∨ ◦ eRr+1 ◦ P0 ◦ ∇ s µt + (cid:229) r+s+t=n r≥1 eR⊥ r+1 ◦ P0 ◦ ∇ k ◦ µn−k+1 + (cid:229) eR⊤ r+1 ◦ P1 ◦ ∇ s µt r+s+t=n r≥1 (cid:229) (cid:229) DOLBEAULT DGA AND L∞-ALGEBROID OF THE FORMAL NEIGHBORHOOD 21 Combine (4.15), (4.16) and (4.17) we get (Dµ)n = ¶ µn + k ◦ µn−k+1 + (cid:229) n(cid:229) k=2 eR⊥ r+s+t=n r,t≥0, s≥1 eR⊤ r+1 ◦ P1 ◦ ∇ s µt (4.18) Finally, to compute the terms P1 ◦ ∇ s µt, we define two derivations of degree 0 with respect to the grading from A0,• X : ⊥ ∇ : A0,• X ( S•(N∨)) → A0,• X (T∗X · S•(N∨)) induced by the connection ∇⊥ on N∨ the same way as we define ∇ in (4.13). ∇⊥ f of a function f is again understood as ¶ f , the (1, 0) differential. The second one fSN : A0,• X (T∗X · S•(N∨)) → A0,• X (T∗X · S•+1(N∨)) induced by the shape operator SN : T∗X → T∗X ⊗ N∨ as in (4.4) yet again with the images symmetrized. Lemma 4.4. With the notations above, we have thus ∀ s ≥ 1, ∀ µ ∈ S•(N∨), P1 ◦ ∇ = ∇ ⊥ ◦ P0 + fSN ◦ P1, P1 ◦ ∇ s µ = (fSN)s−1 ◦ ∇ ⊥ µ. Applying the equality (4.19) to (4.18), we eventually get (4.19) Theorem 4.5. Given µ = (µk)∞ (Dµ)n = ¶ µn + r+s+t=n r,t≥0, s≥1 k=0 ∈ A0,• n(cid:229) k=2 eR⊥ X ( S•(N∨)), the n-th component of Dµ is k ◦ µn−k+1 + (cid:229) ◦ ∇ s−1 ⊥ r+1 ◦ fSN eR⊤ µt. In other words, D = ¶ + (cid:229) k + (cid:229) k≥2 eR⊥ p≥1, q≥0 eR⊤ p ◦ fSN q ◦ ∇ ⊥ . (4.20) Remark 4.6. From Theorem 4.5 we see that, even when D acts on a function f (or a form), higher term would be produced in general. Namely, by (4.20) D f = ¶ f + (cid:229) p≥1, q≥0 eR⊤ p ◦ fSN q ◦ ¶ f (4.21) This is a huge difference between the general situation and the case of diagonal embedding. Remark 4.7. Although we get D2 = 0 for free from how we construct it, it is still an interesting (yet tedious) exercise to verify it by hands and one will observe the Gauss-Codazzi-Ricci equa- tions in classical differential geometry (see [Xin03]). We leave the details to interested readers. 22 SHILIN YU 5. L∞-ALGEBROID OF THE FORMAL NEIGHBORHOOD In this section we repackage the results in § 4 in terms of a L∞-algebroid structure on the shifted cotangent bundle N[−1], or more precisely, on the complex A0,• (N), whose Chevalley-Eilenberg complex is exactly the Dolbeault dga A•(Y(∞) X ). However, to keep the signs simple, we will work with the L∞[1]-algebroid structure on the unshifted normal X (N[−1]) = A0,•−1 X bundle rather than the L∞-algebroid. Since throughout this section the background algebra is the Dolbeault dga (A0,•(X), ¶ ) of the submanifold X, we will just write it as A0,•. 5.1. Conventions and notations. We follow the notations and sign conventions in [Vit14]. For any postitive integers k1, . . . , kl, let Sh(k1, . . . kl) be the set of (k1, . . . , kl)-unshuffles, i.e., permu- tations σ of set of integers {1, 2, . . . , k1 + k2 + · · · + kl} satisfying σ(k1 + · · · + ki−1 + m) < σ(k1 + · · · + ki−1 + n), ∀ 1 ≤ m < n ≤ ki, i = 1, . . . , l. Suppose V = ⊕iVi is a graded vector space over a field K of zero characteristic. Given a list of homogeneous vectors in v = (v1, . . . , vn) in V and a permutation σ ∈ Sn, we denote by α(σ, v) (resp., χ(σ, v)) the sign determined by vσ(1) ⊙ · · · ⊙ vσ(n) = α(σ, v)v1 ⊙ · · · ⊙ vn (resp. vσ(1) ∧ · · · ∧ vσ(n) = χ(σ, v)v1 ∧ · · · ∧ vn) where ⊙ (resp., ∧) is the graded symmetric (resp., graded skew-symmetric) product in the graded symmetric algebra S(V) (resp., graded exterior algebra ∧V). 5.2. L∞- and L∞[1]-algebroids. We recall the definition of L∞-algebras and L∞[1]-algebras ([Vit14]). Definition 5.1. An L∞-algebra is a graded vector space L• equipped with a family of n-ary multilinear operations (n-brackets) ln : (L•)×n → L•, (x1, · · · , xn) 7→ [x1, . . . , xn]n, n ∈ N of degree 2 − n, such that (1) [·, . . . , ·]n is graded skew-symmetric, i.e., for any permutation σ ∈ Sn and vector v = (v1, . . . , vn) of homogeneous elements in L•, [vσ(1), . . . , vσ(n)]n = χ(σ, v)[v1, . . . , vn]n, and (2) the higher Jacobi identity is satisfied: (−1)i jχ(σ, v)[[vσ(1), . . . vσ(i)]i, vσ(i+1), . . . , vσ(n)] j+1 = 0, (5.1) i+ j=n σ∈Sh(i, j) for any n ∈ N. In particular, (L•, d = [ · ]1) is a cochain complex. (cid:229) (cid:229) DOLBEAULT DGA AND L∞-ALGEBROID OF THE FORMAL NEIGHBORHOOD 23 Definition 5.2. An L∞[1]-algebra is a graded vector space L• equipped with a family of n-ary multilinear operations (n-brackets) ℓn : (L•)×n → L•, (x1, · · · , xn) 7→ {x1, . . . , xn}n, n ∈ N of degree 1, such that (1) {·, . . . , ·}n is graded symmetric, i.e., for any permutation σ ∈ Sn and homogeneous vectors v = (v1, . . . , vn) in L•, {vσ(1), . . . , vσ(n)}n = α(σ, v){v1, . . . , vn}n, and (2) the higher Jacobi identity is satisfied: α(σ, v){{vσ(1), . . . vσ(i)}i, vσ(i+1), . . . , vσ(n)} j+1 = 0, (5.2) i+ j=n σ∈Sh(i, j) for any n ∈ N. In particular, (L•, d = { · }1) is a cochain complex. In the case when ln = 0 for all n ≥ 2, we call L• as a shifted differential graded Lie algebra or simply a shifted DGLA. There is a one-to-one correspondence between L∞-algebra structures {[·, · · · , ·]n n ∈ N} on a graded vector space L, and L∞[1]-algebra structures {{·, · · · , ·}n n ∈ N} on L[1], the shifted graded vector space of L, given by {v1, . . . , vn} = (−1)(k−1)v1+(k−2)v2+···+vk−1[v1, · · · , vn], ∀ v1, · · · , vn ∈ L, ∀ k ∈ N. Definition 5.3. A morphism f : L → L′ between the L∞[1]-algebras (L, {·, · · · , ·}k) and (L′, {·, · · · , ·}′ k) is a collection f = ( fn, n ∈ N) of n-ary, multilinear, graded symmetric maps of degree 0, fn : L×n → L′, n ∈ N, α(σ, v) fi+ j+1({vσ(1), . . . , vσ(i)}, vσ(i+1), . . . , vσ(i+ j)) α(σ, v){ fn1 (vσ(1), . . . , vσ(n1)), . . . , fkl (vσ(k−kl +1), . . . , vσ(k))}′, (5.3) such that i+ j=n σ∈Si, j k(cid:229) = l=1 n1+···+nl=n 1≤n1≤···≤nl σ∈Sh(n1,...,nl) for any v ∈ L×k. Definition 5.4. An L∞[1]-algebroid or a strong homotopy (SH) Lie-Rinehart algebra (LR∞[1]-algebra) is a pair (L•, A•), where A• is a unital graded commutative K-algebra and L• is an L∞[1]- algebra with K-multilinear n-brackets {·, · · · , ·}n, n ∈ N, which also possesses the structure of (cid:229) (cid:229) (cid:229) (cid:229) (cid:229) (cid:229) 24 SHILIN YU a graded A•-module. Moreover, there is a family of n-ary K-multilinear operations {·, · · · , · -}n : L×(n−1) × A• → A•, n ∈ N of degree 1 such that (1) Each {·, · · · , · -}n is A•-multilinear subject to the Koszul sign rules and graded sym- metric in the first n − 1 entries and a derivation in the last entry. When n = 1, the operation { -}1 : A• → A• does not depend on L• and determines a derivation dA on A• of degree one. Hence A = (A•, dA) is a unital commutative dga. When n ≥ 1, the induced maps (of degree 0) αn : L×n → Der•(A)[1], (v1, . . . , vn) 7→ {v1, . . . , vn -}n+1, (5.4) form a morphism α of L∞[1]-algebras between L• and shifted DGLA Der•(A)[1] of derivations of A. We call α as the ∞-anchor map (or simply anchor map) and αn the nth anchor map of the L∞[1]-algebroid L. Moreover, αn is A-multilinear. (2) The brackets of L and the anchor map satisfies the equality {v1, . . . , vn−1, a · vn}n = {v1, . . . , vn−1a}n · vn + (−1)a(v1+···+vn−1+1)a · {v1, . . . , vn}n (5.5) for any n ∈ N. Note that when n = 1, this means (L, dL = {·}1) is a dg-module over the dga A = (A•, dA). Remark 5.5. Unlike [Vit14], we always write explicitly the underlying dga A = (A•, dA) and we also call (L•, dL) as an L∞[1]-algebroid over the dga A if the operation {−}1 equals dA. We now describe the (completed) Chevalley-Eilenberg dga of an L∞[1]-algebroid. Instead of the multi-differential algebra structure on the uncompleted symmetric algebra used in [Vit14], we use the completed symmetric algebra. The difference is minor. Let A• be a unital graded commutative K-algebra and L• a graded A•-module. Let Sr A(L, A) be the graded A• module of graded symmetric, A•-multilinear maps with r entries. A homo- geneous element η ∈ Sr A(L, A) is a homogeneous, graded symmetric, K-multilinear map such that η : (L•)×r → A• η(av1, v2, . . . , vr) = (−1)aηaη(v1, v2, . . . , vr), a ∈ A•, v1, . . . , vr ∈ L•. A(L, A) = A• and S1 A(L, A) = L∨• := HomA•(L•, A•). Define the completed In particular, S0 symmetric algebra SA(L, A) = (cid:213) r≥0 Sr A(L, A), DOLBEAULT DGA AND L∞-ALGEBROID OF THE FORMAL NEIGHBORHOOD 25 where the product of two homogenous elements η ∈ Sr the formula A(L, A) and η′ ∈ Sr′ A•(L•, A) is given by (ηη′)(v1, . . . , vr+r′ ) = (cid:229) σ∈Sr,r′ (−1)η′(vσ(1)+···+vσ(r))α(σ, v)η(vσ(1), . . . , vσ(r))η′(vσ(r+1), . . . , vσ(r+r′)), for any v = (v1, . . . , vr+r′) ∈ (L•)×(r+r′). SA(L, A) is a graded commutative unital algebra. Remark 5.6. Suppose L• = A• ⊗A0 L0, where A0 and L0 are the zeroth component of A• and L• respectively, and L0 is a projective and finitely generated A0-module. Then SA(L, A) ≃ A• ⊗A0 SA0((L0)∨) as A•-modules, where ⊗ is the complete tensor product with respect to the projective topology on SA0((L0)∨). The following theorem is an analogue of Theorem 12, [Vit14]. Theorem 5.7. Let A• be an graded commutative unital K-algebra and L• be a projective and finitely generated A•-module. Then an LR∞[1]-algebra structure on (L•, A•) is equivalent to a degree one derivation on SA(L, A). Sketch of proof. We only recall the construction of the derivation from [Vit14], which will be used later. For the detailed proof see the Appendix A of [Vit14]. Since L• is projective and finitely generated, SA(L, A) ≃ SA(L∨), the completed symmetric algebra generated by the A•-module L∨•, and any derivation is hence determined by its action on A• and L∨•. Define degree one derivation Dn on SA(L, A) as follows. For any a ∈ A•, set (Dna)(v1, . . . , vn) := (−1)a(v1+···+vn){v1, . . . , vna}n+1, ∀ v1, . . . , vn ∈ L•, n ≥ 0. We have Dna ∈ Sn A(L∨). Note that D0a = dAa. For any η ∈ L∨•, set n+1 i=1 (Dnη)(v1, . . . , vn+1) := (−1)θ{v1, . . . , bvi, . . . , vn+1η(vi)}n+1 + (−1)ηη({v1, . . . , vn+1}), where θ := η(v1 + · · · + cvi + · · · + vn+1) + vi(vi+1 + · · · + vn+1), v1, . . . , vn+1 ∈ L•, n ≥ 0, and a hat b· stands for omission. We have Dnη ∈ Sn+1(L∨). Note that D0 restricted on L∨• is the differential dL∨ induced from dL on L•. The unique extension of Dn as a degree one derivation on SA(L, A) satisfies the higher Chevalley-Eilenberg formula: (Dnη)(v1, . . . , vn+r) := (cid:229) (−1)η(vσ(1)+···+vσ(n))α(σ, v){vσ(1), . . . , vσ(n)η(vσ(n+1), . . . , vσ(n+r))}n+1 (5.6) (−1)ηα(τ, v)η({vτ(1), . . . , vτ(n+1)}n+1, vτ(n+2), . . . , vτ(n+r)), σ∈Sh(n,r) − τ∈Sh(n+1,r−1) for any η ∈ Sr A(L, A), v = (v1, . . . , vn+k) ∈ (L•)×(n+r). (cid:229) (cid:229) 26 SHILIN YU It is proved in [Vit14] that (cid:229) j+k=n D jDk = 0 for all n ≥ 0. Hence we can define the degree 1 derivation D = (cid:229) n≥0 Dn on SA(L, A) and D2 = 0. Conversely, any degree 1 derivation D on SA(L, A) can be written as D = (cid:229) n≥0 Dn, where Dn maps Sr A(L, A) to Sr+n A (L, A). For any a ∈ A• and v1, . . . , vn ∈ L•, set {v1, . . . , vn−1a}n := (−1)a(v1+···+vn−1)(Dn−1a)(v1, . . . , vn−1) ∈ A• (5.7) and let {v1, . . . , vn}n be the unique element in L• satisfying η({v1, . . . , vn}n) :=(−1)η n(cid:229) n(cid:229) i=1 i=1 (−1)vi(v1+···+vi−1)Dn−1(η(vi))(v1, . . . , bvi, . . . , vn) (5.8) − (−1)η(Dn−1η)(v1, . . . , vn), for any η ∈ L∨• (here we use the projectivity and finiteness of L•). (cid:3) Definition 5.8. The dga ( SA(L, A), D) determined by an L∞[1]-algebroid (L•, A•) in Theorem 5.7 is called the (completed) Chevalley-Eilenberg dga of L•. Remark 5.9. If the dga structure A = (A•, dA) is fixed in advance, an L∞[1]-algebroid structure over A on L• is equivalent to a derivation D on SA(L, A), whose zeroth component D0 acts on A• as dA and acts on L∨• as dL∨ . 5.3. L∞[1]-algebroid of the formal neighborhood. We now repackage the differential D in the formula (4.20) in the language of L∞[1]-algebroids. The underlying dga A is the Dol- beault dga (A0,•(X), ¶ ), hence the shifted DGLA A0,•+1 (TX) of shifted Dolbeaut complex of the tangent bundle TX equipped with the usual Lie bracket is a shifted dg-Lie subaglebra in- side Der•(A)[1]. The underlying A-module of the L∞[1]-algebroid is the Dolbeault complex X (N), ¶ ) of the normal bundle N of X inside Y. The structure maps are given (L•, dL) = (A0,• X (SrN), using as follows. First of all, we compute the values of the anchors on Sr (4.21), (5.4) and (5.7), and get the recursive formulas, A(L, A) = A0,0 X α1 = R⊤ 1 = β : A0,0 X (N) → A0,1 X (TX) ⊂ Der1(A), αn = R⊤ n + σ∈Sh(n−1,1) SN ◦ (αn−1 × 1) ◦ σ : A0,0 X (N)×n → A0,1 X (TX) ⊂ Der1(A), (5.9) n ≥ 2, (5.10) where σ acts as permutation on the n A0,0 operator.) Then we extend αn to an A-multilinear map from A0,• to the Koszul sign rules. X (N)-factors and SN : TX ⊗ N → TX is the shape (TX) subject X (N)×n to A0,•+1 X (cid:229) DOLBEAULT DGA AND L∞-ALGEBROID OF THE FORMAL NEIGHBORHOOD 27 Similarly, by (5.8), (4.20) and the formulas above for α, we get the formulas for the n-ary brackets, ℓn = R⊥ n + σ∈Sh(n−1,1) ∇⊥ ◦ (αn−1 × 1) ◦ σ : A0,0 X (N)×n → A0,1 X (N), n ≥ 2, (5.11) where the connection ∇⊥ on the normal bundle is considered as a map ∇⊥ : A0,• X (TX) ⊗C A0,• X (N) → A0,• X (N) of degree zero. Then we can extend the brackets uniquely such that it satisfies the condition (1) of Definition 5.4. REFERENCES [Blo10] J. Block, Duality and equivalence of module categories in noncommutative geometry, A celebration of the math- ematical legacy of Raoul Bott, CRM Proc. Lecture Notes, vol. 50, Amer. Math. Soc., Providence, RI, 2010, pp. 311–339. [CCT14] D. Calaque, A. Caldararu, and J. Tu, On the Lie algebroid of a derived self-intersection, Adv. Math. 262 (2014), 751–783. MR 3228441 [Gri14] J. Grivaux, The Hochschild-Kostant-Rosenberg isomorphism for quantized analytic cycles, Int. Math. Res. Not. IMRN (2014), no. 4, 865–913. MR 3168398 [Kap99] M. Kapranov, Rozansky-Witten invariants via Atiyah classes, Compositio Math. 115 (1999), no. 1, 71–113. [Kje01a] L. Kjeseth, A homotopy Lie-Rinehart resolution and classical BRST cohomology, Homology Homotopy Appl. 3 (2001), no. 1, 165–192. MR 1854643 (2002i:17030) [Kje01b] , Homotopy Rinehart cohomology of homotopy Lie-Rinehart pairs, Homology Homotopy Appl. 3 (2001), no. 1, 139–163. MR 1854642 (2002i:17029) [SSS12] H. Sati, U. Schreiber, and J. Stasheff, Twisted differential string and fivebrane structures, Comm. Math. Phys. 315 (2012), no. 1, 169–213. MR 2966944 [Vit14] L. Vitagliano, On the strong homotopy Lie-Rinehart algebra of a foliation, Commun. Contemp. Math. 16 (2014), no. 6, 1450007, 49. MR 3277952 [Xin03] Y. Xin, Minimal submanifolds and related topics, Nankai Tracts in Mathematics, vol. 8, World Scientific Pub- lishing Co., Inc., River Edge, NJ, 2003. MR 2035469 (2004m:53112) [Yua] S. Yu, Dolbeault dga of a formal neighborhood, Accepted by Transactions of the American Mathematical Soci- ety. [Yub] [Yu15] , Todd class via homotopy perturbation theory, in preparation. , The Dolbeault dga of the formal neighborhood of the diagonal, J. Noncommut. Geom. 9 (2015), no. 1, 161–184. MR 3337957 DEPARTMENT OF MATHEMATICS, UNIVERSITY OF PENNSYLVANIA, PA 19104-6395, USA E-mail address: [email protected] (cid:229)
1907.03249	1	1907	2019-07-07T08:38:53	Higher order polars of quasi-ordinary singularities	[ "math.AG" ]	A quasi-ordinary polynomial is a monic polynomial with coefficients in the power series ring such that its discriminant equals a monomial up to unit. In this paper we study higher derivatives of quasi-ordinary polynomials, also called higher order polars. We find factorizations of these polars. Our research in this paper goes in two directions. We generalize the results of Casas-Alvero and our previous results on higher order polars in the plane to irreducible quasi-ordinary polynomials. We also generalize the factorization of the first polar of a quasi-ordinary polynomial (not necessary irreducible) given by the first-named author and Gonz\'alez-P\'erez to higher order polars. This is a new result even in the plane case. Our results remain true when we replace quasi-ordinary polynomials by quasi-ordinary power series.	math.AG	math	Higher order polars of quasi-ordinary singularities Evelia R. Garc´ıa Barroso and Janusz Gwo´zdziewicz July 9, 2019 Abstract A quasi-ordinary polynomial is a monic polynomial with coefficients in the power series ring such that its discriminant equals a monomial up to unit. In this paper we study higher derivatives of quasi-ordinary poly- nomials, also called higher order polars. We find factorizations of these polars. Our research in this paper goes in two directions. We generalize the results of Casas-Alvero and our previous results on higher order polars in the plane to irreducible quasi-ordinary polynomials. We also generalize the factorization of the first polar of a quasi-ordinary polynomial (not necessary irreducible) given by the first-named author and Gonz´alez-P´erez to higher order polars. This is a new result even in the plane case. Our results remain true when we replace quasi-ordinary polynomials by quasi- ordinary power series. 1 Introduction In [Me] Merle gave a decomposition theorem of a generic polar curve of an ir- reducible plane curve singularity, according to its topological type. The factors of this decomposition are not necessary irreducible. Merle's decomposition was generalized to reduced plane curve germs by Kuo and Lu [K-Lu], Delgado de la Mata [D], Eggers [Eg], Garc´ıa Barroso [GB] among others. In [GB-GP], Garc´ıa Barroso and Gonz´alez P´erez obtained decompositions of the polar hypersurfaces of quasi-ordinary singularities. On the other hand, Casas-Alvero in [Ca] gener- alized the results of Merle to higher order polars of an irreducible plane curve. In [GB-Gw4] we improved his results giving a finer decomposition in such a way that we are able to determine the topological type of some irreducible factors of the polar as well as their number. 2000 Mathematics Subject Classification: Primary 32S05; Secondary 14H20. Key words and phrases: quasi-ordinary polynomial, higher order polar, factorization, P-contact, self-contact. The first-named author was partially supported by the Spanish Project MTM 2016-80659-P. 1 Our research in this paper goes in two directions. We generalize the results of [Ca] and [GB-Gw4] on higher order polars to irreducible quasi-ordinary sin- gularities (see Theorem 10.11 and Proposition 10.12). We also generalize the factorization of the first polar of a quasi-ordinary singularity (not necessary ir- reducible) from [GB-GP] to higher order polars (see Theorem 10.4). This is a new result even in the plane case. Our approach is based on Kuo-Lu trees, Eggers trees, Newton polytopes and resultants. As it was remarked in [Po] and [GB-GP], the irreducible factors of the polar of a quasi-ordinary singularity are not necessary quasi-ordinary. For that reason, we mesure the relative position of these irreducible factors and those of the quasi-ordinary singularity using a new notion called the P-contact, which plays in our situation the role of the logarithmic distance introduced by P loski in [P l]. The paper is organized as follows. In Section 2 we recall the notion of the Newton polytope of a Weierstrass poly- nomial f ∈ K[[x]][y] and we use it together with the Rond-Schober irreducibility criterium [R-S], in order to give sufficient conditions for the reducibility of f . The most important result in this section is Corollary 2.6, which allows us to characterize, in Theorem 9.1, the irreducible factors of the higher order polars of the polynomial f . In Section 3 we present the notion of the Kuo-Lu tree of a quasi-ordinary Weier- strass polynomial. Then in Section 4 we identify the bars of a Kuo-Lu tree with certain sets of fractional power series called pseudo-balls and we introduce the notion of compatibility of a Weierstrass polynomial with a pseudo-ball. Ev- ery quasi-ordinary Weierstrass polynomial is compatible with every pseudo-ball associated with its Kuo-Lu tree. Moreover if a Weierstrass polynomial is com- patible with a pseudo-ball then any factor of it is compatible too (see Corollary 4.5). In Lemma 4.6 we prove that, under some conditions, the normalized higher derivatives inherit the compatibility property. In Section 5 we introduce, using Galois automorphisms, an equivalence relation in the set of pseudo-balls, called conjugacy, and we explore the compatibility property for conjugate pseudo-balls. We generalize the Kuo-Lu Lemma [K-Lu, Lemma 3.3] to higher derivatives in Section 6. In Section 7 we introduce our main tool, monomial substitutions, that allows us to reduce several questions to the case of two variables. In particular, if f and g are power series in d + 1 variables such that after generic monomials substitutions we obtain power series ¯f , ¯g in two variables with equal Newton polygons, then the Newton polytopes of f and g are also equal (see Corollary 7.3). In Section 8 we extend the notion of Eggers tree introduced in [Eg], to quasi-ordinary settings. Remark that the tree we use here is not exactly the Eggers-Wall tree introduced in [Po] for the quasi-ordinary situation. The main result of Section 9 is Theorem 9.1, where we characterize the irreducible factors of higher derivatives of quasi-ordinary Weierstrass polynomials. Theorem 9.1 allows us to give factorizations of higher derivatives, in terms of the Eggers tree, in Section 10. Theorem 10.4 generalizes the factorization from [Ca] on higher or- 2 der polars to quasi-ordinary singularities (not necessary irreducible) and also the factorization from [GB-GP] to higher order polars. Theorem 10.11 and Propo- sition 10.12 extend the statements of [GB-Gw4, Theorem 6.2] to irreducible quasi-ordinary Weierstrass polynomials. Finally in Section 11 we establish that our results also hold for quasi-ordinary power series. d , with i = (i1, . . . , id). The ≥0. By convention the Newton polytope of the zero power series is the empty set. 2 Newton polytopes Let α = P αixi ∈ S[[x]] be a non zero formal power series with coefficients in a ring S, where x = (x1, . . . , xd) and xi = xi1 Newton polytope ∆(α) ⊂ Rd of α is the convex hull of the set Sαi6=0 i + Rd The Newton polytope of a polynomial f = Pi,j ai,jxiyj ∈ S[[x]][y] is the poly- tope ∆(f ) ⊂ Rd × R of f viewed as a power series in x1, . . . , xd, y. If Γ is a compact face of ∆(f ) then fΓ := P(i,j)∈Γ ai,jxiyj ∈ S[x][y] is called the sym- bolic restriction of f to Γ. 1 ··· xid We say that a subset of Rd+1 is a Newton polytope if it is the Newton polytope of some polynomial in S[[x]][y]. Let q = (q1, . . . , qd) ∈ Qd elementary Newton polytope ≥0 and let k be a positive integer. We define the n q k o := convex hull (cid:0){ (q1, . . . , qd, 0), (0, . . . , 0, k)} + Rd+1 ≥0 (cid:1) . Its inclination is, by definiton, 1 k q. We denote by n∞k o the Newton polytope ∆(yk), which is the first orthant translated by (0, . . . , 0, k). By convention we consider it as an elementary poly- tope. Example 2.1 The elementary Newton polytope (cid:26)(4, 2) 8 (cid:27) is (0,0,8) (4,2,0) 3 A Newton polytope is polygonal if the maximal dimension of its compact faces is one. Remember that the Minkowski sum of A, B ⊂ Rd+1 is the set A + B := {a + b : a ∈ A, b ∈ B}. If a Newton polytope ∆ has a representation of the type ∆ = r Xi=1 (cid:26)qi ki(cid:27) , (1) then summing all the elementary Newton polytopes of the same inclination in (1) we obtain a unique representation, up to the order of the terms, called canonical representation of ∆. If the inclinations can be well-ordered then ∆ is polygonal. 2.1 Newton polytopes and factorizations Let K be a field of characteristic zero. We denote by K[[x1/k ]] the ring of fractional power series in d variables where all the exponents are non- negative rational numbers with denominator k ∈ N\{0}. Put K[[x1/N]] := Sk∈N\{0} K[[x1/k ]]. We will denote by , . . . , x1/k d 1 , . . . , x1/k d 1 αK[[x1/N]] = {αw : w ∈ K[[x1/N]]} the ideal of K[[x1/N]] generated by α ∈ K[[x1/N]]. A Weierstrass polynomial is a monic polynomial where the coefficients different from the leading coefficient are non-units of the formal power series. Notice that, according to this definition, the constant polynomial 1 is a Weierstrass polynomial. The next lemma gives sufficient conditions for reducibility of Weierstrass poly- nomials. One of the consequences of this lemma is that a Weierstrass polynomial with a polygonal Newton polytope admits a decomposition into coprime factors such that the Newton polytope of each factor is elementary (see Theorem 2.4, see also [GB-GP, Theorem 3]). Lemma 2.2 Let g = ym + c1ym−1 + ··· + cm ∈ K[[x]][y] be a Weierstrass poly- nomial. Assume that there exists q ∈ Qd such that ciK[[x1/N]] ⊆ xiqK[[x1/N]] for all 1 ≤ i ≤ m with equality for some i = i0, 1 ≤ i0 < m and strict inclusion for i = m. Then g has at least two coprime factors. Proof. We will apply [R-S, Theorem 2.4]. Without lost of generality we may assume that i0 is the maximal index i ∈ {1, . . . , m− 1} such that ciK[[x1/N]] = xiqK[[x1/N]]. Then the segment Γ with endpoints (0, . . . , 0, m) and (i0q, m−i0) is an edge of ∆(g). The symbolic restriction of g to Γ is the product gΓ = ym−i0 · g, where g ∈ K[x][y] is coprime with y. The associated polyhedron of g, in the sense of Rond-Schober (see [R-S, page 4732] is mq + Rd ≥0. Hence 4 the polynomial g verifies the hypothesis of [R-S, Theorem 2.4] and the lemma follows. Remark 2.3 The assumptions of Lemma 2.2 mean geometrically that the New- ton polytope ∆(g) is included in the elementary polytope (cid:26) mq m (cid:27), and ∆(g) has an edge Γ, which endpoints (0, . . . , 0, m) and (i0q, m− i0), for some 1 ≤ i0 < m. The next picture illustrates the situation: (0, m) (i0q, m − i0) (mq, 0) Theorem 2.4 Let f ∈ K[[x]][y] be a Weierstrass polynomial. Assume that ki(cid:27). i=1(cid:26)qi ∆(f ) is a polygonal Newton polytope with canonical representation Pr Then f admits a factorization f1 ··· fr, where fi ∈ K[[x]][y] are Weierstrass polynomials, not necessarily irreducible, such that ∆(fi) = (cid:26)qi ki(cid:27) for i = 1, . . . , r. Proof. Let f = g1 ··· gs be the factorization of f into irreducible Weier- strass polynomials. Since the Newton polytope of a product is the Minkowski sum of the Newton polytopes of the factors, by hypothesis we get ∆(gj) = for fixed j only one term of the previous sum is nonzero. On the other hand, i=1 bij (cid:26)qi ki(cid:27) for some bij ∈ Q≥0. By Remark 2.3 ∆(gj) is elementary, hence Pr for fixed i, we get Pj bij = 1. Put fi := Q gj, where the product runs over all ki(cid:27) for i = 1, . . . , r. gj such that bij 6= 0. Then f = f1 ··· fr, where ∆(fi) = (cid:26)qi Theorem 2.5 Let f (y), g(y) ∈ L[y] be monic polynomials, where L is a field of characteristic zero. If g(y) is irreducible in the ring L[y] then the polynomial R(T ) = Res y(T − f (y), g(y)), where Res y(−,−) denotes the resultant, is either irreducible in L[T ] or is a power of an irreducible polynomial. 5 Proof. Let y1, . . . , ym be the roots of g(y) in the algebraic closure of the field L. Then R(T ) = Qm i=1(T − f (yi)). Since L is a field of characteristic zero and g(y) is irreducible, the Galois group of the field extension L ֒→ L(y1, . . . , ym) acts transitively on the set {y1, . . . , ym}. It follows that this group acts transitively on the set {f (y1), . . . , f (ym)}. Hence if R = R1 ··· Rs is a factorization of R = R(T ) into irreducible monic polynomials in the ring L[T ] then Ri = Rj for i 6= j. Next corollary will be used in the proof of the main result of the decompositions of higher polars, which is Theorem 9.1. Corollary 2.6 Let f (y), g(y) ∈ K[[x]][y] be Weierstrass polynomials. If the re- sultant Res y(g(y), f (y)−T ) ∈ K[[x]][T ] satisfies the assumptions of Lemma 2.2, then g(y) is not irreducible in the ring K[[x]][y]. Proof. By Lemma 2.2 the polynomial R(T ) has at least two coprime factors. By Theorem 2.5, g(y), considered as a polynomial in K((x))[y], is not irreducible, thus by Gauss Lemma it is not irreducible as a polynomial in K[[x]][y]. Remark 2.7 Beata Hejmej in [He] generalizes Theorem 2.5 to polynomials with coefficients in a field of any characteristic. Hence the results of this section hold for fields of arbitrary characteristic. 3 Kuo-Lu tree of a quasi-ordinary polynomial From now on K will be an algebraically closed field of characteristic zero. Let f (y) ∈ K[[x]][y] be a Weierstrass polynomial of degree n. Such a polynomial is quasi-ordinary if its y-discriminant equals xiu(x), where u(x) is a unit in K[[x]] and i ∈ Nd. After Jung-Abhyankar theorem (see [Pa-R, Theorem 1.3]) the roots of f are in the ring K[[x1/N]] of fractional power series and we may factorize f (y) as Qn i=1(y−αi), where αi is zero or a fractional power series of nonnegative order. Put Zer f := {αi : 1 ≤ i ≤ n}. Since the differences of roots divide the discriminant, for i 6= j we have αi − αj = xqij vij (x), for some qij ∈ Qd and vij (0) 6= 0. (2) The contact of αi and αj is by definition O(αi, αj) := qij . By convention O(αi, αi) = +∞. We introduce in Qd +∞ is bigger than any element of Qd ≥0. After [B-M, Lemma 4.7], for every αi, αj, αk ∈ Zer f one has O(αi, αk) ≤ O(αj , αk) or O(αi, αk) ≥ O(αj , αk). Moreover, we have the strong triangle inequality: ≥0 the partial order: q ≤ q′ if q′ − q ∈ Qd ≥0. By convention O(αi, αj) ≥ min{O(αi, αk), O(αj , αk)}. (STI) 6 In general, we say that the contact between the fractional power series α and β is well-defined if and only if α−β = xqw(x), for some q ∈ Qd and w ∈ K[[x1/N]] such that w(0) 6= 0. In such a case we put O(α, β) = q. Now we construct the Kuo-Lu tree of a quasi-ordinary Weierstrass polynomial f . ≥0 we put αi ≡ αj mod q+ if O(αi, αj) > q, for αi, αj ∈ Zer f . Let Given q ∈ Qd h0 be the minimal contact between the elements of Zer f . We represent Zer f as a horizontal bar B0 and call h(B0) the height of B0. The equivalence relation ≡ mod h(B0)+ divides B0 = Zer f into cosets B1, . . . , Br. We draw r vertical segments from the bar B0 and at the end of the jth vertical segment we draw a horizontal bar which represents Bj. The bar Bj is called a postbar of B0 and in such a situation we write B0 ⊥ Bj. We repeat this construction recursively for every Bj with at least two elements. The set of bars ordered by the inclusion relation is a tree. Following [K-Lu] we call this tree the Kuo-Lu tree of f and denote it T (f ). The bar B0 of minimal height is called the root of T (f ). For every bar B of T (f ) there exists a unique sequence B0 ⊥ B′ ⊥ B′′ ⊥ ··· ⊥ B, starting in B0 and ending in B. In the above construction, we do not draw the bars {αi} ⊂ Zer f . These bars are the leaves of T (f ) and they are the only bars of infinite height. Let B, B′ ∈ T (f ) be such that B ⊥ B′. All fractional power series belonging to B′ have the same term with the exponent h(B). Let c be the coefficient of such term. We say that B′ is supported at the point c on B and we denote it by B ⊥c B′. Observe that different postbars of B are supported at different points. This construction is adapted from [K-Lu] to quasi-ordinary case. Example 3.1 Let f = f1f2 ∈ C[[x1, x2]][y], where f1 = y2 − x3 y − x5 4x9 1x6 The Kuo-Lu tree of f is: 2 and f2 = 2. Observe that f is quasi-ordinary since its y-discriminant equals 1x2 2. 1 x2, β = −x3/2 1x2 2(−1 + x7 1x2 2)2. The roots of f are α = x3/2 1 x2 and γ = x5 1x2 α β γ (cid:0) 3 2 , 1(cid:1) T (f ) In the above picture we draw also a vertical segment supporting T (f ) called by Kuo and Lu in [K-Lu] the main trunk of the tree. 7 4 Compatibility with pseudo-balls Let α ∈ K[[x1/N]] be a fractional power series and h ∈ Qd ≥0. The pseudo- ball centered in α and of height h is the set α + xhK[[x1/N]]. The pseudo-ball centered in α of infinite height is the set {α}. Let f be a quasi-ordinary polynomial f . Consider the bar B = {αi1, . . . , αis} with finite height h of the Kuo-Lu tree T (f ). Set B := α + xhK[[x1/N]], where α ∈ B. As αik − αil ∈ xhK[[x1/N]] for 1 ≤ k ≤ l ≤ s the pseudo-ball B is independent of the choice of α. If B = {αi} is a bar of infinite height then we put B = B. The mapping B → B is a one-to-one correspondence between T (f ) and the set of pseudo-balls T (f ) := {αi + (αi − αj)K[[x1/N]] : αi, αj ∈ Zer f}. For the purposes of this article it is easier to deal with pseudo-balls, hence from now on, we shall identify the elements of T (f ) with corresponding pseudo-balls. Such pseudo-balls will be called quasi-ordinary pseudo-balls. Let B = α + xh(B)K[[x1/N]] be a quasi-ordinary pseudo-ball of finite height. Every γ ∈ B has a form γ = λB(x) + cγxh(B) + ··· , where λB(x) is obtained from any β ∈ B by omitting all the terms of order bigger than or equal to h(B) and ellipsis means terms of higher order. We call the number cγ the leading coefficient of γ with respect to B and denote it lcB(γ). Remark that cγ can be zero. Let L be the field of fractions of K[[x]]. It follows from [GP, Remark 2.3] that any truncation of a root of a quasi-ordinary polynomial is a root of a quasi- ordinary polynomial. Hence the field extensions L ֒→ L(λB(x)) ֒→ L(λB(x), xh(B)) are algebraic and we can associate with B two numbers: • the degree of the field extension L ֒→ L(λB(x)) that we will denote N (B), • the degree of the field extension L(λB(x)) ֒→ L(λB(x), xh(B)) that we will denote n(B). In this section we introduce the notion of compatibility of a Weierstass polyno- mial g with a pseudo-ball B. We define a polynomial GB(z) which will play an important role in the sequel. Definition 4.1 Let g(y) ∈ K[[x]][y] be a Weierstrass polynomial and B be a pseudo-ball of finite height. If g(λB(x) + zxh(B)) = GB(z)xq(g,B) + ··· (3) for some GB(z) ∈ K[z] \ {0} and some exponent q(g, B) ∈ (Q≥0)d then we will say that g is compatible with B. In (3) ··· means terms of higher order. The polynomial GB(z) will be called the B-characteristic polynomial of g. Example 4.2 Return to Example 3.1. Let B = α + x3/2 pseudo-ball of T (f ) of height h(B) = (cid:0) 3 f (λB + zxh(B)) = f (zx( 3 2 , 1(cid:1). Observe that 2 ,1)) = z(z2 − 1)x9/2 1 x3 1 x2K[[x1/N]] be a 2 + ··· 8 Hence the polynomial f is compatible with the pseudo-ball B and its B-characteristic polynomial is FB(z) = z(z2− 1), but for example the polynomial g(y) = y− x1− x2 is not compatible with B. Our next goal is to prove in Corollary 4.5 that if a Weierstrass polynomial is compatible with a pseudo-ball then any factor of it is also compatible. Lemma 4.3 Let g(y) ∈ K[[x]][y] be a Weierstrass polynomial and let B be a pseudo-ball of finite height. Consider g(λB(x) + zxh(B)) as a fractional power series g(x) with coefficients in K[z]. Then g(y) is compatible with B if and only if the Newton polytope of g(x) equals the Newton polytope of a monomial. Proof. If g is compatible with B then by (3) we get ∆(cid:0)g(x)(cid:1) = ∆(cid:0)xq(g,B)(cid:1). Conversely, suppose that the Newton polytope of g(x) equals the Newton poly- tope of the monomial xq. Then g(x) has a form xqPn i=0 ai(x)zn−i, where at least one of the values ai(0) is nonzero. Hence the B-characteristic polynomial of g is GB(z) = Pn GB(z)xq(g,B) +Ph>q(g,B) ah(z)xh, where ah(z) ∈ K[z]. Corollary 4.5 Let g ∈ K[[x]][y] be a Weierstrass polynomial compatible with a pseudo-ball B. Then any factor of g is compatible with B. Remark 4.4 From the proof of Lemma 4.3 we get that g(x) has the form i=0 ai(0)zn−i. Proof. The Newton polytope of the product is the Minkowki sum of Newton polytopes of the factors. Hence, if ∆(g) = ∆(xq) and g = g1g2 then ∆(gi) have the form ∆(xqi) for some q1, q2 such that q = q1 + q2. Next lemma generalizes to d variables [GB-Gw4, Lemma 3.1]. Lemma 4.6 Let f (y) ∈ K[[x]][y] be a Weierstrass polynomial of degree n com- patible with the pseudo-ball B. Then for every k ∈ {1, . . . , deg FB(z)} the dk Weierstrass polynomial g(y) = (n−k)! dyk f (y) is also compatible with B and its B-characteristic polynomial is GB(z) = (n−k)! n! dk dzk FB(z). n! Proof. Differentiating identity f (λB(x) + zxh(B)) = FB(z)xq(f,B) + ··· with respect to z we get f ′(λB(x) + zh(B))xh(B) = F ′ B(z)xq(f,B) + ··· . Hence f ′(λB(x) + zxh(B)) = F ′ B(z)xq(f,B)−h(B) + ··· , which proves the lemma for k = 1. The proof for higher derivatives runs by induction on k. Let f (y) ∈ K[[x]][y] be a Weierstrass polynomial of degree n. The Weier- dk strass polynomial (n−k)! dyk f (y) of Lemma 4.6 will be called the normalized kth derivative of the Weierstrass polynomial f (y) ∈ K[[x]][y] and we will denote it by f (k)(y). The variety of equation f (k) = 0 is called the kth polar of f = 0. Since the normalized nth derivative of f is constant, in the rest of the paper we consider normalized kth derivatives of f for 1 ≤ k < deg f . n! 9 Lemma 4.7 Let f (y) ∈ K[[x]][y] be a Weierstrass polynomial and let B be a pseudo-ball of finite height. 1. If f is compatible with B, then for any γ ∈ B we have f (γ) = FB(lcBγ)xq(f,B) + ··· (4) 2. If f (y) = Qn i=1(y−αi) and we assume that one of the following holds: x is a single variable and B is arbitrary or f is quasi-ordinary and B ∈ T (f ) then f is compatible with B and we have and FB(z) = const Yi:αi∈B (z − lcBαi) q(f, B) = n Xi=1 min(O(λB , αi), h(B)). (5) (6) Proof. Since γ ∈ B we can write γ = λB(x) + zxh(B), where z = lcB(γ) + ··· . By Remark 4.4 we have f (γ) = f (λB(x) + zxh(B)) = FB(z)xq(f,B) + ··· = FB(lcBγ)xq(f,B) + ··· . This proves (4). Suppose γ = λB(x) + zxh, where z is a constant. We have f (γ) = Qn i=1(γ − αi). In order to prove (5) and (6), it is enough to compute the initial term of every factor γ−αi. If αi ∈ B then the initial term of γ−αi equals (lcBγ−lcBαi)xh(B). Otherwise the initial terms of γ − αi and λB − αi are equal. We finish the proof multiplying the initial terms. Corollary 4.8 Let f (y) ∈ K[[x]][y] be a quasi-ordinary Weierstrass polynomial. Then every factor of f (y) is compatible with all bars B ∈ T (f ) of finite height. Lemma 4.9 Let f (y) ∈ K[[x]][y] be a quasi-ordinary Weierstrass polynomial, p(y) be a factor of f (y) and B, B′ be bars of finite heights in T (f ) such that B ⊥ B′. Then q(p, B′) − q(p, B) = ♯(Zer p ∩ B′)[h(B′) − h(B)]. Proof. Put p(y) = Qα∈Zer p(y − α). Let γ ∈ B, γ′ ∈ B′ be such that O(γ, α) = h(B) for all α ∈ B ∩ Zer p and cont(γ′, α) = h(B′) for all α ∈ B′ ∩ Zer p. By the STI we get O(γ′, α) = O(γ, α) for any α ∈ Zer p\B′. If α ∈ Zer p ∩ B′ then O(γ, α) = h(B) and O(γ′, α) = h(B′). Hence q(p, B′) − q(p, B) = Xα∈Zer p O(γ′, α) − Xα∈Zer p O(γ, α) = ♯(Zer p ∩ B′)[h(B′) − h(B)]. Lemma 4.9 is similar in spirit to [GB-Gw-L, Lemma 2.7]. 10 Lemma 4.10 Let B be a quasi-ordinary pseudo-ball and let g(y) ∈ K[[x]][y] be a Weierstrass polynomial compatible with B. Then 1. GB(z) = zk · H(zn(B)), for some k ∈ N and H(z) ∈ K[z]. 2. If g is irreducible and quasi-ordinary then GB(z) = azk or GB(z) = a(zn(B) − c)l, for some non-zero a, c ∈ K and some l ∈ N. Proof. Let L be the field of quotients of K[[x]]. By [Li, Lemma 5.7] and [GP, Remark 2.7] the algebraic extension L(λB(x)) ֒→ L(λB(x), xh(B)) is cyclic. Hence the generator ϕ of the group Gal(L(λB(x) ֒→ L(λB (x), xh(B)) acts as follows: ϕ(λB(x)) = λB(x) and ϕ(xh(B)) = ωxh(B), where ω is a primitive n(B)th root of the unity. Applying ϕ to (3) we get g(λB(x) + zωxh(B)) = GB(z)ωkxq(g,B) + ··· (7) for some 0 ≤ k < n(B). Substituting ωz for z in (3) and comparing with (7) we get GB(z)ωk = GB(ωz). Multiplying this equality by (ωz)n(B)−k and putting W (z) := zn(B)−kGB(z) we obtain W (z) = W (ωz). This implies that W (z) = W (zn(B)), for some W (z) ∈ K[z]. We finish the proof putting H(zn(B)) = z−n(B)W (zn(B)). This proves the first part of the lemma. Suppose now that g is irreducible and quasi-ordinary. Let γ = λB(x) + cxh(B) + ··· ∈ B ∩ Zer g. Since the extension L(λB(x)) ֒→ L(λB(x), xh(B)) ֒→ L(γ) is Galois, any other root of g belonging to B has the form λB(x) + ωicxh(B) +··· , for some 0 ≤ i < n(B). Using the first part of the lemma and the equality (5) we complete the proof. 5 Conjugate pseudo-balls In this section we define an equivalence relation between pseudo-balls called conjugacy relation. This will allow us to introduce, in Section 8, the notion of the Eggers tree of a quasi-ordinary Weierstrass polynomial. Let L be the field of fractions of K[[x]] and M be the field of fractions of K[[x1/N]]. Lemma 5.1 Let ϕ be an L-automorphism of M. Then 1. For any q ∈ Qd there exists a root ω of the unity such that ϕ(xq) = ω· xq, 2. ϕ(K[[x1/N]]) = K[[x1/N]], 3. If u is a unit of the ring ∈ K[[x1/N]] and q ∈ (Q≥0)d then ϕ(u·xq) = u·xq for some unit u ∈ K[[x1/N]]. Proof. Let k be a positive integer. Observe that xi = ϕ(xi) = ϕ(cid:0)(x1/k )k(cid:1) = for some c ∈ K\{0} such that ck = 1. It )k. Hence ϕ(x1/k ) = c · x1/k i i ϕ(x1/k i i 11 follows that for any q ∈ Qd there exists ω ∈ K such that ϕ(xq) = ωxq and ωm = 1 for some positive integer m. Every element of the ring K[[x1/N]] can be represented as a finite sum Pq aqxq where q = (q1, . . . , qd) ∈ (Q≥0)d (0 ≤ qi < 1) and aq ∈ K[[x]]. This together with 1. proves items 2. and 3. of the lemma. Let B, B′ be pseudo-balls. We say that B and B′ are conjugate if there exists an L-automorphism ϕ of M such that B′ = ϕ(B). The conjugacy of pseudo-balls is an equivalence relation. It follows from Lemma 5.1 that conjugate pseudo- balls have the same height. Moreover two quasi-ordinary pseudo-balls B and B′ of the same height are conjugate if any irreducible quasi-ordinary polynomial which has one of its roots in B has another root in B′ (in this way conjugate bars were defined in [K-Pa, Definition 6.1]). If B′ = ϕ(B) then λB′ = ϕ(λB ). The converse is also true; if h(B) = h(B′) and there exists an L-automorphism ϕ of M such that λB′ = ϕ(λB) then B and B′ are conjugate. It follows from the above that the number of pseudo-balls conjugate with B is equal to the degree of the minimal polynomial of λB, which is the degree N (B) of the field extension L ֒→ L(λB(x)). Lemma 5.2 Let B, B′ be quasi-ordinary conjugate pseudo-balls. K[[x]][y] is a Weierstrass polynomial compatible with B then If p(y) ∈ 1. p(y) is compatible with B′. 2. q(p, B) = q(p, B′). 3. The characteristic polynomials PB′ (z) and PB(z) of p(y) verify the equality PB′ (z) = θPB(ωz), for some roots of the unity θ and ω. Proof. Let L be the field of quotients of K[[x]] and let ϕ be a L-automorphism of M such that ϕ(B) = B′. Then ϕ(λB) = λB′ . By Lemma 5.1 we have ϕ(xh(B)) = ω−1xh(B) and ϕ(xq(p,B)) = θxq(p,B) for some roots of the unity θ and ω. Applying ϕ to (3), with g replaced by p, we get p(λB′ + zω−1xh(B)) = PB(z)θxq(p,B) + ··· This gives q(p, B) = q(p, B′) and PB′ (ω−1z) = θPB(z). 6 Kuo-Lu Lemma for higher derivatives Let f (y) ∈ K[[x]][y] be a quasi-ordinary Weierstrass polynomial. We begin with combinatorial results concerning the Kuo-Lu tree T (f ). Remember that we identify any bar of T (f ) with the corresponding quasi-ordinary pseudo-ball. At the end of the section we apply these results to Newton-Puiseux roots of higher derivatives of f (y). Take an integer k such that 1 ≤ k ≤ deg f . With every bar B of T (f ) we associate the numbers: 12 • m(B) which is the number of roots of f (y) which belong to B, • nk(B) = max{m(B) − k, 0}, and • tk(B) = nk(B) −PB⊥B′ nk(B′). Remark 6.1 For 1 ≤ k < m(B) we have nk(B) > 0, tk(B) > 0 and for m(B) ≤ k ≤ deg f we have nk(B) = tk(B) = 0. We denote by Tk(f ) the sub-tree of T (f ) consisting of the bars B ∈ T (f ) such that m(B) ≥ k. Let F ∈ K[z] be a non constant polynomial. Let F (k) denotes the kth derivative of F . Definition 6.2 We will say that F is k-regular if one of the following condi- tions holds: 1. F (k) is zero or 2. F (k) is nonzero and there is not a root of F of multiplicity ≤ k which is a root of F (k). Recall that common roots of a polynomial F and its first derivative are multiple roots of F . Hence any polynomial is 1-regular. In general it is not easy to verify the k-regular property. In this papers polyno- mials of the form F (z) = (zn − c)l ∈ K[z], (8) play an important role. Their k-regularity, for any k, is a consequence of Lemma 10.9. Remark 6.3 Let F (z) = constQr i=1(z − zi)mi , where zi are pairwise different, mi ≥ k for 1 ≤ i ≤ s and mi < k for s < i ≤ r. After differentiating, the multiplicity of any root drops by one. Hence putting F ⊕(z) = Qs i=1(z − zi)mi−k we obtain the decomposition F (k)(z) = F ⊕(z)F ⊖(z) (9) into two coprime polynomials. A polynomial F is k-regular if and only if F and F ⊖ do not have common roots. Definition 6.4 Let f ∈ K[[x]][y] be a quasi-ordinary Weierstrass polynomial. We say that f is Kuo-Lu k-regular if for every B ∈ T (f ) of finite height the polynomial FB(z) is k-regular. 13 We finish this subsection with some results for Weierstrass polynomials with coefficients in the ring of the formal power series in one variable. Let f (y) ∈ K[[x]][y] be a square-free Weierstrass polynomial. Fix B ∈ Tk(f ) and assume that {B1, . . . , Bs} is the set of post-bars of B in Tk(f ). Denote B◦ = B \ (B1 ∪ ··· ∪ Bs). Then Theorem 6.5 Let f (y) ∈ K[[x]][y] be a square-free Weierstrass polynomial over the ring of formal power series in one variable. Let f (y) = Qn i=1(y − αi) and f (k)(y) = Qn−k j=1 (y − βj) be the Newton-Puiseux factorizations of f and f (k). (i) for every B ∈ Tk(f ) the set {j : βj ∈ B} has nk(B) elements, (ii) for every B ∈ Tk(f ) the set {j : βj ∈ B◦} has tk(B) elements, (iii) for every βj there exists a unique B ∈ Tk(f ) such that βj ∈ B◦. (iv) Let B ∈ Tk(f ). If the polynomial FB(z) is k-regular then for every αi ∈ B, βj ∈ B◦ one has O(αi, βj) = h(B). Otherwise there exist αi ∈ B, βj ∈ B◦ such that O(αi, βj) > h(B). Proof. Proof of (i). Suppose first, that B ∈ Tk(B) has finite height. Then by Lemma 4.7 FB(z) = constQi:αi∈B(z − lcB(αi)). By equality (4) of this lemma and Lemma 4.6 we get F (k) B (z) = constQj:βj ∈B(z − lcB(βj )). Hence, the set {j : βj ∈ B} has deg FB − k = nk(B) elements. If the height of B is infinite then B = {αi} for exactly one Newton-Puiseux root αi of f (y). Hence for k = 1 n1(B) = 0 and f ′(y) does not have roots in B and for k > 1 B /∈ Tk(f ). Proof of (ii). using (i). Proof of (iii). Let B0 be the root of the tree T (f ). By (i), {β1, . . . , βn−k} is a subset of B0. It is clear that the sets B◦ for B ∈ Tk(f ) are pairwise disjoint and their union is equal to B0. This proves (iii). Proof of (iv). Assume that B1, . . . , Br are the post-bars of B supported at points z1, . . . , zr respectively, and that m(Bi) ≥ k for i ∈ {1, . . . s}, m(Bi) < k for i ∈ {s + 1, . . . r}. Then by Lemma 4.7 FB(z) = Qr After Remark 6.3 the kth derivative of FB(z) is the product of two coprime polynomials It is enough to count the elements of the set {j : βj ∈ B◦} i=1(z − zi)m(Bi). F (k) B (z) = F ⊕ B (z)F ⊖ B (z), B (z) := Qs i=1(z − zi)nk(Bi). where F ⊕ We get deg F ⊖ B (z) = tk(B). Hence it follows from (ii) and (iii) that all roots of F ⊖ B (z) correspond to those Newton-Puiseux roots of f (k)(y) that belong to B◦. For αi ∈ B, βj ∈ B◦ one has O(αi, βj) > h(B) if and only if lcB(αi) = lcB(βj ), 14 which means that the polynomials FB(z) and F ⊖ FB(z) is k-regular if and only if F ⊖ we get (iv). B (z) have a common root. Since B (z) and FB(z) do not have common roots Remark 6.6 Let f (y) ∈ K[[x]][y] be a square-free Weierstrass polynomial over the ring of formal power series in one variable. Let B ∈ Tk(f ), βi ∈ B∩Zer f (k) and put c = lcBβi. Then F ⊕ B (c) = 0 then there exists a sequence of postbars B ⊥c B1 ⊥ ··· ⊥ Bl such that βi ∈ B◦ l and Bl ∈ Tk(f ). B (c) 6= 0 if and only if βi ∈ B◦. If F ⊕ For quasi-ordinary Weierstrass polynomials which are Kuo-Lu k-regular, the counterpart of [K-Lu, Lemma 3.3] is true: Corollary 6.7 Let f (y) ∈ K[[x]][y] be a square-free Weierstrass polynomial over the ring of formal power series in one variable. Assume that f is Kuo- Lu k-regular. Then under assumptions and notations of Theorem 6.5, for every αi ∈ Zer f , βs ∈ Zer f (k) there exists αj ∈ Zer f such that O(αi, βs) = O(αi, αj). 7 Newton polytopes of resultants In this section we give a formula for the Newton polytope of the resultant Res y(f (k)(y), p(y)− T ), where f (y) is a Kuo-Lu k-regular quasi-ordinary Weier- strass polynomial, p(y) is a factor of f (y) and T is a new variable. We prove that for irreducible p(y), the Newton polytope of the resultant is polygonal. 7.1 Monomial substitutions Let g(x, y) ∈ K[[x, y]]. For any monomial substitution x1 = ur1,. . . , xd = urd, where ri are positive integers, we put ¯g[r](u, y) := g(ur1, . . . , urd , y). (10) We will write simply ¯g(u, y) when no confusion can arise. Observe that for g = xs we get ¯g[r] = uhr,si, where h·,·i denotes the scalar product. Lemma 7.1 Let f (y) ∈ K[[x]][y] be a quasi-ordinary Weierstrass polynomial. There is a one-to-one correspondence between the bars of T (f ) and the bars of T ( ¯f [r]). If B and ¯B are the corresponding bars of T (f ) and T ( ¯f [r]) respectively then 1. h( ¯B) = hr, h(B)i and tk( ¯B) = tk(B). 2. For any factor g of f , the B-characteristic polynomial of g and the ¯B- characteristic polynomial of ¯g[r] are equal and q(¯g[r], ¯B) = hr, q(g, B)i. 15 Proof. Set ur = (ur1, . . . , urd). If Zer f = {αi(x)}n and O(αi(ur), αj(ur)) = hr, O(αi(x), αj (x))i for i 6= j. Hence every bar B = {αij (x)}k T ( ¯f [r]) of height hr, h(B)i. Substituting uri for xi in the equation (3) appearing in Definition 4.1, we get i=1 then Zer ¯f [r] = {αi(ur)}n j=1 of j=1 of T (f ) yields the bar ¯B = {αij (ur)}k i=1 ¯g[r](λ ¯B(u) + zuh( ¯B)) = GB(z)uhr,q(g,B)i + ··· , hence the second part of the lemma follows. The proof of the next lemma is similar in spirit to the proof of [GB-Gw3, Theorem 4.1] and the proof of [GB-Gw3, Theorem 9.2]. The same arguments were used there in special situation. Here we repeat the proof for the convenience of the reader. Lemma 7.2 Let g(x, y) ∈ K[[x, y]] and ∆ ⊆ Rd+1 be a Newton polytope. For any r ∈ (R>0)d let ¯∆[r] be the image of ∆ by the linear mapping πr : Rd×R −→ R2 given by (a, b) 7→ (hr, ai, b). If ∆(¯g[r]) = ¯∆[r] for every r ∈ (N\{0})d, then ∆(g) = ∆. Proof. For every Newton polytope ∆ ⊆ (R≥0)d+1 and every v ∈ (R≥0)d+1 we define the support function l(v, ∆) = min{hv, αi : α ∈ ∆}. To prove the lemma it is enough to show that the support functions l(·, ∆(g)) and l(·, ∆) are equal. As these functions are continuous it suffices to show the equality on a dense subset of Rd+1 ≥0 . Let ~r = (r1, . . . , rd+1) = (r, rd+1) ∈ Rd+1 Perturbing ~r a little we may assume that the hyperplane { α ∈ Rd+1 : h~r, αi = l(~r, ∆(g)} supports ∆(g) at exactly one point α = (α, αd+1). Since after a small change of ~r the support point remains the same, we can assume, perturbing ~r again if necessary, that all ri are positive rational numbers. We will show that ≥0 , where r = (r1, . . . , rd). l(~r, ∆) = l(~r, ∆(g)). (11) Multiplying ~r by the common denominator of r1, . . . , rd+1 we may assume that all ri are positive integers. At this point of the proof we fix ~r. We claim that l(~r, ∆) = l(cid:0)(1, rd+1), ¯∆[r](cid:1) and l(~r, ∆(g)) = l(cid:0)(1, rd+1), ∆(cid:0)¯g[r](cid:1)(cid:1). First equality follows from the definition of πr and the identity h~r, αi = h(1, rd+1), πr(α)i for α ∈ Rd+1. Write α = (α, αd+1) ∈ Rd+1 and g(x, y) = Pα dαxαyαd+1 ∈ K[[x, y]]. Since the hyperplane { α ∈ Rd+1 : h~r, αi = l(~r, ∆(g)} supports ∆(g) at α, the term d αuhr, αiy αd+1 of ¯g[r], satisfies the equality hr, αi + rd+1 αd+1 = l(~r, ∆(g)), while for all other terms dαuhr,αiyαd+1 with dα 6= 0 appearing in ¯g[r], we have hr, αi + rd+1αd+1 > l(~r, ∆(g)). Hence l(cid:0)(1, rd+1), ∆(g)(cid:1) = hr, αi + rd+1 αd+1 = l(~r, ∆(g)), so we get (11). 16 Corollary 7.3 Let g1(x, y), g2(x, y) ∈ K[[x, y]]. Suppose that ∆(¯g[r] for every r ∈ (N\{0})d. Then ∆(g1) = ∆(g2). Theorem 7.4 Assume that f ∈ K[[x]][y] is a Kuo-Lu k-regular quasi-ordinary Weierstrass polynomial and p is a Weierstrass polynomial which is a factor of f in K[[x]][y]. Then the Newton polytope of R(T ) := Res y(f (k)(y), p(y) − T ) ∈ K[[x]][T ] is equal to 1 ) = ∆(¯g[r] 2 ) (cid:26)tk(B)q(p, B) tk(B) (cid:27) . XB∈T (f ) tk (B)6=0 (12) Proof. First we will prove the theorem for d = 1. We use the notation of j=1 (y− βj) be the Newton-Puiseux factorization of f (k)(y). By the well-known properties of the resultants we have Theorem 6.5. Let Qn−k Res y(f (k)(y), p(y) − T ) = ± n−k Yj=1 (p(βj) − T ). (13) By Theorem 6.5, for every βj there exists a unique bar B ∈ T (f ) such that βj ∈ B◦. For such a bar, h(B) is finite and tk(B) 6= 0. By Corollary 4.5 the polynomial p is compatible with B and by (5) of Lemma 4.7 PB(z) is a factor of FB(z). By Theorem 6.5 (iv) we get that O(αi, βj) = h(B) for any αi ∈ B. Hence lcBβj does not belong to the set {lcBαi : αi ∈ B}. So by the equality (5) in Lemma 4.7 we have FB(lcBβj) 6= 0 and consequently PB(lcBβj) 6= 0. Now, using equality (4) of Lemma 4.7 we conclude that the Newton polytope of p(βj) − T is equal to (cid:26)q(p, B) 1 (cid:27). Using the property that the Newton polytope of a product is the Minkowski sum of the Newton polytopes of its factors, and (ii) of Theorem 6.5 we finish the proof for d = 1. Assume now that d > 1. Let x1 = ur1,. . . , xd = urd be a monomial substitution, where ri are positive integers. By Lemma 7.1 f [r] is Kuo-Lu k-regular, hence by the first part of the proof (d = 1) ∆( ¯R[r]) = XB∈T (f ) tk (B)6=0 (cid:26)tk( ¯B)q(¯p[r], ¯B) (cid:27) . tk( ¯B) For any elementary polytope of the above sum, Lemma 7.1 gives (cid:26)tk( ¯B)q(¯p[r], ¯B) tk( ¯B) (cid:27) = (cid:26)tk(B)hr, q(p, B)i (cid:27) = πr(cid:18)(cid:26)tk(B)q(p, B) tk(B) (cid:27)(cid:19) . tk(B) Since the image of the Minkowski sum of Newton polytopes is the Minkowski sum of the images, we get ∆( ¯R[r]) = πr(∆), where ∆ denotes the Newton polytope given in (12). By Lemma 7.2 we get ∆(R) = ∆. 17 8 Eggers tree of a quasi-ordinary Weierstrass polynomial In this section we introduce the Eggers tree of a quasi-ordinary Weierstrass polynomial f , after the conjugacy relation defined in Section 5. Denote by [B] the conjugacy class of the pseudo-ball B of the Kuo-Lu T (f ). By definition, the Eggers tree of f , denoted by E(f ), is the set of conjugacy classes with the natural order induced by the Kuo-Lu tree. This is the natural generalization of the Eggers tree associated with plane curves in [Eg]. The notion of Eggers tree, for quasi-ordinary singularities, was introduced by Popescu-Pampu in [Po]. He defined a slightly different notion of the Eggers tree, since he generalized to quasi-ordinary singularities the version of Eggers tree defined for curves in [W]. The leaves of E(f ) correspond with irreducible factors of f . Following Eggers we draw them in white color. By definition, the root of E(f ) is its vertex of minimum height. The branches of E(f ) are the smallest sub-trees of E(f ) containing the root and one of its leaves. Let [B] be a vertex in the branch of E(f ) corresponding with the irreducible componente fi of f . Eggers draws in a dashed way the edge leaving from the vertex [B] in this branch if there are not two roots of fi with contact h(B). Recall that the number of pseudo-balls conjugate with a quasi-ordinary pseudo- ball B is N (B) (see page 12). Let [B] be a vertex of the Eggers tree of a quasi-ordinary polynomial f . By Lemma 5.2, for any k ∈ {1, . . . , deg f}, the numbers nk(B) and tk(B) do not depend on the representative of [B]. Moreover, if p(y) ∈ K[[x]][y] is a Weier- strass polynomial compatible with B then the number q(p, B) and the degree of its B-characteristic polynomial are also independent of the representative of [B]. The Eggers tree of the quasi-ordinary polynomial f = f1f2 from Example 3.1 is f1 f2 [B] Remark 8.1 If p is an irreducible factor of f then, following Lemma 4.9, the sequence {q(p, B)}[B] is increasing along the branch P of the Eggers tree of f containing the leave representing p. Moreover, if [B] does not belong to P then q(p, B) = q(p, B0), where [B0] is the last common vertex of P and the branches of the Eggers tree containing [B]. Hence, the set {q(p, B)}[B] is well-ordered. After Remark 8.1 we get 18 Corollary 8.2 Let f ∈ K[[x]][y] be a Kuo-Lu k-regular quasi-ordinary Weier- strass polynomial and p a Weierstrass polynomial which is an irreducible factor of f in K[[x]][y]. Then the Newton polytope in (12) is polygonal. 9 Irreducible factors of higher derivatives Let f be a quasi-ordinary Weierstrass polynomial. In this section we study irre- ducible factors of normalized higher derivatives f (k). We show that every such an irreducible factor can be associated with a certain vertex [B] of the Eggers tree of f . By definition an Eggers factor will be the product of all irreducible factors associated with the same vertex of E(f ). The Eggers factorization of a higher derivative is the product of all its Eggers factors. It generalizes to higher derivatives the factorization of the first polar given in [Eg] and [GB] for plane curves and in [GB-GP] for quasi-ordinary polynomials. Let FB(z) be the B-characteristic polynomial of f . After Remark 6.3, the polynomial F (k) B (z), where B (z) is the product of two coprime polynomials F ⊕ B (z) and F ⊖ B (z) = YB⊥zi Bi F ⊕ (z − zi)nk(Bi). Theorem 9.1 Let f (y) be a quasi-ordinary Weierstrass polynomial and let g(y) ∈ K[[x]][y] be a Weierstrass polynomial which is an irreducible factor of f (k)(y). Then there exists [B] ∈ E(f ), with B ∈ Tk(f ), such that: 1. If B′ ∈ Tk(f )\[B] then every root of GB′ (z) is a root of F ⊕ 2. If B′ ∈ Tk(f ) ∩ [B] then GB′ (z) and F ⊕ B′(z) do not have common roots. B′(z). Moreover GB(z) = azl or GB(z) = a(zn(B) − c)l (14) for some l ≥ 1 and a, c ∈ K\{0}. If l = 1 then g(y) is quasi-ordinary. Proof. Let T = { B ∈ Tk(f ) : GB(z) has a root which is not a root of F ⊕ B (z)}. By Remark 6.6, B ∈ T if and only if for any monomial substitution ¯g has a Newton-Puiseux root that belongs to ¯B◦. Let E = { [B] ∈ E(f ) : B ∈ T }. We will show that E has only one element. Suppose that this is not the case, and let [B0] be the infimum of E in the ordered set E(f ) (the infimum exists because E(f ) has the structure of a tree). Let [B′] be an element of E different from [B0] and let p be any irreducible factor of f such that one of its roots belongs to B′. By definition of the Eggers tree there exists B1 ∈ [B0] such that B′ ( B1. Since B′ ∈ Tk(f ), one of the roots of PB1 (z) has multiplicity bigger or equal than k. Hence, by the second statement of Lemma 4.10 all the roots of PB1 (z) have this property. By (9), the polynomial PB1 (z) could only share roots with F ⊕ (z). Hence, by Remark 6.3 the polynomials PB1 (z) and F ⊖ (z) are coprime. B1 B1 19 Let ¯g = Qm i=1(y − ¯βi) be the Newton-Puiseux factorization of g after some monomial substitution. Fix B ∈ [B0]. By Lemmas 5.2 and 7.1 we get q(¯p, ¯B) = q(¯p, ¯B0). Let us define two sets of indexes associated with ¯B: I ¯B = {i : ¯βi ∈ ¯B, PB(lc ¯B J ¯B = {i : ¯βi ∈ ¯B, PB(lc ¯B ¯βi) 6= 0 }, ¯βi) = 0 }. Directly from the definition of PB we have: if i ∈ I ¯B then ord ¯p( ¯βi) = q(¯p, ¯B0), and if i ∈ J ¯B then ord ¯p( ¯βi) > q(¯p, ¯B0). The cardinality of I ¯B is equal to the number of roots of GB(z) counted with multiplicities which are not the roots of PB(z). Similarly the cardinality of J ¯B is equal to the number of roots of GB(z) counted with multiplicities which are the roots of PB(z). Hence the cardinality of these sets does not depend on the choice of the monomial substitution. Let I := SB∈[B0] I ¯B and J := SB∈[B0] J ¯B. Observe that ord ¯p( ¯βi) = q(¯p, ¯B0) for i ∈ I and ord ¯p( ¯βi) > q(¯p, ¯B0) for i ∈ J. The sets I and J depend on the choice of the monomial substitution but their cardinality does not. We will show that the set J is nonempty. Since B′ ∈ T , there exists ¯βi ∈ ¯B′. Any root of ¯p that belongs to ¯B′ has the same leading ¯βi) = 0, which gives i ∈ coefficient with respect to ¯B1 as ¯βi. Hence PB1 (lc ¯B1 J ¯B1 ⊂ J. Now we will prove that the set I is empty. Suppose that it is not the case. Put R(T ) := Res y(g, p − T ) and ¯R(T ) := Res y(¯g, ¯p − T ). We can write R(T ) = ±T m + c1T m−1 + ··· + cm, ¯R(T ) = ±T m + ¯c1T m−1 + ··· + ¯cm, for some ci ∈ K[[x]]. By a well-known formula for the resultant we have ¯R(T ) = ±Qm i=1(¯p( ¯βi) − T ). Since the Newton polygon of a product is the Minkowski sum of the Newton polygons of its factors, ∆( ¯R(T )) has an edge of inclination q(¯p, ¯B0) starting in the point (0, m). The projection of this edge to the vertical axis has length ♯I. This gives ord ¯ci ≥ iq(¯p, ¯B0) for 1 ≤ i < ♯I, ord ¯ci = iq(¯p, ¯B0) for i = ♯I, ord ¯ci > iq(¯p, ¯B0) for ♯I < i ≤ m. Since the monomial substitution was arbitrary, we have ciK[[x1/N]] ⊆ xiq(p,B0)K[[x1/N]] ciK[[x1/N]] = xiq(p,B0)K[[x1/N]] ciK[[x1/N]] ( xiq(p,B0)K[[x1/N]] for 1 ≤ i < ♯I, for i = ♯I, for ♯I < i ≤ m. By Corollary 2.6 g is not irreducible and we get a contradiction. 20 We conclude that I = ∅. This means that for every ¯βi there exists B ∈ [B0] such that ¯βi ∈ ¯B and ord ¯p( ¯βi) > q(¯p, ¯B). By Remark 6.6, ¯βi belongs to a post-bar of ¯B, which has a nonempty intersection with Zer ¯p. All post-bars of B ∈ [B0] that have nonempty intersection with Zer p conjugate. They form the vertex of E(f ), bigger than [B0], which is smaller or equal (with the natural order in E(f )) than any element of E. Hence [B0] cannot be the infimum of E and we arrive again at a contradiction. We have shown that E has only one element. Denote it by [B0]. Hence for any ¯B◦. By Remark 6.6 we get 1. monomial substitution we have Zer ¯g ⊂ SB∈[B0] and the first part of 2. If for every ¯βi ∈ B Now we will find the form of GB(z), for any B ∈ [B0]. ¯βi is 0, then obviously GB(z) = azl. Otherwise by the leading coefficient lc ¯B Lemma 4.10 there exist c 6= 0 and a polynomial G1(z) coprime with zn(B)−cn(B) such that GB(z) = G1(z)(zn(B) − cn(B))l. Let p(y) be the minimal Weierstrass polynomial of λB(x) + cxh(B). Then PB(z) = const · (zn(B) − cn(B)). Proceeding as in the first part of the proof we define again the sets I, J of indexes. By the choice of p(y) the set J is nonempty. If the polynomial G1(z) has positive degree then the set I is nonempty and we arrive at a contradiction. Hence G1(z) is a constant which proves the second part of the theorem. Now we prove that if l = 1 in (14) then g(y) is quasi-ordinary. Let p(y) ∈ K[[x]][y] be the minimal polynomial of λB(x) if GB(z) = az or the minimal polynomial of λB(x) + cxh(B) if GB(z) = a(zn(B) − cn(B)). Then GB(z) is equal to PB(z) up to multiplication by a constant. By Lemma 5.2 for any B′ ∈ [B0] = [B] the characteristic polynomials GB′ (z) and PB′ (z) have the same form, in particular have the same number of roots and all their roots are simple. Take any monomial substitution and let ¯β′, ¯β′′ be different roots of ¯B◦ there exist B′, B′′ ∈ [B0] such that ¯β′ ∈ ¯B′ ¯g(y). Since Zer ¯g ⊂ SB∈[B0] and ¯β′′ ∈ ¯B′′. If B′ = B′′ then O( ¯β′, ¯β′′) = h( ¯B′) because ¯β′ and ¯β′′ have different leading coefficients with respect to ¯B′. If B′ 6= B′′ then O( ¯β′, ¯β′′) = O(λ ¯B′ , λ ¯B′′ ). In both cases the contact O( ¯β′, ¯β′′) depends only on B′ and B′′. The same argument applies to the roots of ¯p(y). As a consequence any bijection Φ : Zer ¯g → Zer ¯p such that Φ( ¯B′ ∩ Zer ¯g) = ¯B′ ∩ Zer ¯p for B′ ∈ [B0] preserves contacts. Since the discriminant of a monic polynomial is the product of differences of its roots, the discriminant of ¯g and the discriminant of ¯p have the same order. Then by Corollary 7.3 the Newton polytopes of the discriminants of g(y) and p(y) are equal and we conclude that g(y) is quasi-ordinary. For k-regular quasi-ordinary Weierstrass polynomials we can say more. Corollary 9.2 Let f (y) be a k-regular Kuo-Lu quasi-ordinary Weierstrass poly- nomial and let g(y) ∈ K[[x]][y] be a Weierstrass polynomial which is an irre- ducible factor of f (k)(y). Then there exists [B] ∈ E(f ) with B ∈ Tk(f ) such that: 21 1. If B′ ∈ T (f ) ∩ [B] then GB′ (z) and FB′ (z) do not have common roots. 2. If B′ ∈ Tk(f )\[B] then every root of GB′ (z) is a root of F ⊕ 3. If B′ ∈ T (f ) \ Tk(f ) then GB′ (z) is a non-zero constant polynomial. B′(z). Proof. Take B′ ∈ Tk(f ). Then by Lemma 4.6 and the definition of k-regularity GB′ (z) and F ⊖ B′(z) do not have common roots. Hence for B′ ∈ Tk(f ) it is enough to use Theorem 9.1. This proves 1. The second statement is the first item of Theorem 9.1. Now let B′ ∈ T (f )\Tk(f ). Consider the chain of bars B0 ⊥c B1 ⊥ ··· ⊥ Bs = B′ of T (f ) such that B0 ∈ Tk(f ) and Bi /∈ Tk(f ) for 1 ≤ i ≤ s. By the k-regularity of FB0 (z), we get GB0 (c) 6= 0. Since g is compatible with B0, after (4) of Lemma 4.7, we have g(λB′ (x) + zxh(B′)) = g(λB0 (x) + cxh(B0) + ··· ) = GB0 (c)xq(g,B0) + ··· , which shows that g is also compatible with B′ and its B′-characteristic polyno- mial GB′ (z) equals GB0 (c). 10 Eggers factorizations of higher derivatives Let f be a quasi-ordinary Weierstrass polynomial. In this section we propose a factorization of the normalized derivative f (k) into factors associated with points of Eggers tree E(f ). Definition 10.1 Let g, p ∈ K[[x]][y] be Weierstrass polynomials. The P-contact between g and p is contP(g, p) := 1 deg g deg p ∆(Res y(g, p)). The notion of P-contact has its counterpart in the theory of plane analytic curves: for y-regular plane branches it is related with the logarithmic distance studied by P loski in [P l], since in such case ∆(Res y(g, p)) equals the New- ton polygon of a monomial xm, where m is the intersection multiplicity of the branches g = 0 and p = 0. If g is compatible with a pseudo-ball B then we put contP(g, B) := 1 deg g ∆(xq(g,B)). Proposition 10.2 Let B be a quasi-ordinary pseudo-ball of finite height and let f be an irreducible quasi-ordinary Weierstrass polynomial compatible with B such that Zer f ∩ B 6= ∅ or equivalently such that FB(z) has positive degree. Then contP(f, B) does not depend on f . 22 Proof. Take any f1, f2 satisfying the assumptions of the proposition and let α1 ∈ Zer f1∩ B, α2 ∈ Zer f2∩ B. Choose a constant c ∈ K such that (Fi)B(c) 6= 0, for i = 1, 2 and let γ = λB + cxh(B). Then O(γ, α1) = O(γ, α2) = h(B) and for any ξ ∈ Zer f1 ∪ Zer f2 we have O(ξ, γ) ≤ h(B). Let G be a finite subgroup of L-automorphisms of M that acts transitively on the sets Zer f1 and Zer f2. By the orbit stabilizer theorem, for i ∈ {1, 2} we get 1 G Xσ∈G O(γ, σ(αi)) = 1 deg fi Xα∈Zer fi O(γ, α) = 1 deg fi q(fi, B). By STI we have O(γ, σ(α1)) = O(γ, σ(α2)) for all σ ∈ G. Thus 1 deg f1 q(f1, B) = 1 q(f2, B). deg f2 After Proposition 10.2 we define the self-contact of a pseudo-ball B of finite height as self-contact(B) := contP(f, B), for any f satisfying the assumptions of this proposition. By Lemma 5.2 conjugate pseudo-balls have the same self-contact, hence the self-contact of [B] is well-defined for any vertex [B] of E(f ), where B is of finite height. In the set of Newton polytopes we define the next partial order: ∆1 (cid:23) ∆2 if and only if ∆1 ⊆ ∆2. Observe that ∆(xq1) (cid:23) ∆(xq2) if and only if q1 ≥ q2. Now we show how the self-contacts of [B] ∈ E(f ) determine the P-contacts between irreducible factors of f . Proposition 10.3 Let f be a quasi-ordinary Weierstrass polynomial. Then the self contacts of vertices of finite height increase along the branches of E(f ). Moreover for any different irreducible factors f1, f2 of f contP(f1, f2) = max{self-contact([B])}, (15) where the maximum is taken over all [B] ∈ E(f ) such that Zer fi ∩ B 6= ∅ for i = 1, 2. Proof. Let B, B′ be pseudo-balls of T (f ) of finite height such that B′ ( B. Choose an irreducible factor fi of f such that Zer fi ∩ B′ 6= ∅. By Lemma 4.9 we get q(fi, B) < q(fi, B′), hence self-contact(B) ≺ self-contact(B′). Let [B] ∈ E(f ) be the maximum (with the order defined in E(f )) of the set of all vertices [B′] ∈ E(f ) such that Zer fi ∩ B′ 6= ∅ for i = 1, 2. The pseudo- ball B has the form γ + (γ − δ)K[[x1/N]], for some γ ∈ Zer f1 and δ ∈ Zer f2 with maximal possible contact. By the choice of γ and δ, we have O(γ, δ′) ≤ h(B) for all δ′ ∈ Zer f2 ∩ B, consequently (F2)B(lcBγ) 6= 0. Then f2(γ) = (F2)B(lcBγ)xq(f2,B) + ··· . 23 Applying the Galois action associated with the irreducible polynomial f2 we get ∆(f2(γ)) = ∆(f2(γ′)), for any γ, γ′ ∈ Zer f1. Hence by the definition of the self-contact and the identity ∆(Res y(f1, f2)) = Pγ∈Zer f1 ∆(f2(γ)) we have self-contact(B) = contP(f2, B) = 1 deg f2 ∆(xq(f2,B)) = = 1 deg f1 deg f2 1 deg f1 deg f2 deg f1 ∆(f2(γ)) ∆(Res y(f1, f2)) = contP(f1, f2). Theorem 10.4 Let f ∈ K[[x]][y] be a quasi-ordinary Weierstrass polynomial. Then f (k) = Y[B]∈E(f ) p[B], where p[B] are Weierstrass polynomials such that 1. The B-characteristic polynomial of p[B] equals F ⊖ B up to multiplication by constants and deg p[B] = N (B)tk(B). 2. For every irreducible factor g of p[B] and every irreducible factor fi of f we get (a) contP(g, B) = self-contact(B). (b) If contP(fi, B) ≺ self-contact(B) then contP(fi, g) = contP(fi, B). (c) If contP(fi, B) = self-contact(B) then contP(fi, g) (cid:23) contP(fi, B). 3. If f is k-regular then the inequalities (cid:23) in (c) become equalities. 4. For every irreducible factor g of p[B] there is an irreducible factor fi of f such that contP(fi, g) = contP(fi, B) = self-contact(B). Proof. We define p[B] as the product of all irreducible factors of f (k) having the same [B] in Theorem 9.1 (by convention the product of an empty family is 1). After some monomial substitution ¯p[B] has N (B)tk(B) roots and all of them are in SB′∈[B] ¯B′0. Consequently deg p[B] = N (B)tk(B). Now we will prove the second statement. Since p[B] has positive degree we may assume that B ∈ Tk(f ). Let fi be an irreducible factor of f . If contP(fi, B) ≺ self-contact(B) then by Proposition 10.2 (Fi)B(z) is a non-zero constant poly- nomial. Hence, for any ¯γ ∈ Zer ¯p[B] we have ord ¯fi(¯γ) = q( ¯fi, ¯B), which proves 2(b). Suppose now that contP(fi, B) = self-contact(B). For every root ¯γ ∈ Zer ¯p[B] we have ord ¯fi(¯γ) ≥ q( ¯fi, ¯B) with equality in the k-regular case. Hence if g is an irreducible factor of p[B] then ord Res y( ¯fi, ¯g) ≥ (deg g)· q( ¯fi, ¯B) 24 with equality in the k-regular case. This gives 2(c) and 3. If the polynomial FB(z) is as in (8) then it is k-regular. In this case for any irreducible factor fi of f , with (Fi)B(z) of positive degree, the polynomials (Fi)B(z) and GB(z) do not have common factors. If FB(z) is not as in (8), then by Lemma 4.10 there is an irreducible factor fi of f such that the polynomials (Fi)B(z) and GB(z) do not have common factors and (Fi)B(z) has positive degree. After any monomial substitution, we have ord ¯fi(¯γ) = q( ¯fi, ¯B), for every ¯γ ∈ Zer ¯g. This gives ord Res ( ¯fi, ¯g) = deg g · q( ¯fi, ¯B). Since the monomial substitution was arbitrary, the fourth statement of the theorem holds true in all cases. It rests to prove 2(a). Choose fi as in the proof of the fourth statament. Then ∆(g(α)) = ∆(xq(g,B)) for any α ∈ B ∩ Zer fi. Applying the same argument as in the end of the proof of Proposition 10.3, we get ∆(Res y(fi, g)) = deg fi∆(g(α)) = deg fi · ∆(xq(g,B)). After the fourth statement and the definition of the P-contact: self-contact(B) = contP(fi, g) = ∆(Res y(fi, g)) = 1 deg g ∆(xq(g,B)) = contP(g, B). 1 deg g deg fi Example 10.5 We consider the example in [GB-GP, Section 10]: f1,1f1,2f2,1f2,2, where fi,j = (y2−ix3 polynomials for i, j ∈ {1, 2}. The Kuo-Lu and the Eggers tree of f are f1,1 f1,2 f2,1 f2,2 let f = 2y are irreducible quasi-ordinary 1x4 1x2 2)2−jx5 [B2] [B3] [B1] The heights of the vertices of the Eggers tree are: h[B1] = (cid:0) 3 4 , 3 2(cid:1); the self-contacts are self-contact([B1]) = 1 h([B3]) = (cid:0) 7 self-contact([B2]) = self-contact([B3]) = 1 4 ∆(cid:0)x(13,10)(cid:1). For any 1 ≤ k ≤ 16, the degrees of polynomials p[Bi] are 2 , 1(cid:1), h([B2]) = 4 ∆(x(6,4)) and deg p[B1] deg p[B2] deg p[B3] f (1) f (2) f (3) f (k) 3 6 9 16-k 6 4 2 0 6 4 2 0 The characteristic polynomials are FB1 (z) = (z2− 1)4(z2− 2)4, FB2 (z) = (4z2− 1)(4z2 − 2) and FB3 (Z) = (8z2 − √2)(8z2 − 2√2). We can verify that these polynomials are k-regular for any k. 25 Theorem 10.4 allows us to compute the P-contact between the irreducible factors of f and the irreducible factors of its higher order polars. For any k and any ir- reducible factor g of p[B1], we have contP(fi,j, g) = self-contact([B1]). For any k and any irreducible factor g of p[B2], we have contP(f1,j, g) = self-contact([B2]) and contP(f2,j, g) = self-contact([B1]), for any j = 1, 2. We have the symmetric situation for the irreducible factors of p[B3]. Example 10.6 The second polar of the quasi-ordinary polynomial f from the Example 3.1 (see page 18 for its Eggers tree) has only one Eggers factor p[B] = y, with contP(f2, y) = ∆(x(5,2)) ≻ contP(f1, y) = ∆(x(3/2,1)) = self-contact(B), hence in item 2. (c) of Theorem 10.4 we have equality for f1 and strict inequality for f2. Now we study the examples of [Ca]: Example 10.7 ([Ca, Example 5.1]) Let f = y3 + x2y. The Eggers tree of f has only one vertex [B] of finite height, where B = xK[[x1/N]]. The B- characteristic polynomial of f is as in the previous example, so it is not 2- regular. We get f (2) = p[B] = y. If f1 = y − x, f2 = y + x and f3 = y then ∅ = contP(f3, y) (cid:23) contP(fi, y) = contP(fi, B) = ∆(x) = self-contact(B) for i = 1, 2. This illustrates the fourth statement of Theorem 10.4. Example 10.8 ([Ca, Example 5.2]) Let fa = y4 + ax2y2 + x2y + x10. We get fa = fa1fa2, where fa1 is irreducible and the contact of any two different roots of it is 2 3 , and fa2 = 0 is a smooth curve tangent to y = 0. The Eggers tree of fa is: fa1 fa2 [B] The characteristic polynomial FB(z) equals z4+z. Hence fa is not 2-regular. For any irreducible factor g of f (2) a we get contP(fa1, g) = ∆(x2/3) = self-contact(B). For fa2 the P-contact depends on a: contP(fa2, g) = (cid:26) ∆(x) ∆(x8) for a 6= 0 for a = 0. Irreducible case 10.1 Assume that f (y) ∈ K[[x]][y] is an irreducible quasi-ordinary Weierstrass poly- nomial of degree n > 1 and Zer f = {αi}n i=1. By [Li] the set {O(αi, αj) : i 6= j} := {h1, . . . , hs} is well-ordered, so we may assume that h1 ≤ h2 ≤ ··· ≤ hs. 26 These values are the finite heights of the bars of T (f ). The sequence h1, . . . , hs is called the sequence of characteristic exponents of f (y). Let Bi be any bar in T (f ) of height hi. By [GP, Remark 2.7] the degree n(Bi) of the field extension L(λBi (x)) ֒→ L(λBi (x), xhi) does not depend on the choice of Bi and will be denoted by ni. Put ei := ni+1 ··· ns for 0 ≤ i ≤ s (by convention the empty product is one). Observe that T (f ) has a special structure: all bars of the same height are conjugate and there are n1 ··· ni−1 conjugate bars of height hi (see [GB-Gw3, Theorem 6.2]). By (12) we get ∆((Res y(f (k), f − T )) = s Xi=1 n1 ··· ni−1tk(Bi)(cid:26)q(f, Bi) 1 (cid:27) , (16) where Bi is any ball of T (f ) of height hi and tk(Bi) =   (ni − 1)k ei−1 − k 0 for 1 ≤ k ≤ ei, for ei ≤ k ≤ ei−1, for ei−1 ≤ k < n. Let ik ∈ {1, . . . , s} be such that eik ≤ k < eik−1. Then tk(Bi) is positive if and only if 1 ≤ i ≤ ik. The Newton polytope of (16) is polygonal (see Corollary 8.2) and has ik edges of different inclinations. After Theorem 2.4 we decompose Res y(f (k), f − T ) = Qik i=1 Ri, where degT Ri = (n1 ··· ni−1)tk(Bi) and any Ri has an elementary Newton polytope of inclination q(f, Bi). Such a decomposition of the resultant can be also obtained from Eggers factor- ization of f (k). By Lemma 4.10 the Bi-characteristic polynomial of f has the form FBi (z) = constant(zni − cBi )ei , (17) for some cBi ∈ K \ {0}. The properties of such polynomials are described in the following lemma, which was proved in [GB-Gw4, Lemma 5.3] for complex polynomials but by Lefschetz Principle it holds true for polynomials over any algebraically closed field of characteristic zero. dk Lemma 10.9 Let K be an algebraically closed field of charactersitic zero. If F (z) = (zn − c)e ∈ K[z] with c 6= 0 then for 1 ≤ k < deg F (z) one has dzk F (z) = Cza(zn − c)bQd (1) 0 ≤ a < n and a + k ≡ 0 (mod n), (2) b = max{e − k, 0}, (3) d = min{e, k} − ⌈ k n⌉, where ⌈x⌉ denotes the smallest integer bigger than i=1(zn − ci), where C 6= 0 and or equal to x, (4) ci 6= cj for 1 ≤ i < j ≤ d and 0 6= ci 6= c for 1 ≤ i ≤ d. 27 Corollary 10.10 Every irreducible quasi-ordinary Weierstrass polynomial is Kuo-Lu k-regular for any positive integer k. Theorem 10.11 Let f (y) ∈ K[[x]][y] be an irreducible quasi-ordinary Weier- strass polynomial of degree n > 1 and characteristic exponents h1, . . . , hs. Then f (k)(y) = ik Yi=1 pi, (18) where 1. pi is a Weierstrass polynomial in K[[x]][y] of degree n1 ··· ni−1tk(Bi). 2. Any irreducible factor g of pi verifies contP(g, f ) = self-contact(Bi). 3. The Bi-characteristic polynomial of pi is (Pi)Bi = constF ⊖ Bi . Proof. The theorem follows from Corollary 10.10 and the first, second and third part of Theorem 10.4. Proposition 10.12 Let f (y) ∈ K[[x]][y] be an irreducible quasi-ordinary Weier- strass polynomial with characteristic exponents h1, . . . , hs. Let a, d be integers such that 0 ≤ a < ni, a + k ≡ 0 (mod ni) and d = min{ei, k} − ⌈ k ni ⌉. Then every pi of (18) admits a factorization of the form pi = pi0pi1 ··· pid, where 1. the corresponding Bi-characteristic polynomials are Pi0(z) = const · za, Pij (z) = const · (zni − cj) with cj 6= cl for 1 ≤ j < l ≤ d and cj 6= 0. 2. pi0 is a Weierstrass polynomial of degree a · n1 ··· ni−1 not necessarily quasi-ordinary. 3. Every pij for 1 ≤ j ≤ d is a quasi-ordinary irreducible Weierstrass poly- nomial of degree n1 ··· ni and characteristic exponents h1, . . . , hi. Proof. After (17) FBi (z) has the form a(zni − c)ei for some nonzero a and c. By the first part of Theorem 10.4 and Lemma 10.9 the polynomial Pi,Bi (z) = const·zaQd j=1(zni −cj). This polynomial is the product of the Bi-characteristic polynomials of the irreducible factors of pi. From the second part of Theo- rem 9.1, we know that pi has d irreducible factors {pij}d j=1 such that Pij (z) = const · (zni − cj). If pi has other irreducible factors, then pi0 is their product. It also follows from Theorem 9.1 that pij are quasi-ordinary for 1 ≤ j ≤ d. By a similar argument as in the first part of the proof of Theorem 10.4 we get deg pij = N (Bi) deg Pi,jBi (z). Since N (Bi) = n1 ··· ni−1, we obtain the state- ments about the degrees of pij. 28 Fix pij for j ∈ {1, . . . , d}. The pseudo-ball Bi has n1 ··· ni−1 conjugate pseudo- balls. Each of these pseudo-balls contains ni roots of pij . Since the roots of Pi,j(z) are simple, any two roots of pij belonging to the same pseudo-ball have different leading coefficients with respect to Bi, so their contact equals hi. Now, if we consider two roots of pij belonging to different conjugate pseudo-balls, then their contact depends only on these two pseudo-balls, hence it is equal to hl for some l ∈ {1, . . . , i − 1}. We conclude that the characteristic exponents of pij are h1, . . . , hi. In Proposition 10.12 the integer a can be 0, in such a case pi0 = 1. If a = 1 then pi0 is quasi-ordinary with characteristic exponents h1, . . . , hi−1. Moreover d can be zero and in such a case pi = pi0. 11 Eggers decomposition for power series In this section we deal with power series in variables x and y. A power series will be called quasi-ordinary if it is a product of a unity and a quasi-ordinary Weierstrass polynomial. We outline how to generalize the results of previous sections to quasi-ordinary power series. For that we need the next generalization of Lemma 4.6: Lemma 11.1 Let f = uf ∗ and ∂k ∂yk f = wg∗, where u, w ∈ K[[x, y]] are unities, f ∗, g∗ ∈ K[[x]][y] are Weierstrass polynomials and 1 ≤ k ≤ n = deg f ∗. Assume that f ∗ is compatible with a pseudo-ball B. Then g∗ is compatible with B and B(z) = (n−k)! G∗ Proof. Substituting x = 0 we get f (0, y) = u(0, 0)yn + ··· . Hence ∂kf (n−k)! u(0, 0)yn−k + ··· . On the other hand ∂kf ∂yk (0, y) = ∂yk (0, y) = w(0, y)g∗(0, y) which dk dzk F ∗ B(z). n! n! implies that w(0, 0) = n! (n − k)! u(0, 0). (19) By the assumption of compatibility of f ∗ we have f ∗(x, λB(x) + zxh(B)) = F ∗ B(z)xq(f ∗,B) + ··· . Hence f1(x, z) := x−q(f ∗,B)f (x, λB(x) + zxh(B)) is a fractional power series such that By the chain rule of differentiation f1(0, z) = u(0, 0)F ∗ B(z). ∂kf1 ∂zk (x, z) = ∂kf ∂yk (x, λB(x) + zxh(B)) · xkh(B)−q(f ∗,B). ∂zk (0, z) = u(0, 0) dk B(z). Thus dzk F ∗ Differentiating (20) yields ∂kf1 ∂kf1 ∂zk (x, z) = u(0, 0) dk dzk F ∗ B(z) + terms of positive degree in x. 29 (20) (21) (22) Comparing (21) and (22) we get ∂kf ∂yk (x, λB(x) + zxh(B)) = u(0, 0) dk dzk F ∗ B(z) · xq(f ∗,B)−kh(B) + ··· By the definition of g∗, the left hand side of the above equality can be written as which gives, after (19) w(0, 0) g∗(x, λB(x) + zxh(B)) + ··· , n! u(0, 0)g∗(x, λB(x) + zxh(B)) = u(0, 0) (n − k)! and finishes the proof. dk dzk F ∗ B(z) · xq(f ∗,B)−kh(B) + ··· Theorem 6.5, Corollary 6.7, Theorem 7.4, Theorem 9.1, Corollary 9.2, The- orem 10.4, Theorem 10.11 and Proposition 10.12, where f (k) stands for the Weierstrass polynomial of kth derivative, remain true for quasi-ordinary power series. For the proofs it is enough to replace the power series by their Weierstrass polynomials and use Lemma 11.1 instead of Lemma 4.6 when required. References [B-M] [Ca] [D] [Eg] [GB] Bierstone, E. and P. D. Milman. Semianalytic and subanalytic sets, Publications Math´ematiques. Institut de Hautes Etudes Sci- entifiques 67, no. 1 (1988), 5-42. Casas-Alvero, E. Higher Order Polar Germs. Journal of Algebra 240, 326-337 (2001). Delgado de la Mata, F. A factorization theorem for the polar of a curve with two branches, Compositio Math. 92 (1994) 327-375. Eggers, H. Polarinvarianten und die Topologie von Kurvensingu- laritaten. Bonner Mathematische Schriften 147, 1983. Garc´ıa Barroso, E.R. Sur les courbes polaires d'une courbe plane r´eduite. Proc.London Math. Soc. (3) 81 (2000) 1-28. doi: 10.1112/S0024611500012430. [GB-GP] E. Garc´ıa Barroso and P.D. Gonz´alez P´erez, Decomposition in bunches of the critical locus of a quasi-ordinary map. Compositio Math. 141 (2005) 461-486. [GB-Gw3] E. Garc´ıa Barroso and J. Gwo´zdziewicz, Quasi-Ordinary Sin- gularities: Tree Model, Discriminant, and Irreducibility. Int Math Res Notices. Volume 2015, Issue 14 (2015), 5783-5805, doi:10.1093/imrn/rnu106 30 [GB-Gw4] E. Garc´ıa Barroso and J. Gwo´zdziewicz, Decompositions of the higher order polars of plane branches, Forum Mathematicum (2017) 29 (2), 357-367. doi: 10.1515/forum-2016-0049 [GB-Gw-L] E. Garc´ıa Barroso, J. Gwo´zdziewicz and A. Lenarcik, Non- degeneracy of the discriminant. Acta Math. Hungar. Volume 147, Issue 1 (2015), 220-246. doi: 10.1007/s10474-015-0515-8. [GP] [He] [K-Lu] [K-Pa] [Li] [Me] P.D. Gonz´alez P´erez, The semigroup of a quasi-ordinary hyper- surface. Journal de l'Institut de Math´ematiques de Jussieu, 2(3), (2003) 383-399. B. Hejmej, A note about irreducibility of a resultant. Bull. Soc. Sci. Lettres L´od´z S´er. Rech. D´eform. 68 (2018), 27-32. T.C. Kuo and C. Lu, On analytic function germ of two complex variables, Topology, 16, 299 -- 310. T.C. Kuo and A. Parusi´nski, Newton-Puiseux roots of jacobian de- terminants. J. Algebraic Geometry 13 (2004) 579-601. Lipman, J. Topological invariants of quasi-ordinary singularities. Memoirs of the American Mathematical Society 74, no. 388 (1988): 1-107. M. Merle, Invariants polaires des courbes planes, Invent. Math. 41, (1977) 103-111. [Pa-R] Parusi´nski, A and Rond, G. The Abhyankar-Jung Theorem, Journal of Algebra 365 (2012) 29-41. [P l] [Po] [R-S] [W] P loski, A. Remarque sur la multiplicit´e d'intersection des branches planes. Bulletin of the Polish Academy of Sciences. Mathematics. Volume 33 (1985), 601-605. Popescu-Pampu, P. Arbres de contact des singularit´es quasi- ordinaires et graphes d'adjacence pour les 3-vari´et´es r´eelles. Th`ese, Univ. Paris 7, 2001. Available at https://tel.archives- ouvertes.fr/tel-00002800v1. Rond, G. and Schober, B. An irreducibility criterion for power se- ries. Proc. Amer. Math. Soc. 145 (2017), 4731-4739. Wall, C. T. C. Chains on the Eggers tree and polar curves. Proc. of the Int. Conf. on Algebraic Geometry and Singularities (Sevilla, 2001). Rev. Mat. Iberoamericana 19 (2003), no. 2, 745 -- 754. Evelia Rosa Garc´ıa Barroso Departamento de Matem´aticas, Estad´ıstica e I.O. Secci´on de Matem´aticas, Universidad de La Laguna 31 Apartado de Correos 456 38200 La Laguna, Tenerife, Espana e-mail: [email protected] Janusz Gwo´zdziewicz Institute of Mathematics Pedagogical University of Krak´ow Podchor¸azych 2 PL-30-084 Cracow, Poland e-mail: [email protected] 32
1609.03706	1	1609	2016-09-13T07:13:32	Surfaces in $\mathbb{P}^4$ lying on small degree hypersurfaces	[ "math.AG" ]	Since the work of Ellingsrud and Peskine at the end of 1980s, it has been known that, with the exception of a finite number of families, smooth compact complex surfaces in $\mathbb{P}^4$ with prescribed Chern classes must lie on hypersurfaces of degree $m\leq 5$. The study of surfaces lying on a small degree hypersurface in $\mathbb{P}^4$---small meaning $\leq5$---seems to be a way of obtaining empirical data leading to a better conceptual understanding of surfaces in $\mathbb{P}^4$. From this perspective, two main issues are considered in the paper: - an analogue of the Hartshorne-Lichtenbaum finiteness results for smooth surfaces of general type contained in a small degree hypersurface in $\mathbb{P}^4$, - a study of the irregularity of smooth surfaces contained in a small degree hypersurface in $\mathbb{P}^4$.	math.AG	math	Surfaces in P4 lying on small degree hypersurfaces Daniel Naie and Igor Reider Abstract Since the work of Ellingsrud and Peskine at the end of 1980s, it has been known that smooth compact complex surfaces in P4 with prescribed Chern classes, with the exception of a finite number of families, must lie on hypersurfaces of degree m ≤ 5. Hence the motivation for the present work: to study smooth surfaces contained in a hypersurface of degree m ≤ 5 (the meaning of 'small degree' in the title). There are two main issues considered in the paper: (I) an analogue of the Hartshorne-Lichtenbaum finiteness results for smooth surfaces of general type contained in a small degree hypersurface in P4, (II) a study of the irregularity of smooth surfaces contained in a small degree hypersur- face in P4. For (I) we show that for m ≤ 4, the number of families is controlled by a function depending on the ratio α = of the Chern invariants (K 2, χ) of surfaces. The same K 2 χ result holds for m = 5, with a possible exception of α = 6. For (II) we determine all irregular surfaces contained in a hypersurface of degree m ≤ 3. We do the same in case m = 4, under the additional assumption that a quartic hypersurface has only isolated double points. In general, we show that the Albanese dimension of surfaces contained in quartic hypersurfaces is at most 1. For m = 5, we show that minimal surfaces of Albanese dimension 2 have the irregu- larity at most 3 and describe the hypothetical surfaces with irregularity 3. Conceptually, the main idea underlying the above results as well as the whole approach of our paper can be termed as a representation of various geometric and cohomological entities attached to a surface in P4 in the category of coherent sheaves on that surface. Mathematics Subject Classification. 14J60, 14M07 Contents 1 Introduction 2 Notation and preliminaries 2.1 Bogomolov instability . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2.2 From hypersurfaces of small degree to extension classes . . . . . . . . . . . . 3 Numerical invariants for surfaces on degree 4 hypersurfaces 3.1 The inequalities (1.5) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3.2 The proof of Theorem 1.3 when mX = 4 . . . . . . . . . . . . . . . . . . . . . 4 Numerical invariants for surfaces with mX = 2, 3, 5 5 The irregularity of surfaces lying on a small degree hypersurface 6 Irregular surfaces on hypersurfaces of degree 2 2 8 8 10 15 15 24 25 31 34 1. Introduction 7 Irregular surfaces on hypersurfaces of degree 3 8 On the elliptic scroll of degree 5 and the Segre cubic 9 Irregular surfaces on hypersurfaces of degree 4 with non-degenerate iso- lated singularities 9.1 The study of H 1(OX (2KX + H)) . . . . . . . . . . . . . . . . . . . . . . . . . 9.2 The study of H 1(JZ (KX + 4H)) . . . . . . . . . . . . . . . . . . . . . . . . . 10 Albanese dimension 10.1 The Albanese dimension of X ⊂ P4 with mX = 4 . . . . . . . . . . . . . . . . 10.2 The surfaces X ⊂ P4 with mX = 5 and of Albanese dimension 2 . . . . . . . . A The projective bundle P(NX(−3H)) and the embedding X ⊂ P4 A.1 On the geometry of the scroll X . . . . . . . . . . . . . . . . . . . . . . . . . A.2 From the morphism ϕ in (A.3) to X ⊂ P(V ∗) . . . . . . . . . . . . . . . . . . A.3 The decomposition (A.14) and the configuration (104, 156) of Segre . . . . . . A.4 Toward a categorification of the configuration (104, 156) of Segre . . . . . . . 36 50 56 57 62 73 74 79 83 84 88 95 98 1. Introduction Smooth compact complex surfaces in P4 constitute an interesting and important part of the study of subvarieties of projective spaces. They are naturally situated at the cross-road of the theories of surfaces, vector bundles, and algebraic cycles. It is well-known that every smooth projective surface can be embedded into P5. To fit into P4, a surface must satisfy an obstruction, known as double point formula, d2 − 5d − 10(g − 1) + (c2 − K 2 X) = 0, (1.1) which ties together the degree d, the sectional genus g, and the basic topological invariants (the Chern numbers) K 2 and c2 of the surface. So one of the basic goals, which is still out of reach, is to find all the surfaces that can be embedded into P4. Another line of inquiry into the geometry of surfaces in P4 has been motivated by a conjecture of Hartshorne and Lichtenbaum stating that rational surfaces in P4 form a finite number of families. The work of Ellingsrud and Peskine, [15], solved a more general problem of finiteness of families of surfaces not of general type in P4. A key observation of [15] is that a surface X lying in a hypersurface of degree m in P4 has the holomorphic Euler characteristic χ(OX ) bounded from below by a certain cubic polynomial Pm(d) in the degree d of X. Since then, this result has been greatly improved and clarified. Notably, Decker and Schreyer's work, [9], gives a precise expression for Pm(d), Pm(d) = m(cid:18) d 3 m + m−3 2 (cid:19) − (m − 1)2 2m d(d − 3) −(cid:18)m − 1 4 (cid:19) + 1, where m = mX is the smallest degree of a hypersurface in P4 containing X, mX := min{k ∈ N h0(JX(k)) > 0}. 2 (1.2) (1.3) 1. Introduction Then, in [9], it is shown that provided d ≥ (m − 1)2 + 2. This result together with the bound on the sectional genus of [15] implies χ(OX ) ≥ Pm(d), (1.4) Theorem 1.1 (Ellingsrud and Peskine; Decker and Schreyer). Given integers χ and m ≥ 2, the surfaces in P4 having holomorphic Euler characteristic χ and lying on a hypersurface of degree m and not on one of a smaller degree form at most a finite number of families. Therefore, the "world" of surfaces in P4 is governed by the pairs of integers (χ, m) as in Theorem 1.1 and emphasis is placed on the understanding of surfaces that can lie on a hypersurface of given degree. From this point of view, the study of surfaces lying on a small degree hypersurface in P4 -- small meaning m ∈ {2, 3, 4, 5} -- would appear, and could perhaps be justified, as a way of obtaining empirical data leading to a better conceptual understanding of surfaces in P4. But in fact, taking the ideas of [15] and [9] a bit further, one can argue that these small degrees really matter precisely for their conceptual significance. To explain this point we fix a pair of integers (K 2, χ) and observe Proposition 1.2 ([26]). There exists a number d(K 2, χ), depending only on K 2 and χ, such that all surfaces in P4 with Chern numbers (K 2, χ) and degree d > d(K 2, χ) lie on a hypersurface of degree ≤ 5. This result tells us that, with a possible exception of finitely many families, the study of surfaces in P4 having prescribed Chern invariants comes down to understanding surfaces lying on hypersurfaces of degree m ∈ {2, 3, 4, 5}. Extrapolating further Proposition 1.2, we suggest Metha-principle. An understanding of a property P for surfaces in P4, with the ex- ception of a finite number of families, comes down to studying the property P for surfaces contained in hypersurfaces of degree m ≤ 5. After clarifying the origins and motivations for studying surfaces in P4 lying on hyper- surfaces of small degree let us give an overview of the main results of this paper. Main results of the paper. There are two main issues considered in this work: (I) an analogue of Hartshorne-Lichtenbaum finiteness results for smooth surfaces of general type contained in a small degree hypersurface in P4, (II) a study of the irregularity of smooth surfaces contained in a small degree hypersurface in P4. We approach (I) as students of the theory of surfaces of general type. To be more precise, let us recall that one of the main problems of that theory is the "geography" problem: characterize the pairs of integers (k, c) which are respectively K 2 X and χ(OX ) of some minimal surface X of general type. The integers χ(OX ) and K 2 X are often referred to as the Chern numbers of X -- terminology1 that we adopt in this paper -- and their ratio αX := K 2 X /χ(OX ) (called the slope in the sequel) provides many important dividing lines in this 2-dimensional 1In view of the Noether formula 12χ = K 2 numbers of X. 3 X + c2, either (K 2 X , c2) or (K 2 X , χ) will be referred to as Chern 1. Introduction world of Chern invariants of surfaces of general type. reference to X, if no ambiguity is likely. In this notation we often omit the It is well known that for a surfaces X of general type χ(OX ) > 0, so it makes sense to speak about the slope αX of the Chern numbers even when the surface X is not minimal. This is what we do for surfaces of general type in P4. It turns out that the number of families of such surfaces contained in a small degree hypersurface can be controlled only by the value of the slope α. With the notation (1.3) in mind, we can formulate a sample result concerning part (I). Theorem 1.3. For all surfaces X of general type in P4 with mX ≤ 4 the following assertions hold. 1) The slope α of the Chern numbers is smaller than 6. 2) For every rational number α < 6, there exists an integer d(α) such that every surface of slope α has the degree ≤ d(α). 3) For every rational number α < 6 and integer d ≤ d(α), there exists χ(α, d) such that every surface of slope α and degree d has the holomorphic Euler characteristic ≤ χ(α, d). A similar but somewhat more involved statement holds for surfaces X with mX = 5, see Proposition 4.5. It should be also pointed out that the expressions d(α) and χ(α, d) in the above theorem are effectively computable. For example, χ(α, d) is an explicit rational function of α and d. The above results, in essence, are obtained by a combination of two ingredients: the bound (1.2) of Decker and Schreyer, and inequalities of the form c2 − K 2 ≥ a H · KX + bd, (1.5) where OX(H) is the line bundle embedding X into P4, d is the degree of X, and a and b are positive rational numbers (depending on mX and explicitly determined in the main body of the paper, see Theorem 3.1, Theorem 4.1 and Theorem 4.2). So our contribution to the Hartshorne-Lichtenbaum problem for surfaces of general type are the inequalities (1.5) above. Of course, one most certainly wonders where those inequalities come from. This and other results of the paper will be explained shortly. For now, let us just say that the existence of these inequalities is an a priori consequence of our approach toward the study of surfaces in P4 contained in a hypersurface of a small degree. We now turn to the results concerning the issue (II), the irregularity of smooth surfaces contained in a small degree hypersurface in P4. Theorem 1.4. Let X ⊂ P4 be a smooth surface and mX be the smallest degree of a hypersurface containing X. 1) If mX = 2, then X is regular. 2) If mX = 3 and X is irregular, then X is an elliptic scroll of degree d = 5. Furthermore, a general cubic hypersurface containing X is a Segre cubic2 and the surface X must pass through the ten nodes of every Segre cubic containing it. 3) If mX = 4, then the Albanese dimension of X is at most 1. Furthermore, if X lies on a quartic hypersurface with only ordinary double points, then X is regular, with a possible exception of X being an elliptic conic bundle of degree d = −K 2 X = 8. 2Such a cubic has ten nodes, the maximal possible number of nodes for a cubic hypersurface in P4. 4 1. Introduction 4) If mX = 5, X is minimal, and its Albanese dimension is 2, then the irregularity of X is 2 or 3. The above statements illustrate several objectives of our inquiry about the irregularity of surfaces in P4: a) determine all irregular surfaces for a given value of mX, b) for a given mX, determine all possible values of the Albanese dimension of irregular surfaces, c) determine the upper bound on the irregularity for every value of the Albanese dimension that may occur. Statements 1) and 2) of Theorem 1.4 are examples of a), while 3) and 4) are partial answers to b) and c). The outline of our approach. In the rest of the introduction we discuss our approach to the study of smooth surfaces contained in a small degree hypersurface in P4. The main idea consists of interpreting the extrinsic datum of a small degree hypersurface as an intrinsic one. This is done first, by thinking of a hypersurface of the minimal degree m = mX containing a surface X ⊂ P4 as a nonzero global section of NX(−KX − (5 − m)H), the normal bundle NX = NX/P4 tensored with OX (−KX − (5 − m)H), and second, by attaching to that global section a cohomology class, call it ξ, in H 1(ΘX(−KX − (5 − m)H)), where ΘX is the holomorphic tangent bundle of X. The last step is achieved via the coboundary homomorphism H 0(NX(−KX − (5 − m)H)) −→ H 1(ΘX(−KX − (5 − m)H)) coming from the normal exact sequence of X ⊂ P4 tensored with OX (−KX −(5−m)H). Next, via the natural identification H 1(ΘX (−KX −(5−m)H)) ∼= Ext1(ΩX, OX (−KX −(5−m)H)), we interpret the cohomology class ξ as the corresponding extension 0 −→ OX (−KX − (5 − m)H) −→ Tξ −→ ΩX −→ 0. (1.6) The above extension sequence can be viewed as a reincarnation of a hypersurface of degree m containing X in the category of complexes of coherent sheaves on X. It is clear that it can be viewed as an independent entity and the understanding of its properties constitutes an important part of our approach. Namely, our strategy is to extract as much information as possible from (1.6) independently of the fact that X is embedded into P4 and then to use the acquired data to gain an additional insight into the embedding X ⊂ P4. As an example, let us take up the question of the (semi)stability of the sheaf Tξ. This is something a priori independent of the fact that X lies on a hypersurface of degree m in P4, and it gives, by the Bogomolov-Gieseker inequality, the constraint on the Chern invariants of Tξ, 3c2(Tξ) ≥ c2 1(Tξ). This together with the Chern invariants of Tξ, determined from the defining sequence (1.6), provide a prototypical example for the inequalities in (1.5), one of the ingredients to prove Theorem 1.3. Of course, there is no reason for Tξ to be semistable. However, even in the unstable case one has a sufficient control of the destabilizing filtration of Tξ to recapture the spirit of those inequalities. Though this constitutes a somewhat technical part of our considerations, the main point is quite transparent: the control of the properties of the destabilizing filtration of Tξ is enabled by 5 1. Introduction • the fact that Tξ is a part of the extension sequence (1.6) and in particular, that the cotangent sheaf ΩX is a quotient bundle of Tξ, and • the geometric origin of the extension sequence (1.6) which allows one to relate certain properties of the destabilizing filtration of Tξ to the embedding of X in P4. These remarks indicate that, though the semistable case provides the strongest form of the inequalities (1.5), it is the unstable case that is more interesting. Not only a numerical constraint in the form of (1.5) is obtained, but a certain amount of geometric data encoded in the destabilizing filtration of Tξ is gained. The results of Theorem 1.3 and Proposition 4.5 exploit only the numerical part of the study of (1.6). In this respect, the problem of the irregularity considered in (II) allows to reveal some more geometric aspects of our approach. Namely, the destabilizing subsheaf of Tξ, call it G, is defined as the saturation of the subsheaf of Tξ generated by its global sections. These sections are connected to the homomorphism H 0(Tξ) −→ H 0(ΩX) induced by the epimorphism in (1.6). It takes no effort to work out conditions for the above homomorphism to be surjective in the case m ≤ 4. When m = 5, we need the hypothesis of minimality for X for our approach to go through. This consideration, for example, immediately implies that for m ≤ 4 the rank of G is at most 2 and the proof of the statement 3) of Theorem 1.4 comes down to ruling out the possibility rank(G) = 2. The geometric destabilization of Tξ just outlined works well as long as the cotangent bundle ΩX is generically generated by its global sections, i.e., when the Albanese dimension of X is 2. In particular, it gives the result on the possible values of the irregularity in Theorem 1.4 as well as establishes a short list of hypothetical surfaces with irregularity 3, see Theorem 10.9. However, for surfaces of Albanese dimension 1, the above approach fails. This brings us to the second way of associating an extension sequence to a reduced irreducible hypersurface of degree m containing a surface X ⊂ P4. Here again we think of such a hypersurface as a nonzero global section of NX(−KX − (5 − m)H) and then take the Koszul sequence associated to it to obtain the extension 0 −→ OX(KX + (5 − m)H) −→ NX −→ JZ(mH) −→ 0. (1.7) The sequence, in general, is not exact, but let us ignore this for now and assume it is3. Its pertinence to the problem of irregularity of X comes from the identifications H 1(NX (KX )) ∼= H 1(N ∗ X)∗ ∼= H 0(ΩX)∗, where the first isomorphism is the Serre duality and the second is a general fact valid for any smooth subvariety of dimension at least 2 in a projective space. From this and (1.7) tensored with OX (KX ) it follows that the the irregularity of X is controlled by two cohomology groups H 1(OX (2KX + (5 − m)H) and H 1(JZ (KX + mH)). For m ≤ 5, the nonvanishing of the first group gives rather strong restrictions on X. The understanding of the nonvanishing of the second group depends largely on the knowledge of the subscheme Z which could be related to the singular locus of the hypersurface we started with. It is clear that this approach can only work when a good understanding of the singular locus of the hypersurface in question 3The interested reader will find a more detailed discussion of this approach in the introduction of §5. 6 1. Introduction is available. This is the case for m = 3, i.e., for cubic hypersurfaces, when everything can be analyzed completely leading to the elliptic scroll as the only irregular surface with m = 3; see Theorem 1.4, 2). Relation to other works. The subject of surfaces in P4 goes back to the classical algebraic geometry, see [30] and the references therein. Most of the results obtained in the subject in the last 30 years are based on the methods of syzygies and of construction of bundles on P4. Our approach of interpreting hypersurfaces containing a surface in P4 as certain extensions of sheaves on the surface itself seems to be a relative newcomer in the subject. It was initiated in our previous work [26] with an eye toward the problem of bounding the irregularity of surfaces in P4. Here we enlarge its scope by addressing the Hartshorne- Lichtenbaum problem as well as the problem of classification of irregular surfaces in P4. If for the first problem our contribution is largely tributary to the works [15] and [9], it is with the second problem that we have tried to be as self-contained in our treatment as possible. In particular, in deriving Theorem 1.4, 2), 3), we have avoided to call upon the results on classification of surfaces of small degree in P4. This is motivated (and hopefully justified) by our objective to show/explore various aspects of using the extension constructions to gain an insight into the geometry of surfaces. More importantly, we wanted to see (and show to the reader) how the extension construction 'pins down' (hypothetical) irregular surfaces in P4. The proofs of Theorem 1.4, 2) - 4) and Theorem 10.9 provide a substantial evidence that our approach is useful for classifying surfaces in P4. From the conceptual point of view, our approach could be termed as representing vari- ous geometric or cohomological entities by (short exact) complexes of coherent sheaves on a surface in question. The complexes or, better, distinguished triangles in the derived category (of coherent sheaves) we are using turn out to be unstable either for numerical (Bogomolov instability) or for more subtle geometrical reasons. This instability gives rise to a new dis- tinguished triangle which carries more geometry than the initial one. It is in relating the two triangles that one is able to obtain new geometric insights. This is very much in line with more recent developments of methods of derived categories in algebraic geometry such as Bridgeland's stability conditions, see e.g., [8], [2]. Organization of the paper. In §2, preliminary material is gathered. We start by recalling some facts about the Bogomolov instability and then go on explaining how to relate hypersurfaces in P4 containing a surface to the extension sequences of sheaves on that surface. One of the technical results used throughout the paper is Lemma 2.3. The sections §§3 - 4 are devoted to the Hartshorne-Lichtenbaum problem for surfaces of general type in P4. The main results here are Proposition 3.6 and Proposition 4.5, see also Theorem 1.3 in the introduction. The rest of the paper is devoted to the problem of the irregularity of surfaces in P4. In §5 we explain how the main ideas of our approach are connected with this problem and in §6 we illustrate some of these ideas in the case of surfaces lying on a quadric hypersurface, see Theorem 6.1. In §7 the surfaces on a cubic hypersurface are treated. Theorem 7.1 is the main result of this section. §8 is an interlude about elliptic scrolls in P4. The subject is well-known, see [18, 4, 5], but we approach it from the point of view of the (twisted) conormal bundle of a scroll. The main result is Theorem 8.1. In §9 we treat the case of surfaces on a quartic hypersurfaces with ordinary double points. The main result here is Theorem 9.1. 7 2. Notation and preliminaries In §10 the Albanese dimension of surfaces contained in a quartic (rep. quintic) hypersur- face is considered: Theorem 10.1 and Theorem 10.9 are the main results of this section. The Appendix of the paper returns again to the case of an elliptic scroll. The main objective here is to show how the classical configuration of 10 nodes of a Segre cubic hyper- surface in P4 is related to the geometry of an elliptic scroll contained in it. In particular, we show how the extension construction lifts the famous configuration (104, 156) of Segre to the category of (short exact complexes of) coherent sheaves on a scroll and we suggest that this should lead to a categorification of (104, 156) configuration of Segre. 2. Notation and preliminaries 2.1. Bogomolov instability Let X be a smooth complex projective surface4. We denote by NS(X) the N´eron-Severi group of X. Its rank ρ is called the Picard number of X and the intersection product defines an integral quadratic form on NS(X), whose real extension to N (X) := NSR(X) is of type (1, ρ − 1), by the Hodge Index Theorem. The positive cone of X is the open cone N +(X) = {D ∈ N (X) D2 > 0, H·D > 0, for some (hence any) ample divisor class H on X}. Note that N +(X) contains the ample cone and is contained in the cone of effective divisors. Let F be a coherent sheaf on X of rank r = rF . The discriminant of F is the expression ∆(F) = 2r c2(F) − (r − 1) c2 1(F). A more geometric way to think about ∆(F) for sheaves of rank r ≥ 1 is to observe that ∆(F) 2r = c2(F) − r − 1 2r c2 1(F) = c2(cid:16)F ⊗ OX(cid:0) − 1 r c1(F)(cid:1)(cid:17). The next result is due to Bogomolov and it is used constantly in the sequel. Bogomolov Theorem. Let F be a torsion free coherent sheaf on a surface X. If ∆(F) < 0, then there exists a maximal non-trivial saturated subsheaf F ′ such that • ∆(F ′) ≥ 0, • c1(F ′) rF ′ − c1(F) rF ∈ N +(X) and(cid:18)c1(F ′) − rF ′ rF c1(F)(cid:19)2 ≥ − ∆(F) 2rF . In particular, if F is D-semistable5 with respect to an ample divisor D, then ∆(F) ≥ 0. A torsion free sheaf is called Bogomolov unstable if ∆(F) < 0 and Bogomolov semistable if ∆(F) ≥ 0. The theorem asserts that a torsion free Bogomolov unstable sheaf contains a maximal Bogomolov semistable subsheaf which destabilizes it with respect to every polar- ization. Such a subsheaf is called a maximal Bogomolov destabilizing subsheaf of the given sheaf. 4These hypotheses are assumed throughout the paper whenerver we speak about surfaces. 5The sheaf F is D-semistable if c1(F ′) rF ′ · D ≤ c1(F) rF · D, for any nonzero subsheaf F ′ ⊂ F. 8 2. Notation and preliminaries Lemma 2.1. Let F be a locally free sheaf on the surface X. There exists a unique Bogo- molov filtration of F, 0 = F0 ⊂ F1 ⊂ · · · ⊂ Fm = F such that for each 1 ≤ i ≤ m, Fi/Fi−1 is the maximal Bogomolov destabilizing subsheaf of Fj/Fi−1 for every j > i. Proof. One can argue by induction on the rank r = rank(F). For r = 1 the statement is obvious, since by definition locally free sheaves of rank 1 are Bogomolov semistable. So we assume r ≥ 2 and suppose that the theorem holds for all locally free sheaves of inferior rank. Furthermore, we can assume that F is Bogomolov unstable (since otherwise there is nothing to prove). Let F1 be a maximal Bogomolov destabilizing subsheaf of F. By assumption, F1 6= F. Since F1 is saturated, the quotient F/F1 is torsion free, and therefore F1 is reflexive (cf. [20, Proposition 5.22]), hence locally free, since X is a surface. Now, if the quotient F/F1 is Bogomolov stable, the filtration reduces to 0 = F0 ⊂ F1 ⊂ F2 = F and we are done. If not, the quotient F/F1 has the rank strictly smaller than r and hence the theorem holds for (the reflexive hull or the double dual of) F/F1. Hence (F/F1)∗∗ admits a unique Bogomolov filtration. Lifting this filtration to F gives the desired filtration of F. It is enough to describe the procedure for the lifting of the maximal Bogomolov destabilizing subsheaf, call it G′, of F/F1 and then apply it inductively for other pieces of the Bogomolov filtration of (F/F1)∗∗. Let G′′ be the quotient of the inclusion G′ ⊂ F/F1. We have the diagram 0 0 F1 F1 0 0 0 F2 F G′′ 0 0 G′ F/F1 G′′ 0 where F2 is the kernel of the epimorphism F → G′′. As before, in this short exact sequence F is locally free and G′′ is torsion free, hence F2 is locally free. Clearly F1 ⊂ F2 and G′ ∼= F2/F1. We must show that F2 is Bogomolov unstable and that F1 is a maximal Bogomolov destabilizing subsheaf of F2. 9 2. Notation and preliminaries − r r2 Set r = rank(F), rj = rank(Fj), and rG ′ = rank(G′). Since c1(F2) c1(F2) c1(F) c1(F) =(cid:18) c1(F1) r1 − − − r1 r1 rG ′ rG ′ c1(F) c1(F) r (cid:19) − r (cid:19) − r − r1(cid:19)(cid:18)c1(F1) =(cid:18) c1(F1) =(cid:18) c1(F1) =(cid:18)1 − r(r − r1)(cid:18) c1(F1) rG ′ r2 r(r − r2) r1 − = r r1 c1(F) r1 r (cid:19) −(cid:18) c1(F1) r2(cid:18) c1(F1) r2(cid:18) c1(F1) r1 − − r1 c1(F) − − r2 (cid:19) c1(G′) c1(F/F1) rG ′ (cid:19) r − r1 (cid:19) + r2(cid:18) c1(G′) rG ′ r (cid:19) + r2(cid:18) c1(G′) rG ′ rG ′ r (cid:19) + rG ′ c1(F/F1) r − r1 (cid:19), − c1(F/F1) r − r1 (cid:19) rG ′ − rG ′ r2(cid:18)c1(G′) r − r1 (cid:19) c1(F/F1) − we see that c1(F2)/r2 − c1(F)/r ∈ N +(X). Hence ∆(F2) < 0, since otherwise F2 would be a Bogomolov destabilizing subsheaf of F and this contradicts the maximality of F1. Thus we now have constructed a Bogomolov unstable subsheaf F2 of F and we claim that F1 is its maximal Bogomolov destabilizing subsheaf. Indeed, if F1 is not a maximal Bogomolov destabilizing subsheaf of F2, then there exists a Bogomolov semistable (locally free) subsheaf F ′ such that F1 ⊂ F ′ ⊂ F2 and such that c1(F ′)/r′ − c1(F2)/r2 ∈ N +(X). But then, by the previous argument, c1(F ′) r′ − c1(F) r =(cid:18) c1(F ′) r′ − c1(F2) r2 (cid:19) +(cid:18) c1(F2) r2 − c1(F) r (cid:19) ∈ N +(X), contradicting the maximality of F1. (cid:3) 2.2. From hypersurfaces of small degree to extension classes Let X ⊂ P4 be a smooth surface. In what follows we denote by NX = NX/P4 the normal bundle of X in P4 and by H a hyperplane section of X. The normal bundle NX is of rank 2 on X with determinant det(NX) = 2NX = OX (KX + 5H). The conormal bundle satisfies JX/J 2 X = N ∗ X ∼= det(N ∗) ⊗ NX = NX(−KX − 5H), (2.1) where JX is the ideal sheaf of X in P4 and the second identification comes from the fact that the rank of N ∗ X is 2. We now assume that X lies on a hypersurface of degree m ≤ 4 and not on one of a smaller X(mH)) does not degree, i.e., m = mX . From the first equality in (2.1), it follows that H 0(N ∗ vanish. Hence, by two other identifications in (2.1), we obtain H 0(NX(−KX − (5 − m)H)) = H 0(N ∗ X(mH)) 6= 0. (2.2) Set t = 5 − m and observe that t ∈ {1, 2, 3}. We wish to interpret nonzero sections of NX(−KX − tH) cohomologically. For this, consider the normal sequence of X in P4 tensored with OX(−KX − tH), 0 −→ ΘX(−KX − tH) −→ ΘP4X(−KX − tH) −→ NX(−KX − tH) −→ 0. This implies that H 0(NX(−KX − tH)) fits into the following exact sequence of cohomology groups H 0(ΘP4X(−KX − tH)) −→ H 0(NX(−KX − tH)) δX−→ H 1(ΘX(−KX − tH)). (2.3) The following result improves a part of Lemma 5.3 in [26]. 10 2. Notation and preliminaries Lemma 2.2. Let s be a nonzero global section of NX(−KX −tH), 1 ≤ t ≤ 3, corresponding to a 3-fold of degree m = 5 − t containing X. If the Kodaira dimension of X is non-negative, then the cohomology class δX (s) 6= 0. Proof. We only prove the case t = 1, i.e., X is contained in a quartic hypersurface, since the other cases are much easier. Assume δX(s) = 0 in H 1(ΘX (−KX − H)), then s is the image of a nonzero global section es of ΘP4X (−KX − H)). From the Euler sequence of ΘP4, we deduce that either H 0(OX (−KX )) 6= 0 or ker(H 1(−KX − H) → H 0(H)∗ ⊗ H 1(−KX)) 6= 0. In both cases we see that H 0(OC (−KX)) 6= 0 for every C ∈ H. Hence H · (−KX ) ≥ 0. If H · (−KX) > 0, then h0(OX (mKX )) = 0 for every positive integer m, hence X rational or irrationally ruled. If H · (−KX) = 0, then OC (−KX ) = OC, for every C ∈ H This tells us that H 0(OX (−KX )) 6= 0 and hence KX = 0. Therefore, X is minimal and it is either a K3 or an abelian surface. If X is a minimal K3 surface, then the exact sequence 0 −→ H 0(OX ) −→ H 0(OX (H)) −→ H 0(OC (H)) −→ 0, with C a general curve in H, implies that g(C) = h0(OC (H)) = h0(OX (H)) − 1 = 4. Hence d = H 2 = 2g(C) − 2 = 6. But then the sequence 0 −→ JX(2) −→ OP4(2) −→ OX (2H) −→ 0 implies h0(JX (2) ≥ h0(OP4(2)) − h0(OX (2H)) =(cid:18)6 2(cid:19) − (2H)2 2 − 2 = 1, i.e., there is a quadric passing through X and this is contrary to our assumption. Thus we are left with the second case: X a minimal abelian surface. From the double point formula (1.1) it follows that d = 10. We show that a minimal abelian surface of degree 10 can not lie on a hypersurface of degree 4. This follows from the following observation. Claim. A quartic hypersurface Q ∈ P4 containing a minimal abelian surface X must be a cone over a quartic surface S ⊂ P3 with at most isolated singularities. Let us assume the claim and derive a contradiction. We view P4 as P(H 0(OX (H))∗) and denote by [v], v ∈ H 0(OX (H))∗, the vertex of Q. Under the projection from [v] the surface X becomes a finite covering of a quartic surface S lying in some P3, complementary to [v]. Let m be the degree of this covering. It is related to the intersection of X with a ruling l of the cone Q as follows. But for a general plane Λ passing through [v], if [v] /∈ X if [v] ∈ X. m − 1 (X · l)Q =(m 10 = d = (X · Λ)P4 =(4m if [v] /∈ X 4(m − 1) + 1 if [v] ∈ X and neither case is possible. We now turn to the proof of the claim. Let us recall the situation: s is a nonzero global section in NX(−H) corresponding to a quartic hypersurface Q containing X, and it is the image of a global section es of ΘP4X (−H). The argument is divided into two steps. 11 2. Notation and preliminaries Step 1. We claim that the scheme of zeros Zs = {s = 0} of the global section s is 0-dimensional. Indeed, let us assume that this is not the case and let Γ be a reduced, irreducible curve in Zs. This means that s = γs′, where γ ∈ H 0(OX (Γ)) is a global section defining Γ and s′ ∈ H 0(NX (−H − Γ)). From the commutative diagram H 0(ΘP4X (−H − Γ)) H 0(NX(−H − Γ)) H 1(ΘX(−H − Γ)) γ· γ· H 0(ΘP4X (−H)) H 0(NX(−H)) it follows that either H 0(ΘP4X (−H −Γ)) 6= 0 or H 1(ΘX(−H −Γ)) 6= 0. The latter possibility leads to H 1(OX (−H − Γ)) 6= 0, since H 1(ΘX(−H − Γ)) ∼= ⊕H 1(OX (−H − Γ)). But then, H 0(OΓ(−H)) ∼= H 1(OX (−H − Γ)) 6= 0 which is impossible. The former possibility leads, using the Euler sequence for ΘP4, to ker(cid:0)H 1(OX (−H − Γ)) → H 0(OX (H))∗ ⊗ H 1(OX (−Γ)(cid:1) 6= 0, which means that H 0(OC (−Γ)) 6= 0 for every C ∈ H, which is not possible either. Step 2. Q is a cone over a quartic surface. Indeed, on the one hand we consider the diagram H 0(OX (H))∗ H 0(ΘP4X (−H)) H 0(NX(−H)) (2.4) where the horizontal arrow comes from the normal sequence of X in P4 and the vertical one is part of the Euler sequence of ΘP4 (tensored with OP4(−1)) restricted to X. Both maps are isomorphisms. This implies that the section s is the image of a unique element v ∈ H 0(OX (H))∗. On the other hand, we have the Koszul sequence associated to s, 0 −→ OX s−→ NX(−H) s∧−→ JZs(3H) −→ 0. which is exact by Step 1. Combining this with the slanted arrow in (2.4) gives the following diagram H 0(OX (H))∗ ⊗ OX e 0 OX s NX(−H) s∧ JZs(3H) 0 Furthermore, the homomorphism ∂ : H 0(OX (H))∗ −→ H 0(JZs(3H)) (2.5) (2.6) induced from the above diagram on the level of global sections is the partial differentiation. Namely, let F be a homogeneous polynomial defining Q, then the homomorphism ∂ is given by H 0(OX (H))∗ ∋ u 7→ ∂(u) = ∂uF X ∈ H 0(JZs(3H)). 12 2. Notation and preliminaries By construction ∂(v) factors via H 0(NX(−H)), i.e., we have ∂vF X = ∂(v) = s ∧ (e(v)) = s ∧ s = 0. This means that the homogeneous polynomial of degree 3, ∂vF ∈ Sym3 H 0(OX (H)), vanishes on X. But since the surface X is not contained in any hypersurface of degree less than 4, we conclude that ∂vF = 0. Equivalently, F ∈ Sym4(ker(v)), i.e., the 3-fold of degree four Q is the cone in P(H 0(OX (H))∗) with vertex [v] and base the quartic surface defined by F in P3 = P(ker(v)∗). The last assertion, stating that the quartic surface S defined by F in P(ker(v)∗) has at most isolated singularities, follows from the observation that a curve, call it Γ, in the singular locus of S produces a surface Σ in Q -- the cone over Γ with vertex at [v] -- and this surface lies in the singular locus of the quartic Q. But then the surfaces X intersects Σ along a curve which is part of the zero-locus of the section s. This contradicts Step 1. (cid:3) From now on we assume that the cohomology class δX (s) ∈ H 1(ΘX(−KX − tH)) in Lemma 2.2 is nonzero. The identification H 1(ΘX (−KX − tH)) ∼= Ext1(ΩX, OX (−KX − tH)) allows us to interpret a cohomology class on the left as an extension sequence of sheaves on X. The following result constitutes one of the main technical ingredients of this study. Lemma 2.3. Let X be a smooth projective surface and let M be a divisor on X. Let ξ ∈ H 1(ΘX (−KX − M )) be a nonzero cohomology class and let 0 −→ OX (−KX − M ) −→ Tξ −→ ΩX −→ 0. (2.7) be the corresponding extension sequence. Assume that Tξ contains a subsheaf F of rank 2 such that the induced morphism F → ΩX is generically an isomorphism. Then the following holds. 1) The canonical divisor of X decomposes as KX = L + E, where L = c1(F) and E is the support of the cokernel coker(F → ΩX), a nonzero effective divisor on X. 2) If e is a section of OX (E) defining E, i.e., E = (e = 0), then the cohomology class ξ is annihilated by e, i.e., e ξ = 0 in H 1(ΘX (E − KX − M )). 3) If, in addition, X ⊂ P4 lies on a 3-fold of degree m ≤ 5 and ξ = δX (s), where δX is the coboundary map in (2.3), then H 0(ΘP4X (E − KX − (5 − m)H)) = H 0(ΘP4X (−L − (5 − m)H)) 6= 0. In particular, H · L ≤ (m − 4)H 2. Proof. Set ϕξ : F −→ ΩX to be the morphism defined by the composition of the inclusion F֒→Tξ together with the epi- morphism of the extension sequence (2.7). By assumption ϕξ is generically an isomorphism. 13 2. Notation and preliminaries This implies that the support of coker(ϕξ) is a nonzero effective divisor, since otherwise ϕξ is an isomorphism and the exact sequence (2.7) splits or, equivalently, ξ = 0. Writing out the exact sequence 0 −→ F ϕξ−→ ΩX −→ coker(ϕξ) −→ 0 we obtain the decomposition of the canonical divisor asserted in 1) of the lemma. To prove 2) we consider the diagram 0 OX (−KX − M ) 0 F Tξ ϕξ e JZ (−L − M ) 0 ΩX 0 (2.8) where the slanted arrow in the lower part of the diagram is the morphism given by multipli- cation with the section e. Dualizing and tensoring the diagram with OX(−L − M ) we arrive at OX e (2.9) 0 ΘX(−L − M ) T ∗ ξ (−L − M ) OX (KX − L) 0 Since the coboundary map H 0(OX (KX − L)) = H 0(OX (E)) −→ H 1(ΘX (−L − M )) in long exact sequence of cohomology groups of the horizontal sequence in (2.9) is given by the cup-product with the class ξ and since the section e is in its kernel, we deduce e ξ = 0 in H 1(ΘX(−L − M )). For 3), we use the fact that ξ = δX (s), where δX is as in (2.3). Set ∆ = KX + (5 − m)H and consider the commutative diagram H 0(NX (−∆)) e H 0(ΘP4X(E − ∆)) H 0(NX (E − ∆)) From this and 2) of the lemma, we obtain δX δ′ X H 1(ΘX(−∆)) e H 1(ΘX(E − ∆)) δ′ X (es) = e δX (s) = e ξ = 0. 14 3. Numerical invariants for surfaces on degree 4 hypersurfaces Hence the global section es ∈ H 0(NX(E − KX − (5 − m)H)), being obviously nonzero, comes from a nonzero section in H 0(ΘP4X(E − KX − (5 − m)H)). This proves the first assertion of 3). To see the inequality H · L ≤ (m − 4)H 2, we restrict the Euler sequence for P4 to X and tensor it with OX (E − KX − (5 − m)H) = OX (−L − (5 − m)H) to arrive at 0 → OX (−L−(5−m)H) → H 0(OX (H))∗⊗OX (−L−(4−m)H) → ΘP4X(−L−(5−m)H) → 0. Since H 0(ΘP4X(E − KX − (5 − m)H)) 6= 0, by the first part of 3), the above sequence implies that either H 0(OX (−L − (4 − m)H)) 6= 0, or ker(cid:16)H 1(OX (−L − (5 − m)H)) −→ H 0(OX (H))∗ ⊗ H 1(OX (−L − (4 − m)H))(cid:17) 6= 0. The first possibility immediately gives the assertion H · L ≤ (m − 4)H 2. The second one h→ H 1(OX (−L − (4 − m)H)) given implies that the homomorphism H 1(OX (−L − (5 − m)H)) by the multiplication by any global section h ∈ H 0(OX (H)) has a nonzero kernel. Since that kernel comes from H 0(OCh (−L − (4 − m)H)), where Ch = (h = 0), we deduce that H 0(OC (−L − (4 − m)H)) 6= 0, for any divisor C in the linear system H. This implies H · (−L − (4 − m)H) ≥ 0 and hence the assertion H · L ≤ (m − 4)H 2. (cid:3) (I) Hartshorne-Lichtenbaum for surfaces of general type 3. Numerical invariants for surfaces on degree 4 hypersurfaces In this section we prove the inequalities (1.5) of the form c2 − K 2 X ≥ a H · KX + bd, stated in the introduction. Let X be a smooth surface in P4 lying on a 3-fold V = V4 of degree four and not on any of a smaller degree. Unless stated otherwise, we assume that the Kodaira dimension of X is non-negative. This assumption, according to Lemma 2.2, gives a nonzero cohomology class δX (s) ∈ H 1(ΘX (−KX − H)) (see Lemma 2.2 for notation) which we denote by ξ. As we already explained, this cohomology class is used to build the extension (2.7) and the focus of study becomes the vector bundle Tξ sitting in the middle of that sequence. The inequalities we are after are a consequence of Bogomolov semistability or instability of Tξ. 3.1. The inequalities (1.5) Theorem 3.1. Let Tξ be the sheaf in the middle of the extension sequence (2.7) associated to ξ = δX (s). 1) If Tξ is Bogomolov semistable, then c2 − K 2 2) If Tξ is Bogomolov unstable, then c2 − K 2 2 H · KX + 1 4 d(cid:17). X ≥ H · KX + 1 3 d. 4 H · KX , 1 X ≥ min(cid:16) 3 Proof. From the exact sequence (2.7) it follows that Tξ is a locally free sheaf of rank 3 X − H · KX . Therefore the with the Chern invariants c1(Tξ) = −H and c2(Tξ) = c2 − K 2 Bogomolov semistability condition for Tξ reads as follows 6c2(Tξ) − 2c2 1(Tξ) = 6(c2 − K 2 X ) − 6H · KX − 2H 2 ≥ 0 15 3. Numerical invariants for surfaces on degree 4 hypersurfaces and this is equivalent to the inequality in 1) of the theorem. We now turn to the case when Tξ is Bogomolov unstable. To analyse the situation we use the Bogomolov filtration of Tξ, see Lemma 2.1. In particular, according to the shape of that filtration, we obtain the following inequalities6: H · KX , 2 H · KX + 1 1 3 4 H · KX, 4 d, if if if 0 ⊂ F1 ⊂ F2 = Tξ and 0 ⊂ F1 ⊂ F2 = Tξ and 0 ⊂ F1 ⊂ F2 ⊂ F3 = Tξ. rank(F1) = 2 rank(F1) = 1 (3.1) (cid:3) c2 − K 2 X ≥ This implies 2) of the theorem. Before we proceed with the proof of (3.1), we would like to provide the reader with the conducting line of the proofs of the lemmas below. The basic idea is to use the Bogomolov filtration of Tξ for writing down a "good" lower bound for the second Chern number of Tξ. "Good" here means that a sought after estimate should imply a lower bound for c2 − K 2 X as a positive function of d or/and H · KX . This is possible in view of the following special features of Tξ: • The subsheaves of rank 2 involved in the filtration of Tξ satisfy the hypotheses of the technical Lemma 2.3. • The subsheaves of rank 1 involved in the filtration of Tξ having positive degree (with respect to some polarization of X) inject into ΩX and hence must be of Iitaka dimension at most one (Bogomolov lemma); furthermore, the generic semi-positivity of ΩX insures that the quotient sheaf must be of non-negative degree. With these remarks in mind, we now consider all possible filtrations of the Bogomolov unstable vector bundle Tξ. Lemma 3.2. If the Bogomolov filtration of Tξ is 0 = F0 ⊂ F1 ⊂ F2 = Tξ with F1 of rank 2, then it gives rise to a divisor B1 in the positive cone of N (X) and to an effective nonzero divisor E such that the following hold: 3 (B1 − 2H) and B1 · H ≤ 2d, 1) c1(F1) = 1 2) KX = 1 3) c2 − K 2 3 (B1 − 2H) + E, X ≥ H · KX . Proof. The bundle Tξ is the middle term of two exact sequences, as in diagram (2.8), where the maximal Bogomolov destabilizing subsheaf F1 takes the place of F. We set L1 = c1(F1) and L2 = c1(Tξ/F1). Using the equality −H = c1(Tξ) = L1 + L2, the Bogomolov destabilizing condition for F1 tells us that the Q-divisor c1(F) 2 − c1(Tξ) 3 = L1 2 − −H 3 = 1 6 (L1 − 2L2), 6The inequality of the last line in (3.1) is strict, see Lemma 3.4 for details. 16 3. Numerical invariants for surfaces on degree 4 hypersurfaces lies in the positive cone N +(X) of N (X). Thus the divisor B1 = L1 − 2L2 lies in the positive cone N +(X) and we write Li, i = 1, 2, as a linear combination of H and B1 as follows: L1 = − L2 = − 2 3 1 3 H + H − 1 3 1 3 B1 B1. (3.2) Recall that c2(Tξ) = c2 − K 2 X − H · KX. Computing that Chern class using the vertical sequence in (2.8) with F = F1 and the quotient sheaf Tξ/F1 = IZ(L2), we obtain c2 − K 2 X − H · KX = c2(F1) + L1 · L2 + deg(Z) ≥ 1 4 L2 1 + L1 · L2 = 1 12 (4H 2 − B2 1). (3.3) where the inequality uses c2(F1) ≥ 1 equality comes from substituting the expressions from (3.2). 4 L2 1, the Bogomolov semistability of F1, and the last Next we claim that the slanted arrow ϕξ : F1 → ΩX in the diagram (2.8) is generically an isomorphism. Indeed, if ϕξ drops its rank everywhere, then we obtain the commutative diagram 0 0 OX (−KX − H) OX (−KX − H) 0 F1 Tξ im(ϕξ) ΩX 0 0 where the sheaf im(ϕξ) is the image of ϕξ. It is a torsion free subsheaf of rank 1 of ΩX with the first Chern class c1(im(ϕξ)) = c1(F1) + (KX + H) = − 2 3 H + 1 3 B1 + KX + H = KX + 1 3 H + 1 3 B1. But this means that ΩX contains a rank 1 subsheaf of Iitaka dimension 2 which is impossible in view of Bogomolov Lemma. Once we know that ϕξ is generically of maximal rank, Lemma 2.3 can be applied. In particular, we obtain the decomposition asserted in 2) of that lemma, where the divisor E is the support of the cokernel of ϕξ, and, from the part 3) of Lemma 2.3, we deduce that H · L1 ≤ 0. This together with the formula for L1 in (3.2) implies 0 ≥ H · L1 = H · (− 2 3 H + 1 3 B1) = 1 3 (B1 · H − 2d). Hence B1 · H ≤ 2d as asserted in 1) of the lemma. The above inequality and the Hodge Index Theorem give B2 inequality in (3.3), we deduce assertion 3) of the lemma. 1 ≤ 4d. Substituting this (cid:3) Lemma 3.3. If the Bogomolov filtration of Tξ is 0 = F0 ⊂ F1 ⊂ F2 = Tξ with F1 of rank 1, then c2 − K 2 X ≥(H · KX + 1 4 d , 2 H · KX + 1 1 4 d , if H · c1(F1) ≤ 0 if H · c1(F1) > 0. 17 3. Numerical invariants for surfaces on degree 4 hypersurfaces Proof. Let Q = Tξ/F1 and set L1 = c1(F1) and L2 = c1(Q). The bundle Tξ becomes the middle term of two exact sequences. 0 OX (−KX − H) 0 OX(L1) Tξ Q 0 ΩX 0 (3.4) The condition that OX (L1) is a Bogomolov destabilizing subsheaf of Tξ and the equality (−H) = L1 + L2 imply that the Q-divisor L1 − c1(Tξ) 3 = L1 − −H 3 = L1 − L1 + L2 3 lies in the positive cone of N (X). Setting = 2 3(cid:16)L1 − 1 2 L2(cid:17) we express L1 and L2 as linear combinations of H and B1: B1 = L1 − 1 2 L2 ∈ N +(X), L1 = − L2 = − 1 3 2 3 H + H − 2 3 2 3 B1 B1. (3.5) From here on we use the same argument as in Lemma 3.2. Namely, we use the vertical sequence in (3.4) to estimate the second Chern class of Tξ: c2 − K 2 X − H · KX = c2(Q) + L1 · L2 ≥ 1 4 L2 2 + L1 · L2 = 1 3 (d − B2 1), (3.6) where the inequality uses the condition that the quotient Q = F2/F1 is Bogomolov semi- stable, i.e., c2(Q) ≥ 1 2, and the last equality comes from substituting the expressions from (3.5). 4 L2 To conclude the argument we need an appropriate upper bound on the self-intersection B2 1. We argue according to the sign of H · L1. First case. If H · L1 ≤ 0, then we are essentially in the same situation as in the proof of Lemma 3.2. Namely, we have 0 ≥ H · L1 = 1 3 H · (−H + 2B1) = 1 3 (2B1 · H − d) , where the first equality uses the formula for L1 in (3.5). Thus we obtain H · B1 ≤ 1 hence, by the Hodge index, the upper bound B2 4 d. The inequality (3.6) then becomes 2 d and 1 ≤ 1 c2 − K 2 X − H · KX ≥ 1 3 (d − B2 1) ≥ 1 3(cid:16)d − 1 4 d(cid:17) = 1 4 d 18 3. Numerical invariants for surfaces on degree 4 hypersurfaces and this is equivalent to the first inequality of the lemma. Second case. If H · L1 > 0, then the morphism OX (L1) → ΩX given by slanted arrow in the diagram (3.4) is nonzero.7 Hence OX (L1) injects into ΩX and by the Bogomolov Lemma, L2 1 ≤ 0. This inequality and the formula (3.5) for L1 give 0 ≥ (−H + 2B1)2 = H 2 − 4H · B1 + 4B2 1 , or equivalently, B2 1 ≤ H · B1 − 1 4 d. (3.7) On the other hand, the generic semi-positivity of ΩX stipulates that the quotient sheaf ΩX/OX (L1) has non-negative degree with respect to any ample divisor on X. In particular, we deduce that 0 ≤ H · c1(ΩX /OX (L1)) = H · KX − H · L1. The above inequality and the formula for L1 in (3.5) imply H · B1 ≤ 1 2 (3H · KX + d). Combining (3.7) and (3.9) we obtain This inequality together with (3.6) gives the estimate B2 1 ≤ 3 2 H · KX + 1 4 d. c2 − K 2 X − H · KX ≥ − 1 2 H · KX + 1 4 d, (3.8) (3.9) which is equivalent to the second inequality of the lemma. (cid:3) Lemma 3.4. If the Bogomolov filtration of Tξ is 0 = F0 ⊂ F1 ⊂ F2 ⊂ F3 = Tξ, then it determines two divisors Bi, i = 1, 2, in the positive cone of N (X) and an effective nonzero divisor E such that the following conditions hold: 1) c1(F1) = 1 2) KX = 1 3 (2B1 + B2 − H), 3 (B1 + 2B2 − 2H) + E, 3) c2 − K 2 X ≥ H · KX + B2 1 + B1 · B2, if H · c1(F1) ≤ 0 3 4 H · KX + 1 8 (H · B1)(H · B2) d if H · c1(F1) > 0. Proof. Set Li = c1(Fi/Fi−1), 1 ≤ i ≤ 3. Hence c1(Fi) = L1 + · · · + Li, for i = 1, 2, 3. In particular, Since F1 is the maximal destabilizing subsheaf of F2 −H = c1(Tξ) = c1(F3) = L1 + L2 + L3. B1 = L1 − L2 = 2(cid:18)L1 − c1(F2) 2 (cid:19) ∈ N +(X), (3.10) 7Otherwise the divisor −(H + KX + L1) must be effective; however substituting the formula for L1 from (3.5), one obtains H + KX + L1 = 3 4 H + KX + 2 3 B1 which is of the Iitaka dimension 2. 19 3. Numerical invariants for surfaces on degree 4 hypersurfaces and similarly, since F2/F1 is the maximal destabilizing subsheaf of F3/F1, B2 = L2 − L3 = 2(cid:18)c1(F2/F1) − c1(F3/F1) 2 (cid:19) ∈ N +(X). (3.11) Combining (3.11) and (3.10) together with the decomposition H = −L1 − L2 − L3, we obtain the following formulas B2 B2 B2. (3.12) L1 = − L2 = − L3 = − 1 3 1 3 1 3 H + H − H − 2 3 1 3 1 3 B1 + B1 + B1 − 1 3 1 3 2 3 As in the two previous lemmas, we use the Bogomolov filtration of Tξ to estimate its X − H · KX . Since the filtration is a maximal ladder, second Chern number c2(Tξ) = c2 − K 2 it yields c2 − K 2 X − H · KX ≥ L1 · L2 + L1 · L3 + L2 · L3. Substituting the formulas from (3.12) leads to c2 − K 2 X − H · KX ≥ 1 3 (d − B2 1 − B2 2 − B1 · B2) = 1 3(cid:18)d − 1 4 (2B1 + B2)2 − 3 4 B2 2(cid:19) . (3.13) The argument continues, as in in the proof of Lemma 3.3, according to the sign of the intersection H · L1. First case. If H · L1 ≤ 0, then the formula for L1 in (3.12) gives d ≥ 2H · B1 + H · B2 = H · (2B1 + B2). This and the Hodge Index Theorem imply (2B1 + B2)2 ≤ d. (3.14) As a consequence, the inequality (3.13) becomes c2 − K 2 X − H · KX ≥ 1 4 (d − B2 2) ≥ B2 1 + B1 · B2, where the last inequality is obtained by substituting the upper bound for B2 Hence the first inequality in part 3) of the lemma. 2 from (3.14). Second case. If H · L1 > 0, then as in the proof of Lemma 3.3, we obtain L2 1 ≤ 0. This and the formula for L1 in (3.12) imply 2H · (2B1 + B2) ≥ d + (2B1 + B2)2. (3.15) Next we exploit the subsheaf F2 of the Bogomolov filtration of Tξ. Namely, combining the inclusion F2 ⊂ Tξ with the epimorphism of the extension sequence (2.7) gives rise to a nonzero morphism u : F2 −→ ΩX. Arguing exactly as in the proof of Lemma 3.2 one shows that this morphism is generically an isomorphism. Thus we can apply Lemma 2.3 to obtain the decomposition KX = c1(F2) + E = L1 + L2 + E = −H − L3 + E, 20 3. Numerical invariants for surfaces on degree 4 hypersurfaces where E is the support of the cokernel of u. This is the decomposition asserted in part 2) of the lemma. Furthermore, the last assertion in Lemma 2.3 tells us that (−H − L3) · H ≤ 0. This inequality combined with the formula for L3 in (3.12) gives the following inequality: 0 ≤ H · (H + L3) = d − 1 3 H · (H + B1 + 2B2) = 1 3 (2d − H · B1 − 2H · B2). Thus H · B1 + 2H · B2 ≤ 2d. (3.16) Rewriting this inequality as H · B2 ≤ d − 1 Hodge Index Theorem, we obtain 2 H · B1, implies that H · B2 < d and hence, by the B2 2 ≤ H · B2 H 2 H · B2 ≤ H · B2 d (cid:18)d − 1 2 H · B1(cid:19) = H · B2 − 1 2d (H · B1)(H · B2). (3.17) We now return to the inequality (3.13). Substituting the upper bounds for the second (resp. third) term from (3.15) (resp. (3.17)), we obtain c2−K 2 X − H · KX ≥ 1 12 (5d − 4H · B1 − 5H · B2) + 1 8d (H · B1)(H · B2) 1 12 (d − 2H · B1 − H · B2) + 1 8d (H · B1)(H · B2) 1 12 1 4 ≥ − = (4d − 2H · B1 − 4H · B2) + H · L1 + 1 8d (H · B1)(H · B2), where the last inequality is obtained by using (3.16) for the first parenthesis, and the formula for L1 in (3.12) for the second one. Thus we obtain c2 − K 2 X ≥ 3 4 H · KX + 1 4 (KX − L1) · H + 1 8d (H · B1)(H · B2) ≥ 3 4 H · KX + 1 8d (H · B1)(H · B2), where the last inequality uses (KX − L1) · H ≥ 0 coming from the generic semi-positivity of ΩX, see the discussion just above (3.8). This completes the proof of the second inequality in part 3) of the lemma. (cid:3) As a corollary of Theorem 3.1 we deduce the following property of surfaces lying on a quartic hypersurface in P4. Theorem 3.5. If X ⊂ P4 is a smooth surface lying on a hypersurface of degree four and not on one of a smaller degree, then K 2 X < 6χ(OX ). Proof. Assume K 2 X ≥ 6χ(OX ). We claim that X must be rational or irrationally ruled. Indeed, if this is not the case, Theorem 3.1 tells us that K 2 X = 6χ(OX ) and H · KX = 0. The second identity implies that KX = 0 in N (X). Hence K 2 X = χ(OX ) = 0. In particular, X must be irregular with irregularity q = 1 or 2, and of degree d = 10 (this is obtained from the double point formula). Furthermore, Theorem 3.1 tells us that the vector bundle Tξ is Bogomolov unstable with the filtration treated in Lemma 3.2. From the proof of that lemma we deduce the isomorphisms H 0(ΩX ) ∼= H 0(Tξ) ∼= H 0(F1). 21 3. Numerical invariants for surfaces on degree 4 hypersurfaces This implies that we must have q = 1 (otherwise X is an abelian surface and the above isomorphisms imply F1 ∼= ΩX, hence a splitting of the extension sequence (2.7)). Thus the Albanese variety Alb(X) is an elliptic curve, call it A, and the Albanese map a : X → A is an elliptic fibration. In addition, since c2(X) = 0, we also see that every reduced fibre is smooth. In particular, there are no smooth rational curves on X. This fact, in turn, implies that any divisor in the positive cone of N (X) is ample. Another aspect of the filtration in Lemma 3.2 is the decomposition of the canonical divisor 0 = KX = L1 + E = 1 3 (B1 − 2H) + E, where L1 = c1(F1) = above equation can be rewritten as follows 1 3 (B1 − 2H), see (3.1), and E is an effective nonzero divisor. The with B1 in the positive cone. Hence, as remarked above, the divisor B1 is ample. Next we investigate the divisor E. Using formula (3.18), we deduce that 2H = 3E + B1, (3.18) C · E + 1 3 C · B1 = 2 3 H · C, for any reduced irreducible component C of E. On the other hand, we know that for every C as above the subsheaf F1 ⊂ Tξ gives rise to a nonzero morphism (see [26]) Using again the expression for L1 from (3.2), we obtain OC(L1 + H) −→ ΘX ⊗ OC. (L1 + H) · C = 1 3 (H + B1) · C > 0. (3.19) (3.20) Hence the composition of the morphism (3.19) with the morphism ΘX ⊗ OC → OC(C) coming from the normal sequence of C ⊂ X yields a nonzero morphism8 This morphism together with formula (3.20) leads to OC (L1 + H) −→ OC(C). C 2 ≥ (L1 + H) · C = 1 3 (H + B1) · C > 0, (3.21) implying that E lies in the positive cone of N (X) and hence it is ample. From (3.18), it follows that 20 = 2H 2 = 3E · H + B1 · H ≥ 3E · H + 4. Thus E · H ≤ 5 and by the Hodge Index Theorem, we obtain E2 ≤ 2. Since the intersection form on NS(X) is even and E is ample, we deduce the equality E2 = 2. Hence E is reduced and irreducible. Thus replacing C by E in (3.21), we obtain 8The fact that X contains no smooth rational curve is used here again. 6 = 3E2 ≥ H · E + B1 · E. 22 3. Numerical invariants for surfaces on degree 4 hypersurfaces But the Hodge index tells us that H · E = 5 and that B1 · E ≥ 2 which contradicts the above inequality. X ≥ 6χ(OX ), implying K 2 Now we know that X is either rational or irrationally ruled. In the latter case χ(OX ) ≤ 0 and 8χ(OX ) ≥ K 2 X = χ(OX ) = 0. This identifies X as the projectivization P(E) of a rank 2 bundle E over an elliptic curve. But then it is well-known that X is an elliptic scroll of degree d = −H · KX = 5, see Lemma 7.8. It is easy to see (and again well-known, see Theorem 8.1) that such an X is contained in cubic hypersurfaces. This of course is contrary to our assumption that 4 is the smallest degree of a hypersurface containing X. We now turn to the remaining possibility: X is rational. In this case χ(OX ) = 1 X ≥ 6χ(OX ) = 6. By Riemann-Roch applied to OX (−KX ) it follows that and hence, K 2 h0(OX (−KX)) ≥ 7. Therefore, (−KX) is an effective nonzero divisor. Next we wish to have an upper bound for (−H · KX). This can be done by observing that which yields h0(OX (H)) ≥ χ(OX (H)) = H · (H − KX ) 2 + 1, d − H · KX ≤ 2h0(OX (H)) − 2. Furthermore, it is well-known that h0(OX (H)) = 5, unless X is the projection of the Veronese surface from a general point in P5. But such a projection lies on a hypersurface of degree 3 and this is contrary to our assumption that the smallest degree of a hypersurface containing X is 4. Thus h0(OX (H)) = 5 and we obtain the upper bound − H · KX ≤ 8 − d. (3.22) However, one observes the following. Claim. H 0(OX (−KX − H)) 6= 0. This concludes the argument, since H 0(OX (−KX − H)) 6= 0 together with the earlier estimate h0(OX (−KX )) ≥ 7 imply (−H · KX) > d. From this inequality and (3.22), one obtains d = 3. It follows immediately that X is the projection of the Veronese surface from a point on it. But then, an easy dimension count implies that X lies on a hypersurface of degree 2, contradicting our hypothesis. We now return to the proof of the claim. It is equivalent to showing that the homomor- phism H 0(OX (−KX)) → H 0(OC(−KX )) induced by the restriction of sections to a smooth irreducible curve C ∈ H is not injective. Assume on the contrary that it is injective. Then 7 ≤ h0(OX (−KX )) ≤ h0(OC (−KX )). If OC(−KX ) is non-special, then we compute the right hand side in the above inequality from the Riemann-Roch and obtain 7 ≤ h0(OX (−KX)) ≤ h0(OC (−KX )) = deg(−KX C) + 1 − g(C) = −3H · KX − d 2 . Equivalently, −H · KX ≥ impossible since d must be at least 3. 3 14 + d , which together with (3.22) tells us that d ≤ 2 which is If OC(−KX ) is special, then by the Clifford inequality 14 ≤ 2h0(OX (−KX )) ≤ 2h0(OC (−KX)) ≤ −H · KX + 2. Hence (−H · KX ) ≥ 12, which again is impossible in view of the upper bound (3.22). (cid:3) 23 3. Numerical invariants for surfaces on degree 4 hypersurfaces 3.2. The proof of Theorem 1.3 when mX = 4 For the reader's convenience we restate the theorem in this case as follows. Proposition 3.6. For all surfaces of general type contained in a quartic hypersurface in P4 and not in one of a smaller degree the following assertions hold. 1) The slope of their Chern numbers α = K 2 2) For every rational number α < 6, there exists d(α) ∈ N such that every surface of slope χ < 6. α has the degree d ≤ d(α). 3) For every rational number α < 6, all surfaces of the slope α and degree d have the holomorphic Euler characteristic χ ≤ χ(α, d), where χ(α, d) = d2 + 20d 8(6 − α) . Proof. Let X be a smooth surface in P4 with mX = 4, see (1.3) for notation. The first assertion of the proposition, αX < 6, is a reformulation of Theorem 3.5. For the second assertion, we use the inequality (1.4) for m = mX = 4 to obtain χ := χ(OX ) ≥ 1 96 d3 − 19 16 d2 + 10 3 d + 5 4 , (3.23) provided d ≥ 11. On the other hand the Chern numbers of X are related to the degree d via the double point formula which we write in the form 12χ − 2K 2 X = c2 − K 2 X = 5H · KX + 10d − d2. Setting K 2 X = αχ and substituting into the above equation give 5H · KX + 10d − d2 = 12χ − 2K 2 X = 12χ − 2αχ = 2(6 − α)χ. (3.24) From [15] the sectional genus of X has the following upper bound 2g(H) − 2 = H 2 + H · KX ≤ 1 4 d2. (3.25) Using this in the equation (3.24) we obtain 2(6 − α)χ = 5H · KX + 10d − d2 = 5(H · KX + d) + 5d − d2 ≤ 1 4 d2 + 5d. (3.26) This inequality combined with the first assertion of the proposition and the lower bound for χ in (3.23) give (3.27) (6 − α)(cid:18) 1 12 d3 − 19 2 d2 + 80 3 d + 10(cid:19) ≤ d2 + 20d, for all d ≥ 11. From this it follows that one can explicitly determine a positive integer d0(α) depending on α only, which is an upper bound for the solutions of the above inequality. Setting d(α) = max(10, d0(α)), one deduces the second assertion of the proposition. The upper bound for χ in the third assertion of the proposition is obtained from (3.26) (cid:3) and the first assertion of the proposition. 24 4. Numerical invariants for surfaces with mX = 2, 3, 5 As a consequence, we deduce the following finiteness result. Corollary 3.7. The number of families of surfaces in P4 of general type with fixed slope α and contained in a quartic hypersurface (and not in one of a smaller degree) is at most finite. Proof. The components of the Hilbert scheme of surfaces in P4 are labeled by Hilbert polynomials. So the proof of the statement comes down to checking that there is at most a finite number of such polynomials for the surfaces subject to the hypotheses of the corollary. Since the Hilbert polynomial of a surface X ⊂ P4 of degree d has the form PX(t) = d 2 t2 − H · KX 2 t + χ(OX ), one needs to see that there is only a finite number of possibilities for the triples (d, H · KX , χ(OX )). This is exactly what Proposition 3.6 tells us: parts 2) and 3) give a finite number of possibilities for d and χ(OX ) respectively. Once d takes a finite number of values, the inequality (3.25) insures that there is only a finite number of values for H · KX as well. (cid:3) Remarks. 1) If α ≤ 5, the inequality (3.27) gives rise to the relation d3 − 126d2 + 80d + 120 ≤ 0, provided the polynomial on the left side of (3.27) is non-negative9. From this it follows that all surfaces X of general type with slope α ≤ 5 and mX = 4 have degree d ≤ 125. 2) For complete intersections (4, a) in P4 with a = d 4 ≥ 2, we have KX = (a − 1)H and χ(OX ) = pg(X) + 1 =(cid:0)a+3 4 (cid:1) −(cid:0)a−1 4 (cid:1). Hence K 2 X = 4a(a − 1)2 χ(OX ) = 1 3 (2a3 − 3a2 + 7a − 3) αX = f (a) = 12a(a − 1)2 2a3 − 3a2 + 7a − 3 −→ 6 as a −→ ∞. It is easily checked that f (a) is an increasing function of a in the interval [2, +∞[. In particular, αX = f (a) > 5, for all a ≥ 7. These examples suggest the following questions: besides complete intersections, a) are there surfaces of general type in P4 which are contained in a quartic 3-fold (and not on one of a smaller degree) and whose slope α = K 2/χ > 5? b) is there an infinite sequence (Xn)n of surfaces of general type in P4 with mXn = 4 such that the slopes αXn = K 2 Xn/χ(OXn) converge to 6? 4. Numerical invariants for surfaces with mX = 2, 3, 5 This section treats the remaining values of the smallest degree mX of a hypersurface con- taining X ⊂ P4. As in Section 3, we assume that the Kodaira dimension of X is non-negative. 9This is the case for d ≥ 112. 25 4. Numerical invariants for surfaces with mX = 2, 3, 5 Consider the extension sequence (1.6) corresponding to ξ = δX (s) ∈ H 1(ΘX (−KX − (5 − m)H)) provided by Lemma 2.2. We begin with cases mX = 2 or 3. Theorem 4.1. If mX = 2, then c2 − K 2 3H · KX + 3 d, 9 3 2 4 H · KX + d, if Tξ is Bogomolov semistable if Tξ is Bogomolov unstable. Theorem 4.2. If mX = 3, then X ≥ min(cid:18)K · H + d, 2H · KX + 4 3 d, c2 − K 2 X ≥ 3 2 K · H + 1 3 d(cid:19), if Tξ is Bogomolov semistable if Tξ is Bogomolov unstable. We omit the proofs since they follow exactly the same pattern as the one of Theorem 3.1. We only mention, that in the case mX = 2 the only possibility for the Bogomolov filtration is as in Lemma 3.3. We now turn to the case mX = 5. The sequence (2.3) takes the form H 0(ΘP4X (−KX)) −→ H 0(NX (−KX)) δX−→ H 1(ΘX (−KX)) and we seek an analogue of Lemma 2.2. Lemma 4.3. Let s be a global section of NX(−KX ) corresponding to a quintic hyper- X ≤ 6χ(OX ). surface containing X. If δX (s) = 0 in H 1(ΘX (−KX)), then d ≤ 14 and K 2 Furthermore, if X is of general type, then pg ≤ 2. Proof. The vanishing of δX(s) implies that the section s of NX(−KX) comes from a nonzero section of ΘP4 ⊗ OX (−KX). Thus H 0(ΘP4 ⊗ OX (−KX)) 6= 0. This together with the Euler sequence of ΘP4 leads to two possibilities: 1. H 0(OX (H − KX)) 6= 0, 2. ker(cid:0)H 1(OX (−KX)) → H 0(OX (H))∗ ⊗ H 1(OX (H − KX ))(cid:1) 6= 0. Both of them tell us that H · KX ≤ H 2 = d. Furthermore, the equality holds if and only if OX (KX ) = OX (H) and then one knows that X must be a complete intersection which is impossible due to the condition mX = 5. Thus we have This and the double point formula give H · KX < d. 2(K 2 X − 6χ) = d2 − 10d − 5H · KX > d2 − 15d. (4.1) (4.2) In particular, one obtains that d ≤ 14, provided K 2 lemma is a consequence of the inequality K 2 X ≤ 6χ. X ≤ 6χ. Thus the first assertion of the Assume K 2 X > 6χ. This and the assumption that the Kodaira dimension of X is non- negative implies χ ≥ 1. Hence K 2 X ≥ 7 and H · KX > 0 tell us that X is of general type. 26 4. Numerical invariants for surfaces with mX = 2, 3, 5 Observe that the inequality H · KX < d and the Hodge index give also the upper bound K 2 X < d. Substituting into (4.2), we deduce 0 > d2 − 17d + 12, and hence d ≤ 16. This and the inequalities 6χ < K 2 X < d imply K 2 X ≤ 15 and χ ≤ 2. Furthermore, in the case K 2 H 2K 2 ≥ 16 · 15 implies H · KX ≥ 16 ≥ d contrary to (4.1). Thus one has K 2 X = 15 the degree d must be 16 and the Hodge index (H · KX )2 ≥ X ≤ 14. We now examine the remaining possibilities according to two possible values of χ = 1 or 2. If χ = 1, then the double point formula reads d2 − 10d − 5H · KX = 2K 2 X − 12. (4.3) By Bogomolov-Miyaoka-Yau inequality, K 2 X ≤ 9χ = 9. Thus K 2 X ∈ {7, 8, 9}, implying that d2 − 10d ≥ 2 + 5H · KX ≥ 32, (4.4) where the last inequality comes from the Hodge index (H ·KX )2 ≥ H 2K 2 X ≥ 5·7 = 35 and where the last inequality uses d ≥ 5, coming from the fact that X is of general type. Hence d ≥ 13. Using this lower bound in the Hodge Index estimate of H · KX once again, we obtain H · KX ≥ 10. Substituting this in (4.4), implies d ≥ 14. With the previously derived upper bound d ≤ 16, we obtain d = 14, 15, 16 are the only possible values. A direct check shows that d = 15 is incompatible with the double point formula (4.3), while for d = 14 and 16, that formula forces K 2 X = 9 and hence, H · KX = 10 and 18, respectively. The latter value contradicts the inequality (4.1), while the former, the Hodge Index inequality. X = d K 2 If χ = 2, then the double point formula becomes d2 − 10d − 5H · KX = 2K 2 X − 24, (4.5) while 13, 14 are the only possible values for K 2 X . Arguing as in the previous case we obtain the lower bound H · KX ≥ 9 which together with (4.5) gives the possibilities d = 14, 15, 16 for the degree of X. But then, going back to the lower bound estimate for H · KX , we obtain H · KX ≥ 14. Substituting into (4.5) gives d ≥ 15 and hence, d = 15 or 16. The value d = 15 is again incompatible with the double point formula (4.5), while for d = 16 that formula tells us that K 2 X = 14 and H · KX = 16. But this contradicts the inequality (4.1). The last statement saying that pg ≤ 2, for X of general type, can be seen as follows. Let X0 be the minimal model of X and set σ : X → X0 to be a sequence of blow-down maps. Then the canonical divisor KX can be written as follows KX = σ∗KX0 + D, where D is an effective divisor composed of curves contracted by σ. From the nonvanishing of H 0(ΘX (−KX)) = H 0(ΘX (−σ∗KX0 − D) it follows that H 0(ΘX(−σ∗KX0)) 6= 0 as well. Running the argument of the first paragraph of the proof for that group and using the fact that H 1(OX (−σ∗KX0)) = 0, we obtain H 0(OX (H − σ∗KX0)) 6= 0. Next we observe that (H − σ∗KX0) 6= 0, since otherwise H = σ∗KX0 = KX and, as it was argued above, this is incompatible with the condition mX = 5. 27 4. Numerical invariants for surfaces with mX = 2, 3, 5 Once (H − σ∗KX0) 6= 0, we take a divisor Γ ∈ H − σ∗KX0 and identify pg as follows pg = h0(OX (σ∗KX0)) = h0(OX (H − Γ)), i.e., pg is the dimension of the space of hyperplanes in P4 containing Γ. From this it follows that pg ≤ 3 and the equality holds if and only if Γ is a line in P4. We claim that this is impossible. Indeed, if Γ is a line then the identity H = Γ + σ∗KX0 implies 1 = Γ · H = Γ2 + Γ · σ∗KX0. (4.6) Hence Γ · σ∗KX0 > 0 and hence Γ is not in the exceptional divisor D. Therefore Γ · D ≥ 0. But then the identity (4.6) can be rewritten as 1 = Γ2 + Γ · σ∗KX0 = Γ2 + Γ · (KX − D) = (Γ2 + Γ · KX) − Γ · D = −2 − Γ · D, which is absurd. Thus pg ≤ 2 as asserted. (cid:3) Question. Can one enumerate all surfaces in P4 with mX = 5 and δX (s) = 0? From now on we set ξ = δX(s) ∈ H 1(ΘX(−KX )) and assume it to be nonzero. The corresponding extension sequence has the form 0 −→ OX (−KX ) −→ Tξ −→ ΩX −→ 0. (4.7) The Bogomolov semistability/instability considerations of the sheaf Tξ above give the follow- ing. Theorem 4.4. Let X be a surface in P4 with mX = 5. Then, either K 2 X ≤ c2, or X is a surface of general type subject to the following properties: i) The canonical divisor KX admits a distinguished decomposition KX = L + E, such that L is in the positive cone of N (X) and E is an effective nonzero divisor. ii) 0 < K 2 X − c2 ≤ 2 3 L2 ≤ 2 3 d. Proof. We may assume that K 2 X > c2, since otherwise there is nothing to prove. This assumption and Lemma 4.3 insure the nonvanishing of the cohomology class ξ = δX (s) defined by a section s ∈ H 0(NX (−KX)) arising from a hypersurface of degree 5 containing X. We associate to ξ the extension sequence (4.7) and observe that the assumption K 2 X > c2 implies that the sheaf Tξ in the middle of that sequence is Bogomolov unstable. Let F be the maximal Bogomolov destabilizing subsheaf of Tξ. Observe that the Bogomolov destabilizing property implies that OX (L), the determinant of F, is in the positive cone of N (X). Putting the inclusion F ⊂ Tξ together with the extension sequence (4.7) we obtain the diagram 0 F Tξ ϕ ΩX 0 0 OX (−KX) (4.8) where the slanted arrow is the composition of the inclusion together with the epimorphism in (4.7). We claim that F must be of rank 2 and the morphism ϕ in the above diagram is 28 4. Numerical invariants for surfaces with mX = 2, 3, 5 generically an isomorphism. Indeed, if the rank of F is one, then it is OX (L) and ϕ must be zero, since by the Bogomolov Lemma, ΩX admits no rank 1 subsheaves of Iitaka dimension 2. But the vanishing of ϕ implies OX(L) = OX (−KX). Hence X is rational with K 2 X > 6χ = 6. This together with the Hodge index gives a lower bound on the degree of −KX : − KX · H ≥ 5. (4.9) On the other hand, the Riemann-Roch for OX (H) yields d − KX · H h0(OX (H)) ≥ + 1. 2 Furthermore, h0(OX (H)) = 5, since X is not the projection of the Veronese surface from a point outside its secant variety10. Substituting this in the above inequality we obtain d − KX · H ≤ 8. This and the lower bound (4.9) imply d = 3 and −KX · H = 5. Therefore, a general hyperplane section, call it C, is a rational normal cubic in P3 and hence C is cut out by quadrics in P3. But then, in view of the isomorphism H 0(JX(2)) ∼= H 0(JC(2)), one obtains quadric hypersurfaces containing X contradicting the assumption mX = 5. We know now that the rank of F is 2 and we need to check that the morphism ϕ in the diagram (4.8) is generically an isomorphism. Assuming that this is not the case and taking the second exterior power of the diagram (4.8), we deduce a nonzero morphism OX(L) → ΘX. This implies once again that X must be rational with K 2 X > 6χ = 6. Thus we are back in the situation considered in the preceding paragraph implying that ϕ in (4.8) is generically an isomorphism. Hence we are in the position to apply Lemma 2.3; in particular, we obtain the decomposition of the canonical divisor KX asserted in i) of the theorem. Furthermore, Lemma 2.3, 3) gives us the upper bound This, by the Hodge Index Theorem, implies L · H ≤ H 2 = d. L2 ≤ d. (4.10) The part ii) of the theorem is obtained by estimating the second Chern number of Tξ from the vertical sequence in the diagram (4.8). Namely, we complete it to the exact sequence 0 −→ F −→ Tξ −→ JZ(−L) −→ 0 from which we obtain c2 − K 2 X = c2(Tξ) = c2(F) − L2 + deg(Z) ≥ c2(F) − L2 ≥ 1 3 L2 − L2 = − 2 3 L2, where the second inequality is a result of Miyaoka, [24, Remark 4.18]. Hence the first upper bound K 2 X − c2 ≤ L2 2 3 asserted in the part ii) of the theorem. This together with (4.10) gives the second upper bound in ii). (cid:3) 10Such a surface is contained in cubic hypersurfaces and this is contrary to our assumption that the minimal degree of a hypersurface containing X is 5. 29 4. Numerical invariants for surfaces with mX = 2, 3, 5 Proposition 4.5. For all surfaces of general type lying on a quintic hypersurface in P4 and not on one of a smaller degree, the following assertions hold: 1) For every rational number α 6= 6, there exists d(α) ∈ N such that every surface of slope α, has the degree d ≤ d(α). 2) For every rational number α 6= 6, all surfaces with slope α and degree d ≤ d(α) have the holomorphic Euler characteristic subject to the inequalities χ ≤( 5d 6−α , d 3(α−6) , if α < 6, if α > 6. 3) The slope of Chern numbers satisfies α = K 2 χ < 6 + 1 2 , with possible exceptions of surfaces having degree ≤ 37 and holomorphic Euler characteristic ≤ 24. Proof. For α < 6 the argument is analogous to the one in the proof of Proposition 3.6. Namely, the Decker-Schreyer polynomial in (1.2) for m = 5 has the form P5(d) = 1 25 d(d2 − 40d + 95) and their result gives the lower bound χ ≥ 1 25 d(d2 − 40d + 95) (4.11) for χ, the holomorphic Euler characteristic, for every surface of degree d ≥ 18 contained in a quintic hypersurface and not in one of a smaller degree. Combining this with the double point formula, we obtain 2 25 (6 − α)d(d2 − 40d + 95) ≤ 2(6 − α)χ = 12χ − K 2 = 5H · KX + 10d − d2. (4.12) From [15] we also know that the genus gH of a (smooth) hyperplane section of X is subject to the inequality From this and (4.12), we deduce d + H · KX = 2gH − 2 ≤ d2 + 5d 5 . (4.13) 1 25 (6 − α)(d2 − 40d + 95) ≤ 5d. Hence one can explicitly determine a positive integer d0(α) depending only on α, which is an upper bound for the integer solutions of the above inequality. Setting d(α) := max(17, d0(α)), we obtain assertion 1) of the proposition in the range α < 6. If α > 6, then Theorem 4.4 ii) tells us that (α − 6)χ = K 2 − 6χ ≤ 1 3 d. Combining this with (4.11) gives the inequality 1 25 (α − 6)(d2 − 40d + 95) ≤ (α − 6)χ d ≤ 1 3 , (4.14) which provides the assertion 1) of the proposition in the range α > 6. The second inequality for χ in 2) is just a restatement of Theorem 4.4, ii), while the first one is obtained by combining (4.12) and the upper bound 5H · KX ≤ d2 from (4.13). 30 5. The irregularity of surfaces lying on a small degree hypersurface For the third part of the proposition which asserts the upper bound for the slope α, we may assume α > 6 and go back to the inequality (4.14). For d ≥ 38, that inequality implies α − 6 ≤ 25 3 · 19 < 1 2 . For d ≤ 37, we use the inequality in Theorem 4.4, ii), to deduce (α − 6)χ = K 2 − 6χ ≤ 1 3 d ≤ 37 3 . Hence (α − 6)χ ≤ 12 and α < 6 + 1 2 , unless χ ≤ 24. (cid:3) As a consequence, we obtain the following finiteness result. Corollary 4.6. The number of families of surfaces in P4 of general type with fixed slope α 6= 6 and contained in a quintic hypersurface (and not in one of a smaller degree) is at most finite. Proof. See the proof of Corollary 3.7. (cid:3) (II) Irregularity of surfaces in P4 5. The irregularity of surfaces lying on a small degree hypersurface One of the outstanding problems about surfaces in P4 is the control of their irregularity. Since the beautiful work of Horrocks and Mumford, [17], which contains a construction of an abelian surface of degree 10 in P4, the irregularity 2 remains the maximal known value for surfaces in P4. Indeed, it is conjectured that no surface of irregularity greater than 2 can be embedded into P4. To our knowledge there is no conceptual reason for this phenomenon. In our previous work, [26], we were able to show that the irregularity of surfaces in P4 is bounded by 3 under a certain precise set of conditions, see [26, Theorem 5.1]. What is perhaps more interesting, is that we have uncovered how to use the cohomological condition H 1(ΘX(−KX )) 6= 0 to bound the irregularity of a surface X ⊂ P4. In the previous sections we have seen that this non-vanishing condition arises naturally whenever X is contained in a hypersurface of a degree m ≤ 5. This phenomenon together with the metha-principle formulated in the introduction suggests that the irregularity of surfaces in P4, with the exception of a finite number of families, is bounded by the irregularity of surfaces contained in hypersurfaces of small degree. Guided by this heuristic principle, this section begins the investigation of the irregularity of surfaces in P4 contained in a small degree hypersurface. Our study of irregular surfaces in P4 lying on a small degree hypersurfaces has the same unifying theme as before: the extension construction. We consistently interpret a small degree hypersurface containing our surface as an extension sequence of sheaves on X. Let m be the smallest degree of a hypersurface containing a surface X ⊂ P4 and let Vm be such a hypersurface. We recall that NX (resp. N ∗ X) denotes the normal (resp. conormal) bundle of X in P4 and one has the following identifications JX/J 2 X = N ∗ X ∼= det(N ∗) ⊗ NX = NX(−KX − 5H), 31 5. The irregularity of surfaces lying on a small degree hypersurface where JX is the ideal sheaf of X in P4 and the second identification is due to the rank of N ∗ X being two. This leads to the non-vanishing H 0(NX(−KX − (5 − m)H)) = H 0(N ∗ X(mH)) 6= 0 already encountered in (2.2). Thus, we associate to Vm a nonzero global section, denoted by s, of the twisted normal bundle NX(−KX − (5 − m)H). Our approach consists of using this section to build up an appropriate extension sequence. In the context of the irregularity, there are two lines of thinking. The first one is to use the coboundary map δX in the normal exact sequence for X ⊂ P4 to produce the cohomology class ξ = δX (s) ∈ H 1(ΘX (−KX − (5 − m)H)) and then, to view it, via the natural identification H 1(ΘX (−KX − (5 − m)H)) ∼= Ext1(ΩX, OX (−KX − (5 − m)H)), as an extension 0 −→ OX (−KX − (5 − m)H) −→ Tξ −→ ΩX −→ 0. This was the idea exploited in [26] and it produces satisfactory results provided that X is of Albanese dimension 2, i.e., the image of the Albanese morphism of X is of dimension 2. However, if X fibers over a curve B of genus g(B) = q(X), the method fails. This brings us to the second way of associating an extension sequence to the section s ∈ H 0(NX(−KX − (5 − m)H)). Namely, we simply take the Koszul sequence associated to s to obtain (5.1) 0 −→ OX(KX + (5 − m)H) s−→ NX ∧s−→ JZ(mH) −→ 0, where Z ⊂ X is the scheme of zeros of s and JZ its ideal sheaf. This is of course very classical and yet efficient in addressing the irregularity problem, provided we have a good control of the subscheme Z. Let us explain the main ingredients of this approach as well as set up the stage for more technical considerations in the subsequent sections. The extension (5.1) fails to be exact precisely when Z = (s = 0) has divisorial part. If this is the case, let Z1 be the divisorial part and Z0 be the residual part of Z1 in Z. If s1 is a section of OX (Z1) defining Z1, then s = s1s0, where s0 is a section of NX(−KX − (5 − m)H − Z1) whose zero-locus is Z0, a 0-dimensional subscheme of X. We now have the short exact sequence 0 −→ OX (KX + (5 − m)H + Z1) s0−→ NX ∧s0−→ JZ0(mH − Z1) −→ 0, (5.2) where JZ0 is the ideal sheaf of Z0. At this stage we bring in the irregularity. The main point in relating the irregularity of X to the extension sequence (5.2) is the following general fact. Lemma 5.1. If X ⊂ Pn is a complex projective manifold of dimension bigger than 1, then H 0(ΩX) ∼= H 1(N ∗ X), where ΩX(resp. N ∗ X) is the cotangent (resp. conormal) bundle of X. Proof. From the conormal exact sequence 0 −→ N ∗ X −→ ΩPnX −→ ΩX −→ 0 32 5. The irregularity of surfaces lying on a small degree hypersurface of X ⊂ Pn we obtain 0 −→ H 0(ΩX ) −→ H 1(N ∗ X ) −→ H 1(ΩPnX) r−→ H 1(ΩX) where the injectivity on the left is the vanishing of H 0(ΩPn(cid:12)(cid:12)X). The asserted isomorphism follows from the injectivity of the homomorphism r : H 1(ΩPnX ) → H 1(ΩX). To estab- lish it we use the dual of the Euler sequence and the assumption dim(X) > 1 to deduce H 1(ΩPnX ) ∼= H 0(X, OX ) ∼= C. At the same time, we have the linear map C ∼= H 1(ΩPn) i∗ −→ H 1(ΩX ) defined by the pullback i∗, where i : X ֒→ Pn is the inclusion morphism. This linear map factors through H 1(ΩPnX ) to give rise to the commutative diagram H 1(ΩPn) i∗ r H 1(ΩX) H 1(ΩPnX) and the injectivity of r follows from the injectivity of i∗. The latter is injective, since it sends the generator c1(OPn(1)) ∈ H 1(ΩPn) to the class of a hyperplane section of X. (cid:3) In case X is a surface, the isomorphism in Lemma 5.1 and the Serre duality yield the identification H 0(ΩX ) ∼= H 1(N ∗ X) ∼= H 1(NX(KX ))∗. Now we can see that the extension sequence (5.2) tensored with OX (KX ) is tied to the irregularity of X via the exact sequence H 1(OX (2KX + (5 − m)H + Z1)) s0−→ H 1(NX (KX )) ∧s0−→ H 1(JZ0(KX + mH − Z1)). (5.3) Thus the problem of computing or bounding the irregularity comes down to the understanding of the cohomology groups H 1(OX (2KX + (5 − m)H + Z1)) and H 1(JZ0(KX + mH − Z1)). This in turn depends on controlling the subscheme Z and the decomposition Z = Z1 + Z0. The subscheme Z is related to the singular locus of the hypersurface Vm which was used to define the section s. Let us spell out this relationship. The normal sequence of X ⊂ P4 and the Euler sequence of ΘP4 give rise to the surjective morphism H 0(OX (H))∗ ⊗ OX (H) −→ NX. This together with the section s ∈ H 0(N ∗ tative diagram X(mH)) -- corresponding to Vm -- yields the commu- H 0(OX (H))∗ ⊗ OX (H) (5.4) NX s∧ JZ(mH) 33 6. Irregular surfaces on hypersurfaces of degree 2 where the composite (slanted) arrow is given by the partial differentiation of a homogeneous polynomial defining Vm. Thus, denoting by IVm the sheaf of ideals defined by the Jacobian ideal of Vm, the diagram (5.4) tells us that we have the inclusion JZ ⊃ IVm ⊗ OX or, equivalently, that Z is a subscheme of the scheme-theoretic intersection of the singular locus Sing(Vm) with X. So it is reasonable to expect that one could control Z and hence, the cohomology groups in (5.3), for small values of m. As an easy illustrative example of the above ideas, let us work out the case m = 2. 6. Irregular surfaces on hypersurfaces of degree 2 The following theorem, no doubt, is well-known to experts, see e.g., [30], p. 152. Theorem 6.1. If X ⊂ P4 is a smooth surface lying on a quadric hypersurface V2, then the irregularity of X vanishes. Proof. Suppose that X ⊂ V2. We may assume that V2 is singular, since otherwise X is a complete intersection and we are done by Lefschetz hyperplane theorem. The singular locus Sing(V2) is either a single point, the vertex p of V2, or a line, L. The case Sing(V2) = {p}. The extension sequence (5.1) takes the form 0 −→ OX (2KX + 3H) s−→ NX(KX ) −→ JZ(KX + 2H) −→ 0 where Z is either empty or p. Thus the sequence of cohomology groups (5.3) becomes H 1(OX (2KX + 3H)) s−→ H 1(NX(KX )) −→ H 1(JZ (KX + 2H)). Thus the irregularity is bounded as follows. q(X) = h1(NX(KX )) ≤ h1(OX (2KX + 3H)) + h1(JZ (KX + 2H)). From [29] it follows that the divisor KX +3H is very ample. Hence, by the Kodaira vanishing and the Serre duality, we obtain h1(OX (2KX + 3H)) = 0. We turn now toward h1(JZ (KX + 2H)). If Z = 0, then H 1(JZ (KX + 2H)) = H 1(OX (KX + 2H)) (SD) ∼= H 1(OX (−2H))∗ = 0, and hence, q(X) = 0 in this case. If Z = p, then the non-vanishing of H 1(Jp(KX + 2H)) means that p is a base point of OX(KX + 2H), but this is ruled out by [29, Theorem 1,(i)]. Hence H 1(Jp(KX + 2H)) = 0. This completes the proof of the theorem when Sing(V2) is a point. The case Sing(V2) = L. In this case V2 is a singular rational scroll over a smooth conic C lying in a plane Π complementary to the line L. It is ruled by the one parameter family of planes {Pt}t∈C , where the plane Pt is the span of t and L. There are two cases to consider according to whether or not the line L is contained in X. 34 6. Irregular surfaces on hypersurfaces of degree 2 • If L ⊂ X, then the projection from L defines the morphism ϕ : X −→ C ∼= P1 with the fibre Ft over a point t ∈ C being the component of the intersection Pt · X com- plementary to L. This implies that the geometric ingredients Z1 and Z0 in the cohomology sequence (5.3) are the line L and the empty set, respectively. Hence that sequence takes the form H 1(OX (2KX + 3H + L)) −→ H 1(NX (KX )) −→ H 1(OX (KX + 2H − L)). We claim that the cohomology groups H 1(OX (2KX + 3H + L)) and H 1(OX (KX + 2H − L)) vanish, since the divisors 2H −L and KX +3H +L are both ample. Indeed, writing 2H −L = H + (H − L) and using the fact that the linear system H − L is base point free (the linear system defines the morphism ϕ), we deduce that 2H − L is very ample. The ampleness of KX +3H +L is checked easily using the very ampleness of KX +3H and the Nakai-Moishezon criterion, see [16]. • If L 6⊂ X, then the sequence (5.1) associated to V2 is exact and has the form 0 −→ OX(KX + 3H) −→ NX −→ JZ(2H) −→ 0, (6.1) where Z is the 0-dimensional subscheme of X, the scheme-theoretic intersection of X and the line L. This gives the relations d2 − 4d − deg(Z) = 4(gH − 1) and deg(Z) < d. (6.2) We will now calculate deg(Z) using the geometry of the singular scroll V2 containing X. Namely, we go back to the plane Π containing the conic C and complementary to L and consider a hyperplane Span(L, Λ) in P4 spanned by L and a general line in Λ ⊂ Π. The hyperplane Span(L, Λ) intersects V2 in the union of two planes Pt, where t ∈ Λ · C. Hence H = Span(L, Λ) · X = 2F , where F is the class of the divisor Pt′ · X, for a closed point t′ ∈ C. From this it follows that d = 4F 2, and, by adjunction, 2(gH − 1) = 2F · KX + 4F 2 = 2 deg(ωF ) + 2F 2 = 2 deg(ωF ) + d 2 , where ωF is the dualizing sheaf of F . On the other hand, since F is a plane divisor, its dualizing sheaf is subject to deg(ωF ) = dF (dF − 3) = d 2(cid:18) d 2 − 3(cid:19) = 1 4 d(d − 6). Substituting into the previous identity, we have 2(gH − 1) = 1 2 d(d − 6) + d 2 = 1 2 d(d − 5). Putting it together with the identity in (6.2), we obtain deg(Z) = d which contradicts the inequality in (6.2). (cid:3) 35 7. Irregular surfaces on hypersurfaces of degree 3 We close this 'warm up' section with a simple general observation concerning the sheaf JZ0(mH −Z1) in the sequence (5.2). This observation will be important in studying irregular surfaces lying on cubic hypersurfaces. Lemma 6.2. The sheaf JZ0((m − 1)H − Z1) (see the sequence (5.2) for notation), is gen- erated by global sections outside the 0-dimensional subscheme Z0. Furthermore, h0(JZ0((m− 1)H − Z1)) ≥ 5, unless the hypersurface Vm is a cone over a surface in P3. Proof. From the sequence (5.2) tensored with OX (−H) it follows that JZ0((m−1)H −Z1) is the quotient of NX(−H) which is globally generated. Hence the first statement of the lemma. For the second statement, we use the diagram (5.4) tensored with OX (−H). The slanted arrow is the morphism H 0(OX (H))∗ ⊗ OX −→ JZ ((m − 1)H) which factors through JZ0((m − 1)H − Z1) and induces, at the level of global sections, the linear map ∂(f )X : H 0(OX (H))∗ −→ H 0(JZ0((m − 1)H − Z1)), (6.3) where f is a homogeneous polynomial defining Vm. The map ∂(f )X takes a vector v ∈ H 0(OX (H))∗ to the partial derivative ∂v(f )X restricted to X and then divides it by a section defining Z1. From this it follows that the kernel of ∂(f )X consists of vectors v ∈ H 0(OX (H))∗ for which ∂v(f ) is a homogeneous polynomial of degree m − 1 vanishing on X. Since by definition, m is the least degree of such polynomials, we deduce that ∂v(f ) = 0 in Symm−1(H 0(OX (H))) or, equivalently, f ∈ Symm−1(v⊥), where v⊥ = {l ∈ H 0(OX (H)) l(v) = 0}. In particular, for v 6= 0, the above means that the point [v] ∈ P(H 0(OX (H))∗) = P4 is the vertex of the cone over the surface in P((v⊥)∗) ∼= P3 defined by f , viewed as a homogeneous polynomial on P((v⊥)∗). Hence, unless Vm is a cone over a surface in P3, the operator ∂(f )X in (6.3) is injective and, therefore, h0(JZ0((m − 1)H − Z1)) ≥ h0(OX (H)) ≥ 5 as asserted in the lemma. (cid:3) 7. Irregular surfaces on hypersurfaces of degree 3 The main result of this section is the following characterization of irregular surfaces contained in a cubic hypersurface in P4. Results concerning surfaces contained in a cubic hypersurface in P4 can be found, e.g., in [21] and [30]. Theorem 7.1. If X ⊂ P4 is a smooth irregular surface contained in a cubic hypersurface V3, then X is an elliptic scroll of degree 5. Moreover, a general cubic hypersurface containing X is a Segre cubic and its ten singular points lie on X. 36 7. Irregular surfaces on hypersurfaces of degree 3 As we have already explained, our main tool is the exact sequence (5.2) which, for m = 3, takes the form 0 −→ OX (KX + 2H + Z1) −→ NX −→ JZ0(3H − Z1) −→ 0. (7.1) The corresponding cohomological sequence (5.3) controlling the irregularity of X becomes H 1(OX (2KX + 2H + Z1)) −→ H 1(NX (KX ) −→ H 1(JZ0(KX + 3H − Z1)). (7.2) To analyse the cohomology group on the left in the above sequence we will need the following. Proposition 7.2. If H 1(OX (2KX + 2H)) 6= 0, then X is an elliptic scroll of degree 5. Proof. Assume that H 1(OX (2KX + 2H)) does not vanish. Using the Serre duality and [29], we deduce that the line bundle OX (KX + 2H) is base point free but not big. Since OX(KX + 2H) is not trivial11, we deduce that the morphism defined by OX (KX + 2H) In other words, there is a morphism ϕ : X → B with connected maps X onto a curve. fibres onto a smooth curve B, and a base point free line bundle OB(D) on B such that OX(KX + 2H) = ϕ∗OB(D). This implies that KX + 2H = deg(D)F, where F is the class of a general fibre of ϕ. Taking the intersection with F on both sides of the above identity implies that H · F = 1. This means that the fibres of ϕ are lines and that X is a minimal ruled surface embedded into P4 by OX (H) as a scroll with irregularity q = q(X) = g(B), the genus of B. It is well known that the only irregular scroll in P4 is an elliptic scroll of degree 5, see Lemma 7.8. (cid:3) Next we turn to the group on the right of the sequence (7.2). Lemma 7.3. If V3 is not a cone and Z1, the divisorial part in (7.2), is non-zero, then H 1(JZ0(KX + 3H − Z1)) = 0. Proof. Assume that H 1(JZ0(KX + 3H − Z1)) 6= 0. From the identification H 1(JZ0(KX + 3H − Z1))∗ ∼= Ext1(JZ0(3H − Z1), OX ), the supposed nonvanishing is interpreted as a nontrivial extension 0 −→ OX −→ E −→ JZ0(3H − Z1) −→ 0. Tensoring it with OX (−H), we obtain H 0(E(−H)) ∼= H 0(JZ0(2H − Z1). This and Lemma 6.2 imply h0(E(−H)) ≥ 5. (7.3) (7.4) (7.5) This is going to play the role of a destabilizing condition for E. Namely, let t be a nonzero global section of E(−H). It gives rise to the exact sequence 0 −→ OX (A) −→ E(−H) −→ JZ ′(H − A − Z1) −→ 0, (7.6) 11Otherwise OX (KX ) = OX (−2H) and hence q(X) = 0. 37 7. Irregular surfaces on hypersurfaces of degree 3 where A is the divisorial part of the zero locus of t and Z ′ is the 0-dimensional residual part of (t = 0). We proceed now to analyse the above sequence according to the dimension of the linear system A. If h0(OX (A)) ≤ 2, then (7.5) implies h0(JZ ′(H − A − Z1)) ≥ 3. Since Z1 6= 0, it follows that Z1 is a line and A = 0. But then, h0(JZ ′(H − A − Z1)) = h0(JZ ′(H − Z1)) = 3 and the sequence (7.6), together with the estimate (7.5), imply 5 ≤ h0(E(−H)) ≤ h0(OX ) + h0(JZ ′(H − Z1)) = 1 + 3 = 4, an obvious contradiction. Thus h0(OX (A)) ≥ 3. Combining the sequence (7.6) with the defining sequence (7.3) tensored with OX (−H), we obtain the nonzero morphism OX(A) −→ JZ0(2H − Z1). Since this morphism can not be an isomorphism12, it is given by a nonzero section e ∈ H 0(JZ0(2H − Z1 − A)) vanishing on the nonzero divisor E = (e = 0) ∈ 2H − Z1 − A. In particular, the image of the morphism H 0(OX (A)) e→ H 0(JZ0(2H −Z1)) consists of global sections of JZ0(2H − Z1)) vanishing on E. In view of the global generation of JZ0(2H − Z1)) outside Z0, see Lemma 6.2, it follows that e H 0(OX (A)) is a proper subspace of H 0(JZ0(2H − Z1)). Combining this and the isomorphism (7.4), we deduce that the image of H 0(OX (A)) under the monomorphism in (7.6) is a proper subspace of H 0(E(−H)). Hence H 0(JZ ′(H − Z1 − A)) 6= 0. Let D = H − Z1 − A. By the above, it is an effective divisor and h0(OX (H − D)) = h0(OX (Z1 + A)) ≥ 3. This tells us that either D = 0 or it is a line and the above inequality must be an equality. The first possibility is equivalent to A = H − Z1 and JZ ′(H − Z1 − A) = OX. This implies h0(E(−H)) ≤ 4 which contradicts (7.5). If D = H − Z1 − A is a line, then the estimate h0(OX (H − Z1 − D) = h0(OX (A) ≥ 3 implies that Z1 + D is a line. But since Z1 6= 0, this is impossible. (cid:3) The above lemma implies the following. Lemma 7.4. If V3 is not a cone and the divisorial part Z1 in (7.2) is nonzero, then X is a regular surface. Proof. From Lemma 7.3 and (7.2) it follows that the irregularity of X is controlled by the group H 1(OX (2KX + 2H + Z1). This group is related to the group H 1(OX (2KX + 2H) considered in Proposition 7.2 via the obvious exact sequence H 1(OX (2KX + 2H) −→ H 1(OX (2KX + 2H + Z1) −→ H 1(OZ1(2KX + 2H + Z1)), (7.7) the understanding of which requires to have a good grasp of Z1. This is the case, since we know that Z1 is contained in the 1-dimensional part of the singular locus Sing(V3). For a cubic hypersurface, which is not a cone over a cubic surface, the 1-dimensional part of its singular locus is known to be 12Otherwise the extension sequence (7.3) is trivial. 38 7. Irregular surfaces on hypersurfaces of degree 3 (i) a line, (ii) a conic (possibly singular), (iii) a rational normal curve of degree 4 in P4. Thus Z1 is one of the above possibilities and we analyse each of them separately. Case (i) -- Z1 is a line. In the exact sequence (7.7), the degree of the line bundle appearing in the group H 1(OZ1 (2KX + 2H + Z1)) satisfies Z1 · (2KX + 2H + Z1) = Z1 · KX + 2 + Z1 · (KX + Z1) = Z1 · KX + 2 − 2 = Z1 · KX = −Z 2 1 − 2. 1 ≤ −1, it follows that h1(OZ1(2KX + 2H + Z1)) = 0. Combining this and (7.7), we If Z 2 obtain h1(OX (2KX + 2H + Z1)) ≤ h1(OX (2KX + 2H) Moreover, from Proposition 7.2, we know that h1(OX (2KX + 2H) 6= 0 only if X is an elliptic scroll. But the only lines on such a surface X are the rulings. Therefore Z 2 1 ≤ −1 implies that h1(OX (2KX + 2H) and hence h1(OX (2KX + 2H + Z1)) are both equal to zero. Thus we may assume Z 2 1 > 0. Hence X, if irregular, is a scroll. The proof of Proposition 7.2 tells us that it must be an elliptic scroll of degree d = 5. In addition, computing the second Chern number of NX from the exact sequence (7.1), we obtain 1 = 0, since X is obviously regular if Z 2 25 = d2 = (KX + 2H + Z1)(3H − Z1) + deg(Z0) = 3d + 3 + deg(Z0) = 18 + deg(Z0) and hence deg(Z0) = 7. (7.8) We decompose the 0-dimensional subscheme Z0, the zero locus of the section s0 in (7.1), into two parts, Z0 = Z 1 0 + Z ′ 0, (7.9) where Z 1 on singular points of V3 belonging to X r Z1. 0 is the part of Z0 lying on the line Z1 and the residual subscheme Z ′ 0 is supported To estimate the degree of Z 1 0 we use the exact sequence 0 −→ OX (−kZ1) −→ JZ 1 0 −→ OkZ1(−Z 1 0 ) −→ 0, where k ≥ 1 is the smallest multiple of Z1 containing the subscheme Z 1 OX(2H − Z1) and using the fact that the linear system JZ 1 dimensional base locus, see Lemma 6.2, we deduce that h0(OkZ1(−Z 1 From this and Z 2 0 . Tensoring with (2H − Z1) has at most a 0- 0 ) ⊗ OX(2H − Z1)) > 0. 1 = 0, we obtain 0 0 ≤ deg(2HZ1 − Z 1 0 ) = 2 − deg(Z 1 0 ) 0 ) ≤ 2. This together with (7.8) tells us that the part Z ′ or, equivalently, deg(Z 1 0 of the decomposition (7.9) has degree at least 5. In particular, V3 must have singular points lying on X and outside the line Z1. However, this is impossible. Indeed, let p be a singular point of V3 lying on X r Z1. Then the plane P spanned by Z1 and p must be contained in V3. Now consider the pencil of hyperplanes {V (t) t ∈ P1} containing P . Each V (t) intersects 39 7. Irregular surfaces on hypersurfaces of degree 3 V3 along PS Qt, where Qt is a quadric surface in V (t) passing through Z1 and the point p. The hyperplane sections Ht = V (t) · X are reducible and have the form Ht = V (t) · X = Z1 + Γt, where Γt is the divisor on X residual to Z1. Furthermore, Γt has degree 4 and is contained in Qt. One sees immediately that Γt can not be irreducible, since otherwise it is a smooth elliptic curve in Qt and its intersection with the ruling Z1 must be 2, but on X the curve Γt is a section of X and Γt · Z1 = 1. Hence Γt = Γ0 + Lt, where Γ0 is a smooth plane cubic, the fixed part of the pencil {Ht}, and Lt is a ruling of X. This means that the rulings Lt of X are rationally equivalent and this is clearly impossible. This completes the proof of the lemma in the case (i). Case (ii) -- Z1 is a conic. If Z1 is a smooth conic, then the argument is as in the previous case: we compute the degree Z1 · (2KX + 2H + Z1) = Z1 · KX + 4 + Z1 · (KX + Z1) = Z1 · KX + 4 − 2 = Z1 · KX + 2 = −Z 2 1 and obtain h1(OZ1(2KX + 2H + Z1)) = 0, unless X is regular. Hence h1(OX (2KX + 2H + Z1)) ≤ h1(OX (2KX + 2H), with the conclusion, in view of Proposition 7.2, that the irregular surface X must be an elliptic scroll. But such a surface can not contain conics. If Z1 is singular, then Z1 = L1 + L2, the sum of two lines Li, i = 1, 2. The cohomology group H 1(OZ1(2KX + 2H + Z1)) fits into the following exact sequence H 1(OL1 (2KX +2H +L1)) → H 1(OZ1 (2KX +2H +Z1)) → H 1(OL2 (2KX +2H +Z1)) (7.10) and we continue as in the case (i). Namely, we compute the degree of the line bundle OL2(2KX + 2H + Z1): L2 · (2KX + 2H + Z1) = L2 · KX + L2 · L1 = −L2 2 + L2 · L1 − 2. This implies that h1(OL2(2KX + 2H + Z1)) = 0, unless L1 = L2 and L2 1 = 0. The latter condition means that X is a scroll and, by the proof of Proposition 7.2, see also Lemma 7.8, it must be an elliptic scroll of degree d = 5. Computing the second Chern number of NX from (7.1), we obtain 25 = d2 = (KX + 2H + Z1)(3H − Z1) = (KX + 2H + 2L1)(3H − 2L1) + deg(Z0) = 3d + 6 + deg(Z0) = 21 + deg(Z0). Hence deg(Z0) = 4 and from here on we repeat the argument form the proof of case (i). Thus we must have H 1(OL2 (2KX + 2H + Z1)) = 0. This and (7.10) imply h1(OZ1(2KX + 2H + Z1)) ≤ h1(OL1(2KX + 2H + L1)) which puts us back into the situation of the case of the line in (i). Case (iii) -- Z1 is a rational normal curve of degree 4. This case is very special since a rational normal curve C of degree 4 that lies in the singular locus of V3 forces the cubic hypersurface V3 to be the secant variety of C. Hence Sing(V3) = C and the Jacobian 40 7. Irregular surfaces on hypersurfaces of degree 3 ideal IV3 is equal to the ideal sheaf JC/P4 of C in P4. From this, the exact sequence (7.1) -- for a smooth surface X contained in V3 and passing through C -- becomes 0 −→ OX (2H + KX + C) −→ NX −→ OX (3H − C) −→ 0, It is now easy to see that X is regular. Indeed, the cohomological sequence controlling the irregularity q of X, H 1(OX (2H + 2KX + C)) −→ H 1(NX (KX )) −→ H 1(OX (3H − C + KX )), tells us that q = h1(NX(KX )) ≤ h1(OX (2H + 2KX + C)), since the line bundle OX(3H − C) is ample. The group H 1(OX (2H + 2KX + C)) is computed by the exact sequence H 1(OX (2H + 2KX )) −→ H 1(OX (2H + 2KX + C)) −→ H 1(OC(2H + 2KX + C)). The group on the right is zero unless C 2 ≥ 6 and this of course is impossible on an irregular surface. If h1(OC(2H + 2KX + C)) = 0, then h1(OX (2H + 2KX + C)) ≤ h1(OX (2H + 2KX )) and Proposition 7.2 tells us that if h1(OX (2H + 2KX )) 6= 0, then X must be a an elliptic scroll. But of course C can not lie on such a surface. Therefore, h1(OX (2H + 2KX )) and hence h1(OX (2H +2KX +C)) are both equal to zero. This completes the proof of the lemma. (cid:3) We know now that if V3 is not a cone and contains an irregular surface X, then the exact sequence (7.1) must have the form 0 −→ OX(KX + 2H) −→ NX −→ JZ(3H) −→ 0, (7.11) where Z is a 0-dimensional scheme, the intersection of X with the singular locus of V3. Hence, the cohomology sequence controlling the irregularity of X becomes H 1(OX (2KX + 2H) −→ H 1(NX (KX )) −→ H 1(JZ(KX + 3H)). Lemma 7.5. If X is irregular, then H 1(JZ (KX + 3H)) = 0. Proof. Assume H 1(JZ(KX + 3H)) 6= 0. Using our approach, we interpret this nonvan- ishing as a nontrivial extension sequence 0 −→ OX −→ E −→ JZ(3H) −→ 0. (7.12) Tensoring it with OX (−H), we obtain H 0(E(−H)) ∼= H 0(JZ (2H)). Combining this and Lemma 6.2 gives h0(E(−H)) ≥ 5. (7.13) From here on we proceed as in the proof of Lemma 7.3 to obtain a 'destabilizing' sequence 0 −→ OX (A) −→ E(−H) −→ JZ ′(H − A) −→ 0. (7.14) 41 7. Irregular surfaces on hypersurfaces of degree 3 Case h0(OX (A)) ≤ 2. In this case (7.13) implies h0(JZ ′(H − A)) ≥ 3 and we claim that A = 0. Indeed, if A 6= 0, then the previous inequality tells us that A is a line, JZ ′(H − A) = OX(H −A), and h0(OX (H −A)) = 3. From this, the exact sequence (7.14), and the inequality (7.13), we obtain 5 ≤ h0(OX (A)) + h0(OX (H − A)) = h0(OX (A)) + 3. Equivalently, h0(OX (A)) ≥ 2, i.e., a line on X moves in a linear system. But this is impossible on an irregular surface. Once we know that A = 0, the exact sequence (7.14) becomes 0 −→ OX −→ E(−H) −→ JZ ′(H) −→ 0. (7.15) Together with (7.13) gives h0(JZ ′(H)) ≥ 4. Hence either Z ′ = p, a single point, or Z ′ = 0. The first possibility is interpreted as p being a base point of OX (KX + H). Since d = H 2 ≥ 5, we can apply [29, Theorem 1] to deduce that X must be a scroll. Furthermore, by the proof of Proposition 7.2, X is an elliptic scroll of degree d = 5. This together with (7.11) tells us that 25 = d2 = c2(NX) = deg(Z) + 3d = deg(Z) + 15 or, equivalently, that deg(Z) = 10. However, from (7.15) and (7.12) it follows that 1 = deg(Z ′) = c2(E(−H)) = deg(Z) − 2H 2 = 10 − 2 · 5 = 0 which is absurd. We turn now to the second possibility, Z ′ = 0. The exact sequence (7.15) takes the form 0 −→ OX −→ E(−H) −→ OX (H) −→ 0 and this implies that E(−H) ∼= OX (H) ⊕ OX since Ext1(OX (H), OX ) ∼= H 1(OX (−H)) = 0. Geometrically, this means that Z is a complete intersection of two effective divisors H1 and H2 in H and 2H respectively. In particular, we obtain H 0(JZ(2H)) = {h1h + γh2 h ∈ H 0(OX (H)), γ ∈ C}, (7.16) where hi is a section corresponding to Hi, for i = 1, 2. Now, from the proof of Lemma 6.2, we recall that H 0(JZ(2H)) contains the 5-dimensional subspace H 0(JV3(2))X = {∂v(f )X v ∈ H 0(OX (H))∗} spanned by the restriction to X of the partial derivatives of f , a homogeneous polynomial defining V3. From this and the description of H 0(JZ(2H)) in (7.16) it follows that the intersection is a 4-dimensional vector space. This implies that we can choose homogeneous coordinate functions Xi, i = 0, . . . , 4, in P4 so that the partial derivatives have the form H 0(JV3(2))X \ h1H 0(OX (H)) ∂f ∂Xi = eh1Tj for 0 ≤ j ≤ 3 42 7. Irregular surfaces on hypersurfaces of degree 3 responding to h1. It follows that the remaining partial derivative ∂f ∂X4 where the Tj's are some linear forms and on H 0(OX (H))∗ and eh1 is the linear form cor- hypersurface Y which intersects the hyperplane (eh1 = 0) along a quadric surface Q con- tained in the singular locus of the cubic hypersurface V3. This means that the secant variety of Q is contained in V3. Since the latter is irreducible, it follows that Q is a double plane. But such a plane must intersect X along a 1-dimensional subscheme which is contrary to our assumption. defines a quadric Case h0(OX (A)) ≥ 3. We go back to the destabilizing sequence (7.14). We argue as in the proof of Lemma 7.3 to deduce that H 0(OX (A)) ֒→ H 0(E(−H)) is a proper subspace. Hence the divisor D = H − A is effective. This gives 3 ≤ h0(OX (A)) = h0(OX (H − D)). Thus D is either a line and the inequality above must be equality, or D = 0. The first possibility implies 5 ≤ h0(E(−H)) ≤ h0(OX (A)) + h0(JZ ′(D)) = 3 + h0(JZ ′(D)) with the conclusion that the line D moves in a linear system on X contradicting the hypothesis X irregular. The second possibility, D = 0, implies A = H. In this case the destabilizing sequence (7.14) takes the form 0 −→ OX(H) −→ E(−H) −→ OX −→ 0. Since the epimorphism above induces a surjection on the level of global sections, we deduce again that E(−H) ∼= OX(H) ⊕ OX , a situation discarded in the first part of the proof. (cid:3) Next we investigate the possibility of V3 being a cone over a cubic surface in P3. Lemma 7.6. If the cubic hypersurface V3 is a cone, then it contains no smooth irregular surface. Proof. We begin with the case of V3 being a cone with vertex at a point x0 over a cubic surface S which is not a cone. We go back to the sequence (7.1) and adapt our arguments thereafter to the case at hand. Claim. If Z1 6= 0, then H 1(JZ0(KX + 3H − Z1)) = 0. We proceed as in the proof of Lemma 7.3 by studying the extension sequence 0 −→ OX −→ E −→ JZ0(3H − Z1) −→ 0. As before we have the isomorphism H 0(E(−H)) ∼= H 0(JZ0(2H − Z1)) and from the proof of Lemma 6.2, it follows that This inequality and the isomorphism just above it give h0(JZ0(2H − Z1)) ≥ 4. h0(E(−H)) ≥ 4. This leads again to the exact sequence (7.6) 0 −→ OX (A) −→ E(−H) −→ JZ ′(H − Z1 − A) −→ 0, 43 7. Irregular surfaces on hypersurfaces of degree 3 where A is an effective divisor and H 0(JZ ′(H − Z1 − A)) 6= 0. Arguing as in the proof of Lemma 7.3 and using the assumption that the base of the cone V3 is not a cone, we obtain h0(OX (A)) = h0(JZ ′(H − Z1 − A)) = 2. (7.17) Set A0 (resp. A′) the fixed (resp. moving) part of A and consider OX (H − Z1 − A0). It is easy to see that h0(JZ ′(H − Z1 − A0)) = 3. From this and (7.17) it follows that A0 = 0, Z1 is a line, Z ′ = 0, and the linear map H 0(OX (A)) ⊗ H 0(OX (H − Z1 − A)) −→ H 0(OX (H − Z1)) has a nontrivial kernel. Hence there is a basis {x, x′} of H 0(OX (A)) and nonzero elements y, y′ ∈ H 0(OX (H − Z1 − A)) such that xy − x′y′ = 0 in H 0(OX (H − Z1)). Furthermore, xy − x′y′ can be viewed as the restriction of an element from Sym2 H 0(OX (H)). Since no quadric hypersurface contains X the above equality implies that y = λx′ and y′ = λx, for some nonzero λ ∈ C. In particular, we obtain OX (A) = OX(H − Z1 − A) or, equivalently, OX (H − Z1) = OX (2A). (7.18) Since h0(OX (2A)) = h0(OX (H − Z1)) = 3, we deduce Sym2 H 0(OX (A)) ∼= H 0(OX (2A)). Hence A is base point free and therefore, A2 = 0. This and (7.18) imply that the linear system 2A = H − Z1 is composed with a pencil, i.e., the morphism defined by H − Z1 factors as follows, ϕH−Z1 : X A −→ P1 OP1 (2) −→ P2. Equivalently, the projection from the line Z1 maps X onto a conic. But this means that X is contained in a quadric hypersurface, contrary to the hypothesis that 3 is the least degree of a hypersurface containing X. This completes the proof of the claim. Next we investigate the cohomology group H 1(OX (2KX +2H +Z1)) in the exact sequence (7.2). For this we need to control the divisor Z1. This divisor depends on the singularities of the cubic surface S, the base of the cone V3. The following two possibilities may occur: 1) S has isolated singular points and then Z1 is composed of one or two rulings of the cone V3; 2) the singular locus L of S is a line and then Z1 is a divisor contained in the plane P spanned by L and the vertex x0 of the cone V3. The first possibility is dealt with in the same way as in the proof of Lemma 7.4, so we turn to the second possibility. A hyperplane V passing through P intersects V3 along the decomposable surface V · V3 = 2P + PV , where PV is the residual plane. Hence the divisor HV = V · X has the form HV = 2Z1 + FV , 44 7. Irregular surfaces on hypersurfaces of degree 3 where FV is the divisor residual to 2Z1 and contained in the plane PV . As in the proof of Lemma 7.4, we relate H 1(OX (2KX + 2H + Z1)) to H 1(OX (2KX + 2H)) via the sequence H 1(OX (2KX + 2H)) −→ H 1(OX (2KX + 2H + Z1)) −→ H 1(OZ1 (2KX + 2H + Z1)) (7.19) and investigate the cohomology group on the right of this sequence. Since the divisor Z1 = P · X is the scheme-theoretic intersection of a plane with the surface X, a result of Ellia and Folegatti, [14], implies that Z1 is a reduced divisor. Hence, for an irreducible component C of Z1, we have h1(OC (2KX + 2H + Z1)) = h1(ωC ⊗ OC(KX + 2H + Z C 1 ) = h0(OC (−(KX + 2H) − Z C 1 ), where ωC is the dualizing sheaf of C, Z C 1 the component of Z1 complementary to C and the second equality is the Serre duality on C. Since OX (KX + 2H) is base point free, see [29], the above identity tells us that h1(OC (2KX + 2H + Z1)) = 0 unless Z1 = C and OC(KX + 2H) = OC . But then H · C = 1 and C 2 = 0. Hence X is a scroll and by Lemma 7.8, it is an elliptic scroll of degree d = 5. Since this possibility was ruled out in the proof of Lemma 7.4, we obtain H 1(OZ1(2KX + 2H + Z1)) = 0. This together with (7.19) imply that the nonvanishing of H 1(OX (2KX + 2H + Z1)) can only occur if H 1(OX (2KX + 2H)) 6= 0. This, in view of Proposition 7.2, implies that X is an elliptic scroll of degree d = 5 and Z1 is a ruling of X. Thus we are back in the situation ruled out in the proof of Lemma 7.4. Our considerations are now reduced to the case when the divisorial part Z1 in the exact sequence (7.1) is zero. Hence that sequence has the form 0 −→ OX (KX + 2H) −→ NX −→ JZ0(3H) −→ 0 and, as before, the irregularity of X is controlled by the groups H 1(OX (2KX + 2H)) and H 1(JZ0(3H + KX )). According to Proposition 7.2, the nonvanishing of H 1(OX (2KX + 2H)) may occur only if X is an elliptic scroll of degree d = 5. This degree is a numerical obstacle for X to be contained in a cubic cone. Indeed, take a general hyperplane V1 in P4 complementary to x0, the vertex of the cone V3, and consider the projection of X from x0 into V1. This gives rise to a (rational) map onto a cubic surface S := V1T V3. The degree of this map is as follows: fx0 : X − − → V1 deg(fx0) =( d if x0 /∈ X, if x0 ∈ X. 3 d−1 3 (7.20) (7.21) For d = 5 none of the above values is an integer. Hence the vanishing of H 1(OX (2KX +2H)). Claim. H 1(JZ0(3H + KX)) = 0. The proof of this assertion follows the same pattern as the one of Lemma 7.5. Namely, the nonvanishing of H 1(JZ0(3H + KX )) is interpreted as a nontrivial extension This implies 0 −→ OX −→ E −→ JZ0(3H) −→ 0. h0(E(−H)) = h0(JZ0(2H)) ≥ 4, (7.22) (7.23) 45 7. Irregular surfaces on hypersurfaces of degree 3 where the inequality comes from the assumption that the surface S, the base of the cone V3, is not a cone; see the proof of Lemma 6.2 for details. In particular, as in the proof of Lemma 7.5, we have an exact sequence 0 −→ OX (A) −→ E(−H) −→ JZ ′(H − A) −→ 0, (7.24) where A is an effective divisor and h0(JZ ′(H − A)) ≥ 1. Putting this sequence together with the sequence (7.22), we obtain the relation deg(Z0) − 2d = deg(Z ′) + A · (H − A). (7.25) We wish to understand the geometric ingredients -- the divisor A and the 0-dimensional sub- scheme Z ′ -- involved in this relation. To begin with, we observe that A is nonzero. Indeed, if A = 0, then the relation (7.25) reads deg(Z ′) = deg(Z0) − 2d. (7.26) To make use of this relation, as well as of (7.25) later on, we give an upper bound for deg(Z0). Namely, from the fact that the divisorial part Z1 is zero, it follows that the cubic surface S, the base of the cone V3, has only isolated singularities and that the map fx0 in (7.20) is finite outside x0. Furthermore, the subscheme Z0 is the scheme-theoretic intersection of X with the rulings of the cone V3 over the singular locus Sing(S) of S. Hence we get the upper bound deg(Z0) ≤ deg(fx0) deg(Sing(S)) ≤ deg(Sing(S)), d 3 where the last inequality comes from (7.21). In addition, from the classification of normal cu- bic surfaces that are not cones, see [12, p. 448], it follows that deg(Sing(S)) ≤ 6. Substituting this estimate into the above inequality, we obtain deg(Z0) ≤ 2d. (7.27) This and the relation (7.26) imply deg(Z ′) = 0. Hence the exact sequence (7.24) takes the form 0 −→ OX −→ E(−H) −→ OX (H) −→ 0, situation we have already considered in the proof of Lemma 7.5. We know now that the divisor A in (7.24) must be nonzero. Assume h0(OX (A)) = 1. From this and (7.23) it follows that h0(JZ ′(H − A)) ≥ 3 implying that A is a line, the above inequality is an equality, and Z ′ = 0. This and (7.25) imply deg(Z0) − 2d = A · (H − A) = 1 − A2, which, together with the upper bound (7.27), implies A2 ≥ 1. Hence X is regular (one sees easily that X is a rational surface). Thus we may assume h0(OX (A)) ≥ 2. In fact the equality must hold, since, if h0(OX (A)) ≥ 3, then the divisor L = H − A is a line and the relation (7.25) reads deg(Z0) − 2d = deg(Z ′) + (H − L) · L = deg(Z ′) + 1 − L2. 46 7. Irregular surfaces on hypersurfaces of degree 3 This and the upper bound (7.27) imply L2 ≥ 1, the situation we have already discarded. From h0(OX (A)) = 2 and the inequality (7.23) we deduce h0(JZ ′(H − A)) ≥ 2. Furthermore, if the inequality is strict, we arrive again at the situation ruled out previously: A is a line moving in a linear system. Thus h0(OX (A)) = h0(JZ ′(H − A)) = 2, a situation we have already encountered in the first part of the proof. Arguing as there, we show that the linear system A has at most a 0-dimensional base locus. Hence A·(H −A) ≥ 0. This, the relation (7.25), and the upper bound (7.27) imply that all inequalities involved must be equalities. In particular, A · (H − A) = 0. Since both A and (H − A) are effective, nonzero divisors that add up to H, a very ample divisor, the above identity is impossible. This completes the proof of the lemma in the case the surface S, the base of the cone V3, is not a cone. We now turn to the case when S is a cone with vertex xS over a plane cubic curve C. If C is smooth, then V3 is a singular scroll ruled by the planes Pt = Span(tS L), for t ∈ C, with the singular locus Sing(V3) being the line L = Span(x0S xS). This is analogous to the situation we arrived at in the proof of Theorem 6.1. Hence the sequence (7.1) takes the form13 0 −→ OX (KX + 2H + L) −→ NX −→ OX (3H − L) −→ 0. It follows that the irregularity of X is controlled by the cohomology group H 1(OX (2KX + 2H + L)). Its nonvanishing, as we have seen on several occasions, may occur only if X is an elliptic scroll of degree d = 5 and L is a ruling of the scroll. But computing the second Chern classes from the above sequence, we obtain 25 = d2 = (KX + 2H + L) · (3H − L) = 18, an obvious contradiction. If C is singular, let c0 be its (unique) singular point. The plane Pc0 = Span(c0S L) is the singular plane of V3. Furthermore, the planes Pt = Span(tS L) give rise to a rational family of curves {Ft = Pt · X t ∈ C}. Let F be the divisor class of this family. Then by taking a general hyperplane in P4 passing through the line L, we have OX(H) = OX (3F ). From h0(OX (F )) = h0(OX (H − 2F )) it follows that h0(OX (F )) = 2. Since the linear map is injective, we deduce Sym2(H 0(OX (F ))) −→ H 0(OX (2F )) 3 ≤ h0(OX (2F )) = h0(OX (H − F )) and hence F is a line. Since F moves in a pencil the surface X must be rational. This completes the proof of the lemma. (cid:3) 13We consider here only the case L ⊂ X, since the other case, L 6⊂ X, is treated in exactly the same way as in the proof of Theorem 6.1. 47 7. Irregular surfaces on hypersurfaces of degree 3 We know now that an irregular X contained in a cubic hypersurface must be an elliptic scroll of degree 5. Furthermore, we have the following. Lemma 7.7. Let X be an elliptic scroll of degree d = 5 in P4 and let V3 be a cubic 10. hypersurface containing it. Then the scheme Z = XT Sing(V3) is 0-dimensional of degree with Lemma 7.4, we deduce that the scheme Z = XT Sing(V3) is 0-dimensional. Hence the Proof. From the proof of Lemma 7.6 we know that V3 is not a cone. Combining this sequence (7.1) has the form 0 −→ OX(KX + 2H) −→ NX −→ JZ(3H) −→ 0. Computing the second Chern classes from this sequence, we obtain deg(Z) = deg(cid:16)X\ Sing(V3)(cid:17) = 10. (cid:3) To make the paper self-contained, we include the following result which is a well-known part of the classification of surfaces in P4, see [22]. Lemma 7.8. The only irrational scrolls in P4 are elliptic scrolls of degree 5. Proof. Let X be a P1-bundle over a smooth connected curve B of genus q ≥ 1 and let OX(H) be a very ample line bundle on X defining an embedding of X into P4 as a scroll. In particular, a smooth divisor in the linear system H is isomorphic to B. This together with the adjunction formula implies H · KX + H 2 = H · KX + d = 2q − 2. Substituting this and the Chern numbers χ(OX ) = 1 − q, K 2 point formula, we obtain X = 8(1 − q) into the double d2 − 5d = 6(q − 1). (7.28) The main point of the argument is to compare the genus q of B in the above formula with the Castelnuovo upper bound on the genus of a smooth curve in a projective space. Namely, the scroll X ⊂ P4 is interpreted as an embedding ϕ : B −→ Gr(1, P4) into the Grassmannian Gr(1, P4) of lines in P4. Setting G to be the pullback under ϕ of the universal subbundle of the Grassmannian, we identify X with the projectivization P(G). Then OX(H), the line bundle embedding X into P4, is such that the direct image π∗(OX (H)) ∼= G∗. In particular, deg(c1(G∗)) = d. Composing ϕ with the Plucker embedding of Gr(1, P4) gives the embedding realized by the subsystem of 2G∗ corresponding to the image of the obvious homomorphism ψ : B ֒→ P9 ρ : 2H 0(G∗) −→ H 0( 2G∗). 48 7. Irregular surfaces on hypersurfaces of degree 3 We claim that ker(ρ) has dimension at least 5. Indeed, assume dim(ker(ρ)) ≤ 4. Then the image of ρ has dimension N + 1 ≥ 6 and ψ embeds B into PN as a nondegenerate curve of degree d. The Castelnuovo bound on the genus q in PN gives q − 1 ≤ sd − s(s + 1)(N − 1) 2 − N − 1, where d = s(N − 1) + r for an integer 0 ≤ r < N − 1. Rewriting the expression on the right as a function of d, r, and N , we obtain q − 1 ≤ 1 2(N − 1) (d2 − r2) − 1 2 d + r 2 − (N + 1), In our situation N ∈ {5, . . . , 9}, hence we can consider the larger bound Putting it together with (7.28) gives q − 1 ≤ d2 − 5d ≤ 1 8 3 4 d2 − 1 2 d − r2 16 + r 2 − 6. d2 − 3d − 3r2 8 + 3r − 36. This can be rewritten in the form (cid:18) 1 2 which is false. d − 2(cid:19)2 = 1 4 d2 − 2d + 4 ≤ − 3r2 8 + 3r − 32 ≤ −26 We know now that the kernel of ρ is at least of dimension 5. Geometrically this means that the projectivized subspace P(ker(ρ)) intersects the Grassmannian variety Gr(2, H 0(G)) ∼= Gr(1, P4) of decomposable tensors in P( 2H 0(G∗)) ∼= P9 along a subscheme of dimension at 2G∗, is least 1. Each decomposable tensor g ∧ g′ in this intersection, viewed as a section of zero. Equivalently, the two sections g and g′ correspond, under the isomorphism H 0(G∗) ∼= H 0(OX (H)), to two hyperplanes Hg and Hg′ intersects the scroll X along a curve Γg,g′ which is a section of the structure projection π : X → B. The above shows that there is a family {Γt}t∈T of such sections parametrized by an irreducible curve T . in P4 such that the plane P = HgT Hg′ Observe that every two planes Pt and Pt′, for t 6= t′, intersect at a single point. This implies that Γt · Γt′ ≤ 1. We easily check that Γt · Γt′ = 1. In particular, setting Γ to be the class of {Γt}t∈T in the N´eron-Severi group of X, we obtain Γ2 = 1. We write H = Γ + aF , where F is the class of a ruling of X. Then d = H 2 = 2a + 1 and the degree of Γ is dΓ = H · Γ = d − a = 2a + 1 − a = a + 1. Since Γ is a plane curve, we have This, the identity d = 2a + 1 and (7.28) imply 2(q − 1) = (a + 1)(a − 2). 3(a + 1)(a − 2) = 6(q − 1) = d2 − 5d = (2a + 1)2 − 5(2a + 1) = (2a + 1)(2a − 4) or, equivalently, (a − 2)(a − 1) = 0. This leads to two solutions d = 5 and 3, which are, respectively, an elliptic scroll of degree 5 and a rational scroll of degree 3. (cid:3) 49 8. On the elliptic scroll of degree 5 and the Segre cubic 8. On the elliptic scroll of degree 5 and the Segre cubic In the previous section we have characterized an elliptic scroll X of degree 5 in P4 as being the only irregular surface lying on a cubic hypersurface. Such scrolls are notorious and they have been subject to extensive study; see, e.g., [18, 4, 5] and the references therein. The main objective of this section is to (re)establish a relationship between two entities related to the embedding of X in P4: • the space IX (3) of cubic hypersurfaces in P4 containing X • the space of global sections H 0(N ∗ X(3H)) of the conormal bundle N ∗ by OX (3H), where OX (H) is a line bundle realizing the embedding of X in P4. X of X in P4 twisted Using the above notation, we formulate the main result of this section. Theorem 8.1. Let X be an elliptic scroll of degree 5. 1) The sheaf N ∗ X(3H) is a rank 2 vector bundle generated by its global sections, with Chern invariants c1(N ∗ X(3H)) = H − KX and c2(N ∗ 2) There is a natural isomorphism H 0(N ∗ 3) Every nonzero global section s of N ∗ X(3H)) = 10. X (3H)) ∼= IX (3) ∼= C5. X(3H) has 0-dimensional zero locus of degree 10. Under the above correspondence, the scheme of zeros Zs = (s = 0) is the scheme- theoretic intersection of X with Sing(V3(s)), the singular locus of the cubic hypersurface V3(s) ∈ IX(3) corresponding to s. In particular, every global section s with Zs = (s = 0) consisting of ten distinct points, corresponds to a Segre cubic V3(s), whose set of nodes Sing(V3(s)) = Zs. Proof. The assertion about the Chern invariants is obvious. Of course the relation between X(3H) and cubic hypersurfaces through X has been at the origin of our X, where JX is the global sections of N ∗ considerations in Section 7 and stems from the identification N ∗ ideal sheaf of X in P4. From the exact sequence X = JX/J 2 0 −→ J 2 X(3) −→ JX(3) −→ N ∗ X(3H) −→ 0 follows the inclusion14 0 −→ IX(3) = H 0(JX(3)) −→ H 0(N ∗ X(3H)). (8.1) From the exact sequence 0 → JX → OP4 → OX → 0 tensored with OP4(3) we obtain the estimate h0(JX (3)) ≥ h0(OP4(3)) − h0(OX (3H)) =(cid:18)7 3(cid:19) − 1 2 (9H 2 − 3H · KX) = 35 − 30 = 5, (8.2) where we used the easily verified property H i(OX (kH)) = 0 for all k > 0 and i = 1, 2. The above estimate is actually an equality.15 This as well as the isomorphism in (8.1), and hence the assertion 2) of the theorem, follow from the next lemma. 14The inclusion comes from H 0(J 2 X(3)) = 0, since we know that X is not contained in a quadric hypersurface, see Theorem 6.1. 15One easily verifies that X ⊂ P4 is a projectively normal embedding. Since we do not use this aspect anywhere, the above mentioned equality is established differently in the proof of Lemma 8.2. 50 8. On the elliptic scroll of degree 5 and the Segre cubic Lemma 8.2. The vector bundle N ∗ X(3H) is globally generated and h0(N ∗ X(3H)) = 5. Proof. Let π : X → E be the structure morphism, i.e., E is an elliptic curve and π is a P1-fibration over E. The line bundle OX (H) defining the embedding of X into P4 as a scroll has degree 1 on the fibres of π, i.e., OX(H) ⊗ OF = OF (1) ∼= OP1(1) on every fibre F of π. We wish to understand the restriction of N ∗ X(3H) to a fibre F . Since det(NX ) = OX (5H + KX), we have From this it follows that det(NX) ⊗ OF = OF (5H + KX ) = OF (3). N ∗ X(3H) ⊗ OF = N ∗ X ⊗ OF (3) = N ∗ X ⊗ det(NX ) ⊗ OF ∼= NX ⊗ OF ∼= OF (1) ⊕ OF (2), (8.3) where the last isomorphism follows from the global generation of NX(−H) ⊗ OF = NX ⊗ OF (−1). In particular, the restriction N ∗ X(3H) ⊗ OF is globally generated on every fibre F of π. So to obtain the global generation of N ∗ X(3H), it is sufficient to show the surjectivity H 0(N ∗ X(3H)) −→ H 0(N ∗ X(3H) ⊗ OF ) = H 0(OF (1) ⊕ OF (2)) for every fibre F of π. For this we use the inclusion (8.1) to obtain the composition 0 IX(3) (8.4) 0 H 0(N ∗ X(3H − F )) H 0(N ∗ X (3H)) H 0(N ∗ X (3H) ⊗ OF ) 0 In addition, from the proof of Theorem 7.1, we know that the sections of N ∗ from IX (3) have 0-dimensional schemes of zeros. Hence, the image of IX(3) in H 0(N ∗ is complementary to the kernel of the restriction homomorphism X(3H) coming X(3H)) H 0(N ∗ X(3H)) −→ H 0(N ∗ X(3H) ⊗ OF ) in (8.4). From this it follows that the slanted arrow in (8.4) is injective. Since the target of that arrow is a space of dimension 5, see (8.3), we deduce the inequality dim(IX (3)) ≤ 5. This and the estimate (8.2) imply dim(IX (3)) = 5. (8.5) Therefore, the slanted arrow in (8.4) is an isomorphism, hence the global generation of N ∗ X(3H). h0(N ∗ X (3H)) ≥ 5. Let us assume that the inequality is strict. From the Koszul sequence We now turn to the assertion h0(N ∗ X (3H)) = 5. From (8.4), we already know that 0 −→ OX s−→ N ∗ X(3H) ∧s−→ JZs(H − KX) −→ 0 (8.6) 51 8. On the elliptic scroll of degree 5 and the Segre cubic of a general global section s of N ∗ X(3H), we have h0(JZs(H − KX)) ≥ h0(N ∗ X(3H)) − 1 ≥ 5. (8.7) Furthermore, considering another general global section of N ∗ curve Γ = (γ = 0), where γ is the section of det(N ∗ s ∧ s′ under the natural homomorphism X(3H), we obtain the smooth X (3H)) = OX (H − KX) corresponding to 2H 0(N ∗ X(3H)) −→ H 0(det(N ∗ X (3H))) = H 0(OX (H − KX)). This gives rise to the exact sequence 0 −→ OX γ −→ JZs(H − KX ) −→ OΓ((H − KX)Γ − Zs) −→ 0. From this and (8.7) we obtain h0(OΓ((H − KX)Γ − Zs)) ≥ h0(JZs(H − KX )) − 1 ≥ 5 − 1 = 4. (8.8) On the other hand, the degree of OΓ((H − KX )Γ − Zs) is (H − KX) · Γ − deg(Zs) = (H − KX )2 − c2(N ∗ X (3H)) = 15 − 10 = 5, while the genus of Γ, by the adjunction formula, is 6. This implies that OΓ((H − KX)Γ − Zs) is special and by the Clifford inequality h0(OΓ((H − KX)Γ − Zs)) ≤ deg((H − KX )Γ − Zs) 2 + 1 = 5 2 + 1. Combining this inequality with (8.8), we obtain 4 ≤ h0(OΓ((H − KX)Γ − Z)) ≤ 5 2 + 1, an obvious contradiction. (cid:3) End of the proof of Theorem 8.1. X(3H)) and IX(3) are both 5-dimensional, see (8.5) for the latter, an immediate corollary of Lemma 8.2 is that the inclusion (8.1) is an isomorphism. Since the spaces H 0(N ∗ Once the isomorphism in Theorem 8.1, 2) is established, we also deduce that every nonzero X(3H) has 0-dimensional scheme of zeros, since this is now equivalent to X(3H) with 0-dimensional zero locus, global section of N ∗ the property of cubics in IX(3) to produce sections of N ∗ see Lemma 7.7. This proves the first assertion of Theorem 8.1, 3). The degree of the scheme of zeros of a nonzero section of N ∗ X(3H) is the value of the X(3H)) and this value is 10 by the part second Chern class (identified with its degree) c2(N ∗ 1) of the theorem. Next we turn to the assertion that Zs, the zero scheme of a nonzero global section s of N ∗ X(3H), is the scheme theoretic intersection of X with the singular locus of the cubic hypersurface V3(s) corresponding to s under the isomorphism in part 2) of the theorem. From the general discussion about the relation of the ideal sheaf JZs of Zs and the restriction of the Jacobian ideal IV3(s) to X we know that JZs ⊃ IV3(s) ⊗ OX , 52 8. On the elliptic scroll of degree 5 and the Segre cubic i.e., Zs is contained in the scheme theoretic intersection of X with the singular locus of the cubic hypersurface V3(s). To see the equality, it is enough to show that the generators of IV3(s), the partial derivatives of a homogeneous polynomial defining V3(s), restricted to X, generate the sheaf JZs(2H). This follows immediately from the epimorphism NX(−H) −→ JZs(2H) coming from the Koszul sequence (8.6) tensored with OX (H + KX ) and the surjective mor- phism H 0(OX (H))∗ ⊗ OX −→ NX(−H). The resulting composition H 0(OX (H))∗ ⊗ OX −→ JZs(2H) is surjective and is described explicitly by the partial derivatives of a homogeneous polynomial defining V3(s), see the proof of Lemma 6.2 for details. Hence the asserted equality JZs = IV3(s) ⊗ OX . (8.9) We are left with the last assertion of 3) of the theorem, stating that sections of N ∗ X(3H) with simple zeros correspond to Segre cubics in IX (3). Indeed, let s be a global section of N ∗ X(3H) with Zs = (s = 0) consisting of ten distinct points. From the equality (8.9), we deduce that the singular locus Sing(V3(s)) of the cubic V3(s) contains ten distinct points. It will be enough to show that Sing(V3(s)) is 0-dimensional, since then (8.9) tells us that the singular locus Sing(V3(s)) = Zs and it is composed of ten ordinary double points. It is well known that such a cubic hypersurface is a Segre cubic (see [11] for an inspiring introduction to the subject). Let us check now that Sing(V3(s)) is 0-dimensional. By Lemma 7.6, V3(s) is not a cone. Then the possibilities for the one dimensional part of Sing(V3(s)) are a line, a conic (possibly singular), or a rational normal curve of degree 4 in P4. If a conic C is a component of Sing(V3(s)), then the plane P spanned by C is contained in V3. We examine the pencil of hyperplanes Vt in P4 passing through P . The intersection Vt · V3(s) is reducible Vt · V3(s) = P ∪ Qt, ∀t, where Qt is a quadric surface residual to the plane P such that Qt ∩ P = C. The hyperplane section Ht = Vt · X = B + Γt (8.10) is also reducible, where B = P · X is the 1-dimensional part of the base locus of the pencil {Ht}. Being a plane divisor, B can be either a ruling of X or its plane cubic section. The latter case implies B · C = 6. So the part ZC of the scheme Zs contained in the intersection B ∩ C has degree at least 6. On the other hand the degree of the subscheme ZB of Zs contained in B is at most B · (H − KX ) = 4 (we use here the fact that Zs is contained in an irreducible divisor in the linear system H − KX). Hence B must be a ruling of X. From this and (8.10) it follows that for a general t, the curve Γt is an elliptic curve of degree 4 contained in Qt. In particular, the scheme-theoretic intersection C ∩ Γt is a 0-dimensional scheme of degree 4. Furthermore, since C is not contained in X, this scheme must be contained in the base locus of the pencil {Γt}. Thus we obtain 4 ≤ Γ2 t = (H − B)2 = 3 53 8. On the elliptic scroll of degree 5 and the Segre cubic a contradiction. If a rational normal curve R of degree 4 is in Sing(V3(s)), then Sing(V3(s)) = R and Zs = X · R. It follows that h0(JZs(2H)) ≥ h0(JR/P4(2)) = 6, (8.11) where JR/P4 is the ideal sheaf of R in P4. Since Zs lies on a smooth curve Γ ∈ H − KX , we can calculate h0(JZs(2H)) from the exact sequence 0 −→ OX (−Γ) −→ JZs −→ OΓ(−Zs) −→ 0 tensored with OX (2H). This gives h0(JZs(2H)) = h0(OΓ(2HΓ − Zs)) = 5 + h1(OΓ(2HΓ − Zs)) = 5 + h0(OΓ(Zs − HΓ). This and (8.11) imply h0(OΓ(Zs − HΓ) = 1 or, equivalently, OΓ(Zs) = OΓ(H) and this contradicts Corollary 8.5 below. We turn now to the remaining case: the 1-dimensional locus of Sing(V3(s)) is a line L. We divide the scheme Zs into two parts, Zs = ZL + Z ′, where ZL is the part of Zs contained in L and Z ′ = Zs \ ZL is the residual part. It is easy to see that deg(ZL) ≤ 3. But then Z ′ consists of at least 7 distinct points. For every point z′ ∈ Z ′ the plane Pz′ = Span(z′ ∪ L) is contained in V3(s). Furthermore, Pz′ for any pair of distinct points z′ 2} intersects L and hence, is a component of the singular locus of V3(s). 2 ∈ Z ′ since otherwise the line L′ = Span {z′ 6= Pz′ 1, z′ 1, z′ 1 2 Every plane Pz′ intersects X along a plane curve. This curve is either a ruling of X or its plane cubic section. Assume there is z′ · X = Γ is a plane cubic section. Then Γ ∩ L = ZL and all other planes Pz′, z′ 6= z′ 0, intersect X along a ruling. Since those rulings must pass through one of the points of ZL, the number of such planes is at most 3. This makes the degree of Z ′ at most 4. This is contrary to the estimate deg(Z ′) ≥ 7. Hence every plane Pz′ intersects X along a ruling. But then 3 ≥ deg(ZL) ≥ deg(Z ′) ≥ 7 which is impossible. This completes the proof of the theorem. (cid:3) 0 ∈ Z ′ such that that Pz′ 0 Remark 8.3. From the proof of Theorem 8.1, 3), it follows that if a cubic hypersurface V3(s) contains X and has 1-dimensional singular locus, then its 1-dimensional part must be a single line. Furthermore, if this possibility occurs, the global section s of N ∗(3H) corresponding to V3(s) under the isomorphism in Theorem 8.1, 2), must have multiple zeros. In the appendix, see (A.17) and the discussion preceding it, we give an explicit geometric construction of a general cubic in IX(3), singular along a line. Hence, that line is precisely the 1-dimensional part of the singular locus of such a cubic. In addition, we show that the isomorphism in Theorem 8.1, 2), matches precisely the global sections of N ∗(3H) having multiple zeros with the cubics in IX (3) having a (unique) line in their singular locus, see Proposition A.10. In the course of the proof of Theorem 8.1, 3), we used the fact that on a smooth curve Γ ∈ H − KX containing Zs = (s = 0), the zero scheme of a nonzero global section s of N ∗(3H), the line bundles OΓ(Zs) and OΓ(H) are not isomorphic; see Corollary 8.5. This is a part of the proof of the following. 54 8. On the elliptic scroll of degree 5 and the Segre cubic Lemma 8.4. For every nonzero global section s ∈ H 0(N ∗ X(3H)) with Zs = (s = 0) one has h0(JZs(2H)) = 5 and h1(JZs(2H)) = 0. Proof. By Theorem 8.1, 3), the subscheme Zs = (s = 0) is 0-dimensional. Hence the Koszul sequence of s 0 −→ OX s−→ N ∗ X(3H) s∧−→ JZs(H − KX) −→ 0 is exact. Tensoring it with OX(KX + H) and using the identification N ∗ NX(−H) we obtain X(KX + 4H) ∼= 0 −→ OX (KX + H) s−→ NX(−H) s∧−→ JZs(2H) −→ 0. From this it follows that h0(JZs(2H)) = h0(NX (−H)) is independent of the choice of s. So to compute h0(JZs(2H)) we choose s with simple zeros and s′ ∈ H 0(N ∗ X (3H)) so that the curve Γ = (s ∧ s′ = 0) ∈ H − KX is smooth. The curve Γ passes through Zs and gives the following exact sequence 0 −→ OX(−Γ) γ −→ JZs −→ OΓ(−Zs) −→ 0, (8.12) where γ is the global section of OX(H − KX ) corresponding to s ∧ s′ under the natural homomorphism 2H 0(N ∗ X (3H))) = H 0(OX (H − KX )). Tensoring the above sequence with OX (2H), we deduce X(3H)) → H 0(det(N ∗ h0(JZs(2H)) = h0(OΓ(2HΓ − Zs)) = 5 + h1(OΓ(2HΓ − Zs)) = 5 + h0(OΓ(Zs − HΓ)), where the second equality is the Riemann-Roch for OΓ(2HΓ − Zs) and the third one is the Serre duality. Thus the first assertion of the lemma is equivalent to OΓ(Zs) 6= OΓ(H). (8.13) Assume the contrary. Then the exact sequence (8.12) tensored with OX (H − KX ) takes the form γ −→ JZs(H − KX ) −→ OΓ(−KX) −→ 0. 0 −→ OX This implies h0(OΓ(−KX )) ≥ h0(JZs(H − KX )) − 1 ≥ h0(N ∗ X (3H)) − 2 = 3. (8.14) On the other hand we have 0 −→ OX(−KX − Γ) −→ OX (−KX ) −→ OΓ(−KX ) −→ 0. Since OX (−KX −Γ) = OX (−H), the above implies H 0(OΓ(−KX )) = H 0(OX (−KX )). How- ever, we know that the last space is 1-dimensional. Hence h0(OΓ(−KX)) = 1 contradicting the estimate in (8.14). The second assertion of the lemma about the vanishing of H 1(JZs(2H)) follows immedi- (cid:3) ately from the first assertion and the Riemann-Roch for JZs(2H). Corollary 8.5. Let s be a nonzero global section of N ∗ X(3H) whose zero locus Zs = (s = 0) is contained in a smooth curve Γ ∈ H − KX. Then the line bundle OΓ(HΓ − Zs) 6= OΓ. Proof. The assertion is a restatement of the identity (8.13) proved in the previous lemma. (cid:3) 55 9. Irregular surfaces on hypersurfaces of degree 4 with non-degenerate isolated singularities 9. Irregular surfaces on hypersurfaces of degree 4 with non-degenerate isolated singularities In this section we consider irregular surfaces X ⊂ P4 contained in a hypersurface of degree 4 and not in one of a smaller degree. Our main result is as follows. Theorem 9.1. Let X ⊂ P4 be a smooth surface with mX = 4 and assume X to be contained in a quartic hypersurface V4 with at most ordinary double points. Then X is regular, with the possible exception of X being a degree 8 elliptic conic bundle with H ·KX = 0 and K 2 X = −8. If such a situation occurs, then X must pass through precisely 32 singular points of V4. The exceptional possibility in the above theorem is an elliptic conic bundle discovered about 20 years ago by Abo, Decker, and Sasakura by using a certain vector bundle of rank 5 on P4, see [1]. Shortly afterwards, Ranestad in [28], gave a geometric construction of a general elliptic conic bundle in P4 as the image of an elliptic scroll in P4 under a certain Cremona transformation. In the sequel we refer to those elliptic conic bundles as ADSR elliptic conic bundle. From the first work cited above one knows that the space of quartics H 0(JX(4)) containing X is of dimension 6. However, at the time of writing this paper, we do not know if there are quartics in H 0(JX(4)) with only ordinary double points. Our proof of Theorem 9.1 follows the same line of thinking as in the case of surfaces contained in hypersurfaces of degree 3. Namely, we assume X ⊂ P4 to be an irregular surfaces with mX = 4 and lying on a quartic hypersurface V4 with only ordinary double points. Our general situation recorded by the sequence (5.1) takes the form 0 −→ OX (KX + H) −→ NX −→ JZ (4H) −→ 0, (9.1) where Z is the 0-dimensional subscheme of X supported on the singular locus of V4 and defined, at each point p of the support of Z, by the restriction to X of the Jacobian ideal IV4,p. In particular, we have16 JZ = IV4 ⊗ OX , (9.2) where IV4 denotes the sheaf of the Jacobian ideal of V4. Expressing the second Chern class of NX from the exact sequence (9.1) provides a new "double point formula" d2 = deg(Z) + 4H · (H + KX ) = deg(Z) + 8(g − 1), (9.3) where g = g(H) is the geometric genus of a general hyperplane section. In the sequel we refer to this identity as the ndp formula. The cohomological sequence (5.3) controlling the irregularity of X becomes H 1(OX (2KX + H)) −→ NX(KX) −→ H 1(JZ (KX + 4H)). (9.4) Therefore, an understanding of the irregularity is reduced to the study of the cohomology groups H 1(OX (2KX + H)) and H 1(JZ (4H + KX )). In particular, we need to control the 16The identity (9.2) is valid as long as a section defining the Koszul sequence (9.1) has a 0-dimensional scheme of zeros. 56 9. Irregular surfaces on hypersurfaces of degree 4 with non-degenerate isolated singularities scheme Z which, in view of the identity (9.2), comes down to controlling the singular locus Sing(V4) of V4. Under our assumption on the isolated singularities of V4, the singular locus Sing(V4) is the set of ordinary double points of V4 and we can quote a result of A. Varchenko, [35], for the estimate deg(Sing(V4)) ≤ 45. This together with (9.2) gives deg(Z) ≤ deg(Sing(V4)) ≤ 45. (9.5) With this estimate of deg(Z) recorded, we turn now to the study of the cohomology groups H 1(OX (2KX + H)) and H 1(JZ (KX + 4H)) in (9.4). 9.1. The study of H 1(OX (2KX + H)) By the Serre duality, H 1(OX (2KX + H))∗ = H 1(OX (−(KX + H))). Thus the question of the (non)vanishing of this group comes down to understanding the geometric properties of the divisor KX + H. The next lemma is an easy consequence of [29]. Lemma 9.2. Let X be an irregular surface and H be a very ample divisor on X. Then the following assertions hold. 1) OX (KX + H) has base points if and only if X is a ruled surface and its embedding by OX (H) is a scroll. 2) If OX (KX +H) is base point free and H 1(OX (−(KX +H))) 6= 0, then X is a birationally ruled surface embedded by OX (H) as a conic bundle over a smooth curve B of genus q = q(X). Proof. The assumption that X is irregular implies H 2 ≥ 5. Then, by [29, Theorem 1,(i)], a base point of OX (KX + H) gives rise to an effective divisor D ⊂ X passing through a base point of KX +H such that H ·D = 1 and D2 = 0. It follows that D is a line in the embedding given by OX (H). Hence the Albanese map a : X → Alb(X) must contract D to a point. The fact that D2 = 0 implies that the map a factors through a smooth curve B ⊂ Alb(X) of genus q = q(X) and a : X → B is a P1-fibration, with D one of the fibres. Thus X is a ruled surface embedded by OX (H) as a scroll. The assertion in the other direction is obvious. We turn now to the assertion 2) of the lemma. The hypotheses imply that OX(KX + H) is nef but not big, i.e., that (KX +H)2 = 0. Since OX (KX +H) 6= OX (otherwise OX(KX ) = OX(−H) and hence q = h1(OX ) = h1(OX (KX)) = h1(OX (−H)) = 0), the linear system KX + H induces a morphism whose image is a curve. More precisely, there is a morphism π : X −→ B (9.6) onto a smooth curve B with connected fibres and a base point free line bundle OB(D) on B such that O(KX + H) = π∗(OB(D)). From this we obtain the relation KX + H = deg(D)F in NS(X), where F stands for the class of a fibre of π. Intersecting with F the above identity, we deduce that H · F = 2, i.e., that π in (9.6) is a conic fibration. Thus a general fibre of π is a smooth conic and there is at most a finite number of singular fibres. A priori, a singular fibre is either the union of two lines intersecting transversely or a double line. The latter, however, is impossible since F = 2L with L a line leads to KX · L = L2 = −1 which contradicts 0 = F 2 = 4L2. (cid:3) 57 9. Irregular surfaces on hypersurfaces of degree 4 with non-degenerate isolated singularities We apply the above result to a surface X subject to the hypotheses of Theorem 9.1 Proposition 9.3. Let X and V4 be as in Theorem 9.1 and assume X to be irregular. Then OX(KX + H) is base point free. Furthermore, OX(KX + H) is big and hence H 1(OX (2KX + H)) = 0 with the possible exception of X being an ADSR elliptic conic bundle. If such a situation occurs, then X must pass through precisely 32 singular points of V4. Proof. By Serre duality H 1(OX (2KX + H)) ∼= H 1(OX (−(KX + H)))∗. According to Lemma 9.2, the latter group is nonzero if X is embedded, either as a scroll, or as a conic bundle. The first possibility implies that X is an elliptic scroll of degree 5, see Lemma 7.8. But such a scroll, as we have seen in the previous section, is contained in hypersurfaces of degree 3, hence it can not occur here. We turn to the second possibility: X is birational to a ruled surface embedded into P4 by OX (H) as a conic bundle. More precisely, from the proof of Lemma 9.2, 2), the line bundle O(KX + H) induces a morphism π : X → B onto a smooth curve B of genus q = q(X), the irregularity of X, such that the H-degree of the fibres of π is 2. A general fibre of π is a smooth plane conic, while the singular fibres are reduced singular conics. If X is not minimal, then π factors trough a minimal model of X, call it X′, and gives the diagram X π X′ π′ σ B where σ is the blow-down of a collection of (−1)-curves on X and X′ is a ruled surface over B with π′ its structure morphism. The collection of exceptional curves is a choice of one out of the two irreducible components of each reducible fibre of π -- these are the only (−1)-curves of X. This implies, in particular, that X is the blow-up of X′ at distinct points. Let δ be the number of blown-up points and let {Cj} be the collection of the exceptional (−1)-curves on X. Then we write KX = σ∗(KX ′ ) + ∆, where KX ′ is the canonical divisor of X′ and ∆ = Pδ (−1)-curves blown-down by σ. This implies j=1 Cj is the sum of the exceptional K 2 X = K 2 X ′ − δ = −8(q − 1) − δ and c2 = −4(q − 1) + δ. (9.7) Since OX (H + KX ) is composed of a pencil, we also have (H + KX)2 = 0, hence From this and the first identity in (9.7), it follows that K 2 X = −d − 2H · KX. H · KX = 4(q − 1) − d − δ 2 . (9.8) Substituting this and the Chern numbers computed in (9.7) into the double point formula, we deduce the identity d2 − 15 2 d − δ = 16(q − 1). (9.9) 1 2 58 9. Irregular surfaces on hypersurfaces of degree 4 with non-degenerate isolated singularities Since X is a conic bundle, we can associate to π : X → B the embedding ϕ : B −→ Gr(2, P4) (9.10) of the base curve B into the Grassmannian Gr(2, P4) of planes in P4, where ϕ sends a point b ∈ B to the plane Pb spanned by the conic Fb = π−1(b), the fibre of π. Let U be the pullback under ϕ of the universal subbundle of Gr(2, P4). It is a rank 3 bundle on B and ϕ induces the morphism defined on its projectivization P(U ). The image of eϕ is a 3-fold which, set-theoretically, is the union of the family of planes {Pb}b∈B. In particular, the line bundle OP(U )(1) := eϕ∗(OP4(1)) is determined by the identity U ∗ = ρ∗(OP(U )(1)) = π∗(OX (H)), (9.11) (9.12) eϕ : P(U ) −→ P4 where ρ : P(U ) → B is the structure projection. We will need to know the degree of U ∗. Claim. deg(U ∗) = 3d − δ 4 . To justify the claim, we start by computing the holomorphic Euler characteristic of U ∗ in (9.12): χ(U ∗) = χ(OX (H)) = H 2 − H · KX 2 + χ(OX ) = d − H · KX 2 − (q − 1). On the other hand, the Riemann-Roch for U ∗ on B gives χ(U ∗) = deg(U ∗)−3(q −1). Putting the two expressions for χ(U ∗) together, we deduce deg(U ∗) = d − H · KX 2 + 2(q − 1). This and the expression for H · KX in (9.8) imply the equality of the claim. Set d′ := 3d − δ 4 . (9.13) The geometric meaning of d′ is two-fold: 1) If ϕ′ : B ֒→ P9 is the composition of ϕ with the Plucker embedding of Gr(2, P4) then d′ is the degree of the image B′ = ϕ′(B). 2) If V is the image of eϕ, then d′ = deg(V ). It will be more convenient at this point to express the double point formula (9.9) in terms of d and d′: d2 − 9d + 2d′ = 16(q − 1). (9.14) We also bring in the ndp formula (9.3) to obtain an upper bound for d′ d2 = 4(H 2 + H · KX ) + deg(Z) = 4(cid:18)d + 4(q − 1) − d − δ 2 (cid:19) + deg(Z) = 16(q − 1) + 8(d − d′) + deg(Z), 59 9. Irregular surfaces on hypersurfaces of degree 4 with non-degenerate isolated singularities where the second equality uses (9.8) and the last one (9.13). This expression for d2 and the identity (9.14) give deg(Z) = d + 6d′. (9.15) From this, the upper bound deg(Z) ≤ 45 in (9.5), and d ≥ 5, we deduce that d′ ≤ 6. This upper bound tells us that the curve B′, the image of ϕ, spans a subspace PN of dimension 3 ≤ N ≤ 5. The value N = 2 is excluded; indeed, if B′ spans a plane, then that plane either intersects Gr(2, P4) along B′ or it is contained in the Grassmannian Gr(2, P4). The first possibility means that B′ is a conic and hence, q = 0, while the second possibility tells us that all planes Pb, b ∈ B, intersect along a line, call it l; but then l ∼= P1 ⊂ X is a multi-section of π : X → B and this forces q to be zero again. Furthermore, if N = 5, then d′ = 6 and B must be an elliptic curve, i.e., q = 1. Substituting these values in (9.14), we obtain d2 − 9d + 12 = 0 which has no integer solutions. Thus N = 3 or 4. If N = 4, then d′ = 5 or 6. The first possibility implies again that q = 1 and the formula (9.14) becomes d2 − 9d + 10 = 0 with no integer solution. The second possibility, d′ = 6, implies that q = 1, 2. The first value has been ruled out in the discussion of the case N = 5. As for the second, the formula (9.14) becomes d2 − 9d − 4 = 0 with no integer solutions. Thus N = 3 is the only admissible value, while d′ = 4, 5 or 6. For d′ = 6, the Castelnuovo upper bound on genus gives q ≤ 4. Only q = 4 is compatible with (9.14), leading to the equation d2 −9d−36 = 0. Hence d = 12. Substituting into (9.15), we obtain the contradiction 45 ≥ deg(Z) = 12 + 36 = 48. For d′ = 5, the Castelnuovo upper bound implies q = 1, 2. The first value was already discarded in the case N = 4, while the second value substituted into (9.14) gives d2−9d−6 = 0 with no integral solutions. Thus we are left with d′ = 4 and hence q = 1. These values substituted in (9.14) yield d2 − 9d + 8 = 0 with the integer solution d = 8. We now go to (9.13) to deduce K 2 X = −δ = −8. This together with (9.8) imply H · KX = 0. Thus X is an ADSR elliptic conic bundle. Furthermore, from the formula (9.15) it follows that deg(Z) = 32. This completes the proof of the proposition. (cid:3) As we have mentioned in the discussion following the statement of Theorem 9.1, we do not know if an ADSR elliptic conic bundle is contained in a quartic hypersurface with isolated ordinary double points only. On the other hand, such a surface, by its very definition, is contained in a distinguished quartic whose singular locus is 2-dimensional. This implicitly appears in the works [1] and [28]. In the proof of Proposition 9.3 this distinguished quartic is the properties of the vector bundle U and its relation to the geometry of V . V , the image of the morphism eϕ : P(U ) → P4 in (9.11). The following statement summarizes of the morphism eϕ : P(U ) → P4 in (9.11). Then Proposition 9.4. Let U ∗ be the vector bundle defined in (9.12) and let V be the image 1) U ∗ has the form U ∗ = OB ⊕ F ∗, (9.16) 60 9. Irregular surfaces on hypersurfaces of degree 4 with non-degenerate isolated singularities where F ∗ is a rank 2 bundle on B fitting into the exact sequence 0 −→ OB(D) −→ F ∗ −→ OB(D′) −→ 0, (9.17) with OB(D) and OB(D′) being line bundles of degree 2. 2) V ⊂ P4 is a hypersurface of degree 4. It is a cone with vertex [v], the image of the section of P(U ) corresponding to the trivial summand in the direct sum decomposition (9.16). 3) The summand F ∗ in (9.16) determines a distinguished divisor P(F) of P(U ). Its image S under eϕ is a base of the cone V . In particular, S is a quartic surface with the singular locus Sing(S) consisting of either one or two (skew) lines in P3. The latter possibility occurs when the exact sequence (9.17) splits. 4) The singular locus Sing(V ) of V is the cone over Sing(S) with vertex at [v]. In partic- ular, set-theoretically, it is composed of either one or two planes depending on whether or not the sequence (9.17) is non split. Proof. Consider the pullback under the morphism ϕ in (9.10) of the dual of the universal sequence on Gr(2, P4) 0 −→ F −→ H 0(U ∗) ⊗ OB −→ U ∗ −→ 0. (9.18) This implies that H 0(F) = 0. The subbundle F ⊂ H 0(U ∗) ⊗ OB defines the morphism ϕ∨ : B −→ Gr(1, (P4) ∨ ), dual of the morphism ϕ in (9.10), where (P4)∨ = P(H 0(U ∗)). Composing ϕ∨ with the Plucker embedding Gr(1, (P4)∨) ⊂ P( 2H 0(U ∗)) ∼= (P9)∨, we obtain the embedding ψ : B ֒→ (P9)∨. The image of ψ, as the image of the embedding ϕ′ in the proof of Proposition 9.3, spans a P3. Hence the image of the linear map w : 2H 0(U ∗)∗ −→ H 0(det(F ∗)) defining the morphism ψ is 4-dimensional, while ker(w) is a 6-dimensional subspace of 2H 0(U ∗)∗. This means that P(ker(w)) intersects the Grassmann variety of decomposable tensors in P( 2H 0(U ∗)∗) along a subscheme of dimension at least 1 implying that the linear map H 0(U ∗)∗ −→ H 0(F ∗) (9.19) has a non-trivial kernel. Indeed, let l ∧ l′ be a nonzero decomposable tensor in 2H 0(U ∗)∗ lying in the kernel of w. We may assume that the pencil Span(l, l′) injects into H 0(F ∗) under the map in (9.19), since otherwise we are done. Thus we can think of l and l′ as two linearly independent global sections of F ∗ which are proportional, i.e., l ∧ l′ is zero as a section of det(F ∗). Hence the Koszul sequence of one of these sections gives rise to an exact sequence 0 −→ OB(D) −→ F ∗ −→ OB(D′) −→ 0, (9.20) where h0(OB(D)) ≥ 2. Hence deg(D) ≥ 2. Furthermore, since deg(F ∗) = deg(U ∗) = 4 and the quotient OB(D′) must be generated by its global sections and nontrivial (the latter 61 9. Irregular surfaces on hypersurfaces of degree 4 with non-degenerate isolated singularities comes from H 0(F) = 0), we deduce deg(D) = deg(D′) = 2. This and the exact sequence (9.20) imply h0(F ∗) = 4. Since H 0(U ∗)∗ ∼= H 0(OX (H) is 5-dimensional, we deduce that the kernel in (9.19) is nontrivial. Considering the dual of (9.18), we see that the kernel in (9.19) is H 0(U ). Now, from H 0(U ) 6= 0 and the global generation of U ∗, we deduce the direct sum decomposition U ∗ ∼= OB ⊕ G. (9.21) This together with (9.18) implies that the direct summand G fits into the exact sequence 0 −→ F −→ H 0(G) ⊗ OB −→ G −→ 0. (9.22) It remains to identify G with F ∗. To this end, we go back to the morphism ϕ′ : B → P9 in the proof of Proposition 9.3 and recall that its image spans a P3. This means that the linear map 3H 0(U ∗) −→ H 0(det(U ∗)) has the image of dimension 4 or, equivalently, the kernel of dimension 6. This together with the decomposition in (9.21) implies that the linear map 2H 0(G) −→ H 0(det(G)) has the kernel, call it W , of dimension at least 2. Combining this and the second exterior power of (9.22) gives h0(F ⊗ G) = h0(End(F ∗, G)) ≥ 2. It follows that a general morphism F ∗ → G is an isomorphism. Thus G ∼= F ∗. Substituting into (9.21), we deduce the de- composition asserted in (9.16). Furthermore, we have h0(End(F ∗, F ∗)) ≥ 2. A nontrivial endomorphism of F ∗ gives rise to the exact sequence (9.17). The remaining assertions of the proposition are obvious geometric analogues of the prop- (cid:3) erties of U ∗ (resp. F ∗) in 1). Remark 9.5. Let X be a smooth surface in P4 such that -- X is birational to an irregular ruled surface, -- mX = 4 and X is contained in a quartic hypersurface with only nondegenerate isolated singularities, -- the degree of X in P4 is not 8. Then from the proof of Proposition 9.3 it follows that the fibres of X are embedded as curves of degree at least 3. 9.2. The study of H 1(JZ(KX + 4H)) Our result here is as follows. Proposition 9.6. Let X and V4 be as in Theorem 9.1 and assume X to be irregular. Then H 1(JZ (KX + 4H)) = 0 with a possible exception of X being an ADSR elliptic conic bundle. If such a situation occurs, then h1(JZ(KX + 4H) = 1 and Z = XT Sing(V4) is a subset of 32 nodes of V4. 62 9. Irregular surfaces on hypersurfaces of degree 4 with non-degenerate isolated singularities We assume the nonvanishing of the cohomology group H 1(JZ (KX + 4H)). The identifi- cation H 1(JZ(KX + 4H))∗ = Ext1(JZ(4H), OX ) provided by the Serre duality, gives rise to a nontrivial extension sequence 0 −→ OX −→ E −→ JZ(4H) −→ 0. (9.23) We may assume the sheaf E in the middle of this sequence to be locally free. The Bogomolov semistability condition for this sheaf reads 0 ≤ 4c2(E) − c2 1(E) = 4 deg(Z) − 16H 2 = 4(deg(Z) − 4d). (9.24) From this and the upper bound deg(Z) ≤ 45, it follows that E is Bogomolov unstable provided d ≥ 12. For this reason, the proof of Proposition 9.6 is naturally divided into two parts: • the first part rules out the case d ≥ 12 by examining the geometric consequences of the Bogomolov instability of E; • the second part deals with the remaining values 5 ≤ d ≤ 11 for the degree of X. First part: d ≥ 12. We begin by recording some geometric consequences of the Bogo- molov instability condition deg(Z) < 4d for the sheaf E in (9.23). Let OX (A) be the maximal Bogomolov destabilizing subsheaf of E. Combining the inclu- sion OX (A) → E with the defining extension sequence (9.23) gives the diagram 0 OX(A) 0 OX E JZ(4H) 0 (9.25) JZ ′(E) 0 where Z ′ is a 0-dimensional subscheme of X and JZ ′ is its ideal sheaf. The Bogomolov destabilizing condition tells us that the divisor is in N +(X), the positive cone of X. In particular, B := A − 2H A = 2H + B and E = 2H − B. (9.26) The first formula implies that the slanted arrows in the above diagram are nonzero. From this it follows that E is an effective, nonzero divisor and that those arrows are defined by multiplication by a section e ∈ H 0(OX (E)) corresponding to E. In particular, from the lower) slanted arrow, we deduce that Z (resp. Z ′) is contained in E. On the upper (resp. 63 9. Irregular surfaces on hypersurfaces of degree 4 with non-degenerate isolated singularities other hand we also know that the linear system JZ(3H) is base point free outside Z. Hence there is a reduced irreducible divisor C ∈ 3H containing Z. Therefore, Z ⊂ C · E and thus subject to the estimate deg(Z) ≤ 3H · E. (9.27) Computing the second Chern number c2(E) = deg(Z) from the vertical sequence in (9.25) gives deg(Z) = A · E + deg(Z ′) = (4H − E) · E + deg(Z ′) = 4H · E − E2 + deg(Z ′). (9.28) Together with the inequality (9.27), this expression implies E2 ≥ H · E + deg(Z ′). This inequality acquires more geometry by observing that the divisor E − H is effective17, allowing us to write E = H + R, where R is an effective divisor. Combining this and the second formula in (9.26), we have H = B + R. (9.29) In addition, by substituting the formulas from (9.26) into (9.28), we obtain the identity deg(Z) = A · E + deg(Z ′) = (2H + B) · (2H − B) + deg(Z ′) = 4d − B2 + deg(Z ′), (9.30) which, together with the bound deg(Z) ≤ 45, enables us to control the degree d and the intersection number H · R. Lemma 9.7. If d ≥ 12, then H · R ≤ 2, i.e., R is empty, a line, or a conic. Proof. Assume that H · R ≥ 3. By the Hodge index we have B2 ≤ (H · B)2 d = (H · (H − R))2 d = (d − H · R)2 d ≤ (d − 3)2 d where the last inequality uses the fact that (d−H·R)2 interval [3, d]. Substituting this upper bound for B2 in (9.30) gives the estimate is a decreasing function of H · R on the d 45 ≥ deg(Z) = 4d − B2 + deg(Z ′) ≥ 4d − (d − 3)2 d + deg(Z ′) = 3d + 6 − 9 d + deg(Z ′). (9.31) From this it follows that the only admissible values for d are 12 and 13. Furthermore, the ndp formula (9.3) yields the divisibility condition deg(Z) ≡ d2 (mod 8). (9.32) For d = 13, it implies deg(Z) ≤ 41, hence (9.31) becomes 41 ≥ deg(Z) ≥ 3 · 13 + 6 − 9 13 + deg(Z ′) > 44 which is absurd. The same argument rules out the other admissible value as well. (cid:3) 17This is seen by tensoring (9.25) with OX (−H) and showing that H 0(JZ ′ (E − H)) 6= 0; the argument is exactly the same as in the proof of Lemma 7.3. 64 9. Irregular surfaces on hypersurfaces of degree 4 with non-degenerate isolated singularities Next we examine the three possibilities for R provided by the previous lemma. If R = 0, then B = H and Z ′ = 0 and the formula (9.30) becomes 45 ≥ deg(Z) = 3d. Hence 12 ≤ d ≤ 15. But none of these values satisfies the divisibility condition (9.32). If R is a line, then B = H − R and B2 = d − 2 + R2. Substituting the self-intersection number into (9.30) gives 45 ≥ deg(Z) = 3d + 2 − R2 + deg(Z ′). Hence d = 12, 13 or 14. The last two values together with the divisibility condition in (9.32) imply that deg(Z) = 41 and 44 respectively. But both values force R2 = deg(Z ′) = 0. Thus an irregular X is a scroll and this is ruled by Remark 9.5. Therefore, we are left with d = 12 and deg(Z) = 40. However, the inequality in (9.27) now reads 40 ≤ 3H · E = 3H · (H + R) = 36 + 3 = 39 which is absurd. If R is a conic, then H · B = 10 and B2 = d − 4 + R2. Substituting into (9.30) gives 45 ≥ deg(Z) = 3d + 4 − R2 + deg(Z ′). This together with the divisibility condition (9.32) implies deg(Z) = 40, d = 12, and deg(Z ′) = R2 = 0. The last equality together with H · R = 2 tells us that an irregular surface X is a conic bundle and this is impossible by Remark 9.5. This completes the treat- ment of the case d ≥ 12 and thus proves the vanishing of H 1(JZ(KX + 4H)) for these values of d. Second part: 5 ≤ d ≤ 11. Though the argument comes down to a case by case consideration, there is a basic feature that is common to all of them. This aspect will be explained next. We begin by recalling that the values of the degree d control the values of deg(Z) via the divisibility condition (9.32), deg(Z) ≡ d2 ( mod 8). Hence we obtain deg(Z) = 8k + 1, 8k + 4, 8k, if d ∈ {5, 7, 9, 11} if d ∈ {6, 10} if d = 8 (9.33) where 1 ≤ k ≤ 5, in view of the upper bound deg(Z) ≤ 45. Thus for every value of d we obtain a list of admissible values for deg(Z). These values are divided into two types according to whether the sheaf E in the extension sequence (9.23) is semistable or unstable in the sense of Bogomolov. When E is unstable, all the possibilities are discarded by exploiting the decomposition H = B + R in (9.29) and by observing that the divisor R has to move in a linear system in order for the hypersurface V4 to have isolated singularities. In the semistable case there are two basic ingredients: − we check that X is birational to a ruled surface of irregularity q, − the two conditions, X birational to a ruled surface and OX (KX +H) base point free and big, exclude all but one possibility: X is an ADSR elliptic conic bundle, see Proposition 9.3. 65 9. Irregular surfaces on hypersurfaces of degree 4 with non-degenerate isolated singularities With these general guidelines in mind, we proceed with the second part of the proof according to the possible values of the degree, 5 ≤ d ≤ 11. • The case d = 11. According to (9.33), we have deg(Z) = 8k +1 ≤ 41. This upper bound insures that the sheaf E in the middle of the extension sequence (9.23) is still Bogomolov unstable, see (9.24). Thus the argument used in the case d ≥ 12 applies and we obtain the decomposition H = B + R as in (9.29). This implies H · B = H · (H − R) = d − H · R ≤ d and, hence, by the Hodge index, B2 ≤ d. Substituting into (9.30) gives deg(Z) = 4d − B2 + deg(Z ′) ≥ 3d + deg(Z ′) ≥ 3d = 33. (9.34) Thus deg(Z) = 33 or 41. Furthermore, for the first value all the above inequalities must be equalities and hence R and Z must both be zero. It follows that E = H and that the vertical sequence in (9.25) takes the form 0 −→ OX (3H) −→ E −→ OX (H) −→ 0. (9.35) We have already encountered a similar situation in the proof of Lemma 7.5 and we use the same argument here. Namely, recall the identification H 0(E(−H)) ∼= H 0(JZ(3H)) (9.36) resulting from the defining extension sequence (9.23) tensored with OX (−H). On the other hand, the destabilizing sequence (9.35) gives 0 −→ H 0(OX (2H)) −→ H 0(E(−H)) −→ H 0(OX ). This together with the fact that the above inclusion H 0(OX (2H)) ֒→ H 0(E(−H)) must be proper, implies that the arrow on the right must be onto. Hence H 0(OX (2H)) ֒→ H 0(E(−H)) is a codimension 1 subspace of H 0(E(−H)). This and the isomorphism (9.36) tell us that the subspace e H 0(OX (2H)), where e ∈ H 0(OX (H)) is a section defining the divisor E, is a codimension 1 subspace of H 0(JZ (3H)). Next recall that the space H 0(JZ (3H)) contains the 5-dimensional subspace W , spanned by the restrictions to X of the partial derivatives of a homogeneous polynomial, call it f , defining the quartic hypersurface V4 that contains X. The above discussion implies that the subspace e H 0(OX (2H)) intersect W along a subspace of codimension at most 1. This means that we can choose homogeneous coordinate functions Xi, i = 0, . . . , 4, on P4 so that ∂f ∂Xi = hγi, for i = 0, . . . , 3, where γi ∈ H 0(OP4(2)) and h ∈ H 0(OP4(1)) is the linear form corresponding to the section e under the identification H 0(OP4(1)) ∼= H 0(OX (H)). From this it follows that V4 is singular Notice that the above argument remains valid as long as the space H 0(JZ ′(R)) -- where the cokernel of H 0(OX (A − H) → H 0(E(−H)) lives -- is 1-dimensional. This will be our tool along the subvariety (cid:0)h = ∂f isolated singularities. ∂X4 = 0(cid:1) which contradicts the assumption that V4 has only 66 9. Irregular surfaces on hypersurfaces of degree 4 with non-degenerate isolated singularities to rule out all the remaining cases whenever the sheaf E in (9.23) is Bogomolov unstable. We give a full account of this for deg(Z) = 41. From the identity B2 = 4d − deg(Z) + deg(Z ′) = 44 − 41 + deg(Z ′) = 3 + deg(Z ′) we deduce the inequality B2 ≥ 3. This and the Hodge index give H · B ≥ 6 or, equivalently, H · R = H · (H − B) = d − H · B ≤ 5. From the inequality (9.28) we also obtain the lower bound H · R ≥ deg(Z) 3 − d = 41 3 − 11, i.e., H · R ≥ 3, and proceed according to the possible values of H · R = 3, 4, or 5. As we have said above, the main idea is, as in the case of deg(Z) = 33, to show that the space H 0(JZ ′(R)) is 1-dimensional. For this we will also need the formula deg(Z ′) = B2 − 3 = (H − R)2 − 3 = d − 3 − 2H · R + R2 = 8 − 2H · R + R2 (9.37) which relates deg(Z ′) to the intersection numbers H · R and R2. 1) H · R = 3. We wish to analyse the possibility h0(OX (R)) ≥ 2. Let R0 (resp. R′) be the fixed (resp. moving) part of the linear system R. If R0 6= 0, then H · R′ ≤ 2 and hence, by Hodge index, R′2 ≤ 0. Since the linear system R′ has at most a 0-dimensional base locus, we deduce the equality R′2 = 0. But this means that X is either a scroll or a conic bundle and, in view of Remark 9.5, neither possibility is allowed. Thus we may assume that the linear system R has at most a 0-dimensional base locus. This and the Hodge index imply R2 = 0. Hence R is base point free. At the same time, observe that h0(OX (B)) = h0(OX (H − R)) ≥ 1, i.e., B is effective and h0(OX (H − B)) = h0(OX (R)) ≥ 2. Since B can not be a line, we deduce that if R moves on X, then h0(OX (R)) = 2. From the formula (9.37), we also find that deg(Z ′) = 2. Hence h0(JZ ′(R)) ≤ h0(OX (R)) − 1 = 1 and from here on we conclude as in the case deg(Z) = 33. 2) H · R = 4. The argument follows the same pattern as in the previous case. Write R = R0 + R′ as above. If R0 6= 0, then H · R′ ≤ 3 and we deduce that R′2 = 0. Hence, R′ is a base point free pencil inducing the morphism f : X −→ P1. (9.38) The fibres of f must be connected (otherwise we are back to the situation of X being a scroll or a conic bundle) with H · R′ = 3 (resp. H · R0 = 1) and hence, f is an elliptic fibration18 with fibres embedded by OX (H) as plane curves of degree 3. The union of the planes spanned 18The other possibility is that f has rational fibres, but then X is rational. 67 9. Irregular surfaces on hypersurfaces of degree 4 with non-degenerate isolated singularities by the fibres of f form a hypersurface, call it V . The degree of this hypersurface is subject to the inequality deg(V ) ≤ deg(f∗(OX (H))) Setting f∗(OX (H)) =L3 3Xi=1 deg(V ) ≤ i=1 OP1(ai), we obtain ai = h0(f∗(OX (H))) − 3 = h0(OX (H)) − 3 = 5 − 3 = 2, contradicting the assumption on the smallest degree of a hypersurface containing X. We have just shown that the linear system R has at most a 0-dimensional base locus. By the Hodge index R2 ≤ 1. Hence either R2 = 0 or R2 = 1. If R2 = 0, then R is base point free and induces a morphism as in (9.38). We may again assume that the fibres of this morphism are connected, since otherwise X is either a scroll or a conic fibration. In view of the degree H · R = 4 of R, the morphism f : X → P1 is either an elliptic fibration or a fibration by plane curves of degree 4. Hence KX · R = 0 or KX · R = 4. (9.39) The ndp formula tells us that KX · H = 9 and together with (9.39) implies KX · B = KX · (H − R) = 9 or 5. Since B2 = (H − R)2 = 11 − 2H · R = 11 − 2 · 4 = 3, we obtain B2 + KX · B = 12 or 8 (9.40) which is the degree of the dualizing sheaf of B. But B is as a divisor of degree dB = H ·B = 7 contained in a plane19. Hence the degree of its dualizing sheaf verifies dB(dB − 3) = 7 · 4 = 28, which does not match any of the values in (9.40). If R2 = 1, then R must have a unique base point and we may assume that a general member of the linear system is either an elliptic curve or a smooth plane curve of degree 4. These possibilities lead to KX · R = −1 and KX · R = 3 respectively. Together with the identity KX · H = 9 (the ndp formula), they imply But B2 = (H − R)2 = 4, hence the degree of the dualizing sheaf of B verifies KX · B = 10 (resp. = 6). B2 + KX · B = 14 (resp. = 10) and, as in the case R2 = 0, neither value agrees with dB(dB − 3) = 28. 3) H · R = 5. Then H · B = 6, B2 = 3, and R2 = 2. Writing R = R0 + R′ as before, we see that R0 6= 0 reduces to the cases previously considered. So we may assume that the linear system R has at most a 0 dimensional base locus. Furthermore, from R2 = 2 it follows that a general member of R is irreducible. If, in addition, a general curve in the linear system is 19This follows from h0(Ox(H − B)) = h0(OX (R)) = 2. 68 9. Irregular surfaces on hypersurfaces of degree 4 with non-degenerate isolated singularities not contained in a hyperplane, then by the Castelnuovo upper bound on the genus of curves in P4, a general member of R is a smooth curve of genus 1. This and the adjunction for R give KX · R = −R2 = −2 implying that X is birational to a ruled surface with q = 1. From the ndp formula in (9.3), we have H · KX = 9 and obtain K 2 X = −17. Next we consider the line bundle OX (KX + 2R). From the Riemann-Roch, h0(OX (KX + 2R)) = (KX + 2R) · R = R2 = 2. X + 4KX · R + 4R2 = K 2 On the other hand, (KX + 2R)2 = K 2 X = −17. Hence the linear system KX + 2R has a fixed part F 6= 0. In particular, there is a dense open subset F ′ of F such that through every point x ∈ F ′ passes an irreducible curve Rx ∈ R. But x is a base point of OX (KX + 2R) and by [29, Theorem 1], there is an irreducible curve C ⊂ X passing through x with the property C · Rx = C · R = 0. Hence the curves C and Rx have a common component. Since both curves are irreducible, we obtain C = Rx and, therefore, 0 = C · Rx = R2 x = R2 = 2, an obvious contradiction. Now we know that a general curve in the linear system R is contained in a hyperplane, hence that H 0(OX (B)) = H 0(OX (H − R)) 6= 0. This tells us that B is an effective divisor. Furthermore, h0(OX (H − B)) = h0(OX (R)) ≥ 2 tells us that equality must hold, i.e., B is a plane curve subject to dB = H · B = 6 and B2 = 3. Hence B2 + KX · B = dB(dB − 3) = 18. It follows that KX · B = 18 − B2 = 15. But then 15 = KX · B = KX · (H − R) = KX · H − KX · R ≤ 9 − (−2 − R2) = 9 + 4 = 13 gives an obvious contradiction. This completes the case d = 11. • The case d = 10. As in the case d = 11, we obtain two possible values for deg(Z): 36 and 44. For the first value, the sheaf E in (9.23) is Bogomolov unstable. The considerations are the same as in the case d = 11 and we leave the details to the reader. For the latter, i.e., if deg(Z) = 44, then we follow the plan outlined in the paragraph prior to the case d = 11. We begin by showing that X is birational to a ruled surface. The ndp formula reads H 2 + H · KX = 14, hence H · KX = 4. Combining this with the Hodge Index, gives K 2 upper bound into the double point formula yields X ≤ 1. Substituting this −20 = 2K 2 X − 12χ ≤ 2 − 12χ. Hence χ ≤ 1 and we obtain K 2 X = −10 + 6χ ≤ −4. (9.41) We are now ready to show that X is birational to a ruled surface. Since we are assuming that X is irregular, it is enough to show that the Kodaira dimension of X is negative. Assume the opposite and let X0 be the minimal model of X with σ : X → X0 the sequence of blow down maps. We write KX = σ∗K0 + ∆, 69 9. Irregular surfaces on hypersurfaces of degree 4 with non-degenerate isolated singularities where K0 is the canonical divisor of X0 and ∆ is the exceptional divisor composed of the blown-down curves. Since K 2 0 ≥ 0, the estimate in (9.41) tells us that σ is the composition of at least four blow-downs. It follows that ∆ has at least four irreducible components, i.e., that H · ∆ ≥ 4. This gives 4 = H · KX = H · σ∗K0 + H · ∆ ≥ H · σ∗K0 + 4 (9.42) or, equivalently, H · σ∗K0 ≤ 0. Since σ∗K0 is nef, it follows that H · σ∗K0 = 0 and X is of Kodaira dimension zero. By the Enriques-Kodaira classification, we must have χ = 0. Returning to the equality in (9.41), we obtain K 2 X = −10. But then ∆ has at least ten irreducible components and the estimate in (9.42) gives H · σ∗K0 ≤ −6 which is impossible. Hence X is birational to a ruled surface and χ(OX ) = 1 − q, which, substituted into the equality (9.41), gives K 2 X = −10 + 6χ(OX ) = −10 − 6(q − 1). (9.43) We pursue by studying the adjoint divisor KX + H. We have, using (9.43), (KX + H)2 = K 2 X + 2H · KX + H 2 = −10 − 6(q − 1) + 2 · 4 + 10 = 8 − 6(q − 1) and recalling that OX (KX + H) is base point free, see Proposition 9.3, we deduce that either q = 1 or q = 2. If q = 2, then (KX + H)2 = 2. By Riemann-Roch, we have h0(OX (KX + H)) = (KX + H) · H 2 + χ(OX ) = 7 − 1 = 6. But, according to Proposition 9.3, the linear system KX + H is base point free and hence defines a morphism X → P5 whose image is a surface of degree at most (KX + H)2 = 2 and this is impossible. If q = 1, then K 2 X = −10 and (KX + H)2 = 8. The latter and Proposition 9.3 tell us that KX + H is base point free and defines a morphism f : X −→ P6 which must be birational onto its image. We also record the projection morphism π : X −→ B (9.44) (9.45) onto an elliptic curve B with rational fibres. In particular, we wish to understand the degree of the fibres of π with respect to H. For this we look at the double adjoint line bundle OX(2KX + H). By Riemann-Roch h0(OX (2KX + H)) = (2KX + H) · (KX + H) 2 = KX · (KX + H) + (KX + H)2 2 = 1. Let D be the divisor defined by a nonzero section of OX (2KX + H). It is a non-zero divisor, since D2 = (2KX + H)2 = −24. Furthermore, since (KX + H) · D = 2, and since KX + H is base point free, an irreducible component C of D verifies one of the following possibilities: (i) (KX + H) · C = 0, (ii) (KX + H) · C = 1, (iii) (KX + H) · C = 2. (9.46) The curves C of type (i) are precisely the curves contracted by the morphism f in (9.44). From (KX +H)·C = 0 it follows that C must be a smooth rational curve with C 2 = H ·C −2. 70 9. Irregular surfaces on hypersurfaces of degree 4 with non-degenerate isolated singularities This implies that H · C = 1 or 2, since C 2 ≤ 0. The second value means that X is a conic bundle which is impossible by Remark 9.5. Thus the curves of type (i) are (−1)-curves on X, contracted by OX (KX + H), and mapped onto lines by OX (H) in the embedding X ⊂ P4. In particular, all these irreducible components of D are contained in the fibres of the projection π : X → B. The curves C of type (ii) are smooth rational curves (they are lines with respect to the morphism f ) with C 2 = H · C − 3. Hence H · C = 1, 2 or 3. The last value is impossible, since otherwise 3 is the H-degree of all fibres of π and they all have intersection −1 with D which is impossible. Thus the curves of type (ii) are smooth rational curves on X of self-intersection, either −1, or −2. Moreover, they are conics and lines respectively in the embedding by OX (H) and are contained in the fibres of the projection π : X → B. Also observe that there are at most 2 distinct such curves in D. We claim that there is no curve of type (iii). Indeed, all points of such a curve C are fixed points of OX (2KX + H) and according to [29, Theorem 1], through every point x ∈ C passes an irreducible curve of type (i) or (ii). This is impossible since those curves are rigid. From the above analysis of the irreducible components of D it follows that D = C1 + C2 + D0, where Ci, i = 1, 2, are the curves (not necessarily distinct) of type (ii) and D0 is the residual part of D composed of curves of type (i). In addition, [29, Theorem 1] stipulates the existence of a unique divisor Ex passing through every point x ∈ Ci, subject to either or (KX + H) · Ex = 0 and E2 x = −1 (KX + H) · Ex = 1 and E2 x = 0 (a) (b) We know that the irreducible components of Ex in (a) are the (−1)-curves of type (i) and no such curve passes through a general point of Ci. So for all points x in the complement of some finite subset of Ci, the corresponding divisor Ex passing through x is of type (b). This implies that Ex must be the fibre of π containing Ci. Thus if F is the class of a fibre of the projection π in (9.45), then or, equivalently, H · F = 3. But then (KX + H) · F = 1. D · F = (2KX + H) · F = 2KX · F + H · F = −4 + 3 = −1 contradicting the fact that D is effective. This completes the proof of the case d = 10. • The case d = 9. The possible values for deg(Z) are 41 and 33, where the first (resp. second) value is 'stable' (resp. 'unstable'). We treat only the stable possibility, i.e., From the ndp formula we obtain H · KX = 1. This and the Hodge index imply K 2 From the double point formula we obtain X ≤ 0. deg(Z) = 41. 0 ≥ K 2 X = 6χ − 7 71 9. Irregular surfaces on hypersurfaces of degree 4 with non-degenerate isolated singularities and hence χ ≤ 1. Arguing as in the case d = 10, deg(Z) = 44, we deduce that X has negative Kodaira dimension or, equivalently, that X is birational to a ruled surface of irregularity q. Thus χ = 1 − q and K 2 X = −7 − 6(q − 1). We proceed with the study of OX (KX + H). The self-intersection (KX + H)2 = 11 + K 2 X = 4 − 6(q − 1) and Proposition 9.3 tell us that the only possibility for the irregularity is q = 1. Hence (KX + H)2 = 4 and K 2 X = −7. Computing the genus of a smooth curve in the adjoint linear system KX + H gives (KX + H)2 + (KX + H) · KX = 4 + 1 − 7 = −2, i.e., such a curve is rational. But then X can not be irregular. • The case d = 8. The possible values for deg(Z) are deg(Z) = 40, 32, and 24, where the last one is the only unstable value. It can be treated easily as in the case d = 11, deg(Z) = 33, so we turn towards the two remaining stable values. From the ndp formula we obtain where k = 4 or 5. The negativity of the Kodaira dimension follows readily. Hence χ = 1 − q and the double point formula gives H · KX = 8 − 2k, K 2 X = −8 − 6(q − 1) − 5(4 − k), where k = 4, 5. We now compute the self-intersection (KX + H)2 = −6(q − 1) − (4 − k), where k = 4, 5. This and Proposition 9.3, according to which OX (KX + H) is base point free, give the only possibility q = 1, k = 4, (KX + H)2 = 0. This corresponds to the exceptional possibility of an ADSR elliptic conic bundle. • The case d ≤ 7. All those values are incompatible with OX (KX + H) being base point free and big. Indeed, the last condition stipulates (KX + H)2 > 0, which we rewrite as K 2 X > −d − 2H · KX. We substitute this into the double point formula and obtain d2 − 8d + 12χ > H · KX. On the other hand, expressing H · KX from the ndp formula (9.3) gives H · KX = 1 4 (d2 − 4d − deg(Z)). Putting it together with the previous inequality, we have 3 4 d2 − 7d + 1 4 deg(Z) + 12χ > 0. 72 (9.47) (9.48) 10. Albanese dimension The left hand side of this inequality is an increasing function of d for d ≥ 5. So, for d ≤ 7, we have 3 4 . Substituting into (9.48) and using deg(Z) ≤ 45 give 4 d2 − 7d ≤ − 49 12χ > 49 4 − 1 4 deg(Z) ≥ 1. Hence Going back to (9.47) we obtain χ ≥ 1. (9.49) H · KX = 1 4 (d2 − 4d − deg(Z)) ≤ 1 4 (72 − 4 · 7 − deg Z) = 1 4 (21 − deg Z), (9.50) for d ≤ 7. In particular, if d ≤ 7 and deg(Z) > 21, then H · KX is negative. Hence the Kodaira dimension of X is negative. This together with the estimate (9.49) gives 1 ≤ χ = 1−q meaning that q = 0. If deg(Z) ≤ 21, then deg(Z) ≤ 21 < 4d, provided 6 ≤ d ≤ 7. Hence the values of deg(Z) ≤ 21 are unstable and one has the estimate deg(Z) > 3d, for d = 6, 7, coming from the first inequality in (9.34) and the divisibility condition (9.32). This implies the only possibility: d = 6 and deg(Z) = 20. Substituting these values into the first equality in (9.50) gives H · KX = Hence, as before, q = 0. 1 4 (d2 − 4d − deg(Z)) = −2. For the remaining value d = 5, the identity (9.47) and the divisibility condition (9.32) give H · KX = 1 4 (5 − deg(Z)) and deg(Z) = 8k + 1. Therefore, as before, the Kodaira dimension of X is negative, except possibly in the case deg(Z) = 1. But then the subscheme Z is a single point, call it p, and the group H 1(Jp(KX + 4H)) = 0, since by [29, Theorem 1], the line bundle OX (KX + 4H) is very ample. This completes the study of cases 5 ≤ d ≤ 11. 10. Albanese dimension In this section we consider the Albanese dimension of surfaces lying on small degree hypersurfaces. Since we have already seen that there is no irregular surface lying on a quadric hypersurface and there is only the elliptic scroll on a cubic hypersurface, our examination will focus on hypersurfaces of degree 4 and 5. From now on we set m = mX = 4 or 5 -- the smallest degree of a hypersurface containing a smooth surface X ⊂ P4 -- and we ask the following questions: 1) Can X be of Albanese dimension 2? 2) If the answer to question 1) is affirmative, can the irregularity of X be bounded? 73 10. Albanese dimension To investigate these questions we assume X of Albanese dimension 2 and consider the extension sequence (2.7) associated to the cohomology class ξ ∈ H 1(ΘX (−KX − (5 − m)H)) arising from a hypersurface of degree m containing X (see Lemma 2.2). Observe: the as- sumption that the Albanese dimension of X is 2 insures that X has non-negative Kodaira dimension and hence the class ξ = δX (s) in Lemma 2.2 is nonzero. In other words the sequence (2.7) is non split. 10.1. The Albanese dimension of X ⊂ P4 with mX = 4 This subsection is devoted to the proof of the following. Theorem 10.1. If X ⊂ P4 is a smooth surface with mX = 4, then the Albanese dimension of X is at most 1. Proof. Assume that X is subject to the hypothesis of the theorem and that it has the Albanese dimension 2. Since m = mX = 4, the exact sequence (2.7) becomes 0 −→ OX (−KX − H) −→ Tξ −→ ΩX −→ 0. (10.1) The assumption on the Albanese dimension means that ΩX is generically generated by global sections. In particular, q(X) ≥ 2 and KX is effective. The last property implies that the homomorphism H 0(Tξ) −→ H 0(ΩX) (10.2) induced by the epimorphism in (10.1) is an isomorphism: the injectivity follows from the obvious vanishing of H 0(OX (−KX − H)) and the surjectivity is insured by the following lemma. Lemma 10.2. If KX is effective, then the divisor KX + H is nef and big. In particular, H 1(OX (−KX − H)) = 0. Proof of Lemma 10.2. Let C be an effective curve such that (KX + H) · C ≤ 0. It follows that KX · C ≤ −H · C < 0. But KX is effective, hence C 2 < 0 forcing C to be a (−1)-curve, i.e., C 2 = KX · C = −1. Since H · C ≥ 1, we deduce (KX + H) · C ≥ 0 and hence (KX + H) · C = 0. Thus KX + H is nef. This implies the non-negativity (KX + H) · KX ≥ 0. Hence (KX + H)2 = (KX + H) · KX + (KX + H) · H ≥ (KX + H) · H ≥ H 2 > 0. (cid:3) Once the isomorphism (10.2) has been established, we define the subsheaf F ⊂ Tξ as the saturation of the subsheaf generated by the global sections of Tξ. In particular, the inclusion F ⊂ Tξ composed with the epimorphism in (10.1) gives the morphism ϕ : F → ΩX which is generically surjective. Hence the rank of F is at least 2. On the other hand (10.1) tells us that the determinant of Tξ is OX (−H). Hence F must be of rank 2 and Lemma 2.3 applies to give us the decomposition KX = L + E (10.3) where L = c1(F) and E = c1(coker(ϕ)) is an effective nonzero divisor. Furthermore, L is effective as well, since by definition H 0(F) ∼= H 0(ΩX) and F is generically generated by its 74 10. Albanese dimension global sections. This and Lemma 2.3, 3), imply that L = 0. Hence F = OX ⊕ OX and therefore, q(X) = 2. In addition, the trivial subsheaf F provides 0 −→ OX ⊕ OX −→ Tξ −→ JZ(−H) −→ 0, (10.4) a destabilizing sequence for the sheaf Tξ. This sequence will play an important role in a later part of the argument. In fact, more can be extracted from part 3) of Lemma 2.3. Namely, since L = 0 (and mX = 4) that assertion also tells us that we have a unique nonzero global section τ of ΘP4 ⊗ OX (−H) such that its image in H 0(NX (−H)) is the section e · s, where s is the global section of NX(−KX − H) defined by a quartic hypersurface containing X and e is a global section of OX(KX ) corresponding to the divisor E in (10.3). Furthermore, putting the normal sequence of X and the Euler sequence of ΘP4 ⊗ OX together and tensoring everything with OX(−H), we obtain the diagram 0 OX (−H) H 0(OX (H))∗ ⊗ OX η (10.5) 0 ΘX(−H) ΘP4 ⊗ OX(−H) NX(−H) 0 0 from which it follows that the global section τ of ΘP4 ⊗ OX (−H) constructed above comes from a unique element v ∈ H 0(OX (H))∗. This vector has the following geometric meaning. Lemma 10.3. Let V4 be a quartic hypersurface containing X and let [v] be the point of P(H 0(OX (H))∗) corresponding to the vector v above. Then V4 is a cone with the vertex [v] over a quartic surface S in P3. Proof of Lemma 10.3. To establish the asserted relation of V4 with the vector v we recall that V4 gives rise to the nonzero global section s of NX(−KX −H). Let Σ be the 1-dimensional part of the zero locus of this section and denote by σ a section of OX (Σ) corresponding to Σ. Then s = σs′, where s′ ∈ H 0(NX(−KX − H − Σ)) and has a 0-dimensional zero locus Zs′. Therefore the Koszul sequence of s′ 0 −→ OX s′ −→ NX(−KX − H − Σ) s′∧−→ JZs′ (3H − KX − 2Σ) −→ 0 is exact. Tensoring it with OX(KX + Σ), we obtain 0 −→ OX (KX + Σ) s′ −→ NX(−H) s′∧−→ JZs′ (3H − Σ) −→ 0. From this sequence, it follows that the global section es = eσs′ ∈ H 0(NX(−H)), which interests us, lies in the kernel of the homomorphism H 0(NX(−H)) → H 0(JZs′ (3H − Σ)). Furthermore, by the definition of v ∈ H 0(OX (H))∗, we have es = η(v), 75 10. Albanese dimension where η is the slanted arrow in the diagram (10.5). On the other hand, the composition H 0(OX (H))∗ ⊗ OX η −→ NX(−H) s′∧·−→ JZs′ (3H − Σ) σ ֒→ JZs′ (3H) is given by the partial derivatives of a polynomial defining V4. More precisely, if F is a homogeneous polynomial defining V4, then the composition σ ◦ (s′ ∧ ·) ◦ η sends every w ∈ H 0(OX (H))∗ to ∂w(F )X , the derivative of F in the direction of w restricted to X. Evaluating the composition on the vector v, we obtain ∂v(F )X = σ · ((s′ ∧ ·) ◦ η)(v) = σ(s′ ∧ η(v)) = σ(s′ ∧ (es)) = σs′ ∧ (eσs′) = 0. Since X is contained in no hypersurface of degree 3, the above implies that ∂v(F ) = 0 in Sym3(H 0(OX (H))). Equivalently, F ∈ Sym4(ker(v)), i.e., the quartic hypersurface V4 is the cone in P(H 0(OX (H))∗) with vertex [v] and base the quartic surface S defined by F in P(ker(v)∗) ∼= P3. (cid:3) In the course of the proof of the preceding lemma we introduced the divisorial part Σ of the section s ∈ H 0(NX(−KX − H)) defined by V4. Geometrically, Σ is the 1-dimensional part of the singular locus of V4 contained in X. Our next result shows Lemma 10.4. Σ = 0. Proof of Lemma 10.4. Assume Σ is not zero. Then we claim that Σ is composed of (−1)- curves which are lines with respect to H. Indeed, let C be a reduced, irreducible component of Σ and let γ be a global section defining C. Then s = s′γ, with s′ ∈ H 0(NX (−KX − H − C)). This section gives the cohomology class δX (s′) = ξ′ ∈ H 1(ΘX(−KX − H − C)) which is related to ξ by the formula γξ′ = ξ. In particular, ξ′ is nonzero and gives rise to a nontrivial extension 0 −→ OX (−KX − H − C)) −→ Tξ′ −→ ΩX −→ 0. (10.6) In contrast to the homomorphism induced by the extension (10.1), the homomorphism H 0(Tξ′) −→ H 0(ΩX ) (10.7) induced by the epimorphism in (10.6) fails to be an isomorphism, since otherwise the argu- ment of the first part of the proof gives a decomposition KX = L′ +E′, with L′ effective, while Lemma 2.3, 3), applied to ξ′ = δX (s′) yields −(L′ + C) · H ≥ 0, an obvious contradiction. The failure of the homomorphism (10.7) to be an isomorphism implies that H 1(OX (−KX− H − C)) 6= 0. But from the exact sequence 0 −→ OX (−KX − H − C) −→ OX(−KX − H) −→ OC (−KX − H) −→ 0 and Lemma 10.2, it follows H 0(OC (−KX − H)) ∼= H 1(OX (−KX − H − C)) 6= 0. Hence (KX + H) · C ≤ 0. (−1)-curves with H · C = 1. In the proof of Lemma 10.2 such a curve C is identified as a Next we wish to identify the configuration of the lines composing Σ inside the quartic cone V4. We begin by observing that no line in Σ can be a ruling of the cone, since otherwise such a ruling L must be contained in the singular locus of V4 and hence connect the vertex 76 10. Albanese dimension [v] to a singular point of a base S of the cone; but then the linear system H − L must have L in its base locus which is absurd since H − L is base point free -- the linear system H − L corresponds to the projection of X from the line L and this is a morphism X → P2. Once we know that the lines composing Σ are not rulings of V4, we deduce that their projection from the vertex [v] are lines contained in the singular locus of a base S of the cone. Thus each line C of Σ is contained in the plane PC spanned by C and [v]. The plane PC is contained in the singular locus of V4. In addition, since the lines composing Σ are disjoint20, C is the only component of Σ contained in PC . From this it follows that the one dimensional part of X ∩ PC is the line C. This means that the pencil of hyperplanes {Vt}t∈P1 in P4 cutting out the plane PC, intersects X along reducible divisors Ht = Vt · X containing the line C with multiplicity at least 2. From this it follows that the plane PC is the embedded tangent plane of X for every x ∈ C. But this contradicts the fact that the Gauss map of X is finite, see [36]. (cid:3) The last lemma implies that the quartic surface S in Lemma 10.3 has at most isolated singularities and the hypersurface V4 is the cone over S with vertex [v]. Our surface X is a smooth divisor in the cone V4 and the projection from the vertex [v] defines a morphism onto a normal quartic surface S ⊂ P3. In particular, the degree of pv is given by the formula pv : X −→ S deg(pv) =( d 4 , d−1 4 , if [v] /∈ X if [v] ∈ X. Since the singular locus of V4 is the union of its rulings joining [v] to the points of Sing(S), the singular locus of S, and by Lemma 10.4 none of these is contained in X, we deduce that the zero-locus Z0 = (s = 0) of the section s ∈ H 0(NX(−KX − H)) defined by V4 is 0-dimensional of degree deg(Z0) =(deg(Sing(S)) · d 4 , deg(Sing(S)) · d−1 4 + 1, if if [v] /∈ X, [v] ∈ X. The above formulas and the well known fact that deg(Sing(S)) ≤ 16 imply deg(Z0) ≤ 4d. (10.8) Next we relate the normal sequence of X with the extension sequence (10.1) defined by ξ = δX (s) ∈ H 1(ΘX (−KX − H)). For this we write the section s as a morphism of sheaves OX (KX + H) s−→ NX. Applying Ext1(•, ΘX ), we have Ext1(NX, ΘX ) s−→ Ext1(OX (KX + H), ΘX ). (10.9) (10.10) In particular, the extension class n ∈ Ext1(NX, ΘX) corresponding to the normal sequence 0 −→ ΘX −→ ΘP4X −→ NX −→ 0 20If there are two intersecting lines C and C ′ in Σ, then (C + C ′)2 = 0 and this contradicts the assumption that X is of Albanese dimension 2. 77 10. Albanese dimension goes under the homomorphism in (10.10) to the extension class n · s. But the coboundary map δX : H 0(NX(−KX − H) → H 1(ΘX (−KX − H)) is precisely the cup-product with n. Hence n · s = δX (s) = ξ. This means that the morphism in (10.9) extends to a morphism of extensions 0 0 ΘX T ∗ ξ OX(KX + H) ΘX ΘP4 ⊗ OX s NX 0 0 where the sequence on the top is the dual of (10.1). This can be completed to the following commutative diagram 0 T ∗ ξ 0 OX(KX + H) ΘX ΘX ΘP4 ⊗ OX s NX s∧ 0 0 JZ0(4H) JZ0(4H) 0 0 0 0 (10.11) This diagram will enable us to control the subscheme Z in the destabilizing sequence (10.4) of Tξ. Lemma 10.5. The subscheme Z is either [v] or empty. In particular, deg(Z) ≤ 1. Proof of Lemma 10.5. Dualizing (10.4) and tensoring it with OX (−H), we obtain a section t ∈ H 0(T ∗ ξ (−H)) whose zero locus is Z. On the other hand the middle column of the diagram (10.11) tensored with OX (−H) tells us that t can be identified with the section τ ∈ H 0(ΘP4 ⊗ OX(−H)) corresponding to the vector v defining the vertex [v] of V4. In particular, the scheme Zτ = (τ = 0) ⊃ Z. To understand Zτ we consider the contraction with τ This implies that the image of τ : H 0(ΩP4(2) ⊗ OX ) −→ H 0(OX (H)). τ is contained in H 0(JZ(H)). But H 0(ΩP4(2) ⊗ OX ) ∼= 2H 0(OX (H)) and under the identification of τ with v, the above contraction becomes simply the contraction with v ∈ H 0(OX (H))∗, This implies v : 2H 0(OX (H)) −→ H 0(OX (H)). im( τ ) = im( v) = v⊥, i.e., im( τ ) is identified with the space of hyperplanes in P4 passing through [v]. Hence, Z ⊂ Zτ is either ∅ or [v]. (cid:3) 78 10. Albanese dimension We now have everything to rule out the existence of X. Indeed, from the destabilizing sequence of Tξ we have deg(Z) = c2(Tξ) = c2(T ∗ ξ ). This and the middle column in (10.11) give 10d = c2(ΘP4 ⊗ OX ) = c2(T ∗ ξ ) + deg(Z0) + 4d = deg(Z) + deg(Z0) + 4d. From this and Lemma 10.5 it follows deg(Z0) ≥ 6d − 1. Putting together this inequality and the upper bound for deg(Z0) in (10.8), we obtain 2d ≤ 1 which is clearly impossible. (cid:3) 10.2. The surfaces X ⊂ P4 with mX = 5 and of Albanese dimension 2 We now turn to the consideration of surfaces X ⊂ P4 of Albanese dimension 2 and mX = 5. To apply our method in this case we need the additional assumption of X being minimal. With this in mind we proceed as in the case mX = 4. Namely, let V5 be a quintic hypersurface containing X and let s be a nonzero global section of NX(−KX ) defined by V5. That section is used to obtain the cohomology class ξ = δX (s), which by Lemma 2.2 is nonzero. The class ξ is interpreted as a nontrivial extension 0 −→ OX (−KX ) −→ Tξ −→ ΩX −→ 0 (10.12) which we use to gain an insight into the geometry of X and V5. Lemma 10.6. X is either of general type or an abelian surface of degree d = 10. Proof. The assumption that X is of Albanese dimension 2 implies that KX is effective. The minimality of X ensures that KX is nef. Hence either K 2 X > 0 and X is of general type, or K 2 X = 0. In the latter case, we deduce that KX = 0, and by the Enriques-Kodaira classification, X is an abelian surface. From this and the double point formula, we obtain that X is of degree d = 10. (cid:3) Abelian surfaces of degree 10 are of course Horrocks-Mumford surfaces and it is well- known that they are contained in a quintic hypersurface and not in one of a smaller degree. From now on we assume that X is of general type. This assumption implies that the homo- morphism H 0(Tξ) −→ H 0(ΩX) induced by the epimorphism in (10.12) is an isomorphism. This allows us to define the saturation G of the subsheaf of Tξ generated by its global sections. Lemma 10.7. The following possibilities may arise: 1) The rank of G is 3 and then G = Tξ is generated by its global sections and K 2 ∼= O⊕3 X . In particular, the irregularity q = 3, ΩX X = c2. 2) The rank of G is 2. 79 10. Albanese dimension Proof. From the definition of G and the assumption that X is of Albanese dimension 2, it follows that the rank of G is at least 2. If rank(G) = 3, then one can choose a subspace V ⊂ H 0(G) of dimension 3 so that the evaluation morphism V ⊗ OX → G is generically an isomorphism. This, followed by the inclusion G ֒→ Tξ, gives a morphism V ⊗ OX −→ Tξ which is generically an isomorphism. Since det(Tξ) = OX , the above morphism must be an isomorphism and we obtain the identifications V ⊗ OX ∼= G = Tξ. This implies H 0(ΩX ) ∼= H 0(G) ∼= H 0(V ⊗ OX ) = V ∼= C3. Hence q = 3 and the exact sequence (10.12) takes the form 0 −→ OX (−KX) −→ V ⊗ OX −→ ΩX −→ 0. In particular, the above sequence implies that the cotangent bundle ΩX is generated by its global sections and the Chern numbers of X are subject to K 2 (cid:3) X = c2. We now investigate the case rank(G) = 2. Lemma 10.8. If rank(G) = 2, then q = h0(G) = 2 or 3. Furthermore, if q = 3, then the following possibilities may arise: 1) det(G) = OX(H). 2) h0(det(G)) = 2 and X admits a fibration p : X → B onto a smooth curve B of genus gB = 2 such that det(G) = OX (p∗KB + R), where KB is the canonical divisor of B and R is the fixed part of the linear system det(G). Proof. The condition rank(G) = 2 allows us to apply Lemma 2.3 and deduce: i) the decomposition of the canonical divisor KX = L + E, where L = c1(G) and E is an effective nonzero divisor supported on the cokernel of the morphism ϕ : G → ΩX, ii) a nonzero global section τ ∈ H 0(ΘP4 ⊗ OX(−L)). To analyse the situation further, we take a nonzero global section g of G and write the exact sequence 0 −→ OX (F ) −→ G −→ JA(L − F ) −→ 0, (10.13) associated to g. In this sequence, F is the divisorial part of the scheme Zg = (g = 0), A is the 0-dimensional subscheme obtained after dividing g by a section of OX (F ) corresponding 80 10. Albanese dimension to F , and JA is the ideal sheaf of A. The assumption that the Albanese dimension of X is 2 implies that the homomorphism H 0(G) −→ H 0(JA(L − F )) induced by the epimorphism in (10.13) is nonzero. In particular, h0(OX (L − F )) ≥ h0(G) − h0(OX (F )) = q − h0(OX (F )) ≥ 1 and we have the estimate h0(OX (L)) ≥ h0(OX (F )) + h0(OX (L − F )) − 1 ≥ h0(OX (F )) + q − h0(OX (F )) − 1 = q − 1, (10.14) implying h0(OX (L)) ≥ 2 if q ≥ 3. From now on we assume q ≥ 3 and write L = M + R, (10.15) where M and R are the moving and the fixed part of L respectively. We divide our considerations according to the dimension h0(OX (L)) = h0(OX (M )). Case h0(OX (L)) = h0(OX (M )) ≥ 3. The nonzero global section τ ∈ H 0(ΘP4 ⊗OX(−L)), see the beginning of the proof, gives rise to a nonzero global section of ΘP4 ⊗ OX(−M )). Since M is nef and big21, the Euler sequence for ΘP4 ⊗ OX tensored with OX (−M ) implies H 0(OX (H − M )) 6= 0. We claim that M = H. Indeed, assuming Γ = H − M 6= 0, we have H 0(OX (H − Γ)) = H 0(OX (M )) ∼= H 0(OX (L)). This implies h0(OX (H −Γ)) = h0(OX (L)) ≥ 3. Hence Γ must be a line and the last inequality must be an equality. Since H = M + Γ = L − R + Γ = KX + Γ − E − R, taking the intersection with Γ on both sides, we deduce 1 = H · Γ = (KX + Γ) · Γ − (E + R) · Γ = −2 − (E + R) · Γ. From this, it follows R · Γ = −3 − E · Γ. In addition, if Γ is not in E, then E · Γ ≥ 0, and if Γ is an irreducible component of E, then E · Γ ≥ −2, see [26]. This and the above identity imply R · Γ ≤ −1. Hence Γ is a component of R and we can rewrite the relation (10.15) as follows L = M + R = M + Γ + R′ = H + R′, where R′ is an effective divisor. But this gives the contradiction 3 = h0(OX (L)) ≥ h0(OX (H)) = 5. 21M is not composed of a pencil, since otherwise the image of the Albanese map is 1-dimensional. 81 10. Albanese dimension Once we know that M = H, the identity (10.15) reads L = H + R and we claim that R = 0. Indeed, if this is not the case, we take C, a reduced, irreducible component of R and use the fact that H 0(ΘP4 ⊗ OX (−H − C)) 6= 0. This together with the Euler sequence implies H 1(OX (−H − C)) 6= 0. But from the exact sequence 0 −→ OX (−C) −→ OX −→ OC −→ 0 tensored with OX (−H), we see that H 1(OX (−H − C)) ∼= H 0(OC (−H)) = 0. Next we show that q = 3. For this we go back to the exact sequence (10.13) which now takes the form 0 −→ OX (F ) −→ G −→ JA(H − F ) −→ 0. (10.16) Furthermore, we may assume that there are two linearly independent global sections g and g′ of G which are proportional, i.e., the exterior product g ∧ g′ is zero, viewed as a section of det(G) (otherwise we are done by [26, Lemma 5.4]). With such a choice of g ∈ H 0(G) in constructing the exact sequence (10.16), we obtain that the line bundle OX(F ) in that sequence has h0(OX (F )) ≥ 2. From the isomorphism H 0(G) ∼= H 0(ΩX ) it also follows that the sections g, g′ correspond to two linearly independent holomorphic 1-forms, call them ω, ω′, subject to ω ∧ ω′ = 0 as a section of OX(KX ). By Castelnuovo -- de Franchis theorem, it follows that X admits a morphism p : X −→ B (10.17) onto a smooth curve B of genus gB ≥ 2 such that ω = p∗(η) and ω′ = p∗(η′), where η, η′ ∈ H 0(OB(KB)). Hence p∗(OB(KB)) = OX (F ). (10.18) Furthermore, from the assumption on the Albanese dimension of X we know that q > gB. Hence H 0(JA(H − F )) ⊃ coker(H 0(OX (F )) −→ H 0(G)) 6= 0. In particular, Γ = H − F is an effective nonzero divisor. From this it follows that h0(OX (F )) = h0(OX (H − Γ)) ≤ 3, with the equality holding if and only if Γ is a line in P4. We claim that the equality is impossible. Indeed, if Γ is a line, then F = H − Γ is base point free and hence, by (10.18), the divisor F is composed of the fibres of the morphism p in (10.17). But Γ is a rational curve and hence, must be contained in a fibre of that morphism. Then 0 = F · Γ = (H − Γ) · Γ = 1 − Γ2. Since Γ2 < 0, the above identity is an obvious contradiction. Thus h0(OX (F )) ≤ 2. This and the hypothesis h0(OX (F )) ≥ 2 imply h0(OX (F )) = 2. which, together with (10.18) leads to gB = 2. Next we claim that h0(OX (H − F )) = 1. Indeed, assume h0(OX (H − F )) ≥ 2. This means that the divisors of the linear system F = p∗(KB) are contained in a plane in P4. 82 A. The projective bundle P(NX (−3H)) and the embedding X ⊂ P4 But a general divisor of F contains two disjoint irreducible curves and this can not happen for plane curves. The case h0(OX (L)) = h0(OX (M )) = 2. In this case the estimate (10.14) tells us that q ≤ 3 and hence, q = 3, in view of the assumption q ≥ 3. Furthermore, the natural homomorphism 2H 0(G) −→ H 0(det(G)) = H 0(OX (L)) has a nontrivial kernel and this implies, as in the previous case, that the sheaf OX(F ) in (10.13) satisfies h0(OX (F )) ≥ 2. By the hypothesis on the Albanese dimension, the sheaf JA(L−F ) in that exact sequence must have h0(JA(L−F )) ≥ 1. This and the first inequality in (10.14) tells us that h0(OX (F )) ≤ h0(OX (L)) = 2. Hence h0(OX (L)) = h0(OX (F )) = 2 and h0(O(L − F )) = 1. From the isomorphism H 0(G) ∼= H 0(ΩX) we also conclude that H 0(OX (F )) defines a two dimensional subspace of H 0(ΩX) contained in the kernel of the natural homomorphism 2H 0(ΩX) −→ H 0(det(ΩX)) = H 0(OX (KX )). By Castelnuovo -- de Franchis theorem, this means that X admits a morphism p : X → B onto a smooth curve B of genus gB ≥ 2. This together with the hypothesis on the Albanese dimension of X imply 2 ≤ gB < q = 3. Hence gB = 2 and OX (F ) = p∗(OB(KB)). In addition, from h0(OX (F )) = h0(OX (L)) = 2, we deduce the formula L = p∗KB + R, where R is the fixed part of L. (cid:3) We summarize the above discussion in the following statement. Theorem 10.9. If X ⊂ P4 is a minimal surface with mX = 5 and of Albanese dimension 2, then its irregularity q(X) = 2 or 3. Furthermore, if q(X) = 3, then one of the following possibilities may occur: 1) The cotangent bundle ΩX is generated by its global sections and K 2 2) The canonical divisor KX admits the decomposition KX = L+E, with L and E effective X = c2, nonzero divisors. The decomposition is subject to the following properties: a) The divisor L is either equal to H, or X admits a fibration p : X → B over a smooth curve B of genus 2 and L = p∗(KB)+R, where KB is the canonical divisor of B and R is the fixed part of the linear system L. b) L · H ≤ H 2 = d and K 2 X − c2 ≤ 2 3 L2. Proof. The assertions 1) and 2), a), are Lemma 10.7 and Lemma 10.8 respectively. The (cid:3) assertion b) is Lemma 2.3, 3) and c) is as in Theorem 4.4, ii). A. The projective bundle P(NX(−3H)) and the embedding X ⊂ P4 In this appendix we return to an elliptic scroll X of degree 5 in P4. we established an isomorphism between the space of global sections of N ∗ In Theorem 8.1 X(3H) and the 83 A. The projective bundle P(NX (−3H)) and the embedding X ⊂ P4 space of cubic hypersurfaces IX(3) containing X, see Section 8 for notation. The subscheme Zs = (s = 0) of zeros of a nonzero global section s of N ∗ X(3H) is identified with the scheme- theoretic intersection of X with the singular locus Sing(V3(s)) of the cubic hypersurface V3(s) corresponding to s under the isomorphism H 0(N ∗ X (3H)) ∼= IX (3). (A.1) This, in particular, allows for a purely geometric way to recover X from the space IX(3) -- a geometric counterpart of a well known algebraic fact that the homogeneous ideal of X is generated in degree 3. More conceptually, the isomorphism (A.1) suggests a sort of 'duality' between N ∗ X(3H) and X embedded in P4 by OX (H). This is the main theme of this section. Of course, one as- pect of the above mentioned duality is well-known -- the famous quadro-cubic transformation of Cremona relating a normal elliptic quintic curve in P4 with an elliptic scroll (in another copy of P4) see [4, 5]. So we do not claim any novelty in the results exposed here. However, placing the vector bundle N ∗ X(3H) in the center of the study, revisiting various aspects of the geometry of the scroll X and of the Segre cubics containing it via the properties of N ∗ X(3H), seem to be new and fruitful. Set E = NX(−3H). The projectivization Y = P(E) with the structure projection p : Y = P(E) −→ X (A.2) is equipped with the line bundle OY (1) chosen so that the direct image p∗OY (1) = E ∗ = N ∗ X(3H). By Lemma 8.2, the vector bundle E ∗ is globally generated. Hence OY (1) is globally generated and defines a morphism ϕ : Y = P(E) −→ P(H 0(E ∗)∗) (A.3) where the target space is, in view of Theorem 8.1, 2), a 4-dimensional projective space. We let V := H 0(OX (H)) and W := H 0(E ∗) = H 0(N ∗ X(3H)) and want to keep a clear distinction between the following two geometric incarnations of the projective space P4: • the projective space P(V ∗), where X lives, • the projective space P(W ∗), the target of the morphism ϕ defined in (A.3). One of our goals here is to describe how to go between these two spaces. A.1. On the geometry of the scroll X To describe the geometry of the morphism ϕ (resp. of the embedding X ⊂ P(V ∗)) we recall that X is a P1-bundle over an elliptic curve which will be denoted by E and we let be the structure projection. It is well known that X can be identified with Sym2(E), the second symmetric power of E. Then the structure projection becomes the Abel-Jacobi map π : X −→ E (A.4) 84 A. The projective bundle P(NX (−3H)) and the embedding X ⊂ P4 which takes a subscheme D ⊂ E of degree 2, viewed as a point of Sym2(E), to (the iso- morphism class of) the line bundle OE(D), viewed as a point of E. We are making here a (non-canonical) identification of E with Jac2(E), the variety of isomorphism classes of line bundles of degree 2 on E. From this it follows that X admits the obvious double covering τ : E × E −→ Sym2(E) = X (A.5) which sends a point (e, e′) ∈ E × E to τ (e, e′), the subscheme of E of degree 2 supported on e and e′. This equips X with a distinguished family of curves Γe := τ (ν−1 1 (e)) = τ (ν−1 2 (e)), (A.6) where e ∈ E and νi denotes the projection of E × E on the i-th factor, for i = 1, 2. These curves will play an important role in the sequel, so we mention some of their properties, immediate consequences of the definition (A.6). i) For every e ∈ E, the curve Γe is a section of the projection π in (A.4). ii) Γe · Γe′ = 1, for every e, e′ ∈ E. iii) h0(OX (Γe)) = 1. (A.7) The curves {Γe} e∈E are used to define embeddings of X into P4. Proposition A.1. Let OE(D) be a line bundle of degree 2 on E. Then for every point e ∈ E the line bundle OX (Γe + π∗(D)) is very ample and it defines an embedding of X into P4 as a scroll of degree 5. Proof. Let l be the class of a fibre of π : X → E in the N´eron-Severi group of X. The canonical divisor has the form KX = −2Γe + l and we set H := Γe + π∗(D) = KX + (H − KX ), (A.8) where H − KX = 3Γe + l is nef and big. The very ampleness of H now follows easily from [29, Theorem 1, 2)]. From the formula (A.8), we see that hi(OX (Γe + π∗(D)) = 0, for i = 1, 2. Hence by Riemann-Roch we obtain h0(OX (Γe + π∗(D)) = χ(OX (Γe + π∗(D)) = 1 2(cid:0)(Γe + 2l)2 − (Γe + 2l) · (−2Γe + l)(cid:1) = 5. Thus OX (H) embeds X into P4. Moreover, since H · l = (Γe + 2l) · l = 1, it follows that OX(H) embeds X as a scroll of degree H 2 = (Γe + 2l)2 = Γ2 e + 4Γe · l = 1 + 4 = 5. (cid:3) From now on we fix a point o ∈ E and a line bundle OE(D) of degree 2 on E. According to the previous proposition, this gives an embedding of X into P(V ∗) defined by OX (H), where H := Γo + π∗(D) and V = H 0(OX (H)). Proposition A.2. Under the embedding of X defined by OX (H), the curves {Γe}e∈E are embedded as plane curves of degree 3. Conversely, a plane cubic curve contained in X is one of the curves of the family {Γe}e∈E. 85 A. The projective bundle P(NX (−3H)) and the embedding X ⊂ P4 Proof. From OX (H − Γe) = OX (Γo + π∗(D) − Γe) = π∗OE(D + (o − e)) it follows that h0(OX (H − Γe)) = h0(π∗OE(D + (o − e))) = h0(OE(D + (o − e))) = 2. This means that under the embedding by OX (H), the image of Γe is contained in a plane. Its degree is computed by the intersection number H · Γe = (Γo + π∗(D)) · Γe = 1 + 2 = 3, where the properties i) and ii) in (A.7) are used. Conversely, let C be a plane cubic contained in X. Then h0(OX (H − C)) = 2 and H · (H − C) = 2. Hence, the effective divisor H − C is composed of rational curves. Since the only rational curves on X are the rulings (the fibres of π) of X, we deduce the linear equivalence for some effective divisor D′ of degree 2 on E. Hence H − C ∼ π∗(D′), C ∼ H − π∗(D′) = Γo + π∗(D − D′) = Γe, where e = o+e′ and e′ is the point of E corresponding to OE(D −D′) under the identification E = Jac0(E). Since h0(OX (Γe)) = 1, see the property iii) in (A.7), it follows that C = Γe. (cid:3) Remark A.3. Viewing X as a subvariety of P(V ∗), one obtains the family {Γe}e∈E as follows: consider two rulings la and la′ of X lying over the points a and a′ of E and consider the hyperplane Va,a′ in P(V ∗) spanned by the rulings la and la′. The hyperplane section Ha,a′ = Va,a′T X is a reducible divisor in H of the form Ha,a′ = C + la + la′, where C is the component of Ha,a′ of degree H · C = 3, residual to la + la′. Hence it must be irreducible and a section of the projection π : X → E. Therefore, C is a plane cubic contained in X and, in view of Proposition A.2, it must be a curve in {Γe}e∈E. The first part of the proof of Proposition A.2 shows that every curve in {Γe}e∈E is obtained in this way. The family of curves {Γe}e∈E gives rise to the family of planes {Πe}e∈E, where Πe is the plane spanned by Γe. To understand the properties of this family of planes we have the following. Proposition A.4. Let F := ν1∗(τ ∗OX (H)). Then F is a rank 3 vector bundle on E of degree 5, generated by its global sections with h0(F) = 5. Proof. By definition, rank(F ⊗ OE,e) = h0(OΓe(H)) = 3 for every e ∈ E. Hence F is locally free of rank 3. By Riemann-Roch deg(F) = χ(F) = χ(ν1∗(τ ∗OX (H))) = χ(τ ∗OX (H)) = H 2 = 5. Furthermore, H 0(F) ∼= H 0(τ ∗OX (H)) ∼= H 0(OX (H)). Hence h0(F) = h0(OX (H)) = 5. The global generation of F follows from the identifications H 0(F) ∼= H 0(OX (H)) −→ H 0(OΓe (H)) ∼= H 0(F ⊗ OE,e) and the fact that, for every e ∈ E, the above restriction homomorphism is surjective. (cid:3) 86 A. The projective bundle P(NX (−3H)) and the embedding X ⊂ P4 Set T := P(F ∗) to be the projectivization of the dual of F and ρ : T → E the structure projection. Let OT (1) be chosen so that ρ∗OT (1) = F. In view of Proposition A.4, the line bundle OT (1) is generated by its global sections and hence gives a morphism ψ : T −→ P(V ∗) By definition, the fibres of T are mapped by ψ to the family of planes {Πe}e∈E. In particular, we obtain Proposition A.5. The image T ′ of ψ is a hypersurface of degree 5 containing X in its singular locus. Proof. From the definition of ψ it follows deg(ψ) deg(T ′) = c3 1(OT (1)) = deg(F) = 5, where the last equality comes from Proposition A.4. The above implies deg(ψ) = 1 and deg(T ′) = 5. The hypersurface T ′ obviously contains X. To see that X is contained in its singular locus, we observe that the planes Pe and Pe′, for e 6= e′ ∈ E, intersect at a single point. Indeed, otherwise the hyperplane spanned by PeS Pe′ gives rise to a divisor D ∈ H of the form D = Γe + Γe′ + R where R is some effective divisor. The divisor D has a wrong degree on the rulings of X. The point of intersection PeT Pe′ is obviously a singular point of T ′. Furthermore, this point is the point of intersection Γe · Γe′ and hence belongs to X. Varying e and e′, the points Γe · Γe′ form a Zariski dense open subset of X. Therefore, X is contained in the singular locus of T ′. (cid:3) Besides being the union of the family of planes {Πe}e∈E, the variety T ′ in Proposition A.5 can be also characterized as follows. Proposition A.6. The variety T ′ in Proposition A.5 is the union of the proper trisecant lines of X ⊂ P(V ∗). ∨ Proof. Let Π e be the dual of the plane Πe. Then the family of the dual planes {Π e }e∈E gives rise to a family of proper trisecants of X in P(V ∗). We need to check that every proper trisecant of X belongs to this family. Let l be a proper trisecant line of X, i.e., Zl := l · X is a 0-dimensional subscheme of X of degree at least 3. Then ∨ h1(JZl(H)) = deg(Zl) − 2 ≥ 3 − 2 = 1. This and the Serre duality Ext1(JZl(H), OX (KX )) ∼= H 1(JZl(H))∗ 6= 0 gives rise to a non- trivial extension sequence 0 −→ OX (KX ) −→ E −→ JZl(H) −→ 0. The sheaf E in the middle of that sequence is torsion free with det(E) = OX (H + KX ) and h0(E) ≥ h0(JZl(H)) − h1(OX (KX )) = 3 − 1 = 2. Hence the saturation of the subsheaf generated by global section of E is a torsion free subsheaf of E of rank one. Hence it has the 87 A. The projective bundle P(NX (−3H)) and the embedding X ⊂ P4 form JZ1(A), where A (resp. Z1) is an effective divisor (resp. 0-dimensional subscheme) on X, and gives rise to the following destabilizing sequence of E 0 −→ JZ1(A) −→ E −→ JZ2(KX + H − A) −→ 0. Combining this together with the defining extension sequence implies that there is an effective nonzero divisor Γ ∈ H − A passing through Zl. From h0(OX (H − Γ)) = h0(OX (A)) ≥ 2 and the fact that Γ can not be a line, it follows that the equality must hold and therefore, Γ is a plane curve. Hence Γ is a plane cubic section of X and, in view of Proposition A.2, Γ = Γe, for some e ∈ E. This implies Zl = l · Γe ⊂ Πe and hence the line l correspond to a point in the dual plane Π (cid:3) ∨ e . A.2. From the morphism ϕ in (A.3) to X ⊂ P(V ∗) The above discussion gives a clear geometric picture on the side of the embedding X ⊂ P(V ∗). We now turn to the other side, the projectivization P(E) and the morphism introduced in (A.3), where W = H 0(E ∗). ϕ : Y = P(E) −→ P(W ∗) Proposition A.7. For ϕ : Y = P(E) → P(W ∗), the following statements hold. 1) The 3-fold Y contains a distinguished divisor Σ ∼= E × E and pΣ : E × E ∼= Σ −→ X ∼= Sym2(E), the restriction to Σ of the structure projection p in (A.2), is the double covering τ defined in (A.5). 2) Let ∆E denote the diagonal of E × E. Then the image ϕ(E × E) = ϕ(∆E) is an embedding of E as an elliptic normal curve of degree 5. 3) The image Y0 = ϕ(Y ) ⊂ P(W ∗) is a quintic hypersurface which is the secant variety of the elliptic normal curve ϕ(∆E). 4) The composition π ◦ p : Y → E is a fibration whose fibres (π ◦ p)−1(e) = Se are isomorphic to the Hirzebruch surface Σ1 = P(OP1 ⊕ OP1(−1)). For every e ∈ E, the restriction ϕFe : Se → P(W ∗) is an embedding. The image S′ e := ϕ(Se) is a rational cubic scroll containing the elliptic curve ϕ(∆E ). This curve is the image under ϕ of the intersection Σ · Se. The intersection, viewed as a divisor in Se, has the form 2Le + 3f , where Le is a unique section of Se with L2 e = −1 and f is the class of a fibre of p. In particular, the image L′ e := ϕ(Le) is a line intersecting ϕ(∆E) transversely at a single point. 5) The projection p : Y → X admits a distinguished section γ : X → Y . The image of the composition ϕ ◦ γ : X → P(W ∗) is a singular scroll of degree 15 containing ϕ(∆E). For the convenience of the reader, the following diagram summarizes all the morphisms appearing in (the proof of) the proposition and we suggest to consult this diagram while 88 A. The projective bundle P(NX (−3H)) and the embedding X ⊂ P4 advancing through the proof. ∆E µ = ϕ∆E P(E) = Y P(W ∗) ϕ P(V ∗) i p E × E τ X π E Proof. We begin by showing that the natural double covering τ : E × E −→ Sym2(E) ∼= X admits a lifting to Y = P(E), i.e., that there is a morphism i : E × E → Y which fits into the following commutative diagram Y = P(E) i τ p X E × E (A.9) It is well known that this amounts to having a line bundle M on E × E together with a surjective morphism τ ∗(E ∗) −→ M, see e.g., [16], Ch. investigate the restriction of E to the curves {Γe}e∈E in (A.6). II, Proposition 7.12. In order to construct M and the morphism, we Lemma A.8. For every e ∈ E, the restriction of E to Γe has the form E ⊗ OΓe = NX(−3H) ⊗ OΓe = OΓe ⊕ OΓe(KX − H). Proof of Lemma A.8. Set Γ = Γe. In view of Proposition A.2, the curve Γ is embedded into P(V ∗) as a plane curve of degree 3. Hence its normal bundle NΓ/P(V ∗) in P(V ∗) has the form NΓ/P4 ∼= OΓ(H) ⊕ OΓ(H) ⊕ OΓ(3H). Using this decomposition in the short exact sequence of normal bundles of the inclusions Γ ⊂ X ⊂ P(V ∗), we obtain 0 −→ OΓ(Γ) −→ OΓ(H) ⊕ OΓ(H) ⊕ OΓ(3H) −→ NX ⊗ OΓ −→ 0. This gives a nonzero morphism OΓ(3H) → NX ⊗ OΓ or, equivalently, a nonzero global section, call it η, of NX(−3H) ⊗ OΓ. This together with the fact that the dual vector bundle N ∗ X(3H) ⊗ OΓ is globally generated, see Theorem 8.1, implies that η is nowhere vanishing and gives a splitting E ⊗ OΓ = NX(−3H) ⊗ OΓ = OΓ ⊕ OΓ(KX − H) as asserted. (cid:3) 89 A. The projective bundle P(NX (−3H)) and the embedding X ⊂ P4 We now take the pullback τ ∗(E) and recall that the curve Γe, for every e ∈ E, is the image under τ of the fibre ν−1 1 (e) of the projection ν1 : E × E → E. This and the decomposition of Lemma A.8 imply that the direct image ν1∗(τ ∗(E)) is a line bundle, call it L. This can 1 (L−1) ⊗ τ ∗(E)) which in turn gives a be interpreted as a nowhere vanishing section of ν1∗(ν∗ 1 (L−1) ⊗ τ ∗(E) or, equivalently, the inclusion of bundles nowhere vanishing global section of ν∗ ν∗ 1 (L) ֒→ τ ∗(E). Dualizing gives the sought after epimorphism τ ∗(E ∗) −→ ν∗ 1 (L−1) −→ 0 (A.10) and hence the diagram (A.9). Furthermore, under the morphism i : E × E → Y associated to the above epimorphism, the line bundle ν∗ 1 (L−1) becomes the pullback i∗(OY (1)). Hence the composition morphism ϕ ◦ i : E × E −→ P(W ∗) factors through the morphism (A.11) defined by L−1. Since the projection ν1 restricted to the diagonal ∆E induces the isomorphism of ∆E with the base E (of the first projection), we deduce µ : E −→ P(W ∗), (ϕ ◦ i)(E × E) = (ϕ ◦ i)(∆E). Geometrically, the morphism i in (A.9), corresponding to the epimorphism (A.10), maps 1 (e) ∼= Γe to the section of the ruled surface P(E ⊗ OΓe) defined by the trivial the fibre pr−1 summand of the decomposition E ⊗ OΓe = OΓe ⊕ OΓe(KX − H) described in Lemma A.8. Under the morphism ϕ that section is contracted to the point µ(e), while the ruled surface P(E ⊗ OΓe ) is mapped by ϕ to the cone over a normal quartic elliptic curve22 with vertex at µ(e). From this and the properties of N ∗(3H) we can draw several conclusions. a) The image µ(E) of the morphism µ in (A.11) can not be contained in a plane. In- deed, otherwise a plane containing µ(E) gives rise to a pair of linearly independent global sections of E ∗ = N ∗(3H) which are proportional, but such sections must have 1-dimensional zero locus and this, according to Theorem 8.1, 3), is impossible. b) The degree of µ onto its image is 1. Indeed, otherwise take two general points x and x′ of µ(E) and let e1 and e2 (resp. e′ 2) be two distinct points lying in the preimage µ−1(x) (resp. µ−1(x′)); now take a plane P passing through the points x and x′ chosen on µ(E). Then all rulings of the cones ϕ(P(E ⊗ Γei)) and ϕ(P(E ⊗ Γe′ )), for i = 1, 2, intersect P . From the identification 1 and e′ i W = H 0(N ∗(3H)) ∼= H 0(OY (1)), we deduce that two hyperplanes in P(W ∗) cutting out the plane P give rise to two linearly independent global sections, say s and s′, of N ∗(3H) such that γ = s ∧ s′, viewed as a global section of det(N ∗(3H)) = OX (H − KX ). The section γ vanishes along the divisor Γe1 + Γe2 + Γe′ 1 + Γe′ 2 = 4Γo which is impossible, since the divisor H − KX − 4Γo = −Γo + l can not be numerically equivalent to an effective divisor. 22The base of the cone is the image of Γe by OΓe (H − KX ) which is a line bundle of degree (H − KX) · Γe = 3 + 1 = 4. 90 A. The projective bundle P(NX (−3H)) and the embedding X ⊂ P4 Let E′ be a Zariski dense open subset of E, where µ is an embedding. Take two distinct points e1 and e2 on E′. The corresponding curves Γe1 and Γe2 intersect at a point, call it x. From the properties established above, we know that the fibre Lx of the projection p : Y → X is mapped by ϕ into the line joining µ(e1) and µ(e2). Hence the secant variety of µ(E′) is contained in the image Y0 = ϕ(Y ). This implies, in view of the irreducibility of Y0, that the curve µ(E) spans P(W ∗). Furthermore, for e ∈ E′, the union of the secant lines of µ(E) passing through µ(e) must be the cone ϕ(P(E ⊗ Γe)) with the vertex µ(e) and base a smooth elliptic curve of degree 4. Hence this base curve is the image of the projection of µ(E) from the point µ(e). This implies that µ must be an embedding, µ(E) is an elliptic normal curve of degree 5, and the secant variety of µ(E) is contained in Y0. Since both are irreducible they must coincide. The assertion that the degree of Y0 is 5 can now be deduced either from the fact that the secant variety of an elliptic normal curve in P4 has degree 5 or from the direct computation deg(Y0) = c3 1(OY (1))) = c2 1(E ∗) − c2(E ∗) = (H − KX)2 − 10 = 15 − 10 = 5. The second equality uses the Chern invariants of E ∗ = N ∗ X(3H) computed in Theorem 8.1. The only remaining statement of Proposition A.7, 1) - 3), to prove is that i : E × E → Y in (A.9) is an embedding. From the work we have already done, this is immediate, since the image of i is a bi-section with respect to the projection p : Y −→ X and it is an isomorphism outside the diagonal ∆E. On the diagonal ∆E, we know that the composition ϕ ◦ i is an embedding. Hence i is an embedding everywhere. Turning to the part 4) of Proposition A.7, we observe that the fibre (π ◦ p)−1(e) = Se over e ∈ E is P(E ⊗ OFe), where Fe = π−1(e). This together with E ⊗ OFe = NX(−3H) ⊗ OFe ∼= OP1(−1) ⊕ OP1(−2) (see (8.3) for the last isomorphism) proves that Se is isomorphic to the Hirzebruch surface Σ1. Furthermore, the line bundle OY (1) restricted to Se is OSe(1) := OY (1) ⊗ OSe which is ∼= OP1(1) ⊕ OP1(2) is obviously very ample. very ample, since p∗(OSe(1)) = N ∗ The morphism ϕSe is defined by OSe(1) and hence it embeds Se into P(W ∗) as a scroll of degree deg(N ∗ e = ϕ(Se) is a rational normal scroll of degree 3 in P(W ∗). Setting Le to be the minimal section of Se, we deduce X(3H) ⊗ OFe) = deg(OP1(1) ⊕ OP1(2)) = 3, i.e., the image S′ X(3H) ⊗ OFe Le = c1(OSe(1)) − 2f, where f is the class of a fibre of p. In particular, Le is a (−1)-curve on Se and its image e = ϕ(Le) is a line in P(W ∗). The rulings of S′ L′ e cut out on the elliptic curve ϕ(∆E) the pencil of degree 2 corresponding to the fibre Fe ⊂ X ∼= Sym2(E). Hence ϕ(∆E) is a divisor on S′ e + bl for some integer b, and where l, the image of f , is the class of a ruling of S′ e of the form 2L′ e. The integer is determined from the equation e(1)) · (2L′ 5 = deg(ϕ(∆E )) = c1(OS ′ e + bl) = 2 + b. Hence ϕ(∆E) = 2L′ e + 3l. From this it follows that e · ϕ(∆E) = L′ L′ e · (2L′ e + 3l) = −2 + 3 = 1. Thus L′ e intersects ϕ(∆E) transversely at a single point. The minimal section Le of Se is a distinguished section of p over the fibre Fe = π−1(e). Varying e in E gives rise to the section γ of p whose existence is claimed in the part 5) of the 91 A. The projective bundle P(NX (−3H)) and the embedding X ⊂ P4 proposition. To be more formal, we seek a sub-linebundle of E = N (−3H) which coincides with OFe(−1) on every fibre Fe of π. The construction is the same as in the proof of part ∼= OP1(−1) ⊕ OP1. Hence, the 1) of the proposition. Namely, we know that N (−2H) ⊗ OFe direct image π∗(N (−2H)) is a line bundle on E which will be denoted OE(D). This gives rise to the exact sequence 0 −→ π∗(OE (D)) −→ N (−2H) −→ OX (KX + H − π∗(D)) −→ 0 and hence, the sought after sub-linebundle of N (−3H) is OX (−H + π∗D). Furthermore, the second Chern class computation from the above exact sequence yields deg(D) = −5. In addition, by construction, the section γ : X −→ Y corresponding to the subbundle OX(−H + π∗D) ⊂ N (−3H) has the property OX (H − π∗D) = γ∗(OY (1)). Hence the image X′ = (ϕ ◦ γ)(X) has degree deg(X′) = (H − π∗D)2 = (H + 5f )2 = 5 + 10 = 15. (cid:3) Remark. All the facts in Proposition A.7, with the possible exception of identifying the It seems to us that taking the vector bundle X(3H) on the scroll X as the starting point makes its relation with the associated P1-bundle P(E), have been proved in [4]. E ∗ = N ∗ elliptic quintic curve more natural. The property of P(E) being a fibration by Hirzebruch surfaces Σ1, which comes almost for free in our exposition, is all one needs to establish the relationship with the space of quadrics passing through the elliptic normal quintic curve ϕ(∆E) and eventually, show that P(E) and P(Nϕ(∆E) ⊗ OP(W ∗)(−2)) are related by a birational morphism23, where Nϕ(∆E) is the normal bundle of ϕ(∆E) in P(W ∗). See [4] for details. With the above description of the morphism ϕ : Y → P(W ∗) we can now explain how to go from ϕ back to the embedding X ⊂ P(V ∗). For this, we use the following notation. • hz, z′i is the line through the distinct points z, z′ ∈ P4 • Yx = p−1(x), the fibre of p over x ∈ X. Proposition A.9. The isomorphism W = H 0(N ∗(3H)) = H 0(E ∗) ∼= H 0(OY (1)), (A.12) establishes the following geometric correspondence between the hyperplanes in P(W ∗) and 0-cycles on X. Let M be a hyperplane in P(W ∗) intersecting the elliptic normal curve ϕ(∆E) along the divisor DM = M · ϕ(∆E) consisting of 5 distinct points24. Then the configuration of the ten secant lines of ϕ(∆E) he, e′i, Xe6=e′∈DM ZM = Xe6=e′∈DM gives rise to the 0-cycle p∗(Yτ (e,e′)) (A.13) on X, where ϕ(Yτ (e,e′)) = he, e′i. That 0-cycle is the scheme of zeros of a unique projective section [sM ] ∈ P(H 0(N ∗(3H))) corresponding to M under the isomorphism (A.12). 23The morphism is the relative blow-down of P(E ) along the image of the section γ in Proposition A.7, 5). 24In what follows we abuse the notation by tacitly identifying ϕ(∆E) with E. 92 A. The projective bundle P(NX (−3H)) and the embedding X ⊂ P4 Proof. Let [sM ] ∈ P(H 0(N ∗(3H))) be the projective section corresponding to a hyper- plane M . The the zero locus (sM = 0) ⊂ X parametrizes the fibres of the projection p : Y → X which are mapped onto the secant lines of ϕ(∆E) contained in M and those are precisely the ones appearing on the right side of the equality (A.13). (cid:3) We can now reconstruct25 X ⊂ P(V ∗) from the correspondence P(W ) ∋ M 7→ ZM = Xe6=e′∈DM p∗(Yτ (e,e′)) described in Proposition A.9. The 0-cycle ZM admits a distinguished decomposition into 5 subcycles Z e M where Z e p∗(Yτ (e,e′)), (A.14) ZM = [e∈DM M = Xe′∈DM r{e} M parametrizes the fibres of p : Y → X mapped by ϕ onto the rulings of the cone i.e., Z e ϕ(P(E ⊗ OΓe)) that are contained in M . Hence we have Z e M ⊂ Γe for every e ∈ DM . The configuration of 5 curves {Γe}e∈DM , on the side of P(V ∗), gives rise to 5 planes {Πe}e∈DM , where Πe is the span of Γe in the embedding X ⊂ P(V ∗). Now one recovers the Segre cubic V3(sM ), the cubic hypersurface corresponding to sM under the isomorphism H 0(N ∗(3H)) ∼= IX(3) in Theorem 8.1, as the union of the lines in P(V ∗) intersecting (any) four of the five planes of the collection {Πe}e∈DM . The 0-cycle ZM = (sM = 0) = Sing(V3(sM )) is seen in P(V ∗) as the cycle ZM = Xe6=e′∈DM Πe\ Πe′. composed of points of pairwise intersections of the planes {Πe}e∈DM . As M varies in the complement of the dual variety of the elliptic normal curve ϕ(∆E), the zero cycles ZM sweep out the scroll X. Let U be the Zariski dense open subset of P(W ) parametrizing regular global sections of N ∗(3H). From the construction above U coincides with the set of hyperplanes in P(W ∗) intersecting the elliptic normal quintic curve ϕ(∆E) ⊂ P(W ∗) along five distinct points. Thus the complement Tϕ(∆E ) := P(W ∗) \ U is the dual variety of ϕ(∆E), i.e., the closed points of Tϕ(∆E ) parametrize the hyperplanes in P(W ∗) which are tangent to ϕ(∆E) at some point. Furthermore, the construction described above and the identification P(W ) ∼= IX (3) (A.15) in Theorem 8.1, 3), imply that the Zariski open subset U , under the above identification, corresponds to Segre cubics in P(V ∗) containing X. Thus the dual variety Tϕ(∆E ), via the identification (A.15), parametrizes cubic hypersurfaces in IX(3) with degenerate singular locus. To see this degeneracy we take a hyperplane M in P(V ∗) which has a contact of order 25We are grateful to I. Dolgachev for pointing out to us the construction that follows; according to him, it was Segre's way to see an elliptic scroll inside a Segre cubic. 93 A. The projective bundle P(NX (−3H)) and the embedding X ⊂ P4 2 with ϕ(∆E ) at a point e0 and is transversal to ϕ(∆E) at the remaining three points which we denote ei, i = 1, 2, 3. Thus the divisor DM = M · ϕ(∆E) has the form DM = e0 + te0 + e1 + e2 + e3, where te0 denotes a tangent vector of ϕ(∆E) at e0. The corresponding configuration of secant lines of ϕ(∆E) contained in M consists of: • he0, eii, i = 1, 2, 3, with multiplicity 2, • the embedded tangent line he0, te0i of ϕ(∆E) at e0, • three lines hei, eji, 1 ≤ i < j ≤ 3. This accounts for ten secant lines counted with multiplicities. Since the scheme of zeros ZM of the projective section [sM ] ∈ P(W ) corresponding M under the isomorphism in (A.12) must parametrize the secants of ϕ(∆E) contained in M and deg(ZM ) = 10, we deduce that no other secant of ϕ(∆E) is contained in M and the 0-cycle ZM has the form ZM = 2 3Xi=1 τ (e0, ei) + τ (e0, e0) + X1≤i<j≤3 τ (ei, ej ), (A.16) where τ : E × E → X is the double covering in (A.5). Now we move on the side of the embedding X ⊂ P(V ∗). Denote by Γi the curves of the family {Γb}b∈E in (A.6) corresponding to the points ei, with i = 0, . . . , 3, (we are making here the obvious identification of E and ϕ(∆E)) and let {Πi = Span(Γi)}0≤i≤3 be the corresponding configuration of planes. By construction, each curve Γi is identified with the projection of ϕ(∆E) from the point ei and comes along with a distinguished divisor Di in the linear system OΓi(H − KX ) = OΓi(H + Γi): D0 = e0 + e1 + e2 + e3 Di = e0 + te0 +Xj6=i ej for i = 1, 2, 3. In particular, e0 = Γ0 · Γ0 and e1 + e2 + e3 = D0 − e0 ∼ H · Γ0. This means that the plane cubic section Γ0 ⊂ Π0 comes along with the line l0 = Span{e1, e2, e3} spanned by the three colinear points ei (i = 1, 2, 3). With the four distinct planes {Πi}0≤i≤3 the construction of Segre associates V (e0, l0) := the union of lines in P(V ∗) intersecting the four planes {Πi}0≤i≤3. (A.17) This is a cubic hypersurface containing X and its singular locus Sing(V (e0, l0)), according to Theorem 8.1, 3), intersects X along the 0-cycle ZM in (A.16). In particular, V (e0, l0) is singular at three colinear points e1, e2, e3 and hence, it must be singular along the line l0. This together with Remark 8.3 implies that the 1-dimensional part of Sing(V (e0, l0)) is the line l0. In addition, the formulas ZM = Sing(V (e0, l0)) · X and (A.16) tell us that V (e0, l0) is singular at four distinct points, τ (e0, e0), τ (e1, e2), τ (e1, e3) and τ (e2, e3), lying outside of l0. Since this is the maximal possible number of isolated singularities for a cubic hypersurface in P4 with precisely one line as the 1-dimensional part of its singular locus, we deduce Sing(V (e0, l0)) = l0 [ {τ (e0, e0), τ (e1, e2), τ (e1, e3), τ (e2, e3)}. 94 A. The projective bundle P(NX (−3H)) and the embedding X ⊂ P4 By the above construction, a Zariski dense open subset of the dual variety Tϕ(∆E ) corre- sponds to the open part of 'degenerate' cubics in IX(3), i.e., cubics having precisely a line as the 1-dimensional part of their singular locus. Since Tϕ(∆E ) is irreducible, we deduce the following. Proposition A.10. Under the isomorphism P(W ) ∼= IX(3) in (A.15), the dual variety Tϕ(∆E ) corresponds to the hypersurface C1 in IX(3) parametrizing the cubic hypersurfaces containing X and having precisely a line as the 1-dimensional part of their singular locus. Furthermore, the singular lines of the cubics in C1 are precisely the trisecant lines of X ⊂ P(V ∗). In particular, Tϕ(∆E ) and C1 are both isomorphic to the variety of trisecants of X. Remark. The desingularization of the dual variety Tϕ(∆E ) is the projectivization P(S) of S := N ∗ ϕ(∆E ) ⊗ OP(W ∗)(1), the twisted conormal bundle of ϕ(∆E) ⊂ P(W ∗), while the desingularization of the variety of trisecants of X ⊂ P(V ∗) is P(F), where F is the bundle defined in Proposition A.4. So the last assertion of Proposition A.10 implies an isomorphism between P(S) and P(F), hence an identification S ∼= F ⊗ L, for some line bundle L on E. Comparing the degrees on both sides, we obtain deg(L) = −5. Thus L = Oϕ(∆E )(−1) ⊗ L′, where L′ ∈ Pic0(E) and Oϕ(∆E )(1) = OP(W ∗)(1) ⊗ Oϕ(∆E ). Thus, the above identification takes the form N ∗ ϕ(∆E ) ⊗ OP(W ∗)(2) ∼= F ⊗ L′. This isomorphism can be viewed as the vector bundle version of the quadro-cubic Cremona transformation, see [4] for another proof of this isomorphism. A.3. The decomposition (A.14) and the configuration (104, 156) of Segre It is well-known that the ten ordinary double points of a Segre cubic hypersurface in P4 have remarkable combinatorial properties: there are fifteen planes, each spanned precisely by four singular points, and through every singular point pass precisely six of the fifteen planes. This is called a (104, 156) configuration. In this subsection we revisit this configuration in the light of the correspondence between the embedding X ⊂ P(V ∗) and the geometry of the twisted conormal bundle N ∗ X(3H) encapsulated in the properties of the morphism ϕ : Y = P(NX(−3H)) → P(W ∗), see Proposition A.7. We keep the notation of the previous subsection unless stated otherwise. On the side of the vector bundle N ∗ X(3H) we take a global section s whose zero locus Zs = (s = 0) consists of ten distinct points on X. According to Theorem 8.1, the section s corresponds to a Segre cubic denoted V3(s) ∈ P(IX(3)), and under this correspondence Sing(V3(s)) = Zs, where Sing(V3(s)) stands for the singular locus of V3(s). So from the Segre's works we know that Zs is a (104, 156) configuration. We wish to be more specific and identify the fifteen planes as well as the 4-point subcycles of Zs spanning these planes in the light of geometry of the morphism ϕ and of the embedding X ⊂ P(V ∗). We know from the previous section that ϕ gives rise to a distinguished decomposition of Zs into subcycles of degree 4. Namely, the isomorphism H 0(N ∗(3H)) ∼= H 0(OY (1)) = W 95 A. The projective bundle P(NX (−3H)) and the embedding X ⊂ P4 identifies s with the hyperplane Ms in P(W ∗). Denote by Ds the divisor cut out by Ms on the elliptic normal curve ϕ(∆E), i.e., Then (A.14) provides the decomposition Ds := Ms\ ϕ(∆E). Zs = [e∈Ds Z e s , (A.18) where, to shorten the notation, we write Z e s is a subcycle of degree 4 in Zs lying on the plane elliptic curve Γe ⊂ P(V ∗) and no other point s and no of Zs is contained in Γe. other point of Zs. Furthermore, since no three points of Zs can be colinear26, the plane Πe is spanned by Z e s . This gives us a collection of five planes {Πe}e∈Ds with the property that for every e 6= e′ the intersection In particular, the plane Πe spanned by Γe contains Z e Ms. We know that each Z e s instead of Z e Πe\ Πe′ = Γe · Γe′ is a single point, see the proof of Proposition A.5. Furthermore, the 0-cycle Zs is seen in the embedding X ⊂ P(V ∗) as the cycle of points in P(V ∗) formed by the above pairwise intersections In the sequel we set Πe\ Πe′. Zs = Xe6=e′∈Ds e · e′ := Πe\ Πe′ for e 6= e′ ∈ Ds. (A.19) (A.20) With this notation the formula (A.19) reads Zs = Xe6=e′∈Ds e · e′, while, for every e ∈ Ds, the subcycle Z e s in the decomposition (A.18) take the form Z e s = Πe\ Zs = Xe′6=e∈Ds e · e′. We now give a description of the remaining ten planes of (104, 156) configuration. Lemma A.11. For every e 6= e′ ∈ Ds, let Z e·e′ s = Zs r(cid:16)Z e s[ Z e′ s(cid:17) + e · e′. Then Z e·e′ s is a 4-degree subcycle of Zs that spans in P(V ∗) a plane denoted by Πe·e′. Proof. The fact that Z e·e′ consists of four points is obvious. To see that these points span a plane in P(V ∗) we use the identification of Zs with the singular locus Sing(V3(s)) of the Segre cubic V3(s), see Theorem 8.1. We fix the point e · e′ = ΠeT Πe′ and consider the projection from this point onto a complementary P3 which intersects V3(s) transversely s 26Otherwise the line through the three colinear points of Zs is a singular line of the cubic V3(s) which is impossible. 96 A. The projective bundle P(NX (−3H)) and the embedding X ⊂ P4 along a smooth cubic surface, call it F . Then the remaining nine points of Zs = Sing(V3(s)) project to the nine nodes of a (3, 3)-divisor where Q is a smooth quadric in P3, the image of the tangent cone of V3(s) at the point e · e′. The nine nodes force the divisor A to be completely reducible, i.e., A has the form A = F\ Q, where fi (resp. gi), for i = 1, 2, 3, are three disjoint rulings of Q such that the 0-cycle A = fi + 3Xi=1 s = X1≤i,j≤3 gi, 3Xi=1 fi\ gj Z ′ is the image of Zs r {e · e′} under the projection from e · e′, see [12] for an excellent account of the above construction. Under the projection from e · e′, the planes Πe and Πe′ go to two skew rulings, say f1 and f2. Hence the points Z e respectively. Therefore, the remaining points Zs r (Z e s and Z e′ s ) of Zs r {e · e′} must go to s ) together with e · e′ lie in the plane spanned by the line f3 and the point e · e′. Using again the fact that no three points of Zs are colinear, we deduce the assertion of the claim. (cid:3) P3 j=1 f3T gj. Hence the three points Zs r (Z e sS Z e′ s are mapped to the points P3 sS Z e′ j=1 f1T gj and P3 j=1 f2T gj The five planes {Πe}e∈Ds together with the planes Πe·e′, (e 6= e′ ∈ Ds), in Claim A.11 account for fifteen planes in the (106, 154) configuration. This collection of planes will be denoted by Ps := {Πe, Πe·e′ e ∈ Ds, e 6= e′ ∈ Ds}. (A.21) The following is obvious from the construction. Lemma A.12. Every point e · e′ ∈ Zs lies precisely on the following six planes of the collection Ps, Pe·e′ s = {Πe, Πe′, Πe·e′, Πe′′·e′′′ e′′ 6= e′′′ ∈ Ds r {e, e′}}. Furthermore, each subset of the partition Pe·e′ s = 1Pe·e′ , where s S 2Pe·e′ s 1Pe·e′ s = {Πe, Πe′, Πe·e′} and 2Pe·e′ s = {Πe′′·e′′′ e′′ 6= e′′′ ∈ Ds r {e, e′}}, consists of three planes of Pe·e′ different subsets intersect along a line. More precisely, if Ds = {e, e′, e′′, e′′′, c}, then s which intersect precisely at e · e′, while the planes taken from Πe ∩ Πe′′·e′′′ = he · e′, e · ci, Πe′\ Πe′′·e′′′ = he · e′, e′ · ci, Πe·e′\ Πe′′·e′′′ = he · e′, e′′ · e′′′i, where for two distinct points x, y in a projective space, hx, yi denotes the line spanned by those points. 97 A. The projective bundle P(NX (−3H)) and the embedding X ⊂ P4 Remark A.13. The above two claims give a precise recipe of how to recover the collection Ps of the fifteen planes of the (104, 156) configuration from the cycle Zs of ten points on X ⊂ V3(s). It should be pointed out that the collection Ps has an extra feature of being 'polarized' into two types of planes: Πe, e ∈ Ds, and Πe·e′, for e 6= e′ ∈ Ds. The presence of this polarization is due of course to the fact that Zs is not just the singular locus of a Segre cubic V3(s), but also lies on the scroll X inside of V3(s). To be even more precise, the above polarization emerges from the fact that Zs is the zero locus of a section of a vector bundle on X. The two types of planes in Ps play different roles with respect to the embedding X ⊂ P(V ∗). By construction, the planes {Πe}e∈Ds are distinguished by the property that the intersection XT Πe is the plane elliptic curve Γe, for every e ∈ Ds. The following lemma gives a similar characterization of the planes Πe·e′. Lemma A.14. For every pair e 6= e′ ∈ Ds, the plane Πe·e′ intersects the scroll X along the subcycle Z e·e′ s and the ruling le·e′ of X passing through the point e · e′. Proof. Let {c, c′, c′′} be the complement Ds r {e, e′}. From Claim A.11 it follows that the subcycle Z e·e′ s = c · c′ + c · c′′ + c′ · c′′ + e · e′ is contained in the intersection XT Πe·e′. Observe that the line L = hc · c′, c · c′′i is contained in the plane Πc and hence it intersects the curve Γc along three points (the degree 3 divisor) T = L · Γc = c · c′ + c · c′′ + t. The same holds for the line L′ = hc · c′, c′ · c′′i (resp. L′′ = hc · c′′, c′ · c′′i) and leads to T ′ = L′ · Γc = c · c′ + c′ · c′′ + t′ (resp. T ′′ = c · c′′ + c′ · c′′ + t′′). It follows that the plane Πe·e′ intersects X along seven points, Z e·e′ + t + t′ + t′′. Since the degree of X is 5, we deduce that the intersection of Πe·e′ and X is not proper, i.e., that the intersection has a 1-dimensional component, call it F . The scheme F can be either a plane cubic or a ruling of X. According to Proposition A.2, the first possibility means that F is one of the curves {Γb}b∈E and this is clearly impossible. Hence F is a ruling of X. To identify this ruling we observe that it must meet all curves {Γb}b∈E. In particular, it must intersect Γe. Hence F must pass through the intersection Πe·e′T Πe and this, in view of Lemma A.12, is the point e · e′. Hence F is the ruling of X passing through the point e · e′ as asserted. (cid:3) A.4. Toward a categorification of the configuration (104, 156) of Segre Conceptually, the whole approach of our paper can be termed as a representation of various geometric or cohomological entities attached to a surface in P4 in the category of complexes of coherent sheaves on that surface. In this subsection we apply this approach to the (104, 156) configuration of Segre con- sidered in the previous section. Namely, with our geometric set up of an elliptic scroll X embedded in P(V ∗), we have seen how the scheme of zeros Zs of a regular27 global section s 27A regular global section is a global section with simple isolated zeros. 98 A. The projective bundle P(NX (−3H)) and the embedding X ⊂ P4 X(3H) acquires the structure of the (104, 156) configuration of Segre. With our notation of N ∗ and results from the previous subsection , Zs = Xe6=e′∈Ds e · e′, where Ds is intrinsically definedby s; it is a set of five distinct points on the elliptic curve E, the base of X. See (A.19) and (A.20) for notation. We have noticed that Zs contains the distinguished subcycles of degree 4, Z e s = Xe′∈Dsr{e} e · e′, Z e·e′ s = Zs r(cid:16)Z e s[ Z e′ s(cid:17) + e · e′, (A.22) which have the geometric property of spanning the planes Πe and Πe·e′ in P(V ∗). These planes form the collection Ps of fifteen planes in (A.21). We suggest that there is a lifting of (Zs, Ps) to the category Comp(X) of (short) exact complexes of torsion free sheaves on X and hence, to the derived category D(X) of the coherent sheaves on X. Before we go on, let us be more precise about our suggestion. The first step of a categorification process is to assign a complex to every subcycle in (A.22). The second one is to turn (Zs, Ps) into a category and then to check that the morphisms of that category go to morphisms of complexes. The main result of this subsection is a realization of the first step. As for the second, let us just indicate here how one could think of (Zs, Ps) as a category. This can be achieved by turning Ps into a graph, call it C(Ps): -- the vertices of C(Ps) are the planes of the collection Ps, -- there is an edge between two vertices if and only if the corresponding planes intersect along a line. Of course, the graph C(Ps) is obviously a category: the objects are the vertices of C(Ps) and the morphisms between two objects, say from Π to Π′, are the paths, composed of edges of the graph, beginning at Π and ending at Π′. With this understood, we define C(Ps) to be the category of the Segre configuration (Zs, Ps) and we propose that there should be geometrically interesting functor(s) F : C(Ps) −→ Comp(X) (resp. D(X)). (A.23) In the sequel, we construct such a functor F on the level of objects. This is essentially Serre construction and is based on the following observation. Lemma A.15. Let Z be one of the subcycles of degree 4 in (A.22). Then there is an extension sequence 0 −→ OX(KX ) −→ FZ −→ JZ(H) −→ 0 (A.24) intrinsically attached to Z, where JZ is the ideal sheaf of Z in X and FZ is a locally free sheaf of rank 2 with Chern invariants c1(FZ ) = KX + H and c2(FZ ) = −1. 99 A. The projective bundle P(NX (−3H)) and the embedding X ⊂ P4 Proof. The geometric condition of Z spanning a plane in P(V ∗) is translated, via the exact sequence to the cohomological condition h1(JZ (H)) = 1. This and the Serre duality 0 −→ JZ (H) −→ OX (H) −→ OZ (H) −→ 0, H 1(JZ (H)))∗ ∼= Ext1(JZ (H), OX (KX )). imply that there is an extension as in (A.24) and such an extension is unique, up to the C×-action of scaling the morphisms in that sequence. Furthermore, since for any proper subscheme Z ′ ⊂ Z the cohomology H 1(JZ ′(H)) = 0, it follows by a lemma of Serre, [27, Lemma 5.1.2], that the sheaf FZ in (A.24) is locally free. Its invariants are immediately deduced from (A.24). (cid:3) Next we investigate the vector bundle FZ in (A.24). Lemma A.16. The vector bundle FZ in (A.24) is H-unstable. More precisely, there is an effective nonzero divisor AZ on X such that FZ fits into the short exact sequence 0 −→ OX (AZ ) −→ FZ −→ JZ ′(KX + H − AZ ) −→ 0, (A.25) where Z ′ is a 0-dimensional subscheme of X and JZ ′ is its ideal sheaf. Furthermore, OX (AZ ) is the H-maximal destabilizing subsheaf of FZ . Proof. From (A.24) it follows that h0(FZ ) ≥ h0(JZ(H)) − h1(OX (KX )) = 2 − 1 = 1. Hence FZ has a nonzero global section, call it f . This and the Chern invariant c2(FZ ) = −1 computed in Lemma A.15, imply that the subscheme of zeros of f must have a divisorial part which is the divisor AZ of the lemma. Hence FZ (−AZ ) has a nonzero global section f ′ whose zero locus is 0-dimensional. The asserted sequence (A.25) is the Koszul sequence of f ′ tensored with OX (AZ ). From H · AZ > 0 = H · (KX + H) = H · c1(FZ ) it follows that OX (AZ ) is H-destabilizing. This together with the fact that the quotient sheaf in (A.25) is torsion free insures the maximality of OX (AZ ). (cid:3) Next we show how the destabilizing sequence (A.25) distinguishes between the two types of subcycles in (A.22). For this we put the defining extension sequence (A.24) together with the destabilizing one to obtain the diagram. 0 OX(AZ ) 0 OX(KX ) FZ JZ(H) 0 (A.26) JZ ′(KX + H − AZ ) 0 100 A. The projective bundle P(NX (−3H)) and the embedding X ⊂ P4 The morphisms defined by the slanted arrows are nonzero and hence give rise to a nonzero JZ ′(H − AZ )). In particular, we obtain the effective divisor BZ ∈ JZ(H − AZ ) (resp. decomposition and claim the following. H = AZ + BZ Lemma A.17. Let π : X → E be the structure projection of X onto the elliptic curve E. Then AZ = π∗(a), where a is a divisor of degree either 1 or 2 on E. Proof. Observe that h0(OX (AZ )) ≤ 2, since otherwise 3 ≤ h0(OX (AZ )) = h0(OX (H − BZ )) implies that BZ is a line containing Z which is impossible since Z spans a plane. On the other hand, the Riemann-Roch for OX (AZ ) gives h0(OX (AZ )) = 1 2 (A2 Z − AZ · KX ) + h1(OX (AZ )). (A.27) From the vertical sequence in (A.26) we obtain −1 = c2(FZ ) = AZ · (KX + H − AZ) + deg(Z ′). This together with (A.27) implies h0(OX (AZ )) = 1 2 (AZ · H + 1 + deg(Z ′)) + h1(OX (AZ )). This identity and the above upper bound h0(OX (AZ )) ≤ 2 give the following possibilities: 1) h0(OX (AZ )) = 1 and AZ · H = 1, deg(Z ′) = 0, 2) h0(OX (AZ )) = 2 and AZ · H = 2, deg(Z ′) = 1, 3) h0(OX (AZ )) = 2 and AZ · H = 3, deg(Z ′) = 0. The third one can not hold, since it implies that BZ must consist of two rulings of X that contain Z. This forces the two rulings to be contained in the plane spanned by Z. The possibility 1) (resp. 2)) implies that AZ is a (resp. the union of two) ruling(s) of X. (cid:3) Hence the assertion of the lemma. Before we proceed, let us recall that the embedding X ⊂ P(V ∗) is defined by OX (H), where H = Γo + π∗(D), (A.28) with o and D being respectively a point and a divisor of degree 2 on E. With this in mind, we can now identify all the ingredients involved in the diagram (A.26) for each type of subcycle in (A.22). Proposition A.18. Let Z be one of the degree 4 subcycles of Zs appearing in (A.22). 1) If Z = Z e s , e ∈ Ds, then the destabilizing sequence (A.25) has the form 0 −→ OX (π∗(ae)) −→ FZ e s −→ Jze(KX + Γe) −→ 0, where ze is the point Γe · Γe and ae is a divisor of degree 2 on E determined by the linear equivalence with D as in (A.28). ae ∼ D + o − e 101 A. The projective bundle P(NX (−3H)) and the embedding X ⊂ P4 2) If Z = Z e·e′ s , e 6= e′ ∈ Ds, then the destabilizing sequence (A.25) has the form 0 −→ OX (lx) −→ FZ e·e′ s −→ OX (KX + H − lx) −→ 0, where lx is the ruling of X passing through a point x ∈ Z e·e′ unique divisor Re·e′ ∈ H − lx passing through Z e·e′ properties: s s . Furthermore, there is a and subject to one of the following • either e · e′ is a unique point of Z e·e′ s lying on the ruling le·e′ and then x = e · e′ and Re·e′ is a smooth elliptic curve of degree 4 and Z e·e′ s is its plane section, • or x 6= e · e′, then the divisor Re·e′ = le·e′ + Γc, for some c ∈ Ds \ {e, e′}, the ruling , where {c′, c′′} = Ds \ {e, e′, c}, le·e′ passes through two points e · e′, c′ · c′′ of Z e·e′ and the point x = c · c′ or c · c′′. s Proof. To prove 1), set Ae := H − Γe and observe that it has the form Ae = π∗(ae), where ae is a divisor of degree 2 on E. Since the subcycle Z = Z e s lies on the curve Γe, we have H 0(JZ e s (H − π∗(ae))) = H 0(JZ e s (Γe)) 6= 0. On the other hand, the defining sequence (A.24) for Z = Z e gives s , tensored with OX(−π∗(ae)), 0 −→ H 0(FZ e s (Γe)) −→ H 1(OX (KX − π∗(ae))). Since by Serre duality H 1(OX (KX − π∗(ae))) = H 1(OX (π∗(ae)))∗ = 0, we deduce s (−π∗(ae))) −→ H 0(JZ e H 0(FZ e s (−π∗(ae))) ∼= H 0(JZ e s (Γe)) ∼= C. s (−π∗(ae)) has, up to a nonzero scalar multiple, a unique nonzero global section. Its Koszul Hence FZ e Furthermore, the scheme of zeros of this section is obviously 0-dimensional. sequence tensored with OX (π∗(ae)) gives 0 −→ OX (π∗(ae)) −→ FZ e s −→ JZ ′(KX + Γe) −→ 0, where Z ′ is a single point, see the proof of Lemma A.17. Call this point ze. From (A.26) it follows that ze ∈ Γe. To identify it, we restrict the diagram (A.26) to the curve Γe and obtain the identity OΓe(ae + ze) = OX(KX ) ⊗ OΓe(Z e s ), (A.29) where we tacitly use the identification of Γe with E. Furthermore, we have OΓe (ae) = OΓe(H − Γe) and OΓe(Z e s ) = OΓe(H − KX), where the first equality is the definition of the divisor ae and the second comes from realizing Z e s as the complete intersection of Γe with a smooth curve in H − KX containing Zs. Substituting into (A.29), we obtain OΓe(ze) ⊗ OΓe(−Γe) = OΓe or, equivalently, ze ∼ Γe · Γe. This together with h0(OΓe(ze)) = 1 imply the equality ze = Γe · Γe. The linear equivalence asserted in 1) of the proposition follows from writing π∗(ae) + Γe ∼ H ∼ Γo + π∗(D), 102 A. The projective bundle P(NX (−3H)) and the embedding X ⊂ P4 where the last equivalence is (A.28). Hence ae ∼ D + o − e as divisors of E. We turn now to the part 2) of the proposition. We know that the subcycle Z e·e′ does not lie on any of the curves of the family {Γb}b∈E. This together with the first part of the proof and Lemma A.17 implies that the destabilizing sequence (A.25) for FZ e·e′ has the form s s 0 −→ OX (lx) −→ FZ e·e′ s −→ OX(KX + H − lx) −→ 0, where lx is the ruling passing through some point x ∈ X. The slanted arrow in the upper right corner of (A.26) tells us that there is an effective divisor, call it Re·e′, in H − lx passing through Z e·e′ (H − lx)) = h0(FZ e·e′ (−lx)) = 1, where the first equality comes from the horizontal sequence in (A.26) and the second one from the destabilizing sequence above. . The uniqueness of this divisor follows from h0(JZ e·e′ s s s It remains to analyse the properties of the divisor Re·e′ as well as the position of the point x. From the previous paragraph, we know already that Re·e′ is a unique effective divisor in H − lx passing through Z e·e′ . From Lemma A.14 we also know that any divisor in H passing through Z e·e′ s must also contain the ruling le·e′. Hence, if lx 6= le·e′, then the divisor Re·e′ has the form s Re·e′ = le·e′ + Γb, , since no three points in Z e·e′ for some b ∈ E, and it must contain Z e·e′ of Z e·e′ ZΓb ⊂ Z e·e′ obtain . But le·e′ is allowed to contain at most two points s are colinear. Therefore, Γb must contain a subscheme consisting of at least two points. Restricting now the diagram (A.26) to Γb, we s s s 0 OΓb(lx) 0 OX (KX ) ⊗ OΓb(ZΓb) FZ ⊗ OΓb OX (H) ⊗ OΓb(−ZΓb) 0 JZ ′(KX + H − AZ ) 0 where the slanted arrows now are zero morphisms. Hence This implies OΓb(lx) = OX(KX ) ⊗ OΓb(ZΓb). lxΓb ∼ ZΓb + KXΓb = ZΓb − ΓbΓb. In particular, deg(ZΓb) = 2. Hence, the remaining two point of Z e·e′ must lie on le·e′. To go further, recall that Z e·e′ = e · e′ + c · c′ + c · c′′ + c′ · c′′, where {c, c′, c′′} = Ds r {e, e′}. Assume now that the ruling le·e′ passes through c′ · c′′ (in addition to the point e · e′). Then Γb passes through c · c′ and c · c′′ and hence Γb must be Γc. 103 The curve Γc intersects the plane Πe·e′, the span of Z e·e′ line hc · c′, c · c′′i, the span of c · c′ and c · c′′. Namely, we have , along the points situated on the Γc\ Πe·e′ = Γc\hc · c′, c · c′′i = c · c′ + c · c′′ + le·e′ · Γc. Since the ruling lx must pass through one of these points, we deduce that x is either c · c′ or c · c′′ as asserted. We now assume that e · e′ is the only point of Z e·e′ lying on the ruling le·e′. Then the preceding argument tells us that lx = le·e′ and Re·e′ ∈ H − le·e′ is a unique divisor passing through Z e·e′ . Let us assume it to be reducible. Then it has the form s s Re·e′ = ly + Γb. Running the argument involving the previous diagram, we deduce that the ruling ly must contain two points of Z e·e′ . Hence ly is contained in the plane Πe·e′ and then it coincides with le·e′. This contradicts the assumption that le·e′ contains only one point of the subcycle Z e·e′ . Hence Re·e′ is irreducible and is a smooth section of π : X → E. Its degree H · Re·e′ = s s H · (H − le·e′) = 4. Since the intersection Re·e′T Πe·e′ ⊃ Z e·e′ s Re·e′\ Πe·e′ = Z e·e′ s . , we deduce the equality (cid:3) Now we define a functor F in (A.23) on the level of objects by F(Πe) = {0 −→ OX(KX ) −→ FZ e F(Πe·e′) = {0 −→ OX (KX ) −→ FZ e·e′ s s −→ JZ e s (H) −→ 0}, for every e ∈ Ds, −→ JZ e·e′ (H) −→ 0}, for every e 6= e′ ∈ Ds. The further study of F and related topics will be considered elsewhere. References [1] H. Abo, W. Decker, N. Sasakura, An elliptic conic bundle in P4 arising from a stable rank-3 vector bundle. Math. Z. 229 (1998), 725 -- 741. [2] D. Arcara, A. Bertram, Reider's Theorem and Thaddeus pairs revisited, [3] M. F. Atiyah, Vector bundles over an elliptic curve. Proc. London Math. Soc. 7 (1957), 414 -- 452. [4] A. Aure, W. Decker, K. Hulek, S. Popescu, K. Ranestad, Syzygies of abelian and bielliptic surfaces in P4. Internat. J. Math. 8 (1997), 849 -- 919. [5] A. Aure, W. Decker, K. Hulek, S. Popescu, K. Ranestad, The geometry of bielliptic surfaces in P4. Internat. J. Math. 4 (1993), 873 -- 902. [6] F. Bogomolov, Holomorphic tensors and vector bundles on projective varieties. Math. URSS Isvestija 13 (1979), 499 -- 555. [7] E. Bombieri, Canonical models of surfaces of general type. Inst. Hautes Etudes Sci. Publ. Math. 42 (1973), 171 -- 219. [8] T. Bridgeland, Stability conditions on K3-surfaces, Duke Math J. 141 (2008), 241 -- 291. [9] W. Decker, F.-O.Schreyer, Non-general type surfaces in P4: some remarks on bounds and constructions J.Symbolic Computation 29 (2000), 545 -- 582. 104 [10] A. Dimca and S. Papadima, Hypersurface complements, Milnor fibers and higher homotopy groups of arrangements. Ann. of Math. 158 (2003), 473 -- 507. [11] I. Dolgachev, Corrado Segre and nodal cubic threefolds. arXiv: 1501.06432 [12] I. Dolgachev, Classical algebraic geometry, Cambridge University Press, 2012. [13] I. Dolgachev, Polar Cremona transformations. Mich. Math. J. 48 (2000), 191 -- 202. [14] F. Ellia, C. Folegatti, On smooth surfaces in P4 containing a plane curve, Illinois Journal of Mathematics 51, (2007), 339 -- 352. [15] G. Ellingsrud, C. Peskine, Sur les surfaces lisses de P4, Invent.math. 95 (1989), 1 -- 11. [16] R. Hartshorne, Algebraic geometry, Graduate Texts in Mathematics 52, Springer-Verlag, New- York, 1977. [17] G. Horrocks, D. Mumford, A rank 2 vector bundle on P4 with 15, 000 symmetries. Topol- ogy 12 (1973), 63 -- 81. [18] K. Hulek, Projective geometry of elliptic curves. Ast´erisque 137, 1986. [19] P. Ionescu, Embedded projective varieties of small invariants. Algebraic geometry, Bucharest 1982, 142 -- 186, Lecture Notes in Math., 1056, Springer, Berlin, 1984. [20] S. Kobayashi, Differential geometry of complex vector bundles, Princeton Univ. Press, 1987. [21] L. Koelblen, Surfaces de P4 traces sur une hypersurface cubique. J. Reine Angew. Math. 433 (1992), 113 -- 141. [22] A. Lanteri, On the existence of scrolls in P4. (Italian summary) Atti Accad. Naz. Lincei Rend. Cl. Sci. Fis. Mat. Natur. (8) 69 (1980),223 -- 227 [23] R. Lazarsfeld, Lectures on linear series (with the assistance of Guillermo Fernndez del Busto). IAS/Park City Math. Ser., 3, Complex algebraic geometry (Park City, UT, 1993), 161 -- 219, Amer. Math. Soc., Providence, RI, 1997. [24] Y. Miyaoka, The maximal number of quotient singularities on surfaces with given numerical invariants. Math. Ann. 268 (1984), 159 -- 171. [25] Y. Miyaoka, On the Mumford-Ramanujam vanishing theorem on a surface. Journ´ees de G´eom´etrie Alg´ebrique d'Angers Sijthoff & Noordhoff, Alphen aan den RijnGermantown, Md., (1980), 239-247. [26] D. Naie, I. Reider, Twisted Kodaira-Spencer classes and the geometry of surfaces of general type. J. Algebraic Geom. 23 (2014), 165 -- 200. [27] Ch. Okonek, M. Schneider, H. Spindler, Vector bundles on Projective spaces, Birkhauser, 1988. [28] K. Ranestad, A geometric construction of elliptic conic bundles in P4. Math. Z. 231 (1999), 771 -- 781. [29] I. Reider, Vector bundles of rank 2 and linear systems on algebraic surfaces. Ann. of Math. 127 (1988), 309 -- 316. [30] L. Roth, On the projective classification of surfaces, Proc. London Math. Soc. 42 (1937), 142 -- 170. [31] R.L.E. Schwarzenberger, Vector bundles on the projective plane. Proc. London Math. Soc.(3) 11 (1961), 623640. [32] F. Severi, Intorno ai punti doppi improri di una superficie generale dello spazio ai quattro dimensioni, e a suoi punti tripli apparenti. Rend. Circ. Math. Palermo 15 (1901), 33 -- 51. 105 A. The projective bundle P(NX (−3H)) and the embedding X ⊂ P4 [33] D. Swinnerton, An enumeration of all varieties of degree 4. American J. of Math., 95 (1973), 403 -- 418. [34] E. Togliatti, Sulle forme cubiche dello spazio a cinque dimensioni aventi il massimo numero finito di punti doppi. Scritti oferti a l.Berzolari, 577-593, Pavia 1936. [35] A. N. Varchenko, Semicontinuity of the spectrum and an upper bound for the number of singular points of the projective hypersurface. (Russian) Dokl. Akad. Nauk SSSR 270 (1983), 1294-1297. [36] F. L. Zak The structure of Gauss mappings. Funct. Anal. Appl. 21 (1987), 32 -- 41. Daniel Naie Universit´e d'Angers 2, Bd Lavoisier France [email protected] Igor Reider Universit´e d'Angers 2, Bd Lavoisier France [email protected] 106
