Datasets:

mlcore

/

arxiv-classifier

like 1

Follow

Kilian's Group 7

1612.08443

2

1612

$G_2$-Grassmannians and derived equivalences

We prove the derived equivalence of a pair of non-compact Calabi-Yau 7-folds, which are the total spaces of certain rank 2 bundles on $G_2$-Grassmannians. The proof follows that of the derived equivalence of Calabi-Yau 3-folds in $G_2$-Grassmannians by Kuznetsov closely.

math.AG

math

G2-GRASSMANNIANS AND DERIVED EQUIVALENCES KAZUSHI UEDA Abstract. We prove the derived equivalence of a pair of non-compact Calabi -- Yau 7- folds, which are the total spaces of certain rank 2 bundles on G2-Grassmannians. The proof follows that of the derived equivalence of Calabi -- Yau 3-folds in G2-Grassmannians by Kuznetsov [Kuz] closely. 1. Introduction The simply-connected simple algebraic group G of type G2 has three homogeneous spaces G := G/P1, Q := G/P2, and F := G/B associated with the crossed Dynkin diagrams × ×, and × ×respectively. The Picard group of F can be identified with the weight lattice of G, which in turn can be identified with Z2 as (a, b) := aω1 + bω2, where ω1 and ω2 are the fundamental weights associated with the long root and the short root respectively. We write the line bundle associated with the weight (k, l) as OF(k, l). , Let ∞ (1.1) R := Mk,l=0 (k,l)(cid:17)∨ be the Cox ring of F, where (cid:16)V G with the highest weight (k, l). H 0 (OF(k, l)) ∼= ∞ (k,l)(cid:1)∨ Mk,l=0(cid:0)V G is the dual of the irreducible representation of G The Z2-grading of R defines a (Gm)2-action on Spec R, which induces an action of Gm embedded in (Gm)2 by the anti-diagonal map α 7→ (α, α−1). We write the geometric invariant theory quotients as (1.2) where (1.3) V+ := Proj R+, V− := Proj R−, V0 := Spec R0, ∞ ∞ Ri,n−i, R+ := Rn, R− := R−n. Mn=0 Mn=0 Rn =Mi∈Z V+ and V− are the total spaces of the dual of the equivariant vector bundles of rank 2 on Q and G associated with irreducible representations of P1 and P2 with the highest weight (1, 1). The structure morphisms φ+ : V+ → V0 and φ− : V− → V0 are crepant resolutions which contract the zero-sections. The same construction for the simply-connected simple algebraic group Sp(2) of type C2, which is accidentally isomorphic to the simply-connected simple algebraic group Spin(5) of type B2, gives the 5-fold flop discussed in [Seg16], where it is attributed to Abuaf. The main result in this paper is the following: Theorem 1.1. V+ and V− are derived-equivalent. Theorem 1.1 provides an evidence for the conjecture [BO02, Conjecture 4.4] [Kaw02, Conjecture 1.2] that birationally equivalent smooth projective varieties are K-equivalent if and only if they are D-equivalent. 1 The proof of Theorem 1.1 closely follows [Kuz], where the derived equivalence of Calabi -- Yau complete intersections in G and Q defined by sections of the equivariant vector bundles dual to V+ and V−. The derived equivalence of these Calabi -- Yau 3-folds in turn follows from Theorem 1.1 using matrix factorizations. Notations and conventions. We work over a field k throughout this paper. All pull- back and push-forward are derived. The complexes underlying Ext•(−, −) and H •(−) will be denoted by hom(−, −) and h(−). Acknowledgements. We thank Atsushi Ito, Makoto Miura, and Shinnosuke Okawa for collaborations [IMOUa, IMOUb, IMOUc] which led to this work. K.U. is supported by Grants-in-Aid for Scientific Research (24740043, 15KT0105, 16K13743, 16H03930). 2. The blow-up diagram The G2-Grassmannian G is the zero locus s−1 λ (0) of the section sλ of the equivariant vector bundle Q∨(1) of rank 5 on Gr(2, V ), obtained as the tensor product of the dual Q∨ of the universal quotient bundle Q and the hyperplane bundle O(1). Here V := V G (0,1) is the 7-dimensional fundamental representation of G2, and sλ corresponds to the G2- invariant 3-form on V under the isomorphism H 0(Gr(2, V ), Q∨(1)) ∼= V3 V ∨. We write the G2-equivariant vector bundle associated with the irreducible representation of P1 with the highest weight (a, b) as E(a,b). The restriction U := SG of the universal subbundle S of rank 2 on Gr(2, V ) is isomorphic to E(−1,1). The G2-flag variety F is isomorphic to the total space of the P1-bundle + : P(U ) → G associated with U (or any other equivariant vector bundle of rank 2, since all of them are related by a twist by a line bundle). We write the relative hyperplane class of + as h, so that (2.1) (+)∗ (OF(h)) ∼= U ∨. The pull-back to F of the hyperplane class H in G will be denoted by H again by abuse of notation. The other G2-Grassmannian Q is a quadric hypersurface in P(V ). We write the equi- variant vector bundle on Q associated with the irreducible representation of P2 with high- est weight (a, b) as F(a,b). The flag variety F has a structure of a P1-bundle − : F → Q, whose relative hyperplane class is given by H. We define a vector bundle K on Q by (2.2) K := ((−)∗ (OF(H)))∨ , so that F ∼= PG(K ). One can show that K is isomorphic to F(1,−3). We write the hyperplane class of Q as h by abuse of notation, since it pulls back to h on F. Let V be the total space of the line bundle OF(−h−H) on F. The structure morphism will be denoted by π : V → F. The Cox ring of V is the N2-graded ring (2.3) S = Sk,l ∞ Mk,l=0 2 given by (2.4) (2.5) (2.6) (2.7) (2.8) (2.9) (2.10) (2.11) (2.12) (2.13) (2.14) Sk,l := H 0 (OV(k, l)) ∼= H 0 (π∗ (OV(k, l))) ∼= H 0 (π∗OV ⊗ OF(k, l)) ∞ ∼= ∼= H 0 ∞ Mm=0 Mm=0 Mm=0(cid:0)V G ∼= U (−H) is given by L∞ ∼= ∞ OF(m, m)! ⊗ OF(k, l)! H 0 (OF(k + m, l + m)) , (k+m,l+m)(cid:1)∨ k=0 H 0(cid:0)OW+(kH)(cid:1) where ∼= H 0(cid:0)π∗OW+ ⊗ OG(kH)(cid:1) Mm=0 Mm=0 Mm=0 H 0(cid:0)(cid:0)Symm E(1,1)(cid:1) ⊗ OG(kH)(cid:1) H 0(cid:0)E(m,m) ⊗ E(k,0)(cid:1) H 0(cid:0)E(m+k,m)(cid:1) . ∼= ∼= ∼= ∞ ∞ ∞ whose multiple Proj recovers V. Similarly, the Cox ring of the total space W+ of the bundle E ∨ (1,1) H 0(cid:0)OW+(kH)(cid:1) ∼= H 0(cid:0)π∗(cid:0)OW+(kH)(cid:1)(cid:1) This is isomorphic to R+, so that W+ is isomorphic to V+, and the affinization morphism (2.15) V → Spec H 0 (OV) ∼= V0 is the composition of the natural projection ϕ+ : V → V+ and the affinization morphism φ+ : V+ → V0. Since V+ is the total space of E ∨ (1,1), the ideal sheaf of the zero-section is the image of the natural morphism from π∗ +E(1,1) to OV+, and the morphism ϕ+ is the blow-up along it. Similarly, the affinization morphism (2.15) also factors into the blow-up ϕ− : V → V− and the affinization morphism φ− : V− → V0, and one obtains the following commutative diagram: (2.16) V+ ϕ− !❈❈❈❈❈❈❈❈ φ− }④④④④④④④④ V− V ϕ+ }④④④④④④④④ !❈❈❈❈❈❈❈❈ φ+ V0 3 } ! ! } The zero-sections and the natural projections fit into the following diagram: 3. Some extension groups (3.1) G V+ + }④④④④④④④④ ~⑤⑤⑤⑤⑤⑤⑤⑤ φ+ F ι V − !❇❇❇❇❇❇❇❇ ❆❆❆❆❆❆❆❆ φ− Q V− S , and UV := π∗UF. By abuse of notation, we use We write UF := ∗ + the same symbol for an object of Db(F) and its image in Db(V) by the push-forward ι∗. Since V is the total space of OV(−h − H), one has a locally free resolution U , SF := ∗ − (3.2) 0 → OV(h + H) → OV → OF → 0 of OF as an OV-module. By tensoring OF(−h) to [Kuz, Equation (5)], one obtains an exact sequence (3.3) 0 → OF(H − 2h) → U ∨ F (−h) → OF → 0. Lemma 3.1 and Proposition 3.2 below are taken from [Kuz]: Lemma 3.1 ([Kuz, Lemma 1]). (i) Line bundles OF(th − H) and OF(tH − h) are acyclic for all t ∈ Z. (ii) Line bundles OF(−2H) and OF(2h − 2H) are acyclic and H •(OF(3h − 2H)) ∼= k[−1]. (iii) Vector bundles UF(−2H), UF(−H), UF(h − H) and UF ⊗ UF(−H) are acyclic, and H •(UF(h)) ∼= k, H •(UF ⊗ UF(h)) ∼= k[−1]. Proposition 3.2 ([Kuz, Proposition 3 and Lemma 4]). One has an exact sequence (3.4) 0 → UF → SF → U ∨ F (−h) → 0. Lemma 3.1 immediately implies the following: Lemma 3.3. OF(−H) is right orthogonal to both U ∨ F (−h) and OF(−h). Proof. We have (3.5) (3.6) and (3.7) (3.8) homOV (OF(−h), OF(−H)) ∼= homOV ({OV(H) → OV(−h)} , OF(−H)) ∼= h ({OF(h − H) → OF(−2H)}) homOV (U ∨ F (−h), OF(−H)) ∼= homOV ({U ∨ V (H) → U ∨ V (−h)} , OF(−H)) ∼= h ({UF(h − H) → UF(−2H)}) , both of which vanish by Lemma 3.1. (cid:3) Lemma 3.4. One has (3.9) homOV (U ∨ F (−h), UF) ∼= k[−1]. 4 }  _   ! _   ~  _   Proof. One has (3.10) (3.11) homOV (U ∨ F (−h), UF) ∼= homOV ({U ∨ V (−h)} , UF) ∼= h ({UF ⊗ UF(h) → UF ⊗ UF(−H)}) . V (H) → U ∨ Lemma 3.1 shows that the first term gives k[−1] and the second term vanishes. (cid:3) Lemma 3.5. One has (3.12) Proof. One has homOV (U ∨ F (−h), OF) ∼= k. (3.13) (3.14) homOV (U ∨ F (−h), OF) ∼= homOV ({U ∨ V (H) → U ∨ ∼= h ({UF(h) → UF(−H)}) . V (−h)} , OF) Lemma 3.1 shows that the first term gives k and the second term vanishes. (cid:3) Lemma 3.6. One has (3.15) Proof. One has homOV (OF(H − 2h), OF(h)) ∼= 0. (3.16) homOV (OF(H − 2h), OF(h)) ∼= homOV ({OV(2H − h) → OV(H − 2h)} , OF(h)) (3.17) ∼= h ({OV(3h − H) → OV(2h − 2H)}) , which vanishes by Lemma 3.1. (cid:3) 4. Derived equivalence by mutation Recall from [Kuz] that (4.1) and (4.2) Db(G) = hOG(−H), U , OG, U ∨, OG(H), U ∨(H)i Db(Q) = hOQ(−3h), OQ(−2h), OQ(−h), S , OQ, OQ(h)i . It follows from [Orl92] that (4.3) and (4.4) where (4.5) and (4.6) (4.1) and (4.3) gives Db(V) =(cid:10)ι∗∗ Db(V) =(cid:10)ι∗∗ +Db(G), Φ+(Db(V+))(cid:11) −Db(Q), Φ−(Db(V−))(cid:11) , Φ+ := φ∗ +(−) ⊗ OV(h) : Db(V+) → Db(V) Φ− := φ∗ −(−) ⊗ OV(H) : Db(V−) → Db(V). By mutating Φ+(Db(V+)) two steps to the left, one obtains (4.7) (4.8) Db(V) =(cid:10)OF(−H), UF, OF, U ∨ Db(V) =(cid:10)OF(−H), UF, OF, U ∨ 5 F , OF(H), U ∨ F (H), Φ+(Db(V+))(cid:11) . F (H)(cid:11) F , Φ1(Db(V+)), OF(H), U ∨ where (4.9) F (H)i ◦ Φ+. By mutating the last two terms to the far left, one obtains Φ1 := LhOF(H),U ∨ Db(V) =(cid:10)OF(−h), U ∨ since ωV affecting other objects: F (−h), OF(−H), UF, OF, U ∨ ∼= OV(−h − H). Lemma 3.3 allows one to move OF(−H) to the far left without F , Φ1(Db(V+))(cid:11) , F , Φ1(Db(V+))(cid:11) . F (−h), UF, OF, U ∨ By mutating UF one step to the left and using Proposition 3.2 and Lemma 3.4, one obtains By mutating OF(−H) to the far right, one obtains By mutating Φ1(Db(V+)) to the right, one obtains Db(V) =(cid:10)OF(−H), OF(−h), U ∨ Db(V) =(cid:10)OF(−H), OF(−h), SF, U ∨ Db(V) =(cid:10)OF(−h), SF, U ∨ Db(V) =(cid:10)OF(−h), SF, U ∨ F (−h), OF, U ∨ F (−h), OF, U ∨ Φ2 := ROF(h) ◦ Φ1. F (−h), OF, U ∨ F , Φ1(Db(V+))(cid:11) . F , Φ1(Db(V+)), OF(h)(cid:11) . F , OF(h), Φ2(Db(V+))(cid:11) (4.10) (4.11) (4.12) (4.13) (4.14) where (4.15) (4.17) (4.18) (4.19) where (4.20) By mutating U ∨ F (−h) one step to the right and using Lemma 3.5 and (3.3), one obtains (4.16) Similarly, by mutating U ∨ Db(V) =(cid:10)OF(−h), SF, OF, OF(H − 2h), U ∨ F one step to the right, one obtains F , OF(h), Φ2(Db(V+))(cid:11) . Lemma 3.6 allows one to exchange OF(H − 2h) and OF(h) to obtain Db(V) =(cid:10)OF(−h), SF, OF, OF(H − 2h), OF(h), OF(H − h), Φ2(Db(V+))(cid:11) . Db(V) =(cid:10)OF(−h), SF, OF, OF(h), OF(H − 2h), OF(H − h), Φ2(Db(V+))(cid:11) . Db(V) =(cid:10)OF(−h), SF, OF, OF(h), Φ3(Db(V+)), OF(H − 2h), OF(H − h)(cid:11) By mutating Φ2(Db(V+)) two steps to the left, one obtains Φ3 := LhOF(H−2h),OF(H−h)i ◦ Φ2. By mutating the last two terms to the far left, one obtains (4.21) By comparing (4.21) with Db(V) =(cid:10)OF(−3h), OF(−2h), OF(−h), SF, OF, OF(h), Φ3(Db(V+))(cid:11) . Db(V) =(cid:10)OF(−3h), OF(−2h), OF(−h), SF, OF, OF(h), Φ−(Db(V−))(cid:11) obtained by combining (4.2) and (4.4), one obtains a derived equivalence (4.22) (4.23) where (4.24) Φ := Φ! − ◦ Φ3 : Db(V+) ∼−→ Db(V−), Φ! −(−) := (φ−)∗ ((−) ⊗ OV(−H)) : Db(V) → Db(V−) 6 is the left adjoint functor of Φ−. Note that the left mutation along an exceptional object E ∈ Db(V) is an integral functor ΦK(−) := (p2)∗ (p∗ 1(−) ⊗ K) along the diagram (4.25) p1 z✉✉✉✉✉✉✉✉✉✉ V p2 $■■■■■■■■■■ V V ×V0 V whose kernel K is the cone over the evaluation morphism ev : E ∨ ⊠ E → ∆V. The functors Φ+ : Db(V+) → Db(V) and Φ! − : Db(V) → Db(V−) are clearly an integral functor, so that the functor (4.23) is also an integral functor, whose kernel is an object of Db(V+ ×V0 V−) obtained by convolution. 5. Matrix factorizations Let s+ be a general section of the equivariant vector bundle E(1,1) on G. The zero X+ of s+ is a smooth projective Calabi -- Yau 3-fold. Since V+ is the total space of the dual bundle E ∨ (1,1) on G, the space of regular functions on V+ which are linear along the fiber can naturally be identified with the space of sections of E(1,1). We write the regular 3-fold in Q. Let ς− be a regular function on V− corresponding to ς+ under the isomorphism is the union of a line sub-bundle of V+ and the inverse image of X+ by the structure morphism π+ : V+ → G. The singular locus of D+ is given by X+. function on V+ associated with s+ ∈ H 0(cid:0)E(1,1)(cid:1) as ς+ ∈ H 0(cid:0)OV+(cid:1). The zero D+ of ς+ H 0(cid:0)OV+(cid:1) ∼= H 0 (OV0) ∼= H 0(cid:0)OV−(cid:1) given by the diagram in (2.16), and X− be the zero of the corresponding section s− ∈ H 0(cid:0)F(1,1)(cid:1) , which is a smooth projective Calabi -- Yau The push-forward of the kernel of Φ on V+ ×V0 V− to V+ ×A1 V− gives a kernel of Φ on V+ ×A1 V−. By taking the base-change along the inclusion 0 → A1 of the origin and applying [Kuz06, Proposition 2.44], one obtains an equivalence Φ0 : Db(D+) ∼= Db(D−) of the bounded derived categories of coherent sheaves. By using either of the charac- terization of perfect complexes as homologically finite objects (i.e., objects whose total Ext-groups with any other object are finite-dimensional) or compact objects (i.e., ob- jects such that the covariant functors represented by them commute with direct sums), one deduces that ΦD preserves perfect complexes, so that it induces an equivalence sing(D+) ∼= Db Φsing sing(D−) of singularity categories (see [KKOY09, Section 7] and [BP10, Theorem 1.1]). : Db 0 Recall that V+, V− and V0 are geometric invariant theory quotient of Spec R by the anti-diagonal Gm-action. The residual diagonal Gm-action on both V+ and V− are dilation action on the fiber. The equivalences Φ, Φ0 and Φsing are equivariant with respect to this Gm-action, and induces an equivalence of Gm-equivariant categories [Hir, Theorem 1.1], which will be denoted by the same symbol by abuse of notation. Now [Isi13, Theorem 3.6] gives equivalences 0 (5.1) and (5.2) Db sing([D+/Gm]) ∼= Db(X+) Db sing([D−/Gm]) ∼= Db(X−) between Gm-equivariant singularity categories and derived categories of coherent sheaves (see also [Shi12] where the case of line bundles is discussed independently and around the same time as [Isi13]). By composing these derived equivalences with Φsing , one obtains a derived equivalence between X+ and X−. It is an interesting problem to compare this 0 7 z $ equivalence with the one obtained in [Kuz]. Another interesting problem is to prove the derived equivalence using variation of geometric invariant theory quotient along the lines of [HHP, Seg11, BFK, HL15, ADS15], and use it to produce autoequivalences of the derived category [DS14, HLS16]. References [ADS15] Nicolas Addington, Will Donovan, and Ed Segal, The Pfaffian-Grassmannian equivalence [BFK] [BO02] [BP10] [DS14] [HHP] [Hir] [HL15] [HLS16] revisited, Algebr. Geom. 2 (2015), no. 3, 332 -- 364. MR 3370126 Matthew Ballard, David Favero, and Ludmil Katzarkov, Variation of geometric invariant theory quotients and derived categories, arXiv:1203.6643. A. Bondal and D. Orlov, Derived categories of coherent sheaves, Proceedings of the Interna- tional Congress of Mathematicians, Vol. II (Beijing, 2002), Higher Ed. Press, Beijing, 2002, pp. 47 -- 56. MR 1957019 Vladimir Baranovsky and Jeremy Pecharich, On equivalences of derived and singular cate- gories, Cent. Eur. J. Math. 8 (2010), no. 1, 1 -- 14. MR 2593258 Will Donovan and Ed Segal, Window shifts, flop equivalences and Grassmannian twists, Compos. Math. 150 (2014), no. 6, 942 -- 978. MR 3223878 Manfred Herbst, Kentaro Hori, and David Page, Phases of N=2 theories in 1+1 dimensions with boundary, arXiv:0803.2045. Yuki Hirano, Equivalences of derived factorization categories of gauged Landau-Ginzburg models, arXiv:1506.00177. Daniel Halpern-Leistner, The derived category of a GIT quotient, J. Amer. Math. Soc. 28 (2015), no. 3, 871 -- 912. MR 3327537 Daniel Halpern-Leistner and Ian Shipman, Autoequivalences of derived categories via geomet- ric invariant theory, Adv. Math. 303 (2016), 1264 -- 1299. MR 3552550 [IMOUa] Atsushi Ito, Makoto Miura, Shinnosuke Okawa, and Kazushi Ueda, Calabi -- Yau complete intersections in G2-grassmannians, arXiv:1606.04076. [IMOUb] , The class of the affine line is a zero divisor in the Grothendieck ring: via G2- Grassmannians, arXiv:1606.04210. [IMOUc] , The class of the affine line is a zero divisor in the Grothendieck ring: via K3 surfaces [Isi13] [Kaw02] of degree 12, arXiv:1612.08497. Mehmet Umut Isik, Equivalence of the derived category of a variety with a singularity cate- gory, Int. Math. Res. Not. IMRN (2013), no. 12, 2787 -- 2808. MR 3071664 Yujiro Kawamata, D-equivalence and K-equivalence, J. Differential Geom. 61 (2002), no. 1, 147 -- 171. MR MR1949787 (2004m:14025) [Kuz] [Kuz06] [KKOY09] Anton Kapustin, Ludmil Katzarkov, Dmitri Orlov, and Mirroslav Yotov, Homological mir- ror symmetry for manifolds of general type, Cent. Eur. J. Math. 7 (2009), no. 4, 571 -- 605. MR 2563433 (2010j:53184) Alexander Kuznetsov, Derived equivalence of Ito -- Miura -- Okawa -- Ueda Calabi -- Yau 3-folds, arXiv:1611.08386. A. G. Kuznetsov, Hyperplane sections and derived categories, Izv. Ross. Akad. Nauk Ser. Mat. 70 (2006), no. 3, 23 -- 128. MR 2238172 (2007c:14014) D. O. Orlov, Projective bundles, monoidal transformations, and derived categories of coher- ent sheaves, Izv. Ross. Akad. Nauk Ser. Mat. 56 (1992), no. 4, 852 -- 862. MR MR1208153 (94e:14024) Ed Segal, Equivalence between GIT quotients of Landau-Ginzburg B-models, Comm. Math. Phys. 304 (2011), no. 2, 411 -- 432. MR 2795327 [Orl92] [Seg11] [Seg16] [Shi12] , A new 5-fold flop and derived equivalence, Bull. Lond. Math. Soc. 48 (2016), no. 3, 533 -- 538. MR 3509912 Ian Shipman, A geometric approach to Orlov's theorem, Compos. Math. 148 (2012), no. 5, 1365 -- 1389. MR 2982435 Graduate School of Mathematical Sciences, The University of Tokyo, 3-8-1 Komaba, Meguro-ku, Tokyo, 153-8914, Japan. E-mail address: [email protected] 8 

1701.08838

1

1701

Dominant classes of projective varieties

We give evidence for a uniformization-type conjecture, that any algebraic variety can be altered into a variety endowed with a tower of smooth fibrations of relative dimension one.