1512.01011	3	1512	2017-08-02T15:55:04	Transcendental Hodge algebra	[ "math.AG", "math.DG", "math.NT" ]	The transcendental Hodge lattice of a projective manifold $M$ is the smallest Hodge substructure in $p$-th cohomology which contains all holomorphic $p$-forms. We prove that the direct sum of all transcendental Hodge lattices has a natural algebraic structure, and compute this algebra explicitly for a hyperkahler manifold. As an application, we obtain a theorem about dimension of a compact torus $T$ admitting a symplectic embedding to a hyperkahler manifold $M$. If $M$ is generic in a $d$-dimensional family of deformations, then $\dim T\geq 2^{[(d+1)/2]}$.	math.AG	math	M. Verbitsky Transcendental Hodge algebra Transcendental Hodge algebra Misha Verbitsky1 Abstract The transcendental Hodge lattice of a projective manifold M is the smallest Hodge substructure in p-th cohomology which contains all holomorphic p-forms. We prove that the direct sum of all transcendental Hodge lattices has a natu- ral algebraic structure, and compute this algebra explicitly for a hyperkahler manifold. As an application, we obtain a theorem about dimension of a compact torus T admit- ting a holomorphic symplectic embedding to a hyperkahler manifold M . If M is generic in a d-dimensional family of deformations, then dim T > 2[(d+1)/2]. Contents 1 Introduction 1.1 Mumford-Tate group and Hodge group . . . . . . . . . . . . . . . 1.2 Hyperkahler manifolds: an introduction . . . . . . . . . . . . . . 1.3 Trianalytic and holomorphic symplectic subvarieties . . . . . . . 2 Hodge structures 2.1 Hodge structures: the definition . . . . . . . . . . . . . . . . . . . 2.2 Special Mumford-Tate group . . . . . . . . . . . . . . . . . . . . 2.3 Special Mumford-Tate group and the Aut(C/Q)-action . . . . . . 3 Transcendental Hodge algebra 3.1 Transcendental Hodge lattice . . . . . . . . . . . . . . . . . . . . 3.2 Transcendental Hodge algebra: the definition . . . . . . . . . . . 4 Zarhin's results about Hodge structures of K3 type 4.1 Number fields and Hodge structures of K3 type . . . . . . . . . . 4.2 Special Mumford-Tate group for Hodge structures of K3 type . . 2 2 3 3 4 4 5 6 7 7 7 8 8 9 5 Transcendental Hodge algebra for hyperkahler manifolds 5.1 Irreducible representations of SO(V ) . . . . . . . . . . . . . . . . 5.2 Transcendental Hodge algebra for hyperkahler manifolds . . . . . 10 10 11 1Misha Verbitsky is partially supported by the Russian Academic Excellence Project '5- 100'. Keywords: hyperkahler manifold, Hodge structure, transcendental Hodge lattice, bira- tional invariance 2010 Mathematics Subject Classification: 53C26, – 1 – version 3.0, 11.07.2017 M. Verbitsky Transcendental Hodge algebra 6 Variations of Hodge structures of K3 type and the special Mumford- Tate group 12 6.1 Variations of Hodge structures and the special Mumford-Tate group 12 6.2 Variations of Hodge structures of K3 type . . . . . . . . . . . . . 14 7 k-symplectic structures and symplectic embeddings 7.1 Non-degeneracy of the map x −→ xn in H ∗ 7.2 k-symplectic structures: definition and applications. tr(M ). . . . . . . . . . . . . . . . . 15 15 15 1 Introduction Transcendental Hodge lattice is the smallest substructure in a Hodge structure of a projective manifold containing the cohomology classes of all holomorphic p-forms. Transcendental Hodge lattices were (as Yu. Zarhin notices) a somewhat neglected subject in Hodge theory. This is somewhat surprising, because, un- like the Hodge structure on cohomology, the transcendental Hodge lattice is a birational invariant. The observation made in this paper was probably made before, but it is still very useful: the direct sum of all transcendental Hodge lattices is natu- rally an algebra (Subsection 3.2). Using Zarhin's classification of transcenden- tal Hodge structures of K3 type, we compute this algebra for all hyperkahler manifolds (Theorem 5.5; see Subsection 1.2 for the definition and basic proper- ties of hyperkahler manifolds). This computation gives a way to several gen- eralization of results which were previously known for general (non-algebraic) hyperkahler manifolds, proving non-existence of low-dimensional holomorphic symplectic tori in projective hyperkahler manifolds (Corollary 7.6). We expect more results to be obtained in the same direction, because the transcendental Hodge algebra has a promise to become a very powerful tool in the study of projective holomorphically symplectic varieties. 1.1 Mumford-Tate group and Hodge group Mumford defined the Mumford-Tate group in [Mu], and called in "Hodge group", and in literature these terms are sometimes considered equivalent. However, in Zarhin's papers [Z1, Z3, Z2] as in [PS], these notions are distinct. We shall call it "special Mumford-Take group". Given a Hodge structure V =L V p,q, we consider action of C∗ on V where z acts on V p,q as multiplication by zp−q. Then "Mumford-Tate group" is the smallest rational algebraic group containing image of C∗ and "special Mumford- Tate group", or "Hodge group", is the smallest rational algebraic group con- taining image of U (1) ⊂ C∗ We shall not pay much attention to this difference. – 2 – version 3.0, 11.07.2017 M. Verbitsky Transcendental Hodge algebra 1.2 Hyperkahler manifolds: an introduction Here we list some basic facts and properties of hyperkahler manifolds; for more details, proofs and history, please see [Bes] and [Bea]. Definition 1.1: A hyperkahler structure on a manifold M is a Riemannian structure g and a triple of complex structures I, J, K, satisfying quaternionic relations I ◦ J = −J ◦ I = K, such that g is Kahler for I, J, K. Remark 1.2: Let ωI , ωJ , ωK be the Kahler symplectic forms associated with I, J, K: ωI (·,·) = g(·, I·), ωJ (·,·) = g(·, J·), ωK(·,·) = g(·, K·). A hyperkahler manifold is holomorphically symplectic: ωJ + √−1 ωK is a holo- morphic symplectic form on (M, I). Converse is also true: Theorem 1.3: (Calabi-Yau; see [Y], [Bes]) A compact, Kahler, holomorphically symplectic manifold admits a unique hyperkahler metric in any Kahler class. Definition 1.4: A compact hyperkahler manifold M is called maximal holon- omy manifold, or simple, or IHS if π1(M ) = 0, H 2,0(M ) = C. Theorem 1.5: (Bogomolov Decomposition Theorem; [Bo1].) Any hyperkahler manifold admits a finite covering which is a product of a torus and several maximal holonomy (simple) hyperkahler manifolds. Remark 1.6: From now on, all holomorphic symplectic manifolds are tacitly assumed to be of Kahler type, "holomorphic symplectic" is used interchangeably with "hyperkahler", and all hyperkahler manifolds are assumed to be of maximal holonomy. 1.3 Trianalytic and holomorphic symplectic subvarieties In [V4] this result was improved: The starting point of the study of holomorphically symplectic subvarieties in hy- perkahler manifolds was the paper [V3], where it was shown that any complex subvariety of a general (non-algebraic) hyperkahler manifold is holomorphically symplectic outside of its singularities. it was shown that for a general complex structure induced by a unit quaternion L = aI +bJ +cK, L2 = −1 on a hyperkahler manifold (M, I, J, K), any complex subvariety of (M, L) is in fact trianalyric, that is, complex analytic with re- spect to I, J, K. The notion of a trianalytic subvariety, introduced in this paper, had many uses further on. In [V5], the notion of an abstract "hypercomplex manifold" was developed, following Deligne and Simpson ([De], [S]). The exam- ples of hypercomplex varieties include all trianalytic subvarieties of hyperkahler manifold. It was shown that hypercomplex varieties admit a natural hypercom- plex desingularization. Applied to trianalytic subvarieties of M , this gives a – 3 – version 3.0, 11.07.2017 M. Verbitsky Transcendental Hodge algebra holomorphically symplectic desingularization immersed to M holomorphically symplectically. In [V7], the results of [V4] were generalized to non-compact hyperkahler manifolds, with further generalizations in [SV1]. Much advance was done towards classifying trianalytic subvarieties in general deformations of hyperkahler manifolds ([V6], [KV98], [K], [SV2]). However, the subject of more general (that is, not necessarily trianalytic) holomorphic symplectic subvarieties was mostly neglected, with the paper [V8] being the only exception (as far as we know). In [V8], the Wirtinger invariantµ(Z, M ) of a holomorphically symplec- tic (possibly, immersed) subvariety Z in a hyperkahler manifold M was de- fined. Wirtinger number measures how far a holomorphically symplectic sub- variety Z ⊂ M is from being trianalytic. This invariant satisfies the inequality µ(Z, M ) > 1, with equality reached if and only if Z is trianalytic. It was shown that µ(Z, M ) is monotonous and multiplicative, in the following sense. Given a chain of symplectic immersions Z1 −→ Z2 −→ M , one has µ(Z1, M ) = µ(Z1, Z2)µ(Z2, M ). Therefore, Z1 is trianalytic in M if and only if it is trianalytic in Z2 and Z2 is trianalytic in M . The formalism of transcendental Hodge algebra seems to be particularly suited for the study of the holomorphically symplectic subvarieties. In this pa- per, we generalize the results of [SV2] from trianalytic to more general symplec- tic subvarieties. In [SV2] it was shown that a trianalytic complex subtorus Z in a very general deformation of a hyperkahler manifold M satisfies dim Z 6 2⌊ d+1 2 ⌋, where d = b2 − 2 is the dimension of the universal family of deformations of M . In this paper, the same result is proven for projective M generic in a d-dimensional family of deformations (Corollary 7.6). 2 Hodge structures In this section we briefly introduce the Hodge structures; for more examples, context and applications, see [Voi] and [G]. 2.1 Hodge structures: the definition Definition 2.1: Let VR be a real vector space. A (real) Hodge struc- ture of weight w on a vector space VC = VR ⊗R C is a decomposition VC = Lp+q=w V p,q, satisfying V p,q = V q,p. It is called rational Hodge structure if one fixes a rational lattice VQ such that VR = VQ ⊗ R. A Hodge structure is equipped with U (1)-action, with u ∈ U (1) acting as up−q on V p,q. Morphism of Hodge structures is a rational map which is U (1)-invariant. Notice that this U (1)-action is complex conjugate to itself. This means that it is well defined on VR. – 4 – version 3.0, 11.07.2017 M. Verbitsky Transcendental Hodge algebra Throughout this paper we often use "Hodge structures" as a shorthand for "rational Hodge structures". Definition 2.2: Polarization on a rational Hodge structrure of weight w is non-degenerate 2-form h ∈ V ∗ Q (symmetric or antisymmetric depending on parity of w) which satisfies Q ⊗ V ∗ − (√−1 )p−qh(x, x) > 0 (2.1) ("Riemann-Hodge relations") for each non-zero x ∈ V p,q, and h(x, y) = 0 for any x ∈ V p,q, y ∈ V p′,q′ , unless p 6= q′, q 6= p′. The later condition is equivalent to U (1)-invariance of h on VR. Remark 2.3: Further on, the Hodge structures we consider are tacitly assumed to be rational and polarized. Example 2.4: Let (M, ω) be a compact Kahler manifold. Then the Hodge decomposition H ∗(M, C) = ⊕H p,q(M ) defines a Hodge structure on H ∗(M, C). If we restrict ourselves to the primitive cohomology space H ∗ prim := {η ∈ H ∗(M ) (∗η) ∧ ω = 0}, and consider h(x, y) :=RM x ∧ y ∧ ωdimC M−w, relations (2.1) become the usual Hodge-Riemann relations. If, in addition, the cohomology class of ω is rational (in this case, by Kodaira theorem, (M, ω) is projective) the space H ∗ prim(M ) is also rational, and the Hodge decomposition H ∗ prim(M ) defines a polarized, rational Hodge structure. prim(M, C) = ⊕H p,q 2.2 Special Mumford-Tate group Results of Subsection 2.2 and Subsection 2.3 are rather standard; please see [Mo] and third part of [Z4]. We repeat these results here for further use. Definition 2.5: A simple object of an abelian category is an object which has no proper subobjects. An abelian category is semisimple if any object is a direct sum of simple objects. Claim 2.6: Category of polarized Hodge structures in semisimple Proof: Orthogonal complement of a Hodge substructure V ′ ⊂ V with re- spect to h is again a Hodge substructure, and this complement does not intersect V ′; both assertions follow from the Riemann-Hodge relations. Definition 2.7: Let V be a Hodge structure over Q, and ρ the correspond- ing U (1)-action. Special Mumford-Tate group (Mumford, [Mu]; Mumford called it "the Hodge group") is the smallest algebraic group over Q containing ρ. – 5 – version 3.0, 11.07.2017 M. Verbitsky Transcendental Hodge algebra Theorem 2.8: Let V be a rational, polarized Hodge structure, and MT(V ) its special Mumford-Tate group. Consider the tensor algebra of V , W = T ⊗(V ) with the Hodge structure (also polarized) induced from V . Let Wh be the space of all ρ-invariant rational vectors in W (such vectors are called "Hodge vectors"). Then MT(V ) coincides with the stabilizer StGL(V )(Wh) := {g ∈ GL(V ) ∀w ∈ Wh, g(w) = w}. Proof: Follows from the Chevalley's theorem on tensor invariants; however, to apply it one needs first to show that the Mumford-Tate group is reductive. Corollary 2.9: Let VQ be a rational Hodge structure, and W ⊂ VC a subspace. Then the following are equivalent: (i) W is a Hodge substructure. (ii) W is special Mumford-Tate invariant. 2.3 Special Mumford-Tate group and the Aut(C/Q)-action Definition 2.10: Let VQ be a Q-vector space, and VC := VQ ⊗Q C its complexi- fication, equipped with a natural Aut(C/Q) action. We call a complex subspace T ⊂ VC rational if T = TQ ⊗Q C, where TQ = VQ ∩ T . Remark 2.11: A complex subspace W ⊂ VC is rational if and only if it is Aut(C/Q)-invariant. This implies Claim 2.12: Consider a subspace W ⊂ VC, and let fWQ ⊂ VQ be the smallest subspace of VQ such that fWQ ⊗Q C ⊃ W . Then WQ ⊗Q C is generated by σ(W ), for all σ ∈ Aut(C/Q). Proposition 2.13: Let VQ be a rational, polarized Hodge structure, ρ the corresponding U (1)-action, and MT its special Mumford-Tate group. Then MT is a Zariski closure of a group generated by σ(ρ), for all σ ∈ Aut(C/Q). Proof: Since MT is rational, it is preserved by Aut(C/Q). The algebraic closure of the group generated by σ(ρ) coincides with MT, because it is the smallest group which is rational, Zariski closed and contains ρ. – 6 – version 3.0, 11.07.2017 M. Verbitsky Transcendental Hodge algebra 3 Transcendental Hodge algebra 3.1 Transcendental Hodge lattice VC = L p+q=w Definition 3.1: Let VQ be a rational, polarized Hodge structure of weight w, V p,q, V tr ⊂ VC a minimal Hodge substructure containing V w,0. We call V tr the transcendental Hodge sublattice of VC, or transcendental Hodge lattice.1 p,q>0 Theorem 3.2: Transcendental Hodge lattice is a birational invariant. Proof: Let ϕ : X −→ Y be a birational morphism of projective varieties. Then ϕ∗ : H d(Y ) −→ H d(X) induces isomorphism on H d,0. Therefore, it is injective on H d tr(Y ) not intersecting H d,0, which is impossible. Applying the same argument to the dual map, we obtain that ϕ∗ is also surjective on H d tr(Y ). Indeed, its kernel is a Hodge substructure of H d tr(Y ). 3.2 Transcendental Hodge algebra: the definition Proposition 3.3: Let M be a projective Kahler manifold, tr(M ) := ⊕dH d H ∗ tr(M ) the direct sum of all transcendental Hodge lattices, and H ∗ complement with respect to the polarization form tr(M )⊥ its orthogonal h(x, y) :=ZM x ∧ y ∧ ωdimC M−ex−ey, where ez = i for any z ∈ Λi(M ). Then H ∗ algebra. tr(M )⊥ is an ideal in the cohomology Proof: The space V ∗ := H ∗ tr(M )⊥ is a maximal Hodge structure contained in A∗ := Mp+q=w p,q>0 V p,q. The space A∗ is clearly an ideal. For any two Hodge substructures X, Y ⊂ H ∗(M ), the product X · Y also rational and U (1)-invariant, hence it is also a Hodge substructure. However, A∗ is an ideal, hence X · V ∗ is a Hodge structure contained in A∗. Therefore, H ∗(M ) · V ∗ is contained in V ∗. 1Let VQ be a rational, polarized Hodge structure of weight w, and V ′ ⊂ VC a minimal rational subspace containing V w,0. In the earlier versions of this paper I claimed that V ′ is a Hodge substructure, but this is false. The space V ′ is generated by vectors vσ ∈ σ(V w,0) which are weight w eigenvectors of operators σ(ρ), where ρ denotes the U (1)-action, and σ ∈ Aut(C/Q). However, the set of such operators is not closed under multiplication, hence V ′ is not preserved by the Mumford-Tate group which is generated by those operators. I am grateful to Pierre Deligne who noticed this error after this paper was already published. – 7 – version 3.0, 11.07.2017 M. Verbitsky Transcendental Hodge algebra Definition 3.4: The quotient algebra H ∗(M )/H ∗ the transcendental Hodge algebra of M . tr(M )⊥ = H ∗ tr(M ) is called Proposition 3.5: Transcendental Hodge algebra is a birational invariant. Proof: Same as for transcendental Hodge lattices. 4 Zarhin's results about Hodge structures of K3 type 4.1 Number fields and Hodge structures of K3 type In this subsection we give a survey of Zarhin's results on simple Hodge structures of K3 type given in a beautiful and very enjoyable paper [Z1]. We use these results in Subsection 5.2. Definition 4.1: A polarized, rational Hodge structure VC = L p+q=2 weight 2 with dim V 2,0 = 1 is called a Hodge structure of K3 type. p,q>0 V p,q of Remark 4.2: Let M be a projective K3 surface, and VQ its transcendental Hodge lattice. Then VQ is simple and of K3 type. Proposition 4.3: (Zarhin) Let VQ be a simple Hodge structure of K3 sype, and E = End(VQ) an algebra of its endomorphisms in the category of Hodge structures. Then E is a number field. Proof: By Schur's lemma, E has no zero divisors, hence it is a division algebra. Since E ⊂ End(VQ), it is countable. To prove that it is a number field, it remains to show that E is commutative. However, E acts on a 1-dimensional space V 2,0. This defines a homomorphism from E to C, which is injective, because E is a division algebra. Theorem 4.4: (Zarhin) Let VQ be a simple Hodge structure of K3 type, and E := End(VQ) its endomorphism field. Then E is either totally real (that is, all its embeddings to C are real) or is an imaginary quadratic extension of a totally real field E0. For the convenience of the reader, we sketch Zarhin's proof here. Proof. Step 1: Let a ∈ E be an endomorphism, and a∗ its conjugate with respect to the polarization h. Since the polarization is rational and U (1)- invariant, the map a∗ also preserves the Hodge decomposition. Then a∗ ∈ E. Denote the generator of V 2,0 by Ω. Then h(a(Ω), a(Ω)) = h(a∗a(Ω), Ω) > 0, – 8 – version 3.0, 11.07.2017 M. Verbitsky Transcendental Hodge algebra hence a∗a(Ω) induced by a −→ a(Ω) Ω is a positive real number. Therefore, the embedding E ֒→ C Ω maps a∗a to a positive real number. Step 2: Since the group Aut(C/Q) acts on V by automorphisms, it preserves E, hence [E : Q] is a Galois extension. This means that Aut(C/Q) acts transi- tively on embeddings from E to C, preserving the map a −→ a∗. Therefore, all embeddings E ֒→ C map a∗a to a positive real number. Step 3: The map a −→ a∗ is either a non-trivial involution or identity. In the second case, all embeddings E ֒→ C map a2 to a positive real number, and E is totally real. In the second case, τ (a) = a∗ is an involution, hence its fixed set Eτ =: E0 is a degree 2 subfield of E, with τ the generator of the Galois group Gal(E/E0). Step 4: Let r be the root of the corresponding quadratic equation x2−u = 0. Then τ (r) = −r, which gives −r2 = rτ (r) > 0, and −r2 is positive and real for all embeddings E ֒→ C. Therefore, [E : E0] is an imaginary quadratic extension. Theorem 4.5: (Zarhin) Let VQ be a Hodge structure of K3 type, and E := End(VQ) the corresponding number field. Denote by SOE(V ) the group of E-linear isometries of V for [E : Q] totally real, and by UE(V ) the group of E0-linear isometries of V for E an imaginary quadratic extension of a totally real field E0. Then the special Mumford-Tate group MT is SOE(V ) in the first case, and UE(V ) in the second. Proof: The group MT is obtained as the Weil restriction of scalars from E to Q for SOE(V ) and from E0 to Q for UE(V ). See [Z1] for original proof, [Z3] for an alternative proof using B. Kostant's theorem and [Z2] for generalizations. Definition 4.6: An imaginary quadratic extension of the totally real field is called CM field ("complex multiplication"), after [ST]. We shall call the two cases considered in Zarhin's theorem "the real case" (for SOE(V )) and "CM case" for UE(V ). 4.2 Special Mumford-Tate group for Hodge structures of K3 type Let VC = VQ ⊗Q C and VC = VR ⊗Q R. We would be interested in the real and complex Lie groups SOE(VC) and UE(VC), obtained as a complexification of M T (V ). As a Lie group, SOE(VC) is isomorphic to SO(Cn), where n is dimension of – 9 – version 3.0, 11.07.2017 M. Verbitsky Transcendental Hodge algebra VQ over E. To see this, consider the tensor product E ⊗Q C = C ⊕ C ⊕ ... ⊕ C } {z k times , where k is degree of E. The k factors of E ⊗Q C correspond to all embeddings of E to C. Let σ1, ..., σk be all such factors, σi : E ⊗Q C −→ C, and denote by Vσi the subset of V = VQ ⊗Q C corresponding to σi, Vσi = {v ∈ VQ ⊗Q C ∀a ∈ E, a(x) = σi(x)}. Clearly, V = Li Vσi , and SOE(V ) embeds to each of SO(Vσi ) tautologically. Since V is generated as a E-module by each of Vσi , this implies that SOE(V ) = SO(Vσi ), for each i. The same argument is used to show that in CM-case, UE(VR) is isomorphic to U (W ), for a complex vector space W obtained as follows. Since E is totally . Then VR is a sum of k copies of the imaginary, E ⊗Q R = C ⊕ C ⊕ ... ⊕ C } same complex vector space W with the Galois group acting transitively, and UE(VR) = U (W ). {z k/2 times The space W ⊗R C can be identified with W ⊕ W , and the Hermitian form h identifies W with W ∗, which are both isotropic subspaces with respect to h. This implies that UE(VC) is a group of isometries of W ⊕ W ∗ preserving this direct sum decomposition, and is identified with GL(W ). 5 Transcendental Hodge algebra for hyperkah- ler manifolds 5.1 Irreducible representations of SO(V ) Before we start describing the transcendental Hodge algebra for a hyperkahler manifold, we have to define a certain SO(V )-invariant quotient algebra of the symmetric algebra Sym∗(V ) of a vector space equipped with a non-degenerate scalar product. Claim 5.1: Let V be a vector space over a field k, chark = 0 equipped with a non-degenerate symmetric product h, and b ∈ Sym2(V ) an SO(V )-invariant bivector dual to h. Let Sym∗ +(V )1 be the quotient of Sym∗(V ) by the ideal generated by b. Then Symi +(V ) is irreducible as a representation of SO(V ). Proof: It is well known (see [W, Theorem V.5.7.F]) that Symi(V ) is decom- posed to a direct sum of irreducible representations of SO(V ) as follows: Symi(V ) = Symi 1We denote Sym∗ +(V ) by Sym∗ i(V ) ⊕ Symi +(V )k when k-dependence is necessary i−2(V ) ⊕ ... ⊕ Symi i−2⌊i/2⌋(V ), – 10 – version 3.0, 11.07.2017 M. Verbitsky Transcendental Hodge algebra where Symi Symi Symi i−2(V ) ⊕ ... ⊕ i−2⌊i/2⌋(V ) is the image of Symi−2(V ) under the map α −→ α · b. Then +(V ) = Symi i−k(V ) and the subrepresentation Symi i−k(V ) ∼= Symi−k i(V ), hence irreducible. Remark 5.2: Since the category of representations of SO(V ) is semisimple, any irreducible quotient of Symi(V ) can be unambiguously realized as a sub- representation of Symi(V ). Further on, we shall always consider Symi +(V ) as a subspace in Symi(V ). 5.2 Transcendental Hodge algebra for hyperkahler mani- folds Theorem 5.3: ([V1, V2]) Let M be a maximal holonomy hyperkahler man- ifold, dimC M = 2n, and H ∗ (2)(M, Q) subalgebra in cohomology generated by H 2(M, Q). Then H 2i (2)(M, Q) = Symi H 2(M, Q) for all i 6 n. Remark 5.4: Let Ω ∈ H 2(M, C) be the cohomology class of the holomorphic symplectic form. By Bogomolov's theorem ([Bo1]), for any hyperkahler manifold of maximal holonomy the space H ∗,0(M ) is spanned (as a vector space) by the powers Ω, Ω2, ..., Ωn. In particular, the transcendental Hodge lattice H i tr(M ) is non-zero only for the even i. The main result of this paper: Theorem 5.5: Let M be a projective maximal holonomy hyperkahler manifold, dimC M = 2n, H 2∗ tr (M ) its transcendental Hodge algebra, and E = End(V ) the number field of endomorphisms of its transcendental Hodge lattice V = H 2 tr(M ). Let Sym∗(V )E denote the E-linear symmetric power. Then nMi=0 H 2i tr (M ) = nMi=0 H 2i tr (M ) = nMi=0 nMi=0 Symi(V )E Symi +(V )E for E a CM-field, and for E totally real. Proof: Using the polarization form on VQ, we can identify Sym∗ Q(VQ) with the space of polylinear symmetric forms on VQ. Since Ω is a common eigenvector of all elements of E ⊂ EndQ(VQ), the complexification of Sym∗ E(VQ) contains Ωi, and therefore H 2i By definition, H 2i tr (M ) belongs to Sym∗ tr (M ) is the smallest Hodge substructure (that is, the special Mumford-Tate subrepresentation) of H 2i(M, C) containing Ωi. By Theorem 5.3, E(VQ). – 11 – version 3.0, 11.07.2017 M. Verbitsky Transcendental Hodge algebra one could replace H 2i(M, C) with Symi(VC). This means that we need to find the smallest UE(V )- and SOE(V )-representations in Symi(VC) containing Ωi. In CM-case V is the standard representation of U (V ), the space Symi(V ) is irreducible. For the totally real case , Mumford-Tate group is SOE(V ). However, the bivector b of Claim 5.1 is of Hodge type (2, 2), and the representations Symi i−k(V ) defined in the proof of Claim 5.1 are orthogonal to Ωi. Therefore, the form Ωi belongs to Symi +(V )E, which is an irreducible representation of SOE(V ) (Claim 5.1) identified with the transcendendental Hodge lattice. Definition 5.6: Let S be a family of deformations of a hyperkahler manifold, s ∈ S a point, Ms its fiber, and TsS −→ H 1(T Ms) the corresponding tangent map, defined by Kodaira and Spencer. The dimension of its image in general s ∈ S is called the essential dimension of the deformation family. Essential dimension of a holomorphic family S of deformations is a smallest d such that S can be locally obtained as a pullback of a d-dimensional family. The dimension of the transcendental Hodge lattice of a general member of a family is determined by the essential dimension of the deformation family: Theorem 5.7: Let M be a projective hyperkahler manifold, which is generic in a family S. Assume that the essential dimension of this deformation family is d. Then dimE H 2 tr(M )) is the corresponding number field. tr(M ) 6 d + 2, where E = End(H 2 Proof: Using the Torelli theorem ([V9]), we may assume that S is a subset of the period space Per of the Hodge structures on H 2(M ). Then Theorem 5.7 is implied by Proposition 6.9 below. 6 Variations of Hodge structures of K3 type and the special Mumford-Tate group 6.1 Variations of Hodge structures and the special Mumford- Tate group Definition 6.1: Let M be a complex manifold, and V a real vector bundle equipped with a flat connection over M . Assume that at each point m ∈ M , the fiber Vm is equipped by a rational, polarized Hodge structure, that is, a Hodge decomposition Vm⊗C =L V p,q m , smoothly depending on M , a rational structure Vm = Vm(Q) ⊗Q R and a polarization s ∈ Vm(Q)∗ ⊗ Vm(Q)∗. Suppose that the following conditions are satisfied. First, the rational lattice Vm(Q) and the polarization are preserved by the connection. Second, the Hodge decomposition – 12 – version 3.0, 11.07.2017 M. Verbitsky Transcendental Hodge algebra m satisfies the Griffiths transversality condition: of Hodge structures, abbreviated to VHS. m , Vm(Q), s) is called a rational, polarized variation Vm ⊗ C =L V p,q ∇(V p,q) ⊂(cid:18)V p,q ⊗ Λ1(M )(cid:19)⊕(cid:18)V p+1,q−1⊗ Λ0,1(M )(cid:19)⊕(cid:18)V p−1,q+1⊗ Λ1,0(M )(cid:19). Then (V, Vm⊗C =L V p,q Example 6.2: Let X −→ S be a holomorphic family of compact Kahler man- ifolds over a base S, and V the corresponding vector bundle of cohomology of the fibers Xt, t ∈ S. Clearly, V is equipped with the tautological connection ∇, called the Gauss-Manin connection. The Hodge decomposition on H ∗(Xt) depends smoothly on t ∈ S. If, in addition, all Xt admit a polarization (that is, an ample line bundle) Lt with c1(Lt) invariant with respect to ∇. Then the sub-bundle of primitive forms is also ∇-invariant, and the corresponding Hodge structures glue together to give a polarized, rational variation of Hodge structures (see [G] or [Voi]). From Theorem 2.8 it is obvious that the special Mumford-Tate group is lower semicontinuous in smooth families of Hodge structures. Indeed, it is the biggest group which fixes all Hodge vectors in all tensor powers of V . The sets Hodge vectors are upper semicontinuous as functions of a base, because for each rational vector, the set of fibers where it is of Hodge type (p, p) is closed. This set has a special name in the theory of variations of Hodge structures. Definition 6.3: Let V be a variation of Hodge structures over S, and ϕ ∈ V ⊗n any tensor. The set of all x ∈ S such that ϕ is of Hodge type (p, p) in x is called the Hodge locus of ϕ. The following claim is easily deduced from Griffiths transversality condition. Claim 6.4: ([G], [CK] [Voi, II, 5.3.1]) Let V be a variation of Hodge structures over a base S, ϕ ∈ V ⊗n any tensor, and Sϕ the corresponding Hodge locus. Then Sϕ is a complex analytic subset of S. Corollary 6.5: Let V be a VHS over a connected base S, and R the set of all rational vectors ϕ ∈ V ⊗n such that the corresponding Hodge locus Sϕ is a proper subset of S, and S0 ⊂ S the complement S\Sϕ∈R Sϕ. Then for each t ∈ S0 the special Mumford-Tate group M T (Vt) is independent from t, and contains M T (Vs) for any other s ∈ S. Q Proof: Follows immediately from Theorem 2.8. Definition 6.6: In assumptions of Corollary 6.5, a point s ∈ S0 is called a Mumford-Tate generic point of S. – 13 – version 3.0, 11.07.2017 M. Verbitsky Transcendental Hodge algebra 6.2 Variations of Hodge structures of K3 type Remark 6.7: Consider the space Per of all K3-type Hodge structures on (VQ, q), where VQ is a rational vector space and q a rational bilinear symmetric form on VQ of signature (2, n). Let VC := VQ ⊗Q C. Then Per = {l ∈ PVC q(l, l) = 0, q(l, l) > 0}. Indeed, for each l in this set, we can define a decomposition V 2,0 = l, V 0,2 = l, V 1,1 = hV 2,0 + V 0,2i⊥. The following claim is well known. Claim 6.8: The trivial local system (VQ ⊗Q R, q) over Per with the Hodge decomposition defined in Remark 6.7 is a VHS. Proof: The only non-trivial part is Griffiths transversality: we need to show that ∇(V 2,0) ⊂ V 2,0 ⊗ Λ1(M ) ⊕ V 1,1 ⊗ Λ1,0(M ). (6.1) Choose any section ξ of the line bundle V 2,0 on Per. Differentiating q(ξ, ξ) = 0, we obtain q(∇(ξ), ξ) = 0, which implies (6.1) immediately. Proposition 6.9: Let V be a VHS of K3 type over a complex manifold S ⊂ Per, and G the special Mumford-Tate group of a generic point s ∈ S. Then G ∼= SOE(VQ) or G ∼= UE(VQ) (Subsection 4.2), and dim S 6 dimE VQ − 2. Proof: Let W ⊂ T ⊗V be the set of all rational tensors ϕ ∈ T ⊗V which are of type (p, p) in s. By Theorem 2.8, W is the space of G-invariants: W = (T ⊗V )G. Replacing S by a bigger complex analytic subvariety of Per, we may assume that S is an open subset of the intersection S1 of all Hodge loci for all w ∈ W ; indeed, this intersection is complex analytic, contains S, and has the same generic special Mumford-Tate group. The space S1 is preserved by G acting on PVC; using Theorem 2.8 again, we identify S1 with an orbit of G. For G = SOE(V ), this orbit has complex dimension dimE VQ − 2, because it contains an open subset of a projectivization of an appropriate quadric, and a quadric in PA has dimension dim A − 2. For G = UE(V ), we have UE(VC) = GL(W ) (Subsection 4.2) acting on V = W ⊕ W ∗. Clearly, S contains a point on a quadric Q = {l ∈ PV l = C · (x1, x2) hx1, x2i = 0}. The whole quadric is an orbit of GL(W ) = UE(VC), hence it contains S, and its dimension is again dimE VQ − 2. – 14 – version 3.0, 11.07.2017 M. Verbitsky Transcendental Hodge algebra 7 k-symplectic structures and symplectic em- beddings 7.1 Non-degeneracy of the map x −→ xn in H ∗ tr(M). Applications of Theorem 5.5 are based on the following observation. Theorem 7.1: Let M be a projective maximal holonomy hyperkahler manifold, dimC M = 2n and x ∈ H 2 tr(M ) a non-zero vector. Then xn 6= 0 in H n tr(M ). tr(M ) = Sym∗ Proof: When E is a CM-field and H ∗ tr(M ) = Sym∗(V ), this is trivial. When E is totally real and H ∗ +(V ), this is proven as follows. The action of SO(V ) on the projectivization PV has only two orbits: the quadric Q := {x h(x, x) = 0} and the rest. The map P (x) = xn is non-zero, hence it is non-zero on the open orbit. To check that it is non-zero on the quadric, notice that Ω belongs to Q, and Ωn is non-zero in Symn +(V ). Corollary 7.2: Let M be a projective maximal holonomy hyperkahler manifold, N ֒→ M a holomorphic symplectic embedding, dimC N = 2n, and x ∈ H 2(N ) a restriction of a non-zero class x ∈ H 2 tr(M ). Then xn 6= 0. Proof: Since the embedding N ֒→ M is holomorphically symplectic, the tr (N ) is non-zero, hence injective. On the other restriction map H 2n hand, xn 6= 0 in H 2n tr (M ) −→ H 2n tr (M ), as shown above. 7.2 k-symplectic structures: definition and applications. Definition 7.3: Let V be a 4n-dimensional vector space, W a k-dimensional vector space, and Ψ : W −→ Λ2(V ) a linear map. Assume that Ψ(ω) is a symplectic form for general ω ∈ W , and has rank 1 2 dim W for ω in a non- degenerate quadric Q ⊂ W . Then Ψ is called k-symplectic structure on V . Remark 7.4: It is not hard to see that the set of degenerate α ∈ W is always a quadric, if the rank of degenerate α is 1 2 dim W (see [SV2]), but this quadric can be degenerate. Theorem 7.5: ([SV2]) Let V be a k-symplectic space. Then V is a Clifford module over a Clifford algebra Cl(W0), with dim W0 = dim W − 1 = k, and dim V is divisible by 2⌊(k−1)/2⌋. Corollary 7.6: Let M be a projective hyperkahler manifold, generic in a de- formation family of dimension d, and N → M a compact torus which is holo- morphically symplectically immersed to M . Then dim N is divisible by 2⌊ d+1 2 ⌋. – 15 – version 3.0, 11.07.2017 M. Verbitsky Transcendental Hodge algebra tr(M ) the number field associated with M . Proof: By Corollary 7.2, H 1(T ) is k-symplectic, where k = dimE H 2 tr(M ), and E = End(H 2 Indeed, fix an embedding E ֒→ C, ad let A := H 2 tr(M ) ⊗ C. This space is equipped with a natural action of the complexification of the special Mumford-Tate group GC, and A has only two non-zero orbits with respect to GC-action (Subsection 4.2). This implies that non-zero 2-forms on H 1(T ) induced from A can pos- sibly be non-degenerate or have rank r < dim H 1(T ) for a fixed r > 0. This implies that for any two symplectic forms ω, ω′ ∈ A, the operator ω−1 ◦ ω′ : H 1(T, C) −→ H 1(T, C) has at most two different eigenvalues, and the corre- sponding eigenspaces have the same dimension: r = dim H 1(T ) . The space of degenerate 2-forms η ∈ A is a quadric, and they all have rank either 0 or r, hence A defines a k-symplectic structure on H 1(T, C). Then k > d+2, as follows from Theorem 5.7. By Theorem 7.5 it is a Clifford 2 module over ClE(d + 1), and dim H 1(T ) is divisible by 2⌊d+1/2⌋. This means, in particular, that 2-dimensional symplectic tori exist only in families of essential dimension 6 3 and 4-dimensional symplectic tori exist only in families of essential dimension 6 5. Also, for a general projective hyperkahler manifold with b2 = 23 and Pic = Z, any symplectic subtorus has dimension divisible by 210 = 1024, because the corresponding deformation space has di- mension b2(M ) − 3 = 20. As another application, we generalize [SV2, Corollary 1.17]. tr(M ) maps injectively to H 2 Theorem 7.7: Let Z ⊂ M be a symplectic subvariety in a hyperkahler mani- fold. Then H 2 tr(Z). In particular, if the pair (M, Z) can be deformed in a family S such that the essential dimension of the corre- sponding deformation of M is d, then dim H 2 tr(Z) > (d + 2)e, where e = dimQ E and E = EndQ(H 2 tr(M )) is the corresponding number field. Proof: The first statement is clear because H 2 tr(M ) is irreducible, and mapped to H 2 tr(Z) non-trivially because Z is holomorphically symplectic. To prove the second statement, notice that dimension of H 2 at least d + 2 (Theorem 5.7), hence the dimension of H 2 (d + 2)e, and this space is embedded to H 2(Z, Q). tr(M ) over E is tr(M ) over Q is at least Acknowledgements: I am grateful to Yuri Zarhin for generously sharing his insight in the transcendental Hodge lattices and to Ljudmila Kamenova for many interesting discussions. Million thanks to Pierre Deligne who found an error in the published version of this paper (see the footnote to Definition 3.1). References [Bea] Beauville, A. Varietes Kahleriennes dont la premi`ere classe de Chern est nulle. J. Diff. Geom. 18, pp. 755-782 (1983). – 16 – version 3.0, 11.07.2017 M. Verbitsky Transcendental Hodge algebra [Bes] Besse, A., Einstein Manifolds, Springer-Verlag, New York (1987). [Bo1] Bogomolov, F. A., On the decomposition of Kahler manifolds with trivial canonical class, Math. USSR-Sb. 22 (1974), 580-583. [CK] [De] [G] Eduardo Cattani and Aroldo Kaplan, Algebraicity of Hodge loci for variations of Hodge structure, Hodge Theory, Complex Geometry, and Representation Theory, Contemp. Math., vol. 608, Amer. Math. Soc., Providence, RI, 2014. Deligne, P., A letter to Carlos Simpson, referred to in [S]. Phillip A. Griffiths (ed.), Topics in Transcendental Algebraic Geometry, Princeton Univ. Press. [KV98] D. Kaledin, M. Verbitsky, Partial resolutions of Hilbert type, Dynkin diagrams, and generalized Kummer varieties. 33 pages, arXiv:math/9812078 [K] [Mo] Nikon Kurnosov, Absolutely trianalytic tori in the generalized Kummer variety, arXiv:1504.08010, 15 pages. Ben http://www.math.ru.nl/~bmoonen/Lecturenotes/CEBnotesMT.pdf Mumford-Tate Moonen, Notes on groups, [Mu] Mumford, David, Families of abelian varieties, Algebraic Groups and Discon- tinuous Subgroups (Proc. Sympos. Pure Math., Boulder, Colo., 1965), Provi- dence, R.I.: American Mathematical Society, pp. 347-351. [PS] [ST] [S] Peters, Chris A. M. and Steenbrink, Joseph H. M., Mixed Hodge structures, Vol. 52, Ergebnisse der Mathematik und ihrer Grenzgebiete. 3. Folge. A Series of Modern Surveys in Mathematics. Berlin: Springer-Verlag, 2008, xiv+470 pages. Shimura, Goro; Taniyama, Yutaka (1961), Complex multiplication of abelian varieties and its applications to number theory, Publications of the Mathemat- ical Society of Japan, 6, Tokyo: The Mathematical Society of Japan. Simpson, C. T., Nonabelian Hodge theory. Proceedings of the International Congress of Mathematicians, (Kyoto, 1990), 747–756, Math. Soc. Japan, Tokyo, 1991. [SV1] Andrey Soldatenkov, Misha Verbitsky, Subvarieties of hypercomplex manifolds with holonomy in SL(n,H), Journal of Geometry and Physics, Volume 62, Issue 11, November 2012, Pages 2234-2240. [SV2] A. Soldatenkov, M. Verbitsky, Misha, k-symplectic structures and absolutely trianalytic subvarieties in hyperkahler manifolds, J. Geom. Phys. 92 (2015), 147-156. [V1] [V2] [V3] [V4] Verbitsky, M., Cohomology of compact hyperkahler manifolds. alg-geom elec- tronic preprint 9501001, 89 pages, LaTeX. Verbitsky, M., Cohomology of compact hyperkahler manifolds and its applica- tions, GAFA vol. 6 (4) pp. 601-612 (1996). Verbitsky M., Hyperkahler and holomorphic symplectic geometry I, Journ. of Alg. Geom., 5 no. 3 (1996) pp. 401-415. Verbitsky, M., Trianalytic subvarieties of hyperkahler manifolds, GAFA vol. 5 no. 1 (1995) pp. 92-104, also published as Hyperkahler embeddings and holo- morphic symplectic geometry II in alg-geom/9403006. – 17 – version 3.0, 11.07.2017 M. Verbitsky Transcendental Hodge algebra [V5] Verbitsky, M., Hypercomplex Varieties, alg-geom/9703016, Comm. Anal. Geom. 7 (1999), no. 2, 355–396. [V6] M. Verbitsky, Trianalytic subvarieties of the Hilbert scheme of points on a K3 surface, GAFA, 8 (1998), 732–782. [V7] [V8] [V9] Verbitsky, M., Subvarieties in non-compact hyperkahler manifolds, math.AG/0312520, Math. Res. Lett. vol. 11 (2004), no. 4, pp. 413-418 Verbitsky M., Wirtinger numbers and holomorphic symplectic immersions, Se- lecta Math. (N.S.) 10 (2004), no. 4, 551–559. Verbitsky, M., A global Torelli theorem for hyperkahler manifolds, Duke Math. J. 162 (2013), 2929-2986. [Voi] Voisin, C., Hodge theory and complex algebraic geometry I,II. Cambridge Univ. Press, Cambridge, 2002. [W] Weyl, H. The classical groups, their invariants and representations, Princeton Univ. Press, New York (1939). [Y] [Z1] [Z2] [Z3] [Z4] Yau, S. T., On the Ricci curvature of a compact Kahler manifold and the complex Monge-Amp`ere equation I, Comm. on Pure and Appl. Math. 31, 339- 411 (1978). Zarhin, Yu.G., Hodge groups of K3 surfaces, Journal fur die reine und ange- wandte Mathematik, Volume 341, page 193-220, 1983. Zarhin, Yu. G. Linear semisimple Lie algebras containing an operator with small number of eigenvalues, Arch. Math. (Basel) 46 (1986), no. 6, 522-532. Zarhin, Yuri G. Linear irreducible Lie algebras and Hodge structures, Alge- braic geometry (Chicago, IL, 1989), 281-297, Lecture Notes in Math., 1479, Springer, Berlin, 1991. Yu. G. Zarkhin, Weights of simple Lie algebras in the cohomology of algebraic varieties Mathematics of the USSR-Izvestiya, 1985, 24:2, 245-281. Misha Verbitsky Universit´e Libre de Bruxelles, CP 218, Bd du Triomphe, 1050 Brussels, Belgium [email protected], also: Laboratory of Algebraic Geometry, National Research University Higher School of Economics, Department of Mathematics, 6 Usacheva street, Moscow, Russia. – 18 – version 3.0, 11.07.2017
1308.4425	2	1308	2016-10-01T04:18:22	The Chern classes of the Verlinde bundles	[ "math.AG" ]	A formula for the first Chern class of the Verlinde bundle over the moduli space of smooth genus g curves is given. A finite-dimensional argument is presented in rank 2 using geometric symmetries obtained from strange duality, relative Serre duality, and Wirtinger duality together with the projective flatness of the Hitchin connection. A derivation using conformal-block methods is presented in higher rank. An expression for the first Chern class over the compact moduli space of curves is obtained.	math.AG	math	THE FIRST CHERN CLASS OF THE VERLINDE BUNDLES A. MARIAN, D. OPREA, AND R. PANDHARIPANDE To the memory of F. Hirzebruch Abstract. A formula for the first Chern class of the Verlinde bundle over the moduli space of smooth genus g curves is given. A finite-dimensional argument is presented in rank 2 using geometric symmetries obtained from strange duality, relative Serre duality, and Wirtinger duality together with the projective flatness of the Hitchin connection. A derivation using conformal-block methods is presented in higher rank. An expression for the first Chern class over the compact moduli space of curves is obtained. Contents 1. Introduction Part I: Finite-dimensional methods 2. Jacobian geometry 3. Slope identities 4. Projective flatness and the rank two case Part II: Representation-theoretic methods 5. The slope of the Verlinde bundles via conformal blocks 6. Extensions over the boundary References 1 5 5 8 16 19 19 25 28 1. Introduction 1.1. The slopes of the Verlinde complexes. Let Mg be the moduli space of non- singular curves of genus g ≥ 2. Over Mg, we consider the relative moduli space of rank r slope-semistable bundles of degree r(g − 1), ν : Ug(r, r(g − 1)) → Mg . The moduli space comes equipped with a canonical universal theta bundle corresponding to the divisorial locus Θr = {(C, E) : h0(E) 6= 0}. 1 2 A. MARIAN, D. OPREA, AND R. PANDHARIPANDE Pushing forward the pluritheta series, we obtain a canonical Verlinde complex 1 r(cid:17) Vr,k = Rν⋆(cid:16)Θk over Mg. For k ≥ 1, Vr,k is a vector bundle. The Verlinde bundles are known to be projectively flat [Hi]. Therefore, their Chern characters satisfy the identity (1) rank Vr,k(cid:19) . ch(Vr,k) = rank Vr,k · exp(cid:18) c1(Vr,k) The rank of Vr,k is given by the well-known Verlinde formula, see [B]. We are interested here in calculating the slope µ(Vr,k) = c1(Vr,k) rank Vr,k ∈ H 2(Mg, Q). Since the Picard rank of Mg is 1, we can express the slope in the form µ(Vr,k) = sr,k λ where λ ∈ H 2(Mg, Q) is the first Chern class of the Hodge bundle. We seek to determine the rational numbers sr,k ∈ Q. By Grothendieck-Riemann-Roch for the push-forward defining the Verlinde bundle, sr,k is in fact a rational function in k. Main Formula. The Verlinde slope is (2) µ(Vr,k) = r(k2 − 1) 2(k + r) λ . The volume of the moduli space UC(r, r(g − 1)) of bundles over a fixed curve with respect to the symplectic form induced by the canonical theta divisor is known to be given in terms of the irreducible representations χ of the group SUr : volr =ZUC (r,r(g−1)) exp(Θ) = cr ·Xχ (cid:18) 1 dim χ(cid:19)2g−2 for the constant cr = (2π)−r(r−1)(g−1)(1! 2! · · · (r − 1)!)−(g−1). Taking the k → ∞ asymptotics in formula (2) and using (1), we obtain as a consequence an expression for the cohomological push-forward: ν⋆ (exp(Θ)) = volr · exp(cid:16) r 2 λ(cid:17) . 1To avoid technical difficulties, it will be convenient to use the coarse moduli schemes of semistable vector bundles throughout most of the paper. Nonetheless, working over the moduli stack yields an equivalent definition of the Verlinde complexes, see Proposition 8.4 of [BL]. THE FIRST CHERN CLASS OF THE VERLINDE BUNDLES 3 This is a higher rank generalization of an equality over the relative Jacobian observed in [vdG]. In Part I of this paper, we are concerned with a finite-dimensional geometric proof of the Main Formula. In Part II, we give a derivation via conformal blocks. We also extend the formula over the boundary of the moduli space. Let us now detail the discussion. For the finite dimensional argument, we note four basic symmetries of the geometry: (i) Relative level-rank duality for the moduli space of bundles over Mg will be shown to give the identity sr,k + sk,r = kr − 1 2 . (ii) Relative duality along the the fibers of SU g(r, O) → Mg leads to sr,k + sr,−k−2r = −2r2. (iii) The initial conditions in rank 1, and in level 0 are µ(V1,k) = k − 1 2 , µ(Vr,0) = − 1 2 . (iv) The projective flatness of the Verlinde bundle. The four features of the geometry will be shown to determine the Verlinde slopes com- pletely in the rank 2 case, proving: Theorem 1. The Verlinde bundle V2,k has slope k2 − 1 k + 2 µ(V2,k) = λ . In arbitrary rank, the symmetries entirely determine the slopes in the Main Formula (2) under one additional assumption. This assumption concerns the roots of the Verlinde polynomial vg(k) = χ(SUC(r, O), Θk) giving the SUr Verlinde numbers at level k. Specifically, with the exception of the root k = −r which should have multiplicity exactly (r − 1)(g − 1), all the other roots of vg(k) should have multiplicity less than g − 2. Numerical evidence suggests this is true. Over a fixed curve C, the moduli spaces of bundles with fixed determinant SUC(2r, OC ) C ) are isomorphic. Relatively over Mg such an isomorphism does not hold. and SUC(2r, ωr Letting Θ denote the canonical theta divisor in ν : SU g(2r, ωr) → Mg , 4 A. MARIAN, D. OPREA, AND R. PANDHARIPANDE we may investigate the slope of W2r,k = Rν⋆(Θk) . The following statement is equivalent to Main Formula (2) via Proposition 3 of Section 3.5. As will be clear in the proof, the equivalence of the two statements corresponds geometrically to the relative version of Wirtinger's duality for level 2 theta functions. Theorem 2. The Verlinde bundle W2r,k has slope µ(W2r,k) = k(2rk + 1) 2(k + 2r) λ . In Part II, we deduce the Main Formula from a representation-theoretic perspective by connecting results in the conformal-block literature. In particular, essential to the derivation are the main statements in [T]. There, an action of a suitable Atiyah algebra, an analogue of a sheaf of differential operators, is used to describe the projectively flat WZW connection. Next, results of Laszlo [L] identify conformal blocks and the bundles of theta functions aside from a normalization ambiguity. An integrality argument fixes the variation over moduli of the results of [L], yielding the main slope formula. This is explained in Section 5. Finally, in the last section, we consider the extension of the Verlinde bundle over the compact moduli space Mg via conformal blocks. The Hitchin connection is known to acquire regular singularities along the boundary [TUY]. The formulas for the first Chern classes of the bundles of conformal blocks are given in Theorem 3 of Section 6. They specialize to the genus 0 expressions of [F] in the simplified form of [Mu]. Related work. In genus 0, the conformal block bundles have been studied in recent years in connection to the nef cone of the moduli space M0,n, see [AGS], [AGSS], [F], [Fe], [GG], [Sw]. In higher genus, the conformal block bundles have been considered in [S] in order to study certain representations arising from Lefschetz pencils. The method of [S] is to use Segal's loop-group results. Unfortunately, the geometry underlying [S] is not uniquely specified. There are at least two perspectives on the study of the higher Chern classes of the Verlinde bundle. Via a version of Thaddeus wall-crossing studied relatively over Mg,1, an approach to the higher Chern class of the Verlinde bundle is pursued in [FMP]. Projective flatness then yields nontrivial relations in the tautological ring R⋆(Mg,1) of the moduli space of curves. Whether these relations always lie in the Faber-Zagier set [PP] is an open question. THE FIRST CHERN CLASS OF THE VERLINDE BUNDLES 5 A completely different point of view is taken in [MOPPZ]. The Chern character of the conformal block bundle defines a semisimple CohFT via the fusion rules. The Givental- Teleman theory provides a classification up to an action of the Givental group. A unique element of the classification is selected by the projective flatness condition and the first Chern class calculation. The outcome is a clean formula for the higher Chern classes extending the first Chern class result of Theorem 3 proven here. However, since the latter formula incorporates the projective flatness as an input, no nontrivial relations in R⋆(Mg,1) are obtained. 1.2. Acknowlegements. We thank Carel Faber for the related computations in [FMP] and Ivan Smith for correspondence concerning [S]. Our research was furthered during the Conference on Algebraic Geometry in July 2013 at the University of Amsterdam. We thank the organizers for the very pleasant environment. A.M. and D.O. were partially supported by the NSF grants DMS 1001604, DMS 1001486, DMS 1150675, as well as by Sloan Foundation Fellowships. R.P. was partially supported by the grant ERC-2012-AdG-320368-MCSK. Part I: Finite-dimensional methods 2. Jacobian geometry In this section, we record useful aspects of the geometry of relative Jacobians over the moduli space of curves. The results will be used to derive the slope identities of Section 3. Let Mg,1 be the moduli space of nonsingular 1-pointed genus g ≥ 2 curves, and let π : C → Mg,1, σ : Mg,1 → C be the universal curve and the tautological section respectively. We set ¯g = g − 1 for convenience. The following line bundle will play an important role in subsequent calculations: L → Mg,1, L = (det Rπ⋆OC(¯gσ))−1 . An elementary Grothendieck-Riemann-Roch computation applied to the morphism π yields where c1(L) = −λ +(cid:18)g 2(cid:19)Ψ, Ψ ∈ H 2(Mg,1, Q) 6 A. MARIAN, D. OPREA, AND R. PANDHARIPANDE is the cotangent class. Consider p : J → Mg,1 the relative Jacobian of degree 0 line bundles. We let be the line bundle associated to the divisor bΘ → J (3) and let {(C, p, L) with H 0(C, L(¯g p)) 6= 0}, be the corresponding divisor class. We show θ = c1(bΘ) nλ 2 . Lemma 1. p⋆(cid:0)enθ(cid:1) = nge Proof. Since the pushforward sheaf p⋆(bΘ) has rank 1 and a nowhere-vanishing section obtained from the divisor (3), we see that p⋆(cid:16)bΘ(cid:17) = OMg,1. The relative tangent bundle of is the pullback of the dual Hodge bundle E∨ → Mg,1, with Todd genus p : J → Mg,1 Todd E∨ = e− λ 2 , see [vdG]. Hence, Grothendieck-Riemann-Roch yields p⋆(eθ) = e λ 2 . The Lemma follows immediately. (cid:3) Via Grothendieck-Riemann-Roch for p⋆(cid:16)bΘk(cid:17), we obtain the following corollary of Lemma 1. Corollary 1. We have s1,k = k − 1 2 . We will later require the following result obtained as a consequence of Wirtinger duality. Let (−1)⋆θ denote the pull-back of θ by the involution −1 in the fibers of p. Lemma 2. p⋆(cid:0)en(θ+(−1)⋆θ)(cid:1) = (2n)ge2n c1(L). THE FIRST CHERN CLASS OF THE VERLINDE BUNDLES 7 Proof. We begin by recalling the classical Wirtinger duality for level 2 theta functions. For a principally polarized abelian variety (A,bΘ), we consider the map µ : A × A → A × A given by We calculate the pullback line bundle µ(a, b) = (a + b, a − b). µ⋆(bΘ ⊠bΘ) = bΘ2 ⊠ (bΘ ⊗ (−1)⋆bΘ). The unique section of bΘ ⊠ bΘ gives a natural section of the bundle (4), inducing by Kunneth decomposition an isomorphism (4) see [M]. H 0(A,bΘ2)∨ → H 0(A,bΘ ⊗ (−1)⋆bΘ), We carry out the same construction for the relative Jacobian Concretely, we let J → Mg,1. µ : J ×Mg,1 J → J ×Mg,1 J be relative version of the map above. The fiberwise identity (4) needs to be corrected by a line bundle twist from Mg,1: (5) We determine µ⋆(bΘ ⊠bΘ) = bΘ2 ⊠(cid:16)bΘ ⊗ (−1)⋆bΘ(cid:17) ⊗ T . by constructing a section for instance T = L−2 s : Mg,1 → J ×Mg,1 J , s(C, p) = (OC , OC ). Pullback of (5) by s then gives the identity L2 = L2 ⊗ L2 ⊗ T yielding the expression for T claimed above. Pushing forward (5) to Mg,1 we obtain the relative Wirtinger isomorphism (cid:16)p⋆(bΘ2)(cid:17)∨ ∼= p⋆(cid:16)bΘ ⊗ (−1)⋆bΘ(cid:17) ⊗ L−2. 8 A. MARIAN, D. OPREA, AND R. PANDHARIPANDE We calculate the Chern characters of both bundles via Grothendieck-Riemann-Roch. We find We have already seen that 2(cid:17)∨ (cid:16)p⋆(e2θ)e− λ = p⋆(eθ+(−1)⋆θ) · e− λ 2 · e−2c1(L). p⋆(e2θ) = 2geλ, hence the above identity becomes p⋆(eθ+(−1)⋆θ) = 2ge2c1(L). The formula in the Lemma follows immediately. (cid:3) 3. Slope identities 3.1. Notation. In the course of the argument, we will occasionally view the spaces of bundles over the moduli space Mg,1 of pointed genus g curves: SUg,1(r, O) = SUg(r, O) ×Mg Mg,1, Ug,1(r, r¯g) = Ug(r, r¯g) ×Mg Mg,1. To keep the notation simple, we will use ν to denote all bundle-forgetting maps from the relative moduli spaces of bundles to the space of (possibly pointed) nonsingular curves. Over the relative moduli space Ug,1(r, r¯g) there is a natural determinant line bundle Θr → Ug,1(r, r¯g), endowed with a canonical section vanishing on the locus θr = {E → C with H 0(C, E) 6= 0}. We construct analogous theta bundles for the moduli space of bundles with trivial determinant, and decorate them with the superscript "+" for clarity. Specifically, we consider the determinant line bundle and corresponding divisor Θ+ r → SUg,1(r, O), θ+ r = {(C, p, E → C) with H 0(C, E(¯gp)) 6= 0}. Pushforward yields an associated Verlinde bundle V+ r,k = Rν⋆(cid:16)(cid:0)Θ+ r(cid:1)k(cid:17) → Mg,1. This bundle is however not defined over the unpointed moduli space Mg. While the first Chern class of Vr,k is necessarily a multiple of λ, the first Chern class of V+ r,k is a combination of λ and the cotangent class Ψ ∈ H 2(Mg,1, Q). THE FIRST CHERN CLASS OF THE VERLINDE BUNDLES 9 3.2. Strange duality. Using a relative version of the level-rank duality over moduli spaces of bundles on a smooth curve, we first prove the following slope symmetry. Proposition 1. For any positive integers k and r, we have sk,r + sr,k = kr − 1 2 . Proof. Let τ : SUg,1(r, O) ×Mg,1 Ug,1(k, k¯g) −→ Ug,1(kr, kr¯g) be the tensor product map, τ (E, F ) = E ⊗ F. Over each fixed pointed curve (C, p) ∈ Mg,1 we have, as explained for instance in [B], τ ⋆Θkr ≃(cid:0)Θ+ r(cid:1)k The natural divisor ⊠ Θr k on SUC(r, O) × UC(k, k¯g). τ ⋆θkr = {(E, F ) with H 0(E ⊗ F ) 6= 0} induces the strange duality map, defined up to multiplication by scalars, H 0(cid:16)SUC(r, O), (Θ+ r )k(cid:17)∨ −→ H 0 (UC(k, k¯g), Θr k) . This map is known to be an isomorphism [Bel], [MO], [P]. Relatively over Mg,1 we write, using the fixed-curve pullback identity (6), (6) (7) (8) τ ⋆Θkr ≃(cid:0)Θ+ r(cid:1)k for a line bundle twist We will determine ⊠ Θr k ⊗ ν⋆T on SUg,1(r, O) ×Mg,1 Ug,1(k, k¯g), T → Mg,1. T = Lkr, so that c1(T ) = kr(cid:18)λ −(cid:18)g 2(cid:19)Ψ(cid:19) . To show this, we pull back (8) via the section s : Mg,1 → SUg,1(r, O) ×Mg,1 Ug,1(k, k¯g), s(C, p) = (O⊕r C , OC(¯gp)⊕k), obtaining Lkr ≃ Lkr ⊗ Lkr ⊗ T , hence the claimed expression for T . Pushing forward (8) now, we note, as a consequence of (7), the isomorphism of Verlinde vector bundles over Mg,1, (cid:16)V+ r,k(cid:17)∨ ≃ Vk,r ⊗ T . 10 A. MARIAN, D. OPREA, AND R. PANDHARIPANDE We conclude hence r,k(cid:17) = µ (Vk,r) + c1(T ), −µ(cid:16)V+ r,k(cid:17) = µ (Vk,r) + kr(cid:18)λ −(cid:18)g −µ(cid:16)V+ 2(cid:19)Ψ(cid:19) . The equation, alongside the following Lemma, allows us to conclude Proposition 1. (cid:3) Lemma 3. We have 2 kr − 1 λ − krc1(L) r,k(cid:17) + r,k(cid:17) + µ (Vr,k) = µ(cid:16)V+ = µ(cid:16)V+ r,k(cid:17) and µ (Vr,k) we use a slightly twisted version of the tensor λ − kr(cid:18)g 2(cid:19)Ψ. 3kr − 1 2 Proof. To relate µ(cid:16)V+ product map τ in the case k = 1. More precisely we have the following diagram, where the top part is a fiber square SUg,1(r, O) ×Mg,1 J t Ug(r, r¯g) . ¯q J ❘ ❘ ❘ ❘ ❘ ❘ ❘ ❘ ❘ r p ❘ ❘ ❘ ❘ ❘ ❘ ❘ )❘ J q p Mg,1 Here, as in the previous section, we write p : J → Mg,1 for the relative Jacobian of degree 0 line bundles, while r denotes multiplication by r on J . Furthermore, for a pointed curve (C, p), t(E, L) = E ⊗ L(¯gp), q(E) = det(E(−¯gp)). Finally, ¯q is the projection onto J . The pullback equation (8) now reads t⋆Θr ≃ Θ+ r ⊠bΘr ⊗ L−r, where, keeping with the previous notation, bΘ → J is the theta line bundle associated with the divisor θ := {(C, p, L → C) with H 0(C, L(¯gp)) 6= 0}. Using the pullback identity and the Cartesian diagram, we conclude (9) r⋆q⋆(Θk r ) = ¯q⋆(cid:16)(cid:0)Θ+ r(cid:1)k ⊠bΘkr ⊗ L−kr(cid:17) = p⋆V+ r,k ⊗bΘkr ⊗ L−kr on J . / / / / ) THE FIRST CHERN CLASS OF THE VERLINDE BUNDLES 11 We are however interested in calculating ch Vr,k = ch ν⋆Θk r = ch p⋆(q⋆Θk r ). We have recorded in Lemma 1 the Todd genus of the the relative tangent bundle of to be p : J → Mg,1 Todd E∨ = e− λ 2 . Grothendieck-Riemann-Roch then gives ch Vr,k = e− λ 2 p⋆(ch (q⋆Θk r )). We further write, on J , ch (q⋆Θk r ) = 1 r2g r⋆ch (q⋆Θk r ) = 1 r2g ch (r⋆q⋆Θk r ) = 1 r2g ekrθ−krc1(L) p⋆ch V+ r,k, where (9) was used. We obtain ch Vr,k = 1 r2g e− λ 2 e−krc1(L) (cid:16)p⋆ekrθ(cid:17) ch V+ r,k on Mg,1. The final p-pushforward in the identity above was calculated in Lemma 1. Substituting, we obtain Therefore, ch Vr,k = kg rg e (kr−1)λ 2 e−krc1(L) ch V+ r,k on Mg,1. which is the assertion of Lemma 3. µ (Vr,k) = µ(cid:16)V+ r,k(cid:17) + kr − 1 2 λ − krc1(L), (cid:3) 3.3. Relative Serre duality. We will presently deduce another identity satisfied by the numbers sr,k using relative Serre duality for the forgetful morphism Proposition 2. We have ν : SUg,1(r, O) → Mg,1 . sr,k + sr,−k−2r = −2r2 . Proof. By relative duality, we have V+ r,k = Rν⋆(cid:16)(cid:0)Θ+ r(cid:1)k(cid:17) ∼= Rν⋆(cid:16)(cid:0)Θ+ r(cid:1)−k ⊗ ων(cid:17)∨ [(r2 − 1)(g − 1)] . We determine the relative dualizing sheaf of the morphism ν. As explained in Theorem E of [DN], the fibers of the morphism ν : SU g,1(r, O) → Mg,1 12 A. MARIAN, D. OPREA, AND R. PANDHARIPANDE are Gorenstein, hence the relative dualizing sheaf is a line bundle. Furthermore, along r )−2r . Thus, up to a line bundle twist the fibers of ν, the canonical bundle equals (Θ+ T → Mg,1, we have (10) The twist T will be found via a Chern class calculation to be ων =(cid:0)Θ+ r(cid:1)−2r ⊗ ν⋆T . c1(T ) = −(r2 + 1)λ + 2r2(cid:18)g 2(cid:19)Ψ. Since r,k(cid:17)∨ (cid:16)V+ we obtain taking slopes that ∼= Rν⋆((Θ+ r )−k ⊗ ων)[(r2 − 1)(g − 1)] = V+ r,−k−2r ⊗ T [(r2 − 1)(g − 1)], −µ(V+ r,k) = µ(V+ r,−k−2r) +(cid:18)−(r2 + 1)λ + 2r2(cid:18)g 2(cid:19)Ψ(cid:19) . The proof is concluded using Lemma 3. To determine the twist T , we begin by restricting (10) to the smooth stable locus of the moduli space of bundles ν : SU s g,1(r, O) → Mg,1. There, the relative dualizing sheaf is the dual determinant of the relative tangent bundle. By Corollary 4.3 of [DN], adapted to the relative situation, the Picard group of the coarse moduli space and the Picard group of the moduli stack are naturally isomorphic. We therefore consider (10) over the moduli stack of stable bundles. (We do not introduce separate notation for the stack, for simplicity.) Let E → SU s g,1(r, O) ×Mg,1 C denote the universal vector bundle of rank r over the stable part of the moduli stack. We write π : SU s g,1(r, O) ×Mg,1 C → SU s g,1(r, O) for the natural projection. Clearly, Θ+ r = (det Rπ⋆(E ⊗ OC(¯gσ)))−1. The relative dualizing sheaf of the morphism ν is expressed as ων = Rπ⋆Hom(E, E)(0) = Rπ⋆Hom(E, E) − Rπ⋆O. We therefore have c1(ων) = c1(Rπ⋆Hom(E, E)) − λ. THE FIRST CHERN CLASS OF THE VERLINDE BUNDLES 13 Using ω for the relative dualizing sheaf along the fibers of π, we calculate c1(ων) + 2rc1(Θ+ ω2 ω 2 r ) = c1 (Rπ⋆EndE) − λ − 2rc1 (Rπ⋆(E ⊗ OC(¯gσ))) = π⋆(cid:20)(cid:18)1 − − 2rπ⋆(cid:20)(cid:18)1 − = −(r2 + 1)λ + 2r2(cid:18)g 12(cid:19)(cid:0)r2 + ((r − 1)c1(E)2 − 2rc2(E)(cid:1)(cid:21)(2) c1(E)2 − c2(E)(cid:19) (1 + ¯gσ − 12(cid:19)(cid:18)r + c1(E) + 2(cid:19)Ψ + π⋆(rω · c1(E) − 2r¯g σ · c1(E) − c1(E)2). + ω 2 − λ ω2 + 1 2 ¯g2 2 σ Ψ)(cid:21)(2) Since the determinant of E is trivial on the fibers of π, we may write det E = π⋆A for a line bundle A → SU s g,1(r, O) with first Chern class α = c1(A). We calculate π⋆(rω · c1(E) − 2r¯g σ · c1(E) − c1(E)2) = 2r¯gα − 2r¯gα − π⋆(α2) = 0, and conclude ν⋆c1(T ) = c1(ων) + 2rc1(Θ+ r ) = −(r2 + 1)λ + 2r2(cid:18)g 2(cid:19)Ψ. This equality holds in the Picard group of the stable locus of the moduli stack and of the coarse moduli space. Since the strictly semistables have codimension at least 2, the equality extends to the entire coarse space SUg,1(r, O). Finally, pushing forward to Mg,1, we find the expression for the twist T claimed above. (cid:3) 3.4. Initial conditions. The next calculation plays a basic role in our argument. Lemma 4. We have sr,0 = − 1 2 . Proof. Since the Verlinde number for k = 0 over the moduli space UC(r, r¯g) is zero, the slope appears to have poles if computed directly. Instead, we carry out the calculation via the fixed determinant moduli space. The trivial bundle has no higher cohomology along the fibers of ν : SU g,1( r, O) → Mg,1 by Kodaira vanishing. To apply the vanishing theorem, we use that the fibers of ν have rational singularities, and the expression of the dualizing sheaf of Proposition 2. Hence, ν⋆ (O) = OMg,1 . 14 A. MARIAN, D. OPREA, AND R. PANDHARIPANDE Therefore which then immediately implies sr,0 = − 1 µ(V+ r,0) = 0 2 by Lemma 3. (cid:3) 3.5. Pluricanonical determinant. We have already investigated moduli spaces of bundles with trivial determinant. Here, we assume that the determinant is of degree equal to rank times g and is a multiple of the canonical bundle. The conditions require the rank to be even. Thus, we are concerned with the slopes of the complexes 2r(cid:17) , W2r,k = Rν⋆(cid:16)Θk where The following slope identity is similar to that of Lemma 3: ν : SUg(2r, ωr) → Mg. Proposition 3. We have In particular, via Theorem 1, we have µ(W2r,k) = µ(V2r,k) + λ 2 . µ(W2,k) = k(2k + 1) 2(k + 2) λ . Proof. Just as in the proof of Lemma 3, we relate µ (W2r,k) and µ (V2r,k) via the tensor product map t: SUg,1(2r, ωr) ×Mg,1 J t Ug(2r, 2r¯g) . ¯q J ❙ ❙ ❙ ❙ ❙ ❙ ❙ 2r p ❙ ❙ ❙ ❙ ❙ ❙ ❙ ❙ ❙ ❙ ❙ )❙ J q p Mg,1 We keep the same notation as in Lemma 3, letting p : J → Mg,1 denote the relative Jacobian of degree 0 line bundles, and writing 2r for the multiplication by 2r on J . Furthermore, for a pointed curve (C, p), t(E, L) = E ⊗ L, q(E) = det E ⊗ ω−r C Finally, ¯q is the projection onto J . Recall that bΘ denotes the theta line bundle on the relative Jacobian associated with the divisor θ := {(C, p, L) with H 0(C, L(¯gp)) 6= 0}. / / / / ) THE FIRST CHERN CLASS OF THE VERLINDE BUNDLES 15 It is clear that (−1)⋆bΘ has the associated divisor (−1)⋆θ = {(C, p, L) with H 0(C, L ⊗ ωC(−¯gp)) 6= 0}. For a fixed pointed curve (C, p), we have the fiberwise identity on SUC(2r, ωr) × JC. Relatively over Mg,1, the same equation holds true up to a twist T → Mg,1: t⋆Θ2r = Θ2r ⊠(cid:16)bΘ ⊗ (−1)⋆bΘ(cid:17)r t⋆Θ2r ≃ Θ2r ⊠(cid:16)bΘ ⊗ (−1)⋆bΘ(cid:17)r T = L−2r. ⊗ T . We claim that Indeed, the twist can be found in the usual way, using a suitable section s : Mg,1 → SUg,1(2r, ωr) ×Mg,1 J , for instance s(C, p) = (ωr C(−¯gp)⊕r ⊕ OC(¯gp)⊕r, OC ). Pulling back by s, we obtain the identity (L ⊗ M)r = (L ⊗ M)r ⊗ (L ⊗ M)r ⊗ T where L = det (Rπ⋆(OC(¯gσ)))−1 , M = det (Rπ⋆ (ωC(−¯gσ)))−1 . In fact, by relative duality, M ∼= L, so we conclude T = L−2r. Using the pullback identity and the Cartesian diagram, we find that over J we have (11) (2r)⋆q⋆Θk Next, we calculate 2r 2r = ¯q⋆(cid:18)Θk ⊠(cid:16)bΘ ⊗ (−1)⋆bΘ(cid:17)kr = p⋆W2r,k ⊗(cid:16)bΘ ⊗ (−1)⋆bΘ(cid:17)kr ch V2r,k = ch ν⋆Θk 2r = ch p⋆(q⋆Θk 2r) ⊗ L−2kr(cid:19) ⊗ L−2kr via Grothendieck-Riemann-Roch: ch V2r,k = e− λ 2 p⋆(ch (q⋆Θk 2r)). We further evaluate, on J , ch (q⋆Θk 2r) = 1 (2r)2g (2r)⋆ch (q⋆Θk 2r) = 1 (2r)2g ch ((2r)⋆q⋆Θk r ) 16 A. MARIAN, D. OPREA, AND R. PANDHARIPANDE where (11) was used. We obtain ch V2r,k = 1 (2r)2g e− λ 1 = (2r)2g ekr(θ+(−1)⋆θ)−2krc1(L) p⋆ch W2r,k, 2 e−2krc1(L) (cid:16)p⋆ekr(θ+(−1)⋆θ)(cid:17) ch W2r,k on Mg,1. ch V2r,k =(cid:18) k 2r(cid:19)g 2 ch W2r,k, e− λ The p-pushforward in the identity above is given by Lemma 2. Substituting, we find and taking slopes it follows that µ(V2r,k) = µ(W2r,k) − λ 2 . (cid:3) 4. Projective flatness and the rank two case 4.1. Projective flatness. By the Grothendieck-Riemann-Roch theorem for singular varieties due to Baum-Fulton-MacPherson [BFM], the Chern character of Vr,k is a poly- nomial in k with entries in the cohomology classes of Mg. (Alternatively, we may transfer the calculation to a smooth moduli space of degree 1 bundles using a Hecke modification at a point as in [BS], and then invoke the usual Grothendieck-Riemann-Roch theorem.) Taking account of the projective flatness identity (1), ch(Vr,k) = rank Vr,k · exp (sr,kλ) , we therefore write chi(Vr,k) = r2¯g+i+1Xj=0 kjαi,j = (rank Vr,k) si r,k i! λi for i ≥ 0, αi,j ∈ H 2i(Mg). As the Vandermonde determinant is nonzero, for each i we can express αi,j in terms of λi. Since λg−2 6= 0, we deduce that (rank Vr,k) si r,k , 0 ≤ i ≤ g − 2, is a polynomial in k of degree r2¯g + i + 1, with coefficients that may depend on r and g. The following is now immediate: (i) For each r we can write sr,k = ar(k) br(k) as quotient of polynomials of minimal degree, with deg ar(k) − deg br(k) ≤ 1. THE FIRST CHERN CLASS OF THE VERLINDE BUNDLES 17 Setting vg,r(k) = rank Vr,k, we also have br(k)g−2 divides vg,r(k) as polynomials in Q[k]. In addition, the following properties of the function sr,k have been established in the previous sections: 2 , sr,0 = − 1 2 , (ii) s1,k = k−1 (iii) sr,k + sk,r = kr−1 (iv) sr,k + sr,−k−2r = −2r2 for all r ≥ 1 and all k. for all k, r ≥ 1, 2 Clearly, the function sr,k = r(k2 − 1) 2(k + r) of formula (2) satisfies symmetries (ii)-(iv). Therefore, the shift s′ r,k = sr,k − r(k2 − 1) 2(k + r) satisfies properties similar to (i)-(iv): (i)′ s′ (ii)′ s′ (iii)′ s′ (iv)′ s′ r,k is a rational function of k, 1,k = 0 for all k, and s′ r,k + s′ r,k + s′ k,r = 0 for r, k ≥ 1, r,−k−2r = 0 for all r ≥ 1 and all k. r,0 = 0 for all r ≥ 1, 4.2. The rank two analysis. To prove Theorem 1, we now show that s′ k. Of course, s′ 2,0 = 0 by (ii)′. Also by (ii)′, we know that s′ 2,k = 0 for all 1,2 = 0, hence by (iii)′ we find Similarly, s′ 2,1 = 0. s′ 2,2 = 0 also by (iii)′. Using (iv)′, we obtain that s′ 2,0 = s′ 2,1 = s′ 2,2 = s′ 2,−4 = s′ 2,−5 = s′ 2,−6 = 0. Finally, we make use of the projective flatness of V2,k. The Verlinde formula reads [B] vg,2(k) = kg(cid:18) k + 2 2 (cid:19)g−1 k+1Xj=1 1 sin2g−2 jπ k+2  . 18 A. MARIAN, D. OPREA, AND R. PANDHARIPANDE The polynomial vg,2(k) admits k = 0 as a root of order g and k = −2 as a root of order (g − 1). Indeed, it was shown by Zagier that k+2!2g−2 sin jπ bvg(k + 2) = k+1Xj=1 1 bvg(0) < 0, is a polynomial in k + 2 such that see Remark 1 on page 4 of [Z]. Let us write for a polynomial B which does not have 0 and −2 as roots. By property (i) above, we b2(k) = (k + 2)mknB(k) obtain Similarly m ≤ g − 1 g − 2 n ≤ g g − 2 =⇒ m ≤ 1. =⇒ n ≤ 1 unless g = 3, 4. Also, B(k)g−2 divides the Verlinde polynomialbvg(k+2) which has degree 4g − 3 − (g − 1) − g = 2g − 2. Thus (g − 2) deg B ≤ 2g − 2 =⇒ deg B ≤ 2 except possibly when g = 3, 4. In conclusion s′ 2,k = a2(k) B(k)(k + 2)mkn − k2 − 1 k + 2 = A(k) B(k)(k + 2)k for a polynomial Since we must have Since A(k) = a2(k)(k + 2)1−mk1−n − (k2 − 1)B(k). lim k→∞ s2,k k < ∞ =⇒ lim k→∞ s′ 2,k k < ∞, deg A − deg B ≤ 3. deg B ≤ 2 =⇒ deg A ≤ 5. Furthermore, we have already observed that A(−6) = A(−5) = A(−4) = A(0) = A(1) = A(2) = 0. This implies A = 0 hence s′ 2,k = 0 as claimed. THE FIRST CHERN CLASS OF THE VERLINDE BUNDLES 19 The cases g = 3 and g = 4 have to be considered separately. First, when g = 4 we obtain m ≤ 1, n ≤ 2 and B(k)2 divides the polynomial bv4(k + 2). By direct calculation via the Verlinde formula we find 2x6 + 21x4 + 168x2 − 191 . This implies B = 1, and thus bv4(x) = with 945 s′ 2,k = A(k) k2(k + 2) deg A ≤ 4. Since A = 0 for 6 different values, it follows as before that A = 0 hence s′ 2,k = 0. When g = 3, the Verlinde flatness does not give us useful information. In this case, one possible argument is via relative Thaddeus flips, for which we refer the reader to the preprint [FMP]. Along these lines, although we do not explicitly show the details here, the genus 3 slope formula was in fact checked by direct calculation. (cid:3) Part II: Representation-theoretic methods 5. The slope of the Verlinde bundles via conformal blocks We derive here the Main Formula (2) using results in the extensive literature on conformal blocks. In particular, the central statement of [T] is used in an essential way. The derivation is by direct comparison of the bundle V+ r,k of generalized theta functions with the bundle of covacua Br,k → Mg,1 defined using the representation theory of the affine Lie algebra bslr. Over pointed curves r,k give the spaces of generalized theta functions (C, p), the fibers of the dual bundle B∨ H 0(SUC(r, O),(cid:0)Θ+ r(cid:1)k). Globally, the identification B∨ The explicit identification of the twist and formula (2) will be deduced together. r,k will be shown below to hold only up to a twist. r,k ≃ V+ 20 A. MARIAN, D. OPREA, AND R. PANDHARIPANDE 5.1. The bundles of covacua. For a self-contained presentation, we start by reviewing briefly the definition of Br,k. Fix a smooth pointed curve (C, p), and write K for the field of fractions of the completed local ring O = bOC,p. For notational simplicity, we set g = slr, and write () for the suitably normalized Killing form. The loop algebra is the central extension of g ⊗ K, endowed with the bracket cLg = g ⊗ K ⊕ C · c [X ⊗ f, Y ⊗ g] = [X, Y ] ⊗ f g + (XY ) · Res (g df ) · c. Two natural subalgebras of the loop algebra cLg play a role: dL+g = g ⊗ O ⊕ C · c ֒→ cLg LC g = g ⊗ OC(C − p) ֒→ cLg. and U ( dL+g) C. V ′ k ֒→ Vk. g acts trivially. We set defined as follows. The one-dimensional vector space C is viewed as a module over the For each positive integer k, we consider the basic representation Hk of cLg at level k, universal enveloping algebra U (dL+g) where the center c acts as multiplication by k, and There is a unique maximal cLg-invariant submodule Vk = U (cLg) ⊗ The basic representation is the quotient Hk = Vk/V ′ k. The finite-dimensional space of covacua for (C, p), dual to the space of conformal blocks, is given in turn as a quotient Br,k = Hk/LC g Hk. When the pointed curve varies, the loop algebra as well as its two natural subalgebras relativize over Mg,1. The above constructions then give rise to the finite-rank vector bundle endowed with the projectively flat WZW connection. Br,k → Mg,1, THE FIRST CHERN CLASS OF THE VERLINDE BUNDLES 21 5.2. Atiyah algebras. The key theorem in [T] uses the language of Atiyah algebras to describe the WZW connection on the bundles Br,k. We review this now, and refer the reader to [Lo] for a different account. An Atiyah algebra over a smooth base S is a Lie algebra which sits in an extension 0 → OS i→ A π→ TS → 0. If L → S is a line bundle, then the sheaf of first order differential operators acting on L is an Atiyah algebra via the symbol exact sequence. AL = Diff1(L), We also need an analogue of the sheaf of differential operators acting on tensor powers Lc for all rational numbers c, even though these line bundles don't actually make sense. To this end, if A is an Atiyah algebra and c ∈ Q, then cA is by definition the Atiyah algebra cA = (OS ⊕ A)/(c, 1)OS sitting canonically in an exact sequence The sum of two Atiyah algebras A and B is given by 0 → OS → cA → TS → 0. A + B = A ×TS B/(iA(f ), −iB(f )) for f ∈ OS. When c is a positive integer, cA coincides with the sum A + . . . + A, but cA is more generally defined for all c ∈ Q. In particular, cAL makes sense for any c ∈ Q and any line bundle L → S. An action of an Atiyah algebra A on a vector bundle V is understood to enjoy the following properties (i) each section a of A acts as a first order differential operator on V with symbol given by π(a) ⊗ 1V ; (ii) the image of 1 ∈ OS i.e. i(1) acts on V via the identity. It is immediate that the action of an Atiyah algebra on V is tantamount to a projectively flat connection in V. Furthermore, if two Atiyah algebras A and B act on vector bundles V and W respectively, then the sum A + B acts on V ⊗ W via (a, b) · v ⊗ w = av ⊗ w + v ⊗ bw. We will make use of the following: 22 A. MARIAN, D. OPREA, AND R. PANDHARIPANDE Lemma 5. Let c ∈ Q be a rational number and L → S be a line bundle. If the Atiyah algebra cAL acts on a vector bundle V, then the slope µ(V) = det V/rank V is determined by µ(V) = c L . Proof. Replacing the pair (V, L) by a suitable tensor power we reduce to the case c ∈ Z via the observation preceding the Lemma. Then, we induct on c, adding one copy of the Atiyah algebra of L at a time. The base case c = 0 corresponds to a flat connection in V. Indeed, the Atiyah algebra of OS splits as OS ⊕ TS and an action of this algebra of V is equivalent to differential operators ∇X for X ∈ TS, such that hence to a flat connection. Consider the rational number [∇X , ∇Y ] = ∇[X,Y ], c = k(r2 − 1) r + k , (cid:3) which is the charge of the Virasoro algebra acting on the basic level k representation Hk of cLg. The representation Hk entered the construction of the bundles of covacua Br,k. The main result of [T] is the fact that the Atiyah algebra c 2 AL acts on the bundle of covacua Br,k where AL is the Atiyah algebra associated to the determinant of the Hodge bundle By Lemma 5, we deduce the slope L = det E. µ(Br,k) = k(r2 − 1) 2(r + k) λ. In fact, by the proof of Lemma 5, the bundle B2(r+k) r,k ⊗ L−k(r2−1) is flat. 5.3. Identifications and the slope calculation. We now explain how the above cal- culation implies the Main Formula (2) via the results of Section 5.7 of [L]. THE FIRST CHERN CLASS OF THE VERLINDE BUNDLES 23 Crucially, Laszlo proves that the projectivization of B∨ r,k coincides with the projec- In fact, Laszlo shows that for a tivization of the bundle V+ suitable line bundle Lr over r,k coming from geometry. SUg,1(r, O) → Mg,1 we have r,k = π⋆(Lk B∨ where fiberwise, over a fixed pointed curve, Lk (Θ+ r )k . Hence, r ), Lk r =(cid:0)Θ+ r(cid:1)k ⊗ Tr,k r coincides with the usual theta bundle for some line for some line bundle twist Tr,k → Mg,1 over the moduli stack. At the heart of this identification is the double quotient construction of the moduli space of bundles over a curve with the theta bundle Θ+ the affine Grassmannian SUC(r, O) = LCG\dLG /[L+G r being obtained by descent of a natural line bundle Qr from Here dLG and [L+G are the central extensions of the corresponding loop groups. The construction is then carried out relatively over Mg,1, such that Qk bundle r descends to the line Qr →dLG /[L+G. It follows from here that fiberwise Lk Lk r → SUg,1(r, O). r coincides with the usual theta bundle (Θ+ r )k. Collecting the above facts, we find that Therefore r,k = V+ B∨ r,k ⊗ Tr,k. −µ(Br,k) = µ(V+ r,k) + c1(Tr,k). Using Lemma 3 we conclude that − k(r2 − 1) 2(r + k) λ = µ(Vr,k) − kr − 1 2 λ + krc1(L) + c1(Tr,k). Simplifying, this yields µ(Vr,k) = r(k2 − 1) 2(r + k) λ − c1(Tr,k) − krc1(L). 24 A. MARIAN, D. OPREA, AND R. PANDHARIPANDE Now, the left hand side is a multiple of λ, namely sr,kλ. The right hand side must be a multiple of λ as well. With we find that s′ r,k = sr,k − r(k2 − 1) 2(r + k) . s′ r,kλ = −c1(Tr,k) − krc1(L). This implies that s′ r,k must be an integer by comparison with the right hand side, because the Picard group of Mg is generated over Z by λ for g ≥ 2, see [AC2]. The fact that s′ r,k ∈ Z is enough to prove which is what we need. s′ r,k = 0, Indeed, as explained in Section 4.1, Grothendieck-Riemann-Roch for the pushforwards giving the Verlinde numbers shows that Writing s′ r,k k lim k→∞ < ∞. s′ r,k = ar(k)/br(k) with deg ar(k) ≤ deg br(k) + 1, we see by direct calculation that lim k→∞ s′ r,k+1 − 2s′ r,k + s′ r,k−1 = 0. Since the expression in the limit is an integer, it must equal zero. By induction, it follows that s′ r,k = Ark + Br for constants Ar, Br that may depend on the rank and the genus. Since s′ r,0 = s′ r,−2r = 0 by the initial condition in Lemma 4 and by Proposition 2, it follows that Ar = Br = 0 hence s′ r,k = 0. As a consequence, we have now also determined the twist Tr,k = L−kr. Therefore, the bundle of conformal blocks is expressed geometrically as r,k = V+ B∨ r,k ⊗ L−kr. We remark furthermore that the latter bundle descends to Mg. To see this, one checks that (Θ+ r )k ⊗ L−kr THE FIRST CHERN CLASS OF THE VERLINDE BUNDLES 25 restricts trivially over the fibers of SUg,1(r, O) → SUg(r, O). This is a straightforward verification. 6. Extensions over the boundary The methods of [T] can be used to find the first Chern class of the bundle of conformal blocks over the compactification Mg. The resulting formula is stated in Theorem 3 below. In particular the first Chern class contains nonzero boundary contributions, contrary to a claim of [S]. In genus 0, formulas for the Chern classes of the bundle of conformal blocks were given in [F], and have been recently brought to simpler form in [Mu]. In higher genus, the expressions we obtain using [T] specialize to the simpler formulas of [Mu]. As it is necessary to consider parabolics, we begin with some terminology on partitions. We denote by Pr,k the set of Young diagrams with at most r rows and at most k columns. Enumerating the lengths of the rows, we write a diagram µ as µ = (µ1, . . . , µr), k ≥ µ1 ≥ · · · ≥ µr ≥ 0. The partition µ is viewed as labeling the irreducible representation of the group SU (r) with highest weight µ, which we denote by Vµ. Two partitions which differ by the augmentation of the rows by a common number of boxes yield isomorphic representations. We will identify such partitions in Pr,k, writing ∼ for the equivalence relation. There is a natural involution where µ⋆ is the diagram whose row lengths are Pr,k ∋ µ 7→ µ⋆ ∈ Pr,k k ≥ k − µr ≥ . . . ≥ k − µ1 ≥ 0. Further, to allow for an arbitrary number of markings, we consider multipartitions µ = (µ1, . . . , µn) whose members belong to Pr,k/ ∼. Finally, for a single partition µ, we write wµ = 1 2(r + k) rXi=1 µ2 i − 1 r rXi=1 µi!2 + (r − 2i + 1)µi rXi=1 for the suitably normalized action of the Casimir element on the representation Vµ. In this setup, we let Bg,µ → Mg,n 26 A. MARIAN, D. OPREA, AND R. PANDHARIPANDE be the bundle of covacua, obtained analogously to the construction of Section 5.1 using representations of highest weight µ, see [T]. To simplify notation, we do not indicate dependence on r, k and n explicitly: these can be read off from the multipartition µ. We set to be the parabolic Verlinde number. vg(µ) = rank Bg,µ We determine the first Chern class c1(Bg,µ) over Mg,n in terms of the natural gener- ators: and the boundary divisors. To fix notation, we write as usual: λ, Ψ1, . . . , Ψn • δirr for the class of the divisor corresponding to irreducible nodal curves; • δh,A for the boundary divisor corresponding to reducible nodal curves, with one component having genus h and containing the markings of the set A. Note that each subset A ⊂ {1, 2, . . . , n} determines a splitting of the multipartition µ corresponding to the markings in A and in its complement Ac. Finally, we define the coefficients µ A ∪ µ Ac and cirr = Xν∈Pr,k/∼ wν · vg−1(µ, ν, ν⋆) vg(µ) ch,A = Xν∈Pr,k/∼ wν · vh(µ A , ν) · vg−h(µ Ac, ν⋆) vg(µ) . Theorem 3. Over Mg,n the slope of the bundle of covacua is (12) slope(Bg,µ) = k(r2 − 1) 2(r + k) λ + nXi=1 wµiΨi − cirrδirr −Xh,A ch,Aδh,A. In the formula, the repetition δh,A = δg−h,Ac is not allowed, so that each divisor appears only once. Proof. The formula written above is correct over the open stratum Mg,n. Indeed, the main theorem of [T], used in the presence of parabolics, shows that the bundle of covacua Bg,µ → Mg,n THE FIRST CHERN CLASS OF THE VERLINDE BUNDLES 27 admits an action of the Atiyah algebra As before k(r2 − 1) 2(r + k) AL + nXi=1 wµiALi. L = det E is the determinant of the Hodge bundle and the Li denote the cotangent lines over Mg,n. Therefore, by Lemma 5, we have slope(Bg,µ) = k(r2 − 1) 2(r + k) λ + wµiΨi nXi=1 over Mg,n. It remains to confirm that the boundary corrections take the form stated above. Since the derivation is identical for all boundary divisors, let us only find the coefficient of δirr. To this end, observe the natural map ξ : Mg−1,n+2 → Mg,n whose image is contained in the divisor δirr. The map is obtained by gluing together the last two markings which we denote • and ⋆. We pull back (12) under ξ. For the left hand side, we use the fusion rules of [TUY]: ξ⋆Bg,µ = Mν∈Pr,k/∼ Bg−1,µ,ν,ν⋆. Thus, the left hand side becomes Xν∈Pr,k/∼ vg−1(µ, ν, ν⋆) vg(µ) · slope(Bg−1,µ,ν,ν⋆) = Xν∈Pr,k/∼ = k(r2 − 1) 2(r + k) λ + vg−1(µ, ν, ν⋆) 2(r + k) · k(r2 − 1) wµiΨi + Xν∈Pr,k/∼ vg(µ) nXi=1 wµiΨi + wν Ψ• + wν⋆Ψ⋆! · (wνΨ• + wν⋆ Ψ⋆) . λ + nXi=1 vg−1(µ, ν, ν⋆) vg(µ) The fusion rules have been used in the third line to compare the ranks of the Verlinde bundles. For the right hand side, we record the following well-known formulas [AC1]: (i) ξ⋆λ = λ; (ii) ξ⋆Ψi = Ψi for 1 ≤ i ≤ n; (iii) ξ⋆δirr = −Ψ• − Ψ⋆; (iv) ξ⋆δh,A = 0. 28 A. MARIAN, D. OPREA, AND R. PANDHARIPANDE These yield the following expression for the right hand side of (12): k(r2 − 1) 2(r + k) λ + nXi=1 wµiΨi − cirr(−Ψ• − Ψ⋆). For g − 1 ≥ 2, Ψ⋆ and Ψ• are independent in the Picard group of Mg−1,n+2, see [AC2], hence we can identify their coefficient cirr uniquely to the formula claimed above. The case of the other boundary corrections is entirely similar. (cid:3) Remark. The low genus case g ≤ 2 not covered by the above argument can be estab- lished by the following approach. Once a correct formula for the Chern class has been proposed, a proof can be obtained by induction on the genus and number of markings. Indeed, with some diligent bookkeeping, it can be seen that the expression of the The- orem restricts to the boundary divisors compatibly with the fusion rules in [TUY]. To finish the argument, we invoke the Hodge theoretic result of Arbarello-Cornalba [AC1] stating the boundary restriction map H 2(Mg,n) → H 2(Mg−1,n+2)Mh,A H 2(Mh,A∪{•} × Mg−h,Ac∪{⋆}) is injective, with the exception of the particular values (g, n) = (0, 4), (0, 5), (1, 1), (1, 2), which may be checked by hand. In fact, the slope expression of the Theorem is certainly correct in the first three cases by [F], [Mu]. When (g, n) = (1, 2), we already know from [T] that the slope takes the form slope(Bµ1,µ2) = k(r2 − 1) 2(r + k) λ + wµ1Ψ1 + wµ2Ψ2 − cirrδirr − c∆, where δirr and ∆ are the two boundary divisors in M1,2. The coefficients cirr and c are determined uniquely in the form stated in the Theorem by restricting Bµ1,µ2 to the two boundary divisors δirr and ∆ (and not only to their interiors as was done above) via the fusion rules. The verification is not difficult for the particular case (1, 2). References [AC1] E. Arbarello, M. Cornalba, Calculating cohomology groups of moduli spaces of curves via algebraic geometry, Inst. Hautes ´Etudes Sci. Publ. Math. 88 (1998), 97 - 127. [AC2] E. Arbarello, M. Cornalba, The Picard groups of the moduli spaces of curves, Topology 26 (1987), 153 - 171. [AGS] V. Alexeev, A. Gibney, D. Swinarski, Conformal blocks divisors on M 0,n from sl2, Proceedings of the Edinburgh Mathematical Society, to appear. [AGSS] M. Arap, A. Gibney, J. Stankewicz, D. Swinarski, sln level 1 conformal blocks divisors on M 0,n, Int. Math. Res. Not. (2012), 1634 -1680. [BFM] P. Baum, W. Fulton, R. MacPherson, Riemann-Roch for singular varieties, Inst. Hautes ´Etudes Sci. Publ. Math 45 (1975), 101 - 145. THE FIRST CHERN CLASS OF THE VERLINDE BUNDLES 29 [B] A. Beauville, Vector bundles on curves and generalized theta functions: recent results and open problems, Current topics in complex algebraic geometry, 17 - 33, Math. Sci. Res. Inst. Publ. 28, Cambridge University Press, Cambridge, 1995. [BL] A. Beauville, Y. Laszlo, Conformal blocks and generalized theta functions, Comm. Math. Phys. 164 (1994), 385 - 419. [BS] A. Bertram, A. Szenes, Hilbert polynomials of moduli spaces of rank 2 vector bundles II, Topology 32, (1993), 599 - 609. [Bel] P. Belkale, Strange duality and the Hitchin/WZW connection, J. Differential Geom. 82 (2009), 445 - 465. [DN] J. M. Dr´ezet, M.S. Narasimhan, Groupe de Picard des vari´et´es de modules de fibr´es semi- stables sur les courbes alg´ebriques, Invent. Math. 97 (1989), 53 - 94. [F] N. Fakhruddin, Chern classes of conformal blocks, Compact moduli spaces and vector bundles, 145 - 176, Contemp. Math., 564, Amer. Math. Soc., Providence, RI, 2012. [Fe] M. Fedorchuk, Cyclic covering morphisms on M 0,n, preprint, arXiv:1105.0655. [FMP] C. Faber, A. Marian, R. Pandharipande, Verlinde flatness and relations in H ⋆(Mg), available at http://www.math.ethz.ch/rahul/vertaut.pdf. [GG] N. Giansiracusa, A. Gibney, The cone of type A, level 1, conformal blocks divisors, Adv. Math. 231 (2012), 798 - 814. [Hi] N. Hitchin, Flat connections and geometric quantization, Comm. Math. Phys. 131 (1990), 347 - 380. [L] Y. Laszlo, Hitchin's and WZW connections are the same, J. Differential Geometry 49 (1998), 547 - 576. [Lo] E. Looijenga, From WZW models to Modular Functors, Handbook of Moduli, Advanced Lectures in Mathematics, vol. 25, Intl. Press, Somerville, MA, 2013. [M] D. Mumford, Prym varieties I, Contributions to Analysis, 325 - 350, Academic Press, New York 1974. [Mu] S. Mukhopadhyay, Rank-Level duality and Conformal Block divisors, preprint, arXiv:1308.0854. [MO] A. Marian, D. Oprea, The level-rank duality for non-abelian theta functions, Invent. Math. 168 (2007), 225 - 247. [MOPPZ] A. Marian, D. Oprea, R. Pandharipande, A. Pixton, D. Zvonkine, The Chern character of the Verlinde bundle, in preparation. [PP] R. Pandharipande, A. Pixton, Relations in the tautological ring of the moduli space of curves, preprint, arXiv:1301.4561. [P] C. Pauly, Strange duality revisited, preprint, arXiv:1204.1186. [S] I. Smith, Symplectic four-manifolds and conformal blocks, J. London Math. Soc. 71 (2005), 503 - 515. [Sw] D. Swinarski, sl2 conformal block divisors and the nef cone of M 0,n, preprint, arXiv:1107.5331. [TUY] A. Tsuchiya, K. Ueno, Y. Yamada, Conformal field theory on universal family of stable curves with gauge symmetries, Integrable systems in quantum field theory and statistical mechanics, Adv. Stud. Pure Math., vol. 19, Academic Press, Boston, MA, 1989, 459 - 566. [T] Y. Tsuchimoto, On the coordinate-free description of the conformal blocks, J . Math. Kyoto Univ. 33 (1993), 29 - 49. [vdG] G. van der Geer, Cycles on the Moduli Space of Abelian Varieties, Moduli of curves and abelian varieties, 65 - 89, Aspects Math., Vieweg, Braunschweig, 1999. [Z] D. Zagier, Elementary aspects of the Verlinde formula and of the Harder-Narasimhan-Atiyah-Bott formula, Proceedings of the Hirzebruch 65 Conference on Algebraic Geometry 445 - 462, Israel Math. Conf. Proc., 9, Bar-Ilan Univ., Ramat Gan, 1996. Department of Mathematics, Northeastern University E-mail address: [email protected] Department of Mathematics, University of California, San Diego E-mail address: [email protected] Department of Mathematics, ETH Zurich E-mail address: [email protected]
1202.6662	2	1202	2013-05-31T12:01:29	Okounkov bodies and Seshadri constants	[ "math.AG" ]	Okounkov bodies, which are closed convex sets defined for big line bundles, have rich information on the line bundles. On the other hand, Seshadri constants are invariants which measure the positivity of line bundles. In this paper, we prove that Okounkov bodies give lower bounds of Seshadri constants.	math.AG	math	OKOUNKOV BODIES AND SESHADRI CONSTANTS ATSUSHI ITO Abstract. Okounkov bodies, which are closed convex sets defined for big line bundles, have rich information on the line bundles. On the other hand, Seshadri constants are invariants which measure the positivity of line bundles. In this paper, we prove that Okounkov bodies give lower bounds of Seshadri constants. 1. Introduction In this paper, we investigate a relation between Okounkov bodies and Seshadri constants. Okounkov bodies were introduced by Lazarsfeld and Mustat¸a [LM] and independently by Kaveh and Khovanskii [KK], based on the work of Ok- ounkov [Ok1, Ok2]. First, recall the definition of Okounkov bodies. Let X be a variety of dimension n, z = (z1, . . . , zn) a local coordinate system at a smooth point p ∈ X, and > a monomial order on Nn. Then we obtain a valuation ν = νz,> : OX,p \ {0} → Nn as follows; for f ∈ OX,p \ {0}, we expand it as a formal power series and set f = Xu∈Nn cuzu ν(f ) := min{ u cu 6= 0}, where the minimum is taken with respect to the monomial order >. Let L be a big line bundle on a projective variety X. Then we can define the Okounkov body ∆(L) of L as follows. ∼= OX,p. Then this isomorphism naturally induces Fix an isomorphism Lp ∼= OX,p for any k ≥ 0, and we have the map Lk p H 0(X, kL) \ {0} ֒→ Lk p \ {0} ∼= OX,p \ {0} ν → Nn. We write ν(H 0(X, kL)) ⊂ Nn for the image of H 0(X, kL) \ {0} by this map. Note that ν(H 0(X, kL)) does not depend on the choice of the isomorphism 2010 Mathematics Subject Classification. 14C20;14M25. Key words and phrases. Okounkov body, Seshadri constant, graded linear series. 1 2 Lp ATSUSHI ITO ∼= OX,p because ν maps any unit element in OX,p to 0 ∈ Nn. Then ΓL = ΓL,z,> := [k∈N {k} × ν(H 0(X, kL)) ⊂ N × Nn is a semigroup by construction. We define the Okounkov body of L with respect to z and > by ∆(L) = ∆z,>(L) : = closed convex hull(cid:16) [k≥1 1 k ν(H 0(X, kL))(cid:17) where Σ(ΓL) is the closed convex cone spanned by ΓL. = Σ(ΓL) ∩ ({1} × Rn), More generally, we can define the Okounkov body for a graded linear series. That is, for a graded linear series W• associated to a line bundle L on a (not necessarily projective) variety X, we set ΓW• = ΓW•,z,> := [k∈N {k} × ν(Wk) ⊂ N × Nn. We define the Okounkov body of W• with respect to z and > by ∆(W•) = ∆z,>(W•) := Σ(ΓW•,z,>) ∩ ({1} × Rn). Remark 1.1. In [LM], they use "admissible flags" instead of local coordinate systems. Essentially, there is no difference if > is the lexicographic order. Local coordinate systems are used in [BC] for instance. On the other hand, Demailly [De] defined an interesting invariant, Se- shadri constant, which measures the local positivity of an ample line bundle at a point. Seshadri constants relate to jet separation of adjoint bundles, Ross-Thomas' slope stabilities of polarized varieties [RT], Gromov width (an invariant in symplectic geometry) [MP], and so on. Nakamaye [Na] defined Seshadri constants at very general points for big line bundles. For a detailed treatment of Seshadri constants, we refer the reader to [La, Chapter 5] or [B+]. We extend the notion of Seshadri constants slightly. Usually, Seshadri constants are defined in a numerical way. However, we adopt an equivalent definition in terms of jet separation since we treat graded linear series in this paper. Let W• be a "birational" graded linear series associated to a line bundle L on a variety X (see Definition 3.2 for the definition of a birational graded linear series). Using jet separation (cf. [De, Theorem 6.4], [ELMNP, Propo- sition 6.6]), we define the Seshadri constant ε(W•; 1) of W• at a very general point by ε(W•; 1) := sup k>0 s(Wk; 1) k ∈ R+ ∪ {+∞}, where s(Wk; 1) is the supremum of integers s ≥ −1 such that the natural map Wk ֒→ H 0(X, Lk) → Lk ⊗ OX/ms+1 is surjective for a very general p OKOUNKOV BODIES AND SESHADRI CONSTANTS 3 point p ∈ X. When L is a big line bundle on a projective variety X and W• = {H 0(X, kL)}k, it is easy to show W• is birational. Hence we can consider the Seshadri constant of W• at a very general point, which we denote by ε(L; 1) or ε(X, L; 1). To relate Seshadri constants with Okounkov bodies, we introduce an in- variant for a convex set. For an integral polytope ∆ ⊂ Rn of dimension n, we can consider an invariant ε(∆; 1) := ε(X∆, L∆; 1), where (X∆, L∆) is the polarized toric variety corresponding to ∆. To generalize this invariant to an n-dimensional convex set ∆, we introduce a "monomial" birational graded linear series W∆,• associated to O(C×)n on (C×)n (see Definition 4.1). Thus we can define ε(∆; 1) := ε((C×)n, W∆,•; 1). In this paper, we show the following theorem, which states that Okounkov bodies give lower bounds on Seshadri constants. Theorem 1.2 (Special case of Theorem 5.8). Let X be a projective variety of dimension n, and fix a local coordinate system z = (z1, . . . , zn) on X at a smooth point and a monomial order > on Nn. For a big line bundle L on X, it holds that ε(L; 1) ≥ ε(∆z,>(L); 1). We explain the idea of the proof of Theorem 1.2 when L is ample. The strategy is to compare two graded linear series {H 0(X, kL)}k and W (ΓL)• := but the degree k-th part of the graded ring C[ΓL]. {Lu∈ν(H 0(X,kL)) Cxu}k. Note that Lu∈ν(H 0(X,kL)) Cxu ⊂ C[Nn] is nothing If ΓL is finitely generated, Anderson [An] showed thatLk H 0(X, kL) de- generates to C[ΓL]. In other words, (X, L) degenerates to the (not necessar- ily normal) toric variety (Proj C[ΓL], O(1)). Since Seshadri constants have a lower semicontinuity, we have ε(X, L; 1) ≥ ε(Proj C[ΓL], O(1); 1). It is easy to check that ε(Proj C[ΓL], O(1); 1) = ε(X∆(L), L∆(L); 1) = ε(∆(L); 1), and Theorem 1.2 follows in this case. Unfortunately, ΓL is not finitely generated in general. Hence we can- Instead, we degenerate linear series as in [CM], that is, we degenerate H 0(X, kL) to not use the degeneration of the section ring Lk H 0(X, kL). W (ΓL)k = Lu∈ν(H 0(X,kL)) Cxu for each k separately. Then we can show s(H 0(X, kL); 1) ≥ s(W (ΓL)k; 1) and compare the Seshadri constants of L and W (ΓL)•. Thus even if we show Theorem 1.2 only for an ample line bundle L, we have to treat the graded linear series W (ΓL)•. This is the reason why we extend the notion of Seshadri constants to graded linear series. The definition of Seshadri constants can be easily generalized to the multi- point case (cf. [La],[B+]). That is, we can define the Seshadri constants ε(W•; m) and ε(∆; m) for a weight m = (m1, . . . , mr) ∈ Rr +. Theorem 1.2 is generalized to the multi-point case and birational graded linear series as follows. 4 ATSUSHI ITO Theorem 1.3 (=Theorem 5.8). Let W• be a birational graded linear series associated to a line bundle L on an n-dimensional variety X. Fix a local coordinate system z = (z1, . . . , zn) on X at a smooth point and a monomial order > on Nn. Then ε(W•; m) ≥ ε(∆z,>(W•); m) holds for any r ∈ Z+ and m ∈ Rr +. This paper is organized as follows. In Section 2, we recall some notations and conventions. In Section 3, we define Seshadri constants of graded linear series and investigate basic properties. In Section 4, we study Seshadri constants of monomial graded linear series. In Section 5, we give the proof of Theorem 1.3. Throughout this paper, we consider varieties or schemes over the complex number field C. Acknowledgments. The author wishes to express his gratitude to his su- pervisor Professor Yujiro Kawamata for his valuable advice, comments and warm encouragement. He is grateful to Professors Robert Lazarsfeld and Ya- sunari Nagai for many valuable comments. He wishes to thank to Professor Marcin Dumnicki for answering his questions about references. He would also like to thank Yoshinori Gongyo, Makoto Miura, Yusuke Nakamura, Shinnosuke Okawa, Taro Sano, and Yusaku Tiba for helpful discussions and comments. The author was supported by the Grant-in-Aid for Scientific Research (KAKENHI No. 23-56182) and the Grant-in-Aid for JSPS fellows. 2. Notations and conventions We denote by N, Z, Q, R, and C the set of all natural numbers, integers, rational numbers, real numbers, and complex numbers respectively. In this paper, N contains 0. Set Z+ = N \ 0, R+ = {x ∈ R x > 0}. For a real number t ∈ R, the round up of t is denoted by ⌈t⌉ ∈ Z. For a subset S ⊂ Rn, we denote by Σ(S) the closed convex cone spanned by S. For t ∈ R, we set tS = { tu u ∈ S}. For another subset S′ ⊂ Rn, S + S′ = {u + u′ u ∈ S, u′ ∈ S′} is the Minkowski sum of S and S′. For simplicity of notation, we denote S + (−S′) by S − S′. For u′ ∈ Rn, S + u′ = {u + u′ u ∈ S} is the parallel translation of S by u′. For a convex set ∆ ⊂ Rn, the dimension of ∆ is the dimension of the affine space spanned by ∆. We denote by ∆◦ the interior of ∆. A subset P ⊂ Rn is called a polytope if it is the convex hull of a finite set in Rn. A polytope P is integral if all its vertices are contained in Zn. For a variety X, we say a property holds at a general point of X if it holds for all points in the complement of a proper algebraic subset. A property holds at a very general point of X if it holds for all points in the complement of the union of countably many proper subvarieties. Throughout this paper, a divisor means a Cartier divisor. Thus we use the words "divisor", "line bundle", and "invertible sheaf " interchangeably. For divisors D and D′, the inequality D ≥ D′ means D − D′ is effective. OKOUNKOV BODIES AND SESHADRI CONSTANTS 5 3. Seshadri constants of graded linear series In this section, we define Seshadri constants of graded linear series. Definition 3.1. Let L be a line bundle on a (not necessarily projective) vari- ety X, and W a subspace of H 0(X, L). For r ∈ Z+ and m = (m1, . . . , mr) ∈ Zr, we say that W separates m-jets at smooth r points p1, . . . , pr in X if the natural map W → L(cid:14)(cid:16)L ⊗ rOi=1 mmi+1 pi (cid:17) = rMi=1 L/mmi+1 pi L is surjective, where we regard mmi+1 = OX for mi ≤ −1. We say W generically separates m-jets if W separates m-jets at general r points in X. By definition, any W generically separates m-jets if mi ≤ −1 for any i. pi For W ⊂ H 0(X, L) and m = (m1, . . . , mr) ∈ Rr +, we define s(W ; m) ∈ R ∪ {+∞} to be s(W ; m) = sup{ t ∈ R W generically separates ⌈tm⌉-jets}, where ⌈tm⌉ = (⌈tm1⌉, . . . , ⌈tmr⌉). When dim W < +∞, we have s(W ; m) = max{ t ∈ R W generically separates ⌈tm⌉-jets} ∈ R. When W = H 0(X, L), we denote s(W, m) by s(L; m). Suppose X is a variety of dimension n and L is a line bundle on X. Let W• = {Wk}k∈N be a graded linear series associated to L, i.e., Wk is a subspace of H 0(X, kL) for any k ≥ 0 with W0 = C, such that Lk≥0 Wk is a graded subalgebra of the section ring Lk≥0 H 0(X, kL). When all Wk are finite dimensional, W• is called of finite dimensional type. Definition 3.2. A graded linear series W• on a variety X is birational if the function field K(X) of X is generated by { f /g ∈ K(X) f, g ∈ Wk, g 6= 0} over C for any k ≫ 0. When W• is of finite dimensional type, this is clearly equivalent to the condition that the rational map defined by Wk is birational onto its image for any k ≫ 0, which is Condition (B) in [LM, Definition 2.5]. Now we define Seshadri constants of graded linear series. Definition 3.3. Let W• be a birational graded linear series associated to a line bundle L on a variety X. For m = (m1, . . . , mr) ∈ Rr +, we define the Seshadri constant of W• at very general points with weight m to be ε(X, W•; m) = ε(W•; m) := sup k>0 s(Wk; m) k ∈ R+ ∪ {+∞}. Note that s(Wk; m) > 0 holds for k ≫ 0 by the birationality of W•. When Wk = H 0(X, kL) for any k, we denote it by ε(X, L; m) or ε(L; m). 6 ATSUSHI ITO Remark 3.4. The definition of Seshadri constants by jet separation is due to Theorem 6.4 in [De]. See also [La] or [ELMNP]. We treat the definition by blowing ups in Lemmas 3.9 and 3.10. ′ ′ ′ Remark 3.5. By definition, s(W ; m) ≤ s(W ′; m) holds for subspaces W ⊂ W ′ in H 0(X, L). Thus ε(W•; m) ≤ ε(W • associated to L if Wk ⊂ W •). •; m) holds for W•, W • for such W•, W k for any k (we write W• ⊂ W ′ ′ For an injection L ֒→ L′ between line bundles on X and a subspace W ⊂ H 0(X, L), s(W ; m) does not change if we regard W as a subspace of H 0(X, L′) because we consider jet separation at general points. Hence ε(W•; m) does not change for W• associated to L if we consider that W• is associated to L′. Let π : X ′ → X be a birational morphism and W• a graded linear series as- sociated to L on X. Then we can consider that W• is a graded linear series on X ′ associated to π∗L by the natural inclusion H 0(X, kL) ⊂ H 0(X ′, kπ∗L). By a similar reason to above, we have ε(X, W•; m) = ε(X ′, W•; m). By the following lemma, we may assume that W• is of finite dimensional type in many cases when we prove properties of ε(X, W•; m). Lemma 3.6. Let W• be a birational graded linear series. Suppose that W1,• ⊂ W2,• ⊂ · · · ⊂ Wl,• ⊂ · · · ⊂ W• is an increasing sequence of birational l=1 Wl,k for each k. Then it graded linear series in W• such that Wk = S∞ holds that ε(W•; m) = supl ε(Wl,•; m) = liml ε(Wl,•; m). Proof. Since ε(Wl,•; m) is monotonically increasing, liml ε(Wl,•; m) exists. The inequalities ε(W•; m) ≥ supl ε(Wl,•; m) ≥ liml ε(Wl,•; m) are clear. Thus it is enough to show ε(W•; m) ≤ liml ε(Wl,•; m). Fix k and a real number t < s(Wk; m). By definition, Wk generically separates ⌈tm⌉-jets. Hence Wl,k also generically separates ⌈tm⌉-jets for l ≫ 0 by the assumption Wk = Sl Wl,k. Thus it holds that s(Wk; m)/k = liml s(Wl,k; m)/k ≤ liml ε(Wl,•; m). By the definition of ε(W•; m), we have ε(W•; m) ≤ liml ε(Wl,•; m). (cid:3) In Definition 3.3, ε(W•; m) is defined by the supremum, but in fact it is the limit. Lemma 3.7. In Definition 3.3, ε(W•; m) = lim s(Wk; m) k holds. Proof. By Lemma 3.6, we may assume that W• is of finite dimensional type. To prove this lemma, it suffices to show that a sequence {s(Wk; m)}k is superadditive, i.e., s(Wk+l; m) ≥ s(Wk; m) + s(Wl; m) holds for k, l > 0 if s(Wk; m), s(Wl; m) ≥ 0. For simplicity, we set sk = s(Wk; m) in this proof. We prove the super- additivity only when r = 1, and write m ∈ R+ instead of m. When r > 1, the proof is essentially the same. Hence we leave the details to the reader. OKOUNKOV BODIES AND SESHADRI CONSTANTS 7 Fix a very general point p ∈ X. For each k, i ≥ 0, set Wk,i = Wk ∩ H 0(X, Lk ⊗ mi p) ⊂ H 0(X, Lk). For any s ∈ N, it is easy to show that Wk → Lk ⊗ OX /ms+1 p is surjective if and only if so is Wk,i → Lk ⊗ mi p/mi+1 p for each i ∈ {0, 1, . . . , s}. Fix an integer 0 ≤ i ≤ ⌈skm⌉ + ⌈slm⌉. Then there exist integers 0 ≤ i1 ≤ ⌈skm⌉, 0 ≤ i2 ≤ ⌈slm⌉ such that i = i1 + i2. Consider the following diagram Wk,i1 ⊗ Wl,i2 α / Lk ⊗ mi1 p /mi1+1 p ⊗ Ll ⊗ mi2 p /mi2+1 p (cid:9) γ Wk+l,i β / Lk+l ⊗ mi p/mi+1 p . In this diagram, α is surjective because i1 ≤ ⌈skm⌉, i2 ≤ ⌈slm⌉, and β is clearly surjective. Hence γ is also surjective for any i ∈ {0, 1, . . . , ⌈skm⌉ + ⌈slm⌉}. Thus Wk+l generically separates ⌈skm⌉ + ⌈slm⌉-jets, which means sk+l ≥ (⌈skm⌉ + ⌈slm⌉)/m ≥ sk + sl. (cid:3) Many properties of Seshadri constants of ample line bundles also hold for graded linear series. Lemma 3.8. For a birational graded linear series W• (resp. W to a line bundles L (resp. L′) on a variety X and m ∈ Rr ′ •) associated +, it holds that (1) ε(W (l) • ; m) = l · ε(W•; m) for l ∈ Z+, where W (l) • := Wkl. series associated to Ll defined by W (l) k (2) ε(W•; tm) = t−1ε(W•; m) for t ∈ R+. (3) ε(W • ; m) ≥ ε(W•; m) + ε(W •; m), where W ′′ ′ series associated to L ⊗ L′ defined by W Wk ⊗ W k → H 0(X, k(L ⊗ L′)). ′ (4) ε(W•; m) ≤ npvol(W•)/mn, where n = dim X, vol(W•) = limk dim Wk kn/n! , and mn =Pr i=1 mn i . is the graded linear ′′ • is the graded linear ′′ k := the image of / / 8 ATSUSHI ITO Proof. By Lemma 3.7, we have ε(W (l) • ; m) = lim k = lim k s(W (l) k ; m) k s(Wkl; m) k s(Wkl; m) kl s(Wk; m) k = l · ε(W•; m). = l · lim k = l · lim k Hence (1) is shown. We can show (2) easily from the definition and the following. s(Wk; tm) = sup{ s ∈ R Wk generically separates ⌈stm⌉-jets } = t−1 sup{ s ∈ R Wk generically separates ⌈sm⌉-jets } = t−1s(Wk; m). To prove (3), it suffices to show s(W k; m) for any k. We can show this by the argument similar to the proof of Lemma 3.7. We leave the details to the reader. k ; m) ≥ s(Wk; m) + s(W ′′ ′ To show (4), we fix a positive number 0 < t < ε(W•; m). By Lemma 3.7, the inequality s(Wk; m) > kt holds for any k ≫ 0. Thus Wk separates ⌈ktm⌉-jets for any k ≫ 0. In other words, Wk →Mi L ⊗ O/m⌈ktmi⌉+1 pi is surjective for very general p1, . . . , pr. This surjection induces dim Wk ≥ dimMi L ⊗ O/m⌈ktmi⌉+1 pi =Xi (cid:18)⌈ktmi⌉ + n n (cid:19) =Xi tnmn i n! kn + O(kn−1). Thus we have vol(W•) ≥ tnmn by k → +∞. We finish the proof by t → ε(W•; m). (cid:3) Definition 3.3 is a natural generalization of the well-known definition of the Seshadri constant (at very general points) for a nef and big line bundle (cf. [La, Theorem 5.1.17]). Lemma 3.9. For a nef and big line bundle L on a projective variety X and m = (m1, . . . , mr) ∈ Rr +, it holds that ε(X, L; m) = max{ t ≥ 0 µ∗L − t rXi=1 miEi is nef }, OKOUNKOV BODIES AND SESHADRI CONSTANTS 9 E1, . . . , Er are the exceptional divisors. Proof. The proof is essentially the same as that of [La, Theorem 5.1.17]. i=1 miEi is nef }, i.e., where µ : eX → X is the blowing up along very general r points on X and First, we show ε(X, L; m) ≤ max{ t ≥ 0 µ∗L − t Pr µ∗L−ε(X, L; m)Pr strict transform of C. It is enough to show (µ∗L−ε(X, L; m)Pr D ∈ Lk ⊗Ni 0. For each k, the line bundle Lk separates ⌈skm⌉-jets at very general points p1, . . . , pr, where sk := s(kL; m). Hence there exists an effective divisor such that D does not contain C. Thus it holds that i=1 miEi is nef. Fix a curve C ⊂ X and let eC ⊂ eX be the i=1 miEi).eC ≥ ⌈skmi⌉ pi m µ∗L.eC = L.C = k−1D.C ≥ k−1Xi ≥ k−1Xi ≥ k−1Xi multpi(D) · multpi(C) ⌈skmi⌉ multpi(C) skmi multpi(C) = k−1skXi miEi.eC, We show the opposite inequality. First, assume L is ample. Let p1, . . . , pr be very general r points in X. Fix a rational number 0 < t = a/b < i=1 miEi is nef } with positive integers a, b. Then where multpi is the multiplicity at pi. By k → +∞, we have µ∗L.eC ≥ ε(X, L; m)Pr i=1 miEi.eC. max{ t ≥ 0 µ∗L − t Pr bµ∗L − aPr by a sufficiently large positive integer, we may assume bµ∗L −Pr i=1 miEi is an ample R-line bundle on eX. Multiplying a and b i=1⌈ami⌉Ei is ample. By Serre's vanishing theorem, there exists a natural number N such that H 1(X, kbL ⊗Oi mk⌈ami⌉ pi ) = H 1(eX, k(bµ∗L −Xi ⌈ami⌉Ei)) = 0 for every k ≥ N . This means kbL separates (k⌈am1⌉−1, . . . , k⌈amr⌉−1)-jets at p1, . . . , pr, that is, s(kbL; m) kb ≥ min i k⌈ami⌉ − 1 kbmi . By k → +∞, we have ε(L; m) ≥ lim k min i k⌈ami⌉ − 1 kbmi = a b = t. By t → max{ t ≥ 0 µ∗L − t Pr max{ t ≥ 0 µ∗L − t Pr i=1 miEi is nef }. i=1 miEi Next we show the nef and big case. Since L is nef and big, there exists an effective divisor E on X such that L − sE is ample for any 0 < s ≪ 1. Fix a rational number 0 < s ≪ 1 and take a sufficiently divisible integer l ∈ Z+ is nef }, we obtain ε(L; m) ≥ 10 ATSUSHI ITO such that ls ∈ Z. Then ε(l(L − sE); m) ≤ ε(lL; m) = l · ε(L; m) holds by Remark 3.5 and Lemma 3.8 (1). Since l(L − sE) is ample, we have ε(l(L − sE); m) = max{ t ≥ 0 µ∗(l(L − sE)) − t Xi = l · max{ t ≥ 0 µ∗(L − sE) − t Xi miEi is nef } miEi is nef }. Hence ε(L; m) ≥ max{ t ≥ 0 µ∗(L − sE) − t Pr s → 0, we have ε(L; m) ≥ max{ t ≥ 0 µ∗L − t Pr we define ε(X, L; m) := max{ t ≥ 0 µ∗L − t Pr For a line bundle L on a projective variety X, W• = {H 0(X, kL)}k is birational if and only if L is big. For a nef but not big line bundle L on X, For projective varieties, we can describe Seshadri constants of graded linear series by using those of nef line bundles as follows. This is a simple generalization of (some part of) [ELMNP, Propositions 6.4, 6.6], which treat ε(L; 1) for a big line bundle L, although their notations are slightly different. i=1 miEi is nef } holds. By i=1 miEi is nef }. (cid:3) i=1 miEi is nef } = 0. Lemma 3.10. Let W• be a birational graded linear series associated to a line bundle L on a projective variety X. For each k > 0, set Mk = µ∗ k(kL) − Fk, where µk : Xk → X is a resolution of the base ideal bk := the image of Wk ⊗ L−k → OX and OXk (−Fk) := µ−1 k bk. Set Mk = 0 if Wk = 0. Then it holds that ε(X, W•; m) = sup k>0 ε(Xk, Mk; m) k = lim k>0 ε(Xk, Mk; m) k . Proof. By definition, Mk is nef and ε(Mk, m) ≥ 0 for k > 0. To prove the existence of limk ε(Xk, Mk; m)/k and the second equality in the above statement, it suffices to show the superadditivity of {ε(Xk, Mk; m)}k, i.e., (†) ε(Xk+l, Mk+l, m) ≥ ε(Xk, Mk; m) + ε(Xl, Ml; m) for k, l > 0 such that Wk, Wl 6= 0. To show (†), fix such k, l > 0. We can take a common resolution of bk, bl, and bk+l. Since Mk+l ≥ Mk + Ml, (†) follows from Lemmas 3.8 (3) or 3.9. kWk ⊂ Mk, it holds that Next we show the first equality. Since µ∗ s(Wk; m) ≤ s(Mk; m) ≤ ε(Mk; m) for any k. Thus we have ε(X, W•; m) = lim s(Wk; m) k ≤ lim ε(Mk; m) k . To show the opposite inequality, we use Lemma 3.9. Since W• is bira- tional, the morphism ϕk : Xk → Pdim Wk defined by µ∗ kWk is birational onto its image for k ≫ 0. Denote the image of ϕk by Yk. By Lemma 3.9, OKOUNKOV BODIES AND SESHADRI CONSTANTS 11 ε(Xk, Mk; m) = ε(Yk, OYk (1); m) holds because ϕk is birational. Further- more W l k → Wkl. This implies s(OYk (l); m) = s(W l k; m) ≤ s(Wkl; m) for l ≫ 0. Thus we have k = H 0(Yk, OYk (l)) for l ≫ 0, where W l k is the image of W ⊗l ε(Yk, OYk (1); m) k s(OYk (l); m) ε(Xk, Mk; m) k = = 1 k lim l l s(Wkl; m) kl ≤ lim = ε(W•; m). (cid:3) 4. Monomial graded linear series on (C×)n Definition 4.1. Let n be a positive integer. For a subset S ⊂ Rn, we set VS := Mu∈S∩Zn Cxu, which is a subspace of Lu∈Zn Cxu = H 0((C×)n, O(C×)n ). For a convex set ∆ in Rn, we define the monomial graded linear series W∆,• associated to O(C×)n by W∆,k := Vk∆ ⊂ H 0((C×)n, O(C×)n ). It is easy to see that W∆,• is birational if and only if dim ∆ = n. Definition 4.2. For an n-dimensional convex set ∆ ⊂ Rn and m ∈ Rr define ε(∆; m) := ε(W∆,•, m) ∈ R+. +, we Remark 4.3. For an integral polytope ∆ ⊂ Rn of dimension n, clearly ε(∆; m) = ε(X∆, L∆; m) holds, where (X∆, L∆) = (ProjLk Vk∆, O(1)) is the polarized toric variety corresponding to ∆. We show some basic properties of ε(∆; m) in this section. Lemma 4.4. The following hold for subsets S1, S2 and n-dimensional con- vex sets ∆1, ∆2 in Rn such that S1 ⊂ S2, ∆1 ⊂ ∆2. (1) s(VS1; m) ≤ s(VS2; m), ε(∆1; m) ≤ ε(∆2; m). (2) s(VS1+u; m) = s(VS1; m) for u ∈ Zn. (3) ε(t∆1; m) = t · ε(∆1; m), ε(∆1 + u; m) = ε(∆1; m) for t ∈ R+ ∩ Q and u ∈ Qn. Proof. (1) is clear. (2) immediately follows from the diagram VS1 ≀ C[x±1 1 , . . . , x±1 n ] (cid:9) ≀ ×xu VS1+u / C[x±1 1 , . . . , x±1 n ]. / / / 12 ATSUSHI ITO To show (3), choose sufficiently divisible l ∈ Z+ such that lt ∈ Z and lu ∈ Zn. Then we have ε(t∆1; m) = lim k s(Vkt∆1; m) k = lim k s(Vklt∆1; m) kl = t lim k s(Vklt∆1; m) klt = t · ε(∆1; m) and ε(∆1 + u; m) = lim k = lim k = lim k = lim k s(Vk(∆1+u); m) k s(Vkl(∆1+u); m) kl s(Vkl∆1+klu; m) kl s(Vkl∆1; m) kl = ε(∆1; m). The last but one equality follows from (2) since klu ∈ Zn. (cid:3) Lemma 4.5. For an n-dimensional convex set ∆ ⊂ Rn, ε(∆; m) = ε(∆◦; m) holds. Proof. It is enough to show ε(∆; m) ≤ ε(∆◦; m). Fix u ∈ ∆◦ ∩ Qn. By the convexity of ∆, we have ∆ − u ⊂ t(∆◦ − u) for t > 1. Thus it holds that ε(∆; m) = ε(∆ − u; m) ≤ ε(t(∆◦ − u); m) = t · ε(∆◦; m) for t > 1 in Q by Lemma 4.4. By t → 1, the lemma is proved. (cid:3) The property (3) in Lemma 4.4 holds for any t ∈ R+ and u ∈ Rn. Lemma 4.6. For an n-dimensional convex set ∆ ⊂ Rn, u ∈ Rn, and t ∈ R+, it holds that ε(∆ + u; m) = ε(∆; m), ε(t∆; m) = t · ε(∆; m). Proof. Fix u′ ∈ ∆◦ ∩ Qn. As in the proof of Lemma 4.5, ∆ − u′ ⊂ (1 + t′)(∆◦ − u′) holds for t′ > 0. Translating the convex sets by u + u′, we have ∆ + u ⊂ (1 + t′)(∆◦ − u′) + u + u′. Choose u′′ ∈ Qn such that (u + u′) − u′′ ∈ t′(∆◦ − u′). Then we have ∆ + u ⊂ (1 + 2t′)(∆◦ − u′) + u′′. Hence for t′ ∈ R+ ∩ Q, it holds that ε(∆ + u; m) ≤ ε((1 + 2t′)(∆◦ − u′) + u′′; m) = (1 + 2t′)ε(∆; m) by Lemmas 4.4 (3) and 4.5. Thus we obtain ε(∆ + u; m) ≤ ε(∆; m) by t′ → 0. Since the opposite inequality ε(∆ + u; m) ≥ ε(∆; m) also holds similarly, ε(∆ + u; m) = ε(∆; m) follows. OKOUNKOV BODIES AND SESHADRI CONSTANTS 13 For t1, t2 ∈ Q such that 0 < t1 ≤ t ≤ t2, we have inclusions t1(∆ − u′) ⊂ t(∆ − u′) ⊂ t2(∆ − u′). Thus it holds that ε(t1(∆ − u′); m) ≤ ε(t(∆ − u′); m) ≤ ε(t2(∆ − u′); m). By Lemma 4.4 (3) and the first equality of this lemma, which we have already proved, ε(t(∆ − u′); m) = ε(t∆; m) and ε(ti(∆ − u′); m) = ti · ε(∆; m) for i = 1, 2. Combining these inequalities, we have t1 · ε(∆; m) ≤ ε(t∆; m) ≤ t2 · ε(∆; m). By t1, t2 → t, ε(t∆; m) = t · ε(∆; m) follows. (cid:3) Lemma 4.7. Let ∆1 ⊂ ∆2 ⊂ · · · ⊂ ∆i ⊂ · · · be an increasing sequence of i=1 ∆i. Then it holds that n-dimensional convex sets in Rn, and set ∆ = S∞ ε(∆; m) = supi ε(∆i; m) = limi ε(∆i; m) for any m ∈ Rr +. Proof. This lemma follows from Lemma 3.6 immediately. (cid:3) 5. Okounkov bodies and Seshadri constants In this section, we prove Theorem 1.3. 5.1. Preliminary. For a subsemigroup Γ in N × Nn and k ∈ N, set ∆(Γ) = Σ(Γ) ∩ ({1} × Rn), Γk = Γ ∩ ({k} × Nn). Recall that Σ(Γ) is the closed convex cone in R × Rn spanned by Γ. We regard ∆(Γ) and Γk as subsets in Rn and Nn respectively in a natural way. Definition 5.1. For a semigroup Γ in N × Nn, we define a graded linear series W (Γ)• on (C×)n associated to O(C×)n by W (Γ)k := VΓk . The birationality of W (Γ)• is interpreted as the following conditions of Γ. Definition 5.2. A semigroup Γ in N × Nn is birational if i) Γ0 = {0} ∈ Nn, ii) Γ generates Z × Zn as a group. These conditions are (2.3) and (2,5) in [LM] respectively. It is easy to check that Γ is birational if and only if so is the graded linear series W (Γ)•. Let > be a monomial order on Nn, i.e., > is a total order on Nn such that (i) for every u ∈ Nn \ 0, u > 0 holds, and (ii) if v > u and w ∈ Nn, then w + v > w + u. In this paper, v > u does not contain the case v = u. Let X be a variety of dimension n and z = (z1, . . . , zn) a local coordinate system at a smooth point p ∈ X. For a birational graded linear series W• associated to a line bundle L on X, we can define a semigroup ΓW• = ΓW•,z,> ⊂ N × Nn by using ν = νz,> : OX,p \ 0 → Nn as in Introduction. In [LM, Lemma 2.12], they assume that W• satisfies "Condition (C)", which seems to be a slightly stronger condition than being birational(= Condition (B) in [LM]), to show that ΓW•,z,> generates Z × Zn as a group for any z. But we can show that it is enough to assume W• is birational. 14 ATSUSHI ITO Lemma 5.3. Let W• be a birational graded linear series associated to a line bundle L on a variety X. Then ΓW•,z,> is birational for any local coordinate system z at any smooth point p ∈ X and any monomial order > on Nn. Proof. The condition i) in Definition 5.2 is clearly satisfied. Thus it is enough to show that ΓW• generates Z×Zn as a group. Fix k ≫ 0. Then the function field K(X) is generated by { f /g ∈ K(X) f, g ∈ Wk, g 6= 0} over C because W• is birational. Hence for any F ∈ K(X) \ {0}, we can write F = G/H, where G, H are written as some polynomials over C of some elements in { f /g ∈ K(X) f, g ∈ Wk, g 6= 0}. Therefore we can write F = G′/H ′ for some G′, H ′ ∈ Wkl for some l ∈ Z+. Thus ν(F ) = ν(G′) − ν(H ′) ∈ ν(Wkl) − ν(Wkl) ⊂ Zn. Since the valuation ν : K(X) \ {0} → Zn is surjective (note that ν is naturally extended to K(X)\{0}), the group Zn is generated by {ν(Wkl)−ν(Wkl)}l∈N. Thus the subgroup {0}× Zn in Z× Zn is generated by {0} × {ν(Wkl) − ν(Wkl)}l ⊂ ΓW• − ΓW•. On the other hand, sk ∈ Wk \ {0} and sk+1 ∈ Wk+1 \ {0} induce the element (1, ν(sk+1) − ν(sk)) ∈ ΓW• − ΓW• ⊂ Z × Zn. Since the group Z × Zn is generated by {0} × Zn and (1, ν(sk+1) − ν(sk)), the semigroup ΓW• generates Z × Zn as a group. (cid:3) Let W• be a birational graded linear series associated to a line bundle L on X. We define the Okounkov body ∆(W•) = ∆z,>(W•) of W• with respect to z and > as in Introduction. That is, ∆z,>(W•) = ∆(ΓW•,z,>). Thus ∆(W•) is an n-dimensional closed convex set in Rn because ΓW• is birational by Lemma 5.3. The following lemma is similar to [An, Lemma 5.2]. Lemma 5.4. For a monomial order > on Nn and a finite set S in Nn, there exists α ∈ Zn + satisfying the following: For u ∈ S and v ∈ Nn such that v > u, it holds that α · v > α · u, where α · u, α · v are the usual inner products. Proof. For each u ∈ S, set Su = {v ∈ Nn v > u}. Let Iu be the ideal in the polynomial ring C[Nn] = C[x1, . . . , xn] generated by { xv v ∈ Su}. By Hilbert's basis theorem, Iu is generated by xvu1, . . . , xvuku for some ku ∈ N and vu1, . . . , vuku ∈ Su. Therefore any v ∈ Su is contained in vuj + Nn for some j. We use the following result by Robbiano. Theorem 5.5 ([Ro, Theorem 2.5]). For a monomial order > on Nn, there exist an integer s ∈ {1, . . . , n} and u1, ..., us ∈ Rn which satisfy the following: For u, v ∈ Nn, v > u if and only if π(v) >lex π(u), where π : Nn → Rs; u 7→ (u1 · u, . . . , us · u) and >lex is the lexicographic order on Rs. Let e1, . . . , en be the standard basis of Zn and consider the above u1, . . . , us and π. For any γ >lex δ in Rs, the following holds from the definition of the OKOUNKOV BODIES AND SESHADRI CONSTANTS 15 lexicographic order: For β = (β1, . . . , βs) ∈ Rs have β · γ > β · δ. + such that β1 ≫ · · · ≫ βs, we Hence we can take β ∈ Rs + such that • β · π(ei) > 0 for 1 ≤ i ≤ n, • β · π(vuj) > β · π(u) for u ∈ S and 1 ≤ j ≤ ku, since π(ei) >lex π(0) = 0 and π(vuj) >lex π(u) (we can take common β because S is a finite set). Since β · π(u′) = (β1u1 + · · · + βsus) · u′ for u′ ∈ Nn, we have (∗) α′ · ei > 0, α′ · vuj > α′ · u for 1 ≤ i ≤ n, u ∈ S and 1 ≤ j ≤ ku if we take α′ ∈ Qn sufficiently close to β1u1 + · · · + βsus. Set α := N α′ ∈ Zn for a sufficiently divisible positive integer N . By (∗), it follows that α ∈ Zn + and α · vuj > α · u for u ∈ S and j. We show this α satisfies the condition in the statement of this lemma. Fix u ∈ S and v ∈ Nn such that v > u, i.e., v ∈ Su. Then v ∈ vuj + Nn for some 1 ≤ j ≤ ku. Thus we have α · v = α · vuj + α · (v − vuj) ≥ α · vuj > α · u. (cid:3) 5.2. Proof of Theorem 1.3. Since Seshadri constants are lower semicon- tinuous (cf. [La, Example 5.1.11]), degenerations are useful to bound Se- shadri constants from below. For example, Biran [Bi] degenerates varieties to reducible schemes, and gives lower bounds of (multi-point) Seshadri con- stants on P2. In [It], the author uses toric degenerations to obtain lower bounds of Seshadri constants on some non-toric varieties. We also use degenerations to prove Theorem 1.3. Although we would like to degenerate Lk Wk to C[ΓW•] = Lk Vν(Wk), the semigroup ΓW• is not finitely generated in general, even if Lk Wk is finitely generated. Instead, we degenerate Wk to Vν(Wk) for each k separately. Then we can use the lower semicontinuity of jet separation, as in [CM]. This is the reason why we define Seshadri constants in terms of jet separation. First, we show that the Seshadri constants of W (Γ)• and W∆(Γ),• coincide for a birational semigroup Γ. Lemma 5.6. For a birational semigroup Γ ⊂ N × Nn and m ∈ Rr ε(∆(Γ); m) = ε(W (Γ)•; m). +, we have Proof. By definition, it holds that ε(∆(Γ); m) = lim s(Vk∆(Γ); m) k , ε(W (Γ)•; m) = lim s(VΓk ; m) k . Since Γk is contained in k∆(Γ) for any k, ε(∆(Γ); m) ≥ ε(W (Γ)•; m) is clear. Hence it is enough to show the opposite inequality. When Γ is finitely generated, there exists u = (l, u) ∈ Γ ⊂ N × Nn such that (Σ(Γ) + u) ∩ (N × Nn) ⊂ Γ by [Kh, §3, Proposition 3] (see also [LM, Subsection 2.1]). This induces (k∆(Γ)∩ Nn)+ u ⊂ Γk+l, from which we have 16 ATSUSHI ITO s(Vk∆(Γ); m) = s(V(k∆(Γ)∩Nn)+u; m) ≤ s(VΓk+l; m). Thus we obtain ε(W (Γ)•; m) = lim k s(VΓk ; m) k = lim k ≥ lim k s(VΓk+l; m) k s(Vk∆(Γ); m) k = ε(∆(Γ); m). In the general case, we take an increasing sequence Γ1 ⊂ Γ2 ⊂ · · · ⊂ Γ such that each Γi is a finitely generated birational subsemigroup of Γ and Si≥1 Γi = Γ. Then it is easy to show Si ∆(Γi)◦ = ∆(Γ)◦. By Lemmas 4.5 and 4.7, we have ε(∆(Γ); m) = ε(∆(Γ)◦; m) = lim ε(∆(Γi)◦; m) = lim ε(∆(Γi); m). Since each Γi is finitely generated, we can apply the first part of the proof of this lemma. Hence we have ε(∆(Γi); m) = ε(W (Γi)•; m) ≤ ε(W (Γ)•; m). Thus we obtain ε(∆(Γ); m) = lim ε(∆(Γi); m) ≤ ε(W (Γ)•; m). (cid:3) Broadly speaking, the geometrical meaning of Lemma 5.6 is that Seshadri constants of ample line bundles (on non-normal toric varieties) at very gen- eral points do not change by normalizations. In fact, when Γ is finitely generated, (ProjLk W∆(Γ),k, O(1)) = (Proj C[Σ(Γ) ∩ (N × Nn)], O(1)) is nothing but the normalization of the toric variety (ProjLk W (Γ)k, O(1)) = (Proj C[Γ], O(1)). Now we show the key proposition by using degeneration of global sections as [CM]. Roughly speaking, this proposition states that W ⊂ H 0(X, L) generically separates no less jets than Vν(W ). Proposition 5.7. Let L be a line bundle on an n-dimensional variety X. Let ν = νz,> be the valuation map defined by a local coordinate system z = (z1, . . . , zn) at a smooth point p ∈ X and a monomial order > on Nn. Then s(W ; m) ≥ s(Vν(W ); m) holds for any subspace W of H 0(X, L) and any m ∈ Rr +. Proof. By considering an increasing sequence of finite dimensional subspaces in W , we may assume dim W < +∞. Let π : U → Cn be the ´etale morphism defined by z1, . . . , zn in an open neighborhood U ⊂ X of p. By the morphism π, we can identify Oan X,p with Oan Cn,0 = C{x1, . . . , xn}, where x1, . . . , xn are the coordinates on Cn such that π∗xi = zi. Then we can regard W as a subspace of C{x1, . . . , xn} by W ֒→ Lp X,p \ {0} = C{x1, . . . , xn} \ {0} → Nn naturally. ∼= OX,p ֒→ C{x1, . . . , xn}. Note that ν is extended to Oan Choose and fix fu ∈ ν−1(u) ∩ W for each u ∈ ν(W ). Then it holds that V =Lu∈ν(W ) Cfu because #ν(W ) = dim W (cf. [LM] or [BC]). Since ν(W ) is a finite set, there exists α ∈ Zn v > u for u ∈ ν(W ) and v ∈ Nn, it holds that α · v > α · u. + satisfying the following by Lemma 5.4; if OKOUNKOV BODIES AND SESHADRI CONSTANTS 17 The vector α induces the action ◦ of C× on C{x1, . . . , xn} by t ◦ xu := tα·uxu for t ∈ C× and u ∈ Nn. For fu = Pv cuvzv = Pv cuvxv (note we identify zi and xi), the regular function t ◦ fu tα·u = t−α·uXv cuvtα·vxv =Xv cuvtα·v−α·uxv on a neighborhood of C× × {0} (in C× × Cn) is naturally extended to a regular function on a neighborhood U of C × {0}. Note that α · v − α · u ≥ 0 if cuv 6= 0. We denote the regular function by Fu. Set W =Lu∈ν(W ) CFu. We prove this proposition only for r = 1. When r > 1, the proof is similar. Thus we leave the details to the reader. Choose a very general section σ of the projection U → C onto the first factor. Let I be the ideal sheaf corresponding to σ(C) on U ⊂ C × Cn. For s ≥ 0, we consider the map φ : W ⊗C C{t} → Oan U → Oan U /I s+1 of flat sheaves over C. For t ∈ C, we write Wt := W ⊗ C{t}{t}×Cn and φt := φ{t}×Cn : Wt → Oan U /I s+1{t}×Cn = OCn/ms+1 σ(t). By the flatness, φt is surjective for very general t if so is φ0. Thus if W0 separates s-jets at σ(0), then Wt also separates s-jets at σ(t) for very general t. Since σ is a very general section, we have s(Wt; m) ≥ s(W0; m) for t in a neighborhood of 0. Since there is a natural identification of Wt and W for t ∈ C× by the action ◦, we have s(W ; m) = s(Wt; m). On the other hand, W0 = Vν(W ) ⊂ C[x1, . . . , xn] since Fu = cuuxu + t · (higher term) for some cuu 6= 0. From these inequalities, we have s(W ; m) ≥ s(Vν(W ); m). (cid:3) Now we can show the main theorem easily. Theorem 5.8 (=Theorem 1.3). Let W• be a birational graded linear series associated to a line bundle L on an n-dimensional variety X. Fix a local coordinate system z = (z1, . . . , zn) on X at a smooth point and a monomial order > on Nn. Then ε(W•; m) ≥ ε(∆z,>(W•); m) holds for any r ∈ Z+ and m ∈ Rr +. Proof. Let Γ := ΓW•,z,> ⊂ N × Nn be the semigroup defined by W•, z, and >. Then we have Γk = ν(Wk) ⊂ Nn and ∆(Γ) = ∆z,>(W•) by definition. By Proposition 5.7, it holds that ε(W•; m) = lim s(Wk; m) k ≥ lim s(Vν(Wk); m) k = lim s(VΓk ; m) k . Since Γ is birational by Lemma 5.3, we have lim s(VΓk ; m) k = ε(W (Γ)•; m) = ε(∆(Γ); m) = ε(∆z,>(W•); m) from Lemma 5.6. Thus ε(W•; m) ≥ ε(∆z,>(W•); m) holds. (cid:3) 18 ATSUSHI ITO Remark 5.9. The inequality in Theorem 5.8 is not equality in general. Let L be an ample line bundle on a projective variety X. Let z be a local coordinate system at p and E the prime divisor on X defined by the first coordinate z1 of z around p. If > is the lexicographic order, ∆z,>(L) is contained in [0, a] × Rdim X−1, where a = sup{ t > 0 L − tE is effective} (cf. [LM]). Hence ε(∆z,>(L); 1) ≤ a holds (cf. [It, Theorem 3.6]). If we choose z so that E ∈ mL for m ≫ 0, we have ε(∆z,>(L); 1) ≤ a = 1/m < ε(L; 1). Remark 5.10. See [LM, Remark 5.5] for another relation between Okounkov bodies and Seshadri constants, though the relation is not written explicitly there. The relation also holds for a birational graded linear series. References [An] D. Anderson, Okounkov bodies and toric degenerations , arXiv:1001.4566, to appear in Math. Ann. [B+] T. Bauer, S. Di Rocco, B. Harbourne, M. Kapustka, A.L. Knutsen, W. Syzdek, and T. Szemberg, A primer on Seshadri constants, Interactions of classical and numerical algebraic geometry, 33 -- 70, Contemp. Math., 496, Amer. Math. Soc., Providence, RI, 2009. [Bi] P. Biran, Constructing new ample divisors out of old ones, Duke Math. J. 98 (1999), no. 1, 113 -- 135. [BC] S. Boucksom and H. Chen, Okounkov bodies of filtered linear series, Compos. Math. 147 (2011), no. 4, 1205 -- 1229. [CM] C. Ciliberto and R. Miranda, Degenerations of planar linear systems, J. Reine Angew. Math. 501 (1998), 191 -- 220. [De] J.P. Demailly, Singular Hermitian metrics on positive line bundles, Complex alge- braic varieties (Bayreuth, 1990), Lect. Notes Math. 1507, Springer-Verlag, 1992, pp. 87 -- 104. [ELMNP] L. Ein, R. Lazarsfeld, M. Mustat¸a, M. Nakamaye, and M. Popa, Restricted volumes and base loci of linear series, Amer. J. Math. 131 (2009), no. 3, 607 -- 651. [It] A. Ito, Seshadri constants via toric degenerations, arXiv:1202.6664, to appear in J. Reine Angew. Math. [KK] K. Kaveh and A. G. Khovanskii, Newton-Okounkov bodies, semigroups of integral points, graded algebras and intersection theory, Ann. of Math. (2) 176 (2012), no. 2, 925 -- 978. [Kh] A. G. Khovanskii, The Newton polytope, the Hilbert polynomial and sums of finite sets, Funct. Anal. Appl. 26 (1992), no. 4, 276 -- 281 (1993). [La] R. Lazarsfeld, Positivity in algebraic geometry I, Springer Verlag, 2004. [LM] R. Lazarsfeld and M. Mustat¸a, Convex bodies associated to linear series, Ann. Sci. Ec. Norm. Super. (4) 42 (2009), no. 5, 783 -- 835. [MP] D. McDuff and L. Polterovich, Symplectic packings and algebraic geometry, Invent. Math. 115 (1994), no. 3, 405 -- 434. [Na] M. Nakamaye, Base loci of linear series are numerically determined, Trans. Amer. Math. Soc. 355 (2003), no. 2, 551 -- 566. [Ok1] A. Okounkov, Brunn-Minkowski inequality for multiplicities, Invent. Math. 125 (1996), no. 3, 405 -- 411. [Ok2] A. Okounkov, Why would multiplicities be log-concave?, in The orbit method in geometry and physics, Progr. Math., 213, 2003, 329 -- 347, [Ro] L. Robbiano, On the theory of graded structures, J. Symbolic Comput. 2 (1986), no. 2, 139 -- 170. OKOUNKOV BODIES AND SESHADRI CONSTANTS 19 [RT] J. Ross and R. P. Thomas, A study of the Hilbert-Mumford criterion for the stability of projective varieties, J. Algebraic Geom. 16 (2007), no. 2, 201 -- 255. Graduate School of Mathematical Sciences, the University of Tokyo, 3-8-1 Komaba, Meguro-ku, Tokyo 153-8914, Japan. E-mail address: [email protected]
1302.0048	1	1302	2013-02-01T00:23:27	Systems of parameters and holonomicity of A-hypergeometric systems	[ "math.AG", "math.AC" ]	The main result is an elementary proof of holonomicity for A-hypergeometric systems, with no requirements on the behavior of their singularities, originally due to Adolphson [Ado94] after the regular singular case by Gelfand and Gelfand [GG86]. Our method yields a direct de novo proof that A-hypergeometric systems form holonomic families over their parameter spaces, as shown by Matusevich, Miller, and Walther [MMW05].	math.AG	math	SYSTEMS OF PARAMETERS AND HOLONOMICITY OF A-HYPERGEOMETRIC SYSTEMS CHRISTINE BERKESCH, STEPHEN GRIFFETH, AND EZRA MILLER Abstract. The main result is an elementary proof of holonomicity for A-hypergeometric systems, with no requirements on the behavior of their singularities, originally due to Adolphson [Ado94] after the regular singular case by Gelfand and Gelfand [GG86]. Our method yields a direct de novo proof that A-hypergeometric systems form holonomic fam- ilies over their parameter spaces, as shown by Matusevich, Miller, and Walther [MMW05]. Introduction An A-hypergeometric system is the D-module counterpart of a toric ideal. Solutions to A-hypergeometric systems are functions, with a fixed infinitesimal homogeneity, on an affine toric variety. The solution space of an A-hypergeometric system behaves well in part because the system is holonomic, which in particular implies that the vector space of germs of analytic solutions at any nonsingular point has finite dimension. This note provides an elementary proof of holonomicity for arbitrary A-hypergeometric systems, relying only on the statement that a module over the Weyl algebra in n variables is holonomic precisely when its characteristic variety has dimension at most n (see [Gab81] or [BGK+87, Theorem 1.12]), along with standard facts about transversality of subvari- eties and about Krull dimension. In particular, our proof requires no assumption about the singularities of the A-hypergeometric system; equivalently, the associated toric ideal need not be standard graded. Holonomicity was proved in the regular singular case by Gelfand and Gelfand [GG86], and later by Adolphson [Ado94, §3] regardless of the be- havior of the singularities of the system. Adolphson's proof relies on careful algebraic analysis of the coordinate rings of a collection of varieties whose union is the character- istic variety of the system. Another proof of the holonomicity of an A-hypergeometric system, by Schulze and Walther [SW08], yields a more general result: for a weight vec- tor L from a large family of possibilities, the L-characteristic variety for the L-filtration is a union of conormal varieties and hence has dimension n; holonomicity follows when L = (0, . . . , 0, 1, . . . , 1) induces the order filtration on the Weyl algebra. The L-filtration method uses an explicit combinatorial interpretation of initial ideals of toric ideals, which requires a series of technical lemmas. Holonomicity of A-hypergeometric systems forms part of the statement and proof, by Matusevich, Miller, and Walther [MMW05], that A-hypergeometric systems determine Date: 22 January 2013. 2010 Mathematics Subject Classification. Primary: 33C70; Secondary: 13N10, 14M25, 16S32, 33C99. EM had support from NSF grant DMS-1001437. SG acknowledges the financial support of Fondecyt Proyecto Regular 1110072. 1 2 CHRISTINE BERKESCH, STEPHEN GRIFFETH, AND EZRA MILLER holonomic families over their parameter spaces. The new proof of that statement here serves as a model suitable for generalization to hypergeometric systems for reductive groups, in the sense of Kapranov [Kap98]. The main step (Theorem 1.2) in our proof is an easy geometric argument showing that the Euler operators corresponding to the rows of an integer matrix A form part of a system of parameters on the product kn × XA, where k is any algebraically closed field and XA is the toric variety over k determined by A. This observation leads quickly in Section 2 to the conclusion that the characteristic variety of the associated A-hypergeometric system has dimension at most n, and hence that the system is holonomic. Since the algebraic part of the proof holds when the entries of β are considered as independent variables that commute with all other variables, the desired stronger consequence is immediate: the A- hypergeometric system forms a holonomic family over its parameter space (Theorem 2.1). 1. Systems of parameters via transversality Fix a field k. Let x = x1, . . . , xn and ξ = ξ1, . . . , ξn be sets of coordinates on kn and let xξ denote the column vector with entries x1ξ1, . . . , xnξn. Given a rectangular matrix L with n columns, write Lxξ for the vector of bilinear forms given by multiplying L times xξ. Lemma 1.1. Let k2n = kn ξ be a subvariety. If L is an ℓ × n matrix with entries in k, then the variety Var(Lxξ) of Lxξ in k2n is transverse to kn × X at any smooth point of kn × X whose ξ-coordinates are all nonzero. ξ have coordinates (x, ξ) and let X ⊆ kn x × kn Proof. It suffices to prove the statement after passing to the algebraic closure of k, so assume k is algebraically closed. Let (p, q) be a smooth point of kn × X that lies in Var(Lxξ) and has all coordinates of q nonzero. The tangent space to kn × X at (p, q) contains kn × {0}. The tangent space T(p,q) to Var(Lxξ) is the kernel of the ℓ × 2n matrix [L(q) L(p)], where L(p) (respectively, L(q)) is the ℓ×n matrix that results after multiplying each column of L by the corresponding coordinate of p (respectively, q). Since the q coordinates are all non-zero, T(p,q) projects surjectively onto the last n coordinates; indeed, if η ∈ kn x with L(q)y + L(p)η = 0. Thus the tangent spaces at (p, q) sum to the ambient space, so the intersection is transverse. (cid:3) ξ is given, then taking yi = −piηi/qi yields y ∈ kn The next result applies the lemma to an affine toric variety X. A fixed d × n integer ξ by matrix A = [a1 a2 · · · an−1 an] defines an action of the algebraic torus T = (k∗)d on kn t · ξ = (ta1ξ1, . . . , tanξn). The orbit Orb(A) of the point 1 = (1, . . . , 1) ∈ kn is the image of an algebraic map T → kn that sends t → t · 1. The closure of Orb(A) in kn is the affine toric variety XA = Var(IA) cut out by the toric ideal IA = hξu − ξv Au = Avi ⊆ k[ξ], of A in the polynomial ring k[ξ] = k[ξ1, . . . , ξn]. The T -action induces an A-grading on k[ξ] via deg(ξi) = ai, and the semigroup ring SA = k[ξ]/IA is A-graded [MS, Chapters 7 -- 8]. SYSTEMS OF PARAMETERS AND HOLONOMICITY OF A-HYPERGEOMETRIC SYSTEMS 3 For any face τ of the real cone R≥0A generated by the columns of A, write τ (cid:22) A and let 1τ ∈ {0, 1}n ⊂ kn be the vector with nonzero entry 1τ i = 1 precisely when A has a nonzero column ai ∈ τ . The variety XA decomposes as a finite disjoint union XA = Fτ (cid:22)A Orb(τ ) of orbits, where Orb(τ ) = T · 1τ . Each orbit has dimension dim Orb(τ ) = rank(Aτ ), where Aτ is the submatrix of A consisting of those columns lying in τ , and dim XA = rank(A). Theorem 1.2. The ring k[x, ξ]/(cid:0)IA + hAxξi(cid:1) has Krull dimension n. In particular, if A has rank d then the forms Axξ are part of a system of parameters for k[x] ⊗k SA. Proof. Let kτ ⊆ kn be the subspace consisting of vectors with 0 in coordinate i if ai /∈ τ , and let τ be its dimension. Since k[x, ξ]/IA = k[x] ⊗k SA has dimension n + rank(A) and the number of k-linearly independent generators of hAxξi is at most rank(A), the Krull dimension in question is at least n. Hence it suffices to prove that (cid:0)kn × Orb(τ )(cid:1) ∩ Var(Axξ) ⊆ kn × kτ has dimension at most n. Let xτ and ξτ denote the subsets corresponding to τ in the variable sets x and ξ, respectively. The projection of the intersection onto the subspace kτ × kτ has image contained in (cid:0)kτ × Orb(τ )(cid:1) ∩ Var(Aτ xτ ξτ ) ⊆ kτ × kτ . It therefore suffices to show that the dimension of this latter intersection is at most τ . By Lemma 1.1, the intersection is transverse in kτ × kτ . But the dimension of Orb(τ ) is the codimension of Var(Aτ xτ ξτ ) in kτ × kτ , which completes the proof. (cid:3) 2. Hypergeometric holonomicity In this section, the matrix A is a d × n integer matrix of full rank d. Let D = Chx, ∂ (cid:12)(cid:12) [∂i, xj] = δij and [xi, xj] = 0 = [∂i, ∂j]i denote the Weyl algebra over the complex numbers C, where x = x1, . . . , xn and ∂i corresponds to ∂ ∂xi . This is the ring of C-linear differential operators on C[x]. For β ∈ Cd, the A-hypergeometric system with parameter β is the left D-module MA(β) = D/D · (I ∂ A, {Ei − βi}d i=1), where I ∂ A = h∂u − ∂v Au = Avi ⊆ C[∂] is the toric ideal associated to A and Ei − βi = n X j=1 aijxj∂j − βi are Euler operators associated to A. The order filtration F filters D by order of differential operators. The symbol of an operator P is its image in(P ) ∈ grF D. Writing ξi = in(∂i), this means grF D is the commutative polynomial ring C[x, ξ]. The characteristic variety of a left D-module M is the variety in A2n of the associated graded ideal grF ann(M) of the annihilator of M. A nonzero D-module is holonomic if its characteristic variety has dimension n; this is equiva- lent to requiring that the dimension be at most n; see [Gab81] or [BGK+87, Theorem 1.12]. 4 CHRISTINE BERKESCH, STEPHEN GRIFFETH, AND EZRA MILLER The rank of a holonomic D-module M is the (always finite) dimension of C(x) ⊗C[x] M as a vector space over C(x); this number equals the dimension of the vector space of germs of analytic solutions of M at any nonsingular point in Cn [SST00, Theorem 1.4.9]. Viewing the A-hypergeometric system MA(β) as having a varying parameter β ∈ Cd, the rank of MA(β) is upper semicontinuous as a function of β [MMW05, Theorem 2.6]. This follows by viewing MA(β) as a holonomic family [MMW05, Definition 2.1] para- metrized by β ∈ Cd. By definition, this means not only that MA(β) is holonomic for each β, but also that it satisfies a coherence condition over Cd, namely: after replacing β with variables b = b1, . . . , bd, the module C(x) ⊗C[x] MA(b) is finitely generated over C(x)[b]. (The definition of holonomic family in [MMW05] allows sheaves of D-modules over arbitrary complex base schemes, but that generality is not needed here.) The derivation of the holonomic family property for MA(b) from the holonomicity of the A-hypergeometric system is more or less the same as in [MMW05, Theorem 7.5], which was phrased in the generality of Euler -- Koszul homology of toric modules. The brief deduction here isolates the steps necessary for A-hypergeometric systems; its brevity stems from the special status of affine semigroup rings among all toric modules [MMW05, Definition 4.5]. Theorem 2.1. The module MA(b) forms a holonomic family over Cd with coordinates b. In more detail, as a D[b]-module the parametric A-hypergeometric system MA(b) satisfies: 1. the fiber MA(β) = MA(b) ⊗C[b] C[b]/hb − βi is holonomic for all β; and 2. the module C(x) ⊗C[x] MA(b) is finitely generated over C(x)[b]. Proof. Since R = C[x, ξ]/hin(IA), Axξi surjects onto grF MA(β), it is enough to show that the ring R has dimension n. If MA(β) is standard Z-graded (equivalently, the rowspan of A over the rational numbers contains the row [1 1 · · · 1 1] of length n), then in(IA) = IA ⊆ C[ξ], and the result follows from Theorem 1.2. When MA(β) is not standard Z-graded, let A be the (d + 1) × (n + 1) matrix obtained by adding a row of 1's across the top of A and then adding as the leftmost column (1, 0, . . . , 0). If ξ0 denotes a new variable corresponding to the leftmost column of A, and ξ = {ξ0} ∪ ξ, then C[ξ]/ in(IA) ∼= C[ ξ]/hI A, ξ0i. In particular, C[x, ξ] hin(IA), Axξi ∼= C[x, ξ] hI A, ξ0, Axξi , where x = {x0} ∪ x. Since hI A, ξ0i is A-graded and A has a row [1 1 · · · 1 1], we have reduced to the case where MA(β) is Z-graded, completing part 1. With R as in part 1, the ring R[b] surjects onto grF MA(b), so it suffices for part 2 to show that R[b] becomes finitely generated over C(x)[b] upon inverting all nonzero polynomials in x. Since the ideal hin(IA), Axξi has no generators involving b variables, it suffices to show that R(x) itself has finite dimension over C(x). The desired result from the statement proved for part 1: any scheme of dimension n has finite degree over Cn x. (cid:3) SYSTEMS OF PARAMETERS AND HOLONOMICITY OF A-HYPERGEOMETRIC SYSTEMS 5 References [Ado94] Alan Adolphson, Hypergeometric functions and rings generated by monomials, Duke Math. J. 73 (1994), 269 -- 290. [BGK+87] A. Borel, P.-P. Grivel, B. Kaup, A. Haefliger, B. Malgrange, and F. Ehlers, Algebraic D- [Gab81] [GG86] [GGZ87] [GKZ90] [GKZ93] [GKZ94] modules, Perspectives in Mathematics 2, Academic Press, Inc., Boston, MA, 1987. Ofer Gabber, The integrability of the characteristic variety, Amer. J. Math. 103 (1981), no. 3, 445 -- 468. I. M. Gelfand and S. I. Gelfand, Generalized hypergeometric equations, (Russian) Dokl. Akad. Nauk SSSR 288 (1986), no. 2, 279 -- 283. English Translation: Soviet Math. Dokl. 33 (1986), no. 3, 643 -- 646. I. M. Gelfand, M. I. Graev, and A. V. Zelevinsky, Holonomic systems of equations and series of hypergeometric type, Dokl. Akad. Nauk SSSR 295 (1987), no. 1, 14 -- 19. I. M. Gelfand, M. Kapranov, and A. V. Zelevinsky, Generalized Euler integrals and A- hypergeometric functions, Adv. Math. 84 (1990), no. 2, 255 -- 271. I. M. Gelfand, A. V. Zelevinsky, and M. M. Kapranov, Hypergeometric functions and toric varieties, Funktsional. Anal. i Prilozhen. 23 (1989), no. 2, 12 -- 26. Correction in ibid, 27 (1993), no. 4, 91. I. M. Gelfand, M. Kapranov, and A. V. Zelevinsky, Discriminants, resultants and multidimen- sional determinants, Mathematics: Theory & Applications. Birkhauser Boston, Inc., Boston, MA, 1994. [Kap98] Mikhail Kapranov, Hypergeometric functions on reductive groups, Integrable systems and al- gebraic geometry (Kobe/Kyoto, 1997), World Sci. Publishing, River Edge, NJ, 1998, pp. 236 -- 281. [MMW05] Laura Felicia Matusevich, Ezra Miller, and Uli Walther, Homological methods for hypergeo- [MS] metric families, J. Amer. Math. Soc. 18 (2005), no. 4, 919 -- 941. Ezra Miller and Bernd Sturmfels, Combinatorial commutative algebra, Graduate Texts in Mathematics, 227. Springer-Verlag, New York, 2005. [SST00] Mutsumi Saito, Bernd Sturmfels, and Nobuki Takayama, Grobner Deformations of Hyperge- [SW08] ometric Differential Equations, Springer -- Verlag, Berlin, 2000. Mathias Schulze and Uli Walther, Irregularity of hypergeometric systems via slopes along coordinate subspaces, Duke Math. J. 142 (2008), no. 3, 465 -- 509. Department of Mathematics, Duke University, Box 90320, Durham, NC 27708 E-mail address: [email protected] Instituto de Mathem´atica y F´ısica, Universidad de Talca, Camino Lircay, Talca, Chile E-mail address: [email protected] Department of Mathematics, Duke University, Box 90320, Durham, NC 27708 E-mail address: [email protected]
1403.3990	2	1403	2016-07-13T19:51:22	Newton-Okounkov polyhedra for character varieties and configuration spaces	[ "math.AG" ]	We construct families of Newton-Okounkov bodies for the free group character varieties and configuration spaces of any connected reductive group. We use this construction to give a proof that these spaces are Cohen-Macaulay.	math.AG	math	NEWTON-OKOUNKOV POLYHEDRA FOR CHARACTER VARIETIES AND CONFIGURATION SPACES CHRISTOPHER MANON Abstract. We construct families of Newton-Okounkov bodies for the free group character varieties and configuration spaces of any connected reductive group. We use this construction to give a proof that these spaces are Cohen- Macaulay. Contents Introduction 1. 2. Background on representation theory 3. Background on valuations and Newton-Okounkov bodies 4. Branching filtrations 5. Valuations from the dual canonical basis 6. Proof of Theorems 1.1 and 1.2 7. Examples References 1 6 7 9 15 20 21 25 Keywords: Character Variety, Configuration Space, Newton-Okounkov Body 1. Introduction The Newton-Okounkov body construction was defined by Okounkov in [Oku97] to study Hilbert functions of spherical varieties. This construction has been reintro- duced in the recent papers of Lazarsfeld, Mustat¸a [LM09] and Kaveh, Khovanskii [KK12], and has become the centerpiece of several efforts to further integrate al- gebra, geometry, and combinatorics. A Newton-Okounkov body is a convex set Cv associated to a commutative domain A of finite Krull dimension by taking the convex hull of the image of a maximal rank valuation v : A → ZM ⊂ RM . Similarly, a compact, convex Newton-Okounkov body Cv(L) can be defined for a projective variety X and a choice of ample line bundle L, for both of these constructions see Section 3. As their name suggests, the Newton-Okounkov body is a generalization (see [Kav11b]) of the Newton polyhedron of an affine or projective toric variety. In keeping with the analogy to the toric case, the Newton-Okounkov body serves as a discrete model of a variety, and a useful tool to study combinatorial and algebraic properties of its coordinate ring. The image v(A) ⊂ ZM of a maximal rank valuation contains 0 and is closed under addition, making it into an affine semigroup. When v(A) is finitely generated the Newton-Okounkov body is polyhedral and there is flat degeneration Spec(A) ⇒ XCv , where XCv is the toric variety attached to Cv. In this way, Newton-Okounkov 1 2 CHRISTOPHER MANON bodies provide a mechanism to produce toric degenerations of varieties, these are a useful tool for studying algebraic presentations of A and studying the singularities of Spec(A) (see Theorem 1.3 below). In this case, Harada and Kaveh have linked Newton-Okounkov body to symplectic geometry by using the degeneration to define integrable systems in X when it is smooth and projective, [MH]. The Newton- Okounkov body is also useful as a combinatorial tool, as the set v(A) ⊂ Cv provides a polyhedral labelling of a basis of A which can be brought to bear when the underlying vector space of A has an enumerative meaning, indeed this was what motivated their introduction by Okounkov in [Oku97]. Newton-Okounkov bodies are one of several ways valuations are used to attach polyhedral geometry to structures in commutative algebra and algebraic geometry. In one of the first results of this type, Morgan and Shalen [MS84] reproduce a mys- terious piecewise-linear compactification of Teichmuller space defined by Thurston as a polyhedral complex of valuations on the coordinate ring of one of the spaces we study in this paper, a complex character variety. The character variety X (π, G) associated to a finitely generated group π and a connected reductive group G is the moduli space of representations ρ : π → G. When π is the fundamental group of a smooth manifold M , X (π, G) is the moduli space of flat, topological principal G bundles on M . For this reason, character varieties appear as natural moduli spaces in subjects which blend topology and geometry, such as Teichmuller theory and the theory of geometric structures [FG06], [Gol86], [Gol84]. We use combinatorial ele- ments from representation theory to build Newton-Okounkov bodies for X (Fg, G), where Fg is a free group. For a graph Γ, always assumed to be connected, we let E(Γ) and V (Γ) be the sets of edges and vertices, respectively. Recall that a leaf is a vertex which is connected to precisely one edge, and a spanning tree T ⊂ Γ is a subtree such that V (T ) = V (Γ). See Section 2 for the definition of the set R(w0) which appears below. Theorem 1.1. To the following information we associate a valuation vi,Γ on C[X (Fg, G)] with a convex, polyhedral Newton-Okounkov body Ci(Γ): (1) A trivalent graph Γ with no leaves and β1(Γ) = g. (2) Total orderings on the edges E(Γ) and vertices V (Γ) of Γ. (3) A spanning tree T ⊂ Γ. (4) An orientation on each edge in the set (cid:126)e = E(Γ) \ E(T ) (5) An assignment i : V (Γ) → R(w0) of reduced decomposition of the longest word w0 in the Weyl group of G to each vertex v ∈ V (Γ). The methods we use to prove Theorem 1.1 apply with little modification to another class of varieties, the configuration spaces P(cid:126)λ(G) of points on flag varieties of G. For what follows see Section 2 and the book of Fulton and Harris [FH91] for the relevant background in representation theory, and the books of Mumford, Fogarty and Kirwan [MFK94] and Dolgachev [Dol03] for the necessary background on Geometric Invariant Theory (GIT ). From now on (cid:126)λ = λ1, . . . , λn ∈ ∆ denotes POLYHEDRA FOR CHARACTER VARIETIES AND CONFIGURATION SPACES 3 a tuple of dominant weights of G with associated highest weight representations V (λ1), . . . , V (λn). A central problem of combinatorial representation theory is to count the dimension of the space of invariant tensors: (1) (V (λ1) ⊗ . . . ⊗ V (λn))G ⊂ V (λ1) ⊗ . . . ⊗ V (λn) Let P1, . . . , Pn be the parabolic subgroups which stabilize the highest weight vectors in the representations V (λ1), . . . , V (λn), respectively. Recall that the flag variety G/Pi has a G−linearized line bundle Lλi, with H 0(G/Pi,Lλi) equal to the dual repesentation V (λ∗ i ). These are the ingredients needed to define P(cid:126)λ(G) as the following projective GIT quotient: (2) P(cid:126)λ(G) = G\Lλ1 (cid:2)...(cid:2)Lλn G/Pi (cid:89) This construction turns the problem of counting the dimension of tensor product invariant spaces into the problem of finding the dimension of the space of global sections H 0(P(cid:126)λ(G),L((cid:126)λ)) of the induced line bundle on P(cid:126)λ(G). To find a solution to this problem, it suffices to find a rule to count the triple tensor product multiplicity spaces (V (λ) ⊗ V (η) ⊗ V (µ))G (see Section 5 ). In type A, this is accomplished by various combinatorial rules: (ordered by increasing generality) the Clebsh-Gordon rule for G = SL2(C), the Pieri rule for SLm(C) when the dominant weight λ is a multiple r1ω1 of the first fundamental weight, and the Littlewood-Richardson rule for general SLm(C) triple tensor product multiplicities. As a common feature, all of these rules can be realized by assigning to λ, η, µ the set of integer points in certain polytopes P (λ, η, µ). Families of polyhedral counting rules were discovered for general semisimple G by Berenstein and Zelevinksy in [BZ01] (their result applies readily to general connected, reductive G). Their proof utilizes a combination of positive geometry and the Lusztig's dual canonical basis. Our second main theorem realizes the polyhedra of Berenstein and Zelevinsky, along with their generalizations to n−fold tensor products, as Newton-Okounkov bodies for the line bundles L((cid:126)λ) on the P(cid:126)λ(G). For a tree T , we let L(T ) be the set of leaves, and we let V o(T ) = V (T )\ L(T ). Similarly, we let Eo(T ) ⊂ E(T ) denote the set of edges which are not connected to a leaf. Theorem 1.2. To the following information we associate a polytope Ci(T , (cid:126)λ∗), re- alized as the Newton-Okounkov body of a valuation vi,T on the projective coordinate ring C[P(cid:126)λ(G)] =(cid:76) m≥0 H 0(P(cid:126)λ(G),L((cid:126)λ)⊗m): (1) A trivalent tree T with an ordering on the set of leaves L(T ). (2) A total ordering on the non-leaf edges Eo(T ) and non-leaf vertices V o(T ). (3) An assignment i : V o(T ) → R(w0) of reduced decomposition of the longest word w0 in the Weyl group of G to each vertex v ∈ V o(T ). In particular, the integer points of Ci(T , (cid:126)λ∗) are in bijection with a basis of the invariant tensors (V (λ∗ 1) ⊗ . . . ⊗ V (λ∗ n))G = H 0(P(cid:126)λ(G),L((cid:126)λ)). 4 CHRISTOPHER MANON The Ci(T , (cid:126)λ) are cross-sections of an affine Newton-Okounkov body Ci(T ) for an affine configuration space Pn(G), defined in Section 4 (Example 4.6). This space acts as a "master" space for the configuration spaces, in that any P(cid:126)λ(G) is obtained from Pn(G) by a right T n GIT quotient. Using the same methods as in the proof of Theorem 1.1, we produce a T n invariant valuation vi,T on C[Pn(G)] with Newton-Okounkov body Ci(T ). As an application of Theorems 1.1 and 1.2, we show that a common structural property is shared by the two families of commutative algebras we consider. Recall that a local ring (R, m) is said to be Cohen-Macaulay (CM) when its dimension coincides with its depth, see Chapter 2 of [BH93]. An affine variety is then Cohen- Macaulay if all of its local rings are CM. Likewise, a projective variety with a chosen embedding is said to be Arithmetically Cohen-Macaulay if its projective coordinate ring is CM. Theorem 1.3. For G a connected, reductive group over C, the affine varieties X (Fg, G) and Pn(G) are Cohen-Macaulay. Moreover, the projective varieties P(cid:126)λ(G) are arithmetically Cohen-Macaulay. Theorem 1.3 follows from Proposition 6.3, which shows that the affine semi- groups obtained as the images of the coordinate rings C[X (Fg, G)], C[Pn(G)] and the projective coordinate ring C[P(cid:126)λ∗ (G)] under the valuations vi,T , vi,Γ are normal. Normal affine semigroup algebras are known to be Cohen-Macaulay by a result of Hochster, (see 6.3 of [BH93]), and this property is preserved under flat degenera- tions. 1.1. Methods. We construct the valuations vi,Γ by building increasing filtrations of the coordinate ring C[X (Fg, G)] by a lexicographically ordered free Abelian group in two steps, given in Sections 4 and 5. These filtrations are shown to be "strong" filtrations (see Section 3), which are naturally associated to valuations by Proposi- tion 3.2. In Section 4 we build a filtration inspired from one of the applications of character varieties to gauge theory. For a maximal compact subgroup K ⊂ G, so-called BF theory on an appropriately triangulated manifold M is quantized by combinatorial objects known as the spin diagrams of K (equivalently G), see [Bae00]. Definition 1.4. Let Γ be an oriented graph, a spin diagram with topology Γ is the following information: (1) An assignment λ : E(Γ) → ∆, of dominant weights to the edges of Γ. (2) An assignment of G−linear maps ρ to the vertices v ∈ V (Γ) which inter- e→v V (λ(e)) with the outgoing repre- twine the incoming representations(cid:78) sentations(cid:78) v→f V (λ(f )). The purpose of Section 4 is to show that for a fixed trivalent Γ with β1(Γ) = g, the spin diagrams with topology Γ define a filtration of C[X (Fg, G)]. We identify the associated graded algebra of this filtration, and note that it is not an affine semigroup algebra unless the semisimple part of G has a product of copies of SL2(C) as its universal cover. In order to enhance these filtrations, we must carefully choose a basis with amenable multiplication and combinatorial properties in the intertwiner spaces at POLYHEDRA FOR CHARACTER VARIETIES AND CONFIGURATION SPACES 5 • µ • ψ β α • η • •φ • λ Figure 1. A spin diagram on a 4−tree. each vertex v ∈ V (Γ). This is provided by the dual canonical basis constructed by Lusztig, [Lus90]. The dual canonical basis can be used to define a basis in each invariant space B(µ, λ, η) ⊂ (V (µ) ⊗ V (λ) ⊗ V (η))G, which are in turn identified with intertwiner spaces, [BZ96], [BZ01]. We study filtrations built from this basis in Section 5. For each choice i ∈ R(w0) of a reduced decomposition of the longest element of the Weyl group of G, there is a labelling of the dual canonical basis by tuples of non-negative integers b → (cid:126)t ∈ ZN called string parameters. In this way, the choice i assigns the elements of (V (µ)⊗ V (λ)⊗ V (η))G to integer points in a convex polytope Ci(µ, λ, η) studied in [BZ01]. We use the inequalities of these polytopes to define the polyhedra in Theorems 1.1 and 1.2. Our use of the dual canonical basis in this role follows previous work of Caldero [Cal02], and Alexeev, Brion [AB04] (see also [Kav11a]), who use a filtration on the string parameters of the dual canonical basis to define toric degenerations of schubert varieties and spherical varieties, respectively. We combine the filtrations from Section 4 with the string parameter filtrations of Section 5 to produce maximal rank valuations on C[X (Fg, G)] and C[Pn(G)] with Proposition 3.3 in Section 3, this proves Theorems 1.1 and 1.2. 1.2. Remarks. Lawton [Law10] and Sikora [Sik12] have given structure theorems for the coordinate rings C[X (Fg, G)], and Lawton, Florentino give descriptions of the topology [FL09] and the singular locus [FL12] of X (Fg, G) in certain cases. It would be interesting to relate the degenerations constructed here to a Grobner theory of their defining equations. Theorem 1.2 gives a construction of a basis for the tensor product invariant spaces (V (λ1) ⊗ . . . ⊗ V (λn))G which is labelled by the lattice points in a convex, rational polytope. Howe, Jackson, Lee, Tan, and Willenbring [HJL+09], [HTW05] use a SAGBI construction to achieve this for tensor product invariant spaces in the case G = SLm(C). The cone Ω3 resulting from their construction on triple tensor products is a slice of the cone of Gel'fand-Tsetlin patterns, and is linearly equivalent to Ci(3) for a particular choice of decomposition i. The algebraic structure of these cones is not very well understood outside the cases SLm(C), m = 2, 3, 4. We also point out that the space Pn(SL2(C)) is the affine cone of the Plucker embedding of the Grassmannian variety Gr2(Cn), and that the toric degenerations we construct in this case coincide with those found by Speyer and Sturmfels in [SS04]. Other enumeration problems in representation theory could plausibly be studied with the methods in this paper. Polyhedra which control Levi branching problems L ⊂ G have been defined by Berenstein and Zelevinsky in [BZ01]. These can x x ' ' 7 7 6 CHRISTOPHER MANON be adapted along the lines of the program used in Sections 4, 5, and realized as Newton-Okounkov bodies. The definition of Newton-Okounkov body we use is more general than the one in [LM09] and [KK12], where the valuation used to construct the Newton-Okounkov body comes from a flag of subspaces of the variety. It would be interesting to realize the tensor product polytope Ci(T , (cid:126)λ) as the Newton-Okounkov body attached to a flag F in a variety birational to P(cid:126)λ(G). This has been carried out in [Man14] for the valuations we construct here for the space X (Fg, SL2(C)). The work of Harada and Kaveh [MH] suggests that each Ci(T , (cid:126)λ) and Ci(Γ) should be the momentum image of an integrable system in P(cid:126)λ∗ (G) and X (Fg, G), would be interesting for the symplectic geometry of X (Fg, G) and P(cid:126)λ∗ (G). It would respectively. A construction of such an integrable system for each of these polyhedra also be interesting to see geometric relationships between the integrable systems associated to different valuations given by our construction. Partial results in this direction appear in [HMM11] and [Man14] for G = SL2(C). The isomorphism ΦT ,(cid:126)e in Proposition 4.13 can be precomposed with the auto- morphism of C[X (Fg, G)] induced by any outer automorphism of the free group α ∈ Out(Fg) to produce a distinct set of maximal rank valuations. In [Man14] we explore this large family of Newton-Okounkov bodies in the case G = SL2(C), and its connections with the outer space construction of Culler and Vogtmann [CV86]. Presumably similar spaces could be defined for all reductive G using the results of this paper. Finally, we remark that Theorem 1.2 essentially appears in the unplublished notes [Man10], along with other remarks on the use of valuations in the study of branching problems. 1.3. Acknowledgements. We thank Kiumars Kaveh for numerous helpful conver- sations about Newton-Okounkov bodies. We also thank Sean Lawton and Adam Sikora for useful conversations about character varieties. We thank the reviewer for useful suggestions to improve the exposition. 2. Background on representation theory We introduce the necessary background on reductive groups and their Lie al- gebras. We refer the reader to the books of Fulton and Harris [FH91], Dolgachev [Dol03], and Grosshans [Gro97] for structural properties of reductive groups, as well as their actions on commutative algebras. a sum of simple Lie algebras gss =(cid:76) gi. We fix a system of Chevallay generators We let G be a connected, complex reductive group with center Z, maximal torus T and root system R. The Lie algebras of Z and T are denoted by z and t, respectively. We choose a set of positive roots R+ ⊂ R ⊂ t∗ with simple roots S = {α1, . . . , αr}. The Lie algebra g of G splits as a direct sum g = z ⊕ gss with z the Lie algebra of Z and gss the semisimple part of g. The Lie algebra gss further decomposes into H1, . . . , Hr, f1, . . . , fr, e1, . . . , er ∈ gss, where r is the rank of gss. The elements Hi are the simple coroots, they form a basis of a Cartan subalgebra h ⊂ gss, and together with z they span the Lie algebra t. The elements fi and ei are called lower- ing and raising operators, and they generate nilpotent Lie subalgebras u−, u+ ⊂ g, respectively. The set {fi, Hi, ei} generates a copy of of sl2(C) ⊂ g, and for any POLYHEDRA FOR CHARACTER VARIETIES AND CONFIGURATION SPACES 7 H ∈ t, [H, fi] = −αi(H)fi, [H, ei] = αi(H)ei. All of this information is captured in the Cartan matrix A = [aij], with aij = αj(Hi). The above information defines a distinguished so-called triangular decomposition g = u− ⊕ t ⊕ u+. The spaces u−, u+ are the Lie algebras of maximal unipotent subgroups U−, U+ ⊂ G, which together with T define a dense, open subset U− × T × U+ → U−T U+ ⊂ G,. When it is important to single out one of these groups we use the notation U±, otherwise we will refer to a general maximal unipotent subgroup U . We let Λ ⊂ t∗ be the character group of T . The weights λ ∈ Λ which satisfy λ(Hi) ≥ 0 are called dominant, and the non-negative real span of the dominant weights is called a Weyl chamber ∆ ⊂ t∗. For Gss the simply-connected semisimple group associated to gss, we can present G as [Gss × L]/K for L a torus and K a finite central subgroup of G × L. In particular, any dominant weight λ ∈ ∆ is a sum ¯λ + τ of a dominant weight ¯λ ∈ ∆ss and a character τ of L such that ¯λ + τ is a trivial character on K. We choose one such presentation for G, and a corresponding isomorphism of the Lie algebra of L with z. In this way, the Weyl chamber can be of gss. The vectors ωi ∈ ∆ss which satisfy λj(Hi) = δij are called the fundamental weights, these are weights of G when G is simply-connected. There is a natural partial ordering < on the set of dominant weights, where λ > η when λ − η can be expressed as a non-negative sum of positive roots. identified with a product z∗×(cid:81) ∆i, where ∆i is a Weyl chamber of a simple part gi The Weyl group N (T )/T = W of G is generated by elements s1, . . . , sr called simple reflections, these are in bijection with the simple roots. The simple reflections define a length function on the elements of W. The unique longest element w0 has length equal to the number N of positive roots of g. We let R(w0) denote the set of reduced decompositions i = {i1, . . . , iN}, of w0 = si1 . . . siN into simple reflections. The category Rep(G) of finite dimensional Representations of G is semisimple, and the set of irreducible representations is in bijection with the dominant weights λ ∈ ∆. We let V (λ) denote the irreducible representation associated to λ ∈ ∆. The dual representation V (λ)∗ of an irreducible has weight λ∗ = −w0(λ). Recall that V (λ) decomposes as a T−representation into weight spaces Vη(λ) ⊂ V (λ), that η < λ for all η which appear in this decomposition, and that the weight space Vλ(λ) is always 1−dimensional. We let bλ be a highest weight vector spanning Vλ(λ). 3. Background on valuations and Newton-Okounkov bodies We introduce basic concepts and operations on valuations, for more on this topic, see Section 1 of [KK12] and Chapter 6 of the book by Zariski and Samuel [ZS76]. Let A be a commutative domain over C of finite Krull dimension dim(A), and let G be an Abelian group with a total ordering ≺. A valuation v : A \ {0} → G is a function which satisfies v(f g) = v(f ) + v(g) and v(f + g) (cid:22) max{v(f ), v(g)} for all f, g ∈ A. Additionally, we assume that v(C) is equal to the neutral element 0 ∈ G for all C ∈ C \ {0}, this means that v is a lift of the trivial valutaion on C to A. For our purposes, the group G is always taken to be a free Abelian group ZM in some guise, and ≺ is the lexicographic ordering. Recall that this means that (a1, . . . , aM ) ≺ (b1, . . . , bM ) if and only if the first nonzero bi − ai is positive. The properties of a valuation imply that the image v(A) ⊂ ZM of v contains 0 and is closed under addition, making it an affine semigroup. The rank of v is the dimension 8 CHRISTOPHER MANON of vector space spanned by v(A) in Qm, or equivalently the rank of the free Abelian subgroup generated by v(A) ⊂ ZM . The Krull dimension of the affine semigroup algebra C[v(A)] is then equal to the rank of v, see Chapter 6 of [BH93]. Definition 3.1. (Newton-Okounkov Body: affine case) With A, v, ZM ,≺ as above and rank(v) = dim(A), the Newton-Okounkov body Cv ⊂ Rm is the positive real cone R≥0v(A) spanned by the image v(A) ⊂ ZM ⊂ RM . It follows that dim(Cv) as a subspace of RM is equal to the dimension of the real space Rv(A), which in turn equals dim(A) = rank(v). Let ∪w∈ZM F(cid:22)w = A be an increasing algebra filtration of A by C vector spaces. We say the spaces F(cid:22)w define a strong filtration if C ⊂ F(cid:22)0, C (cid:54)⊂ F(cid:22)w for any w ≺ 0, and the associated graded algebra grF (A) =(cid:76) w∈ZM F(cid:22)w/F≺w is a domain. Proposition 3.2. Let A, ZM ,≺ and F be as above. The information of a strong increasing filtration A = ∪w∈ZM F≤w is equivalent to a valuation vF : A → ZM . Proof. Starting with a filtration F , define vF by vF (f ) = min{wf ∈ F≤w}. For a w = {fv(f ) ≤ w}. The property v(f g) = v(f )+v(g) valuation v define F v implies that F v is a strong filtration. Similarly, F being a strong algebra filtration implies that vF (f g) = vF (f )+vF (g). We leave it to the reader to check the rest. (cid:3) w ⊂ A by F v The following proposition allows us to construct valuations on A in steps. Proposition 3.3. Let A be as above, with F a strong filtration on A by (ZM , <1), and let G be a strong filtration on grF (A) by (ZL, <2) which is compatible with the induced grading. There is a strong filtration F ◦ G on A by (ZM +L, <1 ◦ <2), where <1 ◦ <2 is the composite order built lexicographically by first ordering by <1 and breaking ties with <2 . This filtration has associated graded algebra grG(grF (A)). Proof. Each space F≤w/F<w ⊂ grF (A) has a filtration . . . ⊂ Gw,u ⊂ . . . . We pull the spaces Gw,u back to a filtration F ◦ Gw,u ⊂ Fw. By construction, each space F ◦ Gw,u contains F<w, this implies that F ◦ Gw(cid:48),u(cid:48) ⊂ F ◦ Gw,u if w(cid:48) < w. If w(cid:48) = w, then u(cid:48) < u and F ◦Gw(cid:48),u(cid:48) ⊂ F ◦Gw,u by construction. It is straightforward to check the identity grF◦G = grG(grF (A)), which proves the strong filtration property. (cid:3) Example 3.4. For a polynomial ring C[x1, . . . , xM ], we can define maximal rank β∈ZM Cβ(cid:126)xβ to the multidegree α of its highest term under the lexicographic ordering. The value semigroup d(C[x1, . . . , xn]) of this valuation is then the set of non-negative mul- tidegrees ZM≥0 ⊂ ZM , and the Newton-Okounkov body Cd is the positive orthant RM≥0 ⊂ RM . valuation d : C[x1, . . . , xM ] → ZM by sending a polynomial p((cid:126)x) =(cid:80) In the presence of a rational action on A by a reductive group G, we say a strong filtration F is G−stable if each space F(cid:22)w is likewise a G−representation. It is straightforward to verify that this is the case if and only if the associated valuation vF is invariant with respect to the G−action. In this case, there is a strong filtration of the algebra of invariants AG by the invariant spaces F G(cid:22)w = F(cid:22)w ∩ AG, and there is a natural induced rational G−action on the associated graded algebra grF (A). Proposition 3.5. Let F be a G−stable filtration on A, and let F also denote the induced filtration on the algebra of invariants AG ⊂ A. We have the following isomorphism: POLYHEDRA FOR CHARACTER VARIETIES AND CONFIGURATION SPACES 9 (3) grF (AG) ∼= grF (A)G. Proof. This is a consequence of the reductivity of G, as it ensures that the following sequence remains exact after taking G−invariants: (4) Let A =(cid:76) 0 → F≺w → F(cid:22)w → F(cid:22)w/F≺w → 0. (cid:3) if f ∈ AL then v(f ) = (. . . , L, . . .). L≥0 AL be a graded domain of finite Krull dimension with a maximal rank valuation v : A → ZM . Suppose further that one component of the functionv returns the grading, i.e. In this case the Newton-Okounkov body Cv carries a natural projection π : Cv → R≥0 and is a cone over the fiber π−1(1). Definition 3.6. With A, v as above, we let ¯Cv = π−1(1), this is a compact, convex set. When A is the projective coordinate ring of a projective variety X with respect to an ample line bundle L, we denote this set ¯Cv(L), and refer to it as the Newton- Okounkov body of L. 4. Branching filtrations We make use of the ordering on dominant weights to construct filtrations of the coordinate rings of character varieties and configuration spaces. These filtrations are not fine enough to give affine semigroup associated graded algebras, however this construction reduces the problem to constructing a T 3−stable filtration on C[P3(G)]. Spin diagrams for the group G emerge from this construction as labels for the graded components of the associated graded algebras we construct. We also give an alternative GIT construction of the character variety X (Fg, G) which makes the connection with spin diagrams more transparent. 4.1. Horospherical contraction and the algebra C[G]. We briefly review the theory of horospherical contraction, due to Popov [Pop86]. We choose a maximal torus T ⊂ G, with triangular decomposition U−T U+ ⊂ G, and Weyl chamber ∆. The coordinate ring C[G] has the following isotypical decomposition as a G × G representation (for proof see e.g. Theorem 12.9 of [Gro97]): (cid:77) λ∈∆ (5) C[G] = V (λ) ⊗ V (λ∗). (cid:76) subspaces F≤η =(cid:76) Horospherical contraction relates G to the affine right GIT quotient G/U of G by any maximal unipotent U ⊂ G. This scheme has coordinate ring C[G/U ] = λ∈∆ V (λ) ⊗ Cbλ∗ ⊂ C[G] with multiplication computed by dualizing the map C : V (λ + η) → V (λ)⊗ V (η) which sends the highest weight vector bλ+η to bλ ⊗ bη. The highest weight ordering induces a G × G-stable filtration on C[G] by the γ≤η V (γ) ⊗ V (γ∗). The next proposition gives the associated graded algebra of this filtration. Proposition 4.1. (Horospherical contraction) The dominant weight filtration on C[G] induced by ∆ has associated graded algebra isomorphic to C[T\(G/U+ × 10 CHRISTOPHER MANON U−\G)], where the T−action is on the right of G/U+ and the left of U−\G. In particular, this T−action has isotypical spaces V (λ) ⊗ V (λ∗) ⊂ C[G/U+ × U−\G]. (cid:3) Proof. This follows from Chapter 3, Section 15 of [Gro97]. As C[T\(G/U+ × U−\G)] is a domain, any prolongation of the partial ordering on ∆ to a total ordering which is compatible with addition of weights defines a η≤λ V (η) ⊗ V (η∗). There are strong filtration ∪λ∈∆F≤λ = C[G], where F≤λ = (cid:76) many such prolongations, we define one below. Definition 4.2. Let G be a simple complex group, with Weyl chamber ∆, and simple coweights H1, . . . , Hr. The total order < is defined by lexicographically organizing the orderings defined by λ(Hi) ∈ Z. Recall that we have chosen a product decomposition ∆ = z∗ ×(cid:81) ∆i. We can define < on ∆ by ordering the gi and using the induced lexicographic organization of the orderings <i. We then break ties with any lexicographic ordering on a basis of z∗. The following is straightforward. Lemma 4.3. The total order < respects addition of weights, and refines the partial dominant weight ordering <. Example 4.4. We consider the ordering < on the Weyl chamber ∆ of GL3(C). A dominant weight λ ∈ ∆ is typically represented as a tuple (λ1, λ2, λ3) satisfying λ1 ≥ λ2 ≥ λ3, and the fundamental coroots are the vectors H1 = (1,−1, 0) and H2 = (0, 1,−1). A weight is seperated into its semisimple part (a dominant weight of SL3(C)) and its central part by writing (λ1, λ2, λ3) = (λ1−λ3, λ2−λ3, 0)+λ3(1, 1, 1). Consider the weights (5, 3, 2), (4, 2, 1). We have H1(5, 3, 2) = 2 = H1(4, 2, 1) and H2(5, 3, 2) = 1 = H2(4, 2, 1), however (5, 3, 2) = (3, 1, 0) + 2(1, 1, 1) and (4, 2, 1) = (3, 1, 0) + (1, 1, 1). This implies that (5, 3, 2)>(4, 2, 1) in the ordering. Remark 4.5. For an irreducible representation V (λ), the orbit G[bλ] ⊂ P(V (λ)) of the highest weight line is a flag variety G/P . The embedding G/P → P(V (λ)) endows G/P with a line bundle Lλ with global section space H 0(G/P,Lλ) = V (λ∗). This linearization can be constructed using an affine GIT quotient of G/U by T with character λ. (cid:77) m≥0 (6) P roj( H 0(G/P,L⊗m λ ) = [G/U ]/λT 4.2. Branching algebras. We apply horospherical contraction to obtain filtra- tions on the coordinate rings of a class of spaces B(φ) called branching varieties. There is one such variety for each map φ : H → G of complex, connected reduc- tive groups. The space Pn(G) is recovered as the branching variety B(δn), where δn : G → Gn−1 is the diagonal embedding. We choose maximal unipotent sub- groups UH ⊂ H, UG ⊂ G, and Weyl chambers ∆H , ∆G, and define B(φ) as the following affine GIT quotient. (7) Here H acts on H/UH × G/UG diagonally on the left, where the action on G/UG is defined through φ. The coordinate ring of B(φ) is graded by the multiplicity spaces B(φ) = H\[H/UH × G/UG] POLYHEDRA FOR CHARACTER VARIETIES AND CONFIGURATION SPACES 11 W (µ, λ) of H irreducible representations in the irreducible representations of G, as branched over the map φ. (8) (9) V (λ) = W (µ, λ) ⊗ V (µ) C[B(φ)] = W (µ, λ) µ,λ∈∆H×∆G (cid:77) µ∈∆H (cid:77) (cid:77) Example 4.6. Let δn : G → Gn−1 be the diagonal map g → (g, . . . , g). The branching variety B(δn) is then a left diagonal G quotient of G/U n, we also knows this space as Pn(G). The coordinate ring C[Pn(G)] is the following direct sum of invariant spaces. (10) C[Pn(G)] = (V (λ1) ⊗ . . . ⊗ V (λn))G (cid:126)λ∈∆n The space Pn(G) carries a residual right hand side T n action. By Remark 4.5, the affine GIT quotient Pn(G)/(cid:126)λT n defined by a tuple of dominant weights (λ1, . . . , λn) = (cid:126)λ can be identified with the configuration space P(cid:126)λ(G). To see this, note that the invariant spaces (V (mλ1)⊗ . . .⊗ V (mλn))G corresponding to the mul- tiples m(cid:126)λ of the character defined by (cid:126)λ are precisely the G invariants in the global (cid:2). . .(cid:2)Lλn on G/P1×. . .×G/Pn. section spaces of the powers of the line bundle Lλ1 The branching algebras C[B(φ)] come with special filtrations defined by diagrams in the category of connected reductive groups. We let φ = π ◦ ψ be a factorization of φ in this category. ψ−−−−→ K π−−−−→ G H The map ψ defines an action of H on the reductive group K, and π defines an action of K on G. Using these actions, we can identify B(φ) with the following GIT quotient. (11) B(φ) = H × K\[H/UH × K × G/UG] φ = π ◦ ψ. The direct sum decomposition C[K] =(cid:76) Here H acts diagonally on the left of H/UH and K, and K acts on the right of K and on the left side of G/UG. The space K\K × G/UG is isomorphic to G/UG, and likewise the resulting action of H on H/UG × G/UG is induced through V (η) ⊗ V (η∗) induces a decomposition of C[B(φ)]. η∈∆K (12) C[B(φ)] = W (µ, η) ⊗ W (η, λ) λ,η,µ∈∆G,∆K ,∆H This defines a TH × TG-stable filtration F ψ,π of C[B(φ)] by the dominant weights η ∈ ∆K. (13) F ψ,π≤η = W (λ, γ) ⊗ W (γ, µ) (cid:77) (cid:77) λ,γ≤η,µ 12 CHRISTOPHER MANON Proposition 4.7. The associated graded algebra of F ψ,π is the affine GIT quotient C[B(π) × B(ψ)/TK], where TK acts on C[B(φ)] ⊗ C[B(ψ)] with isotypical spaces W (λ, γ) ⊗ W (γ, µ). Proof. The filtration F ψ,π is induced from the horospherical filtration on C[K]. The K × K stability of horospherical contraction and the remarks above imply that the associated graded algebra is the domain C[B(π) × B(ψ)]TK . (cid:3) Notice that the associated graded algebra C[B(π)×B(ψ)]TK has a residual algebraic action of TK. We finish this subsection by applying this construction to the diagonal map δn : G → Gn−1. Let T be a tree with two internal vertices and n leaves labelled 0, . . . , n − 1. Let k + 1 be the number of edges incident on the vertex connected to the 0 leaf, and m be the number of leaves connected to the other internal vertex. This structure defines a factorization δn = Ids × δm × Idt ◦ δk : G → Gn−1, where s + t = k − 1 (see Figure 2 for an example). Proposition 4.7 implies there is a filtration F T on C[Pn(G)], with associated graded algebra C[Pm(G) × Pk(G)]T . Example 4.8. The factorization of δ4 below is represented by the tree in Figure 2. δ3−−−−→ (G × G) G δ3×Id−−−−→ ((G × G) × G) • G • δ3 G G • G • G •δ3 • Figure 2. A directed tree of diagonal maps. Given a trivalent tree T with n leaves, and an ordering on Eo(T ), we iterate this construction to obtain a filtration F T on C[Pn(G)]. Proposition 4.9. For a trivalent tree T with n leaves, and an ordering on Eo(T ) there is a T n-invariant filtration on C[Pn(G)] with associated graded algebra the coordinate ring of PT (G) = [(cid:81) v∈V (T ) P3(G)]/T E(T ). Proof. An ordering E(T ) = {e1, . . . , en−3} induces a length n−3 chain of collapsing maps on trees, πi : Ti−1 → Ti where T0 = T , and Ti is obtained from Ti−1 via πi by collapsing the edge ei. The tree Tn−3 has a single internal vertex, and Tn−4 has a single internal edge. By the previous construction there is a T n-stable filtration on C[Pn(G)] with associated graded algebra PTn−4(G) = [Pv(u)(G) × Pv(w)(G)]/T , where v(u) and v(w) are the valences of the two internal vertices u, w ∈ V (Tn−4). The map πn−5 collapses the edge en−2 to either u or w, yielding a corresponding filtration on C[Pv(u)(G)] or C[Pv(w)(G)]. This filtration is invariant with respect to T above, and so induces a filtration on C[Pv(u)(G) × Pv(w)(G)]/T . We can now apply Proposition 3.3 to obtain a filtration on C[Pn(G)]. Continuing this way, we (cid:3) obtain the proposition. O O 8 8 g g O O w w POLYHEDRA FOR CHARACTER VARIETIES AND CONFIGURATION SPACES 13 4.3. Character varieties and the master configuration space. We show that a similar family of filtrations exist for the character variety X (Fg, G). This variety is constructed as the following GIT quotient. (14) X (Fg, G) = Gg/adG The ad subscript indicates the conjugation action of G on the product g ◦ad (x1, . . . , xn) = (gx1g−1, . . . , gxng−1). The coordinate ring C[X (Fg, G)] is therefore the algebra of conjugation invariants in C[Gg]. By Proposition 4.1, the horospherical contraction of C[G] to C[T\(G/U+ × U−\G)] is G × G invariant, therefore we may define a filtration on C[X (Fg, G)] by placing this G × G stable filtration on the coordinate ring of each part G of the product Gg. Proposition 4.10. There is a filtration on C[X (Fg, G)] with associated graded ring equal to the coordinate ring of [T\(G/U+ × U−\G)]g/adG = P2g(G)/T g. Here the invariants of the torus T g are the tensor products (V (λ1) ⊗ . . . ⊗ V (λ2g))G ⊂ C[P2g(G)], where λ∗ 2k−1 = λ2k. Now that we have related the character variety X (Fg, G) to the master configu- ration space P2g(G), we may use the valuations constructed in Proposition 4.9. Proposition 4.11. For every choice of a trivalent graph Γ, spanning tree T ⊂ Γ, an ordering on E(Γ) and an orientation on E(Γ) \ E(T ) = {e1, . . . , eg}, there is a filtration on C[X (Fg, G)] with associated graded ring the coordinate ring of PΓ(G) = [(cid:81) v∈V (Γ) P3(G)]/T E(Γ). Proof. We identify the ordered, oriented edges e1, . . . , eg with the components of Gg, where the orientation distinguishes the left and right hand side G actions on each component. The filtration above then yields an associated graded algebra C[P2g(G)/T g]. We split each edge ei in two: f2i−1, f2i, and build the trivalent tree T (cid:48) using the topology of the spanning tree T , see Figure 3. By Proposition 4.9, this defines a filtration on C[P2g(G)] with associated graded algebra C[PT (cid:48)(G)] = v∈V (Γ) P3(G)]/T E(T (cid:48)). By the T 2g stability of this filtration and Propositions 3.3 and 3.5, we may induce a filtration on C[X (Fg, G)] with associated graded algebra (cid:3) the quotient PΓ(G) = PT (cid:48)(G)/T g. [(cid:81) Figure 3. Splitting the complement of a spanning tree. eeee1324f5f1f2f3f4f7f8f614 CHRISTOPHER MANON Example 4.12. In the case G = SL2(C), the spaces W (r1, r2, r3), ri ∈ Z≥0 ⊂ R≥0 = ∆ are multiplicity-free. The dimension W (r1, r2, r3) is non-zero precisely when r1, r2, r3 satisfy two conditions: r1 + r2 + r3 ∈ 2Z, and r1− r2 ≤ r3 ≤ r1 + r2. This second condition is equivalent to the ri being the sidelengths of a triangle. The affine semigroup BZ3(SL2(C)) ⊂ Z3≥0 (see Section 7 for an explanation of this no- tation) of the (r1, r2, r3) satisfying these conditions is generated by (1, 1, 0), (1, 0, 1) In this case T = C∗, and the action of T 3 = (C∗)3 defines the and (0, 1, 1). structure of an affine 3−dimensional toric variety on P3(SL2(C)), in particular the (C∗)E(Γ) quotient in Proposition 4.11, we identify PΓ(SL2(C)) with an affine toric variety. The associated affine semigroup BZΓ(SL2(C)) is the set of weightings w : E(Γ) → Z≥0 such that for any e, f, g ∈ E(Γ) sharing a common vertex v ∈ V (Γ) we must have w(e) + w(f ) + w(g) ∈ 2Z and w(e) − w(f ) ≤ w(g) ≤ w(e) + w(f ). this space is naturally identified with the affine space (cid:86)2(C3). By working through 4.4. Graph construction of character varieties. We present an alternative construction of the variety X (Fg, G), which motivates the graph filtrations of Propo- sition 4.11. A variation on this construction has also been discovered by Florentino and Lawton in [FL13]. We fix a trivalent graph Γ with no leaves, and consider the forest Γ obtained by splitting each edge in E(Γ). Figure 4. The directed forest associated to a directed graph. (cid:81) We associate a copy of the GIT quotient M3(G) = G\G3 to each connected component of Γ, notice that this space has a residual right G3 action. The product v∈V (Γ) M3(G) then has an action by E(Γ) copies of G, where the component corresponding to e ∈ E(Γ) acts on the two components associated to the pair of edges in Γ which split e. We define MΓ(G) to be the GIT quotient by this action. (cid:89) v∈V (Γ) (15) MΓ(G) = [ M3(G)]/GE(Γ) Proposition 4.13. For each choice of a spanning tree T ⊂ Γ, an ordering on the edges (cid:126)e = E(Γ) \ E(T ), and an orientation of each edge in E(Γ), there is an isomorphism ΦT ,(cid:126)e : MΓ(G) → X (Fg, G). Proof. We split the edges ei ∈ (cid:126)e ⊂ E(Γ) into pairs e2i−1, e2i, ordered using (cid:81) the orientation on ei, this gives a trivalent tree T (cid:48). We construct MT (cid:48)(G) = v∈V (Γ) M3(G)/GEo(T (cid:48)), and note that MT (cid:48)(G)/G(cid:126)e = MΓ(G). We will prove that POLYHEDRA FOR CHARACTER VARIETIES AND CONFIGURATION SPACES 15 MT (cid:48)(G) ∼= G\G2g in Lemma 4.14. We use the isomorphism G2i−1 × G2i/G ∼= G, (g2i−1, g2i) → g2i−1g−1 2i . This intertwines the left G action on G2g with the adjoint (cid:3) action on Gg. Note that the definition of MΓ(G) is sensible for any graph, regardless of the valence of the vertices. We consider MT (G) for possibly non-trivalent trees in the following lemma. Lemma 4.14. For any tree T with n leaves, each orientation on the edges of T gives an isomorphism to the left quotient MT (G) ∼= G\Gn. Proof. It suffices to treat the case where T has one internal edge e, as this calcula- tion can then be iterated to show the result by induction. Let ∂(e) = {u, w}, with the orientation pointing from u to v. We view MT (G) as GV (u) × GV (w) with a left action of G × G and a right action of G on two components Ge,u ⊂ GV (u), Ge,w ⊂ GV (w). We use the map GV (u) × GV (w) → GV (u)+V (w)−2 given by the following. (16) (g1, . . . , gV (u)−1, ge,u) × (ge,w, gV (u)+2, . . . , gV (w)) → (g1, . . . , gV (u)−1, ge,wg−1 (g1, . . . , gV (u)−1, ge,ug−1 e,u, gV (u)+2, . . . , gV (w)) → e,wgV (u)+2, . . . ge,ug−1 e,wgV (w)) This is a G × G × G, where the first component acts diagonally on the left of GV (u)+V (w)−2, and the second and third components act trivially. Quotienting everything by G × G × G then yields the isomorphism. (cid:3) We may also view MΓ(G) as the following quotient. M3(G)/GE(Γ) = GV (Γ)\ (cid:89) (17) (cid:89) = GV (Γ)\ (cid:89) horospherical degeneration GV (Γ)\(cid:81) v∈V (Γ) e∈E(Γ) G3/GE(Γ) v∈V (Γ) [(G × G)/G] = GV (Γ)\GE(Γ). We can now recover Proposition 4.11 by replacing the rightmost term with the e∈E(Γ)[G/U− × U+\G]/T = PΓ(G). 5. Valuations from the dual canonical basis In Propositions 4.9 and 4.11, each graph Γ (respectfully tree T ) with compatible information defines the following direct sum decompositions on the coordinate rings C[X (Fg, G)] and C[Pn(G)]. Here W (λ, η, µ) denotes the invariant vectors in V (λ)⊗ V (η) ⊗ V (µ). C[X (Fg, G)] = C[Pn(G)] = (cid:77) (cid:77) (cid:79) (cid:79) λ:E(Γ)→∆ v∈V (Γ) λ:E(T )→∆ v∈V (T ) (18) (19) [W (λ(v, i), λ(v, j), λ(v, k))] [W (λ(v, i), λ(v, j), λ(v, k))] 16 CHRISTOPHER MANON Here (v, i) denotes a vertex v with incident edge i. In this section we show how to obtain finer combinatorial pictures of C[X (Fg, G)] and C[Pn(G)] by structuring the intertwiner spaces W (λ, η, µ) using the dual canonical basis. 5.1. String parameters and polytopes for tensor product multiplicities. We recall the construction of polyhedral cones Ci(3) which control tensor product multiplicites for a reductive group G. We take G to be semisimple, but we will later remove this restriction. In [Lus90], Lusztig constructs a basis B of the subalgebra Uq(u+) of the quantized universal enveloping algebra Uq(g). Specialization at q = 1 yields the canonical basis for each irreducible representation V (λ) ⊂ U(u+). The dual pairing between U(u+) and C[U+] induces a dual basis B(λ∗) ⊂ V (λ∗) ⊂ (cid:96) C[U+]. A basis B ⊂ C[U−\G] can then be constructed by taking the union B = λ B(λ) × {λ}. We fix a reduced decomposition i ∈ R(w0). The entries of i correspond to to raising operators ei1, . . . , eiN ∈ u+, these act as nilpotent derivations on C[U+] from the differentiation of the left and the right actions of U+ on itself. We use the notation e◦(cid:96) f and e◦r f to distinguish between the left and right actions of such an operator, and repeated applications are denoted with a superscript e ◦ . . . ◦ e = en. For the following see [Cal02] and [Kav11a]. Definition 5.1. For f ∈ C[U+], the string parameters wi(f ) = (t1, . . . , tN ) ∈ ZN ◦(cid:96) f = 0}. The k−th entry are defined inductively, starting with t1 = min{tet+1 ◦(cid:96) f = 0}. This defines the is then defined as tk = min{tet+1 ◦(cid:96) . . . ◦(cid:96) et1 string parametrization wi : C[U+] \ {0} → ZN associated to i ∈ R(w0). Lemma 5.2. The function wi is a valuation on C[U+] when the string parameters are lex ordered first to last. Proof. It is straightforward to show wi(C) = (0, . . . , 0) for any C ∈ C \ {0}. For f, g ∈ C[U+] let eik be the first raising operator for which the string parameters wi(f ), wi(g) differ, with tk(f ) > tk(g). Then by definition, wi(f + g) = wi(f ). If no such k exists, then wi(f + g) ≤ wi(f ) = wi(g). Applying eM to f g yields the following. ◦(cid:96) etk−1 ik−1 ik i1 i1 i1 (cid:88) (cid:18)M (cid:19) p p+q=M (20) eM i1 (f g) = ep i1 (f )eq i1 (g) If M > t1(f ) + t1(g), all terms in this sum must vanish. If M ≤ t1(f ) + t1(g), then the fact that C[U+] is a domain implies that this sum is non-zero, as this is the case for M = t1(f ) + t1(g), where the sum is multiple of et1(f ) (g). We may repeat this calculation with ei2 on this term. By induction this yields (cid:3) wi(f g) = wi(f ) + wi(g). We obtain a valuation vi on the coordinate ring C[U−\G] ⊂ C[T × U+], with image in ZN × ∆ by breaking ties with the < ordering on ∆. Berenstein and Zelevinsky, [BZ96] and Alexeev and Brion, [AB04] give inequalities for the image of this valuation using devices derived from the fundamental representations V (ωi) of the Lie algebra g of the Langlands dual group G, called i−trails. An i−trail from a weight γ to a weight η in the weight polytope of a representation V is a sequence of weights (γ, γ1, . . . , γ(cid:96)−1, η), such that consecutive differences of weights are integer (f )et1(g) i1 i1 POLYHEDRA FOR CHARACTER VARIETIES AND CONFIGURATION SPACES 17 i(cid:96) ◦ . . . ◦ ec(cid:96) multiples of simple roots from i, γi−γi+1 = ckαik , and the application of the raising : Vη → Vγ is non-zero. For any i−trail π, Berenstein and operators ec1 i1 Zelevinsky define dk(π) = 1 2 (γk−1 + γk)(Hik ). For the next proposition see [AB04], [Kav11a], and [BZ96]. Proposition 5.3. The valuation vi defines a bijection between B ⊂ C[U−\G] and the integral points in a convex polyhedral cone Ci ⊂ ZN × ∆ defined by the following inequalities: k dk(π)tk ≥ 0 for any i−trail ωi → w0siωi in V (ωi), for all fundamental (1) (cid:80) (2) tk ≤ λ(Hαik weights ωj of g. ) −(cid:80)N (cid:96)=k+1 ai(cid:96),ik t(cid:96) for k = 1, . . . , N. In particular vi(B) = vi(C[U−\G]), vi is a maximal rank valuation with image equal to a normal affine semigroup, and the Newton-Okounkov body of vi is Ci. Remark 5.4. Any set e1, . . . , em of k−linear nilpotent operators on a k−algebra A defines a valuation in this way. It would be very useful to know general sufficient conditions for the value semigroup v(cid:126)e(A) ⊂ Zm≥0 to be finitely generated. weight vectors, with weights ((cid:80) tiαi − λ, λ). In particular, vi is a T × T -stable For b ∈ B, with vi(b) = ((cid:126)t, λ) ∈ Ci, the tuple (cid:126)t ∈ ZN is called the string parameter of b associated to i. As constructed, the basis B is composed of T × T - valuation. The filtration F i corresponding to this valuation is given by T ×T -stable ⊂ C[U−\G], which has a basis of those b ∈ B with vi(b) ≤ ((cid:126)t, λ). subspaces F i≤((cid:126)t,λ) Now we consider the spaces Vβ,γ(λ) ⊂ V (λ), defined as the collection of those (acting vectors of weight γ which are annhilated by the raising operators e on the right). The basis B has the "good basis" property (see [Mat90]), this implies that Bβ,γ(λ) = B ∩ Vβ,γ(λ) is a basis for the space Vβ,γ(λ). In the case β = η, and γ = µ∗ − η this space is classically known to be isomorphic to the invariant space W (µ, λ, η) (see [Zhe70], we will also provide a proof in the next subsection). Berenstein and Zelevinksy characterize the string parameters (cid:126)t corresponding to the b ∈ Bη,µ∗−η(λ) as follows. Theorem 5.5 (Berenstein, Zelevinksy, [BZ96]). The decomposition i gives a la- belling of Bη,µ∗−η(λ) by the points in ZN≥0 such that the following hold: k dk(π)tk ≥ 0 for any i−trail from ωj to w0sjωj in V (ωj), for all funda- β(Hαi )+1 i mental weights ωj of g. k tkαk + λ + η = µ∗ k dk(π)tk ≥ η(Hαj ) for any i−trail from sjωj to w0ωj in V (ωj), for all (1) (cid:80) (2) −(cid:80) (3) (cid:80) (4) tk +(cid:80) fundamental weights ωj of g. j>k aik,ij tj ≥ λ(Hαik ) These are the integral points in a polytope Ci(µ, λ, η). The first and last conditions say that ((cid:126)t, λ) is a member of C(i) in the fiber over the weight λ, the second condition says that the basis members lie in the weight 18 CHRISTOPHER MANON µ∗ − η subspace of V (λ), and the third condition says that the appropriate raising annihilate. We realize these polytopes as slices of the following operators e polyhedral cone. η(Hαi )+1 i Definition 5.6. For a string parameterization i, the cone Ci(3) is defined by the following inequalities on ((cid:126)t, λ, η) ∈ C(i) × ∆ ⊂ ZN≥0 × ∆ × ∆ : k dk(π)tk ≥ 0 for any i−trail from ωj to w0sjωj in V (ωj), for all funda- mental weights ωj of g. k tkαk + λ + η ∈ ∆ k dk(π)tk ≥ η(Hαj ) for any i−trail from sjωj to w0ωj in V (ωj), for all (1) (cid:80) (2) −(cid:80) (3) (cid:80) (4) tk +(cid:80) We let π1((cid:126)t, λ, η) = (−(cid:80) 5.2. The tensor product ring map. The T ×T−stable subspace(cid:76) Vη,µ∗−η(λ)tη fundamental weights ωj of g. j>k aik,ij tj ≥ λ(Hαij ) k tkαk + λ + η)∗, π2((cid:126)t, λ, η) = λ, and π3((cid:126)t, λ, η) = η. By construction Ci(µ, λ, η) is the fiber of these maps over (µ, λ, η). ⊂ C[U−\G× T ] inherits a basis B3 from B × ∆ ⊂ C[U−\G× T ]. In this subsection we show that this space is an algebra, isomorphic to C[P3(G)]. As a result we will obtain both a basis B3 ⊂ C[P3(G)] and a T 3−invariant valuation vi,3 : C[P3(G)] → Ci(3) with vi,3(B3) equal to the integer points in Ci(3). We use the isomorphism P3(G) ∼= (U−\G × U−\G × G/U+)/G ∼= (U−\G × U−\G)/U+. The residual action of T 3 on P3(G) intertwines under this isomorphism with a torus T 3 ∼= T1 × T2 × T3 action on (U−\G × U−\G)/U+ defined as follows: the tori T2 and T3 act on the left hand sides of the components U−\G × U−\G, and the torus T1 acts diagonally on the right through the automorphism of T corresponding to the dual map −w0 : t∗ → t∗ on characters. We make use of the following commutative diagram of affine varieties. [T × U+ × T × U+]/U+ π←−−−− T × U+ × T (cid:121) (cid:121) [U−\G × U−\G]/U+ ←−−−− U−\G × T The top row is the map π : (s, u, t) → (s, u, t, Id), this is an isomorphism with inverse (s, x, t, y) → (xy−1, s, t). This map intertwines the left U+ × U+ action on (U+ × U+)/U+ with the left and right actions of U+ on itself. The bottom row is defined similarly, and the vertical arrows are given by the map (s, u) → su ∈ T U+ ⊂ U−\G. We identify T × U with the Borel a subgroup B ⊂ G, which is isomorphic as a group to T (cid:110) U . The T1 action is the defined as the right hand side action of the torus on B. Similarly the T2 and T3 actions are defined as the left hand actions on the first and second components T × U+, respectively. These actions make all of the maps in the above diagram T1 × T2 × T3-equivariant. Now we recall that the irreducible representation V (λ) has the following description as a subspace of C[U+] ([Mat90], [Zhe70]). POLYHEDRA FOR CHARACTER VARIETIES AND CONFIGURATION SPACES 19 V (λ) = {f ∈ C[U+]e λ(Hαi )+1 i ◦(cid:96) f = 0} (21) Here 1 ∈ C[U+] is identified with the highest weight vector bλ ∈ V (λ). We let Vη(λ) ⊂ V (λ) denote the space of functions f which satisfy e Lemma 5.7. The following diagram commutes, and the top row is an isomorphism of vector spaces. ◦r f = 0. η(Hαi )+1 i (V (λ) ⊗ V (η))U− −−−−→ (cid:121) Vη(λ) (cid:121) C[(T × U+ × T × U+)/U+] π∗−−−−→ C[T × U+ × T ] η(Hαi )+1 i λ(Hαi )+1 i ◦(cid:96) π∗(f ) = 0, and e ring C[G] =(cid:76) Proof. We take a function f ∈ V (λ)⊗V (η) ⊂ C[(T ×U+×T ×U+)/U+] and analyze ◦(cid:96) f = 0 in the the pullback π∗(f ). The function f satisfies the equations e ◦(cid:96) f = 0 in the second. By definition of π, these η(Hαi )+1 first U+ component and e i ◦r π∗(f ) = λ(Hαi )+1 equations are satisfied if and only if e i (cid:3) 0. Now we consider what happens when f ∈ (V (λ)⊗ V (η))U+ has weight µ∗, this is the case when f represents an intertwiner V (µ) → V (λ) ⊗ V (η). In the coordinate λ∈∆ V (λ)⊗V (λ∗), specialization at Id is contraction of V (λ) against the dual V (λ∗). The coordinate ring C[U−\G] ⊂ C[G] is the subalgebra of spaces Cv−η ⊗ V (η), it therefore follows that π∗(f ) ∈ C[U+] is the coefficient of the vη component of f . Since f was chosen to have weight µ∗, π∗(f ) ∈ Vη(λ) must be a µ∗ − η weight vector. Lemma 5.8. The space W (µ, λ, η) ∼= (V (λ)⊗V (η))U− µ∗ is isomorphic to Vη,µ∗−η(λ). Proof. We have already established a 1−1 map π∗ : (V (λ)⊗V (η))U+ µ∗ → Vη,µ∗−η(λ). To show that this map is also onto, we observe that (V (λ)⊗V (η))U+ is a direct sum of its dominant weight spaces, each of which maps to a distinct Vη,µ∗−η(λ) ⊂ Vη(λ), and that (V (λ) ⊗ V (η))U+ ∼= Vη(λ). (cid:3) The torus T1 × T2 × T3 acts on the space Vη,µ∗−η(λ) with character ((µ∗ − identifying C[P3(G)] with the subspace(cid:76) η + η)∗, λ, η) = (µ, λ, η). We have now established a T 3−stable map of algebras, µ,λ,η Vη,µ∗−η(λ)tη ⊂ C[U−\G × T ]. Theorem 5.9. For each i there is a T 3 stable valuation vi,3 on C[P3(G)], with associated graded ring equal to the affine semigroup algebra C[Ci(3)]. The torus T1 × T2 × T3 acts on C[Ci(3)] with characters π1, π2, π3 : Ci(3) → ∆. Furthermore, vi,3(B3) coincides with the integer points in Ci(3). Proof. The valuation vi,3 is constructed from < on ∆ and vi on C[U−\G]. It is invariant with respect to T1 × T2 × T3, because vi, < : C[U−\G × T → Ci is T 4 invariant. By construction the character with respect to the torus action on C[Ci(3)] corresponds to the maps π1, π2, π3. The image vi,3(C[P3(G)]) is then the image of µ,λ,η Vη,µ∗−η(λ)tη under vi, <. This (cid:3) those b ⊗ tη ∈ C[U−\G × T ] which lie in (cid:76) coincides with the integer points in Ci(3) by construction. 20 CHRISTOPHER MANON This exposition has been for the semisimple case, but as in [AB04], everything can be generalized readily to the reductive case. The weights that define a non- zero W (µ, λ, η) are of the form µ(cid:48) + τ1, η(cid:48) + τ2 and λ(cid:48) + τ3 where τi are characters of the center Z(G) with τ1 + τ2 + τ3 = 0, and µ(cid:48), η(cid:48), λ(cid:48) are dominant weights of the semisimple part of G. The subspace Vη,µ−η(λ) is the same as the subspace Vη(cid:48),µ(cid:48)−β(cid:48)+(τ1−τ2)(λ(cid:48) + τ3) = Vη(cid:48),µ(cid:48)−η(cid:48)+τ3(λ(cid:48) + τ3) = Vη(cid:48),µ(cid:48)−η(cid:48)(λ(cid:48))⊗Cτ3. So this space inherits the subset of the dual canonical basis of the semi-simple part of G coming from Vη(cid:48),µ(cid:48)−β(cid:48)(λ(cid:48)) tensored with the character τ3. 6. Proof of Theorems 1.1 and 1.2 We construct filtrations on C[X (Fg, G)] and C[Pn(G)] with toric associated graded algebras C[Ci(Γ)], C[Ci(T , (cid:126)λ)]. Choose Γ, with a total ordering on E(Γ), a total ordering on V (Γ), and an assignment i : V (Γ) → R(w0). The efforts of Section 4 give a filtration on C[X (Fg, G)] by (∆, <)E(Γ) with asso- ciated graded algebra C[PΓ(G)]. In Section 5 we construct a T 3 invariant valuation on C[P3(G)] with associated graded algebra C[Ci(3)]. We use the ordering on V (Γ) and the assignment i : V (Γ) → R(w0) to define a full rank, T E(Γ)−stable valuation C[Ci(v)(3)]. This C[P3(G)] of v∈V (Γ) T E(Γ) invariants, and by Theorem 5.9 and Proposition 3.5 the associated graded algebra is the semigroup algebra of the following polyhedral cone. on (cid:78) valuation induces a valuation on the sub-algebra C[PΓ(G)] ⊂(cid:78) C[P3(G)] with associated graded algebra (cid:78) v∈V (Γ) v∈V (Γ) We define Ci(Γ) to be the toric fiber product cone in(cid:81) Definition 6.1. We let πv,i be the projection map on Ci(v)(3) defined by the edge i incident on the vertex v ∈ V (Γ). the conditions πv,i = π∗ u,i for all edges i with endpoints u, v. v∈V (Γ) Ci(v)(3) defined by Theorem 1.1 follows from Proposition 3.3. The same program can be carried out on the algebra C[Pn(G)], giving a T n-stable valuation with associated graded algebra C[Ci(T )]. Theorem 1.2 then follows by specializing the weights at the leaves of T to (cid:126)λ, using Theorem 5.9. Proposition 6.2. For every choice of a trivalent tree T with n ordered leaves, and an assignment i : V (T ) → R(w0) we have: (1) a basis B(T , m(cid:126)λ) ⊂ H 0(P(cid:126)λ∗ (G),L(m(cid:126)λ∗)) = (V (λ1) ⊗ . . . ⊗ V (λn))G, (2) a labelling vT ,i : B(T , m(cid:126)λ) → Ci(T , m(cid:126)λ) ⊂ (∆ × ZN≥0 × ∆)V (T ). We conclude by showing that the image of the valuations we have constructed coincide with all of the lattice points in their corresponding Newton-Okounkov bodies. This implies that the corresponding affine semigroup algebras are normal. Proposition 6.3. The integer points in Ci(Γ) and Ci(T ) are in bijection with the images of the induced valuations vi,Γ(C[X (Fg, G)]) and vi,T (C[Pn(G)]), respectively. Proof. Each basis member of C[X (Fg, G)] gives an element of Ci(Γ) by the con- structions in Sections 4, 5. If (cid:126)t ∈ Ci(Γ) is an integer point, it is likewise an integer point in (∆ × ZN≥0 × ∆)V (T ), and therefore corresponds to a product ⊗v∈V (Γ)bv of dual canonical basis elements which defines an regular function in C[X (Fg, G)] POLYHEDRA FOR CHARACTER VARIETIES AND CONFIGURATION SPACES 21 with value equal to (cid:126)t by construction. The same argument applies in the case (cid:3) C[Pn(G)]. 7. Examples We describe the cones Ci(3) for G = SLm(C) and all rank 2 simple, simply- connected groups. For G = SLm(C) we describe particular instances of Bi(Γ), Bi(T ). The inequalities we present are culled from [BZ01], the treatment by Littelman [Lit98], and Theorem 5.5. 7.1. Type A. For G = SLm(C), we take i to be the "good" decomposition w0 = s1(s2s1)(s3s2s1) . . . (sm−1 . . . s1), (see [Lit98]). The polyhedron Ci(3) is then the cone BZ3(SLm(C)) of Berenstein-Zelevinsky triangles [BZ92]. Definition 7.1. For this definition we refer to Figure 5. A BZ triangle T ∈ BZ3(SLm(C)) is an assignment of non-negative integers to vertices of the version of the diagram in Figure 5 with 2(m−1) vertices on a side. If v, w are a pair of vertices which are across a hexagon from a pair u, y, then T (v) + T (w) = T (u) + T (y). We let a1, . . . , a2m−2 = b1, . . . , b2m−2 = c1, . . . , c2m−2 = a1 label the vertices clockwise around the boundary of the diagram. This lets us define the following three projection maps π1, π2, π3 : BZ3(SLm(C)) → ∆SLm(C). (22) π1(T ) = (a1 + a2, . . . , a2m−3 + a2m−2) π2(T ) = (b1 + b2, . . . , b2m−3 + b2m−2) π3(T ) = (c1 + c2, . . . , c2m−3 + c2m−2) Figure 5. Left: A honeycomb weighting. Right: Dual weighting by non-negative integers. We can associate a dual graph to each T ∈ BZ3(SLm(C)) by replacing each entry of weight a with an edge to the center of its adjacent hexagon weighted a. The resulting graphs are also called honeycombs, and have been studied by a number of authors, see e.g. [KTW04], [GP00]. v∈V (Γ) BZ3(SLm(C)) is then de- fined to be those tuples (Tv) with πi,v(Tv) = πi,u(Tu)∗ when an edge i joins u and v, see Figures 6 and 7. The cone of Γ−BZ quilts BZΓ(SLm(C)) ⊂(cid:81) 1111111122 CHRISTOPHER MANON Figure 6. A BZ quilt with dual 4−tree. Figure 7. A BZ quilt with dual genus 3 graph. We represent gluing triangles T1, T2 with matching boundary components as weighted graphs on composite diagrams, as in Figure 6. Paths at the meeting boundaries of T1, T2 are joined by an arrangement of weighted paths in an X configuration. When the path has weight 1, this is either a line (see the left corner of the quilt in Figure 6) or by a crooked line (see the right corner of the quilt in Figure 6). The cones BZT (SL3(C)) are studied by the author and Zhou in [MZ14], see also [KM14] for a relationship between Berenstein-Zelevinsky triangles and mathematical biology. We finish this subsection with an analysis of the generators of the semigroup alge- bra C[BZΓ(SL2(C))]. Recall from Example 4.12 that the semigroup BZ3(SL2(C)) is the free semigroup on three generators, we depict an element of this semigroup as an arrangment of three types of paths in a dual trinode τ , see Figure 8. Figure 8. Three different interpretations of an element of BZ3(SL2(C)). POLYHEDRA FOR CHARACTER VARIETIES AND CONFIGURATION SPACES 23 For an element T ∈ BZ3(SL2(C)), counting the number of endpoints in each edge e, f, g of τ produces an integer weighting wT : {e, f, g} → Z≥0. We associate a planar arrangement of paths P (w) in Γ, by replacing the weights at each trinode with paths as in Figure 8. The endpoints of these paths in an edge e are then connected in the unique planar way. A path γ in the graph Γ has an associated weighting wγ : E(Γ) → Z≥0 obtained by setting wγ(e) equal to the number of times γ passes through e. For a more in depth account of this construction, see [Man14]. Theorem 7.2. The semigroup BZΓ(SL2(C)) is generated by the w : E(Γ) → Z≥0 with w(e) ≤ 2. Figure 9. Using an orientation to factor off a loop. (cid:80) Proof. Fix a w ∈ BZΓ(SL2(C)), and consider the induced planar arrangement of paths P (w) in Γ with multiweight w. Suppose w(e) > 2 for some edge e ∈ E(Γ). We pick a path γ ∈ P (w) which passes through e. If γ passes through e with weight 1, then we may remove wγ to obtain a weighting w(cid:48) with strictly smaller total weight f∈E(Γ) w(cid:48)(f ). If γ weights e greater than 1, we assign an orientation to γ. If two components at e have the same direction, we may alter the weighting as in Figure 9, yielding two closed paths γ(cid:48) ∪ γ(cid:48)(cid:48). Without loss of generality we assume that P (w) = {γ}, so that the weightings wγ(cid:48) and wγ(cid:48)(cid:48) satisfy wγ(cid:48) + wγ(cid:48)(cid:48) = w. In this case we pull off the new closed path γ(cid:48), which has strictly smaller total weight. If w(e) > 2 at least (cid:3) two components through e must have the same direction. As a corollary, a set of regular functions in C[X (Fg, SL2(C))] which represent the w ∈ BZΓ(SL2(C)) with w(e) ≤ 2 form a finite generating set. Such a set is said to be a finite "Khovanskii basis" of the filtration defined by Γ. 7.2. G2. We give inequalities for the tensor product cones for G2, with R(w0) = {α1α2α1α2α1α2, α2α1α2α1α2α1}. These cones have six string parameters t1, t2, t3, t4, t5, t6, and six weight parameters λ = (λ1, λ2), η = (η1, η2), µ = (µ1, µ2). The cone Cα1α2α1α2α1α2(3) is defined by the following inequalities: (23) 6t2 ≥ 2t3 ≥ 3t4 ≥ 2t5 ≥ 6t6 ≥ 0; λ2 ≥ 2t6 +=24 (24) (25) (26) CHRISTOPHER MANON η1 ≥ t1 − 3t2 − t3 − 3t4 − t5 − 3t6, t3 − 3t4 − t5 − 3t6, t5 − 3t6 η2 ≥ t6, t4 − t5 − 3t6, t2 − t3 − 3t4 − t5 − 3t6 2t1 − 3t2 + 2t3 − 3t4 + 2t5 − 3t6 = λ1 + η1 − µ1 −t1 + 2t2 − t3 + 2t4 − t5 + 2t6 = λ2 + η2 − µ2 The alternative cone Cα2α1α2α1α2α1 (3) is defined by the following inequalities: (27) (28) 2t2 ≥ 2t3 ≥ t4 ≥ 2t5 ≥ 2t6 ≥ 0 λ1 ≥ t6; λ2 ≥ t2 + t4 − t5, t2 + t5 − t6, t3 − t4 − t6, t5 − 3t6, t2 + t3 − 2t4, 2t2 − t4, 3t2 − t3, t3 − t5, t4 − 2t6, 2t4 − t5 − t6, 3t4 − 2t5 (29) (30) (31) η1 ≥ t6, t4 − 3t5 − t6, t2 − 3t3 − t4 − 3t5 − t6 η2 ≥ t5 − t6, t3 − t4 − 3t5 − t6, t1 − t2 − 3t3 − t4 − 3t5 − t6 2(t2 + t4 + t6) − 3(t1 + t3 + t5) = λ1 + η1 − µ1 2(t1 + t3 + t5) − (t2 + t4 + t6) = λ2 + η2 − µ2 7.3. Sp4. We give the inequalities for the Sp4 tensor product cones corresponding to the decompositions R(w0) = {α1α2α1α2, α2α1α2α1}. The cone Cα1α2α1α2 (3) is defined by the following inequalities: (32) (33) (34) (35) t2 ≥ t3 ≥ t4, λ2 ≥ t4, 2t3 − t2, t2 − 2t1; λ1 ≥ t1 η2 ≥ t2 + 2t4 − 2t3, t4; η1 ≥ t1 + 2t3 − 2t2, t3 − t2 2t1 − t2 + 2t3 − t4 = λ1 + η1 − µ1 −2t1 + 2t2 − 2t3 + 2t4 = λ2 + η2 − µ2 The cone Cα2α1α2α1(3) is defined by the following inequalities: POLYHEDRA FOR CHARACTER VARIETIES AND CONFIGURATION SPACES 25 (36) (37) (38) t2 ≥ t3 ≥ t4 ≥ 0 ; λ2 ≥ 2t1, t2; λ1 ≥ 2t1 η1 ≥ t3 − t4, t1 − t2 + 2t3; η2 ≥ t2 − 2t3 + 2t4, t4 t1 + t2 + t3 + t4 = λ1 + η1 − µ1; t2 + t4 − t1 − t3 = λ2 + η2 − µ2 References [AB04] [Bae00] Valery Alexeev and Michel Brion. Toric degenerations of spherical varieties. Selecta Math. (N.S.), 10(4):453 -- 478, 2004. John C. Baez. An introduction to spin foam models of BF theory and quantum gravity. In Geometry and quantum physics (Schladming, 1999), volume 543 of Lecture Notes in Phys., pages 25 -- 93. Springer, Berlin, 2000. [BH93] Winfried Bruns and Jurgen Herzog. Cohen-Macaulay rings, volume 39 of Cambridge [BZ92] [BZ96] [BZ01] [Cal02] Studies in Advanced Mathematics. Cambridge University Press, Cambridge, 1993. A. D. Berenstein and A. V. Zelevinsky. Triple multiplicities for sl(r+1) and the spectrum of the exterior algebra of the adjoint representation. J. Algebraic Combin., 1(1):7 -- 22, 1992. Arkady Berenstein and Andrei Zelevinsky. Canonical bases for the quantum group of type Ar and piecewise-linear combinatorics. Duke Math. J., 82(3):473 -- 502, 1996. Arkady Berenstein and Andrei Zelevinsky. Tensor product multiplicities, canonical bases and totally positive varieties. Invent. Math., 143(1):77 -- 128, 2001. Philippe Caldero. Toric degenerations of Schubert varieties. Transform. Groups, 7(1):51 -- 60, 2002. [CV86] Marc Culler and Karen Vogtmann. Moduli of graphs and automorphisms of free groups. [Dol03] [FG06] [FH91] [FL09] [FL12] [FL13] Invent. Math., 84(1):91 -- 119, 1986. Igor Dolgachev. Lectures on invariant theory, volume 296 of London Mathematical Society Lecture Note Series. Cambridge University Press, Cambridge, 2003. Vladimir Fock and Alexander Goncharov. Moduli spaces of local systems and higher Teichmuller theory. Publ. Math. Inst. Hautes ´Etudes Sci., (103):1 -- 211, 2006. R. Fulton and J. Harris. Representation Theory: A First Course, volume 129 of Grad- uate Texts in Mathematics. Springer-Verlag, New York, 1991. Carlos Florentino and Sean Lawton. The topology of moduli spaces of free group rep- resentations. Math. Ann., 345(2):453 -- 489, 2009. Carlos Florentino and Sean Lawton. Singularities of free group character varieties. Pacific J. Math., 260(1):149 -- 179, 2012. Carlos Florentino and Sean Lawton. Character varieties and moduli of quiver represen- tations. In In the tradition of Ahlfors-Bers. VI, volume 590 of Contemp. Math., pages 9 -- 38. Amer. Math. Soc., Providence, RI, 2013. [Gol84] William M. Goldman. The symplectic nature of fundamental groups of surfaces. Adv. in Math., 54(2):200 -- 225, 1984. [Gol86] William M. Goldman. Invariant functions on Lie groups and Hamiltonian flows of sur- [GP00] [Gro97] face group representations. Invent. Math., 85(2):263 -- 302, 1986. Oleg Gleizer and Alexander Postnikov. Littlewood-Richardson coefficients via Yang- Baxter equation. Internat. Math. Res. Notices, (14):741 -- 774, 2000. Frank D. Grosshans. Algebraic homogeneous spaces and invariant theory, volume 1673 of Lecture Notes in Mathematics. Springer-Verlag, Berlin, 1997. [HJL+09] Roger Howe, Steven Jackson, Soo Teck Lee, Eng-Chye Tan, and Jeb Willenbring. Toric degeneration of branching algebras. Adv. Math., 220(6):1809 -- 1841, 2009. [HMM11] Benjamin Howard, Christopher Manon, and John Millson. The toric geometry of tri- angulated polygons in Euclidean space. Canad. J. Math., 63(4):878 -- 937, 2011. [HTW05] Roger E. Howe, Eng-Chye Tan, and Jeb F. Willenbring. A basis for the GLn tensor product algebra. Adv. Math., 196(2):531 -- 564, 2005. 26 CHRISTOPHER MANON [Kav11a] K. Kaveh. Crystal bases and newton-okounkov bodies. arXiv:1101.1687 [math.AG], 2011. [Kav11b] Kiumars Kaveh. Note on cohomology rings of spherical varieties and volume polynomial. J. Lie Theory, 21(2):263 -- 283, 2011. [KK12] Kiumars Kaveh and A. G. Khovanskii. Newton-Okounkov bodies, semigroups of integral points, graded algebras and intersection theory. Ann. of Math. (2), 176(2):925 -- 978, 2012. [KM14] Kaie Kubjas and Christopher Manon. Conformal blocks, Berenstein -- Zelevinsky trian- gles, and group-based models. J. Algebraic Combin., 40(3):861 -- 886, 2014. [Law10] [KTW04] A. Knutson, T. Tao, and C. Woodward. The honeycomb model of GLn(C) tensor products II: Puzzles determine facets of the Littlewood-Richardson cone. J. Amer. Math. Soc., 17(1):19 -- 48, 2004. Sean Lawton. Algebraic independence in SL(3, C) character varieties of free groups. J. Algebra, 324(6):1383 -- 1391, 2010. P. Littelmann. Cones, crystals, and patterns. Transform. Groups, 3(2):145 -- 179, 1998. Robert Lazarsfeld and Mircea Mustat¸a. Convex bodies associated to linear series. Ann. Sci. ´Ec. Norm. Sup´er. (4), 42(5):783 -- 835, 2009. [Lit98] [LM09] [Lus90] G. Lusztig. Canonical bases arising from quantized enveloping algebras. J. Amer. Math. Soc., 3(2):447 -- 498, 1990. [Man10] C. Manon. Toric degenerations and tropical geometry of branching algebras. arXiv:1103.2484 [math.AG], 2010. [Man14] C. Manon. Toric geometry of sl2()¸ free group character varieties from outer space. arXiv:1410.0072 [math.AG], 2014. [MH] [Mat90] Olivier Mathieu. Good bases for G-modules. Geom. Dedicata, 36(1):51 -- 66, 1990. [MFK94] D. Mumford, J. Fogarty, and F. Kirwan. Geometric invariant theory, volume 34 of Ergebnisse der Mathematik und ihrer Grenzgebiete (2) [Results in Mathematics and Related Areas (2)]. Springer-Verlag, Berlin, third edition, 1994. K. Kaveh M. Harada. Toric degenerations, integrable systems and okounkov bodies. arXiv:1205.5249. John W. Morgan and Peter B. Shalen. Valuations, trees, and degenerations of hyper- bolic structures. I. Ann. of Math. (2), 120(3):401 -- 476, 1984. Christopher Manon and Zhengyuan Zhou. Semigroups of sl3(C) tensor product invari- [MZ14] ants. J. Algebra, 400:94 -- 104, 2014. [Oku97] A. Yu. Okun(cid:48)kov. A remark on the Hilbert polynomial of a spherical manifold. Funkt- [MS84] sional. Anal. i Prilozhen., 31(2):82 -- 85, 1997. [Pop86] V. L. Popov. Contractions of actions of reductive algebraic groups. Mat. Sb. (N.S.), [Sik12] [SS04] 130(172)(3):310 -- 334, 431, 1986. Adam S. Sikora. Character varieties. Trans. Amer. Math. Soc., 364(10):5173 -- 5208, 2012. David Speyer and Bernd Sturmfels. The tropical Grassmannian. Adv. Geom., 4(3):389 -- 411, 2004. [Zhe70] D. P. Zhelobenko. Kompaktnye gruppy Li i ikh predstavleniya. Izdat. "Nauka", Moscow, [ZS76] 1970. O. Zariski and P. Samuel. Commutative Algebra volume II, volume 29 of Graduate Texts in Mathematics. Springer-Verlag, New York, 1976. Christopher Manon: Department of Mathematics, George Mason University, Fairfax, VA 22030 USA
1211.6283	1	1211	2012-11-27T12:25:16	A vanishing theorem	[ "math.AG" ]	Let $E$ be a vector bundle and $L$ be a line bundle over a smooth projective variety $X$. In this article, we give a condition for the vanishing of Dolbeault cohomology groups of the form $H^{p,q}(X,\SSS^{\alpha}E\otimes \wedge^{\beta} E\otimes L)$ when $S^{\alpha+\beta}E \otimes L$ is ample. This condition is shown to be invariant under the interchange of $p$ and $q$. The optimality of this condition is discussed for some parameter values.	math.AG	math	A GENERAL VANISHING THEOREM F. LAYTIMI AND W. NAHM Abstract. Let E be a vector bundle and L be a line bundle over a smooth projective variety X. In this article, we give a condition for the vanishing of Dolbeault cohomology groups of the form H p,q(X,S αE ⊗ ∧βE ⊗ L) when Sα+βE ⊗ L is ample. This condition is shown to be invariant under the interchange of p and q. The optimality of this condition is discussed for some parameter values. . G A h t a m [ 1 v 3 8 2 6 . 1 1 2 1 : v i X r a 1. Introduction Throughout this paper X will denote a smooth projective variety of dimension n over the field of complex numbers, E a vector bundle of rank e, and L a line bundle on X. For any non-negative integers α, β we denote by SαE, ∧βE the sym- metric product and the exterior product of E. H p,q(X, SαE⊗∧βE⊗L) will denote the Dolbeault cohomology group H q(X, SαE ⊗ ∧βE ⊗ L ⊗ Ωp X), where Ωp X is the bundle of exterior differential forms of degree p on X. We start with some definitions. Definition 1.1. The function δ : N ∪ {0} −→ N is the one which satisfies δ(x) = m ⇐⇒ (cid:18)m 2 (cid:19). 2(cid:19) ≤ x < (cid:18)m + 1 The last two inequalities imply δ(x) = [ √8x + 1 + 1 2 ], where the symbol [ ] denotes the integral part. i.e., δ(0) = 1, δ(1) = δ(2) = 2, δ(3) = δ(4) = δ(5) = 3, δ(6) = δ(7) = δ(8) = δ(9) = 4, . . . 1991 Mathematics Subject Classification. 14F17. 1 Theorem 1.2. Let α, β ∈ N. If Sα+βE ⊗ L is ample , then H p,q(X, SαE ⊗ ∧βE ⊗ L) = 0 for q + p − n > (r0 + α)(e + α − β) − α(α + 1), where r0 = min{β, δ(n − p), δ(n − q)}. Corollary 1.3. Let β be a positive integer. If SβE ⊗ L is ample, then H p,q(X,∧βE ⊗ L) = 0, for r0 = min{β, δ(n − p), δ(n − q)}. q + p − n > r0(e − β). where This Corollary improve the result of Manivel "theorem 1. p.91" in [13]. Corollary 1.4. Assume SαE ⊗ L is ample. Then H p,q(X, SαE ⊗ L) = 0, for q + p − n > α(e − 1). This article is the final version of several attempts [16], [11]. The result of these latest were used by Chaput in [3] and by Laytimi-Nagaraj in [7]. In [15] Manivel studied the vanishing of Dolbeault cohomology of a product of vector bundles tensored with certain power of their deter- minant. The presence of the latest allowed to deal with the problem by more direct method. 2. The Schur Functor Version of the theorem Our main result is a consequence of a Schur functor version of the theorem, but before giving this version, we need to recall some defini- tions and results: We start by some preparation on partitions and Schur functors (for a definition see [5]). A partition u = (u1, u2, . . . , ur) is a sequence of non increasing posi- tive integers ui. Its length is r and its weight is u = we put ui = 0. The zero-partition is the one where all ui are zero. ui. For i > r For any partition u the corresponding Schur functor is denoted by Su. Let V be a vector space of dimension d. To each partition u corre- sponds an irreducible Gl(V )-module Su(V ) which vanishes iff ud+1 > 0. 2 r Pi=1 For example, S(k)V = SkV . By functoriality the definition of Schur functors carries over to vector bundles E on X. By abuse of language we say that Su has a certain property, if u has this property. For example we will say Sk has weight k. Definition 2.1. The Young diagram Y (u) of a partition u is given by Y (u) = {(i, j) ∈ N2 j ≤ ui}. The transposed partition u is defined by Y (u) = {(i, j) ∈ N2 (j, i) ∈ Y (u)}. We use the notation ∧u = Su. Definition 2.2. The rank of a partition u, is rk(u) = max{ρ (ρ, ρ) ∈ Y (u)}. If rk(u) = 1, then u is called a hook. Notation 2.3. If u is a hook with u1 = α + 1 and u = k, we write In particular, Γ0 k = ∧k and Γk−1 Su = Γα k . k = Sk. Recall that SαE ⊗ ∧βE = Γα α+βE ⊕ Γα−1 α+βE. Definition 2.4. For partitions u, v of the same weight, the dominance partial ordering is defined by u (cid:23) v, iff j Xi=1 ui ≥ j Xi=1 vi for all j. This partial ordering can be extended to a pre-ordering of the set of all non-zero partitions of arbitrary weight u, v with u = n,v = m, by comparing as above the partitions of the same weight mu and nv, where the multiplication mu = m (u1, u2, . . . , ur) = (mu1, . . . , mur) ∀ m ∈ N. More precisely We write u (cid:23) v u ≃ v iff mu (cid:23) nv. iff u (cid:23) v and v (cid:22) u. When it is more convenient we will write Su (cid:23) Sv instead of u (cid:23) v. For example, ∧r ≻ ∧r+1, and Sα ≃ S1 for any α ∈ N. 3 Lemma 2.5. (Dominance Lemma) ([8] "theorem 3.7") For any partition u and v. If u (cid:23) v, then SuE ample =⇒ SvE ample. For example: If ∧2E is ample, then ∧3E is ample. Now we give the Schur presentation of the main theorem under which the main theorem will be shown. With the notation 2.3 we have: Theorem 2.6. Let k ∈ N. If SkE ⊗ L is ample , then H p,q(X, Γα k E ⊗ L) = 0, for q + p − n > (r0 + α)(e − k + 2α) − α(α + 1), where r0 = min{β, δ(n − p), δ(n − q)}. Proposition 2.7. Theorem 2.6 is equivalent to Theorem 1.2 Proof: Since S αE ⊗ ∧k−αE = Γα k E ⊕ Γα−1 we have only to show that for 1 ≤ α ≤ k− 1 the conditions of Theorem 1.2 imply the vanishing of H p,q(X, Γα−1 k E), but this is clear since the function (r0 + α)(e − k + 2α) − α(α + 1) is increasing in α. k E, 3. Some Technical Lemmas We start with some proprieties of the function δ defined in 1.1. Lemma 3.1. For µ ∈ N, x ∈ N such that (x + µδ(x), x − µδ(x)) ∈ N × N, we have 1) δ(x + δ(x)) = δ(x) + 1 2) δ(x + µδ(x)) ≤ δ(x) + µ 3) δ(x − µδ(x)) ≤ δ(x) − µ. Proof: The first assertion and the case µ = 1 in 2) and 3) are obvious. For both remaining assertions we use induction on µ. For 2) δ(x + µδ(x)) = δ(x + δ(x) + (µ − 1)δ(x)), since δ(x) ≤ δ(x + δ(x)) = δ(x) + 1, we have δ(x + δ(x) + (µ − 1)δ(x)) ≤ δ(x + δ(x) + (µ − 1)δ(x + δ(x)). Now induction hypothesis gives δ(x + δ(x) + (µ − 1)δ(x + δ(x)) ≤ δ(x + δ(x)) + µ − 1 = δ(x) + µ. 4 For 3) δ(x − µδ(x)) = δ(x − δ(x) − (µ − 1)δ(x)), δ(x − δ(x) − (µ − 1)δ(x)) ≤ δ(x − δ(x) − (µ − 1)δ(x − δ(x)). Induction hypothesis gives δ(x − δ(x) + (µ − 1)δ(x − δ(x)) ≤ δ(x − δ(x)) − (µ − 1). Now since it is true for µ = 1, we get δ(x − δ(x)) − (µ − 1) ≤ δ(x) − µ. Definition 3.2. Let φ : N × N → N × N × N the following injection φ(x, α) = (φ1(x, α), φ2(x, α), φ3(x, α)), where φ1(x, α) = δ(x) + α 2 (cid:19) φ2(x, α) = x −(cid:18)δ(x) φ3(x, α) = α We define an order on the pairs (x, α) ∈ N × N by the lexicographic order on N × N × N induced by φ, we denote this order by (x′, α′) ≤φ (x, α) The set N × N endowed with the above order will be denoted: (3.1) {N × N, ≤φ} := U Lemma 3.3. For µ ∈ Z − {0} and (x + µδ(x), α − µ) ∈ N × N, then (x + µδ(x), α − µ) ≤φ (x, α). where the order ≤φ is given in Definition 3.2. Proof: By Lemma 3.1 φ1(x + µδ(x), α − µ) ≤ α + δ(x). If δ(x + µδ(x)) = µ + δ(x), then 2 (cid:19) −(cid:18)µ 2 (cid:19). 2(cid:19) ≤ x −(cid:18)δ(x) φ2(x + µδ(x), α − µ) = x −(cid:18)δ(x) 2(cid:1) = 0, which means µ = 1, then If (cid:0)µ φ3(x + µδ(x), α − µ) = α − 1 < α. We need to use these following results Lemma 3.4. Let E an ample vector bundle and G an arbitrary vector bundle on a projective variety X. Then for sufficiently large enough n SnE ⊗ G is ample. 5 Lemma 3.5. Bloch-Gieseker [2] Let L be a line bundle on a projec- tive variety X and d be a positive integer. Then there exist a projective variety Y , a finite surjective morphism f : Y → X, and a line bundle M on Y, such that f ∗L ≃ M d. Lemma 3.6. Let p, q, n, f1, . . . , fr be fixed positive integers and α1, . . . , αr be fixed non-zero partitions. If H p,q(X,⊗r i=1SαiFi) = 0 for all smooth projective varieties X of dimension n and all ample vector bundles F1, . . . , Fr of ranks f1, . . . , fr on X, then this vanishing state- ment remains true if one of the Fi is ample and the others are nef. Proof: We can reorder the Fi such that F1 is ample. Let E = F1 and α = α1. Let N be a sufficiently large number such that SN E ⊗ det E∗ is ample (for the existence of such N see Lemma 3.4, and let a = Pm i=2 αi. By Lemma 3.5 we can find a finite surjective morphism f : Y → X, and a line bundle M on Y, such that f ∗(det E) = M N a. Then Ea = f ∗E ⊗ (M ∗)a is ample since SN Ea is. We have f ∗(SαE ⊗m i=2 SαiFi) = SαEa ⊗m i = M α ⊗ f ∗Fi for i = 2, . . . , m. All F ′ where F ′ i are ample. To finish the proof, we use "lemma 10 in [14] which says, For any vector bundle F on X and any finite surjective morphism f : Y → X, the vanishing of H p,q(Y, f ∗F ) implies the vanishing of H p,q(X,F ). i=2 SαiF ′ i , Lemma 3.7. Fix n, p, q, k, α ∈ N and t ∈ Z. Assume that H p,q(X, Γα k E) vanishes for all smooth projective varieties X of dimension n and all ample vector bundles E of rank e = k + t on X. Let α < k′ < k. Then H p,q(X, Γα k′E′) vanishes for all ample vector bundles E′ of rank e′ = k′ + t on X. Proof: For given E′, put E = E′ ⊕ L⊕(k−k′), where L is any ample k E, we have line bundle. Since Γα k′E′ ⊗ Lk−k′ is a direct summand of Γα H p,q(X, Γα ) = 0 k′E′ ⊗ Lk−k′ for ample vector bundle E′ of rank e′ and ample line bundle L. By Lemma 3.6 , this vanishing result remains true, when L is replaced by the trivial line bundle. (cid:3) Corollary 3.8. Assume that there is an integer k0 such that H p,q(X, Γα k E) = 0 if k > k0, 6 for any projective smooth variety X of dimension n and any ample vector bundle E of rank e, under the condition C(n, p, q, α, e − k). Then under this same condition the vanishing remains true for all k. The Bloch-Gieseker lemma can be used in other way to generalize vanishing theorems. In particular one has Lemma 3.9. Fix n, p, q, e ∈ N and partitions u, v of the same weight. Assume that H p,q(X,SuE) vanishes for all projective varieties X of dimension n and all vector bundles E of rank e for which SvE is am- ple. Let L be a line bundle and F a vector bundle of rank e. Then H p,q(X,SuF ⊗ L) = 0, if SvF ⊗ L is ample. Proof: Let's denote u = v = d. By Lemma 3.5 we can find a finite surjective morphism f : Y → X, and a line bundle M on Y, such that f ∗L = M d. Then (3.2) Due to the analogous equation (3.2) for Su one has by assumption Sv(f ∗F ⊗ M) = f ∗(SvF ⊗ L) is ample. H p,q(Y, f ∗(SuF ⊗ L)) = 0, and the vanishing of H p,q(X,SuF ⊗ L) follows by using "lemma 10 in [14]. (cid:3) The lemma applies for example if SvF is nef and L is ample. Corollary 3.10. To generalize vanishing of type H p,q(X,SuF ⊗ L), from L = OX to arbitrary L, it suffices to use Lemma 3.9. We need to recall Lemma 3.11. ([6] "lemma 1.3") Let X be a projective variety, E, F be vector bundles on X. If E is ample and F nef , then E ⊗ F is ample. 4. The Borel-Le Potier Spectral Sequence To prepare the proof, we need a lemma and some properties of the Borel-Le Potier spectral sequence, which has been made a standard tool in the derivation of vanishing theorems [4]. Let E be a vector bundle over a smooth projective variety X, dim(X) = n. Let Y = Gr(E) be the corresponding Grassmann bundle and Q be the canonical quotient bundle over Y . Lemma 4.1. Let l, r be positive integer and k = lr, if ∧rE is ample. Then for P + q > n + r(e − r) H P,q(Gr(E), det Ql) = 0. 7 Proof: Since det Q = OP(∧rE)(1)Gr(E). Thus ΛrE ample implies that det Q is ample. One conclude by using Nakano-Akizuki-Kodaira van- ishing theorem [1]. (cid:3) Definition 4.2. Let π : Y → X be a morphism of projective manifolds, P a positive integer and F a vector bundle over Y . The Borel-Le Potier spectral sequence P E given by the data π, P,F is the spectral sequence which abuts to H P,q(Y,F ) , it is obtained from the filtration on ΩP Y ⊗F which is induced by the filtration on the bundle ΩP Y of exterior differential forms of degree P . F p(ΩP Y ) = π∗Ωp X ∧ ΩP −p Y The graded bundle which corresponds to the filtration on ΩP Y is given by F p(ΩP Y )/F p+1(ΩP Y ) = π∗Ωp X ⊗ ΩP −p Y /X , where ΩP −p Thus the E1 terms of P E have the form Y /X is the bundle of relative differential forms of degree P − p. P Ep,q−p 1 = H q(Y, π∗Ωp X ⊗ ΩP −p Y /X ⊗ F ). These E1 terms can be calculated as limits groups of the Leray spec- tral sequence associated to the projection π, p,P Eq−j,j 2,L = H p,q−j(X, Rjπ∗(ΩP −p Y /X ⊗ F )) Now we consider the Borel-Le Potier spectral sequence which abuts to H P,q(Gr(E), det Ql). Proposition 4.3. Let π : Gr(E) = Y → X, the E1 terms of the Borel-Le Potier spectral sequence given by π, P, det Ql have the form P Ep,q−p 1 = Mu∈ σ(P −p,r) H q(Gr(E),SuQ∗ ⊗ det Ql ⊗ ∧uS ⊗ π∗Ωp X). Here S is the tautological sub-bundle of π∗E over Y and σ(p, r) is the set of partitions of weight p and length at most r. Proof: One has ΩY /X = Q∗ ⊗ S. Thus Y /X = Mu∈σ(P −p,r) ΩP −p SuQ∗ ⊗ ∧uS. (cid:3) Obviously Leray spectral sequence degenerates at the E2,L level. Using the corollary 1. in ([13] page 94) of Bott formula, Manivel computes the E1 terms under some condition on P, ([13] Proposition 3. page 96). He states his result under the supplementary condition e ≥ k, which is not necessary for the calculation. 8 Proposition 4.4. [13] Assume P ≥ n + (l − 1)(cid:0)r+1 α(p) = 2 (cid:1) − l(r − 1), and k = lr. Let (l − 1)(r + 1) 2 − P − p 2(cid:19) − (r − 1)α(p). 2 j(p) = (l − 1)(cid:18)r Then the E1 terms of the spectral sequence have the form P Ep,q 1 = (cid:26) H p,q−j(p)(X, Γα,kE) 0 for (n − p, α(p)) ∈ U otherwise, where the set U is defined in (3.1). Note that the connecting morphisms of Borel-Le Potier spectral se- quence dm : P Ep,q−p m −→ P Ep+m,q−p+1−m m all vanish, unless m is a multiple of r since under dm the integer α goes to the integer α + m r . 5. Proof of the main theorem Before giving the proof of the main theorem, we will first explain the case r0 = β in the main theorem, which corresponds to Corollary 5.2 bellow. We need to recall these results Theorem 5.1. [9] Let Ei be vector bundles, with rank(Ei) = ei, over a smooth projective variety X of dimension n, and let L be a line bundle on X. If ⊗m i=1ΛriEi ⊗ L is ample, then H p,q(X,⊗m i=1ΛriEi ⊗ L) = 0 for p + q − n > ri(ei − ri). m Xi=1 Corollary 5.2. Let E be a vector bundle of rank e, and let L be a line bundle on a smooth projective variety X of dimension n. If Sα+βE⊗ L is ample, then H p,q(X, SαE ⊗ ΛβE ⊗ L) = 0 for 9 q + p − n > α(e − 1) + β(e − β). Proof: We will apply the Theorem 5.1 to the vector bundle ⊗ΛβE ⊗ L, which SαE ⊗ ΛβE ⊗ L is a direct summand E ⊗ E · · · ⊗ E } of. {z α times (5.1) Let's first show this equivalence of ampleness SαE ⊗ F ≃ E ⊗ E · · · ⊗ E } {z α times ⊗F for any vector bundles F. Indeed: For the first direction, Note that SαE⊗L is direct summand of E ⊗ E · · · ⊗ E } For the second direction, Littlewood-Richardson rules gives, ⊗F. {z α times E ⊗ E · · · ⊗ E } {z α times = SαE ⊕ Xλ=α SλE, we have clearly α ≻ λ in the dominance partial order. Use Remark 2.5 to conclude. Now by Littlewood-Richardson rules SαE ⊗ ΛβE = ⊕ SνE, with ν = α + β, satisfying Sν ≺ Sα+β. Thus the ampleness of Sα+βE ⊗ L implies the ampleness of SαE ⊗ ΛβE ⊗ L by Remark 2.5. Use the equivalence of ampleness (5.1) to conclude. (cid:3) Due to Remark 3.10 one can prove our main theorem without L. U is given in Definition 3.2. We prove Theorem 1.2 by induction on (n− p, α) ∈ U, where the set Assume that the result is true for all pairs (p′, α′) such that with respect to the order introduced in Definition 3.2. (n − p′, α′) ≤φ (n − p, α), Choose r = δ(n − p). Let l be arbitrary if n = p, otherwise let . Choose P such that α(p) = α, and consider the Borel-Le l ≥ rα+n−p Potier spectral sequence. Then for k = lr r−1 P Ep,q+j(p)−p 1 = H p,q(X, Γα,kE). When m is a multiple of r, the morphisms dm connect P Ep,q+j(p)−p 1 with P Ep′,q′+j(p′)−p′ 1 where for the terms on the right of P Ep,q+j(p)−p 1 (5.2) p′ = p + µr, q′ = q + µ(r − 1) + 1, 10 and (5.3) p′ = p − µr, q′ = q − µ(r − 1) − 1 for the terms on the left. Here µ is any positive integer. Lemma 5.3. For any integers p′ and q′ of the form (5.2) or (5.3), P Ep′,q′+j(p′)−p′ 1 = 0, when q > Q(n − p, α), where Q(n − p, α) = n − p + (δ(n − p) + α)(e − k + 2α) − α(α + 1). Proof: The assertion is trivially true for α(p′) < 0 or e−k +α(p′) < We need to prove that the assertion q > Q(n − p, α) implies the The terms on the right of P Ep,q+j(p)−p have α′ = α + µ, and the 0, such that we may assume e − k + 2α(p′) ≥ 0. assertion q′ > Q(n − p′, α′). parameters in (5.2), a straight calculation yields 1 Q(n − p, α) − Q(n − p′, α′) + µ(δ(n − p) − 1) + 1 = (5.4) (e − k + 2(α + µ)(δ(n − p) − µ − δ(n − p − µδ(n − p)) + µ2 + 1. have α′ = α − µ, and the The terms on the left of P Ep,q+j(p)−p parameters in (5.3), the calculation yields 1 Q(n − p, α) − Q(n − p′, α′) − µ(δ(n − p) − 1) − 1 = (5.5) (e − k + 2(α − µ)(δ(n − p) + µ − δ(n − p + µδ(n − p)) + µ2 − 1. 1 By Lemma 3.1 both terms of (5.4) and (5.5) are non negative and positive if µ 6= 1. Thus q > Q(n − p, α) implies q′ > Q(n − p′, α(p′)). By Lemma 3.3 (n − p′, α(p′)) ≤φ (n − p, α), such that the groups vanish by induction hypothesis. Thus all co-bordant is a P Ep′,q′+j(p′)−p′ morphisms of P Ep,q+j(p)−p sub-factor of H P,q+j(p)−p(Y, F ), where F = det(Q)l. vanish. This implies that P Ep,q+j(p)−p Recall that P = p + (l − 1)(cid:0)r+1 Thus the condition q > Q(n − p, α) is equivalent to 2 (cid:1) − αr and dimY = n + r(e − r). 1 1 P + q + j(p) − dimY > α(e − k + α). When the right hand side is non-negative, H P,q+j(p)−p(Y, F ) = 0 by Nakano-Kodaira-Akizuki vanishing theorem. Thereby k E) = 0 for q > Q(n − p, α). H p,q(X, Γα 11 Remember that this proof was under the condition k = rl see Proposi- tion 4.4, but this condition can be removed by Corollary 3.8. To get r0 = δ(n − q) in our theorem, we interchange the role of p and q at every stage of the proof, in particular we use r = δ(n − q). 6. Optimality Proposition 6.1. Let G = Gr(r ,d) be the Grassmannian of all co- dimensional r subspaces of a vector space V of dimension d = f + r. Let Q be the universal sub-bundle of rank r on G, dim G = n = fr . Then, for q = n − f , α = f − 1 H q(G, S αQ ⊗ Q ⊗ det Q ⊗ KX ) 6= 0 Proof: Since Sα+1Q is direct summand of SαQ ⊗ Q, it's enough to show H q(G, S α+1 Q ⊗ det Q ⊗ KX ) 6= 0 . For the universal sub-bundle S on G, we have KG = ((det Q)∗)⊗d = det S ⊗d. Thus since α = f − 1 H q(G, S f Q ⊗ det Q ⊗ KX ) = H q (G, S f Q ⊗ det S ⊗(d−1 ). Now by Bott formula (see corollary 1. page 94 of [13]) H q(G, S f Q ⊗ det S ⊗(d−1 ) = δq,i((a,b)−c(d))Sψ(a,b)V , where a = (f, 0,· · · , 0 {z } r−1 times For any sequence v = (v1, v2, . . .) ), b = (d − 1,· · · , d − 1 } {z d−r times ). where i(v) = card{(i, j) / i < j, vi < vj}, ψ(v) = (v − c(d))≥ + c(d), c(d) = (1, 2, . . . , d), and (v)≥ is the partition obtained by ordering the terms of v in non increasing order. (a, b) = (f, 0,· · · , 0 {z } r−1 times ). ), d − 1,· · · , d − 1 } {z d−r times ((a, b) − c(d)) = (f − 1,−2,−3, . . . ,−r, f − 2, f − 3, . . . , 0,−1), 12 we get i((a, b) − c(d)) = f (r − 1) = n − f, and ). } ψ(a, b) = (f, f, . . . , f d times {z Thus Sψ((a,b)V = (det V )⊗f . Note that the non-vanishing example of the above proposition hap- pens for the limit condition q + p − n = (r0 + α)(e + α − β) − α(α + 1), r0 = min{β, δ(n − p), δ(n − q)}. where (cid:3) References [1] Y. Akizuki, S. Nakano, Note on Kodaira-Spencer's proof of Lefschetz theo- rems, Proc.Jap.Acad. 30 (1954), 266-272. [2] S. Boch, D. Gieseker, The positivity of the Chern classes of an ample vector bundle, Invent. Math. 12 (1971), 112-117 [3] P.E. Chaput, Th´eor`emes d'annulation et Lieux de d´eg´en´erescence en petit corang, Documenta Math. 9 (2004),449-525. [4] J-P. Demailly, Vanishing theorems for tensor powers of an ample vector bun- dle, Invent. Math. 91 (1988),no. 1, 203-220. [5] W. Fulton, J. Harris, Representation theory,, a first course, Graduate texts in Mathematics, Springer Verlag 1991. [6] F. Laytimi, On Degeneracy Loci, International Journal of Mathematics Vol. 7 6 (1996),745-754. [7] F. Laytimi, D.S. Nagaraj vanishing theorems for vector bundles generated by sections, Kyoto Journal of Mathematics vol. 50 Number 3, 469-480 (2010). [8] F. Laytimi, W. Nahm, A generalization of Le Potier's vanishing theorem, Manuscripta math. 113 (2004),165-189. [9] F. Laytimi, W. Nahm, A vanishing theorem Nagoya Math. J.180 (2005), 35-43 . [10] F. Laytimi, W. Nahm, On a Vanishing Problem of Demailly, International Mathematics Research Notices 47 (2005),2877-2889. [11] F. Laytimi, W. Nahm, A vanishing theorem for Product of exterior and sym- metric powers, e-print math.AG/9809064. [12] F. Laytimi Generalization of Peternell, Le Potier and Schneider vanishing theorem Manuscripta Mathematica: Volume 134, Numbers 3-4, (2011), Pages 485-492. [13] L. Manivel, Un th´eor`eme d'annulation pour les puissances ext´erieures d'un fibr´e ample, J. reine angew. Math. 422 (1991), 91-116. [14] L. Manivel, Th´eor`emes d'annulation pour les fibr´es associ´es `a un fibr´e ample, Scuola superiore Pisa (1992), 515-565 [15] L. Manivel, Vanishing theorems for ample vector bundles, Invent. math. 127 (1997), 401-416. [16] W. Nahm, A vanishing theorem for Product of exterior and symmetric powers, preprint, Univ. Bonn, Germany (1995) 13 F. L.: Math´ematiques - bat. M2, Universit´e Lille 1, F-59655 Vil- leneuve d'Ascq Cedex, France E-mail address: [email protected] W. N.: Dublin Institute for Advanced Studies, 10 Burlington Road, Dublin 4, Ireland E-mail address: [email protected] 14
1404.6783	3	1404	2016-11-19T10:43:14	Birational Geometry of Singular Moduli Spaces of O'Grady Type	[ "math.AG" ]	Following Bayer and Macr\`{i}, we study the birational geometry of singular moduli spaces $M$ of sheaves on a K3 surface $X$ which admit symplectic resolutions. More precisely, we use the Bayer-Macr\`{i} map from the space of Bridgeland stability conditions $\mathrm{Stab}(X)$ to the cone of movable divisors on $M$ to relate wall-crossing in $\mathrm{Stab}(X)$ to birational transformations of $M$. We give a complete classification of walls in $\mathrm{Stab}(X)$ and show that every birational model of $M$ obtained by performing a finite sequence of flops from $M$ appears as a moduli space of Bridgeland semistable objects on $X$. An essential ingredient of our proof is an isometry between the orthogonal complement of a Mukai vector inside the algebraic Mukai lattice of $X$ and the N\'{e}ron-Severi lattice of $M$ which generalises results of Yoshioka, as well as Perego and Rapagnetta. Moreover, this allows us to conclude that the symplectic resolution of $M$ is deformation equivalent to the 10-dimensional irreducible holomorphic symplectic manifold found by O'Grady.	math.AG	math	BIRATIONAL GEOMETRY OF SINGULAR MODULI SPACES OF O'GRADY TYPE CIARAN MEACHAN AND ZIYU ZHANG Abstract. Following Bayer and Macr`ı, we study the birational geometry of singular moduli spaces M of sheaves on a K3 surface X which admit symplectic resolutions. More precisely, we use the Bayer-Macr`ı map from the space of Bridgeland stability conditions Stab(X) to the cone of movable divisors on M to relate wall-crossing in Stab(X) to birational transformations of M . We give a complete classification of walls in Stab(X) and show that every minimal birational model of M in the sense of the log minimal model program appears as a moduli space of Bridgeland semistable objects on X. An essential ingredient of our proof is an isometry between the orthogonal complement of a Mukai vector inside the algebraic Mukai lattice of X and the N´eron-Severi lattice of M which generalises results of Yoshioka, as well as Perego and Rapagnetta. Moreover, this allows us to conclude that the symplectic resolution of M is deformation equivalent to the 10-dimensional irreducible holomorphic symplectic manifold found by O'Grady. 6 1 0 2 v o N 9 1 ] . G A h t a m [ 3 v 3 8 7 6 . 4 0 4 1 : v i X r a 1. Introduction 1.1. Background. Let X be a complex projective K3 surface, v ∈ H ∗ alg(X, Z) a Mukai vector and H an ample line bundle which is generic with respect to v (in the sense of [HL10, Section 4.C] and [Yos01, Section 1.4]). We can always write the Mukai vector as a multiple of a primitive class, say v = mvp. If we further assume that v2 > 0 with respect to the Mukai pairing, then there is a precise classification of moduli spaces MH(v) of Gieseker H-semistable sheaves on X with Mukai vector v which goes as follows: • If m = 1, then v is a primitive Mukai vector and MH (v) is smooth. Moreover, in his seminal paper [Muk84], Mukai showed that MH(v) is an irreducible holomorphic symplectic manifold, parametrising H-stable sheaves. • If m > 2, then MH(v) has symplectic singularities (in the sense of [Bea00]). Its smooth locus parametrises H-stable sheaves, while its singular locus parametrises S-equivalence classes of strictly semistable sheaves. This case splits into two radically different situa- tions: – If m = 2 and v2 p = 2, then MH(v) has a symplectic resolution; see [O'G99, LS06]. – If m > 2 or v2 p > 2, then MH (v) does not admit a symplectic resolution; see [KLS06]. The geometry of smooth moduli spaces MH (v), or irreducible holomorphic symplectic manifolds in general, has been the subject of intensive study for many years. In particular, there are many results about their birational geometry in the literature; see [Huy03, HT09] for instance. On the other hand, Bridgeland [Bri08, Section 14] showed that these Gieseker moduli spaces can 2010 Mathematics Subject Classification. 14D20 (Primary); 14F05, 14J28, 18E30 (Secondary). Key words and phrases. Bridgeland stability conditions, derived categories, moduli spaces of sheaves and complexes, wall crossing, symplectic resolutions. 1 2 CIARAN MEACHAN AND ZIYU ZHANG be realised as moduli spaces of σ-semistable objects in the 'large volume limit' of (a certain connected component Stab†(X) of) his stability manifold Stab(X). More precisely, he proved that the stability manifold comes with a wall and chamber decomposition in the sense that the set of σ-semistable objects with some fixed numerical invariants is constant in each chamber and an object of the bounded derived category D(X) of coherent sheaves on X can only become stable or unstable by crossing a wall, that is, a real codimension one submanifold of Stab†(X). Furthermore, Bridgeland conjectured that crossing a wall should induce a birational transfor- mation between the corresponding moduli spaces and this vision was recently crystallised in a revolutionary paper [BM14a] by Bayer and Macr`ı. Before going any further, we should say that these ideas concerning wall-crossing have been successfully applied to the study of the birational geometry of moduli spaces of sheaves on abelian surfaces; see [Yos09, MYY11, Mea12, Mac14, Yos12, AB13, MM13, MYY14, YY14]. However, the focus of this paper will be on moduli spaces of sheaves on K3 surfaces as above. For any Mukai vector v with v2 > 0, on each moduli space of Bridgeland semistable objects Mσ(v), Bayer and Macr`ı used the classical technique of determinant line bundles to construct an ample line bundle ℓσ on Mσ(v). In particular, this gives rise to a 'linearisation map' from any chamber C in Stab†(X) to the nef cone of the moduli space Nef(MC(v)) corresponding to that chamber. This map is an essential part of this paper; we call it the Bayer-Macr`ı map. Moreover, when the Mukai vector v is primitive, they obtained a complete picture relating wall crossing on Stab†(X) to the birational geometry of the corresponding moduli spaces. They studied how the moduli space changes when crossing a wall and classified all the walls in terms of the Mukai lattice. In particular, they verified the conjecture of Bridgeland, by showing that moduli spaces corresponding to two neighbouring chambers are indeed birational. For each type of wall, they could produce an explicit birational map, which identifies the two moduli spaces away from loci of codimension at least two and hence identifies the N´eron-Severi groups of them. Using these birational maps, they could glue the linearisation maps defined on each chamber together to get a global continuous Bayer-Macr`ı map from Stab†(X) to the N´eron-Severi group of any generic moduli space Mσ(v) and prove that, via wall crossings on Stab†(X), every birational minimal model of Mσ(v) appears as a Bridgeland moduli space. 1.2. Summary of main results. This paper grew out of an attempt to first understand [BM14b, BM14a] and then generalise their techniques to the simplest singular case in the hope of obtaining similar results. In particular, we are interested in the case when v = 2vp and vp is a primitive Mukai vector with square equal to 2 with respect to the Mukai pairing. We say that such a Mukai vector v is of O'Grady type. The benefit of considering this type of Mukai vector v is that we will be able to show the existence of a symplectic resolution of the moduli space Mσ(v) under any generic stability condition σ, which will allow us to reuse many arguments in [BM14a]. Our first main result is the following Theorem (5.4). Let X be a projective K3 surface and v be a Mukai vector of O'Grady type. For any two generic stability conditions σ, τ ∈ Stab†(X) with respect to v, there is a birational map Φ∗ : Mσ(v) 99K Mτ (v) induced by a derived (anti-)autoequivalence1 on D(X), which is an isomorphism in codimension one. 1An anti-autoequivalence is an equivalence of categories from D(X)op to D(X), which takes every exact triangle in D(X) to an exact triangle in D(X), but with all the arrows reversed. All examples of this notion encountered in this paper are compositions of autoequivalences with the derived dual functor RHom(−, OX ); see the discussion after [BM14a, Theorem 1.1]. BIRATIONAL GEOMETRY OF SINGULAR MODULI SPACES OF O'GRADY TYPE 3 This result is a generalisation of [BM14a, Theorem 1.1]. Although the statements are quite similar, the singular version is more involved. For instance, in the primitive case, every moduli space under a generic stability condition is smooth and has trivial canonical class. Therefore, any birational map between two such moduli spaces naturally extends to an isomorphism in codimension one [GHJ03, Proposition 21.6]. However, this does not hold in the singular case. To fix this issue, we introduce the notation of a stratum preserving birational map between two singular moduli spaces of O'Grady type and show that the behaviour of such birational maps in codimension one is as nice as it is in the smooth world; see Proposition 2.6. A refinement of Theorem 5.4, which generalises [BM14a, Theorem 1.2], is the following Theorem (7.6). Let v be a Mukai vector of O'Grady type and σ ∈ Stab†(X) be any generic stability condition with respect to v. Then (1) We have a globally defined continuous Bayer-Macr`ı map ℓ : Stab†(X) → NS(Mσ(v)), which is independent of the choice of σ. Moreover, for any generic stability condition τ ∈ Stab†(X), the moduli space Mτ (v) is the birational model corresponding to ℓτ . (2) If C ⊂ Stab†(X) is the open chamber containing σ, then ℓ(C) = Amp(Mσ(v)). (3) The image of ℓ is equal to Big(Mσ(v)) ∩ Mov(Mσ(v)). In particular, every K-trivial Q-factorial birational model of Mσ(v) which is isomorphic to Mσ(v) in codimension 1 appears as a moduli space Mτ (v) for some generic stability condition τ ∈ Stab†(X). In our situation, Mσ(v) is a K-trivial Q-factorial variety with canonical (hence log-terminal) singularities. Therefore Mσ(v) (together with an empty divisor) is a log minimal model. For such symplectic varieties which admit symplectic resolutions, the existence and termination of log- flips have been established in [BCHM10, Corollary 1.4.1] and [LP14, Theorem 4.1] respectively. Therefore the log minimal model program works in this case. In particular, every K-trivial Q-factorial birational model of Mσ(v) which is isomorphic to it in codimension 1 (in other words, every log minimal model which is log-MMP related to Mσ(v)) can be obtained through a finite sequence of log-flops. Theorem 7.6 shows that every such log minimal model has an interpretation as a Bridgeland moduli space Mτ (v) for some generic stability condition τ . This picture is parallel to [BM14a, Theorem 1.2] which deals with the case of primitive Mukai vectors. As an application of Theorem 7.6, we can also formulate a Torelli-type theorem for singular moduli spaces of O'Grady type, which is parallel to [BM14a, Corollary 1.3]; see Corollary 7.8 for more details. The proof of Theorem 5.4 and Theorem 7.6 relies on a complete classification of walls, as stated in Theorem 5.1 and Theorem 5.3, which generalises [BM14a, Theorem 5.7]. Although our proofs follow their approach very closely, the technical details are much more involved. The main difficulty is that the proof of [BM14a, Theorem 5.7] uses many results on irreducible holomorphic symplectic manifolds which are not available to us in the singular world. Each time they use an argument which relies on smoothness in an implicit way, we have to either find a way around it or prove a new version that works in the singular case. One instance of this, which leads to an interesting by-product of the project, goes as follows. The main ingredient for the Bayer-Macr`ı map is the classical construction of determinant line bundles on MH (v); see [HL10, Section 8.1]. Its algebraic version, which is often referred to as the Mukai morphism, is a map of lattices θσ : v⊥ → NS(Mσ(v)) where the orthogonal complement is taken in the algebraic cohomology H ∗ alg(X, Z). When v is primitive, Yoshioka 4 CIARAN MEACHAN AND ZIYU ZHANG proved in [Yos01] that θσ is an isometry with respect to the Mukai pairing on H ∗ the Beauville-Bogomolov pairing on NS(Mσ(v)). alg(X, Z) and However, when v is a Mukai vector of O'Grady type, such a theorem for Bridgeland moduli spaces does not seem to exist in literature. In fact, to make sense of it, we have to find a well-defined pairing on the N´eron-Severi lattice of the moduli space in the first place. Luckily, a special case of it, on the level of cohomology, concerning only Gieseker moduli spaces, was proved in [PR13]. Using their approach, we can similarly define the bilinear pairing on NS(Mσ(v)), or more generally on H 2(Mσ(v)). Although their proof of the isometry does not adapt immediately to the case of Bridgeland moduli spaces, all the key ideas are there and we are able to deduce the desired result using Fourier-Mukai transforms and deformation theory. A necessary and crucial step in this project, as well as an interesting result in its own right, is the analogous result concerning the cohomological version of the Mukai morphism in the singular setting. In particular, we have Theorem (2.7). Let v be a Mukai vector of O'Grady type and σ ∈ Stab†(X) be a generic stability condition with respect to v. Denote Mσ(v) by M and its symplectic resolution by π : fM → M . Then (1) The pullback map π∗ : H 2(M, Z) → H 2(fM , Z) is injective and compatible with the weight two, and the restriction of the Beauville-Bogomolov quadratic form eq(−,−) on H 2(fM , Z) defines a quadratic form q(−,−) on H 2(M, Z). σ : v⊥,tr → H 2(M, Z) induced by the (mixed) Hodge structures. In particular, H 2(M, Z) carries a pure Hodge structure of (2) There exists a well-defined Mukai morphism θtr (quasi-)universal family over M st σ (v), which is a Hodge isometry. We point out that v⊥,tr in the above theorem denotes the orthogonal complement in the total cohomology H ∗(X, Z). If σ lies in the Gieseker chamber then Theorem 2.7 is precisely [PR13, Theorem 1.7]. The more general statement follows by combining the ideas of Perego and Rapagnetta with some new ingredients inspired by [BM14a]. More precisely, for an arbitrary σ ∈ Stab†(X), we use a Fourier-Mukai transform to identify Mσ(v) with a moduli space of twisted sheaves on another K3 surface. Then we deform the underlying twisted K3 surface to an untwisted K3 surface where the isometry has been proved in [PR13, Theorem 1.7]. The observation in [PR13] that this isometry is preserved under these two operations finishes the proof. However, the presence of a non-trivial Brauer class makes the deformation argument more complicated. For instance, the issue of ampleness caused by (−2)-classes and genericness of the polarisation, turn out to be far from straightforward. As an immediate consequence, we obtain the algebraic version of the Theorem 2.7 as follows. Corollary (2.8). Under the assumptions of Theorem 2.7, we have (1) Lefschetz (1, 1) theorem holds for M . That is, NS(M ) = H 1,1(M, Z); (2) The restriction of the pullback map π on NS(X) is an injective map π∗ : NS(M ) → NS(fM ), which is compatible with the Beauville-Bogomolov pairings q(−,−) and eq(−,−); σ on the algebraic Mukai lattice is an isometry θσ : v⊥ → NS(M ). In particular, q(−,−) is a non-degenerate pairing on NS(M ) with signature (1, ρ(X)). (3) The restriction of the Mukai morphism θtr BIRATIONAL GEOMETRY OF SINGULAR MODULI SPACES OF O'GRADY TYPE 5 We remark that it is only the third statement in the above corollary which is needed in our study of the Bayer-Macr`ı map. However, we do not know a direct proof of this statement which does not utilise the cohomological Mukai morphism; the reason being that our proof involves a deformation argument. Unlike cohomology groups which stay constant in a family as a local system, the N´eron-Severi groups do not behave well under deformations. As a consequence of Theorem 2.7, we obtain a generalisation of [PR13, Theorem 1.6] from Gieseker moduli spaces to Bridgeland moduli spaces. The proof combines the arguments in [PR13] and the deformation techniques developed in Section 3 below. Corollary (3.16). Let v be a Mukai vector of O'Grady type and σ ∈ Stab†(X) be any generic stability condition with respect to v. Then the symplectic resolution of Mσ(v) is deformation equivalent to the irreducible holomorphic symplectic manifold constructed by O'Grady in [O'G99]. This result in particular implies that, by resolving singular moduli spaces of Bridgeland semistable objects on K3 surfaces, we cannot get any new deformation types of irreducible holomorphic symplectic manifolds other than the one discovered by O'Grady in [O'G99]. It is somewhat disappointing, but supports the long-standing belief that deformation types of irreducible holo- morphic symplectic manifolds are rare. 1.3. Outline of the paper. In Section 2, we collect together the necessary properties of moduli spaces of O'Grady type that we will need. After briefly mentioning their stratifications and resolutions, we study birational maps between them and state Theorem 2.7. We also give a brief account of various cones of divisors on these moduli spaces. Section 3 is devoted to the proof of Theorem 2.7. We provide some lemmas on the deformations of twisted polarised K3 surfaces, as well as on the existence of local relative moduli spaces. The main proof comes after all these lemmas. In Section 4, we briefly review the Bayer-Macr`ı map constructed in [BM14b], but only from the aspect that will become important in our later discussion. We also use the Bayer-Macr`ı map to generalise the ampleness results proved in [BM14b]. We refer interested readers to [BM14b, Section 3] for the original construction, which provides a very conceptual way to understand the positivity lemma there. We state our classification theorems of potential walls in Section 5; see Theorems 5.1 and 5.3. We also establish the birational maps relating moduli spaces for two chambers separated by a wall; see Theorem 5.4. In this section we only state the results, while leaving all proofs for next section. It is worth pointing out that our first classification theorem is true for arbitrary Mukai vectors, while the second classification theorem only works for Mukai vectors of O'Grady type. Section 6 is devoted to the proof of all results in the section 5. We try to make our proofs short and avoid repeating any existing arguments by making many references to results in [BM14a, Section 6-9]. Readers who are not interested in proofs can safely skip this section without affecting the coherence of logic. In Section 7, we describe and prove our main result. In particular, Theorem 7.6 provides a precise relationship between wall crossings on the stability manifold and the birational geometry of the corresponding moduli spaces. Our main references of the paper are [BM14b, BM14a]. Since our presentation closely follows theirs, all background knowledge required for this paper are already included in the first few sections of those. Nevertheless, we recommend readers the following references for general [HL10] for moduli spaces of sheaves, [GHJ03, Part 3] for knowledge of some relevant topics: 6 CIARAN MEACHAN AND ZIYU ZHANG geometry of irreducible holomorphic symplectic manifolds, [Bri08] for stability conditions on K3 surfaces, and [Cal00] for twisted sheaves. Acknowledgements: We are most grateful to Arend Bayer and Alastair Craw for all their help, support and encouragement throughout this project. Special thanks to Daniel Huybrechts for his invaluable advice and guidance at various stages of this work. We also thank Christian Lehn, Sonke Rollenske and K¯ota Yoshioka for kindly answering our questions, and the referees for their very helpful comments. Z. Z. would also like to thank Jun Li for his initial suggestion of looking into this topic. C. M. is supported by an EPSRC Doctoral Prize Research Fellowship Grant EP/K503034/1 and Z. Z. is supported by an EPSRC Standard Research Grant EP/J019410/1. We also appreciate the support from the University of Bonn, the Max Planck Institute for Mathematics, and the SFB/TR-45 during the initial stage of this project as well as the Hausdorff Research Institute for Mathematics for its conclusion. 2. Singular Moduli Spaces of O'Grady Type We start our discussion by collecting some useful properties of the singular moduli spaces of O'Grady type, which will be the central geometric objects that are studied in the whole paper. We will see that despite of the singularities, these moduli spaces behave very similar to smooth moduli spaces, in the sense that many nice properties of smooth moduli spaces can be generalised to these with some extra care of the singular loci. After introducing necessary notations and background materials, we will mainly focus on two aspects of these moduli spaces: birational maps between them and Mukai morphisms on their cohomology. After that, we will briefly mention various cones of divisors on singular moduli spaces of O'Grady type. 2.1. Moduli spaces and symplectic resolutions. We start by recalling the basic notion of a Bridgeland moduli space. Let X be a projective K3 surface and v ∈ H ∗ alg(X, Z) be a Mukai vector. Throughout this paper we will always assume v2 > 0. Moreover, there is a unique way to write v = mvp for some positive integer m and primitive class vp ∈ H ∗ alg(X, Z). We say m is the divisibility and vp is the primitive part of v. Let σ ∈ Stab†(X) be a Bridgeland stability condition in the distinguished component of the stability manifold. The stability manifold Stab†(X) comes with a wall and chamber structure with respect to v as described above (see also [Bri08, Section 9] and [BM14b, Proposition 2.3]), and we say σ ∈ Stab†(X) is generic if it does not lie on any wall. It was proven in [BM14b, Theorem 1.3] (which generalises [MYY14, Theorem 0.0.2]) that, for a stability condition σ ∈ Stab†(X) which is generic with respect to v, there exists a coarse moduli space MX,σ(v), which parametrises the S-equivalence classes of σ-semistable objects of class v on X. Furthermore, it is a normal projective irreducible variety with Q-factorial singularities. By [BM14a, Theorem 2.15] (or originally [Yos01, Tod08]), MX,σ(v) is non-empty, and a generic point of it represents a σ-stable object. A classical theorem, originally proved by Mukai in [Muk84] (see also [BM14b, Theorem 6.10] and [BM14a, Theorem 3.6]), says that when σ is generic and v is primitive, the moduli space MX,σ(v) is an irreducible holomorphic symplectic manifold, which parametrises σ-stable objects of class v only. However, when v is non-primitive, it is proved in [KLS06, Theorem 6.2] for Gieseker moduli spaces and [BM14a, Theorem 3.10] for Bridgeland moduli spaces, that MX,σ(v) has symplectic singularities. For simplicity, we often drop the K3 surface X from the notation when it is clear from the context. In fact, the moduli space only depends on the choice of the chamber C containing σ. BIRATIONAL GEOMETRY OF SINGULAR MODULI SPACES OF O'GRADY TYPE 7 Therefore we sometimes also denote the moduli space for any stability condition contained in a chamber C by MX,C(v), or simply MC(v). For the convenience of later discussion, we also make the following definition. Definition 2.1. We say a Mukai vector v ∈ H ∗ as v = mvp, where m = 2 and v2 is a generic stability condition, and v is a Mukai vector of O'Grady type. alg(X, Z) is of O'Grady type if it can be written p = 2. We say a moduli space MX,σ(v) is of O'Grady type if σ The importance of this particular type of moduli spaces lies in the study of symplectic resolutions In [O'G99], O'Grady constructed a symplectic resolution of a Gieseker moduli of MX,σ(v). space with Mukai vector v = (2, 0,−2) and showed that it was not deformation equivalent to any existing example of homomorphic symplectic manifolds at the time. In [LS06], Lehn and Sorger generalised the result to arbitrary Gieseker moduli spaces of O'Grady type, and gave a slightly different description of their symplectic resolutions. It was proved in [PR13] that all these symplectic resolutions are in fact deformation equivalent to the one constructed by O'Grady. In [KLS06], it was proved that for a generic Gieseker moduli space with any other non-primitive Mukai vector, a symplectic resolution does not exist. By using the techniques developed in [BM14b, Section 7], we can easily generalise the existence of symplectic resolutions to Bridgeland moduli spaces with the following proposition. Later in Corollary 3.16, we will show that all these symplectic resolutions are still deformation equivalent to the one constructed by O'Grady, hence do not provide new deformation type of irreducible holomorphic symplectic manifolds. Proposition 2.2. Let X be a projective K3 surface and v = mvp ∈ H ∗ vector with m > 2, vp primitive and v2 Then alg(X, Z) be a Mukai p > 0. Let σ ∈ Stab†(X) be generic with respect to v. • If m = 2 and v2 • If m > 2 or v2 p = 2, then Mσ(v) admits a symplectic resolution; p > 2, then Mσ(v) does not admit a symplectic resolution. Proof. By [BM14b, Lemma 7.3] and the discussion after that, there exists a twisted K3 surface (Y, α) where α ∈ Br(Y ) and a derived equivalence Φ : D(X) → D(Y, α) in the form of a Fourier-Mukai transform, such that Φ induces an isomorphism MX,σ(v) ∼= MY,α,H(−Φ(v)), where MY,α,H(−Φ(v)) is the moduli space of α-twisted Gieseker H-semistable locally free sheaves on Y . Hence it suffices to prove the claims for twisted Gieseker moduli spaces. Therefore without loss of generality, we replace the Bridgeland moduli space in question by a twisted Gieseker moduli space MX,α,H (v). By [Lie07, Proposition 2.3.3.6], there is a GIT construction for the twisted Gieseker moduli spaces, which is precisely the same as in the case of untwisted Gieseker moduli spaces. And by [Lie07, Proposition 2.2.4.9], the local deformation theory of twisted sheaves is also the same as that of untwisted sheaves. Therefore the argument in [LS06] shows that MX,α,H (v) admits a symplectic resolution as the blowup of its singular locus when m = 2 and v2 p = 2. And the argument in [KLS06] shows that MX,α,H(v) has no symplectic resolution when m > 2 or v2 (cid:3) p > 2. The existence of symplectic resolutions is critical for most results in the present paper. The following proposition is such an example. One property of irreducible holomorphic symplectic manifolds used in [BM14a] is that, for a divisorial contraction on an irreducible holomorphic symplectic manifold, the image of the contracted divisor has codimension exactly two, which is a 8 CIARAN MEACHAN AND ZIYU ZHANG special case of [Nam01, Proposition 1.4], [Wie03, Theorem 1.2] or [Kal06, Lemma 2.11]. For our purpose, we need a version of this result in singular case. Thanks to the existence of symplectic resolutions, the same result can be easily proved for moduli spaces of O'Grady type. But we nevertheless state it under a more general setup as follows. Proposition 2.3. Let M be a variety with symplectic singularities of dimension 2n admitting a symplectic resolution, and let N be a normal projective variety. Let ϕ : M → N be a birational projective morphism. We denote by Si the set of points p ∈ N such that dim ϕ−1(p) = i. Then dim Si 6 2n−2i. In particular, if ϕ contracts a divisor D ⊂ M , then we have dim ϕ(D) = 2n−2. if V ⊂ M is any Proof. The statement in question is equivalent to the following statement: closed subvariety of dimension 2n − i, then dim ϕ(V ) > 2n − 2i; see for instance the proof of [Nam01, Proposition 1.4]. Without loss of generality, it suffices to prove this statement with the extra assumption that V is irreducible with the generic point ξV . Let the symplectic resolution of M be π : fM → M . We denote the closure of π−1(ξV ) in fM by eV . Then we have π(eV ) = V , hence dimeV > dim V = 2n − i. Then we apply [Nam01, Proposition 1.4], [Wie03, Theorem 1.2] or [Kal06, Lemma 2.11] on the composition ϕ ◦ π : fM → N and conclude dim ϕ(V ) = dim(ϕ ◦ π)(eV ) > 2n − 2i. However, we remind the readers that although many of our results are proved in the context of definition 2.1, some of our results do work in more general situations. We will state very clearly which assumptions are made in every result. (cid:3) 2.2. Stratum preserving birational maps. Whenever σ is generic, there is a stratification of the moduli space MX,σ(v) given by locally closed strata. The stable locus M st X,σ(v), which agrees with the smooth locus, is the unique open stratum. All the other lower dimensional strata are formed by lower dimensional moduli spaces. We refer the readers to the proof of [BM14a, Theorem 2.15] for the general case. Here we only describe the stratification for moduli spaces of O'Grady type. For a moduli space of O'Grady type MX,σ(v), there is a chain of closed subschemes as follows: MX,σ(vp) ⊂ Sym2 MX,σ(vp) ⊂ MX,σ(v), (2.1) where the first inclusion is given by the diagonal morphism, and the second inclusion gives precisely the strictly semistable locus of MX,σ(v), which agrees with the singular locus. This chain of inclusions decomposes MX,σ(v) into the disjoint union of three locally closed strata. X,σ(v) = MX,σ(v)\ Sym2 MX,σ(vp) parametrises all σ-stable objects of class v. More precisely, M st Every point in Sym2 MX,σ(vp) represents the S-equivalent class containing a polystable object E1 ⊕ E2, where E1, E2 ∈ MX,σ(vp) are both σ-stable objects of class vp. Such a point lies in the diagonal MX,σ(vp) if and only if E1 and E2 are isomorphic. With this stratification at hand, we are now ready to discuss birational maps between moduli spaces of O'Grady type, which are compatible with the above stratifications. When talking about birational maps between singular moduli spaces of O'Grady type, we empha- sise a special class of them, which preserve the natural stratifications described above. Almost all birational maps between these moduli spaces which occur in this paper belong to this class. Although the definition could be made for arbitrary moduli spaces under generic stability con- ditions, for the purpose of this paper, we restrict ourselves to moduli spaces of O'Grady type as follows. BIRATIONAL GEOMETRY OF SINGULAR MODULI SPACES OF O'GRADY TYPE 9 Definition 2.4. Let f : MX1,σ1(v1) 99K MX2,σ2(v2) be a birational map between two moduli spaces of O'Grady type. Let MXi,σi(vi,p) ⊂ Sym2 MXi,σi(vi,p) ⊂ MXi,σi(vi) be the standard stratification (2.1) for i = 1, 2, where vi,p denotes the primitive part of vi. If f is defined on the generic point of each stratum in MX1,σ1(v1), and takes each such generic point to the generic point of the corresponding stratum in MX2,σ2(v2), then we say that f is a stratum preserving birational map. Derived (anti-)equivalences are a very natural and particularly rich resource of stratum preserv- ing birational maps. The following lemma gives a criterion for such a derived (anti-)equivalence to induce a stratum preserving birational map. Lemma 2.5. Let Φ : D(X1) → D(X2) be a derived (anti-)equivalence and let MX1,σ1(v1) and MX2,σ2(v2) be two moduli spaces of O'Grady type. Assume that Φ induces a birational map Φ∗ : MX1,σ1(v1) 99K MX2,σ2(v2). Then Φ∗ is stratum preserving if and only if the following condition holds: there exist a σ1-stable object E of class v1 and a σ1-stable object Ep of class v1,p, such that Φ(E) and Φ(Ep) are σ2-stable objects of classes v2 and v2,p respectively. Proof. The necessity is part of the definition of f being stratum preserving, so we only discuss sufficiency. By the openness of stability in [BM14a, Theorem 4.2] (which was originally proved in [Tod08]), the assumptions imply that the induced birational map Φ∗ takes the generic points of the open stratum M st (v1) and the closed stratum MX1,σ1(v1,p) to the generic points of corresponding strata in MX2,σ2(v2). It remains to show that Φ∗ takes the generic point of the singular locus Sym2 MX1,σ1(v1,p) to the generic point of Sym2 MX2,σ2(v2,p). In fact, a generic point Es in Sym2 MX1,σ1(v1,p) can be represented by any extension of two generic stable objects of class v1,p. Since Φ is a derived (anti-)equivalence, it preserves extensions (or switches the direction). Hence Φ(Es) is again the extension of two generic stable objects of class v2,p, which represents a generic point in Sym2 MX2,σ2(v2,p), as desired. (cid:3) X1,σ1 A big advantage of stratum preserving birational maps is that they behave very much like birational maps between smooth symplectic varieties. For example, the following proposition generalises a classical result about a birational map between two K-trivial smooth varieties, for instance, in [GHJ03, Proposition 21.6]. Proposition 2.6. Let f : MX1,σ1(v1) 99K MX2,σ2(v2) be a stratum preserving birational map between two moduli spaces of O'Grady type which is induced by a derived (anti-)equivalence Φ : D(X1) → D(X2). Furthermore, assume that there exists an open subset U ⊂ M st (v1) with complement of codimension at least two, such that the restriction fU is an injective morphism fU : U → M st (v2). (v2). Then f (U ) has complement of codimension at least two in M st X2,σ2 X1,σ1 X2,σ2 X1,σ1 Proof. Since f is a stratum preserving birational map, there exists an open subset Us ⊂ Sym2 MX1,σ1(v1,p), such that fUs : Us → Sym2 MX2,σ2(v2,p) is an injective morphism. Now (v1)\U in MX1,σ1(v1), Z2 to be Sym2 MX1,σ1(v1,p)\Us, we take Z1 to be the closure of M st and Z3 to be the closed stratum MX1,σ1(v1,p). We consider V = MX1,σ1(v1)\(Z1 ∪ Z2 ∪ Z3). Then it is easy to see that V is an open subset of MX1,σ1(v1). Moreover, V is the union of U and an open subset of Us, hence has a complement of codimension at least two, and the restriction fV is an injective morphism fV : V → MX2,σ2(v2). Since f is induced by a derived (anti-)equivalence Φ, we see that Φ−1 defines an inverse of f on f (V ), and therefore f is an isomorphism from V to its image f (V ). Note that since V has no intersection with the closed stratum MX1,σ1(v1,p), f (V ) also has no intersection with the closed stratum MX2,σ2(v2,p). 10 CIARAN MEACHAN AND ZIYU ZHANG 1 (V ) → π−1 1 (V ) → V and π2 : π−1 For i = 1, 2, we write πi : fMi → MXi,σi(vi) for the symplectic resolution constructed in [O'G99, LS06]. The construction there implies that both π1 : π−1 2 (f (V )) → f (V ) are exactly the blowups of the singular loci. Therefore, the isomorphism f : V → f (V ) induces another isomorphism ef : π−1 2 (f (V )). In particular, it is a birational map ef : fM1 99K fM2. 1 (V ) ⊂ fM1 is an open subset with complement of codimension at least two. We claim that π−1 On one hand, from the construction of V we observe that the complement of V in M st (v1) has codimension at least two. Together with the fact that π1 : π−1 (v1)) → M st (v1) is an isomorphism, we conclude that the complement of π−1 (v1)) also has codi- mension at least two. On the other hand, since V contains an open subset of the singular locus Sym2 MX1,σ1(v1,p), we obtain that π−1 1 (V ) contains an open subset of the unique exceptional divisor. Therefore π−1 1 (V ) has a complement of codimension two in fM1. Now we can apply [GHJ03, Proposition 21.6] to the birational map ef : fM1 99K fM2, and conclude 1 (V )) ⊂ fM2 has complement of codimension at least two. This further that π−1 implies f (V ) ⊂ MX2,σ2(v2) also has complement of codimension at least two. Therefore f (U ), as the intersection of f (V ) with the open stratum M st (v2), has complement of codimension at least two as well. (cid:3) 2 (f (V )) = ef (π−1 1 (V ) in π−1 X1,σ1 1 (M st X1,σ1 1 (M st X1,σ1 X1,σ1 X2,σ2 alg(X, Z) and NS(Mσ(v)). 2.3. Mukai morphisms are isomorphisms. The Mukai morphism plays an essential role in [BM14a]. A classical theorem [BM14a, Theorem 3.6], originally proved in [Muk87, Yos01], shows that for a smooth moduli space Mσ(v) of stable objects on a K3 surface X with v2 > 0, the Mukai morphism induced by a (quasi-)universal family is in fact a Hodge isometry between the orthogonal complement of v in the total cohomology v⊥,tr ⊂ H ∗(X, Z) and H 2(Mσ(v), Z). By restricting on the algebraic components on both sides, we get an isometry between the orthogonal complement of v in the algebraic cohomology v⊥ ⊂ H ∗ Perego and Rapagnetta generalised this classical result to generic Gieseker moduli spaces of O'Grady type in [PR13, Theorem 1.7]. Here we will follow their approach to generalise the same result further to generic Bridgeland moduli spaces of O'Grady type. This is the content of the following Theorem 2.7. Let v be a Mukai vector of O'Grady type and σ ∈ Stab†(X) be a generic stability condition with respect to v. Let M = Mσ(v) be the moduli space of σ-semistable objects of class v on X and π : fM → M be its symplectic resolution. Then we have (mixed) Hodge structures. In particular, the Hodge structure on H 2(M, Z) is pure of (1) The pullback map π∗ : H 2(M, Z) → H 2(fM , Z) is injective and compatible with the weight two and the restriction of the Beauville-Bogomolov quadratic form eq(−,−) on H 2(fM , Z) defines a quadratic form q(−,−) on H 2(M, Z); σ : v⊥,tr → H 2(M, Z) induced by the (2) There exists a well-defined Mukai morphism θtr (quasi-)universal family over M st σ (v), which is a Hodge isometry. The proof of Theorem 2.7 is postponed to the next section. We continue our discussion with the following interesting consequence, which will be very important later. Corollary 2.8. Under the assumptions of the Theorem 2.7, we have (1) Lefschetz (1, 1) theorem holds for M . That is, NS(M ) = H 1,1(M, Z); BIRATIONAL GEOMETRY OF SINGULAR MODULI SPACES OF O'GRADY TYPE 11 (2) The restriction of the pullback map π on NS(X) is an injective map π∗ : NS(M ) → NS(fM ), which is compatible with the Beauville-Bogomolov pairings q(−,−) and eq(−,−); σ on the algebraic Mukai lattice is an isometry θσ : v⊥ → NS(M ). In particular, q(−,−) is a non-degenerate pairing on NS(M ) with signature (1, ρ(X)). (3) The restriction of the Mukai morphism θtr Proof. For simplicity, we still denote Mσ(v) by M . By Theorem 2.7, the Hodge structure on H 2(M, Z) is pure of weight two, and π∗ preserves the Hodge structure. Therefore, for any class generic fibre of π within the exceptional divisor is a smooth rational curve. Let C be such a α ∈ H 1,1(M, Z), we have π∗α ∈ H 1,1(fM , Z) and hence π∗α = c1(eL) for some line bundle eL on fM . By O'Grady's construction of the symplectic resolution fM in [O'G99], we know that a rational curve, then eL · C = π∗α · [C] = α · π∗[C] = 0, which implies that the restriction of eL on we must have eL = π∗L for some line bundle L on M . Therefore we have α = c1(L) ∈ NS(M ) C is trivial. Since M is normal by [BM14a, Theorem 3.10] (or originally [KLS06, Theorem 4.4]), and the Lefschetz (1, 1) theorem is true for moduli spaces M of O'Grady type. By taking the (1, 1) components on both sides of the map π∗ between the second cohomology groups, we get the map between the N´eron-Severi lattices. Similarly, we can take the (1, 1) components on both sides of the Hodge isometry θtr (cid:3) σ to get the desired isometry θσ. Remark 2.9. We briefly describe how the map θtr 2.7. Indeed, we will see that it is the unique lift of ΦE along i∗ in the diagram σ will be constructed in the proof of Theorem v⊥,tr H 2(Mσ(v), Z) θtr σ /❴❴❴❴❴❴ )❙❙❙❙❙❙❙❙❙❙❙❙❙❙❙ ΦE i∗ H 2(M st σ (v), Q), where ΦE is the classical Mukai morphism induced by the (quasi-)universal family E on the σ (v) of the moduli space, and i∗ is the pullback along an open embedding. The stable locus M st map θσ, which is the restriction of θtr σ , is sometimes also referred to as the Mukai morphism in the literature. However, we prefer to call it the algebraic Mukai morphism, to distinguish it from the Mukai morphism θtr σ , which includes the (2, 0) and (0, 2) components on both sides. Remark 2.10. We would also like to point out that, although θtr σ and θσ a priori depend on the choice of the generic stability condition σ, they in fact only depend on the choice of the open chamber C ∈ Stab†(X) containing σ. This is because the moduli space is the same for all interior points of C. Therefore, in [BM14b, BM14a], the Mukai morphism and algebraic Mukai morphism are sometimes also denoted by θtr C and θC respectively. We conclude this section by briefly mentioning various cones of divisors on M = Mσ(v). The ample cone Amp(M ), big cone Big(M ) and movable cone Mov(M ) are all well-defined. Due to the existence of the symplectic resolution and Corollary 2.8, we can also define the positive cone of M . We will show that the following definition justifies the name. Definition 2.11. The cone (π∗)−1(Pos(fM )) ⊂ NS(M ) is called the positive cone of M , and is denoted by Pos(M ). We can see from the following proposition that the notion is reasonably defined and agrees with our intuition. / ) 12 CIARAN MEACHAN AND ZIYU ZHANG Proposition 2.12. The positive cone Pos(M ) is one of the two components of {α ∈ NS(M ) : q(α, α) > 0} and contains the ample cone Amp(M ). Proof. In fact, by Corollary 2.8(3), we know that the cone {α ∈ NS(M ) : α2 > 0} has two components. Together with the map π∗ in Corollary 2.8(2), we know they are precisely the restrictions of the two components of {eα ∈ NS(fM ) : eq(eα,eα) > 0} to NS(M ), one of which is Pos(fM ). This proves the first statement. any ample class α ∈ Amp(M ) and note that eα := π∗α is nef and big on fM . By [KM98, Propo- sition 2.61], this implies R fM eα10 > 0. Now the Beauville-Fujiki relation [GHJ03, Proposition 23.14] implies eq(eα,eα) > 0. Thus we have eα ∈ Pos(fM ) and hence α ∈ Pos(M ). Remark 2.13. A priori, the ample cone Amp(M ) could be empty. However, it is proved in [BM14b, Theorem 1.3] that M always carries ample classes. Therefore, between the two com- ponents of the cone of square positive classes on M , Pos(M ) can be simply identified as the one which contains Amp(M ). If there is no ample class then there is nothing to prove for the second statement. Otherwise, take (cid:3) 3. Proof of Theorem 2.7 This whole section is devoted to the proof of Theorem 2.7. We start with some lemmas about deformations of twisted K3 surfaces and local existence of relative moduli spaces of twisted sheaves. The proof of Theorem 2.7 will follow after these preparations. As in [BM14b, BM14a], when talking about twisted sheaves, we always assume that we have a fixed B-field lift of the Brauer class, which was introduced in [HS05]. 3.1. Deformations of twisted polarised K3 surfaces. In this subsection we study defor- mations of a polarised K3 surface which carry a non-trivial Brauer class with a B-field lift. Our main results here are Proposition 3.7 and Proposition 3.9. Roughly speaking, up to changing the B-field by an integral class, we can always deform a twisted polarised K3 surface to an untwisted polarised K3 surface in the period domain. Moreover, if there is a Mukai vector on the initial K3 surface which remains algebraic under deformations, and the initial polarisation is generic with respect to this Mukai vector, then the deformation can be made so that the polarisation is generic with respect to the corresponding Mukai vector on each fibre. Moreover, Lemma 3.6 shows how to find such an integral class so as to amend a given B-field and make the above deformations possible. We briefly recall the necessary notions required for the following discussion. The (cohomological) Brauer group of a K3 surface X is the torsion part of the cohomology group H 2(X,O∗ X ) in the analytic (or ´etale) topology. A twisted K3 surface is a K3 surface X equipped with a Brauer class α. Using the exponential sequence, we can always find a rational class B ∈ H 2(X, Q), such that its (0, 2)-component maps to α under the exponential map, i.e. exp(B0,2) = α. We call such a rational class B a rational B-field lift of the Brauer class α. Note that the B-field lift of any given Brauer class α is not unique. For every α-twisted sheaf E, a twisted Chern character of E, and hence a twisted Mukai vector of E, is defined in [HS05, Proposition 1.2], which depends on the choice of the B-field lift B of α. The construction there guarantees that it is a B-twisted algebraic class, i.e. a class in the alg(X, Q)) ∩ H ∗(X, Z), as B-twisted algebraic cohomology group H ∗ defined in [HS05, Remark 1.3]. alg(X, B, Z) := (exp(B) · H ∗ BIRATIONAL GEOMETRY OF SINGULAR MODULI SPACES OF O'GRADY TYPE 13 We introduce the following notion for simplicity of presentation: assume we have a family of K3 surfaces X → S, and each fibre Xs over the point s ∈ S is equipped with a B-field Bs ∈ H 2(Xs, Q). If these B-fields form a section of the local system over S with fibres given by H 2(Xs, Q), then we say the family of B-fields is locally constant over S. The following lemma shows the locally trivial extension of some cohomology classes could remain algebraic on each fibre of a deformation. Lemma 3.1. Let (X, H) be a polarised K3 surface and B a rational B-field lift of a certain Brauer class on X. Let v ∈ H ∗(X, Z) be of the form v = (r, qh + rB, a) for some positive integer r, rational number q and integer a, where h = c1(H). Then v ∈ H ∗ alg(X, B, Z), i.e. v is a B-twisted algebraic class. Moreover, for any flat deformation of the polarised K3 surface (X, H) with locally constant B- fields extending B on X, v also extends to a locally constant section, such that we get a twisted algebraic class on each fibre. Proof. A simple computation shows the component of exp(−B) · v in H 2(X, Q) is qh, which is a (1, 1)-class. Hence we conclude that v ∈ H ∗ alg(X, B, Z) is an integral B-twisted algebraic class on X. Assume we have a deformation over S, such that for every s ∈ S, the K3 surface Xs comes with an ample line bundle Hs and a B-field Bs, which are both locally constant classes. Then the locally trivial extension of v over each fibre Xs is given by vs = (r, qhs + rBs, a) ∈ H ∗(Xs, Z) where hs = c1(Hs) ∈ H 2(Xs, Z). The same computation shows that it is a Bs-twisted algebraic class. (cid:3) The above lemma leads to the following definition. Definition 3.2. A class v ∈ H ∗(X, Z) satisfying the assumption of Lemma 3.1 is called a deformable B-twisted Mukai vector on X. The following lemma justifies the universality of this notion. Lemma 3.3. Let (X, H) be a polarised K3 surface, B a rational B-field, and v a B-twisted Mukai vector with its degree zero component r > 0. If H is generic with respect to v, then the moduli space MX,B,H(v) of B-twisted H-semistable sheaves with Mukai vector v is always isomorphic to a moduli space MX,B,H ′(v′) where v′ is deformable, and H ′ is generic with respect to v′. Proof. This is also classical (see, for instance, the proof of [HL10, Theorem 6.2.5]). We write v = (r, c, a) and c1(H) = h. By Lemma 3.14, we can replace the Mukai vector v by v′ = v· exp(mh) for any m ∈ Z, without changing the moduli space. Moreover, H is still generic with respect to v′. Note that v′ is B-twisted, therefore the degree two component of v′ · exp(−B) is a rational (1, 1)-class. A simple calculation shows that this class is c + rmh− rB. When m ≫ 0, c + rmh − rB is an ample class and lies in the same chamber as h. We fix such an m and write h′ for the primitive integral class on the ray generated by c + rmh − rB in the ample cone. We denote the corresponding ample line bundle H ′, and write c + rmh − rB = qh′ for some q ∈ Q. Then v′ = (r, qh′ + rB, a′) for some a′ ∈ Z is a deformable Mukai vector on the polarised K3 surface (X, H ′). The construction guarantees that H ′ lies in the interior of a chamber, and hence is generic. (cid:3) We recall the following fact from linear algebra, which will be used in the proof of Lemma 3.6. 14 CIARAN MEACHAN AND ZIYU ZHANG Lemma 3.4. Let V be a real vector space equipped with a (possibly degenerate) real-valued symmetric bilinear pairing, whose signature is (n+, n−, n0) (for the positive definite, negative definite, and isotropic parts respectively). Let V ′ be a codimension one linear subspace of V equipped with induced pairing, with signature (n′ + 6 n+, n− − 1 6 n′ The same statement holds for rational vector spaces with rational-valued pairings. 