math.AG

math

DOMINANT CLASSES OF PROJECTIVE VARIETIES FEDERICO BUONERBA AND FEDOR BOGOMOLOV Abstract. We give evidence for a uniformization-type conjecture, that any algebraic variety can be altered into a variety endowed with a tower of smooth fibrations of relative dimension one. The problem of constructing a resolution of singularities of projective varieties is one of the most fundamental obstructions to our understanding of their analytic, arithmetic and geometric properties. A tremendous amount of work has shed light on the this problem, yet in its full generality it is still wide open over fields of positive characteristic. A cornerstone result, albeit conjecturally not optimal, is de Jong's [dJ96] 3.1, stating that any variety can be altered into a smooth projective one. Allowing alterations, other than birational modifications, comes along with a profusion of new natural questions, in the spirit of: which further properties can we require, for a class of smooth projective varieties, to be dominant? Recall from [BH00] that a class C (k, n) of n-dimensional projective varieties over a field k is dominant if for every projective n-dimensional variety X, there exists Y ∈ C (k, n) and a surjective k-morphism Y → X. In this terminology, de Jong's result says that smooth projective varieties form a dominant class over any field and in any dimension. Constructing minimal classes of dominant varieties is a problem that attracted attention, and a satisfactory answer is still unknown even in the case of curves over fields that are finitely generated over their prime subfield. Some results, questions and speculations in this direction can be found in [BT02],[BT02'],[BT05]. In this paper we give evidence for the following conjecture: Conjecture 1 ([BH00]). For any field k and positive integer n, the class of n-dimensional smooth projective varieties X, endowed with a tower of smooth fibrations X → X1 → ... → Xn with dim Xi = n − i, is dominant. Section I will be devoted to the proof of: 1 2 F.Buonerba & F.Bogomolov Theorem 1. The following classes C (k, n) are dominant: (i) For n = 3, smooth threefolds with a smooth connected morphism onto a smooth curve. (ii) For any n and k a finite field, projective varieties admitting a connected morphism onto a smooth curve, with only one singular fiber whose singular locus consists of one ordinary double point. Let us give a quick indication of the proof. The idea is to construct fibrations using Lefschetz pencils. In fact, the existence of Lefschetz pencils on smooth projective varieties, [SGA7.2] XVII, combined with de Jong's alteration result, [dJ96] 3.1, immediately gives: Fact. For any field k and integer n ≥ 2, the class of projective varieties admitting a connected morphism onto P1 k, with isolated singular fibers whose singular locus consists of one ordinary double point, is dominant. Statement (i) of the Theorem is then an immediate consequence of the Brieskorn- Tyurina's simultaneous resolution of surface ordinary double points, which in fact provides a simultaneous resolution of the fibers of the fibration induced by the Lefschetz pencil. Statement (ii) is more delicate. We are given X a smooth and projective over a finite field, with X ∨ the dual variety of singular hyperplane sections. We construct a curve C, in the space of hyperplanes of X, that intersects X ∨ in a single point, which is a smooth point for both X ∨ and C. In order to perform this construction, the crucial assumption, that k is a finite field, manifests itself in that the Picard group - of degree zero divisor classes - is always finite for projective curves. The total space of the induced family of hyperplane sections has a natural fibration onto C with the required conditions on the fibers, and by construction it has a surjective morphism onto X. In section II we focus our attention on surfaces. In this case the dual variety X ∨ stratifies according to geometric genus of the generic member, and therefore one might try to un- derstand the geometry and the modularity of such strata. The general idea is outlined in Proposition 10. The geometric structure of the stratification provides satisfactory answers for surfaces of negative Kodaira dimension, on which complete families of curves with con- stant geometric genus are easily constructed in Proposition 11. Such families induce, upon normalization of the total space, an equigeneric fibration with smooth general member. Un- fortunately the situation becomes complicated in non-negative Kodaira dimensions, where the method doesn't provide any obvious answer on general K3's and hypersurfaces in P3 Dominant classes of projective varieties 3 of degree at least 4. The difficulty here is that the strata might be nested into each other as divisors, and it seems hard to insert a complete curve between two consecutive ones. Therefore we turn our attention to the modularity of the stratification. Less vaguely, each stratum Sk, the closure of the set of curves with geometric genus gk, admits a rational gk to Sk. Question map Sk 99K M 14 summarizes the difficulty with this approach, due essentially to the non-genericity of gk , and one might try to lift complete curves from M the image of this rational map, whose behaviour near the Deligne-Mumford boundary is, a priori, arbitrary. In section III we continue our discussion by pointing out that the category of surfaces that can be dominated by complete families of smooth curves, which conjecturally is every- thing, is in fact rather flexible. By this we mean that it is somewhat natural and easy to create new smooth fibrations out of old ones, using ideas inspired by Kodaira's construc- tion, [Ko67], of non-isotrivial smooth fibrations. First we show in Proposition 15, that any product of two curves can be dominated by a non-isotrivial smooth fibration. Finally we prove, in Proposition 17, that any two smooth fibrations can be dominated simultaneously by a third one. In section IV, we conclude our discussion by analyzing several aspects of surfaces of general type that carry an everywhere smooth foliation. It turns out, Proposition 19, that such a surface must have positive topological index, which leads one to think about Kodaira fibrations. Indeed Brunella, in [Br97], has set the the foundations of the uniformization theory of such foliated surfaces, by establishing that the universal cover has the structure of a disk bundle over a disk. This, together with Corlette-Simpson's classification, [CS08], of Zariski dense Kahler representations in P SL2(R) leads us to Theorem 21: a smoothly foliated surface of general type is either a Kodaira fibration, or a foliated subvariety of a polydisk quotient. It is extremely reasonable that, in the second case, our initial surface is itself a bidisk quotient. We remark that the existence of ball and bidisk quotients can be used to prove, as in Propo- sition 22, that surfaces with an ´etale cover which is a Stein submanifold in a 3-dimensional ball form a dominant class. Switching our attention to smooth foliations on surfaces over fields of characteristic p > 0, the situation becomes as different from the complex case, as pleasant. Indeed we have: 4 F.Buonerba & F.Bogomolov Theorem 2. Let k be a field with p := char k > 2, and let X/k be an algebraic surface. Then there exists a birational modification of X, followed by an inseparable cyclic cover, such that the resulting surface Y carries a smooth p-closed foliation. It is worth remarking that the construction of Y is extremely generic, in that we start with a general Lefschetz pencil in X, pick a general curve going through the nodes in the pencil, and finally take an inseparable cyclic cover branched along such curve. The foliation defining the Lefschetz pencil is shown to pull back, upon saturation, to a smooth and p-closed foliation, via a trivial local computation. Due to the lack of Brunella's theorem in positive characteristic, the uniformization-type consequences that can be deduced from this statement, if any, are completely mysterious. Acknowledgements We are deeply grateful to Michael McQuillan for his interest in this work, and for fundamental criticism of an attempt to prove the main conjecture over fi- nite fields. We would also like to thank Misha Gromov, for several insights and pleasant conversations. The second author acknowledges that the article was prepared within the framework of a subsidy granted to the HSE by the Government of the Russian Federation for the implementation of the Global Competitiveness Program. The second author was partially supported by EPSRC programme grant EP/M024830, Simons Fellowship and Si- mons travel grant. I. Proof of theorem 1 Let X be a smooth projective variety of dimension n over a field k, and let PX denote the projective space of hyperplane sections of X, with universal family UX −−−−→ X uy PX Let us prove (i). Assume n = 3, and consider a Lefschetz pencil f : P1 k → PX. There is k such that (f ∗UX)s has a single ordinary double point if a Zariski-closed subset S ⊂ P1 and only if s ∈ S. By the Brieskorn-Tyurina's simultaneous resolution of ordinary double points of surfaces, [A74], there is a ramified cover C → P1 k, and a birational morphism Z → C ×P1 (cid:3) f ∗UX such that the composite Z → C is a smooth morphism. k Dominant classes of projective varieties 5 What follows is the proof of (ii). Assume k is a finite field, and denote by X ∨ ⊂ PX the dual variety of singular hyperplanes. It is well known, [SGA7.2] XVII, that upon replacing the projective embedding of X with a multiple, X ∨ is an irreducible divisor inside PX , whose smooth locus corresponds to hyperplane sections with a unique singular point, which is an ordinary double point. Claim 3. In order to conclude the proof of the theorem, it is enough to find an irreducible curve C ⊂ PX such that the intersection C ∩ X ∨ is supported in a single point, which is smooth for both C and X ∨. Proof. Let f : C → PX be such curve, and let uC : f ∗UX → C. uC has a unique singular fiber, whose singular locus is a single ordinary double point, lying over a smooth point of C. Therefore the induced fibration f ∗UX ×C C norm → C norm is the required one. (cid:3) The rest of the proof will be devoted to the construction of such curve C. Let S ∼−→ P2 k be a general linear plane inside PX , intersecting the divisor X ∨ along an irreducible, reduced curve D. Denote by d = degS(D). Recall the well known Fact 4. Since k is finite, the group Pic0 D/k of degree zero divisor classes on D is finite. Let N > d be an integer that kills Pic0 D/k. Denote by H an hyperplane section of S and by x a smooth point of D. By Fact 4 there exists an isomorphism of sheaves where of course M = N d, inducing a commutative diagram OD(N H) ∼−→ OD(M x) 0 −−−−→ OS(N H − D) −−−−→ OS(N H) −−−−→ OD(N H) −−−−→ 0 y Ker γ (cid:13)(cid:13)(cid:13) y 0 −−−−→ −−−−→ OS(N H) −−−−→ OD(M x) −−−−→ 0 γ whose vertical arrows are all isomorphisms. By our choice of N we deduce that Ker γ has non-trivial global sections, and the points of P(H 0(S, Ker γ)) correspond to curves of degree N inside S, whose intersection with D is supported on x. All is left to do is to check that the generic member of the linear system P(H 0(S, Ker γ)) is smooth at x. This is achieved by way of: Lemma 5. Let G, L ∈ k[X, Y, Z] be homogeneous polynomials of degrees d and N > d respectively. Assume that they define irreducible curves intersecting only at x = [0 : 0 : 1] 6 F.Buonerba & F.Bogomolov and that the curve defined by G is smooth at x. If the curve defined by L is singular at x, then F = Z N −dG + L satisfies (i) the curves defined by F and G meet only in x, and (ii) the curve defined by F is smooth at x. Proof. {x} = Supp((G = 0) ∩ (L = 0)) = Supp((G = 0) ∩ (F = 0)), which is (i). In order to prove (ii), observe that x is the origin of the affine space Z = 1, so then denoting by u = X/Z, v = Y /Z and g(u, v) = Z −dG(X, Y, Z), l(u, v) = Z −N L(X, Y, Z) we see that (F = 0) is defined, around x, by the vanishing of f = g + l. Since ∇f (0, 0) = ∇g(0, 0) 6= 0 the proof is complete. (cid:3) Consequently the generic member of P(H 0(S, Ker γ)) is smooth at x, and can be taken to be the curve C we are looking for. (cid:3) II. An approach to the conjecture for surfaces Let us describe some simple examples of surfaces for which Conjecture 1 holds: Example 6. Minimal models of surfaces of negative Kodaira dimension: apart from P2, these are P1-bundles over smooth curves. Example 7. Some surfaces of Kodaira dimension 0: • Abelian varieties: Let A be a d-dimensional abelian variety, and f : C → A any non-constant algebraic curve. The sum morphism C d → A given by (c1, ..., cd) → f (c1) + ... + f (cd) is surjective. • Kummer K3 surfaces: Let C be a genus 2 curve with hyperelliptic involution ι. Consider the sequence of morphisms C × C → Sym2 C → Jac2(C), (c1, c2) → c1 + c2 → [c1 + c2] The graph Γι of the hyperelliptic involution is projected onto a rational curve by the first morphism, and it is contracted by the second. Pulling back the above diagram under a degree 16 cover m2 : Jac2(C) → Jac2(C) - obtained by identifying Jac2(C) ∼−→ Pic(C) via KC - we obtain morphisms m∗ 2(C × C) → m∗ 2 Sym2 C → m∗ 2 Sym2 C/(−1) = K3(C) Since m2 is ´etale, m∗ 2(C × C) is still a product of curves. Dominant classes of projective varieties 7 The previous example generalizes as follows: Proposition 8. Let k be a finite field, and S0 ⊂ P2(k) any finite set. Then there exist curves C1, C2 and a morphism C1 × C2 → BlP2(S0). Proof. Observe that, for any such S0, there exists a finite set S ⊂ P1(k) and a morphism BlP1×P1(S × S) → BlP2(S0). Let E1 be an elliptic curve, and C a curve of genus 2 with a non-constant morphism C → E1. We can assume wlog that the Jacobian of C is isomorphic to E1 × E2, for some elliptic curve E2. Consider non-constant morphisms gi : Ei → P1 and the resulting (g1, g2) : BlE1×E2(g∗ 1S × g∗ 2S) → BlP1×P1(S × S) By our assumption on k, any finite set of points in E1 × E2 is torsion, hence there exists an integer N and a morphism m∗ N BlE1×E2(0 × 0) → BlE1×E2(g∗ 1S × g∗ 2S) where mN is the multiplication by N map on E1 × E2. Finally, composing C × C → Sym2 C ∼−→ BlE1×E2(0 × 0) with the morphisms constructed above, yields m∗ N (C × C) → m∗ N BlE1×E2(0 × 0) → BlE1×E2(g∗ 1S × g∗ 2S) → BlP1×P1(S × S) In order to conclude, we employ [SGA1] X.1.7 to find curves C1, C2 and an isomorphism C1 × C2 N (C × C). ∼−→ m∗ (cid:3) Since every Hirzebruch surface is dominated by a blow-up, in a finite set of points, of P1 × P1, we obtain: Corollary 9. Let X be a Hirzebruch surface over a finite field, and S ⊂ X(k) any finite set of points, then there exist curves C1, C2 and a morphism C1 × C2 → BlX(S). Turning to a more general discussion, let k be a field, X/k a smooth projective surface, and H an ample divisor such that H 1(X, OX (nH)) = H 2(X, OX (nH)) = 0 for all n ≥ 1. We have the linear system PX,n = P(H 0(X, OX (nH))) of dimension dn = dim(PX,n) = dim H 0(X, OX (nH)) − 1 = nH · (nH − KX)/2 + χ(OX) − 1 8 F.Buonerba & F.Bogomolov and its generic member is smooth of genus gn = 1 + nH · (nH + KX)/2. Observe that PX,n admits a natural stratification ∅ =: SN (n)+1 n ( SN (n) n ( ... ( S0 n := PX,n by, not necessarily irreducible, closed subvarieties, such that the generic member of each irreducible component of Sk n, 0 ≤ k ≤ N (n), corresponds to a curve with geometric genus at most gk n = gn. We will call it the Severi stratification. For gk n, and of course 0 ≤ gN (n) < ... < g0 n n ≥ 2 there is rational map into the Deligne-Mumford compactification of the moduli stack of curves. The reason to p : Sk n 99K M gk n introduce the Severi stratification is: Proposition 10. gk n (i) Assume that, for some 0 ≤ k ≤ N (n) such that gk ≥ 2, there exists a smooth proper curve C and a non-constant morphism f : C → Sk n such that the rational extends, after a finite cover C ′/C, to a morphism C ′ → Mgk . map p◦f : C 99K M Then, upon replacing C by a finite cover, there exists a smooth surface Y , a smooth fibration Y → C, and a surjective morphism Y → X. (ii) For some 0 ≤ k ≤ N (n), there exists a smooth proper curve C and a non-constant morphism f : C → Sk . Then, upon replacing C by a finite cover, there exists a smooth surface Y , a fibration Y → C whose fibers have constant geometric genus, and a surjective morphism Y → X. n \ Sk+1 n Proof. (i) As in the previous section, PX,n is equipped with a universal space u : UX,n → PX,n, and we can consider the fibration uC : f ∗UX,n → C. The normalization : (f ∗UX,n)norm → C induces the moduli morphism p ◦ f : C 99K M gk and unorm C therefore after replacing C by C ′, the resulting fibration is smooth. (ii) Since the curves in our family uC are equigeneric - meaning that the geometric is again an equigeneric genus is constant along the fibers - we have that unorm family, with smooth general member. C (cid:3) Dominant classes of projective varieties 9 The natural problem becomes to investigate to what extent Proposition 10 can be ap- plied. The next proposition shows how the geometry of the moduli M3 can be used to construct complete families of smooth curves mapping to the plane, using 10.(i): Fact. There exists a complete family p : Y → C of smooth genus 3 curves, whose fibrewise canonical map p∗ωY /C : Y → C × P2 k realizes the generic curve as a smooth quartic, and the special ones as double conics with smooth support. Proof. The Torelli map M3 → A3 is bijective on closed points, so there is a canonical Baily-Borel compactification M BB , which is projective and with boundary in codimension 2. Therefore, a generic complete intersection curve C ⊂ M BB is contained in M3 and intersects the hyperelliptic locus H3 ⊂ M3 transversely. Moreover the canonical map ωC : C → P2 k realizes points of H3 as plane double conics, and points of M3 \ H3 as smooth plane quartics. (cid:3) 3 3 We refer to [HM98], pp.133 for a discussion around the modular behavior of families plane quartics degenerating to double concics. Something can be said on surfaces with negative Kodaira dimension, indeed 10.