0). Then we have n+ − 1 6 n′ − 6 n−, and n0 − 1 6 n′ +, n′ −, n′ 0 6 n0 + 1. Proof. The proof is completely elementary by looking at the symmetric matrix representing the symmetric bilinear pairing. We leave it to the reader. (cid:3) Remark 3.5. If the symmetric bilinear pairing is non-degenerate, i.e. n0 = 0, then we will also write its signature as (n+, n−) by abuse of notation. We are now ready to prove a technical lemma, which will be used to deal with the subtle potential issues of ampleness caused by (−2)-classes in the deformation, and the genericness of the polarisations. We point out that, despite of the words appearing in the statement, this lemma has nothing to do with ampleness of h or B being a B-field. We state it in this way just to indicate the situation in which we apply it. The proof of this lemma only contains elementary lattice theoretic arguments. Lemma 3.6. Let h ∈ H 2(X, Z) be an ample class, and B ∈ H 2(X, Q) be a B-field. Fix an arbitrary positive integer N0. Then there exists B′ ∈ B + H 2(X, Z) such that B′2 < 0, and for every non-zero class g ∈ SpanQ{h, B′} ∩ H 2(X, Z) ∩ h⊥, we have g2 < −N0. Proof. Without loss of generality, we can assume h to be a primitive class. Otherwise we can replace h by the primitive class in the ray generated by h, which is still an ample class. Moreover, let n by the smallest positive integer such that nB ∈ H 2(X, Z). Note that n is in fact the order of the Brauer class represented by B. Since h is primitive, the exact sequence of lattices 0 → Zh → H 2(X, Z) → H 2(X, Z)/Zh → 0 splits non-canonically. If we choose a lift of H 2(X, Z)/Zh in the ambient lattice H 2(X, Z), say M , then H 2(X, Z) = Zh ⊕ M . We can write nB = αBh + βBmB under this decomposition, where βB is a non-negative integer and mB ∈ M is a primitive class. Moreover, it is easy to see that, for any primitive class pM ∈ M , the lattice SpanQ{h, pM}∩ H 2(X, Z) has an integral basis given by {h, pM}. We do the same thing for the second time. Since mB is primitive, the exact sequence of lattices 0 → ZmB → M → M/ZmB → 0 splits, and we can choose a lift L of M/ZmB in M and write M = ZmB ⊕ L. Again, for any primitive class pL ∈ L, the lattice SpanQ{mB, pL}∩ M has an integral basis given by {mB, pL}. In particular, the divisibility of any integral linear combination γ1mB + γ2pL is the highest common factor of γ1 and γ2. Now we take K := {h, mB}⊥ ∩ L be the sublattice of H 2(X, Z) containing all lattice points in L which are perpendicular to both h and mB under the Poincar´e pairing. In particular, these classes are perpendicular to nB. Note that K ⊂ L has corank at most 2, therefore K ⊂ H 2(X, Z) has corank at most 4. Since the Poincar´e pairing on H 2(X, Z) has signature (3, 19) (see e.g. [BHPVdV04, Proposition VIII.3.2]), we can use Lemma 3.4 repeatedly to see that the restriction of the Poincar´e pairing to K has negative signature at least 15. That is, K contains a negative BIRATIONAL GEOMETRY OF SINGULAR MODULI SPACES OF O'GRADY TYPE 15 definite sublattice of rank at least 15. Therefore, for any N ≫ 0, we can find a primitive integral class c ∈ K such that c2 < −N . Now let us consider the rank two lattice ΛN = SpanQ{h, c + B} ∩ H 2(X, Z). We will show that for N sufficiently large, it does not contain any (−2)-classes which are perpendicular to h. We first try to find an integral basis for ΛN . Note that nc + nB = αBh + (nc + βBmB) where nc + βBmB ∈ M . Since c ∈ L is primitive, the divisibility kN of the class nc + βBmB is the highest common factor of n and βB. In particular, 1 6 kN 6 n. We write nc + βBmB = kN pN where pN is a primitive class in M . By the discussion above, this shows that {h, pN} is an integral basis for ΛN . Therefore, we just need to show that for each pair of integers (x, y) satisfying h · (xh + ypN ) = 0, we have (xh + ypN )2 < −N0 for any given N0. We give an estimate of the intersection matrix of the lattice ΛN with integral basis {h, pN}. Since h is a fixed class, h2 is a fixed number. And we have h · pN = h · 1 kN (nc + βBmB) = βB kN (h · mB). Since 1 6 kN 6 n, we see that h · pN 6 βB(h · mB) where the right hand side is independent of N . Moreover p2 N = 1 k2 N (nc + βBmB)2 = 1 k2 N (n2c2 + β2 Bm2 B). Since 1 6 kN 6 n and c2 < −N , we see that when N ≫ 0 is large and positive, p2 and negative. N ≪ 0 is large Now assume we have a non-zero vector xh + ypN ∈ ΛN with h · (xh + ypN ) = 0. Then we can assume y 6= 0 since otherwise the vector is zero. Thus we have h2 From the above discussion, we see that (h·pN )2 N ≪ 0 when N ≫ 0. Since y2 is a positive integer, the above computation shows that (xh + ypN )2 ≪ 0, hence is smaller than any number −N0 we started with. Therefore, in the statement of the proposition, we can simply choose B′ = c + B. Note that the choice of c makes c · B = 0, hence we also have B′2 = c2 + B2 ≪ 0 for a sufficiently large choice of N . is bounded for all choices of N , while p2 (cid:3) h2 The following proposition shows how to deform a twisted polarised K3 surface to an untwisted polarised K3 surface. Proposition 3.7. Let (X, α) be a twisted projective K3 surface with an ample line bundle H, and B ∈ H 2(X, Q) be a B-field lift of α. We denote h = c1(H) and assume that B2 < 0 and the lattice SpanQ{h, B} ∩ H 2(X, Z) ∩ h⊥ does not contain any (−2)-classes. Then there exists a family of twisted projective K3 surfaces (X ,H) → S over an integral base S, with ample line bundle Hs and locally constant B-field Bs on the fibre Xs of the family at each point s ∈ S, such that at s0 ∈ S, we have (Xs0, Hs0, Bs0) = (X, H, B), and at s1 ∈ S, we have Bs1 ∈ H 1,1(Xs1, Q). In particular, it represents a trivial Brauer class on Xs1. (xh + ypN )2 = ypN · (xh + ypN ) = y(xh · pN + yp2 N ) (h · pN )2 (h · pN )2 + y(p2 )) N − h2 = y(xh · pN + y = y2(p2 h2 (h · pN )2 ). N − 16 CIARAN MEACHAN AND ZIYU ZHANG Proof. We can choose S to be a single point if B is of type (1, 1). Otherwise, we study the period domain for the deformation of polarised K3 surfaces. Consider the lattice Λ = {h}⊥ ⊂ H 2(X, Z), where h = c1(H) ∈ H 2(X, Z). The associated period domain is: D = {σ ∈ P(Λ ⊗ C) : σ2 = 0, σσ > 0}. The period domain of marked polarised K3 surfaces is obtained by removing hyperplanes of the form δ⊥ from D for all (−2)-classes δ. Since Λ has signature (2, 19), D has two connected com- ponents, which can be identified by complex conjugation. We denote (the marking) [H 2,0(X)] by σ0 and observe that σ0 ∈ D. In order to deform X so that the B-field becomes of type (1, 1), we consider Λ′ = {h, B}⊥ ⊂ H 2(X, Z) and set D′ = {σ ∈ P(Λ′ ⊗ C) : σ2 = 0, σσ > 0}. Notice that D′ is in fact a hyperplane section of D. The condition B2 < 0 guarantees that Λ′ has signature (2, 18). Therefore, D′ is non-empty and has two components, which are contained in the two components of D respectively. The assumption also guarantees that D′ is not contained in the hyperplane δ⊥ for any (−2)-class δ ∈ Λ. Therefore any period point in D′ away from these hyperplanes represents a deformation of X on which B has type (1, 1). Let S ⊂ D be any integral curve joining such a period point and σ0. It parametrises a family of projective K3 surfaces after removing at most locally finitely many closed points. (cid:3) Next we deal with the issue of the genericness of polarisations. We remind the reader that for any (twisted) Mukai vector v on a K3 surface with rank r, its discriminant is defined as ∆(v) = v2 + 2r2; see [HL10, Section 3.4] or [Lie07, Definition 3.2.1.1]. Lemma 3.8. Let (X, H) be a polarised K3 surface with a rational B-field B. Let v be a B- twisted algebraic class with its degree zero component r > 0. If there is no class ξ ∈ H 1,1(X, Z) satisfying both ξ · H = 0 and − r2 4 ∆(v) 6 ξ2 < 0, then H is generic with respect to v. Proof. The proof is contained in [HL10, Theorem 4.C.3], and so we only point out what change has to be made to adapt it in the twisted case. Using the notation there, ξ = r·cB 1 (F ) is a class in H 1,1(X, B, Z). However, we realise that cB 1 (F ) − rB is the (1, 1)-component of 1 (F ′) − r′B is also a class in H 1,1(X, Q). exp(−B)v, hence is a class in H 1,1(X, Q). Similarly, cB 1 (F ′) − r′B) − r′ · (cB We note that a different way to write ξ is ξ = r · (cB 1 (F ) − rB), hence ξ is in fact also an untwisted (1, 1) class. So we only need to check classes in H 1,1(X, Z) instead of H 1,1(X, B, Z). On the other hand, we realise that the Bogomolov inequality also holds in twisted case by [Lie07, Proposition 3.2.3.13]. Therefore by the same argument as in the proof of [HL10, Theorem 4.C.3], if H is not v-generic, then there is such a ξ satisfying both ξ · H = 0 and − r2 4 ∆(v) 6 ξ2 < 0. The proof is finished by contradiction. 1 (F ′)−r′·cB (cid:3) The following proposition shows how to keep the genericness of polarisations with respect to a locally constant section of deformable Mukai vectors. Proposition 3.9. Assume we are in the situation of Proposition 3.7. We also assume that v = (r, qh + rB, a) is a deformable Mukai vector with v2 > 0, and H is generic with respect to v. Furthermore suppose that for each non-zero class g ∈ SpanQ{h, B} ∩ H 2(X, Z) ∩ h⊥ we have g2 < − r2 4 ∆(v). Then, in the family constructed in Proposition 3.7, v extends to a locally constant family of deformable Mukai vectors over S with its value vs at the point s ∈ S. Moreover, for every s ∈ S, Hs is generic with respect to vs. BIRATIONAL GEOMETRY OF SINGULAR MODULI SPACES OF O'GRADY TYPE 17 Proof. By Lemma 3.1, it is clear that v can always be extended to a locally constant family of deformable Mukai vectors as long as h and B extend. Therefore, we just need to show that, for a generic period point σ ∈ D′ as constructed in the proof of Proposition 3.7, the class h is generic with respect to the class v on the projective K3 surface corresponding to σ. We use the criterion in Lemma 3.8 to determine the genericness. That is, in the period domain D constructed in the proof of Lemma 3.1, we need to remove hyperplane sections of the form ξ⊥ for every class ξ satisfying both ξ · H = 0 and − r2 4 ∆(v) 6 ξ2 < 0, so that every period point in the remaining part corresponds to a projective K3 surface on which the ample line bundle corresponding to the class h is generic with respect to the locally constant extension of v. By [PR13, Corollary 4.2], we conclude that these hyperplane sections to be removed from are locally finite in D. However, we also have to show that D′ is not contained in any of these hyperplanes, so that we can still find period points to make the B-field have type (1, 1) after In fact, if D′ ⊂ ξ⊥ for some ξ ∈ H 2(X, Z), then ξ ∈ removing these hyperplane sections. SpanQ{h, B} ∩ H 2(X, Z). By assumption, if such a class satisfies further ξ · h = 0, then g2 < − r2 4 ∆(v). Hence ξ⊥ is not one of the hyperplane sections to be removed. (cid:3) 3.2. Local existence of relative twisted moduli spaces and compatibility. In this sub- section, we show that for a family of polarised K3 surfaces with locally constant B-fields and deformable Mukai vectors, we can always construct relative moduli spaces of twisted sheaves in an (´etale or analytic) neighbourhood of an arbitrary point. Moreover, this relative moduli space carries a (quasi-)universal family on its smooth locus. However, this construction would involve a choice of a uniform Cech 2-cocycle representation of the B-fields over that neighbour- hood, hence there is in general no way to glue these (quasi-)universal families together. We will nevertheless show that on any fixed polarised K3 surface, for different choices of the B-field lift of the same Mukai vector, or for different choices of the Cech 2-cocycle representations of the same B-field, the moduli spaces are always non-canonically isomorphic to one another. As we will see later, this is already sufficient for the property of Mukai morphisms being isometries to hold or fail simultaneously over the entire base (if connected). The materials presented in this subsection are probably well-known to experts. However we couldn't find any reference in which these results are explicitly written down. Hence we state these results with full proofs as below. The first result proves the local existence of relative moduli spaces of semistable sheaves.2 Proposition 3.10. Let (X ,H) → S be a family of projective K3 surface equipped with a lo- cally trivial family of B-fields, denoted by Bs on each fibre Xs. We also write hs = c1(Hs) ∈ H 2(Xs, Z). Let vs = (r, qhs + rBs, a) ∈ H ∗(Xs, Z) be a locally constant family of deformable Mukai vectors with v2 s > 0. Then for each point s0 ∈ S, there is an (´etale or analytic) open neighbourhood Us0 ⊂ S, over which there is a relative moduli space M, whose fibre Ms at any point s ∈ Us0 is the moduli space of Bs-twisted Hs-semistable sheaves of Bs-twisted Mukai vector vs, with a (quasi-)universal family on the smooth locus of M. Proof. We first show that, for every fixed s0 ∈ S, over an open neighbourhood of s0, the B-fields have a "uniform" Cech 2-cocycle representation. We write p : X → S for the family, and start with finitely many (´etale or analytic) open subsets of the total space X , which cover the fibre Xs0. We label them by {Vi} for i in a finite index set. The union ∪iVi is an open subset of the 2We thank K¯ota Yoshioka for informing us that he was able to obtain a global relative twisted moduli space without universal families a long time ago, in the category of algebraic (or analytic) spaces by gluing along an ´etale (or analytic) cover of the base. 18 CIARAN MEACHAN AND ZIYU ZHANG total space X containing the fibre Xs0. Therefore p(X\∪i Vi) is a closed subset of S missing the point s0. We denote S\p(X\ ∪i Vi) by Us0, then for every s ∈ Us0, {Vi ∩ Xs} is a finite (´etale or analytic) cover of the fibre Xs. Using these open covers, we realise that the Cech complex ⊕iΓ(Vi, Z) → ⊕i,jΓ(Vi ∩ Vj, Z) → ⊕i,j,kΓ(Vi ∩ Vj ∩ Vk, Z) → ··· computes the integral cohomology H ∗(Xs, Z) for all s ∈ Us0 simultaneously. In fact, its restric- tion on each fibre Xs for s ∈ Us0 is precisely the Cech complex on that fibre. In other words, any Cech 2-cocycle representation of the B-field Bs0 on a single fibre Xs0 using the open cover {Vi ∩ Xs} can be canonically extended over Us0, so that the extension simultaneously represents the B-fields Bs for all s ∈ Bs0. From this point on, everything becomes classical. We can write down a moduli functor for families of twisted semistable sheaves on fibres of p−1(Us0) → Us0. Note that for each such family we have a global Cech 2-cocycle representation of the locally constant B-field. The GIT construction of the relative moduli space in the untwisted situation (see, for instance, [HL10, Theorem 4.3.7]) applies in the twisted situation and gives a relative moduli space over Us0, whose fibre over each point s ∈ Us0 is the moduli space of Bs-twisted Hs-semistable sheaves of class vs over Xs. And a (quasi-)universal family exists on the smooth locus of the relative moduli space. (cid:3) Remark 3.11. For the construction of the associated family of Azumaya algebras and the connec- tion with Simpson's stability, we refer the reader to [Yos06] and [Sim94]. Note that the relative moduli space constructed in Proposition 3.10 is in general not smooth over the parameter space. In particular, in the case of relative moduli spaces of O'Grady type, which is the most interesting case in this paper, the relative moduli space is not smooth over the parameter space. The following three lemmas handle the compatibility issues caused by change of Cech 2-cocycle representation of a fixed B-field, change of the B-field lift of a fixed Brauer class, and change of the Mukai vector by tensoring with multiples of the polarisation. Lemma 3.12. Let (X, H) be a polarised K3 surface, B be a rational B-field on X, and v be a B-twisted Mukai vector of O'Grady type. Let B1 and B2 be two different Cech 2-cocycle representations of the same B-field B, and denote the moduli space of B1-twisted (or B2-twisted) H-semistable sheaves with twisted Mukai vector v by MX,B1,H(v) (or MX,B2,H(v)). Then we can choose an isomorphism f : MX,B1,H(v) → MX,B2,H(v), such that the Mukai morphisms θtr B1,H and θtr B2,H are identified. More precisely, the following diagram commutes v⊥,tr v⊥,tr θtr B1 ,H θtr B2 ,H / H 2(MX,B1,H(v), Z) f ∗ ∼= / H 2(MX,B2,H(v), Z). (3.1) In other words, the Mukai morphism is independent of the choice of the Cech 2-cocycle rep- B2,H is an resentation of the B-field B. isometry. B1,H is an isometry if and only if θtr In particular, θtr Proof. This is probably well-known, but we nevertheless give a proof. We first show that there is an equivalence of the abelian categories F : Coh(X, B1) → Coh(X, B2). Without loss of generality, we can assume that the open cover underlying the two Cech 2-cocycle representations B1 and B2 are the same. Otherwise, we can always find a common refinement, and restriction is / / O O BIRATIONAL GEOMETRY OF SINGULAR MODULI SPACES OF O'GRADY TYPE 19 ijk ∈ Γ(Uijk, Q)} (or {B2 a canonical way to obtain transition functions for the refined open cover from B1 (or B2). Now we assume {Ui} is the open cover for both B1 and B2, then B1 (or B2) can be represented by {B1 ijk ∈ Γ(Uijk, Q)} respectively). Since they represent the same class ijk − B1 in H 2(X, Q), the difference {B2 ijk ∈ Γ(Uijk, Q)} is a coboundary. That is, there exists a ijk − B1 1-cocycle {γij ∈ Γ(Uij, Q)}, such that B2 With the above preparation, we can construct the functor F in the following way: for every E ∈ Coh(X, B1) given by {Ei, ϕij}, where ϕij ∈ Γ(Uij,O∗), we define F (E) to be {Ei, ϕij ·exp(γij)}. A direct computation shows that F (E) ∈ Coh(X, B2). Note that this operation is also invertible and produces F −1. Therefore F : Coh(X, B1) → Coh(X, B2) is an equivalence of categories (in fact in our case both compositions of F and F −1 are the identity functors). It is also easy to see that F preserves the abelian structures on the two categories. ijk = γij + γjk + γki. We also observe that the functor F preserves twisted Chern characters. For this we follow the notations and arguments in proof of [HS05, Proposition 1.2]. Note that the B1-twisted Chern character of E is defined to be the (untwisted) Chern character of (the C∞ sheaf) EB1 = {Ei, ϕ′ ij = ϕij · exp(aij)} for some C∞-coboundary representation {−aij} of B1. We realise that {−aij + γij} becomes a C∞-coboundary representation of B2. Therefore, the B2-twisted Chern character of F (E) can be defined to be the (untwisted) Chern character of (the C∞ sheaf) F (E)B2 = {Ei, ϕ′′ ij = ϕij · exp(γij)· exp(aij − γij) = ϕij · exp(aij). Therefore EB1 and F (E)B2 are in fact the same sheaf, and the twisted Chern characters chB1 (F (E)). In other words, the functor F preserves twisted Chern characters. ij}, where ϕ′′ (E) = chB2 Since F preserves twisted Chern characters, it also preserves (twisted) Euler characteristics and (twisted) Hilbert polynomials, and therefore H-stability. It is also straightforward to see that the construction of F can be carried out in families of B1-twisted sheaves. In particular, F induces a natural transformation between the moduli functors for B1- and B2-twisted H- semistable sheaves of Mukai vector v, and hence induces an isomorphism on the moduli spaces f : MX,B1,H(v) → MX,B2,H(v). Moreover, if E is a (quasi-)universal family of B1-twisted H- stable sheaves of Mukai vector v on the smooth locus of the moduli space M st X,B1,H(v), then F (E) is a (quasi-)universal family of B2-twisted H-stable sheaves of Mukai vector v on the smooth locus of the moduli space M st X,B2,H(v). The above argument shows that the two families E and F (E) have the same twisted Chern character. In particular, the cohomological integral functors ΦE and ΦF (E) given by the two families are the same. More precisely, we have the commutative diagram v⊥,tr ΦE / H 2(M st X,B1,H(v), Q) (3.2) f ∗ ∼= v⊥,tr ΦF (E) / H 2(M st X,B2,H(v), Q). By the construction of the Mukai morphism in [PR13, Section 3], we know that ΦE (resp. ΦF (E)) uniquely determines θtr B2,H), and the commutativity of (3.2) implies the commutativity of (3.1) (see also [PR13, Lemma 3.11]). Therefore θtr B1,H is an isomorphism of lattices if and only if θtr B1,H is an isometry if and only if θtr (cid:3) B2,H is an isomorphism of lattices, and by [PR13, Lemma 3.12], θtr B1,H (resp. θtr B2,H is an isometry. Lemma 3.13. Let (X, H) be a polarised K3 surface, B be a rational B-field on X, and v be a B-twisted Mukai vector of O'Grady type. Pick any c ∈ H 2(X, Q) which is a B-field lift of the zero Brauer class. We denote the moduli space of B-twisted (resp. (B + c)-twisted) H-semistable / / O O 20 CIARAN MEACHAN AND ZIYU ZHANG sheaves with twisted Mukai vector v by MX,B,H (v) (resp. MX,B+c,H(v)). Then we can choose an isomorphism f : MX,B,H(v) → MX,B+c,H(v), such that the Mukai morphisms θtr are identified. More precisely, the following diagram commutes B,H and θtr B+c,H v⊥,tr v⊥,tr θtr B,H / H 2(MX,B,H (v), Z) f ∗ ∼= / H 2(MX,B+c,H (v), Z). θtr B+c,H (3.3) In other words, the Mukai morphism is independent of the choice of the B-field lift of the Brauer class up to a rational (1, 1) class. In particular, θtr B+c,H is an isometry. B,H is an isometry if and only if θtr We point out that the condition c is a B-field lift for the zero Brauer class contains many interesting special cases. The cases that we need are c ∈ H 2(X, Z) or c ∈ H 1,1(X, Z). Proof. We follow the same idea as in proof of Lemma 3.12. We fix an open cover {Ui} on which both B and c has a Cech 2-cocycle representation, say B = {Bijk ∈ Γ(Uijk, Q)} and c = {cijk ∈ Γ(Uijk, Q)}. Since c lifts the zero Brauer class, we have exp(c) = 0 ∈ H 2(X,O∗). In other words, we can find a 1-cocycle {γij ∈ Γ(Uij,O∗)}, such that exp(cijk) = γij · γjk · γki. We define the functor F : Coh(X, B) → Coh(X, B + c) as follows: for every E ∈ Coh(X, B) which can be represented by E = {Ei, ϕij}, we define F (E) = {Ei, ϕij · γij}. It is easy to check that F (E) ∈ Coh(X, B + c). It is easy to see that F is an equivalence of categories, and preserves the abelian structure. We also observe that the same construction can be done in families. However, similar to the computation in the proof of Lemma 3.12, we observe that chB(E) = chB+c(F (E)). That is, F preserves twisted Chern character and hence preserves the twisted Eu- ler characteristic and H-stability. This implies that F induces an isomorphism of moduli spaces f : MX,B,H(v) → MX,B+c,H(v). Moreover, if E is a (quasi-)universal family over M st X,B,H(v), then F (E) is a (quasi-)universal family over M st In particular, the cohomological integral functors ΦE and ΦF (E) given by the two families are the same. Hence the following diagram commutes X,B+c,H(v), with chB(E) = chB+c(F (E)). v⊥,tr ΦE / H 2(M st X,B,H (v), Q) (3.4) f ∗ ∼= v⊥,tr ΦF (E) / H 2(M st X,B+c,H (v), Q). Using a similar reasoning as in the proof of Lemma 3.12, we conclude that the commutativity of (3.4) implies the commutativity of (3.3). In particular, θtr is an isometry. B,H is an isometry if and only if θtr B+c,H (cid:3) Lemma 3.14. Let (X, H) be a polarised K3 surface, B be a rational B-field on X, and v be a B-twisted Mukai vector of O'Grady type. We write h = c1(H). For any m ∈ Z, there is a canonical isomorphism f : MX,B,H (v) → MX,B,H (v · exp(mh)), such that the following diagram / / O O / / O O BIRATIONAL GEOMETRY OF SINGULAR MODULI SPACES OF O'GRADY TYPE 21 commutes v⊥,tr (−)·exp(mh) ∼= θtr B,H,v H 2(MX,B,H (v), Z) (3.5) f ∗ ∼= In particular, θtr B,H,v is an isometry if and only if θtr (v · exp(mh))⊥,tr θtr B,H,v·exp(mh) / H 2(MX,B,H (v · exp(mh)), Z). B,H,v·exp(mh) is an isometry. Proof. This is classical. We state a proof using the language similar to the above lemmas. We observe that F : Coh(X, B) → Coh(X, B) with the assignment F (E) = E ⊗ H ⊗m is an equivalence of abelian categories and is well defined for families. Moreover the functor F preserves H-stability. Therefore it induces an isomorphism of moduli spaces f : MX,B,H (v) → MX,B,H(v·exp(mh)), and takes a (quasi-)universal family E on M st X,B,H (v) to a (quasi-)universal family F (E) on M st X,B,H (v · exp(mh)). We can also write down a diagram as follows. v⊥,tr ΦE H 2(M st X,B,H (v), Q) (3.6) (−)·exp(mh) ∼= f ∗ ∼= (v · exp(mh))⊥,tr ΦF (E) / H 2(M st X,B,H (v · exp(mh)), Q). We just point out that the extra factor exp(mh) in the twisted Mukai vector and the extra factor exp(mh) coming from the twisted Chern character of the (quasi-)universal family cancel out, thanks to the dual operation we have in the Mukai version of cohomological integral functor (see [HL10, Definition 6.1.12], or [Yos01, Section 1.2], or [PR13, Section 3.2]). This is why the commutativity of the diagram holds. By the same reasoning as in Lemma 3.12, this implies the commutativity of (3.5). (cid:3) 3.3. Main proof and consequence. We are now ready to state the proof of Theorem 2.7. The most difficult part in the proof is to show that the Mukai morphism is an isometry. We outline the idea used in this part of proof. The principal idea has already been established in [PR13] where the case of Gieseker moduli spaces is proven. Perego and Rapagnetta proved their result by showing that one particular moduli space has this property, and all the other Gieseker moduli spaces can be related to it by Fourier-Mukai transforms and deformations, under which this property is stable. Our approach for any Bridgeland moduli space comes in two steps. We first use a Fourier-Mukai transform constructed in [BM14b, Lemma 7.3] to relate such a Bridgeland moduli space to a twisted Gieseker moduli space and then use the deformations constructed in Propositions 3.7 and 3.9 to relate the twisted Gieseker moduli space to an untwisted Gieseker moduli space. Since the property of the Mukai morphism being an isometry is stable under both operations, and is already true for the untwisted Gieseker moduli space, we conclude it is true for the Bridgeland moduli space. A complete proof goes as follows. Proof of Theorem 2.7. We observe that the statement (1) follows from the very general result [PR13, Lemma 3.1]. The pure weight two Hodge structure on H 2(M, Z) is described in [PR13, Definition 3.4] and the induced quadratic form is described in [PR13, Definition 3.5]. From now on we focus on statement (2). We first show that the Mukai morphism is always well-defined. Note that using the (quasi-)universal family on the stable locus M st in M , there v : v⊥,tr → H 2(M st, Q). It is just the problem whether this map is always a well-defined map θst can be lifted to H 2(M, Z). / / / O O / / / O O 22 CIARAN MEACHAN AND ZIYU ZHANG If M is the moduli space of B-twisted H-semistable sheaves with Mukai vector v, where B is a rational B-field, v is a Mukai vector of O'Grady type, and H is generic with respect to v, then the moduli space M admits a GIT construction, and the construction in [PR13, Section 3.2] can be applied without change. Nevertheless, we point out that the argument there includes a step which requires the local existence of relative moduli spaces, which is guaranteed by Lemma 3.3 and Proposition 3.10. To show the general case, we use [BM14b, Lemma 7.3] and the discussion thereafter. We can always find a derived equivalence in the form of a Fourier-Mukai transform Φ : D(X) → D(Y, B) for a B-twisted projective K3 surface Y , such that the image Φ(σ) of the generic stability condition σ lies in the Gieseker chamber of Stab†(Y, B), hence is represented by an ample line bundle H, which is generic with respect to the Mukai vector Φ(v). In particular, Φ induces an isomorphism of the moduli spaces Φ : MX,σ(v) → MY,B,H(−Φ(v)). Moreover, every H- semistable object representing a point in MY,B,H(Φ(v)) is locally free, hence Φ(v) has positive rank. We consider the following diagram θtr v ✆ ✾ ♦ s ❑ ❖ ❥ ❧ v⊥,tr θst v H 2(M st X,σ(v), Q) i∗ ❘ H 2(MX,σ(v), Z) Φ ∼= Φ∗ Φ∗ (Φ(v))⊥,tr θst Φ(v) H 2(M st Y,B,H(Φ(v)), Q) θtr Φ(v) i∗ H 2(MY,B,H(Φ(v)), Z) By [Yos01, Proposition 2.4]3, we obtain that the upper square commutes, and the first row is an isometry. It is also easy to see the bottom square also commutes, since both vertical maps in it are functorial and both horizontal maps are induced by the isomorphism of the moduli spaces. Φ(v) for the maps in the right column. Therefore a lift θtr By the above discussion, there is a lift θtr for the left column also exists. This shows that the Mukai morphism θtr v is always well-defined for any Bridgeland moduli space of O'Grady type. v Now we show that θtr v is an isometry. From the above diagram and [PR13, Lemma 3.12], it suffices to show it only for every twisted Gieseker moduli space MX,B,H (v) of O'Grady type. By Lemma 3.3, we can change v and H if necessary, so that v is deformable. By Lemma 3.6, we can add an integral class to B if necessary, so that B2 < 0, and for every non-zero class g ∈ SpanQ{h, B} ∩ H 2(X, Z) ∩ h⊥, we have g2 < min{−2,− r2 4 ∆(v)}, where h = c1(H) and r is the degree zero component of v. By the above discussion, r is a positive integer. By Lemmas 3.12, 3.13, and 3.14, we observe that the above adjustments do not affect the property that θtr is an isometry. That is, it suffices to prove the property of θtr v being an isometry under these additional assumptions. v The key step here, is to deform such a moduli space of twisted sheaves to an untwisted moduli space. By Proposition 3.7, we can find a family of polarised K3 surfaces (X ,H) → S over an integral base S, with fibre (Xs, Hs) over any point s ∈ S. Moreover, the family comes equipped with a locally constant family of B-fields with Bs ∈ H 2(Xs, Q) and such that (Xs0, Hs0, Bs0) = (X, H, B) and (Xs1, Hs1, Bs1) satisfies Bs1 ∈ H 1,1(Xs1, Q) for some s0, s1 ∈ S. By Lemma 3.1, the Mukai vector v on the fibre over s0 extends to a locally constant section over S, such that its value vs at every point s ∈ S is still deformable, hence vs ∈ H ∗ alg(Xs, Bs, Z). By Proposition 3.9, we can further assume that Hs is generic with respect to vs for every point s ∈ S. 3Although this proposition is stated for moduli spaces of sheaves, the calculation in its proof is purely coho- mological, which works for moduli spaces of complexes literally. / / ) ) ✤ u u O O O O BIRATIONAL GEOMETRY OF SINGULAR MODULI SPACES OF O'GRADY TYPE 23 We now define T := {s ∈ S : θtr Bs,Hs,v is an isometry} and claim that T is both open and closed in S. To show it is open, we consider a point t0 ∈ T . By Proposition 3.10, we can find an (´etale or analytic) open subset Ut0 ⊂ S containing t0, such that there is a relative moduli space M over Ut0, whose fibre at each s ∈ Ut0 is the moduli space MX,Bs,Hs(v). As observed in [PR13, Proof of Theorem 1.7], the isomorphisms of lattices and integral bilinear forms are both discrete properties and therefore have to remain constant in families. Since the isometry holds at the point t0, it has to hold for every point s ∈ Ut0, hence Ut0 ⊂ T . Since t0 is an arbitrary point of T , we conclude that T is open. To show T is closed, we take an arbitrary point t0 ∈ (S\T ). Then, again by Proposition 3.10, a relative moduli space exists over an open neighbourhood Ut0 of t0. Since θtr v fails to be an isometry at t0, it must fail to be an isometry for every point of Ut0 . Hence Ut0 ⊂ (S\T ) and we conclude that T is closed. Since S is connected, we either have T = S or T = ∅. However, by the construction of S, there is a point s1 ∈ S at which Bs1 ∈ H 1,1(Xs, Q) and hence represents the zero Brauer class. It has been proved in [PR13, Theorem 1.7] that the Mukai morphism is an isometry for untwisted moduli spaces of O'Grady type. By Lemma 3.13, this implies that Mukai morphism θtr Bs1 ,Hs1 is an isometry. Hence we must have T = S. In particular, the Mukai morphism θtr is an isometry, which is the arbitrary twisted Gieseker moduli space of O'Grady type we started with. This finishes the proof that θtr Bs0 ,Hs0 σ is an isometry. Finally, to show that θtr Theorem 1.7] can be applied without change in our situation. And we are done. σ preserves Hodge structure we observe that step 4 in [PR13, Proof of (cid:3) Remark 3.15. After writing this proof down, it was pointed out to us that one can use an observation of Yoshioka to reduce all of the previous arguments to the untwisted case. Indeed, let Φ : D(X) → D(Y, B) be an equivalence inducing an isomorphism MX,σ(v) → MY,B,H(−Φ(v)) as above. Set w0 := Φ((1, 0, 1))∨ and w1 := Φ((0, 0, 1))∨, where ∨ means the dual, and observe that for a general H ′ which is sufficiently close to H, we have Z := MY,−B,H ′(w1) ≃ X which is an untwisted K3 surface, since hw0, w1i = −1. Now, if we let F be a universal family and Ψ : D(Y, B) → D(Z) be a twisted Fourier-Mukai transform such that Ψ(Oy) = FZ×{y} then we can apply [Yos06, Theorem 1.7 and Remark 1.6] to E(nH ′) for E ∈ MY,B,H ′(−Φ(v)) and n ≫ 0 to reduce to the untwisted case. In fact, the above proof yields the following interesting consequence. Corollary 3.16. Let v be a Mukai vector of O'Grady type and σ ∈ Stab†(X) be any generic stability condition with respect to v. Then the symplectic resolution of Mσ(v) is deformation equivalent to the irreducible holomorphic symplectic manifold constructed by O'Grady in [O'G99]. Proof. We use the same Fourier-Mukai transform and deformation argument as in the proof of Theorem 2.7. Applying [BM14b, Lemma 7.3] we obtain an isomorphism Φ : MX,σ(v) → MY,B,H(−Φ(v)). In particular, their symplectic resolutions are isomorphic as well. By Proposi- tions 3.7 and 3.9, we can deform the underlying twisted polarised K3 surface to an untwisted K3 surface, and the moduli space of twisted semistable sheaves deforms along with the underlying K3 surface locally near every point. Now we can apply [PR13, Proposition 2.17] and conclude that the fibrewise resolutions of this equisingular family of moduli spaces are also deformation equivalent. Although these families of singular moduli spaces do not necessarily glue into a global family due to the relevance of B-field and its Cech 2-cocycle representation, the singular moduli spaces in different local families over the same twisted K3 surface are always isomorphic due to Lemmas 3.12 and 3.13. Therefore, they have isomorphic resolutions. Thus, we can con- clude that the symplectic resolution of Mσ(v) is deformation equivalent to that of a Gieseker 24 CIARAN MEACHAN AND ZIYU ZHANG moduli space of untwisted sheaves, which, by [PR13, Theorem 1.6], is deformation equivalent to the holomorphic symplectic manifold constructed by O'Grady in [O'G99]. (cid:3) Remark 3.17. We have learned from [KLS06, Theorem 6.2] and [PR13, Theorem 1.6] that the only deformation type of irreducible holomorphic symplectic manifold that we can obtain by resolving singular Gieseker moduli spaces of semistable sheaves on K3 surfaces is the one con- structed by O'Grady in [O'G99]. Our Corollary 3.16 shows that, even if we enlarge our scope to resolutions of moduli spaces of Bridgeland semistable objects on K3 surfaces, we still only get the same deformation type. Remark 3.18. There is another more straightforward proof of Corollary 3.16, which does not need to go through twisted Gieseker moduli spaces and their deformations. However, it relies on a result which we will prove later, and a very deep theorem of Huybrechts. In fact, by Theorem 5.4, we know that Mσ(v) is birationally equivalent to a Gieseker moduli space MH (v) for some ample line bundle H which is generic with respect to v. Therefore, their symplectic resolutions are also birationally equivalent. By [Huy03, Theorem 2.5], these two symplectic resolutions are deformation equivalent. By [PR13, Theorem 1.6], they are further deformation equivalent to the irreducible holomorphic symplectic manifold constructed by O'Grady in [O'G99]. We remind the reader that this shorter proof does not result in a circular argument because the proof of Theorem 5.4 does not use Corollary 3.16. 4. The Local Bayer-Macr`ı Map The most essential ingredient in [BM14b, BM14a] is a linearisation map from Stab†(X) to N´eron-Severi group of the moduli space for any primitive Mukai vector v ∈ H ∗ alg(X, Z). In this paper we refer to it as the Bayer-Macr`ı map. Note that the construction of Bayer-Macr`ı map depends on the moduli space, therefore one can define a Bayer-Macr`ı map for each chamber in Stab†(X), which we will call a local Bayer-Macr`ı map. And we are primarily only interested in the restriction of the map in that chamber (or its closure). These local Bayer-Macr`ı maps are already sufficient for the purpose of proving the projectivity of Bridgeland moduli spaces as in [BM14b]. However, in [BM14a], the authors managed to glue the maps defined on various chambers together and obtained a global Bayer-Macr`ı map, which reveals nicely the relation between wall crossings on the stability manifold and birational geometry of the Bridgeland moduli spaces. In this section, we will follow [BM14b] and generalise the definition of the local Bayer-Macr`ı map to any non-primitive Mukai vector v of O'Grady type, which relies greatly on the isomorphism θσ proved in Corollary 2.8(3). By using the local Bayer-Macr`ı map, we will follow the approach in [BM14b] to show the positivity of some determinant line bundles on the singular moduli spaces of O'Grady type. We will discuss the global Bayer-Macr`ı map in a later section. 4.1. Construction of the local Bayer-Macr`ı map. The local Bayer-Macr`ı map was first defined in [BM14b, Section 3 and 4], and plays a vital role in [BM14b, BM14a]. This map establishes a bridge between the two central spaces in question: the stability manifold of the K3 surface, and the N´eron-Severi group of the moduli space. There are mainly two different descriptions of the map. The first description, as defined in [BM14b, Section 3], offers the perfect point of view to understand the positivity theorem [BM14b, Theorem 1.1]. While the second description, defined in [BM14b, Section 4], was used more often thereafter (for instance in [BM14a, Section 10]). The two descriptions are equivalent by [BM14b, Proposition 4.4]. BIRATIONAL GEOMETRY OF SINGULAR MODULI SPACES OF O'GRADY TYPE 25 Here we briefly recall the second description of the local Bayer-Macr`ı map, which emphasises the role of the algebraic Mukai morphism θC from Corollary 2.8, and is more convenient to use for our purposes. Although it could be described in a more general situation, we only restrict ourselves to the case of non-primitive Mukai vectors v of O'Grady type. In short, fix such a Mukai vector v and a chamber C ∈ Stab†(X). Then the composition of the following three maps is called the local Bayer-Macr`ı map for the chamber C with respect to the Mukai vector v and is denoted by ℓC. Stab†(X) Z−→ H ∗ alg(X, Z) ⊗ C I−→ v⊥ θC−→ NS(M ). alg(X, Z) ⊗ C, we have I(Ω) = Im Ω We briefly describe each map in the above composition. The first map is simply a forgetful map. For any σ = (Z,P) ∈ Stab†(X), with the central charge Z(−) = (Ω,−), we have Z(σ) = Ω ∈ H ∗ alg(X, Z) ⊗ C which only remembers the central charge. An important feature is that Z is a covering map on an open subset of the target, which was proved in [Bri08, Section 8] (see also [BM14a, Theorem 2.10]). Note that Z does not depend on v or C. The second map forgets even more. For any Ω ∈ H ∗ −(Ω,v) . In particular, when (Ω, v) = −1, this map simply takes the imaginary part of Ω. This map does not depend on C either. Finally, the third map θC is the algebraic Mukai morphism described in Corollary 2.8. In particular, it is an isometry. Note that it is the only map among the three which depends on the choice of the chamber C. We fix some notations for later convenience (and to keep consistent with the notations in [BM14b, BM14a]). We write wσ := (I ◦ Z)(σ) for the image of any σ ∈ Stab†(X) under the composition of first two maps. The composition of all three maps is denoted by ℓC = θC ◦ I ◦ Z. Although ℓC is defined for the whole Stab†(X), we are mainly interested in its behaviour on the chamber C itself (or rather, the closure of C). When the chamber C is clear from the context, for any σ in the interior of C, we also write ℓσ := ℓC(σ), hence θC(wσ) = ℓσ. And similarly, for any σ0 on the boundary of C, we also write ℓσ0 := ℓC(σ0). 4.2. Positivity via the local Bayer-Macr`ı map. The projectivity of moduli spaces is one of the main results in [BM14b]. In [BM14b, Corollary 7.5], it is proved that ℓσ is ample for a generic σ in the case of primitive Mukai vectors. However, for non-primitive Mukai vectors, the authors used a rather indirect approach to prove the projectivity, due to the lack of Yoshioka's theorem [BM14b, Theorem 6.10] in such cases. As a result, it is proved that ℓσ is ample for σ in a dense subset of C. However, our Theorem 2.7 allows us to mimic the proof of [BM14b, Corollary 7.5] and show the ampleness of ℓσ for all σ ∈ C. alg(X, Z) be a Mukai vector of O'Grady type and C ⊂ Stab†(X) an Proposition 4.1. Let v ∈ H ∗ open chamber with respect to v. Then the image ℓσ under the local Bayer-Macr`ı map defined by the chamber C is ample for all σ ∈ C. Proof. As in Theorem 2.7, we write π : fM → M for the symplectic resolution of M . By fM . Moreover, by Corollary 2.8, we have eq(π∗ℓσ) = q(ℓσ) = w2 23.14], we know that the top self-intersection number of π∗ℓσ on fM is positive. By the bigness criterion [KM98, Proposition 2.61], we see that π∗ℓσ is big and nef on fM . Furthermore, the base σ is globally generated for b ≫ 0. [BM14b, Theorem 1.1], we know that ℓσ is nef on M and therefore its pullback π∗ℓσ is nef on σ > 0, where the last inequality follows from [Bri08, Theorem 1.1]. Now by the Beauville-Fujiki relation [GHJ03, Proposition point free theorem [KM98, Theorem 3.3] tells us that π∗ℓ⊗b 26 CIARAN MEACHAN AND ZIYU ZHANG Now we reduce all the above statements from fM to M . Since M has rational singularities, we have H 0(fM , π∗ℓ⊗b obtained by pulling back global sections of ℓ⊗b globally generated. σ ) for any b, i.e. the global sections of π∗ℓ⊗b σ along π. Hence we conclude that ℓ⊗b σ σ are precisely those is also σ ) = H 0(M, ℓ⊗b Finally, by [BM14b, Theorem 1.1] we see that ℓ⊗b is positive on all curves in M . But any σ globally generated line bundle that is positive on all curves is ample; see [Huy99, Proposition 6.3]. (cid:3) By a similar argument, we can immediately deduce a weaker property for stability conditions on the wall. The following proposition shows that in the O'Grady situation, ℓσ0 for any σ0 on the boundary of C is nef, big, and semiample. Recall that a line bundle is semiample if a certain multiple of it is base point free. In particular, this result shows that the contraction morphisms π± in [BM14b, Theorem 1.4] are still well-defined in the O'Grady situation and are birational morphisms. We will generalise this result to arbitrary Mukai vectors in Proposition 5.2 from a completely different point of view. Proposition 4.2. Let v ∈ H ∗ alg(X, Z) be a Mukai vector of O'Grady type and C ⊂ Stab†(X) an open chamber with respect to v. For any stability condition σ0 in the boundary of the chamber C, its image ℓσ0 under the local Bayer-Macr`ı map defined by the chamber C is big, nef, and semiample on MC(v). In particular, ℓσ0 induces a birational morphism which contracts curves in MC(v) parametrising S-equivalent objects with respect to the stability condition σ0. Proof. In fact, the first two paragraphs in the proof of Proposition 4.1 also work here without any change, which prove the first statement. The second statement is an immediate consequence of [BM14b, Theorem 1.1]. (cid:3) 5. Classification of Walls: Results In this section, we study wall crossing on the stability manifold for non-primitive Mukai vec- tors. More precisely, we prove that the criteria in [BM14a, Theorem 5.7] still gives a complete classification of walls in the O'Grady situation. Moreover, we will prove that [BM14a, Theorem 1.1] also holds in the O'Grady situation, which allows us to glue the local Bayer-Macr`ı maps and study the birational geometry of the singular moduli spaces via wall crossings. Due to the technicality in the proofs, we will only state our results in this section, while leaving all proofs for next section. 5.1. Hyperbolic lattice associated to a wall. One of the main tools in [BM14a] is a rank two hyperbolic lattice associated to any wall. More precisely, if we fix a Mukai vector v ∈ H ∗ alg(X, Z) with v2 > 0, and a wall W of the chamber decomposition with respect to v, then we can define HW ⊂ H ∗ alg(X, Z) to be the set of classes w ∈ HW ⇔ Im = 0 for all σ = (Z,P) ∈ W. Z(w) Z(v) By [BM14a, Proposition 5.1], which also works when v is not primitive, the set HW is a primitive rank two hyperbolic lattice. Conversely, given a primitive rank two hyperbolic sublattice H ⊂ H ∗ alg(X, Z) containing v, we can define a potential wall W in Stab†(X) to be a connected component of the real codimension one submanifold of stability conditions σ = (Z,P) which satisfy the condition that Z(H) is contained in a line. Notice that every (potential) wall is BIRATIONAL GEOMETRY OF SINGULAR MODULI SPACES OF O'GRADY TYPE 27 associated to a unique rank two hyperbolic lattice H whereas each lattice may give rise to many (potential) walls. Roughly speaking, whilst crossing a wall W, all the relevant Mukai vectors lie in the same hyperbolic lattice HW . This simple observation reduces the analysis of wall-crossing to elementary lattice-theoretic computations. We also need to recall the names of some special classes in H. A class u ∈ H is called a spherical class if u2 = −2, or an isotropic class if u2 = 0. We say a hyperbolic lattice H is an isotropic lattice if H contains at least one non-zero isotropic class. We say a (potential) wall W is an isotropic wall if its associated hyperbolic lattice H is an isotropic lattice. After establishing the relation between walls and rank two hyperbolic sublattices, we review the notions of a positive cone and an effective cone. We assume we have a potential wall W and its associated rank two hyperbolic lattice H. The positive cone PH is a cone in H ⊗ R, which is generated by integral classes u ∈ H, with u2 > 0 and (v, u) > 0. We call any integral class in PH a positive class. Note that the positive cone PH only depends on the choice of H, not on the choice of the wall associated to H. In comparison, the effective cone CW depends on the choice of a potential wall. It is a cone in H ⊗ R, which is generated by integral classes u ∈ H, with u2 > −2 and Re Z(u) Z(v) > 0 for a fixed σ = (Z,P) ∈ W. This notion is well-defined. Indeed, by [BM14a, Proposition 5.5], this cone does not depend on the choice of σ ∈ W. We call any integral class in CW an effective class. We also point out that for different walls W associated to the same hyperbolic lattice H, the effective cone CW might differ by spherical classes. However, they all contain the same positive cone PH. We also need to make the same genericity assumption as in [BM14a, Remark 5.6]. 5.2. First classification theorem. We will classify all the walls in the stability manifold from two points of view. Note that various types of walls have been defined in [BM14b, Section 8] and [BM14a, Definition 2.20]. From now on, we always denote a generic stability condition on a wall W by σ0, while two generic stability conditions on two different sides of the W by σ+ and σ−. The chambers which contain σ+ and σ− are denoted by C+ and C− respectively. Moreover, we always assume that σ+ and σ− are sufficiently close to the wall, whose precise meaning is contained in [BM14a, Proof of Proposition 5.1]. We first classify walls according to whether there exists any σ0-stable objects of class v. We say a (potential) wall associated to the class v is totally semistable, if there is no σ0-stable objects of class v for a generic stability condition σ0 ∈ W. Otherwise, W is not totally semistable. In other words, the (potential) wall W is totally semistable if and only Mσ+(v) and Mσ−(v) are completely disjoint from each other, and W is not totally semistable if and only if they contain a common open dense subset Mσ0(v). There is a third way to describe these walls, which reveals more on the reason why the name is obtained. The wall W is totally semistable if every stable σ+-stable object becomes strictly σ0-semistable, which is equivalent to the condition that every stable σ−-stable object is strictly σ0-semistable. Otherwise W is not totally semistable. We are ready to state the following numerical criterion for a wall to be totally semistable for an arbitrary class v ∈ H with v2 > 0, regardless of whether it is primitive or has O'Grady type. The theorem generalises the first half of [BM14a, Theorem 5.7]. Theorem 5.1. Let v ∈ H be a positive class with v2 > 0, and W be a potential wall for v. Then W is a totally semistable wall for v if and only if either of the following two conditions holds 28 CIARAN MEACHAN AND ZIYU ZHANG (TS1) there exists an effective spherical class s ∈ H with (s, v) < 0; (TS2) there exists an isotropic class w ∈ H with (w, v) = 1. In particular, if v is a non-primitive class, then (TS2) cannot happen and (TS1) is the only possibility for a totally semistable wall. As an application of the proof of this theorem, we also obtain the bigness of ℓσ0 for an arbitrary class v, which generalises [BM14b, Theorem 1.4(a)] and Proposition 4.2. Proposition 5.2. Let v ∈ H be a positive class with v2 > 0, and W be a potential wall for v. For a generic σ0 ∈ W, its image ℓC±(σ0) under the local Bayer-Macr`ı map with respect to the chamber C± induces a birational morphism π± : Mσ±(v) → M ± which contracts curves in Mσ±(v) parametrising S-equivalent objects under the stability condition σ0, where M ± is the image of π± in Mσ0 (v). 5.3. Second classification theorem. The second classification result relies on Proposition 5.2 (or Proposition 4.2). Since π+ : Mσ+ (v) → M + is a birational morphism, it could be one of three possible types: a divisorial contraction, a small contraction, or an isomorphism. In fact, in the next theorem, we will see that π+ and π− always correspond to the same contraction type because the numerical criteria do not tell the difference of two sides of W. So it doesn't matter whether the definition is made for π+ or π−. In the first two cases, since π+ and π− induce actual contractions, the two moduli spaces Mσ+ (v) and Mσ−(v) are not the same and hence W is a genuine wall which we call a divisorial wall and a flopping wall respectively; the reason for the name will be explained in Section 6. If π+ and π− are both isomorphisms, we will see that Mσ+ (v) and Mσ−(v) are isomorphic. By the first classification Theorem 5.1, they could be either disjoint from each other, in which case we call W a fake wall, or identical with each other, in which case we say that W is not a wall. We refer the reader to [BM14b, Section 8] and [BM14a, Definition 2.20] for the names of these walls. With these notions at hand, we can now state the second classification result which provides a numerical criterion for each of the three types of contraction discussed above. This generalises the second half of [BM14a, Theorem 5.7], but we can only prove it in the O'Grady situation: Theorem 5.3. Let v ∈ H be a positive class with v = 2vp, where v2 wall for v. Then W can be one of the following three types: p = 2, and W be a potential • W is a wall inducing a divisorial contraction if and only if at least one of the following two conditions holds: (BN) there exists a spherical class s with (s, v) = 0, or (LGU) there exists an isotropic class w with (w, v) = 2. Moreover, the condition (LGU) necessarily implies the condition (BN). • Otherwise, W is a wall inducing a small contraction if and only if the following condition holds: (SC) there exists a spherical class s with 0 < (s, v) 6 v2 2 = 4. • In all other cases, W is either a totally semistable wall, or not a wall. We discuss briefly the differences between [BM14a, Theorem 5.7] in the primitive case and our classification in the O'Grady setting. The main difference shows up in the condition required for a divisorial wall. On the one hand, the non-primitivity of our Mukai vectors excludes the BIRATIONAL GEOMETRY OF SINGULAR MODULI SPACES OF O'GRADY TYPE 29 possibility of having a Hilbert-Chow wall. On the other hand, we will prove in Lemma 6.4 that (LGU) always implies (BN) for moduli spaces of O'Grady type. Therefore, (BN) is the unique condition that is required to characterise a divisorial wall. In other words, every divisorial wall is a wall of Brill-Noether type, while it is also a wall of Li-Gieseker-Uhlenbeck type if both (BN) and (LGU) are satisfied at the same time. However, there is a subtle difference between the cases (BN) and (LGU) which can be seen by looking at the HN filtration of the generic semistable object along the divisor; see Remark 6.5 for more details. Another difference between our Theorem 5.3 and [BM14a, Theorem 5.7] is that there are no flopping walls arising from a decomposition v = a + b into positive classes a and b. Indeed, this case cannot happen because the O'Grady spaces have such small dimension; see Lemma 6.13 for more details. The above classification shows us how a moduli space could change when the stability condition approaches a wall. Although the contraction induced by a wall can be arbitrarily misbehaved, we can always find a birational map, which identifies open subsets in the moduli spaces Mσ+ (v) and Mσ−(v) with complements of codimension at least two. By composing these birational maps, we can obtain the same conclusion for any pair of generic stability conditions. This is the content of the following result, which generalises [BM14a, Theorem 1.1]. Theorem 5.4. Let v ∈ H be a positive class with v = 2vp, where v2 p = 2, and W a potential wall for v. Then there exists a derived (anti-)autoequivalence Φ of D(X), which induces a birational map Φ∗ : Mσ+(v) 99K Mσ−(v), such that • Φ∗ is stratum preserving, and • one can find open subsets U± ⊂ Mσ±(v), with complements of codimension at least two, and Φ∗ : U+ → U− an isomorphism, i.e. Φ∗ is an isomorphism in codimension one. In particular, Φ∗ identifies the N´eron-Severi lattices NS(Mσ+ (v)) ∼= NS(Mσ−(v)), as well as the movable cones Mov(Mσ+ (v)) ∼= Mov(Mσ− (v)) and big cones Big(Mσ+ (v)) ∼= Big(Mσ−(v)). More generally, the same result is true for any two generic stability conditions σ, τ ∈ Stab†(X). Remark 5.5. Let us take this opportunity to clarify why an autoequivalence Φ of D(X), which sends every E ∈ U+ ⊂ Mσ+(v) to an object Φ(E) ∈ U− ⊂ Mσ−(v), induces a birational If we let Mσ±(v) denote the moduli stacks of σ±- map Φ∗ defined as a morphism on U+. semistable objects then, by the GIT construction in [BM14b], we have classifying morphisms f± : Mσ±(v) → Mσ±(v) together with the fact that the moduli spaces Mσ±(v) universally corepresent the moduli stacks Mσ±(v). Now, the autoequivalence Φ defines a morphism from an open substack of Mσ+(v) to an open substack of Mσ−(v), which is in fact an isomorphism (because the functor Φ has an inverse). Note that these two open substacks are in fact preimages of open subschemes of the corresponding moduli spaces (along f+ and f−); the reason is that if Φ(E) is σ−-semistable, then every object in the same σ+-S-equivalence class of E is also mapped to a σ−-semistable object. Finally, the isomorphism between the two open substacks descends to an isomorphism between the two open subschemes by the universal corepresentability of Mσ±(v). In particular, if U± ⊂ Mσ±(v) is an open subset then U± universally corepresents its preimage f −1 ± (U±) since universal corepresentability is preserved under arbitrary base change; see [HL10, Definition 2.2.1 & Theorem 4.3.4] for more details. Remark 5.6. This theorem suggests that we can use the notation NS(M (v)) for the N´eron- Severi lattice of the moduli space without specifying the generic stability condition (or rather its chamber). Similarly, we will also use the notations Mov(M (v)) and Big(M (v)) for the movable cone and big cone of the moduli space associated to any chamber in Stab†(X). 30 CIARAN MEACHAN AND ZIYU ZHANG The fact that all generic moduli spaces have canonically identified N´eron-Severi groups shows that, all chamberwise local Bayer-Macr`ı maps have the same target and hence could possibly be glued together into a global Bayer-Macr`ı map, which will reveal the power of wall crossings in the study of birational geometry of the moduli spaces; see Section 7. 6. Classification of Walls: Proofs In this section, we prove all the results which were stated in Section 5. First, we fix notation. Throughout, we set v = mvp ∈ H ∗(X, Z) to be a Mukai vector, for some positive integer m, and vp a primitive vector with v2 p > 0. Although we will mainly focus on moduli spaces of O'Grady type, where m = 2 and v2 p = 2, some of our results are actually true for arbitrary v. Let H ⊂ H ∗(X, Z) be a primitive hyperbolic rank two sublattice containing v, and W ⊂ Stab†(X) a potential wall associated to H. A generic stability condition on W is denoted by σ0, while generic stability conditions on the two sides of the wall are denoted by σ+ and σ−. The phase functions of the central charges of σ+, σ0, σ− are denoted by φ+, φ0, φ− respectively. We also use π± : Mσ±(v) → M ± for the morphism induced by ℓσ0 on the generic moduli spaces Mσ±(v). Finally, we point out that, in this section, for simplicity of notations, any morphism (or birational map) between two moduli spaces induced by a derived (anti-)autoequivalence Φ of D(X) is still denoted by Φ, by abuse of notation. We should point out that a large portion of our proofs are, in fact, already contained in [BM14a]. Therefore, to avoid repetition we will reference the results and proofs of [BM14a] freely and only point out where the differences are and what additional arguments (if any) need to be added. In particular, our focus will be on explaining the extra effort required to deal with the problems caused by the presence of the singular locus. Moreover, all results that we prove for σ+ also hold for σ− with identical proof, which we will not explicitly mention in every statement. 6.1. Totally semistable walls. In this subsection we prove the criterion stated in Theorem 5.1 for a wall to be totally semistable. We will also show that π± are always birational morphisms. We prove these results for arbitrary Mukai vectors v with v2 > 0. The notion of the minimal class in a GH-orbit will be frequently used; for its definition, we refer the reader to [BM14a, Proposition and Definition 6.6]. We start by proving some lemmas. The first two deal with non-minimal and minimal classes re- spectively, whilst the third one establishes a bridge between these two kinds of classes when they lie in the same GH-orbit. This will be used later to deduce results for non-minimal classes from the existing results for minimal classes. After that, we will prove Theorem 5.1 and Proposition 5.2. Lemma 6.1. If v is non-minimal, then there is no σ0-stable objects of class v, which means W is a totally semistable wall. Proof. This is the first statement in [BM14a, Proposition 6.8]. Note that the proof there works for both primitive and non-primitive classes. (cid:3) Lemma 6.2. Let v be a minimal, non-primitive class. Then there exist σ0-stable objects of class v. Proof. We write v = mvp where m > 1 and vp is a minimal primitive class. There are two cases to consider. BIRATIONAL GEOMETRY OF SINGULAR MODULI SPACES OF O'GRADY TYPE 31 If there is no isotropic class w with (w, vp) = 1, then by [BM14a, Theorem 5.7], there exist σ0-stable objects of class vp. Therefore by [BM14a, Lemma 2.16], there exist σ0-stable objects of class v. If instead there is an isotropic class w with (w, vp) = 1, then by [BM14a, Proposition 8.2] and the discussion above it (see also [LQ14, Lo12]), ℓσ0 induces a Li-Gieseker-Uhlenbeck morphism. Since (w, v) = m > 1, a generic stable object in Mσ±(v) corresponds to a locally free stable sheaf in a Gieseker moduli space, hence remains σ0-stable. (cid:3) Lemma 6.3. Let v be any class not satisfying (TS2), and v0 be the minimal class in the GH- orbit of v. Then there exists a derived autoequivalence Φ+ of D(X) defined as a composition of spherical twists, such that for every σ0-stable object E of class v0, Φ+(E) is a σ+-stable object of class v. Moreover, for any other σ0-stable object E′ of class v0, Φ+(E) and Φ+(E′) are not S-equivalent with respect to σ0. Proof. By [BM14a, Theorem 5.7] (in primitive case) or Lemma 6.2 (in non-primitive case), there always exist σ0-stable objects of class v0. The first statement is contained in [BM14a, Proposition 6.8], and the second statement is contained in the proof of [BM14a, Corollary 7.3]. (cid:3) Now we are ready to prove the criterion for totally semistable walls, and show that the contrac- tion induced by ℓσ0 on Mσ±(v) are always birational for an arbitrary Mukai vector v, which generalises Proposition 4.2. Proof of Theorem 5.1. If v is a primitive class, the statement is proved in [BM14a, Theorem 5.7]. Now we assume v is non-primitive. Note that in this case (TS2) cannot happen, hence we only need to prove that (TS1) is both sufficient and necessary for W to be a totally semistable wall. The sufficiency is the content of Lemma 6.1, and Lemma 6.2 proves the necessity by contradiction. (cid:3) Proof of Proposition 5.2. If v is a primitive class, the statement is [BM14b, Theorem 1.4(a)]. Now we assume v is non-primitive. Note that by [BM14b, Theorem 1.1], ℓσ is nef on Mσ+(v), and contracts curves which generically parametrise S-equivalent objects with respect to σ0; see Proposition 4.2. Therefore, it suffices to show that there exists a dense open subset U ⊂ Mσ0 (v) If v is minimal, we simply take U to be the open in which no curve is contracted by ℓσ0. subset of σ0-stable objects. By Lemma 6.2, U is non-empty and hence dense in Mσ0(v). If v is non-minimal, let v0 be the corresponding minimal class with an open subset U0 ⊂ Mσ+(v0) of σ0-stable objects. Then, by Lemma 6.3, U = Φ(U0) is non-empty and does not contain any curve which can be contracted by ℓσ0. (cid:3) By virtue of Proposition 5.2, the contractions π± : Mσ±(v) → M ± induced by ℓσ0 are birational morphisms, and we will deal with three mutually exclusive cases: a divisorial contraction, a small contraction, or no contraction at all, i.e. an isomorphism. We will prove the numerical criterion for each case, and show that Mσ+(v) and Mσ−(v) are always birational and isomorphic in codimension one. We also point out that from now on, we always restrict ourselves to the O'Grady situation. That is, v = 2vp where vp is a primitive Mukai vector with v2 p = 2. We do this because many of the following proofs will rely heavily on the existence of a symplectic resolution. 6.2. Walls inducing divisorial contractions. In this subsection we prove Theorems 5.3 and 5.4 in the case that ℓσ0 induces a divisorial contraction. We remind readers that the singular 32 CIARAN MEACHAN AND ZIYU ZHANG locus in the O'Grady moduli spaces has codimension two. Therefore, the contraction of a divisor is a phenomenon which happens in the smooth locus. This is the reason why most of the arguments in [BM14a] still apply in the O'Grady situation. This subsection consists mainly of a series of lemmas; each of which will become an integral part in the proof of Theorems 5.3 and 5.4. At the end of this subsection, we summarise all the results with Proposition 6.12. We start with the following special feature of classes of O'Grady type, which will be used later to simplify a few proofs. Lemma 6.4. Let v be any class of O'Grady type. If (LGU) holds for v, then (BN) also holds for v. Moreover, v is a minimal class, and we can choose the class w in (LGU) so that Mσ0(w) = M st σ0(w). Proof. By (LGU) we have (w, vp) = 1 which implies w is a primitive class. Using the notation in [BM14a, Lemma 8.1], we have either (w0, vp) = 1 or (w1, vp) = 1. Suppose (w0, vp) = 1. Then we have (vp− w0)2 = v2 p− 2 = 0 which can only happen when vp− w0 is a multiple of w0 or w1. If it is a multiple of w0, then so is vp, which contradicts v2 p = 2. Thus, we have vp − w0 = kw1 p = (w0 + kw1)2 = 2k(w0, w1) which implies for some non-negative integer k and hence 2 = v2 k = 1 and (w0, w1) = 1. Swapping subscripts shows that starting from (w1, vp) = 1 also yields vp = w0 + w1 and (w0, w1) = 1. Therefore, we always have (vp, w0) = (vp, w1) = 1. Now, the unique effective spherical class is given, up to sign, by s = w0 − w1. In particular, (s, v) = 2(s, vp) = 0 and we see that (BN) must also hold. We also have M st σ0 (w0) = Mσ0(w0) and M st σ0(w1) = ∅, where w1 = w0 + (s, w0)s, and so we can choose the class w in (LGU) to be w0; see [BM14a, Proposition 6.3 and Lemma 8.1] for more details. (cid:3) Remark 6.5. In the case when (LGU), and hence (BN), is satisfied, the HN filtration along the divisor does not quite behave like the normal BN case. Indeed, if we look at the arguments of [BM14a, proof of Theorem 1.1, p.40] then we see that the divisor of semistable objects is still the BN-divisor, but the HN filtration of the generic semistable object does not have the normal BN-form, i.e. it does not contain a spherical factor. Lemma 6.6. Let v be a minimal class of O'Grady type. If (BN) does not hold, then the set of σ0-stable objects in Mσ+(v) has complement of codimension at least two. Proof. If H does not contain any isotropic vector, we follow the proof of [BM14a, Lemma 7.2]. Otherwise, by Lemma 6.4, (LGU) does not hold either. Moreover there is no isotropic class w with (w, v) = 1 since v is non-primitive. Therefore, we can follow steps 1 and 2 of the proof of [BM14a, Proposition 8.6] to get the conclusion. The only change that we have to make in both proofs is that the application of their Theorem 3.8 has to be replaced by our Proposition 2.3. (cid:3) Lemma 6.7. Let v be a non-minimal class of O'Grady type. If (BN) does not hold, then there is an open subset of Mσ+(v) with complement of codimension at least two, on which no curve is contracted by ℓσ0. Proof. We write v0 for the minimal class in the GH-orbit containing v, then (BN) does not hold for v0. By Lemma 6.6, there exists an open subset U0 ⊂ Mσ+ (v0) with complement of codimension at least two, which parametrises objects remaining stable under σ0. By Lemma 6.3, there is a derived autoequivalence Φ+ of D(X) which induces an injective morphism Φ+ : U0 → M st σ+(v), and the image Φ+(U0) does not contain any curve that generically parametrises S-equivalent objects under σ0. Therefore, no curve in Φ+(U0) will be contracted by ℓσ0. It remains to show Φ+(U0) has complement of codimension two in Mσ+(v). BIRATIONAL GEOMETRY OF SINGULAR MODULI SPACES OF O'GRADY TYPE 33 Since v0 is a minimal class and (LGU) does not hold (by Lemma 6.4), Theorem 5.1 tells us that W is not a totally semistable wall for v0,p, i.e. the effective primitive class proportional to v0. By Lemma 6.3, the same derived auto-equivalence Φ+ takes every σ0-stable object of class v0,p to a σ+-stable object of class vp, i.e. the effective primitive class proportional to v. Now by Lemma 2.5, Φ+ : Mσ+ (v0) 99K Mσ+(v) is a stratum preserving birational map and so we can apply Proposition 2.6 to conclude that Φ+(U0) has complement of codimension two in Mσ+(v). (cid:3) Lemma 6.8. Let v be any class of O'Grady type and assume that (BN) holds for v while (LGU) does not hold for v. Then π+ is a divisorial contraction of Brill-Noether type. Proof. We first consider the case that H contains no isotropic vectors. The divisor contracted by π+ is constructed in [BM14a, Lemma 7.4] when v is a minimal class, and in [BM14a, Lemma 7.5] when v is a non-minimal class. Next, we consider the case that H is isotropic. By [BM14a, Proposition 6.3], there is only one effective spherical class in H, hence v is necessarily minimal. The divisor contracted by π+ is constructed in [BM14a, Lemma 8.8], which also shows the contraction has Brill-Noether type. (cid:3) Lemma 6.9. Let v be any class of O'Grady type and assume that (BN) holds for v while (LGU) does not hold. Then there is a derived (anti-)autoequivalence Φ inducing a stratum preserving birational map Φ : Mσ+(v) 99K Mσ−(v), which is an isomorphism in codimension one. Proof. By Lemma 6.8 we know that π+ is a contraction of Brill-Noether type under the given assumption. If v is a minimal class, then we can follow the proof of [BM14a, Theorem 1.1(b)] in the Brill-Noether case and obtain a spherical twist Φ of D(X), inducing a birational map Φ : Mσ+ (v) 99K Mσ−(v), which is an isomorphism in codimension one. It remains to show that it is stratum preserving. Note that under the given assumption, for the primitive class vp (which is the primitive part of v), the wall W also induces a Brill-Noether divisorial contraction and is not totally semistable by [BM14a, Theorem 5.7]. We apply the same part of proof in [BM14a, Theorem 1.1(b)] to conclude that the restriction of Φ on the moduli spaces with primitive classes Φp : Mσ+(vp) 99K Mσ−(vp) is also a birational map. By Lemma 2.5, we conclude that Φ is a stratum preserving birational map between moduli spaces with non-primitive classes. Now we consider the case that v is not a minimal class. We still write v0 for the minimal class in the GH-orbit of v. The above construction gives Φ0 : Mσ+(v0) 99K Mσ−(v0). By [BM14a, Lemma 7.5], there is a composition of spherical twists, say Φ+, inducing a birational map Φ+ : Mσ+(v0) 99K Mσ+(v). Moreover, since W is not a totally semistable wall for the primitive class v0,p (which is the primitive part of v0), we can apply [BM14a, Proposition 6.8] to see that Φ+ also induces a birational map Φ+,p : Mσ+(v0,p) 99K Mσ+(vp) and Lemma 2.5 shows that Φ+ is a stratum preserving birational map. By [BM14a, Lemma 7.5], there exists an open subset U+ ⊂ Mσ+(v0) with complement of codimension at least two, such that Φ+ is an injective morphism on U+ and so by Proposition 2.6, we see that Φ+ is an isomorphism in codimension one. In the same way we can construct Φ− : Mσ−(v0) 99K Mσ−(v). Then the composition Φ− ◦ Φ0 ◦ Φ−1 + : Mσ+ (v) 99K Mσ−(v) is a stratum preserving birational map which is isomorphic in codimension one, as desired. (cid:3) Lemma 6.10. Let v be any class of O'Grady type and assume that (LGU) holds for v. Then π+ is a divisorial contraction of Li-Gieseker-Uhlenbeck type. 34 CIARAN MEACHAN AND ZIYU ZHANG Proof. By Lemma 6.4, v is a minimal class and we can assume Mσ0(w) = M st σ0(w) for the class w in (LGU). We can simply follow the proof of [BM14a, Lemma 8.7] to get the divisor contracted by π+, and it is clear from there that the contraction is of Li-Gieseker-Uhlenbeck type. (cid:3) Lemma 6.11. Let v be any class of O'Grady type and assume (LGU) holds for v. Then there is a derived (anti-)autoequivalence Φ of D(X) inducing a stratum preserving birational map Φ : Mσ+ (v) 99K Mσ−(v), which is an isomorphism in codimension one. Proof. The proof of this lemma is similar to that of Lemma 6.9. However, in this situation, while W is a Li-Gieseker-Uhlenbeck type divisorial wall for the class v, it is a Hilbert-Chow type totally semistable divisorial wall for the class vp. By Lemma 6.4, v and hence vp, are minimal classes. The proof of [BM14a, Theorem 1.1(b)] in the Li-Gieseker-Uhlenbeck case shows that a derived anti-autoequivalence Φ (which is in fact a spherical twist) induces a birational map Φ : Mσ+(v) 99K Mσ−(v) which is isomorphic in codimension one. The same proof as in the Hilbert-Chow case shows that Φ also induces a birational map Φp : Mσ+ (vp) 99K Mσ−(vp). Therefore by Lemma 2.5, Φ is also a stratum preserving birational map. (cid:3) We conclude the discussion about divisorial walls by summarising all the above lemmas and prov- ing the following proposition, which is nothing but the collection of statements about divisorial contractions in Theorems 5.3 and 5.4. Proposition 6.12. The potential wall W associated to a rank two hyperbolic lattice H ⊂ H ∗ alg(X, Z) containing v induces a divisorial contraction if and only if at least one of the following two conditions holds (BN) there exists a spherical class s with (s, v) = 0, or (LGU) there exists an isotropic class w with (w, v) = 2. Moreover, the condition (LGU) necessarily implies the condition (BN). In either case, we can find an (anti-)autoequivalence Φ of D(X), which induces a birational map Φ : Mσ+ (v) 99K Mσ−(v). It is stratum preserving and an isomorphism in codimension one. Proof. Lemma 6.4 shows that (LGU) implies (BN). Therefore it suffices to show W induces a divisorial contraction if and only if (BN) holds. The necessity of (BN) is proved in Lemma 6.6 for minimal classes and in Lemma 6.7 for non-minimal classes by contradiction. The sufficiency of (BN), as well as the type of contraction, are the contents of Lemmas 6.8 and 6.10. Finally, the birational map Φ is constructed in Lemma 6.9 for Brill-Noether contractions and in Lemma 6.11 for Li-Gieseker-Uhlenbeck contractions. (cid:3) 6.3. Walls inducing small contractions or no contractions. In this subsection, we prove Theorems 5.3 and 5.4 in the case that W induces a small contraction or no contraction at all. We point out that most of the results in this subsection work for all effective classes v of divisibility two, because the existence of a symplectic resolution will not be used in the proof. As before, this subsection consists mainly of a series of lemmas; each of which will become an integral part in the proof of Theorems 5.3 and 5.4. All the results will be summarised in Propositions 6.21 and 6.22 at the end. These two propositions, together with Proposition 6.12, cover all the cases in Theorems 5.3 and 5.4. Lemma 6.13. Let M = Mσ(v) be a moduli space of O'Grady type. Then there are no small contractions arising from a decomposition of v into the sum a + b of two positive classes. BIRATIONAL GEOMETRY OF SINGULAR MODULI SPACES OF O'GRADY TYPE 35 Proof. Suppose we have such a decomposition for v. Then 8 = v2 = a2 + 2(a, b) + b2 > 2(a, b) =⇒ (a, b) 6 4. If we suppose further that W does not induce a divisorial contraction, then just as in the proof of [BM14a, Proposition 9.1], we can assume that (a, b) > 2. This leaves (a, b) = 3 or 4 as the only cases. Next, let us observe that we also have the following equation: 8 = v2 = (v, a + b) = (v, a) + (v, b) = 2(vp, a) + 2(vp, b). That is, (v, a) and (v, b) must be even integers. In particular, the following inequality (v, a) = (a + b, a) = a2 + (a, b) > (a, b) > 2 implies that (v, a) = 4; and hence (v, b) = 4 as well. In other words, we are left with two cases when 4 = (v, a) = a2 + (a, b): • If (a, b) = 3 then a2 = 1; contradicting the fact that the lattice is even. • If (a, b) = 4 then a2 = 0 which implies s := vp − a is a spherical class with (s, v) = 0. Hence, by Proposition 6.12, we must be on a BN-wall which contradicts our assumption that W does not induce a divisorial contraction. Thus, writing v as the sum of two positive classes a and b never gives rise to a flopping wall. (cid:3) p = 2. Remark 6.14. Notice that Lemma 6.13 makes our classification of small contractions considerably easier than that of [BM14a]. In particular, for a two-divisible Mukai vector, the parallelogram described in [BM14a, Lemma 9.3] always contains an extra interior lattice point vp. Even though it is possible to modify the proof of [BM14a, Lemma 9.3] in order to deal with two-divisible Mukai vectors, we will not need it here when v2 Lemma 6.15. Let v be a minimal class of O'Grady type, and assume that W does not induce a divisorial contraction for v. If (SC) does not hold for v, then W is not a wall for v. σ±(v) ⊔ Sym2 Mσ±(vp). Proof. We use the decomposition of the moduli space Mσ±(v) = M st Notice that vp is minimal since v is, and so by Lemma 6.13 and [BM14a, Proposition 9.4], we know that W is not a wall for vp. In particular, we can identify the strictly semistable loci Sym2 Mσ+(vp) = Sym2 Mσ−(vp). For the stable loci, we observe that the proof of [BM14a, Proposition 9.4] can be applied without change to show that every σ+-stable object of class v is also σ0-stable, and hence σ−-stable. (In fact, the only extra point is that any σ−-unstable object of class v cannot have two HN-factors of class vp.) The same is true for σ−-stable objects. Therefore, we can also identify the stable loci M st σ−(v) and conclude that W is not a wall for v. Lemma 6.16. Let v be a non-minimal class of O'Grady type, and assume that W does not induce a divisorial contraction for v. If (SC) does not hold for v, then W is a fake wall for v. Moreover, there exists a derived autoequivalence Φ of D(X), which induces an isomorphism Φ : Mσ+ (v) ∼= Mσ−(v). σ+(v) ∼= M st (cid:3) Proof. Let v0 be the minimal class in the same GH-orbit as v. Then as in the proof of [BM14a, Proposition 9.4], we consider the derived autoequivalence Φ+ of D(X) (which is in fact a com- position of spherical twists) constructed in [BM14a, Proposition 6.8]. On the one hand, [BM14a, Proposition 9.4] shows that Φ+ induces an isomorphism Φ+,p : Mσ+(v0,p) ∼−→ Mσ+ (vp) on the moduli spaces with primitive classes. Since Φ+ preserves exten- ∼−→ Sym2 Mσ+(vp) on the strictly sions, it also induces an isomorphism Φ+ : Sym2 Mσ+(v0,p) semistable loci of the moduli spaces with non-primitive classes. Moreover, since the S-equivalence 36 CIARAN MEACHAN AND ZIYU ZHANG relation with respect to σ0 is trivial on Mσ+ (v0,p), it is also trivial on Mσ+(vp), and hence on Sym2 Mσ+(vp), which implies that no curve in Sym2 Mσ+ (vp) is contracted by π+. σ+ (v0) → M st On the other hand, Lemma 6.15 ensures that every σ+-stable object of class v0 is also σ0-stable and so by Lemma 6.3, we have an injective morphism Φ+ : M st σ+(v) whose image does not contain any curve contracted by π+. Now we combine the stable and strictly semistable loci to get an injective morphism Φ+ : Mσ+(v0) → Mσ+(v). Notice that Φ+ is an isomorphism from Mσ+ (v0) to its image in Mσ+ (v) because the inverse derived autoequivalence Φ−1 + induces an inverse of the morphism. Finally, since both moduli spaces are irreducible projective varieties of the same dimension, Φ+ is in fact an isomorphism between the two moduli spaces. Moreover, by the above discussion, no curve in Mσ+ (v) is contracted by π+. By Theorem 5.1, we know that W is totally semistable for v and so it must be a fake wall. Similarly we have a derived autoequivalence Φ− of D(X) inducing an isomorphism Φ− : Mσ−(v0) ∼−→ Mσ−(v). Since W is not a wall for v0, the composition Φ− ◦ Φ−1 desired isomorphism between the two moduli spaces. Lemma 6.17. Let v be a Mukai vector of O'Grady type and assume that W does not induce a divisorial contraction for v. If v (or rather vp) satisfies the following boundary condition: ∼−→ Mσ−(v) gives the + : Mσ+ (v) (cid:3) (BC) there exists a spherical class s ∈ H with 0 < (s, vp) 6 v2 p 2 then W induces a small contraction for v. Proof. First let us observe that a similar argument to the one used in the proof of Lemma 6.13 shows that no small contractions can come from a decomposition of vp into positive classes. Now, if (BC) holds, then by [BM14a, Proposition 9.1], W induces a (divisorial or small) contraction for the primitive class vp. This implies W induces a contraction in the strictly semistable locus Sym2 Mσ+(vp) ⊂ Mσ+ (v), which is a small contraction considered in Mσ+(v). Lemma 6.18. Let v be a Mukai vector of O'Grady type. Assume that W does not induce a divisorial contraction for v, and condition (BC) does not hold for v. If there exists an effective spherical class s with 0 < (s, v) 6 v2 2 , then W induces a small contraction for v. (cid:3) Proof. We first observe that since (BC) does not hold, we in fact have (s, v) = 2(s, vp) > v2 p. Next, following the notation in [BM14a, Proof of Proposition 9.1], we just need to show that and F do not fall in a single S-equivalence class with respect to σ+. To show the inequality, p + 2 or p, depending on whether E1 and E2 are isomorphic or not. In either case, the inequality is (cid:3) ext1(eS, F ) > ext1(E1, E2). This will ensure that all objects obtained by taking extensions of eS we observe that ext1(eS, F ) > (s, v − s) = (s, v) + 2 > v2 v2 satisfied and therefore, the contraction is justified. Lemma 6.19. Let v be a Mukai vector of O'Grady type. Assume that W does not induce a If there exists a non- divisorial contraction for v, and condition (BC) does not hold for v. effective spherical class s with 0 < (s, v) 6 v2 p + 2, while ext1(E1, E2) = v2 2 , then W induces a small contraction for v. Proof. As in Lemma 6.18, the failure of (BC) implies that (s, v) > v2 p. We set t = −s (which is effective) and follow the third and fourth paragraphs in the proof of [BM14a, Proposition 9.1]. There are two steps in the proof which require further justification. As before, we need to show that all the extensions constructed from a fixed F and eT as in [BM14a, Proposition 9.1] cannot lie in the same S-equivalence class with respect to σ+ by a BIRATIONAL GEOMETRY OF SINGULAR MODULI SPACES OF O'GRADY TYPE 37 dimension comparison. The proof of [BM14a, Proposition 9.1] shows that (t, v′) = (s, v) + 2 and all extensions constructed there are parametrised by a Grassmannian with dim Gr((s, v) + 1, ext1(eT , F )) > dim Gr((s, v) + 1, (s, v) + 2) = (s, v) + 1 > v2 p + 1. On the other hand, the dimension of the S-equivalence class of extensions of σ+-stable objects E1 and E2 of class vp is given by ext1(E1, E2)− 1 = v2 p− 1, depending on whether E1 and E2 are isomorphic or not. In either case, we have the desired inequality and the construction in [BM14a, Proposition 9.1] yields a proper contraction under the given assumption. (cid:3) Lemma 6.20. Let v be any class of O'Grady type and assume that W induces a small con- traction for v. Then there is a derived autoequivalence Φ of D(X), which induces a stratum preserving birational map Φ : Mσ+(v) 99K Mσ−(v), which is an isomorphism in codimension one. p + 1 or v2 Proof. When v is a minimal class, we claim that we can simply take Φ to be the identity functor on D(X). Since W induces a small contraction, Lemma 6.6 tells us that Mσ+ (v) and Mσ−(v) contain a common open subset M st σ0(v), with complement of codimension at least two in either moduli space. Therefore, the identify functor induces a birational map from Mσ+ (v) to Mσ−(v) which is an isomorphism in codimension one. Notice that the assumptions imply that W is not a totally semistable wall for the primitive class vp and so the identity functor also induces a birational map between Mσ+(vp) and Mσ−(vp). By Lemma 2.5, we conclude that the birational map induced by the identify functor is stratum preserving. As usual, when v is not a minimal class, we denote the corresponding minimal class by v0. Let Φ0 : Mσ+(v0) 99K Mσ−(v0) be the birational map induced by the identity functor on D(X). By the discussion above, we know that it is stratum preserving and an isomorphism in codimension one. Now we take Φ+ to be the composition of spherical twists constructed in [BM14a, Proposi- tion 6.8]. By Lemma 6.3, Φ+ induces a birational map Φ+ : Mσ+ (v0) 99K Mσ+ (v), such that its restriction on the open subset M st σ0(v0) is an injective morphism. Moreover, since W is not to- tally semistable for v0,p, Φ+ also induces a birational morphism Φ+,p : Mσ+ (v0,p) 99K Mσ+ (vp). Thus, by Lemma 2.5, the birational map induced by Φ+ is stratum preserving. Since M st σ0(v0) has a complement of codimension at least two in Mσ+(v0), we conclude that Φ+ is an isomor- phism in codimension one by Proposition 2.6. We can follow the same procedure to get a derived autoequivalence Φ− which induces a birational map Φ− : Mσ−(v0) 99K Mσ−(v) satisfying both requirements. Then Φ− ◦ Φ0 ◦ Φ−1 + is the autoequivalence which induces the desired birational map. (cid:3) We conclude the discussion by summarising all the lemmas above with the following two propo- sitions; which verify the collection of statements about small contractions and no contractions in Theorems 5.3 and 5.4. Proposition 6.21. The set W is a wall inducing a small contraction if and only if it does not induce any divisorial contraction, and the following condition holds: (SC) there exists a spherical class s with 0 < (s, v) 6 v2 2 . Moreover, we can find an autoequivalence Φ of D(X), which induces a birational map Φ : Mσ+(v) 99K Mσ−(v). It is stratum preserving and an isomorphism in codimension one. Proof. The necessity of (SC) is proved in Lemma 6.15 for minimal classes and in Lemma 6.16 for non-minimal classes by contradiction. The sufficiency is proved in Lemma 6.17 when the extra condition (BC) holds and Lemmas 6.18 and 6.19 when (BC) does not hold and (SC) holds. Finally, the birational map is constructed in Lemma 6.20. (cid:3) 38 CIARAN MEACHAN AND ZIYU ZHANG Proposition 6.22. In the case that W does not induce any contractions at all, we can find an autoequivalence Φ of D(X), which induces an isomorphism Φ : Mσ+ (v) → Mσ−(v). In particular, Φ is just the identity functor when v is a minimal class. In other words, W is a fake wall if v is non-minimal, or is not a wall if v is minimal. Proof. This is contained in Lemma 6.15 for minimal classes and Lemma 6.16 for non-minimal classes. (cid:3) To conclude, we observe that all the cases in Theorems 5.3 and 5.4 have been covered by Proposition 6.12 (in the case of divisorial contractions), Proposition 6.21 (in the case of small contractions), and Proposition 6.22 (in the case of no contractions at all). 7. The Global Bayer-Macr`ı Map In this section we introduce the global Bayer-Macr`ı map and show how it can be used to study the birational geometry of the singular moduli spaces. We start by discussing its construction. 7.1. Construction of the global Bayer-Macr`ı map. We use our Theorems 5.3 and 5.4 to glue the local Bayer-Macr`ı maps defined on chambers to a global map on Stab†(X). The key step is to prove a compatibility result between any two local Bayer-Macr`ı maps ℓC+ and ℓC− which are defined on adjacent chambers C+ and C− separated by a wall W. The following is a generalisation of [BM14a, Lemma 10.1]. Lemma 7.1. The maps ℓC+ and ℓC− agree on the wall W. More precisely, we have • If W induces a divisorial contraction, then the analytic continuations of ℓC+ and ℓC− differ by the reflection of NS(Mσ+(v)) (or NS(Mσ−(v))) at the divisor D contracted by ℓσ0; • In all other cases, the analytic continuations of ℓC+ and ℓC− agree with each other. Proof. Although the Mukai vector v is not primitive in our situation, a (quasi-)universal family still exists on the stable locus M st σ+(v)) which has a complement of codimension two. Therefore, the proof for primitive Mukai vectors in [BM14a, Lemma 10.1] still works without any changes. In fact, our situation is even easier since Hilbert-Chow wall do not exist any more. (cid:3) σ+(v) (and M st As pointed out in [BM14a], we conclude from this lemma that the moduli spaces Mσ±(v) for the two adjacent chambers are isomorphic when W induces a divisorial contraction or no contraction. If W induces a small contraction, then these moduli spaces differ by a flop; this is why these walls are called flopping walls in [BM14b, BM14a]. Remark 7.2. By Theorem 5.4 and Lemma 7.1, the local Bayer-Macr`ı map ℓC : C → MC(v) defined on each chamber C ⊂ Stab†(X) can be glued together to give a continuous map on Stab†(X). By Remark 5.6, we can denote the global Bayer-Macr`ı map by ℓ : Stab†(X) → NS(M (v)). 7.2. Birational geometry via the global Bayer-Macr`ı map. We follow the approach in [BM14a, Section 10] to study the global properties of the Bayer-Macr`ı map. This map allows us to prove a precise relationship between the birational geometry of the moduli spaces and wall-crossing in the stability manifold. Before we state the main theorem of the paper, we need three more lemmas; the first of which partly describes the image of the global Bayer-Macr`ı map. BIRATIONAL GEOMETRY OF SINGULAR MODULI SPACES OF O'GRADY TYPE 39 Lemma 7.3. The image of the global Bayer-Macr`ı map is contained in Big(M (v))∩ Mov(M (v)). Proof. Let σ ∈ Stab†(X) be a generic stability condition with respect to v. Then Proposition 4.1 shows that ℓσ is ample on Mσ(v) and is therefore a big and movable class. By Proposition 4.2 or Proposition 5.2, ℓσ0 is also a big and movable class, because it induces a birational morphism π+ : Mσ+(v) → M +. (cid:3) The second lemma tells us that each open chamber in NS(M (v)) which represents the ample cone of a certain birational model is either completely contained in the image of ℓ, or has no intersection with the image of ℓ at all. Lemma 7.4. For any generic stability condition σ ∈ Stab†(X), the ample cone Amp(Mσ(v)) of the moduli space Mσ(v) is contained in the image of the global Bayer-Macr`ı map ℓ. Proof. By Proposition 4.1, we know that ℓσ ∈ Amp(Mσ(v)). Hence there is at least one point in Amp(Mσ(v)), which lies in the image of ℓ. For any other point α ∈ Amp(Mσ(v)) we can appeal to Proposition 2.12 to see that α ∈ Pos(Mσ(v)). Hence, by Corollary 2.8, we have α = θσ(a) for some class a ∈ v⊥ with a2 > 0. Now we use the argument in [BM14a, Proof of Theorem 1.2 (a)(b)(c)] to conclude that α also lies in the image of ℓ. (cid:3) Now we state the third and final lemma. If a wall W of a chamber C in Stab†(X) induces a contraction of the moduli space MC(v), then the image ℓ(W) ⊂ NS(M (v)) of W under the global Bayer-Macr`ı map ℓ : Stab†(X) → NS(M (v)) is a wall of the nef cone Nef(MC(v)). However, if W is a fake wall, then its image under the global Bayer-Macr`ı map is not a wall in NS(M (v)). This is the content of the following result. Lemma 7.5. Let σ0 be a generic stability condition on a fake wall W. Then its image lies in the interior of the nef cone Nef(Mσ+ (v)) of the moduli space Mσ+(v). The same statement is true for Mσ−(v). Proof. This is just [BM14a, Proposition 10.3]; whose proof works regardless of whether v is primitive or not. (cid:3) We can now state the main theorem of the paper, which crystallises the relation between the birational geometry of singular moduli spaces of O'Grady type and wall crossings in the stability manifold Stab†(X). It is a slight generalisation of [BM14a, Theorem 1.2] in the case of singular moduli spaces which admit symplectic resolutions. Theorem 7.6. Let v be a Mukai vector of O'Grady type and σ ∈ Stab†(X) a generic stability condition with respect to v. Then (1) We have a globally defined continuous Bayer-Macr`ı map ℓ : Stab†(X) → NS(Mσ(v)), which is independent of the choice of σ. Moreover, for any generic stability condition τ ∈ Stab†(X), the moduli space Mτ (v) is the birational model corresponding to ℓτ . (2) If C ⊂ Stab†(X) is the open chamber containing σ, then ℓ(C) = Amp(Mσ(v)). (3) The image of ℓ is equal to Big(Mσ(v)) ∩ Mov(Mσ(v)). In particular, every K-trivial Q-factorial birational model of Mσ(v) which is isomorphic to Mσ(v) in codimension 1 appears as a moduli space Mτ (v) for some generic stability condition τ ∈ Stab†(X). Proof. All the ingredients of the proof have already been presented above. For (1), the exis- tence of the global Bayer-Macr`ı map is contained in Remark 7.2. The ampleness statement in 40 CIARAN MEACHAN AND ZIYU ZHANG Proposition 4.1 implies that Mτ (v) is the birational model corresponding to ℓτ for any generic τ . For (2), we first realise that ℓ(C) has full dimension by Lemma 7.4 and hence must be an open subset of Amp(Mσ(v)) by Proposition 4.1. Now we know that the image of every non-fake wall of C has to be a boundary component of Amp(Mσ(v)) whereas Lemma 7.5 ensures that the image of a fake wall of C never separates Amp(Mσ(v)) into two parts. Therefore, the image of C under ℓ is the whole ample cone. For the first claim in (3), one direction of inclusion is in Lemma 7.3. We just need to show that every open chamber of Big(Mσ(v)) ∩ Mov(Mσ(v)) regarded as the ample cone of some birational model of Mσ(v) is contained in the image of ℓ. Assume we have two such ample cones adjacent to each other, one of which lies in the image of ℓ but not the other. Then the wall separating the two ample cones must be the image ℓ(W) of a wall W in Stab†(X). By Lemma 7.5, W cannot be a fake wall since fake walls are not mapped to walls in the movable cone by ℓ. By Lemma 7.1, W cannot be a flopping wall since the image of ℓ would extend to the other side of ℓ(W). Hence W must be a divisorial wall. Then we claim that the image of W has reached the boundary of Mov(Mσ(v)). In fact, since ℓσ0 induces a divisorial contraction for a generic σ0 ∈ W, ℓσ0 has degree zero on each curve contained in the fibres of this contraction, which implies that any line bundle on the other side of ℓ(W) must have negative degree on these curves contained in fibres. Therefore, its base locus contains the entire contracted divisor. The second claim follows immediately from the surjectivity. (cid:3) Remark 7.7. Careful readers might have found that our proof of surjectivity of the Bayer-Macr`ı map onto the intersection of the big and movable cones is slightly different from that in [BM14a, Proof of Theorem 1.2 (b)]. More precisely, we avoided using the statement that any big and movable class is strictly positive, which is well-known for irreducible holomorphic symplectic manifolds, but not for moduli spaces of O'Grady type. However, using [Bri08, Theorem 1.1], our Theorem 7.6 implies that the same statement is still true for singular moduli spaces of O'Grady type. 7.3. Torelli theorem for singular moduli spaces of O'Grady type. In analogy with [BM14a, Corollary 1.3], we can prove a Torelli-type theorem for the singular moduli spaces of O'Grady type. It gives a Hodge-theoretic criterion for the existence of stratum-preserving birational maps between these moduli spaces. We use the same notation as in [BM14a], which was originally introduced by Mukai: the total cohomology H ∗(X, Z) of a K3 surface X carries a weight two Hodge structure which is polarised by the Mukai pairing. We write v⊥,tr ⊂ H ∗(X, Z) for the orthogonal complement of v in the total cohomology. Corollary 7.8. Let X and X ′ be two smooth projective K3 surfaces with v ∈ H ∗ alg(X, Z) and If σ ∈ Stab†(X) and σ′ ∈ Stab†(X ′) are v′ ∈ H ∗ generic stability conditions with respect to v and v′ respectively, then the following statements are equivalent alg(X ′, Z) Mukai vectors of O'Grady type. (a) There is a stratum preserving birational map ϕ : MX,σ(v) 99K MX ′,σ′(v′). (b) The embedding v⊥,tr ⊂ H ∗(X, Z) of the integral weight two Hodge structures is isomor- phic to the embedding v′⊥,tr ⊂ H ∗(X ′, Z). (c) There is an (anti-)autoequivalence Φ from D(X) to D(X ′) with Φ∗(v) = v′. (d) There is an (anti-)autoequivalence Ψ from D(X) to D(X ′) with Ψ∗(v) = v′ which induces a stratum preserving birational map ψ : MX,σ(v) 99K MX ′,σ′(v′). BIRATIONAL GEOMETRY OF SINGULAR MODULI SPACES OF O'GRADY TYPE 41 Proof. The proof of [BM14a, Corollary 1.3] already contains most of what we need in our situ- ation. We only point out the differences. For (a) ⇒ (b), we observe that MX,σ(vp) ⊂ MX,σ(v) p) ⊂ MX ′,σ′(v′) are the deepest singular strata and so the stratum preserving bi- and MX ′,σ′(v′ rational map ϕ in (a) restricts to a birational map Mσ(vp) 99K Mσ′(v′ p). Therefore, by [BM14a, Corollary 1.3], we can deduce (b). For (b) ⇒ (c), the proof is already in [BM14a, Corollary 1.3]. For (c) ⇒ (d), we can follow the proof of [BM14a, Corollary 1.3] and replace Φ if necessary, so that there exists a generic τ ∈ Stab†(X ′) with Mσ(v) ∼= Mτ (v′). By Theorem 5.4, we can find an (anti-)autoequivalence Φ′ of D(X ′) which induces a stratum preserving birational map Mτ (v′) 99K Mσ′(v′). And the composition Φ′ ◦ Φ does the work. The final part (d) ⇒ (a) is trivial. (cid:3) 7.4. Lagrangian Fibrations. An immediate consequence of the birationality of wall-crossing is the following result: Theorem 7.9. Let M = Mσ(v) be a moduli space of O'Grady type, π : fM → M its symplectic resolution and q(−,−) its Beauville-Bogomolov form. If there is an integral divisor class D with q(D) = 0 then there exists a birational moduli space M ′ = Mσ′ (v′) of O'Grady type whose resolution fM ′ admits a Lagrangian fibration. Proof. This follows from the arguments in [BM14a, Section 11]. Indeed, we just have to replace [BM14a, Theorem 1.2] with our Theorem 7.6 in the proof of [BM14a, Theorem 1.5]. (cid:3) Remark 7.10. Bayer and Macr`ı are able to make a stronger statement in their situation if the divisor class is also assumed to be nef; see [BM14a, Conjecture 1.4(b)]. However, for an analogous statement to hold true for moduli spaces of O'Grady type, we would need to find a replacement for [BM14a, Proposition 3.3] (which is a summary of Markman's results). In particular, we would need to know that the cone Mov(M ) ∩ Pos(M ) of big and movable divisors on M was equal to the fundamental chamber of the Weyl group action on the positive cone Pos(M ) of M . Remark 7.11. There might be square-zero classes on fM which are not the pullback of square-zero we see no reason why the existence of a Lagrangian fibration fM → P5 should guarantee the classes on M and so it is not clear whether the converse of Theorem 7.9 is true. In particular, existence of a square-zero class on M . Remark 7.12. A natural question at this point is whether one can combine the recent techniques of Bayer-Hassett-Tschinkel [BHT13] and Markman [Mar11, Mar13] with similar arguments of [BM14a, Section 12] to obtain a description of the Mori cone of the symplectic resolution fM of a moduli space M = Mσ(v) of O'Grady type. The authors plan to return to this question in future work. 8. Examples of Movable and Nef cones In this section, we examine examples of cones of divisors on moduli spaces of O'Grady type. Let us suppose for simplicity that Pic(X) = Z[H] with H 2 = 2d. We set v = 2vp = (2, 0,−2) and M := MH (v). Then a basis for NS(M ) is given by where θσ : v⊥ ∼−→ NS(M ) is the isometry from Corollary 2.8(3). By Theorem 5.3, divisorial contractions are divided into two cases: eH = θσ(0,−H, 0) and B = θσ(−1, 0,−1) 42 CIARAN MEACHAN AND ZIYU ZHANG (BN) there exists a spherical class s with (s, v) = 0, or (LGU) there exists an isotropic class w with (w, v) = 2. Just as in [BM14a, Section 13], we can solve the following system of equations {s2 = −2, (s, v) = 0 : s = (r, cH, s)} to see that the case of BN-contractions are governed by solutions to the following Pell's equation: x2 − dy2 = 1 via s = (x,−yH, x). (8.1) Similarly, we can solve to see that the case of LGU-contractions are governed by solutions to {w2 = 0, (w, v) = 2 : w = (r, cH, s)} x2 − dy2 = 1 via w =(cid:18) x + 1 2 yH 2 , ,− 2 (cid:19) or w = (x + 1,−yH, x) x − 1 (8.2) depending on whether y is even or odd. The two equations determine the movable cone: Proposition 8.1. Assume Pic(X) = Z[H] with H 2 = 2d. The movable cone of M = MH (2, 0,−2) has the following form: (a) If d = k2 h2 , with k, h > 1, gcd(k, h) = 1, then Mov(M ) = heH, eH − k h Bi, where q(heH − kB) = 0. (b) If d is not a perfect square, and (8.1) has a solution, then where (x1, y1) is the solution to (8.1) with x1, y1 > 0, and with smallest possible x1. Mov(M ) = heH, eH − d y1 x1 Bi, (c) If d is not a perfect square, and (8.1) has no solution, then Mov(M ) = heH, eH − d y′ 1 x′ 1 Bi, where (x′ 1, y′ 1) is the solution to (8.2) with smallest possible y′ > 0. 1 x′ 1 Proof. Setting n = 2 in the proof of [BM14a, Proposition 13.1] is sufficient here. Indeed, their proof relies on [BM14a, Theorem 12.3] whose proof also goes through in our case when we replace [BM14a, Theorem 1.2] with Theorem 7.6. (cid:3) Example 8.2. If d = 1 or 2, then we are in case (a) or (b) of Proposition 8.1 and we have Mov(M ) = heH, eH − Bi or Mov(M ) = heH, eH − 4 3 Bi respectively. By Theorem 5.3, we have a flopping wall if and only if: (SC) there exists a spherical class s with 0 < (s, v) 6 v2 2 = 4. BIRATIONAL GEOMETRY OF SINGULAR MODULI SPACES OF O'GRADY TYPE 43 Notice that the condition becomes 0 < (s, v) = 2(s, vp) 6 4 and so there is only enough room for two walls of this kind: either (s, v) = 2 or 4. Solving the following system of equations {s2 = −2, (s, v) = 2 or 4 : s = (r, cH, s)} shows that flopping walls of type (SC2) are governed by solutions of x2 − 4dy2 = 5 and x2 − dy2 = 2 respectively. (8.3) The associated spherical classes are s =(cid:0) x+1 Lemma 8.3. Let M = MH (2, 0,−2). The nef cone of M has the following form: 2 (cid:1) and s = (x+1,−yH, x−1) respectively. 2 ,−yH, x−1 (a) If (8.3) has no solutions, then Nef(M ) = Mov(M ). (b) Otherwise, let (x1, y1) be the positive solution of (8.3) with x1 minimal. Then where (x1, y1) is the solution to (8.1) with x1, y1 > 0, and with smallest possible x1. Nef(M ) = heH, eH − 2d y1 x1 Bi. Proof. We apply [BM14a, Theorem 12.1] whose proof also works in our case once we replace [BM14a, Theorems 1.2 and 5.7] with Theorems 7.6 and 5.3 respectively. The movable cone and the nef cone agree unless there is a flopping wall, described in Theorem 5.3. By Lemma 6.13, we know that the case v = a + b with a, b positive is impossible. This leaves only the case of a spherical class s with (s, v) = 2 or 4; these exist if and only if (8.3) has a solution. (cid:3) Example 8.4. If d = 1, then x2 − y2 = 2 has no solutions and the only positive solution of x2 − 4y2 = 5 is given by (x, y) = (3, 1) which gives rise to the spherical class s = (2,−H, 1). In order to compute the generators of the nef cone of M , we need an element of v⊥ ∩ s⊥. For example, (2,−3H, 2) is perpendicular to both v and s and can be expressed in our basis for NS(M ) as 3eH − 2B or equivalently eH − 2 3 B. In particular, we have shown that Nef(M ) = heH, eH − 2 3 Bi. Similarly, if d = 2 then x2 − 8y2 = 5 has no solutions and the minimal solution of x2 − 2y2 = 2 is given by (x, y) = (2, 1). Therefore, Remark 8.5. Observe that when (v, s) = 4, we have another spherical class t := v − s with (s, t) = 6 and so this wall is flopping the natural Lagrangian P5 ≃ Ext1(S, T ) inside Mσ0(v) Nef(M ) = heH, eH − Bi. P5 ≃ P Ext1(S, T ) Mσ0 (v) Mσ0(v − s) ≃ Mσ0 (t) ≃ {pt}. However, when (v, s) = 2 we have another spherical class given by t := vp − s with (s, t) = 3. In particular, (t, v − s) < 0 and so by Theorem 5.1, the wall is totally semistable for v − s; similar arguments show that the wall is totally semistable for v − t as well. This means the flopping locus cannot be described as a projective bundle. / / 44 CIARAN MEACHAN AND ZIYU ZHANG Example 8.6. If X is a genus two K3 surface with Pic(X) = Z[H] and H := [C] the class of a genus two curve C, then we can twist our calculations in Example 8.4 by OX(C) to get a description for the nef cone of MH (2, 2H, 0). Moreover, since d = 1 in this case, we know that there is only one flopping wall corresponding to the spherical class s = (2, H, 1). That is, the movable cone has two chambers and hence Mσ(v) has two 'birational models' which are exchanged by the birational map Φ∗ : Mσ(2, 2H, 0) 99K MΦ(σ)(0, 2H,−2) ; E 7→ Φ(E) induced by the spherical twist around OX . To see that the two models really are interchanged by the spherical twist Φ, let h be the upper half plane, β, ω ∈ NS(X)R with ω ample and consider the open subset h0 = h \ {β + iω ∈ h : hexp(β + iω), si = 0 when s is a (−2)-class}. If X has Picard number one then Stab†(X)/]GL+ 2 (R) can be identified with (the universal cover of) h0; see [BB13]. In particular, if X is a generic double cover of P2 with H 2 = 2 and d = 1, then we can set β := sH and ω := tH for (s, t) ∈ R × R>0 and use the formula in [Mea12, p.35] to see that the flopping wall of Mσ(2, 2H, 0) in h0 is given by a semicircle with centre and radius equal to (−1/2,√5/2). Furthermore, the cohomological transform ΦH : H ∗(X, Q) → H ∗(X, Q) implies that Φ : σ0,tH 7→ σ0, 1 t H and so we see that the non-Gieseker chamber with respect to v = (2, 2H, 0) gets identified with the Gieseker chamber with respect to Φ(v) = (0, 2H,−2). If we set eH ′ := θσ(0,−H,−2) and B′ := θσ(−1,−H,−2) (to be the classes eH and B twisted by OX (C)) then we can illustrate our observations as in Figure 1. The LGU wall and hence BN wall (Lemma 6.4) can be realised geometrically by considering the object E := Φ(Iy(−2C))[1] for some point y ∈ X with v(E) = Φ(1,−2H, 2)[1] = (2, 2H, 1) and then splitting off another point x ∈ X. More precisely, turning the torsion sequence F → E → Ox for F ∈ MH(2, 2H, 0) yields Ox → F[1] → E[1]. This is Mukai's morphism θv which contracts the projective space P Ext1(E[1],Ox) ≃ P1. That is, this wall will contract a divisor of non-locally free sheaves F, and the short exact sequence is (generically) induced by injection into a locally free sheaf E. Remark 8.7. The only difference with [BM14a] is that we cannot say the extremal ray eH−B gives rise to a Lagrangian fibration because Mσ(0, 2H,−2) is singular; we would need to precompose with the resolution π : fM → M to get this. Remark 8.8. Recall that the six dimensional O'Grady space [O'G03] is constructed in a very similar way to his ten dimensional space. Indeed, he considers a moduli space of sheaves on an abelian surface, rather than a K3 surface, and takes the fibre of the Albanese map over zero. Since there are no spherical objects on an abelian surface [Bri08, Lemma 15.1], we expect that the analogue of our classification of walls Theorem 5.3 would become a lot simpler in this case. In particular, Yoshioka [Yos12] has already shown that all of the machinery in [BM14b, BM14a] works on abelian surfaces and so if we assume our arguments also work in this case, then we would have to conclude that there are no flopping walls at all; c.f. [Mac14, Proposition 4.6]. That is, the movable cone Mov(Mσ(v)) and the nef cone Nef(Mσ(v)) would actually coincide in this case. It seems rather surprising that the six dimensional O'Grady space Mσ(v) should only have one birational model but this remark should only be regarded as speculation. [AB13] Daniele Arcara and Aaron Bertram. Bridgeland-stable moduli spaces for K-trivial surfaces. J. Eur. Math. Soc. (JEMS), 15(1):1–38, 2013. With an appendix by Max Lieblich. arXiv:0708.2247. References BIRATIONAL GEOMETRY OF SINGULAR MODULI SPACES OF O'GRADY TYPE 45 t MH(2, 2H, 0) 1 flop Mσ(2, 2H, 0) k MH(0, 2H,−2) −2 −1.5 −1 −0.5 0 0.5 s 1 Nef(MH (0, 2H,−2)) Nef(MH (2, 2H, 0)) eH ′ − B′ eH ′ BN=LGU wall flop 3 B′ eH ′ − 2 Mov(Mσ(2, 2H, 0)) = Nef(MH (2, 2H, 0)) ⊔ Nef(MH (0, 2H,−2)) Figure 1. The movable cone of Mσ(2, 2H, 0) when X is a genus two K3 surface. [BB13] [BCHM10] [Bea00] Arend Bayer and Tom Bridgeland. Derived automorphism groups of K3 surfaces of Picard rank 1. Arxiv preprint, 2013. arXiv:1310.8266. Caucher Birkar, Paolo Cascini, Christopher D. Hacon, and James McKernan. Existence of minimal models for varieties of log general type. J. Amer. Math. Soc., 23(2):405–468, 2010. arXiv:math/0610203. Arnaud Beauville. arXiv:math/9903070. Invent. Math., 139(3):541–549, singularities. Symplectic 2000. [BHT13] [BHPVdV04] Wolf Barth, Klaus Hulek, Chris Peters, and Antonius Van de Ven. Compact complex surfaces, volume 4 of Ergebnisse der Mathematik und ihrer Grenzgebiete. 3. Folge. A Series of Modern Surveys in Mathematics [Results in Mathematics and Related Areas. 3rd Series. A Series of Modern Surveys in Mathematics]. Springer-Verlag, Berlin, second edition, 2004. Arend Bayer, Brendan Hassett, and Yuri Tschinkel. Mori cones of holomorphic symplectic varieties of K3 type. Arxiv preprint, 2013. arXiv:1307.2291. Arend Bayer and Emanuele Macr`ı. MMP for moduli of sheaves on K3s via wall-crossing: nef and movable cones, Lagrangian fibrations. Invent. Math., 198(3):505–590, 2014. arXiv:1301.6968v3. Arend Bayer and Emanuele Macr`ı. Projectivity and birational geometry of Bridgeland moduli spaces. J. Amer. Math. Soc., 27(3):707–752, 2014. arXiv:1203.4613v2. Tom Bridgeland. Stability conditions on K3 surfaces. Duke Math. J., 141(2):241–291, 2008. arXiv:math/0307164. [BM14b] [BM14a] [Bri08] 46 [Cal00] [GHJ03] [HL10] [HS05] [HT09] [Huy99] [Huy03] [Kal06] [KLS06] [KM98] [Lie07] [Lo12] [LP14] [LQ14] [LS06] [Mac14] [Mar11] [Mar13] [Mea12] [MM13] [Muk84] [Muk87] [MYY11] [MYY14] [Nam01] [O'G99] [O'G03] CIARAN MEACHAN AND ZIYU ZHANG Cornell University, 2000. Caldararu. of twisted sheaves PhD Thesis, Derived categories Calabi-Yau online at on Andrei manifolds. Available http://www.math.wisc.edu/~andreic/publications/ThesisSingleSpaced.pdf. Mark Gross, Daniel Huybrechts, and Dominic Joyce. Calabi-Yau manifolds and related geometries. Universitext. Springer-Verlag, Berlin, 2003. Lectures from the Summer School held in Nordfjordeid, June 2001. Daniel Huybrechts and Manfred Lehn. The geometry of moduli spaces of sheaves. Cambridge Math- ematical Library. Cambridge University Press, Cambridge, second edition, 2010. Daniel Huybrechts and Paolo Stellari. Equivalences of twisted K3 surfaces. Math. Ann., 332(4):901– 936, 2005. arXiv:math/0409030. Brendan Hassett and Yuri Tschinkel. Moving and ample cones of holomorphic symplectic fourfolds. Geom. Funct. Anal., 19(4):1065–1080, 2009. arXiv:0710.0390. Daniel Huybrechts. Compact hyper-Kahler manifolds: basic results. Invent. Math., 135(1):63–113, 1999. arXiv:alg-geom/9705025. Daniel Huybrechts. The Kahler cone of a compact hyperkahler manifold. Math. Ann., 326(3):499– 513, 2003. arXiv:math/9909109. Dmitry Kaledin. Symplectic singularities from the Poisson point of view. J. Reine Angew. Math., 600:135–156, 2006. arXiv:math/0310186. Dmitry Kaledin, Manfred Lehn, and Christoph Sorger. Singular symplectic moduli spaces. Invent. Math., 164(3):591–614, 2006. arXiv:math/0504202. J´anos Koll´ar and Shigefumi Mori. Birational geometry of algebraic varieties, volume 134 of Cam- bridge Tracts in Mathematics. Cambridge University Press, Cambridge, 1998. With the collabora- tion of Herb (Charles) Clemens and Alessio Corti, Translated from the 1998 Japanese original. Max Lieblich. Moduli of twisted sheaves. Duke Math. J., 138(1):23–118, 2007. arXiv:math/0411337. Jason Lo. On some moduli of complexes on K3 surfaces. Arxiv preprint, 2012. arXiv:1203.1558v1. Christian Lehn and Gianluca Pacienza. On the log minimal model program for irreducible sym- plectic varieties. Arxiv preprint, 2014. arXiv:1405.5649. Jason Lo and Zhenbo Qin. Mini-walls for Bridgeland stability conditions on the derived category of sheaves over surfaces. Asian J. Math., 18(2):321–344, 2014. arXiv:1103.4352v1. Manfred Lehn and Christoph Sorger. La singularit´e de O'Grady. J. Algebraic Geom., 15(4):753–770, 2006. arXiv:math/0504182. Antony Maciocia. Computing the walls associated to Bridgeland stability conditions on projective surfaces. Asian J. Math., 18(2):263–279, 2014. arXiv:1202.4587v2. Eyal Markman. A survey of Torelli and monodromy results for holomorphic-symplectic varieties. In Complex and differential geometry, volume 8 of Springer Proc. Math., pages 257–322. Springer, Heidelberg, 2011. arXiv:1101.4606. Eyal Markman. Prime exceptional divisors on holomorphic symplectic varieties and monodromy reflections. Kyoto J. Math., 53(2):345–403, 2013. arXiv:0912.4981. Ciaran Meachan. Moduli of Bridgeland-stable objects. PhD thesis, University of Edinburgh, 2012. Available online at http://www.maths.ed.ac.uk/~cmeachan/Thesis.pdf. Antony Maciocia and Ciaran Meachan. Rank 1 Bridgeland stable moduli spaces on a principally polarized abelian surface. Int. Math. Res. Not. IMRN, 107(9):2054–2077, 2013. arXiv:1107.5304. Shigeru Mukai. Symplectic structure of the moduli space of sheaves on an abelian or K3 surface. Invent. Math., 77(1):101–116, 1984. Shigeru Mukai. On the moduli In Vector bun- dles on algebraic varieties (Bombay, 1984), volume 11 of Tata Inst. Fund. Res. Stud. Math., 1987. Available online at http://www.kurims.kyoto-u.ac.jp/~mukai/paper/Tata.pdf. Hiroki Minamide, Shintarou Yanagida, and K¯ota Yoshioka. Fourier-Mukai transforms and the wall- crossing behavior for Bridgeland's stability conditions. Arxiv preprint, 2011. arXiv:1106.5217v2. Hiroki Minamide, Shintarou Yanagida, and K¯ota Yoshioka. Some moduli spaces of Bridgeland's stability conditions. Int. Math. Res. Not. IMRN, (19):5264–5327, 2014. arXiv:1111.6187. Yoshinori Namikawa. Deformation theory of singular symplectic n-folds. Math. Ann., 319(3):597– 623, 2001. arXiv:math/0010113. Kieran O'Grady. Desingularized moduli spaces of sheaves on a K3. J. Reine Angew. Math., 512:49– 117, 1999. arXiv:alg-geom/9708009. Kieran O'Grady. A new six-dimensional 12(3):435–505, 2003. arXiv:math/0010187. irreducible symplectic variety. J. Algebraic Geom., Inst. Fund. Res., Bombay, space of bundles on K3 surfaces. I. pages 341–413. Tata BIRATIONAL GEOMETRY OF SINGULAR MODULI SPACES OF O'GRADY TYPE 47 [PR13] [Sim94] [Tod08] [Wie03] [Yos01] [Yos06] [Yos09] [Yos12] [YY14] Arvid Perego and Antonio Rapagnetta. Deformation of the O'Grady moduli spaces. J. Reine Angew. Math., 678:1–34, 2013. arXiv:1008.0190. Carlos T. Simpson. Moduli of representations of the fundamental group of a smooth projective variety. I. Inst. Hautes ´Etudes Sci. Publ. Math., (79):47–129, 1994. Yukinobu Toda. Moduli stacks and invariants of semistable objects on K3 surfaces. Adv. Math., 217(6):2736–2781, 2008. arXiv:math/0703590. Jan Wierzba. Contractions of symplectic varieties. J. Algebraic Geom., 12(3):507–534, 2003. arXiv:math/9910130. K¯ota Yoshioka. Moduli spaces of stable sheaves on abelian surfaces. Math. Ann., 321(4):817–884, 2001. arXiv:math/0009001. K¯ota Yoshioka. Moduli spaces of twisted sheaves on a projective variety. In Moduli spaces and arithmetic geometry, volume 45 of Adv. Stud. Pure Math., pages 1–30. Math. Soc. Japan, Tokyo, 2006. arXiv:math/0411538. K¯ota Yoshioka. Fourier-Mukai transform on abelian surfaces. Math. Ann., 345(3):493–524, 2009. arXiv:math/0605190. K¯ota Yoshioka. Bridgeland's stability and the positive cone of the moduli spaces of stable objects on an abelian surface. Arxiv preprint, 2012. arXiv:1206.4838v1. Shintarou Yanagida and K¯ota Yoshioka. Bridgeland's stabilities on abelian surfaces. Math. Z., 276(1-2):571–610, 2014. arXiv:1203.0884. School of Mathematics, University of Edinburgh, Scotland E-mail address: [email protected] Department of Mathematical Sciences, University of Bath, Bath BA2 7AY, United Kingdom E-mail address: [email protected]
1704.03358	2	1704	2017-11-25T02:06:46	Entire holomorphic curves into projective spaces intersecting a generic hypersurface of high degree	[ "math.AG", "math.CV" ]	In this note, we establish the following Second Main Theorem type estimate for every entire non-algebraically degenerate holomorphic curve $f\colon\mathbb{C}\rightarrow\mathbb{P}^n(\mathbb{C})$, in present of a {\sl generic} hypersuface $D\subset\mathbb{P}^n(\mathbb{C})$ of sufficiently high degree $d\geq 15(5n+1)n^n$: \[ T_f(r) \leq \,N_f^{[1]}(r,D) + O\big(\log T_f(r) + \log r \big)\parallel, \] where $T_f(r)$ and $N_f^{[1]}(r,D)$ stand for the order function and the $1$-truncated counting function in Nevanlinna theory. This inequality quantifies recent results on the logarithmic Green--Griffiths conjecture.	math.AG	math	Entire holomorphic curves into projective spaces intersecting a generic hypersurface of high degree Dinh Tuan HUYNH, Duc-Viet VU and Song-Yan XIE Abstract In this note, we establish the following Second Main Theorem type estimate for every algebraically nondegenerate entire curve f : C → Pn(C), in presence of a generic divisor D ⊂ Pn(C) of sufficiently high degree d ≥ 15(5n + 1)nn: for every r outside a subset of R of finite Lebesgue measure and every real positive constant δ, we have Tf (r) ≤ N [1] f (r, D) + O(cid:0) log Tf (r)(cid:1) + δ log r, where Tf (r) and N [1] theory. This inequality quantifies recent results on the logarithmic Green–Griffiths conjecture. f (r, D) stand for the order function and the 1-truncated counting function in Nevanlinna Keywords: Nevanlinna theory, Second Main Theorem, holomorphic curve, Green–Griffiths' conjecture, alge- braic degeneracy Mathematics Subject Classification 2010: 32H30, 32A22, 30D35, 32Q45, 1 Introduction and the main result We first recall the standard notation in Nevanlinna theory. Let E =Pi αi ai be a divisor on C where αi ≥ 0, ai ∈ C and let k ∈ N ∪ {∞}. Denote by ∆t the disk {z ∈ C, z < t}. Summing the k-truncated degrees of the divisor on disks by the truncated counting function at level k of E is then defined by taking the logarithmic average n[k](t, E) := Xai∈∆t N [k](r, E) := Z r 1 min {k, αi} n[k](t, E) t dt (t > 0), (r > 1). When k = ∞, we write n(t, E), N (r, E) instead of n[∞](t, E), N [∞](r, E). Let f : C → Pn(C) be an entire curve having a reduced representation f = [f0 : · · · : fn] in the homogeneous coordinates [z0 : · · · : zn] of Pn(C). Let D = {Q = 0} be a divisor in Pn(C) defined by a homogeneous polynomial Q ∈ C[z0, . . . , zn] of degree d ≥ 1. If f (C) 6⊂ D, we define the truncated counting function of f with respect to D as where (Q ◦ f )0 denotes the zero divisor of Q ◦ f . N [k] f (r, D) := N [k](cid:0)r, (Q ◦ f )0(cid:1), The proximity function of f for the divisor D is defined as where kQk is the maximum absolute value of the coefficients of Q and dθ 2π , 0 mf (r, D) := Z 2π log(cid:13)(cid:13)f (reiθ)(cid:13)(cid:13)d kQk (cid:12)(cid:12)Q(f )(reiθ)(cid:12)(cid:12) (cid:13)(cid:13)f (z)(cid:13)(cid:13) = max {f0(z), . . . , fn(z)}. Since(cid:12)(cid:12)Q(f )(cid:12)(cid:12) ≤ kQk · kf kd, one has mf (r, D) ≥ 0. 1 Lastly, the Cartan order function of f is defined by Tf (r) : = 2πZ 2π log(cid:13)(cid:13)f (reiθ)(cid:13)(cid:13) dθ = Z r t Z∆t f ∗ωn + O(1), 0 dt 1 1 where ωn is the Fubini–Study form on Pn(C). With the above notations, the Nevanlinna theory consists of two fundamental theorems (for a comprehensive presentation, see Noguchi-Winkelmann [19]). First Main Theorem. Let f : C → Pn(C) be a holomorphic curve and let D be a hypersurface of degree d in Pn(C) such that f (C) 6⊂ D. Then for every r > 1, the following holds whence (1.1) mf (r, D) + Nf (r, D) = d Tf (r) + O(1), Nf (r, D) ≤ d Tf (r) + O(1). Hence the First Main Theorem gives an upper bound on the counting function in terms of the order function. On the other side, in the harder part, so-called Second Main Theorem, one tries to establish a lower bound for the sum of certain counting functions. Such types of estimates were given in several situations. Throughout this note, for an entire curve f, the notation Sf (r) means a real function of r ∈ R+ such that there is a constant C for which Sf (r) ≤ C Tf (r) + δ log r for every positive constant δ and every r outside of a subset (depending on δ) of finite Lebesgue measure of R+. A holomorphic curve f : C → Pn(C) is said to be algebraically (linearly) nondegenerate if its image is not contained in any hypersurface (hyperplane). A family of q ≥ n + 1 hypersurfaces {Di}1≤i≤q in Pn(C) is in general position if any n + 1 hypersurfaces in this family have empty intersection: ∩ i∈I supp(Di) = ∅ (∀ I ⊂ {1,...,q}, I=n+1). We recall here the following classical result [3], with truncation level n. Cartan's Second Main Theorem. Let f : C → Pn(C) be a linearly nondegenerate holomorphic curve and let {Hi}1≤i≤q be a family of q > n + 1 hyperplanes in general position in Pn(C). Then the following estimate holds (q − n − 1) Tf (r) ≤ N [n] f (r, Hi) + Sf (r). qXi=1 In the one-dimensional case, Cartan recovered the classical Nevanlinna theory for nonconstant meromor- phic functions and families of q > 2 distinct points. Since then, many author tried to extend the result of Cartan to the case of (possible) nonlinear hypersurface. Eremenko-Sodin [12] established a Second Main Theorem for q > 2n hypersurfaces Di (1 ≤ i ≤ q) in general position in Pn(C) and for any nonconstant holomorphic curve f : C → Pn(C) whose image is not contained in ∪1≤i≤q supp(Di). Keeping the same assumption on q > n + 1 hypersurfaces, Ru [23] proved a stronger estimate for algebraically nondegenerate holomorphic curves f : C → Pn(C). He then extended this result to the case of algebraically nondegenerate holomorphic mappings into an arbitrary nonsingular complex projective variety [24]. Note that it remains open the question of truncating the counting functions in the above generalizations of Cartan's Second Main Theorem. Some results in this direction are obtained recently but one requires the presence of more targets, see for instance [1], [27]. In the other context, Noguchi-Winkelmann-Yamanoi [20] established a Second Main Theorem for alge- braically nondegenerate holomorphic curves into semiabelian varieties intersecting an effective divisor. Ya- manoi [28] obtained a similar result in the case of abelian varieties with the best truncation level 1, which is extended to the case of semiabelian varieties by Noguchi-Winkelmann-Yamanoi [21]. 2 In the qualitative aspect, the (strong) Green-Griffiths conjecture stipulates that if X is a complex projective space of general type, then there exists a proper subvariety Y ( X containing the image of every nonconstant entire holomorphic curve f : C → X. Following a beautiful strategy of Siu [25], Diverio, Merker and Rousseau [11] confirmed this conjecture . Berczi [2] improved the degree bound to d ≥ n9n. for generic hypersurface D ⊂ Pn+1 of degree d ≥ 2n5 Demailly [8] gave a new degree bound d ≥ n4 3 (cid:18)n log(n log(24n))(cid:19)n . In the logarithmic case, namely for the complement of a hypersurface D ⊂ Pn(C), there is another variant of this conjecture, so-called the logarithmic Green-Griffiths conjecture, which expects that for a generic hyper- surface D ⊂ Pn(C) having degree d ≥ n + 2, there should exist a proper subvariety Y ⊂ Pn(C) containing the image of every nonconstant entire holomorphic curve f : C → Pn(C) \ D. Darondeau [5] gave a positive answer for this case with effective degree bound d ≥ (5n)2nn. In this note, we show that the current method towards the Green-Griffiths conjecture can yield not only qualitative but also quantitative result, namely a Second Main Theorem type estimate in presence of only one generic hypersurface D of sufficiently high degree with the truncation level 1. Main Theorem. Let D ⊂ Pn(C) be a generic divisor having degree d ≥ 15(5n + 1)nn. Let f : C → Pn(C) be an entire holomorphic curve. If f is algebraically nondegenerate, then the following estimate holds Tf (r) ≤ N [1] f (r, D) + Sf (r). For background and standard techniques in Nevanlinna theory, we use the book of Noguchi-Winkelmann [19] as our main reference. The proof of the existence of logarithmic jet differentials in the last part of this note is based on the work of Darondeau [5]. Acknowledgments The authors would like to thank Joel Merker for his encouragements and his comments that greatly improved the manuscript. We would like to thank Nessim Sibony for very fruitful discussions on the paper [22]. We want to thank Junjiro Noguchi, Katsutoshi Yamanoi, Yusaku Tiba and Yuta Kusakabe for their interests in our work and for listening through many technical details. We would like to thank the referee for his/her careful reading of the manuscript and helpful suggestions. The first author is supported by the fellowship of the Japan Society for the Promotion of Science and the Grant-in-Aid for JSPS fellows Number 16F16317. 2 Logarithmic jet differentials 2.1 Logarithmic Green-Griffiths k-jet bundle The general strategy to prove the logarithmic Green-Griffiths conjecture consists of two steps. The first one is to produce many algebraically independent differential equations that all holomorphic curve f : C → Pn(C) \ D must satisfy. The second step consists in producing enough jet differentials from an initial one such that from the corresponding algebraic differential equations, one can eliminate all derivative in order to get purely algebraic equations. The central geometric object corresponding to the algebraic differential equations is the logarithmic Green- Griffiths k-jet bundle constructed as follows. Let X be a complex manifold of dimension n. For a point x ∈ X, consider the holomorphic germs (C, 0) → (X, x). Two such germs are said to be equivalent if they have the same Taylor expansion up to order k in some local coordinates around x. The equivalence class of 3 a holomorphic germ f : (C, 0) → (X, x) is called the k-jet of f , denote jk(f ). A k-jet jk(f ) is said to be regular if f ′(0) 6= 0. For a point x ∈ X, denote by jk(X)x the vector space of all k-jets of holomorphic germs (C, 0) → (X, x). Set Jk(X) := ∪ x∈X Jk(X)x and consider the natural projection πk : Jk(X) → X. Then Jk(X) is a complex manifold which carries the structure of a holomorphic fiber bundle over X, which is called the k-jet bundle over X. When k = 1, J1(X) is canonically isomorphic to the holomorphic tangent bundle TX of X. For an open subset U ⊂ X, for a section ω ∈ H 0(U, T ∗ X ), for a k-jet jk(f ) ∈ Jk(X)U , the pullback f ∗ω is of the form A(z)dz for some holomorphic function A. Since each derivative A(j) (0 ≤ j ≤ k − 1) is well defined, independent of the representation of f in the class jk(f ), the holomorphic 1-form ω induces the holomorphic map (2.1) ω : Jk(X)U → Ck; jk(f ) →(cid:0)A(z), A(z)(1), . . . , A(z)(k−1)(cid:1). Hence on an open subset U , a local holomorphic coframe ω1∧· · ·∧ωn 6= 0 yields a trivialization H 0(U, Jk(X)) → U × (Ck)n by giving new nk independent coordinates: σ → (πk ◦ σ; ω1 ◦ σ, . . . , ωn ◦ σ), where ωi are defined as in (2.1). The components x(j) (1 ≤ i ≤ n, 1 ≤ j ≤ k) of ωi ◦ σ are called jet- coordinates. In a more general setting where ω is a section over U of the sheaf of meromorphic 1-forms, the induced map ω is meromorphic. i Now, in the logarithmic setting, let D ⊂ X be a normal crossing divisor on X. This means that at each point x ∈ X, there exist some local coordinates z1, . . . , zℓ, zℓ+1, . . . , zn (ℓ = ℓ(x)) centered at x in which D is defined by D = {z1 . . . zℓ = 0}. Following Iitaka [14], the logarithmic cotangent bundle of X along D, denoted by T ∗ the locally-free sheaf generated by X (log D), corresponds to dz1 z1 , . . . , dzℓ zℓ , zℓ+1, . . . , zn in the above local coordinates around x. A holomorphic section s ∈ H 0(U, Jk(X)) over an open subset U ⊂ X is said to be a logarithmic k-jetfield if ω ◦ s are holomorphic for all sections ω ∈ H 0(U ′, T ∗ X (log D)), for all open subsets U ′ ⊂ U , where ω are induced maps defined as in (2.1). Such logarithmic k-jet fields define a subsheaf of Jk(X), and this subsheaf is itself a sheaf of sections of a holomorphic fiber bundle over X, called the logarithmic k-jet bundle over X along D, denoted by Jk(X, − log D) [18]. The group C∗ acts fiberwise on the jet bundle as follows. For local coordinates centered at x in which D = {z1 . . . zℓ = 0}, for any logarithmic k-jet field along D represented by some germ f = (f1, . . . , fn), if ϕλ(z) = λz is the homothety with ratio λ ∈ C∗, the action is given by z1, . . . , zℓ, zℓ+1, . . . , zn (ℓ=ℓ(x)) ((cid:0) log(fi ◦ ϕλ)(cid:1)(j) = λj(cid:0) log fi(cid:1)(j) ◦ ϕλ (cid:0)fi ◦ ϕλ(cid:1)(j) = λjf (j) ◦ ϕλ i (1 ≤ i ≤ ℓ), (ℓ+1 ≤ i ≤ n). Now we are ready to introduce the Green-Griffiths k-jet bundle [13] in the logarithmic setting. A logarith- mic jet differential of order k and degree m at a point x ∈ X is a polynomial Q(f (1), . . . , f (k)) on the fiber over x of Jk(X, − log D) enjoying weighted homogeneity: Q(jk(f ◦ ϕλ)) = λmQ(jk(f )). 4 Denote by EGG k,mT ∗ X (log D)x the vector space of such polynomials and set EGG k,mT ∗ X (log D) := ∪ x∈X EGG k,mT ∗ X (log D)x. By Fa`a di bruno's formula [4], [15], EGG logarithmic Green-Griffiths vector bundle. A global section P of EGG type in jet-coordinates x(j) X(log D) carries the structure of a vector bundle over X, called X (log D) locally is of the following k,mT ∗ k,mT ∗ : i Xα1 ,...,αk∈Nn Aα1,...,αk(cid:18) ℓ Yi=1(cid:0)(log xi)(1)(cid:1)α1,i α1+2α2+···+kαk=m n Yi=ℓ+1(cid:0)(xi)(1)(cid:1)α1,i(cid:19) . . .(cid:18) ℓ Yi=1(cid:0)(log xi)(k)(cid:1)αk,i n Yi=ℓ+1(cid:0)(xi)(k)(cid:1)αk,i(cid:19), where are multi-indices of length αλ = (αλ,1, . . . , αλ,n) ∈ Nn (1 ≤ λ ≤ k) αλ = X1≤i≤n αλ,i, and where Aα1,...,αk are locally defined holomorphic functions. By the following classical result, the first step to prove the Green-Griffiths conjecture reduces to finding logarithmic jet differentials valued in the dual of some ample line bundle. Fundamental vanishing theorem. ([9], [10]) Let X be a smooth complex projective variety and let D ⊂ X be a normal crossing divisor on X. If P is a nonzero global holomorphic logarithmic jet differential along D vanishing on some ample line bundle A on X, namely if then all nonconstant holomorphic curves f : C → X \ D must satisfy the associated differential equation 0 6= P ∈ H 0(cid:0)X, EGG k,mT ∗ X(log D) ⊗ A −1(cid:1), (2.2) P(cid:0)jk(f )(cid:1) ≡ 0. In the compact case, the existence of such global sections has been proved recently, first by Merker [15] for the case of smooth hypersurfaces of general type in Pn(C), and later for arbitrary general projective variety by Demailly [7]. Adapting this technique in the logarithmic setting, Darondeau [5] obtained a similar result for smooth hypersurface in projective space, provided that the degree is high enough compared with the dimension. Proposition 2.1. ([5, Th. 1.2]) Let c ∈ N be a positive integer and let D ⊂ Pn(C) be a smooth hypersurface having degree d ≥ 15(c + 2) nn. For jet order k = n, for weighted degrees m ≫ d big enough, the vector space of global logarithmic jet differentials along D of order k and weighted degree m vanishing on the m-th tensor power of the ample line bundle OPn(C)(c) has positive dimension: dim H 0(cid:0)Pn(C), EGG n,mT ∗ Pn(C)(log D) ⊗ OPn(C)(c)−m(cid:1) > 0. 3 Second Main Theorem for logarithmic jet differential Let D ⊂ Pn(C) be a smooth hypersurface in Pn(C). Let f : C → Pn(C) be an entire holomorphic curve, not necessary in the complement of D. If there exists a global logarithmic jet differential P which does not satisfy (2.2), then the fundamental vanishing theorem guarantees that the curve f must intersect the hypersurface D. Furthermore, in the quantitative aspect, based on the proof of the fundamental vanishing theorem, it is known that a Second Main Theorem type estimate Tf (r) ≤ C Nf (r, D) + Sf (r) 5 can be deduced from the existence of such global section P. There are several variants of the above estimate, see for instance in [22], [26]. Here we provide more information about the constant C and truncation of the counting function. Before going to introduce the main result of this section, we need to recall the following lemma on loga- rithmic derivative which is a crucial tool in Nevanlina theory. Logarithmic derivative Lemma. Let g 6≡ 0 be a nonzero meromorphic function on C. For any integer k ≥ 1, we have m g(k) (r) := m g(k) (r, ∞) = Sg(r). g g We refer to [19, Lem. 4.7.1] for a more general version of the above Lemma. Here is our main result in this section. Theorem 3.1. Let f : C → Pn(C) be an entire curve and let D ⊂ Pn(C) be a smooth hypersurface. Let m be a positive integer. If there exists a global logarithmic jet differential such that P ∈ H 0(cid:0)Pn(C), EGG k,mT ∗ Pn(C)(log D) ⊗ OPn(C)(1)− em(cid:1) then the following Second Main Theorem type estimate holds: P(cid:0)jk(f )(cid:1) 6≡ 0, m em Tf (r) ≤ N [1] f (r, D) + Sf (r). Proof. Our proof is partly based on [19, Lem. 4.7.1] and [9]. Let Q be the irreducible homogeneous polynomial defining D. By assumption, P(cid:0)jk(f )(cid:1) is a nonzero meromorphic section of f ∗OPn(C)(1)− em. Let DP,f be the pole divisor of P(cid:0)jk(f )(cid:1). Let(cid:0)V, φ) be a small enough local chart of Pn(C) such that φ : Pn(C) → Cn is a rational map and D is given by D = {z1 = 0}, where z = (z1, · · · , zn) are the natural coordinates on Cn. Put (3.1) fj := φ(f ), which is a meromorphic function on C for 1 ≤ j ≤ n. Then f is written in the local chart V as (f1, · · · , fn) on f −1(V ). Observe that f1/Q(f ) is a nowhere vanishing holomorphic function on f −1(V ). Recall that on V , the section P(cid:0)jk(f )(cid:1) can be written as Xα1,...,αk∈Nn P(cid:0)jk(f )(cid:1) = (3.2) α1+2α2+···+kαk=m Aα1,...,αk kYℓ=1(cid:18)(cid:0)(log f1)(ℓ)(cid:1)αℓ,1 nYj=2(cid:0)f (ℓ) j (cid:1)αℓ,j(cid:19), where Aα1,...,αk are holomorphic functions on f −1(V ) and αj = (αj,1, · · · , αj,n) for 1 ≤ j ≤ n. Hence, the support of DP,f is a subset of the zero set of Q ◦ f on C. Furthermore, since for each 1 ≤ ℓ ≤ k, the pole order of (log f1)(ℓ) at any point z ∈ C is at most ℓ min{ordz f1, 1} (hence at most ℓ min{ordz Q(f ), 1}) and since the degree of P is m, we get Let h be the pullback by f of the Fubini-Study form ωn on Pn(C). Using the Poincar´e-Lelong formula, we have where [DP,f ] is the integration current of DP,f . Combining this fact with the above inequality, we obtain min{ordz(Q ◦ f ), 1}z. DP,f ≤ mXz∈C ddc log kP(cid:0)jk(f )(cid:1)kh ≥ emf ∗ω − [DP,f ], em Tf (r) + O(1) ≤Z∂∆r log kP(cid:0)jk(f )(cid:1)k2 h dθ 2π 6 + m N [1] f (r, D). Thus, it remains to verify (3.3) (3.4) Z∂∆r Z∂∆r h dθ 2π log kP(cid:0)jk(f )(cid:1)k2 log χ(f )P(cid:0)jk(f )(cid:1)2 dθ 2π = Sf (r). = Sf (r), Using a partition of unity on Pn(C), the problem reduces to proving that where χ is a smooth positive function compactly supported on a local chart V as above. Using the following elementary observations with s, s1, · · · , sN ∈ R∗ +: log s = log+ s − log+ 1 s ≤ log+ s log+ log+ NXi=1 NYi=1 si ≤ si ≤ NXi=1 NXi=1 log+ si + log N log+ si, where log+ denotes max{log, 0}, we get Z∂∆r log χ(f )P(cid:0)jk(f )(cid:1)2 dθ 2π ≤ + (3.5) α1+2α2+···+kαk=m Xα1,...,αk∈Nn log+(cid:0)χ(f )f (ℓ) kXℓ=1(cid:18)Z∂∆r αℓ,j(cid:1) dθ j nXj=2Z∂∆r log+(cid:0)χ(f )(log f1)(ℓ)αℓ,1(cid:1) dθ 2π(cid:19) + O(1), 2π Recall from (3.1) that fj are meromorphic functions on C for 1 ≤ j ≤ n. Hence applying the logarithmic derivative Lemma to f1, we infer that (3.6) Z∂∆r log+(cid:0)χ(f )(log f1)(ℓ)αℓ,1(cid:1) dθ 2π = Sf (r). Therefore, it suffices to show that this property still holds for the remaining terms in the right-hand side of (3.5). Continuing to apply the logarithmic derivative Lemma, we obtain for some constant c which is independent of r, f . Hence it remains to check Z∂∆r j 2π ≤ cZ∂∆r αℓ,j(cid:1) dθ log+(cid:0)χ(f )f (ℓ) Z∂∆r log+(cid:0)χ(f )f (1) j j log+(cid:0)χ(f )f (1) 2(cid:1) dθ = Sf (r). 2π + Sf (r), 2(cid:1) dθ 2π This can be done by using the similar arguments as in [19, p. 149]. For the reader's convenience, we present the idea here. Since χ is compactly supported on V , there exists a bounded positive function B for which χ dzj ∧ d¯zj ≤ B(z) ωn on V for 2 ≤ j ≤ n. This yields χ(f )f (1) j 2 dz ∧ d¯z = f ∗(χ dzj ∧ d¯zj) ≤ B(f )f ∗ωn. The pullback f ∗ωn is of the form B1 dz ∧ d¯z. Hence we deduce from the above inequality that Z∂∆r log+(cid:0)χ(f )f (1) j 2(cid:1) dθ 2π ≤ ≤ log+ B(f ) dθ + log+ B1 dθ 1 2πZ∂∆r log+ B1 dθ + suppz∈V B(z). EstimatingR∂∆r Lemma, see [19, (3.2.8)]. The proof is finished. log+ B1 is done by following the same arguments as in the proof of the logarithmic derivative 1 2πZ∂∆r 2πZ∂∆r 1 7 4 Existence of logarithmic jet differentials Let f : C → Pn(C) be an algebraically nondegenerate holomorphic curve. Following the second step in Siu's strategy to prove the Green-Griffith conjecture, morally, if we can produce enough logarithmic jet differentials valued in the dual of some ample line bundle on Pn(C), then among them, we can choose at least one such that f does not satisfy the algebraic differential equation (2.2). Theorem 4.1. Let c be a positive integer with c ≥ 5n − 1. Let D ⊂ Pn(C) be a generic smooth hypersurface in Pn(C) having degree d ≥ 15(c + 2)nn. Let f : C → Pn(C) be an entire holomorphic curve. If f is algebraically nondegenerate, then for jet order k = n and for weighted degrees m > d big enough, there exists an integer 0 ≤ ℓ ≤ m and a global logarithmic jet differential such that (4.1) P ∈ H 0(cid:0)Pn(C), EGG n,mT ∗ Pn(C)(log D) ⊗ OPn(C)(1)−cm+ℓ(5n−2)(cid:1) P(cid:0)jn(f )(cid:1) 6≡ 0. The rest of this section is devoted to proving Theorem 4.1 whose proof is based on [11, 5]. Let S := PH 0(cid:0)Pn(C), O(d)(cid:1) be the projective parameter space of homogeneous polynomials of degree d in Pn(C) which identifies with the projective space PNd(C) of dimension Nd = dim PH 0(cid:0)Pn(C), OPn(C)(d)(cid:1) =(cid:18)n + d d (cid:19) − 1. We then introduce the universal hypersurface parametrizing all hypersurfaces of fixed degree d in Pn(C), defined by the equation H ⊂ Pn(C) × S 0 = Xα∈Nn+1 Aα Z α in the following two collections of homogeneous coordinates Z = [Z0 : · · · : Zn] ∈ Pn(C), A = [(Aα)α∈Nn+1,α=d] ∈ PNd(C), where α = (α0, . . . , αn) ∈ Nn+1 are multiindices. Since H is a smooth hypersurface on Pn(C) × S, we can construct the logarithmic k-jet bundle Jk(Pn(C) × S, − log H) over Pn(C) × S along H. Now, let η be be the natural projection from Jk(Pn(C) × S, − log H) to Pn(C) × S. Let pr1 and pr2 be the natural projections from Pn(C) × S to the first and second part, respectively. Let VH,k be the analytic subset of Jk(Pn(C) × S, − log H) consisting of all vertical logarithmic jet fields of order k which, by definition, are jets jk(f ) such that f lies entirely in some fiber of the second projection pr2. Denote by V reg H,k the open subset consisting of all regular jets. By [6, p. 571-572] (see also [16, p. 1088]), V reg H,k is smooth manifold. Following the method of producing new jet differentials developed by Siu [25], in the logarithmic setting, one needs to construct low pole order frames on V reg H,k. Proposition 4.1. ([6, Main Theorem]) For jet order k ≥ 1, for degree d ≥ k, the twisted tangent bundle is generated over V reg H,k \ η−1H by its global holomorphic sections. TVH,k ⊗ η∗(cid:0)OPn(C)(5 k − 2) ⊗ OS(1)(cid:1) 8 In fact, those global sections mentioned in the above Proposition are global vector fields on the whole logarithmic k-jet bundle and satisfy the canonical tangential conditions described as in [6, 16]. Hence they are true vector fields on the smooth part of VH,k. Moreover by the constructions in [6, 16, 26], the coefficients of those vector fields are polynomials in local logarithmic jet coordinates. Let Z0 be the subset of S consisting of all s whose corresponding hypersurface Ds is not smooth. Observe that Z0 is a proper analytic subset of S. From now on we work with the fixed jet order k = n. Since pr−1 2 s = Pn(C) × {s} and since Ds is smooth for every s outside Z0, one can define Jn(pr−1 2 s, − log Ds) for any s ∈ S\Z0. Let us set L := [s∈S\Z0 Lreg := [s∈S\Z0 Jn(pr−1 2 s, − log Ds), n (pr−1 J reg 2 s, − log Ds). Observe that L has a natural structure of holomorphic fiber bundle over Pn(C)×(S\Z0). Note also that L, Lreg are open subsets of VH,n and V reg H,n, respectively. Set E := [s∈S\Z0 EGG n,mT ∗ pr−1 2 s(log Ds), then E carries the structure of holomorphic vector bundle over Pn(C) × (S\Z0). This fact allows us to extend holomorphically a nonzero jet differential provided by Proposition 2.1. Let us enter the details. Lemma 4.1. Let c ≥ 5n − 1 be a positive integer. For degree d ≥ 15(c + 2)nn, for weighted degree m ≫ d, there exists a proper analytic subset Z of S containing Z0 such that for every s ∈ S \ Z, we can find a Zariski open neighborhood Us of s in S\Z0 and a nonzero holomorphic section P 6≡ 0 of the twisted vector bundle E ⊗ pr∗ 1 OPn(C)(1)−cm over pr−1 2 Us. Proof. By construction, for any s ∈ S\Z0, the restriction of E to pr−1 2 s = Pn(C) × {s} coincides with Hence Proposition 2.1 guarantees the existence of a nonzero global section EGG n,mT ∗ pr−1 2 s(log Ds). 0 6≡ Ps ∈ H 0(cid:0)Pn(C) × {s}, E ⊗ pr∗ 1 OPn(C)(1)−cmpr−1 2 s(cid:1) of the restriction of the twisted vector bundle E ⊗ pr∗ 2 s. By the semi-continuity theorem (c.f. [17, p. 50]), there exists a proper Zariski closed subset Z of S containing Z0 such that for any s ∈ S \ Z, the natural restriction map 1 OPn(C)(1)−cm to pr−1 2 Us, E ⊗ pr∗ H 0(cid:0) pr−1 1 OPn(C)(1)−cm(cid:1) −→ H 0(cid:0) pr−1 2 s, E ⊗ pr∗ 1 OPn(C)(1)−cmpr−1 2 s(cid:1) is onto for some Zariski open subset Us ⊂ S\Z0 containing s. As a consequence, the above section Ps can be extended holomorphically to a section P of E ⊗ pr∗ 1 OPn(C)(1)−cm over pr−1 2 Us. Proof of Theorem 4.1. Let Z, P be as in Lemma 4.1. Let us first describe precisely the generic assumption of D in the statement. By this, we mean that if D corresponds to the element s ∈ S (i.e. D = Ds), then s lies outside Z ∪ HS, where HS is a fixed arbitrary hyperplane of S. Here the condition s 6∈ HS is given in order to get rid of η∗OS(1) in Proposition 4.1 because the line bundle OS(1) is trivial on S\HS. From now on, we fix s ∈ S \ Z and D = Ds. Applying Proposition 4.1 for jet order k = n, the twisted tangent bundle TVH,n ⊗ η∗(cid:0)OPn(C)(5n − 2) ⊗ OS(1)(cid:1) 9 is generated by its global holomorphic sections over V reg H,n \ η−1H. Moreover, the coefficients of those sections are polynomials in the logarithmic n-jet coordinates associated with the canonical coordinates of Pn(C) × S. In particular, the restriction of the bundle to η−1Y , where TVH,n ⊗ η∗OPn(C)(5n − 2) Y := pr−1 2 (Us\HS) \ H is generated on (V reg coordinates as above. H,n ∩ η−1Y ) by its global sections whose coefficients are polynomials in the logarithmic jet For 0 ≤ ℓ ≤ m, let v1, . . . , vℓ be sections of TVH,n ⊗ η∗OPn(C)(5n − 2) over the open subset L ⊂ VH,n. As explained below, the significance of those sections is that they allow to construct new global logarithmic jet differentials. Indeed, we can view P as a holomorphic mapping P : Lpr−1 2 Us → pr∗ 1 OPn(C)(1)−cmpr−1 2 Us , which is locally a homogeneous polynomial of degree m. It follows that the Lie derivative (v1 · · · vℓ) · P is also a holomorphic map from Lpr−1 to 2 Us pr∗ 1 OPn(C)(1)−cm+ℓ(5n−2) and is locally a homogeneous polynomial of the same degree m. The fact that the derivative of P along vj preserves the degree m can be deduced from the fact that the coefficients of vj are polynomials in the logarith- mic jet coordinates and by the chain rule of derivatives, the degree of those polynomials should compensate the losses of degree due to the differentiation with respect to vj, see [26, Sec. 3.7]. In summary, we obtain a holomorphic map (4.2) (v1 · · · vℓ) · P : Lpr−1 2 Us By composing f with the inclusion → pr∗ 1 OPn(C)(1)−cm+ℓ(5n−2) pr−1 2 Us . Pn(C) ֒→ Pn(C) × {s} ⊂ Pn(C) × S, we can consider f as a holomorphic curve into Y ⊂ Pn(C) × S because s ∈ Us \ HS and f is not included in D. Let {P = 0} ⊂ pr−1 2 Us be the zero divisor of P, where we view P as a holomorphic section of E ⊗ pr∗ 2 Us. Since f is algebraically nondegenerate, there exists z0 ∈ C such that f ′(z0) 6= 0 and f (z0) 6∈ D ∩ {P = 0}. Consequently, we get 1 OPn(C)(1)−cm over pr−1 (4.3) jn(f )(z0) ∈ (Lreg ∩ η−1Y ). Now proceeding as in [11], we can show that there exist global slanted vector fields v1, . . . , vℓ for some 0 ≤ ℓ ≤ m such that (v1 · · · vℓ) · P(cid:0)jn(f )(cid:1) 6= 0. For reader's convenience, we briefly recall the idea. Denoted by Ps the restriction of P to Pn(C) × {s}. Consider the logarithmic jet coordinates (z, z(1), . . . , z(n)) ∈ Cn(n+1) around jn(f )(z0) of LPn(C)×{s}. Using a linear change of coordinates, we obtain modified logarithmic jet coordinates (z′, z′(1), . . . , z′(n)) in which jn(f )(z0) is the origin. Since Ps is locally a homogeneous polynomial in logarithmic jet coordinates whose coefficients are holomorphic functions on local charts of Pn(C), so it is in the new logarithmic jet coordinates. By the choice of z0, there exists a coefficient Aα1,...,αn(z) of Ps (see (3.2)) for which Aα1,...,αn(f (z0)) 6= 0. Let be the monomial of Ps associated with Aα1,...,αn , where αj ∈ Nn for 1 ≤ j ≤ n and α1 + 2α2 + · · · + nαn ≤ m. We then choose local vector fields v′ ℓ around the origin jn(f )(z0) for which Aα1,...,αn(z)(cid:0)z′(1)(cid:1)α1 · · ·(cid:0)z′(n)(cid:1)αn ℓ) · P(cid:0)jn(f )(z0)(cid:1) = Aα1,...,αn(cid:0)f (z0)(cid:1) 6= 0. 1, . . . v′ 1, . . . , v′ (v′ 1 · · · v′ As we mentioned above, these vector fields v′ (Lreg ∩ η−1Y ). Combining this with (4.3), we get (v1 · · · vℓ) · P(cid:0)jn(f )(z0)(cid:1) 6= 0. This together with (4.2) implies (4.1). The proof of Theorem 4.1 is completed. ℓ can be generated by global vector fields v1, . . . , vℓ on 10 Corollary 4.1. Let c be a positive integer with c ≥ 5n − 1. Let D ⊂ Pn(C) be a generic hypersurface in Pn(C) having degree d ≥ 15(c + 2)nn. Let f : C → Pn(C) be an entire holomorphic curve. If f is algebraically nondegenerate, then the following Second Main Theorem type estimate holds: Tf (r) ≤ 1 c − 5n + 2 N [1] f (r, D) + Sf (r). In particular, choosing c = 5n − 1, one obtains the Main Theorem. Proof. This is a direct application of Theorem 3.1 to global logarithmic jet differential P supplied by Theorem 4.1, where m = mc − ℓ(5n − 2) ≥ m(c − 5n + 2) ≥ m. 11 References [1] Do Phuong An, Si Duc Quang, and Do Duc Thai. "The second main theorem for meromorphic mappings into a complex projective space". In: Acta Math. Vietnam. 38.1 (2013), pp. 187–205. ISSN: 0251-4184. [2] Gergely Berczi. "Towards the Green-Griffiths-Lang conjecture via equivariant localisation". In: Preprint arxiv:1509.03406 (2015). [3] Henri Cartan. "Sur les z´eros des combinaisons lin´eaires de p fonctions holomorphesdonn´ees". In: Math- ematica 7 (1933), pp. 80–103. [4] G. M. Constantine and T. H. Savits. "A multivariate Fa`a di Bruno formula with applications". In: Trans. Amer. Math. Soc. 348.2 (1996), pp. 503–520. ISSN: 0002-9947. [5] Lionel Darondeau. "On the logarithmic Green-Griffiths conjecture". In: Int. Math. Res. Not. IMRN 6 (2016), pp. 1871–1923. ISSN: 1073-7928. [6] Lionel Darondeau. "Slanted vector fields for jet spaces". In: Math. Z. 282.1-2 (2016), pp. 547–575. [7] [8] [9] Jean-Pierre Demailly. "Holomorphic Morse inequalities and the Green-Griffiths-Lang conjecture". In: Pure Appl. Math. Q. 7.4, Special Issue: In memory of Eckart Viehweg (2011), pp. 1165–1207. ISSN: 1558-8599. Jean-Pierre Demailly. "Hyperbolic algebraic varieties and holomorphic differential equations". In: Acta Math. Vietnam. 37.4 (2012), pp. 441–512. ISSN: 0251-4184. Jean-Pierre Demailly. "Vari´et´es projectives hyperboliques et ´equations diff´erentielles alg´ebriques". In: Journ´ee en l'Honneur de Henri Cartan. Vol. 1997. SMF Journ. Annu. Soc. Math. France, Paris, 1997, pp. 3–17. [10] Gerd-Eberhard Dethloff and Steven Shin-Yi Lu. "Logarithmic jet bundles and applications". In: Osaka J. Math. 38.1 (2001), pp. 185–237. ISSN: 0030-6126. [11] Simone Diverio, Joel Merker, and Erwan Rousseau. "Effective algebraic degeneracy". English. In: In- ventiones mathematicae 180 (2010), pp. 161–223. [12] A. `E. Eremenko and M. L. Sodin. "Distribution of values of meromorphic functions and meromorphic curves from the standpoint of potential theory". In: Algebra i Analiz 3.1 (1991), pp. 131–164. ISSN: 0234-0852. [13] Mark Green and Phillip Griffiths. "Two applications of algebraic geometry to entire holomorphic map- pings". In: The Chern Symposium 1979 (Proc. Internat. Sympos., Berkeley, Calif., 1979). Springer, New York-Berlin, 1980, pp. 41–74. [14] Shigeru Iitaka. Algebraic geometry. Vol. 76. Graduate Texts in Mathematics. An introduction to bira- tional geometry of algebraic varieties, North-Holland Mathematical Library, 24. Springer-Verlag, New York-Berlin, 1982, pp. x+357. ISBN: 0-387-90546-4. [15] Joel Merker. "Algebraic differential equations for entire holomorphic curves in projective hypersurfaces of general type: optimal lower degree bound". In: Geometry and analysis on manifolds. Vol. 308. Progr. Math. Birkhauser/Springer, Cham, 2015, pp. 41–142. [16] Joel Merker. "Low pole order frames on vertical jets of the universal hypersurface". In: Ann. Inst. Fourier (Grenoble) 59.3 (2009), pp. 1077–1104. ISSN: 0373-0956. [17] David Mumford. Abelian varieties. Tata Institute of Fundamental Research Studies in Mathematics, No. 5. Published for the Tata Institute of Fundamental Research, Bombay; Oxford University Press, London, 1970, pp. viii+242. [18] [19] Junjiro Noguchi. "Logarithmic jet spaces and extensions of de Franchis' theorem". In: Contributions to several complex variables. Aspects Math., E9. Friedr. Vieweg, Braunschweig, 1986, pp. 227–249. Junjiro Noguchi and Jorg Winkelmann. Nevanlinna theory in several complex variables and Diophantine approximation. Vol. 350. Grundlehren der Mathematischen Wissenschaften [Fundamental Principles of Mathematical Sciences]. Springer, Tokyo, 2014, pp. xiv+416. ISBN: 978-4-431-54570-5; 978-4-431- 54571-2. 12 [20] [21] Junjiro Noguchi, Jorg Winkelmann, and Katsutoshi Yamanoi. "The second main theorem for holomor- phic curves into semi-abelian varieties". In: Acta Math. 188.1 (2002), pp. 129–161. ISSN: 0001-5962. Junjiro Noguchi, Jorg Winkelmann, and Katsutoshi Yamanoi. "The second main theorem for holomor- phic curves into semi-abelian varieties. II". In: Forum Math. 20.3 (2008), pp. 469–503. ISSN: 0933-7741. [22] Mihai Paun and Nessim Sibony. "Value Distribution Theory for Parabolic Riemann Surfaces". In: Preprint arXiv:1403.6596 (2014). [23] Min Ru. "A defect relation for holomorphic curves intersecting hypersurfaces". In: Amer. J. Math. 126.1 (2004), pp. 215–226. ISSN: 0002-9327. [24] Min Ru. "Holomorphic curves into algebraic varieties". In: Ann. of Math 169 (2009), pp. 255–267. [25] Yum-Tong Siu. "Hyperbolicity in complex geometry". In: The legacy of Niels Henrik Abel. Springer, Berlin, 2004, pp. 543–566. [26] Yum-Tong Siu. "Hyperbolicity of generic high-degree hypersurfaces in complex projective space". In: Invent. Math. 202.3 (2015), pp. 1069–1166. ISSN: 0020-9910. [27] Do Duc Thai and Duc-Viet Vu. "Holomorphic mappings into compact complex manifolds". In: Houston J. Math. 43.3 (2017), pp. 725–762. [28] Katsutoshi Yamanoi. "Holomorphic curves in abelian varieties and intersections with higher codimen- sional subvarieties". In: Forum Math. 16.5 (2004), pp. 749–788. ISSN: 0933-7741. DINH TUAN HUYNH, DEPARTMENT OF MATHEMATICS, GRADUATE SCHOOL OF SCIENCE, OSAKA UNIVERSITY, TOYONAKA, OSAKA 560-0043, JAPAN & DEPARTMENT OF MATHEMATICS, COLLEGE OF EDUCATION, HUE UNIVERSITY, 34 LE LOI ST., HUE CITY, VIETNAM E-mail address: [email protected] DUC-VIET VU, KOREA INSTITUTE FOR ADVANCED STUDY, 85 HOEGIRO, DONGDAEMUN-GU, SEOUL 02455, REPUBLIC OF KOREA E-mail address: [email protected] SONG-YAN XIE, MAX-PLANCK-INSTITUT F UR MATHEMATIK, VIVATSGASSE 7, 53111 BONN, GERMANY E-mail address: [email protected] 13
1903.05949	1	1903	2019-03-14T12:49:08	Polynomial spline spaces of non-uniform bi-degree on T-meshes: Combinatorial bounds on the dimension	[ "math.AG", "math.NA" ]	Polynomial splines are ubiquitous in the fields of computer aided geometric design and computational analysis. Splines on T-meshes, especially, have the potential to be incredibly versatile since local mesh adaptivity enables efficient modeling and approximation of local features. Meaningful use of such splines for modeling and approximation requires the construction of a suitable spanning set of linearly independent splines, and a theoretical understanding of the spline space dimension can be a useful tool when assessing possible approaches for building such splines. Here, we provide such a tool. Focusing on T-meshes, we study the dimension of the space of bivariate polynomial splines, and we discuss the general setting where local mesh adaptivity is combined with local polynomial degree adaptivity. The latter allows for the flexibility of choosing non-uniform bi-degrees for the splines, i.e., different bi-degrees on different faces of the T-mesh. In particular, approaching the problem using tools from homo-logical algebra, we generalize the framework and the discourse presented by Mourrain (2014) for uniform bi-degree splines. We derive combinatorial lower and upper bounds on the spline space dimension and subsequently outline sufficient conditions for the bounds to coincide.	math.AG	math	Polynomial spline spaces of non-uniform bi-degree on T-meshes: Combinatorial bounds on the dimension Deepesh Toshniwala,∗, Bernard Mourrainb, Thomas J. R. Hughesa aInstitute for Computational Engineering and Sciences, University of Texas at Austin, USA bUniversit´e Cote d'Azur, Inria Sophia Antipolis M´editerran´ee, Sophia Antipolis, France Abstract Polynomial splines are ubiquitous in the fields of computer aided geometric design and compu- tational analysis. Splines on T-meshes, especially, have the potential to be incredibly versatile since local mesh adaptivity enables efficient modeling and approximation of local features. Meaningful use of such splines for modeling and approximation requires the construction of a suitable spanning set of linearly independent splines, and a theoretical understanding of the spline space dimension can be a useful tool when assessing possible approaches for building such splines. Here, we provide such a tool. Focusing on T-meshes, we study the dimension of the space of bivariate polynomial splines, and we discuss the general setting where local mesh adaptivity is combined with local polynomial degree adaptivity. The latter allows for the flexibility of choosing non-uniform bi-degrees for the splines, i.e., different bi-degrees on different faces of the T-mesh. In particular, approaching the problem using tools from homo- logical algebra, we generalize the framework and the discourse presented by Mourrain (2014) for uniform bi-degree splines. We derive combinatorial lower and upper bounds on the spline space dimension and subsequently outline sufficient conditions for the bounds to coincide. Keywords: Smooth splines, T-meshes, Non-uniform degrees, Dimension formula, Homological algebra 1. Introduction Standard B-spline parameterizations of surfaces in Computer Aided Geometric Design are defined on a grid of nodes over a rectangular domain. These representations are also the basis of Isogeometric Analysis which provides high order finite element methods in numerical simulations [9]. However, grid structures do not allow complex shapes to be easily resolved. They also preclude the flexibility of performing local refinements for improving the error in numerical simulations. To address these issues, meshes with T-junctions -- also called T- meshes -- and polynomial and rational splines on such meshes have been investigated for performing both geometric modeling and isogeometric analysis. Classically, uniform degree splines, i.e., piecewise polynomial functions of uniform degree on the faces, have been studied and developed on T-meshes with the intent of using T-junctions for locally increasing the resolution offered by the spline space. An alternate strategy to improve the approximation power of splines is to increase the degree in a localized manner. In this paper, motivated by applications for isogeometric finite element methods, we study the space of piecewise poly- nomials functions on a T-mesh with different bi-degrees on its faces and different regularities across its edges. In particular, we analyze the dimension of these functional spaces, thus ∗Corresponding author Email address: [email protected] (Deepesh Toshniwal) providing a tool that can help identify when a given set of linearly independent splines spans the full space. Over the last decades, several works focused on the construction of spline functions and the analysis of spline function spaces on T-meshes have appeared, mainly motivated by applications in isogeometric analysis. So called T-splines, which are B-spline functions defined on domains with a T-mesh structure, have been investigated for their flexibility of representing shapes [24], for isogeometric analysis [1] and for functional approximation [22]. However, linear dependencies of the blending functions involved in T-spline constructions have been observed [8]. To remedy this problem, a special sub-family of T-splines, called Analysis Suitable T-splines has been developed, by imposing sufficient constraints on the T-mesh [19, 23, 7]. The construction of so-called LR-splines defined on T-meshes and based on knot sub-grids has been proposed in [12]. Their use in isogeometric analysis has been further investigated in [16], including an analysis of the linear independency of the blending functions [6]. Another type of splines, so-called hierarchical B-splines, have been investigated in [13, 17, 10, 14]. They are defined by recursive subdivisions of quadrangular faces, producing nested spaces of splines functions and providing simple schemes for performing local refinements. In general, the dimension of the spaces of splines on T-meshes can be unstable, i.e., it can depend on the global geometry of the T-mesh [18, 20]. Since any efficient constructive approach must rely only on local data for building spline functions, this instability in the dimension necessitates identification of configurations where the spline space dimension is a priori guaranteed to be stable. In this direction, a detailed study of spline spaces on general T- meshes has been presented in [20] using homological techniques, which go back to [2]. Results from [20] were used in [11] to devise a refinement strategy for LR-splines that ensures that the entire spline space is spanned by LR B-splines at each stage of refinement. The dimension of Tchebycheffian spline spaces over planar T-meshes, which involve non-polynomial functions, have been investigated in [4], [5], exploiting the same homological techniques as in [20]. In all the works referenced above, only uniform degree splines are considered on the T- meshes. Here, we analyze in detail splines spaces over general T-meshes when non-uniform polynomial bi-degrees are chosen on the faces, thus accounting for local degree adaptivity in conjunction with local mesh adaptivity. We provide combinatorial lower and upper bounds on the dimensions of such spline spaces and outline sufficient conditions for the bounds to coincide. These sufficient conditions are equivalent to geometric conditions that need to be satisfied by the T-meshes. The approach is based on homological techniques and generalizes the framework presented in [20] to the case of non-uniform polynomial bi-degree distributions. As part of the approach, we perform a degree-based decomposition of the mesh into nested regions and this allows us to untangle the contributions of different bi-degrees to the spline space dimension. The main results on the lower and upper bounds of the dimension of these spline spaces (Theorems 6.4, 6.5 and 7.2) involve homological invariants of the nested regions associated to the different bi-degrees. As mentioned previously, the theoretical results presented here can be used to identify when a given set of linearly independent splines spans the full spline space. Conversely, given a constructive approach that aims to produce linearly independent splines over T-meshes using only local data, computation of the associated spline space dimension can help identify cases where the splines produced by the approach cannot be linearly independent. This is crucial for devising constructive approaches that can be robustly employed for performing isogeometric analysis. The layout of the paper is as follows. We start by introducing preliminary concepts and notation about T-meshes and non-uniform bi-degree spline spaces on such meshes in Section 2. Thereafter, we introduce the topological complexes that form the main object of our analysis in Section 3; in particular, Section 3.3 provides an overview of our approach 2 to the problem at hand. Sections 4 and 5 take a closer look at the topological complexes introduced in Section 3, and the results presented therein are used in Section 6 to provide bounds on the spline space dimension (Theorems 6.4 and 6.5). Section 7 contains Theorem 7.2, which outlines sufficient conditions for the bounds derived in Section 6 to coincide. We also discuss the notion of maximal segment weights, generalized from [20]. This notion helps provide a geometric criterion that is useful when computing the spline space dimension. Finally, Section 8 provides examples of the theory developed here. We would like to mention here that computations using Macaulay2 [15] went hand-in-hand with the research presented here. 2. Planar T-meshes and polynomials In the following we define the basic concepts associated with planar T-meshes, and there- after present some preliminary results on polynomials. We will proceed as in [20], albeit in the setting of non-uniform degree spline spaces. 2.1. T-meshes Definition 2.1 (T-mesh). A T-mesh T of R2 is defined as: • a finite set T2 of closed axis-aligned rectangles σ of R2, called 2-cells or faces, • a finite set T1 of closed axis-aligned segments τ , called 1-cells or edges, included in ∪σ∈T2∂σ, and, • a finite set, T0, of points γ, called 0-cells or vertices, included in ∪τ ∈T1∂τ , such that • σ ∈ T2 ⇒ ∂σ is a finite union of elements of T1, • σ, σ′ ∈ T2 ⇒ σ ∩ σ′ = ∂σ ∩ ∂σ′ is a finite union of elements of T1 ∪ T0, and, • τ, τ ′ ∈ T1 with τ 6= τ ′ ⇒ τ ∩ τ ′ = ∂τ ∩ ∂τ ′ ⊂ T0. The domain of the T-mesh is assumed to be connected and is defined as Ω := ∪σ∈T2σ ⊂ R2. Sets of horizontal and vertical edges will be denoted by Th 1, respectively. Edges Ω, and boundary edges otherwise. of the T-mesh are called interior edges if they intersect T1; and the sets of interior horizontal and vertical The set of interior edges will be denoted by edges will be denote by Ω it will be called an interior vertex, and a boundary vertex otherwise. The set of interior vertices will T0. We will denote the number of i-cells with ti := #Ti; the number of interior be denoted by i-cells with 1, respectively. Similarly, if a vertex is in ti := # Ti; and so on. 1 and Tv ◦ Tv ◦ Th 1 and ◦ ◦ ◦ ◦ ◦ ◦ Example 2.2. An example T-mesh is shown in Figure 1. Assumption 2.3. The domain Ω is simply connected, and Ω is connected. ◦ 3 Figure 1: An example of the kind of T-meshes we will consider in this document. 2.2. Splines on T-meshes We will now define spaces of piecewise-polynomial splines on the planar T-meshes intro- duced above. To do so, we will first define a map that specifies relative polynomial degrees on the faces of T , and a second map that specifies the smoothness across its edges. Note that these maps are assumed to be known/fixed throughout this document and, when needed, we will omit mentioning them explicitly in order to simplify notation. Definition 2.4 (Degree deficit distribution). A degree deficit distribution on T is a map ∆m : T2 → Z2 ≥0 , σ 7→ ∆m(σ) . It is assumed that D := {∆m(σ) : σ ∈ T2} can be totally ordered using the relation ≤D defined as (a1, a2) ≤D (b1, b2) ⇔ a1 ≤ b1 ∧ a2 ≤ b2 , and that (0, 0) ∈ D. Given a degree-deficit distribution as defined above, we build the following sequence ∆mi, i = 0, . . . , l, min D = (0, 0) =: ∆m0 < ∆m1 < · · · < ∆ml := max D , such that ∆mi − ∆mi−1 =: ∆ni ∈ {(1, 0), (0, 1), (1, 1)} . (2.1) (2.2) Note that all comparisons carried out above are with respect to the ordering in Definition 2.4. We will denote the components of ∆mi and ∆ni with (∆mi1, ∆mi2) and (∆ni1, ∆ni2), respectively. The map ∆m will help specify the bi-degree of polynomials on a face σ ∈ T2. Given m ∈ Z2 ≥0, we define the bi-degrees m(cid:3), (cid:3) ∈ T2 ∪ T1 ∪ T0, as m(cid:3) := m − ∆m((cid:3)) , where the induced degree deficits on τ ∈ T1 and γ ∈ T0 are defined as ∆m(τ ) := min τ ⊂σ ∆m(σ) , ∆m(γ) := min γ∈σ ∆m(σ) . (2.3) (2.4) Let R := R[s, t] be the polynomial ring with coefficients in R. We define Pm1m2 ≡ P(m1,m2) ⊂ R as the R-linear vector space of polynomials of bi-degree ≤ (m1, m2) spanned by the monomials sitj, 0 ≤ i ≤ m1, 0 ≤ j ≤ m2. If any of m1, m2 are negative, then Pm1m2 := 0. 4 Figure 2: A T-mesh T where the degree deficit on white faces is (2, 2), on blue elements is (1, 1), and on red elements is (0, 0). Definition 2.5 (Smoothness distribution). A smoothness distribution on T is a map r : ◦ T1 → Z≥0 , τ 7→ r(τ ) , such that τ, τ ′ ∈ ◦ Th 1 (or both in ◦ Tv 1) and τ ∩ τ ′ 6= ∅ ⇒ r(τ ) = r(τ ′). ◦ The map r will help us define the smoothness across all interior edges, i.e., for τ ∈ splines will be required to be at least C r(τ ) smooth across τ . For γ ∈ τh ∩ τv, (τh, τv) ∈ respectively, as T1, T0 such that {γ} = 1, the smoothness in horizontal and vertical directions is defined, 1 × Th ◦ Tv ◦ ◦ rγ,h = r(τv) , rγ,v = r(τh) . (2.5) Definition 2.6 (Spline space). Given T-mesh T , degree deficit and smoothness distributions ∆m and r, respectively, and m ∈ Z2 ≥0, we define the spline space Rr ∆m,m(T ) as Rr ∆m,m ≡ Rr ∆m,m(T ) :=(cid:26)f : ∀σ ∈ T2, f σ ∈ Pmσ = Pm−∆m(σ) , ◦ ∀τ ∈ T1, f is C r(τ ) across τ } . (2.6) Example 2.7. Corresponding to the T-mesh shown in Example 2.2, an example degree deficit distribution has been shown in Figure 2. The set D is given by D = {(0, 0), (1, 1), (2, 2)} , and the sequence ∆mi can be chosen to be (0, 0) = ∆m0 < (1, 1) = ∆m1 < (2, 1) = ∆m2 < (2, 2) = ∆m3 . Clearly, the above choice of the sequence is not unique. We could have alternatively chosen the shorter sequence (0, 0) = ∆m0 < (1, 1) = ∆m1 < (2, 2) = ∆m2 . We will employ the following algebraic characterization of smoothness of piecewise-polynomial splines [3]. Lemma 2.8 ([2, 20]). For σ, σ′ ∈ T2, let σ ∩ σ′ = τ ∈ T1. A piecewise polynomial func- tion equaling p ∈ R and p′ ∈ R on σ and σ′, respectively, is at least r times continuously differentiable across τ if and only if ◦ p − p′ ∈(cid:0)lr+1(cid:1) , where l ∈ R is a non-zero linear polynomial vanishing on τ . 5 2.3. Homogenized problem We will translate the problem of investigating the dimension of Rr ∆m,m to the homoge- neous setting [2, 3]. To this end, let us introduce the ring of bi-homogeneous polynomials S := R[u, v] = R[s, t, u, v] which is interpreted as the extension of R by the variables u and v that homogenize s and t, respectively. We denote the associated vector space of bi- homogeneous polynomials of bi-degree exactly m = (m1, m2) ∈ Z2 ≥0 with Sm ≡ Sm1m2 ⊂ S. This vector space is spanned by the monomials sium1−itjvm2−j, 0 ≤ i ≤ m1, 0 ≤ j ≤ m2. If any of m1, m2 are negative then Sm := 0. The ring S is naturally graded by Z2, S = M(i,j)∈Z2 Sij , SijSkl = S(i+k)(j+l) , (2.7) and its graded pieces are shifted in the usual manner: S(−i, −j)kl = S(k−i)(l−j). For convenience, let us define the following notation s := (s, t), u := (u, v), and for any tuple (a, b), (a, b)(i,j) := aibj , for i, j ∈ Z≥0. Using the above, we define the vector space associated to (cid:3) ∈ T2 ∪ T1 ∪ T0 as S(cid:3) :=(cid:16)u∆m((cid:3))(cid:17) = u∆m((cid:3))S(−∆m((cid:3))) . (2.8) In particular, given m ∈ Z2 An algebraic characterization of smoothness for bi-homogeneous piecewise polynomial func- tions follows in the vein of Lemma 2.8, and is stated below. The module of bi-homogeneous splines of interest is defined immediately thereafter. ≥0, we denote its mth graded piece as S(cid:3),m = u∆m((cid:3))S(−∆m((cid:3)))m. Lemma 2.9. For σ, σ′ ∈ T2, let σ ∩ σ′ = τ ∈ T1. A bi-homogeneous piecewise polyno- mial function equaling p ∈ S and p′ ∈ S on σ and σ′, respectively, is at least r(τ ) times continuously differentiable across τ if and only if ◦ p − p′ ∈(cid:16)lr(τ )+1 τ (cid:17) , where lτ is a non-zero u-homogeneous (resp. v-homogeneous) linear polynomial vanishing on τ ∈ Tv 1 (resp. τ ∈ Th 1). Definition 2.10 (Module of bi-homogeneous splines). Given T-mesh T , degree and smooth- ness distributions ∆m and r, respectively, we define the module of bi-homogeneous splines S r ∆m(T ) as S r ∆m ≡ S r ∆m(T ) :=(cid:26)f : ∀σ ∈ T2, f σ ∈ Sσ , ◦ ∀τ ∈ T1, f is C r(τ ) across τ } . (2.9) For a given m ∈ Z2 ≥0, the above S-module of splines is of interest precisely because its m-th graded piece is isomorphic to Rr ∆m,m. Theorem 2.11. dim (S r ∆m)m = dim(cid:0)Rr ∆m,m(cid:1) . 6 3. Topological complexes In this section we will describe the tools from homology that we will use for computing the dimension of graded pieces of S r ∆m but first let us introduce the relevant notation. First, we define the ideal L[i](−(j, k)) ≡ L[i](−j, −k) :=(0 , u∆miS(−∆mi − (j, k)) , i = l + 1 0 ≤ i ≤ l , and denote L[i] ≡ L[i](0, 0). Then, from Equation (2.2), we have for i = 1, . . . , l, L[i](−j, −k) = u∆ni L[i−1](−(j, k) − ∆ni) . We also define the (shifted) quotient M[i](−j, −k) := L[i−1](−j, −k)/L[i](−j, −k) . (3.1) (3.2) (3.3) ◦ T1, it will be convenient to set ∆τ := lr(τ )+1 For τ ∈ , where lτ is the homogeneous linear polynomial from Lemma 2.9. We will use ∆τ to define the edge and vertex associated ideals Iτ and Iγ, τ Iτ :=(cid:16)∆τ u∆m(τ )(cid:17) ⊂ Sτ , Iγ :=Xγ∈τ(cid:16)∆τ u∆m(τ )(cid:17) ⊂ Sγ . (3.4) In general, while Iτ = Sτ ∩ (∆τ ), we have Iγ = Sγ ∩Pγ∈τ (∆τ ) only when γ is a crossing vertex. 3.1. Definitions Orienting the i-cells in a compatible manner, i = 0, 1, 2, they will generate S-modules, and we will index the generators with the respective faces, edges and vertices. These indexed generators will be denoted with [σ], [τ] and [γ]. We will assume that all oriented 2-cells have been assigned a counter-clockwise orientation. For τ ∈ T1 with end points γ, γ′ ∈ T0, the generator associated will be represented as [τ] = [γγ′], with [γ′γ] = −[γγ′] defining the oppositely oriented edge. In the following sections we will only be interested in homology relative to ∂Ω. Therefore, we will assume that [τ] = 0 and [γ] = 0 when τ ⊂ ∂Ω and γ ∈ ∂Ω, respectively. Let εθ,φ ≡ ε([θ], [φ]) denote the orientation function, i.e., the function that takes as inputs one n-dimensional cell [θ], and one (n − 1)-dimensional cell, [φ], and returns • 0 if θ ∩ φ = ∅; • −1 if θ ∩ φ = φ and the orientation endowed by [θ] upon its boundary is incompatible with orientation of [φ]; and, • +1 if θ ∩ φ = φ and the orientation endowed by [θ] upon its boundary is compatible with orientation of [φ]. We will consider the usual boundary maps, ∂, defined as ∂([σ]) = Xτ ∈ T1 ◦ εσ,τ[τ] , ∂([τ]) = Xγ∈ T0 ◦ ετ,γ[γ] , ∂([γ]) = 0 . (3.5) 7 Then, for an element p = Pσ[σ]pσ of the S-module ⊕σ∈T2[σ]Sσ, its image under the action of ∂ will be (3.6) ∂ Xσ∈T2 [σ]pσ  = Xτ ∈ T1 ◦ [τ] Xσ∈T2 εσ,τpσ  . It is clear that Pσ εσ,τpσ ∈ Sτ (= u∆m(τ ) S). Therefore, by Lemma 2.9, for p to be in smoothness class C r we require the following, ∀τ ∈ ◦ T1 , Xσ∈T2 εσ,τpσ ∈ Iτ . (3.7) Then, S r the above requirement, with pσ = f σ. In other words, for m ∈ Z2 ∆m contains all splines f (in all bi-degrees) such that their polynomial pieces satisfy ≥0, dim(cid:0)Rr ∆m,m(cid:1) = dim (S r ∆m)m = dim(cid:0)ker(cid:0)∂(cid:1)(cid:1)m , where ∂, given below, is obtained by composing ∂ with the natural quotient map, ∂ : ⊕ σ∈T2 [σ]Sσ → ⊕ T1 τ ∈ ◦ [τ]Sτ /Iτ . 3.2. Degree-deficit based topological complexes (3.8) (3.9) In light of the above reasoning, we consider the following chain complex of S-modules as the object of our analysis, with the top homology module of Q equaling S r ∆m. Q : [σ]Sσ ⊕ σ∈T2 [τ]Sτ /Iτ ⊕ T1 τ ∈ ◦ ⊕ T0 γ∈ ◦ [γ]Sγ/Iγ 0 . (3.10) We analyze Q by performing a decomposition that untangles the individual contributions from the different degree-deficits from Equation (2.1) to the dimension of S r ∆m. First, for (cid:3) ∈ T2 ∪ T1 ∪ T0, we define S(cid:3),kik = S(cid:3)/(S(cid:3) ∩ L[i]) , S(cid:3),[i] = S(cid:3),kik ∩ L[i−1] , We also define the complex Q[i] as I(cid:3),kik = I(cid:3) · S(cid:3),kik , I(cid:3),[i] = I(cid:3) · S(cid:3),[i] . (3.11) (3.12) ⊕ σ∈T2 [σ]Sσ,[i] [τ]Sτ,[i]/Iτ,[i] ⊕ T1 τ ∈ ◦ ⊕ T0 γ∈ ◦ [γ]Sγ,[i]/Iγ,[i] 0 , (3.13) and the complex Qkik as ⊕ σ∈T2 [σ]Sσ,kik [τ]Sτ,kik/Iτ,kik ⊕ T1 τ ∈ ◦ ⊕ T0 γ∈ ◦ [γ]Sγ,kik/Iγ,kik 0 . (3.14) Let us make a few observations about the setup so far and the motivations behind it: • By construction of these complexes, we have a sequence of complexes 0 Q[i] Qkik Qki−1k 0 . (3.15) In other words, the maps induced on the corresponding components of Q[i], Qkik, Qki−1k form complexes. As we will see in Proposition 3.1, these complexes are exact. 8 • With regards to the analysis of Q from Equation (3.10), it is clear from the above definitions that Qkl+1k = Q. We will perform its analysis using the previous sequence of complexes. • It can be observed that each module in Qk0k is identically zero, thereby implying Qk1k = Q[1]. Therefore, it is only necessary to analyze the complexes Q[l+1] , Q[l] , . . . , Q[2] , Q[1] . • Finally, for a given m ∈ Z2 ≥0, it can be observed that the dimension of the m-th graded piece of the top homology module of Q[l+1] is equal to the dimension of Rr , 0,m−∆ml which is the largest uniform-degree spline space contained in Rr ∆m,m. Therefore, in , Q[1] represent the a rough sense, the contributions from the complexes Q[l] , incremental changes in the dimension of Rr that result from the introduc- tion of non-uniformity in bi-degrees. The net effect of these incremental changes on the dimension of Rr will intuitively bring us close to the quantity of interest, 0,m−∆ml . . . 0,m−∆ml dim(cid:16)Rr ∆m,m(cid:17) = dim (S r ∆m)m. Proposition 3.1. The following is a short exact sequence of complexes 0 Q[i] Qkik Qki−1k 0 . (3.16) Proof. We have to prove that the following is a short exact sequence for all (cid:3) ∈ T2 ∪ T1 ∪ T0: 0 S(cid:3),[i]/I(cid:3),[i] S(cid:3),kik/I(cid:3),kik S(cid:3),ki−1k/I(cid:3),ki−1k 0 , where Iσ,[i] := 0. Given i, if S(cid:3) ( L[i−1], then we must have S(cid:3) ⊆ L[i], which would imply that all spaces of the above complex are identically 0. Therefore, the only non-trivial case to consider is when S(cid:3) ⊇ L[i−1]. Explicitly, the non-trivial cases to analyze yield the following complexes: L[i−1]/L[i] Sσ/L[i] Sσ/L[i−1] 0 , (a) 0 0 0 L[i−1]/(cid:0)Iτ ∩ L[i−1] + L[i](cid:1) L[i−1]/(cid:0)Iγ ∩ L[i−1] + L[i](cid:1) Sτ /(cid:0)Iτ + L[i](cid:1) Sγ/(cid:0)Iγ + L[i](cid:1) Sτ /(cid:0)Iτ + L[i−1](cid:1) Sγ/(cid:0)Iγ + L[i−1](cid:1) 0 , (b) 0 . (c) It is easy to see that (a) is a short exact sequence. The same observation follows for (b) and (c) since, for (cid:3) ∈ T1 ∪ T0, by definition we have L[i−1] ∩(cid:0)I(cid:3) + L[i](cid:1) =(cid:0)I(cid:3) ∩ L[i−1] + L[i](cid:1) , which transforms (b) and (c) into 0 (cid:0)I(cid:3) + L[i−1](cid:1) /(cid:0)I(cid:3) + L[i](cid:1) S(cid:3)/(cid:0)I(cid:3) + L[i](cid:1) S(cid:3)/(cid:0)I(cid:3) + L[i−1](cid:1) 0 . (cid:4) 9 As stated above, Q can be studied by studying the complexes Q[i], i = 1, 2, . . . , l + 1. We do so by analyzing the following short exact sequence of chain complexes for each i. I[i] : 0 C[i] : ⊕ σ∈T2,[i] [σ]Sσ,[i] Q[i] : ⊕ σ∈T2,[i] [σ]Sσ,[i] 0 [τ]Iτ,[i] [τ]Sτ,[i] ⊕ T1,[i] ◦ τ ∈ ⊕ T1,[i] ◦ τ ∈ 0 [γ]Iγ,[i] [γ]Sγ,[i] ⊕ T0,[i] ◦ γ∈ ⊕ T0,[i] ◦ γ∈ [τ]Sτ,[i]/Iτ,[i] ⊕ T1,[i] ◦ τ ∈ [γ]Sγ,[i]/Iγ,[i] ⊕ T0,[i] ◦ γ∈ 0 0 0 0 0 (3.17) Note that the morphisms above are obtained in the obvious way by composing (restrictions of) ∂ with quotient maps. In Equation (3.17), T2,[i], T0,[i] are the active components of the mesh with respect to the index i; this notion was hinted at in the proof of Proposition 3.1 and is defined next. T1,[i] and ◦ ◦ Definition 3.2 (Active T-mesh). For i = 1, 2, . . . , l + 1, the active T-mesh T[i] is composed of • T2,[i] ⊆ T2 such that σ ∈ T2,[i] def. ⇐⇒ Sσ ⊇ L[i−1]; • T1,[i] ⊆ T1 such that τ ∈ T1,[i] def. ⇐⇒ Sτ ⊇ L[i−1]; and, • T0,[i] ⊆ T0 such that γ ∈ T0,[i] def. ⇐⇒ Sγ ⊇ L[i−1]. The domain of this active T-mesh, Ω[i], is defined to be ∪σ∈T2,[i] σ ⊂ R2. The symbols for interior edges, vertices, horizontal and vertical edges etc. are all appended with a subscript of [i] when talking about the active mesh T[i]; see Equation (3.17). Note that "interior" will always mean interior with respect to Ω. It should be noted that T1,[i] is exactly the set of edges that are contained in ∪σ∈T2,[i] ∂σ. Similarly, T0,[i] is exactly the set of vertices that are contained in ∪τ ∈T1,[i] ∂τ . Remark 3.3. Note that Equation (2.2) may introduce more active meshes than strictly nec- essary. However, we choose the degree-deficit sequence in compliance with Equation (2.2) because it simplifies the analysis later on. In particular, the results that are affected by this simplification are Lemmas 6.9 and 6.11 (and those that depend on these lemmas). Example 3.4. Consider the setup in Example 2.7, and let us choose the shorter sequence of degree deficits provided therein. Then, the associated active meshes with respect to i = 1, 2, 3 are shown in Figure 3. The bottom, middle and top layers correspond to T[3], T[2] and T[1], respectively. In other words, the layer corresponding to T[i] is such that for the faces σ, edges τ and vertices γ contained in it, we have the containment S(cid:3) ⊇ L[i−1], (cid:3) ∈ {σ, τ, γ}. 10 H0(Ω[i], ∂Ω[i] ∩ ∂Ω) ∼= Z H1(Ω[i], ∂Ω[i] ∩ ∂Ω) = 0 H0(Ω[i], ∂Ω[i] ∩ ∂Ω) = 0 i = 1 i = 2 i = 3 Figure 3: Active regions of the T-mesh T from Figure 1 for different indices i ∈ {1, 2, 3}; see Example 3.4 for reference. Generators of H0(Ω[i], ∂Ω[i] ∩ ∂Ω) have been shown as gray disks. Bound- aries of the meshes have been emphasized in bold and solid lines correspond to the active boundary, i.e., ∂Ω ∩ Ω[i]. 3.3. Summary of approach Given m ∈ Z2 ≥0, let χ (A)m be the Euler characteristic of the m-th graded piece of the complex A : 0 → Ak → Ak−1 → · · · → A0 → 0, χ (A)m = k (−1)j dim (Hj(A))m = Xj=0 k Xj=0 (−1)j dim (Aj)m . (3.18) Then the Euler characteristic of Q = Qkl+1k helps quantify the homological contribution to the dimension of S r ∆m. Definition 3.5 (Homological contribution to dimension). Given m ∈ Z2 homological contribution to the dimension of S r ∆m in bi-degree m as ≥0, we define the ∆m,m = dim (S r hr ∆m)m − χ (Q)m , so that we have dim (S r ∆m)m = χ (Q)m + hr ∆m,m. It will be shown in Section 3.4 that the Euler characteristic of Q is computable exactly using the rightmost expression in Equation (3.18). This leaves only the computation (or estimation) of hr ∆m. We approach this task as follows. First, using the short exact sequence from Equation (3.15), we build the long exact sequence of homology modules ∆m,m for the purpose of determining (bounds on) the dimension of S r 0 H2(cid:0)Q[i](cid:1) H2(cid:0)Qkik(cid:1) H2(cid:0)Qki−1k(cid:1) H1(cid:0)Q[i](cid:1) ∂i 11 · · · H0(cid:0)Qki−1k(cid:1) 0 (3.19) with ∂1 ≡ 0. The above implies that χ(cid:0)Qkik(cid:1)m = χ(cid:0)Q[i](cid:1)m + χ(cid:0)Qki−1k(cid:1)m . Summing up the terms for i = 1 . . . , l + 1, we obtain χ (Q)m = χ(cid:0)Qkl+1k(cid:1)m = l+1 Xi=1 χ(cid:0)Q[i](cid:1)m since Qk0k = 0 and Qkl+1k = Q. That is, the Euler characteristic of Q decomposes additively in the above manner. The next results simply states a similar additive decomposition does not hold for the 2-homologies. Lemma 3.6. dim (S r ∆m)m = l+1 Xi=1 dim(cid:0)H2(cid:0)Q[i](cid:1)(cid:1)m − dim(cid:16)im ∂i(cid:17)m . Proof. From Equation (3.19), we have the following exact sequence for all i ≥ 1, 0 H2(cid:0)Q[i](cid:1) The above implies that H2(cid:0)Qkik(cid:1) H2(cid:0)Qki−1k(cid:1) ∂i im ∂i 0 . dim(cid:0)H2(cid:0)Qkik(cid:1)(cid:1)m = dim(cid:0)H2(cid:0)Qki−1k(cid:1)(cid:1)m + dim(cid:0)H2(cid:0)Q[i](cid:1)(cid:1)m − dim(cid:16)im ∂i(cid:17)m The claim follows upon summing up the terms for i = 1 . . . , l + 1. . (cid:4) Using Lemma 3.6, we can simplify the expression for χ (Q)m, l+1 χ (Q)m = Xi=1(cid:16)dim(cid:0)H2(cid:0)Q[i](cid:1)(cid:1)m − dim(cid:0)H1(cid:0)Q[i](cid:1)(cid:1)m + dim(cid:0)H0(cid:0)Q[i](cid:1)(cid:1)m(cid:17) , l+1 = dim(S r ∆m)m + Xi=1(cid:16)dim(cid:16)im ∂i(cid:17)m − dim(cid:0)H1(cid:0)Q[i](cid:1)(cid:1)m + dim(cid:0)H0(cid:0)Q[i](cid:1)(cid:1)m(cid:17) . A final simplification is afforded by the following assumption. Assumption 3.7. All complexes C[i] are without holes, i.e., H1(cid:0)C[i](cid:1) = 0 for all i. Proposition 3.8. hr ∆m,m := dim (S r ∆m)m − χ (Q)m , = l+1 Xi=1 dim(cid:0)H0(cid:0)I[i](cid:1)(cid:1)m − dim(cid:0)H0(cid:0)C[i](cid:1)(cid:1)m − dim(cid:16)im ∂i(cid:17)m . Proof. The proof follows from Lemma 3.6 and the diagram in Equation (3.17). following Assumption 3.7, we obtain the long exact sequence of homology modules Indeed, 0 H1(cid:0)Q[i](cid:1) H0(cid:0)I[i](cid:1) H0(cid:0)C[i](cid:1) which yields the claim when combined with Equation (3.18). H0(cid:0)Q[i](cid:1) 0 , (cid:4) 12 γ3 γ5 γ7 γ2 γ12 γ13 γ17 γ14 γ15 σ σ′ γ9 γ16 γ10 γ11 γ4 γ6 γ1 γ8 γ0 Figure 4: A mesh that serves to counter the expectation that the spline space dimension can be additively decomposed in the same manner as the Euler characteristic of Q; see Example 3.9 for details. Therefore, the final problem is the computation (or estimation) of the dimension of im ∂im and the difference in the dimensions of H0(cid:0)I[i](cid:1)m and H0(cid:0)C[i](cid:1)m for all i. The next example shows that in general dim(cid:16)im ∂i(cid:17)m Example 3.9. Consider the T-mesh in Figure 4. Assume that the degree deficit on all faces touching the boundary is (1, 1), and on the remaining faces is (0, 0). Choose the asssociated degree-deficit sequence to be ∆m0 = (0, 0) < ∆m1 = (1, 1). Let us ask for C 2 smoothness across all edges, and let m = (5, 5). Then, based on the definitions, it can be checked that is not equal to zero. dim (S r ∆m)m = 81 . However, we can also compute (using Macaulay2, for instance) that Thus, dim (S r dim(cid:0)H2(cid:0)Q[2](cid:1)(cid:1)m = 81 , In particular, the spline generating H2(cid:0)Q[2](cid:1) in degree m is [σ]p + [σ′]q where i=1 dim(cid:0)H2(cid:0)Q[i](cid:1)(cid:1)m. Equivalently, we have that dim(cid:16)im ∂2(cid:17) = 1. dim(cid:0)H2(cid:0)Q[1](cid:1)(cid:1)m = 1 . ∆m)m 6=P2 q = t5u(3s − 2u)3(9s − 4u) . p = t5u(3s − u)3(9s − 5u) , It can be checked that p − q = 27t5u2(2s − u)3 ∈ Iτ , τ = γ16γ17, and that p, q ∈ Iτ + (uv) where τ is the edge γ13γ17 or γ9γ16. 3.4. Dimension formulas for relevant vector spaces Before proceeding, we present combinatorial formulas for the dimensions of the different vector spaces that have appeared so far and are relevant for our analysis. In the following, a+ := max{a, 0} for a ∈ Z≥0, and m = (m, m′) ∈ Z2 ≥0. We will also need the bi-smoothness associated to an interior edge τ , δ(τ ) :=((0, r(τ ) + 1) , (r(τ ) + 1, 0) , τ ∈ τ ∈ ◦ Th 1 . (3.20) ◦ Tv 1 For an interior vertex γ ∈ τh ∩ τv, (τh, τv) ∈ 1 × ◦ Th 13 ◦ Tv 1, we define δ(γ) := δ(τh) + δ(τv). Proposition 3.10. Let τ ∈ Then the following hold, where for (c) − (e) we assume that 1 ≤ i ≤ l + 1. T0,[i] such that γ = τh ∩ τv, (τh, τv) ∈ T1,[i], and let γ ∈ ◦ ◦ , 0, 0 ≤ i ≤ l dim (S(−j, −k))m = (m − j + 1)+ ×(cid:0)m′ − k + 1(cid:1)+ . dim(cid:0)L[i](−j, −k)(cid:1)m =(dim (S(−j, −k))m−∆mi dim(cid:0)M[i](−j, −k)(cid:1)m = dim(cid:0)L[i−1](−j, −k)(cid:1)m − dim(cid:0)L[i](−j, −k)(cid:1)m . dim(cid:0)Iτ,[i](cid:1)m dim(cid:0)Iγ,[i](cid:1)m = dim (S(−δ(τh) − ∆mi−1))m + dim (S(−δ(τv) − ∆mi−1))m = dim (S(−δ(τ ) − ∆mi−1))m − dim (S(−δ(τ ) − ∆mi))m . + dim (S(−δ(γ) − ∆mi))m − dim (S(−δ(γ) − ∆mi−1))m − dim (S(−δ(τh) − ∆mi))m − dim (S(−δ(τv) − ∆mi))m . i = l + 1 . ◦ Th 1 × ◦ Tv 1. (a) (b) (c) (d) (e) Proof. Claims made in parts (a), (b) and (c) follow by definition. Parts (d) and (e) follow from the following exact sequences, respectively, S(−δ(τ ) − ∆mi) S(−δ(τ ) − ∆mi−1) ⊕ S(−∆mi) Iτ ∩ L[i−1] + L[i] , S(−δ(τh) − ∆mi) S(−δ(τh) − ∆mi−1) ⊕ ⊕ S(−δ(γ) − ∆mi) S(−δ(τv) − ∆mi) S(−δ(τv) − ∆mi−1) Iγ ∩ L[i−1] + L[i] . ⊕ S(−δ(γ) − ∆mi−1) ⊕ S(−∆mi) We can simplify the expression in Proposition 3.10(e) for special choices of ∆m, r, m. (cid:4) Corollary 3.11. Let 1 ≤ i ≤ l and γ ∈ ◦ T0,[i] such that γ = τh ∩ τv, (τh, τv) ∈ (a) If m − ∆mi is entry-wise greater than or equal to (r(τv), r(τh)), then ◦ Th 1 × ◦ Tv 1. dim(cid:0)Iγ,[i](cid:1)m = (m − ∆m(i−1)1 + 1)∆ni2 + (m′ − ∆m(i−1)2 + 1)∆ni1 − ∆ni1∆ni2 . = dim(cid:0)M[i](cid:1)m = dim(cid:0)L[i−1]/L[i](cid:1)m (b) If m − ∆mi−1 is entry-wise smaller than or equal to (r(τv), r(τh)), then dim(cid:0)Iγ,[i](cid:1)m = 0 . The next result follows from the definition of the quotient modules M[i](−j, −k) and Equation 2.2. 14 Lemma 3.12. The following hold for b ∈ Z2 (b) − (d). ≥0, where 1 ≤ i ≤ l for (a) and 1 ≤ i ≤ l for . dim(cid:16)sbM[i](−j, −k)(cid:17)m = dim(cid:0)M[i](−j, −k)(cid:1)m−b dim(cid:0)uM[i](−j, −k)(cid:1)m = dim(cid:0)uL[i−1](−j, −k)(cid:1)m − dim(cid:16)u(1−∆ni1)+ L[i](−j, −k)(cid:17)m dim(cid:0)vM[i](−j, −k)(cid:1)m = dim(cid:0)vL[i−1](−j, −k)(cid:1)m − dim(cid:16)v(1−∆ni2)+ L[i](−j, −k)(cid:17)m dim(cid:0)uvM[i](−j, −k)(cid:1)m = 0 . . . (a) (b) (c) (d) The next result has been adapted from [20] in a form relevant for our analysis; see [20] for its proof. Proposition 3.13. Consider b ∈ Z2 Z≥0. Denote with fk the linear polynomials t − akv. Then, we have ≥0, l distinct numbers a1, . . . , al ∈ R, and d1, . . . , dl ∈ = (m + 1) × min m′ + 1, l Xk=1 m′ − dk + 1! . k S(0, −dk)!m−∆mi−b . f dk dim l Xk=1 dim l Xk=1 dim l Xk=1 f dk k S(0, −dk)!m L[i](−b − (0, dk))!m M[i](−b − (0, dk))!m f dk k f dk k = dim l Xk=1 = dim l Xk=1 f dk k L[i−1](−b − (0, dk))!(cid:14)L[i](−b − (0, dk))!m . (a) (b) (c) Symmetric claims can be made if the linear polynomials are instead chosen to be s − aku. 4. Homology of C[i] In this section we collect main results characterizing the homology of the chain complex C[i]. We will use its following equivalent form that follows from the proof of Proposition 3.1, C[i] : ⊕ σ∈T2,[i] [σ]M[i] Proposition 4.1. ◦ ⊕ T1,[i] τ ∈ [τ]M[i] [γ]M[i] 0 . ◦ ⊕ T0,[i] γ∈ dim(cid:0)H2(cid:0)C[i](cid:1)(cid:1)m =(dim(cid:0)L[l](cid:1)m , 0 , i = l + 1 1 ≤ i ≤ l . Proof. Let p =Pσ[σ]pσ, pσ ∈ L[i−1], be in the kernel of ∂, i.e., T1,[i] , Xσ∈T2,[i] 0 = ∂(p) = Xτ ∈ [τ] Xσ∈T2,[i] εσ,τpσ ⇔ ∀τ ∈ T1,[i] ◦ ◦ εσ,τpσ ∈ L[i] . If any σ and σ′ share an edge τ , εσ,τ = −εσ′,τ . Therefore, if both σ and σ′ also belong to T2,[i], we must have pσ − pσ′ ∈ L[i]. Then, 15 • i = l + 1: Following Assumption 2.3, all edges in T1,[l+1] are shared by exactly two faces in T2,[l+1] ≡ T2. Therefore, all pσ must correspond to the same global polynomial in L[l] for all faces in T2. ◦ • i < l + 1: There exists at least one edge in T1,[i] that is contained in only one face in T2,[i]. Therefore, all pσ must be polynomials in L[i], and therefore must be zero in L[i−1]/L[i]. ◦ (cid:4) Definition 4.2 (Number of relative holes in Ω[i]). We define π[i] to be the number of linearly independent, non-trivial cycles in Ω[i] relative to ∂Ω[i] ∩ ∂Ω, π[i] := rank(cid:16)H1(cid:16)Ω[i], ∂Ω[i] ∩ ∂Ω(cid:17)(cid:17) . Proposition 4.3. dim(cid:0)H1(cid:0)C[i](cid:1)(cid:1)m = π[i] dim(cid:0)M[i](cid:1)m Proof. The entire kernel of . ∂ : ⊕ ◦ [τ]M[i] → ⊕ ◦ [γ]M[i] τ ∈ T1,[i] γ∈ T0,[i] can be generated by (R-linear combinations of) ck,i of the form ck,i = pi−1αk + Xτ ∈ ◦ T1,[i] [τ]pτ,i , where [τ]oτ , oτ ∈ Z, is a relative cycle, i.e., ∂αk = 0; and, • αk =Pτ ∈ ◦ T1,[i] • pi−1 ∈ L[i−1] and pτ,i ∈ L[i]. Then, we only need to see how many such ck,i are linearly independent and not nullhomol- In particular, if pi−1 /∈ L[i], ck,i is nullhomologous iff there exists some dk,i of the ogous. form dk,i = pi−1βk + Xσ∈T2,[i] [σ]pσ,i such that ∂βk = αk, where • pσ,i ∈ L[i]. [σ]oσ, oσ ∈ Z; and, • βk =Pσ∈T2,[i] Then, ck,i is not nullhomologous in H1(cid:0)C[i](cid:1) iff αk is not nullhomologous in H1(cid:16)Ω[i], ∂Ω[i] ∩ ∂Ω(cid:17). (cid:4) Definition 4.4 (Number of relative connected components in Ω[i]). We define N[i] to be the number of connected components in Ω[i] relative to ∂Ω[i] ∩ ∂Ω, N[i] := rank(cid:16)H0(cid:16)Ω[i], ∂Ω[i] ∩ ∂Ω(cid:17)(cid:17) . 16 Proposition 4.5. dim(cid:0)H0(cid:0)C[i](cid:1)(cid:1)m = N[i] dim(cid:0)M[i](cid:1)m . ◦ Proof. All [γ]pγ, γ ∈ T1,[i], pτl ∈ L[i−1], and o1, . . . , ok ∈ Z be such that ◦ T0,[i], pγ ∈ L[i−1], are in the kernel of ∂. Let vertex γ0, edges τ1, . . . , τk ∈ • ∀l ∈ {1, . . . , k}, pτl − pγ ∈ L[i]; and, l=1[τl]ol(cid:17) . Assuming pγ /∈ L[i], [γ]pγ is nullhomologous only if γ and γ0 belong to the same connected • [γ] = [γ0] + ∂(cid:16)Pk component of Ω[i] and γ0 ∈ ∂Ω. Else, [γ0]pγ would be a generator of H0(cid:0)C[i](cid:1) for the particular connected component of Ω[i] that it belongs to. Then, for each pγ, the number of such generators is exactly equal to N[i], and the claim follows. (cid:4) Example 4.6. Consider the setup in Example 3.4 and Figure 3. Then, for m = (5, 5), it can be verified that: 0 , =(16 , dim(cid:0)H2(C[i])(cid:1)m dim(cid:0)H1(C[i])(cid:1)m = 0 , = dim(cid:0)H0(C[i])(cid:1)m  9 , 0 , 11 , 5. The 0-homology of I[i] i = 3 i = 1, 2 , i = 1, 2, 3 , i = 3 i = 2 . i = 1 As per our objectives stated at the end of Section 3.3, only the characterization of the 0-homology of I[i] remains, and we collect the associated results in this section. Similarly to the previous section, we will do so keeping in mind the simplified form of I[i] that follows from Proposition 3.1, I[i] : 0 ◦ ⊕ T1,[i] τ ∈ [τ]Iτ ∩ L[i−1] + L[i]/L[i] [γ]Iγ ∩ L[i−1] + L[i]/L[i] 0 . ◦ ⊕ T0,[i] γ∈ We first provide a lower bound on the dimension of H0(cid:0)I[i](cid:1) that holds for special choices of ∆m, r and m. Proposition 5.1. Let m ∈ Z2 ≥0 and 1 ≤ i ≤ l. (a) If m − ∆mi is entry-wise greater than or equal to (rγ,h, rγ,v) for each γ ∈ ◦ T0,[i], then dim(cid:0)H0(cid:0)I[i](cid:1)(cid:1)m ≥ dim(cid:0)H0(cid:0)C[i](cid:1)(cid:1)m . In particular, the map from H0(cid:0)I[i](cid:1)m to H0(cid:0)C[i](cid:1)m in Equation (3.19) is a surjection and H0(cid:0)Q[i](cid:1)m vanishes. 17 (b) If m − ∆mi−1 is entry-wise smaller than or equal to (rγ,h, rγ,v) for each γ ∈ ◦ T0,[i], then dim(cid:0)H0(cid:0)I[i](cid:1)(cid:1)m = 0 . In particular, the map from H0(cid:0)C[i](cid:1)m to H0(cid:0)Q[i](cid:1)m in Equation (3.19) is an isomor- phism and H1(cid:0)Q[i](cid:1)m vanishes. Proof. (a) If m−∆mi is entry-wise greater than or equal to (rγ,h, rγ,v), then Corollary 3.11 implies that the mth graded piece of ⊕ ◦ γ∈ T0,[i] (b) If m − ∆mi−1 is entry-wise smaller than or equal to (rγ,h, rγ,v), then Corollary 3.11 implies that the mth graded piece of ⊕ [γ]Sγ,[i]/Iγ,[i] vanishes and so does H0(cid:0)Q[i](cid:1)m. [γ]Iγ,[i] vanishes and so does H0(cid:0)I[i](cid:1)m. T0,[i] γ∈ ◦ The claims then follow from Equation (3.19). Let us define the graded multiplication map φ[i],γ for γ ∈ ◦ T0,[i], φ[i],γ : ⊕ T1,[i] ◦ τ ∈ [γτ ]M[i](−δ(τ )) → [γ]Iγ,[i] [γτ ]p 7→ [γ]p∆τ , (cid:4) (5.1) where [γτ ] is a half-edge element, with [γτ ] := 0 when ετ,γ = 0 or when γ ∈ ∂Ω. Lemma 5.2. The map φ[i],γ is surjective. Proof. The claim follows from the isomorphism Iτ ∼= S(−∆m(τ ) − δ(τ )) and the surjective map ⊕ T1,[i] ◦ τ ∈ [γτ ]Iτ ∩ L[i−1] → [γ]Iγ ∩ L[i−1] [γτ ]p 7→ [γ]p . (cid:4) ◦ ◦ Define Eh,[i](γ), Ev,[i](γ) as the sets of horizontal and vertical edges in T1,[i] that contain γ ∈ T0,[i], respectively, and let E[i](γ) = Eh,[i](γ) ∪ Ev,[i](γ). Let P[i](γ) be the set that contains edge-pairs (τ, τ ′) containing γ, both either in Eh,[i](γ) or Ev,[i](γ); we identify (τ, τ ′) with (τ ′, τ ). Note that, depending on the index i, P[i](γ) may be empty, or may contain either one or two elements. When the vertex γ is obvious from the context, we will exclude it from the above notation to simplify the reading (and writing!) of the text. Proposition 5.3. ker(cid:0)φ[i],γ(cid:1) = X(τ,τ ′)∈P[i](cid:0)[γτ ] − [γτ ′](cid:1) M[i](−δ(τ )) + Xτ ∈Ev,[i] τ ′∈Eh,[i] (cid:0)[γτ ]∆τ ′ − [γτ ′]∆τ(cid:1) M[i](−δ(γ)) . 18 Proof. With respect to the T-mesh T[i], let γ be a crossing vertex. (The proof for when γ is a T-junction will proceed analogously.) Let τ1, τ2 ∈ Eh,[i], τ3, τ4 ∈ Ev,[i], and P[i] = (cid:8)(τ1, τ2), (τ3, τ4)(cid:9), and let pj ∈ M[i](−δ(τj )). Then, φ[i],γ  4 [γτj]pj Xj=1 4 pj∆τj , Xj=1  = [γ] = [γ] (p1 + p2) ∆τ1 + [γ] (p3 + p4) ∆τ3 , where ∆τ1 = ∆τ2, and ∆τ3 = ∆τ4, and ∆τ1 and ∆τ3 are relatively prime. Therefore, the kernel of the map is generated by, ([γτ1] − [γτ2]) p12 , ([γτ3] − [γτ4]) p34 , ([γτh]∆τv − [γτv]∆τh) phv , where, p12 ∈ M[i](−δ(τ1)) , p34 ∈ M[i](−δ(τ3)) , phv ∈ M[i](−δ(γ)) . Using φ[i],γ, we can define a map φ[i] as ◦ ⊕ T0,[i] γ∈ φ[i] : with kernel, ◦ ⊕ T1,[i] τ ∈ [γτ ]M[i](−δ(τ )) → ⊕ ◦ [γ]Iγ,[i] , γ∈ T0,[i] ker(cid:0)φ[i](cid:1) = Xγ∈ ◦ T0,[i] ker(cid:0)φ[i],γ(cid:1) . Next, let us consider the diagram (cid:4) (5.2) (5.3) ◦ ⊕ T1,[i] τ ∈ [τ]M[i](−δ(τ )) ∂ ◦ ⊕ T0,[i] γ∈ [γτ ]M[i](−δ(τ )) ◦ ⊕ T1,[i] τ ∈ ψ[i] [τ]Iτ,[i] ∂ , (5.4) φ[i] [γ]Iγ,[i] ⊕ T0,[i] ◦ γ∈ ⊕ T1,[i] ◦ τ ∈ where the maps ∂ and ∂ are the restrictions of the following maps to the active T-mesh, ∂ : [τ] 7→Xγ ετ,γ[γτ ] , and the graded map ψ[i] is defined as ψ[i] : [τ] 7→ [τ]∆τ . ∂ : [τ] 7→Xγ ετ,γ[γ] , (5.5) Lemma 5.4. The map ψ[i] is surjective. Proof. The claim follows from the isomorphism Iτ ∼= S(−∆m(τ ) − δ(τ )). (cid:4) 19 Lemma 5.5. H0(cid:0)I[i](cid:1) ∼= ⊕ γ∈ ◦ T0,[i] ◦ ⊕ T1,[i] τ ∈ [γτ ]M[i](−δ(τ ))(cid:30) ker(cid:0)φ[i](cid:1) + ∂  ◦ ⊕ T1,[i] τ ∈ [τ]M[i](−δ(τ ))    . Proof. The diagram in Equation (5.4) commutes. Indeed, [τ]p [τ]p∆τ Pγ[γτ ]ετ,γ p Pγ[γ]ετ,γ p∆τ . Then, the result follows from surjectivities of φ[i] and ψ[i] (Lemmas 5.2 and 5.4, respectively), and surjectivity of the induced morphism, φ⋆ : ◦ ⊕ T0,[i] γ∈ ◦ ⊕ T1,[i] τ ∈ [γτ ]M[i](−δ(τ )) (cid:14) im(cid:16) ∂(cid:17) → ⊕ γ∈ ◦ T0,[i] [γ]Iγ,[i] (cid:14) im (∂) , Xγ Xτ [γτ ]pγτ + im(cid:16) ∂(cid:17) 7→ φ[i] Xγ Xτ [γτ ]pγτ! + im (∂) . Indeed, the kernel of φ⋆ is exactly ker(cid:0)φ[i](cid:1) + im(cid:16) ∂(cid:17)/ im(cid:16) ∂(cid:17) and we have the isomorphism,  [γ]Iγ,[i](cid:14) im (∂) = H0(cid:0)I[i](cid:1) .  [γτ ]M[i](−δ(τ )) (cid:14) im(cid:16) ∂(cid:17)  (cid:14) ker (φ⋆) ∼= ⊕ ⊕ T1,[i] ⊕ T0,[i] T0,[i] γ∈ γ∈ τ ∈ ◦ ◦ ◦ (cid:4) Before proceeding, we first introduce the concept of maximal segments for the T-mesh T[i]. This will help us further simplify the half-edge based form of H0(cid:0)I[i](cid:1) from Proposition 5.5. Definition 5.6 (Active maximal segments). Given index i ∈ {1, . . . , l + 1}, the set of active horizontal (resp. vertical) maximal segments MSh [i]) is the set containing 1,[i] (resp. Tv maximal connected unions of edges in Th [i] (resp. MSv 1,[i]). ◦ The set of all active maximal segments will be denoted by MS[i] = MSh [i], while the set of active maximal segments that do not intersect the boundary will be denoted by MS[i]; with some abuse of notation, we will refer to these maximal segments as "interior maximal segments." By definition of the smoothness distribution, we can unambiguously define δ(ρ) = δ(τ ) and ∆ρ = ∆τ for any edge τ ⊆ ρ ∈ [i] ∪ MSv ◦ MS[i]. Proposition 5.7. H0(cid:0)I[i](cid:1) ∼= ⊕ ρ∈ ◦ MS[i] [ρ]M[i](−δ(ρ))(cid:30) Xρh∈ ◦ MSh [i] Xρv∈ MSv ◦ [i] Xγ∈ ◦ T0,[i] ([ρh]∆ρv − [ρv]∆ρh) M[i](−δ(γ)) . ρh∩ρv={γ} 20 Proof. Using Equation (5.3) and Lemma 5.5, we define ετ,γ[γτ ]! M[i](−δ(τ )) + Xγ∈ T0,[i] X(τ,τ ′)∈P[i](cid:0)[γτ ] − [γτ ′](cid:1) M[i](−δ(τ )) , ◦ ◦ Xγ K ′ = Xτ ∈ K = K ′ + Xγ∈ T0,[i] Xτ ∈Ev,[i] T1,[i] ◦ B = ⊕ ◦ γ∈ T0,[i] ◦ ⊕ T1,[i] τ ∈ (cid:0)[γτ ]∆τ ′ − [γτ ′]∆τ(cid:1) M[i](−δ(γ)) , τ ′∈Eh,[i] [γτ ]M[i](−δ(τ )) . The first term of K ′ corresponds to relations yielding the identification of [γτ ] with [τ]. The second term of K ′ corresponds to relations yielding the identification of [γτ ] with [γτ ′] whenever τ, τ ′ ⊂ ρ. Therefore, K ′ leads to the identification of all edges that belong to the same maximal segment ρ ∈ MS[i]. Keeping the above in mind, and since B/K ∼= (B/K ′)/(K/K ′), we take the quotient with K ′. The required description is obtained by noticing that, since [γτ ] = 0 if γ ∈ ∂Ω, terms corresponding to [ρ] must be zero in the quotient for all active maximal segments that intersect ∂Ω. (cid:4) 6. Bounds on the dimension of Sr ∆m We will use Proposition 3.8 in this section to provide upper and lower bounds on the dimensions of graded pieces of S r ∆m. We start by providing the following lower and upper bounds on the spline space dimension. Some of the results presented here will assume that the condition of sufficiency in Proposition 5.1(a) is satisfied. Therefore, for the sake of convenience, we define the following configuration so that we can refer to it later. Configuration 6.1. The degree and smoothness distributions are such that, for all i, m − ∆mi is entry-wise greater than or equal to (rγ,h, rγ,v) for each γ ∈ ◦ T0,[i]. Theorem 6.2 (Lower bound for general smoothness distributions). dim (S r ∆m)m ≥ χ (Q)m − l+1 Xi=1 N[i] dim(cid:0)M[i](cid:1)m . Proof. The lower bound can be arrived at in exactly the same way as Proposition 3.8 but with a slightly different point of departure. Instead of using the short exact sequence in Equation 21 (3.15), we embed the complex Q directly in the short exact sequence 0 → I → C → Q → 0, 0 0 0 [σ]Sσ ⊕ σ∈T2 [τ]Iτ ⊕ T1 τ ∈ ◦ [τ]Sτ ⊕ T1 τ ∈ ◦ [γ]Iγ ⊕ T0 γ∈ ◦ [γ]Sγ ⊕ T0 γ∈ ◦ [σ]Sσ ⊕ σ∈T2 [τ]Sτ /Iτ ⊕ T1 τ ∈ ◦ [γ]Sγ/Iγ ⊕ T0 γ∈ ◦ 0 0 0 I : C : Q : 0 0 In a manner similar to the proof of Proposition 4.5, it is easy to establish that dim (H0(C))m = l+1 Xi=1 dim(cid:0)H0(C[i])(cid:1)m , dim (H1(C))m = l+1 Xi=1 dim(cid:0)H1(C[i])(cid:1)m . Then, following the same steps as in Section 3.3 but for the diagram above, one can derive the following equation, dim (S r ∆m)m = χ (Q)m + dim (H0(I))m − dim (H0(C))m . which yields the claimed lower bound. (cid:4) Before presenting a sharper lower bound on the spline space dimension, let us present , which are themselves often pessimistic in practical simple upper bounds on dim(cid:16)im ∂i(cid:17)m configurations. The following can be readily derived using Equation (3.19). Corollary 6.3. (a) In general, dim(cid:16)im ∂i(cid:17)m ≤ min(cid:26) dim(cid:0)H2(Qki−1k)(cid:1)m, dim(cid:0)H0(I[i])(cid:1)m − N[i] dim(cid:0)M[i](cid:1)m + dim(cid:0)H0(Q[i])(cid:1)m(cid:27) . (b) For Configuration 6.1, dim(cid:16)im ∂i(cid:17)m ≤ min(cid:26) dim(cid:0)H2(Qki−1k)(cid:1)m , dim(cid:0)H0(I[i])(cid:1)m − N[i] dim(cid:0)M[i](cid:1)m(cid:27) . Theorem 6.4 (Lower bound for practical smoothness distributions). For Configuration 6.1, dim (S r ∆m)m ≥ χ (Q)m . 22 Proof. Since the conditions of sufficiency in Proposition 5.1(a) is assumed to be satisfied for all 1 ≤ i ≤ l, H0(I[i])m surjects onto H0(C[i])m. For i = l + 1, the dimension of H0(I[i])m is trivially bounded from below by 0 = dim(cid:0)H0(C[i])(cid:1)m. From Corollary 6.3, this implies that − dim(cid:16)im ∂i(cid:17)m ≥ N[i] dim(cid:0)M[i](cid:1)m − dim(cid:0)H0(I[i])(cid:1)m . The claim follows from Proposition 3.8 and Proposition 4.5. (cid:4) Theorem 6.5 (Upper bound for general smoothness distributions). dim (S r ∆m)m ≤ χ (Q)m + l+1 Xi=1 dim(cid:0)H0(I[i])(cid:1)m − N[i] dim(cid:0)M[i](cid:1)m . ≥ 0, the claim follows from Proposition 3.8. Proof. Since dim(cid:16)im ∂i(cid:17)m It only remains to derive upper bounds on dim(cid:0)H0(I[i])(cid:1)m and we do so next. Given a particular i, we bound the dimensions of graded pieces of H0(cid:0)I[i](cid:1) from above by introducing utilizing the representation of H0(cid:0)I[i](cid:1) from Proposition 5.7. an ordering on the active interior maximal segments, i.e., on the elements of MS[i], i.e., an injective MS[i], define Γ[i](ρ) ⊂ MS[i] as the set of maximal map from segments ρ′ that intersect ρ non-trivially and such that either ξ[i](ρ) > ξ[i](ρ′) or ρ′ ∩ ∂Ω 6= ∅. MS[i]). Given i, let ξ[i] be an ordering on MS[i] to N. Given ξ[i] and ρ ∈ Definition 6.6 (Ordering of MS[i] and by (cid:4) ◦ ◦ ◦ ◦ ◦ Hereafter, we will assume that given i the ordering ξ[i] is fixed. We will abuse the notation by using ρ > ρ′ to mean the same thing as ξ[i](ρ) > ξ[i](ρ′). Let us define the modules M[i] := ⊕ ◦ [ρ]M[i](−δ(ρ)) , ρ∈ MS[i] D[i] := Xρh∈ ◦ MSh [i] Xρv ∈ MSv ◦ [i] Xγ∈ ◦ T0,[i] ([ρh]∆ρv − [ρv]∆ρh) M[i](−δ(γ)) . (6.1) (6.2) ρh∩ρv={γ} For p =Pρ[ρ]pρ ∈ M[i], we define its initial, denoted In p, as [ρ′]pρ′ if, out of all ρ such that pρ 6= 0, ρ′ has the biggest index according to ξ[i]. Lemma 6.7. H0(cid:0)I[i](cid:1) ∼= M[i]/In D[i] . Proof. The claim is a standard result; see [21], for example. Proposition 6.8. (cid:4) dim(cid:0)H0(I[i])(cid:1)m ≤ Xρ∈ ◦ MS[i]  dim(cid:0)M[i](−δ(ρ))(cid:1)m − dim  Xρ′∈Γ[i](ρ) ∆ρ′M[i](−δ(ρ) − δ(ρ′)) m   . Proof. From Lemma 6.7, if we can provide a lower bound on the dimension of In D[i], then we can provide an upper bound on dim(cid:0)H0(I[i])(cid:1)m Notice that In D[i] is going to be at least partially generated by the initials of its generators. MS[i] the contributions only come from the ρ′ ∈ Γ[i](ρ). . ◦ Looking at the generators, for each ρ ∈ The claim follows. 23 (cid:4) In practice, we have observed that if each connected component of Ω[i] intersects ∂Ω, the upper bound in Proposition 6.8 is usually optimal. Here we use "optimal" in the sense of the upper bound coinciding with the exact dimension of H0(I[i]) in bi-degree m. However, if ∂Ω[i] does not intersect ∂Ω, it turns out that the upper bound in Proposition 6.8 is a poor estimate. In other words, the initials of the generators of D[i] used in Proposition 6.8 do a bad job of approximating its dimension. Nonetheless, we can significantly improve upon this estimate by systematically adding some new generators to the ones used previously. In particular, doing so will allow us to compute the dimension of H0(I[i]) exactly even when there exist connected components of Ω[i] that do not intersect ∂Ω; see Sections 7 and 8 for examples. In turn, this will enable the use of the sufficient conditions outlined later in Section 7 for computing the exact spline space dimension in many cases. Following the above comments, for a given maximal segment ρ ∈ MS[i], consider the particular connected component of Ω[i] that ρ belongs to. Let us focus on enlarging the set of generators involving ρ in Equation (6.2). Define Υ[i](ρ) as the following set of maximal segment pairs, Υ[i](ρ) :=n(ρ1, ρ2) ∈ ◦ ◦ MS[i] × MS[i] : ρ2 ∩ ρ 6= ∅ 6= ρ2 ∩ ρ1, r(ρ) ≥ r(ρ1), ρ > ρ1o . (6.3) Note that in the above definition ρ and ρ1 must be parallel and ρ2 must be perpendicular to both. In Equation (6.2), the generators of D[i] already contain explicit relations between [ρ] and [ρ2], and between [ρ1] and [ρ2] for (ρ1, ρ2) ∈ Υ[i](ρ). Then, these generators can be manipulated to give a generator involving only [ρ] and [ρ1]. Lemma 6.9. Let 1 ≤ i ≤ l. Given maximal segment ρ ∈ (ρ1, ρ2) ∈ Υ[i](ρ). Recalling Equation (2.2), define ∆pρ as ◦ MS[i], let j (∆ni1, 0) , ρ ∈ MSh [i] , j (0, ∆ni2) , ρ ∈ MSv [i] . ∆pρ :=  Then, there is a polynomial ∆ such that ρ2 ρ ρ1 u∆pρ∆ρ2(cid:16)[ρ] − [ρ1] ∆(cid:17) M[i](−δ(ρ) − δ(ρ2) − ∆pρ) ⊂ D[i] . Proof. Without loss of generality, let ρ, ρ1 ∈ MSv definition of D[i], the following are two of its generators, [i] and define rρρ1 := r(ρ) − r(ρ1) ≥ 0. By ([ρ2]∆ρ1 − [ρ1]∆ρ2) M[i](−δ(ρ1) − δ(ρ2)) , ([ρ2]∆ρ − [ρ]∆ρ2) M[i](−δ(ρ) − δ(ρ2)) . (⋆1) (⋆2) Let ∆ρ1 = sr(ρ1)+1 and ∆ρ = (s + au)(r(ρ)+1), a ∈ R. We can write ∆ρ = ∆ρ1 ∆ + ∆′, where the term with the highest power of s in ∆′ is a multiple of sr(ρ1)urρρ1 +1. Then, we can combine the two generators in Equations (⋆1) and (⋆2) to yield (⋆2) − (⋆1) × ∆ =(cid:16)[ρ2]∆′ − [ρ]∆ρ2 + [ρ1]∆ρ2 We can further reduce the above relation to ∆(cid:17) M[i](−δ(ρ) − δ(ρ2)) ⊂ D[i] . u∆pρ∆ρ2(cid:16)[ρ] − [ρ1] ∆(cid:17) M[i](−δ(ρ) − δ(ρ2) − ∆pρ) ⊂ D[i] because L[i](−δ(ρ2)) = u∆ni L[i−1](−δ(ρ2) − ∆ni) from Equation (3.2) and ∆′ is a multiple of u. (cid:4) 24 Lemma 6.9 shows that, even if ρ2 > ρ, it may be a part of the contribution that [ρ] makes toward In D[i]. Given enough new generators of D[i] of this form, we can go a step further and identify some additional generators. To this end, given a bi-degree m, define the set Υ[i](ρ, m) as Υ[i](ρ, m) :=(cid:26)Υ ⊂ Υ[i](ρ) : X(·,ρ′)∈Υ ∆ρ′M[i](−δ(ρ′) − δ(ρ) − ∆pρ)m = M[i](−δ(ρ) − ∆pρ)m(cid:27) . (6.4) ◦ MS[i], 1 ≤ i ≤ l, and consider Υ ⊂ Υ[i](ρ). Corollary 6.10. Let m = (m, m′) ∈ Z2 Then Υ ∈ Υ[i](ρ, m) if ≥0, ρ ∈ (m − ∆mi1 − r(ρ′))+ ≥ m − ∆mi1 + 1 , ρ ∈ MSh [i] , (m′ − ∆mi2 − r(ρ′))+ ≥ m′ − ∆mi2 + 1 , ρ ∈ MSv [i] . (·,ρ′)∈Υ Xρ′ Xρ′ (·,ρ′)∈Υ   Proof. The proof follows directly from Proposition 3.13 since ∆m(i−1)j + ∆nij = ∆mij, j = 1, 2. (cid:4) Using the above definition, we define the set Θ[i](ρ, m) as Θ[i](ρ, m) :=(cid:26)(ρ1, ρ2) ∈ ◦ ◦ MS[i] × MS[i] : ρ1 ∩ ρ 6= ∅ 6= ρ2 ∩ ρ , r(ρ2) ≥ r(ρ1), ∃Υ ∈ Υ[i](ρ, m), ∀(ρ3, ·) ∈ Υ, ρ2 > ρ1 > ρ3(cid:27) . (6.5) Given a maximal segment ρ such that Υ[i](ρ, m) is not empty, we can identify further con- tributions to the initial In D[i],m from maximal segments that intersect ρ. The next result elucidates our reasoning. Lemma 6.11. For bi-degree m ∈ Z2 (ρ1, ρ2) ∈ Θ[i](ρ, m). Define ≥0 and fixed maximal segment ρ, let αρ1ρρ2 :=(1 , ∆pρ = 0 , 0 , ∆pρ 6= 0 , αρ2ρρ1 := 1 . Then, we have [ρ1]αρ1ρρ2∆ρM[i](−δ(ρ1) − δ(ρ))m ⊂ In D[i],m , [ρ2]αρ2ρρ1∆ρM[i](−δ(ρ2) − δ(ρ))m ⊂ In D[i],m . ρ ρ2 αρ2ρρ1 αρ1ρρ2 ρ1 Proof. From Lemma 6.9, there exists Υ ∈ Υ[i](ρ, m) and polynomials ηρ3 such that g0 = [ρ]u∆pρM[i](−δ(ρ) − ∆pρ)m − X(ρ3,ρ4)∈Υ [ρ3]u∆pρηρ3∆ρ4 M[i](−δ(ρ4) − δ(ρ) − ∆pρ)m is a relation in D[i],m. Then, as in the proof of Lemma 6.9, consider the relations between [ρ1] and [ρ], and between [ρ2] and [ρ]; denote these with g1 and g2, respectively. Thereafter, 25 if ∆pρ = 0, eliminate [ρ] from both g1 and g2 using g0 to get new relations g1 and g2; the claim will follow since ρ1, ρ2 > ρ3 for all (ρ3, ·) ∈ Υ. If, on the other hand, ∆pρ 6= 0, combine g1 and g2 to get a new relation g3 = g2 − g1 ∆. For an appropriate choice of ∆ (as in the proof of Lemma 6.9), it will be possible to eliminate [ρ] from g3, and the initial of g3 will thus be a part of the contribution that [ρ2] makes toward In D[i],m. (cid:4) Finally, we can toss in all the new generators identified in Lemmas 6.9 and 6.11 with the original set of generators in Equation (6.2), and define Dρ [i],m := M[i](−δ(ρ) − δ(ρ1) − ∆pρ)m (6.6) Pρ′∈Γ[i](ρ) ∆ρ′M[i](−δ(ρ) − δ(ρ′))m +P(·,ρ1)∈Υ[i](ρ) u∆pρ∆ρ1 +P(ρ,ρ2)∈Θ[i](ρ3,m) αρρ3ρ2∆ρ3 +P(ρ4,ρ)∈Θ[i](ρ5,m) αρ4ρ5ρ∆ρ5 M[i](−δ(ρ) − δ(ρ3))m M[i](−δ(ρ) − δ(ρ5))m . This is the contribution of [ρ] toward the initial of D[i] corresponding to the generators identified in Equation (6.2) and Lemmas 6.9 and 6.11. Then, we can use the above definition and Lemma 6.7 to provide an upper bound on the dimension of H0(I[i]) that improves upon the one presented in Proposition 6.8. Corollary 6.12. dim(cid:0)H0(I[i])(cid:1)m ≤ Xρ∈ ◦ MS[i] dim(cid:0)M[i](−δ(ρ))(cid:1)m − dim(cid:16)Dρ [i],m(cid:17) . 7. Configurations with stable dimension In this section, we outline sufficient conditions that guarantee that the dimension of the spline space S r ∆m,m can be computed exactly. We will work in a setting where Configuration 6.1 is true, i.e., where the bounds from Theorems 6.4 and 6.5 hold. Note that supplementary Macaulay2 scripts accompanying some of the examples presented in this section and the next can be downloaded from [28]. Assumption 7.1. For the rest of this section, we assume that Configuration 6.1 holds, i.e., the lower and upper bounds from Corollary 6.4 and Theorem 6.5 hold. Theorem 7.2. For all i, if dim(cid:0)H0(cid:0)I[i](cid:1)(cid:1)m ∆m)m = χ (Q)m . dim (S r = dim(cid:0)H0(cid:0)C[i](cid:1)(cid:1)m , then hr ∆m,m = 0, i.e., As per Theorem 7.2, if the upper bound in Corollary 6.12 equals dim(cid:0)H0(cid:0)C[i](cid:1)(cid:1)m then the dimension of the spline space can be exactly determined. Before stating sufficient conditions on T , ∆m, r, m that make this possible, let us first consider an example where we explicitly use the results from the previous section to compute the spline space dimension. Example 7.3. Consider the T-mesh shown in Figure 5 and let r(τ ) = 1 for all interior edges. Let us consider two different degree deficit distributions on this mesh and find the dimension of the resulting spline space. In all of the following cases, the bi-smoothness for each maximal segment is simply δ(ρ) = (2, 0) or (0, 2). We will also use the following fact that is implied by Proposition 3.13 for real numbers a1 6= a2 and m ≥ (3, 3), dim(cid:0)M[i](0, −2)(cid:1)m = dim(cid:0)(s + a1u)2M[i](−2, −2) + (s + a2u)2M[i](−2, −2)(cid:1)m , dim(cid:0)M[i](−2, 0)(cid:1)m = dim(cid:0)(t + a1v)2M[i](−2, −2) + (t + a2v)2M[i](−2, −2)(cid:1)m . 26 (⋆) γ3 γ5 γ6 γ2 γ7 γ9 γ10 γ11 γ12 γ0 γ4 γ8 γ1 Figure 5: The T-mesh used in Example 7.3 to explicitly demonstrate how results from Section 3 can be used to find the dimension of S r ∆m,m. (a) Let ∆m(σ) = (1, 1) for all faces σ ∈ T except for the face bounded by vertices γ4, γ8, γ12, γ11; on the latter face the degree deficit is chosen to be (0, 0). Let us choose the associated degree-deficit sequence as ∆m0 = (0, 0) < ∆m1 = (1, 1) so that l = 1. We will choose m = (3, 3). The following results follow on the different active meshes T[i], 1 ≤ i ≤ l + 1 = 2. • [i = 2]: From Proposition 4.5, dim(cid:0)H0(cid:0)C[i](cid:1)(cid:1)m tion 5.7, there are no interior maximal segments therefore dim(cid:0)H0(cid:0)I[i](cid:1)(cid:1)m = 0. • [i = 1]: From Proposition 4.5, dim(cid:0)H0(cid:0)C[i](cid:1)(cid:1)m = 0. Furthermore, there is a single active interior maximal segment ρ = γ11γ12, and the non-interior maximal segments ρ1 = γ4γ11 and ρ2 = γ8γ12 intersect it. Therefore, ρ1, ρ2 ∈ Γ[i](ρ) and from Proposition 6.8 and Equation (⋆) above, we get = 0. Furthermore, from Proposi- dim(cid:0)H0(cid:0)I[i](cid:1)(cid:1)m ≤ dim(cid:0)M[i](0, −2)(cid:1)m−dim  Xρ′∈{ρ1,ρ2} ∆ρ′M[i](−2, −2) m = 0 . From the above we can see that Theorem 7.2 applies. Thus, the dimension of S r ∆m,m can be determined to be exactly 17. It can be observed using the Macaulay2 script [28, ex0a.m2] that m = (3, 3) is also the smallest degree for which we get an increase in the dimension because of the non-uniformity in polynomial degrees. (b) Let ∆m(σ) = (1, 1) for all faces σ ∈ T except for the face bounded by vertices γ6, γ9, γ12, γ11; on the latter face the degree deficit is chosen to be (0, 0). Let us choose the associated degree-deficit sequence as ∆m0 = (0, 0) < ∆m1 = (1, 1) so that l = 1. We will choose m = (4, 4). The following results follow on the different active meshes T[i], 1 ≤ i ≤ l + 1 = 2. • [i = 2]: From Proposition 4.5, dim(cid:0)H0(cid:0)C[i](cid:1)(cid:1)m = 0. Furthermore, from Proposi- tion 5.7, there are no interior maximal segments therefore dim(cid:0)H0(cid:0)I[i](cid:1)(cid:1)m = 0. • [i = 1]: From Proposition 4.5, dim(cid:0)H0(cid:0)C[i](cid:1)(cid:1)m = 9. In this case, there are 4 active maximal segments ρ1 = γ11γ12, ρ2 = γ12γ9, ρ3 = γ9γ6, ρ1 = γ6γ11. Let us order these maximal segments as ρ3 > ρ2 > ρ4 > ρ1. Therefore, Γ[i](ρ1) = ∅ , Γ[i](ρ2) = Γ[i](ρ4) = {ρ1} , Γ[i](ρ3) = {ρ2, ρ4} . Once again, from Proposition 6.8 and Equation (⋆), there is no contribution to dim(cid:0)H0(cid:0)I[i](cid:1)(cid:1)m from ρ3. Furthermore, from Lemma 6.9 and Corollary 6.10, 27 we can also verify that ∆pρ3 = (1, 0) and Υ[i](ρ3, m) = {{(ρ1, ρ2), (ρ1, ρ4)}}. Therefore, from Lemma 6.11, we can state that, [ρ2]∆ρ3 M[i](−2, −2)m ⊂ In D[i] . However, since ρ1 ∈ Γ[i](ρ2), we also have the containment [ρ2]∆ρ1 M[i](−2, −2)m ⊂ In D[i] . Therefore, again from Equation (⋆), there is no contribution to dim(cid:0)H0(cid:0)I[i](cid:1)(cid:1)m from ρ2. Thus, the only contributions to the upper bound in Proposition 6.8 come from ρ1 and ρ4, dim(cid:0)H0(cid:0)I[i](cid:1)(cid:1)m Thus, dim(cid:0)H0(cid:0)I[i](cid:1)(cid:1)m ≤ dim(cid:0)[ρ1]M[i](0, −2)(cid:1)m = 7 + 7 − 5 = 9 = dim(cid:0)H0(cid:0)C[i](cid:1)(cid:1)m = dim(cid:0)H0(cid:0)C[i](cid:1)(cid:1)m − dim(cid:0)[ρ4]∆ρ1 . + dim(cid:0)[ρ4]M[i](−2, 0)(cid:1)m M[i](−2, −2)(cid:1)m . From the above we can see that Theorem 7.2 applies. Thus, the dimension of S r ∆m,m can be found to be 41. As in part (a), it can be observed using the Macaulay2 script [28, ex0b.m2] that m = (4, 4) is again the smallest degree for which we get an increase in the dimension because of the non-uniformity in polynomial degrees. Following Equation (6.6), let us define the set Λ[i](ρ, m) to be Λ[i](ρ, m) :=(cid:26)ρ′ : ρ′ ∈ Γ[i](ρ) , or (·, ρ) ∈ Θ[i](ρ′, m) , or (ρ, ρ′′) ∈ Θ[i](ρ′, m) and αρρ′ρ′′ = 1(cid:27) . (7.1) Definition 7.4 (Maximal-segment weights). For ρ ∈ weight of ρ is denoted by ωρ [i],m, and it is defined to be ◦ MS[i] and m = (m, m′) ∈ Z2 ≥0, the ωρ [i],m :=  Xρ′∈Λ[i](ρ,m) Xρ′∈Λ[i](ρ,m) Lemma 7.5. For m = (m, m′) ∈ Z2 ≥0, (m − ∆m(i−1)1 − r(ρ′))+ , ρ ∈ MSh [i] , (m′ − ∆m(i−1)2 − r(ρ′))+ , ρ ∈ MSv [i] . ωρ [i],m ≥(m − ∆m(i−1)1 + 1 , ρ ∈ MSh m′ − ∆m(i−1)2 + 1 , ρ ∈ MSv [i] [i] ⇒ Dρ [i],m = M[i](−δ(ρ))m . Example 7.6. Let us revisit Example 7.3(b). For i = 2 there are no interior maximal segments; so, let us look at the case of i = 1. We have Γ[i](ρ3) = {ρ2, ρ4} , Γ[i](ρ2) = {ρ1} , Γ[i](ρ4) = {ρ1} , Θ[i](ρ3, m) = {(ρ2, ρ4)} . Thus, from Equation 7.1, we have Λ[i](ρ1, m) = ∅ , Λ[i](ρ2, m) = {ρ1, ρ3} , Λ[i](ρ3, m) = {ρ2, ρ4} , Λ[i](ρ4, m) = {ρ1} , and from Definition 7.4, for ρ ∈ {ρ2, ρ3} we obtain ωρ [i],m = 2 × (4 − 1 − 1) = 4 . Then, from Lemma 7.5, we see that ρ2 and ρ3 do not contribute to dim(cid:0)H0(cid:0)I[i](cid:1)(cid:1)m. 28 γ3 γ5 γ7 γ9 γ2 γ15 γ16 γ28 γ17 γ18 γ19 γ23 γ27 γ24 γ31 γ25 γ10 γ11 γ26 γ12 γ30 γ13 γ14 γ20 γ21 γ29 γ22 γ0 γ4 γ6 γ8 γ1 Figure 6: A non-uniform degree C1 spline space consisting of quadratic and cubic polynomial pieces is built on the above mesh. Example 8.1 shows that the dimension of the spline space can be computed using Theorem 7.2. 8. Examples In this section, we provide three examples that illustrate how the theory developed in this document can be used to compute the spline space dimension in the presence of non-uniform degrees. The first two examples show configurations where Theorem 7.2 applies, i.e., where the dimension can be computed exactly. The last example serves to counter the expectation that Theorem 7.2 will apply in all circumstances. Example 8.1. Consider the T-mesh shown in Figure 6. Let us build a C 1 spline space on this mesh, i.e., r(τ ) = 1 for all interior edges τ . The degree deficit on the shaded faces is chosen to be (0, 0) and on the white faces it is chosen to be (1, 1). We choose ∆m0 = (0, 0) and ∆m1 = (1, 1), i.e., l = 1, and choose m = (3, 3). Let us examine the active T-meshes T[i] in the following for i = 1, 2. • [i = 2]: The only interior maximal segments in T[i] are ρ1 = γ20γ22 , ρ2 = γ23γ25 , ρ3 = γ26γ28 , ρ4 = γ29γ31 . Let us order these interior maximal segments as ρ1 < ρ2 < ρ3 < ρ4. Then notice that the cardinality of Γ[i](ρj) is 3 for each ρj. Therefore, we can compute the weight of each ρj [i],m ≥ 3 for all j. Then, using Lemma 7.5, we see interior maximal segment to be ω that dim(cid:0)H0(cid:0)I[i](cid:1)(cid:1)m = dim(cid:0)H0(cid:0)C[i](cid:1)(cid:1)m = 0 . • [i = 1]: All maximal segments in T[i] are interior maximal segments, let us denote them as below, ρ1 = γ11γ13 , ρ2 = γ22γ18 , ρ3 = γ29γ31 , ρ4 = γ12γ17 , ρ5 = γ26γ28 , ρ6 = γ11γ23 , ρ7 = γ29γ22 , ρ8 = γ23γ25 , ρ9 = γ28γ18 . 29 Let us order these interior maximal segments so that ρj < ρk when j < k. The sets Γ[i](ρj) can be seen to be Γ[i](ρ1) = ∅ , Γ[i](ρ2) = Γ[i](ρ3) = Γ[i](ρ4) = Γ[i](ρ5) = Γ[i](ρ6) = {ρ1} , Γ[i](ρ7) = {ρ2, ρ3} , Γ[i](ρ8) = {ρ2, ρ3, ρ4, ρ5, ρ6} , Γ[i](ρ9) = {ρ2, ρ4, ρ5} . Next, it can be seen from Corollary 6.10 that Υ[i](ρ8, m) ∋ Υ := {(ρ1, ρ2), (ρ1, ρ3), (ρ1, ρ4), (ρ1, ρ5), (ρ1, ρ6)} . Then, from Lemma 6.11, we can build the sets Λ[i](ρj) for j = 3, . . . , 6 such that {ρ1, ρ8} ⊆ Λ[i](ρ3) , Λ[i](ρ4) , Λ[i](ρ5) , Λ[i](ρ6) . Then, we see that the weight ωρj [i],m ≥ 4 for all j = 3, . . . , 9, and thus the upper bound dim(cid:0)H0(cid:0)I[i](cid:1)(cid:1)m ≤ dim(cid:0)[ρ1]M[i](0, −2)(cid:1)m + dim(cid:0)[ρ2]M[i](−2, 0)(cid:1)m on the dimension of H0(cid:0)I[i](cid:1) can be computed to be the following, M[i](−2, −2)(cid:1)m = 5 + 5 − 3 = 7 = dim(cid:0)H0(cid:0)C[i](cid:1)(cid:1)m . − dim(cid:0)[ρ2]∆ρ1 Therefore, dim(cid:0)H0(cid:0)I[i](cid:1)(cid:1)m = dim(cid:0)H0(cid:0)C[i](cid:1)(cid:1)m. From the above we can see that Theorem 7.2 applies and dim (S r ∆m)m = χ (Q)m = 37 . In particular, there are 30 splines on T[2] and 7 on T[1]. The reader can use the accompanying Macaulay2 script [28, ex1.m2] to confirm that m = (3, 3) is the smallest bi-degree for which non-uniformity in degrees leads to an increase in the dimension. Example 8.2. Consider the T-mesh shown in Figure 6. Let us build a C 2 spline space on this mesh, i.e., r(τ ) = 2 for all interior edges τ . The degree deficit on the shaded faces is chosen to be (0, 0) and on the white faces it is chosen to be (1, 1). We choose ∆m0 = (0, 0) and ∆m1 = (1, 1), i.e., l = 1, and choose m = (4, 4). Let us examine the active T-meshes T[i] in the following for i = 1, 2. • [i = 2]: The only interior maximal segments in T[i] are ρ1 = γ35γ39 , ρ2 = γ40γ44 , ρ3 = γ45γ49 . We can order them in any manner with respect to each other since they don't intersect each other. The weight of each interior maximal segment can be computed to be ωρj [i],m = 5 for all j. Then, using Lemma 7.5, we see that dim(cid:0)H0(cid:0)I[i](cid:1)(cid:1)m = dim(cid:0)H0(cid:0)C[i](cid:1)(cid:1)m = 0 . • [i = 1]: All maximal segments in T[i] are interior maximal segments, let us denote them as below, ρ1 = γ12γ32 , ρ2 = γ36γ18 , ρ3 = γ12γ13 , ρ4 = γ37γ23 , ρ5 = γ38γ28 , ρ6 = γ39γ44 , ρ7 = γ36γ39 , ρ8 = γ40γ44 , ρ9 = γ13γ28 . 30 γ3 γ5 γ7 γ9 γ2 γ30 γ25 γ20 γ15 γ10 γ31 γ39 γ32 γ44 γ33 γ49 γ34 γ26 γ38 γ27 γ43 γ28 γ48 γ29 γ21 γ37 γ22 γ42 γ23 γ47 γ24 γ16 γ36 γ17 γ41 γ18 γ46 γ19 γ11 γ35 γ12 γ40 γ13 γ45 γ14 γ0 γ4 γ6 γ8 γ1 Figure 7: A non-uniform degree C2 spline space consisting of cubic and quartic polynomial pieces is built on the above mesh. Example 8.2 shows that the dimension of the spline space can be computed using Theorem 7.2. Let us order these interior maximal segments so that ρj < ρk when j < k. The sets Γ[i](ρj) can be seen to be Γ[i](ρ1) = ∅ , Γ[i](ρ2) = Γ[i](ρ3) = Γ[i](ρ4) = Γ[i](ρ5) = Γ[i](ρ6) = {ρ1} , Γ[i](ρ7) = {ρ2, ρ4, ρ5, ρ6} , Γ[i](ρ8) = {ρ2, ρ3, ρ4, ρ5, ρ6} , Γ[i](ρ9) = {ρ2, ρ3, ρ4, ρ5} . Next, it can be seen from Corollary 6.10 that Υ[i](ρ7, m) ∋ {(ρ1, ρ2), (ρ1, ρ4), (ρ1, ρ5), (ρ1, ρ6)} , Υ[i](ρ8, m) ∋ {(ρ1, ρ2), (ρ1, ρ3), (ρ1, ρ4), (ρ1, ρ5), (ρ1, ρ6)} , Υ[i](ρ9, m) ∋ {(ρ1, ρ2), (ρ1, ρ3), (ρ1, ρ4), (ρ1, ρ5)} . Then, from Lemma 6.11, we can build the sets Λ[i](ρj) for j = 3, . . . , 6 such that {ρ1, ρ8, ρ9} ⊆ Λ[i](ρ3) , {ρ1, ρ7, ρ8} ⊆ Λ[i](ρ4) , Λ[i](ρ5) , Λ[i](ρ6) . Then, we see that the weight ωρj [i],m > 5 for all j = 3, . . . , 9, and thus the upper bound dim(cid:0)H0(cid:0)I[i](cid:1)(cid:1)m on the dimension of H0(cid:0)I[i](cid:1) can be computed to be the following, + dim(cid:0)[ρ2]M[i](−3, 0)(cid:1)m M[i](−3, −3)(cid:1)m ≤ dim(cid:0)[ρ1]M[i](0, −3)(cid:1)m − dim(cid:0)[ρ2]∆ρ1 = 6 + 6 − 3 = 9 = dim(cid:0)H0(cid:0)C[i](cid:1)(cid:1)m Therefore, dim(cid:0)H0(cid:0)I[i](cid:1)(cid:1)m = dim(cid:0)H0(cid:0)C[i](cid:1)(cid:1)m. . From the above we can see that Theorem 7.2 applies and dim (S r ∆m)m = χ (Q)m = 75 . In particular, there are 66 splines on T[2] and 9 on T[1]. The reader can use the accompanying Macaulay2 script [28, ex2.m2] can be used to confirm that m = (4, 4) is the smallest bi-degree for which non-uniformity in degrees leads to an increase in the dimension. 31 γ3 γ5 γ7 γ2 γ12 γ8 γ0 γ13 γ20 γ18 γ9 γ17 γ21 γ19 γ16 γ14 γ15 γ10 γ11 γ4 γ6 γ1 Figure 8: A non-uniform degree C2 spline space consisting of cubic and quartic polynomial pieces is built on the above mesh. Example 8.3 shows that the dimension of the spline space coincides with the upper bound implied by Theorem 7.2. Example 8.3. Consider the T-mesh shown in Figure 8. Let us build a C 2 spline space on this mesh, i.e., r(τ ) = 2 for all interior edges τ . The degree deficit on the shaded faces is chosen to be (0, 0) and on the white faces it is chosen to be (1, 1). We choose ∆m0 = (0, 0) and ∆m1 = (1, 1), i.e., l = 1, and choose m = (6, 6). Let us examine the active T-meshes T[i] in the following for i = 1, 2. • [i = 2]: The only interior maximal segments in T[i] are ρ1 = γ16γ17 , ρ2 = γ20γ21 , ρ3 = γ18γ19 . Let us order them as ρ1 < ρ2 < ρ3. Then, for m = (6, 6), it can be computed that ωρj [i],m ≥ 6. Then, using Lemma 7.5, we see that dim(cid:0)H0(cid:0)I[i](cid:1)(cid:1)m = dim(cid:0)H0(cid:0)C[i](cid:1)(cid:1)m = 0 . Note that m = (6, 6) is the smallest bi-degree in which H0(cid:0)I[i](cid:1) vanishes. • [i = 1]: All maximal segments in T[i] are interior maximal segments, let us denote them as below, ρ1 = γ16γ17 , ρ2 = γ16γ10 , ρ3 = γ17γ14 , ρ4 = γ10γ14 , ρ5 = γ18γ19 , ρ6 = γ20γ21 , ρ7 = γ18γ20 . Let us order these interior maximal segments so that ρj < ρk when j < k. The sets Γ[i](ρj) can be seen to be Γ[i](ρ1) = ∅ , Γ[i](ρ2) = Γ[i](ρ3) = Γ[i](ρ5) = Γ[i](ρ6) = {ρ1} , Γ[i](ρ4) = {ρ2, ρ3} , Γ[i](ρ7) = {ρ5, ρ6} . Next, it can be seen from Corollary 6.10 that Υ[i](ρ4, m) ∋ {(ρ1, ρ2), (ρ1, ρ3)} , Υ[i](ρ7, m) ∋ {(ρ1, ρ5), (ρ1, ρ6)} . 32 Then, from Lemma 6.11, we can build the sets Λ[i](ρ3) and Λ[i](ρ6) such that Λ[i](ρ3) = {ρ1, ρ4} , Λ[i](ρ6) = {ρ1, ρ7} . Then, we see that the weight ωρj [i],m ≥ 8 for j = 3, 4, 6, 7. Thus the upper bound on the dimension of H0(cid:0)I[i](cid:1) can be computed to be the following, dim(cid:0)H0(cid:0)I[i](cid:1)(cid:1)m ≤ dim(cid:0)[ρ1]M[i](0, −3)(cid:1)m + dim(cid:0)[ρ2]M[i](−3, 0)(cid:1)m + dim(cid:0)[ρ5]M[i](0, −3)(cid:1)m − dim(cid:0)[ρ2]∆ρ1 − dim(cid:0)[ρ5]∆ρ1 M[i](−3, −3)(cid:1)m = 10 + 10 + 10 − 7 − 7 = 16 . M[i](−3, −3)(cid:1)m On the other hand, using Proposition 4.5, we can compute that dim(cid:0)H0(cid:0)C[i](cid:1)(cid:1)m = 13 . Therefore, dim(cid:0)H0(cid:0)I[i](cid:1)(cid:1)m − dim(cid:0)H0(cid:0)C[i](cid:1)(cid:1)m ≤ 3. From the above we see that Theorem 7.2 does not apply. Therefore, let us use Theorems 6.4 and 6.5 to bound the spline space dimension from below and above. Computing the Euler characteristic of Q to be χ (Q)m = 143, we use those theorems to obtain the following, 143 ≤ dim (S r ∆m)m ≤ 146 . The reader can use the accompanying Macaulay2 script [28, ex3.m2] to confirm that the dimension of the spline space is exactly 146 for this configuration and thus coincides with the computed upper bound. This example serves to show that there exist configurations where a maximal segment ordering that allows us to use Theorem 7.2 does not exist. Indeed, the accompanying script can be used to verify that dim(cid:0)H0(cid:0)I[i](cid:1)(cid:1)m − dim(cid:0)H0(cid:0)C[i](cid:1)(cid:1)m = 3 for all bi-degrees greater than or equal to (6, 6). 9. Conclusions Splines have been used for geometric modeling for several decades, and they are now rapidly becoming indispensable tools for performing approximation. In order to efficiently alter the local resolution offered by a spline space, it is important to be able to perform local adaptivity. While local mesh adaptivity on quadrilateral meshes has been widely studied since the introduction of T-splines [25], theoretical studies focused on splines that allow local degree adaptivity have been missing heretofore from the bivariate spline literature. Since the possibility of using non-uniform bi-degree splines on T-meshes would enable powerful new paradigms of local refinement, we take a first step in this direction by analyzing the the dimension of such spline spaces. In particular, using tools from homological algebra, we provide combinatorial bounds on the dimension. We also outline sufficient conditions that guarantee that the spline space dimension is stable, i.e., the dimension of the space is independent of the geometry of the T-mesh for a fixed topology. Several examples are provided to show applicability of the theory developed here. The results presented in this paper can be used for classifying spline spaces with stable dimension. This is important for avoiding geometry-based linear dependency issues that may arise when performing approximation. The ability to combinatorially compute the spline space dimension is also important because it can be used to determine when a given set of linearly independent splines spans the full spline space. Conversely, given a constructive approach that aims to produce linearly independent splines over T-meshes using only local 33 data, computation of the associated spline space dimension can help identify cases where the splines produced by the approach cannot be linearly independent. This is crucial for devising constructive approaches that can be robustly employed for performing isogeometric analysis. Current research on this topic is progressing along several lines of inquiry. A follow up paper will discuss local refinement algorithms for both mesh sizes and polynomial degrees that ensure stability of the spline space dimension. The construction of a suitable basis that possesses B-spline-like properties remains an open and essential question and will be a part of future efforts focused on formulation of constructive approaches. A practical construction of this nature has been successfully devised for univariate non-uniform degree splines [30, 26, 29]. A generalization of this univariate approach to the bivariate setting has been recently conjectured [27]. References References [1] Y. Bazilevs, V.M. Calo, J.A. Cottrell, J.A. Evans, T.J.R. Hughes, S. Lipton, M.A. Scott, and T.W. Seder- berg. Isogeometric analysis using T-splines. Computer Methods in Applied Mechanics and Engineering, 199(5-8):229 -- 263, 2010. [2] Louis J Billera. Homology of smooth splines: generic triangulations and a conjecture of strang. Trans- actions of the american Mathematical Society, 310(1):325 -- 340, 1988. [3] Louis J. Billera and Lauren L. Rose. A dimension series for multivariate splines. 6(2):107 -- 128. [4] Cesare Bracco, Tom Lyche, Carla Manni, Fabio Roman, and Hendrik Speleers. On the dimension of Tchebycheffian spline spaces over planar T-meshes. Computer Aided Geometric Design, 45:151 -- 173, 2016. [5] Cesare Bracco, Tom Lyche, Carla Manni, and Hendrik Speleers. Tchebycheffian spline spaces over pla- nar T-meshes: Dimension bounds and dimension instabilities. Journal of Computational and Applied Mathematics, 349:265 -- 278, 2019. [6] Andrea Bressan. Some properties of LR-splines. Computer Aided Geometric Design, 30(8):778 -- 794, 2013. [7] Andrea Bressan, Annalisa Buffa, and Giancarlo Sangalli. Characterization of analysis-suitable T-splines. Computer Aided Geometric Design, 39:17 -- 49, 2015. [8] A. Buffa, D. Cho, and G. Sangalli. Linear independence of the T-spline blending functions associated with some particular T-meshes. Computer Methods in Applied Mechanics and Engineering, 199(23-24):1437 -- 1445, 2010. [9] J. Austin Cottrell, Thomas J. R. Hughes, and Yuri Bazilevs. Isogeometric Analysis: Toward Integration of CAD and FEA. Wiley, Chichester, West Sussex, U.K. ; Hoboken, NJ, 1 edition edition, 2009. [10] Jiansong Deng, Falai Chen, Xin Li, Changqi Hu, Weihua Tong, Zhouwang Yang, and Yuyu Feng. Poly- nomial splines over hierarchical T-meshes. Graphical Models, 70(4):76 -- 86, 2008. [11] T. Dokken, T. Lyche, and K.F. Pettersen. Polynomial splines over locally refined box-partitions. Com- puter Aided Geometric Design, 30:331 -- 356, 2013. [12] Tor Dokken, Tom Lyche, and Kjell Fredrik Pettersen. Polynomial splines over locally refined box- partitions. Computer Aided Geometric Design, 30(3):331 -- 356, 2013. [13] David R. Forsey and Richard H. Bartels. Hierarchical B-spline refinement. In ACM SIGGRAPH Computer Graphics, volume 22, pages 205 -- 212. ACM, 1988. [14] Carlotta Giannelli, Bert Juttler, and Hendrik Speleers. THB-splines: The truncated basis for hierarchical splines. Computer Aided Geometric Design, 29(7):485 -- 498, 2012. [15] Daniel R. Grayson and Michael E. Stillman. Macaulay2, a software system for research in algebraic geometry. Available at https://faculty.math.illinois.edu/Macaulay2/. [16] Kjetil Andr´e Johannessen, Trond Kvamsdal, and Tor Dokken. Isogeometric analysis using LR B-splines. Computer Methods in Applied Mechanics and Engineering, 269:471 -- 514, 2014. [17] R. Kraft. Adaptive and linearly independent multilevel b-splines. In A.L. M´ehaut´e, C. Rabut, and L.L. Schumaker, editors, Surface Fitting and Multiresolution Methods, pages 209 -- 218. Vanderbilt University Press, 1997. [18] Xin Li and Falai Chen. On the instability in the dimension of splines spaces over t-meshes. 28(7):420 -- 426. [19] Xin Li, Jianmin Zheng, Thomas W. Sederberg, Thomas J. R. Hughes, and Michael A. Scott. On linear independence of t-spline blending functions. 29(1):63 -- 76. [20] Bernard Mourrain. On the dimension of spline spaces on planar t-meshes. Mathematics of Computation, 83(286):847 -- 871, 2014. [21] Hal Schenck. Computational algebraic geometry, volume 58. Cambridge University Press, 2003. 34 [22] Larry L. Schumaker and Lujun Wang. Approximation power of polynomial splines on T-meshes. Computer Aided Geometric Design, 29(8):599 -- 612, 2012. [23] M.A. Scott, X. Li, T.W. Sederberg, and T.J.R. Hughes. Local refinement of analysis-suitable T-splines. Computer Methods in Applied Mechanics and Engineering, 213-216:206 -- 222, 2012. [24] Thomas W. Sederberg, Jianmin Zheng, Almaz Bakenov, and Ahmad Nasri. T-splines and T-NURCCs. In ACM SIGGRAPH 2003 Papers, SIGGRAPH '03, pages 477 -- 484, New York, NY, USA, 2003. ACM. [25] Thomas W Sederberg, Jianmin Zheng, Almaz Bakenov, and Ahmad Nasri. T-splines and t-nurccs. In ACM transactions on graphics (TOG), volume 22, pages 477 -- 484. ACM, 2003. [26] H. Speleers. Computation of multi-degree B-splines. Preprint, arXiv:1809.01598, 2018. [27] D. Thomas, L. Engvall, S. Schmidt, K. Tew, and M. Scott. U-splines: Splines over unstructured meshes. Preprint, https://coreform.com/usplines, 2018. [28] Deepesh Toshniwal. Supplementary Macaulay2 scripts for splines on T-meshes. Available at http://users.ices.utexas.edu/~deepesh/tmesh_splines.html. [29] Deepesh Toshniwal, Hendrik Speleers, Rene Hiemstra, Carla Manni, and Thomas J. R. Hughes. ICES Report 18-22, Multi-degree B-splines: Algorithmic computation and properties. Preprint, https://www.ices.utexas.edu/media/reports/2018/1822.pdf, 2018. [30] Deepesh Toshniwal, Hendrik Speleers, Ren´e R Hiemstra, and Thomas JR Hughes. Multi-degree smooth polar splines: A framework for geometric modeling and isogeometric analysis. Computer Methods in Applied Mechanics and Engineering, 316:1005 -- 1061, 2017. 35
1505.01890	2	1505	2015-07-10T21:38:08	Convergence polygons for connections on nonarchimedean curves	[ "math.AG", "math.NT" ]	This is a survey article on ordinary differential equations over nonarchimedean fields based on the author's lecture at the 2015 Simons Symposium on nonarchimedean and tropical geometry. Topics include: the convergence polygon associated to a differential equation (or a connection on a curve); links to the formal classification of differential equations (Turrittin-Levelt); index formulas for de Rham cohomology of connections; ramification of finite morphisms; relations with the Oort lifting problem on automorphisms of curves. The appendices include some new technical results and an extensive thematic bibliography.	math.AG	math	CONVERGENCE POLYGONS FOR CONNECTIONS ON NONARCHIMEDEAN CURVES KIRAN S. KEDLAYA In classical analysis, one builds the catalog of special functions by repeatedly adjoining solutions of differential equations whose coefficients are previously known functions. Conse- quently, the properties of special functions depend crucially on the basic properties of ordi- nary differential equations. This naturally led to the study of formal differential equations, as in the seminal work of Turrittin [167]; this may be viewed retroactively as a theory of differ- ential equations over a trivially valued field. After the introduction of p-adic analysis in the early 20th century, there began to be corresponding interest in solutions of p-adic differential equations; however, aside from some isolated instances (e.g., the proof of the Nagell-Lutz theorem; see Theorem 3.4), a unified theory of p-adic ordinary differential equations did not emerge until the pioneering work of Dwork on the relationship between p-adic special functions and the zeta functions of algebraic varieties over finite fields (e.g., see [57, 58]). At that point, serious attention began to be devoted to a serious discrepancy between the p-adic and complex-analytic theories: on an open p-adic disc, a nonsingular differential equation can have a formal solution which does not converge in the entire disc (e.g., the exponential series). One is thus led to quantify the convergence of power series solutions of differential equations involving rational functions over a nonarchimedean field; this was originally done by Dwork in terms of the generic radius of convergence [59]. This and more refined invariants were studied by numerous authors during the half-century following Dwork’s initial work, as documented in the author’s book [93]. At around the time that [93] was published, a new perspective was introduced by Baldas- sarri [13] (and partly anticipated in prior unpublished work of Baldassarri and Di Vizio [15]) which makes full use of Berkovich’s theory of nonarchimedean analytic spaces. Given a dif- ferential equation as above, or more generally a connection on a curve over a nonarchimedean field, one can define an invariant called the convergence polygon; this is a function from the underlying Berkovich topological space of the curve into a space of Newton polygons, which measures the convergence of formal horizontal sections and is well-behaved with respect to both the topology and the piecewise linear structure on the Berkovich space. One can translate much of the prior theory of p-adic differential equations into (deceptively) simple statements about the behavior of the convergence polygon; this process was carried out in a series of papers by Poineau and Pulita [141, 133, 135, 136], as supplemented by work of this author [98] and upcoming joint work with Baldassarri [16]. Date: July 10, 2015. Thanks to Francesco Baldassarri, Andrew Obus, Andrea Pulita, Junecue Suh, Daniele Turchetti, the anonymous referee, and the participants of the Simons Symposium for detailed feedback. The author was supported by NSF grant DMS-1101343 and UC San Diego (Warschawski Professorship), and by a Simons Visiting Professorship grant during November 2014 (visiting Pulita). 1 In this paper, we present the basic theorems on the convergence polygon, which provide a number of combinatorial constraints that may be used to extract information about con- vergence of formal horizontal sections at one point from corresponding information at other points. We include numerous examples to illustrate some typical behaviors of the conver- gence polygon. We also indicate some relationships between convergence polygons, index formulas for de Rham cohomology, and the geometry of finite morphisms, paying special attention to the case of cyclic p-power coverings with p equal to the residual characteristic. This case is closely linked with the Oort lifting problem for Galois covers of curves in char- acteristic p, and some combinatorial constructions arising in that problem turn out to be closely related to convergence polygons. There are additional applications to the study of integrable connections on higher-dimensional nonarchimedean analytic spaces, both in the cases of zero residual characteristic [95] and positive residual characteristic [96], but we do not pursue these applications here. To streamline the exposition, we make no attempt to indicate the techniques of proof in some cases, quite sophisticated arguments are required. underlying our main results; We limit ourselves to saying that the basic tools are developed in a self-contained fashion in [93] (but without reference to Berkovich spaces), and the other aforementioned results are obtained by combining results from [93] in an intricate manner. (Some exceptions are made for results which do not occur in any existing paper; their proofs are relegated to an appendix.) We also restrict generality by considering only proper curves, even though many of the results we discuss can be formulated for open curves, possibly of infinite genus. At the suggestion of the referee, we include (as Appendix B) a thematic bibliography in the style of [147] listing additional references germane to the topics discussed in the article. We also include references for the following topics which we have omitted out of space considerations, even though in practice they are thoroughly entangled with the proofs of the results which we do mention. • Decomposition theorems, which give splittings of connections analogous to the fac- torizations of polynomials given by various forms of Hensel’s lemma. • Monodromy theorems, which provide structure theorems for connections at the ex- pense of passing from a given space to a finite cover. • Logarithmic growth, i.e., secondary terms in the measurement of convergence of hor- izontal sections. 1. Newton polygons As setup for our definition of convergence polygons, we fix some conventions regarding Newton polygons. Definition 1.1. For n a positive integer, let P[0, n] be the set of continuous functions N : [0, n] → R satisfying the following conditions. (a) We have N (0) = 0. (b) For i = 1, . . . , n, the restriction of N to [i − 1, i] is affine. For i = 1, . . . , n, we write hi : P[0, n] → R for the function N 7→ N (i); we call hi(N ) the i-th height of N . The product map h1 × · · · × hn : P[0, n] → Rn is a bijection, using which we equip P[0, n] with a topology and an integral piecewise linear structure. We sometimes refer to hn simply as h and call it the total height. 2 Definition 1.2. Let N P[0, n] be the subset of P[0, n] consisting of concave functions. For i = 1, . . . , n, we write si : P[0, n] → R for the function N 7→ N (i) − N (i − 1); we call si(N ) the i-th slope of N . For N ∈ N P[0, n], we have s1(N ) ≥ · · · ≥ sn(N ). Definition 1.3. Let I ⊆ R be a closed interval. A function N : I → N P[0, n] is affine if it has the form N (t) = N0 + tN1 for some N0, N1 ∈ P[0, n]. In this case, we call N1 the slope of N . We say that N has integral derivative if N1(i) ∈ Z for i = n and for each i ∈ {1, . . . , n − 1} such that for all t in the interior of I, N (t) has a change of slope at i. This implies that the graph of N1 has vertices only at lattice points, but not conversely. (It would be natural to use the terminology integral slope, but we avoid this terminology to alleviate confusion with Definition 1.2.) 2. PL structures on Berkovich curves We next recall the canonical piecewise linear structure on a Berkovich curve (e.g., see [8]). Hypothesis 2.1. For the rest of this paper, let K be a nonarchimedean field, i.e., a field complete with respect to a nonarchimedean absolute value; let X be a smooth, proper, geometrically connected curve over K; let Z be a finite set of closed points of X; and put U = X − Z. Convention 2.2. Whenever we view Qp as a nonarchimedean field, we normalize the p-adic absolute value so that p = p−1. Remark 2.3. Recall that the points of the Berkovich analytification X an may be identified with equivalence classes of pairs (L, x) in which L is a nonarchimedean field over K and x is an element of X(L), where the equivalence relation is generated by relations of the form (L, x) ∼ (L′, x′) where x′ is the restriction of x along a continuous K-algebra homomorphism L → L′. As is customary, we classify points of X an into types 1,2,3,4 (e.g., see [98, Proposition 4.2.7]). To lighten notation, we identify Z with Z an, which is a finite subset of X an consisting of type 1 points. Definition 2.4. For ρ > 0, let xρ denote the generic point of the disc z ≤ ρ in P1 K. A segment in X an is a closed subspace S homeomorphic to a closed interval for which there exist an open subspace V of X an, a choice of values 0 ≤ α < β ≤ +∞, and an isomorphism of V with {z ∈ P1,an K : α < z < β} identifying the interior of S with {xρ : ρ ∈ (α, β)}. A virtual segment in X an is a connected closed subspace whose base extension to some finite extension of K is a disjoint union of segments. A strict skeleton in X an is a subspace Γ containing Z an equipped with a homeomorphism to a finite connected graph, such that each vertex of the graph corresponds to either a point of Z or a point of type 2, and each edge corresponds to a virtual segment, and X an retracts continuously onto Γ. Using either tropicalizations or semistable models, one may realize X an as the inverse limit of its strict skeleta; again, see [8] for a detailed discussion. Definition 2.5. Note that χ(U) = 2 − 2g(X) − length(Z), so χ(U) ≤ 0 if and only if either g(X) ≥ 1 or length(Z) ≥ 2. In this case, there is a unique minimal strict skeleton in X an, which we denote ΓX,Z. Explicitly, if K is algebraically closed, then the underlying set of ΓX,Z is the complement in X an of the union of all open discs in U an; 3 for general K, the underlying set of ΓX,Z is the image under restriction of the minimal strict L for L a completed algebraic closure of K. In particular, if K ′ is the completion skelelon in X an of an algebraic extension of K, then the minimal strict skeleton in X an K ′ is the inverse image of ΓX,Z in X an K ′. However, the corresponding statement for a general nonarchimedean field extension K ′ of K is false; see Definition 8.5 for a related phenomenon. 3. Convergence polygons: projective line We next introduce the concept of the convergence polygon associated to a differential equation on P1. Hypothesis 3.1. For the rest of this paper, we assume that the nonarchimedean field K is of characteristic 0, as otherwise the study of differential operators on K-algebras has a markedly different flavor (for instance, any derivation on a ring R of characteristic p > 0 has the subring of p-th powers in its kernel). By contrast, the residue characteristic of K, which we call p, may be either 0 or positive unless otherwise specified (e.g., if we refer to Qp then we implicitly require p > 0). Hypothesis 3.2. For the rest of §3, take X = P1 differential equation K, assume ∞ ∈ Z, and consider the (3.2.1) y(n) + fn−1(z)y(n−1) + · · · + f0(z)y = 0 for some rational functions f0, . . . , fn−1 ∈ K(z) with poles only within Z. If Z = {∞}, let m be the dimension of the K-vector space of entire solutions of (3.2.1); otherwise, take m = 0. Definition 3.3. For any nonarchimedean field L over K and any x ∈ U(L), let Sx be the set of formal solutions of (3.2.1) with y ∈ LJz −xK. By interpreting (3.2.1) as a linear recurrence relation of order n on the coefficients of a power series, we see that every list of n initial conditions at z = x corresponds to a unique formal solution; that is, the composition is a bijection. In particular, Sx is an L-vector space of dimension n. Sx → LJz − xK → LJz − xK/(z − x)n Theorem 3.4 (p-adic Cauchy theorem). Each element of Sx has a positive radius of con- vergence. Proof. This result was originally proved by Lutz [107, Th´eor`eme IV] somewhat before the emergence of the general theory of p-adic differential equations; Lutz used it as a lemma in her proof of the Nagell-Lutz theorem on the integrality of torsion points on rational elliptic curves. One can give several independent proofs using the modern theory; see [93, Proposition 9.3.3, Proposition 18.1.1]. (cid:3) Definition 3.5. For i = 1, . . . , n − m, choose si(x) ∈ R so that e−si(x) is the supremum of the set of ρ > 0 such that U an contains the open disc z − x < ρ and Sx contains n − i + 1 linearly independent elements convergent on this disc. Note that this set is nonempty by Theorem 3.4 and bounded above by the definition of m, so the definition makes sense. In particular, s1(x) is the joint radius of convergence of all of the elements of Sx, while sn−m(x) is the maximum finite radius of convergence of a nonzero element of Sx. Since s1(x) ≥ · · · ≥ sn−m(x), the si(x) are the slopes of a polygon Nz(x) ∈ N P[0, n − m], which we call the convergence polygon of (3.2.1) at x. (We include z in the notation to 4 remind ourselves that Nz depends on the choice of the coordinate z of X.) This construction is compatible with base change: if L′ is a nonarchimedean field containing L and x′ is the image of x in U(L′), then Nz(x) = Nz(x′). Consequently, we obtain a well-defined function Nz : U an → N P[0, n − m]. Definition 3.6. Suppose that Z 6= {∞}. By definition, e−s1(Nz(x)) can never exceed the largest value of ρ for which the disc z − x < ρ does not meet Z. When equality occurs, we say that (3.2.1) satisfies the Robba condition at x. Theorem 3.7. The function Nz : U an → N P[0, n − m] is continuous; more precisely, it factors through the retraction of P1,an K onto some strict skeleton Γ, and the restriction of Nz to each edge of Γ is affine with integral derivative. Proof. See [141] or [98] or [16]. (cid:3) One can say quite a bit more, but for this it is easier to shift to a coordinate-free inter- pretation, which also works for more general curves; see §5. 4. A gallery of examples To help the reader develop some intuition, we collect a few illustrative examples of con- vergence polygons. Throughout §4, retain Hypothesis 3.1. Example 4.1. Take K = Qp, Z = {∞}, and consider the differential equation y′ − y = 0. The formal solutions of this equation with y ∈ LJz − xK are the scalar multiples of the exponential series which has radius of convergence p−1/(p−1). Consequently, exp(z − x) = (z − x)i i! , ∞Xi=0 in particular, Nz is constant on U an. s1(Nz(x)) = 1 p − 1 log p; In this next example, we illustrate the effect of changing Z on the convergence polygon. Example 4.2. Set notation as in Example 4.1, except now with Z = {0, ∞}. In this case we have s1(Nz(x)) = max(cid:26)− log x, 1 log p(cid:27) . (p − 1) K onto the path from 0 to ∞. For In particular, Nz factors through the retraction of P1,an x ∈ U an, the Robba condition holds at x if and only if x ≤ p−1/(p−1). Example 4.3. Take K = Qp, Z = {0, ∞}, and consider the differential equation y′− 1 pz−1y = 0. The formal solutions of this equation with y ∈ LJz − xK are the scalar multiples of the binomial series ∞Xi=0(cid:18)1/p i (cid:19)x−i+p−1 5 (z − x)i, which has radius of convergence p−p/(p−1)x. Consequently, s1(Nz(x)) = p p − 1 log p − log x, so again Nz factors through the retraction of P1,an the Robba condition holds nowhere. K onto the path from 0 to ∞. In this case, Example 4.4. Assume p > 2, take K = Qp, Z = {0, ∞}, and consider the Bessel differential equation (with parameter 0) This example was studied by Dwork [61], who showed that y′′ + z−1y′ + y = 0. s1(Nz(x)) = s2(Nz(x)) = max(cid:26)− log x, 1 p − 1 log p(cid:27) . Again, Nz factors through the retraction of P1,an ple 4.2, for x ∈ U an, the Robba condition holds at x if and only if x ≤ p−1/(p−1). K onto the path from 0 to ∞. As in Exam- Our next example illustrates a typical effect of varying a parameter. Example 4.5. Let K be an extension of Qp, take Z = {0, ∞}, and consider the differential equation y′ − λz−1y = 0 for some λ ∈ K (the case λ = 1/p being Example 4.3). Then s1(Nz(x)) = c + 1 p − 1 log p − log x where c is a continuous function of namely, by [63, Proposition IV.7.3] we have c0 = min{λ − t : t ∈ Zp}; log c0 − pm−1 − 1 (p−1)pm log p + 1 p−1 log p pm+1 log(pmc0) if c0 ≥ 1 if p−m−1 ≤ c0 ≤ p−m if c0 = 0. (m = 0, 1, . . . ) c = In particular, the Robba condition holds everywhere if λ ∈ Zp and nowhere otherwise. In either case, Nz factors through the retraction of P1,an K onto the path from 0 to ∞. Example 4.6. Take K = Qp, Z = {∞}, and consider the differential equation y′ −aza−1y = 0 for some positive integer a not divisible by p (the case a = 1 being Example 4.1). The formal solutions of this equation are the scalar multiples of exp(za − xa). This series converges in the region where za − xa < p−1/(p−1); consequently, s1(Nz(x)) = max(cid:26) 1 p − 1 log p + (a − 1) log x, 1 a(p − 1) log p(cid:27) . In this case, Nz factors through the retraction onto the path from xp−1/(p−1) to ∞. 6 Example 4.7. Take K = C((t)) (so p = 0), Z = {∞}, and consider the differential equation y′′′ + zy′′ + y = 0. It can be shown that s1(Nz(x)) = max {0, log x} , s2(Nz(x)) = s3(Nz(x)) = min(cid:26)0, − 1 2 log x(cid:27) . In this case, Nz factors through the retraction onto the path from x1 to ∞. Note that this provides an example where the slopes of Nz(x) are not bounded below uniformly on (P1 K − Z)an; that is, as x approaches ∞, two linearly independent local solutions have radii of convergence growing without bound, but these local solutions do not patch together. Example 4.8. Assume p > 2, take K = Qp, Z = {0, 1, ∞}, and consider the Gaussian hypergeometric differential equation y′′ + (1 − 2z) z(1 − z) y′ − 1 4z(1 − z) y = 0. This example was originally studied by Dwork [58] due to its relationship with the zeta functions of elliptic curves. Using Dwork’s calculations, it can be shown that s1(Nz(x)) = s2(Nz(x)) = max{log x, − log x, − log x − 1}. In this case, Nz factors through the retraction from P1,an 0 to ∞ and from 1 to ∞, and the Robba condition holds everywhere. K onto the union of the paths from Remark 4.9. One can compute additional examples of convergence polygons associated to first-order differential equations using an explicit formula for the radius of convergence at a point, due to Christol–Pulita. This result was originally reported in [36] but with an error in the formula; for a corrected statement, see [142, Introduction, Th´eor`eme 5]. 5. Convergence polygons: general curves We now describe an analogue of the convergence polygon in a more geometric setting. Hypothesis 5.1. Throughout §5, assume that χ(U) ≤ 0, i.e., either g(X) ≥ 1 or length(Z) ≥ 2. Let E be a vector bundle on U of rank n equipped with a connection ∇ : E → E ⊗OU ΩU/K. As is typical, we describe sections of E in the kernel of ∇ as being horizontal. Remark 5.2. For the results in this section, one could also allow X to be an analytic curve which is compact but not necessarily proper. To simplify the discussion, we omit this level of generality; the general results can be found in any of [16], [98], [133]. Definition 5.3. Let L be a nonarchimedean field containing K and choose x ∈ U(L), which we identify canonically with a point of U an L contains a neighborhood of x isomorphic to an open disc over L. Thanks to our restrictions on X and Z, the union Ux of all such neighborhoods in U an L is itself isomorphic to an open disc if L′ is a over L. The construction is compatible with base change in the following sense: nonarchimedean field containing L, then Ux,L′ is the union of all neighborhoods of x in U an L′ isomorphic to an open disc over L′. For each ρ ∈ (0, 1], let Ux,ρ be the open disc of radius ρ centered at x within Ux (normalized so that Ux,1 = Ux). L . Since X is smooth, U an 7 Let bOXL,x denote the completed local ring of XL at x; it is abstractly a power series ring in one variable over L. Let Ex denote the pullback of E to bOXL,x, equipped with the induced connection. One checks easily that Ex is a trivial differential module; more precisely, the space ker(∇, Ex) is an n-dimensional vector space over L and the natural map ker(∇, Ex) ⊗L bOXL,x → Ex is an isomorphism. For i = 1, . . . , n, choose si(x) ∈ [0, +∞) so that e−si(x) is the supremum of the set of ρ ∈ (0, 1] such that Ex contains n − i + 1 linearly independent sections convergent on Ux,ρ. Again, this set of such ρ is nonempty by Theorem 3.4. Since s1(x) ≥ · · · ≥ sn(x), the si(x) are the slopes of a polygon N (x) ∈ N P[0, n], which we call the convergence polygon of E at x. Again, the construction is compatible with base change, so it induces a well-defined function N : U an → N P[0, n]. Definition 5.4. For x ∈ U an, we say that E satisfies the Robba condition at x if N (x) is the zero polygon. We have the following analogue of Theorem 3.7. Theorem 5.5. The function N : (X − Z)an → N P[0, n] is continuous. More precisely, there exists a strict skeleton Γ such that N factors through the retraction of X an onto Γ, and the restriction of N to each edge of Γ is affine with integral derivative. Proof. See [133] or [98] or [16] (and Remark 5.6). (cid:3) Remark 5.6. It is slightly inaccurate to attribute Theorem 5.5 to [133] or [16], as the results proved therein are slightly weaker: they require an uncontrolled base extension on K, which creates more options for the strict skeleton Γ. In particular, Theorem 5.5 as stated implies that N is locally constant around any point of type 4, which cannot be established using the methods of [133] or [16]; one instead requires some dedicated arguments found only in [98]. These extra arguments are crucial for the applications of Theorem 5.5 in the contexts described in [95] and [96]. Remark 5.7. Suppose that X = P1 K and ∞ ∈ Z. Given a differential equation as in (3.2.1), we can construct an associated connection E of rank n whose underlying vector bundle is free on the basis e1, . . . , en and whose action of ∇ is given by ∇(e1) = f0(z)en ∇(e2) = f1(z)en − e1 ... ∇(en−1) = fn−2(z)en − en−2 ∇(en) = fn−1(z)en − en−1. A section of E is then horizontal if and only if it has the form ye1 + y′e2 +· · ·+ y(n−1)en where y is a solution of (3.2.1). If length(Z) ≥ 2, each of Nz and N can be computed in terms of the other; this amounts to changing the normalization of certain discs. In particular, the statements of Theorem 3.7 and Theorem 5.5 in this case are equivalent. 8 If length(Z) = 1, we cannot define N as above. However, if K is nontrivially valued, one can recover the properties of Nz by considering N with Z replaced by Z ∪ {x} for some x ∈ U(K) with x sufficiently large (namely, larger than the radius of convergence of any nonentire formal solution at 0). We refer to [16] for further details. Remark 5.8. One can extend Remark 5.7 by defining Nz in the case where X = P1 K and ∞ ∈ Z, and using Theorem 5.5 to establish an analogue of Theorem 3.7. With this modification, we still do not define either N or Nz in the case where X = P1 K and Z = ∅, but this case is completely trivial: the vector bundle E must admit a basis of horizontal sections (see [16]). Theorem 5.9. Suppose that x ∈ Γ ∩ U an is the generic point of a open disc D contained in X and the Robba condition holds at x. (a) If D ∩ Z = ∅, then the restriction of E to D is trivial (i.e., it admits a basis of horizontal sections). (b) If D ∩ Z consists of a single point z at which ∇ is regular, then the Robba condition holds on D − {z}. Proof. Part (a) is a special case of the Dwork transfer theorem; see for instance [93, Theo- rem 9.6.1]. Part (b) follows as in the proof of [93, Theorem 13.7.1]. (cid:3) Remark 5.10. Theorem 5.9(b) is a variant of a result of Christol, which has a slightly stronger hypothesis and a slightly stronger conclusion. In Christol’s result, one must assume either that p = 0, or that p > 0 and the pairwise differences between the exponents of ∇ at z are not p-adic Liouville numbers (see Example 7.19). One however gets the stronger conclusion that the “formal solution matrix” of ∇ at z converges on all of D. Both parts of Theorem 5.9 are examples of transfer theorems, which can be viewed as transferring convergence information from one disc to another. Remark 5.11. Let k be the residue field of K. Suppose that X = P1 K, ∇ is regular everywhere, and the reduction map from Z to P1 k is injective. Then Theorem 5.9 implies that if the Robba condition holds at x1, then it holds on all of U an. For instance, this is the case for (the connection associated via Remark 5.7 to) the hypergeometric equation considered in Example 4.8; more generally, it holds for the hypergeometric equation y′′ − c − (a + b + 1)z z(1 − z) y′ − ab z(1 − z) y = 0 if and only if a, b, c ∈ Zp (the case of Example 4.8 being a = b = 1/2, c = 1). This example and Example 4.5, taken together, suggest that for a general differential equation with one or more accessory parameters, the Robba condition at a fixed point is likely to be of a “fractal” nature in these parameters. For some additional examples with four singular points, see the work of Beukers [21]. Remark 5.12. One can also consider some modified versions of the convergence polygon. For instance, one might take e−si(x) to be the supremum of those ρ ∈ (0, 1] such that the restriction of E to Ux,ρ splits off a trivial submodule of rank at least n − i + 1; the resulting convergence polygons will again satisfy Theorem 5.5. It may be that some modification of this kind can be used to eliminate some hypotheses on p-adic exponents, as in Theorem 7.23. 9 6. Derivatives of convergence polygons We now take a closer look at the local variation of convergence polygons. Throughout §6, continue to retain Hypothesis 5.1. Definition 6.1. For x ∈ X an, a branch of X at x is a local connected component of X −{x}, that is, an element of the direct limit of π0(U − {x}) as U runs over all neighborhoods of x in X. Depending on the type of x, the branches of X can be described as follows. Type 1: A single branch. Type 2: One branch corresponding to each closed point on the curve Cx (defined over the residue field of K) whose function field is the residue field of H(x). Type 3: Two branches. Type 4: One branch. For each branch ~t of X at x, by Theorem 5.5 we may define the derivative of N along ~t (away from x), as an element of P[0, n] with integral vertices; we denote this element by ∂~t(N ). For x of type 1, we also denote this element by ∂x(N ) since there is no ambiguity about the choice of the branch. We may similarly define ∂~t(hi(N )) ∈ R for i = 1, . . . , n, optionally omitting i in the case i = n; note that ∂~t(h(N )) ∈ Z. Theorem 6.2. For z ∈ Z, −∂z(N ) is the polygon associated to the Turrittin-Levelt-Hukuhara decomposition of Ez (see [63, Chapter 4], [83, §11], or [93, Chapter 7]). In particular, this polygon belongs to N P[0, n], its slopes are all nonnegative, and its height equals the irregu- larity Irrz(∇) of ∇ at z. Proof. See [134, §5.7] and [135, §3.6], or see [16]. (cid:3) Corollary 6.3. For z ∈ Z, N extends continuously to a neighborhood of z if and only if ∇ has a regular singularity at z (i.e., its irregularity at z equals 0). In particular, N extends continuously to all of X an if and only if ∇ is everywhere regular. Remark 6.4. Using a similar technique, one can compute the asymptotic behavior of N in a neighborhood of z ∈ Z in terms of the “eigenvalues” occurring in the Turrittin-Levelt- Hukuhara decomposition of Ez. For example, ∇ satisfies the Robba condition on some neighborhood of z if and only if ∇ is regular at z with all exponents in Zp. Remark 6.5. In the complex-analytic setting, the decomposition of Ez does not typically extend to any nonformal neighborhood of z; this is related to the discrepancy between local indices in the algebraic and analytic categories (see Remark 7.8). Nonetheless, there are still some close links between the formal decomposition and the asymptotic behavior of local solutions on sectors at z (related to the theory of Stokes phenomena). In the nonarchimedean setting, by contrast, the decomposition of Ez always lifts to some nonformal neighborhood of z; this follows from a theorem of Clark [46], which asserts that formal solutions of a connection at a (possibly irregular) singular point always converge in some punctured disc. Theorem 6.6. Assume p = 0. For x ∈ U an and ~t a branch of X at x not pointing along ΓX,Z, we have ∂~t(N ) ≤ 0. Proof. This requires somewhat technical arguments not present in the existing literature; Theorem A.6. (cid:3) 10 Remark 6.7. In the setting of Theorem 6.6, the statement that ∂~t(h1(N )) ≤ 0 is equivalent to the Dwork transfer theorem (again see [93, Theorem 9.6.1]). In general, Theorem 6.6 is deduced by relating ∂~t(N ) to local indices, as discussed in §7. This argument cannot work for p > 0 due to certain pathologies related to p-adic Liouville numbers (see Remark 7.18). We are hopeful that one can use a perturbation argument to deduce the analogue of Theorem 6.6 for p > 0, by reducing to the case where N (x) has no slopes equal to 0 and applying results of [93] (especially [93, Theorem 11.3.4]); however, a proof along these lines was not ready at the time of this writing. Remark 6.8. By Theorem 5.5, for each x ∈ U an, there exist only finitely many branches ~t at x along which N has nonzero slope. If x is of type 1 or 4, there are in fact no such branches. If x is of type 3, then the slopes along the two branches at x add up to 0. 7. Subharmonicity and index Using the piecewise affine structure of the convergence polygon, we formulate some addi- tional properties, including local and global index formulas for de Rham cohomology. The local index formula is due to Poineau and Pulita [136], generalizing some partial results due to Robba [143, 144, 146, 148] and Christol-Mebkhout [40, 41, 42, 44]. Unfortunately, in the case p > 0 one is forced to interact with a fundamental pathology in the theory of p-adic differential equations, namely the effect of p-adic Liouville numbers; consequently, the global formula we derive here cannot be directly deduced from the local formula (see Remark 7.14 and Remark 7.18). Hypothesis 7.1. Throughout §7, continue to retain Hypothesis 5.1, but assume in addition that K is algebraically closed. (Without this assumption, one can still formulate the results at the expense of having to keep track of some additional multiplicity factors.) Definition 7.2. For x ∈ U an, let (∆N )x ∈ P[0, n] denote the sum of ∂~t(N ) over all branches ~t of X at x (oriented away from x); by Remark 6.8, this sum can only be nonzero when x is of type 2. Define the Laplacian of N as the P[0, n]-valued measure ∆N taking a continuous function f : U an → R toPx∈U an f (x)(∆N )x. For i = 1, . . . , n, we may similarly define the multiplicities (∆hi(N ))x ∈ R and the Laplacian ∆hi(N ); we again omit the index i when it equals n. Remark 7.3. The definition of the Laplacian can also be interpreted in the context of Thuillier’s potential theory [161], which applies more generally to functions which need not be piecewise affine. Lemma 7.4. We have Z ∆h(N ) =Xz∈Z Irrz(∇). Proof. For e an edge of Γ, we may compute the slopes of N along the two branches pointing into e from the endpoints of e; these two slopes add up to 0. If we add up these slopes over all e, then regroup this sum by vertices, then the sum at each vertex z ∈ Z equals − Irrz(∇) by Theorem 6.2, while the sum at each vertex x ∈ U an is the multiplicity of x in ∆N . This proves the claim. (cid:3) 11 Definition 7.5. For any open subset V of X an, consider the complex 0 → E ∇→ E ⊗ Ω → 0 of sheaves, keeping in mind that if V ∩ Z 6= ∅, then the sections over V are allowed to be meromorphic at V ∩ Z (but not to have essential singularities; see Remark 7.8). We define χdR(V, E) to be the index of the hypercohomology of this complex, i.e., the alternating sum of K-dimensions of the hypercohomology groups. Lemma 7.6. We have (7.6.1) χdR(X an, E) = nχ(U) −Xz∈Z Irrz(∇) = n(2 − 2g(X) − length(Z)) −Xz∈Z Irrz(∇). Proof. Let K0 be a subfield of K which is finitely generated over Q to which X, Z, E, ∇ can be descended. Then choose an embedding K0 ⊂ C and let XC be the base extension of the descent of X, again equipped with a meromorphic vector bundle E and connection ∇. Note that χdR(X an, E) is computed by a spectral sequence in which one first computes the coherent cohomology of E and E ⊗ Ω separately. By the GAGA principle both over C [76, Expos´e XII] and K [48, Example 3.2.6], these coherent cohomology groups can be computed equally well over any of X an, X (or its descent to K0), XC, or X an C . Consequently, despite the fact that the connection is only K-linear rather than O-linear, we may nonetheless conclude that χdR(X an, E) = χdR(X an C , E). (As an aside, this argument recovers a comparison theorem of Baldassarri [11].) To compute χdR(X an C , E), we may either appeal directly to [53, (6.21.1)] or argue directly as follows. Form a finite open covering {Vi}i∈I of XC such that for each nonempty subset S of I, the set VS =Ti∈S Vi satisfies the following conditions. • If nonempty, VS is isomorphic to a simply connected domain in C. • The set VS ∩ Z contains at most one element. χdR(X an C , E) = XS⊆I,S6=∅ (−1)#S−1χdR (VS, E) . It then suffices to check that for each nonempty subset S of I, We then have (7.6.2) χdR(VS, E) = − Irrz(∇) n 0 (VS ∩ Z = {z}) (VS ∩ Z = ∅, VS 6= ∅) (VS = ∅). In case VS = ∅, there is nothing to check. In case VS 6= ∅ but VS ∩ Z = ∅, this is immediate because the restriction of E to VS is trivial. In case VS ∩ Z = {z}, we may similarly replace VS with a small open disc around z, and then invoke the Deligne-Malgrange interpretation of irregularity as the local index of meromorphic de Rham cohomology on a punctured disc [109, Th´eor`eme 3.3(d)]. (cid:3) Theorem 7.7 (Global index formula). We have (7.7.1) χdR(X an, E) = nχ(U) −Z ∆h(N ) = n(2 − 2g(X) − length(Z)) −Z ∆h(N ). Proof. This follows by comparing Lemma 7.4 with Lemma 7.6. (cid:3) 12 Remark 7.8. It may not be immediately obvious why Theorem 7.7 is of value, i.e., why it is useful to express the index of de Rham cohomology in terms of convergence polygons instead of irregularity. As observed by Baldassarri [11, (0.13)], there is a profound difference between the behavior of the index in the complex-analytic and nonarchimedean settings. In the complex case, for any open analytic subspace V of X an C , we have (7.8.1) χdR(V, E) = nχ(V ∩ U an) − Xz∈V ∩Z Irrz(∇) by the same argument as in the proof of (7.6.2). In particular, χdR(V, E) 6= χdR(V − Z, E); that is, the index of de Rham hypercohomology depends on whether we allow poles or essential singularities at the points of Z. By contrast, in the nonarchimedean case, these two indices coincide under a suitable technical hypothesis to ensure that they are both defined; see Example 7.9 for a simple example and Corollary 7.13 for the general case (and Example 7.19 and Remark 7.20 for a counterexample failing the technical hypothesis). This means that in the nonarchimedean case, the “source” of the index of de Rham cohomology is not irregularity, but rather the Laplacian of the convergence polygon (see Theorem 7.12). Example 7.9. Consider the connection associated to Example 4.2 as per Remark 5.7; note that Irr0(∇) = 0, Irr∞(∇) = 1, so χdR(X an, E) = −1. For 0 < α < β, let Vβ, Wα be the subspace z < β, z > α of X an. By Mayer-Vietoris, (7.9.1) χdR(Vβ, E) + χdR(Wα, E) − χdR(Vβ ∩ Wα, E) = χdR(X an, E) = −1. On the other hand, χdR(Vβ, E) (resp. χdR(Wα, E)) equals the index of the operator y 7→ y′ −y on Laurent series in z convergent for z < β (resp. in z−1 convergent for z−1 < α−1). If f =Pn fnz−n, g =Pn gnz−n are two such series, then the equation g = f ′ − f is equivalent to (7.9.2) fn = −gn − (n − 1)fn−1 (n ∈ Z). Suppose first that p−1/(p−1) < α < β. Then given g ∈ K((z−1)), we may solve uniquely for f ∈ K((z−1)), and if g converges on Wα − {∞}, then so does f . We thus compute that χdR(Vβ, E) = −1, χdR(Wα, E) = 0, χdR(Vβ ∩ Wα, E) = 0; namely, the second equality is what we just computed, the third follows from Robba’s index formula [148] (see also [98, Lemma 3.7.5]), and the first follows from the other two plus (7.9.1). In particular, the local index at ∞ equals 0, whereas in the complex-analytic setting it equals −1 by the Deligne-Malgrange formula (see the proof of Lemma 7.6). Suppose next that α < β < p−1/(p−1). Then by contrast, we have χdR(Vβ, E) = 0, χdR(Wα, E) = −1, χdR(Vβ ∩ Wα, E) = 0; namely, the first and third equalities follow from the triviality of ∇ on Vβ, and the second follows from the other two plus (7.9.1). With Example 7.9 in mind, we now describe a local refinement of Theorem 7.7, in which we dissect the combinatorial formula for the index into local contributions. Definition 7.10. For x ∈ U an ∩ ΓX,Z, let valΓ(x) be the valence of x as a vertex of ΓX,Z, taking valΓ(x) = 2 when x lies on the interior of an edge. (We refer to valence instead of 13 degree to avoid confusion with degrees of morphisms.) For x ∈ U an, define (7.10.1) χx(E) =(n(2 − 2g(Cx) − valΓ(x)) − (∆h(N ))x −∆h(N )x if x ∈ ΓX,Z otherwise. Let χ(E) be the R-valued measure whose value on a continuous function f : U an → R is Px∈U an f (x)χx(E). Lemma 7.11. We have χdR(X an, E) = nχ(U) −Z ∆h(N ) =Z χ(E). Proof. This follows from Theorem 7.7 plus the identity (2 − 2g(Cx) − valΓ(x)) = 2 − 2g(X) − length(Z), Xx∈U an∩ΓX,Z which amounts to the combinatorial formula for the genus of an analytic curve [8, §4.16]. (cid:3) Theorem 7.12 (Local index formula). Let V be an open subspace of X an which is the retraction of an open subspace of ΓX,Z. If p > 0, assume some additional technical hypotheses (see Remark 7.14). Then χdR(V, E) =ZV ∩U an χ(E). Proof. See [136, Theorem 3.5.2]. (cid:3) Corollary 7.13. With hypotheses as in Theorem 7.12, χdR(V, E) = χdR(V ∩ U an, E); that is, the index of de Rham hypercohomology is the same whether we allow poles or essential singularities at Z. Remark 7.14. Let Γ be a strict skeleton for which the conclusion of Theorem 5.5 holds. For v a vertex of Γ, define the star of v, denoted ⋆v, as the union of v and the interiors of the edges of Γ incident to v. Let πΓ : X → Γ be the retraction onto Γ, through which N factors. Under the hypotheses of Theorem 7.12, we have χdR(π−1 Γ (⋆v ∩ ⋆w), E) = 0. Γ (⋆v), E) = χv(E), χdR(π−1 (7.14.1) We can then recover Theorem 7.7 from (7.14.1) by using Mayer-Vietoris (and GAGA over K; see Remark 7.8) to write χdR(X an, E) =Xv χdR(π−1 Γ (⋆v), E) −Xv6=w χdR(π−1 Γ (⋆v ∩ ⋆w), E); one may similarly deduce Theorem 7.12 from (7.14.1). Remark 7.15. For a given connection, one can extend the range of applicability of Theo- rem 7.12 by enlarging the set Z; however, this depends on an understanding of how the two sides of the formula depend on Z. To this end, let us rewrite the equality as with Z included in the notation. χdR(V, E, Z) =ZV ∩U an χ(E, Z) 14 Put Z ′ = Z ∪ {z′} for some z′ ∈ U(K). Let x′ ∈ ΓX,Z be the generic point of the open disc Uz ′; by subdividing if necessary, we may view x′ as a vertex of ΓX,Z. Then ΓX,Z ′ is the union of ΓX,Z with a single edge joining x′ to z′ within Uz ′. Let V be an open subset of X an containing z′ which is the retraction of an open subspace of ΓX,Z; it is then also a retraction of an open subspace of ΓX,Z ′. We then have χdR(V, E, Z ′) − χdR(V, E, Z) = −n; namely, by Mayer-Vietoris this reduces to the case where V is a small open disc around z′, in which case we may assume E is trivial on V and make the computational directly. By this computation, Theorem 7.12, and the fact that χz ′(E, Z) = 0, we must also have Z(V ∩U an)\{z ′} (χ(E, Z ′) − χ(E, Z)) = −n. From (7.10.1), we see that χ(E, Z ′) − χ(E, Z) is supported within Uz ′ ∪ {x′}. However, while one can easily compute the convergence polygon associated to Z ′ from the one associated to Z (see Example 4.2 and Remark 8.13 for examples), we do not know how to predict a priori where the support of χ(E, Z ′) − χ(E, Z) will lie within Uz ′ ∪ {x′}. Remark 7.16. One of the main reasons we have restricted attention to meromorphic con- nections on proper curves is that in this setting, Theorem 5.5 ensures that χx(E) = 0 for all but finitely many x ∈ U an. It is ultimately more natural to state Theorem 7.12 for con- nections on open analytic curves, as is done in [136, Theorem 3.8.10]; however, this requires some additional hypotheses to ensure that χ(E) is a finite measure. Definition 7.17. Assume p > 0. A p-adic Liouville number is an element x ∈ Zp − Z such that lim inf m→∞ (cid:26) y : y ∈ Z, y − x ∈ pmZp(cid:27) < +∞. m As in the classical case, p-adic Liouville numbers are always transcendental [63, Proposition VI.1.1]. Remark 7.18. Assume p > 0. The technical hypotheses of Theorem 7.12 are needed In case ∇ has a regular to guarantee the existence of the indices appearing in (7.14.1). singularity at z ∈ Z with all exponents in Zp, these hypotheses include the condition that no two exponents of ∇ at z differ by a p-adic Liouville number; see Example 7.19 for a demonstration of the necessity of such a condition. (Such hypotheses are not needed in Theorem 7.7 because there we only poles rather than essential singularities.) Unfortunately, the full hypotheses are somewhat more complicated to state. They arise from the fact that with notation as in Remark 7.14, one can separate off a maximal compo- nent of E on π−1 Γ (⋆v ∩ ⋆w) which satisfies the Robba condition, to which one may associate some p-adic numbers playing the role of exponents; the hypothesis is that (for any particular v, w) no two of these numbers differ by a p-adic Liouville numbers. The difficulty is that the definition of these p-adic exponents, due to Christol and Mebkhout (and later simplified by Dwork) is somewhat indirect; they occur as “resonant frequencies” for a certain action by the group of p-power roots of unity, which are hard to control except in some isolated cases where they are forced to be rational numbers (e.g., Picard-Fuchs equations, a/k/a Gauss- Manin connections, or connections arising from F -isocrystals in the theory of crystalline cohomology). See [93, Chapter 13] for more discussion. 15 Example 7.19. Assume p > 0 and take X = P1 with the action of ∇ given by K, Z = {0, ∞}. Take E to be free of rank 1 ∇(f ) = λf dz z + df 1-forms on V are seriesP∞ for some λ ∈ K. For α, β with 0 < α < β, let V be the open annulus α < z < β. The z such that for each ρ ∈ (α, β), cnρn → 0 as n → ±∞. If λ = 0, then cnρn → 0 for all ρ ∈ (α, β) if and only if cn/nρn → 0 for all ρ ∈ (α, β), so every 1-form on V with c0 = 0 is in the image of ∇. Note that multiplying the generator of E by t has the effect of replacing λ by λ + 1; it follows that if λ ∈ Z, then the kernel and cokernel of ∇ on V are both 1-dimensional, so χdR(V, E) = 0. n=−∞ cnzn dz If λ ∈ K − Z, then a 1-form is in the image of ∇ on V if and only if for each ρ ∈ (α, β), cn/(n − λ)ρn → 0 as n → ±∞. This holds if λ is not a p-adic Liouville number (see [63, §VI.1] or [93, Proposition 13.1.4]); otherwise, one shows that ∇ has infinite-dimensional cokernel on V , so χdR(V, E) is undefined. Remark 7.20. Example 7.19 provides an example showing that the equality χdR(V, E) = χdR(V ∩U an, E) of Corollary 7.13 cannot hold without conditions on p-adic Liouville numbers. In this example, for all λ, χdR(X an, E) = 0 by Lemma 7.6; but when λ is a p-adic Liouville number, χdR(U an, E) is undefined. Remark 7.21. The net result of Remark 7.18 is that in general, one can only view χx(E) as a virtual local index of E at x, not a true local index. Nonetheless, this interpretation can be used to predict combinatorial properties of the convergence polygon which often continue to hold even without restrictions on p-adic exponents. For example, Theorem 6.6 corresponds to the fact that if V is an open disc in U an, then the dimension of the cokernel of ∇ on V is nonnegative; this argument appears in the proof of Theorem 6.6 in the case p = 0 (see Theorem A.6). Remark 7.22. In light of Remark 7.21, one might hope to establish some inequalities on χx(E). One might first hope to refine Theorem 7.23 by analogy with (7.8.1), by proving that the measure ∆h(N ) is nonnegative; however, this fails already in simple examples such as Example 7.25. On the other hand, since our running hypothesis is that χ(U) ≤ 0, Theorem 7.7 and Lemma 7.11 imply that χx(E) = χdR(X an, E) ≤ nχ(U) ≤ 0. Xx∈U an One might thus hope to refine Theorem 7.23 by proving that χx(E) ≤ 0 for all x ∈ U an. Unfortunately, this is not known (and may not even be safe to conjecture) in full generality, but see Theorem 7.23 for some important special cases. Theorem 7.23. Choose x ∈ U an. (a) Let R(x) be the infimum of the radii of open discs in Ux containing x. If N (x) has no slopes equal to − log R(x), then χx(E) = 0. (b) If x ∈ ΓX,Z, then χx(E) ≤ 0. (c) If p = 0, then χx(E) ≤ 0. 16 Proof. For (a), see [134, Proposition 6.2.11] or [16] (or reduce to the case where N (x) has all slopes greater than − log R(x) and then apply [98, Theorem 5.3.6]). A similar argument implies (b) because in this case, R(x) = 1 and the zero slopes are forced to make a nonpositive contribution to the index. For (c), see Theorem A.9. (cid:3) Remark 7.24. The proof of Theorem 7.23 can also be used to quantify the extent to which N fails to factor through the retract onto ΓX,Z, and hence to help identify a suitable skeleton Γ for which the conclusion of Theorem 5.5 holds. To be precise, for x ∈ ΓX,Z, if the restriction of N to ΓX,Z is harmonic at x, then N is constant on the fiber at x of the retraction of X an onto ΓX,Z. See also [134, §6.3]. Example 7.25. Let h be a nonzero rational function on X, and take Z to be the pole locus of h. Let E be the free bundle on a single generator v equipped with the connection ∇(f v) = v ⊗ (df + f dh). For each z ∈ Z, Irrz(∇) equals the multiplicity of z as a pole of h. By Lemma 7.11, χx(E) = χ(U) − m Xx∈U an where m is the number of poles of h counted with multiplicity. Suppose now that there exists a point x ∈ U an with g(Cx) > 0. By multiplying h by a suitably large element of K, we can ensure that N (x) has positive slope. In this case, by Theorem 7.23 we must have χx(E) = 0 > 2 − 2g(Cx) − valΓ(x). In particular, while R ∆h(N ) = Pz∈Z Irrz(∇) ≥ 0, the measure ∆h(N ) is not necessarily nonnegative. 8. Ramification of finite morphisms Hypothesis 8.1. Throughout §8, let f : Y → X be a finite flat morphism of degree d of smooth, proper, geometrically connected curves over K (a nonarchimedean field of characteristic 0), and suppose that the restriction of f to U (the complement of a finite set Z of closed points of X) is ´etale and Galois with Galois group G. Let π : G → GL(V ) be a faithful representation of G on an n-dimensional vector space V over K. Note that since G is finite and π is faithful (and K is of characteristic 0), the Tannakian category generated by V contains all finite-dimensional K-linear representations of G, including the regular representation. Definition 8.2. Equip EV = OY ⊗K V ∨ with the diagonal action of G induced by the Galois action on OY and the action via π∨ on V ∨. By faithfully flat descent, G-invariant sections of EV can be identified with sections of a vector bundle E on U; moreover, the trivial connection on EV induces a connection ∇ on E. We are thus in the situation of Hypothesis 5.1. Remark 8.3. One way to arrive at Hypothesis 8.1 is to start with a non-Galois cover g : W → X, let f be the Galois closure, and let π be the representation of G induced by the trivial representation of the subgroup of G fixing W . In this case, E is just the pushforward of the trivial connection on g−1(U). 17 Remark 8.4. Following up on the previous remark, note that (E, ∇) is always a subobject of the pushforward of the trivial connection on f −1(U). Since the latter is the Gauss-Manin connection associated to a smooth proper morphism (of relative dimension 0), it is everywhere regular by virtue of the Griffiths monodromy theorem [83, Theorem 14.1]. This implies in turn that for z ∈ Z, our original connection satisfies Irrz(∇) = 0 and the exponents of ∇ at z are rational. To say more, let N be the order of the inertia subgroup of G corresponding to some element of f −1({z}); we then claim that the exponents of ∇ at z belong to 1 N Z and generate this abelian group. To check this, we may assume that K is algebraically closed. By the My∈f −1({z}) Frac(bOY,y), Cohen structure theorem, the field Frac(bOX,z) may be identified with K((t)). The vector space Ez = E ⊗OU Frac(bOX,z) over this field splits compatibly with ∇ as a direct sum LN −1 each summand of which is isomorphic to K((t1/N )). If we split this summand further as i=0 ti/N K((t)), then the i-th summand contributes an exponent congruent to i/N modulo Z. In particular, when π is the regular representation, the exponents of ∇ at z belong to 1 Z/Z. For a general π, on one hand, π occurs as a summand N of the regular representation, so the exponents of ∇ at z again belong to 1 N Z. On the other hand, the regular representation occurs as a summand of some tensor product of copies of V and its dual (see Hypothesis 8.1), so the group generated by the exponents must also fill out the quotient 1 N Z/Z (although the exponents themselves need not). Z and fill out the quotient 1 N Definition 8.5. For L a nonarchimedean field over K, let CL denote a completed algebraic closure of L, and let BL be the subset of XL for which x ∈ BL if and only if the preimage of x in Y an CL does not consist of d distinct points. (For a demonstration of this definition, see Example 8.12.) We have the following remark about BL echoing an earlier observation about ΓX,Z (see Definition 2.5). Remark 8.6. Let L′ be a nonarchimedean field containing L. Then BL′ is contained in the inverse image of BL under the restriction map X an L , but this containment is typically strict if L′ is not the completion of an algebraic extension of L. This is because BL′ only contains points of types 2 or 3 (except for the points of Z), whereas the inverse image of BL typically also contains some points of type 1. L′ → X an Example 8.7. Take K = Qp(ζp), X = P1 K, let f : Y → X be the map z 7→ zp, identify G with Z/pZ so that 1 ∈ Z/pZ corresponds to the map z 7→ ζz on f −1(U), and let π : G → GL1(K) be the character taking 1 to ζ −1 p . Then E is free on a single generator v satisfying ∇(v) = − 1 p z−1v ⊗ dz. This is also the connection obtained from Example 4.3 by applying Remark 5.7. K, Z = {0, ∞}, Y = P1 Put ω = p−1/(p−1) = ζp − 1. For each x ∈ U, we may define the normalized diameter of x as an element of [0, 1] defined as follows: choose an extension L of K such that x lifts to some x ∈ U(L), then take the infimum of all ρ ∈ (0, 1] such that Ux,ρ meets the inverse image of x in U an L (or 1 if no such ρ exists). With this definition, for any L, the set BL consists of Z plus all points with normalized diameter in [ωp, 1] (see for example [93, Lemma 10.2.2]). 18 Lemma 8.8. Suppose that L is algebraically closed. Let V be an open subset of X an L which is disjoint from BL. Then the map f −1(V ) → V of topological spaces is a covering space map. L . Since Y an Proof. For each x ∈ V , the local ring of X an L at x is henselian [17, Theorem 2.1.5]; hence for each y ∈ f −1({x}), we can find a neighborhood Wy of y in f −1(V ) which is finite ´etale over its image in X an L is Hausdorff, we can shrink the Wy so that they are pairwise disjoint and all have the same image V ′ in X an L . The maps Wy → V ′ are then finite ´etale of some degrees adding up to deg(f ) = d. Since #f −1({x}) = d by hypothesis, these degrees must all be equal to 1; hence each map Wy → V ′ is actually an isomorphism of analytic spaces, and in particular a homeomorphism of topological spaces. (cid:3) Theorem 8.9. For L a nonarchimedean field over K and x ∈ U(L), e−s1(N (x)) equals the supremum of ρ ∈ (0, 1] such that Ux,ρ ∩ BL = ∅. Proof. We may assume without loss of generality that L is itself algebraically closed. If Ux,ρ ∩ BL = ∅, then by Lemma 8.8, the map f −1(Ux,ρ) → Ux,ρ is a covering space map of topological spaces. Since Ux,ρ is contractible and hence simply connected, f −1(Ux,ρ) splits topologically as a disjoint union of copies of Ux,ρ. From this, it follows easily that the restriction of E to Ux,ρ is trivial. Conversely, suppose that the restriction of E to Ux,ρ is trivial. Then the same holds when V is replaced by any representation in the Tannakian category generated by V ; since G is finite and π is faithful, this includes the regular representation. That is, the pushforward of the trivial connection on f −1(U) has trivial restriction to Ux,ρ. In addition to the connection, the sections of this restriction inherit from OY a multipli- cation map, and the product of two horizontal sections is again horizontal. Consequently, the horizontal sections form a finite reduced L-algebra of rank d; since L is algebraically closed, this algebra must split as a direct sum of d copies of L. This splitting corresponds to a splitting of f −1(Ux,ρ) into d disjoint sets, proving that Ux,ρ ∩ BL = ∅. (cid:3) Remark 8.10. One may view Theorem 8.9 as saying that under a suitable normalization, e−s1(N (x)) measures the distance from x to BL. This suggests the interpretation of BL as an “extended ramification locus” of the map f ; for maps from P1 K to itself, this interpretation has been adopted in the context of nonarchimedean dynamics (e.g., see [68, 69]). However, this picture is complicated by Remark 8.6; roughly speaking, even for x ∈ BL, the “distance from x to itself” must be interpreted as a nonzero quantity. In any case, Theorem 8.9 suggests the possibility of relating the full convergence polygon to more subtle measures of ramification, such as those considered recently by Temkin [47, 159]. One older result in this direction is the theorem of Matsuda and Tsuzuki; see Theorem 8.11 below. Theorem 8.11. Assume that K has perfect residue field of characteristic p > 0. Suppose that x ∈ ΓX,Z is of type 2, choose y ∈ f −1({x}), and suppose that H(y) is unramified over H(x). Let κx, κy be the residue fields of the nonarchimedean fields H(x), H(y). Let π be the restriction of π along the identification of H := Gal(κy/κx) ∼= Gal(H(y)/H(x)) with the stabilizer of y in G. (a) The polygon N (x) is zero. (b) Let ~t be a branch of X at x, and let v be the point on Cx corresponding to ~t (see Definition 6.1). Then ∂~t(N ) computes the wild Hasse-Arf polygon of π at v, i.e., the 19 Newton polygon in which the slope s ≥ 0 occurs with multiplicity dimK(V H s+/V H s). In particular, ∂~t(h(N )) computes the Swan conductor of π at v. Proof. See [163], [114], or [87]. (cid:3) Example 8.12. Set notation as in Example 8.7, except take Z = {0, 1, ∞}. Let x ∈ ΓX,Z be the generic point of the disc z − 1 ≤ ωp, which we can also write as u ≤ 1 for u = (z − 1)/(ζp − 1)p. Let y be the unique preimage of x in Y . Let t be the coordinate (z − 1)/(ζp − 1) on Y ; then κx = Fp(u) while κy = Fp(t) where tp − t = u. For ~t the branch of x towards x1, we have ∂~t(h(N )) = 1; as predicted by Theorem 8.11, this equals the Swan conductor of the residual extension at ∞. Remark 8.13. In Example 8.12, it is necessary to include 1 in Z to force x into ΓX,Z, so that Theorem 8.11 applies; otherwise, we would have ∂~t(h(N )) = 0 as in Example 4.3. This provides an explicit example of the effect of enlarging Z on the behavior of N , as described in Remark 7.15. Remark 8.14. To generalize Theorem 8.11 to cases where H(y) is ramified over H(x), it may be most convenient to use Huber’s ramification theory for adic curves [80], possibly as refined in [47, 159]. In a similar vein, the global index formula (Theorem 7.7) is essentially the Riemann-Hurwitz formula for the map f , in which case it should be possible to match up the local contributions appearing in Theorem 7.7 with ramification-theoretic local contributions. Remark 8.15. Suppose that p > 0, X = P1 K, and f extends to a finite flat morphism of smooth curves over oK with target P1 oK . By [51, Theorem 3.1], E admits a unit-root Frobenius structure in a neighborhood of x1, from which it follows that E satisfies the Robba condition at x1. Let k be the residue field of K. If in addition the points of Z have distinct projections to P1 k, then by Remark 5.11, E satisfies the Robba condition everywhere. By Remark 6.4, this implies that E has exponents in Zp; by Remark 8.4, this implies that f is tamely ramified (i.e., the inertia group of each point of Y has order coprime to p). By contrast, if the points of Z do not have distinct projections to P1 k, then f need not be tamely ramified; see §11. Remark 8.16. For a connection derived from a finite morphism, in case p > 0 the technical conditions of Remark 7.18 are always satisfied. An important consequence of this statement for our present work is that the conclusion of Theorem 7.23(c) can be established for such a connection even when p > 0; see Remark A.11. 9. Artin–Hasse exponentials and Witt vectors We will conclude by specializing the previous discussion to cyclic covers of discs in con- nection with the Oort local lifting problem. In preparation for this, we need to recall some standard constructions of p-adic analysis. Hypothesis 9.1. Throughout §9, fix a prime p and a positive integer n. In some algebraic closure of Qp, fix a sequence of primitive pn-th roots of unity ζpn such that ζ p Definition 9.2. The Artin-Hasse exponential series at p is the formal power series pn = ζpn−1. Ep(t) := exp ∞Xi=0 20 tpi pi! . Lemma 9.3. We have Ep(t) ∈ Z(p)JtK. In particular, Ep(t) converges for t < 1. Proof. See for instance [93, Proposition 9.9.2]. (cid:3) Lemma 9.4. Let Z(p)hti be the subring of Z(p)JtK consisting of seriesP∞ ci converge p-adically to 0 as i → ∞. Then if we define the power series i=0 citi for which the f (z, t) := using Lemma 9.3, we have Ep(zt)Ep(tp) Ep(t)Ep(ztp) ∈ Z(p)Jt, 1 − zK Proof. Write f (z, t) = exp g(z, t) with f (z, t) ∈ z + (t − 1)Z(p)htiJ1 − zK. pi g(z, t) = (zpi − 1)(tpi − tpi+1) (−1)j(cid:18)pi ∞Xj=1 ∞Xi=0 j(cid:19)(1 − z)j tpi − tpi+1 ∞Xi=0 j − 1(cid:19)(tpi ∞Xi=0(cid:18)pi − 1 ∞Xj=1 (1 − z)j (cid:18) 0 j − 1(cid:19)t + ∞Xj=1 j−1(cid:1) converges p-adically to (cid:0) −1 For each fixed j, (cid:0)pi−1 g(z, t) ∈ (Z(p)hti ⊗Z Q)J1 − zK and (1 − z)j (−1)j = = = j j (−1)j pi − tpi+1 ) j − 1 (cid:19)(cid:19) tpi! . j − 1(cid:19) −(cid:18)pi−1 − 1 ∞Xi=1(cid:18)(cid:18)pi − 1 j−1(cid:1) = (−1)j−1 as i → ∞. It follows that g(z, t) ≡ − (1 − z)j j (mod (t − 1)(Z(p)hti ⊗Z Q)J1 − zK). ∞Xj=1 This then implies that f (z, t) ∈ (Z(p)hti ⊗Z Q)J1 − zK and f (z, t) ≡ z (mod (1 − t)(Z(p)hti ⊗Z Q)J1 − zK). Since f (z, t) also belongs to Z(p)Jt, 1−zK by Lemma 9.3, we may deduce the claimed inclusion. (cid:3) Definition 9.5. Let Wn denote the p-typical Witt vector functor. Given a ring R, the set Wn(R) consists of n-tuples a = (a0, . . . , an−1), and the arithmetic operations on Wn(R) are characterized by functoriality in R and the property that the ghost map w : Wn(R) → Rn given by (9.5.1) (a0, . . . , an−1) 7→ (w0, . . . , wn−1), wi = pjapi−j j iXj=0 is a ring homomorphism for the product ring structure on Rn. For any ideal I of R, let Wn(I) denote the subset of Wn(R) consisting of n-tuples with components in I; since Wn(I) = ker(Wn(R) → Wn(R/I)), it is an ideal of Wn(R). 21 Various standard properties of Witt vectors may be derived by using functoriality to reduce to polynomial identities over Z, then checking these over Q using the fact that the ghost map is a bijection if p−1 ∈ R. Here are two key examples. (i) Define the Teichmuller map sending r ∈ R to [r] := (r, 0, . . . , 0) ∈ Wn(R). Then this map is multiplicative: for all r, s ∈ R, [rs] = [r][s]. (ii) Define the Verschiebung map sending a ∈ Wn(R) to V (a) := (0, a0, . . . , an−2) ∈ Wn(R). Then this map is additive: for all a, b ∈ Wn(R), Vn(a + b) = Vn(a) + Vn(b). Definition 9.6. In case R is an Fp-algebra, the Frobenius endomorphism ϕ : R → R extends by functoriality to Wn(R) and satisfies (9.6.1) pa = (V ◦ ϕ)(a) = (ϕ ◦ V )(a) (a ∈ Wn(R)); It follows that for general R, if we define the map σ sending see for instance [81, §0.1]. a ∈ Wn(R) to σ(a) = (ap (9.6.2) 0, . . . , ap n−1) ∈ Wn(R), then pa − (V ◦ σ)(a) ∈ Wn(pR) Beware that σ is in general not a ring homomorphism. Definition 9.7. By Lemma 9.3, we have (a ∈ Wn(R)). (ζpn−i − 1) tpi pi! ∈ Z(p)[ζpn]JtK. En,p(t) := We may also define which we may also write as (9.7.1) Ep(ζpnt) Ep(t) = exp n−1Xi=0 n−1Yi=0 En,p(a) = exp n−1Xi=0 (ζpn−i − 1) wi pi! En,p(a) := En−i,p(ai) ∈ Z(p)[ζpn]Ja0, . . . , an−1K, for wi as in (9.5.1). Consequently, in Z(p)[ζpn]Ja0, . . . , an−1, b0, . . . , bn−1K, (9.7.2) En,p(a)En,p(b) = En,p(a + b). Definition 9.8. By Lemma 9.3, we may define the formal power series Fn,p(t) := En,p(t) En,p(tp) and Fn,p(a) := = exp n−1Xi=0 n−1Yi=0 = En,p(a) En,p(σ(a)) (ζpn−i − 1)(tpi − tpi+1) pi ! ∈ Z(p)[ζpn]JtK Fn−i,p(ai) ∈ Z(p)[ζpn]Ja0, . . . , an−1K. By (9.7.1), in Z(p)[ζpn]Ja0, . . . , an−1, b0, . . . , bn−1K we have (9.8.1) Fn,p(a)Fn,p(b) = Fn,p(a + b)En,p(σ(a + b) − σ(a) − σ(b)). We will see shortly (Lemma 9.10) that Fn,p(t) has radius of convergence greater than 1, which implies an analogous assertion for Fn,p(a). For p > 2, this is shown in [114, Proposition 1.10] 22 using a detailed computational argument; our argument follows the more conceptual ap- proach given in [140, Theorem 2.5]. Definition 9.9. By Lemma 9.3, we may define the formal power series By (9.7.2), (9.9.1) We also have (9.9.2) Gn,p(a) = En,p(pa) En−1,p(a) ∈ Z(p)[ζpn]Ja0, . . . , an−1K. Gn,p(a)Gn,p(b) = Gn,p(a + b). Gn,p(a) = En,p(pa − (0, ap 0, . . . , ap n−2))Fn−1,p(a)−1. Lemma 9.10. (a) We have Gn,p(a) ∈ Z(p)[ζpn]J(ζpn − 1)a0, . . . , (ζp − 1)an−1K. (b) The power series Fn,p(a) converges on a polydisc with radius of convergence strictly greater than 1. Proof. Write Note that p−1−[ζpn−j ]−· · ·−[ζ p−1 1)Z(p)[ζpn−j ]). This proves (a). pn−j ] maps to zero in Wn−j(Fp) and so belongs to Wn−j((ζpn−j − To prove (b), apply (9.9.2) to write Fn,p(a) = Gn+1,p(a)−1En+1,p(pa − (V ◦ σ)(a)). Since Gn+1,p(0) = 1, (a) implies that Gn+1,p(a)−1 converges on a polydisc with radius of convergence strictly greater than 1. Since En+1,p(a) ∈ Z(p)[ζpn+1]Ja0, . . . , an−1K and (9.6.2) implies that pa − (V ◦ σ)(a) ∈ Wn(pR), En+1,p(pa − (V ◦ σ)(a)) also converges on a polydisc with radius of convergence strictly greater than 1. This proves (b). (cid:3) Lemma 9.11. For m ∈ Z, we have (9.11.1) Fn,p(a + b + m)Fn,p(a)−1En,p(σ(a + b + m) − σ(a) − σ(b) − m) = ζ m pn as an equality of elements of Zp[ζpn]Ja0, . . . , an−1, b0, . . . , bn−1K. Note that the presence of m prevents us from heedlessly applying (9.7.2), because for instance En,p(m) does not make sense. (We like to think of this as an example of conditional convergence in a nonarchimedean setting.) Similarly, we must work over Zp rather than Z(p). 23 (ζpn−j−i − 1)(p − 1 − ζpn−j−i − · · · − ζ p−1 pn−j−i) api j pi! api j pi! (p(ζpn−j−i − 1) − (ζpn−j−1−i − 1)) Gn,p(a) = = = exp n−j−1Xi=0 exp n−j−1Xi=0 En−j,p(cid:16)(p − 1 − [ζpn−j ] − · · · − [ζ p−1 n−1Yj=0 n−1Yj=0 n−1Yj=0 pn−j ])[aj](cid:17) . Proof. Convergence of Fn,p(a + b + m) is guaranteed by Lemma 9.10. Convergence of En,p(σ(a + b + m) − σ(a) − σ(b) − m) is guaranteed by Lemma 9.3 and the fact that σ(a + b + m) − σ(a) − σ(b) − m ∈ Wn((ζpn − 1)Zp[ζpn]Ja0, . . . , an−1, b0, . . . , bn−1K). Thus the left side of (9.11.1) is well-defined. Using (9.7.2), we see that this quantity is constant as a power series in a0, . . . , an−1, b0, . . . , bn−1; it thus remains to prove that (9.11.2) Fn,p(m)En,p(σ(m) − m) = ζ m pn. Using (9.7.2) and (9.8.1), we see that for m, m′ ∈ Z, Fn,p(m)En,p(σ(m) − m)Fn,p(m′)En,p(σ(m′) − m′) = Fn,p(m + m′)En,p(σ(m + m′) − σ(m) − σ(m′))En,p(σ(m) − m)En,p(σ(m′) − m′) = Fn,p(m + m′)En,p(σ(m + m′) − m − m′), so both sides of (9.11.2) are multiplicative in m. It thus suffices to check (9.11.2) for m = 1 = (1, 0, . . . ), in which case σ(m) = m and so the desired equality becomes Fn,p(1) = ζpn. This follows from Lemma 9.4 by evaluating at z = ζpn. (cid:3) 10. Kummer-Artin-Schreier-Witt theory In further preparation for discussion of the Oort local lifting problem, we describe a form of Kummer-Artin-Schreier-Witt theory for cyclic Galois extensions of a power series field. Hypothesis 10.1. Throughout §10, fix a positive integer n, assume that K is discretely valued, the residue field k of K is algebraically closed of characteristic p > 0, and K contains a primitive pn-th root of unity ζpn. Put F = k((z)). Definition 10.2. For ρ ∈ (0, 1), let A(ρ, 1) be the annulus ρ < z < 1 in P1 K; the analytic functions on A(ρ, 1) can be viewed as certain Laurent series in z. The union of the rings O(A(ρ, 1)) over all ρ ∈ (0, 1) is called the Robba ring over K and will be denoted R. (This ring can be interpreted as the local ring of the adic point of P1,an K specializing x1 in the direction towards 0.) Let Rbd be the subring of R consisting of formal sums with bounded coefficients; these are exactly the elements of R which define bounded analytic functions on A(ρ, 1) for some ρ ∈ (0, 1). The ring Rbd carries a multiplicative Gauss norm defined by (cid:12)(cid:12)(cid:12)(cid:12)(cid:12)Xi∈Z aizi(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) = max i {ai}; let Rint be the subring of Rbd consisting of elements of Gauss norm at most 1. Lemma 10.3. The ring Rint is a henselian discrete valuation ring. Consequently, Rbd is a henselian local field with residue field F . Proof. See [113, Proposition 3.2]. (cid:3) We next prepare to formulate the comparison between Kummer theory and Artin-Schreier- Witt theory by introducing the two sides of the comparison. 24 Definition 10.4. For any field L of characteristic not equal to p, for Lsep a separable closure of L, taking Galois cohomology on the exact sequence 1 → µpn → (Lsep)× •pn → (Lsep)× → 1 of GL-modules gives the Kummer isomorphism L×/L×pn ∼= H 1(GL, µpn) because H 1(GL, (Lsep)×) = 0 by Noether’s form of Hilbert’s Theorem 90. In the case L = Rbd, by Lemma 10.3 we have a surjection GL → GF identifying GF with the quotient of the maximal unramified extension of L. We thus obtain a restriction map H 1(GF , µpn) → H 1(GL, µpn) and thus a map H 1(GF , µpn) → L×/L×pn. Note that GF acts trivially on µpn, so we may identify µpn as a GF -module with Z/pnZ by identifying our chosen primitive pn-th root of unity ζpn ∈ µpn with 1 ∈ Z/pnZ. We thus end up with a homomorphism (10.4.1) (Beware that the opposite sign is used to normalize the isomorphism µpn ∼= Z/pnZ in [114], [140], leading to some minor differences in the formulas.) H 1(GF , Z/pnZ) → (Rbd)×/(Rbd)×pn . Definition 10.5. Consider the exact sequence 0 → Z/pnZ = Wn(Fp) → Wn(F sep) 1−ϕ→ Wn(F sep) → 0 where ϕ denotes the Frobenius endomorphism of Wn(F sep). The additive group Wn(F sep) is a successive extension of copies of the additive group of F sep; since H 1(GF , F sep) = 0 by the additive version of Theorem 90, we also have H 1(GF , Wn(F sep)) = 0. We thus obtain the Artin-Schreier-Witt isomorphism (10.5.1) coker(1 − ϕ, Wn(F )) ∼= H 1(GF , Z/pnZ). Combining this isomorphism with the map (10.4.1) derived from the Kummer isomorphism, we obtain a homomorphism (10.5.2) coker(1 − ϕ, Wn(F )) → (Rbd)×/(Rbd)×pn . We note in passing how Swan conductors appear in the Artin-Schreier-Witt isomorphism. See also [82, Theorem 3.2], [71, Theorem 1.1], [160, Proposition 4.2]; see especially [28, Theorem 4.9] for the full computation of coker(1 − ϕ, Wn(F )). Lemma 10.6. For a ∈ Wn(F ), let π : GF → K × be the character corresponding via (10.5.1) to the class of a in coker(ϕ − 1, Wn(F )). For j = 0, . . . , n − 1, let mj be the negation of the z-adic valuation of aj, and assume that mj is not a positive multiple of p. Then the Swan conductor of π equals max{0, m0pn−1, m1pn−2, . . . , mn−1}. Proof. By hypothesis, if mj is positive then it is not divisible by p, so mjpn−1−i has p-adic valuation n−i−1. Consequently, any two of the quantities m0pn−1, m1pn−2, . . . , mn−1, if they are nonzero, must be distinct. It thus suffices to check the claim in case a is a Teichmuller element [a] for some a ∈ F of z-adic valuation −m for some integer m which is positive and not divisible by p. By splitting a into powers of z, we may further reduce to the case 25 a = cz−m. Using the compatibility of Swan conductors with tame base extensions, we may further reduce to the case a = z−1. For j = 1, . . . , n, let Fj be the extension of F obtained by adjoining the coordinates b0, . . . , bj−1 of a Witt vector b satisfying b − ϕ(b) = [a]. p p It is clear that b0 − b 0 = a. By direct computation, one sees that for j = 1, . . . , n − 1, j has the same z-adic valuation as bj−1apj−pj−1 and that this valuation is −(pj − pj−1 + bj − b · · · + p−j). It follows that the breaks in the lower numbering filtration of Gal(Fn/F ) occur at (p2j+1 + 1)/(p + 1) for j = 0, . . . , n − 1; by Herbrand’s formula [156, Chapter IV], the breaks in the upper numbering filtration occur at 1, p, . . . , pn−1. The last of these breaks is the Swan conductor of π, proving the claim. (cid:3) Corollary 10.7. With notation as in Lemma 10.6, if bj is the Swan conductor of π⊗pn−j , then bj ≥ pbj−1 for j = 1, . . . , n. Proof. The representation π⊗pn−j corresponds via (10.5.1) to the class of pn−ja in coker(ϕ − 1, Wj(F )). By (9.6.1), this class is also represented by V n−j(a). We may now apply Lemma 10.6 to deduce the claim. (cid:3) We now make explicit the relationship between the Kummer and Artin-Schreier-Witt isomorphisms. Theorem 10.8 (Matsuda). The homomorphism (10.5.2) is induced by a homomorphism (10.8.1) Wn(Rint) → (Rint)×, a 7→ En,p(pna). Proof. We first clarify the interpretation of the expression En,p(pna) as an element of Rint. i=0 En−i,p(ai), this amounts to evaluating the formal power Since by definition En,p(a) =Qn−1 series En,p(pnt) at t = g for some g =Pi∈Z gizi ∈ Rint. By Lemma 9.3, there exists ρ1 > 1 such that the series En,p(pnt) converges for t < ρ1. By the definition of R, there exists ρ2 ∈ (0, 1) such that for any z in any nonarchimedean field containing K with ρ2 < z < 1, gizi < ρ1 for all but finitely many i < 0. The same remains true if we replace ρ2 by any larger value in (0, 1); by so doing, we can force the inequality gizi < ρ1 to hold for all i < 0. The same inequality also holds for i ≥ 0 because gi ≤ 1 by the definition of Rint; consequently, g(z) belongs to the region of convergence of En,p(pnt), and the evaluation En,p(png) is well-defined as an analytic function on the annulus A(ρ2, 1). Since both En,p(pnt) and g have coefficients in oK, the same is true of the composition, so En,p(png) ∈ Rint. By (9.7.2), the map (10.8.1) is a homomorphism. Given a, choose a minimal finite sepa- rable extension S of F such that there exists b ∈ Wn(S) with b − ϕ(b) = a. (This amounts to forming a tower of Artin-Schreier extensions over F .) Apply Lemma 10.3 to construct a finite ´etale algebra S int over Rint with residue field S. Choose a lift b ∈ S int of b; then a + σ(b) − b ∈ Wn(pS int). By Lemma 9.10(b), we may define an element f := Fn,p(b)En,p(a + σ(b) − b) ∈ S int. Then f pn = En,p(pna). 26 On one hand, the image of a in coker(ϕ − 1, Wn(F )) corresponds to the element of H 1(GF , Z/pnZ) which factors through H 1(Gal(S/F ), Z/pnZ) and sends g ∈ Gal(S/F ) to the integer m ∈ Z/pnZ for which ϕ(b) = b + m. On the other hand, we have g(b) = b + m + c for some c ∈ W ((ζpn − 1)Rint n ), and so g(f ) = Fn,p(b + c + m)En,p(a + σ(b + c + m) − b − c − m) = Fn,p(b + c + m)Fn,p(b)−1En,p(σ(b + c + m) − σ(b) − c − m)Fn,p(b)En,p(a + σ(b) − b) = ζ m pnf by Lemma 9.11. It follows that (10.8.1) induces (10.5.2) as desired. (cid:3) Remark 10.9. By comparing a character with its p-th power, we may deduce from Theo- rem 10.8 that En,p(pna) En−1,p(pn−1a) ∈ (Rint)×pn−1 . This may also be seen directly from Lemma 9.10 by rewriting the left side as Gn−1,p(a)pn−1. 11. Automorphisms of a formal disc To conclude, we use Kummer-Artin-Schreier-Witt theory to translate the Oort local lifting problem into a question about the construction of suitable connections on P1 K, and use this interpretation to describe existing combinatorial invariants connected with the Oort problem in terms of convergence polygons. Hypothesis 11.1. Throughout §11, retain Hypothesis 10.1 and additionally fix a ∈ Wn(Rint). Definition 11.2. Let π : GF → µpn be the character corresponding to a via the maps Wn(Rint) → coker(ϕ − 1, Wn(F )) ∼= H 1(GF , µpn) (the latter isomorphism being (10.5.1)). For i = 1, . . . , n, let bi be the Swan conductor of π⊗pn−i. As before, let x1 ∈ P1,an K denote the generic point of the disc z < 1. The residue field H(x1) is the fraction field of a Cohen ring for the field k(z). We have an embedding of H(x1) into the completion of Rbd for the Gauss norm, arising from an inclusion of Cohen rings lifting the inclusion k(z) ⊂ F . Apply the Katz-Gabber construction [84] to lift π to a representation of Gk(z) unramified away from {0, ∞}, then identify the latter with a representation π : GH(x1) → µpn. By Crew’s analogue of the Katz-Gabber construction for p-adic differential equations [52], the character π arises from a finite Galois cover of A(ρ1, ρ2) for some ρ1 < 1 < ρ2. We may then proceed as in Definition 8.2 to obtain a rank 1 bundle En with connection on this subspace. As in [140, Theorem 3.1] (modulo a sign convention; see Definition 10.4), this connection can be described explicitly as the free vector bundle on a single generator v equipped with the connection ∇(v) = (ζpn−j − 1)apj−1 i v ⊗ dai. n−1Xi=0 n−i−1Xj=0 Formally, we have dv = v ⊗ d log En,p(a). For i = 1, . . . , n, put Ei = E ⊗pn−i n in a similar fashion. . Then Ei corresponds to the character π⊗pn−i of order pi 27 Remark 11.3. The construction given in Definition 11.2 is precisely the (ϕ, ∇)-module associated to π by the work of Fontaine and Tsuzuki [164]. In particular, it is unique in the sense that given any other construction, the two become isomorphic on A(ρ1, ρ2) for some convenient choice of ρ1 < 1 < ρ2. Similarly, the restriction of the connection to A(1, ρ2) is the (ϕ, ∇)-module associated to the restriction of π to the decomposition group at ∞, and enjoys a similar uniqueness property. In fact, we can say something stronger. Recall that the Katz extension of a representation of GF has only tame ramification at ∞. Since π factors through a p-group, its Katz extension must in fact be unramified at ∞, so it descends to a representation of the ´etale fundamental group of P1 K − {0}. Such a representation gives rise to an overconvergent F -isocrystal, as shown by Crew [50, Theorem 3.1]; this means that for some choice of ρ1 < 1 < ρ2, the connection on A(ρ1, ρ2) described above extends to the subspace z > ρ1 of P1 K. Definition 11.4. Let S be the fixed field of ker(π); we may identify S with k((u)) for some parameter u. The action of GF defines a continuous k-linear action τ of µpn on kJuK. A solution of the lifting problem for π is a lifting of τ to a continuous oK-linear action τ of µpn on oKJuK. Conjecture 11.5 (Oort). A solution of the lifting problem exists for every π. A spectacular breakthrough on this problem has been made recently in work of Obus– Wewers and Pop [124, 139]. Theorem 11.6. For fixed π, a solution of the lifting problem exists over some finite extension of K (that is, the lifting problem is solved if we do not insist on the field of definition). We will not say anything more here about the techniques used to prove Theorem 11.6. Instead, we describe an equivalence between solutions of the lifting problem for π and ex- tensions of the connection on En. Definition 11.7. Suppose that τ is a solution of the lifting problem. Then τ gives rise to a finite Galois cover of the disc z < 1, which thanks to Remark 11.3 may be glued together with the cover from Definition 11.2 to give a finite ramified cover fn : Yn → X with X = P1 K, such that x1 has a unique preimage in Yn. (This cover is constructed a priori at the level of analytic spaces, but descends to a cover of schemes by rigid GAGA; see Remark 7.8.) For i = 1, . . . , n, let fi : Yi → X be the cover corresponding to ρpn−i in similar fashion, and let Zi be the ramification locus of fi; also put Z0 = ∅. For each x ∈ Zi − Zi−1, En is regular at x with exponent m/pn−i+1 for some m ∈ Z − pZ. Using the Riemann-Hurwitz formulas in characteristic 0 and p, we obtain the following relationship between the ramification of τ and of fn. Lemma 11.8. With notation as in Definition 11.7, for i = 1, . . . , n, (11.8.1) 2 − 2g(Yi) = 2pi − (11.8.2) = 2pi − iXj=1 iXj=1 (pi − pj−1)(length(Zj) − length(Zj−1)) (pj − pj−1)bj. 28 Consequently, (11.8.3) length(Zi) = bi + 1 (i = 1, . . . , n). Proof. The Rieman-Hurwitz formula for fi asserts that 2 − 2g(Yi) = deg(fi)(2 − 2g(P1 (deg(fi) − length(f −1 i (z))). K)) −Xz∈X For z ∈ X, we have deg(fi) − length(f −1 If z ∈ Zj − Zj−1 for some j ∈ {1, . . . , i}, then each preimage of z in Yi is fixed by a group of order pi−j+1, so length(f −1 (z)) = pj−1. This yields (11.8.1). (z)) = 0 unless z ∈ Zi. i Let f i : Y i → P1 k be the reduction of fi; then g(Y i) = g(Yi). Set notation as in the proof of Lemma 10.6. Since f i is Galois and only ramifies above 0, the Riemann-Hurwitz formula for f i (see [158, Corollary 3.4.14, Theorem 3.8.7]) can be written in the form i 2 − 2g(Y i) = deg(fi)(2 − 2g(P1 k)) − (# Gal(Fi/F )m − 1) ∞Xm=0 where Gal(Fi/F )m denotes the m-th subgroup of Gal(Fi/F ) in the lower numbering filtration. By identifying these subgroups as in the proof of Lemma 10.6, we obtain (11.8.2). By combining (11.8.1) and (11.8.2), we may solve for bi for i = 1, . . . , n in succession to obtain (11.8.3). (cid:3) We have the following explicit version of Theorem 5.5 in this setting. Theorem 11.9. Retain notation as in Definition 11.7. Let Γ = ΓX,Zn∪{∞} be the union of the paths from ∞ to the elements of Zn. Let Nn be the the convergence polygon of En. (a) The function Nn factors through the retraction of X an onto Γ, and is affine on each edge of Γ. (b) Let Γ∞ ⊂ Γ be the path from x1 to ∞. For each x ∈ Γ∞, s1(Nn(x)) = 0. (c) The measure χ(E) (more precisely, the measure χ(E, Zn) in the notation of Re- mark 7.15) is discrete, supported at x1, of total measure 1 − bn. (d) For each x ∈ Γ − Γ∞, for ~t the branch of x towards x1, ∂~ts1(Nn) = ℓ − 1 where ℓ is the length of the subset of Zn dominated by x. (e) For i = 1, . . . , n, for each x ∈ Zi − Zi−1, let Γx be the pendant edge of Γ terminating at x. For each y ∈ Γx, (11.9.1) s1(Nn(y)) =(cid:18)n − i + 1 + 1 p − 1(cid:19) log p. Proof. By Theorem 8.11(a), we have s1(Nn(x1)) = 0; by Theorem 5.9(a), Nn is identically zero on all x outside of the closed unit disc. In particular, we deduce (b). By Theorem 7.23(c) and Remark 8.16, we must have χx(En) ≤ 0 for all x ∈ X an. By Theorem 7.7 (as reformulated in Theorem 7.12), Lemma 7.6, and Lemma 11.8, we have χ(En) = χdR(X an, En) = χdR(U an) = 2 − length(Zn) = 1 − bn. By the previous paragraph plus Theorem 8.11(b), χx1(E) = 1 −bn; we thus deduce (c), which in turn immediately implies (a). Using Theorem 7.12 (which again applies in this situation thanks to Remark 8.16), we deduce (d). To obtain (e), in light of (a) we need only check 29 ZX an (11.9.1) for y in some neighborhood of x. This follows by noting that as in Example 4.3, the binomial series (1 + z)1/pn has radius of convergence p−n−1/(p−1). (cid:3) Remark 11.10. Retain notation as in Definition 11.7 and Theorem 11.9. The union of the paths from x1 to the points of Z then form a tree on which s1(Nn) restricts to a harmonic function which is affine on each edge of the tree, with prescribed values and slopes at the pendant vertices (i.e., x1 and the points of Z). However, the existence of such a function imposes strong combinatorial constraints on the relative positions of the points of Z; the resulting data are well-known in the literature on the Oort lifting problem, under the rubric of Hurwitz trees [24]. Similar data arise in [47, 159]. Remark 11.11. Conversely, suppose X = P1 K; Z equals {∞} plus a subset of the open unit disc; E is a rank 1 vector bundle with connection on U = X − Z; for each z ∈ Z, ∇ is regular at z with exponent in p−nZ; and for some ρ ∈ (0, 1), the restriction of E to the space z > ρ is isomorphic to En. The connection E ⊗pn has only removable singularities, so it can be shown to be trivial either by passing to C as in the proof of Lemma 7.6 and invoking complex GAGA (using the Riemann-Hilbert correspondence and the fact that P1,an is simply connected), or by applying the Dwork transfer theorem (Theorem 5.9(a)) to the disc z < σ for some σ > ρ. C From this, we may see that E arises as En for some data as in Definition 11.2, i.e., that E arises from a solution of the lifting problem. For example, to construct the finite map f : Y → X, we may work analytically over C (again as in the proof of Lemma 7.6), using complex GAGA to descend to the category of schemes. In the complex analytic situation, we may locally choose a generator v of E, then form Y from X by adjoining s1/p where sv⊗p is a nonzero horizontal section of E ⊗pn. By similar considerations, we see that f is Galois with group µpn and that E ∼= En. Remark 11.12. For a given π, it should be possible to construct a moduli space of solu- tions of the lifting problem in the category of rigid analytic spaces over K. Theorem 11.6 would then imply that this space is nonempty. Given this fact, it may be possible to derive additional results on the lifting problem, e.g., to resolve the case of dihedral groups. For p > 2, this amounts to showing that if τ anticommutes with the involution z 7→ −z, then the action of the involution z 7→ −z fixes some point of the moduli space. Appendix A. Convexity In this appendix, we give some additional technical arguments needed for the proofs of Theorem 6.6 and Theorem 7.23(c), which do not appear elsewhere in the literature. These arguments are not written in the same expository style as the rest of the main text; for instance, they assume much more familiarity with the author’s book [93]. Definition A.1. For I a subinterval of [0, +∞), let RI be the ring of rigid analytic functions on the space t ∈ I within the affine t-line over K, as in [98, Definition 3.1.1]. Hypothesis A.2. Throughout Appendix A, assume p = 0, and let (M, D) be a differential module of rank n over R[0,β). For I a subinterval of [0, β), write MI as shorthand for M ⊗R[0,β) RI. 30 Definition A.3. Let Dβ be the Berkovich disc t < β over K. For x ∈ Dβ, define the real numbers si(M, x) for i = 1, . . . , n as in [98, Definition 4.3.2]; note that they are invariant under extension of K [98, Lemma 4.3.3]. For r > − log β, define the functions gi(M, r) = − log si(M, xe−r ) Gi(M, r) = g1(M, r) + · · · + gi(M, r). We begin with a variant of Theorem 7.23. Lemma A.4. The right slope of Gn(M, r) at r = − log β equals − dimK H 1(M), provided that at least one of the two is finite. Proof. Apply [136, Theorem 3.5.2]. (cid:3) Lemma A.5. For ρ ∈ (0, β), let xρ be the generic point of the disc t ≤ ρ. Then for i = 1, . . . , n, the function ρ 7→ s1(M, xρ) · · · si(M, xρ) is nonincreasing in ρ. Proof. The case i = n is immediate from Lemma A.4. To deduce the case i < n, it suffices to work locally around some ρ0 = e−r0, and to check only those values of i for which gi(M, r0) > gi+1(M, r0). For λ ∈ K, let Mλ be the differential module obtained from M by adding λ to D; we may also view Mλ as the tensor product of M with the rank one differential module Nλ on a single generator v satisfying D(v) = λv. From the tensor product description, for j = 1, . . . , n we have (A.5.1) gj(Mλ, r) ≤ max{gj(M, r), g1(Nλ, r)}, gj(M, r) ≤ max{gj(Mλ, r), g1(N ∨ λ , r)}. In addition, by Example 4.1, we have g1(Nλ, r) = g1(N ∨ λ , r) = max(cid:26)− log β , 1 p − 1 log p + log λ(cid:27) independently of r. Choose an open subinterval I of (gi+1(M, r0), gi(M, r0)). After replacing K with a finite extension (which does not affect Gj(M, r)), we may choose λ in such a way that g1(Nλ, r) = g1(N ∨ λ , r) is equal to a constant value contained in I. We then have gi(M, r) > g1(Nλ, r) > gi+1(M, r) for all r in some neighborhood of r0. For such r, for j ≤ i we apply (A.5.1) to see that gj(Mλ, r) = gj(M, r). For j > i, (A.5.1) implies gj(Mλ, r) ≤ g1(Nλ, r), but we claim that in fact equality must hold. To wit, assume to the contrary that the inequality is strict; then the restriction of M to the disc t < e−gj (Mλ,r) would have a submodule isomorphic to N ∨ λ . This would mean that M has a local horizontal section whose exact radius of convergence is equal to e−g1(Nλ,r), which would imply that there exists some value of k for which gk(M, r) = g1(Nλ, r). However, this contradicts our choice of the interval I. To summarize, for r in a neighborhood of r0 we have gj(Mλ, r) =(gj(M, r) gj(Nλ, r) (j = 1, . . . , i) (j = i + 1, . . . , n). We thus deduce the original claim by applying the case i = n to Mλ. (cid:3) We now deduce Theorem 6.6. 31 Theorem A.6. For x, y ∈ Dβ such that x is the generic point of a disc containing y, for i = 1, . . . , n, we have In particular, the conclusion of Theorem 6.6 holds. s1(M, x) · · · si(M, x) ≤ s1(M, y) · · · si(M, y). Proof. This follows from Lemma A.5 thanks to the invariance of the si under base extension. (cid:3) We next proceed towards Theorem 7.23(c). Definition A.7. Define the convergence polygon N of M as in Definition 5.3, using the same disc at every point. Define the modified convergence polygon N ′ using maximal discs not containing 0. Lemma A.8. Assume p = 0. For i = 1, . . . , n, for x ∈ Dβ, we have ∆hi(N )x ≥ 0. Proof. By base extension, we may reduce to the case x = x1; we may also assume that the norm on K is nontrivial. As in the proof of Lemma A.5, we may further reduce to the case i = n. We may further reduce to the case where H 0(M[0,δ)) = 0 for all δ ∈ (1, β). Choose a nonempty set W of K-rational points of Dβ with the following properties. (a) For each w ∈ W , Uw,1 is a branch of P1 (b) The branches ~tw for w ∈ W are pairwise distinct. (c) Let ~t∞ be the branch of P1 K at x1, which we also denote by ~tw. K at x in the direction of ∞. Then for all branches ~t of P1 K at x1, we have ∂~t(N ) = 0 unless ~t = ~t∞ or ~t = ~tw for some w ∈ W . Let Rw,I be the ring of rigid analytic functions on the space z − w ∈ I, and put Mw,I = M ⊗R[0,β) Rw,I. Then by Lemma A.4, to prove the desired result, it suffices to check that for any γ ∈ (0, 1), δ ∈ (1, β) sufficiently close to 1, M[0,δ) → Mw∈W Mw,[0,γ) has dense image, so then does (A.8.1), proving the claim. (cid:3) Theorem A.9. The conclusion of Theorem 7.23(c) holds. Proof. By Theorem 7.23(b), we may assume x /∈ ΓX,Z; the claim thus reduces to Lemma A.8. (cid:3) 32 dimK H 1(Mw,[0,γ)). dimK H 1(M[0,δ)) ≥ Xw∈W H 1(M[0,δ)) → Mw∈W It would hence also suffice to prove surjectivity of the map (A.8.1) H 1(Mw,[0,γ)). Since p = 0, we may invoke Theorem 7.12 to see that H 1(Mw,[0,γ)) is a finite-dimensional K-vector space, and so is complete with respect to its natural topology as a K-vector space (i.e., the one induced by the supremum norm with respect to some basis). Since the map Mw,[0,γ) → H 1(Mw,[0,γ)) is a K-linear surjection from a Fr´echet space over K to a Banach space over K, the Banach open mapping theorem implies that this map is a quotient map of topological spaces. In particular, since the map Remark A.10. In the proof of Lemma A.8, to deduce the finite-dimensionality of H 1(Mw,[0,γ)), one may replace the invocation of Theorem 7.12 with the following argument. For δ1 ∈ (0, γ), δ2 ∈ (δ1, γ) sufficiently large, by [98, Lemma 3.7.6] we have dimK H 0(Mw,(δ1,γ)) = dimK H 1(Mw,(δ1,γ)), dimK H 0(Mw,(δ1,δ2)) = dimK H 1(Mw,(δ1,δ2)). By this calculation plus Mayer-Vietoris, the map H 1(Mw,[0,γ)) → H 1(Mw,[0,δ2)) has finite- dimensional kernel and cokernel. Since Rw,[0,γ) → Rw,[0,δ2) is a compact map of topological K-vector spaces, the same is true of Mw,[0,γ) → Mw,[0,δ2); we may thus apply the Schwartz- Cartan-Serre lemma [101, Satz 1.2] to conclude. Remark A.11. At this point, it is natural to consider what happens when p > 0. If we also add suitable hypotheses on p-adic non-Liouville exponents, then all of the preceding statements remain true (as in Remark 8.16). Absent such hypotheses, we cannot rely on finite-dimensionality of cohomology groups, but it may nonetheless be possible to adapt the proof of Theorem A.9 to the case p > 0 by establishing a relative version of Lemma A.4. To be precise, in the notation of the proof of Lemma A.8, one might hope to prove that H 1(M[0,δ)/Mw∈W Mw,[0,γ)) = 0 and to relate this vanishing directly to the Laplacian, bypassing the potential failure of finite-dimensionality for the individual cohomology groups. Appendix B. Thematic bibliography As promised in the introduction, we include an expansive but unannotated list of references related to key topics in the paper. For those topics discussed in [93], the chapter endnotes therein may be consulted for additional context. • Underlying structure of Berkovich spaces: [7], [8], [9, Chapter 1], [18], [19], [25], [54], [55], [56], [77], [79], [131], [132], [168]. See also the survey [172] in this volume. • Potential theory for Berkovich curves: [6, chapter by Baker], [9], [161]. • Formal structure of singular connections: [63, Chapter 4], [72], [83, §11], [93, Chap- ter 7], [94], [95], [103], [104], [109], [110], [119], [120], [157, Chapter 3], [167]. • Convergence of solutions of p-adic differential equations: [10], [13], [15], [16], [33], [36], [39], [45], [46], [65], [93, Chapters 9–11], [100], [107], [133], [134], [137], [138], [141], [176]. For a retrospective circa 2000, see also [35]. • Transfer principles and effective convergence bounds: [34], [38], [64], [66], [93, Chap- ters 9, 13, 18]. • Logarithmic growth of p-adic solutions: [3], [29], [30], [33], [59], [60], [93, Chapter 18], [111], [125], [127]. • Stokes phenomena for complex connections: [105], [106], [108], [121], [149], [150], [157, Chapters 7–9], [169]. • Index formulas for nonarchimedean differential equations: [37], [40], [41], [42], [44], [87], [88], [114], [143], [144], [146], [148], [136], [163], [176]. See also the thematic bibliography of [147], and [43] for a survey of [40, 41, 42, 44]. • Comparison between algebraic and complex analytic cohomology of connections: [4], [5], [53], [75], [116]. 33 • Comparison between algebraic and nonarchimedean analytic cohomology of connec- tions: [2], [4], [11], [12], [26], [27], [6, chapter by Kedlaya], [102]. • Decomposition theorems for nonarchimedean connections: [33], [37], [40], [41], [42], [44], [64], [93, Chapter 12], [98], [135], [143], [145]. • Monodromy for p-adic connections (Crew’s conjecture): [1], [51], [86], [90], [91], [92], [93, Chapters 20, 21], [96], [98], [117], [164], [165]. • p-adic Liouville numbers and p-adic exponents: [37], [40], [41], [42], [44], [46], [62], [93, Chapter 13], [98]. • Ramification of maps of Berkovich curves: [47], [68], [69], [80], [93, Chapter 19], [159]. • Measures of ramification and p-adic differential equations: [14], [28], [87], [89], [93, Chapter 19], [97], [112], [113], [114], [115], [126], [163], [173], [174], [175]. • Kummer theory in mixed characteristic (Kummer-Artin-Schreier-Witt theory): [73], [113], [118], [129], [140], [153], [154] (unpublished; see [118] instead), [155], [162], [171]. • Oort lifting problem: [20], [22], [24], [31], [32], [49], [70], [73], [74], [78, §9], [122], [123], [124], [128], [129], [130], [139], [151]. See also the lecture notes [23], the introductions to [124, 139], and the PhD thesis [166]. References [1] Y. Andr´e, Filtrations de type Hasse-Arf et monodromie p-adique, Invent. Math. 148 (2002), 285–317. [2] Y. Andr´e, Comparison theorems between algebraic and analytic de Rham cohomology, J. Th´eor. Nombres Bordeaux 16 (2004), 335–355. [3] Y. Andr´e, Dwork’s conjecture on the logarithmic growth of solutions of p-adic differential equations, Compos. Math. 144 (2008), 484–494. [4] Y. Andr´e and F. Baldassarri, De Rham Cohomology of Differential Modules on Algebraic Varieties, Progress in Math. 189, Birkhauser, Basel, 2001. [5] M.F. Atiyah and W.V.D. Hodge, Integrals of the second kind on an algebraic variety, Ann. Math. 62 (1955), 56–91. [6] M. Baker et al., p-adic Geometry: Lectures from the 2007 Arizona Winter School, Amer. Math. Soc., 2008. [7] M. Bayer, S. Payne, and J. Rabinoff, Nonarchimedean geometry, tropicalization, and metrics on curves, arXiv:1104.0320v2 (2012). [8] M. Baker, S. Payne, and J. Rabinoff, On the structure of non-Archimedean analytic curves, in Tropical and Non-Archimedean Geometry, Contemp. Math., 605, Amer. Math. Soc., Providence, RI, 2013, 93–121. [9] M. Baker and R. Rumely, Potential Theory and Dynamics on the Berkovich Projective Line, AMS Surveys and Monographs 159, Amer. Math. Soc., Providence, 2010. [10] F. Baldassarri, Differential modules and singular points of p-adic differential equations, Adv. Math. 44 (1982), 155–179. [11] F. Baldassarri, Comparaison entre la cohomologie alg´ebrique et la cohomologie p-adique rigide `a coef- ficients dans un module diff´erentiel, I. Cas des courbes, Inv. Math. 87 (1987), 83–99. [12] F. Baldassarri, Comparaison entre la cohomologie alg´ebrique et la cohomologie p-adique rigide `a co- efficients dans un module diff´erentiel, II. Cas des singularit´es r´eguli`eres `a plusieures variables, Math. An. 280 (1988), 417–439. [13] F. Baldassarri, Continuity of the radius of convergence of differential equations on p-adic analytic curves, Invent. Math. 182 (2010), 513–584. [14] F. Baldassarri, Radius of convergence of p-adic connections and the p-adic Rolle theorem, Milan J. Math. 81 (2013), 397–419. [15] F. Baldassarri and L. Di Vizio, Continuity of the radius of convergence of p-adic differential equations on Berkovich spaces, arXiv:07092008v1 (2007). 34 [16] F. Baldassarri and K.S. Kedlaya, Harmonic functions attached to meromorphic connections on non- archimedean curves, in preparation. [17] V. Berkovich, ´Etale cohomology for non-Archimedean analytic spaces, Publ. Math. IH ´ES 78 (1993), 5–161. [18] V. Berkovich, Smooth p-adic analytic spaces are locally contractible, Invent. Math. 137 (1999), 1–84. [19] V. Berkovich, Smooth p-adic analytic spaces are locally contractible, II, in Geometric Aspects of Dwork Theory, Vol. I, de Gruyter, Berlin, 2004, 293–370. [20] J. Bertin, Obstructions locales, au rel`evement de revetements galoisiens de courbes lisses, C.R. Acad. Sci. Paris S´er. I Math. 326 (1998), 55–58. [21] F. Beukers, On Dwork’s accessory parameter problem, Math. Z. 241 (2002), 425–444. [22] I.I. Bouw and S. Wewers, The local lifting problem for dihedral groups, Duke Math. J. 134 (2006), 421–452. [23] I. Bouw and S. Wewers, Group actions on curves and the lifting problem, course notes for the 2012 Arizona Winter School, available at http://math.arizona.edu/~swc/aws/2012/. [24] L.H. Brewis and S. Wewers, Artin characters, Hurwitz trees, and the lifting problem, Math. Ann. 345 (2009), 711–730. [25] A. Chambert-Loir and A. Ducros, Formes diff´erentielles r´eelles et courants sur les espaces de Berkovich, preprint available at http://www.math.jussieu.fr/~ducros/. [26] B. Chiarellotto, Sur le th´eor`eme de comparaison entre cohomologies de de Rham alg´ebrique et p-adique rigide, Ann. Inst. Fourier (Grenoble) 38 (1988), 1–15. [27] B. Chiarellotto, A comparison theorem in p-adic cohomology, Ann. Mat. Pura Appl. 153 (1988), 115– 131. [28] B. Chiarellotto and A. Pulita, Arithmetic and differential Swan conductors of rank 1 representations with finite local monodromy, Amer. J. Math. 131 (2009), 1743–1794. [29] B. Chiarellotto and N. Tsuzuki, Logarithmic growth and Frobenius filtrations for solutions of p-adic differential equations, J. Inst. Math. Jussieu 8 (2009), 465–505. [30] B. Chiarellotto and N. Tsuzuki, Log-growth filtration and Frobenius slope filtration of F -isocrystals at the generic and special points, Doc. Math. 16 (2011), 33–69. [31] T. Chinburg, R. Guralnick, and D. Harbater, Oort groups and local lifting problems, Compos. Math. 144 (2008), 849–866. [32] T. Chinburg, R. Guralnick, and D. Harbater, The local lifting problem for actions of finite groups on curves, Ann. Scient. ´Ec. Norm. Sup. 44 (2011), 537–605. [33] G. Christol, Modules Diff´erentielles et ´Equations Diff´erentielles p-adiques, Queen’s University, Kingston, 1983. [34] G. Christol, Un th´eor`eme de transfert pour les disques singuliers r´eguliers, Cohomologie p-adique, Ast´erisque 119–120 (1984), 151–168. [35] G. Christol, Thirty years later, in Geometric Aspects of Dwork Theory, Vol. I, de Gruyter, Berlin, 2004, 419–436. [36] G. Christol, The radius of convergence function for first order differential equations, in Advances in non-Archimedean analysis, Contemp. Math. 551, Amer. Math. Soc., Providence, RI, 2011, 71–89. [37] G. Christol, Le Th´eor`eme de Turrittin p-adique, in preparation; version of 11 Jun 2011 downloaded from http://www.math.jussieu.fr/~christol. [38] G. Christol and B. Dwork, Effective p-adic bounds at regular singular points, Duke Math. J. 62 (1991), 689–720. [39] G. Christol and B. Dwork, Modules differentiels sur les couronnes, Ann. Inst. Fourier (Grenoble) 44 (1994), 663–701. [40] G. Christol and Z. Mebkhout, Sur le th´eor`eme de l’indice des ´equations diff´erentielles, Annales Inst. Fourier 43 (1993), 1545–1574. [41] G. Christol and Z. Mebkhout, Sur le th´eor`eme de l’indice des ´equations diff´erentielles, II, Annals of Math. 146 (1997), 345–410. [42] G. Christol and Z. Mebkhout, Sur le th´eor`eme de l’indice des ´equations diff´erentielles, III, Annals of Math. 151 (2000), 385–457. 35 [43] G. Christol and Z. Mebkhout, p-adic differential equations, in Algebra and number theory (Fez) Lect. Notes Pure Appl. Math., 208, Dekker, New York, 2000, 105–116. [44] G. Christol and Z. Mebkhout, Sur le th´eor`eme de l’indice des ´equations diff´erentielles, IV, Invent. Math. 143 (2001), 629–672. [45] G. Christol and S. Remmal, Irregular p-adic linear differential equations, in Algebra and number theory (Fez) Lect. Notes Pure Appl. Math., 208, Dekker, New York, 2000, 195–206. [46] D. Clark, A note on the p-adic convergence of solutions of linear differential equations, Proc. Amer. Math. Soc. 17 (1966), 262–269. [47] A. Cohen, M. Temkin, and D. Trushin, Morphisms of Berkovich curves and the different function, arXiv:1408.2949v2 (2014). [48] Brian Conrad, Relative ampleness in rigid geometry, Annales de l’Institut Fourier (Grenoble) 56 (2006), 1049–1126. [49] S. Corry, Galois covers of the open p-adic disc, Manuscripta Math. 131 (2010), 43–61. [50] R. Crew, F -isocrystals and p-adic representations, in Algebraic Geometry—Bowdoin 1985, Part 2, Proc. Sympos. Pure Math. 46.2, Amer. Math. Soc, 1987, 111–138. [51] R. Crew, Finiteness theorems for the cohomology of an overconvergent isocrystal on a curve, Ann. Scient. ´Ec. Norm. Sup. 31 (1998), 717–763. [52] R. Crew, Canonical extensions, irregularities, and the Swan conductor, Math. Ann. 316 (2000), 19–37. [53] P. Deligne, ´Equations Diff´erentielles `a Points Singuliers R´eguliers, Lecture Notes in Math. 163, Springer-Verlag, Berlin, 1970. [54] A. Ducros, Les espaces de Berkovich sont mod´er´es, d’apr`es E. Hrushovski and F. Loeser, S´eminaire Bourbaki expos´e 1056, Ast´erisque 352 (2013). [55] A. Ducros, About Hrushovski and Loeser’s work on the homotopy type of Berkovich spaces, arXiv:1309.0340v1 (2013). [56] A. Ducros, La Structure des Courbes Analytiques, in preparation; version of 12 Feb 2014 downloaded from http://www.math.jussieu.fr/~ducros/livre.html. [57] B. Dwork, On the zeta function of a hypersurface, II, Annals of Math. 80 (1964), 227–299. [58] B. Dwork, p-adic cycles, Publ. Math. IH ´ES 37 (1969), 27–115. [59] B. Dwork, On p-adic differential equations, II: The p-adic asymptotic behavior of solutions of linear differential equations with rational function coefficients, Annals of Math. 98 (1973), 366–376. [60] B. Dwork, On p-adic differential equations, III: On p-adically bounded solutions of ordinary linear differential equations with rational function coefficients, Invent. Math. 20 (1973), 35–45. [61] B. Dwork, Bessel functions as p-adic functions of the argument, Duke Math. J. 41 (1974), 711–738. [62] B. Dwork, On exponents of p-adic differential equations, J. reine angew. Math. 484 (1997), 85–126. [63] B. Dwork, G. Gerotto, and F. Sullivan, An Introduction to G-Functions, Ann. Math. Studies 133, Princeton Univ. Press, Princeton, 1994. [64] B. Dwork and P. Robba, On ordinary linear p-adic differential equations, Trans. Amer. Math. Soc. 231 (1977), 1–46. [65] B. Dwork and P. Robba, On natural radii of p-adic convergence, Trans. Amer. Math. Soc. 256 (1979), 199–213. [66] B. Dwork and P. Robba, Effective p-adic bounds for solutions of homogeneous linear differential equa- tions, Trans. Amer. Math. Soc. 259 (1980), 559–577. [67] D. Eisenbud, Commutative Algebra with a View Toward Algebraic Geometry, Graduate Texts in Math. 150, Springer, New York, 1995. [68] X. Faber, Topology and geometry of the Berkovich ramification locus for rational functions, I, Manuscripta Math. 142 (2013), 439–474. [69] X. Faber, Topology and geometry of the Berkovich ramification locus for rational functions, II, Math. Ann. 356 (2013), 819–844. [70] M. Garuti, Prolongement de revetements galoisiens en g´eom´etrie rigide, Compos. Math. 104 (1996), 305–331. [71] M. Garuti, Linear systems attached to cyclic inertia, in Arithmetic Fundamental Groups and Noncom- mutative Algebra (Berkeley, CA, 1999), Proc. Symp. Pure Math. 70, Amer. Math. Soc., Providence, RI, 2002, 377–386. 36 [72] R. G´erard and A.M. Levelt, Invariants mesurant l’irr´egularit´e en un point singulier des syst`emes d’´equations diff´erentielles lin´eaires, Ann. Inst. Fourier (Grenoble) 23 (1973), 157–195. [73] B. Green and M. Matignon, Liftings of galois covers of smooth curves, Compos. Math. 113 (1998), 237–272. [74] B. Green and M. Matignon, Order p automorphisms of the open disc of a p-adic field, J. Amer. Math. Soc. 12 (1999), 269–303. [75] A. Grothendieck, On the de Rham cohomology of algebraic varieties, Publ. Math. IH ´ES 29 (1966), 95–103. [76] A. Grothendieck et al., SGA 1: Revetements ´Etales et Groupe Fondamental, corrected edition, Docu- ments Math´ematiques 3, Soc. Math. France, Paris, 2003. [77] W. Gubler, J. Rabinoff, and A. Werner, Skeletons and tropicalizations, arXiv:1404.7044v1 (2014). [78] D. Harbater, A. Obus, R. Pries, and K. Stevenson, Abhyankar’s conjectures in Galois theory: Current status and future directions, arXiv:1408.0859v1 (2014). [79] E. Hrushovski and F. Loeser, Non-archimedean tame topology and stably dominated types, Annals of Mathematics Studies, Princeton Univ. Press, Princeton, to appear. [80] R. Huber, Swan representations associated with rigid analytic curves, J. reine angew. Math. 537 (2001), 165–234. [81] L. Illusie, Complexe de de Rham–Witt et cohomologie crystalline, Ann. Sci. ´Ec. Norm. Sup. 12 (1979), 501–661. [82] K. Kato, Swan conductors for characters of degree one in the imperfect residue field case, in Algebraic K-theory and algebraic number theory (Honolulu, HI, 1987), 101–131, Contemp. Math. 83, Amer. Math. Soc., Providence, 1989. [83] N.M. Katz, Nilpotent connections and the monodromy theorem: applications of a result of Turrittin, Publ. Math. IH ´ES 39 (1970), 175–232. [84] N.M. Katz, Local-to-global extensions of representations of fundamental groups, Ann. Inst. Fourier 36 (1986), 69–106. [85] K.S. Kedlaya, Semistable reduction for overconvergent F -isocrystals on a curve, Math. Res. Lett. 10 (2003), 151–159. [86] K.S. Kedlaya, A p-adic local monodromy theorem, Annals of Math. 160 (2004), 93–184. [87] K.S. Kedlaya, Local monodromy of p-adic differential equations: an overview, Int. J. Number Theory 1 (2005), 109–154; errata posted at http://kskedlaya.org/papers. [88] K.S. Kedlaya, Fourier transforms and p-adic “Weil II”, Compos. Math. 142 (2006), 1426–1450. [89] K.S. Kedlaya, Swan conductors for p-adic differential modules, I: A local construction, Alg. Number Theory 1 (2007), 269–300. [90] K.S. Kedlaya, Semistable reduction for overconvergent F -isocrystals, I: Unipotence and logarithmic extensions, Compos. Math. 143 (2007), 1164–1212. [91] K.S. Kedlaya, Semistable reduction for overconvergent F -isocrystals, II: A valuation-theoretic ap- proach, Compos. Math. 144 (2008), 657–672. [92] K.S. Kedlaya, Semistable reduction for overconvergent F -isocrystals, III: Local semistable reduction at monomial valuations, Compos. Math. 145 (2009), 143–172. [93] K.S. Kedlaya, p-adic Differential Equations, Cambridge Univ. Press, Cambridge, 2010; errata posted at http://kskedlaya.org/papers. [94] K.S. Kedlaya, Good formal structures for flat meromorphic connections, I: Surfaces, Duke Math. J. 154 (2010), 343–418; erratum, ibid. 161 (2012), 733–734. [95] K.S. Kedlaya, Good formal structures for flat meromorphic connections, II: Excellent schemes, J. Amer. Math. Soc. 24 (2011), 183–229. [96] K.S. Kedlaya, Semistable reduction for overconvergent F -isocrystals, IV: Local semistable reduction at nonmonomial valuations, Compos. Math. 147 (2011), 467–523. [97] K.S. Kedlaya, Swan conductors for p-adic differential modules, II: Global variation, J. Inst. Math. Jussieu 10 (2011), 191–224. [98] K.S. Kedlaya, Local and global structure of connections on meromorphic curves, Compos. Math. 151 (2015), 1096–1156. 37 [99] K.S. Kedlaya and J. Tuitman, Effective bounds for Frobenius structures, Rend. Sem. Mat. Padova 128 (2012), 7–16. [100] K.S. Kedlaya and L. Xiao, Differential modules on p-adic polyannuli, J. Inst. Math. Jussieu 9 (2010), 155–201; erratum, ibid. 9 (2010), 669–671. [101] R. Kiehl, Der Endlichkeitsatz fur eigentliche Abbildungen in der nichtarchimedischen Funktionenthe- orie, Invent. Math. 2 (1967), 191–214. [102] R. Kiehl, Die de Rham Kohomologie algebraischer Mannigfaltigkeiten uber einem bewerteten Korper, Publ. Math. IH ´ES 33 (1967), 5–20. [103] A.H.M. Levelt, Jordan decomposition for a class of singular differential operators, Ark. Mat. 13 (1975), 1–27. [104] A.H.M. Levelt and A.R.P. van den Essen, Irregular Singularities in Several Variables, Mem. Amer. Math. Soc. vol. 40, no. 270, Amer. Math. Soc., Providence, 1982. [105] M. Loday-Richaud, Stokes phenomenon, multisummability and differential Galois groups, Ann. Inst. Fourier (Grenoble) 44 (1994), 849–906. [106] M. Loday-Richaud, Solutions formelles des syst`emes diff´erentiels lin´eaires m´eromorphes et sommation, Expo. Math. 13 (1995), 116–162. [107] E. Lutz, Sur l’´equation y2 = x3 + Ax + B sur les corps p-adiques, J. reine angew. Math. 177 (1937), 238–247. [108] H. Majima, Asymptotic Analysis for Integrable Connections with Irregular Singular Points, Lect. Notes in Math. 1075, Springer-Verlag, 1984. [109] B. Malgrange, Sur les points singuliers des ´equations diff´erentielles, Enseign. Math. 20 (1974), 147–176. [110] B. Malgrange, Connexions m´eromorphes 2: Le r´eseau canonique, Invent. Math. 124 (1996), 367–387. [111] S. Manjra, A note on non-Robba p-adic differential equations, Proc. Japan Acad. Ser. A 87 (2011), 40–43. [112] A. Marmora, Irr´egularit´e et conducteur de Swan p-adiques, Doc. Math. 9 (2004), 413–433. [113] S. Matsuda, Local indices of p-adic differential operators corresponding to Artin-Schreier-Witt cover- ings, Duke Math. J. 77 (1995), 607–625. [114] S. Matsuda, Katz correspondence for quasi-unipotent overconvergent isocrystals, Compos. Math. 134 (2002), 1–34. [115] S. Matsuda, Conjecture on Abbes-Saito filtration and Christol-Mebkhout filtration, in Geometric As- pects of Dwork Theory, Vol. II, de Gruyter, Berlin, 2004, 845–856. [116] Z. Mebkhout, Le th´eor`eme de comparaison entre cohomologies de de Rham d’une vari´et´e alg´ebrique complexe et le th´eor`eme d’existence de Riemann, Publ. Math. IH ´ES 69 (1989), 47–89. [117] Z. Mebkhout, Analogue p-adique du th´eor`eme de Turrittin et le th´eor`eme de la monodromie p-adique, Invent. Math. 148 (2002), 319–351. [118] A. M´ezard, M. Romagny, and D. Tossici, Sekiguchi-Suwa theory revisited, J. Th´eor. Nombres Bordeaux 26 (2014), 163-200. [119] T. Mochizuki, Good formal structure for meromorphic flat connections on smooth projective surfaces, in Algebraic analysis and around, Advanced Studies in Pure Math. 54, Math. Soc. Japan, 2009, 223– 253. [120] T. Mochizuki, Wild harmonic bundles and wild pure twistor D-modules, Ast´erisque 340 (2011). [121] T. Mochizuki, The Stokes structure of a good meromorphic flat bundle, J. Inst. Math. Jussieu 10 (2011), 675–712. [122] A. Obus, The (local) lifting problem for curves, in Galois-Teichmuller Theory and Arithmetic Geom- etry, Adv. Stud. Pure Math. 63, Math. Soc. Japan, Tokyo, 2012, 359–412. [123] A. Obus, A generalization of the Oort conjecture, arXiv:1502.07623v1 (2015). [124] A. Obus and S. Wewers, Cyclic extensions and the local lifting problem, Ann. of Math. 180 (2014), 233–284. [125] S. Ohkubo, A note on logarithmic growth Newton polygons of p-adic differential equations, Int. Math. Res. Notices 2014, article ID rnu017. [126] S. Ohkubo, On differential modules associated to de Rham representations in the imperfect residue field case, arXiv:1307.8110v3 (2014). 38 [127] S. Ohkubo, On the rationality and continuity of logarithmic growth filtration of solutions of p-adic differential equations, arXiv:1502.03804v1 (2015). [128] F. Oort, Lifting algebraic curves, abelian varieties, and their endomorphisms to characteristic zero, in Algebraic Geometry, Bowdoin, 1985 (Brunswick, Maine, 1985), Proc. Symp. Pure Math. 46, Amer. Math. Soc., Providence, RI, 1987, 165–195. [129] F. Oort, T. Sekiguchi, and N. Suwa, On the deformation of Artin-Schreier to Kummer, Ann. Scient. ´Ec. Norm. Sup. 22 (1989), 345–375. [130] G. Pagot, Fp-espaces vectoriels de formes diff´erentielles logarithmiques sur la droite projective, J. Number Theory 97 (2002), 58–94. [131] S. Payne, Analytification is the limit of all tropicalizations, Math. Res. Lett. 16 (2009), 543–556. [132] J. Poineau, Les espaces de Berkovich sont ang´eliques, Bull. Soc. Math. France 141 (2013), 267–297. [133] J. Poineau and A. Pulita, The convergence Newton polygon of a p-adic differential equation II: Con- tinuity and finiteness on Berkovich curves, arXiv:1209.3663v1 (2012). [134] J. Poineau and A. Pulita, Continuity and finiteness of the radius of convergence of a p-adic differential equation via potential theory, arXiv:1209.6276v1 (2012). [135] J. Poineau and A. Pulita, The convergence Newton polygon of a p-adic differential equation III: global decomposition and controlling graphs, arXiv:1308.0859v1 (2013). [136] J. Poineau and A. Pulita, The convergence Newton polygon of a p-adic differential equation IV: local and global index theorems, arXiv:1309.3940v1 (2013). [137] E. Pons, Polygone de convergence d’un module diff´erentiel p-adique, C.R. Acad. Sci. Paris S´er. I Math. 327 (1998), 77–80. [138] E. Pons, E. Modules diff´erentiels non solubles. Rayons de convergence et indices, Rend. Sem. Mat. Univ. Padova 103 (2000), 21–45. [139] F. Pop, Lifting of curves: the Oort conjecture, Ann. of Math. 180 (2014), 285–322. [140] A. Pulita, Rank one solvable p-adic differential equations and finite Abelian characters via Lubin–Tate groups, Math. Ann. 337 (2007), 489–555. [141] A. Pulita, The convergence Newton polygon of a p-adic differential equation I: Affinoid domains of the Berkovich affine line, arXiv:1208.5850v4 (2014). [142] A. Pulita, ´Equations diff´erentielles p-adiques, m´emoire d’habilitation `a diriger des recherches, Univ. Montpellier, 2014. [143] P. Robba, On the index of p-adic differential operators, I, Annals of Math. 101 (1975), 280–316. [144] P. Robba, On the index of p-adic differential operators, II, Duke Math. J. 43 (1976), 19–31. [145] P. Robba, Lemmes de Hensel pour les op´erateurs diff´erentiels. Application a la r´eduction formelle des ´equations diff´erentielles, Ens. Math. 26 (1980), 279–311. [146] P. Robba, On the index of p-adic differential operators, III, Application to twisted exponential sums, Ast´erisque 119–120 (1984), 191–266. [147] P. Robba, Une introduction naıve aux cohomologies de Dwork, M´em. Soc. Math. France 23 (1986), 61–105. [148] P. Robba, Indice d’un operateur diff´erentiel p-adique, IV: Cas des syst`emes. Mesure de l’irr´egularit´e dans un disque, Ann. Inst. Fourier, Grenoble 35 (1985), 13–55. [149] C. Sabbah, ´Equations diff´erentielles `a points singuliers irr´eguliers et phenomene de Stokes en dimension 2, Ast´erisque 263 (2000). [150] C. Sabbah, Introduction to Stokes Structures, Lecture Notes in Math. 2060, Springer-Verlag, Berlin, 2013. [151] M. Saıdi, Fake liftings of Galois covers between smooth curves, in Galois-Teichmuller Theory and Arithmetic Geometry, Adv. Stud. Pure Math. 63, Math. Soc. Japan, Tokyo, 2012, 457–501. [152] P. Schneider and J. Teitelbaum, Algebras of p-adic distributions and admissible representations, Invent. Math. 153 (2003), 145–196. [153] T. Sekiguchi and N. Suwa, Th´eorie de Kummer-Artin-Schreier et applications, J. Th´eorie Nombres Bordeaux 7 (1995), 177–189. [154] T. Sekiguchi and N. Suwa, On the unified Kummer-Artin-Schreier-Witt theory, Pr´epublications du laboratoire de Math´ematiques Pures de Bordeaux, no. 111 (1999). 39 [155] T. Sekiguchi and N. Suwa, A note on extensions of algebraic and formal groups. IV. Kummer-Artin- Schreier-Witt theory of degree p2, Tohoku Math. J. 53 (2001), 203–240. [156] J.-P. Serre, Local Fields, Graduate Texts in Math. 67, Springer-Verlag, New York, 1979. [157] M. Singer and M. van der Put, Galois Theory of Linear Differential Equations, Grundlehren der math. Wissenschaften 328, Springer-Verlag, Berlin, 2003. [158] H. Stichtenoth, Algebraic Function Fields and Codes, Graduate Texts in Math. 254, Springer-Verlag, Berlin, 2009. [159] M. Temkin, Metric uniformization of morphisms of Berkovich curves, arXiv:1410.6892v1 (2014). [160] L. Thomas, Ramification groups in Artin-Schreier-Witt extensions, J. Th´eor. Nombres Bordeaux 17 (2005), 689–720. [161] A. Thuillier, Th´eorie du potentiel sur les courbes en g´eom´etrie analytique non archim´edienne, PhD thesis, Universit´e de Rennes, 2005. [162] D. Tossici, Models of µp2,K over a discrete valuation ring (with an appendix by X. Caruso), J. Algebra 323 (2010), 1908–1957. [163] N. Tsuzuki, The local index and the Swan conductor, Comp. Math. 111 (1998), 245–288. [164] N. Tsuzuki, Finite local monodromy of overconvergent unit-root F -isocrystals on a curve, Amer. J. Math. 120 (1998), 1165–1190. [165] N. Tsuzuki, Morphisms of F -isocrystals and the finite monodromy theorem for unit-root F -isocrystals, Duke Math. J. 111 (2002), 385–419. [166] D. Turchetti, Contributions to arithmetic geometry in mixed characteristic: Lifting covers of curves, non-Archimedean geometry and the ℓ-modular Weil representation, PhD thesis, Univ. de Versailles St.-Quentin, 2014; available online at http://webusers.imj-prg.fr/~daniele.turchetti/. [167] H.L. Turrittin, Convergent solutions of ordinary linear homogeneous differential equations in the neigh- borhood of an irregular singular point, Acta Math. 93 (1955), 27–66. [168] I. Tyomkin, Tropical geometry and correspondence theorems via toric stacks, Math. Ann. 353 (2012), 945–995. [169] V.S. Varadarajan, Linear meromorphic differential equations: a modern point of view, Bull. Amer. Math. Soc. 33 (1996), 1–42. [170] W. Wasow, Asymptotic Expansions for Ordinary Differential Equations, reprint of the 1976 edition, Dover, New York, 1987. [171] W. Waterhouse, A unified Kummer-Artin-Schreier sequence, Math. Ann. 277 (1987), 447–451. [172] A. Werner, Analytification and tropicalization over non-archimedean fields, Simons Symposium 2015 lecture notes available at http://users.math.yale.edu/~sp547/SimonsSymposium2015.html. [173] L. Xiao, On ramification filtrations and p-adic differential equations, I: equal characteristic case, Alg. and Number Theory 4 (2010), 969–1027. [174] L. Xiao, On ramification filtrations and p-adic differential equations, II: mixed characteristic case, Compos. Math. 148 (2012), 415–463. [175] L. Xiao, On the refined ramification filtrations in the equal characteristic case, Alg. and Num. Theory 6 (2012), 1579–1667. [176] P.T. Young, Radii of convergence and index for p-adic differential operators, Trans. Amer. Math. Soc. 333 (1992), 769–785. 40
1208.4584	2	1208	2012-12-05T07:45:01	Monodromies at infinity of non-tame polynomials	[ "math.AG", "math.CV" ]	We consider the monodromy at infinity and the monodromies around the bifurcation points of polynomial functions $f : \CC^n \longrightarrow \CC$ which are not tame and might have non-isolated singularities. Our description of their Jordan blocks in terms of the Newton polyhedra and the motivic Milnor fibers relies on two new issues: the non-atypical eigenvalues of the monodromies and the corresponding concentration results for their generalized eigenspaces.	math.AG	math	MONODROMIES AT INFINITY OF NON-TAME POLYNOMIALS KIYOSHI TAKEUCHI AND MIHAI TIB AR Abstract. We consider the monodromy at infinity and the monodromies around the bifurcation points of polynomial functions f : Cn −→ C which are not tame and might have non-isolated singularities. Our description of their Jordan blocks in terms of the Newton polyhedra and the motivic Milnor fibers relies on two new issues: the non- atypical eigenvalues of the monodromies and the corresponding concentration results for their generalized eigenspaces. 1. Introduction For a polynomial map f : Cn −→ C, it is well-known that there exists a finite subset B ⊂ C such that the restriction (1.1) of f is a locally trivial fibration. We denote by Bf the smallest subset B ⊂ C satisfying this condition. We call the elements of Bf bifurcation points of f . Let CR = {x ∈ C x = R} (R ≫ 0) be a sufficiently large circle in C such that Bf ⊂ {x ∈ C x < R}. Then by restricting the locally trivial fibration Cn\f −1(Bf ) −→ C\Bf to CR we obtain a geometric monodromy automorphism Φ∞ Cn \ f −1(B) −→ C \ B f : f −1(R) ∼−→ f −1(R) and the linear maps : H j(f −1(R); C) ∼−→ H j(f −1(R); C) (j = 0, 1, . . .) (1.2) Φ∞ j associated to it, where the orientation of CR is taken to be counter-clockwise as usual. We call Φ∞ j 's the (cohomological) monodromies at infinity of f . In the last few decades many mathematicians studied Φ∞ j 's from various points of view. In particular in [1] Broughton proved that if f is tame at infinity (see Definition 2.1) one has the concentration H j(f −1(R); C) = 0 (j 6= 0, n − 1) (1.3) for the generic fiber f −1(R) (R ≫ 0) of f . In this case Libgober-Sperber [13] obtained a beautiful formula which expresses the semisimple part (i.e. the eigenvalues) of Φ∞ n−1 in terms of the Newton polyhedron at infinity of f (see [15] for its generalizations). Recently in [17] (see also [6]) the first author proved formulae for its nilpotent part, i.e. its Jordan normal form, by using the motivic Milnor fiber at infinity of f . However, the methods of the above cited papers do not apply beyond the tame case since one cannot insure the concentration of the cohomology (1.3) for non-tame polynomials. In what concerns the evaluation of the bifurcation set Bf , non-tame polynomials were studied by N´emethi- Zaharia [21], Zaharia [33] and many other mathematicians. We overcome here the above problem by showing that the desired cohomological con- for "good" eigenvalues. Namely if centration holds for the generalized eigenspaces of Φ∞ j 2010 Mathematics Subject Classification. 14E18, 14M25, 32C38, 32S35, 32S40. 1 2 KIYOSHI TAKEUCHI AND MIHAI TIB AR we avoid some "bad" eigenvalues associated to f , we can successfully generalize the results in [17] to non-tame polynomials and completely determine the Jordan normal forms of Φ∞ n−1. In order to explain our results more precisely, let us recall some basic definitions. + av 6= 0} in Rn the Newton polytope of f and denote it by NP (f ). After Kushnirenko [12], the + of {0}∪NP (f ) in Rn is called the Newton polyhedron at infinity convex hull Γ∞(f ) ⊂ Rn of f . avxv (av ∈ C) we call the convex hull of suppf = {v ∈ Rn For f (x) =Pv∈Zn + Definition 1.1. We say that f is convenient if Γ∞(f ) intersects the positive part of the i-th axis of Rn for any 1 ≤ i ≤ n. If f is convenient and non-degenerate at infinity (for the definition, see Definition 2.3), then by a result of Broughton [1] it is tame at infinity. However here we do not assume that f is convenient. In Definition 2.12 by using the Newton polyhedron at infinity Γ∞(f ) we define a finite subset Af ⊂ C of "bad" eigenvalues which we call atypical engenvalues of f . Then the following result plays a key role in this paper. For λ ∈ C and j ∈ Z let H j(f −1(R); C)λ ⊂ H j(f −1(R); C) be the generalized eigenspace for the eigenvalue λ of the monodromy at infinity Φ∞ j . Theorem 1.2. Let f ∈ C[x1, . . . , xn] be a non-convenient polynomial such that dimΓ∞(f ) = n. Assume that f is non-degenerate at infinity. Then for any non-atypical eigenvalue λ /∈ Af of f we have the concentration H j(f −1(R); C)λ ≃ 0 (1.4) for the generic fiber f −1(R) ⊂ Cn (R ≫ 0) of f . (j 6= n − 1) This theorem allows non-isolated singularities of f and also the situation where the fibers may have cohomological perturbation "at infinity". Indeed, some of its atypical fibers f −1(b) (b ∈ Bf ) e.g. f −1(0) have non-isolated singularities in general. This is the main reason why the monodromies of non-tame polynomials could not be studied In the "tame" case one has only isolated singularities in Cn and successfully before. either vanishing cycles at infinity do not occur at all or they occur at isolated points only (in the sense of [27], [31]), and then the concentration of cohomology (1.3) follows. Theorem 1.2 will be proved by refining the proof of Sabbah's theorem [26, Theorem 13.1] in our situation. More precisely we construct a nice compactification fXΣ of Cn and study the "horizontal" divisors at infinity for f in fXΣ \ Cn very precisely to prove the concentration. With Theorem 1.2 at hand, by using the results in [17, Section 2] we can easily prove the generalizations of [17, Theorems 5.9, 5.14 and 5.16] to non- tame polynomials and completely determine the λ-part of the Jordan normal form of Φ∞ n−1 for any λ /∈ Af . Let us explain one of our results, which generalizes [17, Theorem 5.9]. Denote by Cone∞(f ) the closed cone R+Γ∞(f ) ⊂ Rn + generated by Γ∞(f ). Let q1, . . . , ql (resp. γ1, . . . , γl′) be the 0-dimensional (resp. 1-dimensional) faces of Γ∞(f ) such that qi ∈ Int(Cone∞(f )) (resp. the relative interior rel.int(γi) of γi is contained in Int(Cone∞(f ))). For each qi (resp. γi), denote by di > 0 (resp. ei > 0) its "lattice distance" from the origin 0 ∈ Rn (see Section 2 for the precise definition). For 1 ≤ i ≤ l′, let ∆i be the convex hull of {0} ⊔ γi in Rn. Then for λ 6= 1 and 1 ≤ i ≤ l′ such that MONODROMIES AT INFINITY OF NON-TAME POLYNOMIALS 3 λei = 1 we set (1.5) n(λ)i = ♯{v ∈ Zn ∩ rel.int(∆i) ht(v, γi) = k} + ♯{v ∈ Zn ∩ rel.int(∆i) ht(v, γi) = ei − k}, ei (we set ζd := exp(2π√−1/d) ∈ where k is the minimal positive integer satisfying λ = ζ k C) and for v ∈ Zn ∩ rel.int(∆i) we denote by ht(v, γi) the lattice height of v from the base γi of ∆i. Then we have the following extension of [17, Theorem 5.9] from tame to non-tame polynomials. Theorem 1.3. Assume that dimΓ∞(f ) = n and f is non-degenerate at infinity. Then for any λ /∈ Af we have (1) The number of the Jordan blocks for the eigenvalue λ with the maximal possible size n−1 : H n−1(f −1(R); C) ∼−→ H n−1(f −1(R); C) (R ≫ 0) is equal to ♯{qi λdi = (2) The number of the Jordan blocks for the eigenvalue λ with the second maximal n in Φ∞ 1}. possible size n − 1 in Φ∞ n−1 is equal toPi : λei =1 n(λ)i. We can treat in a similar manner the monodromies at bifurcation points of f . Let Bf ∩ {x ∈ C x − b ≤ ε} = {b} j : H j(f −1(b + ε); C) ∼−→ H j(f −1(b + ε); C) (j = 0, 1, . . .) Φb b ∈ Bf be such a bifurcation point. Choose sufficiently small ε > 0 such that (1.6) and set Cε(b) = {x ∈ C x − b = ε} ⊂ C. Then we obtain a locally trivial fibration f −1(Cε(b)) −→ Cε(b) over the small circle Cε(b) ⊂ C and the monodromy automorphisms (1.7) around the atypical fiber f −1(b) ⊂ Cn associated to it. In Section 5 we apply our methods to the Jordan normal forms of Φb j's. If f is not convenient, then for some b ∈ Bf the atypical fiber f −1(b) ⊂ Cn may have "singularities at infinity". Even in such cases, we can define a finite subset A◦ f,b ⊂ C of "bad" eigenvalues for b ∈ Bf and completely determine the λ-part of the Jordan normal form of Φb f,b. In fact we obtain these results more generally, for polynomial maps f : U −→ C of affine algebraic varieties U. See Section 5 for the details. n−1 for any λ /∈ A◦ 2. Preliminaries Let f : Cn −→ C be the polynomial map in Section 1. To study its monodromies at j , we often impose the following natural condition. infinity Φ∞ Definition 2.1 ([12]). Let ∂f : Cn −→ Cn be the map defined by ∂f (x) = infinity if the restriction (∂1f (x), . . . , ∂nf (x)). (∂f )−1(B(0; ε)) −→ B(0; ε) of ∂f to a sufficiently small ball B(0; ε) centered at the origin 0 ∈ Cn is proper. Then we say that f is tame at The following result is fundamental in the study of monodromies at infinity. Theorem 2.2 (Broughton [1] and Siersma-Tibar [27]). Assume that f is tame at infinity (more generally, with isolated W-singularities, see [27]). Then the generic fiber f −1(c) 4 KIYOSHI TAKEUCHI AND MIHAI TIB AR (c ∈ C \ Bf ) of f has the homotopy type of the bouquet of (n − 1)-spheres. In particular, we have (2.1) H j(f −1(c); C) = 0 (j 6= 0, n − 1). ✷ If f is tame at infinity, then of course Φ∞ n−1 is the unique non-trivial monodromy at infinity. avxv (av ∈ C) is non-degenerate at infinity if for any face γ of Γ∞(f ) such that 0 /∈ γ the complex hypersurface {x ∈ (C∗)n fγ(x) = 0} in (C∗)n is smooth and reduced, where we set Definition 2.3 ([12]). We say that the polynomial f (x) = Pv∈Zn fγ(x) =Pv∈γ∩Zn If f is convenient and non-degenerate at infinity, then by a result of Broughton [1] it is avxv. + + tame at infinity. However in this paper, we do not assume that f is convenient. Definition 2.4. Assume that dimΓ∞(f ) = n. Then we say that a face γ ≺ Γ∞(f ) is atypical if 0 ∈ γ and there exists a facet i.e. an (n − 1)-dimensional face Γ of Γ∞(f ) containing γ whose non-zero inner conormal vectors are not contained in the first quadrant + of Rn. Rn Remark 2.5. Our definition above is closely related to that of bad faces of NP (f ) in N´emethi-Zaharia [21]. If γ ≺ NP (f ) is a bad face of NP (f ), then the convex hull of {0} ∪ γ in Rn is an atypical one of Γ∞(f ). However, not all the atypical faces of Γ∞(f ) are obtained in this way. Example 2.6. Let n = 3 and consider a non-convenient polynomial f (x, y, z) on C3 whose Newton polyhedron at infinity Γ∞(f ) is the convex hull of the points (2, 0, 0), (2, 2, 0), (2, 2, 3) ∈ R3 + and the origin 0 = (0, 0, 0) ∈ R3. Then the line seg- ment connecting the point (2, 2, 0) and the origin 0 ∈ R3 is an atypical face of Γ∞(f ). However the triangle whose vertices are the points (2, 0, 0), (2, 2, 0) and the origin 0 ∈ R3 is not so. Example 2.7. Let n = 3 and consider a non-convenient polynomial f (x, y, z) on C3 whose Newton polyhedron at infinity Γ∞(f ) is the convex hull of the points (2, 0, 0), (0, 2, 0), (1, 1, 2) ∈ R3 + and the origin 0 = (0, 0, 0) ∈ R3. Then the line segment connecting the point (2, 0, 0) and the origin 0 ∈ R3 is an atypical face of Γ∞(f ). If dimΓ∞(f ) = n, to the n-dimensional integral polytope Γ∞(f ) in Rn we can naturally associate a subdivision of (the dual vector space of) Rn into rational convex polyhedral cones as follows. For an element u ∈ Rn of (the dual vector space of) Rn define the supporting face γu ≺ Γ∞(f ) of u in Γ∞(f ) by (2.2) γu =(cid:26)v ∈ Γ∞(f ) hu, vi = min w∈Γ∞(f )hu, wi(cid:27) . Then we introduce an equivalence relation ∼ on (the dual vector space of) Rn by u ∼ u′ ⇐⇒ γu = γu′. We can easily see that for any face γ ≺ Γ∞(f ) of Γ∞(f ) the closure of the equivalence class associated to γ in Rn is an (n − dimγ)-dimensional rational convex MONODROMIES AT INFINITY OF NON-TAME POLYNOMIALS 5 polyhedral cone σ(γ) in Rn. Moreover the family {σ(γ) γ ≺ Γ∞(f )} of cones in Rn thus obtained is a subdivision of Rn and satisfies the axiom of fans (see [7] and [22] etc.). We call it the dual fan of Γ∞(f ). Then we have the following characterization of atypical faces of Γ∞(f ). Lemma 2.8. Assume that dimΓ∞(f ) = n and let γ ≺ Γ∞(f ) be a face of Γ∞(f ) such that 0 ∈ γ. Then γ is atypical if and only if the cone σ(γ) which corresponds to it in the dual fan of Γ∞(f ) is not contained in Rn +. ✷ For a subset S ⊂ {1, 2, . . . , n} we define a coordinate subspace RS ≃ RS of Rn by (2.3) RS = {v = (v1, . . . , vn) ∈ Rn vi = 0 for any i /∈ S}. The following lemma should be obvious. Lemma 2.9. Assume that dimΓ∞(f ) = n and let γ ≺ Γ∞(f ) be a face of Γ∞(f ) such that 0 ∈ γ. Let RS ⊂ Rn be the minimal coordinate subspace of Rn containing γ and assume that dimγ < dimRS = S. Then γ is an atypical face of Γ∞(f ). ✷ By this lemma we can easily prove the following proposition. Proposition 2.10. Assume that dimΓ∞(f ) = n and a face γ ≺ Γ∞(f ) of Γ∞(f ) such that 0 ∈ γ is not atypical. Let RS ⊂ Rn be the minimal coordinate subspace of Rn containing γ. Then we have dimγ = dimRS = S and there exist exactly n − S facets i.e. (n − 1)- dimensional faces Γi ≺ Γ∞(f ) (i /∈ S) of Γ∞(f ) containing γ. Moreover they are explicitly given by (2.4) Γi = Γ∞(f ) ∩ {v = (v1, . . . , vn) ∈ Rn vi = 0} (i /∈ S). ✷ Definition 2.11. We say that a face γ ≺ Γ∞(f ) of Γ∞(f ) is at infinity if 0 /∈ γ. We say that such a face γ is moreover admissible if it is not contained in any atypical face of Γ∞(f ). For a face at infinity γ ≺ Γ∞(f ) of Γ∞(f ), let ∆γ be the convex hull of {0} ⊔ γ in Rn. Denote by L(∆γ) the (dimγ + 1)-dimensional linear subspace of Rn spanned by ∆γ and consider the lattice Mγ = Zn ∩ L(∆γ) ≃ Zdimγ+1 in it. Then there exists a unique non-zero primitive vector uγ in its dual lattice which takes its maximum in ∆γ exactly on γ ≺ ∆γ: (2.5) γ =(cid:26)v ∈ ∆γ huγ, vi = max w∈∆γhuγ, wi(cid:27) . We set (2.6) dγ = max w∈∆γhuγ, wi ∈ Z>0 and call it the lattice distance of γ from the origin 0 ∈ Rn. The following definition will be used in Sections 3 and 4. 6 KIYOSHI TAKEUCHI AND MIHAI TIB AR Definition 2.12. Assume that dimΓ∞(f ) = n. Then we say that a complex number λ ∈ C is an atypical eigenvalue of f if either λ = 1 or there exists a non-admissible face at infinity γ ≺ Γ∞(f ) of Γ∞(f ) such that λdγ = 1. We denote by Af ⊂ C the set of the atypical eigenvalues of f . Example 2.13. Let n = 2 and consider a non-convenient polynomial f (x, y) on C2 whose Newton polyhedron at infinity Γ∞(f ) is the convex hull of the points (1, 3), (3, 0), (3, 2) ∈ + and the origin 0 = (0, 0) ∈ R2. Then the line segment connecting the point (1, 3) and R2 the origin is the unique atypical face of Γ∞(f ) and we have Af = {1}. i Let −→ei =t (0, . . . , 0, 1, 0, . . . , 0) ∈ Rn (i = 1, 2, . . . , n) be the standard basis of Rn. The following result, whose proof immediately follows from Proposition 2.10, will be used in Section 3. Proposition 2.14. Assume that dimΓ∞(f ) = n and let γ ≺ Γ∞(f ) be a face at infinity of Γ∞(f ). Let RS ⊂ Rn be the minimal coordinate subspace of Rn containing γ and σ ⊂ Rn the cone in the dual fan of Γ∞(f ) which corresponds to γ ≺ Γ∞(f ). Then γ is admissible if and only if there exist some integral vectors −→f1, . . . ,−→fk ∈ (Rn \ Rn +) ∩ Zn such that (2.7) and (2.8) v∈Γ∞(f )h−→fj , vi < 0 min (1 ≤ j ≤ k) σ = (Xi /∈S R+−→ei ) + ( R+−→fj ). kXj=1 ✷ 3. Motivic Milnor fibers at infinity From now, following Denef-Loeser [3], [4] and Guibert-Loeser-Merle [9] we introduce motivic reincarnations of global (Milnor) fibers of polynomial maps. For the details, see Matsui-Takeuchi [17], Esterov-Takeuchi [6] and Raibaut [24]. We also follow the termi- nologies in [5], [10] and [11] etc. Let f : Cn −→ C be a general polynomial map. First, take a smooth compactification X of Cn. Next, by eliminating the points of indeterminacy of the meromorphic extension of f to X we obtain a commutative diagram (3.1) ι−−−→ eX yg P1 C −−−→j Cn fy such that horizontal arrows are open embeddings, g is a proper holomorphic map and eX \ Cn, Y := g−1(∞) are normal crossing divisors in eX. Take a local coordinate h of P1 in a neighborhood of ∞ ∈ P1 such that ∞ = {h = 0} and set eg = h ◦ g. Note MONODROMIES AT INFINITY OF NON-TAME POLYNOMIALS 7 (3.2) (3.3) where ψh and ψeg are nearby cycle functors (for the definition, see [5] and [11] etc.). Let c (f −1(R); C) ≃ H jψh(j!Rf!CCn) ≃ H jψh(Rg!ι!CCn) ≃ H j(Y ; ψeg(ι!CCn)), H j that eg is a holomorphic function defined on a neighborhood of the closed subvariety Y =eg−1(0) = g−1(∞) ⊂ eX \ Cn of eX. Then for R ≫ 0 we have us define an open subset Ω of eX by and set U = Ω ∩ Y . Then U (resp. the complement of Ω in eX) is a normal crossing divisor in Ω (resp. eX). By using this very special geometric situation we can easily prove where ι′ : Ω ֒→ eX is the inclusion. Now let E1, E2, . . . , Ek be the irreducible components of the normal crossing divisor U = Ω ∩ Y in Ω ⊂ eX. For each 1 ≤ i ≤ k, let bi > 0 be the order of the zero ofeg along Ei. For a non-empty subset I ⊂ {1, 2, . . . , k}, let us set H j(Y ; ψeg(ι!CCn)) ≃ H j(Y ; ψeg(ι′ Ω = Int(ι(Cn) ⊔ Y ) !CΩ)) ≃ H j c (U; ψeg(C eX )), the isomorphisms (3.4) (3.5) Ei, E◦ EI =\i∈I I = EI \[i6∈I Ei i bi dI i I of E◦ I −→ E◦ I is covered by such affine ]E◦ I,W = {(t, z) ∈ C∗ × (E◦ I as follows. First, for a point p ∈ E◦ and gg2,W =Qi∈I ξ open subsets W of Ω \ (∪i /∈IEi). Then as in [4, Section 3.3] by gluing the varieties (3.6) and dI = gcd(bi)i∈I > 0. Then, as in [4, Section 3.3], we can construct an unramified I we take an affine open neighborhood W ⊂ Ω\ (∪i /∈IEi) of p on which there exist regular functions ξi (i ∈ I) Galois covering fE◦ such that Ei ∩ W = {ξi = 0} for any i ∈ I. Then on W we have eg = gg1,W (gg2,W )dI , where we set gg1,W =egQi∈I ξ−bi . Note that gg1,W is a unit on W and gg2,W : W −→ C is a regular function. It is easy to see that E◦ together in the following way, we obtain the variety fE◦ open subset and eg = ]g1,W ′(]g2,W ′)dI is the decomposition of eg on it, we patch ]E◦ ]E◦ I,W ′ by the morphism (t, z) 7−→ (]g2,W ′(z)(gg2,W )−1(z) · t, z) defined over W ∩ W ′. Now for d ∈ Z>0, let µd ≃ Z/Zd be the multiplicative group consisting of the d-roots in C. We µd of the projective system {µi}i≥1 with morphisms denote by µ the projective limit lim ←−d µid −→ µi given by t 7−→ td. Then the unramified Galois covering fE◦ I admits a natural µdI -action defined by assigning the automorphism (t, z) 7−→ (ζdI t, z) of fE◦ I to the generator ζdI := exp(2π√−1/dI) ∈ µdI . Namely the variety fE◦ I is equipped with a good µ-action in the sense of [4, Section 2.4]. Following the notations in [4], denote by Mµ C the ring obtained from the Grothendieck ring Kµ 0 (VarC) of varieties over C with good µ- 0 (VarC). Recall that L ∈ Kµ actions by inverting the Lefschetz motive L ≃ C ∈ Kµ 0 (VarC) is endowed with the trivial action of µ. I ∩ W ) tdI = (gg1,W )−1(z)} I . If W ′ is another such I,W and I over E◦ I of E◦ 8 KIYOSHI TAKEUCHI AND MIHAI TIB AR Definition 3.1 ([17] and [24]). We define the motivic Milnor fiber at infinity S∞ polynomial map f : Cn −→ C by f =XI6=∅ S∞ (3.7) I ] ∈ Mµ C. (1 − L)I−1[fE◦ Remark 3.2. By Guibert-Loeser-Merle [9, Theorem 3.9], the motivic Milnor fiber at infinity S∞ f of f does not depend on the compactification X of Cn. This fact was informed to us by Schurmann (private communication) and Raibaut [24]. f of the As in [4, Section 3.1.2 and 3.1.3], we denote by HSmon the abelian category of Hodge structures with a quasi-unipotent endomorphism. Then, to the object ψh(j!Rf!CCn) ∈ Db c({∞}) and the semisimple part of the monodromy automorphism acting on it, we can associate an element [H ∞ f ] ∈ K0(HSmon) f ] are "relative" monodromy filtrations. To describe the element [H ∞ as in [3] and [4]. Recall that the weight filtrations of H jψh(j!Rf!CCn) in the construction f ] ∈ K0(HSmon) of [H ∞ in terms of S∞ (3.9) f ∈ Mµ C, let χh : Mµ C −→ K0(HSmon) be the Hodge characteristic morphism defined in [4] which associates to a variety Z with a good µd-action the Hodge structure (3.8) (3.10) χh([Z]) =Xj∈Z (−1)j[H j c (Z; Q)] ∈ K0(HSmon) with the actions induced by the one z 7−→ exp(2π√−1/d)z (z ∈ Z) on Z. Then as in [17, Theorem 4.4], by applying the proof of [3, Theorem 4.2.1] to our situation (3.2) and (3.4), we obtain the following result. Theorem 3.3. In the Grothendieck group K0(HSmon), we have the equality (3.11) [H ∞ f ] = χh(S∞ f ). ✷ By using Newton polyhedrons at infinity, we can rewrite Theorem 3.3 more explicitly as follows. Let f ∈ C[x1, . . . , xn] be a "non-convenient" polynomial such that dimΓ∞(f ) = n. Assume that f is non-degenerate at infinity. Now let us consider Cn as a toric variety associated with the fan Σ0 in Rn formed by all the faces of the first quadrant Rn + ⊂ Rn. Denote by T ≃ (C∗)n the open dense torus in it. Let Σ1 be a subdivision of the dual fan of Γ∞(f ) which contains Σ0 as its subfan. Then we can construct a smooth subdivision Σ of Σ1 without subdividing the cones in Σ0 (see e.g. [23, Lemma (2.6), Chapter II, page 99]). This implies that the toric variety XΣ associated with Σ is a smooth compactification of Cn. Our construction of XΣ coincides with the one in Zaharia [33]. Recall that T acts on XΣ and the T -orbits are parametrized by the cones in Σ. For a cone σ ∈ Σ denote by Tσ ≃ (C∗)n−dimσ the corresponding T -orbit. We have also natural affine open subsets Cn(σ) ≃ Cn of XΣ associated to n-dimensional cones σ in Σ as follows. Let σ MONODROMIES AT INFINITY OF NON-TAME POLYNOMIALS 9 be an n-dimensional cone in Σ and {w1, . . . , wn} ⊂ Zn the set of the (non-zero) primitive vectors on the edges of σ. Then by the smoothness of Σ the semigroup ring C[Zn ∩ σ] is isomorphic to the polynomial ring C[y1, . . . , yn]. This implies that the affine open subset Cn(σ) := Spec(C[Zn ∩ σ]) of XΣ is isomorphic to Cn y the function f has the following form: y . Moreover, on Cn(σ) ≃ Cn avyhw1,vi 1 · · · yhwn,vi n = yb1 1 · · · ybn n × fσ(y), f (y) = Xv∈Zn + avxv, (3.12) where we set f =Pv∈Zn (3.13) + bi = min v∈Γ∞(f )hwi, vi ≤ 0 (i = 1, 2, . . . , n) y . In Cn(σ) ≃ Cn and fσ(y) is a polynomial on Cn(σ) ≃ Cn y the hypersurface Z := f −1(0) ⊂ XΣ is explicitly written as {y ∈ Cn(σ) fσ(y) = 0}. The variety XΣ is covered by such affine open subsets. Let τ be a d-dimensional face of the n-dimensional cone σ ∈ Σ. For simplicity, assume that w1, . . . , wd generate τ . Then in the affine chart Cn(σ) ≃ Cn y the T -orbit Tτ associated to τ is explicitly defined by Tτ = {(y1, . . . , yn) ∈ Cn(σ) y1 = · · · = yd = 0, yd+1, . . . , yn 6= 0} ≃ (C∗)n−d. Hence we have (3.14) XΣ = [dimσ=n Cn(σ) = Gτ ∈Σ Tτ . Now f extends to a meromorphic function on XΣ, which may still have points of indeter- minacy. For simplicity we denote this meromorphic extension also by f . From now on, we will eliminate its points of indeterminacy by blowing up XΣ (see [15, Section 3] and [17, Section 3] for the details). For a cone σ in Σ by taking a non-zero vector u in the relative interior rel.int(σ) of σ we define a face γ(σ) of Γ∞(f ) by (3.15) γ(σ) =(cid:26)v ∈ Γ∞(f ) hu, vi = min w∈Γ∞(f )hu, wi(cid:27) . This face γ(σ) does not depend on the choice of u ∈ rel.int(σ) and is called the supporting face of σ in Γ∞(f ). Following [13], we say that a T -orbit Tσ in XΣ (or a cone σ ∈ Σ) is at infinity if its supporting face γ(σ) ≺ Γ∞(f ) is at infinity i.e. 0 /∈ γ(σ). We can easily see that f has poles on the union of T -orbits at infinity. Let ρ1, ρ2, . . . , ρm be the 1-dimensional cones at infinity in Σ and set Ti = Tρi. We call the cones ρi rays at infinity in Σ. Then T1, T2, . . . , Tm are the (n − 1)-dimensional T -orbits at infinity in XΣ. For any i = 1, 2, . . . , m the toric divisor Di := Ti is a smooth hypersurface in XΣ and the poles of f are contained in D1 ∪ · · · ∪ Dm. Let us denote the (unique non-zero) primitive vector in ρi ∩ Zn by ui. Then the order ai > 0 of the pole of f along Di is given by (3.16) ai = − min v∈Γ∞(f )hui, vi. Moreover by the non-convenience of f , there exist some cones σ ∈ Σ such that σ /∈ Σ0 and 0 ∈ γ(σ) i.e. γ(σ) is an atypical face of Γ∞(f ). For such σ the function f extends holomorphically to a neighborhood of Tσ ⊂ XΣ \ Cn. For this reason we call them "horizontal" T -orbits in XΣ. Note also that by the non-degeneracy at infinity of f , for 10 KIYOSHI TAKEUCHI AND MIHAI TIB AR DI :=Ti∈I Di transversally (or the intersection is empty). At such intersection points, any non-empty subset I ⊂ {1, 2, . . . , m} the hypersurface Z = f −1(0) in XΣ intersects f has indeterminacy. Now, in order to eliminate the indeterminacy of the meromorphic function f on XΣ, we first consider the blow-up π1 : X (1) Σ −→ XΣ of XΣ along the (n− 2)- dimensional smooth subvariety D1 ∩ Z. Then the indeterminacy of the pull-back f ◦ π1 of f to X (1) Σ is improved. If f ◦ π1 still has points of indeterminacy on the intersection of the exceptional divisor E1 of π1 and the proper transform Z (1) of Z, we construct the blow-up π2 : X (2) Σ along E1 ∩ Z (1). By repeating this procedure a1 times, we obtain a tower of blow-ups Σ −→ X (1) Σ of X (1) (3.17) X (a1) Σ −→πa1 · · ·· · · −→π2 X (1) Σ −→π1 XΣ. Σ has no indeterminacy over T1. Then the pull-back of f to X (a1) It also extends to a holomorphic function on (an open dense subset of) the exceptional divisor of the last blow-up πa1 whose monodromy at infinity is trivial. For this reason we call it a horizontal exceptional divisor. For the details see the figures in [15, page 420]. Next we apply this construction to the proper transforms of D2 and Z in X (a1) Σ . Then we obtain also a tower of blow-ups (3.18) X (a1)(a2) −→ · · ·· · · −→ X (a1)(1) −→ X (a1) and the indeterminacy of the pull-back of f to X (a1)(a2) is eliminated over T1 ⊔ T2. By applying the same construction to (the proper transforms of) D3, D4, . . . , Dm, we finally Σ Σ Σ Σ obtain a birational morphism π : fXΣ −→ XΣ such that g := f ◦ π has no point of indeterminacy on the whole fXΣ. Note that the smooth compactification fXΣ of Cn thus obtained is not a toric variety any more. On fXΣ there are m horizontal exceptional divisors. By eliminating the points of indeterminacy of the meromorphic extension of f to XΣ we have constructed the commutative diagram: (3.19) ι−−−→ fXΣ yg P1. C −−−→j Cn fy proper transforms D′ Take a local coordinate h of P1 in a neighborhood of ∞ ∈ P1 such that ∞ = {h = 0} f . Then the divisor U = Y ∩ Ω in Ω contains not only the and set eg = h ◦ g, Y =eg−1(0) = g−1(∞) ⊂ fXΣ and Ω = Int(ι(Cn) ⊔ Y ) as before. For simplicity, let us set eg = 1 m of D1, . . . , Dm in fXΣ but also the exceptional divisors of the blow-up: fXΣ −→ XΣ. So the motivic Milnor fiber at infinity S∞ defined by this compactification fXΣ of Cn contains also unramified Galois coverings of some subsets of these exceptional divisors. However they are not necessary to compute the Hodge realization of S∞ f as follows. For each non-empty subset I ⊂ {1, 2, . . . , m}, set f of f : Cn −→ C 1, . . . , D′ MONODROMIES AT INFINITY OF NON-TAME POLYNOMIALS 11 DI =Ti∈I Di, (3.20) D◦ I = DI \( [i /∈I Di! ∪ f −1(0)) ⊂ XΣ av (v ∈ γ), −1 (v = 0), 0 (otherwise), bv = it as 1 I of D◦ f is regular on D◦ I and we can decompose and dI = gcd(ai)i∈I > 0. Then the functioneg = 1 I in XΣ, where eg1 is a unit on W and eg2 : W −→ C is regular. Therefore we can construct an unramified Galois covering fD◦ ring Mµ result. f = eg1(eg2)dI globally on a Zariski open neighborhood W of D◦ I with a natural µdI -action as in (3.6). Let [fD◦ C which corresponds to fD◦ χh(cid:0)S∞ Theorem 3.4. Assume that dimΓ∞(f ) = n and f is non-degenerate at infinity. Then we have the equality I ] be the element of the I . Then as in [17, Theorem 4.7] we obtain the following χh(cid:16)(1 − L)I−1[fD◦ I ](cid:17) f (cid:1) =XI6=∅ in the Grothendieck group K0(HSmon). (3.21) ✷ For a face at infinity γ ≺ Γ∞(f ) of Γ∞(f ), by using the lattice Mγ = Zn ∩ L(∆γ) ≃ Zdimγ+1 in L(∆γ) ≃ Rdimγ+1 we set T∆γ := Spec(C[Mγ]) ≃ (C∗)dimγ+1. Moreover let L(γ) be the smallest affine linear subspace of Rn containing γ and for v ∈ Mγ define their lattice heights ht(v, γ) ∈ Z from L(γ) in L(∆γ) so that we have ht(0, γ) = dγ > 0. Then to the group homomorphism Mγ −→ C∗ defined by v 7−→ ζ ht(v,γ) we can naturally associate an element τγ ∈ T∆γ . We define a Laurent polynomial gγ =Pv∈Mγ bvxv on T∆γ by dγ (3.22) + where f =Pv∈Zn ∆γ ⊂ T∆γ is invariant by the multiplication lτγ : T∆γ avxv. Then the Newton polytope NP (gγ) of gγ is ∆γ, suppgγ ⊂ {0}⊔ γ and the hypersurface Z ∗ ∆γ = {x ∈ T∆γ gγ(x) = 0} is non-degenerate (see [17, Section 4]). Since Z ∗ ∆γ admits an action of µdγ . We thus obtain an element [Z ∗ C. For a face at infinity γ ≺ Γ∞(f ) let sγ > 0 be the dimension of the minimal coordinate subspace of Rn containing γ and set mγ = sγ − dimγ − 1 ≥ 0. Finally, for λ ∈ C and an element H ∈ K0(HSmon) denote by Hλ ∈ K0(HSmon) the eigenvalue λ-part of H. Then by applying the proof of [17, Theorem 5.7 (i)] to the geometric situation in Proposition 2.14, we obtain the following result. ∼−→ T∆γ by τγ, Z ∗ ∆γ ] of Mµ Theorem 3.5. Assume that dimΓ∞(f ) = n and f is non-degenerate at infinity. Then for any λ /∈ Af we have the equality f ]λ = χh(S∞ (3.23) χh((1 − L)mγ · [Z ∗ ∆γ ])λ [H ∞ in K0(HSmon), where in the sum Pγ the face γ of Γ∞(f ) ranges through the admissible ones at infinity. f )λ =Xγ 12 KIYOSHI TAKEUCHI AND MIHAI TIB AR Proof. The proof is similar to that of [17, Theorem 5.7 (i)]. By Proposition 2.14 the argument at the end of the proof of [17, Theorem 5.7 (i)] holds for admissible faces at infinity of Γ∞(f ). But it does not hold for non-admissible ones by the presence of horizontal T -orbits in XΣ. Hence it suffices to avoid atypical eigenvalues λ ∈ Af . (cid:3) 4. Main results In this section, we consider non-convenient polynomials f : Cn −→ C such that dimΓ∞(f ) = n. For λ ∈ C and j ∈ Z let H j(f −1(R); C)λ ⊂ H j(f −1(R); C) be the general- : H j(f −1(R); C) ∼−→ ized eigenspace for the eigenvalue λ of the monodromy at infinity Φ∞ j H j(f −1(R); C) (R ≫ 0). Denote by Φ∞ to H j(f −1(R); C)λ. As- suming also that f is non-degenerate at infinity, for non-atypical eigenvalues λ /∈ Af of f we will prove the concentration j,λ the restriction of Φ∞ j (j 6= n − 1) H j(f −1(R); C)λ ≃ 0 (4.1) for the λ-parts H j(f −1(R); C)λ of the cohomology groups of the generic fiber f −1(R) (R ≫ 0) of f . This implies that the Jordan normal forms of the λ-parts Φ∞ j,λ of the monodromies at infinity of f can be completely determined by Γ∞(f ) as in [17, Section 5]. For this purpose we first consider Laurent polynomials on T = (C∗)n. Let f ′ ∈ C[x±1 1 , . . . , x±1 n ] be a Laurent polynomial on T = (C∗)n. We define its Newton polytope NP (f ′) ⊂ Rn as usual and let Γ∞(f ′) ⊂ Rn be the convex hull of {0} ∪ NP (f ′) in Rn. We say that a face γ ≺ Γ∞(f ′) is at infinity if 0 /∈ γ. By using faces at infinity of Γ∞(f ′) we define also the non-degeneracy at infinity of f ′ as in Definition 2.3. Definition 4.1. Assume that dimΓ∞(f ′) = n. Then we say that a face γ ≺ Γ∞(f ′) is atypical if 0 ∈ γ. Moreover a face at infinity γ ≺ Γ∞(f ′) is called admissible if it is not contained in any atypical one. As in Definition 2.12, by using non-admissible faces at infinity of Γ∞(f ′) we define the subset Af ′ ⊂ C of the atypical eigenvalues of f ′ such that 1 ∈ Af ′. Finally let us recall the following result of Libgober-Sperber [13] on the monodromies at infinity : H j((f ′)−1(R); C) ∼−→ H j((f ′)−1(R); C) (R ≫ 0) of f ′ : T = (C∗)n −→ C. We define Ψ∞ j the monodromy zeta function at infinity ζ ∞ f ′ (t) ∈ C((t)) of f ′ by (4.2) ζ ∞ f ′ (t) = n−1Yj=0 det(id − tΨ∞ j )(−1)j ∈ C((t)). For a face at infinity γ ≺ Γ∞(f ′) let L(γ) ≃ Rdimγ be the minimal affine subspace of Rn containing γ. Proposition 4.2. (Libgober-Sperber [13]) Assume that dimΓ∞(f ′) = n and f ′ is non- degenerate at infinity. Then we have (4.3) ζ ∞ f ′ (t) =Yγ (1 − tdγ )(−1)n−1VolZ(γ) ∈ C((t)), MONODROMIES AT INFINITY OF NON-TAME POLYNOMIALS 13 1 , . . . , x±1 where in the product Qγ the face γ ≺ Γ∞(f ′) ranges through those at infinity such that dimγ = n − 1 and VolZ(γ) ∈ Z>0 is the normalized (n − 1)-dimensional volume of γ with respect to the lattice L(γ) ∩ Zn ≃ Zn−1. Proposition 4.3. Let f ′ ∈ C[x±1 n ] be a Laurent polynomial on T = (C∗)n such that dimΓ∞(f ′) = n. Assume that f ′ is non-degenerate at infinity. For λ ∈ C and j ∈ Z denote the generalized eigenspace for the eigenvalue λ of its monodromy at in- : H j((f ′)−1(R); C) ∼−→ H j((f ′)−1(R); C) (R ≫ 0) by H j((f ′)−1(R); C)λ ⊂ finity Ψ∞ j H j((f ′)−1(R); C). Then for any λ /∈ Af ′ we have the concentration (4.4) for the generic fiber (f ′)−1(R) ⊂ T (R ≫ 0) of f ′. If λ /∈ Af ′ satisfies the condition H n−1((f ′)−1(R); C)λ 6= 0, then there exists a facet at infinity γ ≺ Γ∞(f ′) such that λdγ = 1. Moreover for such λ the relative monodromy filtration of H n−1((f ′)−1(R); C)λ (R ≫ 0) coincides with the absolute one (up to some shift). H j((f ′)−1(R); C)λ ≃ 0 (j 6= n − 1) ✷ Proof. We will prove the proposition by induction on n. If n = 1 the assertion is obvious. Assume that we already proved it for the lower dimensions 1, 2, . . . , n − 1. Let Σ′ 1 be the dual fan of Γ∞(f ′) in Rn and Σ′ its smooth subdivision. Then the toric variety XΣ′ associated to Σ′ is a smooth compactification of T . By eliminating the points of indeterminacy of the meromorphic extension of f ′ to XΣ′ as in Section 3 we obtain a commutative diagram: (4.5) ι′ −−−→ gXΣ′ yg′ P1 C −−−→j T f ′y of holomorphic maps, where j and ι′ are open embeddings and g′ is proper. Now restricting the map g′ : gXΣ′ −→ P1 to C ⊂ P1 we set K ′ = (g′)−1(C) = gXΣ′ \ (g′)−1(∞). Let κ′ : K ′ −→ C be the restriction of g′ to K ′. Set D′ = K ′ \ T and let iD′ : D′ −→ K ′ and iT : T −→ K ′ be the inclusions. Then we obtain also a commutative diagram: (4.6) T f ′y C iT−−−→ K ′ yκ′ C. Note that the normal crossing divisor D′ in K ′ is a union of horizontal T -orbits (which correspond to atypical faces γ ≺ Γ∞(f ′)) and the horizontal exceptional divisors on gXΣ′. By our induction hypothesis and Proposition 4.2, for λ /∈ Af ′ the monodromies at in- finity of the restrictions of κ′ to these horizontal T -orbits have no λ-part. Moreover the corresponding monodromies at infinity over the horizontal exceptional divisors have only the eigenvalue 1 ∈ Af ′. On the other hand, by applying the functor Rκ′ ! to the ∗ = Rκ′ 14 KIYOSHI TAKEUCHI AND MIHAI TIB AR distinguished triangle (4.7) (iT )!CT −→ R(iT )∗CT −→ (iD′)∗i−1 D′ (R(iT )∗CT ) −→ +1 we obtain a distinguished triangle R(f ′)!CT −→ R(f ′)∗CT −→ R(κ′D′)∗i−1 (4.8) Then by using the above description of κ′D′ : D′ −→ C, for λ /∈ Af ′ we can easily show (4.9) D′ (R(iT )∗CT ) −→ +1. ψh,λ(j!R(κ′D′)∗i−1 D′ (R(iT )∗CT )) ≃ 0, where ψh,λ is the λ-part of the nearby cycle functor ψh. This implies that there exists an isomorphism ψh,λ(j!R(f ′)!CT ) ≃ ψh,λ(j!R(f ′)∗CT ). c ((f ′)−1(R); C)λ ≃ H j((f ′)−1(R); C)λ H j (4.10) Namely, for any λ /∈ Af ′ and j ∈ Z we have an isomorphism (4.11) Since the generic fiber (f ′)−1(R) ⊂ T (R ≫ 0) of f ′ is affine, the left (resp. right) hand side is zero for j < n − 1 (resp. j > n − 1). Hence we obtain the desired concentration (4.12) for R ≫ 0. Now the second assertion follows immediately from Proposition 4.2. Also the last assertion follows from the proof of Sabbah [26, Theorem 13.1] by using the isomor- phism (4.10). This completes the proof. (cid:3) Remark 4.4. If 0 ∈ Int(NP (f ′)) then the Laurent polynomial f ′ : T = (C∗)n −→ C is cohomologically tame at infinity in the sense of N´emethi-Sabbah [20] and Sabbah [26] on H j((f ′)−1(R); C)λ ≃ 0 (j 6= n − 1) (R ≫ 0). our compactification gXΣ′ of T . In this case the first assertion of Proposition 4.3 is due to [20]. Theorem 4.5. Let f ∈ C[x1, . . . , xn] be a non-convenient polynomial such that dimΓ∞(f ) = n. Assume that f is non-degenerate at infinity. Then for any non-atypical eigenvalue λ /∈ Af of f we have the concentration (4.13) for the generic fiber f −1(R) ⊂ Cn (R ≫ 0) of f . Moreover for such λ the relative monodromy filtration of H n−1(f −1(R); C)λ (R ≫ 0) coincides with the absolute one (up to some shift). H j(f −1(R); C)λ ≃ 0 (j 6= n − 1) Proof. We will freely use the notations in Section 3. For example, we consider the com- mutative diagram: (4.14) ι−−−→ fXΣ yg P1. C −−−→j Cn fy MONODROMIES AT INFINITY OF NON-TAME POLYNOMIALS 15 By restricting the map g : fXΣ −→ P1 to C ⊂ P1 we set K = g−1(C) = fXΣ \ g−1(∞) and κ = gK : K −→ C. Set D = K \ Cn and let iD : D −→ K and i : Cn −→ K be the inclusions. Then we obtain also a commutative diagram: Cn fy C i−−−→ K yκ C. (4.15) Note that the normal crossing divisor D in K is a union of horizontal T -orbits (which correspond to atypical faces γ ≺ Γ∞(f )) and the horizontal exceptional divisors on fXΣ. By Proposition 4.3, for λ /∈ Af the monodromies at infinity of the restrictions of κ to these horizontal T -orbits have no λ-part. Moreover the corresponding monodromies at infinity over the horizontal exceptional divisors have only the eigenvalue 1 ∈ Af . On the other hand, by applying the functor Rκ∗ = Rκ! to the distinguished triangle (4.16) i!CCn −→ Ri∗CCn −→ (iD)∗i−1 D (Ri∗CCn) −→ +1 we obtain a distinguished triangle Rf!CCn −→ Rf∗CCn −→ R(κD)∗i−1 (4.17) Then by using the above description of κD : D −→ C, for λ /∈ Af we can easily show (4.18) D (Ri∗CCn) −→ +1. ψh,λ(j!R(κD)∗i−1 This implies that there exists an isomorphism D (Ri∗CCn)) ≃ 0. c (f −1(R); C)λ ≃ H j(f −1(R); C)λ H j ψh,λ(j!Rf!CCn) ≃ ψh,λ(j!Rf∗CCn). (4.19) Namely, for any λ /∈ Af and j ∈ Z we have an isomorphism (4.20) Since the generic fiber f −1(R) ⊂ Cn (R ≫ 0) of f is affine, the left (resp. right) hand side is zero for j < n − 1 (resp. j > n − 1). Hence we obtain the desired concentration (4.21) for R ≫ 0. Moreover the last assertion follows from the proof of Sabbah [26, Theorem 13.1] by using the isomorphism (4.19). This completes the proof. H j(f −1(R); C)λ ≃ 0 (j 6= n − 1) (R ≫ 0). (cid:3) Remark 4.6. As is clear from the proof above, Theorem 4.5 can be easily generalized to arbitrary polynomial maps f : U −→ C of affine algebraic varieties U. We leave the precise formulation to the reader. If n = 2 the first assertion of Theorem 4.5 can be improved as follows. Lemma 4.7. Assume that a polynomial f (x, y) ∈ C[x, y] of two variables is non- degenerate at infinity and satisfies the condition dimΓ∞(f ) = 2. Then the generic fiber of f : C2 −→ C is connected and hence H 0(f −1(R); C) ≃ C for R ≫ 0. In particular, for any λ 6= 1 we have the concentration (4.22) (R ≫ 0). H j(f −1(R); C)λ ≃ 0 (j 6= 1) 16 KIYOSHI TAKEUCHI AND MIHAI TIB AR Proof. By the classification of open connected Riemann surfaces, it suffices to show that there is no decomposition of f of the form f (x, y) = f ( f (x, y)) (4.23) by polynomials f (t) and f (x, y) such that deg f (t) ≥ 2. Assume that there exists such a decomposition f = f ◦ f and set m = deg f ≥ 2. Then we have Γ∞(f ) = mΓ∞( f ). Take a face at infinity γ ≺ Γ∞(f ) of Γ∞(f ) satisfying dimγ = 1 and let γ ≺ Γ∞( f ) be the corresponding one of Γ∞( f ) such that γ = mγ. Denote by fγ (resp. fγ) the γ-part of f (resp. the γ-part of f ). Then we have fγ = ( fγ)m (up to some non-zero constant multiple) for m ≥ 2. This contradicts the non-degeneracy at infinity of f . (cid:3) For an element [V ] ∈ K0(HSmon), V ∈ HSmon with a quasi-unipotent endomorphism Θ : V ∼−→ V , p, q ≥ 0 and λ ∈ C denote by ep,q([V ])λ the dimension of the λ-eigenspace of the morphism V p,q ∼−→ V p,q induced by Θ on the (p, q)-part V p,q of V . Then by Theorem 4.5 we immediately obtain the following result. Corollary 4.8. Assume that dimΓ∞(f ) = n and f is non-degenerate at infinity. Let λ /∈ Af . Then we have ep,q([H ∞ f ])λ = 0 for (p, q) /∈ [0, n − 1] × [0, n − 1]. Moreover for any (p, q) ∈ [0, n − 1] × [0, n − 1] we have the Hodge symmetry (4.24) f ])λ. f ])λ = en−1−q,n−1−p([H ∞ ep,q([H ∞ By using the notations in Section 3 we thus obtain the following theorem, whose proof is similar to that of [17, Theorem 5.7 (ii)]. ✷ Theorem 4.9. Assume that dimΓ∞(f ) = n and f is non-degenerate at infinity. Let λ /∈ Af and k ≥ 1. Then the number of the Jordan blocks for the eigenvalue λ with sizes ≥ k in Φ∞ n−1 : H n−1(f −1(R); C) ∼−→ H n−1(f −1(R); C) (R ≫ 0) is equal to ∆γ ]))λ) , (−1)n−1 Xp+q=n−2+k,n−1+k(Xγ where in the sumPγ the face γ of Γ∞(f ) ranges through the admissible ones at infinity. ep,q(χh((1 − L)mγ · [Z ∗ (4.25) ✷ By this theorem and the results in [17, Section 2] we immediately obtain the general- izations of [17, Theorems 5.9, 5.14 and 5.16] to non-tame polynomials. Here we introduce only that of [17, Theorem 5.9]. Denote by Cone∞(f ) the closed cone R+Γ∞(f ) ⊂ Rn + generated by Γ∞(f ). Let q1, . . . , ql (resp. γ1, . . . , γl′) be the 0-dimensional (resp. 1- dimensional) faces at infinity of Γ∞(f ) such that qi ∈ Int(Cone∞(f )) (resp. the relative interior rel.int(γi) of γi is contained in Int(Cone∞(f ))). For each qi (resp. γi), denote by di > 0 (resp. ei > 0) its lattice distance from the origin 0 ∈ Rn. For 1 ≤ i ≤ l′, let ∆i be the convex hull of {0} ⊔ γi in Rn. Then for λ 6= 1 and 1 ≤ i ≤ l′ such that λei = 1 we set (4.26) n(λ)i = ♯{v ∈ Zn ∩ rel.int(∆i) ht(v, γi) = k} + ♯{v ∈ Zn ∩ rel.int(∆i) ht(v, γi) = ei − k}, MONODROMIES AT INFINITY OF NON-TAME POLYNOMIALS 17 where k is the minimal positive integer satisfying λ = ζ k ei and for v ∈ Zn ∩ rel.int(∆i) we denote by ht(v, γi) the lattice height of v from the base γi of ∆i. Then we have the following generalization of [17, Theorem 5.9]. Theorem 4.10. Assume that dimΓ∞(f ) = n and f is non-degenerate at infinity. Let λ /∈ Af . Then we have (1) The number of the Jordan blocks for the eigenvalue λ with the maximal possible size n−1 : H n−1(f −1(R); C) ∼−→ H n−1(f −1(R); C) (R ≫ 0) is equal to ♯{qi λdi = (2) The number of the Jordan blocks for the eigenvalue λ with the second maximal n in Φ∞ 1}. possible size n − 1 in Φ∞ n−1 is equal toPi : λei =1 n(λ)i. ✷ Remark 4.11. By Proposition 4.3 we can similarly obtain the analogues of [17, Theorems 5.9, 5.14 and 5.16] for Laurent polynomials f ′ ∈ C[x±1 n ]. The results on the Jordan normal forms of their monodromies at infinity for the eigenvalues λ /∈ Af ′ are explicitly described by the admissible faces at infinity of Γ∞(f ′). We omit the details. 1 , . . . , x±1 Moreover in the situation above, we can obtain also a closed formula for the multiplicities of the non-atypical eigenvalues λ /∈ Af in the monodromy at infinity n−1 : H n−1(f −1(R); C) ∼−→ H n−1(f −1(R); C) (R ≫ 0) as follows. We define the mon- Φ∞ odromy zeta function at infinity ζ ∞ Then by our compactification fXΣ of Cn we obtain the following refinement of the previous results in [13] and [15]. In particular here we can remove the condition (∗) in [15]. Theorem 4.12. Assume that dimΓ∞(f ) = n and f is non-degenerate at infinity. Then we have (4.28) ζ ∞ f (t) =Yγ (1 − tdγ )(−1)sγ −1VolZ(γ) ∈ C((t)), where in the product Qγ the face γ ≺ Γ∞(f ) ranges through those at infinity satisfying the condition mγ = sγ − dimγ − 1 = 0 and VolZ(γ) ∈ Z>0 is the normalized (dimγ)- dimensional volume of γ with respect to the lattice L(γ) ∩ Zn ≃ Zdimγ. Example 4.13. Let n = 3 and consider a non-convenient polynomial f (x, y, z) on C3 whose Newton polyhedron at infinity Γ∞(f ) is the convex hull of the points (2, 0, 0), (0, 2, 0), (1, 1, 1) ∈ R3 + and the origin 0 = (0, 0, 0) ∈ R3. Then the line segment connecting the point (2, 0, 0) and the origin 0 ∈ R3 is an atypical face of Γ∞(f ). Hence the 0-dimensional face at infinity γ = {(2, 0, 0)} ≺ Γ∞(f ) of Γ∞(f ) contained in it is not admissible. However it satisfies the condition mγ = sγ − dimγ− 1 = 1− 0− 1 = 0. For the proof of Theorem 4.12 we have to consider also the contribution from such non-admissible faces at infinity of Γ∞(f ). ✷ (4.27) ζ ∞ f (t) = f (t) ∈ C((t)) of f by j )(−1)j det(id − tΦ∞ n−1Yj=0 ∈ C((t)). 18 KIYOSHI TAKEUCHI AND MIHAI TIB AR If we restrict ourselves to the non-atypical eigenvalues λ /∈ Af for which we have the concentration H j(f −1(R); C)λ ≃ 0 (4.29) (R ≫ 0) in Theorem 4.5, we have the following result. Corollary 4.14. Assume that dimΓ∞(f ) = n and f is non-degenerate at infinity. Then for any λ /∈ Af the multiplicity of the eigenvalue λ in the monodromy at infinity Φ∞ n−1 of f is equal to that of the factor (1 − λt) = λ · (1/λ − t) in the rational function (4.30) (j 6= n − 1) Yγ (1 − tdγ )(−1)n−sγ VolZ(γ) ∈ C((t)), where in the productQγ the face γ ≺ Γ∞(f ) of Γ∞(f ) ranges through the admissible ones at infinity satisfying the condition mγ = sγ − dimγ − 1 = 0. ✷ 5. Monodromies around atypical fibers Let f : U −→ C be a polynomial map of an affine algebraic variety U and Bf ⊂ C the set of its bifurcation points. For a point b ∈ Bf we choose sufficiently small ε > 0 such that Bf ∩ {x ∈ C x − b ≤ ε} = {b} (5.1) and set Cε(b) = {x ∈ C x − b = ε} ⊂ C. Then we obtain a locally trivial fibration f −1(Cε(b)) −→ Cε(b) over the small circle Cε(b) ⊂ C and the monodromy automorphisms j : H j(f −1(b + ε); C) ∼−→ H j(f −1(b + ε); C) (j = 0, 1, . . .) Φb (5.2) around the atypical fiber f −1(b) ⊂ U associated to it. We can construct Φb j's functorially as follows. Let hb be a holomorphic local coordinate of C on a neighborhood of b ∈ Bf such that b = {hb(x) = 0}. Then to the object ψhb(Rf!CU ) ∈ Db c({b}) and the semisimple part of the monodromy automorphism acting on it, we can associate an element [H b (5.3) f ] ∈ K0(HSmon). Recall that the weight filtration of [H b f ] is a relative one. In this situation, we can apply our methods in previous sections to the Jordan normal forms of Φb j. For the sake of simplicity, let us assume here that the central fiber f −1(b) ⊂ U is reduced and has only isolated singular points p1, p2, . . . , pl ∈ f −1(b) ⊂ U. When f is not tame at infinity, we have to consider also the singularities at infinity of f . For this purpose, let X be a smooth compactification of U for which there exists a commutative diagram (5.4) U fy C −−−→j ι−−−→ X yg P1 of holomorphic maps. Here ι and j are inclusion maps and g is proper. We may assume also that the divisor at infinity D = X \ U ⊂ X is normal crossing and all its irreducible MONODROMIES AT INFINITY OF NON-TAME POLYNOMIALS 19 components are smooth. We call the irreducible components of D contained in g−1(∞) ⊂ D (resp. in D \ g−1(∞) ⊂ D) "vertical" (resp. "horizontal") divisors at infinity of f in X. For the normal crossing divisor D let us consider the standard (minimal) stratification. Then for simplicity we assume also that the restriction gD\g−1(∞) : D \ g−1(∞) −→ C of g to the horizontal part D \ g−1(∞) of D has only stratified isolated singular points pl+1, . . . , pl+r in g−1(b) ⊂ X and all of them are contained in the smooth part of D\g−1(∞). By our assumption on f −1(b) ⊂ U this implies that the hypersurface f −1(b) = g−1(b) ⊂ X in X has also an isolated singular point at each pi (l + 1 ≤ i ≤ l + r). Remark 5.1. If U = Cn, b 6= f (0) and in addition to the conditions in Theorems 4.5 and 4.9 (i.e. dimΓ∞(f ) = n and f is non-degenerate at infinity) we assume that for any atypical face γ ≺ Γ∞(f ) such that dimγ < n− 1 the γ-part fγ : (C∗)n −→ C of f does not have the critical value b, then the meromorphic extension g of f to the compactification X = fXΣ satisfies the above-mentioned property in general (see also N´emethi-Zaharia [21], Zaharia [33]). In this case the stratified isolated singular points pl+1, . . . , pl+r are on the (n− 1)-dimensional horizontal T -orbits which correspond to the atypical facets of Γ∞(f ). For l + 1 ≤ i ≤ l + r, in a neighborhood of pi the divisor D is smooth and the function gD : D ≃ Cn−1 −→ C has an isolated singular point at pi ∈ D. Therefore we may consider the (local) Milnor monodromies of gD : D ≃ Cn−1 −→ C at pi ∈ D. Denote by Af,b ⊂ C the union of their eigenvalues and 1 ∈ C. Then by applying the proof of Theorem 4.5 to this situation, we obtain the following result. For λ ∈ C and j ∈ Z let H j(f −1(b + ε); C)λ ⊂ H j(f −1(b + ε); C) be the generalized eigenspace for the eigenvalue λ of the monodromy Φb Theorem 5.2. In the situation as above, for any λ /∈ Af,b we have the concentration (5.5) Moreover for such λ the relative monodromy filtration of H n−1(f −1(b + ε); C)λ coincides with the absolute one (up to some shift). ✷ H j(f −1(b + ε); C)λ ≃ 0 j around f −1(b). (j 6= n − 1). Corollary 5.3. Let λ /∈ Af,b. Then we have ep,q([H b Moreover for any (p, q) ∈ [0, n − 1] × [0, n − 1] we have the Hodge symmetry (5.6) f ])λ = en−1−q,n−1−p([H b f ])λ = 0 for (p, q) /∈ [0, n−1]×[0, n−1]. ep,q([H b f ])λ. ✷ From now on we shall use Theorem 5.2 and Corollary 5.3 to describe explicitly the Jordan normal form of Φb n−1 in terms of some Newton polyhedra associated to f . For this purpose, assume moreover that for any 1 ≤ i ≤ l + r there exists a local coordinate y = (y1, y2, . . . , yn) of X on a neighborhood Wi of pi such that pi = {y = 0} and the local defining polynomial fi(y) ∈ C[y1, . . . , yn] of the hypersurface f −1(b) = g−1(b) (for which we have f −1(b) = {fi(y) = 0}) is convenient and non-degenerate at y = 0 (see [32] etc.). We assume also that for l + 1 ≤ i ≤ l + r we have D = {yn = 0} in Wi. For 1 ≤ i ≤ l + r let Γ+(fi) ⊂ Rn + be the Newton polyhedron of fi at y = 0. Moreover for l + 1 ≤ i ≤ l + r we set (5.7) Γ◦ +(fi) = Γ+(fi) ∩ {v = (v1, . . . , vn) ∈ Rn vn = 0} . 20 KIYOSHI TAKEUCHI AND MIHAI TIB AR +(fi) is nothing but the Newton polyhedron of the restriction fiD of fi to Note that Γ◦ D = {yn = 0}. Definition 5.4. In the situation as above, we say that a complex number λ ∈ C is an atypical eigenvalue for b ∈ Bf if either λ = 1 or there exists a compact face γ ≺ Γ◦ +(fi) of Γ◦ f,b ⊂ C the set of the atypical eigenvalues for b ∈ Bf . +(fi) for some l + 1 ≤ i ≤ l + r such that λdγ = 1. We denote by A◦ By the main theorem of Varchenko [32] we have Af,b ⊂ A◦ f,b. On the other hand, as in [4], [17] and [18], for 1 ≤ i ≤ l + r by a toric modification πi : Yi −→ Wi of Wi we can explicitly construct the motivic Milnor fiber Sfi,pi ∈ Mµ C of fi at pi. See [18] for the details. For l + 1 ≤ i ≤ l + r let (Wi ∩ D)′ ⊂ Yi be the proper transform of Wi ∩ D = {yn = 0} by πi and S◦ C the base change of Sfi,pi by the inclusion map Yi \ (Wi ∩ D)′ ֒→ Yi. Let [Zf,b] ∈ Mµ C be the class of the variety Zf,b = f −1(b) \ {p1, p2, . . . , pl} with the trivial action of µ and set fi,pi ∈ Mµ (5.8) S b f = [Zf,b] + Sfi,pi + lXi=1 l+rXi=l+1 fi,pi ∈ Mµ S◦ C. Then as in [17, Theorem 4.4], by the proof of [3, Theorem 4.2.1] we obtain the following result. Theorem 5.5. In K0(HSmon) we have the equality f ] = χh(S b f ). (5.9) [H b ✷ By Theorems 5.2 and 5.5 and Corollary 5.3, for any λ /∈ A◦ f,b we can describe explicitly the λ-part of the Jordan normal form of Φb n−1 as follows. For 1 ≤ i ≤ l + r let γ ≺ Γ+(fi) be a compact face of Γ+(fi). Denote by ∆γ the convex hull of {0} ⊔ γ in Rn. Let L(∆γ) be the (dimγ + 1)-dimensional linear subspace of Rn spanned by ∆γ and consider the lattice Mγ = Zn ∩ L(∆γ) ≃ Zdimγ+1 in it. Then we set T∆γ := Spec(C[Mγ]) ≃ (C∗)dimγ+1. Moreover for the points v ∈ Mγ we define their lattice heights ht(v, γ) ∈ Z from the affine hyperplane L(γ) in L(∆γ) so that we have ht(0, γ) = dγ > 0. Then to the group homomorphism Mγ −→ C∗ defined by v 7−→ ζ −ht(v,γ) we can naturally associate an element τγ ∈ T∆γ . We define a Laurent polynomial gγ =Pv∈Mγ bvyv on T∆γ by dγ (5.10) (v ∈ γ), av −1 (v = 0), 0 (otherwise), bv = + where fi =Pv∈Zn avyv. Then we have NP (gγ) = ∆γ, suppgγ ⊂ {0} ⊔ γ and the hyper- ∆γ = {y ∈ T∆γ gγ(y) = 0} is non-degenerate by [17, Proposition 5.3]. Moreover ∼−→ T∆γ by τγ, and hence we obtain C. Finally we define mγ ∈ Z+ as in Section 3. Then in the same surface Z ∗ Z ∗ ∆γ ⊂ T∆γ is invariant by the multiplication lτγ : T∆γ an element [Z ∗ way as [17, Theorem 5.7] and [18, Theorem 4.3] we obtain the following results. ∆γ ] of Mµ MONODROMIES AT INFINITY OF NON-TAME POLYNOMIALS 21 Theorem 5.6. In the situation as above, for any λ /∈ A◦ f,b we have the equality (5.11) [H b f ]λ = χh(S b f )λ = χh((1 − L)mγ · [Z ∗ ∆γ ])λ in K0(HSmon), where in the sumPγ≺Γ+(fi) for l + 1 ≤ i ≤ l + r the face γ of Γ+(fi) ranges through compact ones not contained in Γ◦ +(fi). ✷ Theorem 5.7. In the situation as above, let λ /∈ A◦ Jordan blocks for the eigenvalue λ with sizes ≥ k in Φb f,b and k ≥ 1. Then the number of the n−1 is equal to l+rXi=1 Xγ≺Γ+(fi)  l+rXi=1 Xγ≺Γ+(fi) , ∆γ ]))λ (5.12) (−1)n−1 Xp+q=n−2+k,n−1+k ep,q(χh((1 − L)mγ · [Z ∗ where in the sum Pγ≺Γ+(fi) for l + 1 ≤ i ≤ l + r the face γ of Γ+(fi) ranges through compact ones not contained in Γ◦ +(fi). ✷ By this theorem and the results in [17, Section 2], for λ /∈ A◦ n−1. More precisely it suffices to neglect the compact faces of Γ◦ f,b we immediately obtain the analogues of [17, Theorems 5.9, 5.14 and 5.16] for the λ-part of the Jordan normal form of Φb +(fi) for l + 1 ≤ i ≤ l + r. We omit the details. References [1] Broughton, S. A. "Milnor numbers and the topology of polynomial hypersurfaces", Invent. Math., 92 (1988): 217-241. [2] Danilov, V. I. and Khovanskii, A. G. "Newton polyhedra and an algorithm for computing Hodge- Deligne numbers", Math. Ussr Izvestiya, 29 (1987): 279-298. [3] Denef, J. and Loeser, F. "Motivic Igusa zeta functions", J. Alg. Geom., 7 (1998): 505-537. [4] Denef, J. and Loeser, F. "Geometry on arc spaces of algebraic varieties", Progr. Math., 201 (2001): 327-348. [5] Dimca, A. Sheaves in topology, Universitext, Springer-Verlag, Berlin, 2004. [6] Esterov, A. and Takeuchi, K. "Motivic Milnor fibers over complete intersection varieties and their virtual Betti numbers", Int. Math. Res. Not., Vol. 2012, No. 15 (2012): 3567-3613. [7] Fulton, W. Introduction to toric varieties, Princeton University Press, 1993. [8] Garc´ıa L´opez, R. and N´emethi, A. "Hodge numbers attached to a polynomial map", Ann. Inst. Fourier, 49 (1999): 1547-1579. [9] Guibert, G., Loeser, F. and Merle, M. "Iterated vanishing cycles, convolution, and a motivic analogue of a conjecture of Steenbrink", Duke Math. J., 132 (2006): 409-457. [10] Hotta, R., Takeuchi, K. and Tanisaki, T. D-modules, perverse sheaves, and representation theory, Birkhauser Boston, 2008. [11] Kashiwara, M. and Schapira, P. Sheaves on manifolds, Springer-Verlag, 1990. [12] Kouchnirenko, A. G. "Poly´edres de Newton et nombres de Milnor", Invent. Math., 32 (1976): 1-31. [13] Libgober, A. and Sperber, S. "On the zeta function of monodromy of a polynomial map", Compositio Math., 95 (1995): 287-307. [14] Matsui, Y. and Takeuchi, K. "Milnor fibers over singular toric varieties and nearby cycle sheaves", Tohoku Math. J., 63 (2011): 113-136. [15] Matsui, Y. and Takeuchi, K. "Monodromy zeta functions at infinity, Newton polyhedra and con- structible sheaves", Mathematische Zeitschrift, 268 (2011): 409-439. 22 KIYOSHI TAKEUCHI AND MIHAI TIB AR [16] Matsui, Y. and Takeuchi, K. "A geometric degree formula for A-discriminants and Euler obstructions of toric varieties", Adv. in Math., 226 (2011): 2040-2064. [17] Matsui, Y. and Takeuchi, K. "Monodromy at infinity of polynomial maps and Newton polyhedra, with Appendix by C. Sabbah", to appear in Int. Math. Res. Not. [18] Matsui, Y. and Takeuchi, K. "Motivic Milnor fibers and Jordan normal forms of Milnor mon- odromies", arXiv:1202.5076v1, submitted. [19] Matsui, Y. and Takeuchi, K. "On the sizes of the Jordan blocks of monodromies at infinity", arXiv:1202.5077v1, submitted. [20] N´emethi, A. and Sabbah, C. "Semicontinuity of the spectrum at infinity", Abh. Math. Sem. Univ. Hamburg, 69 (1999): 25-35. [21] N´emethi, A. and Zaharia, A. "On the bifurcation set of a polynomial function and Newton bound- ary", Publ. Res. Inst. Math. Sci., 26 (1990): 681-689. [22] Oda, T. Convex bodies and algebraic geometry. An introduction to the theory of toric varieties, Springer-Verlag, 1988. [23] Oka, M. Non-degenerate complete intersection singularity, Hermann, Paris (1997). [24] Raibaut, M. "Fibre de Milnor motivique `a l'infini", C. R. Acad. Sci. Paris S´er. I Math., 348 (2010): 419-422. [25] Sabbah, C. "Monodromy at infinity and Fourier transform", Publ. Res. Inst. Math. Sci., 33 (1997): 643-685. [26] Sabbah, C. "Hypergeometric periods for a tame polynomial", Port. Math., 63 (2006): 173-226. [27] Siersma, D. and Tibar, M. "Singularities at infinity and their vanishing cycles", Duke Math. J., 80 (1995): 771-783. [28] Steenbrink, J. H. M. "Mixed Hodge structures on the vanishing cohomology", Real and Complex Singularities, Sijthoff and Noordhoff, Alphen aan den Rijn, (1977): 525-563. [29] Steenbrink, J. H. M. and Zucker, S. "Variation of mixed Hodge structure I", Invent. Math., 80 (1985): 489-542. [30] Tibar, M. "Topology at infinity of polynomial mappings and Thom regularity condition", Compositio Math., 111, no.1 (1998), 89-109. [31] Tibar, M. Polynomials and vanishing cycles, Cambridge University Press, 2007. [32] Varchenko, A. N. "Zeta-function of monodromy and Newton's diagram", Invent. Math., 37 (1976): 253-262. [33] Zaharia A. "On the bifurcation set of a polynomial function and Newton boundary II", Kodai Math. J., 19 (1996): 218-233. Institute of Mathematics, University of Tsukuba, 1-1-1, Tennodai, Tsukuba, Ibaraki, 305-8571, Japan. E-mail address: [email protected] Math´ematiques, Laboratoire Paul Painlev´e, Universit´e Lille 1, 59655 Villeneuve d'Ascq, France. E-mail address: [email protected]
1607.00895	1	1607	2016-07-04T14:03:54	Ulrich bundles on surfaces with $q=p_g=0$	[ "math.AG" ]	We prove that any surface with q=p_g=0 embedded by a sufficiently large linear system admits a rank 2 Ulrich bundle. In particular every Enriques surface admits a rank 2 Ulrich bundle.	math.AG	math	ULRICH BUNDLES ON SURFACES WITH q = pg = 0 ARNAUD BEAUVILLE ABSTRACT. We prove that any surface with q = pg admits a rank 2 Ulrich bundle. In particular every Enriques surface admits a rank 2 Ulrich bundle. = 0 embedded by a sufficiently large linear system Let X ⊂ Pn be a complex projective variety of dimension d. A Ulrich bundle on X is a vector bundle E on X satisfying H •(X, E(−1)) = . . . = H •(X, E(−d)) = 0 . This notion was introduced in [ES], where various other characterizations are given; let us just mention that it is equivalent to say that E admits a linear resolution as a OPn -module, or that the pushforward of E onto Pd by a general linear projection is a trivial bundle. In [ES] the authors ask whether every projective variety admits a Ulrich bundle. The answer is known only in a few cases: hypersurfaces and complete intersections [HUB], del Pezzo surfaces [ES, Corollary 6.5], abelian surfaces [B], sufficiently general K3 surfaces [AFO]. We show here that the Lazarsfeld-Mukai construction produces a rank 2 Ulrich bundle on any surface with q = pg = 0 , provided the linear system OS(1) is large enough : Theorem. Let S ⊂ Pn be a surface with q = pg = 0 . Assume that the linear system K −1 irreducible curve. Then S admits a rank 2 Ulrich bundle. S (1) contains an In particular, every Enriques surface S ⊂ Pn admits a rank 2 Ulrich bundle. Proof : Let C be an irreducible curve in K −1 S (1); we have OS(1)C = ωC . Let A be a general line bundle of degree g(C) + 1 on C . Then h0(A) = 2 , h1(A) = 0 , and A is generated by its global sections. Let E be the kernel of the evaluation map H 0(C, A) ⊗ OS ։ A. This is a rank 2 vector bundle on S . From the exact sequence 0 → E → H 0(C, A) ⊗ OS → A → 0 we get H •(S, E) = 0 . Since det E = OS(−C), E(1) is isomorphic to E∗ ⊗ KS , thus H •(S, E(1)) = 0 . Therefore E(2) is a Ulrich bundle for S . Remarks .− 1) It might be that every Enriques surface carries actually a Ulrich line bundle. In [BN] the authors conjecture that Enriques surfaces with no (−2)-curves admit a Ulrich line bundle, and prove that conjecture in a significant case. 2) Minimal surfaces with Kodaira dimension 0 fall into 4 classes, namely K3, Enriques, abelian and bielliptic surfaces. A bielliptic surface S admits a finite ´etale covering π : A → S by an abelian surface A; if L is a very ample line bundle on S and E a Ulrich bundle for (A, π∗L), then π∗E is a Ulrich bundle for (S, L). Thus all minimal surfaces with Kodaira dimension 0 admit a Ulrich bundle, except possibly some special K3 surfaces. 3) Let S ⊂ Pn be a surface with q = pg = 0 , such that OS(1) ∼= K r S , with r ≥ 3 . One can show that the linear system K r−1 S contains an irreducible curve, hence S admits a rank 2 Ulrich bundle. 1 2 ARNAUD BEAUVILLE REFERENCES [AFO] M. Aprodu, G. Farkas, A. Ortega : Minimal resolutions, Chow forms and Ulrich bundles on K 3 surfaces. J. Reine [B] [BN] [ES] [HUB] Angew. Math., to appear. Preprint arXiv:1212.6248. A. Beauville : Ulrich bundlesonabeliansurfaces. Proc. Amer. Math. Soc., to appear. Preprint arXiv:1512.00992. L. Borisov, H. Nuer : Ulrich BundlesonEnriquesSurfaces. Preprint arXiv:1606.01459. D. Eisenbud, F.-O. Schreyer : Resultants and Chow forms via exterior syzygies. J. Amer. Math. Soc. 16 (2003), no. 3, 537-579. J. Herzog, B. Ulrich, J. Backelin : LinearmaximalCohen-Macaulaymodulesoverstrict completeintersections. J. Pure Appl. Algebra 71 (1991), no. 2-3, 187-202. LABORATOIRE J.-A. DIEUDONN ´E, UMR 7351 DU CNRS, UNIVERSIT ´E DE NICE, PARC VALROSE, F-06108 NICE CEDEX 2, FRANCE E-mail address: [email protected]
1603.04480	1	1603	2016-03-14T21:20:22	The Halphen cubics of order two	[ "math.AG" ]	For each $m\ge 1$, Roulleau and Urz\'ua give an implicit construction of a configuration of $4(3m^2-1)$ complex plane cubic curves. This construction was crucial for their work on surfaces of general type. We make this construction explicit by proving that the Roulleau-Urz\'ua configuration consists precisely of the Halphen cubics of order $m$, and we determine specific equations of the cubics for $m=1$ (which were known) and for $m=2$ (which are new).	math.AG	math	The Halphen cubics of order two Thomas Bauer, Brian Harbourne, Joaquim Ro´e, Tomasz Szemberg March 21, 2018 Abstract. For each m > 1, Roulleau and Urz´ua give an implicit con- struction of a configuration of 4(3m2 − 1) complex plane cubic curves. This construction was crucial for their work on surfaces of general type. We make this construction explicit by proving that the Roulleau-Urz´ua configuration consists precisely of the Halphen cubics of order m, and we determine specific equations of the cubics for m = 1 (which were known) and for m = 2 (which are new). Introduction For each n = 3m, we study certain arrangements of 4 3 (n2−3) plane cubic curves; each curve is isomorphic to the Fermat cubic x3 + y3 + z3 (i.e., to T = C/Z[ζ], ζ = e2πi/6). For n = 3, the eight curves in the arrangement were known from invariant theory [2] and give the "inscribed and circumscribed" cubics [1, Proposition 5.2] for the four singular cubics in the Hesse pencil hx3 + y3 + z3, xyzi, but for m > 1 the arrangements are described explicitly here for the first time. These arrangements come from arrangements of 4n2 elliptic curves on the abelian surface T × T , studied by Hirzebruch in [11]. Using a quotient by a finite group action followed by a blow up to resolve singularities, Roulleau and Urz´ua obtain in [14, section 3] corresponding arrangements of 4 3 (n2 − 3) elliptic curves on a rational surface they call H, whose images (see [13, Section 1]) under a birational morphism to P2 give the arrangements of plane cubics we study here. The Roulleau-Urz´ua arrangements on H were crucial for their construction of surfaces of general type in [14]. In [13], Roulleau shows that the corresponding arrangements of plane cubics are interesting for another reason: the Harbourne index of the union Cn of the Acknowledgements: The understanding necessary for this paper grew out of discussions that oc- curred in working groups at two workshops in 2015, namely "Recent advances in Linear series and Newton-Okounkov bodies", February 9–14 at the University of Padua, and "Ideals of Linear Subspaces, Their Symbolic Powers and Waring Problems", February 15–21 at the Mathematisches Forschungsinstitut Oberwolfach. We thank the participants in these workshops and the institu- tions that hosted them. We also thank the University of Freiburg for hosting visits by Bauer and Harbourne to work on this paper with Szemberg in person for a week in the summer of 2015. In addition, the research of Bauer was partially supported by DFG grant BA 1559/6-1, the research of Ro´e was partially supported by MTM 2013-40680-P (Spanish MICINN grant) and 2014 SGR 114 (Catalan AGAUR grant), and the research of Szemberg was partially supported by National Science Centre, Poland, grant 2014/15/B/ST1/02197 Keywords: Automorphisms, abelian surfaces, Halphen cubics, Hesse arrangement, rational surfaces Mathematics Subject Classification (2010): 14C20, 14E05, 14H52, 14J26, 14K12. 2 4 3 (n2 − 3) cubics has limit −4 as n → ∞. We note that no reduced plane curve is yet known with Harbourne index less than or even equal to −4. The arrangements of cubics as given in [13] are not given explicitly; they are described only as images under birational maps. In this paper we construct the same arrangements of cubics by elementary methods directly on P2, which allows us to give explicit equations for these interesting cubics. In the case n = 3, the corresponding eight cubics coincide with the classical Halphen cubics [9, 1, 2, 8]. For n = 3m with m > 1, they are members of four pencils, where each pencil is spanned by two of the eight Halphen cubics. We call these pencils Hesse singular point cubic pencils. (The reason for this name is that these pencils can also be obtained from the singular points of the four singular cubics in the Hesse pencil. Each of these singular cubics has three singular points. Thus each choice of three of the four singular cubics in the Hesse pencil defines 9 points, and these 9 points are the base points of a cubic pencil. The four pencils defined this way are precisely the four pencils obtained from the eight Halphen cubics.) The specific members chosen from each Hesse singular point cubic pencil depend on the 3m-torsion points of T , so we refer to the specific cubics chosen as the Halphen cubics of order m. Our construction relies on classical facts known for the so-called dual Hesse configuration of 9 lines in the plane meeting by threes on 12 points, and on the geometry of the curve T . The Weierstrass function ℘′ is a morphism to P1 of degre 3, which in the case of T is triply ramified at 3 points. For each positive integer m, we denote by P1[n] the images of the n-torsion points of T . As noted above, the Hesse singular point cubic pencils are defined by taking as base points subsets of 9 points among the 12 vertices of the dual Hesse configuration; each pencil has 3 reducible members which are composed of lines of the configuration. We parameterize the pencils so that the reducible members correspond to parameters u ∈ P1 belonging to the branch locus of ℘′. Then the Halphen cubics of order m are defined as the cubics in the Hesse singular point cubic pencils with parameters in P1[3m]. Our first result is the following: Theorem 1 For each n ∈ 3N, let H(n) be the union of all Halphen cubics of order n/3. The singularities of H(n) are: 12 points of multiplicity n2 − 3 at the vertices of the dual Hesse configuration, with n2/3 − 1 triple points infinitely near to them, and (n2 − 3)(n2/3 − 3) quadruple points. As a corollary the Harbourne index of H(n) tends to -4 as n grows. The con- figuration of Halphen cubics therefore behaves like the Roulleau-Urz´ua configura- tion over which it is modelled. Our second goal is to understand the rational map T × T 99K P2 used by Roulleau and Urz´ua to construct their configuration. In The- orem 2.1 we determine the linear series associated to the map, and using the action of the theta group, we prove that both configurations agree: Theorem 2 The curves that form the Roulleau-Urz´ua configuration corresponding to the n-torsion points for n = 3m are the Halphen cubics of m-th order. The paper is organized as follows. In section 1 we recall the classical construction of Hesse line configurations and Halphen cubics, and we prove Theorem 1 along with additional information on the position of the singularities (Theorem 1.5). Section 2 is devoted to the study of the Roulleau-Urz´ua configuration and the map T ×T 99K P2, and culminates in the proof of Theorem 2. 3 1 The Hesse configurations and Halphen cubics Recall the construction of the so-called Hesse configurations (a modern account of this classical subject can be found in the book [7, Section 3.1], see also [1] and references therein for their history and attributions). Given a smooth plane cubic C and a line ℓ joining two flexes of C, the third intersection of ℓ and C is another flex. Altogether, there are 12 such lines of flexes, and at each of the 9 flexes of C exactly 4 of the 12 lines meet. This configuration of lines and points is classically called the Hesse line arrangement (123, 94); it does not depend on the choice of a cubic, i.e., the sets of 9 flex points of any smooth plane cubic are projectively equivalent. A triangle containing all 9 flexes is called a triangle of flexes; there are 4 such triangles which together form the Hesse arrangement. They can be obtained as follows. Fix one of the flexes p0 as the zero for the group law on C; then the 9 flexes of C form the 3-torsion subgroup C[3] ∼= (Z/3Z)2. Each line of flexes ℓ through p0 cuts on C one of its 4 cyclic subgroups of order 3. The two cosets of this subgroup in C[3] correspond to the two lines of flexes which do not meet ℓ on C, which toghether with ℓ form one of the triangles. The polar curve of C with respect to a flex p is the degenerate conic consisting of the tangent TpC and another line, called the harmonic polar of p. If ℓ is a line of flexes, the harmonic polars of the three flexes on ℓ are concurrent, and their point of intersection is the vertex opposite to ℓ in the triangle of flexes to which it belongs. Thus, each of the 9 harmonic polars goes through 4 vertices, one on each triangle of flexes, and at each of the 12 vertices exactly 3 of the 9 lines meet. They form the dual Hesse arrangement (94, 123), which again does not depend on the cubic. The given curve C and each triangle of flexes are cubics through the 9 flex points. It follows that the 9 points of the Hesse configuration are the base points of a pencil of cubics, called the Hesse pencil; all cubics in the pencil have the same flex points, and every plane cubic is projectively equivalent to one of the curves in the pencil. The Hesse pencil has 4 singular members, namely the 4 triangles of flexes. We denote them Tv, Th, Tδ, Tγ. For convenience we also fix the following notations for the whole paper: T = {v, h, δ, γ} will be the set of indices for the triangles; Vv = {v0, v1, v∞} will be the vertices of Tv; and similarly for Vh = {h0, h1, h∞}, Vδ = {δ0, δ1, δ∞} and Vγ = {γ0, γ1, γ∞} (the symbols v, h, δ, γ, and 0, 1, ∞, are chosen to match with constructions to appear later on). Moreover we take V = St∈T Vt to denote the whole set of vertices of the dual Hesse configuration, and for each t ∈ T, Λt = V \ Vt to be the complement in V of the set of vertices of Tt. The group of projective transformations of the plane which preserve the Hesse arrangement (or equivalently the Hesse pencil) is a finite group G216 called the Hesse group. It obviously acts on the set {Tv, Th, Tδ, Tγ}, and the image of the corresponding representation G216 → S4 is the alternating group A4. Since it is not the full S4 group, the order of the indices v, h, δ, γ is not entirely innocuous; we now introduce coordinates in order to be precise, and for later use in the explicit determination of the Halphen cubics. Take C to be the Fermat cubic x3 +y3+z3, and let ε denote a primitive third root of unity. The 9 flex points of C are the points (−1, 1, 0), (−1, ε, 0) and (−1, ε2, 0) and the 6 other points obtained from these by permutation; it is customary to take as generators for the Hesse pencil the Fermat cubic C and the triangle Tv = xyz. 4 Thus the three coordinate points v0 = (1 : 0 : 0), v1 = (0 : 1 : 0), v∞ = (0 : 0 : 1) are three of the vertices of the dual Hesse configuration; the remaining points are h0 = (1 : 1 : 1), δ0 = (ε : 1 : 1), γ0 = (ε2 : 1 : 1), h1 = (1 : ε : ε2), δ1 = (1 : ε : 1), γ1 = (1 : ε2 : 1), h∞ = (1 : ε2 : ε) δ∞ = (1 : 1 : ε) γ∞ = (1 : 1 : ε2) The equations of the harmonic polar lines forming the dual Hesse configuration can be obtained using the coordinates of the points. For example the line through the points v0 and h0 has equation z − y and also goes through δ0 and γ0. We denote by Li,j,k,l the line through vui, huj , δuk and γul, so the line just considered is L0000. Evaluating all collinearities we obtain L0000 = z − y, L1011 = z − x, L∞0∞∞ = y − x, L01∞1 = z − εy, L110∞ = εz − x, L∞110 = y − εx, L0∞1∞ = εz − y L1∞∞1 = z − εx L∞∞01 = εy − x (1) For each t ∈ T, the linear system of cubics through the 9 points in Λt is a pencil Ct, which we call a Hesse singular point cubic pencil. It has three singular members, namely, for each vertex tu of Tt, u ∈ {0, 1, ∞}, the union of the three harmonic polars concurrent at tu is a member of the pencil, which we call Ctu. All nonsingular members of each pencil have j-invariant equal to zero. Halphen showed that the locus of 9-torsion points of all curves in the Hesse pencil consists of two members of these pencils Ch, Cv, Cδ, Cγ (so-called Halphen cubics); the configurations of plane cubics described by Roulleau and Urz´ua consist of members of the same pencils (which we call higher order Halphen cubics). Later on we shall give explicit equations of the Halphen cubics; for this purpose we fix a coordinate u in each pencil Ct, such that {0, 1, ∞} correspond to the singular members. Since the harmonic polars concurrent at tu, in the notation above, are the Luvuhuδuγ with ut = u, the three singular members of Cv are Cv0 = ε2Y L0∗∗∗ = ε2(z − y)(z − εy)(εz − y) = z3 − y3 Cv1 = ε2Y L1∗∗∗ = ε2(z − x)(εz − x)(z − εx) = z3 − x3 Cv∞ = ε2Y L∞∗∗∗ = ε2(y − x)(y − εx)(εy − x) = y3 − x3 where the factor ε2 serves only a simplification purpose. Similarly, the three singular memebers of Ch are − x2y + xy2 + x2z − y2z − xz2 + yz2 Ch0 = Y L∗0∗∗ = (z − y)(z − x)(y − x) = Ch1 = −Y L∗1∗∗ = −(z − εy)(εz − x)(y − εx) = Ch∞ = Y L∗∞∗∗ = (εz − y)(z − εx)(εy − x) = ε2(x2y − ε2xy2 − ε2x2z + y2z + xz2 − ε2yz2) − ε(x2y − εxy2 − εx2z + y2z + xz2 − εyz2) where the signs are chosen so that Ch0 + Ch∞ = Ch1. Finally, the singular members of Cδ and Cγ are 5 − x2y + εxy2 + x2z − ε2y2z − εxz2 + ε2yz2 ε2(x2y − xy2 − ε2x2z + ε2y2z + εxz2 − εyz2) − ε(x2y − ε2xy2 − εx2z + ε2y2z + εxz2 − yz2) Cδ0 = Y L∗∗0∗ = (z − y)(εz − x)(εy − x) = Cδ1 = −Y L∗∗1∗ = −(z − εy)(z − εx)(y − x) = Cδ∞ = Y L∗∗∞∗ = (εz − y)(z − x)(y − εx) = Cγ,0 = Y L∗∗∗0 = (εz − y)(εz − x)(y − x) = Cγ,1 = −Y L∗∗∗1 = −(z − εx)(y − εx)(z − y) = Cγ,∞ = Y L∗∗∗∞ = (z − εy)(z − x)(εy − x) = − x2y + xy2 + εx2z − εy2z − ε2xz2 + ε2yz2 ε2(x2y − ε2xy2 − x2z + εy2z + ε2xz2 − εyz2) − ε(x2y − εxy2 − ε2x2z + εy2z + ε2xz2 − yz2) With these choices, the member of the pencil Ct corresponding to the parameter u ∈ C will be Ctu = Ct0 + u Ct∞ for every t ∈ T. Halphen's Theorem. Fixing any of the flex points of a plane cubic curve C as the zero point in the group law, the set of n-torsion points for n = 3m coincides with the set of points p such that there exists a (possibly reducible) curve of degree m meeting C only at p; thus the set of 3m-torsion points of a cubic is well defined and independent of the choice of a flex. Halphen [9] studied the locus of points of 3m-torsion of all cubics in the Hesse pencil. For m = 1 it consists on the 9 base points, as already said. For m = 2, it is the union of the 9 harmonic polars. For m = 3 it is the union of 8 cubics, two in each pencil Cu, which we shall describe next. For m > 3, it is the union of 8 or 9 curves of higher degrees, depending on the divisibility of m by 3. Lemma 1.1 Let T be an elliptic curve and f : T → P1 a morphism of degree 3 with triple ramification at 3 points. Then the j-invariant of T is 0, and f is unique up to translation and inversion on T , and up to automorphisms of P1. Proof. By the Riemann-Hurwitz formula, there is no more ramification than the three triple points, so E is a 3 sheeted cover of P1 away from these points. Permuting the sheets of the cover induces an automorphism of E of order 3. But the only elliptic curves with an order 3 automorphism are those of j-invariant 0, and these support a single order 3 automorphism (up to translation and inversion), namely complex multiplication by a cube root of 1 [15, Theorem 10.1]. (cid:3) The automorphisms of P1 which appear in lemma 1.1 must obviously preserve the branch locus of f . To take care of these we call marked line a pair P1 M = (P1, M ) where M = {p, q, r} is a set of 3 distinct points in P1. A morphism of marked lines is an algebraic morphism which preserves the markings (and so any two marked 6 lines are isomorphic). The group of automorphisms of the marked line is finite, isomorphic to S3, and by choosing a coordinate u on P1 such that the marking is M = {0, 1, ∞}, it is generated by the morphisms u 7→ u−1 and u 7→ 1 − u. Then, given an equianharmonic elliptic curve T and a marked line P1 M , there exixts a morphism f : T → P1 of degree 3, with triple ramification over M , unique up to translation and inversion on T , and up to the action of S3 on P1 M . In classical terminology, elliptic curves with j-invariant equal to zero are called In the setting of lemma 1.1, consider the grup law on T with equianharmonic. the zero at one of the ramification points. Such a curve T is obtained as T = C/Z[ζ], where ζ = e2πi/6; a degree 3 map is given by the Weierstrass derivative function ℘′ : T → P1; and the corresponding order three automorphism is induced by multiplication by ε = ζ 2. Its 3 fixed points, which are the ramification points of ℘′, are pk = k 3 ζ for 0 6 k 6 2. Their images form the branch locus M = branch(f ) = {i, −i, ∞} ⊂ P1. 3 + k For every choice of the zero at one of the ramification points of f , the other two ramification points are of 3-torsion (in fact they form a cyclic subgroup of order 3, as can be seen in the description via the ℘′ function) hence for every positive integer m the set T [3m] of 3m-torsion points on T independent on which ramification point is chosen as zero. Remark 1.2 Since the permuting of the 3 sheets is an automorphism, if a (non- ramification) point in T is of 3m-torsion, then the remaining two points in its fiber are 3m-torsion as well. So by lemma 1.1 the set of images P1 M [3m] = f (T [3m]) \ M (where M = branch(f ) is the branch locus) is a well defined set (independent of any choices and invariant under isomorphisms of marked lines) consisting of ((3m)2 − 3)/3 = 3m2 − 1 distinct points. We call P1 M [3m] the set of equianharmonic 3m- torsion parameters relative to the markings M , or simply the set of equianharmonic 3m-torsion parameters (denoted P1[3m]) if M = {0, 1, ∞}. Corollary 1.3 Let T be an elliptic curve and f : T → P1 a morphism of degree 3 with triple ramification at 3 points, and let M = branch(f ). The set f −1(M ∪ P1 M [3m]) is a translate of the subgroup of 3m-torsion. We will be interested in the explicit determination of P1[3m]. The first case cor- responds to 3-torsion points, m = 1, for which there are 3·12 −1 = 2 equianharmonic torsion parameters. Invariance under the action of S3 is enough to determine P1 M [3] as the Hessian pair of the marking M , which can also be characterized as the two points which together with M form an equianharmonic set (i.e., with cross ratio a cube root of 1). For M = {0, 1, ∞} one gets P1[3] = {−ε, −ε2} (this is the only set of 2 points invariant for both t 7→ t−1 and t 7→ 1 − t). P1[6], although more involved, can be computed using the S3-action as well, but this is not the case for higher m. Below we compute P1[6] from the definition, a method that does generalise to all m. Theorem 1.4 (Halphen, [9, §3]) Consider each pencil Ct, t ∈ T as a marked line, where the marking consists of the three singular members Ctu with u ∈ {0, 1, ∞}. Denote Ct[3] the set of two cubics in the pencil that correspond to equianharmonic 3-torsion parameters. Then Ch[3] ∪ Cv[3] ∪ Cδ[3] ∪ Cγ[3] 7 is the locus of 9-torsion points of curves in the Hesse pencil. Using the fact that P1[3] = {−ε, −ε2}, the 8 Halphen cubics can be explicitly written as follows. Cv−ε = x3 + ε2y3 + εz3 Ch−ε = x2y + y2z + xz2 Cδ−ε = x2y + ε2y2z + εxz2 Cδ−ε2 = xy2 + ε2x2z + εyz2 Cγ−ε = x2y + εy2z + ε2xz2 Cγ−ε2 = xy2 + εx2z + ε2yz2 Cv−ε2 = x3 + εy3 + ε2z3 Ch−ε2 = xy2 + x2z + yz2 This same list also arises in a somewhat different way in [2, 8] and [1, Proposition 5.2], where a modern account of Halphen's theorem is given. In addition to computing the 9-torsion of members of the Hesse pencil, the Halphen cubics are special in their Hesse singular point cubic pencils, in the following ways. First, the base points of the pencil are a translate (with respect to the group law of the Halphen cubic) of its set of 3-torsion points; and second, they intersect one another only in base points, tangently in sets of three (from distinct pencils). It is these latter properties that we seek to generalize in higher order Halphen cubics. Denote Ct[3m] the union of the 3m2 − 1 cubics in the pencil Ct corresponding to equianharmonic 3m-torsion parameters, and call them the Halphen cubics of order m. In order to better describe their intersections we consider the blow up X → P2 at the 12 points of V , and denote Ct[3m] the strict transforms. Denote L the pullback to X of the class of a line, Etu the exceptional divisor above the point tu, E the sum of all 12 exceptionals, and Et = Xt′ u∈Λt Et′ u = Xt′6=t u∈{0,1,∞} Et′ u the divisor above the 9 point set Λt; thus each Hesse singular point cubic pencil can be written Ct = 3H − Et. Each exceptional divisor Etu carries a natural marking Mtu = Etu ∩ Ctu (2) consisting of the directions of the three lines in the dual Hesse configuration going through the point tu (these are the component lines of Ctu). In the sequel, unless ex- plicitly specified, the equianharmonic torsion points on Etu will always be considered with respect to the natural marking, and we denote them Etu[3m] = (Etu )Mtu [3m]. Note that, for each t′ 6= t we have (3) Etu ∩(cid:16) Ct′ 0 ∪ Ct′ 1 ∪ Ct′ ∞(cid:17) = Etu ∩ Ctu = Mtu . Theorem 1.5 The reducible curve Hm = Ch[3m] ∪ Cv[3m] ∪ Cδ[3m] ∪ Cγ[3m] belongs to the linear system 12(3m2 −1)L−9(3m2−1)E, and has the following singularities: 1. 3m2 − 1 ordinary triple points on each Etu , for a total of 12(3m2 − 1) triple points; 2. 9(3m2 − 1)(m2 − 1) ordinary quadruple points off E. Moreover, each component of the curve passes through 9 of the triple points and 9(m2 − 1) of the quadruple points, which together constitute a translate of its 3m- torsion subgroup. 8 Proof. The linear equivalence class is clear from the fact that Ct[3m] ∼ (3m2 − 1) (3L − Et) . Each pencil Ct induces an elliptic fibration φt : X → P1 for which the 9 ex- ceptional components above Λt are sections. For every exceptional component Etu , three of the fibrations have it as a section (those φt′ with t′ 6= t), and for the remaining fibration φt, it is a component of a fiber, with multiplicity 3. Indeed, t (φt(Ctu )) = Ctu + 3Etu φ−1 (4) because Ctu has multiplicity 3 at tu. Any two pencils among the 4 share 6 base points, and hence have intersection number 32 − 6 = 3. The restriction of φh, φδ and φγ to any smooth curve Cvs, s ∈ C\{0, 1} in the pencil Cv gives therefore a morphism of degree 3, φt Cvs : Cv,s → Ev0 . Since tu is a base point of Cv for each t 6= v and u ∈ {0, 1, ∞} Cvs meets Etu at a point p = ptu and by (4), (cid:16)φt Cvs(cid:17)−1 (Ctu ∩ Ev0) = Chs ∩(cid:16) Ctu + 3Etu(cid:17) > 3p, but since the degree of the map is 3, the inequality must in fact be an equality (so Chs meets no point on Ctu and the intersection with Etu is transversal). This holds for all has triple ramification above Mh0. As the action of the Hesse u ∈ {0, 1, ∞}, so φt Chs group on T is 2-transitive, for every t 6= t′ the restriction of φt to a fiber Ct′ s has triple has no additional ramification (by Riemann- ramification above Mt′ . Since φtCt′ Hurwitz) all intersections between Ct′ s and nonsingular fibers of φt are transversal. Therefore, intersections between components of Hm (which are nonsingular fibers of the pencils) are transversal; this means that all multiple points of Hm, which are generated by such intersections, are ordinary. Moreover, by corollary 1.3, 0 s s (cid:16)Et′ 0 0(cid:17) Ct′ s ∩ ( Ct[3m] ∪ Et0 ∪ Et1 ∪ Et∞ ) = φt−1 Ct′ [3m] ∪ Mt′ consists of the 3m-torsion points of Ct′ s up to translation, for each t′ 6= t. By construction, for each t 6= t′, the intersection Ct[3m] ∩ Et′ consists of the equianharmonic 3m-torsion parameters Et′ [3m] there are three components of Hm, one in each pencil Ct, t 6= t′. These points are therefore triple points of Hm, and there are 3m2 − 1 such points on each of the 12 exceptional components. [3m], i.e., through each point of Et′ 0 0 0 Taking into account the linear equivalence class of the Ct and the intersection product on X, we see that besides the triple points, each pair Ct[3m], Ct′[3m] intersect at 9(3m2 − 1)(m2 − 1) additional points. The proof will be complete by showing that these belong to the two remaining Ct′′[3m]'s. We accomplish this by proving that, for every component Cvs of Cv[3m], the sets Ah = Cvs ∩ ( Ch[3m] ∪ Eh0 ∪ Eh1 ∪ Eh∞) Aδ = Cvs ∩ ( Cδ[3m] ∪ Eδ0 ∪ Eδ1 ∪ Eδ∞) are equal; then Cv[3m] ∩ Ch[3m] \ E = Cv[3m] ∩ Cδ[3m] \ E and by symmetry all pairs Ct[3m], Ct′[3m] intersect at the same set of 9(3m2 − 1)(m2 − 1) points, which finishes the proof. 9 Indeed, by corollary 1.3 both Ah and Aδ are obtained from the set of 3m-torsion points of Cvs by suitable translations according to the group law in Cvs, therefore Ah = tp(Aδ) for some point p ∈ Cvs. Since Cvs meets Eγ0 at one of the triple points, which must also belong to Ch[3m] and Cδ[3m], it follows that Ah ∩ Aδ is nonempty. Therefore p is of 3m-torsion, and Ah = Aδ as claimed. (cid:3) Explicit computation of the equianharmonic torsion parameters. Our method to compute the higher order Halphen cubics geometrically is based on the following remark. Remark 1.6 The plane cubic curve C = x3 + y3 − z3 is a j-invariant 0 curve, and each of the three lines L1∗∗∗ in the dual Hesse configuration going through the point v1 = (0, 1, 0) is a flex line for C (i.e., tangent to C at a flex point). For brevity, in this section we denote these lines simply L1 = −ε2(z − x), L0 = εz − x, L∞ = z − εx (With respect to the equations L1∗∗∗ above, a product with adequate constant co- efficients was done so that L0 + L∞ = L1). Thus the linear series on C given by the pencil of lines through v1 defines a morphism C → Ev1 with triple ramification above the points corresponding to the directions of the Lu, which form exactly the set Mv1; and the equianharmonic torsion parameters (with respect to Mv1) can be computed as the projections to Ev1 of the torsion points on C. Once the n-torsion points on C are known, their projections to Ev1 (which means their (x, z) coordinates) are the equianharmonic n-torsion parameters. The torsion points can be found in principle solving algebraic equations involving so-called di- vision polynomials [12], so the method works uniformly for all n. In this section we will determine the equianharmonic 6-torsion parameters, where we can find the required 6-torsion points using a more geometric method. We then produce the 44 Halphen cubics of order 2. These include the 8 order 1 cubics, so we have 36 still to find. Lemma 1.7 The 6-torsion points of C are obtained by adding (using the group law on C) each of the nine points (1, 0, 1), (1, 0, ε), (1, 0, ε2 ), (1, −1, 0), (1, −ε, 0), (1, −ε2 , 0), (0, 1, 1), (0, 1, ε), (0, 1, ε2 ) to each of the four points (1, 0, 1), (1, −b, −1), (1, −εb, −1), (1, −ε2 b, −1) where b3 − 2 = 0. Proof. The 6-torsion points can be obtained by adding 3-torsion points and 2-torsion points. The 3-torsion points are the flex points, which as noted above, are the nine points given. As for the 2-torsion points, note first that (1, 0, 1) is a flex. Regarding it as the identity for the group law on C, the 2-torsion points on C are the lines tangent to C which go through the identity (i.e., through (1, 0, 1)). One can check that the required points are the four points given. Using the group law on C one can now find all 36 of the 6-torsion points. (cid:3) Projections from 6-torsion points. Three of the 6-torsion points are the ram- ification points of the projection, and 6 of them correspond to the n = 3 case. The other 27 come from taking a line through a 2-torsion point and one of the nine 3-torsion points. One then finds the line through each of these 27 points and the point v1. This gives 9 lines through v1: 10 x − (1/2)b2z, x − bz, x + z, x − εbz, x + εz, x − (1/2)εb2z, x − (1/2)ε2b2z x − ε2bz x + ε2z The ramification points map to the lines Li, and the lines above (up to product with a constant, in the same order) can be written as: (bε − 1)L0 − ε(bε2 − 1)L∞, (bε2 − 1)L0 − ε2(bε − 1)L∞, (b − 1)L0 − ε(bε − 1)L∞, (b − 1)L0 − ε2(bε2 − 1)L∞, (bε2 − 1)L0 − ε(b − 1)L∞ (bε − 1)L0 − ε2(b − 1)L∞ L0 − L∞, 2L0 + L∞, L0 + 2L∞ Now the equianharmonic torsion parameters P1[6] \ P1[3] can be obtained as the ratios between the coefficients of L∞ and L0 in the previous list. In the same order again: , τ = −ε bε2 − 1 bε − 1 τ −1 = −ε2 bε − 1 bε2 − 1 bε − 1 b − 1 (1 − τ )−1 = −ε , (1 − τ −1)−1 = −ε2 bε2 − 1 b − 1 , − 1, 1 2 , 1 − τ −1 = −ε b − 1 bε2 − 1 1 − τ = −ε2 b − 1 bε − 1 , , , 2 As an aside we note that 3 of them are defined over Q; they form one orbit under the action of S3 generated by t 7→ 1 − t and t 7→ t−1. The remaining 6 form another orbit, which consists of the roots of the irreducible invariant polynomial P (x) = x6 − 3x5 + 5x3 − 3x + 1 = (x2 − x − 1)3 + 2, with the property that Q[τ ] = Q[ε, b] is the splitting field of x3 + 2 (see [6, page 59]). The 36 new Halphen cubics. Now here are the 36 cubics we get, normalized to obtain a simple expression (i.e., the polynomial Cvτ as given on the list is actually a scalar multiple of Cv0 + τ Cv∞). Cvτ = (b − ε)x3 + ε2(b − 1)y3 + (bε − 1)z3 Chτ = x2y − ε2bxy2 − ε2bx2z + y2z + xz2 − ε2byz2 Cδτ = x2y − bxy2 − ε2bx2z + ε2y2z + εxz2 − εbyz2 Cγτ = x2y − ε2bxy2 − bx2z + εy2z + ε2xz2 − εbyz2 Cv(1−τ )−1 = ε(bε − 1)x3 + ε(b − ε)y3 + (b − 1)z3 Ch(1−τ )−1 = x2y − εbxy2 − εbx2z + y2z + xz2 − εbyz2 Cδ(1−τ )−1 = x2y − ε2bxy2 − εbx2z + ε2y2z + εxz2 − byz2 Cγ(1−τ )−1 = x2y − εbxy2 − ε2bx2z + εy2z + ε2xz2 − byz2 11 Cv1−τ −1 = ε(b − 1)x3 + (b − ε2)y3 + (bε2 − 1)z3 Ch1−τ −1 = x2y − bxy2 − bx2z + y2z + xz2 − byz2 Cδ1−τ −1 = x2y − εbxy2 − bx2z + ε2y2z + εxz2 − ε2byz2 Cγ1−τ −1 = x2y − bxy2 − εbx2z + εy2z + ε2xz2 − ε2byz2 Cvτ −1 = (b − ε2)x3 + ε(b − 1)y3 + (bε2 − 1)z3 Chτ −1 = 2x2y − ε2b2xy2 − ε2b2x2z + 2y2z + 2xz2 − ε2b2yz2 Cδτ −1 = 2x2y − b2xy2 − ε2b2x2z + 2ε2y2z + 2εxz2 − εb2yz2 Cγτ −1 = 2x2y − ε2b2xy2 − b2x2z + 2εy2z + 2ε2xz2 − εb2yz2 Cv(1−τ −1)−1 = ε(b − ε)x3 + ε(bε − 1)y3 + (b − 1)z3 Ch(1−τ −1)−1 = 2x2y − εb2xy2 − εb2x2z + 2y2z + 2xz2 − εb2yz2 Cδ(1−τ −1)−1 = 2x2y − ε2b2xy2 − εb2x2z + 2ε2y2z + 2εxz2 − b2yz2 Cγ(1−τ −1)−1 = 2x2y − εb2xy2 − ε2b2x2z + 2εy2z + 2ε2xz2 − b2yz2 Cv(1−τ )−1 = ε2(b − 1)x3 + (b − ε)y3 + (bε − 1)z3 Ch(1−τ )−1 = 2x2y − b2xy2 − b2x2z + 2y2z + 2xz2 − b2yz2 Cδ(1−τ )−1 = 2x2y − εb2xy2 − b2x2z + 2ε2y2z + 2εxz2 − ε2b2yz2 Cγ(1−τ )−1 = 2x2y − b2xy2 − εb2x2z + 2εy2z + 2ε2xz2 − ε2b2yz2 Cv−1 = x3 − 2y3 + z3 Ch−1 = x2y + ε2xy2 + ε2x2z + y2z + xz2 + ε2yz2 Cδ−1 = x2y + xy2 + ε2x2z + ε2y2z + εxz2 + εyz2 Cγ−1 = x2y + εxy2 + x2z + εy2z + ε2xz2 + εyz2 = x3 + y3 − 2z3 = x2y + εxy2 + εx2z + y2z + xz2 + εyz2 = x2y + ε2xy2 + εx2z + ε2y2z + εxz2 + yz2 = x2y + εxy2 + ε2x2z + εy2z + ε2xz2 + yz2 Cv 1 2 Ch 1 2 Cδ 1 2 Cγ 1 2 Cv2 = 2x3 − y3 − z3 Ch2 = x2z + xz2 + x2y + z2y + xy2 + zy2 Cδ2 = x2y + εxy2 + x2z + ε2y2z + εxz2 + ε2yz2 Cγ2 = x2y + xy2 + εx2z + εy2z + ε2xz2 + ε2yz2 12 As explained in the proof of Theorem 1.5, the singular points of the configuration Hm are exactly the 3m-torsion points on each of its components, translated by one (arbitrary) base point of the pencil to which it belongs. In the case of H2, these 6-torsion points can be computed for each Cts by the same method above; we leave the details to the interested reader. 2 The linear series of the Roulleau-Urz´ua map Recall now the construction by Roulleau and Urz´ua of their planar configuration of cubics (which we will eventually show to be equal to the Halphen cubics of order m). Let as before T = C/Z[ζ], where ζ = e2πi/6, be the equianharmonic elliptic curve, and let pi = i 3 ζ for 0 6 i 6 2 be the 3 fixed points of multiplication by ζ 2. Let A be the abelian surface A = T × T , and denote σ : A → A the induced automorphism defined by (x, y) 7→ (ζ 2x, ζ 2y), which has 9 fixed points, namely 3 + i pij = ( i 3 + i 3 ζ, j 3 + j 3 ζ) (0 6 i, j 6 2), so that pij = (pi, pj). Divisors of particular importance on A are V = 0 × T , H = T × 0, the diagonal ∆ and the graph Γ of the complex multiplication by ζ. In fact, these curves span the N´eron-Severi group of A. Translating V, H, ∆, Γ by the fixed points pij gives twelve curves: each of the fixed points is on four of the translates and each translate contains 3 of the fixed points, as suggested in Figure 1. Of the diagonal lines, only ∆ and Γ can be shown properly as going through three of the points pij, but the line through points p01 and p10 also goes through point p22, and in general if (i1, j1) + (i2, j2) + (i3, j3) add up as vectors in Z2 3 to (0, 0), then the points pi1j1, pi2j2 and pi3j3, are collinear. Moreover, the points of intersection actually occur only at the points pij, contrary to how it might look in the drawing. p21 Γ p20 H p01 ∆ p00 V p11 p10 p22 p02 p12 Figure 1: The curves V, H, ∆, Γ ⊂ A and their translates by the fixed points pij Now let B → A be the blow up at the nine points pij. Since these are fixed points for σ, the automorphism lifts to B. By a slight abuse of notation, we denote the lift of σ to B again by σ (and we denote the total transforms of V, H, ∆, Γ to 13 B with the same letters). Thus we have the quotient B → X = B/ hσi. Moreover, because σ acts diagonally on A, σ fixes tangent directions at p00 (and hence also at each fixed point pij), so the fixed points for σ acting on B are exactly the points of the exceptional curves for the nine points pij. Thus the ramification locus for the quotient map B → X is the union of these nine exceptional curves. Roulleau and Urz´ua show in [14] that X is smooth and rational, and that un- der the quotient B → X the images of the 12 curves obtained from V, H, ∆, Γ by translation are disjoint (−1)-curves whose contraction gives a birational morphism X → P2, representing X as the blow-up of P2 at the twelve points of the dual Hesse configuration. So we have a diagram B = Bl9(A) −→ A = T × T ϕ ց y X = Bl12(P2) −→ P2 where the vertical map is of degree 3 and its branch locus is the union of the nine exceptional curves for the upper horizontal map, whose images are the 9 harmonic polar lines of the dual Hesse configuration. In this section we describe the induced map ϕ : B → P2 in terms of linear series. Then, using the action of the theta group, we determine the coordinates of the images of the 12 translates of V, H, ∆, Γ, which will justify the choice of indices v, h, δ, γ in the previous section. Theorem 2.1 a) The morphism ϕ : B → P2 is the map ϕL defined by the complete linear series L associated with the line bundle L = V + H + ∆ + Γ − E where E is the sum of the nine exceptional divisors Eij of the blow-up B → A. b) Consider the translates Vi = V + (pi, 0) = V + pi0. Then there are elliptic curves N01, N02, N12 ⊂ A such that the divisors D0 := V1 + V2 + N12 D1 := V0 + V2 + N02 D2 := V0 + V1 + N01 belong to the linear series L. If we define the map ϕL : B → P2 by suitably scaled sections corresponding to the divisors D0, D1, D2, then the images of the 12 translates of V, H, ∆, Γ are the 12 Hesse dual points vu, hu, δu, γu. The proof of Theorem 2.1 is split into several intermediate steps filling the rest of the present section. We start by showing: Proposition 2.2 The morphism ϕ : B → P2 is the map defined by the complete linear series V + H + ∆ + Γ − E where E is as above. 14 Proof. With respect to the blow up B → A, the proper transforms of the curves V, H, ∆, Γ are V ′ = V − E02 − E00 − E01, H ′ = H − E20 − E00 − E10, ∆′ = ∆ − E22 − E00 − E11 and Γ′ = Γ − E21 − E00 − E12, where Eij is the exceptional curve for the blow up of pij. These curves are mutually disjoint and meet E00 transversely. Since V ′, H ′, ∆′ and Γ′ are preserved curvewise by σ and E00 is fixed pointwise, the images V ′′, H ′′, ∆′′ and Γ′′ of V ′, H ′, ∆′ and Γ′ under the quotient B → X are 00 of E00 transversely. Since V ′′, H ′′, ∆′′ and Γ′′ are disjoint and meet the image E′ exceptional curves which map to points under X → P2, E′ 00 maps to a smooth plane 00 + V ′′ + H ′′ + ∆′′ + Γ′′)2 = rational curve C, hence of self-intersection C 2 = (E′ (E′ 00)2, and has triple ramification along E00, so ϕ∗(E′ 00)2, so (E′ 00)2 + 4(2) + 4(−1). But B → X is a triple cover, so (ϕ∗(E′ 00) = 3E00. Thus −9 = (3E00)2 = 3(E′ 00)2 = −3 and C 2 = 1, hence C is a line. 00))2 = 3(E′ The pullback of C to B is V ′ + H ′ + ∆′ + Γ′ + 3E′ 5 = V + H + ∆ + Γ − E, which we denote by L. I.e., the map ϕ is defined by a 3 dimensional subspace of H 0(B, L), and the argument so far shows that L = V + H + ∆ + Γ − E. We will now prove that h0(B, L) = 3, which then implies that ϕ is defined by the complete linear series L. First, we have (V + H + ∆ + Γ)2 = 12, and therefore by Riemann-Roch on A we get h0(A, V + H + ∆ + Γ) = 6. It is therefore enough to find three fixed points q1, q2, q3 of σ that impose independent conditions on V + H + ∆ + Γ, i.e., such that there is a divisor in the linear series V + H + ∆ + Γ passing through q1 and q2, but not through q3 (this suffices since the divisor V + H + ∆ + Γ is very ample by [4, Theorem 2.3]). Consider to this end the point x = ( 1 3 , 0) on A. Lemma 2.3 implies that there is a point z ∈ A such that the divisor t∗ z(∆ + Γ) belongs to the linear series V + H + ∆ + Γ. Let q1 and q2 be any two of the three fixed points lying on H. Clearly none of the nine fixed points lies on t∗ xV , and there can be at most five of them on t∗ z(∆ + Γ). Therefore there exists a fixed point q3 that lies neither on H nor on t∗ z(∆ + Γ). The triple of points q1, q2, q3 thus satisfies the required condition. (cid:3) xV + H + t∗ Lemma 2.3 For every pair of points x, y ∈ A there exists a unique point z ∈ A such that t∗ xV + t∗ yH + t∗ z(∆ + Γ) ≡lin V + H + ∆ + Γ The analogous statement holds for any permutation of the curves V, H, ∆, Γ. Proof. Consider the homomorphism of groups Φ : A × A × A → Pic0(A) xV + t∗ (x, y, z) 7→ t∗ yH + t∗ z(∆ + Γ) − (V + H + ∆ + Γ) For every pair (x, y) ∈ A × A, the map Φ(x, y, ·) is a translate of the canonical homomorphism A → Pic0(A), z 7→ t∗ z(∆ + Γ) − (∆ + Γ), associated with the line bundle ∆ + Γ. Since this line bundle is of self-intersection 2, it gives a principal polarization and therefore its canonical homomorphism is in fact an isomorphism (see [5, Prop. 2.4.9]) and thus the intersection ker Φ ∩ ({(x, y)} × A) consists of exactly one point. (cid:3) Proposition 2.4 The map ψ that assigns to given points x, y ∈ A the point z as in the preceding lemma is given by ψ : A × A → A ((x1, x2), (y1, y2)) 7→(cid:0) − 2x1 − (1 + ζ)y2, −(1 + ζ)x1 − 2y2(cid:1) 15 Remark 2.5 For the special case where x and y are among the nine fixed points of σ, we get with a calculation In other words, we have ψ(pij, pkl) = pil t∗ pij V + t∗ pklH + t∗ pil(∆ + Γ) ≡lin V + H + ∆ + Γ Proof of the proposition. For a line bundle M on A denote as usual by φM the canonical homomorphism A → Pic0(A), x 7→ t∗ xM − M . The point z = ψ(x, y) is characterized by the condition t∗ z(∆ + Γ) ≡lin V + H + ∆ + Γ, which is equivalent to φV (x) + φH(y) + φ∆+Γ(z) = 0. This in turn implies that yH + t∗ xV + t∗ ψ(x, y) = φ−1 ∆+Γ(cid:16) − φV (x) − φH(y)(cid:17) (5) The issue therefore is to explicitly determine the canonical maps. As V + H gives a principal polarization, φV +H is an isomorphism, and hence we can use its inverse to identify Pic0(A) with A. In that sense, we will, by slight abuse of notation, denote φ−1 V +H−→ A = T × T again by φM . φM−→ Pic0(A) the composed homomorphism T × T = A In this setup, φV and φH are given by the matrices (cid:16)1 0 0 0(cid:17) and (cid:16)0 0 0 1(cid:17) respectively. We now determine the map φ∆+Γ in these terms. Consider to this end the isomorphism g : T × T → T × T , (x, y) 7→ (x, y − x). The analytic representation of g and its dual map g are We have g−1(H) = ∆, thus 0 (cid:16) 1 −1 1(cid:17) and (cid:16)1 −1 1 (cid:17) 0 φ∆ = gφHg =(cid:16) 1 −1 1 (cid:17) −1 We can proceed in the analogous way for φΓ using the isomorphism h : T ×T → T ×T , (x, y) 7→ (x, y − ζx). The analytic representations of h and h are Since h−1(H) = Γ, we get (cid:16) 1 −ζ 0 1 (cid:17) 1(cid:17) and (cid:16)1 −ζ 0 In conclusion we find −ζ φΓ = hφHh =(cid:16) 1 −ζ 1 (cid:17) φ∆+Γ =(cid:16) 2 2 (cid:17) −1 − ζ −1 − ζ The assertion follows now from (5) using the matrices we just found. (cid:3) Lemma 2.6 The divisor H + ∆ + Γ − V is numerically equivalent to an elliptic curve N . We have φN = 1 −1 − ζ −1 − ζ 3 ! 16 Proof. The line bundle H + ∆ + Γ − V has self-intersection 0 and it has positive intersection with the ample bundle ∆ + Γ. This implies that its numerical class belongs to a sum of numerically equivalent elliptic curves. As its intersection with H is 1, it is in fact the class of a single elliptic curve. The second assertion follows from the equation φN = φH + φ∆ + φΓ − φV upon using the explicit matrix representations of the maps that were worked out above. (cid:3) The following statement can be useful in understanding the map B → P2, or in the construction of a basis of H 0(B, L). Lemma 2.7 Consider the line bundle M = V + H + ∆ + Γ on A. All nine fixed points of the automorphism σ = (ζ 2, ζ 2) are contained in the kernel K(M ) of φM . In other words, if D ∈ M , then t∗ xD ∈ M for every x in Fix(σ) Proof. From the equation φM = φV + φH + φ∆ + φΓ we get φM = 3 −1 − ζ −1 − ζ 3 ! and one checks that φM · pij is contained in (Z + Zζ) × (Z + Zζ) for every i and j. (cid:3) Preimages of lines. As we know, the three translates Vi = V + pi0 map to points in P2. We would like to see the curves on A which correspond to the lines ℓij through any two of those points. As the preimage of ℓij contains Vi and Vj, we have V + H + ∆ + Γ = Vi + Vj + Nij, where the residual curve Nij is an elliptic curve (by Lemma 2.6). Its intersection numbers with the generators are Nij · V = 3, Nij · H = 1, Nij · ∆ = 1, Nij · Γ = 1 On the other hand, every elliptic curve on A that passes through the origin arises as the image of a homomorphism T → A, x 7→ (ax + bζx, cx + dζx) for suitable integers a, b, c, d (see [10]). Using the method from [3, Sect. 4.2] one finds that the elliptic curve N corresponding to (a, b, c, d) = (1, 1, 0, 1), i.e, the image of the map x 7→ (x + ζx, ζx) has the same intersection numbers as Nij and is therefore numerically equivalent to Nij. So Nij can be obtained from N by a translation – and we determine now explicitly such a translation. The idea is this: We know that the divisor Vi + Vj + Nij passes through all 9 points pij. Since Vi and Vj cover 6 of them, Nij must pass through the remaining 3. Now, a computation shows that the intersection points of N and V0 are p00, p01, p02. This implies that N = N12. The other cases are obtained via translation by suitable fixed points – altogether we have N12 = N N02 = N + p10 = t∗ N01 = N + p20 = t∗ −p10N = t∗ −p20N = t∗ p20N p10N 17 The images of the contracted translates. We know that the 12 translates of V, H, ∆, Γ by fixed-points of σ map to points in P2. We will use the notation Vi = V + pi0, Hi = H + p0i, ∆i = ∆ + pi,2i, Γi = Γ + pii for these translates (where 0 6 i 6 2) and we will determine the coordinates of their image points. Consider in the linear series L the divisors D0 := V1 + V2 + N12 D1 := V0 + V2 + N02 D2 := V0 + V1 + N01 We choose sections si ∈ H 0(A, M ) defining them and use these to define the map φL : B → P2. Clearly the vertical curves V0, V1, V2 then map to the points v0 = (1 : 0 : 0), v1 = (0 : 1 : 0), v∞ = (0 : 0 : 1) (6) respectively. We will now use the projective representation K(M ) → PGL(H 0(A, M )) (see [5, Chap. 6]) in order to determine the coordinates of the images of the remain- ing nine curves Hi, ∆i, Γi. By Lemma 2.7 we have Fix(ζ 2) ⊂ K(M ), and we know that translation by fixed points leaves the condition of vanishing in these points invariant. Therefore the representation restricts to Fix(ζ 2) → PGL(H 0(B, L)). In other words, each of the nine fixed points gives rise to a projective transformation of P2. Let Mij denote the projective transformation corresponding to pij. As the translation tp10 cycles the vertical translates, V0 → V1 → V2 → V0, we know that M10 must be of the form M10 = 0 λ1 0 0 0 λ2 λ3 0 0! with non-zero entries λi. Note now that scaling the sections si corresponds to a diagonal transformation on P2. We can therefore scale the si (which leaves the coordinates in (6) invariant) in such a way that in fact λ1 = λ2 = λ3 = 1, so that M10 = 0 0 1 0 1 0! 1 0 0 The key is now the fact that the horizontal curves H0, H1, H2 are fixed under p10. Their coordinate vectors must therefore be eigenvectors of M10. These are h0 = (1 : 1 : 1), h1 = (1 : ε : ε2), h∞ = (1 : ε2 : ε) (7) where ε denotes a primitive third root of unity. (After possibly rechoosing the origin in A they are in this order.) Consider now the translation tp01. It fixes V0, V1, V2 and therefore M01 is of the form M01 = µ1 0 0 0 µ2 0 0 µ3! 0 with non-zero entries µi. Since tp01 maps H0 to H1, we have in fact M01 = 1 0 0 0 ε2! 0 ε 0 0 after scaling M01 if necessary. (Here we use that the images of the Hi are given by the coordinates, and the order, in (7).) With this information at hand, we can now also determine the coordinates of the images of the ∆i and Γi. First, the diagonal translates ∆i are fixed under tp11, and therefore their images are given by the eigenvectors of the matrix 18 M11 = M10 · M01 = 0 ε 0 0 ε2 0 0 1 0! Thus we get the points δ0 = (ε : 1 : 1), δ1 = (1 : ε : 1), δ∞ = (1 : 1 : ε) (8) Here the first of these points is the image of ∆0, because it is this point among the three which is collinear with the images of V0 and H0. And finally, the graph translates Γi are fixed under tp12, which leads us to the matrix M12 = M 2 01 · M10 = 0 ε2 0 0 0 0 1 ε 0! and the coordinates γ0 = (ε2 : 1 : 1), γ1 = (1 : ε2 : 1), γ∞ = (1 : 1 : ε2) (9) The first of these points is the image of Γ0 (again by collinearity with V0 and H0). Summing up, we found that in the chosen basis of H 0(L) the image points of the 12 translates are given by (6), (7), (8) and (9), and these coincide with the points in the dual Hesse configuration in standard form. The Roulleau-Urz´ua configuration. Let n = 3m for some integer m ≥ 1. Using the group of n-torsion points on A, we can translate the curves V, H, ∆ and Γ to obtain an a priori count of 4n4 curves, 4 each at each of the n4 n-torsion points. Since each of the divisors V, H, ∆ and Γ contain n2 of the torsion points, and thus are their own images under translation by this subgroup, there are only actually 4n4/n2 = 4n2 curves. The images under ϕL : B → P2 of the proper transforms under B → A of these curves form the Roulleau-Urz´ua configuration. We can now prove that these are exactly the Halphen cubics of order m. Proof of Theorem 2. The n-torsion subgroup contains the order 9 subgroup con- sisting of the 9 points fixed with respect to the action of σ on A. Each of the 12 curves through these 9 points (these are the curves shown in Figure 1) map to points of P2, and we found these points above. The orbits under applications of σ among the remaining 4(n2 − 3) curves consist of 3 curves each. Thus under ϕL these curves map 3 to 1 to cubic curves, and the images of the 4(n2 − 3) curves in A are 4(n2 − 3)/3 = 4(3m2 − 1) cubic curves in P2. The curves in the pencils (Hu), (∆u) and (Γu) meet V0, V1, V2, so their images pass through the points v0, v1, v∞. Similarly, the curves in the pencils (Vu), (∆u) and (Γu) meet H0, H1, H2, so their images pass through the points h0, h1, h∞, and so on. All together, the images of curves in (Vu) pass through all 9 points in Λv, so they are members of Cv, and similarly the pencils (Hu), (∆u) and (Γu) map to the pencils Ch, Cδ and Cγ. 19 Note that n2 − 3 curves on A come from each of the 4 pencils. The ones which meet V come from the pencils (Hu), (∆u) and (Γu), with one from each pencil meeting V at each of the n2 − 3 n-torsion points on V not fixed by σ. Now, V maps to the point v0 ∈ P2, and the triples of curves at each torsion point of V thus map to curves with the same tangent direction at v0 = (1 : 0 : 0). The tangent directions are the (n2 − 3)/3 infinitely near images in X of the n-torsion points on V not fixed by σ. We also get 3 tangent directions corresponding to the infinitely near images of the 3 n-torsion points on V fixed by σ; but these we know to be the tangent directions of the 3 lines of the dual Hesse configuration that pass through v0, which are the images of the 3 exceptional divisors E0i above the fixed points on V . So the restriction of the quotient map B → X to V is a degree 3 morphism V → Ev0 triply ramified above the directions of the dual Hesse lines, and therefore the tangent directions to the Roulleau-Urza cubics in Ch, Cδ and Cγ are exactly the points in Ev0 [3m]. So, they are indeed the Halphen cubics of order m in the pencils Ch, Cδ and Cγ. The same argument applied to the restriction of B → X to H proves that the Roulleau-Urz´ua cubics in the remaining pencil Cv are the Halphen cubics as well. (cid:3) References [1] Artebani, M., Dolgachev, I.: The Hesse pencil of plane cubic curves. L'Enseignement Math´ematique. Revue Internationale. 2e S´erie 55 (2009) 235–273. [2] Aure, A., Decker, W., Popescu, S., Hulek, K., Ranestad, K.: The geometry of bielliptic surfaces in P4. Internat. J. Math., 4 (1993) 873–902. [3] Bauer, Th., Schulz, C.: Seshadri constants on the self-product of an elliptic curve. Journal of Algebra 320 (2008), 2981–3005. [4] Bauer, Th., Szemberg, T.: On tensor products of ample line bundles on abelian varieties. Math. Z. 223, 79-85 (1996) [5] Birkenhake, C., Lange, H.: Complex abelian varieties. Springer, 2004. [6] Cohen, H.: Advanced topics in computational number theory. Graduate Texts in Mathematics, 193. Springer-Verlag, New York, 2000. xvi+578 pp. ISBN: 0-387-98727-4 [7] Dolgachev, I. V.: Classical algebraic geometry. A modern view. Cambridge University Press, 2012. [8] Frium, H.: The group law on elliptic curves on Hesse form, in "Finite fields with applications to coding theory, cryptography and related areas (Oaxaca, 2001), 123–151, Springer, Berlin, 2002. [9] Halphen, G.: Recherches sur les courbes planes du troisieme degr´e. Math. Ann. 15, 359–379 (1879) [10] Hayashida, T., Nishi, M.: Existence of curves of genus two on a product of two elliptic curves. J. Math. Soc. Japan 17, No. 1, 1-16 (1965) [11] Hirzebruch, F.: Chern numbers of algebraic surfaces: an example. Math. Ann. 266, no. 3, 351–356 (1984) [12] Lang, S.: Elliptic curves: Diophantine Analysis, Springer-Verlag, 1978. [13] Roulleau, X.: Bounded negativity, Miyaoka-Sakai inequality and elliptic curve configurations. Preprint 2014, arXiv:1411.6996. [14] Roulleau, X., Urz´ua, G.: Chern slopes of simply connected complex surfaces of general type. Ann. Math. 182, 287–306 (2015) [15] Silverman, J. H., The arithmetic of elliptic curves. Graduate Texts in Mathematics, 106. Springer-Verlag, New York, 1986. xii+400 pp. ISBN: 0-387-96203-4 Thomas Bauer, Fachbereich Mathematik und Informatik, Philipps-Universitat Marburg, Hans-Meerwein-Strasse, D-35032 Marburg, Germany E-mail address: [email protected] 20 Brian Harbourne, Department of Mathematics, University of Nebraska, Lincoln, NE 68588-0130 USA E-mail address: [email protected] Joaquim Ro´e, Departament de Matem`atiques, Universitat Aut`onoma de Barcelona, 08193 Bellaterra (Barcelona), Spain E-mail address: [email protected] Tomasz Szemberg, Department of Mathematics, Pedagogical University of Cra- cow, Podchor¸azych 2, PL-30-084 Krak´ow, Poland Current Address: Polish Academy of Sciences, Institute of Mathematics, ´Sniadeckich 8, PL-00-656 Warszawa, Poland E-mail address: [email protected]
1605.09657	1	1605	2016-05-31T15:08:10	Irreducible Theta Divisors of PPAV's are Strongly F-regular	[ "math.AG" ]	We study the birational geometry of irregular varieties and the singularities of Theta divisors of PPAV's in positive characteristic by applying recent generic vanishing results of Hacon and Patakfalvi. In particular, we prove that irreducible Theta divisors of principally polarized abelian varieties are strongly F-regular, which extends an old result of Ein and Lazarsfeld to fields of positive characteristic. In order to prove this, we formulate a positive characteristic analogue of another result of Ein and Lazarsfeld, to the effect that the Albanese image of a smooth projective variety of maximal Albanese dimension with vanishing holomorphic Euler characteristic is fibered by abelian subvarieties.	math.AG	math	IRREDUCIBLE THETA DIVISORS OF PRINCIPALLY POLARIZED ABELIAN VARIETIES ARE STRONGLY F-REGULAR ALAN MARC WATSON Abstract. We study the birational geometry of irregular varieties and the singularities of Theta divisors of PPAV's in positive characteristic by applying recent generic vanishing results of Hacon and Patakfalvi. In particular, we prove that irreducible Theta divisors of principally polarized abelian varieties are strongly F-regular, which extends an old result of Ein and Lazarsfeld to fields of positive characteristic. In order to prove this, we formulate a positive characteristic analogue of another result of Ein and Lazarsfeld, to the effect that the Albanese image of a smooth projective variety of maximal Albanese dimension with vanishing holomorphic Euler characteristic is fibered by abelian subvarieties. Contents Introduction 1. Acknowledgements 2. Preliminaries 2.1. Derived categories and Fourier-Mukai transforms 2.2. F-singularities and linear subvarieties of abelian subvarieties 2.3. The Frobenius morphism on Abelian varieties 2.4. Generalities on inverse systems and spectral sequences 2.5. Generic vanishing in positive characteristic 3. Main technical result 4. Fibering of the Albanese image 5. Singularities of Theta divisors 5.1. Case of simple abelian varieties 5.2. General case References 1 4 4 4 6 8 8 11 14 20 24 24 26 27 1. Introduction The purpose of this paper is to apply recent generic vanishing results in positive characteristic due to Hacon and Patakfalvi [HP13] to the study of the birational geometry of irregular varieties and the singularities of Theta divisors of principally polarized abelian varieties. Over fields of characteristic zero, seminal work of Ein and Lazarsfeld [EL97] applied generic vanishing techniques over the complex num- bers to settle a number of questions concerning the geometry of irregular varieties. Date: March 9, 2016. 1 2 ALAN MARC WATSON One of their main results states that irreducible Theta divisors on principally po- larized abelian varieties have mild singularities: Theorem 1.1. (c.f. [EL97, Theorem 1]) Let A be an abelian variety and let Θ ⊂ A be a principal polarization (i.e. an ample divisor such that h0(A, ØA(Θ)) = 1). If Θ is irreducible, then it is normal and has rational singularities. The conclusion of the theorem is captured by the adjoint ideal of Θ: given any log resolution µ : A′→A of the pair (A, Θ) and writing µ∗Θ = Θ′ + F with Θ′ smooth and F µ-exceptional, one may define adj(A, Θ) = µ∗ØA′(KA′/A − F ). Standard arguments show that adj(A, Θ) = ØA is equivalent to Θ being normal and having rational singularities (see section 9.3.E in [Laz04]). Bearing this in mind, Ein and Lazarsfeld's argument breaks into the following steps: let X→Θ be a resolution of singularities. (i) If Θ is irreducible, then X is of general type. This relies on a classical argument due to Ueno (see [Mor00], Theorem 3.7), characterizing the Itaka fibration and the Kodaira dimension of an irreducible subvariety of an abelian variety. (ii) If X is of general type, then χ(ωX ) > 0. More concretely, if X is a smooth projective variety of maximal Albanese dimension and χ(X, ωX ) = 0, then the image of the Albanese map is fibred by tori. In particular, this shows that if X is birational onto its image under the Albanese map, then X is not of general type. (iii) Generic vanishing theorems and Nadel vanishing yield adj(Θ) = ØA ⇐⇒ χ(X, ωX) > 0. Therefore if Θ is irreducible, its adjoint ideal must be trivial, and by the charac- terization described above, it must be normal an have rational singularities. Work of Abramovich [Abr95] shows that the statement in (i) remains valid in positive characteristic once an appropriate notion of Kodaira dimension is defined for pos- sibly singular varieties. In this paper we provide positive characteristic analogues of items (ii) and (iii). The only known results in this direction are due to Hacon [Hac11], where he proved that for principally polarized abelian varieties (A, Θ) over algebraically closed fields of positive characteristic, the pair (A, Θ) is a limit of strongly F-regular pairs. More precisely: Theorem 1.2. (c.f. [Hac11, Theorem 1.1]) Let (A, Θ) be a principally polarized abelian variety over an algebraically closed field of characteristic p > 0. If D ∈ mΘ, then(cid:0)A, 1−ǫ m D(cid:1) is strongly F-regular for any rational number 0 < ǫ < 1 We summarize briefly our main results. The arguments employed in the proofs bear a strong resemblance to their characteristic zero analogues, albeit plenty of technicalities arise. Not only is resolution of singularities unavailable in general, but also generic vanishing for canonical sheaves is known to fail in positive characteristic (c.f. [HK12]). Nevertheless, recent work of Hacon and Patakfalvi [HP13] provides strong generic vanishing statements for objects arising from Cartier modules (see section 2.4 for precise statements): given a coherent Cartier module Ω0 ∈ Coh(A), the traces of the Frobenius iterates yield an inverse system · · · →F e ∗ Ω0→F e−1 ∗ Ω0→ · · · IRREDUCIBLE THETA DIVISORS OF PPAV'S ARE STRONGLY F-REGULAR 3 ∗ Ω0 its inverse limit, there exists a closed subset Z ⊂ A F e and denoting by Ω = lim ←− α ) = 0 for every i > 0 and every α ∈ Z such that peα /∈ Z for such that H i(A, Ω ⊗ P ∨ all e >> 0 (c.f. Corollary 3.3.1 in [HP13]). This grounds on the following Theorem, which is the main result in [HP13]. Theorem 1.3. (c.f. [HP13, Theorem 3.1.1 and Lemma 3.1.2]) Let k be an alge- braically closed field of characteristic p > 0 and A be an abelian variety over k. Let {Ωe} be Cartier module on A. If for any sufficiently ample line bundle L on A and any e ≫ 0, H i(A, Ωe ⊗ L∨) = 0 for all i > 0, then the complex1 is a quasi-coherent sheaf in degree 0, i.e., Λ = H0(Λ). Λ = hocolim→RSA, A(DA(Ωe)) This result is generalized further in [WZ14], where a notion of M-regularity in positive characteristic is also introduced. Concretely, one has the following: Theorem 1.4 (c.f. Theorem 4.2 in [WZ14]). Let A be an abelian variety and {Ωe} be a GV-inverse system of coherent sheaves on A such that {Ωe} is M-regular, in the sense that H0(Λ) is torsion-free. Then for any scheme-theoretic point P ∈ A, if dim P > i, then P is not in the support of for any e. Im(Ri S(Ω)→Ri S(Ωe)) Our main technical result is a partial converse to the previous theorem: the presence of torsion in H0( L) induces the following non-vanishing statement: Theorem 1.5 (c.f. Theorem 3.1). Let {Ωe} be an inverse system of coherent sheaves on a g-dimensional abelian variety satisfying the Mittag-Leffler condition and let Ω = lim Ωe. Let Λe = RSA, A(DA(Ωe)) and Λ = hocolim→Λe. Suppose that ←− {Ωe} is a GV-inverse system, in the sense that Hi(Λ) = 0 for any i 6= 0. If H0( L) is has a torsion point P of dimension g − k, then the maps lim ←−(cid:16)RkSA, A(Ωe) ⊗ k(P )(cid:17) →RkSA, A(Ωe) ⊗ k(P ) are non-zero for every e >> 0. Using this, we can derive a fibration statement similar to that of Ein and Lazars- feld: Theorem 1.6 (c.f. Theorem 4.2). Let X be a smooth projective variety of maximal Albanese dimension and denote by a : X→A the Albanese map. Let g = dim A. Consider the inverse system {Ωe := F e Ωe. Define Le = RSA, ADA(Ωe) and assume that the sheaf H0( L) = lim H0( Le) has torsion. −→ Then the image of the Albanese map is fibered by abelian subvarieties of A. ∗ S0a∗ωX }e and denote Ω = lim ←− An identical argument to the one we employ to prove the previous theorem also yields the following result describing the singularities of Theta divisors in positive characteristic. 1Here SA, A denotes the Fourier-Mukai functor with kernel given by the Poincar´e bundle of A × A 4 ALAN MARC WATSON Theorem 1.7. Let A be an ordinary abelian variety over an algebraically closed field of positive characteristic and let Θ be an irreducible Theta divisor. Then Θ is strongly F-regular. Over fields of positive characteristic, work of Smith and Hara (c.f. [Smi97], [Har98]) shows that F-rationality is the positive characteristic counterpart to ratio- nal singularities, and the former is implied by strong F-regularity, so in this sense Theorem 1.7 is stronger than one might expect. This paper is structured as follows. We start by recording all the background results we need in section 2: in 2.1 we recall the main definitions and some useful properties of the Fourier-Mukai transform in the context of abelian varieties, in 2.2 we record the relevant definitions of F-singularities; in 2.3 we record results of Pink and Roessler characterizing subvarieties of abelian varieties in positive characteris- tic; in 2.4 we outline a few useful facts concerning inverse systems that will ease the exposition of the proofs and in 2.5 we collect the generic vanishing statements in positive characteristic that will be needed in the sequel. Sections 3 and 4 constitute the technical core of the paper: section 3 contains the proof of the non-vanishing statement in the presence of torsion of H0( L) and in section 4 we generalize Ein and Lazarsfeld's fibration statement. Finally in section 5 we present the proof of Theorem 1.7 on the singularities of Theta divisors. We start with the case of simple abelian varieties in section 5.1, which is a simple computation following easily from the results in [HP13] that does not require the arguments from sections 3 and 4. The general case is more involved and is presented section 5.2. Acknowledgements. The author would like to thank his advisor Christopher Hacon for suggesting this problem, for invaluable discussions and for sharing an early draft of [HP13]. 2. Preliminaries 2.1. Derived categories and Fourier-Mukai transforms. Let A be a g-dimensional abelian variety, denote by A = Pic0(A) its dual and let F ∈ Coh(A). Let P ∈ Pic(A × A) be the Poincar´e bundle and denote and consider the usual Fourier- Mukai functors: RSP A, A : D(A)→D( A), RSP A, A(•) = Rp A∗(p∗ A(•) ⊗ P) even though we will most often omit P from the notation and simply write RSA, A(•). We start by stating Mukai's inversion theorem in the derived category of quasi- coherent sheaves: Theorem 2.1 ( [Muk81]). If [−g] denotes the rightwise shift by g places and −1A is the inverse on A, the following equalities hold on Dqc(A) and Dqc( A) RS A,A ◦ RSA, A = (−1A)∗[−g], RSA, A ◦ RS A,A = (−1 A)∗[−g] We will also be using the following two results: IRREDUCIBLE THETA DIVISORS OF PPAV'S ARE STRONGLY F-REGULAR 5 Lemma 2.2 ( [Muk81], Proposition 3.8). The Fourier-Mukai transform commutes with the dualizing functor in Dqc( A) up to inversions and shifts, namely DA ◦ RS A,A ≃(cid:16)(−1A)∗ ◦ RS A,A ◦ D A(cid:17) [g] Lemma 2.3. (c.f, [Muk81, Lemma 3.4]) Let φ : A→B be an isogeny between abelian varieties and denote by φ : B→ A the dual isogeny. Then the following equalities hold on Dqc(B) and Dqc(A) respectively. φ∗ ◦ RS B,B ≃ RS A,A ◦ φ∗, φ∗ ◦ RS A,A ≃ RS B,B ◦ φ∗ In particular, this holds for the (e-th iterate) Frobenius map F e and its dual isogeny, namely the Verschiebung map V e = F e. We will also be using the following simple remark. Lemma 2.4. (c.f. [Huy06, Exercise 5.12]) Let π : B→A be a morphism between abelian varieties and let P be a locally free sheaf on A × A. Denote Pπ = (π × 1 A)∗(P). Then SPπ ≃ SP ◦ π∗ We next record the notions of homotopy limits and colimits in the derived cate- gory. Given a direct system of objects Ci ∈ D(A) C1→C2→ . . . its homotopy colimit hocolim→Ci is defined by the triangle ⊕Ci−→ ⊕ Ci−→hocolim→Ci [+1] −→ where the first map is the homomorphism given by id−shif t where shif t : ⊕Ci→⊕ Ci is given on Ci by the composition Ci→Ci+1 ֒→ ⊕Cj. Given an inverse system of objects Ci ∈ Dqc(X) C1 ←− C2 ←− · · · its homotopy limit holim←Ci is given by the triangle holim←Ci−→Y Ci−→Y Ci +1−→ where the map between products isQ(id − shif t) and where by product we mean product of chain complexes as opposed to the product inside Dqc(X). If X is an n-dimensional variety over a field k and ω• Note that if Ci are coherent sheaves, then hocolim→Ci = lim −→ X denotes its dualizing X ) ≃ ωX , we define the dualizing functor DX on X). In this context, Grothendieck duality reads complex, so that H−dim X (ω• Dqc(X) as DX (F ) = RHom(F , ω• as follows: Ci. 6 ALAN MARC WATSON Theorem 2.5. Let f : X→Y be a proper morphism of quasi-projective varieties over a field k. Then for any complex F ∈ Dqc(X) we have an isomorphism Rf∗DX (F ) ≃ DY Rf∗(F ) Assuming that X and Y are smooth, then we equivalently have that for any F ∈ Dqc(X) and E ∈ Dqc(Y ), if ωf = ωX ⊗ f ∗ωY denotes the relative dualizing sheaf, there is a functorial isomorphism Rf∗RHom(F , Lf ∗(E) ⊗ ωf [dim X − dim Y ]) ≃ RHom(Rf∗F , E) 2.2. F-singularities and linear subvarieties of abelian subvarieties. In this section we recall the basic notions from the theory of F-singularities following [Sch12] and [BST12]. Let X be a separated scheme of finite type over an F-finite perfect field of characteristic p > 0. A variety is a connected reduced equidimen- sional scheme over k. We denote the canonical sheaf of X by ωX = H−dim X (ω• X ), where ω• X = η∗k is the dualizing complex of X and η : X→k is the structural map. If X is normal, a canonical divisor on X is any divisor KX such that ωX ≃ ØX (KX ). By a pair (X, ∆) we mean the combined information of a normal integral scheme X and an effective Q-divisor ∆. Denote by F e : X→X the e-th iterated absolute Frobenius, where the source has structure map η ◦ F e : X→k. Since (F e)!ω• X = (F e)!η!k = η!(F e)!k = η!k = ω• X . In general for a finite morphism f : X→Y , a coherent sheaf F on X and a quasi-coherent sheaf G on Y , we have the duality X ≃ (F e)!ω• Hom(f∗F , G) ≃ f∗Hom(F , f !G), so the identity ω• X yields a trace map F e ∗ ωX →ωX . Given a variety X, the parameter test submodule τ (ωX ) of X is the unique smallest ØX -submodule M ⊆ ωX , non-zero on any component of X, such that Φ1(F∗M ) ⊆ M . X →ω• X and taking cohomology we obtain Φe : F e ∗ ω• X →ω• Assume that (X, ∆) is a pair such that KX + ∆ is Q-Cartier with index not divisible by p. Choose e > 0 such that (pe − 1)(KX + ∆) is Cartier and define the line bundle Le,∆ = ØX ((1 − pe)(KX + ∆)). By [Sch09], there is a canonically determined map φe,∆ : F e ∗ Le,∆→ØX . We define the test ideal τ (X, ∆) of the pair (X, ∆) to be the smallest non-zero ideal J ⊆ ØX such that φe,∆(F e ∗ (J · Le,∆)) ⊆ J. Similarly one defines the non-F-pure ideal σ(X, ∆) of (X, ∆) to be the the largest such ideal J ⊆ ØX . Ever since Hochster and Huneke introduced test ideals and tight closure theory in [HH90], deep connections have been established between the classes of singular- ities defined in terms of Frobenius splittings and those arising within the minimal model program. For instance, a normal domain (R, m) of characteristic p > 0 is ∗ R ≡ R1/pe said to be F-pure if the inclusion induced by the Frobenius R ֒→ F e splits for every e. Similarly, a pair (R, ∆) is said to be F-pure if the inclusion R ֒→ R1/pe splits for every e and it was shown in [HW02] that F-pure pairs are the analogues of log canonical pairs in characteristic zero, in the sense that if (X, ∆) is a log canonical pair, then its reduction mod p (Xp, ∆p) is F-pure for all p >> 0. ֒→ R (⌈(pe − 1)∆⌉)1/pe IRREDUCIBLE THETA DIVISORS OF PPAV'S ARE STRONGLY F-REGULAR 7 In this paper we shall be concerned with the two classes of F-singularities that we define next. We will be recording the original definition in terms of Frobenius splittings and we will then state their description in terms of test ideals that will be used in the sequel. Definition 2.6. (i) A pair (X = Spec R, ∆) is strongly F-regular if for every ֒→ splits as an R-module homo- non-zero element c ∈ R, there exists e such that the map R ֒→ R1/pe R((pe − 1)∆)1/pe morphism. that sends 1 7→ c1/pe 7→ c1/pe (ii) A reduced connected variety X is F-rational if it is Cohen-Macaulay and there is no non-zero submodule M ( ωR such that the Grothendieck trace map ΦX : F e ∗ ωX →ωX satisfies Φ(F e ∗ M ) ⊆ M . Strongly F-regular pairs are the analog of log terminal pairs in characteristic zero (c.f. [HW02]) and F-rational varieties are the analogue of varieties with rational singularities (c.f. [Smi97]). The notion of strong F-regularity is also captured by the test ideal, as the following well-known result shows. Lemma 2.7. (c.f. [HW02, Proposition 2.4]) A pair (X, ∆) is strongly F-regular if, and only if, τ (X, ∆) = ØX . Assume that (X, ∆) is a pair, where X is a normal proper variety over an al- gebraically closed field of characteristic p > 0 and ∆ ≥ 0 is a Q-divisor such that KX + ∆ is Q-Cartier with index not divisible by p. The map φe ∗ Le,∆→ØX defined in [Sch09] restricts to surjective maps ∆ : F e F e ∗ (σ(X, ∆) ⊗ Le,∆)−→σ(X, D), F e ∗ (τ (X, ∆) ⊗ Le,∆)−→τ (X, D). The power of vanishing theorems in characteristic zero relies on the fact that they allow us to lift global sections of adjoint bundles. The full space of global sections is not so well behaved in positive characteristic, so one instead focuses on a subspace of it that is stable under the Frobenius action. If M is any Cartier divisor, one thus defines the subspace S0(X, τ (X, ∆) ⊗ ØX(M )) as S0(X, τ (X, ∆) ⊗ ØX (M )) := \n≥0 Im(cid:0)H 0(X, F ne ∗ τ (X, ∆) ⊗ Lne,∆(pneM ))−→H 0(X, τ (X, ∆) ⊗ ØX (M ))(cid:1) ⊆ H 0(X, ØX (M )) Among the many applications of these subspaces, for instance, they can be used to prove global generation statements: concretely, suppose that X is a d- dimensional variety of characteristic p > 0 and that ∆ is a Q-divisor such that KX + ∆ is Q-Cartier with index not divisible by p. It was shown in [Sch09] that if L and M are Cartier divisors such that L − KX − ∆ is ample and M is ample and globally generated, then the sheaf τ (X, ∆) ⊗ ØX (L + nM ) is globally generated for all n ≥ d by S0(X, τ (X, ∆) ⊗ ØX (L + nM )). 8 ALAN MARC WATSON 2.3. The Frobenius morphism on Abelian varieties. Throughout this paper, A will denote an abelian variety of dimension g over a field k and A = Pic0(A) will denote the dual abelian variety Lemma 2.8. (c.f. [HP13, Proposition 2.13]) For a g-dimensional abelian variety A over a field k, the following conditions are equivalent. (i) There are pg p-torsion points. (ii) The Frobenius action H g(A, ØA)→H g(A, ØA) is bijective (iii) The Frobenius action H i(A, ØA)→H i(A, ØA) is bijective for all 0 ≤ i ≤ g (iv) S0(A, ωA) = H 0(A, ωA) If any of these equivalent conditions is satisfied we say that A is ordinary. Given an isogeny ϕ : A→B between abelian varieties of dimension g, A is ordinary if and only if B is ordinary. Given a surjective morphism ϕ : A→B of abelian varieties, if A is ordinary then so is B (see Lemmas 2.14 and 2.14 in [HP13]). We finally record a characterization of linear subvarieties of abelian varieties following [PR03]. Let A be an abelian variety endowed with an isogeny ϕ : A→A. We say that A is pure of positive weight if there exist integers r, s > 0 such that ϕs = Fpr for some model of A over Fpr . If A is defined over a finite field, we say A is supersingular if and only if it is pure of positive weight for the isogeny given by multiplication by p; in general, we say that A is supersingular if it is isogenous to a supersingular variety defined over a finite field. We say that A has no supersingular factors is there exist no non-trivial homomorphism to an abelian variety which is pure of positive weight for the isogeny given by multiplication by p. One sees that A has no supersingular factors if there does not exist a non-trivial homomorphism to a supersingular abelian variety. In particular, if A is an ordinary abelian variety, it follows from the observations in the previous paragraph that A has no supersingular factors (see Lemma 2.16 in [HP13]). The following result of Pink and Roessler characterizing linear subvarieties of abelian varieties will be crucial in our proof: Theorem 2.9. (c.f. [PR03, Theorem 2.2]) Let A be an abelian variety over a field K of characteristic p > 0 and let X ⊂ A be a reduced closed subscheme p(X) ⊂ X, where p denotes the isogeny given by multiplication by p. If A has no supersingular factors, then all the maximal dimensional irreducible components of X are completely linear (namely, torsion translates of subabelian varieties). 2.4. Generalities on inverse systems and spectral sequences. Mittag-Leffler inverse systems. We start by recording a few results that will be used in the se- quel, most of which are taken directly from [Har78]. Recall that a sheaf is countably quasi-coherent if it is quasi-coherent and locally countably generated. Also recall that an inverse system of coherent sheaves {Ωe} is said to satisfy the Mittag-Leffler condition if for any e ≥ 0, the image of Ωe′ →Ωe stabilizes for e′ sufficiently large. The inverse limit functor is always left exact in the sense that if 0 0 / Fe Ge He / Fe−1 / Ge−1 / He−1 0 / 0 / / / / / / / / / / / IRREDUCIBLE THETA DIVISORS OF PPAV'S ARE STRONGLY F-REGULAR 9 is an exact sequence of inverse systems, then 0→ lim ←− Fe→ lim ←− Ge→ lim ←− He is exact in the category of quasi-coherent sheaves. By a theorem of Roos (c.f. Proposition I.4.1 in [Har78]), the right derived functors Ri lim are 0 for i > 2. ←− Hence, we have a long exact sequence 0→ lim ←− Fe→ lim ←− Ge→ lim ←− He→R1 lim ←− Fe→R1 lim ←− Ge→R1 lim ←− He→0. We start by recording a characterization of the Mittag-Leffler condition in terms of the first right-derived inverse limit. Lemma 2.10. (c.f. [Har78, Proposition I.4.9]) Let {Ωe}e be an inverse system of countably quasi-coherent sheaves on a scheme X of finite type. Then the following conditions are equivalent: (i) {Ωe}e is satisfies the Mittag-Leffler condition. (ii) R1 lim ←− (iii) R1 lim ←− Ωe = 0 Ωe is countably quasi-coherent. The following is basic result about the cohomology of an inverse system of sheaves: Proposition 2.11. (c.f. [Har78, Theorem I.4.5]) Let {Ωe} be an inverse system of coherent sheaves on a variety X. Let T be a functor on D(X) which commutes with arbitrary direct products. Suppose that {Ωe} satisfies the Mittag-Leffler condition. Then for each i, there is an exact sequence 0→R1 lim ←− Ri−1T (Ωe)→RiT (lim ←− Ωe)→ lim ←− RiT (Ωe)→0. In particular, if for some i, {Ri−1T (Ωe)} satisfies the Mittag-Leffler condition, then RiT (lim ←− RiT (Ωe) (by Lemma 2.10). Ωe) ∼= lim ←− We will be applying this theorem to the push-forward f∗ under a proper mor- phism of schemes. We finally record a standard statement about the commutation of inverse limits and tensor products. Recall that a sheaf is countably quasi-coherent if it is quasi-coherent and locally countably generated. Then one has the following: Lemma 2.12. (see [Har78, Proposition 4.10]) Let {Fe}e be an inverse system of countably quasi-coherent sheaves on a scheme X of finite type and let E be a flat ØX-module. Consider the natural map α : (lim ←− Fe) ⊗ E → lim ←− (Fe ⊗ E) Fe is countably quasi-coherent then α is injective and if furthermore R1 lim ←− Ωe If lim ←− is countably quasi-coherent, then α is surjective. In particular, if {Fe} is an inverse system of coherent sheaves satisfying the Mittag-Leffler condition on a scheme X with generic point w, then there is an isomorphism lim ←− e Fe! ⊗ k(w)−→ lim ←− e Fe ⊗ k(w)! 10 ALAN MARC WATSON Inverse systems of convergent spectral sequences. The following observa- tion is taken from [Car08]. Let {E(n)} be an inverse system of spectral sequences with morphisms of spectral sequences E(n)→E(n − 1) and consider the tri-graded abelian groups Er p,q(n), with differentials given by the inverse limits of the differentials in the E(n). Concretely, if dr(n) is the r-th differential in E(n) and x(n) ∈ Er p,q is given by dr(x(n)) = dr(n)(x(n)). The resulting object is a spectral sequence provided that H(Er p,q, then the r-th differential dr in the limit sequence Er p,q , which is in turn equivalent to showing that p,q, dr) = Er+1 p,q = lim ←−n Er H(lim ←− n Er p,q(n), dr) = lim ←− n H(Er p,q(n), dr) and this is precisely the statement of Proposition 2.11 above (for the functor of global sections). Note that, in particular, if the terms Er p,q are all finite-dimensional vector spaces, (1) terms are 0. Besides, the hypotheses of Proposition 2.11 hold and all the lim ←− given that the inverse limit of the spectral sequences is again a spectral sequence and provided that every spectral sequence in the inverse system is bounded and convergent, one observes that the limit spectral sequence is also convergent: for fixed p, q, there is a fixed N such that E∞ p,q(n) = EN p,q(n). Morphisms between spectral sequences. We recall the definition of a spec- tral sequence from EGA III [0III, 1.1]. Let C be an abelian category. A (biregular) spectral sequence E on C consists of the following ingredients: (1) A family of objects {Ep,q r } in C , where p, q, r ∈ Z and r > 2, such that for any fixed pair (p, q), Ep,q stabilizes when r is sufficiently large. We denote r the stable objects by Ep,q ∞ . (2) A family of morphisms dp,q r r →Ep+r,q−r+1 satisfying : Ep,q dp+r,q−r+1 r r ◦ dp,q r = 0. r )/Im(dp−r,q+r−1 r (3) A family of isomorphisms αp,q (4) A family of objects {En} in C . For every En, there is a bounded decreasing filtration {F pEn} in the sense that there is some p such that F pEn = En and there is some p such that F pEn = 0. : ker(dp,q r ) →∼ Ep,q r+1. (5) A family of isomorphisms βp,q : Ep,q ∞ →∼ F pEp+q/F p+1Ep+q. We say that the spectral sequence {Ep,q r } converges to {En} and write Ep,q 2 ⇒ Ep+q. A morphism φ : E→H between two spectral sequences on C is a family of morphisms φp,q and φn : En→H n such that φ is compatible with d, α, the filtration and β. The following result is useful in order to obtain information about the limiting map φn from the maps φp,q 2 . r →H p,q : Ep,q r r Lemma 2.13 (see Lemma 2.15 in [WZ14]). Let Ei,j 2 Ei+j φi,j 2 φi+j H i,j 2 3 H i+j + 3 + IRREDUCIBLE THETA DIVISORS OF PPAV'S ARE STRONGLY F-REGULAR 11 be two spectral sequences with commutative maps. Let l and a be integers. Suppose that Ei,l−i = 0 for i > a and φa,l−a = 0 for i < a, H i,l−i = 0. Then φl = 0. 2 2 2 2.5. Generic vanishing in positive characteristic. A smooth projective variety X over an algebraically closed field is said to have maximal Albanese dimension if it admits a generically finite morphism to an abelian variety X→A. Over fields of characteristic zero, the main tool that is employed when studying properties of varieties of maximal Albanese dimension is the generic vanishing theorem of Green and Lazarsfeld ( [GL87], [GL91]). Even though it is shown in [HK12] that the obvious generalization of this result to fields of positive characteristic if false, recent work of Hacon and Patakfalvi [HP13] provides a weaker generic vanishing statement in positive characteristic which albeit necessarily weaker, is strong enough prove positive characteristic versions of Kawamata's celebrated characterization of abelian varieties. In this subsection we collect the results of [HP13] that we shall be using throughout. The following is the main theorem in [HP13]: Theorem 2.14. ( c.f. [HP13, Theorem 3.1, Lemma 3.2]) Let A be an abelian variety defined over an algebraically closed field of positive characteristic and let Ωe+1→Ωe be an inverse system of coherent sheaves on A. (i) If for any sufficiently ample line bundle L ∈ Pic( A) and for any e >> 0 we have H i(A, Ωe ⊗ RS A,A(L)∨) = 0 for every i > 0, then the com- plex Λ = hocolim→RSA, A(DAΩe) (which in general is concentrated in de- grees [−g, . . . , 0]), is actually a quasi-coherent sheaf concentrated in degree 0, namely Λ = H0(Λ). Besides Ω = lim ←− Ωe =(cid:16)(−1A)∗DARS A,A(Λ)(cid:17) [g]. (ii) The condition in (i) is satisfied for coherent Cartier modules: if F∗Ω0→Ω0 is ∗ Ω0, then for any ample line a coherent Cartier module and we denote Ωe = F e bundle L ∈ Pic( A) we have H i(A, Ωe ⊗ RS A,A(L)∨ ⊗ Pα) = 0, ∀e >> 0, ∀i > 0, ∀α ∈ A From the above result and the cohomology and base change theorem one derives the following corollary: Corollary 2.15. (c.f. [HP13, Corollaries 3.5 and 3.6]) With the same notations as above we have the following: (i) For every α ∈ A we have Λ ⊗ k(α) ≃ lim −→ integer e ≥ 0, H0(Λe) ⊗ k(α) ≃ H 0(A, Ωe ⊗ P ∨ α )∨. H 0(A, Ωe ⊗ P ∨ α )∨, and for every where pZ ′ ⊂ Z ′ If besides A has no supersingular factors, then the top dimen- sional components of Z ′ are a finite union of torsion translates of subtori of A. (ii) There exists a proper closed subset Z ⊂ A such that if i > 0 and pey /∈ Z α )∨ = 0. Furthermore, if W i = {α ∈ H i(A, Ωe ⊗ P ∨ for all e >> 0, then lim −→ A, H i(A, Ωe ⊗ P ∨ α )∨ 6= 0}, then lim ←− W i ⊂ Z ′ = [e≥0(cid:16)[pe A ]−1(Z)(cid:17)red 12 ALAN MARC WATSON We quote two more results from [HP13] that will provide a simple proof of a special case of our main theorem: Proposition 2.16. (c.f. [HP13, Proposition 3.17]) Let A be an ordinary abelian variety and consider the same notations as above. Then each maximal dimensional irreducible component of the set Z of points α ∈ A such that the image of the natural map H0(Λ0) ⊗ Ø A,α−→H0(Λ) ⊗ Ø A,α ≃ Λ ⊗ Ø A,α is non-zero, is a torsion translate of an abelian subvariety of A and Λ ⊗ Ø A,α 6= 0 if and only if P e α ∈ Z. Proposition 2.17. (c.f. [HP13, Lemma 3.9, Corollary 3.10]) Let Ω0 be a coher- ent sheaf on an abelian variety A and assume that F∗Ω0→Ω0 is surjective. Then Supp Ω = Supp Ω0, so that Supp Ω is a closed subvariety. Let B ⊂ A be an abelian subvariety such that V 0(Ω0) = {α ∈ A : h0(Ω0 ⊗ Pα) 6= 0} xΩ ≃ Ω for every x ∈ [A/ B. is contained in finitely many translates of B. Then t∗ In particular, Supp Ω is fibered by the projection A→B, namely Supp Ω is a union of fibers of A→B. Note that, in particular, Proposition 2.17 applies to the subvariety Z =nα ∈ A : Im(cid:16)H0(Λ0) ⊗ Ø A,α−→Λ ⊗ Ø A,α(cid:17) 6= 0o from Proposition 2.16. Finally, grounding on the results in [HP13], part of Pareschi and Popa's generic vanishing theory was extended to positive characteristic in [WZ14]. The main result in that paper is the following: Theorem 2.18. Let A be an abelian variety. Let {Ωe} be an inverse system of coherent sheaves on A satisfying the Mittag-Leffler condition and let Ω = lim Ωe. ←− Let Λe = R S(DA(Ωe)) and Λ = hocolim→Λe. The following are equivalent: (1) For any ample line bundle L on A, H i(A, Ω ⊗ L∨) = 0 for any i > 0. (1') For any fixed positive integer e and any i > 0, the homomorphism H i(A, Ω ⊗ L∨)→H i(A, Ωe ⊗ L∨) is 0 for any sufficiently ample line bundle L. (2) Hi(Λ) = 0 for any i 6= 0. These conditions imply the following: (3) For any scheme-theoretical point P ∈ A, if dim P > g − i, then P is not in the support of the image of for any e. Ri S(Ωe)→Ri S(Ωe) lim ←− If {Ri S(Ωe)} satisfies the Mittag-Leffler condition for any i ≥ 0, then (3) also implies (1), (1') and (2) and, moreover, the support of the image of the map in (3) is a closed subset. IRREDUCIBLE THETA DIVISORS OF PPAV'S ARE STRONGLY F-REGULAR 13 We also record the following variant of the implication (2) ⇒ (3) in the previous theorem. Proposition 2.19. Let π : A։W be a projection between abelian varieties with generic fiber dimension f and with dim W = k. Let {Ωe}e be a Cartier module on A P W × W . Denote If P ∈ W is a scheme-theoretic point with dim (P ) > and let SA, W be the Fourier-Mukai functor with kernel(cid:0)π × 1 W(cid:1)∗ Le = RSA, W (DA(Ωe)). k + f − ℓ, then P is not in the support of the image of the map RℓSA, W (Ωe)−→RℓSA, W (Ωe) lim ←− e Moreover, if the inverse system {RℓSA, W (Ωe)}e satisfies the Mittag-Leffler condi- tion, then the support of the image of the above map is closed and its codimension is ≥ ℓ − f . Proof. The proof is identical (modulo shifts) to that of Theorem 4.2 in [WZ14], but we include it for the sake of completeness. We need to show that if P ∈ W is a scheme-theoretic point with dim (P ) > k + f − ℓ, then (cid:16)RℓSA, W (lim ←− Ωe)(cid:17)P 0−→(cid:16)RℓSA, W (Ωe)(cid:17)P Note in the first place that for any ℓ, we have the following isomorphisms Extp( Le, O W ) ≃ Hp−k(D W ( Le)) ≃ Hp−k(D W (RSA, W (DA(Ωe)))) [∗] ≃ Hp(R SA, W (DA(DA(Ωe)))) ≃ Hp(R SA, W (Ωe)) (1) and Extp( L, O W ) ≃ Hp−g(D W ( L)) ≃ Hp−g(D W (hocolim→eRSA, W (DA(Ωe)))) (2) ≃ Hp−g(holim←eD W (RSA, W (DA(Ωe)))) ≃ Hp(holim←e(−1 W )∗RSA, W ( Ωe)) ≃ (−1 W )∗Hp(cid:16)holim←eRSA, W ( Ωe)(cid:17) where in [∗] we used Lemma 2.2 in [PP11]. Here, if S is the Fourier-Mukai functor with kernel P, S denotes the Fourier-Mukai functor with kernel P ∨; the codimension computation is unaffected by this, so we omit the tildes in the remainder of the proof. From the following factorization of the map RℓSA, W (lim ←− Ωe)→RℓSA, W (Ωe) RℓSA, W (lim ←− e Ωe)−→Hℓ(cid:16)holim←RSA, W (Ωe)(cid:17) −→Hℓ(cid:16)RSA, W (Ωe)(cid:17) we see that it suffices to show that the map Extℓ( L, Ø W )P −→Extℓ( Le, Ø W )P 14 ALAN MARC WATSON is zero for dim (P ) > k + f − ℓ. In order to see this, we may proceed as in the proof of Theorem 4.2 in [WZ14], computing the above map via the commutative diagram Exti(Hj ( L), Ø W )P 3 Extℓ( L, Ø W )P φi,j φℓ Exti(Hj( Le), Ø W )P 3 Extℓ( Le, Ø W )P with i − j = ℓ. We seek to apply Lemma 2.13 with a = ℓ − 1 − f . If i > a, then i ≥ ℓ − f > [k + f − dim (P )] − f = k − dim (P ) and hence Exti(Hj( Le), Ø W )P = 0 and if i ≤ a, then j = i − ℓ ≤ (ℓ − 1 − f ) − ℓ = −1 − f , so that Hj( L) = 0 (c.f. proof of Theorem 3.1.1 in [HP13]). Lemma 2.13 then implies that φℓ = 0 as claimed. (cid:3) 3. Main technical result Let {Ωe} be an inverse system of coherent sheaves on a g-dimensional abelian variety satisfying the Mittag-Leffler condition and let Ω = lim Ωe. Let Λe = ←− RSA, A(DA(Ωe)) and Λ = hocolim→Λe. Suppose that {Ωe} is a GV-inverse sys- tem, in the sense that Hi(Λ) = 0 for any i 6= 0. It was shown in Theorem 4.2 of [WZ14] that if H0( L) is torsion-free, then the maps (cid:16)lim ←− RkSA, A(Ωe)(cid:17)P −→RkSA, A(Ωe)P are zero for any point P ∈ A such that dim (P ) ≥ g − k. We next show a partial converse to this statement. In the sequel, we will say that H0( L) has torsion if it is not torsion-free. More concretely, we will say that H0( L) has torsion at a point P if there exists a section s ∈ Ø A such that the multiplication map H0( L)P ×sP−→ H0( L)P is not injective. Before stating our main result we need to introduce some notation. Consider the following commutative diagram (3) 0 0 / Lt / L t e / L Le / F = L/ Lt Fe / 0 / 0 e = ker(cid:16) Le→F(cid:17). It is easy to see that the second row is exact and that L where Lt denotes the torsion subsheaf of L, Le = Im( Le→ L), Fe = Im( Le→F ) and t t L e is the torsion subsheaf of Le. It is also clear by construction that L = lim Le (c.f. −→e Theorem 2.14). Since the direct limit is exact, by its universal property there is also a commu- tative diagram 0 0 / Lt / L ≃ / F = L/ Lt L t e / lim −→e Le lim −→e lim −→e Fe / 0 / 0 + + / / / / / / / O O / / O O O O / / / / / / / / O O / / O O O O / IRREDUCIBLE THETA DIVISORS OF PPAV'S ARE STRONGLY F-REGULAR 15 L t e. Indeed, an element η ∈ Lt ֒→ L lifts to a class [ηe] ∈ Le = L and this class maps to 0 under the composition lim Le→F , so it lies in −→ t L e (by exactness of the direct limit). By the 5-lemma, we have that the right Note that Lt ≃ lim −→e lim −→ lim −→ vertical map is also an isomorphism. We are now ready to state our main technical result: Theorem 3.1. Let {Ωe} be an inverse system of coherent sheaves on a g-dimensional abelian variety satisfying the Mittag-Leffler condition. Let Λe = RSA, A(DA(Ωe)) and Λ = hocolim→Λe. Denote as above Le = Im( Le→ L) and define Ωe = RS A,A(D A( Le)). Suppose that {Ωe} is a GV-inverse system, in the sense that Hi(Λ) = 0 for any i 6= 0. If H0( L) has a torsion point P of maximal dimension g − k, then lim ←−(cid:16)RkSA, A( Ωe) ⊗ k(P )(cid:17) 6= 0. Equivalently, the maps lim ←−(cid:16)RkSA, A( Ωe) ⊗ k(P )(cid:17) →RkSA, A( Ωe) ⊗ k(P ) are non-zero for every e ≫ 0. Proof. We start by performing a sequence of reductions. Reduction 1: With the notation introduced in diagram (3), in order to show that it is sufficient to show that lim ←−(cid:16)RkSA, A( Ωe) ⊗ k(P )(cid:17) 6= 0 e, Ø A) ⊗ k(P )i 6= 0. e hExtk( L lim ←− t Indeed, consider the long exact sequence for Ext induced by the short exact sequence namely 0−→ L t e−→ Le−→Fe−→0 · · · −→Extk( Le, Ø A)−→Extk( L t e, Ø A)−→Extk+1(Fe, Ø A)−→ · · · By Lemma 6.3 in [PP08b] it follows that Extk+1(Fe, Ø A) ⊗ k(P ) = 0 for all e, so since ⊗k(P ) is right-exact, we obtain a surjection Extk( Le, Ø A) ⊗ k(P )−→Extk( L t e, Ø A) ⊗ k(P ) and hence2 a surjection Therefore, if lim t lim ←− ←− e hExtk( Le, Ø A) ⊗ k(P )i −→ lim ←−ehExtk( L e hExtk( Le, Ø A) ⊗ k(P )i [∗] e, Ø A) ⊗ k(P )i . e hExtk( L e, Ø A) ⊗ k(P )i 6= 0, then ←−(cid:16)RkSA, A( Ωe) ⊗ k(P )(cid:17) 6= 0 ≃ lim t lim ←− as claimed, where the isomorphism [∗] follows from the following computation 2Note that the system nExtk( L t e, Ø A) ⊗ k(P )oe satisfies the ML-condition. 16 ALAN MARC WATSON Extp( Le, O A) ≃ Hp−g(cid:16)D A( Le)(cid:17) ≃ Hp−g((−1 A)∗RSA, A z ≃ Hp(cid:16)(−1 A)∗RSA, A( Ωe)(cid:17) . (4) RS A,A(D A( Le))[g]) := Ωe } { Reduction 2: Denoting i : Z := {P } ֒→ A, we next reduce to showing that lim ←− e hExtk(Li∗ L t e, ØZ) ⊗ k(P )i 6= 0 In order to see this, it suffices to show that for every e there is an isomorphism (5) Extk( L t e, Ø A) ⊗ k(P ) ≃ Extk(Li∗ L t e, ØZ) ⊗ k(P ). But observe that Extk( L t e, Ø A)Z⊗k(P ) ≃ L0i∗Extk( L t e, Ø A)⊗k(P ) [1] ≃ Hk(cid:16)Li∗RHom( L [2] ≃ Extk(Li∗ L t e, Ø A)(cid:17)⊗k(P ) t e, ØZ) ⊗ k(P ) where: (i) The isomorphism in [1] follows from Grothendieck's spectral sequence3 Ep,q 2 = Lpi∗Extq( L t e, Ø A) ⊗ k(P ) =⇒ Lp+qi∗RHom( L t e, Ø A) ⊗ k(P ) t Note in the first place that Extq( L e, Ø A) = 0 near P for all q < k. Indeed, t since L e is supported on Z 4 near P , which has codimension k, our claim follows from the fact that Extq(•, Ø A) = 0 for all q < codimSupp (•) (c.f. Lemma 6.1 in [PP08b]). The differentials coming out of E0,k and the differential targeting E0,k 2 2 = L0i∗Extk( L is t e, Ø A) are hence trivial d−2,k+1 2 : L−2i∗Extk+1( L t e, Ø A) ⊗ k(P )−→L0i∗Extk( L t e, Ø A) ⊗ k(P ) which is also trivial since there is an open neighborhood U of P over which hL−2i∗Extk+1( L t e, Ø A)iU ≃ L−2i∗ U Extk+1( L t e, Ø A)U = 0 where iU : Z ∩U ֒→ U and where the last vanishing follows from the coherence of L t e and the fact that P has codimension k. (ii) The isomorphism in [2] follows from the Lf ∗RHom(F •, G•) ≃ RHom(Lf ∗F •, Lf ∗G•) (c.f. equation (3.17) in [Huy06]). 3C.f. equation (3.10) in [Huy06]. Also note that since tensoring by k(P ) is exact on D(Z), there is a spectral sequence as written. 4Note that if Z is an irreducible component of Supp L, then it is also an irreducible component of Supp H0( Le) for every e >> 0. Let Z be one such component, denote by IZ its ideal sheaf and take a section η = {ηe} ∈ L supported on Z. It is then clear that Z ⊂ Supp H0( Le) for every e >> 0. Now, if f ∈ IZ is in ann(η), it is clear that f ∈ ann(ηe) for e >> 0, so that Z = Supp ηe for all e >> 0, as claimed. IRREDUCIBLE THETA DIVISORS OF PPAV'S ARE STRONGLY F-REGULAR 17 Finally, in order to show that is non-zero, we will use the isomorphism t lim ←− e (cid:16)Extk(Li∗ L e, ØZ) ⊗ k(P )(cid:17) = lim e, ØZ) ⊗ k(P )(cid:17) k(P )(cid:16)Li∗ L Extk ←− e t lim ←− e (cid:16)Extk ØZ (Li∗ L t e ⊗ k(P ), k(P )(cid:17) 6= 0 where we used that k(P ) ≃ ØP and Proposition III.6.8 in [Har77]. Denote by i : Z = {P } ֒→ A the inclusion. By Grothendieck duality (c.f. discussion in section 2.3 from [BST12]), since all the higher direct images of a closed immersion are zero, we have a functorial isomorphism (6) RHom A(cid:16)i∗hLi∗ Le ⊗ k(P )i , k(P )[−k](cid:17) ≃ Ri∗RHomZ(cid:16)Li∗ Le ⊗ k(P ), Li!k(P )(cid:17) Homk(P )(cid:16)i∗hLi∗ L e ⊗ k(P )i , k(P )(cid:17) ≃ Hk(cid:16)Ri∗RHomk(P )(Li∗ L Taking k-th cohomology, we obtain e ⊗ k(P ), Li!k(P ))(cid:17) t t ≃ i∗Extk k(P )(Li∗ L t e ⊗ k(P ), Li!k(P )). (7) Note that the inverse limit of the left hand side is lim ←− Hom A(cid:16)i∗hLi∗ L t e ⊗ k(P )i , k(P )(cid:17) = Hom A (i∗ [Li∗ L ⊗ k(P )] , k(P )) . We claim that the latter sheaf is non-zero. Indeed, note that (8) Hom A (i∗ L ⊗ k(P ), k(P )) 6= 0 since it is simply the k(P )-dual of the non-zero k(P )-vector space LZ ⊗ k(P ). The natural map Li∗ L→L0i∗ L induces R0i∗ [Li∗( L) ⊗ k(P )] ≃−→ R0i∗ [i∗( L) ⊗ k(P )] 6=0 −→ k(P ) where the last map is just a non-zero morphism from (8) with the source sheaf ex- tended by zero. Since L is locally free in a neighborhood of P and closed immersions have no higher direct images, the fact that the first map is an isomorphism follows from the degeneration of the spectral sequence (c.f. equation (3.10) in [Huy06]) Rsi∗(cid:0)Lti∗ L ⊗ k(P )(cid:1) ≃ Rsi∗(cid:0)Ht (Li∗ L ⊗ k(P ))(cid:1) s+t=p since Lti∗ L ⊗ k(P ) = 0 for all t < 0. Hence, the inverse limit on the right hand side of (7) is also non-zero, and in particular that =⇒ Rpi∗ (Li∗ L ⊗ k(P )) Extk Z (Li∗ Le ⊗ k(P ), Li!k(P )) 6= 0 lim ←− But recall that Z is a torsion translate of an abelian subvariety of A, so ωZ ≃ ØZ ≃ ØP ≃ k(P ), so Li!k(P ) ≃ k(P ) and we may hence conclude that as claimed. (cid:3) Extk Z (Li∗ Le ⊗ k(P ), k(P )) 6= 0 lim ←− 18 ALAN MARC WATSON Remark 3.1. Theorem 3.1 has shown that lim ←−(cid:16)RkSA, A( Ωe) ⊗ k(P )(cid:17) 6= 0. where recall, we defined Ωe = RS A,A(D A( Le)) where Le = Im( Le→ L). In what follows we will drop the tildes in order to ease de notation: all we need is a projective system of coherent sheaves satisfying the generic vanishing property and inducing a non-zero limit as stated in the theorem. If A has no super-singular factors, we know by Proposition 3.3.5 in [HP13] that for every e ≥ 0, the top dimensional components of the set of points P ∈ A such that the map H0( Le)P →H0( L)P is non-zero is a torsion translate of an abelian subvariety of A. Let P ∈ A be a torsion point of maximal dimension (namely, dim (P ) is maximal such that H0( L)P has torsion) and consider W = {P }. In particular, W is a component of Supp H0( L), and we already argued earlier that W must also be an irreducible component of Supp H0( Le) for e >> 0, so W is also a top dimensional component of the support of the image of the map H0( Le)→H0( L). We thus conclude that if P ∈ A is a torsion point of maximal dimension, then W = {P } is a torsion translate of an abelian subvariety of A. In this context, Theorem 3.1 yields the following. Corollary 3.2. Let {Ωe} be a Mittag-Leffler inverse system of coherent sheaves on a g-dimensional abelian variety A with no supersingular factors, and let Ω = lim Ωe. ←− Let Λe = RSA, A(DA(Ωe)) and Λ = hocolim→Λe. Suppose that {Ωe} is a GV- inverse system, in the sense that Hi(Λ) = 0 for any i 6= 0. If P is a torsion point of H0( L) of maximal dimension (so that W = {P } is a torsion translate of an abelian subvariety of A), then there are non-zero maps lim ←−(cid:16)Rg−kSA, W (Ωe) ⊗ k(P )(cid:17) →Rg−kSA, W (Ωe) ⊗ k(P ) for e ≫ 0, where the Fourier-Mukai kernel of SA, W is given by P A× W = (id × ι)∗P A× A, P A× A being the normalized Poincar´e bundle of A × A. Proof. By Theorem 3.1 we have a non-zero map lim ←−(cid:16)Rg−kSA, A(Ωe) ⊗ k(P )(cid:17) →Rg−kSA, A(Ωe) ⊗ k(P ) for some e > 0. Consider the base-change maps Rg−kSA, A(Ωe)⊗k(P )→H g−k(A, Ωe⊗P A× A Note that the second map is an isomorphism by flat base change: indeed, denoting by ι : W ֒→ A the inclusion, we have A×{P }), Rg−kSA, W (Ωe)⊗k(P )→H g−k(A, Ωe⊗P A× W A×{P }) Rg−kSA, W (Ωe) ⊗ k(P ) def = F BC ≃ def ≃ Rg−kp′ Rg−kp W ∗(cid:16)p∗ W ∗(cid:16)p∗ W ∗(cid:16)p∗ Rg−kp′ AΩe ⊗ P A× W(cid:17) ⊗ k(P ) AΩe ⊗ P A× W ⊗ k(P )(cid:17) AΩe ⊗ (id × ι)∗P A× A ⊗ k(P )(cid:17) [∗] ≃ H g−k(A, Ωe ⊗ P A× W A×{P }) IRREDUCIBLE THETA DIVISORS OF PPAV'S ARE STRONGLY F-REGULAR 19 where in [∗] we used Proposition III.8.5 in [Har77] and where p′ W of the projection, as illustrated in the diagram is the base change A × W Spec k(P ) p′ W A × W p W Spec k(P ) flat / W However note that we have (9) A×{P }) ≃ H g−k(A, Ωe ⊗ P A× W so we can write both base change maps in the following diagram H g−k(A, Ωe ⊗ P A× A A×{P }) 6=0 Rg−kSA, A(Ωe) ⊗ k(P ) [∗] A×{P }(cid:17) H g−k(cid:16)A, Ωe ⊗ P A× A A×{P }(cid:17) / H g−k(cid:16)A, Ωe ⊗ P A× W ≃ ≃ / Rg−kSA, W (Ωe) ⊗ k(P ) lim lim ←− ←−(cid:16)Rg−kSA, A(Ωe) ⊗ k(P )(cid:17) A×{P }(cid:17) H g−k(cid:16)A, Ωe ⊗ P A× A A×{P }(cid:17) H g−k(cid:16)A, Ωe ⊗ P A× W ←−(cid:16)Rg−kSA, W (Ωe) ⊗ k(P )(cid:17) lim ←− lim ≃ ≃ where the top horizontal map is non-zero for e ≫ 0 by Theorem 3.1, the middle isomorphisms are the ones in (9), and where the isomorphisms at the bottom follow from flat base change as described above. We seek to show that the bottom horizontal map is non-zero for some e. Nev- ertheless, note that if it this were not the case, then the top horizontal map could not possibly be non-zero, since in any case the base change maps [∗] are injective by the proof of Proposition III.12.5 in [Har77]5. (cid:3) 5In a nutshell, let f : X→Y = Spec A be a projective morphism and let F be a coherent sheaf on X. For any A-module M , define T i(M ) := H i(X, F ⊗A M ), which is a covariant additive functor from A-modules to A-modules which is exact in the middle (by Proposition III.12.1 in [Har77]). Writing Ar→As→M →0 we have a diagram T i(A) ⊗ Ar T i(A) ⊗ As T i(A) ⊗ M 0 ≃ ≃ T i(Ar) / T i(As) ϕ / T i(M ) where ϕ : T i(A) ⊗ M →T i(M ) is the base change map and where the two first vertical arrows are isomorphisms. A straight-forward diagram chase then shows that ϕ is injective. / / / / / / / / O O / O O / / / / / / / / 20 ALAN MARC WATSON Remark 3.2. We will be using two different Fourier-Mukai kernels on A × W . If ι : W ֒→ A denotes the inclusion and π : A։W is the dual projection, we have a diagram A × W idA×ι A × A π×id W W × W If P A× A and P W × W denote the normalized Poincar´e bundles on A × A and W × W respectively, on A × W we may consider the locally-free sheaves (idA × ι)∗P A× A and (π × id W )∗P W × W . In Corollary 3.2 we proved a non-vanishing statement for the derived Fourier- Mukai transform with respect to the former kernel and in what follows we need an analogous statement for the transform with respect to the latter. Nevertheless, note that we are simply looking at fibers over points w ∈ W ⊂ A (concretely over the generic point of W ), and over these points both sheaves are isomorphic. Indeed, w ∈ W ⊂ A determines P W × W A×{w} ∈ Pic(A), with P A× A A×{w} ≃ π∗P W × W W ×{w} ∈ Pic(W ) and P A× A W ×{w}, and therefore ≃P A× A A×{w} { z h(idA × ι)∗P A× AiA×{w} } ≃ ≃π∗PW × W W ×{w} z { h(π × id W )∗P W × WiA×{w} } 4. Fibering of the Albanese image In [EL97], Ein and Lazarsfeld showed the following statement: Theorem 4.1 (see [EL97], Theorem 3). X is a smooth projective variety of maxi- mal Albanese dimension over a field of characteristic zero and χ(ωX ) = 0, then the image of the Albanese map is fibered by subtori of A. Sketch of proof. We sketch the proof given in [PP08] (Theorem E). If χ(ωX ) = 0, it follows that V 0(ωX ) ⊂ A is a proper subvariety (c.f. Lemma 1.12(b) in [Par11]). Choose an irreducible component W ⊂ V 0(ωX ) of codimension p > 0, which is a torsion translate of an abelian subvariety of A that we also denote by W . Let π : A։ W denote the dual projection and consider the diagram X a / Y := a(X) / A ✉ ✉ ✉ ✉ h=πY W ✉ ✉ π ✉ ✉ ✉ ✉ z✉ Since the fibers of the projection A։ W are abelian subvarieties of dimension p, the conclusion of the theorem will follow provided that f ≥ p, where f denotes the dimension of the generic fiber of h. In order to see this, recall the following standard facts: (a) a∗ωX is a GV-sheaf on Y = a(X) and V 0(ωX ) = V 0(a∗ωX ). / / / / z z z IRREDUCIBLE THETA DIVISORS OF PPAV'S ARE STRONGLY F-REGULAR 21 (b) If W is an irreducible component of V 0(a∗ωX) of codimension p, then it is also a component of V p(a∗ωX ). (c) a∗ωX is a GV−f -sheaf with respect the the Fourier-Mukai functor with kernel (π × 1W )∗P W × W , so in particular codimW V p(a∗ωX ) ≥ p − f for every p ≥ 0. By (b) we have W ⊆ V p(a∗ωX) ⊆ W so that codimW V p(a∗ωX ) = 0, and finally (c) yields f ≥ p, which is what we sought to show. (cid:3) Our goal in this section is to prove a positive characteristic analogue of Theorem 4.1. Let X be a smooth projective variety of maximal Albanese dimension and denote by a : X→A the Albanese map. Then a∗ωX is a Cartier module and we may consider the inverse system {Ωe = F e ∗ S0a∗ωX }e. Define Le = RSA, ADA(Ωe) and set L = hocolim→ Le. By Corollary 3.3.1 in [HP13] we have that H i(Ω ⊗ Pα) = 0 for every i > 0 and very general α ∈ A. Thus, defining as above χ(Ω) := χ(Ω ⊗ Pα) for very general α ∈ A, we see that χ(Ω) = h0(Ω ⊗ Pα) and it seems that in trying to extend Theorem 4.1 to positive characteristic, one should assume that h0(Ω ⊗ Pα) = 0. This leaves us in a setting which is similar to the one we encountered in the proof of Theorem 5.2. If rk( L) = h0(Ω ⊗ Pα) = 0, then in particular Λ must be a torsion sheaf. In light of this observation, we show the following: Theorem 4.2. Let X be a smooth projective variety of maximal Albanese dimen- sion and let a : X→A be a generically finite map to an abelian variety A with no supersingular factors. Let g = dim A. Consider the inverse system {Ωe = ∗ S0a∗ωX }e and denote Ω = lim F e Ωe. Define Le = RSA, ADA(Ωe) and assume that ←− the sheaf H0( L) = lim H0( Le) has torsion. Then the image of the Albanese map is −→ fibered by linear subvarieties of A. Proof. Let w ∈ A be a torsion point of H0( L) of maximal dimension k; by the remark preceding Corollary 3.2, we have that W := {w} ⊂ A is a torsion translate of an abelian subvariety of A, which we still denote by W . Denote by π : A→W the projection dual to the inclusion W ֒→ A. Y := a(X) y A (cid:26) h=πY (cid:26) π (cid:26) (cid:26) (cid:26)(cid:29) W By Corollary 3.2 we know that the map lim ←−(cid:16)Rg−kSA, W (Ωe) ⊗ k(w)(cid:17) →Rg−kSA, W (Ωe) ⊗ k(w) is non-zero for every e ≫ 0, where the Fourier-Mukai kernel of SA, W is given by (id × ι)∗P, P being the normalized Poincar´e bundle of A × A. Recall that, in general, even though the system {Ωe} satisfies the Mittag-Leffler condition, the inverse system {RtS(Ωe)}e may fail to do so (c.f. Example 3.2 w 22 ALAN MARC WATSON in [WZ14]). We handle the Mittag-Leffler case first, however, since the proof is neater and the subsequent generalization does not rely on new ideas. Case in which {Rg−kSA, W (Ωe)}e satisfies the ML-condition. The proof in this case goes along the lines of that of Theorem E in [PP08]. Note that if w is the generic point of W ֒→ A, we have the following: w [1] [2] lim ←− ∈ (w ∈ W : = (w ∈ W : lim ⊆ (w ∈ W : lim e (cid:16)Rg−kSA, W (Ωe) ⊗ k(w)(cid:17) 6=0 Rg−kSA, W (Ωe)! ⊗ k(w) Rg−kSA, W (Ωe)!w ←− ←− e e ⊆ W −→ Rg−kSA, W (Ωe) ⊗ k(w)) −→ Rg−kSA, W (Ωe) ⊗ k(w)) 6=0 6=0 −→(cid:16)Rg−kSA, W (Ωe)(cid:17)w) where w lies in the first set by Corollary 3.2 and where the equality [2] follows from Lemma 2.12, since we are under the assumption that the system {Rg−kSA, W (Ωe)}e satisfies the Mittag-Leffler condition. This implies that the codimension (in W ) of the support of image of the map Rg−kSA, W (Ωe)−→Rg−kSA, W (Ωe) lim ←− e is zero (this support is closed - under the Mittag-Leffler assumption - by Propo- sition 4.3 in [WZ14]). At the same time, by Proposition 2.19 we know that this codimension must be ≥ g − k − f , where f is the dimension of a general fiber of h, so in particular f ≥ g − k and this concludes the proof under the Mittag-Leffler assumption (indeed, the fibers of the projection A։W are abelian subvarieties of dimension g − k). General case. We finally observe that, in our setting, we can actually do without the Mittag-Leffler assumption on {Rg−kSA, W (Ωe)}e. We only used this assumption in order to guarantee the closedness of the support of the image of the and in order to ensure that 6=0 map(cid:16)lim ←−e Rg−kSA, W (Ωe)(cid:17)w the inverse limit commutes with ⊗k(w). −→(cid:16)Rg−kSA, W (Ωe)(cid:17)w Note in the first place that we don't need the support of the image of the above map to be closed for the previous argument to work. Proposition 2.19 shows that in order for w to belong to the support, we need codim{w} ≥ g − k − f , and this suffices in order to conclude that f ≥ g − k. With regards to the commutation of the inverse limit and ⊗k(w), note in the first place that there is always an inclusion ⊇ induced by the natural map lim ←− e (cid:16)Rg−kSA, W (Ωe)(cid:17) ⊗ k(w)−→ lim e (cid:16)Rg−kSA, W (Ωe) ⊗ k(w)(cid:17) ←− Moreover, in our setting, the opposite inclusion ⊆ follows from the flatness of k(w) as an Ø W -module and the fact that the projection formula and its consequences still hold in the category of quasi-coherent sheaves under some perfection assumptions then lim ←− e (cid:16)Rg−kSA, W (Ωe) ⊗ k(w)(cid:17) 6=0 e (cid:16)Rg−kSA, W (Ωe)(cid:17) ⊗ k(w) 6=0 lim ←− −→ Rg−kSA, W (Ωe) ⊗ k(w) −→ Rg−kSA, W (Ωe) ⊗ k(w) IRREDUCIBLE THETA DIVISORS OF PPAV'S ARE STRONGLY F-REGULAR 23 (c.f. Lemma 71 in [Mur06]). We state this below as a lemma, and the proof is hence complete. (cid:3) Lemma 4.3. With the same notations as above, if In particular, equality [2] in the above chain still holds. Proof. As in the proof of Proposition 2.19, let Le = R SA, W (DA(Ωe)), where Note in the first place that we have the following isomorphisms of Ø W -modules: SA, W denotes the Fourier-Mukai transform with kernel P ∨, with P =(cid:0)π × 1 W(cid:1)∗ e (cid:16)Rg−kSA, W (Ωe) ⊗ k(w)(cid:17) [1] ( Le, Ø W ) ⊗ k(w)(cid:17) ≃ lim ←− lim ←− Ø W P W × W . e [3] ←− ≃ lim ←− [2] ≃ lim ←− e (cid:16)Extg−k e (cid:16)Hg−k(cid:16)RHomØ W Hg−k(cid:16)RHomØ W e (cid:16)RHomØ W ( Le, Ø W )(cid:17) ⊗ k(w)(cid:17) ( Le, Ø W ) ⊗ k(w)(cid:17) ( Le, Ø W ) ⊗ k(w)(cid:17)! ≃ Hg−k lim ( Le, k(w))! ≃ Hg−k lim ( L, k(w))(cid:17) ≃ Hg−k(cid:16)RHomØ W ( L, Ø W ) ⊗ k(w)(cid:17) ≃ Hg−k(cid:16)RHomØ W ( L, Ø W )(cid:17) ⊗ k(w) ≃ Hg−k(cid:16)RHomØ W ≃ Hg−k(cid:16)holim←RSA, W (Ωe)(cid:17) ⊗ k(w) RHomØ W ←− [8] [5] [6] [7] [4] e where [1] and [8] follow from the computations (1) and (2) in the proof of Proposi- tion 2.19, [2] and [7] follow from the flatness of ⊗k(w) as an Ø W -module, [3] follows from Proposition 2.11, since the system (cid:8)Extp−1( Le, Ø W ) ⊗ k(w)(cid:9)e satisfies the ML-condition, and [4] and [6] follow from the isomorphism6 RHom(F , G) ⊗ H ≃ RHom(F , G ⊗ H). Hence, by assumption we have a non-zero map Hg−k(cid:16)holim←RSA, W (Ωe)(cid:17) ⊗ k(w) 6=0 −→ Rg−kSA, W (Ωe) ⊗ k(w) 6This isomorphism holds for complexes of sheaves of modules F , G, H provided that either F or H are perfect (c.f. Lemma 71 in [Mur06]). Note that k(w) is a perfect complex, being a coherent sheaf. 24 ALAN MARC WATSON and the conclusion of the lemma then follows from the following commutative dia- gram lim ←−e(cid:16)Rg−kSA, W (Ωe)(cid:17) ⊗ k(w) Hg−k(cid:16)holim←RSA, W (Ωe)(cid:17) ⊗ k(w) 4✐✐✐✐✐✐✐✐✐✐✐✐✐✐✐✐✐✐ 6=0 / Rg−kSA, W (Ωe) ⊗ k(w) (cid:3) In particular, within the context of principally polarized abelian varieties, the same argument yields the following statement: Corollary 4.4. Let (A, Θ) is a principally polarized abelian variety with no super- singular factors defined over an algebraically closed field of characteristic p > 0. Assume further that Θ is irreducible. Consider the inverse system {Ωe := F e ∗ (ωΘ ⊗ τΘ)}e on A and set L = hocolim→eRSA, ADA(Ωe). Then L is a torsion-free quasi- coherent sheaf concentrated in degree 0. Proof. The fact that L = H0( L) is a quasi-coherent sheaf concentrated in degree zero follows from Theorem 2.14(i), since ωΘ is a Cartier module. Assume for a contradiction that H0( L) is not torsion-free and fix an irreducible component W := {w} ֒→ A of maximal dimension of the closure of the set of torsion points of H0( L). Denote by π : A։W the dual projection. A ⑥ ⑥ ⑥ ⑥ π ⑥ ⑥ ~⑥ Θ h W We may then argue as in the proof of Theorem 4.2 to conclude that Θ is fibered by abelian subvarieties of A, but this is not possible given that Θ is irreducible (and hence of general type) in light of Abramovich's work (c.f. Section 2.3 or [Abr95]). (cid:3) 5. Singularities of Theta divisors We now focus on the singularities of Theta divisors and embark on the proof of Theorem 1.7. As a warm-up, we focus on simple abelian varieties to start with, namely those which do not contain smaller dimensional abelian varieties. 5.1. Case of simple abelian varieties. The crux of the argument resides in the construction of sections of ØA(Θ) which vanish along the test ideal τ (Θ) and, in the case of simple abelian varieties, it is a direct consequence of the results in [HP13]. The proof of the general case will follow the same pattern, albeit further work will be required to prove that the required sections exist. Theorem 5.1. Let (A, Θ) be a PPAV over an algebraically closed field K of char- acteristic p > 0 such that A is simple and ordinary. Then Θ is strongly F-regular. (In particular, Θ is F-rational, and by Lemma 2.34 in [BST12], it is normal and Cohen-Macaulay). / O O 4 / / ~ ~ ~ IRREDUCIBLE THETA DIVISORS OF PPAV'S ARE STRONGLY F-REGULAR 25 Proof. On A consider the inverse system Ωe = F e ∗ Ω0, where Ω0 = ωΘ ⊗ τ (Θ). This yields a direct system Λe = R SDAΩe equipped with natural maps H0(Λe)→H0(Λ) = Λ = hocolim→Λe. By 2.14, we know that Λ is quasi-coherent sheaf in degree 0. Consider the set Z =nα ∈ A : Im(cid:16)H0(Λ0) ⊗ Ø A,α−→Λ ⊗ Ø A,α(cid:17) 6= 0o By Proposition 2.16 we know that Z is a finite union of torsion translates of subtori. Since A is simple by assumption, this implies that either Z = A or Z is a finite set. Assume for a contradiction that Z is finite. By Proposition 2.17 we have t∗ xΩ = Ω for every x ∈ dA/Z = bA, so that Supp Ω = A. Since the maps in the inverse system F e Supp Ω = Supp Ω0 = Supp ωΘ ⊗ τ (Θ) = A, which is absurd. ∗ Ω0 = F e ∗ (ωΘ ⊗ τ (Θ)) are surjective, we know by Proposition 2.17 that We must thus have Z = A, so that Supp H0(Λ0) = A, and hence cohomology and base change yields H 0(A, ωΘ ⊗ τ (Θ) ⊗ Pα) 6= 0 for all α ∈ A. Consider the following commutative diagram: H 0(A, Pα) / H 0(A, ØA(Θ) ⊗ Pα) / H 0(Θ, ØA(Θ)Θ ⊗ Pα) H 0(A, K ⊗ Pα) H 0(A, ØA(Θ) ⊗ Pα ⊗ τ ) H 0(Θ, ØA(Θ)Θ ⊗ τ (Θ) ⊗ Pα) where K is defined so that the diagram commutes. In the top row we have H 1(A, Pα) = 0 for α 6= 0 since Pα is topologically trivial. The polarization in- duced by Θ is principal, so h0(ØA(Θ) ⊗ Pα) = 1. Since by the above discussion H 0(Θ, ØA(Θ)Θ ⊗ Pα ⊗ τ (Θ)) 6= 0, it follows that H 0(Θ, ØA(Θ)Θ ⊗ Pα) 6= 0 and hence both the right inclusion and the top right restriction are equalities. The ideal sheaf τ on A is defined as follows: fix an open subset U = Spec R ⊆ A and assume that Θ is given by an ideal sheaf I = I(Θ). Let J = τΘ(U ) be the test ideal of Θ and let J ⊂ R be an ideal such that J = J/I. Omitting the twist by Pα, the diagram above locally boils down to R/ J ≃ / / (R/I)/( J/I) ≃ R/ J 0 0 / I / I / R J R/I J 0 0 so τ (U ) = J. Now taking cohomology, a section s ∈ H 0(J) embeds as ¯s ∈ H 0(R/I) and maps to zero in H 0(R/ J) by exactness. By exactness of the second row, ¯s lifts to ¯s ∈ H 0(R), which still projects to zero in H 0(R/ J) by commutativity of the top square, so ¯s must lift to a non-zero section of H 0( J). / / / / ? O O / / ? O O ? O O / / / / O O / / O O / / / O O / / O O / / O O 26 ALAN MARC WATSON Finally, since H 0(Θ, ØA(Θ)Θ ⊗ Pα ⊗ τ (Θ)) 6= 0 for every α ∈ A and these sections lift to sections of H 0(A, ØA(Θ) ⊗ Pα) vanishing along τ , we conclude that h0(A, ØA(Θ) ⊗ Pα ⊗ τ ) = 1. Hence if τ were not trivial, we would have Zeros(τ ) ⊂ Θ + αP for every αP ∈ A (where αP ∈ A is the point corresponding to Pα ∈ Pic0(A)), which is absurd since these translates of Θ don't have any points in common. We thus conclude that τ = ØA, and hence τ (Θ) = ØΘ so that Θ is strongly F-regular. (cid:3) In the proof of Theorem 5.1 we used the simplicity of A in order to show that H 0(A, ωΘ ⊗ τ (Θ) ⊗ Pα) 6= 0 for all α ∈ A. The same argument we employed above will work in the general case provided that we can show the existence of sections in H 0(A, ωΘ ⊗ τ (Θ) ⊗ Pα), and it turns out that this is somewhat more involved. 5.2. General case. We finally study singularities of Theta divisors in the gen- eral setting. As we mentioned earlier, the argument will be analogous to the one employed to prove the theorem in the case of simple abelian varieties, although ad- ditional work will be required to prove that there exist sections in H 0(Θ, ØA(Θ)Θ ⊗ τ (Θ) ⊗ Pα). The main ingredient in Ein and Lazarsfeld's proof over fields of characteristic zero was that given a smooth projective variety X of maximal Albanese dimension such that χ(X, ωX) = 0, the image of its Albanese morphism is fibered by tori (c.f. Theorem 3 in [EL97]). Our proof will rely on Corollary 4.4, where we proved that if the sheaf H0( L) associated to the inverse system {F e ∗ (ωΘ ⊗ τΘ)} was not torsion- free, then Θ would be fibered by tori, which is impossible since Θ is irreducible. In a nutshell, and as in the case of simple abelian varieties, Θ will be strongly F-regular provided that there exist non-trivial sections in H 0(Θ, ØA(Θ) ⊗ ØΘ ⊗ τ (Θ) ⊗ Pα) and we will show that if that was not the case, then the sheaf H0( L) would have torsion, a contradiction. Theorem 5.2. Let (A, Θ) be an ordinary principally polarized abelian variety over an algebraically closed field k of characteristic p > 0. If Θ is irreducible, then Θ is strongly F-regular. Proof. The proof goes along the lines of Theorem 5.1: consider again the com- mutative diagram: H 0(A, Pα) / H 0(A, ØA(Θ) ⊗ Pα) / H 0(Θ, ØA(Θ)Θ ⊗ Pα) H 0(A, K ⊗ Pα) H 0(A, ØA(Θ) ⊗ Pα ⊗ τ ) H 0(Θ, ØA(Θ)Θ ⊗ τ (Θ) ⊗ Pα) In the proof of Theorem 5.1 we used the simplicity of A to conclude easily that z ωΘ } { H 0(Θ, ØA(Θ) ⊗ ØΘ ⊗τ (Θ) ⊗ Pα) 6= 0 It then followed from the commutative diagram above that H 0(A, ØA(Θ)⊗Pα⊗τ ) 6= 0 and this in turn forced τ to be trivial, whence τ (Θ) = ØΘ. We shall now use the previous results in order to conclude that H 0(Θ, ωΘ ⊗ τ (Θ) ⊗ Pα) 6= 0. In a nutshell, assuming for a contradiction that τ (Θ) is not trivial / / / / ? O O / / ? O O ? O O IRREDUCIBLE THETA DIVISORS OF PPAV'S ARE STRONGLY F-REGULAR 27 we will show that 0 6= S0(Θ, ωΘ ⊗ τ (Θ) ⊗ Pα) ⊆ H 0(Θ, ωΘ ⊗ τ (Θ) ⊗ Pα) for general α ∈ A. The diagram above then yields H 0(A, ØA(Θ) ⊗ Pα ⊗ τ ) 6= 0 and since by assumption τ (Θ) 6= ØΘ, we conclude that Zeroes(τ ) ⊂ Θ + αP for general αP ∈ A (as before, αP denotes the point in A corresponding to Pα ∈ Pic0A), but this is not possible since general translates of Θ do not have points in common. Therefore we must have τ (Θ) = ØΘ, and hence Θ is strongly F-regular. We now argue by contradiction: let Ω = lim ←− ∗ S0ωΘ = lim F e ←− F e ∗ (ωΘ ⊗ τ (Θ)). By Corollary 3.2.1 in [HP13] and its proof, for all closed α ∈ A we have that H0( L) ⊗ k(α) ≃ H 0(Θ, lim ←− Ωe ⊗ P ∨ α )∨ so assuming for a contradiction that S0(Θ, ωΘ ⊗ τ (Θ) ⊗ Pα) = 0 for general α ∈ A, it follows that rk(Λ) = 0, and hence that H0(Λ) has torsion. Nevertheless, this is not possible by Corollary 4.4, since we are assuming that Θ is irreducible, and this concludes the proof. (cid:3) References [Abr95] D. Abramovich, Subvarieties of semiabelian varieties, 1995. [BL92] C. Birkenhake and H. Lange, Complex abelian varieties, Vol. 302. Springer, 1992. [BS12] M. Blickle and K. Schwede, p−1-linear maps in algebra and geometry, arXiv:1205.4577v2, 2012 [BST12] M. Blickle, K. Schwede and K. Tucker, F-singularities via alterations, arXiv:1107.3807, [Car08] [EL97] 2012 J. Carter, The Morava K-Theory Eilenberg-Moore spectral sequence, New York J. Math. 14, 495-515, 2008 L. Ein and R. Lazarsfeld Singularities of theta divisors and the birational geometry of irregular varieties, J. Amer. Math. Soc 10:1, 243-258, 1997 [GL87] M. Green and R. Lazarsfeld, Deformation theory, generic vanishing theorems and some conjectures of Enriques, Catanese and Beauville, Invent. Math. 90, 389-407, 1987 [GL91] M. Green and R. Lazarsfeld, Higher obstructions to deforming cohomology groups of line bundles, Jour. of. A.M.S. 4, 87-103, 1991 [Hac04] C. Hacon. A derived category approach to generic vanishing, J. Reine Angew. Math. 575, 173-187, 2004 [Hac11] C. Hacon. Singularities of pluri-theta divisors in characteristic p > 0, arXiv:1112.2219v1, 2011 [HH90] M. Hochster and C. Huneke. Tight closure, invariant theory, and the Brianon-Skoda theorem, J. Amer. Math. Soc. 3 (1990), no. 1, 31-116 [HK12] C. Hacon and S. Kov´acs. Generic vanishing fails for singular varieties and in charac- teristic p > 0, arXiv:1212.5105 [HP13] C. Hacon and Z. Patakfalvi. Generic vanishing in characteristic p > 0 and the charac- terization of ordinary abelian varieties, arXiv:1310.2996, 2013 [Har77] R. Hartshorne. Algebraic geometry. Graduate Texts in Mathemaics, 52, Springer, 1977. [Har78] R. Hartshorne. On the de Rham cohomology of algebraic varieties, Publ. Math. de l'IHES, vol. 45, 5-99, 1978, 2007 [Har98] N. Hara. A characterization of F-rational singularities in terms of injectivity of Frobe- nius maps, American Journal of Mathematics, Volume 120, Number 5, pp. 981-996, 1998 [HW02] N. Hara and K.I. Watanabe. F-regular and F-pure rings vs. log terminal and log canon- ical singularities, J. Algevraic Geom. 11, no. 2, 363-392, 2002 [Huy06] D. Huybrechts, Fourier-Mukai transforms in algebraic geometry, Oxford Mathematical Monographs, 2006 [KV81] Y. Kawamata and E. Viehweg. On a characterization of abelian varieties in the classi- fication theory of algebraic varieties, Comp. Math. 41, 355-360, 1981 [Laz04] R. Lazarsfeld. Positivity in Algebraic Geometry I and II, 48-49, Springer-Verlag, 2004 28 ALAN MARC WATSON [Mor00] S. Mori. Classification of higher-dimensional algebraicd varieties. Proc. Symp. Pue Math. 46 (1987), 269-332 [Muk81] S. Mukai. Duality between D(X) and D( X) with its application to Picard scheaves, Nagoya Math. J. 81, 133-175, 1981 [Mur06] D. Murfet. Derived categories of quasi-coherent sheaves, 2006 [Nee96] A. Neeman. The Grothendieck duality theorem via Bousfield's techniques and Brown representability, J. Amer. Math. Soc. 9, 205-235, 1996 [OSS80] C. Okonek, M. Schneider and H. Spindler. Vector bundles on complex projective spaces, Progress in Mathematics 3, Birkhiser, Boston, 1980. [Par11] G. Pareschi. Basic results on irregular varieties via Fourier-Mukai methods, Current Developments in Algebraic Geometry, MRSI Publications, 59, 2011 [PP03] G. Pareschi and M. Popa, Regularity on abelian varieties I, arXiv:math/0110003 [PP08] G. Pareschi and M. Popa, Regularity on abelian varieties III: relationship with generic vanishing and apploications, arXiv:0802.1021, 2008 [PP08b] G. Pareschi and M. Popa, Strong generic vanishing and a ahigher dimensional Castelnuovo-de Franchis inequality, arXiv:0808.2444, 2008 [PP11] G. Pareschi and M. Popa. GV-sheaves, Fourier-Mukai transform and generic vanishing, Smer. J. Math. J. 133:1, 235-271, 2011 [PR03] R. Pink and D. Roesddler. A conjecture of Beauville and Catanese revisited, Mathema- tische Annalen, Volume 330, Issue 2, pp 293-308, 2004 [Sch09] K. Schwede. F-adjunction, arXiv:0901.1154 [Sch12] K. Schwede. A canonical linear system associated to adjoint divisors in characteristic p > 0, arXiv:1107.3833 [Smi97] K. Smith. F-rational rings have rational singularities, Amer. J. Math. 119 (1997), no. 1, 159180, 1997. [Tak04] S. Takagi. An interpretation of multiplier ideals via tight closure, J. Algebraic. Geom. 13, 393415, 2004 [Wan11] J. Wang. Quotients of algebraic groups, 2011 [Wei94] C. A. Weibel. An introduction to homological algebra, Cambridge University Press, 1994 [WZ14] A. Watson and Y. Zhang. On the generic vanishing theorem of Cartier modules, arXiv:1404.2669 Department of Mathematics, University of Utah, Salt Lake City, Utah 84112 E-mail address: [email protected]
1602.01227	3	1602	2016-09-27T13:51:28	Determinantal variety and normal embedding	[ "math.AG", "math.DG" ]	The space of matrices of positive determinant GL^+_n inherits an extrinsic metric space structure from R^{n^2}. On the other hand, taking the infimum of the lengths of all paths connecting two points in GL^+_n gives an intrinsic metric. We prove bilipschitz equivalence for intrinsic and extrinsic metrics on GL^+_n, exploiting the conical structure of the stratification of the space of n by n matrices by rank.	math.AG	math	DETERMINANTAL VARIETY AND NORMAL EMBEDDING KARIN U. KATZ, MIKHAIL G. KATZ, DMITRY KERNER, AND YEVGENY LIOKUMOVICH Abstract. The space GL+ n of matrices of positive determinant in- herits an extrinsic metric space structure from Rn . On the other hand, taking the infimum of the lengths of all paths connecting a pair of points in GL+ n gives an intrinsic metric. We prove bilips- chitz equivalence between intrinsic and extrinsic metrics on GL+ n , exploiting the conical structure of the stratification of the space of n × n matrices by rank. 2 1. Introduction Consider the group GL+(n, R) of n × n matrices of positive deter- minant. It is an open submanifold of the space Rn2 of n × n matrices. Here GL+(n, R) carries two metrics: its extrinsic ambient Euclidean metric with distance function dext, and the induced intrinsic metric (i.e. least length of path) with distance function dint. Our main result is the following. Theorem 1.1. There is a constant C = C(n) such that dint < Cdext. Thus the determinantal variety is normally embedded ; in other words, the intrinsic and extrinsic metrics are bilipschitz equivalent. We prove Theorem 1.1 first for n = 2, 3 and then for general n. The proof uses in an essential way the conical structure of the stratification of the space Rn2 Indeed, the extrinsic and in- trinsic metrics on a set as simple as {(x, y) : x2 − y3 > 0} are clearly inequivalent, due to the fact that the curve x2 − y3 = 0 has a cusp. of n × n matrices by rank. Bilipschitz equivalence has been studied by a number of authors; see e.g., [1], [2], [4], [7]. For a study of the Lipschitz condition in an infinitesimal context see [3], [6]. 2010 Mathematics Subject Classification. 53C23, 58A35, 32S60. Key words and phrases. Determinantal variety, intrinsic metric, bilipschitz equivalence, conical stratification. 1 2 K. K., M. K., D. K., AND Y. L. 2. Solution in dimension 2 For 2 × 2 matrices with coordinates given by the entries (aij), the condition a11a22 − a12a21 = 0 for the determinantal variety in the space of matrices translates into (a11 + a22)2 − (a11 − a22)2 − (a12 + a21)2 + (a12 − a21)2 = 0 or to simplify notation x2 + y2 = z2 + w2. (2.1) Definition 2.1. Let X2 = {A ∈ R22 minantal variety. : det(A) = 0} denote the deter- Lemma 2.2. The intersection X2 ∩ S3 with the unit sphere S3 ⊆ R22 is a flat 2-torus, namely the Clifford torus T ⊆ S3. Indeed, by (2.1) the complement S3 \ T is a union of two solid tori, consisting respectively of matrices of positive and negative determi- nant. Note that X2 is a linear cone over T ⊆ R22 Lemma 2.3. The metrics dint and dext on T ⊆ R22 equivalent. are bilipschitz . Proof. By rescaling, we can assume that the furthest of the two points is at unit distance from the origin. By compactness, the only problem for a pair of points p, q can arise when the points collide, i.e., d(p, q) tends to 0. By the smoothness of the Clifford torus, one has dint(p,q) dext(p,q) → 1 as d(p, q) → 0. (cid:3) Lemma 2.4. Bilipschitz equivalence holds for the intrinsic and the extrinstic metrics on X2. Proof. Let p, q ∈ X2. If the apex O ∈ X2 is one of the points p, q then the intrinsic and the Euclidean distances between them coincide by linearity of the cone. Thus we can assume that p 6= O and q 6= O. To connect a pair of points p, q, at different levels in the cone by a path lying in the cone, assume without loss of generality that p is further than q from the apex of the cone. We slide p along the ray toward the apex until it reaches the level of q, and then connect it to q by a shortest path contained in that level. The length of the combined path is clearly bilipschitz with the extrinsic distance in the ambient R22 (cid:3) . Proposition 2.5. The intrinsic and the ambient metrics on the mani- fold of 2 × 2 matrices of positive determinant are bilipschitz equivalent. DETERMINANTAL VARIETY AND NORMAL EMBEDDING 3 Proof. The closure of GL+(2, R) is a linear cone over the solid torus. Meanwhile, GL+(2, R) itself is the cone without the apex over the in- terior of the solid torus. We consider a straight line path in R22 con- necting a pair of points in GL+(2, R). The length of this path is the extrinsic distance by definition. The path does not necessarily lie entirely inside GL+(2, R). As the path is straight and the determinantal variety X2 is algebraic, the path splits into a finite number of segments satisfying the following: (1) the interior of each segment lies fully either inside GL+(2, R) or inside the component GL−(2, R); (2) the endpoints are in X2. Applying Lemma 2.4, we replace every segment that lies in the com- ponent GL−(2, R) by an arc that lies in X2. Finally, we push out the arc in X2 into GL+(2, R), i.e., replace it by a nearby arc "just inside" GL+(2, R). We will explain the push-out procedure in detail since a similar argument will be used in the general case. The vertex O ∈ X2 is the unique singular point of the cone, i.e., the complement X2\{O} is a smooth manifold. Hence a path in X2 disjoint from O can be pushed out infinitesimally into GL+(2, R) by following the normal direction, without significantly affecting its length. Suppose a path joining P, Q ∈ X2 passes through O. Using the linear structure of the cone, the path can be replaced by the shorter path given by the union of the straight line segments P O ∪ OQ ⊆ X2. Next, P and Q can be replaced by nearby points P ′, Q′ ∈ GL+(2, R). We now form a path P ′O ∪ OQ′ lying entirely within GL+(2, R) except for the single point O. Let p ∈ OP ′ be the point of intersection of the segment with a sphere S(O, ǫ) of small radius ǫ > 0 centered at O, and similarly for point q ∈ OQ′. The intersection S(O, ǫ) ∩ GL+(2, R) is an open solid torus and therefore connected. Hence there is a short path γ ⊆ S(O, ǫ) ∩ GL+(2, R) joining p to q. The resulting path P ′p ∪ γ ∪ qQ′ is only slightly longer than the original path joining P and Q. This completes the proof of the bilipschitz property for GL+(2, R). (cid:3) 3. Singularities in dimension 3 For 3 × 3 matrices, the singular locus of the determinantal vari- : det A = 0} consists of matrices of rank ≤ 1. The ety X3 = {A ∈ R32 4 K. K., M. K., D. K., AND Y. L. variety X3 can be stratified as follows: X3 = X3,0 ∪ X3,1 ∪ X3,2. Here the stratum X3,i consists of matrices of rank i. The stratum X3,1 lies in the closure of X3,2, while X3,0 lies in the closure of X3,1. Each X3,i is smooth. Here X3,2 is of codimension 1 in the space of matrices, while X3,1 is of codimension 4 in the space of matrices, and X3,0 is a single point. The closure X 3,2 is a cone on a smooth manifold away from X3,1. Thus the only potential obstruction to bilipschitz equivalence is the singularity of X 3,2 along X 3,1, which we now analyze. Lemma 3.1. The 5-dimensional closure X 3,1 is a linear cone over the smooth compact 4-manifold (S2 × S2)/{±1}. Proof. Here S2 × S2/{±1} parametrizes the intersection of X3,1 with the unit 8-sphere in the space of matrices. The manifold S2 ×S2 can be parametrized by pairs of unit column vectors v, w ∈ S2 ⊆ R3 producing a rank 1 matrix v tw ∈ X3,1. The element −1 acts simultaneously on both factors of S2 × S2 by the antipodal map. (cid:3) Fix an element x ∈ X3,1. We would like to understand the bilips- chitz property for X3 in a neighborhood of x. Exploiting the action of GL(3, R) × GL(3, R) on X3,1 by right and left multiplication, we may assume that x = 1 ⊕(cid:18)0 0 0 0(cid:19) . Definition 3.2. Let Lx = {1 ⊕ A : A ∈ R22 } ⊆ R32 be a transverse slice to X3,1 ⊆ R32 at x, so that Lx ∩ X 3,2 = {1 ⊕ A ∈ Lx : det(A) = 0}. (3.1) Note that of course transversality is weaker than orthogonality. The ambient Euclidean metric on R9 plays no role here. Lemma 3.3. A transverse slice for X3,1 ⊆ X 3,2 in local coordinates is a linear cone over a Clifford torus. Proof. The defining equation (3.1) of the slice is det(A) = 0. By Lemma 2.2 this is a cone over a torus. (cid:3) DETERMINANTAL VARIETY AND NORMAL EMBEDDING 5 Now let UT ⊆ R5 be the unit ball. Choose open sets UN ⊆ Lx and U ⊆ R32 , each containing x, and a bilipschitz diffeomorphism φ : U → UT × UN (3.2) such that φ(U ∩ X 3,1) = UT × {x} and φ(U ∩ X 3,2) = UT × (UN ∩ X 3,2). This can be done using the GL(3, R) × GL(3, R) action, as explained in [5, p. 138]. Hence Lemmas 3.3 and 2.4 imply that the intrinsic and extrinsic metrics on U ∩ X 3,2 are bilipschitz equivalent. Corollary 3.4. The intrinsic and extrinsic metrics on X 3,2 = X3 are bilipschitz equivalent. This follows from Lemma 3.1 by a compactness argument. Theorem 3.5. The intrinsic and extrinsic metrics on GL+(3, R) are bilipschitz equivalent. Proof. The extrinstic distance between a pair of matrices of positive determinant is the length of the straight line path in R32 joining them. We partition the path into finitely many segments, where the interior of each segment lies entirely in a connected component of GL(3, R) while the endpoints are in X3 ⊆ R32 . Then we apply Corollary 3.4 to replace each segment belonging to the component GL−(3, R) by an arc in X3. If the arc in X3 lies in the smooth part X3,2 ⊆ X3 then it can be pushed out into GL+(3, R) by a small deformation in the normal direction, as in Section 2. Unlike the case of Proposition 2.5, the determinantal variety is not a cone on a manifold, so that an additional argument is required to push the arc out of X3 and into GL+ 3 while retaining bilipschitz control. If an arc in X3 joining points P, Q passes through the apex O ∈ X3 then it can be replaced by the union P O ∪ OQ and pushed out into GL+(3, R) as in the proof of Proposition 2.5. Otherwise we exploit the local product structure on X3 \ {O} as in (3.2). A path in X = X3 that dips into the singular locus X3,1 ⊆ X can be handled as follows. Given a path γ : [0, 1] → X, let a ∈ [0, 1] be the least parameter value t such that γ(t) is contained in the singular locus X3,1 ⊆ X, and b ∈ [0, 1] the greatest such value. Step 1. Consider the restriction of the path γ to [a, b] ⊆ [0, 1]. We replace it by a path that lies entirely in the singular locus X3,1. Section 4 proves that X3,1 is embedded in X in a bilipschitz fashion. Hence this replacement can be performed in a bilipschitz-controlled way. Step 2. Once the path γ([a, b]) is in the singular locus, we ex- ploit a local trivialisation to push it in a constant direction as follows. 6 K. K., M. K., D. K., AND Y. L. Choose δ > 0 sufficiently small to be specified later. Then γ(a − δ) and γ(b + δ) are in the smooth part X3,2 ⊆ X. The path can easily be shortened to a smooth one, still denoted γ. Over a sufficiently small neightborhood of the smooth path, we can choose a smooth (and in particular bilipschitz) trivialisation of X over the path γ([a, b]). Let I0,1(t) = I0,1 = (cid:18)1 0 0 0(cid:19) (3.3) be a constant section of the bundle over the path. Relative to the trivialisation of the bundle we can form a new path ¯γǫ(t) = γ(t) + ǫI0,1(t) which pushes γ([a, b]) in the constant direction (3.3) for each t. By construction, the new path is contained in the nonsingular part X3,2 of the determinantal variety, where ǫ > 0 is chosen small enough so that the length of the path stays close to the original length of the path in X3,1. Step 3. It remains to check that the path can be patched up with the value of the path γ at the parameter values a − δ and b + δ. Since the determinantal variety in the fiber is a cone over a connected space, the two values can be connected by an arbitrarily short path, provided ǫ and δ are chosen small enough. Step 4. Once the path lies in the nonsingular part X3,2 ⊆ X of the determinantal variety, it can be pushed out into GL+(3, R) by following the normal direction as in the case n = 2. (cid:3) The higher dimensional case is treated inductively in Section 5. 4. The general determinantal variety Let Xn,k ⊆ Rn2 be the stratum of rank k matrices. Thus Xn is the closure of Xn,n−1 and each Xn,k is a smooth connected manifold of dimension n2 −(n−k)2. For each x ∈ Xn,k choose an open set Ux ⊆ Rn2 x ⊆ R(n−k)2 containing x, a metric ball U T x ×U N containing 0, and a bilipschitz diffeomorphism φ = φx : Ux → U T x , such that , an open set U N x ⊆ Rn2 −(n−k)2 φ(Ux ∩ Xn,k) = U T x × {0}, φ(Ux ∩ Xn) = U T x × (U N x ∩ Xn−k). (4.1) See [5, p. 138] for the construction of the diffeomorphism φx. To ensure the bilipschitz property, it suffices to shrink slightly Ux, U N x and U T x . Assume by induction that we have proved the bilipschitz equivalence of the intrinsic and extrinsic metrics on Xℓ for ℓ < n. Generalizing Lemma 3.1 to dimension n, we see X n,1 is a linear cone over a compact DETERMINANTAL VARIETY AND NORMAL EMBEDDING 7 smooth submanifold. So, the intrinsic and extrinsic metrics on X n,1 are bilipschitz equivalent. For each x ∈ Xn,1, the bilipschitz diffeomor- phism φx and the induction hypothesis for Xn−1 show the intrinsic and extrinsic metrics are bilipschitz equivalent on Ux ∩ Xn. Then, a com- pactness argument shows the bilipschitz property holds for a sufficiently is defined to be a cone over an ǫ-neighborhood in the unit sphere. is a cone on a compact smooth manifold with boundary. For each small open neighborhood Uǫ(cid:0) X n,1(cid:1) of X n,1 in Xn. More precisely, Uǫ Next we consider the stratum Xn,2. The complement Xn,2\ Uǫ(cid:0) X n,1(cid:1) point x ∈ Xn,2 \ Uǫ(cid:0) X n,1(cid:1), the bilipschitz diffeomorphism φx and the induction hypothesis for Xn−2 show the intrinsic and extrinsic met- rics are bilipschitz equivalent on Ux ∩ Xn. Arguing by compactness as before, we obtain the bilipschitz property for a sufficiently small neighborhood Uǫ(cid:0) X n,2(cid:1) of X n,2 in Xn. Here ǫ may have to be chosen smaller than the one chosen for Xn,1. We proceed in this way until we obtain the bilipschitz property for a neighborhood of Xn,n−2 in the determinantal variety Xn. By compact- ness, the bilipschitz property holds for Xn itself. 5. Bilipschitz property for the set of matrices of positive determinant We prove Theorem 1.1 by pushing a path in the determinantal va- riety out into the component GL+(n, R), and mimicking the proof of Theorem 3.5. If the path γ passes via the apex O ∈ X, it can be replaced by a pair of straight line segments and pushed out into C as in Section 2. Otherwise let k ≥ 1 be the least rank of a matrix γ(t) along the path, and let ak ≤ bk ∈ [0, 1] be respectively the first and last occurrences of a matrix of rank k. As in Section 3, we push the path γ([ak, bk]) into Xn,k ⊆ X, by applying the results of Section 4. We then push it out into Xn,k+1 by following a constant direction, and patch it up at the endpoints as in Step 3 of the proof of Theorem 3.5. The new path is in Xn,k+1. We now choose the corresponding param- eter values ak+1, bk+1 ∈ [0, 1] and proceed inductively. Thus the path can be pushed out into Xn,n−1. Finally we follow the normal direction to push the path out into GL+(n, R). Acknowledgments M. Katz was partially funded by the Israel Science Foundation grant no. 1517/12. D. Kerner was partially supported by the Israel Science Foundation grant 844/14. This paper answers a question posed by Asaf 8 K. K., M. K., D. K., AND Y. L. Shachar at MO1 and we thank him for posing the question. We are grateful to Yves Cornulier for a helpful comment posted there. We are grateful to Jake Solomon for providing the proof in Section 4 of the general case of the bilipschitz property for the determinantal variety. We thank Jason Starr for a helpful comment posted at MO.2 We thank Amitai Yuval for pointing out a gap in an earlier version of the article, and Alik Nabutovsky and Kobi Peterzil for useful suggestions. References [1] Birbrair, L.; Fernandes, A.; Le, D.; Sampaio, J. "Lipschitz regular complex algebraic sets are smooth." Proc. Amer. Math. Soc. 144 (2016), no. 3, 983– 987. [2] Birbrair, L.; Mostowski, T. "Normal embeddings of semialgebraic sets." Michigan Math. J. 47 (2000), no. 1, 125–132. [3] Kanovei, V.; Katz, K.; Katz, M.; Nowik, T. "Small oscillations of the pendu- lum, Euler's method, and adequality." Quantum Studies: Mathematics and Foundations 3 (2016), no. 3, 231–236. See http://dx.doi.org/10.1007/s40509-016-0074-x and http://arxiv.org/abs/1604.06663 [4] Katz, K.; Katz, M. "Bi-Lipschitz approximation by finite-dimensional imbed- dings." Geom. Dedicata 150 (2011), 131–136. [5] MacPherson, R.; Procesi, C. "Making conical compactifications wonderful." Selecta Math. (N.S.) 4 (1998), no. 1, 125–139. [6] Nowik, T., Katz, M. "Differential geometry via infinitesimal displacements." Journal of Logic and Analysis 7:5 (2015), 1–44. See http://dx.doi.org/10.4115/jla.2015.7.5 and http://arxiv.org/abs/1405.0984 [7] Pedersen, H.; Ruas, M. "Lipschitz Normal Embeddings and Determinantal Singularities." See http://arxiv.org/abs/1607.07746 K. Katz, Department of Mathematics, Bar Ilan University, Ramat Gan 52900 Israel E-mail address: [email protected] M. Katz, Department of Mathematics, Bar Ilan University, Ramat Gan 52900 Israel E-mail address: [email protected] D. Kerner, Department of Mathematics, Ben Gurion University of the Negev, Be'er Sheva 8410501 Israel E-mail address: [email protected] Y. Liokumovich, Imperial College, London E-mail address: [email protected] 1See http://mathoverflow.net/questions/222162 2See http://mathoverflow.net/questions/230668
1804.06366	3	1804	2018-09-07T05:04:09	Higher Obstructions of Complex Supermanifolds	[ "math.AG" ]	In this article we introduce the notion of a 'good model' in order to study the higher obstructions of complex supermanifolds. We identify necessary and sufficient conditions for such models to exist. Illustrations over Riemann surfaces are provided.	math.AG	math	Symmetry, Integrability and Geometry: Methods and Applications SIGMA 14 (2018), 094, 12 pages Higher Obstructions of Complex Supermanifolds Kowshik BETTADAPURA Yau Mathematical Sciences Center, Tsinghua University, Haidian, Beijing, 100084, China E-mail: [email protected] Received April 29, 2018, in final form August 30, 2018; Published online September 07, 2018 https://doi.org/10.3842/SIGMA.2018.094 Abstract. In this article we introduce the notion of a 'good model' in order to study the higher obstructions of complex supermanifolds. We identify necessary and sufficient conditions for such models to exist. Illustrations over Riemann surfaces are provided. Key words: complex supergeometry; supermanifolds; obstruction theory 2010 Mathematics Subject Classification: 32C11; 58A50 1 Introduction Complex supermanifolds are spaces modelled on the data of a complex manifold X and holomor- phic vector bundle E → X. Accordingly, we refer to the pair (X, E) as a 'model'. Supermanifolds can be either split or non-split. A splitting is a global isomorphism to some fixed model space. Berezin in [2, pp. 163 -- 169] observed that to any complex supermanifold, there will be associated a hierarchy of inductively defined cohomology classes representing obstructions to the existence of a splitting.1 These classes are called obstruction classes to splitting, or simply 'obstructions'. To any supermanifold one can always define its primary obstruction. If this vanishes, another will appear in its place, sitting at a higher degree (in a suitable sense). Accordingly, these classes are referred to as 'higher obstructions'. Donagi and Witten in [7] observed that the higher obstructions to splitting a supermanifold X might fail to represent a genuine obstruction to splitting X. That is, it might well be that X is split even if it supports an atlas in which a higher obstruction to the existence of a splitting does not vanish. Such a phenomenon seems to be difficult to illustrate in practice. For instance, it was shown by the author in [3] that certain deformations of super Riemann surfaces will not support such atlases. In this article we are motivated by the following question: When does a higher obstruction to splitting X represent a genuine obstruction? (1.1) Following the terminology of Donagi and Witten, if a higher obstruction is not genuine it is referred to as exotic. We summarise the main ideas and results in this article below. 1.1 Article summary Obstruction classes are associated to supermanifolds, while obstruction spaces are associated to models. In our attempt to address the question in (1.1), we will look to classify models (X, E). In Definition 4.1 we introduce the notion of a 'good model' (X, E). This is a model on which every higher obstruction for any supermanifold modelled on (X, E) will be genuine. Lemma 4.2 1In the terminology of Berezin, a splitting is referred to as a 'retraction' and a split supermanifold is referred to as 'simple'. 2 K. Bettadapura clarifies the relation to the splitting problem. Our main result is Theorem 4.3 where we obtain necessary and sufficient conditions for (X, E) to be a good model. Section 5 is devoted to illustrations of models which are good and otherwise. We restrict our attention to holomorphic vector bundles on Riemann surfaces. Over the projective line P1C, when the bundle has rank 3, a general characterisation of good models is derived in Theorem 5.3. It is based essentially on an example by Donagi and Witten in [7]. This is then extended to higher rank for a particular class of bundles on P1C in Theorem 5.5. The relation of higher obstructions to the splitting problem is subtle. In Appendix A we submit a proof of the assertion in Theorem 2.12, being: if the primary obstruction of a super- manifold X does not vanish, then X is non-split. There is no reason for the analogous statement for higher obstructions to hold. It is for this reason that we introduce variants of splitting such as 'strong splitting' and 'weak non-splitting' in Definition 3.6. 2 Complex supermanifolds There are a number of different categories of supermanifolds and each comes with its own level of subtlety as detailed in [1]. Those considered in this article fall under the umbrella of the 'algebro- geometric' supermanifolds. These are defined as locally ringed spaces which are modelled on a given space. Standard reference include [5, 10]. One can however bypass that language and give an equivalent definition in terms of Cech cohomology sets. This is the approach we consider here since it is more convenient for the purposes of this article. 2.1 Green's automorphism groups Let X be a complex manifold and E → X a holomorphic vector bundle. Throughout this article we will refer to the pair (X, E) as a model. Denote by E the sheaf of local sections of E. We can then form the sheaf of exterior algebras ∧•E on X. Remark 2.1. Note that ∧•E is a sheaf of Z2-graded, supercommutative algebras. It also admits a compatible Z-grading. We will however only consider automorphisms of ∧•E which preserve the Z2-grading. The sheaf of such automorphisms will be denoted AutZ2(∧•E). However we will suppress the subscript 'Z2' so as to avoid cumbersome notation and simply write Aut∧•E. Let J(X,E) ⊂ ∧•E be the ideal generated by local sections of E. As an OX -module, J(X,E) = ⊕j>0 ∧j E. The following sheaf of non-abelian groups were defined by Green in [8]: (X,E) :=(cid:8)α ∈ Aut∧•E α(u) − u ∈ J k G(k) (X,E) (cid:9). (2.1) The Cech cohomology of X valued in G(k) working definition of a supermanifold. Definition 2.2. A supermanifold modelled on (X, E) is an element of the Cech cohomology set (X,E) is of particular importance. We present now our H 1(cid:0)X,G(2) (cid:1). (X,E) Generally H 1(X, G), for G a sheaf of non-abelian groups on X, will be a pointed set. The basepoint corresponds to the identity section of G. That is, under the inclusion {e} → G, for {e} the trivial group, one obtains a map H 1(X,{e}) → H 1(X, G) of pointed sets. Now H 1(X,{e}) is a one-point set. The basepoint in H 1(X, G) is the image of this one-point set. Definition 2.3. The basepoint in H 1(cid:0)X,G(2) The 1-cohomology set H 1(cid:0)X,G(2) (cid:1) is referred to as the split supermanifold model- (cid:1) will be referred to as the set of supermanifolds modelled led on (X, E), or simply the split model. (X,E) (X,E) on (X, E). Higher Obstructions for Complex Supermanifolds 3 2.2 Isomorphisms of supermanifolds We will make use of an early observation of Grothendieck in [9], discussed also in [4, p. 160], regarding sheaves of non-abelian groups. Theorem 2.4. Let X be a (paracompact) topological space and suppose the following is a short exact sequence of sheaves of (not necessarily abelian) groups on X e −→ A −→ B −→ C −→ e. Then there exists a long exact sequence e −→ H 0(X, A) −→ H 0(X, B) −→ H 0(X, C) −→ H 1(X, A) −→ H 1(X, B) −→ H 1(X, C), (2.2) (2.3) where the sequence in (2.2) is as groups whereas that in (2.3) is as pointed sets. Furthermore, there exists an action of H 0(X, C) on the set H 1(X, A) such that the following diagram commutes H 1(X, A) / H 1(X, B) H 1(X, A)/H 0(X, C). We have so far defined supermanifolds as elements of a certain Cech cohomology set. As for when two supermanifolds are isomorphic, this is defined as follows. Firstly observe for k = 2 the following short exact sequence of sheaves of groups {e} −→ G(2) (X,E) −→ Aut∧•E −→ AutE −→ {e}. (2.4) Applying Theorem 2.4 gives a long exact sequence on cohomology. Consider the following piece ··· −→ H 0(cid:0)X,AutE(cid:1) −→ H 1(cid:0)X,G(2) (cid:1) −→ H 1(cid:0)X,Aut∧•E(cid:1) −→ ··· . (2.5) (X,E) Isomorphisms are defined as follows. Definition 2.5. Two supermanifolds modelled on (X, E) are isomorphic if and only if their image in H 1(cid:0)X,Aut∧•E(cid:1) coincide. latter part of Theorem 2.4 we have an action of the group Aut(E) on H 1(cid:0)X,G(2) (cid:1) and a well- defined map from the orbits to H 1(cid:0)X,Aut∧•E(cid:1). This leads to the following result, first described Note, H 0(X,AutE) coincides with the group Aut(E) of global automorphisms. From the (X,E) by Green in [8], on the general classification of supermanifolds. As per the definitions we have made it follows from Theorem 2.4. Theorem 2.6. There exists a bijective correspondence {supermanifolds modelled on (X, E)} up to isomorphism ∼= H 1(cid:0)X,G(2) (X,E) Aut(E) (cid:1) . ( ( / 6 6 4 2.3 Obstruction theory Recall that in Definition 2.3 we termed the basepoint in H 1(cid:0)X,G(2) (X,E) the notion of isomorphism in Definition 2.5 we have: K. Bettadapura (cid:1) the split model. With Definition 2.7. A supermanifold if said to be split if and only if it is isomorphic to the split model. Obstruction theory for supermanifolds is concerned with the following question which is referred to as the splitting problem: Question 2.8. Given a supermanifold, how can one tell if it is split? The splitting problem has a long history dating back to its original formulation by Berezin in [2], framed using the term 'retraction'. Examples of non-split supermanifolds were provided by Berezin and contemporaries including Palamodov in [13], Green in [8] and Manin in [10]. Outside of mathematical curiosity however, the relevance of the splitting problem to theoretical physics is not so clear. Recent advances by Donagi and Witten in [6, 7] serve to show that non-splitting is a phenomenon that, at least in superstring theory, cannot be ignored. As our starting point in addressing the splitting problem, we have the following result by Green in [8]. Proposition 2.9. For each k ≥ 2 there exists a short exact sequence of sheaves of groups {e} −→ G(k+1) (X,E) −→ G(k) (X,E) −→ Q(k) (X,E) −→ {e}, where the inclusion G(k+1) (X,E) → G(k) (X,E) is normal and the factor Q(k) (X,E) is abelian. An important utility of Proposition 2.9 is in relating non-abelian cohomology to the coho- mology of abelian sheaves. By Theorem 2.4 the short exact sequence in Proposition 2.9 will induce, for each k, a long exact sequence on cohomology. The piece of most relevance for our present purposes is ··· −→ H 1(cid:0)X,G(k+1) (cid:1) −→ H 1(cid:0)X,G(k) (X,E) (X,E) (cid:1) ω∗−→ H 1(cid:0)X,Q(k) (cid:1), (X,E) (2.6) where in the latter we have used the identification of Cech cohomology and sheaf cohomology for abelian sheaves. Definition 2.10. To any model (X, E) we term the following important constructs for each k ≥ 2: • the sheaf Q(k) • the cohomology group H 1(cid:0)X,Q(k) (cid:1) is referred to as the k-th obstruction space. (X,E) in Proposition 2.9 is referred to as the k-th obstruction sheaf ; (X,E) Inspecting (2.6) shows: there will exist a class in the second obstruction space associated to any supermanifold X modelled on (X, E). We denote this class by ω∗(X). Definition 2.11. Let X be a supermanifold modelled on (X, E). The class ω∗(X) in the second obstruction space will be referred to as the primary obstruction to splitting X. The terminology in the above definition is justified in the following classical result. Theorem 2.12. Let X be a supermanifold. If its primary obstruction is non-vanishing, then X is non-split. Higher Obstructions for Complex Supermanifolds 5 Proof . A proof of this theorem, stated using the notion of 'retractibility,' can be found in [2, Theorem 4.6.2, p. 158]. For completeness, we present a proof based on our formulation of isomorphisms of supermanifolds in Definition 2.5. As the proof is a little involved, we defer it (cid:4) to Appendix A. Note that Theorem 2.12 above cleanly addresses the splitting problem, as posed in Ques- tion 2.8. If the primary obstruction to splitting vanishes however, it far more subtle to adequately address this problem. 3 Higher obstructions 3.1 Higher atlases manifold modelled on (X, E). Definition 3.2. If a supermanifold X lies in the image of a (k − 1)-split atlas X(k), then X is (X,E) We are interested in the splitting problem in instances where the primary obstruction vanishes. To that extent we begin with the following string of definitions. Definition 3.1. An element in H 1(cid:0)X,G(k) said to be (k − 1)-split. Conversely, the image of X(k) in H 1(cid:0)X,G(2) (cid:1). Following Definition 2.11 we have: the cohomology group H 1(cid:0)X,Q(k) (cid:1) is referred to as a (k − 1)-split atlas for a super- (cid:1) will be referred to as Associated to any (k − 1)-split atlas will be a class in the k-th obstruction space, which is the supermanifold associated to X(k). (X,E) (X,E) Definition 3.3. The class in the k-th obstruction space associated to any (k − 1)-split atlas will be referred to as the primary obstruction of the (k − 1)-split atlas. Definition 3.4. If the primary obstruction of a (k−1)-split atlas X(k) is non-vanishing, then X(k) is said to be obstructed. The primary obstructions associated to (k− 1)-split atlases are precisely the 'higher' obstruc- tion classes associated to supermanifolds, as described by Berezin (see [2, p. 164]) and more recently by Donagi and Witten in [7, p. 15]. In these texts the higher obstructions were defined inductively as follows. Starting with a (k−1)-split atlas X(k) for a supermanifold X, one wants to lift this to a k-split atlas for X. The obstruction to doing so is precisely the primary obstruction of X(k). The primary obstruction of X(k) is the (k − 1)-th obstruction to splitting X. Hence, the k-th obstruction to splitting X is defined if and only if the (k − 1)-th obstruction to splitting X vanishes. 3.2 The weak splitting problem To see the relation of higher obstructions to the splitting problem, note that the groups G(k) defined in (2.1), will be trivial for k sufficiently large. holomorphic vector bundle E → X, then J k k > q. This leads to the following result, which can also be found (albeit phrased differently) in [2, 13]: Theorem 3.5. A supermanifold X is split if and only if it admits a (k− 1)-split atlas for k > q. (X,E), Indeed, if q denotes the rank of the (X,E) = {e} for all (X,E) = (0) for all k > q. Hence G(k) We present now a weaker notion of splitting for higher atlases suited to the purposes of this article. 6 K. Bettadapura Definition 3.6. A (k − 1)-split atlas X(k), k > 2, is said to be weakly non-split if its associated supermanifold does not coincide with the basepoint. Otherwise, it is said to be strongly split. We say X(k) is split or non-split if its associated supermanifold is split or non-split. Note that non-split will imply weakly non-split but not necessarily conversely. Similarly, strongly split will imply split, but not necessarily conversely. These notions lead to the following variant of the splitting problem. Question 3.7. When will a given (k − 1)-split atlas be weakly non-split? We refer to our proof of Theorem 2.12 in Appendix A for an illustration of the subtleties involved in resolving the splitting problem as outlined in Question 2.8. 3.3 Exotic atlases In [7] it was observed that the map H 1(cid:0)X,G(k) (cid:1) → H 1(cid:0)X,G(2) (cid:1) will generally fail to be (X,E) either injective or surjective. This means the existence of an obstructed, (k − 1)-split atlas for a supermanifold X does not imply X will be non-split -- in contrast with the k = 2 case in Theorem 2.12. The following definition captures precisely those obstructed atlases which represent split supermanifolds. Definition 3.8. A (k − 1)-split atlas X(k), for k > 2, is said to be exotic if it entertains the following two properties: (X,E) (i) X(k) is obstructed; (ii) X(k) is strongly split. With the sequence in Proposition 2.9 we can describe an exotic atlas more concretely. In- specting the cohomology sequence in (2.6) one step to the left and for the sequence of groups G(k−1) (X,E) → G(k) (X,E) gives (X,E) → Q(k−1) ··· −→ H 0(cid:0)X,Q(k−1) β−→ H 1(cid:0)X,G(k−1) (X,E) (X,E) (cid:1) α−→ H 1(cid:0)X,G(k) (cid:1) −→ H 1(cid:0)X,Q(k−1) (X,E) (cid:1) (cid:1). (X,E) (3.1) (3.2) (X,E) We prove: Lemma 3.9. Let X(k) be an obstructed, (k − 1)-split atlas for k > 2. If there exists a non-zero φ ∈ H 0(cid:0)X,Q(k−1) φ ∈ H 0(cid:0)X,Q(k−1) H 1(cid:0)X,G(k) map to the basepoint in H 1(cid:0)X,G(k−1) (cid:1) such that α(φ) = X(k), then X(k) will be exotic. (cid:1). (cid:1). Now by exactness of the sequence in (3.1) -- (3.2) note, under β, that X(k) will (cid:1). (cid:1). Hence it will map to the basepoint in H 1(cid:0)X,G(2) for some non-zero In assuming X(k) is obstructed, it cannot represent the basepoint in Proof . Let X(k) be a (k − 1)-split atlas and suppose α(φ) = X(k), (X,E) (X,E) (X,E) Thus X(k) is obstructed and strongly split. Hence X(k) is exotic. (X,E) (cid:4) Further assumptions on the global sections of the obstruction sheaves yield the following sufficient conditions forbidding the existence of exotic atlases. Proposition 3.10. Suppose that H 0(cid:0)X,Q(k) (cid:1) = (0) for all k > 2. Then there do not exist (X,E) exotic atlases for any supermanifold modelled on (X, E). (X,E) (X,E) (X,E) sets (3.3) (cid:0)ιk+1∗ Higher Obstructions for Complex Supermanifolds Proof . Suppose H 0(cid:0)X,Q(k) {e} −→ H 1(cid:0)X,G(k+1) (cid:1) = (0). Then from (3.1) we have the exact sequence of pointed (cid:1) ιk+1∗−→ H 1(cid:0)X,G(k) (cid:1). following implication from exactness of (3.3): Unlike for rings or modules, an exact sequence as in (3.3) above need not imply ιk+1∗ (cid:1) we can nevertheless conclude the (cid:1) = (0) for all k. Then the implication (3.4) will hold for all k. (cid:1). It will then (cid:1) and suppose it maps to the basepoint in H 1(cid:0)X,G(2) (cid:0)X(k+1)(cid:1) = {e}(cid:1) =⇒(cid:0)X(k+1) = {e}(cid:1). injective as a map of sets. But for X(k+1) ∈ H 1(cid:0)X,G(k+1) Assume now that H 0(cid:0)X,Q(k) Let X(k+1) ∈ H 1(cid:0)X,G(k+1) H 0(cid:0)X,Q(q) we already have {e} → H 1(cid:0)X,G(q) (cid:1) → H 1(cid:0)X,G(q−1) under the slightly weakened assumption that H 0(cid:0)X,Q(k) follow from (3.4) that X(k+1) = {e}. Hence X(k+1) cannot be obstructed and so, by definition, (cid:4) cannot be exotic. Remark 3.11. For E → X a holomorphic, rank q vector bundle it is superfluous to assume (X,E) = {e} for all k > q, meaning (cid:1) = (0) in Proposition 3.10. This is because G(k) (cid:1). That is, Proposition 3.10 will be valid (cid:1) = (0) for all 2 < k < q. will be (3.4) (X,E) (X,E) (X,E) (X,E) (X,E) (X,E) (X,E) 7 (X,E) 4 Good models The existence of an exotic atlas depends essentially on the model (X, E). This leads to the following definition. Definition 4.1. A model (X, E) is said to be good if there do not exist any exotic atlases for any supermanifold modelled on (X, E). Otherwise, it is said to support exotic atlases. The prime motivation for introducing good models lies in the following result which follows from the definitions. Lemma 4.2. Let (X, E) be a good model. Then if a (k − 1)-split atlas modelled on (X, E) is obstructed, it will be weakly non-split. It is a meaningful endeavour to understand when a given model (X, E) will be good. In Proposition 3.10 we found sufficient conditions in this vein. We arrive now at our main result in this article, concerning both necessary and sufficient conditions for a model to be good. Theorem 4.3. A model (X, E) will be a good if and only if taking global sections preserves exactness of the sequence {e} −→ G(k+1) (X,E) −→ G(k) (X,E) −→ Q(k) (X,E) −→ {e} for all k ≥ 2. Proof . From the exact sequence of sheaves of groups G(k+1) following piece of the long exact sequence on cohomology {e} −→ H 0(cid:0)X,G(k+1) α−→ H 1(cid:0)X,G(k+1) (X,E) (cid:1) −→ H 0(cid:0)X,G(k) (cid:1) −→ ··· . (X,E) (X,E) (X,E) → G(k) (X,E) → Q(k) (cid:1) (cid:1) −→ H 0(cid:0)X,Q(k) (X,E) (X,E) we have the (4.1) H 0(cid:0)X,Q(k) (X,E) (cid:1) (cid:101)α α (X,E) 8 K. Bettadapura In the direction (⇐) in the statement of this theorem, assume (4.1) is exact for all k ≥ 2. (cid:1) = (0) for all k ≥ 2 or the map α is constant, sending every φ ∈ (cid:1) to the basepoint {e}. In either case, the proof of Proposition 3.10 will apply (cid:1), the atlas α(φ) must be an Then either H 0(cid:0)X,Q(k) H 0(cid:0)X,Q(k−1) not exist any exotic atlases. Therefore, for any φ ∈ H 0(cid:0)X,Q(k) a result we have the following lifting (cid:101)α of α, represented by the dashed arrow since we will obtain exact sequences in (3.3) for all k. In the direction (⇒), suppose (X, E) is a good model. We will show that α is constant. Firstly, by definition we know that there will unobstructed, k-split atlas, for otherwise it would be exotic by Lemma 3.9. Observe that as (X,E) (X,E) (X,E) (X,E) (cid:1) (cid:1) (cid:1) (cid:1). / H 1(cid:0)X,G(k) H 1(cid:0)X,G(k+2) / H 1(cid:0)X,G(k+1) H 1(cid:0)X,Q(k+1) (cid:1). As such it will map to the basepoint in H 1(cid:0)X,G(2) (X,E) ω∗ (X,E) The lifting (cid:101)α exists since the vertical maps in (4.2) are exact and unobstructedness of α(φ) means ω∗(cid:0)α(φ)(cid:1) = 0. Hence from φ we obtain a (k + 1)-split atlas (cid:101)α(φ). We claim that (cid:101)α(φ) is also unobstructed. To see this note that (4.2) will commute and so (cid:101)α(φ) will map to the basepoint in H 1(cid:0)X,G(k) (cid:1) and so be split. This means, if (cid:101)α(φ) were obstructed, it would be exotic, contradicting that (X, E) is a good model. Hence (cid:101)α(φ) must be unobstructed. Applying this argument again to (cid:101)α(φ) will (cid:1) and that to φ will be associated show that (cid:101)α will itself lift to some (cid:101)α(cid:48) valued in H 1(cid:0)X,G(k+3) (X,E) (X,E) (X,E) an unobstructed, (k + 2)-split atlas. We can proceed inductively now and keep lifting the map α to get unobstructed, k(cid:48)-split atlases for any k(cid:48) > k, resulting in the commutative diagram H 1(cid:0)X,G(k(cid:48)) / H 1(cid:0)X,G(k+1) (X,E) (cid:1) (cid:1). (X,E) H 0(cid:0)X,Q(k) (X,E) (cid:1) (cid:101)αk(cid:48) α / (4.2) (4.3) (5.1) Since G(k(cid:48)) (X,E) is trivial for k(cid:48) sufficiently large, commutativity of (4.3) requires α map φ to the basepoint {e} for all φ. Thus α is constant. This argument is independent of k ≥ 2 and depends (cid:4) only on (X, E) being a good model. The theorem now follows. Remark 4.4. Concerning the diagram (4.2), it is argued in [2, Theorem 4.7.1, pp. 163 -- 164] and in [7, p. 16] that there will always exist a lift (cid:101)α of α when k is odd. 5 Illustrations To convince the reader that exotic atlases and good models exist we present some illustrations. It will be convenient to firstly present the following more explicit characterisation of the obstruction sheaves which we will make use of in our applications. A justification can be found in [8]. Lemma 5.1. To any model (X, E) there exists an isomorphism of sheaves (cid:40)∧kE ⊗ TX , k is even, E∗ ⊗ ∧kE, k is odd. Q(k) (X,E) = ' ' 6 6 / / 7 7 Higher Obstructions for Complex Supermanifolds 9 5.1 On Riemann surfaces Let X be a complex manifold and E → X a holomorphic vector bundle of rank 3. Then G(k) (X,E) = {e} for all k > 3. With the identifications of the obstruction sheaves in (5.1) we obtain the following long exact sequence, corresponding to the more general case in (4.1) {e} −→ H 0(cid:0)X,G(3) α−→ H 1(cid:0)X,G(3) (cid:1) a non-trivial condition to check. It will be true however if h0(cid:0)X,∧2E ⊗ TX (cid:1) −→ H 0(cid:0)X,G(2) (cid:1) −→ ··· . (cid:1) −→ H 0(cid:0)X,∧2E ⊗ TX (X,E) (X,E) (X,E) (cid:1) = 0.2 (5.2) Theorem 4.3 says (X, E) will be a good model if and only if α in (5.2) vanishes. This is generally Proposition 5.2. Let X be a Riemann surface of genus g and E → X a rank 3, holomor- phic vector bundle. Suppose it splits into a sum of non-negative, holomorphic line bundles and deg E < 3g − 3. Then (X, E) will be a good model. (cid:1) = 0. Let X be a genus g Riemann surface. Assu- ming E is holomorphically split we can write E = Ld1 ⊕ Ld2 ⊕ Ld3, for Ldi → X a line bundle on X of degree di. If Ldi is non-negative, then di ≥ 0. Note that ∧2E = Ld1+d2⊕Ld1+d3⊕Ld2+d3; and deg ∧2E = 2 deg E = 2(d1 + d2 + d3). With deg TX = 2 − 2g and the above description Proof . It suffices to show h0(cid:0)X,∧2E ⊗ TX of ∧2E we see that h0(cid:0) ∧2 E ⊗ TX (cid:1) = 0 when d1 + d2 < 2g − 2, d2 + d3 < 2g − 2. These conditions are equivalent to deg E < 3g − 3 when di ≥ 0 for each i. d1 + d3 < 2g − 2 and (cid:4) 5.2 In genus zero We can be considerably more specific on the projective line. Theorem 5.3. Let E = Ld1⊕Ld2⊕Ld3 be a rank 3, holomorphic vector bundle on P1C, where Ldi is a line bundle on P1C of degree di. Then (P1C, E) will be a good model if and only if at least one of di (cid:54)= −1. (X,E) ∼= Q(q) Before giving a proof of Theorem 5.3 we provide some preliminary remarks about exotic atlases. Regarding their existence, this was addressed in [7, pp. 15 -- 16] for the present situation, i.e., when E has rank 3. We comment on this here, but in a little more generality. Let E → X be a holomorphic vector bundle of rank q. Note that G(k) (X,E) = {e} for all k > q. Hence G(q) (X,E) which implies H 1(cid:0)X,G(q) (cid:1) −→ H 1(cid:0)X,G(q) (cid:1). With the exact sequence (cid:1) −→ ··· (cid:1) in the image of a non-trivial section in H 0(cid:0)X,Q(q−1) we deduce: Lemma 5.4. Let E → X be a rank q, holomorphic vector bundle. Any (q − 1)-split atlas in (cid:1) ∼→ H 1(cid:0)X,Q(q) (cid:1) −→ H 1(cid:0)X,G(q−1) (cid:1) will be obstructed and ··· −→ H 0(cid:0)X,Q(q−1) H 1(cid:0)X,G(q) (X,E) (X,E) (X,E) (X,E) (X,E) (X,E) (X,E) hence, by Lemma 3.9, exotic. We now present a proof of Theorem 5.3. 2For any abelian sheaf A on X we denote hi(X,A) := dim H i(X,A). 10 K. Bettadapura Proof of Theorem 5.3. Write E = O(d1) ⊕ O(d2) ⊕ O(d3), where E denotes the sheaf of sections of E. Then Q(2) (P1C,E) Q(3) (P1C,E) = O(d1 + d2 + 2) ⊕ O(d1 + d3 + 2) ⊕ O(d2 + d3 + 2) = O(d1 + d2) ⊕ O(d1 + d3) ⊕ O(d2 + d3). In order to construct an exotic atlas for a supermanifold modelled on(cid:0)P1C, E(cid:1) it will be necessary (i) h0(cid:0)Q(2) (ii) h1(cid:0)Q(3) (cid:1) (cid:54)= 0 (by Proposition 3.10 and Remark 3.11) and (cid:1) (cid:54)= 0 (since G(3) ∼= Q(3) (P1C,E) here). for and (P1C,E) (P1C,E) (P1C,E) From Bott's formula on the dimension of the cohomology of line bundles on projective space,3 both (i) and (ii) will be satisfied if and only if (d1, d2, d3) = (−1,−1,−1). Now in this case, where (d1, d2, d3) = (−1,−1,−1), we have by Serre duality H 0(cid:0)P1C,Q(2) (cid:1) ∼= H 1(cid:0)P1C,Q(3) (cid:1). Upon observing H 1(cid:0)P1C,Q(3) (P1C,E) ma 5.4 applied to q = 3. (P1C,E) (P1C,E) (cid:1) ∼= H 1(cid:0)P1C,G(3) (P1C,E) (cid:1), the theorem will then follow from Lem- (cid:4) We will now consider those bundles of higher rank which decompose into copies of a single line bundle, i.e., E = ⊕qLd. As we will see, the case d = −1 will be similar to that in Theorem 5.3. Theorem 5.5. Let E → P1C be a rank q, holomorphic vector bundle. Suppose E = ⊕qLd where Ld → P1C is a line bundle of degree d. Then (i) if d = −1 then(cid:0)P1C, E(cid:1) will support exotic atlases; (ii) if d < −1, then(cid:0)P1C, E(cid:1) will be a good model. Proof . Let O(d) denote the sheaf of sections of Ld so that E = ⊕qO(d). We begin by pro- ving (ii). We have generally ∧kE = ⊕(q k)O(kd). As such the obstruction sheaves are given by = Q(k) (P1C,E) k)O(kd + 2), k)O(cid:0)(k − 1)d(cid:1), k is odd. bundles of negative degree. Hence H 0(cid:0)P1C,Q(k) ⊕k(q k is even, (P1C,E) Note that 2 ≤ k ≤ q here. Evidently, if d < −1 the obstruction sheaves will be sums of line (cid:1) = (0). Part (ii) then follows from Proposi- (cid:40)⊕(q (5.3) (5.4) tion 3.10. As for part (i), firstly note that ∧•E = ∧0E ⊕ ∧1E ⊕ ∧2E ⊕ ··· ⊕ ∧qE = O(0) ⊕(cid:2) ⊕q O(d)(cid:3) ⊕(cid:2) ⊕(q 2) O(2d)(cid:3) ⊕ ··· ⊕ O(qd). We have AutE ⊂⊕q2O(0) and Aut∧•E ⊂⊕0≤a≤b≤qO((b−a)d). Note in particular that b − a ≥ 0. Regarding the sheaf G(k) observe that it can be realised as a subsheaf of ⊕0≤a<b≤qO((b− a)d) where now b − a ≥ k − 1. Therefore if d < 0: (P1C,E) H 0(cid:0)P1C,G(k) (P1C,E) (cid:1) = {1}, 3See, e.g., [11, p. 4]. Higher Obstructions for Complex Supermanifolds for all k ≥ 2. Now if d = −1 then h0(cid:0)P1C,Q(2) (cid:1) α−→ H 1(cid:0)P1C,G(3) {1} −→ H 0(cid:0)P1C,Q(2) {e} from (5.4) we have the exact sequence (cid:1) =(cid:0)q (cid:1), (P1C,E) 2 (P1C,E) (P1C,E) (cid:1) (cid:54)= 0 from (5.3). Since H 0(cid:0)P1C,G(2) 11 (cid:1) = (P1C,E) where α is the boundary map from (5.2). In particular α does not vanish and any atlas in its (cid:4) image will be exotic. 5.3 In genus one Here we can relax some of conditions in Proposition 5.2, although we cannot be as general as in the genus zero case. Proposition 5.6. Let C be a Riemann surface of genus one and E → C a holomorphic, rank 3 vector bundle. Then (C, E) will be a good model when either: (i) deg Q(2) (ii) deg Q(2) Proof . In genus one note that TC = OC is a line bundle of degree zero. Hence we have a natural isomorphism Q(2) (C,E). The proposition now follows from the Riemann -- Roch (cid:4) theorem. (C,E) = h0(cid:0)Q(2) (C,E) = −h1(cid:0)Q(2) (cid:1) or (cid:1). ∼= Q(3) (C,E) (C,E) (C,E) A Proof of Theorem 2.12 We will firstly present an erroneous proof as it will be instructive in illustrating the subtlety involved in the splitting problem. False proof . Let X be a supermanifold modelled on (X, E) and suppose its primary obstruc- tion ω∗(X) does not vanish. The map ω∗ : H 1(cid:0)X,G(2) sets, sending the basepoint {e} ∈ H 1(cid:0)X,G(2) global symmetries Aut(E) acting on H 1(cid:0)X,G(2) ω∗(X) (cid:54)= 0, X cannot be split. (X,E) ··· / Aut(E) δ / / H 1(cid:0)X,G(2) H 1(cid:0)X,Q(2) ω∗ (X,E) (cid:1) (cid:1). (X,E) The fault in the above reasoning lies in our failure to consider isomorphisms induced by the If Aut(E) fixes the basepoint {e} then the above argument would be valid, but there is of course no reason for it to fix the basepoint in general. Consider instead the following (X,E) (X,E) (X,E) (cid:1) → H 1(cid:0)X,Q(2) (cid:1) is a map of pointed (cid:1) to the basepoint 0 ∈ H 1(cid:0)X,Q(2) (cid:1). Since (cid:1) (cf. Definition 2.5 and Theorem 2.6). / H 1(cid:0)X,Aut∧•E(cid:1) / ··· (X,E) (cid:4) (A.1) Since (A.1) is exact we have: if a supermanifold X is split, then X = δ(φ) for some φ ∈ Aut(E). Observe that Theorem 2.12 will then follow from: Proposition A.1. The composition of maps ω∗δ in (A.1) vanishes. The proof we submit of Proposition A.1 is based on the discussion in [7, p. 16]. We will cite the following result whose proof can be found in [12]. / / / (X,E) . j (X,E) (cid:1). (X,E) Proof of Proposition A.1. From Theorem 2.4 we know that Aut(E) will act on 12 K. Bettadapura Lemma A.2. Let (X, E) be a model. The subgroup C× · 1E < Aut(E) acts on the k-th obstruc- (cid:1) by sending v (cid:55)→ λkv for any vector v ∈ H 1(cid:0)X,Q(k) trivially on the image of Aut(E). To see this, let (Ui → X) be an open cover and φ ∈ Aut(E). (cid:1). Denote this action by (cid:63). Our first claim is that the subgroup C× · 1E will act tion space H 1(cid:0)X,Q(k) H 1(cid:0)X,G(2) If (cid:101)φi ∈(cid:0)Aut∧•E(cid:1)(Ui) are local lifts of φi, then δ(φ)ij = (cid:101)φi(cid:101)φ−1 (cid:1) (cid:63) δ(φ)ij =(cid:0)λ(cid:101)φi (cid:0)λ · 1E (cid:1) = (cid:101)φi(cid:101)φ−1 The group Aut(E) will act on H 1(cid:0)X,G(2) (cid:1). The action of C× · 1E mentioned in Lemma A.2 is compatible with this in- H 1(cid:0)X,Q(2) duced Aut(E) action. In particular we can conclude: for a supermanifold X ∈ H 1(cid:0)X,G(2) (cid:1) In comparing (A.2) and (A.3) we see that if X = δ(φ), we must have ω∗(cid:0)X(cid:1) = 0. This proves Now clearly any λ · 1E ∈ C× · 1E < Aut(E) will act by sending φi (cid:55)→ λφi. Hence, and any λ · 1E ∈ C× · 1E, ω∗(cid:0)(λ · 1E) (cid:63) X(cid:1) = λ2ω∗(cid:0)X(cid:1). j = δ(φ)ij. (cid:1) and this action will (cid:1)(cid:0)λ−1(cid:101)φ−1 j induce an action on (X,E) (X,E) (A.2) (X,E) (A.3) (cid:4) Proposition A.1 from whence Theorem 2.12 follows. Acknowledgements The author would like to acknowledge the helpful feedback of the anonymous referees. References [1] Bartocci C., Bruzzo U., Hern´andez Ruip´erez D., The geometry of supermanifolds, Mathematics and its Applications, Vol. 71, Kluwer Academic Publishers Group, Dordrecht, 1991. [2] Berezin F.A., Introduction to superanalysis, Mathematical Physics and Applied Mathematics, Vol. 9, D. Rei- del Publishing Co., Dordrecht, 1987. [3] Bettadapura K., On the problem of splitting deformations of super Riemann surfaces, Lett. Math. Phys., to appear, arXiv:1610.07541. [4] Brylinski J.-L., Loop spaces, characteristic classes and geometric quantization, Modern Birkhauser Classics, Birkhauser Boston, Inc., Boston, MA, 2008. [5] Deligne P., Morgan J.W., Notes on supersymmetry (following Joseph Bernstein), in Quantum Fields and Strings: a Course for Mathematicians, Vols. 1, 2 (Princeton, NJ, 1996/1997), Amer. Math. Soc., Providence, RI, 1999, 41 -- 97. [6] Donagi R., Witten E., Super Atiyah classes and obstructions to splitting of supermoduli space, Pure Appl. Math. Q. 9 (2013), 739 -- 788, arXiv:1404.6257. [7] Donagi R., Witten E., Supermoduli space is not projected, in String-Math 2012, Proc. Sympos. Pure Math., Vol. 90, Amer. Math. Soc., Providence, RI, 2015, 19 -- 71, arXiv:1304.7798. [8] Green P., On holomorphic graded manifolds, Proc. Amer. Math. Soc. 85 (1982), 587 -- 590. [9] Grothendieck A., A general theory of fibre spaces with structure sheaf, University of Kansas, 1995. [10] Manin Yu.I., Gauge field theory and complex geometry, Grundlehren der Mathematischen Wissenschaften, Vol. 289, Springer-Verlag, Berlin, 1988. [11] Okonek C., Schneider M., Spindler H., Vector bundles on complex projective spaces, Modern Birkhauser Classics, Birkhauser/Springer Basel AG, Basel, 2011. [12] Onishchik A.L., On the classification of complex analytic supermanifolds, Lobachevskii J. Math. 4 (1999), 47 -- 70. [13] Palamodov V.P., Invariants of analytic Z2-manifolds, Funct. Anal. Appl. 17 (1983), 68 -- 69.
1711.05854	1	1711	2017-11-15T23:51:18	Logarithmic Kodaira dimension and zeros of holomorphic log-one-forms	[ "math.AG", "math.DG" ]	In this paper, we prove that the zero-locus of any global holomorphic log-one-form on a projective log-smooth pair $\left(X,D\right)$ of log-general type must be non-empty. Applying this result, we give an answer to the algebraic hyperbolicity part of Shafarevich's conjecture, with the generic fiber being Kawamata-log-terminal (klt) and of log-general type.	math.AG	math	LOGARITHMIC KODAIRA DIMENSION AND ZEROS OF HOLOMORPHIC LOG-ONE-FORMS CHUANHAO WEI 1. Introduction In this paper, we prove the conjecture that appeared in [Wei16], which is a natural generalization of the work of M. Popa and C. Schnell, [PS14]. Theorem 1. The zero-locus of any global holomorphic log-one-form on a projective log-smooth pair (X, D) of log-general type must be non-empty. Actually, we can prove a more general result: Theorem 2. Let (X, D) be a projective log-smooth pair, and let W ⊂ H 0(cid:0)X, Ω1 X (log D)(cid:1) be a linear subspace that consists of global holomorphic log-one-forms with empty zero-locus. Then the dimension of W can be at most dim X − κ (X, D), where κ stands for the logarithmic Kodaira dimension. This is a corollary to the following main theorem of this paper. Before stating the theorem, we recall the notations and some propositions about qusi-abelian varieties in the sense of Iitaka, given in [Wei16]. See also [Iit76] and [Fuj14] for more details. Definition 1. T r,d is a quasi-abelian variety (in the sense of Iitaka), if it is an extension of a d-dimensional abelian variety Ad by an algebraic torus Gr m, i.e. it is a connected commutative algebraic group which has the following Chevalley decomposition 1 → Gr In particular, T r,d is a principal Gr m → T r,d → Ad → 1. m-bundle over Ad. Consider the following group homomorphism: ρ : Gr m → P GL(r, C), given by ρ(λ1, ..., λr) =   1 0 λ1 . . . 0 λr   Let P r,d := T r,d ×ρ Pr = T r,d × Pr/Gr m, which is a Pr-bundle over Ad. We can view P r,d as a compactification of T r,d which naturally carries the T r,d action on it and denote the boundary divisor L, which is of simple normal crossing and is T r,d-invariant for each stratum. We say that (X, D) is a log-smooth pair if X is a smooth variety and D is a reduced divisor on X with normal crossing support. Given two log smooth pairs 1 2 CHUANHAO WEI (cid:0)X, DX(cid:1) and (cid:0)Y, DY(cid:1), and a morphism f : X → Y , we say that f is a morphism of log-pairs and denoted by f :(cid:0)X, DX(cid:1) →(cid:0)Y, DY(cid:1), if f −1DY := Supp(cid:0)f ∗DY(cid:1) ⊂ DX . Given a holomorphic log-one-form θ on a log smooth pair (X, D), we use Z (θ) X (log D). to denote the zero-locus of θ as a global section of the locally free sheaf Ω1 Pr-bundle compactification of a quasi-abelian variety T r,d, with the boundary divisor L. Denote by p : P r,d → Ad, the natural projection. Given a morphism of log-pairs Theorem 3. Let (X, D) be a projective log-smooth pair, (cid:0)P r,d, L(cid:1) be the canonical f : (X, D) →(cid:0)P r,d, L(cid:1), if there is a positive integer k and an ample line bundle A on Ad, such that H 0(cid:16)X, (ωX (D))⊗k ⊗ f ∗(cid:0)p∗A−1 ⊗ OP r,d (−L)(cid:1)(cid:17) 6= 0, then Z (f ∗ω) 6= ∅, for any ω ∈ H 0(cid:0)Ω1 Further, for generic such ω, we have Z (f ∗ω) ∩(cid:0)X \ f −1 (L)(cid:1) 6= ∅. P r,d (log L)(cid:1) . Remark 1. Actually, we only need to show the case for r = 0 or 1. We refer to the proof of Theorem 4 below for the details. Remark 2. The two statements in the previous theorem are actually equivalent to each other. In particular, it is a problem that only depends on the morphism over T r,d. See Step 0 of the proof of this theorem in Section 4 for details. There are two important applications of Theorem 2. One of them is that we get an affirmative answer (in a much more general setting) to a question posed by F. Catanese and M. Schneider, Corollary 8 and Corollary 9. Another application is that we give an answer to the algebraic hyperbolicity part of Shafarevich's con- jecture, with the generic fiber being klt and of log-general type, Corollary 11. E. Viehweg and K. Zuo have shown the previous two questions in the case that the base is one dimensional and the general fibers are projective and smooth, [VZ01]. In Section 2, we first prove Theorem 2 assuming Theorem 3. Then we show the geometric applications mentioned above. In Section 3, we give a simplified proof of the main theorem in [PS14] without applying the generic vanishing of mixed Hodge modules on abelian varieties, [PS13]. In Section 4, assuming Claim 2 (Main Claim), we prove Theorem 3. In Section 5, we first recall some results about logarithmic comparison in mixed Hodge modules in [Wei17]. Then we show a vanishing result that will be used to prove the Main Claim. In Section 6, we relate the Main Claim with the theory of mixed Hodge modules and prove it. All the varieties that appear in this paper are assumed to be reduced but possibly reducible separated schemes of finite type over the field of complex number C. We use strict right D-modules to represent mixed Hodge modules, forgetting the weight filtration. All mixed Hodge modules in this paper are assumed to be algebraic. In particular, they are assumed to be extendable and polarizable, [Sai90]. Acknowledgment. The author would like to express his gratitude to his advisor Christopher Hacon for suggesting this topic and useful discussions. The author thanks Honglu Fan, Kalle Karu, Mihnea Popa, Christian Schnell, Lei Wu and Ziwen Zhu for answering his questions, and especially Schnell for suggesting a better choice of Hodge module which is used in the paper that essentially simplifies the proof of the main theorem. LOG-KODAIRA DIMENSION AND ZEROS OF HOLOMORPHIC LOG-ONE-FORMS 3 During the preparation of this paper, the author was partially supported by DMS-1300750, DMS-1265285 and a grant from the Simons Foundation, Award Number 256202. 2. Proof of Theorem 2 and applications We start by showing that Theorem 3 implies the following Theorem 4. Given a morphism of log smooth pairs f : (X, D) → (cid:0)P r,d, L(cid:1), where (cid:0)P r,d, L(cid:1) is the canonical Pr-bundle compactification of a quasi-abelian va- H 0(cid:0)P r,d, Ω1 P r,d (log L)(cid:1) of co-dimension dim X − κ (X, D), such that Z (f ∗ω) 6= ∅, riety T r,d, with the boundary divisor L, then there exists a linear subspace W ⊂ for any ω ∈ W . z \ D′ Proof. Denote U = X \ D. The statement is vacuous when κ (X, D) = −∞, so we can assume that κ (X, D) ≥ 0. Let µ : (X ′, D′) → (X, D) be a birational modification, such that g : X ′ → Z is a smooth model for the Iitaka fibration associated to the log-canonical bundle ωX (D). It is not hard to see that, if (X ′ z, D′ z) is a very general fiber over Z, then it is a log-smooth pair of dimension δ (U ) = dim U −κ (X, D) with log-Kodaira dimension 0. Hence due to [Kaw81, Theorem 28], the quasi-albanese map of X ′ z is an open algebraic fiber space, and by [Kaw81, Theorem 27], it is not hard to see that its image in T r,d U is a quasi-abelian variety. Further, by Proposition 5 below, we know that there are at most countably many quasi-abelian sub-algebraic group in T r,d U of every fiber of g is a dense subset of a translate of a single quasi-abelian sub-algebraic group. Letting T be the quotient of T r,d U by that sub quasi-abelian sub-algebraic group, we have dim T ≥ m − δ (U ), where m = r + d = dim T r,d U . Hence we get an induced rational map Z 99K T . Let (P, L) be the projective-space-bundle compactification of T . Though P r,d 99K P is just a rational map in general, after a toroidal log- resolution (cid:16) P r,d, L(cid:17) →(cid:0)P r,d, L(cid:1), the induced map of log-pairs(cid:16) P r,d, L(cid:17) → (P, L) is a morphism. After making a toroidal log resolution U . Hence, the image in T r,d we can get a morphism of log-pairs(cid:16) X, D(cid:17) →(cid:16) P r,d, L(cid:17) ([AK00] Section 1, Remark 1.4). Composing it with the previous one, we get a morphism of log-pairs τ :(cid:16) X, D(cid:17) → (X, D) , f :(cid:16) X, D(cid:17) → (P, L) . (cid:16) X, D(cid:17) (X ′, D′) Consider the following commutative diagram: Z f (P, L) . Now for any line bundle A on P , we have H 0(cid:18) X,(cid:16)ω X(cid:16) D(cid:17)(cid:17)⊗k ⊗ f ∗ A(cid:19) 6= 0, 4 CHUANHAO WEI by Theorem 3. For the rest of the proof, we show that we actually just need the case that r = 0 for some integer k. Since τ :(cid:16) X, D(cid:17) → (X, D) is a toroidal log-resolution, we have X (log D)(cid:17) has zero-locus if and only if τ ∗θ ∈ H 0(cid:16) X, Ω1 that θ ∈ H 0(cid:16) X, Ω1 X(cid:16)log D(cid:17)(cid:17) has zero-locus. Hence now we only need to show that, for any ω ∈ H 0(cid:0)P, Ω1 P (log L)(cid:1), X(cid:16)log D(cid:17)(cid:17) will have zero-locus. Now we can conclude the proof f ∗ω ∈ H 0(cid:16) X, Ω1 or 1 in Theorem 3. Denote (P, L) by (cid:0)P r,d, L(cid:1) again to specify the dimension of the base quasi-abelian variety and projective space fiber, and T r,d := P r,d \ L. If r ≥ 2, we will argue that the statement can be reduced to the r = 1 case. Consider m of T r,d, which is the fiber over 0, the unit of Ad. Fix {x1, ..., xr}, the subgroup Gr a global algebraic coordinate system on Gr m, and x1 = ... = xr = 1 gives the unit e of Gr m. Given a vector a := [a1, ..., ar] ∈ Zr, assuming a 6= 0, let's consider the divisor Ga of Gr r = 1. It is evident that it is actually a subgroup of Gr m, hence of T r,d. Let's denote the algebraic quotient by m defined by xa1 1 · ... · xar qa : T r,d → T r,d/Ga. a , for some quasi-abelian variety T 1,d It is evident that T r,d/Ga ≃ T 1,d cated dimensions. Denote by (cid:0)P 1,d a , L(cid:1), the canonical P1-bundle compactification a . As in the argument above, replacing (cid:16) X, D(cid:17) and (cid:0)P r,d, L(cid:1) by a toroidal of T 1,d log-resolution, we obtain morphisms of log-pairs: a with indi- a , L(cid:1) qa :(cid:0)P r,d, L(cid:1) →(cid:0)P 1,d ga :(cid:16) X, D(cid:17) →(cid:0)P 1,d a , L(cid:1) Denote P 1,d a a , Ω1 Wa = q∗ aH 0(cid:16)P 1,d straight forward to check that Wa is the co-set (log L)(cid:17) ⊂ H 0(cid:0)P r,d, Ω1 Note that ∪aWa is a dense subset of H 0(cid:0)P r,d, Ω1 {dx1/x1, ..., dxr/xr} gives a basis of H 0(cid:0)Ω1 P r,d(log L)(cid:1) . P r,d(log L)(cid:1) /p∗H 0(cid:0)Ad, Ω1 Ad(cid:1) . Since f ∗ω has no zero-locus is an open condition on ω ∈ H 0(cid:0)P r,d, Ω1 now it suffices to show the statement for each ga : (cid:16) X, D(cid:17) → (cid:0)P 1,d C · (a1dx1/x1 + ... + ardxr/xr) + p∗H 0(cid:0)Ad, Ω1 implied by Theorem 3. P r,d (log L)(cid:1). This is because Ad(cid:1), and it is P r,d(log L)(cid:1), a , L(cid:1), which is (cid:3) Remark 3. For a fixed quasi-abelian variety T r,d, fixing any smooth projective r- dimensional toric variety P r, we have the canonical P r-bundle compactification of T r,d, with a natural simply normal crossing boundary divisor L, and we also denote it by (cid:0)P r,d, L(cid:1). All the (cid:0)P r,d, L(cid:1) that appear in this paper can be viewed in this setting. Proposition 5. There are at most countably many quasi-abelian sub-algebraic groups in a quasi-abelian variety. Proof. It is evident that we only need to show the case for any abelian variety and for any algebraic torus. For the abelian variety case, it is obvious by noting that there are at most countably many sub-lattice in a given lattice of finite dimension. LOG-KODAIRA DIMENSION AND ZEROS OF HOLOMORPHIC LOG-ONE-FORMS 5 For the algebraic torus case, it is not hard to see that we only need to show that there are at most countably many algebraic group auto-morphism G1 m. Fix a global coordinator t, the auto-morphism will be given by the map of function t 7→ P [t, t−1], where P [t, t−1] is a two-variable polynomial. Since it needs to be a group morphism, we have m → G1 P(cid:2)tk, t−k(cid:3) =(cid:0)P(cid:2)t, t−1(cid:3)(cid:1)k , for any k ∈ Z. Hence it is not hard to conclude that P [t, t−1] = tm for some m ∈ Z, which has countably many choices. (cid:3) Recall that given a smooth quasi-projective variety U and a log-smooth com- pactification (X, D) of U , we have that T1 (U ) := H 0(cid:0)X, Ω1 not depend on the compactification, [Fuj14, 2.4]. Further, we canonically have a quasi-albanese map aU : U → TU , such that a∗ U (T1 (TU )) = T1 (U ). TU is the quasi-albanese variety of U which is a quasi-abelian variety and the quasi-albanese map aU is algebraic. We refer to [Fuj14], [Iit76] for the details of the quasi-albanese map. X (log D)(cid:1) , which does To prove Theorem 2 by applying the argument in the proof of the previous theorem, we need to construct such algebraic morphism of log-smooth pairs f : (X, D) → (cid:0)P r,d, L(cid:1). Ideally, we want to directly use the quasi-albanese map aU : U → T r,d, and then compactify it and perform a log-resolution to get f . However, taking log-resolution may introduce new zero-loci for holomorphic log-one-forms. To keep track the zero-loci, we are only allowed to perform toroidal log-resolutions. That is the reason that we consider the following Lemma 6. Fix a log smooth pair (X, D), and denote X \ D = U . Assume that we have an algebraic morphism f : U → T r,d. Then there exists a toroidal log- resolution τ :(cid:16) X, D(cid:17) → (X, D) such that f can be extended to get a morphism of log-pairs: f :(cid:16) X, D(cid:17) →(cid:0)P r,d, L(cid:1). Proof. We first note that although we may not be able to extended f onto X, the morphism p ◦ f : U → Ad is well defined on X. This is because if not, by a sequence of blowing-ups of smooth center, we can get a morphism. However, each rational curve will map to a point on an abelian variety, hence these blow-ups are not needed. Then we argue that it suffices to show the case that r = 1. Actually, it is not hard to see that we only need to show the case that (cid:0)P r,d, L(cid:1) is the (P1)r-bundle compactification of T r,d, Remark 3. Note that in this case, we can decompose P r,d = P 1,d 1 ×Ad P 1,d 2 ×Ad ... ×Ad P 1,d r , i where P 1,d is the P1-bundle compatification of a quasi-abelian variety T 1,d indicated dimension. Hence to get a morphism of log-pairs f :(cid:16) X, D(cid:17) →(cid:0)P r,d, L(cid:1), we only need to get a morphism of log-pairs(cid:16) X, D(cid:17) →(cid:16)P 1,d , L(cid:17) for each i. From , with the now on, we only consider the case that r = 1. i i Denote the two components of L by L1 and L2. It is evident that L separates into two components by the construction of P 1,d. We also have that L1 and L2 are linearly equivalent ([Har77, Exercise II.7.9]). Since f is well-defined on X except a co-dimension 2 locus. Hence the zero order of f ∗L1 and f ∗L2 are well defined along 6 CHUANHAO WEI 1 ∩ D′′ 1 and D′′ 1 , with D′ each irreducible component Di of D. We denote by D′ and D′′ the two sub-divisors of D that consists of those irreducible components with positive zero order of f ∗L1 and f ∗L2 respectively. Since L1 ∩L2 = ∅, D′ and D′′ have no common components. We say that a co-dimension-2 stratum S of D is bad if S is the generic locus of D′ 1 are irreducible components of D′ and D′′ respectively. For the rest of the proof, we will show that (1)If we have no bad stratum, then f is well defined on X. (2)We can find a sequence of toroidal resolutions on (X, D) to eliminate bad strata. For (1), if f is not well defined on X, by a sequence of blowing-ups of smooth locus on D, we can get a morphism of log-pairs f : ( X, D) → (P 1,d, L). Note that since we start with no bad stratum, it is evident that for each step of blow-up, we will not introduce new bad stratum. Let's consider the last step of the blow-ups and we denote it by π : ( X, D) → (X, D), with E ⊂ D being the exceptional divisor. We only need to show that each P1 on E maps to a point on P 1,d, which implies that this blow-up is not needed. Hence we can conclude (1) by induction. Otherwise, since each P1 on E will map to a point on Ad, we can find a P1 on E maps to a P1 fiber over a point a ∈ Ad surjectively. In particular, there is one point e1 ∈ E that maps to L1 and another point e2 ∈ E maps to L2. Consider the two sub-divisors of D: D′ := f −1L1 and D′′ := f −1L2. Obviously they have no common components, and E does not belong to either one. However, we can find two irreducible components D1 and D2 such that e1 ∈ D1 ⊂ D′, and e2 ∈ D2 ⊂ D′′. We denote by D1 and D2 the two irreducible components of D with their strict transform on X being D1 and D2 respectively. Let S be the stratum given by the generic locus of D1 ∩ D2 which is not empty by the construction. However it is not hard to see that S is a bad stratum. Hence we get a contradiction. 1 · ... · xan For (2), since L1 and L2 are linearly equivalent, we can find a rational function y on P 1,d such that its zero-locus and pole locus are L1 and L2 respectively. Since f is well defined on U and its image is in T 1,d, the zero-locus and the pole locus of the rational function f ∗y on X has to be contained in D. Hence, around a stratum S, the generic locus of ∩1≤i≤nDi, up to a multiplication of a unit, we locally have f ∗y = xa1 n , with xi are local functions that defines Di, the irreducible components of D, and ai ∈ Z. Note that ai does not depend on the choice of the stratum. Pick a bad stratum S of co-dimension-2. Assume S is the generic locus of D1 ∩ D2. S being bad is equivalent to that a1 · a2 < 0. We can find two co-prime integers b1, b2, such that a1 · b2 + a2 · b1 = 0. By a blow-up of the ideal locally being 2 (cid:17), which is toroidal, we get π : (X ′, D′ + E) → (X, D), where D′ is the strict transform of D and E is the exceptional divisor. Now by induction, we only need to show that those newly introduced co-dimension-2 strata of D′ + E are not bad. It is evident by noticing that the zero order of (π ◦ f )∗y along E is 0. (cid:16)xb1 1 , xb2 Note that (X ′, D′ + E) we construct above is not log-smooth in general, but after we eliminate all bad strata, we can perform a toroidal log-resolution to get a log-smooth pair we are after. (cid:3) Now, we are ready to show the proof of Theorem 2 assuming Theorem 3. Proof of Theorem 2. Following the argument of the proof of Theorem 4, we are reduced to consider the rational map ga : X 99K P 1,d a , which is well defined on LOG-KODAIRA DIMENSION AND ZEROS OF HOLOMORPHIC LOG-ONE-FORMS 7 U , ga(cid:12)(cid:12)U : U → T 1,d resolution (cid:16) X, D(cid:17) → (X, D) such that we honestly get a morphism of log pairs: a . Now apply the previous lemma, we can find a toroidal log- ga :(cid:16) X, D(cid:17) →(cid:0)P 1,d a , L(cid:1) . Hence we can conclude the proof by the r = 0 or 1 case of Theorem 3 as in the proof of Theorem 4. (cid:3) For the rest of the section, we state couple corollaries that follow by Theorem 2. Corollary 7. Given a smooth projective variety X, a global holomorphic one-form θ on X with no zero-locus, and a smooth divisor D. If ωX (D) is big, then θD must have non-empty zero-locus as a global holomorphic one-form on D. Proof. Let x1, ..., xn be a local holomorphic coordinate system of X and x1 defines D. Then {dx1/x1, dx2, ..., dxn} gives a local basis of Ω1 X (log D). Since we know that it has no zero-locus as holomorphic one-form, according to Theorem 1, θ has a zero-locus at some point p ∈ D as a global section of Ω1 X (log D). Hence it locally looks like θ = x1g1dx1/x1 + g2dx2... + gndxn, where g2 (p) = ... = gn (p) = 0. Hence θD ∈ H 0 (D, ΩD) has zero-locus at p. (cid:3) Definition 2. Given a projective morphism of log pairs f : (cid:0)X, DX(cid:1) → (cid:0)Y, DY(cid:1) and assume that (cid:0)X, DX red(cid:1) are log-smooth. We say that f is log- red(cid:1) and (cid:0)Y, DY smooth if the cokernel sheaf of the natural map f ∗Ω1 Y (cid:0)log DY(cid:1) → Ω1 X(cid:0)log DX(cid:1) , is locally-free. Corollary 8. If f : (X, D) → (cid:0)P r,d, L(cid:1) is a log-smooth morphism, from a log- smooth pair (X, D) onto the canonical Pr-bundle compactification of a quasi-abelian variety T r,d, of dimension m = r + d, then m ≤ dim X − κ (X, D). Proof. It follows immediately by Theorem 2. See also Remark 3. (cid:3) Corollary 9. Let (X, D) be a projective log-smooth pair of log general type. Assume that we have a surjective projective morphism f : X \ D → T r,d, where T r,d is a quasi-abelian variety. Then, f is not smooth. Proof. We can find a log-resolution (cid:16) X, D(cid:17) → (X, D) such that it induces a mor- phism of log-pairs f :(cid:16) X, D(cid:17) →(cid:0)P r,d, L(cid:1) , with f −1L = D. Now by Theorem 4, we have Z(cid:16) f ∗ω(cid:17) 6= ∅, for any ω ∈ H 0(cid:0)Ω1 P r,d (log L)(cid:1). However, by Lemma 10, (see also Step 0 of the proof of Theo- rem 3 in Section 4,) for general such ω, Z(cid:16) f ∗ω(cid:17) ∩ ( X \ D) 6= ∅. We refer [Wei16] for more details, and a different proof based on the structure of Higgs bundles. (cid:3) We use the following lemma in the proof above, [Wei16, Lemma 8]. Lemma 10. Given a morphism f : X → P r,d, where X is smooth and quasi- projective and P r,d is the Pr-bundle compactification of a quasi-abelian variety T r,d, with L being the boundary divisor on P r,d. Let D = (f ∗L)red and as- sume that (X, D) is log-smooth. Then for general θ ∈ H 0(cid:0)Ω1 P r,d (log L)(cid:1), f ∗θ ∈ 8 CHUANHAO WEI H 0(cid:0)Ω1 X (log D)(cid:1) has a simple log pole on every point of D. In particular, it does not vanish at any point of D. Another application of Theorem 2 is that we can give an answer to the algebraic hyperbolicity part of Shafarevich's conjecture, with general fibers being of log- general type, which is the special case of the following theorem by taking T r,d = G1 m or Ad. We refer to [HK10, §16] for details of this topic. Note that a projective morphism of log-smooth pairs f : (U, DU ) → G1 m being log-smooth is equivalent to the existence of (X, DX ), a projective log-smooth compactification of (U, DU ) such that g : (X, DX ) → (P1, L), the induced morphism of log-smooth pairs by f , is log-smooth. It can be very complicated in the higher dimensional cases. A similar question has been asked and we refer to [AKMWo02] for the details on this topic. Definition 3. Given a flat dominant projective morphism of log-smooth pairs f : (X, DX ) → (Y, DY ), with connected fibers and the generic fiber (Xη, DX η ) being Kawamata-log-terminal (klt) and of log-general type. We say that it is birationally isotrivial if the two log-fibers (Xa, DX b ) have the same log-canonical model ([BCHM10]), for any two general closed points a and b on Y . a ) and (Xb, DX Corollary 11. Given a log-smooth surjective morphism of log-pairs f : (X, D) → (cid:0)P r,d, L(cid:1), with (X, Dred) being log-smooth and the generic fiber (Xη, Dη) being klt and of log-general type, then f is birationally isotrivial. Proof. We only need to show Var(f ) = 0, in the sense of [KP16, Definition 9.3]. By Corollary 8, we have By Theorem 12, we have κ (X, D) ≤ dim Xη. κ (X, D) ≥ κ (Xη, Dη) + Var(f ). Since dim Xη = κ (Xη, Dη) by the assumption of the theorem, we get that Var(f ) = 0. (cid:3) Remark 4. If we only assume that the generic fiber (Xη, Dη) is log-canonical and of log-general type, since (Xη, Dη) being of log-general type is an open condition on the coefficients of the boundary divisor, we can consider f ′ : (X, (1 − ǫ) D) →(cid:0)P r,d, L(cid:1) , for 0 < ǫ ≪ 1. Then the generic fiber of f ′ is klt and of log-general type. Following the same arguement above, we can show that f ′ is birationally isotrivial. In the proof of the previous theorem, we used the following theorem, which can be easily deduced from [KP16, Theorem 9.5, Theorem 9.6]. Theorem 12. Let f : (X, D) → (Y, E) be a surjective morphism of projective log canonical pairs with both (X, Dred) and (Y, E) are log-smooth. Assume that ⌊D⌋ contains f −1Ered and the generic fiber (Xη, Dη) is of log-general type. Then κ (X, D) ≥ κ (Xη, Dη) + κ (Y, E) . Further, if (Xη, Dη) is klt, then κ (X, D) ≥ κ (Xη, Dη) + max{κ (Y, E) , Var(f )}. LOG-KODAIRA DIMENSION AND ZEROS OF HOLOMORPHIC LOG-ONE-FORMS 9 3. A Simplified proof of the non-log version of Theorem 3 In this section, we show a simplified proof of the the following theorem, which is the main result of [PS14], without using the generic vanishing of Hodge modules on abelian varieties, which they introduced in [PS13]. The method is still mainly based on [PS13]. Theorem 13. Given a morphism f : X → A, if there is a positive integer k and an ample line bundle A on A, such that X ⊗ f ∗A−1(cid:17) 6= 0, H 0(cid:16)X, ω⊗k then Z (f ∗ω) 6= ∅, for any ω ∈ H 0(cid:0)A, Ω1 A(cid:1) . : A → A which is defined by Proof. Step 1. We have an ´etale morphism [k] multiplying by k. Apply the finite base change and let f ′ : X ′ → A be the fiber product. Since [k] is ´etale, we have that X ′ is smooth and it suffices to prove the theorem for f ′ : X ′ → A. φ [k] X ′ f ′ A X f A We have an inclusion A⊗k → [k]∗A [Mum08, II, 6]. Hence =φ∗ω⊗k X ⊗ f ∗(cid:0)A−1(cid:1)(cid:17) φ∗(cid:16)ω⊗k ⊂(cid:0)ωX ′ ⊗ f ′∗A−1(cid:1)⊗k X ⊗ f ′∗[k]∗A−1 which, by assumption, has a global section. Hence we reduce to the case that there exists a positive integer k such that H 0(cid:16)X,(cid:0)ωX ⊗ f ∗A−1(cid:1)⊗k(cid:17) 6= 0. Step 2. Let B = ωX ⊗ f ∗A−1, a line bundle over X. Do the cyclic cover induced by a section s of B⊗k and resolve it. We get ψ : Y → X. Hence we have the following commutative diagram Y ψ g X f A. By the construction, there is a tautological section of ψ∗B on Y . Hence we have a natural injection, ψ∗B−1 → OY , which induces the following injection (1) ψ∗(cid:0)B−1 ⊗ Ωk X(cid:1) → Ωk Y . Actually, both of them are isomorpisms over the complement of the zero-locus of s. 10 CHUANHAO WEI Step 3. Let d = dim A. We have that Ω1 over A. Denote V = H 0(cid:0)A, Ω1 Hence we have A(cid:1) , and denote the space of cotangent of A by T ∗ A is a trivial d-dimensional vector bundle A. T ∗ A = A × V. Consider the following commutative diagram, which contains all the morphisms we will need in the proof: Y ψ g X f A pY pX pA Y × V ψ X × V g f A × V pV V, where all the morphisms are natural projections or naturally induced by f and ψ. Let n = dim X = dim Y . Denote CY = [p∗ Y OY → p∗ Y Ω1 Y → ... → p∗ Y Ωn Y ], placed in cohomological degrees −n, −n + 1, ..., 0, which is the Koszul complex given by the pullback of the tautological section of p∗ AΩ1 A. Note that pY is an affine morphism, and pY ∗ (CY ) is the following graded complex CY,• = [OY ⊗ S•−d → Ω1 Y ⊗ S•−d+1 → ... → Ωn Y ⊗ S•−d+n], where S• := SymV ∗. The differential in the complex is induced by the evaluation morphism V ⊗ OY → Ω1 Y . We define CX , CX,• in a similar way. Denote L = g∗A and L = p∗ Y (L) . Claim 1. is torsion free. pV ∗R0g∗(cid:16) L−1 ⊗ CY(cid:17) Step 4. We continue the proof assuming the claim. We set F as the image of the map The restriction of p∗ X B−1 ⊗ CX on the fiber of pV ◦ f : X × V → V over θ is just B−1 → B−1 ⊗ Ω1 X → ... → B−1 ⊗ Ωn X , the Koszul complex defined by f ∗ (θ) and twisted by B−1. Since the zero-locus of f ∗ (θ) is empty if and only if the above complex is exact, to prove the theorem, it suffices to prove pV (suppF ) = V . Note that pA and pY are affine, so we have that pA∗ and pY ∗ are exact, and pA∗ ◦ R0g∗ = R0g∗ ◦ pY ∗. Hence, pA∗F is a graded pA∗OA×V -module F• given by the image of pA∗R0g∗(cid:16) L−1 ⊗ ψ∗(cid:0)p∗ X B−1 ⊗ CX(cid:1)(cid:17) → pA∗R0g∗(cid:16) L−1 ⊗ CY(cid:17) , R0g∗(cid:16) L−1 ⊗ ψ∗(cid:0)p∗ induced by the injection (1). Pick any global one-form X B−1 ⊗ CX(cid:1)(cid:17) → R0g∗(cid:16) L−1 ⊗ CY(cid:17) θ ∈ H 0(cid:0)A, Ω1 A(cid:1) = V. LOG-KODAIRA DIMENSION AND ZEROS OF HOLOMORPHIC LOG-ONE-FORMS 11 which is the same as the image of R0g∗(cid:0)L−1 ⊗ ψ∗(cid:0)B−1 ⊗ CX,•(cid:1)(cid:1) → R0g∗(cid:0)L−1 ⊗ CY,•(cid:1) . If pV (SuppF ) 6= V , the subsheaf pV ∗F ⊂ pV ∗R0g∗(cid:16) L−1 ⊗ CY(cid:17) would then be torsion and hence zero. Therefore, since V is a vector space, H 0 (A, pA∗F ) = H 0 (A, F•) = 0. Recall that B = ωX ⊗ f ∗A−1 and L = g∗A. We have This forces H 0 (g∗OY ) = 0, which is absurd. F−n+r = g∗(cid:0)L−1 ⊗ ψ∗(cid:0)B−1 ⊗ ωX(cid:1)(cid:1) = g∗OY Denote E = RpV ∗R0g∗(cid:16) L ⊗ CY(cid:17) . We first show that for l > 0, Step 5. Finally, we show the proof of Claim 1. (2) HlE = 0. In particular, E is a sheaf. Since both pA and pY are affine, we have H l(cid:16)A × V, R0g∗(cid:16) L ⊗ CY(cid:17)(cid:17) ≃H l(cid:16)A, pA∗R0g∗(cid:16) L ⊗ CY(cid:17)(cid:17) ≃H l(cid:16)A, R0g∗pY ∗(cid:16) L ⊗ CY(cid:17)(cid:17) =H l(cid:0)A, R0g∗ (CY,•) ⊗ A(cid:1) =0 The last vanishing is due to Laumon's formula ([PS14, Lemma 15.1], or Proposition 23) and [PS14, Lemma 2.5] or Proposition 24. More precisely, we have and for l > 0. R0g∗ (CY,•) ≃ GrF • R0g+eωY , • R0g+eωY ⊗ A(cid:17) = 0, H l(cid:16)GrF Hence, due to the degeneration of the Leray spectral sequence induced by pV ∗, we have H 0(cid:0)V, HlE(cid:1) =H 0(cid:16)V, RlpV ∗R0g∗(cid:16) L ⊗ CY(cid:17)(cid:17) ≃H l(cid:16)P × V, R0g∗(cid:16) L ⊗ CY(cid:17)(cid:17) =0, and so that (2) follows. pV ∗R0g∗(cid:16) L−1 ⊗ CY(cid:17) being torsion free is due to the following relation: pV ∗R0g∗(cid:16) L−1 ⊗ CY(cid:17) = (−1V )∗R0Hom (E, OV ) , (3) 12 CHUANHAO WEI where (−1V ) is the involution on V by multiplying −1. To show (3), by Grothendieck duality, we have that RHom (E, OV ) =DV (cid:16)RpV ∗R0g∗(cid:16) L ⊗ CY(cid:17)(cid:17) [−d] ≃RpV ∗DA×V (cid:16)R0g∗(cid:16) L ⊗ CY(cid:17)(cid:17) [−d]. =RpV ∗RHom(cid:16)R0g∗(cid:16) L ⊗ CY(cid:17) , OA×V [d](cid:17) . By [PS13, Proposition 2.11], (or Proposition 16 by taking DY = 0 and r = 0,) we have R0g∗ (CY ) ≃ G(cid:0)H0g+eωX(cid:1) , the corresponding coherent sheaf on T ∗ A = A × V of the mixed Hodge module Theorem 2.3], (or Proposition 19 by taking H = 0,) and considering that the dual H0g+eωX on A. By the formula of taking duality in mixed Hodge modules [PS13, Hodge module of R0g+eωX is itself up to a Tate twist, we obtain RHom (E, OV ) = (−1V )∗ RpV ∗R0g∗(cid:16) L−1 ⊗ CY(cid:17) , which implies (3). (cid:3) 4. Proof of Theorem 3 The proof is inspired by M. Popa and C. Schnell's work [PS14]. The idea of the proof is similar to the proof of Theorem 13. Proof. Step 0. Let's consider the set S = {ω ∈ H 0(cid:0)Ω1 It is closed as a subset of H 0(cid:0)Ω1 P r,d (log L)(cid:1) Z (f ∗ω) 6= ∅}. P r,d (log L)(cid:1). Considering Lemma 10, it is easy to see that the two statements in the theorem are equivalent to each other. Hence, we can focus on the part over T r,d. In other words, we only need to prove the theorem for any log-smooth model (X ′, D′) over (X, D), that is identical over T r,d. In the rest of the proof, we use P and T to replace P r,d and T r,d respectively to simplify the notations. Actually, we can also restrict ourselves to the case that r = 0 or 1, which suffices to show the statement in general. See the second half of the proof of Theorem 4 in Section 2 for details. Step 1. We have a finite morphism [n] : P → P which is defined in the following way. In the interior part T , it is defined by multiplying by n (if we use addition for the group structure on T ). It can be canonically extended onto the boundary L [Wei16]. Note that it is only ramified over the boundary L of degree n. Take n = k, and apply the finite base change, and let (Xn, Dn) be the nomalization of the fiber product. fn : Xn → P is the induced morphism. Let (X ′, D′) be a log-resolution of (Xn, Dn), which can be achieved by only blowing up loci contained in f −1 n (L). According to Step 0, it suffices to prove the theorem for f ′ : X ′ → P . LOG-KODAIRA DIMENSION AND ZEROS OF HOLOMORPHIC LOG-ONE-FORMS 13 φ X ′ π Xn f ′ X ×P P fn [k] P X f P Note that we have φ∗ωX (D) ⊂ ωX ′ (D′), and we have an inclusion p∗A⊗k → [k]∗p∗A [Mum08, II, 6]. Hence which, by assumption, has a global section. Denote E′ = D′ − f ′−1L. Note that OX ′ (D′ − E′) ⊂ f ′∗OP (L), we have that Hence we reduce to the case that there exists a positive integer k such that = (φ∗ωX (D))⊗k ⊗ f ′∗[k]∗OP (−L) ⊗ f ′∗[k]∗p∗A−1 ⊗ f ′∗OP (−kL) ⊗ f ′∗p∗A−k φ∗(cid:16)(ωX (D))⊗k ⊗ f ∗(cid:0)p∗A−1 ⊗ OP r,d (−L)(cid:1)(cid:17) ⊂(cid:0)ωX ′(cid:0)D′(cid:1)(cid:1)⊗k =(cid:0)ωX ′(cid:0)D′(cid:1) ⊗ f ′∗OP (−L) ⊗ f ′∗p∗A−1(cid:1)⊗k H 0(cid:16)X ′,(cid:0)ωX ′(cid:0)E′(cid:1) ⊗ f ′∗p∗A−1(cid:1)⊗k(cid:17) 6= 0. H 0(cid:16)X,(cid:0)ωX (E) ⊗ f ∗p∗A−1(cid:1)⊗k(cid:17) 6= 0, , where E = D − f −1L. Step 2. Let B = ωX (E) ⊗ f ∗p∗A−1, a line bundle over X. Do the cyclic cover induced by a section s of B⊗k and resolve it. We get ψ : Y → X. Denote DY = ψ−1D which can also be assumed to have normal crossings after further blowing up if necessary. We denote by g :(cid:0)Y, DY(cid:1) → (P, L) the induced morphism. Hence we have the following commutative diagram (cid:0)Y, DY(cid:1) ψ g (X, D) f (P, L) . By the construction, there is a tautological section of ψ∗B on Y . Hence we have a natural injection, ψ∗B−1 → OY , which induces the following injection (4) ψ∗(cid:0)B−1 ⊗ Ωk X (log D)(cid:1) → Ωk Y (cid:0)log DY(cid:1) . Actually, both of them are isomorpisms over the complement of the zero-locus of s. P (log L) is a trivial m-dimensional Step 3. Let m = r+d = dim P. We have that Ω1 vector bundle over P . Denote V = H 0(cid:0)P, Ω1 cotangent of (P, L) by T ∗ (P,L). Hence we have P (log L)(cid:1) , and denote the space of log- Consider the following commutative diagram, which contains all the morphisms we will need in the proof: T ∗ (P,L) = P × V. 14 CHUANHAO WEI Y ψ g X f P p Ad pY pX pP pA Y × V ψ X × V g f P × V pV V p Ad × V where all the morphisms are natural projections or naturally induced by f and ψ. Let n = dim X = dim Y . Denote Y Ω1 CY,DY = [p∗ Y OY → p∗ Y (cid:0)log DY(cid:1) → ... → p∗ Y Ωn Y (cid:0)log DY(cid:1)], placed in cohomological degrees −n, −n+1, ..., 0, which is the Koszul complex given by the tautological section of p∗ P (log L). Note that pY is an affine morphism, P Ω1 and pY ∗(cid:0)CY,DY(cid:1) is the following graded complex CY,DY ,• = [OY ⊗ S•−m → Ω1 where S• := SymV ∗. The differential in the complex is induced by the evaluation morphism V ⊗ OY → Ω1 Y . We define CX,D, CX,D,• in a similar way. Denote Y (cid:0)log DY(cid:1) ⊗ S•−m+n], Y (cid:0)log DY(cid:1) ⊗ S•−m+1 → ... → Ωn L = g∗p∗A ⊗ OY (cid:0)g−1L(cid:1) , and Claim 2 (Main Claim). If r = 0 or 1, L = p∗ Y (L) . pV ∗R0g∗(cid:16) L−1 ⊗ CY,DY(cid:17) is torsion free. Step 4. We continue the proof assuming the claim. We set F as the image of the map R0g∗(cid:16) L−1 ⊗ ψ∗(cid:0)p∗ X B−1 ⊗ CX,D(cid:1)(cid:17) → R0g∗(cid:16) L−1 ⊗ CY,DY(cid:17) induced by the injection (4). Pick any global log-one-form θ ∈ H 0(cid:0)P, Ω1 P (log L)(cid:1) = V. The restriction of p∗ X B−1 ⊗ CX,D on the fiber of pV ◦ f : X × V → V over θ is just B−1 → B−1 ⊗ Ω1 X (log D) → ... → B−1 ⊗ Ωn X (log D) , the Koszul complex defined by f ∗ (θ) and twisted by B−1. Since the zero-locus of f ∗ (θ) is empty if and only if the above complex is exact, to prove the theorem, it suffices to prove pV (SuppF ) = V . Note that pP and pY are affine, so we have that pP ∗ and pY ∗ are exact, and pP ∗ ◦ R0g∗ = R0g∗ ◦ pY ∗. Hence, pP ∗F is a graded pP ∗OP ×V -module F• given by the image of pP ∗R0g∗(cid:16) L−1 ⊗ ψ∗(cid:0)p∗ X B−1 ⊗ CX,D(cid:1)(cid:17) → pP ∗R0g∗(cid:16) L−1 ⊗ CY,DY(cid:17) , LOG-KODAIRA DIMENSION AND ZEROS OF HOLOMORPHIC LOG-ONE-FORMS 15 which is the same as the image of R0g∗(cid:0)L−1 ⊗ ψ∗(cid:0)B−1 ⊗ CX,D,•(cid:1)(cid:1) → R0g∗(cid:0)L−1 ⊗ CY,DY ,•(cid:1) . If pV (SuppF ) 6= V , the subsheaf pV ∗F ⊂ pV ∗R0g∗(cid:16) L−1 ⊗ CY,DY(cid:17) would then be torsion and hence zero. Therefore, since V is a vector space, H 0 (P, pP ∗F ) = H 0 (P, F•) = 0. Recall that B = ωX (E) ⊗ f ∗p∗A−1 and L = g∗p∗A ⊗ OY (cid:0)g−1L(cid:1) . We have F−n+r =g∗(cid:0)L−1 ⊗ ψ∗(cid:0)B−1 ⊗ ωX (D)(cid:1)(cid:1) =g∗(cid:0)OY (cid:0)ψ∗ (D − E) − g−1L(cid:1)(cid:1) =g∗(cid:0)OY (cid:0)ψ∗f −1L − g−1L(cid:1)(cid:1) H 0(cid:0)P, g∗(cid:0)OY (cid:0)ψ∗f −1L − g−1L(cid:1)(cid:1)(cid:1) = 0, This forces which is absurd, since ψ∗f −1L − g−1L is effective over Y . (cid:3) 5. Log comparison of mixed Hodge modules and related vanishings In this section, we first recall some results about the logarithmic comparison proved in [Wei17]. Then we deduce some vanishing theorems that will be used in proving the Main Claim modules, forgetting the weight filtration. Given a mixed Hodge module M on a smooth variety X, and given a normal Notations as in [Wei17], we use strict right eD-modules to represent mixed Hodge crossing boundary divisor D = D1 + ... + Dm on X, we introduce two strict eD(X,D)- modules: M∗D = VD M!D = VD 0 M[∗D] := ∩iV Di <0M[!D] := ∩iV Di 0 M[∗D], <0 M[!D], where M[∗D] (resp. M[!D]) is the localization (resp. dual localization) of the mixed Hodge module M along D [SS16, 9], and V Di is the V -filtration respect to Di. Note that V Di . In particular, we have <0 only depends on M(cid:12)(cid:12)X\Di <0M[!D] = VD VD <0M = VD <0M[∗D]. Proposition 14 ([Wei17]). Notations as above, assume that M is of normal cross- ing type respect to D or D is smooth. Then we have and M∗D ⊗L M!D ⊗L eD(X,D) eDX ≃ M[∗D], eD(X,D) eDX ≃ M[!D]. In particular, we have the following quasi-isomorphisms in the derived category of graded eC-modules: Sp(X,D)M∗D ≃SpXM[∗D], Sp(X,D)M!D ≃SpXM[!D]. 16 CHUANHAO WEI Let f : (cid:0)X, DX(cid:1) →(cid:0)Y, DY(cid:1) be a projective morphism between two log-smooth pairs. Assume that DY is smooth, which means DY has only one component or all components of DY do not intersect each other. Proposition 15 ([Wei17]). Notations as above and assume DX = f −1DY . Given a mixed Hodge module M of normal crossing type respect to a normal crossing divisor D′ that contains DX on X, we have that the direct image functor f# is strict on both M∗DX and M!DX , and Hif# (M∗DX ) =(cid:0)Hif+M(cid:1)∗DY , Hif# (M!DX ) =(cid:0)Hif+M(cid:1)!DY . Proposition 16. Given f : (cid:0)X, DX(cid:1) → (cid:0)Y, DY(cid:1) as above, denote EX = DX − f −1DY . We have both f#eωX(cid:0)EX(cid:1) and f#eωX(cid:0)DX − EX(cid:1) are strict, where eω(cid:0)EX(cid:1) andeωX(cid:0)DX − EX(cid:1) are strict eD(X,DX )-modules that induced by the trivial filtration. Further, we have Hif#(cid:0)eωX(cid:0)EX(cid:1)(cid:1) ≃(cid:0)Hif+eωX [∗DX ](cid:1)!DY , Hif#(cid:0)eωX(cid:0)DX − EX(cid:1)(cid:1) ≃(cid:0)Hif+eωX [!DX ](cid:1)∗DY . Proof. We just show the first identity here, the second one follows similarly. Decompose f : (cid:0)X, DX(cid:1) → (cid:0)Y, DY(cid:1) into two morphisms of log pairs: (cid:0)X, DX(cid:1) →(cid:0)X, f −1DY(cid:1), and f ′ :(cid:0)X, f −1DY(cid:1) →(cid:0)Y, DY(cid:1). Now, by the definition of push forward functor, id : Note that id#eωX(cid:0)EX(cid:1) ≃ eωX(cid:0)EX(cid:1) ⊗L eD eωX(cid:0)EX(cid:1) = VEX 0 Vf −1DY <0 (X,DX ) eD(X,f −1DY ). eωX [∗DX ]. Apply the comparison theorem with normal crossing boundary divisor in [Wei17] inductively, we have VEX 0 Vf −1DY <0 eωX[∗DX ] ⊗L eD (X,DX ) eD(X,f −1DY ), ≃Vf −1DX <0 Hence we are left to compute f ′ previous proposition. eωX [∗DX ] =eωX[∗DX ]!f −1DY . #eωX[∗DX ]!f −1DY , but this follows directly from the Proposition 17 ([Wei17]). Given f : (cid:0)X, DX(cid:1) → (cid:0)Y, DY(cid:1) as above, let M be a strict eD(X,DX )-module. Assume that f#M is strict. Then we have The following proposition can be viewed as a log version of Laumon's formula. (cid:3) Hif e#GrF M := Rif∗(cid:18)GrF M ⊗L A (X,DX ) f ∗A (Y,DY )(cid:19) = GrF Hif#M. For the dual functor, we have the following. LOG-KODAIRA DIMENSION AND ZEROS OF HOLOMORPHIC LOG-ONE-FORMS 17 Proposition 18 ([Wei17]). Given a mixed Hodge module M on a log smooth pair (X, H) with H being smooth, we have that D(X,H)M∗H = M′ !H , D(X,H)M!H = M′ ∗H , where M′ = DXM, the dual mixed Hodge module. The following proposition is a log version of [PS13, Theorem 2.3]. Proposition 19 ([Wei17]). Under the same condition as in the previous proposi- tion, we denote by G(M∗H ) and G(M!H ) the associated graded coherent OT ∗ - module of M∗H and M!H respectively, supported on T ∗ (X,H), the space of log- cotangent bundle of (X, H). Then we have (X,H) G(cid:0)M′ G(cid:0)M′ ∗H(cid:1) ≃ (−1)∗ !H(cid:1) ≃ (−1)∗ T ∗ (X,H) T ∗ (X,H) RHomO RHomO (X,H) (cid:16)G (M!H ) , p∗ (X,H) (cid:16)G (M∗H ) , p∗ T ∗ T ∗ X ωX [dX ] ⊗ OT ∗ X ωX [dX ] ⊗ OT ∗ (X,H)(cid:17) , (X,H)(cid:17) . Now we are ready to show some vanishing results that will be used to prove the main claim. The main vanishing result is the following theorem. Note that by taking D = 0, it is Kodaira-Saito vanishing . Theorem 20. Fix a mixed Hodge module M, a possibly un-reduced effective divisor D and a semi-ample line bundle L on a smooth projective variety X. Assume fur- ther that OX (D) ⊗ L is an ample line bundle, then we have the following vanishing: Hi(cid:16)GrF Hi(cid:16)GrF p SpXM[∗D] ⊗ L(cid:17) =0, for i > 0, p SpXM[!D] ⊗ L−1(cid:17) =0, for i < 0. Proof. The proof is similar to the proof of Kodaira-Saito vanishing in [Sai90, 2.33. Proposition]. See also [Pop16]. Since both taking (dual) localization and GrF p SpX are exact, by a standard reduction, we only need to show the case that M is a pure Hodge module with strict support Z ⊂ X. If Z ⊂ D, the vanishings are trivial. Further, by Grothendieck- Serre duality and its compatibility with the dual functor in mixed Hodge modules, it suffices to show the second vanishing. Since Lm is globally generated for some integer m, by Bertini's theorem, we can find a global section s ∈ H 0 (X, Lm) which defines a smooth hypersurface Y that is non-characteristic for both M and M[!D]. We have a finite covering: π : X := SpecX(cid:0)⊕0≤i<mL−i(cid:1) → X, ramified along Y . Denote U = X \ Y and j : V = X \ (D + Y ) → X be the open embedding. Note that V is affine by the ampleness assumption. Because π is non-characteristic by the construction respect to M, the Hodge module π∗π∗M is well defined. Further, we have a natural injection M → π∗π∗M. Denote the Hodge module M = Coker (M → π∗π∗M) . Take the dual localization at D, we have the following short exact sequence: 0 → M[!D] → π∗π∗M[!D] → M[!D] → 0. 18 CHUANHAO WEI By the construction, M is the unique extension of M(cid:12)(cid:12)U onto X with strict support, since the eigenspace with eigenvalue 0 of the monodromy operator on ψY SpXM, the nearby cycle of SpXM respect to Y , is empty. In particular, we have M[!D] = M[!(D + Y )]. Hence by Artin's vanishing, we have c(cid:0)SpV M(cid:12)(cid:12)V(cid:1) = 0, for i < 0. Hi(cid:0)SpXM[!D](cid:1) = Hi Further, by the strictness of a+M[!D], where a : X → pt, Hi(cid:16)GrF a sub-quotient of Hi(cid:0)SpXM[!D](cid:1). Hence Hi(cid:16)GrF p SpXM[!D](cid:17) = 0, for i < 0. (5) On the other hand, we have the short exact sequence p SpXM[!D](cid:17) is 0 → M[!D] → (M[!D]) [∗Y ] → i∗H1i! (M[!D]) → 0, with H1i!M[!D] being a mixed Hodge module by the non-charactericity of i, where i : Y → X is the close embedding. We have i∗H1i! (M [!D]) =(cid:0)i∗H1i!M(cid:1) [!D] , which can be checked at the level of perverse sheaf by [Sai90, 2.11. Proposition]. Note that i∗H1i!M has support on Z ∩ Y . By induction on dimension, now we only need to show Hi(cid:16)GrF p SpX ((M[!D]) [∗Y ]) ⊗ L−1(cid:17) = 0, for i < 0. However, by replacing fM in [Sai90, (2.33.2)] by M[!D], (see also [Pop16, (8.8)],) we have (6) (7) where GrF p (cid:0)M[!D](cid:1) ≃ GrF L := Coker(cid:0)OX → π∗O p ((M[!D]) [∗Y ]) ⊗ L, X(cid:1) ≃ ⊕0<i<mL−i. Now (6) follows by (5) and (7), which concludes the proof. (cid:3) Corollary 21. Fix (cid:0)P r,d, L(cid:1), p : P r,d → Ad as in Theorem 3. Fix a mixed Hodge module M on P r,d and an ample line bundle A over Ad. Assume that r = 0 or 1, or M is of normal crossing type respect to a divisor L′ that contains L. Then we have H i(cid:16)P r,d, GrF H i(cid:16)P r,d, GrF p M∗L ⊗ p∗A(cid:17) = 0, for i > 0, p M!L ⊗ p∗A−1(cid:17) = 0, for i < 0, for any integer k. Proof. Note that if r = 0, it is just [PS13, Lemma 2.5]. We use P and A to replace P r,d and Ad to simplify the notations. We only show the vanishing of the i > 0 case here, the other case follows similarly. Let m = r + d = dim P . We have that GrF p Sp(P,L)M∗L = [GrF p M∗L ⊗ ∧mT(P,L) → GrF p+1M∗L ⊗ ∧m−1T(P,L) → ... → GrF p+mM∗L], LOG-KODAIRA DIMENSION AND ZEROS OF HOLOMORPHIC LOG-ONE-FORMS 19 placed in cohomological degree −m, ..., 0. Note that OP (L) ⊗ p∗A is ample and p∗A is semi-ample. According to the previous vanishing theorem and Proposition 14, for i > 0, we have that (8) Hi(cid:16)GrF p Sp(P,L)M∗L ⊗ p∗A(cid:17) = 0. We have that T(P,L) ≃ O ⊕m smallest integer such that GrF P . Since FpM∗L = 0, for p ≪ 0, take p + m be the p+mM∗L is not trivial. Then by (8), we get that p+mM∗L ⊗ p∗A(cid:17) = 0, for i > 0. By induction, we can conclude the H i(cid:16)P, GrF vanishing of GrF p M∗L for any p. (cid:3) Corollary 22. Assume that we have a morphism of log smooth pairs f : (X, D) → (cid:0)P r,d, L(cid:1), with r = 0 or 1 as in Theorem 3, with p : P r,d → Ad, the natural projection. Then we have H i(cid:16)P r,d, Hkf e#ωX(cid:0)f −1L(cid:1) ⊗ p∗A(cid:17) = 0, for any ample line bundle A over Ad, all k ∈ Z and i > 0. Proof. By Proposition 16 and Proposition 17, we have Now it follows by the previous Corollary. Hkf e#ωX(cid:0)f −1L(cid:1) ≃ GrF (cid:0)Hkf#eωX(cid:0)f −1L(cid:1)(cid:1) ≃ GrF (cid:0)(cid:0)Hkf+ (eωX [!D])(cid:1)∗L(cid:1) . Proposition 23. Let g : (cid:0)Y, DY(cid:1) → (cid:0)P r,d, L(cid:1) be a morphism between two log- smooth pairs, with r = 0 or 1, as in Theorem 3. Denote EY = DY − g−1L. We have 6. Connection to mixed Hodge modules and the proof of the claim (cid:3) and dually g e#ωY (cid:0)EY(cid:1) ≃ Rg∗(cid:0)CY,DY ,• ⊗ OY (cid:0)EY − DY(cid:1)(cid:1) , g e#ωY (cid:0)DY − EY(cid:1) ≃ Rg∗(cid:0)CY,DY ,• ⊗ OY (cid:0)−EY(cid:1)(cid:1) , where CY,DY ,• is defined in the proof of Theorem 3. In particular, we have and GrF (cid:0)(cid:0)Hig+eωY [∗DY ](cid:1)!L(cid:1) ≃ Rig∗(cid:0)CY,DY ,• ⊗ OY (cid:0)EY − DY(cid:1)(cid:1) , GrF (cid:0)(cid:0)Hig+eωY [!DY ](cid:1)∗L(cid:1) ≃ Rig∗(cid:0)CY,DY ,• ⊗ OY (cid:0)−EY(cid:1)(cid:1) . Proof. This follows exactly as [PS13, 2.11] by considering Proposition 16 and Propo- sition 17. More precisely, we have A (Y,DY ) g e#ωY (cid:0)EY(cid:1) ≃Rg∗(cid:18)ωY (cid:0)EY(cid:1) ⊗L g∗A(P r,d,L)(cid:19) ≃Rg∗(cid:18)(cid:2)ωY (cid:0)EY(cid:1) ⊗OY ∧−•T(Y,DY ) ⊗OY ≃Rg∗(cid:18)(cid:2)OY (cid:0)EY − DY(cid:1) ⊗OY Ωn+• ≃Rg∗(cid:0)CY,DY ,• ⊗ OY (cid:0)EY − DY(cid:1)(cid:1) , Y A(Y,DY )(cid:3) ⊗L A(Y,DY ) g∗A (P r,d,L)(cid:19) (cid:0)log DY(cid:1) ⊗OY A (Y,DY )(cid:3) ⊗L A (Y,DY ) g∗A (P r,d,L)(cid:19) 20 CHUANHAO WEI where the second identity is due to the canonical resolution [Wei17, (1)] and the third identity is due to [Har77, Exercise II 5.16 (b)]. The second identity of the proposition can be shown similarly. (cid:3) Proposition 24. Notations as in the previous proposition, let p : P r,d → Ad be the natural projection, and A be an ample line bundle over Ad. Then we have that for any i and l > 0, H l(cid:0)X, Rig∗(cid:0)CY,DY ,• ⊗ OY (cid:0)−EY(cid:1) ⊗ p∗A(cid:1)(cid:1) = 0. Proof. It follows by the previous proposition and Proposition 22. (cid:3) Proof of Claim 2 (Main Claim). Denote E = RpV ∗R0g∗(cid:16) L ⊗ p∗ Y OY (cid:0)−DY(cid:1) ⊗ CY,DY(cid:17) . We first show that for l > 0, (9) HlE = 0. In particular, E is a sheaf. Since both pP and pY are affine and by Proposition 24, we have Y H l(cid:16)P × V, R0g∗(cid:16) L ⊗ p∗ OY (cid:0)−DY(cid:1) ⊗ CY,DY(cid:17)(cid:17) ≃H l(cid:16)P, p1∗R0g∗(cid:16) L ⊗ p∗ OY (cid:0)−DY(cid:1) ⊗ CY,DY(cid:17)(cid:17) ≃H l(cid:16)P, R0g∗pY ∗(cid:16) L ⊗ p∗ OY (cid:0)−DY(cid:1) ⊗ CY,DY(cid:17)(cid:17) ≃H l(cid:0)P, R0g∗(cid:0)CY,DY ,• ⊗ OY (cid:0)−EY(cid:1)(cid:1) ⊗ p∗A(cid:1) Y Y =0, where EY = DY − f −1L. Hence, due to the degeneration of the Leray spectral sequence induced by pV ∗, we have H 0(cid:0)V, HlE(cid:1) =H 0(cid:16)V, RlpV ∗R0g∗(cid:16) L ⊗ p∗ ≃H l(cid:16)P × V, R0g∗(cid:16) L ⊗ p∗ Y =0, Y OY (cid:0)−DY(cid:1) ⊗ CY,DY(cid:17)(cid:17) OY (cid:0)−DY(cid:1) ⊗ CY,DY(cid:17)(cid:17) and so that (9) follows. To prove that pV ∗R0g∗(cid:16) L−1 ⊗ CY,DY(cid:17) is torsion free, now it suffices to show that (10) pV ∗R0g∗(cid:16) L−1 ⊗ CY,DY(cid:17) = R0Hom (E, OV ) . By Grothendieck Duality, we have that RHom (E, OV ) ≃DV (cid:16)RpV ∗R0g∗(cid:16) L ⊗ p∗ ≃RpV ∗DP ×V (cid:16)R0g∗(cid:16) L ⊗ p∗ OY (cid:0)−DY(cid:1) ⊗ CY,DY(cid:17)(cid:17) [−m] OY (cid:0)−DY(cid:1) ⊗ CY,DY(cid:17)(cid:17) [−m]. Y Y (11) LOG-KODAIRA DIMENSION AND ZEROS OF HOLOMORPHIC LOG-ONE-FORMS 21 Note that by definition, L ⊗ p∗ Y By Proposition 23 and T ∗ OY (cid:0)−DY(cid:1) = p∗ Y (cid:0)OY (cid:0)−EY(cid:1) ⊗ g∗p∗A(cid:1) . (P,L) = P × V , we have R0g∗(cid:0)CY,DY ⊗ p∗ R0g∗(cid:0)CY,DY ⊗ p∗ OY (cid:0)EY − DY(cid:1)(cid:1) ≃G(cid:0)(cid:0)H0g+eωY [∗DY ](cid:1)!L(cid:1) , OY (cid:0)−EY(cid:1)(cid:1) ≃G(cid:0)(cid:0)H0g+eωY [!DY ](cid:1)∗L(cid:1) . Y Y Since up to a Tate twist, we have by Proposition 18 and Proposition 19, we have DP (cid:0)H0g+eωY [!DY ](cid:1) = H0g+eωY [∗DY ], OY (cid:0)−EY(cid:1)(cid:1) , p∗ RHom(cid:0)R0g∗(cid:0)CY,DY ⊗ p∗ ≃G(cid:0)D(P,L)(cid:0)H0g+eωY [!DY ](cid:1)∗L(cid:1) ≃G(cid:0)(cid:0)H0g+eωY [∗DY ](cid:1)!L(cid:1) ≃R0g∗(cid:0)CY,DY ⊗ p∗ OY (cid:0)EY − DY(cid:1)(cid:1) . Y Y P ωP [m](cid:1) Comparing it with (11), and by definition L−1 = p∗ Y (cid:0)OY (cid:0)EY − DY(cid:1) ⊗ g∗p∗A−1(cid:1) , RHom (E, OV ) ≃ RpV ∗R0g∗(cid:16) L−1 ⊗ CY,DY(cid:17) , we obtain which implies (10). (cid:3) References [AK00] D. Abramovich and K. Karu. Weak semistable reduction in characteristic 0. Invent. Math., 139(2):241–273, 2000. [AKMWo02] Dan Abramovich, Kalle Karu, Kenji Matsuki, and Jaros l aw W l odarczyk. Torifi- cation and factorization of birational maps. J. Amer. Math. Soc., 15(3):531–572, 2002. [Iit76] [HK10] [Fuj14] [Har77] [BCHM10] Caucher Birkar, Paolo Cascini, Christopher D. Hacon, and James McKernan. Ex- istence of minimal models for varieties of log general type. J. Amer. Math. Soc., 23(2):405–468, 2010. Osamu Fujino. On quasi-albanese maps. 2014. Robin Hartshorne. Algebraic geometry. Springer-Verlag, New York-Heidelberg, 1977. Graduate Texts in Mathematics, No. 52. Christopher D. Hacon and S´andor J. Kov´acs. Classification of higher dimensional algebraic varieties, volume 41 of Oberwolfach Seminars. Birkhauser Verlag, Basel, 2010. Shigeru Iitaka. Logarithmic forms of algebraic varieties. J. Fac. Sci. Univ. Tokyo Sect. IA Math., 23(3):525–544, 1976. Yujiro Kawamata. Characterization of abelian varieties. Compositio Math., 43(2):253–276, 1981. S´andor Kov´acs and Zsolt Patakfalvi. Projectivity of the moduli space of stable log- varieties and subadditivity of log-kodaira dimension. Journal of the American Math- ematical Society, 2016. David Mumford. Abelian varieties, volume 5 of Tata Institute of Fundamental Re- search Studies in Mathematics. Published for the Tata Institute of Fundamental Research, Bombay; by Hindustan Book Agency, New Delhi, 2008. With appendices by C. P. Ramanujam and Yuri Manin, Corrected reprint of the second (1974) edition. Mihnea Popa. Kodaira-Saito vanishing and applications. Enseign. Math., 62(1-2):49– 89, 2016. [Mum08] [Kaw81] [Pop16] [KP16] 22 [PS13] [PS14] [Sai90] [SS16] [VZ01] [Wei16] [Wei17] CHUANHAO WEI Mihnea Popa and Christian Schnell. Generic vanishing theory via mixed Hodge modules. Forum Math. Sigma, 1:e1, 60, 2013. Mihnea Popa and Christian Schnell. Kodaira dimension and zeros of holomorphic one-forms. Ann. of Math. (2), 179(3):1109–1120, 2014. Morihiko Saito. Mixed Hodge modules. Publ. Res. Inst. Math. Sci., 26(2):221–333, 1990. Claude Sabbah and Christian Schnell. The MHM project. 2016. Eckart Viehweg and Kang Zuo. On the isotriviality of families of projective manifolds over curves. J. Algebraic Geom., 10(4):781–799, 2001. Chuanhao Wei. Fibration of log-general type space over quasi-abelian varieties. arXiv preprint arXiv:1609.03089, 2016. Chuanhao Wei. Logarithmic comparison theorems in mixed hodge modules. preprint, 2017.