(ii) quickly proves: Proposition 11. Let X be a smooth surface with negative Kodaira dimension. Then for n n n \ Sk+1 . Therefore, X can be dominated by an equigeneric family of curves. n sufficiently big, there exists a smooth proper curve C and a non-constant morphism f : C → Sk Proof. Consider the following inductive construction: set S0 := S0 n, and assuming defined Sk, let Sk+1 ⊂ Sk be an irreducible component of Sk+1 of maximal dimension. We claim that, for some k, dim Sk − dim Sk+1 ≥ 2: by assumption the canonical bundle of X is not pseudoefective, therefore KX · H < 0, so then dn − gn = −nH · KX − 2 + χ(OX ) is positive for n sufficiently big. Hence it is impossible to have dim Sk = dim S0− k = dn − k for every k. (cid:3) Remark 12. Unfortunately, this is still not enough to prove the conjecture in negative Ko- daira dimension: obviously, there exist equigeneric families of curves with smooth generic member, yet carrying singular, necessarily not irreducible, members. The situation becomes more interesting in non-negative Kodaira dimension, where the genus gn tends to be bigger than the dimension dn, and the stratification might consist of strata of consecutive codimension one. In fact this happens on general hypersurfaces: 10 F.Buonerba & F.Bogomolov Theorem 13. [CC98] Let X be a general hypersurface of P3 of degree d ≥ 4. For n ≥ d and any 0 ≤ k ≤ dn, the variety Sk n contains an irreducible component of dimension dn − k whose generic point parametrizes curves with k nodes. In this situation, the best we can hope in order to construct a smooth family is a positve answer to: Question 14. Can we find n, k such that the boundary divisor ∆gk the closure of p(Sk n ⊂ M gk n n), is not ample ? Even better, admitting a contraction ? , restricted to A positive answer to the above would provide us with a curve to which apply Proposition 10 (i), and hence prove the conjecture in dimension 2. The problem is that the image of p in the moduli M gk n is going to be of high codimension and extremely non-generic. For example, consider the natural rational map p : M gn 99K M BB gn Then, unless X is dominated by an isotrivial surface, p(Sk albeit p does restrict to a contraction along ∆gk . n n) ∩ ∆gk n is not contracted by p, III. flexibility of Kodaira fibrations In this section we emphasize that the class of Kodaira fibrations, i.e. those surfaces admitting a non-isotrivial smooth morphism onto a smooth curve, is remarkably flexible, and there plenty of smooth fibrations that can be constructed out of given ones. First, let us review Kodaira's original construction, [Ko67]: given any curve C0 of genus at least 2, let C → C0 be any non-trivial, finite ´etale cover with Galois group Γ. Consider, for any mΓ, the natural quotient π1(C) → H1(C, Z/mZ), and the corresponding ´etale cover f : C ′ → C with Galois group H1(C, Z/mZ). The crucial observation is that, by the Kunneth formula, the class of the graph Γf inside H 2(C ′ × C, Z/mZ) depends uniquely on the morphism f ∗ : H q(C, Z/mZ) → H q(C ′, Z/mZ). By construction, this morphism is trivial when q = 1, 2, while it is an isomorphism when q = 0. In particular, the cohomology class of Γγ◦f , γ ∈ Γ is independent of γ. Since mΓ, we deduce that D := ∪γΓγ◦f is m- divisible in H 2(C ′ × C, Z). Let X = X(C, m) → C ′ × C be the cyclic covering of order m, branched along D. Since the Γγ◦f are pairwise disjoint and each of them is an ´etale multisection of the second projection p2 : C ′ × C → C, we deduce that the composition Dominant classes of projective varieties 11 X → C is a non-isotrivial smooth fibration. We now employ Kodaira's construction as follows: Proposition 15. Given two curves C1, C2, there exists a curve C, a smooth non-isotrivial fibration in smooth curves Y → C and a finite morphism Y → C1 × C2. Proof. Let C0 be a curve of genus at least 2 with two surjective morphisms fi : C → Ci. Running the above construction, we obtain a Kodaira fibration X, with a natural sequence of finite morphism X → C ′ × C → C × C → C1 × C2. (cid:3) The class of surfaces of general type that are finite quotients of products of curves is vast. [BCF15], and references therein, point to a detailed study of the class of so-called product-quotient surfaces, which are, by the previous proposition, also dominated by non- isotrivial smooth fibrations. Similarly to what has been done in Proposition 15, let C of genus g ≥ 2, and consider the divisor ∆m ⊂ Symm C inside the m-fold symmetric product of C, of points (c1, ..., cm) with ci = cj for some i 6= j. Proposition 16. Let f : D → Symm C \ ∆m be a non-constant morphism from a complete curve D. Then there exists a smooth non-isotrivial fibration Y → D and a finite morphism Y → C × D. Proof. The morphism f defines a divisor Df ⊂ C × D, whose fiber over d ∈ D is the m-tuple of points f (d) ⊂ C, and by construction the projection p : Df → D is ´etale. As before, upon replacing D by a non-trivial Galois cover, one constructs Kodaira fibrations as cyclic covers of Df × D, branched along the Galois orbit of the graph of p. (cid:3) Observe that this construction provides many examples when m = 2, since ∆ ⊂ Sym2 C can be contracted via Sym2 C → Jac0(C)/(−1), (c1, c2) → [c1 − c2]. The case m ≥ 3 is more subtle, since the diagonal ∆m ⊂ Symm C lies on the boundary of the effective cone of Symm C, and even its numerical properties seem rather mysterious. In a deeper vein, the next proposition shows that pairs of Kodaira fibrations admit common refinements, meaning that they can dominated simultaneously by a third Kodaira fibration. Proposition 17. Let q1 : X1 → C1 and q2 : X2 → C2 be smooth fibrations in curves. Then there exists a smooth curve C0 with surjective morphisms ti : C0 → Ci, a fibration in curves X → C0 and finite C0-morphisms X → t∗ i Xi, i = 1, 2. 12 F.Buonerba & F.Bogomolov Proof. We first fix some notation: for a given curve D, denote by Mg(D) the substack of Mg parametrizing smooth genus g curves C that admit a surjective morphism C → D. Similarly, for a given pair of curves D1, D2, denote by Mg(D1, D2) the substack of Mg parametrizing smooth genus g curves C that admit surjective morphisms C → D1 and C → D2. The above admits a relative analogue, in that if D1 → B and D2 → B are two families of curves, we have an induced fibration Mg(D1, D2/B) → B, whose fiber over b ∈ B is Mg(D1,b, D2,b). A trick by Kodaira, [HM98] 2.33, shows that for every n, there exists g such that Mg(D) contains complete subvarieties of dimension n, and therefore the same holds for Mg(D1, D2). We are now ready to offer a proof of the proposition: let C be a curve with surjective morphisms ti : C → Ci and replace qi by pi : Yi := t∗ i Xi → C. Consider the fibration π : Mg(Y1, Y2/C) → C. By the above remarks, for g big enough there exists a surface Z ⊂ Mg(Y1, Y2/C) such that πZ : Z → C is surjective and the fiber Z ∩ π−1(c) is a complete curve for generic c ∈ C. Such morphism πZ clearly admits a multisection C0 - for example by taking an ample divisor in Z, the closure of Z in Mg(Y1, Y2/C), missing the isolated boundary points - and such multisection C0 induces, by definition, a smooth fibration in curves X → C0 dominating Y1/C and Y2/C. (cid:3) IV. Smooth foliations on surfaces of general type Motivated by the study of surfaces that carry smooth fibrations, we dedicate this final section to surfaces of general type carrying smooth foliations. In particular we look for restrictions the existence of a smooth foliation imposes on the ambient surface, and then try to understand what they might possibly look like. Recall that a smooth foliation F on an algebraic surface X defines an exact sequence of vector bundles The next fact recollects some well known numerical properties of KF := T F ∨. 0 → T F → T X → N F → 0 Fact 18. (i) c2(X) = KF · (KX − KF ). (ii) (KX − KF )2 = 0. (iii) c2(X) = K 2 (iv) If X is of general type then KF is pseudoeffective, and it is big unless F is either X − KF · KX . an isotrivial fibration, or a Hilbert modular foliation. Dominant classes of projective varieties 13 Proof. (i) This follows by taking Chern numbers in the defining exact sequence of T F , plus c2(X) > 0. (ii) This is the Baum-Bott index formula, N F 2 = 0. (iii) This is a formal consequence of (i) and (ii). (iv) The pseudoeffectivity of KF follows from the Main Theorem of [BM16], since X has general type. By the birational classification of foliations on surfaces, [McQ08], if a foliation on a surface of general type is such that KF is not big, then the foliation is an isotrivial or Hilbert modular. (cid:3) Therefore, granted a decent amount of understanding of isotrivial and Hilbert modular surfaces, we concentrate on the strong topological and algebraic restrictions, for a surface of general type to carry a smooth foliation. Proposition 19. Let X a smooth surface of general type, and F an everywhere smooth foliation on X with KF big. Then we have • rank Pic(X) ≥ 2. • X has positive topological index, i.e. c1(X)2 > 2c2(X). Proof. Let a := KX ·KF index theorem, that R2 < 0 unless R is numerically trivial. , and R := aKX − KF . Since R · KX = 0 we deduce, by Hodge K 2 X Claim 20. R 6= 0 Proof. We have 0 = R · KX = −KF · KX + K 2 X + (a − 1)K 2 X = c2(X) + (a − 1)K 2 X so then If R = 0 then aKX = KF , and (KX − KF )2 = 0 implies a = 1, from which c2(X) = 0, impossible since X has general type. (cid:3) c2(X) = (1 − a)K 2 X From which we deduce that rank Pic(X) ≥ 2. Let P ⊆ N S(X) be the plane spanned by KX and R. We know that Amp(X) = {D : D2 > 0, D · KX > 0} 14 F.Buonerba & F.Bogomolov so then Fact 18 (i) and (ii) imply that KX − KF = (1 − a)KX + R lies on the boundary of Amp(X) ∩ P . It follows that Amp(X) ∩ P is bounded by the rays (1 − a)KX + R, (1 − a)KX − R Since KF = aKX − R ∈ Amp(X) ∩ P by assumption, we have a > 1 − a, or a > 1/2. The identity c2(X) = (1 − a)K 2 derived in the proof of Claim 20, concludes the proof. X (cid:3) More interestingly, Brunella [Br97] has initiated the uniformization theory on smoothly foliated surfaces (X, F ) of general type: the universal cover X of such a surface, admits a smooth holomorphic fibration pF : X → ∆ onto the unit disk, with disks as fibers, such that F becomes tangent to pF on X. In particular, X is a K(Γ, 1), for Γ := π1(X). With the aim of better understanding the geometry of X, observe that the fibration pF is naturally Γ-equivariant, and hence there is a natural representation ρ : Γ → P SL2(R). Corlette and Simpson, in [CS08], have given a complete classification of what such a ρ can be, and deduced a beautiful dichotomoy for its geometric origin. We can summarize all of this in: Theorem 21 (Brunella, Corlette & Simpson). Let (X, F ) be a smoothly foliated surface of general type. Then at least one of the following happens: • X admits a smooth fibration p : X → C onto a smooth curve C and F is tangent to p; • ρ is rigid and integral, there exists a quasiprojective polydisk quotient, Y , and a natural morphism X → Y such that F is induced by one of tautological codimension one foliations on Y . The only issue to be settled here, is the relation between the dimensions of X and Y . It seems plausible, thinking about the wild behaviour of Hilbert modular foliations, that indeed the dimensions must be the same, and X itself is a bidisk quotient. Let us remark that bounded symmetric domains can be used in our problem of finding dominant classes as follows: Dominant classes of projective varieties 15 Proposition 22. Let X be a smooth algebraic surface. Then there exists a finite morphism Y → X, such that an ´etale cover Y ′ of Y is a Stein submanifold of a 3-dimensional ball. Proof. Let Z be a ball quotient, consider projections X → P2, Z → P2, and let Y := X ×P2 Z. For sufficiently generic projections, the branch loci in P2 are transverse, hence Y is smooth. Let p : B2 → Z be the universal cover, then Y ′ := p∗Y is a complex manifold which is naturally a finite ramified cover of the ball B2, and as such embeds into a product B2 × ∆. (cid:3) We wish to conclude by turning our attention to the corresponding problem in charac- teristic p > 0, of understanding smooth foliations on algebraic surfaces. Perhaps not too surprisingly, the situation is drastically different from the complex case, and it turns out that it is extremely simple to construct a covering of a given surface that carries a smooth foliation. Before proceeding to our main result, let us recall a key difference between fo- liations in characteristic 0 and p, that lies in the notion of integrability: the celebrated Frobenius integrability theorem, as well known for integrable distributions over C, fails over fields of positive characteristic: integrable distributions need not have a formal first integral. Observe, indeed, that the notion of leaf is rather badly behaved, and the closest we can get to formal integrability is p-closedness - that is, the algebra generated by the vector fields defining our foliation is closed under p-powers. This implies that the kernel of such algebra of differential operators defines a factorization of the Frobenius morphism on the ambient variety. And a modicum of thought shows that this is the best integrability condition one can hope for. In practice, probably the quickest way of appreciating the role of p-closedness in remedying the failure of formal integrability is: Fact 23. [McQ08, II.1.6] Let A be a complete regular local ring over an algebraically closed field of characteristic p, and let ∂ be a non-singular derivation of A. Then there exists a regular system x, y1, ..., yn of parameters such that ∂ = ∂ ∂x + xp−1 n X i=1 fi(xp, y) ∂ ∂yi Moreover the ideal defining the vanishing of ∂ ∧ ∂p is generated by f1, ..., fn. In particular ∂ is p-closed iff ∂ = ∂ ∂x , i.e. a smooth foliation by curves is p-closed iff it is formally integrable. 16 F.Buonerba & F.Bogomolov IV.1. Proof of Theorem 2. Let p : Y0 → P1 k be a general Lefschetz pencil in X, whose general member is smooth, and whose singular members have exactly one node. Observe that Y0 is obtained by blowing up X along the base points of the pencil. Let S ⊂ P1 parametrize the singular fibers, and for each s ∈ S let ys ∈ Ys be the node in the fiber. Let H be a very ample divisor, and consider the sheaf O(pH) ⊗ I∪sys whose local sections are those of O(pH) vanishing along ∪sys. Its generic global section defines a curve, D, which is smooth, transverse to the branches of Ys at ys, and has simple tangencies with the fibers of p outside their nodes. The following claim will conclude: k Claim 24. Let rD : Y := Y0( p√D) → Y0 denote the inseparable cyclic p-cover, branched along D. If F denotes the foliation defined by p, then the saturation of r∗ D F is smooth and p-closed. Proof. First we deal with smoothness. We need to worry about what happens at: (i) The nodes in the fibers of p, and (ii) The simple tangencies between D and the smooth locus of the fibers of p. Let us consider (i). In the local ring of Y0 completed in a node in a singular fiber, the branches of the singular fiber give us local coordinates x, y, while D is defined by the vanishing of a third local function, z. F is defined by the vanishing of the form d(xy), and our assumptions on D imply that there exists a formal function G such that y = G(x, z) holds. The cyclic cover rD is defined by r∗ Dz = zp, hence r∗d(xy) = d(x · G(x, zp)) = (G + x ∂G ∂x )dx and therefore, upon saturation, r∗ D in the fibers of p. Let us deal with (ii). In the local ring of Y0 completed in a point of simple tangency, we can pick local coordinates x, y such that our fibration is defined by the vanishing of the F is smooth around the pre-image, under rD, of nodes form dy. If the vanishing of z defines D in such coordinates, simple tangency means that there exists a formal function g such that z = y − g(x) and ordx g(x) = 2. In particular, dg(x) dx The cyclic cover rD is defined by is not identically zero, since p > 2. r∗ Dz = zp, r∗ Dx = x, r∗ Dy = zp + g(x) Dominant classes of projective varieties 17 and the pullback foliation is then r∗ Ddy = d(zp + g(x)) = dg(x) dx dx which is again smooth upon saturation. Observe, as an output of our local computations, that the saturation of r∗ D smooth, but also formally integrable. By Fact 23, it is p-closed. F is not only (cid:3) References [A74] M. Artin, Algebraic construction of Brieskorn's resolutions, J. Alg., Vol. 29, Issue 2, May 1974, pp. 330-348, http://www.sciencedirect.com/science/article/pii/0021869374901021 [BCF15] I. Bauer, F. Catanese, D. Frapporti, The fundamental group and torsion group of Beauville surfaces, Beauville surfaces and groups, 114, Springer Proc. Math. Stat., 123, Springer, Cham, 2015. 14Fxx (14J25), http://arxiv.org/abs/1402.2109 [BH00] F. Bogomolov, D. Husemoller, Geometric properties of curves defined over number fields, MPI preprint , 2000-1, http://www.mpim-bonn.mpg.de/preprints [BM16] F. Bogomolov, M. McQuillan, Rational curves on foliated varieties, Foliation Theory in Algebraic Geometry. Springer International Publishing, 2016. 21-51, http://link.springer.com/content/pdf/10.1007%2F978-3-319-24460-0_2.pdf [BT02] F. Bogomolov, Y. Tschinkel, Unramified correspondences, Algebraic number theory and alge- braic geometry, Contemp. Math., vol. 300, Amer. Math. Soc., Providence, RI, 2002, pp. 17-25, http://arxiv.org/abs/math/0202223 [BT02'] F. Bogomolov, Y. Tschinkel, On curve correspondences, Communications in Arithmetic Fundamental Groups and Galois Theory, RIMS, pp.157-166, 2002, [BT05] http://www.math.nyu.edu/~tschinke/papers/yuri/02genram/genram.pdf F. Bogomolov, Y. Tschinkel, Couniformization of curves over number fields, Geometric meth- ods in algebra and number theory, pp. 43-57, Progress in Math. 235, Birkhuser, 2004 , [Br97] http://www.math.nyu.edu/~tschinke/papers/yuri/04covers/cover4.pdf M. Brunella, Feuilletages holomorphes sur les surfaces complexes compactes, Annales scientifiques de l'´Ecole Normale Sup´erieure (1997) Volume: 30, Issue: 5, page 569-594 http://archive.numdam.org/ARCHIVE/ASENS/ASENS_1997_4_30_5/ASENS_1997_4_30_5_569_0/ASENS_1997_4_30_5_569_0.pdf [CC98] L. Chiantini, C. Ciliberto, On the Severi varieties of surfaces in P3, http://arxiv.org/abs/math/9802009 [CS08] K. Corlette, C. Simpson, On the classification of rank-two representations of quasiprojective fundamental groups, Compos. Math. 144 (2008), 12711331, https://arxiv.org/pdf/math/0702287.pdf [SGA7.2] P. Deligne, N. Katz, S´eminaire de G´eom´etrie Alg´ebrique du Bois Marie - 1967-69 - Groupes de monodromie en g´eom´etrie alg´ebrique - (SGA 7) - 18 F.Buonerba & F.Bogomolov vol. 2. LNM 340. Berlin, New York: Springer-Verlag. pp. x+438, 1973, https://publications.ias.edu/sites/default/files/Number12.pdf [dJ96] [SGA1] A.J. de Jong, Smoothness, semi-stability and alterations, Publications Math´ematiques de l'IH´ES 83, 1996, pp. 51-93, http://www.math.uiuc.edu/K-theory/0081 A. Grothendieck, S´eminaire de G´eom´etrie Alg´ebrique du Bois Marie - 1960-61 - Revetements ´etales et groupe fondamental - (SGA 1). LNM 224. Springer-Verlag. pp. xviii+327, 1971, https://arxiv.org/abs/math/0206203 [HM98] J. Harris, I. Morrison, Moduli of curves, Graduate Texts Mathematics, 187. Springer-Verlag, New York, 1998. xiv+366 in pp, http://link.springer.com/content/pdf/10.1007%2Fb98867.pdf [Ko67] K. Kodaira, A certain type of irregular algebraic surfaces Journal d'Analyse Mathematique, vol. 19 (1967), pp. 207-215 [McQ08] M. McQuillan, Canonical models of foliations, Pure Appl. Math. Q. 4 (2008), no. 3, part 2, 8771012, http://www.mat.uniroma2.it/~mcquilla/files/canmod.pdf Federico Buonerba Courant Institute of Mathematical Sciences, New York University, 251 Mercer Street, New York, NY 10012, USA E-mail address: [email protected] Fedor Bogomolov Courant Institute of Mathematical Sciences, New York University, 251 Mercer Street, New York, NY 10012, USA National Research University Higher School of Economics, Russian Federation, AG Laboratory, HSE, 7 Vavilova str., Moscow, Russia, 117312 E-mail address: [email protected] 

0905.1418

3

0905

On Grothendieck--Serre's conjecture concerning principal G-bundles over reductive group schemes:I

Let k be an infinite field. Let R be the semi-local ring of a finite family of closed points on a k-smooth affine irreducible variety, let K be the fraction field of R, and let G be a reductive simple simply connected R-group scheme isotropic over R. We prove that for any Noetherian k-algebra A, the map of etale cohomology sets H^1(A\otimes_k R,G)-> H^1(A\otimes_ k K,G), induced by the inclusion of R into K, has trivial kernel. This implies the Serre-Grothendieck conjecture for such groups G. The main theorem for A=k and some other results of the present paper are used significantly in arXiv:1211.2678 to prove the Serre-Grothendieck conjecture for all reductive groups over a regular semi-local ring containing an infinite field.

math.AG

math

On Grothendieck -- Serre's conjecture concerning principal G-bundles over reductive group schemes: I I. Panin∗, A. Stavrova†, N. Vavilov‡ October 25, 2018 Abstract Let k be an infinite field. Let R be the semi-local ring of a finite family of closed points on a k-smooth affine irreducible variety and let K be the fraction field of R and let G be a reductive simple simply connected R-group scheme isotropic over R. Our Theorem 1.1 states that for any Noetherian k-algebra A the kernel of the map H 1 ´et(R ⊗k A, G) → H 1 ´et(K ⊗k A, G) induced by the inclusion of R into K is trivial. Theorem 1.2 for A = k and some other results of the present paper are used significantly in [FP] to prove the Grothendieck-Serre's conjecture for regular semi-local rings R containing an infinite field. 1 Introduction Recall that an R-group scheme G is called reductive (respectively, semi-simple or simple), if it is affine and smooth as an R-scheme and if, moreover, for each ring homomorphism s : R → Ω(s) to an algebraically closed field Ω(s), its scalar extension GΩ(s) is a connected reductive (respectively, semi-simple or simple) algebraic group over Ω(s). The class of reductive group schemes contains the class of semi-simple group schemes which in turn contains the class of simple group schemes. This notion of a simple R-group scheme coincides with the notion of a simple semi-simple R-group scheme from Demazure -- Grothendieck [SGA3, Exp. XIX, Defn. 2.7 and Exp. XXIV, 5.3]. Throughout the paper R denotes an integral domain and G denotes a semi-simple R-group scheme, unless explicitly stated otherwise. All commutative rings that we consider are assumed to be Noetherian. ∗The author acknowledges support of the RFBR grant 13-01-00429-a †The author acknowledges support of the J.E. Marsden postdoctoral fellowship of the Fields Institute for Research in Mathematical Sciences, the RFBR grants 12-01-33057, 12-01-31100, 10-01-00551, 09-01- 00878, and of the research program 6.38.74.2011 "Structure theory and geometry of algebraic groups and their applications in representation theory and algebraic K-theory" at St. Petersburg State University. ‡The author acknowledges support of the RFFI projects 08-01-00756, 09-01-00762, 09-01-00784, 09- 01-00878 and 09-01-91333 1 A semi-simple R-group scheme G is called simply connected (respectively, adjoint), provided that for an inclusion s : R ֒→ Ω(s) of R into an algebraically closed field Ω(s) the scalar extension GΩ(s) is a simply connected (respectively, adjoint) Ω(s)-group scheme. This definition coincides with the one from [SGA3, Exp. XXII. Defn. 4.3.3]. A well-known conjecture due to J.-P. Serre and A. Grothendieck [Se, Remarque, p.31], [Gr1, Remarque 3, p.26-27], and [Gr2, Remarque 1.11.a] asserts that given a regular local ring R and its field of fractions K and given a reductive group scheme G over R the map H 1 ´et(R, G) → H 1 ´et(K, G), induced by the inclusion of R into K, has trivial kernel. The following theorem, which is one of the main result of the present paper, asserts that for simple and simply connected isotropic group schemes over certain rings R this is indeed the case (recall that a simple R-group scheme is called isotropic if it contains a split torus Gm,R). Actually, we prove something significantly stronger. Namely, Theorem 1.1. Let k be an infinite field. Let O be the semi-local ring of finitely many closed points on a k-smooth irreducible affine k-variety X and let K be its field of frac- tions. Let G be an isotropic simple simply connected group scheme over O. Then for any Noetherian k-algebra A the map H 1 ´et(O ⊗k A, G) → H 1 ´et(K ⊗k A, G), induced by the inclusion O into K, has trivial kernel. In other words, under the above assumptions on O and G each principal G-bundle P over O ⊗k A which is trivial over K ⊗k A is itself trivial. In the case A = k the main result of [FP] is much stronger, since there are no anisotropy assumptions on G there. However, it seems that there is no reason to expect that Theorem 1.1 holds for anisotropic G. Theorem 1.6 extends easily to the case of simply connected semi-simple group schemes, using the Faddeev -- Shapiro lemma. However, in this generality its statement is a bit more technical and we postpone it till Section 11 (see Theorem 11.2). All other results stated below extend to semi-simple simply connected group schemes as well. Theorem 1.1 is an easy consequence of the following two theorems. Theorem 1.2. Let k, O, K, A be the same as in Theorem 1.1. Let G be a not necessarily isotropic simple simply connected group scheme over O. Let G be a principal G-bundle over O ⊗k A which is trivial over K ⊗k A. Then there exists a principal G-bundle Gt over O[t] ⊗k A and a monic polynomial f (t) ∈ O[t] such that (i) the G-bundle Gt is trivial over (O[t]f ) ⊗k A, (ii) the evaluation of Gt at t = 0 coincides with the original G-bundle G, (iii) f (1) ∈ O is invertible in O. Theorem 1.3. Let k be a not necessarily infinite field. Let B be a Noetherian k-algebra. Let G be an isotropic simple simply connected group scheme over B, that is G contains 2 a torus Gm,B. Let Pt be a principal G-bundle over B[t] and let h(t) ∈ B[t] a monic polynomial such that (i) the G-bundle Pt is trivial over B[t]h, (ii) h(1) ∈ B is invertible in B. Then the principal G-bundle Pt is trivial. Remark 1.4. To prove Theorem 1.1 one needs to substitute in Theorem 1.3 B = O ⊗k A, Pt := Gt, h(t) = f (t) ⊗ 1 from Theorem 1.2. By Theorem 1.3 the G-bundle Gt is trivial. Now by the item (ii) of Theorem 1.2 the original G-bundle G is trivial. Remark 1.5. Theorem 1.3 does not hold in the case of anisotropic G even for B = O with O as in Theorem 1.2. There are various conterexamples. Only a weaker form of Theorem 1.3 holds in the case of anisotropic G and B = O as it is proved in [FP, Thm. 2]. This is why we are skeptical that Theorem 1.1 holds in the anisotropic case. Theorem 1.6. Let R be a regular semi-local domain containing an infinite field and let K be the fraction field of R. Let G be an isotropic simple simply connected group scheme over R, containing a split rank 1 torus Gm,R. Then for any Noetherian commutative ring A the map H 1 ´et(R ⊗Z A, G) → H 1 ´et(K ⊗Z A, G), induced by the inclusion R into K, has trivial kernel. Combining Theorem 1.6 with the well-known result of M.S. Raghunathan and A. Ramanathan [RR], we obtain the following Corollary 1.7. Let R be a regular domain containing Q. Let G be an isotropic simple simply connected group scheme over R, containing a split rank 1 torus Gm,R. Then the map ´et(R[t], G) → H 1 induced by evaluation at t = 0, has trivial kernel. H 1 ´et(R, G), To put these statements into context, let us recall other known results on the Serre -- Grothendieck conjecture. • The case where the group scheme G comes from the ground field k is completely solved by J.-L. Colliot-Th´el`ene, M. Ojanguren, M. S. Raghunatan and O. Gabber: in [C-T/O] and [R1], [R2] when k is infinite; O. Gabber [Ga] announced a proof for an arbitrary ground field k. • The case of an arbitrary reductive group scheme over a discrete valuation ring is completely solved by Y. Nisnevich in [Ni2]. • The case where G is an arbitrary torus over a regular local ring was settled by J.-L. Colliot-Th´el`ene and J.-J. Sansuc in [C-T/S]. • For most simple group schemes of classical series the Serre -- Grothendieck conjecture was solved in works of the first author, A. Suslin, M. Ojanguren and K. Zainoulline [PS], 3 [OP2], [Z], [OPZ]. In fact, unlike our Theorem 1.6, no isotropy hypotheses was imposed there. • The first author, the second author, and V. Petrov proved the Serre -- Grothendieck conjecture for strongly inner adjoint groups of type E6 or E7 [PPeSt], and for groups of type F4 with trivial f3 invariant [PeSt2], under the same assumptions on R. • V. Chernousov [Ch] established the Serre -- Grothendieck conjecture for groups of type F4 with trivial g3 invariant, under the assumption that R is a regular local ring containing a field of charactersitic 0. • In [P2] the first author reduced the Serre -- Grothendieck conjecture to the case of semi-simple simply connected group schemes (assuming that R is the semilocal ring of finitely many closed points on a k-smooth affine variety with an infinite field k). • R. Fedorov and the first author in [FP] prove the Serre -- Grothendieck conjecture for an arbitrary reductive group scheme over a semilocal regular ring containing an infinite field. Their work relies heavily on the results of the present paper and of [P2]. The authors would like to thank Vladimir Chernousov, Philippe Gille, Victor Petrov, and Konstantin Pimenov for useful discussions on the subject of the present paper. 2 Almost elementary fibrations In this Section we modify a result of M. Artin from [A] concerning existence of nice neighborhoods. The following notion is a modification of the one introduced by Artin in [A, Exp. XI, D´ef. 3.1]. Definition 2.1. An almost elementary fibration over a scheme S is a morphism of schemes p : X → S which can be included in a commutative diagram j X &▼▼▼▼▼▼▼▼▼▼▼▼▼ p / X p S Y i q xqqqqqqqqqqqqq (1) of morphisms satisfying the following conditions: (i) j is an open immersion dense at each fibre of p, and X = X − Y ; (ii) p is smooth projective all of whose fibres are geometrically irreducible of dimension one; (iii) q is a finite flat morphism all of whose fibres are non-empty; (iv) the morphism i is a closed embedding and the ideal sheaf IY ⊂ OX defining the closed subscheme Y in X is locally principal. 4 & /   o o x Remark 2.2. This definition is motivated by the following example. Take a field k and S = Spec(k), take X = P1 k − {y}, Y = y. Then the structure morphism X → S is an almost elementary fibration. If the field extension k(y)/k is purely inseparable, then X → S is not an elementary fibration in the sense of Artin [A, Exp. XI, D´ef. 3.1]. k, take a closed point y ∈ P1 k and set X = P1 We prove the following result, which is a slight modification of Artin's result [A, Exp. XI, Prop. 3.3]. Proposition 2.3. Let k be an infinite field, X be a smooth geometrically irreducible variety over k, x1, x2, . . . , xn ∈ X be closed points. Then there exists a Zariski open neighborhood X 0 of the family {x1, x2, . . . , xn} and an almost elementary fibration p : X 0 → S, where S is an open subscheme of the projective space PdimX−1. If, moreover, Z is a closed co-dimension one subvariety in X, then one can choose X 0 and p in such a way that pZ T X 0 : ZT X 0 → S is finite surjective. The proofs of the above Proposition and of the following one are provided in Ap- pendix, 12.1. Proposition 2.4. Let p : X → S be an almost elementary fibration. If S is a regular semi-local irreducible scheme, then there exists a commutative diagram of S-schemes X π A1 × S j in X π / P1 × S i i Y (2) {∞} × S such that the left hand side square is Cartesian. Here j and i are the same as in Definition 2.1, while prS ◦ π = p, where prS is the projection A1 × S → S. In particular, π : X → A1 × S is a finite surjective morphism of S-schemes, where X and A1 × S are regarded as S-schemes via the morphism p and the projection prS, respectively. 3 Nice triples In the present section we introduce and study certain collections of geometric data and their morphisms. The concept of a nice triple is very similar to that of a standard triple introduced by Voevodsky [Vo, Defn. 4.1], and was in fact inspired by the latter notion. Let k be an infinite field, let X/k be a smooth geometrically irreducible variety, and let x1, x2, . . . , xn ∈ X be its closed points. Further, let O = OX,{x1,x2,...,xn} be the correspond- ing geometric semi-local ring. Definition 3.1. Let U := Spec(OX,{x1,x2,...,xn}). A nice triple over U consists of the following data: 5 / /     o o   / o o (i) a smooth morphism qU : X → U, where X is an irreducible scheme, (ii) an element f ∈ Γ(X, OX), (iii) a section ∆ of the morphism qU , subject to the following conditions: (a) each irreducible component of each fibre of the morphism qU has dimension one, (b) the module Γ(X, OX)/f · Γ(X, OX) is finite as a Γ(U, OU ) = O-module, (c) there exists a finite surjective U-morphism Π : X → A1 × U, (d) ∆∗(f ) 6= 0 ∈ Γ(U, OU ). Definition 3.2. A morphism of two nice triples (q′ : X′ → U, f ′, ∆′) → (q : X → U, f, ∆) is an ´etale morphism of U-schemes θ : X′ → X such that (1) q′ U = qU ◦ θ, (2) f ′ = θ∗(f ) · h′ for an element h′ ∈ Γ(X′, OX′), (3) ∆ = θ ◦ ∆′. Two observations are in order here. • Item (2) implies in particular that Γ(X′, OX′)/θ∗(f ) · Γ(X′, OX′) is a finite O-module. • It should be emphasized that no conditions are imposed on the interrelation of Π′ and Π. Let U be as in Definition 3.1. Let (X, f, ∆) be a nice triple over U. Then for each finite surjective U-morphism σ : X → A1 × U and the corresponding O-algebra inclusion O[t] ֒→ Γ(X, OX) the algebra Γ(X, OX) is finitely generated as an O[t]-module. Since both rings O[t] and Γ(X, OX) are regular, the algebra Γ(X, OX) is finitely generated and projective as an O[t]-module by theorem [E, Cor. 18.17]. Let T r − an−1T r−1 + · · · ± N(f ) f be the characteristic polynomial of the O[t]-module endomorphism Γ(X, OX) −→ Γ(X, OX), and set gf,σ := f r−1 − an−1f r−2 + · · · ± a1 ∈ Γ(X, OX). (3) Lemma 3.3. f · gf,σ = ±N(f ) ∈ Γ(X, OX). Proof. Indeed, the characteristic polynomial of the operator Γ(X, OX) ishes on f . f −→ Γ(X, OX) van- Let us state two crucial results which will be used in our main construction. Their proofs are given in Sections 4 and 5 respectively. 6 Theorem 3.4. Let U be as in Definition 3.1. Let (X, f, ∆) be a nice triple over U, such that f vanishes at every closed point of ∆(U). There exists a distinguished finite surjective morphism σ : X → A1 × U of U-schemes which enjoys the following properties. (1) σ is ´etale along the closed subset {f = 0} ∪ ∆(U). (2) For gf,σ and N(f ) defined by the distinguished σ, one has σ−1(cid:16)σ(cid:0){f = 0}(cid:1)(cid:17) = {N(f ) = 0} = {f = 0} ⊔ {gf,σ = 0}. (3) Denote by X0 ֒→ X the largest open subscheme where the morphism σ is ´etale. Write g for gf,σ in this item. Then the square X0 N (f ) = X0 f g σ0 f g (A1 × U)N (f ) inc inc X0 g σ0 g / A1 × U (4) is an elementary Nisnevich square. More precisely, this square is Cartesian and the morphism of the reduced closed subschemes σ0 g{f =0}red : {f = 0}red → {N(f ) = 0}red of the schemes X0 g and A1 × U is an isomorphism. (4) One has ∆(U) ⊂ X0 g. Remark 3.5. One readily sees that if in Theorem 3.4 we let X0 be any open subscheme of X such that σ is ´etale on X0 and X0 contains the closed subset {f = 0} ∪ ∆(U), then all the claims of this theorem are still valid. In particular, if needed, one can assume that X0 is an affine scheme. Theorem 3.6. Let U be as in Definition 3.1. Let (X, f, ∆) be a nice triple over U. Let GX be a simple simply connected X-group scheme, and let GU := ∆∗(GX). Finally, let Gconst be the pull-back of GU to X. Then there exists a morphism θ : (X′, f ′, ∆′) → (X, f, ∆) of nice triples and an isomorphism of X′-group schemes such that (∆′)∗(Φ) = idGU . Φ : θ∗(Gconst) → θ∗(GX) 7 / /     / 4 Proof of Theorem 3.4 The nearest aim is to prove Theorem 3.4. We will use analogues of three lemmas from [P1] making them characteristic free. Lemma 4.3 is a refinement of [OP1, Lemma 5.2]. Lemma 4.1. Let k be an infinite field and let S be an k-smooth equidimensional k-algebra of dimension one. Let f ∈ S be a non-zero divisor. Let m0 be a maximal ideal with S/m0 = k. Let m1, m2, . . . , mn be pairwise distinct maximal ideals of S (possibly m0 = mi for some i). Then there exists a non-zero divisor ¯s ∈ S such that S is finite over k[¯s] and (1) the ideals ni := mi ∩ k[¯s], 1 ≤ i ≤ n, are pairwise distinct. If m0 is distinct from all mi's, then n0 := m0 ∩ k[¯s] is distinct from all ni's; (2) the extension S/k[¯s] is ´etale at each mi, i = 1, 2, . . . , n, and at m0; (3) k[¯s]/ni = S/mi for each i = 1, 2, . . . , n; (4) n0 = ¯sk[¯s]. Proof. Let xi, 0 ≤ i ≤ n, be the points on Spec(S) corresponding to the ideals mi. Consider a closed embedding Spec(S) ֒→ An k and find a generic linear projection p : An k, defined over k and such that the following holds: k → A1 (1) for all i, j ≥ 0 one has p(xi) 6= p(xj), provided that xi 6= xj; (2) for each index i ≥ 0 the map pSpec(S) : Spec(S) → A1 k is ´etale at the point xi; (3) for each i, the separable degree of the extension k(xi)/k(p(xi)) is one. These items imply equalities k(p(xi)) = k(xi), for all i. Indeed, the extension k(xi)/k(p(xi)) is separable by (2). By (3) we conclude that k(p(xi)) = k(xi). Lemma follows. Lemma 4.2. Under the hypotheses of Lemma 4.1 let f ∈ S be a non-zero divisor which does not belong to a maximal ideal distinct from m0, m1, . . . , mn. Let ¯s ∈ S be an element satisfying (1) to (4) of Lemma 4.1. Let N(f ) = NS/k[¯s](f ) be the norm of f . Then one has (a) N(f ) = f g for an element g ∈ S; (b) f S + gS = S; (c) the map k[¯s]/(N(f )) → S/(f ) is an isomorphism. Proof. Straightforward. 