1701.01986	1	1701	2017-01-08T17:16:44	Picard curves with small conductor	[ "math.AG", "math.NT" ]	We study the conductor of Picard curves over $\mathbb{Q}$, which is a product of local factors. Our results are based on previous results on stable reduction of superelliptic curves that allow to compute the conductor exponent $f_p$ at the primes $p$ of bad reduction. A careful analysis of the possibilities of the stable reduction at $p$ yields restrictions on the conductor exponent $f_p$. We prove that Picard curves over $\mathbb{Q}$ always have bad reduction at $p=3$, with $f_3\geq 4$. As an application we discuss the question of finding Picard curves with small conductor.	math.AG	math	Picard curves with small conductor Michel Borner, Irene I. Bouw, and Stefan Wewers Abstract We study the conductor of Picard curves over Q, which is a product of local factors. Our results are based on previous results on stable reduction of su- perelliptic curves that allow to compute the conductor exponent f p at the primes p of bad reduction. A careful analysis of the possibilities of the stable reduction at p yields restrictions on the conductor exponent f p. We prove that Picard curves over Q always have bad reduction at p = 3, with f3 ≥ 4. As an application we discuss the question of finding Picard curves with small conductor. Key words: 2010 Mathematics Subject Classification. Primary 14H25. Secondary: 11G30, 14H45. 1 Introduction Let Y be a smooth projective curve of genus g over a number field K. To simplify the exposition, let us assume that K = Q. With Y we can associate an L-function L(Y,s) and a conductor NY ∈ N. Conjecturally, the L-function satisfies a functional equation of the form By definition, both L(Y,s) and NY are a product of local factors. In this paper we are really only concerned with the conductor, which can be written as NY = (cid:213) p f p. p Institut fur Reine Mathematik, Universitat Ulm e-mail: [email protected], e-mail: [email protected] 1 where L (Y,s) = ±L (Y,2− s), L (Y,s) := √NY s · (2p )−gs·G (s)g · L(Y,s). 2 Borner, Bouw, Wewers The exponent f p is called the conductor exponent of Y at p. It is known that f p only depends on the ramification of the local Galois representation associated with Y . In particular, if Y has good reduction at p then f p = 0. If Y has bad reduction at p then the computation of f p can be quite difficult. Until recently, an effective method for computing f p was only known for elliptic curves ([24], §IV.10) and for genus 2 curves if p 6= 2 ([11]). It was shown in [2] that f p can effectively be computed from the stable reduction of Y at p. Moreover, for certain families of curves (the superelliptic curves) we gave a rather simple recipe for computing the stable reduction. The latter result needed the assumption that p does not divide the degree n. In [18] this restriction is removed for superelliptic curves of prime degree. In the present paper we systematically study the case of Picard curves. These are superelliptic curves of genus 3 and degree 3, given by an equation of the form Y : y3 = f (x) = x4 + a3x3 + a2x2 + a1x + a0, with f ∈ Q[x] separable. Picard curves form in some sense the next family of curves to study after hyperelliptic curves. They are interesting for many reasons and have been intensively studied, see e.g. [14], [8], [9], and [16]. Our main results classify all possible configurations for the stable reduction of a Picard curve at a prime p, and use this to determine restrictions on the conductor exponents. For instance, we prove the following. Theorem 1.1. Let Y be a Picard curve over Q. (a) Then Y has bad reduction at p = 3, and f3 ≥ 4. (b) For p = 2 we have f2 6= 1. (c) For p ≥ 5 we have f p ∈ {0,2,4,6}. Theorem 3.6 is a somewhat stronger version of the first statement. Theorem 4.4 contains the last two statements. We also give explicit examples, showing that at least part of our results are sharp. Our result can be seen as a complement, for Picard curves, to a result of Brumer -- Kramer ([3], Theorem 6.2), who prove an upper bound for f p for abelian varieties of fixed dimension. Since the conductor of a curve coincides with that of its Jacobian, the result applies to our situation, as well. A more careful case-by-case analysis, combined with ideas from [3], could probably be used to obtain a more precise list of possible values for the conductor exponent at p = 2,3, as well. In the last section we discuss the problem of constructing Picard curves with small conductor. As a consequence of the Shafarevich conjecture (aka Faltings' The- orem), there are at most a finite number of nonisomorphic curves of given genus and of bounded conductor. But except in very special cases, no effective proof of this theorem is known. In his recent PhD thesis, the first named author has made an extensive search for Picard curves with good reduction outside a small set of small primes, and computed their conductor. The Picard curve with the smallest conductor that was found is the curve Picard curves with small conductor 3 which has conductor Y : y3 = x4 − 1, NY = 2636 = 46656. We propose as a subject for further research to either prove that the above example is the Picard curve over Q with the smallest possible conductor, or to find (one or all) counterexamples. We believe that the methods presented in this paper may be very helpful to achieve this goal. 2 Semistable reduction We first introduce the general setup concerning the stable reduction and the con- ductor exponents of Picard curves. As explained in the introduction, the conductor exponent is a local invariant, encoding information about the ramification of the lo- cal Galois representation associated with the curve. Therefore, we may replace the number field K by its strict henselization. In other words, we may work from the start over a henselian field of mixed characteristic with algebraically closed residue field. 2.1 Setup and notation Throughout Section 2 - 4 the letter K will denote a field of characteristic zero that is henselian with respect to a discrete valuation. We denote the valuation ring by OK, the maximal ideal of OK by p and the residue field by k = OK/p. We assume that k is algebraically closed of characteristic p > 0. The most important example for us is when K = Qnr p is the maximally unramified extension of the p-adic numbers. Then p = (p) and k = ¯Fp. Let Y /K be a Picard curve, given by the equation Y : y3 = f (x), (1) K and interpret where f ∈ K[x] is a separable polynomial of degree 4. We set X := P1 (1) as a finite cover f : Y → X, (x,y) 7→ x, of degree 3. By the Semistable Reduction Theorem (see [5]), there exists a finite extension L/K such that the curve YL := Y ⊗K L has semistable reduction. Since g(Y ) = 3 ≥ 2, there even exists a (unique) distinguished semistable model Y → Spec OL of YL, the stable model ([5], Corollary 2.7). The special fiber ¯Y := Ys of Y is called the stable reduction of Y . It is a stable curve over k ([5], § 1), and it only depends on Y , up to unique isomorphism. It is no restriction to assume that the extension L/K is Galois and contains a third to L) is a root of unity z 3 ∈ L. Then the cover f L : YL → XL (the base change of f 4 Borner, Bouw, Wewers Galois cover. Its Galois group G is cyclic of order 3, generated by the element s which is determined by s (y) = z 3y. Let G := Gal(L/K) denote the Galois group of the extension L/K. The group G acts faithfully and in a natural way on the scheme YL = Y ⊗K L. We denote by G the subgroup of Aut(YL) generated by G and the image of G . By definition, G is a semidirect product, G = G ⋊G . The action of G on G via conjugation is determined by the following formula: for t in G we have tst −1 =(s s 2 if t (z 3) = z 3, if t (z 3) = z 2 3 . (2) (b) Let ¯X := X ⊗ k denote the special fiber of X and ¯f Because of the uniqueness properties of the stable model, the action of G on YL extends to an action on Y . By restriction, we see that G has a natural, k-linear action on ¯Y . This action will play a decisive role in our analysis of the stable reduction ¯Y . For the rest of this subsection we focus on the action of the subgroup G ⊂ G. The role of the subgroup G ⊂ G will become important later. Remark 2.1. (a) The quotient scheme X := Y /G is a semistable model of XL = P1 L, see e.g. [17], Cor. 1.3.3.i. Since the map Y → X is finite and Y is normal, Y is the normalization of X in the function field of YL. This means that Y is uniquely determined by a suitable semistable model X of XL. : ¯Y → ¯X the induced map. We note that ¯f is a finite G-invariant map. It is not true in general that ¯Y /G = ¯X. However, the natural map ¯Y /G → ¯X is radicial and in particular a homeomor- phism (see e.g. [17], Prop. 2.4.11). (c) Every irreducible component W ⊂ ¯Y is smooth. To see this note that the quotient of W by its stabilizer in G is homeomorphic to an irreducible component Z ⊂ ¯X, which is a smooth curve of genus 0. If W has a singular point, then s acts on W and permutes the two branches of W passing through this point. But since s has order 3, this is impossible. Let D ¯Y denote the component graph of ¯Y : the vertices are the irreducible components of ¯Y and the edges correspond to the singular points. The stability condition for ¯Y means that an irreducible component of genus 0 corresponds to a vertex of D ¯Y of degree ≥ 3. The number of loops of D ¯Y is given by the well known formula ¯Y , Q) = r− s + 1, where r is the number of edges and s the number vertices of D g ( ¯Y ) := dimQ H1(D ¯Y . The curve ¯X is also semistable, but in general not stable. Since ¯X has arithmetic ¯X is a tree, and every vertex corresponds to a smooth genus 0, the component graph D curve of genus 0. It follows from Remark 2.1 that D Lemma 2.2. If W ⊂ ¯Y is an irreducible component, then s (W ) = W . ¯X = D ¯Y /G. (3) Picard curves with small conductor 5 Proof. To derive a contradiction, we assume that W1,W2,W3 ⊂ ¯Y are three distinct components that form a single G-orbit. Then Wi ∼→ Z := ¯f (Wi). Since Z is a com- ponent of ¯X, we conclude that g(Wi) = 0, for i = 1,2,3. The stability condition on ¯Y implies that each Wi contains at least three singular points of ¯Y . Hence Z also contains at least three singular points of ¯X. Let ¯Y → ¯Y0 denote the unique morphism which contracts all components of ¯Y except the Wi and which is an isomorphism on the intersection of ∪iWi with the smooth locus of ¯Y . Similarly, let ¯X → ¯X0 be the map contracting all components of ¯X except Z. These maps fit into a commutative diagram ¯Y ¯X ¯Y0 / ¯X0, where the vertical arrows are quotient maps by the group G (at least for the under- lying topological spaces). Also, ¯X0 ∼= Z. Let ¯x ∈ Z be one of the singular point of ¯X lying on Z, and let T ⊂ ¯X the closed subset which is contracted to ¯x ∈ Z = ¯X0. Then T is a nonempty and connected union of irreducible components of ¯X and hence a semistable curve of genus 0. In particular, the component graph of T is a tree. Let Z′ ⊂ T be a tail component. As a component of ¯X, Z′ intersects the rest of ¯X in at most two points. Let W′ ⊂ ¯Y be an irreducible component lying above Z′. The stability of ¯Y implies that s (W ′) = W′ and that the action of s on W′ is nontrivial. (Otherwise W′ would be homeomorphic to Z′, and hence W′ would be a component of genus 0 intersecting the rest of ¯Y in at most two points. ) It follows that the inverse image S ⊂ ¯Y of T is connected. Note that S meets the component Wi in the unique point on Wi above ¯x. Since S is connected, it follows that the map ¯Y → ¯Y0 contracts S to a single point. We conclude that the curve ¯Y0 has at least three distinct singular points where all three components Wi meet. Equation (3) implies that g ( ¯Y0) is at least 1. It follows that the arithmetic genus of ¯Y0 is ≥ 4, and hence g( ¯Y ) ≥ 4 as well. This is a contra- diction, and the lemma follows. ⊓⊔ 2.2 The conductor exponent Let cp be the conductor of the Gal( ¯K/K)-representation H1 definition, this is an ideal of OK of the form et(Y ¯K, Qℓ), see [22]. By cp = p fp, with fp ≥ 0. The integer fp is called the conductor exponent of Y /K.1 1 When working in a local context, fp is often simply called the conductor of Y . / / / 6 Borner, Bouw, Wewers We recall from [2] an explicit formula for fp, in terms of the action of G = Gal(L/K) on ¯Y . For this we let G u ⊂ G , for u ≥ 0, denote the uth higher ramification group (in the upper numbering). We set ¯Y u := ¯Y /G u. Note that ¯Y u is a semistable curve for all u. Note also that G = G 0 because the residue field k is assumed to be algebraically closed. Proposition 2.3. The conductor exponent of the curve Y /K is given by where and fp = e + d , et( ¯Y 0, Qℓ) e := 6− dim H1 d :=Z 0 (cid:0)6− 2g( ¯Y u)(cid:1)du. (4) (5) (6) ⊓⊔ Proof. See [2], Theorem 2.9 and [1], Corollary 2.14. The ´etale cohomology group H1 et( ¯Y u, Qℓ) decomposes as et( ¯Y u, Qℓ) = ⊕W H1 H1 et(W, Qℓ)⊕ H1(D ¯Y u, Qℓ), where the first sum runs over the set of irreducible components W of the normal- ization of ¯Y u and D ¯Y u is the graph of components of ¯Y u. (See [2], Lemma 2.7.(1).) Therefore, the second term in (5) can be written as dim H1 et( ¯Y 0, Qℓ) = (cid:229) W dim H1 et(W, Qℓ) + dim H1(D ¯Y 0). The arithmetic genus of ¯Y u, which occurs in (6), is given by the formula g( ¯Y u) = (cid:229) W g(W ) + dim H1(D ¯Y u). (7) (8) For future reference we note that dim H1 dim H1(D bounded by g( ¯Y 0), and hence by g(Y ) = 3. ¯Y 0) can be interpreted as the number of loops of the graph D et(W, Qℓ) = 2g(W ). The integer g ( ¯Y 0) := ¯Y 0. It is Lemma 2.4. The following statements are equivalent. (a) d = 0. (b) G u acts trivially on ¯Y , for all u > 0. (c) The curve Y has semistable reduction over a tamely ramified extension of K. Proof. Assume that d = 0. By (6) this means that 3 = g( ¯Y ) = g( ¯Y u) for all u > 0. Using (8) one easily shows that this means that G u acts trivially on the component graph D ¯Y of ¯Y . Moreover, for every component W ⊂ ¯Y we have g(W ) = g(W /G u). It follows that G u acts trivially on ¯Y . We have proved the implication (a)⇒(b). The implication (b)⇒(c) follows from [12], Theorem 4.44. The implication (c)⇒(a) fol- lows immediately from the definition of d . ⊓⊔ ¥ Picard curves with small conductor 3 The wild case: p = 3 7 In this section we assume that p = 3. We first analyze the special fiber of the stable model of YL, and show that there are essentially five reduction types. From § 3.2 we consider the case where K is absolutely unramified, and derive a lower bound for the conductor exponent f3. 3.1 The stable model We keep all the notation introduced in § 2. In addition, we assume that p = 3. Lemma 2.2 implies that we can distinguish between two types of irreducible components of ¯Y . Definition 3.1. An irreducible component W ⊂ ¯Y is called ´etale if the restriction s W ∈ Autk(W ) is nontrivial. If s W is the identity, then W is called an inseparable component. Let W ⊂ ¯Y be an irreducible component, and let Z := ¯f (W ) ⊂ ¯X be its image. Then Z is an irreducible component of ¯X and hence a smooth curve of genus 0. Lemma 2.2 shows that s (W ) = W . It follows that W /G → Z is a homeomorphism. If W is an inseparable component, then W → Z is a purely inseparable homeomor- phism (since W → Z has degree 3, this can only happen when p = 3). It follows that every inseparable component has genus zero. If W is an ´etale component, then Z ∼= W /G, and W → Z is a G-Galois cover. For future reference we recall that the Riemann -- Hurwitz formula for wildly ramified Galois covers of curves yields 2g(W )− 2 = −2· 3 +(cid:229) z 2(hz + 1), (9) where the sum runs over the branch points of W → Z and hz is the (unique) jump in the filtration of the higher ramification groups in the lower numbering. We have that hz ≥ 1 is prime to p ([21], § IV.2, Cor. 2 to Prop. 9). Theorem 3.2. We are in exactly one of the following five cases. (a) The curve ¯Y is smooth and irreducible. (b) There are exactly two components W1,W2 which are both ´etale, meet in a single point, and have genus g(W1) = 2, g(W2) = 1. (c) There are three ´etale components W1,W2,W3 of genus one, and one inseparable component W0 of genus zero. For i = 1,2,3, Wi intersects W0 in a unique point, and these intersection points are precisely the singular points of ¯Y . (d) There are two components W1,W2 which are ´etale of genus g(W1) = 1, g(W2) = 0. There are exactly three singular points, which form an orbit under the action of G, and where W1 and W2 meet. 8 Borner, Bouw, Wewers (e) There are three components W1,W2,W3, which are ´etale and of genus g(W1) = g(W2) = 0 and g(W3) = 1, and four singular points. Three of the singular points are points of intersection of W1 and W2, and form an orbit under the action of G. The fourth singular point is the point of intersection of W2 and W3. Proof. Let r1 (resp. s1) be the number of singular points (resp. irreducible compo- nents) of ¯Y which are fixed by s , and let r2 (resp. s2) be the number of orbits of singular point (resp. irreducible components) of ¯Y of length 3. Lemma 2.2 states that s2 = 0. Therefore, (3) becomes Because D ¯X = D g ( ¯Y ) = r− s + 1 = r1 + 3r2 − s + 1. ¯Y /G is a tree, we have g ( ¯X) := dim H1(D ¯X ) = r1 + r2 − s + 1 = 0. Combining (10) and (11) we obtain g ( ¯Y ) = 2r2. Since 0 ≤ g ( ¯Y ) ≤ 3, we conclude that g ( ¯Y ) ∈ {0,2} and r2 ∈ {0,1}. (10) (11) (12) Case 1: r2 = 0 and g ( ¯Y ) = 0. In this case D ¯Y is a tree, and the sum of the genera of all irreducible components is 3. In particular, there are at most 3 components of genus > 0. Moreover, the stability condition implies that every component of genus zero contains at least three singular points of ¯Y . It is an easy combinatorial exercise to see that this leaves us with exactly four possibilities for the tree D ¯Y . Going through these four cases we will see that one of them is excluded, while the remaining three correspond to Case (a), (b), and (c) of Theorem 3.2. The first case is when ¯Y has a unique irreducible component. Then ¯Y is smooth. This is Case (a) of the lemma. Secondly, there may be two irreducible components, of genus 1 and 2, and a unique singular point. This corresponds to Case (b). Thirdly, there may be three irreducible components, each of genus 1, and two singular points. We claim that this case cannot occur. Indeed, one of the three com- ponents would contain two singular points, and each of these two points must be a fixed point of s . It follows that the G-cover W → Z = W /G is ramified in at least two points. The Riemann -- Hurwitz formula (9) implies that g(W ) ≥ 2. This yields a contradiction, and we conclude that this case does not occur. Finally, in the last case, there are four singular points and four irreducible compo- nents. Three of them have genus 1 and one has genus zero. The component of genus zero necessarily contains all three singular points. A similar argument as in the pre- vious case shows that the genus-0 component cannot be ´etale. This corresponds to Case (c). Case 2: g ( ¯Y ) = 2 and r2 = 1. In this case the sum of the genera of all components is equal to 1. Therefore, there Picard curves with small conductor 9 must be a unique component of genus 1, and all other components have genus 0. Let W1 and W2 be two components which meet in a singular point ¯y such that s ( ¯y) 6= ¯y. Since s (Wi) = Wi for i = 1,2 (Lemma 2.2), W1 and W2 are ´etale components and intersect each other in exactly three points (the G-orbit of ¯y). If there are no further components, we are in Case (d). Assume that there exists a third component W3. Let T ⊂ ¯Y be the maximal connected union of components which contains W3 but neither W1 nor W2. Then T contains a unique component W0 which meets either W1 or W2 in a singular point. The component graph of T is a tree, and we consider W0 as its root. By the stability condition, every tail component of T must have positive genus, so T has a unique tail. If W0 is not this tail, it has genus 0 and intersects the rest of ¯Y in exactly 2 points. This contradicts the stability condition. We conclude that ¯Y has exactly three components, of genus g(W1) = g(W2) = 0 and g(W3) = 1. This is Case (e) of the lemma. Now the proof is complete. ⊓⊔ 3.2 A lower bound for f3 We continue with the assumptions from the previous subsection. In addition, we assume that K is absolutely unramified. By this we mean that p = (3). Under this assumption, we prove a lower bound for the conductor exponent f3 := fp. In fact, we will give a lower bound for e , where f3 = e + d is the decomposition from Proposition 2.3. If L/K is at most tamely ramified, then d = 0 (Lemma 2.4). In this case, our bounds are sharp. Since K is absolutely unramified, the third root of unity z 3 ∈ L is not contained in K. Therefore, there exists an element t ∈ G = Gal(L/K) such that t (z 3) = z 2 3 . Let m be the order of t . After replacing t by a suitable odd power of itself we may assume that m is a power of 2. We keep this notation fixed for the rest of this paper. Recall that the semidirect product G = G ⋊G acts on ¯Y in a natural way. The following observation is crucial for our analysis of the conductor exponent. Lemma 3.3. Let W ⊂ ¯Y be an ´etale component such that t (W ) = W . Then inside the automorphism group of W we have t ◦ s ◦ t −1 = s 2 6= s . (13) In particular, t W is nontrivial. Proof. The statement follows immediately from Equation (2) and Definition 3.1. ⊓⊔ Despite its simplicity, Lemma 3.3 has the following striking consequence. Note that we consider potentially good but not good reduction as bad reduction in this paper. Proposition 3.4. Assume that p = (3). Then every Picard curve Y over K has bad reduction. 10 Borner, Bouw, Wewers Proof. Lemma 3.3 implies that Y acquires semistable reduction only after passing to a ramified extension L ∋ z 3. Therefore Y /K does not have good reduction. The fact that f3 6= 0 follows from Proposition 2.3, together with the fact that t acts nontrivially on each irreducible component of ¯Y (Lemma 3.3). ⊓⊔ In order to prove more precise lower bounds for f3, we need to analyze the action of s and t on ¯Y in more detail. Lemma 3.5. Let W ⊂ ¯Y be an ´etale component. Then one of the following cases occurs: g(W /G 0) h 1 2 g(W ) r 1 1 2 (1,1) 1 0 1 2 3 0 0 1 0 4 Here r is the number of ramification points of the G-cover W → Z := W /G and h lists the set of lower jumps. The fourth column gives an upper bound for the genus of W /G 0. Proof. Recall that we have assumed that the order m of t is a power of 2. Assume that W is an ´etale component of ¯Y such that f The Riemann -- Hurwitz formula (9) immediately yields the cases for g(W ),r, and h stated in the lemma, together with one additional possibility: the curve W has genus 3 and f : W → Z ∼= P1 is branched at two points, with lower jump 1 and 2, respectively. We claim that this case does not occur. : W → Z is branched at 2 points. Lemma 3.3 implies that t acts nontrivially on W . Since t normalizes s and the two ramification points have different lower jumps, it follows that t fixes both ramification points wi of f . We conclude that H := hs ,t i acts on W as a nonabelian group of order 6 fixing the vi. We write hi for lower jump of wi. Lemma 2.6 of [15] implies that gcd(hi,m) is the order of the prime-to-3 part of the centralizer of H. Since gcd(h1,m) 6= gcd(h2,m) we obtain a contradiction, and conclude that this case does not occur. cases. This is also an upper bound for g(W /G 0). We compute an upper bound for the genus of W /ht i in each of the remaining In the case that g(W ) = 0 there is nothing to prove. In the case that g(W ) = 1, the automorphism t fixes the unique ramification point of f , hence g(W /G 0) = 0. Assume that g(W ) = 2. The Riemann -- Hurwitz formula immediately implies that that g(W /ht i) ≤ 1. Finally, we consider the case that g(W ) = 3, i.e. Y has potentially good reduction. As before, we have that t fixes the unique fixed point of s . Put H = hs ,t i. Lemma 3.3 together with the assumption that the order m of t is a power of 2 implies that the order of the prime-to-p centralizer of H is gcd(h = 4,m) = m/2. It follows that m = 8. Since t has at least one fixed point on W , namely the point at ¥ , the Riemann -- Hurwitz formula implies that g(W /ht i) = 0. This finishes the proof of ⊓⊔ the lemma. Picard curves with small conductor 11 We have now all the necessary tools to prove our main theorem. Theorem 3.6. Assume p = (3), and let Y be a Picard curve over K. The conductor exponent f3 of Y /K satisfies f3 ≥ 4. L/K. Moreover: (a) If f3 ≤ 6 then Y achieves semistable reduction over a tamely ramified extension (b) If f3 = 4 then we are in Case (b) or Case (c) from Theorem 3.2. (c) If f3 = 5 then we are in Case (d) or in Case (e) of Theorem 3.2. Proof. We use the assumptions and notations from the beginning of § 3.2. Recall that the inertia subgroup G 0 ⊂ G := Gal(L/K) acts on the geometric special fiber ¯Y of the stable model of YL and that .the quotient ¯Y 0 = ¯Y /G 0 is again a semistable curve. Claim: We have that dim H1 et( ¯Y 0, Qℓ) ≤ 2. (14) ¯Y ) to dim H1 et( ¯Y 0, Qℓ) is 2g(W ). The contribution of H1(D Note that (14), together with (4) and (5), immediately implies the first statement f3 ≥ 4 of the theorem. Recall from (8) and (3) that the contribution of a smooth component W of ¯Y 0 to et( ¯Y 0, Qℓ) is g ( ¯Y 0), dim H1 which is less than or equal to g( ¯Y 0). Let W ⊂ ¯Y be an irreducible component, and denote by W 0 ⊂ ¯Y0 its image in ¯Y 0. Clearly, g(W 0) ≤ g(W ). Moreover, if t (W ) = W then Lemma 3.5 shows that g(W 0) ≤ 1. Let us consider each case of Theorem 3.2 separately. In Case (a), ¯Y is smooth and irreducible of genus 3. Then ¯Y 0 is also smooth and irreducible, and Lemma et( ¯Y 0, Qℓ) = 0, 3.5 shows that g( ¯Y 0) = 0. So in Case (a) we have proved dim H1 which is strictly stronger than (14). Similarly, in Case (b) Lemma 3.5 shows that ¯Y 0 consists of two irreducible components which meet in a single point. One of these components has genus zero, the other one has genus ≤ 1. Therefore, (14) holds in Case (b). Assume that we are in Case (c). Let W1,W2,W3 denote the three components of genus 1, and W 0 is a power of two, t fixes exactly one of these components (say W1), or all three. In the first case, et( ¯Y 0, Qℓ) = 1. In the g(W 0 et( ¯Y 0, Qℓ) = 0. In both cases, (14) second case, g(W 0 holds. Now assume that we are in Case (d). The action of G 0 must fix both components W1,W2, since g(W1) 6= g(W2). Lemma 3.5 shows that g(Wi/G 0) = 0, for i = 1,2. Also, t permutes the three singular points of ¯Y . But these points form one orbit under the action of G. Hence it follows from (13) that t fixes exactly one singular point and permutes the other two. We conclude that the curve ¯Y 0 has two smooth components i , i = 1,2,3, their images in ¯Y 0. Since the order of t i ) = 0 for i = 1,2,3, and dim H1 1 ) = 0 by Lemma 3.5, and W 0 2 = W 0 3 . Therefore, dim H1 12 Borner, Bouw, Wewers et( ¯Y 0, Qℓ) ≤ 1. of genus 0 which meet in at most two points. We conclude that dim H1 A similar analysis shows that the same conclusion holds in Case (e). This proves the claim (14). While proving the claim, we have shown the following stronger conclusion: dim H1 et( ¯Y 0, Qℓ) ∈  Case (a), {0}, {0,2}, Case (b), (c), {0,1}, Case (d), (e). (15) It follows that e = 6 in Case (a), e ∈ {4,6} in the Cases (b) and (c), and e ∈ {5,6} in the Cases (d) and (e). The remaining statement that Y acquires stable reduction over a tamely ramified ⊓⊔ extension L of K in the case that f3 ≤ 6 follows from Lemma 2.4. Corollary 3.7. If p = (3) and Y has potentially good reduction, then f3 ≥ 6. 3.3 Examples In this section we discuss two explicit examples of Picard curves over Qnr 3 in some detail. These examples show, among other things, that the lower bounds for f3 given by Theorem 3.6 are sharp. Let us fix some notation. We set K := Qnr 3 . Given a suitable finite extension L/K, we denote by vL the unique extension of the 3-adic valuation to L (which is normal- ized such that vL(3) = 1). We let F(XL) denote the function field of XL := P1 L, and identify F(XL) with the rational function field L(x). For a Picard curve Y over K given by y3 = f (x) for a quartic polynomial f ∈ K[x] the function field F(YL) of YL is the degree-3 extension of F(XL) obtained by adjoining the function y. Let X be a semistable model of XL, and let Z1, . . . ,Zn ⊂ ¯X := X ⊗ FL denote the irreducible components of the special fiber. Since each Zi is a prime divisor on X , it gives rise to a discrete valuation vi on F(XL), extending vL. It has the property that the residue field of vi can be naturally identified with the function field of Zi. Since XL is simply a projective line and X is a semistable model, the valuations vi have a simple description, as follows. For all i, there exists a coordinate xi ∈ F(XL) such that vi is the Gauss valuation on F(XL) = L(xi) with respect to xi. The coordinate xi is related to x by a fractional linear transformation x = aixi + bi cixi + di , with aidi − bici 6= 0. It can be shown that the model X is uniquely determined by the set {v1, . . . ,vn}, see [2] or [19]. Let Y denote the normalization of X inside the function field F(YL). Then Y is a normal integral model of YL. In general, Y has no reason to be semistable, and it is not clear in general how to describe its special fiber ¯Y := Y ⊗ k. However, each Picard curves with small conductor 13 irreducible component W ⊂ ¯Y corresponds again to a discrete valuation w on F(YL) extending vL, such that the residue field of w is the function field of W . It can be shown that this gives a bijection between the irreducible components of ¯Y and the set of discrete valuations on F(YL) extending one of the valuations vi (see e.g. [19], § 3). In many situations, the knowledge of all extensions of the vi to F(YL) will give enough information to decide whether the model Y is semistable and to describe its special fiber. We need one more piece of notation. For m > 1 prime to 3 we set Lm := K(p )/K where p m = −3. Then Lm/K is a tamely ramified Galois extension of degree m. The Galois group G := Gal(Lm/K) is cyclic and generated by the element t ∈ G m := Gal(Lm/K) determined by t (p ) = z mp , where z m ∈ K is a primitive mth root of unity (which exists because k is algebraically closed). Note also that Lm contains the third root of unity z 3 := −1 + p m/2 2 . We remark that the choice of t and m agrees with the notation chosen in § 3.2 Example 3.8. Let Y be the Picard curve over K given by the equation y3 = x4 + 1. (16) We claim that Y has potentially good reduction, which is attained over the tame extension L := L8 = K(p )/K, with p 8 = −3. To prove this, we apply the coordinate changes x = p 3x1, y = 1 + p 4y1 to (16). After a brief calculation, we obtain the new equation 1 − p 4y2 y3 1 − y1 = x4 1. (17) Equation (17) is equivalent to (16) in the sense that it defines a curve over K which is isomorphic to Y . Also, (17) defines an integral model Y of YL. Its special fiber is the curve over k = ¯F3 given by the (affine) equation 1 − y1 = x4 1. ¯Y : y3 This is a smooth curve of genus 3. It follows that Y has good reduction over L, as claimed. Since Y acquires stable reduction over a tame extension L/K, Lemma 2.4 implies that f3 = e . Equations (5) and (15) imply that f3 = 6. 14 Borner, Bouw, Wewers For completeness, we compute the action of G 0 = ht i on ¯Y explicitly. We con- sider t as an automorphism of the structure sheaf of Y . By definition, we have t (p ) = z 8p , t (x) = x, t (y) = y. It follows that t (x1) = z 5 8 x1, t (y1) = −y1. This describes t ¯Y as an automorphism of ¯Y of order 8, as expected from the proof of Lemma 3.5. Example 3.9. Let Y /K be the Picard curve Y : y3 = f (x) := 3x4 + x3 − 54. (18) We claim that Y has semistable reduction over the tame extension L := L4/K. More- over, the stable reduction ¯Y is as in Case (b) of Theorem 3.2, and f3 = 4. First we define a semistable model X of XL := P1 L by specifying two discrete valuations v1,v2 on F(XL) which extend vL. We then show that the normalization Y of X in F(YL) is the stable model of YL, and determine its special fiber ¯Y and the action of the inertia group of L/K on ¯Y . The valuation v1 is defined as the Gauss valuation on F(XL) = F(x1) with respect to the coordinate x1, which is related to x by x = p 2x1. (19) We claim that v1 has a unique extension w1 to F(YL) that is unramified. To show this, we need a so-called p-approximation of f with respect to v1, see [18]. In fact, we can write Here we have used the relation p 4 = −3. This suggests the coordinate change f = p 6(cid:0)x3 1 + p 6(2− x4 1)(cid:1). y = p 2(x1 + p 2y1). (20) (21) After a short calculation we obtain a new equation for YL: 1y1 = 2− x4 1. 1 − p 2x1y2 y3 1 − x2 If we consider (21) as defining an affine curve over OL, its special fiber is the affine curve over k with equation 1 − ¯x2 ¯y3 1 ¯y1 = −1− ¯x4 1. (22) In fact, (22) defines an irreducible affine curve with a cusp singularity in ( ¯x1, ¯y1) = (0,−1). It follows that the inverse image in ¯Y of ¯X1 is an irreducible component W1 of multiplicity one birationally equivalent to the curve given by (22). To compute the geometric genus of W1 we substitute ¯y1 = −1 + ¯x1 ¯z1 into (22) and obtain the Artin -- Schreier equation Picard curves with small conductor Using the Riemann -- Hurwitz formula, one sees that W1 has geometric genus 2. The valuation v2 of F(XL) corresponds to the choice of the coordinate x2 given by ¯z3 1 − ¯z1 = − ¯x−1 1 − ¯x1. x = 3(1 + p x2). 15 (23) (24) (25) (26) (27) (28) After a short calculation we can write f = 33(cid:0) (−1 + p x2)3 − 2p 6x2 2 + 32(. . .)(cid:1). This suggests the change of coordinate Plugging in (26) into (18) and using (25) we arrive at the equation y = 3(cid:0) (−1 + p x2) + p 2y2(cid:1). 2 − (−1 + p x2)2y2 = −2x2 Reducing (27) modulo p we obtain the irreducible equation 2 + p 2(−1 + p x2)y2 y3 2 + p 2(. . .). ¯y3 2 − ¯y2 = ¯x2 2, which defines a curve of genus 1. It follows that the inverse image of ¯X2 in ¯Y is an irreducible projective curve W2 of geometric genus 1. So ¯Y consists of two irreducible components W1 and W2 of geometric genus 2 and 1. On the other hand, Ys is known to have arithmetic genus 3. By a standard argument (see e.g. ) we can conclude that W1, W2 are smooth and meet transversely in a single point. This shows that Y has semistable reduction over the tame extension L4/K, with a stable model of type (b). Let us try to analyze the action of G = G 0 = ht i on ¯Y . By definition, t (p ) = z 4p , t (x) = x and t (y) = y. From (19) and (20) we deduce that t W1 is given by t ( ¯x1) = − ¯x1, t ( ¯y1) = ¯y1, t (¯z1) = −¯z1. From (24) and (26) we see that t ( ¯x2) = z 3 4 ¯x2, t ( ¯y2) = − ¯y2. 1 = 2 = W2/G 0, meeting in a single point. An easy calculation (compare 2 ) = 0. It follows that It follows that the curve ¯Y 0 := ¯Y /G 0 has two irreducible smooth components, W 0 W1/G 0 and W 0 with the proof of Lemma 3.5) shows that g(W 0 g( ¯Y 0) = 1 and dim H1(D Remark 3.10. The two examples discussed above are quite special. Typically, the extension L/K needs to be wildly ramified, and have rather large degree. It is then hard (and often practically impossible) to do computations as above by hand. Most of the examples in [4] and this paper have been computed with the help of (earlier ¯Y 0) = 0 and hence f3 = 6− 2 = 4. 1 ) = 1 and g(W 0 16 Borner, Bouw, Wewers versions of) Julian Ruth's Sage packages mac lane and completion (available at https://github.com/saraedum), and the algorithms from [2] and [18]. 4 The tame case: p 6= 3 In this section we assume that the residue characteristic p of our ground field K is different from 3. In this case it is much easier to analyze the semistable reduction of Picard curves and to compute the conductor exponent fp than for p = 3. The theoretical background for this are the admissible covers, see [7], [12], § 10.4.3, or [26]. In the case of superelliptic curves the computation of fp has already been described in detail in [2], hence we can be much briefer than in the previous section. 4.1 The stable model Let K be as in § 2.1, with p 6= 3. Let Y /K be a Picard curve, given by an equation Y : y3 = f (x), where f ∈ K[x] is a separable polynomial of degree 4. Let L0/K denote the splitting field of f . Let L/L0 be a finite extension with ramification index 3 such that L/K is a Galois extension. Then [2], Corollary 4.6 implies that Y acquires semistable reduction over L. We note in passing that L/K is tamely ramified unless p = 2. This follows from the definition of the Galois extension L0/K, whose degree divides 4! = 24. K denote the branch divisor of the cover f A semistable model Y of YL may be constructed as follows, see [2], § 4. Let D ⊂ X = P1 : Y → X, consisting of the set of zeros of f and ¥ . Since L contains the splitting field of f , the pullback DL ⊂ YL consists of 5 distinct L-rational points. Let (X , D) denote the stably marked model of (XL,DL). By this we mean that X is the minimal semistable model of XL with the property that the schematic closure D ⊂ X of DL is ´etale over Spec OL and contained inside the smooth locus of X → Spec OL. Let ¯X := X ⊗ FL denote the special fiber of X and ¯D = D ∩ ¯X the specialization of DL. Then ( ¯X, ¯D) is a stable 5-marked curve of genus zero. This means that ¯X is a tree of projective lines, where every irreducible component has at least three points which are either marked (i.e. lie in the support of ¯D) or are singular points of ¯X. Let Y denote the normalization of X with respect to the cover YL → XL. Theo- rem 3.4 from [2] shows that Y is a quasi-stable model of YL. A priory, it is not clear whether Y is the stable model of Y . The following case-by-case analysis will show that it is. We will use the fact that the natural map Y → X is an admissible cover with branch locus D. In particular, the induced map Picard curves with small conductor 17 ¯f : ¯Y → ¯X Assume that x between the special fiber of Y and of X is generically ´etale and identifies ¯X with the quotient scheme ¯Y /G. We describe the restriction of the map ¯f to an irreducible component ¯Xi of ¯X. Without loss of generality we may assume that K (and hence L) contains a primitive 3rd root of unity z 3, which we fix. For each branch point x of ¯f ¯Xi the canonical generator of inertia g ∈ G is characterized by g∗u ≡ z 3u (mod u2), where u is a local parameter at ¯f −1 (x i). A branch point of ¯f ¯Xi is either the specialization of a branch point of f or a singular point of ¯X. ¯Xi tion shows that the canonical generator of inertia is s of x is the specialization of a branch point. An elementary calcula- is the specialization of and s 2 otherwise. Now let x be a singularity of ¯X, and denote the irreducible components intersecting in x by ¯X1 and ¯X2. Then the canonical generators gi of the restrictions ¯f ¯Xi at x satisfy (This last condition says that ¯f The upshot is that the map ¯f is an admissible cover.) : ¯Y → ¯X is completely determined and easily de- The following lemma lists the 5 possibilities for ¯X. Note that we need to dis- and the other 4 branch points. The proof is elementary, and scribed by the stably marked curve ( ¯X, ¯D). g1 = g−1 2 . tinguish between ¥ therefore omitted. Lemma 4.1. With assumptions and notations as in the beginning of the section, we have the following 5 possibilities for ¯X. (a) The curve ¯X is irreducible. (b) The curve ¯X consists of two irreducible components ¯X1 and ¯X2. Three of the branch points of f including ¥ specialize to ¯X1, the other two to ¯X2. (c) The curve ¯X consists of three irreducible components ¯X1, ¯X2, and ¯X3, where ¯X1 specializes to ¯X2, two other branch and ¯X3 intersect ¯X2. The branch point ¥ points specialize to ¯X1, and two to ¯X3. (d) The curve ¯X consists of two irreducible components ¯X1 and ¯X2. Three of the branch points of f different from ¥ specialize to ¯X1, the other two to ¯X2. (e) The curve ¯X consists of three irreducible components ¯X1, ¯X2, and ¯X3, where ¯X1 specialize to ¯X1, two other and ¯X3 intersect ¯X2. Two branch points including ¥ branch points specialize to ¯X3, and the last one to ¯X2. The following result immediately follows from the possibilities for ¯X, together with the fact that ¯f Theorem 4.2. Let K be as in §2.1, with p 6= 3. Let Y be a Picard curve over K, L/K a finite Galois extension over which Y has semistable reduction. Let Y denote the stable model of YL over OL and ¯Y := Y ⊗ k the special fiber. Then ¯Y is as in one of the following five cases. (a) The curve ¯Y is smooth. is an admissible cover. ¥ 18 Borner, Bouw, Wewers (b) The curve ¯Y consists of two irreducible components, of genus 2 and 1, which intersect in a unique singular point. (c) The curve ¯Y has three irreducible components W1,W2,W3 which are each smooth of genus 1. There are two singular points where W1 (resp. W3) intersects W2. (d) There are two irreducible components W1,W2 of genus 0 and 1, respectively, and three singular points where W1 and W2 intersect. (e) There are three irreducible components W1,W2,W3, of genus 0, 0 and 1, respec- tively, and 4 singular points. The components W1, W2 meet in three of these singular points, while W2 and W3 meet in the fourth. 4.2 The conductor exponent in the tame case In the tame case, there are no useful lower bounds for the conductor exponent. In particular, Y may have good reduction in which case we have fp = 0. Also, unlike for p = 3, nothing is gained by assuming that the ground field K is totally unramified. Still, some useful restrictions on fp can be proved (see Theorem 4.4 below). We start by recalling a well known criterion for good reduction, see e.g. [8], § 7. Let Y : y3 = f (x) = a4x4 + a3x3 + a2x2 + a1x + a0 4 x,a−1 be a Picard curve over K. Replacing (x,y) by (a−1 4 ) and multiplying both sides 4, we may assume that a4 = 1. Let D ( f ) ∈ K× denote of the defining equation by a3 the discriminant of f . (Since we assume that f is separable, we have D ( f ) 6= 0.) After replacing (x,y) by (u−3x,u−4y) and multiplying by u12 on both sides, for a suitable u ∈ K×, we may further assume that all coefficients ai ∈ OK are integral. In particular, it follows that D ( f ) ∈ OK. Since D (u12 f (u−3x)) = u36D ( f ), by the right choice of u, we may assume that 0 ≤ ordp(D ( f )) < 36. (29) Lemma 4.3. Assume that the Picard curve Y is given by a minimal equation over OK, as above. Then Y has good reduction if and only if D ( f ) ∈ O×K . Proof. See [8], Lemma 7.13. ⊓⊔ Note that the forwards direction of Lemma 4.3 also follows from Theorem 4.2. Here is what we can say in general about the conductor exponent. Theorem 4.4. Let K be as before, with p 6= 3, and Y a Picard curve over K. Let fp denote the conductor exponent for Y , relative to the prime ideal p of OK. Then the following holds. (a) If fp = 0 then the stable reduction of Y is as in Case (a), (b), or (c) of Theorem 4.2. Furthermore, the splitting field L0/K of f is unramified at p. Picard curves with small conductor 19 (b) If p = 2 then fp 6= 1. (c) If p ≥ 5 then fp ∈ {0,2,4,6}. Proof. We start be proving Statement (a). Note that fp = 0 if and only if d = 0 et( ¯Y 0, Qℓ) = 6. The second condition, together with the discussion after and dim H1 Proposition 2.3, implies that g ( ¯Y 0). Statement (a) now follows immediately from Theorem 4.4. Claim: The integer e , defined in Proposition 2.3, is even. The discussion follow- ing Proposition 2.3 implies that fp is odd if and only if dim H1(D ¯Y 0) is odd. The case distinction in Theorem 4.2 implies that dim H1(D ¯Y 0 ) is at most 2. Therefore to prove the claim, it suffices to show that g ( ¯Y 0) = dim H1(D ¯Y 0) 6= 1. We prove this in the case that ¯Y is as in (d) of Theorem 4.2. The argument in the case that ¯Y is as in (e) is very similar. In the other cases there is nothing to prove. acts on the component ¯X2 to which ¥ Assume that ¯Y is as in (d) of Theorem 4.2. Then ¯X is as (d) of Lemma 4.1 and ¯f maps Wi to ¯Xi. Since ¥ is K-rational, the monodromy group G fixes it. It follows that G specializes. (This is similar to the argument in the proof of [2], Lemma 5.4.) Since there is exactly one other branch point specializing to ¯X2, this point is fixed by G , as well. Similarly, G fixes the unique singularity. Since G fixes at least 3 points on the genus-0 curve ¯X2, it acts trivially on ¯X2. Equation (2) implies that the action of G on ¯Y descends to ¯X. It follows that G acts on W2 via a subgroup of G. We conclude that G either fixes the three singularities of ¯Y or cyclically permutes them. It follows that g ( ¯Y 0) is 2 or 0. This proves the claim. 6= 0 then d ≥ 2. For Statement (c) recall that L/K is at most tamely ramified for p ≥ 5. It follows is bounded by 2g(Y ) = 6. Statement (c) now ⊓⊔ Remark 4.5. (a) The condition fp = 0 in Theorem 4.4.(a) is equivalent to the condi- tion that the Jacobian variety of Y has good reduction over K. This is the case if and only if Y has stable reduction already over K, and the graph of components D ¯Y is a tree. This observation is similar to the statement of Lemma 2.4. Assume that p = 2. Using Equation (6) one shows that if d that d = 0, and hence that fp = e follows from the claim. Therefore Statement (b) follows from the claim. (b) For p = 2 the conductor exponent f2 may be odd. An example can be found in Example 5.5. (c) The bound on fp for p = 5,7 in Theorem 4.4.(c) is slightly sharper than the bound for fp for general abelian varieties of dimension 3 from [3], Thm. 6.2. The reason is that Brumer and Kramer obtain an upper bound for d . For Picard curves and p = 5,7 we have d = 0, whereas this is not necessarily the case for general curves of genus 3. For p = 2 the result of [3] yields the upper bound fp ≤ 28. Distinguishing the possibilities for the stable reduction and combining our arguments with those of [3] it might be possible to improve the bound in this case. Example 4.6. Consider the Picard curve Y : y3 = f (x) = x4 + 14x2 + 72x− 41 20 Borner, Bouw, Wewers over K := Qnr tion type is as in Case (b) of Theorem 4.2. Therefore, f5 = 0. 5 . We claim that Y has semistable reduction over K, and that the reduc- We will argue in a similar way as in § 3.3, see in particular Example 3.9, see also [2], § 6 and § 7. The first observation is that f = x4 + 14x2 + 72x− 41 ≡ (x + 3)2(x2 + 4x + 1) (mod 5). (30) By Hensel's Lemma, f has two distinct roots a 1,a 2 ∈ OK with a 2 i + 4a i + 1 ≡ 0 (mod 5). The other two roots of f are congruent to −3 (mod 5). Substituting x = −58 + 53x1 into f , we see that f ≡ 56(3x2 1 + 4x1 + 2) (mod 57). (31) i + 4b i + 2 ≡ 0 (mod 5). So f splits over K. It follows that f has two more roots a 3,a 4 ∈ K of the form a i = −58 + 53b i, with b i ∈ OK and 3b 2 Let (X , D) be the stably marked model of (X,D), where X = P1 K and D = ,a 1, . . . ,a 4}. The calculation of the a i above show that X is the OK-model of {¥ X corresponding to the set of valuations {v0,v1}, where v0 (resp. v1) is the Gauss valuation on K(x) with respect to the parameter x (resp. to x1). Let Y be the nor- malization of X in the function field of Y . We claim that the special fiber ¯Y of Y consists of two irreducible components W0,W1 of geometric genus 2 and 1, re- spectively. By the same argument as in Example 3.9, this already implies that Y is semistable and that the special fiber is as in Case (b) of Theorem 4.2. To prove the claim it suffices to find generic equations for W0 and W01. For W0 we just have to reduce the original equation for Y modulo 5. By (30) we obtain W0 : ¯y3 = ( ¯x + 3)2( ¯x2 + 4 ¯x + 1), which shows that g(W0) = 2. For W1 we write f as a polynomial in x1, substitute y = 52w, divide by 56 and reduce modulo 5. By (31) we obtain W1 : ¯w3 = 3 ¯x2 1 + 4 ¯x1 + 2, which shows that g(W1) = 1. Now everything is proved. ⊓⊔ Remark 4.7. The example above is again rather special, since f5 = 0 even though Y has bad reduction at p = 5. (See also Definition 5.4). 5 Searching for Picard curves over Q with small conductor In this last section we briefly address the problem of constructing Picard curves with small conductor. We think this is an interesting problem which deserves further investigation. The main background result here is the Shafarevic conjecture (which is a theorem due to Faltings). We use this theorem via the following corollary. Theorem 5.1 (Faltings). Fix a number field K and an integer g ≥ 2. Picard curves with small conductor 21 (a) For any finite set S of finite places of K there exist at most a finite number of isomorphism classes of smooth projective curves of genus g over K with good reduction outside S. (b) For any constant N > 0 there exists at most a finite number of isomorphism classes of curves of genus g over K with conductor ≤ N. Proof. Satz 6 in [6] states that there are at most a finite number of d-polarized abelian varieties of dimension g over K with good reduction outside S, for fixed K, g, d and S. Statements (a) and (b) follow from this. For (a), one simply uses Torelli's theorem (see [6], p. 365, Korollar 1). To deduce (b) we use that the conductor of a curve Y is the same as the conductor of its Jacobian, and that an abelian variety over K has bad reduction at a finite place p of K if and only if fp = 0 (see e.g. [23], Theorem 1). ⊓⊔ Unfortunately, no effective proof of Theorem 5.1 is known in general.2 However, for some special classes of curves effective proofs are known, see e.g. [10]. The problem we wish to discuss here is whether the statement of Theorem 5.1 can be made computable in the case of Picard curves. More precisely: given a finite set S of rational primes (or a bound N > 0), can we compute the finite set of curves with good reduction outside S (resp. with conductor ≤ N)? Note that this is not equivalent to (and may be much easier than) having an effective proof of Theorem 5.1 for Picard curves. For the first problem, the answer is known to be affirmative: Proposition 5.2. There exists an algorithm which, given as input a number field K and finite set S of finite places of K, computes the set of isomorphism classes of all Picard curves Y /K with good reduction outside S. Proof. This is an adaption to Picard curves of the algorithm given by Smart for hy- perelliptic curves, see [25] and [13]. The idea is that it suffices to determine the finite set of equivalence classes of binary forms of degree 4 over K whose discriminant is an S-unit (corresponding to the polynomial f (x)). The latter problem can be reduced to solving an S-unit equation, for which effective algorithms are known. ⊓⊔ Example 5.3. Let K = Q and S = {3}. Then there are precisely 63 isomorphism classes of Picard curves over Q with good reduction outside S. See [13]. For example, the curve Y : y3 = f (x) = x4 − 3x3 − 24x2− x has good reduction outside S = {3} (the discriminant of f is D ( f ) = 310). The stable reduction ¯Y of Y at p = 3 is as in Case (c) of Theorem 3.2, the exponent conductor is f3 = 10 (see [4], Appendix A1.1). This is the lowest value for the conductor which occurs for the curves in the list of [13]. The conductor exponents of all 63 Picard curves from [13] have been computed in [4], Appendix A1.2. From 2 The precise meaning of an effective proof is that it provides an explicitly computable bound on the height of the curve or abelian variety in question. 22 Borner, Bouw, Wewers this calculation it follows that the conductor exponent f3 only takes the values f3 = 10,11,12,13,15,17,19,21. The upper bound on the conductor exponent from abelian varieties of genus 3 from [3], Theorem 6.2 yields f3 ≤ 21. The result stated above therefore implies that this bound is also obtained for Picard curves. Unfortunately we do not know any algorithm for solving (b), i.e. for finding all Picard curves with bounded conductor. The reason that the method for (a) does not solve (b) is the existence of exceptional primes. Definition 5.4. Let Y be a Picard curve over Q and p a prime number. Then p is called exceptional with respect to Y if Y has bad reduction at p and f p = 0. Exceptional primes are rather rare. It can easily be shown, using the arguments from this paper, that if p is a exceptional prime for Y then the splitting field of the polynomial f is unramified at p, and Example 5.5. We consider the Picard curve over Q ordp(D ( f )) ∈ {6,12}. Y : y3 = f (x) = x4 + 14x2 + 72x− 41. The discriminant of f is D ( f ) = −2103456. So Y has good reduction outside S = {2,3,5}. We have shown in Example 4.6 that f5 = 0, i.e. that 5 is an exceptional prime. Using the methods of [2] and [18] one can prove that f2 = 19 and f3 = 13 (see e.g. this SageMathCloud worksheet: http://tinyurl.com/hp3qzmo, [20]). All in all, the conductor of Y is NY = 219313 = 835884417024. Although S is small and p = 5 is an exceptional prime, NY is relatively large. We have tried but were not able to find a similar example with exceptional primes and a significantly smaller conductor. Nevertheless, the fact that exceptional primes exist means that we cannot easily bound the size of the set S while searching for Picard curves with bounded conductor. Here is an example of a Picard curve with a relatively small conductor. Example 5.6. Consider the Picard curve Y /Q : y3 = f (x) = x4 − 1. The discriminant of f is D ( f ) = −256 = −28. It follows that Y has good reduction outside S = {2,3}. By [4], § 5.1.3, we have f2 = 6 and f3 = 6. Therefore, NY = 2636 = 46656. Picard curves with small conductor 23 The first named author has made an extensive search for Picard curves over Q with small conductor ([4], § 5.3). Among all computed examples, the curve Y was the one with the smallest conductor. A remarkable property of the curve Y is that for every (rational) prime p it admits a map to P1 of order prime to p, which becomes Galois over an extension: besides the degree-3 map f given by (x,y) 7→ x, we have the map (x,y) 7→ y, which has degree 4. In fact, the full automorphism group of Y has order 48, and is maximal in the sense that Y / AutC(Y ) is a projective line, and the natural cover is branched at three points. It is instructive to compare the above example with the curve Y ′ : y3 = x4 + 1. This is a twist of Y . The curve Y and Y ′ become isomorphic over Q[i], yet have different conductors. In fact, NY′ = 21636, see [4], § 5.1.2. We propose to study the following problem. Problem 5.7. Prove that the curve from Example 5.6 is the only Picard curve (up to isomorphism) with conductor NY ≤ 46656, or find explicit counterexamples. Proposition 5.2 and our main results (Theorem 3.6 and Theorem 4.4) suggest the following strategy for construction Picard curves with small conductor and thereby finding counterexamples. If we ignore the possibility of exceptional primes, a Picard curve with conductor ≤ 2636 must have good reduction outside S, where S is one of the following sets: • {2,3, p}, • {3, p}, To find all such curves looks challenging but within reach. It should also be very useful to take into account the local restrictions on the polynomial f imposed by our results on curves with a specific value for f p. On the other hand, without an effective proof of Theorem 5.1 (b) for Picard curves, it is not clear at the moment how one could actually prove that the curve from Example 5.6 (or any other curve we may find) has minimal conductor. p ≤ 13, p ≤ 23. References 1. Bouw, I.I., Wewers, bad of S.: Euler Semistable computation https://www.uni-ulm.de/fileadmin/website_uni_ulm/mawi.inst.100/mitarbeiter/wewers/course_notes.pdf. Notes for a minicourse at ICERM factors of reduction of L-functions. curves and URL 24 Borner, Bouw, Wewers 2. Bouw, I.I., Wewers, S.: Computing L-functions and semistable reduction of superelliptic curves. Glasgow Math. J. 59, 77 -- 108 (2017) 3. Brumer, A., Kramer, K.: The conductor of an abelian variety. Compositio Math. 92(2), 227 -- 248 (1994) 4. Borner, M.: L-Functions of curves of genus ≥ 3. Ph.D. thesis, Universitat Ulm (2016). URL 5. Deligne, P., Mumford, D.: The irreducibility of the space of curves of given genus. Publ. http://dx.doi.org/10.18725/OPARU-4137 Math. IHES 36, 75 -- 109 (1969) 6. Faltings, G.: Endlichkeitssatze fur abelsche Varietaten uber Zahlkorpern. Inventiones Math. 73, 349 -- 366 (1983) 7. Harris, J., Mumford, D.: On the Kodaira dimension of the moduli space of curves. Invent. Math. 67, 23 -- 86 (1982) 8. Holzapfel, R.P.: The Ball and Some Hilbert Problems. Birkhauser (1995) 9. Koike, K., Weng, A.: Construction of CM Picard curves. Math. Comp. 74(249), 499 -- 518 (2005) 10. von Kanel, R.: An effective proof of the hyperelliptic Shafarevich conjecture. Journal de Th´eorie des Nombres de Bordeaux 26(2), 507 -- 530 (2014) 11. Liu, Q.: Conducteur et discriminant minimal de courbes de genre 2. Compositio Math. 94(1), 51 -- 79 (1994) 12. Liu, Q.: Algebraic geometry and arithmetic curves. Oxford University Press (2006) 13. Malmskog, B., Rasmussen, C.: Picard curves over Q with good reduction away from 3. arXiv:1407.7892 14. Picard, E.: Sur des fonctions de deux variables ind´ependantes analogues aux fonctions modu- laires. Acta Math. 2(1) (1883) 15. Pries, R.: Wildly ramified covers with large genus. J. Number Theory 119(2), 194 -- 209 (2006) 16. Quine, J.R.: Jacobian of the Picard curve. In: Extremal Riemann surfaces (San Francisco, CA, 1995), Contemp. Math., vol. 201, pp. 33 -- 41. Amer. Math. Soc., Providence, RI (1997) 17. Raynaud, M.: Sp´ecialisation des revetements en caract´eristique p > 0. Annales scientifiques de l'Ecole normale sup´erieure 32(1), 87 -- 126 (1999) 18. Ruth, J., Wewers, S.: Semistable reduction of superelliptic curves of degree p. In preparation 19. Ruth, J.: Models of curves and valuations. Ph.D. thesis, Universitat Ulm (2014). URL http://dx.doi.org/10.18725/OPARU-3275 20. SageMath, I.: SageMathCloud Online Computational Mathematics (2016). https://cloud.sagemath.com/ 21. Serre, J.P.: Corps locaux. Hermann, Paris (1968). Troisi`eme ´edition, Publications de l'Universit´e de Nancago, No. VIII 22. Serre, J.P.: Facteurs locaux des fonctions zeta des vari´et´es alg´ebriques (d´efinitions et conjec- tures). S´eminaire Delange-Pisot-Poitou (Th´eorie des Nombres) 19(2), 1 -- 15 (1969) 23. Serre, J.P., Tate, J.: Good reduction of abelian varieties. Annals of Math. 88(3), 492 -- 517 (1968) 24. Silverman, J.H.: Advanced topics in the arithmetic of elliptic curves. No. 151 in Graduate Text in Math. Springer-Verlag (1994) 25. Smart, N.P.: S-unit equations, binary forms and curves of genus 2. Proceedings of the London Mathematical Society 75(2), 271 -- 307 (1997) 26. Wewers, S.: Deformation of tame admissible covers of curves. In: H. Volklein (ed.) Aspects of Galois theory, no. 256 in LMS Lecture Note Series, pp. 239 -- 282 (1999)
1702.03623	2	1702	2017-05-23T06:18:41	Connections on parahoric torsors over curves	[ "math.AG", "math.DG" ]	We define parahoric $\cG$--torsors for certain Bruhat--Tits group scheme $\cG$ on a smooth complex projective curve $X$ when the weights are real, and also define connections on them. We prove that a $\cG$--torsor is given by a homomorphism from $\pi_1(X\setminus D)$ to a maximal compact subgroup of $G$, where $D\, \subset\, X$ is the parabolic divisor, if and only if the torsor is polystable.	math.AG	math	CONNECTIONS ON PARAHORIC TORSORS OVER CURVES VIKRAMAN BALAJI, INDRANIL BISWAS, AND YASHONIDHI PANDEY ABSTRACT. We define parahoric G -- torsors for certain Bruhat -- Tits group scheme G on a smooth complex projective curve X when the weights are real, and also define connections on them. We prove that a G -- torsor is given by a homomorphism from π1(X \ D) to a max- imal compact subgroup of G, where the finite subset D ⊂ X is the parabolic divisor, if and only if the G -- torsor is polystable. CONTENTS 1. Introduction 1.1. Origins 2. Preliminaries 2.1. Apartment of G 3. Parahoric group scheme and torsors 3.1. Invariant direct image 3.2. Parahoric torsors 3.3. Rational weights 3.4. Rational weights and parahoric G -- torsors as (Γ,G) -- bundles 4. Change of weights under a homomorphism 4.1. The local homomorphism problem 4.2. Facets of a homomorphism 5. Associated Constructions 5.1. The construction 5.2. Tensor product of parabolic vector bundles 5.3. Lie algebra bundle of a G -torsor 6. Semistability and stability of torsors 6.1. First definition 6.2. Second definition 7. (Semi)stability and polystability under variation of weights 7.1. Facets of a quasi -- parahoric torsor 8. Connections on parahoric G -- torsors 2010 Mathematics Subject Classification. 14F22, 14D23, 14D20. Key words and phrases. Bruhat -- Tits group scheme; parahoric torsor; connection; polystability. 1 2 3 4 5 5 5 6 7 8 9 9 9 11 11 11 12 12 12 13 14 14 17 2 V. BALAJI, I. BISWAS, AND Y. PANDEY 8.1. D X -- modules 8.2. Logarithmic connections on curves 8.3. Connection on parabolic vector bundles 8.4. Lie connection on a principal G -- bundle 8.5. Connection on parahoric G -- torsors 8.6. A Tannakian description of connection 9. Connections on (Γ,G) -- bundles and rational weights 10. Flat unitary connections on parahoric torsors and stability 10.1. Polystable parahoric torsors 10.2. Polystable parahoric torsors from unitary representations 10.3. Polystable parahoric torsors and unitary representations 10.4. The reductive case Acknowledgements References 17 17 18 18 19 19 20 22 22 23 25 26 27 27 1. INTRODUCTION Let X be an irreducible smooth projective curve defined over the field of complex num- bers. Let G be a simple and simply connected affine algebraic group defined over C. Fix a finite subset D ⊂ X . Let G be a Bruhat -- Tits group scheme with parabolic points at D (see Definition 3.3). In [BS], an analogue of the Mehta -- Seshadri theorem in [MS] was proved relating stable parahoric torsors under Bruhat -- Tits group schemes with irreducible homomorphisms of certain Fuchsian groups into a maximal compact subgroup of G. This was done under the assumption that the parabolic weights are rational, or equivalently, the fixed points of the Fuchsian group are all elliptic. Recall that in [BS] it is shown that if the weights are chosen rational then one can recover the G -- torsors as invariant direct images of orbifold principal bundles with respect to suitable ramified covers of X ramified over the parabolic points. An obstruction to cover real weights in the setting of parahoric torsors is that the classical Bruhat -- Tits group scheme is defined on spectra of discrete valuation rings while the phe- nomenon of parabolic bundles with real weights naturally lies in the setting of an analytic neighborhood of the origin. Further, the notion of invariant direct image fails to generalize directly to the analytic setting when the covering map is no longer algebraic. We address this issue by working with a natural analogue of Deligne extensions to the parahoric torsor setting; to a pair (E , bd ) of a principal G -- bundle E equipped with a flat connection bd on the punctured disc, we give a canonical extension to a torsor under Bruhat -- Tits group scheme on the compact Riemann surface (see Section 10.2). We use this to construct the parahoric torsor associated to a representation of the fundamental group of the punctured Riemann surface. CONNECTIONS ON PARAHORIC TORSORS OVER CURVES 3 Another obstruction in case of real weights is the notion of stability for parahoric tor- sors. Filling the lacuna in [BS], we have a definition of stability for parahoric torsors (6.2) which covers real weights as well. Then, following the approach in the paper of Mehta -- Seshadri, [MS, p. 217], in Section 7 we develop the theme of variation of weights in the setting of parahoric torsors, where the notions of (semi)stability for the case of real weights are shown to coincide for "nearby" rational weights. Although the broad lines are the same this generalization is not entirely straight-forward. We then go on to define the notion of a connection on a G -- torsor over a smooth pro- jective curve over C and prove the analogue of the Donaldson -- Uhlenbeck -- Yau correspon- dence in the parahoric setting when the weights are real. The basic idea is to use the Tan- nakian formalism and the argument reduces to the case of the associated parabolic Lie algebra bundle. The more general reductive group case is then an easy generalization. 1.1. Origins. In the early 60s, Mumford defined the notion of stability for vector bundles on curves as a tool to get Hausdorff moduli spaces; using geometric invariant theory, he then constructed the moduli space of stable vector bundles. In ([NS]) Narasimhan and Seshadri gave an alternative characterization of stability using flat connections; more pre- cisely, they proved that a holomorphic vector bundle E on a compact Riemann surface is stable if and only if E arises from an irreducible projective unitary representations of the fundamental group of the Riemann surface. This correspondence between flat pro- jective unitary connections and stable vector bundles has been generalized in several di- rections. A. Ramanathan ([Ra]) extended the correspondence to the case of holomorphic principal G -- bundles, where G is a complex reductive affine algebraic group [Ra] On the other hand, Mehta and Seshadri ([MS]) generalized the Narasimhan -- Seshadri construction to include unitary logarithmic connections, or equivalently, to classify irreducible unitary representations of general Fuchsian groups with fixed conjugacy classes. These logarith- mic connections are those which have regular singularity at finitely many points and were already apparent in the early work of Weil ([We]); their importance was emphasized by Deligne in ([De]). The objects replacing stable bundles in the Mehta -- Seshadri correspon- dence are stable parabolic vector bundles. Following Donaldson's reinterpretation of the Narasimhan -- Seshadri correspondence, Biquard ([Biq]) gave a differential geometric inter- pretation of this Mehta -- Seshadri correspondence. It is a very natural problem to generalize the Mehta -- Seshadri correspondence from the setting of parabolic vector bundles to that of principal G -- bundles, where G is a complex reductive affine algebraic group. On the side of representations, the objects were easy to define; they were homomorphisms of Fuchsian groups taking values in a maximal com- pact subgroup of G such that for each puncture of the Riemann surface the associated conjugacy class in the fundamental group of the surface is mapped to a fixed conjugacy class of the maximal compact subgroup of G. Similarly, on the side of connections, the corresponding objects were quite well-understood since the work of Deligne. The central problem was to generalize the notion of a stable parabolic vector bundle to the setting of principal G -- bundles. From a Tannakian perspective ([BBN1]) it became apparent that a naive generalization in terms parabolic G -- bundles, i.e., principal G bundles with parabolic structures, was insufficient for setting up a comprehensive analogue of the Mehta -- Seshadri correspondence. 4 V. BALAJI, I. BISWAS, AND Y. PANDEY At a technical level, it was not clear what the correct notion of weight should be in the general setting. A breakthrough came in the work of Boalch ([Bo]) and Balaji -- Seshadri ([BS]). These papers introduced the correct and very natural notion of weight, namely, as a point in the affine apartment of G. As a consequence of this point view, the reason for the inadequacy of parabolic G -- bundles also became clear. It was realized that instead of parabolic G -- bundles one should consider torsors or principal homogeneous spaces under parahoric group schemes in the sense of Bruhat and Tits. Balaji and Seshadri ([BS]) ex- tended the Mehta -- Seshadri theorem to the case of parahoric torsors with rational weights. Prior to these works, there were at least two partial approaches to generalize the Mehta -- Seshadri theorem, both around the turn of the millennium. The approach in ([BBN1]) was Tannakian in spirit and followed the method of Nori [No]; this Tannakian approach identi- fied the basic problem, namely that the object associated to a representation of a Fuchsian group into the maximal compact of G, such that for each puncture of the Riemann surface the associated conjugacy class in the fundamental group of the surface is mapped to a fixed conjugacy class of the maximal compact subgroup of G, in general will not be a principal G -- bundle on the Riemann surface. The approach in ([TW]) again gave a partial solution to the problem; in the language of Weyl alcoves, the solution was for weights in the interior of the Weyl alcove which corresponds to the subclass of parabolic G -- bundles. Parahoric group schemes G and G -- torsors over smooth projective curves were first de- fined and studied by Pappas -- Rapoport [PR1]. Subsequently, in [PR2], they made several conjectures on the moduli stacks of G -- torsors. Most of these conjectures were answered by Heinloth ([He1]) providing a precise setting for the study of these moduli stacks. The paper of Seshadri ([Se2]), takes up the question of the analogue of the Mehta -- Seshadri the- orem; the main emphasis in ([Se2]) was again was to point out that for a solution to the problem of obtaining analogues of the Mehta -- Seshadri theorem, one has to go beyond the category of principal G -- bundles. Evidence to the role of Bruhat -- Tits theory was also given in a few illustrative examples. The paper of Boalch ([Bo]) studies logarithmic connections on G -- bundles and the notion of parahoric torsors comes along with the first appearance of the notion of weights for these torsors. Around the same time and independently in [BS], a similar notion of weights was defined towards providing a satisfactory analogue of the Mehta -- Seshadri theorem in the general setting for semisimple groups G, thereby complet- ing the broad picture outlined in [Se2]. The notion of "invariant direct images" of torsors in [BS] plays a key role analogous to the one in [MS] for the case of vector bundles. 2. PRELIMINARIES In this section we collect together some standard notions and notation that will be used throughout this paper. See [BT1], [BT2], [BS], [He1] for this section and the next one. The base field will always be C. Define A := C[[t ]] and K := C((t )) , (2.1) where t denote a uniformizing parameter. Let G be a semisimple simply connected affine algebraic group defined over C. The Lie algebra of G will be denoted by g. Fix a maxi- mal torus T ⊂ G; let Y (T ) = Hom(Gm, T ) be the group of all holomorphic one -- parameter subgroups of T . CONNECTIONS ON PARAHORIC TORSORS OVER CURVES 5 2.1. Apartment of G. For each maximal torus T of G, we have the standard affine apart- T . It is an affine space under Y (T )⊗Z R. In general, there is no origin in the apart- ment A ment (cf. [BT1]). But for purposes of this paper, we shall identify A T with Y (T )⊗Z R (see [BS, § 2]). Let V be a real affine space. A function f : V −→ R is said to be an affine functional if f (r x + (1− r )y) = r f (x)+ (1− r )f (y) for all x , y ∈ V and r ∈ R. Thus, for a root α of G and an integer n ∈ Z there is the affine functional α+ n : A T −→ R , x 7−→ α(x − 0)+ n . We note that these are called the affine roots of G. The zero locus of α+n will be denoted by Hα+n; it is called an affine wall. The set of affine walls is known to be locally finite, meaning T intersects only finitely many affine walls. For any point x ∈ A T , any compact subset of A let Zx denote the set of affine functionals vanishing on x. For an integer n ≥ 0, define Hn = {x ∈ A Zx = n} , which is the set of all points where n of the affine functionals vanish. A facet Ω of A T is defined to be a connected component of Hn for some n. The dimension of a facet is its dimension as a real manifold. We then have a decomposition of the apartment A T = G n Hn . (2.2) Although, as mentioned above, almost always Θ will be a point of A T , sometimes Θ will also be a facet of A T . 3. PARAHORIC GROUP SCHEME AND TORSORS 3.1. Invariant direct image. The base field is C. Let p : W −→ U be a finite flat surjective morphism of normal, integral Noetherian schemes which is Galois. So the Galois group Gal(p), which we will denote by Γ, acts on W with U = W /Γ being the quotient. Such a morphism p is called a Galois covering with Galois group Γ. Let G be an affine group scheme over W . For the above Galois covering map p, the direct image p ∗ G is defined as follows: for each U -- scheme S, set (p G )(S) = HomW (S ×U W, G ) ; ∗ (3.1) this is representable by a group scheme [BLR, Theorem 4 and Proposition 6]. Assume that the group scheme G is equipped with an action of the Galois group Γ that lifts the ac- tion of Γ on W ; in particular, the "multiplication map" and the "inverse map" on G are Γ -- equivariant. Such a G will be called a Γ -- group scheme over W . There is a left action of Γ on S ×U W induced by the action of Γ on W . This and the left action of Γ on G together induce the following right action of Γ on (p G )(S): ∗ (f .γ)([s, w ]) := γ−1.f (γ.[s, w ]) , Consider the fixed point subscheme under the above action of Γ on p G . The general re- sults on fixed point subschemes given in [Ed, Section 3] can be applied to our situation [s, w ] ∈ S ×U W , γ ∈ Γ . ∗ (3.2) 6 V. BALAJI, I. BISWAS, AND Y. PANDEY since the characteristic is 0. Consequently, a canonically defined smooth closed X -- subgroup scheme G ) G is representable because p G := (p p Γ ∗ ∗ Γ ∗ G ⊂ p G is representable. ∗ is obtained. This p Γ ∗ Definition 3.1 (Invariant direct image). Let p : W −→ U be as above, and let Γ = Gal(W /U ). Let G be a smooth affine Γ -- group scheme over W . Define the invariant direct image of G to be G )Γ , ∗ (3.3) More generally, if E is any affine scheme over W with a lift of the Γ -- action on W , then (G ) := (p G )(S) = G (S ×U W )Γ for any U -- scheme S. p Γ ∗ so (p Γ ∗ define the invariant direct image of E to be p Γ ∗ E := (p ∗ E )Γ . 3.2. Parahoric torsors. Notation of Section 2.1 will be followed. Let R = R(T, G) denote the root system of G (cf. [Sp, p. 125]). Thus for every r ∈ R, there is the root homomorphism ur : Ga −→ G [Sp, Proposition 8.1.1]. For any non-empty subset Θ ⊂ AT , the parahoric subgroup P Θ := 〈T (A), {ur (t mr (Θ) A)}r∈R〉 . is the subgroup generated by T (A) and {ur (t mr (Θ) A)}r∈R, where P Θ ⊂ G(K ) (3.4) mr = mr (Θ) = −[infθ∈Θ(θ,r )] , and A is defined in (2.1) [BS, Page 8]. Moreover, by [BT2, Section 1.7] we have an affine flat Θ −→ Spec(A) corresponding to Θ. The set of K -- valued (respec- smooth group scheme G tively, A -- valued) points of G Θ). The group scheme Θ is uniquely determined by its A -- valued points. Here we will often take Θ to be just a G Θ is identified with G(K ) (respectively, P point of A T . Remark 3.2. We remark that the notion of a parahoric subgroup is defined in the greatest generality in the basic papers of Bruhat and Tits. For our purposes, the definition given above is sufficient. Let X be a smooth complex projective curve. We fix once for all a nonempty finite set of closed points D = {x j }m j=1 ⊂ X . (3.5) These will play the role of parabolic points. Let A j := (cid:129)O be the complete discrete val- uation ring with function field K j ≃ C((t )) and residue field C, obtained by completing the local rings OX ,x j . We shall denote Spec (A j ) by D j . Definition 3.3. Let G be a flat, affine group scheme on X of finite type. We call G a Bruhat -- Tits group scheme with parabolic points D, if X ,x j (1) restricted to X \ D, it is isomorphic to the split group scheme G × (X \ D), and (2) GD j −→ D j is a Bruhat -- Tits group scheme for each j . We shall denote G by GΩ, where Ω = {Ωj }m building with GD j corresponding to Ωj . j=1 is a collection of facets of the Bruhat -- Tits CONNECTIONS ON PARAHORIC TORSORS OVER CURVES 7 G Θ is identified with G(K ) (respectively, P By the general theory due to Bruhat and Tits, one has an affine flat smooth group scheme Θ on Spec(A) corresponding to Θ. The set of K -- valued (respectively, A -- valued) points Θ is uniquely deter- of G mined by its A -- valued points. These notions were defined and the moduli stack of G Θ -- torsors studied extensively in the papers of Pappas -- Rapoport (cf. [PR1], [PR2]) and Hein- loth ([He1]). To construct one on the whole of X one can proceed as in [BS, 5.2]. Existence of such group schemes also follows from the invariant direct images constructed above (see also [BS, Theorem 5.2.7]). Θ). The group scheme G Definition 3.4 ([BS, Section 6]). A quasi -- parahoric torsor E is a G Ω,X -- torsor on X . Definition 3.5 ([BS, Section 6], [Bo]). A parahoric torsor is a pair (E , θ) consisting of (1) a G Ω,X -- torsor E −→ X , and (2) weights, meaning elements θ = {θi }m i=1 ∈ (Y (T )⊗ R)m in the interior of Ωi . Remark 3.6. The above notion of weight is the precise analogue of the classical weight for a parabolic vector bundle with multiplicity (cf. [MS, page 211, Definition 1.5]). Remark 3.7. It should be noted that as in [BS], the theory of Bruhat -- Tits group schemes that are used here assumes that the group G is semisimple and simply connected. On the other hand, for the case of GL(V ), which satisfies neither of these conditions, the parabolic bundles is classical (cf. [Se1], [MS], [Bis]). In [BS, Example 2.3.4] and [BS, Remark 6.1.5], it is noted that the torsors under Bruhat -- Tits group schemes for GL(V ) are same as the par- abolic vector bundles. Let us spell this out for the convenience of the reader. Let G L (V ) be a Bruhat -- Tits group scheme on Spec(A) with generic fiber GL(V ). Fix a maximal torus T (V ) ⊂ GL(V ). Let E be a G L (V ) -- torsor, and let θ(V ) ∈ Al (T (V ))R be a weight as a point in the so-called Weyl alcove (see [BS, p. 9]). Then, the associated vector bundle gets a canoni- cal parabolic structure with quasi -- parabolic type determined by the group scheme G L (V ) while the parabolic weights are given by θ(V ). We observe that in the case of GL(n, C), or SL(n, C), giving a point θ(V ) ∈ Al (T (V ))R in the Weyl alcove is equivalent to giving n -- tuples , such that every αi ≥ 0 and αi ≤ αi+1 , for 1 ≤ i ≤ n − 1; in the case of (α1, α2, ··· , αn ) ∈ R SL(n, C), the condition of the determinant translates to the further condition n nX i=1 αi ∈ Z . By a Weyl group conjugation, we can also arrange the αi 's in increasing order, i.e., 0 ≤ α1 ≤ α2 ≤ ··· ≤ αn < 1. Once the parahoric subgroup is chosen, this determines a flag type and hence we may even order the parabolic weights in a strictly increasing sequence. 3.3. Rational weights. Starting with a m -- tuple of weights θ ∈ (Y (T )⊗ Q)m, following the proof of the converse in [BS, Theorem 2.3.1], we get positive integers d1, d2, ··· , dm such that di · θi ∈ Y (T ). By choosing these di to be smallest with this property, we see that a choice of θ entails a choice of ramification index dx at each point x ∈ D. There exists a ramified Galois cover of curves p : Y −→ X which is • unramified over X \ D, and • the ramification index over x ∈ D is dx , if and only if exactly one of the following conditions hold: 8 V. BALAJI, I. BISWAS, AND Y. PANDEY (1) the genus of X is nonzero, (2) X = P1 and #D ≥ 3, (3) X = P1, #D = 2, and dx = dy , where D = {x, y}. (See [Na, p. 26, Proposition 1.2.12].) When we consider parahoric torsors as invariant direct images of (Γ,G) bundles, it will always be with respect to such a Galois cover. See also [Bis]. 3.4. Rational weights and parahoric G -- torsors as (Γ,G) -- bundles. Let (X , D) be as above, and let G −→ X be a Bruhat -- Tits group scheme over X . By [BS, Theorem 5.3.1], there exists a (possibly ramified) finite Galois cover p : Y −→ X branched over D, and a principal G -- bundle F over Y (cf. [BS, Notation 5.1.0.1]) equipped with a lift of the action of the Galois group Γ := Gal(p), such that the following statements hold: (1) Let FG := I somY (F,F ) be the twisting of the constant group scheme G −→ Y by F . The invariant direct image satisfies the condition p Γ ∗F G = G . (3.6) (2) Let DY := p−1(D) ⊂ Y denote the ramification points of the covering p. For each y ∈ DY , let Γy ⊂ Γ be the isotropy subgroup that fixes y. Let τy : Γy −→ Aut (Fy ) be the action of Γy on the fiber Fy . Its conjugacy class is called the local type of F at y; this local type will be denoted by [τy ]. By the type τ of F , we shall mean the set of conjugacy classes [τy ] of τy : τ := {[τy ] y ∈ DY }. Let M τ Y (Γ,G) denote the moduli stack of (Γ,G) bundles over Y of type τ, and let MX (G ) denote the moduli stack of G torsors on X . Then there is an isomorphism of moduli stacks αF : M τ (3.7) given by the (Γ,G) bundle F as follows: Denote by F op the left G -- bundle defined by g f := f g −1, where g ∈ G and f ∈ F . The above group scheme FG acts on the right of F op . The isomorphism in (3.7) is: Y (Γ,G) −→ MX (G ) , E 7−→ p Γ ∗(E ×Y ,G F op ) = p Γ ∗(I somY (E ,F )). The inverse of the map in (3.7) is given by E 7−→ p∗E ×p∗G F . (3.8) (3.9) (3) Let y ∈ DY , and x := p(y). Let Ny = Spec(B ), where B = (cid:129)OY ,y , and also Dx = Spec(A) with A = (cid:129)OX ,x . Let Uy denote the group of local Γy -- G automorphisms of FNy (cf. [BS, Definition 2.2.7]). Then by [BS, Proposition 5.1.2] we have For the equality in (3.10), we need the existence of (Γ,G) bundle F only locally on Ny and not on entire Y . GDx (A) = Uy . (3.10) A different (Γ,G) bundle F ′ of type τ will, in general, provide a different (cf. (3.7)) iso- morphism αF′ : M τ Y (Γ,G) −→ MX (G ). CONNECTIONS ON PARAHORIC TORSORS OVER CURVES 9 4. CHANGE OF WEIGHTS UNDER A HOMOMORPHISM 4.1. The local homomorphism problem. Consider the following problem. Let A be an ar- bitrary discrete valuation ring with quotient field K . Let G be semisimple and simply con- nected group. Let ρ : G −→ H be a homomorphism. Fix a maximal torus T ⊂ G. Then T for T . It should be emphasized that θ may not fix a weight θ in the affine apartment A be a rational point, meaning it may not lie in the image of Y (T )⊗Z Q. Fix a maximal torus TH ⊂ H such that ρ(T ) ⊂ TH . Via the identification between A T and Y (T )⊗Z R mentioned in Section 2.1, we have a linear map between the apartments. Let θH denote the image of θ under this map. We wish to con- struct, through ρ, a canonical homomorphism of group schemes over Spec (A): from Gθ corresponding to θ to GθH corresponding to θH . A T −→ A TH (4.1) TH 4.2. Facets of a homomorphism. Let ρ : G −→ H be a homomorphism. The affine roots of H give affine functionals on A (cf. Section 2). These functionals are defined over the corresponding rationals. By (4.1) it follows that the linear map of apartments A to ρ is also defined over the rationals. Indeed, it is induced by the algebraic map ρT : T −→ TH . Via A , associated to the affine roots of H, as affine functionals on A T . Take the union of these affine functionals with the T corresponding to the affine roots of G. Note that all these usual affine functionals on A functionals are defined over rationals. We shall call them as ρ -- functionals. Definition 4.1. An affine wall of ρ is the zero locus of a ρ -- functional. For any x ∈ A define Zx to be the set of all affine functionals on A , we view the above affine functionals on A T vanishing at x. Define T −→ A T −→ A T , TH TH TH A facet of ρ is a connected component of Hn for some n ≥ 0. Hn := {x ∈ A Zx = n} . Notice that for the identity homomorphism G −→ G, we have just got the usual facets. By the theory of buildings, one knows that the usual affine walls of G provide a decomposition of A T (cf. (2.2)). We claim that more generally, the facets of a homomorphism ρ : G −→ H provide an even finer decomposition of A T = G n Hn ; T −→ A T meets finitely many usual affine walls in A , the image of C being compact, meets finitely many affine walls of A here n varies over a finite set. To prove this claim, first note that the set of all ρ -- affine walls corresponding to ρ -- functionals form a locally finite set. Indeed, this follows because any compact set C of A T . Now under the map . Thus A C meets only finitely many walls of ρ. Proposition 4.2. Let ρ : G −→ H be a homomorphism of algebraic groups. Given a weight T lying in the same facet as θ such that ηH and θ ∈ A θH also lie in a common facet of A T , there exists a rational weight η ∈ A TH TH . TH Proof. It can be shown that if the element θ ∈ A T lies in a zero dimensional facet, then it must be a rational point. Indeed, this follows from that fact that we are looking at common 10 V. BALAJI, I. BISWAS, AND Y. PANDEY zero locus of a set of rational functionals. Taking contrapositive of the last statement, if θ is not rational then a ρ -- facet of θ cannot be zero dimensional. So we can find a rational weight η in it. Now by construction, both η and θ lie in the same facet of A a common facet of A . TH T , and also θH and ηH lie in (cid:3) Remark 4.3. Henceforth, when we say that a rational weight η is close to θ with respect to ρ : G −→ H, we mean a rational weight in the sense of Proposition 4.2. The constructions in the proof of Proposition 4.2 can be extended to the context of a finite set of representations {ρi : G −→ GL(Vi )}i≤n of G. Now further assume that η and θ are actually interior points of one ρ -- facet. This as- sumption implies that they define isomorphic Bruhat -- Tits group schemes with generic fiber G and also group schemes with generic fiber H. There are canonical homomorphisms between them by the following proposition. Proposition 4.4. Let ρ : G −→ H be given. Let θ ∈ A map of apartments A G θ −→ G Proof. Let us recall a characterization of parahoric groups when the weight η lies in A Theorem 2.3.1]. In this case, by (3.10), there exists a ramified Galois cover T be a weight with image θH under the . Then there is a canonical homomorphism of group schemes over Spec(A). T −→ A Q [BS, TH θH with Galois group say Γ, and a (Γ,G) bundle F −→ Spec(B ), such that p : Spec(B ) −→ Spec(A) G Ω(A) = G η(A) = Aut(Γ,G)(F ) . (4.2) Choose a rational weight η close to θ with respect to ρ. This gives an element ηH close to θH . Further, let be the (Γ, H) bundle obtained by extending structure group using ρ. Via the natural homomorphism Aut(Γ,G)(F ) −→ Aut(Γ,H)(F (H)) and the equalities G η(A) = Aut(Γ,G)(F ) and G ηH (A) = Aut(Γ,H)(F (H)), we obtain a map of parahoric groups F (H) := F ×G H ρK : G η (A) −→ G ηH (A) . Now the characterizing property of the Bruhat -- Tits group schemes is that they are étoffé, which means that any morphism at the level of group schemes over Spec(A) is determined completely by the A -- valued points alone (cf. [BT2, Definition 1.7.1]). Thus we get a morphism over Spec (A) of group schemes G η −→ G ηH extending (4.2). Since ηH is close to θH , it follows that G have an induced isomorphism of group schemes G ηH −→ G homomorphism ρθ : G . ηH θH θ −→ G θH (4.3) θH (A), and hence we (A) = G on Spec (A) which gives the (cid:3) To the best of our knowledge, Proposition 4.4 is not available in the papers of Bruhat and Tits. Alternatively it can be proved using the general framework of functoriality of buildings as in [La, Theorem 2.1.8] and [La, Theorem 2.2.1]. CONNECTIONS ON PARAHORIC TORSORS OVER CURVES 11 5. ASSOCIATED CONSTRUCTIONS As before, G is semisimple and simply connected. Let ρ : G −→ GL(V ) be a rational representation. Fix a maximal torus T ⊂ G, and also fix a maximal torus TV ⊂ GL(V ) con- taining ρ(T ). Let E be a parahoric G θ,X -- torsor on X with weights θ. group scheme G 5.1. The construction. As in (3.5), fix a finite subset D = {x j }m j −→ D j = Spec(A j ), 1 ≤ j ≤ m, we get a facet Ω ing B(G). Any weight θ ∈ A m induces a weight θV in A m T (V ) T j=1 ⊂ X . For each parahoric j in the Bruhat -- Tits build- (see Section 4.2). The group scheme G θ,X on X is obtained by gluing G , for every 1 ≤ j ≤ m, with θ j GX \D ≃ (X \ D)×G . Using the same gluing data via the representation ρ one immediately gets a Bruhat -- Tits group scheme G LθV ,X , and using locally Proposition 4.4, we see that ρ gives a natural global homomorphism of group schemes over X Using ρX one gets the standard construction of extension of structure groups. By Remark 3.7, via ρ one obtains an associated parabolic vector bundle ρX : G θ,X −→ G LθV ,X . (5.1) EV,θV := (E (V ), θV ) with parabolic weights θV . We will mostly need to apply associated constructions under the adjoint homomorphism Ad : G −→ GL(g). 5.2. Tensor product of parabolic vector bundles. Assume that we have homomorphisms ρ1 : G −→ GL(V ) and ρ2 : G −→ GL(W ), and let ρ1 ⊗ ρ2 : G −→ GL(V ⊗W ) be their tensor product. Therefore, for a parahoric torsor (E , θ) we get parabolic vector bundles EV,θV , EW,θW , and EV ⊗W,θV ⊗W . When the weights θ are rational numbers then, by (Γ,G) bundle theory, we have a canon- ical isomorphism of parabolic vector bundles [MY]: EV,θV ⊗p EW,θW ≃ EV ⊗W,θV ⊗W , where ⊗p is parabolic tensor product. The following proposition, which extends this iso- morphism to the case of real weights, is immediate by observing that the quasi -- parabolic bundle for a parabolic tensor product does not change under sufficiently small change of the parabolic weights of the factors, while the parabolic weights of the parabolic tensor product are given by the parabolic weights of the factors using a standard algebraic for- mula. Proposition 5.1. Let θ ∈ A T be arbitrary. Then there is a canonical isomorphism EV,θV ⊗p EW,θW ≃ EV ⊗W,θV ⊗W of parabolic vector bundles. 12 V. BALAJI, I. BISWAS, AND Y. PANDEY Let V and W be two representations of G and ψ : V −→ W a G -- equivariant homomor- phism. Let (E , θ) be a parahoric torsor. By identifying the G -- module Hom(V,W ) with W ⊗V ∗ it follows that ψ induces a homomorphism (E , θ)(ψ) : EV,θV −→ EW,θW . (5.2) 5.3. Lie algebra bundle of a G -torsor. Consider a parahoric Gθ -- torsor (E , θ), where θ is a system of real weights. We define a parabolic Lie bracket operation on E (g) as follows. The Lie bracket [, ] : g⊗ g −→ g is a G -- equivariant homomorphism for the adjoint action of G, and hence by (5.2) we have a homomorphism of parabolic vector bundles Now Proposition 5.1 gives an isomorphism E[,] : E (g⊗ g) −→ E (g) . E (g)⊗p E (g) ∼−→ E (g⊗ g) of parabolic vector bundles. Combining with E[,], the Lie bracket can be defined as the parabolic homomorphism [., .] : E (g)⊗ E (g) −→ E (g) . (5.3) p X 6. SEMISTABILITY AND STABILITY OF TORSORS G Ω,X be a Bruhat -- Tits group scheme on X as in Definition 3.3, and let θ be such that θ,X . Let (E , θ) be a parahoric G Let G Ω,X ≃ G Remark 6.1. Let GK be a split group scheme over a field K . Let EK be a GK -- torsor. Consider the twisted group scheme E (G)K . Then giving a parabolic subgroup scheme θ,X -- torsor with arbitrary real weights θ ∈ A m T . (of fiber type P ) is equivalent to giving a reduction of structure group of EK to P (cf. [SGA, Exposé XXVI, Cor. 3.6]). PK ⊂ E (G)K 6.1. First definition. Let E (G ) denote the group scheme of automorphisms of E obtained by taking the quotient E ×X G by the left G -- action on E and the right G -- action on itself by conjugation. Let Lie(E (G )) denote the Lie algebra bundle of E (G ). One has the following well-known identification of Lie algebra bundles: Lie(E (G )) = E (g) . (6.1) Since (E , θ) is a parahoric torsor, by Section 5.3, the above Lie algebra bundle E (g) gets a natural parabolic Lie algebra bundle structure (E (g), θg). Thus via the isomorphism (6.1) the bundle Lie(E (G )) gets a parabolic vector bundle structure with a Lie bracket operation compatible with the parabolic structure: By Remark 6.1, giving a generic parabolic reduction is equivalent to giving a parabolic sub- group scheme (Lie(E (G )) , θg , [. , . ]) . (6.2) For each 1 ≤ j ≤ m, we take the flat closure of PK in E (G scheme P ⊂ E (G ). ). This will give a subgroup A j PK ⊂ E (G K ). CONNECTIONS ON PARAHORIC TORSORS OVER CURVES 13 The extended subgroup scheme P also gives a Lie subalgebra bundle Lie(P) ⊂ Lie(E (G )) ≃ E (g) . Then endow the bundle Lie(P) with the canonical induced parabolic structure on the divi- sor D and denote this parabolic subbundle by Lie(P) ∗ . Definition 6.2. We say that the parahoric torsor (E , θ) with arbitrary real weights θ ∈ A T θ,X -- torsor if for every generic parabolic reduction is semistable (respectively, stable) as a G datum as above, m par.deg(Lie(P) ∗ ) ≤ 0 (respectively, par.deg(Lie(P) ∗ ) < 0). 6.2. Second definition. Now assume that (E , θ) is a parahoric torsor with rational weights. Let PK ⊂ GK be a maximal parabolic subgroup of the generic fiber GK of G Ω,X . Let χ : PK −→ Gm,K be a strictly anti-dominant character of the parabolic subgroup PK . There- fore, the associated line bundle on GK /PK is ample. Since the quotient map EK −→ EK /PK defines a principal PK -- bundle, it follows that χ defines a line bundle Lχ on EK /PK = EK (GK /PK ). For any reduction of structure group we have the pulled back line bundle s∗K (Lχ) on X \ D. This line bundle s∗K (Lχ) extends X by the following proposition. sK : X \ D −→ EK (GK /PK ) , Proposition 6.3 ([BS, Proposition 6.3.1]). Let (E , θ) be a parahoric torsor with rational weights. Let sK be a generic reduction of structure group of EK to PK . Then the line bun- dle s∗K (Lχ) on X \ D has a canonical extension Lθ to X as a parabolic line bundle. χ Definition 6.4 ([BS, Definition 6.3.4]). A parahoric torsor (E , θ) with rational weights is called stable (respectively, semistable) if for every maximal parabolic P K ⊂ GK , for all strictly anti-dominant character χ of P K , and for every reduction of structure group sK as above, par.deg(Lθ χ ) < 0 (respectively, par.deg(Lθ χ ) ≤ 0). We observe that the two definitions, namely Definition 6.2 and Definition 6.4, are equiv- alent when the weights are rational; its proof is identical to the proof of [Ra, Lemma 2.1]. Remark 6.5. In a recent paper, Heinloth ([He2]) studies the Hilbert-Mumford criterion in terms of algebraic stacks. The point of view developed there allows natural choices of test objects for the verification of stability, leading to criteria for the existence of separated coarse moduli spaces. Let E is a G -torsor for a parahoric group scheme G and let B ⊂ AutG (E ). Let E B be a reduction of structure group to B. To a character χ of B, one can then associate a line B (λ) on X . In the language of Section 6.2 this will be a parabolic line bundle. In bundle E B (λ) is recovered in the new [He2, Section 3.5] the classical notion of a parabolic degree of E setting which leads to the definition of stability as in Section 6.2. An advantage with his new definition is that it works in positive characteristics as well and moreover, the parahoric group scheme G need not be assumed to be generically split as is done in the present paper. 14 V. BALAJI, I. BISWAS, AND Y. PANDEY 7. (SEMI)STABILITY AND POLYSTABILITY UNDER VARIATION OF WEIGHTS m Let V be the category of all quasi -- parahoric G θ,X -- torsors along D with weights θ varying T . For convenience, we shall work with a single parabolic point P ∈ X . The general- in A ization to finitely many points follows without any difficulty. Let r be the rank of E (g)∗. The degree of the underlying vector bundle is denoted by d1. g denote the space Note that its parabolic degree is 0 because g = g∗ as G -- modules. Let V of all parabolic vector bundles such that • the rank is the fixed integer r , • the quasi -- parabolic structure at P is given by that of E (g)∗, • the degree of the underlying vector bundle is d1, • the parabolic degree is zero, and • the parabolic weights 0 ≤ α1 < α2 < ··· < αr ′ < 1 are not fixed, but the length r ′ and the multiplicities m1, ··· , mr ′ are fixed. Take any V∗ ∈ V plicity of αk for the parabolic structure on W induced by V of V∗ is g. For a subbundle W of V , if nk for each 1 ≤ k ≤ r ′ denote the multi- g, the condition for α∗ -- stability (7.1) degree(W )+ X k≤r ′ nik αik < 0 for all subbundles W of V ; for semistability the strict inequality is replaced by inequality. Let χ(V,W, α) denote the left hand side of (7.1). There exists a constant C2 ≥ 0 such that if degree(W ) > C2, then χ(V,W, α) ≥ 0 for all g, if the underlying vector bundle V admits a subbundle W with α. Thus for any V ∈ V degree(W ) ≥ C2, then V∗ can never be parabolic semistable for any choice of weights α∗. We note that the quasi -- parabolic structure of V ∈ V 7.1. Facets of a quasi -- parahoric torsor. As before, G is simple and simply connected. g alone determines such bundles. In this subsection we shall only consider parahoric torsors E ∈ V such that the quasi -- parabolic bundle E (g) admits no subbundle of degree greater than C2: V g(C2) := {V∗ ∈ V g V has no subbundle of degree greater than C2} . (7.2) Proposition 7.1. The set of inequalities required to verify the (semi)stability of any bundle in V g(C2) has finite cardinality. Proof. From (7.1) we see that there exists a constant C1 ≤ 0 such that degree(W ) ≤ d1 +C1 =⇒ χ(V,W, α) < 0 . In other words, subbundles of degree at most C1 will never be destabilizing with respect to any inequality. Thus to check (semi)stability of (E (g), θg) we may restrict ourselves to subbundles W of E (g) such that C1 ≤ degree(W ) ≤ C2 . (7.3) The ranks of subbundles W vary between 1 and r − 1. Let m1, ··· , mr ′ be the multiplic- g. The multiplicities ni1, ··· , nik of the induced g, only finitely many (cid:3) ities of the quasi -- parabolic structure on V parabolic structure are positive integers. Thus as one varies over V inequalities appear. CONNECTIONS ON PARAHORIC TORSORS OVER CURVES 15 If E (g) ∈ V g(C2), it follows that to check (semi)stability of (E , θ) we need to consider only finitely many inequalities corresponding to a (possibly proper) subset of the set of inequalities seen in Proposition 7.1. This is because we need to check these inequalities for subbundles which are Lie algebra bundles of certain subgroup schemes (see Definition 6.2). We fix a maximal torus T of G and also fix a maximal torus Tg of GL(g) such that Ad (TG ) ⊂ Tg. For every inequality, degree(W )+ X k≤r ′ nik αik < 0 (respectively, ≤ 0) in (7.1) and for every integer c between C1 (as in the proof of Proposition 7.1) and C2, we associate a functional ℓc : A Tg −→ R as follows: for any η = (α1, ··· , αr ) ∈ A , Tg ℓc (η) = c rank(W ) + Pk≤r ′ nik αik rank(W ) . T −→ R by f (θ) = ℓc(θ(g)). These are finitely many in number. We denote the Define f : A set of these functionals by S E T (or ST for notational convenience). Further they are defined over rationals, since clearly the definition of f only involves rational numbers and the map is defined over rationals and is linear. For any functional in f in ST , define the T −→ A A Tg f -- wall in A T as W f := {x ∈ A T f (x) = 0} . The collection {W f }f ∈S E Definition 7.2. Fix a quasi -- parahoric torsor E . For any θ ∈ A will be called the walls of E . T T , let Let Hn = {x ∈ A for some n ≥ 0. T Sθ 1 = {f ∈ S E T x ∈ W f } . Sθ 1 = n}. Define a facet of E to be a connected component of Hn Thus the facets of E provide a decomposition T = G T . Note that only finitely many n appear here. A n Hn of A For any weight θ ∈ A the following three are equivalent: T there is a unique n ≥ 0 such that θ ∈ Hn. Thus for any θ ∈ A T , 1 is empty, and (1) θ does not belong to any E -- wall, (2) Sθ (3) θ ∈ H 0. The facet of θ is the unique facet of E containing it. The following propositions generalize [MS, Section 2, page 217]. We note that the quasi -- parabolic bundles E (g) cannot admit any subbundle of degree greater than C2 (cf. (7.2)). 16 V. BALAJI, I. BISWAS, AND Y. PANDEY T = Y (T )⊗ R. Then there exists an element θ′ in the rational Proposition 7.3. Let θ lie in A apartment AQ = Y (T )⊗ Q such that for all E ∈ V , the pair (E , θ) is semistable (respectively, stable) if and only if (E , θ′) is semistable (respectively, stable). Proof. Let Sθ 2 be the complement of Sθ 1 in ST . We note that θ has a strictly positive distance from each W f , where f ∈ Sθ minimum distance if Sθ d > 0. Let U be the ball in the alcove of radius d around θ. 2 is nonempty and set d to be ∞ if Sθ 2 . Let d be the 2 is empty. Thus in all cases Let I denote the E -- facet of θ. Let I1 be connected component of I ∩ U containing θ, where U is the above ball. Now if I1 is not reduced to a single point, then we can take a rational weight θ′ in it. If I1 is just a point, then θ must be rational because d > 0 and all the functionals are defined over Q. In this case, we take θ′ to be θ itself. Let us check that the (semi)stability conditions for θ and θ′ coincide. For each functional J in ST , (1) if J ∈ Sθ 2. if J ∈ S2 then sign(J (θ′)) = sign(J (θ)), because θ′ ∈ I1 ⊂ U . 1 then J (θ′) = J (θ), because θ′ ∈ I1 ⊂ I , and So for E , one has θ′ -- (semi)stability is equivalent to θ -- (semi)stability. (cid:3) We return to the setting of m -- marked points on X noting that the above discussion im- mediately goes through for multiple marked points. Lemma 7.4. Let θ ∈ A T . Let ρi : G −→ GL(Vi ) for i ≤ m be finitely many representations. Then there exists θ′ ∈ A Q such that for any E ∈ V , and any i ≤ m, the parahoric torsor (E , θ(Vi )) is stable (respectively, semistable) if and only if (E , θ′(Vi )) is stable (respectively, semistable). m m Proof. This is immediate from Proposition 7.3. (cid:3) The following proposition is a generalization of [MS, Proposition 2.1]. m m 0 ∈ A T , there exists a neighborhood U of θ Proposition 7.5. Given any θ T with the property that for all E ∈ V such that (E , θ 0) is stable, the pair (E , θ) is stable for all θ ∈ U . Proof. Now θ may as well be rational. Owing to the stability condition, f (θ) < 0 for all f ∈ ST . Thus we have Sθ 2 = ST . Let d be the minimum distance between θ and any f -- wall. Now we take U to be the ball around θ radius d . 0 in A (cid:3) Definition 7.6. A parahoric torsor (E , θ) for the linear parahoric group scheme G L (V ) is called polystable if the associated parabolic vector bundle E (V ) is polystable (i.e., a direct ∗ sum of stable parabolic bundles of parabolic degree 0). It is straight-forward to check that Lemma 7.4 remains valid if stability in the lemma is substituted by polystability. Corollary 7.7. Let ρ : G −→ GL(V ) be a representation and ρX : G −→ G L (V ) the induced homomorphism of parahoric group schemes as in (5.1). Let E be a G -- torsor. Then, for a weight θ ∈ A T such that (E , θ) is stable, the pair (E (V ) ∗ , θ(V )) is polystable. m CONNECTIONS ON PARAHORIC TORSORS OVER CURVES 17 Proof. By Proposition (7.5), we can assume θ is rational. For a stable equivariant principal G -- bundle, the associated bundles are polystable. Consequently, in view of the equivalence of categories in Section 3.4, The stability of (E , θ) implies that (E (V ), θ(V ) is polystable. (cid:3) 8. CONNECTIONS ON PARAHORIC G -- TORSORS The main aim of this section is to define connections on a parahoric G θ,X -- torsor. 8.1. D X -- modules. We briefly recall the definition of D X -- modules. Definition 8.1. Let X −→ S be a S -- scheme. Let d x denote the image of x under the canon- ical de Rham differentiation map d : OX −→ Ω1 X /S. Let F be a coherent sheaf of OX -- modules over X . By a D X -- module structure on F we mean a OS -- linear homomorphism of sheaf of abelian groups ∇ : F −→ F ⊗OX X /S satisfying Leibniz rule which says that Ω1 where f and x are local sections of F and OX respectively. ∇(x f ) = f ⊗ d x + x∇(f ) , (8.1) Definition 8.2. Let ∇F : F −→ F ⊗ Ω1 F and E respectively. Define their tensor product ∇F ⊗∇E : F ⊗ E −→ F ⊗ E ⊗ Ω1 X /S and ∇E : E −→ E ⊗ Ω1 X /S be two connections over X /S to be ∇F⊗E (f ⊗ e) = ∇F (f )⊗ e + f ⊗∇E (e) , where f and e are local sections of F and E respectively. Similarly define ∇Hom : Hom(E , F ) −→ Hom(E , F )⊗ Ω1 X /S to be ∇Hom(E ,F )(Φ)(e) = ∇F (Φ(e))− Φ(∇E (e)) , where Φ and e are local sections of Hom(E , F ) and E respectively. (8.2) (8.3) 8.2. Logarithmic connections on curves. The canonical line bundle of the smooth com- plex projective curve X will be denoted by K X . Fix a finite subset D = {xi }1≤i≤m ⊂ X ; define K X (logD) = K X ⊗ OX (D) . A logarithmic connection on a vector bundle V −→ X singular on D is a first order algebraic differential operator ∇ : V −→ V ⊗ K X (logD) satisfying the Leibniz rule. For a point x ∈ D, the fiber K X (logD)x is identified with C using the Poincaré adjunction formula. For a logarithmic connection (V, ∇), the composition V ∇−→ V ⊗ K X (logD) −→ (V ⊗ K X (logD))x −→ Vx , which is a C -- linear endomorphism of Vx , is called the residue of ∇ at x [De, page 53], and it is denoted by Res(∇, x). The monodromy of ∇ around x is conjugate to exp(−2πp −1Res(∇, x)) (8.4) [De, p. 53, Théorème 1.17]. 18 V. BALAJI, I. BISWAS, AND Y. PANDEY 8.3. Connection on parabolic vector bundles. Let V −→ X be a vector bundle on X . A quasi -- parabolic structure on V over D is a filtration, for each x ∈ D, of subspaces (8.5) Vx = F x 1 ) F x 2 ) ··· ) F x ax ) F x ax+1 = {0} . A parabolic structure on V over D is a quasi -- parabolic structure as above together with real numbers 0 ≤ αx 1 < ··· < αx i < ··· < αx ax < 1 ∗ , αx ∗}x∈D ) by V∗. associated to the quasi -- parabolic flags. We shall often abbreviate a parabolic vector bundle (V, {F x Definition 8.3. A connection on V∗ is a logarithmic connection ∇ on V such that for all x ∈ D, • the residue Res(∇, x) is semisimple and preserves the quasi -- parabolic flag at x, meaning Res(∇, x)(F x i+1) = αx • Res(∇, x)(F x i ) ⊆ F x i i IdF x i /F x i+1 for all i , and i /F x . A connection on V∗ induces a connection on the dual parabolic vector bundle V ∗ ∗ . To see this, given a logarithmic connection ∇ on V defining a connection on V∗, consider the log- arithmic connection on V ∗ ⊗ OX (D) induced by ∇. This logarithmic connection preserves the subsheaf of V ∗⊗ OX (D) identified with the vector bundle underlying the parabolic vec- tor bundle V ∗ ∗ . The logarithmic connection on this subsheaf obtained this way defines a connection on V ∗ ∗ . ∗ be parabolic vector bundles with underlying vector bundles V 1 and V 2 ∗ and V 2 respectively. Let ∇1 and ∇2 be connections on V 1 ∗ respectively. Consider the loga- rithmic connection on V 1 ⊗V 2 ⊗ OX (D) induced by ∇1 and ∇2. It preserves the subsheaf of ∗ ⊗p V 2 V 1 ⊗ V 2 ⊗ OX (D) corresponding to the parabolic tensor product V 1 ∗ . The logarithmic ∗ ⊗p V 2 connection on this subsheaf obtained this way defines a connection on V 1 ∗ . ∗ and V 2 Let V 1 8.4. Lie connection on a principal G -- bundle. For a principal G -- bundle E −→ X , let E (g) = EG ×G g be its adjoint bundle. The fibers of E (g) are equipped with a Lie bracket structure [., .] : E (g)⊗ E (g) −→ E (g) induced by the Lie algebra structure of g. Definition 8.4. A Lie connection on E is a connection such that following diagram is commutative ∇ : E (g) −→ E (g)⊗ Ω1 X E (g)⊗ E (g) ∇⊗ E (g)⊗ E (g)⊗ Ω1 X [.,.] [.,.]⊗IdΩ1 X E (g) ∇ / E (g)⊗ Ω1 X (8.6) where ∇⊗ is the connection on E (g)⊗ E (g) induced by ∇. The above commutativity condition means that the section of E (g)⊗ (E (g)⊗ E (g))∗ given by the Lie bracket operation on E (g) is flat with respect to the connection on E (g)⊗ (E (g)⊗ E (g))∗ induced by ∇. / / / CONNECTIONS ON PARAHORIC TORSORS OVER CURVES 19 A connection on EG produces a Lie connection on EG . Therefore, we get a map from the space of all connections on EG to the space of all Lie connection on EG . Since G is semisimple, the adjoint homomorphism G −→ GL(g) has finite kernel and its image is the connected component, containing the identity element, of the group of all automorphisms of the Lie algebra g. From this it follows that the above map from the space of all connec- tions on EG to the space of all Lie connection on EG is a bijection. This fact motivates the definition of a connection on a parahoric torsor. θ,X . 8.5. Connection on parahoric G -- torsors. We refer to Section 5.3 for the notation. Take any G = G Definition 8.5. A connection on a G -- torsor E is a logarithmic parabolic connection (see Definition 8.3) ∇ on E (g) satisfying the condition that the section of Homp (E (g) ⊗p E (g), E (g)) given by the homomorphism in (5.3) is flat with respect to the connection ∇hom on the parabolic vector bundle Homp (E (g)⊗p E (g), E (g)) induced by ∇. 8.6. A Tannakian description of connection. Let M be a smooth complex variety. Let G be a complex reductive algebraic group and EG −→ M a principal G -- bundle. Take any pair (H, f ), where H is a complex algebraic group and f : G −→ H an algebraic homomorphism such that corresponding homomorphism of Lie algebras d f : Lie(G) −→ Lie(H) is injective. Let EH := EH ×f H −→ M be the principal H -- bundle obtained by extending the structure group of EG using f . Let ef : EG −→ EH be the natural morphism. A connection on EG induces a connection on EH . The converse is also true. To see this, fix a G -- equivariant splitting σ : Lie(H) −→ Lie(G) , meaning σ◦ d f = IdLie(G) (such a splitting exists because G is reductive). If D is a Lie(H) -- valued 1 -- form on EH defining a connection on H, then σ◦ ef ∗D is a connection on EG . If D0 is a connection on EG and D the connection on EH induced by D0, then the connection σ◦ ef ∗D on EG coincides with D0. Indeed, this follows immediately from the fact that σ◦ d f = IdLie(G). Therefore, the map from connections on EG to connections on EH is injective. The image of this map from connections on EG admits a group theoretic description. This will be explained below. We remove the assumption that the algebraic group G is reductive. As before f : G −→ H to be any algebraic homomorphism such that d f is injective. A theorem of C. Chevalley (see [Hu, p. 80]) says that there is a finite dimensional left H -- module and a complex line ℓ ⊂ W such that f (G) is exactly the isotropy subgroup, of the point in the projective space P (W ) representing the line ℓ, for the action of H on the projective space P (W ) of lines in W induced by the action of H on W . Let ρ : H −→ GL(W ) be the vector bundle associated to EH for the H -- module W . For a connection D on EH , the connection on EW induced by D will be denoted by DW . Note that EW is identified with the EW := EH ×H W −→ M 20 V. BALAJI, I. BISWAS, AND Y. PANDEY vector bundle associated to EG for the action ρ ◦ f of G on W . The condition on ℓ implies that the action of G on W preserves it. Let Eℓ ⊂ EW be the line subbundle associated to the G -- submodule ℓ ⊂ W . A connection D on EH is induced by a connection on EG if and only if the corresponding connection DW on EW preserves the above line subbundle Eℓ ⊂ EW . This characterizes the connections on EH that are induced by connections on EG . We recall that the Tannakian theory involves describing properties of principal bundles in terms of properties of associated vector bundles. For a Tannakian description of con- nections on EG , take H = GL(V ), so V is a finite dimensional G -- module. Let EV := EG ×f V −→ M be the vector bundle associated to EG for the G -- module V . From the above observation we know that a connection on EG is a connection D on the vector bundle EV such that the connection on the vector bundle EW induced by D preserves the line subbundle Eℓ ⊂ EW . 9. CONNECTIONS ON (Γ,G) -- BUNDLES AND RATIONAL WEIGHTS Let F be a principal G -- bundle on a curve Y with adjoint bundle where G acts on itself by conjugation. Given a principal G -- bundle E on Y , define the "twist- ing" by F Ad(F ) = F (G) = F ×G G , (9.1) where the equivalence relation identifies all pairs (e, f ) , (ez, f z) ∈ E ×Y F , with z ∈ G. Con- sider the map E ×G F op := (E ×Y F )/ ∼ , ξ : E ×Y F ×Y F ×G −→ E ×Y F , (e, f , f z, z1) 7−→ (ezz1, f z) , where (e, f ) ∈ E ×Y F and z, z1 ∈ G. There is a unique map bξ : (E ×G F op )×Y F (G) −→ E ×G F op such that the following diagram is commutative (E ×Y F )×Y (F ×G) y (E ×G F op )×Y F (G) ξ −→ E ×Y F y bξ −→ E ×G F op ; recall that E ×G F op and F (G) are quotients of E ×Y F and F × G respectively. This map bξ makes E ×G F op a F (G) -- torsor on Y . Let αF be the map from the space of principal G -- bundles to the space of F (G) -- torsors on Y defined by E 7−→ E ×G F op . This αF is an equivalence of categories. Consider the adjoint action of G on Lie(G) = g. Let ad(F ) = F ×G g −→ Y be the associ- ated adjoint vector bundle. We note that ad(F ) is the Lie algebra bundle associated to the group scheme F (G). Let ∇0 be a connection F . Using ∇0 we will define connections on a F (G) -- torsor. CONNECTIONS ON PARAHORIC TORSORS OVER CURVES 21 The connection ∇0 induces a connection on every fiber bundle associated to F . In par- ticular, it produces a connection on F (G); this connection on F (G) given by ∇0 will be de- noted by ∇G 0 . The kernel of the differential d π : T F (G) −→ π∗T Y of the map π in (9.1) is identified with π∗ad(F ). So the above connection ∇G 0 on F (G) gives a homomorphism ∇G 0 : T F (G) −→ π∗ad(F ) . (9.2) Take any F (G) -- torsor ϕ : E −→ Y . Consider the action E × F (G) −→ E of F (G) on E . Let (9.3) δ : T E ⊕ ϕ∗T F (G) −→ T E be the differential of this map giving the action. Consider the differential of ϕ d ϕ : T E −→ ϕ∗T Y . Let be the relative tangent bundle for the projection ϕ. The action of F (G) on E identifies Tϕ with ϕ∗ad(F ). Tϕ := kernel(d ϕ) ⊂ T E A connection on a F (G) -- torsor ϕ : E −→ Y is a holomorphic homomorphism of vector bundles over Y such that β : T E −→ ϕ∗ad(F ) = ad(ϕ∗F ) (1) the restriction of β to Tϕ coincides with the above identification of Tϕ with ϕ∗ad(F ), and (2) for the homomorphism δ in (9.3), δ(kernel(β)⊕ ϕ∗kernel(∇G 0 )) ⊂ kernel(β) , where ∇G 0 is the homomorphism in (9.2). Note that the above definition of a connection on E depends on ∇0. If F is the trivial principal G -- bundle Y × G, then F (G) = Y × G, and a F (G) -- torsor is in fact a principal G -- bundle on Y . If we choose ∇0 to be the trivial connection on Y ×G, then connections on a F (G) -- torsor are same as connections on the corresponding principal G -- bundle. The following lemma is straight-forward to check. Lemma 9.1. Given a F (G) -- torsor ϕ : E −→ Y , a homomorphism β : T E −→ ϕ∗ad(F ) = ad(ϕ∗F ) defines a connection on E if and only if δ∗β = β⊕∇G 0 on T E ⊕ ϕ∗T F (G), where δ is constructed in (9.3) and ∇G Proposition 9.2. Twisting by F defines an equivalence between principal G -- bundles equipped with a connection and F (G) -- torsors equipped with a connection. 0 is the homomorphism in (9.2). 22 V. BALAJI, I. BISWAS, AND Y. PANDEY Proof. Let E be a principal G -- bundle on Y . Let D be a connection on E . Consider D as a g -- valued 1 -- form on E . Let D′ denote the g -- valued 1 -- form on F corresponding to the connection ∇0 on F . So (D, D′) is a g -- valued 1 -- form on the fiber product E ×Y F . The pullback of ad(F ) to F is identified with the trivial vector bundle F × g −→ F . Therefore, (D, D′) defines a 1 -- form with values in the pullback of ad(F ) to E ×Y F . This form on E ×Y F descends to the quotient F (G) -- torsor E ×G F op as a 1 -- form with values in the pullback of ad(F ) to E ×G F op . It is straight-forward to check that this form defines a connection on the F (G) -- torsor corresponding to E . Conversely, let β be a connection on a F (G) -- torsor ϕ : E −→ Y . Consider the pullback β′ of β to E ×Y F as a 1 -- form with values in the pullback of ad(F ). As noted above, the pullback of ad(F ) to E ×Y F is identified with the trivial vector bundle with fiber g. So β′ is a 1 -- form on E ×Y F with values in g. Let D′ be the pullback of the connection form D to E ×Y F . Then β′ − D′ descends to E by the projection E ×Y F −→ E , and this descended form defines a connection on the principal G -- bundle E . The above two constructions are evidently inverses of each other. (cid:3) Assume that Y is equipped with the action of a finite group Γ. A Γ -- connection on a (Γ,G) -- bundle E on Y is a connection on E which is preserved by the action of Γ. Proposition 9.3. Let E −→ Y be a (Γ,G) -- bundle on some Galois cover p : Y −→ X with Galois group Γ. Let E be the parahoric torsor on X with rational weights corresponding to E. Then there is a natural bijection between the connections on E and the Γ -- connections on E. Proof. This follows from the fact that the connections on a Γ -- equivariant vector bundle on Y are in bijection with the connections on the corresponding parabolic vector bundle on X . (cid:3) 10. FLAT UNITARY CONNECTIONS ON PARAHORIC TORSORS AND STABILITY 10.1. Polystable parahoric torsors. Lemma 10.1. Let V∗ be a polystable parabolic vector bundle of parabolic degree zero with real weights θ. Then the parabolic vector bundle (V∗)⊗m ⊗p ((V∗)∗)⊗n is also polystable. Proof. A parabolic vector bundle of parabolic degree zero is polystable if and only if it is given by a unitary representation of π1(X \ D), where D is the parabolic divisor [MS], [Biq]. Since V∗ is polystable, it is given by a representation ρ of π1(X \ D). The parabolic vector bundle (V∗)⊗m ⊗p ((V∗)∗)⊗n is given by the representation ρ⊗m ⊗ (ρ)⊗n. This implies that (V∗)⊗m ⊗p ((V∗)∗)⊗n is polystable. Corollary 10.2. Take V∗ as in Lemma 10.1. Take any homomorphism (cid:3) ρ : GL(r, C) −→ GL(N , C) , where r is the rank of V∗. Let W∗ be the parabolic vector bundle associated to V∗ for ρ. Then W∗ is also polystable. Proof. Consider CN as a GL(r, C) -- module using ρ and the standard representation of GL(N , C). This GL(r, C) -- module CN is a direct summand of a direct sum of GL(r, C) -- modules of the form (Cr )⊗mi ((Cr )∗)⊗ni [DMOS, p. 40, Proposition 3.1(a)]. Therefore, from CONNECTIONS ON PARAHORIC TORSORS OVER CURVES 23 Lemma 10.1 we conclude that W∗ is a direct summand of a polystable parabolic vector bundle of parabolic degree zero. Hence W∗ is polystable. (cid:3) We define polystability for parahoric torsors. θ,X −→ X be a Bruhat -- Tits group scheme with generic fiber G. A Definition 10.3. Let G parahoric G θ,X -- torsor E with real weights θ is said to be polystable if for every representa- tion ρ : G −→ GL(V ), the corresponding parabolic vector bundle (E , θ(V )) is polystable in the sense of Definition 7.6. 10.2. Polystable parahoric torsors from unitary representations. In this subsection we will first assume that D = {x} is a single point. The multi-point case is actually a straight- forward generalization. The complement X \ {x} would be denoted by Y . For a base point y0 ∈ Y , set Γ = π1(Y , y0). Choose an analytic disc U ⊂ X around x such that y0 ∈ U . The inclusion of U \ {x} in Y produces an inclusion π1(U \ {x}, y0) ,→ Γ. Using the orientation of U \ {x}, the group π1(U \ {x}, y0) gets identified with Z. The element of π1(U \ {x}, y0) corresponding to 1 ∈ Z will be denoted by γ. We now recall a description of the set of conjugacy classes in a compact semisimple and simply connected group in terms of the Weyl alcove (see [BS, page 9]). Let KG ⊂ G be a fixed maximal compact subgroup and T a fixed maximal torus in KG . The corresponding Weyl group in the quotient NG (T )/T . The set of conjugacy classes of element in KG gets identified with the T /W which is in fact the Weyl alcove because any element of KG is conjugate to an element in the maximal torus up to an element of the Weyl group (cf. [Mor, page 151]). Given any t ∈ KG, let θt denote the point in the Weyl alcove corresponding to t . Given any homomorphism ρ : Γ −→ KG , let Eρ be the flat principal G -- bundle on Y associated to it. To construct Eρ, let (eY , ey0) be the pointed universal cover of Y corresponding to the base point y0; note that Γ acts on eY . Identify two points (y1, g1), (y2, g2) ∈ eY × G if there is an element γ ∈ Γ such that (y2, g2) = (y1γ, ρ(γ−1)g2). The quotient of eY × G is a principal G -- bundle on Y , which is denoted by Eρ; the right translation action of G on eY ×G produces an action of G on Eρ. The flat connection on the trivial principal G -- bundle eY ×G −→ eY descends to a flat connection on Eρ. For any h ∈ KG , the map eY ×G −→ eY ×G , (y, z) 7−→ (y, ρ(h)z) descends to an isomorphism as flat principal G -- bundles on Y . Let Ehρh−1 ∼−→ E ρ (10.1) t = ρ(γ) ∈ KG be the image of γ. Since h.ρ.h−1(γ) = hth−1 for all h ∈ KG , there is the map Hom(Γ, KG )/KG −→ A , ρ 7−→ θt . After conjugating ρ by an element of KG, we may assume that t belongs to a fixed maximal torus T of KG . 24 V. BALAJI, I. BISWAS, AND Y. PANDEY Let t ∈ Lie(T ) be such that (10.2) where exp denotes the exponential map on the Lie algebra Lie(T ). Consider the trivial prin- cipal G -- bundle (U \{x})×G over U \{x}. The trivial connection on it given by this trivializa- tion will be denoted by d0. On (U \ {x})×G, we now have the flat connection −1t) = t , exp(−2πp bd = d0 + td z z , (10.3) where z is a holomorphic coordinate function on U with z(x) = 0. Restrict the representation ρ to the subgroup π1(U \{x}, y0). This produces a flat principal G -- bundle Eρ(∞) −→ U \ {x}. Note that E ρU \{x} ≃ E ρ (∞) (10.4) as flat principal G -- bundles. Both the flat principal G -- bundles in (10.4) are isomorphic to the flat principal G -- bundle ((U \{x})×G, bd ) constructed in (10.3). This is because all of them have the same residue, namely t, at x. Recall that the residue determines the conjugacy class of the monodromy (see (8.4)). To the element t ∈ KG ∩ T , we have the associated conjugacy class θt ∈ A , and hence by on a formal neighborhood bU = Spec C[[t ]] of is the trivial group scheme Spec C((t ))×G over bU \x = Spec C((t )). uniquely to X by setting it to be the trivial group scheme Y ×G Bruhat -- Tits theory have a group scheme G x. This group scheme G Therefore, we can extend G over Y [BS, Section 5.2, page 28]. We denote this group scheme on X by Gt,X . θt θt θt We observe that there is a morphism bU ,→ U , where one identifies the formal power -- torsor on bU which we denote by Pt . Note that the connection on E ρ (∞) restricts to a series ring with the completion of the convergent power series ring. Consider the trivial G θt natural connection on Pt bU \{x}. We now patch together (using for example [BS, 5.2.3]) the trivial G -- torsor Pt on bU and the principal G -- bundle Eρ over Y along the intersection bU \ {x} = Spec C((t )) such that the patching is connection preserving; as noted above, on Spec C((t )), both the principal G -- bundles are the trivial principal G -- bundle Spec C((t ))×G equipped with the connection bd . θt The above construction is summarized in the following proposition. Proposition 10.4. Given any homomorphism ρ : Γ −→ KG, and any t ∈ Lie(T ) satisfying (10.2), the flat principal G -- bundle on Y has a canonical extension to a Gt,X -- torsor over X . It should be clarified that the Gt,X -- torsor in Proposition (10.4) depends on the choice of t (a branch of the logarithm), while the isomorphism class of the Bruhat -- Tits group scheme Gt,X depends only on the conjugacy class [t ] of t . For G = GL(r, C), if the logarithm t is chosen as done in [MS] (meaning t is semisimple and eigenvalues are nonnegative and less than 1), then the construction in (10.4) coincides with the construction in [MS] of a parabolic vector bundle from a homomorphism Γ −→ U(r ). This follows by comparing the two constructions. The Gt,X -- torsor in Proposition (10.4) will be denoted by Eρ(t ). It may be mentioned that we may restrict the connection bd in (10.3) to bU \ {x} ⊂ U \ {x}, where bU as before is the formal completion along x. A Gt,X -- torsor on X can be trivialized CONNECTIONS ON PARAHORIC TORSORS OVER CURVES 25 over both Y and bU . Conversely, given Gt,X -- torsors on Y and bU , and an isomorphism be- tween them over bU \ {x}, we get a Gt,X -- torsor on X . Therefore, the connection over bU \ {x} is enough to construct the Gt,X -- torsor Eρ(t ). 10.3. Polystable parahoric torsors and unitary representations. As before, fix a maximal compact subgroup KG of G. m θ,X -- torsor on X with arbitrary real weights θ ∈ T . Then (E , θ) is polystable if and only if (E , θ) is given by a homomorphism from π1(X \D) Theorem 10.5. Let (E , θ) be a parahoric G A to KG as described in Section 10.2. Proof. First assume that (E , θ) is given by a homomorphism β : π1(X \ D) −→ KG . Let ρ : G −→ GL(V ) be any homomorphism. Fix a maximal compact subgroup KGL(V ) of GL(V ) such that β(KG ) ⊂ KGL(V ). Then the parabolic vector bundle W∗ associated to (E , θ) for ρ is given by the homomorphism Therefore, this associated parabolic vector bundle W∗ is polystable. ρ ◦ β : π1(X \ D) −→ KGL(V ) . To prove the converse, assume that (E , θ) is polystable. Let E (g) and E (g ⊗ g) be the parabolic vector bundles associated to (E , θ) for the G -- modules g and g⊗ g respectively. From Definition 10.3 we know that both E (g) and E (g⊗g) are polystable of parabolic degree zero. If E (g) is given by a homomorphism β from π1(X \D) to a maximal compact subgroup of GL(g), then E (g⊗ g) is given by β⊗ β. For any two parabolic vector bundles given by unitary representations of π1(X \ D), any homomorphism between them is given by a homomorphism of π1(X \ D) -- modules. Let γ : E (g⊗ g) −→ E (g) be the homomorphism of parabolic vector bundles given by the Lie bracket g ⊗ g −→ g. From the above statement we conclude that γ is given by a homomorphism of π1(X \ D) -- modules. This implies that the connection on E (g) is induced by a connection on (E , θ) (see Definition 8.5). Therefore, (E , θ) is given by a homomorphism from π1(X \ D) to KG . (cid:3) A homomorphism ρ : π1(X \ D) −→ KG is called irreducible if ρ(π1(X \ D)) is not con- tained in some proper parabolic subgroup of G. A homomorphism ρ is irreducible if and only if the space of invariants in g for the adjoint action of ρ(π1(X \D)) is the zero element. θ,X -- torsor on X with arbitrary real weights θ ∈ T . Then (E , θ) is stable if and only if (E , θ) is given by an irreducible homomorphism Corollary 10.6. Let (E , θ) be a parahoric G A from π1(X \ D) to KG as described in Section 10.2. m Proof. Assume that (E , θ) is stable. Therefore, (E , θ) is polystable. If the homomorphism ρ : π1(X \D) −→ KG corresponding to (E , θ) has the property that ρ(π1(X \D)) is contained in a proper parabolic subgroup P of G, then the reduction of E to P over X \ D contradicts the stability of (E , θ). Therefore, ρ is irreducible. Conversely, for a polystable (E , θ), if the corresponding homomorphism ρ : π1(X \D) −→ KG is irreducible, then the polystable parabolic vector bundle E (g) does not admit any 26 V. BALAJI, I. BISWAS, AND Y. PANDEY holomorphic section [Si, p. 744, Theorem 3]. Consequently, the polystable torsor (E , θ) (cid:3) is stable. Remark 10.7. The Corlette -- Donaldson -- Hitchin -- Simpson correspondence between flat G -- bundles and G -- Higgs bundles also extends to the parahoric case. When G = GL(n, C) this was proved by Simpson in [Si]. This result of [Si] is the key ingredient in this extension for general G. Using this result the question for G is reduced to one on vector bundles using the adjoint representation of G. The approach in the present paper then goes through without any essential difficulty. Remark 10.8. The paper [BGM] considers the problem of parabolic Higgs G -- bundle and the Corlette -- Donaldson -- Hitchin -- Simpson correspondence on curves from a somewhat differ- ent perspective and also consider real representations. Remark 10.9. The Atiyah -- Weil criterion, [At], [We], [AB], for the existence of a holomorphic connection on a holomorphic principal G -- bundle generalizes to G -- torsors. The proof in [AB] has a straight-forward generalization. Similarly, the Atiyah -- Krull -- Schmitt reduction of a holomorphic principal G -- bundle, [BBN2], generalizes to G -- torsors. Remark 10.10. Theorem 10.5 evidently generalizes to the situation where G is a product of simple and simply connected groups. The more general case of semisimple groups G that are not simply connected is covered by using twisted bundles as in [BLS]. For a reductive group G, the natural map G −→ G/Z0(G)× (G/[G, G]) is surjective with finite kernel, where Z0(G) is the connected component of the center of G containing the identity element. Since G/Z0(G) is semisimple and /[G, G] is a product of copies of Gm, to prove Theorem 10.5 it suffices to prove it for Gm. But this was done in [Si]. 10.4. The reductive case. We now indicate briefly how to extend the consideration of (semi)stability of torsors in the case when the structure group G is a connected reductive algebraic group and identify it with the space of homomorphisms from π to KG . However, the corresponding relationship with parahoric torsors for reductive G needs a closer analy- sis of Bruhat -- Tits theory for reductive groups. Let S = [G, G] be the derived group, i.e. the maximal connected semisimple subgroup of G. Let Z0 be the connected component of the center of G (which is a torus) and one know that S and Z0 together generate G. Let H = Z0 × S. Then in fact, H −→ G is a finite covering map. It is easy to see (following [Ra, page 145]) that (Γ, H) -- bundles gives rise to (Γ,G) -- bundles and the stability and semistability of the associated (Γ,G) -- bundles follows immediately from that of the (Γ, H) -- bundles. Thus, the problem of handling the reductive group H reduces to the problem of han- dling the semisimple group G but which is not simply connected. On the side of Bruhat -- Tits group schemes and parahoric group schemes, for a general connected reductive G the existence of Bruhat -- Tits group schemes are well known and would give the existence of similar group schemes on the whole of X . Several technical issues which one has avoided are in the setting of the Bruhat -- Tits buildings. Canonical choices of apartments and alcoves which give a transparent meaning to the association of conjugacy classes with weights in the alcove would need technical modifications which led us too far afield; future consider- ations would therefore need a careful discussion on a "canonical" choice of apartment as for example indicated in ([Tits, page 32]). CONNECTIONS ON PARAHORIC TORSORS OVER CURVES 27 ACKNOWLEDGEMENTS We are very grateful to the referee for very detailed comments. A portion of Section 1.1 was formulated by the referee. The first two authors acknowledge partial support by the J. C. Bose Fellowship. The first author thanks Department of Mathematics, Pondicherry University for its hospitality where this work was completed. REFERENCES [At] [AB] [BLS] M.F. Atiyah, Complex analytic connections in fibre bundles, Trans. Amer. Math. Soc. 85 (1957) 181 -- 207. H. Azad and I. Biswas, On holomorphic principal bundles over a compact Riemann surface admit- ting a flat connection, Math. Ann. 322 (2002), 333 -- 346. A. Beauville, Y. Laszlo and C. Sorger, The Picard group of the moduli of G-bundles on a curve, Com- positio Math. 112 (1998), 183 -- 216. [BBN1] V. Balaji, I. Biswas and D. S. Nagaraj, Principal bundles over projective manifolds with parabolic structure over a divisor, Tohoku Math. Jour. 53 (2001), 337 -- 368. [BBN2] V. Balaji, I. Biswas and D. S. Nagaraj, Krull -- Schmidt reduction for principal bundles, Jour. Reine [BS] [Bis] [Biq] Angew. Math. 578 (2005), 225 -- 234. V. Balaji and C.S Seshadri, Moduli of parahoric G -torsors on a compact Riemann surface, Jour. Alg. Geom. 24 (2015), 1 -- 49. (arxiv:math.AG 1009.3485) I. Biswas, Parabolic bundles as orbifold bundles, Duke Math. Jour. 88 (1997), 305 -- 325. O. Biquard, Fibrés Paraboliques Stables et Connexions Singulières Plates, Bull. Soc. Math. Fr. 119 (1991), 231 -- 257. [BGM] O. Biquard, Oscar Garcia-Prada, Ignasi Mundet i Riera, Parabolic Higgs bundles and representations [Bo] [BLR] [BT1] [BT2] [De] of the fundamental group of a punctured surface into a real group, arXiv:1510.04207v1. P. Boalch, Riemann-Hilbert for tame complex parahoric connections, Transform. Gr. 16 (2011), 27 -- 50. (math arxiv DG 1003.3177). S. Bosch, W. Lutkebohmert and M. Raynaud, Neron models, Ergebnisse 21, Springer Verlag, Berlin- New York, 1990. F. Bruhat and J. Tits, Groupes réductifs sur un Corps Local, I: Données radicielles valuées, Inst. Hautes Éludes Sci. Publ. Math. 41 (1972), 5 -- 252. F. Bruhat and J. Tits, Groupes réductifs sur un corps local II, Schémas en groupes, Inst. Hautes Éludes Sci. Publ. Math. 60 (1984), 197 -- 376. P. Deligne, Équations différentielles à points singuliers réguliers, Lecture Notes in Mathematics, Vol. 163, Springer-Verlag, Berlin-New York, 1970. [DMOS] P. Deligne, J. S. Milne, A. Ogus and K.-y. Shih, Hodge cycles, motives, and Shimura varieties, Lecture [Ed] [He1] [He2] [Hu] [La] [MY] [MS] [Mor] [Na] [NS] Notes in Mathematics, 900, Springer-Verlag, Berlin-New York, 1982. B. Edixhoven, Néron models and tame ramification, Compos. Math. 81 (1992), 291 -- 306. J. Heinloth, Uniformization of G -bundles, Math. Ann. 347 (2010), 499 -- 528. J. Heinloth, Hilbert-Mumford stability on algebraic stacks and applications to G -bundles on curves, arXiv:1609.06058. J. E. Humphreys, Linear algebraic groups, Graduate Texts in Mathematics, Vol. 21, Springer-Verlag, New York, Heidelberg, Berlin, 1987. E.Landvogt, Some functorial properties of the Bruhat-Tits building, Jour. Reine Angew. Math. 518 (2000), 213 -- 241. M. Maruyama and K.Yokogawa, Moduli of parabolic stable sheaves. Math. Ann. 293 (1992), 77 -- 99. V.Mehta and C.S. Seshadri, Moduli of vector bundles on smooth curves with parabolic structures, Math. Ann. 248 (1980), 205 -- 239. J.W. Morgan, Holomorphic bundles over elliptic manifolds, ICTP Lecture Notes, Volume 1, (2000). M. Namba, Branched coverings and algebraic functions, Pitman Research Notes in Mathematics Se- ries, 161, Longman Scientific & Technical, Harlow; John Wiley & Sons, Inc., New York, 1987. M. S. Narasimhan and C. S. Seshadri, Stable and unitary vector bundles on a compact Riemann surface, Ann. of Math. 82 (1965), 540 -- 567. 28 [No] [PR1] [PR2] [Ra] [RS] [Se1] [Se2] [SGA] [Si] [Sp] [TW] [Tits] [We] V. BALAJI, I. BISWAS, AND Y. PANDEY M.V. Nori, The fundamental group scheme, Proc. Ind. Acad. Sci. Math. Sci. 91 (1982), 73 -- 122. G. Pappas and M. Rapoport, Twisted loop groups and their affine flag varieties, Advances in Math 219 (2008), 118 -- 198. G. Pappas and M. Rapoport, Some questions about G -bundles on curves, Algebraic and arithmetic structures of moduli spaces (Sapporo 2007), pp. 159 -- 171, Adv. Stud. Pure Math., vol. 58, Math. Soc. Japan, Tokyo, 2010. A. Ramanathan, Stable principal bundles on a compact Riemann surface, Math. Ann. 213 (1975), 129 -- 152. A. Ramanathan and Subramanian, Einstein-Hermitian connections on principal bundles and stabil- ity, Jour. Reine Angew. Math. 390 (1988), 21 -- 31. C. S. Seshadri, Moduli of π -- vector bundles over an algebraic curve, Questions On al- gebraic Varieties, C.I.M.E, Varenna, (For a new corrected version see (1969), 139-261. http://www.cmi.ac.in/∼balaji/Publications/CSS1.19.pdf). C. S. Seshadri, Remarks on Parabolic Structures, Contemporary Mathematics, Volume 522, Papers in Honour of S.Ramanan, 2010. Séminaire de Géométrie Algèbrique du Bois-Marie SGA3, Séminaire 1962/1964 - Schémas en groupes - (SGA 3) - vol. 1 (Lecture notes in mathematics 151) (in French). Berlin; New York: Springer- Verlag. C. T. Simpson, Harmonic bundles on noncompact curves, Jour. Amer. Math. Soc. 3 (1990), 713 -- 770. T. A. Springer, Linear algebraic groups, Second edition, Progress in Mathematics, 9, Birkhäuser Boston, Inc., Boston, MA, 1998. C. Teleman and C. Woodward, Parabolic bundles, products of conjugacy classes, and quantum co- homology, Ann. Institut Fourier 3 (2003), 713 -- 748. J. Tits, Reductive groups over local fields, Proceedings of Symposia in Pure Mathematics, Vol 33, (1979), 29-69. A. Weil, Généralisation des fonctions abéliennes, Jour. Math. Pures Appl. 17 (1938), 47 -- 87. CHENNAI MATHEMATICAL INSTITUTE, SIPCOT IT PARK, SIRUSERI 603103, INDIA E-mail address: [email protected] SCHOOL OF MATHEMATICS, TATA INSTITUTE OF FUNDAMENTAL RESEARCH, HOMI BHABHA ROAD, MUMBAI 400005, INDIA E-mail address: [email protected] INDIAN INSTITUTE OF SCIENCE EDUCATION AND RESEARCH, MOHALI KNOWLEDGE CITY, SECTOR 81, SAS NAGAR, MANAULI PO 140306, INDIA E-mail address: [email protected]
1711.10215	1	1711	2017-11-28T10:19:30	Stratifying quotient stacks and moduli stacks	[ "math.AG" ]	Recent results in geometric invariant theory (GIT) for non-reductive linear algebraic group actions allow us to stratify quotient stacks of the form [X/H], where X is a projective scheme and H is a linear algebraic group with internally graded unipotent radical acting linearly on X, in such a way that each stratum [S/H] has a geometric quotient S/H. This leads to stratifications of moduli stacks (for example, sheaves over a projective scheme) such that each stratum has a coarse moduli space.	math.AG	math	STRATIFYING QUOTIENT STACKS AND MODULI STACKS GERGELY B ´ERCZI, VICTORIA HOSKINS, FRANCES KIRWAN ABSTRACT. Recent results in geometric invariant theory (GIT) for non-reductive linear algebraic group actions allow us to stratify quotient stacks of the form [X/H], where X is a projective scheme and H is a linear algebraic group with internally graded unipotent radical acting linearly on X, in such a way that each stratum [S/H] has a geometric quotient S/H. This leads to stratifi- cations of moduli stacks (for example, sheaves over a projective scheme) such that each stratum has a coarse moduli space. 7 1 0 2 v o N 8 2 ] . G A h t a m [ 1 v 5 1 2 0 1 . 1 1 7 1 : v i X r a 1. INTRODUCTION Let H = U ⋊ R be a linear algebraic group over an algebraically closed field of characteristic 0 with internally graded unipotent radical U ; that is, the Levi subgroup R of H has a central one-parameter subgroup (1-PS) λ : Gm → R which acts on Lie U with all weights strictly posi- tive. Of course any reductive group G has this form with both U and the central one-parameter subgroup λ being trivial. Suppose that H acts linearly on an irreducible projective scheme X with respect to an ample line bundle L over X. The aim of this paper is to describe a strati- fication of the quotient stack [X/H] such that each stratum [S/H] (where S is an H-invariant quasi-projective subscheme of X) has a geometric quotient S/H. When H = R is reductive this stratification refines the Hesselink–Kempf–Kirwan–Ness (HKKN) stratification associated to the linear action on X (cf. [18, 22, 23, 27]). Potential applications of this construction include moduli stacks which can be filtered by quotient stacks with compatible linearisations; for exam- ple, it can be applied to moduli of sheaves of fixed Harder–Narasimhan type over a projective scheme [6], and moduli of unstable projective curves [21]. When H is reductive Mumford's geometric invariant theory (GIT) [26] allows us to find open subschemes X s ⊆ X ss of X, the stable and semistable loci, such that X s has a geometric quotient X s/H and X ss has a good quotient φ : X ss → X//H = Proj Mm>0 H 0(X, L⊗m)H! . Here X s = X s,H = X s,H,L and X ss = X ss,H = X ss,H,L depend on the choice of linearisation L (that is, the ample line bundle L and the lift of the action of H to an action on L) and the GIT quotient X//H = X//LH is a projective scheme with X s/H as an open subscheme. Moreover when x, y ∈ X ss then φ(x) = φ(y) if and only if the closures of the H-orbits of x and y meet in X ss. The Hilbert–Mumford criteria allow us to determine the stable and semistable loci in a simple way without needing to understand the algebra of invariantsLm>0 H 0(X, L⊗m)H. The best situation occurs when X ss = X s 6= ∅; then X//H = X s/H is both a projective scheme and a geometric quotient of the open subscheme X s of X. More generally if X s 6= ∅ V.H. is supported by the Excellence Initiative of the DFG at the Freie Universitat Berlin. 1 STRATIFYING QUOTIENT STACKS AND MODULI STACKS 2 then the projective completion X//H of the geometric quotient X s/H has a 'partial desingular- there exists x ∈ X ss whose stabiliser in H is reductive of dimension strictly bigger than 0. To isation' eX//H = eX ss/H where ψ : eX ss → X ss is obtained as follows [24]. If X ss 6= X s then construct eX ss we first blow up X ss along its closed subscheme where the stabiliser has maxi- mal dimension in X ss (such stabilisers are always reductive) or equivalently blow up X along the closure of this subscheme in X. We then remove the complement of the semistable locus for a small ample perturbation of the pullback linearisation. The maximal dimension of a stabiliser in this new semistable locus is strictly less than in X ss. When X s 6= ∅, repeating this process finitely many times leads to eX ss = eX s 6= ∅ and the partial desingularisation eX//H = eX s/H. When H is non-reductive, then the graded algebraLm>0 H 0(X, L⊗m)H is not necessarily finitely generated and in general the attractive properties of Mumford's GIT fail [2]. However when the unipotent radical U of H = U ⋊ R is graded in the sense described above by a central 1-PS λ : Gm → R of the Levi subgroup R, then after twisting the linearisation by an appropriate rational character, so that it becomes 'graded' itself in the sense of [5], some of the desirable properties of classical GIT still hold [3, 4]. More precisely, we first quotient R/λ(Gm). In the best case, when the U -action is free on a certain open subscheme X 0 min of X (cf. the condition (∗) in Definition 2.11 and Theorem 2.13), one can construct a geometric quotient by the linear action of the graded unipotent group bU := U ⋊ λ(Gm) using the results of [3, 4] described in §2.3, then we quotient by the residual action of the reductive group H/bU ∼= of an open subscheme of 'stable points' for thebU -action such that the quotient is projective and sequence of blow-ups of X along H-invariant subschemes to obtain ψ : eX → X such that eX bU -quotient of an open subscheme of eX that contains as an open subscheme a geometric bU - this stable set has a Hilbert–Mumford type description. If the U -action has positive dimensional stabilisers generically, one can conclude the same results if we assume a weaker condition (∗ ∗ ∗) (cf. Theorem 2.17). Even when this weaker condition fails, one can perform an iterated quotient of an open subscheme of X, as ψ is an isomorphism away from the exceptional divisor. Now suppose that G is a reductive group acting linearly on a projective scheme X with respect to an ample line bundle L. Associated to this linear G-action and an invariant inner product on Lie G, there is a stratification (the 'HKKN stratification' cf. [18, 22, 23, 27]) has an induced linear H-action satisfying (∗ ∗ ∗). Hence, there is a projective and geometric of X by locally closed subschemes Sβ, indexed by a partially ordered finite set B, such that X = Gβ∈B Sβ (1) if X ss 6= ∅, then B has a minimal element 0 such that S0 = X ss, (2) for β ∈ B, the closure of Sβ is contained inSβ ′>β Sβ ′, and (3) for β ∈ B, we have Sβ β , where G ×Pβ Y ss β of X. β by the diagonal action of a parabolic subgroup Pβ of G acting on the right on G and on the left on a Pβ-invariant locally closed subscheme Y ss β is the quotient of G × Y ss ∼= G ×Pβ Y ss In fact, Y ss β is an open subscheme of a projective subscheme Y β of X that is determined by the action of a Levi subgroup Lβ of Pβ with respect to the restriction of the G-linearisation L → X to the Pβ-action on Y β twisted by a rational character χβ of Pβ. The index β determines a central (rational) 1-PS λβ : Gm → Lβ and χβ is the corresponding rational character, where the choice of invariant inner product allows us to identify characters and co-characters of a fixed STRATIFYING QUOTIENT STACKS AND MODULI STACKS 3 maximal torus (cf. Remark 2.2). Furthermore Pβ is the parabolic subgroup P (λβ) determined by the 1-PS λβ, which grades the unipotent radical Uβ of Pβ. Thus by Property (3) above, to construct a G-quotient of (an open subset of) an unstable stratum Sβ, we can study the linear Pβ-action on Y β and apply the results described above for the action of bU := Uβ ⋊ λβ(Gm). The G-action on the stratum Sβ has a categorical quotient Zβ//Lβ induced by the morphism pβ : Y ss β → Z ss ss,Lβ/λβ(Gm) β = Z β y 7→ pβ(y) := lim t→0 λβ(t)y, where Zβ is the union of those connected components of the fixed point set for the action of λβ(Gm) on X over which λβ acts on the fibres of L with weight given by the restriction of β. However this categorical quotient is in general far from being a geometric quotient; it identifies y with pβ(y), which lies in the orbit closure but typically not the orbit of y. In this article we will show that, applying the blow-up sequence needed to construct a quo- tient by an action of a linear algebraic group with internally graded unipotent radical to the Pβ-action on the projective subscheme Y β of X, we can refine the stratification {Sββ ∈ B} to obtain a stratification of X such that each stratum is a G-invariant quasi-projective subscheme of X with a geometric quotient by the action of G. This refined stratification is a further refine- ment of the construction described in [25]. The quotient stack [X/G] has an induced stratifica- tion {Σγγ ∈ Γ} such that each stratum Σγ has the form Σγ ∼= [Wγ/Hγ] where Wγ is a quasi-projective subscheme of X acted on by a linear algebraic subgroup Hγ of G with internally graded unipotent radical, and this action has a geometric quotient Wγ/Hγ. Moreover under appropriate hypotheses (involving condition (∗ ∗ ∗)) for the actions of the sub- groups Hγ, the geometric quotients Wγ/Hγ are themselves projective. This will follow from the following theorem, which is proved in §3. Theorem 1.1. Let H = U ⋊ R be a linear algebraic group with internally graded unipotent radical U acting on a projective scheme X over an algebraically closed field k of characteristic 0 with respect to an ample linearisation and fix an invariant inner product on Lie R. Then X has a stratification {Sγγ ∈ Γ} induced by the linearisation L and grading λ : Gm → R for the action of H on X, such that the following properties hold. i) The index set Γ is finite and partially ordered such that for all γ ∈ Γ, we have Sγ ⊆ Sγ ∪ [δ∈Γ,δ>γ Sδ. ii) Each Sγ is a H-invariant quasi-projective subscheme of X with a geometric quotient Sγ/H. iii) If Y is an H-invariant projective subscheme of X then the stratification {S Y γ γ ∈ ΓY } of Y induced by the restriction LY of the linearisation L to Y is (up to taking connected components) the restriction to Y of the stratification {Sγγ ∈ Γ} of X, so that there is a map of indexing sets φY : ΓY → Γ such that if γ ∈ ΓY then S Y γ is a connected component of SφY (γ) ∩ Y ; Moreover, if H = G is reductive, then this stratification satisfies the following additional properties. iv) The stratification {Sγγ ∈ Γ} is a refinement of the HKKN stratification {Sββ ∈ B} for the linearisation L (cf. §2.1.1). v) If β ∈ B (which we recall determines a 1-PS λβ : Gm → Pβ ≤ G) satisfies (†) x ∈ Z ss β ⇒ dim(StabG(x)/λβ (Gm)) = 0, STRATIFYING QUOTIENT STACKS AND MODULI STACKS 4 then the connected components of GZ ss refined stratification {Sγγ ∈ Γ}. β and Sβ \ GZ ss β (if these are nonempty) are strata in the As a consequence, we obtain a stratification of the quotient stack [X/H] by locally closed substacks, each of which admits a coarse moduli space (cf. Corollary 3.5). Inspired by the reductive GIT notion of a good quotient, Alper introduces a notion of a good moduli space for a stack [1]. However, in general the strata appearing in this stratification of [X/H] will not admit good moduli spaces, because a necessary condition for a stack to admit a good moduli space is that its closed points have reductive stabiliser groups (cf. [1, Proposition 12.14]). In general (even when H = G is reductive) the points in the strata of [X/H] will have non-reductive stabiliser groups. If H = G is reductive, then a stacky version of the HKKN stratification has been studied by Halpern-Leister, and by abstracting this concept he obtains a notion of a Θ-stratification [16]. Indeed, the linearisation of the G-action on X and the choice of invariant inner product is pre- cisely the data required to construct a Θ-stratification of [X/G], and this Θ-stratification is the stratification {[Sβ/G] : β ∈ B} obtained from the HKKN stratification of X. The stratification described above thus refines this Θ-stratification without depending on any additional data. Since the construction of the refined stratification involves studying the blow-up procedures used in partial desingularisations of reductive GIT quotients [24] and for constructing geomet- ric quotients by linear algebraic groups with internally graded unipotent radical [4], one can ask how this compares with the stack-theoretic blow-up constructions. The ideas in [24] have been generalised to stacks by Edidin and Rydh [15] to show that for a smooth Artin stack X admitting a stable good moduli space, there is a sequence of birational morphisms of smooth Artin stacks Xn → · · · X1 → X such that the good moduli space of Xn is an algebraic space with only tame quotient singularities and is a partial desingularisation of the good moduli space of X. However, for H non-reductive, it is often the case that [X/H] will not be a good moduli space, and so one cannot apply this result. The picture provided by Theorem 1.1 has potential applications to moduli stacks which are filtered by quotient stacks, and to the construction of moduli spaces of 'unstable' objects (for example, moduli of sheaves over a projective scheme [6] or moduli of projective curves [21]). Suppose that M is a moduli stack which can be expressed as an increasing union of open substacks of the form M = [n>0 Un Un ∼= [Vn/Gn] where [Vn/Gn] is the quotient stack associated to a linear action on a quasi-projective scheme Vn by a group Gn which is reductive (or more generally has internally graded unipotent radical). We can look for suitable 'stability conditions' on M: linearisations (Ln)n>0 for the actions of Gn on projective completions Vn of Vn and invariant inner products on Lie Gn which are com- patible in the sense that the stratification induced by Ln on [Vn/Gn] restricts to the stratification induced by Lm on [Vm/Gm] when n > m. This situation arises for sheaves over a projective scheme [20, 6], for example, and also for projective curves [21], and we obtain a stratification {Σγγ ∈ Γ} of the stack M such that each stratum Σγ is isomorphic to a quotient stack [Wγ/Hγ], where Wγ is quasi-projective acted on by a linear algebraic group Hγ with internally graded unipotent radical, and there is a geometric STRATIFYING QUOTIENT STACKS AND MODULI STACKS 5 quotient Wγ/Hγ which is a coarse moduli space for Σγ. The geometric quotient Wγ/Hγ will be projective if semistability coincides with stability in an appropriate sense for the action of Hγ on a suitable projective completion of Wγ with respect to an induced linearisation. The layout of this article is as follows. In §2, we will review classical and non-reductive GIT, describing how to construct quotients by actions of linear algebraic groups with internally graded unipotent radical. The heart of the paper is §3, in which we describe how to stratify a quotient stack [X/H] into strata Σγ = [Wγ/Hγ] where the action of Hγ on Wγ has a geometric quotient Wγ/Hγ. The argument is an inductive one, so the assumption on X and H is that H is a linear algebraic group with internally graded unipotent radical, and X is a projective scheme which has an amply linearised action of H. In §4, this construction is applied to stacks which are suitably filtered by quotient stacks, and §5 contains a brief discussion of examples including moduli of unstable curves and moduli of sheaves of given Harder–Narasimhan type over a fixed projective scheme. We would like to thank Brent Doran, Daniel Halpern-Leistner, Eloise Hamilton and Joshua Jackson for helpful discussions about this material. 2. CLASSICAL AND NON-REDUCTIVE GIT 2.1. Classical GIT for reductive groups. In Mumford's GIT [26], a linearisation of an action of a reductive group G on a projective scheme X over an algebraically closed field k of character- istic 0 is given by a line bundle L (which we will always assume to be ample) on X and a lift of the action to L. Since G is reductive, the algebra of G-invariant sections bOL(X)G is finitely generated as a graded algebra with associated projective scheme X//G = Proj(bOL(X)G). k=0 H 0(X, L⊗k) := L∞ The inclusion of bOL(X)G in bOL(X) determines a rational map X − − → X//G which fits into a X −− → X//G projective diagram algebra of invariants. ↓ (X, L) ; bOL(X) S X//G ; bOL(X)G S S semistable X ss −−−−։ X//G stable X s −−−−։ X s/G open S where X s (the stable locus) and X ss (the semistable locus) are open subschemes of X, there is a geometric quotient X s/G for the action of G on X s, the GIT quotient X//G is a good quotient for the action of G on X ss via the G-invariant surjective morphism φ : X ss → X//G, and φ(x) = φ(y) ⇔ Gx ∩ Gy ∩ X ss 6= ∅. The semistable and stable loci X ss and X s of X are characterised by the following properties ([26, Chapter 2], [28]). Proposition 2.1 (Hilbert–Mumford criteria for reductive groups). Let T be a maximal torus of G. (1) A point x ∈ X is semistable (respectively stable) for the G-action on X if and only if for every g ∈ G the point gx is semistable (respectively stable) for the T -action. STRATIFYING QUOTIENT STACKS AND MODULI STACKS 6 (2) A point x ∈ X ⊂ Pn with homogeneous coordinates [x0 : . . . : xn] is semistable (respectively stable) for a diagonal T -action on Pn with weights α0, . . . , αn if and only if (respectively 0 is contained in the interior of this convex hull). 0 ∈ Conv{αi : xi 6= 0} 2.1.1. The HKKN stratification. Associated to the linear action of G on X and an invariant inner product on the Lie algebra of G, there is a stratification (the 'HKKN stratification', which in the case k = C is the Morse stratification for the norm-square of an associated moment map [18, 22, 23, 27]). Remark 2.2. Let us clarify what is meant by this invariant inner product, whose associated norm we denote by − . If k = C, then G is the complexification of its maximal compact group K; then the Lie algebra of K is a real vector space, and we choose an inner product on this Lie algebra that is invariant under the adjoint action of K. In fact, we will also assume that we fix a maximal compact torus Tc ⊂ K such that the inner product is integral on the co- character lattice X∗(Tc) ⊂ Lie K. For an arbitrary algebraically closed field k of characteristic zero, one can fix a maximal torus T of G and choose an inner product on the co-character space X∗(T ) ⊗Z R that is invariant for the Weyl group of T and is integral on the co-character lattice (for example, see [19, §2]). Then this inner product gives an identification between characters and co-characters (i.e. 1-PSs) of T . The HKKN stratification associated to the action of G on X with respect to L and the norm − is a stratification X = Gβ∈B Sβ of X by locally closed subschemes Sβ, indexed by a partially ordered finite subset B of a rational elements in a positive Weyl chamber for the reductive group G, with the following properties. (1) If 0 ∈ B, then this is the minimal element and S0 = X ss. Moreover, for each β ∈ B, we additionally have the following properties. (2) the closure of Sβ is contained inSβ ′>β Sβ ′ where γ > β if and only if γ = β or γ > β; β )/Pβ where this quotient is of the diagonal action of Pβ on (3) Sβ ∼= G ×Pβ Y ss β := (G × Y ss the right on G and on the left on Y ss β . Here Pβ is a parabolic subgroup of G which acts on a locally closed subscheme Y ss β of X. More precisely, β ∈ B determines a (rational) 1-PS λβ : Gm → G and an associated parabolic subgroup Pβ = P (λβ) = Uβ ⋊ Lβ with Levi subgroup Lβ = StabG(β) such that the conju- gation action of λβ(Gm) on Lie Uβ has strictly positive weights; thus Pβ has internally graded unipotent radical. Let Zβ be the union of components in the fixed locus X λβ (Gm) on which this 1-PS acts on the fibres of L with weight given by the restriction of β, and let Yβ ⊂ X be the subscheme of points x ∈ X such that limt→0 λβ(t)y ∈ Zβ; thus there is a retraction pβ : Yβ → Zβ y 7→ pβ(y) := lim t→0 λβ(t)y which is equivariant with respect to the quotient homomorphism qβ : Pβ → Lβ obtained by identifying Lβ with Pβ/Uβ. Let Lβ denote the restriction of the G-linearisation L on X to the Pβ-action on Y β twisted by the (rational) character χβ of Pβ corresponding to the 1-PS λβ (via the norm − ). We also let Lβ denote the restriction of this linearisation to the Lβ-action on STRATIFYING QUOTIENT STACKS AND MODULI STACKS 7 (respectively Z ss Zβ. Then Y ss β the Lβ/λβ(Gm)-action on Zβ) linearised by the twisted linearisation Lβ; furthermore, Y ss p−1 β (Z ss β ). Finally, we make the following observation about quotienting the unstable strata (cf. [20]). β ) is the semistable locus for the Pβ-action on Y β (respectively β = Remark 2.3. The G-action on Sβ has a categorical quotient Zβ//Lβ induced by the map pβ : Y ss β → Z ss β . In general, this quotient is far from being a geometric quotient (even after restriction to the pre-image of any nonempty open subscheme of Zβ//Lβ), as y ∈ Y ss β is identified with pβ(y) ∈ Gy. By (3), constructing a quotient of the G-action on a G-invariant open subset of Sβ is equiv- alent to constructing a Pβ-quotient of a Pβ-invariant open subset of Y ss (or its closure); the β latter perspective will lead to a geometric quotient by using GIT for the non-reductive group Pβ, whose unipotent radical Uβ is internally graded by λβ (cf. §2.3). 2.1.2. Partial desingularisations of reductive GIT quotients. The geometric quotient X s/G has at most orbifold singularities when X is nonsingular, since the stabiliser subgroups of stable If X ss 6= X s 6= ∅, the singularities of X//G are typically points are finite subgroups of G. eX//G = eX ss/G which is also a projective completion of X s/G and is itself a geometric quotient more severe even when X is itself nonsingular, but X//G has a 'partial desingularisation' eX//G by G of an open subscheme eX ss = eX s of a G-equivariant blow-up eX of X [24]. Here eX ss For the construction of the partial desingularisation eX//G in [24], it is assumed that X s 6= ∅. is obtained from X ss by successively blowing up along the subschemes of semistable points stabilised by reductive subgroups of G of maximal dimension and removing the complement of the semistable locus from the resulting blow-up. There exist semistable points of X which are not stable if and only if there exists a non-trivial connected reductive subgroup of G fixing a semistable point. Let r > 0 be the maximal dimen- sion of a reductive subgroup of G fixing a point of X ss and let R(r) be a set of representatives of conjugacy classes of all connected reductive subgroups R of dimension r in G such that Z ss R := {x ∈ X ss : Rx = x} is non-empty. Then Z ss R(r) := [R∈R(r) GZ ss R (1) in X(1) is the proper transform of the closed subscheme π−1(π(GZ ss is a disjoint union of closed G-invariant subschemes of X ss. The action of G on X ss lifts to an action on the blow-up X(1) of X ss along Z ss R(r), and this action can be linearised so that the complement of X ss R )) of X ss where π : X ss → X//G is the quotient map (see [24, 7.17]). The G-linearisation on X(1) used here is (a tensor power of) the pullback of the ample line bundle L on X along ψ(1) : X(1) → X perturbed by a sufficiently small multiple of the exceptional divisor E(1); then, if the perturbation is sufficiently small, we have (1) ⊆ X ss (1)(X ss) = X(1), ψ−1 (1)(X s) ⊆ X s (1) ⊆ ψ−1 STRATIFYING QUOTIENT STACKS AND MODULI STACKS 8 and the stable and semistable loci X s tion. Moreover, no point x ∈ X ss and x ∈ X ss belongs to the proper transform of the closed subscheme Z ss (1) will be independent of the choice of perturba- (1) is fixed by a reductive subgroup of G of dimension at least r, (1) is fixed by a reductive subgroup R of dimension less than r in G if and only if it (1) and X ss Remark 2.4. In [24], X itself is blown up along the closure Z ss R(r) in X (or in a projective completion of X ss with a G-equivariant morphism to X which is an isomorphism over X ss). This gives a projective scheme X (1) and blow-down map ψ(1) : X (1) → X restricting to ψ(1) : X(1) → X where (ψ(1))−1(X ss) = X(1). For a sufficiently small perturbation of the pullback to ss X (1) of the linearisation on X, we have (ψ(1))−1(X s) ⊆ X (1) ⊆ (ψ(1))−1(X ss) = X(1), and moreover the restriction of the linearisation to X(1) is obtained from the pullback of L by perturbing by a sufficiently small multiple of the exceptional divisor E(1). s (1) ⊆ X R of X ss. R(r) of Z ss If r > 1, we can apply the same procedure to X ss (2) such that no point of X ss (2) is fixed by a reductive subgroup of G of dimension at least r − 1. If X s 6= ∅ then repeating this (1) to obtain X ss process at most r times gives us ψ : eX ss → X ss such that ψ is an isomorphism over X s and no positive-dimensional reductive subgroup of G fixes a point of eX ss. The partial desingular- isation eX//G = eX ss/G can be obtained by blowing up X//G along the proper transforms of R ) ⊂ X//G in decreasing order of the dimension of R. π(GZ ss Remark 2.5. Suppose for simplicity that X is irreducible. a) If X s 6= ∅, then this is the situation considered in [24] and the partial desingularisation construction is described above. If X ss = X s, then eX = X. b) If X ss = ∅, then there is an unstable stratum Sβ with β 6= 0 in the HKKN stratification (cf. §2.1.1) which is a non-empty open subscheme of X, and thus when X is irreducible X = Sβ. Then constructing a quotient of a non-empty open subscheme of X reduces to non-reductive GIT for the action of the parabolic subgroup Pβ on Yβ as described in §2.3 below, where a blow-up sequence may also need to be performed. c) If X s = ∅ 6= X ss then the partial desingularisation construction can be applied to X ss, and there are different ways in which it can terminate. (i) If X ss = GZ ss R R for a positive-dimensional connected reductive sub- group R of G with normaliser NR in G, then NR and NR/R are also reductive, and ∼= G ×NR Z ss X//G ∼= ZR//NR ∼= ZR//(NR/R) (ii) If GZ ss where ZR is the closed subscheme of X which is the fixed point set for the action of R. Then we can apply induction on the dimension of G to study this case. R 6= X ss for each positive-dimensional connected reductive subgroup R of G, then we can perform the first blow-up in the partial desingularisation construction to obtain ψ(1) : X(1) → X ss such that X ss (1) = ∅ as above (as ∼= X s = ∅, where E(1) is the exceptional divisor). If X s (1) is open and X s X ss (1) = ∅, then X(1) has a dense open stratum S(1),β for β 6= 0 as in Case b). If we have X ss (1),R for a positive-dimensional connected reductive subgroup R of G, where Z ss (1) : Rx = x}, then we proceed as in Case (i) above. Otherwise we can repeat the process, until it terminates in one of these two ways. (1) ⊆ X(1) and X s (1),R = {x ∈ X ss (1) = GZ ss (1) \ E(1) STRATIFYING QUOTIENT STACKS AND MODULI STACKS 9 2.2. GIT for non-reductive groups. Now suppose that X is a projective scheme over an alge- braically closed field k of characteristic 0 and let H be a linear algebraic group, with unipotent radical U , acting on X with respect to an ample linearisation L. Definition 2.6. (Semistability for the unipotent group cf. [14, §4] and [14, 5.3.7]). For an invari- which f does not vanish, and let O(Xf ) denote its coordinate ring. ant section f ∈ I =Sm>0 H 0(X, L⊗m)U , let Xf be the U -invariant affine open subset of X on (1) The semistable locus for the U -action on X linearised by L is X ss,U =Sf ∈I fg Xf where (2) The (locally trivial) stable locus for the linearised U -action on X is X lts,U = Sf ∈I lts Xf where I lts := {f ∈ I fg qU : Xf → Spec(O(Xf )U ) is a locally trivial geometric quotient}. I fg = {f ∈ I O(Xf )U is finitely generated}. (3) The enveloped quotient of X ss,U by the linear U -action is qU : X ss,U → qU (X ss,U ), where qU : X ss,U → Proj(bOL(X)U ) is the natural morphism of schemes and qU (X ss,U ) is a dense constructible subset of the enveloping quotient X≈U = [f ∈I f g Spec(O(Xf )U ). Remark 2.7. (1) Even when bOL(X)U is finitely generated, so that X≈U = Proj(bOL(X)U ), the enveloped quotient qU (X ss,U ) is not necessarily a subscheme of X≈U (for example, see [14, §6]). (2) The enveloping quotient X ≈ U has quasi-projective open subschemes ('inner envelop- ing quotients' X//◦U ) that contain the enveloped quotient qU (X ss) and have ample line bundles which under the natural map qU : X ss → X ≈ U pull back to positive tensor powers of L (see [2] for details). The H-semistable locus X ss = X ss,H and enveloped and (inner) enveloping quotients qH : X ss → qH(X ss) ⊆ X//◦H ⊆ X≈H for the linear action of H are defined as for the unipotent case in Definition 2.6 and Remark 2.7 (cf. [2]), but the definition of the stable locus X lts = X lts,H for the linear action of H combines (and extends) the definitions for unipotent and reductive groups. Definition 2.8. (Stability for linear algebraic groups). Let H be a linear algebraic group acting linearly on X with respect to an ample line bundle L; then the (locally trivial) stable locus is the open subscheme X lts = Sf ∈I lts Xf of X ss, where I lts ⊆ Sr>0 H 0(X, L⊗r)H is the subset of H-invariant sections f satisfying the following conditions: (1) the H-invariant open subscheme Xf is affine; (2) the H-action on Xf is closed with finite stabiliser groups; and (3) the restriction of the U -enveloping quotient map is a locally trivial geometric quotient for the U -action on Xf . qU : Xf → Spec((bOL(X)U )(f )) STRATIFYING QUOTIENT STACKS AND MODULI STACKS 10 2.3. GIT for linear algebraic groups with internally graded unipotent radicals. Now sup- pose that H = U ⋊ R is a linear algebraic group with internally graded unipotent radical U in the following sense. Definition 2.9. We say H = U ⋊ R has internally graded unipotent radical U if there is a central 1-PS λ : Gm → R whose conjugation action on the Lie algebra of U has strictly positive weights. We write bU = U ⋊ λ(Gm) 6 H for the associated semi-direct product. in [3, 4] that the algebra of H-invariant sections is finitely generated provided: Suppose also that H acts linearly on X with respect to an ample line bundle L. It is shown a) L is replaced with a suitable tensor power L⊗m, with m ≥ 1 sufficiently divisible, and the linearisation of the action of H is twisted by a suitable (rational) character, and b) condition (∗) described below (also known as 'semistability coincides with stability for the unipotent radical U ') holds. Moreover, in this situation the natural quotient morphism qH from the semistable locus X ss,H to the enveloping quotient X ≈ H is surjective, and expresses the projective scheme X//H = X//◦H = X ≈ H as a good quotient of X ss,H. Furthermore this locus X ss,H can be described using Hilbert–Mumford criteria. It is also shown in [4] that when condition (∗) is not satisfied, but is replaced with a slightly weaker condition, such as (∗∗) below, then there is a sequence of blow-ups of X along H-invariant subschemes (similar to that of [24] when H is reductive) resulting in a projective scheme bX with an induced linear action of H satisfying condition (∗). In fact, these results can be generalised to allow actions where the U -action has positive dimensional stabilisers (cf. Theorems 2.17 and 2.18 below). Before giving a precise description of the condition (∗) and its variants, we define the notion of an adapted linearisation, which is also needed for the statement of the main results of [3, 4]. can be identified naturally with an integer so that the integer 1 corresponds to the character high power, we can without loss of generality assume that L is very ample. Let ωmin := ω0 < Let χ : H → Gm be a character of H; the restriction to bU of χ contains U in its kernel and of bU which fits into the exact sequence U ֒→ bU ։ λ(Gm). By replacing L with a sufficiently ω1 < · · · < ωmax be the weights with which the 1-PS λ : Gm → bU ≤ bH acts on the fibres of the tautological line bundle OP((H 0(X,L)∗)(−1) over points of the fixed locus P((H 0(X, L)∗)λ(Gm). Without loss of generality we may assume that there exist at least two distinct such weights, as otherwise the U -action on X is trivial, in which case we can take a quotient by the action of the reductive group R = H/U . Definition 2.10. For a character χ of H as above and a positive integer c, we say the rational (2.1) χ c ωmin := ω0 < character χ/c is adapted to the linear action of bU if Furthermore, we say L is adapted to the bU -action if ωmin := ω0 < 0 < ω1. If the rational character χ/c is adapted to the linear action of bU , then the H-linearisation L⊗c χ on X given by twisting the ample line bundle L⊗c by the character χ (that is, so that the weights ωj are replaced with ωjc−χ) is adapted. Let X s,Gm min + denote the stable set in X for the linear action of Gm via λ with respect to the adapted linearisation L⊗c χ and, for a maximal torus T of H containing λ(Gm), let X s,T min + denote the stable set in X for the linear action of T with min + = X s,λ(Gm) < ω1. STRATIFYING QUOTIENT STACKS AND MODULI STACKS 11 respect to the adapted linearisation L⊗c the stable set X s,λ(Gm) by the Hilbert–Mumford criterion, we have X s,λ(Gm) weight space of ωmin in V = H 0(X, L)∗, then χ . By the theory of variation of (classical) GIT [13, 34], is independent of the choice of adapted rational character χ/c. In fact, min \ Zmin, where if Vmin denotes the min + = X 0 min + Zmin := X ∩ P(Vmin) =nx ∈ X λ(Gm) : λ(Gm) acts on L∗x with weight ωmino and X 0 min := {x ∈ X lim t→0, t∈Gm λ(t) · x ∈ Zmin}. Definition 2.11. (Conditions (∗)−(∗∗∗) generalising 'semistability equals stability' cf. [4]) With the above notation, we define the following conditions for the bU -action on X. StabU (z) = {e} for every z ∈ Zmin. (∗) (∗∗) StabU (x) = {e} for generic x ∈ X 0 min. Moreover, if U > U (1) > ... > U (s) > {e} denotes the derived series of U , we define condition (∗ ∗ ∗) for 1 ≤ j ≤ s, there exists dj ∈ N such that dim StabU (j)(x) = dj for all x ∈ X 0 Note that condition (∗) holds if and only if we have StabU (x) = {e} for all x ∈ X 0 min. This condition is also referred to in [4] as the condition 'semistability coincides with stability' for the min. X s, bU min \ U Zmin. X s,H hX s,T min +. min + := \h∈H Definition 2.12. Let T ≤ R be a maximal torus containing λ(Gm). The minimal stable set for a action of bU (or, when the 1-PS λ : Gm → R is fixed, for the linear action of U ). linear H-action for which condition (∗) holds for the action of the graded unipotent group bU is We note that the minimal stable set for the bU-action satisfies Theorem 2.13. (bU -Theorem when semistability coincides with stability for bU ) [4] Suppose that the linearisation for the action of H on X is adapted as in Definition 2.10 and that the bU -action on X min + of X has a projective geometric quotient X//bU = X s, bU min +/bU by bU . min + of X has a good quotient X//H = (X//bU )//R by H = U ⋊ R, (i) The open subscheme X s, bU (ii) The open subscheme X s,H satisfies condition (∗). Then the following statements hold. min + = \u∈U Remark 2.14. In the proof of Theorem 2.13 (and its variants below), one replaces the adapted linearisation by a 'well adapted' linearisation (which can be achieved by twisting by a ra- tional character); this is a slightly stronger notion. This strengthening does not alter X s, bU uX s,λ(Gm) min + = X 0 which is also projective. or its quotient X//bU , but it affects what can be said about induced ample line bundles on min + STRATIFYING QUOTIENT STACKS AND MODULI STACKS 12 X//bU and X//H = (X//bU )//(R/λ(Gm)). The proofs in [3, 4] that the algebras of invariants ⊕m≥0H 0(X, L⊗cm mχ ) bU and H 0(X, L⊗cm mχ ) bU )(R/λ(Gm)) Mm≥0 H 0(X, L⊗cm mχ )H = (Mm≥0 are finitely generated, and that the enveloping quotients X//bU = X s, bU the associated projective schemes, require that the linearisation is twisted by a well adapted rational character χ/c. More precisely, it is shown in [3, 4] that, given a linear action of H on X with respect to an ample line bundle L, there exists ǫ > 0 such that if χ/c is a rational character of Gm (lifting to H) with c sufficiently divisible and χ : H → Gm a character of H such that min +/bU and X//H are then the algebras of invariants ⊕m≥0H 0(X, L⊗cm mχ ) bU and ⊕m≥0H 0(X, L⊗cm mχ )H are finitely gener- ωmin < χ c < ωmin + ǫ, 2.13. Theorem 2.13 describes the good case when semistability coincides with stability for the the conditions of Theorem 2.13 still hold for a suitable ample linearisation, and such that the quotient on X is adapted and satisfies condition (∗∗). Then there exists a sequence of blow-ups along H-invariant Remark 2.16. This blow-up sequence is constructed in two stages: one first blows up by consid- ering the stabiliser subgroups for the unipotent group U , and then one blows up by considering the stabiliser subgroups for the reductive group R. ated, and the associated projective schemes X//bU and X//H satisfy the conclusions of Theorem linear action of bU . The following versions proved in [4] apply more generally. Theorem 2.15. (bU -Theorem giving projective completions) [4] Suppose that the linear action of bU projective subschemes, resulting in a projective scheme eX (with blow-down map eψ : eX → X) such that given by that theorem is a geometric quotient of an open subscheme eX s,H of eX. In the first step, one performs a blow-up sequence to obtain bψ : bX → X such that the bU - action on bX satisfies condition (∗) with respect to a linearisation bL, which is an arbitrarily small perturbation of bψ∗(L). The centres of the blow-ups used to obtain bX from X are determined by Then one can construct a projective and geometric quotient of thebU -action on bX s, bU by Theorem 2.13 and, as bψ is an isomorphism away from the exceptional divisor, one obtains a geometric quotient of a bU -invariant open subset X s, bU of X as an open subscheme of bX// bLbU , where X s, bU is the image under bψ of the intersection of bX s, bU with the complement of the exceptional divisor in bX. Another characterisation of this stable locus is as X s, bU = {x ∈ ψ(bX s, bU ) dim StabU (lim If one is only interested in obtaining a good quotient for the H-action, then the second stage of the blow-up procedure is not needed: one can then take a reductive GIT quotient of the the dimensions of the stabilisers in U of the limits limt→0 λ(t)·x for x ∈ X 0 min (for details, see [4]). λ(t) · x) = 0}. t→0 residual action on bX//bLbU of H/bU = R/λ(Gm), and thus one obtains a good quotient of the H-action on an open subset of X as an open subscheme of (bX// bLbU )//(R/λ(Gm)). Moreover, this good quotient restricts to a geometric quotient on an open subscheme of stable points. STRATIFYING QUOTIENT STACKS AND MODULI STACKS 13 This gives us an H-invariant open subscheme X s,H of X with a geometric quotient by H sequence by considering the stabiliser groups for the action of the reductive group R/λ(Gm) as in the partial desingularisation procedure described in §2.1.2. Finally, to go from bX to the blow-up eX in Theorem 2.15, one performs an additional blow up which is an open subscheme of the projective scheme eX//eLH (and also of bX//bLH). Here X s,H is the image under bψ of the intersection of bX s,H with the complement of the exceptional divisor in bX. Theorem 2.17. (bU -Theorem with positive-dimensional stabilisers in U ) [4] Suppose that condi- tion (∗ ∗ ∗) holds for an adapted linear bU -action on X. Then the conclusions of Theorem 2.13 hold. In fact, this theorem still holds if we replace the derived series in condition (∗ ∗ ∗) with any series U > U (1) > ... > U (s) > {e} which is normalised by H and whose successive quotients U (j)/U (j+1) are abelian, provided that (∗ ∗ ∗) holds for this series. Theorem 2.13 can be generalised by weaking the condition (∗) further to (∗ ∗ ∗) to allow for actions with positive dimensional stabiliser groups generically. Finally, there is a version of the theorem without requiring any hypothesis related to semista- bility coinciding with stability. pletions) [4] For a linear H-action on X with respect to an adapted ample linearisation L, there is a Theorem 2.18. (bU -Theorem with positive-dimensional stabilisers in U , giving projective com- sequence of blow-ups along H-invariant projective subschemes, resulting in a projective scheme eX such 2.13 is a geometric quotient of an open subscheme eX s,H of eX. that condition (∗ ∗ ∗) holds for a suitable linearisation, and such that the H-quotient given by Theorem This theorem provides a non-reductive analogue of the partial desingularisation construc- tion for reductive GIT described at the end of §2.1. As before, this gives us an H-invariant open subscheme X s,H of X with a geometric quotient by H which is open in the projective scheme X//LH; here X s,H is the image under ψ of the considering stabiliser subgroups for the unipotent action, then one constructs a further blow- reductive subgroup R/λ(Gm). The slight difference is that in the first step, the centres of the intersection of eX s,H with the complement of the exceptional divisor in eX. Remark 2.19. As for Theorem 2.15 (cf. Remark 2.16), one first constructs a blow-up bX → X by up sequence eX → bX by considering the stabiliser subgroups for the residual action of the blow-ups used to obtain bX from X are determined by the dimensions of the U (j)-stabilisers of the limit limt→0 λ(t) · x and of x itself for x ∈ X 0 min. We will see in the next section that by applying Theorem 2.18 to the closures of the subschemes where these dimensions take dif- ferent values, and combining this with the partial desingularisation construction of [24] for reductive GIT quotients, X can be stratified so that each stratum is a locally closed H-invariant subscheme of X with a geometric quotient by the action of H. 3. STRATIFYING QUOTIENT STACKS Let H = U ⋊ R be a linear algebraic group with internally graded unipotent radical U (cf. Definition 2.9) acting on a projective scheme X over an algebraically closed field k of charac- teristic 0 with respect to an ample linearisation L. We fix an invariant inner product on the Lie algebra Lie R of the Levi factor, just as in the construction of the HKKN stratification (cf. §2.1.1). STRATIFYING QUOTIENT STACKS AND MODULI STACKS 14 The aim of this section is to prove Theorem 1.1 stated in §1. We will prove this result using a recursive argument involving the dimensions of X and of H and the number of irreducible components of X. The idea will be to start by defining a 'minimum' stratum, which will be a non-empty H-invariant open subscheme of X, and then proceed recursively. We will assume that H is connected and that X is reduced. Remark 3.1. These assumptions do not involve any significant loss of generality. (1) The HKKN stratification {Sβ β ∈ B} is usually indexed by a finite subset B of a positive Weyl chamber. Then the strata are not necessarily connected even when X is irreducible, and it is often useful to refine the stratification so that the strata are the connected components of Sβ, or are unions of some but not all of these connected com- ponents (cf. [23]). There is a similar ambiguity in the construction of the refined strati- fication defined in this section: at some points we take connected components, but this is not crucial to the definition. Indeed if we wish to allow the group H to be discon- nected then we cannot assume that the strata are connected since they are required to be H-invariant. Then instead of taking connected components (which will be invariant under the component H0 of the identity in H), we can take their H-sweeps, which will be disjoint unions of at most H/H0 of these connected components. (2) If X is non-reduced, then we can define the stratification on X by using a positive power of L to define a H-equivariant embedding of X into a projective space Pn, and then take the fibre product of X with the stratification on Pn. Indeed, we will see that the stratification is functorial for equivariant closed immersions (this is essentially the third statement in Theorem 1.1). This follows as the reductive notions of GIT (semi)stability are functorial and since in the non-reductive case, we are assuming that H has inter- nally graded unipotent radical. The stable loci when H has internally graded unipotent radical and adapted linearisation are also functorial, as they have Hilbert–Mumford style descriptions (see Definition 2.12). We note that in the more general non-reductive GIT set up described in §2.2 the notions of (semi)stability are not functorial, as taking H-invariants is not exact, and so there can be invariants which do not extend to the ambient space. The advantage of assuming that X is reduced is that the complement to the open stratum then has a canonical scheme structure; thus it is easier to recursively define the stratification. Let us describe the recursive construction when H is connected and X is reduced. For each linear action on X of H = U ⋊ R with internal grading λ : Gm → R and linearisation L with underlying ample line bundle L, we will first use recursion to define a nonempty H-invariant open subscheme S0(X, H, λ, L) of X that admits a geometric quotient S0(X, H, λ, L)/H; this will be done by considering seven different cases. After defining the open stratum, we will define the stratification {Sγγ ∈ Γ} of X with strata Sγ = Sγ(X, H, λ, L) and index set Γ = Γ(X, H, λ, L) by letting X1, . . . , Xk be the connected components of the projective subscheme of X equal to the complement of S0(X, H, λ, L), letting S0,i(X, H, λ, L) for 1 6 i 6 m be the connected components of S0(X, H, λ, L), and setting (3.1) Γ(X, H, λ, L) := {0} × {1, . . . , m} ∪ [16j6k {Xj} × Γ(Xj, H, λ, LXj ). The strata indexed by (0, i) ∈ Γ(X, H, λ, L) for 1 6 i 6 m are then the connected components of the open subscheme the open subscheme S0(X, H, λ, L) constructed using the case by case STRATIFYING QUOTIENT STACKS AND MODULI STACKS 15 argument below, whereas the stratum indexed by an element (Xj , γ) for 1 ≤ j ≤ k and γ ∈ Γ(Xj, H, λ, LXj ) is (3.2) S(Xj ,γ)(X, H, λ, L) := Sγ(Xj , H, λ, LXj ), where the strata Sγ(Xj, H, λ, LXj ) are constructed by induction. The partial order on Γ then naturally comes from the partial orders on each Γ(Xj, H, λ, LXj ) with (0, i) < (Xj, γ) for all 1 6 i 6 m and 1 6 j 6 k and γ ∈ Γ(Xj, H, λ, LXj ). Let us now describe how to define S0(X, H, λ, L) in the seven different cases. Let U = U (0) > U (1) = [U, U ] > ... > U (s) > {e} be the derived series of U . Let Z d0 min = {x ∈ Zmin dim(StabU (x)) = d0} where d0 is the minimal value of dim(StabU (x)) for x ∈ Zmin. Then Z d0 invariant open subscheme of Zmin, and U Z d0 Case 1. First assume that the central one-parameter subgroup λ : Gm → R of R has at least two distinct weights for the linear action of H on X; equivalently Zmin is a projective subscheme of X with Zmin 6= X. Under this assumption we have two possibilities to consider. Case 1(a). Suppose that U Z d0 min is a nonempty locally closed subscheme of X. min is a nonempty H- min is open in X. Then we define S0(X, H, λ, L) = U {x ∈ S0(Zmin, R/λ(Gm), λ0, LZmin) dim(StabUj (x)) = dj for 1 6 j 6 s} where λ0 is the trivial one-parameter subgroup that grades the trivial unipotent radical of the reductive group R/λ(Gm) and dj := min{dim(StabUj (x)) : x ∈ S0(Zmin, R/λ(Gm), λ0, LZmin)}. By induction we can assume that S0(Zmin, R/λ(Gm), λ0, LZmin) is a nonempty R-invariant open subscheme of Zmin with a geometric quotient by the action of R/λ(Gm) (or equivalently by the action of R, since the central one-parameter subgroup λ(Gm) of R acts trivially on Zmin). Indeed, we can construct such an open subscheme as in Case 2 described below. Thus S0(X, H, λ, L) is an H-invariant nonempty open subscheme of X, and by the proof of Theorem 2.18 (see [4, Remark 2.10]) it has a geometric quotient S0(X, H, λ, L)/H ∼= S0(Zmin, R/λ(Gm), λ0, LZmin)/R = S0(Zmin, R/λ(Gm), λ0, LZmin)/(R/λ(Gm)). Case 1(b). Suppose that U Z d0 subscheme of X and that the morphism p : X 0 min is not open in X. Recall that X 0 min → Zmin defined by min is an H-invariant open p(x) = lim t→0 λ(t)x satisfies p(urx) = rp(x) for all x ∈ X 0 min, u ∈ U and r ∈ R (cf. [4]). We set S0(X, H, λ, L) =(cid:26)x ∈ X 0 min \ U Zmin p(x) ∈ S0(Zmin, R/λ(Gm), λ0, LZmin) and for 1 6 j 6 s dim(StabUj (x)) = δj and dim(StabUj (p(x))) = dj where dj and δj are defined inductively, in decreasing order of j, as follows. Assuming that di and δi are defined for i > j, we define dj := min(cid:26)dim(StabUj (z)) ∃ x ∈ X 0 min with z = p(x) ∈ S0(Zmin, R/λ(Gm), λ0, LZmin) and dim(StabUi(z)) = di for all i > j (cid:27) (cid:27) STRATIFYING QUOTIENT STACKS AND MODULI STACKS 16 and δj := min(cid:26)dim(StabUj (x)) x ∈ X 0 min such that p(x) ∈ S0(Zmin, R/λ(Gm), λ0, LZmin) and dim(StabUi(x)) = δi for all i > j (cid:27) , where for j = s, the conditions appearing in the second line of these definitions are empty. Note that if X and Zmin are irreducible then we can define dj and δj much more simply as the min, but in generic values taken by dim(StabUj (z)) for z ∈ Zmin and by dim(StabUj (x)) for x ∈ X 0 this approach to defining the stratification {Sγγ ∈ Γ} we need to allow X to be reducible. Using the proof of Theorem 2.18 again (cf. [4, Remark 2.10]), we have that S0(X, H, λ, L) is an H-invariant nonempty open subscheme of X with a geometric quotient S0(X, H, λ, L)/bU , and by the inductive construction a geometric quotient S0(X, H, λ, L)/H = (S0(X, H, λ, L)/bU )/(R/λ(Gm)). Case 2. Now assume that the central one-parameter subgroup λ : Gm → R which grades U acts trivially on X, so the unipotent radical U of H = U ⋊ R must also act trivially on X. Then without loss of generality we can assume that H = R is reductive and that λ = λ0 is trivial. Case 2(a). Suppose that the stable locus X s,R for the linear action of R on X is nonempty. Then we let S0(X, H, λ, L) = X s,R and by classical GIT this has a geometric quotient X s,R/H = X s,R/R. Case 2(b). Suppose that the semistable locus X ss,R for the linear action of R on X is empty. Then there is a stratum Sβ from the HKKN stratification for X (associated to our invariant inner product on R) such that β 6= 0 and Sβ = HY ss β = RY ss β ∼= R ×Pβ Y ss β is nonempty and open in X (see §2.1.1 and Remark 2.5). Then we have the following two subcases to consider. Case 2(b)i). Suppose that Y ss components of X, β 6= X. Then by induction on dim X and the number of irreducible S0(X, H, λ, L) = R(S0(Y ss β , Pβ, λβ, LY ss β ) ∩ Y ss β ) ∼= R ×Pβ (S0(Y ss β , Pβ, λβ, LY ss β ) ∩ Y ss β ) is a nonempty R-invariant (and hence H-invariant) open subscheme of X and has a geometric quotient S0(X, H, λ, L)/H = S0(X, H, λ, L)/R ∼= (S0(Y ss β , Pβ, λβ, LY ss β ) ∩ Y ss β )/Pβ. Case 2(b)ii). Suppose that Y ss β = X. Then Pβ = R so β defines a rational character of R and corresponds to a nontrivial central one-parameter subgroup λβ : Gm → R. If Zβ 6= X then the one-parameter subgroup λβ(Gm) acts nontrivially on X, and thus we can use Case 1 and induction to define S0(X, H, λ, L) = S0(X, R, λβ, L) so that it is a nonempty R-invariant (and hence H-invariant) open subscheme of X and has a geometric quotient S0(X, H, λ, L)/H = S0(X, R, λβ, L)/R. STRATIFYING QUOTIENT STACKS AND MODULI STACKS 17 If Zβ = X then λβ(Gm) acts trivially on X and we can use induction on the dimension of H to define S0(X, H, λ, L) = S0(X, R/λβ(Gm), λ0, L) which is a nonempty R-invariant (and hence H-invariant) open subscheme of X with geomet- ric quotient S0(X, H, λ, L)/H = S0(X, R/λβ (Gm), λ0, L)/(R/λβ(Gm)). Case 2(c). Suppose now that X s,R = ∅ 6= X ss,R. Recall from Remark 2.5 that the partial desingularisation construction [24] for the linear action of the reductive group R on X can be applied to X ss,R, although, for simplicity, X was assumed to be irreducible in Remark 2.5, which is no longer the case here. This construction terminates with a birational projective morphism ψ : eX → X and R-linearisation eL for an ample line bundle eL on eX which is a positive integer multiple of ψ∗L ⊗ O(−ǫE) where E is the exceptional divisor and 0 < ǫ << 1 is rational. Then we have the following two subcases to consider. Case 2(c)i). Suppose that there is a positive-dimensional connected reductive subgroup R′ of R such that R′ ∼= R ×NR′ eZ ss ReZ ss R′ eZ ss R′ = {x ∈ eX ss,R R′x = x} = eZR′ ∩ X ss,R R′ coincides with both the semistable locus and the stable locus for the induced action of NR′ /R′ is open and nonempty in eX ss,R. Here NR′ is the normaliser of R′ in R and with eZR′ the union of those connected components of the fixed point set eX R′ trivially on the fibres of eL. Moreover NR′ and its quotient group NR′/R′ are reductive, and eZ ss on eZR′. In this situation ReZ ss invariant) open subscheme of eX \ E with a geometric quotient Since ψ restricts to an H-equivariant isomorphism from eX \ E to an open subscheme of X, it R′ \ E ∼= R ×NR′ (eZ ss R′ \ E) is a nonempty R-invariant (and hence H- over which R′ acts follows that if we set R′ \ E)/NR′ . R′ \ E) then S0(X, H, λ, L) is a nonempty R-invariant (and hence H-invariant) open subscheme of X with a geometric quotient R′ \ E)/R ∼= (eZ ss (ReZ ss S0(X, H, λ, L) = ψ(ReZ ss S0(X, H, λ, L)/H ∼= (eZ ss R′ \ E)/NR′ . Case 2(c)ii). By Remark 2.5, if we are not in the situation of 2(c)i), then the semistable locus eX ss,R is empty, and as in Case 2(b) there is a nonempty open stratum with β 6= 0 in the HKKN stratification of eX. Then eSβ \ E ∼= R ×Pβ (eY ss open in eX. ∼= R ×Pβ eY ss eSβ β β \ E) is nonempty and 18 ) \ E) STRATIFYING QUOTIENT STACKS AND MODULI STACKS If eY ss β \ E 6= eX then as in Case 2(b)i) S0(X, H, λ, L) = ψ(R S0(eY ss is a nonempty R-invariant (and hence H-invariant) open subscheme of X with a geometric quotient β \ E, Pβ, λβ, eL eY ss β \E S0(X, H, λ, L)/H ∼= (S0(eY ss β \ E, Pβ, λβ, eL eY ss β \E ) \ E)/Pβ . on the dimension of H. Finally if eY ss β \ E = eX then as in Case 2(b)ii) we can either reduce to Case 1 or use induction This completes the recursive definition of S0(X, H, λ, L); then we define the stratification {Sγγ ∈ Γ} and its indexing set Γ as at (3.1) and (3.2), and Theorem 1.1 follows from the con- struction. Of course, if the dimension of X or H is 0, we set S0(X, H, λ, L) = X. Example: Ordered points on the projective line. This is a familiar example (cf. [23, 28]). Let H = SL(2) so that U is trivial and R = H. Fix n ≥ 1 and consider the diagonal action of H on (P1)n. The linearisation is the nth tensor power of the standard representation on C2 via the Segre embedding, and the indexing set of the HKKN stratification is B = {(2r − n) n ≥ r > n/2} ∪ {0}. If β = 2r − n with r > n/2 then a sequence lies in Zβ = Z ss β if and only if it contains r entries equal to [1 : 0] and n − r entries equal to [0 : 1], while Yβ = Y ss β consists of sequences with precisely r entries equal to [1 : 0], and finally the HKKN stratum Sβ consists of sequences such that exactly r entries coincide. Thus Zβ, Yβ and Sβ all have(cid:0)n semistable stratum corresponds to β = 0 and consists of sequences in which no point of P1 occurs strictly more than n/2 times. r(cid:1) connected components. The In order to describe the refined stratification, note first that the Uβ-stabilisers in Zβ when β = 2r − n are trivial for n/2 < r < n. Therefore in the refined stratification, Sβ decomposes as the disjoint union of Sr,n−r 2r−n = SL(2)Zβ, consisting of sequences supported at two points of multiplicity r and n − r respectively, and its complement Sr,<n−r in Sβ; when r = n − 1 this complement is empty. When r = n the Un-stabilisers are not trivial, but they have the same dimension for all points in Zn and the stratum Sn remains intact. 2r−n The semistable stratum S0 behaves differently for even and odd n. For odd n, the SL(2)- semistable locus S0 coincides with the stable locus and no blow-up is needed. However, for even n we need to split S0 into smaller strata to get geometric quotients: the stable locus S<n where strictly fewer than n/2 points coincide, the (connected components of the) centre Sn/2,n/2 of the blow-up where half the entries of the sequence coincide at each of two points in P1, and the (connected components of the) strata Sn/2,<n/2 where half the entries coincide but the other half do not. 0 0 0 Thus modulo taking connected components the refined stratification has the following form: n odd: (P1)n = S0 ⊔ Sn−2 ⊔ Sn ⊔ Gn/2<r<n−1 n even: (P1)n = S<n 0 ⊔ Sn/2,n/2 S0 0 ⊔ Sn/2,<n/2 0 } {z Sr,n−r 2r−n ⊔ Sr,<n−r 2r−n ; S2r−n {z ⊔Sn−2 ⊔ Sn ⊔ Gn/2<r<n−1 } Sr,n−r 2r−n ⊔ Sr,<n−r 2r−n . S2r−n {z } STRATIFYING QUOTIENT STACKS AND MODULI STACKS 19 Remark 3.2. If we only require good quotients rather than geometric quotients of the strata Sγ, then we can simplify the algorithm for constructing the stratification {Sγγ ∈ Γ} somewhat, by combining the Cases 2(a) and (c) in the definition of S0(X, H, λ, L) into the following single case. Case 2(a'). Suppose that the semistable locus X ss,R for the linear action of R on X is nonempty. Then we let and by classical GIT this has a good quotient X//R. S0(X, H, λ, L) = X ss,R We note that this stratification is not intrinsic to the stack X := [X/H], as it depends on the presentation of X as a quotient stack [X/G] and a choice of ample linearisation L on X, as well as an invariant inner product on Lie R. Since for reductive groups H = R, this stratification refines the HKKN stratification, which itself depends on a choice of such a linearisation and an inner product, this dependence is to be expected. Remark 3.3. Note that when H = R is reductive then condition (†) in Theorem 1.1 holds for β = 0 if and only if X ss = X s. Moreover (†) holds for β 6= 0 if and only if semistability coin- cides with stability for the induced linear action of Stab(β)/λβ (Gm) on Zβ and also condition subgroup Pβ ≤ R. (∗) holds for the action of the graded unipotent radical bUβ = Uβ ⋊ λβ(Gm) of the parabolic There is a notion of a coarse moduli space of an Artin stack X (over k), which is a morphism π : X → Y to an algebraic space Y that is a categorical quotient (i.e. π is initial among all morphisms from X to an algebraic space) and such that π induces a bijection on point sets [X(k)] ∼= Y (k). Remark 3.4. In general, if G is an algebraic group acting on a scheme X with a categorical quotient q : X → Y , then π : [X/G] → Y is not necessarily a coarse moduli space, as π only induces a bijection on points sets if Y is an orbit space. However, if q is a geometric quotient, then π is a coarse moduli space for X := [X/G]. Indeed, since geometric quotients are orbit spaces, π clearly induces a bijection [X(k)] = X(k)/G(k) ∼= Y (k). To prove it is also a categorical quotient, one can apply a result of Rydh [29, Theorem 3.16], which says that if q is universally open and strongly geometric (in the sense of [29, Definition 2.2]), then π is a categorical quotient. Since q is a geometric quotient it is surjective, universally open and the fibres of q are single orbits; furthermore, the natural morphism j : G × X → X ×Y X is surjective. In particular, j, q and qcons (the induced map for the constructible topology) are all universally submersive; thus q is a strongly geometric quotient by [29, Lemma 2.3]. Alper [1] introduced a notion of a good moduli space for stacks, which is inspired by the notion of a good quotient in reductive GIT. However, this notion is suited to reductive groups and is less favourable for our set up where we have stacks with potentially non-reductive sta- biliser groups. Indeed BG = [Spec k/G] → Spec k is a good moduli space if and only if G/k is linearly reductive, as the stabiliser groups of the closed points of a stack admitting a good moduli space are all linearly reductive by [1, Proposition 12.14]. By Remark 3.4, we obtain the following corollary of Theorem 1.1. Corollary 3.5. Let X = [X/H] be a quotient stack, where X and H are as in the statement of Theorem 1.1, and so we have a stratification {Sγγ ∈ Γ} of X into H-invariant quasi-projective subschemes which STRATIFYING QUOTIENT STACKS AND MODULI STACKS 20 each admit geometric quotients Sγ/H. Then there is an induced stratification {Xγ := [Sγ/H] : γ ∈ Γ} of X such that each stratum has a coarse moduli space Xγ → Sγ/H. We note that, in general, it is not the case that [Sγ/H] is isomorphic to Sγ/H; this only hap- pens if additionally the action of H on Sγ is free. 4. STACKS FILTERED BY QUOTIENT STACKS In §3, we constructed a stratification {Sγγ ∈ Γ} of a projective scheme X acted on linearly by a linear algebraic group H = U ⋊ R with internally graded unipotent radical U , such that each stratum is an H-invariant quasi-projective subscheme of X with a geometric quotient by the action of H. The quotient stack [X/H] then has an induced stratification {Σγγ ∈ Γ} such that each stratum Σγ has the form Σγ := [Sγ/H] ∼= [Wγ/Hγ] where Wγ is a quasi-projective subscheme of X acted on by a linear algebraic subgroup Hγ of H with internally graded unipotent radical, and this action has a geometric quotient Wγ/Hγ. Moreover under appropriate hypotheses of 'semistability coinciding with stability' for the ac- tions of the subgroups Hγ, the geometric quotients Wγ/Hγ are themselves projective. We can apply this to Artin stacks which are filtered by quotient stacks admitting stratifica- tions that are compatible in the following sense. Definition 4.1. Let M be an Artin stack. (1) We say M is filtered by quotient stacks if M can be expressed as an increasing union M = [n>0 Mn of open substacks with presentations of the form Mn ∼= [Vn/Hn] where [Vn/Hn] is the quotient stack associated to a linear action on a quasi-projective scheme Vn by a group Hn = Un ⋊ Rn with internally graded unipotent radical. (2) We say M is filtered by quotient stacks admitting compatible stratifications if M is filtered as above and moreover, for each n ∈ N, there is an ample Hn-linearisation Ln on projective completion Vn of Vn and a choice of invariant inner product on the Lie algebra of the Levi factor Rn ≤ Hn such that for n′ > n, the pullback along Mn ֒→ Mn′ of the stratification of Mn′ (obtained by restricting the stratification on [Vn′/Hn′] defined by Ln′ and this norm to Mn′) coincides with the corresponding stratification on Mn. (3) We additionally say that M is filtered by quotient stacks admitting compatible stratifications which are asymptotically stable, if for each n and each stratum Σγ,n of Mn ∼= [Vn/Hn] there exists N ≥ n such that if l > m > N and Σγ,l ֒→ Ml and Σγ,m ֒→ Mm are the strata whose restrictions to Mn are Σγ,n, then Σγ,l is the image of Σγ,m under the inclusion Mm ֒→ Ml. Then we obtain the following corollary to Theorem 1.1. Corollary 4.2. Let M be an Artin stack that is filtered by quotient admitting compatible stratifications which are asymptotically stable; then there is a stratification {Σγγ ∈ Γ} of the stack M such that each stratum Σγ is isomorphic to a quotient stack [Wγ/Hγ], where Wγ is quasi-projective and acted on by a STRATIFYING QUOTIENT STACKS AND MODULI STACKS 21 linear algebraic group Hγ = Uγ ⋊ Rγ with internally graded unipotent radical Uγ. Moreover, there is a geometric quotient Wγ/Hγ which is a coarse moduli space for Σγ. In particular, M has a stratification by quotient stacks which all have coarse moduli spaces. Remark 4.3. In fact, the same result holds if we consider stacks which admit an increasing open cover indexed by a partially ordered filtered set that contains a cofinal copy of N with analogous properties. We have the following criterion for these moduli spaces to be projective (cf. Remark 3.3). Remark 4.4. In the situation of Corollary 4.2, for a stratum Σγ ∼= [Wγ/Hγ] where the unipotent radical Uγ is graded by λγ : Gm → Rγ, the coarse moduli space Wγ/Hγ will be projective if (∗) holds for the action of Uγ ⋊ λγ(Gm) on a suitable projective completion of Wγ with respect to an induced linearisation, and semistabiity coincides with stability for the induced action of Rγ/λγ(Gm) on the analogue of Zmin in this projective completion. 5. APPLICATIONS We can apply the previous section to the construction of moduli spaces of 'unstable' objects, including moduli of sheaves over a projective scheme [6], projective curves [21]) and hypersur- faces in a complete toric variety. 5.1. Moduli of hypersurfaces in a complete toric variety. Let Y be a complete simplicial toric variety. It was observed in [3, §4] (using the description of Aut(Y ) given in [11]) that the auto- morphism group H = Aut(Y ) of Y is a linear algebraic group with internally graded unipotent radical. By considering suitable Hilbert schemes, each component of the moduli stack of hypersur- faces in Y can be expressed itself as a quotient stack [V /H] where V is a projective scheme with a linear action of H := Aut(Y ) with an ample linearisation. Thus we can apply Theorem 1.1 directly to the moduli stack M to obtain a stratification {Σγγ ∈ Γ} of M such that each stratum Σγ is isomorphic to a quotient stack [Wγ/Hγ], where Wγ is a quasi-projective scheme acted on by a linear algebraic group Hγ with internally graded unipotent radical, and there is a geometric quotient Wγ/Hγ which is a coarse moduli space for Σγ. A simple example is the automorphism group of the weighted projective plane Y = P(1, 1, 2) = (k3 \ {0})/Gm, for Gm acting linearly on k3 with weights 1, 1, 2. This is given by Aut(P(1, 1, 2)) ∼= U ⋊ R where R ∼= GL(2) is reductive and U ∼= (Ga)3 is unipotent, with elements (λ, µ, ν) ∈ k3 acting on P(1, 1, 2) via [x, y, z] 7→ [x, y, z + λx2 + µxy + νy2]. The central one-parameter subgroup Gm of R ∼= GL(2) acts on the Lie algebra of U with all positive weights, providing an internal grading of the unipotent radical U . 5.2. Moduli of sheaves over a projective scheme. Let (Y, OY (1)) denote a polarised projective scheme over an algebraically closed field k of characteristic 0. For a fixed Hilbert polynomial P , we let M := MY,P denote the moduli stack of coherent sheaves over Y with Hilbert polynomial P with respect to OY (1). We can filter this stack by quotient stacks in several ways: for example, STRATIFYING QUOTIENT STACKS AND MODULI STACKS 22 one can use Castelnuevo–Mumford regularity or the Harder–Narasimhan type. In both filtra- tions, we will filter by quotient stacks which are quotients by general linear groups of open subschemes of Quot schemes, which appear in Simpson's GIT construction [33] of the moduli space of semistable sheaves on Y with Hilbert polynomial P . In this section, we will describe the two ways to filter M and explain how to use a filtration involving Harder–Narasimhan types to apply the results in §4. Let us recall Simpson's reductive GIT construction of moduli spaces of semistable sheaves. By the Le Potier estimates [33, Theorem 1.1], the family of semistable sheaves over Y with Hilbert polynomial P is bounded; hence, for n ≫ 0, all such sheaves are n-regular in the sense of Castelnuevo-Mumford. For an n-regular sheaf E, the evaluation map H 0(E(n)) ⊗ OY (−n) → E is surjective and dim H 0(E(n)) = P (E, n), as all the higher cohomology groups of E(n) vanish; hence, any n-regular sheaf with Hilbert polynomial P determines a closed point in the Quot scheme Qn := Quot(OY (−n)⊕P (n), P ) parameterising quotient sheaves of OY (−n)⊕P (n) with Hilbert polynomial P . In fact, it deter- mines a point in the open subscheme Qn−reg consisting of quotients q : OY (−n)⊕P (n) ։ F such that F is n-regular and H 0(q(n)) is an isomorphism. If Qss n−reg denotes the open subset, where additionally the quotient sheaf is semistable, then Simpson constructs the moduli space of semistable sheaves as a quotient of the natural SLP (n)-action on the closure of Qss n−reg. The action is linearised by choosing m ≫ n to construct an embedding of the Quot scheme into a Grassmannian and then composing with a Pl ucker embedding of the Grassmannian into pro- jective space; we let Lm,n denote this linearisation. Since n-regularity is an open condition, we obtain an increasing open filtration of M by the substacks Mn−reg of n-regular sheaves; furthermore, there are isomorphisms Mn−reg ∼= [Qn−reg/ GLP (n)] for all n. Hence M is filtered by quotient stacks using n-regularity. Let us explain a second way to filter M using Harder–Narasimhan (HN) types [17]. We recall that every coherent sheaf E has a unique HN filtration 0 = E (0) ( E (1) · · · E (r) = E (where here we use a Rudakov type notion of semistability to define HN filtrations for any coherent sheaf; for example, see [19, Definition 4.11]). The HN type of E is τ (E) := (P (E1), . . . , P (Er)), where Ei := E (i)/E (i−1). The set of HN types for sheaves with Hilbert polynomial P is partially ordered, such that the semistable HN type given by the single polynomial (P ) is minimal. We let Mτ and M<τ denote the substacks of sheaves of HN type τ and of HN type less than τ respectively. By work of Shatz [32], the HN type is upper semi-continuous, and so M<τ is an open substack of M. Furthermore, as there are only finitely many HN types less than a given HN type τ , these sheaves form a bounded family, and so M<τ is a quotient stack of the form [Q<τ n−reg/ GLP (n)] for n ≫ 0 (depending on τ ); here, Q<τ n−reg is the open subscheme of Qn−reg parametrising quotient sheaves with HN type less than τ . Remark 5.1. The relationship between the HKKN stratifications and the HN stratifications on Qn−reg is studied in [19, 20], where it is shown that for a HN type τ and for m ≫ n ≫ 1 (depending on τ ), there is an HKKN index βn,m(τ ) for the SLP (n)-action on Qn with respect to Ln,m (and the Killing form on SLP (n)) such that the following statements hold. (1) All sheaves with HN type τ are n-regular, and thus Mτ ∼= [Qτ n−reg/ GLP (n)], where Qτ n−reg is the HN stratum in Qn−reg indexed by τ . STRATIFYING QUOTIENT STACKS AND MODULI STACKS 23 (2) The HN stratum Qτ n−reg is contained as a closed subscheme in the HKKN stratum Sβn,m(τ ). For the precise definition of βn,m(τ ), see [19, Definition 4.18]. For fixed n and m, the assignment τ 7→ βn,m(τ ) is not injective; however, for two HN types τ 6= τ ′ and m ≫ n ≫ 0, we have βn,m(τ ) 6= βn,m(τ ′). In fact, for a finite number of HN types τ1, . . . , τl, we can pick m ≫ n ≫ 1 so that all the indices βn,m(τi) are distinct and both the properties (1) and (2) above hold for all of these HN types. Furthermore, in [19, Theorem 1.1] an asymptotic HKKN stratification on M is constructed using the HKKN stratifications on each Mn−reg and it is shown that this asymptotic HKKN stratification coincides with the HN stratification on M. In fact, the HKKN stratifications on Mn−reg are not compatible stratifications which are asymptotically stable in the sense of Defini- tion 4.1, and so this asymptotic construction of a HKKN stratification is slightly more involved (for details, see [19, §4]). We claim that if we consider the filtration given by the open substacks M<τ and choose an appropriate presentation of these stacks, then the HKKN stratifications on each M<τ co- incide with the HN stratifications, and thus in particular, these stratifications are compatible and asymptotically stable in the sense of Definition 4.1. For each HN type τ , as there are only finitely many HN types less than τ , we can pick m ≫ n ≫ 1 as in Remark 5.1, so that both statements in this remark hold for all these HN types and they all have different HKKN in- dices. Indeed, if we compare the stratifications on Q<τ n−reg ֒→ Sβn,m(τ ′) for all τ ′ < τ , and as Q<τ n−reg, we see that in fact the HN and the HKKN stratifications must agree. n−reg is the disjoint union of the HN strata Qτ ′ n−reg, then we have Qτ ′ Finally let us consider the refinement of the HKKN stratification on each M<τ provided by Theorem 1.1. Since the refinement will come from considering various stabiliser groups and stable loci, and these can be studied intrinsically for the sheaves themselves, we see that the stratifications on each M<τ ∼= [Q<τ n−reg/ GLP (n)] are compatible and asymptotically stable in the sense of Definition 4.1. Therefore, there is an induced stratification {Σγγ ∈ Γ} of M by quotient stacks that refines the HN stratification and such that each quotient stack has a presentation Σγ ∼= [Wγ/Hγ], where Wγ is a quasi-projective scheme acted on by a linear algebraic group Hγ with internally graded unipotent radical, and there is a geometric quotient Wγ/Hγ which is a coarse moduli space for Σγ. In fact, we can say what the open stratum is: provided that there exists a stable sheaf with Hilbert polynomial P , the open stratum is the moduli stack Ms of stable sheaves. The precise construction and description of these coarse moduli spaces is described in [6]. 5.3. Moduli of unstable curves. Another classical family of moduli spaces constructed as re- ductive GIT quotients is given by the moduli spaces of stable curves of fixed genus g ≥ 2. Here we can use special linear group actions on Hilbert schemes of curves embedded in projective spaces, and exploit the Rosenlicht–Serre description of singular curves in terms of their nor- malisations together with additional data describing the singularities, to look for an associated stratification {Σγγ ∈ Γ} of the moduli stack of projective curves, allowing us to construct geo- metric quotients of the unstable strata [21]. In this case of projective curves the linear actions can be set up so that (almost always) the condition that semistability coincides with stability is satisfied. Thus we expect the coarse moduli spaces of 'unstable curves' given by the geometric quotients Wγ/Hγ of the strata Σγ ∼= [Wγ/Hγ] to be projective. STRATIFYING QUOTIENT STACKS AND MODULI STACKS 24 REFERENCES [1] J. Alper, Good moduli spaces for Artin stacks, Ann. Inst. Fourier (Grenoble), 63 (6), 2349–2402, 2013. [2] G. B´erczi, B. Doran, T. Hawes, F. Kirwan, Constructing quotients of algebraic varieties by linear algebraic group actions, arXiv:1512.02997, to appear in "Handbook of Group Actions" Vol 3, editors Lizhen Ji, Athanase Pa- padopoulos, Shing-Tung Yau. [3] G. B´erczi, B. Doran, T. Hawes, F. Kirwan, Geometric invariant theory for graded unipotent groups and applica- tions, arXiv: 1601.00340 . [4] G. B´erczi, B. Doran, T. Hawes, F. Kirwan, Projective completions of graded unipotent quotients, arXiv:1607.04181. [5] G. B´erczi, B. Doran, F. Kirwan, Graded linearisations, arXiv:1703.05226, to appear in Modern Geometry a cele- bration of the work of Simon Donaldson, editors Vincente Munoz, Ivan Smith, Richard Thomas, AMS Proceed- ings of Symposia in Pure Mathematics. [6] G. B´erczi, V. Hoskins, J. Jackson, F. Kirwan, Quotients of unstable subvarieties and moduli spaces of sheaves of fixed Harder–Narasimhan type II, in preparation. [7] G. B´erczi, J. Jackson, F. Kirwan, Variation of non-reductive GIT, in preparation. [8] G. B´erczi, F. Kirwan, Graded unipotent groups and Grosshans theory, arXiv:1511.06983, to appear in Forum of Mathematics, Sigma. [9] L. Brambila-Paz, O. Mata, N. Nitsure, Moduli stacks and moduli schemes for rank 2 unstable bundles, arXiv:0911.2301. [10] M. Brion, On the geometry of algebraic groups and homogeneous spaces, J. Algebra, 329, 52–71, 2011. [11] D.A. Cox, The homogeneous coordinate ring of a toric variety, J. Algebraic Geometry 4 (1995) 17-50. [12] Jacques Dixmier. Enveloping algebras, volume 11 of Graduate Studies in Mathematics. American Mathematical Society, Providence, RI, 1996. Revised reprint of the 1977 translation. [13] I. Dolgachev and Y. Hu, Variation of geometric invariant theory quotients, Inst. Hautes ´Etudes Sci. Publ. Math., 87(1), 5–56, 1998 (with an appendix by N. Ressayre). [14] B. Doran and F. Kirwan, Towards non-reductive geometric invariant theory, Pure Appl. Math. Q. 3 (2007), 61– 105. [15] D. Edidin and D. Rydh, Canonical reduction of stabilizers for Artin stacks with good moduli spaces, arXiv: 1710.03220. [16] D. Halpern-Leistner, Theta-stratifications, Theta-reductive stacks, and applications, to appear in the proceedings of the 2015 AMS summer institute in Salt Lake City. [17] G. Harder and M. S. Narasimhan. On the cohomology groups of moduli spaces of vector bundles on curves. Math. Ann., 212 (3) 215–248, 1975. [18] W.H. Hesselink, Uniform instability in reductive groups, J. Reine Angew. Math. 304 (1978), 299-316. [19] V. Hoskins, Stratifications for moduli of sheaves and moduli of quiver representations, to appear in Algebraic Geometry (arXiv: 1407.4057). [20] V. Hoskins and F. Kirwan, Quotients of unstable subvarieties and moduli spaces of sheaves of fixed Harder– Narasimhan type, Proc. London Math. Soc., 105(4), 852–890, 2012. [21] J. Jackson, Oxford DPhil thesis, in preparation. [22] G. Kempf, Instability in invariant theory, Ann. of Math. 108 (2), 299–316, 1978. [23] F. Kirwan, Cohomology of quotients in symplectic and algebraic geometry. Mathematical Notes, 31. Princeton University Press, Princeton, NJ, 1984. [24] F. Kirwan, Partial desingularisations of quotients of nonsingular varieties and their Betti numbers. Ann. of Math. (2) 122 (1985), no. 1, 41-85. [25] F. Kirwan, Refinements of the Morse stratification of the normsquare of the moment map, in 'The breadth of symplectic and Poisson geometry', Progr. Math., 232, 327–362, 2005. [26] D. Mumford, J. Fogarty and F. Kirwan, Geometric invariant theory, 3rd edition, Springer, 1994. [27] L. Ness, A stratification of the null cone via the moment map, Amer. J. Math. 106, 1984, 1281-1329. [28] P.E. Newstead, Introduction to moduli problems and orbit spaces, Tata Institute Lecture Notes, Springer, 1978. [29] D. Rydh, Existence and properties of geometric quotients, J. Algebraic Geom. 22 (2013), no. 4, 629669. [30] The Stacks Project Authors. Stacks Project. http://stacks.math.columbia.edu, 2015. [31] J. -P. Serre, Espaces fibr´es alg´ebriques, S´eminaire C. Chevalley; 2e ann´ee: 1958. Anneaux de Chow et applica- tions seminaire Claude Chevalley, 1-37. STRATIFYING QUOTIENT STACKS AND MODULI STACKS 25 [32] S. Shatz, The decomposition and specialization of algebraic families of vector bundles, Compositio Math. 35, 163–187, 1977. [33] C. Simpson, Moduli of representations of the fundamental group of a smooth projective variety, Publications Math´ematiques de l'IH ´ES, 79(1), 47–129, 1994. [34] M. Thaddeus, Geometric invariant theory and flips, J. Amer. Math. Soc., 9(3), 691–723, 1996. (B´erczi) DEPARTEMENT MATHEMATIK, ETH, SWITZERLAND E-mail address: [email protected] (Hoskins) FACHBEREICH MATHEMATIK UND INFORMATIK, FREIE UNIVERSIT AT BERLIN, GERMANY E-mail address: [email protected] (Kirwan) MATHEMATICAL INSTITUTE, OXFORD UNIVERSITY, UK E-mail address: [email protected]
1803.05524	1	1803	2018-03-14T22:32:56	Positivity Cones under Deformations of Complex Structures	[ "math.AG", "math.CV", "math.DG" ]	We investigate connections between the sGG property of compact complex manifolds, defined in earlier work by the second author and L. Ugarte by the requirement that every Gauduchon metric be strongly Gauduchon, and a possible degeneration of the Fr\"olicher spectral sequence. In the first approach that we propose, we prove a partial degeneration at $E_2$ and we introduce a positivity cone in the $E_2$-cohomology of bidegree $(n-2,\,n)$ of the manifold that we then prove to behave lower semicontinuously under deformations of the complex structure. In the second approach that we propose, we introduce an analogue of the $\partial\bar\partial$-lemma property of compact complex manifolds for any real non-zero constant $h$ using the partial twisting $d_h$, introduced recently by the second author, of the standard Poincar\'e differential $d$. We then show, among other things, that this $h$-$\partial\bar\partial$-property is deformation open.	math.AG	math	Positivity Cones under Deformations of Complex Structures Houda Bellitir and Dan Popovici Abstract. We investigate connections between the sGG property of compact complex manifolds, defined in earlier work by the second author and L. Ugarte by the requirement that every Gauduchon metric be strongly Gauduchon, and a possible degeneration of the Frolicher spectral sequence. In the first approach that we propose, we prove a partial degeneration at E2 and we introduce a positivity cone in the E2- cohomology of bidegree (n − 2, n) of the manifold that we then prove to behave lower semicontinuously under deformations of the complex structure. In the second approach that we propose, we introduce an analogue of the ∂ ¯∂-lemma property of compact complex manifolds for any real non-zero constant h using the partial twisting dh, introduced recently by the second author, of the standard Poincar´e differential d. We then show, among other things, that this h-∂ ¯∂-property is deformation open. 1 Introduction In this paper, we explore a few connections between the metric geometry of compact complex manifolds and some Hodge-theoretic aspects related to the Frolicher spectral sequence (a classical object linking the differential and the complex structures, see e.g. a reminder of the definition in §.2 below) of these manifolds. It is well known that the existence of a Kahler metric implies the best possible degeneration (i.e. at the first page E1) of this spectral sequence. However, compact complex manifolds admit only rarely Kahler metrics and no other metric property is currently known to imply the degeneration at some page of this spectral sequence. The following conjecture was proposed and solved in a special case in [Pop16]. Conjecture 1.1. Let X be a compact complex manifold. If an SKT metric (i.e. a C ∞ positive defi- nite (1, 1)-form ω such that ∂ ¯∂ω = 0) exists on X, the Frolicher spectral sequence of X degenerates at the second page E2. The general case of this conjecture, that will certainly have a role to play in the classification theory of compact complex manifolds, remains open. In this paper, we investigate a possible variant of it in which SKT metrics are replaced by strongly Gauduchon (sG) metrics and the degeneration issue is often confined to the cohomology of X in a degree close to the maximal one. We take our cue from a result of Ceballos-Otal-Ugarte-Villacampa [COUV16, Theorem 5.6] asserting that the Frolicher spectral sequence of any 6-real-dimensional nilmanifold endowed with an invariant complex structure and carrying an sG metric degenerates at E2 and from a possible generalisation of this statement they wondered about in Question 1.2. ([COUV16, Question 5.7]) Does the Frolicher spectral sequence of any 3-dimensional compact complex manifold carrying an sG metric degenerate at E2? Let X be a compact complex manifold with dimCX = n. Recall that Hermitian metrics, defined as C ∞ positive definite (1, 1)-forms ω, always exist on X. Even Gauduchon metrics ω, defined as 1 Hermitian metrics satisfying the extra condition ∂ ¯∂ωn−1 = 0, always exist (cf. [Gau77]). However, strongly Gauduchon (sG) metrics, introduced in [Pop13] in the context of deformations of complex structures and defined by the stronger requirement that ∂ωn−1 be ¯∂-exact, need not exist. Compact complex manifolds that admit sG metrics were called strongly Gauduchon (sG) manifolds and the notion covers a wide range of manifolds and considerably enlarges the class of compact Kahler manifolds and their bimeromorphic models (= the so-called Fujiki class C manifolds). Section 3 of this paper investigates elements of the E2 degeneration of the Frolicher spectral se- quence of sGG manifolds, a class of compact complex manifolds introduced and studied in [PU18a]. They are defined by the requirement that every Gauduchon metric be strongly Gauduchon. In particular, sG metrics exist on every sGG manifold. Moreover, thanks to [PU18a, Lemma 1.3], an n-dimensional compact complex manifold X is sGG if and only if the following special case of the ∂ ¯∂-lemma holds on X: for every d-closed (n, n − 1)-form Γ on X, if Γ is ∂-exact, then Γ is also ¯∂-exact. We exploit this fact in at least two ways in section 3: (1) by showing that every sGG manifold has the partial Frolicher E2 degeneration property that the map d2 acting in bidegree (n−2, n) on the second page of its Frolicher spectral sequence vanishes identically (cf. Proposition 3.3); 2 (2) by introducing three versions SX ⊂ En−2, n 2 bSX ⊂ En−2, n DR (X, R) (cf. Definition 3.5) and (X) (cf. (15)) of a positivity cone that we call the E2sG-cone of a given compact complex n-dimensional manifold X. The term E2sG refers to the fact that this cone consists of (double) cohomology classes of bidegree (n − 2, n) that arise on the second page of the Frolicher spectral sequence of X and are thus a refinement of the Dolbeault cohomology. (X), eSX ⊂ H 2n−2 The surprising aspect is that we thus effectively introduce a notion of positivity for cohomology classes of bidegree (n − 2, n) that runs counter to the familiar notions of positivity that exist in all the bidegrees (p, p) (but not (p, q) with p 6= q) in complex geometry. This is done by using strongly Gauduchon metrics that enable the existing notion of positivity in bidegree (n − 1, n − 1) to carry over to the bidegree (n − 2, n) in a natural way. The E2sG-cone is empty if the manifold X is not strongly Gauduchon. It depends on the complex structure of X. We study this dependence by proving the following Theorem 1.3. Let π : X −→ ∆ be a holomorphic family of compact complex n-dimensional man- ifolds over a ball ∆ ⊂ CN centred at the origin. Suppose that the fibre X0 := π−1(0) is an sGG manifold. Then, the De Rham E2sG-cone eSXt ⊂ H 2n−2 lower semicontinuous way with t ∈ ∆ varying in a small enough neighbourhood of 0 ∈ ∆. DR (X, R) of the fibre Xt := π−1(t) ⊂ X varies in a By X we mean the smooth manifold that underlies all the fibres Xt of the family (known to be C ∞ trivial by the classical Ehresmann Theorem, but in general not holomorphically trivial), so the real De Rham cohomology group H 2n−2 DR (X, R) of degree 2n − 2 is independent of the (complex structure of the) fibre Xt. The meaning of lower semicontinuous in connection with the dependence on t of the De Rham E2sG-cone eSXt is made precise in Theorem 3.9. 2 We also use the Serre-type duality (proved in the forthcoming joint paper [PU18b] of the second author with L. Ugarte by means of the pseudo-differential Laplacian introduced in [Pop16]) between any pair (Ep, q (X)) of vector spaces of complementary bidegrees featuring on the sec- ond page of the Frolicher spectral sequence of X. In this way, we prove the following analogue in our context of Lamari's duality [Lam99, Lemma 3.3] between the pseudo-effective cone and the closure of the Gauduchon cone of any compact complex manifold. (X), En−p, n−q 2 2 Proposition 1.4. Let X be an n-dimensional sGG compact complex manifold on which an arbitrary Hermitian metric γ has been fixed. The dual of the closure of the cone bSX =(cid:26)(cid:20)[Γn−2, n] ¯∂(cid:21)d1 ∃ ω Hermitian metric such that ∂Γn−2, n = − ¯∂ωn−1(cid:27) ⊂ En−2, n 2 (X) (1) under the duality E2, 0 E2-classes [[θ2, 0] ¯∂]d1 "representable" by γ-positive currents τ 2, 0 : C ∞ (X) −→ C is the closed convex cone in E2, 0 (X) × En−2, n 2 2 2 n−2, n(X, R)γ −→ R. (X) consisting of the We refer to Proposition 3.16 for a more precise statement and to Definition 3.13 for the no- tion of γ-positive current of bidegree (2, 0). This is another extension of the classical notions of positivity in complex geometry to a bidegree (p, q) with p 6= q. The bidegree (2, 0) is especially important on holomorphic symplectic (not necessarily Kahler) manifolds and this lead will hopefully be investigated in future work. Section 4 of this paper takes up the Frolicher degeneration issue and the variation of the complex structure from a different point of view that still ties in with the theory of sGG manifolds. Let X be a compact complex manifold with dimCX = n. In [Pop17], for every positive constant h, the differential operator d = ∂ + ¯∂ associated with the smooth structure of X was modified to dh := h∂ + ¯∂ : C ∞ k (X, C) → C ∞ k+1(X, C), k ∈ {0, . . . , 2n}, 1 (2) by rescaling its (1, 0)-part in the splitting induced by the complex structure of X. Unlike d, the operators dh depend on the complex structure of X while also sharing some properties with d. The most striking of these is that the dh-cohomology of X, relying on the integrability property d2 h = 0 and defined for every k ∈ {0, . . . , 2n} by H k dh(X, C) = ker(dh : C ∞ k (X, C) → C ∞ k−1(X, C) → C ∞ k (X, C)), k+1(X, C))(cid:30)Im (dh : C ∞ is isomorphic to the De Rham cohomology of X via the isomorphism H k {θhu}dh ∈ H k dh (X, C) induced by the pointwise isomorphism DR(X, C) ∋ {u}DR 7→ θh : Λp, qT ⋆X → Λp, qT ⋆X, u 7→ θhu := hp u, at the level of pure-type differential forms on X. The operators dh capture in a certain sense the relationships between the smooth and the complex structures of X. On the other hand, there exist in the literarure at least two notions of the ∂ ¯∂-lemma being satisfied by a compact complex manifold that capture the smooth structure-complex structure relationship. 1Throughout the paper, C∞ k (X, C) will stand for the space of C-valued C∞ k-forms on X. 3 (a) An early notion that has been used by many authors appeared in [DGMS75, Lemmas 5.11 and 5.15]. It defines the fulfilment of the ∂ ¯∂-lemma (or of the equivalent ddc-lemma) on a compact complex manifold by the requirement that any smooth differential form u of any degree (but not necessarily of pure type) that is both ∂-closed and ¯∂-closed satisfies the following implication: u ∈ Im d =⇒ u ∈ Im (∂ ¯∂). (3) This implication is actually an equivalence since the reverse implication holds trivially on any man- ifold. For example, this is what was meant by the ∂ ¯∂-lemma holding on a given manifold in [AT12]. (b) In [Pop14], the term ∂ ¯∂-manifold was introduced to mean that a given compact complex manifold X satisfies the ∂ ¯∂-lemma if for any d-closed pure-type form u on X the following exactness properties are equivalent: u ∈ Im d ⇐⇒ u ∈ Im ∂ ⇐⇒ u ∈ Im ¯∂ ⇐⇒ u ∈ Im (∂ ¯∂). (4) The last property trivially implies the others, so the above equivalences reduce to each of the other three forms of exactness implying (∂ ¯∂)-exactness. Since u is of pure type, the d-closedness assumption on u is equivalent to u being assumed both ∂-closed and ¯∂-closed. On the face of it, condition (4) is more restrictive than (3), but (4) is only required to apply to pure-type forms. It is implicit in [DGMS75, Lemma 5.15] that version (a) of the ∂ ¯∂-lemma condition (required to hold on all, not necessarily pure-type forms) actually implies version (b). For every constant h ∈ R \ {0}, we introduce in this paper the notion of h-∂ ¯∂-manifold that implies the above version (a) (hence also version (b)) of the ∂ ¯∂-condition. The idea is to use the operators dh to capture the interplay between the smooth structure and the complex structure. Definition 1.5. Let h ∈ R \ {0} be an arbitrary constant. A compact complex manifold X with dimCX = n is said to be an h-∂ ¯∂-manifold if for every k ∈ {0, 1, . . . , 2n} and every k-form u ∈ ker dh ∩ ker d−h−1, the following exactness conditions are equivalent: u ∈ Im dh ⇐⇒ u ∈ Im d−h−1 ⇐⇒ u ∈ Im d ⇐⇒ u ∈ Im (dh d−h−1) = Im (∂ ¯∂). We prove in Corollary 4.19 that an equivalent property is obtained by the removal of the condition u ∈ Im d from the above sequence of equivalences. Note that the forms u are not required to be of pure type in Definition 1.5. Unlike ∂, ¯∂ and ∂ ¯∂, the operators dh do not map pure-type forms to pure-type forms, so the proof of the implication "(a)n =⇒ (b)n " in [DGMS75, Lemma 5.15] does not seem to adapt easily to yield a proof of the possible implication "version (a) of the ∂ ¯∂-property =⇒ the h-∂ ¯∂-property". Actually, we do not know whether this last implication holds. A priori, the h-∂ ¯∂-property is stronger when h /∈ {−1, 1}. Introducing dh-analogues of the standard Bott-Chern and Aeppli cohomologies and following the pattern of Wu's proof in [Wu06] of the stability under small deformations of the complex structure of the standard ∂ ¯∂-property, we prove that the analogous statement holds for our h-∂ ¯∂-property. Theorem 1.6. Fix an arbitrary constant h ∈ R \ {0}. The h-∂ ¯∂-property of compact complex manifolds is open under deformations of the complex structure. 4 See Theorem 4.16 for a more precise statement. One last explanation is in order about the choice of pairing dh with d−h−1, rather than with the more natural-looking dh = h dh−1. This choice is forced on us by a formula of the Bochner- Kodaira-Nakano type (the h-BKN identity) that we establish in Theorem 4.5 as a consequence of what we call h-commutation relations that we compute for the operators dh and for an arbitrary Hermitian metric in Lemma 4.1. The pair (dh, d−h−1) generalises the classical pair (d, dc) since for h = −1, d−1 is a constant multiple (which does not change either the kernel or the image) of dc = −JdJ = i( ¯∂ − ∂) = i d−1. Acknowledgments. This paper is part of the first author's PhD thesis under the co-supervision of Saıd Asserda and the second author. The first author is grateful to Professor Saıd Asserda for his constant support, patient guiding of her work throughout the duration of her thesis and careful reading of the paper. Both authors are grateful to Ahmed Zeriahi for fruitful discussions and for his interest in this work. 2 Preliminaries Let X be a compact complex manifold with dimCX = n. Let (Er, dr) be the Frolicher spectral sequence of X. It relates the De Rham cohomology of X to its Dolbeault cohomology in the following way. The 0th page features the (infinite-dimensional) C-vector spaces Ep, q p, q(X, C) of C ∞ forms of arbitrary bidegree (p, q) with 0 ≤ p, q ≤ n and the linear maps 0 = C ∞ · · · d0−→ Ep, q 0 (X) d0−→ Ep, q+1 0 (X) d0−→ . . . , where d0 = ¯∂. The 1st page is defined as the cohomology of the 0th page and consists of the (finite-dimensional) Dolbeault cohomology groups Ep, q 1 (X) = H p, q(X, C) = ker(d0 : Ep, q 0 (X) → Ep, q+1 0 (X))/Im (d0 : Ep, q−1 0 (X) → Ep, q 0 (X)) and the linear maps · · · d1−→ Ep, q 1 (X) d1−→ Ep+1, q 1 (X) d1−→ . . . defined by ∂ in cohomology in the following way: d1([α] ¯∂) = [∂α] ¯∂ for all [α] ¯∂ ∈ Ep, q . 1 We then continue by induction and define the rth page as the cohomology of the (r − 1)st page, namely Ep, q r (X) = ker(dr−1 : Ep, q r−1(X) → Ep+r−1, q−r+2 r−1 (X))/Im (dr−1 : Ep−r+1, q+r−2 r−1 (X) → Ep, q r−1(X)), where the linear maps dr on each page r are of bidegree (r, −r + 1). In particular, the 2nd page is of the form · · · d2−→ Ep, q (X) d2−→ Ep+2, q−1 (X) d2−→ . . . , 2 2 where each space Ep, q 2 satisfying the conditions (X) consists of double cohomology classes (cid:20)[α] ¯∂(cid:21)d1 of smooth (p, q)-forms α 5 while each map d2 : Ep, q 2 ¯∂α = 0 (X) −→ Ep+2, q−1 2 and ∂α ∈ Im ¯∂, (X) is defined by d2(cid:18)(cid:20)[α] ¯∂(cid:21)d1(cid:19) =(cid:20)[∂u1] ¯∂(cid:21)d1 where u1 is any form such that ∂α = ¯∂u1. (5) (6) The definition of d2 is independent of the choice of the ¯∂-potential u1. It is easy to check that the vanishing condition for an arbitrary element (cid:20)[α] ¯∂(cid:21)d1 (cid:20)[α] ¯∂(cid:21)d1 p−1, q(X, C) ∩ ker ¯∂ and v ∈ C ∞ = 0 ⇐⇒ ∃u ∈ C ∞ p, q−1(X, C) such that α = ∂u + ¯∂v. (7) ∈ Ep, q 2 (X) is 3 The E2sG cone Let X be a compact complex manifold with dimCX = n. 3.1 Complex-valued cohomology We shall first deal essentially with the De Rham cohomology of X with values in C and the standard space En−2, n (X) on the second page of the Frolicher spectral sequence. 2 Proposition 3.1. The following canonical linear map T : H 2n−2 DR (X, C) −→ En−2, n 2 (X), {α}DR 7→(cid:20)[αn−2, n] ¯∂(cid:21)d1 , is well defined and its image is given by Im T = ker dn−2, n 2 , (8) where dn−2, n 2 : En−2, n 2 (X) → En, n−1 2 (X) is the d2-map acting in bidegree (n − 2, n). Proof. Let {α}DR ∈ H 2n−2 DR (X, C) be an arbitrary class and let α = αn, n−2 + αn−1, n−1 + αn−2, n be an arbitrary representative, where the αp, q's are the components of α of types (p, q). The condition dα = 0 is equivalent to ∂αn−1, n−1 + ¯∂αn, n−2 = 0 and ¯∂αn−1, n−1 + ∂αn−2, n = 0. Since ¯∂αn−2, n = 0 and ∂αn−2, n = − ¯∂αn−1, n−1 ∈ Im ¯∂, αn−2, n defines a class [[αn−2, n] ¯∂]d1 in En−2, n (X). 2 To show well-definedness for T , we still have to show that the definition is independent of the choice of representative α of the De Rham class {α}DR. This is equivalent to showing that T maps 0 ∈ H 2n−2 2n−2(X, C) be d-exact. Then, there exists β = βn, n−3 + βn−1, n−2 + βn−2, n−1 + βn−3, n a (2n − 3)-form such that α = dβ. This amounts to DR (X, C) to 0 ∈ En−2, n (X). Let α ∈ C ∞ 2 6 αn, n−2 = ∂βn−1, n−2+ ¯∂βn, n−3, αn−1, n−1 = ∂βn−2, n−1+ ¯∂βn−1, n−2 and αn−2, n = ∂βn−3, n+ ¯∂βn−2, n−1. (9) Since ¯∂βn−3, n = 0 for bidegree reasons, the last identity in (9) shows, thanks to (7), that [[αn−2, n] ¯∂]d1 = 0 in En−2, n (X). Let us now prove the inclusion ker dn−2, n (X) such 2 that d2([[αn−2, n] ¯∂]d1) = 0. Thanks to (5), (6) and (7), there exist forms Ωn−1, n−1 ∈ C ∞ n−1, n−1(X, C), u ∈ C ∞ n−1, n−1(X, C) with u ∈ ker ¯∂ and v ∈ C ∞ ⊂ Im T in (8). Let [[αn−2, n] ¯∂]d1 ∈ En−2, n n, n−2(X, C) such that 2 2 ∂αn−2, n = − ¯∂Ωn−1, n−1 = − ¯∂(Ωn−1, n−1 − u) and ∂Ωn−1, n−1 = ∂u + ¯∂v. If we put α := αn−2, n + (Ωn−1, n−1 − u) − v, we see that dα = 0 and T ({α}DR) = [[αn−2, n] ¯∂]d1. Let us now prove the reverse inclusion ker dn−2, n ⊃ Im T in (8). Let [[αn−2, n] ¯∂]d1 ∈ Im T . This means that αn−2, n is the (n − 2, n)-component of a d-closed (2n − 2)-form α = αn, n−2 + αn−1, n−1 + αn−2, n. As already noticed, the condition dα = 0 is equivalent to 2 ∂αn−1, n−1 + ¯∂αn, n−2 = 0 and ∂αn−2, n + ¯∂αn−1, n−1 = 0. On the other hand, d2([[αn−2, n] ¯∂]d1) = −[[∂αn−1, n−1] ¯∂]d1 = [[ ¯∂αn, n−2] ¯∂]d1 = 0 ∈ En, n−1 [ ¯∂αn, n−2] ¯∂ = 0 ∈ H n, n−1 (X, C). 2 ¯∂ since even (cid:3) As a consequence, we obtain the following criterion for partial degeneration at E2 of the Frolicher spectral sequence of X. Corollary 3.2. The canonical map T defined in Proposition 3.1 is surjective if and only if the map d2 vanishes identically in bidegree (n − 2, n). We now show that the sGG assumption on the ambient manifold X (see [PU18a] for the definition and a study of the class of sGG manifolds) suffices to guarantee the partial degeneration property mentioned above. : En−2, n Proposition 3.3. Let X be a compact complex manifold with dimCX = n. If X is sGG, the map dn−2, n (X) on the 2nd page of the Frolicher spectral sequence of X vanishes 2 identically (equivalently, the canonical linear map T : H 2n−2 (X) of Proposition 3.1 is surjective). DR (X, C) −→ En−2, n (X) −→ En, n−1 2 2 2 Proof. To prove that T is surjective, let [[αn−2, n] ¯∂]d1 ∈ En−2, n (X) and let αn−2, n be an arbitrary (n − 2, n)-form representing this double class. Then ∂αn−2, n is ¯∂-exact, so there exists an (n − 1, n − 1)-form Ωn−1, n−1 such that 2 (10) Now, the sGG assumption on X implies that ¯∂Ωn−1, n−1 is ∂-exact. Indeed, this is equivalent to ∂Ωn−1, n−1 being ¯∂-exact. Meanwhile, ∂Ωn−1, n−1 is a d-closed and ∂-exact (n, n − 1)-form, so (iii) of Lemma 1.3. in [PU18a] implies that ∂Ωn−1, n−1 is also ¯∂-exact (thanks to X being sGG). ∂αn−2, n = − ¯∂Ωn−1, n−1. 7 Thus, there exists βn−2, n ∈ C ∞ n−2, n(X, C) such that ∂βn−2, n = − ¯∂Ωn−1, n−1. (11) Consequently, ∂(αn−2, n + βn−2, n) = − ¯∂(Ωn−1, n−1 + Ωn−1, n−1), hence Γ1 := (αn−2, n + βn−2, n) + (Ωn−1, n−1 + Ωn−1, n−1) + (αn−2, n + βn−2, n) is a (2n − 2)-form such that dΓ1 = 0 and (12) Note that βn−2, n defines indeed an E2-class since it is ¯∂-closed (for bidegree reasons) and ∂βn−2, n is ¯∂-exact (by construction). T ({Γ1}DR) = [[αn−2, n + βn−2, n] ¯∂]d1 ∈ En−2, n (X). 2 We also get ∂(αn−2, n − βn−2, n) = − ¯∂(Ωn−1, n−1 − Ωn−1, n−1), hence Γ2 := (βn−2, n − αn−2, n) + (Ωn−1, n−1 − Ωn−1, n−1) + (αn−2, n − βn−2, n) is a (2n − 2)-form such that dΓ2 = 0 and T ({Γ2}DR) = [[αn−2, n − βn−2, n] ¯∂]d1 ∈ En−2, n 2 (X). (13) Putting (12) and (13) together, we finally get a d-closed (2n − 2)-form satisfying the condition Γ1 + Γ2 2 = βn−2, n + Ωn−1, n−1 + αn−2, n T(cid:18)(cid:26) Γ1 + Γ2 2 (cid:27)DR(cid:19) =(cid:20)[αn−2, n] ¯∂(cid:21)d1 ∈ En−2, n 2 (X). Thus, T is surjective. (cid:3) When the manifold X is sGG, we can take the surjectivity of the canonical linear map T : DR (X, C) −→ En−2, n (X) of Proposition 3.1 further by showing that every Hermitian metric ω DR (X, C) that is a section of T . We will need H 2n−2 on X defines a natural injection of En−2, n the following Laplace-type pseudo-differential operator (X) into H 2n−2 2 2 e∆ := ∂p′′∂⋆ + ∂⋆p′′∂ + ¯∂ ¯∂⋆ + ¯∂⋆ ¯∂ : C ∞ p, q(X, C) −→ C ∞ p, q(X, C) induced in every bidegree (p, q) by any fixed Hermitian metric ω on X. (All the formal adjoints are computed w.r.t. the L2 inner product defined by ω and so is the orthogonal projection p′′ = p′′ ω onto the harmonic space ker ∆′′, where ∆′′ := ¯∂ ¯∂⋆ + ¯∂⋆ ¯∂ is the usual ¯∂-Laplacian induced by ω.) This operator was introduced in [Pop16] where it was shown that every double class [[αp, q] ¯∂]d1 ∈ Ep, q (X) 2 has a unique representative lying in the kernel of e∆ = e∆ω (cf. [Pop16, Theorem 1.1]). 8 Proposition 3.4. Let X be a compact complex sGG manifold with dimCX = n and let ω be an arbitrary Hermitian metric on X. For any class [[αn−2, n] ¯∂]d1 ∈ En−2, n harmonic representative of [[αn−2, n] ¯∂]d1, let Ωn−1, n−1 solution of the equation ¯∂Ωn−1, n−1 = −∂αn−2, n minimal L2 ω-norm solution of the equation ∂βn−2, n = − ¯∂ Ωn−1, n−1 ω-norm n−2, n(X, C) be the n−1, n−1(X, C) be the minimal L2 be the e∆ω- (cf. (10)) and let βn−2, n (X), let αn−2, n (cf. (11)). ∈ C ∞ ∈ C ∞ ω ω ω ω ω 2 The linear map jω : En−2, n 2 (X) −→ H 2n−2 DR (X, C), jω([[αn−2, n] ¯∂]d1) = {βn−2, n ω + Ωn−1, n−1 ω + αn−2, n ω }DR, is injective and T ◦ jω is the identity map of En−2, n Proof. Suppose that jω([[αn−2, n] ¯∂]d1) = 0 ∈ H 2n−2 (X). Then, there exists a smooth (2n − 3)-form u = un, n−3 + un−1, n−2 + un−2, n−1 + un−3, n such that = ∂un−3, n + ¯∂un−2, n−1. Since ¯∂un−3, n = 0, βn−2, n ω this further implies that [[αn−2, n] ¯∂]d1 = [[αn−2, n DR (X, C) for some [[αn−2, n] ¯∂]d1 ∈ En−2, n = du. This implies that αn−2, n +Ωn−1, n−1 +αn−2, n (X). ω ω ω 2 2 ] ¯∂]d1 = 0. Thus, jω is injective. (X) follows immediately from the definitions. ω The equality T ◦ jω = IdEn−2, n 2 (cid:3) 3.2 Cohomology and sG metrics We now introduce strongly Gauduchon (sG) metrics into our discussion. Definition 3.5. Let X be a compact complex manifold with dimCX = n. (a) For every strongly Gauduchon metric (if any) ω > 0 on X, ¯∂ωn−1 is ∂-exact. Let us denote by Γn−2, n ω ∈ C ∞ n−2, n(X, C) the (unique) solution of minimal L2 ω-norm of the equation Since ¯∂Γn−2, n ω = 0 (for bidegree reasons) and ∂Γn−2, n ω ∈ Im ¯∂, Γn−2, n ω defines an element in En−2, n 2 (X). We consider the following subset ∂Γn−2, n ω = − ¯∂ωn−1. (14) SX :=(cid:26)(cid:20)[Γn−2, n ω ] ¯∂(cid:21)d1 ω is an sG metric on X(cid:27) ⊂ En−2, n 2 (X) that we call the E2sG-cone of X. The real (2n − 2)-form Γω := Γn−2, n ω + ωn−1 + Γn−2, n ω is d-closed, so it defines a real De Rham cohomology class {Γω}DR. We consider the following subset eSX :=(cid:26){Γω}DR ω is an sG metric on X(cid:27) ⊂ H 2n−2 DR (X, R) that we call the De Rham E2sG-cone of X. (b) We also define the following variant of the E2sG-cone of X by dropping the L2 ω-norm mini- mality requirement on the solution Γn−2, n of equation (14): ∃ ω Hermitian metric such that ∂Γn−2, n = − ¯∂ωn−1(cid:27) ⊂ En−2, n bSX =(cid:26)(cid:20)[Γn−2, n] ¯∂(cid:21)d1 Every metric ω involved in (15) is sG, so we obviously have SX ⊂ bSX. We do not know whether the reverse inclusion holds. (X). (15) 2 9 The manifold X is strongly Gauduchon if and only if SX is non-empty. We shall now prove that SX (resp. eSX) is indeed a cone (i.e. a subset that is stable under DR (X, R)). We need a few preliminaries. multiplications by positive scalars) in En−2, n (X) (resp. H 2n−2 2 Lemma 3.6. Let X be a compact complex manifold with dimCX = n. (i) For any Hermitian metric ω on X and any positive real λ, the formal adjoints of ¯∂ w.r.t. the metrics λ ω and ω, as well as the corresponding ¯∂-Laplacians, are related by the formulae ¯∂⋆ λ ω = 1 λ ¯∂⋆ ω and ∆′′ λ ω = 1 λ ∆′′ ω (16) in all bidegrees. (ii) For any strongly Gauduchon metric ω on X and any positive real λ, the forms Γn, n−2 λ ω := Γn−2, n λ ω and Γn, n−2 ω := Γn−2, n ω (see Definition 3.5) are related by the formula Consequently, we also have Γλ ω = λn−1 Γω for any sG metric ω on X. Γn, n−2 λ ω = λn−1 Γn, n−2 ω . Proof. (i) Let us fix an arbitrary bidegree (p, q). For any forms α, β of respective bidegrees (p, q −1) and (p, q), we have hh ¯∂α, βiiλ ω = λn Z X h ¯∂α, βiλ ω ωn n! = λn λp+q Z X h ¯∂α, βiω ωn n! = λn λp+q hh ¯∂α, βiiω = λn λp+q hhα, ¯∂⋆ ωβiiω and Since hh ¯∂α, βiiλ ω = hhα, ¯∂⋆ hhα, ¯∂⋆ λn λp+q−1 hhα, ¯∂⋆ λ ωβiiλ ω, the above formulae imply λ ωβiiλ ω = λ ωβiiω. λn λp+q hhα, ¯∂⋆ ωβiiω = λn λp+q−1 hhα, ¯∂⋆ λ ωβiiω, i.e. hhα, 1 λ ωβiiω = hhα, ¯∂⋆ ¯∂⋆ λ ωβiiω for all forms α and β. This proves the first formula in (16). Since ∆′′ = ¯∂ ¯∂⋆ + ¯∂⋆ ¯∂ (when ∆′′ and ¯∂⋆ are computed w.r.t. the same metric), the latter formula in (16) follows immediately from the former. (ii) By Definition 3.5, Γn, n−2 ω is the minimal L2 ω-norm solution of equation ¯∂Γn, n−2 = −∂ωn−1. Consequently, the Neumann formula spells Γn, n−2 ω ′′−1 = −∆ ω ¯∂⋆ ω (∂ωn−1). (17) Thus, we get: Γn, n−2 λ ω = −∆ ′′−1 λ ω ¯∂⋆ λ ω (∂(λ ω)n−1) = −λn−1 ∆ ′′−1 ω ¯∂⋆ ω (∂ωn−1) = λn−1 Γn, n−2 ω , 10 where we used the analogue of (17) for λ ω to get the first identity and (17) again to get the third identity. (cid:3) cones in the C-vector space En−2, n Lemma 3.7. Let X be a compact complex manifold with dimCX = n. The sets SX and bSX are Moreover, the cone bSX is convex. (X), while the set eSX is a cone in H 2n−2 ] ¯∂]d1 ∈ SX and µ > 0 be arbitrary. Let λ > 0 be the unique positive real such DR (X, R). 2 Proof. Let [[Γn−2, n that λn−1 = µ. We have ω µ [[Γn−2, n ω ] ¯∂]d1 = [[λn−1 Γn−2, n ω ] ¯∂]d1 = [[Γn−2, n λ ω ] ¯∂]d1, where we used (ii) of Lemma 3.6 to get the last identity. Now, λ ω is a strongly Gauduchon metric if ω is one, so [[Γn−2, n ] ¯∂]d1 ∈ SX . λ ω Consequently, SX is stable under multiplications by positive scalars, hence it is a cone. The same goes for eSX since (ii) of Lemma 3.6 also applies to Γω. That bSX is a cone is trivial. To prove the convexity of bSX , it suffices to show that bSX is stable under additions. This is + , where ω0 > 0 is the unique positive definite C ∞ (1, 1)-form on X such that (cid:3) for i ∈ {1, 2} and ωi Hermitian metrics on X, then ∂(Γn−2, n immediate since if ∂Γn−2, n Γn−2, n 2 ωn−1 0 = ωn−1 2 > 0. Therefore, [[Γn−2, n ) = − ¯∂ωn−1 1 + ωn−1 = − ¯∂ωn−1 + Γn−2, n 0 1 1 2 i i The De Rham E2sG-cone eSX depends on the complex structure of X and we now show this dependence to be lower semicontinuous in the sense described below in families of sGG manifolds. It actually suffices to assume that one fibre is sGG as all the nearby fibres are then sGG by the deformation openness of the sGG property (cf. [PU18, Corollary 1.7]). ] ¯∂]d1 ∈ bSX . Definition 3.8. Let π : X −→ ∆ be a holomorphic family of compact complex n-dimensional manifolds over a ball ∆ ⊂ CN centred at the origin. Suppose that the fibre X0 := π−1(0) is strongly Gauduchon. Let X stand for the C ∞ manifold that underlies the fibres Xt with t ∈ ∆. With every sG metric ω on X0, we associate a local section τω of the constant real vector bundle H2n−2 R −→ ∆ whose fibre is the real De Rham cohomology space H 2n−2 As in Definition 3.5, we let Γω := Γn−2, n +ωn−1+Γn−2, n ω-norm solution Γn−2, n by the minimal L2 ¯∂ := ¯∂0.) The component (Γω)n−1, n−1 positive definite if t is close enough to 0, by the continuity of the dependence on t of (Γω)n−1, n−1 the positivity of (Γω)n−1, n−1 definite smooth (1, 1)-form ωt on Xt such that ωn−1 be the real d-closed (2n−2)-form defined = − ¯∂ωn−1. (We put ∂ := ∂0 and of Γω of type (n − 1, n − 1) for the complex structure of Xt is and = ωn−1 > 0. Hence, for every t close to 0, there exists a unique positive > 0. In particular, ω0 = ω. of the equation ∂Γn−2, n t = (Γω)n−1, n−1 DR (X, R) as follows. ω ω ω ω 0 t t t Since dΓω = 0, the form ¯∂tωn−1 C ∞ n−2, n(Xt, C) be the minimal L2 t ωt-norm solution of the equation is ∂t-exact (so ωt is an sG metric on Xt). We let Γn−2, n ∈ ωt and we consider the real d-closed (2n − 2)-form on X defined as ∂tΓn−2, n ωt = − ¯∂tωn−1 t (18) Γω(t) := Γn−2, n ωt + ωn−1 t + Γn−2, n ωt 11 for t close to 0. In particular, Γω(0) = Γω. Finally, we put for all t in a sufficiently small neighbourhood U (depending on ω) of 0 in ∆. τω(t) := {Γω(t)}DR ∈ eSXt ⊂ H 2n−2 DR (X, R) The lower semicontinuity result for the De Rham E2sG-cone eSX ⊂ H 2n−2 complex structure of X varies is the following DR (X, R) when the Theorem 3.9. Let π : X −→ ∆ be a holomorphic family of compact complex n-dimensional man- ifolds over a ball ∆ ⊂ CN centred at the origin. Suppose that the fibre X0 := π−1(0) is an sGG manifold. For every sG metric ω on X0, the section τω of the constant real vector bundle H2n−2 R −→ ∆ on a small neighbourhood of 0 in ∆ constructed in Definition 3.8 is C ∞. there is such an extension for every representative Γω of the given De Rham class {Γω}DR defined In particular, every element {Γω}DR of the De Rham E2sG-cone eSX0 of X0 extends to a C ∞ family of elements {Γω(t)}DR of the De Rham E2sG-cones eSXt of the nearby fibres Xt. Moreover, by an sG metric ω. So, in this sense, the De Rham E2sG-cone eSX0 of X0 can only be "smaller" than the De Rham E2sG-cones eSXt of the nearby fibres Xt. Proof. By the well-known Neumann formula, the minimal L2 ωt-norm solution of equation (18) is Γn−2, n ωt = −(∂t)⋆ ωt∆ ′−1 ωt ( ¯∂tωn−1 t ), t ωt + (∂t)⋆ ωt and (∂t)⋆ where (∂t)⋆ ∆′ ωt = ∂t(∂t)⋆ Now, the (n − 1, n)-form ¯∂tωn−1 ωt∂t is the ∂-Laplacian induced by ωt and ∆′−1 ωt ωt. Moreover, the classical Kodaira-Spencer theory (cf. ωt is the formal adjoint of ∂t w.r.t. the L2 inner product induced by the metric ωt, while stands for its Green operator. varies in a C ∞ way with t and so do the differential operators [KS60]) applied to the C ∞ ωt)t∈∆ of elliptic differential operators acting in bidegree (n − 1, n) ensures that the family ωt (that are (Xt, C) by the Hodge isomorphism) are independent (Xt, C) by conjugation, while (Xt, C), so its dimension equals the Hodge number ∆′ family (∆′ (∆′−1 isomorphic to the ∂-cohomology spaces H n−1, n of t. However, H n−1, n the latter vector space is Serre-dual to H 0, 1 h0, 1 ¯∂ (t) of the fibre Xt for every t. ωt )t∈∆ of their Green operators is again C ∞ if the dimensions of the kernels ker ∆′ (Xt, C) is C-anti-linearly isomorphic to H n, n−1 ¯∂ Here is where the sGG assumption on the fibre X0 comes in. By [PU18, Corollary 1.7], it ensures that the Hodge numbers h0, 1 ¯∂ (t) are independent of t when t varies in a small enough neighbourhood of 0. Thus, the Green operators ∆′−1 in bidegree (n − 1, n), hence also the (n − 2, n)-forms Γn−2, n , ωt vary in a C ∞ way with t near 0. Since so also do (for trivial reasons) the (n − 1, n − 1)-forms ωn−1 , we infer that the smooth (2n − 2)-forms ωt t ¯∂ ∂ ∂ Γω(t) := Γn−2, n ωt + ωn−1 t + Γn−2, n ωt vary in a C ∞ way with t in a small enough neighbourhood of 0. The application of the De Rham cohomology class being a smooth operation, we conclude that τω(t) := {Γω(t)}DR depends in a C ∞ way on t varying in a small enough neighbourhood of 0 ∈ ∆. (cid:3) 12 3.3 Real-valued cohomology We shall now deal with the real version of some of the objects introduced in §.3.1 for the sake of enhanced flexibility. Definition 3.10. Let X be a compact complex manifold with dimCX = n. (a) For any element [[αn−2, n] ¯∂]d1 ∈ En−2, n n−2, n(X, C) of this double class, we know from (5) that ¯∂α = 0 (trivial here for bidegree reasons) and that there exists a (not necessarily real and non-unique) form Ωn−1, n−1 ∈ C ∞ n−1, n−1(X, C) such that ∂αn−2, n = − ¯∂Ωn−1, n−1. (X) and any representative αn−2, n ∈ C ∞ 2 We refer to any such form Ωn−1, n−1 as an (n − 1, n − 1)-potential of αn−2, n. (b) We define the real part En−2, n (X)R of the C-vector space En−2, n 2 2 (X) by selecting the classes representable by forms admitting a real (n − 1, n − 1)-potential: En−2, n 2 (X)R :=(cid:26)[[αn−2, n] ¯∂]d1 ∈ En−2, n 2 (X) ∃αn−2, n representative having a real potential Ωn−1, n−1(cid:27). By definition, En−2, n 2 (X)R is a real vector subspace of En−2, n 2 (X). We shall now consider the real version of the map T introduced in §.3.1. Lemma 3.11. Let X be a compact complex manifold with dimCX = n. (i) The following inclusion holds: En−2, n (ii) The restriction to H 2n−2 2 (X)R ⊂ ker dn−2, n 2 . DR (X, R) of the map T defined in Proposition 3.1, namely the map TR : H 2n−2 DR (X, R) −→ En−2, n 2 (X)R, {α}DR 7→(cid:20)[αn−2, n] ¯∂(cid:21)d1 , assumes its values in the real space En−2, n 2 (X)R and is surjective. Proof. (i) Let [[αn−2, n] ¯∂]d1 ∈ En−2, n form Ωn−1, n−1. Thanks to (6), we have 2 (X)R with ∂αn−2, n = − ¯∂Ωn−1, n−1 for some real (n − 1, n − 1)- d2([[αn−2, n] ¯∂]d1) = −[[∂Ωn−1, n−1] ¯∂]d1 = [[ ¯∂αn−2, n] ¯∂]d1 = 0 ∈ En, n−1 2 (X), because by conjugating the identity defining Ωn−1, n−1 and using the fact that Ωn−1, n−1 is real, we get ∂Ωn−1, n−1 = − ¯∂αn−2, n. Therefore, [[αn−2, n] ¯∂]d1 ∈ ker dn−2, n (ii) Let {α}DR ∈ H 2n−2 2 . DR (X, R) and pick a real representative α = αn, n−2 + αn−1, n−1 + αn−2, n. Since α is real, αn−1, n−1 is real. Since α is d-closed, ∂αn−2, n = − ¯∂αn−1, n−1. Thus, αn−1, n−1 is a real (n − 1, n − 1)-potential of αn−2, n, so T ({α}DR) = [[αn−2, n] ¯∂]d1 ∈ En−2, n (X)R. 2 To prove that T is surjective, let [[αn−2, n] ¯∂]d1 ∈ En−2, n 2 some real (n − 1, n − 1)-form Ωn−1, n−1. Then, the (2n − 2)-form (X)R with ∂αn−2, n = − ¯∂Ωn−1, n−1 for α := αn−2, n + Ωn−1, n−1 + αn−2, n 13 is real, d-closed and T ({α}DR) = [[αn−2, n] ¯∂]d1 ∈ En−2, n 2 (X)R. (cid:3) 2 2 2 En−2, n X defined in §.3.1 as an open cone. We are now in a position to see that the real space En−2, n (X)R hold and the cone bSX is open in En−2, n (X)R contains the E2sG-cone bSX of Lemma 3.12. Let X be a compact complex manifold with dimCX = n. The inclusions SX ⊂ bSX ⊂ Proof. The inclusion SX ⊂ bSX is obvious and has already been noticed. Let [[Γn−2, n] ¯∂]d1 ∈ bSX be the inclusion bSX ⊂ En−2, n arbitrary. So, there exists a representative Γn−2, n ∈ C ∞ ω on X such that ∂Γn−2, n = − ¯∂ωn−1. Since ωn−1 is real, [[Γn−2, n] ¯∂]d1 ∈ En−2, n n−2, n(X, C) of this E2-class and an sG metric (X)R. This proves (X)R and let ε > 0. Then, there exists a real form Ωn−1, n−1 ∈ Let [[αn−2, n] ¯∂]d1 ∈ En−2, n n−1, n−1(X, C) such that ¯∂Ωn−1, n−1 = −∂αn−2, n. We get C ∞ (X)R. (X)R. 2 2 2 ∂(Γn−2, n + ε αn−2, n) = − ¯∂(ωn−1 + ε Ωn−1, n−1). On the other hand, the (n−1, n−1)-form ωn−1 +ε Ωn−1, n−1 is real for every ε and is positive definite if ε > 0 is small enough. Therefore, for every small ε > 0, there exists a unique positive definite ε = − ¯∂(Γn−2, n + ε αn−2, n), hence (1, 1)-form ρε > 0 such that ρn−1 ∂ρn−1 is ¯∂-exact, so ρε is a strongly Gauduchon metric on X. Consequently, ε = ωn−1 + ε Ωn−1, n−1. We get ∂ρn−1 ε [[Γn−2, n] ¯∂]d1 + ε [[αn−2, n] ¯∂]d1 ∈ bSX (X)R. for every small ε > 0. This proves that bSX is open in En−2, n 2 3.4 Duality of positive cones in the E2 cohomology (cid:3) It was proved in [PU18b] (by means of the pseudo-differential Laplacian e∆ introduced in [Pop16] that gives a Hodge theory for the second page of the Frolicher spectral sequence) that for any compact complex n-dimensional manifold X and every p, q ∈ {0, . . . , n}, the canonical bilinear pairing Ep, q 2 (X) × En−p, n−q 2 (X) −→ C, (cid:18)[[α] ¯∂]d1, [[β] ¯∂]d1(cid:19) 7→Z X α ∧ β, (19) is well defined (i.e. volved) and non-degenerate. Hence, it defines a Serre-type duality between Ep, q independent of the choices of representatives of the E2-cohomology classes in- (X) and En−p, n−q (X). 2 2 Under this duality, the closure of our E2sG cone bSX ⊂ En−2, n (X), consisting of those E2-classes of type (n − 2, n) that are "positive" in the sense of Definition 3.5, has a dual cone in E2, 0 (X) that we will now describe. To this end, we will introduce ad hoc notions of real and positive (2, 0)-forms and currents that run counter to the standard definitions of real and positive forms and currents of bidegree (p, p), but propose a not so far-fetched analogue thereof in this bidegree that is relevant to holomorphic symplectic geometry. A possible extension of this geometry on sGG manifolds is one of our motivations and will hopefully be attempted in future work. 2 2 In this subsection, we will establish an E2 analogue for the bidegrees (2, 0) and (n − 2, n) [Lam99, lemme 3.3]) between Demailly's pseudo-effective cone E(X) ⊂ of Lamari's duality (cf. 14 H 1, 1 BC(X, R) (consisting of the Bott-Chern cohomology classes of all the d-closed, positive (1, 1)- currents T ≥ 0 on X, see [Dem92]) and the closure of the Gauduchon cone GX ⊂ H n−1, n−1 (X, R) introduced in [Pop15] (consisting of the Aeppli cohomology classes of all the Gauduchon metrics ωn−1 > 0 on X). A We will assume throughout this subsection that X is an sGG manifold. This will guarantee that every Gauduchon metric is actually strongly Gauduchon (see [PU18a]). Definition 3.13. Let X be an sGG compact complex n-dimensional manifold. (i) We consider the following sets: V = (cid:26)Γn−2, n ∈ C ∞ E = (cid:26)Γn−2, n ∈ C ∞ ER = (cid:26)Γn−2, n ∈ C ∞ n−2, n(X, C) ∃ ω Hermitian metric such that ∂Γn−2, n = − ¯∂ωn−1(cid:27), n−2, n(X, C) ∂Γn−2, n ∈ Im ¯∂(cid:27), n−2, n(X, C) ∃ Ωn−1, n−1 real form such that ∂Γn−2, n = − ¯∂Ωn−1, n−1(cid:27). Thus, V ⊂ ER ⊂ E ⊂ C ∞ n−2, n(X, C) and E consists of the smooth (n−2, n)-forms that are E2-closed (their ¯∂-closedness is automatic for bidegree reasons), while ER consists of the real (in this ad hoc sense) such forms and V consists of the positive (in this ad hoc sense) such forms. Note that any metric ω featuring in the definition of V is automatically strongly Gauduchon (or, equivalently, Gauduchon since X is assumed sGG). (ii) Fix an arbitrary Hermitian metric γ on X. Let p(γ) n−2, n(X, C) −→ Im ¯∂ be the γ-inner product onto the closed subspace of ¯∂-exact (n − 2, n)- Im ¯∂ : C ∞ orthogonal projection w.r.t. forms, induced by the standard Hodge-theoretical L2 the L2 γ-orthogonal 3-space decomposition n−2, n(X, C) = ker ∆′′ ⊕ Im ¯∂ ⊕ Im ¯∂⋆. C ∞ We consider the following sets: n−2, n(X, C) ∃ ω Hermitian metric such that p(γ) Im ¯∂(∂Γn−2, n) = − ¯∂ωn−1(cid:27), Uγ = (cid:26)Γn−2, n ∈ C ∞ C ∞ n−2, n(X, R)γ = (cid:26)αn−2, n ∈ C ∞ n−2, n(X, C) ∃ βn−1, n−1 real form such that p(γ) Im ¯∂(∂αn−2, n) = − ¯∂βn−1, n−1(cid:27). n−2, n(X, R)γ ⊂ C ∞ Thus, Uγ ⊂ C ∞ n−2, n(X, C) and Uγ consists of the γ-positive (in this ad hoc sense) smooth (n − 2, n)-forms, while C ∞ n−2, n(X, R)γ consists of the γ-real (in this ad hoc sense) such forms. Unlike the sets defined under (i), these sets are not subjected to any E2-closedness condition. In particular, the following inclusions hold: V ⊂ Uγ and ER ⊂ C ∞ n−2, n(X, R)γ 15 for every Hermitian metric γ on X. (iii) In the context of (ii), by a γ-real current of bidegree (2, 0) on X we mean any continuous R-linear form τ 2, 0 : C ∞ n−2, n(X, R)γ −→ R. By such a current being γ-positive we mean that τ 2, 0 evaluates non-negatively on every element of Uγ. (So, in particular, the zero current τ 2, 0 = 0 is γ-positive.) By a γ-real current τ 2, 0 of bidegree (2, 0) being E2-exact we mean that τ 2, 0 vanishes identically on the R-vector space ER of "real" E2-closed (n − 2, n)-forms defined under (i). 2 The following properties of the above sets are immediate to check. Lemma 3.14. (a) The set E is a closed C-vector subspace of C ∞ C ∞ cones in ER, respectively C ∞ n−2, n(X, R)γ are closed R-vector subspaces of C ∞ n−2, n(X, C), the sets ER and n−2, n(X, C), while V and Uγ are open convex n−2, n(X, R)γ. (b) The following identity holds: Uγ ∩ ER = V. j γ Proof. (a) The closedness conclusion for E follows from the well-known fact (itself a consequence of standard elliptic theory on compact manifolds) that Im ¯∂ is closed in the space of C ∞ forms in which it lies. j j = −∂Γn−2, n j = −∆′′−1 γ ¯∂⋆ γ∂Γn−2, n To see that ER is closed in C ∞ n−1, n−1(X, C) and the hypothesis Γn−2, n j ∈ C ∞ n−1, n−1(X, C) be the solution of minimal L2 n−2, n(X, C), we need one further step. Let Γn−2, n n−2, n(X, C) in the C ∞ topology as j → +∞, where Γn−2, n j . (So, the set of all the solutions is the affine subspace Γn−1, n−1 → Γn−2, n ∈ ∈ ER for every j ∈ N. For every j, let γ-norm of the equation + ker ¯∂ ⊂ j n−1, n−1(X, R) 6= ∅, n−1, n−1(X, C) is the real vector subspace of real forms.) Then Im ¯∂ ∋ n−1, n−1(X, R) ⊂ C ∞ → ∂Γn−2, n in the C ∞ topology as j → +∞, so ∂Γn−2, n ∈ Im ¯∂ since Im ¯∂ is closed. γ-norm of → Γn−1, n−1 in the C ∞ topology as j → +∞ ¯∂⋆ γ is continuous in the C ∞ topology. Since n−1, n−1(X, C), n−1, n−1(X, R) 6= ∅. This means that Γn−2, n ∈ ER. Thus, ER is closed C ∞ Γn−1, n−1 j ¯∂Γn−1, n−1 j C ∞ where C ∞ ∂Γn−2, n Moreover, Γn−1, n−1 = −∆′′−1 the equation ¯∂Γn−1, n−1 = −∂Γn−2, n, so Γn−1, n−1 because the restriction to Im ¯∂ of the operator ∆′′−1 (Γn−1, n−1 we get (Γn−1, n−1 + ker ¯∂) ∩ C ∞ in C ∞ n−1, n−1(X, C) is the solution of minimal L2 n−1, n−1(X, R) 6= ∅ for all j ∈ N and C ∞ n−1, n−1(X, R) is closed in C ∞ ∈ ER means that (Γn−1, n−1 +ker ¯∂)∩C ∞ ¯∂⋆ γ∂Γn−2, n ∈ C ∞ +ker ¯∂)∩C ∞ j j j γ n−2, n(X, C). The closedness of C ∞ n−2, n(X, R)γ in C ∞ n−2, n(X, C) can be proved in the same way since the projection p(γ) Im ¯∂ is continuous w.r.t. the C ∞ topology. The convexity of V and Uγ follows from the linearity of the operators ∂, ¯∂ and p(γ) Im ¯∂ involved in their definitions and from the convexity of the set of Gauduchon metrics (itself a consequence of the existence of a unique positive definite (n − 1)st root for every positive definite (n − 1, n − 1)-form). 2This last notion is in keeping with the usual duality according to which a current is exact (w.r.t. a given coho- mology) if and only if it vanishes identically on the closed (w.r.t the same cohomology) C∞ forms of complementary bidegree. 16 Let us prove that Uγ is open in C ∞ similar way.) Let Γn−2, n ∈ Uγ and αn−2, n ∈ C ∞ a Hermitian metric ω and a real form βn−1, n−1 ∈ C ∞ n−2, n(X, R)γ. (The openness of V in ER can be proved in a n−2, n(X, R)γ be arbitrary. By definition, there exist n−1, n−1(X, R) such that p(γ) Im ¯∂(∂Γn−2, n) = − ¯∂ωn−1 Thus, for every constant ε > 0, we get p(γ) Im ¯∂(∂(Γn−2, n + ε αn−2, n)) = − ¯∂(ωn−1 + ε βn−1, n−1). Since βn−1, n−1 is real and ωn−1 is positive definite, ωn−1 + ε βn−1, n−1 is positive definite for all sufficiently small ε > 0. Therefore, Γn−2, n + ε αn−2, n ∈ Uγ. This proves that Uγ is open in C ∞ p(γ) Im ¯∂(∂αn−2, n) = − ¯∂βn−1, n−1. and n−2, n(X, R)γ. (b) To prove the inclusion "⊂", let Γn−2, n ∈ Uγ ∩ ER. Since Γn−2, n ∈ ER, ∂Γn−2, n ∈ Im ¯∂, so p(γ) Im ¯∂(∂Γn−2, n) = ∂Γn−2, n. Thus, ∂Γn−2, n = − ¯∂ωn−1 for some Hermitian metric ω thanks to Γn−2, n lying in Uγ. Therefore, Γn−2, n ∈ V . This proves the inclusion "⊂". The reverse inclusion is obvious. (cid:3) The last preliminary remark that we make serves as an example pointing out a very particular way of constructing real-valued linear maps on ER. Such maps occur below in a more general form. (See the last hypothesis of Proposition 3.16.) Lemma 3.15. If a form θ2, 0 ∈ C ∞ ¯∂ξ1, 0 is real, then RX θ2, 0 ∧ Γn−2, n is real for every Γn−2, n ∈ ER. 2, 0(X, C) is of the shape θ2, 0 = ∂ξ1, 0 such that the (1, 1)-form Proof. Let Γn−2, n ∈ ER. Then, ∂Γn−2, n = − ¯∂Ωn−1, n−1 for some real (n − 1, n − 1)-form Ωn−1, n−1. Applying Stokes's Theorem twice, we get Z X θ2, 0 ∧ Γn−2, n =Z ∂ξ1, 0 ∧ Γn−2, n = −Z ξ1, 0 ∧ ¯∂Ωn−1, n−1 = −Z X X X ¯∂ξ1, 0 ∧ Ωn−1, n−1. The last quantity is real since both forms ¯∂ξ1, 0 and Ωn−1, n−1 are real. (cid:3) We are now in a position to prove the duality result we have been aiming for. Both the statement and the proof parallel those of Lemma 3.3 in [Lam99]. Proposition 3.16. Let X be an sGG compact complex n-dimensional manifold on which an arbi- trary Hermitian metric γ has been fixed. Let θ2, 0 ∈ C ∞ 2, 0(X, C) satisfy the condition Z X θ2, 0 ∧ Γn−2, n ≥ 0 for every Γn−2, n ∈ C ∞ ∂Γn−2, n = − ¯∂ωn−1 (i.e. for every Γn−2, n ∈ ER. n−2, n(X, C) for which there exists a Hermitian metric ω on X such that for every Γn−2, n ∈ V ). Suppose, moreover, that RX θ2, 0 ∧ Γn−2, n ∈ R Then, there exists a γ-positive current τ 2, 0 : C ∞ n−2, n(X, R)γ −→ R of bidegree (2, 0) on X such that θ2, 0 − τ 2, 0 is E2-exact in the sense that it vanishes identically on ER. 17 Note that if we assume θ2, 0 to be E2-closed (i.e. ¯∂θ2, 0 = 0 and ∂θ2, 0 ∈ Im ¯∂, which in bidegree and the integral (2, 0) is equivalent to assuming that dθ2, 0 = 0), it defines a class [[θ2, 0] ¯∂]d1 ∈ E2, 0 RX θ2, 0 ∧ Γn−2, n is independent of the choice of representative of this class. Thus, Proposition 3.16 implies that the dual of the closure of the E2sG-cone bSX defined in (15) under the duality E2, 0 [[θ2, 0] ¯∂]d1 "representable" by γ-positive currents τ 2, 0 : C ∞ (X) −→ C is the closed convex cone in E2, 0 (X) consisting of the E2-classes n−2, n(X, R)γ −→ R. (X) × En−2, n 2 2 2 2 Proof of Proposition 3.16. We follow closely Lamari's arguments of the proof of Lemma 3.3. [Lam99]. The form θ2, 0 defines a C-linear map in θ2, 0 : C ∞ n−2, n(X, C) −→ C, Γn−2, n 7→Z X θ2, 0 ∧ Γn−2, n. The hypothesis imposed on θ2, 0 translates to θ2, 0 V ≥ 0. Thus, there are two cases. θ2, 0 ER Case 1. Suppose there exists Γn−2, n ≡ 0. Indeed, fix an arbitrary Γn−2, n ∈ ER and let Γn−2, n 0 t ∈ V ⊂ ER such that RX θ2, 0 ∧ Γn−2, n 0 = 0. This implies that Γn−2, n t ∈ ER for all t ∈ [0, 1] since ER is convex. Moreover, for all t ∈ R, we have := (1 − t) Γn−2, n 0 + t Γn−2, n for t ∈ R. Then f (t) :=Z X θ2, 0 ∧ Γn−2, n t = (1 − t) Z X θ2, 0 ∧ Γn−2, n 0 + t Z X θ2, 0 ∧ Γn−2, n = t Z X θ2, 0 ∧ Γn−2, n. t ∈ V for all t close enough to 0. Since θ2, 0 In particular, f (0) = 0. Meanwhile, V is open in ER (cf. Lemma 3.14) and Γn−2, n Γn−2, n ∈ V , so V ≥ 0, we infer that f (t) ≥ 0 for all t ∈ [−ε, ε] for some small ε > 0. This means that t RX θ2, 0 ∧ Γn−2, n ≥ 0 for all t ∈ [−ε, ε], which is impossible unlessRX θ2, 0∧Γn−2, n = 0. This proves that θ2, 0 ≡ 0, so we can choose τ 2, 0 = 0 (which is γ-positive). ER 0 Case 2. Suppose that θ2, 0 V > 0. Let F ⊂ ER be the kernel of the restriction θ2, 0 ER : ER −→ R. Thus, F has real codimension 1 in ER and Uγ ∩ F = ∅. To see the last identity, suppose there exists Γn−2, n ∈ Uγ ∩ F . Then, RX θ2, 0 ∧ Γn−2, n = 0 because ER). Meanwhile, RX θ2, 0 ∧ Γn−2, n > 0 because F ⊂ ER, so Γn−2, n ∈ Uγ ∩ ER = V Γn−2, n ∈ F = ker(θ2, 0 (see (b) of Lemma 3.14 for the last identity) and θ2, 0 V > 0. This is a contradiction. Since Uγ is a convex open subset of C ∞ n−2, n(X, R)γ and Uγ and F are disjoint, the Hahn-Banach Separation Theorem allows us to separate them. Consequently, there exists a continuous R-linear form n−2, n(X, R)γ, F is a convex closed subset of C ∞ such that l2, 0 Uγ > 0 and l2, 0 F = 0. The first condition implies that the γ-real current l2, 0 of bidegree (2, 0) is γ-positive. l2, 0 : C ∞ n−2, n(X, R)γ −→ R 18 Let Γn−2, n 1 ∈ V . Then RX θ2, 0 ∧ Γn−2, n λ > 0 such that RX θ2, 0 ∧ Γn−2, n > 0, so there exists a constant = 0. But we also have (θ2, 0 − λ l2, 0)F = 0. Since the real codimension of F in ER is 1, we get that (θ2, 0 − λ l2, 0)ER = 0. If we put τ 2, 0 := λ l2, 0, we are done. (cid:3) = λ RX l2, 0 ∧ Γn−2, n . This means that (θ2, 0 − λ l2, 0)R Γn−2, n > 0 and RX l2, 0 ∧ Γn−2, n 1 1 1 1 1 4 The h-∂ ¯∂ property of compact complex manifolds Let X be a compact complex manifold with dimCX = n. We now consider the adiabatic limit construction of the differential operator dh = h∂ + ¯∂ (cf. (2)) that was introduced in [Pop17] for every constant h > 0. Allowing now h to be negative, some obvious properties include the following: dh = h dh−1; (i) (iii) dh1dh2 = (h1 − h2) ∂ ¯∂; (ii) d−h = −h d−h−1; in particular, dh d−h−1 = (h + ) ∂ ¯∂; 1 h (iv) h + 1 h2 + 1 dh + h(h − 1) h2 + 1 d−h−1 = d, (20) for all h ∈ R \ {0}. When a Hermitian metric ω has been fixed on X, the formal adjoint d⋆ h of dh w.r.t. ω induces together with dh a Laplace-type operator in the usual way: ∆h : dh d⋆ h + d⋆ hdh : C ∞ k (X, C) −→ C ∞ k (X, C), for every k ∈ {0, . . . , 2n}. This h-Laplacian is elliptic (cf. [Pop17]). Identity (ii) in (20) implies ∆−h = h2 ∆−h−1, for all h ∈ R \ {0}. (21) We shall now continue the study of the operators dh, both from a metric and an intrinsic angle. 4.1 Commutation relations and BKN identity for the operators dh Let us fix an arbitrary Hermitian metric ω on X. All the formal adjoints will be computed w.r.t. ω, as will the (pointwise and formal) adjoint Λ = Λω of the multiplication operator L = Lω := ω ∧ ·. Recall the standard torsion operator of type (1, 0) (cf. [Dem84]) τ = [Λ, ∂ω ∧ ·] and the Hermitian commutation relations (cf. again [Dem84]): ∂⋆ + τ ⋆ = i [Λ, ¯∂] and ¯∂⋆ + ¯τ ⋆ = −i [Λ, ∂]. We will infer the following Lemma 4.1. Let (X, ω) be a complex Hermitian manifold. For every h ∈ R \ {0}, we define the h-torsion operator of type (1, 0) induced by ω by τh := [Λ, dhω ∧ ·]. The following Hermitian h-commutation relations hold on differential forms of any degree: (a) (dh + τh)⋆ = −i [Λ, d−h]; ⋆ −h, ω ∧ ·]; (c) dh + τh = i [d (b) (dh + τ h)⋆ = i [Λ, d−h]; (d) dh + τ h = −i [d⋆ −h, ω ∧ ·]. 19 Proof. Since (b) is the conjugate of (a), while the implications (a) =⇒ (c) and (b) =⇒ (d) are obtained by taking adjoints, it suffices to prove (a). Using the above definitions and the standard Hermitian commutation relations, we get h = h∂⋆ + ¯∂⋆ = i [Λ, h ¯∂] − hτ ⋆ − i [Λ, ∂] − ¯τ ⋆ = −i [Λ, d−h] − (hτ ⋆ + ¯τ ⋆) d⋆ and hτ ⋆ + ¯τ ⋆ = [(h ∂ω ∧ ·)⋆, ω ∧ ·] + [( ¯∂ω ∧ ·)⋆, ω ∧ ·] = [(dhω ∧ ·)⋆, ω ∧ ·] = τ ⋆ h . Summing up these identities, we get (a). (cid:3) An immediate consequence is the following Corollary 4.2. Let (X, ω) be a complex Hermitian manifold. For every h ∈ R \ {0}, the following rough h-Bochner-Kodaira-Nakano (h-BKN) identity holds on differential forms of any degree: Proof. Using the h-commutation relation (a) of Lemma 4.1 for the second identity below, we get ∆h = ∆−h + [d−h, τ ⋆ −h] − [dh, τ ⋆ h ]. On the other hand, the Jacobi identity spells: ∆h = [dh, d⋆ h] = −i [dh, [Λ, d−h]] − [dh, τ ⋆ h ]. −[dh, [Λ, d−h]] + [Λ, [d−h, dh]] + [d−h, [dh, Λ]] = 0. Since [d−h, dh] = 0 whenever h 6= 0, the second term above vanishes. Meanwhile, [dh, Λ] = i (d−h + τ −h)⋆ as follows from the h-commutation relation (b) of Lemma 4.1 after replacing h with −h. Therefore, we get −i [dh, [Λ, d−h]] = [d−h, (d−h + τ −h)⋆] = ∆−h + [d−h, τ ⋆ −h] and the formula follows. (cid:3) Another immediate consequence is the following anti-commutation statement in the Kahler case. Corollary 4.3. Let (X, ω) be a compact Kahler manifold. For every h ∈ R \ {0}, the following identities hold: [dh, d⋆ −h−1] = 0 and [d−h−1, d⋆ h] = 0. Proof. The latter identity is the adjoint of the former, so it suffices to prove the former one. When h has been replaced by −h−1, the h-commutation relation (a) of Lemma 4.1 spells d⋆ −h−1 = −i [Λ, dh−1] = − i h [Λ, dh] since τh = 0 for every h when ω is Kahler and identity (i) in (20) has been used to infer the last identity. Therefore, [dh, d⋆ h [dh, [Λ, dh]] when ω is Kahler. −h−1] = − i Now, the Jacobi identity yields: Since [dh, dh] = 0 and [dh, Λ] = −[Λ, dh], we get [dh, [Λ, dh]] = 0. −[dh, [Λ, dh]] + [Λ, [dh, dh]] + [dh, [dh, Λ]] = 0. Consequently, [dh, d⋆ −h−1] = − i h [dh, [Λ, dh]] = 0 and we are done. (cid:3) 20 Taking our cue from [Dem84], we shall now refine the above BKN formula by incorporating the 1st order terms into a twisted Laplace-type operator on the r.h.s. of the identity so that the discrepancy terms become of order zero. We begin with some preliminary computations. Lemma 4.4. Let (X, ω) be a complex Hermitian manifold. For every h ∈ R \ {0}, the following identities hold: (i) [L, τh] = 3 dhω ∧ ·, (ii) [Λ, τh] = 2i τ ⋆ −h, ⋆ (iii) [dh, d −h] = −[dh, τ ⋆ −h], (iv) [dh, d⋆ h] + [dh, τ ⋆ h ] − [d−h, τ ⋆ −h] = [dh + τh, d⋆ h + τ ⋆ h ] + S(h) ω , where S(h) ω := i 2 [Λ, [Λ, d−hdhω ∧ ·]] − [dhω ∧ ·, (dhω ∧ ·)⋆]. Proof. (i) The definition of τh and the Jacobi identity yield the first and respectively the second identities below: [L, τh] = [L, [Λ, dhω ∧ ·]] = −[Λ, [dhω ∧ ·, L]] − [dhω ∧ ·, [L, Λ]]. Now, [dhω ∧ ·, L] = dhω ∧ (ω ∧ ·) − ω ∧ dhω ∧ · = 0, so the first term on the r.h.s. above vanishes. Meanwhile, it is standard that [L, Λ] = (k − n) Id on k-forms. So for any k-form u, we get [dhω∧·, [L, Λ]] u = dhω∧([L, Λ] u)−[L, Λ] (dhω∧u) = (k−n) dhω∧u−(k+3−n) dhω∧u = −3 dhω∧u. Thus, [dhω ∧ ·, [L, Λ]] = −3 dhω ∧ · and (i) follows. (ii) We know from the h-commutation relation (c) of Lemma 4.1 that τh = i [d ⋆ −h, ω ∧ ·] − dh. Hence, using also (b) of Lemma 4.1, we get ⋆ −h, ω ∧ ·]] − [Λ, dh] = i [Λ, [d [Λ, τh] = i [Λ, [d ⋆ −h, ω ∧ ·]] + i (d−h + τ −h)⋆. The Jacobi identity spells [Λ, [d ⋆ −h, ω ∧ ·]] + [d ⋆ −h, [ω ∧ ·, Λ]] + [ω ∧ ·, [Λ, d ⋆ −h]] = 0. ⋆ Since [ω ∧ ·, Λ] = (k − n) Id on k-forms, we get [d −h, [ω ∧ ·, Λ]] = d [[d−h, ω ∧ ·], Λ]⋆, so we get ⋆ −h. Meanwhile, [ω ∧ ·, [Λ, d ⋆ −h]] = ⋆ [Λ, τh] = −i [[d−h, ω ∧ ·], Λ]⋆ − i d −h + i (d−h + τ −h)⋆. Moreover, for an arbitrary form u, we get [d−h, ω ∧ ·] u = (∂ − h ¯∂) (ω ∧ u) − ω ∧ (∂u − h ¯∂u) = (∂ω − h ¯∂ω) ∧ u = d−hω ∧ u. Thus, [d−h, ω ∧ ·] = d−h ω ∧ · and we finally get ⋆ [Λ, τh] = −i [d−h ω ∧ ·, Λ]⋆ − i d −h + i (d−h + τ −h)⋆ = i τ ⋆ ⋆ −h − i d −h + i (d−h + τ −h)⋆ = 2i τ ⋆ −h, 21 where the second identity followed from the definition of τh by replacing h with −h and then taking conjugates and adjoints. This proves (ii). (iii) The Jacobi identity yields −[dh, [Λ, dh]] + [Λ, [dh, dh]] + [dh, [dh, Λ]] = 0. h = 0), and [dh, Λ] = −[Λ, dh], we get [dh, [Λ, dh]] = 0. Using the h- ⋆ −h] = −[dh, d −h], or equivalently that ⋆ ⋆ −h] = −[dh, d −h], where the latter identity was obtained from the former by taking adjoints, Since [dh, dh] = 0 (because d2 commutation relation (b) of Lemma 4.1, this means that [dh, τ ⋆ [τh, d conjugates and replacing h with −h. In other words, we have [dh, d ⋆ −h] = −[τh, d ⋆ −h] = −[dh, τ ⋆ −h]. (22) This proves (iii). (iv) Applying part (ii) and then the Jacobi identity, we get [d−h, τ ⋆ −h] = − i 2 [d−h, [Λ, τh]] = − i 2 [Λ, [τh, d−h]] − i 2 [τh, [d−h, Λ]]. (23) On the other hand, [τh, d−h] (a) = [d−h, τh] (d) = [Λ, d−hdhω ∧ ·] − i [dhω ∧ ·, d⋆ (b) = [d−h, [Λ, dhω ∧ ·]] (c) = [Λ, [dhω ∧ ·, d−h]] + [dhω ∧ ·, [d−h, Λ]] h + τ ⋆ h ], where (a) follows from τh and d−h being operators of odd degrees, (b) follows from the definition of τh, (c) follows from the Jacobi identity, while the latter term in (d) follows from the h-commutation relation (b) of Lemma 4.1 and the former term in (d) follows from the following easy computation: [dhω ∧ ·, d−h] u = dhω ∧ d−hu + d−h(dhω ∧ u) = d−hdhω ∧ u, for any form u. Taking the bracket with Λ in the above formula for [τh, d−h], we get [Λ, [τh, d−h]] = [Λ, [Λ, d−hdhω ∧ ·]] − i [Λ, [dhω ∧ ·, d⋆ h + τ ⋆ h ]]. Applying again the Jacobi formula for the last term, we get [Λ, [dhω ∧ ·, d⋆ h + τ ⋆ h ]] = −[dhω ∧ ·, [d⋆ h + τ ⋆ = −[dhω ∧ ·, [ω ∧ ·, dh + τh]⋆] + [d⋆ h + τ ⋆ = −2 [dhω ∧ ·, (dhω ∧ ·)⋆] + [d⋆ h, Λ]] + [d⋆ h + τ ⋆ h , [Λ, dhω ∧ ·]] h + τ ⋆ h , τh], h , τh], (24) (25) where the first term on the last line is given by the following simple computation. For any form u, we have [ω ∧ ·, dh] u = ω ∧ dhu − dh(ω ∧ u) = −dhω ∧ u. Thus, [ω ∧ ·, dh] = −dhω ∧ ·. Combined with identity (i), this yields [ω ∧ ·, dh + τh] = 2 dhω ∧ ·. 22 Putting (24) and (25) together, we get [Λ, [τh, d−h]] = [Λ, [Λ, d−hdhω ∧ ·]] + 2i [dhω ∧ ·, (dhω ∧ ·)⋆] − i [d⋆ h + τ ⋆ h , τh], which, in turn, combines with (23) to yield [d−h, τ ⋆ −h] = − i 2 [Λ, [Λ, d−hdhω ∧ ·]] + [dhω ∧ ·, (dhω ∧ ·)⋆] − 1 2 [d⋆ h + τ ⋆ h , τh] − i 2 [τh, [d−h, Λ]]. Since −i [Λ, d−h] = d⋆ h + τ ⋆ h by the h-commutation relation (a) of Lemma 4.1, we get −[d−h, τ ⋆ −h] = [d⋆ h + τ ⋆ h , τh] + i 2 [Λ, [Λ, d−hdhω ∧ ·]] − [dhω ∧ ·, (dhω ∧ ·)⋆]. Adding [dh, d⋆ h] + [dh, τ ⋆ h ] on either side of the above identity, we get [dh, d⋆ h] + [dh, τ ⋆ h ] − [d−h, τ ⋆ −h] = [dh + τh, d⋆ h + τ ⋆ h ] + S(h) ω , where S(h) ω := i 2 [Λ, [Λ, d−hdhω ∧ ·]] − [dhω ∧ ·, (dhω ∧ ·)⋆]. This proves (iv). (cid:3) We can now state the main result of this subsection. Theorem 4.5. Let (X, ω) be a complex Hermitian manifold. For every h ∈ R \ {0}, the follow- ing refined h-Bochner-Kodaira-Nakano (h-BKN) identity holds on differential forms of any degree: where T (h) ω is the zero-th order operator defined by ∆h = [d−h + τ −h, d ⋆ −h + τ ⋆ −h] + T (h) ω , T (h) ω := − i 2 [Λ, [Λ, dhd−hω ∧ ·]] − [d−hω ∧ ·, (d−hω ∧ ·)⋆]. In particular, if the metric ω is Kahler, dhω = 0 hence τh = 0 and T (h) ω = 0, so we get ∆h = ∆−h (for every h ∈ R) and ∆h = h2 ∆−h−1 (for every h ∈ R \ {0}). The latter identity follows from the former thanks to (21). Proof. Combining (iv) of Lemma 4.4 with the rough BKN formula of Corollary 4.2, we get ∆h + [dh + τh, d⋆ h + τ ⋆ h ] + S(h) ω = ∆−h + [d−h, τ ⋆ −h] − [dh, τ ⋆ h ] + [dh, d⋆ h] + [dh, τ ⋆ h ] − [d−h, τ ⋆ −h]. Since [dh, d⋆ h] = ∆h, the last formula reduces to ∆−h = [dh + τh, d⋆ h + τ ⋆ h ] + S(h) ω . The refined h-BKN identity follows from this by taking conjugates and replacing h with −h. (cid:3) 23 h-∂ ¯∂-manifolds 4.2 The standard ∂ ¯∂-lemma asserts that every compact Kahler manifold is a ∂ ¯∂-manifold. We will now investigate the analogue of this statement in our dh-cohomology context. Theorem 4.6. Let (X, ω) be a compact Kahler manifold. As usual, we let ∆ := dd⋆ + d⋆d. For every h ∈ R \ {0}, the following identity holds: ∆ = (h + 1)2 (h2 + 1)2 ∆h + (h − 1)2 (h2 + 1)2 (h2 ∆−h−1). Proof. Using (iv) of (20) and the obvious identity ∆h = [dh, d⋆ h] for every h, we get ∆ = [d, d⋆] = + (h + 1)2 (h2 + 1)2 ∆h + (h + 1)h(h − 1) (h2 + 1)2 h2 (h − 1)2 (h2 + 1)2 ∆−h−1 [dh, d⋆ −h−1] + (h + 1)h(h − 1) (h2 + 1)2 [d−h−1, d⋆ h]. Since the metric ω is supposed to be Kahler, [dh, d⋆ The statement follows. −h−1] = 0 and [d−h−1, d⋆ h] = 0 by Corollary 4.3. (cid:3) An immediate consequence of Theorems 4.5 and 4.6 is the following proportionality statement. Corollary 4.7. Let (X, ω) be a compact Kahler manifold. For every h ∈ R \ {0}, the following identities hold on differential forms of any degree: ∆ = 2 h2 + 1 ∆h = 2h2 h2 + 1 ∆−h−1 = 2 h2 + 1 ∆−h. We pause briefly to notice that the above proportionality statement reproves, in conjunction with the main result of [Pop17], the standard fact that the Kahler property of compact complex manifolds implies the degeneration at E1 of the Frolicher spectral sequence. Yet another proof will be implicit further down by putting together Theorems 4.9 and 4.11. Corollary 4.8. (standard) Let (X, ω) be a compact Kahler manifold. The Frolicher spectral se- quence of X degenerates at E1. Proof. We know from Corollary 4.7 that ∆h = h2+1 In particular, ker ∆h = ker ∆ for all h 6= 0. Let δ(k) ∆h : C ∞ eigenvector normalised such that its L2 is also orthogonal on ker ∆ for all h 6= 0. For every h > 0, we get k (X, C) acting on k-forms and let u(k) k (X, C) −→ C ∞ 2 ∆ for every h ∈ R \ {0} in every degree k. h > 0 be the smallest positive eigenvalue of k (X, C) be a corresponding is orthogonal on ker ∆h, it h ∈ C ∞ h equals 1. Since u(k) ω-norm u(k) h δ(k) h = hh∆hu(k) h , u(k) h iiω = h2 + 1 2 hh∆u(k) h , u(k) h iiω ≥ h2 + 1 2 δ(k) ≥ 1 2 δ(k), (26) 24 where δ(k) > 0 is the smallest positive eigenvalue of ∆ : C ∞ (So, δ(k) is independent of h.) k (X, C) −→ C ∞ k (X, C) acting on k-forms. Now, we know from Theorem 1.3 (and its corollary, Proposition 5.3) in [Pop17] that the Frolicher spectral sequence of any compact Hermitian manifold (X, ω) degenerates at E1 if and only if δ(k) does not converge to zero at least as fast as O(h2) when h ↓ 0 for every k. In our case, since the metric ω is Kahler, (26) shows that for every k, δ(k) even remains uniformly bounded below by a positive constant when h ↓ 0. (cid:3) h h We can now infer the dh-cohomology analogue of the standard ∂ ¯∂-lemma. Theorem 4.9. (the h-∂ ¯∂-lemma) Let (X, ω) be a compact Kahler manifold with dimCX = n. For every k ∈ {0, 1, . . . , 2n}, every h ∈ R \ {0} and every k-form u ∈ ker dh ∩ ker d−h−1, the following exactness conditions are equivalent: u ∈ Im dh ⇐⇒ u ∈ Im d−h−1 ⇐⇒ u ∈ Im d ⇐⇒ u ∈ Im (dh d−h−1) = Im (∂ ¯∂). Proof. The equality Im (dh d−h−1) = Im (∂ ¯∂) follows from (iii) of (20), while the property u ∈ Im (dh d−h−1) obviously implies all the other exactness properties. Since d1 = d and we allow any h 6= 0, it suffices to prove the implication "u ∈ Im dh =⇒ u ∈ Im (dh d−h−1)" for an arbitrary h 6= 0. Since ∆h and ∆−h−1 are self-adjoint elliptic operators with d2 is compact, standard Hodge theory yields the following L2 require ω to be Kahler): h = d2 −h−1 = 0 and the manifold X ω-orthogonal decomposition (that does not C ∞ k−1(X, C) = ker ∆−h−1 ⊕ Im d−h−1 ⊕ Im d⋆ −h−1 (27) in which ker d−h−1 = ker ∆−h−1 ⊕ Im d−h−1. Let u ∈ C ∞ k (X, C) such that u ∈ ker dh ∩ ker d−h−1 and u = dhv for some v ∈ C ∞ k−1(X, C). The 3-space decomposition (27) yields a unique decomposition v = v0 + d−h−1u1 + d⋆ −h−1u2, where the (k − 1)-form v0 lies in ker ∆−h−1 and u1, u2 are of respective degrees k − 2 and k. We get u = dhv = dhv0 + dhd−h−1u1 + dhd⋆ −h−1u2 = −d−h−1dhu1 − d⋆ −h−1dhu2. Indeed, the last identity above follows from v0 ∈ ker ∆−h−1 = ker ∆h = ker dh ∩ ker d⋆ h (where the Kahler assumption on ω was used to guarantee the proportionality of the Laplacians ∆−h−1 and ∆h – see Theorem 4.5 – hence the equality of their kernels), from the anti-commutation of dh and d−h−1 (which holds trivially for any, not necessarily Kahler, metric ω – see (iii) of (20)) and from the anti-commutation of dh and d⋆ −h−1 (which is a consequence of the Kahler assumption on ω via the h-commutation relations – see Corollary 4.3). Now, u + d−h−1dhu1 ∈ ker d−h−1 while −d⋆ −h−1, so the form u + d−h−1dhu1 = −d⋆ to Im d⋆ particular, u = −d−h−1dhu1 ∈ Im (dh d−h−1). −h−1dhu2 ∈ Im d⋆ −h−1. However, ker d−h−1 is orthogonal −h−1dhu2, that lies in both subspaces, must vanish. In (cid:3) The above theorem leads naturally to the following 25 Definition 4.10. Let h ∈ R \ {0} be an arbitrary constant. A compact complex manifold X with dimCX = n is said to be an h-∂ ¯∂-manifold if for every k ∈ {0, 1, . . . , 2n} and every k-form u ∈ ker dh ∩ ker d−h−1, the following exactness conditions are equivalent: u ∈ Im dh ⇐⇒ u ∈ Im d−h−1 ⇐⇒ u ∈ Im d ⇐⇒ u ∈ Im (dh d−h−1) = Im (∂ ¯∂). Note that when h = 1, dh = d and d−h−1 = d−1 coincides (up to a multiplicative constant) with dc. The h-∂ ¯∂-property introduced above does not require the form u to be of pure type. In the cases h /∈ {−1, 1}, it is meant to reinforce the standard ∂ ¯∂-property. Like the standard ∂ ¯∂-property, the h-∂ ¯∂-property is implied by the Kahler condition and implies the degeneration at the first page of the Froilicher spectral sequence (cf. Theorems 4.9 above and 4.11 below). Actually, this last implication follows from the well-known implication with the ∂ ¯∂-property in place of the h-∂ ¯∂-property, but we prefer to give a self-contained proof. Theorem 4.11. Let h ∈ R \ {0} be an arbitrary constant. The Frolicher spectral sequence of any h-∂ ¯∂-manifold degenerates at E1. 1 Proof. Let X be an h-∂ ¯∂-manifold with dimCX = n. For any bidegree (p, q), pick any class [α] ¯∂ ∈ Ep, q (X) and any representative α of [α] ¯∂. We have d1([α] ¯∂) = [∂α] ¯∂. Moreover, since ¯∂α = 0, we have ∂α = h∂(h−1 α) + ¯∂(h−1 α) = dh(h−1 α) ∈ Im dh. In particular, ∂α ∈ ker dh and d−h−1(dh(h−1 α)) = ((h2 + 1)/h2) ∂ ¯∂α = 0, so dh(h−1 α) ∈ ker dh ∩ ker d−h−1. Thus, thanks to the h-∂ ¯∂ assumption on X, the dh-exactness of ∂α = dh(h−1 α) implies its ∂ ¯∂-exactness. In particular, ∂α ∈ Im ¯∂, hence d1([α] ¯∂) = [∂α] ¯∂ = 0 ∈ Ep+1, q (X). 1 This proves that all the differentials d1 vanish identically, so Ep, q Furthermore, since ∂α is ∂ ¯∂-exact, there exists a (p, q − 1)-form u such that ∂α = ¯∂∂u, so (X) for (X) and dr([. . . [[α] ¯∂]d1 . . . ]dr−1) = 0 ∈ Ep+r, q−r+1 (X) for all p, q. (X) = Ep, q 1 2 r d2([[α] ¯∂]d1) = [[∂(∂u)] ¯∂]d1 = 0 ∈ Ep+2, q−1 all r ≥ 2. 2 Thus, all the differentials dr with r ≥ 1 vanish identically. Hence, the Frolicher spectral sequence (cid:3) of X degenerates at E1. 4.3 The h-Bott-Chern and h-Aeppli cohomologies We start by defining the h-twisted analogues of the Bott-Chern and Aeppli cohomologies and by observing some basic properties of them. Unlike their standard counterparts, they are not defined in a given bidegree, but in a given total degree. Definition 4.12. Let X be a compact complex n-dimensional manifold. For every h ∈ R \ {0} and every k ∈ {0, . . . , 2n}, we define the kth degree h-Bott-Chern and h-Aeppli cohomology groups by the formulae H k h−BC(X, C) = ker dh ∩ ker d− 1 h Im (dhd− 1 h ) and H k h−A(X, C) = ker(dhd− 1 h ) Im dh + Im d− 1 h , where all the vector spaces involved are subspaces of the space C ∞ k (X, C) of smooth k-forms on X. We now observe some basic properties of these spaces that parallel their standard counterparts. 26 Lemma 4.13. Let X be an n-dimensional compact complex manifold. (a) For every h ∈ R \ {0} and every k ∈ {0, . . . , 2n}, the canonical map T (k) h : H k h−BC(X, C) −→ H k h−A(X, C), [α]h−BC 7→ [α]h−A, is well defined. Moreover, if X is an h-∂ ¯∂-manifold for some fixed h ∈ R \ {0}, the map T (k) isomorphism for every k ∈ {0, . . . , 2n}. h is an (b) For every h ∈ R \ {0} and every k ∈ {0, . . . , 2n}, the following identities hold: H k h−BC(X, C) = Mp+q=k h−A(X, C) = Mp+q=k H k H p, q BC(X, C), H p, q A (X, C). Hence, the dimensions of the vector spaces H k h−BC(X, C) and H k h−A(X, C) are independent of h. (c) For every h ∈ R \ {0} and every k ∈ {0, . . . , 2n}, the canonical maps H k h−BC(X, C) −→ H k dh (X, C) −→ H k h−A(X, C), [α]h−BC 7→ [α]dh 7→ [α]h−A, are well defined. Moreover, if X is an h-∂ ¯∂-manifold for some fixed h ∈ R \ {0}, they are isomor- phisms, in particular their dimensions equal the kth Betti number bk of X, for every k ∈ {0, . . . , 2n}. Proof. (a) Let [α]h−BC ∈ H k h−BC(X, C) be an arbitrary class and let α be an arbitrary representative h−A(X, C). To of it. Then dhα = 0 and d− 1 show that [α]h−A is independent of the choice of representative α of the original class [α]h−BC, we ) ⊂ have to show that [α]h−A = 0 whenever [α]h−BC = 0. However, this is obvious since Im (dhd− 1 Im dh + Im d− 1 α = 0, so α defines a class in H k α = 0, hence dhd− 1 . h h h Suppose now that X is an h-∂ ¯∂-manifold for some fixed h ∈ R \ {0}. Fix any k. To show that T (k) is injective, suppose that dhα = 0, d− 1 α = 0 (i.e. α defines a class [α]h−BC) h h h h ([α]h−BC) = 0). In particular, α = dhu + d− 1 and [α]h−A = 0 (i.e. T (k) Then α − dhu = d− 1 Meanwhile, α − d− 1 Consequently, α = dhu + d− 1 h h v ∈ ker dh ∩ Im d− 1 v = dhu ∈ ker d− 1 h , so d− 1 v ∈ Im (dhd− 1 ∩ Im dh, so dhu ∈ Im (dhd− 1 h h h v ∈ Im (dhd− 1 h h h v for some forms u, v. ) thanks to the h-∂ ¯∂-assumption. ) thanks to the h-∂ ¯∂-assumption. ), so [α]h−BC = 0. k (X, C) such that dhd− 1 v) = 0 and d− 1 h h To show that T (k) h is surjective, let α ∈ C ∞ existence of (k−1)-forms u, v such that dh(α+dhu+d− 1 we will then have [α]h−A = [α+dhu+d− 1 well defined.) These identities are equivalent to dhd− 1 dhα ∈ Im dh and d− 1 they must be (dhd− 1 v]h−A = T (k) h h h h h α ∈ Im d− 1 )-exact thanks to the h-∂ ¯∂-assumption. The surjectivity statement follows. while both forms are simultaneously dh-closed and d− 1 h h h ([α+dhu+d− 1 v]h−BC) with [α+dhu+d− 1 v = −dhα and d− 1 dhu = −d− 1 h h v]h−BC α. Since -closed, h α = 0. We need to prove the v) = 0. (Indeed, (α+dhu+d− 1 h h h (b) For every h ∈ R \ {0} and every k ∈ {0, . . . , n}, the following equalities of subspaces of C ∞ k (X, C) hold: ker dh ∩ ker d− 1 Im dh + Im d− 1 h h = ker ∂ ∩ ker ¯∂ = Im ∂ + Im ¯∂. 27 (28) h ∂α + ¯∂α = 0, whose difference yields (h + 1 Indeed, for any k-form α, the relation α ∈ ker dh ∩ ker d− 1 and − 1 inclusion ker ∂ ∩ ker ¯∂ ⊂ ker dh ∩ ker d− 1 equivalent to the existence of (k − 1)-forms u, v such that α = dhu + d− 1 is equivalent to having h∂α + ¯∂α = 0 h) ∂α = 0, hence ∂α = 0 and ¯∂α = 0. The reverse is h v) + ¯∂(u + v). αp, q (written with its pure-type splitting apparent), the re- is obvious. Meanwhile, the relation α ∈ Im dh + Im d− 1 v = ∂(h u − 1 h h h h Thus, for every k-form α = Pp+q=k quirement α ∈ ker dh ∩ ker d− 1 h = ker ∂ ∩ ker ¯∂ is equivalent to the requirements Xp+q=k ∂αp, q = 0 and Xp+q=k ¯∂αp, q = 0, which, in turn, are equivalent to requiring αp, q ∈ ker ∂ ∩ker ¯∂ = ker dh ∩ker d− 1 for all p, q. Similarly, ) = Im(∂ ¯∂) is equivalent to requiring αp, q ∈ Im(∂ ¯∂) thanks to (iii) of (20), requiring α ∈ Im (dhd− 1 for every p, q. We thus get the first decomposition of vector spaces stated under (b). The second decomposition is obtained in a similar way from the second identity in (28) and from (iii) of (20). h h h h ) and Im (dhd− 1 ) ⊂ Im dh ⊂ (Im dh+Im d− 1 (c) The well-definedness of these maps follows at once from the inclusions ker dh ∩ ker d− 1 ⊂ ker dh ⊂ ker (dhd− 1 ). The bijectivity of these maps when X is supposed to be an h-∂ ¯∂-manifold follows from straightforward applications of this hypothesis, from the proof of (a) and from the following lemma. (cid:3) Lemma 4.14. If X is an h-∂ ¯∂-manifold, every dh-cohomology class [α]dh (of any degree) contains a representative lying in ker dh ∩ ker d− 1 Proof. Let α be a smooth k-form such that dhα = 0. We wish to prove the existence of a smooth (k − 1)-form β such that d− 1 α. However, ), proving the d− 1 existence of β. (cid:3) , so the h-∂ ¯∂-hypothesis ensures that d− 1 (α + dhβ) = 0. This amounts to d− 1 dhβ = −d− 1 α ∈ Im (dhd− 1 α ∈ ker dh ∩ Im d− 1 . h h h h h h h h h h Recall that it was proved by Angella and Tomassini in [AT12] that on every compact complex manifold X, the inequality holds for every k. Thanks to (b) of Lemma 4.13, this translates in our language to 2bk ≤ Xp+q=k hp, q BC + Xp+q=k hp, q A (29) (30) for all k ∈ {0, . . . , 2n} and all h ∈ R \ {0}, 2bk ≤ hk h−BC := dimH k h−A, h−BC + hk h−BC(X, C) and hk h−A := dimH k where hk (Recall that we always have bk = dimH k (X, C) for all k and h 6= 0, see e.g. Introduction.) Moreover, the second main result dh of [AT12] states that equality holds in (29) for every k if and only if X satisfies version (a) of the ∂ ¯∂-lemma (see Introduction). Corollary 4.15. (a)3 Let X be an n-dimensional compact complex manifold. If X is an h-∂ ¯∂- manifold for some h ∈ R \ {0}, then X satisfies version (a) of the ∂ ¯∂-lemma (see Introduction). h−A(X, C). 3This is already obvious from the definitions, but we give a new argument to show the consistency of several results with one another. 28 (b) Let (Xt)t∈∆ be a holomorphic family of compact complex manifolds. If some fibre X0 is an h−BC(Xt, C) h−A(Xt, C) remain constant in a neighbourhood of X0: h-∂ ¯∂-manifold for some h ∈ R \ {0}, then the h-Bott-Chern numbers hk and the h-Aeppli numbers hk h−BC(t) := dimH k h−A(t) := dimH k for all k ∈ {0, . . . , 2n} and all t ∈ ∆ close enough to 0. hk h−BC(t) = hk h−BC(0) and hk h−A(t) = hk h−A(0) Proof. (a) If X is an h-∂ ¯∂-manifold for some h ∈ R \ {0}, it follows from (c) of Lemma 4.13 that h−A for every k. This is equivalent to X satisfying version (a) of the ∂ ¯∂-lemma by 2bk = hk the above discussion and the second main result of [AT12]. h−BC + hk (b) It follows from (b) of Lemma 4.13 that hk for all k and all t. Since the Bott-Chern and the Aeppli numbers are known to satisfy the semicon- tinuity property hp, q A (t) for all t close enough to 0 and all p, q, we infer the analogous property for the h-Bott-Chern and the h-Aeppli numbers: BC(t) and hp, q BC(0) ≥ hp, q A (0) ≥ hp, q h−BC(t) = Pp+q=k hp, q BC(t) and hk h−A(t) = Pp+q=k hp, q A (t) hk h−BC(0) ≥ hk h−BC(t) and hk h−A(0) ≥ hk h−A(t) (31) for all t close enough to 0 and all k. On the other hand, since X0 is an h-∂ ¯∂-manifold for some h ∈ R \ {0}, we have 2bk = hk h−A(0) for every k (cf. (c) of Lemma 4.13). Hence, 2bk ≥ hk hk t close enough to 0 thanks also to (31). However, the reverse inequality 2bk ≤ hk also holds for all t and k thanks to [AT12] (cf. (30)). The result follows. h−BC(t) + hk h−BC(0) + h−A(t) for every k and every h−A(t) (cid:3) h−BC(t) + hk 4.4 Deformation openness of the h-∂ ¯∂-property We now prove the following analogue for h-∂ ¯∂-manifolds of Wu's openness result for ∂ ¯∂-manifolds (cf. [Wu06]). Theorem 4.16. Let π : X −→ ∆ be a proper holomorphic submersion from a complex manifold X to a ball ∆ ⊂ CN containing the origin. For every t ∈ ∆, let Xt := π−1(t) be the fibre above t. Fix an arbitrary constant h ∈ R \ {0}. If X0 is an h-∂ ¯∂-manifold, then Xt is an h-∂ ¯∂-manifold for all t ∈ ∆ sufficiently close to 0. The proof will follow the pattern of the one given by Wu in [Wu06] for the deformation openness of the standard ∂ ¯∂-property. For the sake of consistency, we will follow the presentation in §.4.3 of [Pop14] where Wu's arguments and some of those in [DGMS75] were re-expalined, while pointing out the changes needed in our current h-∂ ¯∂-context. We start with some ad hoc terminology that parallels Definition 4.7. in [Pop14]. Definition 4.17. Let h ∈ R \ {0} be an arbitrary constant. For any given k = 0, 1, . . . , 2n, a given n-dimensional compact complex manifold X is said to satisfy property: (Ak) if the canonical map H k h−BC(X, C) −→ H k h−A(X, C) is injective. This property is equivalent to the property 29 (A′ k) ker dh ∩ ker d− 1 h ∩ (Im dh + Im d− 1 h ) = Im (dh d− 1 h ) as subspaces of C ∞ k (X, C). (Bk) if the canonical map H k h−BC(X, C) −→ H k h−A(X, C) is surjective. This property is equivalent to the property (B′ k) Im dh + Im d− 1 h + (ker dh ∩ ker d− 1 h ) = ker(dh d− 1 h ) as subspaces of C ∞ k (X, C). (Ck) if the canonical maps H k h−BC(X, C) −→ H k d (X, C) and H k h−BC(X, C) −→ H k dh (X, C) are 1 h − injective. This property is equivalent to the simultaneous occurence of (C ′ k)(i) Im d− 1 h ∩ ker dh = Im (dh d− 1 h ) and as subspaces of C ∞ k (X, C). (C ′ k)(ii) Im dh ∩ ker d− 1 h = Im (dh d− 1 h ) (D′ k) if (i) Im dh + ker d− 1 = ker(dh d− 1 h ) and (ii) Im d− 1 h + ker dh = ker(dh d− 1 h ) h as subspaces of C ∞ k (X, C). (Lk) if for every k-form u ∈ ker dh ∩ ker d−h−1, the following exactness conditions are equivalent: u ∈ Im dh ⇐⇒ u ∈ Im d−h−1 ⇐⇒ u ∈ Im (dh d−h−1) = Im (∂ ¯∂). Property (Lk) is a restatement of the pair of properties (C ′ k)(i) and (C ′ k)(ii). As already pointed out, the following equivalences are immediate: (Ak) ⇐⇒ (A′ k), (Bk) ⇐⇒ (B′ k), k), ⊃ in (C ′ k)(i), (ii) and ⊂ in (D′ (Ck) ⇐⇒ (C ′ k). Meanwhile, the inclusions ⊃ in (A′ k)(i), (ii) always hold trivially. The following statement is the h-∂ ¯∂ analogue of a fact implicitly proved in [DGMS75] in the standard ∂ ¯∂ context and will provide a key ingredient for the proof of Theorem 4.16. k), ⊂ in (B′ Proposition 4.18. (the h-∂ ¯∂-analogue of Lemma 5.15 in [DGMS75]) Let h ∈ R\{0} be an arbitrary constant. Let X be a compact n-dimensional complex manifold. For every k = 1, . . . , 2n, the following equivalences hold: (Lk) ⇐⇒ (Ak) ⇐⇒ (Ck) ⇐⇒ (D′ k−1) ⇐⇒ (Bk−1). Proof. Fix an arbitrary k ∈ {1, . . . , 2n}. Given the above explanations, it suffices to prove the equivalences (A′ k) ⇐⇒ (C ′ k) ⇐⇒ (D′ k−1) ⇐⇒ (B′ k−1). Proof of (A′ v. Then u ∈ ker dh ∩ ker d− 1 k) =⇒ (C ′ k). Let u ∈ C ∞ k (X, C) such that dhu = 0 and u = d− 1 ∩ (Im dh + Im d− 1 h h ). So (A′ k) forces u ∈ Im (dh d− 1 h v for some (k − 1)-form ). This proves (i) of h 30 (C ′ k). The proof of (ii) of (C ′ k) is similar with dh and d− 1 h reversed. Proof of (C ′ some (k − 1)-forms v and w. Then k). Let u ∈ C ∞ k) =⇒ (A′ k (X, C) such that dhu = 0, d− 1 u = 0 and u = dhv + d− 1 h w for h · Im dh ∋ dhv = u − d− 1 w ∈ ker d− 1 , so dhv ∈ Im dh ∩ ker d− 1 h h = Im (dhd− 1 h ), the last identity of h subspaces being given by the hypothesis (C ′ k)(i). · Im d− 1 w = u − dhv ∈ ker dh, so d− 1 of subspaces being given by the hypothesis (C ′ ∋ d− 1 h h h k)(ii). w ∈ Im d− 1 h ∩ ker dh = Im (dhd− 1 h ), the last identity We now get u = dhv + d− 1 h w ∈ Im (dhd− 1 h ). This proves (A′ k). Proof of (C ′ k) =⇒ (D′ k−1). Let u ∈ C ∞ k−1(X, C) such that dhd− 1 h u = 0. Then: · d− 1 h u is a k-form and d− 1 being given by the hypothesis (C ′ to u − dhζ ∈ ker d− 1 . h h u ∈ ker dh ∩ Im d− 1 k)(i). So d− 1 u = d− 1 h = Im (dhd− 1 ), the last identity of subspaces dhζ for some (k − 2)-form ζ. This amounts h h h We get u = dhζ + (u − dhζ) ∈ Im dh + ker d− 1 . This proves (D′ k−1)(i). h · dhu is a k-form and dhu ∈ ker d− 1 ∩ Im dh = Im (dhd− 1 h k)(ii). Hence dhu = dhd− 1 h ), the last identity of subspaces being w for some (k − 2)-form w. This amounts to h given by the hypothesis (C ′ u − d− 1 w ∈ ker dh. We get u = d− 1 h w + (u − d− 1 h w) ∈ Im d− 1 h + ker dh. This proves (D′ k−1)(ii). h k−1) =⇒ (C ′ Proof of (D′ v. Then v ∈ ker(dhd− 1 we can find a (k − 2)-form w and a (k − 1)-form ζ such that k). Let u ∈ C ∞ ) = Im dh + ker d− 1 h h k (X, C) such that dhu = 0 and u = d− 1 , where the last identity of subspaces is (D′ h v for some (k − 1)-form k−1)(i). Thus, Applying d− 1 , we get: u = d− 1 v = d− 1 dhw ∈ Im (dhd− 1 h h h Reversing the roles of dh and d− 1 , we get (C ′ h ). This proves (C ′ k)(i). k)(ii) in a similar way from (D′ h k−1)(ii). v = dhw + ζ and d− 1 h ζ = 0. Proof of (D′ can find a (k − 2)-form v and a (k − 1)-form w such that k−1). Let u ∈ C ∞ k−1) =⇒ (B′ k−1(X, C) such that dhd− 1 h u = 0. Thanks to (D′ k−1)(ii), we Thus dhd− 1 h w = 0, so by (D′ k−1)(i) we can write u = d− 1 h v + w and w ∈ ker dh. w = dhζ + ρ with ρ ∈ ker d− 1 h for some (k − 2)-form ζ and some (k − 1)-form ρ. We get ρ = w − dhζ ∈ ker dh (because w ∈ ker dh). Given the choice of ρ, this implies that ρ ∈ ker dh ∩ ker d− 1 . h 31 Putting the bits together, we have u = d− 1 h v + dhζ + ρ ∈ Im d− 1 h + Im dh + (ker dh ∩ ker d− 1 h ). This proves (B′ k−1). Proof of (B′ and Im dh + (ker dh ∩ ker d− 1 k−1) =⇒ (D′ ) ⊂ ker dh. h k−1). This implication is trivial because Im d− 1 + (ker dh ∩ ker d− 1 h ) ⊂ ker d− 1 h h The proof of Proposition 4.18 is complete. (cid:3) Note that the simultaneous occurence of properties (Lk) for all k ∈ {0, . . . , 2n} is an a priori weaker condition than the h-∂ ¯∂-property since it does not include the equivalence u ∈ Im d ⇐⇒ u ∈ Im (dh d−h−1). However, we can easily see as a consequence of Proposition 4.18 that these two conditions are actually equivalent. Corollary 4.19. Let h ∈ R \ {0} be an arbitrary constant. Let X be a compact complex manifold with dimCX = n. Fix an arbitrary k ∈ {0, . . . , 2n} and suppose that for every k-form u ∈ ker dh ∩ ker d−h−1, the following exactness conditions are equivalent: u ∈ Im dh ⇐⇒ u ∈ Im d−h−1 ⇐⇒ u ∈ Im (dh d−h−1) = Im (∂ ¯∂). Then, for every k-form u ∈ ker dh∩ker d−h−1, the equivalence "u ∈ Im d ⇐⇒ u ∈ Im (dh d−h−1) = Im (∂ ¯∂)" also holds. In particular, if the assumption is made for all k ∈ {0, . . . , 2n}, then X is an h-∂ ¯∂-manifold. Proof. Since the implication u ∈ Im (dh d−h−1) =⇒ u ∈ Im d is trivial, we only have to prove the reverse implication for every k-form u ∈ ker dh ∩ ker d−h−1. Let u ∈ Im d be such a k-form. Then u ∈ Im dh + Im d−h−1 thanks to identity (iv) in (20), so u defines a class [u]h−BC ∈ H k h−BC(X, C) that maps to the zero class in H k h−BC(X, C) −→ H k h−A(X, C). Then by (Ak), which holds because it is equivalent to our assumption (Lk) thanks to Proposition 4.18, this map is injective. Hence, [u]h−BC = 0 ∈ H k (cid:3) h−A(X, C) under the canonical map H k h−BC(X, C), so u ∈ Im (dh d−h−1). We are now well equipped to prove the deformation openness of the h-∂ ¯∂-property of compact complex manifolds. Proof of Theorem 4.16. The arguments are analogues in the h-∂ ¯∂ context of those given by Wu in the classical ∂ ¯∂ context. As in [Wu06], the main idea is to exploit, for every fixed k, the equivalence (Lk) ⇐⇒ (Ak) ⇐⇒ (Bk−1), namely the discrepancy of one degree between the characterisation of the h-∂ ¯∂-property for k-forms in terms of the injectivity of the h-BC→ h-A-map and in terms of its surjectivity. This prompts an argument by induction on k, since the h-∂ ¯∂-property holds trivially in degree k = 0 (i.e. for functions). 32 To show that Xt is an h-∂ ¯∂-manifold for all t ∈ ∆ sufficiently close to 0, suppose that the h- (t) induced by the complex ∂ ¯∂-property holds in degree k on Xt for the operators dh(t) and d− 1 structure of Xt for all t close to 0. We will prove that the same is true in degree k + 1. The h-∂ ¯∂ assumption on X0 in degree k implies the following identities thanks to (c) of Lemma h 4.13 and respectively (b) of Corollary 4.15: dimCH k dimCH k dimCH k h−BC(X0, C) = dimCH k h−BC(X0, C) = dimCH k h−A(X0, C) = dimCH k h−A(X0, C), h−BC(Xt, C) h−A(Xt, C) for all t close to 0, for all t close to 0. Hence, dimCH k h−BC(Xt, C) = dimCH k h−A(Xt, C) for all t ∈ ∆ close to 0. h−BC(Xt, C) → H k Meanwhile, by Proposition 4.18, the induction hypothesis (Lk) on Xt is equivalent to the canonical linear map H k h−A(Xt, C) being injective (property (Ak)). Since these are finite- dimensional vector spaces of equal dimensions, the linear map H k h−A(Xt, C) must also be surjective. Thus, property (Bk) holds on Xt for all t ∈ ∆ close to 0. However, thanks to Proposition 4.18, this is equivalent to property (Lk+1), i.e. to the h-∂ ¯∂-property in degree k + 1, holding on Xt for all t ∈ ∆ close to 0. (cid:3) h−BC(Xt, C) → H k References. [AT12] D. Angella, A. Tomassini - On the ∂ ¯∂-Lemma and Bott-Chern Cohomology - Invent. Math. [COUV16] M. Ceballos, A. Otal, L. Ugarte, R. Villacampa - Invariant Complex Structures on 6- Nilmanifolds: Classification, Frolicher Spectral Sequence and Special Hermitian Metrics - J. Geom. Anal. 26 (2016), no. 1, 252–286. [Dem84] J.-P. Demailly - Sur l'identit´e de Bochner-Kodaira-Nakano en g´eom´etrie hermitienne - S´eminaire d'analyse P. Lelong, P. Dolbeault, H. Skoda (editors) 1983/1984, Lecture Notes in Math., no. 1198, Springer Verlag (1986), 88-97. [Dem92] J.-P. Demailly - Regularization of Closed Positive Currents and Intersection Theory - J. Alg. Geom., 1 (1992), 361-409. [DGMS75] P. Deligne, Ph. Griffiths, J. Morgan, D. Sullivan - Real Homotopy Theory of Kahler Manifolds - Invent. Math. 29 (1975), 245-274. [Gau77] P. Gauduchon - Le th´eor`eme de l'excentricit´e nulle - C.R. Acad. Sc. Paris, S´erie A, t. 285 (1977), 387-390. [KS60] K. Kodaira, D.C. Spencer - On Deformations of Complex Analytic Structures, III. Stability Theorems for Complex Structures - Ann. Math. 71, No. 1 (1960), 43-76. [Lam99] A. Lamari - Courants kahl´eriens et surfaces compactes - Ann. Inst. Fourier, Grenoble, 49, 1 (1999), 263-285. [Pop13] D. Popovici - Deformation Limits of Projective Manifolds: Hodge Numbers and Strongly Gauduchon Metrics - Invent. Math. 194 (2013), 515-534. 33 [Pop14] D. Popovici - Deformation Openness and Closedness of Various Classes of Compact Com- plex Manifolds; Examples - Ann. Sc. Norm. Super. Pisa Cl. Sci. (5), Vol. XIII (2014), 255-305. [Pop15] D. Popovici - Aeppli Cohomology Classes Associated with Gauduchon Metrics on Compact Complex Manifolds - Bull. Soc. Math. France 143 (3), (2015), p. 1-37. [Pop16] D. Popovici - Degeneration at E2 of Certain Spectral Sequences - International Journal of Mathematics 27, no. 14 (2016), 1650111, 31 pp. [Pop17] D. Popovici - Adiabatic Limit and the Frolicher Spectral Sequence - arXiv e-print CV 1709.04332v1 [PU18a] D. Popovici, L. Ugarte - Compact Complex Manifolds with Small Gauduchon Cone - Proceedings of the London Mathematical Society (3) (2018) doi:10.1112/plms.12110. [PU18b] D. Popovici, L. Ugarte - Symmetry and Duality for a 5-Dimensional Nilmanifold - in preparation. [Wu06] C.-C. Wu - On the Geometry of Superstrings with Torsion - thesis, Department of Math- ematics, Harvard University, April 2006. Ibn Tofail University, Faculty of Sciences, Departement of Mathematics, PO 242 Kenitra, Morocco Email: [email protected] and Universit´e Paul Sabatier, Institut de Math´ematiques de Toulouse 118, route de Narbonne, 31062, Toulouse Cedex 9, France Email: [email protected] 34
1210.3339	2	1210	2013-06-10T19:08:40	Exceptional collections of line bundles on the Beauville surface	[ "math.AG", "math.KT" ]	We construct quasi-phantom admissible subcategories in the derived category of coherent sheaves on the Beauville surface $S$. These quasi-phantoms subcategories appear as right orthogonals to subcategories generated by exceptional collections of maximal possible length 4 on $S$. We prove that there are exactly 6 exceptional collections consisting of line bundles (up to a twist) and these collections are spires of two helices.	math.AG	math	EXCEPTIONAL COLLECTIONS OF LINE BUNDLES ON THE BEAUVILLE SURFACE SERGEY GALKIN, EVGENY SHINDER Abstract. We construct quasi-phantom admissible subcategories in the derived category of coherent sheaves on the Beauville surface S. These quasi-phantoms subcategories appear as right orthogonals to subcategories generated by exceptional collections of maximal possible length 4 on S. We prove that there are exactly 6 exceptional collections consisting of line bundles (up to a twist) and these collections are spires of two helices. Keywords: exceptional collection, quasi-phantom category, Beauville surface 1. Introduction Bounded derived categories of coherent sheaves on algebraic varieties, their admissible subcategories and semiorthogonal decompositions have been studied intensively by Bondal, Kapranov, Kuznetsov, Orlov, and others [Bon], [BK], [BO], [Kap], [Kuz06], [Kuz09]. It has been questioned which additive invariants of admissible geometric triangulated categories are conservative, that is do not vanish for non-zero categories. Non-vanishing of the Hochschild homology of geometric admissible categories has been conjectured by Kuznetsov in [Kuz09] and non-vanishing of the Grothendieck group has been conjectured by Bondal in early 90's (unpublished). On the contrary, existence of geometric categories with van- ishing Hochschild homology (quasi-phantoms) has been indicated by Katzarkov in [Kat] and existence of geometric categories with vanishing Grothendieck group (phantoms) has been conjectured by Diemer, Katzarkov and Kerr [DKK], both motivated by considerations from mirror symmetry. Let us consider the simplest interesting case, that of a complex smooth projective surface S of general type. On one hand such a surface is not expected to admit a full exceptional collection in its bounded derived category Db(S). On the other hand exceptional collections of maximal possible length dim H ∗(S, Q) seem to exist at least in some cases when pg(S) = q(S) = 0. In such a case the orthogonal complement to the category generated by the exceptional collection has vanishing Hochschild homology [Kuz09], torsion Grothendieck group and generally rather mysterious structure. The first counterexample to Kuznetsov's conjecture was given by Bohning, Graf von Bothmer and Sosna, who constructed exceptional collections of length 11 on the classical Godeaux surface (pg = q = 0, K 2 = 1, b2 = 9) [BBS]. Alexeev and Orlov [AO] came up with exceptional collections of length 6 on Burniat surfaces (pg = q = 0, K 2 = 6, b2 = 4). Some of the fake projective planes (pg = q = 0, K 2 = 9, b2 = 1) are expected to admit exceptional collections of length 3 [GKMS][Section 3]. In this paper we consider yet another surface with similar properties, the Beauville surface S [Bea]. S is a surface of general type with pg = q = 0, K 2 = 8, b2 = 2, constructed as follows. Let C and C ′ be two copies of the Fermat quintic X 5 + Y 5 + Z 5 = 0, acted upon by G = (Z/5)2 in two different ways. We consider the product surface T = C × C ′ with the diagonal G-action. The latter action turns out to be free for an appropriate choice of G-actions on C and C ′. The Beauville surface S is defined as a quotient T /G. According to [BaC], Theorem 3.7 there are two non-isomorphic surfaces that can be obtained this way. We chose one of these two models which we describe in detail in Section 1. One can find useful the analogy between the Beauville surface S and the quadric surface, that is to think of Beauville surface as a sort of a fake quadric. First of all these two surfaces have the same numerical invariants (pg = q = 0, K 2 = 8, b2 = 2). Furthermore, we prove in Section 2.3 that the Picard group of S is generated modulo Sergey Galkin: Laboratory of Algebraic Geometry, National Research University Higher School of Economics, 7 Vavilova Str., Moscow, Russia, 117312; Universitat Wien, Fakultat fur Mathematik; Independent University of Moscow; Moscow Institute of Physics and Technology [email protected] Evgeny Shinder: Corresponding author. Max Planck Institute for Mathematics, Vivatsgasse 7, Bonn 53177, Germany, [email protected]. 1 2 SERGEY GALKIN, EVGENY SHINDER torsion by the bundles O(1, 0), O(0, 1) which come as pull-backs from the factors C and C ′. The Riemann-Roch formula on S implies that χ(O(i, j)) = (i − 1)(j − 1) also in analogy with the quadric on which we have minus signs replaced by the plus signs. A line bundle L ∈ P ic(S) is called acyclic if H ∗(S, L) = 0, for example line bundles O(−1, k) and O(k, −1) are acyclic line bundles on a quadric P1 × P1 for any k. However unlike the quadric case there are only finitely many isomorphism classes of acyclic line bundles on S. We list these line bundles in Section 3.2 (Lemma 3.3) and use them to construct six exceptional collections on S of length 4. We prove that this list exhausts all the exceptional collections consisting of 4 line bundles up to a common twist by a line bundle (Theorem 3.5). We compute dimensions of Ext-groups between elements of the collections in Proposition 3.7. All of our exceptional collections in question are non-strict. Moreover in all of them both Ext1 and Ext2 are present unlike the case of the Burniat surfaces where only Ext2 appears ([AO], Lemma 4.8). We also note that unlike the case of Burniat surfaces the exceptional collections we present have no blocks, that is no groups of pairwise orthogonal elements. Confirming the analogy between the Beauville surface and the quadric, it turns out furthermore that line bundles in the exceptional collections on S are all products of powers of square roots K(1, 0), K(0, 1) of canonical classes coming from the factors C and C ′. We expect the existence of exceptional collections of line bundles to hold for other product-quotient surfaces with pg = q = 0, K 2 = 8 (see e.g. [BaP]) as well. However we do not see at the moment whether there could be a uniform proof for that (see Remark 3.6). We plan to return to this question in the future. We thank Alexander Kuznetsov for reading a draft of the paper and providing us with many useful comments and remarks. We thank Ingrid Bauer, Arend Bayer, Ludmil Katzarkov, Mateusz Michalek, Dmitry Orlov, Yuri Prokhorov, Nicolo Sibilla, Maxim Smirnov for helpful discussions. We thank the referee for pointing out some typos and inaccuracies in the previous version of the paper. The first author is partially supported by NSF Grant DMS0600800, NSF FRG Grant DMS-0652633, FWF Grant P20778, and an ERC Grant (GEMIS). The second author is supported by the Max-Planck-Institut fur Mathematik and the SFB / Transregio 45 "Periods, moduli spaces and arithmetic of algebraic varieties" Bonn - Mainz - Essen. 2. The Beauville surface and its properties 2.1. Generalities on G-equivariant line bundles. We list general facts on G-linearized line bundles and their cohomology (see [Mum] for details). Let G be a finite group acting on a smooth projective variety X/C. The equivariant Picard group P icG(X) is the group of isomorphism classes of G-linearized line bundles on X. The equivariant Picard group is related to the ordinary Picard group via an exact sequence (2.1) 0 → bG → P icG(X) → P ic(X)G, where bG = Hom(G, C∗) is the group of characters and the first arrow associates to a character χ : G → C∗ a trivial line bundle with the G-action induced by χ. Suppose G is abelian; then we can describe the equivariant Picard group in terms of G-invariant divisors on X. Lemma 2.1. Let G be a finite abelian group. Then the image of P icG(X) in P ic(X)G consists of equivalence classes of G-invariant divisors and (2.1) rewrites as (2.2) 0 → bG → P icG(X) → Div(X)G rational equivalence → 0. Proof. We need to prove that for a G-linearized line bundle L there exists a section s with a G-invariant divisor div(s). Let W be an arbitrary finite-dimensional invariant subspace of meromorphic sections of L. Since G is abelian, we may assume W is one-dimensional, W = C · s. Now s is a G-eigensection, which is (cid:3) equivalent to div(s) being G-invariant. If G is a finite group (not necessarily abelian) acting freely on X, then we have an etale covering of smooth projective varieties π : X → X/G. EXCEPTIONAL COLLECTIONS OF LINE BUNDLES ON THE BEAUVILLE SURFACE 3 In this case specifying a line bundle L on X/G is the same as specifying a line bundle eL = π∗L on X together with additional structure of G-linearization. This way we get an identification P icG(X) = P ic(X/G). For any line bundle L on X/G the groups H i(X, π∗L) have a natural structure of G-representations and we have canonical isomorphisms H i(X/G, L) = H i(X, π∗L)G. For our computations we need an equivariant version of the Serre duality. For any G-linearized line bundle on X we have an isomorphism of G-representations: (2.3) H k(X, L) ∼= (H dim(X)−k(X, L∗ ⊗ ωX ))∗. Lemma 2.2. Let V be an n + 1-dimensional representation of a finite group G. Then we have an isomorphism of G-linearized line bundles on P(V ): ωP(V ) ∼= O(−n − 1)(det V ∗). Proof. The claim follows by taking the determinant of the Euler exact sequence of G-linearized line bundles on P(V ) 0 → Ω1 P(V ) → O(−1) ⊗ V ∗ → O → 0. (cid:3) In the notation of Lemma 2.2 let F be an invariant section of O(d) on P(V ) and X be the hypersurface F = 0. Then there is a standard adjunction formula giving an isomorphism of G-linearized line bundles on X: (2.4) ωX ∼= O(d − n − 1)(det V ∗). 2.2. Equivariant Fermat quintics. In what follows G is an abelian group G = (Z/5)2 = Z/5 · e1 ⊕ Z/5 · e2 acting on a three dimensional vector space V with induced action on P2 = P(V ) given by e1 · (X : Y : Z) = (ζ5X : Y : Z) e2 · (X : Y : Z) = (X : ζ5Y : Z), where ζ5 is the 5-th root of unity. Let C be the plane G-invariant Fermat quintic curve X 5 + Y 5 + Z 5 = 0. We consider the scheme-theoretic quotient C/G which is isomorphic to P1 and the quotient map π : C → P1 of degree 25. Explicitly we may pick coordinates on P1 such that π is given by the formula π(X : Y : Z) = (X 5 : Y 5). One easily checks that there are three ramification points on P1 corresponding to the orbits where G acts non-freely: (2.5) D1 = {(0 : −ζ j D2 = {(−ζ j D3 = {(ζ j 5 : 1), j = 0 . . . 4} 5 : 0 : 1), j = 0 . . . 4} 5 : 0), j = 0 . . . 4} 5 : −ζ j Stabilizers of the points in Di, i = 1, 2, 3 are equal to (2.6) respectively. G1 = Z/5 · e1 G2 = Z/5 · e2 G3 = Z/5 · (e1 + e2) 4 SERGEY GALKIN, EVGENY SHINDER Lemma 2.3. The equivariant Picard group P icG(C) splits as a direct sum Proof. The claim follows from the exact sequence (2.2). Indeed any G-invariant divisor is a combination of G-orbits on C. Any orbit is either a smooth fiber of π consisting of 25 points or one of the divisors (2.5) consisting of 5 points. Since D1, D2, D3 are hyperplane sections of C they give rise to the same element O(1) in the Picard group P ic(C). All the generic fibers are of π are linearly equivalent to each other, and also equivalent to O(5). P icG(C) = bG ⊕ Z · O(1). Therefore the third term in the exact sequence (2.2) is Z·O(1) and (2.2) splits giving the required decomposition. (cid:3) We introduce some notation which will help us to keep track of characters appearing in the cohomology repre- sentations. Note that the Grothendieck ring of the category of Z+-graded representations of G is isomorphic to Z[q, x, y]/(x5 − 1, y5 − 1). Thus to any Z+-graded G-representation W we can attach a polynomial (2.7) [W ] ∈ K0(RepZ+(Z/5)2) = Z[q, x, y]/(x5 − 1, y5 − 1). By definition we have the following properties of the polynomial [W ]: [W ⊕ W ′] = [W ] + [W ′] [W ⊗ W ′] = [W ] · [W ′] [W ∗] = [W ](cid:12)(cid:12)(cid:12)x=x4,y=y4 . Later we will use the same bracket notation [i, j], i, j ∈ Z/5 for the character e1 7→ ζ i 5, e2 7→ ζ j 5 which will hopefully not lead to a confusion. For example we have [W [i, j]] = [W ] · xiyj. We now proceed to computing cohomology groups of line bundles O(n), n ≤ 5 on C taking into account the G-action. For n ≤ 4 we have H 0(C, O(n)) ∼= H 0(P2, O(n)) = M i,j≥0, i+j≤n C · X iY jZ n−i−j. For n = 5 we quotient out the representation space H 0(P2, O(5)) by the relation X 5 + Y 5 + Z 5 = 0. Thus we have [H 0(C, O(n))] = X [H 0(C, O(5))] = X i,j≥0, i+j≤n i,j≥0, i+j≤5 xi yj, 0 ≤ n ≤ 4 xi yj − 1. (2.8) In order to compute H 1(C, O(n)) we first use the adjunction (2.4): V ∗ = Γ(P(V ), O(1)) = C · X ⊕ C · Y ⊕ C · Z ∼= [1, 0] ⊕ [0, 1] ⊕ [0, 0] det(V ∗) = [1, 0] ⊗ [0, 1] ⊗ [0, 0] = [1, 1] ωC = O(2)[1, 1], so that by Serre duality (2.3) we have H 1(C, O(n)) ∼= H 0(C, O(2 − n)[1, 1])∗ = H 0(C, O(2 − n))∗[4, 4], which in terms of polynomials implies that [H 1(C, O(n))](x, y) = [H 0(C, O(2 − n))](x4, y4) · x4 y4. A short computation shows that [H 1(C, O)] = q(x4y4 + x4y3 + x3y4 + x4y2 + x3y3 + x2y4) (2.9) [H 1(C, O(1))] = q(x4y4 + x4y3 + x3y4) [H 1(C, O(2))] = qx4y4 [H 1(C, O(n))] = 0, n ≥ 3. EXCEPTIONAL COLLECTIONS OF LINE BUNDLES ON THE BEAUVILLE SURFACE 5 We introduce the curve C ′ which is defined by the same equation X 5 + Y 5 + Z 5 = 0 as C but has a different G-action. We pick the G-action on C ′ to be defined as e1 · (X : Y : Z) = (ζ2 5 X : ζ4 e2 · (X : Y : Z) = (ζ5X : ζ3 5 Y : Z) 5 Y : Z) For this action points in divisors Di, i = 1, 2, 3 defined as in (2.5) have stabilizers (2.10) respectively. G′ G′ G′ 1 = Z/5 · (e1 + 2e2) 2 = Z/5 · (e1 + 3e2) 3 = Z/5 · (e1 + 4e2) It follows from the construction that for any n ∈ Z we have a formula (2.11) [H ∗(C ′, O(n))](q, x, y) = [H ∗(C, O(n))](q, x2y, x4y3) and that the canonical class on C ′ is equal to O(2)[1, 4]. We introduce the notation KC (1) = OC (1)[3, 3] KC ′ (1) = OC ′(1)[3, 2] for the unique square roots of the canonical classes on C and C ′ respectively. 2.3. Line bundles and cohomological invariants of the Beauville surface. We let T = C × C ′ with the diagonal G-action. Since the stabilizers in (2.6) and (2.10) are distinct, the G-action on T is free. One can check that the corresponding smooth quotient Beauville surface S = T /G is of general type with pg = q = 0, K 2 = 8 (Chapter X, Exercise 4 in [Bea]). The Noether formula gives b2 = 2. Since pg = q = 0, the exponential exact sequence gives an identification P ic(S) = H 2(S, Z). Modulo torsion P ic(S) is an indefinite unimodular lattice of rank 2, that is a hyperbolic plane. We introduce G-linearized line bundles O(i, j) and K(i, j) for i, j ∈ Z as follows: O(i, j) = p∗ K(i, j) = p∗ 1(O(i)) ⊗ p∗ 1(K(i)) ⊗ p∗ 2(O(j)) 2(K(j)) = O(i, j)[3i + 3j, 3i + 2j]. We will often prefer to work with the lattice K(i, j) since the exceptional collections we write down in Section 3 are all contained in this lattice. We note however that K(i, j) and O(i, j) differ by a torsion line bundle hence are equivalent from the point of view of intersection pairing. In particular in the following Proposition O(i, j) can be replaced by K(i, j) (with an obvious exception of the second claim). Proposition 2.4. 1. The Picard group of S splits as 2. The canonical class ωS is equal to K(2, 2) = O(2, 2)[2, 0]. 3. The intersection pairing is given by P ic(S)(= P icG(T )) = bG · [O] ⊕ Z · [O(1, 0)] ⊕ Z · [O(0, 1)]. 4. The Euler characteristic of a line bundle L = O(i, j)(χ) is equal to (i − 1)(j − 1). (O(i1, j1)(χ1) · O(i2, j2)(χ2)) = i1j2 + j1i2. Proof. Let us first prove that (2.12) (O(1, 0) · O(0, 1)) = 1. For that we pull-back the intersection to T : 25 · (O(1, 0) · O(0, 1))S = (π∗O(1, 0) · π∗O(0, 1))T = (5[pt × C ′] · 5[C × pt])T = 25, 6 SERGEY GALKIN, EVGENY SHINDER which implies (2.12). Since we also obviously have (2.13) (O(1, 0)2) = (O(0, 1)2) = 0, it follows that O(1, 0) and O(0, 1) span a hyperbolic plane and therefore generate the whole Picard group modulo torsion. To prove the first claim we use the fact that H1(S) = (Z/5)2 [BaC], Theorem 4.3, (4), which implies that P ic(S)tors = H 2(S, Z)tors = H1(S, Z)tors = (Z/5)2. Since by (2.2) bG ∼= (Z/5)2 is contained in P ic(S), P ic(S)tors ∼= bG and we get a decomposition P ic(S) = bG · [O] ⊕ P ic(S)/tors = bG · [O] ⊕ Z · [O(1, 0)] ⊕ Z · [O(0, 1)]. The second claim follows from ωS = p∗ 1ωC ⊗ p∗ 2ωC ′ = K(2, 0) ⊗ K(0, 2) = O(2, 0)[1, 1] ⊗ O(0, 2)[1, 4]. The third claim of the Lemma follows from (2.12),(2.13), and the fact that twisting by torsion classes does not affect the intersection form. To check the fourth claim we use Riemann-Roch formula: χ(L) = 1 + (L · L ⊗ ω∗ S) 2 = 1 + (O(i, j)(χ) · O(i − 2, j − 2)(χ − [2, 0])) 2 (i(j − 2) + j(i − 2) = 1 + 2 = (i − 1)(j − 1). We have a Kunneth-type formula for isomorphism classes of graded representations (recall the notation from (2.7)): (2.14) [H ∗(T, K(i, j))](q, x, y) = [H ∗(C, K(i))](q, x, y) · [H ∗(C ′, K(j))](q, x, y), and the analogous formula with K(i, j) replaced by O(i, j). This is simply a reformulation of the Kunneth formula (cid:3) H ∗(C × C ′, p∗ 1L1 ⊗ p∗ 2L2) = H ∗(C, L1) ⊗ H ∗(C ′, L2). with the G-action on both sides taken into account. In the following Lemma we perform necessary computations which will be used later for computing Hochschild homology of S as well as cohomology of dg-algebras of the exceptional collections on S. Lemma 2.5. Some cohomology ranks h0(K(i, j)) + qh1(K(i, j)) + q2h2(K(i, j)) are given in the table: i j 4 3 2 1 0 -1 -2 -3 -2 -1 3q2 + 3q 0 3q2 6q2 9q2 3q2 + q 4q2 6q2 8q2 0 0 1 1 0 2 3 3 + 3q 3 + q q2 0 0 0 0 5 8 3 6 4 3 + q 4 9 6 3 0 3 + 3q 3q2 + q 3q2 + 3q 3q2 0 Proof. The entries of the table are in agreement with the Serre isomorphism hp(S, K(i, j)) = h2−p(S, K(2 − i, 2 − j)), therefore it is sufficient to consider i, j from the table with i, j ≥ 1. The Euler characteristic of K(i, j) is equal to (i − 1)(j − 1). By Kodaira vanishing theorem there is no higher cohomology for i, j ≥ 3. The rest is done using the EXCEPTIONAL COLLECTIONS OF LINE BUNDLES ON THE BEAUVILLE SURFACE 7 Kunneth formula (2.14) and (2.8), (2.9), (2.11) which we use to compute: (2.15) [H ∗(C, K(1))] = x4y3 + x3y4 + x3y3 + qx2y2 + qx2y + qxy2 [H ∗(C ′, K(1))] = x3y2 + y3 + x2 + qx3 + qy2 + qx2y3 [H ∗(C, K(2))] = x3y + x2y2 + xy3 + x2y + xy2 + xy + q [H ∗(C ′, K(2))] = x2y3 + xy4 + x4 + x3 + y2 + y + q [H ∗(C, K(3))] = x4y4 + x4y2 + x2y4 + x4y + xy4 + x4 + y4 + x + y + 1 [H ∗(C ′, K(3))] = x4y3 + x3y4 + x3y3 + x4y + x2y2 + y4 + x2y + xy2 + x + 1 [H ∗(C, K(4))] = x4y4 + x4y3 + x3y4 + x4y2 + x3y3 + x2y4 + x3y2 + x2y3+ + x2y2 + x3 + x2y + xy2 + y3 + x2 + y2 [H ∗(C ′, K(4))] = x4y4 + x4y2 + x2y4 + x4y + x3y2 + x2y3 + x3y + xy3+ + y4 + x3 + y3 + x2 + xy + y2 + x. Lemma 2.6. The Hochschild cohomology groups HH ∗(S, C) = ⊕p+q=∗H p(S, ΛqTS) of S are given below. HH 0(S) = C HH 1(S) = 0 HH 2(S) = 0 HH 3(S) = H 2(S, TS) = C6 HH 4(S) = H 2(S, Λ2TS) = C9. H p(S, ΛqTS) = H p(T, ΛqTT )G TT = p∗ Λ2TT = p∗ 1TC ⊕ p∗ 1TC ⊗ p∗ 2TC ′ = K(−2, 0) ⊕ K(0, −2) 2TC ′ = K(−2, −2). Proof. We have and Now the cohomology groups in question are found in the table of Lemma 2.5. (cid:3) (cid:3) Next we would like to compute the Grothendieck group K0(S) of the Beauville surface S. By the results of Kimura [Kim], Bloch conjecture is known for all surfaces with pg = 0 which admit a covering by a product of curves, hence CH0(S) = Z for the Beauville surface S. Therefore by Lemma 2.7 below the Grothendieck group of S has a decomposition K0(S) = Z4 ⊕ (Z/5)2. Lemma 2.7. Let X be a smooth projective surface such that the degree morphism CH0(X) → Z is an isomorphism. Then we have a (non-canonical) isomorphism K0(X) ∼= Z2 ⊕ P ic(X). Proof. Consider the topological filtration F i ⊂ K0(X) given by the codimension of support [Ful]. By Riemann-Roch theorem without denominators [Ful] we have F 0/F 1 ∼= Z F 1/F 2 ∼= P ic(X) F 2 ∼= CH0(X) ∼= Z. Extension 0 → F 1 → F 0 → Z → 0 always splits for group-theoretic reasons, so We have a short exact sequence (2.16) 0 → Z i→ F 1 → P ic(X) → 0. K0(S) = Z ⊕ F 1. 8 SERGEY GALKIN, EVGENY SHINDER We have to prove that (2.16) splits, that is there exists a retraction F 1 → Z. In general such a retraction exists whenever the image of i(1) in F 1/tors is not divisible by any integer a > 1. Recall that i(1) = [OP ] ∈ F 1 ⊂ K0(S) where P is a point of S. Assume that [OP ] = a · A + α, where α is a torsion element. Then since a is positive integer and χ(O, A) is integer last equality implies a = 1. 1 = χ(O, OP ) = χ(O, aA + α) = aχ(O, A) (cid:3) 3.1. Numerically exceptional collections and helices. We call a sequence of line bundles 3. Exceptional collections on the Beauville surface on a variety numerically exceptional if for all j > i χ(Lj, Li) = X l L1, . . . , Ln (−1)l dim Extl(Lj, Li) = 0. Any exceptional collection is obviously numerically exceptional as well. We note that in order to speak about numerically exceptional collections we only need to consider classes of Li's modulo torsion. This implies that a sequence L1, . . . , Ln forms a numerically exceptional collection on S if and only if any twist L1(χ1), . . . , Ln(χn) does. In particular we will not make a distinction between O(i, j) and K(i, j) when investigating numerically exceptional collections. Lemma 3.1. A sequence O, L1, L2, L3 of line bundles on S is numerically exceptional if and only if it belongs to one of the following four numerical types: (Ic) O, O(−1, 0), O(c − 1, −1), O(c − 2, −1), c ∈ Z (IIc) O, O(0, −1), O(−1, c − 1), O(−1, c − 2), c ∈ Z (IIIc) O, O(−1, c), O(−1, c − 1), O(−2, −1), c ∈ Z (IVc) O, O(c, −1), O(c − 1, −1), O(−1, −2), c ∈ Z. We note that I0 = III0, II0 = IV0 and also that types Ic and IIc are characterized by the property L3 ∼= L1 ⊗L2. Proof. By Proposition 2.4(4) the sequence O, O(a1, b1), O(a2, b2), O(a3, b3) is numerically exceptional if and only if all of the vectors (ai, bi), (aj − ai, bj − bi), j > i have one of the coordinates equal to -1. The rest of the proof is left to the reader. (cid:3) If we consider a general sequence of line bundles (3.1) L0, L1, L2, L3 on S, then it is (numerically) exceptional if and only if (3.2) O, L1 ⊗ L∗ 0, L2 ⊗ L∗ 0, L3 ⊗ L∗ 0 is (numerically) exceptional. We say that the sequence (3.1) is of type Ic, IIc, IIIc or IVc if (3.2) is of this type. In order to study exceptional collections on S more efficiently we will use so-called helices ([GR], [Bon], [BP]). We call a sequence E q = (Ei, i ∈ Z) of sheaves on a smooth variety X a helix of period n if Ei−kn = Ei ⊗ ω⊗k X for all 0 ≤ i ≤ n − 1, k ∈ Z. 1 Given a sequence E0, . . . , En−1 of sheaves on X we can extend it to a helix by the formula above. Any subsequence of a helix consisting of n consecutive elements Ea, Ea+1, . . . , Ea+n−1 will be called a spire. By Serre duality an arbitrary spire of a helix is a (numerically) exceptional collection if and only if E0, . . . , En−1 is a 1The definition of helix we use coincides with that from [Bon] up to shifts which we have dropped for convenience. The definition of helix in [BP] which is given in terms of mutations rather than the Serre functor differs from ours since the collections we consider are not full. EXCEPTIONAL COLLECTIONS OF LINE BUNDLES ON THE BEAUVILLE SURFACE 9 (numerically) exceptional collection. We will sometimes represent a helix E q as a sequence of n + 1 consecutive spires Ea → Ea+1 → · · · → Ea+n, for some a ∈ Z where Ej = {Ej, Ej+1, . . . , Ej+n−1}. Note that since n is the period of E q, Ea+n differs from Ea by a twist by ωX. We now may ask what are the helices formed by numerically exceptional collections of Lemma 3.1. The proof of the following Lemma is straightforward from definitions. Lemma 3.2. Numerically exceptional helices on S formed by line bundles belong to one of the two families: Ic → IVc → I−c → IV−c → Ic, c ∈ Z IIc → IIIc → II−c → III−c → IIc, c ∈ Z. 3.2. Acyclic line bundles and exceptional collections. We will now investigate which of the numerically exceptional collections of Lemma 3.1 can be lifted to exceptional collections. Here by a lift we mean a lift with respect to the morphism that is a choice of a character χ ∈ bG. We will need a detailed study of the characters that may appear in the cohomology groups of sheaves on T . Z2 ⊕ bG = P ic(S) → P ic(S)/tors = Z2, For a G-linearized line bundle on T we define the acyclic set of L as A(L) := {χ ∈ Hom(G, C∗) : χ /∈ [H ∗(T, L)]} By definition L(χ) is acyclic if and only if −χ ∈ A(L). Since by Proposition 2.4(1), any line bundle on S is isomorphic to some K(i, j)(χ), we see from the next lemma that there are 39 isomorphism classes of acyclic line bundles on S. Lemma 3.3. The only nonempty acyclic sets of line bundles K(i, j) on S are: A(K(1, −2)) = {[0, 0]} A(K(1, −1)) = {[0, 3], [2, 0], [3, 2]} A(K(1, 0)) = {[0, 0], [0, 1], [0, 2], [1, 4], [2, 3], [3, 0], [4, 0]} A(K(1, 1)) = {[0, 0], [1, 2], [2, 1], [2, 2], [3, 3], [3, 4], [4, 3]} A(K(1, 2)) = {[0, 0], [0, 3], [0, 4], [1, 0], [2, 0], [3, 2], [4, 1]} A(K(1, 3)) = {[0, 2], [2, 3], [3, 0]} A(K(1, 4)) = {[0, 0]} A(K(−1, 1)) = {[0, 0]} A(K(0, 1)) = {[0, 0], [3, 3], [3, 4], [4, 3]} A(K(2, 1)) = {[0, 0], [1, 2], [2, 1], [2, 2]} A(K(3, 1)) = {[0, 0]}. Proof. Since by Proposition 2.4(4) any bundle K(i, j)(χ) with i 6= 1 and j 6= 1 is not acyclic we restrict to the cases i = 1 or j = 1. We note in addition that our claim is consistent with the Serre duality: A(K(i, j)) is in duality with A(K(2 − i, 2 − j)); therefore we only need to consider the cases K(1, j), K(i, 1), i, j ≥ 1. For i, j ≥ 3 we have an implication A(K(i, j)) = ∅ =⇒ A(K(i + 1, j)) = ∅, A(K(i, j + 1)) = ∅, therefore it is sufficient to prove that (3.3) and to compute A(L) for line bundles A(K(1, 5)) = ∅ A(K(4, 1)) = ∅ This is done by looking at the terms of the products of the polynomials in (2.15). (cid:3) K(1, 1), K(1, 2), K(1, 3), K(1, 4), K(2, 1), K(3, 1). 10 SERGEY GALKIN, EVGENY SHINDER Lemma 3.4. Let L1,L2,L3 be line bundles on S. A sequence O, L1(χ1), L2(χ2), L3(χ3) forms an exceptional collection if and only if the following conditions hold: χ1 ∈ A(L∗ 1) χ2 ∈ A(L∗ 2) χ3 ∈ A(L∗ 3) χ2 − χ1 ∈ A(L1 ⊗ L∗ 2) χ3 − χ1 ∈ A(L1 ⊗ L∗ 3) χ3 − χ2 ∈ A(L2 ⊗ L∗ 3). Proof. The statement is a reformulation of the definition of exceptional collection. (cid:3) Theorem 3.5. The following list contains all exceptional collections of length 4 consisting of line bundles on S (up to a common twist by a line bundle): (3.4) (I1) O, K(−1, 0), K(0, −1), K(−1, −1) (IV1) O, K(1, −1), K(0, −1), K(−1, −2) (I−1) O, K(−1, 0), K(−2, −1), K(−3, −1) (IV−1) O, K(−1, −1), K(−2, −1), K(−1, −2) (II0 = IV0) O, K(0, −1), K(−1, −1), K(−1, −2) (I0) O, K(−1, 0), K(−1, −1), K(−2, −1). These six collections are spires of the two helices (3.5) (H1) I1 → IV1 → I−1 → IV−1 → I1 (H2) I0 → II0 → I0. Proof. Because of Remark 3.2 we only need to consider numerically exceptional collections of types Ic, c ≥ 0, IIc, c > 0 and all helices formed by them. Let us start by listing all numerically exceptional collections of line bundles of the types as above satisfying the properties: O, L1, L2, L3 = L1 ⊗ L2 A(L∗ 3) 6= ∅ 1) 6= ∅; A(L∗ 2) 6= ∅; A(L∗ A(L1 ⊗ L∗ 2) 6= ∅. By Lemma 3.3 these properties are necessary for O, L1, L2, L3 to form an exceptional collection. With the help of Lemmas 3.1 and 3.3 we get the following list: I0, I1, II1, II2. Finally we check whether there are characters χ1, χ2, χ3 for each of these types of collections that will satisfy the conditions of Lemma 3.4. Type I0: O, K(−1, 0)(χ1), K(−1, −1)(χ2), K(−2, −1)(χ3) with conditions χ1, χ3 − χ2 ∈ A(K(1, 0)) = {[0, 0], [0, 1], [0, 2], [1, 4], [2, 3], [3, 0], [4, 0]} χ2, χ3 − χ1 ∈ A(K(1, 1)) = {[0, 0], [1, 2], [2, 1], [2, 2], [3, 3], [3, 4], [4, 3]} χ3 ∈ A(K(2, 1)) = {[0, 0], [1, 2], [2, 1], [2, 2]} χ2 − χ1 ∈ A(K(0, 1)) = {[0, 0], [3, 3], [3, 4], [4, 3]} For each choice of χ3 we find possible χ1, χ2 from conditions (3.6) χ1 ∈ A(K(1, 0)) ∩ χ3 − A(K(1, 1)) χ2 ∈ A(K(1, 1)) ∩ χ3 − A(K(1, 0)) and look for those χ1, χ2 that satisfy (3.7) χ2 − χ1 ∈ A(K(0, 1)). EXCEPTIONAL COLLECTIONS OF LINE BUNDLES ON THE BEAUVILLE SURFACE 11 1. χ3 = [0, 0]. Using (3.6) we find the only set of characters χ1 = χ2 = [0, 0] and it obviously satisfies the condition (3.7) as well. Thus we obtain the collection (I0) O, K(−1, 0), K(−1, −1), K(−2, −1) and the one in the same helix 2. χ3 = [1, 2]. (3.6) reads as: (II0 = IV0) O, K(0, −1), K(−1, −1), K(−1, −2). and none of these pairs satisfies (3.7). 3. χ3 = [2, 1]. (3.6) reads as: and none of these pairs satisfies (3.7). 4. χ3 = [2, 2] (3.6) reads as: χ1 ∈ {[0, 0], [4, 0]} χ2 ∈ {[1, 2], [2, 2]} χ1 ∈ {[0, 0], [1, 4]} χ2 ∈ {[1, 2], [2, 1]} χ1 ∈ {[0, 1], [0, 0]} χ2 ∈ {[2, 1], [2, 2]} and none of these pairs satisfies (3.7). Type I1: O, K(−1, 0)(χ1), K(0, −1)(χ2), K(−1, −1)(χ3) with conditions χ1, χ3 − χ2 ∈ A(K(1, 0)) = {[0, 0], [0, 1], [0, 2], [1, 4], [2, 3], [3, 0], [4, 0]} χ2, χ3 − χ1 ∈ A(K(0, 1)) = {[0, 0], [3, 3], [3, 4], [4, 3]} χ3 ∈ A(K(1, 1)) = {[0, 0], [1, 2], [2, 1], [2, 2], [3, 3], [3, 4], [4, 3]} χ2 − χ1 ∈ A(K(−1, 1)) = {[0, 0]} From the conditions on χ1, χ2 we find that χ1 = χ2 = [0, 0]. Then χ3 ∈ A(K(1, 1)) ∩ A(K(1, 0)) ∩ A(K(0, 1)) = {[0, 0]}. This way we get exceptional collection (I1) O, K(−1, 0), K(0, −1), K(−1, −1) and three others lying in the same helix (IV1) O, K(1, −1), K(0, −1), K(−1, −2) (I−1) O, K(−1, 0), K(−2, −1), K(−3, −1) (IV−1) O, K(−1, −1), K(−2, −1), K(−1, −2). Type II1: O, K(0, −1)(χ1), K(−1, 0)(χ2), K(−1, −1)(χ3) with conditions χ1, χ3 − χ2 ∈ A(K(0, 1)) = {[0, 0], [3, 3], [3, 4], [4, 3]} χ2, χ3 − χ1 ∈ A(K(1, 0)) = {[0, 0], [0, 1], [0, 2], [1, 4], [2, 3], [3, 0], [4, 0]} χ3 ∈ A(K(1, 1)) = {[0, 0], [1, 2], [2, 1], [2, 2], [3, 3], [3, 4], [4, 3]} χ2 − χ1 ∈ A(K(1, −1)) = {[0, 3], [2, 0], [3, 2]} There exist no χ1, χ2 satisfying the respective conditions. Type II2: O, O(0, −1)(χ1), O(−1, 1)(χ2), O(−1, 0)(χ3) with conditions χ1, χ3 − χ2 ∈ A(K(0, 1)) = {[0, 0], [3, 3], [3, 4], [4, 3]} χ2, χ3 − χ1 ∈ A(K(1, −1)) = {[0, 3], [2, 0], [3, 2]} χ3 ∈ A(K(1, 0)) = {[0, 0], [0, 1], [0, 2], [1, 4], [2, 3], [3, 0], [4, 0]} χ2 − χ1 ∈ A(K(1, −2)) = {[0, 0]} 12 SERGEY GALKIN, EVGENY SHINDER There exist no χ1, χ2 satisfying the respective conditions. Remark 3.6. All six exceptional collections in (3.4) span the same torsion-free subgroup in P ic(S) with generators (cid:3) K(1, 0) = O(1, 0)[3, 3], K(0, 1) = O(0, 1)[3, 2]. We do not have a conceptual proof for this statement. For a helix E q of period n we introduce a matrix M(E q) with entries consisting of the Ext-groups in the spires of E q: Mi,j = X l dim Extl(Ei, Ei+j ) · ql; 0 ≤ i, j ≤ n − 1. Proposition 3.7. For the helices (3.5) we have: M(H1) = M(H2) =     1 1 1 1 1 1 1 1   3q2 + q 3q2 + 3q 3q2 + q 4q2 3q2 + q 3q2 + q 3q2 + q 3q2 + q 3q2 + q 3q2 + q 6q2 6q2 4q2 4q2 4q2 4q2 6q2 6q2 6q2 6q2 4q2 6q2 8q2 6q2   In particular we see that all our collections have endomorphism dg-algebras with non-vanishing first and second cohomology groups. Proof. The entries are found in the table given in Lemma 2.5. (cid:3) Proposition 3.8. The A∞-algebra of the exceptional collection (I−1) O, K(−1, 0), K(−2, −1), K(−3, −1) is formal and moreover the usual product m2 is trivial. Proof. The Ext-groups of the collection E q = I−1 are all found in M(H1) from the previous Proposition. In fact we have: (cid:16)X l dim Extl(Ei, Ej) · ql(cid:17)i,j =   1 0 0 0 3q2 + q 1 0 0 6q2 4q2 1 0 8q2 6q2 3q2 + q 1   . In order to prove formality we check that the higher A∞-operations mk, k ≥ 3 of the collection E q vanish. Using a standard argument (see [Sei08] Lemma 2.1 or [Lef02] Th 3.2.1.1), we may assume that ml(. . . , idEi, . . . ) = 0 for all objects Ei and all l > 2. Now the third A∞-operation m3 vanishes for grading reasons and the products mk, k ≥ 4 also vanish since our graded quiver has only 4 vertices. The product of the two non-trivial elements of degree 1 vanishes since these elements are not composable. All other products are trivial for grading reasons. (cid:3) Let (E q) be one of the collections in (3.4). By [BK], Theorem 2.10 the subcategory hE0, E1, E2, E3i generated by coh(S) = the collection is admissible and has a right orthogonal A, i.e. there is a semiorthogonal decomposition Db hE0, E1, E2, E3, Ai. Proposition 3.9. Right orthogonals to two spires of a helix are equivalent categories. Proof. By transitivity it is enough to prove the statement for two consecutive spires. Denote E4 = E0 ⊗ ω−1 S . Let A be the right orthogonal to hE0, E1, E2, E3i, and A′ be the right orthogonal to hE1, E2, E3, E4i. We want to show that categories A and A′ are equivalent. Denote by C the right orthogonal to hE1, E2, E3i. Second decomposition Db(S) = hE1, E2, E3, E4, A′i implies C = hE4, A′i. First decomposition Db(S) = hE0, E1, E2, E3, Ai is equivalent to Db(S) = hE1, E2, E3, A, E4i by Serre duality, so C = hA, E4i. Hence both A and A′ are subcategories in C EXCEPTIONAL COLLECTIONS OF LINE BUNDLES ON THE BEAUVILLE SURFACE 13 orthogonal to E4: A is the left orthogonal and A′ is the right orthogonal. So (left/right) mutations in E4 establish the equivalence between A and A′. (cid:3) We denote two equivalence classes of subcategories obtained by taking right orthogonals to H1 and H2 by A1 and A2 respectively. We note that a choice of a spire gives rise to a fully faithful embedding Ai → Db(S). Proposition 3.10. We have In particular we see that Ai's are non-trivial. Proof. We have K0(Ai) = (Z/5)2 HH∗(Ai) = 0 HH 0(Ai) = C. Z4 ⊕ (Z/5)2 = K0(S) = K0(Db(S)) = K0(hE0, E1, E2, E4i) ⊕ K0(Ai) = Z4 ⊕ K0(Ai), thus K0(Ai) ∼= (Z/5)2. For the homology we use the additivity theorem [Ke] (see also[Kuz09], Corollary 7.5): C4 = H ∗(S) = HH∗(Db(S)) = HH∗(hE0, E1, E2, E4i) ⊕ HH∗(Ai) = C4 ⊕ HH∗(Ai). The statement about Hochschild cohomology is proved by the following approach of Kuznetsov [Kuz12]. Define e(F, F ′) = min{p Extp(F, F ′) 6= 0} For any increasing sequence a0 < a1 < · · · < ak = a0 + n (n is the period of the helix E q, in our case n = 4) define Finally the anticanonical height of the exceptional collection is defined as δa q(E q) = e(Ea0 , Ea1) + · · · + e(Eak−1, Eak ) + 1 − k. We now use the following result: h(E q) = min a q δa q (E q) Proposition 3.11. [Kuz12] Let A be right orthogonal to exceptional collection E q. For k ≤ h(E q) + (dim S − 2) the natural map HH k(S) → HH k(A) is isomorphism. For our helices we have and hence we see that HH 0(Ai) = HH 0(S) = C. h(H1) = 2 h(H2) = 1 References (cid:3) [AO] V. Alexeev, D. Orlov, Derived categories of Burniat surfaces and exceptional collections, arXiv:1208.4348v2 [BaC] Bauer, Ingrid C., Catanese, Fabrizio, Some new surfaces with pg = q = 0, The Fano Conference, 123 -- 142, Univ. Torino, Turin, 2004. [BaP] Bauer, I., Pignatelli, R., The classification of minimal product-quotient surfaces with pg = 0, Math. Comp. 81 (2012), no. 280, 2389 -- 2418. [Bea] A.Beauville, Complex Algebraic Surfaces, Second edition. London Mathematical Society Student Texts, 34. Cambridge University Press, Cambridge, 1996 [BBS] Christian Bohning, Hans-Christian Graf von Bothmer, Pawel Sosna, On the derived category of the classical Godeaux surface, arXiv:1206.1830v1 [Bon] Alexey Bondal, Representations of associative algebras and coherent sheaves, Izv. Akad. Nauk SSSR Ser. Mat., 53:1 (1989), [BK] Alexey Bondal, Mikhail Kapranov, Representable functors, Serre functors, and mutations, Izv. Akad. Nauk SSSR Ser. Mat., 53:6 (1989), [BO] Alexey Bondal, Dmitri Orlov, Reconstruction of a variety from the derived category and groups of autoequivalences, arXiv:alg-geom/9712029 [BP] Alexey Bondal, Alexander Polishchuk, Homological properties of associative algebras: the method of helices, Russian Academy of Sciences. Izvestiya Mathematics, 1994, 42:2, 219-260. [DKK] Colin Diemer, Ludmil Katzarkov, Gabriel Kerr, Compactifications of spaces of Landau-Ginzburg models, arXiv:1207.0042, To appear in Shafarevich volume 14 SERGEY GALKIN, EVGENY SHINDER [Ful] Fulton W., Intersection theory. Second edition. Ergebnisse der Mathematik und ihrer Grenzgebiete. 3. Folge. A Series of Modern Surveys in Mathematics [Results in Mathematics and Related Areas. 3rd Series. A Series of Modern Surveys in Mathematics], 2. Springer-Verlag, Berlin, 1998. [GKMS] S.Galkin, L.Katzarkov, A.Mellit, E.Shinder, Minifolds and Phantoms, arXiv:1305.4549, preprint IPMU 13-0102 [GR] Gorodentsev, A. L.; Rudakov, A. N., Exceptional vector bundles on projective spaces, Duke Math. J. 54 (1987), no. 1, 115 -- 130. [Kap] Michail Kapranov, On the derived categories of coherent sheaves on some homogeneous spaces, Invent. Math. 92 (1988), no. 3, 479 -- 508. [Kat] L. Katzarkov, Homological Mirror Symmetry and Algebraic Cycles, Riemannian topology and geometric structures on manifolds, 63 -- 92, Progr. Math., 271, Birkhauser Boston, Boston, MA, 2009. [Ke] B. Keller, Invariance and localization for cyclic homology of DG algebras, J. Pure Appl. Algebra 123 (1998), no. 1 -- 3, 223 -- 273. [Kim] Kimura, Shun-Ichi, Chow groups are finite dimensional, in some sense. Math. Ann. 331 (2005), no. 1, 173 -- 201 [Kuz06] Alexander Kuznetsov: Hyperplane sections and derived categories, arXiv:math.AG/0503700, Izvestiya RAN: Ser. Mat. 70:3 (2006) p. 23 -- 128 (in Russian); translation in Izvestiya: Mathematics 70:3 (2006) p. 447 -- 547. [Kuz09] A. Kuznetsov, Hochschild homology and semiorthogonal decompositions, arXiv:0904.4330v1 [Kuz12] A. Kuznetsov, Height of exceptional collections and Hochschild cohomology of quasiphantoms categories, arXiv:1211.4693 [Lef] K. Lefevre, Sur les A∞-categories, Ph.D. thesis, Universite Paris 7, (2002). [Mum] Mumford, D.; Fogarty, J.; Kirwan, F., Geometric Invariant Theory, Third edition. Ergebnisse der Mathematik und ihrer Grenzgebiete (2) [Results in Mathematics and Related Areas (2)], 34. Springer-Verlag, Berlin, 1994 [Sei] P. Seidel, Fukaya categories and Picard-Lefschetz theory, Zurich Lectures in Advanced Mathematics, European Mathematical Society (EMS), Zurich, 2008.
1011.5484	2	1011	2013-06-20T06:59:38	The Hilbert Stack	[ "math.AG" ]	Let \pi : X -> S be a morphism of algebraic stacks that is locally of finite presentation with affine stabilizers. We prove that there is an algebraic S-stack, the Hilbert stack, parameterizing proper algebraic stacks mapping quasi-finitely to X. This was previously unknown, even for a morphism of schemes.	math.AG	math	THE HILBERT STACK JACK HALL AND DAVID RYDH ABSTRACT. Let π : X → S be a morphism of algebraic stacks that is locally of fi- nite presentation with affine stabilizers. We prove that there is an algebraic S-stack-the Hilbert stack-parameterizing proper algebraic stacks mapping quasi-finitely to X. This was previously unknown, even for a morphism of schemes. INTRODUCTION Let π : X → S be a morphism of algebraic stacks. Define the Hilbert stack, HSX/S, to be the S-stack that sends an S-scheme T to the groupoid of quasi-finite and representable πT−−→ T is proper, morphisms (Z s−→ X ×S T ), such that the composition Z s−→ X ×S T flat, and of finite presentation. Let HSmono X/S ⊂ HSX/S be the S-substack whose objects are those (Z s−→ X ×S T ) such X/S is never an algebraic stack. that s is a monomorphism. The main results of this paper are as follows. Theorem 1. Let π : X → S be a non-separated morphism of noetherian algebraic stacks. Then HSmono Theorem 2. Let π : X → S be a morphism of algebraic stacks that is locally of finite presentation, with quasi-compact and separated diagonal, and affine stabilizers. Then HSX/S is an algebraic stack, locally of finite presentation over S, with quasi-affine diago- nal over S. Theorem 3. Let X → S be a morphism of algebraic stacks that is locally of finite presen- tation, with quasi-finite and separated diagonal. Let Z → S be a morphism of algebraic stacks that is proper, flat, and of finite presentation with finite diagonal. Then the S-stack T 7→ HOMT (Z ×S T, X ×S T ) is algebraic, locally of finite presentation over S, with quasi-affine diagonal over S. Let f : Y → Z and p : Z → W be morphisms of stacks. Define the fibered category p∗Y , the restriction of scalars of Y along p, by (p∗Y )(T ) = Y (T ×W Z). Theorem 4. Let f : Y → Z and p : Z → W be morphisms of algebraic stacks. Assume that p is proper, flat, and of finite presentation with finite diagonal and that f is locally of finite presentation with quasi-finite and separated diagonal. Then the restriction of scalars p∗Y is an algebraic stack, locally of finite presentation over W , with quasi-affine diagonal over W . Theorem 1 is similar to the main conclusion of [LS08], and is included for complete- ness. In the case that the morphism π : X → S is separated, the Hilbert stack, HSX/S, is equivalent to the stack of properly supported algebras on X, which was shown to be algebraic in [Lie06]. Thus the new content of this paper is in the removal of separatedness assumptions from similar theorems in the existing literature. The statement of Theorem 2 Date: 2013-06-20. 2010 Mathematics Subject Classification. Primary 14C05; Secondary 14A20, 14D15, 14D23. Key words and phrases. Hilbert stack, non-separated, pushouts, Generalized Stein factorizations. We would like to sincerely thank Brian Conrad, Jacob Lurie, Martin Olsson, Jason Starr, and Ravi Vakil for their comments and suggestions. We would also like to express our gratitude to the referee for their careful reading and excellent suggestions. 1 2 J. HALL AND D. RYDH for algebraic spaces appeared in [Art74, Appendix 1], but was left unproved due to a lack of foundational results. It is important to note that Theorems 2, 3, and 4 are completely new, even for schemes and algebraic spaces. We wish to point out that if X is an algebraic S-stack with affine stabilizer groups, and X 0 ⊂ X denotes the open locus where the inertia stack IX/S is quasi-finite, then there is an isomorphism of S-stacks HSX0/S → HSX/S. In particular, Theorem 2 is really about algebraic stacks with quasi-finite diagonals. Theorems 3 and 4 generalize [Ols06, Thm. 1.1 & 1.5] and [Aok06a, Aok06b] to the non-separated setting, and follow easily from Theorem 2. In the case where π is not flat, as was remarked in [Hal12b], Artin's Criterion is difficult to apply. Thus, to prove Theorem 2 we use the algebraicity criterion [op. cit., Thm. A]. The results of [op. cit., §9] show that it is sufficient to understand how infinitesimal deformations can be extended to global deformations (i.e. the effectivity of formal deformations). The difficulty in extending infinitesimal deformations of the Hilbert stack lies in the dearth of "formal GAGA" type results-in the spirit of [EGA, III.5]-for non-separated schemes, algebraic spaces, and algebraic stacks. In this paper, we will prove a generaliza- tion of formal GAGA to non-separated morphisms of algebraic stacks. The proof of our version of non-separated formal GAGA requires the development of a number of founda- tional results on non-separated spaces, and forms the bulk of the paper. 0.1. Background. The most fundamental moduli problem in algebraic geometry is the Hilbert moduli problem for PN Z : find a scheme that parameterizes flat families of closed Z . It was proven by Grothendieck [FGA, IV.3.1] that this moduli problem subschemes of PN has a solution which is a disjoint union of projective schemes. In general, given a morphism of schemes X → S, one may consider the Hilbert moduli problem: find a scheme HilbX/S parameterizing flat families of closed subschemes of X. It is more precisely described by its functor of points: for any scheme T , a map of schemes T → HilbX/S is equivalent to a diagram: Z X ×S T ● ● ● ● ● ● ● ● ● #● T, Z /Z. where the morphism Z → X ×S T is a closed immersion, and the composition Z → T is proper, flat, and of finite presentation. Grothendieck, using projective methods, constructed the scheme HilbPN In [Art69], M. Artin developed a new approach to constructing moduli spaces. It was proved, by M. Artin in [Art69, Cor. 6.2] and [Art74, Appendix], that the functor HilbX/S had the structure of an algebraic space for any separated and locally finitely presented morphism of algebraic spaces X → S. The algebraic space HilbX/S is not, in general, a scheme-even if X → S is a proper morphism of smooth complex varieties. In more recent work, Olsson–Starr [OS03] and Olsson [Ols05] showed that the functor HilbX/S is an algebraic space in the case of a separated and locally finitely presented morphism of algebraic stacks X → S. A separatedness assumption on a scheme is rarely restrictive to an algebraic geometer. Indeed, most schemes algebraic geometers are interested in are quasi-projective or proper. Let us examine some spaces that arise in the theory of moduli. Example 0.1 (Picard Schemes). Let C → A1 be the family of curves corresponding to a conic degenerating to a node. Then the Picard scheme PicC/A1, which parameterizes families of line bundles on C/A1 modulo pullbacks from the base, is not separated. This is worked out in detail in [FGI+05, Ex. 9.4.14]. / / # THE HILBERT STACK 3 Example 0.2 (Curves). Let U be the stack of all curves. That is, a morphism T → U from a scheme T , is equivalent to a morphism of algebraic spaces C → T that is proper, flat, finitely presented, and with one-dimensional fibers. In particular, U parameterizes all singular curves, which could be non-reduced and have many irreducible and connected components. In [Smy13, Appendix B], it was shown that U is an algebraic stack, locally of finite presentation over Z. The stack U is interesting, as Hassett [Has03], Schubert [Sch91], and Smyth [Smy13] have constructed modular compactifications of Mg, different from the classical Deligne–Mumford compactification [DM69], that are open substacks of U. The algebraic stack U is not separated. Unlike schemes, the norm for interesting moduli spaces is that they are non-separated. Indeed, families of interesting geometric objects tend not to have unique limits. This is precisely the reason why compactifying moduli spaces is an active, and very difficult, area of research. Lundkvist and Skjelnes showed in [LS08] that for a non-separated morphism of noe- therian algebraic spaces X → S, the functor HilbX/S is never an algebraic space. We will provide an illustrative example of this phenomenon. Example 0.3. Consider the simplest non-separated scheme: origin. Let k be a field and set S = Spec k. Let X = A1 s ∐A1 as an S-scheme. Now, for a y-line, A1 A1 the line with the doubled t , which we view s=t−(0) A1 y, we have a D := Spec k[[x]]-morphism Ty : D → y ×S D given by 1 ⊗ x 7→ x, y ⊗ 1 7→ x. Thus, we have an induced D-morphism: Ts−→ (A1 i : D s) ×S D → X ×S D. Now, the fiber in of i over Dn := Spec k[[x]]/(xn+1) is topologically the inclusion of one of the two origins, which is a closed immersion. Note, however, that the map i is not a closed immersion. The closed immersions in : Dn → X ×S Dn induce compatible S-morphisms Dn → HilbX/S. If HilbX/S is an algebraic space, then this data induces a unique S-morphism D → HilbX/S. That is, there exists a closed immersion j : Z → X ×S D whose fiber over Dn is in. One immediately deduces that j is isomorphic to i. But i is not a closed immersion, thus we have a contradiction. Note that if X → S is separated, any monomorphism Z → X ×S T , such that Z → T is proper, is automatically a closed immersion. Thus, for a separated morphism X → S, the stack HSmono X/S is equivalent to the Hilbert functor HilbX/S. In the case that the morphism X → S is non-separated, they are different. Note that in Example 0.3, the deformed object was still a monomorphism, so will not prove Theorem 1 for the line with the doubled- origin. Let us consider another example. Example 0.4. Consider the line with doubled-origin again, and retain the notation and conventions of Example 0.3. Thus, we have an induced map over D: D ∐ D Ts∐Tt−−−−→ (A1 s ∐ A1 t ) ×S D → X ×S D. Where x = 0, this becomes the inclusion of the doubled point, which is a closed immer- sion. Where x 6= 0, this becomes non-monomorphic. The proof of [LS08, Thm. 2.6] is based upon Example 0.3. Arguing similarly, but with the last example, one readily obtains Theorem 1. Thus, for a non-separated morphism of schemes, the obstruction to the existence of a Hilbert scheme is that a monomorphism Z ֒→ X can deform to a non-monomorphism. So, one is forced to parameterize non- monomorphic maps Z → X. Such variants of the Hilbert moduli problem have been considered previously in the literature. We now list those variants that the authors are aware of at the time of publication and that are known to be algebraic. 4 J. HALL AND D. RYDH • Vistoli's Hilbert stack, which parameterizes families of finite and unramified mor- phisms to a separated stack [Vis91]. • The stack of coherent algebras on a separated algebraic stack [Lie06]. • The stack of branchvarieties, which parameterizes geometrically reduced alge- braic stacks mapping finitely to a separated algebraic stack. It has proper compo- nents when the target stack has projective coarse moduli space or admits a proper flat cover by a quasi-projective scheme [AK10, Lie06]. • There is a proper algebraic space parameterizing Cohen–Macaulay curves with fixed Hilbert polynomial mapping finitely, and birationally onto its image, to pro- jective space [Høn04]. • The Hilbert stack of points for any morphism of algebraic stacks [Ryd11]. To subsume the variants of the Hilbert moduli problem listed above, we parameterize the quasi-finite morphisms Z → X. Example 0.5. Closed immersions, quasi-compact open immersions, quasi-compact un- ramified morphisms, and finite morphisms are all examples of quasi-finite morphisms. By Zariski's Main Theorem [EGA, IV.18.12.13], any quasi-finite and separated map of schemes Z → X factors as Z → Z → X where Z → Z is an open immersion and Z → X is finite. We now define the Generalized Hilbert moduli problem: for a morphism of algebraic stacks π : X → S, find an algebraic stack HSX/S such that a map T → HSX/S is equiva- lent to the data of a quasi-finite and representable map Z → X ×S T , with the composition Z → T proper, flat, and of finite presentation. This is the fibered category that appears in Theorem 2. The main result in this paper, Theorem 2, is that this stack is algebraic. Example 0.6. Let X → S be a separated morphism of algebraic stacks. Then every quasi- finite and separated morphism Z → X, such that Z → S proper, is finite. Hence, HSX/S is the stack of properly supported coherent algebras on X. Lieblich [Lie06] showed that HSX/S is algebraic whenever X → S is locally of finite presentation and separated. 0.2. Outline. In §5, we will prove Theorem 2 using the algebraicity criterion [Hal12b, Thm. A]. To apply this criterion, like Artin's Criterion [Art74, Thm. 5.3], it is necessary to know that formally versal deformations of objects in the Hilbert stack can be effectivized. Note that effectivity results for moduli problems related to separated objects usually follow from the formal GAGA results of [EGA, III.5], and the relevant generalizations to algebraic stacks [OS03, Ols05]. Since we are concerned with non-separated objects, no previously published effectivity result applies. In §4, we prove a generalization of formal GAGA for non-separated algebraic stacks. This is the main technical result of this paper. To be precise, let R be a noetherian ring that is separated and complete for the topology defined by an ideal I ⊆ R (i.e., R is I-adic [EGA, 0I.7.1.9]). Set S = Spec R and for each n ≥ 0 let Sn = Spec(R/I n+1). Now let π : X → S be a morphism of algebraic stacks that is locally of finite type, with quasi- compact and separated diagonal, and affine stabilizers. For each n ≥ 0 let πn : Xn → Sn denote the pullback of the map π along the closed immersion Sn ֒→ S. Suppose that for each n ≥ 0, we have compatible quasi-finite Sn-morphisms sn : Zn → Xn such that the composition πn ◦ sn : Zn → Sn is proper. We show (Theorem 4.3) that there exists a unique, quasi-finite S-morphism s : Z → X, such that the composition π ◦ s : Z → S is proper, and that there are compatible Xn-isomorphisms Z ×X Xn → Zn. In §§1-3, we will develop techniques to prove the afforementioned effectivity result in §4. To motivate these techniques, it is instructive to explain part of Grothendieck's proof of formal GAGA for properly supported coherent sheaves [EGA, III.5.1.4]. So, let R be an I-adic noetherian ring, let S = Spec R, and for each n ≥ 0 let Sn = Spec(R/I n+1). For a morphism of schemes f : Y → S that is locally of finite type and separated, let fn : Yn → Sn denote the pullback of the morphism f along the closed THE HILBERT STACK 5 immersion Sn ֒→ S. Suppose that for each n ≥ 0, we have a coherent Yn-sheaf Fn, prop- ∼= Fn. Then Grothendieck's erly supported over Sn, together with isomorphisms Fn+1Yn formal GAGA [EGA, III.5.1.4] states that there is a unique coherent Y -sheaf F, with sup- port proper over S, such that FYn ∼= Fn. For various reasons, it is better to think of adic systems of coherent sheaves {Fn} as a coherent sheaf F on the formal scheme bY [EGA, I.10.11.3]. We say that a coherent sheaf F on the formal scheme bY is effectivizable, if there exists a coherent sheaf F on the scheme Y and an isomorphism of coherent sheaves bF ∼= F on the formal scheme bY . The effectivity problem is thus recast as: any coherent sheaf F on the formal scheme bY , with support proper over bS, is effectivizable. This is proven using the method of d´evissage on the abelian category of coherent sheaves with proper support, Cohp/S(Y ), on the scheme Y . The proof consists of the following sequence of observations. (1) Given coherent sheaves H and H′ on Y , with H′ properly supported over S, the natural map of R-modules: Exti OY (H, H′) → Exti O bY (bH, bH′) is an isomorphism for all i ≥ 0. In particular, the "i = 0" statement shows that the functor Cohp/S(Y ) → Cohp/S(bY ) is fully faithful. (2) By (1), it remains to prove that the functor Cohp/S(Y ) → Cohp/S(bY ) is es- sentially surjective. It is sufficient to prove this essential surjectivity when Y is quasi-compact. By noetherian induction on Y , we may, in addition, assume that if F ∈ Cohp/S(bY ) is annihilated by some coherent ideal J ⊆ OY with supp(OY /J) ( Y , then F is effectivizable. (3) If we have an exact sequence of bS-properly supported coherent sheaves on bY : 0 / F′ / F / F′′ / 0 and two of F′, F, F′′ are effectivizable, then the third is. This follows from the exactness of completion and the i = 0,1 statements of (1). (4) Combine (2) and (3) to deduce that if α : F → F′ is a morphism in Cohp/S(bY ) such that F′ is effectivizable and ker α and coker α are annihilated by some coher- ent ideal J ⊆ OY with supp(OY /J) ( Y , then F is effectivizable. (5) The result is true for quasi-projective morphisms Y → S. This is proved via a direct argument. (6) The Chow Lemma [EGA, II.5.6.1] gives a quasi-projective S-scheme Y ′ and a projective S-morphism p : Y ′ → Y that is an isomorphism over a dense open subset U of Y . (7) Use (5) for the quasi-projective morphism Y ′ → S to show that bp∗F ∼= bG for some (8) Use the Theorem on Formal Functions [EGA, III.4.1.5] to show that (p∗G)∧ ∼= S-properly supported coherent sheaf G on Y ′. bp∗bp∗F. Thus bp∗bp∗F is effectivizable with bS-proper support. (9) Note that we have an adjunction morphism η : F → bp∗bp∗F. Pick a coherent ideal J ⊆ OY defining the complement of U . By [EGA, III.5.3.4], we may choose J so that it annihilates the kernel and cokernel of η. By (4), F is effectivizable, and the result follows. The proof of the non-separated effectivity result in §4 will be very similar to the technique outlined above, once the steps are appropriately reinterpreted. For a non-separated mor- phism of schemes X → S, instead of the abelian category Coh (X) (resp. Cohp/S(X)), we consider the non-abelian category QFs(X) (resp. QFp/S(X)) that consists of quasi- finite and separated morphisms Z → X (resp. quasi-finite and separated morphism Z → X such that Z → S is proper). Instead of extensions, which do not make sense in a non-abelian category, we will use finite limits. / / / / 6 J. HALL AND D. RYDH In §1, we will show that QFs(X) is closed under finite limits along finite morphisms. In §1 we also reinterpret (1) and (3) in terms of the preservation of these finite limits under completion. The analogue of quasi-projective morphisms Y → S in (5) and (7) are morphisms that factor as Y → Y ′ → S, where Y → Y ′ is ´etale, and Y ′ → S is projective-note that it is essential that we do not assume that Y → Y ′ is separated. The analogue of the Chow Lemma used in (6), is a generalization, for non-separated schemes and algebraic spaces, due to Raynaud–Gruson [RG71, Cor. 5.7.13]. In §3, for a proper morphism q : Y ′ → Y , we construct an adjoint pair (q!, q∗) : QFs(Y ′) ⇆ QFs(Y ) which takes the role of the adjoint pair (q∗, q∗) : Coh (Y ) ⇆ Coh (Y ′). We also show that the adjoint pair can be constructed for locally noetherian formal schemes, and prove an analogue of (8) in this setting. In §4, we will combine the results of §§1 and 3 to obtain analogues of (4) and [EGA, III.5.3.4] in order to complete the analogue of step (9). In §2, we introduce the generalized Stein factorization: every separated morphism of finite type with proper fibers factors canonically as a proper Stein morphism followed by a quasi-finite morphism. The adjoint pair of §3 is an immediate consequence of the existence of the generalized Stein factorization. 0.3. Notation. We introduce some notation here that will be used throughout the paper. For a category C and X ∈ Obj C, we have the slice category C/X, with objects the morphisms V → X in C, and morphisms commuting diagrams over X, which are called X-morphisms. If the category C has finite limits and f : Y → X is a morphism in C, then for (V → X) ∈ Obj(C/X), define VY := V ×X Y . Given a morphism p : V ′ → V in C/X there is an induced morphism pY : V ′ Y → VY in C/Y . There is an induced functor f ∗ : C/X → C/Y given by (V → X) 7→ (VY → Y ). Given a ringed space U := (U , OU ), a sheaf of ideals I ⊆ OU , and a morphism of ringed spaces g : V → U , we define the pulled back ideal IV = im(g∗I → OV ) ⊆ OV . Let S be a scheme. An algebraic S-space is a sheaf F on the big ´etale site (Sch/S)´Et, such that the diagonal morphism ∆F : F → F ×S F is represented by schemes, and there is a smooth surjection U → F from an S-scheme U . An algebraic S-stack is a stack H on (Sch/S)´Et, such that the diagonal morphism ∆H : H → H ×S H is represented by algebraic S-spaces and there is a smooth surjection U → H from an algebraic S- space U . A priori, we make no separation assumptions on our algebraic stacks. We do show, however, that all algebraic stacks figuring in this paper possess quasi-compact and separated diagonals. Thus all of the results of [LMB] apply. We denote the (2, 1)-category of algebraic stacks by AlgStk. 1. FINITE COLIMITS OF QUASI-FINITE MORPHISMS In this section we will prove that the colimit of a finite diagram of finite morphisms be- tween quasi-finite objects over an algebraic stack exists (Theorem 1.10). We will also prove that these colimits are compatible with completions (Theorem 1.18 and Corollary 1.19). Let X be an algebraic stack. Denote the 2-category of algebraic stacks over X by AlgStk/X. Define the full 2-subcategory RSch(X) ⊂ AlgStk/X to have those objects Y s−→ X, where the morphism s is schematic. We now make three simple observations. (1) The 1-morphisms in RSch(X) are schematic. (2) Automorphisms of 1-morphisms in RSch(X) are trivial. Thus RSch(X) is nat- urally 2-equivalent to a 1-category. (3) If the algebraic stack X is a scheme, then the natural functor Sch/X → RSch(X) is an equivalence of categories. Definition 1.1. Let X be an algebraic stack. Define the following full subcategories of RSch(X): (1) RAff (X) has objects the affine morphisms to X; (2) QFs(X) has objects the quasi-finite and separated maps to X. THE HILBERT STACK 7 We also let QFfin QFs(X) but only has the morphisms f : (Z s−→ X) → (Z ′ s′ finite. This is not a full subcategory. s (X) ⊆ QFs(X) denote the subcategory that has the same objects as −→ X) such that Z → Z ′ is Recall that quasi-finite, separated, and representable morphisms of algebraic stacks are schematic [LMB, Thm. A.2]. The subcategories QFs(X), RAff (X) and RSch(X) have all finite limits and these coincide with the limits in AlgStk/X. The subcategory QFfin s (X) has fiber products and these coincide with the fiber products in AlgStk/X. However, there is not a final object in QFfin s (X), except when every quasi-finite morphism to X is finite. 1.1. Algebraic Stacks. For the moment, we will be primarily concerned with the exis- tence of colimits in the category QFs(X), where X is a locally noetherian algebraic stack. Colimits in algebraic geometry are usually subtle, so we restrict our attention to finite col- imits in QFfin s (X). This includes the two types that will be useful in §4-pushouts of two finite morphisms and coequalizers of two finite morphisms. We will, however, pay attention to some more general types of colimits in the 2-category of algebraic stacks, as the added flexibility will simplify the exposition. Remark 1.2. Note that because we have a fully faithful embedding of categories QFs(X) ⊂ RSch(X), a categorical colimit in RSch(X), that lies in QFs(X), is automatically a cat- egorical colimit in QFs(X). This will happen frequently in this section. Moreover, if a finite diagram {Zi}i∈I in QFfin s (X) has a categorical colimit Z in QFs(X) and Zi → Z is finite for every i ∈ I, then Z is also a categorical colimit in QFfin s (X). Indeed, it is readily seen that `i∈I Zi → Z is surjective so any morphism (Z → W ) ∈ QFs(X), such that Zi → Z → W is finite for every i ∈ I, is finite. Remark 1.3. It is important to observe that a categorical colimit in AlgStk/X is, in gen- eral, different from a categorical colimit in RSch(X). Indeed, let X = Spec k, and let G be a finite group, then the X-stack BG is the colimit in AlgStk/X of the diagram [GX ⇒ X]. The colimit of this diagram in the category RSch(X) is just X. The definitions that follow are closely related to those given in [Ryd07, Def. 2.2]. Recall that a map of topological spaces g : U → V is submersive if a subset Z ⊆ V is open if and only if g−1(Z) is an open subset of U . Definition 1.4. Consider a diagram of algebraic stacks {Zi}i∈I, an algebraic stack Z, and suppose that we have compatible maps φi : Zi → Z for all i ∈ I. Then we say that the data (Z, {φi}i∈I ) is a (1) Zariski colimit if the induced map on topological spaces φ : lim−→i Zi → Z is a homeomorphism (equivalently, the map ∐i∈I φi : ∐i∈I Zi → Z is submersive and the map φ is a bijection of sets); (2) weak geometric colimit if it is a Zariski colimit of the diagram {Zi}i∈I, and the (φi)∗OZi is an isomor- canonical map of lisse-´etale sheaves of rings OZ → lim←−i phism; (3) universal Zariski colimit if for any algebraic Z-stack Y , the data (Y, {(φi)Y }i∈I) is a Zariski colimit of the diagram {(Zi)Y }i∈I; (4) geometric colimit if it is a universal Zariski and weak geometric colimit of the diagram {Zi}i∈I; (5) uniform geometric colimit if for any flat, algebraic Z-stack Y , the data (Y, {(φi)Y }i∈I ) is a geometric colimit of the diagram {(Zi)Y }i∈I. The exactness of flat pullback of sheaves and flat base change easily proves Lemma 1.5. Suppose that we have a finite diagram of algebraic stacks {Zi}i∈I, then a geometric colimit (Z, {φi}i∈I ) is a uniform geometric colimit. 8 J. HALL AND D. RYDH Also, note that weak geometric colimits in the category of algebraic stacks are not In the setting of schemes, unique and thus are not categorical colimits (Remark 1.3). however, we have the following Lemma. Lemma 1.6. Let {Zi}i∈I be a diagram of schemes. Then every weak geometric colimit (Z, {φi}i∈I ) is a colimit in the category of locally ringed spaces and, consequently, a colimit in the category of schemes. Proof. Fix a ringed space W = (W , OW ), together with compatible morphisms of ringed spaces ψi : Zi → W . We will produce a unique map of ringed spaces ψ : Z → W satis- fying ψφi = ψi for all i ∈ I. Since (Z, {φi}i∈I ) is a Zariski colimit, it follows that there exists a unique, continuous, map of topological spaces ψ : Z → W such that ψφi = ψi. (φi)∗OZi is an isomorphism. On the lisse-´etale site, we also know that the map OZ → lim←−i The natural functor from the lisse-´etale site of Z to the Zariski site of Z also preserves limits. In particular, we deduce that as sheaves of rings on the topological space Z, the (φi)∗OZi is an isomorphism. The functor ψ∗ preserves limits, natural map OZ → lim←−i thus the maps OW → (ψi)∗OZi induce a unique map to ψ∗OZ. Hence, we obtain a unique morphism of ringed spaces ψ : Z → W that is compatible with the data. To complete the proof it remains to show that if, in addition, W = (W , OW ) is a locally ringed space and the morphisms ψi are morphisms of locally ringed spaces, then ψ is a morphism of locally ringed spaces. Let z ∈ Z. We must prove that the induced morphism OW,ψ(z) → OZ,z is local. Since ∐i∈I Zi → Z is surjective, there exists an i ∈ I and zi ∈ Zi such that φi(zi) = z. By hypothesis, the morphism OW,ψ(z) → OZi,zi is local. Also, Zi → Z is a morphism of schemes so OZ,z → OZi,zi is also local. We immediately deduce that the morphism OW,ψ(z) → OZ,z is local, and the result is proven. (cid:3) The following criterion will be useful for verifying when a colimit is a universal Zariski colimit. Lemma 1.7. Consider a diagram of algebraic stacks {Zi}i∈I, an algebraic stack Z, and suppose that we have compatible maps φi : Zi → Z for all i ∈ I. If the map ∐iφi : ∐i Zi → Z is surjective and universally submersive and also (1) for any geometric point Spec K → Z, the map φK : lim−→i (2) there is an algebraic stack X such that Z, Zi ∈ QFs(X) for all i ∈ I, the maps φi (Zi)K → Spec K is an injection of sets; or are X-maps, and for any geometric point Spec L → X, the map φL : lim−→i ZL is an injection of sets, (Zi)L → then (Z, {φi}i∈I ) is a universal Zariski colimit of the diagram {Zi}i∈I. Proof. To show (1), we observe that the universal submersiveness hypothesis on the map ∐iφi : ∐i Zi → Z reduces the statement to showing that for any morphism of algebraic stacks Y → Z, the map φY : lim−→i (Zi)Y → Y is a bijection of sets, which will follow if the map φK : lim−→i (Zi)K → Spec K is bijective for any geometric point Spec K → Y . By assumption, we know that φK is injective. For the surjectivity, we note that we have a commutative diagram of sets: `i (Zi)K αK / ◆ ◆ ◆ ◆ ◆ ◆ ∐iφi ◆ ◆ ◆ lim−→i (Zi)K , φK ◆ '◆ Spec K where αK and ∐iφi are surjective, thus φK is also surjective. For (2), given a geometric point Spec L → X, then Z(Spec L) = ZL and Zi(Spec L) = (Zi)L, thus we may apply the criterion of (1) to obtain the claim. (cid:3) / ' THE HILBERT STACK 9 To obtain similarly useful results for stacks, we will need some relative notions, unlike the previous definitions which were all absolute. Definition 1.8. Fix an algebraic stack X and a diagram {Zi}i∈I in RSch(X). Suppose that Z ∈ RSch(X), and that we have compatible X-morphisms φi : Zi → Z for all i ∈ I. We say that the data (Z, {φi}i∈I ) is a uniform categorical colimit in RSch(X) if for any flat, algebraic X-stack Y , the data (ZY , {(φi)Y }i∈I) is the categorical colimit of the diagram {(Zi)Y }i∈I in RSch(Y ). Combining Lemmas 1.5 and 1.6 we obtain Proposition 1.9. Let X be an algebraic stack, and suppose that {Zi}i∈I is a finite dia- gram in RSch(X). Then a geometric colimit (Z, {φi}i∈I ) in RSch(X) is also a uniform geometric and uniform categorical colimit in RSch(X). Proof. By Lemma 1.5, it suffices to show that Z is a categorical colimit. Let U → X be a smooth surjection from a scheme. By Lemmas 1.5 and 1.6, we see that (ZU , {(φi)U }i∈I) is a categorical colimit of the diagram {(Zi)U }i∈I in RSch(U ), and remains so after flat schematic base change on U . Suppose that we have compatible X-morphisms αi : Zi → W . Then we want to show that there is a unique X-morphism α : Z → W that is compati- ble with this data. Let R = U ×X U , and let s, t : R → U denote the two projections. Since ZU is the categorical colimit in RSch(U ), there is a unique X-morphism βU : ZU → WU that is compatible with (αi)U : (Zi)U → WU . Suppose for the moment that X is an algebraic space. Then R is a scheme and the maps s, t : R → U are morphisms of schemes. In particular, we know that ZR is a categorical colimit in RSch(R) and so there is a unique morphism βR : ZR → WR that is compatible with the morphisms (αi)R : (Zi)R → WR. Noting that (βU ) ×U,s R and (βU ) ×U,t R are also such maps, we conclude that these maps are actually equal (as maps of algebraic spaces) and so by smooth descent, we conclude that there is a unique X-morphism α : Z → W , and so (Z, {φi}i∈I ) is a categorical colimit if X is an algebraic space. Applying Lemma 1.5, one is able to conclude that (Z, {φi}i∈I ) remains a categorical colimit after flat representable base change on X, for all algebraic spaces X. Now, returning to the case that X is an algebraic stack, we know that s, t : R → U are flat representable morphisms, repeating the descent argument given above shows that (Z, {φi})i∈I is a categorical colimit in RSch(X). (cid:3) The following result is a generalization of [Kol11, Lem. 17], which treats the case of a finite equivalence relation of schemes. Theorem 1.10. Let X be a locally noetherian algebraic stack. Let {Zi}i∈I be a finite diagram in QFfin s (X). Moreover, this colimit is a uniform categorical colimit in RSch(X) and a uniform geometric colimit. s (X). Then a colimit (Z, {φi}i∈I ) exists in QFfin We will need some lemmas to prove Theorem 1.10. We first treat the affine case, then the finite case and finally the quasi-finite case. Lemma 1.11. Let X be an algebraic stack. Let {Zi}i∈I be a finite diagram in RAff (X). Then this diagram has a categorical colimit in RAff (X), whose formation commutes with flat base change on X. Proof. By [LMB, Prop. 14.2.4], there is an anti-equivalence of categories between RAff (X) and the category of quasi-coherent sheaves of OX-algebras, which commutes with arbi- s♯ trary change of base, and is given by (Z s−→ X) 7→ (OX −→ s∗OZ ). Since the category of quasi-coherent OX-algebras has finite limits, it follows that if si : Zi → X denotes the (si)∗OZi. Since flat structure map of Zi, then the categorical colimit is Z = SpecX lim←−i pullback of sheaves is exact, the formation of this colimit commutes with flat base change on X. (cid:3) 10 J. HALL AND D. RYDH Lemma 1.12. Let X be a locally noetherian algebraic stack. Let {Zi}i∈I be a finite diagram in RSch(X) such that Zi → X is finite for every i ∈ I. Then this diagram has a uniform categorical and uniform geometric colimit Z in RSch(X) which is finite over X. Proof. By Lemma 1.11 the diagram has a categorical colimit (Z, {φi}i∈I ) in RAff (X). If si : Zi → X and s : Z → X denote the structure maps, then s∗OZ ⊆ Q(si)∗OZi. Since X is locally noetherian, it follows that Z is finite over X. By Proposition 1.9, it remains to show that Z is a geometric colimit. Note that since `i∈I Zi → Z is dominant and finite, it is surjective, universally closed and thus universally submersive. We will now show that Z is a universal Zariski colimit using the criterion of Lemma 1.7(2). To apply this criterion we will employ a generalization of the arguments given in [Kol11, Lem. 17]. Let x : Spec k → X be a geometric point. By [EGA, 0III.10.3.1] this map factors as Spec k −→ X, where p is flat and X 1 is the spectrum of a maximal- adically complete, local noetherian ring with residue field k. Since the algebraic stack Z is a uniform categorical colimit in RAff (X), we may replace X by X 1, and we denote the unique closed point of X by x. x1 −→ X 1 p sU−−→ ∐m∈π0(U)X → X such that (sU )x = hU . For a finite X-scheme U , let π0(U ) be its set of connected components. The scheme X is henselian [EGA, IV.18.5.14], so there is a unique universal homeomorphism hU : Ux → ∐m∈π0(U){x} which is functorial with respect to U . In particular, there is a unique factor- ization U Let Wi = `m∈π0(Zi) X. Since π0(−) is a functor, we obtain a diagram {Wi}i∈I in RAff (X) and we let W be the categorical colimit of this diagram in RAff (X). It is readily seen that W = `m∈lim−→ π0(Zi) X so that there is a bijection of sets lim−→i π0(Zi) → π0(W ). Since Z is a categorical colimit, there is a canonical map µ : Z → W . In par- ψx−−→ Zx µx−−→ Wx ticular, the bijection νx : lim−→i and thus ψx : lim−→i (Zi)x → Zx is injective. Hence, we have shown that Z is a universal Zariski colimit and it remains to show that Z has the correct functions. (Zi)x → Wx factors as lim−→i (Zi)x There is a canonical morphism of sheaves of OX-algebras ǫ : OZ → lim←−i (φi)∗OZi, which we have to show is an isomorphism. By functoriality, we have an induced morphism of sheaves of OX-algebras: ǫ2 : s∗OZ s∗ǫ−−→ s∗(cid:0)lim←− (φi)∗OZi(cid:1) ǫ1−→ lim←− i i s∗(φi)∗OZi. Since the functor s∗ is left exact, ǫ1 is an isomorphism; by construction of Z the map ǫ2 is an isomorphism and so s∗ǫ is an isomorphism. Since s is affine, the functor s∗ is faithfully exact and we conclude that the map ǫ is an isomorphism of sheaves. (cid:3) Proof of Theorem 1.10. By hypothesis, the structure morphisms si : Zi → X are quasi- finite, separated, and representable. By Zariski's Main Theorem [LMB, Thm. 16.5(ii)], there are finite X-morphisms si : Zi → X and open immersions ui : Zi ֒→ Zi. Let Wi = Qi→j Zj where the product is fibered over X. There is an induced morphism uj−→ Zj over the factor corresponding Zi → Wi given by the composition Zi to the arrow h : i → j. Using the composition in I, there is a natural diagram {Wi}i∈I such that the morphisms Zi → Wi induce a morphism of diagrams. Since Zi → Zi is an immersion, so is Zi → Wi. Now, replace Wi with the schematic image of Zi → Wi. Then Zi → Wi is an open dense immersion and {Wi}i∈I is a diagram with finite structure morphisms Wi → X. Thus, the diagram {Wi}i∈I has a uniform geometric and uniform categorical colimit W ∈ RSch(X) and W → X is finite (Lemma 1.12). h∗−→ Zj We will now show that for any arrow h : i → j, the open substack Zj ⊂ Wj is pulled β back to the open substack Zi ⊂ Wi. We have canonical maps Zi −→ Wi and since the maps β ◦ α and β are open immersions, so is α : Zi → Zj ×Wj Wi. Similarly, γ −→ Zj where γ ◦ α and γ are finite morphisms. Thus α : Zi → we have Zi α−→ Zj ×Wj Wi α−→ Zj ×Wj Wi THE HILBERT STACK 11 Zj ×Wj Wi is open and closed. Since ui = β ◦ α : Zi → Wi is dense, we conclude that Zi = Zj ×Wj Wi. Let Z be the set-theoretic image of `i Zi in W . As W is a Zariski colimit, it follows that Z = lim−→i Zi is open in W . We let Z ⊂ W be the open substack with underlying topological space Z. Then the diagram {Zi}i∈I is obtained as the pull-back of the diagram {Wi}i∈I along the open immersion Z → W . Since W is a uniform geometric colimit, the pull-back Z is a uniform geometric colimit. As Z ∈ RSch(X), it is a uniform categorical colimit in RSch(X) by Proposition 1.9. (cid:3) 1.2. Completions of schemes. Here we will show that the colimits constructed in Theo- rem 1.10 remain colimits after completing along a closed subset. Denote the category of formal schemes by FSch. We require some more definitions that are analogous to those given in §1.1. Definition 1.13. Consider a diagram of formal schemes {Zi}i∈I, a formal scheme Z, and suppose that we have compatible maps ϕi : Zi → Z for every i ∈ I. Then we say that the data (Z, {ϕi}i∈I) is a is a homeomorphism; (1) formal Zariski colimit if the induced map on topological spaces lim−→i (2) formal weak geometric colimit if it is a formal Zariski colimit and the canonical (ϕi)∗OZi is a topological isomorphism, where map of sheaves of rings OZ → lim←−i we give the latter sheaf of rings the limit topology (this is nothing other than Z being the colimit in the category of topologically ringed spaces). Zi → Z If, in addition, the formal scheme Z is locally noetherian, and the maps ϕi : Zi → Z are topologically of finite type, then we say that the data (Z, {ϕi}i∈I) is a (3) universal formal Zariski colimit if for any adic formal Z-scheme Y, the data (Y, {(ϕi)Y}i∈I) is the formal Zariski colimit of the diagram {(Zi)Y}i∈I; (4) formal geometric colimit if it is a universal formal Zariski and a formal weak geometric colimit of the diagram {Zi}i∈I; (5) uniform formal geometric colimit if for any adic flat formal Z-scheme Y, the data (Y, {(ϕi)Y}i∈I ) is a formal geometric colimit of the diagram {(Zi)Y}i∈I. We have two lemmas which are the analogues of Lemmas 1.5 and 1.6 for formal schemes. Lemma 1.14. Let {Zi}i∈I be a diagram of formal schemes. Then every weak geometric colimit (Z, {ϕi}i∈I ) is a colimit in the category of topologically locally ringed spaces and, consequently, a colimit in the category FSch. Lemma 1.15. Consider a finite diagram of locally noetherian formal schemes {Zi}i∈I, and a locally noetherian formal scheme Z, together with finite morphisms ϕi : Zi → Z. If the data (Z, {ϕi}i∈I ) is a formal geometric colimit of the diagram {Zi}i∈I, then it is a uniform formal geometric colimit. Proof. Let : Y → Z be an adic flat morphism of locally noetherian formal schemes. It remains to show that the canonical map ϑY : OY → lim←−i [(ϕi)Y]∗O(Zi)Y is a topological isomorphism. Since ϕi is finite for all i, and we are taking a finite limit, we conclude that it suffices to show that ϑY is an isomorphism of coherent OY-modules. By hypothesis, the (ϕi)∗OZi is an isomorphism, and since is adic flat, ∗ is an exact map ϑZ : OZ → lim←−i functor from coherent OZ-modules to coherent OY-modules, thus commutes with finite limits. Hence, we see that ϑY factors as the following sequence of isomorphisms: OY ∼= ∗OZ ∼= lim←− i ∗[(ϕi)∗OZi ] ∼= lim←− i [(ϕi)Y]∗O(Zi)Y . (cid:3) 12 J. HALL AND D. RYDH Definition 1.16. Fix a locally noetherian formal scheme X and let {Zi}i∈I be a diagram in FSch/X, where each formal scheme Zi is locally noetherian. Let Z ∈ FSch/X also be lo- cally noetherian, and suppose that we have compatible X-morphisms ϕi : Zi → Z that are topologically of finite type. Then we say that the data (Z, {ϕi}i∈I ) is a uniform categori- cal colimit in FSch/X if for any adic flat formal X-scheme Y, the data (ZY, {(ϕi)Y}i∈I) is the categorical colimit of the diagram {(Zi)Y}i∈I in FSch/Y. Definition 1.17. Let X be a formal scheme. Define QFs(X) to be the category whose σ−→ X). A morphism in QFs(X) is an objects are adic, quasi-finite, and separated maps (Z X-morphism f : (Z σ−→ X) → (Z′ σ′ −→ X). For a scheme X, and a closed subset V ⊂ X, we define the completion functor cX,V : Sch/X → FSch/ bX/V , (Z s−→ X) 7→ (cid:0)bZ/s−1V → bX/V (cid:1). Note that restricting cX,V to QFs(X) has essential image contained in QFs( bX/V ). Theorem 1.18. Let X be a locally noetherian scheme, let V ⊂ X be a closed sub- set, and let {Zi}i∈I be a diagram in QFfin s (X). If the scheme Z denotes the categorical colimit of the diagram in Sch/X, which exists by Theorem 1.10, then cX,V (Z) is a uni- form categorical and uniform formal geometric colimit of the diagram {cX,V (Zi)}i∈I in FSch/ bX/V . Proof. Let X = bX/V , Z = cX,V (Z), and Zi = cX,V (Zi). By Lemmas 1.14 and 1.15, it suffices to show that Z is a formal geometric colimit of the diagram {Zi}i∈I. We first show that Z is a universal formal Zariski colimit. Let : Y → X be an adic morphism of locally noetherian formal schemes. Let I be a coherent sheaf of radical ideals defining V ⊂ X so that (−1I)OY is an ideal of definition of Y. Let X0 = V (I) ⊂ X and Y0 = Y ×X X0. Then by Theorem 1.10 the map of topological spaces lim−→i∈I Zi ×X Y0 → Z ×X Y0 is a homeomorphism. Noting that ZY = Z ×X Y0 and ZiY = Zi ×X Y0, we conclude that Z is a universal formal Zariski colimit and it remains to show that Z has the correct functions. Let φi : Zi → Z denote the canonical morphisms. By Theorem 1.10 we have an iso- (φi)∗OZi. Hence, since φi is finite, and morphism of sheaves of rings ǫ : OZ → lim←−i completion is exact on coherent modules [EGA, I.10.8.9], by [EGA, I.10.8.8(i)] we have an isomorphism of coherent OZ-algebras: OZ ǫ∧ −→ (lim←− i φi∗OZi )∧ ∼= lim←− i (φi∗OZi )∧ ∼= lim←− i bφi∗OZi. It remains to show that this isomorphism is topological (where we endow the right side with the limit topology). A general fact here is that the topology on the right is the subspace topology of the product topology on Qi∈I bφi∗OZi. Since φi is finite, Krull's Theorem [EGA, 0I.7.3.2] implies that the limit and adic topologies coincide. The result follows. (cid:3) 1.3. Completions of algebraic stacks. Here we observe that the colimits constructed in Theorem 1.10 remain colimits after completing along a closed subset. To avoid developing the theory of formal algebraic stacks, we will work with adic systems of algebraic stacks. This has the advantage of being elementary as well as sufficient for our purposes. Let X be a locally noetherian algebraic stack and suppose that V ⊆ X is a closed subset, defined by a coherent OX-ideal I. For each n ≥ 0 let Xn = V (I n+1) and define: QFs( bX/V ) := lim←− n QFs(Xn). This is an abuse of notation, and for that we apologize. We firmly believe, however, that this notation will be sufficiently convenient to outweigh any potential confusion that may THE HILBERT STACK 13 arise. Also, there is a completion functor: cX,V : QFs(X) → QFs( bX/V ), (Z → X) 7→ (Z ×X Xn → Xn)n≥0. We conclude this section with a corollary, which is an immediate consequence of Theorem 1.10, smooth descent, and Theorem 1.18. Corollary 1.19. Let X be a locally noetherian algebraic stack. Suppose that V ⊆ X is a closed subset. Let {Zi}i∈I be a diagram in QFfin s (X) and let Z ∈ QFs(X) be the categorical colimit, which exists by Theorem 1.10. Then cX,V (Z) is a categorical col- imit of the diagram {cX,V (Zi)}i∈I in QFs( bX/V ), and remains so after flat and locally noetherian base change on X. 2. GENERALIZED STEIN FACTORIZATIONS A morphism of schemes f : X → Y is Stein if the morphism f ♯ : OY → f∗OX is an isomorphism. If f is a proper morphism of locally noetherian schemes that is Stein, then Zariski's Connectedness Theorem [EGA, III.4.3.2] implies that f is surjective with geometrically connected fibers. Note that if f is quasi-compact and quasi-separated then it factors as: X r−→ eX f −→ Y, where r is Stein and f is affine. Indeed, we simply take eX = SpecY (f∗OX ). In general, this factorization says little about f . In the case where Y is locally noetherian and f is proper, however, one obtains the well-known Stein factorization of f [EGA, III.4.3.1]. In this case, f is finite and r is proper and surjective with geometrically connected fibers. Note that the Stein factorization is also unique and compatible with flat base change on Y . We would like to generalize these properties of the Stein factorization to non-proper morphisms. We also wish to point out that the above discussion is perfectly valid for locally noe- therian formal schemes. That is, if ϕ : X → Y is a proper morphism of locally noetherian ϕ formal schemes, then it admits a Stein factorization: X −→ Y, where ϕ is finite and ρ is proper and Stein. The Stein factorization is compatible with (not necessarily adic) flat base change on Y (Proposition A.1) and ρ has geometrically connected fibers (Corol- lary A.2). −→ eX ρ In this section we will address the following question: suppose that ϕ : X → Y is an adic morphism of locally noetherian formal schemes that is locally of finite type. Is there ϕ a factorization of ϕ as X −→ Y, where ϕ is locally quasi-finite and ρ is proper and Stein? If such a factorization exists, we call it a generalized Stein factorization. A desirable property of a generalized Stein factorization will be that it is compatible with flat base change on Y. −→ eX ρ Note that a necessary condition for ϕ : X → Y to admit a generalized Stein factorization is that every w ∈ X lies in a proper connected component of the fiber Xϕ(w). We will call such morphisms locally quasi-proper. If ϕ is separated (resp. quasi-compact), then so is ϕ in any generalized Stein factorization since ρ is proper and surjective. In §2.4 we prove Theorem 2.1. Let Y be a locally noetherian formal scheme. Suppose that ϕ : X → Y is locally quasi-proper and separated. Then ϕ admits a generalized Stein factorization ϕ −→ Y which is unique and compatible with flat base change on Y. Moreover, ϕ is X separated and if ϕ is quasi-compact, so is ϕ. −→ eX ρ In future work, we will prove Theorem 2.1 for algebraic stacks in order to compare our generalized Stein factorizations with the connected component fibrations of [LMB, §6.8] and [Rom11, Thm. 2.5.2]. We wish to point out that Theorem 2.1 should extend to a reasonable theory of formal algebraic spaces or stacks. Unfortunately, the only treatise on formal algebraic spaces that we are aware of is D. Knutson's [Knu71, V], and this requires 14 J. HALL AND D. RYDH everything to be separated [Knu71, V.2.1]-an assumption we definitely do not wish to impose. We are not aware of any written account of a theory of formal algebraic stacks. Remark 2.2. We will construct the generalized Stein factorization by working ´etale-locally on Y. If f : X → Y is a separated and locally quasi-proper morphism of schemes that is quasi-compact, then there is a more explicit construction of the generalized Stein factor- ization. If Y denotes the integral closure of Y in f∗OX, then the image of X → Y is open and equals eX. This follows from Theorem 2.1 and [EGA, IV.8.12.3]. Proving directly that the image is open, that X → eX is proper and that eX → Y is quasi-finite is possible but not trivial. This construction also has the deficiency that we were unable to easily adapt it to locally noetherian formal schemes. 2.1. Uniqueness and base change of generalized Stein factorizations. In this subsec- tion we address the uniqueness and base change assertions in Theorem 2.1. Both assertions will also be important for the proof of the existence of generalized Stein factorizations in Theorem 2.1. Lemma 2.3. Let ϕ : X → Y be a morphism of locally noetherian formal schemes that ϕ admits a factorization X −→ Y where ρ is proper and ϕ is locally quasi-finite and separated. Then ρ is Stein if and only if the natural map: −→ eX ρ HomY(eX, Z) → HomY(X, Z) is bijective for every locally quasi-finite and separated morphism s : Z → Y. In particular, for locally quasi-proper and separated morphisms, the generalized Stein factorization, if it exists, is unique. Proof. A Y-morphism eX → Z is equivalent to a section of the projection Z ×Y eX → eX. In particular, we may now replace Y with eX and Z with Z ×Y eX. Thus we have reduced the result to the situation where ϕ is proper and eX = Y. β′ −→ StY(X) b−→ Y be the Stein factorization of ϕ (which exists because ϕ is proper). Fix a Y-morphism α : X → Z. Since Z is separated over Y and ϕ is proper, α′ −→ StZ(X) a−→ it follows that α is proper. Thus, there is a Stein factorization of α : X Z. Note, however, that γ : StZ(X) → Y is quasi-finite and proper, thus finite [EGA, III.4.8.11]. Next observe that we have natural isomorphisms of coherent sheaves of OY- algebras: Now let X ∗OX ∼= γ∗α′ ∗OX ∼= γ∗OStZ(X). b∗OStY(X) ∼= b∗β′ Hence, there exists a unique Y-isomorphism δ : StY(X) → StZ(X) that is compatible with the data. If ϕ is Stein, then StY(X) = X and b = idX. We deduce that there is a uniquely δ−→ StZ(X) a−→ Z. Conversely, we have a Stein factorization induced Y-morphism Y X → St eX(X) → eX. By the Stein case already considered, we see that St eX(X) and eX both satisfy the same universal property for locally quasi-finite and separated Y-schemes. The result follows. (cid:3) Fix a locally noetherian formal scheme Y and let LQPs(Y) (resp. LQFs(Y)) denote the category of locally quasi-proper (resp. locally quasi-finite) and separated morphisms X → Y. Lemma 2.3 allows us to reinterpret Theorem 2.1 as the existence of a left adjoint: StY : LQPs(Y) → LQFs(Y) to the inclusion LQFs(Y) ⊆ LQPs(Y) such that if Z ∈ LQPs(Y), then: (1) the natural morphism Z → StY(Z) is proper; (2) if Y′ → Y is flat, then we have a natural isomorphism: StY′(Z ×Y Y′) ∼= StY(Z) ×Y Y′. We address property (2) immediately. Lemma 2.4. Fix a cartesian diagram of locally noetherian formal schemes: THE HILBERT STACK 15 X′ p′ ϕ′ Y′ p X ϕ / Y. Suppose that ϕ is separated and admits a generalized Stein factorization X → StY(X) → Y. Then ϕ′ admits a generalized Stein factorization X′ → StY′(X′) → Y′ and there is a naturally induced Y′-morphism: h : StY′ (X′) → StY(X) ×Y Y′, which is a finite universal homeomorphism. In addition, if p is flat, then the morphism h is an isomorphism. Proof. The morphism ϕ′ factors as X′ γ −→ StY(X) ×Y Y′ σ−→ Y′ where σ is locally quasi- finite and separated and γ is proper. Next observe that the Stein factorization of γ is the generalized Stein factorization of ϕ′. Thus γ factors as: X′ → StY′(X′) h−→ StY(X) ×Y Y′. By Zariski's Connectedness Theorem (Corollary A.2) X′ → StY′(X′) and γ have geo- metrically connected fibers. It follows that h is a finite universal homeomorphism (it is finite with geometrically connected fibers). In addition, if p is flat, then since cohomol- ogy commutes with flat base change (Proposition A.1), γ is already Stein, and so h is an isomorphism. (cid:3) 2.2. Local decompositions. Before we prove Theorem 2.1 it will be necessary to under- stand the ´etale local structure of locally quasi-proper morphisms. We accomplish this by generalizing the well-known structure results that are available for locally quasi-finite and separated morphisms [EGA, IV.18.5.11c, IV.18.12.1]. Proposition 2.5. Let ϕ : X → Y be an adic morphism of locally noetherian formal schemes that is locally of finite type and separated. Let y : Spec(k) → Y be a point and suppose that Z is a connected component of the fiber Xy. If Z is proper, then there exists an adic ´etale neighborhood (Y′, y′) → (Y, y) and an open and closed immersion i : XZ → X ×Y Y′ where XZ → Y′ is proper and (XZ)y′ ∼= Z. Proof. We first prove the result in the setting of a morphism of locally noetherian schemes f : X → Y . We now immediately reduce to the case where Y is affine and f is quasi- compact. Let k0 ⊆ k denote the separable closure of the residue field k(y) in k. Then the resulting morphism Xy → Xk0 induces a bijection on connected components [EGA, IV.4.3.2]. So, we may replace k with k0 and Z by the corresponding connected component in Xk0. By standard limit and smearing out arguments [EGA, IV.8] we further reduce to the case where Y is local and henselian with closed point y and residue field k. By Chow's Lemma [EGA, II.5.6.1], there exists a quasi-projective Y -scheme W and a proper and surjective Y -morphism p : W → X. If the result has been proved for W → Y , then we claim that this implies the result for f : X → Y . Indeed, each connected compo- nent of p−1(Z) is proper, thus we may write W = Wp−1(Z) ∐ W ′ where Wp−1(Z) → Y is proper, and the connected components of Wp−1(Z) coincide with the connected compo- nents of (Wp−1(Z))y = p−1(Z). Now take XZ = p(Wp−1(Z)) and X ′ = p(W ′). Then XZ is proper over Y and it remains to show that XZ ∩ X ′ = ∅. Note that XZ ∩ X ′ is a closed subset of XZ. If it is non-empty, then it must have non-empty intersection with Z. But this would imply that Z ∩ X ′ 6= ∅, hence p−1(Z) ∩ W ′ 6= ∅-a contradiction. We deduce that it remains to prove the result when X is also quasi-projective over Y . In this case, there is an open and dense immersion i : X ֒→ X where X is a projective Y -scheme. Note that if V is a connected component of X y and V ∩Z 6= ∅, then V ∩Z = V / / / 16 J. HALL AND D. RYDH since Z is proper. By Corollary B.4, there is also a bijection between the set of connected components of X y and X. Set XZ to be the union of those connected components of X that meet Z. Observe that XZ ∩ X is an open subset of XZ which contains (XZ )y = Z. Since XZ → Y is proper and Y is a local scheme, we have that XZ ∩ X = XZ and we conclude that XZ ⊆ X. In particular, we readily deduce that XZ is an open and closed subset of X and we have the claimed decomposition X = XZ ∐ X ′ with (XZ)y = Z. Now let ϕ : X → Y be as in the Proposition. Let f : X → Y denote the induced morphism on the underlying schemes. By what we have proven for schemes, there is an ´etale neighborhood (Y ′, y′) → (Y, y) and an open and closed immersion i : XZ → X ×Y Y ′ where XZ → Y ′ is proper and (XZ )y′ ∼= Z. By [EGA, IV.18.1.2], there is a unique lift of the ´etale neighborhood (Y ′, y′) → (Y, y) to an adic ´etale neighborhood (Y′, y′) → (Y, y) whose underlying morphism of pointed schemes coincides with (Y ′, y′) → (Y, y). The result follows. (cid:3) Remark 2.6. In future work we will prove Proposition 2.5 for non-separated morphisms of non-noetherian algebraic stacks. We do not treat this here as the necessary reformulations in the non-locally-separated case would take us too far afield. We conclude this subsection with a trivial, but important, corollary. Corollary 2.7. Let ϕ : X → Y be a morphism of locally noetherian formal schemes that is separated and locally quasi-proper. Then there exists an adic ´etale morphism Y′ → Y, and an open and closed immersion i : U ֒→ X ×Y Y′ such that the composition U → X ×Y Y′ → X is surjective and the composition U → X ×Y Y′ → Y′ is proper. Thus Corollary 2.7 implies that any separated and locally quasi-proper morphism ϕ : X → Y admits an ´etale slice that has a generalized Stein factorization. Thus in order to construct the generalized Stein factorization of ϕ, we can use ´etale descent. The problem now is thus the description of the relevant descent data. In order to this, we will address a more general problem in §2.3: for a proper and Stein morphism, which ´etale morphisms are pulled back from the base? 2.3. Pullbacks of ´etale morphisms. Let ϕ : X → Y be a proper and Stein morphism of locally noetherian formal schemes. In this subsection we will identify those ´etale mor- phisms to X that are pulled back from Y. So, let S be a locally noetherian formal scheme. Define Ets(S) to be the category of morphisms (V → S) that are adic ´etale and separated. Also, define OC(S) to be the set of open and closed immersions into S. If S ⊆ S denotes the underlying scheme of S, then the natural map OC(S) → OC(S) is bijective. More generally, an easy consequence of [EGA, IV.18.1.2] is that the functor Ets(S) → Ets(S) is an equivalence of categories. Let ϕ : X → Y be an adic morphism of locally noetherian formal schemes. There is an induced functor ϕ∗ : Ets(Y) → Ets(X), (V → Y) 7→ (V ×Y X → X). Let Ets,t/Y(X) be the full subcategory of Ets(X) with objects those W → X such that for any geometric point y of Y, and any connected component Z of Wy, the induced morphism Z → Xs is an open and closed immersion. Note that OC(X) ⊆ Ets,t/Y(X). In the following Lemma, we characterize the essential image of ϕ∗. Lemma 2.8. Let ϕ : X → Y be a proper and Stein morphism of locally noetherian formal schemes. (1) The functor ϕ∗ : Ets(Y) → Ets(X) is fully faithful with image Ets,t/Y(X). (2) The induced map OC(Y) → OC(X) is bijective. (3) If W ∈ Ets,t/Y(X), then it admits a generalized Stein factorization W → StY(W) → Y, with StY(W) → Y ´etale and separated. Moreover, W = Y ×X StY(W). THE HILBERT STACK 17 Proof. For (1) and (2), by passing to the underlying morphism of schemes we obtain, via Zariski's Connected Theorem (Corollary A.2), a proper and surjective morphism of lo- cally noetherian schemes f : X → Y with geometrically connected fibers. By Proposition B.1 and Remark B.2, the functor f ∗ : Ets(Y ) → Ets(X) is fully faithful with image Ets,t/Y (X) and the induced map OC(Y) → OC(X) is bijective. The claim (3) follows from (1) and Lemma 2.4. (cid:3) 2.4. Existence of generalized Stein factorizations. In this subsection we will prove The- orem 2.1. Before we can accomplish this we require one more Lemma. This Lemma is well-known for schemes, though requires some care for formal schemes. Lemma 2.9. Let X be a locally noetherian formal scheme and let U → X be an adic, locally quasi-finite, and separated morphism. Suppose that [R ⇒ U] is an adic flat equiv- alence relation over X such that R → U×X U is a closed immersion. Then this equivalence relation has a formal geometric quotient Z that is adic, locally quasi-finite, and separated over X. Moreover, taking the quotient commutes with arbitrary (not necessarily adic) base change, the quotient map U → Z is adic and faithfully flat, and R = U ×Z U. Proof. Let I be an ideal of definition for X and denote XI = V (I), UI = U ×X XI, and RI = R ×X XI. Then, the hypotheses ensure that [RI ⇒ UI] is an fppf equivalence relation over XI. We observe that the quotient of this fppf equivalence relation in the category of algebraic spaces is a scheme ZI, as it is locally quasi-finite and separated over XI [Knu71, II.6.16]. Note that for any other ideal of definition J ⊃ I there is a natural isomorphism ZI ×X XJ ∼= ZJ. Hence, the directed system {ZI}I is adic over X. Taking the colimit of this system in the category of topologically ringed spaces produces a locally noetherian formal scheme Z, which is adic, locally quasi-finite, and separated over X. It is now immediate that Z is a formal geometric quotient of the given equivalence relation. That Z has the stated properties is an easy consequence of the corresponding results for its truncations ZI. (cid:3) We now proceed to the proof of Theorem 2.1. Proof of Theorem 2.1. The latter claims are consequences of Lemmas 2.3 and 2.4, thus it remains to address the existence of generalized Stein factorizations. Note that if Y → T is locally quasi-finite and separated, then StY(X) ∼= StT(X) over T. Moreover, Ets,t/Y(X) ≃ Ets,t/T(X). Thus if ϕ : X → Y factors as X → Y′ → Y, where X → Y′ is proper and Y′ → Y is locally quasi-finite and separated, then the Stein factorization of X → Y′ gives the generalized Stein factorization of X → Y. So, we fix a locally quasi-proper and separated morphism ϕ : X → Y. By Corollary 2.7 there is an adic, ´etale, and separated morphism Y′ → Y, together with an open and closed immersion i : U ֒→ X′ := X ×Y Y′, such that the induced morphism U → Y′ is proper, and the induced morphism U → X is adic, ´etale, separated, and surjective. By the remarks above we obtain a generalized Stein factorization U → StY(U) → Y (note that StY(U) = StY′(U)). Let Y′′ = Y′ ×Y Y′ and take s1, s2 : Y′′ → Y′ to denote the two projections. Let X′′ = X′ ×X X′ and denote by t1 and t2 the two projections. For j = 1 and 2 let Rj denote the pullback of U along tj. Note that the morphisms ij : Rj → X′′ are open and closed immersions. Let R = R1 ∩ R2 = U ×X U. Then we obtain an adic ´etale equivalence relation [R ⇒ U] with quotient X. Note that both morphisms R → U belong to Ets,t/Y(U). The morphism U → StY(U) is also proper and Stein. By Lemma 2.8(1) we see that the functor Ets(StY(U)) → Ets,t/Y(U) is an equivalence of categories (note that Ets,t/Y(U) ≃ Ets,t/StY(U)(U)). Functoriality, together with Lemma 2.8(3), produces an adic ´etale groupoid [StY(R) ⇒ StY(U)] in LQFs(Y) which pulls back to the equivalence relation [R ⇒ U]. We will now verify that StY(R) → StY(U) ×Y StY(U) is a closed immersion. 18 J. HALL AND D. RYDH Note that for j = 1 and 2 we have that R ∈ OC(Rj). By Lemma 2.8(2,3) we thus see that StY(R) ∈ OC(StY(Rj)). We also know that StY(Rj) is the pullback of StY(U) → Y′ along sj : Y′′ → Y′ (Lemma 2.4). Thus we obtain a Y-isomorphism: StY(R1) ×Y′′ StY(R2) → StY(U) ×Y StY(U). We have already seen, however, that the map StY(R) → StY(R1) is an open and closed immersion. Since everything is separated, it follows that the natural map StY(R) → StY(R1) ×Y′′ StY(R2) = StY(U) ×Y StY(U) is a closed immersion. Thus, we have an adic ´etale equivalence relation [StY(R) ⇒ StY(U)] in LQFs(Y) satisfying the hypothe- ses of Lemma 2.9. Let eX ∈ LQFs(Y) be the quotient of this equivalence relation. Since StY(U) → eX is adic ´etale and StY(R) = StY(U) × eX StY(U), the proper and Stein morphism U → StY(U), with descent data given by R → StY(R), descends to a proper and Stein mor- phism X → eX. This is the generalized Stein factorization of ϕ by Lemma 2.3. (cid:3) 3. ADJUNCTIONS FOR QUASI-FINITE SPACES If p : X ′ → X is a morphism of algebraic stacks, then there is a pullback functor p∗ : QFs(X) → QFs(X ′) given by (Z → X) 7→ (Z ×X X ′ → X ′). More generally, given a closed subset V ⊆ X, set V ′ = p−1V . Then there is an induced pullback /V ′). In this short section we will use general- functor bp∗ ized Stein factorizations to construct left adjoints to these functors when p is proper and schematic (i.e., those proper morphisms that are representable by schemes). This is cer- tainly true in greater generality, though we limit ourselves to this situation for simplicity. /V : QFs( bX/V ) → QFs( bX ′ Definition 3.1. Suppose that p : X ′ → X is a proper morphism of locally noetherian alge- braic stacks. Let V ⊆ X be a closed subset and let V ′ = p−1V . Let QFs,qp/X ( bX ′ denote the full subcategory of QFs( bX ′ n)n≥0 such that the composition Z0 → X ′ X ′ We now have the following Theorem. /V ′) /V ′) consisting of those adic systems (Zn → 0 → X0 has proper fibers. Theorem 3.2. Suppose that p : X ′ → X is a proper and schematic morphism of locally noetherian algebraic stacks. Let V ⊆ X be a closed subset and let V ′ = p−1V . Then the functor bp∗ /V ′) and admits a left adjoint: /V ′) factors through QFs,qp/X ( bX ′ /V : QFs( bX/V ) → QFs( bX ′ (bp/V )! : QFs,qp/X ( bX ′ /V ′) → QFs( bX/V ), which is compatible with locally noetherian and flat base change on X. Moreover, for (Z → X ′) ∈ QFs,qp/X (X ′), the induced natural map: (bp/V )! ◦ cX ′,V ′(Z → X ′) → cX,V ◦ p!(Z → X) is an isomorphism. Proof. We wish to point out that if V = X, then bX/V = X and QFs( bX/V ) = QFs(X). Thus we denote bp∗ /V by p∗ and the functor p! (which appears at the end of the statement) is a left adjoint to p∗. Now the first claim about the factorization is trivial. For the existence and properties of (bp/V )!, by smooth descent, it is sufficient to prove the result in the case where p : X ′ → X is a morphism of locally noetherian schemes. Set X = bX/V and X′ = bX ′ /V ′ and let π : X′ → X be the induced morphism. Let (Zn → X ′ n)n≥0 ∈ QFs,qp/X(X′) which we may also view as a quasi-finite and separated morphism of locally noetherian formal schemes Z → X′ such that the compo- sition Z → X′ → X is locally quasi-proper and separated. By Theorem 2.1, there is a THE HILBERT STACK 19 ρ −→ StX(Z) → X and StX(Z) → X is quasi-finite and generalized Stein factorization Z separated. We set π!Z = StX(Z) and note that π! is a left adjoint to π∗ by Lemma 2.3. The compatibility with flat base change and completions is implied by the flat base change property of StX, together with the fact that X → X is a (non-adic) flat morphism [EGA, I.10.8.9]. (cid:3) 4. THE EXISTENCE THEOREM Definition 4.1. Fix a morphism of algebraic stacks π : X → Y and a closed subset Y0 ⊆ Y . Let X0 = π−1Y0. Define QFp/Y ( bX/X0) to be the full subcategory of QFs,qp/Y ( bX/X0) consisting of those (Zn → X)n≥0 such that the composition Z0 → X → Y is proper. s ( bX/X0). That is, every Note that QFp/Y ( bX/X0) is also a full subcategory of QFfin morphism in QFp/Y ( bX/X0) is finite. Throughout this section we fix a noetherian ring R and an ideal I ⊆ R, such that R is I-adic. Let S = Spec R and Sn = Spec(R/I n+1). Consider a morphism of algebraic stacks π : X → S that is locally of finite type. For n ≥ 0 let Xn = X ×S Sn. It will be convenient for us to abbreviate the symbol bX/X0 to bX. Consider the I-adic completion functor: Ψπ,I : QFp/S(X) −→ QFp/S( bX), (Z → X) 7→ (Z ×X Xn → Xn)n≥0. In this section we prove an Existence Theorem, which gives sufficient conditions on the morphism π : X → S for the functor Ψπ,I to be an equivalence of categories. In the case that π : X → S is separated, the category QFp/S(X) is equivalent to the category of finite algebras over X with S-proper support. An immediate consequence of [Ols06, Thm. A.1] is that the functor Ψπ,I is an equivalence in this situation. Note that [loc. cit.] also gives a straightforward proof that Ψπ,I is fully faithful when π is no longer assumed to be separated. Indeed, we have Lemma 4.2. Let π : X → S be a morphism of algebraic stacks that is locally of finite type with quasi-compact and separated diagonal. Then the functor Ψπ,I is fully faithful. Proof. We begin with the following general observation: let f : V → W be a representable morphism of algebraic S-stacks. Let Sec(V /W ) denote the set of sections to f . Denote by fn : Vn → Wn the pullback of f along Sn ֒→ S. If W is a proper algebraic stack over S, and f is also separated and locally of finite type, then the natural map Sec(V /W ) → Sec(Vn/Wn) is bijective. Indeed, a section t : W → V of f is equivalent to a closed lim←−n immersion W → W ×S V such that the composition with the first projection is the identity. The algebraic S-stack W ×S V is also separated and [Con, 4.6] now gives the claim. To see that the functor Ψπ,I is fully faithful, we simply observe that if Z, eZ ∈ QFp/S(X), then HomQFp/S (X)(Z, eZ) = Sec([Z ×X eZ]/Z). (cid:3) Proving that Ψπ,I is an equivalence when π is not necessarily separated is the main technical contribution of this paper. In this section, we will prove Theorem 4.3. Let π : X → S be a morphism of algebraic stacks that is locally of finite type with quasi-compact and separated diagonal and affine stabilizers. Then Ψπ,I is an equivalence. Some interesting special cases of stacks with affine stabilizers are: (1) global quotient stacks, (2) stacks of global type [Ryd09, Defn. 2.1], (3) stacks with quasi-finite diagonal, (4) algebraic spaces, and 20 J. HALL AND D. RYDH (5) schemes. We wish to point out that Theorem 4.3 is even new for schemes. As outlined in §0.2, we will prove Theorem 4.3 by a d´evissage on the non-abelian category QFp/S(X). This d´evissage is combined with the Raynaud–Gruson Chow Lemma [RG71, Cor. 5.7.13] to reduce to proving that Ψπ,I is essentially surjective for a very special class of morphisms of schemes π : X → S. Before we get into our main lemmas to set up the d´evissage, the following definitions will be useful. Let π : X → S be a morphism of algebraic stacks that is locally of finite type with quasi-compact and separated diagonal. We say that (Zn → Xn)n≥0 ∈ QFp/S( bX) is effectivizable if it lies in the essential image of Ψπ,I. If (Zn → Xn)n≥0 ∈ QFp/S( bX) is effectivizable, we say that an object (Z → X) ∈ QFp/S(X) together with an isomor- phism e : Ψπ,I (Z → X) → (Zn → Xn)n≥0 is an effectivization of (Zn → Xn)n≥0. Note that given two effectivizations ((Z → X), e), ((Z ′ → X), e′) of an effectivizable (Zn → Xn)n≥0, then there exists a unique isomorphism α : (Z ′ → X) → (Z → X) such that e ◦ Ψπ,I (α) = e′ (Lemma 4.2). Let (ϕn)n≥0 : (Z ′ n → Xn)n≥0 → (Zn → Xn)n≥0 be a morphism in QFp/S( bX). Let J ⊆ OX be a coherent sheaf of ideals. If X is a scheme, then (ϕn) corresponds to a finite morphism of locally noetherian formal schemes ϕ : Z′ → Z. Let ϕ♯ : OZ → ϕ∗OZ′ denote the corresponding homomorphism. We say that ϕ is J-admissible if JZ ∩ ker ϕ♯ = 0 and coker ϕ♯ is annihilated by JZ. The last condition is equivalent to ϕ∗JZ′ ⊆ im ϕ♯. We say that ϕ is strongly J-admissible if JZ ∩ ker ϕ♯ = 0 and ϕ∗JZ′ = ϕ♯(JZ). If X is any algebraic stack, locally of finite type over S, then we say that (ϕn)n≥0 is J-admissible (resp. strongly J-admissible) if there exists a smooth surjection from a scheme V → X such that the adic system (ϕn)V : Z ′ Vn → ZVn induces a J-admissible (resp. strongly J-admissible) morphism ϕV : Z′ V → ZV of formal schemes. This notion does not depend on V → X. Our d´evissage methods hinge on the following two lemmas. Our first Lemma forms the analogue of Step (4) given in §0.2. It is here that we first utilize Corollary 1.19. Lemma 4.4. Let π : X → S be a morphism of algebraic stacks that is locally of finite type with quasi-compact and separated diagonal. Let J ⊆ OX be a coherent ideal and let (ϕn)n≥0 : (Z ′ n → Xn)n≥0 → (Zn → Xn)n≥0 be a J-admissible morphism in QFp/S( bX). Suppose that (Z ′ n → Xn)n≥0 and (Zn ×X V (J) → Xn)n≥0 are effectiviz- able. Then (Zn → Xn)n≥0 is effectivizable. Proof. There exists a cocartesian diagram in QFs( bX) (Corollary 1.19): n → Xn)n≥0 n ×X V (J) → Xn)n≥0 (Z ′ (Z ′ (Zn ×X V (J) → Xn)n≥0 / (Wn → Xn)n≥0, (ϕn)n≥0 where (Wn → Xn)n≥0 ∈ QFp/S(X) is effectivizable to (W → X). The universal properties show that there is a uniquely induced morphism (bn)n≥0 : (Wn → Xn)n≥0 → (Zn → Xn)n≥0 such that each bn is finite. We will show that (bn) is strongly J-admissible and that if (ϕn) is strongly J-admissible, then (bn) is an isomorphism. It follows that (Zn) is the pushout of an effectivizable diagram and is thus effectivizable. That (bn) is strongly J-admissible (resp. an isomorphism) can be verified smooth- locally on X. Thus after replacing X by an affine and noetherian scheme, we may pass from the adic systems above to noetherian formal schemes. That is, (ϕn)n≥0 is given by a finite morphism of noetherian formal schemes ϕ : (Z′ → bX) → (Z → bX). / / / THE HILBERT STACK 21 By assumption, ϕ is J-admissible (resp. strongly J-admissible), so we also know that JZ ∩ ker ϕ♯ = 0 and that ϕ∗JZ′ ⊆ im ϕ♯ (resp. ϕ∗JZ′ = ϕ♯(JZ)). Moreover, (bn)n≥0 is given by a finite morphism of noetherian formal schemes β : W → Z. That β is strongly J-admissible (resp. an isomorphism) is Zariski local on Z. Thus, we may assume that Z = Spf IA A, Z′ = Spf IA′ A′, and that ϕ is given by a finite morphism f : A → A′. Let B = A/JA×A′/JA′ A′. By Theorem 1.18, there is a natural isomorphism W ∼= Spf IA B. It remains to prove that the natural map A → B is strongly J-admissible (resp. an isomorphism). To this end we form the commutative diagram with exact rows: 0 0 / JA f / JA′ m7→(m,0) A × A′ / A/JA × A′ d1 / A′ d2 / A′/JA′ / 0 / 0, where d1(a, a′) = f (a) − a′ and d2(a, a′) = f (a) − a′. By the Snake Lemma, there is an exact sequence of A-modules: 0 / JA ∩ ker(f ) / A / B / JA′/f (JA) / 0. By hypothesis, JA ∩ ker(f ) = 0 and JA′ ⊆ f (A) (resp. JA′ = f (JA)). Thus, if ϕ is strongly J-admissible, then A → B is an isomorphism. If ϕ is only J-admissible, then A → B is injective and JB ⊆ A. Furthermore, JB ⊆ (0, JA′) and B/(0, JA′) = A/JA. It follows that A/JA → B/JB is injective so that JB = JA and A → B is strongly J- admissible. (cid:3) In order to apply Lemma 4.4 we have our next Lemma. This is the analogue of [EGA, III.5.3.4] in our situation; forming part of Step (9) in the outline given in §0.2. To prove this Lemma, we must combine the main techniques of the paper developed thus far- specifically Corollary 1.19 and Theorem 2.1. Lemma 4.5. Let X be an algebraic stack of finite type over S with quasi-compact and sep- arated diagonal. Let (Zn → Xn)n≥0 ∈ QFp/S( bX). Suppose that there is a proper and schematic morphism p : X ′ → X and an open substack U ⊆ X such that the morphism p−1(U ) → U is finite and flat. Assume, in addition, that bp∗(Zn → Xn)n≥0 ∈ QFp/S( bX ′) is effectivizable. Then there exists a coherent ideal J ⊆ OX, with supp(OX /J) equal to the complement of U in X, and a J-admissible morphism (ϕn)n≥0 : (eZn → Xn)n≥0 → (Zn → Xn)n≥0 with (eZn → Xn)n≥0 effectivizable. Proof. Let X ′′ = X ′ ×X X ′. For j = 1 and 2 let sj : X ′′ → X ′ denote the jth projection. Let q = p ◦ s1 and observe that there is a 2-morphism α : q ⇒ p ◦ s2. By hypothesis, bp∗(Zn → Xn)n≥0 admits an effectivization (Z ′ → X ′) ∈ QFp/S(X ′). By Theorem 3.2, for j = 1 and 2, there exist natural morphisms ηj : (sj )!(sj)∗Z ′ → Z ′ in QFp/S(X ′). By Theorem 3.2, there are also natural morphisms p!(ηj) : p!(sj )!(sj)∗Z ′ → p!Z ′ in QFp/S(X). Moreover, there is a natural isomorphism of functors p!(s1)! ⇒ q!. The 2-morphism α also induces a natural isomorphism of functors p!(s2) ⇒ q! and an iso- 2) in QFp/S( bX ′′), where π : X → S denotes the morphism Ψπ◦q,I(s∗ structure morphism of X. By Lemma 4.2, we obtain an isomorphism ν : s∗ 2 in QFp/S(X ′′). Let Z ′′ = s∗ 1Z ′. Combining the afforementioned isomorphisms, we obtain for j = 1 and 2 natural morphisms tj : q!Z ′′ → p!Z ′ in QFp/S(X). The morphisms tj are finite, thus the coequalizer diagram [q!Z ′′ ⇒ p!Z ′] has a colimit, eZ in QFp/S(X) (Theorem 1.10). By Theorem 3.2 there are natural isomorphisms: 1) ∼= Ψπ◦q,I (s∗ 2Z ′ 1 → s∗ 2Z ′ 1Z ′ 1Z ′ Ψπ,I(p!Z ′) ∼= bp!(Z ′ ×X ′ X ′ Ψπ,I(q!Z ′) ∼= bq!(Z ′′ ×X ′′ X ′′ n → X ′ n → X ′′ n)n≥0 ∼= bp!bp∗(Zn → Xn)n≥0 n)n≥0 ∼= bq!bq∗(Zn → Xn)n≥0. / / / / / / / / / / / / / / 22 J. HALL AND D. RYDH Moreover, the morphism bp!bp∗(Zn → Xn)n≥0 → (Zn → Xn)n≥0 coequalizes the two morphisms bq!bq∗(Zn → Xn)n≥0 ⇒ bp!bp∗(Zn → Xn)n≥0. By Corollary 1.19, there is a uniquely induced morphism (ϕn)n≥0 : (eZ ×X Xn → Xn)n≥0 → (Zn → Xn)n≥0 in QFp/S( bX). It remains to prove that (ϕn)n≥0 is J-admissible for a suitable ideal J. First, let J ⊆ OX be any coherent ideal defining the complement of U . We will show that (ϕn)n≥0 is J N -admissible for sufficiently large N . Note that the verification of this is local on X for the smooth topology. Indeed, by Corollary 1.19 and Theorem 3.2, the construction of (ϕn)n≥0 is compatible with smooth base change on X. So, we may henceforth assume that X = Spec B, where B is an R-algebra of finite type, and that (Zn → Xn)n≥0 is given by a morphism of locally noetherian formal schemes (Z → bX). We next observe that we may work Zariski locally on Z. To see this, we note that the morphism bp!bp∗Z → Z is just the Stein factorization of bp∗Z → Z (Lemma 2.3) and this is compatible with flat base change on Z (Lemma 2.4), and similarly for the morphisms bq!bq∗Z ⇒ bp!bp∗Z. Also, let eZ ∧ = eZ and let (ϕn)n≥0 be given by the finite morphism ϕ : eZ → Z. Then Theorem 1.18 1−bt♯ bt♯ 2−−−→ Obq! bq∗Z). Thus the formation of eZ is also compatible implies that OeZ = ker(O bp! bp∗Z with flat base change on Z. Consequently, we may assume that Z = Spf IA A, for some B-algebra A that is IA-adically complete. Let Z = Spec A, then the morphism bZ → bX induces a morphism of schemes Z → X. Let pZ : Z ′ → Z (resp. qZ : Z ′′ → Z) denote the pullback of p (resp. q) along Z → X. Let A′ = Γ(Z ′, OZ′) and A′′ = Γ(Z ′′, OZ′′ ) (which are both coherent A-modules because pZ and qZ are proper). Then there are two induced morphisms d1, d2 : A′ → A′′ and eA := Γ(eZ, O eZ) = ker(A′ d1−d2−−−−→ A′′). The morphism ϕ : eZ → Z induces a natural finite map f : A → eA. It remains to show that we can find an N ≫ 0 such that (J N A) ∩ ker f = 0 and J N A annihilates coker f . Now, on the complement, UZ, of the support of A/JA on Spec A, the morphism p−1 g. By finite flat descent, ϕg : Ag → eAg is an isomorphism. Since A is noetherian, we deduce immediately that the kernel and cokernel of ϕ are annihilated by J N A for some N ≫ 0. By the Artin–Rees Lemma [EGA, 0I.7.3.2.1], and by possibly increasing N , we can ensure that (J N A) ∩ ker f = 0. The result follows. (cid:3) Z (UZ ) → UZ is finite and flat. So, if g ∈ JA, then A′′ g ⊗Ag A′ g = A′ We may finally prove Theorem 4.3. g f −→ P Proof of Theorem 4.3. By Lemma 4.2, it remains to show that the functor Ψπ,I is essen- tially surjective. We divide the proof of this into a number of cases. Basic case. Let π : X → S be a morphism of schemes that is of finite type and factors as −→ S, where f is ´etale and g is projective. Let (Zn → Xn)n≥0 ∈ QFp/S( bX), X which we may regard as a quasi-finite and separated morphism of locally noetherian formal schemes (Z → bX). Then the composition Z → bX → bP is quasi-finite and proper, since P is separated and Z is S-proper. Hence, the morphism Z → bP is finite [EGA, III.4.8.11]. As g : P → S is projective, there is a finite P -scheme Z such that (bZ → bP ) ∼= (Z → bP ) in QFp/S( bP ) [EGA, III.5.4.4]. The morphism f is ´etale, thus Corollary B.5 implies that there is a unique P -morphism Z → X lifting the given P0-morphism Z0 → X0. Thus we have obtained (Z → X) ∈ QFp/S(X) with completion isomorphic to Z over bX. We conclude that the functor Ψπ,I is an equivalence of categories in this case. Quasi-compact with quasi-finite and separated diagonal case. Let π : X → S be a morphism of algebraic stacks that is of finite type with quasi-finite and separated diagonal. We now prove that the functor Ψπ,I is an equivalence by noetherian induction on the closed subsets of X. Thus, for any closed immersion i : V ֒→ X with V ( X, we may assume that the functor Ψπ◦i,I : QFp/S(V ) → QFp/S(bV ) is an equivalence. THE HILBERT STACK 23 Let (Zn → Xn)n≥0 ∈ QFp/S(X). Then there exists a 2-commutative diagram: X π S p g X ′ f P, such that g is projective, f is ´etale, X ′ is a scheme, and p is proper, schematic, surjective, and over a dense open subset U ⊆ X the morphism p−1(U ) → U is finite and flat [Ryd09, Thm. 8.9]. By the Basic Case considered above, bp∗(Zn → Xn)n≥0 ∈ QFp/S( bX ′) is effectivizable. By Lemma 4.5, there exists a coherent ideal J ⊆ OX, with supp(OX /J) equal to the complement of U in X, and a J-admissible morphism (ϕn)n≥0 : (eZn → Xn)n≥0 → (Zn → Xn)n≥0 with (eZn → Xn)n≥0 effectivizable. By noetherian induction, (Zn×X V (J) → Xn)n≥0 ∈ QFp/S( bX) is effectivizable. By Lemma 4.4, (Zn → Xn)n≥0 is effectivizable. General case. We now assume that π : X → S is as in the statement of the Theorem. It remains to prove that the functor Ψπ,I is essentially surjective. To show this, we first reduce to the case where π is quasi-compact. Let (Zn → Xn)n≥0 ∈ QFp/S( bX). Let OX denote the set of quasi-compact open substacks of X. The set OX is ordered by inclusion, and we note that {W ×X Z0}W ∈OX is an open cover of the quasi-compact topological space Z0. Thus, there is a W ∈ OX such that the canonical X-morphism W ×X Z0 → Z0 is an isomorphism. In particular, the map Z0 → X factors uniquely as Z0 → W ֒→ X. Since the open immersion W ֒→ X is ´etale, it follows that the maps Zn → X also factor compatibly through W . We can thus replace X with W and assume henceforth that X is quasi-compact. Next, we note that there is an open substack X qf ⊆ X with the property that a point x of X belongs to X qf if and only if the stabilizer at x is finite. Indeed, by Chevalley's Theorem [EGA, IV.13.1.4], the locus of points in the inertia stack IX/S → X that are isolated in their fibers is an open subset of IX/S. Intersecting this open set with the identity section e : X → IX/S gives X qf as IX/S → X is a relative group stack (the dimension of a finite type group algebraic space over a field is its local dimension at the identity). Arguing as before and applying the previous case, it now remains to show that the map Z0 → X factors through X qf. So we let IX → X (resp. IZ0 → Z0) denote the inertia stack of X (resp. Z0) over S. The morphism Z0 → X is separated and representable thus IZ0 → IX ×X Z0 is a closed immersion over Z0. By assumption, IX → X has affine fibers and IZ0 → Z0 is proper. Hence the fibers of IZ0 → Z0 are proper and affine, thus finite. By Zariski's Main Theorem [LMB, Cor. A.2.1], IZ0 → Z0 is finite, and so Z0 has finite diagonal over S (because Z0 is separated over S). We now form the 2-commutative diagram: Z0 ×X Z0 Z0 ×S Z0 Z0. The morphism Z0 → X is also quasi-finite, thus the morphism Z0 ×X Z0 → Z0 is quasi- finite. The morphism Z0 ×S Z0 → Z0 has finite diagonal, thus Z0 ×X Z0 → Z0 ×S Z0 is quasi-finite. We may now conclude that if z is a point of Z0, then its image in X has finite stabilizer. Thus Z0 factors through X qf and we deduce the result. (cid:3) Remark 4.6. It is possible that Theorem 4.3 extends to any stack X with locally quasi- finite diagonal, that is, without requiring that ∆X is separated and quasi-compact. For such X, we consider the full subcategory LQFp/S( bX/X0) ⊆ LQF( bX/X0) consisting of systems of locally quasi-finite representable morphisms (Zn → X)n≥0 such that the composition Z0 → X → S is proper. o o o o / / 0 0 24 J. HALL AND D. RYDH The generalization of Lemma 4.2, that Ψπ,I : LQFp/S(X) −→ LQFp/S( bX) is fully faithful, follows as for stacks with separated diagonals, but using Theorem 4.3 instead of [Ols05, Thm. 1.4]. The corresponding generalizations of Lemmas 4.4 and 4.5 would follow if we extend the results of Sections 1 and 2 to include certain non-separated quasi- finite morphisms. This is probably not too difficult to accomplish, although one may have to develop a theory of non-separated formal algebraic spaces. Finally, to obtain the generalization of Theorem 4.3 we would also need a suitable Chow lemma for stacks with locally quasi-finite diagonals. It would suffice to find a proper generically finite and flat morphism p : X ′ → X together with an ´etale, not necessarily representable, morphism f : X ′ → P with P → S projective. Whether such a proper covering exists is not clear to the authors. 5. ALGEBRAICITY OF THE HILBERT STACK Before we get to the proof Theorem 2, we will require an analysis of some spaces of sections. Let T be a scheme and let s : V ′ → V be a representable morphism of algebraic T -stacks. Define SecT (V ′/V ) to be the sheaf that takes a T -scheme W to the set of sections of the morphism V ′ ×T W → V ×T W . We require an improvement of [Ols06, Prop. 5.10] and [Lie06, Lem. 2.10], which we prove using [Hal12a, Thm. D]. Proposition 5.1. Let T be a scheme and let p : Z → T be a morphism of algebraic stacks that is proper, flat, and of finite presentation. Let s : Q → Z be a quasi-finite, separated, finitely presented, and representable morphism. Then the T -sheaf SecT (Q/Z) = p∗Q is represented by a quasi-affine T -scheme which is of finite presentation. Proof. By a standard limit argument [Ryd09, Prop. B.3], we can assume that T is noether- ian. By Zariski's Main Theorem [LMB, Thm. 16.5(ii)], there is a finite morphism Q → Z and an open immersion Q ֒→ Q over Z. In particular, we see that there is a natural trans- formation of T -sheaves SecT (Q/Z) → SecT (Q/Z) which is represented by open immer- sions. Hence, we may assume for the remainder that the morphism s : Q → Z is finite. Next, observe that SecT (Q/Z) is a subfunctor of the T -presheaf HomOZ /T (s∗OQ, OZ). By [Hal12a, Thm. D], the sheaf HomOZ /T (s∗OQ, OZ) is represented by a scheme that is affine and of finite type over T . Similarly, HomOZ /T (OZ, OZ) and HomOZ /T (s∗OQ ⊗OZ s∗OQ, OZ) are affine. It follows that SecT (Q/Z) ֒→ HomOZ /T (s∗OQ, OZ) is represented by closed immersions and the result follows. (cid:3) Corollary 5.2. Let T be a scheme and let X → T be a morphism of algebraic stacks that is locally of finite presentation. Suppose that we have quasi-finite, separated and representable morphisms si : Zi → X for i = 1, 2 such that Z1 and Z2 are proper, flat and of finite presentation over T . Then the functor HomX (Z1, Z2) on Sch/T , given by T ′ 7→ HomXT ′(cid:0)(Z1)T ′ , (Z2)T ′(cid:1), is represented by a scheme that is quasi-affine over T . In particular, the open subfunctor IsomX (Z1, Z2) ⊂ HomX (Z1, Z2) parameterizing isomorphisms is represented by a scheme that is quasi-affine over T . Proof. Note that HomX (Z1, Z2) = SecT ((Z1 ×X Z2)/Z1). Thus, by Proposition 5.1 the functor HomX (Z1, Z2) has the asserted properties. (cid:3) Proof of Theorem 2. The results of [Hal12b, §9], together with Theorem 4.3, show that the stack HSX/S is algebraic and locally of finite presentation over S. We now apply Corollary 5.2 to obtain the asserted separation property. (cid:3) APPENDIX A. COHERENT COHOMOLOGY OF FORMAL SCHEMES In this appendix, we prove that cohomology commute with flat base change of formal schemes. This is well-known for schemes, but we could not find a suitable reference for formal schemes. Note that in the adic case, this result follows from the more general result [AJL99, Prop. 7.2(b)]. Proposition A.1. Consider a cartesian diagram of locally noetherian formal schemes: THE HILBERT STACK 25 X′ p′ ′ X Y′ p / Y, where is proper and p is flat. Let F be a coherent OX-module, then for any q ≥ 0 the base change morphism p∗Rq∗F → Rq′ ∗p′∗F is a topological isomorphism. Proof. By [EGA, III.3.4.5.1], the statement is Zariski local on Y and Y′ so we may assume that Y = Spf I R for some I-adic noetherian ring R, Y′ = Spf I ′ R′ for some I ′-adic noetherian ring R′, IR′ ⊆ I ′, and R → R′ is a flat morphism of rings. It suffices to prove that the morphism H q(X, F)b⊗RR′ → H q(X′, p′∗F) is a topological isomorphism. Note that p factors as Spf I ′ R′ r−→ Spf IR′ R′ q −→ Spf I R. Since q is adic, we are reduced to proving the Proposition when p is also adic or when R = R′. Note that cohomology can be shown to commute with adic flat base change by minor modifications to the statements and arguments of [EGA, 0III.13.7.8], combined with [EGA, 0III.13.7.7] and [EGA, III.3.4.3] to deal with flat base change instead of localizations. It remains to prove the Proposition when R = R′. For each k, l ≥ 0, set Rk,l = R/(I k+1, I ′l+1), Rk = R/I k+1, and R′ l = R/I ′l+1. l (because I ⊆ I ′). Also, for each fixed k, the noetherian If k ≥ l, then Rk,l = R′ Rk,l is a topological ring Rk is I ′-adically complete, and the natural map Rk → lim←−l isomorphism. Let Xk,l = X ⊗R Rk,l, Xk = X ⊗R Rk (which we view as noetherian schemes), Fk,l = F ⊗R Rk,l, and Fk = F⊗RRk. By [EGA, III.4.1.7], for each k ≥ 0, there are natural induced topological isomorphisms: H q(Xk, Fk)∧ ∼= lim←− l H q(Xk,l, Fk,l), where H q(Xk, Fk)∧ denotes the completion of the Rk-module H q(Xk, Fk) with respect to the I ′-adic topology. Note, however, that Xk → Spec Rk is proper, thus H q(Xk, Fk) is a coherent Rk-module [EGA, III.3.2.1]. Consequently, H q(Xk, Fk) is I ′-adically complete [EGA, 0I.7.3.6], and we have topological isomorphisms: H q(Xk, Fk) ∼= lim←− l H q(Xk,l, Fk,l). l = X ⊗R R′ l and F ′ l = F ⊗R R′ l. Applying the functor lim←−k Let X ′ isomorphism above we obtain natural isomorphisms of R′-modules: lim←− H q(Xk,l, Fk,l) ∼= lim←− l By [EGA, III.3.4.4], we have naturally induced topological isomorphisms: H q(Xk,l, Fk,l) ∼= lim←− H q(Xk, Fk) ∼= lim←− lim←− k k lim←− l k l to both sides of the H q(X ′ l , F ′ l ). H q(X, F) → lim←− k and we deduce the result. H q(Xk, Fk) and H q(X′, F′) → lim←− l H q(X ′ l , F ′ l ), (cid:3) As a Corollary, we can prove Zariski's Connectedness Theorem for formal schemes, using the same argument as [EGA, III.4.3.2 and 4.3.4]. Corollary A.2. Let : X → Y be a proper and Stein morphism of locally noetherian formal schemes. Then has geometrically connected fibers. Proof. Use [EGA, 0III.10.3.1] and Proposition A.1 to reduce to Y = Spf IR R where R is a complete local ring with maximal ideal I and algebraically closed residue field R/I. Then X equals the unique fiber of . Finally observe that ∗OX = OY is local so X is connected. (cid:3) / / / 26 J. HALL AND D. RYDH APPENDIX B. HENSELIAN PAIRS In this appendix, we address some simple results about ´etale morphisms that we could not locate in the literature. As an application, we strengthen a well-known result about the stability of henselian pairs under proper morphisms. For an algebraic space S, let Et(S) denote the category of morphisms of algebraic spaces V → S that are ´etale. Let S´et denote the small ´etale site of S: this has underlying category Et(S) and the coverings are jointly surjective families of morphisms. It is well- known that the natural functor: Et(S) → Sh(S´et), (V → S) 7→ HomS(−, V ) is an equivalence of categories [Mil80, V.1.5], where Sh(S´et) denotes category of S´et- sheaves of sets (even if S is a scheme, we need to allow V to be an algebraic space). If f : X → S is a morphism of algebraic spaces, there are thus natural functors f ∗ : Et(S) → Et(X) and f∗ : Et(X) → Et(S), with f ∗ left adjoint to f∗. We define Ett/S(X) to be the full subcategory of Et(X) with objects W → X such that for every geometric point s of S and every connected component Z of Ws, the induced morphism Z → Xs is an open and closed immersion. We now have the main technical result of this appendix. Proposition B.1. Let f : X → S be a proper and surjective morphism of schemes with geometrically connected fibers. Then the functor f ∗ : Et(S) → Et(X) is fully faithful with image Ett/S(X). Proof. To show that f ∗ is fully faithful, it is sufficient to prove that if V ∈ Et(S), then the natural map V → f∗f ∗V is an isomorphism. This may be verified over the geometric points of s of S. Note, however, that the functor f ∗ is trivially compatible with arbitrary base change on S. Since f is proper, it is a basic case of the Proper Base Change Theorem [SGA4, XII.5.1(i)] that f∗ is also compatible with arbitrary base change on S. Thus, it suffices to prove the result where S = Spec k with k an algebraically closed field. Since f is surjective, this is trivial. To classify the image of f ∗, it remains to show that if W ∈ Ett/S(X), then the natural map f ∗f∗W → W is an isomorphism. This may be verified at geometric points x of X. The remarks above about base change also apply here, so we are again reduced to the situation where S = Spec k with k an algebraically closed field. In this case, the connected components of W are all open and closed subschemes of X. Since X is geometrically connected, we conclude that W = ∐w∈π0(W )X. It remains to show that the natural map ∐w∈π0(W )S → f∗W is an isomorphism. This may be checked on global sections (since k is algebraically closed) and there we have a natural bijection π0(W ) → HomX (X, W ). (cid:3) Remark B.2. Let f : X → S be a proper morphism of schemes with geometrically con- nected fibers and let γ : V1 → V2 be a morphism in Ett/S(X). Then γ is an open im- mersion (resp. an open and closed immersion, resp. separated) if and only if the induced morphism f∗γ : f∗V1 → f∗V2 is such. The first is an easy consequence of the fact that open immersions are categorical monomorphisms in Et(X), and since f∗ preserves products, it preserves monomorphisms. The other two cases are even easier. For a scheme S, let OC(S) denote its set of open and closed subsets. Note that OC(X) ⊆ Ett/S(X). A henselian pair (S, S0) consists of a scheme S and a closed immersion S0 ֒→ S such that for any finite morphism g : S′ → S, the natural map OC(S′) → OC(S′ ×S S0) is bijective [EGA, IV.18.5.5]. Example B.3. Note that if B is a noetherian ring, separated and complete for the topol- ogy defined by an ideal I ⊆ B, then (Spec B, Spec B/I) is a henselian pair [EGA, IV.18.5.16(ii)]. THE HILBERT STACK 27 We now obtain the following improvement of [SGA4, XII.5.5] and [EGA, IV.18.5.19], where it is proved for henselian pairs of the form (Spec A, {m}), where A is a henselian local ring and m is the maximal ideal of A. Corollary B.4. Let (S, S0) be a henselian pair and let f : X → S be a proper morphism of noetherian schemes. Then (X, X ×S S0) is a henselian pair. Proof. Let X0 = X ×S S0 and take f0 : X0 → S0 to be the induced morphism. It is If X → eX → S denotes sufficient to prove that OC(X) → OC(X0) is bijective. the Stein factorization of f , then Proposition B.1 and Remark B.2 implies that we have a bijection OC( eX) → OC(X). Since eX → S is finite, ( eX, eX ×S S0) is a henselian pair f0−→ [EGA, IV.18.5.6]. A similar analysis applies to the Stein factorization of f0, X0 → eX0 S0, where we also obtain a bijection OC( eX0) → OC(X0). Denote by g : X0 → eX ×S S0 and g : eX ×S S0 → S0 the induced morphisms. There are now natural maps of coherent OS0-algebras: ∼= ( f0)∗O eX0 . g∗O eX×SS0 → g∗g∗OX0 ∼= (f0)∗OX0 Whence we obtain a finite S0-morphism h : eX0 → eX ×S S0. But the morphisms X0 → eX0 and X0 → eX ×S S0 are both surjective with geometrically connected fibers, thus h also is surjective with geometrically connected fibers. Consequently, h is a universal homeomorphism [EGA, IV.18.12.11] so that OC( eX ×S S0) → OC( eX0) is bijective. The result follows. (cid:3) Note that one of the strengths of henselian pairs is that they enable the computation of sections of sheaves. Indeed, if (S, S0) is a henselian pair, with S quasi-compact and quasi- separated, and V ∈ Et(S), then the natural map HomS(S, V ) → HomS0(S0, V ×S S0) is bijective [SGA4, XII.6.5(i)]. Corollary B.5. Let (S, S0) be a henselian pair where S is a noetherian scheme. Fix a scheme Y over S. Let Z and X be schemes over Y such that X → Y is ´etale and Z → S is proper. Then the natural map: HomY (Z, X) → HomY (Z ×S S0, X) is bijective. Proof. The map in the statement is identified with the natural map: HomZ (Z, V ) → HomZ×S S0 (Z ×S S0, V ×S S0) where V = X ×Y Z is ´etale over Z. Since (Z, Z ×S S0) is a henselian pair (Corollary B.4), the result follows. (cid:3) Remark B.6. Note that [SGA4, XII.6.5(i)] and [SGA4, XII.5.1(i)] are quite elementary. The first result follows from the following facts: (i) every sheaf is a direct limit of con- structible sheaves [SGA4, IX.2.7.2], and (ii) every constructible sheaf embeds in a product of push-forwards of constant sheaves along finite morphisms [SGA4, IX.2.14]. To prove the second result, one reduces to the case where S is henselian, then to X = Pn S using a suitable Chow lemma, and finally to S noetherian and henselian using approximation. Then one concludes using [SGA4, XII.5.1(i)] and the Stein factorization [SGA4, XII.5.8]. For completeness, let us mention some generalizations of the results in this section to algebraic spaces and stacks. The first result, [SGA4, XII.6.5(i)], is easily extended to noetherian stacks and to quasi-compact and quasi-separated Deligne–Mumford stacks. The second result, [SGA4, XII.5.1(i)], then follows for noetherian stacks using the Chow lemma [Ols05, Thm. 1.1] and for quasi-compact and quasi-separated Deligne–Mumford stacks using the Chow lemma [Ryd09, Thm. B]. Proposition B.1, Corollary B.4 and Corol- lary B.5 thus follow for such stacks. The noetherian assumption in Corollary B.4 can be removed for Deligne–Mumford stacks using the fact that there exists Stein factorizations 28 J. HALL AND D. RYDH for proper morphisms of non-noetherian stacks, although one gets an integral morphism instead of a finite morphism. However, this is not a problem as integral morphisms can be approximated by finite morphisms [Ryd09, Thm. A]. REFERENCES [AJL99] L. Alonso Tarr´ıo, A. Jerem´ıas L´opez, and J. Lipman, Studies in duality on Noetherian formal schemes and non-Noetherian ordinary schemes, Contemporary Mathematics, vol. 244, American Mathematical Society, Providence, RI, 1999. [AK10] V. Alexeev and A. Knutson, Complete moduli spaces of branchvarieties, J. Reine Angew. Math. 639 (2010), 39–71. [Aok06a] M. Aoki, Hom stacks, Manuscripta Math. 119 (2006), no. 1, 37–56. [Aok06b] [Art69] M. Artin, Algebraization of formal moduli. I, Global Analysis (Papers in Honor of K. Kodaira), Univ. , Hom stacks: erratum, Manuscripta Math. 121 (2006), no. 1, 135. Tokyo Press, Tokyo, 1969, pp. 21–71. , Versal deformations and algebraic stacks, Invent. Math. 27 (1974), 165–189. [Art74] [Con] [DM69] P. Deligne and D. Mumford, The irreducibility of the space of curves of given genus, Inst. Hautes B. Conrad, Formal GAGA for Artin stacks, Available on homepage. ´Etudes Sci. Publ. Math. (1969), no. 36, 75–109. [EGA] A. Grothendieck, ´El´ements de g´eom´etrie alg´ebrique, I.H.E.S. Publ. Math. 4, 8, 11, 17, 20, 24, 28, 32 (1960, 1961, 1961, 1963, 1964, 1965, 1966, 1967). [FGA] , Fondements de la g´eom´etrie alg´ebrique. Extraits du S´eminaire Bourbaki, 1957–1962, Secr´etariat math´ematique, Paris, 1962. [FGI+05] B. Fantechi, L. Gottsche, L. Illusie, S. L. Kleiman, N. Nitsure, and A. Vistoli, Fundamental alge- braic geometry, Mathematical Surveys and Monographs, vol. 123, American Mathematical Society, Providence, RI, 2005, Grothendieck's FGA explained. [Hal12a] J. Hall, Cohomology and base change for algebraic stacks, Preprint, June 2012, arXiv:1206.4179v2. [Hal12b] [Has03] B. Hassett, Moduli spaces of weighted pointed stable curves, Adv. Math. 173 (2003), no. 2, 316–352. [Høn04] M. Hønsen, A compact moduli space parameterizing Cohen-Macaulay curves in projective space, , Openness of versality via coherent functors, Preprint, June 2012, arXiv:1206.4182v2. Ph.D. thesis, MIT, 2004. [Knu71] D. Knutson, Algebraic spaces, Lecture Notes in Mathematics, Vol. 203, Springer-Verlag, Berlin, 1971. J. Koll´ar, Quotients by finite equivalence relations, Current Developments in Algebraic Geometry, [Kol11] Math. Sci. Res. Inst. Publ., vol. 59, Cambridge Univ. Press, Cambridge, 2011, pp. 227–256. [Lie06] M. Lieblich, Remarks on the stack of coherent algebras, Int. Math. Res. Not. (2006), Art. ID 75273, 12. [LMB] G. Laumon and L. Moret-Bailly, Champs alg´ebriques, Ergebnisse der Mathematik und ihrer Grenzge- biete. 3. Folge., vol. 39, Springer-Verlag, Berlin, 2000. [LS08] C. Lundkvist and R. Skjelnes, Non-effective deformations of Grothendieck's Hilbert functor, Math. Z. [Mil80] 258 (2008), no. 3, 513–519. J. S. Milne, ´Etale cohomology, Princeton Mathematical Series, vol. 33, Princeton University Press, Princeton, N.J., 1980. [Ols05] M. Olsson, On proper coverings of Artin stacks, Adv. Math. 198 (2005), no. 1, 93–106. [Ols06] [OS03] M. Olsson and J. M. Starr, Quot functors for Deligne-Mumford stacks, Comm. Algebra 31 (2003), , Hom-stacks and restriction of scalars, Duke Math. J. 134 (2006), no. 1, 139–164. no. 8, 4069–4096, Special issue in honor of Steven L. Kleiman. [RG71] M. Raynaud and L. Gruson, Crit`eres de platitude et de projectivit´e. Techniques de "platification" d'un module, Invent. Math. 13 (1971), 1–89. [Rom11] M. Romagny, Composantes connexes et irr´eductibles en familles, Manuscripta Math. 136 (2011), no. 1- 2, 1–32. [Ryd07] D. Rydh, Existence and properties of geometric quotients, Aug 2007, arXiv:0708.3333v2, To appear in J. Alg. Geom. [Ryd09] , Noetherian approximation of algebraic spaces and stacks, Preprint, April 2009, arXiv:0904.0227v3. [Ryd11] , Representability of Hilbert schemes and Hilbert stacks of points, Comm. Algebra 39 (2011), no. 7, 2632–2646. [Sch91] D. Schubert, A new compactification of the moduli space of curves, Compositio Math. 78 (1991), no. 3, 297–313. [SGA4] M. Artin, A. Grothendieck, J. L. Verdier, P. Deligne, and B. Saint-Donat, Th´eorie des topos et coho- mologie ´etale des sch´emas, Lecture Notes in Mathematics, Vol. 305, Springer-Verlag, 1973. [Smy13] D. I. Smyth, Towards a classification of modular compactifications of Mg,n, Inventiones mathematicae 192 (2013), no. 2, 459–503. THE HILBERT STACK 29 [Vis91] A. Vistoli, The Hilbert stack and the theory of moduli of families, Geometry Seminars, 1988–1991 (Italian) (Bologna, 1988–1991), Univ. Stud. Bologna, Bologna, 1991, pp. 175–181. DEPARTMENT OF MATHEMATICS, KTH ROYAL INSTITUTE OF TECHNOLOGY, SE-100 44 STOCKHOLM, SWEDEN E-mail address: [email protected] DEPARTMENT OF MATHEMATICS, KTH ROYAL INSTITUTE OF TECHNOLOGY, SE-100 44 STOCKHOLM, SWEDEN E-mail address: [email protected]
1812.09918	1	1812	2018-12-24T13:26:08	Basics of jet modules and algebraic linear differential operators	[ "math.AG" ]	In this paper, we collect the fundamental basic properties of jet modules in algebraic geometry and related properties of differential operators. We claim no originality but we want to provide a reference work for own research and the research of other people.	math.AG	math	Basics of jet modules and algebraic linear differential operators December 27, 2018 Stefan Gunther Abstract In this paper, we collect the fundamental basic properties of jet- modules in algebraic geometry and related properties of differential operators. We claim no originality but we want to provide a reference work for own research and the research of other people. Contents 1 Notation and Conventions 2 Introduction 1 3 3 Fundamentals of jet-modules and differential operators in al- 3 gebraic geometry 3 3.1 Definition of a linear partial differential operator . . . . . . . . 4 3.2 Supplements to the calculus of jet modules . . . . . . . . . . . 3.3 Fundamental properties of the jet-modules . . . . . . . . . . . 4 3.4 Comparison with the C∞-category . . . . . . . . . . . . . . . . 13 3.5 The global case . . . . . . . . . . . . . . . . . . . . . . . . . . 14 3.6 Basic properties of differential operators . . . . . . . . . . . . 23 3.7 Existence of differential operators in the affine case . . . . . . 26 1 Notation and Conventions Remark 1.1 This is a slightly advanced introduction to the theory of jet modules in algebraic geometry. For the elementary facts see e.g. [5]. Convention 1 By N we denote the natural numbers, by N0 the set of non- negative integers. 1 We use multi index notation: if x1, ..., xn is a set of variables, we denote xm : xm1 1 · xm2 2 . . . xmn n where m := (m1, m2, . . . , mn) is a multiindex of lenght n. By m we denote the number m1 + ...mn . The partial derivatives of a function f (x1, ..., xm) in the variables xi we denote by ∂m/∂xm(f (x1, ..., xm)) . Notation 1 Let X −→ S be a morphism of schemes. By Ω(1)(X/S) we denote the usual sheaf of Kahler differentials. We use this notation, because there exist higher Kahler differential modules Ω(n)(X/S) see [3]. The direct Ω(n)(X/S) is a graded sheaf of OX -algebras whose SpecX is the sum Ln∈N0 well known space of relative arcs (arcs in fibre direction). Notation 2 Let k −→ A be a homomorphism of commutative rings. By IA/k we denote the ideal in the ring A ⊗k A which is the kernel of the mul- tiplication map µ : A ⊗k A −→ A . By pi : A −→ A ⊗k A, i = 1, 2 we denote the maps p1 : a 7→ a ⊗ 1 and p2 : a 7→ 1 ⊗ a and likewise for the residue rings (jet algebras) A ⊗k A/I N +1 A/k . Notation 3 Let k −→ A be a homomorphism of commutative rings and let M be an A-module. For each N ∈ N0 ∪ {N} we denote the N th jet-module of M relative to k by J N (M/k) which is by definition the module J N (M/k) := A ⊗k M/I N +1 A/k · (A ⊗k M) J N(M/k) := \A ⊗k M IA/k if N ∈ N0 and , which is the completed module with respect to the diagonal ideal IA/k . If M = A , this is a ring, which has two A-algebra structures in the obvious way. J N (M/k) is an J N (A/k)-module in a canonical way. The derivation M −→ J N (M/k), m 7→ 1 ⊗ m M/k . is denoted by dN Because of Lemma 3.2 proven in section 2.1, it is called the universal deriva- tion for the A-module M relative to k . If M = A , for a ∈ A we denote by d1a the element 1 ⊗ a−a ⊗ 1 ∈ J N (A/k) . Thus dN If X −→ S is a morphism of schemes an F is a quasi coherent OX-module, we denote the N th-jet module by J N (F /S) with universal derivation A/k(a) = a + d1(a). dN F /S : F −→ J N (F /S). Observe that in [2][EGAIV,chapter 16.3-16.4, pp.14-27], these are called the bundles of higher order principal parts. If F = OX we denote J N (X/S) := SpecX J N (X/S) pX−→ X the associated affine bundle over X which is some kind of higher order relative tangent bundle. 2 2 Introduction The calculus of jet-modules and jet bundles in algebraic geometry is basic for understanding linear partial differential operators and for a given extension of commutative rings k −→ A and an A-module M, the N th jet module J N (M/k) provides infinitesimal information about the A-module M. If M = A , and A is a local ring, essentially of finite type over a base field k, the N th jet-algebra J N (A/k) provides further information about the singularity (A, m, k). This algebra is again a k-algebra, essentially of finite type over k, and, e.g., the Hilbert-Samuel polynom of this algebra gives higher order information about the singularity (A, m, k) . So, studying jet modules can be very fruitful and in this paper we want to give basic elementary properties, e.g. generalizing properties of the classical module of Kahler differentials. We claim no originality but want to collect some basic facts that do not occur in the basic textbooks as [2][EGAIV,chapter 16.3-16.4, pp.14-27] in order to provide a reference for further own research and the research of other people. 3 Fundamentals of jet-modules and differen- tial operators in algebraic geometry 3.1 Definition of a linear partial differential operator Recall that if k is a field and An partial linear differential operator of order N, k is affine n-space over k, then a homogenous D : mM i=1 k[x1, ..., xn]ei −→ mM i=1 k[x1, ..., xn]ei corresponds to a k[x1, ..., xn]-linear map eD : (k[x1, ..., xn][d1x1, ..., d1xn]/(d1x1, ..., d1xn)N +1)⊕m −→ k[x1, ..., xn]⊕m under the natural correspondence eD 7→ eD ◦ (dN k /k is the AN N-truncated Taylor series expansion, k /k)⊕m where dN An k[x1, ..., xn] −→ k[x1, ..., xn][d1x1, ..., d1xn]/(d1x1, ..., d1xn)N +1, sending xi to xi + d1xi . This is a standard calculation. The k-algebra k[x][d1x]/(d1xN +1) is the N th jet module J N (k[x]/k) of k[x]/k and is a k[x]-algebra. The inverse limit J N(k[x]/k) := proj lim n∈N J N (k[x]/k) = k[x1, ..., xn][[d1x1, ..., d1xn]] ∼= k[x1, ..., xn]b⊗k[x1, ..., xn] is the universal jetalgebra of k[x] where the last expression is the tensor product completed with respect to the ideal Ik[x]/k which is the kernel of the algebra multiplication map. Over k[x] each projective module is free, so 3 the standard definition from the C∞-case carries over to the algebraic case. For k = R, C these are just the linear partial differential operators with polynomial coefficients. Via the above corresondence each partial differential operator corresponds to an A-linear map eD : J N (A⊕m/k) −→ A⊕m (A = k[x]). Since the formation of the jet modules commutes with localization, they give rise to a coherent sheaf on An and we define the algebraic differential opera- tors for a free A = Γ(U, OAn)-module A⊕m for U ⊂ An k a Zariski open subset to be the A-linear homomorphisms Γ(U, J N (k[x]/k))⊕m −→ A⊕m . In particular, if U = Spec k[x]f for some polynomial f , this definition gives linear partial differential operators with rational function coefficients g f n . 3.2 Supplements to the calculus of jet modules Definition 3.1 Let k −→ A be a homomorphism of commutative rings, let N ∈ N0 ∪ {N}, M be an A-module and Q be a J N (A/k)-module. A filtered derivation tM : M −→ Q is an A-linear homomorphism where Q is given the A-module structure via the second factor A −→ A ⊗k A/I N +1 A/k . The notation filtered derivation, introduced in [3], is used because for each a ∈ A, m ∈ M, t(am) − am ∈ IA/k · Q . Lemma 3.2 Let A −→ B be a homomorphism of rings, M be a B-module, and Q be a J N (B/A) -module and t : M −→ Q be a B-linear map with respect to the second B-module structure on J N (B/A), i.e., a filtered deriva- tion. Then, there is a unique homomorphism of J N (B/A) -modules φ : J N (M/A) −→ Q such that t = φ ◦ dN M/A. Proof: We have J N (B/A) = B ⊗A B/I N +1 p1, p2 : B −→ J N (B/A) . Then, by definition B/A and natural homomorphisms J N (M/A) = M ⊗B,p2 J N (B/A). The statement reduces to the easy fact, that, given a homomorphism of rings k −→ l, given a k-module Mk and an l-module Ml, and a k-linear homomorphism Mk −→ Ml , there is a unique l-linear homomorphism Mk ⊗k l −→ Ml which follows by the adjunction of restriction and extension of scalars. (cid:4) 3.3 Fundamental properties of the jet-modules Recall e.g. from [3], that if k −→ A is a k-algebra with presentation A ∼= k[xii ∈ I]/(fjj ∈ J), then the N th jet algebra J N (A/k) possesses the presentation J N (A/k) := k[xi, d1xii ∈ I]/(I N +1 A/k + (fj, d1fjj ∈ J)). 4 This can also be taken as the definition of the N th-jet-algebra. One then has to show that this is independend of the choosen presentation for A. We have the following easy consequences. Lemma 3.3 (Base change 0) Let k −→ A and k −→ A′ be homomorphisms of commutative rings. Then for each N ∈ N0 ∪ {N} , there is an isomorphism J N (A ⊗k A′/A′) ∼=−→ J N (A/k) ⊗A A′. Proof: This follows from the fact, that if A = k[xii ∈ I]/(fjj ∈ J) is a presentation of A/k, then A ⊗k A′ = A′[xii ∈ I]/(fjj ∈ J) is a presentation for A ⊗k A′/A′ and the corresponding presentation of the jet-algebras. This isomorphism can be made canonical by observing that A ⊗k A′ dN A/k⊗kidA′ −→ J N (A/k) ⊗k A′ −→ J N (A/k) ⊗A A′ is a filtered module derivation which by the universal property of the jet- modules induces a canonical isomorphism J N (A⊗k A′/A′) ∼= J N (A/k)⊗A A′. (cid:4) Corollary 3.4 If A is a finitely generated k-algebra, then for each N ∈ N0 , J N (A/k) is a finitely generated A-algebra. Proof: (cid:4) Proposition 3.5 Let k −→ A be a homomorphism of noetherian rings that makes A a smooth, finite type k-algebra. Let M be a finitely generated projective A-module. Then for each N ∈ N0 , J N (M/k) is a projective A- module. Proof: First, consider the case M = A . We prove the result by induction on N ∈ N0 . For N = 0, J 0(A/k) ∼= A the result is trivially true. From the jet-bundle exact sequence jN A/k, 0 −→ I N A/k/I N +1 A/k −→ J N (A/k) −→ J N −1(A/k) −→ 0, suppose we know that J N −1(A/k) is a projective A-module. Since A is a smooth k-algebra, the diagonal ideal IA/k corresponds to the regular embed- ding X = Spec A ֒→ X ×Spec k X. In this case, I N A/k/I N +1 A/k ∼= (IA/k/I 2 A/k)⊗sN ∼= Ω(1)(A/k)⊗sN which is well known to be a projective A-module. The induction step is then complete by observing that an extension of projective A-modules is again a 5 projective A-module. For the general case, if M is a free A-module, the claim is true since taking jet-modules commute with taking direct sums. For arbitrary projective M, since by Corollary 3.18 taking jet modules com- mute with Zariski localizations, choosing a Zariski-open covering Spec A = Si∈I Spec Ai such that M is free on Spec Ai , we know that J N (M/k) is Zariski-locally a projective A-module, hence a projective A-module. We now prove some fundamental properties. (cid:4) Proposition 3.6 (Exterior products I) Let k −→ A and k −→ B be ho- momorphisms of commutative rings. Then, there is a canonical isomorphism γA,B : J N(A/k) ⊗k J N(B/k) ∼= J N(A ⊗k B/k). Proof: This follows from the explicite presentations of the jet modules for a given presentation of A and B, respectively (see [3][chapter 6.5, pp. 101- 119]). Choose N ∈ N0 . If A = k[xii ∈ I]/(fjj ∈ J) and B = k[ykk ∈ K]/(gll ∈ L) are presentations for A and B, then A ⊗k B = k[xi, yk i ∈ I, k ∈ K]/(fj, gl j ∈ J, l ∈ L) is a presentation for A ⊗k B and J N (k[xi i ∈ I]/(fj j ∈ J)/k) ⊗k J N (k[yk k ∈ K]/(gl l ∈ L)/k) ∼= k[xi, d1xii ∈ I]/((fj, d1fj j ∈ J) + I N +1 A/k ) ⊗kk[yk, d1ykk ∈ K]/((gl, d1gl l ∈ L) + I N +1 B/k ) and J N (A ⊗k B/k) ∼= k[xi, yk, d1xi, d1yki ∈ I, k ∈ K]/((fj, d1fj, gl, d1glk ∈ K, l ∈ L) + I N +1 A⊗kB/k) We have the identity IA⊗kB/k = (IA/k + IB/k) which follows from the well known fact that the diagonal ideal is generated by all 1⊗a−a⊗1 = d1a . Thus IA/k is generated by all 1 ⊗ xi − xi ⊗ 1, IB/k is generated by all 1 ⊗ yj − yj ⊗ 1. Obviously, we have an inclusion (I N +1 A/k + I N +1 B/k ) · (A ⊗k B) ⊆ I N +1 A⊗kB. Conversely, there is an inlcusion I 2N +1 A⊗kB/k ⊆ (I N +1 A/k + I N +1 B/k ) · (A ⊗k B) because 2N + 1-fold products of elements xi ⊗ 1 − 1 ⊗ xi and yj ⊗ 1 − 1 ⊗ yj must either contain an N + 1 fold product of the xi ⊗ 1 − 1 ⊗ xi or an N + 1 fold product of the yj ⊗ 1 − 1 ⊗ yj . It follows, taking the projective limit over the natural homomorphisms γN A,B : J N (A/k) ⊗k J N (B/k) −→ J N (A ⊗k B/k), 6 we get an isomorphism γA,B : J N(A/k) ⊗k J N(B/k) ∼=−→ J N(A ⊗k B/k). (cid:4) Lemma 3.7 Let k −→ A be a homomorphism of commutative rings. Then the functor J N (−/k) : (A − Mod) −→ (J N (A/k) − Mod), M 7→ J N (M/k) is right exact. Proof: This follows from the functorial isomorphism J N (M/k) ∼= M ⊗A,p2 J N (A/k) and the right exactness of the tensor product. (cid:4) Corollary 3.8 With the previous notation, if A⊕J φ −→ A⊕I ։ M is a presentation for the A-module M, with φ given by the matrix (aij), then J N (M/k) is given by the presentation J N (A/k)⊕J J N (φ/k) −→ J N (A/k)⊕I ։ J N (M/k), where J N (φ/k) is given by the matrix (aij + d1aij) . Proof: This follows from the easy fact, that the functor J N (−/k) com- mutes with direct sums. (cid:4) Proposition 3.9 (Exterior Products II) Let k −→ A and k −→ B be homomorphisms of commutative rings. Let M be an A-module and N be a B-module. Then, there is a canonical isomorphism γM,N : J N(M/k) ⊗k J N(N/k) ∼= J N(M ⊗k N/k). Proof: The standard arguement shows that there is a canonical transfor- mation of bi-functors γM,N : J N(M ⊗k N/k) −→ J N(M/k) ⊗k J N(N/k) that comes from the fact the the tensor product dN N/k is a filtered derivation and the universal representing property of the jet-modules. If M and N are free A- and B-modules respectively, the fact that γM,N is an isomorphism, follows from (Exterior products I) and the fact that the jet- modules commute with direct sums. Then both sides are right exact in each variable M, N and choosing free presentations of M and N respectively, the claim follows by the five lemma. (cid:4) M/k ⊗k dN 7 Proposition 3.10 (Base change I) Let A −→ B, A −→ A′ be homo- morphisms of commutative rings and M be a B-module. Then, for each N ∈ N0 ∪ {N} , there is a canonical isomorphism βM : J N (M ⊗A A′/A′) ∼= J N (M/A) ⊗A A′. Proof: Both sides are right exact functors from B −Mod to J N (B ⊗A A′)− Mod . This follows from (Base change 0)( see Lemma 3.3). There is an A′-linear map tN : dN M/A ⊗A IdA′ : M ⊗A A′ −→ J N (M/A) ⊗A A′ which is a filtered module-derivation. If IB/A is the diagonal ideal, we have that tN (b · m ⊗ a′) − b · m ⊗ a′ = dN M (bm) ⊗ a′ − bm ⊗ a′ ∈ IB/A · J N (M/A) ⊗A A′ ⊆ IB⊗AA/A · (J N (M/A) ⊗A A′), which is the definition of a filtered module derivation. By the universal property of the jet modules, there is a functorial homomorphism (natural transformation) βM : J N (M ⊗A A′/A′) −→ J N (M/A) ⊗A A′ of functors from B − Mod to J N (B ⊗A A′) − Mod . If M ∼= B, by (Base change 0) (see Lemma 3.3), there is an isomorphism βB : J N (B ⊗A A′) ∼= J N (B/A) ⊗A A′. Since the jet-modules commute with direct sums, βM is an isomorphism for each free B-module B⊕I . If M is arbitrary, choose a presentation B⊕I −→ B⊕J −→ M −→ 0 Since βB⊕I and βB⊕J are isomorphisms, βM is an isomorphism by the five- lemma. (cid:4) Proposition 3.11 (Base change II) Let A −→ B be a homomorphism of rings, M be a B-module and N be an A-module. Then, there is a canonical isomorphism αN : J N (M/A) ⊗A N ∼= J N (M/A) ⊗A N. Proof: Fixing the B-module M, both sides can be considered as functors from A − Mod to J N (B/A) − Mod . For each A-module N, there is an A-linear map M ⊗A N dN M ⊗AIdN−→ J N (M/A) ⊗A N. This is a filtered module derivation, i.e., if IB/A is the diagonal ideal, then dN (b · (m ⊗ n)) − b · (m ⊗ n) = dN M/k(b · m) ⊗ n − b · m ⊗ n ∈ IB/A · (J N (M/A) ⊗A N) ⊆ IB⊗AA′ · (J N (M/A) ⊗A N, 8 because dN jet-modules, there is a unique homomorphism of J N (B/A)-modules M/A is a module derivation. By the representing property of the αN : J N (M ⊗A N/A) −→ J N (M/A) ⊗A N. This homomorphism is in fact a natural transformation of functors from A − Mod to J N (B/A) − Mod. Both functors are right exact functors. If N = A⊕I is a free A-module, then both sides are isomorphic to J N (M/A)⊕I because the jet-modules commute with direct sums. In the general case, choose a presentation A⊕I −→ A⊕J −→ N −→ 0. We know that αA⊕I and αA⊕J are isomorphisms, so by the five lemma , it follows that αN is an isomorphism. (cid:4) Proposition 3.12 (Tensor Products) Let k −→ A be a homomorphism of commutative rings and M, N be two A-modules. Then, for each N ∈ N0 ∪ {N} , there is a canonical functorial isomorphism θM,N : J N (M ⊗A N/k) ∼=−→ J N (M/k) ⊗J N (A/k) J N (N/k), in the sense that both sides are bi-functors to J N (A/k) − Mod and θM,N is a natural transformation of bifunctors that is for each object (M, N) an isomorphism. Proof: The arguement is standard. There is a canonical homomorphism tN M ⊗AN : M ⊗A N M ⊗AdK dK N−→ J N (M/k)(2) ⊗A J N (N/k)(2) ։ J N (M/k) ⊗J N (A/k) J N (N/k), where the superscript (−)(2) indicates that the jet-modules are considered with respect to the second A-module structure. By the universal property of J N (M ⊗A N/k), there is a unique homomorphism of J N (A/k)-modules θM,N : J N (M ⊗A N/k) −→ J N (M/k) ⊗J N (A/k) J N (N/k). That this is a natural transformation of bi-functors follows from the unique- ness of θM,N . Now, if M = A⊕I is free, then θM,N is an isomorphism (both sides are iso- morphic to Li∈I J N (N/k) . Furthermore, both sides are right exact functors in the M-variable for fixed N, so the result follows by choosing a presentation A⊕I −→ A⊕J ։ M. (cid:4) Combining (Base change I) and (Base change II) we get Proposition 3.13 (Base change III) Let A −→ B and A −→ A′ be homo- morphisms of commutative rings, M be a B-module and N be an A′-module. Then, for each N ∈ N0 ∪ {N}, functorial in N, there are isomorphisms of J N (B ⊗A A′/A′)-modules αM,N : J N (M ⊗A N/A′) ∼=−→ J N (M/A) ⊗A N. 9 Proof: The arguement is now standard. We fix the B-module M . One checks that dN M/A ⊗A idN : M ⊗A N −→ J N (M/A) ⊗A N is a filtered module dervation relative to A′, giving rise to a functorial homo- morphism of J N (B ⊗A A′)-modules αN : J N (M ⊗A N/A′) ∼=−→ J N (M/A) ⊗A N. is an isomorphism. It then follows αN is an iso- By (Base change II), αA′ morphism for a free A′-module N, and αN is then an isomorphism for each N by taking free presentations and application of the five-lemma. (cid:4) Lemma 3.14 Let A −→ B be a smooth homomorphism (of finite type) of noetherian rings and M be a projective B-module. If 0 −→ N1 −→ N −→ N2 −→ 0 is an exact sequence of A-modules, then for each N ∈ N0 ∪ {N}, there is an exact sequence of J N (B/A) modules (∗)M : 0 −→ J N (M⊗AN1/A) −→ J N (M⊗AN/A) −→ J N (M⊗AN2/A) −→ 0. Furhtermore, the exact sequence (∗)M is functorial in M i.e., (∗)M is a functor from A−Mod to the category of exact sequences in J N (B/A)−Mod . Proof: Follows from (Base change III) and the functor properties of the jet- modules and the fact, that in this case J N (M/A) is a projective B-module, hence a flat A-module. (see Proposition 3.5) (cid:4) Remark 3.15 We have only proved the fundamental properties of the jet- modules (tensor products, base change ...) for N ∈ N0 . But the result for N = N follows by taking projective limits of the isomorphisms obtained for N ∈ N0. Because of lack of reference, we want to prove the following inocuous gener- alization of the formal inverse function theorem. Lemma 3.16 Let A be a noethrian ring and let C := A[[y1, ..., yn]]/(f1, ..., fm) with m ≥ n be a formal power series ring such that some (n × n)-minor has a determinant which is a unit in A. Then C ∼= K . Proof: Without loss of generality, let this be the left upper most minor . But then, in the power series ring A[[d1x1, ..., d1xn]] , by the formal inverse function, theorem d1f1, ..., d1fn are formal coordinates and A[[d1x1, ..., d1xn]]/(d1f1, ..., d1fn) ∼= A 10 and a fortiori A[[d1x1, ..., d1xn]]/(d1f1, ..., d1fm) ∼= A . The proof in [4] in the introductory chapter given for the case where A is a field carries over verbatim. One has to develop d1xi into a formal power series in the d1fj, d1xi = nX j=1 bjid1xi +X J (d1f )J . By the Cramer rule one determines the bji and then, inductively one deter- mines the bJ , J ≥ 2 . (cid:4) Lemma 3.17 ( etale invariance of the jet modules). Let k −→ A −→ B be homomorphism of finite type of noetherian rings with A −→ B being etale. Let M be an A-module. Then, there is a canonical isomorphism αM : J N (M ⊗A B/k) ∼=−→ J N (M/A) ⊗A B which is a natural transformation of right exact functors from (A − Mod) to J N (B/k) − Mod . Proof: We first treat the case M = A . It suffices to show the claim for the full jet module. For finite N ∈ N0 , the result follows by taking truncations. We choose a presentation B = A[x1, ..., xn]/(f1, ..., fm) . By elementary di- mension theory, we must have m ≥ n . We then have J N(B/k) = J N(A/k) ⊗A B ⊗B B[[d1x1, ..., d1xn]]/(d1f1, ..., d1fm), by choosing an appropriate presentation of A/k. So it suffices to show, that B[[d1x1, ..., d1xn]]/(d1f1, ..., d1fm) ∼= B and we get J N(B/k) ∼= J N(A/k) ⊗A B. We use the Jacobian criterion for smoothness. Considering the Jacobian matrix ∂1/∂1x1(f1) ∂1/∂1x1(f2) ∂1/∂1x2(f1) ∂1/∂1x2(f2) ... ...   . . . ∂1/∂1xn(f1) . . . ∂1/∂1xn(f2) . . . . . . ∂1/∂1xn(fm) ...   ∂1/∂1x1(fm) ∂1/∂1x2(fm) We then have that the nth Fitting ideal of this matrix, generated by the (n × n)-minors is equal to the unit ideal in B, because it is nonzero modulo each maximal ideal of B, and otherwise, if the fitting ideal where not the unit ideal, there would be a maximal ideal containing it, a contradiction. We want to consider the nth-fitting ideal of the Jacobian matrix J (d1f /d1x). To make sense of this, recall that by definition for the jet algebras for free polynomial algebras (see [3])[chapter 6.5, pp. 116-119], and simply by the fact that the universal derivation is a k-algebra homomorphism, fi + d1fi = fi(x1 + d1x1, ..., xn + d1xn) = X I ∂I/∂xI(f ) · d1xI. 11 Observe that this sum is finite, since the fj are polynomials and considering the B-algebra B[d1x1, ..., d1xn]/(d1f1, ..., d1fm) makes sense. So we can write d1fi as a polynomial in the d1xi with zero constant term and coefficients in B. By the above formula, if I = (0, ..., 1, ..., 0) , we get that the first partial derivative of d1fi with respect to the free variable d1xj is just ∂1/∂1xj(fi) ∈ B . For an arbitrary multi-index I, this equality only holds up to a constant factor c ∈ N . Thus, we can apply the Jacobian criterion for smoothness in order to conclude that B[d1x1, ..., d1xn]]/(d1f1, ..., d1fm) is etale over B, i.e., smooth of relative dimension zero. We consider the nth fitting ideal Fittn of J (d1f /d1x) . The coefficients of the n × n-minors lie actually in B. Let p ∈ Spec(B) be given with Bp/p · Bp = K being the residue field. We consider the reduced ring K[[d1x1, ..., d1xn]]/(d1f1, ..., d1fm). The nth fitting ideal of the Jacobian J (d1f /d1x) modulo p is then the unit ideal in K which precisely means that the determinant of some n × n-minor must be nonzero. But then, the determinant of this minor is a unit in Bp and there exists an open affine Spec C ⊂ Spec B such that this determinant is a unit in A. By the previous lemma, we conclude A[[d1x1, ..., d1xn]/(d1f1, ..., d1fm) ∼= A. This holds for each prime ideal p ∈ Spec(B) . Hence, there is a finite Zariski open affine covering Spec B = Sn i=1 Spec Ai such that Ai[[d1x1, ..., d1xn]]/(d1f1, ..., d1fm) ∼= Ai and the claim follows. This shows, that the canonical homomorphism φM : J N(M/k) ⊗A,p1 B −→ J N(M ⊗A B/B), which is simply the map \A ⊗k M ⊗A,p1 B −→ \B ⊗k M −→ \B ⊗k (M ⊗A B) is an isomorphism for M = B . Now, the proof is standard. φM is a natural transformation of right exact functors from B − Mod to J N (B/k) − Mod . The result follows for free A-modules M, since taking jet-modules commutes with taking direct sums and, choosing a free presentation for general M, the result follows by the five-lemma. (cid:4) Corollary 3.18 (invariance under Zariski- localization) Let k be a noethe- rian ring and k −→ A be k-algebra of finite type. Let S ⊂ A be a multiplica- tively closed subset. Then, there is a canonical isomorphism J N (AS/k) ∼= J N (A/k)S, where S = S ⊗ 1 in J N (A/k). Corollary 3.19 If (A, m, κ) is a local ring that is a k-algebra essentially of finite type, then for each N ∈ N0 , J N (A/k) is an A-algebra essentially of finite type. 12 3.4 Comparison with the C∞-category It is well known that for C∞-manifolds and vector bundles on them, being a differential operator is a local property, which is in this category one way to define them. In this subsection we show that in the algebraic category, an analogous statement holds, if we use the etale topology on a smooth algebraic scheme X. Proposition 3.20 Let S be a noetherian scheme and π : X −→ S be a smooth S- scheme of finite type of dimension n over S and E be a locally free coherent OX-module. Let D : E −→ E be a homomorphism of etale sheaves of π−1OS-modules. Suppose, that for each scheme point x ∈ X, there is an etale neighbourhood px : Ux −→ X such that there is a trivialization Ux plus an etale surjective morphism qx : Ux −→ Vx ⊆ An φx : p∗ S . Then Vx is Zariski- open in An Ux be the section over Ux of D with respect to the trivialization of φx of E around x ∈ X . We say that D is a classical linear partial differential operator if there is a partial differential operator Dx : O⊕r Vx , that pulls back under qx to S . Let Γ(Ux, D) : O⊕r Vx −→ O⊕r Ux −→ O⊕r xE ∼= O⊕r Γ(Ux, D) . Then, there is an OX -linear homomorphism eD : J N (E/S) −→ E such that D = eD ◦ dN E/S is of this form. Furthermore, every OX-linear homomorphism from Ω≤N (E/S) −→ E (see [3]) corresponds to a classical differential operator E −→ E. E/S , and conversely, every D = eD ◦ dN Proof: Under these assumptions for each scheme point x ∈ X the classical operator Dx : O⊕r Vx corresponds to a section for some Nx ∈ N. Since the jet bundles are invariant under etale pull back (Lemma 3.17) , for each x ∈ X we get a the pulled back section fDx ∈ Γ(Ux, HomUx(J Nx(E/S), E)), using the trivialization of E over Ux, such that fDx composed with Γ(Ux, dNx is Γ(Ux, D) . I claim that the fDx glue to a global section of DON E/S) X/S(E, E) over X for some N ∈ N. First since X is quasicompact, we can find an etale finite subcovering {Uxi −→ X, i ∈ I} with I a finite set. So we can take as our N the number N = maxi∈I Nxi. Now the global differential operators are a subalgebra of the π−1(OS) linear endomorphism algebra of E . We know, that etale locally, the endomorphism D is given by an OX -linear map fDx. On etale overlaps Ux ×X Uy, D is certainly given by an element of Γ(Uxy, DON (E), E) But since each element in Γ(Uxy, DON (E, E)) determines uniquely an element in HomUxy(J N (E/S), E) the two elements obtained by restrictions from Ux and Uy to Uxy = Ux ×X Uy must agree. Thus since J N (E/S) and E are etale sheaves (since they are coherent on X), we get a global section in DON (E, E) over X. The converse of the statement follows from the fact, that if πx : Ux −→ Vy is an etale morphism, then π∗ xDON (O⊕r Vx /S) ∼= DON (O⊕r Ux /S), 13 Vx −→ O⊕r fDx ′ ∈ Γ(Vx, HomVx(J Nx(⊕r i=1OVx/S), ⊕r i=1Ovx) which is a simple consequence of the etale pull back property of the jet- modules (Theorem 3.45). Thus each differential operator on Ux is the pull back of a differential operator on Vx , so the local description of a globally defined differential operator on E/S is always satisfied. The last statement follows from the local description of differential operators in the Ω-formalizm on An k (see [3][chapter 6.4. Theorem 6.55,p.97, chapter 8, Corollary 8.11(2),p. 146], namely locally on An k they give classical partial linear differential operators, and the fact, that they form an etale sheaf . (cid:4) Remark 3.21 In the same situation, we can prove in the same way, that if E1 and E2 are locally free sheaves on X, then each linear partial differential operator between E1 and E2 has either a description via the jet bundle or the etale local description. 3.5 The global case By the etale invariance property of the jet-module (and hence invariance under Zariski-localizations), if q : X −→ S is a morphism of finite type be- tween noetherian schemes or noetherian algebraic spaces, if F is a coherent sheaf on X, if we choose affine Zariski-open covers of X and S , the locally defined jet-modules glue to a global jet-module J N (F /S) . Under the as- sumptions made, this is a coherent sheaf on X. This follows from the fact, that the localization isomorphisms are canonical (follows from the universal representing properties of the jet-modules) and hence, the cocycle conditions are satisfied). If X and S are noetherian algebraic spaces, one defines the jet sheaf first in the case, where the morphism is representable, i.e. we can find an etale cover {Spec Ai −→ S} such that X ×S Spec Ai is a scheme. Then, etale locally over S, the jet-modules are defined by the scheme case. If the morphism q is not representable we can assume that S = Spec A is a noetherian affine scheme. Then choose an etale cover {Spec Bj −→ X} and the jet-modules J N (Mj/A) , where F Spec Bj = fMj glue to a globally defined jet sheaf J N (F /S) . All we need is the base change -and etale invariance property of the jet-modules. Also, the universal filtered derivations dN Mj/Ai : Mj −→ J N (Mj/Ai), i ∈ I, j ∈ J glue to a universal derivation dN F /S : F −→ J N (F /S). Furthermore, for a fixed quasi coherent sheaf F , the universal represent- ing property of the pair dN F /S, J N (F /S) for the moduli problem, sending a J N (X/S)-module Q to the set of all filtered derivations t : F −→ Q is satisfied, because the required homomorphism of J N (X/S)-modules φ : J N (F /S) −→ Q can be constructed etale-or Zariski-locally, and by the uni- versal property in the affine case, these locally construced φi glue to a global φ . If φ1, φ2 are two homomorphisms of J N (X/S)-modules with t = φi◦dN F /S , then they locally agree, hence by the sheaf property they agree globally. We have proved the following 14 Theorem 3.22 Let q : X −→ S be a morphism of finite type of noetherian schemes, or, more generally of neotherian algebraic spaces and let F be a quasi coherent sheaf on X. Then, for each N ∈ N0 ∪ {N}, there is a quasi coherent OX-module J N (F /S) plus a filtered derivation dN F /S : F −→ J N (F /S) with respect to the diagonal ideal sheaf IX/S that represents the functor, send- ing a J N (X/S)-module Q to the set of all filtered derivations t : F −→ Q with respect to the diagonal ideal sheaf IX/S . Proof: (cid:4) We thus make the following (basically standard) definition. Definition 3.23 Let X −→ S be an arbitrary morphism of finite type of noetherian schemes, or more generally of noetherian algebraic spaces and Fi, i = 1, 2 be quasi coherent sheaves on X. Then, a differential operator of order ≤ N is an OS-linear map D : F1 −→ F2 that can be factored as dN F /S−→ J N (F1/S) eD−→ F2, F1 where the homomorphism eD is OX-linear, where J N (X/S) is regarded with respect to the OX -module structure coming from the first tensor factor. A differential operator of order N is a differential operator that is of order ≤ N but not of order ≤ N − 1 . Thus, in this situation, there is a 1-1 correspondence between differential operators F1 −→ F2 relative to S and OX-linear maps J N (F1/S) −→ F2 . Proposition 3.24 (arbitrary push-forwards) Let X −→ S be mor- phisms of schemes and Fi, i = 1, 2 be quasi coherent sheaves on X. Let f∗D D : F1 −→ F2 be a differential operator relative to S. Then f∗F1 −→ f∗F2 is a differential operator between the quasi coherent sheaves f∗Fi relative to S, where f∗D is taken in the category of sheaves of (π ◦ f )−1OS-modules on X. −→ Y f p Proof: Let D be given by F1/S : F1 −→ J N (F1/S) −→ F2, eD ◦ dN where the first map is (π ◦ f )−1OS-linear and eD is OX -linear. Then f∗dN is an π−1OS-linear map from f∗F1 to f∗J N (F1/S) and f∗eD is f∗OX, and thus OY linear via the structure homomorphism OY −→ f∗OX . The morphism f induces a morphism F1/S J N (f /S) : J N (X/S) −→ J N (Y /S), where J N (X/S) := SpecX J N (X/S) with projection pX : J N (X/S) −→ X , such that pY ◦ J N (f /S) = f ◦ pX . 15 and thus we have a homomorphism of sheaves J N (Y /S) −→ f∗J N (X/S) . Hence we have that f∗J N (F /S) is an f∗J N (X/S)-module and thus an J N (Y /S)-module. By Lemma 3.2, there is a unique homomorphism φ : J N (f∗F1/S) −→ f∗J N (F1/S) such that f∗dN as F1/S = φ ◦ dN f∗F1/S . The π−1OS-linear map f∗D can be written f∗D : f∗F1 dN f∗F1/S −→ J N (f∗F1/S) eD◦φ) (f∗ −→ f∗F2 and is a partial linear differential operator on Y over S. (cid:4) Remark 3.25 If q : X −→ S is a morphism of noetherian schemes, D : E1 −→ E2 is a differential operator relative to S and F is a quasi coherent OS-module, it follows from the global version of Lemma 3.39 that D ⊗q−1OS IdF is a differential operator on X relative to S. π−→ S be morphisms of Proposition 3.26 (etale pull back) Let X schemes where f is etale. If D : F −→ F is a differential operator on the quasi coherent OY -module F , then f ∗D : f ∗F −→ f ∗F is a differential operator on the quasi coherent OX -module f ∗F . −→ Y f Proof: This follows from the etale invariance property of the jet modules, i.e. f ∗J N if D is given as F /S , the homomorphism of sheaves of (π−1 ◦ f )(OS)-modules X (f ∗F /S) (Lemma 3.17.) Then, Y (F /S) ∼= J N D = eD ◦ dN f ∗D, is given by f ∗D : f ∗F f ∗dN F /S =dN −→ f ∗F /S f ∗J N (F /S) ∼= J N (f ∗F /S) f ∗ eD −→ f ∗F . (cid:4) To avoid confusion, for each N ∈ N0, the OS-module J N (X/S) can be regarded as the structure sheaf of the higher tangent bundle J N (X/S) . This is an OX -bi-module with respect to the two tensor factors, sloppily written as J N (X/S) = OX ⊗OS OX /I N +1 X/S . Denote by p1,X, p2,X the two projections J N (X/S) −→ X , where we defined in the introduction p1,X = pX . If the scheme X under consideration is clear from the context, we drop the subscript (−)X . To be more precise, there are two OS-linear homomorphisms p♯ 1,p♯ 2−→ J N(X/S). OX If locally Spec A ⊂ S and Spec B ⊂ X are open affine subsets, Spec B mapping to Spec A , then Γ(Spec A, J N (X/S)) = B ⊗A B/I N +1 B/A The two maps p♯ 1, p♯ 2 correspond to the natural maps B −→ B ⊗A B −→ B ⊗A B/I N +1 B/A , b 7→ b ⊗ 1, 1 ⊗ b. 16 Both homomorphisms give J N (X/S) the structure of a quasi coherent OX - algebra We have defined in the section Notation and Conventions J N (X/S) = SpecX p1,,X∗J N (X/S) with natural projection p1,X = pX : J N (X/S) −→ X which is a morphism of schemes over S. There is a second morphism over S, p2,X : J N (X/S) −→ X whose structure homomorphism OX −→ p2,X,∗J N (X/S) corresponds to the universal filtered derivation. This holds for all N ∈ N0 ∪ {N}. This we want to make clear by the following Lemma 3.27 Let q : X −→ S be a morphism of schemes and Q be a J N (X/S) -module for some N ∈ N0. Then p1,∗Q ∼= p2,∗Q as q−1OS-modules. In particular, pX,1,∗OJ N (X/S) = pX,2,∗OJ N (X/S) = J N (X/S). Proof: The morphisms p1 and p2 are affine and finite and on the underlying scheme points a topological isomorphism J N (X/S) ∼= X . Let Spec B ⊂ X and Spec A ⊂ S be open affine subschemes with Spec B mapping to Spec A . Then p−1 1 (Spec B) = p−1 2 (Spec B) = Spec(B ⊗A B/I N +1 B/A ). Then by definition of push forward of a sheaf, the claim follows. (cid:4) Remark 3.28 If Q is an J N (X/S)-module, the sheaf p1,∗Q = p2,∗Q sim- ply regarded as a sheaf of q−1OS-modules on X, possesses two OX-module structures. Restricting to an open affine Spec B ⊂ X , if Q corresponds to the B ⊗A B/I N +1 B/A -module M , this is simply the A-module M, and the two OX -module structures on M correspond to the two B-algebra structures on J N (B/A) . Lemma 3.29 With notation as above, suppose that supp(F ) = Y ( X with DY−→ F on Y , i.e., IY = ann(F ) . Then, if there is a differential operator F F regarded as a sheaf on Y , then too on F regarded as a sheaf on X. fDY : J N Proof: By assumption, there is some N ∈ N plus an OY -linear map Y (F /k) −→ F . By looking at the local description of the jet-modules, Y (F /k) there is always an OX -linear surjection pXY : J N (which is locally of the form X (F /k) −→ J N (A ⊗k M ։ (A/IY ) ⊗k M) which is (A, p1)-linear. Composing with pXY , we get gDX = fDY ◦ pXY : X (F /k) −→ F , that, composed with dN J N over X, DX : F −→ F . In order to study the behavior of a differential operator with respect to the natural torsion filtration on a coherent sheaf, we prove the following X,F /k, gives the differential operator (cid:4) 17 Lemma 3.30 Let π : X −→ S be a morphism of algebraic schemes and F be a quasi coherent sheaf on X and F ′ ⊂ F be a coherent subsheaf. For each N ∈ N, let J N (F ′/S)′ be the subsheaf of J N (F /S) which is the image under the natural homomorphism J N (F ′/S) −→ J N (F/S), where i : F ′ ֒→ F is the inclusion. We have on X J N (i/S) Ann(J N (F ′/S)′) ⊇ ann(F ′)N +1 ⊗OS OX , where we regard J N (F ′/S) as a coherent sheaf on J N (X/S). Thus, if dim(F ′) ≤ d, then also dim(J N (F ′/S)′) ≤ d. Proof: The question is local, so let A −→ B be a homomorphism of finitely generated k-algebras and let M be an A- module. I claim that ann(M)N +1⊗A B ⊆ ann(J N (M/A)) . We have J N (M/A) = B ⊗A M/I N +1 B/A · (B ⊗A M). Let a = ann(M) and a ∈ a . By definition, we know, that J N (M/A) is annihilated by B ⊗ a . We have a ⊗ 1 − 1 ⊗ a ∈ IB/A and for all m ∈ M, 0 = (a ⊗ 1 − 1 ⊗ a)N +1 · 1 ⊗ m = (aN +1 ⊗ 1 + (1 ⊗ a) · ω) · 1 ⊗ m and it follows (a ⊗ 1)N +1 · 1 ⊗ m = 0 ∀m ∈ M and a ∈ a . Thus ann(M)N +1 ⊗A B ⊆ Ann(J N (M/A)). If now M ′ ⊂ M corresponds over Spec B to F ′ ⊂ F and a′ = ann(M)′ then ′N +1 ⊗A B and so the image J N (F ′/S)′ ⊆ J N (F /S) is Ann(J N (M ′/A) ⊇ a also (locally over X annihilated by a′N +1 ⊗A B . Now the statement about the dimension follows from the fact that if dim(OX / Ann(F ′)) ≤ d , then also d ≥ dim(F ′) = dim(OX / Ann(F ′)N +1) ≥ dim(OJ N (X/S)/ Ann(J N (F ′/S))′) = dim(J N (F ′/S)′). (cid:4) We have the following important Corollary 3.31 Let π : X −→ S be a morphism of algebraic schemes and E be a coherent OX-module. Let T iE, i = 0, ..., dim(E) be the torsion filtration of E and for some N ∈ N and D : E −→ E be a differential operator relative to S of order ≤ N . Then, D respects the torsion filtration of E , i.e. D(T i(E)) ⊆ T iE. Proof: The question is local so let be as above A −→ B be a homomorphism of rings and M be a B module and D : M −→ M be a differential operator of order ≤ N . Let M ′ ⊂ M be a submodule of M of dimension ≤ d . Let I = ann(M ′) . Then by the previous proposition I N +1⊗B ⊆ ann(J N (M ′/A) . The differential operator D restricted to M ′ factors over J N (M ′/A)′ ⊂ J N (M/A) . Let eD : J N (M/A) −→ M be the B-linear map corresponding 18 to D. Then the image of J N (M ′/A) in J N (M/A) is likewise annihilated by I N +1 ⊗ B and so is eD(J N (M ′/A)) ⊂ M . Thus, the image of M ′ under D in M is contained in a submodule annihilated by I N +1 . Since dim(M ′) ≤ d dim(B · D(M ′)) ≤ d. T d(M) is the maximal submodule of M of dimension ≤ d and we have proved that dim(B · D(T d(M)) ≤ d which implies the claim. (cid:4) Let π : X −→ S be a smooth morphism of finite type of noetherian schemes and E be a coherent sheaf on X. Let J (N )(E/S) = IX/S · J N (E/S) so in particular J (N )(X/S) = IX/S/I N +1 X/S . There is a short exact sequence (∗) 0 −→ J (N )(E/k) −→ J N (E/k) −→ E −→ 0. For a smooth morphism, it is well known that I N X/S/I N +1 X/S ∼= Ω(1)(X/S)⊗sN . Lemma 3.32 With notation as above, the homomorphism X/S/I N +1 I N X/S ⊗OX J N (E/S) −→ J N (E/S), coming from the J N (X/S)-module structure of J N (E/S) descends to a OX - linear map, X/S/I N +1 I N X/S ⊗OX E −→ J N (E/S). Proof: This follows from the exact sequence (*) and the fact that I N IX/S · J N (X/S) −→ J N (E/S) is the zero map. We now state here the following basic fact about jet-modules in the global case. X/S/I N +1 X/S ⊗ (cid:4) Proposition 3.33 Let π : X −→ S be a morphism of finite type of noethe- rian schemes. 1 for each N ∈ N0 ∪ {N} , let J N (−/S) be the functor from quasi co- herent OX -modules to quasi coherent J N (X/S)-modules, sending F to J N (F /S) . Then, this functor is right exact and there is a canonical nat- ural isomorphism J N (−/S) ∼=−→ p∗ 2(−). 2 If π : X −→ S is flat, then J N (−/S) is an exact functor. 3 If π : X −→ S is a smooth morphism of noetherian schemes, then for each N ∈ N0 , the functor J N (−/S), sending quasi coherent OX-modules to quasi coherent J N (X/S)-modules, is exact and equal to (pN 2 )∗ . Proof: 1 This follows from the local definition of the jet-modules. Here, of coarse, p2 : J N (X/S) −→ X is the finite affine morphism which corresponds to the second OX-module structure. 2 This immediately follows from (1). 3 This follows from the fact, that for X/S smooth, the N th-jet algebra J N (X/S) is a projective, hence flat OX -module (see Proposition 3.5). So the assertion follows from (2). 19 Proposition 3.34 (Exact sequence I) Let π : X −→ S be a morphism of finite type between noetherian schemes and Y ⊂ X be a closed subscheme and F be a quasi coherent OY -module. Let IY be the defining ideal sheaf of Y . Then for all N ∈ N ∪ {∞} there is an exact sequence (cid:4) 0 −→ (IY ) · J N X (F /S) −→ J N X (F /S) −→ J N Y (F /S) −→ 0. Here multiplication with IY is via the first OX-module structure. Proof: This is a local question , so let A −→ B be a homomorphism of rings, I ⊂ B be an ideal and M be a B/I-module. We have J N B/A(M/A) = B ⊗A M/I N +1 B/A · (B ⊗A M) and J(B/I)/A(M/A) = (B/I) ⊗A M/I N +1 (B/I)/A · (B/I ⊗A M). B ⊗A M is already a B ⊗A (B/I) module, so tensoring with B/I via the first B-module structure, we obviously get B/I ⊗A M which is J(B/I)/A(M/A) . (cid:4) Proposition 3.35 (Exact sequence II) Let as above π : X −→ S be a mor- phism of finite type between noetherian schemes and Y ֒→ X be a closed sub- scheme with defining ideal sheaf IY . Let F be a quasi coherent OY -module. Then for all N ∈ N0 ∪ {N} there is an exact sequence 0 −→ IY · J N X (F /S) Y −→ J N X (F /S) Y −→ J N Y (F Y /S) −→ 0, where multiplication with IY is via the second OX -module structure on the jet bundle. Proof: The question is again local, so let A −→ B be a homomorphism of rings and M be a B-module and I ⊂ B be the ideal corresponding to Y ⊂ X . The module J N Y /S(F Y /S) then corresponds to the module (B/I) ⊗A (M/IM) (modulo the ideal I N +1 X (F /S) Y corresponds to the module B/I ⊗A M so tensoring with B/I via the second B-module structure we get B/I ⊗ M/IM which was the claim. (cid:4) (B/I)/A) and the jet bundle J N Proposition 3.36 (First cotangential sequence) i ֒→ X π−→ S be morphisms of finite type of noetherian schemes such Let E that i is a closed immersion. Let F be a quasi coherent OX -module. Then there is and exact sequence (∗1) IE · F /I 2 E · F −→ J (1) X (F /S) E−→ J (1) E (F E /S) −→ 0. If π and π ◦ i are smooth and F is locally free, the sequence (∗1) is exact on the left. In this case, the first module is isomorphic to IE/I 2 These exact sequences are functorial in F ∈ QCoh(X) . E ⊗OX F . 20 Proof: In view of the exact sequence II, it is enough to construct a functorial homomorphism φ : IE · F /I 2 E · F −→ IE · J (1)(F /S) E . The question is local, so let k −→ A be a homomorphism of commutative rings, I ⊂ A be an ideal, and M be an A-module. We have to construct an A-linear homomorphism ψM : IM/I 2M −→ (A/I) ⊗k M/(I 2 A/k · ((A/I) ⊗k M)), which is functorial in M, i.e., is a natural transformation from A − Mod to A − Mod . Let f ∈ I, m ∈ M be given. We let ψM (f m) := 1 ⊗ f m. This is obviously additive. Secondly, since 0 = (f ⊗ 1 − 1 ⊗ f )2 = f 2 ⊗ 1 + 2f ⊗ f + 1 ⊗ f 2 = 1 ⊗ f 2 , it follows that I 2M maps to zero. We have to show that this is A-linear. So for a ∈ A we have to show that ψM (af m) = 1 ⊗ af m = a ⊗ f m ∀a ∈ A, f ∈ I, m ∈ M. Now,J 1(M/k) is an J 1(A/k) = A ⊕ Ω(1)(A/k)-module. There is the stan- dard A-linear homomorphism I/I 2 −→ Ω(1)(A/k) ⊗A A/I which is in our notation the homomorphism ψA with ψA(f ) = 1 ⊗ f − f ⊗ 1 = 1 ⊗ f . Thus we have in J 1(A/k) ⊗A A/I the identity 1 ⊗ af = a ⊗ f . Then from the J 1(A/k)-module structure of J 1(M/k) it follows that 1 ⊗ af m = a ⊗ f m which is the A-linearity of the map ψM . Thus we have in any case an exact sequence IE · F /I 2 E · F −→ J (1) X/S(F /S) E−→ J (1) E/S(F E /S) −→ 0. So let now π and π ◦ i be smooth. From the standard cotangential sequence, the claim is true for F ∼= OX . Since the assertion is local, the claim is true for locally free OX -modules in which case IE/I 2 E ⊗OX F ∼= IE · F /I 2 E · F . Both sides are right exact functors from OX -modules to OE-modules, so the claim follows by taking locally a free presentation and the five lemma. (cid:4) Proposition 3.37 (Second cotangential sequence)) Let X E be a quasi coherent OY -module. Then there is an exact sequence −→ Z be morphisms of finite type of noetherian schemes and −→ Y f g (∗2) f ∗J (1)(E/Z) αE−→ J (1)(f ∗E/Z) βE−→ f ∗E ⊗OX J (1)(X/Y ) −→ 0. If f and g are smooth and E is locally free, then the left hand map is injective. These sequences are functorial in E ∈ QCoh(Y ). Proof: 21 1 Construction of the exact sequence The question is local so, let X = Spec A, Y = Spec B Z = Spec C and let E correspond to the B-module M. The generalized second cotangential sequence is then a subsequence (the trivial direct summand f ∗E deleted) of the following sequence C ⊗B,p1 B ⊗A M = C ⊗A M −→ C ⊗A (C ⊗B M) −→ C ⊗B (C ⊗B M) = (C ⊗B M) ⊗B B ⊗B C, −→ 0 where the last identity comes from the fact that J N (B/A) ⊗A M ∼= J N (B ⊗A M/A) which is a special case of base change for jet modules. Since the homo- morphisms of the sequence are canonical, these glue to the sequence (∗2) . Observe that in J 1(M/k) , J (1)(M/k) sits as a direct summand. Observe furthermore, that each three terms in the sequence (∗2) are functors from OY −Mod to OX −Mod that are all right exact and αE and βE are natural transformations of functors from OY − Mod to OX − Mod . 2 Exactness If E ∼= OY this is the standard cotangential exact sequence. For arbi- trary locally free E the question is local , so we may assume that X = Spec A, Y = Spec B, Z = Spec C are affine such that E is free on Spec B . Then from the standard cotangential sequence for the situation Spec A −→ Spec B −→ Spec C and taking direct sums, we know that the sequence is locally exact, hence globally exact. For arbitrary coherent E , we only need to show that the sequence is locally exact, since this is a statement about sheaves. If Γ(Spec B, E) = M choose a presentation B⊕I −→ B⊕J −→ M −→ 0. Then the second cotangential sequence for M is the cokernel of the homo- morphism of the second cotangential exact sequence for A⊕I to A⊕J . (all three terms are right exact functors from OY − Mod to OX − Mod . Since the cokernel functor is right exact, the claim follows. If f and g are smooth and E is locally free, this reduces locally to E ∼= OY where this is the standard second cotangential exact sequence. (cid:4) Lemma 3.38 Let q : X −→ S be a morphism of finite type of noetherian schemes where X possesses an ample invertible sheaf and D : F1 −→ F2 be a differential operator of order ≤ N relative to S for some N ∈ N0 between quasi coherent sheaves F1 and F2 , Then, there is a filtration F •F1 and a filtration F •(F2) such that F j(Fi), i = 1, 2 is coherent and D(F j 1) ⊆ F j 2 , where j ∈ J and J is the directed index set of the filtration. 22 Proof: Under the assumptions made, each quasi-coherent sheaf E is the direct limit of its coherent subsheaves E i ⊂ E . Write D as dN F1−→ J N (F1/S) eD−→ F2. F1 1 ⊂ F1 is a coherent subsheaf, then dN if F j the image of the canonical homomorphism J N (F j F1(F j 1 ) ⊂ J N (F1/S) is contained in 1 /S) −→ J N (F1/S) since F j 1 −→ J N (OX/S) · dN F1(F j 1 ) is a filtered module derivation and the homomorphism exists by the univer- sal property of the jet-modules (see Lemma 3.2). It follows that D(F j 1) is contained in the coherent subsheaf F j 2 := eD(J N (X/S) · dN F1(F j 1 )) ⊂ F2. Thus we can write D as the filtered direct limit of differential operators Dj : F j (cid:4) 2 for some directed index set J . 1 −→ F j 3.6 Basic properties of differential operators In this subsection, we collect the basic properties of linear algebraic differen- tial operators which directly follow from the corresponding basic properties of the jet-modules. Lemma 3.39 Let A −→ B be homomorphisms of rings, M1, M2 be B-modules N be an A-module, and D : M1 −→ M2 be a differential operator relative to A. Then, D ⊗A idN : M1 ⊗A N −→ M2 ⊗A N is a differential operator of B-modules relative to A. Proof: This follows from the properties of the jet modules J N (M1 ⊗A N/A) ∼= J N (M/A) ⊗A N (see Proposition 3.11). The operator D is given by a B-linear map eD : J N (M1/A) −→ M2 , so we get a B-linear map eD ⊗A idN : J N (M1/A) ⊗A N = J N (M1 ⊗A N/A), −→ M2 ⊗A N where the B-module structures are given by the first tensor factor. We can generalize the previous lemma slightly to (cid:4) Proposition 3.40 (Base change 0) Let A −→ B be homomorphisms of commutative rings, D : M1 −→ M2 be a differential operator between B- modules relative to A. Let A −→ A′ be homomorphism of commutative rings. Then D ⊗A idA′ : M1 ⊗A A′ −→ M2 ⊗A A′ is a differential operator on B ⊗A A′-modules relative to A′ . Proof: This follows from the base change properties of the jet-modules : we have an isomorphism J N (M/A) ⊗A A′ ∼= J N (M ⊗A A′/A′) 23 see Proposition 3.10 (Base change I for jet- modules). D corresponds to a B-linear map eD : J N (M1/A) −→ M2 . We get a B ⊗A A′- linear map eD ⊗A IdA′ : J N (M1 ⊗A A′/A′) ∼= J N (M1/A) ⊗A A′ −→ M2 ⊗A A′ and the claim follows. (cid:4) Proposition 3.41 (Base change I) Let A −→ B and A −→ A′ be homo- morphisms of commutative rings and N be an A′-module. Given a differential operator D : M1 −→ M2 of B-modules relative to A, the A′-linear map D ⊗A idN : M1 ⊗A N −→ M2 ⊗A N is a differential operator of B ⊗A A′-modules Proof: This follows in the standard way from the identity J N (M ⊗A N/A′) ∼= J N (M/A) ⊗A N (see Proposition 3.13, Base change III). (cid:4) Proposition 3.42 (exterior products) Let k −→ A and k −→ B be ho- momorphisms of commutative rings, D : M1 −→ M2 and E : N1 −→ N2 be differential operators between A-modules M1, M2 and B-modules N1, N2 respectively. Then, the tensor product over k: D ⊗k E : M1 ⊗k N1 −→ M2 ⊗k N2 is a differential operator between the A⊗k B-modules M1⊗k N1 and M2⊗k N2 . Proof: This follows from the properties of the jet-modules (exterior prod- ucts II), namely J N(M/k) ⊗k J N(N/k) ∼= J N(M ⊗k N/k). If for some N ∈ N0 there is an A-linear map eD : J N (M1/k) −→ M1 and a B-linear map eE : J N (N1/k) −→ N2, for some large N ′ >> N , there is an A ⊗K B-linear map J N ′ (M1 ⊗k N1/k) −→ J N (M1/k) ⊗k J N (N1/k) eE eD⊗k −→ M2 ⊗k N2. The proof in the opposite direction is the same, one has only to bear in mind that the order of the differential operators may change. (cid:4) Lemma 3.43 Let k −→ A be a homomorphism of rings and M1, M2, N1, N2, be A-modules with differential operators Di : Mi −→ Ni, i = 1, 2 and let φi : Mi −→ Ni, i = 1, 2 be A-linear maps commuting with the Di . Then, the k-linear maps induced by Di, ker(φ1) −→ ker(φ2) and coker(φ1) −→ coker(φ2) are differential operators relative to k. Also an arbitrary direct sum M i∈I i∈I i∈I Di : M Mi −→ M Ni is a differential operator for an arbitrary index set I and differential operators Di : Mi −→ Ni . 24 Proof: The last statement follows from the fact, that the jet-modules com- mute with direct sums. For the first statement, observe, that, more generally, if D : G −→ F is a dif- ferential operator relative to k and G1 ⊂ G and F1 ⊂ F are A-sub-modules with D(G1) ⊂ F1 , then the restriction of D to G1 is a differential operator relative to k . Namely, consider the J N (A/k)- module J N (A/k) · dN Obviously, the restriction of dN G/k(G1) ⊂ J N (G/k) . G/k to G1 gives a module derivation tG1/k : G1 −→ J N (A/k) · dN G/k(G1). By the universal property, ther exists a J N (A/k)-linear homomorphism φ : J N (G1/k) −→ J N (A/k) · dN G/k(G1), such that tG1/k = φ ◦ dN G1/S. If eD : J N (G/k) −→ F corresponds to D, I claim that eD(J N (A/k) · dN G/k(G1)) ⊂ F1. Indeed, eD is A-linear with respect to the first A-module structure of J N (G/k) . The quantity dN G/k(G1) is an A-submodule with respect to the second A- module structure of J N (A/k) . Thus, J N (A/k) · dN G/k(G1) the superscript (−)(1) indicates that we take the A-module generated by dN G/k(G1) with respect to the first module structure. But this is then an A-bi-submodule, (since the bi-module structure is commutative), or, equiva- G/k(G1) = A ·(1) dN lently a J N (A/k)-submodule of J N (G/k) . Since eD is A-linear with respect to the first A-module structure, the claim follows. Now, the statement for the cokernel. Let πG : G ։ G2 and πF : F ։ F2 be A-linear surjective maps such that the given differential operator D : G −→ F descends to a k-linear map D2 : G2 −→ F2 . The claim is that D2 is a differential operator of A-modules relative to k . Let G1 and F1 be the kernels of πG and πF respectively, with inclusions iG and iF . By the first case, the induced map D1 : G1 −→ F1 is a differential operator relative to k. By the right exactness of the jet-module-functor we have J N (G2/k) ∼= J N (G/G1/k) ∼= J N (G/k)/im(J N (iG/k)(J N (G1/k))) We know, that D1 factors over a map fD1 : im(J N (iG//k))(J N (G1/k))) −→ F1. We can form the quotient map ∼= F2. fD2 : J N (G2/k) ∼= J N (G/k)/im(J N (iG/k))(J N (G1/k)) −→ F/F1 Then, obviously, D2 factors over fD2 since the quotient map has to factor over the quotient module of the modules over which the first two maps factor. This shows that D2 : G2 −→ F2 is a differential operator relative to k. The case of the map induced on cokernels is a special case of this. (cid:4) 25 3.7 Existence of differential operators in the affine case A/k(M1, M2) ( A/k (M1, M2) , where k −→ A is a homomorphism of commutative rings We want to investigate the question, under which conditions DON DON +1 M1, M2 are A-modules. We start with a proposition. Proposition 3.44 Let k −→ A be a homomorphism of noetherian rings and M1, M2 be finitely generated A-modules. 1 Let Ann(M2) = I . Given any A-linear map eD : J N (M1/k) −→ M2 for some N ∈ N0 , there is some K ∈ N such that eD factors over fD′ : J N ((M1/I K · M1)/k) −→ M2 . 2 Let now Ann(M1) = I . Then there is some K ∈ N such that the image of eD is a submodule annihilated by I K . Proof: Denote by IA/k the diagonal ideal, generated by all elements (a ⊗ 1 − 1 ⊗ a) . 1 Obviously the kernel of eD contains I · J N (M1/k) = I ⊗k M1 . We show that the ideal I ⊗k A contains A ⊗k I K for some K ∈ N . We know that I is finitely generated, I = (f1, ..., fk) . have I N +1 A/k = 0 containing the elements In J N (A/k) we (fi ⊗ 1 − 1 ⊗ fi)N +1 = fi ⊗ ω + 1 ⊗ f N +1 i and i ∈ I ⊗k A . 1 ⊗ f N +1 It follows that there exists some K ∈ N such that I ⊗k A contains A ⊗k I K (one can take K = (N + 1) · k.) Then, I ⊗k M contains A ⊗k M/I K · A ⊗ M . Thus, the A-linear map eD factors through A ⊗ M1/A ⊗ I K · M1 = J N ((M1/I KM1)/k). 2 By the same arguement, if Ann(M1) = I , there is K ∈ N such that Ann(J N (M1/k)) contains I K ⊗ A . Then, since A ⊗ I is contained in the annihilator of J N (M1/k) = A ⊗k M , we have I K ⊗ A ⊆ A ⊗ I ⊂ Ann(J N (M/k)). Since eD is A-linear, the result follows. We prove the following basic (cid:4) Theorem 3.45 Let k −→ A be a homomorphism of finite type of noetherian rings of characteristic zero of relative dimension (fibre dimension) ≥ 1 and M1, M2 be finitely generated A-modules. Then, the inclusion DON A/k(M1, M2) ( DON +1 A/k (M1, M2) is for all N ≥ 0 strict in the following cases: 26 1 M1 = M2, or more generally Ann(M1) = Ann(M2) = I and dim(supp(M1)) has fibre dimension greater than zero over k . 2 M1, M2 are nontorsion modules on A. 3 The module M1 is nontorsion and M2 is a torsion module with V (Ann(M2)) = V (I) having fibre dimension greater than zero over k. Proof: 1 The first case is easily reduced to the case (2). By Lemma 3.29, there is a surjection DON A/k(M1, M2) ։ DON (A/I)/k(M1, M2) which induces a surjection DON A/k(M1, M2)/DON −1 A/k (M1, M2) ։ DON (A/I)/k(M1, M2)/DON −1 (A/I)/k(M1, M2). 2 First, we treat for reasons of intuition the case where A is an integral k-algebra with k being a field and M1, M2 being torsion free, or more generally nontorsion A-modules. If Mi ∼= A⊕ri is free, DON (M1, M2) = HomA(J N (A/k)⊕r1, A⊕r2) = DON (A, A)⊕r1·r2 and we are reduced to the case r1 = r2 = 1 . We have the standard exact sequence jN (A/k) : 0 −→ I N ∆ /I N +1 ∆ −→ J N (A/k) −→ J N −1(A/k) −→ 0 A/k/I N +1 A/k Let Xns ⊂ X = Spec A be the set of nonsingular points, which is a nonempty Zariski-open subset. Over Xns , the A-modules J N (A/k), J N −1(A/k) are projective of different ranks, since, considering the exact sequence jN (A/k) we know that I N is isomorphic to the N th sym- metric power of the relative cotangential sheaf. This is nonzero because the relative dimension of A/k was assumed to be greater than one. Thus, the duals DON (A/k) and DON −1(A/k) have over Xns different ranks. It follows that the inclusion DON −1(A/k) ( DON (A/k) is strict, since the inclusion of the corresponding Zariski- sheaves on Spec A is strict. This settles the free and projective case. If M1 and M2 are any nontorsion modules , there is an open subset U ⊂ Spec A where M1 and M2 are free and DON U/k (M1, M2) is nonzero. Since DON lows that this module is nonzero. We now treat the general nontorsion case. So let k −→ A be a homomor- phism of noetherian rings which makes A a finitely generated k-algebra and let M1, M2 be nontorsion modules on A. Let ηi, i = 1, ..., l be the generic points of A. We want to show that the A-module DON A/k (M1, M2) is finitely generated and coherent, it fol- A/k (M1, M2) is nonzero. To this aim, it suffices to A/k(M1, M2)/DON −1 A/k(M1, M2)/DON −1 U/k(M1, M2)/DON −1 27 show that this module is nonzero at the generic points of A, which re- duces by the localization properties of the jet modules to the case, where (A, m, κ) is an Artinian k-algebra, essentially of finite type over k such that the resiude field κ has transcendence degree ≥ 1 over k. First, we show that l(J N −1(A/k)) < l(J N (A/k)) for each N ≥ 1 , where l(−) de- notes the length function. By the standard jet-module sequence and the additivity of the lenght function, we only have to show that the A-module A/k/I N +1 I N is nonzero. Let A ։ κ be the canonical surjection. There is a surjection IA/k ։ Iκ/k which induces surjections A/k I N A/k/I N +1 A/k ։ I N κ/k/I N +1 κ/k ∀n ∈ N. The last module is equal to Ω(1)(κ/k)⊗sN since we work in characteristic zero and this module is nonzero, since trdeg(κ/k) is greater or equal to one. The case for general M1 is basically the same. There is a surjection A/k · M/I N +1 I N A/k · M ։ I N κ/k · M /I N +1 κ/k · M , where M := M ⊗A κ . Since κ is a field, M is free, and the last quantity is isomorphic to Ω(1)(κ/k)⊗sN ⊗κ M which is also nonzero. Thus for each A-module M and each N ∈ N , we have strict inequality l(J N −1(M/k)) < l(J N (M/k)) coming from the exact sequence of nonzero A-modules jN A/k(M) : 0 −→ I N A/k · M/I N +1 A/k · M −→ J N (M/k) −→ J N −1(A/k) −→ 0. By [1][chapter 18, Proposition 18.4, p. 454], for given A-modules N1, N2 , the minimal number r such that Ext1 A(N1, N2) 6= 0 is given by r = depth(Ann(N1), N2). The ideal Ann(N1) is contained in the maximal ideal, unless N1 = 0 which we want to exclude, and, the maximal ideal consists of zero divisors, since A was assumed to be Artinian. Thus the depth is always equal to zero. Consequently, taking HomA(−, M2) of the exact sequence jN A/k(M1) , we get an exact sequence doN A/k(M1, M2) : 0 −→ DON −1 A/k (M1, M2) −→ DON A/k(M1, M2) −→ HomA(I N A/k · M1/I N +1 A/k · M1, M2) −→ 0, where surjectivity on the right comes from the vanishing of the Ext1 and the last A-module is nonzero by what has been just said. Thus, the quotient module DON A/k(M1, M2)/DON −1 A/k (M1, M2) is nonzero. The general case where A/k is an arbitrary finite type k-algebra follows from the fact, that the last quotient is a coherent sheaf on Spec A which is nonzero at the generic points of Spec A and thus nonzero. 28 3 Next, if M2 is a torsion module with Ann(M2) = I, by the previous propo- sition , for each N ∈ N there is K2 = K(N) ∈ N such that eD factors over J N ((M1/I K2 · M1)/k) −→ M2 . But then, D : M1 −→ M2 factors through a differential operator D′ : M1/I K2 · M1 −→ M2. Thus, we have shown that for fixed M1, M2 with Ann(M2) = I and M1 being a nontorsion module, for each N ∈ N there is a surjection DON A/k(M1, M2) ։ DON (A/I K(N)A)/k((M1/I K(N ) · M1), M2). Observe, that if D : M1 −→ M2 factors over M1/I K · M1 , then it certainly factors over M1/I L · M1 for each L ≥ K . Putting K = max(K(N), K(N − 1)) , we get a surjection DON A/k(M1, M2)/DON −1 A/k (M1, M2) ։ DON (A/I K )/k((M1/I KM1), M2)/DON −1 (A/I K )/k(M1/I KM1, M2). Fixing N ∈ N0, we know by case (2), that the module on the left hand side is nonzero and the claim follows. (cid:4) Remark 3.46 Observe, that, if A/k is a finitely generated Artinian k-algebra, there is an N ∈ N such that I N ∆ = 0 in A⊗k A and so J M (A/k) = A⊗k A for M ≥ N . In particular DOM (A, A) = HomA(A ⊗k A, A) and the statement is not true. References [1] David Eisenbud. Commutative Algebra with a View Toward Algebraic Geometry. Springer Verlag New York-Berlin-Heidelberg, 1995. 28 [2] Alexandre Grothendieck. ´El´ement de G´eom´etrie Alg´ebrique I-IV. Pub- lication de IHES No.32, Le Bois-Marie-Bures-sur Yvette, 1967. 2, 3 [3] Stefan Gunther. Higher differential calculus in commutative algebra and algebraic geometry. Universitatsbibliothek Osnabruck, Thesis, 2014. 2, 4, 6, 11, 13, 14 [4] William Ted Martin Salomon Bochner. Several complex variables. Prince- ton University press, 1948. 11 [5] R. Vakil. A beginners guide to jet bundles from the point of view of algebraic geometry. math. stanford.edu, 1998. 1 E-Mail-adress:[email protected] 29
1010.3465	2	1010	2012-02-08T15:44:44	Nonnegative Polynomials and Sums of Squares	[ "math.AG" ]	In the smallest cases where there exist nonnegative polynomials that are not sums of squares we present a complete explanation of this distinction. The fundamental reason that the cone of sums of squares is strictly contained in the cone of nonnegative polynomials is that polynomials of degree $d$ satisfy certain linear relations, known as the Cayley-Bacharach relations, which are not satisfied by polynomials of full degree 2d. For any nonnegative polynomial that is not a sum of squares we can write down a linear inequality coming from a Cayley-Bacharach relation that certifies this fact. We also characterize strictly positive sums of squares that lie on the boundary of the cone of sums of squares and extreme rays of the cone dual to the cone of sums of squares	math.AG	math	NONNEGATIVE POLYNOMIALS AND SUMS OF SQUARES GRIGORIY BLEKHERMAN Abstract. In the smallest cases where there exist nonnegative polynomials that are not sums of squares we present a complete explanation of this distinction. The fundamental reason that the cone of sums of squares is strictly contained in the cone of nonnegative polynomials is that polynomials of degree d satisfy certain linear relations, known as the Cayley-Bacharach relations, which are not satisfied by polynomials of full degree 2d. For any nonnegative polynomial that is not a sum of squares we can write down a linear inequality coming from a Cayley-Bacharach relation that certifies this fact. We also characterize strictly positive sums of squares that lie on the boundary of the cone of sums of squares and extreme rays of the cone dual to the cone of sums of squares 1. Introduction A real polynomial in n variables is called nonnegative if it is greater than or equal to 0 at all points in Rn. It is a central question in real algebraic geometry, whether a non-negative polynomial can be written in a way that makes its nonnegativity apparent, i.e. as a sum of squares of polynomials (or more general objects). Algorithms to obtain such representations, when they are known, have many applications in polynomial optimization [9],[10],[11]. The investigation of the relation between nonnegativity and sums of squares began in the seminal paper of Hilbert from 1888. Hilbert showed that every nonnegative polynomial is a sum of squares of polynomials only in the following 3 cases: univariate polynomials, quadratic polynomials and bivariate polynomials of degree 4. In all other cases Hilbert showed existence of nonnegative polynomials that are not sums of squares. Hilbert’s proof used the fact that polynomials of degree d satisfy linear relations, known as the Cayley-Bacharach relations, which are not satisfied by polynomials of full degree 2d [14],[15]. Hilbert then showed that every bivariate nonnegative polynomial is a sum of squares of rational functions and Hilbert’s 17th problem asked whether this is true in general. In 1920’a Artin and Schreier solved Hilbert’s 17th problem in the affirmative. However, there is no known algorithm to obtain this representation. In particular we may need to use numerators and denominators of very large degree, thus representing a simple object (the polynomial) as a sum of squares of significantly more complex objects [3]. It should be noted that Hilbert did not provide an explicit nonnegative polynomial that is not a sum of squares of polynomials, he only proved its existence. The first explicit example appeared only eighty years later and is due to Motzkin. Since then many explicit examples of nonnegative polynomials that are not sums of squares have appeared [14]. For some low dimensional, symmetric families there are also descriptions of the exact differences between nonnegative polynomials and sums of squares [5]. However even in the smallest cases where nonnegative polynomials are different from sums of squares, 3 variables degree 4 and 2 variables degree 6, we have not had a complete understanding of what makes nonnegative polynomials different from sums of squares. We show that, in these cases, all linear inequalities that separate nonnegative polynomials from sums of squares come from the Cayley-Bacharach relations. The Cayley-Bacharach relations were already used by Hilbert in the original proof of existence of nonnegative polynomials that are not sums of squares. We show that, in fact, these relations are the fundamental reason underlying the existence of any such polynomial, and we provide explicit structure for the linear inequalities 1 separating nonnegative polynomials from sums of squares. The algebra and geometry involved in these two cases is quite similar and we give a complete unified geometric description of the differences between nonnegative polynomials and sums of squares. 1.1. Main Results. By analogy with quadratic forms we will refer to nonnegative polynomials as positive semidefinite or psd for short and sums of squares will be called sos. Any psd polynomial can be made homogeneous by adding an extra variable and it will remain nonnegative. The same holds for sums of squares. We will therefore work with homogeneous polynomials (forms). Our goal is to investigate the cases of forms in 3 variables of degree 6, known as ternary sextics, and forms in 4 variables of degree 4, known as quaternary quartics. We will denote these cases as (3, 6) and (4, 4) respectively. Let Hn,d be the vector space of real forms in n variables of degree d. Nonnegative forms and sums of squares both form full dimensional closed convex cones in Hn,2d, which we call Pn,2d and Σn,2d respectively: and Pn,2d = {p ∈ Hn,2d p(x) ≥ 0 for all x ∈ Rn} , Σn,2d =np ∈ Hn,2d (cid:12)(cid:12) p(x) =X q2 i for some qi ∈ Hn,do . (3, 6) and (4, 4). It is clear that Σn,2d ⊆ Pn,2d and by Hilbert’s theorem this inclusion is actually strict in the cases The defining linear inequalities for the psd cone Pn,2d are easy to describe, they are given by f (v) ≥ 0 for all v ∈ Rn. By homogeneity of forms it suffices to only consider points v in the unit sphere Sn−1. We remark that with this characterization and an appropriate choice of the inner product it is not hard to show that the dual cone to Pn,2d is the conic hull of the real Veronese variety of degree 2d and thus the dual cone of Pn,2d is essentially equivalent to the Veronese Orbitope [16]. The above inequalities are clearly satisfied by all sums of squares but when the sos cone is strictly smaller, it must satisfy additional linear inequalities. We prove the following characterization for (3, 6): Theorem 1.1. Suppose that p ∈ P3,6 and p is not sos. Then there exist two real cubics q1, q2 ∈ H3,3 intersecting in 9 (possibly complex) projective points γ1, . . . , γ9 such that the values of p on γi certify that p is not a sum of squares. More precisely, let z1, . . . , z9 be affine representatives of γi. Then there exists a real linear functional ℓ : H3,6 → R given by ℓ(f ) =X µif (zi), for some µi ∈ C such that ℓ(r) ≥ 0 for all r ∈ Σ3,6 and ℓ(p) < 0. Furthermore at most 2 of the points γi are complex. We also prove a similar theorem for the case (4, 4): Theorem 1.2. Suppose that p ∈ P4,4 and p is not sos. Then there exist three real quadrics q1, q2, q3 ∈ H4,2 intersecting in 8 (possibly complex) projective points γ1, . . . , γ8 such that the values of p on γi certify that p is not a sum of squares. More precisely, let z1, . . . , z8 be affine representatives of γi. Then there exists a real linear functional ℓ : H4,4 → R given by for some µi ∈ C such that ℓ(r) ≥ 0 for all r ∈ Σ4,4 and ℓ(p) < 0. Furthermore at most 2 of the points γi are complex. ℓ(f ) =X µif (zi), 2 These theorems are proved at the end of Section 5. The cases (3, 6) and (4, 4) are quite similar, and we provide a unified presentation of the proofs. The main ingredient in the proofs is the Cayley- Bacharach theorem [6], which shows that the values of forms in H3,3 (resp. H4,2) on the points zi defined above are linearly related and this relation is unique. It was already observed by Hilbert in his original proof that the Cayley-Bacharach relations can be used to construct nonnegative polynomials that are not sums of squares. A modern exposition of Hilbert’s construction along with generalizations is given by Reznick in [15]. We show that the Cayley-Bacharach relations are more than just a way of constructing examples and in the fact they are the fundamental reason that prevents sums of squares from filling out the entire psd cone. We note that for the cases where Pn,2d = Σn,2d the Cayley-Bacharach relations do not exist and it is possible to prove the equality of the psd and sos cones based on non-existence of the relations. Complex zeroes of real forms come in conjugate pairs. In Section 4.1 we show how to exclude the cases of the intersection containing more than one conjugate pair of complex zeroes. We also show how to explicitly derive the inequalities ℓ, given the Cayley-Bacharach relation. This is done for a fully real intersection case in Section 6 and in Section 7 for the case of one conjugate pair of complex zeroes. We also obtain the following interesting corollaries: Corollary 1.3. Suppose that p ∈ Σ3,6 lies on the boundary of the cone of sums of squares and p is a strictly positive form. Then p is a sum of 3 squares and cannot be written as a sum of 2 squares. And for the case (4, 4): Corollary 1.4. Suppose that p ∈ Σ4,4 lies on the boundary of the cone of sums of squares and p is a strictly positive form. Then p is a sum of 4 squares and cannot be written as a sum of 3 squares. The Corollaries 1.3 and 1.4 were used as a starting point to investigate the algebraic boundary of the cones Σ3,6 and Σ4,4 in [2]. Here we briefly note that sextics that are sums of three squares of cubics and quartics that are sums of four squares of quadratics form hypersurfaces in H3,6 and H4,4. One of the main results of [2] is establishing the degree of these hypersurfaces with a connection with K3 surfaces. In Section 3 we examine in detail the case of an arbitrary completely real transverse intersection of two cubics for the case (3, 6) and three quadratics for the case (4, 4). We provide a complete de- scription of the differences between attainable values of psd forms and sos forms on the intersection points zi. Let E : Hn,2d → Rm be the evaluation map, sending p ∈ Hn,2d to its values on zi: E(p) = (p(z1), . . . , p(zm)). Here m = 9 for the case (3, 6) and m = 8 for the case (4, 4). Let Rm + be the nonnegative orthant of Rm, and let Rm ++ denote the (open) strictly positive orthant. Let P ′ and Sq′ be the images of Pn,2d and Σn,2d under E. We show that in the cases (3, 6) and (4, 4) with zi coming from any completely real transverse intersection of two cubics or three quadratics the image P ′ of Pn,2d contains the positive orthant Rm ++. In other words any combination of strictly values on the points zi is realizable by psd forms. However the Cayley-Bacharach relation forces restrictions on values of sos forms. We show the following (Theorem 3.4): Theorem 1.5. We can choose affine representatives z1, . . . , zm for the projective points γi so that the image Sq′ of the sos cone Σn,2d under E is given by: Sq′ =((x1, . . . , xm) ∈ Rm for all k) . If we intersect the images P ′ and Sq′ with the hyperplane L = {x ∈ Rm Pm P ′ ∩ L is essentially just a simplex since Rm √xi ≥ 2√xk ++ ⊂ P ′, while Sq′ is a simplex with cut off corners. i=1 xi = 1}, then + m Xi=1 (cid:12)(cid:12)(cid:12) 3 The proofs for main theorems are obtained by analyzing the dual cone Σ∗ n,2d. Let K be a convex cone in a real vector space V . Its dual cone K ∗ is defined as the set of all linear functionals in the dual space V ∗ that are nonnegative on K: K ∗ = {ℓ ∈ V ∗ ℓ(x) ≥ 0 for all x ∈ K} . Let’s consider the dual space H ∗ n,2d of linear functionals on Hn,2d. To every linear functional n,2d we can associate a quadratic form Qℓ defined on Hn,d by setting ℓ ∈ H ∗ Qℓ(f ) = ℓ(f 2) for all f ∈ Hn,d. We classify the extreme rays of the dual cone Σ∗ n,2d which provides us with the description of all linear inequalities that define the sos cone. We prove the following theorems, which we think are interesting in themselves: Theorem 1.6. Suppose that ℓ spans an extreme ray of Σ∗ 3,6. Then rank Qℓ = 1 or rank Qℓ = 7. And for the case (4, 4): Theorem 1.7. Suppose that ℓ spans an extreme ray of Σ∗ We remark that in real analysis the functionals ℓ ∈ H ∗ n,2d are represented by their values on the monomial basis and are called truncated moment sequences. The matrix of the quadratic form Qℓ, when written with respect to the monomial basis of Hn,d has several names: it is called the moment matrix or Generalized Hankel matrix in real analysis, and symmetric catalecticant matrix in algebraic geometry. We prefer to keep a basis-free approach, but our results have interesting consequences when stated in terms of moment terminology. 4,4. Then rank Qℓ = 1 or rank Qℓ = 6. 2. Dual Cones Let Sn,d be the vector space of real quadratic forms on Hn,d. We can view the dual space H ∗ n,2d as a subspace of Sn,d by identifying the linear functional ℓ ∈ H ∗ n,2d with its quadratic form Qℓ defined by Qℓ(f ) = ℓ(f 2). If we choose the basis of monomials for Hn,2d then H ∗ n,2d is identified with the subspace of generalized Hankel matrices in Sn,d [13]. However, it is advantageous in our approach to not work with a fixed basis. Let S+ n,d be the cone of positive semidefinite forms in Sn,d: S+ n,d = {Q ∈ Sn,d Q(f ) ≥ 0 for all f ∈ Hn,d} . n,2d and S+ The following lemma is a well-known connection between Σ∗ n,d, that allows sums of squares problems to be solved by semidefinite programming. Viewed with the monomial basis it says that Σ∗ n,d with the subspace of generalized Hankel matrices, thus Σ∗ n,2d is the intersection of S+ n,2d is the Hankel spectrahedron. Lemma 2.1. The cone Σ∗ n,2d is the section of the cone of psd matrices S+ n,d with the subspace H ∗ n,2d: Σ∗ n,2d = S+ n,d ∩ H ∗ n,2d. Proof. Suppose that ℓ ∈ Σ∗ must be psd. Thus Σ∗ n,2d ⊆ S+ Now suppose that Qℓ ∈ S+ n,d ∩ H ∗ n,d ∩ H ∗ n,2d. Thus S+ n,d ∩ H ∗ n,2d ⊆ Σ∗ ℓ ∈ Σ∗ n,2d. Then ℓ(f 2) ≥ 0 for all f ∈ Hn,d. By definition of Qℓ we see that it n,2d. n,2d. Then it follows that ℓ(f 2) ≥ 0 for all f ∈ Hn,d and thus (cid:3) n,2d and the lemma follows. 4 We now need a general lemma about extreme rays of sections of the cone of positive semidefinite forms. Let S be the vector space of quadratic forms on a real vector space V . Let S+ be the cone of psd forms in S. The following lemma is from [12] (Corollary 4), we provide a proof for completeness: Lemma 2.2. Let L be a linear subspace of S and let K be the section of S+ with L: K = S+ ∩ L. Suppose that a quadratic form Q spans an extreme ray of K. Then the kernel of Q is maximal for all quadratic forms in L: if P ∈ L and ker Q ⊆ ker P then P = λQ for some λ ∈ R. Proof. Suppose not, so that there exists an extreme ray Q of K and a quadratic form P ∈ L such that ker Q ⊆ ker P and P 6= λQ. Since ker Q ⊆ ker P it follows that all eigenvectors of both Q and P corresponding to non-zero eigenvalues lie in the orthogonal complement (ker Q)⊥ of ker Q. Furthermore, Q is positive definite on (ker Q)⊥. It follows that Q and P can be simultaneously diagonalized to matrices Q′ and P ′ with the additional property that whenever the diagonal entry Q′ ii is also 0. Therefore, for sufficiently small ǫ ∈ R we have Q + ǫP and Q − ǫP are positive semidefinite and therefore Q + ǫP, Q − ǫP ∈ K. Thus Q is not an extreme ray of K, which is a contradiction. ii is 0 the corresponding entry P ′ (cid:3) Combining Lemma 2.1 and Lemma 2.2 we obtain the following corollary, which will be a critical tool for describing the extreme rays of Σ∗ n,2d: n,2d. Then either rank Q = 1, or the Corollary 2.3. Suppose that Q spans an extreme ray of Σ∗ forms in the kernel of Q have no common projective zeroes, real or complex. Proof. Let W ⊂ Hn,d be the kernel of Q and suppose that the forms in W have a common real zero v 6= 0. Let ℓ ∈ H ∗ n,2d be the linear functional given by evaluation at v: ℓ(f ) = f (v) for all f ∈ Hn,2d. Then Qℓ is a rank 1 positive semidefinite quadratic form and ker Q ⊆ ker Qℓ. By Lemma 2.2 it follows that Q = λQℓ and thus Q has rank 1. n,2d be the linear functional given by taking the real part of the value at z: ℓ(f ) = Re f (z) for all f ∈ Hn,2d. It is easy to check that the kernel of Qℓ includes all forms that vanish at z and therefore W ⊆ ker Qℓ. Therefore by applying Lemma 2.2 we again see that Q = λQℓ. However, we claim that Qℓ is not a psd form. Now suppose that the forms in W have a common nonreal zero z 6= 0. Let ℓ ∈ H ∗ The quadratic form Qℓ is given by Qℓ(f ) = Re f 2(z) for f ∈ Hn,d. However, there exists f ∈ Hn,d such that f (z) is purely imaginary and therefore Qℓ(f ) < 0. The Corollary now follows. (cid:3) We note that if we can find a nonzero psd quadratic form Qℓ such that the forms in its kernel Wℓ have no common real zeroes then ℓ will indeed provide a linear inequality that holds for all sos forms but fails for some psd forms. Since Qℓ is psd, we know that ℓ ∈ Σ∗ n,2d and we need to construct a nonnegative f ∈ Hn,2d such that ℓ(f ) < 0. Since forms in Wℓ have no common real zeroes we can find fi ∈ Wℓ such that q =Pi f 2 i ) = 0 for all i. Therefore ℓ(q) = 0 and q is strictly positive on the unit sphere. For sufficiently small ǫ > 0 we know that n)d) < 0. f = q−ǫ(x2 n,2d. For v ∈ Rn let ℓv be the linear functional in H ∗ n)d is nonnegative. On the other hand we have ℓ(f ) = −ǫℓ((x2 We will also need the following classification of all rank 1 forms in H ∗ i is strictly positive. We have Qℓ(fi) = ℓ(f 2 n,2d given by evaluation at v: 1+. . .+x2 1 +. . .+x2 and let Qv be the quadratic form associated to ℓv: Qv(f ) = f 2(v). In this case we say that Qv (or ℓv) corresponds to point evaluation. Note that the inequalities ℓv ≥ 0 are the defining inequalities 5 ℓv(f ) = f (v) for f ∈ Hn,2d, n,2d . n,2d. The following lemma shows that all rank 1 forms in H ∗ of P ∗ n,2d correspond to point evaluations. Since we are interested in the inequalities that are valid on Σn,2d but not valid on Pn,2d it allows us to disregard rank 1 forms Q ∈ H ∗ Lemma 2.4. Suppose that Q is a rank 1 form in H ∗ λ ∈ R. Proof. Let Q be a rank 1 form in H ∗ Therefore it suffices to show that if Q = s2(f ) for some s ∈ H ∗ n,2d and therefore ℓ(f 2) = s2(f ) for all f ∈ Hn,d. We have Q(f + g) = ℓ((f + g)2) = ℓ(f 2) + 2ℓ(f g) + ℓ(g2) = (s(f ) + s(g))2 = s2(f ) + 2s(f )s(g) + s2(g) and it follows that ℓ(f g) = s(f )s(g) for all f, g ∈ Hn,d. n . If we take monomials xα, xβ, xγ, xδ in Hn,d such that xαxβ = xγxδ then we must have s(xα)s(xβ) = s(xγ)s(xδ). n,2d. Then Q(f ) = λs2(f ) for some linear functional s ∈ H ∗ n,d then Q = Qv for some v ∈ Rn. n,d we know that Q is defined by Q(f ) = ℓ(f 2) for a linear functional ℓ ∈ H ∗ n,2d. Then Q = λQv for some v ∈ Rn and Let xα denote the monomial xα1 Since Q ∈ H ∗ 1 ··· xαn n,d. Suppose that s(xd x2 j ) = 0 and continuing in similar fashion we have s(xα) = 0 for all monomials. Then ℓ is the zero functional and Q does not have rank one. Contradiction. i ) = 0 for all i. Then we see that s(xd−1 xj)2 = s(xd i )s(xd−2 i i We may assume without loss of generality that s(xd we can work with −s, if necessary, and thus we may assume that s(xd for 1 ≤ i ≤ n. We will express s(xα) in terms of si for all xα ∈ Hn,d. Since (xd (xd−1 1)(xd−2 1 xixj) = sisj/s1. Continuing in this fashion we find that 1 xj) we have s(xd−2 1) 6= 0. Since we are interested in ℓ(f 2) = s2(f ) 1 xi) 1 xixj) = 1) > 0. Let si = s(xd−1 1 xi)(xd−1 s(xα1 1 ··· xαn n ) = sα2 2 ··· sαn n sd−1−α1 1 . Now let v ∈ Rn be the following vector , s v = (s1/d −(d−1)/d 1 s2, . . . , s −(d−1)/d 1 sn). 1 Let sv be the linear operator on Hn,d defined by evaluating a form at v: sv(f ) = f (v). Then we have sv(xd−1 1 xi) = si and sv(xα1 1 ··· xαn n ) = sα2 2 ··· sαn n sα1/d−(d−1)(d−α1 )/d 1 = sα2 2 ··· sαn n sd−1−α1 1 . Since s agrees with sv on monomials it follows that s = sv and thus ℓ(f 2) = s2(f ) = f (v)2 = (cid:3) f 2(v). Therefore ℓ indeed corresponds to points evaluation and we are done. 2.1. Kernels of Extreme Rays. Let Qℓ span an extreme ray of Σ∗ n,2d that does not correspond to point evaluation. Let Wℓ be the kernel of Qℓ and let J(ℓ) be the ideal generated by Wℓ. By Corollary 2.3 and Lemma 2.4 we know that the forms in Wℓ have no common projective zeroes real or complex, i.e. VC(Wℓ) = ∅. We now investigate the kernel Wℓ further. projective complex zeroes: Forms p1, . . . , pn ∈ Hn,d are said to form a sequence of parameters if they have no common VC(p1, . . . , pn) = ∅. It follows that we can find a sequence of parameters p1, . . . , pn ∈ Wℓ. Let I be the ideal generated by p1, . . . , pn. We will need the following theorem (special case of [6, Theorem CB8]): Theorem 2.5. Suppose that p1, . . . , pn ∈ Hn,d are a sequence of parameters and let I be the ideal generated by p1, . . . , pn in C[x1, . . . , xn]. Then I is a Gorenstein ideal with socle of degree n(d− 1). We also prove a simple but very useful characterization of kernels of forms Qℓ ∈ H ∗ n,2d: 6 Lemma 2.6. Let Qℓ be a quadratic form in H ∗ q ∈ Hn,d. Proof. In order to investigate Wℓ need to define the associated bilinear form Bℓ: n,2d. Then p ∈ Wℓ if and only if ℓ(pq) = 0 for all Bℓ(p, q) = Qℓ(p + q) − Qℓ(p) − Qℓ(q) 2 for p, q ∈ Hn,d. By definition of Qℓ we have Qℓ(p) = ℓ(p2). Therefore it follows that (2.1) A form p ∈ Hn,d is in the kernel of Qℓ if and only if Bℓ(p, q) = 0 for all q ∈ Hn,d. Using (2.1) the lemma follows. Bℓ(p, q) = ℓ(pq). (cid:3) We are now in position to prove Theorems 1.6 and 1.7, which we restate in a unified way: Theorem 2.7. Suppose that ℓ is an extreme ray of Σ∗ not correspond to point evaluation. Then rank of Qℓ is equal to dim Hn,d − n. Proof. Let p1, . . . , pn be a sequence of parameters in Wℓ and let I be the ideal generated by p1, . . . , pn. We claim that Wℓ = Id, or in other words, linear combinations of p1, . . . , pn gener- ate Wℓ. We note that this claim implies the desired Corollary, since it shows that the kernel of Qℓ has dimension exactly n. n,2d in the cases (3, 6) and (4, 4) and ℓ does By Theorem 2.5 we know that the socle of I has degree n(d − 1) = 2d in the cases (3, 6) and (4, 4). Suppose that Wℓ is strictly larger than Id. The ideal I is Gorenstein with socle of degree 2d, and hence J(ℓ)2d is strictly larger than I2d, which means that J(ℓ)2d = Hn,2d. It follows from Lemma 2.6 that f ∈ J(ℓ)2d. Therefore ℓ is the zero linear functional, which is a contradiction. ℓ(f ) = 0 for all (cid:3) Given an extreme ray Qℓ of Σ∗ Therefore instead of directly studying the extreme rays ℓ of Σ∗ n,2d that does not correspond to point evaluation we can pass to its kernel Wℓ and in the cases (3, 6) and (4, 4) the kernel Wℓ has dimension exactly n and further VC(Wℓ) = ∅. It follows from Theorem 2.5 that an n-dimensional subspace W with VC(W ) = ∅ uniquely determines (up to a constant multiple) the linear functional ℓ such that the kernel of Qℓ is W . The linear functional ℓ is the unique linear functional vanishing on the degree 2d part hWi2d of the ideal generated by W . This correspondence is a special case of the global residue map [4, §1.6]. n,2d we can look instead for n- dimensional subspaces W of Hn,d, with VC(W ) = ∅, whose corresponding linear functionals are extreme rays of Σ∗ n,2d have the defining property of being non- negative on squares. In order to see when a subspace W of Hn,d gives rise to an extreme ray of Σ∗ n,2d we need to get a handle on the linear functional ℓ ∈ H ∗ n,2d that W defines. We do this by passing to point evaluations. We need the following general Lemma, which allows us to extract a transverse zero-dimensional intersection from forms in W . Lemma 2.8. Suppose that p1, . . . , pn ∈ Hn,d are a sequence of parameters. Then there exist f1, . . . , fn−1 in the real linear span of pi such that the forms f1, . . . , fn−1 intersect transversely in dn−1 (possibly complex) points. n,2d. The linear functionals ℓ ∈ Σ∗ Proof. Let W be the linear span of p1, . . . , pn with complex coefficients. We begin by showing that there exist linear combinations f1, . . . , fn−1 ∈ W such that f1, . . . fn−1 intersect transversely in CPn−1. By Bertini’s theorem a general form in W is smooth. Let f1 be such a form. Let V1 be the smooth variety defined by f1 and let W1 be a subspace of W complementary to f1. Then W1 defines a linear 7 system of divisors on V1 and by Bertini’s Theorem the intersection of V1 with a general element of W1 is a smooth variety of dimension n − 2. Let f2 be such an element of W1. Now we can let V2 be the smooth variety defined by f1 and f2, let W2 be the complementary subspace to f1 and f2 and repeatedly apply Bertini’s Theorem until we get a 0-dimensional smooth intersection. Hence the forms f1, . . . , fn−1 we constructed intersect transversely. Suppose not and let fi =Pn Now we argue that there exist real linear combinations f1, . . . , fn−1 which intersect transversely. j=1 αijpj. Then for all αij ∈ R the forms fi do not intersect transversely. This is an algebraic conditions on the coefficients αij, given by vanishing of some polynomials in the variables αij. However, if a polynomial vanishes on all real points then it must be identically zero. Therefore, no complex linear combinations of pi intersect transversely, which is a contradiction. (cid:3) Now suppose that we have an n-dimensional subspace W of Hn,d with VC(W ) = ∅ and we locate f1, . . . , fn−1 ∈ W that intersect transversely. Let s = dn−1 and let Γ be the complex projective variety defined by f1, . . . , fn−1: Γ = VC(f1, . . . , fn−1) = {γ1, . . . , γs} ⊂ CPn−1 . Let S = {z1, . . . , zs} be a set of affine representatives for projective points γi. The functional ℓ determined by W is the unique linear functional vanishing on hWi2d. In particular ℓ vanishes on hf1, . . . , fn−1i2d. Since the forms f1, . . . , fn−1 intersect transversely, the ideal generated by fi is radical [8]. It follows therefore that ℓ can be expressed as a linear combination of evaluations at points vi: s s ℓ = µiℓvi; ℓ(p) = µip(vi), p ∈ Hn,2d. Xi=1 Xi=1 The coefficients µi are determined uniquely from any form fn so that f1, . . . , fn form a basis of W . In order to see how this occurs we need to introduce Cayley-Bacharach relations. 2.2. Cayley-Bacharach Relations. We now recall the Cayley-Bacharach theorem as applicable to ternary cubics and quaternary quadrics [6, Theorem CB6]: Lemma 2.9. For the cases (3, 6) and (4, 4) let f1, . . . , fn−1 ∈ Hn,d be forms intersecting transversely in s = dn−1 complex projective points γ1, . . . , γs. Let z1, . . . , zs be affine representatives of the projective points γi. Then there is a unique linear relation on the values of any form in Hn,d on zi: with nonzero ui ∈ C. u1p(z1) + . . . + usp(zs) = 0 for all p ∈ Hn,d, As we will see later evaluation on transverse intersections will be enough to distinguish nonneg- ative forms from sums of squares. Before we proceed with that we would like to show explicitly the geometry of values on transverse intersections, when all of the intersections points are real. 3. Cones of Point Evaluations Since the geometry of the cases (3, 6) and (4, 4) is very similar we will give a unified presentation. For these cases let f1, . . . , fn−1 be forms in Hn,d intersecting transversely in s = dn−1 real projective points γ1, . . . , γs. Let v1, . . . , vs ∈ Rn be arbitrary nonzero affine representatives for γ1, . . . , γs with vi corresponding to γi. Then by Lemma 2.9 there exists a unique linear relation u1p(z1) + . . . + usp(zs) = 0 for all p ∈ Hn,d and since we have all points vi ∈ Rn coming from intersection of real forms, all the coefficients ui must be real. 8 We first look at general real zero-dimensional intersections. Suppose that Γ = {γ1, . . . , γm} are real projective points that can be given as the complete set of common real zeroes of some forms f1, . . . , fk ∈ Hn,d: Γ = VR(f1, . . . , fk). Let S = {v1, . . . , vm} ⊂ Rn be a set of affine representatives for γi. Let E be the evaluation map that sends p ∈ Hn,2d to its values on the points vi: E : Hn,2d −→ Rm, E(p) = (p(v1), . . . , p(vm)). Note that E is defined on forms of degree 2d. Let P ′ and Sq′ be the images of Pn,2d and Σn,2d under E respectively and let H ′ be the image of Hn,2d. We observe that H ′ does not have to be all of Rm, since the values of forms in Hn,2d on points vi may be linearly dependent. Since we are evaluating nonnegative forms it follows that P ′ lies inside the intersection of H ′ and Rm + : The following theorem shows that this inclusion is almost an equality. P ′ ⊆ H ′ ∩ Rm + . Theorem 3.1. Let Rm orthant is contained in P ′: ++ be the positive orthant of Rm. The intersection of H ′ with the positive H ′ ∩ Rm ++ ⊂ P ′. Proof. Let s = (s1, . . . , sm) be a point in the intersection of H ′ and Rm ++. Since s ∈ H there exists a form p ∈ Hn,2d such that p(vi) = si. Let g = f 2 k . We claim that for large enough λ ∈ R the form ¯p = p + λg will be nonnegative, and since each fi is zero on S we will also have E (¯p) = s. By homogeneity of ¯p it suffices to show that it is nonnegative on the unit sphere Sn−1. Further- more, we may assume that the evaluation points vi lie on the unit sphere. Since we are dealing with forms, evaluation on the points outside of the unit sphere amounts to rescaling of the values on Sn−1. 1 + . . . + f 2 Let Bǫ(S) be the open epsilon neighborhood of S in the unit sphere Sn−1. Since p(vi) > 0 for all i, it follows that for sufficiently small ǫ the form p is strictly positive on Bǫ(S): p(x) > 0 for all x ∈ Bǫ(S). The complement of Bǫ(S) in Sn−1 is compact, and therefore we can let m1 be the minimum of g and m2 be the minimum of p on Sn−1 \ Bǫ(S). If m2 ≥ 0 then p itself was nonnegative and we are done. Therefore, we may assume m2 < 0. We also note that since g vanishes on S only, it follows that m1 is strictly positive. Now let λ ≥ − m2 . The form ¯p = p + λg is positive on Bǫ(S). By construction of Bǫ(S) we also see that the minimum of ¯p on the complement of Bǫ(S) is at least 0. Therefore ¯p is nonnegative on the unit sphere Sn−1 and we are done. (cid:3) m1 We obtain the following corollary for the cases (3, 6) and (4, 4): Corollary 3.2. Suppose that Γ comes from transverse intersection of two ternary cubics (resp. 3 quaternary quadrics) then R9 ++ ⊂ P ′ (resp. R8 ++ ⊂ P ′). Proof. We need to show that in our two cases the cone P ′ is full dimensional. This happens if and only if the values on the points vi are linearly independent for forms in H3,6 (resp. H4,4). This is an easy special case of Cayley-Bacharach Theorem [6, Theorem CB6]. (cid:3) We now show how the presence of a Cayley-Bacharach relation impacts the values attainable by sums of squares. Suppose now that the points v1, . . . , vm ∈ Rn are such that there exists a unique Cayley-Bacharach relation satisfied by all forms p ∈ Hn,d: u1p(v1) + . . . + ump(vm) = 0 with nonzero coefficients ui ∈ R. 9 We first describe Sq′ if the affine representatives are chosen so that the coefficients in the Cayley- Bacharach relation have absolute value 1. Let wi = ui1/dvi. Then p(wi) = uip(vi) for all f ∈ Hn,d. Thus we see that the values of forms in Hn,d on wi satisfy a unique relation a1f (w1) + . . . + amf (wm) = 0 with ai = ±1. Now redefine the map E using this particular set of affine representatives wi and let Sq′ be the image of Σn,2d under E. Let Tm be the subset of the nonnegative orthant Rm + defined by the following m inequalities: + Tm =((x1, . . . , xm) ∈ Rm + where Pm m Xi=1 (cid:12)(cid:12)(cid:12) √xi ≥ 2√xk for all k) . Lemma 3.3. The set Tm is a closed convex cone. Moreover, Tm is the convex hull of the points x = (x1, . . . , xm) ∈ Rm Proof. Tm is defined as a subset of Rm by the following 2m inequalities: xk ≥ 0 and √x1 + . . . + √xm ≥ 2√xk for all k. Therefore it is clear that Tm is a closed set. For x = (x1, . . . , xm) ∈ Rm + let x1/2 denote the L1/2 norm of x: i=1 √xi = 2√xk for some k. x1/2 = (√x1 + . . . + √xm)2. We can restate inequalities of Tm as xk ≥ 0 and x1/2 ≥ 4xk for all k. Now suppose that x, y ∈ Tm and let z = λx + (1 − λ)y for some 0 ≤ λ ≤ 1. It is clear that zk ≥ 0 for all It is known by the Minkowski inequality ([7] p.30) that L1/2 norm is a concave function: k. λx + (1 − λ)y1/2 ≥ λx1/2 + (1 − λ)y1/2. Therefore z1/2 ≥ λx1/2 + (1 − λ)y1/2 ≥ 4λxk + 4(1 − λ)yk = 4zk. Therefore Tm is a convex cone. To show that Tm is the convex hull of the points where x1/2 = 4xk for some k we proceed by induction. The base case m = 2 is simple since T2 is just a ray spanned by the point (1, 1). For the induction step we observe that any convex set is the convex hull of its boundary. For any point in the boundary of Tm one of the defining 2m inequalities must be sharp. If a point x is in the boundary of Tm and xi 6= 0 for all i then the inequalities xi ≥ 0 are not sharp at x and therefore the inequality x1/2 ≥ 4xk must be sharp for some k and we are done. If xi = 0 for some i then the point x lies in the set Tm−1 in the subspace spanned by the m − 1 standard basis vectors excluding ei and we are done by induction. (cid:3) Theorem 3.4. With the choices of affine representatives wi, so that the coefficients in the unique Cayley-Bacharach relation are of absolute value 1, we have Sq′ = Tm. Proof. By slight abuse of notation we will also use E as the evaluation map at wi for forms in Hn,d. Let a = (a1, . . . , am) be the vector of coefficients in the Cayley-Bacharach relation a1p(w1) + . . . + amp(wm) = 0 with ai = ±1 and f ∈ Hn,d. By uniqueness of the Cayley-Bacharach relation by know that L = E(Hn,d) is the hyperplane in Rm perpendicular to a. To show that Sq′ ⊆ Tm it suffices to show that E(q2) ∈ Tm for any q ∈ Hn,d. Let s = E(q) and t = E(q2). We know that E(q2) = (t1, . . . , tm) = (s2 m). By the Cayley- Bacharach relation we have a1s1 + . . . + amsm = 0 with ai = ±1. Without loss of generality, we may assume that s1 has the maximal absolute value among si. Multiplying the Cayley-Bacharach relation by −1, if necessary, we can make a1 = −1. Then we have s1 = a2s2 + . . . + amsm. We can now write √t1 = ±√t2 ± √t3 ± . . . ± √tm with the exact signs depending on ai and signs of si. Therefore we see that 2√t1 ≤ √t1 + . . . + √tm. Since s1 had the largest absolute value among si it follows that Sq′ ⊆ Tm. To show the reverse inclusion Tm ⊆ Sq′ we use Lemma 3.3. It suffices to show that all points in x ∈ Tm with 2√xk = √x1 + . . . + √xm for some k, are also in Sq′. Without loss of generality may 1, . . . , s2 10 assume that k = 1 and we have √x1 = √x2 + . . . + √xm. Let y = (y1, . . . , ym) with y1 = −√x1/a1 and yi = √xi/ai for 2 ≤ i ≤ m. It follows that a1y1 + . . . + amym = 0. Therefore y ∈ E(Hn,d) and y = E(q) for some quadratic form q. Then E(q2) = x and we are done. (cid:3) Note that this already proves Hilbert’s Theorem that there exist nonnegative polynomials that are not sums of squares for the cases (3, 6) and (4, 4). In fact Hilbert’s proof, by different methods, established that the standard basis vectors are not in Sq′, while we provide a complete description of Sq′. We now describe what happens if we do not rescale the affine representatives and the coefficients in the Cayley-Bacharach relation are arbitrary real numbers. Suppose that the unique Cayley- Bacharach relation satisfied by all forms f ∈ Hn,d is given by: u1f (v1) + . . . + umf (vm) = 0 with nonzero coefficients ui. Let E be the evaluation map at the points vi and let Sq′ be the image of Σn,2d under E. Corollary 3.5. The cone Sq′ is the subset of Rm + satisfying the following m inequalities: u1√x1 + . . . + um√xm ≥ 2uk√xk, for all 1 ≤ k ≤ m. Proof. Let a ∈ Rm be a vector with ai = ui/ui and let D be the diagonal m × m matrix with Dii = ui. Let La be the hyperplane of vectors in Rm perpendicular to a and Lu be the hyperplane of vectors perpendicular to u. The linear transformation ¯D sending x ∈ Rm to Dx sends Lu to La. Since the Cayley-Bacharach relation is unique it follows that Sq′ is the convex hull of the points (y2 1, . . . , y2 m) with y = (y1, . . . , ym) ∈ Lu. We have shown in Theorem 3.4 that the convex hull of squares from La is Tm. Since ¯D sends Lu to La it follows that ¯D2 sends Sq′ to Tm. + satisfying inequalities √x1 + . . . + √xm ≥ By Lemma 3.3 we know that Tm is the set of x ∈ Rm 2√xk for all 1 ≤ k ≤ m. Now suppose that x = D2y with x ∈ Tm and y ∈ Sq′. Then xi = ui2yi and it follows that y satisfies u1√x1 + . . . + um√xm ≥ 2uk√xk for all 1 ≤ k ≤ m. Since D2 is an invertible linear transformation (ui 6= 0) it follows that all y satisfying these inequalities are in Sq′. (cid:3) 4. Structure of Extreme Rays of Σ∗ 3,6 and Σ∗ 4,4 We now return to the study of extreme rays of Σ∗ n,2d for the cases (3, 6) and (4, 4). Let W be an n-dimensional linear subspace of Hn,d with VC(W ) = ∅. Let f1, . . . , fn−1 ∈ W be forms intersecting transversely in s = dn−1 points γ1, . . . , γs. Let z1, . . . , zs be affine representatives for points γi. By Lemma 2.9 there is a unique linear relation for values of forms in Hn,d on the points zi: u1f (z1) + ··· + usf (zs) = 0, for all f ∈ Hn,d. The unique (up to a constant multiple) linear functional ℓ vanishing on hWi2d can be written as a linear combination of point evaluations on the points zi: s ℓ = Xi=1 µiℓzi, for some µi ∈ C. Let fn be a form in W such that f1, . . . , fn form a basis of W . Note that the values fn(zi) are the same regardless of which fn ∈ W we choose. We now explain how to determine the coefficients µi from the knowledge of the Cayley-Bacharach coefficients ui and the values fn(zi). Lemma 4.1. µi = ui fn(zi) ; ℓ = ui fn(zi) s Xi=1 11 ℓzi i = 1 . . . s. Proof. Let ℓ be defined as above. We need to show that ℓ vanishes on all forms in hWi2d. We observe that for any form q ∈ Hn,d we have since ℓ is defined by values at common zeroes of f1, . . . , fn−1. Also ℓ(f1q) = ··· = ℓ(fn−1q) = 0 ℓ(fnq) = by the Cayley-Bacharach relation. ui fn(zi) s Xi=1 fn(zi)q(zi) = uiq(zi) = 0, s Xi=1 (cid:3) Since the forms f1, . . . fn−1 are real, the set Γ = {γ1, . . . , γs} is invariant under conjugation. Hence we can choose affine representatives zi so that the set S = {z1, . . . , zs} is invariant under conjugation. By uniqueness of the Cayley-Bacharach relation it follows that if zi = ¯zj then ui = ¯uj. We now show that if the functional ℓ is nonnegative on squares, then we can restrict the number of complex points zi, forcing most of the intersection points to be real. 4.1. Number of Complex Points. Suppose that S is a finite set of points in Cn that is invariant under conjugation: ¯S = S. Let S be given by S = {r1, . . . , rk, z1, . . . , zm, ¯z1, . . . , ¯zm} with ri ∈ Rn and zi, ¯zi ∈ Cn, zi 6= ¯zi. Let ℓ : Hn,2d → R be a linear functional given as a combination of evaluations on S: Xi=1 with λi ∈ R, µi ∈ C and αi, µi 6= 0. ℓ(p) = k λip(ri) + m Xi=1 (µip(zi) + ¯µip(¯zi)) , p ∈ Hn,2d Let ER : Hn,d → Rk+2m be the real evaluation projection of forms in Hn,d given by ER(p) = (p(r1), . . . , p(rk), Re p(z1), Im p(z1), . . . , Re p(zm), Im p(zm)) , p ∈ Hn,d. Let c be the dimension of the image of ER: c = dim ER(Hn,d). Lemma 4.2. Suppose that the quadratic form Qℓ is positive semidefinite. Then c ≤ k + m. Proof. The quadratic form Qℓ : Hn,d → R is defined by Qℓ(q) = λiq2(ri) + k Xi=1 m Xi=1(cid:0)µiq2(zi) + ¯µiq2(¯zi)(cid:1) , q ∈ Hn,d. Let ¯Qℓ be the quadratic form on Ck+2m given by: k m λix2 i + Xi=1 Xi=1 µi(cid:0)x2i−1 + √−1x2i(cid:1)2 + ¯µi(cid:0)x2i−1 − √−1x2i(cid:1)2 . By its definition, the form Qℓ is a composition on ER and ¯Qℓ: Qℓ = ¯Qℓ ◦ ER. Each of the 2 variable blocks µi(cid:0)x2i−1 + √−1x2i(cid:1)2 has one positive and one negative eigenvalue, since µi 6= 0. Therefore the form ¯Qℓ has at least m negative eigenvalues, and thus Qℓ is strictly negative on a subspace of dimension at least m. Recall that the form Qℓ is positive semidefinite, which implies that ¯Qℓ is psd on the image of (cid:3) + ¯µi(cid:0)x2i−1 − √−1x2i(cid:1)2 ER. Thus the image of ER has codimension at least m and the Lemma follows. 12 We can restate the Lemma as follows: note that S = k + 2m and S − c is the number of linearly independent relations that evaluation on S imposes on forms in Hn,d. Hence we get the following Corollary: Corollary 4.3. Suppose that Qℓ is positive semidefinite, then the number of complex conjugate pairs in S is at most equal to the number of linearly independent relations on S for forms of degree d. Applying this to transversal intersections in the cases (3, 6) and (4, 4) we get: Corollary 4.4. Suppose that ℓ is an extreme ray of Σ∗ n,2d that does not correspond to point eval- uation, and let f1, . . . , fn−1 be forms in the kernel Wℓ of Qℓ intersecting transversely in s = dn−1 points, Γ = {γ1, . . . , γs}. Then the set Γ includes at most 1 complex conjugate pair and the rest of the points in Γ are real. We now prove Theorem 1.1 and Theorem 1.2 in a unified manner. 5. Proofs of Main Theorems Proof of Theorems 1.1 and 1.2. Suppose that p ∈ Pn,2d and p is not sos. Then there exists an extreme ray ℓ of Σ∗ n,2d such that ℓ(p) < 0 and ℓ(q) ≥ 0 for all q ∈ Σn,2d. Since ℓ(p) < 0 and p is nonnegative it follows that ℓ does not correspond to point evaluation. Let Qℓ be the quadratic from associated with ℓ and let Wℓ be the kernel of Qℓ. Then by Theorem 2.7 we have dim Wℓ = n and by Lemma 2.8 we can find f1, . . . , fn−1 ∈ Wℓ intersecting transversely in s = dn−1 projective points γ1, . . . , γs. By Corollary 4.4 we know that at most two of γi are complex and by Lemma 4.1 the linear functional ℓ has the desired form. (cid:3) Proof of Corollaries 1.3 and 1.4. Let p be a strictly positive form on the boundary of Σn,2d. Then there exists an extreme ray ℓ of the dual cone Σ∗ n,2d, such that ℓ(p) = 0. Now suppose that It follows that Qℓ(fi) = 0 for all i and since Qℓ is a positive semidefinite quadratic form we see that all fi lie in the kernel Wℓ of Qℓ. By Theorem 2.7 we know that dim Wℓ = n and therefore p is a sum of squares of forms coming from a n dimensional subspace of Hn,2d. It follows that p is a sum of at most n squares. for some fi ∈ Hn,d. p = P f 2 i Now suppose that p is a sum of n− 1 or fewer squares, p = f 2 n−1 with some fi possibly zero. Since p is strictly positive we know that the forms fi have no common real zeroes. Therefore we found n − 1 forms fi ∈ Wℓ that have no common real zeroes (if p was a sum of fewer than n− 1 squares then we can add arbitrary fi to get their number up to n − 1). By the proof of Lemma 1.7 we know that n − 1 generic forms in Wℓ intersect transversely, and hence we can find forms i intersect transversely in dn−1 complex points. This is f ′ i ∈ Wℓ in a neighborhood of fi such that f ′ a contradiction by Corollary 4.4. (cid:3) 1 +··· + f 2 We now examine the two cases of Corollary 4.4: The case of fully real intersection and the case of one complex conjugate pair. In each of these cases there exist psd forms Qℓ corresponding to extreme rays of Σ∗ n,2d. We provide explicit equations of these extreme rays, based on the Cayley- Bacharach relation, thus giving us a complete description of the extreme rays of Σ∗ n,2d. 6. Fully Real Intersection We have already examined the difference between attainable values on fully real intersection for psd and sos forms in Section 3. Now we describe the dual picture of all the linear inequalities that come from fully real intersections, which hold on Σn,2d but fail on Pn,2d. Suppose that for the cases (3, 6) and (4, 4) a linear functional ℓ ∈ H ∗ n,2d spans an extreme ray of n2d that does not correspond to point evaluation. Let Wℓ be the kernel of Qℓ and suppose that Σ∗ 13 f1, . . . , fn−1 ∈ Wℓ intersect transversely in s = dn−1 real projective points γ1, . . . , γs. Let v1, . . . , vs be affine representatives for γ1, . . . , γs and let u1p(v1) + . . . + usp(vs) = 0 with ui ∈ R be the unique Cayley-Bacharach relation on the points vi. Theorem 6.1. The form Qℓ can be uniquely written as Qℓ(f ) = a1f (v1)2 + . . . + asf (vs)2 for f ∈ Hn,d, with a single negative coefficient ak, the rest of the ai positive and Furthermore any such form is extreme in Σ∗ s Xi=1 n,2d u2 i ai = 0. The key to the unified description in these cases is the uniqueness of the Cayley-Bacharach relation, which holds for both (3, 6) and (4, 4). Let E : Hn,d → Rs be the evaluation map that sends f ∈ Hn,d to its values at the points vi: E(f ) = (f (v1), . . . , f (vs)). Let L be the image of Hn,d under E. Since the forms in Hn,d satisfy a unique relation, it follows that L is the following hyperplane: L = {x ∈ Rs u1x1 + . . . + usxs = 0} . We would like to classify all positive semidefinite quadratic forms Qℓ on Hn,d with Qℓ = a1f 2(v1) + . . . + asf 2(vs), and coefficients ai ∈ R. By Lemma 4.1 the extreme rays ℓ of Σ∗ n,2d are guaranteed to have this form with points vi coming from transverse intersection of 2 cubics or 3 quadratics. In terms of the evaluations map we would like to find all quadratic forms Q : Rs → R given by Q = a1x2 1 +. . .+asx2 s that are positive semidefinite on the hyperplane L. Let SL be the cone of diagonal quadratic forms Q = a1x2 1+. . .+asx2 s that are positive semidefinite on the hyperplane L. Theorem 6.1 follows immediately from the following proposition: Proposition 6.2. Suppose that Q spans an extreme ray of SL. Then either Q = aix2 and ai > 0 or Q has the form specified in Theorem 6.1. Proof. Let Q = a1x2 then since Q spans an extreme ray it follows that Q = aix2 1 + . . . + asx2 i for some i and ai > 0. s span an extreme ray of SL. If all coefficients ai are nonnegative i for some i Suppose now that one of the coefficients ai is zero. Without loss of generality we may assume as = 0. Then we claim that Q = aix2 i for some i < s and ai > 0. First we show that all other coefficients must be nonnegative. Suppose that as = 0 and a1 < 0. From the equation of L we can write x1 = −(u2x2 + . . . + usxs)/u1. Therefore the form (u2x2 + . . . + usxs)2 Q = a1 + a2x2 2 + . . . + as−1x2 s−1 u2 1 is positive semidefinite for all values of x2, . . . , xs. However, the coefficient of x2 which is a contradiction. Therefore we can write Q = a1x2 spans an extreme ray it follows that Q = aix2 s is strictly negative, s−1 with ai ≥ 0. Since Q Next we claim that if one of ai is negative then the rest are strictly positive. Suppose that a1 < 0 2 + 1 + . . . + as−1x2 i for some i < s and ai > 0. and a2 ≤ 0. Then again write x1 = −(u2x2 + . . . + usxs)/u1 and Q = a1 . . . + asx2 u2 2 is strictly negative, which is a contradiction. s. Now the coefficient of x2 (u2x2+...+usxs)2 + a2x2 1 14 Now we have only one case left: one ai is negative and the rest are strictly positive. Suppose that as < 0. Write xs = −(u1x1 + . . . + us−1xs−1)/us and Q = a1x2 1 + . . . + as−1x2 s−1 + as u2 s (u1x1 + . . . + us−1xs−1)2 . Let’s maximize (u1x1+...+us−1xs−1)2 s−1 = 1. Applying Lagrange multipliers we see that xi = λui/ai for some λ and all i ≤ s − 1. Now we find the value of as that makes Q(u1/a1, . . . , us−1/as−1) = 0. We see that this happens for 1 + . . . + as−1x2 subject to a1x2 u2 s a⋆ s = s −u2 + . . . + 1 u2 a1 . u2 s−1 as−1 It is clear that any as ≥ a⋆ s then the form Q is positive definite on L and therefore it does not lie on the boundary of SL and does not span an extreme ray. s will result in a psd form Q. However, if as > a⋆ With as = a⋆ s the kernel of Q is spanned by the vector v = (u1/a1, . . . , us/as). We see that (up to a constant multiple) Q is the only form in SL with kernel that includes v. Therefore Q is extreme in SL. (cid:3) 7. One Complex Pair We now examine the last case of intersection with one complex conjugate pair of zeroes. Suppose that ℓ spans an extreme ray of Σ∗ n,2d that does not correspond to point evaluation. Let Wℓ be the kernel of Qℓ and suppose that f1, . . . , fn−1 ∈ Wℓ intersect transversely in s = dn−1 projective points γ1, . . . , γs with a single complex conjugate pair and the rest of γi real. Let v1, . . . , vs−2 be affine representatives for the real γi and z, ¯z be affine representatives for the complex roots chosen such that u1f (v1) + . . . + us−2f (vs−2) + f (z) + f (¯z) = 0, with ui ∈ R, is the unique Cayley-Bacharach relation on the points vi, z, ¯z. Theorem 7.1. The form Qℓ can be uniquely written as Qℓ(f ) = a1f (v1)2 + . . . + as−2f (vs−2)2 + 4m (Re z)2 − 4t (Im z)2 for f ∈ Hn,d, with all ai > 0 and m and t satisfying Furthermore any such form is extreme in Σ∗ 2m m2 + t2 + n,2d. u2 i ai s−2 Xi=1 = 0. Again we give a unified presentation based on the uniqueness of the Cayley-Bacharach relation. We construct the real evaluation map ER : Hn,d → Rs as follows: ER(f ) = (f (v1), . . . , f (vs−2), 2 Re f (z), 2 Im f (z)). Let L be the image of Hn,d under E. Then L is the following hyperplane: u1x1 + . . . + us−2xs−2 + xs−1 = 0} . L = {x ∈ Rs Note that L does not depend on xs = 2 Im f (z). We would like to classify all positive semidefinite quadratic forms Qℓ on Hn,d with Qℓ = a1f 2(v1) + . . . + as−2f 2(vs−2) + bf 2(z) + ¯bf 2(¯z), 15 Q = a1x2 1 + . . . + as−2x2 s−2 + m 2 (cid:0)x2 s−1 − x2 s(cid:1) − txs−1xs. and coefficients ai ∈ R and b ∈ C. By Lemma 4.1 the extreme rays ℓ of Σ∗ n,2d are guaranteed to have this form with points vi and z coming from transverse intersection of 2 cubics or 3 quadratics. Let b = m + t√−1. In terms of the evaluation map we would like to find all quadratic forms Q : Rs → R given by that are positive semidefinite on the hyperplane L. Let SL be the convex cone of all such quadratic forms. Proposition 7.2. Suppose that Q spans an extreme ray of SL. Then either Q = aix2 i for some i and ai > 0 or Q has the form specified by Theorem 7.1. Conversely, all such forms span extreme rays of SL. 1 + . . . + as−2x2 Proof. Let Q = a1x2 m = 0 then in order for Q to be psd on L we must have t = 0 and then all the coefficients ai are nonnegative and Q is a nonnegative combination of point evaluations. Since Q is extreme in SL it follows that Q = aix2 s(cid:1) − txs−1xs span an extreme ray of SL. If i for some i and ai > 0. 2 (cid:0)x2 s−1 − x2 s−2 + m s is not constrained by L and its coefficient is If m 6= 0 then it must be strictly negative since x2 −m/2. We can complete the square in Q and write Q = a1x2 1 + . . . + as−2x2 s−2 + m2 + t2 2m x2 s−1 − 1 2m (txs−1 + mxs)2. Since the term − 1 m x2 2m (txs−1 + mxs)2 is always nonnegative and xs is unconstrained, we can always make it equal to zero by taking xs = −txs−1/m. Therefore Q is psd if and only if Q′ = a1x2 1 + . . . + as−2x2 u1x1 + . . . + us−2xs−2 + xs−1 = 0}. We are in exactly the same situation as the case of fully real intersection and since the coefficient of xs−1 is guaranteed to be negative we know from Proposition 6.2 that all ai are positive and s−1 is psd on L′ = {x ∈ Rs−1 s−2 + m2+t2 m2 + t2 2m = 1 u2 a1 −1 + . . . + . u2 s−2 as−2 The resulting quadratic form will have a unique projective zero in L at v =(cid:18) u1 a1 , . . . , us−2 as−2 , 2m m2 + t2 , −2t m2 + t2(cid:19) . It is easy to verify that up to a constant multiple there is a unique form Q in SL with v in the (cid:3) kernel, which guarantees that Q is extreme and completes the proof. We would like to close the paper with a conjecture. Let Wℓ be the kernel of an extreme ray Qℓ of Σ∗ n,2d, in the cases (3, 6) and (4, 4), that does not correspond to point evaluation. It follows from Corollary 4.4 that any transversal intersection of n − 1 forms in the kernel Wℓ has at most one complex pair of zeroes. We have examined extreme rays that can be defined on such transverse intersections in Sections 6 and 7. However, we conjecture that the case of one complex pair of zeroes is not truly necessary, as we can always find a fully real intersection inside Wℓ: Conjecture 7.3. Suppose that for the cases (3, 6) and (4, 4), W is an n-dimensional subspace of Hn,2d such that VC(W ) = ∅ and any collection of forms f1, . . . , fn−1 intersecting transversely in W has at most 1 complex pair of zeroes. Then there exist forms f1, . . . , fn−1 ∈ W intersecting transversely in only real points. 16 acknowledgments The author would like to thank anonymous referees whose suggestions greatly helped in improving the exposition and clarity of the paper, and the hospitality of the Institute for Pure and Applied Math at UCLA, where the paper was written. References [1] G. Blekherman, There are significantly more nonnegative polynomials than sums of squares, Israel J. of Math., vol 183, 355-380, (2006). [2] G. Blekherman, J. Hauenstein, J. C. Ottem, K. Ranestad, B. Sturmfels,Algebraic Boundaries of Hilbert’s SOS Cones, submitted for publication, arXiv:1107.1846. [3] J. Bochnak, M. Coste, M.-F. Roy, Real Algebraic Geometry, Springer-Verlag, Berlin, (1998). [4] E. Cattani and A. Dickenstein: Introduction to residues and resultants, in Solving Polynomial Equations: Foun- dations, Algorithms, and Applications (eds. A. Dickenstein and I.Z. Emiris), Algorithms and Computation in Mathematics 14, Springer, 2005. [5] M.D. Choi, T.Y. Lam, B. Reznick Even symmetric sextics. Math. Z. 195, no. 4, 559-580, (1987). [6] D. Eisenbud, M. Green, J. Harris, Cayley-Bacharach theorems and conjectures, Bull. Amer. Math. Soc. vol. 33, no. 3, 295-324, (1996). [7] G. Hardy, E. Littlewood, G. Polya, Inequalities, Cambridge University Press, Cambridge, (1988). [8] J. Harris, Algebraic Geometry. A First Course, Graduate Texts in Mathematics, vol. 133. Springer-Verlag, New York, (1995). [9] J. Nie, M. Schweighofer, On the complexity of Putinar’s Positivstellensatz, J. of Complex., vol. 23, no. 1, 135-150, (2007). [10] J.B. Lasserre, Global optimization with polynomials and the problem of moments, SIAM J. Optim. vol. 11, no. 3, 796-817 (electronic), (2000/01). [11] P. Parrilo, Semidefinite programming relaxations for semialgebraic problems, Math. Program., vol. 96, no. 2, Ser. B, 293-320, (2000/01). [12] M. Ramana, A.J. Goldman, Some geometric results in semidefinite programming. J. Global Optim., vol. 7, no. 1, 33-50, (1995). [13] B. Reznick, Sums of Even Powers of Real Linear Forms, Mem. Amer. Math. Soc., vol. 96, no. 463, (1992). [14] B. Reznick, Some concrete aspects of Hilbert’s 17th Problem, Contemp. Math., no. 253, 251-272, (2000). [15] B.Reznick, On Hilbert’s construction of positive polynomials, arXiv:0707.2156. [16] R. Sanyal, F. Sottile, B. Sturmfels Orbitopes, Mathematika, vol. 57 (2011) 275-314. [17] R. Schneider, Convex Bodies: the Brunn-Minkowski Theory, Cambridge University Press, Cambridge, (1993). Grigoriy Blekherman, Georgia Inst. of Technology, Atlanta, USA, [email protected] 17
0912.2623	3	0912	2011-08-31T09:17:12	On the S-fundamental group scheme II	[ "math.AG" ]	The S-fundamental group scheme is the group scheme corresponding to the Tannaka category of numerically flat vector bundles. We use determinant line bundles to prove that the S-fundamental group of a product of two complete varieties is a product of their S-fundamental groups as conjectured by V. Mehta and the author. We also compute the abelian part of the S-fundamental group scheme and the S-fundamental group scheme of an abelian variety or a variety with trivial etale fundamental group.	math.AG	math	On the S-fundamental group scheme II ∗ Adrian Langer October 21, 2018 ADDRESS: 1. Institute of Mathematics, Warsaw University, ul. Banacha 2, 02-097 Warszawa, Poland, 2. Institute of Mathematics, Polish Academy of Sciences, ul. ´Sniadeckich 8, 00-956 Warszawa, Poland. e-mail: [email protected] Abstract The S-fundamental group scheme is the group scheme corresponding to the Tannaka category of numerically flat vector bundles. We use determinant line bundles to prove that the S-fundamental group of a product of two complete varieties is a product of their S- fundamental groups as conjectured by V. Mehta and the author. We also compute the abelian part of the S-fundamental group scheme and the S-fundamental group scheme of an abelian variety or a variety with trivial ´etale fundamental group. Introduction Let X be a complete, reduced and connected scheme defined over an algebraically closed field k. A vector bundle E on X is called numerically flat if both E and its dual E∗ are nef vector bundles. The category C nf(X ) of numerically flat vector bundles is a k-linear abelian rigid tensor category. Fixing a closed point x ∈ X endows C nf(X ) with a fiber functor E → E(x) which makes C nf(X ) into a neutral Tannaka category. Hence there is an equivalence between C nf(X ) and the category of represen- tations of some affine group scheme p S 1 (X , x) that we call an S-fundamental group scheme of X with base point x. The S-fundamental group scheme appeared in the curve case in [BPS] and then in general in [La2] and [Me]. The strategy of constructing a similar fundamental group scheme using another Tannaka category of essentially finite vector bundles goes back to Nori's influential paper [No1]. In this paper Nori defined a smaller group scheme, called Nori's fundamental group scheme, which is a pro-finite completion of the S-fundamental group scheme. Therefore the S-fundamental group ∗2000 Mathematics Subject Classification codes: 14J60; 14F05; 14L15 1 scheme plays with respect to Nori's fundamental group scheme a similar role as the fundamental group scheme for stratified sheaves with respect to the ´etale fundamental group scheme. In [La2] we exploited another interpretation of the S-fundamental group scheme in case of smooth projective varieties. Namely, numerically flat vector bundles are precisely strongly semistable torsion free sheaves with vanishing Chern classes. Using it one can apply vanishing theorems to establish, e.g., Lefschetz type theorems for the S-fundamental group scheme (see [La2, Theorems 10.2, 10.4 and 11.3]). This paper further exploits an observation that all the known properties of Nori's fundamental group scheme should still be valid for the more general S-fundamental group scheme. This allows in particular to obtain interesting properties of numerically flat bundles. As a main result of this paper we prove that the S-fundamental group of a product of two complete varieties is a product of their S-fundamental groups. This result was conjectured both in [La2] and [Me, Remark 5.12]. It implies in particular the corresponding result for Nori's fundamental group that was conjectured by Nori in [No1] and proven by Mehta and Subramanian in [MS]. Our proof of Nori's conjecture is completely different from that in [MS]. On the other hand, our proof gives also the corresponding result for the ´etale fundamental group scheme that was used in the proof of Mehta and Subramanian. As a corollary to our theorem we prove that the reduced scheme underlying the torsion com- ponent of the identity of the Picard scheme of a product of projective varieties is a product of the corresponding torsion components of its factors (see Corollary 4.7). The author does not know any other proof of this fact. The proof of the main theorem follows easily from the fact that the push forward of a numer- ically flat sheaf on the product X ×Y to X is also numerically flat (and in particular locally free). To prove this result we proceed by induction. In the curve case we employ some determinants of line bundles. In the remaining part of the paper we compute the abelian part of the S-fundamental group scheme (cf. [dS, Lemma 20 and Theorem 21] for a similar result for the fundamental group scheme for stratified sheaves). This allows us to compute the S-fundamental group scheme for abelian varieties (this can also be done using an earlier result of Mehta and Nori in [MN] for which we again have a completely different proof). The last part of the paper is based on [EM1] and [EM2] and it contains computation of the S-fundamental group for varieties with trivial ´etale fundamental group. 1 Preliminaries In this section we gather a few auxiliary results. 1.1 Numerical equivalence Let X be a complete, reduced and connected d-dimensional scheme defined over an algebraically closed field k. We say that a rank r locally free sheaf E on X is numerically trivial if for every coherent sheaf F on X we have c (X , F ⊗ E) = rc (X , F). 2 Now assume that X is smooth. Then we define the numerical Grothendieck group K(X )num as the Grothendieck group (ring) K(X ) of coherent sheaves modulo numerical equivalence, i.e., modulo the radical of the quadratic form given by the Euler characteristic (a, b) 7→ c (a · b) = RX ch(a) ch(b) td(X ). We say that a coherent sheaf has numerically trivial Chern classes if there exists an integer r such that the class [E] − r[OX] is zero in K(X )num. By the Riemann -- Roch theorem this is equivalent to the vanishing of numerical Chern classes. 1.2 Nefness Let us recall that a locally free sheaf E on a complete k-scheme X is called nef if and only if OP(E)(1) is nef on the projectivization P(E) of E. A locally free sheaf E is nef if and only if for any morphism f : C → X from a smooth projective curve C each quotient of f ∗E has a non-negative degree. We say that E is numerically flat if both E and E∗ are nef. A locally free sheaf E is nu- merically flat if and only if for any morphism f : C → X from a smooth projective curve C the pull-back f ∗E is semistable of degree zero. If X is projective then any numerically flat sheaf is numerically trivial and a line bundle is numerically flat if and only if it is numerically trivial. 1.3 Picard schemes Let X be an (integral) variety defined over an algebraically closed field k. Let Pic X denote the Picard group scheme of X. By Pic0 X we denote its connected component of the identity and by Pict X we denote the torsion component of the identity. Note that all these group schemes can be non-reduced. If X is projective then Pict X is of finite type and the torsion group Pict X / Pic0 X is finite. The scheme Pict X represents the functor of numerically trivial line bundles. If X is smooth and projective then Pict X is the fine moduli space of torsion free rank 1 sheaves with numer- ically trivial Chern classes (e.g., because by Theorem 2.2 every torsion free rank 1 sheaf with numerically trivial Chern classes is in fact a line bundle). 1.4 Cohomology and base change Let f : X → Y be a proper morphism of noetherian schemes and let E be a Y -flat coherent OX- module. Then we say that cohomology and base change commute for E in degree i if for every base change diagram v u / X f / Y X ×Y Y ′ g Y ′ 3 / / the natural map u∗Ri f∗E → Rig∗(v∗E) is an isomorphism. If for every point y ∈ Y the natural map Ri f∗E ⊗ k(y) → Hi(Xy, Ey) is an isomorphism then cohomology and base change commute for E in degree i. Indeed, our assumption implies that u∗Ri f∗E → Rig∗(v∗E) is an isomorphism at every point of Y ′. 1.5 Boundedness of numerically flat sheaves The following theorem is a corollary of the general boundedness result [La1, Theorem 4.4]. THEOREM 1.1. The set of numerically flat sheaves on a d-dimensional normal projective variety X is bounded. Proof. Let E be a rank r numerically flat sheaf on X. Let us fix a very ample line bundle OX (1) on X. Then we can write the Hilbert polynomial of E as P(E)(m) = c (X , E(m)) = d(cid:229) i=0 c (ET j≤iH j)(cid:18)m + i − 1 i (cid:19), where H1, ..., Hd ∈ OX (1) are general hyperplane sections. First let us note that c (ET j≤iH j) = rc (OT j≤iH j) for i = d, d − 1, d − 2. In cases i = d or d − 1 the assertion is clear since ET j≤iH j is a numerically flat vector bundle on a set of points or on a smooth curve, so the assertion follows from the usual Riemann -- Roch theorem. In the case when i = d − 2 it is sufficient to show that a rank r numerically flat sheaf F on a normal projective surface Y satisfies equality c (Y, F) = rc (Y, OY ). To prove this let us take a resolution of singularities f : Y → Y . Then c ( Y , f ∗F) = rc ( Y , O Y ), because f ∗F is a vector bundle with trivial Chern classes (as follows, e.g., from Theorem 2.2). But using the Leray spectral sequence H j( Y , Ri f∗F) ⇒ Hi+ j(Y, F) and the projection formula we see that Similarly, we have c ( Y , f ∗F) = c (Y, F) + c (Y, F ⊗ R1 f∗O Y ). c ( Y , O Y ) = c (Y, OY ) + c (Y, R1 f∗O Y ). Since R1 f∗O Y is supported on a finite number of points and F is locally free of rank r we see that c (Y, F ⊗ R1 f∗O Y ) = rc (Y, R1 f∗O Y ), which proves the required equality. Now let us write P(E)(m) = d(cid:229) i=0 ai(cid:18)m + d − i d − i (cid:19). Then our assertion implies that a0(E), a1(E) and a2(E) depend only on X. But the set of reflexive semistable sheaves with fixed a0(E), a1(E) and a2(E) on a normal projective variety is bounded by [La1, Theorem 4.4]. In case of smooth varieties the above theorem is an immediate corollary of Theorem 2.2 and [La1, Theorem 4.4]. 4 2 Fundamental groups in positive characteristic Let X be a complete connected reduced scheme defined over an algebraically closed field k. 2.1 S-fundamental group scheme Let C nf(X ) denote the full subcategory of the category of coherent sheaves on X, which as objects contains all numerically flat (in particular locally free) sheaves. Let us fix a k-point x ∈ X. Then we can define the fiber functor Tx : C nf(X ) → k-mod by sending E to its fiber E(x). One can show that (C nf(X ), ⊗, Tx, OX ) is a neutral Tannaka category (see [La2, Section 6]). Therefore by [DM, Theorem 2.11] the following definition makes sense: Definition 2.1. The affine k-group scheme Tannaka dual to this neutral Tannaka category is denoted by p S 1 (X , x) and it is called the S-fundamental group scheme of X with base point x. This group scheme was first defined in the curve case by Biswas, Parameswaran and Subra- manian in [BPS, Section 5], and then independently in [La2] and [Me]. The following characterization of numerically flat bundles as semistable sheaves with van- ishing Chern classes appears in [La2, Theorem 4.1 and Proposition 5.1]. THEOREM 2.2. Let X be a smooth projective k-variety of dimension d. Let H be an ample divisor on X and let E be a coherent sheaf on X. Then the following conditions are equivalent: 1. E is a strongly H-semistable torsion free sheaf and its Hilbert polynomial is the same as that of the trivial sheaf of the same rank. 2. E is a strongly H-semistable torsion free sheaf and it has numerically trivial Chern classes. 3. E is a strongly H-semistable reflexive sheaf with ch1(E) · Hd−1 = 0 and ch2(E) · Hd−2 = 0. 4. E is locally free, nef and c1(E)Hd−1 = 0. 5. E is numerically flat. 2.2 Nori's and ´etale fundamental group schemes Let us consider the category C N(X ) of bundles which are trivializable over a principal bundle under a finite group scheme. For a k-point x ∈ X we can define the fiber functor Tx : C N(X ) → k-mod by sending E to its fiber E(x). This makes C N(X ) a neutral Tannaka category which is equivalent to the category of representations of an affine group scheme p N 1 (X , x) called Nori's fundamental group scheme. If instead of C N(X ) we consider the category C ´et(X ) of bundles which are trivializable over a principal bundle under a finite ´etale group scheme then we get an ´etale fundamental group p ´et 1 (X , x). 5 Note that both these group schemes can be recovered from p S 1 (X , x) as inverse limits of some directed systems (see, e.g., [La2, Section 6]). In particular, any theorem proved for the S- fundamental group scheme implies the corresponding theorems for ´etale and Nori's fundamental group schemes. 2.3 The unipotent part of the S-fundamental group scheme The largest unipotent quotient of the S-fundamental group scheme of X is called the unipotent part of p S(X , x) and denoted by p U (X , x). By the standard Tannakian considerations we know that the category of finite dimensional k-representations of p U (X , x) is equivalent (as a neutral Tannakian category) to the category N C nf(X ) defined as the full subcategory of C nf(X ) whose objects have a filtration with all quotients isomorphic to OX . Nori proves that in positive characteristic p U (X , x) is a pro-finite group scheme (see [No1, Chapter IV, Proposition 3]) and hence p U (X , x) is the unipotent part of p N(X , x). This is no longer true in the characteristic zero case. However, in arbitrary characteristic we know that Hom(p U (X , x), Ga) = H1(X , OX). In particular, in characteristic zero the abelian part of p U (X , x) is equal to the group of the dual vector space H1(X , OX)∗ (see [No1, Chapter IV, Proposition 2]). In positive characteristic, the abelian part of p U (X , x) is the inverse limit of Cartier duals of finite local group subschemes of Pic X (cf. [No1, Chapter IV, Proposition 6]). 3 Numerically flat sheaves on products of curves In this section we keep the following notation. Let X and Y be complete k-varieties. Then p, q are the projections of X ×Y onto X and Y , respectively. Let E be a coherent sheaf on X ×Y . For a point y ∈ Y we set Ey = p∗(E ⊗ OX×{y}). Similarly, Ex = q∗(E ⊗ O{x}×Y ) for a point x ∈ X. Let us consider the following proposition in the special case when X and Y are curves. PROPOSITION 3.1. Let X and Y be smooth projective curves. Let F be a locally free sheaf on X ×Y , such that Fx1 is semistable for some x1 ∈ X. Assume that F is numerically trivial. Then for any closed points y1, y2 ∈ Y the corresponding locally free sheaves Fy1 and Fy2 are isomorphic. Moreover, the bundles Fx are semistable and S-equivalent for all closed points x ∈ X. Proof. Let us fix a point x1 ∈ X. By Faltings's theorem [Fa, Theorem I.2] (see also [Se, Remark 3.2 (b)]) there exists a rank r′ vector bundle E on Y such that H∗(Y, Fx1 ⊗ E) = 0. Note that this condition implies that E is semistable (see [Se, Theorem 6.2]). This bundle defines a global section Q E of a line bundle L = det p!(F ⊗ q∗E)−1. Set-theoretically, the zero set of this section is equal to {x ∈ X : H∗(Y, Fx ⊗ E) 6= 0}. But by the Grothendieck -- Riemann -- Roch theorem (see, e.g., [Ha, Appendix A, Theorem 5.3]) for every u ∈ K(Y ) we have ch(p!(F · q∗u)) = p∗(ch(F) · q∗(ch(u) · td(X ))) = rp∗(q∗(ch(u) · td(X ))) and the only non-zero part in the last term is in degree zero. Therefore L has degree 0. Since L has the section Q E non-vanishing at x1 it follows that L is trivial and H∗(Y, Fx ⊗ E) = 0 for all 6 x ∈ X. By [Se, Theorem 6.2] this implies that Fx is semistable for every x ∈ X. To prove that Fx are S-equivalent one can use determinant line bundles on the moduli space of semistable vector bundles on X. Since we do not use this fact in the following we omit the proof. An alternative proof can be found in [Se, Lemma 4.2]. The rest of the proof is similar to part of proof of [Fa, Theorem I.4] (see also Step 5 in proof of [He, Theorem 4.2]). Let us fix a point y1 ∈ Y and take any non-trivial extension 0 → E → E′ → Oy1 → 0. The Quot-scheme Q of rank 0 and degree 1 quotients of E′ is isomorphic to P(E′). Let us set Ep = kerp for a point [p : E′ → Oy] ∈ Q. The set U of points [p ] ∈ Q such that H∗(Y, Fx1 ⊗Ep ) = 0 is non-empty and open. The same arguments as before show that H∗(Y, Fx ⊗ Ep ) = 0 for all x ∈ X and [p ] ∈ U. Applying p∗ to the sequence 0 → F ⊗ q∗Ep → F ⊗ q∗E′ idF ⊗p −→ F ⊗ q∗Oy → 0 for [p : E′ → Oy] ∈ U we see that p∗(F ⊗ q∗E′) ≃ Fy. Therefore Fy1 ≃ Fy for points y in some non-empty open subset of Y . Since similar arguments apply to any other point y2 ∈ Y we see that Fy1 and Fy2 are isomorphic for all points y1, y2 ∈ Y . The following corollary is analogous to [Gi, Proposition 2.4] in case of stratified sheaves: COROLLARY 3.2. Let X be a normal complete variety and let Y be a complete variety. Let E be a numerically flat sheaf on X ×Y . Then for any closed points y1, y2 ∈ Y the corresponding locally free sheaves Ey1 and Ey2 are isomorphic. In particular, the sheaf q∗E is locally free. Proof. Using Chow's lemma it is easy to see that there exists an irreducible curve on Y containing both y1 and y2. Taking its normalization we can replace Y by a smooth projective curve and prove the assertion in this special case. So in the following we assume that Y is a smooth projective curve. Now we prove the assertion assuming that X is projective. The proof is by induction on the dimension d of X. For d = 1 the assertion follows from Proposition 3.1. For d ≥ 2 let us fix a divisor D on X and consider the following exact sequence Hom(Ey1, Ey2) a −→ Hom(Ey1D, Ey2D) −→ H1(H om(Ey1, Ey2) ⊗ OX (−D)). Taking a sufficiently ample divisor D, we can assume that D is smooth and H1(H om(Ey1, Ey2) ⊗ OX (−D)) = 0 (here we use d ≥ 2; see [Ha, Chapter III, Corollary 7.8]). But then the map a is surjective and the isomorphism Ey1D ≃ Ey2D (coming from the inductive assumption) can be lifted to a homomorphism j is injective at the points of D and Ey1 is torsion-free, it follows that j : Ey1 → Ey2. Since j is an injection. But then it must be an isomorphism. To prove the assertion in case when X is non-projective we can use Chow's lemma. Namely, there exists a normal projective variety X and a birational morphism f : X → X. Let us set g = f ×idY : X ×Y → X ×Y . By the previous part of the proof we know that f ∗(Ey1) = (g∗E)y1 ≃ (g∗E)y2 = f ∗(Ey2). But by Zariski's main theorem we know that f∗O X = OX. Hence by the projection formula we have Ey1 ≃ f∗ f ∗(Ey1) ≃ f∗ f ∗(Ey2) ≃ Ey2. The last part of the corollary follows from Grauert's theorem (see [Ha, Chapter III, Corollary 12.9]). 7 4 S-fundamental group scheme of a product The following result was conjectured both by the author in [La2, Section 8] and by V. Mehta in [Me, Remark 5.12]: THEOREM 4.1. Let X and Y be complete k-varieties. Let us fix k-points x0 ∈ X and y0 ∈ Y . Then the natural homomorphism p S 1 (X ×k Y, (x0, y0)) → p S 1 (X , x0) ×k p S 1 (Y, y0) is an isomorphism. Proof. In characteristic zero the assertion follows from the corresponding fact for topological fundamental groups of complex varieties and the Lefschetz principle. More precisely, in the case of complex varieties the assertion follows from the corresponding isomorphism of topological fundamental groups by passing to a pro-unitary completion. In general, the fact follows from the Lefschetz principle if one notes that Theorem 4.1 follows from Lemma 4.2 and if we use the Lefschetz principle for this special assertion (note that to apply Lefschetz principle we need to reformulate Theorem 4.1 as it involves group schemes which are not of finite type over the field). So in the following we can assume that the characteristic of k is positive (but we need it only to prove the next lemma which is the main ingredient in proof of Theorem 4.1). LEMMA 4.2. Let us assume that X and Y are normal and projective. Let E be a numerically flat sheaf on X ×Y . Then p∗E is numerically flat. Proof. By Corollary 3.2 p∗E is locally free. Let us fix a point x0 ∈ X. Then the sheaf Gn = p∗((Fn X )∗(p∗E) is locally free of rank a = h0(Y, Ex0). X × idY )∗E) = (Fn Now let us consider the set A of all numerically flat sheaves on X ×Y . This set is bounded by Theorem 1.1, so there exist a scheme S of finite type over k and an S-flat sheaf E on X ×Y × S such that the set of restrictions {Es}s∈S contains all the sheaves in the set A. Let us consider a subscheme S′ ⊂ S defined by S′ = {s ∈ S : h0(Y, (Es)x0) ≥ a}. By semicontinuity of cohomology, S′ is a closed subscheme of S. Let us consider an open subset U ⊂ S′ that corresponds to points s ∈ S′ where h0(Y, (Es)x0) = a. We consider U with the reduced scheme structure. By abuse of notation, the restriction of E to U will be again denoted by E . We claim that the set {p∗(Es)}s∈U is a bounded set of sheaves. To prove this let us consider the following diagram Y px Speck jx ix j(x,s) X ×Y p / X i(x,s) js is X ×Y ×U p / X ×U in which the vertical maps are canonical projections and the horizontal maps are embeddings cor- responding to fixed points x ∈ X and s ∈ U. Let us recall that p∗( j∗ s E ) is locally free by Corollary 8 & & / / / / 8 8 / / 3.2 and the definition of U. Moreover, p∗E is locally free as h0(Y, (Es)x) = h0(Y, (Es)x0) = a for every x ∈ X by Corollary 3.2. Therefore the above diagram induces the commutative diagram x p∗( j∗ i∗ s E ) / (px)∗ j∗ x( j∗ s E ) ≃ i∗ (x,s)( p∗E ) (px)∗ j∗ (x,s) E in which the horizontal maps are isomorphisms by Grauert's theorem (see [Ha, Chapter III, Corollary 12.9]). Hence the remaining vertical map is also an isomorphism. This shows that ( p∗E )s ≃ p∗(Es), which gives the required claim. By our assumptions there exists a sequence (sn)n∈N of points of U such that p∗(Esn) ≃ Gn. Therefore the set {Gn}n∈N is bounded. Now let us take any morphism f : C → X from a smooth projective curve C. Since (Fn C )∗( f ∗p∗E) ≃ f ∗(Gn) and the set { f ∗(Gn)}n∈N is bounded, we see that f ∗ p∗E is semistable of degree 0. More precisely, semistability follows from the fact that sequences of slopes of maximal destabilizing subsheaves and minimal destabilizing quotients of Frobenius pull-backs of f ∗ p∗E are bounded. Similarly, the sheaf f ∗ p∗E has degree 0 since its Frobenius pull-backs have bounded degree. Therefore p∗E is numerically flat. 1 (X ×k Y, (x0, y0)) → p S Now we can go back to the proof of Theorem 4.1. The first part of proof is the same as the analogous part of proof of [No1, Chapter IV, Lemma 8]. Namely, the homomorphism p S 1 (Y, y0) is induced by projections p : X ×Y → X and q : X ×Y → Y . Let i : X → X ×Y be the embedding onto X × {y0} and let j : Y → X ×Y be the embedding onto {x0} ×Y . Since p i = idX and q i is constant, the composition p S 1 (Y, y0) is an embedding onto the first component. Similarly q j = idY is the embedding onto the second component, so the homomorphism p S 1 (Y, y0) can be split and in particular it is faithfully flat. 1 (X ×k Y, (x0, y0)) → p S 1 (X ×k Y, (x0, y0)) → p S 1 (X , x0) ×k p S 1 (X , x0) → p S 1 (X , x0) ×k p S 1 (X , x0) ×k p S Hence we only need to prove that it is a closed immersion. Let us first assume that X and Y are normal and projective. Note that the sheaf F = H om(q∗Ex0, E) is numerically flat. By our assumptions and Lemma 4.2 the sheaf p∗F is numerically flat. The induced map p∗ p∗F ⊗ q∗Ex0 → E is surjective as its restriction to {x0} × Y corresponds to the surjective map Hom(Ex0, Ex0) ⊗ Ex0 → Ex0. Now [DM, Proposition 2.21 (b)] implies that the natural homomorphism p S 1 (X ×k Y, (x0, y0)) → p S 1 (Y, y0) is a closed immersion and therefore it is an isomorphism. 1 (X , x0) ×k p S Now we need the following lemma: LEMMA 4.3. Let X and Y be complete k-varieties such that the natural map p S 1 (X ×kY, (x0, y0)) → p S 1 (X , x0) ×k p S 1 (Y, y0) is an isomorphism. Let E be a numerically flat sheaf on X ×Y . Then for every integer i the sheaf Ri p∗E is numerically flat and cohomology and base change commute for E in all degrees. 9 / / / O O O O 1 (X ×kY, (x0, y0)) ≃ p S Proof. Since p S 1 (Y, y0), E is a subsheaf of a sheaf p∗E0 ⊗q∗F0 for some numerically flat sheaves E0 and F0. This is an easy fact from representation theory as every G1 ×G2-module is a submodule of the tensor product of G1 and G2-modules. Alternatively, [DM, Proposition 2.21 (b)] implies that every numerically flat sheaf on X ×Y is a quotient of a sheaf of the form p∗E0 ⊗ q∗F0, which implies the required statement by taking the duals. 1 (X , x0) ×k p S The quotient (p∗E0 ⊗ q∗F0)/E is also numerically flat, so it is a subsheaf of a sheaf of the form p∗E1 ⊗ q∗F1 for some numerically flat sheaves E1 and F1. Inductively we can therefore construct the following acyclic complex of sheaves on X ×Y : (∗) 0 → E → p∗E0 ⊗ q∗F0 → p∗E1 ⊗ q∗F1 → ... → p∗Ei ⊗ q∗Fi → ... Let us set C i = p∗Ei ⊗ q∗Fi. Note that Ri p∗C j ≃ E⊕hi(Y,Fj) following spectral sequence j is numerically flat and consider the Ei j 1 = Ri p∗C j =⇒ Ri+ j p∗C • ≃ Ri+ j p∗E. Since the category of numerically flat sheaves on X is abelian, kernels and cokernels of objects from this category are also numerically flat. This implies that the limit Ri+ j p∗E of the above spectral sequence is also numerically flat. Now note that the complex (∗) restricted to {x} ×Y remains acyclic, as all the sheaves in this complex are locally free. Therefore we have a commutative diagram of spectral sequences Ri p∗C j ⊗ k(x) =⇒ Ri+ j p∗E ⊗ k(x) Hi(Y, C j x ) =⇒ Hi+ j(Y, Ex). Since C i = p∗Ei ⊗ q∗Fi, the left vertical map in this diagram is an isomorphism. Therefore the right vertical map is also an isomorphism. But this implies that cohomology and base change commute for E in all degrees (see 1.4). Now let us return to the proof of Theorem 4.1. Let us first assume that X is normal and projective. By Chow's lemma there exists a projective variety Y and a birational morphism f : Y → Y . Passing to the normalization, we can assume that Y is normal. Consider the base change diagram X × Y q Y g f X ×Y . q / Y Let us take two closed points y1, y2 ∈ Y . Let us choose closed points y1, y2 ∈ Y mapping onto y1, y2, respectively. Then h0((g∗E) y1) = h0((g∗E) y2) and hence h0(Ey1) = h0(Ey2). By Grauert's theorem (and 1.4) this implies that q∗E is locally free and cohomology and base change commute 10 / / / for E in all degrees. In particular, we have and f ∗(q∗E) ≃ q∗(g∗E). But q∗(g∗E) is numerically flat, so q∗E is also numerically flat. In this case the same proof as in the previous case shows that p S 1 (X ×kY, (x0, y0)) → p S 1 (X , x0)×k p S 1 (Y, y0) is an isomorphism. Now we can again apply Chow's lemma to prove that if X and Y are complete and E is numerically flat on X ×Y then p∗E is numerically flat. As in the previous case this implies that p S 1 (X ×k Y, (x0, y0)) → p S 1 (Y, y0) is a closed immersion. 1 (X , x0) ×k p S Remark 4.4. Most of the proof of Theorem 4.1 works in an arbitrary characteristic. But the proof of Lemma 4.2 uses positive characteristic. The characteristic zero version of Theorem 4.1 would follow from the positive characteristic case if one knew that the reduction of a semistable com- plex bundle for some characteristic is strongly semistable (this is a weak version of Miyaoka's problem; see [Miy, Problem 5.4]). This problem seems to be open even for numerically flat bundles on a product of two curves of genera ≥ 2. Applying [La2, Lemma 6.3] as a corollary to Theorem 4.1 we obtain the following result of Mehta and Subramanian (conjectured earlier by Nori in [No1]). COROLLARY 4.5. (see [MS, Theorem 2.3]) Let X and Y be complete k-varieties. Let us fix k-points x0 ∈ X and y0 ∈ Y . Then the natural homomorphism p N 1 (X ×k Y, (x0, y0)) → p N 1 (X , x0) ×k p N 1 (Y, y0) is an isomorphism. Note that Theorem 4.1 and Lemma 4.3 imply the following corollary. COROLLARY 4.6. Let X and Y be complete k-varieties and let E be a numerically flat sheaf on X ×Y . Then for every integer i the sheaf Ri p∗E is numerically flat and cohomology and base change commute for E in all degrees. COROLLARY 4.7. Let X and Y be projective k-varieties. Then and t (Pic t (X ×Y ))red ≃ (Pic t X )red × (Pic Y )red (Pic0(X ×Y ))red ≃ (Pic0X )red × (Pic0Y )red. Proof. The category of representations of the S-fundamental group scheme p S 1 (X , x) is equivalent to the category C nf(X ) of numerically flat sheaves. Since a line bundle is numerically flat if and only if it is numerically trivial, the group (Pict (X ))red is the group of characters of p 1(X , x) (here we use projectivity of X). Now the first isomorphism follows directly from Theorem 4.1. The second isomorphism follows immediately from the first one (if X and Y are smooth projective varieties one can also give another proof using comparison of dimensions of Pic0(X × Y ) and Pic0 X × Pic0Y ). 11 5 The abelian part of the S-fundamental group scheme We say that a numerically flat sheaf is irreducible if it does not contain any proper numerically flat subsheaves (or, equivalently, if it corresponds to an irreducible representation of p S 1 (X , x)). In case of projective varieties a numerically flat sheaf is irreducible if and only if it is slope stable (with respect to some fixed polarization; or, equivalently, with respect to all polarizations). If E is an irreducible numerically flat sheaf then it is simple. This follows from the fact that any endomorphism of such E is either 0 or an isomorphism (otherwise the image would give a proper numerically flat subsheaf). THEOREM 5.1. Let E be an irreducible numerically flat sheaf on a product X ×Y of complete varieties X and Y . Then there exist irreducible numerically flat sheaves E1 on X and E2 on Y such that E ≃ p∗E1 ⊗ q∗E2. Proof. Let us fix a point x0 ∈ X and an irreducible numerically flat subsheaf K ⊂ Ex0. Let us set F = H om(p∗K, E). Since the sheaf F is numerically flat, by Corollary 4.6 we know that q∗F is numerically flat. It is easy to see that the induced map j : q∗q∗F ⊗ p∗K → E is injective on {x0} ×Y . Since (q∗F) ⊗ k(x0) = Hom(K, Ex0) is non-zero, q∗q∗F ⊗ p∗K is nu- merically flat and E is irreducible, j is an isomorphism. The following lemma follows from the proof of Theorem 5.1 but we give a slightly different proof without using Theorem 4.1 (so in particular the proof is completely algebraic) in a case sufficient for the applications in the next section. LEMMA 5.2. Let X and Y be complete varieties. Let E be a numerically flat sheaf on X ×k Y . Assume that for some point y0 ∈ Y , Ey0 is simple (i.e., End(Ey0) = k). Then there exists a numerically trivial line bundle L on Y such that E ≃ p∗Ey0 ⊗ q∗L. Proof. For simplicity let us assume that X is smooth and projective. Let us set F = H om(p∗Ey0, E). Since the sheaf F is numerically flat, by Corollary 3.2 we know that q∗F is locally free. Therefore the induced map q∗q∗F ⊗ p∗Ey0 → E is surjective on all fibers {x} ×Y and hence it is surjective. Since Ey0 is simple, L = q∗F is a line bundle. Therefore the above map is a surjective map of locally free sheaves of the same rank and hence it is an isomorphism. This also shows that L is numerically trivial. PROPOSITION 5.3. Let X and Y be complete varieties. Let E be a numerically flat sheaf on X ×k Y . Assume that for some point x0 ∈ X and y0 ∈ Y , Ex0 and Ey0 are simple. Then both Ex0 and Ey0 are line bundles and E ≃ p∗Ey0 ⊗ q∗Ex0. Proof. By the above lemma we know that E ≃ p∗Ey0 ⊗ q∗L for some line bundle L. Therefore Ex0 ≃ q∗((p∗Ey0 ⊗ q∗L) ⊗ p∗Ox0) ≃ L⊕rk Ey0 . Since Ex0 is simple, Ey0 is a line bundle and Ex0 ≃ L. 12 The following definition is an analogue of [Gi, Definition 2.5] where the corresponding no- tion is defined for stratified sheaves. Definition 5.4. Let E be a numerically flat sheaf on a complete variety X /k. We say that E is abelian if there exists a numerically flat sheaf E′ on X ×k X and a closed point x0 ∈ X such that E′ restricted to both p−1(x0) and q−1(x0) is isomorphic to E. Proposition 5.3 implies that a simple abelian numerically flat sheaf on a complete variety has rank one. Since every numerically flat sheaf has a filtration with irreducible quotients, we get the following corollary describing the category of representations of p S ab(X , x). COROLLARY 5.5. Let E be an abelian numerically flat sheaf on a complete variety X. Then E has a filtration 0 = E0 ⊂ E1 ⊂ . . . ⊂ Er = E in which all quotients Ei/Ei−1 are numerically trivial line bundles. Moreover, if E is simple then it is a line bundle. Note that the last part of the corollary is non-trivial as even on complex projective varieties, a simple, semistable bundle need not be stable. Remark 5.6. In the next section we will see that every numerically flat sheaf on an abelian variety is abelian. Therefore the above corollary generalizes [MN, Theorem 2] which describes numerically flat sheaves on abelian varieties assuming boundedness of semistable sheaves with fixed numerical invariants (this is proven in [La1]). Mehta and Nori proved this theorem using reduction to the case of abelian varieties defined over an algebraic closure of a finite field. Our proof is completely different. LEMMA 5.7. Let G be an affine k-group scheme. Then the category of representations of the largest abelian quotient Gab of G is isomorphic to the full subcategory of the category of repre- sentation of G, whose objects are those G-modules V for which there exists a G × G-module W such that we have isomorphisms of G-modules W G×{e} ≃ W {e}×G ≃ V . Proof. We give a naive proof in the case of algebraic groups. The reader is asked to rewrite it in terms of group schemes using a precise definition of derived subgroup scheme (see [Wa, 10.1]). Let V be a Gab-module. Since Gab is abelian, the multiplication map Gab × Gab → Gab is a homomorphism of affine group schemes. Hence we can treat V as a G × G-module obtaining the required module W . Now let V be a G-module for which there exists r G-modules W G×{e} ≃ W {e}×G ≃ V . The representation r given by r (g) = r (g, e) = r (e, g). Since : G × G → GL(W ) and isomorphisms of : G → GL(V ) corresponding to V is r (g, h) = r (g, e) · r (e, h) = r (e, h) · r (g, e), we have r (ghg−1h−1) = e. Therefore r vanishes on the derived subgroup of G and hence it defines the required representation of Gab. 13 The largest abelian quotient of the S-fundamental group scheme of X is called the abelian part of p S(X , x) and denoted by p S ab(X , x). Lemma 5.7, together with Theorem 4.1, implies that p S ab(X , x) is Tannaka dual to the (neu- tral Tannakian) category of abelian numerically flat sheaves on X. This also explains the name "abelian" in Definition 5.4. Remark 5.8. Note that Corollary 5.5 follows immediately from the interpretation of abelian nu- merically flat bundles given above and the fact that irreducible representations of abelian group schemes are one-dimensional (see [Wa, Theorem 9.4]). In fact, this also proves that all sub- sheaves Ei in the filtration are abelian. Let G be an abelian group. Then we define the k-group scheme Diag(G) by setting (Diag(G))(SpecA) = Homgr(G, A×) for a k-algebra A, where Homgr(G, A×) denotes all group homomorphisms from G to the group of units in A. This is the same as taking Spec k[G] with the natural k-group scheme structure. THEOREM 5.9. Let X be a smooth projective variety defined over an algebraically closed field k. If the characteristic of k is positive then we have an isomorphism p S ab(X , x) ≃ lim ← t G × Diag((Pic X )red), where G denotes the Cartier dual of G and the inverse limit is taken over all finite local group subschemes G of Pic0 X. If k has characteristic zero then we have t p S ab(X , x) ≃ H1(X , OX)∗ × Diag(Pic X ). Proof. By [Wa, Theorem 9.5] p S ab(X , x) is a product of its unipotent and multiplicative parts. By Nori's results mentioned in Subsection 2.3 (see [No1, Chapter IV, Proposition 6]) we know that in positive characteristic the abelian part of the unipotent part (or equivalently, the unipotent part of the abelian part) p U (X , x) of the S-fundamental group is isomorphic to the inverse limit of Cartier duals of finite local group subschemes of PicX. In characteristic zero, p U ab(X , x) = H1(X , OX)∗. underlying Pict X (in characteristic zero Pict X is already reduced), the diagonal part of p S is given by Diag((Pict On the other hand, since the character group of p S(X , x) is isomorphic to the reduced scheme ab(X , x) X )red), which finishes the proof. Let us recall that the Neron -- Severi group NS(X ) = (PicX )red/(Pic0 X )red is finitely gener- ated. We have a short exact sequence t 0 → (Pic0X )red → (Pic X )red → NS(X )tors → 0, where NS(X )tors is the torsion group of NS(X ) (it is a finite group). Therefore we have a short exact sequence t 0 → Diag(NS(X )tors) → Diag((Pic X )red) → Diag((Pic0X )red) → 0. 14 6 Numerically flat sheaves on abelian varieties Let A be an abelian variety defined over an algebraically closed field of characteristic p and let An be the kernel of the multiplication by n map nA : A → A. Let Ar pn be the reduced part of Apn. Then Apn is a product of Ar pn, its Cartier dual Ar pn ≃ (Z/pnZ)r, Ar The p-adic discrete Tate group T d pn (which is a local and diagonalizable group scheme) and a local -- local group scheme A0 pn (see [Mu, Section 15]; local -- local means that both the group scheme and its Cartier dual are local). Let r be the pn is a group scheme of order p2(dim A−r)n. p-rank of A. Then Ar pn (see [Mu, Section p (A) as the pn. For an arbitrary prime l (possibly l = p) we also define the l-adic Tate 18]). This is a Zp-module. Similarly, we define the p-adic local -- local Tate group T 0 inverse limit lim ←− group scheme as Tl(A) = lim ←− pn ≃ (m pn)r and A0 p (A) is defined as the inverse limit lim ←− Aln (note that for l = p our notation differs from that in [Mu]). A0 Ar The following theorem is analogous to [dS, Theorem 21]: THEOREM 6.1. Let A be an abelian variety defined over an algebarically closed field k. Then p S(A, 0) is abelian and it decomposes as a product of its unipotent and diagonal parts. Moreover, if the characteristic of k is positive then p S(A, 0) ≃ T 0 p (A) × Diag(Pic0A). p (A) × T d In characteristic zero we have p S(A, 0) ≃ H1(A, OA)∗ × Diag(Pic0A). Proof. Let us first remark that every numerically flat bundle E on A is abelian. To show this one can take the addition map m : A ×k A → A. Then E′ = m∗E restricted to either p−1(0) or q−1(0) is isomorphic to E. This shows that p S(A, 0) is abelian and we can use Theorem 5.9. In positive characteristic, local group subschemes of the dual abelian variety A = Pic0 A are of the form A0 p (A). On the other hand, since Pict A = Pic0 A, the diagonal part of p S(A, 0) is given by Diag(Pic0A). pn. So the inverse limit of their Cartier duals is isomorphic to T 0 p (A) × T d pn × Ar We can also give another proof of the above theorem without using that p S(A, 0) is abelian (which uses the rather difficult Theorem 4.1). Namely, as in the proof of [dS, Theorem 21] it is sufficient to show that for every indecomposable numerically flat sheaf E on A there exists a unique line bundle L ∈ Pic0(A), such that L ⊗ E has a filtration by subbundles with each succes- sive quotient trivial. But by Corollary 5.5 every numerically flat bundle E on A has a filtration 0 = E0 ⊂ E1 ⊂ . . . ⊂ Er = E in which all quotients Ei/Ei−1 are numerically trivial line bundles (in fact, this filtration is just a Jordan -- Holder filtration for an arbitrary polarization, but Corollary 5.5 says a bit more about this filtration). Now if E is indecomposable then it is easy to see that all the quotients in the filtration are isomorphic (otherwise one can prove by induction that the filtration would split as H1(A, L) = 0 for a non-trivial line bundle L on the abelian variety A), which finishes the proof. 15 COROLLARY 6.2. Let A be an abelian variety defined over an algebraically closed field of posi- tive characteristic. Then Tl(A). l prime p N(A, 0) ≃ lim ←− An ≃ (cid:213) The above corollary also follows from [No2, Remark 3] and [MS, Theorem 2.3]. It is well known that the above corollary implies the Serre -- Lang theorem (see [Mu, Section 18]), although this sort of proof is much more complicated than the original one. In [No2] Nori had to use the Serre -- Lang theorem to prove the corollary. 7 The Albanese morphism Let X be a smooth projective variety and let x ∈ X be a fixed point. Let albX : X → AlbX be the Albanese morphism mapping x to 0. The variety Alb X is abelian and it is dual to the re- duced scheme underlying Pic0 X. Since the S-fundamental group scheme of an abelian variety is abelian, we have the induced homomorphism p S 1 (AlbX , 0). The following proposi- tion proves that this homomorphism is faithfully flat and it describes its kernel. ab(X , x) → p S THEOREM 7.1. We have the following short exact sequence: 0 → lim G⊂Pic0 X \G/Gred × Diag(NS(X )tors) → p S ab(X , x) → p S 1 (AlbX , 0) → 0, where the limit is taken over all local group schemes G ⊂ Pic0 X. Proof. The Albanese morphism induces alb∗ natural closed embedding. By Theorem 5.9 we have the commutative diagram X : Pic0(AlbX ) ≃ (Pic0 X )red → Pic0 X, which is the p S ab(X , x) ≃ p S 1 (AlbX , 0) ≃ G × Diag((Pict X )red) lim G⊂Pic0 X lim G⊂(Pic0 X)red G × Diag((Pic0X )red), where the limits are taken over local group schemes G and the lower horizontal map is induced by alb∗ 1 (AlbX , 0) is faithfully flat and one can easily describe its kernel. X . In particular, the homomorphism p S ab(X , x) → p S Theorem 7.1 implies in particular that if Pic0 X is reduced then the sequence 0 → Diag(NS(X )tors) → p S ab(X , x) → p S 1 (AlbX , 0) → 0 is exact. Moreover, if Pict X is connected and reduced (e.g., if X is a curve or a product of two curves as in proof of Proposition 3.1) then p S 1 (AlbX , 0) is an isomorphism. ab(X , x) → p S 16 / / / / COROLLARY 7.2. (cf. [Mi, Chapter III, Corollary 4.19]) We have 0 → lim G⊂Pic0 X and \G/Gred × Diag(NS(X )tors) → p N ab(X , x) → p N 1 (AlbX , 0) → 0 0 →(cid:18) lim G⊂Pic0 X \G/Gred(cid:19)red × Diag(NS(X )tors) → p ´et ab(X , x) → p ´et 1 (AlbX , 0) → 0. Proof. For a group scheme H let us consider the directed system of quotients H → G, where G is a finite group scheme (or an ´etale finite group). Let us consider the functor F (F ′) from the category of affine group schemes to the category of pro-finite group schemes (pro-finite groups, respectively), which to H associates the inverse limit of the above directed system. Note that these functors are exact, as the directed systems that we consider satisfy the Mittag -- Leffler con- dition. Moreover, the functor F (F ′) applied to p S 1 (X , x), respectively). Therefore the required assertions follow by applying the functors F, F ′ to the sequence from Theorem 7.1. 1 (X , x) gives p N 1 (X , x) (p ´et 8 Varieties with trivial ´etale fundamental group This section contains computation of the S-fundamental group scheme for varieties with trivial ´etale fundamental group. It contains a generalization of the main result of [EM2] and it is based on the method of [EM1]. Let X be a smooth projective variety defined over an algebraically closed field of character- istic p > 0. Let Mr be the moduli space of rank r slope stable bundles with numerically trivial Chern classes. It is known that it is a quasi-projective scheme. A closed point [E] ∈ Mr is called torsion if there exists a positive integer n such that (Fn X )∗E ≃ E. LEMMA 8.1. Assume that p ´et(X , x) = 0. If E is a strongly stable numerically flat vector bundle then its rank is equal to 1 and there exists n such that (Fn Proof. By assumption all vector bundles En = (Fn X )∗E are stable. Let N be the Zariski closure of the set {[E0], [E1], . . . , } in Mr. If N has dimension 0 then some Frobenius pull back E′ = (Fn X )∗E is a torsion point of M = Mr. In this case by the Lange -- Stuhler theorem (see [LS, Satz 1.4]) there exists a finite ´etale covering f : Y → X such that f ∗E′ is trivial. But by assumption there are no non-trivial ´etale coverings so E′ is trivial. X )∗E ≃ OX. Therefore we can assume that N has dimension at least 1. Note that the set N′ of irreducible components of N of dimension ≥ 1 is Verschiebung divisible (see [EM1, Definition 3.6]), since V N is defined at points En for n ≥ 1. Now we can proceed exactly as in proof of [EM1, Theorem 3.15] to conclude that the trivial bundle is dense in N′, a contradiction. PROPOSITION 8.2. Assume that p ´et(X , x) = 0. Let E be a rank r numerically flat vector bundle on X. Then there exists some integer n ≥ 0 such that (Fn X )∗E ≃ O r X . 17 Proof. Proof is by induction on the rank r of E. When r = 1 then the assertion follows from the above lemma. Let us recall that there exists an integer n such that (Fn X )∗E has a Jordan Holder filtration E0 = X )∗E in which all quotients Ei = Ei/Ei−1 are strongly stable numerically 0 ⊂ E1 ⊂ . . . ⊂ Em = (Fn flat vector bundles (see [La2, Theorem 4.1]). By taking further Frobenius pull backs and using the above lemma we can also assume that Ei ≃ OX. By our induction assumption taking further Frobenius pull backs we can also assume that Em/E1 ≃ O r−1 X . Now we need to show that there exists some integer s such that the extension 0 → (Fs X )∗E1 → (Fs X )∗E → (Fs X )∗(Em/E1) → 0 splits. To prove that it is sufficient to note that the endomorphism F ∗ : H1(X , OX) → H1(X , OX) is nilpotent. But we know that F ∗ induces the Fitting decomposition H1(X , OX) = H1(X , OX)s ⊕ H1(X , OX)n into stable and nilpotent parts and the assertion follows from equality H1(X , OX)s = Hom(p ´et 1 (X , x), Z/p) ⊗Fp k = 0. COROLLARY 8.3. If p ´et(X , x) = 0 then p S(X , x) ≃ p N(X , x). As a special case we get [EM2, Theorem 1.2]: if p N(X , x) = 0 then p S(X , x) = 0. Acknowledgements. The author was partially supported by a Polish MNiSW grant (contract number NN201265333). The author would like to thank the referee for the remarks that helped to improve the paper. References [BPS] I. Biswas, A. J. Parameswaran, S. Subramanian, Monodromy group for a strongly semistable principal bundle over a curve, Duke Math. J. 132 (2006), 1 -- 48. [DM] P. Deligne, J. S. Milne, Tannakian categories, pp. 101 -- 228; in P. Deligne, J. S. Milne, A. Ogus, K. Shih, Hodge cycles, motives, and Shimura varieties, Lecture Notes in Mathe- matics 900. Springer-Verlag, Berlin-New York, 1982. [dS] J. P. P. dos Santos, Fundamental group schemes for stratified sheaves, J. Algebra 317 (2007), 691 -- 713. [EM1] H. Esnault, V. Mehta, Simply connected projective manifolds in characteristic p > 0 have no nontrivial stratified bundles, Invent. Math. 181 (2010), 449 -- 465. [EM2] H. Esnault, V. Mehta, Weak density of the fundamental group scheme, Int. Math. Res. Notices (2010), doi: 10.1093/imrn/rnq187. 18 [Fa] G. Faltings, Stable G-bundles and projective connections, J. Algebraic Geom. 2 (1993), 507 -- 568. [Gi] D. Gieseker, Flat vector bundles and the fundamental group in non-zero characteristics, Ann. Scuola Norm. Sup. Pisa Cl. Sci. 2 (1975), 1 -- 31. [Ha] R. Hartshorne, Algebraic geometry, Graduate Texts in Mathematics 52, Springer-Verlag, New York-Heidelberg, 1977. [He] G. Hein, Generalized Albanese morphisms, Compos. Math. 142 (2006), 719 -- 733. [HL] D. Huybrechts, M. Lehn, The geometry of moduli spaces of sheaves, Aspects of Mathe- matics 31, 1997. [LS] H. Lange, U. Stuhler, Vektorbundel auf Kurven und Darstellungen der algebraischen Fun- damentalgruppe, Math. Z. 156 (1977), 73 -- 83. [La1] A. Langer, Semistable sheaves in positive characteristic, Ann. of Math. 159 (2004), 251 -- 276; Addendum, Ann. of Math. 160 (2004), 1211 -- 1213. [La2] A. Langer, On the S-fundamental group scheme, to appear in Ann. Inst. Fourier 61 (2011), 43pp. [Me] V. B. Mehta, Some remarks on the local fundamental group scheme and the big funda- mental group scheme, unpublished manuscript. [MN] V. B. Mehta, M. V. Nori, Semistable sheaves on homogeneous spaces and abelian vari- eties, Proc. Indian Acad. Sci. Math. Sci. 93 (1984), 1 -- 12. [MS] V. B. Mehta, S. Subramanian, On the fundamental group scheme, Invent. Math. 148 (2002), 143 -- 150. [Mi] J. S. Milne, ´Etale cohomology, Princeton Mathematical Series 33, Princeton University Press, Princeton, N.J., 1980. [Miy] Y. Miyaoka, The Chern classes and Kodaira dimension of a minimal variety, Algebraic geometry, Sendai, 1985, 449 -- 476, Adv.Stud.PureMath.10, North-Holland, Amsterdam, 1987. [Mu] D. Mumford, Abelian varieties, Tata Institute of Fundamental Research Studies in Math- ematics 5, Bombay; Oxford University Press, London 1970. [No1] M. V. Nori, The fundamental group-scheme, Proc. Indian Acad. Sci. Math. Sci. 91 (1982), 73 -- 122. [No2] M. V. Nori, The fundamental group-scheme of an abelian variety, Math. Ann. 263 (1983), 263 -- 266. 19 [Se] C. S. Seshadri, Vector bundles on curves, Contemp. Math. 153 (1993), 163 -- 200. [Wa] W. C. Waterhouse, Introduction to affine group schemes, Graduate Texts in Mathematics 66, Springer-Verlag, New York-Berlin, 1979. 20
1512.04686	1	1512	2015-12-15T09:15:08	Smooth projective horospherical varieties with nef tangent bundles	[ "math.AG" ]	We show that smooth projective horospherical varieties with nef tangent bundles are rational homogeneous spaces.	math.AG	math	Smooth projective horospherical varieties with nef tangent bundles Qifeng Li July 30, 2018 Abstract We show that smooth projective horospherical varieties with nef tangent bundles are rational homogeneous spaces. Keywords: Horospherical varieties; Campana-Peternell conjecture; VMRT. 1 Introduction We work over the field of complex numbers. A famous conjecture of Hartshorne proved by Mori [16] in 1979 shows that projective spaces are the only smooth projective varieties with ample tangent bundles. A natural question is whether there is a similar characterization for smooth projective varieties with some semipositive conditions. Demailly, Peternell and Schneider [7] proved that if X is a compact Kahler manifold with nef tangent bundle, then there is a finite ´etale cover eX such that the Albanese map α : eX → A(eX) is a smooth fiberation with fibers being Fano manifolds with nef tangent bundles. The question above will be answered if the following conjecture of Campana and Peternell [4] holds. Conjecture 1.1 (Campana-Peternell Conjecture). Smooth projective Fano varieties with nef tangent bundles are rational homogeneous spaces. From now on, we simply call a Fano manifold with nef tangent bundle a CP-manifold. De- mailly, Peternell and Schneider [7] confirm Campana-Peternell Conjecture for Fano manifolds with dimension at most three. Through the works of Campana and Peternell [5], Mok [15], Hwang [9], Watanabe [24][25], Kanemitsu [10][11], Campana-Peternell Conjecture has been proved for Fano manifolds X with dim(X) ≤ ρ(X) + 4 , where ρ(X) is the Picard number of X. The case when all the elementary Mori contractions of the manifold are smooth P1-fibrations have been solved by the series work of Munoz, Occhetta, Sol´a Conde, Watanabe and Wi´sniewski in [17] and [19]. In particular, [19] characterizes complete flag variety as the Fano manifolds all of whose elementary Mori contractions are smooth P1-fibrations. In higher dimension very little is known except for the complete flag varieties cases. In fact, it is not known whether Campana-Peternell conjecture is true for quasi-homogeneous varieties. The aim of this paper is to verify the Campana-Peternell Conjecture for an important class of quasi-homogeneous varieties. We will show the following Theorem 1.2. Smooth projective horospherical varieties with nef tangent bundles are rational homogeneous spaces. 1 The sketch of the proof is as follows. In Section 2, we study some special Mori contractions on horospherical varieties and reduce the proof of Theorem 1.2 to the cases of Picard number one. In Section 3, we study the indices of smooth projective non-homogeneous horospherical varieties of Picard number one, and show that except for two cases they do not have nef tangent bundles. In Section 4, we study the singularity of the VMRTs of the two exceptions, which imply that these varieties do not have nef tangent bundles. Notations and Conventions. Denote by G a connected reductive algebraic group. Let B be a Borel subgroup of G, T be a maximal torus in B, and Ru(B) be the unipotent radical of B. Denote by S the set of simple roots of G. For a subset I of S, denote by PI the corresponding parabolic subgroup containing B such that under that correspondence P∅ = B and PS = G. When mentioning a rational homogeneous space G/PI , we always mean its image under the minimal embedding unless otherwise stated. When G is a simple group, we use the standard notation for roots as in Bourbaki [2]. More precisely, ωi is denoted for the i-th fundamental dominant weight, and αi be the i-th simple root. We also write P (ωi) = PS\{αi}. For a linear algebraic group H, denote by X (H) the group of characters of H. For a subset A of X (H), denote by KerH (A) = {h ∈ H χ(h) = 0, ∀χ ∈ A}. For an H-variety X and a point x ∈ X, denote by Hx the isotropy group of x and H · x the H-orbit of x. Denote by C(X)(H) = {f ∈ C(X)(H)\{0} ∃χ ∈ X s.t. ∀h ∈ H, hv = χ(h)f }, i.e. C(X)(H) is the set of H-semiinvariant rational functions on X. 2 Mori contractions on horospherical manifolds 2.1 Horospherical varieties In this subsection, we will review some properties on horospherical varieties. One can consult [20] for more details or [22] for related results on spherical varieties. Definition 2.1. A G-variety X is called G-horospherical if there exists a point x0 ∈ X such that G · x0 is an open G-orbit on X and the isotropy group Gx0 ⊇ Ru(B). In this situation, we also say X is a horospherical G/H-embedding, where H = Gx0. Irreducible B-stable divisors having nonempty intersection with the open G-orbit are called colors of X. Denote by D(G/H) the set of colors of X. Let X be a G-horospherical variety. Let x0 ∈ X and H ⊆ G be as in the definition above. Then the normalizer NG(H) of H is a parabolic subgroup of G (see [20, Propostion 2.2]). Denote by I the subset of S such that NG(H) = PI . From the Bruhat decomposition on G/PI , we can get a natural bijective correspondence S\I → D(G/H), α 7→ Dα. When there is no confusions, we do not distinguish S\I and D(G/H). Denote by MG/H = X (PI /H). The character group MG/H is naturally a subgroup of X (B), and it is isomorphic to C(X)(B)/C∗. Let NG/H = Hom(MG/H , Z). For any G-orbit Z on X, denote by CZ the associated cone in (NG/H )Q, and DZ = {D ∈ D(G/H) Z ⊆ D}. Denote by FX the associated colored fan of X, i.e. FX = {(CZ, DZ ) Z is a G-orbit on X}. Denote by DX =S DZ, where Z runs over the set of G-orbits on X. The following lemma is collected from [1, Proposition 2.4] and [13, Corollary 7.8]. 2 Lemma 2.2. Let X be a horospherical G/H-embedding. Assume H ⊇ Ru(B). Denote by I the subset of S such that PI = NG(H). Take a G-orbit Z on X. Then the following hold. (a) MCZ ⊆ X (PI∪DZ ), where MCZ = {χ ∈ MG/H ∀v ∈ CZ , hχ, vi = 0}. (b) Z is G-equivariantly to G/K, where K = KerPI∪DZ (c) MCZ = MG/K and NG(K) = PI∪DZ . (d) dim(Z) = rank(MCZ ) + dim(G/PI∪DZ ). MCZ ⊇ H ⊇ Ru(B). Denote by D0(G/H) = {D ∈ D(G/H) ∀f ∈ C(X)(B), νD(f ) = 0}. The following lemma is collected from Proposition 4.6, Theorem 4.7, Lemma 7.17, and Corol- lary 7.23 in [13]. Note that for any parabolic subgroup P of G containing B, the intersection P ∩ P − is a connected reductive algebraic group. Lemma 2.3. Let X be a smooth projective horospherical G/H-embedding, where H ⊇ Ru(B) and NG(H) = PI . (i) Set D0 = D, and J the subset of S corresponding to D0(G/H). Then the PD∈D0(G/H) linear system D0 is base point free, and it induces a G-equivariant morphism π : X → G/PS\J . (ii) The Mori cone N E(X) = F × F ′, where F is the extremal face corresponding to the Mori contraction π and F ′ is another extremal face. (iii) Set L = PS\J ∩ P − horospherical variety Y . S\J . Then all fibers of π are isomorphic to a same irreducible L- (iv) If all effective divisors on X are nef, then S is the disjoint union of I, J and DX , and Xi, where each Xi is a smooth projective L-horospherical Y is L-equivariantly isomorphic to variety of Picard number one. mQi=1 2.2 Reduction to Picard number one cases The following results on manifolds with nef tangent bundles are basically due to [7, Proposition 3.7, Theorem 5.2]. See also [23, Theorem 4.4] and [18, Theorem 3.3] for modification of proofs of the conclusion (ii). Proposition 2.4. Let X be a CP-manifold. Then the following hold. (i) All effective divisors on X are nef. (ii) Let f : X → Y be the contraction of an extremal ray. Then Y and f are smooth. Moreover, Y and the fibers of f are CP-manifolds. Now we can state and prove the main result in this section. This proposition reduces the proof of Theorem 1.2 to the case when the horospherical CP-manifold is of Picard number one. Proposition 2.5. Let X be a smooth projective G-horospherical variety with nef tangent bundle. Then X is isomorphic to ( Xi) × G/P , where P is a parabolic subgroup of G, and each Xi is a smooth projective horospherical variety of Picard number one with nef tangent bundle. mQi=1 3 Proof. Step 1. In this step, we will construct two Mori contractions on X. Horospherical varieties are rational. Hence, X is simply connected and all finite ´etale cover over X is trivial (see [6, Corollary 4.18]). By considering the Albanese map (see [7, Theorem 3.14] or the statement of this result before Conjecture 1.1 in our introduction), X is a CP-manifold. Hence, all effective divisors on X are nef by Proposition 2.4(i). Thus, we can apply Lemma 2.3. There is a Mori contraction π : X → G/PS\J with fibers isomorphic to a same irreducible L-horospherical variety Xi, where J = D(G/H) ⊆ S, and each Xi is a smooth projective mQi=1 L-horospherical variety of Picard number one. The fact X is a CP-manifold implies that Y and all Xi are CP-manifolds by Proposition 2.4(ii). So we only need to prove that X ∼= Y × (G/PS\J ). By Lemma 2.3(ii), the Mori cone N E(X) = F × F ′, where F is the extremal face corre- sponding to the Mori contraction π and F ′ is another extremal face. Denote by Ψ : X → W the Mori contraction associated with the extremal face F ′. For the existence of such Ψ, see [3, Theorem 3.1(i)]. From the same theorem, we know that W is a projective G-horospherical variety and Ψ is a G-equivariant morphism. Moreover, Ψ is a smooth fibration by Proposition 2.4(ii). Then for any point w ∈ Ψ(Z), dim Ψ−1(w) = dim X − dim W. (2.1) By the description of Mori contractions on horospherical varieties in terms of colored fans (see [3, Proposition 3.4]) and Lemma 2.2, we can get the conclusions (a)(b)(c) as follows. (a) Take any G-orbit Z on X, Ψ−1(Ψ(Z)) = Z, and DΨ(Z) = DZ . (b) There exists horospherical subgroups H ′ and K ′ of G such that K ′ ⊇ H ′ ⊇ H, and Z and Ψ(Z) are G-equiviriantly isomorphic to G/H ′ and G/K ′ respectively. (c) NG(H ′) = PI∪DZ , NG(K ′) = PI∪J∪DZ , and PI∪DZ /H ′ ∼= PI∪J∪DZ /K ′. In summary, there is a commutative diagram as follows: Z G/H ′ G/PI∪DZ Ψ(Z) G/K ′ / G/PI∪J∪DZ . Step 2. The aim of this step is to show a claim on the Dynkin diagram of G. For any subset A of S, denote by ΓA the full subdiagram of the Dynkin diagram Γ of G with vertices A. Claim: Each element in S\(I ∪J) and each element in J lie in different connected components of Γ. Take any G-orbit Z on X. Then by the conclusions (a)(b)(c) in Step 1, for any point w ∈ Ψ(Z), dim Ψ−1(w) = dim PI∪J∪DZ − dim PI∪DZ . By equations (2.1) and (2.2), dim PI∪J∪DZ − dim PI∪DZ = dim X − dim W. 4 (2.2) (2.3) / / / Now apply formula (2.3) to the open G-orbit. Then for any G-orbit Z on X, dim(PI∪J∪DZ ) − dim(PI∪DZ ) = dim(PI∪J ) − dim(PI ). (2.4) Note that the union I ∪ J ∪ DZ is a disjoint union. By formula (2.4), for any G-orbit Z on X, each element in DZ and each element in J lie in different connected components of ΓI∪J∪DZ . By Lemma 2.3(iv), S is the disjoint union of I, J and DX . Recall that DX =S DZ , where Z runs over the set of G-orbits on W . Thus, the claim holds. Step 3. In this step, we will combine the two Mori contractions to get the isomorphism. Recall that N E(X) = F × F ′, and π and Ψ are the contractions corresponding to F and F ′ respectively. Hence, the natural morphism Φ : X → W × (G/PS\J ) x 7→ (Ψ(x), π(x)) is finite onto the image. On the other hand, by the Claim in Step 2, dim(G/PI ) = dim(G/PS\J ) + dim(G/PI∪J ). (2.5) Denote by G/K = Ψ(G/H) the open G-orbit on W with K ⊇ H. Then by the conclusion (c) in Step 1, NG(H) = PI , NG(K) = PI∪J , PI /H ∼= PI∪J /K. Hence, dim(X) = dim(G/PI ) + dim(PI /H) = dim(G/PS\J ) + dim(G/PI∪J ) + dim(PI∪J /K) = dim(G/PS\J ) + dim(W ). Thus, the morphism Φ : X → W × (G/PS\J ) is surjective. Take any point x ∈ X. Denote by w = Ψ(x) ∈ W . Then Ψ−1(w) is naturally a Gw- variety. By the conclusion (a) in Step 1, Ψ−1(w) is Gw-homogeneous, and it is Gw-equivariantly isomorphic to Gw/Gx. By the conclusions (a)(b)(c) in Step 1, Gw/Gx ∼= PI∪J∪DG·w /PI∪DG·x and DG·w = DG·x. Thus, Ψ−1(w) is irreducible, it is isomorphic to Go w is a parabolic subgroup of the linear algebraic group Go w is the connected component of Gw that contains the identity. Hence, the finite surjective Go w-equivariant morphism πΨ−1(w) : Ψ−1(w) → G/PS\J is an isomorphism. So Φ : X → W × (G/PS\J ) is an isomorphism. By considering the Mori contraction π, we know W is isomorphic to Y . Then the conclusion follows. w, and Gx ∩ Go w/Gx ∩ Go w, where Go 3 Indices of smooth projective horospherical varieties of Picard number one In this section, we turn to the cases of Picard number one. There is a classification of smooth projective horospherical varieties of Picard number one due to Pasquier. Proposition 3.1. ([21, Theorem 0.1, Lemma 1.19]) Let X be a smooth projective G-horospherical variety of Picard number one. Assume that X is not homogeneous. Then the following hold. (1) G acts on X with three orbits, one open orbit and two closed orbits. Identify them with G/H, G/P1 and G/P2 respectively, where P1 and P2 are parabolic subgroups of G containing 5 B. If necessary, reorder P1 and P2. Then the automorphism group Aut(X) acts on X with two orbits, G/H ∪ G/P1 and G/P2. (2) The variety X is uniquely determined by the triple (G, P1, P2). The triple (G, P1, P2) is one of the following list: (i) (Bm, P (ωm−1), P (ωm)) with m ≥ 3; (ii) (B3, P (ω1), P (ω3)); (iii) (Cm, P (ωk+1), P (ωk)) with m ≥ 2 and 1 ≤ k ≤ m − 1; (iv) (F4, P (ω2), P (ω3)); (v) (G2, P (ω2), P (ω1)), where G = Bm (resp. Cm, F4, G2) means G is a simple group of that type. For smooth projective nonhomogeneous horospherical varieties of Picard number one, we have a formula as follows. Proposition 3.2. Keep notations X, G, P1, P2 as in Proposition 3.1. Denote by rX the Fano index of X, and set P = P1 ∩ P2. Then rX = dim(P1/P ) + dim(P2/P ) + 2. (3.1) Proof. Denote by Xi the closed G-orbit on X that is identified with G/Pi, and U the open G-orbit on X. Consider the following diagram π / / G/P eX Φ X, (3.2) where Φ : eX → X is the morphism of blowing-up X along X1 ∪ X2, and the existence of the morphism π follows from the fact that X is a smooth horospherical variety. Moreover, both Φ and π are G-equivariant morphisms, and π is a P1-bundle. Denote by Ei = Φ−1(Xi). Then π sends both E1 and E2 G-equivariantly isomorphically to G/P . Set the colored fan of X to be FX = {(U, ∅), (X1, D2), (X2, D1)}, where D1 and D2 are the Now we will try to express the anticanonical divisor −KG/P using the information of X. colors of X. Denote by eDi ⊆ eX the proper transform of Di. Let Dip = π(eDi). Since X1 * D1 and X2 ⊆ D1, we know that Φ∗(D1) = eD1 + a2E2 in Pic(eX)Q for some a2 ∈ Q. Now consider the following commutative diagram: E2 πE2 / / G/P ΦE2 p2 ψ2 X2 / G/P2, (3.3) where p2 is the natural morphism induced by the inclusion P ⊆ P2, and ψ2 is a G-equivariant isomorphism. Since Φ : eX → X is the blow-up morphism of X along X1 ∪ X2, the restriction ΦE2 : E2 → X2 is a Pc2−1-bundle, where c2 is the codimension of X2 in X. Thus, dim(P2/P ) = c2 − 1, (3.4) 6 / and each fiber of p2 is isomorphic to a projective space. Similarly, we know that dim(P1/P ) = c1 − 1, (3.5) where c1 is the codimension of X1 in X. By classical results on intersection theory on rational homogeneous spaces, we know that Pic(G/P ) = Z(D1p) ⊕ Z(D2p), N E(G/P ) = Z+(l1) + Z+(l2), and Dip · lj = δij . Moreover, p2 is induced by the linear system D2p. Hence, any line l in an arbitrary fiber F of p2 is numerically equivalent to l1. In particular, D1p · l = 1 and D1pF is a generator of Pic(F ). Denote by l = (πE2 )−1(l). Since π : eX → G/P is a P1-bundle, eDi = π∗Dip for i = 1, 2. Thus, Φ∗(D1) · l = (eD1 + a2E2) · l = D1p · l + a2E2 · l = 1 − a2. Since Φ contracts l, Φ∗(D1) · l = 0. Hence, a2 = 1 and Similarly, we have Since D1 = D2 in Pic(X), we have The fact E1 ∩ E2 = ∅ implies that Φ∗(D1) = eD1 + E2 in Pic(eX). Φ∗(D2) = eD2 + E1 in Pic(eX). eD1 − eD2 = E1 − E2 in Pic(eX). E1 · E2 is a zero cycle on eX. (3.6) (3.7) (3.8) (3.9) (3.10) (3.11) The anticanonical divisor −KX = rX D1 = rX D2 in Pic(X). Since Φ is the blow-up morphism, the anticanonical divisor Moreover, the anticanonical divisor − K eX = Φ∗(−KX) − (c1 − 1)E1 − (c2 − 1)E2 in Pic(eX). − KE1 = (−K eX − E1)E1 in Pic(E1). Consider the following commutative diagram: j1 E1 ❈ ❈ ❈ ❈ πE1 ❈ ❈ ❈ / eX π !❈ G/P, where j1 is the natural inclusion morphism. Since the diagram is commutative and πE1 is an isomorphism, we know that (πE1 )∗j∗ Hence, 1 π∗(Dip) = Dip in Pic(G/P ). Recall that π∗(Dip) = eDi. (πE1 )∗(eDiE1 ) = Dip in Pic(G/P ), for i = 1, 2. 7 (3.12) ! / Note that πE1 : E1 → G/P is an isomorphism. Combining with equations (3.6) − (3.12), we can get that −KG/P = (rX − c1)D1p + c1D2p in Pic(G/P ). Similarly, by considering E2 instead of E1, we can get that −KG/P = c2D1p + (rX − c2)D2p in Pic(G/P ). Since Pic(G/P) is freely generated by D1p and D2p, we know that rX = c1 + c2 = dim(P1/P ) + dim(P2/P ) + 2, where the second equality follows from (3.4) and (3.5). Proposition 3.3. Keep notations as in Proposition 3.2. Assume moreover that X is a CP- manifold. Denote by ri the Fano index of G/Pi for i = 1, 2. Then rX = dim(P1/P ) + dim(P2/P ) + 2 ≥ max{r1, r2}. (3.13) Proof. Keep notation as in the proof of Proposition 3.2. Let Ci be a line on Xi ∼= G/Pi, i.e. Ci is an irreducible curve on Xi with smallest anticanonical degree. Thus, Ci is a smooth rational curve. Denote by fi : Ci → X the natural inclusion. By assumption, the tangent bundle TX is nef. Thus, NXi/X Ci, the restriction of the normal bundle, is nef. Consider the short exact sequence: By the nefness of NXi/X Ci, 0 → TXi Ci → TX Ci → NXi/X Ci → 0. −KX · Ci = deg(TX Ci) ≥ deg(TXi Ci) = −KXi · Ci = ri. To complete the proof, we only need to show that Di · Ci = 1 for i = 1, 2. This is equivalent to show that DiXi is the ample generator of Pic(Xi), where the later is known to be true. For the convenience of the readers, we give a proof in detail as follows. We consider the following commutative diagram: G/P p2 πE2 j2 E2 ΦE2 / eX Φ G/P2 X2 / X, i2 ψ2 (3.14) where i2 and j2 are natural inclusions. 2bD2 = D2p. Thus, Let bD2 = p2(D2p). Then bD2 is the ample generator of Pic(G/P2) and p∗ 2 Φ∗D2 = (eD2 + E1)E2 = eD2E2 2bD2, (ΦE2)∗i∗ = (πE2 )∗D2p = (πE2 )∗p∗ 2bD2 = (ΦE2)∗ψ∗ 2D2 = j∗ where the fourth equality follows from an analogue of the formula (3.12). Hence, i∗ which implies that D2X2 is the ample generator of Pic(X2). Similarly, D1X1 is the ample generator of Pic(X1). Then the conclusion follows. 2D2 = ψ∗ 2bD2, 8 o o / o o / The inequality (3.13) can be checked for the five series of horospherical varieties (i) − (v) in Proposition 3.1(2). Now keep notations as above. Denote by di = dim(Pi/P ). Then we have the following table. It should be noticed that in Case (i) of Proposition 3.1(2), we have m ≥ 3. Case (i) (ii) (iii) (iv) (v) d1 1 3 2m − 2k − 1 2 1 d2 m − 1 2 k 2 1 r1 m + 1 5 r2 2m 6 2m − k + 1 2m − k 5 3 7 5 Does (3.13) hold? No Yes Yes No No As a direct consequence, we get the following Corollary 3.4. Let X be a CP-manifold as well as a horospherical variety of Picard one. Then X is either a rational homogeneous space or a variety as in Proposition 3.1(2)(ii) or (iii). Remark 3.5. varieties in Proposition 3.1(2)(ii) and (iii) have geometric explanations. Firstly, let X be as in Case (ii). By Proposition 3.3 and the table above, dim(X) = 9 and rX = 7. Thus, X is a Mukai variety. In fact, X is isomorphic to a nonsingular section of the 10-dimensional spinor variety S10 ⊆ P15 (see also [8, Subsection 3.3]). Now assume X to be as in Case (iii). Then X is isomorphic to the odd symplectic Grass- mannian Gω(k + 1, 2m + 1) with m ≥ 2 and 1 ≤ k ≤ m − 1 (see [21, Proposition 1.12]). The odd symplectic Grassmannian is defined as follows. Let V be a linear space of dimension 2m + 1. Equip a skew bilinear form ω of rank 2m on V . Then Gω(i, 2m+1) is the variety of i-dimensional isotropic linear subspaces of V , where 1 ≤ i ≤ m + 1. 4 Lines on horospherical varieties of Picard number one Let X ⊆ PN be an n-dimensional (not necessarily irreducible or smooth) projective variety that are covered by lines. Denote by F (X) the variety of lines in X. For any point x ∈ X, denote by F (x, X) ⊆ P(TxX) the variety of lines in X passing through x, and we call it the variety of minimal rational tangents (VMRT for short) of X at x. If moreover F (x, X) ⊆ Pn−1 is covered by lines, then for any [l] ∈ F (x; X), where [l] stands for a line l in X passing through x, we denote F (x, l, X) = F ([l], F (x, X)). Remark 4.1. (i) The VMRT has the definition in a more general version. But we only need this restrictive version of definition here. (ii) Let X be a smooth projective horospherical variety of Picard number one. Then the ample generator of Pic(X), denoted by OX (1), is very ample. Denote by Φ the closed embedding defined by the linear system OX (1). Then the image Φ(X) is covered by lines. We always identify X with the image Φ(X). In particular, the VMRTs are naturally defined. Sometimes we also regard X as an closed subvariety of a natural projective variety Y (for example, in Proposition 4.7). If we can show F (x, X) 6= ∅ for some x ∈ X ⊆ Y , then the inclusion X ⊆ Y is compatible with the closed embedding induced by OX (1). The aim of this section is to show the following Proposition 4.2. Let X be a variety as in Case (ii) or (iii) in Proposition 3.1(2). Then there exists a point x ∈ X such that F (x, X) is singular. 9 Remark 4.3. If X is a CP-manifold, then the nefness of the tangent bundle implies that any irreducible rational curve on X is free. Hence, for any x ∈ X, F (x, X) is smooth. Then by Proposition 4.2, varieties in Proposition 3.1(2)(ii) and (iii) do not have nef tangent bundles. This completes the proof of Theorem 1.2. Let V be an n-dimensional linear space. Denote by F(k1, . . . , km; V ) the flag variety parame- terizing the sequences (Vk1 , . . . , Vkm ) such that Vk1 ⊆ Vk2 ⊆ . . . ⊆ Vkm and Vki is a ki-dimensional linear subspace of V . We also denote the Grassmannian G(k, n) = G(k, V ) = F(k; V ). Remark 4.4. Keep notation as above. By [12, Theorem 4.9], F (G(k, V )) = F(k − 1, k + 1; V ) and the VMRT of G(k, V ) ⊆ P(Vk V ) is isomorphic to Pk−1 × Pn−k−1 ⊆ Pk(n−k)−1. Thus, the natural projection φ : F(k − 1, k, k + 1; V ) → F(k − 1, k + 1; V ) is the universal family of lines on G(k, V ), and the natural projection ψ : F(k − 1, k, k + 1; V ) → G(k, V ) is the evaluation morphism. Each line l ⊆ G(k, n) corresponds to a point x = (Vk−1, Vk+1) ∈ F(k − 1, k + 1; V ), l = ψ(φ−1(x)) and φ−1(x) = G(1, Vk+1/Vk−1). So we can regard l as a line in G(1, V /Vk−1). In particular, for any k-dimensional linear subspace Vk of V , the VMRT F ([Vk], G(k, V )) admits a morphism Π : F ([Vk], G(k, V )) → G(k − 1, Vk) [l] 7→ [Vk−1], where [l] = (Vk−1, Vk+1) ∈ F(k − 1, k + 1; V ). The morphism Π is a trivial Pn−k−1- bundle and for each (k − 1)-dimensional linear subspace Vk−1 of Vk, the fiber Π−1([Vk−1]) = F ([Vk/Vk−1], G(1, V /Vk−1)). Let G = SO(7). Consider the following diagram: G/(P (ω1) ∩ P (ω3)) p1 / G/P (ω1) (4.1) p3 G/P (ω3). By the classical results on rational homogeneous spaces, G/P (ω1) and G/P (ω3) are smooth quadric hypersurfaces of dimension 5 and 6 respectively, p1 is a P3-bundle, and p3 is a P2- bundle. For any point x ∈ G/P (ω1), p3 sends p−1 1 (x) isomorphically to a 3-dimensional linear subspace contained in the 6-dimensional quadric hypersurface G/P (ω3). Lemma 4.5. Keep notations as above. Then there exists a line l in G/P (ω3) and a line C in G/P (ω1) such that for any q ∈ C, p3(p−1 1 (q)) is a 3-dimensional linear space containing l. Proof. Let V be a 7-dimensional linear space equipped with a nondegenerate symmetric bilinear form. Then for 1 ≤ i ≤ 3, the variety G/P (ωi) can be identified with the variety of i-dimensional isotropic linear subspaces of V . Take V2 to be a 2-dimensional isotropic linear subspace of V . Set C := G(1, V2), and l := {[V3] ∈ G(3, V ) V2 ⊆ V3 ⊆ V ⊥ 3 }. 10 / By regarding G/P (ω1) as a subvariety of the Grassmannian G(1, V ), we can see that C is a line in G/P (ω1) (see Remark 4.4). Moreover, l is a rational curve in G/P (ω3). Now consider the following diagram: G/(P (ω2) ∩ P (ω3)) p2 / G/P (ω2) p3 G/P (ω3). Then l = p3(p−1 in G/P (ω3). Hence, for any q ∈ C, p3(p−1 2 ([V2])). By the intersection theory on rational homogeneous spaces, l is a line 1 (q)) is a 3-dimensional linear space containing l. Proposition 4.6. Let X ⊆ P14 be a nonsingular hyperplane section of the 10-dimensional spinor variety S10 ⊆ P15. Then there exists x ∈ X such that F (x, X) is not smooth. Proof. Firstly, X is a horospherical variety as in Proposition 3.1(2)(ii). The automorphism group Aut(X) acts on X with two orbits. Denote by Z the closed orbit. Then Z ⊆ X is a 6-dimensional smooth quadric hypersurface. Denote by Φ : eX → X the blow-up of X along Z. Let E be the exceptional divisor of Φ. Then there is a morphism π : eX → Q5 making eX to be a P4-bundle over a 5-dimensional smooth quadric hypersurface. In summary, we have the following commutative diagram: E Z Q5 eX π Φ / X. Moveover, the restriction πE : E → Q5 is a P3-bundle. Let G = SO(7). Then Z ∼= G/P (ω3), G/P (ω1) ∼= Q5, E ∼= G/(P (ω1) ∩ P (ω3)), and the morphisms πE : E → Q5 and ΦE : E → Z coincide with p1 and p3 in the diagram (4.1) respectively. By Lemma 4.5, there exists a line C ⊆ Q5 and a line l ⊆ Z such that for each q ∈ C, Φ((πE)−1(q)) ⊆ Z is a 3-dimensional linear subspace containing the line l. Take a point z ∈ l. Since X is a nonsingular hyperplane section of S10 ⊆ P15, F (z, X) is If [l] is a singular point of F (z, X), then F (z, l, X) = If [l] is a smooth point of F (z, X), then F (z, l, X) is a hyperplane section of a hyperplane section of G(2, 5) ⊆ P9. P1 × P2 ⊆ P5. P1 × P2 ⊆ P5. Now take any point q ∈ C. Consider the following commutative diagram: (πE)−1(q) π−1(q) Φ((πE)−1(q)) / Φ(π−1(q)). By considering the diagram (4.1), we get that Φ sends (πE)−1(q) isomorphically to a 3-dimensional linear subspace containing the line l. Recall that π−1(q) is isomorphic to P4. Thus, Φ sends 11 / / / / / / / / / π−1(q) isomorphically to a 4-dimensional linear subspace containing the line l. Take any point q′ ∈ C\{q}, by the construction of the diagram (4.1), Φ((πE)−1(q′)) 6= Φ((πE)−1(q)). Thus, Φ(π−1(q′)) 6= Φ(π−1(q)). Hence, there is a 1-dimensional family of planes contained in F (z, l, X). By the discussion in last paragraph, F (z, l, X) = P1 × P2 ⊆ P5 and F (z, X) is singular at the point [l]. Proposition 4.7. Let X be the odd symplectic Grassmannian Gω(k, 2m + 1) with m ≥ 2 and 2 ≤ k ≤ m. Regard it as a subvariety of the Grassmannian G(k, 2m+1) ⊆ P(Vk C2m+1). Denote by Z the closed orbit under the action of the automorphism group of X. Take a point x ∈ X. Then there is a surjective morphism π : F (x, X) → Pk−1. (i) If x /∈ Z, then π is a P2m−2k+1-bundle. (ii) If x ∈ Z, then there exists a hyperplane H ⊆ Pk−1 such that π−1(v) ∼=(cid:26) P2m−2k+2, P2m−2k+1, v ∈ H, v ∈ Pk−1\H. In particular, when x ∈ Z, F (x, X) is not smooth. Proof. Let V be a (2m + 1)-dimensional linear space equipped with a skew bilinear form ω of rank 2m. Denote by M = G(k, 2m + 1) ⊆ P(Vk V ). Let Vk be the k-dimensional linear subspace of V represented by x ∈ X ⊆ M . By Remark 4.4, there exists a P2m−k-bundle Π : F (x, M ) → G(k − 1, Vk) such that for each (k − 1)-dimensional linear subspace Vk−1 of Vk, Π−1([Vk−1]) = F ([Vk/Vk−1], G(1, V /Vk−1)). Restrict Π on F (x, X), we get the morphism π : F (x, X) → G(k − 1, Vk) = Pk−1. For any linear subspace W of V , define W ⊥ := {v ∈ V ω(v, W ) = 0}. Take an arbitrary (k − 1)-dimensional linear subspace Vk−1 of Vk. The form ω induces a skew bilinear form ω on V ⊥ k−1/Vk−1. Moreover, k ⊇ Vk ⊇ Vk−1, V ⊥ k−1 ⊇ V ⊥ X ∩ G(1, V /Vk−1) = G w(1, V ⊥ π−1([Vk−1]) = F ([Vk/Vk−1], Gω(1, V ⊥ k−1/Vk−1), k−1/Vk−1)). Note that R := V ⊥ is a 1-dimensional linear subspace and the induced form on V /R of the skew bilinear form ω is skew bilinear and nondegenerate. Hence, dim(V ⊥ k−1) =(cid:26) 2m − k + 3, R ⊆ Vk−1, 2m − k + 2, R * Vk−1, and Gω(1, V ⊥ k−1/Vk−1)) = G(1, V ⊥ k−1/Vk−1)) ∼=(cid:26) P2m−2k+3, R ⊆ Vk−1, P2m−2k+2, R * Vk−1. By [14, Proposition 4.3], Z = {[W ] ∈ X R ⊆ W }. Then the conclusion (ii) holds, where H is defined to be the hyperplane G(k − 2, Vk/R) in G(k − 1, Vk) = Pk−1. Note that when x ∈ Z, 12 F (x, X) has two irreducible components with the same dimension and the intersection of them is not empty. So F (x, X) is not smooth for x ∈ Z. Now assume x /∈ Z. Then π is a P2m−2k+1-fibration over Pk−1. It is known that such fiberation are projective bundles. So the conclusion (i) holds. Finally, we summarize the proof of Theorem 1.2 as follows. Proof of Theorem 1.2. Let X be a smooth projective horospherical variety with nef tangent bundle. By Proposition 2.5, X ∼= ( Xi)×G/P and each Xi is a smooth projective horospherical mQi=1 variety of Picard number one as well as a CP-manifold. It suffices to show that each Xi is a rational homogeneous space. Now assume that there exists some i0 such that Xi0 is not homogeneous. By Corollary 3.4, Xi0 is as in Proposition 3.1(2)(ii) or (iii). By Proposition 4.2, in both cases we can find x ∈ Xi0 such that the variety F (x, Xi0 ) is singular. This contradicts Remark 4.3. Then the conclusion follows. Acknowledgements. I want to thank Professor Baohua Fu and Professor Jun-Muk Hwang for communications related with this paper. References [1] V. Batyrev, A. Moreau, The arc space of horospherical varieties and motivic integration. Compos. Math. 149 (2013), no. 8, 1327-1352. [2] N. Bourbaki, Groupes et alg`ebres de Lie. Chapitres 4, 5, 6. C.C.L.S., Paris (1975). [3] M. Brion, Vari´et´es sph´eriques et th´eorie de Mori. Duke Math. J. 72 (1993), no. 2, 369-404. [4] F. Campana, T. Peternell, Projective manifolds whose tangent bundles are numerically ef- fective. Math. Ann. 289, 169-187 (1991). [5] F. Campana, T. Peternell, 4-folds with numerically effective tangent bundles and second Betti numbers greater than one. Manuscr. Math. 79(3-4), 225-238 (1993). [6] O. Debarre, Higher-dimensional algebraic geometry. Universitext. Springer-Verlag, New York, 2001. [7] J. P. Demailly, T. Peternell, M. Schneider, Compact complex manifolds with numerically effective tangent bundles. J. Algebraic Geom. 3 (1994), no. 2, 295-345. [8] B. Fu, J.-M. Hwang, Classification of non-degenerate projective varieties with non-zero pro- longation and application to target rigidity. Invent. Math. 189 (2012), no. 2, 457-513. [9] J. M. Hwang, Rigidity of rational homogeneous spaces. In: Proceedings of ICM. 2006 Madrid, vol. II, pp. 613-626. European Mathematical Society (2006). [10] A. Kanemitsu, Fano 5-folds with nef tangent bundles. arXiv:1503.04579, 2015. [11] A. Kanemitsu, Fano n-folds with nef tangent bundle and Picard number greater than n-5. arXiv:1505.02952, 2015. 13 [12] J. M. Landsberg, L. Manivel, On the projective geometry of rational homogeneous varieties. Comment. Math. Helv. 78 (2003), no. 1, 65-100. [13] Q. Li, Pseudo-effective and nef cones on spherical varieties. Math. Z. 280 (2015), no. 3-4, 945-979. [14] I. A. Mihai, Odd symplectic flag manifolds. Transform. Groups 12 (2007), no. 3, 573-599. [15] N. Mok, On Fano manifolds with nef tangent bundles admitting 1-dimensional varieties of minimal rational tangents. Trans. Am. Math. Soc. 354(7), 2639-2658 (2002). [16] S. Mori, Projective manifolds with ample tangent bundles. Ann. of Math. (2) 110 (1979), no. 3, 593-606. [17] R. Munoz, G. Occhetta, L. E. Sol´a Conde, K. Watanabe, Rational curves, Dynkin diagrams and Fano manifolds with nef tangent bundle. Math. Ann. 361 (2015), no. 3-4, 583-609. [18] R. Munoz, G. Occhetta, L. E. Sol´a Conde, K. Watanabe, J. A. Wi´sniewski, A survey on the Campana-Peternell Conjecture, arXiv:1407.6483, 2014. [19] G. Occhetta, L. E. Sol´a Conde, K. Watanabe, J. A. Wi´sniewski, Fano manifolds whose elementary contractions are smooth P1-fibrations. arXiv:1407.3658, 2014. [20] B. Pasquier, Vari´et´es horosph´eriques de Fano. Bull. Soc. Math. France 136 (2008), no. 2, 195-225. [21] B. Pasquier, On some smooth projective two-orbit varieties with Picard number 1. Math. Ann. 344 (2009), no. 4, 963-987. [22] N. Perrin, On the geometry of spherical varieties. Transform. Groups 19 (2014), no. 1, 171-223. [23] L. E. Sol´a Conde, J. A. Wi´sniewski, On manifolds whose tangent bundle is big and 1-ample. Proc. London Math. Soc. (3) 89 (2004), no. 2, 273-290. [24] K. Watanabe, Fano 5-folds with nef tangent bundles and Picard numbers greater than one. Math. Z. 276 (2014), no. 1-2, 39-49. [25] K. Watanabe, Fano manifolds with nef tangent bundle and large Picard number. Proc. Japan Acad. Ser. A Math. Sci. 91 (2015), no. 6, 89-94. Korea Institute for Advanced Study, 85 Hoegiro, Dongdaemun-gu, Seoul, 130-722, Republic of Korea E-mail address: [email protected] 14
1001.4865	2	1001	2010-02-03T00:21:29	Thomae type formula for K3 surfaces given by double covers of the projective plane branching along six lines	[ "math.AG", "math.CV" ]	In this paper, we give Thomae type formula for \KK surfaces $\cS$ given by double covers of the projective plane branching along six lines. This formula gives relations between theta constants on the bounded symmetric domain of type $I_{22}$ and period integrals of $X$. Moreover, we express the period integrals by using the hypergeometric function $F_S$ of four variables. As an application of our main theorem, we define $\R^4$-valued sequences by mean iterations of four terms, and express their common limits by the hypergeometric function $F_S$.	math.AG	math	Thomae type formula for K3 surfaces given by double covers of the projective plane branching along six lines Keiji Matsumoto and Tomohide Terasoma November 4, 2018 Dedicated to Professor Takayuki Oda on his sixtieth birthday Abstract In this paper, we give Thomae type formula for K3 surfaces X given by double covers of the projective plane branching along six lines. This formula gives relations between theta constants on the bounded symmetric domain of type I22 and period integrals of X . Moreover, we express the period integrals by using the hypergeometric function FS of four variables. As applications of our main theorem, we define R4-valued sequences by mean iterations of four terms, and express their common limits by the hypergeometric function FS. MSC2000: Primary 33C70; Secondary 11F55. Keywords: Hypergeometric Functions, Theta Functions. Contents 1 Introduction 2 2 Certain family of K3 surfaces 5 2.1 Double coverings of P2 branching along 6 lines . . . . . . . . . 5 6 2.2 Relative invariants and a global two form . . . . . . . . . . . . 8 2.3 Topological cycles at a reference point . . . . . . . . . . . . . 2.4 Period integrals and Hypergeometric functions . . . . . . . . . 9 2.5 Preparation for the association involution . . . . . . . . . . . . 11 1 Thomae type formula for K3 surfaces 2 3 Bounded symmetric domains and theta functions. 13 3.1 Period map and symmetric domains DH and D . . . . . . . . 13 3.2 Homomorphisms of discrete groups . . . . . . . . . . . . . . . 15 3.3 Monodromy actions on the spaces DH and D . . . . . . . . . . 16 3.4 Theta functions and their functional equations . . . . . . . . . 18 4 Thomae type formula for K3 surfaces 21 4.1 Main Theorem . . . . . . . . . . . . . . . . . . . . . . . . . . 21 5 Mean iterations 22 5.1 Mean iteration associated to D4 degeneration . . . . . . . . . 22 . . . . . . . . . . 27 5.2 Mean iteration associated to Kummer locus 6 Functional equations for FS 29 1 Introduction Let us consider period integrals ωA(λ) =Z 1/λ 1 dt pt(1 − t)(1 − λt) , ωB(λ) =Z 1 0 dt pt(1 − t)(1 − λt) , (1) of an elliptic curve s2 = t(1 − t)(1 − λt) with λ ∈ C − {0, 1}. If λ belongs to the open interval (0, 1), then they are expressed by the Gauss hypergeometric function F (a, b, c; z) = ∞Pn=0 1 2 (a)n(b)n (c)nn! zn : , 1 2 , 1; 1 − λ), ωB(λ) = πF ( 1 2 , 1 2 , 1; λ), ωA(λ) = ıπF ( where ı = √−1. The function τ = ωA(λ)/ωB(λ) of λ is continued to a map per : eX → H = {z ∈ C Im(z) > 0} from the universal covering eX of C − {0, 1} to H, which is called the period map. The inverse of the map per can be described as λ = ϑ4 [10](τ )/ϑ4 [00](τ ), Thomae type formula for K3 surfaces 3 where ϑ[ab](τ ) =Xn∈Z exp[πı{(n + a 2 )2τ + (n + a 2 )b}] ([ab] = [00], [01], [10]) is Jacobi's theta constant. Under these correspondences of variables λ ∈ C − {0, 1} and τ ∈ H, the theta constant and the elliptic integral are related as ϑ4 [ab](τ ) = Λ[ab] π2 ωB(λ)2, (2) where Λ[00] = 1, Λ[01] = 1 − λ, Λ[10] = λ. The identity (2) is called Jacobi's formula. On the other hand, we have the 2τ -formulas for the theta constants ϑ2 [00](2τ ) = [00](τ ) + ϑ2 ϑ2 [01](τ ) 2 , ϑ2 [01](2τ ) = ϑ[00](τ )ϑ[01](τ ). These formulas are applied to the study of the arithmetic-geometric mean as follows. Let c1, c2 ∈ R×+ be positive real numbers. We define vector valued sequence {mn(c1, c2)}n∈N by z where the map m : (R×+)2 → (R×+)2 is mn(c1, c2) = m ◦ · · · ◦ m(c1, c2) n } { m(u1, u2) = ( u1 + u2 2 ,√u1u2). Both components have a common limit and it is called the arithmetic-geometric mean and denoted by m∞ (c1, c2). Using Jacobi's formula and 2τ -formulas, ∗ we have a relation between the arithmetic-geometric mean and the hyperge- ometric function: m∞ ∗ (c1, c2) = 2, 1; 1 − ( c2 By the above relation and the invariance property m∞ we have the Gauss transformation formula 2, 1 F ( 1 c1 . )2) c1 ∗ (m(c1, c2))) = m∞ ∗ (c1, c2), F(cid:16)1 2 , 1 2 , 1; 1 − 4z (1 + z)2(cid:17) = 1 + z 2 F ( 1 2 , 1 2 , 1; 1 − z2). (3) Thomae type formula for K3 surfaces 4 Thomae studies period integrals of a hyperelliptic curve of arbitrary genus and generalizes Jacobi's formula to Thomae's formulas in [To]. Based on 2τ - formulas of theta constants defined on the Siegel upper half space H2 of degree 2, Borchardt introduces a vector valued sequence {mn(c1, . . . , c4)}n∈N with initial (c1, . . . , c4) ∈ (R×+)4 given by the iteration of the map m : (R×+)4 ∋ u = (u1, . . . , u4) 7→ (m1(u), . . . , m4(u)) ∈ (R×+)4, where m1(u) = m3(u) = u1 + u2 + u3 + u4 √u1u3 + √u2u4 4 , m2(u) = m4(u) = , 2 √u1u2 + √u3u4 , √u1u4 + √u2u3 2 . 2 By using Thomae's formulas, he expresses the common limit of the compo- nents of {mn(c1, . . . , c4)}n∈N by period integrals of a genus 2 hyperelliptic curve. For related studies, refer to [B], [MT] and [Me]. In this paper, we give Thomae type formula for K3 surfaces which are dou- ble covers of the complex projective plane P2 branching along normal crossing six lines. The configurations of normal crossing six lines are parametrized by 3 × 6 matrices x and the corresponding K3 surface is denoted by X (x). Pe- riod integrals of X (x) are expressed in terms of two kinds of hypergeometric functions FS and FT of four variables defined in (8) and (9), respectively. In §3.1, we define a normalized period matrix τ of X (x) in the 4-dimensional bounded symmetric domain D of type I22. Let P3,3 be the set of unorderd pair hJi = (J, J c), such that #J = #J c = 3 and J ∪ J c = {1, . . . , 6}. Then we have #P3,3 = 10. To state the main theorem, we introduce the following notations: 1. ΘhJi(τ ) is theta functions on D evaluated at the normalized period matrix τ indexed by hJi ∈ P3,3, 2. xhJi is the product of two 3 × 3-minors of a 3 × 6-matrix x in the configuration space X(3, 6) also indexed by hJi ∈ P3,3, 3. ω34(x) is the period integral of the K3 surface X (x) given in (6). Then the main theorem is the identity Θ 2 hJi(τ ) = 1 4π4 xhJiω34(x)2 (4) Thomae type formula for K3 surfaces 5 for any hJi ∈ P3,3. The subfamily consisting of Kummer varieties of principally polarized abelian varieties is called the Kummer locus. This locus corresponds to the Siegle upper half space H2 realized as a closed subdomain of D. Our identity becomes Thomae's formula for genus two curves on this locus. In the paper [MSY] and [Ma], they prove that the point [ΘhJi(τ )]hJi∈P3,3 in P9 is equal to [xhJi]hJi∈P3,3. The key for our proof of the main theorem is the study of the relation between a period of X and the automorphic factor of ΘhJi by the action of the monodromy group of per via the isomorphism between D and DH defined in §3.1. As an application of our main theorem, we study a vector valued sequence obtained by mean iteration of a map from R4 + which is different from that defined by Borchardt in §5.1. We show that this vector valued sequence has a common limit and that it can be expressed by the hypergeometric function FS. This formula is obtained by the main theorem and 2τ formulas for the theta functions ΘhJi in Theorem 1. We also give an explanation on the relation between Borchardt's arithmetic-geometric mean m∞ (c1, . . . , c4) ∗ and the hypergeometric function FS in §5.2. In the last section, we prove several functional equations of the hypergeometric function FS arising from the invariance property for vector valued mean iterations. These are analogs of the Gauss transformation formula (3) for the hypergeometric function FS. + to R4 2 Certain family of K3 surfaces 2.1 Double coverings of P2 branching along 6 lines Let M×(3, 6) be the open subset of M(3, 6) defined by the determinants of (3, 3)- minors are non-zero M×(3, 6) =nx = (ℓ1, . . . , ℓ6) ∈ M(3, 6)(cid:12)(cid:12)(cid:12) For ℓi = t(ℓ0i, ℓ1i, ℓ2i), we define a linear function (t, ℓi) by P3 bX ∗ be the variety defined by bX ∗ = {(t : y) × x ∈ P(1, 1, 1, 3) × M×(3, 6) y2 = (t, ℓ1) · · · (t, ℓ6)}, where (t : y) = (t0 : t1 : t2 : y), x = (ℓ1, . . . , ℓ6) and P(1, 1, 1, 3) is the weighted projective space of weight (1, 1, 1, 3). Then by the natural map o j=1 ℓj,itj. Let Thomae type formula for K3 surfaces 6 M×(3, 6). By resolving singularities, we have a family of K3 surfaces pr2 : pr1 : bX ∗ → P2 × M×(3, 6), bX ∗ is a family of branched covering of P2 over bX → M×(3, 6) on M×(3, 6). Let T be a torus defined by T = {λ = (λ0, λ1, . . . , λ6) ∈ C×6 × C× λ2 0 = λ1 · · · λ6}. The group GL3(C) × T acts on bX by t 7→ tg−1, x 7→ g · x · diag(λ1, . . . , λ6), y 7→ λ0y for (g, λ) ∈ GL3(C)× T and it induces an action of GL3(C)× T on M×(3, 6). The quotients of bX and M×(3, 6) by GL3(C) × T are denoted by X and X, respectively. The natural map pr2 : bX → M×(3, 6) induces a map X → X, which is also denoted by pr2. The variety X is equal to the double coset space: X = X(3, 6) = GL3(C)\M×(3, 6)/(C×)6, which is called the configuration space. The fiber of X at x ∈ X is denoted by X (x). There are 15 rational curves ljk(x) (1 ≤ j < k ≤ 6) in X (x) coming from the resolutions of nodes at ljk(x) = lj(x)∩lk(x), where li(x) is the line defined by (t, ℓi) = 0. Let l0(x) be a pull back of a generic line in P2 by pr1. Let S(x) be the subgroup of H2(X (x), Z) generated by algebraic cycles ljk(x) and l0(x). Its orthogonal complement T (x) in H2(X (x), Z) with respect to the intersection form ( · ) is called the transcendental lattice of X and its rank is 22 − 16 = 6. In §2.3, we give a basis of T (x) and its dual in H2(X (x), Z). These bases are slightly different from those defined in [MSY] and [Y]. 2.2 Relative invariants and a global two form The characters ρ : GL3(C) × T ∋ (g, λ) 7→ deg(g)λ0 ∈ C× : and ρ2 define linearlizations of GL3(C)×T of OM ×(3,6) and O bX . The invariant line bundles on X and X under these actions are denoted by L and M, respectively. We have L⊗2 = pr∗2M. Thomae type formula for K3 surfaces 7 We construct elements of H 0(X,M). Let J be a subset of the set {1, . . . , 6} with cardinality 3 and J c be its complement. By reordering ele- ments, we may write J and J c as J = {j1, j2, j3}, j1 < j2 < j3, J c = {j4, j5, j6}, j4 < j5 < j6. A pair hJi = (J, J c) = (J c, J) of J and J c is called a (3, 3)-partition of the set {1, . . . , 6}. The set of (3, 3)-partitions is denoted by P3,3. Note that #P3,3 = 10. For x = (xij) ∈ M(3, 6) and hJi ∈ P3,3, we set xhJi = det(xi,jk)1≤i,k≤3 det(xi,jk+3)1≤i,k≤3. Then xhJi is an element of H 0(X,M). By Plucker relations, we have the following. Lemma 1 Let St be the set of (2,2,2)-standard tableaux i.e. ({j1, j2, j3}, {j4, j5, j6}) with j1 < j4 ∧ ∧ j2 < j5 ∧ ∧ j3 < j6. (J, J c) = Then #St = 5 and {xhJi hJi ∈ St} forms a basis of a linear system in H 0(X,M) generated by the polynomials xhijki (1 ≤ i < j < k ≤ 6). where we arrange xhJi lexicographically for J = {j1, j2, j3} with j3 ≤ 5. By Lemma 1, the image of pl is contained in a 5-dimensional linear subspace of Let bpl be the map from M×(3, 6) to C10 defined by bpl : M×(3, 6) ∋ x 7→ (. . . , xhJi, . . .)hJi∈P3,3 ∈ C10, C10. The map X → P4 induced from bpl is denoted by pl. The space of relative global holomorphic 2-forms H 0(bX , Ω2 generated by ) is bX /M ×(3,6) ϕ = t0dt1 ∧ dt2 − t1dt0 ∧ dt2 + t2dt0 ∧ dt1 . y Proposition 1 The form ϕ satisfies the equality: (g, λi, λ)∗ϕ = ρ−1(g, λi, λ)ϕ. Therefore it defines a global section of H 0(X , Ω2 X /X ⊗ L−1). Thomae type formula for K3 surfaces 8 2.3 Topological cycles at a reference point We take a reference point x in M×(3, 6) as x = p2 2 p2 p2 1 6 −p1 −p2 −p3 −p4 −p5 −p6 1 1 p2 4 p2 3 1 p2 5 1 1 1  with p1 = −3, p2 = −2, p3 = −1, p4 = 1, p5 = 2, p6 = 3. We consider the affine coordinates s1 = t1/t0 and s2 = t2/t0 of P2. We construct topological 2-cycles of X ( x) using the isomorphism of the Kummer surface of C and X ( x) given in [Te]. Let C be a hyperelliptic curve defined by u2 = 6Yi=1 (w − pi), and C1, C2 be copies of C. Let sym : C1 × C2 → X ( x) be a map defined by ((w1, u1), (w2, u2)) 7→ (s1, s2, y) = (w1 + w2, w1w2, u1u2). For a < b ∈ R, the 1-chain in P1 defined by the segment from a to b is denoted by (a, b). We define chains A′1, A′2, B′1 and B′2 in C by the liftings of (p1, p2), (p5, p6), (p2, p3) and (p4, p5) on which 1/u is in ıR+, ıR+, R+ and R+. Then Ai = A′i − σ(A′i) and Bi = B′i − σ(B′i) become 1-cycles on C. We set A1 = γ1, A2 = γ2, B1 = γ3 and B2 = γ4. We define a topological cycle γij∗ by sym∗(γi × γj). The proper inverse image of γij∗ in X ( x) is denoted by γij. Then γij = γij/2 is an element in H2(X ( x), Z). Let γ′ij be the orthogonal projection of 2γij to T (x). Then {γ′12, γ′13, γ′14, γ′23, γ′24, γ′34} is a basis of T (x). Since sym∗(ϕ( x)) = (w1 − w2)dw1 ∧ dw2 u1u2 , we have ZZγ13 ZZγ12 ϕ( x), ZZγ14 ϕ( x) ∈ −R×+, ϕ( x) ∈ ıR×+, ZZγ34 ϕ( x) ∈ R×+, ϕ( x), ZZγ24 ZZγ23 ϕ( x) ∈ −ıR×+. (5) Thomae type formula for K3 surfaces 9 Proposition 2 We set γ = t(γ12, γ13, γ14, γ23, γ24, γ34), γ′ = t(γ′12, γ′13, γ′14, γ′23, γ′24, γ′34). Then the intersection matrix is equal to where (γ · tγ′) = H, (γ′ · tγ′) = 2H, H = − 1 1 1 1 1 1  .  (6) (7) Thus the lattice structure of T (x) is equal to U(2)⊕U(2)⊕A1(−1)⊕A1(−1). (For details, see [MSY] and [Y], Chapter VIII.) For any x ∈ M×(3, 6), take a path ρx in M×(3, 6) connecting x and x, and define bases γij(x) and γ′ij(x) as the continuations of γij( x) and γ′ij( x) along the path ρx by the local triviality. They depend only on the homotopy tγ′(x)) are equal to that in class of ρx. The intersection matrix for (γ′(x) · Proposition 2. Let ω be a vector defined by ω(x) = t(ω12(x), ω13(x), ω14(x), ω23(x), ω24(x), ω34(x)), ϕ(x). ωij(x) =ZZγij (x) fper : eX ∋ x 7→ [ω(x)] ∈ P5, The map z3 2.4 Period integrals and Hypergeometric functions is called the period map, where eX is the universal covering of X. (cid:18) z1 We define two hypergeometric series F α z4(cid:19) with parameters α = (α1, . . . , α6) satisfyingP6 S (z) = Xn∈N4 T (z) = Xn∈N4 (1−α1)n1+n3(1−α2)n2+n4(α5)n1+n2(α6)n3+n4 (2−α1−α3)n1+n3(2−α2−α4)n2+n4n1!n2!n3!n4! (1 − α1)n1+n3(1 − α2)n2+n4(α5)n1+n2(α6)n3+n4 (3 − α1 − α2 − α3)n1+n2+n3+n4n1!n2!n3!n4! j=1 αj = 3 as F α F α z2 S (z) and F α T (z) of variables z = zn, (8) zn, (9) Thomae type formula for K3 surfaces 10 where N = {0, 1, 2, . . .}, zn = zn1 αj(αj + 1) · · · (αj + nj − 1) = Γ (αj + nj)/Γ (αj), and we assume that for n = (n1, . . . , n4), (αj)nj = 3 zn4 1 zn2 2 zn3 4 α1 + α3 − 2, α2 + α4 − 2 /∈ N for F α for F α S (z), T (z). α1 + α2 + α3 − 3 /∈ N They absolutely converge on the domain {z ∈ C4 z1 + z2 < 1, z3 + z4 < 1}. By the standard argument for Euler type integrals, we have the following Proposition. Proposition 3 The hypergeometric series F α resentations: S and F α T admit the integral rep- F α S (z) = F α T (z) = 1 B(1 − α1, 1 − α3)B(1 − α2, 1 − α4)Z 1 0Z 1 Γ (1 − α1)Γ (1 − α2)Γ (1 − α3)ZZ∆ Γ (3 − α1 − α2 − α3) Lα 0 Lα S(z, s) ds1ds2, T (z, s)ds1ds2, where B denotes the beta function, S(z, s) = s−α1 Lα 1 s−α2 2 (1 − s1)−α3(1 − s2)−α4 ×(1 − z1s1 − z2s2)−α5(1 − z3s1 − z4s2)−α6, T (z, s) = s−α1 Lα 1 s−α2 2 (1 − s1 − s2)−α3 ×(1 − z1s1 − z2s2)−α5(1 − z3s1 − z4s2)−α6, ∆ = {(s1, s2) ∈ R2 s1 > 0, s2 > 0, s1 + s2 < 1}, arg(sj) = arg(1 − sj) = arg(1 − s1 − s2) = 0 on each interior area of the integrations, arg(1 − z1s1 − z2s2), arg(1 − z3s1 − z4s2) become 0 at (s1, s2) = (0, 0), and we assume that Re(α1), Re(α2), Re(α3), Re(α4) < 1 for F α for F α Re(α1), Re(α2), Re(α3) < 1 S (z), T (z). From now on, we put α = (1/2, . . . , 1/2) and we set FS(z) = F α S (z), FT (z) = F α T (z). We can regard xhJiωij(x)2 as a multivalued function on X for any hJi. Proposition 1 and 3 imply the following. Thomae type formula for K3 surfaces 11 Proposition 4 For a fixed hJi ∈ P3,3, the product xhJi· ωij(x)2 is invariant under the action of GL3(C) × T, i.e., (g, λi, λ)∗(xhJi · ω2 ij) = xhJi · ω2 ij. As a consequence, if x is in a neighborhood of x ∈ M×(3, 6), ωij(x)2 can be expressed as z3 z2 (i, j) = (1, 2), (1, 4), (2, 3), (3, 4), (i, j) = (1, 3), (2, 4), 16π2FT (ζij(x))2νijhJi, are equivalent to x as elements of X: ωij(x)2xhJi =( 4π4FS(ζij(x))2νijhJi, where we set ζij(x) =(cid:18) z1 ν12 = ν14 = ν24 = 0 1 1 −1 −z1 −z3 0 1 −1 0 1 1 0 −z1 −z3 −1 0 −z2 −z4 −1 1 0 0 −z2 −z4 −1 0 1 1 −z1 0 −z2 1 1 0 1 1 0 −z3 0 1 −1 −z4 0 0 1  ,  ,  , z4(cid:19) so that the following ν12, . . . , ν34 ∈ M×(3, 6)  ,  ,  . 0 1 1 −1 −z1 −z3 0 −1 −z2 −z4 1 0 0 −z3 −1 1 −z4 0 0 1 0 ν13 = ν23 = ν34 = 0 0 1 1 1 −z1 0 −z2 −1 1 1 −z1 −1 −z2 0 1 1 0 0 −z3 0 1 −1 −z4 1 0 0 0 1 0 1 1 1 1 1 0 0 1 0 1 0 2.5 Preparation for the association involution Let as be an automorphism of M×(3, 6) given by as : M×(3, 6) ∋ (y1, y2) 7→ ( t(y−1 1 y2y1), ty1) ∈ M×(3, 6), where y1, y2 ∈ GL3(C). By a straightforward calculation, we have 1 y2y1y−1 as2(y1, y2) = (y−1 1 y2y1) = y−1 1 y2y1y−1 2 y1, y−1 2 (y1, y2). Therefore as induces an involution on X, which is called the association involution and also denoted by as. By the above equality, we have an induced morphism epl ◦ as = epl, pl∗ : X/hasi → P4. Thomae type formula for K3 surfaces 12 Proposition 5 (Chapter VII of [Y]) The morphism pl∗ is an open im- mersion. Let X be the normalization of P4 in X. Then we have the diagram: X → X ↓ ↓ X/hasi → P4. The induced map X → P4 is denoted as pl. Let x = (xij)ij ∈ M×(3, 6). We define the following polynomials Q = det(x2 1i, x2 2i, x2 3i, x2ix3i, x3ix1i, x1ix2i)i=1,...,6, D(ijk) = det(xpi, xpj, xpk)p=1,...,3, {ij; kl} = D(ijm)D(ijn)D(mkl)D(nkl), T (ijklmn) = D(ijk)D(klm)D(mni)D(nlj) for {i, j, k, l, m, n} = {1, . . . , 6}. Then we have {ij; kl} = ±xhijmixhijni and as(Q) = −Q. We give an explicit description of the normal form 1 1 0 0 1 1 0 1 0 1 x1 y1 0 0 1 1 x2 y2   of the inverse image under pl. By Lemma A6.8 of [MSY], we have x2 = D(125)D(234) D(235)D(124) = T (125364) {35; 14} . We have as(T (125364)) = T (364125) and by Lemma A7.3 of [MSY], we have T (125364) + as(T (125364)) = {14; 53} − {52; 16} + {63; 54} T (125364) · as(T (125364)) = {16; 23} · {12; 36} −{23; 15} + {24; 56} Therefore T (125364) is defined by a quadratic equation with the coefficients in polynomials of xhijki's. The values x1, y2, y2 can be obtained by substi- tuting indices 2 ↔ 3, 5 ↔ 6. Since T (125364) − as(T (125364)) = Q, if {16; 23}· {12; 36} = 0, then Q = T (125364) or −T (364125) and the values x2, x1, y2, y1 are polynomials of xhijki. Thus we have the following lemma. Thomae type formula for K3 surfaces 13 Lemma 2 The inverse image of the divisor {xh164i = 0} on P4 under the map pl consists of two irreducible components. On X, we can express x1, x2, y1, y2 as rational functions of xhijki on each irreducible component. For an explicit description of the inverse image, see Proposition 8. 3 Bounded symmetric domains and theta func- tions. 3.1 Period map and symmetric domains DH and D In this section, we introduce two symmtric domains DH and D. The target space for the natural period map for K3 surfaces is the symmetric domain DH of type IV . We use theta functions on D and an isomorphism between DH and D to construct automorphic functions on DH by using results in [Ma]. By Riemann bilinear relations for the K3 surface X (x), and the choice of orientations of T (x) in (5), the class [ω(x)] of (6) in P5 belongs to the subset DH = {[w] ∈ P5 tω H ω = 0, ω∗ H ω > 0, Im( ω14 ω34 ) > 0}. Here y∗ = ty denotes the adjoint of a matrix y. Therefore we regard the map (7) as fper : eX → DH. Let D be the bounded symmetric domain of type I22 defined by D = {τ ∈ M(2, 2) τ − τ∗ 2ı is positive definite }. In this subsection, we define an isomorphism D → DH. Let τ be an element in D. We set τ =(cid:18) τ E2(cid:19) . Let τhi1i2i be the (i1, i2) × (1, 2)-minor of the 4 × 2 matrix τ . They satisfy the Plucker relation τh12iτh34i − τh13iτh24i + τh14iτh23i = tv(τ ) H′ v(τ ) = 0, Thomae type formula for K3 surfaces 14 where H′ =  1 −1 1 1 −1 1  Since the matrix (τ − τ∗)/2ı is positive definite, we have ) > 0. v(τ )∗ H v(τ ) > 0, Im( τh14i(τ ) τh34i(τ ) We set Q = Then by the equality 1  1+ı 2 −1+ı 2 1 1 −1+ı 2 1+ı 2 1 , v(τ ) = τh12i τh13i τh14i τh23i τh24i τh34i .    . Q∗ H Q = H, H′ = tQ H Q, H = tQ H′ Q, the class [Qv(τ )] of Qv(τ ) in P5 is contained in DH . Thus we have an isomorphism (10) D ([ω(x)]) ∈ pD : D ∋ τ 7→ [Qv(τ )] ∈ DH . We define the normalized period matrix of X (x) by τ = τ (x) = p−1 D. Then we have τ (x) = 1 ω34(x)(cid:18) ω14(x) − ω13(x)+ıω24(x) 1−ı − ω13(x)−ıω24(x) −ω23(x) (cid:19) , 1+ı (11) where ωij(x) are defined by (6). Thus we have a map Since τh34i(τ ) = 1, we have fper : eX ∋ x 7→ τ (x) ∈ D. D(τ (x)) = ω(x)/ω34(x) (12) as elements of C6. Thomae type formula for K3 surfaces 15 3.2 Homomorphisms of discrete groups We define a discrete group ΓH in GL6(Z) by ΓH = {R ∈ GL6(Z) tRHR = H, Im( [Rω( x)]14 [Rω( x)]34 ) > 0}, where [Rω( x)]ij denotes the (ij)-component of Rω( x). Its center consists of ±E6. The the group ΓH acts on DH from the left. Since the monodromy action preserves the intersection forms, the monodromy group is contained in ΓH. We define U22(Z[ı]) and the principal congruence subgroup of level (1 + ı) by U22(1 + ı) = {g ∈ U22(Z[ı]) g ≡ E4 mod (1 + ı)}. U22(Z[ı]) =ng ∈ GL4(Z[ı])(cid:12)(cid:12)(cid:12) g I22 g∗ = I22o , g22(cid:19) in U22(Z[ı]) acts on D by g21 g12 E2 g · τ = (g11τ + g12)(g21τ + g22)−1, O (cid:19) and gij ∈ M(2, 2). Then an element g =(cid:18) g11 where I22 =(cid:18) O −E2 mains D → DH defined in (10). For g = (gij) =(cid:18) g11 gi2j2(cid:12)(cid:12)(cid:12)(cid:12)(cid:19)(i1i2),(j1j2) ∧2g =(cid:18)(cid:12)(cid:12)(cid:12)(cid:12) set a 6 × 6 matrix ∧2g by gi1j1 gi2j1 gi1j2 g21 , In this subsection, we define a homomorphism U22(Z[ı]) → ΓH/h±1i of discrete group which is compatible with the isomorphism of symmetric do- g12 g22(cid:19) ∈ U22(Z[ı]), we where 1 ≤ i1 < i2 ≤ 4, 1 ≤ j1 < j2 ≤ 4, and they are arranged lexicographi- cally. Then we have and det(∧2g) = det(g)3, as elements of C6, where τ ∈ D, det(g21τ + g22)v(τ′) = v(gτ ) = (∧2g)v(τ ) E2(cid:19) , τ′ =(cid:18) τ′ τ′ = g · τ = (g11τ + g12)(g21τ + g22)−1, gτ = g(cid:18) τ E2(cid:19) . Thomae type formula for K3 surfaces 16 Thus we have det(g21τ + g22)Qv(τ′) = {Q(∧2g)Q−1}{Qv(τ )}, and by the definition of D, we have det(g21τ + g22)D(g · τ ) = Q(∧2g)Q−1D(τ ). (13) The matrix Q(∧2g)Q−1 belongs to the orthogonal group with respect to the quadratic form H. Moreover, a straight forward calculation shows Q(∧2g)Q−1 ∈(cid:26) SL6(Z) ı SL6(Z) if det(g) = 1, if det(g) = −1, for g ∈ U22(Z[ı]). We set Rg =pdet(g)Q(∧2g)Q−1, (14) which is determined modulo sign. Since ∧2(ıg) = −∧2 g, we have Rıg = −Rg. Thus Rg defines a homomorphism U2,2(Z[ı])/hiE4i → ΓH/h±E6i → Aut(DH) Aut(D) ∩ ∩ . 3.3 Monodromy actions on the spaces DH and D Each center of U22(Z[ı]) and U22(1 + ı) is the group hıE4i generated by the scalar matrix ıE4. We have t(¯g · τ ) = g · tτ for any g ∈ U22(Z[ı]) and τ ∈ D. Let tp be the transpose operator acting on D and htpi be the group generated by tp. The fixed locus of tp is the Siegel upper half space H2 = {τ ∈ D 22(Z[ı]) acting on D as the group generated by U22(Z[ı])/hıE4i and htpi with relations tτ = τ} of degree 2. We define U tp (tp)g = ¯g(tp) for any g ∈ U22(Z[ı]). This group is a semi-direct product (U22(Z[ı])/hıE4i) ⋊ htpi. We set U tp 22(1 + ı) = (U22(1 + ı)/hıE4i) ⋊ htpi. Thomae type formula for K3 surfaces 17 Proposition 6 ([Ma],[KiM],[Y]) (1) We define the principally congru- ence subgroup ΓH(2) of level 2 by ΓH(2) = {R ∈ ΓH R ≡ E6 mod 2}. Then the monodromy group for fper : eX → DH is equal to ΓH (2). (2) The monodromy group of fper : eX → D over X is equal to 22 (1 + ı) = {(g, tpk) ∈ U tp U M 22(1 + ı) det(g) = (−1)k}. We note that det(g) is well defined on U22(Z[ı])/hıE4i. The monodromy group over X/hasi is equal to U tp 22(1 + ı). The map per ◦ pl∗−1 induces the isomorphism from P4 to the Satake compactification of the quotient D/U tp 22(1 + ı). (3) Let As be an element in GL6(Z) defined by As(ωij) = ωij ω13 ω24 (i, j) 6= (1, 3), (2, 4) (i, j) = (2, 4) (i, j) = (1, 3), and as the association involution defined in §2.5. Then we have ω(as(x)) = As(ω(x)) (15) for x in a small neighborhood of x ∈ X. (4) Under the isomorphism Aut(DH) ≃ Aut(D), the matrix As defined in (3) corresponds to tp. Therefore we have τ (as(x)) = tp(τ (x)) for x in a small neighborhood of x ∈ X and an isomorphism eΓH(2) ≃ U tp where eΓH(2) = ΓH (2)/h±E6i · hAsi. By the above proposition, we have a map 22(1 + ı), per : X → DH/ΓH (2) ≃ D/U M 22 (1 + ı). (16) (17) Thomae type formula for K3 surfaces 18 Since the last components D(τ ) and D(g · τ ) are 1, we have ± det(g21τ (x) + g22) = ±pdet(g)[Rgω(x)]34 ω34(x) by the equality (13) together with (12) and (14), where [Rgω(x)]34 denotes the (34)-component of the column vector Rgω(x). By squaring this equality, we have det(g) det(g21τ (x) + g22)2 = for g ∈ U22(Z[ı])/hıE4i. [Rgω(x)]2 34 ω34(x)2 (18) 3.4 Theta functions and their functional equations The theta function Θab with characteristic a, b on D is defined as Θab(τ ) = Xn∈Z[ı]2 e[ 1 2 (n + a)τ (n + a)∗ + Re((n + a)b∗)], (19) where x∗ = t ¯x, e[x] = exp(2πıx), τ ∈ D, n = (n1, n2) ∈ Z[ı]2, a = (a1, a2), b = (b1, b2) ∈ Q[ı]2. Remark 1 This Θab is different from that defined in [MY] and [Ma] by the factor of e[Re(ab∗)] . If τ belongs to the Siegel upper half space H2 of degree 2, then Θab decomposes into the product of Riemann's theta constants: Θab(τ ) = ϑRe(a)Re(b)(τ )ϑIm(a)Im(b)(τ ), where for a′, b′ ∈ Q2. ϑa′b′(τ ) = Xn∈Z2 e[ 1 2 (n + a′)τ t(n + a′) + (n + a′) tb′] This function satisfies Θab( tτ ) = Θ¯a,¯b(τ ), Θa+n,b(τ ) = Θab(τ ), Θa,b+n(τ ) = e[Re(an∗)]Θab(τ ), for any n ∈ Z[ı]2. For a, b ∈ Z[ı]2, Θ a Θ 2 [ab](τ ) depends only on the class of a and b in F2 set (τ ) is denoted by Θ[ab](τ ). Then 2 ≃ (Z[ı]/(1 + ı)Z[ı])2. We 1+ı b 1+ı Ev = {(a, b) ∈ (Z[ı]/(1 + ı)Z[ı])2 ab∗ = 0 mod (1 + ı)}. Thomae type formula for K3 surfaces 19 h123i ↔ [1111] h124i ↔ [0011] h125i ↔ [0010] h134i ↔ [0001] h135i ↔ [0000] h145i ↔ [1100] h234i ↔ [1001] h235i ↔ [1000] h245i ↔ [0100] h345i ↔ [0110] Table 1: Correspondence between hJi and [ab] Then we have #Ev = 10 and Θ[ab](τ ) = 0 if (a, b) /∈ Ev. We identify the sets P3,3 and Ev by the rule given in Table 1. Under this correspondence, [ab](τ ) is denoted by Θ 2 Θ 2 hJi (τ ). Remark 2 The correspondence between P3,3 and Ev is different from that in [Ma], since the bases of the transcendental lattice T (x) are different. Proposition 7 ([Ma]) (1) They satisfy for any g =(cid:18) g11 g21 (2) We define a map θ : D/U tp Θ 2 [ab]( tτ ) = Θ 2 Θ 2 [ab](g · τ ) = det(g) det(g21τ + g22)2Θ 2 [ab](τ ), [ab](τ ), g12 g22(cid:19) ∈ U22(1 + ı). τ 7→ [. . . , Θ 2 hJi 22(1 + ı) → P9 by (τ ), . . .]hJi∈P3,3 ∈ P9. Then the map pl is equal to the composite θ ◦ per from X to P9. Theorem 1 (2τ -formula) We have 4Θab(2τ ) =Xq∈F2 2 e[−Re(aq∗)]Θ(1+ı)a, b+q 1−ı (τ ), where q runs over the set F2 2 = {(0, 0), (0, 1), (1, 0), (1, 1)}. In particular, Θ[0000](2τ ) = Θ[0100](2τ ) = Θ[1000](2τ ) = Θ[1100](2τ ) = 1 4 1 4 1 4 1 4 (Θ[0000](τ ) + Θ[0001](τ ) + Θ[0010](τ ) + Θ[0011](τ )), (Θ[0000](τ ) − Θ[0001](τ ) + Θ[0010](τ ) − Θ[0011](τ )), (Θ[0000](τ ) + Θ[0001](τ ) − Θ[0010](τ ) − Θ[0011](τ )), (Θ[0000](τ ) − Θ[0001](τ ) − Θ[0010](τ ) + Θ[0011](τ )). Thomae type formula for K3 surfaces 20 Proof. We consider the summation Xn′∈LXq∈F2 2 e[ 1 2 Re(n′q∗)] · e[(n′ + a)τ (n′ + a)∗ + Re((n′ + a)b∗)], (20) 1+ı Z[ı]2. Since Z[ı]2 ⊂ L, L/Z[ı]2 ≃ F2 where L = 1 2Re(n′q∗)] (q ∈ F2 2) are the characters of the quotient group L/Z[ı]2, this summation reduces to the four times of the summation over the subgroup Z[ı]2: 2 and e[ 1 e[(n + a)τ (n + a)∗ + Re((n + a)b∗)] = 4Θab(2τ ). On the other hand, the summation (20) is e[ 1 2Re( n 1+ı q∗)] · e[( n 1+ı + a)τ ( n 1+ı + a)∗ + Re(( n 1+ı + a)b∗)] 4 Xn∈Z[ı] Xq∈F2 Xn∈Z[ı]2 = Xq∈F2 = Xq∈F2 e[− e[− 1 2 1 2 Re(aq∗)] Xn∈Z[ı]2 e[ 1 2 (n + (1 + ı)a)τ (n + (1 + ı)a)∗] ×e[Re((n + (1 + ı)a)( b+q 1−ı )∗)] Re(aq∗)]Θ(1+ı)a, b+q (τ ). 1−ı For a ∈ 1 1+ı F2 2 and b = (0, 0), we have the rests. Corollary 1 Θ 2 [0001](2τ ) + Θ 2 [1111](2τ ) = Θ 2 [0010](2τ ) + Θ 2 [1111](2τ ) = Θ 2 [0011](2τ ) + Θ 2 [1111](2τ ) = Θ[0000](τ ) + Θ[0010](τ ) Θ[0001](τ ) + Θ[0011](τ ) 2 2 Θ[0000](τ ) + Θ[0001](τ ) Θ[0010](τ ) + Θ[0011](τ ) 2 2 Θ[0000](τ )Θ[0011](τ ) + Θ[0001](τ )Θ[0010](τ ) 2 . Proof. By Proposition 7 and Plucker relations, we have (cid:3) , , Θ 2 Θ 2 Θ 2 [0001](2τ )+Θ 2 [0010](2τ )+Θ 2 [0011](2τ )+Θ 2 [1111](2τ ) = Θ 2 [1111](2τ ) = Θ 2 [1111](2τ ) = Θ 2 [0000](2τ )−Θ 2 [0000](2τ )−Θ 2 [0000](2τ )−Θ 2 [0100](2τ ), [1000](2τ ), [0100](2τ )−Θ 2 [1000](2τ )+Θ 2 [1100](2τ ). By Theorem 1, we have the corollary. (cid:3) Thomae type formula for K3 surfaces 21 4 Thomae type formula for K3 surfaces 4.1 Main Theorem Theorem 2 Suppose that x is in a neighborhood U of our reference point x ∈ M×(3, 6). Let τ be an element of D defined in (11). Then we have Θ 2 hJi(τ ) = 1 4π4 xhJiω34(x)2 (21) for any hJi ∈ P3,3 Remark 3 Using the notations in Proposition 4, the above value is equal to ν34hJiFS(z)2. Proof. By the first statement of Proposition 4, xhJiω34(x)2 is a holomorphic function on DH. We use actions of eΓH (2) (defined in Proposition 6) and U tp 22(1 + ı) on the domains DH and D to compare two functions Θ 2 (τ ) and hJi xhJiω34(x)2. 2,2(1+ı), τ ∈ D) and J2(R, ω) (R ∈eΓH(2), ω ∈ Lemma 3 Let J1(g, τ ) (γ ∈ U tp DH) be two cocycles defined by J1(g, τ ) = Θ 2 hJi Θ 2 hJi (gτ ) (τ ) , J2(R, ω) = Rω2 34 ω2 34 . (22) Then they coincide via the isomorphisms (10) and (17). Proof. Since the group U tp it is enough to show the identity (22) for 2,2(1 + ı) is generated by U22(1 + ı)/hıE4i and tp, (1) g ∈ U22(1 + ı)/hıE4i and R = Rg, and (2) g = tp and R = T . In the case (1), the statement follows from the equality (18), and that for (2) follows from Proposition 7 (1) and Proposition 6 (4). (cid:3) Thomae type formula for K3 surfaces 22 By the above lemma, the function f (x) = Θ 2 hJi (τ (x))/(xhJiω34(x)2) be- comes a function on D/U tp 22 ≃ X/hasi. The space X/hasi can be compactified by the embedding pl∗ : X/hasi → P4. It is shown in [Ma] that the zero of ΘhJi(τ (x))2 coincides with that of xhJi. Hence f (x) is a constant map. We evaluate this constant by taking the degeneration for z2 → 0, z3 → 0 in the affine open set of X defined by n 1 1 −z1 −1 0 −z2 0 1 0 0 1 1 0 0 −z3 1 −1 −z4  z1, . . . , z4 ∈ Co. Under this limit, we have ω13(x), ω24(x) → 0 and ω34 → 2ωB(z1)ωB(z4), ω14 → 2ωA(z1)ωB(z4), ω23 → 2ωB(z1)ωA(z4), where ωA(λ), ωB(λ) are defined in (1). We set ωij = limz2,z3→0 ωij(x). Then we have τ (x) = diag( lim z2,z3→0 ω14 ω34 , ω23 ω34 ), and by xh135i → 1, f (x) = Θ 2 [0000](diag( lim z2,z3→0 ωA(z1) ωB(z1) = ϑ4 [00]( = 1 4π4 ω14 ω34 , )ϑ4 [00]( ω23 ω34 ωA(z4) ωB(z4) ))/(xh135iω2 34) )/(4ω2 B(z1)ω2 B(z4)) by Jacobi's formula (2). (cid:3) 5 Mean iterations 5.1 Mean iteration associated to D4 degeneration In this and next subsections, we apply the main identity (21) to the study of mean iterations. In this subsection, we consider configurations of six lines which contain three lines intersecting at one point. In this degeneration, three A1 singularities on bX ∗ confluent to one D4 singularity. This degeneration is obtained by taking the limit xhJi → 0. We consider the case hJi = h123i. Thomae type formula for K3 surfaces 23 Proposition 8 The two preimages of the map pl on the subvariety defined by xh123i = 0 are expressed as − xh124i−xh125i−xh134i+xh135i −xh134i+xh135i 0 1 1 1 1 0 − xh124ixh135i−xh125ixh134i (−xh134i+xh135i)xh125i − xh125i−xh124i 1 −1 0 0 0 1 xh134i − xh134i−xh124i − xh135ixh124i−xh125ixh134i (xh135i−xh125i)xh134i Let m be a map from (R×+)4 to (R×+)4 given by 0 0 0 1 −1 − xh124i−xh125i−xh134i+xh135i xh135i−xh125i 1 −1 0 0 0 1 −1 1 0  ,  .   xh125i 1 0 0 1 , m : (R×+)4 ∋ u = (u1, . . . , u4) 7→ (m1(u), . . . , m4(u)) ∈ (R×+)4, (23) where m1(u) = u1 + u2 + u3 + u4 4 m3(u) = p(u1 + u2)(u3 + u4) 2 m2(u) = p(u1 + u3)(u2 + u4) , m4(u) =r u1u4 + u2u3 2 2 . , For an element c = (c1, . . . , c4) ∈ (R×+)4 with c1 > c2 > c3 > c4, we define a vector valued sequence {mn(c) = (mn 4 (c))}n∈N by 1 (c), . . . , mn mn(c) = m ◦ · · · ◦ m(c). z n } { Lemma 4 (1) The components of the sequence {(mn 4 (c))}n∈N (c). The convergence is quadratic. 1 (c), . . . , mn converge and have a common limit m∞ ∗ (2) lim n→∞ Proof. (1) Since mn 1 (c)2 − mn mn 3 (c)2 − mn 2 (c)2 4 (c)2 = lim n→∞ mn 1 (c)2 − mn mn 2 (c)2 − mn 3 (c)2 4 (c)2 = 1. m1(c)2 − m2(c)2 = m2(c)2 − m3(c)2 = m3(c)2 − m4(c)2 = , 16 (c1 − c2 + c3 − c4)2 (c1 − c4)(c2 − c3) (c1 − c2)(c3 − c4) 4 , , 4 Thomae type formula for K3 surfaces 24 we have m1(c) > m2(c) > m3(c) > m4(c) for c1 > c2 > c3 > c4 > 0. We can easily see that c4 < m4(c) < · · · < mn 1 (c)} and {mn 4 (c) < mn 1 (c) < · · · < m1(c) < c1, 4 (c)} converge. We set µ1 = lim n→∞ mn the sequences {mn µ4 = lim n→∞ m1(c)2 − m4(c)2 = m1(c)2 − m2(c)2 + m2(c)2 − m2 4 (c). Since 4(c) mn 1 (c) and (c1 − c2 + c3 − c4)2 16 (c1 − c4)2 + (c1 − c4)2 + = < (c1 − c3)(c2 − c4) 4 (c1 − c4)2 , = 4 4 4 ≤ 1 1 − µ2 2 2(µ1 − µ4)2. If µ1 > µ4, then µ1 + µ4 ≤ 1 2(µ1 − µ4), which By the inequality (24), the convergence of m1(c) − m4(c) = m1(c)2−m4(c)2 m1(c)+m4(c) , we have µ2 implies µ1 + 3µ4 ≤ 0. This is a contradiction. is quadratic. (2) We have (24) m1(c)2 − m2(c)2 m3(c)2 − m4(c)2 = We set sn = 1 (c)2 − mn mn mn 3 (c)2 − mn 2 (c)2 4 (c)2 , tn = Then sn and tn satisfy 1 4(cid:18)c1 − c2 c3 − c4 + 2 + c3 − c4 c1 − c2(cid:19) ≥ 1. mn mn 1 (c) + mn 3 (c) + mn 2 (c) 4 (c) , f (s, t) = 1 4 ( s t + 2 + t s ). sn+1 = f (sn, tn), sn, tn ≥ 1, tn = 1. lim n→∞ Note that f (hs, ht) = f (s, t) for any h ∈ R×+, f (s, s) = 1 and that (s + 2s + s) < s f (s, 1) = f (1, s) = (s + 2 + ) < 1 4 1 s 1 4 for any s > 1. If sn > tn then (sn/tn) > 1 and sn+1 = f (sn, tn) = f ( sn tn , 1) < sn tn < sn. Thomae type formula for K3 surfaces 25 If sn ≤ tn then (tn/sn) ≥ 1 and tn sn sn+1 = f (sn, tn) = f ( tn , 1) ≤ sn ≤ tn. Thus we have sn+1 ≤ max(sn, tn). Since lim tn = 1, for any ε > 0 there n→∞ exists N ∈ N such that tn < 1 + ε for any n > N. If there exists n0 > N sn = 1. such that sn0 ≤ tn0, then sn < 1 + ε for any n ≥ n0; this means lim n→∞ Otherwise, i.e. sn > tn for any n > N, then sn is monotonously decreasing. Thus the limit lim sn exists. Let n → ∞ for sn+1 = f (sn, tn), then we have n→∞ lim n→∞ sn = 1. Similarly we can show lim n→∞ mn mn 1 (c)2−mn 2 (c)2−mn 3 (c)2 4 (c)2 = 1. (cid:3) Theorem 3 The common limit m∞ m∞ ∗ (c) can be expressed as c2 c3 =s c2 3 1 − c2 c2 2 − c2 4 FS(w) 4 2 FS(z) 1 − c2 c2 3 − c2 ∗ (c) =s c2 z4(cid:19) and w =(cid:18) w1 w3 1 − c2 z3 2 1(c2 c2 3(c2 3−c2 4) 1−c2 2) 1 − c2 4 c2 3 w2 w4(cid:19) are given as ! , w = 1 − c2 1 − c2 4 c2 2 2−c2 1(c2 4) 1−c2 c2 2(c2 3) , ! . 0 2−c2 c2 1−c2 4 3 1 − c2 z2 where z =(cid:18) z1 z = 1 − c2 4 3−c2 1−c2 c2 0 Remark 4 For a given c = (c1, . . . , c4), the hypergeometric series FS in Theorem 3 may not converge. By Lemma 4 (2), there exists n ∈ N such that it converges for mn(c) instead of c. Proof. There exists τ ∈ D such that Θ[1111](τ ) = 0 and Θ[0000](τ ) : Θ[0001](τ ) : Θ[0010](τ ) : Θ[0011](τ ) = c1 : c2 : c3 : c4. By Corollary 1, we have (Θ[0000](2τ ), Θ[0001](2τ ), Θ[0010](2τ ), Θ[0011](2τ )) = m(Θ[0000](τ ), Θ[0001](τ ), Θ[0010](τ ), Θ[0011](τ )), since Θ[1111](τ ) = 0. By the homogeneity of m1, . . . , m4, m∞ ∗ satisfies m∞ ∗ (c) = c1m∞ ∗ (1, c2 c1 , c3 c1 , c4 c1 ). Thomae type formula for K3 surfaces 26 Thus = = = = → m∞ ∗ (c) = c1m∞ c1 m∞ ∗ Θ[0000](τ ) Θ[0001](τ ) Θ[0000](τ ) Θ[0010](τ ) Θ[0000](τ ) , ∗ (1, (Θ[0000](τ ), Θ[0001](τ ), Θ[0010](τ ), Θ[0011](τ )) ) , Θ[0011](τ ) Θ[0000](τ ) c1 Θ[0000](τ ) c1 Θ[0000](τ ) c1 Θ[0000](τ ) c1 Θ[0000](τ ) m∞ ∗ m∞ ∗ (m(Θ[0000](τ ), Θ[0001](τ ), Θ[0010](τ ), Θ[0011](τ ))) (Θ[0000](2τ ), Θ[0001](2τ ), Θ[0010](2τ ), Θ[0011](2τ )) (Θ[0000](2nτ ), Θ[0001](2nτ ), Θ[0010](2nτ ), Θ[0011](2nτ )) m∞ ∗ as n → ∞, since Θ[ab](2nτ ) converge to 1 for a = (0, 0) and any b ∈ F2 8, the preimages of 2. By Proposition [xh123i, xh124i, xh125i, xh134i, xh135i] = [0, c2 4, c2 3, c2 2, c2 1] for the map pl : X → P4 ⊂ P9 are given by 0 0 1 0 1 −1 1 1 −z1 −1 0 0 x = 4 3 − c2 c2 c2 1 − c2 2 , z2 = 0, z3 = 1 −  , 1 1 0 −z3 0 −z4 3 − c2 1(c2 c2 4) c2 3(c2 1 − c2 2) , z4 = 1 − c2 4 c2 3 , z1 = 1 − with and z1 = 1 − c2 4 c2 2 , pxh135i = c1 c3sc3 4 3 − c2 c2 1 − c2 2 , x = 0 0 1 1 1 0 −z1 −1 1 −1 0 −z2 2 − c2 4) z2 = 1 − 1 − c2 3) c2 1(c2 2(c2 c2 , 1 1 0 0 0 −z4  , z3 = 0, z4 = 1 − 4 c2 2 − c2 1 − c2 c2 3 , Thomae type formula for K3 surfaces with pxh135i = Theorem 2 implies this theorem. . 27 (cid:3) c1 c2sc2 4 2 − c2 c2 1 − c2 3 5.2 Mean iteration associated to Kummer locus The Borchardt's mean iteration is obtained from the restriction of Thomae type formula for K3 surfaces to the Kummer locus. In this subsection, we explain how to recover limit formulas in [B], [MT] and [Me] from our main theorem. Proposition 9 Let c1 > c2 > c3 > c4 be real numbers such that c1 − c2 − c3 + c4 > 0. We set Q1 = (c1+c2+c3+c4)(c1+c2−c3−c4)(c1−c2+c3−c4)(c1−c2−c3+c4), 4−√Q1) 4−√Q1) 2+c2 3−c2 3−c2 1−c2 2c2(c1c3−c2c4) 1+c2 2c2(c1c2−c3c4) 2−c2 1 − c1(c2 1 − c4(c2 2+c2 3−c2 3−c2 1−c2 2c3(c1c3−c2c4) 1+c2 2c3(c1c2−c3c4) 2−c2 4−√Q1) 4−√Q1) 1 1 −z1 −1 0 −z2 0 1 0 0 1 1 0 0 −z3 1 −1 −z4  ! . (25) (26) and (cid:18) z1 z2 Then z3 z4(cid:19) = 1 − c4(c2 1 − c1(c2 x = lies on the Kummer locus. In this case, we have where [xh123i : xh135i : xh134i : xh125i : xh124i] = [c2 4 + √Q1 c2 0 = c2 1 − c2 3 + c2 2 − c2 2 . 0 : c2 1 : c2 2 : c2 3 : c2 4], Let m be a map from (R×+)4 to (R×+)4 given by m : (R×+)4 ∋ u = (u1, . . . , u4) 7→ (m1(u), . . . , m4(u)) ∈ (R×+)4, (27) Thomae type formula for K3 surfaces 28 where m1(u) = m3(u) = u1 + u2 + u3 + u4 √u1u3 + √u2u4 4 , m2(u) = m4(u) = , 2 √u1u2 + √u3u4 , √u1u4 + √u2u3 2 . 2 Note that if u1 > u2 > u3 > u4 then m1(u) > m2(u) > m3(u) > m4(u). For an element c = (c1, . . . , c4) ∈ (R×+)4 with c1 > c2 > c3 > c4, we define a vector valued sequence {mn(c) = (mn 4 (c))}n∈N by 1 (c), . . . , mn mn(c) = m ◦ · · · ◦ m(c). z n } { In [B] and [Me], they prove that the common limit m∞ ∗ (c) is expressed in terms of period integrals of a hyperelliptic curve of genus 2. In [MT], they give its expression in terms of the period integral ω34(x) of the K3 surface X (x). Here, we give its expression by the hypergeometric series FS. Theorem 4 We can express the common limit m∞ ∗ (c) by m∞ ∗ (c) = = where 4pc2c3(c1c2 − c3c4)(c1c3 − c2c4) (√d1d2 − √d3d4)(√d1d3 − √d2d4) 4pc2c3(c1c2 − c3c4)(c1c3 − c2c4) (√d1d2 + √d3d4)(√d1d3 + √d2d4) 1 FS(z) 1 FS(w) z2 z =(cid:18) z1 w =(cid:18) z1 z2 c2 z3 1 − c1 z4(cid:19) = 1 − c4 z4(cid:19) = 1 − c4 1 − c1 z3 c2 c2 c2 √d1d3−√d2d4 √d1d3+√d2d4 √d1d2−√d3d4 √d1d2+√d3d4 √d1d3+√d2d4 √d1d3−√d2d4 √d1d2+√d3d4 √d1d2−√d3d4 c3 c3 1 − c1 1 − c4 1 − c1 1 − c4 c3 c3 √d1d3−√d2d4 √d1d3+√d2d4 √d1d2−√d3d4 √d1d2+√d3d4! , √d1d2−√d3d4! , √d1d3+√d2d4 √d1d3−√d2d4 √d1d2+√d3d4 d1 = c1 + c2 + c3 + c4, d3 = c1 − c2 + c3 − c4, d2 = c1 + c2 − c3 − c4, d4 = c1 − c2 − c3 + c4. Thomae type formula for K3 surfaces 29 Proof. Let c1, . . . , c4 be elements in R×+ with c1 > · · · > c4. Though the value c1 − c2 − c3 + c4 may be negative, by the inequality 1 4 m1(c)−m2(c)−m3(c)+m4(c) = (√c1−√c2−√c3+√c4)2 > 0, we may assume c1 − c2 − c3 + c4 ≥ 0 by applying the map m. Let(cid:18) z1, z2 z3, z4 (cid:19) be 2 × 2 matrix defined as (25). Then x defined in (26) lies on the Kummer locus. In this case, the theta constants Θ[a,b](τ ) coincide with the square of Riemann's theta constants. Using 2τ -formulas for Riemann's theta con- stants, and similar argument as in Theorem 3, we have m∞ ∗ (c) = c1 Θ[0000](τ ) . By Theorem 2 and Proposition 4, we have the first expression of m∞ ∗ FS. By putting 1 − c2 c2 we have the other expression of m∞ c2 0 = 3 + c2 2 − c2 2 ∗ (c). 4 − √Q1 , (c) by (cid:3) 6 Functional equations for FS The common limit m∞ ∗ (c) in Theorem 3 (resp. 4) satisfies m∞ ∗ (c) = m∞ ∗ (m(c)). This property implies functional equations for the hypergeometric function FS. Theorem 5 We have the following functional equations for FS: FS(m(z)) = FS(m(w)) = (c1−c2+c3−c4)(c3+c4) (c1+c2−c3−c4)(c2+c4) 4(c1−c2)c3 4(c1−c3)c2 FS(z), FS(w), Thomae type formula for K3 surfaces 30 where z and w are given in Theorem 3 and (c1−c2+c3−c4)2 2(c1c4−c2c3)(c2 3+c2 4) (c1+c2)(c3+c4)(c1−c2+c3−c4)2 m(z) = (c1−c2−c3+c4)2 m(w) = 0 (c1−c3)(c2−c4) (c1+c3)(c2+c4) 1−c2 2−c2 2−c2 1−c2 (c1−c2)(c3−c4) (c1+c2)(c3+c4) 0 ! , (c1+c2−c3−c4)2! . √d2d3(√d1d2 − √d3d4)(√d1d3 − √d2d4) 1 √c2c3(√c1c2 − √c3c4)(√c1c3 − √c2c4) 16 (√d1d2 + √d3d4)(√d1d3 + √d2d4) 2(c1c4−c2c3)(c2 3+c2 4) (c1+c3)(c2+c4)(c1+c2−c3−c4)2 (c1−c2−c3+c4)2 FS(w), Theorem 6 We have the following functional equations for FS: FS(m(z)) = FS(z), FS(m(w)) = 4√c2c3d2d3 where z and w are given in Theorem 4 and m(z) = 1 − 2(√c1c4+√c2c3)(√c1c3−√c2c4) m(w) = 1 − 1 − (√c1c2−√c3c4)d1 2(√c1c2+√c3c4)(√c1c3−√c2c4) (√c1c2+√c3c4)d3 (√c1c2+√c3c4)d2 (√c1c4+√c2c3)d3 d1d2 1 − 4(c1c2−c3c4) (√c1c3+√c2c4)d3 (√c1c3+√c2c4)d2 ! , 1 − (√c1c3−√c2c4)d1 1 − 2(√c1c4+√c2c3)(√c1c2−√c3c4) 2(√c1c3+√c2c4)(√c1c2−√c3c4)! . (√c1c4+√c2c3)d2 4(c1c3−c2c4) 1 − 1 − d1d3 Proof. To obtain the expression of m(z) and m(w), we use the equalities: m1(c) + ǫ1m2(c) + ǫ2m3(c) + ǫ1ǫ2m4(c) = (√c1 + ǫ1√c2 + ǫ2√c3 + ǫ1ǫ2√c4)2 1 4 for ǫ1, ǫ2 = ±1 and mi(c)mj(c)−mk(c)ml(c) = 1 8 (√cicj−√ckcl)(ci +cj−ck−cl) for (i, j, k, l) = (1, 2, 3, 4), (1, 3, 2, 4), (1, 4, 2, 3). (cid:3) References [B] C.W. Borchardt, Uber das arithmetisch-geometrische Mittel aus vier Elementen, Berl. Monatsber, 53 (1876), 611-621. Thomae type formula for K3 surfaces 31 [F] E. Freitag, Modulformen zweiten Grades zum rationalen und Gaussschen Zahlkorper, Sitzungsber. Heidelb. Akad. Wiss., 1 (1967), 1 -- 49. [I] J. Igusa, Theta functions, Die Grundlehren der mathematischen Wis- senshaften in Einzeldarstellungen 194, Springer-Berlin-Heidelberg, New York, 1972. [IKSY] K. Iwasaki, H. Kimura, S. Shimomura and M. Yoshida, From Gauss to Painlev´e, Vieweg, Braunschweig, Wiesbaden, 1991. [KY] M. Kita and M. Yoshida, Intersection theory for twisted cycles I, Math. Nachr. 166 (1994), 287 -- 304. [KaM] T. Kato and K. Matsumoto, The common limit of a quadruple se- quence and the hypergeometric function FD of three variables, Nagoya Math. J. 195 (2009). 113 -- 124. [KiM] M. Kita and K. Matsumoto, Duality for hypergeometric functions and invariant Gauss-Manin systems, Compositio Math. 108 (1997), 77 -- 106. [MSY] K. Matsumoto, T. Sasaki and M. Yoshida, The monodromy of the period map of a 4-parameter family of K3 surfaces and the Aomoto- Gel'fand hypergeometric function of type (3,6), Internat. J. of Math., 3 (1992), 1 -- 164. [MT] K. Matsumoto and T. Terasoma, Arithmetic-geometric means for hy- perelliptic curves and Calabi-Yau varieties, to appear in Internat. J. of Math. [MY] K. Matsumoto and M. Yoshida, Invariants for some real hyperbolic groups, Internat. J. of Math., 13 (2002), 415 -- 443. [Ma] K. Matsumoto, Theta functions on the bounded symmetric domain of type I2,2 and the period map of 4-parameter family of K3 surfaces, Math. Ann., 295 (1993), 383 -- 408. [Me] J. Mestre, Moyenne de Borchardt et integrales elliptiques, C. R. Acad. Sci. Paris Ser. I Math. 313 (1991), no. 5, 273 -- 276. [Mu] D. Mumford, Tata lectures on Theta I, progress in Math 28. Birkhauser, Boston-Basel-Berlin, 1983. Thomae type formula for K3 surfaces 32 [Te] T. Terasoma, Exponential Kummer coverings and determinants of hy- pergeometric functions. Tokyo J. Math. 16 (1993), no. 2, 497 -- 508. [To] J. Thomae, Beitrag zur Bestimmung von θ(0, 0, ..., 0) durch die Klassen- moduln algebraischer Funktionen, J. Reine Angew. Math. 71 (1870), 201 -- 222. [Y] M. Yoshida, Hypergeometric Functions, My Love, Aspects of Mathe- matics, E32, Friedr Vieweg & Sohn, Braunschweig, 1997. Keiji Matsumoto Department of Mathematics Hokkaido University Sapporo 060-0810, Japan e-mail:[email protected] Tomohide Terasoma Graduate School of Mathematical Sciences The University of Tokyo Komaba, Meguro, Tokyo, 153-8914, Japan e-mail:[email protected] ∆13 l4( x) ∆4 l5( x) s2 ∆ij = ∆i ∩ ∆j ∆14 ∆34 ∆12 ∆23 ∆2 ∆1 l6( x) l1( x) ∆24 s1 l3( x) ∆3 l2( x) A1 A2 p1 p2 p3 p4 p5 p6 B1 B2