8 Lemma 4.3. Let k be an infinite field, and let R be a domain which is a semi-local essentially smooth k-algebra with maximal ideals pi, 1 ≤ i ≤ m. Let A ⊇ R[t] be another domain, smooth as an R-algebra and finite over R[t]. Assume that for each i the R/pi- algebra Ai = A/piA is equidimensional of dimension one. Let ǫ : A → R be an R- augmentation and I = Ker(ǫ). Given an f ∈ A with 0 6= ǫ(f ) ∈ m\ i=1 pi ⊂ R and such that the R-module A/f A is finite, one can find an element u ∈ A satisfying the following conditions: (1) A is a finite projective module over R[u]; (2) A/uA = A/I × A/J for some ideal J; (3) J + f A = A; (4) (u − 1)A + f A = A; (5) set N(f ) = NA/R[u](f ), then N(f ) = f g ∈ A for some g ∈ A; (6) f A + gA = A; (7) the composition map ϕ : R[u]/(NA/R[u](f )) → A/(NA/R[u](f )) → A/(f ) is an iso- morphism. Proof. Replacing t by t − ǫ(t) we may assume that ǫ(t) = 0. Since A is finite over R[t], it follows from a theorem of Grothendieck [E, Cor. 17.18] that it is a finite projective R[t]-module. Since A is finite over R[t] and A/f A is finite over R we conclude that R[t]/(NA/R[t](f )) is finite over R, and hence R/(tNA/R[t](f )) is finite over R. Setting v = tNA/R[t](f ), we get an integral extension R[t] over R[v]. Thus A is finite over R[v]. By the theorem of Grothendieck [E, Cor. 17.18] A is a finite projective R[v]- module. Applying Lemma 3.3 to A over R[t] (not over R[v]) one gets an equality NA/R[t](f ) = f · gf,t ∈ A for an element gf,t ∈ A. Thus v = t · NA/R[t](f ) = t · f · gf,t ∈ f A, and ǫ(v) = ǫ(t) · ǫ(NA/R[t](f )) = 0 Below, we use the bar ¯ to denote reduction modulo an ideal, and the subscript i to indicate that reduction is modulo piA, 1 ≤ i ≤ m. Let li = ¯Ri = R/pi. By the assumption of the lemma, the li-algebra ¯Ai is li-smooth equidimensional of dimension 1. The element ¯fi ∈ ¯Ai is a non-zero divisor since ¯Ai/ ¯fi ¯Ai = (A/f A)i is a finite li-module. (i) 0 = Ker(¯ǫi). Let m Let ¯si ∈ ¯Ai be such that the extension ¯Ai/li[¯si] satisfies conditions (1) to (4) of Lemma 4.1. ni be distinct maximal ideals of ¯Ai dividing ¯fi and let m (i) (i) 1 , m (i) 2 , . . . , m 9 Let s ∈ A be a common lifting of ¯si's, in other words, s = ¯si in ¯Ai for all i = 1, . . . , m. Replacing s by s − ǫ(s) we may assume that ǫ(s) = 0 and, as above, s = ¯si for all i = 1, . . . , m. Let sn + p1(v)sn−1 + · · · + pn(v) = 0 be an integral dependence relation for s. Let N be an integer larger than max{2, deg(pj(t))}, where j = 1, 2, . . . , n. Then for any r ∈ k× the element u = s − rvN has the following property: v is integral over R[u]. Thus, for any r ∈ k× the ring A is integral over R[u]. (i) On the other hand, one has ¯vi ∈ m j v ∈ f A and ǫ(v) = 0. It implies that each element ¯ui = ¯si − r ¯vi (1) to (4) of Lemma 4.1. for all 1 ≤ i ≤ m and all 0 ≤ j ≤ ni, since N still satisfies conditions We claim that the element u ∈ R has all the properties listed in the statement of the present lemma, for almost all r ∈ k×. Indeed, for almost all r ∈ k× the element u satisfies Conditions (1) to (4) of Lemma 4.3. It remains to show that Conditions (5) to (7) hold for all r ∈ k×. Since A is finite over R[u], the same theorem of Grothendieck [E, Cor. 17.18] implies that it is a finite projective R[u]-module. To prove (5), consider the characteristic polyno- f mial of the operator A −→ A as an R[u]-module operator. This polynomial vanishes on f and its free term equals ±NA/R[u](f ), the norm of f . Thus, f n−a1f n−1+· · ·±NA/R[u](f ) = 0 and NA/R[u](f ) = f · gf,u for some gf,u ∈ A. To prove (6), one has to verify that the above g is a unit modulo the ideal f A. It suffices to check that for each index i the element ¯gi ∈ ¯Ai is a unit modulo the ideal ¯fi ¯Ai. To that end observe that the field li = R/pi, the li-algebra Si = ¯Ai, its maximal (i) ideals m ni and the element ¯ui satisfy the hypotheses of Lemma 4.2, with u replaced by ¯ui. Now, by Item (b) of Lemma 4.2 the reduction ¯gi is a unit modulo the ideal ¯fi ¯Ri. (i) 1 , . . . , m (i) 0 , m To prove (7), observe that R[u]/(NA/R[u](f )) and A/f A are finite R-modules. Thus, it remains to check that the map ϕ : R[u]/(NA/R[u](f )) → A/f A is an isomorphism modulo each maximal ideal pi. To that end it suffices to verify that the map ¯ϕi : li[¯ui]/(N( ¯fi)) → ¯Ai/ ¯fi ¯Ai is an isomorphism for each index i, where N( ¯fi) := N ¯Ai/li[¯u]( ¯fi). Now, by Item (c) of Lemma 4.2 the map ¯ϕi is an isomorphism. This finishes the proof. Proof of Theorem 3.4. Let U = Spec(OX,{x1,x2,...,xr}) be as in Definition 3.1. Write R for OX,{x1,x2,...,xr}. It is a domain which is a semi-local essentially smooth k-algebra with maximal ideals pi, 1 ≤ i ≤ r. Let (X, f, ∆) be a nice triple over U. We show that it gives rise to certain data subject to the hypotheses of Lemma 4.3. Let A = Γ(X, OX). It is a domain, since X is irreducible. It is an R-algebra via the ring homomorphism q∗ U : R → Γ(X, OX). Furthermore, it is smooth as an R-algebra. The triple (X, f, ∆) is a nice triple. Thus, there exists a finite surjective U-morphism Π : X → A1 U . It induces an R-algebra inclusion R[t] ֒→ Γ(X, OX) = A such that A is finitely generated as an R[t]-module. Also, for all i = 1, . . . , r, the R/pi-algebra A/piA is equidimensional of dimension one. Let ǫ = ∆∗ : A → R 10 be an R-algebra homomorphism induced by the section ∆ of the morphism qU . Clearly, this ǫ is an augmentation; set I = Ker(ǫ). Further, since (X, f, ∆) is a nice triple, ǫ(f ) 6= 0 ∈ R and A/f A is finite as an R-module. Finally, f vanishes at every closed point of ∆(U) by the assumption of the Theorem. Summarising the above, we conclude that we are in the setting of Lemma 4.3, and may use the conclusion of that Lemma. Thus, there exists an element u ∈ A subject to Conditions (1) through (7) of Lemma 4.3. This u induces an R-algebra inclusion R[u] ֒→ A such that A is finite as an R[u]- module. Let σ : X → A1 × U be the U-scheme morphism induced by the above inclusion R[u] ֒→ A. Clearly, σ is finite and surjective. In the rest of the proof we write t instead of u, and consider A as an R[t]-module via σ. Let N(f ) := NA/R[t](f ) ∈ R[t] ⊆ A and gf,σ ∈ A be the elements defined just above Lemma 3.3. We claim that this morphism σ and the chosen elements N(f ) and gf,σ satisfy con- clusions (1) to (4) of Theorem 3.4. Let us verify this claim. Since A is finite as an R[t]-module and both rings R[t] and A are regular, the R[t]-module A is finitely gener- ated and projective, see [E, Corollary 18.17]. Thus, σ is ´etale at a point x ∈ X if and only if the k(σ(x))-algebra k(σ(x)) ⊗R[t] A is ´etale. If the point x belongs to the closed subscheme Spec(A/piA) for some maximal ideal pi of R, then k(σ(x)) ⊗R[t] A = k(σ(x)) ⊗(R/pi)[t] A/piA. We can conclude that σ is ´etale at a specific point x if and only if the (R/pi)[t]-algebra A/piA is ´etale at the point x. It follows from the proof of Lemma 4.3 that the morphism σ induces a morphism Spec(A/piA) σi−→ A1 li on the closed fibre Spec(A/piA) for each i. This induced morphism is ´etale along the vanishing locus of the function ¯fi and along each point ∆i(Spec li). Indeed, for the vanishing locus of the function ¯fi this follows from Items (6) and (7) of Lemma 4.3. It follows from the hypotheses of Lemma 4.3 that the function f vanishes at each maximal ideal containing I. Thus σ is ´etale along the closed subscheme X defined by the ideal I, that is along ∆(U). This settles Item (1) of Theorem 3.4. Consider Item (2). Write g for gf,σ. The first of the following equalities σ−1(σ({f = 0})) = {N(f ) = 0} = {f = 0} ⊔ {g = 0} is a commonplace. The second one follows from the equality N(f ) = ±f · g, proved in Lemma 3.3 and Item (6) of Lemma 4.3. Clearly, the square (4) is Cartesian and the morphism σ0 contains a closed subscheme ∆(U), and hence is non-empty. shows that the morphism of the reduced closed subschemes g is ´etale. The scheme X0 g Item (7) of Lemma 4.3 σ0 g{f =0}red : {f = 0}red → {N(f ) = 0}red is an isomorphism. Thus, we have checked Item (3) of Theorem 3.4. It remains only to check Item (4). We already know that {f = 0} ⊂ X0 g. Both schemes ∆(U) and {f = 0} are semi-local and the set of closed points of ∆(U) is contained in the 11 set of closed points of the closed set {f = 0} by the assuptions of the theorem. Thus, ∆(U) ⊂ X0 g. This concludes the proof of Item (4) of Theorem 3.4 and thus of the theorem itself. 5 Proof of Theorem 3.6 The aim of this Section is to proof of Theorem 3.6. We begin with the following Propo- sition which is a straightforward analogue of [OP2, Prop. 7.1] Proposition 5.1. Let S be a regular semi-local irreducible scheme and let G1, G2 be two semi-simple simply-connected S-group schemes which are twisted forms of each other. Further, let T ⊂ S be a closed sub-scheme of S and ϕ : G1T → G2T be an S-group scheme isomorphism. Then there exists a finite ´etale morphism eS π−→ S together with its section δ : T → eS over T and an eS-group scheme isomorphism Φ : π∗G1 → π∗G2 such that δ∗(Φ) = ϕ. Since the proof of the Proposition 5.1 is rather long we first give an outline. Clearly, G1 and G2 are of the same type. By [SGA3, Exp. XXIV, Cor. 1.8] there exists an S-scheme IsomS(G1, G2) representing the functor that sends an S-scheme W to the set of all W -group scheme isomorphisms from W ×S G1 to W ×S G2. The isomorphism ϕ from the hypothesis of Proposition 5.1 determines a section δ : T → IsomS(G1, G2) of the structure map IsomS(G1, G2) → S. By Lemmas 5.4 and 5.2 below there exists a closed subscheme eS of IsomS(G1, G2) which is finite ´etale over S and contains δ(T ). So, we have a commutative diagram of S-schemes T δ / eS %▲▲▲▲▲▲▲▲▲▲▲▲▲ i IsomS(G1, G2) (5) π v❧❧❧❧❧❧❧❧❧❧❧❧❧❧❧❧ S . such that the horizontal arrows are closed embeddings. Thus we get an isomorphism Φ : π∗(G1) → π∗(G2) such that δ∗(Φ) = ϕ. The precise proof of the Proposition requires some auxiliary results and will be given right below Lemma 5.4. Clearly, G1 and G2 are of the same type. Let G0 be a split semi- simple simply connected algebraic group over the ground field k such that G1 and G2 are twisted forms of the S-group scheme S ×Spec(k) G0. Let Autk(G0) be the automorphism scheme of the algebraic k-group G0. It is known that Autk(G0) is a semi-direct product of the algebraic k-group Gad is a group adjoint to G0. Also, Autk(G0) is a smooth affine algebraic k-group (for example, by [SGA3, Exp. XXIV, Cor. 1.8]). Set for short Aut := Autk(G0) and AutS for the S-group scheme S ×Spec(k) Aut. 0 and a finite group, where Gad 0 Consider an S-scheme IsomS(G0,S, G2) constructed in [SGA3, Exp. XXIV, Cor.1.8] and representing a functor that sends an S-scheme W to the set of all W -group scheme isomorphisms ϕ2 : W ×S G0,S → W ×S G2. Similarly, consider an S-scheme AutS(G2) 12 % / / /   v constructed in [SGA3, Exp. XXIV, Cor. 1.8] and representing a functor that sends an S-scheme W to the set of all W -group scheme automorphisms α : W ×S G2 → W ×S G2. The functor transformation (ϕ2, α2) 7→ ϕ2 ◦ α−1 2 defines an S-scheme morphism IsomS(G0,S, G2) ×S AutS → IsomS(G0,S, G2) which makes the S-scheme IsomS(G0,S, G2) a principal right AutS-bundle. The functor transformation (β2, ϕ2) 7→ β2 ◦ ϕ2 defines an S-scheme morphism AutS(G2) ×S IsomS(G0,S, G2) → IsomS(G0,S, G2) which makes the S-scheme IsomS(G0,S, G2) a principal left AutS(G2)-bundle. Analogously, the functor transformation (α1, ϕ1) 7→ α1 ◦ ϕ1 makes the S-scheme IsomS(G1, G0,S) a principal left AutS-bundle and the functor transformation (ϕ1, β1) 7→ ϕ1 ◦ β1 makes the S-scheme IsomS(G1, G0,S) a principal right AutS(G1)-bundle. Let 2Pr be a left principal AutS(G2)-bundle and at the same time a right principal AutS-bundle such that the two actions commute. Let lP1 be a left principal AutS-bundle and at the same time a right principal AutS(G1)-bundle such that the two actions com- mute. Let Y be a k-variety equipped with a left and a right Autk-actions which commute. Then the k-scheme is equipped with a left Autk × Autk-action given by (2Pr) ×S (YS) ×S (lP1) (α2, α1)(p2, y, p1) = (p2α−1 2 , α2yα−1 1 , α1p1). The orbit space does exist (it can be constructed by descent). Denote it by 2Y1. We now show that it is an S-scheme. Indeed, the structure morphism Y → Spec(k) defines a morphism (2Pr) ×S (YS) ×S (lP1) → (2Pr) ×S (lP1) respecting the Aut × Aut-actions on both sides. Thus it defines a morphism of the orbit spaces 2Y1 → (2Spec(k)1) = S. The latter equality holds since (2Pr) ×S (lP1) is a principal left Aut × Aut-bundle with respect the left action given by (α2, α1)(p2, p1) = (p2α−1 2 , α1p1). The construction Y 7→ 2Y1 has several nice properties. Namely, (i) it is natural with respect to k-morphisms of k-varieties Y → Y ′ commuting with the given two-sided Aut × Aut-actions on Y and Y ′, (ii) it takes closed embeddings to closed embeddings, (iii) it takes open embeddings to open embeddings, (iv) it takes k-products to S-products, 13 (v) locally in the ´etale topology on S, the S-schemes YS and 2Y1 are isomorphic. Set 2Pr = IsomS(G0,S, G2) and lP1 = IsomS(G1, G0,S). The functor transformation (ϕ2, α, ϕ1) 7→ ϕ2 ◦ α ◦ ϕ1 gives a morphism of representable S-functors IsomS(G0,S, G2) ×S (AutS) ×S IsomS(G1, G0,S) Φ −→ IsomS(G1, G2). The equality ϕ2 ◦ α ◦ ϕ1 = (ϕ2 ◦ α−1 2 ) ◦ (α2 ◦ α ◦ α−1 1 ) ◦ (α1 ◦ ϕ1) shows that the morphism Φ induces a morphism ¯Φ : 2(Aut)1 → IsomS(G1, G2). Lemma 5.2. The S-morphism ¯Φ : 2(Aut)1 → IsomS(G1, G2) is an isomorphism. Proof. It suffices to prove that ¯Φ is an isomorphism locally in the ´etale topology on S. The latter follows from the property (v). Now let G0 and Aut be as above. There is a closed embedding of algebraic groups ρ : Aut ֒→ GLV,k for an n-dimensional k-vector space V . Replacing ρ with ρ ⊕ det−1 ◦ ρ we get a closed embedding of algebraic k-groups ρ1 : Aut ֒→ SLW,k, where W = V ⊕ k. ρ1−→ SLW,k ֒→ End is a closed Let End := Endk(W ). Clearly, the composition in : Aut embedding. We will identify Aut with its image in End. Let Aut be the closure of Aut in the projective space P(k ⊕ End). Set Aut∞ := Aut − Aut regarded as a reduced scheme. So, we get a commutative diagram of k-varieties Aut in End j J Aut in / P(k ⊕ End) i I Aut∞ in∞ P(End) (6) where the left square is Cartesian. All varieties are equipped with the left Aut × Aut- action induced by Aut × Aut-action on the affine space k ⊕ End given by (g1, g2)(α, c) = (c, g1αg−1 2 ). All the arrows in this diagram respect this action. Applying to this diagram the above construction Y 7→ 2Y1, we obtain a commutative diagram of S-schemes 2Aut1 in 2End1 j J 2(Aut)1 in / P(OS ⊕ 2End1) i I 2(Aut∞)1 in∞ P(2End1) (7) where the square on the left is Cartesian. From now on we assume that S is a semi-local irreducible scheme. Then the vector bundle 2End1 is trivial. Since it is trivial, we may choose homogeneous coordinates Yi's 14 / /     o o   / o o / /     o o   / o o on P(OS ⊕ 2End1) such that the closed subschemes {Y0 = 0} and P(2End1) of the scheme P(OS ⊕ 2End1) coincide and the S-scheme P(OS ⊕ 2End1) itself is isomorphic to the projective space Pn2 S . Thus the diagram (7) of S-schemes and of S-scheme morphisms can be rewritten as follows 2Aut1 in {Y0 6= 0} j J 2(Aut)1 in / Pn2 S i I 2(Aut∞)1 in∞ {Y0 = 0} (8) where the square on the left is Cartesian. Since 2(Aut∞)1 = 2(Aut)1 − 2Aut1, the set-theoretic intersection 2(Aut)1 ∩ {Y0 = 0} in Pn2 S coincides with 2(Aut∞)1. The following Lemma is the lemma [OP2, Lemma 7.2]. Lemma 5.3. Let S = Spec(R) be a regular semi-local scheme and T a closed subscheme of S. Let ¯X be a closed subscheme of PN S , where S is the affine space defined by Y0 6= 0. Let X∞ = ¯X \ X be the intersection of ¯X with AN the hyperplane at infinity Y0 = 0. Assume further that S = P roj(S[Y0, . . . , YN ]) and X = ¯X ∩ AN (1) X is smooth and equidimensional over S, of relative dimension r. (2) For every closed point s ∈ S the closed fibres of X∞ and X satisfy dim(X∞(s)) < dim(X(s)) = r . (3) Over T there exists a section δ : T → X of the canonical projection X → S. δ(T ). Then there exists a closed subscheme eS of X which is finite ´etale over S and contains The diagram (8) shows that the S-schemes X = 2Aut1, ¯X = 2(Aut)1 and X∞ = 2(Aut∞)1 satisfy all the hypotheses of Lemma 5.3 except possibly the conditions (2) and (3). To check (2), observe that the diagram of S-schemes 2Aut1 j / 2(Aut)1 i 2(Aut∞)1 locally in the ´etale topology on S is isomorphic to the diagram of S-schemes Aut × S j / (Aut) × S i (Aut∞) × S. (9) (10) This follows from the property (v) of the construction Z → 2Z1. Since Aut is equidimen- sional and Aut is the closure of Aut in P(End ⊕ k), one has dim(Aut∞) < dim(Aut) = dim Aut. Thus the assumption (2) of Lemma 5.3 is fulfilled. Whence we have proved the following 15 / /     o o   / o o / o o / o o Lemma 5.4. Assume S is a regular semi-local irreducible scheme and assume we are given with a closed subscheme T ⊂ S equipped with a section δ : T → 2Aut1 of the structure map 2Aut1 → S. Then there exists a closed subscheme eS of 2Aut1 which is finite and ´etale over S and contains δ(T ). Proof of Proposition 5.1. By Lemma 5.2 the S-schemes IsomS(G1, G2) and 2Aut1 are nat- urally isomorphic as S-schemes. The isomorphism ϕ from the hypotheses of the Propo- sition 5.1 determines a section δ : T → IsomS(G1, G2) = 2Aut1 of the structure map IsomS(G1, G2) = 2Aut1 → S. By Lemma 5.4 there exists a closed subscheme eS of 2Aut1 = IsomS(G1, G2) which is finite ´etale over S and contains δ(T ). So, we have morphisms (even closed inclusions) of S-schemes T δ / eS %▲▲▲▲▲▲▲▲▲▲▲▲▲ i IsomS(G1, G2) (11) π v❧❧❧❧❧❧❧❧❧❧❧❧❧❧❧❧ S . Thus we get an isomorphism Φ : π∗(G1) → π∗(G2) such that δ∗(Φ) = ϕ. Proof of Theorem 3.6. We can start by almost literally repeating arguments from the proof of [OP2, Lemma 8.1], which involve the following purely geometric lemma [OP2, Lemma 8.2]. For reader's convenience below we state that Lemma adapting notation to the ones of Section 3. Namely, let U be as in Definition 3.1 and let (X, f, ∆) be a nice triple over U. Further, let GX be a simple simply-connected X-group scheme, GU := ∆∗(GX), and let Gconst be the pull-back of GU to X. Finally, by the definition of a nice triple there exists a finite surjective morphism Π : X → A1 × U of U-schemes. Lemma 5.5. Let Y be a closed nonempty sub-scheme of X, finite over U. Let V be an open subset of X containing Π−1(Π(Y)). There exists an open set W ⊆ V still containing q−1 U (qU (Y)) and endowed with a finite surjective morphism Π∗ : W → A1 × U (in general 6= Π). Let Π : X → A1 ×U be the above finite surjective U-morphism. The following diagram summarises the situation: Z X − Z  / X Π / / A1 × U ∆ qU U Here Z is the closed sub-scheme defined by the equation f = 0. By assumption, Z is finite over U. Let Y = Π−1(Π(Z ∪ ∆(U))). Since Z and ∆(U) are both finite over U and since Π 16 % / / /   v    /   O O is a finite morphism of U-schemes, Y is also finite over U. Denote by y1, . . . , ym its closed points and let S = Spec(OX,y1,...,ym). Set T = ∆(U) ⊆ S. Further, let GU = ∆∗(GX) be as in the hypotheses of Theorem 3.6 and let Gconst be the pull-back of GU to X. Finally, let ϕ : GconstT → GXT be the canonical isomorphism. Recall that by assumption X is U-smooth, and thus S is regular. By Proposition 5.1 there exists a finite ´etale covering θ0 : eS → S, a section δ : T → eS of θ0 over T and an isomorphism Φ0 : θ∗ 0(Gconst,S) → θ∗ 0(GXS) such that δ∗Φ0 = ϕ. Replacing eS with a connected component of eS which contains δ(T ) = δ(∆(U)) we may and will assume that eS is irreducible. We can extend these data to a neighborhood V of {y1, . . . , yn} and get the diagram δ @✁✁✁ ✁✁✁✁✁ eS θ0 S  T  / eV θ / V  / X (12) where π : eV → V finite ´etale, and an isomorphism Φ : θ∗(Gconst) → θ∗(GX). Since T isomorphically projects onto U, it is still closed viewed as a sub-scheme of V. Note that since Y is semi-local and V contains all of its closed points, V contains Π−1(Π(Y)) = Y. By Lemma 5.5 there exists an open subset W ⊆ V containing Y and endowed with a finite surjective U-morphism Π∗ : W → A1 × U. Let X′ = θ−1(W), f ′ = θ∗(f ), q′ U = qU ◦ θ, and let ∆′ : U → X′ be the section of q′ U obtained as the composition of δ with ∆. We claim that the triple (X′, f ′, ∆′) is a nice triple. Let us verify this. Firstly, the structure morphism q′ U : X′ → U coincides with the composition X′ θ−→ W ֒→ X qU−→ U. Thus, it is smooth. The element f ′ belongs to the ring Γ(X′, OX′), the morphism ∆′ is a section of q′ U . Each component of each fibre of the morphism qU has dimension θ−→ W ֒→ X is ´etale. Thus, each component of each fibre of the one, the morphism X′ U is also of dimension one. Since {f = 0} ⊂ W and θ : X′ → W is finite, morphism q′ {f ′ = 0} is finite over {f = 0} and hence also over U. In other words, the O-module Γ(X′, OX′)/f ′ · Γ(X′, OX′) is finite. The morphism θ : X′ → W is finite and surjective. We have constructed above in Lemma 5.5 the finite surjective morphism Π∗ : W → A1 × U. It follows that Π∗ ◦ θ : X′ → A1 × U is finite and surjective. Clearly, the ´etale morphism θ : X′ → X is a morphism of nice triples, with g = 1. Denote the restriction of Φ to X′ simply by Φ. The equality (∆′)∗Φ = idGU holds by the very construction of the isomorphism Φ. Theorem follows. 