1312.4822	1	1312	2013-12-17T15:22:09	N\'eron models of algebraic curves	[ "math.AG", "math.NT" ]	Let S be a Dedekind scheme with field of functions K. We show that if X_K is a smooth connected proper curve of positive genus over K, then it admits a N\'eron model over S, i.e., a smooth separated model of finite type satisfying the usual N\'eron mapping property. It is given by the smooth locus of the minimal proper regular model of X_K over S, as in the case of elliptic curves. When S is excellent, a similar result holds for connected smooth affine curves different from the affine line, with locally finite type N\'eron models.	math.AG	math	N ´ERON MODELS OF ALGEBRAIC CURVES QING LIU AND JILONG TONG Dedicated to Michel Raynaud on the occasion of his seventy-fifth birthday Abstract. Let S be a Dedekind scheme with field of functions K . We show that if XK is a smooth connected proper curve of positive genus over K , then it admits a N´eron model over S , i.e., a smooth separated model of finite type satisfying the usual N´eron mapping property. It is given by the smooth locus of the minimal proper regular model of XK over S , as in the case of elliptic curves. When S is excellent, a similar result holds for connected smooth affine curves different from the affine line, with locally finite type N´eron models. 1. Introduction In 1964, A. N´eron [11] introduced the notion of N´eron models (see Definition 2.2) of abelian varieties over the fraction field of a Dedekind domain, and proved the existence of these models (see the introduction of [1] for a detailed presentation). Since then, this notion has been generalized to smooth commutative algebraic groups and to torsors under these groups (see [1], §6.5). In this work we investigate the case of smooth proper or affine curves. Let S be a Dedekind scheme, that is, a noetherian regular connected scheme of dimension 1. Let K = K (S ) be its field of functions. Let XK be a smooth connected curve over K . When XK is proper of positive genus, a canonical smooth model is the smooth locus Xsm of the minimal proper regular model of XK over S . When XK is an elliptic curve, it is well known that Xsm is the N´eron model of XK (see [11] or [8], 10.2.14). The first main result of this work is a generalization of the latter fact to higher genus. Theorem 1.1. (Theorem 4.1) Let XK be a proper smooth connected curve of positive genus over K . Then the smooth locus Xsm of the minimal proper regular model of XK over S is the N´eron model of XK . Date : December 18, 2013. 2000 Mathematics Subject Classification. 14H25, 14G20, 14G40, 11G35. Key words and phrases. N´eron model, curve, good reduction. 1 N ´ERON MODELS OF ALGEBRAIC CURVES 2 When S is excellent, we actually prove a slightly more general re- sult: for a regular proper connected curve XK /K of positive arithmetic genus, the smooth locus of XK admits a N´eron model over S , equal to the smooth locus of the minimal proper regular model of XK over S . As an immediate consequence of Theorem 1.1, we have the next corollary. Corollary 1.2. Let XK be as in Theorem 1.1. Let Y be a smooth scheme over S and let fK : YK → XK be a morphism of K -schemes. Then (1) the morphism fK extends uniquely to a morphism of S -schemes Y → Xsm ; (2) (Corollary 4.7) if Y is proper over S (i.e., YK has good reduction over S ) and fK is dominant, then Xsm is proper over S and XK has good reduction over S . In the second part of this paper (§5–§7), we consider N´eron lft-models (N´eron model locally of finite type) of smooth affine curves. The main result of this second part is Theorem 1.3. (Theorem 7.10) Suppose S is excel lent. Let UK be an affine smooth and geometrical ly connected curve over K , not isomor- phic to A1 K . Then UK admits a N´eron lft-model U over S . Note that in general, the scheme U in the above theorem is not of finite type. However, necessary and sufficient conditions (in terms of the points at infinity of UK ) can be found in Proposition 7.11 to insure that U is of finite type over S . The paper is organized as follows. Some basic properties of N´eron lft-models are assembled in §2. In §3, we prove a crucial technical result (Proposition 3.8) on the image of a morphism f : Y → X from a smooth S -scheme to a normal relative curve over S . In §4, we prove the main theorem 4.1 on the existence of N´eron models for proper smooth curves of positive genus. In §5, we study the N´eron lft-models of open subschemes of a curve having already a N´eron lft-model (Theorem 5.1). §6 is devoted to the existence of the N´eron model of some special affine open subschemes of a smooth conic over local Dedekind schemes S . Finally we prove Theorem 1.3 in §7. The present work grew from a partial answer to a question asked by an anonymous poster at mathoverflow.net/questions/110359/, on the existence of N´eron models of pro jective curves. We would like to thank A. Javanpeykar and M. Raynaud for their interests in this work and especially for their comments which improve the presentation of the paper. N ´ERON MODELS OF ALGEBRAIC CURVES 3 Notation: In all this paper, unless explicitly mentioned, the letter S denotes a Dedekind scheme (that is, a noetherian regular connected scheme of dimension 1), K denotes its field of functions K (S ). Symbols such as XK , YK , UK usually denote a scheme over K . On the other hand, for any S -scheme X , XK also denotes the generic fiber of X . 2. Basic properties Let S, K be as above. Definition 2.1 Let XK be a separated algebraic variety (i.e. sep- arated scheme of finite type) over K . A model of XK over S is a locally finite type, separated and flat1 scheme over S endowed with an isomorphism from its generic fiber to XK . Definition 2.2 ([1] 10.1/1, 1.2/1) Let XK be a separated smooth algebraic variety over K . A N´eron lft-model of XK over S or an S - N´eron lft-model of XK is a smooth model X of XK over S satisfying the following universal property, called N´eron mapping property : for any smooth scheme Y → S , the canonical map (once the generic fiber of X is identified with XK ) (2.1) MorS (Y , X ) → MorK (YK , XK ) is a bijection. A N´eron lft-model of finite type is called a N´eron model. Remark 2.3 (1) The universal property above implies the unique- ness (up to a unique isomorphism) of the N´eron lft-model if it exists. (2) Let X be a model of XK . As X is separated, the map (2.1) is always injective. So it is enough to check the surjectivity for the N´eron mapping property. By the injectivity, when S is local, it is also enough to check the surjectivity with Y smooth of finite type having irreducible fibers over S . (3) Let XK be a separated smooth algebraic variety over K . If each connected component of XK admits an S -N´eron lft-model, then XK has also an S -N´eron lft-model given by the disjoint union of the S -N´eron lft-models of the connected components. This holds similarly for N´eron models. Proposition 2.4. Let X be an S -model of XK . Let S ′ → S be a morphism of Dedekind schemes. Denote by K ′ the function field of S ′ . 1In this work, we do not require the models to be faithfully flat. N ´ERON MODELS OF ALGEBRAIC CURVES 4 (1) Assume that the morphism S ′ → S is faithful ly flat, and that XS ′ := X ×S S ′ is the N´eron lft-model (resp. N´eron model) of XK ′ := X ×Spec(K ) Spec(K ′ ) over S ′ . Then X is the N´eron lft-model (resp. N´eron model) of XK over S . (2) Assume that S ′ → S can be written as a filtered inverse limit of affine smooth schemes of finite type over S , and that X is the N´eron lft-model (resp. N´eron model) of XK over S . Then the base change XS ′ is the N´eron lft-model (resp. N´eron model) of XK ′ over S ′ . (3) Assume that S ′ → S is an extension of local Dedekind schemes of ramification index 12 with S excellent, and that X is the N´eron lft-model over S . Then, XS ′ is the N´eron lft-model over S ′ . Proof. (1) This is a direct application of fpqc descent, so we omit the details here. (2) Let Y ′ be a smooth S ′ -scheme, and let fK ′ : Y ′ K ′ → XK ′ be a morphism of K ′ -schemes. Without loss of generality, we may and do assume that the scheme Y ′ is quasi-compact, hence is of finite presen- tation over S ′ . Consequently, the S ′ -scheme Y ′ (resp. the morphism fK ′ : Y ′ K ′ → XK ′ ) descends to an S0 -scheme Y0 (resp. to a morphism of S0,K -schemes f0,K : Y0×S0 S0,K → XS0×S0 S0,K ) for some (affine) smooth morphism S0 → S ([5], IV 8.8.2). Thus, we only need to prove that the morphism f0,K can be extended to a morphism f0 : Y0 → X ×S S0 . Con- sider the composite Y0 → S0 → S , which defines a smooth S -scheme. The morphism f0,K gives a morphism of K -schemes from Y0×S Spec(K ) to XK . As X is the S -N´eron lft-model of XK , the N´eron mapping prop- erty implies a morphism of S -schemes g : Y0 → X making the following diagram commutative: g Y0 / XS0 ∃f0 !❇❇❇❇❇❇❇❇ S0 / X / S We obtain a morphism f0 : Y0 → XS0 of S0 -schemes extending f0,K , as required. If X is the S -N´eron model of XK , then XS ′ is of finite type over S ′ and is the S ′ -N´eron model of XK ′ . (3) Write S = Spec(R) and S ′ = Spec(R′ ). As S is excellent, by [1], 3.6/2, the morphism R → R′ is regular. Consequently, by a result of N´eron ([1], 3.6/8, see also [12], 2.5 for a more general statement), 2see [1], 3.6/1 for the definition, the residue extension is required to be (not necessarily algebraic) separable. ! ( ( / / / N ´ERON MODELS OF ALGEBRAIC CURVES 5 R′ can be written as a filtered inductive limit of smooth R-algebras of finite type. Therefore, XS ′ is the N´eron lft-model over S ′ by (2). (cid:3) Corollary 2.5. Let S be a Dedekind scheme with field of functions K , and let X be an S -scheme local ly of finite type (resp. an S -scheme of finite type) with smooth finite type generic fiber XK . Then X is the S -N´eron lft-model of XK (resp. S -N´eron model of XK ) if and only if for any closed point s ∈ S , X ×S Spec(OS,s) is the Spec(OS,s )-N´eron lft-model of XK (resp. Spec(OS,s )-N´eron model of XK ). Proof. If X/S is of finite type, then this corollary is proved in [1], 1.2/4. The same proof applies in the locally finite type case. For the convenience of the readers, we reproduce the proof in this case here. The direct implication follows from Proposition 2.4 (2). Conversely, if X ×S Spec(OS,s) is the N´eron lft-model of XK over Spec(OS,s ) for any closed point s ∈ S , it follows that X/S is smooth and separated ([5], IV.8.10.5(v)). It remains to check the N´eron mapping property for X/S (Definition 2.2). Consider Y a smooth S -scheme, and fK : YK → XK a morphism of K -schemes, we need to extend fK to a morphism of S -schemes from Y to X . We may assume that Y is quasi-compact, hence is of finite presentation over S . Now, for any closed point s ∈ S , our assumptions imply that one can find an extension fOS,s of fK over Spec(OS,s ). As Y /S is of finite presentation and X/S is locally of finite presentation, we may descend fOS,s to a morphism fs : Y ×S V → X ×S V over some open neighborhood V ⊆ S of s ([5], IV 8.8.2). As X/S is separated, and Y /S is flat, these local extensions of fK are compatible with each other. As a result, they can be glued together to a morphism f : Y → S extending fK , as desired. (cid:3) The next lemma says that we can restrict ourselves to geometrically connected varieties XK without loss of generality. Denote by K an algebraic closure of K . For any closed point s ∈ S we denote by K sh s the fraction field of the strict henselization of the local ring OS,s . Lemma 2.6. Let XK be a separated smooth connected variety over K . Let K ′ = K (XK ) ∩ K be the field of constants of K (XK ), let S ′ be the integral closure of S in K ′ and let T ⊆ S ′ be the ´etale locus of S ′ → S . Then (1) XK is canonical ly a separated smooth and geometrical ly connected variety over K ′ ; (2) XK /K admits an S -N´eron lft-model (resp. S -N´eron model) if and only if dim T = 0, or dim T = 1 and XK /K ′ admits a T -N´eron lft-model (resp. T -N´eron model). N ´ERON MODELS OF ALGEBRAIC CURVES 6 Proof. (1) As K ′ ⊆ OXK (XK ) ⊆ K (XK ) by the normality of XK , the latter has a canonical structure of K ′ -variety. As K ′ is algebraically closed in K (XK ), XK /K ′ is geometrically connected. As XK ×Spec(K ′ ) Spec(K ) is a connected component of XK ×Spec(K ) Spec(K ), XK /K ′ is separated and smooth. (2) Note that because XK is smooth over K , K ′ /K is separable, so S ′ → S is finite. If dim T = 0, then S is semi-local and S ′ → S is ramified at all closed points. This implies that for all closed points s ∈ S , XK has no K sh s -point, so XK is its own S -N´eron model. Suppose now that dim T = 1 and that XK /K ′ has a T -N´eron lft-model X . Then the composition with T → S makes X into a smooth separated S -scheme, with generic fiber XK /K . Let us check that it satisfies the N´eron mapping property. Let Y → S be a smooth scheme and let fK : YK → XK be a K -morphism. Then YK → Spec(K ) also factors through YK → Spec(K ′ ) via XK → Spec(K ′ ). In particular, fK is a K ′ -morphism. On the other hand, as Y is normal, Y → S factors through Y → S ′ . The latter has image in T by Corollary 3.2 and makes Y a smooth T -scheme. So fK extends to a T -morphism f : Y → X , which is a fortiori an S -morphism. Conversely, if XK /K has an S -N´eron lft-model X , the above argu- ments show that X is canonically a smooth separated T -scheme, and the generic fiber of XT is nothing but XK viewed as a scheme over K ′ . If dim T = 1, the N´eron mapping property of X → T is immediate to verify. Finally, as T → S is of finite type and separated, X → T is of finite type if and only if X → S is of finite type. (cid:3) Corollary 2.7. Let XK be a smooth K -variety of dimension zero. Then XK admits a N´eron model over S . Proof. By Proposition 2.3 (3) and Lemma 2.6, we can suppose XK /K is geometrically connected, smooth of dimension 0. Then XK = Spec(K ), and S is clearly the N´eron model of XK over S . (cid:3) The following proposition will be used to see when a N´eron lft-model is a N´eron model (e.g. in Proposition 7.11). Proposition 2.8. Let X → S be a separated morphism local ly of finite type, such that X ×S Spec(OS,s) is of finite type for al l s. (1) If X is irreducible and XK is proper over K , then X is of finite type over S . (2) If XK is affine, Xs is irreducible for al l s ∈ S and S is excel lent, then X is of finite type over S . N ´ERON MODELS OF ALGEBRAIC CURVES 7 (3) If X is of finite type and if XK /K is geometrical ly connected, then Xs is geometrical ly connected for al l s in some dense open subset of S . Proof. (1) Let U be a quasi-compact open subset of X such that XK ⊆ U , and let U ′ be a Nagata compactification of U → S . Then U and U ′ are of finite type over S sharing the same generic fiber. So there exists a dense open subset V of S such that U ×S V ∼= U ′ ×S V is proper over V . The inclusion U ×S V → X ×S V is then open and closed. As X ×S V is irreducible, X ×S V = U ×S V , thus is of finite type over V . The remaining part X ×S (S \ V ) is a finite union of quasi-compact subsets, so X is quasi-compact. (2) As in (1), to prove X is quasi-compact, we are allowed to shrink S . Let U be a quasi-compact open subset of X containing XK , and let W be an affine finite type S -scheme such that WK = XK . As U, W are of finite type over the noetherian scheme S with the same generic fiber, shrinking S if necessary, we can suppose that U = W is affine, and that U → S is surjective (with S affine). We claim that X = U . Let U ′ be any affine open subset of X not contained in U and F := U ′ \ (U ∩ U ′ ). Then F 6= ∅, so it has pure codimension 1 in U ′ because U ∩ U ′ is affine (see [8], Exercise 4.1.15. The hypothesis S excellent implies that the normalization map of U ′ is finite). As FK = ∅, F is then a finite union of vertical divisors, which is impossible since Us is dense in Xs . Consequently, X = U is of finite type over S . (3) This follows from [5], IV.9.7.7 after noticing that, as S is ir- reducible, a dense locally constructible subset of S contains an open dense subset. (cid:3) Remark 2.9 In general, the condition X ×S Spec(OS,s) of finite type over OS,s for all s is not sufficient to conclude that X is of finite type over S , as an example of Oesterl´e ([1], 10.1/11) shows. See also Re- mark 7.12. 3. Image of smooth schemes Let f : Y → X be a morphism of S -schemes. In this section, we study geometric properties of X at the points of f (Y ), when Y is smooth over S . The main result is Proposition 3.6 which states that X is smooth at points of f (Y ) under some mild hypothesis. Its Corollary 3.8 is a principal ingredient of the proof of Theorem 4.1. Let Y be a scheme. For any morphism locally of finite type Z → Y , we denote by sm(Z/Y ) ⊆ Z the smooth locus of Z → Y . This is an open subset of Z if Y is locally noetherian. N ´ERON MODELS OF ALGEBRAIC CURVES 8 Lemma 3.1. Let Z → Y be a morphism local ly of finite type. Sup- pose that Y , Z are local ly noetherian and regular. Then for any section σ : Y → Z , the image σ(Y ) is contained in the smooth locus sm(Z/Y ) of Z → Y . Proof. See [1], 3.1/2. (cid:3) Corollary 3.2. Let S be a local ly noetherian regular scheme. Let f : Y → X be a morphism between two S -schemes local ly of finite type. Suppose that Y is smooth over S , and that X is regular. Then f (Y ) is contained in the smooth locus of X → S . Proof. Consider the Y -scheme Z := X ×S Y . Notice that Z is regular, being smooth over X . The morphism f induces a section σ : Y → Z , y 7→ (f (y ), y ), of the second pro jection Z → Y . By Lemma 3.1, σ(Y ) ⊆ sm(Z/Y ) = sm(X/S ) ×S Y , hence f (Y ) ⊆ sm(X/S ). (cid:3) Corollary 3.2 does not hold in general if we remove the regular- ity hypothesis on X . However, in the situation of relative curves, we can weaken the regularity hypothesis to the normality of X (Proposi- tion 3.6). We first prove some preliminary results. Lemma 3.3. Let S be an irreducible local ly noetherian scheme, and let X, Y be irreducible flat S -schemes local ly of finite type. Let f : Y → X be a dominant S -morphism. Let s ∈ S . Suppose that f is quasi-finite at some point y0 ∈ Ys , and that Ys is irreducible at y0 . Then Xs is irreducible at x0 := f (y0). Proof. The property is local on X and Y . In particular, shrinking X and Y if necessary, we can suppose that f : Y → X is quasi-finite and separated. Thus f can be factorized as Y → Y → X with an open (dense) immersion followed by a finite surjective morphism ([5], IV.8.12.6). Let Z1 , Z2 be two irreducible components of Xs passing through x0 . By the going-down property of Y → X ([10], 5.E.(v)), there exist irreducible closed subschemes F1 , F2 of Y passing through y0 such that the induced maps Fi → Zi (i = 1, 2) are finite and dom- inant (thus surjective). Let η be the generic point of S . As X, Y are irreducible and flat over S , both Xs and Ys are equidimensional of dimension dim Xη = dim Yη ([5], IV.14.2.3). Consequently, dim Fi = dim Zi = dimx0 Xs = dim Xη = dim Yη = dimy0 Ys and Fi ∩ Ys = Ys . Therefore Z1 = Z2 and Xs is irreducible at x0 . (cid:3) N ´ERON MODELS OF ALGEBRAIC CURVES 9 Lemma 3.4. Let S be a local ly noetherian scheme, let X, Y be two S -schemes local ly of finite type, and let f : Y → X be a morphism of S -schemes. Consider s ∈ S , and y0 ∈ Ys a closed point of Ys . Let x0 = f (y0). Suppose that Ys is regular at y0 and and codimy0 (f −1(x0 ), Ys) > 0. dimy0 Ys > 1 (The second inequality means that f is non-constant on the irreducible component of Ys containing y0). Then there exists a subscheme Z of Y passing through y0 such that dimy0 Zs < dimy0 Ys and codimy0 (f −1(x0 ) ∩ Z, Zs) > 0, and Zs is regular at y0 . If furthermore Y is regular (resp. if Y → S is flat) at y0 , we can assume that the same property holds for Z . Proof. We construct Z locally at y0 as a hypersurface defined by some u ∈ my0 OY ,y0 which must avoid some ideals of OY ,y0 . We notice the following facts: (i) Let Γ1 , . . . , Γn be the irreducible components of f −1(x0 ) of codi- mension 1 in Ys passing through y0 . Locally at y0 , each Γi is defined by a prime principal ideal ¯tiOYs ,y0 ⊆ my0 OYs ,y0 because Ys is regular at y0 . As dimy0 Ys > 1, we have my0 OYs ,y0 6⊆ ¯tiOYs ,y0 ; (ii) We have my0 OYs ,y0 6⊆ m2 y0 OYs ,y0 because Ys has positive dimension at y0 . By prime avoidance lemma ([10], 1.B), there exists ¯u ∈ my0 OYs ,y0 \ (cid:0)m y0 OYs ,y0 ∪ (∪i≤n ¯tiOYs ,y0 )(cid:1) . 2 Lift ¯u to some u ∈ my0 OY ,y0 and let Z := V (u) be the subscheme of Y defined in some open neighborhood of y0 . Then: if Y → S is flat) at y0 , (1) Zs is regular at y0 ; if Y is regular (resp. then the same holds for Z because u /∈ m2 y0 OY ,y0 and ¯u is not a zero divisor); (2) dimy0 Zs < dimy0 Ys ; (3) and codimy0 (f −1(x0 ) ∩ Z, Zs) > 0, because otherwise V ( ¯u) would be contained in, hence equal to, some irreducible component of f −1(x0 ), so ¯u ∈ ¯tiOYs ,y0 for some i ≤ n. Contradiction. Therefore Z satisfies the desired properties. (cid:3) Lemma 3.5. Let S be a Dedekind scheme. Let X be a normal relative curve over S ,3 and let Y be a regular scheme which is flat and local ly 3By a relative curve over S , we mean a flat, locally finite type S -scheme with generic fiber of dimension 1. N ´ERON MODELS OF ALGEBRAIC CURVES 10 of finite type over S . Suppose Ys is regular. Let f : Y → X be an S -morphism, and x0 = f (y0) for some y0 ∈ Ys . Suppose that codimy0 (f −1(x0), Ys) > 0. Then Xs is irreducible and reduced at x0 . If Ys is geometrical ly re- duced in a neighborhood of y0 , then Xs is geometrical ly reduced in a neighborhood of x0 . Proof. We can suppose X, Y are integral. Using repeatedly Lemma 3.4, we find a subscheme Z of Y containing y0 , flat over S , such that Zs is irreducible and regular of dimension 1 and f Zs is non-constant. This implies that f Zs is quasi-finite and f Z is dominant. By Lemma 3.3, Xs is irreducible at x0 . Shrinking X and Y if necessary, we can suppose Xs is irreducible. As f Zs is non-constant, Ys → Xs , and hence Y → X , are dominant. Thus Xs is reduced at its generic point ξ because the ramification index of OS,s → OX,ξ is 1. But X is normal, hence S2 , Xs is S1 . This implies that Xs is reduced and Ys → Xs is scheme- theoretically dominant. Then the same property holds over k(s), which implies that Xs is geometrically reduced if Ys is geometrically reduced. (cid:3) Proposition 3.6. Let S be a Dedekind scheme, and let X be a normal relative curve over S with smooth generic fiber. Let f : Y → X be a morphism with Y smooth and Ys irreducible for some closed point s ∈ S . Then either f (Ys) is one point, or X is smooth at every point of f (Ys). Proof. We may assume that S = Spec(R) is local, and that f (Ys) is not one point. Then for all y0 ∈ Ys and x0 := f (y0), we have codimy0 (f −1(x0), Ys) > 0. Let R′ be any discrete valuation ring dom- inating R. Then XR′ := X ⊗R R′ has smooth generic fiber, and its special fiber is reduced at any point x′ 0 lying over x0 by Lemma 3.5. So XR′ is normal at x′ 0 . Therefore, to prove X is smooth at x0 , we can enlarge R and suppose it is complete with algebraically closed residue field. Using Lemma 3.4, we can suppose that Y is a relative smooth curve. Then f is quasi-finite, and Spec bOY ,y0 → Spec bOX,x0 is finite. By a result of Raynaud ([13], Appendice, p. 195), X is smooth at x0 . (cid:3) Remark 3.7 Let X be an integral relative curve over S with smooth generic fiber. Then X admits a minimal desingularization X ′ → X , made of a finite sequence of normalizations and blowing-ups of closed singular points. See e.g. [8], 8.3.50 and 9.3.32 when X is proper over N ´ERON MODELS OF ALGEBRAIC CURVES 11 S . As the construction of X ′ is local on X , and the minimal desin- gularization is unique, the same result holds for any integral relative curve X over S with smooth generic fiber. Corollary 3.8. Let S be a Dedekind scheme, and let X be an integral relative curve over S . Let f : Y → X be an S -morphism from a smooth S -scheme Y to X . Let s ∈ S be a closed point. (1) If f (Ys) is reduced to one point x0 ∈ Xs , then f factors as Y → eX → X , where the second morphism is the blowing-up of X along the reduced center x0 . (2) Suppose that XK is smooth, Y is irreducible and that fK : YK → XK is dominant. Let X ′ → X be the minimal desingularization of X (Remark 3.7). Then f : Y → X factors through X ′ → X . Proof. (1) We have to show that mx0 OY is an invertible sheaf of ideals of OY . Let y ∈ f −1(x0 ) = Ys , and let π be a generator of msOS,s . We have pmx0 OY ,y = pπOY ,y = πOY ,y (because Ys is reduced). As πOY ,y ⊆ mx0 OY ,y , we find mx0 OY ,y = πOY ,y . Therefore mx0 OY is an invertible sheaf of ideals, and Y → X factors through Y → eX → X . (2) As the minimal desingularization commutes with restriction to open subsets, and because the property to prove is local at Y , we can suppose Y is quasi-compact. Then f (Y ) is contained in a quasi- compact open subset of X . Therefore we can also suppose X is quasi- compact. As Y is normal and fK is dominant, f factors through the normal- ization of X . Furthermore, the normalization map of X is finite (see [8], 8.3.49(d)). So we can suppose X is normal. Let F be the singular locus of X , which is a finite closed subset of Xs . Let X1 → X be the blowing-up along F (with the reduced structure). If F ∩ f (Ys) = ∅, then f trivially factors as Y → X1 → X . In general, for any x0 ∈ F ∩ f (Ys), it follows easily from Proposition 3.6 that f −1(x0) is a union of irreducible components of Ys and, by (1), f factors through X1 → X . Similarly as above, the morphism Y → X1 factors through the normalization map X ′ 1 → X1 . Now we start again with Y → X ′ 1 and the process will stop at the minimal desingularization of X . (cid:3) 4. N´eron models of proper smooth curves Let S be a Dedekind scheme with field of functions K . Let XK be a proper regular and connected curve over K , of positive arithmetic N ´ERON MODELS OF ALGEBRAIC CURVES 12 genus (i.e., dim H1(XK , OXK ) > 0). When the base S is excellent or XK is smooth over K , XK admits a unique minimal proper regular model Xmin over S (see [2], Theorem 1.2 or [8], 8.3.45 and 9.3.21). Let Xsm denote the smooth locus of Xmin/S . The aim of this section is to prove the next theorem. See also Proposition 4.12 for a partial result in higher dimension. Theorem 4.1. Let S be a Dedekind scheme with field of functions K . Let XK be a proper regular connected curve of positive arithmetic genus over K . Assume either S is excel lent or XK /K is smooth. Then Xsm is the N´eron model of the smooth locus XK,sm of XK over S . We will deduce Theorem 4.1 from the next proposition. Proposition 4.2 (see also [1], 7.1/6). Let PK be a separated con- nected smooth K -scheme of finite type, and let UK ⊆ PK be a connected smooth closed subscheme of dimension one. Assume that PK admits a N´eron lft-model (resp. N´eron model) P over S . Then UK admits a N´eron lft-model (resp. N´eron model) over S . Proof. Let U0 denote the scheme-theoretic closure of UK inside P . Let p : U → U0 be the minimal desingularization of U0 (Remark 3.7). We want to prove that the smooth locus Usm of U → S is the N´eron lft-model of UK over S . As the formation of Usm commutes with lo- calization and strict henselization ([8], Proposition 9.3.28), by Propo- sition 2.4 we can suppose S = Spec(R) is strictly local (i.e., R is a strictly henselian discrete valuation ring). Let Y be a smooth scheme over S and let fK : YK → UK be a morphism of K -schemes. We want to extend fK to a morphism of S -schemes Y → Usm . We can suppose Ys is irreducible (Remark 2.3 (2)). If YK (K ) = ∅, then Y = YK , and fK is a morphism from Y to U . Suppose YK (K ) 6= ∅. If fK : YK → UK is not dominant, the image fK (YK ) consists of a rational point q of UK . The Zariski closure {q} of {q} in U is contained in Usm (Lemma 3.1) and is the image of a section σ : S → Usm . Then fK extends to Y → Usm as composition of the structure morphism Y → S and the section σ : S → Usm . Now suppose fK is dominant. The morphism YK → UK → PK extends to a dominant morphism Y → U0 . By Corollary 3.8(2), the latter induces a dominant morphism Y → U . Therefore fK extends to Y → U , hence to Y → Usm (Proposition 3.2). This shows that Usm is the N´eron lft-model of UK over S . If P is of finite type over S , then U0 and U above are of finite type over S , thus Usm is of finite type. (cid:3) N ´ERON MODELS OF ALGEBRAIC CURVES 13 The next proposition is well known. Proposition 4.3. Let k be a field, and let C be a projective geometri- cal ly integral curve over k of arithmetic genus ≥ 1. Let U ⊆ C be the smooth locus of C/k . Then the canonical morphism U → Pic 1 C/k , x 7→ OC (x) (given by the invertible sheaf OC×k U (D), where D is the graph of the inclusion U → C ) is a closed immersion. Proof. Let DivC/k be the scheme of effective Cartier divisors on C (see [1], §8.2). Let Div1 C/k be the subscheme corresponding to effective Cartier divisors of degree 1. Then the canonical morphism U → Div1 C/k , x 7→ x, is an isomorphism ([6], Exercise 9.3.8). Let f : Div1 C/k → Pic 1 C/k be the restriction of the canonical morphism DivC/k → Pic C/k (corresponding to E 7→ OC (E )). It will be enough to show that f is a closed immersion. It is known that f can be identified C/k for some coherent sheaf F on Pic 1 to P(F ) → Pic 1 C/k ([1], Proposition 8.2/7). In particular, f is proper and its fibers are pro jective spaces. On the other hand, the map U → Pic 1 C/k is injective because pa(C ) > 0. So the fiber of f at any y ∈ Im(f ) is a pro jective space of dimension 0 over k(y ), hence isomorphic to Spec(k(y )). This implies that f is a proper (hence closed) immersion. (cid:3) Proof of Theorem 4.1: As Xsm is a finite type scheme over S , it is enough to show Xsm is the N´eron lft-model of XK,sm . By Corollary 2.5, to show our theorem we can suppose S is local. Consider the closed immersion f : XK,sm → Pic 1 XK /K defined in Proposition 4.3. On the other hand, Pic1 XK /K is a torsor under JK := Pic0 XK /K , and JK has no subgroup isomorphic to Ga,K or Gm,K ([14], Proposition 1.1). Hence JK has a N´eron model over S ([1], Theorem 10.2/1) as well as Pic1 XK /K ([1], Corollary 6.5/4). By Proposition 4.2, XK,sm has a N´eron model N over S . Embed N into a proper model, and resolve the singularities without modifying the regular locus (which contains N ). Then we get a proper regular model N ′ containing N as an open subset. The identity on XK extends to a morphism N ′ → Xmin . By Corollary 3.2, this morphism induces a morphism N → Xsm which is an isomorphism on the generic fiber. Therefore Xsm satisfies the N´eron mapping property. (cid:3) Remark 4.4 Keep the notation of Theorem 4.1. (1) The N´eron model Xsm is not necessarily faithfully flat over S . if Xmin has a multiple fiber above some point s ∈ Indeed, N ´ERON MODELS OF ALGEBRAIC CURVES 14 S , then (Xsm )s = ∅. However, the faithful flatness holds if XK (K ) 6= ∅. (2) Assume XK is smooth over K and XK (K ) 6= ∅, and embed XK into its Jacobian JK by using some rational point of XK . The proof of Theorem 4.1 shows that the smooth locus of the minimal desingularization of the scheme-theoretic closure XK ⊂ J is isomorphic to Xsm , where J denotes the N´eron model of JK . Note that in general, even when Xmin is semi-stable, Xsm → J is not an immersion, see [3], Proposition 9.5. Remark 4.5 Suppose S = Spec(R) is local. Let R → R′ be an exten- sion of discrete valuation rings of index 1 (i.e., ramification index 1 and separable residue extension), let K ′ = Frac(R′ ). Then under the con- dition of Theorem 4.1, Xsm ×S Spec(R′ ) is the N´eron model of XK ′ ,sm over R′ . Indeed, the formation of Xsm commutes with completion ([8], 9.3.28). Applying Proposition 2.4 (3) to the extension bR → bR′ , we see that Xsm ×S Spec( bR′ ) is the bR′ -N´eron model, hence Xsm ×S Spec(R′ ) is the N´eron R′ -N´eron model by Proposition 2.4 (1). Remark 4.6 For any smooth connected curve UK over K , if it can be embedded as a closed subscheme into a semi-abelian K -variety (or more generally, into a smooth K -group scheme of finite type admitting a S - N´eron lft-model), Proposition 4.2 implies immediately that UK admits also an S -N´eron lft-model. For example, this works if UK is proper of positive genus, or if UK is affine such that the reduced divisor at infinity of UK in its regular compactification is separable of degree > 1 over K . But this method does not apply for al l curves: for instance the complement of a rational point in a proper smooth connected curve can not be embedded as a closed subscheme into any semi-abelian variety over K . In the second part of this work (§5–§7), we propose a different approach which works for any affine curve. Even better, our method allows a very explicit description of the N´eron lft-model with the help of minimal proper regular model of the regular compactification of the affine curve. Corollary 4.7. Let S be a Dedekind scheme with field of functions K , and let XK be a connected smooth proper curve over K of positive genus. Suppose that there exist a proper smooth variety YK having good reduction (i.e., having a proper smooth model Y /S ) and a dominant morphism fK : YK → XK over K . Then XK has good reduction. Proof. Let Xsm be the N´eron model of XK over S . Then fK extends to f : Y → Xsm . As Y is proper over S , the image f (Y ) is closed in N ´ERON MODELS OF ALGEBRAIC CURVES 15 Xsm and is dense because it contains XK . So f (Y ) = Xsm . Therefore Xsm is proper over S , and XK has good reduction. (cid:3) Remark 4.8 One can give a direct proof of Corollary 4.7 if the smooth S -scheme Y /S is assumed to be projective. Indeed, it is enough to show Xsm = Xmin . So one can suppose S is local. Using Bertini-type result, we may find a smooth closed subscheme of Y /S such that its generic fiber dominates again XK . Then we can repeat this argument to lower the relative dimension of the scheme Y /S until we find a smooth relative curve over S whose generic fiber still dominates XK . Finally we only need to apply [9], Corollary 4.10 to conclude. Remark 4.9 Corollary 4.7 does not hold in general if g (XK ) = 0. Let us consider the following example. Let R be a henselian discrete valu- ation ring with finite residue field k . Let k ′ /k be a quadratic extension and let T 2 + aT + b ∈ R[T ] be a lifting of the minimal polynomial of a generator of k ′ /k . Consider the scheme X = Pro j R[x, y , z ]/(x2 + axy + by 2 + πz 2 ) where π is a uniformizing element of R. Let R′ = R[T ]/(T 2 + aT + b). This is a finite ´etale extension of R. Let K = Frac(R), K ′ = Frac(R′ ). Then XK ′ ∼= P1 K ′ . So XK ′ has good reduction over R′ , hence over R when it is viewed as a K -scheme (a smooth pro jective model over R′ is also smooth pro jective over R). We have a surjective K -morphism XK ′ → XK . However, XK does not have good reduction over S . Indeed, suppose XK has a proper smooth model P → S . Then Pk is a smooth conic over a finite field, hence has a rational point. As R is henselian, then XK = PK has a rational point. The latter then specializes to the (unique) rational point of Xk . But this is a singular point of Xk . As X is regular, we have a contradiction by Lemma 3.1. Corollary 4.10. Let XK be a proper smooth connected curve of positive genus over K . Suppose that for some proper smooth connected variety YK over K , the product XK ×K YK has good reduction over S . Then XK has good reduction over S . Proof. Apply Corollary 4.7 to the pro jection XK ×K YK → XK . (cid:3) Let fK : YK → XK be a finite morphism of proper smooth and con- nected curves over K . Suppose g (XK ) ≥ 1. Let Ymin , Xmin be the respective minimal proper regular models of YK , XK over S . In gen- eral, fK does not extend to a morphism Ymin → Xmin ([9], Remark 4.5). However, Theorem 4.1 immediately implies the next corollary, answering positively a question raised by A. Pirutka. N ´ERON MODELS OF ALGEBRAIC CURVES 16 Corollary 4.11. Let fK : YK → XK be a finite morphism of proper smooth and connected curves over K , with g (XK ) ≥ 1. Let Ysm (resp. Xsm) be the smooth locus of the minimal proper regular model of YK (resp. XK ) over S . Then fK extends to a morphism f : Ysm → Xsm . For the sake of completeness, let us consider the N´eron lft-models of curves of genus 0. Proposition 4.12. Let XK be a smooth projective conic over K . (1) If XK = P1 K , then XK does not have N´eron lft-model over S . (2) In general, XK has a N´eron lft-model over S if and only if S is semi-local and if XK (K sh s ) = ∅ for any closed point s ∈ S , where K sh s denotes the fraction field of the strict henselization of OS,s . In this case, XK is its own N´eron model over S . Proof. (1) We will argue by contradiction: assume that P1 K admits a N´eron lft-model P over S . By the N´eron mapping property, there exists a morphism of S -schemes f : P1 S → P extending the canonical identification between the generic fibers. We claim that f is an iso- morphism. As P1 S is proper, and P /S is separated, the image f (P1 S ) is a closed subset of P . Its closed fiber has dimension 1 by Chevalley’s semi-continuity theorem ([5], IV.13.1.1). Therefore P1 k(s) → Ps is quasi- finite. By Zariski’s Main Theorem, f is an open immersion. But f is proper and P is irreducible, thus f : P1 S → P is an isomorphism. On the other hand, there are many endomorphisms of P1 K that can S , hence P1 not extend to an endomorphism of P1 S is not the N´eron lft- model of P1 K . Contradiction. (2) Assume first that XK admits a N´eron lft-model over S . By s admits a N´eron lft-model over Spec(Osh S,s). Proposition 2.4 (2), XK sh 6∼= P1 . In other words, XK (K sh s ) = ∅ for all closed As a result, XK sh K sh s s points s ∈ S . On the other hand, this condition implies that S is semi- local. Conversely, assume XK (K sh s ) = ∅ for any closed point s ∈ S . As S is semi-local, XK is of finite type over S and is its own S -N´eron model. (cid:3) Higher dimension. Let V be an algebraic variety over an alge- braically closed field k . We say that V contains a rational curve if V has a subscheme isomorphic to an open dense subscheme of P1 k . It is easy to see that if V does not contain any rational curve, then for any algebraically closed field extension K/k , VK does not contain any rational curve. Proposition 4.13. Let S be a Dedekind scheme with field of functions K . Let XK be a smooth proper algebraic variety over K . Suppose XK N ´ERON MODELS OF ALGEBRAIC CURVES 17 has a proper regular model X over S such that no geometric fiber X¯s , s ∈ S , contains a rational curve. Then the smooth locus Xsm of X is the N´eron model of XK . Proof. Let Y be an irreducible smooth scheme of finite type over S and let fK : YK → XK be a morphism of K -schemes. Consider the Y - scheme Z := X ×S Y → Y . Its geometric fibers are Z ¯y = Xs ⊗k(s) k(y ) and they do not contain any rational curve. The morphism fK induces a section YK → ZK , y 7→ (fK (y ), y ), which extends to a section Y → Z by [4], Proposition 6.2. Composing this section with the pro jection Z → X gives a morphism f : Y → X extending fK . By Proposition 3.2, f (Y ) ⊆ Xsm . This proves that Xsm is the N´eron model of XK over S . (cid:3) Remark 4.14 If AK is an abelian variety having good reduction over S , then we recover the well-known fact that a proper smooth model A of AK is the N´eron model of AK ([1], Proposition 1.2/8). 5. N´eron lft-models of open curves in the local case Let XK be a separated smooth connected curve having a N´eron lft- model X over S (e.g. if XK is proper of positive genus, see Theorem 4.1). Let UK be an open dense subscheme of XK . A natural question is whether UK has a N´eron lft-model U and, if it exists, how it is related to X . In this section, we restrict ourselves to the case where S = Spec(R) is local. Then we will show that the answer is positive under mild hypothesis and we describe explicitly the construction of U . Let Rsh denote the strict henselization of R, dRsh the completion of Rsh , K sh = Frac(Rsh ) and dK sh = Frac(dRsh ). 5.1. Main statement. Theorem 5.1. Let XK be a separated smooth connected curve over K and let UK be a dense open subscheme of XK . Suppose that XK has a smooth model X over S such that Xd Rsh is the N´eron lft-model of K sh . Then UK has a N´eron lft-model U over S . Moreover, denoting X d by ∆K = XK \ UK the boundary of UK in XK , U satisfies the fol lowing properties. (1) The scheme U is of finite type over S if and only if X is of finite type and if ∆K ∩ XK (dK sh ) = ∅. If S is excel lent, the latter condition is also equivalent to ∆K ∩ XK (K sh ) = ∅; N ´ERON MODELS OF ALGEBRAIC CURVES 18 (2) Let ∆ be the Zariski closure of ∆K in X . Then the identity on the generic fiber UK extends to an open immersion X \ ∆ → U , and the open immersion UK → XK extends to a morphism U → X . (3) Let s be the closed point of S and let k(s)sep be a separable closure of the residue field k(s) of s. Then U = X \ ∆ if and only if ∆ ∩ Xs(k(s)sep ) = ∅. Remark 5.2 In the statement of Theorem 5.1, the S -model X is nec- essarily the N´eron lft-model of XK over S (Proposition 2.4(1)), but the requirement in Theorem 5.1 is slightly stronger than this, except when S is excellent (Proposition 2.4 (3)). 5.2. Construction of U . Let k = k(s). First we construct a (possibly infinite) sequence of blow-ups of X , then define U as a suitable open subset of the resulting scheme. Note that the following construction can be done for any smooth S -scheme X such that UK is a dense open subset of XK , and Lemmas 5.3 and 5.4 hold just under these assumptions. Put X0 := X and ∆0 := ∆. 0 = (∆0)s∩sm(X0 )s(k sep ), i.e., the subset of points of (∆0)s∩ (i) Let ∆′ sm(X0 )s with separable residue field over k . If ∆′ 0 = ∅, we stop. (ii) Otherwise, let X1 → X0 be the blowing-up of X0 along ∆′ 0 (en- dowed with the reduced structure). Let ∆1 be the Zariski closure 1 = ∆1 ∩ sm(X1)s (k sep ). If ∆′ of ∆K in X1 and let ∆′ 1 = ∅, we stop. (iii) Otherwise blow up X2 → X1 along ∆′ 1 (reduced) and start again with the Zariski closure ∆2 of ∆K in X2 . We construct in this way a (possibly infinite) sequence of models locally of finite type Xn of XK , with ∆n ⊂ Xn the Zariski closure of ∆K in Xn , and n := ∆n ∩ sm(Xn )s (k sep ). Xn+1 → Xn is the blowing-up of ∆′ Let Un = sm(Xn ) \ ∆n . The identity map on UK extends to an open immersion Un → Un+1 . More precisely Un is Un+1 minus the the exceptional divisor of Xn+1 → Xn . Let U = Sn≥0 Un . This is a smooth, separated scheme locally of finite type over S , with generic fiber isomorphic to UK . We will show that U is the N´eron lft-model of UK over S . By construction, we have canonical morphisms (5.1) X \ ∆ ֒→ U → X, and the second morphism has image in (X \ ∆) ∪ (∆′ 0 )s . Note that the formation of U commutes with any flat extension of dis- crete valuation rings R → R′ because taking Zariski closure, blowing-up and taking the smooth locus are all compatible with such an extension. N ´ERON MODELS OF ALGEBRAIC CURVES 19 To prove U is the N´eron lft-model, we can replace R by dRsh and sup- pose R is strictly henselian and excellent (Proposition 2.4). After this reduction, X is the N´eron lft-model of XK in the situation of Theo- rem 5.1. But in general the curve UK might be no longer connected (the curve XK and its N´eron lft-model decompose accordingly). In this situation, if we can deal with each connected component of UK (which is an open subscheme of a connected component of XK ), the general case follows (Remark 2.3 (3)). Hence, we may assume that UK is still connected. 5.3. Dilatation. ([1], §3.2). Let R be a discrete valuation ring. Let X be any flat R-scheme of finite type. Let E be a closed subscheme of the special fiber Xs defined by a sheaf of ideals I ⊂ OX . Recall that the dilatation of E on X is obtained by blowing-up u : eX → X along E , and then taking the open subset of eX where u∗I is generated by a uniformizing element of R. Denote by X ′ the dilatation of E . It satisfies the following universal property: for any flat R-scheme Z , a morphism Z → X factors through X ′ → X if and only if Zs → Xs factors through E → Xs as morphism. If X/R is smooth and if the center E is smooth over k , by a local computation, one can show that the dilatation X ′ of E on X is smooth over R, and is equal to the complement in eX of the strict transform of Xs . 5.4. Some technical lemmas. Lemma 5.3. Keep the notation of §5.2 and assume R is strictly hense- lian and excel lent. Suppose that the sequence of blowing-ups Xn+1 → Xn is infinite. Let p1 , p2 ∈ XK be closed points such that for al l n ∈ N, they specialize to a same point xn ∈ ∆′ n . Then pi ∈ XK (K ) and p1 = p2 . Proof. Let us first prove pi ∈ XK (K ). Let p ∈ {p1 , p2}. Let X ′ n+1 ⊆ Xn+1 → Xn be the dilatation of ∆′ n on Xn and let Pn = {p, xn} be the reduced Zariski closure of p in Xn . By the construction of Xn , xn+1 n and it is a smooth point of Xn+1 . So xn+1 ∈ X ′ maps to xn ∈ ∆′ n+1 and for all n ≥ 0, we have a commutative diagram Pn+1 X ′ n+1 ⊂ Xn+1 Pn / Xn The first vertical arrow is birational and finite. Moreover, Pn+1 is nothing but the strict transform of Pn in Xn+1 . By the embedded / / / N ´ERON MODELS OF ALGEBRAIC CURVES 20 resolution of singularities (see e.g., [8], 9.2.26, recall that S is excellent), there exists m ≥ 1 such that Pm is a regular scheme. Then Pm+1 → Pm is an isomorphism and we have a factorization X ′ m+1 <②②②②②②②② σm / Xm . Pm By the universal property of the dilatation (§5.3), the closed immersion σm,s : Pm,s → Xm,s factors through Spec k(xm ) ⊂ Xm,s . Hence Pm,s → Spec k(xm ) = Spec k is an isomorphism. Therefore Pm → S is an isomorphism and we have K (p) = K . If p1 6= p2 , by the embedded resolution of singularities of the Zariski closure of {p1 , p2} in X0 , the Zariski closure becomes a disjoint union of sections in some Xn . Contradiction with the hypothesis in the lemma. (cid:3) Lemma 5.4. Keep the notation of §5.2 and suppose R is strictly hense- lian and excel lent. (a) If (Xn )n is an infinite sequence, and if (xn )n is such that xn ∈ ∆′ n and xn+1 7→ xn by Xn+1 → Xn for al l n ≥ 0, then there exists p ∈ ∆K such that xn ∈ Pn (Zariski closure of p in Xn ) for al l n ≥ 0. (b) Let Y be a flat integral S -scheme of finite type with irreducible closed fiber Ys . Let fK : YK → XK be a morphism. Suppose that for al l n ≥ 1, fK extends to fn : Y → Xn and that fn (Ys) ⊆ ∆′ n . Then fK (YK ) ∩ ∆K 6= ∅. (c) The scheme U is of finite type over S if and only if X is of finite type and ∆K ∩ XK (K ) = ∅. Proof. (a) Let Fn ⊆ ∆K be the set of points specializing to xn . Then (Fn )n is a decreasing sequence of non-empty finite sets, so ∩nFn 6= ∅ and any point p in the intersection satisfies the required property. (b) As Ys is irreducible, fn (Ys) is reduced to one point xn and xn+1 7→ xn by Xn+1 → Xn . Let p ∈ ∆K be such that xn ∈ Pn for all n ≥ 0. Let q ∈ YK be a lifting of some closed point of Ys . Then fK (q) and p are closed points of XK having the same specialization xn ∈ ∆′ n for all n. By Lemma 5.3, p = fK (q) ∈ ∆K ∩ fK (YK ). (c) Suppose there exists p ∈ ∆K ∩ XK (K ). For any n ≥ 0 such that Xn is constructed, p specializes to a point of ∆n ∩ sm(Xn )s(k), so Xn+1 exists in the construction of § 5.2. As Xn+1 → Xn consists in blowing-up some smooth points, Un+1 contains a dense open subset of the exceptional locus of Xn+1 → Xn . So Un ( Un+1 and U is not of / < N ´ERON MODELS OF ALGEBRAIC CURVES 21 finite type. On the other hand, if X is not of finite type, then U1 , thus U , is not of finite type. Conversely, if ∆K does not contain rational point of XK , there exists m ≥ 1 such that no point of ∆K specializes to a point of sm(Xm )s(k) (Lemma 5.3), and the construction of § 5.2 stops at this step. Thus U = Um is of finite type over S if X is of finite type. (cid:3) Lemma 5.5. Let S be a local ly noetherian scheme, and let f : Z → T be a morphism of local ly noetherian flat S -schemes. Let z0 ∈ Zs and t0 = f (z0 ). Suppose Zs → Ts is flat at z0 . Then the canonical morphism fz0 : Spec OZ,z0 → Spec OT ,t0 is surjective. In particular, for any P ∈ T such that t0 ∈ {P }, there exists Q ∈ f −1(P ) such that z0 ∈ {Q}. Proof. By the fiberwise flatness criterion ([5], IV.11.3.10.1), f is flat at z0 and fz0 is a flat morphism of local schemes, hence faithfully flat. In particular, fz0 is surjective. The last assertion results from the usual interpretation of the images of Spec OT ,t0 and Spec OZ,z0 in T and Z respectively. (cid:3) 5.5. Proof of Theorem 5.1. Now we prove that U is the N´eron lft- model of UK over S . As noticed previously in §5.2, we can suppose S is strictly local and excellent. In particular k sep = k . Let Y be a smooth scheme over S and let fK : YK → UK be a morphism of K -schemes. We want to extend fK to a morphism Y → U . We can suppose Ys is irreducible (Remark 2.3). First fK extends to a morphism f0 : Y → X by the hypothesis on X . Consider the sequence (Xn)n constructed in §5.2. (A) Suppose that for some m ≥ 0, f factors through fm : Y → Xm and that fm (Ys) 6⊆ ∆′ m . Let us show fm (Y ) ⊆ Um or, equivalently, that fm(Ys ) ∩ ∆m = ∅ because fm (Y ) ⊆ sm(Xm). If this is true, then fK extends to fm : Y → Um ⊆ U and we are done. We distinguish two cases: Case 1: fm(Ys ) = {xm} is a singleton. By hypothesis, xm /∈ ∆′ m . As Ys is smooth and Ys → Spec k factors through Ys → Spec k(xm ), we have xm ∈ sm(Xm )s(k). So xm /∈ ∆m and fm (Ys) ⊆ Um . Case 2: fm(Ys ) is not a singleton. If there exists xm ∈ fm(Ys ) ∩ ∆m , let ym ∈ f −1 m (xm ) and let p ∈ ∆K specializing to xm . As sm(Xm)s is a smooth curve and (fm )s is dominant, (fm )s is flat at ym . By Lemma 5.5 applied to Z = Y and T = sm(Xm), we find a closed point q ∈ YK such that fK (q) = p. This contradicts the hypothesis that fK (YK ) ⊆ UK . N ´ERON MODELS OF ALGEBRAIC CURVES 22 (B) Now we show that the condition in (A) is satisfied for some m ≥ 0. We start with f0 : Y → X0 . If f0(Ys ) ⊆ ∆′ 0 , then f0(Ys ) = {x0}. As in Corollary 3.8(1), f0 factors through f1 : Y → X1 . If f1(Ys ) 6⊆ ∆′ 1 , we are done. Otherwise, we have a f2 : Y → X2 . Repeating this construction, we see that if (A) is never satisfied, then for all n ≥ 0, fK extends to fn : Y → Xn with fn (Ys) ⊆ ∆′ n . This is impossible by Lemma 5.4(b). Hence U is the N´eron lft-model of UK . It remains to prove the various properties of U . We first remark that if S is excellent, then K sh is algebraically closed in its completion dK sh , hence the two conditions in Part (1) are indeed equivalent. Part (1) is a direct consequence of Lemma 5.4(c). Parts (2) and (3) follow from the construction in 5.2: X \ ∆0 = U0 is open in U and its generic fiber is XK \ ∆K = UK ; and U = X \ ∆0 is equivalent to ∆′ 0 = ∅. (cid:3) As an application of Theorem 5.1, we have the following result which will be used in Proposition 6.1 in the next section. Proposition 5.6. Let S be local. Let P be a regular, proper semi-stable model of P1 K over S . Let Γ1 , · · · , Γn be disjoint sections, n ≥ 2, in the smooth locus sm(P /S ) such that Γ := ∪iΓi meets every exceptional divisor of P /S when Ps is not irreducible. Let VK = P1 K \ ΓK . Let V be the S -model over S obtained by the process described in §5.2 starting with X := sm(P /S ). Then V is the S -N´eron lft-model of VK over S . Moreover, the identity VK → VK and the inclusion VK → PK extend to sm(P /S ) \ Γ → V → sm(P /S ) and the first morphism is an open immersion. Proof. As the formation of V commutes with completion and strict henselization, by Proposition 2.4 (1), we can suppose S is strictly local and excellent. We prove the result by induction on the number of irreducible components of Ps . First suppose Ps is irreducible. Start with the case n = 2. Then P ∼= P1 S . One can see easily that V , which is isomorphic to the N´eron lft-model of Gm,K , is obtained by the process described in §5.2 with X := P . See [1], 10.1. If Ps is irreducible and n ≥ 3, we consider UK = PK \ {(Γ1)K , (Γ2)K } with its N´eron lft-model U . Then V can be obtained by the process of §5.2 starting with X := U . As P \ (Γ1 ∪ Γ2) is open in U by the above discussions, and the Zariski closure of (Γi )K in P is Γi ⊂ P \ (Γ1 ∪ Γ2) when i ≥ 3, Theorem 5.1 says that V is the S -N´eron lft-model of VK , P \ Γ is open in V and the latter maps to U , hence to P . Now suppose Ps has more than one component. Let E be an ex- ceptional divisor in P . Up to renumbering, we can suppose Γ1 , . . . , Γr , N ´ERON MODELS OF ALGEBRAIC CURVES 23 r ≤ n − 1, are exactly the sections of P among the Γi ’s not meeting E . Let π : P → Q be the contraction of E et let q = π(E ) ∈ Qs . Consider UK = P1 K \ {(Γ1)K , . . . , (Γr )K , (Γr+1)K }. Then Q is regular, proper and semi-stable and, if we still denote by Γi the Zariski closure of (Γi )K in Q, Γ1 , . . . , Γr , Γr+1 correspond to r+1 disjoint sections of Q/S whose union meets every exceptional divisor of Q/S when Qs is not irreducible. By the induction hypothesis, the N´eron lft-model U of UK is obtained by the process of §5.2 starting from X := sm(Q/S ). As (Γr+1)s = {q} and π : P → Q is the blow-up of Q along {q}, by the explicit construction of U , sm(P ) \ (∪i≤r+1Γi ) is open in U . For any i ≥ r + 2, the point of (Γi )K specializes to a point of (sm(P ) \ (∪i≤r+1Γi ))s . Then Theorem 5.1 tells us that VK admits an S -N´eron lft-model V , and the latter is ob- tained from U by blowing-up the closed points ∪i≥r+2 (Γi )s contained in the open subset sm(P ) \ (∪i≤r+1Γi ) ⊂ U , taking the smooth locus and start again etc. In particular sm(P /S ) \ Γ is an open subscheme of V and the latter maps to sm(P /S ). (cid:3) 6. N´eron models of open subsets of a smooth conic In this section, we suppose S = Spec(R) is local and excellent. We prove the existence of the N´eron model for affine open subsets of a smooth pro jective conic, whose complement is non-empty and consists of ramified points. Proposition 6.1. Let S be an excel lent local Dedekind scheme with field of functions K . Let CK be a projective smooth conic over K . Let ∆K be a non-empty finite closed subset of CK (endowed with the reduced structure) such that ∆K (K sh ) = ∅. Then UK := CK \ ∆K admits a N´eron model over S . ∼= P1 Proof. Assume first CK (K ) 6= ∅, or equivalently, CK K . Consider a smooth proper model isomorphic to P1 S of P1 K . After finitely blowing- ups along smooth separable points of the special fiber of the latter, the construction of §5.2 gives us a regular proper semi-stable model P of ∼= P1 K such that the intersection ∆ ∩ sm(P /S )s(k(s)sep ) is empty. CK By successively blowing-down exceptional divisors of P which do not meet the Zariski closure ∆ of ∆K , we can suppose that: (1) ∆ meets every exceptional divisor of P /S if Ps is not irreducible and (2) ∆ meets Ps only at singular points or smooth inseparable points. Claim: under the above conditions, U := sm(P /S ) \ ∆ is the N´eron model of UK . N ´ERON MODELS OF ALGEBRAIC CURVES 24 To prove the claim we can suppose S is strictly local. In particular, each (reduced) irreducible component of Ps is isomorphic to P1 k . Let Y be a smooth S -scheme with connected fibers and let fK : YK → UK be a morphism of K -schemes. Let Γ1 , . . . , Γn ⊂ U (S ) be disjoint sections such that n ≥ 2, Γ := ∪iΓi is ample in P and such that fK (YK ) 6⊂ ΓK . Set H = f −1 K (ΓK ) ⊂ Y . This is a closed subset of Y , empty or of codimension 1. Let Y ′ := Y \ H . We claim that the restriction : Y ′ K → UK \ ΓK extends to a morphism Y ′ → U . fK Y ′ Indeed, K let V be the N´eron lft-model of VK := PK \ ΓK . Then fK induces a morphism f ′ : Y ′ → V . By Proposition 5.6, we have a morphism V → sm(P /S ) \ Γ = U . Consequently, f ′ extends to a morphism f ′′ : Y ′′ → U where Y ′′ = Y ′ ∪ YK . As Ys is connected, the image f ′′ s : Y ′′ s → Us is contained in some connected component. Let U ′ denote the union of the latter connected component of Us with UK . Then f ′′ factors through U ′ ⊂ U . On the other hand, the scheme U ′ can be obtained by first blowing- down successively all the irreducible components of Ps other than the one containing U ′ s , then removing from the resulting proper smooth ∼= P1 K the closure of ∆K . In particular, the scheme U ′ is model of CK affine. Thus f ′′ extends to a morphism Y → U ′ ⊂ U since Y \ Y ′′ is a closed subset of codimension ≥ 2 of the normal scheme Y ([1], 4.4/2). Therefore U satisfies N´eron mapping property, as desired. For the general case, if XK (K sh ) = ∅, then XK is the S -N´eron model of itself. Otherwise, there exists a finite unramified extension K ′ of K such that CK (K ′ ) 6= ∅. Since K ′/K is unramified, the complement in its smooth compactification CK ′ consists of closed points of UK ′ which are still ramified of degree > 1 over K ′ . Therefore, UK ′ admits a N´eron model U ′ over the semi-local ring S ′ , the normalization of S in K ′ . Hence UK admits also a S -N´eron model by Proposition 7.4 and Proposition 7.5. (cid:3) Remark 6.2 Keep the notation of Proposition 6.1, and let U be the S - ∼= P1 K with a proper smooth S -model N´eron model of UK . Assume CK C such that the Zariski closure ∆ of ∆K (with the reduced structure) is regular. Then the canonical morphism obtained from N´eron mapping property C \ ∆ → U is an open immersion. Indeed, as ∆ is regular, after blowing up all separable closed points of ∆, we obtain a proper semi-stable model P of CK such that the conditions (1) (2) in the proof of Proposition 6.1 are satisfied. By the claim in the same proof, U = sm(P /S ) \ ∆K . Thereby C \ ∆ is an open subset of U . Remark 6.3 In Proposition 6.1, we can not drop the assumption that S is excellent. For example, let k be a field of characteristic 2, and N ´ERON MODELS OF ALGEBRAIC CURVES 25 K := k(t, u2) ⊂ k [[t]] ⊂ k((t)) with u ∈ k [[t]] transcendental over k(t). The discrete valuation on k((t)) induces a discrete valuation on K , and let R be the corresponding (discrete) valuation ring (with t ∈ R a uniformizer). The completion bR of R is k [[t]] ⊂ k((t)), hence bK = Frac( bR) = k((t)). Note that u ∈ bK is purely inseparable of degree 2 over K (in particular, R is not excellent). Consider v = u2 ∈ K and let UK = Spec(K [X1 , X2 ]/(X 2 1 − vX 2 0 − X0). This is the underlying scheme of a unipotent group. As v /∈ K 2 , UK 6∼= A1 K . On the other ∼= Ga, bK . Therefore, by [1], 10.2/2, ∼= A1 . It follows that U bK hand, U bK bK UK does not admit N´eron lft-model over S . 7. N´eron lft-models of affine curves The aim of this section is to prove the existence of N´eron lft-models of affine curves over K different from the affine line (Theorem 7.10). Let S be a Dedekind scheme with field of functions K . 7.1. Globalizing local N´eron lft-models. Lemma 7.1. Let S be a Dedekind scheme with K its field of functions. Let UK be a separated connected smooth curve over K . Suppose that (i) for any closed point s ∈ S , UK admits a N´eron lft-model U (s) over OS,s ; (ii) there exists a model of finite type U 0 of UK over S such that for al l s ∈ S , the isomorphism U 0 K → U (s)K extends to an open immersion U 0 ×S Spec(OS,s) → U (s). Then UK admits a N´eron lft-model over S . Proof. The proof is inspired from that of [1], 10.1/7. The N´eron lft- model of UK over S will be obtained by gluing the local N´eron lft- models U (s) for all closed points s ∈ S . We will first extend U (s) to a model V (s) over S which coincides with U 0 above S \ {s} and then glue the various V (s) in a natural way to obtain the N´eron lft-model U of UK over S . Fix a closed point s. As U (s) is locally of finite type, any connected component U (s)s,α of U (s)s is open in the closed fiber U (s)s , hence its union with UK is a quasi-compact open subset Uα (s) of U (s). We can extend Uα(s) to a separated scheme of finite type Uα over S (use [5], IV.8.10.5). As Uα and U 0 are both of finite type over S and have the same generic fiber, they are S -isomorphic over a dense open subset Sα ⊆ S \ {s} (and the S -isomorphism is unique once the isomorphisms U 0 K → UK , (Uα)K → UK are fixed because Uα is separated over S ). N ´ERON MODELS OF ALGEBRAIC CURVES 26 Now we glue the separated morphisms of finite type Uα ×S (Sα ∪ {s}) → Sα ∪ {s}, U 0 ×S (S \ {s}) → S \ {s} above (Sα ∪ {s}) ∩ (S \ {s}) = Sα . The resulting S -scheme Vα is separated and of finite type because these properties are satisfied above Sα ∪ {s} and S \ {s}. By construction, we have canonically Vα ×S (S \ {s}) = U 0 ×S (S \ {s}), Vα ×S Spec(OS,s) = Uα (s). Next we glue the various Vα (when Uα (s) runs through the connected components of U (s)s ) with the condition Vα ∩ Vα′ = U 0 if α 6= α′ . The resulting S -scheme V (s) satisfies canonically V (s) ×S (S \ {s}) = U 0 ×S (S \ {s}), V (s) ×S Spec(OS,s ) = U (s). Hence V (s) is separated and locally of finite type over S . Moreover, Condition (ii) implies that the isomorphism U 0 K → V (s)K = U (s)K extends to an open immersion U 0 → V (s). Finally, we glue the various V (s) when s runs through the closed points of S with the condition V (s)∩V (s′ ) = U 0 if s 6= s′ . The resulting S -scheme U is locally of finite type and U ×S Spec(OS,s ) ∼= U (s) for all s ∈ S . By Corollary 2.5, U is the N´eron lft-model of UK overS , as required. (cid:3) Lemma 7.2. Let S be a Dedekind scheme. Let XK be a smooth con- nected separated curve over K and let UK be an open dense subscheme of XK . Suppose that XK has a smooth model X over S such that for al l closed points s ∈ S , X ×S Spec(OS,s ) satisfies the property in Theorem 5.1. Then UK admits a N´eron lft-model over S . Proof. Let ∆K := XK \ UK , X 0 ⊂ X any quasi-compact open subset containing XK , and ∆0 := ∆K ⊂ X 0 . By Theorem 5.1, the N´eron model of UK over Spec(OS,s) exists for all closed points s ∈ S , and by Theorem 5.1 (2), U 0 := X 0 \ ∆0 verifies the hypothesis of Lemma 7.1. Thus we can apply Lemma 7.1 to conclude. (cid:3) Proposition 7.3. Let S be a Dedekind scheme. Let XK be a connected regular proper curve over K of arithmetic genus ≥ 1. Let UK be a dense open subscheme of XK contained in the smooth locus of XK /K . Suppose either S is excel lent or the scheme XK is smooth over K . Then UK admits a N´eron lft-model over S . Proof. Let X = Xsm be the smooth locus of the minimal proper regular model of XK over S . Let s ∈ S be a closed point. Then X ×S dOsh S,s is the N´eron model of XK over dOsh S,s by Theorem 4.1 and because in both N ´ERON MODELS OF ALGEBRAIC CURVES 27 cases the minimal proper regular as well as the its smooth locus com- mute with strict henselization and completion ([8], 9.3.28). Applying Lemma 7.2 to UK ⊆ XK,sm , we see that UK has a N´eron lft-model U over S . (cid:3) 7.2. Weil restriction. Proposition 7.4 (see also [1], 10.1/4). Let S be a Dedekind scheme with field of functions K and let XK be a separated smooth connected curve over K . Let K ′/K be a finite extension, and let S ′ be the nor- malization of S in K ′ . Assume that (i) S ′ → S is finite (e.g., if S is excel lent or K ′ /K is separable); (ii) XK ′ admits a N´eron lft-model (resp. N´eron model) X ′ over S ′ ; (iii) any quasi-compact open subset of X ′ is quasi-projective over S ′ . Then XK admits also a N´eron lft-model (resp. N´eron model) over S . Proof. Let s ∈ S . Then any finite subset F of X ′ s := X ′ ×S Spec k(s) is contained in an affine open subset of X ′ . Indeed, F is contained in a quasi-compact open subset W of X ′ . Let V be an affine open neighborhood of s. As S ′ ×S V is finite over V , W ×S V = W ×S ′ (S ′ ×S V ) is quasi-pro jective over V , hence F is contained in an affine open subset of W ×S V ⊆ X ′ . The morphism S ′ → S is finite and locally free, hence the above prop- erty implies that the Weil restriction functor Y := ResS ′ /S X ′ is repre- sentable by a smooth S -scheme locally of finite type ([1], 7.6/4). Fur- thermore, by the functoriality of the Weil restriction, one checks easily ∼= ResK ′/K (XK ′ ). that Y is the S -N´eron lft-model of its generic fiber YK Finally, remark that as XK /K is separated, the adjunction map XK → ResK ′ /K (XK ′ ) = YK is a closed immersion, hence it suffices to apply Proposition 4.2 to conclude the existence of N´eron lft-model or N´eron model of XK . (cid:3) The following result is useful when we want to check the condi- tion (iii) of Proposition 7.4. Proposition 7.5. (Quasi-pro jectivity) Let S be a Dedekind scheme with field of functions K , and let U be a connected regular relative curve over S local ly of finite type. Suppose that either UK has a smooth compactification or S is excel lent. Then any quasi-compact open subset of U is quasi-projective over S (in the sense of [5], II.5.3.1). Proof. Let U0 be a quasi-compact open subset of U . Let U0 ⊆ U ′ 0 be a Nagata compactification. The hypothesis on UK or S implies that there exists a desingularization morphism Z → U ′ 0 which is an isomorphism N ´ERON MODELS OF ALGEBRAIC CURVES 28 above U0 . So U0 is isomorphic to an open subscheme of a regular proper flat S -scheme Z . It is enough to show that Z → S is pro jective. This is a theorem of Lichtenbaum when S is affine ([7], Theorem 2.8 or [8], 8.3.16), but the proof works exactly in the same way in the general case: find a positive horizontal Weil divisor H on Z which meets all irreducible components of all fibers of Z → S . As Z is regular, H is defined by an invertible sheaf L on Z . The hypothesis on H implies that L is fiberwise ample, hence L is relatively ample for Z → S . (cid:3) Remark 7.6 If the Dedekind scheme S is separated, then any quasi- pro jective scheme over S is a subscheme of some PN S . Indeed, S then has an invertible ample sheaf ([15, Proposition 09NZ]), and one can conclude with [5], II.5.3.3. 7.3. Affine open subsets of a conic. In this Subsection, we discuss the existence of N´eron lft-models of an affine open subscheme UK of a smooth pro jective conic CK /K . Observe first that A1 K does not admit N´eron lft-model over S ([1], 10.1/8). Proposition 7.7. Let S be an excel lent Dedekind scheme with field of functions K . Let UK be an affine open subscheme of a smooth projective conic CK over K . Suppose UK is not isomorphic to A1 K . Then UK admits a N´eron lft-model U over S . Proof. If over an algebraic closure K of K , CK \ UK contains at least two points, then there exists a finite extension K ′/K such that UK ′ is isomorphic to an open subscheme of Gm,K ′ . It follows from Proposi- tion 7.4 that we can suppose UK is an open subscheme of Gm,K . The latter has a N´eron lft-model G over S , locally on S compatible with any index 1 extension (see the construction of [1], 10.1/5). Consequently, by Lemma 7.2, UK admits a N´eron lft-model over S . For the rest of the proof, we can therefore suppose that UK is CK minus one point. So ∆K := CK \ UK consists of a single point q∞ which is purely inseparable of degree > 1 over K because UK 6∼= A1 K . As CK is smooth over K , there exists a separable extension K ′/K such that CK ′ ∼= P1 K ′ . The point of CK ′ \UK ′ is still purely inseparable of degree > 1 over K ′ because K ′ /K is separable. Using Proposition 7.4, K . Let P ∼= P1 we can reduce to the case CK = P1 S be a smooth proper model of P1 K over S , and let ∆ = {q∞} ⊂ P . We know (Proposition 6.1) that for all closed points s ∈ S , UK admits a N´eron model U (s) over OS,s . To find a global N´eron lft-model, it is enough to show that for U 0 := P \ ∆, the canonical morphism U 0 ×S Spec(OS,s ) → U (s) (7.1) N ´ERON MODELS OF ALGEBRAIC CURVES 29 is an open immersion for all s contained in a dense open subset V ⊂ S : the base change U 0 ×S V satisfies then Condition (ii) of Lemma 7.1. As S is excellent, so is ∆. Thus the regular locus of ∆ is open in ∆. Shrinking S if necessary, we can assume ∆ is regular. Then for any closed point s ∈ S , the morphism (7.1) is an open immersion by Proposition 6.1 and Remark 6.2, and the proposition is proved. (cid:3) Corollary 7.8. Let S be an excel lent Dedekind scheme of characteristic p > 0, with K its field of functions. Let GK be a connected smooth K -wound unipotent group of dimension 1. Then GK admits a N´eron lft-model over S . Remark 7.9 Let S be a Dedekind scheme with field of functions K . One can deduce from Proposition 4.12 and Proposition 4.2 that if XK is a connected separated smooth K -variety admitting N´eron lft-model over S , then XK does not contain any closed subscheme isomorphic to P1 K or A1 K . Conversely, if XK is the underlying scheme of a smooth commutative algebraic group over K , the latter condition is also suffi- cient for the existence of N´eron lft-model when S is local and excellent ([1], 10.2/2). When S is global and excellent, whether this latter condition is suffi- cient is still an open question. It is conjectured ([1], 10.3, Conjecture I) that the answer is yes. Some positive examples are known in [1], Chap. 10. Corollary 7.8 provides some evidence in favor of this conjecture. Together with the well-known results for abelian varieties and for tori, we deduce that when S is excellent, any smooth connected K -algebraic group GK of dimension one admits a N´eron lft-model over S if and only if GK is not isomorphic to Ga,K . In other words, Conjecture I of [1], 10.3 holds when dim(GK ) = 1. When the unipotent group scheme GK in Corollary 7.8 admits a regular compactification of genus ≥ 1, or equivalently when uni(GK ) = 0, Corollary 7.8 is a special case of [1], 10.3/5. So the new case provided here is when uni(GK ) > 0, or equiv- alently when GK admits a smooth compactification of genus 0. The latter happens only when char(K ) = 2 (see last paragraph of [1], 10.3, p. 316). In this case, GK is the subgroup of G2 a,K = Spec(K [X, Y ]) defined by the equation X 2 = Y + aY 2 for some a ∈ K \ K 2 , which, as a scheme, is isomorphic to Pro j(K [T , T ′ ]) \ V+ (T 2 − aT ′2 ). 7.4. N´eron lft-models for affine curves. We are now in the position to prove the existence of N´eron lft-models for affine curves. Theorem 7.10. Let S be an excel lent Dedekind scheme with field of functions K . Let UK be an affine smooth connected curve over K . N ´ERON MODELS OF ALGEBRAIC CURVES 30 Then UK admits a N´eron lft-model over S if UK 6∼= A1 L for any finite extension L/K . Proof. By Lemma 2.6, we can suppose UK is geometrically connected. If the regular compactification of U has positive arithmetic genus, then UK has a N´eron lft-model by Proposition 7.3. Otherwise, UK is an affine open subset of a smooth pro jective conic over K , not isomorphic to A1 K . So UK admits a N´eron lft-model over S by Proposition 7.7. (cid:3) Next we examine when the N´eron lft-model is of finite type. Proposition 7.11. Let S be an excel lent Dedekind scheme with field of functions K . Let UK be an affine smooth geometrical ly connected curve of K , not isomorphic to A1 K , and let CK be its regular compactification. Denote by ∆K := CK \ UK . Let C be a relatively minimal regular model of CK over S and let ∆ be the reduced Zariski closure of ∆K in C . Let U be the N´eron lft-model of UK over S . Then the fol lowing properties are true. (1) The scheme U/S is of finite type if and only if ∆K (K sh s ) = ∅ for al l closed points s ∈ S and if ∆s ∩ Csm,s (k(s)sep ) = ∅ for almost al l s ∈ S . (2) Assume S is infinite. For each closed point s ∈ S , set U (s) := U ×S Spec(OS,s), the local N´eron lft-model of UK over Spec(OS,s ). Let K sep denote a separable closure of K . (i) ∆K (K sep ) = ∅ if and only if al l the local N´eron lft-models U (s) are of finite type. (ii) If ∆K (K sep ) 6= ∅, the local N´eron lft-models U (s) are not of finite type for al l but finitely many closed points s ∈ S . In particular, if K has characteristic 0, then U is never of finite type. Proof. (1) The S -scheme U is of finite type if and only if (a) for all closed point s ∈ S , U (s) is of finite type over OS,s ; and if (b) U is of finite type over some open dense subset of S , or equiv- alently, Us is connected for all but finitely many s (Proposi- tion 2.8). Therefore we only need to show that the conditions of (1) are equivalent to the conditions (a) and (b) above. First assume that CK is of arithmetic genus > 0. Then UK ⊆ XK := sm(CK /K ). Let X be the N´eron model of XK over S , equal to the smooth locus Csm of C/S (Theorem 4.1). The minimal regular model C commutes with strict henselization and completion ([8], 9.3.28), so by Theorem 5.1, for any closed point s ∈ S , U (s) is of finite type over N ´ERON MODELS OF ALGEBRAIC CURVES 31 OS,s if and only if ∆K (K sh s ) = ∅. Now consider the connectedness at a closed point s ∈ S . Shrinking S if necessary, we can suppose Xs= Csm,s is connected. If ∆s ∩ Csm,s(k(s)sep ) = ∅, then Us = Xs \ ∆s (Proposition 5.1 (3)) is connected. Conversely, suppose Us is connected. By separatedness, Us → Xs is an open immersion. Then over OS,s , U → X is an open immersion, as X \ U contains ∆K , hence U ⊆ X \ ∆. By (5.1), we have U = X \ ∆ (over OS,s ), thus ∆s ∩ Csm,s(k(s)sep ) = ∅ (Theorem 5.1 (3)). Now suppose CK is a smooth conic. Let s ∈ S be a closed point. If ∆K (K sh s ) = ∅, then U (s) is of finite type over OS,s by Proposition 6.1. Conversely, suppose ∆K (K sh s ) 6= ∅. It is enough to show U (s) is not of finite type over Osh S,s . Thus we can suppose K strictly henselian and ∆K (K ) 6= ∅. If there are two rational points in ∆K , then UK is isomorphic to an open subscheme of Gm,K , hence U (s) is not of finite type by Theorem 5.1 (1). Otherwise, there exists a non-rational point q∞ ∈ ∆K . We apply again Theorem 5.1 (1) to UK ⊂ CK \ {q∞} to conclude that U (s) is not of finite type. So again U (s) is of finite type if and only if ∆K (K sh s ) = ∅. To see the equivalence between Condition (b) above and the second condition of (1), we are allowed to shrink S and suppose that C/S is smooth with connected fibers. The condition ∆s ∩ Csm,s(k(s)sep ) = ∅ then implies that C \ ∆ is the N´eron model of UK over OS,s (see Claim in the proof of Proposition 6.1). Thus U = C \ ∆ is of finite type with connected fibers over S . Conversely, suppose U/S is of finite type (up to replace S by some open dense subset). Shrinking S if necessary, the isomorphism UK → (C \ ∆)K extends to an isomorphism U ∼= C \ ∆. If there exists x ∈ ∆s ∩ Csm,s (k(s)sep ), then there exists y ∈ UK (K sh s ) = (C \ ∆)K (K sh s ) which does not specializes to Us . Contradiction with the N´eron mapping property of U . (2) As S is infinite, ∆K (K sep ) 6= ∅ if and only if ∆K (K sh s ) 6= ∅ for some (hence for almost all) s ∈ S . Thereby (2) follows directly from (1). (cid:3) Remark 7.12 In general it is not true that if ∆K (K sep ) = ∅ then U is of finite type. One can construct examples similar to that of Oesterl´e ([1], 10.1/11), showing that the existence of local N´eron models does not imply the existence of global N´eron model. Let S be an excellent Dedekind scheme of characteristic p > 0 with infinitely many closed points, such that each closed point of S has perfect residue field (for example, take S a smooth algebraic curve over a perfect field of characteristic p). Let q∞ ∈ P1 K be a purely inseparable closed point of degree > 1. Let UK = P1 K \ {q∞} and let U be its S -N´eron model. N ´ERON MODELS OF ALGEBRAIC CURVES 32 Then U ×S Spec(OS,s ) is of finite type. But as Us has two connected components for almost all s ∈ S , U is not of finite type over S by Proposition 2.8 (3). References [1] S. Bosch, W. Lutkebohmert and M. Raynaud, N´eron models, Ergebnisse der Math., 3. Folge, Bd. 21, 1990. [2] T. Chinburg, Minimal models for curves over Dedekind rings, in Arithmetic Geometry edited by G. Cornell and J. H. Silverman. Springer-Verlag, 1986. [3] B. Edixhoven, On N´eron models, divisors and modular curves, J. Ramanujan Math. Soc. 13 (1998), 157–194. [4] O. Gabber, Q. Liu and D. Lorenzini, Hypersurfaces in projective schemes and a moving lemma, preprint (2011). [5] A. Grothendieck and J. Dieudonn´e, ´El´ements de g´eom´etrie alg´ebrique, Publ. Math. IH ´ES 4 (Chapter 0, 1–7, and I, 1–10), 8 (II, 1–8), 11 (Chapter 0, 8–13, and III, 1–5), 17 (III, 6–7), 20 (Chapter 0, 14–23, and IV, 1), 24 (IV, 2–7), 28 (IV, 8–15), and 32 (IV, 16–21), 1960–1967. [6] S. Kleiman, The Picard scheme, Fundamental algebraic geometry, 235–321, Math. Surveys Monogr., 123, Amer. Math. Soc., Providence, RI, 2005. [7] S. Lichtenbaum, Curves over discrete valuation rings, Amer. J. Math. 90 (1968), 380–405. [8] Q. Liu, Algebraic geometry and arithmetic curves, Oxford Graduate Texts in Mathematics 6, Oxford University Press, paperback new edition (2006). [9] Q. Liu and D. Lorenzini, Models of curves and finite covers, Compositio Math. 118 (1999), no. 1, 61–102. [10] H. Matsumura, Commutative algebra (second edition), The Benjamin /Cum- mings Company, Inc., 1980. [11] A. N´eron, Mod`eles minimaux des vari´et´es ab´eliennes sur les corps locaux et globaux, Publ. Math. IHES, 21 (1964), 128 pp. [12] D. Popescu, General N´eron desingularization and approximation, Nagoya Math. J., 104 (1986), 85-115. [13] M. Raynaud: p-groupes et r´eduction semi-stable des courbes, The Grothendieck Festschrift, Vol. III, 179–197, Progr. Math., 88, Birkhauser Boston, Boston, MA, 1990. [14] M. Raynaud: Groupes vectoriels et sch´ema de Picard, C. R. Acad. Sci. Paris, Ser. I, 338 (2004), 223–227. [15] Stacks Pro ject Authors: Stacks Project, stacks.math.columbia.edu, (2013). Universit´e de Bordeaux 1, Institut de Math´ematiques de Bordeaux, CNRS UMR 5251, 33405 Talence, France E-mail address : [email protected] E-mail address : [email protected]
1803.09008	1	1803	2018-03-23T23:11:46	The Jordan property of Cremona groups and essential dimension	[ "math.AG", "math.GR" ]	We use a recent advance in birational geometry to prove new lower bounds on the essential dimension of some finite groups.	math.AG	math	THE JORDAN PROPERTY OF CREMONA GROUPS AND ESSENTIAL DIMENSION ZINOVY REICHSTEIN Abstract. We use a recent advance in birational geometry to prove new lower bounds on the essential dimension of some finite groups. 1. Introduction A abstract group Γ is called Jordan if there exists an integer j such that every finite subgroup G ⊂ Γ has a normal abelian subgroup A of index [G : A] 6 j. This definition, due to V. L. Popov [Po11, Po14], was motivated by the classical theorem of Camille Jordan [J1878] which asserts that GLn(k) is Jordan, and by a theorem of J.-P. Serre [Se10] which asserts that the Cremona group Cr2(k) is also Jordan. Here and throughout this note k denotes a base field of characteristic 0. The Cremona group Cr2 = Bir(P2) is the group of birational automorphisms of the projective plane. Serre asked whether the higher Cremona groups Crn = Bir(Pn) are Jordan as well. Our starting point is the following remarkable theorem of Y. Prokhorov, C. Shramov and C. Birkar, which asserts that groups of birational isomorphisms of rationally connected varieties of fixed dimension are "uniformly Jordan". Theorem 1. ([Bi16, Corollary 1.3]) For every positive integer n > 1 there exists a positive integer j(n) with the following property. Let G be a finite subgroup of the group Bir(X) of birational automorphisms of an n-dimensional rationally connected variety X. Then G has a normal abelian subgroup A such that [G : A] 6 j(n). Prokhorov and Shramov [PS16] proved this theorem assuming the Borisov-Alexeev- Borisov (BAB) conjecture. The BAB conjecture was subsequently proved by Birkar [Bi16]. In this note we will deduce some consequences of Theorem 1 concerning essential dimen- sion of finite groups. Let G be a finite group. Recall that the representation dimension rdimk(G) is the minimal dimension of a faithful representation of G defined over k, i.e., the smallest positive integer r such that G is isomorphic to a subgroup of GLr(k). The essential dimension edk(G) is the minimal dimension of a faithful linearizable G-variety defined over k. Here by a faithful G-variety we mean an algebraic variety X with a faithful action of G. We say that X is linearizable if there exists a G-equivariant dominant rational map V 99K X, where V is a vector space with a linear action for G. 2010 Mathematics Subject Classification. 14E07, 20C05. Key words and phrases. Jordan property, Cremona group, essential dimension, representation dimension. Partially supported by National Sciences and Engineering Research Council of Canada Discovery grant 253424-2017. 1 2 ZINOVY REICHSTEIN It is clear from these definitions that (1) edk(G) 6 rdimk(G). We will write ed(G) and rdim(G) in place of edk(G) and rdimk(G), respectively, when the reference of k is clear from the context. Equality in (1) holds in two interesting cases: • if G is abelian and k contains a primitive eth root of unity, where e is the exponent of G (see [BR97, Theorem 6.1]), or • if G is a p-group and k contains a primitive pth root of unity (see [KM08]). For other finite groups, ed(G) and rdim(G) can diverge. Our first main result shows that they do not diverge too far, assuming k contains suitable roots of unity. Theorem 2. Let r(n) = nj(n), where j(n) is the Jordan constant from Theorem 1. Suppose G is a finite group of exponent e and the base field k contains a primitive eth root of unity. (a) If edk(G) 6 n, then rdimk(G) 6 r(n). (b) Moreover, if edk(G) 6 n, then G is isomorphic to a finite subgroup of Gr(n) m ⋊ Sr(n), where the symmetric group Sr(n) acts on Gr(n) m by permuting the factors. To place Theorem 2(a) into the context of what is currently known about essential dimension of finite groups, let us assume for simplicity that k is algebraically closed. Let G be a finite group, p be a prime, and G[p] be a Sylow p-subgroup of G. As we mentioned above, ed(G[p]) = rdim(G[p]) by the Karpenko-Merkurjev theorem [KM08], and rdim(G[p]) can be computed, at least in principle, by the methods of representation theory of finite groups. This way we obtain a lower bound (2) ed(G) > max p rdim(G[p]). One can then try to prove a matching upper bound by constructing an explicit d- dimensional faithful linearizable G-variety of dimension maxp rdim(G[p]). In most of the cases where the exact value of ed(G) is known, it was established using this strategy. There are, however, finite groups G for which the inequality (2) is strict. All known proofs of stronger lower bounds of the form ed(G) > d appeal to the classification of finite subgroups of Bir(X), where X ranges over the d-dimensional unirational (or rationally connected) varieties. Such classifications is available only for d = 1 (see [Kl1884, Chapter 1]) and d = 2, and the latter is rather complicated; see [DI09]. For d = 3 there is only a partial classification (see [Pr12]), and for d > 4 even a partial classification is currently out of reach. Lower bounds of the form ed(G) > d proved by this method (for suitable finite groups G), can be found • in [BR97, Theorem 6.2] for d = 1, • in [Se10, Proposition 3.6], [Dun13] for d = 2, and • in [Dun10], [Be14], [Pr17] for d = 3. For an overview, see [Rei10, Section 6]. This paper is in a similar spirit, with Theorem 1 used in place of the above-mentioned classifications. As a consequence of Theorem 2(a), we will obtain the following. JORDAN PROPERTY AND ESSENTIAL DIMENSION 3 Theorem 3. Let Z/nZ be a cyclic group of order n and Hn be a subgroup of Aut(Z/nZ) = (Z/nZ)∗ for n = 1, 2, 3, . . .. If limn→∞ Hn = ∞, then limn→∞ edk((Z/nZ) ⋊ Hn) = ∞ for any field k of characteristic 0. In particular, limn→∞ edC((Z/nZ)⋊(Z/nZ)∗) = ∞. In the case where n = p is a prime, all Sylow subgroups of (Z/pZ) ⋊ (Z/pZ)∗ are cyclic, so (2) reduces to the vacuous lower bound edC((Z/pZ) ⋊ (Z/pZ)∗) > 1. It was not previously known that ed((Z/pZ) ⋊ (Z/pZ)∗) > 3 for any prime p. 2. Proof of Theorem 2 Let G → GL(V ) be a faithful linear representation of G. By the definition of essential dimension there exists a G-equivariant dominant rational map V 99K X such that G acts faithfully on X and dim(X) = ed(G) 6 n. By Theorem 1, there exists a normal abelian subgroup A ⊳ G such that [G : A] 6 j(n). As we mentioned above, when A is abelian, and k has a primitive eth root of unity, we have rdim(A) = ed(A). Since A is a subgroup of G, ed(A) 6 ed(G) 6 n. Thus there exists a faithful representation W of A of dimension d 6 n. The induced representation V = IndG A(W ) of G is clearly faithful, and dim(V ) = d[G : A] 6 nj(n) = r(n). Thus rdim(G) 6 dim(V ) 6 r(n). This proves (a). To prove (b), note that in some basis e1, . . . , ed of W , A acts on W by diagonal matrices. Choosing a set of representatives g1, . . . , gs for the cosets of A in G, we see that the vectors giej form a basis of V , as i ranges from 1 to s and j ranges from 1 to d. The group G permutes the lines Spank(giej) in V . The subgroup of GL(V ) that fixed each of these lines individually is a maximal torus T = Gds m . The subgroup of GL(V ) that preserves this set of lines is the normalizer N of T in GL(V ), where N ≃ T ⋊ Sds. Thus our faithful representation G → GL(V ) embeds G in N. Since s = [G : A] 6 j(n) and thus ds 6 nj(n) = r(n), we can further embed N into (Gm)r(n) ⋊ Sr(n). (cid:3) Remark 4. Without the assumption that k contains a primitive root ζe of unity of degree e, our proof of Theorem 2(a) only shows that if there exists a number ak(n) such that for every finite abelian group A, then edk(A) 6 n =⇒ rdimk(A) 6 ak(n) edk(G) 6 n =⇒ rdimk(G) 6 ak(n)j(n). Note that Bir(X)(k) ⊂ Bir(X)(k), where k is the algebraic closure of k, we may assume that the Jordan constant j(n) is the same for k as for k. On the other hand, ak(n) may depend on k. Moreover, if k is an arbitrary field of characteristic 0, we do not know whether or not a(n) exists. For example, if p is a prime, then rdimQ(Z/pZ) = p − 1 is not bounded from above, as p increases, but it is not known whether or not edQ(Z/pZ) is bounded from above. Remark 5. Finite subgroups of G2 m ⋊ S, for certain small finite groups S play a promi- nent role in the classification of finite groups of essential dimension 2 (over C), due to A. Duncan; see [Dun13, Theorem 1.1]. Theorem 2(b) suggests that this is not an accident. 4 ZINOVY REICHSTEIN Remark 6. In the definition of the Jordan group we could have dropped the assumption that A is normal: Γ is Jordan if and only if there exists an integer j such that every finite subgroup G ⊂ Γ contains an abelian subgroup A ⊂ G of index [G : A] < j. One usually refers to j and j as the Jordan constant and the weak Jordan constant for Γ, respectively. These constants are related by the inequalities j 6 j 6 j2; see [PS17, Remark 1.2.2]. Indeed, if G has an abelian subgroup of index 6 i, then it has a normal abelian subgroup of index 6 i2; see [I08, Theorem 1.41]. Now observe that our proof of Theorem 2(a) does not use the fact that A is normal. Thus we could have defined r(n) as nj(n), rather than nj(n), in the statement of Theo- rem 2(a). This will not make a difference in this paper, but may be helpful if one tries to find an explicit value for r(n) for some (or perhaps, even all) n. The constants j(n) and j(n) are largely mysterious, but some explicit values for n = 2 and 3 can be found in [PS17]. Remark 7. It is not true that [G : Z(G)] is bounded from above, as G ranges over the groups of essential dimension 6 n. Here Z(G) denotes the center of G. For example let D2n be the dihedral group of order 2n. Then edC(D2n) = 1 for every odd integer n (see [BR97, Theorem 6.2]), but Z(D2n) = 1, and thus [D2n : Z(D2n)] = 2n is unbounded from above, as n ranges over the odd integers. Let l be the field obtained from k by adjoining all roots of unity. Since 3. Proof of Theorem 3 edk(G) > edl(G) for every finite group G, we may replace k by l and thus assume that k contains all roots of unity. Under this assumption, we can restate Theorem 2(a) as follows. Let G1, G2, . . . be a sequence of finite groups. (3) If lim n→∞ rdim(Gn) = ∞, then lim n→∞ ed(Gn) = ∞. The following lemma is elementary; we include a short proof for the sake of complete- ness. Lemma 8. Let q = pa be a prime power, and φ : H → Aut(Z/qZ) = (Z/qZ)∗ be a group homomorphism. Then rdim((Z/qZ) ⋊φ H) > φ(H). Proof. Suppose ρ : (Z/qZ) ⋊φ H → GL(V ) is a d-dimensional faithful representation. Our goal is to show that d > φ(H). By our assumption on k, V as a direct sum of 1-dimensional character spaces V = Vχ1 ⊕ · · · ⊕ Vχd for the cyclic group Z/qZ, where the characters χ1, . . . , χd are permuted by H. Since ρ is faithful, the restriction of one of these characters, say of χi, to the unique subgroup of order p in Z/qZ is non-trivial. Hence, χi : Z/qZ → k∗ is faithful. This implies that the H-orbit of χi has exactly φ(H) elements. Thus d > φ(H), as desired. (cid:3) We are now ready to proceed with the proof of Theorem 3. Set Gn = (Z/nZ) ⋊ Hn. By (3), it suffices to show that limn→∞ rdim(Gn) = ∞. In other words, for any positive JORDAN PROPERTY AND ESSENTIAL DIMENSION 5 real number R, we want to show that there are at most finitely many integers n > 1 such that (4) Write (5) rdim(Gn) 6 R. Hn = (Z/q1Z)a1 × · · · × (Z/qrZ)ar as a product of cyclic groups, where q1, . . . , qr are distinct prime powers. Claim: If (4) holds, then (a) ai 6 R, and (b) qi 6 R, for every i = 1, . . . , r. Assume for a moment that this claim is established. For a fixed R, there are only finitely many groups of the form (Z/q1Z)a1 × · · · × (Z/qrZ)ar satisfying (a) and (b) (recall that q1, . . . , qr are required to be distinct). Since limn→∞ Hn = ∞, we conclude that the inequality rdim(Gn) 6 R holds for only finitely many integers n > 1, as desired. It remains to prove the claim. For part (a), note that R > rdim(Gn) > rdim(Hn) > rdim(Z/qiZ)ai = ai . To prove part (b), by symmetry it suffices to show that q1 6 R. Let n = pe1 prime decomposition of n and let φj : Hn → Aut(Z/pej j Z) be the projection of 1 . . . pes s be the Hn ⊂ Aut(Z/nZ) = Aut(Z/pe1 1 Z) × · · · × Aut(Z/pes s Z) to the jth factor. Since q1 is a prime power, at least one of the projections φj maps the first factor Z/q1Z in (5) isomorphically onto its image. Thus Gn contains a subgroup isomorphic to (Z/pej j Z) ⋊φj (Z/q1Z) and by Lemma 8, R > rdim(Gn) > rdim((Z/pej j Z) ⋊φj (Z/q1Z)) > φj(Z/q1Z) = q1. This completes the proof of the claim and thus of Theorem 3. (cid:3) Example 9. Fix a prime p. For each n > 1, choose a prime qn so that qn − 1 is divisible by pn. There are infinitely many choices of such qn for each n by Dirichlet's theorem of primes in arithmetic progressions. Embed Z/pnZ into the cyclic group (Z/qnZ)∗ of order qn − 1 and form the semidirect product Γn = (Z/qnZn) ⋊ (Z/pnZ). It is shown in [BRV18, Example 3.5] that a conjecture of Ledet implies that edC(Γn) > n. Theorem 3 yields an unconditional proof of a weaker assertion: As is pointed out in [BRV18], it was not previously known that edC(Γn) > 3 for any n. lim n→∞ edC(Γn) = ∞. The author is grateful to Alexander Duncan and Fei Hu for helpful comments. Acknowledgments 6 ZINOVY REICHSTEIN References [Be14] A. Beauville, Finite simple groups of small essential dimension, in Trends in contemporary [Bi16] [BR97] mathematics, 221 -- 228, Springer INdAM Ser., 8, Springer, 2014. MR3586401 C. Birkar, Singularities of arXiv:1609.05543. J. Buhler and Z. Reichstein, On the essential dimension of a finite group, Compositio Math. 106 (1997), no. 2, 159 -- 179. systems and boundedness of Fano varieties, 2016, linear [BRV18] P. Brosnan, Z. Reichstein, and A. Vistoli, Essential dimension in mixed characteristic, [DI09] arXiv:1801.02245. I. V. Dolgachev and V. A. Iskovskikh, Finite subgroups of the plane Cremona group, in Alge- bra, arithmetic, and geometry: in honor of Yu. I. Manin. Vol. I, 443 -- 548, Progr. Math., 269, Birkhauser Boston, Inc., Boston, MA. MR2641179 [Dun10] A. Duncan, Essential dimensions of A7 and S7, Math. Res. Lett. 17 (2010), no. 2, 263 -- 266. MR2644373 [Dun13] A. Duncan, Finite groups of essential dimension 2, Comment. Math. Helv. 88 (2013), no. 3, [I08] 555 -- 585. I. M. Isaacs, Finite group theory, Graduate Studies in Mathematics, 92, American Mathematical Society, Providence, RI, 2008. MR2426855 [J1878] C. Jordan, M´emoire sur les ´equations differentielles lin´eaires `a int´egrale alg´ebrique, J. Reine Angew. Math. 84 (1878) 89 -- 215. [KM08] N. A. Karpenko and A. S. Merkurjev, Essential dimension of finite p-groups, Invent. Math. 172 (2008), no. 3, 491 -- 508. MR2393078 [Kl1884] F. Klein, Vorlesungen uber das Ikosaeder und die Auflosung der Gleichungen vom 5ten Grade, 1884. English translation: Lectures on the icosahedron and the solution of equations of the fifth degree, translated into English by George Gavin Morrice, second and revised edition, Dover Publications, Inc., New York, NY, 1956. MR0080930 [Po11] V. L. Popov, On the Makar-Limanov, Derksen invariants, and finite automorphism groups of algebraic varieties, in Affine algebraic geometry, 289 -- 311, CRM Proc. Lecture Notes, 54, Amer. Math. Soc., Providence, RI. MR2768646 [Po14] V. L. Popov, Jordan groups and automorphism groups of algebraic varieties, in Automorphisms in birational and affine geometry, 185 -- 213, Proc. Math. Stat., 79, Springer, 2014. MR3229352 [Pr12] Y. Prokhorov, Simple finite subgroups of the Cremona group of rank 3, J. Algebraic Geom. 21 (2012), no. 3, 563 -- 600. MR2914804 [Pr17] Y. Prokhorov, Quasi-simple finite groups of essential dimension 3, arXiv: 1703.10780 [PS16] Y. Prokhorov and C. Shramov, Jordan property for Cremona groups, Amer. J. Math. 138 (2016), no. 2, 403 -- 418. MR3483470 [PS17] Y. Prokhorov and C. Shramov, Jordan constant for Cremona group of rank 3, Mosc. Math. J. 17 (2017), no. 3, 457 -- 509. MR3711004 [Rei10] Z. Reichstein, Essential dimension, Proceedings of the International Congress of Mathemati- [Se10] cians. Volume II, Hindustan Book Agency, New Delhi, 2010, pp. 162 -- 188. MR 2827790 J.-P. Serre, Le groupe de Cremona et ses sous-groupes finis, Ast´erisque No. 332 (2010), Exp. No. 1000, vii, 75 -- 100. MR2648675 Department of Mathematics, University of British Columbia, Vancouver, BC V6T 1Z2, Canada E-mail address: [email protected]
1703.00636	1	1703	2017-03-02T06:43:13	Non-geodesic variations of Hodge structure of maximum dimension	[ "math.AG", "math.DG" ]	There are a number of examples of variations of Hodge structure of maximum dimension. However, to our knowledge, those that are global on the level of the period domain are totally geodesic subspaces that arise from an orbit of a subgroup of the group of the period domain. That is, they are defined by Lie theory rather than by algebraic geometry. In this note, we give an example of a variation of maximum dimension which is nowhere tangent to a geodesic variation. The period domain in question, which classifies weight two Hodge structures with $h^{2,0} = 2$ and $h^{1,1} = 28$, is of dimension $57$. The horizontal tangent bundle has codimension one, thus it is an example of a holomorphic contact structure, with local integral manifolds of dimension 28. The group of the period domain is $SO(4,28)$, and one can produce global integral manifolds as orbits of the action of subgroups isomorphic to $SU(2,14)$. Our example is given by the variation of Hodge structure on the second cohomology of weighted projective hypersurfaces of degree $10$ in a weighted projective three-space with weights $1, 1, 2, 5$	math.AG	math	NON-GEODESIC VARIATIONS OF HODGE STRUCTURE OF MAXIMUM DIMENSION JAMES A. CARLSON AND DOMINGO TOLEDO Abstract. There are a number of examples of variations of Hodge structure of maximum dimension. However, to our knowledge, those that are global on the level of the period domain are totally geodesic subspaces that arise from an orbit of a subgroup of the group of the period domain. That is, they are defined by Lie theory rather than by algebraic geometry. In this note, we give an example of a variation of maximum dimension which is nowhere tangent to a geodesic variation. The period domain in question, which classifies weight two Hodge structures with h2 1 = 28, is of dimension 57. The horizontal tangent bundle has codimension one, thus it is an example of a holomorphic contact structure, with local integral manifolds of dimension 28. The group of the period domain is SO(4, 28), and one can produce global integral manifolds as orbits of the action of subgroups isomorphic to SU (2, 14). Our example is given by the variation of Hodge structure on the second cohomology of weighted projective hypersurfaces of degree 10 in a weighted projective three-space with weights 1, 1, 2, 5 0 = 2 and h1 , , 1. Introduction Period domains D = G/V for G a (semi-simple, adjoint linear Lie group with a compact Cartan subgroup T ⊂ G and V the centralizer of a sub-torus of T ) occur in many interesting situations. It is known that there is a unique maximal compact subgroup K ⊂ G containing V , so that there is a fibration (1) K/V −→ G/V π−→ G/K of the homogeneous complex manifold G/V onto the symmetric space G/K with fiber the homogeneous projective variety K/V . The tangent bundle T D has a distinguished horizontal sub-bundle ThD (also called Date: October 8, 2018. 2010 Mathematics Subject Classification. 32G20, 32M10. Key words and phrases. Hodge theory, period domains, horizontal maps. Second author supported by Simons Foundation Collaboration Grant 208853. 1 2 CARLSON AND TOLEDO the infinitesimal period relation). It is a sub-bundle of the differential- geometric horizontal bundle (the orthogonal complement of the tangent bundle to the fibers). It usually, but not always a proper sub-bundle. When it is a proper sub-bundle, it is not integrable. Typically, suc- cessive brackets of vector fields in ThD generate all of T D. We are interested in the case where the symmetric space G/K is not Hermit- ian symmetric. In that case, the complex manifold D admits invariant pseudo-Kahler metrics, but no invariant Kahler metric. These manifolds were introduced by Griffiths as a category of manifolds that contains the classifying spaces of Hodge structures. For example, if (H, h , i) is a real vector space of dimension 2p + q with a symmetric bilinear form of signature 2p, q, the manifolds SO(2p, q)/U(p) × SO(q) classify Hodge decompositions of weight two. Thus, we have a direct sum decomposition H C = H 2,0 ⊕ H 1,1 ⊕ H 0,2 (2) with Hodge numbers (dimensions) h2,0 = h0,2 = p, h1,1 = q, and polar- ized by h , i: The real points of H 2,0 ⊕ H 0,2 form a maximal positive subspace, H 1,1 is the complexification of its orthogonal complement (a maximal negative subspace), and so (H 2.0)⊥ = H 2,0 ⊕ H 1,1. Therefore the filtration H 2,0 ⊂ (H 2,0)⊥ ⊂ H C (3) of H C is the same as the Hodge filtration. Therefore H 2,0 determines the Hodge filtration, hence the Hodge decomposition. Note that hu, vi is a positive Hermitian inner product on H 2,0 The special orthogonal group of h , i, isomorphic to SO(2p, q), acts transitively on the choices of H 2,0, and the subgroup fixing one choice is isomorphic to U(p)×SO(q). Thus, the homogeneous complex manifold D = SO(2p, q)/U(p) × SO(q) classifies polarized Hodge structures on a fixed vector space (H, h , i). Over D, there are tautological Hodge bundles H2,0, H1,1, H0,2. The tangent bundle T D and horizontal sub- bundle are (4) T D = Homh , i(H2,0, H1,1 ⊕ H0,2), ThD = Hom(H2,0, H1,1), where Homh , i means homomorphisms X which preserving h , i in- finitesimally, that is, hXu, vi + hu, Xvi = 0 for all u, v ∈ H 2,0. If X : H 2,0 → H 1,1 this condition is vacuous, since hH 2,0, H 1,1i = 0. Therefore Homh , i(H2,0, H1,1) = Hom(H2,0, H1,1). Whenever p > 1, the horizontal tangent bundle is a proper sub-bundle of the tangent bundle. The first interesting case is p = 2. If in addition NON-GEODESIC VARIATIONS OF MAXIMUM DIMENSION 3 q = 2r is even, then the horizontal distribution locally a contact distri- bution, i.e., is the null space of a form ω = dz − (x1dy1 + · · · + xrdyr) in suitable local coordinates (x, y, z). Our example of weighted hyper- surfaces yields a variation of Hodge structure of this type. 1.1. Construction of horizontal maps. The two main sources of horizontal holomorphic maps to period domains are • Totally geodesic maps: these come from Lie group theory, as orbits of suitable Lie subgroups of G. For example, for the domains SO(2p, 2q)/U(p) × SO(2q), we can put a complex structure J on the underlying R-vector space H, compatible with < , >. Let H +, H − denote the underlying real spaces of H 2,0 ⊕ H 0,2 and H 1,1 respectively. Consider the variation in which all H + are J-invariant. This gives an embedding SU(p, q)/S(U(p) × U(q)) F−→ SO(2p, 2q)/U(p) × SO(2q) of the Hermitian symmetric space D1 for SU(p, q) in the domain D. Since H + always remains J-invariant, the tangent vector to its motion, an element of Hom(H +, H −) commutes with J. Let V ⊂ H 1,1 be the space of (1, 0)-vectors for J, that is, V = {X − iJX X ∈ H 1,1}. Then dF : T D1 → Hom(H 2,0, V ) ⊂ Hom(H 2,0, H 1,1) = ThD in particular F is horizontal and holomorphic. • Periods of families of algebraic varieties This may be called the geometric method. We proceed to explain it by describing the special case of SO(2p, 2q): Let S → B be smooth algebraic family of smooth projective algebraic surfaces over a smooth connected algebraic base B, fix a base point b0 ∈ B, and fix (H, h , i) to be the pair (H 2(Sb0, R)prim, intersection form). For any b ∈ B and a path λ from b0 to b, there is an isomorphism λ# : H 2(Sb) → H 2(Sb0), where different paths give different isomorphisms related by an element of the image of the monodromy representation ρ : π1(B, b0) → Aut(H 2 prim(Sb0)). The period map F is defined by the rule: F (b) is the Hodge structure λ#(Hodge structure on H 2(Sb)). In this way, F (b) is a Hodge structure on a fixed vector space, hence an element of D, well defined up to the action of the monodromy group. We could look at this as a function of b and λ, in which case we are 4 CARLSON AND TOLEDO lifting F to a map eF on a covering space of B. Thus we have two equivalent formulations F, eF of the period map related as follows: (5) eB py B F−→ Γ\D eF−→ D y where p : eB → B is the covering corresponding to the kernel of ρ and Γ is a suitable monodromy group (containing the image of ρ). Locally, the two maps F, eF are the same, except when F (b) is fixed by some non-identity element of Γ. Griffiths showed that F is holomorphic and horizontal, in other words, deF : T eB → F ∗ThD ⊂ T D. Under suitable assumptions, the closure F (B) is an analytic subvariety of Γ\D, hence is a closed horizontal analytic subvariety of Γ\D. 1.2. A concrete example. The preceding discussion can be applied to the family of smooth hypersurfaces in P3 of a fixed degree d. In order to get non-constant variations and for the period domain not to be Hermitan symmetric we need to take d ≥ 5. For d = 5 we have that the Hodge numbers are (4, 44, 4), hence D = SO(8, 44)/U(4) × SO(44) has dimension 182, the horizontal tangent space has dimension 176 and the maximum dimension of an integral submanifold is 88, the dimension of the horizontal SU(4, 22) orbit, see [1] We therefore find two horizontal maps: • Horizontal SU(4, 22) orbits of maximum dimension 88. • Periods of quintic surfaces, a maximal integral manifold, see [2] of dimension 40 (the dimension of the moduli space of quintic surfaces). In general, period domains, can have maximal integral manifolds of many different dimensions. Hypersurfaces generally yield integral man- ifolds of rather small dimension compared to the the maximum possible. We would like to see geometric examples of maximum, or close to max- imum, dimension that come from geometry as opposed to Lie theory. Hypersurfaces in weighted projective spaces provide such examples. NON-GEODESIC VARIATIONS OF MAXIMUM DIMENSION 5 2. The example Let us consider the weighted projective space P(1, 1, 2, 5) with coordi- nates x1, x2, x3, x4 with weights 1, 1, 2, 5 respectively. One may think of P(1, 1, 2, 5) as the quotient of C4 by the C∗-action λ ∈ C∗ which acts by (6) λ · (x1, x2, x3, x4) −→ (λx1, λx2, λ2x3, λ5x4) A weighted homogeneous polynomial of degree d is a linear combination of monomials xk1 1 xk2 4 of total weighted degree d = k1 + k2 + 2k3 + 5k4 3 xk4 2 xk3 (7) For fixed d, the collection of weighted polynomials of degree d forms a vector space that we will denote Sd(1, 1, 2, 5), or, simply Sd. The direct sum S(1, 1, 2, 5) = ⊕dSd(1, 1, 2, 5) is the algebra of weighted homogeneous polynomials. Any f ∈ Sd defines a subvariety Vf ⊂ P (1, 1, 2, 5), namely Vf = {(x1 : x2 : x3 : x4)f (x1, x2, x3, x4) = 0}. If the only common solution of ∂f ∂x1 = 0, . . . , ∂f ∂x4 = 0 is (0, 0, 0, 0), then Vf is called a quasi-smooth subvariety. It is smooth except possibly for quotient singularities. Topologically it is a rational homology manifold, and in particular satisfies Poincar´e duality over Q. Its second cohomology has a pure Hodge structure of weight two, polarized by the intersection form. Fix d and let S0 d ⊂ Sd denote the set, possibly empty, of all f ∈ Sd for which Vf is quasi-smooth. For example, if f ∈ S4, then no monomial in f can contain the variable x4 of weight 5, so ∂f = 0 for all f ∈ S4. Therefore S0 4 = ∅ since (0 : 0 : 0 : 1) is a singular point of all f ∈ S4. On the other hand, a polynomial in Sd is a sum of powers of all of the variables defines a Fermat hypersurface. These are always quasi- smooth. In our case, one has the Fermat surface ∂x4 (8) f0(x1, x2, x3, x4) = x10 1 + x10 2 + x5 3 + x2 4 ∈ S0 10, It has a rich structure, and, in particular, is double cover of the 2- dimensional weighted projective plane with weights 1, 1, 2, branched over a curve of degree ten. 6 CARLSON AND TOLEDO d is a subvariety of Sd. It is a proper d 6= ∅. The complement ∆d = Sd \ S0 subvariety if S0 Assume S0 quently, S0 fibration V → S0 d 6= ∅. Then ∆d has complex codimension 1 in Sd. Conse- d is connnected and we obtain a topologically locally trivial d where the fiber over f is the variety Vf : (9) V = {(f, x)f (x) = 0} ⊂ S0 d × P(1, 1, 2, 5) y S0 d y S0 d = Fix a base point f0 ∈ S0 d. Then there is a monodromy representa- tion ρ : π1(S0 d, f0) → Aut(H 2(Vf0)), where Aut is the group of au- tomorphisms respecting all topological structures, in particular, the intersection form. As f varies, we transport the Hodge structure on H 2(Vf , C)prim = H 2,0(Vf ) ⊕ H 1,1(Vf )prim ⊕ H 0,2(Vf ) to H 2(Vf0)prim, as explained in §1, thus obtaining a point F (f ) ∈ D, well defined up to the action of the image of ρ, where D is the classifying space of Hodge structures on H 2(Vf0)prim. This defines holomorphic period maps as in (5), namely (10) d fS0 py S0 d eF−→ D y F−→ Γ\D where Γ denotes the image of the monodromy representation ρ, and which is horizontal in the sense that (11) deF : T eS0 d −→ eF ∗ThD. We must look carefully at some local properties of the period map F . Let U be a simply connected neighborhood of the base point f0. The inverse image of U in fS0 d is a disjoint union of open sets isomorphic to U. On such a component of the inverse image, we can replace the map eF of (10) by its restriction to a that connected component. Identifying it with U, we may replace (10) by the simpler diagram (12) U D eF ր y F−→ Γ\D Thus the period map F to Γ\D is locally liftable to D. This is only an issue in the presence of fixed points. NON-GEODESIC VARIATIONS OF MAXIMUM DIMENSION 7 Our example of a horizontal non-geodesic V ⊂ Γ\D will be F (S0 d), the closure of the image of F , for suitable d. We proceed to the necessary computations. 2.1. The Jacobian Ring. First of all, choose d = 10, and consider the space S10(1, 1, 2, 5) of weighted homogeneous polynomials of degree 10 with weights (1, 1, 2, 5). Some computer experimentation led us to this choice. As noted above, the "Fermat hypersurface"Vf0 is defined by an element of S10, and so S0 10 6= ∅. Given f ∈ S0 10, let (i ) J(f ) ⊂ S denote the Jacobian ideal of f , namely the ideal generated by the partial derivatives of f . (ii ) R(f ) = S/J(f ) be the Jacobian ring of f . The Hodge decomposition and the differential of the period map have very explicit descriptions in terms of the graded ring R(f ) for f ∈ S0 10. Since the dimensions of the graded components Rk(f ) are independent of f , we often write simply Rk for Rk(f ). Proposition 1. Let f ∈ S0 10 and let J and R be as just defined. Then ∼= H 2,0 ∼= H 1,1 ∼= H 0,2 ∼= C (i ) R1 (ii ) R11 (iii ) R21 (iv ) R22 (v ) Rk = 0 for k > 22 (vi ) For 0 ≤ i ≤ 22, the pairing Ri ⊗ R22−i → R22 is non- degenerate. Proof. Statements of this type for projective hypersurfaces are conse- quences of the Griffiths residue calculus. The analogous statements for weighted projective hypersurfaces are proved in Theorem 1 of [8] and in §4.3 of [5]. (cid:3) Applying the above to our situation, and using the polynomial f0 to do computations, we find Lemma 2. (i ) h2,0 = 2, h1,1 = 28, h0,2 = 2 (ii ) D = SO(4, 28)/U(2) × SO(28) (iii ) D has dimension 57. (iv ) The horizontal sub-bundle ThD = Hom(H2,0, H1,1) has fiber dimension 56, hence is a holomorphic contact structure on D. 8 CARLSON AND TOLEDO Proof. Since the Hodge numbers are independent of f , we can compute them for f0. Using Proposition 1, this is the same as computing the spaces Rk(f0), which amounts to a straightforward exercise of counting monomials. First of all, J is the ideal generated by x9 3, x4. We find that 2, x4 1, x9 (i ) R1 = S1 = hx1, x2i is the vector space with basis x1, x2, so that h2,0 = h0,2 = 2. 1, x9 (ii ) R11: to find a basis for this space, list all monomials that do not contain any of the above generators of J. In particular, x4 does not appear, so a basis consists of monomials in x1, x2, x3 that do not contain x9 2, x4 3. These can be conveniently grouped by powers of x3: 3i = 0, . . . 5(cid:11) is six-dimensional (a) G3 = (cid:10)xi 1x5−i 2 x3 (b) G2 = (cid:10)xi 3i = 0, . . . 5(cid:11) is eight-dimensional 1x7−i 2 x2 2 x3i = 1, . . . 8(cid:11) is eight-dimensional (c) G1 = (cid:10)xi 1x9−i i = 3, . . . 8(cid:11) is six-dimensional (d) G0 = (cid:10)xi 1x11−i Therefore dim R11 = h1,1 = 28 2 (iii ) It follows that D classifies polarized Hodge structures with Hodge numbers 2, 28, 2. From the discussion in the introduc- tion, it follows that D = SO(4, 28)/U(2) × SO(28), which has dimension 57 and its sub-bundle ThD = Hom(H2,0, H1,1) has fiber dimension h2,0h1,1 = 56. The easiest way to visualize D, and to see its dimension and the structure of the horizon- tal sub-bundle, is to use its fibration (1) over the symmetric space. In this case the symmetric space has real dimension 4 · 28 and the fiber is a projective line: SO(4)/U(2) −→ SO(4, 28)/U(2) × SO(28) (13) yπ SO(4, 28)/S(O(4) × O(28)) It is easy to see that dπ maps the fibers of ThD isomorphically (as real vector spaces) to the tangent spaces to the symmetric space. Thus ThD coincides, in this case, with the differential- geometric horizontal bundle. (iv ) To see that ThD is a holomorphic contact structure, recall the identification (4), T D ∼= Homh , i(H2,0, H1,1 ⊕ H0,2). Under this identification, ThD is identified with Hom(H2,0, H1,1) as the kernel of the projection to Hom< , >(H2,0, H0,2). Since NON-GEODESIC VARIATIONS OF MAXIMUM DIMENSION 9 Homh , i(H 2,0, H 0,2) is a space of skew-symmetric endomor- phisms, and since dim H 2,0 = 2, we see that dim Homh , i(H 2,0, H 0,2) = 1 The projection is a one-form ω with values in the line bundle TvD = Homh , i(H 2,0, H 0,2) whose kernel is ThD. Here TvD stands for the vertical bundle. To be a contact structure means that it is totally non-integrable. This means the following: if X, Y are horizontal vector fields, then, for all p ∈ D, ω([X, Y ])p depends only on Xp, Yp, hence defines a bundle map Λ2ThD → TvD. To be a contact structure then means that this is a non- degenerate pairing. In other words, the resulting map ThD → Hom(ThD, TvD) is an isomorphism. This is a reformulation of the local coordinate condition ω ∧ (dω)28 6= 0 at every point. Under our identification ThD ∼= Hom(H2,0, H1,1), it is easy to check that ω([X, Y ]) = X tY − Y tX, where the transpose is with respect to < , >, see §6 of [3] for details. One easily checks that this paring is non-degenerate, so that we indeed have a contact structure. (cid:3) Next, we compute dF , where F : S0 10 → Γ\D is the period map of (10). The group G(1, 1, 2, 5) of automorphisms of P (1, 1, 2, 5) acts on S0 10 and F is constant on orbits, so it should factor through an appropriate quotient. Since the group is not reductive, we avoid the technicalities of forming quotients, by working mostly on the infinitesimal level. Given f ∈ S0 10, the tangent space at f to its G(1, 1, 2, 5)-orbit is J10(f ). When we have a quotient, R10(f ) can be identified with the tangent space to the quotient at the orbit of f . We use this fact as a guiding principle, relying on the fact that df F vanishes on J10(f ) and hence factors through R10(f ). Thus we avoid working with the quotient di- rectly. To be more precise, fix f ∈ S0 10 and a simply connected neighborhood U of f . Since Γ\D need not be a manifold (and will not be at points fixed by non-identity elements of Γ), what we actually want to compute is df eF , where eF : U → D is a local lift of F as in (12). Since U is an open subset of the vector space S10, there is a canonical identification (14) Tf U ∼= S10 by translation. 10 CARLSON AND TOLEDO Under this identification, J10(f ) is the tangent space to the orbit of f . Consequently, df eF : S10 → ThD vanishes on J10(f ), hence factors through R10(f ). Keeping in mind the exact sequence (15) 0 −→ J10(f ) −→ S10 p −→ R10(f ) −→ 0, we can state the main tool for computing differentials of period maps: Proposition 3. Under the isomorphisms of Proposition 1, the isomor- phism (14), and p as in (15), we have a commutative diagram (16) Tf U ∼=y S10 eF df −→ p −→ R10(f ) ThD ∼= Hom(H 2,0, H 1,1) y∼= m−→ Hom(R1(f ), R11(f )) where, for φ ∈ R10, m(φ) : R1 → R11 is multiplication by φ: if x ∈ R1, then m(φ)(x) = φx Proof. This is the content of the residue calculus. The isomorphisms between holomorphic objects and elements of the Jacobian ring pre- serve all natural products and pairings. (cid:3) The above proposition will allow us to compute the rank of d eF at the point f0 of (8). We remark that, up to this point, the residue calcu- lus and the corresponding algebraic facts about the Jacobian ring have closely paralleled the projective case. But the failure of Macauley's theorem in the weighted projective case forces us to look carefully at the remaining statements. Most results in the literature require as- sumptions on the weights, and on the degree, that are not satisfied for degree 10 and weights (1, 1, 2, 5). See the introduction and §1 of [6] for a general discussion of the possible difficulties that can appear in the weighted case. Proposition 4. imum possible rank of a horizontal holomorphic map. (i ) The rank of d eF at f0 is 28, which is the max- (ii ) Let W ⊂ ThD denote the image of d eF . Under the identifica- tion ThD ∼= Hom(H 2,0, H 1,1), we have: (a) For each v ∈ H 2,0, the subspace W v =def {Xv X ∈ W } ⊂ H 1,1 has dimension 26. (b) {Xv v ∈ H 2,0, X ∈ W } = H 1,1 Proof. By Proposition 3 we need to compute the multiplication map R10 → Hom(R1, R11). In the proof of Lemma 2 we found a basis for R11, and we can do a similar calculation with R10: a basis will be given NON-GEODESIC VARIATIONS OF MAXIMUM DIMENSION 11 by the monomials xa 3 of total weight 10 with 0 ≤ a, b ≤ 8 and 0 ≤ c ≤ 3. These can again be conveniently grouped by the powers of x3: 1, xb 2, xc (i ) G′ (ii ) G′ (iii ) G′ (iv ) G′ 2 3 = (cid:10)xi 2 = (cid:10)xi 1 = (cid:10)xi 0 = (cid:10)xi 3i = 0, . . . 5(cid:11) is five-dimensional 1x4−i 2 x3 3i = 0, . . . 5(cid:11) is seven-dimensional 1x6−i 2 x2 2 x3i = 1, . . . 8(cid:11) is nine-dimensional 1x8−i i = 2, . . . 8(cid:11) is seven-dimensional 1x10−i Therefore dim R10 = 28, as claimed. 2 ⊕ G′ 3 ⊕ G′ 2 ⊕ G′ 0 and R11 = G3 ⊕ G2 ⊕ G1 ⊕ G0, Next, we examine the map m : R10 → Hom(R1, R11), where m(φ) is the homomorphism m(φ)(x) = φx. We claim that m is injective. Since R1 = hx1, x2i, it suffices to show that if φ ∈ R10 and both φx1 = φx2 = 0, then φ = 0. We have (17) R10 = G′ it is easy to see that multiplication by R1 maps G′ i to Gi, that mul- tiplication by x1 is injective for i = 2, 3, and that the same holds for multiplication by x2. Moreover multiplication by either x1 or x2 is surjective for i = 0, 1 and the intersection of their kernels is zero. Writing φ = φ3 + · · · + φ0 and applying this information we see that φx1 = φx2 = 0 implies φ = 0. Combining these two facts, we see that df0 eF has rank 28. Since its image is an integral element of the holomorphic contact structure ThD, its dimension can be at most half of 56, the fiber dimension of ThD. Therefore eF has the highest possible rank of a horizontal holomorphic map, namely 28. The second part is easily verified using the above bases of monomials. For v = x1 or x2, both assertions are clear, and they are easily checked for linear combinations v = ax1 + bx2. (cid:3) 2.2. A closed horizontal subvariety of maximum dimension. Consider now the horizontal holomorphic map F : S0 10 → Γ\D. Fol- lowing Griffiths (see §9 of [7]) we can embed S0 10 ⊂ S′, where S′ is a smooth complex manifold containing S0 10 as the complement of an analytic subset. One does this by compactifying with normal crossing divisors. One can then extend over the branches of the compactifying divisor for which the monodromy is finite to obtain a proper horizon- tal holomorphic map F : S′ → Γ\D. Then F (S′) is a closed analytic 12 CARLSON AND TOLEDO subvariety of Γ\D containing F (S0 subvariety. 10) as the complement of an analytic 10, we found that a local lift eF : U → D has At the point f0 ∈ S0 maximum rank 28. Consequently, there is a neighborhood U ′ of f0, where U ′ ⊂ U, eF has rank 28, and eF U ′ is a submersion onto its image. Therefore eF (U ′) is a 28-dimensional horizontal submanifold of D containing eF (f0). We now examine the local structure of Γ\D. Since f0 has symmetries, eF (f0) is fixed by some element γ ∈ Γ, γ 6= id. Let Γ0 denote the subgroup of Γ fixing eF (f0). It is necessarily a finite group. If N is a Γ0-invariant neighborhood of eF (f0), then Γ0\N is an orbifold neigh- borhood of F (f0) in the orbifold Γ\D, and F (f0) is a singular point of this orbifold. Strictly speaking, we do not have a tangent space at F (f0). But we can move away from f0 in the above neighborhood U ′ to find non-singular points: Lemma 5. Let W ⊂ (Th) eF (f0)D denote the image of d(f0) eF . Then (i ) W is not fixed by any γ ∈ Γ0, γ 6= id. (ii ) W is not tangent to any horizontal geodesic embedding of SU(2, 14)/S(U(2) × U(14)) passing through eF (f0). Proof. As usual, identify ThD with Hom(H 2,0, H 1,1), and let V = H 2,0, V ′ = H 1,1. The group Γ0 acts on ThD through the action of the isotropy group U(2) × SO(28) of eF (f0). Namely (A, B), where A ∈ U(2) and B ∈ SO(28) acts on X ∈ Hom(V, V ′) by X → BXA−1. Let us prove the stronger statement that W is not fixed by any element of U(2)×SO(28): Suppose X is fixed by (A, B) 6= id, say A 6= id. Then BX = XA. Let λ1, λ2 be the eigenvalues of A (roots of unity), and assume, first, that λ1 6= λ2 and neither eigenvalue is real. Let V1, V2 be the corresponding eigenspaces, it is easy to see that, for vi ∈ Vi, Xvi is an eigenvector for B with eigenvalue λi. From this we see that V ′ = V ′ 1⊕V ′ 2 are the eigenspaces of B for λ1, λ2 respectively, and V ′ i for i = 1, 2. In other words, W ⊂ Hom(V1, V ′ 2). Observe that dim V ′ 2 ≤ 14, since B is real and its eigenvalues come in complex conjugate pairs. Therefore, if v1 ∈ V1, {Xv1 X ∈ W } ⊂ V ′ 1. 2 ⊕V ′ 3 is their orthogonal complement. If X ∈ W , then X(Vi) ⊂ V ′ 1) ⊕ Hom(V2, V ′ 3 , where V ′ 1, V ′ 1, dim V ′ Since dim V ′ 1 ≤ 14, this contradicts Proposition 4. The remaining possibilities for λ1, λ2 are handled by similar arguments. This proves NON-GEODESIC VARIATIONS OF MAXIMUM DIMENSION 13 that W is not fixed by any element of the isotropy group of eF (f0). The first part of the Lemma is proved. For the second part, recall from §1.1 that the tangent space to a geo- desic embedding of the symmetric space of SU(2, 14) through the point V = H 2,0 is determined by a complex structure J on V ′ = H 1,1 and is the subspace of X ∈ Hom(V, V ′) satisfying JX = Xi, in other words, the fixed point set of the element (i, J) of U(2) × SO(28), which we have already excluded. (cid:3) An immediate consequence of this lemma is that eF (U ′) is not fixed by any γ ∈ Γ0, so there exist f ∈ U ′ with F (f ) a smooth point of Γ\D. The same must be true in a neighborhood U ′′ ⊂ U ′ of f , so F U ′′ : U ′′ → (Γ\D)0 (the regular points of Γ\D) and rank of dF must be 28 on U ′′. In summary: Theorem 6. Let S′, F : S′ → Γ\D and eF : fS0 Then 10 → D be as above. (i ) F is a proper horizontal holomorphic map. (ii ) There is a proper analytic subvariety Z ⊂ S′ so that, if S′′ = S′ \ Z, then F S ′′ : S′′ → (Γ\D)0 and dF has rank 28 on S′′. (iii ) F (S′) is a closed horizontal subvariety of Γ\D of maximum possible dimension 28. (iv ) If x ∈ S′′, the tangent space to F (S′) at F (x) is not the tangent space to any totally geodesic immersion of the symmetric space of SU(2, 14) in Γ\D. (v ) Alternatively, if x ∈ fS0 10 lies in the dense open set where dx eF has maximum rank 28, the image of dx eF is not the tangent space to a geodesic embedding of the symmetric space SU(2, 14) in D. 3. Geodesic submanifolds and integral elements We close with some remarks on integral elements of contact structures. The period domains for which the horizontal bundle gives a contact structure are the twistor spaces of the quaternionic-Kahler symmetric spaces, also called the Wolf spaces, see [9] for their classification. We briefly discuss two examples from this point of view: our example D, 14 CARLSON AND TOLEDO associated to the symmetric space SO(4, 28)/S(O(4) × O(28)), and another example we call D′ associated to quaternionic hyperbolic space. Whenever the horizontal sub-bundle ThD of a domain D is a contact structure, we know that each fiber of ThD has a symplectic structure, and the integral elements in that fiber are the Lagrangian subspaces of this symplectic structure. Lagrangian subspaces of a 2g-dimensional symplectic space are parametrized by Sp(g)/U(g), the compact dual of the Siegel upper half plane of genus g. If D = SO(4, 28)/U(2) × SO(28) is the domain we have been study- ing, of dimension 57, ThD of dimension 56, the integral elements in a fiber of ThD are parametrized by Sp(28)/U(28), a manifold of complex dimension (28 · 29)/2 = 406. On the other hand, the totally geodesic embeddings of D1, the symmetric space for SU(2, 14) through a fixed point in D are parametrized by the choice of complex structure J on the space H + as in §1.1. These are in turn parametrized by the space SO(28)/U(14) of dimension 28 · 27 − 142 = 14 · 13 = 182. Thus we see that the space of tangents to geodesic embeddings of SU(2, 14) is a rather small subset of the space of Lagrangian subspaces. We therefore expect the generic horizontal map to miss these embeddings. In a way, this is what made our example possible. 3.1. The quaternionic hyperbolic space. We conclude with a re- lated problem, which was the motivation for writing this paper. Con- sider the period domain D′ associated to the quaternionic hyperbolic space, namely (18) Sp(1)/U(1) −→ D′ = Sp(1, n)/U(1) × Sp(n) yπ Sp(1, n)/Sp(1) · Sp(n) We can think of this domain as classifying Hodge structures on R4n+4 ∼= Hn+1 with Hodge numbers 2, 4n, 2 which are stable under right multi- plication by quaternions. Equivalently, we can think of points in this domain as pairs L, J where L ⊂ Hn+1 is a positive right-quaternionic line and J : L → L is a right quaternionic linear complex structure on L orthogonal with respect to the polarizing form h , i. Let L⊥ denote the orthogonal complement of L in Hn=1 and LC, L⊥ C their complexifi- cations. Then the horizontal tangent space to the domain D′ is C ⊕ L0,1) C ) ⊂ T D′ =C HomH(L1,0, L⊥ TnD′ =C HomH(L1,0, L⊥ where CHomH denotes left C-linear and right H-linear homomorphisms. See §6 of [4] for a more detailed discussion. NON-GEODESIC VARIATIONS OF MAXIMUM DIMENSION 15 Once again, D′ has complex dimension 2n + 1 and ThD′ has fiber di- mension 2n, so it is a holomorphic contact structure on D′. Each fiber of ThD′ has a symplectic structure, and the integral elements of the contact structure in a fixed fiber coincide with the Lagrangians of this symplectic structure, and are therefore parametrized by Sp(n)/U(n). We also have horizontal totally geodesic embeddings of the symmetric space of SU(1, n) in D′, namely the unit ball or complex hyperbolic space SU(1, n)/U(n). The group Sp(n) acts transitively on the embed- dings passing through a point (L, J), corresponding to orthogonal right H-linear complex structures on L⊥, hence parametrized by the same homogeneous space Sp(n)/U(n) that parametrizes the Lagrangians. Thus, for D′, every horizontal subvariety of maximum dimension n is tangent, at each smooth point, to a horizontal totally geodesic complex hyperbolic n-space. (We used this fact in §6 of [4] to give a structure theory for harmonic maps of Kahler manifolds to manifolds covered by quaternionic hyperbolic space). Problem. Find examples of discrete groups Γ ⊂ Sp(1, n) and closed horizontal subvarieties V ⊂ Γ\D′ that are not totally geodesic. References [1] J. A. Carlson, Bounds on the dimension of a variation of Hodge Structure, Trans. AMS 294 (1986), 45 -- 64. [2] J. A. Carlson and R. Donagi, Hypersurface variations are maximal, I, Invent. Math. 89 (1987) 371 -- 374. [3] J. A. Carlson and D. Toledo, Variations of Hodge structure, Legendre subman- ifolds, and accessibility, Trans. AMS 311 (1989), 391 -- 411 [4] J.A. Carlson and D. Toledo, Harmonic mappings of Kahler manifolds to locally symmetric spaces, Pub. Maths. IHES 69 (1989), 173 -- 201. [5] I. Dolgachev, Weighted projective varieties, in Lecture Notes in Mathematics 956, 34 -- 71, Springer, 1982. [6] R. Donagi and L. W. Tu, Generic Torelli for weighted hypersurfaces, Math. Annalen 276 (1987), 399 -- 413. [7] P. A, Griffiths, Periods of integrals on algebraic manifolds, III, Pub. Maths. IHES, 38 (1970), 125 -- 180. [8] J. Steenbrink, Intersection form for quasi-homogeneous singularities, Comp. Math. 34 (1977), 211 -- 223. [9] J. A Wolf, Complex homogeneous contact manifolds and quaternionic sym- metric spaces, Jour. Math. Mechanics 14 (1965), 1033 -- 1048. 16 CARLSON AND TOLEDO Department of Mathematics, University of Utah, Salt Lake City, UT 84112 E-mail address: [email protected] URL: http://www.math.utah.edu/~carlson Department of Mathematics, University of Utah, Salt Lake City, UT 84112 E-mail address: [email protected] URL: http://www.math.utah.edu/~toledo
1210.3980	8	1210	2017-06-29T00:58:42	On the Cartier Duality of Certain Finite Group Schemes of order $p^n$, II	[ "math.AG" ]	We explicitly describe the Cartier dual of the $l$-th Frobenius kernel $N_l$ of the deformation group scheme, which deforms the additive group scheme to the multiplicative group scheme. Then the Cartier dual of $N_l$ is given by a certain Frobenius type kernel of the Witt scheme. Here we assume that the base ring $A$ is a $Z_{(p)}/(p^n)$-algebra, where $p$ is a prime number. The obtained result generalizes a previous result by the author which assumes that $A$ is a ring of characteristic $p$.	math.AG	math	On the Cartier duality of certain finite group schemes of order pn, II Michio Amano Revised June 29, 2017∗ Abstract We explicitly describe the Cartier dual of the l-th Frobenius kernel Nl of the group scheme G(λ), which deforms Ga to Gm. Then the Cartier dual of Nl is given by a certain Frobenius type kernel of the Witt scheme. Here we assume that the base ring A is a Z(p)/(pn)-algebra, where p is a prime number. The obtained result generalizes a previous result by the author [1] which assumes that A is an Fp-algebra. 1 Introduction Throughout this paper, we denote by p a prime number. Let A be a commutative ring with unit and λ a suitable element of A. We consider the group scheme G(λ) which deforms the additive group scheme Ga,A to the multiplicative group scheme Gm,A determined by λ (we recall the group structure of G(λ) in Section 3 below). The group scheme G(λ) has been treated by F. Oort, T. Sekiguchi and N. Suwa [5] and by W. Waterhouse and B. Weisfeiler [10] in detail. The group scheme G(λ) is useful for studying the deformation of Artin-Schreier theory to Kummer theory. In particular, the surjective homomorphism ψ : G(λ) → G(λp); x 7→ λ−p{(1 + λx)p − 1} plays an important role in the unified Kummer-Artin-Schreier theory. In this paper, we explicitly describe the Cartier dual of a certain kernel given by a homomorphism ψ(l) generalized ψ. We remark that ψ is nothing but the Frobenius homomorphism over the base ring of the characteristic p. Under this assumption, Y. Tsuno [9] has shown the following: Theorem 1 ([9]) Assume that A is an Fp-algebra. Then the Cartier dual of Ker(ψ) is canonically isomorphic to Ker[F − λp−1 : Ga,A → Ga,A], where F is the Frobenius endomorphism. ∗A corrigendum was added at the end of the previous version [arXiv:1210.3980v7]. 1 Note that Tsuno's result is a special case of the result obtained by F. Oort and J. Tate [6]. Tsuno's result, however, is embedding certain classified finite group schemes of order p into G(λ) over A[ p−1√b], as λ = p−1√b for an element b ∈ A. The author has generalized Tsuno's theorem as follows. Let A be an Fp-algebra and l a positive integer. We consider the surjective homomorphism ψ(l) : G(λ) → G(λpl ); x 7→ λ−pl {(1 + λx)pl − 1}. Then we have ψ(l)(x) = xpl by our assumption. Put Nl := Ker(ψ(l)). Suppose that WA is the Witt ring scheme over A. Let F : WA → WA be the Frobenius endomorphism and [λ] : WA → WA the Teichmuller lifting of λ ∈ A. Set F (λ) := F − [λp−1]. We restrict F (λ) to the Witt ring scheme Wl,A of length l. The result of the previous paper [1] is: Theorem 2 ([1]) Assume that A is an Fp-algebra. Then the Cartier dual of Nl is canon- ically isomorphic to Ker[F (λ) : Wl,A → Wl,A]. To prove Theorem 2, we have used the deformations of Artin-Hasse exponential series introduced by T. Sekiguchi and N. Suwa [8] and a duality between Ker[F (λ) : W (A) → W (A)] with a formal completion of G(λ) proved by them [Ibid]. Theorem 2 has been constructed by assuming the characteristic p. We do not assume it. Our arguments are as follows. Let n be a positive integer. Suppose that Z(p) is a localization of rational integers Z at p. Let A be a Z(p)/(pn)-algebra and λ a suitable element of A. Here, for each integer 0 ≤ k ≤ l − 1, we assume that pl−kλpk is divided by λpl (if λ = 0, we put pl−kλpk/λpl := 0) and that pl−kλpk/λpl is nilpotent. Then the homomorphism ψ(l) is well-defined and Nl = Ker(ψ(l)) is a finite group scheme of order pl, since ψ(l)(X) is a monic polynomial of the degree pl. For a ∈ W (A), T. Sekiguchi and N. Suwa [7] have introduced an endomorphism Ta on W (A) (we recall the definition of Ta a := F (λ) ◦ Ta. in Section 2 below). Put W (A)/Ta := Coker[Ta : W (A) → W (A)]. Set T ′ Put W (A)/T ′ a := Coker[T ′ a : W (A) → W (A)]. We consider the diagram W (A) −−−→ W (A)/Ta W (A) −−−→ W (A)/T ′ a. F (λ)y yF (λ) Here F (λ) is defined by F (λ)(x ) := F (λ)(x ). It is shown that the homomorphism F (λ) is well-defined and that the above diagram is commutative. Put a := λ−plpl[λ] ∈ W (A). Then the result of this paper is: Theorem 3 With the above notations, the Cartier dual of Nl is canonically isomorphic to Ker[F (λ) : WA/Ta → WA/T ′ a ]. 2 The case n = 1 of Theorem 3 is nothing but Theorem 2 except restricting λ ∈ A. In fact, if n = 1, we have Ta = V l ([1, Lemma 1, p.123]), where V is the Verschiebung endomorphism. Then Theorem 3 is stated by Ker[F (λ) : WA/Ta → WA/T ′ a ] ≃ Ker[F (λ) : Wl,A → Wl,A ⊂ WA/T ′ a]. The framework of our proof is similar to the previous paper [1]. But we do not assume the characteristic p. Then the equality Ker(F (λpl )) = Ker(F (λ) ◦ Ta ) is our important tool (we prove this equality in Subsection 4.1 below). The contents of this paper are as follows. The next two sections are devoted to recalling the definitions and the some properties of the Witt scheme and of the deformed Artin-Hasse exponential series. In Section 4 we give our proof of Theorem 3. Acknowledgments The author express gratitude to Professor Tsutomu Sekiguchi for his kind advice and suggestions. He also would like to thank Dr. Yuji Tsuno for suggesting the representability of the quotient group schemes. Furthermore he is grateful to Dr. Takayuki Yamada for his advice to improve the presentations, and the referee for a number of suggestions improving the paper. Finally he should express hearty thanks to people of high school attached to Chiba university of commerce for hospitality. Notations Ga,A : additive group scheme over A Gm,A : multiplicative group scheme over A bGm,A : multiplicative formal group scheme over A Wn,A : group scheme of Witt vectors of length n over A WA : group scheme of Witt vectors over A F : Frobenius endomorphism of WA [λ] : Teichmuller lifting (λ, 0, 0, . . .) ∈ W (A) of λ ∈ A F (λ) : = F − [λp−1] Ta : homomorphism decided by a ∈ W (A) (recalled in Section 2) a (p) : = (ap for a = (a0, a1, . . .) ∈ W (A) 0, ap 1, . . .) W (A)F (λ) : = Ker[F (λ) : W (A) → W (A)] W (A)/F (λ) : = Coker[F (λ) : W (A) → W (A)] W (A)/Ta : = Coker[Ta : W (A) → W (A)] W (A)/T ′ a : W (A) → W (A)] a : = Coker[T ′ 3 2 Witt vectors In this short section we recall necessary facts on Witt vectors for this paper. For details, see [3, Chap. V] or [4, Chap. III]. 2.1 Let X = (X0, X1, . . .) be a sequence of variables. For each n ≥ 0, we denote by Φn(X) = Φn(X0, X1, . . . , Xn) the Witt polynomial Φn(X) = X pn 0 + pX pn−1 1 + · · · + pnXn in Z[X] = Z[X0, X1, . . .]. Let Wn,Z = Spec(Z[X0, X1, . . . , Xn−1]) be an n-dimensional affine space over Z. The phantom map Φ(n) is defined by Φ(n) : Wn,Z → An Z; x 7→ (Φ0(x ), Φ1(x ), . . . , Φn−1(x )), Z is the usual n-dimensional affine space over Z. The scheme An where An Z has a natural ring scheme structure. It is known that Wn,Z has a unique commutative ring scheme structure over Z such that the phantom map Φ(n) is a homomorphism of commutative ring schemes over Z. Then A-valued points Wn(A) are called Witt vectors of length n over A. 2.2 We define a morphism F : W (A) → W (A) by Φi(F (x )) = Φi+1(x ) for x ∈ W (A). If A is an Fp-algebra, F is nothing but the usual Frobenius endomor- phism. Let [λ] be the Teichmuller lifting [λ] = (λ, 0, 0, . . .) ∈ W (A) for λ ∈ A. Set the endomorphism F (λ) := F − [λp−1] on W (A). For a = (a0, a1, . . .) ∈ W (A), we also define a morphism Ta : W (A) → W (A) by Φn(Ta(x )) = a0 pn Φn(x ) + pa1 pn−1 Φn−1(x ) + · · · + pnanΦ0(x ) for x ∈ W (A) ([7, Chap.4, p.20]). 3 Deformed Artin-Hasse exponential series In this short section we recall necessary facts on the deformed Artin-Hasse exponential series for this paper. 4 3.1 Let A be a ring and λ an element of A. Put G(λ) := Spec(A[X, 1/(1 + λX)]). We define a morphism α(λ) by α(λ) : G(λ) → Gm,A; x 7→ 1 + λx. It is known that G(λ) has a unique commutative group scheme structure such that α(λ) is a group scheme homomorphism over A. Then the group scheme structure of G(λ) is given by x · y = x + y + λxy. If λ is invertible in A, α(λ) is an A-isomorphism. On the other hand, if λ = 0, G(λ) is nothing but the additive group scheme Ga,A. 3.2 The Artin-Hasse exponential series Ep(X) is given by We define a formal power series Ep(U, Λ; X) in Q[U, Λ][[X]] by Ep(U, Λ; X) = (1 + ΛX) (1 + Λpk X pk ) 1 pk (( U Λ )pk −( U Λ )pk−1 ). X pr pr ! ∈ Z(p)[[X]]. Ep(X) = exp Xr≥0 ∞Yk=1 U Λ As in [8, Corollary 2.5.] or [7, Lemma 4.8.], we see that the formal power series Ep(U, Λ; X) is integral over Z(p). Note that Ep(1, 0; X) = Ep(X). Let A be a Z(p)-algebra. For λ ∈ A and v = (v0, v1, . . .) ∈ W (A), we define a formal power series Ep(v , λ; X) in A[[X]] by Ep(v , λ; X) = Ep(vk, λpk ; X pk ) = (1 + λX) v0 λ (1 + λpk 1 pk λpk Φk−1(F (λ)(v )) X pk ) . (1) ∞Yk=0 ∞Yk=1 Moreover we define a formal power series Fp(v , λ; X, Y ) as follows: Fp(v , λ; X, Y ) = . (2) 1 + λpk(X + Y + λXY )pk! 1 ∞Yk=1 (1 + λpkX pk)(1 + λpkY pk) pk λpk Φk−1(v ) As in [8, Lemma 2.16.] or [7, Lemma 4.9.], we see that the formal power series Fp(v , λ; X, Y ) is integral over Z(p). T. Sekiguchi and N. Suwa [8, Theorem 2.19.1.] have shown the following isomorphisms with the formal power series (1) and (2): W (A)F (λ) ∼−→ Hom(bG(λ),bGm,A); v 7→ Ep(v , λ; X), W (A)/F (λ) ∼−→ H 2 0 (bG(λ),bGm,A); w 7→ Fp(w , λ; X, Y ). Here H 2 0 (G, H) denotes the Hochschild cohomology group consisting of symmetric 2- cocycles of G with coefficients in H for formal group schemes G and H ([3, Chap. II.3 and Chap. III.6]). (3) (4) 5 4 Proof of Theorem 3 In this section we prove Theorem 3. Subsection 4.1 is a technical part in our proof. In Subsection 4.2 we complete our proof of Theorem 3. 4.1 Suppose that A is a ring. Let λ be an element of A and l a positive integer. Assume that pl−kλpk is divided by λpl for each integer 0 ≤ k ≤ l − 1. Put a := λ−plpl[λ] ∈ W (A). Lemma 1 With the above notations, we have Ker(F (λ) ◦ Ta ) = Ker(F (λpl )). Proof As a preparation, we calculate the components of b := pl[λ] ∈ W (A) by using the phantom map. For b = (b0, b1, . . .), we have b0 = plλ by Φ0(b) = Φ0(pl[λ]). Similarly, we have b1 = pl−1λp(1 − p(p−1)l). Put α1 := (1 − p(p−1)l). For k ≥ 2, the components of b is inductively given by bk = pl−kλpk where we put (1 − p(pk−1)l − p(pk−1−1)(l−1)αpk−1 1 − p(pk−2−1)(l−2)αpk−2 2 − · · · − pp−1αp k−1) αk := 1 − p(pk−1)l − k−1Xi=1 p(pk−i−1)(l−i)αpk−i i (k ≥ 2). Note that we have the congruences bk ≡ λpl Therefore b is stated by (mod p) if k = l and bk ≡ 0 (mod p) if k 6= l. (5) (6) b =pl[λ] = (plλ, pl−1λpα1, pl−2λp2 α2, . . . , λpl αl, p−1λpl+1 αl+1, . . .). (7) Moreover we also obtain the components of a = λ−pl b ∈ W (A). Next, we show the equality of Lemma 1. Ker(F (λpl )) ⊂ Ker(F (λ) ◦ Ta) is proved as follows. For x ∈ Ker(F (λpl )), we have Φk+1(x ) = λpl+k(p−1)Φk(x ) since F (x ) = [λpl(p−1)]·x . We must show F (λ) ◦ Ta (x ) = o. The claim is proved by induction on k. Put y := F (λ) ◦ Ta (x ). For y = (y0, y1, y2, . . .), we have y0 = Φ0(y ) = Φ0(F ◦ Ta (x )) − λp−1Φ0(Ta (x )) = (ap 0λpl(p−1) + pa1 − λp−1a0)Φ0(x ). 6 By components of a, we have λpl(p−1)ap 0 + pa1 − λp−1a0 = 0. Hence y0 = 0. Assume yj = 0 for 1 ≤ j ≤ k − 1. Then we have Φk−1(F (λ) ◦ Ta (x )) = o, i.e., Φk(Ta (x )) = λpk−1(p−1)Φk−1(Ta (x )). By using the phantom map and the relations (5), we have Φk(F (λ) ◦ Ta (x )) = Φk+1(Ta(x )) − λpk(p−1)λpk−1(p−1) · · · λp−1Φ0(Ta (x )) = λpl+k(p−1)λpl+k−1(p−1) · · · λpl(p−1)apk+1 + λpl+k−1(p−1)λpl+k−2(p−1) · · · λpl(p−1)papk = (λpl+k+1−pl = {pk+1ak+1 − plλk+1/λpl = plλpk+1 1 Φ0(x ) + · · · + pk+1ak+1Φ0(x ) − λpk+1−1a0Φ0(x ) apk 1 + · · · + pk+1ak+1 − λpk+1−1a0)Φ0(x ) (1 − p(pk+1−1)l − p(pk−1)(l−1)αpk 1 − · · · − p(p−1)(l−k)λpk+1 p(pk+1−i−1)(l−i)αpk+1−i apk+1 0 + pλpl+k−pl 0 Φ0(x ) αp k)}Φ0(x ) /λpl {αk+1 − (1 − p(pk+1−1)l − i )}Φ0(x ) = 0. kXi=1 Therefore, for x ∈ Ker(F (λpl )), we have F (λ) ◦ Ta (x ) = o. We consider the following diagram in order to prove the reverse inclusion: Ta W (A) W (A) ∆ W (A) × W (A) (F,−[λpl(p−1)]) W (A) × W (A) ❯ ❯ ❯ ❯ ❯ ❯ ❯ t′ a ❯ ❯ ❯ ❯ ❯ ❯ ❯ ❯ ❯ ❯ m W (A) W (A) × W (A) ❱ ❱ ❱ ❱ ❱ ❱ ❱ ❱ m ❱ ❱ ❱ ❱ ❱ ❱ ❱ ❱ ❱ ❱ ❱ ❱ F (λ) W (A), where homomorphisms m, ∆ and t′ a are defined by m :W (A) × W (A) → W (A); (x 1, x 2) 7→ x 1 + x 2, ∆ :W (A) → W (A) × W (A); x 7→ (x , x ) t′ a :W (A) × W (A) → W (A) × W (A); and (x 1, x 2) 7→ (Ta (p)(x 1), Tc(p) ◦ F (x 2) − F ◦ Tc(x 2) + [λp−1] ◦ Tc(x 2)). Here we put c := λ−pl+1pl[λ]. Note that the homomorphism t′ (Im(F )) × (Im(−[λpl(p−1)])). Hence we obtain a is well-defined over F (λ) ◦ Ta = m ◦ t′ a ◦ (F,−[λpl(p−1)]) ◦ ∆ and F (λpl ) = m ◦ (F,−[λpl(p−1)]) ◦ ∆. 7 / / * * o o By the above equalities, we have W (A)/Ker(F (λ) ◦ Ta) ≃ Im(F (λ) ◦ Ta) ⊂ Im(F (λpl )) ≃ W (A)/Ker(F (λpl Therefore, if x ∈ Ker(F (λ) ◦ Ta ), then we have x = o ∈ W (A)/Ker(F (λpl x ∈ Ker(F (λpl 4.2 )). Thus we obtain the result. )). )). Hence ✷ Let n be a positive integer. Suppose that A is a Z(p)/(pn)-algebra. Let λ be an element of A. For each integer 0 ≤ k ≤ l − 1, we assume that pl−kλpk is divided by λpl and that pl−kλpk/λpl is nilpotent. In particular, if λ = 0, we set pl−kλpk/λpl := 0. Let G(λ) be the deformation group scheme defined in Subsection 3.1 and bG(λ) the formal completion of G(λ) along the zero section. We consider the homomorphism Then we have ψ(l) : bG(λ) → bG(λpl {(1 + λX)pl ψ(l)(X) = λ−pl ); x 7→ λ−pl {(1 + λx)pl − 1}. − 1} = λ−pl pl−1Xk=1(cid:18)pl k(cid:19)λkX k + X pl . Note that ψ(l) is well-defined under our assumptions (even λ = 0). By the nilpotency of pl−kλpk/λpl, the class X is nilpotent ([2, Chap. 1, Ex. 2]). If λ = 0, we have X = 0. In particular, if pl−kλpk/λpl is divided by p, the nilpotency of p is used in the coordinate ring. Hence the kernel of ψ(l) has the equalities pl Nl := Ker(ψ(l)) = Spf(A[[X]]/(ψ(l)(X))) = Spec(A[X]/(ψ(l)(X))). Note that Nl is a finite group scheme of order pl of G(λ). The following short exact sequence (8) is induced by ψ(l): ψ(l) where ι is a canonical inclusion. The exact sequence (8) deduces the long exact sequence (ψ(l))∗ (ψ(l))∗ ) −−−→ 0, 0 −−−→ Nl −−−→ bG(λpl ι−−−→ bG(λ) ),bGm,A) −−−−→ Hom(bG(λ),bGm,A) ),bGm,A) −−−−→ Ext1(bG(λ),bGm,A) −−−→ · · · 0 (bG(λpl ),bGm,A) with H 2 −−−−→ Hom(bG(λ),bGm,A) ),bGm,A) ([1, Lemma 3]). Therefore the exact ),bGm,A). −−−→ Hom(Nl,bGm,A) −−−→ Hom(Nl,bGm,A) 0 (bG(λpl ∂−−−→ H 2 (ι)∗ 8 0 −−−→ Hom(bG(λpl ∂−−−→ Ext1(bG(λpl can replace Ext1(bG(λpl Hom(bG(λpl ),bGm,A) sequence (9) is given by (ψ(l))∗ Since the image of the boundary map ∂ is given by direct product of formal schemes, we (ι)∗ . On the other hand, we show that the following sequence (10) is exact: W (A)F (λpl ) Ta−−−→ W (A)F (λ) π−−−→ Ml ∂−−−→ 0, where we put Ml := Ker[F (λ) : W (A)/Ta → W (A)/T ′ a] and π is a homomorphism induced by the natural projection W (A) ։ W (A)/Ta. We show that Im(Ta) = Ker(π) and Im(π) = Ml. Im(Ta ) ⊂ Ker(π) is obvious. To prove the reverse inclusion, if π(x ) = o ∈ Ml (x ∈ W (A)F (λ)), then we have x ∈ Im(Ta ). Hence x = Ta (z ) (z ∈ W (A)). Then we have F (λ)(x ) = F (λ) ◦ Ta (z ) = o. Therefore we have z ∈ Ker(F (λ) ◦ Ta) = Ker(F (λpl )) = W (A)F (λpl . ) Next, we prove the surjectivity of π. Let x (6= 0) ∈ Ml. Hence x /∈ Im(Ta ). Since F (λ)(x ) = F (λ)(x ) = 0 and F (λ)(x ) 6= F (λ) ◦ Ta(z ) (z ∈ W (A)), we have F (λ)(x ) /∈ a ) = Im(F (λ) ◦ Ta) and F (λ)(x ) = o. Hence x ∈ W (A)F (λ). Therefore π is surjective, Im(T ′ i.e., we have W (A)F (λ)/Im(Ta ) ≃ Ml. Now, by combining the exact sequences (9), (10) and the isomorphisms (3), (4), we have the following diagram (11) consisting of exact horizontal lines and vertical isomor- phisms except φ: ),bGm,A) Hom(bG(λpl φ1x W (A)F (λpl ) (ψ(l))∗ −−−−→ Hom(bG(λ),bGm,A) Ta−−−→ φ2x W (A)F (λ) (ι)∗ −−−→ Hom(Nl,bGm,A) π−−−→ φx Ml ∂−−−→ H 2 0 (bG(λpl ),bGm,A) φ3x ∂−−−→ W (A)/F (λpl ), where φ is the following homomorphism induced from the exact sequence (8) and the isomorphism (3): φ : Ml → Hom(Nl,bGm,A); x 7→ Ep(x , λ; x) := Ep(x , λ; x). ). By z ∈ W (A)F (λpl We must show the well-definedness of φ. For x ∈ Ml, we choose an inverse image x + Ta (z ) ∈ W (A), where x ∈ W (A)F (λ) and z ∈ W (A)F (λpl ), we can use the equality Ep(z , λpl; ψ(l)(x)) = Ep(Ta (z ), λ; x) ([1, Lemma 1, p.123]). Hence we have Ep(x , λ; x) = Ep(x , λ; x) · Ep(Ta (z ), λ; x) = Ep(x , λ; x) · Ep(z , λ; ψ(l)(x)) ≡ Ep(x , λ; x) (mod ψ(l)(x)). If the diagram (11) is commutative, by using the five lemma, φ becomes an isomor- phism, i.e., Ml ≃ Hom(Nl,bGm,A). Therefore we must prove the equalities (12) (ψ(l))∗ ◦ φ1 = φ2 ◦ Ta , (13) (ι)∗ ◦ φ2 = φ ◦ π, (14) ∂ ◦ φ = φ3 ◦ ∂. 9 For (12), we must show the equality Ep(x , λpl; ψ(l)(x)) = Ep(Ta(x ), λ; x). This is nothing but the equality in [1, Lemma 1, p.123]. The equality of (13) follows from the definition of φ. The calculation of the boundary ∂(Ep(x , λ; x)) (x ∈ Ml) is similar to the previous paper [1, Lemma 3]. Hence we have ∂(Ep(x , λ; x)) = Fp(F (λ)(x + Ta(z )), λ; x1, x2), where x + Ta (z ) is an inverse image of x for π : W (A) → W (A)/Ta. Note that z ∈ W (A)F (λpl ). Since z ∈ W (A)F (λpl ) = Ker(F (λ) ◦ Ta ), we have Fp(F (λ)(x + Ta(z )), λ; x1, x2) = Fp(F (λ)(x ) + F (λ) ◦ Ta (z ), λ; x1, x2) = Fp(o, λ; x1, x2) = 1. Therefore the equality (14) is a conclusion from ∂(Ep(x , λ; x)) = 1 and ∂(Ml) = {o}. Hence we obtain Theorem 3. References [1] M. Amano, On the Cartier duality of certain finite group schemes of order pn, Tokyo J. Math. 33 (2010), 117 -- 127. [2] M. F. Atiyah and I. G. Macdonald, Introduction to commutative algebra, Addison- Wesley, Reading, Mass, 1969. [3] M. Demazure and P. Gabriel, Groupes alg´ebriques, Tome I, Masson-North-Holland, Paris-Amsterdam, 1970. [4] M. Hazewinkel, Formal groups and applications, Academic Press, New York, 1978. [5] F. Oort, T. Sekiguchi, and N. Suwa, On the deformation of Artin-Schreier to Kum- mer, Ann. Sci. ´Ecole Norm. Sup. 22 (1989), no. 4, 345 -- 375. [6] F. Oort and J. Tate, Group schemes of prime order, Ann. Sci. ´Ecole Norm. Sup. 4 (1970), no. 3, 1 -- 21. [7] T. Sekiguchi and N. Suwa, On the unified Kummer-Artin-Schreier-Witt theory, Uni- versit´e de Bordeaux (1999), no. 111. [8] , A note on extensions of algebraic and formal groups IV, Tohoku Math. J 53 (2001), 203 -- 240. [9] Y. Tsuno, Degeneration of the Kummer sequence in characteristic p > 0, Journal de th´eorie des nombres de Bordeaux 22 (2010), 219 -- 257. [10] W. Waterhouse and B. Weisfeiler, One-dimensional affine group schemes, Journal of Algebra 66 (1980), 555 -- 568. 10 Corrigendum to "On the Cartier duality of certain finite group schemes of order pn, II" There is an error in the proof of Lemma 1, which is amended as follows. On Page 7, line −1, it is claimed that the diagram there given were commutative. But it is false. The only consequence of this wrong claim that is used in the subsequent argument is the inclusion Ker(F (λ) ◦ Ta ) ⊂ Ker(F (λpℓ )). See Page 8, line 2. Therefore, one has only to reprove this inclusion. Suppose x ∈ Ker(F (λ) ◦ Ta ), or equivalently, Φk+1(Ta(x )) = λpk(p−1)Φk(Ta (x )), k ≥ 0. See Page 7, line 4. To show x ∈ Ker(F (λpℓ )), we wish to prove the equivalent (Φk(F (λpℓ )(x )) =)Φk+1(x ) − λpℓ+k(p−1)Φk(x ) = 0, k ≥ 0 (C1) (C2) by induction on k. Suppose k = 0. The desired equality then follows by direct computation using Eqs. (5) and (7) on Page 6. Suppose k > 0. The induction hypothesis Φi(F (λpℓ implies )(x )) = 0, 0 ≤ i < k, immediately Φi+1(x ) = λpℓ(pi+1−1)Φ0(x ), 0 ≤ i < k. (C3) Using (5) again, we compute Φk+1(Ta (x )) = apk+1 = apk+1 0 Φk+1(x ) + papk 0 Φk+1(x ) +(cid:16)pℓλpk+1 kXi=1 + 1 − p(pk+1−1)ℓ − 0 Φk+1(x ) +n(cid:16)pℓλpk+1 1 Φk(x ) + · · · + pk+1ak+1Φ0(x ) k(cid:17) /λpℓ(cid:17)n(cid:16)p(ℓ−1)(pk −1)αpk 1 + · · · + p(ℓ−k)(p−1)αp oΦ0(x ) /λpℓ(cid:17)o Φ0(x ). /λpℓ(cid:17) −(cid:16)pℓpk+1 p(pk+1−i−1)(ℓ−i)αpk+1−i i = apk+1 λpk+1 Similarly we have Φk(Ta (x )) =(cid:16)pℓλpk/λpℓ(cid:17) Φ0(x ). The last two results, combined with There is a misprint. On page 9, line −5 should read "Ep(z , λpℓ; ψ(ℓ)(x))" instead of (C1), show that the equality (C3) holds when i = k, as well. The equalities (C3) for i = k − 1, k show the desired (C2). "Ep(z , λ; ψ(ℓ)(x))." 11
1912.11895	1	1912	2019-12-26T16:33:51	Positive Populations	[ "math.AG", "math.CO", "math.RA" ]	A positive structure on the varieties of critical points of master functions for KZ equations is introduced. It comes as a combination of the ideas from classical works by G.Lusztig and a previous work by E.Mukhin and the second named author.	math.AG	math	POSITIVE POPULATIONS VADIM SCHECHTMAN ◦ AND ALEXANDER VARCHENKO ⋆ ◦ Institut de Math´ematiques de Toulouse -- Universit´e Paul Sabatier 118 Route de Narbonne, 31062 Toulouse, France ⋆Department of Mathematics, University of North Carolina at Chapel Hill Chapel Hill, NC 27599-3250, USA ⋆Faculty of Mathematics and Mechanics, Lomonosov Moscow State University Leninskiye Gory 1, 119991 Moscow GSP-1, Russia Abstract. A positive structure on the varieties of critical points of master functions for KZ equations is introduced. It comes as a combination of the ideas from classical works by G. Lusztig and a previous work by E. Mukhin and the second named author. Keywords: Totally positive matrices, Whitney-Lusztig charts, Wronskian differential equation, Wronski map, master functions, Bethe cells, positive populations 2010 Mathematics Subject Classification: 13F60 (14M15, 82B23) 9 1 0 2 c e D 6 2 ] . G A h t a m [ 1 v 5 9 8 1 1 . 2 1 9 1 : v i X r a Contents Introduction: Whitney-Lusztig patterns and Bethe populations 1. 1.1. Big cell and critical points 1.2. What is done in Introduction 1.3. Whitney-Loewner charts and Lusztig transition maps 1.4. Bethe Ansatz equations 1.5. Reproduction, or Wronskian evolution (bootstrap) 1.6. Wronskian bootstrap: the details 1.7. Example, [MV, Section 3.5] 1.8. Not necessarily reduced words 1.9. Main point 1.10. Contents of the paper 2. Generalities on Wronskians 2.1. Wronskian differential equation 2.2. Univaluedness 2.3. Wronskian 2.4. W5 Identity 3. Normalized Wronskian bootstrap 3.1. Generic and fertile tuples 3.2. Normalized mutations 3.3. Normalized population related to a word ◦E -mail: [email protected]. ⋆E -mail: [email protected] , supported in part by NSF grant DMS-1665239 2 2 2 3 4 5 6 7 8 8 8 9 9 9 9 10 10 10 10 11 2 VADIM SCHECHTMAN AND ALEXANDER VARCHENKO 4. Whitney-Lusztig charts and the comparison theorem 4.1. Birational isomorphisms 4.2. Comparison Theorem 4.3. Cell N 4.4. N -Y correspondence 5. Triangular coordinates on the Bethe cell 5.1. Example of evolutions, group SL3 5.2. Example of evolutions, group SL4 5.3. Triangular coordinates 5.4. Polynomials ϕij 6. Bethe cells from subspaces of C[x] and from critical points 6.1. From a vector space of polynomials to a population ZV , see [MV] 6.2. From a critical point to a population ZV , [MV] 6.3. Bethe cells associated with (V ; z) 6.4. Normalized mutations and N -Y correspondence 6.5. Positive populations 6.6. Coordinates on the Bethe cell 7. Base affine space and fat populations 7.1. Base affine space 7.2. Fat population 7.3. Fat Bethe cell 7.4. Example of a cluster structure on a fat Bethe cell 8. Appendix: Fourteen and Eightfold Ways 8.1. Group SL3 and a 2-category 8.2. Group SL4 and a tetrahedron equation References 12 12 12 13 13 14 14 16 19 20 20 20 22 23 24 25 25 26 26 26 27 27 27 27 28 31 1. Introduction: Whitney-Lusztig patterns and Bethe populations 1.1. Big cell and critical points. The aim of the present note is to introduce a positive structure on varieties of critical points of master functions arising in the integral representation for solutions of KZ equations. Let N = Nr+1 ⊂ G = SLr+1(C) denote the group of upper triangular matrices with 1's on the diagonal. It may also be considered as a big cell in the flag variety SLr+1(C)/B−, where B− is the subgroup of lower triangular matrices. Let Sr+1 denote the Weyl group of G, the symmetric group. In this note two objects related to N will be discussed: on the one hand, what we call here the Whitney-Loewner-Lusztig data on N, on the other hand, a construction, introduced in [MV], which we call here the Wronskian evolution along the varieties of critical points. In this note we consider only the case of the group SLr+1(C), although the other reductive groups can be considered similarly. 1.2. What is done in Introduction. In Section 1.3 we recall the classical objects: Whitney- Loewner charts, these are collections of birational coordinate systems on N indexed by reduced decompositions of the longest element w0 ∈ Sr+1, and Lusztig transition maps between them. POSITIVE POPULATIONS 3 In Sections 1.4 - 1.8 the main ideas from [MV] are introduced. Namely, it is a reproduction recipe, called here a Wronskian evolution, which produces varieties of critical points for master functions appearing in integral representations for solutions of KZ equations, [SV]. In Section 1.10 the content of Sections 2 - 8 is described. 1.3. Whitney-Loewner charts and Lusztig transition maps. In the seminal papers [L, BFZ] Lusztig and Berenstein-Fomin-Zelevinsky have performed a deep study of certain remark- able coordinate systems on N, i.e. morphisms of algebraic varieties Lh : Cq −→ N, q = r(r + 1)/2, with dense image, which are birational isomorphisms. The main feature of these morphisms is that the restriction of them to Rq >0 induces an isomorphism Lh : Rq >0 ∼−→ N>0, where N>0 is the subspace of totally positive upper triangular matrices. Recall that a matrix g ∈ N is called totally positive if all its minors are strictly positive, except for those who are identically zero on the whole group N, see [BFZ]1. The set of such coordinate systems, which we will be calling the Whitney-Loewner charts, is in bijection with the set Red(w0) of reduced decompositions (1.1) h : w0 = siq . . . si1 of the longest element w0 ∈ Sr+1. For example, for r = 2 there are two such coordinate systems, L121 and L212 corresponding to the reduced words s1s2s1 and s2s1s2 respectively. The construction of maps Lh will be recalled below, see Section 4.1. To every h ∈ Red(w0) and a = (aq, . . . , a1) ∈ Cq there corresponds a matrix Nh(a) = Lh(a) ∈ N. A theorem of A. Whitney2, as reformulated by Ch. Loewner, see [W, Lo], says; Theorem 1.1. For every h each matrix A ∈ N>0 is of the form Nh(a) for some a ∈ Rq >0. For any two words h, h′ Lusztig has defined a birational self-map of Aq, i.e. an automorphism of the field of rational functions F := C(a) = C(a1, . . . , aq) ∼= C(N) (here we consider ai as independent transcendental generators), (1.2) such that Rh,h′ : F ∼−→ F , a′ = Rh,h′(a), if Nh(a) = Nh′(a′) . 1The notion of a totally positive matrix first appeared in the works of I. Schoenberg [S] and Gantmacher-Krein [GK]. 2Anne M. Whitney (1921 -- 2008), a student of Isaac Schoenberg (1903 -- 1990). 4 VADIM SCHECHTMAN AND ALEXANDER VARCHENKO For example N121(a1, a2, a3) =   3) =   1, a′ and 1 a′ 2 0 1 0 0 where e1(a) = 1 + ae12, e2(a) = 1 + ae23. N212(a′ 2, a′ a′ 2a′ 3 1 + a′ a′ 3 1 1 a1 + a3 a1a2 0 a2 1 0 1 0   = e1(a1)e2(a2)e1(a3)   = e2(a′ 1)e1(a′ 2(a′ 2)e′ 3) , It follows that provided (1.3) This is equivalent to (1.4) N121(a1, a2, a3) = N212(a′ 3) , 1, a′ 2, a′ a′ 1 = a2a3 a1 + a3 a1 = 2a′ a′ 3 a′ 1 + a′ 3 , , a′ 2 = a1 + a3 , a′ 3 = a2 = a′ 1 + a′ 3 , a3 = a2a1 a1 + a3 2a′ a′ 1 a′ 1 + a′ 3 . . The transformation (1.3) is involutive, that is, its square is the identity. 1.4. Bethe Ansatz equations. On the other hand, in the work [MV] it was discovered that the variety N is closely connected with the varieties of critical points of certain master functions Φk. Namely, for a sequence k = (k1, . . . , kr) ∈ Nr , consider a function Φk(u) depending on k := Pr 1 , . . . , u(1) u = (u(1) k1 ; . . . ; u(r) 1 , . . . , u(r) kr ) . i=1 ki variables subdivided into r groups: By definition, Φk(u) = r Yi=1 Y16m<l6ki (u(i) m − u(i) l )aii · Y16i<j6r ki Ym=1 kj (u(i) m − u(j) l )aij . Yl=1 Here A = (aij) is the Cartan matrix for the root system of type Ar, in other words, Φk(u) = r Yi=1 Y16l<m6ki (u(i) l − u(i) m )2 · r−1 Yi=1 ki Yl=1 ki+1 Ym=1 (u(i) l − u(i+1) m )−1 . Functions of this kind first appeared in [SV] in the study of integral representations for solutions of KZ differential equations. A point t = (t(1) 1 , . . . , t(1) k1 ; . . . ; t(r), . . . , t(r) kr ) is critical for the function log Φk(u) if it satisfies the system of k equations ∂ log Φk ∂u(i) m (t) = (cid:20) ∂Φk ∂u(i) m Φ−1 k (cid:21)(t) = 0 , 1 6 i 6 r, 1 6 m 6 ki , POSITIVE POPULATIONS 5 or, equivalently, (1.5) Xl6=m aii m − t(i) t(i) l kj +Xj6=i Xl=1 aij m − t(j) t(i) l = 0 , 1 6 i 6 r, 1 6 m 6 ki . This system of critical point equations is also called the system of Bethe Ansatz equations in the Gaudin model. 1.5. Reproduction, or Wronskian evolution (bootstrap). The following procedure of re- production for constructing critical points has been proposed in [ScV], [MV]. Let us identify the group Zr with the root lattice Q of the group G using the base of standard simple roots α1, . . . , αr. Introduce the usual shifted action of W = Sr+1 on Q: where ρ is the half-sum of positive roots. Let w ∈ Sr+1 and let w ∗ v = w(v − ρ) + ρ , (1.6) h : w = sim . . . si1 be a reduced decomposition of w. For any 0 6 j 6 q we define an r-tuple k(j) = sij . . . si1 ∗ (0) ∈ Nr , where 0 = (0, . . . , 0) = k(0). Starting from the r-tuple of polynomials one defines inductively a sequence of r-tuples of polynomials y(0) = (1, . . . , 1) ∈ C[x]r , (1.7) where 0 6 j 6 m, with yh = (y(0), y(1)(v1), y(2)(v1, v2), . . . , y(m)(v1, . . . , vm)) , y(j)(v) = (y(j) 1 (v; x), . . . , y(j) r (v; x)) ∈ C[v1, . . . , vj; x]r, deg y(j)(v) := (deg y(j) 1 (v; x), . . . , deg y(j) r (v; x)) = k(j) , where deg is the degree with respect to x. The sequence (1.4.2) is called the population associated with a reduced word h. We consider a polynomial y(j) i (v; x) as a family y(j) depending on a parameter c = c(j) = (c1, . . . , cj) ∈ Cj. i (c; x) of polynomials of one variable x Let c ∈ Cj and ti,1, . . . , ti,k(j) i be the roots of y(j) i (c; x) ordered in any way. Consider the tuple t(j)(c) = (t1,1, . . . , t1,k(j) 1 ; . . . ; tr,1, . . . , tr,k(j) r ) . The main property of the sequence yh is: 6 VADIM SCHECHTMAN AND ALEXANDER VARCHENKO Theorem 1.2 ([MV]). For each j = 1, . . . , m, there exists a Zarisky open dense subspace U ⊂ Cj such that for every c = c(j) ∈ U , the tuple of roots t(j)(c) is a critical point of the master function Φ(j)(u) := Φk(j)(u), i.e. it satisfies the Bethe Ansatz equations (1.5). Moreover, if Φk is a master function for some index k and t a critical point of log Φk as in Section 1.4, then t appears in this construction for a reduced decomposition h of some element w ∈ Sr+1. The construction of yh for h ∈ Red(w0) see in Section 1.6 below. 1.6. Wronskian bootstrap: the details. Starting from the r-tuple of polynomials one constructs the sequence y(0) = (1, . . . , 1) ∈ C[x]r , yh = (y(0), y(1)(v1), y(2)(v1, v2), . . . , y(m)(v1, . . . , vm)) (1.5.1) by induction. Assume that the sequence y(j)(v) = (y(j) 1 (v; x), . . . , y(j) r (v; x)) ∈ C[v1, . . . , vj; x]r has been constructed. Then the sequence y(j+1)(v) = (y(j+1) 1 (v; x), . . . , y(j+1) r (v; x)) ∈ C[v1, . . . , vj, vj+1; x]r is such that y(j+1) i (v; x) = y(j) i (v; x), ∀i 6= ij+1 . and y(j+1) ij+1 (v1, . . . , vj, vj+1; x) is constructed in two steps. First one shows that there is a unique polynomial y(v1, . . . , vj; x) such that (i) Wr(y(j) ij+1, y) = const y(j) ij+1−1y(j) ij+1+1 , where Wr(f, g) = f g′ − f ′g denotes the Wronskian of two functions in x and the constant does not depend on v, x; (ii) the polynomial y(v1, . . . , vj; x) is monic with respect to the variable x; (iii) the coefficient in y(v1, . . . , vj; x) of the monomial x k(j) ij+1 equals zero, where k(j) ij+1 is the ij+1-st coordinate of the vector k(j). Then we define y(j+1) ij+1 (v1, . . . , vj, vj+1; x) := y(v1, . . . , vj; x) + vj+1 y(j) ij+1(v1, . . . , vj; x) . Consider all coordinates of the resulting family y(m)(v1, . . . , vm) = (y(m) 1 (v; x), . . . , y(m) r (v; x)) ∈ C[v1, . . . , vm; x]r up to multiplication by nonzero numbers. This gives a map Fh : Cm → P(C[x])r , where P(C[x]) is the projective space associated with C[x]. Denote by Zh = Fh(Cm) ⊂ P(C[x])r its image. POSITIVE POPULATIONS 7 Let V be an r + 1-dimensional complex vector space, X = G/B− the space of all complete flags in V . Let F0 ∈ X be a point. The choice of F0 gives rise to a decomposition of X into W = (r + 1)! Bruhat cells, which are in bijection with W = Sr+1: We have X = aw∈W Xw . dim Xw = ℓ(w). For example, for the identity e ∈ W , the cell Xe = {F0} is the zero-dimensional cell, and for the longest element w0 ∈ W , the cell Xw0 is the open cell, the space of all flags in general position with F0. Theorem 1.3 ([MV]). The union Z = [w,h Zh ⊂ P(C[x])r over all reduced decompositions h of all elements w ∈ Sr+1 is an algebraic subvariety of P(C[x])r isomorphic to the variety of complete flags X = G/B−. The subspace Zh ⊂ Z does not depend on a choice of h ∈ Red(w), so it may be denoted by (cid:3) Zw. It is identified with the Bruhat cell Xw ⊂ X. In particular, the subset Zw0 The algebraic subvariety Z ⊂ P(C[x])r is what was called in [MV] the population of critical ∼= Xw0 is isomorphic to the big Bruhat cell N ⊂ G/B−. points originated from y(0). The proof of Theorem 1.3 in [MV] identifies the variety Z with the space of complete flags of a particular r + 1-dimensional vector space V , where V = Vr = C[x]6r ⊂ C[x] is the vector space of polynomials of degree 6 r, and the flag F0 is the standard complete flag F0 = (V0 ⊂ . . . ⊂ Vr−1 ⊂ Vr), Vi = C[x]6i . 1.7. Example, [MV, Section 3.5]. For r = 2, let us see how the populations give rise to a decomposition of X = SL3(C)/B− into six Bruhat cells. We have the zero-dimensional Bruhat cell We have two one-dimensional Bruhat cells: Zid = {(1 : 1)} ∈ P(C[x])2 . Zs1 = {(x + c1 : 1) c1 ∈ C} ⊂ P(C[x])2 , 1 ∈ C} ⊂ P(C[x])2 , Zs2 = {(1 : x + c′ 1) c′ two two-dimensional Bruhat cells: Zs2s1 = {(x + c1 : x2 + 2c1x + c2) c1, c2 ∈ C} ⊂ P(C[x])2 , 2 ∈ C} ⊂ P(C[x])2 . Zs1s2 = {(x2 + 2c′ 2 : x + c′ 1x + c′ 1) c′ 1, c′ 8 VADIM SCHECHTMAN AND ALEXANDER VARCHENKO The longest element w0 ∈ S3 has two reduced decompositions s1s2s1 and s2s1s2, which give two parametrizations of the same three-dimensional Bruhat cell: Zs1s2s1 = {(x2 + c3x + c1c3 − c2 : x2 + 2c1x + c2) c1, c2, c3 ∈ C} ⊂ P(C[x])2 , 3 ∈ C} ⊂ P(C[x])2 . Zs2s1s2 = {(x2 + 2c′ 2 : x2 + c′ 1x + c′ 3 − c′ The two coordinate systems are related by the equations 3x + c′ 2) c′ 2, c′ 1c′ 1, c′ (1.8) c′ 1 = c3/2 , c′ 2 = c1c3 − c2 , c′ 3 = 2c1 . Notice that this transformation is involutive. The union of the six Bruhat cells gives the subvariety Z ⊂ P(C[x])2 isomorphic to SL3(C)/B−. The subvariety Z consists of all pairs of quadratic polynomials (a2x2 + a1x + a0 : b2x2 + b1x + b0) such that (1.9) a2b0 − 1 2 a1b1 + a0b2 = 0 . 1.8. Not necessarily reduced words. One can associate to an arbitrary, not necessarily re- duced word h of length m a sequence of m r-tuples (1.3) as well, see [MV]. Namely, using the Wronskian differential equation W (y(j) ij+1, y) = const · y(j) ij+1−1y(j) ij+1+1 as in (i) above alone, but without normalizing conditions (ii) and (iii) one gets for each j = 1, . . . , m a tuple y(j)(v) = (y(j) 1 (v; x) : . . . : y(j) r (v; x)) ∈ P (C[x])r , where v belongs to a variety of parameters B(j), which is an iterated C-torsor. This means that B(j) is included into a sequence of fibrations B(j) −→ B(j−1) −→ . . . −→ B(1) ∼= C , where each step B(p) −→ B(p−1) is an analytic C-torsor, locally trivial in the usual topology. But the corresponding cohomology H 1(Cp; C) vanishes which implies that the torsors are trivial, and this provides a global isomorphism B(j) ∼= Cj. 1.9. Main point. The main new point of the present work is a definition of a certain modified reproduction, which we call the normalized reproduction. It provides the varieties of r-tuples of polynomials Y Bethe, to be called the Bethe cells, equipped with a system of coordinate charts isomorphic to N equipped with the Whitney-Lusztig charts. We also define the totally positive subspace Y Bethe >0 ⊂ Y Bethe isomorphic to the subspace N>0 ⊂ N of totally positive upper triangular matrices. 1.10. Contents of the paper. Section 2 contains some preparations. In Section 3 a modifica- tion of the mutation procedure of Section 1.6 is introduced. Based on it, in Section 4 the Bethe cell is defined, and Comparison Theorem 4.4 is proven, which is one of our main results. In Section 5 we first describe in full detail Wronsky evolution for the case of groups SL3 and SL4. In particular we compute explicitly the positive part Y Bethe of the Bethe cell, see Theorem 5.2. Af- terwards we prove Triangular Theorem 5.3, which establishes an explicit isomorphism between N and the Bethe cell. Section 6 contains a generalization of the previous constructions. Namely, we define the Wronskian evolution, the Bethe cell, etc. associated with a finite-dimensional subspace V ⊂ C[x] and a distinguished complete flag F0 in V . The previous consideration corresponded >0 POSITIVE POPULATIONS 9 the subspace C[x]6r of polynomials of degree 6 r. In Section 7 we present a version of the above considerations for the base affine space. Namely, we define a variety Y Bethe, which is related to the previous Bethe variety Y Bethe in the same way as the big cell in the base affine space G/N− is related to the big cell in the flag space G/B−. In Section 8 we interpret the Whitney-Lusztig data in the language of higher Bruhat orders, [MS], in particular give its complete description in the crucial case of SL4. We are grateful to E. Mukhin and M. Shapiro for useful discussions. 2. Generalities on Wronskians 2.1. Wronskian differential equation. The Wronskian of two functions f (x), g(x) is the func- tion Wr(f, g) = f g′ − f ′g = f 2(g/f )′ . Given f (x) and h(x), the equation (2.1) Wr(f, g) = h with respect to the function g(x) has a solution g(x) = f (x)Z h(x)f (x)−2dx . The general solution is (2.2) g(x, c) = f (x)Z x x0 h(t)f (t)−2dt + cf (x), c ∈ C . 2.2. Univaluedness. Let f (x), g(x) ∈ C[x] be polynomials. Then h(x) := Wr(f, g) is a polyno- mial. Hence the indefinite integral of a rational function, Z h(x)f (x)−2dx , has no logarithmic terms, which is equivalent to the condition: (U) The function hf −2 has zero residues at its poles. 2.3. Wronskian. Let f1(x), . . . , fn(x) be holomorphic functions. Define the Wronskian matrix (f (a−1) )n a,b=1 and the Wronskian b The Wronskian is a polylinear skew-symmetric function of f1, . . . , fn. Wr(f1, . . . , fn) = det (f (a−1) b )n a,b=1 . Example 2.1. We have More generally, (2.3) Wr(cid:16)1, x, x2 2 , . . . , xn n!(cid:17) = 1 . Wr(xd1, . . . , xdk) = Yi<j (dj − di) · xP di−k(k+1)/2, see this formula in [MTV]. 10 VADIM SCHECHTMAN AND ALEXANDER VARCHENKO 2.4. W5 Identity. Let f1(x), f2(x), . . . be a sequence of holomorphic functions. For an ordered finite subset we write Denote A = {i1, . . . , ia} ⊂ N := {1, 2, . . .}, Wr(A) = Wr(fi1, . . . , fia) . [a] = {1, 2, . . . , a}. Proposition 2.1 (W5 Identity, [MV, Section 9]). Let A = [a + 1], B = [a] ∪ {a + 2}. Then (2.4) Wr(Wr(A), Wr(B)) = Wr(A ∩ B) · Wr(A ∪ B). Example 2.2. We have (2.5) Wr(Wr(f1, f2), Wr(f1, f3)) = f1 Wr(f1, f2, f3), as one can easily check. 3. Normalized Wronskian bootstrap 3.1. Generic and fertile tuples. Let y = (y1(x), . . . , yr(x)) ∈ C[x]r be a tuple of polynomials. Define y0 = yr+1 = 1. We say that the r-tuple y is fertile, if for every i the equation Wr(yi(x), yi(x)) = yi−1(x) yi+1(x) with respect to yi(x) admits a polynomial solution. We say that y is generic, if for every i the polynomial yi(x) has no multiple roots and the polynomials yi(x) and yi−1(x) yi+1(x) have no common roots. Let ti,1, . . . , ti,ki be the roots of yi(x) ordered in any way. Consider the tuple t = (t1,1, . . . , t1,k1; . . . ; tr,1, . . . , tr,kr) . Lemma 3.1 ([MV]). The tuple y is generic and fertile if and only if the tuple t is a critical point of the master function Φk, where k = (deg y1, . . . , deg yr). 3.2. Normalized mutations. Let y = (y1, . . . , yr) ∈ C[x]r be an r-tuple of polynomials such (3.1) yi(0) = 1 , i = 1, . . . , r . Lemma 3.2. There exists a unique solutions yi(x) of the differential equation (3.2) such that (3.3) Wr(yi(x), yi(x)) = yi−1(x) yi+1(x) , yi(x) = x + O(x2) as x → 0 . Proof. We have (3.4) POSITIVE POPULATIONS yi(x) = yi(x)Z x 0 yi−1(u)yi+1(u) y2 i (u) du . 11 (cid:3) Assume that the tuple y is generic and fertile, then the function yi(x) is a polynomial by Lemma 3.1. The polynomial yi(x) can be determined either by formula (3.4) or by the method of undetermined coefficients, see examples in Section 5. For c ∈ C, denote (3.5) yi(c; x) = yi(x) + cyi(x) . Notice that for any c we have yi(c; 0) = 1. Define a new r-tuple of polynomials (3.6) νi(c)y := (y1(x), . . . , yi−1(x), yi(c; x), yi+1(x), . . . , yr(x)) and call it the i-th normalized mutation of the fertile generic tuple y. Equation (3.2) with the normalizing condition (3.3) will be called the normalized Wronskian evolution, or bootstrap, equation. 3.3. Normalized population related to a word. Let (3.7) h = sim . . . si1 be any word in Sr+1. Sometimes we will write for brevity simply (3.8) instead of (3.7). h = (im . . . i1) Now we proceed as in Section 1.5, but will use the normalized mutations νi. Namely, we start with y(0) = (1, . . . , 1) ∈ C[x]r and for each c = (c1, . . . , cm) ∈ Cm define an r-tuples of polynomials by the formula: (3.9) yh(c) = (yh,1(c; x), . . . , yh,r(c; x)) := νim(cm) . . . νi1(c1)y(0) . Notice that for any i = 1, . . . , r and c ∈ Cm we have (3.10) yh,i(c; 0) = 1 . Theorem 3.3 ([MV]). For any h and c ∈ Cm, the tuple yh(c) is a tuple of polynomials. Moreover, there exists a Zarisky open dense subspace U ⊂ Cm such that for every c ∈ U , the tuple of roots th(c) of the polynomials (yh,1(c; x), . . . , yh,r(c; x)) is a critical point of the corresponding master function Φk. 12 VADIM SCHECHTMAN AND ALEXANDER VARCHENKO 4. Whitney-Lusztig charts and the comparison theorem 4.1. Birational isomorphisms. Recall the group N from Section 1.1. Let eij ∈ glr+1(C) denote the elementary matrix, (eij)ab = δiaδjb . Define the matrices (4.1) ei(c) = 1 + cei,i+1 ∈ N , i = 1, . . . , r , c ∈ C . Given a word h = sim . . . si1, define a map (4.2) Lh : Cm −→ N , (c1, . . . cm) 7→ eip(cm) . . . ei1(c1) . Suppose that the word h is a reduced decomposition of the longest element w0 as in (1.1). Then h is of length q = r(r + 1)/2. The corresponding map (4.3) Lh : Cq −→ N is called the Whitney-Lusztig chart corresponding to h. The map Lh is a birational isomorphism. Remark. The map Lh is not epimorphic, and N is not even equal to the union of the images of Lh for h ∈ Red(w0). For example for r = 2 the set Red(w0) has two elements, the words (121) and (212). It is easy to see that the matrices of the form 1 + ae13, a 6= 0, are inaccessible, 1 + ae13 /∈ L121(C3) ∪ L212(C3) . 4.2. Comparison Theorem. Denote M = C[x]r+1. We will write the elements of M as se- quences (f1, . . . , fr+1) but will think of them as column vectors. Then the algebra glr+1(C) of (r + 1) × (r + 1)-matrices acts on M from the left in the standard way. We introduce a distinguished element (4.4) b(0) = (cid:16)1, x, x2 2 , . . . , xr r!(cid:17). For a word h = sim . . . si1 in Sr+1 and c = (c1, . . . , cm) ∈ Cm, define an element bh(c) = (bh,1(c), . . . , bh,r+1(c)) ∈ M , by the formula (4.5) bh(c) := Lh(c1, . . . cm)b(0) . Recall the r-tuple of polynomials (4.6) yh(c) = (yh,1(c; x), . . . , yh,r(c; x)) = νim(cm) . . . νi1(c1)y(0) , defined in (3.9). Theorem 4.1 (Comparison Theorem). For any j = 1, . . . , r, we have Wr(bh,1(c; x), . . . , bh,j(c; x)) = yh,j(c; x) . Proof. The theorem is a consequence of a general statement, see Theorem 4.3 below. (cid:3) POSITIVE POPULATIONS 13 4.3. Cell N . Denote the orbit of the element b(0) under the action of N. An element b = (b1, . . . , br+1) ∈ M belongs to N if and only if N = Nr := N b(0) ⊂ M , (4.7) for some bij ∈ C. bi = xi−1 (i − 1)! + r Xj=i bijxj, i = 1, . . . , r + 1, For any b = (b1, . . . , br+1) ∈ N we have (4.8) Wr(b1, . . . , br+1) = 1 . We define the totally positive subvariety N>0 ⊂ N as (4.9) N>0 = N>0b(0) . 4.4. N -Y correspondence. Define the Wronski map (4.10) W : N −→ C[x]r , b 7→ (b1, Wr(b1, b2), Wr(b1, b2, b3), . . . , Wr(b1, . . . , br)) . Lemma 4.2. If b = (b1, . . . , br+1) ∈ N and y = (y1, . . . , yr) = W (b), then (4.11) (4.12) yi = Wr(b1, . . . , bi) = 1 + O(x) , Wr(b1, . . . , bi−1, bi+1) = x + O(x2) , i = 1, . . . , r , i = 1, . . . , r − 1 , as x → 0. (cid:3) We define the Bethe cell or variety of Bethe r-tuples as (4.13) Y Bethe := W (N ) ⊂ C[x]r , the image of the Wronski map, and the totally positive Bethe subvariety or positive population as Y Bethe >0 := W (N>0) ⊂ W (N ) = Y Bethe . Theorem 4.3. The Wronski map induces an isomorphism W : N ∼−→ Y Bethe , N>0 ∼−→ Y Bethe >0 . This theorem is a consequence of the Triangular Theorem below, see Section 5.3. Theorem 4.4. The Lusztig's mutations, i.e. mutiplications by ei(c) on the left are translated by W to the Wronskian mutations νi(c) on the right: W (ei(c)b) = νi(c)W (b), i = 1, . . . , r . (4.8.1) Proof. The theorem follows from the W5 Identity. Let us treat the case r = 3. We have b = (b1, b2, b3, b4), bi = xi−1 (i − 1)! The i = 1 case. We have + . . . , Wr(b) = 1 , W (b) = y = (y1, y2, y3, 1) . e1(c)b = (b1 + cb2, b2, b3, b4), y1 = b1 , Wr(y1, b2) = Wr(b1, b2) = y2 . 14 VADIM SCHECHTMAN AND ALEXANDER VARCHENKO Hence y1 = b2 is the solution of the normalized Wronskian equations (3.2), (3.3), and ν1(c)y = (y1 + cy1, y2, y3) = W (e1(c)y) , cf. formula (3.6). The i = 2 case. We have Denote y2 := Wr(b1, b3) . Then e2(c)b = (b1, b2 + cb3, b3, b4). by formula (2.3), and y2(x) = x + O(x2) as x → 0 , Wr(y2, y2) = Wr(Wr(b1, b2), Wr(b1, b3)) = Wr(b1) Wr(b1, b2, b3) = y1y3 , by the W5 Identity. Hence y2 is the solution of the normalized Wronskian equations (3.2), (3.3), and ν2(c)y = (y1, y2 + cy2, y3) = W (e2(c)y) . The i = 3 case. We have Denote y3 := Wr(b1, b2, b4) . Then e3(c)b = (b1, b2, b3 + cb4, b4) y3(x) = x + O(x2) as x → 0 , by formula (2.3), and Wr(y3, y3) = Wr(Wr(b1, b2, b3), Wr(b1, b2, b4)) = Wr(b1, b2) Wr(b1, b2, b3, b4) = y2 , by the W5 Identity and formula (4.8). Hence y3 is the solution of the normalized Wronskian equations (3.2), (3.3), and The case of arbitrary r is similar. (cid:3) ν3(c)y = (y1, y2, y3 + cy3) = W (e3(c)y) . 5. Triangular coordinates on the Bethe cell 5.1. Example of evolutions, group SL3 . 5.1.1. Wronskian evolution. We start with y(0) = (1, 1) and a reduced decomposition of the longest element in S3, h = (121) : w0 = s1s2s1 . The pair y(0) evolves by means of the normalized mutations: (5.1) (1, 1) ν1(b1) −→ (1 + b1x, 1) ν1(b3) −→ (1 + b1x + b3(x + b2x2/2), 1 + b2(x + b1x2/2)) . ν2(b2) −→ (1 + b1x, 1 + b2(x + b1x2/2)) The last pair is ν1(b3)ν2(b2)ν1(b1) y(0) . The second reduced word, h′ = (212), gives rise to another evolution: (5.2) (1, 1) ν2(c1) −→ (1, 1 + c1x) ν2(c3) −→ (1 + c2(x + c1x2/2), 1 + c1x + c3(x + c2x2/2)) . ν1(c2) −→ (1 + c2(x + c1x2/2), 1 + c1x) POSITIVE POPULATIONS 15 The change of coordinates on Y Bethe from (b1, b2, b3) to (c1, c2, c3) is c1 + c3 = b2, c2c3 = b1b2 , c2 = b1 + b3, 3 c1c2 = b2b3, whence (5.3) c1 = b2b3 b1 + b3 , c2 = b1 + b3, c3 = b1b2 b1 + b3 , cf. [L, Proposition 2.5]. 5.1.2. Bethe cell and positive population. Analyzing formulas (5.1) and (5.2) we observe that the Bethe cell Y Bethe consists of the polynomials (1 + e1x + e2/x2/2, 1 + f1x + f2x2/2) such that 3 (5.4) cf. (1.9), and the positive population Y Bethe e2 + f2 = e1f2 , 3, >0 ⊂ Y Bethe 3 is cut from Y Bethe 3 by the inequalities (5.5) e1, e2, f1, f2 > 0 . 5.1.3. Whitney-Lusztig evolution. We start with b(0) = (b1, b2, b3) = (1, x, x2/2) and act on it by the elementary unipotent matrices in the order dictated by the reduced word h = (121): ei(t) = 1 + tei,i+1, i = 1, 2 , (1, x, x2/2) e1(t1) −→ (1 + t1x, x, x2/2) e1(t3) −→ (1 + t1x + t3(x + t2x2/2), x + t2x2/2, x2/2) . e2(t2) −→ (1 + t1x, x + t2x2/2, x2/2) The resulting triple is (5.6) bh(t) = (bh,1(t; x), bh,2(t; x), bh,3(t; x)) := (1 + t1x + t3(x + t2x2/2), x + t2x2/2, x2/2) , cf. Section 4.2. One easily checks that applying the Wronski map W to the evolution (5.6) we obtain the evolution (5.1). 5.1.4. Remark. Note a useful formula (5.7) Wr(1 + t1x, x + t2x2/2) = 1 + t2(x + t1x2/2) , which could be written symbolically as In general let Wr(e1(t1), e2(t2)) = e1(t2)e2(t1) . b = (1 + ax + bx2/2, x + cx2/2, x2/2) . Using the formulas (Wr(1, x), Wr(1, x2/2), Wr(x, x2/2)) = (1, x, x2/2) , we get y := W (b) = (1 + ax + bx2/2, 1 + cx + (ac − b)x2/2) . Thus, the 2 × 2-matrix of nontrivial coefficients of y has the form b (cid:18)a c ac − b(cid:19) . 16 VADIM SCHECHTMAN AND ALEXANDER VARCHENKO The variety Y Bethe 3 may be identified with an exchange relation familiar in the theory of cluster algebras. Cf. Example 7.4 below. {(aij) ∈ gl2(C) a11a21 = a12 + a22} ⊂ gl2(C), 5.2. Example of evolutions, group SL4. 5.2.1. Wronskian evolution. There are 16 distinct reduced decompositions of the longest element w0 ∈ S4. We choose one of them: h = (121321) . We start with y(0) = (1, 1, 1) and perform the sequence of normalized mutations corresponding to the reduced decomposition h : (1, 1, 1) ν1(a1) −→ (1 + a1x, 1, 1) ν2(a2) −→ (1 + a1x, 1 + a2(x + a1x2/2), 1) ν3(a3) −→ (1 + a1x, 1 + a2(x + a1x2/2), 1 + a3(x + a2(x2/2 + a1x3/6))) ν1(a4) −→ (1 + a1x + a4(x + a2x2/2), 1 + a2(x + a1x2/2), 1 + a3(x + a2(x2/2 + a1x3/6))) . To find the next normalized mutation, ν2(a5), we have to solve an equation (5.8) Wr(1 + a2(x + a1x2/2, x + b2x2/2 + b3x3/6 + b4x4/24) = with respect to b2, b3, b4. We calculate inductively the coefficients of x, x2, x3 in (5.8) and obtain: (cid:0)1 + a1x + a4(x + a2x2/2)(cid:1) ·(cid:0)1 + a3(x + a2(x2 + a1x3/6))(cid:1) b2 = a1 + a3 + a4, b3 = 2(a1 + a4)a3, b4 = 2a2a3a4 . Note that the coefficients of x4 and x5 in the left-hand and right-hand sides of (5.8) should be equal as well, but the corresponding additional equations on b2, b3, b4 will be satisfied identically -- these are incarnations of the "Bethe equations". Thus, ν2(a5)ν1(a4)ν3(a3)ν2(a2)ν1(a1)y(0) = (1 + a1x + a4(x + a2x2/2), 1 + a2(x + a1x2/2) + a5(x + (a1 + a3 + a4)x2/2 + 2(a1 + a4)a3x3/6 + 2a2a3a4x4/24), 1 + a3(x + a2(x2/2 + a1x3/6))) . Finally, to apply ν1(a6) to this 3-tuple, we have to find c2, c3 from the equation Wr(1 + a1x + a4(x + a2x2/2), x + c2x2/2 + c3x6/6) = 1 + a2(x + a1x2/2) + a5(x + (a1 + a3 + a4)x2/2 + 2(a1 + a4)a3x3/6 + 2a2a3a4x4/24 . We find c2, c3 by equating the coefficients of x, x2: c2 = a2 + a5, c3 = a3a5 . Then the coefficients of x3, x4 will be equal as well by the "Bethe equations". Thus, yh(a) = (yh,1(a; x), yh,2(a; x), yh,3(a; x)) = ν1(a6)ν2(a5)ν1(a4)ν3(a3)ν2(a2)ν1(a1)y(0) = (1 + a1x + a4(x + a2x2/2) + a6(x + (a2 + a5)x2/2 + a3a5x3/6), 1 + a2(x + a1x2/2) + a5(x + (a1 + a3 + a4)x2/2 + 2(a1 + a4)a3x3/6 + 2a2a3a4x4/24), 1 + a3(x + a2(x2/2 + a1x3/6))) , POSITIVE POPULATIONS 17 cf. formula (3.9). 5.2.2. Lusztig evolution. Consider the Lusztig evolution corresponding to the same word h = (121321): e1(a1) −→ (1 + a1x, x, x2/2, x3/6) b(0) = (1, x, x2/2, x3/6) e2(a2) −→ (1 + a1x, x + a2x2/2, x2/2, x3/6) e1(a4) −→ (1 + a1x + a4(x + a2x2/2), x2/2 + a3x3/6, x3/6) e2(a5) −→ (1 + a1x + a4(1 + a2x2/2), x + a2x2/2 + a5(x2/2 + a3x3/6), x2/2 + a3x3/6, x3/6) e1(a6) −→ (1 + a1x + a4(1 + a2x2/2) + a6(x + a2x2/2 + a5(x2/2 + a3x3/6)), e3(a3) −→ (1 + a1x, x + a2x2/2, x2/2 + a3x3/6, x3/6) x + a2x2/2 + a5(x2/2 + a3x3/6), x2/2 + a3x3/6, x3/6) = bh(a) = (bh,1(a; x), bh,2(a; x), bh,3(a; x), bh,4(a; x)) , cf. formula (4.5). 5.2.3. Comparison. We have Namely, yh(a) = W (bh(a)) . yh,1(a; x) = bh,1(a; x) , yh,2(a; x) = Wr(bh,1(a; x), bh,2(a; x)) , yh,3(a; x) = Wr(bh,1(a; x), bh,2(a; x), bh,3(a; x)) . 5.2.4. Wronskian map and the minors of g. Let b = (1 + a1x + a2x2/2 + a3x3/6, x + b2x2/2 + b3x3/6, x2/2 + c3x3/6) ∈ N (5.9) and (5.10) g = 1 a1 a2 a3 b3 0 0 c3 1 0 b2 1 0 1 0 0   ∈ N .   Then b = gb(0). Let us compute y = (y1, y2, y3) = W (b) . Clearly, y1 = 1 + a1x + a2x2/2 + a3x3/6 , (5.11) 1 a2 1 a3 a1 a3 1 0 b2(cid:12)(cid:12)(cid:12)(cid:12) y2 = 1 +(cid:12)(cid:12)(cid:12)(cid:12) 0 b3(cid:12)(cid:12)(cid:12)(cid:12) x +(cid:18)(cid:12)(cid:12)(cid:12)(cid:12) x3/3 +(cid:12)(cid:12)(cid:12)(cid:12) b3(cid:12)(cid:12)(cid:12)(cid:12) +(cid:12)(cid:12)(cid:12)(cid:12) x +(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) (cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) y3 = 1 +(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) 1 a1 a3 b3 0 0 c3 1 0 +(cid:12)(cid:12)(cid:12)(cid:12) b3(cid:12)(cid:12)(cid:12)(cid:12) a1 a2 1 b2(cid:12)(cid:12)(cid:12)(cid:12) (cid:19)x2/2 x2/2 +(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) (cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) a2 a3 b2 x4/12 , 1 a2 a3 b3 0 b2 0 1 c3 a1 a2 a3 b3 1 0 c3 b2 1 x3/6 . (cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) 18 VADIM SCHECHTMAN AND ALEXANDER VARCHENKO Notice that all these determinants are minors of the matrix g. In particular, if the matrix g is totally positive, then all these determinants are positive. 5.2.5. Wronskian map in matrix form. By formula (5.11) : y1 = 1 + a1x + a2x2/2 + a3x3/6 , y2 = 1 + b2x + (b3 + a1b2 − a2)x2/2 + (a1b3 − a3)x3/3 + (a2b3 − a3b2)x4/12 , y3 = 1 + c3x + (b2c3 − b3)x2/2 + (a1(b2c3 − b3) − (a2c3 − a3))x3/6 . These formulas show that the inverse map W −1 : Y Bethe ∼−→ N assigns to a triple y = (y1, y2, y3) with y1 = 1 + α1x + α2x2/2 + α3x3/6 , y2 = 1 + β2x + β3x2/2 + β4x3/6 + β5x4/24 , y3 = 1 + γ3x + γ4x2/2 + γ5x3/6 , the triple b = W −1(y), as in (5.9), with (5.12) a1 = α1, a2 = α2, b2 = β2, a3 = α3, b3 = β3 − (α1β2 − α2), c3 = γ3 . In a matrix form, the map W is given by the formula (5.13) 0 1 α1 α2 α3 0 0 0 β2 β3 β4 β5 1 γ3 γ4 γ5 1 0   0 0 a3   =   1 a1 a2 b2 0 0 1 1 0 b3 + (a1b2 − a2) 2(a1b3 − a3) 2(a2b3 − a3b2) c3 b2c3 − b3 a1(b2c3 − b3) − (a2c3 − a3)   , and the inverse map W −1 is given by the formula (5.14) 1 a1 a2 a3 0 b3 c3 0 b2 1 1 0     =   1 α1 α2 0 0 1 0 α3 β2 β3 − (α1β2 − α2) 1 γ3   . Theorem 5.1. The coefficients α1, α2, α3, β2, β3, γ3 of the polynomials y1, y2, y3 serve as global coordinates on the Bethe cell Y Bethe , the other coefficients β4, β5, γ4 are polynomial functions of the global coordinates due to formulas (5.13) and (5.14). (cid:3) 4 Theorem 5.2. The positive population Y Bethe 4, >0 ⊂ Y Bethe 4 is cut from Y Bethe 4 by the inequalities (5.15) all αi, βi, γi > 0 and β3 > α1β2 − α2 > 0 . Proof. The proof follows from (5.11), (5.13), (5.14). (cid:3) 5.3. Triangular coordinates. For an arbitrary r, consider an element b = (b1, . . . , br+1) ∈ N , where POSITIVE POPULATIONS 19 bi = xi−1/(i − 1)! + r Xj=i bijxj, i = 1, . . . , r + 1 . Denote by the triangular array of nontrivial coefficients. Let y be the corresponding Bethe tuple, M(b) = (bij)16i6j6r with W (b) = y = (y1, . . . , yr) ∈ Y Bethe, yi(x) = 1 +Xj>1 aijxj/j! . Define the triangular part y△ of y, y△ = (y6r 1 , y6r−1 2 , . . . , y61 r ) . Here for a polynomial f (x) = Pi>0 aixi/i! ∈ C[x] we use the notation n f 6n(x) = aixi/i! . Xi=0 Denote A(y△) = (aij)16i6r;16j6r+1−i , the triangular array of the nontrivial coefficients of y△; thus we take all r nontrivial coefficients of y1(x); the first r − 1 nontrivial coefficients of y2(x), etc. The following statement describes the relationship between the two arrays M(b) and A(y). Theorem 5.3 (Triangular Theorem). (i) For all 1 6 i 6 j 6 r, we have (5.16) ai,j−i+1 = bij + ϕij((bkl)k<i) , where ϕij is a polynomial with all monomials of degree at least 2. (ii) Conversely, for all 1 6 i 6 j 6 r, we have (5.17) bij = ai,j−i+1 + ψij((akl)k<i) , where ψij is a polynomial with all monomials of degree at least 2. (iii) The map is an isomorphism. (iv) The Wronski map is an isomorphism. Y Bethe −→ Cq, y 7→ A(y△) , W : N −→ Y Bethe Proof. All statements are corollaries of statement (i). We leave a proof of (i) to the reader, cf. formulas (5.13) and (5.14). (cid:3) 20 VADIM SCHECHTMAN AND ALEXANDER VARCHENKO Corollary 5.4. The coefficients (aij)16i6r;16j6r+1−i of the polynomials y = (y1, . . . , yr) serve as global coordinates on the Bethe cell Y Bethe , the other coefficients of y are polynomial functions of the global coordinates due to formulas (5.16) and (5.17). 5.4. Polynomials ϕij. The next statement gives information on polynomials ϕij(b). Let b(0) be given by (4.4). An arbitrary b ∈ B has the form b = gb(0) for some unique g ∈ N. r Theorem 5.5. The polynomials ϕij(b) are linear combinations, with strictly positive coefficients, of some minors of the matrix g. Consequently, if g is totally positive, then all the polynomials yi of the tuple W (b) = (y1, . . . , yr) have strictly positive coefficients. Proof. The proof follows from formula (2.3), cf. formula (5.11). (cid:3) 6. Bethe cells from subspaces of C[x] and from critical points In this section we describe a generalization of the previous correspondence. 6.1. From a vector space of polynomials to a population ZV , see [MV]. Let V ⊂ C[x] be an r + 1-dimensional vector space of polynomials in x. We assume that V has not base points, that is for any z ∈ C there is f (x) ∈ V such that f (z) 6= 0. For any z ∈ C there exists a unique r + 1-tuple of integers λ = (λ0 = 0 < λ1 < · · · < λr) such that for any i = 0, . . . , r, there exists f (x) ∈ V with the property: djf dxj (z)(z) = 0, The tuple λ is called the tuple of exponents of V at z. dλif dxλi (z) 6= 0, j < λi . Having λ introduce an r-tuple of nonnegative integers µ = (µ1, . . . , µr) by the formula µ1 + · · · + µi + i = λi, The tuple µ is called the slr+1 weight of V at z. i = 1, . . . , r . A point z is regular for V if λ = (0, 1, . . . , r) and hence, µ = (0, . . . , 0), otherwise the point z is singular. Denote by ΣV the set of singular points. The set ΣV is finite. Denote ΣV = {z1, . . . , zn} ⊂ C . Denote µ(a) = (µ(a) 1 , . . . , µ(a) r ) the weight vector at a singular point za. Introduce an r-tuple of polynomials T = (T1, . . . , Tr), (6.1) Ti(x) = n Ya=1 (x − za)µ(a) i . Let b = (b1, . . . , br+1) be a basis of V , then the Wronskian Wr(b1, . . . , br+1) does not depend on the choice of the basis up to multiplication by a nonzero constant. Moreover, Wr(b1, . . . , br+1) = const T r 1 T r−1 2 . . . T 2 r−1Tr . This formula shows that the set ΣV of singular points of V is the set of zeros of the Wronskian of a basis of V . For any i = 2, . . . , r and b1, . . . , bi ∈ V introduce the reduced Wronskian Wr† V (b1, . . . , bi) = Wr(b1, . . . , bi) T 1−i 1 T 2−i 2 . . . T −1 i−1 . For any i = 1, . . . , r and b1, . . . , bi ∈ V , the reduced Wronskian is a polynomial. POSITIVE POPULATIONS 21 Introduce the reduced Wronski map W † V , which maps the variety of bases of V to the space of r-tuples of polynomials. If b = (b1, . . . , br+1) is a basis of V , then V (b1, b2), . . . , Wr† V : b 7→ y = (y1, . . . , yr) := (b1, Wr† (6.2) W † V (b1, . . . , br)) , cf. (4.10). We set y0 = yr+1 = 1. Denote (6.3) Then (6.4) yi = Wr† V (b1, b2, . . . , bi−1bi+1) , i = 1, . . . , r . Wr(yi, yi) = const Ti yi−1 yi+1 , i = 1, . . . , r . The generalized Wronski map W † V induces a map of the variety of bases of V to the direct product of projective spaces P(C[x])r. The bases defining the same complete flag of V are mapped to the same point of P(C[x])r. Hence the generalized Wronski map induces a map, W † V : XV → P(C[x])r , of the variety XV of complete flags of V to P(C[x])r. Theorem 6.1 ([MV]). This map is an embedding. The image (6.5) ZV ⊂ P(C[x])r of this map is isomorphic to the variety of complete flags of V . The variety ZV is called the population associated with V , see [MV]. See in Sections 1.5-1.7 the example of this construction corresponding to the case, where V = C[x]6r is the space of all the polynomials of degree 6 r. In that case the set of singular points of V is empty and T1 = · · · = Tr = 1. A tuple y = (y1, . . . , yr) ∈ C[x]r is called fertile with respect to T1, . . . , Tr, if for any i = 1, . . . , r the equation (6.6) Wr(yi, yi) = Ti yi−1 yi+1 with respect to yi(x) admits a polynomial solution. For example, all tuples (y1 : · · · : yr) ∈ ZV are fertile due to (6.4). A tuple y = (y1, . . . , yr) ∈ C[x]r is called generic with respect to T1, . . . , Tr, if for any i the polynomial yi(x) has no multiple roots and the polynomials yi(x) and Ti(x)yi−1(x) yi+1(x) have no common roots. Lemma 6.2 ([MV]). All y ∈ ZV are fertile. Also there exists a Zariski open subset U ⊂ ZV such that any y ∈ U is generic. Let y = (y1, . . . , yr) be a tuple of polynomials. Denote k = (k1, . . . , kr) := (deg y1, . . . , yr). Let ki yi(x) = (x − t(i) j ) , Yj=1 i = 1, . . . , r , 22 VADIM SCHECHTMAN AND ALEXANDER VARCHENKO where t(i) ordered in any way. j are the roots of yi. Denote t = (t(1) 1 , . . . , t(1) k1 ; . . . ; t(r) 1 , . . . , t(r) kr ), the tuple of roots of y Theorem 6.3 ([MV]). A tuple y is generic and fertile with respect to T1, . . . , Tr if and only if the tuple of roots t is a critical points of the master function (6.7) Φk(u, z, µ) = r ki n Yi=1 Yj=1 Ya=1 (u(i) j − za)−µ(a) i r × Yi=1 Y16l<m6ki (u(i) l − u(i) m )2 · r−1 Yi=1 ki Yl=1 ki+1 Ym=1 (u(i) l − u(i+1) m )−1 . These master functions appear in the integral representations of the KZ equations associated with glr+1 , see [SV]. The following statement is converse to the statement of Theorem 6.3. Given a finite set z = (z1, . . . , zn), a collection of µ(a) = (µ(a) 1 , . . . , µ(a) r ) ∈ Zr >0, a = 1, . . . , n, a vector k = (k1, . . . , kr) ∈ Zr >0, we define the master function Φk(u, z, µ) by formula (6.7). Theorem 6.4 ([MV]). If t is a critical point of a master function Φk(u, z, µ) with respect to the u-variables, then there exists a unique r + 1-dimensional vector space V ⊂ C[x], a basis b of V , such that the roots of the r-tuple of polynomials W † V (b) give t. Cf. Theorem 1.2. 6.2. From a critical point to a population ZV , [MV]. Given z = (z1, . . . , zn), a collec- tion of µ(a) = (µ(a) let t = (t(1) kr ), be a critical point of the master function Φk(u, z, µ). Define >0, a = 1, . . . , n, a vector k = (k1, . . . , kr) ∈ Zr 1 , . . . , µ(a) 1 , . . . , t(1) 1 , . . . , t(r) k1 ; . . . ; t(r) r ) ∈ Zr >0, (6.8) (6.9) and the tuple Ti(x) = yi(x) = n Ya=1 Yj=1 ki (x − za)µ(a) i , i = 1, . . . , r , (x − t(i) j ) , i = 1, . . . , r , y = (y1 : · · · : yr) ∈ P(C[x])r . The tuple y is generic and fertile with respect to T1, . . . , Tr. Hence for every i, there exists a polynomial yi satisfying the equation Wr(yi, yi) = const Ti yi−1 yi+1 . Choose one solution yi of this equation, denote yi(c; x) = yi(x) + cyi(x) , c ∈ C , and define a curve in P(C[x])r by the formula (6.10) y(i)(c; x) = (y1(x) : · · · : yi(c; x) : · · · : yr(x)) ∈ P(C[x])r . This curve is called the generation from y in the i-th direction. POSITIVE POPULATIONS 23 Thus starting from the point y in P(C[x])r we have constructed r curves in P(C[x])r. Now starting from any point of the constructed curves we may repeat this procedure and generate new r curves in P(C[x])r in any of the r directions. Repeating this procedure in all possible direction in any number of steps we obtain a subset Z ⊂ P(C[x])r of all points appearing in this way. The subset Z is called the population generated from the critical point t Theorem 6.5 ([MV]). The set Z is an algebraic variety isomorphic to the variety X of complete flags in an r+1-dimensional vector space. Moreover, starting with given t one can also determine uniquely an r + 1-dimensional vector space V and a basis b of V , such that y = W † V (b) and Z = ZV . 6.3. Bethe cells associated with (V ; z). Let V be an r + 1-dimensional vector space as in Section 6.1. Let ΣV = {z1, . . . , zn} ⊂ C be the set of singular points of V . Fix a complex number z /∈ ΣV , a regular point of V . We say that a basis b = (b1, . . . , br+1) of V is a unipotent basis of V with respect to z, if for any i = 1, . . . , r + 1, we have (6.11) bi = (x − z)i−1 (i − 1)! + O((x − z)i) as x → z . Denote by N (V ; z) the set of all unipotent bases of V at z. If we consider each basis b of V as an r + 1-column vector, then the group N freely acts on N (V ; z) from the left with one orbit. We call N (V ; z) the cell of bases of V unipotent at z. For any i = 2, . . . , r and b1, . . . , bi ∈ V introduce the reduced Wronskian normalized at z by the formula Wr† V,z(b1, . . . , bi) := Wr† V (b1, . . . , bi) T i−1 1 T i−1 1 T i−1 1 (z)T i−2 2 (z)T i−2 (x)T i−2 2 2 = Wr(b1, . . . , bi) (z) . . . Ti−1(z) (z) . . . Ti−1(z) (x) . . . Ti−1(x) . For any i = 1, . . . , r and b1, . . . , bi ∈ V , this is a polynomial. Introduce the reduced Wronski map W † V,z, which maps the variety of bases of V to the space of r-tuples of polynomials. If b = (b1, . . . , br+1) is a basis of V , then (6.12) W † V : b 7→ y = (y1, . . . , yr) := (b1, Wr† V,z(b1, b2), . . . , Wr† V,z(b1, . . . , br)) , cf. (4.10). We set y0 = yr+1 = 1. If b = (b1, . . . , br+1) ∈ N (V ; z), then (6.13) yi(z) = 1 , i = 1, . . . , r . Introduce the Bethe cell Y Bethe(V ; z) as the image of the cell N (V ; z) under the reduced Wronski map W † V,z , (6.14) Y Bethe(V ; z) := W † V,z(N (V ; z)) ⊂ C[x]r . Theorem 6.6. The reduced Wronski map W † V,z induces an isomorphism (6.15) W † V,z : N (V ; z) → Y Bethe(V ; z) . 24 VADIM SCHECHTMAN AND ALEXANDER VARCHENKO Proof. The theorem is deduced from Theorem 6.1, or it can be proved along the lines of the proof of Triangular Theorem 5.3, which corresponds to the subspace V = C[x]6r of polynomials of degree 6 r and z = 0. (cid:3) Corollary 6.7. The action of N on N (V ; z) and the Wronski map W † on the Bethe cell Y Bethe(V ; z). V,z induce an action of N Lemma 6.8. There exists a Zariski open subset U ⊂ Y Bethe(V ; z), such that any y ∈ U is generic and fertile. Proof. The theorem follows from Lemma 6.2. (cid:3) Recall that by Theorem 6.3, if y = (y1, . . . , yr) is generic and fertile, then roots of these polynomials determine a critical point of a suitable master function. 6.4. Normalized mutations and N -Y correspondence. Let y = (y1, . . . , yr) ∈ Y Bethe(V ; z) and i = 1, . . . , r. Define the normalized mutation of y in the i-th direction. Consider the differential equation (6.16) Wr(yi(x), yi(x)) = with respect to yi with initial condition Ti(x) Ti(z) yi−1(x) yi+1(x), (6.17) It has the unique solution yi(z) = 0, dyi dx (z) = 1. (6.18) yi(x) = yi(x)Z x z Ti(u)yi−1(u)yi+1(u) Ti(z)yi(u)2 du . This solution is a polynomial, cf. equation (6.4). For c ∈ C, denote (6.19) yi(c; x) = yi(x) + cyi(x) . Notice that yi(c; z) = 1. Define a new r-tuple of polynomials (6.20) νi(c)y := (y1(x), . . . , yi−1(x), yi(c; x), yi+1(x), . . . , yr(x)) . We call it the i-th normalized mutation of the tuple y ∈ N (V ; z). Lemma 6.9. For any i, c the tuple νi(c)y lies in Y Bethe(V ; z). Proof. The lemma follows from formula (6.4). By this lemma we have a map νi(c) : Y Bethe(V ; z) → Y Bethe(V ; z) . Recall the unipotent matrices ei(c), i = 1, . . . , r, c ∈ C, introduced in (4.1). Theorem 6.10 (Comparison Theorem). Let b ∈ N (V ; z), i = 1, . . . , r, c ∈ C. Then (6.21) W † V,z(ei(c)b) = µi(c)W † V,z(b) . Proof. The proof follows from formula (6.4). Cf. the proof of Theorem 4.4. (cid:3) (cid:3) POSITIVE POPULATIONS 25 Corollary 6.11. By Corollary 6.7 the group N acts on the Bethe cell Y Bethe(V ; z). By Theorem 6.10 a unipotent matrix ei(c) acts on the Bethe cell Y Bethe(V ; z) as the normalized mutation µi(c). 6.5. Positive populations. Fix any point b(0) ∈ N (V ; z), that is any basis of V unipotent at z. Then N b(0) = N (V, z) . Define the totally positive part of N (V, z) as N>0(V ; z; b(0)) := N>0b(0) . The totally positive part depends on the choice of b(0). Denote y(0) := W † V,z(b(0)) ∈ Y Bethe(V ; z) . Since b(0) is any point of N (V ; z), the point y(0) could be an any point of the Bethe cell Y Bethe(V ; z). Define the totally positive Bethe subvariety or positive population as (V ; z; b(0)) := N>0(y(0))) ⊂ Y Bethe(V ; z) . Y Bethe (6.22) >0 We also have (6.23) Y Bethe >0 (V ; z; b(0)) = W † V,z(N>0(V ; z; b(0)) . 6.6. Coordinates on the Bethe cell. Let b(0) be a point of the Bethe cell Y Bethe(V ; z). Let h = siq . . . si1 ∈ Red(w0) be a reduced decomposition of the longest element w0 ∈ Sr+1 . We call the map νh : Cq −→ Y(V ; z)Bethe, (cq, . . . , c1) 7→ νiq (cq) . . . νi1(c1)y(0) , the Wronskian chart corresponding to h. Its image is Zariski open. The map νh is a birational isomorphism. For any two reduced words h, h′ ∈ Red(w0) we have the transition function ν−1 h′ ◦ νh, which defines an automorphism (6.24) of the field F := C(c1, . . . , cq) . Rh,h′ : F ∼−→ F Recall the Whitney-Lusztig charts on N, see (4.3), Lh : Cq −→ N and transition function automorphisms Rh,h′ : F ∼−→ F , defined for any two words h, h′ ∈ Red(w0), see Section 1.3. Theorem 6.12. For any h, h′ ∈ Red(w0) we have (6.25) Rh,h′ = Rh,h′ . Proof. The theorem follows from Comparison Theorem 6.10. (cid:3) 26 VADIM SCHECHTMAN AND ALEXANDER VARCHENKO Corollary 6.13. For any h, h′ ∈ Red(w0), the map νh′ ◦ ν−1 population Y Bethe (V ; z; b(0)) and defines an isomorphism >0 is well defined on the positive h νh′ ◦ ν−1 h : Y Bethe >0 (V ; z; b(0)) → Y Bethe >0 (V ; z; b(0)) . Proof. The corollary follows from Theorem 1.1 and Comparison Theorem 6.10. (cid:3) 7. Base affine space and fat populations 7.1. Base affine space. Let N, N− ⊂ G := SLr+1(C) be the subgroups of the upper and lower triangular matrices with 1's on the diagonal. Let T ⊂ G be the subgroup of diagonal matrices. Let B− = N−T be the subgroup of lower triangular matrices and B = NT the subgroup of upper triangular matrices The quotient G/N− is called the base affine space of G. It is fibered over the flag space G/B− with fiber T . The image B of B in G/N− is called the big cell of the base affine space G/N− . 7.2. Fat population. Let V be an r + 1-dimensional vector space as in Section 6.1. The vector space V has a volume form. The volume of a basis b = (b1, . . . , br+1) of V is defined to be the number (7.1) Wr† V (b1, . . . , br+1) . Denote by BV the set of all bases of V of volume 1. We consider every basis as an r + 1-column vector. Then the group SLr+1 acts on BV on the left freely with one orbit. The quotient BV /N− is isomorphic to the base affine space of SLr+1. The reduced Wronski map W † V defined in (6.2) induces a map (7.2) W † V : BV /N− → C[x]r . Theorem 7.1. The reduced Wronski map W † V : BV /N− → C[x]r is an embedding. The image of the map, denoted by Fat(ZV ), is isomorphic to the affine base space G/N− of the group SLr+1. The image Fat(ZV ) will be called the fat population associated with V . Proof. The proof is parallel to the proof of Theorem 6.1 in [MV]. (cid:3) We have another description of the fat population as a bundle (7.3) Fat(ZV ) → ZV over the population ZV , defined in Theorem 6.1, with fiber isomorphic to (C×)r. Namely, if y = (y1 : · · · : yr) ∈ ZV ⊂ P(C[x])r and (y1, . . . , yr) ∈ C[x]r is a representative of y, then the fiber over y consist of the following r-tuples of polynomials {(d1y1, . . . , dryr) ∈ C[x]r (d1, . . . , dr) ∈ (C×)r} . POSITIVE POPULATIONS 27 7.3. Fat Bethe cell. Let ΣV = {z1, . . . , zn} ⊂ C be the set of singular points of V . Fix a complex number z /∈ ΣV , a regular point of V . Let Y Bethe(V ; z) ⊂ C[x]r be the Bethe cell defined in (6.14). Define the fat Bethe cell Fat(Y Bethe(V ; z)) by the formula (7.4) Fat(Y Bethe(V ; z)) = = {(d1y1, . . . , dryr) ∈ C[x]r (y1, . . . , yr) ∈ Y Bethe(V ; z) , (d1, . . . , dr) ∈ (C×)r} . Recall N (V ; z), the cell of bases of V unipotent at z. Define the fat cell Fat(N (V ; z)) by the formula (7.5) Fat(N (V ; z)) = {(b1d1, b2d2/d1, . . . , brdr/dr−1, br+1/dr) ∈ C[x]r (b1, . . . , br+1) ∈ N (V ; z) , (d1, . . . , dr) ∈ (C×)r} . Clearly the fat cell is isomorphic to the big cell B of the base affine space G/N−. Theorem 7.2. The reduced Wronski map induces an isomorphism V : Fat(N (V ; z)) → Fat(Y Bethe(V ; z)) . (7.6) Hence the fat Bathe cell Fat(Y Bethe(V ; z)) is isomorphic to the big cell B of the base affine space G/N−. W † Proof. This theorem is a corollary of Theorem 6.6. (cid:3) 7.4. Example of a cluster structure on a fat Bethe cell. Consider the example of the 3- dimensional vector space V = C[x]62 of quadratic polynomials. In this case the set ΣV of singular points of V is empty. We choose z = 0, a regular point for V , and consider the corresponding fat Bethe cell Fat(Y Bethe(V ; z)). It consist of pairs of polynomials (α0 + α1x + α2x2/2, β0 + β1x + β2x2/2) (7.7) such that and such that the Plucker equation holds, α0 6= 0, β0 6= 0 α1β1 = α0β2 + α2β0 . This is a familiar relation in the cluster algebra structure of type A1 on the ring C[SL3 /N−], where the cluster variables are α1, β1, cf. [Z, Section 3.1] and [FZ]. In fact, in this case the coefficients of the polynomials in (7.7) are nothing else but the Plucker coordinates on C[SL3 /N−]. 8. Appendix: Fourteen and Eightfold Ways 8.1. Group SL3 and a 2-category. One can reformulate the Whitney-Lusztig data from Sec- tion 1.3 in the language of [MS]. Namely, consider a 2-category3 S3 whose objects are in bijection with S3; the 1-arrows corre- spond to the weak Bruhat order on this group: there are 6 elementary arrows: (8.1) (123) τ12−→ (213) τ23−→ (231) τ ′ 12−→ (321) 3for a definition of (globular) n-categories see [St] 28 and (8.2) VADIM SCHECHTMAN AND ALEXANDER VARCHENKO (123) τ ′ 23−→ (132) τ ′′ 12−→ (312) τ ′′ 23−→ (321). Finally, there is one nontrivial 2-arrow between two compositions (8.3) h : τ ′ 12τ23τ12 −→ τ ′′ 23τ ′′ 12τ ′ 23 . This structure is conveniently visualized in a hexagon, to be denoted P3: its 6 vertices corre- spond to objects of S3, 6 oriented edges - to elementary 1-morphisms τ , and the unique 2 cell - to the 2-morphism h. The Whitney-Lusztig data described above may be called a 2-representation ρ of S3: to each object we assign the same vector space V(ijk) = F3 with a fixed standard base. To the elementary arrows τ (′,′′) ij we assign elementary matrices: (8.4) ρ(τ12) = e1(a1), ρ(τ23) = e2(a2), etc. To products of elementary arrows we assign the products of the corresponding matrices. Finally, ρ(h) will be an automorphism of F given by (8.5) ρ(h) = R121;212 =: R , see formulas (1.3) and (1.4). Note that (8.6) see Section 1.3. R2 = IdF, Thus, the matrix ρ(τ ′′ 23τ ′′ 12τ ′ 23) is obtained from the matrix ρ(τ ′ 12τ23τ12) by applying the field automorphism ρ(h): (8.7) ρ(τ ′′ 23τ ′′ 23) = ρ(h)(cid:8)ρ(τ ′ 12τ ′ 12τ23τ12)(cid:9). Similarly, for any r one defines in [MS] an r-category Sr, whose 1-coskeleton is a usual 1- category corresponding to the symmetric group Sr+1 with the weak Bruhat order. Informally speaking, Sr is an r-category structure on the r + 1-th permutohedron Pr+1, the r-dimensional polyhedron in Rr+1, the convex hull of a generic Sr+1-orbit of a point x ∈ Rr+1. In particular for each w ∈ Sr+1 we have an r − 1-category HomSr (e, w); its 1-skeleton Hom61 Sr (e, w) := Sk1 HomSr (e, w) is a usual (1-)category; the set of its objects is by definition the set Red(w) which is obtained from Red(w) by identifying any two words h and h′ if their only distinction is a couple ij in h vs ji in h′ somewhere in the middle, with i − j > 1. According to [MS], Red(w) is equipped with a partial order, the 2nd Bruhat order 62. The Sr (e, w) corresponds to this order, i.e. two words h, h′ ∈ Red(w) are connected by category Hom61 a unique arrow if and only if h 62 h′. 8.2. Group SL4 and a tetrahedron equation. POSITIVE POPULATIONS 29 8.2.1. The Forteenfold Way. 4 Consider the case r = 3. The 3-category S4 is related to a polyhedron P4 which looks as follows. We take a regular octahedron and cut from it 6 little square pyramids at its vertices; we get a polyhedron with 6 × 4 = 24 vertices. It has 8 hexagons and 6 squares as its 2-faces. The vertices of P4 are in bijection with S4. An oriented edge x → y connects elements x, y ∈ S4 such that x 6 y and ℓ(y) = ℓ(x) + 1 where 6 is the weak Bruhat order, and ℓ(x) is the usual length (the number of factors in a reduced decomposition). Its eight hexagonal 2-faces correspond to identities of the form (8.8) sisi+1si = si+1sisi+1 , which give rise to eight "elementary" 2-morphisms of type (1.3). Its six square edges correspond to identities of the form (8.9) sisj = sjsi, i − j > 1. which give rise to the identity 2-morphisms in S4. It is convenient to imagine the vertices e and w0 as the "North" and "South" poles of P4. The set Red(w0) consists of 16 longest paths on P4, of length 6, going downstairs, which connect e and w0. For example, the path ℓ1 is (1234) τ1−→ (2134) τ2−→ (2314) τ ′ 1−→ (3214) τ3−→ (3241) τ ′ 2−→ (3421) τ ′′ 1−→ (4321), whereas the path ℓ5 is (1234) τ ′ 3−→ (1243) τ ′′ 2−→ (1423) τ ′′ 3−→ (1432) τ ′′′ 1−→ (4132) τ ′′′ 2−→ (4312) τ ′′′ 3−→ (4321). Fourteen paths are depicted below, it is a 14-fold Way: 212321 −→ 213231 = 231231 = 231213 −→ 232123 (8.10) ↑ 121321 123121 ↓ ↓ 323123 . 321323 ↑ 123212 −→ 132312 = 132132 = 312132 −→ 321232 The elementary paths which are equal as 1-morphisms are connected by the signs =. They correspond to 6 mutations of type (8.9) and are in bijection with the square 2-faces of P4. If we identify the paths related by =, we are left with the set Hom(1)(e, w0) = Red(w0) , which contains 8 elements. 4Cf. [GN]. 30 VADIM SCHECHTMAN AND ALEXANDER VARCHENKO Let us number the elements of Red(w0) as ℓc, c ∈ Z/8Z. These elements are connected by 8 mutations of type (8.8), which are geometrically given by hexagons: (8.11) ℓ0 ℓ0 h1(1) −→ ℓ1 h2(3) −→ ℓ2 h1(3) −→ ℓ3 h2(1) −→ ℓ4 . h1(4) −→ ℓ−1 h2(2) −→ ℓ−2 h1(2) −→ ℓ−3 h2(4) −→ ℓ−4 Thus, the 2-nd Bruhat order on the set Red(w0) converts it to an octagon. Finally we have one 3-morphism (homotopy) k : h2(1)h1(3)h2(3)h1(1) −→ h2(4)h1(2)h2(2)h1(4) which corresponds to the single 3-cell of P4, its body. 8.2.2. Representation of the 3-category. The Whitney - Lusztig data give rise to a 2-representation of the 3-category S4. It is visualized on the permutohedron P4 as follows. Our base field will be a purely transcendental extension of C, F = C(a) = C(a1, a2, a3, a4, a5, a6) ∼= C(N4). At each vertex we put the based F-vector space V = F6. At the edges we put the elementary matrices ei(aj) ∈ N3(F), 1 6 i 6 3, 1 6 j 6 6. For example, at the 3 edges going down from e we put the matrices e1(a1), e2(a1), e3(a1), and to the last three edges coming to w0 we put the matrices e1(a6), e2(a6), e3(a6). To any path we assign the product of the corresponding matrices. For example whereas ρ(ℓ1) = e1(a6)e2(a5)e3(a4)e1(a3)e2(a2)e1(a1), ρ(ℓ2) = e3(a6)e2(a5)e1(a4)e3(a3)e2(a2)e3(a1), On 2-faces of P4 we put certain automorphisms of the base field F. Namely, let us introduce involutive operators L(i) : F ∼−→ F, 1 6 i 6 5, by (8.12) L(i)(ai) = ai+1, L(i)(ak) = ak L(i)(ai+1) = ai, if k 6= i, i + 1, and R(j) : F ∼−→ F, 1 6 j 6 4, given by (8.13) cf. formula (1.3). R(j)(aj) = aj+1aj+2/(aj + aj+2), R(j)(aj+1) = aj + aj+2, R(j)(aj+2) = ajaj+1/(aj + aj+2), if k 6= i, i + 1, R(j)(ak) = ak On eight hexagons (resp. on six squares) we put the involutions R (resp. L) according to the POSITIVE POPULATIONS 31 picture: (8.14) F(212321) R(1) ↑ F(121321) L(3) ↓ F(123121) R(4) ↓ F(123212) R(3) −→ F(213231) L(2) −→ F(231231) L(5) −→ F(231213) R(3) −→ F(232123) ↓ R(1) F(323123) ↓ L(3) . F(321323) ↑ R(4) R(2) −→ F(132312) L(4) −→ F(132132) L(1) −→ F(312132) R(2) −→ F(321232) Here F(h), h ∈ Red(w0), is a copy of the field F. This way we have associated to any path ℓ = ℓ(h), h ∈ Red(w0), an upper triangular matrix ρ(ℓ) ∈ N(F), and to every homotopy ℓ h−→ ℓ′ an automorphism R(h) : F ∼−→ F such that (8.15) ρ(ℓ′) = R(h){ρ(ℓ)} Theorem 8.1. The diagram (8.14) is commutative, i.e., L(3)R(1)R(3)L(2)L(5)R(3)R(1) = R(4)R(2)L(1)L(4)R(2)R(4)L(3) . 8.2.3. Eightfold Way. We can rewrite this assertion as follows. Define 8 automorphisms (involutions or compositions of two involutions): R1(1) = R(1), R1(2) = L(4)R(2), R1(3) = L(5)R(3), R1(4) = R(4)L(3), R2(1) = L(3)R(1), R2(2) = R(2)L(1), R2(3) = R(3)L(2), R2(4) = R(4). Then we have a tetrahedron equation (8.16) R2(1)R2(3)R1(3)R1(1) = R2(4)R2(2)R1(2)R1(4) . References [BFZ] A. Berenstein, S. Fomin, A. Zelevinsky, Parametrizations of canonical bases and totally [FZ] [GK] positive matrices, Adv. Math., 122 (1996), 49 -- 149 S. Fomin, A. Zelevinsky, Cluster algebras I: foundations, JAMS, 15 (2002), 497 -- 529 F. Gantmacher, M. Krein, Sur les matrices oscillatoires, C.R. Acad. Sci. Paris, 201 (1935), 456 -- 477 [GN] M. Gell-Mann, Y. Ne'eman, The Eightfold Way, W. A. Benjamin, 1964 [Lo] Ch. Loewner, On totally positive matrices, Math. Zeitschr. Bd., 63 (1955), 338 -- 340 32 [L] [MS] VADIM SCHECHTMAN AND ALEXANDER VARCHENKO G. Lusztig, Total positivity in reductive groups, in: Lie theory and geometry: in honor of Bertram Konstant, Progress in Math., 123, Birkhauser, 1994 Yu.I. Manin, V.V. Schechtman, Arrangements of hyperplanes, higher braid groups and higher Bruhat orders, Algebraic number theory - in honor of K. Iwasawa, Adv. Stud. Pure Math., 17 (1989), 289 -- 308 [MTV] E. Mukhin, V. Tarasov, A. Varchenko, Schubert calculus and representations of general [MV] [SV] [ScV] [S] [St] [W] [Z] linear group, J. Amer. Math. Soc. 22 (2009), no. 4, 909 -- 940 E. Mukhin, A. Varchenko, Critical points of master functions and flag varieties, Com- mun. Contemp. Math., 6 (2004), 111 -- 163 V. Schechtman, A. Varchenko, Arrangements of hyperplanes and Lie algebra homology, Inv. Math., 106 (1991), 139 -- 194 I. Scherbak and A. Varchenko, Critical points of functions, sl2 representations, and Fuch- sian differential equations with only univalued solutions, Moscow Math. J., 3, n. 2 (2003), 621 -- 645 I. Schoenberg, Uber variationsvermindernde lineare Transformationen, Math. Zeitschr., 32 (1930), 321 -- 322 R. Street, The algebra of oriented simplices, J. Pure Appl. Alg., 49 (1987), 283 -- 335 A.M. Whitney, A reduction theorem for positive matrices, J. d'Analyse Math., 2 (1952), 88 -- 92 A. Zelevinsky, From Littlewood - Richardson coefficients to cluster algebras in three lec- tures, Symmetric Functions 2001: Surveys of Developments and Perspectives, NATO Science Series (Series II: Mathematics, Physics and Chemistry), 74, Springer, Dordrecht.
1604.06228	3	1604	2018-01-08T11:50:42	A note on homogeneous ideals of polarized abelian surfaces	[ "math.AG" ]	Gross and Popescu conjectured that the homogeneous ideal of an embedded $(1,d)$-polarized abelian surface is generated by quadrics and cubics for $d\geq 9$. We prove this using the projective normality of the embedding. It follows that the homogeneous ideal of an abelian surface embedded by a complete linear system is generated by quadrics and cubics, with three exceptions.	math.AG	math	A NOTE ON HOMOGENEOUS IDEALS OF POLARIZED ABELIAN SURFACES DANIELE AGOSTINI Abstract. Gross and Popescu conjectured that the homogeneous ideal of an embedded (1, d)- polarized abelian surface is generated by quadrics and cubics for d ≥ 9. We prove a generalization of this using the projective normality of the embedding. It follows that the homogeneous ideal of an abelian surface embedded by a complete linear system is generated by quadrics and cubics, with three exceptions. 1. Introduction We will always work over C, but the same arguments work over an algebraically closed field of characteristic 0. In their paper [GP98] Gross and Popescu proved that if (A, L) is a general polarized abelian surface of type (1, d) with d ≥ 10, then its homogeneous ideal in the embedding A ⊆ P(H0(A, L)) is generated by quadrics. At the end of the same paper, they formulated the following conjecture: Conjecture 1. [GP98, Conjecture (a)] Let (A, L) be a polarized abelian surface of type (1, d) such that L is very ample and d ≥ 9. Then the homogeneous ideal of A in the embedding A ⊆ P(H0(A, L)) is generated by quadrics and cubics. This result was already proven for d = 7 by Manolache and Schreyer in [MS01, Corollary 2.2], where they show that the ideal is generated by cubics and compute the whole minimal free resolution. The case d = 8 was proven by Gross and Popescu [GP01, Theorem 6.13] for a general abelian surface and the cases d ≥ 23 were recently proved by K uronya and Lozovanu [KL, Theorem 1.3] as a consequence of their Reider-type result for higher syzygies on such surfaces. The purpose of this note is to give a simple unified proof of the following: Proposition 1.1. Let (A, L) be a polarized abelian surface of type (1, d) such that L is very ample and d ≥ 7. Then the homogeneous ideal of A in the embedding A ⊆ P(H0(L)) is generated by quadrics and cubics. The proof we give is based on a simple application of Koszul cohomology, together with the following result of Lazarsfeld and Fuentes Garc´ıa: Theorem 1.1 (Lazarsfeld [Laz], Fuentes Garc´ıa [Gar04]). Let (A, L) be a polarized abelian surface of type (1, d) such that L is very ample and d ≥ 7. Then the embedding A ⊆ P(H0(A, L)) is projectively normal. Proposition 1.1 and previous results give the following: Theorem 1.2. Let (A, L) be a polarized abelian surfaces with L is very ample and not of type (1, 5), (1, 6), (2, 4). Then the embedding A ⊆ P(H0(A, L)) is projectively normal and the homogeneous ideal of A is generated by quadrics and cubics. Acknowledgments: I am very grateful to Robert Lazarsfeld for having made available to me the paper [Laz] and for allowing me to reproduce a part of it here. I would like to thank Ciro Ciliberto and Edoardo Sernesi for helpful discussions. I am happy to thank my advisor Gavril Farkas for his suggestion to study this question and for his advice. I would also like to thank the referee for the helpful comments. 2. Koszul cohomology Koszul cohomology is a language introduced by Green [Gre84] to study minimal free resolutions. We recall here its basic definitions and properties. Let X be an integral projective variety of positive 1 A NOTE ON HOMOGENEOUS IDEALS OF POLARIZED ABELIAN SURFACES 2 dimension and L a very ample line bundle on X. For every vector bundle E on X we can form the group Γ(X, E, L) = H0(X, Lq ⊗ E) (cid:77) q∈Z which has a natural structure of a finitely generated S = Sym•(H0(X, L))-module. In particular we can take its minimal free resolution: 0 ←− Γ(X, E, L) ←− F0 ←− F1 ←− · · · ←− Fp ←− . . . Fn ←− 0 where every Fp is a graded free S-module: (cid:77) q∈Z Fp = Kp,q(X, E, L) ⊗C S(−p − q). The vector spaces Kp,q(X, E, L) are called the Koszul cohomology groups of (X, E, L) and they can also be computed as the middle cohomology of the Koszul complex (see [Gre84, Theorem 1.b.4]): ∧p+1H0(X, L)⊗ H0(X, E⊗ Lq−1) −→ ∧pH0(X, L)⊗ H0(X, E⊗ Lq) −→ ∧p−1H0(X, L)⊗ H0(X, E⊗ Lq+1) If E = OX we denote Kp,q(X, L) :def= Kp,q(X, E, L). Remark 2.1. From the Koszul complex we see immediately that Kp,q(X, L) = 0 for q < 0 or q = 0, p > 0. Moreover we also see that the embedding X (cid:44)→ P(H0(X, L)) is projectively normal if and only if K0,q(X, L) = 0 for all q ≥ 2. In this case, the minimal free resolution of the ideal IX is given by 0 ←− IX ←− F1 ←− · · · ←− Fp ←− . . . Fn ←− 0 so that IX is generated by quadrics and cubics if and only if the only factors appearing in F1 are S(−2) and S(−3) which means that K1,q(X, L) = 0 for all q ≥ 3. We will need a duality theorem for Koszul cohomology: Theorem 2.1. [Gre84, Theorem 2.c.6] Let X be a smooth integral projective variety, L a very ample line bundle on X and E a vector bundle on X. Set r = h0(X, L) − 1 and suppose that (1) Then there is an isomorphism (2) Kp,q(X, E, L)∨ ∼= Kr−dim X−p,dim X+1−q(X, ωX ⊗ E∨, L). Hi(X, E ⊗ Lq−i) = Hi(X, E ⊗ Lq−1−i) = 0 for all i = 1, . . . , dim X − 1 In the case of abelian surfaces, the canonical bundle is trivial, so that we get the following lemma. Lemma 2.1. Let A ⊆ P(H0(L)) be an abelian surface embedded by a complete linear system. Then if the embedding is projectively normal, the homogeneous ideal of A is generated by quadrics and cubics. Proof. By Remark 2.1, we just need to prove that K1,q(A, L) = 0 for all q ≥ 3. Set r = h0(A, L) − 1, then Theorem 2.1 gives an isomorphism K1,q(A, L)∨ ∼= Kr−3,3−q(A, L) and the group on the right is zero because of Remark 2.1. 3. Projective normality of polarized abelian surfaces (cid:3) The key result that we are going to use is Theorem 1.1: this was proven by Lazarsfeld [Laz] in the cases d = 7, 9, 11 and d ≥ 13. The remaining cases d = 8, 10, 12 were solved by Fuentes Garc´ıa [Gar04]. We would like to sketch here the proof, in particular because the original preprint [Laz] is quite hard to find. We would like to emphasize that all the results and the ideas of this section are due to Lazarsfeld and Fuentes Garc´ıa. Our only contribution is in the presentation of the argument. First observe the following: A NOTE ON HOMOGENEOUS IDEALS OF POLARIZED ABELIAN SURFACES 3 Lemma 3.1. Let A be an abelian surface and L a very ample line bundle on it. Then the embedding A ⊆ P(H0(A, L)) is projectively normal if and only if the multiplication map Sym2H0(A, L) −→ H0(A, L2) is surjective. Proof. This is proven for an abelian variety of any dimension by Iyer in [Iye99, Proposition 2.1]. In the case of abelian surfaces, we can give a quick proof via Koszul cohomology as follows: suppose that the above multiplication map is surjective, then by definition K0,2(A, L) = 0 and by Remark 2.1 we just need to prove that K0,q(A, L) = 0 for all q ≥ 3. In this case we see that H1(A, Lq−1) = H1(A, Lq−2) = 0 so that Theorem 2.1 gives K0,2(A, L)∨ ∼= Kr−2,3−q(A, L) where r = h0(L) − 1. Now, it is enough to observe that r − 2 > 0 (since L is very ample) and 3 − q ≤ 0 by hypothesis, so Remark 2.1 gives Kr−2,3−q(A, L) = 0. (cid:3) Now, let us fix an embedded polarized abelian surface A ⊆ P(H0(A, L)) with L of type (1, d) and d ≥ 7. There is an exact sequence (3) 0 −→ I −→ Sym2H0(A, L) −→ H0(A, L2) −→ U −→ 0 and we want to prove U = 0. Lazarsfeld [Laz] makes the following observation: Lemma 3.2. With the above notations, we have dim U ≤ 6. Proof. The argument given here is different from the one in [Laz] and it is a slight clarification of the one given in [Gar]. It is based on the following lemma. Lemma 3.3. Let X ⊆ PN be a nondegenerate integral surface of degree t. Then dim H0(PN, IX/PN (2)) ≤ N(N − 1) − min{t, 2N − 5} 2 Proof. First recall that t ≥ N − 1 (see [EH87, Proposition 0]). Now choose H ⊆ PN to be a general linear subspace of codimension 2. Then H ∩ X consists of t distinct points in linearly general position in H: in particular, they span H, since t ≥ dim H + 1. Now we observe that there is no quadric Q ⊆ PN containing both X and H. Indeed, suppose that there is such a Q: then, since X is nondegenerate it would have rank at least 3 so that its singular locus Sing(Q) is a linear subspace of codimension at least 3, which cannot contain H ∩ X. This shows that X ∩ H ∩ (Q \ Sing(Q)) (cid:54)= ∅ and since H ∩ (Q \ Sing(Q)) is a Cartier divisor on Q \ Sing(Q) it follows from Krull's principal ideal theorem that every irreducible component of X ∩ H ∩ (Q \ Sing(Q)) has positive dimension, which gives a contradiction. This shows that the restriction map H0(PN, IX/PN (2)) −→ H0(H, IX∩H/H(2)) is injective. To conclude, we can just apply Castelnuovo's argument for which t ≥ N − 1 points in linearly general position in PN−2 impose at least min{t, 2N − 5} independent conditions on (cid:3) quadrics (see [ACGH85, Lemma p.115]). Now it is immediate to prove Lemma 3.2: we have A ⊆ P(H0(A, L)) ∼= Pd−1 of degree 2d, so that applying Lemma 3.3 we obtain dim U = dim H0(A, L2) − dim Sym2H0(A, L) + dim I ≤ 4d − (cid:3) Then it is enough to show that if U (cid:54)= 0, then dim U ≥ 7. Lazarsfeld's idea is to use the Heisenberg group associated to (A, L). (cid:18)d + 1 (cid:19) 2 d(d − 7) + 2 + 6 = 6 A NOTE ON HOMOGENEOUS IDEALS OF POLARIZED ABELIAN SURFACES 4 3.1. The Heisenberg group. Recall that to (A, L) we can associate the two groups K(L) = {x ∈ xL isomorphism }. The group G(L) is called the Xt∗ Heisenberg group of L and it is a central extension of K(L) by C∗ [Mum66, Theorem 1]: xL ∼= L} and G(L) = {(α, x)α : L → t∗ 1 −→ C∗ −→ G(L) −→ K(L) −→ 0 A linear representation of G(L) where C∗ acts by the character λ (cid:55)→ λk is called a representation of weight k. The space H0(A, L) has a natural linear action of G(L) given by (α, x) · σ = t∗−x(α(σ)) and, up to isomorphism, this is the unique irreducible representation of G(L) of weight 1 (see [Mum66, Proposition 3, Theorem 2]). This representation induces other representations of weight 2 on Sym2H0(A, L) and H0(A, L2) such that the multiplication map (3) is G(L)-equivariant. In particular, I and U can be regarded as G(L)-representations of weight 2. The irreducible ones have been classified by Iyer [Iye99, Proposition 3.2]. Proposition 3.1 (Iyer). Let (A, L) be a polarized abelian surface of type (1, d). Then (1) if d is odd, there is, up to isomorphism, a unique irreducible G(L)-representation of weight 2. This representation has dimension d. (2) if d = 2m is even, then there are, up to isomorphism, four distinct G(L)-representations of weight 2. Each irreducible representation has dimension m. This proves Theorem 1.1 for most cases: Proof of Theorem 1.1. Suppose that d is odd and greater than 7 or even and greater than 14. Assume by contradiction that in 3 we have U (cid:54)= 0: then, since U is a G(L)-representation it must be by Proposition 3.1 that dim U ≥ 7. However, this is impossible because dim U ≤ 6 by Lemma 3.2. This leaves the cases d = 8, 10, 12, which were solved by Fuentes Garc´ıa in [Gar04] using the involutions in G(L) coming from the 2-torsion points of K(L), together with geometric results (cid:3) about polarized abelian surfaces of small degree. Remark 3.1. Theorem 1.1 can be proven for d ≥ 10 also using the results of K uronya and Lozovanu: indeed in this case (L)2 ≥ 20, so that by [KL, Theorem 1.1] the embedding is not projectively normal if and only if there exists an elliptic curve E ⊆ A such that (E)2 = 0 and (E · L) = 1, 2. In particular, LE is not very ample so that L cannot be very ample. 4. Homogeneous ideals of polarized abelian surfaces Now it is very easy to prove Proposition 1.1: Proof of Proposition 1.1. Follows immediately from Theorem 1.1 and Lemma 2.1. (cid:3) Theorem 1.2 instead is a consequence of the following: Theorem 4.1 (Koizumi [Koi76], Lazarsfeld [Laz], Fuentes Garc´ıa [Gar04], Ohbuchi [Ohb93]). Let A ⊆ P(H0(L)) be an abelian surface embedded by a complete linear system. Then the embedding is projectively normal, unless L is of type (1, 5), (1, 6), (2, 4). In these cases, the embedding is never projectively normal. Proof. Suppose that L is of type (d1, d1m): if d1 ≥ 3, then the result was first proven by Koizumi [Koi76]. Another proof can be found in [BL04, Theorem 7.3.1]. If d1 = 2 and m ≥ 3 then projective normality follows from a result by Ohbuchi [Ohb93]. Alternatively, we can reason as in Remark 3.1. Ohbuchi also shows in [Ohb93, Lemma 6] that if m = 2, then L is not projectively normal. For another proof of this, Barth has shown in [Bar87, Theorem 2.11] that the ideal IA contains precisely 6 linearly independent quadrics. Hence Sym2H0(A, L) −→ H0(A, L2) has image of dimension 36 − 6 = 30, which is less than the dimension of H0(A, L2). If d1 = 1, then this is Theorem 1.1. Observe that there cannot be projective normality for L of type (1, 5) or (1, 6) because in these cases Sym2H0(A, L) has dimension smaller than H0(A, L2). (cid:3) (cid:3) In all the other cases, it is easy to see that the line bundle cannot be very ample. Proof of Theorem 1.2. This follows immediately from Theorem 4.1, Lemma 2.1. A NOTE ON HOMOGENEOUS IDEALS OF POLARIZED ABELIAN SURFACES 5 Remark 4.1. We can also say something about the exceptional cases: for a very ample line bundle of type (1, 5) Manolache has proven [Man88, Theorem 1] that the homogeneous ideal is generated by 3 quintics and 15 sextics. For the case (1, 6) Gross and Popescu [GP01, Remark 4.8.(2)] have proven that the ideal sheaf of a general such abelian surface is generated by cubics and quartics. For the case (2, 4) Barth [Bar87, Theorem 2.14,Theorem 4.9] gives explicit quadrics which generate the ideal sheaf of the surface: it is then easy (for example with Macaulay2 [M2]) to compute examples where the homogeneous ideal is generated by quadrics and quartics. References [ACGH85] Enrico Arbarello, Maurizio Cornalba, Phillip A. Griffiths, and Joseph Harris, Geometry of algebraic curves. Vol. I, Grundlehren der Mathematischen Wissenschaften [Funda- mental Principles of Mathematical Sciences], vol. 267, Springer-Verlag, New York, 1985. [Bar87] Wolf Barth, Abelian surfaces with (1, 2)-polarization, Algebraic geometry, Sendai, 1985, Adv. Stud. Pure Math., vol. 10, North-Holland, Amsterdam, 1987, pp. 41–84. [BL04] Christina Birkenhake and Herbert Lange, Complex abelian varieties, 2nd ed., Grundlehren der Mathematischen Wissenschaften [Fundamental Principles of Mathe- matical Sciences], vol. 302, Springer-Verlag, Berlin, 2004. [EH87] David Eisenbud and Joe Harris, On varieties of minimal degree (a centennial account), Algebraic geometry, Bowdoin, 1985, Proc. Sympos. Pure Math., vol. 46, Amer. Math. Soc., Providence, 1987. [Gar04] Luis Fuentes Garc´ıa, Projective normality of abelian surfaces of type (1, 2d), Manuscripta [Iye99] Jaya N. Iyer, Projective normality of abelian surfaces given by primitive line bundles, Manuscripta Math. 98 (1999), no. 2, 139–153. [Koi76] Shoji Koizumi, Theta relations and projective normality of Abelian varieties, Amer. J. Math. 98 (1976), no. 4, 865–889. [KL] Alex K uronya and Victor Lozovanu, A Reider-type theorem for higher syzygies on abelian surfaces, available at http://arxiv.org/abs/1509.08621. [Laz] Robert Lazarsfeld, Projectivit´e normale des surfaces ab´eliennes. (Redige par O. De- barre),Preprint No. 14, Europroj CIMPA, 1990. [M2] Daniel R. Grayson and Michael E. Stillman, Macaulay2, a software system for research in algebraic geometry, available at http://www.math.uiuc.edu/Macaulay2/. [Man88] Nicolae Manolache, Syzygies of abelian surfaces embedded in P4(C), J. Reine Angew. Math. [MS01] Nicolae Manolache and Frank-Olaf Schreyer, Moduli of (1, 7)-polarized abelian surfaces via syzygies, Math. Nachr. 226 (2001), 177–203. [Mum66] David Mumford, On the equations defining abelian varieties. I, Invent. Math. 1 (1966), 384 (1988), 180–191. 287–354. [Ohb93] Akira Ohbuchi, A note on the normal generation of ample line bundle on abelian surface, Proc. Amer. Math. Soc. 117 (1993), 275–277. Humboldt-Universit at zu Berlin, Institut f ur Mathematik, Unter den Linden 6, 10099, Berlin Germany E-mail address: [email protected] [Gar] , Projective normality of abelian surfaces of type (1, 2d), available at http://arxiv. Math. 114 (2004), no. 3, 385–390. org/abs/math/0306058. Geom. 19 (1984), no. 1, 125–171. 310 (1998), no. 2, 333–377. (2001), no. 2, 169–228. [Gre84] Mark L. Green, Koszul cohomology and the geometry of projective varieties, J. Differential [GP98] Mark Gross and Sorin Popescu, Equations of (1, d)-polarized abelian surfaces, Math. Ann. [GP01] , Calabi-Yau threefolds and moduli of abelian surfaces. I, Compositio Math. 127
1103.5482	1	1103	2011-03-28T20:46:45	Deformation of quotients on a product	[ "math.AG" ]	We consider the general problem of deforming a surjective map of modules $f : E \to F$ over a coproduct sheaf of rings $B=B_1 \otimes_A B_2$ when the domain module $E = B_1 \otimes_A E_2$ is obtained via extension of scalars from a $B_2$-module $E_2$. Assuming $B_1$ is flat over $A$, we show that the Atiyah class morphism $F \to \LL_{B/B_2} \otimes^{\bL} F[1]$ in the derived category $D(B)$ factors naturally through (the shift of) a morphism $\beta : \Ker f \to \LL_{B/B_2} \otimes^{\bL} F$. We describe the obstruction to lifting $f$ over a (square zero) extension $B_1' \to B_1$ in terms of $\beta$ and the class of the extension. As an application, we use the reduced Atiyah class to construct a perfect obstruction theory on the Quot scheme of a vector bundle on a smooth curve (and more generally).	math.AG	math	DEFORMATION OF QUOTIENTS ON A PRODUCT W. D. GILLAM Abstract. We consider the general problem of deforming a surjective map of modules f : E → F over a coproduct sheaf of rings B = B1 ⊗A B2 when the domain module E = B1 ⊗A E2 is obtained via extension of scalars from a B2-module E2. Assuming L F [1] in the B1 is flat over A, we show that the Atiyah class morphism F → LB/B2 derived category D(B) factors naturally through (the shift of) a morphism β : Ker f → L F . We describe the obstruction to lifting f over a (square zero) extension LB/B2 B ′ 1 → B1 in terms of β and the class of the extension. As an application, we use the reduced Atiyah class to construct a perfect obstruction theory on the Quot scheme of a vector bundle on a smooth curve (and more generally). ⊗ ⊗ Introduction Let Y be a scheme, E a quasi-coherent sheaf on Y . Let X0 ֒→ X be a square zero closed immersion with ideal I and let (1) 0 → N0 → π∗ 2E → F0 → 0 be an exact sequence of sheaves on X0 × Y with F0 flat over X0. Consider the problem of finding an exact sequence (2) 0 → N → π∗ 2E → F → 0 on X×Y with F flat over X restricting to (1) on X0×Y . Assuming Quot E is representable, this is equivalent to finding a map X → Quot E restricting to the map X0 → Quot E determined by (1). It is "well-known" that there is an obstruction ω ∈ Ext1 X0×Y (N0, π∗ 1I ⊗ F0) whose vanishing is necessary and sufficient for the existence of such a sequence and that, when ω = 0, the set of such sequences is a torsor under HomX0×Y (N0, π∗ 1I ⊗ F0). A simple proof is provided for the reader's convenience in (1.6). Following a trick of Nagata, Illusie (IV.3.2)1 constructs the obstruction ω by translating this lifting problem into a problem of G := Gm equivariant algebra extensions, which can be handled using the machinery of the G equivariant cotangent complex. The basic link between the G equivariant cotangent complex and the usual cotangent complex is provided by the equivalence of the two different descriptions of the Atiyah class in (IV.2.3.7.3). This discussion motivates the main results of this paper, which we now briefly summarize. For a scheme X and a quasi-coherent sheaf F on X, let X[F ] := SpecX OX ⊕ F denote the trivial square zero extension of X by F . The scheme X[F ] has a natural G := Gm action obtained by restricting the action of scalars on F to the invertible scalars. (X, F ) 7→ Date: October 19, 2018. 1For a roman numeral N , a reference in the present paper of the form (N.a) always refers to section a of chapter N in Illusie's book [Ill]. 1 2 W. D. GILLAM X[F ] is a contravariant functor from the category (stack) of quasi-coherent sheaves to the category of G schemes. Let X, Y be schemes over a field k so that the projection π2 : X × Y → Y is flat and LX. Let E be a quasi-coherent sheaf on Y , we have a natural isomorphism L X×Y /Y = π∗ 1 (3) 0 / N / π∗ 2E / F / 0 an exact sequence of sheaves on X × Y . From the morphism of distinguished triangles of G equivariant cotangent complexes associated to the diagram (4) (X × Y )[F ] / (X × Y )[π∗ 2E] / Y [E] (X × Y )[F ] / X × Y Y of G schemes, we obtain a commutative diagram (5) LG (X×Y )[F ]/(X×Y )[π∗ 2 E] / LG (X×Y )[π∗ 2 E]/Y [E] ⊗L OX×Y [π∗ 2 E] OX×Y [F ][1] LG (X×Y )[F ]/(X×Y ) / π∗ 1 LX ⊗L OX×Y OX×Y [F ][1] in the G equivariant derived category DG((X × Y )[F ]). Pushing forward to X × Y and taking the weight one subcomplex is an exact functor k1 : DG((X × Y )[F ]) → D(X × Y ). Applying k1 to (5) we obtain a commutative diagram in D(X × Y ) that can be written (6) β[1] N [1] / π∗ 1 LX ⊗L F [1] at(F ) F / π∗ 1 LX ⊗L F [1]. The bottom horizontal arrow is the Atiyah class of F relative to π2 and the D(X × Y ) morphism β : N → π∗ LX ⊗L F whose shift appears on the top line is called the reduced 1 Atiyah class of the quotient π∗ 2E → F . It is studied extensively in Section 2 where we give a description of the reduced Atiyah class in terms of principal parts, thereby avoiding the equivariant cotangent complex. Consider the functors HomD(X)(LX , ), HomD(X×Y )(N, π∗ 1 ⊗L F ) : D(X) → Vectk and the (k linear) natural transformation β : HomD(X)(LX , ) → HomD(X×Y )(N, π∗ 1 ⊗L F ) β(M )(g) := (π∗ 1g ⊗L F )β. Suppose that the Quot scheme of E is representable, X is an open subset of it, and (3) is the (restriction of the) universal sequence. In Theorem 4.2, we show that, for any quasi- coherent sheaf I on X (regarded as a complex I ∈ D(X) supported in degree zero), the k / / / / / / / / / O O O O / O O / O O / O O / DEFORMATION OF QUOTIENTS ON A PRODUCT 3 vector space map β(I) : Hom(LX, I) → Hom(N, π∗ 1I ⊗L F ) is an isomorphism and the map β(I[1]) : Ext1(LX, I) → Ext1(N, π∗ 1I ⊗L F ) is injective. If Y is Gorenstein projective, then under some technical hypotheses, we show in The- ) is represented orem 4.6, using Serre duality, that the functor HomD(X×Y )(N, π∗ by 1F ⊗L E := R H om(R π1∗ R H om(N, F ), OX ) ∈ D(X) and that the D(X) morphism E → LX obtained from Yoneda's Lemma is an isomorphism on H 0 and surjective on H −1. If we assume furthermore that N is locally free and Y is a curve, then we will show that E is of perfect amplitude ⊆ [−1, 0] and hence E → LX is a perfect obstruction theory (POT) in the sense of [BF]. Similar results have appeared in [CFK] and [MO], but neither of these references ex- plains the general mechanism for producing the derived category morphism E → LQ "intrinsically" using only the universal sequence on X × Y . In any case, the methods used here are completely different and should serve to clarify the construction of this POT and its variants, which are used in various places in the literature ([Gil], [MO], [MOP], . . . ). Acknowledgements. Johan de Jong suggested the general idea of the reduced Atiyah class to me. After thinking about it for a while, I realized that this was already implicit in (IV.3.2). I owe a debt of gratitude to Luc Illusie since most of the results presented here are simple applications of the theory developed in his book [Ill]. Conventions. We work in a fixed topos T , as in Chapter IV, throughout; in all the applications T is the topos of sheaves on a topological space. We abbreviate "ring object of T " to "ring," etc. When we speak of H i of a double (triple, etc.) complex, we mean H i of the associated total complex; we define "quasi-isomorphism" of such complexes accordingly. For a ring homomorphism A → B, we write I∆ ⊆ B ⊗A B for the kernel of the multiplication map. For a B module M , let P 1 B/A(M ) := (B ⊗A B/I 2 ∆) ⊗B M where we use the right B module structure (restriction of scalars along b 7→ 1 ⊗ b) on (B ⊗A B/I 2 ∆) to define the tensor product over B, then we regard the result as a B module via the left B module structure (restriction of scalars along b 7→ b⊗1). "Extension" always means "square zero extension" (surjective map of algebras whose kernel is a square zero ideal). We use kn to denote the functor from graded modules (resp. complexes of graded modules, the derived category of the category of graded modules, . . . ) over a graded ring B to modules over the degree zero part of B given by taking the degree n part of a module (resp. complex of modules; this is exact functor). A retract of a morphism f : A → B (in any category) is a morphism s : B → A such that sf = IdA, while a section of f is a morphism s : B → A such that f s = IdB. These are dual: s is a retract of f iff sop : A → B is a section of f op : B → A in the opposite category. 4 W. D. GILLAM 1. Deformation of quotients on a coproduct 1.1. Setup. The following basic setup will be used throughout this paper. Let A → B0 be a ring homomorphism and assume B0 = B1 ⊗A B2 is the coproduct of A algebras B1, B2. Let E2 be a B2 module and let E0 := B0 ⊗B2 E2 = B1 ⊗A E2 be the B0 module obtained from E2 by extension of scalars. Let (1.1.1) 0 / N0 / E0 f0 / F0 / 0 be an exact sequence of B0 modules and let (1.1.2) B1 = 0 / I1 / B′ 1 / 0 / B1 `BBBBBBBB A be an A extension of B1 by a B1 module I1. Assume that (1.1.3) Let I = I1 ⊗A B2 and let T orA 1 (B1, B2) = 0. (1.1.4) B = 0 / I / B / 0 / B0 `AAAAAAAA A be the A extension of B0 by I obtained from (1.1.2) by applying ⊗AB2. Let E = B′ (1.1.5) 1 ⊗A B2 and assume that T orA 1 (B1, E2) = 0 so we have an exact sequence (1.1.6) E = ( 0 / I ⊗B0 E0 / E / E0 / 0 ) of B modules obtained from (1.1.2) by taking ⊗AE2. (As usual, we identify B0 modules with B modules annihilated by I.) 1.2. Problem. Fix a B0 module K and a morphism u0 : I ⊗B0 F0 → K of B0 modules. Consider the problem of finding a B module extension F of F0 by K and a morphism of extensions f : E → F of the form: 0 0 / I ⊗B0 E0 E / E0 u0(I⊗f0) f0 / K / F / F0 0 / 0 The basic result concerning this problem is the following "classical" theorem, which we will prove in (1.6) and again in (1.8). / / / / / / / / ` O O / / / / ` O O / / / / / / / / / / / / / / DEFORMATION OF QUOTIENTS ON A PRODUCT 5 1.3. Theorem. There is an obstruction ω ∈ Ext1 B0(N0, K) whose vanishing is necessary and sufficient for the existence of a solution to (1.2). When ω = 0, the set of solutions to (1.2) is a torsor under HomB0(N0, K). 1.4. Remark. An extreme case of the problem (1.2) occurs when u0 = 0 (equivalently u0(I ⊗ f0) = 0). In this case f0 always factors through F → F0 and we see that a solution to the problem is the same thing as a morphism of extensions (1.4.1) 0 0 / N0 E0 / F0 / 0 / K / F / F0 / 0. Of course, every such morphism is obtained by pushing out the top row along a unique morphism N0 → K so in this case it is clear that the obstruction vanishes and, in fact, we have a natural trivialization of the torsor under Hom(N0, K) because we have a natural base point obtained from the split extension. 1.5. Remark. Assume F0 is flat over B1. Then by the flatness criterion, F will be flat over B′ 1 iff the natural map u(F ) : I ⊗B0 F0 → K (IV.3.1) is an isomorphism (note that I ⊗B0 F0 = I1 ⊗B1 F0). Using this isomorphism to make the identification I ⊗B0 F0 = K, we see that the problem of lifting f0 : E0 → F0 to a surjection f : E → F with F flat over B′ 1 is equivalent to the special case of (1.2) where K = I ⊗B0 F0 and u0 is the identity. In this case, we refer to a solution of (1.2) as a flat deformation of f0 over B. 1.6. The obstruction ω ∈ Ext1 L be the kernel of the composition E → E0 → F0. We have a commutative diagram B0(N0, K) can be constructed elementarily as follows. Let (1.6.1) 0 0 0 / I ⊗B0 E0 / I ⊗B0 E0 0 / L E 0 / F0 0 / N0 E0 f0 F0 / 0 0 / 0 with exact columns and rows. By pushing out the top row along u0(I ⊗ f0) we obtain a morphism of extensions 0 0 (1.6.2) 0 0 / I ⊗B0 E0 / L N0 u0(I⊗f0) / K / M / N0 / 0 / 0 / / / / / / / / / / / / / / / / / / / / / / / / / / / / / / / 6 W. D. GILLAM where M := K ⊕I⊗E0 L. Note that M is a B module (i.e. IM = 0): [ik, il] = [0, il] = [if0(l), 0] = 0, so the bottom row defines a B0 module extension. Let ω ∈ Ext1 B0(N0, K) be the class of this extension. I claim that the vanishing of ω is necessary and sufficient for the existence of a solution to (1.2), and that, in fact, splittings of the bottom row in (1.6.2) (which are a pseudo-torsor under HomB0(N0, K)) correspond bijectively to solutions of (1.2). Given a solution f : E → F to (1.2) we obtain a commutative diagram (1.6.3) 0 0 0 / I ⊗B0 E0 / N / N0 / I ⊗B0 E0 u0(I⊗f0) / K / E f / F E0 f0 / F0 0 / 0 / 0 with exact rows by pulling back the extension in the middle row along N0 → E0. Note that N → E is monic and N ⊂ L ⊂ E, but N need not be contained in the kernel of f . Given n ∈ N0, choose a (local) lift n ∈ N . I claim that n 7→ [−f (n), n] ∈ M is well-defined (independent of the choice of lift n ∈ N ), in which case it will clearly provide a splitting of the bottom row in (1.6.2). Indeed, if n′ is another (local) lift, then n′ = n + ǫ for some ǫ ∈ I ⊗B0 E0 and we have [−f (n + ǫ), n + ǫ] = [−f (n) − u0(I ⊗ f0)(ǫ), n + ǫ] = [−f (n), n] ∈ M. Conversely, notice that a splitting of the bottom row in (1.6.2) is the same thing as a map s : L → K making the diagram (1.6.4) I ⊗B0 E0 / L zvvvvvvvvvv s u0(I⊗f0) K commute. By pushing out the sequence defining L along s we obtain a commutative diagram (1.6.5) 0 0 0 / I ⊗B0 E0 E / E0 / L s / K / E / F f0 / F0 / F0 / 0 / 0 / 0 with exact rows, where we have set F := K ⊕L E. The map from the top row to the bottom row is evidently a solution to (1.2). It is straightforward to check that the two constructions are inverse. / / / / / / / / / / / / / / / / / z / / / / / / / / / / / / / DEFORMATION OF QUOTIENTS ON A PRODUCT 7 1.7. Remark. Assume u0 is surjective, hence u0(I ⊗ f0) is also surjective. Let S be its kernel. Given any solution f : E → F to (1.2), set N := Ker f . We have a commutative diagram (1.7.1) 0 0 0 0 / S 0 0 / N / N0 / I ⊗B0 E0 / E / E0 / K 0 F 0 F0 0 / 0 / 0 0 with exact rows and columns. Suppose f ∗ : E → F ∗ is some fixed solution to (1.2). Then by following through the above discussion we see that the bijection between solutions to (1.2) and HomB0(N0, K) (using f ∗ as a basepoint to trivialize the torsor) is realized by (1.7.2) (f : E → F ) 7→ S / N F ∗ / F0. Here, the top row on the right is a complex quasi-isomorphic to N0 and the bottom row on the right is a complex quasi-isomorphic to K. The map N → F ∗ is the composition of the inclusion N → E and f : E → F ∗. 1.8. Illusie constructs the obstruction ω, following a trick of Nagata, by translating the problem (1.2) into a problem of graded algebra extensions and then using the machinery of the graded cotangent complex. Let B[E] := Sym∗ B E/ Sym>1 B E be the algebra of dual numbers on E over B, graded as usual so that E is the degree one component; define graded algebras B0[E0] and B0[F0] similarly. There is an obvious diagram (1.8.1) B[E] / B0[E0] / B0[F0] of graded rings. Regard K as a graded B0[F0] module supported in degree one (hence annihilated by F0). A solution to (1.2) is the same thing as an extension of graded algebras (1.8.2) 0 / K / G / 0 / B0[F0] bEEEEEEEEE B[E] where induced map I ⊗B0 E0 → E → K is u0(I ⊗ f0). Indeed, the bijection is given by taking the degree one part of G (the extension (1.8.2) is uninteresting in all other degrees / / / / / / / / / / / / / / / / / / / / / / / b O O 8 W. D. GILLAM since K is supported in degree one, so we can always write G = B0[F ] for some B module F ). Graded algebra extensions (1.8.2) form a graded B0[E0] module Exalgr B[E](B0[F0], K). By the fundamental theorem of the graded cotangent complex (IV.2.4.2) and (IV.1.2.2.1) there are isomorphisms Exalgr (1.8.3) B[E](B0[F0], K) = Ext1 = Ext1 B0[F0](Lgr B0(k1Lgr B0[F0]/B[E], K)gr B0[F0]/B[E], K). From (IV.2.2.5) and the triangle associated to B0 → B0[E0] → B0[F0] we obtain a natural isomorphism (1.8.4) k1Lgr B0[F0]/B0[E0] = N0[1]. Using (1.8.4), the degree one part of the distinguished triangle of cotangent complexes associated to (1.8.1) is a triangle in D(B0) that can be written (1.8.5) k1Lgr B0[E0]/B[E] ⊗L B0[E0] B0[F0] / k1Lgr B0[F0]/B[E] / N0[1] . Since B[E] → B0[E0] is a square zero extension with kernel I ⊕ (I ⊗B0 E0) (in gradings 0, 1) we have τ≥−1Lgr B0[E0]/B[E] = (I ⊕ (I ⊗B0 E0))[1] and hence (1.8.6) τ≥−1Lgr B0[E0]/B[E] ⊗L B0[E0] B0[F0] = (I ⊕ (I ⊗B0 F0))[1] (in gradings 0, 1). Applying Hom( a long exact sequence , K) to (1.8.5) and using (1.8.3) and (1.8.6) we obtain (1.8.7) 0 / Hom(N0, K) / Exalgr B[E](B0[F0], K) / Hom(I ⊗B0 F0, K) δ / Ext1(N0, K) · · · of B0 modules. A solution to (1.2) is an element of Exalgr B[E](B0[F0], K) whose image in Hom(I ⊗B0 F0, K) is u0. Evidently such a solution exists iff ω := δu0 vanishes, in which case solutions are a torsor under Hom(N0, K). As a morphism in D(B0), ω is the composition (1.8.8) N0 / k1LB0[E0]/B[E] ⊗L B0[E0] B0[F0] / I ⊗B0 F0[1] u0[1] / K[1] where the first map is the one from the triangle (1.8.5) and the second is the truncation. 1.9. As the notation suggests, the obstruction ω constructed in (1.6) coincides with the one constructed in (1.8). To see this, first recall that we natural isomorphisms k1L B0[F0]/B0[E0] = τ≥−1k1L B0[F0]/B0[E0] = N0[1], so the morphism N0 → I ⊗B0 F0[1] constructed in (1.8) using the degree one part of the transitivity triangle associated to (1.8.1) coincides with the morphism τ≥−1k1LB0[F0]/B0[N0][−1] → τ≥−1k1LB0[E0]/B[E] ⊗L B0[E0] B0[F0] = I ⊗B0 F0[1] / / / / / / / / / DEFORMATION OF QUOTIENTS ON A PRODUCT 9 obtained from the degree one part of the truncated transitivity triangle (1.9.1) τ≥−1LB0[E0]/B[E] ⊗L B0[E0] B0[F0] / τ≥−1L B0[F0]/B[E] / τ≥−1L B0[F0]/B0[E0]. The graded algebra maps appearing in (1.8.1) are all surjective, with kernels as indicated below. B0[E0] / B0[F0] 0⊕N0 :ttttttttt I⊕L I⊕(I⊗B0 E0) B[E] In each case, the direct summand decomposition corresponds to the splitting into the degree 0, 1 parts, respectively. Recall that (I ⊕ (I ⊗B0 E0)) ⊗B0[E0] B0[F0] = I ⊕ (I ⊗B0 E0) = N0 · (I ⊕ (I ⊗B0 E0)) I ⊕ (I ⊗B0 E0) 0 ⊕ (I ⊗B0 N0) = I ⊕ (I ⊗B0 F0). By standard facts about the truncated cotangent complex (III.1.3), the triangle (1.9.1) is naturally identified, after applying [−1], with (the triangle associated to) the short exact sequence (1.9.2) 0 / I ⊕ (I ⊗B0 F0) I ⊕ L (I ⊕ L)2 / 0 ⊕ N0 / 0 of graded B0[F0] modules. Note that I ⊕ (I ⊗B0 E0) ⊂ B[E] and 0 ⊕ N0 ⊆ B0[E0] are square zero ideals, while (I ⊕ L)2 = 0 ⊕ IL ⊆ I ⊕ L. The degree one part of the exact sequence (1.9.2) is therefore (1.9.3) 0 / I ⊗B0 F0 / L/IL / N0 / 0. There is a natural isomorphism I ⊗B0 N0 = IL obtained by mapping i ⊗ n to il where l is a (local) lift of n under the surjection L → N0 (the choice of local lift is irrelavant since the kernel I ⊗B0 E0 of L → N0 is annihilated by I). In fact, (1.9.3) is part of the / / / O O : / / / / / / / / / 10 W. D. GILLAM commuatative diagram (1.9.4) 0 / I ⊗B0 N0 / I ⊗B0 E0 I⊗f0 0 0 0 0 IL / L / 0 / N0 / 0 0 / I ⊗B0 F0 / L/IL / N0 0 0 0 with exact columns and rows. As in any map of extensions with isomorphic cokernels, the bottom left square in (1.9.4) is a pushout. Therefore, upon pushing out the bottom row along u0 : I ⊗B0 F0 → K, the resulting extension of N0 by K coincides with the one on the bottom row of (1.6.2). Note that L/IL is manifestly a B0 module (annihilated by I), so this gives another proof that the bottom row of (1.6.2) is a B0 module extension. On the other hand, the map N0 → K[1] obtained from this extension coincides with the composition of the map N0 → I ⊗B0 F0[1] (obtained from the truncated triangle (1.9.1)) and u0[1]. 1.10. So far we have only reviewed a special case of (IV.3.2). We have not yet made use of the fact that B0 = B1 ⊗A B2 is a coproduct or the fact that the extension B (1.1.4) of B0 is obtained by base change from the extension B1 (1.1.2) of B1. The rest of this section is devoted to giving an alternative description of the morphism ω, assuming B1 is flat over A, in terms of the class (1.10.1) e(B1) ∈ Ext1 B1(LB1/A, I1) corresponding to the extension (1.1.2) under the isomorphism provided by the fundamental theorem of the cotangent complex (III.1.2.3) and the "reduced Atiyah class" of the quotient map f0, to be introduced below. 1.11. Assumption. We assume throughout that B1 is flat over A, and hence B0 is flat over B2. Note that this assumption implies the assumptions (1.1.3) and (1.1.5) of (1.1). It also implies that B1 and B2 are tor-independent over A, in the sense that T orA >0(B1, B2) = 0. It is possible to get away with significantly less, but the concomitant complications would make the exposition even more obtuse than it already is. In the applications of Section 4, the ring A will have tor dimension zero, hence B1 will be trivially A flat. / / / / / / / / / / / DEFORMATION OF QUOTIENTS ON A PRODUCT 11 1.12. By taking the degree one part k1 of the map of cotangent complex transitivity triangles associated to (1.12.1) B[E] / B0[E0] / B0[F0] we obtain a commutative diagram B2[E2] / B0[E0] / B0[F0] (1.12.2) in D(B0). N0 N0 / k1Lgr B0[E0]/B[E] ⊗L B0[E0] B0[F0] / k1Lgr B0[E0]/B2[E2] ⊗L B0[E0] B0[F0] From the commutative diagram of graded rings (1.12.3) B :tttttttttt $JJJJJJJJJJ B0 B2 B[E] :vvvvvvvvv B2[E2] $HHHHHHHHH B0[E0] and the naturality of the (graded) cotangent complex, we obtain a commutative diagram (1.12.4) L B0/B ⊗L B0 B0[E0] / Lgr B0[E0]/B[E] LB0/B2 ⊗L B0 B0[E0] / Lgr B0[E0]/B2[E2] in D(B0[E0])gr. Since the front square in (1.12.3) is cocartesian: B0[E0] = B0 ⊗B2 B2[E2], it follows from the base change properties of the cotangent complex (II.2.2) and the fact that B0 is flat over B2 (1.11) that the bottom arrow in (1.12.4) is an isomorphism.2 By applying ⊗L B0[E0]B0[F0] to (1.12.4) and taking k1, we obtain a commutative diagram (1.12.5) L B0/B ⊗L B0 F0 L B0/B2 ⊗L B0 F0 / k1Lgr B0[E0]/B[E] ⊗L B0[E0] B0[F0] / k1Lgr B0[E0]/B2[E2] ⊗L B0[E0] B0[F0] 2All that is really needed is that T orB2 >0(B0, E2) = 0. / / O O / / / / O O $ : / / $ / / : / O O / O O / O O / O O 12 W. D. GILLAM in D(B0), where the bottom arrow is an isomorphism. From the maps of cotangent complex transitivity triangles associated to the diagram of rings (1.12.6) A A A / B′ 1 / B1 / B / B0 / B2 B0 we obtain a commutative diagram (1.12.7) L B1/A ⊗L B1 B0 L B1/B′ 1 ⊗L B1 B0 I1 ⊗B1 B0[1] L B0/A (QQQQQQQQQQQQQQ L B0/B / I[1] L B0/B2 in D(B0). The two right horizontal arrows are the truncations τ≥−1 and the composition given by the top row is e(B1) ⊗B1 B0, where e(B1) ∈ Ext1(LB1/A, I1) corresponds to the algebra extension B1 (1.1.2) under the Fundamental Theorem (III.1.2.3) . Under the tor-independence assumption (1.11), the base change properties of the cotan- gent complex (II.2.2) imply that the composition LB1/A ⊗L B1 B0 → LB0/B2 in this diagram is an isomorphism. Combining this isomorphism with the one given by the bottom row of (1.12.5) yields a natural isomorphism B0[E0]/B2[E2] ⊗L B0[E0] B0[F0] = L B1/A ⊗L (1.12.8) B1 F0, k1Lgr hence the bottom arrow in (1.12.2) may be viewed as a morphism (1.12.9) ra(f0) : N0 → LB1/A ⊗L B1 F0 which we will call the reduced Atiyah class of the quotient f0. It will be further studied in Section 2. By assembling the diagrams (1.12.2), (1.12.5), and (1.12.7)⊗L B0F0 and using the isomor- phisms noted above, we obtain a commutative diagram (1.12.10) N0 k1Lgr B0[E0]/B[E] ⊗L B0[E0] B0[F0] / I ⊗L B0 F0[1] (QQQQQQQQQQQQQQQQ ra(f0) 5jjjjjjjjjjjjjjj e(B 1)⊗B1 F0 L B1/A ⊗L B1 F0 in D(B0). The composition of the top row and u0[1] is the obstruction ω (1.8.8). We have proved the following (c.f. (IV.3.1.8), (IV.3.2.14), and Theorem 2.9 in [HT]): / / / / / / / O O / / / / / / ( / O O / / ( / 5 DEFORMATION OF QUOTIENTS ON A PRODUCT 13 1.13. Theorem. Assume B1 is flat over A. Then the obstruction ω ∈ Ext1(N0, K) of (1.3) can be written as the composition N0 / L B1/A ⊗L B1 F0 / I ⊗B0 F0[1] u0[1] / K[1], where the first arrow is given by the reduced Atiyah class of f0 (1.12.9) and the second is e(B1) ⊗B1 F0, where e(B1) ∈ Ext1(LB1/A, I1) corresponds to B1 under the Fundamental Theorem (III.1.2.3). 1.14. For later use, we will now determine the map induced on H 0 by the reduced Atiyah class N0 → L B1 F0 (1.12.9). Recall that this map was constructed from the natural diagram of solid arrows B1/A ⊗L (1.14.1) N0 / k1Lgr B0[E0]/B2[E2] ⊗L B0[E0] B0[F0] ∼= ra(f0) LB0/B2 ⊗L B0 F0 ∼= LB1/A ⊗L B1 F0 and the fact that the indicated arrows are isomorphisms under the assumption (1.11) that B1 is flat over A. These isomorphisms were obtained from the base change property (II.2.2) of the cotangent complex. Note that, even without the flatness assumption (1.11) the indicated arrows induce isomorphisms on H 0 because Kahler differentials commute with all direct limits, in particular with pushouts. Taking H 0 of (1.14.1), we obtain the diagram: (1.14.2) N0 / k1ΩB0[E0]/B2[E2] ⊗B0[E0] B0[F0] ΩB0/B2 ⊗B0 F0 ΩB1/A ⊗B1 F0 The horizontal arrow is just the "connecting homomorphism" from the long exact se- quence obtained from the triangle associated to the bottom row of (1.12.1) (note N0 = H −1(Lgr B0[F0]/B0[E0])). Viewing N0 as a submodule of E0 = B1 ⊗A E2, this map is given by b1 ⊗ e 7→ d(b1 ⊗ e) ⊗ 1. Note b1 ⊗ e is in the degree one part of B0[E0], so d(b1 ⊗ e) is in k1ΩB0[E0]/B0[F0]. Notice that d(b1 ⊗ e) ⊗ 1 = d((1 ⊗ e)(b1 ⊗ 1)) ⊗ 1 = (1 ⊗ e) · d(b1 ⊗ 1) ⊗ 1 = d(b1 ⊗ 1) ⊗ f0(1 ⊗ e) / / / " " / O O O O / 14 W. D. GILLAM is the image of db1 ⊗ f0(1 ⊗ e) under the vertical isomorphisms, so the map induced on H 0 by the reduced Atiyah class is given by (1.14.3) H 0(ra(f0)) : N0 → ΩB1/A ⊗B1 F0 b1 ⊗ e 7→ db1 ⊗ f0(1 ⊗ e). Notice that this map is B0 linear because b1 ⊗ e ∈ N0 ⊂ E0 = B1 ⊗A E2. 2. Reduced Atiyah class In this section, we give an independent treatment of the reduced Atiyah class, first introduced in (1.12.9) above. The results of this section are not needed elsewhere; our purpose here is only to further examine the reduced Atiyah class and give an alternative construction of it via principal parts. 2.1. Let B → C be a ring homomorphism and let E be a C module. Recall (III.1.2.6.3) the principal parts sequence (2.1.1) 0 → ΩC/B ⊗C E → P 1 C/B(E) → E → 0 of E. The cokernel map in (2.1.1) admits a natural B linear section (2.1.2) s : E → P 1 C/B(E) e 7→ 1 ⊗ 1 ⊗ e, which yields a B module splitting of (2.1.1). The map s is generally not C linear, since s(c · e) − c · s(e) = 1 ⊗ 1 ⊗ c · e − c · (1 ⊗ 1 ⊗ e) = 1 ⊗ c ⊗ e − c ⊗ 1 ⊗ e = dc ⊗ e. Now suppose that E = C ⊗B M is obtained by extension of scalars from a B module M . Then I claim the map t : E → P 1 C/B(E) given by (2.1.3) t : C ⊗B M → ((C ⊗B C)/I 2 c ⊗ m 7→ c ⊗ 1 ⊗ 1 ⊗ m ∆) ⊗C (C ⊗B M ) provides a C linear splitting (2.1.4) P 1 C/B(E) = (ΩC/B ⊗ E) ⊕ E of (2.1.1). Obviously this is a splitting; it is C linear by the following computation: t(c′ · c ⊗ m) − c′ · t(c ⊗ m) = t(cc′ ⊗ m) − c′ · (c ⊗ 1 ⊗ 1 ⊗ m) = c′c ⊗ 1 ⊗ 1 ⊗ m − c′c ⊗ 1 ⊗ 1 ⊗ m = 0. DEFORMATION OF QUOTIENTS ON A PRODUCT 15 2.2. Continue to assume E = C ⊗B M and suppose (2.2.1) 0 / N / E f / F / 0 is an exact sequence of C modules. From the naturality of the principal parts sequence we obtain a commutative diagram (2.2.2) 0 0 0 0 0 / ΩC/B ⊗C F / P 1 C/B(F ) ΩC/B ⊗f / ΩC/B ⊗C E / P 1 C/B(E) / ΩC/B ⊗C N / P 1 C/B(N ) 0 / F f E N 0 / 0 / 0 / 0 with exact rows and columns. Using the splitting (2.1.4) and the diagram (2.2.2) we obtain a morphism of double complexes of C modules (2.2.3) ΩC/B ⊗C E ΩC/B ⊗C E ΩC/B ⊗C E β ΩC/B ⊗C N / P 1 C/B(N ) ΩC/B ⊗C N (where ΩC/B ⊗C N is placed in degree (0, −1)) by projecting onto the quotient complex. The induced map on H 0 is given by (2.2.4) H 0(β) : N → ΩC/B ⊗C F c ⊗ m 7→ dc ⊗ f (1 ⊗ m). Indeed, H 0(β)(c ⊗ m) can be computed by choosing (locally) a lift of c ⊗ m to P 1 C/B(E), then applying ΩC/B ⊗ f to the component of this lift in ΩC/B ⊗B E under the splitting (2.1.4) (the result will be independent of the lift of c ⊗ m ∈ N ⊂ E and will yield a C linear map). We may as well choose the lift systematically be means of the B linear map s in (2.1.2). Since the splitting (2.1.4) is defined via t, the ΩC/B ⊗ E component of s(c ⊗ m) is given by s(c ⊗ m) − t(c ⊗ m). Now we simply compute H 0(β)(c ⊗ m) = (ΩC/B ⊗ f )(s(c ⊗ m) − t(c ⊗ m)) = (ΩC/B ⊗ f )(1 ⊗ 1 ⊗ c ⊗ m − c ⊗ 1 ⊗ 1 ⊗ m) = (ΩC/B ⊗ f )(1 ⊗ c ⊗ 1 ⊗ m − c ⊗ 1 ⊗ ⊗1 ⊗ m) = (ΩC/B ⊗ f )(dc ⊗ 1 ⊗ m) = dc ⊗ f (1 ⊗ m). Notice that this map is the same as the one in (1.14.3). / / / / / O O / O O / O O / / O O / / / O O O O / / O O / / / O O O O / O O O O / O O / / O O 16 W. D. GILLAM 2.3. In order to generalize the results of (2.1) and (2.2) by replacing Kahler differentials with the cotangent complex it becomes necessary to impose some additional hypotheses. This is because, even though P 1 C/B(E) splits naturally when E is pulled back from B, there is no reason to think that P 1 P/B(E) splits when P = PBC is the standard simplicial resolution of C by free B algebras and E is viewed as a P module by restriction of scalars along the augmentation P → C. What is needed is a particularly nice choice of resolution of C. We adopt, for the remainder of this section, the setup of (1.1) and we assume, as in (1.11) that B1 is flat over A. To simplify notation, we will just write N, E = B1 ⊗A B2, F (as in the previous section) instead of N0, E0, F0, since we will not consider here any problem of lifting module maps over algebra extensions. Let P1 = PAB1 be the standard simplicial resolution of B1 by free A algebras. Since the homology B1 of P1 (viewing P1 just as a complex of A modules) is A flat, taking homology commutes with tensoring with any A module. In particular, P := P1 ⊗A B2 is a simplicial resolution of B0 = B1 ⊗A B2 by free B2 algebras and hence the cotangent complex LB0/B2 is the image in D(B0) of ΩP/B2 ⊗P B0 (II.2.1.2). Note that the natural isomorphism ΩP/B2 = ΩP1/A ⊗A B2 corresponds to the natural derived category isomorphism (2.3.1) L B0/B2 = L B1/A ⊗L B1 B0 obtained from the base change theorem (II.2.2). (Actually, this is more-or-less how one proves the base change properties of the cotangent complex.) Set V = P1 ⊗A E2, so V is obtained from the B2 module E2 by extension of scalars along B2 → P and hence, just as in (2.1), we obtain a natural map t : V → P 1 (V ) yielding a splitting P/B2 (2.3.2) P 1 P/B2 (V ) = ΩP/B2 ⊗P V ⊕ V of the principal parts sequence of V . The augmentation map V → B1 ⊗A E2 = E is a quasi-isomorphism of P modules (viewing E as a P module via restriction of scalars along P → B0), again because the homology of P1 is flat over A. We have a morphism of P module extensions (2.3.3) 0 0 / ΩP/B2 ⊗P V / (ΩP/B2 ⊗P V ) ⊕ V ≃ ≃ / ΩP/B2 ⊗P E / P 1 P/B2 (E) 0 / V ≃ / E / 0, where the vertical arrows are quasi-isomorphisms (by the Five Lemma and the fact that ΩP/B2 is a flat P module) so that the parts sequence of E is quasi-isomorphic to a split sequence of P modules. / / / / / / / / / DEFORMATION OF QUOTIENTS ON A PRODUCT 17 As in (2.2), the exact sequence (2.2.1) yields a commutative diagram (2.3.4) 0 0 0 0 0 / ΩP/B2 ⊗P F / P 1 P/B2 (F ) / ΩP/B2 ⊗P E / P 1 P/B2 (E) / ΩP/B2 ⊗P N / P 1 P/B2 (N ) 0 0 0 / F E N 0 / 0 / 0 / 0 of P modules with exact rows and columns (except here ΩP/B2 is a flat P module, so the complex is "more exact"). Consider the two term complex (2.3.5) W := [ ΩP/B2 ⊗P F ⊕ V / P 1 P/B2 (F ) ] of P modules (in degrees 0, 1), where the map V → P 1 (F ) is the composition of the map V → P 1 (F ). From the exactness of (2.3.4) and the quasi-isomorphisms in (2.3.3) it follows that W is quasi-isomorphic to N .3 This is clear once we note that W is quasi-isomorphic to the double complex below. (E) appearing in (2.3.3) and the natural map P 1 P/B2 P/B2 (E) → P 1 P/B2 P/B2 (2.3.6) (ΩP/B2 ⊗P E) ⊕ V / P 1 P/B2 (E) ΩP/B2 ⊗P N / P 1 P/B2 (N ) There is an obvious morphism of complexes β : W → ΩP/B2 ⊗P F obtained by projecting to the quotient complex. After extending scalars along P → B0 and using the isomorphism (2.3.1), we may view β as a D(B0) morphism (2.3.7) β : N → LB1/A ⊗L B1 F called, suggestively, the reduced Atiyah class of the quotient f : E → F . 2.4. We now explain the appellation "reduced Atiyah class." Notice that the projection map (2.4.1) [ ΩP/B2 ⊗P F / P 1 P/B2 (F ) ] ΩP/B2 ⊗P F obviously factors through β by including the domain complex as a subcomplex of W . 3This is probably bad terminology. We don't mean that the homology of W is isomorphic to N as a P module, but only that the homology of W is quasi-isomorphic to N as simplicial abelian groups. In other words, the corresponding double complex of abelian groups has homology N . / O O / O O / O O / / O O / / / O O O O / / O O / / / O O O O / O O O O O O / / O O / O O / / / 18 W. D. GILLAM By (IV.2.3.7.3) the image of this projection map in the derived category D(B0) coincides (up to a shift) with the Atiyah class (2.4.2) atB0/B2(F ) : F → LB0/B2 ⊗L B0 F [1] of F relative to B2 → B0 (see (IV.2.3) for the construction of the Atiyah class via the graded cotangent complex). Note that the aforementioned inclusion of subcomplexes is naturally isomorphic in D(B0) to the map F [−1] → N obtained from (2.2.1). Hence, in D(B0), we obtain a commutative diagram (2.4.3) F [−1] N atB0/B2 (F )[−1] &MMMMMMMMMM β / L B0/B2 ⊗L F factoring the (shift of the) Atiyah class of F through the reduced Atiyah class β. 2.5. The relationship between the reduced Atiyah class of (2.3) and the one introduced in (1.12.9) can be explained as follows. From the transitivity triangle of graded cotangent complexes associated to the diagram of graded rings (2.5.1) B2[E2] / B0[E] / B0[F ] B2 / B0 B0[F ] we obtain a commutative diagram (2.5.2) N / k1LB0[E]/B2[E2] ⊗L B0[E] B0[F ] ∼= F [−1] atB0/B2 (F )[−1] / LB0/B2 ⊗L B0 F where the bottom horizontal arrow is the Atiyah class of F relative to B2 → B0. Since B0[E] = B2[E2] ⊗B2 B0 and B0 is flat over B2, it follows from the base change theorem (II.2.2) that the right vertical arrow is an isomorphism. 2.6. Theorem. The diagram (2.6.1) N k1LB0[E]/B2[E2] ⊗L B0[E] B0[F ] )RRRRRRRRRRRRRRRRR β ∼= F [−1] / LB0/B2 ⊗L B0 F obtained from (2.5.2) by inserting the reduced Atiyah class is commutative. & / / / O O / / / O O / O O / O O / / ) O O / O O DEFORMATION OF QUOTIENTS ON A PRODUCT 19 Proof. We already saw in (2.4.3) that the lower triangle commutes, so we focus now on the upper triangle. Consistently with the previous notation, set := PAB1 P1 P := P1 ⊗A B2 V := P1 ⊗A E2 Q := P ∆gr P P [F ] R := P ∆gr P [V ]P [F ], where, for example, P ∆gr of P [F ] by graded free P algebras. The diagram of graded simplicial rings P P [F ] is the diagonal of the standard bisimplicial graded resolution (2.6.2) B2[E2] / P [V ] / R B2 / P Q admits a natural augmentation quasi-isomorphism to the diagram (2.5.1) and the transi- tivity triangle of graded cotangent complexes associated to (2.5.1) is naturally identified with the map of D(B0[F ]) triangles associated to the map of short exact sequences of R modules (2.6.3) 0 / ΩP [V ]/B2[E2] ⊗P [V ] R / ΩR/B2[E2] / ΩR/P [V ] 0 / ΩP/B2 ⊗P R / ΩQ/B2 ⊗Q R ΩQ/P ⊗Q R / 0 / 0 after extending scalars along the composition R → P [F ] → B0[F ] of the augmentation maps. Note that the left vertical arrow in (2.6.3) is an isomorphism because the left square in (2.6.2) is cocartesian and note that the sequences are exact on the left because R is free term-by-term over P [V ] and Q is free term-by-term over P (II.1.1.2.13). The composition N → L B0/B2 ⊗L B0 F of the top horizontal arrow and the right vertical arrow in (2.6.1) is the image in D(B0) of the quotient complex projection (2.6.4) [ ΩP/B2 ⊗P R1 / k1ΩR/B2[E2] ] ΩP/B2 ⊗P R1 obtained from the degree one part of the diagram (2.6.3) (after extending scalars from R0 := k0R to B0 = k0B0[F ]). Note that the terms in the domain complex of (2.6.4) should be placed in degrees 0, 1 since we have natural isomorphisms k1ΩR/P [V ] ⊗R0 B0 = k1LB0[F ]/B0[E] = N [1]. From the graded ring map B2 → B2[E2] → R, the fact that ΩB2[E2]/B2 = k1ΩB2[E2]/B2 = E2 and the fact that R is free term-by-term over B2[E2], we obtain an exact sequence (2.6.5) 0 / E2 ⊗B2 R1 / k1ΩR/B2[E2] / k1ΩR/B2 / 0 / / O O / / / O O O O / / / / / / / / O O O O / / / / / / / / 20 W. D. GILLAM of R0 modules. From (II.1.2.6.7) and the proof of (IV.2.3.7.3) we obtain a morphism (2.6.6) 0 0 / ΩP/B2 ⊗P R1 / P 1 P/B2 (R1) ≃ R1 d≃ / ΩP/B2 ⊗P R1 / k1(ΩR/B2) / ΩR/P / 0 / 0 of exact sequences of R0 modules. Since P → R is a quasi-isomorphism in degree zero and P is supported in degree zero, the right vertical arrow is a quasi-isomorphism (IV.2.2.5), hence so is the middle vertical arrow. We now have a sequence of quasi-isomorphisms compatible with the projections to ΩP/B2 ⊗P R1 as follows. [ ΩP/B2 ⊗P R1 / k1ΩR/B2[E2] ] ≃ [ (ΩP/B2 ⊗P R1) ⊕ (E2 ⊗B2 R0) / k1ΩR/B2 ] ≃ [ (ΩP/B2 ⊗P R1) ⊕ (E2 ⊗B2 R0) / P 1 P/B2 (R1) ] ≃ [ (ΩP/B2 ⊗P F ) ⊕ V / P 1 P/B2 (F ) ]. Note that the last quasi-isomorphism is compatible with the quasi-isomorphism ΩP/B2 ⊗P R1 ≃ ΩP/B2 ⊗P F. The first is obtained using the sequence (2.6.5) and the second is obtained using (2.6.6). The other quasi-isomorphisms are obtained from the natural augmentation maps: for ex- ample R0 ≃ P and R1 ≃ F are the degree 0, 1 parts of the augmentation quasi-isomorphism R ≃ P [F ]. This completes the proof. 2.7. The reduced Atiyah class is functorial in various ways that we leave to the reader to make precise. We only point out that, given a morphism of B0 module extensions of the form (2.7.1) 0 0 / N E f / F / N ′ / E / F ′ 0 / 0, we have a commutative diagram in D(B0): (2.7.2) β β′ N N ′ / L B0/B2 ⊗L B0 F / L B0/B2 ⊗L B0 F ′ It is a good exercise to prove this using both constructions of the reduced Atiyah class. 2.8. Exercise. Show that the Atiyah class factors through the reduced Atiyah class as in (2.4.2) by using the constructions of the two classes via the graded cotangent complex (instead of using principal parts as we have done). / / / / / / / / / / / / / / / / / / / / / / / / / DEFORMATION OF QUOTIENTS ON A PRODUCT 21 3. Perfect quotients 3.1. We continue with the setup (1.1) and the assumptions (1.11). We also assume that the extension (1.1.2) is trivialized: (3.1.1) B1 = 0 / I1 / B1[I1] / B1 / 0 bFFFFFFFF A, and that F0 is flat over B1. The extension (1.1.4) is also naturally trivialized: B = B0[I]. Given an A algebra section s : B1 → B1[I1] of B1[I1] → B1, we obtain a flat deformation (c.f. (1.5)) s∗f0 := B1[I1] ⊗s f0 of f0 over B. Here we write B1[I1]⊗s instead of B1[I1]⊗B1 to emphasize that B1[I1] is regarded as a B1 module via restriction of scalars along s. Note that s∗E0 is naturally identified with E via the isomorphism (3.1.2) B1[I1] ⊗s (B1 ⊗A E2) → B1[I1] ⊗A E2 7→ (b1 + i1)s(b′ (b1 + i1) ⊗ b′ 1 ⊗ e2 1) ⊗ e2 of B modules. Set Fs := B1[I1] ⊗s F0. The solution s∗f0 to (1.2) can be explicitly written (3.1.3) 0 0 / I1 ⊗A E2 B1[I1] ⊗A E2 B1 ⊗A E2 I1⊗f0 / I1 ⊗B1 F0 / Fs f0 / F0 / 0 / 0 Since F0 is flat over B1, note that the kernel of E → Fs is Ns := B1[I1] ⊗s N0 and the kernel of I1 ⊗B1 E0 → I1 ⊗B1 F0 is I1 ⊗B1 N0. Let s0 : B1 → B1[I1] denote the zero section b1 7→ b1 of B1[I1] → B1. We say that the surjection f0 : E0 → F0 is a perfect quotient if the map {A algebra sections of B1[I1] → B1 } → { flat deformations of f0 over B1[I1] } is bijective for every B1 module I1. s 7→ s∗f0 Given an A algebra section s of B1[I1] → B1, let ds : B1 → I1 be the corresponding A linear derivation and let gs : ΩB1/A → I1 be the corresponding map of B1 modules. 3.2. Lemma. The map N0 → I1 ⊗B0 F0 defined by the map of complexes (3.2.1) I1 ⊗B1 N0 / Ns as in Remark 1.7 coincides with the map induced by the composition (3.2.2) N0 / LB1/A ⊗L F0 / ΩB1/A ⊗L F0 / I1 ⊗B1 F0 Fs0 / F0. / / / / b O O / / / / / / / / / / / / / / / 22 W. D. GILLAM of the truncated reduced Atiyah class of f0 and gs ⊗ F0. Both maps are given by (3.2.3) N0 → I1 ⊗B1 F0 b1 ⊗ e2 7→ dsb1 ⊗ f0(1 ⊗ e2). Proof. The maps s, gs, and ds are related by the formulas (3.2.4) s(b1) = b1 + dj(b1) gs(db1) = dj(b1), so the fact that (3.2.2) is given by (3.2.3) follows from (2.2.4). To calculate the map defined by (3.2.1), we will make use of the natural B2 linear section n 7→ 1 ⊗ n of Ns → N0. Note that the isomorphism Es ∼= Es0 obtained by identifying both Es and Es0 with E via the natural isomorphism (3.1.2) can be explicitly written: (3.2.5) (B1[I1] ⊗A B2) ⊗s (B1 ⊗A E2) → (B1[I1] ⊗A B2) ⊗s0 (B1 ⊗A E2) (b1 + i1) ⊗ b2 ⊗ b′ 1 ⊗ e2 7→ (b1 + i1)s(b′ 1) ⊗ b2 ⊗ 1 ⊗ e2. The map defined by (3.2.1) can be computed on n = b1 ⊗ e2 ∈ N = Ker f0 by following 1 ⊗ 1 ⊗ b1 ⊗ e2 through the sequence of maps (B1[I1] ⊗A B2) ⊗s (B1 ⊗A E2) ∼= (3.2.5) (B1[I1] ⊗A B2) ⊗s0 (B1 ⊗A E2) (B1[I1] ⊗A E2) ⊗s0 F0 and noting that the result is in (I1 ⊗A B2) ⊗B0 F0. Carrying this out, we find: 1 ⊗ 1 ⊗ b1 ⊗ e2 7→ s(b1) ⊗ 1 ⊗ 1 ⊗ e2 = b1 ⊗ 1 ⊗ 1 ⊗ e2 + dsb1 ⊗ 1 ⊗ 1 ⊗ e2 = 1 ⊗ 1 ⊗ b1 ⊗ e2 + dsb1 ⊗ 1 ⊗ 1 ⊗ e2 7→ 1 ⊗ 1 ⊗ f0(b1 ⊗ e2) + dsb1 ⊗ 1 ⊗ f0(1 ⊗ e2) = dsb1 ⊗ 1 ⊗ f0(1 ⊗ e2). Of course we drop the 1 in the middle when we make the natural identification (I1 ⊗A B2) ⊗B0 F0 = I1 ⊗B1 F0. We have now shown that the map induced by (3.2.1) is also given by (3.2.3), so the proof is complete. 4. Applications to the Quot scheme 4.1. Setup. We work throughout in the category Sch of schemes over a field k, which we refer to simply as "schemes". All relative constructions done without explicit reference to a morphism are assumed to be relative to the terminal object, so, for example, LX = LX/k, DEFORMATION OF QUOTIENTS ON A PRODUCT 23 X × Y = X ×k Y , etc. Let Y be a scheme, E a quasi-coherent sheaf on Y . We assume that the functor (4.1.1) Schop → Ens X 7→ { quotients of π∗ 2E on X × Y flat over X } is representable by a scheme Q (the Quot scheme). This holds, for example, if Y is projective and E is coherent, in which case the components of Q are projective [Gro]. The Quot functor (4.1.1) is representable by an algebraic space in much more generality. Assuming the basic machinery of Serre duality, etc. for algebraic spaces, the results of this section carry over to that setting as well. Since we work over a field, for schemes X and Y we have LX×Y = π∗ 1 X×Y /Y = π∗ 1 LX (the projections are flat). For an exact sequence L LX ⊕ π∗ 2 LY and (4.1.2) 0 / N f / π∗ 2E / F / 0 on X × Y (with F flat over X) we let β(f ) : N → π∗ LX ⊗L F (or just β if f is clear from 1 context) denote the reduced Atiyah class of f (1.12.9).4 Recall the natural transformation β : HomD(X)(LX , ) → HomD(X×Y )(N, π∗ 1 ⊗L F ) β(M )(g) := (π∗ 1g ⊗L F )β. from the introduction. By abuse of notation, we write f : X → Q for the morphism corresponding to the quotient f . Since all the technical results are already in place, the proof of the next theorem will amount to little more than unwinding various definitions. I view this theorem as the main result of this paper. 4.2. Theorem. The following are equivalent: (1) f is a perfect quotient. (2) β(I) is an isomorphism for every quasi-coherent sheaf I on X. If f : X → Q is formally ´etale, then these two equivalent conditions hold, and, furthermore, β(I[1]) : Ext1(LX , I) → Ext1(N, π∗ 1I ⊗ F ) is injective for every quasi-coherent sheaf I on X. Proof. For a quasi-coherent sheaf I, let ιI : X ֒→ X[I] be the trivial square zero closed immersion with ideal I. By Lemma 3.2 we have a commutative diagram (4.2.1) Hom(ΩX , I) β(I) / Hom(N, π∗ 1I ⊗ F ) { retracts of ιI } / { flat deformations of f over X[I] } 4It is not necessary to derive the tensor product in π∗ 1 M ⊗ F = π−1 M 7→ π∗ 1 M ⊗ L F because 1 M ⊗ π−1 1 OX F is an exact functor from X modules to X × Y modules since F is flat over X / / / / / / 24 W. D. GILLAM where the bottom arrow is given by pullback. By defintion, the latter is an isomorphism for every I iff f is a perfect quotient, so the equivalence of the two conditions is clear. If f : X → Q is formally ´etale, then by defintion of "formally ´etale" s 7→ f s gives a bijection between retracts of ιI and completions of the solid diagram ιI / X[I] X f Q in Sch. On the other hand, by definition of Q, such completions are the same thing as flat deformations of f over X[I]. This proves the second statement. For the final statement, note that the Fundamental Theorem of the Cotangent Complex (III.1.2.3) identifies Ext1(LX , I) with (isomorphism classes of) square zero thickenings X ֒→ X ′ of X with ideal sheaf I. According to (1.13), the map takes the class e(X ′) : LX → I[1] of the thickening X ֒→ X ′ to the obstruction β(I[1]) : Ext1(LX , I) → Ext1(N, π∗ 1I ⊗ F ) to finding a flat deformation of f over X ′. But by definition of the Quot scheme Q, finding such a flat deformation is the same thing as finding a commutative diagram of solid arrows ω = (π∗ 1e(X ′) ⊗ F )β X X ′ X f Q and by defintion of formally ´etale every such diagram can be completed to a commutative diagram as indicated by the dotted arrow. Such a dotted arrow is the same thing as a retract of X ֒→ X ′, which is the same thing as an identification of X ′ with the trivial thickening: X ′ = X[I]. If β(I[1]) failed to be injective we would therefore find some nontrivial thickening X ′ with a retract, which is absurd. 4.3. Lemma. Let Z be a scheme. Suppose F ∈ D(Z) is a complex with finite perfect am- plitude whose rank r is prime to the characteristic of k. Let L, N be arbitrary complexes of OX modules and let β : N → L ⊗L F be a morphism in D(Z). Set F ∨ := R H om(F, OZ ). For any E ∈ D(Z), there is an isomorphism HomD(Z)(N, E ⊗L F ) → HomD(Z)(N ⊗L F ∨, E), natural in E. Proof. Since F has finite perfect amplitude, we have F ⊗L F ∨ = R H om(F, F ), and we have a trace morphism and a "scalar multiplication" morphism tr : R H om(F, F ) → OZ id : OZ → R H om(F, F ) satisfying (tr)(id) = ·r. The desired isomorphism is given by g 7→ (E ⊗ tr)(g ⊗ F ∨). Its inverse is given by h 7→ (h ⊗ F )(N ⊗ r−1id). / } } _ / / > > DEFORMATION OF QUOTIENTS ON A PRODUCT 25 4.4. Lemma. Let A be an abelian category with enough injectives and let B be a full abelian subcategory of A. Let f : E → F be a map in D(A) between complexes with cohomology in B and vanishing in positive degrees. Then the following are equivalent: (1) H 0(f ) is an isomorphism. (2) Hom(F, I) → Hom(E, I) is an isomorphism for every I in B. If these equivalent conditions are satisfied, then the following are equivalent: (1) H −1(f ) is surjective. (2) Ext1(F, I) → Ext1(E, I) is injective for every I in B. Proof. This is standard. Let I → J be an injective resolution of I. Then we can compute Extn(E, I) as H n of the double complex Hom•(E, J). By first taking cohomology in the E direction, we obtain a spectral sequence Extp(H −q(E), I) ⇒ Extp+q(E, I) natural in E, I. Since H >0(E) = H >0(F ) = 0, the map f ∗ : Hom(F, I) → Hom(E, I) is identified with the B morphism H 0(f )∗ : Hom(H 0(F ), I) → Hom(H 0(E), I), so the first statement follows from Yoneda's Lemma. From the naturality of the low order terms in this spectral sequence, we obtain a commutative diagram 0 0 / Ext1(H 0(E), I) / Ext1(E, I) / Hom(H −1(E), I) / Ext2(H 0(E), I) H 0(f )∗ ∼= f ∗ H −1(f )∗ H 0(f )∗ ∼= / Ext1(H 0(F ), I) / Ext1(F, I) / Hom(H −1(F ), I) / Ext2(H 0(F ), I) with exact rows and natural in I. Assuming H 0(f ) is an isomorphism, so are the indicated arrows. By the Subtle Five Lemma,5 H −1(f )∗ is injective iff f ∗ is injective. On the other hand, H −1(f )∗ is injective for every I in B iff the B morphism H −1(f ) is surjective. 4.5. Remark. One could also use the long exact sequence obtained by applying Hom( to the map of truncation triangles , I) τ≤−1E[1] / E / τ≥0E τ≤−1F [1] / F / τ≥0F instead of the low order terms in the spectral sequence, though this amounts to the same thing. 5This isn't so subtle since we have much more exactness than necessary for the desired conclusion. / / / / / O O / O O / O O / O O / / / / 26 W. D. GILLAM 4.6. Theorem. Suppose Y is a projective Gorenstein scheme of pure dimension d, E is a coherent sheaf on Y , and f : π∗ 2E → F is a surjection of sheaves on X ×Y with F flat over X and of finite perfect amplitude. Assume that the rank of F is prime to the characteristic of k and that N := Ker f is also of finite perfect amplitude. Then the functor D(X) → Vectk I 7→ HomD(X×Y )(N, π∗ 1I ⊗L F ) is represented by E := R H om(R π1∗ R H om(N, F ), OX ) ∈ D(X). If f defines a formally ´etale map from X to the Quot scheme, then the map H 0(E → LX) is an isomorphism and the map H −1(E → LX) is surjective. Proof. By the hypotheses on Y , there is an invertible sheaf ωY on Y and an isomorphism (4.6.1) R H om(R π1∗A, B) = R π1∗ R H om(A, π∗ 1B ⊗ π∗ 2ωY [d]) (Grothendieck-Serre duality) in D(X) natural in A ∈ D(X × Y ) and B ∈ D(X). There are natural isomorphisms HomD(X×Y )(N, π∗ 1I ⊗L F ) = HomD(X×Y )(N ⊗L F ∨, π∗ 1I) = HomD(X×Y )(N ⊗L F ∨ ⊗ π∗ = HomD(X)(R π1∗(N ⊗L F ∨ ⊗ π∗ 2ωY [d], π∗ 1I ⊗ π∗ 2ωY [d]), I), 2ωY [d]) where the first isomorphism is obtained from Lemma 4.3, the second isomorphism simply reflects the fact that ⊗π∗ 2ωY is an invertible sheaf, and the third isomorphism is obtained from the Serre duality isomorphism (4.6.1) by applying R Γ and taking H 0. 2ωY [d] is an automorphism of D(X ×Y ) since π∗ Using Serre duality and the various perfection hypotheses, we obtain a sequence of natural isomorphisms E = R H om(R π1∗ R H om(N, F ), OX ) = R π1∗(R H om(R H om(N, F ), ωY [d])) = R π1∗(R H om(R H om(N, F ), OX×Y ) ⊗ π∗ = R π1∗(R H om(N ∨ ⊗L F, OX×Y ) ⊗ π∗ 2ωY [d]) = R π1∗(N ⊗L F ∨ ⊗ π∗ 2ωY [d]). 2ωY [d]) Putting this together with the isomorphism from the previous paragraph proves the first part of the theorem. The second statement follows from Theorem 4.2 and Lemma 4.4 (applied with A = ModY , B = quasi-coherent sheaves). 4.7. Lemma. Suppose Y = C is a projective curve. Then for any quasi-compact scheme X and any exact sequence (4.1.2) with F flat on X × C and N locally free, the complex E = R H om(R π1∗ R H om(N, F ), OX ) ∈ D(X) is of perfect amplitude ⊆ [−1, 0]. DEFORMATION OF QUOTIENTS ON A PRODUCT 27 Proof. It suffices to show that the complex R π1∗ R H om(N, F ) is of perfect amplitude ⊆ [0, 1]. Since N is locally free we have R H om(N, F ) = H om(N, F ). Let D ⊂ C be an effective Cartier divisor. Then we have an exact sequence (4.7.1) 0 / H om(N, F ) / H om(N, F )(D) / H om(N, F )(D)X×D / 0 on X × C (writing D for π∗ 2D to save notation). Since F is flat over X and X is quasi- compact, it follows from the basic cohomology and base change theorems that for any sufficiently positive such D we have R>0π1∗H om(N, F )(D) = 0. In this case, by Grauert's Criterion, π1∗H om(N, F )(D) will be a vector bundle on X. Since H om(N, F )(D)X×D is supported in relative dimension zero over X, we conclude similarly that R>0π1∗H om(N, F )(D)X×D = 0 and that π1∗H om(N, F )(D)X×D is a vector bundle on X. From the usual spectral sequences argument we conclude that R π1∗ R H om(N, F ) = π1∗H om(N, F )(D) → π1∗H om(N, F )(D)X×D in D(X), where the two term complex on the right is placed in degrees 0, 1. Since the two terms in this complex are vector bundles, the proof is complete. 4.8. Corollary. Suppose C is a projective Gorenstein curve, E is a vector bundle on C, and X is an open subset of Quot E on which the universal kernel N ⊆ π∗ 2E is locally free (this always holds if C is smooth). Then the map R H om(R π1∗ R H om(N, F ), OX ) → LX obtained from the reduced Atiyah class defines a perfect obstruction theory on X. Proof. Note that if C is a smooth curve, OC is a sheaf of PIDs so any (quasi-coherent) subsheaf of a vector bundle is a vector bundle. If N is locally free, then from the exact sequence (4.1.2) it is clear that F has finite perfect amplitude (⊆ [−1, 0] in fact), so this follows immediately from the previous results of this section. 4.9. Remark. If Y is smooth, then π1 : X × Y → X is also smooth, so flatness of F (hence N ) over X implies perfection of F and N by [SGA6.III.3.6]. It may be possible to weaken some of the perfection hypotheses. References [BF] K. Behrend and B. Fantechi, The intrinsic normal cone. Invent. Math. 128 (1997) 45-88. [CFK] I. Ciocan-Fontanine and M. Kapranov, Virtual fundamental classes for dg-manifolds. Unpublished manuscript. [Gil] W. Gillam, Maximal subbundles, Quot schemes, and curve counting. [Gro] A. Grothendieck, Techniques de construction et theor'emes dexistence en geometrie algebrique IV: Les schemas de Hilbert. Seminaire Bourbaki 221 (1960/61). [HT] D. Huybrechts and R. Thomas, Deformation-obstruction theory for complexes via Atiyah and Kodaira-Spencer classes. arXiv:0805.3527 L. Illusie, Complexe cotangent et d´eformations, I. Lec. Notes Math. 239. Springer-Verlag, 1971. [Ill] [MO] A. Marian and D. Oprea, Virtual intersection numbers on the Quot scheme and Vafa-Intriligator formulas. [MOP] A. Marian, D. Oprea, and R. Pandharipande, Stable quotients. [SGA6] A. Grothendieck, P. Berthelot, and L. Illusie, Seminaire de Geometrie Algebrique du Bois-Marie. Lec. Notes Math. 255, Springer-Verlag. / / / / 28 W. D. GILLAM Department of Mathematics, Brown University, Providence, RI 02912 E-mail address: [email protected]
1501.04713	2	1501	2015-01-21T02:04:09	Dual fans and mirror symmetry	[ "math.AG" ]	We show that the mirror constructions of Greene-Plesser, Berglund-Hubsch, Batryev-Borsov, Givental and Hori-Vafa can be expressed in terms of what we call dual fans. To do this, we associate to a pair of dual fans a pair of toric Landau-Ginzburg models, and we describe a process by which each of the mirror constructions listed also produces a pair of toric Landau-Ginzburg models. Replacing mirror pairs by toric Landau-Ginzburg models is reversible, and our main result is the dual fan models and the mirror pairs models coincide.	math.AG	math	Dual fans and mirror symmetry Patrick Clarke January 22, 2015 Abstract We show that the mirror constructions of Greene-Plesser, Berglund- Hübsch, Batryev-Borsov, Givental and Hori-Vafa can be expressed in terms of what we call dual fans. To do this, we associate to a pair of dual fans a pair of toric Landau-Ginzburg models, and we describe a process by which each of the mirror constructions listed also produces a pair of toric Landau-Ginzburg models. Replacing mirror pairs by toric Landau- Ginzburg models is reversible, and our main result is the dual fan models and the mirror pairs models coincide. 1 Introduction In this paper, we show that the mirror constructions of Greene-Plesser [GP90], Berglund-Hübsch [BH92], Batyrev-Borisov [BB97]1, Givental [Giv98], and Hori- Vafa [HV00] can all be cleanly described in terms of what we call dual fans. These are fans Σ ⊂ N, Σ′ ⊂ M of strongly convex rational polyhedral cones such that • M and N dual finite rank free abelian groups, and • 0 ≤ 〈m, n〉 for any m ∈ Σ′ and n ∈ Σ. Associated to an ordered pair of dual fans (Σ, Σ′) is a toric Landau-Ginzburg model. This is a morphism W(Σ′) : X(Σ) × C(Σ′) → A1 C(Σ′) 1As a consequence, Batyrev [Bat94] and Borisov [Bor93] are also described this way. 1 over C(Σ′) = Spec Z[cρ′ ]ρ′∈Σ′(1). Here X(Σ) is a toric “variety” (over Z, or C, or ...) formed in the usual way from Σ, and W(Σ′) = Xρ′∈Σ′(1) cρ′ χuρ′ for the character χuρ′ on X(Σ) corresponding to the integral generator uρ′ ∈ M of the ray ρ′. Reversing the roles of the fans (Σ′, Σ) produces the dual toric Landau-Ginzburg model W(Σ) : X(Σ′) × C(Σ) → A1 C(Σ). Depending on the mirror construction, the mirror pair produced is either • a pair or toric Landau-Ginzburg models W : X → A1 and W′ : X′ → A1 [BH92], • complete intersections Z, Z′ in a toric varieties Y and Y′ [GP90, Bat94, Bor93, BB97], or • a complete intersection Z in a toric variety Y and a toric Landau-Ginzburg model W′ : X′ → A1 [Giv98, HV00]. The complete intersections are presented as a global section of a split bundle over an ambient toric variety. For instance, Z ⊆ Y would be specified as the zero locus of some g ∈ Γ(Y, V ), where V =Lc i =1 In order to make our comparisons, to any member of a mirror pair we Li for invertible sheaves Li . introduce an auxiliary toric Landau-Ginzburg model: W : X(ΣX) × Γ → A1 Γ where • X =½ X Spec Sym• V if the member is W : X → A1, if the member is specified by g ∈ Γ(Y, V ), • Γ = Spec Z[γm]m∈Ξ, and • W =Pm∈Ξ γmχm, where • Ξ =½ characters which appear in the expansion W =Pm gmχm, characters on Spec Sym• V from Γ(Y, V ). 2 This toric Landau-Ginzburg model comes with equipped with a specialization2 given by γm 7→ gm where the gm’s come from either the expansion in characters of W =Pm gmχm or the section g =Pm gmχm. It is somewhat surprising that, provided the specialization γm 7→ gm is re- membered, nothing is lost by replacing the complete intersection or Landau- Ginzburg model by the auxiliary Landau-Ginzburg model. For instance, given W : X → A1 it is easy to see that W = Wg . In the case of the complete intersec- tion, observe that from Wg : X(ΣX)g → A1 we can recover Y, V , g and thus Z as • Y is the maximal proper toric stratum in X(ΣX)g , • V is the conormal bundle, IY/I 2 Y , of Y in X(ΣX)g , and • g = Wg since Wg ∈ IY. This way we treat any mirror pair in the same way: as a pair (W : X(ΣX) × Γ → A1 Γ, γ 7→ g ) and (W′ : X(ΣX′ ) × Γ′ → A1 Γ′, γ′ 7→ g ′). Finally, our theorem is this: Theorem 1.1. For any mirror pair, ΣX and ΣX′ are dual fans, Γ and Γ′ are affine spaces (by construction), and there are inclusions as coordinate subspaces C(ΣX′) → Γ and C(ΣX) → Γ′ such that the toric Landau-Ginzburg models given by the fans W(ΣX′ ) : X(ΣX) × C(ΣX′ ) → A1 C(ΣX′ ) and W(ΣX) : X(ΣX′ ) × C(ΣX) → A1 C(ΣX) are obtained from W : X(ΣX) × Γ → A1 Γ and W′ : X(ΣX′) × Γ′ → A1 Γ′ by base change. Furthermore, if the mirror partner producing ΣX is a Landau-Ginzberg model (as opposed to a complete intersection) then then C(ΣX) → Γ′ is an isomorphism. 2Here the word “specialization” just means “morphism,” but using it allows us to refer specifically to these morphisms. 3 Remark 1.2. Exhibiting ΣX and ΣX′ as dual fans implies that the character group MX and the one-parameter subgroups NX′ are identified. In all the con- structions here, except Greene-Plesser, this identification is built-in. For the Greene-Plesser construction, there is a unique identification that makes the the- orem true. Remark 1.3. There is no guarantee that the map given by the specialization γ 7→ g factors through C(ΣX′). Nevertheless, these constructions are described by dual fans. If the mirror object is a Landau-Ginzburg model, then C(Σ) = Γ′ so the specialization factors. In the case of a complete intersection, there is a process by which one can recover Γ from (Σ, Σ′). Simply ask “Does W(Σ′) : X(Σ)×C(Σ′) → A1 for some Y, and V ?” C(Σ′) look like it came from g ∈ Γ(Y, V ), When the answer is “Yes,” then we can recover recover Y and V as before, and enlarge the base from C(Σ′) to Γ. We conclude the introduction some speculative remarks about dual fan mir- ror symmetry. Remark 1.4. In light of Theorem 1.1 an immediate question is “To what extent do the computations normally compared in mirror symmetry match for the dual toric Landau-Ginzburg models formed from dual fans?” Theorem 1.1 provides a partial answer to this question in that in many cases both Hodge diamonds [BH92, GP90, BB97, Kra10] and Gromov-Witten invari- ants match [Giv98]. An additional question that arises when one considers the birational trans- formations that happen as one moves about within a toric variety’s GKZ fan [GKZ94] is “To what extent do mirror computations depend on Σ(1) and Σ′(1) alone?” 2 Toric geometry We review here some basic constructions in toric geometry, and give a descrip- tion of the fan of a toric variety of the form X = SpecSym• V in Corollary 2.17. 4 2.1 Fans Definition 2.1. A rational polyhedral cone σ is a subset of a Q-vector space V that can be written {r1v1 + · · · + rs vs ri ≥ 0} for a finite set of vectors v1, . . . , vs ∈ V. The vectors v1, . . . , vs are said to generate σ. Definition 2.2. A face of a rational polyhedral cone σ in V is a subset of the form {v ∈ σ ℓ(v ) = 0} for a linear functional ℓ : V → Q with the property {v ∈ σ ℓ(v ) < 0} is empty. A maximal proper face is called a facet. Definition 2.3. The dimension of a cone σ is the dimension of the vector space spanQ σ. Definition 2.4. A cone σ is called strongly convex if 0 ∈ V is a face of σ. Definition 2.5. A 1-dimensional strongly convex cone ρ ⊂ N is called ray. The monoid ρ ∩ N is isomorphic to N, and we denote its generator by uρ. Definition 2.6. Given a finitely generated free abelian group N, a fan Σ in N is a collection of strongly convex rational cones in NQ = N ⊗Z Q such that • The face of any cone is a member of Σ, and • The intersection of any two cones in Σ is a face of each. Definition 2.7. The set of k-dimensional cones in a fan Σ is denoted Σ(k). 2.2 The toric variety of a fan We recall the standard procedure [Dem70] by which one produces a scheme from a fan. Definition 2.8. Given a cone σ in V, the dual cone σ∨ is the subset of HomQ(V, Q) made up of elements { ℓ ℓ(v ) ≥ 0, ∀v ∈ σ}. 5 Definition 2.9. If (τ, +) is a commutative monoid, we denote the monoid algebra by Z[τ]. For an element of τ, we often write the associated monomial with the element as the exponent of a base variable. For instance, if v, v ′ ∈ τ, and x as the base variable we have xv · xv ′ = xv +v ′ . Definition 2.10. Denote M = HomZ(N, Z). Associated to a fan Σ ⊂ N there is an integral Noetherian scheme X(Σ) with rational functions and an open cover Q[M] Uσ = Spec Z[σ∨ ∩ M] for σ ∈ Σ. This is called the toric scheme3 of Σ. Remark 2.11. Setting Uσ = Spec C[σ∨ ∩ M] in the above definition yields the usual construction of a toric variety over C. This can also be achieved by chang- ing base from Z to C. 2.2.1 Quotient fans It is often convenient to describe a fan Σ ⊂ N as the image under a homomor- phism Q : N → N of a fan Σ ⊂ N. This makes sense provided Σ = {P(σ) ⊂ N σ ∈ Σ} is a fan. Under this assumption, the resulting toric scheme X(Σ) is the product X( Σ)/G × Spec Z[K], where K is the kernel and C is the cokernel of the map Qt : M → M, and G = Spec Z[C]. This follows easily from the fact that the assignment Λ 7→ Spec Z[Λ] is a version of Cartier duality. If one defines the Cartier dual of Z to be Gm, then this fits in perfectly with the fact that the big torus a toric scheme is the spectrum of the group ring of its characters. When the abelian group is finite, the Cartier dual equals the Pontryagin dual. 3These properties determine X(Σ). 6 2.3 Normal fans Definition 2.12. A rational convex polyhedral set P ⊆ MR is any subset of the form P = {m ∈ MR A(m) + ℓ ≥ 0} for a homomorphism A : M → Zn and an element ℓ ∈ Rn. The dimension of P is the dimension of the vector space spanR P. If dim P = dim MR, then P is called full dimensional. Definition 2.13. A full dimensional rational convex polyhedral set P defines an normal fan ΣP ⊂ N. The cones of ΣP are labeled by faces F of P : σF = {n ∈ NQ n(p − pF) ≥ 0 for all p ∈ P and pF ∈ F}. 2.4 Completely split bundles over a toric variety. Definition 2.14. Given a fan in Σ ⊂ N, to each ray ρ ∈ Σ(1) we have a torus invariant Weil divisor Dρ whose valuation on characters M → Z. is given by m 7→ 〈m, uρ〉. Theorem 2.15. [CLS11, Theorem 4.2.8 (a) and (c)] An element D =Pρ∈Σ(1) aρDρ ∈ ZΣ(1) defines a Cartier divisor on Y if and only if for each cone σ ∈ Σ there is a homomorphism aσ : σ ∩ N → Z such that aσ(uρ) = aρ for all 1-cones ρ ⊆ σ. Proposition 2.16. (line bundle fan) Let D =Pρ∈Σ(1) aρDρ ∈ ZΣ(1) be a Cartier divisor on the toric variety Y. Set V = O (D), and X = Spec Sym• V . Then X is toric with fan ΣX ⊂ N ⊕ Z. The cones of ΣX can be described in terms of those of ΣY: For each σ ∈ Σ, we have the lifted cone σ0 = the cone generated by {(uρ, aρ) ρ ∈ σ(1)} and the over cone where the second summand is the vertical ray σ = σ0 + ρvert ρvert = Q≥0 · (0, 1). 7 Proof. Choose a rational section p of O (D) whose divisor is D. Since Y is integral so is X, and we can identify the rational functions on X with Q[M ⊕ Z], where the Z summand corresponds to powers of p. Now consider X over Uσ ⊆ Y for some σ ∈ Σ. The regular functions on the affine scheme X ×Y Uσ are given by a subring of Q[M ⊕ Z]. Indeed, we may immediately restrict to Z[M ⊕Z]. The divisor of the monomial y m p ℓ on X ×Y Uσ is ℓ (Y ∩ Uσ) + Xρ∈σ(1) (uρ(m) − ℓaρ) L ×Y (Dρ ∩ Uσ) where we have written Y ∩ Uσ for the copy of Uσ in the zero section. Thus it is regular on X ×Y Uσ if and only if ℓ ≥ 0 and uρ(m) − ℓaρ ≥ 0 for all ρ ∈ σ(1). These conditions cut out a monoid in M ⊕ Z, and thus a subring of Z[M ⊕ Z]. The elements σ in (N ⊕ Z)Q which evaluate positively on this monoid are exactly the elements in the cone generated by (0, 1) and (uρ, aρ). This is the cone σ. The cone σ0 is a face of this cone, and it is straightforward to check that these cones form a fan, and the corresponding open sets U σ = X ×Y Uσ cover X. Corollary 2.17. (split bundle fan) Iterating this procedure, one can describe for V =Lc a fan i =1 X = Spec Sym• V Li where Li = O (D(i )) for a Cartier divisor D(i ) =Pρ∈Σ(1) a(i ) ρ Dρ by ΣX ⊂ N ⊕ Zc whose cones can be written in term of those of ΣY by setting ρ , . . . , a(c) σ(0,...,0) = the cone generated by {(uρ, a(1) ρ ) ρ ∈ σ(1)} and then σ(b1,...,bc ) = σ(0,...,0) +Xi Q≥0 · bi (0, e ∨ i ) where bi ∈ {0, 1} and e ∨ i is the i th standard basis vector of Qc . 2.5 Global functions on toric varieties. Proposition 2.18. (regularity) A rational function W =Pχ cχχ expanded in char- acters on a toric variety X is an element of OX(X) if and only if for each χ with non-zero coefficient in the expansion and all ρ ∈ ΣX(1) 〈χ, uρ〉 ≥ 0. 8 Corollary 2.19. (dual fans and regularity) The following are equivalent: • (Σ, Σ′) are dual fans. • W(Σ′) is a regular function on X(Σ) × C(Σ′). • W(Σ) is a regular function on X(Σ′) × C(Σ). 3 Comparison to existing constructions We now make our way through the mirror constructions, verifying Theorem 1.1. In each case the main steps we take are to • recall the original mirror construction, • identify the auxiliary data (ΣX, Γ, W, γ 7→ g , and ΣX′ , Γ′, W′, γ′ 7→ g ′), • verify ΣX and ΣX′ are dual fans, • define inclusions C(Σ′) → Γ and C(Σ) → Γ′, and finally • we verify that the LGs given by the dual fans ΣX and ΣX′ are produced by base change to C(Σ′) and C(Σ) from the auxiliary ones. 3.1 The quintic threefold [CdlOGP91] Here both sides of the mirror are hypersurfaces, so Γ and Γ′ are the affine spaces of global sections, and W and W′ are the universal sections considered as functions. Definition 3.1. (Candelas-de la Ossa-Green-Parks family of quintics and its mirror) Candelas-de la Ossa-Green-Parks [CdlOGP91] considers the family depending on the parameter ψ of subschemes of P4 defined by the polynomial g = Y5 0 + Y5 1 + Y5 2 + Y5 3 + Y5 4 − 5ψY0Y1Y2Y3Y4. (1) The mirror family is formed by taking a (Z/5Z)3 quotient of these quintics. Explicitly, the group action is (ζ1, ζ2, ζ3) ⋆ (Y0, Y1, Y2, Y3) = (Y0, ζ1Y1, ζ2Y2, ζ3Y3, (ζ1ζ2ζ3)4Y4) (2) where the ζ’s are 5th roots of unity. 9 Definition 3.2. (base fan) Writing F(−) for the free group, the fan of P4 is ΣY ⊂ N = F({uρ0, uρ1, uρ2, uρ3, uρ4})/〈uρ0 + uρ1 + uρ2 + uρ3 + uρ4〉 with cones given by the Q≥0-spans of the subsets of size ≤ 4 of the generators. Proposition 3.3. (ΣX) With V = O (5), the fan of the toric variety X = SpecSym• V is ΣX ⊂ N = N ⊕ Z with cones given by the Q≥0-spans of the subsets of {(0, 1), (uρ0, 1), (uρ1, 1), (uρ2, 1), (uρ3, 1), (uρ4, 1)} that exclude at least one element of the form (uρi , 1). Proof. Following the method of Proposition 2.16, we choose the divisor D = D0 + D1 + D2 + D3 + D4 where Di = {[Y0 : · · · : Y4] Yi = 0}. Then by excluding at least one of the lifted rays in forming our cones, we get all the cones described in the proposition. Definition 3.4. (invariant characters) The characters M on X are naturally thought of as those Laurent monomials in Y0, . . . , Y4 with degree divisible by 5. ′ ⊆ M for those monomials invariant under the action of (Z/5Z)3 given Write M in Equation (2), and set N ′ = HomZ(M , Z). ′ Proposition 3.5. (ΣX′) The fan ΣX′ ⊂ N ΣX under the map N → N ′ transpose to the inclusion M ′ ′ of X′ = X/(Z/5Z)3 is the quotient fan of ,→ M. Proof. This is an example of a situation in which the quotient is presented by the quotient fan as described in Subsection 2.2.1. Proposition 3.6. (N ′ = M) The assignment (n, k) 7→ (Y0Y1Y2Y3Y4)kYi (Y5 i /Y0Y1Y2Y3Y4)ni for n = n0uρ0 +· · ·+n5uρ5 ∈ N and k ∈ Z injectively maps N into M and with image ′ M ′ . The transpose isomorphism N → M allows us to meaningfully write ΣX′ ⊂ M. 10 Proof. To be sure, M′ is made up of monomials with degree 0, and is freely generated by those with exponent vectors {(−1, 4, −1, −1, −1), (−1, −1, 4, −1, −1), (−1, −1, −1, 4, −1), (−1, −1, −1, −1, 4)}. A computational check shows the assignment uρi 7→ Y5 i /Y0Y1Y2Y3Y4 (3) (4) gives a well defined homomorphism and identifies N and M′. Finally, sending (0, 1) 7→ Y0Y1Y2Y3Y4 and ′ M =[k (Y0Y1Y2Y3Y4)k M′ completes the isomorphism M ′ = N. Proposition 3.7. (Γ, W, γ 7→ g ) The global sections Γ(P4, V ) considered as functions on X has a basis of monomials Ξ = {degree 5 monomials in Y0, · · · , Y4}. Thus, Γ = Spec Z[γm]m∈Ξ and the auxiliary Landau-Ginzburg model of the quintic is where The specialization W : X(ΣX)Γ → A1 Γ W = Xm∈Ξ γmYm. γm 7→  −5ψ if m = Y0Y1Y2Y3Y4, i for some i, if m = Y5 otherwise 1 0 A1 ψ → Γ given by sending sets the coefficients to those used in (1). Proof. These statements are immediate. 11 Proposition 3.8. (Γ′, W′, γ′ 7→ g ′) Functions on X′ which pullback to elements of Γ(Y, V ) have a basis of monomials Ξ′ = {degree 5 monomials in Y0, · · · , Y4 invariant under the action (Z/5Z)3}. Thus and the auxiliary Landau-Ginzburg model of the quintic-mirror is Γ′ = Spec Z[γ′ m ]m∈Ξ′ where The specialization W′ : X(Σ′)Γ′ → A1 Γ′ W′ = Xm∈Ξ′ mYm. γ′ A1 ψ → Γ′ given by sending γ′ sets the coefficients to those used in (1). m 7→  −5ψ if m = Y0Y1Y2Y3Y4, i for some i, if m = Y5 otherwise 1 0 Proof. These statements are immediate. Theorem 3.9. (quintic main) ΣX and ΣX′ are dual fans, • C(ΣX), C(ΣX′) and Γ′ are isomorphic, • there is a closed immersion Γ′ ,→ Γ, and • a closed immersion A1 ψ ,→ C(ΣX), such that the dual toric Landau-Ginzburg models of ΣX and ΣX′ are obtained from the associated toric Landau-Ginzburg models of Proposition 3.7 via base change, and the specializations of Proposition 3.7 factor though the map A1 ψ ,→ C(ΣX). Proof. The map in Equation (4) produces the map C(ΣX) ,→ Γ′ by sending • c(uρi ,1) 7→ γ′ Y5 i and • c(0,1) 7→ γ′ Y0Y1Y2Y3Y4 . 12 Most of these statements follow from explicitly writing the maps involved. The m if m is invariant under the (Z/5Z)3 closed immersion Γ′ ,→ Γ sends γm 7→ γ′ action, and 0 otherwise. The isomorphism C(ΣX′) → C(ΣX) comes from the fact that the 1-cones are in bijection under the quotient map ΣX → ΣX′. The expression c(ρ0,1)Y5 0 + c(ρ1,1)Y5 1 + c(ρ2,1)Y5 2 + c(ρ3,1)Y5 3 + c(ρ4,1)Y5 4 + c(0,1)Y0Y1Y2Y3Y4 for W(ΣX) produces most of the rest. In light of the identification C(ΣX′) → C(ΣX) this is also an expression for W(Σ′), and this immediately yields the base-change statements. The regularity of these functions guarantees the fans are dual by the observation made in Proposition 2.18. The remaining question whether C(ΣX) ,→ Γ′ is an isomorphism is settled by noticing the matrix with rows from Equation (3) is the map M′ → M in the exact sequence 0 → M′ → M → (Z/5Z)3 → 0. 3.2 Berglund-Hübsch-Krawitz [BH92] [Kra10]. In this case, both sides of the mirror are Landau-Ginzburg models so W and W′ are formed by simply inserting variables for the coefficients of W and W′. Γ and Γ′ are the affine spaces with these coefficient variables as coordinates. Definition 3.10. (polynomials and phase symmetries of a matrix) The dual- ity of Berglund-Hübsch [BH92] is based on invertible matrices with nonnegative integer entries. Given such an (n +1)×(n +1)-matrix P = (pi j )i j , one can define the surjective group homomorphism F (P) : (C×)n+1 → (C×)n+1 given by the assignment y j 7→ x this we have a polynomial p0 j 0 · · · x pn j n . Abstractly, F (−) = Spec C[−]. From WP = ( nXj =0 y j ) ◦ F (P) = nXj =0 nYi =0 x pi j i , and a group of phase symmetries SP = ker F (P). 13 The transpose matrix Pt defines what is referred to as the transpose poly- nomial WPt and the transpose phase symmetries SPt . Definition 3.11. (quantum symmetries and BH mirror criterion) The Berglund- Hübsch [BH92] mirror criterion is formulated in terms of choices of groups of quantum symmetries QP ≤ SP and QPt ≤ SPt . These can be any subgroups and constitute a dual pair if the cokernels GP in and GPt in 1 → QP → SP → GP → 1 1 → QPt → SPt → GPt → 1 satisfy QP ∼= GPt and QPt ∼= GP. Typically such a pair is presented as (WP, QP), (WPt , QPt ). Definition 3.12. (Krawitz’s dual group) Krawitz [Kra10] discovered a way to systematically produce a group Qt ≤ SPt dual to any given group of quantum P symmetries QP. A reformulation Krawitz’s dual by Clarke [Cla] first considers the diagram of character groups M A Bt Zn+1 P Zn+1. (5) associated to the homomorphisms (C×)n+1 F (A) (C×)n+1/QP F (P) (C×)n+1 F (Bt ) . Qt P is then defined to be the the kernel of F (B) : (C×)n+1 → U′ U′ = Spec C[N] for N = HomZ(M, Z). 14 Definition 3.13. (BHK mirror LG models) Berglund-Hubsch-Krawitz mirror pairs arise as Landau-Ginzburg models associated to (P, QP) and (Pt , Qt P). The Landau-Ginzburg model associated P and a group of quantum symmetries QP, is WP : Cn+1/QP → C. In the same way Pt and Qt two constitute a BHK mirror pair. P defines a Landau-Ginzburg model. Together, these Proposition 3.14. (Σ, Σ′) The fan Σ of Cn+1/QP is the image of the positive orthant fan under the map At : Zn+1 → N, and WP is given by the image Σ′ of the positive orthant fan under the map B : Zn+1 → M of Equation (5). Proof. This is an example of the situation in which the geometric quotient is presented by the quotient fan as described in Subsection 2.2.1. Definition 3.15. (W, W′, Γ, Γ′ and γ 7→ g , γ′ 7→ g ′) For the auxiliary Landau- Ginzburg models we have • W =Pn • W′ =Pn j =0 γ jQn i =0 x iQn j =0 x′pi j i =0 γ′ where pi j i , and j • Γ = Spec Z[γ j ]j , and • Γ′ = Spec Z[γ′ i ]i . The specializations to recover the BHK mirrors set all the γ and γ′ variables to 1. Theorem 3.16. (BHK main) The fans (Σ, Σ′) are dual with pi j = uρ′ j (uρi ), there are isomorphisms C(Σ′) = Γ, C(Σ) = Γ′, W(Σ′) = W, W(Σ) = W′ and WP and WPt are obtained from these via base change. Proof. These statements are immediate from the definition and considerations above. 15 3.3 Batyrev-Borisov [BB97] Some work is required to explicitly extract the mirror complete intersections in this construction. Γ and Γ′ are the affine spaces of global sections, and W and W′ are the universal sections considered as functions. 3.3.1 Gorenstein cones and splittings. Batyrev-Borisov mirror symmetry [BB97] is centered on complete splittings of reflexive Gorenstein cones. Definition 3.17. A full dimensional cone K in a the reification MR of a finite rank free abelian group M is called Gorenstein if there is a homomorphism ℓ∨ : K ∩ M → N such that K ∩ M is generated as a monoid by (ℓ∨)−1({1}). Definition 3.18. A Gorenstein cone K is called reflexive if the dual cone K∨ is also Gorenstein. The number ℓ∨(ℓ) = r is called the index of K. Definition 3.19. A complete splitting of an index r reflexive Gorenstein cone K ⊆ M is a set of non-zero elements E = {e1, . . . , er } ⊆ K ∩ M such that ℓ = e1 + · · · + er . 3.3.2 Dual splittings, support and partitions If a Gorenstein cone has a splitting, then so does its dual. The proof of this uses the notions of both the support of a Gorenstein cone and the support partition given by a complete splitting. Definition 3.20. A dual complete splitting E∨ to a given complete splitting E is a complete splitting E∨ = {e ∨ 1 , . . . , e ∨ r } of K∨. Remark 3.21. (non-uniqueness of splittings) A splitting need not be unique. Indeed, Batyrev-Nill give a simple explicit example where it is not [BN08, Ex- ample 5.1]. However, if r = 1 it’s clear that there is only one “splitting,” e1 = ℓ, so uniqueness is guaranteed. Definition 3.22. Associated to a Gorenstein cone K is a convex polyhedral set ∆ = {k ∈ K ℓ∨(k) = 1 ∈ R}. called the support of K. We will denote the support of K∨ by ∇. 16 Definition 3.23. (support partition) A complete splitting E∨ determines sets ∆i = {k ∈ K e ∨ i (k) = 1 and e ∨ j (k) = 0 for all j 6= i } ⊆ ∆. Proposition 3.24. (support partitions partition) The sets ∆i partition the in- tegral elements of ∆, each ∆i is nonempty and is the convex hull of vertices of ∆ it contains. Proof. The partition assertion follows from the fact that any integral point p in i (p) = 1 for exactly i (p) ∈ N. This means that e ∨ ∆ has ℓ∨(p) =Pi e ∨(p) = 1 and e ∨ one i and zero for all others, i.e. p ∈ ∆i . Nonemptyness can be seen by considering ℓ ∈ K and the fact that ℓ = p1 + i (p j ) = 0 for all j , i ) = 0. This is a contradiction. In fact, we have shown that in any expression of ℓ as a sum of element · · · + pr for not necessarily distinct p j ∈ ∆. If ∆i = ∅ then e ∨ and consequently ℓ(e ∨ each ∆i contains exactly one p j from ∆, since any other number would lead to ℓ(e ∨ The final claim, that each vertex of ∆i is a vertex of ∆, is an immediate con- sequence of the fact that a point in ∆i cannot be written as a convex combina- tion of points in ∆ if there is a non-zero term whose point p has e ∨ i ) 6= 1. i (p) = 0. Corollary 3.25. (summand partition) For an index r reflexive Gorenstein cone, each ∆i contains exactly one p j in any expression of as a sum of non-zero lattice points in K. ℓ = p1 + · · · + pr Proposition 3.26. (dual splittings exist) If K admits a complete splitting, then so does K∨. Proof. The proof is essentially a restatement of the middle of the proof of Propo- sition 3.24. If we fix a splitting E, the integral points of ∇ generate K∨ ∩ N and are partitioned by the ∇i ’s. So ℓ∨ ∈ ∇1 +· · · + ∇r and any choice ℓ∨ = e ∨ +· · ·+e ∨ r 1 gives a complete splitting of K∨. 3.3.3 Dual splittings, complete intersections, and corresponding LG’s Definition 3.27. (the base toric variety Y) Given dual complete splittings E and E∨, define M = {m ∈ M e ∨ j (n) = 0 ∀e ∨ j ∈ E∨}. 17 Within MR we have the convex polyhedral set ∆i = {m ∈ MR m + ei ∈ ∆i } and the Minkowski sum ∆ =Xi ∆i . Denote by ΣY ⊆ N = HomZ(M, Z) the normal fan of ∆, and write Y for the toric variety of this fan. Proposition 3.28. (polar=convex) [BN08, Proposition 3.18] The polar polytope ∆∗ of ∆ is the convex hull of the ∇ j ’s: ∆∗ = conv({∇ j } j ). Definition 3.29. The proposition above and the integrality of the vertices of the ∇ j ’s allow us to define for each ei ∈ E a divisor Di = Xuρ∈∇i Dρ on Y. Theorem 3.30. (Cartier-ity) Di is Cartier with global section polytope ∆i . Proof. The theorem follows from establishing the polytope associated to Di is ∆i . From this, Di is Cartier by the integrality of the vertices of ∆i and Theorem 6.1.7 from Cox-Little-Schenck [CLS11]. Recall that m ∈ MR is in the polytope associated to Di if and only if Working ρ by ρ this is Di + Xρ∈ΣY(1) uρ(m)Dρ ≥ 0. uρ(m) +½ 1 if uρ ∈ ∇i , 0 otherwise ≥ 0. Proposition 3.28 means the uρ’s equal the vertices of the convex hull of the ∇i ’s and this provides enough to complete the proof. Indeed, the transpose N → N 18 of the inclusion M ,→ M sends ∇ to the convex hull of the uρ’s So uρ(m) = v (m) for an appropriate vertex of ∇, and becomes v (m + ei ) ≥ 0 uρ(m) ≥ 0 for those uρ which do not come from ∇i , and uρ(m) + 1 ≥ 0 for the others. Definition 3.31. (V , X, Σ) With V =Li • X = SpecSym• V , and O (Di ), we set • Σ = ΣX. A complete description of the fan Σ depends on knowing more information about the vertices ∆. However, we can determine what we need with what we know so far. Proposition 3.32. (NX = N, Σ(1) and Γ(Y, V )) The one-parameter subgroups of X are naturally identified with N. Under this identification the primitive generators of the 1-cones in Σ are exactly vert( ∇) ∪ {e ∨ i }i . Thus Σ = K∨, and Γ(Y, V ), considered as functions on X, have a basis made up of elements ∆ ∩ M. Proof. To produce Y via the construction according to Definition 3.27, we begin with a pair of dual nef-partitions E and E∨. First there is a polytope ∆ ⊆ M = (E∨)⊥, defined in terms of E. The fan of Y is the inward normal fan of ∆ and lives in the group N which is Z-dual to M. Writing F(−) for the free group, and following the construction of the fan Σ O (−Di ) over a toric variety of Corollary 2.17 the fan lives of a split bundle Li in N = N ⊕ F({τ∨ i }i ) (= N ⊕ Z{τi }i ) where τ∨ i 1-cones are either the “vertical” 1-cones generated by the is Z dual to the rational section τi of O (Di ) with divisor Di . The τ∨ i . 19 or those the lifted from the 1-cones of ΣY. The generators of the lifts can be written abstractly as where the coefficients come from the expansion νi ,ρτ∨ i , buρ = uρ +Xi Di =Xρ νi ,ρDρ. For us this is for the unique ∇i containing uρ. buρ = uρ + τ∨ i , Identification of these points with corresponding points in K∨ begins with identifying the ambient spaces. Writing F(−) for the free group, the dual nef- partitions provide a splitting M = M ⊕ F(E). M is already equal to the dual of N, and we can further identify the characters on the total space with M by the assignments ei = τi . To be sure, this is legitimate because the divisor of τi divisors and it (up to constant rescaling) a character. is supported on toric This identification immediately matches buρ ∈ ∇i with a vertex of ∇i , and the elements of E∨ corresponds with the τ∨ i ’s. Definition 3.33. (Γ, Γ′, W, W′, Σ′, Z, Z∨, γ 7→ g , γ′ 7→ g ′) In light of the theorem above, we set • Γ = Spec Z[γm]m∈ ∆∩M, • W =Pm∈ ∆∩M γm xm , and • Z ⊆ Y × Γ is the zero scheme of the universal section of V . In addition we denote by Σ′, Γ′, W′ and Z∨ the analogous objects obtained by reversing the roles of K and K∨ above. Finally, the specializations are the identity maps Γ → Γ and Γ′ → Γ′. 20 Theorem 3.34. (BB main) The fans Σ and Σ′ are dual, and the toric Landau- Ginzburg models they define are obtained from the auxiliary Landau-Ginzburg mod- els of Z and Z∨ via base change along the inclusions C(Σ′) ,→ Γ and C(Σ) ,→ Γ′ which set to 0 any coefficient corresponding to a non-ray generating character. Proof. These statements are immediate from the observation that the primitive generators of the 1-cones of Σ′ lie in ∆. 3.4 Givental [Giv98] In this case, one side of the mirror is a complete intersection and the other is a Landau-Ginzburg model, so Γ is the affine spaces of global sections, and W is the universal section considered as a function. On the other side, W′ is formed by simply inserting variables for the coefficients of W′, and Γ′ is the affine space with these coefficient variables as coordinates. Most of the work is expressing X′ as a toric variety, and expanding W′ in characters. Definition 3.35. (Givental’s mirror) As input data, we have • a smooth projective toric variety Y with n torus invariant divisors, • ℓ nonnegative (i.e. basepoint free [Giv98, pg. 23]) line bundles (La)a • a choice of integral symplectic basis (p1, . . . , pk ) of H2(Y, Z). such that ω∨ L ∨ a is nonnegative, and a ⊗Na From this we have integers ℓi a defined by and integers mi ρ defined by ch(La) =Xi ℓi a pi ch(O (Dρ)) =Xi mi ρpi for the toric divisor Dρ corresponding to ρ ∈ ΣY(1). 21 Givental’s mirror Landau-Ginzburg model is given by the scheme E ⊆ Cn w × Cℓ v × (C×)k q which is the closure of the zero locus in (C×)n q of the equations u × (C×)ℓ v × (C×)k Yρ w mi ρ ρ = qiYa v ℓi a a i = 1, . . . , k equipped with the function W′ : E → C given by W′ =Pa va −Pρ wρ. Definition 3.36. (V , X, N, Σ, Γ, W, γ 7→ g ) With V =La • X = SpecSym• V , La , we set (6) (7) • N = NY ⊕ Z{ea }a , • Σ = ΣX ⊂ N, • Γ = Spec Z[γm]m∈Ξ where Ξ = {µ + ea µ ∈ ∆a ∩ MY} for the global section polytopes ∆a of the La’s, • W =Pm∈Ξ γmxm, and • the specialization Spec C → Γ depends on the equations of the complete intersection chosen. Lemma 3.37. (F-split) Writing F(−) for the free group, H2(Y, Z) is the cokernel of the map from characters to divisors on Y 0 → MY → F({Dρ ρ ∈ ΣY(1)}) → H2(Y, Z) → 0. Furthermore, one can find k divisors Dρi such Li Z Dρi maps isomorphically to H2(Y, Z). Proof. Both statements follow from the facts that Y is smooth and complete. For the first, see, for instance, Theorems 4.1.3, 4.2.1, and 12.3.2 of Cox-Little- Schenck [CLS11]. For the second, since Y is smooth one can choose a torus fixed point, and the k divisors which do not pass through it map to a basis for H2(Y, Z). 22 Lemma 3.38. (torus of E) Since there is a natural injection ΣY(1) → ΣX(1), we can combine the inclusion ι : MX → F({Dρ ρ ∈ ΣX(1)}) and the graph of the map s : H2(Y, Z) → F({Dρ ρ ∈ ΣY(1)}) formed from an isomorphism H2(Y, Z) →Li Z Dρi to yield an inclusion · ι 0 1 ¸ : MX ⊕ H2(Y, Z) ,→ F({Dρ ρ ∈ ΣY(1)}) ⊕ H2(Y, Z). s Applying the functor Spec C[HomZ(−, Z)] identifies • Spec C[HomZ(H2(Y, Z), Z)] with (C×)k q , • Spec(C[NX]) × (C×)k q with E ∩ (C×)n u × (C×)ℓ v × (C×)k q, and • the inclusion with E ∩ (C×)n u × (C×)ℓ v × (C×)k q ,→ (C×)n u × (C×)ℓ v × (C×)k q . Proof. Note H2(X, Z) = H2(Y, Z). If H2(X, Z) is identified with Zk with the (pi )i basis, then matrix [(mi ρ)i ρ − (ℓi a )i a] is the map from the group of torus invariant divisors if X to H2(X, Z). To verify this, we check that it is the kernel of the transpose Ut : Zn ⊕ Zk → NX of the first non-zero map in the exact sequence 0 → MX → Zn ⊕ Zℓ → Zk → 0 The map Ut sends the standard bases vectors to the primitive generators of rays in ΣX(1). Following Corollary 2.17, we have explicitly U =· (uρ)ρ 0 (αρa)ρa I ¸ . where we write (uρ)ρ for the matrix with rows uρ for ρ ∈ ΣY(1) and (αρa)ρa is obtained from the expansions Da =Xρ αρaDρ 23 of toric divisors Da such that La = OY(Da). The α’s satisfy ℓi a =Xρ mi ραρa. (8) Finally, [(mi ρ)i ρ − (ℓi a)i ]U = 0 by virtue of equation (8) and relations for the rays of Y Xρ mi ρuρ = 0. The identification of Spec C[NX]×(C×)k q with E∩(C×)n u ×(C×)ℓ v ×(C×)k q comes from plugging the exact sequence 0 → MX ⊕ H2(X, Z) → F({Dρ ρ ∈ ΣY(1)}) ⊕ H2(X, Z) → H2(X, Z) → 0 into the functor Spec C[HomZ(−, Z)]. The first non-zero map is · ι 0 1 ¸ s (9) and the second is [(mi ρ)i ρ−(ℓi a)i −1]. This yields an exact sequence of groups 1 → Spec C[NX] × (C×)k q → (C×)n u × (C×)ℓ v × (C×)k q → (C×)k q ′ → 1 where the map to (C×)k q ′ is given by q ′ i = (Yρ mi ρ ρ w )/(qiYa v ℓi a a ). Proposition 3.39. (ΣX′) Denote H = Spec C[HomZ(H2(Y, Z), Z)]. Then MX = MY ⊕ F({ea}a), and E = Spec C[K∨ ∩ NX] × H where K∨ is the cone dual to the global function cone of X : K = {t1(µ1 + e1) + · · · + tℓ(µℓ + eℓ) ∈ (MX)Q ta ≥ 0 and µa ∈ ∆a for all a}. Proof. Functions which are global on E are exactly those on E ∩ (C×)n (C×)k elements of q which are restrictions from Cn v × (C×)k v × q . For characters, these are u × (C×)ℓ u × Cℓ NX ⊕ Zk 24 which come from Nn ⊕ Nℓ ⊕ Zk under the transpose of · ι 0 1 ¸ s from (9). An element (n, z) ∈ NX ⊕ Zk defines a global characters if and only if both (n, 0) and (0, z) are global characters. The 1 in the lower righthand corner guarantees 0 ⊕ Zk is made up of global characters. So the question reduces to knowing for which elements n of NX the character (n, 0) is global. The n’s we are interested in are exactly those which are global on X′, and this translates into the condition n is in the cone dual the global functions on X. An element n ∈ NX comes from Nn ⊕ Nℓ if and only if n is non-negative on any element of MX which maps into Nn ⊕ Nℓ. These are exactly the elements of MX which define a global function on X. Corollary 3.40. (ΣX′ (1) ⊂ Γ(Y, V )) Take X′ = Spec C[K∨ ∩NX]. Then ΣX′ is the cone K and all its facets. Furthermore, the primitive integral generators of elements of ΣX′(1) are all of the form v + ea for v ∈ ∆a. Proof. The only statement that isn’t immediate is the one about the generators. However, we know that ∆a has integral vertices, so this guarantees result. Proposition 3.41. (W′ on X′ × H) Enumerate ΣX(1) so that the first k rays corre- spond to the divisors in Lemma 3.37. Then as a function on E = X′ × H we have W′ =Xa x′ a − kXj =1Yi q −s j i i x′ ρ j − nXj =k+1 x′ ρ j (10) where • x′ a = va, =Qi q • x′ ρ j • x′ ρ j s j i i wρ j for j ≤ k, and = wρ j for j > k. are characters on X′ via Zn ⊕ Zℓ → NX. The integers s j i come from the map s of Lemma 3.38 in the (pi )i basis. Furthermore, the characters x′ ρ are exactly the primitive generators for the rays in ΣX(1). a and x′ 25 Proof. The way functions pull back under the inclusion X′ × H ,→ Cn × Cℓ(C×)k is described by the transpose of the map in (9). Both the characters x′ a and x′ ρ and the primitive generators for the rays in ΣX(1) are the images of the standard basis vectors under Zn ⊕ Zℓ → NX. Definition 3.42. (Γ′, W′, γ′ 7→ g ′) The expansion of W′ in (10) and the identifi- cation of its terms with the primitive generators for the rays in ΣX(1) gives us the definitions • Ξ′ = {n ∈ NX n is a primitive generator of ΣX(1)}, • Γ′ = Z[γ′ n]n∈Ξ′ , • W′ =Pn∈Ξ′ γ′ nx′ n, and • the specialization H → Γ′ sets the coefficients of W′ to match those in (10). Theorem 3.43. (Givental’s mirror main) There are inclusions as coordinate sub- spaces C(Σ′) → Γ and C(Σ) → Γ′ which produce the dual pair of toric Landau-Ginzburg models formed from Σ and Σ′ via base change. Furthermore, C(Σ) → Γ′ is an isomorphism. Proof. Omitted. Remark 3.44. (independence of choice of p’s) Notice that one consequence of this theorem is that Givental’s mirror is independent of the choice of basis (pi )i of H2(Y, Z). Indeed, one need not even assume the existence. 3.5 Hori-Vafa [HV00] Hori-Vafa gives an expression [HV00, Equation (7.78)] for the BPS mass of a certain D-brane related to complete intersection Z in a toric variety Y. Their expression is the integral (11).This integral is the pullback of and integral on a algebraic torus. This torus the function W′ appearing in the integrand match Givental’s Landau-Ginzburg mirror. Definition 3.45. (Hori-Vafa BPS mass [HV00, Equation (7.78)]) Given a toric variety Y, ℓ global sections Gβ of line bundles Lβ, and a relative homology class, [γ] ∈ Hn(CN × Cℓ; B) where B = W′−1(r ) is a fiber of W′ = NXi =1 exp(−Yi ) + ℓXβ=1 exp(−YPβ ) 26 over a sufficiently large r ∈ R, the BPS mass for a D-brane wrapping k is Πγ = Rγ QN Qℓ Qk a=1 δ(PN exp(−W′) i =1 d Yi β=1 exp(−YPβ )d YPβ i =1 Qi aYi −Pβ dβaYPβ (11) − ta ) where • ΣY(1) = {ρ1, . . . , ρN}, • {η1, . . . , ηk} is an integral basis of H2(Y, Q), and • the class of Lβ is Pa dβa ηa. The delta functions in the integrand of Equation (11) are interpreted to mean a restriction of the integral to the subset on which the arguments vanish. Definition 3.46. (δ functions and forms) If φ is a differential form, the ex- pression δ(g1) · · · δ(gk )φ Zγ Zγ∩{g1 =···=gk =0} ψ means where φ = d g1 ∧ · · · ∧ d gk ∧ ψ. The apparent ambiguity in this definition is addressed in Lemma 3.47. Lemma 3.47. (unambiguity of δ valued forms) Given functions f1, · · · , fk, write V(g ) = {p f1(p) = · · · = fk (p) = 0} and U(d g ) = {p d f1p , . . . , d fk p are linearly independent}. Any forms ψ and ψ′ such that d f1 ∧ · · · ∧ d fk ∧ ψ = d f1 ∧ · · · ∧ d fk ∧ ψ′, 27 satisfy Zγ ψ =Zγ ψ′ for any simpex γ : ∆n → V(g ) such that γ−1(U(d g )) is dense in ∆n. Proof. Consider a point p ∈ U(d g ) ∩ V(g ). About this point we can find co- ordinates g : U → Rn such that the first k coordinates are f1, . . . , fk . On this neighborhood, ψ = ψ′ + d f1 ∧ ψ1 + · · · + d fk ∧ ψk for forms ψ1, . . . , ψk. However, any d fi for i = 1, . . . , k annihilates any tangent vector in V(g ). So provided the image of γ is in V(g ) we know Zγ−1(U(d g )) γ∗ψ =Zγ−1(U(d g )) γ∗ψ′. Finally, γ−1(U(d g )) is dense in ∆n, so its complement has measure zero and its omission doesn’t affect the integral. integrand of Equation (11) is the pullback of Proposition 3.48. (variables on the torus) For W′ = −PN (−1)N+ℓ exp(−W′) Ya −qaYβ δ(Yi w Qi a dβa β v i i =1 wi −Pℓ β=1 vβ, the ) d ln(w1)∧· · ·∧d ln(wN)∧d v1∧· · ·∧d vℓ (12) → (C×)k q × (C×)N w × (C×)ℓ v along given by Ck t × CN Y × Cℓ YP • wi = exp(−Yi ), • vβ = exp(−YPβ ), and • qa = exp(−ta ). Proof. Direct substitution. Definition 3.49. (Hori-Vafa mirror LG) Denote by Z the subsheme of Y de- fined by Gβ = 0 for β = 1, . . ., ℓ. The Hori-Vafa mirror Landau-Ginzburg mirror to Z is (X′, W′) where X′ ⊆ (C×)k v is given by the k equations q × CN w × Cℓ Yi w Qi a i − qaYβ v dβa β and W′ = −PN i =1 wi −Pℓ β=1 vβ. 28 Theorem 3.50. (computation of BPS mass on the dual) Using the interpreta- tion of Definition 3.46, the BPS mass given in Equation (11) equals Zγ ψ for the form in (12) and the class γ ⊆ X′ that is the image in X′ of γ ∩ {g1 = . . . = gk = 0}, where the g ’s are the arguments of the δ functions. Remark 3.51. (comparison to Givental’s mirror) Notice that X′ and Givental’s Eq in Equation (6) are the same. However, Givental’s W′ in Equation (7) has a different sign in front of the w-variables than the Hori-Vafa W′. This difference means that both are obtained from the same auxiliary Landau-Ginzburg model via different base change morphisms. Acknowledgements The author would like to thank E. Sharpe for his interest and for first pointing out Berglund-Hübsch mirror symmetry, and M. Krawitz for explaining his dual group construction. Correspondence and conversations with L. Borisov, D. Cox and B. Nill on several aspects of toric geometry have been invaluable. Finally, we thank L. Katzarkov, E. Gasparim and Y. Ruan for their great encouragement and patience. 29 References [Bat94] [BB97] [BH92] [BN08] Victor V. Batyrev. Dual polyhedra and mirror symmetry for Calabi- J. Algebraic Geom., 3(3):493– Yau hypersurfaces in toric varieties. 535, 1994. Victor V. Batyrev and Lev A. Borisov. Dual cones and mirror sym- metry for generalized Calabi-Yau manifolds. In Mirror symmetry, II, volume 1 of AMS/IP Stud. Adv. Math., pages 71–86. Amer. Math. Soc., Providence, RI, 1997. Per Berglund and Tristan Hübsch. A generalized construction of mirror manifolds. In Essays on mirror manifolds, pages 388–407. Int. Press, Hong Kong, 1992. Victor Batyrev and Benjamin Nill. Combinatorial aspects of mirror symmetry. In Integer points in polyhedra—geometry, number theory, representation theory, algebra, optimization, statistics, volume 452 of Contemp. Math., pages 35–66. Amer. Math. Soc., Providence, RI, 2008. [Bor93] Lev Borisov. Towards the Mirror Symmetry for Calabi-Yau Complete intersections in Gorenstein Toric Fano Varieties, 1993, arXiv:alg-geom/9310001. [CdlOGP91] Philip Candelas, Xenia C. de la Ossa, Paul S. Green, and Linda Parkes. A pair of Calabi-Yau manifolds as an exactly soluble su- perconformal theory. Nuclear Phys. B, 359(1):21–74, 1991. [Cla] [CLS11] [Dem70] Patrick Clarke. A proof of the birationality of certain BHK-mirrors. Complex Manifolds. David A. Cox, John B. Little, and Henry K. Schenck. Toric va- rieties, volume 124 of Graduate Studies in Mathematics. American Mathematical Society, Providence, RI, 2011. Michel Demazure. Sous-groupes algébriques de rang maximum du groupe de Cremona. Ann. Sci. École Norm. Sup. (4), 3:507–588, 1970. 30 [Giv98] Alexander Givental. A mirror theorem for toric complete intersec- tions. In Topological field theory, primitive forms and related topics (Kyoto, 1996), volume 160 of Progr. Math., pages 141–175. Birkhäuser Boston, Boston, MA, 1998. [GKZ94] I. M. Gel′fand, M. M. Kapranov, and A. V. Zelevinsky. Discrimi- nants, resultants, and multidimensional determinants. Mathematics: Theory & Applications. Birkhäuser Boston, Inc., Boston, MA, 1994. [GP90] [HV00] [Kra10] B.R. Greene and M.R. Plesser. Duality in calabi-yau moduli space. Nuclear Physics B, 338(1):15 – 37, 1990. Kentaro Hori and Cumrun Vafa. Mirror symmetry, 2000, arXiv:hep-th/0002222. Marc Krawitz. FJRW rings and Landau-Ginzburg mirror symmetry. ProQuest LLC, Ann Arbor, MI, 2010, arXiv:0906.0796. Thesis (Ph.D.)–University of Michigan. Department of Mathematics, Drexel University, Philadelphia, PA 19104 [email protected] 31
1502.00618	3	1502	2015-11-19T21:47:53	Isotrivial unfoldings and structural theorems for foliations on Projective spaces	[ "math.AG", "math.CV", "math.DS" ]	Following T. Suwa, we study unfoldings of algebraic foliations and their relationship with families of foliations, making focus on those unfoldings related to trivial families. The results obtained in the study of unfoldings are then applied to obtain information on the structure of foliations on projective spaces.	math.AG	math	Isotrivial unfoldings and structural theorems for foliations on Projective spaces. Federico Quallbrunn Abstract. Following T. Suwa, we study unfoldings of algebraic foli- ations and their relationship with families of foliations, making focus on those unfoldings related to trivial families. The results obtained in the study of unfoldings are then applied to obtain information on the structure of foliations on projective spaces. The objective of this work is twofold. 1. Introduction In a first stance, we want to investigate the relation between unfold- ings and families of foliations in algebraic varieties (see definitions below), specially those unfoldings related with trivial families. In this respect our principal result is Theorem 2.12, which can be viewed as a generalization of [11, chapter 7]. In a second stance, we apply our results on unfoldings and deformations In this to study foliations on Pn whose degree is low with respect to n. respect our main results are Theorem 5.2 and Corollary 5.3. The result of Theorem 5.2 is obtained here by considering unfoldings of foliations on P2. The concept of unfoldings of foliations and its relations with deformation theory of foliations were first introduced and studied by Suwa in a series of papers (see [8, 9, 11] and references therein). An unfolding of a foliation F0 on a variety X is a "bigger" foliation F on a variety X × S whose leaves contains those of F0 (see below for a precise definition). So we can always think of a foliation of codimension q on Pn as being birationally equivalent to an unfolding of a foliation by curves on Pq+1. This, in general, does not provide us with much information, but under certain hypotheses we can control what kind of unfolding we may consider. The special kind of unfolding that will give us structural information on the foliation will be Isotrivial unfoldings. Unfoldings may be thought of as special kinds of families of foliations, a family on which not only the differential equations defining the foliation vary continuously but also the solutions i.e.: the leaves vary continuously as well (this is, of course, very Isotrivial unfoldings are vague, again, see bellow for precise definitions). 1 2 FEDERICO QUALLBRUNN those that are related to trivial families of foliations, with this vague intu- ition, they are families in which the equations stays still while the solutions vary. They were studied by Suwa in the infinitesimal case, see [11]. Here we generalize Suwa's results on isotrivial unfolding to be able to It deal with unfoldings and families parametrized by arbitrary schemes. turns out that isotrivial unfoldings have a rather rigid structure and so being able to identify a foliation on Pn with an isotrivial unfolding says something about the structure of the foliation. In Section 2 we give the principal definitions and state the theorems on unfoldings that will serve us as tools to conclude statements about foliations on Pn. The main result is Theorem 2.12 which can be viewed as a first step in determining the representation of the functor that to every scheme S associates the unfoldings of a given foliation parametrized by S, although such a line of investigation will not be pursued in this work. In Section 3 we explain how foliations on Pn may be considered as giving rise to unfoldings of foliations by curves, and when this unfoldings may be taken isotrivial. In Section 4 we deal with the technical issue of transversality, which is a condition we need to have to be able to apply the results of Section 2. Finally, in Section 5 we apply the results of the previous sections to conclude some structural statements on foliations on Pn(C). In particular, Theorem 5.2 follows as a particular case of a statement (Proposition 5.1) valid for arbitrary dimensional foliations, although restrictions on the degree are always required. The author was supported by a post-doctoral grant of CONICET. The author is grateful to Jorge Vitorio Pereira for useful suggestions, and for corrections of crucial mistakes in earlier versions of this paper. Also grati- tude is due to Ariel Molinuevo for introducing the author to the concept of unfolding of a foliation, and to the Algebraic Geometry Seminar of Buenos Aires for its nurturing atmosphere and overall support. 2. Isotrivial unfoldings In order to treat infinitesimal unfoldings and its relations with deforma- tions of foliations we will need to consider non-reduced schemes, and gen- eralize the notion of a foliation to this setting. Luckily, a straightforward generalization will serve our purposes just fine. Definition 2.1. Let X be a (separated) scheme of finite type over an al- gebraically closed field k. An involutive distribution over X is a subsheaf T F ⊆ T X that is closed under the Lie bracket, i.e.: for every open sub- scheme U ⊂ X we have [T F(U ), T F(U )] ⊆ T F(U ). UNFOLDINGS AND STRUCTURAL THEOREMS 3 The annihilator of an involutive distribution I(F) := ann(T F) ⊆ Ω1 integrable Pfaff system. An integrable pfaff system I verifies the equation X is an d(I(F)) ∧ r^ I(F) = 0 ⊂ Ωr+2 X ; X → Ωj+1 where d : Ωj of the sheaf Ω1 the foliation and noted dim F. X is the de Rham differential, and r is the generic rank X /I. The generic rank of T F will be called the dimension of Although there is no form of the Frobenius integrability theorem valid for any scheme (possibly non-reduced) of finite type over a field, we will still refer to the data of an integrable Pfaff system or an involutive distribution as a foliation. Note that with this definition foliations may have singularities. Definition 2.2 (Pull-back of foliations). Given a morphism p : X → Y and a foliation on Y defined by a Pfaff system I(F) ⊆ Ω1 Y . By considering X (p−1U ) the pull-back of 1-forms we define a foliation p∗F p∗ : Ω1 on X by defining its Pfaff system to be the one generated by the pull-backs of forms in I(F). We call this the pull-back foliation of the foliation F. Y (U ) → Ω1 Definition 2.3 (Unfolding). Let X be a non-singular algebraic variety over an algebraically closed field and let F0 be a foliation on X. Let S be any scheme (of finite type over the base field of X) and s ∈ S a closed point. We denote by π1 and π2 the projections of X × S to X and S respectively. We denote by Dπ2 the differential map Dπ2(x.s) : T (X × S) ⊗ k((x, s)) → T S ⊗ k(s). An unfolding of F0 parametrized by (S, s) is a foliation F on X × S such that (1) The restriction of F to X × s is F0 i.e.: if we take the pull-back foliation ι∗(F) as in the above definition we get ι∗(F) = F0, here ι : X × s → X × S is the inclusion. (2) Dimensions of F and F0 are related as thus: dim F = dim F0 + dim S. In the case where X and S are non-singular varieties over C, the leaves of the foliation F0 are contained in the larger dimensional leaves of the unfolding F. Now we remind the definition of the relative tangent sheaf. Given a morphism f : X → Y the relative tangent sheaf Tf X is the dual of Ω1 X Y . Remember that Tf F is naturally a sub-sheaf of the tangent sheaf T X , its sections are the vector fields on T X that are tangent to the fibers of f . In the case where f = π2 : X × S → S is the projection we note Tπ2(X × S) = TS(X × S) and similarly with the other projection. 4 FEDERICO QUALLBRUNN Remark 2.4. In the case of the product of X and S we have the decompo- sition of sheaves. T (X × S) ∼= TS(X × S) ⊕ TX(X × S). Moreover we have TS(X × S) ∼= π∗ 1(T X), TX(X × S) ∼= π∗ 2(T S), where π1 and π2 are the projections. Remark 2.5. Let F0 be a foliation over a variety X and F an unfolding of F0 parametrized by (S, s0). Then F induces a family of foliations over X parametrized by S (see [7]) in the following way: Let I(F) be the Pfaff system associated with F. Let p : Ω1 X×S → Ω1 X×SS be the projection from the sheaf of differential to the sheaf of relative differentials. We set IS(F) := p(I(F)) ⊆ Ω1 X×SS. Note that IS(F) is a family of integrable Pfaff systems in the sense of [7] such that its restriction to s0 is I(F0). We can calculate the annihilator TSF of IS(F) which will be, of course, a family of involutive distributions. Indeed we obtain TS(F) as the intersection of the subsheaves T F and TS(X × S) of T (X × S). In general, given a family of foliations, it is not possible to glue the leaves of the different foliations in the family to higher dimensional leaves. So not every family of foliation comes from an unfolding, as a matter of fact, that is almost never the case. Definition 2.6 (Isotriviality). We say the unfolding F of F0 is isotrivial if it induces a trivial family of Pfaff systems (equivalently of distributions), i.e.: if IS(F) = π∗ X×SS, where π1 : X × S → X is the projection and the morphism π∗ X×S is the pull-back of forms. 1I(F0) as subsheaves of Ω1 1 : Ω1 X ⊗ OX×S → Ω1 Definition 2.7 (Transversality). Given a foliation F0 and an unfolding F of F0 parametrized by (S, s0) we have exact sequences. 0 0 / TSF TS(X × S) / NSF / 0 / T F / T (X × S) / N F / 0. Where TSF is defined as in Remark 2.5. We say F is transversal to S if N F/NSF = 0. Example 2.8. Let X = An = Spec(k[x1, . . . , xn]) and S = A1 = Spec(k[y]) be affine spaces. Let F0 be a foliation on An with tangent sheaf T F0 = k[x1, . . . , xn] · (X1, . . . , Xr), the Xi's being vector fields on An. Lets see what an isotrivial and transversal unfolding F of F0 parametrized by A1 should look like. / / / _ _ / _ / / / / / UNFOLDINGS AND STRUCTURAL THEOREMS 5 In the first place, being isotrivial, the subsheaf TA1F ⊆ T F will be generated by the vector fields X1, . . . , Xr, viewed as vector fields on An ×A1. So we can write T F as generated by vector fields X1, . . . , Xr, Y1, . . . , Ys. Now as F is an unfolding of F0, then Y1, . . . , Yr must span a space of dimension ≤ 1 on the tangent space to each point. So we have T F = k[x1, . . . , xn, y] · (X1, . . . , Xr, Y ). Moreover, as F is transversal over A1, we can write Y = eY + ∂ ∂ ∂y , where eY = f1(x, y) ∂ ∂x1 + · · · + fn(x, y) . ∂xn Then the involutivity of T F implies that, for each fixed y0 ∈ A1 we have a vector field eY (−, y0) on An that verifies Note that every family eY of vector fields on An satisfying the above condition [T F0,eY (−, y0)] ⊆ T F0. give rise to an isotrivial transversal unfolding of F0. Essentially, the same is true for any isotrivial transversal unfolding of a foliation on a non-singular variety X, although the role of the family of vector fields eY will be taken by a section of a certain sheaf on the parameter space S. The rest of the section is devoted to the generalization of the above example. Definition 2.9. Remember that, given an involutive distribution T F, the bracket of vector fields defines a map [−, −] : T F ⊗ N F → N F known as Bott connection. Using the Bott connection we will define a subsheaf of NSF, which we call u(F), as the subsheaf of NSF whose local sections on an open set V ⊂ X × S are the following: u(F)(V ) :=Def {s ∈ NSF(V ) s.t.: [TSF, s] = 0}. Note that u(F) is not a coherent subsheaf of NSF but only a subsheaf of k-vector spaces. Notation 1. We denote Υ(F0) :=Def Γ(X, u(F0)). Remark 2.10. The sheaf u(F) inherits from TS(X × S) the structure of a sheaf of Lie algebras. Indeed, if W ⊆ TS(X × S) is the preimage of u(F) by the projection TS(X × S) → NS(F), then TSF is an ideal inside of W . In particular Υ(F) is a Lie algebra over the field of definition k. The sheaf u(F) will be useful to study the relation between infinitesimal unfoldings (i.e.: those parametrized by the Spec of artinian algebras) and unfoldings parametrized by varieties such as P1 k. Let us begin by considering an unfolding F of a foliation F0 parametrized by S, that is a foliation over X × S. By Remark 2.4 we have a projection T (X × S) → TS(X × S), from this we get a projection T (X × S)/TSF → TS(X × S)/TSF = NSF. 6 (1) FEDERICO QUALLBRUNN We will focus now on the properties of the map υF one gets by composing T F/TSF → T (X × S)/TSF → NSF. Note that, when the unfolding is transversal, we have an isomorphism 2T S ∼= T F/TSF, so we can consider υF to be a morphism π∗ υF : π∗ 2T S → NSF. Proposition 2.11. If F is transversal to S then υF (π−1 of u(F) ⊆ NSF. 2 T S) is a subsheaf Proof. Note that the statement is making reference to π−1 2 T S ⊂ π∗ that is the sheaf of vector fields that are constant along the fibers of π2. So, given a local section s ∈ υF (π−1 induces an isomorphism between T F/TSF and π∗ To do this we take a pre-image of s, say es ∈ T F. As p2 : T (X × S) → π∗ we can take es of the form Y + Z with Y ∈ π∗ 2 T S) we need to compute [TSF, s]. 2T S 2 T S), 1 T S. Given 2T S and s ∈ υF (π−1 1T X and Z ∈ π−1 W ∈ TSF we compute 2T S, as W (Z) = 0, being Z in π−1 so it is in TSF. Therefore [TSF, s] = 0. [W,es] = [W, Y + Z] = [W, Y ] − Z(W ), 1 T S. Then [W,es] is in π∗ 1T X, and also in T F, (cid:3) Theorem 2.12. Let X be a non-singular variety and F0 a foliation on X. There is, for each scheme S, a 1 to 1 correspondence: (cid:26) isotrivial transversal unfoldings of F parametrized by S(cid:27) ←→  Proof. sections υ ∈ H 0(S, Ω1 S) ⊗ Υ(F0) 1 2 s.t.: dυ + [υ, υ] = 0 .   1.Form associated to an unfolding: Given an isotrivial transver- sal unfolding F of F0 we associate to it the map T F/TSF → NSF, as in eq. (1). As F is transversal we have T F/TSF ∼= TX (X × S) ∼= π∗ 2(T S). So we have a map υF : π∗ 2(T S) → NSF. By Proposition 2.11, the image of π−1 of u(F). Now, F being isotrivial implies that NSF ∼= π∗ that u(F) ∼= π∗ 2 (T S) under υF is a subsheaf 1(N F0) and 1(u(F0)). Then we have a global section υF ∈ H 0(π∗ 2Ω1 by Kunneth isomorphism this is a section in H 0(Ω1 and by Proposition 2.11 this section actually belongs to H 0(Ω1 H 0(u(F0)). So we get a global form S ⊗ π∗ 1(N F0)), S) ⊗ H 0(N F0) S) ⊗ υF ∈ H 0(S, Ω1 S) ⊗ Υ(F0) UNFOLDINGS AND STRUCTURAL THEOREMS 7 associated with an isotrivial transversal unfolding F parametrized by S. Now to establish the first part of the correspondence we need the following: Lemma 2.13. Let F0 be a foliation on a nonsingular variety X, and F an isotrivial and transversal unfolding of F0 parametrized by a variety S. The global 1-form υF verifies the Maurer-Cartan equation: dυF + 1 2 [υF , υF ] = 0. Proof of Lemma 2.13. Set υ = υF , now given two vector fields Y and Z defined on S we use Cartan's formula for the differ- ential dυ(Y, Z) = Y (υ(Z)) − Z(υ(Y )) − υ([Y, Z]). Now we want to calculate υ([Y, Z]), for this we look at the definition of υF . Given a local section Y of T S we look at it as a section of π∗(T S). As was noted above, transversality of F gives us an isomorphism T F/TSF ∼= π∗ 2(T S), so we can associate to Y a corresponding global section of T F/TSF. We then apply the restriction of the morphism p1 : T (X × 1(T X) of Remark 2.4 to the above section obtaining by S) → π∗ Proposition 2.11 a section of u(F). with Y ′ in π∗ So eY viewed as a section of T (X × S) is of the form Y + Y ′, 1(u(F)). Then we have Y ′ = υ(Y ). To calculate υ([Y, Z]) we first observe that Indeed, the isomorphism between T F/TS and π∗ 2(T S) that we use ^[Y, Z] = [eY , eZ]. to define eY comes from the inclusion of T F in T (X × S), so it respects Lie brackets. So now υ([Y, Z]) is simply p1([eY , eZ]). We can write eY = Y + Y ′ and similarly with Z and, noting p1([eY , eZ]) = p1([Y + Y ′, Z + Z ′]) = [Y ′, Z ′] + p1([Y, Z ′]) − p1([Z, Y ′]). that p1([Y, Z]) = 0, compute Now, as p1([Y, Z ′]) = Y (Z ′) we have υ([Y, Z]) = [υ(Y ), υ(Z)] + Y (υ(Z)) − Z(υ(Y )), from which the Maurer-Cartan equation, and therefore the lemma, follows. (cid:3) 8 FEDERICO QUALLBRUNN 2.Unfolding associated with a form: Given υ ∈ H 0(S, Ω1 S)⊗Υ(F0) we have an associated morphism 2(T S) → π∗ υ : π∗ 1(N F0). So we consider the morphism 2(T S) → π∗ υ ⊕ id : π∗ 1(N F0) ⊕ π∗ 2(T S) given by υ ⊕ id(s) = (υ(s), s). We also have the projection φ : T (X × S) → π∗ 2(T S). So we consider the diagram 1(N F0) ⊕ π∗ 2(T S) υ⊕id−−−→ π∗ π∗ 1(N F0) ⊕ π∗ 2(T S) φ ←− T (X × S). We then take T Fυ ⊆ T (X × S) to be the sub-sheaf generated by φ−1(υ(π∗ 2(T S))). The fact that the image of υ is within Υ(F0) and that υ satisfies the Maurer-Cartan equation implies that T Fυ is involutive. Indeed, let eY and eZ be local sections of T Fυ, we need to check that [eY , eZ] is also a section of T Fυ. So we take have section Y and 2(T S) such that Z in π∗ φ(eY ) = (υ(Y ), Y ) φ(eZ) = (υ(Z), Z)). Moreover, we may assume that Z and Y are local sections of π−1 2 (T S), as this latter sheaf generates π∗ 2(T S) and the general result will follow from the fact that the Lie bracket is a derivation on both of its inputs. So, given that Y, Z ∈ π−1 2 (T S) we have φ([eY , eZ]) = [φ(eY ), φ(eZ)] = = [Y, Z] + Y (υ(Z)) − Z(υ(Y )) + [υ(Y ), υ(Z)] = = [Y, Z] + υ([Y, Z]). So, T Fυ is involutive. Also, by construction, Fυ induces a trivial family. Hence Fυ is an isotrivial, transversal unfolding of F0, associated to an υ ∈ H 0(S, Ω1 S) ⊗ Υ(F0). It follows routinely that the constructions of υF and of Fυ are inverse to each other. (cid:3) We thus see that the space Υ(F) will be an important ingredient in studying isotrivial unfoldings of a foliation F. This space is acted upon by the group of automorphism of the foliation, that is the group Aut(F) := {g ∈ Aut(X) s.t.: g∗(T F) = T F}. The Lie algebra of this group may be naturally identified with the global sections of the sheaf aut(F) whose local sections are aut(F)(V ) := {θ ∈ T X(V ) s.t.: [T F, θ] ⊆ T F}, UNFOLDINGS AND STRUCTURAL THEOREMS 9 so Lie(Aut(F)) = H 0(X, aut(F)). Remark 2.14. Note also that we have a short exact sequence of sheaves 0 → T F → aut(F) → u(F) → 0. In the particular case when H 0(X, T F) = H 1(X, T F) = 0, which will be important to us later, we have the equality (2) Lie(Aut(F)) = H 0(X, aut(F)) = H 0(X, u(F)) = Υ(F). In particular, in the case where X a complex variety and H 0(X, T F) = H 1(X, T F) = 0, there is an unfolding associated to the Maurer-Cartan form of the group Aut(F), which in this case take values in Υ(F). So there is, by Theorem 2.12, an unfolding associated to the Maurer-Cartan form. In the situation where X and S be varieties over C, F0 a foliation on X such that H 0(X, T F0) = H 1(X, T F0) = 0 and F a transversal isotrivial unfolding of F0 parametrized by S. Denote π : eS → S the universal covering of S, υM C the Maurer-Cartan form of Aut(F0). Let FM C the unfolding as- sociated to the Maurer-Cartan form (cf.: Remark 2.14), and υ ∈ Ω1 ⊗Υ(F0) eS the pull-back of υF by the universal covering map. A direct application of Darboux's existence theorem gives us: Corollary 2.15. With hypotheses as in the above paragraph, there is a mor- phism f : eS → Aut(F0) such that υ is the pull-back of υM C. Equivalently, f ∗(FM C ) = π∗(F) as unfoldings of F0. Example 2.16. Let F0 be a foliation by curves on the complex projective plane P2(C) such that H 0(X, aut(F)) 6= 0. Then T F0 is a line bundle on P2, so H 0(X, T F0) = H 1(X, T F0) = 0. Also, being a foliation by curves in P2 implies that dim Aut(F0) ≤ 1. So there is an infinitesimal symmetry Y such that H 0(X, aut(F)) = (Y ) and Aut(F0)0 = exp(tY ) ∼= C∗, where Aut(F0)0 is the connected component of the identity. In particular, we have that the universal covering A of Aut(F0)0 is isomorphic as a complex variety to C. On A we have the Maurer-Cartan form υM C = dz ⊗ Y, here we are taking z to be a coordinate of A ∼= C and we are identifying Lie(A) = Lie(Aut(F0)) = (Y ). Considering that Lie(Aut(F0)) = Υ(F0) and applying Theorem 2.12, this gives us the unfolding F such that: T F = (π∗ 1T F0 ⊕ ( ∂ ∂z + Y )) 10 FEDERICO QUALLBRUNN on P2 × C. If ω is the rational 1-form on P2 annihilating T F0, then 1-form annihilating T F is = ω + ω(Y )dz ∈ Ω1 P2×C. so F posses the integrating factor ω(Y ) (considered as a rational function on P2 × C). Note that ω(Y ) considered as a rational function on P2 is an integrating factor for F0. Remark 2.17. In a sense, the previous results generalize those of Suwa in [11]. There is proven, in the context of codimension 1 foliations on Pn, a correspondence between infinitesimal isotrivial unfoldings (i.e.: isotrivial unfoldings parametrized by k[x]/(x2)) and rational integrating factors of the foliation. In our context this is understood the following way: An infinitesimal isotrivial unfolding of a foliation F will be transversal on some open set ι : U ֒→ Pn. As we have seen, a transversal isotrivial unfolding of ι∗F give rise to a global section s ∈ Υ(ι∗F ). Restricting the open set U further if needed we can take a section Y in Lie(Aut(ι∗F )) representing s modulo T ι∗F . In other words Y is a rational symmetry of F . It is well known, see [6], that to every rational symmetry of a codimension 1 foliation in a complete variety corresponds a rational integrating factor. Thus we can recover Suwa's theorem from [11]. 3. Foliations on Pn viewed as unfoldings In this section we will be interested in foliations on Pn(C) up to birational equivalence. Moreover we will study the relations of foliations of arbitrary dimension on Pn(C) with foliations by curves on projective spaces of lower dimension. First we recall an important definition. Definition 3.1. A codimension q foliation F on Pn is said to be of degree d if and only if the associated integrable Pfaff system I(F) have the property that ∧qI(F) ∼= OPn(−d − q − 1). Given a foliation F on Pn(C) of codimension q, we fix a (rational, linear) q+1 be the Grassmannian of q + 1- projection p : Pn 99K Pq+1. Now, let Grn dimensional linear spaces on Pn, define the open set U = {P ∈ Grn q+1 s.t.: pP : P → Pq+1 is a (regular) isomorphism} ⊂ Grn q+1. Then we have for every P ∈ U a foliation in Pq+1 given by first restricting F to P and then applying the isomorphism p. Note that if F is of degree d so is every foliation on Pq+1 obtained this way. In other words, what we are doing here is considering the incidence q+1, intersecting with correspondence Z = {(x, P ) s.t.: x ∈ P } ⊂ Pn × Grn Pn × U gives us Z ∩ (Pn × U ) ∼= Pq+1 × U UNFOLDINGS AND STRUCTURAL THEOREMS 11 and hence we have a diagram Pq+1 × U π1 ztttttttttt Pn π2 $■■■■■■■■■■ U. 1F as a foliation on Pq+1 × U . Now, So taking the pull-back of F we have π∗ we can take the family of foliations over U induced by π∗ 1F, as in Section 2. Restricting U if necessary, we may assume that π∗ 1F induces a flat fam- ily of involutive distributions over Pq+1, parametrized by U , lets call this family FU . We can characterize FU as follows, to each point u ∈ U corre- sponds a q + 1−linear space P and π2 is the projection from the incidence correspondence, so FU Pq+1×{u} = p(FP ). By the results of [7, Proposition 6.3] such a family defines a morphism between U and the moduli space of involutive distributions over Pq+1 of codimension q and degree d. This later space is PH 0(Pq+1, T Pq+1(d − 1)). Then we have a morphism φF : U → PH 0(Pq+1, T Pq+1(d − 1)), the later being a projective space of dimension (q + 2)(cid:0)d+q+1 particular if the dimension of U (which is that of Grn of the target space, the morphism will have fibers of positive dimension. q+1) is greater than that (cid:1) −(cid:0)d+q d−1(cid:1). In d Remark 3.2. Suppose the dimension of Grn q+1 is greater than the dimension of PH 0(Pq+1, T Pq+1(d − 1)). Set V to be a fiber of φF , so φ−1 F (x) = V ⊆ U , note that V must have positive dimension. Then it follows from the universal property of the moduli space that the family FU V is trivial. In particular the foliation π∗ 1FPq+1×V defines an isotrivial unfolding parametrized by V . 4. Generic transversality Now we investigate conditions under which an isotrivial unfolding turns out to be also transversal, so we can apply to it the theory of Section 2. Definition 4.1. Let F be a foliation on X × S viewed as an unfolding of foliations on X. We define TF to be the schematic support of the sheaf N F/NSF. The subscheme TF will be of interest as it is the locus of points where transversality fails. Recall that sing(F), the singular locus of a foliation F on a scheme X , is defined to be the schematic support of the sheaf Ext1 X (N F, OX ) (local Ext). Similarly, if we have a family of involutive distributions parametrized by a scheme S its singular locus is the scheme theoretic support of the sheaf Ext1 X (NSF, OX×S ). z $ 12 FEDERICO QUALLBRUNN Lemma 4.2. Let X and S be non-singular varieties. Let F be an unfold- ing of foliations on X parametrized by S. Suppose that F is an isotriv- ial unfolding of a foliation F0 on X, and that TF ∩ sing(F) = ∅. Then TF ⊆ sing(F0) × S. Proof. As F is an isotrivial unfolding sing(F0) × S is the singular locus of the (trivial) family of involutive distributions induced by F. So let p /∈ sing(F0) × S, then the localization (NSF)p is a free OX×S -module and the short sequence 0 → TSF ⊗ k(p) → TS(X × S) ⊗ k(p) → NSF ⊗ k(p) → 0 is exact. If p /∈ sing(F) then the sequence 0 → T F ⊗ k(p) → T (X × S) ⊗ k(p) → N F ⊗ k(p) → 0 is exact. On the other hand we always have an immersion TS(X × S) ⊗ k(p) ֒→ T (X × S) ⊗ k(p). Then if p is a point neither in sing(F0) × S nor in sing(F) we have an immersion So we have another immersion TSF ⊗ k(p) ֒→ T F ⊗ k(p). T F/TSF ⊗ k(p) ֒→ TX (X × S). As dim T F = dim T F0 + dim S the dimension of the above vector spaces are equal to dim S, so N F/NSF ⊗ k(p) = 0. Then, if p /∈ sing(F0) × S and p /∈ sing(F), the point p is not in TF . Lemma 4.3. Let X and S be non-singular varieties over C, denote π1, π2 the projections of X × S to the first and second factor, respectively. Let F be an isotrivial unfolding of a foliation F0 on X parametrized by S. Sup- pose that F0 is non-singular and that dim(sing(F)) ≤ dim(F) − 1. Then π2sing(F ) : sing(F) → S is not dominant. (cid:3) Proof. This follows from Theorem 2.7 of [10], actually we will use the following weaker version of the theorem in [10]: If we take the reduced structure sing(F)red ⊆ sing(F) then, at a regular point p of sing(F)red, the image of the map T F ⊗ k(p) → T (X × S) ⊗ k(p) falls within the tangent space to sing(F)red at p. To prove our assertion suppose that sing(F) is dominant over S. Take p to be a regular point of sing(F)red. Then sing(F)red is dominant over S as well, so it has dimension at least that of S. On the other hand by Theorem 2.7 of [10] the image of (T F0) ⊗ k(π1(p)) ∼= π∗ 1(T F0) ⊗ k(p) ∼= TSF ⊗ k(p) → T (X × S) ⊗ k(p) UNFOLDINGS AND STRUCTURAL THEOREMS 13 falls within the tangent space to sing(F)red at p. And, as F0 is a non-singular foliation, we have that the dimension of that image is that of the leaves of T F0. As TSF is tangent to the fibers of π2 the differential of the projection Dπ2 satisfy Dπ2(TSF) = 0. Then, comparing dimensions of tangent spaces we get dim(sing(F)) ≥ dim F0 + dim S = dim F, obtaining a contradiction, so sing(F) cannot be dominant over S. (cid:3) Putting together the last two lemmas we obtain the following. Proposition 4.4. Let F be an isotrivial unfolding of a foliation F0 on a non-singular variety X parametrized by a non-singular variety S, such that dim(sing(F)) ≤ dim(F) − 1. Set Y = X \ sing(F0). Then there is an open set U ⊂ S such that the restriction of F to Y × U is a transversal isotrivial unfolding. 5. Unfoldings of Foliations by curves Now we are in conditions of applying the results of Section 2 to foliations If the on Pn. Let F be a foliation of degree d in projective space Pn. condition (n − q − 1)(q + 1) > q + 2(cid:18)d + q + 1 d (cid:19) −(cid:18)d + q d − 1(cid:19) is satisfied, we are in the situation of Remark 3.2. Then there is an open set U of Grn q+1 and a projection p : Pn 99K Pq+1 trivializing the incidence correspondence, in such a way that we get a diagram Pq+1 × U π1 ztttttttttt Pn π2 $■■■■■■■■■■ U. And through each point in U passes a closed subscheme V ⊆ U such that π∗FPq+1×V is an isotrivial unfolding of a foliation by curves F0 in Pq+1 parametrized by V . Indeed, we can get F0 by taking a plane P representing a point in V and so we have F0 = p(FP ). We can further restrict V to its reduced structure V red and then even more to the non-empty open set of regular points of V red in order to be able to apply Proposition 4.4. By Proposition 4.4, there is an open subset Y of Pq+1 and an open subset W of V red such that the restriction of π∗FPq+1×V to Y × W is a transversal and isotrivial unfolding of F0Y . Now, we have two possibilities, either F0 have rational infinitesimal symmetries or not. If F0 have no rational symmetries, then the sheaf aut(F0) is trivial. Then, because of the short exact sequence 0 → T F0 → aut(F0) → u(F0) → 0, z $ 14 FEDERICO QUALLBRUNN also u(F0) = 0. As was said before, by Proposition 4.4 we can restrict the unfolding π∗FPq+1×V to Y × W in such a way that the unfolding is now transversal and isotrivial on Y × W . Note that u(F0Y ) = u(F0)Y = 0. So the unfolding is trivial and thus π∗FPq+1×V is birationally equivalent to the pull-back of a foliation by curves on Pq+1. Notice also that foliations with no rational symmetries form a dense open set on PH 0(Pq+1, T Pq+1(d − 1)), so either a dense open set of P ∈ U verify that FP have no rational symmetries or, on the contrary, every P ∈ U is such that FP have a rational symmetry. Summarizing we have proved the following. Proposition 5.1. Let F be a degree d foliation in Pn of codimension q. Suppose the condition (n − q − 1)(q + 1) > (q + 2)(cid:18)d + q + 1 d is satisfied. Then there is an open set U ⊂ Grn (cid:19) −(cid:18)d + q d − 1(cid:19) q+1 and a rational morphism Pq+1 × U π→ Pn such that the following alternative holds. Either one have that for every q + 1-linear subspace P ∈ U , the restriction FP have rational symmetries; or there is, for each P in a dense open subset of U , a subvariety VP ⊆ U of codimension at most (q + 2)(cid:0)d+q+1 pull-back of a foliation by curves on Pq+1. (cid:1) −(cid:0)d+q d−1(cid:1) such that π∗FPq+1×VP is the d In codimension 1 we can improve this result. Theorem 5.2. Let F be a foliation of codimension 1 and degree d on Pn(C). Suppose dim(sing(F)) ≤ n − 2 and the condition 2(n − 2) > 3(cid:18)d + 2 d (cid:19) −(cid:18)d + 1 d − 1(cid:19) is satisfied. Then we have the following alternative: 2 (1) Either there exist an open subset U of Grq+1 , and a map π : U × P2 → Pn such that for each P ∈ U there is a subvariety VP ⊆ U π∗FP2×VP is pull-back of a foliation by curves on P2. containing P and of codimension at most 3(cid:0)d+2 d−1(cid:1) such that (2) Or there are holomorphic varieties eS and eY and a meromorphic map with discrete (not necessarily finite) generic fiber, φ : eS ×P2 → Pn, such that φ∗F has a meromorphic first integral. d (cid:1) −(cid:0)d+1 Proof. The inequality in the hypotheses allow us to apply Proposi- tion 5.1. Then, either we have the map π, and subvarieties VP as in Propo- sition 5.1 or we have that for every 2-dimensional linear subspace P ⊂ Pn the restriction FP is a foliation with rational symmetries. UNFOLDINGS AND STRUCTURAL THEOREMS 15 In the second case FP , has a rational integrating factor f . Then, by has a first integral. Then (p × id)∗φ∗F is a transveral isotrivial unfolding of a foliation over [3], if we take p : eY → P \ Div(f ) to be the universal covering map, p∗FP eY parametrized by S. It is an unfolding of a foliation with a first integral. As (p × id)∗φ∗F is the unfolding of a foliation with a first integral defined on a simply connected space, then by [11, 5.3], (p × id)∗φ∗F has itself a first integral (Suwa's original result implies the existence of a local first integral, which we can extend globally on account of Y being simply connected). (cid:3) We can express this result more succinctly as follows. Corollary 5.3. Let F be a foliation in Pn(C) satisfying the hypotheses of Theorem 5.2. Then either F is given by a closed rational form or a generic leaf of F contains algebraic varieties of codimension at most 3(cid:0)d+2 d (cid:1) −(cid:0)d+1 d−1(cid:1). References [1] Dominique Cerveau and Alcides Lins Neto, A structural theorem for codimension one foliations on Pn , n ≥ 3, with application to degree three foliations, Ann. Sc. Norm. Super. Pisa Cl. Sci. 12 (2013), no. 1, 1 -- 41. [2] Dominique Cerveau, Alcides Lins Neto, Frank Loray, Jorge Vitorio Pereira, and Fr´ed´eric Touzet, Complex codimension one singular foliations and Godbillon-Vey se- quences, Moscow mathematical journal 7 (2007), 21 -- 54. [3] Dominique Cerveau and J.F. Matt´ei, Formes int´egrables holomorphes singuli`eres, Soci´et´e Math´ematique de France, 1982. [4] Gael Cousin and Jorge Vit´orio Pereira, Transversely affine foliations on projective manifolds, arXiv preprint arXiv:1305.2175 (2013). [5] Frank Loray, Jorge Vitorio Pereira, and Fr´ed´eric Touzet, Singular foliations with trivial canonical class, arXiv preprint arXiv:1107.1538 (2011). [6] Jorge Vit´orio Pereira and Percy Fern´andez S´anchez, Transformation groups of holo- morphic foliations, Comm. Anal. Geom 10 (2002), no. 5, 1115 -- 1123. [7] F. Quallbrunn, Families of distributions and Pfaff systems under duality, arXiv:1305.3817. [8] Tatsuo Suwa, A theorem of versality for unfoldings of complex analytic foliation sin- gularities, Inventiones mathematicae 65 (1981), no. 1, 29 -- 48. [9] [10] [11] , Unfoldings of complex analytic foliations with singularities, Japanese journal of mathematics. New series 9 (1983), no. 1, 181 -- 206. , Structure of the singular set of a complex analytic foliation, Preprint series in mathematics. Hokkaido University 33 (1988). , Unfoldings of codimension one complex analytic foliation singularities (1992). Departamento de Matem´atica, FCEyN, Universidad de Buenos Aires, Ciu- dad Universitaria, Pabell´on 1, Buenos Aires (Argentina) E-mail address: [email protected]