17     /    / / @  /  / 6 A basic nice triple With Propositions 2.3 and 2.4 at our disposal we may form a basic nice triple, namely the triple (16) below. This is the main aim of the present section. Namely, fix a smooth geometrically irreducible affine k-scheme X, a finite family of points x1, x2, . . . , xn on X, and a non-zero function f ∈ k[X]. We always assume that the set {x1, x2, . . . , xn} is contained in the vanishing locus of the function f. By Proposition 2.3 there exist a Zariski open neighborhood X 0 of the family {x1, x2, . . . , xn} and an almost elementary fibration p : X 0 → S, where S is an open subscheme of the projective space PdimX−1, such that p{f=0}∩X 0 : {f = 0} ∩ X 0 → S is finite surjective. Let si = p(xi) ∈ S, for each 1 ≤ i ≤ n. Shrinking S, we may assume that S is affine and still contains the family {s1, s2, . . . , sn}. Clearly, in this case p−1(S) ⊆ X 0 contains the family {x1, x2, . . . , xn}. We replace X by p−1(S) and f by its restriction to this new X. In this way we get an almost elementary fibration p : X → S such that {x1, . . . , xn} ⊂ {f = 0} ⊂ X, S is an open affine subscheme in the projective space PdimX−1, and the restriction p{f=0} : {f = 0} → S of p to the vanishing locus of f is a finite surjective morphism. In other words, k[X]/(f) is finite as a k[S]-module. As an open affine subscheme of the projective space PdimX−1 the scheme S is regular. By Proposition 2.4 one can shrink S in such a way that S is still affine, contains the family {s1, s2, . . . , sn} and there exists a finite surjective morphism π : X → A1 × S such that p = prS ◦π. Clearly, in this case p−1(S) ⊆ X contains the family {x1, x2, . . . , xn}. We replace X by p−1(S) and f by its restriction to this new X. In this way we get an almost elementary fibration p : X → S such that {x1, . . . , xn} ⊂ {f = 0} ⊂ X, S is an open affine subscheme in the projective space PdimX−1, and the restriction p{f=0} : {f = 0} → S is a finite surjective morphism. Eventually we conclude that there exists a finite surjective morphism π : X → A1 × S such that p = prS ◦ π. Now, set U := Spec(OX,{x1,x2,...,xn}), denote by can : U ֒→ X the canonical inclusion of schemes, and let pU = p ◦ can : U → S. Further, we consider the fibre product X := U ×S X. 18 Then the canonical projections qU : X → U and qX : X → X and the diagonal morphism ∆ : U → X can be included in the following diagram qX / X 8qqqqqqqqqqqqq can ∆ X qU U where and qX ◦ ∆ = can qU ◦ ∆ = idU . (13) (14) (15) Note that qU is a smooth morphism with geometrically irreducible fibres of dimension one. Indeed, observe that qU is a base change via pU of the morphism p which has the desired properties. Note that X is irreducible. Indeed, U is irreducible and the fibre of qU over the generic point of U is irreducible. Taking the base change via pU of the finite surjective morphism π : X → A1 × S, we get a finite surjective morphism Π : X → A1 × U such that qU = prU ◦ Π, where prU : A1 × U → U is the natural projection. Set f := q∗ X (f). The OX,{x1,x2,...,xn}-module Γ(X, OX)/f · Γ(X, OX) is finite, since the k[S]-module k[X]/f · k[X] is finite. Now the data (qU : X → U, f, ∆) (16) form an example of a nice triple as in Definition 3.1. Moreover, we have Claim 6.1. The schemes ∆(U) and {f = 0} are both semi-local and the set of closed points of ∆(U) is contained in the set of closed points of {f = 0}. This holds since the set {x1, x2, . . . , xn} is contained in the vanishing locus of the function f. 7 Main construction The main result of this Section is Corollary 7.2. Fix a k-smooth irreducible affine k-scheme X, a finite family of points x1, x2, . . . , xn on X, and set O := OX,{x1,x2,...,xn} and U := Spec(O). Let A be the Noetherian k-algebra from Theorem 1.1 and T = Spec(A). Further, consider a simple simply connected U- group scheme G and a principal G-bundle P over O ⊗k A which is trivial over K ⊗k A for the field of fractions K of O. We may and will assume that for certain f ∈ O the principal G-bundle P is trivial over Of ⊗k A. 19   / 8 Z Z Shrinking X if necessary, we may secure the following properties (i) The points x1, x2, . . . , xn are still in X and X is affine. (ii) The group scheme G is defined over X and it is a simple group scheme. We will often denote this X-group scheme by GX and write GU for the original G. (iii) The principal GU -bundle P is the restriction to U ×Spec(k) T of a principal GX- bundle PX over X ×Spec(k) T and f ∈ k[X]. We often will write PU for the original principal GU -bundle P over U ×Spec(k) T . (iv) The restriction Pf of the bundle PX to the principal open subset Xf ×Spec(k) T is trivial and f vanishes at each xi's. After substituting k by its algebraic closureek in k[X], and T by eT = Spec(ek)×Spec(k)T , we can assume that X is a ek-smooth geometrically irreducible affine ek-scheme. Note that U ×Spec(ek) eT ∼= U ×Spec(k) T as U-schemes, and the same holds for X instead of U. To simplify the notation, we will continue to denote this new ek by k and eT by T . In particular, we are given now the smooth geometrically irreducible affine k-scheme X, the finite family of points x1, x2, . . . , xn on X, and the non-zero function f ∈ k[X] vanishing at each point xi. Recall that starting from these data we constructed at the very end of Section 6 the nice triple (16) of the form (qU : X → U, f, ∆) with X = U ×S X. We did that shrinking X and securing properties (i) to (iv) at the same time. Recall that qX : X = U ×S X → X is the projection to X. Set GX := (qX )∗(GX ) and Gconst := (qU )∗(GU ). By Theorem 3.6 there exists a morphism of nice triples θ : (q′ U : X′ → U, f ′, ∆′) → (qU : X → U, f, ∆) and an isomorphism Φ : θ∗(Gconst) → θ∗(GX) =: GX′ of X′-group schemes such that (∆′)∗(Φ) = idGU . Set Recall that q′ X = qX ◦ θ : X′ → X. U = qU ◦ θ : X′ → U, q′ since θ is a morphism of nice triples. (17) (18) (19) Note that, since by Claim 6.1 f vanishes on all closed points of ∆(U), and θ is a morphism of nice triples, f ′ vanishes on all closed points of ∆′(U) as well. Therefore, the nice triple (q′ U : X′ → U, f ′, ∆′ : U → X′) is subject to Theorem 3.4. By Theorem 3.4 there exists a finite surjective morphism σ : X′ → A1×U of U-schemes satisfying (1) to (3) from that Theorem. In particular, one has σ−1(cid:16)σ(cid:0){f ′ = 0}(cid:1)(cid:17) = N(f ′) = {f ′ = 0} ⊔ {gf ′,σ = 0} 20 with N(f ′) and gf ′,σ defined in the item (2) of Theorem 3.4. Thus, replacing for brevity gf ′,σ by g′, one gets the following elementary distinguished square in the category of U-smooth schemes (see [Def. 2.1, Vo]): (X′)0 N (f ′) = (X′)0 σ0 f ′ g′ f ′g′ (A1 × U)N (f ′) inc inc (X′)0 g′ σ0 g′ / A1 × U (20) The base change of this square by means of the morphism U ×Spec(k) T → U is an elementary distinguished square in the category of smooth U ×Spec(k) T -schemes. Thus this new square can be used to build up principal GU -bundles over (A1 × U) ×Spec(k) T beginning with certain data over the three other corners. This is what we are going to do below in this Section. Set X = q′ Q′ Q′ U = q′ X × idT : X′ ×Spec(k) T → X ×Spec(k) T, U × idT : X′ ×Spec(k) T → U ×Spec(k) T. X)∗(PX) as a principal (q′ Consider (Q′ Recall that PX is trivial as a principal GX-bundle over Xf×Spec(k)T . Therefore, (Q′ is trivial as a principal ρ∗ X′ U )∗(GU ) = θ∗(Gconst)-bundle via the isomorphism Φ. X)∗(PX) f ′ ×Spec(k) T . So, (Q′ X )∗(PX ) is trivial over S(Gconst)-bundle via the isomorphism ΨS. X)∗(PX) over X′ ×Spec(k) T f ′g′ ×Spec(k) T . Now, taking the S(GX)-bundle over X′ f ′ ×Spec(k) T , when regarded as a principal ρ∗ Thus, regarded as a principal GU -bundle, the bundle (Q′ becomes trivial over X′ f ′ ×Spec(k) T , and a fortiori over (X′)0 trivial GU -bundle over (A1 × U)N (f ′) and an isomorphism ψ : GU ×U [(X′)0 N (f ′) ×Spec(k) T ] → (Q′ X)∗(PX)[(X′)0 N(f ′) ×Spec(k)T ] (21) of principal GU -bundles, we get a principal GU -bundle Gt over (A1 × U) ×Spec(k) T such that (1) Gt[(A1×U )N(f ′)×Spec(k)T ] = GU ×U [(A1 × U)N (f ′) ×Spec(k) T ] (2) there is an isomorphism ϕ : [(σ0 principal GU -bundles, where (Q′ the X′-group scheme isomorphism Φ from (17); g′) × idT ]∗(Gt) → (Q′ g′ ×Spec(k)T ] of the X )∗(PX) is regarded as a principal GU -bundle via X)∗(PX)[(X′)0 (3) (inc × idT )∗(ϕ) = ψ. Finally, form the following diagram (A1 × U) ×Spec(k) T (X′)0 g′ ×Spec(k) T / X ×Spec(k) T (22) σ0 g′ ×id *❱❱❱❱❱❱❱❱❱❱❱❱❱❱❱❱❱❱❱ prU ×id q′ U ×id ∆′×id U ×Spec(k) T Q′ X =q′ X ×id 4❥❥❥❥❥❥❥❥❥❥❥❥❥❥❥❥❥ can×id 21 / /     / *   o o   / 4 V V This diagram is well-defined, since by Item (4) of Theorem 3.4 the image of the morphism ∆′ lands in (X′)0 g′. Theorem 7.1. The principal GU -bundle Gt over (A1×U)×Spec(k)T , the monic polynomial N(f ′) ∈ O[t], the diagram (22), and the isomorphism Φ from (17) constructed above, satisfy the following conditions (1*) -- (6*). U = prU ◦ σ0 g′, g′ is ´etale, U ◦ ∆′ = idU , X ◦ ∆′ = can, (1*) q′ (2*) σ0 (3*) q′ (4*) q′ (5*) the restriction of Gt to (A1 × U)N (f ′) ×Spec(k) T is a trivial GU -bundle, (6*) (σ0 X )∗(PX) are isomorphic as GU -bundles over (X′)0 g′ ×id)∗(Gt) and (Q′ g′ ×Spec(k)T . X)∗(PX) is regarded as a principal GU -bundle via the group scheme isomorphism Here (Q′ Φ from (17). Proof. By the very choice of σ it is an U-scheme morphism, which proves (1*). By the choice of (X′)0 ֒→ X′ in Theorem 3.4, the morphism σ is ´etale on this subscheme, hence one gets (2*). Property (3*) holds for ∆′ since (q′ X : X′ → U, f ′, ∆′) is a nice triple and, in particular, ∆′ is a section of q′ U . Property (4*) can be established as follows: q′ X ◦ ∆′ = (qX ◦ ρS) ◦ ∆′ = qX ◦ ∆ = can. The first equality here holds by the definition of q′ X , the second one holds since ρS is a morphism of nice triples; the third one follows from equality (14). Property (5*) is just Property (1) in the above construction of Gt. Property (6*) is precisely Property (2) in the construction of Gt. The composition s′ := σ0 g′ ◦ ∆′ : U → A1 × U is a section of the projection prU by the properties (1*) and (3*). Recall that GU over U is the original group scheme G introduced in the very beginning of this Section. Since U is semi-local, we may assume that s′ is the zero section of the projection A1 U → U. Further- U → U, we may assume that N(f ′)(1) ∈ O is more, making an affine transformation of A1 invertible. Corollary 7.2 (=Theorem 1.2). The principal GU -bundle Gt over A1 monic polynomial N(f ′) ∈ O[t] are subject to the following conditions [U ×Spec(k)T ] and the (i) the restriction of Gt to [(A1 × U)N (f ′) ×Spec(k) T ] is a trivial GU -bundle, (ii) the restriction of Gt to {0} × U ×Spec(k) T is the original GU -bundle PU . (iii) N(f ′)(1) ∈ O is invertible. 22 Proof. The property (i) is just the property (5*) above. Now by (6*) the GU -bundles Gt{0}×U ×Spec(k)T = (s′ × id)∗(Gt) = (∆′ × id)∗((σ0 g′ × id)∗(Gt)) and (∆′ × id)∗(Q′ X)∗(PX ) = (can × id)∗(PX) are isomorphic, since ∆′∗(Φ) = idGU . It remains to recall that the principal GU -bundle (can × id)∗(PX) is the original GU -bundle PU by the choice of PX. Whence the Corollary. 8 Group of points of an isotropic simple group In this section we establish several results concerning groups of points of simple groups, in particular, Lemma 8.2, Proposition 8.7 and Lemma 8.8, which play crucial role in the rest of the paper. Definition 8.1. Let G be a reductive group scheme over a commutative ring A. Assume that G has a proper parabolic subgroup P = P + over A, and denote by U + its unipotent radical. By [SGA3, Exp. XXVI Cor. 2.3, Th. 4.3.2] there exists a parabolic subgroup P − of G opposite to P +, and by [SGA3, Exp. XXVI Cor. 1.8] any two such subgroups are conjugate by an element of U +(A). Let U − be the unipotent radical of P −. For any commutative A-algebra B we define the P -elementary subgroup EP (B) of the group G(B) as follows: EP (B) = hU +(B), U −(B)i. Lemma 8.2. Let B → ¯B be a surjective A-algebra homomorphism. Then the induced homomorphism of elementary groups EP (B) → EP ( ¯B) is also surjective. Proof. By [SGA3, Exp.XXVI Cor. 2.5] the A-schemes U + and U − are isomorphic to A-vector bundles of finite rank. Thus, the maps U ±(B) → U ±( ¯B) are surjective. Let l be a field and Gl be an isotropic simple simply connected l-group scheme. Recall that an isotropic scheme contains an l-split rank one torus Gm,l. Choose and fix two opposite parabolic subgroups Pl = P + l and U − l be their unipotent radicals. We will be interested mostly in the group of points Gl(l(t)). The following definition originates from [T, Main theorem]. l of the l-group scheme Gl. Let U + l and P − Definition 8.3. Define Gl(l(t))+ as the subgroup of the group Gl(l(t)) generated by l(t)- points of unipotent radicals of all parabolic subgroups of Gl defined over the field l. Remark 8.4. Clearly, l(t) = l(t−1). Thus, Gl(l(t)) = Gl(l(t−1)) and Gl(l(t))+ = Gl(l(t−1))+. By definition the group Gl(l(t))+ is generated by unipotent radicals of all l-parabolic subgroups, and thus contains the elementary group EPl(l(t)), introduced in Definition 8.1. In fact they coincide. 23 Proposition 8.5. The group Gl(l(t))+ is generated by l(t)-points of unipotent radicals of any two opposite parabolic subgroups of the l-group scheme Gl. In particular, one has the equality Gl(l(t−1))+ = DU + l (l(t−1)), U − l (l(t−1))E = EPl(l(t−1)). (23) Proof. Set Gl(t) = Gl ×Spec l Spec l(t). The group Gl(l(t))+ is contained in the subgroup of Gl(l(t)) = Gl(t)(l(t)) generated by l(t)-points of unipotent radicals of all parabolic sub- groups of the group scheme Gl(t) defined over the field l(t). By [BT3, Prop.6.2.(v)] the latter group is generated by l(t)-points of unipotent radicals of any two opposite parabolic subgroups of Gl(t), in particular, by l(t)-points of U + l(t)(l(t)) = U ± l(t). Since U ± l(t) and U − l (l(t)), we have (23). Remark 8.6. For any commutative ring A and an A-algebra B, and any reductive A- group scheme G we can define the group GA(B)+ as in the Definition 8.3, that is, as the subgroup generated by B-points of unipotent radicals of all A-parabolic subgroups of G. The question, whether this subgroup coincides with EP (B) for an A-parabolic subgroup P of G, is in general rather subtle. See the paper [PeSt1] by V. Petrov and the second author for details. The following result is crucial for the sequel. Proposition 8.7. One has the equality Gl(l(t−1)) = Gl(l(t−1))+ · Gl(l), (24) where Gl(l(t−1))+ is the group defined in Definition 8.3 (see also Remark 8.4). Proof. This is proved in [Gi, Th´eor`eme 5.8]. Let f (t) ∈ l[t] be a polynomial of degree n = deg(f ) in t such that f (0) 6= 0. We consider the reciprocal polynomial f ∗(t−1) := f (t)/tn ∈ l[t−1]; clearly, f ∗(0) 6= 0. Conversely, if we are given a polynomial g(t−1) ∈ l[t−1] of degree n = deg(g) in t−1 such that g(0) 6= 0, we define the reciprocal polynomial g∗(t) := g(t−1) · tn ∈ l[t]. The above correspondences are mutually inverse. Further, when f (t) ∈ l[t] runs over all polynomials in t with f (0) 6= 0, then the reciprocal polynomial f ∗(t−1) runs over all polynomials g(t−1) ∈ l[t−1] with g(0) 6= 0. Now let us return to the setting considered in (23) and (24) and Remark 8.4. Each non-constant f (t) ∈ l[t] admits a unique factorisation of the form f (t) = tr · g(t), where g(t) ∈ l[t] and g(0) 6= 0. Clearly, for each h(t) ∈ l[t] with h(0) 6= 0 one gets the following inclusions Gl(cid:0)l[t]f(cid:1) ≤ Gl(cid:0)l[t−1, t]g∗(cid:1) ≤ Gl(cid:0)l[t−1, t]g∗h∗(cid:1). This leads us to the following Lemma. 24 Lemma 8.8. Let Pl be an arbitrary parabolic l-subgroup of Gl. For each α ∈ Gl(l[t]f (t)) one can find a polynomial h(t) ∈ l[t], h(0) 6= 0, and elements u ∈ EPl(cid:0)l[t−1, t]g∗h∗(cid:1), β ∈ Gl(cid:0)l(cid:1) such that The chain of l-algebra inclusions l[t]f h ⊆ l[t]tf h = l[t, t−1]gh = l[t−1, t]g∗h∗ shows that α = uβ ∈ Gl(cid:0)l[t−1, t]g∗h∗(cid:1). (25) u ∈ EPl(cid:0)l[t]tf h(cid:1), and α ∈ Gl(cid:0)l[t]tf h(cid:1). Proof. As observed above, inclusions α ∈ Gl(cid:0)l[t−1, t]g∗(cid:1) ≤ Gl(cid:0)l[t−1, t]g∗h∗(cid:1) are obvious. The equalities (24) and (23) imply that there exists a polynomial h(t) ∈ l[t], h(0) 6= 0, and elements u ∈ EPl(cid:0)[t−1, t]g∗h∗(cid:1), β ∈ Gl(cid:0)l(cid:1), such that α = uβ in Gl(cid:0)l[t−1, t]g∗h∗(cid:1). The last assertion of the Lemma follows from the obvious l-algebra inclusions l[t]f h ⊆ l[t]tf h = l[t, t−1]gh = l[t−1, t]g∗h∗. 9 Principal G-bundles on a projective line The main result of the present section is Corollary 9.8, which implies Theorem 1.3. Let B be a Noetherian commutative ring, and let A1 B be the affine line and the projective line over B, respectively. Usually we identify the affine line with a subscheme B − ({∞} × Spec(B)), where ∞ = [0 : 1] ∈ P1. of the projective line as follows A1 Let G be a semi-simple B-group scheme, let P a principal G-bundle over A1 B, and let p : P → A1 B be the corresponding canonical projection. B and P1 B = P1 For a monic polynomial f = f (t) = tn + an−1tn−1 + · · · + a0 ∈ B[t] we set Pf = p−1((A1 denote by B)f ). Clearly, it is a principal G-bundle over (A1 B)f . Further, we F (t0, t1) = tn 1 + an−1tn−1 1 t0 + · · · + a0tn 0 the corresponding homogeneous polynomial in two variables. Note that the intersection B defined by the inequality F 6= 0 with the affine line A1 of the principal open set in P1 B equals the principal open subset (A1 B)f . As in the previous section in the case where a0 6= 0 we consider the reciprocal polynomial f ∗(t−1) ∈ B[t−1] equal to f (t)/tn. Definition 9.1. Let ϕ : G(A1 P (ϕ, f ) for a principal G-bundle over the projective line P1 B)f via the principal G-bundle isomorphism ϕ. G(P1 B )f → Pf be a principal G-bundle isomorphism. We write B obtained by gluing P and B )F over (A1 25 Remark 9.2. For any ϕ and f as above, the principal G-bundles P (ϕ, f ) and P (ϕ, f g) coincide for each monic polynomial g ∈ B[t]. For any ϕ and f , any monic polynomial h(t) ∈ B[t] such that h(0) ∈ B× is invertible, and any β ∈ G(B[t−1]h∗) the principal G-bundles P (ϕ, f ) and P (ϕ◦β, tf h) are isomorphic. In fact, they differ by a co-boundary. Lemma 9.3. Let l be a field and Gl be a semi-simple l-group scheme. Let f ∈ l[t] be a non-constant polynomial. Let P be a principal Gl-bundle over A1 l such that Pf is trivial over A1 → Pf be a principal Gl-bundle isomorphism. Let P (ϕ, f ) be the corresponding principal G-bundle over P1 l . Then there exists an α ∈ G(l[t]f ) such that the principal G-bundle P (ϕ ◦ α, f ) is trivial over P1 l . f . Let ϕ : GA1 f Proof. By [C-T/O, Prop. 2.2] one has ker[H1 ´et(l[t], Gl) → H1 ´et(l(t), Gl)] = ∗ . In this case the So, we may assume that there is an isomorphism GA1 above isomorphism ϕ coincides with the right multiplication by an element β ∈ Gl(l[t]f ). Clearly, P (β ◦ β−1, f ) is trivial over P1 l . Thus, P (ϕ ◦ α, f ) is trivial for α = β−1. = P over A1 l . l Corollary 9.4. Let l be a field, and let Gl be an isotropic simply connected semi-simple l-group scheme with a parabolic l-subgroup Ql. Let P be a Gl-bundle over A1 l . Further, be a principal Gl-bundle let f (t) ∈ l[t] be a non-constant polynomial, ϕ : GA1 isomorphism and let P (ϕ, f ) be the corresponding principal Gl-bundle on the projective line P1 l . Then there exist h(t) ∈ l[t] and u ∈ EQl(l[t]tf h) such that the principal Gl-bundle P (ϕ ◦ u, tf h) is trivial over P1 l . → PA1 f f Proof. By Lemma 9.3 there exists an α ∈ Gl(l[t]f ) such that the principal Gl-bundle P (ϕ ◦ α, f ) is trivial. Let f (t) = trg(t) be the unique factorisation such that g(t) ∈ l[t] and g(0) 6= 0. By Lemma 8.8 there exist an element h(t) ∈ l[t] with h(0) 6= 0 and elements u ∈ Gl(l[t]tf h)+, β ∈ Gl(l) such that α = uβ ∈ Gl(l[t−1, t]g∗h∗). (26) The following chain of principal Gl-bundle isomorphisms completes the proof: Gl ×Spec(l) P1 l = P (ϕ ◦ α, f ) = P (ϕ ◦ α, tf h) = P (ϕ ◦ u ◦ β, tf h) ∼= P (ϕ ◦ u, tf h). Here all the equalities are obvious. The last isomorphism holds by Remark 9.2, since β ∈ Gl(l[t−1]g∗h∗). 26 Let B′ be a Noetherian semi-local ring. Let G be a simple simply connected B′-group scheme. Let mi ⊆ B′, i = 1, 2, . . . , n, be all maximal ideals of B′. Let J be the intersection of all mi, 1 ≤ i ≤ n. Then l := B′/J = l1 × l2 × · · · × ln, where li = B′/mi. Let Gl = G ⊗B′ l be the fibre of G over Spec(l). In the sequel we write P1 and A1 for P1 l denote the projective line and the affine line over l. B′ respectively, whereas P1 B′ and A1 l and A1 Let f ∈ B′[t] be a monic polynomial, and let P be a principal GB′-bundle over A1 be a principal G-bundle isomorphism, and such that PA1 let P (ϕ, f ) be the corresponding principal G-bundle on P1 (see Definition 9.1). is trivial. Let ϕ : GA1 → PA1 f f f Theorem 9.5. Assume that the group scheme G over B′ is isotropic, simple and sim- ply connected. Then there exist a monic polynomial h(t) ∈ B′[t] and an element α ∈ G(B′[t]tf h) such that the principal G-bundle P (ϕ ◦ α, tf h) satisfies the condition (i) P (ϕ ◦ α, tf h)P1 l is a trivial principal Gl-bundle over the projective line P1 l . Proof. We denote by f the image of f in l[t], by P the restriction of P to A1 l , by P (ϕ, f ) the restriction of P (ϕ, f ) to the projective line P1 l , etc. Let Q be a parabolic B′-subgroup of G. By Corollary 9.4 there exist a monic polynomial h(t) ∈ l[t] such that h(0) ∈ l×, and an element u ∈ EQl(cid:0)l[t]tf h(cid:1) ≤ Gl(cid:0)l[t]tf h(cid:1) such that the principal Gl-bundle P (ϕ ◦ u, tf h) is trivial over P1 l . Choose a monic polynomial h(t) ∈ B′[t] of degree equal to the degree of h(t) and such that h(t) modulo J coincides with h(t). Clearly, the homomorphism of B′-algebras B′[t]tf h → l[t]tf h is surjective. By Lemma 8.2 it induces a surjective group homomorphism EQ(cid:0)B′[t]tf h(cid:1) → EQ(cid:0)l[t]tf h(cid:1) = EQl(cid:0)l[t]tf h(cid:1). Thus, there exists an α ∈ EQ(B′[t]tf h) ≤ G(B′[t]tf h) such that α equals u modulo J; we write α = u. Consider the G-bundle P (ϕ ◦ α, tf h). We claim that its restriction to the projective line P1 l is trivial. Indeed, one has the following chain of equalities of principal Gl-bundles over P1 l : P (ϕ ◦ α, tf h) = P (ϕ ◦ α, tf h) = P (ϕ ◦ u, tf h), where the principal Gl-bundle P (ϕ ◦ u, tf h) is trivial over P1 l . We keep the same notation as introduced before Theorem 9.5. Theorem 9.6. Assume that the Noetherian semi-local ring B′ contains a field k. Let G be a not necessarily isotropic simple simply connected B′-group scheme. Let E be a principal G-bundle over P1 whose restriction to the closed fibre EP1 is trivial. Then E is of the form: E = pr∗(E0), where E0 is a principal G-bundle over Spec(B′) and pr : P1 → Spec(B′) is the canonical projection. l 27 Proof. See Appendix, 12.2. Let us state an important corollary of the above theorems. Corollary 9.7. Let k be a field, and let B′ be a semi-local Noetherian algebra over k. Let G be an isotropic simple simply connected B′-group scheme. Further, let P be a principal G-bundle over A1. Assume that there exists a monic polynomial f ∈ B′[t] such that the principal G-bundle PA1 is trivial. Then the principal G-bundle P is trivial. f Proof. Let f ∈ B′[t] be a monic polynomial such that the principal G-bundle PA1 is . By Theorem 9.5 trivial. Choose a principal G-bundle isomorphism ϕ : GA1 there exists a monic polynomial h(t) ∈ B′[t] and an element α ∈ G(B′[t]tf h) such that the of the principal G-bundle P (ϕ ◦ α, tf h) to the projective line restriction P (ϕ ◦ α, tf h)P1 P1 l is a trivial principal Gl-bundle. f l → PA1 f f By Theorem 9.6 the principal G-bundle P (ϕ ◦ α, tf h) is of the form: P (ϕ ◦ α, tf h) = pr∗(P0), where P0 is a principal G-bundle over Spec(B′). Note that G{∞}×Spec(B′) ∼= P (ϕ ◦ α, tf h){∞}×Spec(B′), that is the restriction of P (ϕ ◦ α, tf h) to {∞} × Spec(B′) is trivial. Thus GP1 ∼= P (ϕ ◦ α, tf h). Since the original principal G-bundle P over A1 is isomorphic to P (ϕ ◦ α, tf h)A1, it follows that P is trivial. This finishes the proof. Corollary 9.8 (=Theorem 1.3). Let k be a field and let B be a Noetherian k-algebra. Assume that a group scheme G over B is simple, simply connected and isotropic. Further, let P be a principal G-bundle over A1 B. Assume that there exists a monic polynomial B )f is trivial and f (1) ∈ B is invertible. f ∈ B[t] such that the principal G-bundle P(A1 Then the principal G-bundle P is trivial. Proof. It is routine to prove that there is a closed B-group scheme embedding G ֒→ GLN,B for an N > 0. Since f (1) is invertible, the principle G-bundle P is trivial at the closed subscheme {1} × Spec(B) ⊂ A1 B. By Corollary 9.7, for any maximal ideal m of B, the bundle PA1 is trivial too. Now [Mo, Korollar 3.5.2] completes the proof of Corollary 9.8. Bm 10 Proofs of Theorems 1.1 and 1.6, and of Corol- lary 1.7 Proof of Theorem 1.1. Substitute to Theorem 1.3 B = O ⊗k A, Pt := Gt, h(t) = f (t) ⊗ 1 from Theorem 1.2. By Theorem 1.3 the G-bundle Gt is trivial. Now by the item (ii) of Theorem 1.2 the original G-bundle G is trivial. 28 Proof of Theorem 1.6. Let G be a principal G-bundle over R ⊗Z A that becomes trivial over K ⊗Z A. Clearly, there is a non-zero f ∈ R such that G is trivial over Rf ⊗Z A. Let k′ be the prime subfield of R. It follows from Popescu's theorem [Po, Sw] that R is a filtered inductive limit of smooth k′-algebras Rα. Then there exist an index α, a reductive group scheme Gα over Rα, a principal Gα-bundle Gα over Rα ⊗Z A, and an element fα ∈ Rα such that G = Gα ×Spec(Rα) Spec(R), G ∼= Gα ×Spec(Rα⊗ZA) Spec(R ⊗Z A) as principal G-bundles, f is the image of fα under the map ϕα : Rα → R, and Gα is trivial over (Rα)fα ⊗Z A. If the field k′ is infinite, then for each maximal ideal mi in R (i = 1, . . . , n) set pi = ϕ−1 α (mi). The map ϕα induces a map of semi-local rings (Rα)p1,...,pn → R. Since the principal Gα-bundle Gα is trivial over (Rα)fα ⊗Z A ∼= (Rα)fα ⊗k′ (k′ ⊗Z A), by Theorem 1.1 the bundle Gα is trivial over (Rα)p1,...,pn ⊗k′ (k′ ⊗Z A) ∼= (Rα)p1,...,pn ⊗Z A. Whence the G-bundle G is trivial over R ⊗Z A. Now consider the case where the field k′ is finite. Since R contains an infinite field by the assumption of the theorem, R also contains a field k′(t) of rational functions in one α = Rα ⊗k′ k′(t), then the map ϕα can be decomposed as follows variable t over k′. Set R′ Rα → Rα ⊗k′ k′(t) = R′ α ψα−→ R. α = Gα ×Spec(Rα) Spec(R′ α), G′ α = fα ⊗ 1 ∈ α is trivial over α ⊗Z A. Arguing exactly as in the previous case with the field k′(t) instead of k′, we Set G′ R′ (R′ conclude, by means of Theorem 1.1, that the G-bundle G is trivial over R ⊗Z A. α is a smooth k′(t)-algebra, and the principal G′ α = Gα ×Spec(Rα⊗ZA) Spec(R′ α. Then R′ α)f ′ α ⊗Z A), f ′ α-bundle G′ Proof of Corollary 1.7. Consider the commutative diagram H 1 ´et(R[t], G) H 1 ´et(K[t], G) t=0 t=0 H 1 ´et(R, G) / H 1 ´et(K, G). (27) Since K is perfect, the bottom arrow is bijective by the main result of [RR]. Therefore, any element ξ ∈ H 1 ´et(R, G) also has trivial image in H 1 ´et(K[t], G). By Theorem 1.6, for any maximal ideal m ⊆ R the map ´et(R[t], G) having trivial image in H 1 H 1 ´et(Rm[t], G) → H 1 ´et(K[t], G) has trivial kernel. Therefore, for any maximal ideal m, the image of ξ in H 1 trivial as well. By [Mo, Korollar 3.5.2] this implies that ξ is trivial. ´et(Rm[t], G) is 11 Semi-simple case In the present Section we show how Theorem 1.1 extends to the case of semi-simple simply connected groups; this is Theorem 11.2 below. One readily sees that Theorem 1.6 and 29 / /     / Corollary 1.7 extend to semi-simple simply connected groups as well, once we substitute the isotropy condition imposed in these statements by the same one as in Theorem 11.2. By [SGA3, Exp. XXIV 5.3, Prop. 5.10] the category of semi-simple simply connected group schemes over a Noetherian domain R is semi-simple. In other words, each object has a unique decomposition into a product of indecomposable objects. Indecomposable objects can be described as follows. Take a domain R′ such that R ⊆ R′ is a finite ´etale extension and a simple simply connected group scheme G′ over R′. Now, applying the Weil restriction functor RR′/R to the R-group scheme G′ we get a simply connected R-group scheme RR′/R(G′), which is an indecomposable object in the above category. Conversely, each indecomposable object can be constructed in this way. Definition 11.1. We say that an indecomposable semi-simple simply connected group schemes H = RR′/R(H ′) over a Noetherian domain R is isotropic if H ′ is isotropic. Theorem 11.2. Let k be an infinite field. Let O be the semi-local ring of finitely many closed points on a smooth irreducible k-variety X and let K be its field of fractions. Let G be a semi-simple simply connected O-group scheme all of whose indecomposable factors are isotropic in the sense of Definition 11.1. Then for any Noetherian k-algebra A the map H 1 ´et(R ⊗k A, G) → H 1 ´et(K ⊗k A, G), induced by the inclusion R into K, has trivial kernel. Proof. Take a decomposition of G into indecomposable factors G = G1 × G2 × · · · × Gr. Clearly, it suffices to check that for each index i the kernel of the map H 1 ´et(R ⊗k A, Gi) → H 1 ´et(K ⊗k A, Gi) is trivial. We know that there exists a finite ´etale extension R′ and the Weil restriction RR′ i) coincides with Gi. i/R(G′ i/R such that R′ i is a domain The Faddeev -- Shapiro Lemma [SGA3, Exp. XXIV Prop. 8.4] states that there is a canonical isomorphism H 1 ´et(cid:0)R ⊗k A, RR′ i/R(G′ i)(cid:1) ∼= H 1 ´et(cid:0)R′ ⊗k A, Gi(cid:1) that preserves the distinguished point. To complete the proof, it only remains to apply Theorem 1.6 to the semi-local regular ring R′ i-group scheme G′ i. i, its fraction field Ki and the simple R′ 12 Appendix 12.1 Almost elementary fibration In this section we prove Propositions 2.3 and 2.4. 30 contains the points x1, x2, . . . , xn. Set x := Proof of Proposition 2.3. The proof almost literally follows the proof of the original Artin's result [A, Exp. XI, Prop. 3.3]. Shrinking X, may assume that X ⊂ Ar k is affine and still k. Let ¯X is the normalization of X0 and set Y = ¯X − X with the induced reduced structure. Let Z ⊂ X be a subset of ¯X consisting of all non-regular points of ¯X. By [EGAIV, Cor. 6.12.5] the set Z is Zariski closed in ¯X. Since ¯X is normal we conclude that dim Z ≤ n−2. Since X is k-smooth one has an inclusion Z ⊂ Y . Summarizing one has xi. Let X0 be the closure of X in Pr n`j=1 (i) Z ⊂ Y , (ii) dim ¯X = dim X = n, (iii) dim Y = n − 1, (iv) dim Z ≤ n − 2. Shrinking X and following Artin's procedure one can construct a diagram of the form X j &▼▼▼▼▼▼▼▼▼▼▼▼▼▼ p p / X S Y ′ ′ i q xqqqqqqqqqqqqqq (28) k subject to conditions (i), (ii) of Definition 2.1 and such that x ⊂ X and S is an affine open subset in Pn−1 and i is a closed imbedding and Y ′ is a regular scheme. Moreover, the restriction of q ⊗k ¯k : Y ′ ⊗k ¯k → S ⊗k ¯k to the reduced subscheme (Y ′ ⊗k ¯k)red is a finite ´etale morphism all of whose fibres are non-empty (here ¯k is the algebraic closure of k). In this case for each irreducible component Y ′ r → S is a finite surjective morphism. Since Y ′ r , S are regular irreducible schemes of the same dimension, the morphism qY ′ r is finite flat (see Grothendieck [E, Cor. 17.18]). Thus q is subject to the condition (iii) of Definition 2.1. Finally, the ideal sheaf IY ′ defining the closed subscheme Y ′ in X is locally principal. In fact, S is regular and p is smooth. So, X . Thus IY ′ is locally principal. Whence the Proposition. is regular. The closed subscheme Y ′ is regular of pure codimension one in X r of Y ′ the restriction qY ′ r : Y ′ ′ ′ ′ Proof of Proposition 2.4. To prove this Proposition it suffices to construct a finite surjec- tive S-morphism ¯π : X → P1 × S such that Yred = ¯π−1({∞} × S) set-theoretically. To do that, we first note that, under the hypotheses of the Proposition, the closed subscheme Y of X is a locally principal divisor. We will construct a desired ¯π using two sections t0 and t1 of the sheaf O(nY ) for a sufficiently large n. Assume that t0 and t1 are such that the vanishing locus of t0 is nY and the vanishing locus of t1 does not intersect Y . Then the pair t0, t1 defines a regular map ϕ := [t0 : t1] : X → P1. Set ¯π = (ϕ, ¯p) : X → P1 × S. Clearly, ¯π is 31 & /   o o x It is a quasi-finite surjective morphism. an S-morphism of the S-schemes. It is a projective morphism since both S-schemes are projective S-schemes. In fact, for each point s ∈ S the morphism ¯π induces a non-constant morphism X s → P1 s of two k(s)-smooth geometrically irreducible projective k(s)-curves. Thus ¯π is finite surjective as a quasi- finite projective morphism. It remains to find an appropriate integer n and two sections t0 and t1 with the above properties. Firstly, for each point s of the scheme S set X s := (¯π)−1(s) scheme-theoretically, and note that X s is a k(s)-smooth geometrically irreducible projective k(s)-curve. The morphism ¯π is smooth. In particular, it is flat. Whence the function s 7→ χ(X s, OX s) is constant by [Mu, Ch. II, Sect. 5, Cor. 1]. The latter means that the genus g(X s) is the same for all points s ∈ S. Set g = g(X s). By the assumption Y is finite flat over S and S is semi-local. Let r be the rank of the free Γ(S, OS)-module Γ(Y, OY ). Assume that n ≥ 2g − 1, then h0(X s, OX s(nYs)) = χ(X s, OX s(nYs)) = rn − g + 1. Let En := ¯p∗(OX s(nYs))). By [Mu, Ch. II, Sect. 5, Cor. 1] and [Mu, Ch. II, Sect. 5, Lem. 1] the sheaf En on S is locally free of rank rn − g + 1, and for each point s ∈ S the canonical map En ⊗OS k(s) can−−→ H 0(X s, OX s(nYs)) is an isomorphism. Let s = ` si, where si are all closed points of the semi-local scheme S. Let k(s) = Q k(si), where k(si) denotes the residue field of the point si. Consider the commutative diagram H 0(S, En) id H 0(X, OX (nY )) (29) α β En ⊗OS k(s) can / H 0(X s, OX s(nYs)), where α, β, and can are the canonical homomorphisms. As mentioned in the previous paragraph, the map can is an isomorphism. The map α is surjective, since s = ` si is a closed subscheme of the affine scheme S. Whence the map β is surjective. For each si ∈ s the curve X si is a k(si)-smooth geometrically irreducible k(si)-curve of genus g. Whence there exists an integer n0 such that for each n ≥ n0 and each si ∈ s has H 1(X si, O((n − 1)Ysi)) = 0. Thus there exists for any i a section t1,i of OX si (nYsi) that does not vanish on Ysi. By the surjectivity of β we may choose a section t1 of OX (nY ) such that β(t1)X si = t1,i for any i. The vanishing locus of t1 does not intersect Ys, whence it does not intersect Y . Clearly, t1 is the desired section of OX (nY ). It remains to take for t0 a section of OX (nY ) with the vanishing locus nY . 12.2 Proof of the Horrocks type Theorem 9.6 In this section we give a proof of Theorem 9.6. This proof is rather standard, and for the most part follows [R1]. However, our group scheme G does not come from the ground field k. Therefore, we have to somewhat modify Raghunathan's arguments. We will use the following lemma. 32 / /     / W the projective line over W . Let F ∈ H 1(P1 Lemma 12.1. Let W be a semi-local irreducible Noetherian scheme over an arbitrary field k. Let H and H ′ be two reductive group schemes over W , such that H is a closed W - subgroup scheme of H ′, and denote by j : H ֒→ H ′ the corresponding embedding. Denote by P1 W , H) be a principal H-bundle, and let M := j∗(F ) ∈ H 1(P1 W , H ′) be the corresponding principal H ′-bundle. If M is a trivial H ′-bundle, then there exists a principal H-bundle F0 over W such that pr∗(F0) ∼= F , where pr : P1 W → W is the canonical projection. Proof. Set X = H ′/j(H). Locally in the ´etale topology on W this scheme is isomorphic to the W -scheme W ×Spec(k) H ′ 0,k are the split reductive k-group schemes of the same types as H and H ′ respectively. By results of Haboush [Hab] and Nagata [Na] (see Nisnevich [Ni1, Corollary]) the k-scheme H ′ 0,k/H0,k is an affine k-scheme. Thus X is an affine W -scheme. Consider the long exact sequence of pointed sets 0,k/H0,k, where H0,k and H ′ 1 → H(P1 W ) j∗−→ H ′(P1 W ) → X(P1 W ) ∂−→ H 1 ´et(P1 W , H) j∗−→ H 1 ´et(P1 W , H ′). Since j∗(F ) is trivial, there is ϕ ∈ X(P1 W ) such that ∂(ϕ) = F . The W -morphism ϕ : P1 W → X is a W -morphism of a W -projective scheme to a W -affine scheme. Thus ϕ is "constant", that is, there exists a section s : W → X such that ϕ = s ◦ pr. Consider another long exact sequence of pointed sets, this time the one corresponding to the scheme W , and the morphism of the first sequence to the second one induced by the projection pr. We get a big commutative diagram. In particular, we get the following commutative square X(W ) pr∗ W X(P1 W ) ∂ ∂ We have pr∗ W (s) = ϕ. Hence / H 1 ´et(W, H) pr∗ W (30) / H 1 ´et(P1 W , H). F = ∂(ϕ) = ∂(pr∗ W (s)) = pr∗ W (∂(s)). Setting F0 = ∂(s) we see that F = pr∗ W (F0). The Lemma is proved. Proof of Theorem 9.6. It is routine to prove that there is a closed B′-group scheme em- bedding j : G ֒→ GLN,B′ for an N > 0. By the assumption of the theorem, the G-bundle E is trivial on P1 l ⊂ P1. Hence the GLN,B′-bundle j∗(E) over P1 is trivial over P1 l . The B′-group scheme GLN,B′ is just the ordinary general linear group. Thus j∗(E) corresponds to a vector bundle M over P1. Moreover, this vector bundle is trivial on P1 l . Using the equality H 1(P1, OP1) = 0 and [EGAIII, Cor. 4.6.4], we see that M is of the form M = pr∗(M0) for a vector bundle M0 over Spec(B′). Since B′ is semi-local, M0 is trivial over Spec(B′). Thus M is trivial on P1. Thus j∗(F ) is a trivial GLN,B′-bundle. Now, applying Lemma 12.1 to the embedding j : G ֒→ GLN,B′, we see that E = pr∗(E0) for some E0 ∈ H 1(B′, G). Theorem 9.6 is proved. 33   /   / References [A] [BT1] [BT2] [BT3] [Ch] Artin, M. Comparaison avec la cohomologie classique: cas d'un pr´esch´ema lisse, in Th´eorie des topos et cohomologie ´etale des sch´emas (SGA 4). Tome 3. Lect. Notes Math., vol. 305, Exp. XI, Springer-Verlag, Berlin-New York, 1973. Borel, A.; Tits, J. Groupes r´eductifs, Publ. Math. IH´ES 27 (1965), 55 -- 151. Borel, A.; Tits, Publ. Math. IH´ES 41 (1972), 253 -- 276. J. Compl´ements `a l'article "Groupes r´eductifs", Borel, A.; Tits, J. Homomorphismes "abstraits" de groupes algebriques simples, Ann. Math. 97 (1973), no. 3, 499 -- 571. Chernousov, V. Variations on a theme of groups splitting by a quadratic ex- tension and Grothendieck-Serre conjecture for group schemes F4 with trivial g3 invariant, Doc. Math., Extra Volume: Andrei A. Suslin's Sixtieth Birthday (2010), 147 -- 169. [ChP] Chernousov, V.; Panin, I. Purity of G2-torsors, C. R. Math. Acad. Sci. Paris 345 (2007), no. 6, 307 -- 312. [C-T/O] Colliot-Th´el`ene, J.-L.; Ojanguren, M. Espaces Principaux Homog`enes Locale- ment Triviaux, Publ. Math. IH´ES 75 (1992), no. 2, 97 -- 122. [C-T/S] Colliot-Th´el`ene, J.-L.; Sansuc, J.-J. Principal homogeneous spaces under flasque tori: Applications, Journal of Algebra 106 (1987), 148 -- 205. [SGA3] Demazure, M.; Grothendieck, A. Sch´emas en groupes, Lect. Notes Math., vol. 151 -- 153, Springer-Verlag, Berlin-Heidelberg-New York, 1970. [E] [FP] [Ga] [Gi] [Gr1] Eisenbud, D. Commutative algebra with a view toward algebraic geometry. Graduate Texts in Mathematics 150, Springer-Verlag, New York, 1995. Fedorov, R.; Panin, I. A proof of Grothendieck -- Serre conjecture on principal bundles over a semilocal regular ring containing an infinite field, Preprint, April, 2013, http://www.arxiv.org/abs/1211.2678v2. Gabber, O. announced and still unpublished. Gille, Ph. Le probl`eme de Kneser-Tits, S´em. Bourbaki 983 (2007), 983-01 -- 983- 39. Grothendieck, A. Torsion homologique et in An- neaux de Chow et applications, S´eminaire Chevalley, 2-e ann´ee, Secr´etariat math´ematique, Paris, 1958. section rationnalles, 34 [EGAIII] Grothendieck, A. ´El´ements de g´eom´etrie alg´ebrique (r´edig´es avec la collabora- tion de Jean Dieudonn´e) : III. ´Etude cohomologique des faisceaux coh´erents, Premi`ere partie, Publ. Math. IH´ES 11 (1961), 5 -- 167. [EGAIV] Grothendieck, A. ´El´ements de g´eom´etrie alg´ebrique (r´edig´es avec la collabora- tion de Jean Dieudonn´e) : IV. ´Etude locale des sch´emas et des morphismes de sch´emas, Seconde partie, Publ. Math. IH´ES 24 (1965), 5 -- 231. [SGA1] Grothendieck, A. Revetements ´etales et groupe fondamental (SGA 1). Fasc. I: Expos´es 1 `a 5. S´eminaire de G´eom´etrie Alg´ebrique, 1960/61, Inst. Hautes ´Etudes Sci., Paris, 1963. [Gr2] [Hab] [Ha] [Mo] [Mu] [Na] [Ni1] [Ni2] [OP1] [OP2] [OPZ] Grothendieck, A. Le group de Brauer II, in Dix expos´es sur la cohomologique de sch´emas, Amsterdam, North-Holland, 1968. Haboush, W.J. Reductive groups are geometrically reductive, Ann. Math. 102 (1975), no. 1, 67 -- 83. Hartshorne, R. Algebraic geometry. Graduate Texts in Mathematics 52, Springer-Verlag, New York-Heidelberg, 1977. Moser, L.-F. Grothendiecksche http://www.mathematik.uni-muenchen.de/ lfmoser/da.pdf. Vermutung, Torseure Rational triviale und Diplomarbeit, die Serre- 2008, Mumford, D. Abelian Varieties, Oxford University Press, Oxford, 1970. Nagata, M. Invariants of a group in an affine ring, J. Math. Kyoto Univ. 3 (1964), no. 3, 369 -- 377. Nisnevich, E.A. Affine homogeneous spaces and finite subgroups of arithmetic groups over function fields, Functional Analysis and Its Applications 11 (1977), no. 1, 64 -- 66. Nisnevich, Y. Rationally Trivial Principal Homogeneous Spaces and Arithmetic of Reductive Group Schemes Over Dedekind Rings, C. R. Acad. Sci. Paris, S´erie I, 299 (1984), no. 1, 5 -- 8. Ojanguren, M.; Panin, I. A purity theorem for the Witt group, Ann. Sci. Ecole Norm. Sup. (4) 32 (1999), no. 1, 71 -- 86. Ojanguren, M.; Panin, I. Rationally trivial hermitian spaces are locally trivial, Math. Z. 237 (2001), 181 -- 198. Ojanguren, M.; Panin, I.; Zainoulline, K. On the norm principle for quadratic forms, J. Ramanujan Math. Soc. 19 (2004), no. 4, 1 -- 12. 35 [PS] [P1] [P2] Panin, I.; Suslin, A. On a conjecture of Grothendieck concerning Azumaya algebras, St. Petersburg Math. J. 9 (1998), no. 4, 851 -- 858. Panin, I. A purity theorem for linear algebraic groups, Preprint, 2005, http://www.math.uiuc.edu/K-theory/0729. I. On Grothendieck -- Serre's Panin, pal G-bundles over http://www.math.org/0905.1423v3. princi- reductive group schemes:II, Preprint, April, 2013, concerning conjecture [PPeSt] Panin, I.; Petrov, V.; Stavrova, A. On Grothendieck -- Serre's for sim- 2009, and E7, Preprint, types E6 of adjoint ple http://www.math.uiuc.edu/K-theory/. group schemes [PeSt1] Petrov V.; Stavrova A. Elementary subgroups in isotropic reductive groups, St. Petersburg Math. J. 20 (2009), no. 4, 625 -- 644. [PeSt2] Petrov groups http://www.mathematik.uni-bielefeld.de/LAG/man/374.html. Grothendieck -- Serre invariant, type F4 with Stavrova, V.; of trivial A. f3 conjecture Preprint, for 2009, [Po] [R1] [R2] [RR] [Se] [Sw] [T] [Z] [Vo] Popescu, D. General N´eron desingularization and approximation, Nagoya Math. J. 104 (1986), 85 -- 115. Raghunathan, M.S. Principal bundles admitting a rational section, vent. Math. 116 (1994), no. 1 -- 3, 409 -- 423. In- Raghunathan, M.S. Erratum: Principal bundles admitting a rational section, Invent. Math. 121 (1995), no. 1, 223. Raghunathan, M.S.; Ramanathan, A. Principal bundles on the affine line, Proc. Indian Acad. Sci., Math. Sci. 93 (1984), 137 -- 145. Serre, J.-P. Espaces fibr´es alg´ebriques, in Anneaux de Chow et applications, S´eminaire Chevalley, 2-e ann´ee, Secr´etariat math´ematique, Paris, 1958. Swan, R.G. N´eron -- Popescu desingularization, Algebra and Geometry (Taipei, 1995), Lect. Algebra Geom. 2, Internat. Press, Cambridge, MA, 1998, 135 -- 192. Tits, J. Algebraic and abstract simple groups, Ann. Math. 80 (1964), no. 2, 313 -- 329. Zainoulline, K.V. On Grothendieck's conjecture on principal homogeneous spaces for some classical algebraic groups, St. Petersburg Math. J. 12 (2001), no. 1, 117 -- 143. Voevodsky, V. Cohomological theory of presheaves with transfers, in Cycles, Transfers, and Motivic Homology Theories, Ann. Math. Studies, 2000, Prince- ton University Press. 36 