1604.00726	4	1604	2018-03-01T09:25:12	Boundary combinatorics of orthogonal modular 4-folds	[ "math.AG" ]	We study combinatorial problems related to the singularities and boundary components of toroidal compactifications of orthogonal modular varieties. In particular, those associated with the moduli of algebraic deformation generalised Kummer 4-folds.	math.AG	math	BOUNDARY COMBINATORICS OF ORTHOGONAL MODULAR 4-FOLDS MATTHEW DAWES Abstract. We study combinatorial problems related to the singularities and boundary components of toroidal compactifications of orthogonal modular varieties. In particular, those associated with the moduli of algebraic deformation generalised Kummer 4-folds. 1. Introduction Toroidal compactifications of modular varieties provide a rich source of combinatorics. Two natural problems are to count boundary components, and to describe the singular locus. Both of these problems have been studied in detail for Siegel modular 3-folds [HKW91,HKW93], but there are fewer results for orthogonal modular varieties. In particular, while there are some results for orthogonal modular varieties of large dimension [Sca87], less is known about those of small dimension. The purpose of this article is to study these problems for orthogonal modular 4-folds F2d associated with the moduli of deformation generalised Kummer varieties of dimension 4, with polarisation of split type, and degree 2d, where d = p2 for an odd prime p. In Theorem 5.7 we produce bounds for the number of boundary curves of F2p2. In Theorem 5.12 we bound the number of components of the singular locus of a toroidal compactification F tor 2p2 in the neighbourhood of a boundary curve. To the best of the author's knowledge, these are the first such results for F2d, and the first such results for orthogonal modular 4-folds. (Li( ∗ ∗ ),Li Cai) The finite quadratic formLi( ∗ 2. Notation χX Cr div(x) D(M ) DN F2d eΓ Γ+ h⊥ L L2x,2y M ∨ O(L, h) O(M ) O+(M ) eO(M ) φ Φr ∗ ) (with values in Q/2Z) on the abelian group ⊕iCai. The characteristic polynomial of the matrix X. The cyclic group of order r. The divisor of x ∈ M for a lattice M . (i.e. the positive generator of the ideal (x, M ).) The discriminant group of the (even) lattice M [Nik79]. A connected spinor component of the domain ΩN = {[x] ∈ P(N ⊗ C) (x, x) = 0, (x, x) > 0} for a lattice N of signature (2, n). The orthogonal modular variety Dh⊥/ O+(L, h) where h ∈ L is primitive, of split type, and degree 2d. The intersection Γ ∩ eO(M ) for Γ ⊂ O(M ). The intersection Γ ∩ O+(M ) for Γ ⊂ O(M ). The orthogonal complement h⊥ ⊂ L for h ∈ L. The lattice U ⊕3 ⊕ h−2(n + 1)i. The lattice U ⊕2 ⊕ h−2xi ⊕ h−2yi. The dual of the lattice M . The group {g ∈ O(L) gh = h} for h ∈ L. The orthogonal group of a lattice M . The kernel of the real spinor norm on O(M ⊗ R). The stable orthogonal group eO(M ) = {g ∈ O(M ) gD(M ) = id}. The Euler φ-function. The r-th cyclotomic polynomial. 2010 Mathematics Subject Classification: Primary 14G35; Secondary 14M27. Key words and phrases: orthogonal modular variety; generalised Kummer variety; toroidal compactification. 1 2 MATTHEW DAWES 3. Moduli of Deformation generalised Kummer varieties Let A be an abelian surface, and let A[n+1] be the Hilbert scheme parametrising (n + 1)-points on A. The Hilbert scheme A[n+1] inherits an addition from A, and so there is a projection p : A[n+1] → A. The fibre p−1(0) is an irreducible symplectic (compact hyperkahler) manifold known as a generalised Kummer variety [Bea83]. A deformation X of p−1(0) is an irreducible symplectic manifold known as a deformation generalised Kummer variety. There is a lattice structure L on H 2(X, Z) defined by the Beauville-Bogomolov-Fujiki [Bea83] form. By the results of Rapagnetta [Rap08], L is equal to L = U ⊕3 ⊕ h−2(n + 1)i, where h−2(n + 1)i is the rank 1 lattice generated by a vector of square length −2(n + 1) and U is the hyperbolic plane. A basis {e, f } for U is said to be standard or a standard basis for U if its Gram matrix is given by (cid:18)0 1 1 0(cid:19) . A choice of ample line bundle L ∈ Pic(X) defines a polarisation for X. The first Chern class of L defines a vector h := c1(L) ∈ L. The degree 2d of L is defined as the square length h2, and the polarisation type of L is defined as the O(L)-orbit of h. We assume all polarisations are primitive; that is, the vector h is primitive in the lattice L. Let O+(L, h) be the subgroup of O(L, h) consisting of all elements of spinor norm 1, and let Dh⊥ be a connected component of the quadric Ωh⊥ = {[x] ∈ P(h⊥ ⊗ C) (x, x) = 0, (x, x) > 0} preserved by the kernel of the real spinor norm on O(h⊥ ⊗ R). There is a GIT quotient M parametrising deformation generalised Kummer varieties of fixed di- mension and polarisation type. By Theorem 3.8 of [GHS13], for every connected component M′ of M there is a finite-to-one dominant morphism ψ : M′ → F to an orthogonal modular variety F (i.e. a quotient of a Hermitian symmetric domain of type IV by an arithmetic subgroup of O(2, m)). Here, the orthogonal modular variety F is given by F = O+(L, h)\Dh⊥ . Proposition 3.1. Suppose h ∈ L is primitive of length 2d > 0 with div(h) = f . Let g =(cid:0)2(n + 1)f −1, 2df −1(cid:1), w = (g, f ), g = wg1, and f = wf1. Then 2(n + 1) = f gn1 = w2f1g1n1 and 2d = f gd1 = w2f1g1d1, where (n1, d1) = (f1, g1) = 1. (1) If g1 is even, then h exists if and only if (d1, f1) = (f1, n1) = 1 and d1/n1 is a quadratic residue (2) If g1 is odd and f1 is even, or f1 and d1 are both odd, then such an h exists if and only if where w = w+(f1)w−(f1), w+(f1) is the product of all powers of primes dividing (w, f1), ρ(n+1) is the number of prime factors of n + 1, and φ is the Euler function. modulo f1. Moreover, the number of eO(L)-orbits of h with fixed f is equal to w+(f1)φ(w−(f1)).2ρ(f1), (d1, f1) = (t1, 2f1) = 1 and −d1/n1 is a quadratic residue modulo 2f1. The number of eO(L)- orbits is equal to w+(f1)φ(w−(f1)).2ρ(f1/2) if f1 is even, and to w+(f1)φ(w−(f1)).2ρ(f1) if f1 and d1 are both odd. (3) If g1 and f1 are both odd and d1 is even, then such an h exists if and only if (d1, f1) = (n1, 2f1) = 1, −d1/(4t1) is a quadratic residue modulo f1, and w is odd. In such a case, the (4) If c ∈ Z (determined modulo f ) satisfies (c, f ) = 1 and b = (d + c2(n + 1))/f 2, then number of eO(L)-orbits of h is equal to w+(f1)φ(w−(f1)).2ρ(f1). h⊥ ∼= 2U ⊕ B, where B = −2b c 2(n+1) f c 2(n+1) f −2t  . BOUNDARY COMBINATORICS OF ORTHOGONAL MODULAR 4-FOLDS 3 Proof. Identical to Proposition 3.6 of [GHS10]. (Noting that the Beauville lattice of a deformation generalised Kummer variety and an irreducible symplectic manifold of K3[2(n+1)]-type differ by a factor of 2E8(−1).) (cid:3) Definition 3.2. A polarisation determined by a primitive vector h ∈ L is said to be split, or h is said to be split, if div(h) = 1. Corollary 3.3. If h ∈ L is split then, (1) the polarisation type of h is uniquely determined by the length h2; (2) the lattice h⊥ is isomorphic to L2(n+1),2d := 2U ⊕ h−2(n + 1)i ⊕ h−2di. Definition 3.4. Let F2d denote the modular variety F2d = O+(L, h)\Dh⊥ where h ∈ L is split and h2 = −2d. From now on, we assume all polarisations are split and fix 2(n + 1) = 6. Throughout, we shall use 4. The group O(L, h) hx to denote split h ∈ L of degree x. Proposition 4.1. If h = h2d and d > 2, then O(L, h) ∼= {g ∈ O(L6,2d) gv∗ = v∗ + L6,2d}, where v generates the h−2di factor of L6,2d and v∗ = (2d)−1v ∈ L∨. Moreover, if d = p2 for an odd prime p, then O(L, h) ≤ O(L6,2). Proof. (c.f. Proposition 3.12 of [GHS10].) By noting that O(L, h) acts on both hhi and hhi⊥, but trivially on hhi, we can immediately identify O(L, h) with a subgroup of O(L6,2d). There is an inclusion of abelian groups L/(hhi ⊕ hhi⊥) ⊂ hhi∨/hhi ⊕ (hhi⊥)∨/(hhi⊥) = D(hhi) ⊕ D(hhi⊥) defined by the series of overlattices hhi ⊕ h⊥ ⊂ L ⊂ L∨ ⊂ hh∨i ⊕ (h⊥)∨, and so the isotropic subgroup H = L/(hhi ⊕ h⊥) can be regarded as a subgroup of D(hhi) ⊕ D(hhi⊥). Let ph and ph⊥ denote the corresponding projections ph : H → D(hhi), ph⊥ : H → D(hhi⊥). By 1 = (2d)−1k1, Proposition 3.1, we can assume that h is given by h = e3 + df3 ∈ U ⊕ h−6i. Let k′ 3 = (2d)−1h, and k1 = e3 − df3, where k2 is a generator of the h−6i factor of L. Take 2 = 6−1k2, k′ k′ 3)i + hhi ⊕ h⊥, a basis {e1, f1, e2, f2, k′ ph⊥(H) = hk′ 1, d(k′ 2i. By applying Corollary 1.5.2 of [Nik79], 2} for (h⊥)∨. By direct calculation, H = hk′ 1i, and D(h⊥) = hk′ 3 − k′ 1 + k′ 1, k′ 1i ⊕ hk′ O(L, h) ∼= {g ∈ O(h⊥) gph⊥ (H) = id}, and the first part of the claim follows. For the second part of the claim, embed L6,2p2 ⊂ L6,2 by identifying factors of 2U ⊕ h−6i and mapping L6,2p2 ∋ t + ak1 7→ t + bu + apk ∈ L6,2, where t ∈ 2U ⊕ h−6i, k generates h−2i ⊂ L6,2, and a, b ∈ Z. Define the totally isotropic subspace M ⊂ D(L6,2p2) by M = L6,2/L6,2p2 ⊂ D(L6,2p2). By the above, if g ∈ O(L, h), then g(k′ 1 +L6,2p2. As M ⊂ hk′ 1i + L6,2p2 ⊂ D(L6,2p2) and g(L6,2p2) = L6,2p2, then g extends to a unique element of O(L6,2). (cid:3) 1) = k′ Let p be an odd prime. We use an idea in [Kon93] (who attributes it to O'Grady) to bound the index O(L6,2) : O(L6, h2p2). This involves considering the finite quadratic space Qp, defined by and a number of classical results on orthogonal groups of finite type (which can be found in [Die71], but are stated below for the convenience of the reader). Qp = L6,2/pL6,2 ⊂ L6,2p2/pL6,2, For i ∈ N, let Hi denote hyperbolic planes over Fp, and let Vθ denote the quadratic space hu, vi whose bilinear form is given by (u, u) = 1, (u, v) = 0, and (v, v) = θ where −θ /∈ (F∗ q)2. A non-degenerate quadratic space V over a finite field Fq of odd order q is uniquely determined by dim V and the discriminant ∆ = det B ∈ F∗ q/(F∗ q)2, where B is the bilinear form on V . 4 MATTHEW DAWES If dim V = 2m and ǫ = (−1)m∆ ∈ F∗ q/(F∗ q)2, then V is isomorphic to (V 2m ǫ = H1 ⊕ . . . ⊕ Hm V 2m ǫ = Vθ ⊕ H1 ⊕ H2 ⊕ . . . ⊕ Hm−1 if ǫ = 1 if ǫ = −1. If dim V = 2m + 1, there is a single isomorphism class for V given by V 2m+1 = H1 ⊕ . . . Hm ⊕ hθi for 0 6= θ ∈ Fq. We also need to know the order of O+(V ). As in [Die71], ( O+(V 2m+1) = (q2m − 1)q2m−1(q2m−2 − 1) . . . (q2 − 1)q ) = (q2m−1 − ǫqm−1)(q2m−2 − 1)q2m−3 . . . (q2 − 1)q. O+(V 2m ǫ Lemma 4.2. If u, v ∈ Qp and u2 = v2 ∈ F∗ O(L6,2). p/(F∗ p)2 for p > 3, then u and v are equivalent under Proof. Let {e1, f1, e2, f2, v1, v2} be a basis for L6,2 where v1, v2 are the respective generators of h−6i, h−2i and {ei, fi} are standard bases for the two copies of U . We define elements of O(L6,2). For e ∈ L6,2 isotropic and a ∈ e⊥ ⊂ L6,2 there are elements t(e, a) ∈ O(L) (known as Eichler transvections) defined by t(e, a) : v 7→ v − (a, v)e + (e, v)a − 1 2 (a, a)(e, v)e ( [Eic74] or §3 of [GHS09]). The action of t(e2, v1) and t(e2, v2) on w = (w1, w2, w3, w4, w5, w6) ∈ L6,2 are given by (t(e2, v1) : w 7→ (w1, w2, w3 + 3w4 + 6w5, w4, w5 + w4, w6) t(e2, v2) : w 7→ (w1, w2, w3 + w4 + 2w6, w4, w5, w6 + w4). One can also obtain elements of O(L6,2) by the trivial extension of elements in O(2U ) for an embedding 2U ⊂ L6,2. In particular, if (w, x, y, z) is taken on the standard basis {ei, fi} of U ⊕ U then the map (1) U ⊕ U ∋ (w, x, y, z) 7→(cid:18)w −y x (cid:19) , z identifies M2(Z) with U ⊕ U (where the inner product on M2(Z) is defined by det). An element (A, B) ∈ SL(2, Z) × SL(2, Z) defines an element in O(U ⊕ U ) by the mapping (A, B) :(cid:18)w −y x (cid:19) 7→ A(cid:18)w −y x (cid:19) B−1. z z Let x = (x1, x2, x3, x4, x5, x6) ∈ L6,2/pL6,2 be non-zero. We can assume that x4 6= 0 by (if required) applying t(e2, v1) or t(e2, v2), or permuting {x1, x2, x3, x4} by elements in O(2U ). By rescaling x so that x4 = 1, and by repeated application of t(e2, v1) and t(e2, v2), the element x can be transformed to an element of the form (x′ 4, 0, 0), and so can be identified with an element of 2U . Because of the existence of a Smith normal form for the associated matrix (1), x can be mapped to an element of the form (r, s, 0, 0, 0, 0) by using the image of SL(2, Z) × SL(2, Z) in O(2U ). One can assume x is of the form (1, a, 0, 0, 0, 0) by rescaling (if necessary). 3, x′ 2, x′ 1, x′ Now suppose u, v ∈ L6,2/pL6,2 are given by u = (1, a, 0, 0, 0, 0) and v = (1, b, 0, 0, 0, 0). If ab−1 ∈ (F∗ p)2, then there exists µ, λ ∈ Fp such that (µu)2 = (λv)2. Define u and v by u = µu = (u1, u2, 0, 0, 0, 0), v = λv = (v1, v2, 0, 0, 0, 0) and suppose (without loss of generality) that u − v = (r, s, 0, 0, 0, 0) is non- zero. Take representatives for r, s modulo p, and let r s gcd(r, s) if s = 0 if r = 0 otherwise. d = If r1, r2, s1, s2 are solutions to r2u1 + r1u2 = d and s2v1 + y2v2 = d taken modulo p, define the 1 ⊂ L6,2 by u′ = (r1, r2, 0, 0, 0, 0), v′ = (s1, s2, 0, 0, 0, 0), and w = elements u′, v′, w ∈ e⊥ (d−1r, d−1s, 0, 0, 0, 0) where r′ ≡ r mod p and s′ ≡ s mod p. Then, over Fp, (u, u′) = d, (v, v) = d, and t(e2, v′)t(f2, w)t(e2, u′) : u 7→ v. The result follows. (cid:3) 1 ∩ f ⊥ BOUNDARY COMBINATORICS OF ORTHOGONAL MODULAR 4-FOLDS 5 Theorem 4.3. Let p > 3 be prime. Then O+(L, h2p2 ) is of finite index in O+(L, h2) and O+(L, h2) : O+(L, h2p2 ) ≤ 16(p5 + p2). Proof. There is a natural homomorphism O(L6,2) → O(L6,2/pL6,2). If v, w ∈ Qp and v2 = w2 mod (F∗ p)2 then, by Lemma 4.2, v ∼ w under the action of O(L6,2) and so v ∼ w under the action of O(L6,2/pL6,2). The group O(L, h2p2) ⊂ O(L6,2) stabilises a hyperplane Π ⊂ Qp and, as StabO(L6,2/pL6,2)(Π) = O(L6,2p2/pL6,2), by the orbit-stabiliser theorem, O(L6,2) : StabO(L6,2)(Π) = O(L6,2/pL6,2) : O(L6,2p2/pL6,2) and O+(L6,2) : StabO+(L6,2)(Π) = O+(L6,2/pL6,2) : O+(L6,2p2/pL6,2). By Lemma 4.1, O(L, h2p2 ) ⊂ O(L6,2) and so, + (L6,2p2) ≤ O+(L, h2p2) ≤ StabO+(L6,2)(Π) ≤ O+(L6,2p2). eO As O(D(L6,2p2)) ∼= C ⊕3 2 , then (2) and so, StabO(L6,2)(Π) : O(L, h2p2 ) ≤ O(L6,2p2) : eO(L6,2p2) = 8, O+(L6,2) : O+(L, h2p2 ) ≤ 8 O+(L6,2/pL6,2) : O+(L6,2p2/pL6,2) ≤ 8 (p5 − ǫp2)(p4 − 1)p3(p2 − 1)p (p4 − 1)p3(p2 − 1)p ≤ 8(p5 + p2). (cid:3) 5. Toroidal compactifications 5.1. Overview. Toroidal compactifications are defined fully in [AMRT10]. We state only the results we require, following [GHS13]. Let F denote the orthogonal modular variety DM /Γ, where M is a lattice of signature (2, n) and (without loss of generality) Γ is a neat normal subgroup of O+(M ). There is a partial compactification D∗ M ) to which the action of Γ extends. The compactification D∗ M of DM (taken in the compact dual D∨ M admits a decomposition (3) D∗ M = DM ⊔GΠ FΠ ⊔Gℓ Fℓ, where the rational boundary components FΠ (boundary curves) and Fℓ (boundary points) are sym- metric spaces corresponding to Γ-equivalence classes of totally isotropic planes and isotropic lines in M ⊗ Q. For a rational boundary component F (given by some Fℓ or FΠ in (3)), let N (F ) ⊂ O(2, n) be the stabiliser of F , W (F ) be the unipotent radical of N (F ), and U (F ) be the centre of W (F ). Denote the intersections of N (F ), U (F ), and W (F ) with Γ by N (F )Z, U (F )Z, and W (F )Z. Let DM (F ) denote the domain U (F ).DM ⊂ D∨ M . Because of the Langlands decomposition for the parabolic subgroup N (F ), the domain DM (F ) decomposes as where V (F ) is the complex vector space W (F )/U (F ). DM (F ) = F × V (F ) × U (F )C, 6 MATTHEW DAWES If DM (F )′ is the quotient DM (F )′ = DM (F )/U (F )C, then the spaces DM (F ), DM (F )′, and F are related by the diagram π′ F pF DM (F ) πF F DM (F )′ where πF , pF , and π′ F are the natural projections onto F , F , and DM (F ), respectively. The space (4) π′ F : DM (F ) → DM (F )′ is a principal homogeneous space for U (F )C and admits an N (F )Z-action. By taking the quotient of (4) by U (F )Z ⊂ N (F )Z, one obtains the principal fibre bundle DM (F )/U (F )Z → DM (F )′ (5) whose fibre is equal to the algebraic torus T (F ) := U (F )C/U (F )Z. There is a real cone C(F ) ⊂ U (F ). By taking a fan Σ in the closure of C(F ) and then replacing the torus T (F ) in the fibre bundle (5) with the toric variety XΣ(F ) one obtains a new bundle over DM (F ) with fibre XΣ(F ). One constructs a partial compactification F(F )tor for F in a neighbourhood of F by taking the closure of DM /U (F )Z in the new bundle, and then taking the quotient by N (F )Z. A toroidal com- pactification for F is obtained by taking the set of partial compactifications over all rational boundary components F and gluing by identifying the copies of F contained in each one. 5.2. Invariants associated with F . Definition 5.1. If M is a lattice and E ⊂ M is a primitive totally isotropic sublattice, let HE := E⊥⊥/E ⊂ D(M ) where E⊥⊥ ⊂ M ∨. Lemma 5.2. Let E ⊂ L6,2p2 be a primitive totally isotropic sublattice of rank 2 corresponding to the boundary component F . Then there exists a Z-basis {v1, . . . , v6} of L6,2p2 so that {v1, v2} is a basis for E and {v1, . . . , v4} is a basis for E⊥. Furthermore, the basis can be chosen so that the Gram matrix (6) where Moreover, Q = ((vi, vj)) = A =(cid:18)a1 0 0 A 0 0 B C tA tC D  0 a1a2(cid:19) , (a1, a1a2) ∈ {(1, 1), (1, p), (1, 2p)}. a1, a2 are the elementary divisors of the group D(L6,2p2)/H ⊥ E , and B is the quadratic form on E⊥/E. Proof. As the lattices E and E⊥ are primitive in L6,2p2, the existence of a basis on which Q assumes the form of (6) is immediate. The Smith normal form of the matrix A embeds hv5, v6i in the dual hv∗ 6i. Therefore, the elementary divisors of A correspond to the elementary divisors of the abelian group hv∗ 6i/hv5, v6i (c.f. [GHS07]). 5, v∗ 5, v∗ If HE = E⊥⊥/E ⊂ D(L6,2p2), then H ⊥ E +L6,2p2 = hv∗ 1, . . . , v∗ 4i in D(L6,2p2), and so hv∗ 5, v∗ 6i/hv5, v6i ∼= D(L6,2p2)/H ⊥ E . As E is totally isotropic in L6,2p2, then HE is totally isotropic in D(L6,2p2). If D(L6,2p2) is identified with ((−1/6) ⊕ (−1/2p2), C6 ⊕ C2p2), then (x, y) ∈ D(L6,2p2) is isotropic if and only if As (3, p) = 1 then py, p2x2 + 3p2y2 p2x2 + 3y2 = 0 mod 12p2. 1 = 0 mod 12p2, and x2 + py2 1 = 0 mod 6. BOUNDARY COMBINATORICS OF ORTHOGONAL MODULAR 4-FOLDS 7 By considering squares modulo 6, we conclude that x = 0 or 3, and that x and y must have equal parities. Therefore, the isotropic elements of D(L6,2p2) are given by the set (x, y) ∈ {(0, 2kp), (3, (2k + 1)p) k ∈ Z}; the primitive isotropic subspaces of rank 1 in D(L6,2p2) are generated by x1 := (0, 2p) and x2 := (3, p); and the single primitive totally isotropic subspace of rank 2 is generated by hx1, x2i. If HE = hx1i then H ⊥ E = {(a, b) ∈ D(L6,2p2) 6b ≡ 0 mod 6p}, and so, pb and H ⊥ If HE = hx2i then E = h(1, 0), (0, p)i ∼= C6 ⊕ C2p. H ⊥ and so, pb, 2(a + b), and H ⊥ E = {(a, b) ∈ D(L6,2p2) pa + b ≡ 0 mod 2p}, E = h(1, p), (2, 0)i. If y1 = (1, p) and y2 = (2, 0), we also have the relations 6py1 = 0, 3y2 = 0, and p(2y1 − y2) = 0. As p ≡ ±1 mod 6, then 2py1 = ±y2 and so H ⊥ E = hy1i = h(1, p)i ∼= C3 ⊕ C2p. E = hy1i = h(1, p)i ∼= C3 ⊕ C2p. If HE = hx1, x2i, then H ⊥ We conclude that, (1) if HE = {0}, then H ⊥ (2) if HE = hx1i, then H ⊥ (3) if HE = hx2i, then H ⊥ (4) if HE = hx1, x2i, then H ⊥ E = D(L6,2p2) and D(L6,2p2)/H ⊥ E E = h(1, 0), (0, p)i ∼= C6 ⊕ C2p and D(L6,2p2)/H ⊥ E = h(1, p)i ∼= C3 ⊕ C2p and D(L6,2p2)/H ⊥ ∼= Cp; ∼= C2 ⊕ Cp; ∼= {0}; E E E = h(1, p)i ∼= C3 ⊕ C2p and D(L6,2p2)/H ⊥ E ∼= C2 ⊕ Cp, and the result follows. (cid:3) Lemma 5.3. There exists a basis {v1, . . . , v6} for L6,2p2 ⊗ Q so that {v1, v2} is a Z-basis for E, {v1, . . . , v4} is a Z-basis for E⊥, and Q = ((vi, vj)) = 0 0 A 0 B 0 A 0 0  where A and B are as in Lemma 5.2. Proof. (Essentially as in Lemma 2.24 of [GHS07].) Suppose C and D are as in Lemma 5.2. R := −B−1C ∈ det B−1M2(Z) and R′ ∈ det B−1M2(Z) satisfies If D − tCB−1C + tR′A + tAR′ = 0, then the required base change is given by the matrix (7) M = 0 R′ I 0 I R I 0 0  . (cid:3) 5.3. Counting boundary components. We determine the O(L6,2)-orbits of primitive totally isotropic rank 2 sublattices of L6,2 along the lines of [Sca87]. As in [Sca87], the O(L6,2)-orbits of primitive isotropic vectors in L6,2 can be determined in a straightforward manner by using the Eichler criterion (§10 [Eic74]), and so we omit the calculation here. Lemma 5.4. (Lemma 4.1 of [Bri83]) If M is a non-degenerate even lattice, and E ⊂ M is a primitive totally isotropic sublattice, then the discriminant form of the lattice E⊥/E is isomorphic to H ⊥ E /HE ⊂ D(M ). Lemma 5.5. If the rank 2 sublattice E ⊂ L6,2 is primitive and totally isotropic, then E⊥/E ∼= h−6i ⊕ h−2i or E⊥/E ∼= A2(−1). Proof. The lattice E⊥/E is negative definite and, by Lemma 5.4, D(E⊥/E) ∼= H ⊥ Identify D(L6,2) with C6⊕C2. If (a, b) ∈ D(L6,2) is isotropic, then a2/6+b2/2 = 0 mod 2Z and so, (a, b) = (0, 0) E /HE = D(L6,2) with discriminant form ((−1/6) ⊕ or (a, b) = (3, 1). ∼= h(2, 0)i with discriminant form (1/2), C6 ⊕ C2). If HE = h(3, 1)i, then H ⊥ ((1/3), C3). By tables in [CS99], the two negative definite even lattices of determinant 12 are If HE = {(0, 0)}, then H ⊥ E = h(1, 1)i and H ⊥ E /HE. E /HE h−6i ⊕ h−2i 8 and (8) MATTHEW DAWES (cid:18)−4 −2 −2 −4(cid:19) . The discriminant form of (8) is inequivalent to ((1/2)⊕2⊕(−1/3), C ⊕2 then E⊥/E ∼= h−6i ⊕ h−2i; if HE = h(3, 1)i then, from tables in [CS99], E⊥/E ∼= A2(−1). 2 ⊕C3). Therefore, if HE = h(0, 0)i (cid:3) Lemma 5.6. If E ⊂ L6,2 is a primitive totally isotropic sublattice of rank 2, then there exists a Z-basis {v1, . . . , v6} of L6,2 so that {v1, v2} is a basis for E, {v1, . . . , v4} is a basis for E⊥ ⊂ L6,2, and the Gram matrix (9) Moreover, (1) if HE = h(1, 1)i, then B = h−6i ⊕ h−2i, C = D = 0, and 0 0 P 0 B C P tC D  . Q = ((vi, vj)) = P =(cid:18)0 1 1 0(cid:19) ; c 0(cid:19), D =(cid:18)2d 0 0 C =(cid:18)0 0 3 0(cid:19), 0(cid:19), P =(cid:18)0 1 (2) if HE = h(3, 1)i, then B = A2(−1), for c ∈ {0, 1, 2}, d ∈ {0, 1, 2}. Proof. As in Lemma 5.2, take a basis {v1, . . . , v6} for L6,2 so that {v1, v2} is a basis for E, and {v1, . . . , v4} is a basis for E⊥. Suppose that By Lemma 5.5, HE = h(0, 0)i or HE = h(3, 1)i. If HE = h(0, 0)i then, because of the existence of a Smith normal form, there exist integral matrices U and Z so that By Lemma 5.5, there exists X ∈ GL(2, Z) so that tXB0X = B = h−6i ⊕ h−2i. Therefore, the matrix g1 := diag(tU, X, Z) transforms Q to Q′ where 0 0 A0 0 B0 C0 tC0 D0 A0  . 0 A 0 0 B C1 tA tC1 D1  , Q = ((vi, vj)) = U A0Z =(cid:18)0 1 1 0(cid:19) . Q′ = tg1Qg1 = A =(cid:18)0 1 1 0(cid:19) . g2 = Q′′ = g3 = 0 0 A 0 B 0 tA 0 D2 I −tAtC1 0 0 0 0 I I 0 I 0 0 0 W 0 I   .  , I 0 and The integral matrix g2 defined by transforms Q′ to Q′′, where If g3 is an integral matrix of the form BOUNDARY COMBINATORICS OF ORTHOGONAL MODULAR 4-FOLDS 9 then g3 sends D2 7→ D2 + tW A + tAW . One checks that the set {tW A + tAW W ∈ M2(Z)} contains all matrices of the form for a, b, c ∈ Z. Therefore, there exists integral W and d′ d′ 22 ≡ d22 mod 2, and 11, d′ 22 ∈ {0, 1} so that d′ 11 ≡ d11 mod 2, b b (cid:18)2a 2c(cid:19) g3 : D2 7→(cid:18)d′ 11 0 0 d′ 22(cid:19) . As the form Q is even, both d11 and d22 are even, and so there exists W so that g3 sends D2 to 0. Therefore, the matrix g3g2g1 gives the base change required in the statement of the theorem. If HE = h(3, 1)i then, because of the Smith normal form, there exist U, Z ∈ GL(2, Z) such that Moreover, there exists X ∈ GL(2, Z) such that tXB0X = B = A2(−1). Therefore the matrix g4 given by g4 = diag(tU, X, Z) ∈ GL(2, Z) transforms Q to Q′ where  A 0 0 0 B C1 tA tC1 D1 U A0Z =(cid:18)0 1 3 0(cid:19) . Q′ = tg1Qg1 = A =(cid:18)0 1 3 0(cid:19) . g5 =  , tSA + BT + C1 =(cid:18)0 0 a 0(cid:19) tSA + T Q + C1 =(cid:18)3s21 − 2s11 − t21 + c11 s12 − t12 − 2t22 + c22(cid:19) , I S 0 0 I T I 0 3s22 − 2t21 − t11 + c21 s11 − 2t12 − t22 + c12 0 for some S, T ∈ M2(Z). We claim that (sij) := S and (tij) := T can be chosen so that where a is determined modulo 3. As and (10) Let g5 be an integral matrix of the form then the claim about the second column is immediate, as s11 and t12 are both free. Let δ := 2t11 + t21 and s21 = 0. By taking t11 and t21 so that δ = −c11, the first column of (10) can be mapped to t(0, 3s22 − c11 + c21). Therefore, with an appropriate choice of s22, the matrix g5 transforms Q′ to where C0 is as in the statement of the theorem. We next put D2 in the correct form. If g6 is an integral matrix of the form then g6 sends One checks that the set {tW A + tAW W ∈ M2(Z)} contains all matrices of the form for a, b, c ∈ Z. Therefore if (dij) := D, there exists W so that Q′′ = g6 = 0 A 0 B C0 0 tA tC0 D2 I 0 I 0 0 0 W 0 I  ,  , D2 7→ D2 + tW A + tAW. b b (cid:18)6a 2c(cid:19) g3 : D2 7→(cid:18)d′ 11 0 0 d′ 22(cid:19) , 10 MATTHEW DAWES 11 ∈ {0, . . . , 5}, d′ 22 ∈ {0, 1}, and where d′ for d′ is even, then both d11 and d22 are even. Therefore, there exists W so that d′ Therefore, the matrix g6g5g4 gives the base change required in the statement of the theorem. Theorem 5.7. For prime p > 3, the modular variety F2p2 has at most 160(p5 + p2) boundary curves. 22 ≡ d22 mod 2. As the form Q 22 = 0. (cid:3) 11 ≡ d11 mod 6 and d′ 11 is even and d′ Proof. If E1, E2 ⊂ L6,2 are totally isotropic primitive sublattices of rank 2 with the same normal form (9) then, by Lemma 5.6, there exists g ∈ O(L6,2) so that g(E1) = E2. Therefore, by counting normal forms, there are at most 20 totally isotropic primitive rank 2 sublattices of L6,2 up to O+(L6,2)- equivalence. By Theorem 4.3, O+(L6,2) : O+(L, h2p2 ) ≤ 8(p5 + p2), and so, up to O+(L, h2p2 )-equivalence, there are at most 160(p5 + p2) boundary curves. (cid:3) 5.4. Counting singularities. In this section we count singularities in the boundary of a toroidal compactification of F2p2. We shall only consider the compactification in a neighbourhood of a boundary curve. The structure of the boundary in a neighbourhood of a boundary point is different, and presents additional toric considerations. Throughout, we assume the boundary component F corresponds to a rank 2 primitive totally isotropic sublattice E ⊂ L6,2p2 and define N = a1a2 det B, where a1, a2, B are as in Lemma 5.2. Lemma 5.8. On the basis given in Lemma 5.3, the groups N (F ), W (F ) and U (F ) are given by tU AZ = A, tXBX = B, tV AZ = 0, tXBY + tV AZ = 0, tY BY + tZAW + tW AZ = 0, det(U ) > 0, U V W 0 X Y 0 0 Z I V W Y 0 0 I I 0 0 (cid:18) 0 −x   BY + tV A = 0, tY BY + AW + tW A = 0 0 (cid:19) a1a2x . ,  0 I 0 0 0 I  x ∈ R N (F ) =  W (F ) =  U (F ) =  I Furthermore, if g ∈ N (F ), then g ∈ N (F ) ∩ O(L6,2p2) if and only if U V −V B−1C + W + U R′ − R′Z 0 X 0 0 Y − XB−1C + B−1CZ Z M −1gM =  ∈ GL(6, Z) where M is as in (7). Proof. The first part follows from direct calculation (as in Section 2.12 of [Kon93]). The second part follows as in Proposition 2.27 of [GHS07]. (cid:3) As in [Kon93], DL6,2p2 (F ) can be identified with a Siegel domain of the third kind inside C×C2×H+. The identification proceeds by taking homogeneous coordinates [t1 : . . . : t6] for P(L6,2p2 ⊗ C) and mapping DL6,2p2 (F ) → P(L6,2p2 ⊗ C) by setting t6 = 1 and (11) Proposition 5.9. If  t1 7→ z ∈ C t3 7→ w1 ∈ C t4 7→ w2 ∈ C t5 7→ τ ∈ H+ t2 7→ −2a1zτ −t(w1,w2)B(w1,w2) 2a1a2 . g = U V W 0 X Y 0 0 Z  ∈ N (F ) BOUNDARY COMBINATORICS OF ORTHOGONAL MODULAR 4-FOLDS 11 is as in Lemma 5.8, where Z =(cid:18)a b c d(cid:19), then the action of g on DL(F ) is given by twBw + V1w + W11τ + W12(cid:17) c det Z + (cτ + d)−1(cid:16) z 7→ z w 7→ (cτ + d)−1 (Xw + Y ( τ τ 7→ aτ +b cτ +d . 2a1 det Z 1 ))  Proof. As in [GHS07]. Lemma 5.10. Let (12) (cid:3) where ΓN ⊂ SL(2, Z) is the principal congruence subgroup of level N . Then, on the basis given in Lemma 5.3, the map where (13) If then defines an embedding of ΓN in N (F ) ∩ O(L6,2p2). Proof. Suppose A, B, C, and R′ are as in Lemma 5.2. Let P := Z ′R′ − R′Z, Q := −B−1C + B−1CZ, and let Z, Z ′ be given by (12), (13), respectively. As Z ′ = t(AZ −1A−1) then, by Lemma 5.8, gZ ∈ N (F ) ∩ O(L6,2p2) if and only if Z ′, P , Q ∈ M2(Z). As and Z ∈ ΓN , then Z ′ ∈ M2(Z). −ca2 −b/a2 Z ′ =(cid:18) d a (cid:19) R′ =(cid:18)w x z(cid:19) , −by + az − dz − bx/a2(cid:19) . −a2cz − bw y P =(cid:18)−a2cy − aw + dw − cx −cz − bw/a2 As Z ∈ ΓN , then P ∈ M2(Z). As Z ≡ I mod N , then C − CZ ≡ 0 mod N . As det BN and R′, B−1 ∈ det B−1M2(Z), then and the result follows. (cid:3) Q = B−1(CZ − C) ∈ M2(Z), Lemma 5.11. If Y ∈ M2(N Z) then, on the basis given in Lemma 5.3, there exists gY ∈ W (F ) ∩ O(L6,2p2) of the form Proof. Fix Y ∈ M2(N Z). Let A, B, and C be as in Lemma 5.2 and M be as in Lemma 5.3. By Lemma 5.8, c d(cid:19) ∈ ΓN , Z =(cid:18)a b Z 7→ gZ = Z ′ =(cid:18) d −b/a2 Z ′ 0 0 I 0 0 0 Z 0  , a (cid:19) , −ca2 gY = gY = ∗ I ∗ 0 I Y I 0 0 I V W Y 0 0 I I 0  .  belongs to W (F ) if and only if both (14) and (15) BY + tV A = 0 tY BY + AW + tW A = 0 12 MATTHEW DAWES are satisfied. If gY ∈ W (F ) then, by Lemma 5.8, gY ∈ W (F ) ∩ O(L6,2p2) if and only if V , Y , and P are integral matrices, where P := W − V B−1C. Equation (14) has a solution in V if Y ∈ M2(a1a2Z), and the matrix P is integral if V ∈ M2((det B)Z). Therefore, by (14), both conditions are satisfied if Y ∈ M2(N Z). If W =(cid:18)w11 w12 w21 w22(cid:19) , then (15) is equivalent to −tY BY = AW + tW A =(cid:18) 2a1w11 a1w12 + a1a2w21 a1w12 + a1a2w21 2a1a2w22 (cid:19) , and so has a solution in W if Y ∈ M2(N Z). (cid:3) Suppose (z, w, τ ) are coordinates for DL6,2p2 (F ), as in (11). Local coordinates for a cover D2p2(F )tor of a toroidal compactification F tor 2p2 in a neighbourhood of F are obtained from (z, w, τ ) by replacing z with u = expa1(z) = e2πi/a1 , and allowing u = 0. The group acting on the cover is given by G(F ) := (N (F ) ∩ O+(L, h))/(U (F ) ∩ O+(L, h)). Theorem 5.12. If g′ ∈ G(F ) fixes a point x = (0, w, τ ) ∈ D2p2(F )tor then, in local coordinates around x, g′ acts by (16) diag(ω0, ξω1, ξω2, ξ2), where ω1, ω2, and ξ are 4-th or 6-th roots of unity, and ω0 is a 12-th root of unity. For given (ω1, ω2, ξ), bounds for the number of connected components of D2p2(F )tor fixed by some g ∈ G(F ) acting as in (16) are given in Table 1 and Table 2, where JN,B = N det B and ω = e2πi/3. Assuming g′ acts non-trivially, the image of x in F tor 2p2 is singular. For invariants (a1, a2) corresponding to F , the values of N , det B, and KN are given in Table 3. χX Φ2 1 Φ1Φ2 Φ2 2 Φ3 Φ4 Φ6 28KN J 4 28KN J 4 28KN J 4 24KN J 4 24KN J 2 24KN J 4 28KN J 4 N,B N,B 25JN,B 28KN J 4 28KN J 4 N,B 24KN J 4 24KN J 2 24KN J 4 - 1 1 1 N,B N,B N,B N,B N,B N,B N,B N,B 25JN,B N,B ξ −i i −1 1 1 - 1 1 1 1 Table 1. ξ ω ω2 N,B 24.34KN J 4 24KN J 4 24KN J 4 24KN J 2 24KN J 4 28KN J 4 N,B 24.34KN J 4 24KN J 4 N,B N,B 24KN J 4 24KN J 2 24KN J 4 28KN J 4 N,B N,B N,B N,B N,B N,B N,B N,B Table 2. N,B −ω2 24KN J 4 24KN J 4 N,B 24.34KN J 4 28KN J 4 24KN J 4 24KN J 2 N,B N,B N,B N,B χX Φ2 1 Φ1Φ2 Φ2 2 Φ3 Φ4 Φ6 N,B −ω 24KN J 4 24KN J 4 N,B 24.34KN J 4 28KN J 4 24KN J 4 24KN J 2 N,B N,B N,B N,B BOUNDARY COMBINATORICS OF ORTHOGONAL MODULAR 4-FOLDS 13 (a1, a2) N det B KN 12p2 (1, 1) 12p (1, p) (1, 2p) 6p 12p2 12 3 Table 3. 1 p 2p Proof. Throughout, we assume g ∈ N (F ) represents a finite order element of G(F ), and that g fixes the point x = (0, w, τ ) ∈ D(F )tor. By Corollary 2.29 of [GHS07], no element of G(F ) acts as a quasi- reflection. Therefore, if g acts non-trivially, by a theorem of Chevalley [Che55], each point in the fixed locus of G(F ) on D2p2(F )tor is singular. As in Lemma 5.8, let where  U V W 0 X Y 0 Z 0 g = Z =(cid:18)a b c d(cid:19) , and let ξ = (cτ + d)−1, T = (I − ξX). As g ∈ G(F ) is of finite order, then U , X, Z are of finite order. By considering rational representations, o(U ), o(X), o(Z) ∈ {1, 2, 3, 4, 6} and so the basis of Lemma 5.5 can be chosen so that each of U , X, Z can be represented by one of ±(cid:18)1 0 0 1(cid:19), (cid:18)−1 0 1(cid:19), (cid:18) 0 −1 −1(cid:19), (cid:18)0 −1 0 (cid:19), (cid:18)0 −1 1 (cid:19). 0 1 1 1 Indeed, as Z ∈ M2(Z) (Lemma 5.8) and Z acts on H+, then Z ∈ SL(2, Z) (as noted in [GHS07]). Suppose o(Z) ∈ {3, 4, 6}. By standard results on the elliptic elements of SL(2, Z) [DS05], τ is SL(2, Z)-equivalent to i, ω or ω2. If Gk is the Eisenstein series of weight k, then G4(i) 6= 0, and G6(ω) 6= 0 (p. 10 [DS05]), and so ξ is a 4-th root of unity if Z is of order 4, and a 6-th root of unity if Z is of order 3 or 6. If cτ + d = ±1 then, as both {1, ω} and {1, i} are linearly independent over Q, Z =(cid:18)±1 0 ±1(cid:19) b which implies the contradiction o(Z) = 1 or 2. Therefore, h1, τ i =(h1, ii h1, ωi if o(Z) = 4 if o(Z) = 3 or 6. The values of det T against χX and ξ are given in Table 4. By (11), Φ2 1 Φ1Φ2 Φ2 2 Φ3 Φ4 Φ6 i −2i 2 2i i 0 −i χX ξ −ω ω ω ω2 −ω2 ω2 −3ω −3ω2 −1 −i 4 2i 0 2 0 −2i 4 −3ω 1 −i 3 −2ω −ω 2 0 2 3 i 1 0 Table 4. 1 0 0 1 − ω2 1 − ω2 1 − ω 1 − ω −3ω2 −2ω −ω2 −ω2 ω 0 −ω −2ω −2ω2 ω2 0 0 and so if det T 6= 0, by Table 4, T w = Y ( τ 1 ) ⊂ h1, τ i det B × h1, τ i det B , w ∈ h1, τ i K det B × h1, τ i K det B , for appropriate K ∈ {1, 2, 3}. Therefore, by using elements of the form gY defined in Lemma 5.11, w can be reduced to one of (KN det B)4 points modulo N (F ) ∩ O(L6,2p2). 14 MATTHEW DAWES If o(Z) = 1 or 2, then τ ∈ H+ is free and ξ = ±1. By Lemma 5.8, V = 0, Y = 0, and so T w = 0. If det T 6= 0 then w = 0. We consider each of the cases det T = 0 separately. If (χX , ξ) = (Φ2 1, 1) or (χX, ξ) = (Φ2 2, −1), then g acts as the identity (as it cannot act as a quasi-reflection); if (χX, ξ) = (Φ1Φ2, 1), then o(Z) = 1, τ is free, w1 ∈ (2 det B)−1Z, and w can be reduced to one of 2N det B lines by using elements of the form gY ∈ N (F ) ∩ O(L6,2p2). (The case (Φ1Φ2, −1) proceeds as for (Φ1Φ2, 1).) The remaining cases proceed as for (χX, ξ) = (Φ4, −i). If (χX , ξ) = (Φ4, −i), then w1 − iw2 ∈ (det B)−1h1, ii and w can be reduced to one of (N det B)2 lines using elements of the form gY ∈ N (F ) ∩ O(L6,2p2). By standard results on congruence subgroups (p.13 [DS05]), the index KN := SL(2, Z) : ΓN is given by KN = N 3YpN(cid:18)1 − 1 p2(cid:19) . Therefore, if o(Z) = 3, 4, 6, then τ can be reduced to one of KN cases modulo N (F ) ∩ O(L6,2p2) by using elements of the form gZ ∈ ΓN ⊂ N (F )Z defined in Lemma 5.10. (L6,2p2) ⊂ O+(L, h) and one obtains a final bound from (2) by noting that By Proposition 4.1, eO O+(L, h) : O(L6,2p2) ≤ 24. + The statement about the action of g in local coordinates follows as in [Kon93], and the values in Table 3 can be calculated from Lemma 5.2. The statement about ω0 follows by noting that o(ω0) divides lcm(o(U ), o(X), o(Z)). (cid:3) Remark 5.13. If g ∈ G(F ) acts as diag(ω0, ξω1, ξω2ξ2) in a neighbourhood of (0, w, τ ) and ω0 6= 1, then (0, w, τ ) is not contained in the closure of the singular locus of F2p2. The singular locus of F2p2 can be studied by different methods, as in [Daw17]. 6. Acknowledgements This paper originates from my PhD thesis. I thank Professor G. K. Sankaran for his supervision, and the University of Bath for financial support in the form of a research studentship. I also thank the Riemann Center for Geometry and Physics for providing excellent working conditions as I edited the paper, and for financial support in the form of a Riemann fellowship. References [AMRT10] A. Ash, D. Mumford, M. Rapoport, and Y-S. Tai. Smooth compactifications of locally symmetric varieties. [Bea83] [Bri83] [Che55] [CS99] Cambridge Mathematical Library. Cambridge University Press, Cambridge, second edition, 2010. A. Beauville. Vari´et´es Kahleriennes dont la premi`ere classe de Chern est nulle. J. Differential Geom., 18(4):755–782, 1983. E. Brieskorn. Die Milnorgitter der exzeptionellen unimodularen Singularitaten. Bonn. Math. Schr. 150, 1983. C. Chevalley. Invariants of finite groups generated by reflections. Amer. J. Math., 77:778–782, 1955. J. H. Conway and N. J. A. Sloane. Sphere packings, lattices and groups, volume 290 of Grundlehren der Mathematischen Wissenschaften. Springer-Verlag, New York, 1999. [Daw17] M. Dawes. On the Kodaira dimension of the moduli of deformation generalised Kummer varieties. arXiv [Die71] [DS05] [Eic74] preprint arXiv:1710.01672, 2017. J. A. Dieudonn´e. La g´eom´etrie des groupes classiques. Springer-Verlag, Berlin-New York, 1971. Troisi`eme ´edition, Ergebnisse der Mathematik und ihrer Grenzgebiete, Band 5. F. Diamond and J. Shurman. A first course in modular forms, volume 228 of Graduate Texts in Mathematics. Springer-Verlag, New York, 2005. M. Eichler. Quadratische Formen und orthogonale Gruppen. Springer-Verlag, Berlin-New York, 1974. Zweite Auflage, Die Grundlehren der mathematischen Wissenschaften, Band 63. [GHS07] V. Gritsenko, K. Hulek, and G. K. Sankaran. The Kodaira dimension of the moduli of K3 surfaces. Invent. Math., 169(3):519–567, 2007. [GHS09] V. Gritsenko, K. Hulek, and G. K. Sankaran. Abelianisation of orthogonal groups and the fundamental group of modular varieties. J. Algebra, 322(2):463–478, 2009. [GHS10] V. Gritsenko, K. Hulek, and G. K. Sankaran. Moduli spaces of irreducible symplectic manifolds. Compos. Math., 146(2):404–434, 2010. [GHS13] V. Gritsenko, K. Hulek, and G. K. Sankaran. Moduli of K3 surfaces and irreducible symplectic manifolds. In Handbook of moduli. Vol. I, volume 24 of Adv. Lect. Math. (ALM), pages 459–526. Int. Press, Somerville, MA, 2013. [HKW91] K. Hulek, C. Kahn, and S. H. Weintraub. Singularities of the moduli spaces of certain abelian surfaces. Compos. Math., 79(2):231–253, 1991. BOUNDARY COMBINATORICS OF ORTHOGONAL MODULAR 4-FOLDS 15 [HKW93] K. Hulek, C. Kahn, and S. H. Weintraub. Moduli spaces of abelian surfaces: compactification, degenerations, [Kon93] [Nik79] [Rap08] [Sca87] and theta functions. Berlin: Walter de Gruyter, 1993. S. Kond¯o. On the Kodaira dimension of the moduli space of K3 surfaces. Compos. Math., 89(3):251–299, 1993. V. V. Nikulin. Integer symmetric bilinear forms and some of their geometric applications. Izv. Akad. Nauk SSSR Ser. Mat., 43(1):111–177, 238, 1979. A. Rapagnetta. On the Beauville form of the known irreducible symplectic varieties. Math. Ann., 340(1):77–95, 2008. F. Scattone. On the compactification of moduli spaces for algebraic K3 surfaces. Mem. Amer. Math. Soc., 70(374), 1987. Riemann Center for Geometry and Physics Leibniz Universitat Hannover Appelstrasse 2 30167 Hannover Deutschland [email protected]
1502.03936	2	1502	2017-10-25T14:56:21	Rational curves on compact K\"ahler manifolds	[ "math.AG", "math.CV" ]	We present an inductive strategy to show the existence of rational curves on compact Kaehler manifolds which are not minimal models but have a pseudoeffective canonical bundle. The tool for this inductive strategy is a weak subadjunction formula for lc centers associated to certain big cohomology classes. This subadjunction formula is based, as in the projective case, on positivity arguments for relative adjoint classes.	math.AG	math	RATIONAL CURVES ON COMPACT KÄHLER MANIFOLDS JUNYAN CAO AND ANDREAS HÖRING Abstract. Mori's theorem yields the existence of rational curves on projec- tive manifolds such that the canonical bundle is not nef. In this paper we study compact Kähler manifolds such that the canonical bundle is pseudoeffective, but not nef. We present an inductive argument for the existence of rational curves that uses neither deformation theory nor reduction to positive char- acteristic. The main tool for this inductive strategy is a weak subadjunction formula for lc centres associated to certain big cohomology classes. 7 1 0 2 t c O 5 2 ] . G A h t a m [ 2 v 6 3 9 3 0 . 2 0 5 1 : v i X r a 1. Introduction 1.A. Main results. Rational curves have played an important role in the classifi- cation theory of projective manifolds ever since Mori showed that they appear as a geometric obstruction to the nefness of the canonical bundle. 1.1. Theorem. [Mor79, Mor82] Let X be a complex projective manifold such that the canonical bundle KX is not nef. Then there exists a rational curve C ⊂ X such that KX · C < 0. This statement was recently generalised to compact Kähler manifolds of dimension three [HP16], but the proof makes crucial use of results on deformation theory of curves on threefolds which are not available in higher dimension. Mori's proof uses a reduction to positive characteristic in an essential way and thus does not adapt to the more general analytic setting. The aim of this paper is to develop a completely different, inductive approach to the existence of rational curves. Our starting point is the following 1.2. Conjecture. Let X be a compact Kähler manifold. Then the canonical class KX is pseudoeffective if and only if X is not uniruled (i.e. not covered by rational curves). This conjecture is shown for projective manifolds in [MM86, BDPP13] and it is also known in dimension three by a theorem of Brunella [Bru06] using his theory of rank one foliations. Our main result is as follows: 1.3. Theorem. Let X be a compact Kähler manifold of dimension n. Suppose that Conjecture 1.2 holds for all manifolds of dimension at most n − 1. If KX is pseudoeffective but not nef, there exists a KX-negative rational curve f : P1 → X. Date: October 25, 2017. 2000 Mathematics Subject Classification. 32J27, 14E30, 14J35, 14J40, 14M22, 32J25. Key words and phrases. MMP, rational curves, Kähler manifolds, relative adjoint classes, subadjunction. 1 Our statement is actually a bit more precise: the KX-negative rational curve has zero intersection with a cohomology class that is nef and big, so the class of the curve lies in an extremal face of the (generalised) Mori cone. Theorem 1.3 is thus a first step towards a cone and contraction theorem for Kähler manifolds of arbitrary dimension. In low dimension we can combine our theorem with Brunella's result: 1.4. Corollary. Let X be a compact Kähler manifold of dimension at most four. If KX is pseudoeffective but not nef, there exists a rational curve f : P1 → X such that KX · f (P1) < 0. 1.B. The strategy. The idea of the proof is quite natural and inspired by well- known results of the minimal model program: let X be a compact Kähler manifold such that KX is pseudoeffective but not nef. We choose a Kähler class ω such that α := KX + ω is nef and big but not Kähler. If we suppose that X is projective and ω is an R-divisor class we know by the base point free theorem [HM05, Thm.7.1] that there exists a morphism µ : X → X′ such that α = µ∗ω′ with ω′ an ample R-divisor class on X′. Since α is big the morphism µ is birational, and we denote by Z an irreducible component of its exceptional locus. A general fibre of Z → µ(Z) has positive dimension and is covered by rational curves, in particular Z is uniruled. More precisely, denote by k ∈ N the dimension of µ(Z). Since α = µ∗ω′ we have (αZ )k+1 = 0 and (αZ )k is represented by some multiple of F where F is an irreducible component of a general fibre of Z → µ(Z). Since F is an irreducible component of a µ-fibre the conormal sheaf is "semipositive", so we expect that F F KF ′ · π∗ωdim Z−k−1 ≤ π∗KXF · π∗ωdim Z−k−1 (1) where π : F ′ → F is a desingularisation of F . Since αF is trivial and KX = α − ω we see that the right hand side is negative, in particular KF ′ is not pseudoeffective. Thus we can apply [MM86, BDPP13] to F ′ and obtain that F is uniruled. Since F is general we obtain that Z is uniruled. The key idea of our approach is to prove a numerical analogue of (1) that does not assume the existence of the contraction. Indeed if X is Kähler we are far from knowing the existence of a contraction. However we can still consider the null-locus Null(α) = [RZ αdim Z Z Z. =0 It is easy to see that if a contraction theorem holds also in the Kähler setting, then the null-locus is exactly the exceptional locus of the bimeromorphic contraction. We will prove that at least one of the irreducible components Z ⊂ Null(α) is covered by α-trivial rational curves: let π : Z′ → Z be a desingularisation, and let k be the numerical dimension of π∗αZ (cf. Definition 2.5). We will prove that (2) ≤ π∗KXZ · π∗αk Z · π∗ωdim Z−k−1 KZ′ · π∗αk Z · π∗ωdim Z−k−1 Z Note that the right hand side is negative, so Conjecture 1.2 yields the existence of rational curves. Recall also that if the contraction µ exists, then π∗αk Z is a multiple of a general fibre, so this inequality is a refinement of (1). The inequality (2) follows from a more general weak subadjunction formula for maximal lc centres (cf. Definition 4.4) of the pair (X, cα) (for some real number c > 0) which we will Z . 2 explain in the next section. The idea of seeing the irreducible components of the null locus as an lc centre for a suitably chosen pair is already present in Takayama's uniruledness of stable base loci [Tak08], in our case a recent result of Collins and Tosatti [CT15, Thm.1.1] and the work of Boucksom [Bou04] yield this property without too much effort. While (2) and Conjecture 1.2 imply immediately that Z is uniruled it is a priori not clear if we can choose the rational curves to be KX-negative (or even α-trivial): for the simplicity of notation, let us suppose that Z is smooth. If Z was projective and αZ an R-divisor class we could argue as in [HP16, Prop.7.11] using Araujo's In the Kähler case we need a description of the mobile cone [Ara10, Thm.1.3]. new argument: let Z → Y be the MRC-fibration (cf. Remark 6.10) and let F be a general fibre. Arguing by contradiction we suppose that F is not covered by α-trivial rational curves. A positivity theorem for relative adjoint classes (Theorem 5.2) shows that KZ/Y + αZ is pseudoeffective if KF + αF is pseudoeffective. Since KY is pseudoeffective by Conjecture 1.2 this implies that KZ + αZ is pseudoeffective, a contradiction to (2). Thus we are left to show that KF + αF is pseudoeffective, at least up to replacing αF by λαF for some λ ≫ 0. Since αF is not a rational cohomology class this is a non-trivial property related to the Nakai-Moishezon criterion for R-divisors by Campana and Peternell [CP90]. Using the minimal model program for the projec- tive manifold F and Kawamata's bound on the length of extremal rays [Kaw91, Thm.1] we overcome this problem in Proposition 6.9. 1.C. Weak subadjunction. Let X be a complex projective manifold, and let ∆ be an effective Q-Cartier divisor on X such that the pair (X, ∆) is log-canonical. Then there is a finite number of log-canonical centres associated to (X, ∆) and if we choose Z ⊂ X an lc centre that is minimal with respect to the inclusion, the Kawamata subadjunction formula holds [Kaw98] [FG12, Thm1.2]: the centre Z is a normal variety and there exists a boundary divisor ∆Z such that (Z, ∆Z ) is klt and KZ + ∆Z ∼Q (KX + ∆)Z . If the centre Z is not minimal the geometry is more complicated, however we can still find an effective Q-divisor ∆ Z on the normalisation ν : Z → Z such that1 K Z + ∆ Z ∼Q ν∗(KX + ∆)Z . We prove a weak analogue of the subadjunction formula for cohomology classes: 1.5. Theorem. Let X be a compact Kähler manifold, and let α be a cohomology class on X that is a modified Kähler class (cf. Definition 4.1). Suppose that Z ⊂ X is a maximal lc centre of the pair (X, α), and let ν : Z → Z be the normalisation. Then we have K Z · ω1 · . . . · ωdim Z−1 ≤ ν∗(KX + α)Z · ω1 · . . . · ωdim Z−1, where ω1, . . . , ωdim Z−1 are arbitrary nef classes on Z. Our proof follows the strategy of Kawamata in [Kaw98]: given a log-resolution µ : X → X and an lc place E1 dominating Z we want to use a canonical bundle formula for the fibre space µE1 : E1 → Z to relate µ∗(KX + α)E1 and K Z. As in 1This statement is well-known to experts, cf. [BHN15, Lemma 3.1] for a proof. 3 [Kaw98] the main ingredient for a canonical bundle formula is the positivity theorem for relative adjoint classes Theorem 3.4 which, together with Theorem 5.2, is the main technical contribution of this paper. The main tool of the proofs of Theorem 3.4 and Theorem 5.2 is the positivity of the fibrewise Bergman kernel which is established in [BP08, BP10]. Since we work with lc centres that are not necessarily minimal the positivity result Theorem 3.4 has to be stated for pairs which might not be (sub-)klt. This makes the setup of the proof quite heavy, but similar to earlier arguments (cf. [BP10, Pău12b] and [FM00, Tak06] in the projective case). The following elementary example illustrates Theorem 1.5 and shows how it leads to Theorem 1.3: 1.6. Example. Let X′ be a smooth projective threefold, and let C ⊂ X′ be a smooth curve such that the normal bundle NC/X ′ is ample. Let µ : X → X′ be the blow-up of X′ along C and let Z be the exceptional divisor. Let D ⊂ X′ be a smooth ample divisor containing the curve C, and let D′ be the strict transform. By the adjunction formula we have KZ = (KX + Z)Z, in particular it is not true that KZ · ω1 ≤ KXZ · ω1 for every nef class ω1 on Z. Indeed this would imply that −ZZ is pseudoeffective, hence N∗C/X ′ is pseudoeffective in contradiction to the construction. However if we set α := µ∗c1(D), then α is nef and represented by µ∗D = D′ + Z. Then the pair (X, D′ + Z) is log-canonical and Z is a maximal lc centre. Moreover we have KZ · ω1 = (KX + Z)Z · ω1 ≤ (KX + D′ + Z)Z · ω1 = (KX + α)Z · ω1 since D′Z is an effective divisor. Now we set ω1 = αZ , then αZ · ω1 = α2 KX is anti-ample on the µ-fibres we have Z = 0 since it is a pull-back from C. Since Thus KZ is not pseudoeffective. KZ · αZ = KXZ · αZ < 0. 1.D. Relative adjoint classes. We now explain briefly the idea of the proof of Theorem 3.4 and Theorem 5.2. In view of the main results in [BP08] and [Pău12a], it is natural to ask the following question : 1.7. Question. Let X and Y be two compact Kähler manifolds of dimension m and n respectively, and let f : X → Y be a surjective map with connected fibres. Let F be the general fiber of f . Let αX be a Kähler class on X and let D be a klt Q-divisor on X such that c1(KF ) + [(αX + D)F ] is a pseudoeffective class. Is c1(KX/Y ) + [αX + D] pseudoeffective ? In the case c1(KF ) + [(αX + D)F ] is a Kähler class on F , [Pău12a, Gue16] confirm the above question by studying the variation of Kähler-Einstein metrics (based on [Sch12]). In our article, we confirm Question 1.7 in two special cases: Theorem 3.4 and Theorem 5.2 by using the positivity of the fibrewise Bergman kernel which is established in [BP08, BP10]. Let us compare our results to Păun's result [Pău12a, Thm.1.1] on relative adjoint classes: while we make much weaker assumptions on the geometry of pairs or the positivity of the involved cohomology classes we are always in a situation where locally over the base we only have to deal with R-divisor classes. Thus the transcendental character of the argument is only apparent on the base, not along the general fibres. 4 More precisely, in Theorem 3.4, we add an additional condition that c1(KX/Y + [αX +D]) is pull-back of a (1, 1)-class on Y (but we assume that D is sub-boundary). Then we can take a Stein cover (Ui) of Y such that (KX/Y + [αX + D])f −1(Ui) is trivial on f−1(Ui). Therefore [αX + D]f −1(Ui) is a R-line bundle on f−1(Ui). We assume for simplicity that D is klt (the sub-boundary case is more complicated). We can thus apply [BP10] to every pair (f−1(Ui), KX/Y + [αX + D]). Since the fibrewise Bergman kernel metrics are defined fiber by fiber, by using ∂∂-lemma, we can glue the metrics together and Theorem 3.4 is thus proved. In Theorem 5.2, we add the condition that F is simply connected and H 0(F, Ω2 F ) = 0 2. Then we can find a Zariski open set Y0 of Y such that Rif∗(OX ) = 0 on Y0 for every i = 1, 2. By using the same argument as in Theorem 3.4, we can construct √−1 a quasi-psh function ϕ on f−1(Y0) such that 2π Θ(KX/Y ) + αX + ddcϕ ≥ 0 on f−1(Y0). Now the main problem is to extend ϕ to be a quasi-psh function on X. Since c1(KF + αXF ) is not necessary a Kähler class on F , we cannot use directly In fact, thanks the method in [Pău12a, 3.3] . Here we use the idea in [Lae02]. to [Lae02, Part II, Thm 1.3], we can find an increasing sequence (km)m∈N and hermitian line bundles (Fm, hm)m∈N (not necessarily holomorphic) on X such that (3) Let Xy be the fiber over y ∈ Y0. As we assume that H 0(Xy, Ω2 ) = 0, FmXy can be equipped with a holomorphic structure JXy,m. Therefore we can define the Bergman kernel metric associated to (FmXy , JXy ,m, hm). Thanks to ∂∂-lemma, we can compare ϕXy and the Bergman kernel metric associated to (FmXy , JXy ,m, hm). Note that (3) implies that Fm is more and more holomorphic. Therefore, by using standard Ohsawa-Takegoshi technique [BP10], we can well estimate the Bergman kernel metric associated to FmXy when y → Y \ Y0. Theorem 5.2 is thus proved by combining these two facts. Acknowledgements. This work was partially supported by the A.N.R. project CLASS3. Θ(KX/Y ) + αX )kC∞(X) → 0. Θhm(Fm) − km( √−1 2π Xy √−1 2π k 2. Notation and terminology For general definitions we refer to [Har77, KK83, Dem12]. Manifolds and normal complex spaces will always be supposed to be irreducible. A fibration is a proper surjective map with connected fibres ϕ : X → Y between normal complex spaces. 2.1. Definition. Let X be a normal complex space, and let f : X → Y be a proper surjective morphism. A Q-divisor D is f -vertical if f (Supp D) ( Y . Given a Q-divisor D it admits a unique decomposition D = Df -hor + Df -vert such that Df -vert is f -vertical and every irreducible component E ⊂ Supp Df -hor surjects onto Y . 2.2. Definition. Let X be a complex manifold, and let F be a sheaf of rank one on X that is locally free in codimension one. The bidual F∗∗ is reflexive of rank one, so locally free, and we set c1(F ) := c1(F∗∗). 2If F is rational connected these two conditions are satisfied. 3ANR-10-JCJC-0111 5 Throughout this paper we will use positivity properties of real cohomology classes of type (1, 1), that is elements of the vector space H 1,1(X)∩H 2(X, R). The definitions can be adapted to the case of a normal compact Kähler space X by using Bott- Chern cohomology for (1, 1)-forms with local potentials [HP16]. In order to simplify the notation we will use the notation N 1(X) := H 1,1(X) ∩ H 2(X, R). Note that for the purpose of this paper we will only use cohomology classes that are pull-backs of nef classes on some smooth space, so it is sufficient to give the definitions in the smooth case. 2.3. Definition. [Dem12, Defn 6.16] Let (X, ωX ) be a compact Kähler manifold, and let α ∈ N 1(X). We say that α is nef if for every ǫ > 0, there is a smooth (1, 1)-form αǫ in the same class of α such that αǫ ≥ −ǫωX. We say that α is pseudoeffective if there exists a (1, 1)-current T ≥ 0 in the same class of α. We say that α is big if there exists a ǫ > 0 such that α − ǫωX is pseudoeffective. 2.4. Definition. Let X be a compact Kähler manifold, and let α ∈ N 1(X) be a nef and big cohomology class on X. The null-locus of α is defined as Null(α) = [RZ αdim Z Z Z. =0 Remark. A priori the null-locus is a countable union of proper subvarieties of X. However by [CT15, Thm.1.1] the null-locus coincides with the non-Kähler locus EnK(α), in particular it is an analytic subvariety of X. 2.5. Definition. [Dem12, Defn 6.20] Let X be a compact Kähler manifold, and let α ∈ N 1(X) be a nef class. We define the numerical dimension of α by 2.6. Remark. A nef class α is big if and only if RX αdim X > 0 [DP04, Thm.0.5] closed positive (nd(α), nd(α))-current T . Therefore RX αnd(α) ∧ ωdim X−nd(α) which is of course equivalent to nd(α) = dim X. By [Dem12, Prop 6.21] the cohomology class αnd(α) can be represented by a non-zero > 0 nd(α) := max{k ∈ N αk 6= 0 in H 2k(X, R)}. for any Kähler class ωX . 2.7. Definition. Let X be a normal compact complex space of dimension n, and let ω1, . . . , ωn−1 ∈ N 1(X) be cohomology classes. Let F be a reflexive rank one sheaf on X, and let π : X′ → X be a desingularisation. We define the intersection number c1(F ) · ω1 · . . . · ωn−1 by X c1((µ∗F )∗∗) · µ∗ω1 · . . . · µ∗ωn−1. Remark. The definition above does not depend on the choice of the resolution π: the sheaf F is reflexive of rank one, so locally free on the smooth locus of X. Thus µ∗F is locally free in the complement of the µ-exceptional locus. Thus π1 : X′1 → X and π2 : X′2 → X are two resolutions and Γ is a manifold dominating X′1 and X′2 via bimeromorphic morphisms q1 and q2, then q∗1 π∗1F and q∗2π∗2F coincide in the complement of the π1 ◦ q1 = π2 ◦ q2-exceptional locus. Thus their biduals coincide in the complement of this locus. By the projection formula their intersection with classes coming from X are the same. 6 3. Positivity of relative adjoint classes, part 1 Before the proof of the main theorem in this section, we first recall the construction of fibrewise Bergman kernel metric and its important property, which are estab- lished in the works [BP08, BP10]. The original version [BP10] concerns only the projective fibration. However, thanks to the optimal extension theorem [GZ15] and an Ohsawa-Takegoshi extension theorem for Kähler manifolds [Yi14, Cao14], we know that it is also true for the Kähler case : 3.1. Theorem. [BP10, Thm 0.1], [GZ15, 3.5], [Yi14, Thm 1.1][Cao14, Thm 1.2] Let p : X → Y be a proper fibration between Kähler manifolds of dimension m and n respectively, and let L be a line bundle endowed with a metric hL such that: 1) The curvature current of the bundle (L, hL) is semipositive in the sense of cur- rent, i.e., √−1ΘhL(L) ≥ 0; 2) there exists a general point z ∈ Y and a non zero section u ∈ H 0(Xz, mKXz + L) such that (4) ZXz u 2 m hL < +∞. Then the line bundle mKX/Y + L admits a metric with positive curvature current. Moreover, this metric is equal to the fibrewise m-Bergman kernel metric on the general fibre of p. 3.2. Remark. Here are some remarks about the above theorem. (1): Note first that as u ∈ H 0(Xz, mKXz + L), u Therefore the integral (4) is well defined. (2): The fibrewise m-Bergman kernel metric is defined as follows : Let x ∈ X be a point on a smooth fibre of p. We first define a hermitian metric h on −(mKX/Y +L)x by is a volume form on Xz. 2 m hL kξk2 h := sup τ (x) · ξ2 (RXp(x) τ 2 m hL , )m where ξ is a basis of −(mKX/Y + L)x and the 'sup' is taken over all sections τ ∈ H 0(Xp(x), mKX/Y +L). The fibrewise m-Bergman kernel metric on mKX/Y +L is defined to be the dual of h. It will be useful to give a more explicit expression of the Bergman kernel type metric. Let ωX and ωY be Kähler metrics on X and Y respectively. Then ωX and ωY induce a natural metric hX/Y on KX/Y . Let Y0 be a Zariski open set of Y such that p is smooth over Y0. Set h0 := hm X/Y · hL be the induced metric on mKX/Y + L. Let ϕ be a function on p−1(Y0) defined by ϕ(x) = sup τ∈A 1 m lnτh0 (x), where A := {f f ∈ H 0(Xp(x), mKX/Y + L) and ZXp(x) f 2 m h0 (ωm X /p∗ωn Y ) = 1}. We can easily check that the metric h0 · e−2mϕ on mKX/Y + L coincides with In particular, h0 · e−2mϕ the fibrewise m-Bergman kernel metric defined above. 7 is independent of the choice of the metrics ωX and ωY . Sometimes we call ϕ the fibrewise m-Bergman kernel metric. (3): Note that, by construction, if we replace hL by f ⋆c(y) · hL for some smooth strictly positive function c(y) on Y , the corresponding weight function ϕ in un- changed. For readers' convenience, we recall also the following version of the Ohsawa- Takgoshi extension theorem which will be used in the article. 3.3. Proposition.[BP10, Prop 0.2] Let p : X → ∆ be a fibration from a Kähler manifold to the unit disc ∆ ∈ Cn. and let L be a line bundle endowed with a possible singular metric hL such that √−1ΘhL(L) ≥ 0 in the sense of current. Let m ∈ N. We suppose that the center fiber X0 is smooth and let f ∈ H 0(X0, mKX0 + L) such that ZX0 f 2 m hL < +∞. Then there exists a F ∈ H 0(X, mKX/Y + L) such that (i) FX0 = f (ii) The following L 2 m bound holds ZX F 2 m hL ≤ C0ZX0 f 2 m hL . where C0 is an absolute constant as in the standard Ohsawa-Takegoshi the- orem. Moreover, thanks to [GZ15], we can take C0 as the volume of the unit disc ∆. Here is the main theorem in this section. 3.4. Theorem. Let X and Y be two compact Kähler manifolds of dimension m and n respectively, and let f : X → Y be a surjective map with connected fibres. kPj=2 −djDj be a Q-divisor on X such Let αX be a Kähler class on X. Let4 D = that the support has simple normal crossings. Suppose that the following properties hold: (a) If dj ≤ −1 then f (Dj) has codimension at least 2. (b) The direct image sheaf f∗OX (⌈−D⌉) has rank one. Moreover, if D = Dh + Dv is the decomposition in a f -horizontal part Dh (resp. f -vertical part Dv) then we have (f∗OX (⌈−Dv⌉))∗∗ ≃ OY . (c) c1(KX/Y + αX + D) = f∗β for some real class β ∈ H 1,1(Y, R). Let ω1, ω2,··· , ωdim Y −1 be nef classes on Y . Then we have (5) β · ω1 ··· ωdim Y −1 ≥ 0. Proof. Step 1: Preparation. We start by interpreting the conditions (a) and (b) in a more analytic language. We can write the divisor D as D = B − F v − F h, 4The somewhat awkward notation will be become clear in the proof of Theorem 1.5. 8 where B, F v, F h are effective Q-divisors and F v (resp. F h) is f -vertical (resp. f -horizontal). We also decompose F v as F v = F v 1 + F v 2 2 ) ≥ 2 and codimY f (E) = 1 for every irreducible component such that codimY f (F v E ⊂ F v 1 . Let Xy be a general f -fibre. Since dj > −1 for every Dj mapping onto Y (cf. condition (a)), the divisors ⌈−D⌉ and ⌈F h⌉ coincide over a non-empty Zariski open subset of Y . Thus the condition rank f∗OX (⌈−D⌉) = 1 implies that h0(Xy,⌈F h⌉Xy ) = 1. Therefore, for any meromorphic function ζ on Xy, we have ⇒ ζ is constant. div(ζ) ≥ −⌈F h⌉Xy (6) Since dj > −1 for every Dj mapping onto a divisor in Y (cf. condition (a)), the divisors ⌈−Dv⌉ and ⌈F v⌉ coincide over a Zariski open subset Y1 ⊂ Y such that codimY (Y \ Y1) ≥ 2. In particular the condition (f∗OX (⌈−Dv⌉))∗∗ ≃ OY implies that (f∗OX (⌈−Dv⌉))Y1 = OY1 . So for every meromorphic function ζ on any small Stein open subset of U ⊂ Y1, we have div(ζ ◦ f ) ≥ −⌈F v⌉f −1(U) (7) ζ is holomorphic. ⇒ Step 2: Stein cover. Select a Stein cover (Ui)i∈I of Y such that H 1,1(Ui, R) = 0 for every i. Let θ be a smooth closed (1, 1)-form in the same class of c1(KX/Y + αX + D + ⌈F v + F h⌉). Thanks to (c), we have c1(KX/Y + αX + D)f −1(Ui) ∈ f−1(H 1,1(Ui, R)) = 0. There exists thus a line bundle Li on f−1(Ui) such that KX/Y + Li ≃ ⌈F v + F h⌉ on f−1(Ui). Moreover, we can find a smooth hermitian metric hi on KX/Y + Li over f−1(Ui) such that √−1 2π Θhi(KX/Y + Li) = θ on f−1(Ui). (8) (9) Step 3: Local construction of metric. We construct in this step a canonical function ϕi on f−1(Ui) such that θ + ddcϕi ≥ ⌈F v 1 + F h⌉ over f−1(Ui) for every i. The function is in fact just the potential of the fibrewise Bergman kernel metric mentioned in Remark 3.2. A more explicit construction is as follows: Note first that c1(Li) = αX + D + ⌈F v + F h⌉, we can find a metric hLi on Li such that iΘhLi = αX + [D] + ⌈F v + F h⌉ = αX + [B] + (⌈F v + F h⌉ − [F v + F h]) ≥ 0 in the sense of current. Moreover, we can ask that hi/hLi is a global metric on KX/Y , i.e., hi/hLi = hj/hLj on f−1(Ui ∩ Uj). Thanks to the sub-klt condition (a) and the construction of the metric hLi, we can find a Zariski open subset Ui,0 of Ui such that for every y ∈ Ui,0, f is smooth over y and there exists a sy ∈ H 0(Xy, KX/Y + Li) such that (10) = 1. ZXy sy2 hLi 9 Recall that sy2 (11) hLi is a volume forme on Xy (cf. Remark 3.2). Using the fact that h0(Xy, KX/Y + Li) = h0(Xy,⌈F h⌉) = 1 for every y ∈ Ui,0, we know that sy is unique after multiplying by a unit norm complex number. There exists thus a unique function ϕi on f−1(Ui,0) such that its restriction on Xy equals to lnsyhi. We have the following key property. Claim: ϕi can be extended to be a quasi-psh function (we still denote it as ϕi) on f−1(Ui), and satisfies (9). The claim will be proved by using the methods in [BP08, Thm 0.1]. We postpone the proof of the claim later and first finish the proof of the theorem. The properties (6) and (7) will be used in the proof of the claim. Step 4: Gluing process, final conclusion. We first prove that (12) ϕi = ϕj on f−1(Ui ∩ Uj). Let y ∈ Ui,0∩ Uj,0. Since both (KX/Y + Li)Xy ≃ (KX/Y + Lj)Xy ≃ ⌈F v + F h⌉Xy , we have LiXy ≃ LjXy . Under this isomorphism, the curvature condition (8) and ∂∂-lemma imply that hLiXy = hLjXy · e−cy (13) where the constant cy depends on y ∈ Y . As hi/hLi is a metric on KX/Y indepen- dent of i, we have for some constant cy on Xy, (14) By (11), there exist unique elements sy,i ∈ H 0(Xy, KX/Y + Li) and sy,j ∈ H 0(Xy, KX/Y + Lj) (after multiply by a unit norm complex number) such that hiXy = hjXy · e−cy on Xy. ZXy sy,i2 hLi = 1 and ZXy sy,j2 hLj = 1. Thanks to (13), we have (after multiply by a unit norm complex number) Together with (14), we get sy,i = e cy 2 · sy,j. ϕiXy = lnsy,ihi = lnsy,jhj = ϕjXy . (15) Since (15) is proved for every y ∈ Ui,0 ∩ Uj,0, we have ϕi = ϕj on f−1(Ui,0 ∩ Uj,0). Combining this with the extension property of quasi-psh functions, (12) is thus proved. Thanks to (12), (ϕi)i∈I defines a global quasi-psh function on X which we denote by ϕ. By (9), we have θ + ddcϕ ≥ ⌈F v 1 + F h⌉ over f−1(Ui) for every i. Therefore θ + ddcϕ ≥ ⌈F v 1 + F h⌉ over X. Then c1(KX/Y + αX + D + ⌈F v codimY f∗(F v 2 ) ≥ 2, the theorem is proved. 10 2 ⌉) is pseudoeffective on X. Together with the fact (cid:3) The rest part of this section is devoted to the proof of the claim in Theorem 3.4. The main method is the Ohsawa-Takegoshi extension techniques used in [BP10]. Before the proof of the claim, we need the following lemma which interprets the property (7) in terms of a condition on the metric hi. 3.5. Lemma. Fix a Kähler metric ωX (resp. ωY ) on X (resp. Y ). Let sB (resp. sF v , sF h ) be the canonical section of the divisor B (resp. F v and F h). Let ψ be the function of the form ψ = lnsB − lnsF v − lnsF h + C∞, (16) where · is with respect to some smooth metric on the corresponding line bundle. Let Y1 be the open set defined in Step 1 of the proof of Theorem 3.4 and let Y0 ⊂ Y1 be a non-empty Zariski open set satisfying the following conditions : (a) f is smooth over Y0; (b) f (Dv) ⊂ Y \ Y0; (c) F hXy is snc for every y ∈ Y0; (d) The property (6) holds for every y ∈ Y0. ZXy Then for any open set ∆ ⋐ Y1 ∩ Ui (i.e., the closure of ∆ is in Y1 ∩ Ui), there exists some constant C(∆, Y1, Ui) > 0 depending only on ∆, Y1 and Ui, such that (17) e−2ψωm X /f∗ωn Y ≥ C(∆, Y1, Ui) for every y ∈ ∆ ∩ Y0, where m (resp. n) is the dimension of X (resp. Y ). 3.6. Remark. The meaning of (17) is that, for any sequence (yi)i≥1 converging to a point in Y1 \ Y0, the sequence (RXyi Proof. Fix an open set ∆1 such that ∆ ⋐ ∆1 ⋐ Y1 ∩ Ui. Let y0 be a point in ∆∩ Y0 and let cy0 be a constant such that Y )i≥1 will not tend to 0. X /f∗ωn e−2ψωm cy02ZXy0 e−2ψωm X /f∗ωn Y = 1. (18) and (19) Let s⌈F⌉ be the canonical section of ⌈F v + F h⌉. By applying Proposition 3.3 to (f−1(∆1), KX + Li, hLi) and the section cy0 ⊗ s⌈F⌉ ∈ H 0(Xy0, KX + Li), we can find a holomorphic section τ ∈ H 0(f−1(∆1), KX + Li) such that τXy0 hLi ≤ C1ZXy0 τ2 = cy0 ⊗ s⌈F⌉ = C1cy02ZXy0 hLi Zf −1(∆1) τ2 e−2ψωm X /f∗ωn Y = C1 where C1 is a constant independent of y0 ∈ ∆ ∩ Y0. Set eτ := τ it by eτ ) on f−1(∆1) and (19) implies that . Then eτ can be extended to a meromorphic function (we still denote s⌈F ⌉ (20) Zf −1(∆1) eτ2e−2ψ ≤ C1 Therefore (21) div(eτ ) ≥ −⌈F h⌉ − ⌈F v⌉ 11 on f−1(∆1). We now prove thateτ is in fact holomorphic on f−1(∆1). For every point y ∈ ∆1∩Y0, thanks to (b), F v ∩ Xy = ∅. Together with (21) and (c), we have div(eτXy ) ≥ −⌈F hXy⌉ on Xy for every y ∈ ∆1 ∩ Y0. Combining this with (d), eτXy is constant for every y ∈ ∆1 ∩ Y0. Therefore eτ comes from a meromorphic function on ∆1. Then eτ does not have poles along Supp(F h) and (21) implies that div(eτ ) ≥ −⌈F v⌉. Together with (7), we can find a holomorphic function ζ on ∆1 such that eτ = ζ ◦ f . We now prove the lemma. Let M ∈ N large enough such that the Q-divisor M−1 F v + 1 M−1 F h is klt. Thanks to (20) and the Hölder inequality, we have 1 Zf −1(∆1) eτ 2 M ≤ (Zf −1(∆1) eτ2e−2ψ) (22) M ≤ C2 for some uniform constant C2. Since eτ = ζ ◦ f and ζ is holomorphic on ∆1 and ∆ ⋐ ∆1, by applying maximal principal to ζ, (22) implies that sF v sF h M −1 ) 2 1 M (Zf −1(∆1) 2 M −1 sB M −1 sup z∈∆ζ(z) ≤ C3 · (C2)M where C3 is a constant depending only on ∆ and ∆1. In particular, the norm of = ζ(y0) is less than C3 · (C2)M . Combining this with (18) and the fact cy0 = τXy0 that C2 and C3 are independent of the choice of y0 ∈ ∆, the lemma is proved. (cid:3) Now we prove the claim in the proof of Theorem 3.4. Proof of the claim. Let Ui,0 be the open set defined in Step 3 of the proof of Theorem 3.4. Thanks to Theorem 3.1, ϕi can be extended as a quasi-psh function on f−1(Ui) and satisfying (23) θ + ddcϕi ≥ 0 on f−1(Ui). s⌈F ⌉ is well defined on Let s⌈F⌉ be the canonical section of ⌈F v + F h⌉. Then eϕi f−1(Ui,0) \ (F v + F h). We next prove that is uniformly upper bounded near the generic point of div(F v + F h). Let y be a generic point in Ui,0. By the construction of sy and (6), is uniformly bounded on Xy. Therefore eϕi s⌈F ⌉ For any ∆ ⋐ Y1 ∩ Ui, thanks to Lemma 3.5, there exists a constant c > 0, such that is uniformly bounded near the generic point of div(F h). is a constant on Xy. Then eϕi s⌈F ⌉Xy = syhi eϕi s⌈F ⌉ s⌈F ⌉ s⌈F ⌉ sy ZXy e−2ψ(ωm Together with the facts that for every y ∈ ∆ ∩ Y0. X /f∗(ωY )n) ≥ c s⌈F⌉ 2e−2ψ =ZXy sy2 ZXy sy 12 = 1 hLi is constant on Xy, we see that eϕi and sy s⌈F ⌉ s⌈F ⌉ Y0). Since codimY (Y \ Y1) ≥ 2 and f∗(F v function eϕi s⌈F ⌉ is uniformly upper bounded on f−1(∆∩ 1 ) is of codimension 1 by assumption, the is uniformly upper bounded near the generic point of div(F v 1 ). Now we can prove the claim. Since eϕi s⌈F ⌉ near the generic point of div(F v points of div(F v at the generic points of div(F v is proved to be uniformly upper bounded 1 + F h), the Lelong numbers of ddcϕi at the generic 1 + F h⌉ 1 + F h) is not less than the Lelong numbers of the current ⌈F v 1 + F h). Together with (23), we have (24) θ + ddcϕi ≥ ⌈F v 1 + F h⌉. on f−1(Ui), and the claim is proved. (cid:3) 4. Weak subadjunction 4.1. Definition. [Bou04, Defn.2.2] Let X be a compact Kähler manifold, and let α be a cohomology class on X. We say that α is a modified Kähler class if it contains a Kähler current T such that the generic Lelong number ν(T, D) is zero for every prime divisor D ⊂ X. By [Bou04, Prop.2.3] a cohomology class is modified Kähler if and only if there exists a modification µ : X → X and a Kähler class α on X such that µ∗ α = α. For our purpose we have to fix some more notation: 4.2. Definition. Let X be a compact Kähler manifold, and let α be a modified Kähler class on X. A log-resolution of α is a bimeromorphic morphism µ : X → X from a compact Kähler manifold X such that the exceptional locus is a simple normal crossings divisor Pk j=1 Ej and there exists a Kähler class α on X such that µ∗ α = α. The definition can easily be extended to arbitrary big classes by using the Bouck- som's Zariski decomposition [Bou04, Thm.3.12]. 4.3. Remark. If µ : X → X is a log-resolution of α one can write µ∗α = α + kXj=1 rj Ej and rj > 0 for all j ∈ {1, . . . , k}. For R-divisors this is known as the the negativity lemma [BCHM10, 3.6.2], in the analytic setting we proceed as follows: let T ∈ α be a current with analytic singularities such that the generic Lelong ν(T, D) is zero for every prime divisor D ⊂ X. Resolving the ideal sheaf defining T and pulling back we obtain µ∗α = α′ + kXj=1 r′j Ej ≥ µ∗ω where ω is a Kähler form, r′j > 0 for all j ∈ {1, . . . , k} and α′ is semi-positive with null locus equal to ∪k j=1 εjEj is Kähler, so the statement holds by setting rj := r′j + εj. j=1Ej. For 0 < εj ≪ 1 the class α := α′−Pk 13 4.4. Definition. Let X be a compact Kähler manifold, and let α be a modified Kähler class on X. A subvariety Z ⊂ X is a maximal lc centre if there exists a log- resolution µ : X → X of α with exceptional locus Pk j=1 Ej such that the following holds: • Z is an irreducible component of µ(SuppPk • if we write kXj=1 K X + α = µ∗(KX + α) + j=1 Ej); djEj , then dj ≥ −1 for every Ej mapping onto Z and (up to renumbering) we have µ(E1) = Z and d1 = −1. Following the terminology for singularities of pairs we call the coefficients dj the discrepancies of (X, α). Note that this terminology is somewhat abusive since dj is not determined by the class α but depends on the choice of α (hence implicitly on the choice of a Kähler current T in α that is used to construct the log-resolution). Similarly it would be more appropriate to define Z as an lc centre of the pair (X, T ) with [T ] ∈ α. Since most of the time we will only work with the cohomology class we have chosen to use this more convenient terminology. We can now prove the weak subadjunction formula: Proof of Theorem 1.5. Step 1. Geometric setup. Since Z ⊂ X is a maximal lc centre of (X, α) there exists a log-resolution µ : X → X of α with exceptional locus j=1 Ej such that Z is an irreducible component of µ(SuppPk Pk j=1 Ej) and (25) K X + α = µ∗(KX + α) + djEj , kXj=1 satisfies dj ≥ −1 for every Ej mapping onto Z and (up to renumbering) we have µ(E1) = Z and d1 = −1. Let π : X′ → X be an embedded resolution of Z, then (up to blowing up further X) we can suppose that there exists a factorisation ψ : X → X′. Let Z′ ⊂ X′ be the strict transform of Z. Since π is an isomorphism in the generic point of Z′, the divisors Ej mapping onto Z′ via ψ are exactly those mapping onto Z via µ. Denote by Ql ⊂ Z′ the prime divisors that are images of divisors E1 ∩ Ej via ψE1 . Then we can suppose (up to blowing up further X) that the divisor Xl (ψE1 )∗Ql + kXj=2 E1 ∩ Ej has a support with simple normal crossings. We set f := ψE1, and D = − kXj=2 djDj where Dj := Ej ∩ E1. Note also that the desingularisation πZ′ factors through the normalisation ν : Z → Z, so we have a bimeromorphic morphism τ : Z′ → Z such 14 that πZ′ = ν ◦ τ . We summarise the construction in a commutative diagram: ψ f :=ψE1 ~⑥⑥⑥⑥⑥⑥⑥⑥⑥⑥⑥⑥⑥⑥⑥⑥⑥⑥⑥ ~⑤⑤⑤⑤⑤⑤⑤⑤ (PPPPPPPPPPPPPPPP X′ τ Z′ µ ❅❅❅❅❅❅❅❅❅❅❅❅❅❅❅❅❅❅❅ ❆❆❆❆❆❆❆❆ 7♦♦♦♦♦♦♦♦♦♦♦♦♦♦♦ / X ν Z? E1 X π Z A priori there might be more than one divisor with discrepancy −1 mapping onto Z, but we can use the tie-breaking technique which is well-known in the context of singularities of pairs: recall that the class α is Kähler which is an open property. Thus we can choose 0 < εj ≪ 1 for all j ∈ {2, . . . , k} such that the class α + Pk j=2 εjEj is Kähler. The decomposition K X + (α + kXj=2 εjEj) = µ∗(KX + α) − E1 + kXj=2 (dj + εj)Ej still satisfies the properties in Definition 4.4 and E1 is now the unique divisor with discrepancy −1 mapping onto Z. Note that up to perturbing εj we can suppose that dj + εj is rational for every j ∈ {1, . . . , k}. In order to simplify the notation we will suppose without loss of generality, that these properties already holds for the decomposition (25). Outline of the strategy. The geometric setup above is analogous to the proof of Kawamata's subadjunction formula [Kaw98, Thm.1] and as in Kawamata's proof our aim is now to apply the positivity theorem 3.4 to f to relate KZ′ and (πZ′ )∗(KX + α)Z . However since we deal with an lc centre that is not minimal we encounter some additional problems: the pair (E1, D) is not necessarily (sub-)klt and the centre Z might not be regular in codimension one. In the end this will not change the relation between KZ′ and (πZ′ )∗(KX + α)Z , but it leads to some technical computations which will be carried out in the Steps 3 and 4. Step 2. Relative vanishing. Note that the Q-divisor −K X − E1 +Pk j=2 djEj is µ-ample since its class is equal to α on the µ-fibres. Thus we can apply the relative Kawamata-Viehweg theorem (in its analytic version [Anc87, Thm.2.3] [Nak87]) to obtain that R1µ∗O X (−E1 + kXj=2 ⌈dj⌉Ej) = 0. Pushing the exact sequence 0 → O X (−E1 + kXj=2 kXj=2 ⌈dj⌉Ej) → O X ( 15 ⌈dj⌉Ej) → OE1(⌈−D⌉) → 0 ~ _ ~ / / ( / _ o o 7 down to X, the vanishing of R1 yields a surjective map (26) kXj=2 µ∗(O X ( ⌈dj⌉Ej)) → (µE1 )∗(OE1 (⌈−D⌉)). Since all the divisors Ej are µ-exceptional, we see that µ∗(O X (Pk j=2⌈dj⌉Ej)) is an ideal sheaf I. Moreover, since dj > −1 for all Ej mapping onto Z the sheaf I is isomorphic to the structure sheaf in the generic point of Z . In particular (µE1)∗(OE1 (⌈−D⌉)) has rank one. Step 3. Application of the positivity result. By the adjunction formula we have kXj=2 dj(Ej ∩ E1) = f∗(πZ′ )∗(KX + α)Z . KE1 + αE1 − (27) Since f coincides with µE1 over the generic point of Z′, we know by Step 2 that the direct image sheaf f∗(OE1(⌈−D⌉)) has rank one. In particular f has connected fibres. In general the boundary D does not satisfy the conditions a) and b) in Theorem 3.4, however we can still obtain some important information by applying Theorem 3.4 for a slightly modified boundary: note first that the fibration f is equidimen- sional over the complement of a codimension two set. In particular the direct image sheaf f∗(OE1 (⌈−D⌉)) is reflexive [Har80, Cor.1.7], hence locally free, on the complement of a codimension two set. Thus we can consider the first Chern class c1(f∗(OE1 (⌈−D⌉))) (cf. Definition 2.2). Set L := (πZ′ )∗(KX + α)Z − KZ′, then we claim that (L + c1(f∗(OE1 (⌈−D⌉)))) · ω′1 · . . . · ω′dim Z−1 ≥ 0 (28) for any collection of nef classes ω′j on Z′. Proof of the inequality (28). In the complement of a codimension two subset B ⊂ Z′ the fibration ff −1(Z′\B) is equidimensional, so the direct image sheaf OE1(⌈−Dv⌉) is reflexive. Since it has rank one we thus can write f∗(OE1 (⌈−Dv⌉)) ⊗ OZ′\B = OZ′\B(X elQl) where el ∈ Z and Ql ⊂ Z′ are the prime divisors introduced in the geometric setup. If el > 0 then el is the largest integer such that (ff −1(Z′\B))∗(elQl) ⊂ ⌈−Dv⌉. In particular if Dj maps onto Ql, then dj > −1. If el < 0 there exists a divisor Dj that maps onto Ql such that dj ≤ −1. Moreover if wj is the coefficient of Dj in the pull-back (ff −1(Z′\B))∗Ql, then el is the largest integer such that dj − elwj > −1 for every divisor Dj mapping onto Ql. Thus if we set D := D +X elf∗Ql, then D has normal crossings support (cf. Step 1) and satisfies the condition a) in Theorem 3.4. Moreover if we denote by D = Dh + Dv the decomposition in horizontal and vertical part, then Dh = Dh and Dv = Dv +P elf∗Ql. Since we 16 did not change the horizontal part, the direct image f∗(OE1 (⌈− D⌉)) has rank one. Since P elf∗Ql has integral coefficients, the projection formula shows that (f∗(OE1 (⌈− Dv⌉)))∗∗ ≃ (f∗(OE1 (⌈−Dv⌉)))∗∗ ⊗ OZ′ (−X elQl) ≃ OZ′ . Thus we satisfy the condition b) in Theorem 3.4. Finally note that KE1/Z + αE1 + D = f∗(L +X elQl). So if we set L := L +P elQl, then (29) L + c1(f∗(OE1 (⌈− D⌉))) = L + c1(f∗(OE1 (⌈−D⌉))). Now we apply Theorem 3.4 and obtain L · ω′1 · . . . · ω′dim Z′−1 ≥ 0. Yet by the conditions a) and b) there exists an ideal sheaf I on Z′ that has cosupport of codimension at least two and f∗(OE1 (⌈− D⌉)) ≃ I ⊗ OZ′ (B) with B an effective divisor on Z′. Thus c1(f∗(OE1 (⌈− D⌉))) is represented by the effective divisor B and (28) follows from (29). Step 4. Final computation. In view of our definition of the intersection product on Z (cf. Definition 2.7) we are done if we prove that L · τ∗ω1 · . . . · τ∗ωdim Z−1 ≥ 0 where the ωj are the nef cohomology classes from the statement of Theorem 1.5. We claim that (30) where ∆1 is an effective divisor and ∆2 is a divisor such that πZ′ (Supp ∆2) has codimension at least two in Z. Assuming this claim for the time being let us see how to conclude: by (28) we have c1(f∗(OE1 (⌈−D⌉))) = −∆1 + ∆2 (L + c1(f∗(OE1 (⌈−D⌉)))) · τ∗ω1 · . . . · τ∗ωdim Z−1 ≥ 0. (31) Since the normalisation ν is finite and πZ′ (Supp ∆2) has codimension at least two in Z, we see that τ (Supp ∆2) has codimension at least two in Z. Thus we have c1(f∗(OE1 (⌈−D⌉))) · τ∗ω1 · . . . · τ∗ωdim Z−1 = −∆1 · τ∗ω1 · . . . · τ∗ωdim Z−1 ≤ 0. Hence the statement follows from (31). Proof of the equality (30). Applying as in Step 2 the relative Kawamata-Viehweg vanishing theorem to the morphism ψ we obtain a surjection kXj=2 ψ∗(O X ( ⌈dj⌉Ej)) → (ψE1 )∗(OE1 (⌈−D⌉)) In order to verify (30) note first that some of the divisors Ej might not be ψ- j=2⌈dj⌉Ej)) is an ideal sheaf. However if exceptional, so it is not clear if ψ∗(O X (Pk we restrict the surjection (26) to Z we obtain a surjective map (32) where I is the ideal sheaf introduced in Step 2. There exists an analytic set B ⊂ Z of codimension at least two such that I ⊗OX OZ → (πZ′ )∗(f∗(OE1 (⌈−D⌉))), Z′ \ π−1(B) → Z \ B 17 is isomorphic to the normalisation of Z \ B. In particular the restriction of π to Z′ \ π−1(B) is finite, so the natural map (πZ′ )∗(πZ′ )∗(f∗(OE1(⌈−D⌉))) → f∗(OE1(⌈−D⌉)) is surjective on Z′ \ π−1(B). Pulling back is right exact, so composing with the surjective map (32) we obtain a map from an ideal sheaf to f∗(OE1(⌈−D⌉)) that is surjective on Z′ \ π−1(B). An ideal sheaf is torsion-free, so this map is an isomorphism onto its image in J ⊂ f∗(OE1 (⌈−D⌉)). In the complement of a codimension two set the sheaf J corresponds to an antieffective divisor −∆′1. Since the inclusion J ⊂ f∗(OE1 (⌈−D⌉)) is an isomorphism on Z′ \ π−1(B), there exists an effective divisor ∆′2 with support in π−1(B) such that c1(f∗(OE1 (⌈−D⌉))) = −∆′1 + ∆′2. We denote by ∆1 the part of ∆′1 whose support is not mapped into B (hence maps into the non-normal locus of Z \ B) and set ∆2 := ∆′2 + ∆1 − ∆′1. Then we have c1(f∗(OE1(⌈−D⌉))) = −∆1 + ∆2 and the support of ∆2 maps into B. Since B has codimension at least two this proves the equality (30). (cid:3) 4.5. Remark. In Step 3 of the proof of Theorem 1.5 above we introduce a "bound- ary" c1(f∗(OM (⌈−D⌉))) so that we can apply Theorem 3.4. One should note that this divisor is fundamentally different from the divisor ∆ appearing in [Kaw98, Thm.1, Thm.2]. In fact for a minimal lc centre Kawamata's arguments show that c1(f∗(OM (⌈−D⌉))) = 0, his boundary divisor ∆ is defined in order to obtain the stronger result that L − ∆ is nef. We have to introduce c1(f∗(OM (⌈−D⌉))) since we want to deal with non-minimal centres. 5. Positivity of relative adjoint classes, part 2 Convention : In this section, we use the following convention. Let U be a open set and (fm)m∈N be a sequence of smooth functions on U . We say that if for every open subset V ⋐ U and every index α, we have kfmkC∞(U) → 0, k∂αfmkC 0(V ) → 0. Similarly, in the case (fm)m∈N are smooth formes, we say that kfmkC∞(U) → 0 if every component tends to 0 in the above sense. Before giving the main theorem of this section, we need two preparatory lemmas. The first comes from [Lae02, Part II, Thm 1.3] : 5.1. Lemma.[Lae02, Part II, Thm 1.3] Let X be a compact Kähler manifold and let α be a closed smooth real 2-form on X. Then we can find a strictly increas- ing sequence of integers (sm)m≥1 and a sequence of hermitian line bundles (not necessary holomorphic) (Fm, DFm, hFm )m≥1 on X such that √−1 2π (33) lim m→+∞k ΘhFm (Fm) − smαkC∞(X) = 0. Here DFm is a hermitian connection with respect to the smooth hermitian metric hFm and ΘhFm (Fm) = DFm ◦ DFm. Moreover, let (Wj ) be a small Stein cover of X and let eFm,j be a basis of an iso- metric trivialisation of Fm over Wj i.e., keFm,jkhm = 1. Then we can ask the 18 hermitian connections DFm (under the basis eFm,j) to satisfy the following addi- tional condition: for the (0, 1)-part of DFm on Wj : D′′Fm = ∂ + β0,1 m,j, we have (34) β0,1 m,jkC∞(Wj ) ≤ CkαkC∞(X), where C is a uniform constant independent of j and m. k 1 sm Proof. Thanks to [Lae02, Part II, Thm 1.3], we can find a strictly increasing integer sequence (sm)m≥1 and closed smooth 2-forms (αm)m≥1 on X, such that αm ∈ H 2(X, Z). m→+∞kαm − smαkC∞(X) = 0 and lim Since (Wj ) are small Stein open sets, we can find some smooth 1-forms βm,j on Wj such that (35) 1 2π · dβm,j = αm on Wj and 1 sm k βm,jkC∞(Wj ) ≤ CkαkC∞(X) for a constant C independent of m and j. By using the standard construction (cf. for example [Dem, V, Thm 9.5]), the form (βm,j)j induces a hermitian line bundle (Fm, Dm, hFm ) on X such that Dm = d + √−1 2π βm,j with respect to an isometric trivialisation over Wj. Then ΘhFm (Fm) − smαkC∞(X) = kαm − smαkC∞(X) → 0. Let β0,1 m,j be the (0, 1)-part of βm,j. Then (35) implies (34). (cid:3) Now we can prove the main theorem of this section. 5.2. Theorem. Let X and Y be two compact Kähler manifolds and let f : X → Y be a surjective map with connected fibres such that the general fibre F is simply connected and H 0(F, Ω2 F ) = 0. Let ω be a Kähler form on X such that c1(KF ) + [ωF ] is a pseudoeffective class. Then c1(KX/Y ) + [ω] is pseudoeffective. √−1 2π k Proof. Being pseudoeffective is a closed property, so we can assume without loss of generality that c1(KF ) + [ωF ] is big on F . Step 1: Preparation, Stein Cover. Fix two Kähler metrics ωX , ωY on X and Y respectively. Let h be the smooth her- √−1 mitian metric on KX/Y induced by ωX and ωY . Set α := 2π Θh(KX/Y ). Thanks to Lemma 5.1, there exist a strictly increasing sequence of integers (sm)m≥1 and a sequence of hermitian line bundles (not necessary holomorphic) (Fm, DFm, hFm)m≥1 on X such that (36) k ΘhFm (Fm) − sm(α + ω)kC∞(X) → 0. √−1 2π By our assumption on F we can find a non empty Zariski open subset Y0 of Y such that f is smooth over Y0 and Rif∗OX = 0 on Y0 for every i = 1, 2. Let (Ui)i∈I be a Stein cover of Y0. Therefore (37) H 0,2(f−1(Ui), R) = 0 19 for every i ∈ I. Step 2: Construction of the approximate holomorphic line bundles. Let Θ(0,2) hFm is ∂-exact on f−1(Ui) and (Fm) be the (0, 2)-part of ΘhFm (Fm). Thanks to (37) and (36), Θ(0,2) hFm (Fm) (38) kΘ(0,2) hFm (Fm)kC∞(f −1(Ui)) → 0. We first construct a sequence of (0, 1)-formes βm on f−1(Ui) such that Θ(0,2) hFm and (Fm) = ∂βm kβmkC∞(f −1(Ui)) → 0. (39) In fact, for every y ∈ Ui, as Xy is compact and H 0,2(Xy) = 0, we can find smooth (0, 1)-forms θm on f−1(Ui) such that for every y ∈ Ui (40) (Θ(0,2) (Fm) − ∂θm)Xy = 0 hFm (Fm) − ∂θm =Pj f ⋆(dtj) ∧ γm,j, where (dtj) is a basis of ∧0,1(Ui) Therefore Θ(0,2) hFm and kγm,jkC∞(f −1(Ui)) → 0. Note that Θ(0,2) (Fm) − ∂θm is ∂-closed. Then ∂γm,jXy = 0. As H 0,1(Xy) = 0, we can find θ′m,j on f−1(Uj) such that (γm,j − ∂θ′m,j)Xy = 0 and kθ′m,jkC∞(f −1(Ui)) → 0. As a consequence, kθmkC∞(f −1(Ui)) → 0. and hFm Θ(0,2) hFm (Fm) − ∂(θm +Xj f ⋆(dtj) ∧ θ′m,j) = f ⋆γ for some closed (0, 2)-form γ on Ui and kγkC∞(Ui) → 0. Together with the fact that Ui is Stein, we can thus find βm satisfies (39). Thanks to (39), we can find holomorphic line bundles Li,m on f−1(Ui) equipped with smooth hermitian metrics hi,m such that 2π = ( ΘhFm (Fm)) + ( Θhi,m(Li,m) − ΘhFm (Fm) − sm(α + ω)) + smω. Thanks to the estimates (36) and (41), the first two terms of the right-hand side of the above equality tends to 0. Therefore we can find a sequence of open sets Ui,m ⋐ Ui, such that ∪m≥1Ui,m = Ui, Ui,m ⋐ Ui,m+1 for every m ∈ N, and (42) on f−1(Ui,m). √−1 √−1 2π Θhi,m(Li,m) − sm 2π Θh(KX/Y ) ≥ 0 2π Step 3: Construction of Bergman kernel type metrics. Let ϕi,m be the sm-Bergman kernel associated to the pair (cf. Remark 3.2) (Li,m = smKX/Y + (Li,m − smKX/Y ), hi,m) (43) (44) i.e., ϕi,m(x) := sup g∈A 1 sm lnghi,m(x), where A := {g g ∈ H 0(Xf (x), Li,m),ZXf (x) g 2 sm hi,m 20 ωdim X X /f∗ωdim Y Y = 1}. (41) k √−1 2π By construction, we have ΘhFm (Fm) − √−1 2π Θhi,m(Li,m) − sm √−1 √−1 2π √−1 2π √−1 2π Θhi,m(Li,m)kC∞(f −1(Ui)) → 0. √−1 2π Θhi,m(Li,m) − smα Θh(KX/Y ) = √−1 Thanks to (42), we can apply Theorem 3.1 to the pair (43) over f−1(Ui,m). particular, we have In (45) (α + ω) + ddcϕi,m ≥ 0 on f−1(Ui,m). We recall that ϕi,m is invariant after a normalisation of hi,m, namely, if we replace the metric hi,mXy by c · hi,mXy for some constant c > 0, the associated Bergman kernel function ϕi,mXy is unchanged cf. Remark 3.2 (3). Let y ∈ Ui be a generic point. Thanks to the above remark and (41), we can find a constant cy > 0 independent of m, such that cy ≤ hi,mXy ≤ c−1 y . Therefore, by mean value inequality, ϕi,mXy is uniformly upper bounded. Therefore we can define ϕi := lim k→+∞ ( sup m≥k ϕi,m)⋆, where ⋆ is the u.s.c regularization. Thanks to (44), ϕi cannot be identically −∞. Therefore ϕi is a quasi-psh. As ∪m≥1Ui,m = Ui, (45) implies (46) on f−1(Ui) in the sense of currents. α + ω + ddcϕi ≥ 0 Step 4: Final conclusion. We claim that Claim 1. ϕi = ϕj on f−1(Ui ∩ Uj) for every i, j. Claim 2. For every small Stein open set V in X, we can find a constant CV depending only on V such that ϕi(x) ≤ CV for every i and x ∈ V ∩ f−1(Ui). We postpone the proof of these two claims and finish first the proof of the theorem. Thanks to Claim 1, (ϕi)i∈I defines a global quasi-psh function ϕ on f−1(Y0) and (46) implies that α + ω + ddcϕ ≥ 0 on f−1(Y0). Thanks to Claim 2, we have ϕ ≤ CV on V ∩ f−1(Y0). Therefore ϕ can be extended as a quasi-psh function on V . Since Claim 2 is true for every small Stein open set V , ϕ can be extended as a quasi-psh function on X and satisfies As a consequence, c1(KX/Y ) + [ω] is pseudoeffective and the theorem is proved. (cid:3) α + ω + ddcϕ ≥ 0 on X. We are left to prove the two claims in the proof of the theorem. 5.3. Lemma. The claim 1 holds, i.e., ϕi = ϕj on f−1(Ui ∩ Uj) for every i, j. Proof. Let y ∈ Ui ∩ Uj be a generic point. Thanks to (41), we have (47) √−1 √−1 Θhj,m(Lj,m)XykC∞(Xy) = 0. 2π lim m→+∞k Θhi,m(Li,m)Xy − 2π When m is large enough, (47) implies that c1(Li,mXy ) = c1(Lj,mXy ) ∈ H 1,1(Xy) ∩ H 2(Xy, Z). As Xy is simply connected, Pic0(Xy) = 0. Therefore (48) Li,mXy = Lj,mXy 21 for m ≫ 1. Under the isomorphism of (48), by applying ∂∂-lemma, (47) imply the existence of constants cm ∈ R and smooth functions τm ∈ C∞(Xy) such that hi,m = hj,mecm+τm on Xy and lim m→+∞kτmkC∞(Xy) = 0. Combining with the construction of ϕi,m and ϕj,m, we know that Therefore kϕi,m − ϕj,mkC 0(Xy) ≤ kτmkC 0(Xy) → 0. (49) As (49) is proved for every generic point y ∈ Ui ∩ Uj, we have ϕiXy = ϕjXy The lemma is proved. ϕi = ϕj on f−1(Ui ∩ Uj). (cid:3) It remains to prove the claim 2. Note that (Li,m, hi,m) is defined only on f−1(Ui), we can not directly apply Proposition 3.3 to (Li,m, hi,m).The idea of the proof is as follows. Thanks to the construction of Fm and Li,m, by using ∂∂-lemma, we can prove that, after multiplying by a constant (which depends on f (x) ∈ Y ), the difference between hFmXf (x) and hi,mXf (x) is uniformly controlled for m ≫ 1 5. Therefore (FmXf (x) , hFm) is not far from (Li,mXf (x) , hi,m). Note that, using again (36), FmV is not far from a holomorphic line bundle over V . Combining Proposition 3.3 with these two facts, we can finally prove the claim 2. 5.4. Lemma. The claim 2 holds, i.e., for every small Stein open set V in X, we can find a constant CV depending only on V such that ϕi(x) ≤ CV for every i and x ∈ V ∩ f−1(Ui). Proof. Step 1: Global approximation. Fix a small Stein cover (Wj )N j=1 of X. Without loss of generality, we can assume that V ⋐ W1. Let (Fm, DFm, hFm)m≥1 be the hermitian line bundles (not necessary holomorphic) constructed in the step 1 of the proof of Theorem 5.2. Let eFm,j be a basis of a isometric trivialisation of Fm over Wj i.e., keFm,jkhFm = 1. Under this = ∂ + β0,1 trivialisation, we suppose that the (0, 1)-part of DFm on Wj is D′′Fm m,j, where β0,1 m,j is a smooth (0, 1)-form on Wj. By Lemma 5.1, we can assume that (50) 1 sm k β0,1 m,jkC∞(Wj ) ≤ C1kα + ωkC∞(X) for a uniform constant C1 independent of m and j. Step 2: Local estimation near V . Thanks to (36), we know that Fm is not far from a holomorphic line bundle. In this step, we would like to give a more precise description of this on W1. Since W1 is a small Stein open set, thanks to (36), we can find {σ0,1 such that ∂σ0,1 m }m≥1 on W1 (Fm) and lim m kC∞(W1) = 0. Then we have m = −Θ(0,2) hFm m→+∞kσ0,1 m )2 = 0 on W1, (D′′F,m + σ0,1 (51) 5The bigness of m ≫ 1 depends on f (x). 22 and √−1 F,m+σ0,1 ΘhFm ,D′′ (Fm) is the curvature for the Chern connection on Fm with (Fm) − sm(α + ω)kC∞(X) → 0, m 2π k where ΘhFm ,D′′ F,m+σ0,1 respect to complex structure D′′F,m + σ0,1 Note that we can find smooth functions {ψm}m≥1 on W1 such that √−1 2π ΘhFm ,D′′ F,m+σ0,1 m m m and the metric hFm . (Fm) is a closed (1, 1)-form on W1. By ∂∂-lemma, (i) √−1 2π ΘhFm e−ψm ,D′′ lim F,m+σ0,1 m (Fm) = sm(α + ω) on W1 for every m ∈ N. 6 (ii) m→+∞ (kσ0,1 Thanks to (51), β0,1 m is ∂-closed. Applying standard L2-estimate, by re- stricting on some a little bit smaller open subset of W1 (we still denote it by W1 for simplicity), there exists a smooth function ηm on W1 such that m kC∞(W1) + kψmkC∞(W1)) = 0. m,1 + σ0,1 (52) and ∂ηm = β0,1 m,1 + σ0,1 m on W1 1 smkηmkC∞(W1) ≤ C2 smkβ0,1 m,1 + σ0,1 m kC∞(W1) for a constant C2 independent of m. Combining this with (50) and (ii), we get (53) limm→+∞ 1 smkηmkC∞(W1) ≤ C1 · C2. Moreover, by (52), e−ηm · eFm,1 is a holomorphic basis of (W1, Fm, D′′Fm Step 3: Final conclusion. Let x ∈ V ∩ f−1(Ui) and set y := f (x). Claim. For m large enough, there exists a bg ∈ H 0(Xy ∩ W1, Fm, D′′Fm that + σ0,1 m ). + σ0,1 m ). such (54) and ZXy∩W1 bg 2 sm hFm ωdim X X /ωdim Y Y ≤ 2 (55) ϕi,m(x) ≤ lnbghFm (x) + 2. As e−ηm · eFm,1 is a holomorphic basis of (W1, Fm, D′′Fm 1 sm + σ0,1 m ), we have We postphone the proof of the claim later and first finish the proof of our lemma. for some holomorphic function f on W1∩Xy. Thanks to (53), we can find a uniform constant C3 > 0 independent of m such that bg = f · e−ηm · eFm,1 (56) C−1 3 ≤ e−ηm · eFm,1 2 sm hFm ≤ C3 on W1. 6Here Θ 0,1 +σ m spect to complex structure D′′ hFm e−ψm ,D′′ F,m F,m + σ0,1 m and the metric hFm · e−ψm . 23 (Fm) is the curvature for the Chern connection on Fm with re- Together with (54), we have ZXy∩W1 f 2 sm ωdim X X /ωdim Y Y ≤ 2C3. 2 By applying the Ohsawa-Takegoshi extension theorem [BP10, Prop 0.2], we know that f by a uniform constant C4. Combining this with (55), the lemma is proved. sm is uniformly controled. Together with (56), 1 sm lnbghFm (x) is controled (cid:3) It remains to prove the claim in Lemma 5.4. Proof of the claim in Lemma 5.4. By (36) and Pic0(Xy) = 0, when m is large enough, we can find a smooth (0, 1)-forms τ 0,1 (57) lim m→+∞kτ 0,1 m kC∞(Xy) = 0 and m on Xy such that (Fm, D′′Fm + τ 0,1 m )Xy ≃ Li,mXy . Let ΘhFm ,τ 0,1 respect to hFm and the complex structure D′′Fm + τ 0,1 Thanks (36) and (57) imply that (FmXy ) be the curvature calculated for the Chern connection with m for the line bundle FmXy . m (58) lim m→+∞kΘhFm ,τ 0,1 m (FmXy ) − Θhi,m(Li,mXy )kC∞(Xy ) = 0. By using ∂∂-lemma over Xy, under the holomorphic isomorphism of (57), (58) implies the existence of a constant cm,y and a smooth function eψm on Xy such that on Xy, hFm · e− eψm = hi,m · e−cm,y (59) and (60) lim m→+∞keψmkC∞(Xy ) = 0. Here cm,y is a constant on Xy which depends only on m and y. By the definition of ϕi,m, there exists a g ∈ H 0(Xy, Li,m) such that (61) ωdim X ϕi,m(x) = and 2 sm hi,m X 1 sm lnghi,m(x) ZXy g /ωdim Y Y = 1. Using the holomorphic isomorphism (57) and the metric estimations (60) and (59), we can thus find a eg ∈ H 0(Xy, Fm, D′′Fm ωdim X X /ωdim Y Y (62) = 1 2 sm hFm + τ 0,1 m )7 such that and ϕi,m(x) ≤ 1 sm ZXy eg lneghFm (x) + 1 where m is large enough. Here we use Remark 3.2 (3) and the fact that cm,y is constant on Xy (although it might be very large). Now we prove the claim. Thanks to (57) and the fact that τ 0,1 m is ∂-exact on the Stein open set Xy ∩ W1, there exists some smooth functions ζm on Xy ∩ W1, such that m − σ0,1 and (63) ∂ζm = τ 0,1 m − σ0,1 m on Xy ∩ W1 lim m→+∞ 1 smkζmkC∞(Xy∩W1) ≤ lim m→+∞ Cy smkτ 0,1 m − σ0,1 m kC∞(Xy∩W1) = 0. 7It means that eg is a holomorphic section of Fm on Xy with respect to the complex structure D′′ Fm + τ 0,1 m . 24 for a constant Cy independent of m, but depending on y. Set bg := eζm ·eg. Then bg ∈ H 0(Xy ∩ W1, Fm, D′′Fm when m is large enough, we have + σ0,1 m ). Thanks to (63) and (62), (64) and (65) The claim is proved. ZXy∩W1 bg ωdim X X /ωdim Y Y ≤ 2 2 sm hFm ϕi,m(x) ≤ 1 sm lnbghFm (x) + 2. 6. Proof of the main theorem (cid:3) We start with an easy, but important lemma relating null locus and lc centres. 6.1. Lemma. Let X be a compact Kähler manifold, and let α be a nef and big class such that the null locus Null(α) has no divisorial components. Let Z ⊂ X be an irreducible component of Null(α). Then there exists a positive real number c such that Z is a maximal lc centre for (X, cα). Remark. The coefficient c depends on the choice of Z, so in general the other irreducible components of Null(α) will not be lc centres for (X, cα). Proof. By a theorem of Collins of Tosatti [CT15, Thm.1.1] the non-Kähler locus EnK(α) coincides with the null-locus of Null(α). Moreover by [Bou04, Thm.3.17] there exists a Kähler current T with analytic singularities in the class α such that the Lelong set coincides with EnK(α). Since the non-Kähler locus has no divisorial components the class α is a modified Kähler class [Bou04, Defn.2.2]. By [Bou04, Prop.2.3] the class α has a log-resolution µ : X → X such that µ∗ α = α. In fact the proof proceeds by desingularising a Kähler current with analytic singularities in the class α, so, using the current T defined above, we see that the µ-exceptional locus maps exactly onto Null(α). Up to blowing up further the exceptional locus is a SNC divisor. By Remark 4.3 we have µ∗α = α + kXj=1 rjDj. with rj > 0 for all j ∈ {1, . . . , k}. Since α is nef and big, the class α + mµ∗α is Kähler for all m > 0. Thus up to replacing the decomposition above by µ∗α = α + mµ∗α m + 1 + kXj=1 rj m + 1 Dj for m ≫ 0 we can suppose that rj < 1 for all j ∈ {1, . . . , k}. Since X is smooth we have K X = µ∗KX +Pk j=1 ajEj with aj a positive integer. Since rj < 1 we have aj − rj > −1 for all Ej mapping onto Z. Thus we can choose a c ∈ R+ such that aj − crj ≥ −1 for all Ej mapping onto Z and equality holds for at least one divisor. (cid:3) As a first step toward Theorem 1.3 we can now prove the following: 25 6.2. Theorem. Let X be a compact Kähler manifold of dimension n. Suppose that Conjecture 1.2 holds for all manifolds of dimension at most n − 1. Suppose that KX is pseudoeffective but not nef, and let ω be a Kähler class on X such that α := KX + ω is nef and big but not Kähler. Let Z ⊂ X be an irreducible component of maximal dimension of the null-locus Null(α), and let π : Z′ → Z be the composition of the normalisation and a resolution of singularities. Let k be the numerical dimension of π∗αZ (cf. Definition 2.5). Then we have KZ′ · π∗αk Z · π∗ωdim Z−k−1 Z < 0. In particular Z′ is uniruled. Proof of Theorem 6.2. Since α = KX + ω and π∗αk+1 Z = 0 we have π∗KXZ · π∗αk Z = −π∗ωZ · π∗αk Z . By hypothesis k < dim Z so dim Z−k−1 is non-negative. Since π∗αk nef class and ω is Kähler this implies by Remark 2.6 that = −π∗ωdim Z−k Z · π∗ωdim Z−k−1 π∗KXZ · π∗αk · π∗αk (66) Z Z Z < 0. Z is a non-zero Our goal will be to prove that KZ′ · π∗αk Z · π∗ωdim Z−k−1 Z < 0. This inequality implies the statement: since KZ′ is not pseudoeffective and Conjec- ture 1.2 holds in dimension at most n − 1 ≥ dim Z′ we obtain that Z′ is uniruled. We will make a case distinction: Step 1. The null-locus of α contains an irreducible divisor. Since Z has maximal dimension, it is a divisor. Since KX is pseudoeffective we can consider the divisorial Zariski decomposition [Bou04, Defn.3.7] c1(KX ) =X eiZi + P (KX ), where ei ≥ 0, the Zi ⊂ X are prime divisors and P (KX ) is a modified nef class [Bou04, Defn.2.2]. Arguing as in [HP16, Lemma 4.1] we see that the inequality (66) implies (up to renumbering) that Z1 = Z and (67) Thus the normal bundle NZ/X ≃ OZ (Z) is negative with respect to these nef classes. Moreover there exist effective Q-divisors on D1 and D2 on Z′ such that π∗(c1(OZ (Z))) · π∗αk Z · π∗ωn−k−2 < 0. Z and π(D1) has codimension at least two in Z (cf. [Rei94, Prop.2.3]). Thus we have KZ′ = π∗(KX + Z) + D1 − D2 Z KZ′ · π∗αk Z · π∗ωn−k−2 ≤ π∗(KX + Z) · π∗αk Z · π∗ωn−k−2 Combining (66) and (67) we obtain that the right hand side is negative. Step 2. The null-locus of α has no divisorial components. In this case we know by Lemma 6.1 that there exists a c > 0 such that Z is a maximal lc centre for (X, cα). The classes π∗αZ and π∗ωZ are nef, so by Theorem 1.5 we have Z . KZ′ · π∗αk Z · π∗ωdim Z−k−1 Z ≤ π∗(KX + cα)Z · π∗αk Since k is the numerical dimension of π∗αZ we have c π∗αk+1 Thus (66) yields the claim. Z Z Z · π∗ωdim Z−k−1 . · π∗ωdim Z−k−1 Z = 0. (cid:3) 26 6.3. Remark. We used the hypothesis that Z has maximal dimension only in Step 1, so our proof actually yields a more precise statement: Null(α) contains a uniruled divisor or all the components of Null(α) are uniruled. We come now to the technical problem mentioned in the introduction: 6.4. Problem. Let X be a compact Kähler manifold, and let α ∈ N 1(X) be a nef cohomology class. Does there exist a real number b > 0 such that for every (rational) curve C ⊂ X we have either α · C = 0 or α · C ≥ b ? 6.5. Remark. If α is the class of a nef Q-divisor, the answer is obviously yes: some positive multiple mα is integral, so we can choose b := 1 m . If α is a Kähler class the answer is also yes: by Bishop's theorem there are only finitely many deformation families of curves C such that α · C ≤ 1, so α · C takes only finitely many values in ]0, 1[. However, even for the class of an R-divisor on a projective manifold X it seems possible that the values α · C accumulate at 0 [Laz04, Rem.1.3.12]. In the proof of Theorem 1.3 we will use that α is an adjoint class to obtain the existence of the lower bound b. The problem 6.4 is invariant under certain birational morphisms: 6.6. Lemma. Let π : X → X′ be a holomorphic map between normal projective varieties X and X′. Let α′ be a nef R-divisor class on X′ and set α := π∗α′. a) Suppose that there exists a real number b > 0 such that for every (rational) curve C′ ⊂ X′ we have α′ · C′ = 0 or α′ · C′ ≥ b. Then for every (rational) curve C ⊂ X we have α · C = 0 or α · C ≥ b. b) Suppose that there exists a real number b > 0 such that for every (rational) curve C ⊂ X we have α · C = 0 or α · C ≥ b. Suppose also that X has klt singularities and π is the contraction of a KX-negative extremal ray. Then for every (rational) curve C′ ⊂ X′ we have α′ · C′ = 0 or α′ · C′ ≥ b. Proof. Proof of a) Let C ⊂ X be a (rational) curve such that α · C 6= 0. the image C′ := π(C) ⊂ X′ is a (rational) curve and the induced map C → C′ has degree d ≥ 1. Thus the projection formula yields α · C = π∗α′ · C = α′ · π∗(C) = dα′ · C′ ≥ db ≥ b. Proof of b) Let C′ ⊂ X′ be an arbitrary (rational) curve such that α′ · C′ 6= 0. By [HM07, Cor.1.7(2)] the natural map π−1(C′) → C′ has a section, so there exists a (rational) curve C ⊂ X such that the map πC : C → C′ has degree one. Thus the projection formula yields α′ · C′ = α′ · π∗(C) = π∗α · C ≥ b. (cid:3) 6.7. Remark. It is easy to see that statement a) also holds when X and X′ are compact Kähler manifolds and α′ is a nef cohomology class on X′. 6.8. Corollary. Let X be a normal projective Q-factorial variety with klt singu- larities, and let α be a nef R-divisor class on X. Suppose that there exists a real number b > 0 such that for every (rational) curve C ⊂ X we have α · C = 0 or α · C ≥ b. Let µ : X 99K X′ be the divisorial contraction or flip of a KX-negative extremal ray Γ such that α · Γ = 0. Set α′ := µ∗(α). Then α′ is a nef R-divisor class on X′ and for every (rational) curve C ⊂ X we have α · C = 0 or α · C ≥ b. 27 Proof. If µ is divisorial the condition α · Γ = 0 implies that α = µ∗α′ [KM98, If µ is a flip, let f : X → Y be the Cor.3.17]. Thus Lemma 6.6, b) applies. contraction of the extremal ray and f′ : X′ → Y the flipping map. Since α · Γ = 0 there exists an R-divisor class αY on Y such that α = f∗αY [KM98, Cor.3.17]. Moreover we have α′ = (f′)∗αY since they coincide in the complement of the flipped locus. Thus we conclude by applying Lemma 6.6,b) to f and Lemma 6.6,a) to f′. (cid:3) 6.9. Proposition. Let F be a projective manifold, and let α be a nef R-divisor class on F . Suppose that there exists a real number b > 0 such that for every rational curve C ⊂ F such that α · C 6= 0 we have (68) α · C > b. Then one of the following holds • F is dominated by rational curves C ⊂ F such that α · C = 0; or • the class KF + 2 dim F b α is pseudoeffective. Proof. Note that, up to replacing α by 2 dim F b α, we can suppose that α · C > 2 dim F (69) for every rational curve C ⊂ F that is not α-trivial. Suppose that KF + α is not pseudoeffective, then our goal is to show that F is covered by α-trivial rational curves. Since KF + α is not pseudoeffective, there exists an ample R-divisor H such that KF + α + H is not pseudoeffective. Since H and α + H are ample we can choose effective R-divisors ∆H ∼R H and ∆ ∼R α + H such that the pairs (F, ∆H ) and (F, ∆) are klt. By [BCHM10, Cor.1.3.3] we can run a KF + ∆-MMP (F, ∆) =: (F0, ∆0) µ0 99K (F1, ∆1) µ1 99K . . . µk 99K (Fk, ∆k), that is for every i ∈ {0, . . . , k − 1} the map µi : Fi 99K Fi+1 is either a divisorial Mori contraction of a KFi + ∆i-negative extremal ray Γi in NE(Xi) or the flip of a small contraction of such an extremal ray. Note that for every i ∈ {0, . . . , k} the variety Fi is normal Q-factorial and the pair (Fi, ∆i) is klt. Moreover Fk admits a Mori contraction of fibre type ψ : Fk → Y contracting an extremal ray Γk such that (KFk + ∆k) · Γk < 0. Set ∆H,0 := ∆H , α0 := α and for all i ∈ {0, . . . , k − 1} we define inductively ∆H,i+1 := (µi)∗(∆H,i), αi+1 := (µi)∗(αi). KFi + ∆i ≡ KFi + ∆H,i + αi. Note that for all i ∈ {0, . . . , k} we have (70) We claim that for all i ∈ {0, . . . , k} the R-divisor class αi is nef and αi · Γi = 0. Moreover the pairs (Xi, ∆H,i) are klt. Assuming this for the time being, let us see how to conclude: since ψ : Fk → Y is a Mori fibre space and the extremal ray Γk is αk-trivial, we see that Fk is dominated by αk-trivial rational curves (Ct)t∈T . A general member of this family of rational curves is not contained in the exceptional locus of F0 99K Fk, so the strict transforms define a dominant family of rational curves (C′t)t∈T of F0. Since all the birational contractions in the MMP F0 99K Fk are α•-trivial, we easily see (cf. the proof of Corollary 6.8) that α · C′t = αk · Ct = 0. 28 Proof of the claim. Since α0 is nef, we have 0 > (KF0 + ∆0) · Γ0 = (KF0 + ∆H,0 + α0) · Γ0 ≥ (KF0 + ∆H,0) · Γ0. Thus the extremal ray Γ0 is KF0 + ∆H,0-negative, in particular the pair (F1, ∆1) is klt [KM98, Cor.3.42, 3.43]. Moreover there exists by [Kaw91, Thm.1] a rational curve [C0] ∈ Γ0 such that (KF0 + ∆H,0) · C0 ≥ −2 dim F . Thus if α0 · C0 6= 0, the inequality (69) implies that (KF0 + ∆0) · C0 = (KF0 + ∆H,0) · C0 + α0 · C0 > 0. In particular the extremal ray Γ0 is not KF0 + ∆0-negative, a contradiction to our assumption. Thus we have α0 · C0 = 0. By Corollary 6.8 this implies that α1 is nef and satisfies the inequality (69). The claim now follows by induction on i. (cid:3) 6.10. Remark. For the proof of Theorem 1.3 we will use the MRC fibration of a uniruled manifold. Since the original papers [KMM92, Cam92] are formulated for projective manifolds, let us recall that for a compact Kähler manifold M that is uniruled the MRC fibration is defined as an almost holomorphic map f : M 99K N such that the general fibre F is rationally connected and the dimension of F is maximal among all the fibrations of this type. The existence of the MRC fibration follows, as in the projective case, from the existence of a quotient map for covering families [Cam04]. The base N is not uniruled : arguing by contradiction we consider a dominating family (Ct)t∈T of rational curves on N . Let Mt be a desingularisation of f−1(Ct) for a general Ct, then Mt is a compact Kähler manifold with a fibration onto a curve Mt → Ct such that the general fibre is rationally connected. In particular H 0(Mt, Ω2 ) = 0 so Mt is projective by Kodaira's criterion. Thus we can apply the Graber-Harris-Starr theorem [GHS03] to see that Mt is rationally connected, a contradiction. Mt Proof of Theorem 1.3. Let ω be a Kähler class such that α := KX + ω is nef and big, but not Kähler. By Theorem 6.2 there exists a subvariety Z ⊂ X contained in the null-locus Null(α) that is uniruled. More precisely let π : Z′ → Z be a desingularisation, and denote by k the numerical dimension of α′ := π∗αZ . Then we know by Theorem 6.2 that Since α′k+1 = 0 this actually implies that KZ′ · α′k · π∗ωdim Z−k−1 Z < 0. (71) (KZ′ + λα′) · α′k · π∗ωdim Z−k−1 Z < 0 ∀ λ > 0. Our goal is to prove that this implies that Z contains a KX-negative rational curve. Arguing by contradiction we suppose that KX · C ≥ 0 for every rational curve C ⊂ Z. Since ω is a Kähler class this implies by Remark 6.5 that there exists a b > 0 such that for every rational curve C ⊂ Z we have α · C = (KX + ω) · C ≥ ω · C ≥ b. (72) By Lemma 6.6a) and Remark 6.7 this implies that for every rational curve C′ ⊂ Z′ we have α′ · C′ = 0 or α′ · C′ ≥ b. Since Z′ is uniruled we can consider the MRC-fibration f : Z′ 99K Y (cf. Remark 6.10). The general fibre F is rationally connected, in particular we can consider α′F as a nef R-divisor class. Moreover the inequality above shows that α′F satisfies the condition (68) in Proposition 6.9. If F is dominated by α′F -trivial rational 29 curves, then Z′ is dominated by α′-trivial rational curves. A general member of this dominating family is not contracted by π, so Z is dominated by α-trivial rational curves. This possibility is excluded by (72), so Proposition 6.9 shows that there exists a λ > 0 such that KF + λα′F is pseudoeffective. We will now prove that KZ′ + λα is pseudoeffective, which clearly contradicts (71). If ν : Z′′ → Z is a resolution of the indeterminacies of f such that KZ′′ + ν∗(λα) is pseudoeffective, then KZ′ + λα = (ν)∗(KZ′′ + ν∗(λα)) is pseudoeffective. Thus we can assume without loss of generality that the MRC-fibration f is a holomorphic map. Let ω′ be a Kähler class on Z′, then for every ε > 0 the class λα′ + εω is Kähler and KF + (λα + εω)F is pseudoeffective. Thus we can apply Theorem 5.2 to f : Z′ → Y to see that KZ′/Y + λα + εω is pseudoeffective. Note now that Y has dimension at most dim X − 2 is not uniruled (Remark 6.10) Since we assume that Conjecture 1.2 holds in dimension up to dim X−1, we obtain that KY is pseudoeffective. Thus we see that KZ′ +λα+εω is pseudoeffective for all ε > 0. The statement follows by taking the limit ε → 0. (cid:3) References [Anc87] [Ara10] Vincenzo Ancona. Vanishing and nonvanishing theorems for numerically effective line bundles on complex spaces. Ann. Mat. Pura Appl. (4), 149:153–164, 1987. Carolina Araujo. The cone of pseudo-effective divisors of log varieties after Batyrev. Math. Z., 264(1):179–193, 2010. [BCHM10] Caucher Birkar, Paolo Cascini, Christopher D. Hacon, and James McKernan. Ex- istence of minimal models for varieties of log general type. J. Amer. Math. Soc., 23(2):405–468, 2010. [BDPP13] Sébastien Boucksom, Jean-Pierre Demailly, Mihai Păun, and Thomas Peternell. The pseudo-effective cone of a compact Kähler manifold and varieties of negative Kodaira dimension. Journal of Algebraic Geometry, 22:201–248, 2013. [BHN15] Mauro C. Beltrametti, Andreas Höring, and Carla Novelli. Fano varieties with small [Bou04] [BP08] [BP10] [Bru06] [Cam92] [Cam04] [Cao14] [CP90] [CT15] [Dem] [Dem12] non-klt locus. Int. Math. Res. Not. IMRN, (11):3094–3120, 2015. Sébastien Boucksom. Divisorial Zariski decompositions on compact complex manifolds. Ann. Sci. École Norm. Sup. (4), 37(1):45–76, 2004. Bo Berndtsson and Mihai Păun. Bergman kernels and the pseudoeffectivity of relative canonical bundles. Duke Math. J., 145(2):341–378, 2008. Bo Berndtsson and Mihai Păun. Bergman kernels and subadjunction. ArXiv e-prints, February 2010. Marco Brunella. A positivity property for foliations on compact Kähler manifolds. Internat. J. Math., 17(1):35–43, 2006. Frédéric Campana. Connexité rationnelle des variétés de Fano. Ann. Sci. École Norm. Sup. (4), 25(5):539–545, 1992. Frédéric Campana. Orbifolds, special varieties and classification theory: an appendix. Ann. Inst. Fourier (Grenoble), 54(3):631–665, 2004. Junyan Cao. Ohsawa-Takegoshi extension theorem for compact Kähler manifolds and applications. ArXiv e-prints, April 2014. Frédéric Campana and Thomas Peternell. Algebraicity of the ample cone of projective varieties. J. Reine Angew. Math., 407:160–166, 1990. Tristan C. Collins and Valentino Tosatti. Kähler currents and null loci. Invent.Math., 202(3):1167–1198, 2015. Jean-Pierre Demailly. http://www-fourier.ujf-grenoble.fr/∼demailly/documents.html. Jean-Pierre Demailly. Analytic methods in algebraic geometry, volume 1 of Surveys of Modern Mathematics. International Press, Somerville, MA; Higher Education Press, Beijing, 2012. differential geometry. Complex analytic and 30 [DP04] [FG12] [FM00] [GHS03] [Gue16] [GZ15] [Har77] [Har80] [HM05] [HM07] [HP16] [Kaw91] [Kaw98] [KK83] [KM98] Jean-Pierre Demailly and Mihai Păun. Numerical characterization of the Kähler cone of a compact Kähler manifold. Ann. of Math. (2), 159(3):1247–1274, 2004. Osamu Fujino and Yoshinori Gongyo. On canonical bundle formulas and subadjunc- tions. Michigan Math. J., 61(2):255–264, 2012. Osamu Fujino and Shigefumi Mori. A canonical bundle formula. J. Differential Geom., 56(1):167–188, 2000. Tom Graber, Joe Harris, and Jason Starr. Families of rationally connected varieties. J. Amer. Math. Soc., 16(1):57–67 (electronic), 2003. H. Guenancia. Families of conic K\"ahler-Einstein metrics. ArXiv e-prints, May 2016. Qi'an Guan and Xiangyu Zhou. A solution of an l2 extension problem with an optimal estimate and applications. Annals of Mathematics, 181:1139–1208, 2015. Robin Hartshorne. Algebraic geometry. Springer-Verlag, New York, 1977. Graduate Texts in Mathematics, No. 52. Robin Hartshorne. Stable reflexive sheaves. Math. Ann., 254(2):121–176, 1980. Christopher Hacon and James McKernan. On the existence of flips. arXiv preprint, 0507597, 2005. Christopher D. Hacon and James Mckernan. On Shokurov's rational connectedness conjecture. Duke Math. J., 138(1):119–136, 2007. Andreas Höring and Thomas Peternell. Minimal models for Kähler threefolds. Invent. Math., 203(1):217–264, 2016. Yujiro Kawamata. On the length of an extremal rational curve. Invent. Math., 105(3):609–611, 1991. Yujiro Kawamata. Subadjunction of log canonical divisors. II. Amer. J. Math., 120(5):893–899, 1998. Ludger Kaup and Burchard Kaup. Holomorphic functions of several variables, vol- ume 3 of de Gruyter Studies in Mathematics. Walter de Gruyter & Co., Berlin, 1983. János Kollár and Shigefumi Mori. Birational geometry of algebraic varieties, volume 134 of Cambridge Tracts in Mathematics. Cambridge University Press, Cambridge, 1998. With the collaboration of C. H. Clemens and A. Corti. [KMM92] Janos Kollár, Yoichi Miyaoka, and Shigefumi Mori. Rational connectedness and bound- [Lae02] [Laz04] [MM86] [Mor79] [Mor82] [Nak87] edness of Fano manifolds. J. Diff. Geom. 36, pages 765–769, 1992. Laurent Laeng. Estimations spectrales asymptotiques en géométrie hermitienne. Theses, Université Joseph-Fourier - Grenoble I, October 2002. https://tel.archives- ouvertes.fr/tel-00002098. Robert Lazarsfeld. Positivity in algebraic geometry. I, volume 48 of Ergebnisse der Mathematik und ihrer Grenzgebiete. Springer-Verlag, Berlin, 2004. Classical setting: line bundles and linear series. Yoichi Miyaoka and Shigefumi Mori. A numerical criterion for uniruledness. Ann. of Math. (2), 124(1):65–69, 1986. Shigefumi Mori. Projective manifolds with ample tangent bundles. Ann. of Math. (2), 110(3):593–606, 1979. Shigefumi Mori. Threefolds whose canonical bundles are not numerically effective. Ann. of Math. (2), 116(1):133–176, 1982. Noboru Nakayama. The lower semicontinuity of the plurigenera of complex varieties. In Algebraic geometry, Sendai, 1985, volume 10 of Adv. Stud. Pure Math., pages 551–590. North-Holland, Amsterdam, 1987. [Pău12a] Mihai Păun. Relative adjoint transcendental classes and Albanese maps of compact Kähler manifolds with nef Ricci curvature. arXiv preprint, 1209.2195, 2012. [Pău12b] Mihai Păun. Relative critical exponents, non-vanishing and metrics with minimal sin- [Rei94] [Sch12] [Tak06] [Tak08] gularities. Invent. Math., 187(1):195–258, 2012. Miles Reid. Nonnormal del Pezzo surfaces. Publ. Res. Inst. Math. Sci., 30(5):695–727, 1994. Georg Schumacher. Positivity of relative canonical bundles and applications. Invent. Math., 190(1):1–56, 2012. Shigeharu Takayama. Pluricanonical systems on algebraic varieties of general type. Invent. Math., 165(3):551–587, 2006. Shigeharu Takayama. On the uniruledness of stable base loci. J. Differential Geom., 78(3):521–541, 2008. 31 [Yi14] Li Yi. An Ohsawa-Takegoshi theorem on compact Kähler manifolds. Sci. China Math., 57(1):9–30, 2014. Junyan Cao, Institut de Mathématiques de Jussieu, Université Pierre et Marie Curie, Case 247, 4 Place Jussieu, 75252 Paris Cedex, France E-mail address: [email protected] Andreas Höring, Laboratoire de Mathématiques J.A. Dieudonné, UMR 7351 CNRS, Université de Nice Sophia-Antipolis, 06108 Nice Cedex 02, France E-mail address: [email protected] 32
1408.6698	3	1408	2015-09-30T15:12:47	Approximation of sheaves on algebraic stacks	[ "math.AG" ]	Raynaud--Gruson characterized flat and pure morphisms between affine schemes in terms of projective modules. We give a similar characterization for non-affine morphisms. As an application, we show that every quasi-coherent sheaf is the union of its finitely generated quasi-coherent subsheaves on any quasi-compact and quasi-separated algebraic stack.	math.AG	math	APPROXIMATION OF SHEAVES ON ALGEBRAIC STACKS DAVID RYDH Abstract. Raynaud -- Gruson characterized flat and pure morphisms between affine schemes in terms of projective modules. We give a similar characterization for non-affine morphisms. As an application, we show that every quasi-coherent sheaf is the union of its finitely generated quasi-coherent subsheaves on any quasi-compact and quasi-separated algebraic stack. 1. Introduction It is well-known that on a noetherian scheme every quasi-coherent sheaf is the union of its coherent subsheaves [EGAIa, Cor. 9.4.9]. This is also true for noetherian algebraic stacks [LMB00, Prop. 15.4]. For a non-noetherian scheme or algebraic stack X, this question splits up into two questions. (i) Is every quasi-coherent OX -module the union of its quasi-coherent submodules of finite type? (ii) Is every quasi-coherent OX-module a directed colimit of finitely presented OX -modules? When these questions have positive answers, we say that X has the partial completeness property and completeness property respectively. The second property implies the first (take images). It is known that quasi-compact and quasi-separated schemes have the completeness property [EGAIb, §6.9]. In [Ryd15, Thm. A] it was shown that many stacks, including quasi-compact and quasi-separated algebraic spaces and Deligne -- Mumford stacks, have the completeness property. With current technology, this result only applies to relatively few algebraic stacks with infinite stabilizer groups. The main result of this paper settles the partial completeness property for every reasonable stack. Theorem. Let X be a quasi-compact and quasi-separated algebraic stack. Then every quasi-coherent OX -module is the union of its quasi-coherent sub- modules of finite type. An important application of the theorem is that when X in addition has affine stabilizer groups, then there exists a finitely presented filtration of X with strata that are global quotient stacks [HR14a, Prop. 2.6 (i)]. Date: 2015-05-07. 2010 Mathematics Subject Classification. Primary 14A20. Key words and phrases. Noetherian approximation, pure, projective, algebraic stacks. Supported by the Swedish Research Council grant no 2011-5599. 1 2 DAVID RYDH This is used to obtain a criterion for an algebraic stack to have finite coho- mological dimension [HR14a, Thm. 2.1] and to extend Tannaka duality to non-noetherian stacks [HR14b, Thm. 1.4]. The key idea in the proof of the main theorem is to use projective mod- ules instead of flat modules. The main lemma (6.2) on existence of minimal modules goes back to Serre [SGA3, Exp. VIB, 11.8, 11.10.1] in the context of coalgebras and comodules. Here projectivity cannot be replaced with flat- ness. The bulk of the paper extends this result to non-affine pure morphisms (Theorem 6.3). For this, we use a new characterization of pure morphisms between stacks in terms of projectivity (Theorem 5.3). This generalizes the characterization of affine pure morphisms due to Raynaud -- Gruson [RG71, Thm. I.3.3.5]. The main result naturally leads to the following conjectures. Conjecture A. If X is a quasi-compact and quasi-separated algebraic stack, then X has the completeness property. Conjecture B. If X is a quasi-compact and quasi-separated algebraic stack, then X has an approximation, that is, there exists a factorization X → X0 → Spec Z where X → X0 is affine and X0 is of finite presentation over Spec Z. The second conjecture implies the first conjecture. Our proof of the main theorem first reduces the question to the case when there is a pure pre- sentation. The conjectures can also be reduced to this seemingly simpler situation (see Remark 7.6). In Sections 2 -- 4, we recall and extend some notions from schemes to alge- braic stacks. This includes (1) locally free and locally projective modules, (2) assassins and schematically dominant morphisms, and (3) pure morphisms. In Section 5, we give a characterization of pure morphisms in terms of pro- jectivity (Theorem 5.3). In Section 6, we prove the existence of minimal subsheaves for pure morphisms. In Section 7, we prove the main theorem. In the last section, we give some applications to the main theorem. We follow the terminology of [SP] and do not impose any separation conditions on a general algebraic stack. An algebraic stack is quasi-separated if its diagonal is quasi-compact and quasi-separated, that is, if the diagonal and the double diagonal are quasi-compact. 1.1. Acknowledgments. It is my pleasure to acknowledge useful discus- sions with Jack Hall and useful comments from Martin Brandenburg and Matthieu Romagny. I would also like to express my gratitude to the referees for their many useful suggestions and corrections that improved the paper. 2. Locally free and locally projective modules In this section, we recall some standard results on infinitely generated projective modules due to Kaplansky, Bass and Raynaud -- Gruson. Definition (2.1). Let X be an algebraic stack. We say that a quasi- coherent sheaf F is locally free (resp. locally projective) if there exists a APPROXIMATION OF SHEAVES 3 jointly surjective family of flat morphisms pi : Spec Ai → X, locally of finite presentation, such that p∗ i F is free (resp. projective) for every i. We do not require that p∗ i F has finite rank in the definition of locally free. Note that the properties locally free and locally projective are stable under arbitrary pull-back and are local for the fppf-topology. We have the implications: locally free =⇒ locally projective =⇒ flat. If x ∈ X is a point, then we define the rank rkF (x) of F at x as the cardinality of a basis of the k-vector space ϕ∗F for any representa- tive ϕ : Spec k → X of x. Since flat morphisms that are locally of finite presentation are open, the rank of F is locally constant on X if F is locally free. The rank does not behave so well for flat modules that are not finitely generated. If A = Z and M = Q, then the rank of M is not upper semicon- tinuous. The rank of projective modules is more well-behaved. Lemma (2.2) (Kaplansky [Kap58]). If A is a local ring, then every projec- tive A-module is free. Thus, if X is a quasi-separated1 algebraic stack and F is a locally projec- tive OX-module, then (i) the rank of F is constant on irreducible components of X; and (ii) if X has a finite number of irreducible components (e.g., X noe- therian), then the rank is locally constant. Nevertheless, even if M is projective and has finite rank at every point, the rank need not be locally constant. Bass gives an example, due to Kaplansky, of a projective module of rank ≤ 1 such that the locus where the module has rank 0 is closed but not open [Bas63, p. 31, (2)]. We now give a similar example. Example (2.3). Let k be an algebraically closed field and let A = T (k[x]) be the absolutely flat ring associated to the polynomial ring k[x] [Oli68, Prop. 5]. Then Spec A is zero-dimensional and reduced and its underlying topological space is the one-point compactification of k with its discrete topology. For every λ ∈ k, the corresponding quotient A ։ κ(λ) = k is a locally free and finitely generated A-module, hence projective. The direct sum M = ⊕λ∈kκ(λ) is a projective A-module with rank 1 over the open subset k and rank 0 over its complement, which consists of a single point ξ. The discrete additive group G = (k, +) acts freely on Spec A and the quotient X = Spec A/G is an algebraic space consisting of two points {x, ξ} where x is open and ξ is closed. Note that X is not quasi-separated since the orbit of x is not quasi-compact. The module M descends to a locally projective OX -module F such that the rank over x is one and the rank over ξ is zero. The topological space X is irreducible and hence the rank is not constant over irreducible components in the usual sense. A flat module that has constant rank need not be so nice either as the following example shows. 1This condition is necessary with the naive notion of irreducible components, cf. Ex- ample (2.3). 4 DAVID RYDH Example (2.4). If M ⊆ Q is the Z-submodule generated by all p−1, for prime numbers p, then M is flat of constant rank 1 but neither projective nor finitely generated. Proposition (2.5). Let X be an algebraic stack and let F be a quasi- coherent sheaf on X. (i) If X is an affine scheme, then F is locally projective if and only if F is projective. (ii) If X is a noetherian affine scheme and ℵ ≥ ℵ0 is an infinite car- dinal, then F is projective with constant rank ℵ if and only if F is free of rank ℵ. (iii) If X is noetherian, then F is locally projective of finite rank if and only if F is finitely generated and locally free. (iv) If X is noetherian, then F is locally projective if and only if F is locally free. (v) If X is a noetherian scheme, then F is locally free if and only if F is Zariski-locally free. Proof. In each case, the "if" part is trivial. The necessity of the first con- dition follows from [RG71, I.3.1.4] (countable rank) or [RG71, II.2.5.1 and II.3.1.3] (general case). That conditions (ii) and (iii) are necessary is [Bas63, Cor. 3.2 & Prop. 4.2] respectively. Since the rank of a locally projective sheaf is locally constant on a noetherian stack, the necessity of conditions (iv) and (v) follow from (i), (ii) and (iii). (cid:3) Remark (2.6). Without the noetherian assumptions, statements (iii) and (iv) are false. If statement (ii) holds without the noetherian assumption, then so does (v). In particular, this would imply that on any stack X, a quasi- coherent sheaf F is locally free if and only if F is locally projective, has locally constant rank and is finitely generated over the open locus of finite rank. 3. Relative assassins and relative faithfulness In this section, we extend the notions of relative assassins [RG71, 3.2.2] and schematically dominant morphisms [EGAIV, 11.9 -- 11.10] from schemes to algebraic stacks. (3.1) Associated points -- There is a unique notion of associated points of coherent sheaves on locally noetherian algebraic stacks such that (i) it coincides with the usual one for schemes; and (ii) if f : X → Y is a flat morphism between locally noetherian stacks and F is a coherent OY -module, then f (AssX(f ∗F)) ⊆ AssY (F) with equality if f is surjective. The usual assassin satisfies (ii) for morphisms between schemes. more precisely we have that Indeed, (3.1.1) AssX(f ∗F) = [ AssXy (OXy ) for any flat morphism f : X → Y between locally noetherian schemes [EGAIV, Prop. 3.3.1]. We may thus simply define AssX (F) for a coherent sheaf F y∈AssY (F ) APPROXIMATION OF SHEAVES 5 on X as AssX(F) := p(AssU (p∗F)) where p : U → X is a presentation. One can also give a more intrinsic definition, cf. [Lie07, 2.2.6.3 -- 2.2.6.7]. We abbreviate Ass(X) = AssX (OX ). In particular, if f : X → Y is locally of finite type and ξ ∈ Y is a point, then we may define Ass(Xξ) ⊆ f −1(ξ) as the image of Ass(Xy) → X for any representative y : Spec k → Y of ξ. Definition (3.2) ([RG71, D´ef. 3.2.2]). Let f : X → Y be a morphism of algebraic stacks that is locally of finite type. The relative assassin Ass(X/Y ) is the subset Sy∈Y Ass(Xy) of X. Note that X and Y need not be noetherian in the definition above, but the finiteness condition ensures that the fibers are locally noetherian. If f is flat and X and Y are locally noetherian, then Ass(X) = Sy∈Ass(Y ) Ass(Xy) ⊆ Ass(X/Y ) by (3.1.1). The advantage of Ass(X/Y ) is that it behaves well with respect to any base change Y ′ → Y , whereas Ass(X) does not behave well with respect to non-flat base change, e.g., passage to a fiber. If p : X ′ → X is flat and locally of finite type, then p(Ass(X ′/Y )) ⊆ Ass(X/Y ) with equality if p is surjective; this follows from property (ii) above. Definition (3.3). Let f : X → Y be a morphism of algebraic stacks. We say that f is schematically dominant if OY → f∗OX is injective as a morphism of lisse-´etale sheaves. This agrees with the usual definition for schemes [EGAIV, 11.10.2] since that notion is stable under base change by flat morphisms that are locally of finite presentation [EGAIV, 11.10.5 (ii) b)]. It follows that our notion for algebraic stacks also is stable under base change by flat morphisms that are locally of finite presentation. When f is quasi-compact, the notion is stable under arbitrary flat base change [EGAIV, 11.10.5 (ii) a)]. If p : X ′ → X is another morphism and f ◦ p is schematically dominant, then so is f . If f and p are schematically dominant, then so is f ◦ p. In par- ticular, morphisms that are covering in the fppf topology are schematically dominant. Definition (3.4). Let S be an algebraic stack and let f : X → Y be a mor- phism of algebraic stacks over S. We say that f is S-universally schemati- cally dominant if f ′ : X ×S S′ → Y ×S S′ is schematically dominant for every morphism S′ → S. Proposition (3.5). Let S, X and Y be algebraic stacks and let f : X → Y and Y → S be flat morphisms that are locally of finite presentation. The following are equivalent. (i) The morphism f is S-universally schematically dominant. (ii) The image f (X) contains the relative assassin Ass(Y /S). Proof. Since f is open and faithfully flat onto its image, we may assume that f is an open immersion. As the question is fppf-local on Y and S we may assume that Y and S are affine schemes. The result is then [EGAIV, Prop. 11.10.10] (or [RG71, Cor. 3.2.6]). (cid:3) 6 DAVID RYDH Definition (3.6). Let f : X → Y and g : Y → S be morphisms, locally of finite presentation, between algebraic stacks such that g is flat. We say that f is S-faithfully flat if f is flat and the equivalent conditions of Proposi- tion (3.5) hold. This terminology is explained by the following lemma. Lemma (3.7). Let f : X → Y and π : Y → S be morphisms of algebraic stacks. Assume that f is S-universally schematically dominant. Given F ∈ QCoh(Y ) and G ∈ QCoh(S), we have that (i) the unit map ηπ∗G : π∗G → f∗f ∗π∗G is injective; and (ii) a morphism θ : F → π∗G is zero if and only if f ∗θ is zero. Proof. Consider S′ = Spec(OS ⊕ G), where G is square-zero, and let X ′ = X ×S S′ and Y ′ = Y ×S S′. Then f ′ : X ′ → Y ′ is schematically dominant, that is, OY ⊕ π∗G → f∗(OX ⊕ f ∗π∗G) is injective. It follows that ηπ∗G is injective. If θ is zero, then so is f ∗θ. Conversely, if f ∗θ is zero, then so is ηπ∗G ◦ θ = (cid:3) (f∗f ∗θ) ◦ ηF . It follows that θ is zero since ηπ∗G is injective. Lemma (3.8). Let f : X → Y and π : Y → S be flat morphisms, that are locally of finite presentation, between algebraic stacks. Let F0 ⊆ F be quasi-coherent OS-modules and let G0 ⊆ π∗F be a quasi-coherent OY - submodule. Assume that f is S-faithfully flat. Then G0 ⊆ π∗F0 if and only if f ∗G0 ⊆ f ∗π∗F0. Proof. Let F ′ = F/F0. Consider the map θ : G0 ֒→ π∗F ։ π∗F ′. Then G0 ⊆ π∗F0 if and only if θ = 0 and f ∗G0 ⊆ f ∗π∗F0 if and only if f ∗θ = 0. Thus, the result follows from the previous lemma. (cid:3) 4. Pure morphisms of algebraic stacks We begin by recalling the definition of pure morphisms of schemes [RG71, D´ef. 3.3.3]. Definition (4.1). Let f : X → S be a morphism of schemes, locally of finite type. Let s ∈ S be a point and let (cid:0)eS,es(cid:1) → (S, s) be the henselization and eX = X ×S eS. We say that f is (i) pure along Xs if for every point s1 ∈ eS, every associated point (ii) pure if f is pure along Xs for every s ∈ S; and (iii) universally pure, if f ′ : X ×S S′ → S′ is pure for every morphism x1 ∈ Ass( eXs1) is the generization of a point in Xs; S′ → S. (4.2) Examples -- The two key examples of pure morphisms are [RG71, Ex. I.3.3.4]: (i) proper morphisms, and (ii) faithfully flat morphisms, locally of finite type, with fibers that are geometrically irreducible without embedded components. (4.3) Base change: descent -- If S′ → S is faithfully flat and f ′ is pure, then f is pure. Indeed, for every s′ ∈ S′ with image s ∈ S, the morphism APPROXIMATION OF SHEAVES 7 between henselizations (eS′, s′) → (eS, s) is surjective. If x1 ∈ Ass( eXs1 ), then there exists x′ ) above x1 (3.1, (ii)) and, by purity, a specialization x′ ∈ X ′ s′. Its image x ∈ Xs, is a specialization of x1. 1 ∈ Ass( eX ′ s′ 1 (4.4) Base change: stability -- If f is flat, pure and of finite presentation, then f is universally pure [RG71, 3.3.7]. Also, every pure morphism of finite presentation is universally pure when S is locally noetherian [SP, 05J8] but not for general S [SP, 05JJ]. (4.5) Composition -- Let f : X → Y and g : Y → S be morphisms of schemes, locally of finite type. If f and g are pure, then g ◦ f need not be pure, e.g., the composition Spec(k[x, y]/xy − 1) ֒→ Spec k[x, y] → Spec k[x] is not pure. On the other hand, if f is flat and pure and g is pure, then Xy for every y ∈ Ys since the henselization of Y at any point of Ys factors g ◦ f is pure. Indeed, the map f : eX = X ×S eS → eY = Y ×S eS is pure along through eY . Moreover, since f is flat, we have that f (Ass( eXs1)) ⊆ Ass(eYs1) for every s1 ∈ eS. Also, if f is faithfully flat and g ◦ f is pure, then g is pure. Indeed, for every point s1 ∈ eS, we have that f (Ass( eXs1)) = Ass(eYs1). To extend purity to morphisms of stacks, we give a slightly different def- inition. Definition (4.6). Let f : X → S be a morphism between algebraic stacks that is quasi-separated and locally of finite type. When S is quasi-separated, we say that f is weakly closed if f (Z) is closed for every closed irreducible subset Z ⊂ X, such that the generic point of Z is associated in its fiber. We say that f is universally weakly closed, if f ′ : X ×S S′ → S′ is weakly closed for every morphism S′ → S where S′ is quasi-separated. The remarks in (4.2), (4.3) and (4.5) hold for "pure" replaced by "weakly closed". For Remark (4.3), note that f is weakly closed if and only if f ({z}) is stable under specialization for every z ∈ Ass(X/S), and this can be checked flat-locally on S. The analogue of Remark (4.4) is false, which is not surprising: the good notion is universally weakly closed for which we have the following valuative criterion. Proposition (4.7). Let f : X → S be a quasi-separated morphism, locally of finite type, between algebraic stacks. Then the following are equivalent: (i) f is universally weakly closed; (ii) for every valuation ring V and morphism Spec V → S, the base change X ×S Spec V → Spec V is weakly closed; and (iii) for every valuation ring V , morphism Spec V → S, and associated point z in the generic fiber X ×S Spec K(V ), the closure of z in X ×S Spec V surjects onto Spec V . If f is a morphism of schemes, then this is equivalent to: (i′) f is universally pure. Proof. Clearly, (i) =⇒ (ii) =⇒ (iii). If f is a morphism of schemes, then trivially (i) =⇒ (i′) and we note that (i′) =⇒ (ii) since it is enough to verify (ii) for henselian valuation rings. 8 DAVID RYDH To see that (iii) =⇒ (i) it is enough to prove that f is weakly closed. Let z ∈ X be a point that is associated in its fiber and let Z = {z}. It is enough to prove that f (Z) = {f (z)}. This can be verified after the base change S′ = Spec V → S for every valuation ring V and every dominant morphism Spec V → {f (z)}. Then f (Z) = Spec V by (iii) and the result follows. (cid:3) Definition (4.8). Let f : X → Y be a flat morphism of finite presentation between algebraic stacks. We say that f is pure if it is universally weakly closed. This definition coincides with the usual definition for flat morphisms of schemes by (4.4). It also coincides with the definition of pure in [Rom11, B.1]. The following lemma, which is a direct transcription of an argument in [RG71, proof of Prop. 3.3.6], shows that a flat morphism X → S of finite presentation is weakly closed if and only if the map Ass(X/S) → S is closed under specializations, i.e., if subsets closed under specialization in Ass(X/S) maps to subsets closed under specialization in S. Lemma (4.9). Let S be a scheme and let X be an algebraic stack that is flat and of finite presentation over S. Let s, s1 ∈ S and x1 ∈ Ass(Xs1). If Xs ∩ {x1} 6= ∅, then Ass(Xs) ∩ {x1} 6= ∅. Proof. We may assume that S = Spec A is affine. Pick a smooth presen- tation p : U = Spec B → X. If Xs ∩ {x1} 6= ∅, then there exists a point u1 ∈ U above x1 such that Us ∩ {u1} 6= ∅. We may assume that u1 is maximal in p−1(x1) and then u1 ∈ Ass(Us1). Since p(Ass(Us)) = Ass(Xs), it is enough to prove that Ass(Us) ∩ {u1} 6= ∅. Let u ∈ Us ∩ {u1} and let Σ ⊆ OU,u be the set of elements whose images in OU,u ⊗ κ(s) are non-zero divisors. Then OU,u → Σ−1OU,u is A-universally injective and Σ−1OU,u is a semi-local ring whose maximal ideals are associated points of Us [RG71, 3.2.5]. In particular, the morphism OU,u ⊗ κ(s1) → (Σ−1OU,u) ⊗ κ(s1) is injective. Since u1 is associated in Spec(OU,u ⊗ κ(s1)), this means that u1 ∈ Spec(Σ−1OU,u ⊗ κ(s1)); hence u1 is a generization of an associated point u0 of Us. (cid:3) 5. Homological projectivity The main theorem of [RG71, §I.3] is the following relation between purity and projectivity for affine morphisms. Theorem (5.1) (Raynaud -- Gruson). Let f : X → Y be an affine finitely presented morphism of schemes. The following are equivalent: (i) f is flat and pure; (ii) f∗OX is locally projective; and (iii) f∗OX is locally free. Proof. The equivalence between (i) and (ii) is [RG71, Thm. I.3.3.5]. The equivalence between (ii) and (iii) is [RG71, Cor. I.3.3.12]. Note that if Y is noetherian, then the latter equivalence follows directly from Propo- sition (2.5) (iv). The non-noetherian case follows from the noetherian case APPROXIMATION OF SHEAVES 9 using the equivalence between (i) and (ii) and using that pure morphisms behave well under approximation [RG71, Cor. I.3.3.10]. (cid:3) Local projectivity of f∗OX is not local on X. To obtain a non-affine analogue of the theorem above, we introduce the following definition. Definition (5.2). Let f : X → Y be a flat morphism of finite presentation between algebraic stacks. We say that f is homologically projective (resp. strongly homologically projective) if there exists (i) an fppf-covering {Spec(Ai) → Y }; and (ii) flat morphisms qi : Spec(Bi) → X ×Y Spec(Ai), locally of finite presentation; such that for every i (a) the composition Spec(Bi) qi−→ X ×Y Spec(Ai) → Spec(Ai) makes Bi into a projective Ai-module; and (b) qi is Spec(Ai)-faithfully flat (resp. faithfully flat), cf. Definition (3.6). Here "homological" is to indicate that projective is interpreted as in homo- logical algebra and not as in algebraic geometry. It should not be confused with the notion of cohomologically projective morphisms in [Alp13, 3.18]. By definition, the notion of (strong) homological projectivity is stable under base change and fppf-local on the target. If p : X ′ → X is faithfully flat and locally of finite presentation and f ◦ p is (strongly) homologically projective, then f is (strongly) homologically projective but the converse does not hold. It is, a priori, not clear whether the composition of two (strongly) homologically projective morphisms is (strongly) homologically projective. Recall that X has the resolution property if every quasi-coherent sheaf of finite type on X admits a surjection from a vector bundle. Theorem (5.3). Let f : X → Y be a morphism of algebraic stacks that is flat and of finite presentation. Consider the following conditions: (i) f is affine and f∗OX is locally projective; (ii) f is strongly homologically projective; (iii) f is homologically projective; and (iv) f is pure. Then (i) =⇒ (ii) =⇒ (iii) ⇐⇒ (iv). If f is affine, then all four conditions are equivalent. If X has the resolution property fppf-locally on Y (e.g., if f is quasi-affine), then (ii) ⇐⇒ (iii). Proof. From the definitions, it follows that (i) =⇒ (ii) =⇒ (iii). To prove that (iii) =⇒ (iv), we may assume that Y = Spec A and that there is a Y -faithfully flat and finitely presented morphism U = Spec B → X such that B is a projective A-module. By Theorem (5.1), we have that U → Y is pure. Since the image of U contains Ass(X/Y ), it follows that X → Y is pure. When f is affine, (iv) =⇒ (i) by Theorem (5.1). For (iv) =⇒ (iii), suppose that f is pure. As before we may assume that Y is affine. Pick a smooth presentation U = Spec B → X. Let y ∈ Y be a 10 DAVID RYDH point. Then, by [RG71, Prop. 3.3.2], there exists a commutative diagram U ′ Y ′ U Y and a point y′ ∈ Y ′ above y such that • U ′ → U and Y ′ → Y are ´etale, and κ(y) = κ(y′); • U ′ = Spec B′ and Y ′ = Spec A′ are affine and B′ is a projective A′-module; and • the image of U ′ → U contains Ass(Uy). In particular, the image of U ′ → U → X contains Ass(Xy). After replacing X, Y and U by their pull-backs along the base change Y ′ → Y , we may assume that Y ′ = Y . We now claim that the image of U ′ → U → X contains Ass(Xy1) for every generization y1 of y. To see this, let x1 ∈ Ass(Xy1 ). By the definition of purity, there exists a point x ∈ Xy ∩ {x1}. By Lemma (4.9), there exists a point x0 ∈ Ass(Xy) ∩ {x1}. Since x0 is in the image of U ′, so is its generization x1. By [RG71, Lem. 3.3.9], there is then an open neighborhood y ∈ V ⊆ Y such that the image of U ′ → U → X contains Ass(Xy1) for every y1 ∈ V . This means that U ′ → U → X is Y -faithfully flat over V , that is, X → Y is homologically projective over V . As the question is local on Y , it follows that X → Y is homologically projective. Under the additional assumption on X, we will we prove that (iv) =⇒ (ii). For this, we may work locally on Y and assume that Y = Spec A is affine and that X has the resolution property. Then X = [U/GLn] for some quasi- affine scheme U [Tot04, Gro13]. By Jouanolou's trick, there is an affine vector bundle torsor E → U [Jou73, Lem. 1.5] (also see [Wei89, 4.3 -- 4.4]). Since E → X is flat with geometrically integral fibers, hence flat and pure, it follows that E → X → Y is pure (4.5). Since E = Spec B is affine, we have that B is A-projective; thus, f is strongly homologically projective. (cid:3) 6. Existence of minimal subsheaves Let f : X → Y be a faithfully flat morphism between quasi-compact al- gebraic stacks and let F ∈ QCoh(Y ). Assume that G0 ⊆ G := f ∗F is a quasi-coherent subsheaf of finite type. If F is the union of its quasi-coherent subsheaves Fλ of finite type, then, for sufficiently large λ, we have that G0 ⊆ f ∗Fλ. Conversely, if G = f ∗F is the union of its quasi-coherent subsheaves Gλ of finite type and for every Gλ there exists Fλ ⊆ F of finite type such that Gλ ⊆ f ∗Fλ, then F is the union of its subsheaves of finite type. We will see that, under suitable hypotheses, for every Gλ of finite type as above there is a minimal Fλ as above and it is of finite type. This is, however, not always the case: Example (6.1). Let A be a discrete valuation ring with fraction field K and uniformizing parameter t. Let B = A × K, which is a faithfully flat APPROXIMATION OF SHEAVES 11 A-algebra. Let M = A and consider the submodule N0 = (0 × K) ⊆ M ⊗A B = B. For every non-trivial ideal Mn = (tn) ⊆ A = M , we then have that N0 ⊆ Mn ⊗A B = (tn) × K. But the intersection is T Mn = 0 and N0 * (T Mn) ⊗A B = 0. Hence, there is no minimal submodule M ′ of M such that N0 ⊆ M ′ ⊗A B. The problem in Example (6.1) is that infinite intersections do not com- mute with flat pull-back. This does not happen if we replace flatness with projectivity. Lemma (6.2) (Serre). Let A be a ring and let B be an A-algebra which is projective as an A-module. Let M be an A-module and let N0 ⊆ M ⊗A B be a B-submodule. Then there is a unique minimal A-submodule M0 ⊆ M such that N0 ⊆ M0 ⊗A B. Moreover, 0 is the minimal A′-submodule of M ′ such that N ′ 0 ⊆ M ′ (i) if N0 is of finite type, then so is M0; and (ii) if A′ is an A-algebra and we let B′ = B⊗AA′, M ′ = M ⊗AA′, M ′ im(M0 ⊗A A′ → M ′) and N ′ M ′ 0 := 0 = im(N0 ⊗B B′ → M ′ ⊗A′ B′), then 0 ⊗A′ B′. Proof. Choose a free A-module F such that B is a direct summand of F and pick a basis {ei} of F . Let M0 ⊆ M be an A-submodule. Then M0 ⊗A B ⊆ M0⊗AF ⊆ M ⊗AF and M ⊗AB ⊆ M ⊗AF . Let x ∈ N0 be an element. Then x = Pi xi ⊗ ei in M ⊗A F , and, using the retraction F → B, we may also write x = Pi xi ⊗ bi in M ⊗A B. Thus x ∈ M0 ⊗A B if and only if xi ∈ M0 for every i. It follows that the minimal submodule M0 is the submodule generated by the xi's when x ranges over a set of generators of N0. The remaining claims follows immediately from the construction of M0. (cid:3) Using purity, we give the following global version. Theorem (6.3). Let f : X → Y be a flat morphism of finite presentation between algebraic stacks. Assume that f is pure. Let F ∈ QCoh(Y ) and let G0 ⊆ G := f ∗F be a quasi-coherent submodule. Then there is a unique minimal quasi-coherent submodule F0 ⊆ F such that G0 ⊆ f ∗F0. Moreover, (i) if G0 is of finite type, then so is F0; and (ii) if f ′ : X ′ → Y ′ is the base change of f along a morphism g : Y ′ → Y , 0 of g∗F0 → g∗F is the minimal quasi-coherent then the image F ′ submodule such that f ′∗F ′ 0 contains the image of g′∗G0 → g′∗G. Proof. By Theorem (5.3), f is homologically projective. By fppf descent, it is enough to prove the statement after replacing Y with an fppf cover. We may thus assume that Y = Spec A and that there exists a Y -faithfully flat morphism q : X ′ = Spec B → X of finite presentation such that B is a projective A-module. If F0 ⊆ F is a submodule, then G0 ⊆ f ∗F0 if and only if q∗G0 ⊆ q∗f ∗F0 (Lemma 3.8). We may thus replace X with X ′ and assume that X and Y are affine. The theorem is then Lemma (6.2). (cid:3) 7. Approximation of quasi-coherent sheaves Let X be a quasi-compact and quasi-separated algebraic stack. We recall that X has the completeness property if every quasi-coherent OX -module is a directed colimit of finitely presented OX -modules and that X has the partial 12 DAVID RYDH completeness property if every quasi-coherent OX-module is the union of its finitely generated quasi-coherent submodules. In the terminology of [Ryd15, §4], these two conditions are the conditions (C1) and (C2) for the category QCoh(X) and they imply the corresponding facts for quasi-coherent OX- algebras. We also make the following definition that extends [Ryd15, Def. 4.7]. Definition (7.1). An algebraic stack X is semi-noetherian (resp. pseudo- noetherian) if it is quasi-compact, quasi-separated and X ′ has the partial completeness property (resp. completeness property) for every finitely pre- sented morphism X ′ → X of algebraic stacks. Every pseudo-noetherian algebraic stack is semi-noetherian. Noether- ian algebraic stacks, quasi-compact and quasi-separated schemes, algebraic spaces and Deligne -- Mumford stacks are examples of pseudo-noetherian al- gebraic stacks [Ryd15, Thm. A]. Proposition (7.2). Let f : X → Y be a faithfully flat and pure morphism of finite presentation between quasi-compact and quasi-separated algebraic stacks. If X has the partial completeness property, then so has Y . In par- ticular, X is semi-noetherian if and only if Y is semi-noetherian. Proof. Let F ∈ QCoh(Y ) and write f ∗F as a union S Gλ of quasi-coherent submodules of finite type. By Theorem (6.3), for every λ there exists a minimal quasi-coherent subsheaf Fλ ⊆ F of finite type such that Gλ ⊆ f ∗Fλ. If we let F ′ = S Fλ ⊆ F, then f ∗F ′ contains every Gλ. It follows that f ∗F ′ = f ∗F and thus F ′ = F since f is faithfully flat. We conclude that Y has the partial completeness property. (cid:3) Proposition (7.3). Let X be an algebraic stack and let p : X ′ → X be ´etale, representable, surjective and of finite presentation. Then X is semi- noetherian if and only if X ′ is semi-noetherian. Proof. This is proven exactly as [Ryd15, Prop. 4.11]: ´etale d´evissage [Ryd11, Thm. D] is used to reduce the question to where p is either finite, surjective and ´etale or an ´etale neighborhood. These two cases follow from simplified versions of [Ryd15, Lem. 4.9 and 4.10] where "completeness property" is replaced with "partial completeness property". (cid:3) The main theorem will follow from the previous two propositions together with the following factorization result. For our main theorem we will only apply it to a smooth and representable morphism (the presentation of a stack). Theorem (7.4) ([LMB00, 6.8],[Rom11]). Let f : X → Y be a faithfully flat morphism of finite presentation with geometrically reduced fibers (e.g., f smooth) between algebraic stacks. Then there exists an open substack U ⊆ X and a factorization f U = h ◦ g such that (i) g and h are faithfully flat of finite presentation; (ii) h is representable and ´etale; and (iii) g has geometrically integral fibers. APPROXIMATION OF SHEAVES 13 In particular, g is pure. If f is smooth, then g is smooth and we can take U = X. Proof. First assume that f is smooth. Consider the connected factoriza- tion X → π0(X/Y ) → Y , which is described for morphisms of schemes in [LMB00, 6.8] and for an algebraic stack over an algebraic space in [Rom11, Thm. 2.5.2]. Since the construction commutes with base change, it general- izes to our situation as well. In this factorization g : X → π0(X/Y ) is smooth with geometrically connected fibers and h : π0(X/Y ) → Y is ´etale, repre- sentable and of finite presentation, but not necessarily separated [Rom11, Thm. 2.5.2 (i), (ii)]. In the general case we use the functor of irreducible components of Ro- magny. The unicomponent locus U ⊆ X is the subset of points that belong to exactly one irreducible component of their fibers. It is open and quasi- compact and there is a factorization U → Irr(X/Y ) → Y where the first morphism has geometrically integral fibers and the second is surjective, ´etale, representable and of finite presentation [Rom11, Thm. 2.5.2 (i), (iii)]. (cid:3) We now obtain the following equivalent form of the main theorem. Theorem (7.5). Let X be a quasi-compact and quasi-separated algebraic stack. Then X is semi-noetherian. Proof. Pick a smooth presentation Spec B → X. Theorem (7.4) gives a factorization Spec B → W → X where Spec B → W is smooth, surjective and pure and W → X is ´etale, surjective and of finite presentation. The result now follows from Propositions (7.2) and (7.3). (cid:3) Remark (7.6). To answer Conjectures A and B, we may argue as in the proof of Theorem (7.5) using [Ryd15, Prop. 4.11 and Lem. 7.9]. This re- duces the situation to where X has a smooth presentation U → X with geometrically connected fibers. The author hopes that the purity of U → X and its characterization as homological projectivity can be used to settle the conjectures. We conclude with some applications of the main theorem. 8. Applications Theorem (8.1) (Zariski's main theorem). Let f : X → Y be a morphism between quasi-compact and quasi-separated algebraic stacks. Then the fol- lowing are equivalent: (i) f is representable, separated and quasi-finite; and (ii) there is a factorization f = f ◦ j where j is a quasi-compact open immersion and f is finite. Proof. This follows from [LMB00, Thm. 16.5 (ii)] and the main theorem (taking into account that the finite presentation assumption of loc. cit. can be avoided by replacing the reference to [EGA] IV 8.12.6 with [EGA] IV 18.12.13). An essentially identical proof is given in [Ryd15, Thm. 8.6 (ii)] (use the partial completeness property instead of the completeness prop- erty). (cid:3) 14 DAVID RYDH Proposition (8.2). Let X be a quasi-compact and quasi-separated algebraic stack and let U ⊆ X be a quasi-compact open substack. Then there exists a closed immersion Z ֒→ X of finite presentation such that U = X r Z. Proof. Let I ⊆ OX be the quasi-coherent sheaf of ideals defining Zred = (X r U )red. Write I = S Iλ as a union of quasi-coherent ideals of finite type. If Zλ denotes the finitely presented closed substack corresponding to Iλ, then ∩Zλ = Zred. Since U is quasi-compact it follows that Zλ = Zred for all sufficiently large λ. We may take Z = Zλ for any such λ. (cid:3) As a third application, we have the existence of flattening stratifications for finitely presented morphisms. Theorem (8.3). Let X be a quasi-compact and quasi-separated algebraic stack and let W → X be a morphism of finite presentation. Then there exists a sequence of finitely presented closed substacks ∅ = X0 ֒→ X1 ֒→ . . . ֒→ Xn such that Xn = X and the restriction of W → X to Xk r Xk−1 is flat for every k = 1, 2, . . . , n. Proof. The result is well-known when X is noetherian: let Xn = Xred; pick a smooth presentation p : Spec(A) → Xn; choose a non-empty open subscheme V ⊆ Spec(A) over which W is flat (generic flatness); let Xn−1 = (X r p(V ))red. The result now follows by noetherian induction. If X is affine, the result follows by standard limit methods: there is a noetherian affine scheme X0, a morphism X → X0 and a morphism W0 → X0 of finite presentation that pull-backs to W → X. The pull-back of a solution to the problem for W0 → X0 gives a solution for W → X. 0 ֒→ X ′ In the general case, we pick a smooth presentation p : X ′ = Spec(A) → X and choose a filtration X ′ n that solves the problem over X ′. We will prove that X has a filtration of length n that solves the problem. Set-theoretically, we will have Xk = X r p(X ′ r X ′ k). If n = 0, the problem is trivial. By induction on n, we may assume that there exists a filtration of length n − 1 on every closed substack Q ֒→ X such that p−1(Q) ⊆ X ′ 1 ֒→ . . . ֒→ X ′ n−1. The subset p(X ′ r X ′ finitely presented closed substack Z ֒→ X such that X r Z = p(X ′ r X ′ (Proposition 8.2). n−1) is open and quasi-compact, hence there is a n−1) Since p is smooth, we have that p−1(Xred) = X ′ red and hence p−1(Xred) ֒→ X ′ factors through X ′ n. Writing the nilradical of OX as a union of quasi- coherent ideals of finite type, we may write the nil-immersion Xred ֒→ X as an intersection of finitely presented nil-immersions Xλ ֒→ X. For sufficiently large λ, we have that p−1(Xλ) ֒→ X ′ factors through X ′ n. Then W → X is flat over Xλ r Z for such λ since p−1(Xλ) r X ′ n−1 → Xλ r Z is smooth and surjective. We let Xn = Xλ and Q = Z ∩ Xλ. Then, by induction there is a filtration X0 ֒→ X1 ֒→ . . . ֒→ Xn−1 ֒→ Q with Xn−1 = Q such that W → X is flat over the strata. The result follows. (cid:3) As a fourth application, we have the existence of stratifications into gerbes for stacks with finitely presented inertia. APPROXIMATION OF SHEAVES 15 Corollary (8.4). Let X be a quasi-compact and quasi-separated algebraic stack with inertia of finite presentation. Then there exists a sequence of finitely presented closed substacks ∅ = X0 ֒→ X1 ֒→ . . . ֒→ Xn such that Xn = X and Xk r Xk−1 is an fppf gerbe over an affine scheme for every k = 1, 2, . . . , n. Proof. Apply Theorem (8.3) on IX → X to obtain a stratification into fppf gerbes over quasi-compact and quasi-separated algebraic spaces. By Propo- sition (8.2), it remains to prove that a quasi-compact and quasi-separated algebraic space S can be stratified into affine schemes. Pick an approxima- tion S → S0 → Spec Z, that is, an algebraic space S0 of finite presentation over Spec Z and an affine morphism S → S0 [Ryd15, Thm. D]. It is enough to stratify S0 into affine schemes. This can be done by noetherian induction since S0 has an open subspace that is a scheme. (cid:3) For a general quasi-compact and quasi-separated algebraic stack, the in- In this case, it is not always possible to find In fact, sometimes ertia is only of finite type. finitely presented stratifications as in Corollary (8.4). even an infinite number of strata is required [SP, 06RE]. As a final application, we see that two different definitions of projectiv- ity and quasi-projectivity over algebraic stacks are equivalent. Our main definition is analogous to that for schemes in EGA [EGAII, D´efs. 5.3.1 and 5.5.2]. Definition (8.5). A representable morphism f : X → Y of algebraic stacks is (i) quasi-projective if f is of finite type and there exists an f -ample invertible OX -module; and (ii) projective if X is Y -isomorphic to a closed substack of a projective bundle PY (E) where E is a quasi-coherent OY -module of finite type. Note that being f -ample is an fppf-local property on the target [EGAIV, Cor. 2.7.2] and hence makes sense for representable morphisms. Similarly, projective bundles is a local construction on the base. Theorem (8.6) (cf. [EGAII, Prop. 5.3.2 & Thm. 5.5.3]). Let Y be a quasi- compact and quasi-separated algebraic stack and let f : X → Y be a repre- sentable morphism. Then (i) f is quasi-projective if and only if there exists a quasi-compact im- mersion X ֒→ PY (E) over Y , where E is a quasi-coherent OY - module of finite type. (ii) f is projective if and only if it is proper and quasi-projective. Proof. If i : X → PY (E) is a quasi-compact immersion, then i∗OP(E)(1) is very ample and f is quasi-projective. Conversely, assume that f is quasi- projective and let L be an f -ample invertible sheaf. There is a natural map σ : f ∗f∗L → L and when this map is surjective, we have an induced mor- phism rL,σ : X → P(f∗L). Choose a presentation g : Y ′ → Y . After replac- ing L with a sufficiently large power, the invertible sheaf g′∗(L) becomes very ample which implies that σ is surjective and rL,σ is an immersion [EGAII, Prop. 4.4.4]. Write f∗L as the union of its finitely generated submodules Eλ. Then for sufficiently large λ, the map σλ : f ∗Eλ → L is surjective and 16 DAVID RYDH the induced morphism rL,σλ : X → P(Eλ) is an immersion [EGAII, pf. of Prop. 4.4.1 (ii)]. If f is projective, then f is quasi-projective (as before) and proper (check locally on Y ). Conversely, if f is quasi-projective and proper, then by (i), there is an immersion X ֒→ PY (E) which is closed since f is proper. (cid:3) References [Alp13] Jarod Alper, Good moduli spaces for Artin stacks, Ann. Inst. Fourier (Grenoble) 63 (2013), no. 6, 2349 -- 2402. [Bas63] Hyman Bass, Big projective modules are free, Illinois J. Math. 7 (1963), no. 1, 24 -- 31. [EGAIa] A. Grothendieck, ´El´ements de g´eom´etrie alg´ebrique. I. Le langage des sch´emas, Inst. Hautes ´Etudes Sci. Publ. Math. (1960), no. 4, 228. [EGAIb] A. Grothendieck, ´El´ements de g´eom´etrie alg´ebrique. I. Le langage des sch´emas, second ed., Die Grundlehren der mathematischen Wissenschaften in Einzel- darstellungen, vol. 166, Springer-Verlag, Berlin, 1971. [EGAII] A. Grothendieck, ´El´ements de g´eom´etrie alg´ebrique. II. ´Etude globale ´el´emen- taire de quelques classes de morphismes, Inst. Hautes ´Etudes Sci. Publ. Math. (1961), no. 8, 222. [EGAIV] A. Grothendieck, ´El´ements de g´eom´etrie alg´ebrique. IV. ´Etude locale des sch´e- mas et des morphismes de sch´emas, Inst. Hautes ´Etudes Sci. Publ. Math. (1964 -- 67), nos. 20, 24, 28, 32. [Gro13] Philipp Gross, Tensor generators on schemes and stacks, Preprint, Jun 2013, arXiv:1306.5418. [HR14a] Jack Hall and David Rydh, Algebraic groups and compact generation of their derived categories of representations, Indiana Univ. Math. J. (2014), arXiv:1405.1890v2, accepted for publication. [HR14b] Jack Hall and David Rydh, Coherent Tannaka duality and algebraicity of Hom- [Jou73] [Kap58] stacks, Preprint, May 2014, arXiv:1405.7680v2. J. P. Jouanolou, Une suite exacte de Mayer-Vietoris en K-th´eorie alg´ebrique, Algebraic K-theory, I: Higher K-theories (Proc. Conf., Battelle Memorial Inst., Seattle, Wash., 1972), Springer, Berlin, 1973, pp. 293 -- 316. Lecture Notes in Math., Vol. 341. Irving Kaplansky, Projective modules, Ann. of Math (2) 68 (1958), no. 2, 372 -- 377. [Lie07] Max Lieblich, Moduli of twisted sheaves, Duke Math. J. 138 (2007), no. 1, 23 -- 118. [LMB00] G´erard Laumon and Laurent Moret-Bailly, Champs alg´ebriques, Springer-Verlag, [Oli68] Berlin, 2000. Jean-Pierre Olivier, Anneaux absolument plats universels et ´epimorphismes `a buts r´eduit, S´eminaire d'Alg`ebre Commutative dirig´e par Pierre Samuel: 1967 -- 1968. Les ´epimorphismes d'anneaux, Exp. No. 6, Secr´etariat math´ematique, Paris, 1968, p. 12. [RG71] Michel Raynaud and Laurent Gruson, Crit`eres de platitude et de projectivit´e. Techniques de "platification" d'un module, Invent. Math. 13 (1971), no. 1, 1 -- 89. irr´eductibles en familles, [Rom11] Matthieu Romagny, Composantes connexes et Manuscripta Math. 136 (2011), no. 1-2, 1 -- 32. [Ryd11] David Rydh, ´Etale d´evissage, descent and pushouts of stacks, J. Algebra 331 (2011), 194 -- 223. [Ryd15] David Rydh, Noetherian approximation of algebraic spaces and stacks, J. Algebra 422 (2015), 105 -- 147, arXiv:0904.0227v4. [SGA3] M. Demazure and A. Grothendieck (eds.), Sch´emas en groupes, Springer-Verlag, Berlin, 1970, S´eminaire de G´eom´etrie Alg´ebrique du Bois Marie 1962/64 (SGA 3). Dirig´e par M. Demazure et A. Grothendieck. Lecture Notes in Mathematics, Vol. 151 -- 153. APPROXIMATION OF SHEAVES 17 [SP] The http://stacks.math.columbia.edu/. Project Stacks Authors, Stacks project, [Tot04] Burt Totaro, The resolution property for schemes and stacks, J. Reine Angew. Math. 577 (2004), 1 -- 22. [Wei89] Charles A. Weibel, Homotopy algebraic K-theory, Algebraic K-theory and al- gebraic number theory (Honolulu, HI, 1987), Contemp. Math., vol. 83, Amer. Math. Soc., Providence, RI, 1989, pp. 461 -- 488. KTH Royal Institute of Technology, Department of Mathematics, SE-100 44 Stockholm, Sweden E-mail address: [email protected]
1208.6517	1	1208	2012-08-31T15:09:46	Glicci ideals	[ "math.AG" ]	A central problem in liaison theory is to decide whether every arithmetically Cohen-Macaulay subscheme of projective $n$-space can be linked by a finite number of arithmetically Gorenstein schemes to a complete intersection. We show that this can be indeed achieved if the given scheme is also generically Gorenstein and we allow the links to take place in an $(n+1)$-dimensional projective space. For example, this result applies to all reduced arithmetically Cohen-Macaulay subschemes. We also show that every union of fat points in projective 3-space can be linked in the same space to a union of simple points in finitely many steps, and hence to a complete intersection in projective 4-space.	math.AG	math	GLICCI IDEALS JUAN MIGLIORE∗ AND UWE NAGEL+ Abstract. A central problem in liaison theory is to decide whether every arithmetically Cohen-Macaulay subscheme of projective n-space can be linked by a finite number of arithmetically Gorenstein schemes to a complete intersection. We show that this can be indeed achieved if the given scheme is also generically Gorenstein and we allow the links to take place in an (n + 1)-dimensional projective space. For example, this result applies to all reduced arithmetically Cohen-Macaulay subschemes. We also show that every union of fat points in projective 3-space can be linked in the same space to a union of simple points in finitely many steps, and hence to a complete intersection in projective 4-space. 2 1 0 2 g u A 1 3 ] . G A h t a m [ 1 v 7 1 5 6 . 8 0 2 1 : v i X r a 1. Introduction A central problem is liaison theory has always been to describe the ideals/subschemes that are linked to a complete intersection in a finite number of steps. When the links are all by complete intersections (CI-liaison) these are named licci ideals, whereas when the links are Gorenstein (G-liaison) they are dubbed glicci ideals. It is well-known that each glicci (and in particular licci) ideal is Cohen-Macaulay. However, the converse is false for CI-liaison. There are many arithmetically Cohen-Macaulay schemes that are not licci, for instance a set of four points in P3 or a fat point in P3 (see [6] and [14], respectively). These two examples are, however, glicci. In [8] the question was raised of whether the converse is true in general for G-liaison, i.e., whether every Cohen-Macaulay ideal is glicci. While there are many affirmative partial results in this direction (see, e.g., [8], [2], [11], [5], [3], [13]), there has also been scepticism. Most notably, [4] and [5] discuss certain sets of points which cannot be shown to be glicci with current methods (see Remark 2.4). In this note we propose to shift focus by enlarging the ambient space of the embedded subschemes. Henceforth, "link" will be understood to mean by Gorenstein ideals (but recall that in codimension two, all Gorenstein ideals are complete intersections). Let X ⊂ Pn be a closed subscheme. Considering Pn as a hyperplane of some Pn+1, we can view X as a subscheme of Pn+1. If X is any finite set of points, then we will see that X is glicci as a subscheme of Pn+1. In fact, much more is true. A special case of our Theorem 2.3 below is: Theorem 1.1. Let X ⊂ Pn be a reduced arithmetically Cohen-Macaulay subscheme. Considering X as a subscheme of Pn+1, X can be linked in Pn+1 to a complete intersection in Pn+1. This motivates the following: ∗ The work for this paper was done while the first author was sponsored by the National Security Agency under Grant Numbers H98230-09-1-0031 and H98230-12-1-0204, and by the Simons Foundation under grant #208579. + The work for this paper was done while the second author was sponsored by the National Security Agency under Grant Numbers H98230-09-1-0032 and H98230-12-1-0247, and by the Simons Foundation under grant #208869. 1 2 JUAN MIGLIORE AND UWE NAGEL Problem 1.2. Let X ⊂ Pn be an arithmetically Cohen-Macaulay subscheme. Can X be G-linked in Pn to a reduced subscheme in a finite number of steps? Is it at least true if we allow the links to take place in Pn+1? Of course, combined with Theorem 1.1, an affirmative answer would imply that every arithmetically Cohen-Macaulay subschema of Pn can be linked to a complete intersection in some higher-dimensional projective space. Another motivation for the above problem is the result of Rao [15] that says that every even liaison class of curves in P3 contains a smooth element. On the other hand, in higher codimension we have the striking result of Polini and Ulrich [14] that says that in some situations, there is a non-reduced locus that is contained in every element of the CI-liaison class. For example, if Z is any union of (non-reduced) fat points in Pn, then Z is a subscheme of every element of its even liaison class, which hence contains no reduced element. Nevertheless, it was shown in [8] that a single fat point (as a very special case of a deeper result) can be G-linked in a finite number of steps to a single point. This result will be generalized in this paper to an arbitrary union of fat points in P3, even of mixed multiplicities. The technique that proves Theorem 1.1 is not limited to reduced subschemes. As example, we discuss monomial ideals. In [11], we showed that strongly stable artinian ideals are glicci. Using a similar technique, Huneke-Ulrich [7] proved that every monomial artinian ideal is glicci in the localization with respect to the homogeneous maximal ideal. This result is not useful for applications when one is interested in information about the Hilbert function. Thus, it is still an important open problem whether each such ideal can be linked homogenously to a complete intersection. Again, we have an affirmative answer if we pass to a larger polynomial ring. Theorem 1.3. Let I ⊂ R := K[x0, . . . , xn] be a monomial ideal such that R/I is Cohen- Macaulay. Let t be a new variable. Then the ideal (I, t) ⊂ R[t] is glicci in R[t]. Since R/I ∼= R[t]/(I, t), this result should be as useful for applications as an affirmative answer to the original question in R would be. We remark that analogous results hold for ideals in regular local rings. 2. Reduced schemes Let R = K[x0, . . . , xn] be a polynomial ring over a field K with its standard grading. The basis of our results is the following observation. Lemma 2.1. Let I ⊂ R be a homogeneous ideal such that R/I is Cohen-Macaulay and generically Gorenstein. If f ∈ R is a homogeneous nonzero divisor of R/I, then the ideal (I, f ) ⊂ R is glicci. Proof. The assumptions on A := R/I guarantee that its canonical module KA can be identified with an ideal of A (cf., for example, [1], Proposition 3.3.18). It follows that there is a Gorenstein ideal J ⊂ R containing I such that codim J = codim I + 1 and KA(−t) ∼= J/I for some integer t. We claim that the ideal (I, f ) is directly G-linked to J. Indeed, by [8, Lemma 4.8], the ideal I + f J also is a Gorenstein ideal with the same codimension as J. Thus it suffices to show that However, since I : f = I this follows easily because (I + f J) : (I, f ) = J. (I + f J) : (I, f ) = (I + f J) : f = (I : f ) + J = J. GLICCI IDEALS 3 (The second equality follows by a standard argument.) We have shown that (I, f ) is di- rectly G-linked to the Gorenstein ideal J. Therefore (I, f ) is glicci because every Goren- stein ideal is glicci by [2]. Note that the proof in [2] is for subschemes of Pn. However, by using graded modules and local cohomology instead of sheaves, the arguments can be adapted so that they also apply to artinian ideals. (cid:3) An immediate consequence is the following: Corollary 2.2. Let V1 be a generically Gorenstein, arithmetically Cohen-Macaulay sub- scheme of Pn and let V2 be a complete intersection that meets V1 properly, that is, dim(V1 ∩ V2) = dim V1 + dim V2 − n. Then V1 ∩ V2 is glicci as a subscheme of Pn. Our first main result is a generalization of Theorem 1.1. Theorem 2.3. Let X ⊂ Pn be an arithmetically Cohen-Macaulay subscheme that is generically Gorenstein. Then, considering X as a subscheme of Pn+1, X can be linked in Pn+1 to a complete intersection of Pn+1. Proof. Let t be a new variable so that S := R[t] is the coordinate ring of Pn+1. By as- sumption, R/IX is Cohen-Macaulay and generically Gorenstein, thus so is S/IX S. Hence, Lemma 2.1 provides that (IXS, t) ⊂ S is glicci in S. Since S/(IX S, t) ∼= R/IX this con- cludes the argument. (cid:3) Notice that this result establishes Theorem 1.1 because every reduced subscheme is generically Gorenstein. Remark 2.4. In [4], Hartshorne investigated the problem of whether arithmetically Cohen-Macaulay subschemes are glicci in the case of general sets of points in P3. He showed that the answer is affirmative if the number of points is at most 19 and proposed a set of 20 general points as a possible counterexample. The problem has been studied further in [5], where it is shown that every general set of at least 56 points in P3 cannot be linked to a complete intersection by strictly descending Gorenstein liaison or biliaison. In contrast, our Theorem 1.1 guarantees that every set of points in projective space can be linked to a complete intersection if we pass to a higher-dimensional space. Now we turn to considering monomial ideals that are not necessarily generically Goren- stein. Proof of Theorem 1.3. Again, let t be a new variable, and set S := R[t]. By [10], we can lift the ideal I ⊂ R to a reduced ideal J ⊂ S such that S/(J, t) ∼= R/I and J : t = J. In particular, S/J is Cohen-Macaulay. Hence, Lemma 2.1 yields that (J, t) is glicci in S, and our claim follows because (J, t) = (I, t)S. (cid:3) 3. Unions of fat points In this section we will give an affirmative answer to Problem 1.2 in the case where X is a union of fat points in P3 (not necessarily of the same degree). We recall that if P is a point in Pn with saturated ideal p then a fat point supported at P is a zero-dimensional scheme defined by the saturated ideal pk for some k ≥ 1. Of course if k > 1 then the fat point is non-reduced. Since a fat point is defined by a standard determinantal ideal, it is glicci by the Gaeta-type result in [8]. However, we begin by giving a new proof for this result, which illustrates the arguments for our more general result in a particularly transparent manner. It uses ideas from [12]. Recall the the h-vector of a d-dimensional arithmetically Cohen-Macaulay subscheme is the d+1-st difference of its Hilbert function. 4 JUAN MIGLIORE AND UWE NAGEL Lemma 3.1. A fat point of Pn is glicci. Proof. Let P ∈ Pn be a point, and without loss of generality assume that IP = hx1, . . . , xni. Let Z ⊂ Pn be the fat point defined by the saturated ideal hx1, . . . , xnia. It is well-known that the h-vector of Z is (cid:18)1, n,(cid:18)n + 1 a − 1 (cid:19)(cid:19) 2 (cid:19), . . . ,(cid:18)n + a − 2 By [12, Theorem 4.3], taking 2a + n − 2 general linear forms in IP , one can define an h-vector arithmetically Gorenstein curve G ⊂ Pn that is a union of 2(cid:0)n+a−2 a − 1 (cid:19), . . . ,(cid:18)n 2(cid:19), . . . ,(cid:18)n + a − 3 (cid:18)1, n − 1,(cid:18)n a−1 (cid:1) lines through P with 2(cid:19), n − 1, 1(cid:19) . Comparing with [12, Theorem 4.1], the curve G contains an arithmetically Cohen-Macaulay a − 1 (cid:19),(cid:18)n + a − 3 a−1 (cid:1) lines with h-vector (cid:18)1, n − 1,(cid:18)n 2(cid:19), . . . ,(cid:18)n + a − 3 a − 1 (cid:19)(cid:19) . subcurve C consisting of (cid:0)n+a−2 Note that both, G and C, are cones over suitable sets of points in the hyperplane {x0 = 0} ∼= Pn−1. Claim. Z ⊂ C. We want to show IC ⊂ IZ = I a P = (x1, . . . , xn)a. From the h-vector we see that )i, where the Fi are homogeneous polynomials of degree a in the IC = hF1, . . . , F(n+a−2 variables x1, . . . , xn. The latter is true by construction of C. Thus all the ith partial derivatives of the Fj, for 1 ≤ i ≤ a − 1, vanish at P . Hence IC ⊂ IZ as claimed. a Let D be the complement of C in G. This means that the curves C and D are geomet- rically linked by G. Furthermore, the curve D is also a reduced union of lines through P , having the same h-vector as C. In the same way as above, Z ⊂ D. Hence IC + ID defines an arithmetically Gorenstein zero-dimensional scheme X, sup- ported at P . Using [12, Lemma 2.5], we find that the h-vector of X is: deg 0 1 2 . . . a − 2 a − 1 a . . . 2a − 4 2a − 3 2a − 2 n 1 Thus X links Z to a zero-dimensional scheme Z ′ supported at P with h-vector h-vector 1 n (cid:0)n+1 2 (cid:1) . . . a−2 (cid:1) (cid:0)n+a−2 (cid:0)n+a−3 (cid:18)1, n,(cid:18)n + 1 a−1 (cid:1) (cid:0)n+a−3 2 (cid:19), . . . ,(cid:18)n + a − 3 a−2 (cid:1) . . . a − 2 (cid:19)(cid:19) . (cid:0)n+1 2 (cid:1) By construction, the generators of IX are polynomials only in x1, . . . , xn. Thus IZ ′ = I a−1 . Continuing inductively, we see that this construction provides a series of links from Z (cid:3) down to P itself, which is reduced. This completes the proof. P If n = 3, the constructions in the above proof become a lot easier. In fact, then the above arithmetically Gorenstein scheme G is simply a complete intersection of forms of degree a and a + 1, respectively, both of which are products of linear forms in IP . This simplicity is the reason why we restrict ourselves to the case n = 3 for showing how to pass from the case of one fat point to a union of fat points. GLICCI IDEALS 5 Theorem 3.2. Let P1, . . . , Ps be points in P3, with defining ideals p1, . . . , ps respectively. Choose positive integers b1, . . . , bs and let Z be the scheme defined by the saturated ideal b1 1 ∩ · · · ∩ pbs s . Then Z is G-linked in P3 to a reduced scheme in a finite number of steps. p Proof. In this proof we will abuse notation somewhat and use the same name for a plane and for the linear form defining it. Our approach will be to apply a series of evenly many Gorenstein links with the following strategy. We will focus on one point at a time, and locally at that point we will be reducing it via the links described in Lemma 3.1 for s = 1. However, globally we will leave any other existing points the way they are (i.e. any reduced points stay reduced, and any fat points retain their multiplicity), and will will pick up new points, but they will all be reduced. Thus at the end we will have removed any non-reduced structure on the original points, and picked up many new reduced points, but we will have produced a reduced scheme in the even Gorenstein class of Z. For compatibility with Lemma 3.1, set b1 = a, P1 = P , and IP = p. Let Z be the scheme defined by IZ = pa ∩ p b2 2 ∩ · · · ∩ pbs s , where a ≥ 2 and bi ≥ 1 for 2 ≤ i ≤ s. Let A = product of a general linear forms in p B = product of a + 1 general linear forms in p The ideal hA, Bi is a complete intersection consisting of a2 + a lines through P . As pointed out above, there is a subset C consisting of (cid:0)a+1 (1, 2, 3, . . . , a). Note that the curve C is arithmetically Cohen-Macaulay (being a cone over a finite set of points). Let D be the residual in the complete intersection defined by hA, Bi. Notice that D is also arithmetically Cohen-Macaulay and has the same h-vector as C. By the claim above, we have IC ⊂ pa and ID ⊂ pa. 2 (cid:1) of these lines with h-vector Choose general planes Li,j ∈ pi, where 2 ≤ i ≤ s and 1 ≤ j ≤ bi. Notice that bi i . Let Qj Li,j ∈ p Notice that deg F = a +Ps that F ∈ IC (hence also in IZ). F = A ·Yi,j Li,j. i=2 bi is the sum of the powers of the ideals of the points, and Choose general planes Mi,j ∈ pi, where 2 ≤ i ≤ s and 1 ≤ j ≤ bi. Let Mi,j. Q =Yi,j Define a curve Y with defining ideal IY = Q · IC + hF i. This is a basic double link (see for instance [9]), and since C is arithmetically Cohen- Macaulay, it follows that so is Y . Furthermore, by the genericity of Li,j and Mi,j and comparing degrees, we see that IY = IC ∩ hF, Qi. Thus, Y is a union of lines with at most two passing through any point other than the Pi or P . It also follows that Z ⊂ Y . Through each point Pi (2 ≤ i ≤ s) pass b2 i lines of Y , namely the complete intersection ofQj Li,j andQj Mi,j. The components of Y containing P form the curve C. 6 JUAN MIGLIORE AND UWE NAGEL Choose general planes Ni,j ∈ pi, where 2 ≤ i ≤ s and 1 ≤ j ≤ bi. Let We make the following observations about the ideal hF, Gi. G = QB ·Yi,j Ni,j. (1) It is a complete intersection defining a union of lines. (2) An any point other than P or one of the Pi, no more than two of these lines pass. (3) At any of Pi ∈ {P2, . . . , Ps}, locally it is a complete intersection of type (bi, 2bi), namely the complete intersection of Qj Li,j and Qj Mi,j ·Qj Ni,j. (4) At P it is a complete intersection of type (a, a + 1), namely that defined by hA, Bi. (5) F ∈ IY by construction and G ∈ IY since B ∈ IC so QB ∈ Q · IC; hence hF, Gi ⊂ IY . Thus hF, Gi links Y to a residual curve W , which is arithmetically Cohen-Macaulay since Y is. W is again a union of lines, with at most two meeting at any point away from P or one of the Pi. Locally at P , the components of W coincide with the components of D. At any Pi ∈ {P2, . . . , Ps}, W is a complete intersection of type (bi, bi); namely it is the complete intersection of Qj Li,j and Qj Ni,j. By a similar reasoning as above, Z ⊂ W . Furthermore, the link joining Y and W is a [9]), IY + IW is the saturated geometric link. Thus by standard liaison theory (see e.g. ideal of a codimension three Gorenstein scheme, which we will call Gor; and since IY ⊂ IZ and IW ⊂ IZ, we obtain Z ⊂ Gor. Locally at any Pi ∈ {P2, . . . , Ps}, Gor is the complete intersection *Yj Li,j,Yj Mi,j,Yj Ni,j+ . Locally at P , Gor is precisely the Gorenstein scheme X described in the proof of Lemma 3.1 in the case n = 3. Everywhere else, Gor is reduced. Since IGor ⊂ IZ, Gor links Z to a residual zero-dimensional scheme Z ′. Thanks to the work done in the case s = 1, the component of Z ′ supported at P is just defined by pa−1. The components supported at P2, . . . , Ps are non-reduced in general (and it will not matter to us what they look like there), but everywhere else Z ′ is reduced. Denote by {Rk} the remaining points other than P and P2, . . . , Ps. It does not matter how many points Rk there are; let us suppose 1 ≤ k ≤ τ . Now we will perform a second link, from Z ′ to a new zero-dimensional scheme Z ′′ which will have important properties for us. The approach is similar, but with small differences. Let A′ = product of a general linear forms in p B ′ = product of a − 1 general linear forms in p Then hA′, B ′i is a complete intersection defining a reduced union of a2 − a lines through P . Let C ′ be a subset of (cid:0)a arithmetically Cohen-Macaulay curve. Let D′ be the residual. D′ is also arithmetically Cohen-Macaulay, with the same h-vector. As earlier, IC ′ ⊂ p 2(cid:1) of these lines with h-vector (1, 2, 3, . . . , a − 1). C ′ is an a−1 and ID′ ⊂ p a−1. For each 1 ≤ k ≤ τ , let L′ k, M ′ k and N ′ k be general linear forms vanishing on Rk. Let k F ′ = A′ ·Qi,j Li,j ·Qk L′ Q′ = Q ·Qk M ′ IY ′ = Q′ · IC ′ + hF ′i k GLICCI IDEALS 7 As before, Y ′ is an arithmetically Cohen-Macaulay union of lines. The components of Y ′ passing through P are precisely C ′. The components of Y ′ passing through P2, . . . , Ps form a complete intersection of lines (namely at Pi they are defined by hQj Li,j,Qj Mi,ji). Through each Rk there is just one line (namely that defined by hL′ ki). k, M ′ Now let G′ = Q′ · B ′ ·Yi,j Ni,j ·Yk N ′ k. Since both F ′ and G′ are in IY ′, and by the genericity of their choices, hF ′, G′i provides a geometric link of Y ′ to a residual curve, W ′. Locally at P , W ′ consists of the components of D′. Locally at Pi, 2 ≤ i ≤ s, W ′ consists of the same complete intersection that we had for W , namely the complete intersection of type (bi, bi) given by Qj Li,j and Qj Ni,j. Locally at Rk, W ′ is the complete intersection of type (1, 1) (i.e. a single line) given by hLk, Nki. Now, since Y ′ and W ′ are geometrically linked, the sum of their ideals is an arith- metically Gorenstein ideal, Gor′. At P it is the intersection of the linked arithmetically Cohen-Macaulay curves C ′ and D′, hence the component of Gor′ at P is itself arithmeti- cally Gorenstein, with h-vector (cid:18)1, 3, 6, . . . ,(cid:18)a − 1 2 (cid:19),(cid:18)a 2(cid:19),(cid:18)a − 1 2 (cid:19), . . . , 6, 3, 1(cid:19) . The component of Gor′ supported at Pi is the complete intersection hYj Li,j,Yj Mi,jYj Ni,ji. The component of Gor′ supported at Rk is reduced. There are also other reduced com- ponents of Gor′. Now, Gor′ provides a link of Z ′ to a residual zero-dimensional scheme Z ′′. The compo- nent of Z ′′ supported at P has defining ideal pa−2 thanks to the calculations in the proof bi of Lemma 3.1. The component of Z ′′ at Pi, 2 ≤ i ≤ s, are the original fat points, p i , since locally we have linked twice using the same complete intersection. Z ′′ has no components supported at any of the Rk. Everywhere else, Z ′′ is reduced. Thus in two links we have passed from the ideal pa ∩ p scheme consisting of the form b2 2 ∩ · · · ∩ pbs s to a zero-dimensional a−2 ∩ p p b2 2 ∩ · · · ∩ p bs s ∩ J where J is the ideal of a reduced set of points. Thus, depending on whether a is even or odd, we can reduce in a finite (even) number of Gorenstein links to a zero-dimensional scheme where the components supported at P2, . . . , Ps are the original fat points, the component at P is either reduced or empty, and all other points are reduced. We can then do the same thing at each of P2, . . . , Ps, obtaining in the end a reduced zero-dimensional scheme (of very large degree). (cid:3) In the above result it is crucial that we are not bound to use complete intersections for the links. In fact, by the results in [14] it not possible to link any union of fat points that is not reduced to a union of simple points by using only complete intersections for the links. Corollary 3.3. Every union of fat points in P3 can be linked in P4 to a complete inter- section of P4 in a finite number of steps. 8 JUAN MIGLIORE AND UWE NAGEL Proof. Combine Theorems 3.2 and 2.3. (cid:3) Using the arguments for establishing Lemma 3.1, we believe that the ideas in the proof of Theorem 3.2 can be extended to show that every union of fat points in Pn can be linked to a reduced subscheme of Pn. However, the details seem to become very demanding. References [1] W. Bruns and J. Herzog: Cohen-Macaulay Rings, Cambridge Studies in Advanced Mathematics 39, Cambridge University Press, Cambridge, 1998. [2] M. Casanellas, E. Drozd, and R. Hartshorne, Gorenstein liaison and ACM sheaves, J. Reine Angew. Math. 584 (2005), 149 -- 171. [3] E. Gorla, A generalized Gaeta's Theorem, Compositio Math. 144 (2008), 689 -- 704. [4] R. Hartshorne, Some examples of Gorenstein liaison in codimension three, Collect. Math. 53 (2002), 21 -- 48. [5] R. Hartshorne, I. Sabadini, and E. Schlesinger, Codimension 3 arithmetically Gorenstein subschemes of projective N -space, Ann. Inst. Fourier (Grenoble) 58 (2008), 2037 -- 2073. [6] C. Huneke, B. Ulrich, The structure of linkage, Ann. of Math. 126 (1987), 221 -- 275. [7] C. Huneke and B. Ulrich, Liaison of monomial ideals, Bull. London Math. Soc. 39 (2007), 384 -- 392. [8] J. Kleppe, J. Migliore, R.M. Mir´o-Roig, U. Nagel, and C. Peterson, Gorenstein liaison, complete intersection liaison invariants and unobstructedness, Mem. Amer. Math. Soc. 154 (2001), no. 732. [9] J. Migliore, Introduction to Liaison Theory and Deficiency Modules, Progress in Mathematics 165, Birkhauser, 1998. [10] J. Migliore and U. Nagel, Lifting monomial ideals, Comm. Algebra 28 (2000), Special volume in honor of R. Hartshorne, 5679-5701. [11] J. Migliore and U. Nagel, Monomial ideals and the Gorenstein liaison class of a complete intersection, Compositio Math. 133 (2002), 25 -- 36. [12] J. Migliore and U. Nagel, Reduced arithmetically Gorenstein schemes and simplicial polytopes with maximal Betti numbers, Adv. Math. 180 (2003), 1 -- 63. [13] U. Nagel and T. Romer, Glicci simplicial complexes, J. Pure Appl. Algebra 212 (2008), 2250 -- 2258. [14] C. Polini and B. Ulrich, Linkage and reduction numbers, Math. Ann. 310 (1998), 631 -- 651. [15] P. Rao, Liaison among Curves in P3, Invent. Math. 50 (1979), 205 -- 217. Department of Mathematics, University of Notre Dame, Notre Dame, IN 4655, USA E-mail address: [email protected] Department of Mathematics, University of Kentucky, 715 Patterson Office Tower, Lexington, KY 40506-0027, USA E-mail address: [email protected]
math/0212279	5	0212	2010-02-23T00:45:36	Poisson deformations of symplectic quotient singularities	[ "math.AG", "math.DG", "math.QA", "math.SG" ]	We establish a connection between smooth symplectic resolutions and symplectic deformations of a (possibly singular) affine Poisson variety. In particular, let V be a finite-dimensional complex symplectic vector space and G\subset Sp(V) a finite subgroup. Our main result says that the so-called Calogero-Moser deformation of the orbifold V/G is, in an appropriate sense, a versal Poisson deformation. That enables us to determine the algebra structure on the rational cohomology H^(X) of any smooth symplectic resolution X \to V/G (multiplicative McKay correspondence). We prove further that if G is an irreducible Weyl group in GL(h) and V=h+ h^ then no smooth symplectic resolution of V/G exists unless G is of types A,B, or C.	math.AG	math	Poisson deformations of symplectic quotient singularities Victor Ginzburg and Dmitry Kaledin To the memory of Andrey Tyurin Abstract We establish a connection between smooth symplectic resolutions and sym- plectic deformations of a (possibly singular) affine Poisson variety. In particular, let V be a finite-dimensional complex symplectic vector space and G ⊂ S p(V ) a finite subgroup. Our main result says that the so-called Calogero-Moser deformation of the orbifold V /G is, in an appropriate sense, a versal Poisson deformation. That enables us to determine the algebra structure on the cohomology H q (X, C) of any smooth symplectic resolution X ։ V /G (multiplicative McKay correspondence). We prove further that if G ⊂ GL(h) is an irreducible Weyl group and V = h ⊕ h∗ , then no smooth symplectic resolution of V /G exists unless G is of types A, B, C. Contents 2 1 Main results. 2 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1.1 Introduction. 4 1.2 Orbifold cohomology and Quantization. . . . . . . . . . . . . . . . . . . . . . . 6 1.3 Poisson deformations. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8 1.4 Deforming symplectic resolutions. . . . . . . . . . . . . . . . . . . . . . . . . . . 1.5 Comparison. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10 1.6 Calogero-Moser deformation. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11 1.7 Proofs of Theorem 1.1 and Theorem 1.2. . . . . . . . . . . . . . . . . . . . . . . 13 2 Generalities on Poisson deformations. 13 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14 2.1 Poisson cohomology. 2.2 Globalization. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15 3 The case of dim V = 2. 4 The computation of H P 2 (V /G). 17 20 24 5 Resolutions. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24 5.1 Geometry of the resolution. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26 5.2 Globalizing the deformation. 5.3 Comparison with the Calogero-Moser deformation. . . . . . . . . . . . . . . . . 29 5.4 Restriction to strata and the end of the proof. . . . . . . . . . . . . . . . . . . . 30 6 Applications of Hochschild cohomology. 33 1 38 7 Appendix. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 38 A.1 Harrison cohomology. A.2 Poisson cohomology. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 39 A.3 Relation to Hochschild cohomology and the Kunneth formula. . . . . . . . . . . 41 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 44 A.4 Graded algebras. A.5 Modules and ´etale descent. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 44 A.6 Poisson schemes. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 46 A.7 The smooth case. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 49 A.8 Deformations. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 50 1 Main results. 1.1 Introduction. Let V be a finite-dimensional symplectic vector space over C, and G ⊂ S p(V ) a finite subgroup. The quotient V /G has a natural structure of an irreducible affine algebraic variety with coordinate ring C[V /G] := C[V ]G , the subalgebra of G-invariant polynomials on V . The algebraic variety V /G is normal and Gorenstein. However, it is always singular whenever G 6= {1}. The symplectic structure on V gives rise to a Poisson algebra structure on C[V ] which induces, by restriction, a Poisson algebra structure on C[V ]G . Thus, the space V /G becomes a Poisson variety. Recall that a resolution of singularities X → Y of an irreducible Goren- stein variety Y is called crepant provided the smooth manifold X has trivial canonical class. Assume now Y has a Poisson structure which is generically non-degenerate (i.e., symplectic). Then a resolution of singularities X → Y is called symplectic provided the pull-back of the symplectic 2-form on the generic locus of Y extends to a (non-degenerate) symplectic 2-form on the whole of X . Any symplectic resolution is crepant. Conversely, it was shown in [K1] that any crepant resolution X → Y is necessarily symplectic, i.e., the pull-back of symplectic 2-form on the generic locus of Y automatically extends to a non- degenerate 2-form on X . In this paper we study crepant (equivalently, symplectic) resolutions X → V /G of a symplectic quotient singularity V /G. Our first result concerns the existence of such resolutions. It has been shown by M. Verbitsky [Ve] that if the quotient V /G admits a crepant resolution, then the group G must be generated by the so-called symplectic reflections (see [Ve] or Definition 1.15 below). If dim V = 2 then any finite subgroup G ⊂ SL(V ) is generated by Example. symplectic reflections, and V /G is the so-called du Val surface singularity. It is well-known that there exists a canonical minimal resolution of singularities X → V /G, which is at the same time a symplectic resolution. An important series of groups generated by symplectic reflections are pro- vided by Coxeter groups. Specifically, let h be the complexification of a euclidean real vector space with root system of some Dynkin type, and let G ⊂ GL(h) be the corresponding Weyl group. Further let C2 be the standard space with a fixed volume 2-form, and set V := h ⊗ C2 . The tensor product of the symmetric bilinear form (coming from the euclidean inner product) on h and the 2-form 2 on C2 gives a nondegenerate 2-form on V , hence, makes V a symplectic vector space. The group G acts on V = h ⊗ C2 , via the action on the first factor, by symplectic automorphisms. For types An , Bn , Cn the quotient V /G is known to admit a crepant reso- lution of singularities (of Hilbert scheme type, see e.g. [Ku]). Using the results of I. Gordon [Go], we resolve the existence question for other Dynkin graphs to the negative: Theorem 1.1. For a root system of type Dn , En , F4 or G2 , the quotient V /G where V = h ⊗ C2 , does not admit a resolution of singularities X → V /G with trivial canonical bund le. Following [Al], define an increasing filtration F q (C[G]) on the group algebra C[G] by letting Fk (C[G]), k ≥ 0, be the C-linear span of the elements g ∈ G such that rk(idV −g ) ≤ k . (Thus, 1 ∈ F0 (C[G]) and symplectic reflections belong to F2 (C[G])). This filtration is clearly compatible with the algebra structure on C[G]. Let F q (ZG) denote the induced filtration on ZG, the center of C[G]. Write grF q (ZG) for the corresponding associated graded algebra. Let X → V /G be a symplectic resolution. Our second result describes the algebra structure in the cohomology H q (X, C) of the manifold X . Theorem 1.2 (Multiplicative McKay correspondence). Let X ։ V /G be a resolution of singularities with trivial canonical bund le. Then there is a canonical graded algebra isomorphism: H q (X, C) ∼= grF q (ZG). In particular, X has no odd (rational) cohomology. We note that the dimension equality: dim H i (X, C) = dim gri (ZG) has been known for some time, see [BD]. Later on, an explicit basis in the Borel-Moore homology H BM q (X, C) parametrized by conjugacy classes in G was constructed in [K2]. This constitues the so-called generalized McKay correspondence and d− q (X, C) ∼−→ gr q (ZG), amounts, in our notation, to a linear bijection γ∗ : H BM where d = dimR X . The multiplication structure in the cohomology was first computed indepen- dently by Vasserot [Vas] and Lehn-Sorger [LS] (cf. also [LQW]) in the case when X is the Hilbert scheme of n points on C2 . At the same time, Vasserot (and the first author) conjectured the existence of a natural algebra isomorphism: γ ∗ : grF q (ZG) ∼−→ H q (X, C) for arbitrary symplectic quotient singularities; our result proves this conjecture. The natural C∗ -action on V by dilations induces a C∗ -action on V /G. The latter may be canonically lifted to the symplectic resolution X , cf. [K1] and Proposition 5.2 below. Recall further that C∗ -equavariant cohomology of a C∗ - variety is an algebra over C[u], the cohomology of the classifying space. Further, using the filtration F q (ZG) one forms the corresponding graded Rees algebra Rees q (ZG) := Pi Fi (ZG) · ui ⊂ ZG[u]. Conjecture 1.3 (C∗ -equavariant cohomology). There is a canonical graded C∗ (X, C) ∼= Rees q (ZG). C[u]-algebra isomorphism: H q 3 In the special case of the Hilbert scheme of n points in C2 a proof of this conjecture is implicitly contained in [Vas]. Theorem 1.2 leads further to an interesting question of computing the Poin- car´e duality isomorphism for X in terms of the group G; in other words we propose Problem 1.4. Compute the composite map grF q (ZG) γ ∗ ∼ / H q (X, C) γ∗ Poincar´e duality / H BM d− q (X, C) ∼ ∼ / gr q (ZG). This map seems to be closely related to the character table of the group G. 1.2 Orbifold cohomology and Quantization. Let M be an arbitrary smo- oth symplectic algebraic variety, and G a finite group of symplectic automor- phisms of M . There are many examples, cf. [Ba], [BKR], in which, given a crepant resolution X → M /G one has a canonical equivalence of triangulated categories: Db (C oh(X )) ∼= Db (C ohG (M )). In such a case, taking the Grothen- dick groups of both categories, one obtaines an isomorphism: K (X ) ∼= KG (M ) of the algebraic K-groups. In the special case where V = M is a symplectic vector space and G ⊂ S p(V ), the Borel-Moore homology group H BM q (X, C) is known, by [K2], to be spanned by the algebraic cycles. Hence the Chern character map gives a ring isomorphism ch : C⊗K (X ) ∼−→ H q (X, C). One also has the Thom isomorphism: KG (V ) ∼= R(G), where R(G) stands for the representation ring of G. Thus, in addition to Problem 1.4 we arrive at the following Problem 1.5. Compute the composite map (cf. also [EG, Problem 17.11]): ZG ∼→ C ⊗ R(G) ∼→ C ⊗ KG (V ) ∼→ C ⊗ K (X ) ∼ ch / H q (X, C) ∼ 1.4 / gr(ZG). Note that the group ZG on the left is viewed as the algebra of class-functions on G with pointwise multiplication. Note further, that the isomorphism K (X ) ∼= KG (V ) is not compatible with the ring structures. Associated with a finite group action on a Calabi-Yau manifold M , one can introduce an orbifold (= ‘stringy’) cohomology H q orb (M ; G), see [Ba]. We consider the special case where M is a holomorphic symplectic manifold and the group G acts by symplectic automorphisms. In such a case the definition reads: orb (M ; G) := (cid:16)Mg∈G (M g )(cid:17)G H q where M g denotes the fixed point set of g ∈ G, and where H q−dim M g (M g ) is a shorthand notation for Lα H q−dim Mα (Mα ), a direct sum ranging over the set of connected components, Mα , of the manifold M g . Further, for any g , h ∈ G, 4 H q−dim M g (1.1) , / / / / / one has a cup-product pairing ∪ : H q (M g ) × H q (M h ) → H q (M g ∩ M h), and the Gysin map ı∗ : H q (M g ∩ M h ) → H q (M gh ), induced by the imbedding ı : M g ∩ M h ֒→ M gh . Following Ruan, cf. [R1], [R2], Fantechi and Gottsche [FG] have introduced a certain cohomology class c(g , h) ∈ H q (M g ∩ M h), and showed that the assignment H q (M g ) × H q (M h ) −→ H q (M gh ), a, b 7−→ ı∗ (a ∪ b ∪ c(g , h)) gives rise to an associative product on the direct sum in the RHS of (1.1). This product puts a graded algebra structure on orbifold cohomology. It is known further, see [Ba], [DL] and also [Ba], that, given an arbitrary Calabi-Yau orbifold M /G and a smooth crepant resolution X → M /G, there orb (M ; G) ∼= H q (X ). Moreover, it has been is a graded space isomorphism H q conjectured in [CR],[R1] and [FG], that there is a graded algebra isomorphism orb (M ; G) ∼= H q (X ), provided M is a holomorphic symplectic manifold with H q symplectic G-action and X → M /G is a symplectic resolution. The above conjecture is supported by the special case where M = V is a symplectic vector space and G ⊂ S p(V ). Then each fixed point set V g is contractible, and formula (1.1) reduces to H q orb (V ; G) = gr q (ZG) (lemma A.16 from §6 below insures compatibility of the gradings on both sides). Thus, the orb (M ; G) ∼= H q (X ) becomes nothing but conjectured algebra isomorphism H q our Theorem 1.2 above. For a general algebraic symplectic manifold M , the above conjecture may be approached as follows. The symplectic form on M makes the structure sheaf into a sheaf, (OM , {−, −}M ), of Poisson algebras. Consider a deformation- quantization of OM , i.e., a sheaf QuantM of locally free complete C[[ε]]-algebras with a G-equivariant star-product a, b 7→ a ⋆ b, such that a ⋆ b − b ⋆ a ≡ ε · {a, b}M (mod ε2 ). Let QuantM [ 1 ε ] be the localization of QuantM with respect to ε, a sheaf of C((ε))-algebras on M . We form the cross-product algebra ε ] # G. Let (QuantM [ 1 QuantM [ 1 ε ] # G) -mod be the abelian category of coher- ent QuantM [ 1 ε ] # G-modules, that is, the category of G-equivariant coherent QuantM [ 1 ε ]-modules. On the other hand, given a smooth symplectic resolution X → M /G, we consider similarly a sheaf QuantX of C[[ε]]-algebras, which is a deformation- quantization of the structure sheaf (OX , {−, −}X ), viewed as a sheaf of Poisson algebras on X . Let QuantX [ 1 ε ] be its localization with respect to ε, a sheaf of C((ε))-algebras on X . Write QuantX [ 1 ε ] -mod for the abelian category of sheaves of coherent QuantX [ 1 ε ]-modules. We propose the following conjecture that may be thought of as a ‘quantum analogue’ of Bridgeland-King-Reid theorem [BKR]. Conjecture 1.6. For appropriate choices of deformation-quantizations of sym- plectic manifolds X and M , respectively, there is a category equivalence ε ] -mod ≃ (QuantM [ 1 QuantX [ 1 ε ] # G) -mod . 5 Remark 1.7. Notice that, unlike the situation considered in [BKR], the conjec- ture above involves no derived categories. This is somewhat reminiscent of the ‘D-affineness’ property, proved by Beilinson-Bernstein [BB] for flag manifolds. Now, given a space U and a sheaf of algebras AU on U , define the Hochschild cohomology of AU by the formula HH q (AU ) := Ext q (AU , AU ), where AU ⊠A op U AU is viewed as a sheaf on the diagonal U ⊂ U × U . It was shown in [EG, §15] that Kontsevich’s Formality theorem [Kon] yields a graded C((ε))-algebra isomorphism: (1.2) (1.3) C((ε)) ⊗ H q (X ) ∼= HH q (QuantX [ 1 ε ]). Further, we expect that there is a graded C((ε))-algebra isomorphism orb (M ; G) ∼= HH q ((QuantM [ 1 C((ε)) ⊗ H q ε ]) # G), that may be viewed as a ‘quantization’ of the isomorphism of Proposition 6.2 (see §6 below). Assuming this, the equivalence of Conjecture 1.6 would yield an isomorphism between the Hochschild cohomology algebras in the right-hand sides of (1.2) and (1.3). Hence, the corresponding left-hand sides should also be orb (M ; G) ∼= isomorphic, and we would get the desired algebra isomorphism H q H q (X ). 1.3 Poisson deformations. The main idea of this paper, used in particular in the proofs of Theorem 1.1 and Theorem 1.2, is to relate smooth symplectic resolutions of V /G to smooth symplectic deformations of V /G. Specifically, a certain canonical Poisson deformation of the Poisson algebra C[V ]G , which we propose to call the Calogero-Moser deformation, has been introduced in [EG]. Our main Theorem 1.18 claims that – under some assumptions, and in an appropriate sense – the Calogero-Moser deformation is a versal deformation of V /G in the class of Poisson algebras. In the course of proving the Theorem, we establish some general basic results on Poisson deformations, which may be of independent interest. We will now introduce the necessary definitions and state, one by one, the technical results leading up to and including Theorem 1.18. At the the end of this section, we will show how the stated results imply Theorem 1.1 and Theorem 1.2. Let A be a commutative unital C-algebra with product (a, b) 7→ a · b, and R ⊂ A a (unital) subalgebra. Recall that A is said to be a Poisson algebra over R, or a Poisson R-algebra, if A is equipped with an R-linear skew-symmetric bracket {−, −} that satisfies the Leibniz rule {a, (b · c)} = {a, b} · c + {a, c} · b (1.4) and the Jacobi identity for all a, b, c ∈ A, (1.5) {a, {b, c}} + {b, {c, a}} + {c, {a, b}} = 0 for all a, b, c ∈ A. 6 Geometrically, the embedding R ֒→ A corresponds to a scheme morphism f : Spec A → Spec R, and the Poisson R-algebra structure on A makes each fiber of f a Poisson scheme. In particular, if R is a local Artin algebra with maximal ideal m, then the fiber A/m over the special point o ∈ S = Spec R is a Poisson algebra over C. Definition 1.8. A Poisson deformation of a Poisson algebra A over the spec- trum S = Spec R of a local Artin algebra R with maximal ideal m ⊂ R is a pair of a flat Poisson R-algebra AR and a Poisson algebra isomophism AR /m ∼= A. It is useful to introduce gradings into the picture. Say that a unital asso- ciative algebra A = Ln≥0 An is positively graded if for all n, m ≥ 0 we have An · Am ⊂ An+m , and A0 = C · 1. Let A be a commutative positively graded C-algebra, and let R ⊂ A be a graded subalgebra. Fix l > 0, a positive integer. We will say that A is a graded Poisson R-algebra of degree l if A is a Poisson R- algebra and, for any n, m ≥ 0, we have {An , Am} ⊂ An+m−l . Repeating literally Definition 1.8, we obtain the notion of a graded deformation of a graded Poisson algebra A over the spectrum of a graded local Artin algebra R. Note that our assumption implies in particular that the maximal ideal m ⊂ R coincides with the augmentation ideal Ln≥1 Rn ⊂ R. A Poisson R-algebra AR over a complete local C-algebra hR, mR i equipped with a Possion isomorphism AR /mR ∼= A is called a universal formal Poisson deformation of the Poisson algebra A if for every Poisson deformation AS over a local Artin base S there exists a unique map τ : S → Spec R such that the isomorphism A ∼= AR /mR extends to a Poisson isomorphism AS ∼= τ ∗AR . (1.6) Analogously, say that a graded Poisson R-algebra AR over a positively graded C-algebra R equipped with a graded Poisson isomorphism AR/R>0 ∼= A is a universal graded Poisson deformation of the Poisson algebra A if for every graded Poisson deformation AS over a local Artin base S with an action of C∗ there exists a unique C∗ -equivaraint map τ : S → Spec R such that the isomorphism A ∼= AR /R>0 extends to a graded Poisson isomorphism (1.6). To study Poisson deformations of a given Poisson C-algebra A, we develop a cohomology theory, H P q (A), the Poisson cohomology of A. Very much like Hochshcild cohomology of an associative algebra, first order deformations of a Poisson algebra A are controlled by the group H P 2 (A), while obstructions to deformations are controlled by the group H P 3 (A). The Poisson cohomology groups are sufficiently functorial; in particular, for a graded Poisson algebra A, each space H P k (A) carries a natural additional grading. Remark 1.9. For Poisson algebras such that Spec A is smooth, the theory of Poisson cohomology goes back to J.-L. Koszul [Ko], and J.-L. Brylinski [Br]. In the singular case, the approach must be quite different; it is based on the general operadic formalism (see [Fr], and Appendix below). 7 Assume now that H P 1(A) = 0 and dim H P 2 (A) < ∞. Then, the result below says that there exists a formal universal Poisson deformation of A; more- over, under additional assumptions, in the graded setting the formal universal deformation comes from a deformation over a base of finite type over C. To make a precise statement, let \H P 2 (A) denote the completion of the affine space H P 2(A). Theorem 1.10 (Universal Poisson deformations). Let A be a Poisson al- gebra. Assume that H P 1 (A) = 0 and that H P 2 (A) is a finite-dimensional vector space over C. (i) There exists a closed subscheme S ⊂ \H P 2 (A) and a Poisson C[S ]-algebra AS which is a universal formal Poisson deformation of the algebra A. (ii) Assume in addition that the Poisson algebra A is positively graded of some positive degree l, and that the induced grading on H P 2 (A) is also positive. View H P 2(A) as a C∗ -variety via the grading. Then there exists a C∗ -stable closed subvariety S ⊂ H P 2(A) and a pos- itively graded Poisson C[S ]-algebra AS of degree l which is a universal graded Poisson deformation of the algebra A. The proof of Theorem 1.10 is contained in the Appendix, in Subsection A.8. Remark 1.11. We would like to note that the isomorphism (1.6) is not canon- ical, nor indeed is it unique. Thus what we obtain is more than a “coarse deformation space” and less than a “fine moduli space”. In the language of the stack formalism, our deformation problems are classified by the quotient stack of some variety S by a trivial action of some group G. The group in question is the connected component of unity in the group of Poisson automorphisms of the algebra A. 1.4 Deforming symplectic resolutions. Let X aff := Spec H 0 (X, OX ) de- note the affinization of an algebraic variety X . This is an affine algebraic variety. We first introduce Definition 1.12. An irreducible smooth algebraic variety X will be called con- vex if the canonical morphism X → X aff is pro jective and birational. The affine algebraic variety X aff is then a normal variety. For a convex symplectic manifold X , the symplectic structure on X induces a Poisson bracket on the algebra H 0 (X, OX ), hence, a Poisson structure on X aff . Our next result concerns deformations of symplectic convex manifolds. Re- lated results on infinitesemal and formal deformations have been studied in the paper [KV]. In order to generalize the main theorem of [KV] to a global set- ting, we need to assume that the affinization Y = X aff is equipped with an expanding action of the multiplicative group C∗ . Equivalently, we assume that 8 the algebra C[Y ] = H 0 (X, OX ) carries a grading by non-negative integers such that the Poisson bracket on C[Y ] is of some fixed degree l > 0, that is, for any homogeneous functions a1 , a2 ∈ C[Y ], we have deg{a1 , a2} = deg a1 + deg a2 − l. With these assumptions, the result below says that any convex symplectic manifold can be deformed nicely1 to a smooth affine symplectic manifold without changing the rational cohomology algebra. Theorem 1.13. Let X be a symplectic convex variety, and X → Y = X aff the corresponding resolution. Assume that the algebra C[Y ] carries a grading by non-negative integers such that the Poisson bracket has degree l > 0. Put B := H 2 (X, C) and equip the affine space B with a C∗ -action by z · b = z−l b, z ∈ C∗ , b ∈ B = H 2 (X, C). Then there exists a smooth C∗ -variety XB and a smooth C∗ -equivariant mor- phism π : XB → B such that (i) XB is a relative symplectic manifold over B , i.e., we have a relative 2-form ω ∈ H 0 (XB , Ω2 (XB /B )) which induces a symplectic structure on each fiber Xb , b ∈ B = H 2 (X, C). (ii) The relative cohomology sheaves Rk π∗Q are constant sheaves on B for al l k ≥ 0, and the canonical base change morphism H k (Xb , Q) → (cid:0)Rk π∗Q(cid:1)b is an isomorphism for every point b ∈ B . (iii) The affinization, (XB )aff , is flat over B . The canonical map XB → (XB )aff is projective, birational, and it is an isomorphism over the generic point b ∈ B . (iv) The special fiber X0 over 0 ∈ B = H 2 (X, C) is isomorphic to X as a symplectic algebraic variety. The special fiber (X0 )aff of the affinization (XB )aff is isomorphic to X aff = Y as a Poisson algebraic variety with C∗ -action. 1 this is similar to the deformation of the Springer resolution N → N (of the nilpotent variety N in a semisimple Lie algebra g) provided by Grothendieck’s simultaneous resolution g → g. Here X = N , X aff = N , and XB in the Theorem below plays the role of g. 9 In other words, there exists a Zariski open, dense subset B generic ⊂ B , such that one has a commutative diagram X Bgeneric π / XB π Xo π (X Bgeneric )aff (XB )aff (Xo )aff X Y / B B generic {o} {o} In this diagram, the subscript ‘B generic ’ indicates restricion to B generic of a scheme over B . Note, in particular, that according to Theorem 1.13. the map Bgeneric )aff , on the left of Bgeneric → B generic is affine, so the arrow X Bgeneric → (X X the diagram above is an isomorphism. The proof of Theorem 1.13 is contained in Subsection 5.2. 1.5 Comparison. Let X be a convex symplectic manifold whose affinization Y = X aff is a positively-graded Poisson algebra of degree l, so that the as- sumptions of Theorem 1.13 are satisfied. Assume in addition that H P 1 (C[Y ]) = 0, dim H P 2 (C[Y ]) < ∞, and the natural grading on H P 2 (C[Y ]) is positive, so that we can apply Theorem 1.10 to get the universal graded Poisson defor- mation (of Y ) over a base S . On the other hand, let XB /B be the deformation provided by Theorem 1.13. Then the affinization (XB )aff is by construction a positively-graded flat Poisson OB -algebra. Hence, by Theorem 1.10 (ii), we get the classifying map τ : B → S := base of universal deformation. In Subsection 5.2 we will prove Proposition 1.14. The classifying map τ : B → S is a finite map onto an irreducible component of the variety S . Now, fix a symplectic vector space V and a finite group G ⊂ S p(V ). Definition 1.15 ([EG]). An element g ∈ G is called a symplectic reflection if rk(idV −g ) = 2. Let Σ denote the set of symplectic reflections in G. The group G acts on Σ by conjugation, and we put n := number of G-conjugacy classes in Σ. Introduce a grading on the polynomial algebra C[V ] by assigning degree 1 to all linear functions. This turns C[V ] into a positively graded Poisson algebra of degree l = 2. Set Y = V /G. Both the grading and the Poisson bracket descend to the algebra C[Y ] = C[V ]G . In Section 4 we will prove 10 / ? _ o o / / ? _ o o / ? _ o o Proposition 1.16. We have H P 1 (C[V ]G ) = 0 and dim H P 2 (C[V ]G ) = n. Moreover, the natural grading on H P 2 (C[V ]G ) is positive. Let X → V /G be a symplectic resolution. The Betti numbers of X are known by [BD] (see [K2] for an alternative proof ), in particular, we have dim H 2 (X, C) = n. Applying Proposition 1.14 we obtain a sequence of maps H 2 (X, C) τ→ S ֒→ H P 2 (C[V /G]), where the first map is a finite surjection onto an irreducible component, and the second map is a closed embedding. Since dim H P 2 (C[V /G]) = n = dim H 2 (X, C), we conclude that S is the whole H P 2 (C[V /G]) and τ : H 2 (X, C) → S is a finite dominant map. Remark 1.17. The composite map H 2 (X, C) → H P 2 (C[V /G]) is not a bijec- tion. Although both the source and the target of this map are affine spaces, the map is not linear; it is a ramified covering, and its differential at 0 ∈ H 2 (X, C) usually vanishes. 1.6 Calogero-Moser deformation. Let V and G be as above. Let Σ ⊂ G be the set of all symplectic reflections in G, and let C be the vector space of all G-invariant functions c : Σ → C. Clearly, dim C = n. We regard C[C ], the polynomial algebra, as a positively graded algebra by assigning degree 2 to linear polynomials. In [EG], P. Etingof and the first author have constructed a certain flat graded Poisson C[C ]-algebra B of degree 2, which gives a graded n-parameter defor- mation of C[V ]G . Let Bc denote the specialization of B at a point c ∈ C , and put Mc = Spec Bc . Thus, {Mc}c∈C is a flat family of Poisson affine al- gebraic varieties and, by construction, we have M0 = V /G. We will refer to Mc as a Calogero-Moser variety with parameter c, and we call the pro jection M := Spec B ։ C the Calogero-Moser deformation of V /G.2 According to Proposition 1.16, there exists a canonical classifying map κ : C → S ⊂ H P 2 (C[V ]G ) such that the Calogero-Moser deformation is obtained by pull-back from the universal deformation. We will see that the map κ is never bijective. The best we can prove is provided by the following Theorem 1.18. If the Calogero-Moser space Mc is smooth for generic values of the parameter c ∈ C then: (i) The base S of the universal Poisson deformation of V /G coincides with the whole vector space H P 2 (C[V ]G ). (ii) The classifying map κ : C → H P 2 (C[V ]G ) is surjective and generical ly ´etale. 2Our terminology is motivated by the special case where G = Sn is the symmetric group, acting diagonally on V = Cn⊕Cn by permutation of coordinates. In that case, the deformation Mc of V /Sn is known (see [EG] and references therein) to be the usual Calogero-Moser space. 11 The proof of this theorem is contained in Section 6. Part (ii) of the Theorem says, roughly speaking, that generically, every fiber of the universal deformation is isomorphic to Bc for a suitable value of the parameter c ∈ C . Moreover, for general c, there exists only a finite number of other values c′ ∈ C such that Bc ∼= Bc′ . This is what we mean by saying that the Calogero-Moser deformation is “versal”. It is not unreasonable to conjecture that the claim of Theorem 1.18 holds in the general situation. Even stronger, we propose the following. Conjecture 1.19. For an arbitrary symplectic quotient singularity V /G, the classifying map C → S of the Calogero-Moser deformation M/C is finite. The conjecture is motivated by analogy with the case of a symplectic reso- lution X → V /G and the corresponding deformation (XB )aff . The parameter spaces of these two deformations are the same vector spaces with the same grading. Moreover, we will see in Section 3 that they actually coincide when dim V /G = dim V = 2. In general, view V /G as a C∗ -variety, with C∗ -action being induced from the natural one on the vector space V . Then, the relation between symplec- tic resolutions of V /G and the Calogero-Moser deformation is provided by the following theorem. Theorem 1.20 (Symplectic resolutions and Calogero-Moser). Let X → Y = V /G be a symplectic resolution of a symplectic quotient singularity V /G. • Let Y /S be the universal Poisson deformation of the Poisson variety Y , • Let XB be the deformation over B = H 2 (X, C) provided by Theorem 1.13, • Let M/C be the Calogero-Moser deformation, and denote by κ : C → S its classifying map. Then for c ∈ C general enough, the image κ(c) ∈ S is generic in the sense of Theorem 1.13 (iii). More precisely, for every point b ∈ B lying over κ(c) ∈ S , the canonical map Xb → Yκ(c) is an isomorphism. Put eC = C ×S B and let ψ : eC → B and ϕ : eC → C be the natural pro jections. Then the statement of Theorem 1.20 can be summarized by the following commutative diagram. π eC M eC := eC × C )aff (X eC ∼ M )SSSSSSSSSSSSSSSSS woooooooooooooo (XB )aff 'PPPPPPPPPPPPPPPP ukkkkkkkkkkkkkkkkkkkk ψ ϕ / S = H P 2 (C[V ]G ) τ κ eC B ϕ×idM Calogero- Moser M C X eC := eC × B (XB ) π=π B XB 12 ) w u ' o o / o o The proof of Theorem 1.20 is contained in Subsection 5.3. Corollary 1.21. The existence of a symplectic resolution X → V /G implies that the generic fiber Mc of the Calogero-Moser deformation M/C is smooth. Proof. By definition we have Mc ∼= Yκ(c) , where YB = (XB )aff . The latter is isomorphic, for b ∈ B such that κ(c) = τ (b) ∈ S is general enough, to Xb , which is smooth. (cid:3) 1.7 Proofs of Theorem 1.1 and Theorem 1.2. Let V be a symplectic vector space, let G ⊂ S p(V ) be a finite group, and let X → Y = V /G be a crepant, hence symplectic, resolution. To prove Theorem 1.1, note that in the assumptions of the Theorem, the ex- istence of X implies by Corollary 1.21 that the generic fiber Mc of the Calogero- Moser deformation M/C is smooth. This contradicts [Go, Proposition 7.3], see also [EG, Proposition 16.4(ii)]. To prove Theorem 1.2, consider the universal deformation Y /S and the deformation X/B provided by Theorem 1.13. Consider a general fiber Mc of the Calogero-Moser deformation M/C . Choose a point b ∈ B lying over κ(c) ∈ S . By definition, we have isomorphisms of fibers Mc ∼= Yκ(c) ∼= Xb , which induce algebra isomorphisms of the rational cohomology: H q (Mc , Q) ∼= H q (Yκ(c) , Q) ∼= H q (Xb , Q) ∼= H q (X, Q), where the last isomorphism is due to Theorem 1.13 (ii). By Corollary 1.21, the generic fiber Bc is smooth. By [EG, Theorem 1.8(i)], this implies that the left-hand-side is isomorphic to gr q (ZG). (cid:3) Acknowledgments. We are grateful to V. Baranovsky, R. Bezrukavnikov, P. Etingof, and V. Ostrik for many useful discussions. We also thank Y. Ruan for bringing the question of the validity of Lemma A.16 to our attention. The second author was partially supported by CRDF Award RM1-2354-MO02. 2 Generalities on Poisson deformations. In this section, we review basic results on the deformation theory of Poisson algebras that will be used later. We have been unable to find an adequate reference in the literature, so in the Appendix to this paper the reader may find some details of proofs and precise definitions. Convention. Given a commutative C-algebra A, we write HomA , ⊗A , and Λk A for Hom, tensor product and k -th wedge product over A, respectively. If no subscript A is indicated, then the corresponding functors are understood to be taken over C, e.g. ⊗ = ⊗C . 13 We let TM denote the tangent sheaf (or tangent bundle) of a smooth algebraic variety M . Further, given a finitely generated commutative algebra A we write Ω1A and T (A) = T (Spec A) for the A-modules of (global) Kahler differentials and vector fields on the scheme Spec A, respectively. 2.1 Poisson cohomology. For smooth Poisson manifolds, the notion of Pois- son cohomology (in the differential geometric setting) is due to J.-L. Koszul [Ko] and J.-L. Brylinski [Br]. In the algebraic setting, the Poisson cohomology has been introduced by B. Fresse [Fr]. We review and extend it below (relations with deformation theory were not discussed in [Fr]). For a more general for- malism of deformation theory of an algebra over an arbitrary operad the reader may consult [KS]. Let A be a finitely generated commutative algebra. The standard Lie bracket of vector fields extends to the so-called Schouten bracket {−, −} on Λ q AT (A), the space of polyvector fields. It is well-known that if Spec A is smooth then any Poisson structure on A defines (and is defined by) a bivector field Θ ∈ Λ2 AT (A) which satisfies the integrability condition {Θ, Θ} = 0. Given a smooth Poisson algebra A one defines a map AT (A) → Λ q+1 d : Λ q a 7→ da := {Θ, a} . A T (A) We have d ◦ d = 0. Taking d as the differential makes Λ q AT (A) into a com- plex. The cohomology groups of this complex are called the Poisson cohomology groups of the algebra A and denoted by H P q (A). , It turns out that this formalism can be extended to arbitrary, not necessarily smooth Poisson algebras A. The precise constructions are a little bit technical; we give them in full in the Appendix. Here we only describe the end result. Recall that for an arbitrary finite-type commutative algebra A, one can define the so-called Harrison complex Har q (A), a certain canonical complex of free A-modules respresenting the cotangent complex of the algebra A. We recall the precise construction of Har q (A) in Subsection A.1. Here we only note that when the algebra A is smooth, the complex Har q (A) has non-trivial cohomology only in degree 0, and this non-trivial cohomology module is isomorphic to the module Ω1A of Kahler differentials. Let now A be an arbitrary Poisson algebra, not necessarily smooth. Heuris- tically, to extend the Brylinski construction to A, one replaces everywhere the module Ω1A of Kahler differentials with the Harrison complex Har q (A). More precisely, one considers the exterior powers Λk A Har q (A) of the complex Har q (A) of flat A-modules and sets (2.1) DP q,k (A) ∼= HomA (Λk k ≥ 0. A Har q (A), A), This defines a canonical bigraded vector space DP q, q (A) and a differential d : DP q, q (A) → DP q+1, q (A). In particular, we have DP 0,2 (A) ∼= HomA (Λ2A ⊗ A, A) = HomC (Λ2A, A). 14 There is a natural Gerstenhaber bracket on DP q, q (A), see (A.8): (A) → DP p+p′ ,q+q′ −1 (A), {−, −} : DP p,q (A) ⊗ DP p′ ,q′ that makes DP q, q (A) a DG Lie algebra (with shifted grading). The Pois- son structure on A defines (and is defined by) an element Θ ∈ DP 0,2 (A) = Hom(Λ2A, A) satisfying dΘ = 0 and {Θ, Θ} = 0. We call this element the Pois- son cochain. Given such an element, one defines the differential δ : DP q, q (A) → DP q, q+1 (A) by setting , δ : DP q, q (A) → DP q, q+1 (A) a 7→ {Θ, a}. Thus we have a bicomplex DP q, q (A) with differentials d, δ . We define Poisson cohomology of A to be the cohomology groups H P q (A) of the total complex associated with this bicomplex, with respect to the total differential d + δ . If the algebra A is smooth, then we have a quasiisomorphism Har q (A) ∼= Ω1A, so that DP q,k (A) ∼= Hom(Λk AΩ1A, A) ∼= Λk AT (A), and the general Poisson cohomology complex DP q (A) is quasiisomorphic to the Brylinski complex hΛ q AT (A), di. If the algebra A is a Poisson graded algebra in the sense of Subsection 1.3, then the Poisson cohomology bicomplex DP q, q (A) acquires an additional grad- ing, called the A-grading. However, since the Poisson cochain Θ is of degree l with respect to the grading, the differential δ : DP q, q (A) → DP q+1, q (A) does not preserve the A-grading but rather shifts it by l. To cure this, we redefine the A-grading by shifting it by (k − 1)l on DP k, q (A). 2.2 Globalization. The Poisson cohomology complex DP q (A) becomes very simple if the scheme A is not only smooth but also symplectic. In that case the canonical isomorphism Ω1A ∼= T (A) extends to an isomorphism Ω q (A) ∼= Λ qT (A) between the Brylinski complex Λ qT (A) and the de Rham complex Ω q (A). Thus the Poisson cohomology H P q (A) coincides with the de Rham cohomology H q (Spec A) of the scheme Spec A. It turns out that several features of the de Rham cohomology formalism extends to the general case. Firstly, one can define Poisson cohomology with coefficients, which is anal- ogous to the de Rham cohomology with coefficients in a local system (or, more generally, in a D-module). For this one defines a Poisson module M over an arbitrary Poisson algebra A in an natural way. Then to every Poisson module M one associates a canonical bicomplex DP q, q (A, M ) called the cohomology complex with coefficients in M . The algebra A is a Poisson module over itself, and we have DP q, q (A, A) = DP q, q (A). Secondly, one generalizes the notion of Poisson cohomology to the scheme case. One defines a Poisson scheme X and a Poisson sheaf F of OX -modules in the natural way. To a Poisson scheme X with a Poisson sheaf F (or, more generally, to a complex F q of Poisson sheaves) one associates a canonical com- plex HP q (X, F q ) of Zariski sheaves on X called the local Poisson cohomology 15 complex with coefficients in F q . In the particular case F = OX , one obtains the local Poisson cohomology complex HP q (X ) = HP q (X, OX ). When the Pois- son scheme X is smooth and symplectic, a Poisson sheaf on X is the same as a D-module, the Poisson cohomology complex HP q (X ) is the de Rham com- plex of the scheme X , and the Poisson cohomology complex with coefficients HP q (X, F ) is the de Rham complex of the D-module F . Taking the hyperhomology, one obtains the groups H P q (X ) = H q (X, HP q (X )) of global Poisson cohomology of the scheme X and the groups (2.2) k ≥ 0. H P p (Y ) ⊗ H P q (Z ), H P q (X, F q ) = H q (X, HP q (X, F q )) of global Poisson cohomogoly of X with coefficients in the complex F q . When the scheme X = Spec A is affine and the complex F q comes from a complex M q of Poisson A-modules, we have H P q (X, F q ) ∼= H P q (A, M q ). When the scheme X is smooth and symplectic, H P q (X, F q ) is the singular cohomology H q (X, F q ) with coefficients in the D-module F q . When X = Y × Z is a product of two Poisson schemes Y and Z , we have the Kunneth formula (Proposition A.11(i)) H P k (X ) ∼= Mp+q=k The notion of a Poisson sheaf is sufficiently functorial; in particular, for any morphism f : X → Y between Poisson schemes, the direct image f∗F of a Poisson sheaf F on X is a Poisson sheaf on Y . Moreover, one can represent the direct image R q f∗F in the derived category by a complex of Poisson sheaves. In the particular case of an open embedding j : U ֒→ X , we obtain a Poisson structure on the direct image R q j∗OU . Moreover, in this particular case we have HP q (X, R q j∗OU ) ∼= R q j∗ (HP q (U )). One can also obtain a Poisson structure on the third term i!OZ in the exact triangle i!OZ −−−−→ OX −−−−→ R q j∗OU −−−−→ where Z ⊂ X is the closed complement to U ⊂ X and i : Z ֒→ X is its embedding. We define the Poisson cohomology H P q Z (X ) of the scheme X with Z (X ) ∼= H P q (X, i!OZ ). When X is smooth supports in Z ⊂ X by setting H P q and symplectic, this is the ordinary singular cohomology H q Z (X ) with supports in Z . By definition, for a general Poisson X we have the canonical exact triangle (2.3) Z (X ) −−−−→ H P q (X ) −−−−→ H P q (U ) −−−−→ H P q Note that the scheme Z actually enters into this construction only through its open complement U ⊂ X . In particular, there is no need to assume that Z ⊂ X is a Poisson subscheme in any sense. 16 The reader will find the precise definitions and statements on Poisson coho- mology in the Apenndix, with all the proofs. The only statements that we will actually use in the main body of the paper are contained in Proposition A.11, Corollary A.12, and in Lemma A.9. The reason we are interested in Poisson cohomology is its role in the study of Poisson deformations of a Poisson scheme X . This role is completely anal- ogous to the role of the standard cotangent complex Ω q (X ) and the groups Ext q (Ω q , OX ) in the standard deformation theory of a scheme X . The analogy can be pushed quite far. We will use Poisson deformation theory in Subsec- tion A.8 to prove Theorem 1.10. 3 The case of dim V = 2. Before we proceed to the study of general symplectic quotient singularities V /G, we need to consider the particular case dim V = 2. This is the case of the so-called McKay correspondence. Starting with the paper [McK], it has been studied excessively by many authors. We recall here some of the results. Let V be a complex vector space of dimension dim V = 2. To every finite subgroup G ⊂ S p(V ), one canonically associates a simply-laced root system whose rank is equal to the number of non-trivial conjugacy classes in G. Let g be the simple Lie algebra associated to this root system. The Cartan subalgebra h ⊂ g is naturally dual, h ∼= C ∗ , to the base C of the Calogero-Moser deformation M/C of the quotient variety V /G. The conjugacy classes of elements g ∈ G define a basis in the vector space C ∼= h∗ , which is in fact a basis of simple roots. The dual space g∗ is naturally a Poisson scheme. The nilpotent cone N ⊂ g∗ is a Poisson subscheme consisting of a finite number of coadjoint orbits, all of which are symplectic. There exists a unique orbit Nsubreg ⊂ N of dimension dim Nsubreg = dim N − 2, called the subregular nilpotent orbit. Take an arbitrary element n ∈ Nsubreg and let p ⊂ g∗ be an affine space passing through n ⊂ g∗ and transversal to Nsubreg ⊂ g∗ . The Poisson structure on g∗ induces a Poisson structure on the affine space p. The intersection p ∩ N ⊂ p is a Poisson subscheme. It turns out that the Poisson scheme p ∩ N is naturally isomorphic to the quotient V /G. Moreover, the Poisson algebra of functions on p has a large center, so that in fact we have a Poisson deformation p/S over a base S of dimension dim S = rk g. The base S of the deformation p/S is canonically isomorphic to the quotient S = h∗/W of the dual Cartan algebra h∗ by the Weyl group W . In particular S is smooth. The subscheme Y ∼= p ∩ N ⊂ p is the fiber of p/S over the point 0 ∈ S ∼= h∗/W . In other words, we obtain a Poisson deformation p/S of the Poisson scheme V /G over a smooth base S of dimension dim S = rk g. The natural C∗ -action on the vector space g∗ by dilatations induces the natural grading on the Poisson algebra C[V /G]. The deformation p/S is a graded deformation. The grading on S ∼= h∗ /W is induced by the C∗ -action by dilatations on h∗ . From now on, set Y = V /G. It is known that the deformation p/S is a 17 miniversal deformation of Y in the category of affine schemes, with the Poisson structure forgotten. In particular, the canonical map ToS → Ext1 (Ω q (Y ), OY ) between the Zariski tangent space ToS and the group which classifies infinitese- mal deformations of the scheme Y is an isomorphism. Moreover, the total space p and the base S of the deformation p/S are smooth; therefore the scheme Y is a complete intersection, and the cotangent complex Ω q (Y ) only has non-trivial cohomology in degree 0. Using these facts, we can now compute the Poisson cohomology groups H P 1 (Y ) and H P 2 (Y ). Lemma 3.1. We have H P 1 (Y ) = 0, and H P 2 (Y ) ∼= Ext1 (Ω(Y ), OY ) ∼= ToS. Moreover, the degrees of the natural grading on the group H P 2 (Y ) coincide with the exponents of the Weyl group W . Proof. Denote A = C[Y ]. By definition of the Poisson cohomology bicomplex DP q, q (A) we have DP 0,0 (A) ∼= A, DP 0,k (A) = 0 for k ≥ 1 and DP 1, q (A) ∼= RHom q (Ω1A, A). In particular, there exists a canonical map κ : H P 2(A) → Ext1 (Ω1A, A) induced by the natural pro jection DP q,k (A) → DP 1,k (A), k ≥ 1. On the level of deformation theory, the map κ corresponds to forgetting the Poisson structure. Since the universal deformation p/S of the scheme Y does admit a Poisson structure, the map κ : H P 2 (A) → Ext1 (Ω1A, A) is surjective. Thus to prove the Lemma, it suffices to prove that H P 1 (A) = 0 and that κ is an injective map. Consider the spectral sequence associated to the bicomplex H P q, q (A). We have 1 = Extq (Λp E p,q p, q ≥ 0. AΩ q (A), A), The only term which contributes to H P 1 (A) is the term E 1,0 ∞ . The only non- trivial terms which contribute to HP 2 (Y ) are E 2,0 and E 0,2 1 , E 1,1 1 . The term 1 1 = Ext2 (A, A) vanishes. The term E 1,1 E 0,2 is precisely Ext1 (Ω1A, A). More- 1 over, the claim about the gradings on this term follows from the identification S ∼= h∗/W . To prove that the map H P 2(Y ) → E 1,1 is injective, it suffices to 1 ∞ = E 1,0 ∞ vanishes. Thus is suffices to prove that E 2,0 prove that the term E 2,0 ∞ = 0. We will prove that already E p,0 2 = 0 for every p ≥ 1. Indeed, denote by j : U ֒→ Y the embedding of the open complement U = Y \ {0} ⊂ Y to the origin o ∈ Y . By definition we have 1 ∼= Hom(ΛpΩY , OY ) ∼= Hom(ΛpΩY , j∗OY ) E p,0 ∼= HomU (ΛpΩU , OU ) ∼= H 0 (U, ΛpTU ). 18 The quotient map π : V → Y = V /G is ´etale over U , so that H 0 (U, ΛpTU ) ∼= H 0 (π−1 (U ), ΛpTV )G . Moreover, since π−1 (U ) ⊂ V is the complement to a point in a smooth scheme, the right-hand side is isomorphic to H 0 (V , ΛpTV ). The differential d1 : E p,0 1 → E p+1,0 in the spectral sequence is induced by the Poisson differential on the space 1 2 ∼= H P p (V )G . Since V is smooth and symplectic, and H 0 (V , ΛpTV ). Hence, E p,0 H p (V , C) = 0 for p ≥ 1, this implies that E p,0 2 = 0 for p ≥ 1. This finishes the proof. (cid:3) In particular we see that p/S is the universal Poisson deformation of the Poisson scheme Y = V /G. Recall the Calogero-Moser deformation M/C introduced in [EG]. The de- formation M/C does not coincide with the universal deformation p/S but there is a Cartesian square M −−−−→ p y y C −−−−→ S Here C ∼= h, S ∼= h/W , and the bottom row in the square may be identified with the canonical pro jection C = h ։ h/W = S (which is not an isomorphism). Thus, Lemma 3.1 provides a canonical identification H P 2 (Y ) ∼= C/W (this is an identification of algebraic varieties, not of vector spaces). The identification H P 2 (Y ) ∼= C/W yields: dim H P 2 (Y ) = dim C = n, the equality claimed in Propositon 1.16. To study the higher-dimensional case we will need the following twisted version of this equality. Lemma 3.2. Let G ⊂ G′ ⊂ S p(V ) be two finite groups acting on the two- dimensional symplectic space V . Assume that G ⊂ G′ is normal, and let H = G′/G be the quotient group, which act natural ly on the quotient variety V /G and on the vector space C . Then we have: dim H P 2 (Y )H = dim C H . Proof. The group H preserves the root system ∆ ∈ C corresponding to G ⊂ S p(V ) and the base of simple roots defined by the conjugacy classes in G. More- over, H commutes with the action of the Weyl group w. We have H P 2 (Y ) ∼= C/W . The space H P 2(Y )H of H -invariant vectors is the subvariety (C/W )H ⊂ C/W of H -fixed points in C/W . Therefore it suffices to use the following stan- dard result. Lemma 3.3. Let ∆ ∈ CR be a root system with Weyl group W , and let H be a finite group of automorphisms of the root system ∆ which preserves a Weyl chamber C + R ⊂ CR . Let C = CR ⊗R C be the complexification of the real vector space CR . Then the quotient map C → C/W induces a surjective finite map C H → (C/W )H onto the set of H -fixed points in the quotient variety C/W . 19 R + √−1CR is a fundamental It is well-known that the set C + = C + Proof. domain for the W -action on the vector space C , so that we have an isomorphism C + ∼= C/W . Since H preserves C + , it induces an isomorphism H ∼= C + ∩ C H ∼= (C/W )H , C + which gives a section of the map C H → (C/W )H . To describe the Calogero-Moser deformation M/C more functorially, we recall that the quotient V /G admits a canonical smooth symplectic resolution X → Y (coming from the Springer resolution N → N of the nilpotent cone in g∗ ). Moreover, the scheme X has a universal deformation X/C , and the total space X is symplectic over C . Therefore the algebra H 0 (X, OX ) of global functions on X has a natural Poisson structure. This algebra coincides with the algebra of functions on the Calogero-Moser deformation M/C . Geometrically, we have a natural birational pro jective map π : X → M and an isomorphism π∗OX ∼= OM . The cohomology group H 2 (X, C) of the resolution X is naturally identified with the base C of the deformation X/C . The map π is compatible with the pro jections to the base C . Moreover, it is an isomorphism over a generic point c ∈ C . These facts taken together immediately imply Propositon 1.16, Theorem 1.18 and Theorem 1.20 in the case dim V = 2. (cid:3) 4 The computation of H P 2(V /G). Let V be an arbitrary finite-dimensional symplectic vector space, and let G ⊂ S p(V ) be a finite subgroup. In this section we will compute the Poisson co- homology groups H P 1 (V /G) and H P 2 (V /G) of the quotient Y = V /G. In particular, we will prove Proposition 1.16. First, we introduce some notation. Notice that the vector space V is nat- urally stratified by subspaces V H ⊂ V of H -invariant vectors for various sub- groups H ⊂ G. All these subspaces are symplectic, hence even-dimensional. This induces a stratification of the quotient variety Y = V /G. The strata of the stratification are known to be smooth symplectic locally-closed subvarieties in V /G, in particular have even dimension. Moreover, these strata turn out to be exactly the symplectic leaves of the standard Poisson structure on V /G, cf. e.g. [BG]. In more detail, let Γ ⊂ G be a subgroup which is the isotropy group of an element of V . Then V Γ ⊂ V is a nonzero vector subspace such that the symplectic form on V restricts to a nondegenerate 2-form on V Γ . Further, let U Γ ⊂ V be the set of points of V whose stabilizer (in G) is equal to Γ. It is clear that U Γ ⊂ V Γ , moreover, it is known from the theory of finite group actions that U Γ is a non-empty Zariski open, hence dense, subset in V Γ . The strata of the stratification of V /G that has been mentioned in the previous paragraph are defined to be the images under the pro jection V ։ V /G of the sets of the form U Γ , as Γ varies inside G. 20 It is straightforward to verify that N (Γ), the normalizer of Γ in G, preserves the set U Γ . Furthermore, the resulting N (Γ)/Γ-action on U Γ is free, and the pro jection V ։ V /G induces an isomorphism of U Γ/(N (Γ)/Γ) with its image in V /G, that is, with the corresponding stratum of the stratification. Now, recall the notation Y = V /G and write U ⊂ Y for the complement to the union of all the strata in Y of codimension ≥ 4. Thus we have U = U0 ` (∪i≥1 Ui ), where U0 is the unique open stratum in Y , and Ui , i = 1, 2, . . . , are all the codemension two strata of the stratification. By the earlier discussion, each stratum is the image under the pro jection V ։ V /G of the set U Gi ⊂ V := V Gi where Gi ⊂ G is a subgroup such that the fixed point set Vi is a codimension 2 vector subspace in V and, moreover, such that the set U Gi = {v ∈ V isotropy group of v = Gi } is a Zariski dense subset in V Gi . Two subgroups as above give rise to the same stratum Ui if and only if they are conjugate within G. Let G′ i := N (Gi ) ⊂ G denote the normalizer of the subgroup Gi , and Hi := G′ i /Gi the quotient group. As we have explained above, the group G′ i preserves the set U Gi , the induced Hi -action on U Gi is free; furthermore, we have an isomorphism U Gi /Hi ∼−→ Ui . For each i ≥ 1, let Wi denote the annihilator of Vi with respect to the symplectic form. Thus, dim Wi = 2, and there is a canonical G′ i -stable direct sum decomposition V = Vi L Wi . Let Yi = Wi /Gi be the quotient variety. The group Hi acts on the variety Yi and preserves the origin o ∈ Yi . The direct sum decomposition V ∼= Vi ⊕ Wi induces a map ηi : (Vi × Yi )/Hi → Y . We have a Zariski open subset U ′ i ⊂ Vi such that ηi maps (U ′ i × {o})/Hi isomorphically onto Ui and, moreover, is ´etale in a Zariski open neighborhood i /Hi ) × {o} ⊂ (Vi × Yi )/Hi . We note for further record that Vi r U ′ of Ui = (U ′ i is a finite union of vector subspaces in Vi of (complex) codimension ≥ 2. It follows that H l (U ′ i , C) = 0, for l = 1, 2. As in Subsection 1.5, let C be the space of all G-invariant functions on the set Σ of symplectic reflections in G. By definition we have dim C = n. For every i, denote by Ci the vector space of Gi -invariant C-valued functions on the group Gi , and let Bi ⊂ Ci be the subspace of G′ i -invariant functions. The group Hi acts on the space Ci and we have Bi = C Hi i ⊂ Ci . Every element g ∈ Gi r {1} pointwise fixes the codimension two subspace Vi , hence is a symplectic reflection. Therefore we have a natural restriction map C → Ci . This map factors through a map C → Bi = C Hi i ⊂ Ci . The map C → Bi is surjective. Indeed, it suffices to check that two elements g1 , g2 ⊂ Gi conjugate in G are already conjugate in G′ i ; and if g · g1 · g−1 = g2 , then g preserves the subspace Vi = V g1 ⊂ V , hence lies in G′ i . Moreover, every symplectic reflection g ∈ G lies in one of the subgroups Gi ⊂ G, – namely, the stabilizer of the invariant subspace V g ⊂ V . Therefore the pro jections C → Bi give a natural splitting C ∼= Mi Bi = Mi C Hi (4.1) . i We recall that the quotient singularity Y = V /G carries a natural Poisson structure, so that we have the Poisson cohomology groups H P q (Y ). Moreover, 21 k = 1, 2, 3. C[Y ] is a positively-graded Poisson algebra of degree 2, and this grading induces a grading on H P q (Y ). Proof of Proposition 1.16. The scheme Y = V /G is normal, and the smooth locus U0 ⊂ Y carries a non-degenerate symplectic form. Moreover, by [W] the scheme Y is Gorenstein, hence Cohen-Macaulay. By definition, the complement to the open subset U ⊂ Y is of codimension ≥ 4. Therefore by Lemma A.9, the natural map H P k (Y ) → H P i (U ) is an isomorphism for k = 1, 2. Since the open subscheme U0 ⊂ U is smooth and symplectic, we have H P q (U0 ) ∼= H q (U0 , C). But U0 = (V \ Z )/G is the quotient of the comple- ment in the vector space V to some closed subscheme Z ⊂ V of codim ≥ 2. Therefore the singular cohomology H q (U, C) = H q (V \ Z )G vanishes in low degrees, H P k (U0 ) ∼= H k (U0 , C) ∼= H k (V \ Z, C)G ∼= H k (V , C)G = 0, (U ) → H P k (U ) is We conclude that for k = 1, 2 the natural map H P k U \U0 an isomorphism, and the isomorphism H P k (Y ) ∼= H P k (U ) factors through a canonical isomorphism H P k (Y ) ∼= H P k U \U0 (U ). The complement U \ U0 is the disjoint union of the closed strata Ui ⊂ U , i ≥ 1, and for every i ≥ 1, we have a map ηi : (Vi × Yi )/Hi → Y . The map ηi identifies (U ′ i /Hi ) ⊂ (Vi × Yi )/Hi with the stratum Ui ⊂ Y , and it is ´etale in an open neighborhood of Ui ∼= (U ′ i /Hi ) ⊂ (Vi × Yi )/Hi . By Corollary A.12(i), we have Ui (U ) ∼= H P q Ui ((U ′ H P q i × Yi )/Hi ). By Proposition A.11(iii), the right-hand hand side is naturally isomorphic to i × Yi )/Hi ) ∼= (cid:16)H P q i × Yi )(cid:17)Hi Ui ((U ′ (U ′ H P q U ′ i Hence, we obtain a canonical direct sum decomposition i × Yi )(cid:17)Hi U \U0 (U ) ∼= Mi (cid:16)H P k H P k (Y ) ∼= H P k (U ′ U ′ i The Kunneth formula yields i × Yi ) ∼= M0≤l≤k (U ′ i ) ⊗ H P k−l H P l (U ′ o Further, recall that by construction U ′ i is an open subset in Vi whose comple- In particular, U ′ ment has complex codimension ≥ 2. i is smooth symplectic and connected. Therefore, H P l (U ′ i ) = H l (U ′ i , C); moreover, this group is 1- dimensional (and has trivial action of the group Hi ) if l = 0, and is equal to zero if l = 1, 2. We conclude that the product decomposition (A.16) reduces to an Hi -equivariant isomorphism i × Yi ) ∼= H P k i ) ∼= H P k (U ′ o (Yi ) ⊗ H P 0 (U ′ o (Yi ), k ≤ 2. H P k U ′ i . . (Yi ). H P k U ′ i 22 H P k o (Yi )Hi . Therefore for k = 1, 2 we have H P k (Y ) ∼= Mi But for every i ≥ 0, the complement Yi \ {o} ∼= (Wi \ {0})/Gi is smooth, symplectic, and satisfies H k (Yi \ {0}) ∼= H k (Wi \ {0})Gi = 0, k = 1, 2, 3. Therefore H P k (Yi \ {o}) = 0 for k = 1, 2, 3, and we have an isomorphism o (Yi ) ∼= H P k (Yi ). H P k Collecting together all of the above, we conclude that for k = 1, 2 there exists a natural isomorphism H P k (Y ) ∼= Mi (cid:0)H P k (Yi )(cid:1)Hi . But we have already computed H P k (Yi ), k = 1, 2 in Section 3. Lemma 3.1 immediately implies that H P 1 (Y ) = 0, and Lemma 3.2 shows that dim H P 2(Yi )Hi = Xi dim H P 2(Y ) = Xi By (4.1) the right-hand side is equal to dim C . Moreover, the natural grading on each of the H P 2 (Yi ) has positive degrees. (cid:3) This Proposition of course applies to any pair hV , Gi. In particular, one can take the trivial group G = {e}, and obtain, as expected, the equality H P 2 (V ) = 0. Taking the quotient by some non-trivial G ⊂ S p(V ) increases the second Poisson cohomology group and creates new deformations. These deformations are indeed new: they do not lift to G-equivariant deformations of the symplectic vector space V . We do not formulate this fact precisely because we will not need it. However, in Section 5 we will need the following claim. dim C Hi i . Lemma 4.1. Let Y /S be a Poisson deformation of a symplectic quotient sin- gularity V /G over a local Artin base S . Assume that there exists a Poisson deformation V /S of the vector space V and a map η : V /S → Y /S which ex- tends the quotient map η : V → Y . Then the Poisson cocycle ΘY ∈ H P 2 (Y /S ) is trivial. Proof. Assume that ΘY 6= 0. Let m ⊂ C[S ] be the maximal ideal, and let k be the largest integer such that ΘY = 0 mod mk . Then ΘY mod mk+1 is a non-trivial element in the Poisson cohomology group H P 2 (Y /S, mk /mk+1 ) ∼= H 2 (Y ) ⊗C (cid:0)mk /mk+1 (cid:1) . But the group H P 2 (Y ) admits a natural map η : H 2 (Y ) → H 2 (Y , η∗OV ) which identifies it with the direct summand H 2 (Y , η∗OV )G ⊂ H 2 (Y , η∗OV ). Therefore η(ΘY ) must be a non-trivial class in the group H 2 (Y , η∗OV ) ⊗C (cid:0)mk /mk+1 (cid:1) . 23 This is impossible. Indeed, by the functoriality of the Poisson cohomology, the class η(ΘY ) comes from the class ΘV ∈ H P 2 (V ) of the Poisson cocycle on the vector space V , and the group H P 2 (V ) is a trivial group. (cid:3) In truth, under the assumptions of the Lemma the whole deformation Y /S must be trivial, not only its Poisson cocycle ΘY . But the proof of this fact is slightly harder. We settle for the weaker version to save space. One final remark is the following: since the Poisson algebra C[Y ] is positively graded, Lemma A.15 immediately shows that all the results of this Section are valid not only for Y , but also for its completion bY at the origin o ∈ Y . 5 Resolutions. In this section we will prove Theorem 1.13 and Theorem 1.20. Below, we use the notation H q (M ) = H q (M , C) for the singular cohomology of a variety M with complex coefficients. The homology of M is usually taken with rational coefficients, i.e., H q (M ) = H q (M , Q). 5.1 Geometry of the resolution. We will need several general facts on the geometry of symplectic resolutions. Most of them are well-known. The reader can find the proofs, for instance, in the papers [K1], [K3]. Let X be an irreducible smooth variety over C equipped with a closed nowhere-degenerate 2-form ω , and let Y = X aff be its affinization. Assume that the natural map π : X → Y is pro jective and birational. By definition Y is normal, and we have π∗OX ∼= OY . The canonical bundle KX of the mani- fold X is trivial (it is trivialized by the top power of the symplectic form ω ). Therefore map π : X → Y is one-to-one over the smooth locus Yo ⊂ Y . By the Grauert-Riemenschneider Vanishing Theorem, we have (5.1) H i (X, OX ) = H 0 (Y , Riπ∗OX ) = H 0 (Y , Riπ∗KX ) = 0 for i ≥ 1. Considering the exponential exact sequence on X , one easily deduces that R1π∗ZX = 0, where ZX is the constant sheaf on X . The symplectic form ω on the smooth variety X induces a Poisson structure on the algebra C[Y ] ∼= H 0 (X, OX ). Further, the resolution X → Y is known, see e.g. [K2], to be semismall (that is, dim X ×Y X = dim X ). This implies that the Leray spectral sequence which computes the group H 2 (X, Q) degenerates, and we have a canonical short exact sequence of rational cohomology groups (5.2) 0 −−−−→ H 2 (Y , Q) −−−−→ H 2 (X, Q) −−−−→ H 0 (Y , R2π∗Q) −−−−→ 0 Set H2 (X/ Y ) := Ker(H2 (X, Q) → H2 (Y , Q)) the subgroup in the homology group H2 (X, Q) dual to H 0 (Y , R2π∗Q). The following fact will be very important. 24 Lemma 5.1. Let [ω ] ∈ H 2 (X, C) be the cohomology class of the symplectic form ω . Then for every homology class α ∈ H2 (X/ Y ), we have hα, [ω ]i = 0, and [ω ] = π∗ [ω ]Y , (5.3) for some cohomology class [ω ]Y ∈ H 2 (Y , C). Moreover, the class [ω ]Y depends only on the Poisson structure on the variety Y , not on the resolution X . Proof. This is not new, see e.g. [CF], [K3, Lemma 2.2], [WW]. We only give a sketch of the proof. By (5.2), to establish (5.3) it suffices to prove that [ω ] vanishes on all the fibers of the map Y → X . Taking a resolution of singularities, it suffices to prove that f ∗ [ω ] = 0 for every map f : Z → X from a smooth pro jective manifold Z . By Hodge theory, this is equivalent to proving that (5.4) f ∗ [ω ] = 0 ∈ H 0,2(Z ) = H 2 (Z, OZ ), where [ω ] is the complex-conjugate cohomology class. But already [ω ] = 0, since H 2 (X, OX ) = 0 by (5.1). To prove that the class [ω ]Y is canonical, note that its restriction to the non-singular part U ⊂ Y is represented by an explicit 2-form ωY which is inverse to the non-degenerate Poisson bivector Θ. Thus [ω ]Y U depends only on Θ. Moreover, we know that the form ωY extends to some, hence to an arbitrary smooth resolution X → Y and represents a cohomology class [ω ]Y ⊂ H 2 (X, C) coming from H 2 (Y , C). The class [ω ]Y is completely determined by the form ωY since the map H 2 (Y , C) → H 2 (X, C) is injective. (cid:3) In order to study deformations of the pair hX, ω i, we follow the paper [KV] and introduce the period map. To do this, note that for every smooth deforma- tion X/B over a local Artin base B , the Gauss-Manin connection trivializes the relative de Rham cohomology H 2 DR (X/B ), DR (X/B ) ∼= H 2 (X, C) ⊗C C[B ]. H 2 Therefore the class [ω ] can be canonically considered as a class in the group H 2 (X, C) ⊗C C[B ]. This induces a map P : B → H 2 (X, C) called the period map. The fundamental theorem of [KV] claims that the period map completely determines the deformation. More precisely, for every local Artin scheme B and a map P : B → H 2 (X, C) which sends the closed point o ∈ B to the class [ω ] ∈ H 2 (X, C), there exists a deformation X/B of the pair hX, ω i with the period map P . Moreover, such a deformation X/B is unique up to a non-canonical isomorphism. Set B := H 2 (X, C), and let bB denote the formal neighborhood of [ω ] ∈ H 2 (X, C). This way, one obtains a universal deformation bX/ bB of the pair hX, ω i. Since one has to pass to the limit, the universal deformation bX/ bB is only a formal scheme. 25 5.2 Globalizing the deformation. We can now start proving Theorem 1.13. Our method will be to extend the formal deformation from [KV] to an actual deformation defined over a global base. Throughout the subsection, let X be a convex symplectic manifold with symplectic form ω . We write B := H 2 (X, C), and equip the vector space B with C∗ -action so that z ∈ C∗ acts via multiplication by z−l . Further, let bB denote the formal neighborhood of [ω ] ∈ B , and let m ⊂ C[ bB ] be the maximal ideal of the complete algebra C[ bB ]. Write bX/ bB for the universal formal deformation of the pair hX, ω i. Put Y := X aff . Assume that the Poisson algebra C[Y ] = H 0 (X, OX ) is a finitely-generated positively graded Poisson algebra of degree l > 0, and con- sider the corresponding C∗ -action on the scheme Y . We split the proof of Theorem 1.13 into three Propositions below. Proposition 5.2. The C∗ -action on Y lifts uniquely to a C∗ -action on X/ Y , and the resulting action on X extends to a C∗ -action on bX/ bB . Moreover, we have [ω ] = 0 in H 2 (X ). This result is known, and we refer to [Fu], [K1], and [Ve] for the proofs. We also have the following infinitesimal version of the above proposition. Lemma 5.3. Let Y be an irreducible algebraic variety, and let X → Y be smooth projective semismal l resolution of Y . If the canonical bund le is trivial then every vector field ξ (a derivation of OY ) on Y lifts canonical ly to a vector field on X . Proof. (Compare [CF], where a similar statement is proved for isolated singu- larities, but without the semismallness asumption.) The vector field ξ canoni- cally lifts to a vector field defined outside of the exceptional locus of the map π : X → Y . Bince the variety X is smooth, the sheaf T (X ) of vector fields is reflexive. Therefore it suffices to prove that ξ to the generic point of an arbi- trary exceptional Weil divisor E ⊂ X . Since π : X → Y is semismall, the image π(E ) ⊂ Y of the divisor E is a subvariety of codimension 2. Therefore near the generic point of the subvariety π(E ) ⊂ Y , the variety Y is the product of a smooth variety and an isolated surface singularity, and it suffices to prove the Lemma in the case dim Y = 2. In this case, the triviality of the canonical bundle implies that Y is a Du Val point. Then Lemma easily follows from an explicit construction of the minimal resolution as a transversal slice to a subregular nilpotent orbit, see Section 3. (cid:3) Proposition 5.4. The formal scheme bX extends to an actual scheme X/B de- fined over the whole affine space B = H 2 (X ). The map σ : X → H 2 (X ) is smooth and C∗ -equivariant, and the scheme X is symplectic over B . Moreover, the relative cohomology sheaves R qσ∗Q are constant sheaves on H 2 (X ), and the canonical base change morphism H k (Xb , Q) → (cid:0)Rk π∗Q(cid:1)b 26 is an isomorphism for every point b ∈ B . Proof. To extend the formal scheme bX/ bB to a scheme over the whole B , we repeat the argument of [K3, Lemma 4.2]. Take an ample line bundle L on X and note that, since by (5.1) H 1 (X, OX ) = H 2 (X, OX ) = 0, the bundle L canoni- cally extends to the formal scheme bX/ bB . Moreover, using the same cohomology vanishing it has been shown in [K3] that H 0 ( bX, O bX ), the algebra of global func- tions on bX, is a filtered algebra whose associated graded algebra is isomorphic It follows that H 0 ( bX, O bX ) is a to H 0 (X, OX ) ⊗ C[t], a Noetherian algebra. Noetherian algebra, which is also complete with respect to the m-adic topology. Thus, bY = Spf (H 0 ( bX, O bX )), the formal spectrum, is an affine Noetherian formal scheme flat over bB , and bX is pro jective over bY , with an ample line bundle L. Therefore we can apply the Grothendieck algebraization theorem [EGA, III, Th´eor`eme 5.4.5] to bX/ bY and conclude that the formal scheme bX extends to an actual scheme X over bY (and a posteriori, over bB ). Since the algebra C[B ] of functions on the vector space B = H 2 (X ) is positively graded, we are done by Lemma A.15. Moreover, by construction the scheme X/B is symplectic and smooth, and the map X → B is C∗ -equivariant. DR ( bX/ bB ). Let us now analyze the relative de Rham cohomology sheaves H q For every k ≥ 0, denote by Bk = Spec O bB /mk ⊂ bB the k -th infinitesimal neigh- borhood of the special point in the local scheme bB , and let Xk = bX × bB Bk . By [EGA, III, Th´eor`eme 4.1.5], for every p, q ≥ 0 we have canonical isomorphisms H q ( bX, Ωp ( bX/ bB )) ∼= lim H q (Xk , Ωp (Xk /Bk )), (5.5) ← and the pro jective system on the right-hand side satisfies the Mittag-Leffler condition [EGA, 0, 13.1]. Moreover, for every k , n the de Rham cohomology module H n DR (Xk /Bk ) is a finitely-generated module over C[Bk ], an Artinian algebra. Therefore for every n ≥ 0, the pro jective system H n DR (Xk /Bk ) also satisfies the Mittag-Leffler condition, and by [EGA, 0, Proposition 13.2.3] the DR ( bX/ bB ) ∼= lim← H n maps (5.5) induce canonical isomorphisms H n DR (Xk /Bk ). DR ( bX/ bB ) is finitely We conclude that for every n ≥ 0, the C[ bB ]-module H n generated, complete and separated with respect to the m-adic topology. Since it carries a flat connection – namely, the Gauss-Manin connection – it must be a free C[ bB ]-module. By Lemma A.15, this implies that the relative de Rham cohomology sheaves DR (X/B ) are also free sheaves of C[B ]-modules. Therefore the D-modules H q R qσ∗OX are constant D-modules on B . This means that the sheaves R qσ∗Q are constant sheaves. Finally, let b ∈ B be an arbitrary point with embedding ib : b → B , and let Xb ⊂ X be the fiber of X over b, with embedding Xb ֒→ X denoted by the same letter ib . Since the scheme X/B is smooth, we can apply Poincare 27 duality together with the Proper Base Change Theorem and obtain a canonical isomorphism H q (Xb , Q) ∼= H q (X, i! bQ)[dim B ] ∼= i! bR qσ∗Q[dim B ]. Since the sheaves Rkσ∗Q are locally constant on B , the right-hand side is iso- morphic to (R qσ∗Q)b . (cid:3) Below, by ‘general enough’ or ‘generic’ point of B we mean a point from an appropriately chosen Zariski open dense subset of B . To finish the proof of Theorem 1.13, it remains to prove the following. Proposition 5.5. The map π : X → Y is semismal l. Moreover, for a general enough point b ∈ B , the induced map Xb → Yb of fibers over b in the varieties X/B , resp., Y /B , is an isomorphism. Proof. (Compare [K3, Proposition 4.6].) To prove that π : X → Y is semismall, it suffices to prove that for every closed subvariety Z ⊂ X , the dimension of the generic fiber of the map Z → π(Z ) does not exceed the codimension codim(Z , X). Let Z = X ∩ Z . Since X ⊂ X is the fiber of a smooth map X → H 2 (X ), we have codim(Z, X ) ≤ codim(Z , X). Since the map π : X → Y is semismall, the dimension of the generic fiber of the map π : Z → π(Z ) does not exceed the codimension codim(Z, X ). Therefore the same is true for the generic fiber of the map π : Z → π(Z ). Let b ∈ B . Then, since Y is normal, the variety Yb is also normal. By construction, the map Xb → Yb is pro jective and generically one-to-one. The set of points b ∈ B such that the map Xb → Yb has fibers of dimension > 1 is clearly a closed subset B BAD ⊂ B . Hence, to complete the proof of the Proposition it suffices to prove that B BAD is a proper subset in B . To this end, we choose b ∈ B such that the cohomology class [ω ] = b ∈ H 2 (X ) = H 2 (X, C) does not annihilate any rational homology class [Z ] ∈ H2 (X/ Y ) ⊂ H2 (X, Q) (the latter obviously form a countable subset H2 (X, Q) ⊂ H2 (X, C)). We claim that b 6∈ B BAD , hence, B BAD 6= B . Indeed, if not, then there is a fiber of the map Xb → Yb that contains a pro jective curve Z ⊂ Xb . The corresponding homology class [Z ] ∈ H2 (X/ Y ) ⊂ H2 (X, Q) annihilates ω , i.e., we have h[Z ], [ω ]i = 0, by Lemma 5.1. But that would contradict the choice of b which was made so that [ω ] does not annihilate any rational homology class [Z ] ∈ H2 (X/ Y ). This completes the proof. (cid:3) Remark 5.6. We included the claim that X → Y semismall because it was used in the proof of Lemma 5.3 (if applied to the map X → Y ); this case of the Lemma will be used at one point of the argument later, in Subsection 5.4. We have completed the proof of Theorem 1.13. Proof of Proposition 1.14. Assume that the graded Poisson algebra C[Y ] satisfies the assumptions of Theorem 1.10, so that we have the universal graded Poisson 28 deformation YS /S . Let τ : B → S be the classifying map of the deformation Y /B = Xaff/B . We have to prove that the map τ is a finite map onto an irreducible component of the variety S . Let b ∈ B = H 2 (X ) be a point sufficiently generic so that the map Xb → Yb is an isomorphism. Let S ′ ⊂ S be the irreducible component which contains the image τ (c) of the point b ∈ B . Since Xb → Yb is one-to-one, the Poisson scheme Yb is in fact symplectic, and the Poisson deformation theory for Yb reduces to the deformation theory of the pair hXb = Yb , ω i. Since the deformations X/B , Y /S are universal, this implies that the map τ : B → S ′ is ´etale at the point b. We see that the map τ : B → S ′ is generically ´etale. To prove that it is in fact finite, it suffices to prove that all of its geometric fibers are finite. Moreover, since the map τ is C∗ -equivariant, it suffices to consider the fiber over the origin point o ∈ S ′ . We claim that this fiber in fact consists of the point 0 ∈ H 2 (X ), and this follows from Proposition 5.2. Indeed, for every point b ∈ H 2 (X ) with τ (b) = o, we apply Proposition 5.2 to Xb/ Y and deduce [ω ] = 0 ∈ H 2 (Xb , C). This implies b = 0 by the definition of the period map. (cid:3) 5.3 Comparison with the Calogero-Moser deformation. We now turn to the proof of Theorem 1.20. Let Y = V /G be a symplectic quotient singularity with a resolution X , let X/B be the deformation provided by Theorem 1.13, and let M/C be the Calogero-Moser deformation. Consider also the universal deformation Y /S of the Poisson scheme Y . By construction, we have a multi-valued map κ : C 99K B . As we have noted after stating Conjecture 1.19, the domain and the range of this map are in fact canonically isomorphic as vector spaces. If dim V = 2, then the map κ is single-valued and provides an isomorphism between C and B . It is natural to conjecture that the same is true in the general case: κ : C 99K B is single-valued and an isomorphism. It would immediately imply Theorem 1.20. Unfortunately, we were not able to prove it, and we have to settle for less. Namely, we consider the direct sum decomposition C = Li Bi , see (4.1), and the associated decomposition of the vector space B = H 2 (X, C) ∼= C . As the main technical step in the proof of Theorem 1.20, we show that the map κ : C 99K B is compatible with these decompositions. Proposition 5.7. Let c ∈ Bi be a point in the subspace Bi ⊂ C of the base of the Calogero-Moser deformation M/C . Let κ : C → S be the classifying map of the deformation M/C , and let b ∈ B be an arbitrary point lying over κ(c) ∈ S . (i) The cohomology class [ω ]b ⊂ B ∼= H 2 (X, C) of the symplectic form ω ∈ Ω2 (Xb ) lies in the subspace Bi ⊂ B . (ii) Assume that the point c ∈ Bi is generic. Then a rational homology class α ∈ H2 (X, Q) satisfies hα, [ω ]b i = 0 if and only if α is orthogonal to the whole subspace Bi ⊂ B . 29 By definition of the period map, Proposition 5.7(i) means that the point b ∈ B lies in Bi . Proposition 5.7(ii) means that for a generic c ∈ Bi , the point b ∈ Bi is generic. The proof of Proposition 5.7 is rather technical. We will give it in the next subsection. But first, we will deduce Theorem 1.20 from the proposition. Proof of Theorem 1.20. Let C GOOD ⊂ C be the set of points such that the claim of the Theorem holds for c. This is clearly a Zariski open subset in C . Hence, we must only show that C GOOD is nonempty. To this end, for every rational homology class 0 6= ϕ ∈ H2 (X, Q), denote by Cϕ ⊂ C the subset of elements c ∈ C such that • For some point b ∈ B lying over κ(c) ∈ S , the class ϕ ∈ H2 (X, Q) ∼= H2 (Xb , Q) = H2 (Xb ) lies in the subgroup H2 (Xb/Yκ(c) ) ⊂ H2 (Xb , Q). All the subsets Cϕ ⊂ C are closed algebraic subvarieties. Assume first that there exists ϕ ∈ H2 (X, Q) , ϕ 6= 0, such that Cϕ = C . Then by definition, for any c ∈ C , the class ϕ belongs to the subgroup H2 (Xb/Yκ(c) ) ⊂ H2 (Xb , Q). It follows, by the first claim of Lemma 5.1, that ϕ is orthogonal to [ω ]b , for al l elements c ∈ C . In particular, this applies to generic elements in Bi ⊂ C for each i ≥ 1. By Proposition 5.7, we deduce that ϕ ∈ H2 (X, Q) is orthogonal to all the subspaces Bi ⊂ C . Thus, ϕ = 0, contradicting the assumption ϕ 6= 0. The contradiction implies that Cϕ 6= C , for any nonzero class ϕ ∈ H2 (X, Q). Since there is only a countable set of homology classes ϕ ∈ H2 (X, Q), it follows that there exists c ∈ C which is not contained in Cϕ , for any ϕ. It follows that, for such a c, no nonzero element in H2 (X, Q) belongs to H2 (Xb/Yκ(c) ) for any b lying over κ(c). This means H2 (Xb/Yκ(c) ) = 0. We deduce, similarly to the argument in the proof of Proposition 5.5, that the fibers of Xb → Yκ(c) do not contain pro jective curves. As in the proof of Theorem 1.13, this implies that the map Xb → Yκ(c) is an isomorphism. Thus, c ∈ C GOOD , and we are done. (cid:3) 5.4 Restriction to strata and the end of the proof. It remains to prove Proposition 5.7. To do this, we need some information on the Calogero-Moser deformation M/C and on the geometry of the resolution X/ Y . Somewhat surprisingly, we need to know very little about the Calogero-Moser deformation – it suffices to know how it behaves with respect to the change of the subgroup G ⊂ S p(V ). The precise statement is as follows. Let G1 ⊂ G be a subgroup, let Y1 = V /G1 be the associated symplectic quotient singularity, and let M1/C1 be its Calogero-Moser deformation with its base C1 . We have a canonical restriction map C → C1 . Denote by B1 ⊂ C1 its image, and consider the splitting ι : B1 → C of the pro jection C → B1 defined by ι(c)(g ) = 0 unless g ∈ G is conjugate to an element g1 ∈ G1 . 30 Proposition 5.8. The canonical projection η : Y ′ → Y extends to a commuta- tive diagram (5.6) η −−−−→ M M1 ×C1 B1 y y ι−−−−→ C B1 where M1 ×C1 B1 is the restriction of the Calogero-Moser deformation M1/C1 to B1 ⊂ C1 . (cid:3) , Proposition 5.8 immediately follows from the definition of the Calogero- Moser deformation, see [EG]. We will aply it to the terms in the direct sum decomposition (4.1). Remark 5.9. (Added on Feb. 16, 2010.) As G. Bellamy kindly indicated to us, the above Proposition is not contained in [EG], nor does it follow directly from the definition, and in fact it is not clear if the fact is true or not. Fortunately, in his beautiful recent paper arXiv:1001.0239, I. Losev has proved that if one chooses a point v ∈ V , and lets G1 ⊂ G be the stabilizer of this point, then a slightly stronger statement becomes true after completing both sides near v . This is Theorem 1.2.1 of arXiv:1001.0239. Losev’s Theorem is sufficient to save our proof of Proposition 5.7, the only place where Proposition 5.8 is used. Recall (see Section 4) that the subspaces Bi ⊂ C correspond to codimension- 2 strata Ui in the natural stratification of the quotient variety Y = V /G. Every i /G′ stratum Ui is of the form Ui = V o i , where Vi ⊂ V is a symplectic vector subspace of codimension 2, G′ i ⊂ G is the subgroup of elements which preserve Vi ⊂ V , and V o i ⊂ Vi is the open subset of elements with minimal possible stabilizer. This stabilizer is a subgroup Gi ⊂ G′ i ⊂ G. All elements of the subgroup Gi are symplectic reflections. They act trivially on the subspace Vi ⊂ V . This induces a natural action of the group Gi on the 2-dimensional quotient Wi = V /Vi . The Calogero-Moser deformation of the quotient V /Gi is simply the product of the Calogero-Moser deformation of the quotient Yi = Wi /Gi and the symplectic vector space Vi . Its base Ci is the space of Gi -invariant C-valued functions on the set of all non-trivial elements g ∈ Gi . The image Bi ⊂ Ci of the restriction map C → Ci is the subspace Bi = C Hi i ⊂ Ci of functions invariant with respect to the natural action of the group Hi = G′ i /Gi . By Proposition 5.8, for every stratum Ui we have a commutative dia- gram (5.6), where the scheme in the left top corner is the product Mi × Vi of the Calogero-Moser deformation of the quotient Yi and the symplectic vector space Vi . The Calogero-Moser deformation Mi is the standard deformation described in Section 3. 31 These are all the properties of the Calogero-Moser deformation that we will need. Proposition 5.7, hence also Theorem 1.20 will hold for every Poisson deformation M/C of the quotient singularity Y = V /G which admits a diagram (5.6) for every Gi ⊂ G, with M′ ∼= Mi × Vi . Aside from Proposition 5.8, the decomposition (4.1) has a very clear geo- metric interpretation in terms of the isomorphism H 2 (X, C) ∼= B provided by the generalized McKay correspondence [K2]. Namely, let y ∈ Ui be an arbitrary point in the stratum Ui ⊂ Y . The map Yi × Ui → Y induced by the pro jection V /Gi → V /G is ´etale in y . Therefore the formal neighborhood bYy of the point y ∈ Y naturally decomposes bYy = bYi × bVi , (5.7) product in the sense of formal schemes, where bYi is the completion of the quotient Yi = Wi /Gi at the origin o ∈ Yi , and bVi is the completion of the vector space Vi at 0 ⊂ Vi . Let bXy be the completion of the variety X at the fiber π−1 (y ) ⊂ X . Then by [K1, Proposition 5.2] the decomposition (5.7) lifts to the formal scheme bXy . Namely, we have a canonical decomposition bXy = cXi × bVi , where cXi is the completion of a smooth symplectic resolution Xi / Yi of the 2- dimensional quotient Yi at the exceptional divisor E ⊂ Xi of the map Xi → Yi . The cohomology space H 2 (cXi ) ∼= H 2 (Xi ) ∼= H 2 (E ) is naturally identified with the space Ci , and the natural map B → Ci is given by the restriction map H 2 (X, C) → H 2 (E , C). Proof of Proposition 5.7. We begin with (i). Consider the restriction M/Bi of the Calogero-Moser deformation M/C to the subspace Bi ⊂ C . Let η : Mi × Vi → M/Bi be the pro jection provided by Proposition 5.8. It suffices to prove that [ω ]b maps to 0 under every pro jection B → Bj with j 6= i. Choose an arbitrary point y ∈ Uj and consider the completion bYy together with the induced deformation cMy /cBi (here cBi is the completion of the space Bi at 0 ⊂ Bi . Since j 6= i, there exists a point y ′ ∈ Yi × Vi such that η(y ′ ) = y , the map η is ´etale at y ′ , and the scheme Yi × Vi is smooth at y ′ ∈ Yi × Vi . Then the completion \(Yi × Vi )y ′ of Yi × Vi at y ′ is isomorphic to the completion bV of the vector space V at 0, and the pro jection bV ∼= \(Yi × Vi )y ′ −→ bYy ∼= cYj × cVj is induced by the quotient map V → Yj × Vj ∼= V /Gj . By Lemma 4.1, this im- plies that the Poisson cocycle Θ ∈ H P 2 ( cMy /cBi ) of the Poisson scheme cMy /cBi is a coboundary. In other words, we have Θ = dξ for some vertical vector field ξ on cMy /cBi . Now replace, if necessary, the base Bi of the deformation M/Bi with a finite cover B ′ i , and consider the resolution π : X → M/B ′ i provided by Theorem 1.13. Let bXy be its completion at the fiber π−1 (y ) ⊂ X. The map B → Bj ⊂ Cj is given by the restriction from B ∼= H 2 (X, C) to H 2 (Xj , C) ∼= Bj and factors 32 through the restriction to H 2 ( bXy , C). By Lemma 5.3, the vector field ξ extends to a vertical vector field on bXy /B ′ i . Since Θ = dξ , the Cartan homotopy formula gives ω = d(ω y θ), where ω ∈ Ω2 ( bXy /B ′ i ) is the relative symplectic form. This implies that [ω ] = 0 tautologically on bXy . Therefore the same is true on the whole variety X/B ′ j and for every fiber Xb . This proves (i). To prove (ii), denote by C ⊥ i ⊂ H2 (X, Q) the orthogonal to the subspace Bi ⊂ B ∼= H 2 (X, C), and denote by ω⊥ ⊂ H2 (X, Q) the orthogonal to the cohomology class [ω ]b ∈ H 2 (X, C). We have to prove that the embedding i ⊂ ω⊥ is in fact an equality. To prove this, it suffices to show that C ⊥ dim ω⊥ ≤ dim C ⊥ i = dim B − dim Bi . To prove this inequality, it suffices to exhibit a Q-vector space subspace P of dimension dim P = dim Bi and a map f : P → H2 (X, Q) such that the pairing with [ω ]b induces an embedding P → C. We claim that we can take P = (H2 (Xi , Q))Hi = (H2 (Ei , Q))Hi and f = η∗ : H2 (Ei , Q) → H2 (X, Q). Indeed, by Lemma 5.1 the class [ω ]b ∈ H 2 (Xb , C) comes from a class [ω ] ∈ H 2 (Mc , C). Moreover, the class [ω ] ∈ H 2 (Mc , C) is canonical. In particular, the restriction η∗ [ω ] ∈ H 2 ((Mi )c × Vi , C) coincides with the class [ω ]c ⊗ 1, where [ω ]c ∈ H 2 ((Mi )c , C) is the class corresponding to the Calogero-Moser deformation Mi . But the variety Yi is 2-dimensional. Therefore the Calogero- Moser deformation Mi comes from the universal symplectic deformation of the resolution Xi , and the class [ω ]c ∈ Ci ∼= H 2 (Xi , C) coincides with c ∈ Bi = i ∼= P ∗ ⊗ C, i ⊂ Ci . Thus we have to show that for a generic element c ∈ C Hi C Hi the pairing with c induces an embedding P → C. This is clear. (cid:3) 6 Applications of Hochschild cohomology. The primary goal of this section is to prove Theorem 1.18. We begin how- ever with some general results that relate Hochschild cohomology to orbifold cohomology. Let A be an associative algebra. Recall the Gerstenhaber bracket [−, −] : HH p (A) × HH q (A) → HH p+q−1 (A) on Hochschild cohomology (see [Lo] and §7 below). Given a Hochschild cocyle Θ ∈ HH 2 (A) such that [Θ, Θ] = 0, we introduce a twisted Hochschild cohomology algebra of A as follows. Definition 6.1. For Θ ∈ HH 2(A) such that [Θ, Θ] = 0, define HH q Θ (A) to be the cohomology of the complex (cid:0)HH q (A), dΘ (cid:1) where the differential dΘ : HH q (A) → HH q+1 (A) is given by dΘ (a) = [Θ, a]. Let A = C[M ] be the coordinate ring of a smooth affine algebraic variety M . Then we have HH q (A) = Γ(M , Λ qTM ), by Hochschild-Kostant-Rosenberg 33 theorem, cf. [Lo]. Furthermore, the Gerstenhaber bracket on HH q (A) reduces, in this case, to the Schouten bracket on polyvector fields, cf. §A.7. Thus, the cocycle Θ ∈ HH 2 (A) may be viewed as a bivector. The equation [Θ, Θ] = 0 says that this bivector gives a Poisson structure on A. Moreover, according to (A.17) Θ (A) ∼= H P q (A). Note that of the Appendix, there is a natural isomorphism HH q if M is not smooth, both sides in the isomorphism are still well-defined, but they are not necessarily isomorphic any more. Assume that the smooth affine variety M is a symplectic manifold. We let the 2-cocycle Θ be the bivector on M corresponding to (the inverse of ) the symplectic 2-form. In such a case, the Poisson cohomology reduces to De Rham cohomology, see [Br] and section A.7 below, hence we obtain graded algebra isomorphisms (6.1) Θ (C[M ]) ∼= H P q (C[M ]) ∼= H q (M , C). HH q Assume next that a finite group G acts on M , a smooth affine symplectic variety, by symplectic automorphisms. The action on M induces one on C[M ], and we form the cross-product algebra A := C[M ] # G. The bivector corresponding to (the inverse of ) the symplectic form on M is G-invariant, and the class Θ ∈ HH 2 (C[M ]) = Γ(M , Λ qTM ) gives rise, by the standard deformation theory, to a G-equivariant first-order infinitesimal deformation of the associative algebra C[M ]. Moreover, the equation [Θ, Θ] = 0 insures that this first-order deformation may be extended (not uniquely) to a G-equivariant second order deformation. The latter gives rise, via the cross- product construction, to a second order deformation of C[M ] # G, hence, to a class eΘ ∈ HH 2 (A) = HH 2 (C[M ] # G), such that [ eΘ, eΘ] = 0. We form the (C[M ] # G). corresponding twisted Hochschild cohomology algebra HH q eΘ On the other hand, associated with the G-action on M one has the orbifold cohomology algebra H q orb (M ; G), see (1.1). We have HH q eΘ H q−dim M g Proposition 6.2. Given a symplectic G-action on an affine symplectic mani- fold, there is a natural graded algebra isomorphism orb(M ; G) (cid:18)= (cid:16)Mg∈G (M g )(cid:17)G(cid:19) . (C[M ] # G) ∼= H q Proof. The action of the finite group G on C[M ] being semisimple, the Hochschild cohomology of the cross-product may be expressed in terms of Ext-groups of C[M ]-bimodules as follows, cf. [AFLS]: C[M ]−bimod(cid:0)C[M ] , C[M ]#G(cid:1)G HH q (C[M ] ⋊ G) = Ext q C[M ×M ]−mod (cid:0)C[M ] , C[M ] · g(cid:1)(cid:17)G = (cid:16)Mg∈G Ext q 34 (6.2) . Here G-invariants are taken with respect to the adjoint G-action, and in the rightmost term of the formula we identify C[M ]-bimodules with C[M × M ]- modules. Thus C[M ] · g stands for the C[M × M ]-module arising from the coordinate ring of the graph-subvariety Graph(g : M → M ) ⊂ M × M . In general, let W be a smooth affine variety containing two smooth (closed) subvarieties E , F ⊂ W, such that the intersection E ∩ F is clean. The latter means that E ∩ F is smooth and that the equation TE∩F = TE E∩F ∩ TF E∩F holds for the corresponding tangent bundles. In such a case, we put (6.3) d := dim W − dim E − dim F + dim(E ∩ F ). A standard argument based on Koszul complexes shows that the group Exti C[W ]−mod (C[E ] , C[F ]) vanishes for all i < d, and for i ≥ d we have Exti C[W ]−mod(C[E ] , C[F ]) = = Γ E ∩ F, Λi−d (cid:0) (cid:1)! , + TF E∩F (cid:1) ⊗ det(cid:0) TE E∩F TW E∩F + TF E∩F TE E∩F TE E∩F where det(. . .) denotes the top wedge power. For any g ∈ G, there is a natural g -action on TM M g by vector bundle endomorphisms, and we have a canonical direct sum decomposition TM M g = Image(id −g ) M TM g , where Image(id −g ) denotes the subbundle (of locally constant rank) formed by the images of the fiberwise action of the operator id −g . Writing E ⊂ M × M for the diagonal, we get E ∩ Graph(g ) ∼= M g . We deduce canonical isomorphisms TE M g + TGraph(g) M g ∼= Image(id −g ), and, TE M g TM ×M M g ∼= (TM M g )(cid:14)Image(id −g ) ∼= TM g . TE M g + TGraph(g) M g Thus, M g is a symplectic submanifold and d = dim M g . Further, the symplectic form restricts to a non-degenerate 2-form on the fibers of the vector bundle Image(id −g ), and this gives canonical trivializations det(cid:0)Image(id −g )(cid:1) ∼= C. Being canonical, the trivializations are compatible with G-action (that permutes the fixed point sets M g for various elements g ). Thus, from (6.2) and (6.3) we obtain HH i (C[M ] ⋊ G) ∼= (cid:16)Mg∈G TM g )(cid:17)G Further, the symplectic form on M g induces an vector bundle isomorphism TM g ∼= (TM g )∗ , hence a graded algebra isomorphism Γ(M g , Λ qTM g ) ∼= Ω q (M g ). Combining the formulas above, we obtain HH q (C[M ] ⋊ G) ∼= (cid:16)Mg∈G (M g )(cid:17)G 35 Γ(M g , Λi−dim M g (6.4) . Ω q−dim M g . Observe next that the cochain eΘ ∈ HH 2 (C[M ] ⋊ G) is given, essentially, by the bivector corresponding to the symplectic 2-form. Therefore, it is easy in [Br]), that the differential [ eΘ, −] on Hochschild cohomology to see (as e.g. gets transported under the isomorphism (6.4) to the direct sum of the de Rham (M g ) → Ω q+1−dim M g differentials d : Ω q−dim M g (M g ). The cohomology of the latter is nothing but H q−dim M g (M g , C), the singular cohomology of M g (up to shift). This establishes the isomorphism of the Proposition. Compatibility of the isomorphism with the algebra structures is more difficult (it involves Kontse- vich’s theorem, see [Ko, §8.4], on the cup-product on tangent cohomology), and it will not be given here. Below, see (6.5), we will only use a very special case of Proposition 6.2 where such a compatibility is immediate from definitions. (cid:3) (6.5) For the rest of this section, assume M = V a symplectic vector space, and G ⊂ S p(V ), a finite group, so that the de Rham cohomology of each fixed point set V g is trivial in all degrees but zero. Thus, writing Chki for a 1-dimensional graded vector space concentrated in degree k , the orbifold cohomology algebra in the RHS of the isomorphism of Proposition 6.2 reads orb (V ; G) ∼= (cid:16)Mg∈G Chdim V g i(cid:17)G q C[G](cid:1)G ∼= grF ∼= (cid:0)grF H q q (ZG), where the associated graded algebra is taken with respect to the filtration F q (C[G]) considered in §1. In [EG], the authors construct a certain deformation Ht,c of the cross-product algebra H0,0 := C[V ] # G, which is parametrized by an affine line with coordi- nate t and the space C ∼= grF 2 (ZG) of G-invariant functions on the set on sym- plectic reflections in G. When t = 0, the algebra H0,c has a large center Zc . The Calogero-Moser space Mc is obtained by taking its spectrum, Mc = Spec Zc . The deformation in the t-direction induces a Poisson structure on the algebra Zc and on the variety Mc . When, on the other hand, t is generic, the center of the algebra Ht,c is the one-dimensional C-vector space spanned by the unit element. In [EG], the authors consider the Hochschild cohomology HH q (Ht,c ). They construct a canonical map χ : grF q (ZG) → HH q (Ht,c ), which is shown to be an isomorphism for generic t. Moreover, the map χ is compatible with the deformation Ht,c , i.e., for every pair ht0 , c0 i, the composite map: χ −→ HH 2 (Ht0 ,c0 ) C ∼−→ grF 2 (ZG) is the Kodaira-Spencer map for the family Ht,c near the point ht0 , c0 i. In par- itucular, if one fixes a generic enough t = t0 , then the family Ht0 ,c/C is the universal deformation of the associative algebra Ht0 ,c0 . When t = 0, the map χ is still defined, but its image no longer generates the Hochschild cohomology groups HH q (H0,c ). In fact, these groups become infinite-dimensional as C-vector spaces. Now, fix a c ∈ C , and consider Ht,c as a family depending on t. Applying the Kodaira-Spencer map at the point t = 0, we obtain a cohomology class Θc ∈ 36 HH 2 (H0,c ) which satisfies [Θc , Θc ] = 0. Therefore, we may apply Definition 6.1 and form twisted Hochschild cohomology HH q Θc (H0,c ). q, q HH q Θ0 (H0,0 ) = E 1 Lemma 6.3. The canonical map χ : grF q (ZG) → HH q (H0,c ) descends to a map χ : grF q (ZG) → HH q Θc (H0,c ), and the latter map is an isomorphism. Proof. The map χ : grF q (ZG) → HH q (Ht,c ) is defined for all (t, c). It follows that, for t = 0, the image of this map commutes with Θ0 , hence, the map χ induces a well-defined map grF q (ZG) → HH q Θc (H0,c ). First, consider the case c = 0 where H0,0 = C[V ]#G. Applying Proposition Θ0 (H0,0 ) ∼= 6.2 and using (6.5), we obtain a graded algebra isomorphism HH q gr(ZG), in particular, all odd twisted Hochschild cohomology groups of H0,0 vanish. Moreover, a calculation carried out in [Al] for a Weyl algebra instead of the polynomial algebra C[V ] shows that the isomorphisms above is the inverse of the map χ : grF q (ZG) → HH q Θ0 (H0,0 ) considered in [EG]. This completes the proof in the special case: c = 0. To complete the proof in the general case recall from [EG] that, for any c, the algebra H0,c comes equipped with a canonical increasing filtration such that the associated graded algebra gr H0,c is isomorphic to the algebra H0,0 , see [EG]. This filtration induces a filtration on the twisted Hochschild complex which is compatible with the differential [Θ, −], and therefore gives rise to a spectral sequence Θc (H0,c )(cid:1) . ∞ = gr q(cid:0)HH q =⇒ E q, q Since H0,0 has no odd twisted Hochschild cohomology, all the differentials in q, q 1 = E q, q ∞ . Thus, it follows from this spectral sequence vanish, and we have: E q (ZG) → gr(cid:0)HH q Θc (H0,c )(cid:1) the case c = 0 of the Lemma that the map gr(χ) : grF is a bijection. Therefore, for any c, the map χ : grF q (ZG) → HH q Θc (H0,c ) is also a bijection, and the Lemma is proved. (cid:3) We now apply the above result assuming in addition that c ∈ C is such that the Calogero-Moser space Mc is smooth. Lemma 6.4. Let c ∈ C be such that the Calogero-Moser space Mc = Spec Bc is smooth. Then the Calogero-Moser family M/C considered over the formal neighborhood of the point c gives a universal Poisson deformation of the Poisson variety Mc . Proof. It is proved in [EG] that whenever Mc = Spec Bc is smooth, the algebra H0,c is Morita-equivalent to Bc . Therefore we have HH q (H0,c ) ∼= HH q (Bc ) and Θc (H0,c ) ∼= HH q Θc (Bc ). Using the smoothness of Mc again, we conclude, HH q see (A.17), that the right-hand side is isomorphic to H P q (Bc ). Combining this with Lemma 6.3, we see that the Kodaira-Spencer map for the family M/C computed at the point c ∈ C induces an isomorphism C ∼= H P 2(Bc ). (cid:3) We can now prove Theorem 1.18 using the gradings and the dimension esti- mate obtained in Section 4. 37 Proof of Theorem 1.18. Denote S ′ = H P 2(C[Y ]), and let S ⊂ S ′ be the base of the universal graded Poisson deformation of the graded Poisson algebra C[Y ] = C[V ]G . By Lemma 6.4, the classifying map κ : C → S of the Calogero-Moser family M/C is ´etale at a generic point c ∈ C . In particular, its differential dκ : TcC → Tκ(c)S ′ is injective. Since, dim S ′ ≤ n = dim C , this differential is also surjective, and S = S ′ . It remains to prove that the map κ : C → S = S ′ is surjective. The Calogero-Moser deformation M/C is graded, with grading given by assigning degree 2 to every element in the space C . Therefore the map κ : C → S ⊂ S ′ is compatible with the gradings. Taking quotients with respect to C∗ , we obtain a rational map κ : P(C ) 99K P(S ′ ), where P(C ) ∼= Pn−1 is the pro jectivization of C , and P(S ′ ) is the pro jectivization of the vector space S ′ (with its grading, whatever it may be). Its image κ(P(C )) ⊂ P(S ′ ) is a closed subvariety. Since κ : C → S = S ′ is generically ´etale, the image κ(P(C )) ⊂ P(S ) coincides with the whole P(S ). (cid:3) 7 Appendix. A.1 Harrison cohomology. Consider an arbitrary vector space A over C (or, more generally, over an arbitrary field of characteristic 0). Let L qA be the free graded Lie coalgebra generated by the vector space A placed in degree −1, so that L1A = A and L2A = S 2A, the symmetric square of A. Denote by DL q (A) the graded Lie algebra of coderivations of the Lie coalgebra L qA. Since L q is free, we have DLk (A) ∼= Hom(LkA, A). In particular, every map m : S 2A → A extends to a coderivation d ∈ DL1 . The following is easily checked by a direct computation. Lemma A.1. A commutative product m : S 2A → A is associative if and only if the corresponding coderivation m ∈ DL1 satisfies {m, m} = 0 : L3A → A. (cid:3) Assume that A is a commutative associative algebra. Then we can apply the Lemma and obtain a canonical coderivation m ∈ DL1 (A) satisfying {m, m} = 0. Setting a 7→ {m, a} defines a differential d : DL q (A) → DL q+1 (A) and turns DL q (A) into a DG Lie algebra. The differential d : DL0 (A) → DL1 (A) ∼= Hom(S 2A, A) can be explicitly described in the following way: , (A.1) a ⊗ b ∈ S 2A. (df )(a ⊗ b) = f (ab) − af (b) − bf (a) The spaces DL q (A) = Hom(L qA, A) carry natural A-module structure – A acts on the target space A. It is easy to check that d : DL q (A) → DL q+1 (A) is an A-module map. Moreover, we have DLk (A) ∼= Hom(LkA, A) ∼= HomA (LkA ⊗ A, A), 38 and the the differential d is dual to an A-module map d : L q+1A ⊗ A → L qA ⊗ A. The complex hL q ⊗ A, di is denoted by Har q (A) and called the Harrison complex of the commutative associative algebra A. Its first two terms are d−−−−→ A ⊗ A −−−−→ 0 −−−−→ −−−−→ S 2A ⊗ A with the differential given by d : ab ⊗ c 7−→ a ⊗ bc + b ⊗ ac − ab ⊗ c. It is well- known that the Harrison homology complex is quasiisomorphic to the so-called cotangent complex Ω q (A) of the algebra A3 . A.2 Poisson cohomology. Let A be a unital commutative C-algebra with a Poisson bracket {−, −} : Λ2 C (A) → A that satisfies the Leibniz formula (1.4), the Jacobi identity (1.5), and such that {1, a} = 0, ∀a. Let P q (A) be the free graded Poisson coalgebra generated by the vector space A placed in degree −1 (here ‘free’ means that our coalgebra is a universal ob ject in the category of graded Poisson coalgebras). It is easy to see, cf. e.g. [Fr], that the coalgebra P q (A) may be constructed as the free super-symmetric (co)algebra on the vector space L q (A)(= free graded Lie coalgebra generated by A). In more details, for each m ≥ 0, one has a canonical decomposition Pm (A) = Mp+q=m Pp,q where P q,k = Λk (L q (A)), (A.2) k = 0, 1, . . . . This gives a bigrading P (A) = L Pp,q , Pp,q = Pp,q (A) such that P q,1 (A) = L q (A). The comultiplication in P q (A) preserves the bigrading, while the co- bracket is of bidegree (0, 1). Denote by DP q, q (A) the graded Lie algebra of coderivations of the Pois- son coalgebra P q, q . For reasons of convenience, we will shift the bigrading on DP q, q (A) by (1, 0), so that the first non-trivial term is DP 0,0 (A). Since P q is free, we have k ≥ 0. DP k (A) ∼= Hom(Pk (A), A), Moreover, (A.2) gives an identification DP k, q (A) ∼= Hom(Pk, q (A), A) ∼= Hom(Λk (DL q (A)), A). In particular, we have DP 0,2 (A) = Hom(Λ2 (A), A), the space of all skew- commutative binary operations on A. Assume now that A is an associative commutative algebra. Then by defi- nition, we have the Harrison complex Har q (A) of flat (in fact, free) A-modules. The identification (A.3) can be re-written as DP q,k (A) ∼= HomA (Λk k ≥ 0, A (Har q (A)), A), and the differential in the Harrison complex Har q (A) induces a differential d : DP q, q (A) → DP q, q+1 (A). This turns the graded Lie algebra DP q (A) into a DG Lie algebra. (A.3) (A.4) 3 In degree 0, this is essentially the standard representation of the Kahler differentials module Ω1A as the quotient of the diagonal ideal I ⊂ A ⊗ A by its square. 39 Lemma A.2. (i) A skew-linear operation {−, −} : Λ2 (A) → A satisfies the Leibnitz rule (1.4) if and only if for the corresponding element Θ ∈ DP 0,2 (A) ∼= Hom(Λ2 (A), A) we have dΘ = 0. (ii) The operation {−, −} : Λ2 (A) → A satisfies the Jacobi identity (1.5) if and only if [Θ, Θ] = 0. (iii) Let 1 ∈ A ∼= DP 0,0 (A) be the unit element. We have {−, 1} = 0 if and only if [Θ, 1] = 0. Proof. The first claim immediately follows from (A.1). To prove (ii), note that DP q,0 (A) = Hom(Λ q (A), A) is in fact the graded Lie algebra of coderivations of the free skew-commutative coalgebra Λk (A) generated by A. Therefore [Θ, Θ] = 0 if and only if the associated map {−, −} : Λ2 (A) → A extends to a coerivation δ : Λ q+1 (A) → Λ q (A) satisfying δ ◦ δ = 0. It is well-known (and easily checked) that the latter is equivalent to the Jacobi identity (1.5) on the operation {−, −} : Λ2 (A) → A. Finally, (iii) is clear. (cid:3) By Lemma A.2, giving a Poisson structure on a commutative associative algebra A is equivalent to giving an element Θ ∈ DP 0,2 (A) such that [Θ, 1] = 0. [Θ, Θ] = 0, dΘ = 0, We will call the element Θ ∈ DP 2 (A) the Poisson cochain corresponding to the Poisson structure on A. For every Poisson algebra A, the map (A.5) , a 7→ δ(a) := [Θ, a] δ : DP q, q (A) → DP q+1, q (A) satisfies δ ◦ δ = 0. Definition A.3. The Poisson cohomology H P q (A) of a Poisson algebra A is the cohomology of the total complex (DP q , d + δ) associated to the Poisson (bi)complex DP q, q (A) with differentials d : DP q, q (A) → DP q, q+1 (A) and δ : DP q, q (A) → DP q+1, q (A). Here are the first few terms in the bicomplex DP q, q (A). (A.6) x x δ −−−−−→ Hom(Λ2 (S 2A), A) ⊕ Hom(L3A ⊗ A, A) Hom(L3A, A) x x Hom(S 2A ⊗ A, A) Hom(S 2A, A) x x δ Hom(Λ2 (A), A) −−−−−→ Hom(A, A) δ −−−−−→ δ −−−−−→ δ −−−−−→ d d d d d d δ −−−−−→ δ −−−−−→ A 40 Here the leftmost vertical column is the Harrison complex Hom(L qA, A), the 2-d vertical column is formed by graded pieces of Λ2 (cid:0)Hom(L qA, A)(cid:1), etc., and the bottom row is the standard cochain complex for A viewed as a Lie algebra with respect to the Poisson bracket. In particular we get (i) H P 0 (A) = (cid:8)z ∈ A (cid:12)(cid:12) {z , a} = 0 , ∀a ∈ A(cid:9) = Poisson center of A. (ii) H P 1 (A) = Z P 1/BP 1 where Z P 1 , BP 1 ⊂ Hom(A, A) are defined by Z P 1 := nf (cid:12)(cid:12) f (ab) = af (b) + f (a)b , f ({a, b}) = {f (a), b} + {a, f (b)} , ∀a, bo and BP 1 := nf = fc : a 7→ {c, a} (cid:12)(cid:12) c ∈ Ao. Next, let ϕ ∈ Hom(S 2A, A) and ψ ∈ Hom(Λ2A, A). Define two C-bilinear maps A ⊗ A → A[ε]/(ε2 ) by the formulas (A.7) a, b 7−→ a ·ε b = a · b + ε · ϕ(a, b) a, b 7−→ {a, b}ε = {a, b} + ε · ψ(a, b) , It is straightforward to verify that (iii) ϕ ⊕ ψ is a Poisson 2-cocycle if and only if formulas (A.7) give a 1-st order infinitesimal deformation of A as a Poisson algebra. Furthermore, the group H P 2 (A) classifies 1-st order infinitesimal deformations up to equivalence. Remark A.4. The Poisson bracket gives a canonical cocycle Θ ∈ DP 2 (A), which may or may not be trivial in Poisson cohomology, depending on the algebra A. A.3 Relation to Hochschild cohomology and the Kunneth formula. It is well-known that the Poisson operad may be viewed as a ‘degeneration’ of the Associative operad. In this way, the Poisson cohomology bicomplex of a Poisson algebra may be viewed as a ‘degeneration’ of the Hochschild cochain complex. To explain the analogy, recall that, given a vector space A, one defines three graded Lie algebras DL q (A), DP q (A), and DT q (A) as the Lie algebras of coderivations of the free Lie colagebra L q (A), free Poisson colagebra P q (A) or free associative coalgebra (with counit) T q (A), respectively. As in the case of DP q (A), we shift the grading on DT q (A) by one, so that DT k (A) ∼= Hom(Tk (A), A) = Hom(A⊗k , A), k ≥ 0. Under this identification, the Lie bracket in DT q (A) becomes the Gerstenhaber bracket given, for every f ∈ DT k (A), g ∈ DT l(A), by the standard formula [f , g ](a1 ⊗ · · · ⊗ ak+l−1 ) = (A.8) X1≤i≤l (−1)i g (a1 ⊗ · · · ⊗ f (ai ⊗ · · · ⊗ ai+k−1 ) ⊗ · · · ⊗ ak+l−1 ) − X1≤i≤k (−1)i f (a1 ⊗ · · · ⊗ g (ai ⊗ · · · ⊗ ai+l−1 ) ⊗ · · · ⊗ ak+l−1 ). 41 Recall that the free associative coalgebra T q (A) is known to be isomorphic to the universal enveloping coalgebra of the free Lie coalgebra L q (A). This means that there is a canonical decreasing filtration on the associative coalgebra T q (A), such that the corresponding associated graded gr q T q (A) is isomorphic, by Poincar´e-Birkhoff-Witt theorem, to the free super-symmetric coalgebra on the vector space L q (A). As we have mentioned earlier, this free super-symmetric coalgebra is nothing but the free Poisson coalgebra P q, q (A) generated by A. In other words, there is a canonical Poisson coalgebra (bigraded) isomorphism gr q T q (A) ∼= P q, q (A) . Further, the decreasing filtration on T q (A) induces an increasing filtration on the graded Lie algebra DT q (A) which we call the PBW filtration. The associated graded gr q DT q (A) acts on the associated graded gr q T q (A) ∼= P q, q (A) by Poisson coderivations; therefore we have a Lie algebra map gr q DT q (A) → DP q, q (A). It is easy to check that this map is an isomorphism. A similar relation be- tween DT q (A) and DP q, q (A), formulated in terms of eulerian idempotents, was considered in [Fr]. For every vector space A we have DT 1(A) ∼= Hom(A ⊗ A, A). An element m ∈ Hom(A ⊗ A, A) satisfies [m, m] = 0 if and only if the corresponding binary operation A ⊗ A → A is associative. In this case, one defines a differential DT q (A) → DT q+1 (A) and obtains the Hochschild cochain complex DT q (A). Its cohomology groups are called the Hochschild cohomology and denoted by HH q (A). If the associative algebra A is commutative, then the Hochshchild differential DT q (A) → DT q+1 (A) preserves the PBW filtration. The complex gr q DT q (A) with induced differential is then isomorphic to the Poisson coho- mology bicomplex H P q (A) of the Poisson algebra A, which is the commutative algebra A equipped with trivial Poisson bracket {−, −} = 0 : Λ2 (A) → A. Note that the PBW filtration on the complex DT q (A) gives rise to a PBW filtration F PBW q HH q (A) on the Hochschild cohomology. We will use the relationship between Hochschild and Poisson complexes to prove the Kunneth formula for Poisson cohomology. To do this, we recall that for every two associative unitary algebras A, B , there exists a canonical quasi- isomorphism (A.9) sh : DT q (A ⊗ B ) ∼= DT q (A) ⊗ DT q (B ), given by the shuffle product, see [Lo, §4.2]. In more detail, consider the cat- egory ∆ of finite linearly ordered sets, and denote by [n] ∈ ∆, n ≥ 1 the set of cardinality n. By an (l, n)-shuffle we will understand an order-preserving embedding Φ : [l] → [n]. Given an (l, k + l)-shuffle Φ, one defines the comle- mentary (k , k + l)-shuffle Φ : [k ] → [k + l]. Taken together, Φ and Φ define a permutation σΦ : [l] ∪ [k ] → [k + l] of the set of k + l elements. For every shuffle 42 Φ : [l] → [k + l], we define a map shΦ : A⊗l ⊗ B⊗k → (A ⊗ B )⊗k+l by shΦ (a1 ⊗ · · · ⊗ al ⊗ b1 ⊗ · · · ⊗ bk ) = c1 ⊗ · · · ⊗ ck+l , ci = (ap ⊗ 1, i = Φ(p), p ∈ [l], 1 ⊗ bq , i = Φ(q), p ∈ [k ], The map sh : DT k+l (A ⊗ B ) → DT l(A) ⊗ DT k (B ) ∼= Hom(A⊗l ⊗ B⊗k , A ⊗ B ) is given by (A.10) sh(f )(a1 ⊗ · · · ⊗ ak ⊗ b1 ⊗ · · · ⊗ bl ) = XΦ:[l]→[k+l] sign(σΦ )f (shΦ (a1 ⊗ · · · ⊗ ak ⊗ b1 ⊗ · · · ⊗ bl )) for every f ∈ DT k+l (A ⊗ B ) = Hom((A ⊗ B )⊗k+l , A ⊗ B ). In addition to the quasiisomorphism sh, we also have an embedding κA : DT q (A) → DT q (A ⊗ B ) given by f ∈ DT k (A), κA (f )(a1 ⊗ b1 ⊗ · · · ⊗ ak ⊗ bk ) = f (a1 ⊗ · · · ⊗ ak ) ⊗ b1 · . . . · bk , and an analogously defined embedding κB : DT q (B ) → DT q (A ⊗ B ). Looking at the formula (A.8) for the Gerstenhaber bracket, we immediately see that both κA and κB are Lie algebra maps. Moreover, say that a cochain ω ∈ DT k (A) is reduced if f (a1 ⊗ · · · ⊗ ak ) = 0 whenever at least one of the elements a1 , . . . , ak ∈ is equal to 1. Then one can easily derive from (A.8) and (A.10) that for every reduced cochain ω ∈ DT q (A) and an arbitrary cochain f ∈ DT q (A ⊗ B ) we have (A.11) sh([κAω , f ]) = [ω , sh(f )] (here the bracket on the right-hand side acts on the first factor in DP q (A) ⊗ DP q (B )). The same statement holds for A replaced by B . If the algebras A and B are commutative, then the quasiisomorphism (A.10) is compatible with the PBW filtrations. The associated graded map (A.12) sh : DP q (A) ⊗ DP q (B ) → DP q (A ⊗ B ) is also a quasiisomorphism; it is induced by the standard quasiisomorphism (Ω q (A) ⊗ B ) ⊕ (A ⊗ Ω q (B )) ∼= Ω q (A ⊗ B ) between cotangent complexes. Lemma A.5. For every two Poisson algebras A, B , the shuffle map (A.10) induces a natural quasiisomorphism sh : DP q (A ⊗ B ) ∼= DP q (A) ⊗ DP q (B ). 43 Proof. It suffices to prove that the map sh is compatible with the differentials; since the associated graded map (A.12) is a quasiisomorphism, sh : DP q (A ⊗ B ) ∼= DP q (A) ⊗ DP q (B ) will be a quasiisomorphism as well. By definiton, the differential in the Poisson cohomology complex is the sum of two differential, which we denoted by d and δ . Compatibility with d is contained in (A.12). Thus we have to prove that sh(δA⊗B f ) = (δA ⊗ id + id ⊗δB ) sh(f ) for every cochain f ∈ DP q (A ⊗ B ). But by definition, the Poisson cochain ΘA⊗B is of the form ΘA⊗B = κAΘA + κB ΘB . Since both Poisson cochains ΘA , ΘB are clearly reduced, we can apply (A.11) and obtain sh([ΘA⊗B , f ]) = [ΘA , sh(f )] + [ΘB , sh(f )]. By (A.5), this is precisely what we had to prove. (cid:3) A.4 Graded algebras. As in Subsection 1.3, by a graded Poisson algebra of degree l we will understand a Poisson algebra A equipped with a grading such that the multiplication in A is compatible with the grading, and the Poisson bracket is of degree −l. For a graded Poisson algebra A of degree l 6= 0, the Poisson cocycle Θ ∈ DP 2 (A) is canonically a coundary. Indeed, the grading A = L A q induces a canonical derivation ξ : A → A by setting ξ = k id on Ak . The formula c 7→ [ξ , c] extends this derivation to a derivation of the Lie algebra DP q, q (A). Since the Poisson bracket is of degree −l with respect to the grading, we have [ξ , Θ] = −l · Θ, which can be rewritten as Θ = − 1 l dξ . The grading on A induces an additional grading on the Poisson cohomology bicomplex DP q, q (A), called the A-grading. The differential d : DP q, q (A) → DP q, q+1 (A) preserves the A-grading, and the differential δ : DP q, q (A) → DP q+1, q (A) shifts it by l. It will be convenient to redefine the A-grading by shifting it by (k − 1)l on DP k, q (A), so that it is preserved by both differen- tials d, δ . We obtain a Poisson cohomology graded bicomplex DP q, q (A). Note that after the shift, the Poisson cochain Θ ∈ DP 2, q (A) has A-degree 0, and the A-grading becomes compatible with the Lie bracket on DP q, q (A). A.5 Modules and ´etale descent. Let A be a Poisson algebra. By a Pois- son module M over A we will understand an A-module M equipped with an additional operation {−, −} : A ⊗ M → M such that {ab, m} = a{b, m} + b{a, m}, {a, bm} = {a, b}m + b{a, m}, {{a, b}, m} = {a, {b, m}} − {b, {a, m}}. Given a Poisson module M one defines, following [Fr], its Poisson cohomology bicomplex DP q, q (A, M ) (with coefficients in M ) by DP p,q (A, M ) = Hom(Pp,q (A), M ), p, q ≥ 0, 44 To get the differentials, it is convenient to treat bA := A ⊕ M as a Poisson algebra by letting both operations to be zero on M ⊂ bA (the trivial square- zero extension). We grade this algebra by assigning deg A = 0, deg M = 1, so that bA becomes a graded Poisson algebra of degree l = 0. The differentials on DP q, q ( bA) preserve the induced grading, and DP q, q (A, M ) is nothing but the graded component in DP q, q ( bA) of minimal degree. In the special case M = A, we get back the original Poisson cohomology bicomplex DP q, q (A). Indeed, the algebra bA ∼= Ahεi = A[ε]/(ε2) is simply the truncated polynomial algebra over A. The bicomplex DP q, q (A) naturally lies in the degree-0 part of DP q, q ( bA), and it is easy to check that, for M = A, multiplication by ε provides an isomorphism DP q, q (A, M ) ∼= DP q, q (A). Cohomology with coefficients appear naturally if one wants to consider the functoriality properties of Poisson cohomology groups. Let f : A → B be a Poisson map between Poisson algebras. Then there is no natural map betwen the groups H P q (A), H P q (B ) – just as there is no natural map between the spaces of derivations of the algebras A, B . However, there exist obvious natural maps f : DP q, q (A) → DP q, q (A, f∗B ) and df : DP q, q (B ) → DP q, q (A, f∗B ). If ΘA , ΘB are the Poisson cochains of the algebra A, B , then we have f (ΘA ) = df (ΘB ). More generally, for every Poisson module M over B , there exists a natural map f : DP q, q (B , M ) → DP q, q (A, f∗M ). The main reason we want to consider functoriality is the following compati- bility result. Lemma A.6. Let A → B be an ´etale map between Poisson algebras, and let M be a Poisson module over the algebra B . Then the natural map (A.13) f : DP k, q (B , M ) → DP k, q (A, f∗M ) induced by the map A → B is a quasiisomorphism for every k ≥ 0. Conse- quently, the induced map DP q (B , M ) → DP q (A, f∗M ) of total complexes is a quasiisomorphism. Proof. By (A.4), it suffices to prove that the canonical map Har q (A) ⊗ B → Har q (B ) of Harrison complexes is a quasiisomorphism – in other words, that the pull- back Ω q (A) ⊗ B of the cotangent complex Ω q (A) of the algebra A is naturally quasiisomorphic to the cotangent complex Ω q (B ) of the algebra B . This is very well-known4 . (cid:3) Notice that for every Poisson algebra A and every multiplicative system S ⊂ A, the localization A[S−1 ] carries a natural Poisson structure. The natural map A → A[S−1 ] is an example of a Poisson ´etale map. This example will be used later in constructing Poisson cohomology complexes of Poisson schemes. 4 In degree 0, this is the definition of an ´etale map. 45 A.6 Poisson schemes. We define a Poisson scheme as a scheme X equipped with a Poisson bracket on the structure sheaf OX . For any Poisson algebra A, the affine scheme X = Spec A is a Poisson scheme. Poisson morphisms between Poisson schemes, sheaves of Poisson OX -modules and complexes of such sheaves are defined in the natural way. For every Poisson module M over a Poisson algebra A, the corresponding coherent sheaf M on X = Spec A is a sheaf of Poisson modules. Moreover, for every Poisson morphism f : X → Y and a sheaf F of Poisson OX -modules on X , the direct image sheaf f∗F is a sheaf of Poisson OY -modules. The same is true for the higher direct image sheaves Rpf∗F , p ≥ 1, and for the whole complex R qf∗F . The construction of the Poisson cohomology complex DP q (A) generalizes straightforwardly to the scheme case. Namely, we notice that the functors Pk , k ≥ 0 from the category of vector spaces into itself can be easily defined in an arbitrary symmetric tensor category with the unit ob ject. The terms DP k (A) = Hom(Pk (A), A) of the Poisson cohomology complex can be defined in any symmetric monoidal category which admits internal Hom ’s. Moreover, the differential in DP q (A) for a Poisson algebra A is defined essentially by linear algebra, and the definition also works just as well in the general categorical set- ting. The same is true for the cohomology complex with coefficients H P q (A, M ). We use this and define the Poisson cohomology complex HP q (X, F ) of the Pois- son scheme X with coefficients in an injective sheaf F of Poisson OX -modules by setting HP k (X ) = Hom (Pk (OX ), F ), where all tensor products and the Hom are taken in the category of Zariski sheaves of groups on X . This complex carries the same combinatorial structures as in the case of algebras, in particular, the PBW filtration. This definition makes sense for an arbitrary sheaf F of Poisson modules – in particular, for the structure sheaf OX – but it gives the wrong result: the resulting functor does not have reasonable exactness properties, and it becomes impossible to prove a crucial compatibility condition (Proposition A.10 below). To get the correct definition, we use canonical injective resolutions. Namely, let X be the union of all scheme-theoretic points of X , and let j : X → X be the canonical map. Recall that for every sheaf F of abelian groups on X , the canon- ical map F → j∗ j ∗F is an embedding, and the sheaf j∗ j ∗F is injective. Iterating this construction, one obtains the so-called canonical Godement resolution F q of the sheaf F . If F is a sheaf of Poisson OX -modules, then all the terms in the Godement resolution carry canonical structures of Poisson OX -modules. Definition A.7. Let F be a sheaf of Poisson OX -modules on a Poisson scheme X . The local Poisson cohomology complex HP q (X, F ) is the total complex of the bicomplex HP q (X, F q ) = Hom (Pk (OX ), F q ), (A.14) where F q is the canonical Godement resolution of the sheaf F . The Poisson cohomology H P q (X, F ) = H q (X, HP q (X, F )) 46 with coefficients in F is the hypercohomology of the local Poisson cohomology complex HP q (X, F ). We note that the functors P q (OX ) = P q, q (X ) carry an additional grading, so that in fact the local Poisson cohomology complex HP q (X, F ) = HP q, q (X, F ) is a bicomplex. For every k ≥ 0, we have a quasiisomorphism HP k, q (X, F ) ∼= RHom q (ΛkΩ q (X ), F ), where Ω q (X ) is the cotangent complex of the scheme X , and Λk is understood in the derived-category sense. Moreover, the complex HP k, q (X, F ) is concentrated in degrees ≥ 0. Therefore we obtain a canonical embedding Hom(ΛkΩ(X ), OX ) → HP k, q (X, F ), where Ω(X ) is the cotangent sheaf. Since the Godement resolution is functorial, Definition A.7 immediately gen- eralizes to the case when F is itself not a single sheaf of Poisson OX -modules, but a complex of such sheaves. Then it descends to a triangulated functor on the corresponding derived category. In the case when F = OX is the structure sheaf, we will denote the local Poisson cohomology complex HP q (X, OX ) sim- ply by HP q (X ), and we will denote the Poisson cohomology H P q (X, OX ) by H P q (X ). Another important situation is the following. Let Z ⊂ X be a closed sub- scheme in a Poisson scheme X , and let U ⊂ X be the open complement. Let i : Z ֒→ X , j : U ֒→ X be the embeddings. Then there exists a canonical map O q U , where O q X → j∗O q X is the Godement resolution of the sheaf OU , and U is the Godement resolution of the sheaf OU . This map is compatible with O q the natural Poisson structures on both sides. Therefore we also have a Poisson module structure on its cone, denoted by i!O q Z ⊂ O q X (this is a particular model of the canonical quasicoherent complex i!OZ used in the local duality theory). Z (X ) = H P q (X, i!O q Definition A.8. The Poisson cohomology H P q Z ) is called the Poisson cohomology of the scheme X with supports in Z ⊂ X . Note that the closed subscheme Z ⊂ X enters into this construction only through its open complement U ⊂ X . In particular, there is no need to assume that Z ⊂ X is a Poisson subscheme in any sense. We will need the following easy vanishing result. Lemma A.9. Assume that the Poisson scheme X is Cohen-Macaulay, and the closed subscheme Z ⊂ X is of codimension codim Z ≥ k for some integer k ≥ 2. Then H P i Z (X ) = 0 for al l i ≤ k − 2. Proof. Since X is Cohen-Macaulay, the complex i!OZ is trivial in degrees ≤ k−1. (cid:3) 47 Proposition A.10. Let X = Spec A be a Poisson scheme with a Poisson sheaf F of OX -modules. Then the canonical map H P q (X, F ) → H P q (A, RΓ q (X, F )) is a quasiisomorphism. Proof. The claim is functorial in F . By Definition A.7, it suffices to consider sheaves of the form F = i∗M , where i : x ֒→ X is the embedding of a scheme- theoretic point x ∈ X , and M ∼= RΓ q (A, F ) is a Poisson module over the local ring OX,x . Let F = i∗M be such a sheaf. Then by definition the local Poisson cohomology complex HP q (X, F ) is quasiisomorphic HP q (X, F ) ∼= i∗H P q (OX,x , M ), so that H P q (X, F ) ∼= H P q (OX,x , M ). To prove the claim, it suffices to note that the natural map A → OX,x is ´etale, and apply Lemma A.6. (cid:3) Proposition A.11. (i) For every two Poisson schemes X , Y , there exists a canonical quasiismor- phism (A.15) H P q (X × Y ) ∼= H P q (X ) ⊗ H P q (Y ). (ii) Let f : X → Y be an ´etale map of Poisson schemes. Then for every complex F q of sheaves of Poisson modules on X , there exists a canonical quasiisomprhism H P q (Y , f∗F q ) ∼= H P q (X, F q ). (iii) Let f : X → Y be an ´etale Galois cover with Galois group G. Then there exists a canonical quasiisomorphism H P q (Y ) ∼= (H P q (X ))G . Proof. By taking affine covers, we reduce (i) and (ii) to the case of affine schemes X , Y . Then Proposition A.10 reduces both statements to their algebraic ver- sions Lemma A.5 and Lemma A.6. Finally, (iii) is a direct corollary of (ii): we have (H P q (X ))G ∼= (H P q (Y , f∗OX ))G ∼= H P q (Y , (f∗OX )G ) ∼= H P q (Y , OY ). (cid:3) Corollary A.12. (i) Let f : X → Y be a map between Poisson schemes, and assume the restriction of f to a closed subscheme Z ⊂ X is injective. Then, there exists a canonical quasiisomorphism Z (Y ) → H P q H P q Z (X ). (ii) Let Z ⊂ X be a closed subscheme in a Poisson scheme X , and let Y be another Poisson scheme. Then there exists a canonical isomorphism Z×Y (X × Y ) ∼= H P q H P q Z (X ) ⊗ H P q (Y ). (A.16) 48 Proof. (ii) immediately follows from Proposition A.11(i). To prove (i), let iX : Z ֒→ X , iY = f ◦ iX : Z ֒→ Y be the embeddings. Notice that the map f must be ´etale in a Zariski neighborhood of the subscheme Z ⊂ X , and we have ! OZ ∼= iY f∗ iX ! OZ . The claim then follows from Proposition A.11(ii). (cid:3) A.7 The smooth case. Let now X be a smooth Poisson scheme. In this case, we have the original Koszul-Brylinski formalism for local Poisson cohomology. Namely, one considers the graded Lie algebra Λ qT (X ) of polyvector fields on X with the so-called Schouten bracket. The Poisson structure is defined by a bivector field Θ ∈ Λ2T (X ) such that [Θ, Θ] = 0. Taking commutator with Θ defines a differential d : Λ qT (X ) → Λ q+1T (X ); we will call the complex Θ (X ) ∼= hΛ qT (X ), di the Koszul-Brylinski complex. HP q For a general smooth Poisson scheme X , we have canonical embeddings Λk T (X ) → HP k, q (X ). These embeddings are compatible with the Koszul-Brylinski differential and de- fine a quasiisomorphism HP q Θ (X ) → HP q (X ). Indeed, since the embeddings are canonical, it suffices to check this fact locally, so that we can assume that the scheme X = Spec A is affine. Then since A is smooth, the Harrison ho- mology Har q (A) is quasiisomorphic to the module Ω1A of Kahler differentials of the algebra A, and this module is flat over A. Therefore the complex DP k, q (A) only has non-trivial cohomology in degree 0, and H P k,0 (A) ∼= Λk T (A), cf. [Fr]. We have DP k,l (A) = 0 for l < 0 and a canonical embedding Λk T (A) ∼= H P k,0 (A) → DP k,0 (A) sends the Poisson bivector Θ into the Poisson cochain, and it also sends the Schouten bracket in Λ qT (A) into the Gerstenhaber bracket in DP q (A). Therefore it sends the Koszul-Brylinski differential into the differ- ential δ : DP q, q (A) → DP q+1, q (A), see also [Fr]. Recall that in case of a smooth affine Poisson variety X = Spec A there are Hochschild-Kostant-Rosenberg isomorphisms HH q (A) ∼= Λ q A (Har1 (A)) ∼= Λ q AT (A) , and the Koszul-Brylinski complex of X may be identified via this isomorphism with the ‘twisted’ Hochschild complex of A considered in §6, cf. Definition 6.1. Thus, we have (A.17) Θ (X ) ∼= HH q H P q Θ (A) . Also, in the smooth case, for the PBW filtration on Hochschild cohomology we get i HH j (A) = (HH j (A) F PBW 0 Furthermore, the spectral sequence associated to the bicomplex (A.6) collapses. If the Poisson scheme X is not only smooth but also symplectic, then the symplectic form provides an identification T (X ) ∼= Ω(X ). This identification if i ≥ j if i < j. 49 Θ (X ) ∼= Ω q (X ), the standard de Rham complex extends to an isomorphism HP q of the smooth scheme X . Thus for symplectic schemes we have HP q (X ) ∼= HP q Θ (X ) ∼= Ω q (X ) and H P q (X ) ∼= H q (X, C), the ordinary singular cohomology of the scheme X . We can also obtain iso- morphisms for cohomology with coefficients: Poisson modules over X are the same as D-modules, and the Poisson cohomology H P q (X, F ) with coefficients in a Poisson module is isomorphic to the de Rham cohomology with coefficients in the corresponding D-module. In particular, the Poisson cohomology with supports is isomorphic to the singular cohomology with supports. A.8 Deformations. The shortest route to the classification of Poisson defor- mations (in particular, to the proof of Theorem 1.10) is through the formalism of deformation groupoids. This consists of obtaining deformed structures from solutions of the so-called Maurer-Cartan equation dγ = {γ , γ } in a certain differential-graded Lie algebra L q . More precisely, assume given a DG Lie algebra L q with differential d : L q → L q+1 and commutator {−, −}, and a local Artin algebra S with maximal ideal m ⊂ S . Consider the DG Lie algebra L q ⊗ S over S . Set g = L0 . For any g ∈ g ⊗ m, a ∈ L1 ⊗ S set g · a = dg + [g , a] ∈ L1 ⊗ S. Consider g · a as a tangent vector to the vector space L1 ⊗ S at the point a ∈ L1 ⊗ S . Then the collection g · a, a ∈ L1 ⊗ S defines a vector field on L1 ⊗ S for every g ∈ g ⊗ m. These vector fields glue together to an action of the Lie algebra g ⊗ m on L1 ⊗ S . Since the Lie algebra g ⊗ m is nilpotent, this action extends to an action of the corresponding nilpotent Lie group Gm . Definition A.13. The Maurer-Cartan groupoid M C (L q , S ) associated to the pair hL q , S i is defined in the following way: (i) Ob jects of M C (L q , S ) are elements γ ∈ L1 ⊗ m satisfying the Maurer- Cartan equation (A.18) dγ = {γ , γ }. (ii) Morphisms between ob jects γ1 , γ2 ∈ M C (L q , S ) are elements g ∈ Gm in the group Gm such that g · γ1 = γ2 . To check that this definition is consistent, one has to prove that the Gm - action preserves the Maurer-Cartan equation, which is an elementary computa- tion. The Maurer-Cartan formalism admits an obvious graded version. If the DG Lie algebra L q is equipped with a grading, – or, equivalently, witha C∗ -action, – then one can consider local Artin algebras S equipped with a C∗ -action and form 50 the grouppoid M Cgr (L q , S ) of the C∗ -invariant solutions to the Maurer-Cartain equation (A.18). Application of the Maurer-Cartan formalism to the Poisson deformation the- ory is an immediate corollary of Lemma A.2 and Lemma A.1. k ≥ 0, k < 0. Lemma A.14. Let A be a Poisson algebra over C. Consider the Poisson co- homology complex DP q (A), and let Lk = (DP k+1 (A), 0, Let S be a local Artin algebra with maximal ideal m. Then the category Def (A, S ) of flat Poisson algebras eA over S equipped with an isomorphism eA/m · eA ∼= A is equivalent to the Maurer-Cartan groupoid M C (L1 , S ). Proof. By Lemma A.2 and Lemma A.1, defining a Poisson algebra struc- ture on a vector space V is equivalent to giving a cochain α = m + Θ ∈ DP 2 (V ) = DL1 (V ) ⊕ DP 2,0 (V ) which satifies [α, α] = 0. To define a func- tor M C (Der q (A), S ) → Def (A, S ), it suffices to notice that for every solution γ ∈ L1 = DP 2 (A) ⊗ m of the Maurer-Cartan equation, the element 2 γ ∈ L1 ⊗ S eα = α − 1 satisfies { eα, eα} = 0, hence defines a Poisson algebra structure on A ⊗ S . This functor is surjective on ob jects – indeed, by definition for every deformation eA we have eA ∼= A ⊗ S as S -modules. To prove that it is an equivalence, it suffices to prove that it is surjective on morphisms. But the Lie algebra L0 = DP 1 (A) ∼= Hom(A, A) by definition consists of al l linear endomorphisms of the vector space A. Therefore morphisms in the Maurer-Cartan groupoid are all S -module morphisms A ⊗ S → A ⊗ S which reduce to identity modulo m ⊂ S . This is the same as morphisms in Def (A, S ). (cid:3) For any DG Lie algebra L q , the Maurer-Cartan groupoids M C (L q , S ) form a stack over the category of local Artin algebras S . The general machinery shows that this stack is (pro)-representable. The representing ob ject is a certain canonically defined subscheme in the vector space H 1 (L q ) completed at zero, modulo the action of the (pro)-algebraic group corresponding to the Lie algebra L0 . The machinery works in the case of graded DG Lie algebras without any change: the (pro)-representing ob ject carries a natural C∗ -action compatible with the induced grading on H 1 (L q ), and it (pro)-represents the Maurer-Cartain stack M Cgr (L q , S ) on the category of local Artin schemes equipped with a C∗ - action. However, if the degrees of the induced gradings on H 1 (L q ), H 0 (L q ) are strictly positive, we can obtain more. Say that a commutative algebra A = L A q is positively graded if dim Ak = 0 for k < 0, dim A0 = 1 and dim Ak < ∞ for k > 0. Further, we say that a 51 complete local algebra hA, mi is equipped with a good C∗ -action if the group C∗ acts on bA, and this action induces a positive grading on every Artin quotient A/mkA and on the associated graded quotient gr q bA = ⊕k mk /mk+1 . For every positively graded algebra A q , denote by bA the completion of the algebra A q with respect to the maximal ideal m = A≥1 ⊂ A q . We note the following standard fact (essentially, this is a toy version of the Grothendieck algebraization, [EGA, III, 5.4]). (i) The correspondence A q 7→ bA is an equivalence between the Lemma A.15. category of positively graded commutative algebras A q and the category of complete local algebras h bA, m ⊂ bAi equipped with a good C∗ -action. (ii) For any positively graded algebra A q , the completion induces an equiva- lence between the category of finitely generated graded A q -modules and the category of finitely generated C∗ -equivariant bA-modules. (iii) If the algebra A q is finitely generated, then the completion induces an equiv- alence between the category of projective C∗ -equivariant schemes of finite type over Spec A q and the category of projective C∗ -equivariant schemes of finite type over Spec bA. (iv) Let X be a projective C∗ -equivariant scheme of finite type over Spec A q , and let X/ Spec bA be its completion. Then completion induces an equiva- lence between the category of coherent C∗ -equivariant sheaves on X and coherent C∗ -equivariant sheaves on X. A sheaf F on X is flat if and only if its completion bF is flat. Moreover, for every sheaf F on X , the completion \H k (X, F ) of the cohomology group H k (X, F ) with respect to the m-adic topology is canonical ly isomorphic to the cohomology group H k (X, bF ) of the completion bF Sketch of proof. To prove (i), we will just give the inverse to the equivalence of categories. It is given by hA, mi 7→ Mp By our definition of a good C∗ -action, for every integer p the pro jective system of the corresponding graded components (A/mk )p actually stabilizes at some finite level k ≥ 0, and the limit is a finite-dimensional vector space. Thus the right-hand side is indeed a positively graded commutative algebra. To prove (ii), we define the inverse equivalence by the same formula (A.19). To prove (iii), we write X = Pro j bB q for a scheme bX pro jective over Spec bA, apply (A.19) and obtain a scheme X = Pro j B over Spec A whose completion is bX (here bB has two independent gradings: the first one is related to the C∗ - action, and the second one comes from the definition of Pro j). This defines the inverse equivalence on the level of ob jects. To lift it to morphisms, one identifies morphisms with their graphs. (A/mk )p . lim ← (A.19) 52 Finally, to prove (iv) one uses Serre’s Theorem and identifies the category of coherent sheaves on a pro jective scheme B with the quotient of the category of finitely-generated graded B -modules by the fat subcategory of finitely-generated B -modules with finite number of non-trivial graded components. Again using (A.19) – and keeping track of the additional grading – one checks easily that the completion is an equivalence both between the categories of modules and between the fat subcategories one has to mod out. Therefore it also induces an equivalence between the quotient categories. This equivalence is clearly compatible with the tensor products, hence pre- serves flatness. To prove the statement on cohomology, it suffices to identify the graded component H k (X, F )p of some degree p with the Ext-group Extk (Op , F ), where Op is the structure sheaf on X with grading shifted by p. (cid:3) If the degrees of the natural gradings on H 1 (L q ), H 0 (L q ) are strictly positive, then the C∗ -action on the universal ob ject for the Maurer-Cartan stack is good in the sense of Lemma A.15. Therefore this ob ject falls within the scope of the Lemma, and we automatically obtain a universal solution for the Maurer-Cartan equation over a positively graded algebra S , not only over its completion bS . In the case of Poisson deformations of a Poisson algebra A, we have H 1 (L q ) ∼= H P 2 (A). The Lie algebra H 0 (L q ) is the algebra of all Poisson cocycles in DP 1 (A) = Hom(A, A) which coincides with the Lie algbera of all Poisson derivations. This Lie algebra is rather big; in particular, it contains all the Hamiltonian vector fields on A. However, we do not need to know the precise stack structure on M C (L q , S ) – our goal is the weaker Theorem 1.10. To prove it, it suffices to prove that the Lie algebra H 0 (L q ) acts trivially on the vector space H 1 (L q ). Proof of Theorem 1.10. Assume that H P 1(A). We have to prove that the Lie algebra H 0 (L1 ) of Poisson derivations of the algebra A acts trivially on the cohomology group H 1 (L q ). Consider the map L q → DP q−1 (A). This is a Lie algebra map and a isomorphism in all non-negative degrees. In particular, it induces an isomorphism H 1 (L q ) ∼= H P 2 (A), and the action of the algebra H 0 (L q ) on H P 2 (A) ∼= H 1 (L q ) is induced by the action of the Lie algebra H P 1 (A). By assumption, the Lie algebra H P 1 (A) is trivial. (cid:3) A linear algebra result. Let V be a finite dimensional complex symplectic vector space, and G ⊂ S p(V ) a finite subgroup. Write V g for the fixed point space of g ∈ G. Lemma A.16. For any elements g , h ∈ G such that V = V g + V h , we have V g·h = V g ∩ V h . Proof. It is clear that V g ∩ V h ⊆ V g·h . To prove the opposite inclusion, let x ∈ V g·h . The assumption V = V g + V h , implies that x = u + v , for some u ∈ V g and v ∈ V h . Thus, gh(u + v) = u + v , hence, hu + v = u + g−1v , or equivalently (h − 1)u = (g−1 − 1)v . But, for any a ∈ S p(V ) of finite order, there is an orthogonal direct sum decomposition V = Image(id −a) ⊕ V a . We 53 deduce that (h − 1)u in the LHS of the last equation is orthogonal to V h , resp. (g−1 − 1)v in the RHS is orthogonal to V g , with respect to the symplectic form. Thus, each side is orthogonal to V = V g + V h , hence, vanishes. It follows that (h − 1)u = (g−1 − 1)v = 0. Therefore u, v ∈ V g ∩ V h , hence x ∈ V g ∩ V h , and we obtain V g·h ⊆ V g ∩ V h . (cid:3) References [AFLS] J. Alev, M.A. Farinati, T. Lambre, and A.L. Solotar, Homologie des invariants d’une alg`ebre de Weyl sous l’action d’un groupe fini. J. of Algebra 232 (2000), 564–577. [Al] M.S. Alvarez, Algebra structure on the Hochschild cohomology of the ring of invariants of a Weyl algebra under a finite group. J. Algebra 248 (2002), 291- 306. [Ba] V. Baranovsky, Orbifold Cohomology as Periodic Cyclic Homology. arXiv:math.AG/0206256 [BB] A. Beilinson, J. Bernstein, Localisation de g-modules. C. R. Acad. Sci. Paris 292 (1981), 15–18. [BD] V. Batyrev and D. Dais, Strong McKay correspondence, string-theoretic Hodge numbers and mirror symmetry, Topology, 35 (1996), 901–929. Preprint alg-geom/9410001 . [BKR] T. Bridgeland, A. King, and M. Reid, The McKay correspondence as an equiv- alence of derived categories. J. Amer. Math. Soc. 14 (2001), 535–554. [BG] K. Brown, I. Gordon, Poisson orders, symplectic reflection algebras and repre- sentation theory, arXiv:math.RT/020104. [Br] J.-L. Brylinski, A differential complex for Poisson manifolds. J. Diff. Geom. 28 (1988), 93–114. [CF] F. Campana and H. Flenner, Contact singularities, [ arXiv:math.AG/010906 ]. [CR] W. Chen, Y. Ruan, A new cohomology theory for orbifold, arXiv:math.AG/0004129. [DL] J. Denef, F. Loeser, Germs of arcs on singular algebraic varieties and motivic integration. Invent. Math. 135 (1999), 201–232. [EG] P. Etingof and V. Ginzburg, Symplectic reflection algebras, Calogero-Moser space, and deformed Harish-Chandra homomorphism, Invent. Math. 147 (2002), 243-348, [ arXiv:math.AG/0011114 ]. [FG] B. Fantechi, L. Gottsche, Orbifold cohomology for global quotients. Duke Math. J. 117 (2003), 197–227. [arXiv:math.AG/0104207]. [Fl] [Fr] [Fu] H. Flenner, Extendability of differential forms on non-isolated singularities, Inv. Math. 94 (1988), 317-326. B. Fresse, Homologie de Quil len pour les alg`ebres de Poisson, C.R. Acad. Sci. Paris 326 (1998), 1053–1058. B. Fu, Symplectic Resolutions for Nilpotent Orbits, Invent. Mathem. 151 (2003), 167–186. [arXiv:math.AG/0205048] 54 [Go] I. Gordon, Baby Verma modules for rational Cherednik algebras, Bull. London Math. Soc. 35 (2003), 321–336, [arXiv:math.RT/0202301]. [EGA] A. Grothendieck, EGA, 0, III, Publ. Math. IHES 11. [K1] D. Kaledin, Dynkin diagrams and crepant resolutions of quotient singularities, arXiv:math.AG/9903157, to appear in Selecta Math. [K2] D. Kaledin, McKay correspondence for symplectic quotient singularities, In- vent. Math. 148 (2002), 151–175, [ arXiv:math.AG/9907087 ]. [K3] D. Kaledin, Symplectic resolutions: deformations and birational maps, preprint arXiv:math.AG/0012008. [KV] D. Kaledin and M. Verbitsky, Period map for non-compact holomorphical ly symplectic manifolds, Geom. Funct. Anal. 12 (2002), 1265–1295. [KS] M. Kontsevich, Y. Soibelman, Deformations of algebras over operads and Deligne’s conjecture. [arXiv:math.QA/0001151]. [Kon] M. Kontsevich, Deformation quantization of Poisson manifolds. arXiv:q-alg/9709040. [Ko] J.-L. Koszul, Crochet de Schouten-Nijenhuis et cohomologie. Ast´erisque 1985, Numero Hors Serie, 257–271. [Ku] A. Kuznetsov, Quiver varieties and Hilbert schemes, arXiv:math.AG/0111092. [LS] M. Lehn, Ch. Sorger, Symmetric groups and the cup product on the cohomology of Hilbert schemes, Duke Math. J. 110 (2001), 345–357, [arXiv:math.AG/0009131]. [Lo] J.-L. Loday, Cyclic homology. Grundlehren der Mathematischen Wis- senschaften 301 Springer-Verlag, Berlin, 1992. [LQW] W. Li, Z. Qin, and W. Wang, Vertex algebras and the cohomology ring structure of Hilbert schemes of points on surfaces, Math. Ann. 324 (2002), 105–133, [arXiv:math.AG/0009132]. [McK] J. McKay, Graphs, singularities and finite groups, in The Santa Cruz Confer- ence on Finite Groups, Proc. Symp. Pure Math. 37 (1980), 183–186. [Ri] [R1] D.S. Rim, Formal Deformation Theory. SGA 7 I, Lect. Notes in Mathem. 288 Springer-Verlag 1972, pp.32–132. Y. Ruan, Stringy Geometry and Topology of Orbifolds, Symposium in Honor of C. H. Clemens, 187–233, Contemp. Math., 312, Amer. Math. Soc., Providence, RI, 2002, arXiv:math.AG/0011149. [R2] Y. Ruan, Stringy Orbifolds. Orbifolds in mathematics and physics (Madison, WI, 2001), 259–299, Contemp. Math., 310, Amer. Math. Soc., Providence, RI, 2002, [arXiv:math.AG/0201123]. [Vas] E. Vasserot, Sur l’anneau de cohomologie du sch`ema de Hilbert de C2 , C.R. Acad. Sci. Paris, 332 (2001), 7–12, [ arXiv:math.AG/0009127 ]. [Ve] M. Verbitsky, Holomorphic symplectic geometry and orbifold singularities, Asian Journal of Mathematics, 4 (2000), 553-564. [W] K. Watanabe, Certain invariant subrings are Gorenstein, I, II, Osaka J. Math. 11 (1974), 1–8; 379–388. [WW] J. Wierzba and J. Wiszniewski, Smal l Contractions of Symplectic 4-folds, preprint arXiv:math.AG/020102. 55 V.G.: Department of Mathematics, University of Chicago, Chicago, IL 60637, USA; [email protected] D.K.: Steklov Mathematical Institute, Moscow, Russia [email protected] 56

Subsets and Splits

No community queries yet

The top public SQL queries from the community will appear here once available.