1611.00971

2

1611

Explicit description of jumping phenomena on moduli spaces of parabolic connections and Hilbert schemes of points on surfaces

In this paper, we investigate the apparent singularities and the dual parameters of rank 2 parabolic connections on $\mathbb{P}^1$ and rank 2 (parabolic) Higgs bundle on $\mathbb{P}^1$. Then we obtain explicit descriptions of Zariski open sets of the moduli space of the parabolic connections and the moduli space of the Higgs bundles. For $n=5$, we can give global descriptions of the moduli spaces in detail.

math.AG

math

EXPLICIT DESCRIPTION OF JUMPING PHENOMENA ON MODULI SPACES OF PARABOLIC CONNECTIONS AND HILBERT SCHEMES OF POINTS ON SURFACES ARATA KOMYO AND MASA-HIKO SAITO Abstract. In this paper, we investigate the apparent singularities and the dual parameters of rank 2 parabolic connections on P1 and rank 2 (parabolic) Higgs bundle on P1. Then we obtain explicit descriptions of Zariski open sets of the moduli space of the parabolic connections and the moduli space of the Higgs bundles. For n = 5, we can give global descriptions of the moduli spaces in detail. 1. Introduction The purpose of this paper is to give explicit descriptions of the moduli space of rank 2 (parabolic) Higgs bundles on P1 and the moduli space of rank 2 parabolic connections on P1 by the apparent singularities and their dual parameters. It is well-known that the apparent singularities and their dual parameters are coordinates on Zariski open sets of these moduli spaces. Historically, Okamoto [11] described Hamiltonians of the Garnier systems by the apparent singularities and their duals, for the Garnier systems are obtained by the isomonodromic deformations of rank 2 parabolic connections on P1. The apparent singularities and their duals are introduced as coordinates for a Zariski open set of Okamoto's space of initial conditions, which are nothing but the moduli space of rank 2 parabolic connections on P1. Arinkin -- Lysenko [1] and Oblezin [12] studied the moduli space of rank 2 parabolic connections on P1 more systematically, and they also introduced the apparent singularities and their duals as coordinates for Zariski open sets of moduli spaces. Oblezin also showed that the moduli space of rank 2 parabolic connections on P1 is birational to the Hilbert scheme of points on the blowing up of the total space of a certain line bundle on P1. Dubrovin and Mazzoco discussed the apparent singularities and their duals for higher rank cases on P1 in detail [4]. For the moduli spaces of parabolic connections on arbitrary genus curves, Inaba -- Iwasaki -- Saito [6] and Inaba [5] established the existence of good moduli spaces of stable parabolic connections, and it is interesting to describe their geometric structures. Saito and Szabo are developing a systematic treatment of apparent singularities and their duals for general parabolic connections on higher genus curves [13]. One can show the similar geometric description of the moduli spaces of parabolic Higgs bundles. The main purpose of this paper is to give more explicit description of the total space of the moduli spaces of rank 2 parabolic connections on P1. For the purpose, we need treat the following particular cases. The first case is that the apparent singularities approaches to the regular singularities of Higgs fields or connections (This case is already treated in [12, Section 3.7]). The second case is that the apparent singularities have multiplicities. The third case is that the type of the underlying bundle is jumping. The jumping phenomenon happens to n-regular singularities cases where n ≥ 5. For the first and second cases, we can give an explicit description of families for the n-point regular singularities case. For the third case, we give an explicit description of jumping families parameterized by the apparent singularities and their duals for the 5-point regular singularities case. Oblezin [12] considered the stratification of the moduli spaces of rank 2 parabolic connections on P1 associated to the bundle type of underlying bundles, and gave geometric descriptions of each strata of the moduli spaces, separately. On the other hand, in this paper, we try to give a global geometric description of the moduli spaces including the jumping 2010 Mathematics Subject Classification. Primary 14D20, Secondary 34M55 32G34. Key words and phrases. Parabolic connection, Parabolic Higgs bundle, Apparent singularity. This research was partly supported by JSPS Grant-in-Aid for Scientific Research (S)24224001, challenging Exploratory Research 15K13427. 1 2 A. KOMYO AND M.-H. SAITO phenomena of bundle type. As the result, we obtain a global description of the moduli spaces for the 5-point regular singularities case, and give an explicit description of universal families of Higgs bundles and connections. i(ξ+ i + ξ− Fix points t1, . . . , tn ∈ P1 (ti (cid:54)= tj), and set D = t1 + ··· + tn. We consider pairs (E,∇) where E is a rank 2 vector bundle on P1 and ∇ : E → E ⊗ Ω1P1(D) a connection having simple poles supported on D. (cid:80) i } of ∇, i = 1, . . . , n; they satisfy Fuchs relation At each pole, we have two residual eigenvalues {ξ+ i ) + d = 0 where d = deg(E). Moreover, we introduce parabolic structures l = {li}1≤i≤n such that li is a one dimensional subspace of Eti which corresponds to an eigenspace of the residue of ∇ at ti (cid:54)= ξ− with the eigenvalue ξ+ i , the parabolic structure l is determined uniquely by the connection (E,∇). Fixing a spectral data ξ = (ξ± i ) with integral sum −d and introducing the weight w, t,ξ(gl2) of w-stable ξ-parabolic connections (E,∇, l) by Geometric we can construct the moduli space M w Invariant Theory and the moduli space M w t,ξ(gl2) turns to be a smooth irreducible quasi-projective variety of dimension 2n − 6 for generic weight w (see [6]). Note that, when i . Note that when ξ+ i , ξ− i (1) n(cid:88) i /∈ Z ξ
i i=1 for any (
i),
i ∈ {+,−}, every parabolic connection (E,∇, l) is irreducible, hence stable. Therefore the moduli space M w t,ξ(gl2) does not depend on the choice of the weight w in such cases. It is known that the moduli spaces coincide with the spaces of initial conditions for Garnier systems, and the case n = 4 corresponding to the Pinlev´e VI equation, for such differential equations are nothing but isomonodromic deformations for linear connections. Next, we fix ξ = (ξ± i ) = 0. In the same way as above, we can define w-stable ξ-parabolic Higgs bundle (E, Φ, l). Here E is a rank 2 vector bundle on P1, Φ : E → E ⊗ Ω1P1 (D) is an OP1-morphism, and l is the parabolic structure. At each point ti, residual eigenvalues of Φ are {ξ+ H,t,ξ(gl2) be the moduli space of w-stable ξ-parabolic Higgs bundles. i )1≤i≤n where(cid:80) i } of Φ. Let M w i + ξ− i , ξ− i(ξ+ By suitable transformations, we may assume that d = deg(E) = −1 and ξ can be normalized as follows. For connection cases, we can putξ+ i = νi i = −νi ξ− n = 1 − νn, ξ− (i = 1, . . . , n) (i = 1, . . . , n − 1) i = νi and ξ− i = −νi (i = 1, . . . , n), for some ν = (ν1, . . . , νn) ∈ Cn. Let and for Higgs cases, we can put ξ+ M and MH be the moduli space of ν-sl2-parabolic connections and the moduli space of ν-sl2-parabolic Higgs bundles, respectively. By these normalizations, we have natural isomorphisms M ∼= M w t,ξ(gl2) H,t,ξ(gl2). (Note that the moduli space M is noting but the moduli space of (modified) and MH (ν1, . . . , νn)-bundles on P1 treated in [1] and [12].) ∼= M w For the moduli space MH , we obtain the following results. First, we consider the Zariski open set H which is the locus where the type of the underlying bundle is O ⊕ O(−1). Let K(cid:48) M 0 n be some Zariski open set of some blowing-up of the Hirzebruch surface of degree n − 2. (See Figure 1). By the explicit computation of the apparent singularities and their dual parameters, we have the following Theorem 1.1 (Theorem 3.1). By the apparent singularities and dual parameters, we have a map H −→ Hilbn−3(K(cid:48) M 0 n), and this map is injective. Moreover, we can give an explicit description of the universal family ( E(0), Φ(0)) → H × P1. M 0 (cid:99)MH be the moduli space of ν-sl2-parabolic Higgs bundles with a cyclic vector σ ∈ H 0(P1, E). The moduli Suppose n = 5. We consider the total moduli space MH , which also includes the jumping locus. The type of the underlying bundle of members are O⊕O(−1) (generic) or O(1)⊕O(−2) (jumping locus). Let space (cid:99)MH is the blowing-up of MH along the jumping locus. We take some blowing-up of Hilb2(K(cid:48) denoted by (cid:103)Hilb EXPLICIT DESCRIPTION OF JUMPING PHENOMENA (K(cid:48) (K(cid:48) 2 2 5). 3 5), 5). Then we have the map (cid:99)MH → (cid:103)Hilb (cid:99)MH −→ (cid:103)Hilb 2 Theorem 1.2 (Theorem 3.2 and Section 4.1). Suppose that n = 5. The map (K(cid:48) 5) For the moduli space M of connections, which is isomorphic to M w is injective. Moreover, we can give an explicit description of the universal family ( E, Φ, σ) → (cid:99)MH × P1. is O ⊕ O(−1). Let (cid:101)K(cid:48) t,ξ(gl2), we have the following results. First, we consider the Zariski open set M 0 which is the locus where the type of the underlying bundle n be some Zariski open set of some blowing-up of the Hirzebruch surface of degree n − 2. By the same argument as in the Higgs case, we have the following Theorem 1.3 (Theorem 5.2). By the apparent singularities and dual parameters, we have a map (2) and this map is injective. Moreover, we can give an explicit description of the universal family ( E(0), ∇(0)) → M 0 × P1. n), M 0 −→ Hilbn−3((cid:101)K(cid:48) Parts of Theorem 1.1 and Theorem 1.3 are already contained in [1], [7], [12], and [16]. For n = 4, H (resp. M 0) to a (n − 3)-th n), which is an isomorphism on a certain open set. For n = 5, the the results are discussed in [1] and [7]. Oblezin [12] gives a map from M 0 symmetric product of K(cid:48) injectivities are discussed in [16]. n (resp. (cid:101)K(cid:48) Suppose n = 5. We consider the moduli space M , which includes the jumping locus. The type of of M along the jumping locus. Let C∞ be the ∞-section of the Hirzebruch surface of degree n − 2. the underlying bundle of members of the jumping locus is O(1) ⊕ O(−2). Let (cid:99)M be the moduli space of ν-sl2-parabolic connections with a cyclic vector σ ∈ H 0(P1, E). The moduli space (cid:99)M is the blowing-up Theorem 1.4 (Theorem 5.5). Let φ : (cid:99)M (cid:57)(cid:57)(cid:75) Hilb2((cid:101)K(cid:48) apparent singularities and the dual parameters. By taking some sequence of blowing-ups (cid:103)Hilb ((cid:101)K(cid:48) n ∪ C∞) be the birational map constructed by the n ∪ C∞), we have the injective map φ : (cid:99)M → (cid:103)Hilb n ∪ C∞) for φ. The moduli space (cid:99)M is ((cid:101)K(cid:48) Hilb2((cid:101)K(cid:48) n∪C∞) → biregular to its image Image φ((cid:99)M ) ⊂ (cid:103)Hilb ((cid:101)K(cid:48) n ∪ C∞). 2 2 2 The organization of this paper is as follows. In Section 2, we introduce definitions and notations which are necessary in this paper. In Section 3, 3.1, we consider the moduli spaces of ν-sl2-parabolic Higgs bundles with bundle type O ⊕ O(−1). We show Theorem 1.1 (Theorem 3.1) by explicit calculations of In 3.2, we consider the moduli spaces of ν-sl2-parabolic Higgs bundles with a apparent singularities. cyclic vector for n = 5. We show the injectivity of the map in Theorem 1.2 (Theorem 3.2) by explicit calculations of apparent singularities and spectral curves. In Section 4, we construct an explicit jumping families of ν-sl2-parabolic Higgs bundles by the lower and upper modifications. In particular, in 4.1, we give an explicit description of the universal family of (cid:99)MH In Section 5, we consider the moduli spaces of ν-sl2-parabolic connections. In 5.1, we show Theorem 1.3 (Theorem 5.2) by the same way as in the Higgs case. In 5.2, we construct an explicit jumping family of connections for n = 5, and in 5.3, we analyze the behavior of the apparent singularities and their duals when the parameters of the jumping family approach to the jumping locus. Finally, we obtain Theorem 1.4 (Theorem 5.5). 2. Preliminaries In this section, first, we define ν-sl2-parabolic connections and ν-sl2-parabolic Higgs bundles, and recall the well-known facts of the connections and Higgs bundles. In 2.2, we describe some blowing-ups of the Hirzebruch surface Σn−2, which are target spaces of the map defined by the apparent singularities and their duals. In 2.3, we discuss descriptions of Higgs fields. Since we consider Higgs bundles on P1, 4 A. KOMYO AND M.-H. SAITO the underlying vector bundles split into the direct sum of line bundles. Then we can describe the Higgs fields explicitly. By the automorphisms of vector bundles, we can normalize the Higgs fields to reduce the number of parameters. In 2.4, we discuss the apparent singularities and the dual parameters of Higgs bundles, which give a map from the moduli space of Higgs bundles (with a cyclic vector) to the symmetric product of the total space of Ω1P1 (D). In 2.5, we discuss the transformations called the lower modification and the upper modification, and we use the transformations for a construction of a universal family of the moduli space of Higgs bundles (with a cyclic vector). The contents of 2.1, 2.2, and 2.5 basically follow the expositions from [1] and [12]. 2.1. sl2-connections and sl2-Higgs bundles. We introduce sl2-parabolic connections and sl2-parabolic Higgs bundles, and we consider relations between the moduli space M w t,ξ(gl2) and these moduli spaces. Fix complex numbers ν1, . . . , νn ∈ C. Suppose that ν1 ··· νn (cid:54)= 0 and n(cid:88)
iνi /∈ Z (3) for any (
i),
i ∈ {1,−1}. Definition 2.1. A ν-sl2-parabolic connection on P1 is a triplet (E,∇, ϕ) such that i=1 (3) ϕ : (cid:86)2 E ∼= OP1(−1) is a horizontal isomorphism, (1) E is a rank 2 vector bundle on P1, (2) ∇ : E → E ⊗ Ω1P1 (D) is a connection, (4) the residue resti(∇) of the connection ∇ at ti has eigenvalues ν± i , 1 ≤ i ≤ n. Here we put ν± i := ±νi (i = 1, . . . , n − 1), ν+ n := νn, ν− n := 1 − νn. Denote by M the moduli stack of ν-sl2-parabolic connections on P1, and by M its coarse moduli space. Let ∇(cid:48) : OP1 → OP1 ⊗ Ω1P1(D) be the connection defined by (cid:32) n−1(cid:88) i=1 1 2 f (cid:55)−→ df + n−1(cid:88) i=1 dz z − ti + 1 2 i + ξ− (ξ+ i ) dz z − tn f. (−ξ+ i − ξ− i ) (cid:54)= ξ− (cid:33) Suppose that the condition (1) holds and ξ+ i n for i = 1, . . . , n. Then we have an isomorphism t,ξ(gl2) −→ M = Mν M w (E,∇, l) (cid:55)−→ ((E,∇) ⊗ (OP1,∇(cid:48)), ϕ), i )/2 (i = 1, . . . , n − 1), νn = ξ+ n +(cid:80)n−1 i − ξ− where ν = (ν1, . . . , νn), νi = (ξ+ Definition 2.2. A ν-sl2-parabolic Higgs bundle on P1 is a triplet (E, Φ, ϕ) such that i − ξ− i=1 (ξ+ i )/2. (3) ϕ : (cid:86)2 E ∼= OP1(−1) is an isomorphism and tr(Φ) = 0, (1) E is a rank 2 vector bundle on P1, (2) Φ : E → E ⊗ Ω1P1 (D) is an OP1 -morphism, (4) the residue resti(Φ) of the connection Φ at ti has eigenvalues ±νi, 1 ≤ i ≤ n. Denote by MH the moduli stack of ν-sl2-parabolic Higgs bundles on P1, and by MH its coarse moduli space. We have a stratification of MH as follows. By the irreducibility of (E, Φ, ϕ) ∈ MH , we have the following Proposition 2.3. For (E, Φ, ϕ) ∈ MH , we have E ∼= O(k) ⊕ O(−k − 1) where 0 ≤ k ≤ (cid:20) n − 3 (cid:21) . 2 Let M k H be the subvariety of MH where E ∼= O(k) ⊕ O(−k − 1). Then EXPLICIT DESCRIPTION OF JUMPING PHENOMENA 5 MH = M 0 H ∪ ··· ∪ M [(n−3)/2] H . Note that the stratum M 0 H is a Zariski open dense of MH . Moreover, we introduce ν-sl2-parabolic connection on P1 with a cyclic vector and ν-sl2-parabolic Higgs bundle on P1 with a cyclic vector. Definition 2.4. A ν-sl2-parabolic connection on P1 with a cyclic vector is a tuple (E,∇, ϕ, [σ]) such that (1) E is a rank 2 vector bundle on P1, (3) ϕ : (cid:86)2 E ∼= OP1(−1) is a horizontal isomorphism, (2) ∇ : E → E ⊗ Ω1P1 (D) is a connection, (4) the residue resti(∇) of the connection ∇ at ti has eigenvalues ν± i , 1 ≤ i ≤ n. (5) [σ] ⊂ H 0(P1, E) is a 1-dimensional subspace generated by a nonzero section σ ∈ H 0(P1, E). Denote by (cid:99)M the moduli stack of ν-sl2-parabolic connections on P1 with a cyclic vector, and by (cid:99)M its coarse moduli space. Since dim H 0(P1,O ⊕ O(−1)) = 1, M 0 is contained in (cid:99)M . Definition 2.5. A ν-sl2-parabolic Higgs bundle on P1 with a cyclic vector is a tuple (E, Φ, ϕ, [σ]) such that (1) E is a rank 2 vector bundle on P1, (3) ϕ : (cid:86)2 E ∼= OP1(−1) is an isomorphism and tr(Φ) = 0, (2) Φ : E → E ⊗ Ω1P1 (D) is an OP1 -morphism, (4) the residue resti(∇) of the connection ∇ at ti has eigenvalues ±νi, 1 ≤ i ≤ n. (5) [σ] ⊂ H 0(P1, E) is a 1-dimensional subspace generated by a nonzero section σ ∈ H 0(P1, E). Denote by (cid:99)MH the moduli stack of ν-sl2-parabolic Higgs bundles on P1 with a cyclic vector, and by (cid:99)MH its coarse moduli space. For n = 5, the moduli space (cid:99)MH is the blowing-up of MH along M 1 H . 2.2. Hirzebruch surfaces and the blowing-ups. For description of the moduli spaces M and MH , we introduce some blowing-ups of the Hirzebruch surface Σn−2. Put L := Ω1P1 (D). Let L be the total space of the line bundle L. Note that L = Σn−2 \ C∞ where C∞ is the infinity section (C∞)2 = −(n− 2). First, we construct a blowing-up of the Hirzebruch surface Σn−2 corresponding to MH . Let π : L → P1 := −νi for i = 1, . . . , n, ∼=−→ C be the residue map. Put ν+ := νi, ν− be the projection and let τi : π−1(ti) and ν± := τ−1 (ν± i ). Set i i K(cid:48) n := L is the blowing-up of L at ν± where Blν Fi i = 1, . . . , n. We denote by Kn the image of K(cid:48) ± i i i L(cid:17) \ ((cid:101)F1 ∪ ··· ∪ (cid:101)Fn) (cid:16) for i = 1, . . . , n, and (cid:101)Fi are the proper pre-images of the fiber Blν ± i i n by the projection K(cid:48) n → L (see Figure 1). Second, we construct a blowing-up of the Hirzebruch surface Σn−2 corresponding to M . Let π : L → P1 be the projection and let τi : π−1(ti) ∼=−→ C be the residue map. Set (cid:101)K(cid:48) L(cid:17) \ ((cid:101)F1 ∪ ··· ∪ (cid:101)Fn) (cid:16) n := := νi, ν− ± i Blν := −νi for i = 1, . . . , n − 1 and ν+ where ν± denote by (cid:101)Kn the image of (cid:101)K(cid:48) := τ−1 i ). Here, ν+ n by the projection (cid:101)K(cid:48) (ν± i i i i n → L. n := νn, ν− n := 1 − νn. We 6 A. KOMYO AND M.-H. SAITO Figure 1. Kn and K(cid:48) n (cid:18) zk 0 (cid:19) 2.3. Description of Higgs fields. Put U0 := P1\{∞}, U∞ := P1\{0}. Let z and w be the coordinates on U0 and U∞, respectively. Put (4) ωz := z(z − 1)(z − x1)··· (z − xn−3) and Rk := dz 0 1 zk+1 , 0 ≤ k ≤ We consider an explicit description of the Higgs field of (E, Φ, ϕ) ∈ MH . Suppose that E ∼= O(k) ⊕ O(−k − 1) where 0 ≤ k ≤ [(n − 3)/2]. We can describe the Higgs field Φ as follows: f (n−2) f (n+2k−1) (z) f (n−2k−3) (z) −f (n−2) (z) on U0 on U∞, z ⊗ ωz Ak R−1 k (Ak z ⊗ ωz)Rk Ak z := (cid:40) (cid:32) Φ = (z) (5) 11 21 12 11 where f (l) ij (z) is a polynomial in z of degree at most l. By the irreducibility, we have f (n−2k−3) (z) (cid:54)= 0. We consider automorphisms of the vector bundle E ∼= O(k) ⊕ O(−k − 1). Any element P ∈ 21 HomOP1 (E, E) ∼= H 0(P1,O ⊕ O(2k + 1) ⊕ O) is described as follows: (cid:20) n − 3 (cid:21) . 2 (cid:33) (cid:18) s (cid:19) PU0 = p(2k+1)(z) 0 t on U0, PU∞ = (cid:18) s w2k+1p(2k+1)(1/w) (cid:19) 0 t on U∞ where s, t ∈ C and p(2k+1)(z) is a polynomial in z of degree at most 2k + 1. If st (cid:54)= 0, then P ∈ Aut(E). We take P ∈ Aut(E). Then we have (6) (7) the (1, 1)-entry of P −1 U0 AzPU0 = the (2, 1)-entry of P −1 U0 AzPU0 = tf (n−2) 11 (z) − p(2k+1)(z)f (n−2k−3) 21 sf (n−2k−3) 21 t t (z) . We consider simple descriptions of Higgs fields by the automorphisms of E. First, we consider the (z) on P1 where (2, 1)-entry. Let {[s1 : 1], . . . , [si : 1], [1 : qi+1], . . . , [1 : qn−2k−3]} be the zeros of f (n−2k−3) 0 ≤ i ≤ n − 2k − 3. By the automorphisms of E, we can put 21 f (n−2k−3) 21 (z) := (s1z − 1)··· (siz − 1)(z − qi+1)··· (z − qn−2k−3). Second, we consider the (1, 1)-entry. We assume that the coefficient of zl in the polynomial f (n−2k−3) is nonzero for some l (0 ≤ l ≤ n − 2k − 3). By the automorphisms of E, we can put 21 (z) (8) f (n−2) 11 (z) := an−2zn−2 + ··· + al+2k+2zl+2k+2 + al−1zl−1 + . . . + a0. EXPLICIT DESCRIPTION OF JUMPING PHENOMENA 7 In particular, if i = 0, that is, f (n−2k−3) in the polynomial f (n−2k−3) (9) f (n−2) 21 21 11 (z) is nonzero. In this case, we can put (z) := an−2k−4zn−2k−4 + . . . + a0. (z) := (z − q1)··· (z − qn−2k−3), then the coefficient of zn−2k−3 2.4. Apparent singularities and the dual parameters. We recall the apparent singularities and the dual parameters introduced by Saito-Szabo [13]. Let (E, Φ, ϕ) ∈ MH . apparent singularities and the dual parameters coincide with the geometric Darboux coordinates due to Oblezin [12, Section 3], which gives a geometric interpretation of the Sklyanin formulas from [15]. We fix a section σ ∈ H 0(P1, E). For the section σ, we define the following composition If E ∼= O ⊕ O(−1), then the The composition OP1 → (E/OP1) ⊗ L is injective. Then we can define a subsheaf F 0 ⊂ E such that OP1 → (F 0/OP1) ⊗ L is isomorphic. By the isomorphism F 0/OP1 ∼= L−1, we have F 0 ∼= OP1 ⊕ L−1. OP1 σ−−→ E Φ−−→ E ⊗ L −→ (E/OP1 ) ⊗ L. Therefore, we have the following exact sequence. (10) 0 −→ OP1 ⊕ L−1 −→ E −→ TA −→ 0 where TA is a torsion sheaf. By the Riemann-Roch theorem, we have that the torsion sheaf TA is length n − 3. The exact sequence (10) is called a Frobenius -- Hecke sheaf originally introduced by Drinfeld (see [3] and [12, Section 3.3]). Definition 2.6. For (E, Φ, ϕ) ∈ MH and a nonzero section σ ∈ H 0(P1, E), we call the support of TA apparent singular points of a ν-sl2-parabolic Higgs bundle with a cyclic vector (E, Φ, ϕ, [σ]). Next, we define dual parameters of (E, Φ, ϕ, [σ]). Let Cs be the spectral curve of (E, Φ, ϕ). Let G be a torsion free sheaf of rank 1 on Cs corresponding to (E, Φ, ϕ), which satisfies E = π∗G. Since H 0(Cs, G) ∼= H 0(P1, E), for a section σ ∈ H 0(P1, E), we have the short exact sequence 0 −→ OCs σ−−→ G −→ TB −→ 0 where TB is a torsion sheaf on Cs of length n − 3. We take the direct image of the short exact sequence. Since π∗(OCs) = OP1 ⊕ L−1 and π∗G = E, we have 0 −→ OP1 ⊕ L−1 π∗σ−−−−→ E −→ π∗(TB) −→ 0. We may show that this short exact sequence coincides with the short exact sequence (10). In particular, we have π∗(TB) = TA, whose support is the apparent singularities of (E, Φ, ϕ, [σ]). Set Supp(TB) = {p1, . . . , pn−3}, where pi is a point on Cs. Put pi = (qi, pi) where qi = π(pi), which is an apparent singularity, and pi ∈ Lqi. Definition 2.7. For (E, Φ, ϕ) ∈ MH and a nonzero section σ ∈ H 0(P1, E), we call {p1, . . . , pn−3} dual parameters of a ν-sl2-parabolic Higgs bundle with a cyclic vector (E, Φ, ϕ, [σ]). We consider the apparent singularities and the dual parameters of (E, Φ, ϕ, [σ]) where E ∼= O⊕O(−1). In this case, the Higgs field Φ is described as follows: Φ = z ⊗ ωz A0 R−1 0 (A0 z ⊗ ωz)R0 on U0 on U∞ where A0 z := f (n−2) f (n−3) 11 21 f (n−1) (z) (z) (z) −f (n−2) 12 11 (z) The apparent singularities are the zeros of f (n−3) eters are {p1, . . . , pn−3} where we put pi := f (n−2) 21 (z) on P1, denoted by {q1, . . . , qn−3}. The dual param- (qi). Then, for (O ⊕ O(−1), Φ, ϕ), we have 11 {(q1, p1), . . . , (qn−3, pn−3)} ∈ Symn−3(Kn). Next, we consider the case E ∼= O(k) ⊕ O(−k − 1) where k > 0. In this case, the Higgs field Φ is (cid:40) (cid:40) described as follows: z ⊗ ωz Ak R−1 0 (Ak Φ = z ⊗ ωz)R0 on U0 on U∞ where Ak z := f (n−2) f (n+2k−1) (z) f (n−2k−3) (z) −f (n−2) (z) 11 12 (z) 21 11 (cid:32) (cid:32) (cid:33) . (cid:33) . 8 A. KOMYO AND M.-H. SAITO Let σ ∈ H 0(P1, E) ∼= H 0(P1,O(k)) be a section of E, and let {q1, . . . , qk} ∈ Symk(P1) be the zeros of the section σ. We denote by {q2k+1, . . . , qn−3} ∈ Symn−2k−3(P1) the zeros of f (n−2k−3) (z) on P1. The apparent singularities of (E, Φ, ϕ, [σ]) are the following 21 {2q1, . . . , 2qk, q2k+1, . . . , qn−3} ∈ Symn−3(P1). We compute the dual parameters of (E, Φ, ϕ, [σ]). We take σ ∈ H 0(Cs, G) ∼= H 0(P1, π∗G) corresponding to σ ∈ H 0(P1, E). The section σ ∈ H 0(Cs, G) has the following zeros {(q1, p1), (q1,−p1), . . . , (qk, pk), (qk,−pk)} where det(piI−Φqi) = 0 for i = 1, . . . , k. The dual parameters are {p1,−p1, . . . , pk,−pk, p2k+1, . . . , pn−3} where we put pi := f (n−2) (qi) for i = 2k + 1, . . . , n − 3. Then, for (E, Φ, ϕ, [σ]), we have 11 {(q1, p1), (q1,−p1), . . . , (qk, pk), (qk,−pk), (q2k+1, p2k+1), . . . , (qn−3, pn−3)} ∈ Symn−3(Kn). 2.5. Lower and upper modifications. In this subsection, following [12, Section 2], we describe the lower and the upper modifications. Let E be an algebraic vector bundle on P1 of rank 2 and of degree d. Fix a point a ∈ P1. Let l ⊂ Ea be a 1-dimensional subspace. Definition 2.8. We call (a, l)low(E) := {s ∈ E s(a) ∈ l}, (a, l)up(E) := (a, l)low(E) ⊗ O(a) the lower and the upper modifications of E, respectively. The lower and the upper modifications provide the following exact sequences 0 −→ (a, l)low(E) −→ E −→ Ea/l −→ 0, 0 −→ E −→ (a, l)up(E) −→ l ⊗ O(a) −→ 0, respectively. In other words, we change our bundle rescaling the basis of sections in the neighborhood of a point a as follows. Given a local decomposition V = l ⊕ l(cid:48) of E ∼= V ⊗ O, we put the local basis {s1(z), s2(z)} with l ⊗ O ∼= (cid:104)s1(z)(cid:105) and l(cid:48) ⊗ O ∼= (cid:104)s2(z)(cid:105). Then the basis of the lower modification (x, l)low of the bundle is generated by the sections {s1(z), (z − x)s2(z)}, and of the upper one (x, l)up by {(z − x)−1s1(z), s2(z)}. Consequently, in the punctured neighborhood, we may represent the action of the modifications by the following gluing matrices (cid:18) 1 0 (cid:19) (a, l)low = 0 (z − a) , (a, l)up = (cid:18) (z − a)−1 0 (cid:19) . 0 1 3. Geometric description of the moduli spaces Suppose that ν satisfies the condition (3) and ν1 ··· νn (cid:54)= 0. We put (t1, t2, tn) := (0, 1,∞), (ν± 1 , . . . , ν± n ) := (±ν1,±ν2, . . . ,±νn), and νi := νi(ti − t1)··· (ti − ti−1)(ti − ti+1)··· (ti − tn−1) for i = 1, . . . , n − 1. H for n ≥ 4. Then First, we consider the apparent singularities and the dual parameters of members of M 0 and the dual parameters of members of (cid:99)MH . Then we have the first assertion of Theorem 1.2 (Theorem we have Theorem 1.1 (Theorem 3.1). Second, we assume n = 5. We consider the apparent singularities 3.2). EXPLICIT DESCRIPTION OF JUMPING PHENOMENA 9 H for n ≥ 4. Let (O ⊕ O(−1), Φ, ϕ) ∈ M 0 3.1. Geometric description of M 0 n be the Zariski open set of the blowing-up of Hirzebruch surface of degree n − 2 defined in 2.2, and Kn be the contraction K(cid:48) n → Kn. Since dim H 0(P1,O ⊕ O(−1)) = 1, sections are determined uniquely up to constant. Then the apparent singularities and the dual parameters are determined by (E, Φ, ϕ). Let {(q1, p1), . . . , (qn−3, pn−3)} ∈ Symn−3(Kn) be the apparent singularities and the dual parameters of (E, Φ, ϕ). We consider the map H , and K(cid:48) (11) H −→ Symn−3(Kn) M 0 (E, Φ, ϕ) (cid:55)−→ {(q1, p1), . . . , (qn−3, pn−3)}, which is essentially constructed in [12, Section 3]. Since this map is not injective, we consider the composite of the Hilbert-Chow morphism and the blowing-up Hilbn−3(K(cid:48) n) −→ Symn−3(K(cid:48) n) −→ Symn−3(Kn). Then we have the following Theorem 3.1. The map (11) is extended to M 0 H −→ Hilbn−3(K(cid:48) (12) The map is injective. Moreover, we can give an explicit description of the universal family ( E(0), Φ(0)) → H × P1. M 0 n). The image of the map M 0 Symn−3(K(cid:48) n) is H → Hilbn−3(K(cid:48) n) is described as follows. The image of the map M 0 H → {x = {n1 p1, . . . , nr pr} ∈ Symn−3(K(cid:48) n) π(pi) (cid:54)= π(pj) for i (cid:54)= j} where nj are integers such that n1 + ··· + nr = n − 3, nj ≥ 1 (j = 1, . . . , r) and π is the projection K(cid:48) n → P1. For x = {n1 p1, . . . , nr pr}, let M 0 n). Then ∼= Cn1−1 × ··· × Cnr−1 where Cn1−1 × ··· × Cnr−1 is an affine open set of the fiber of x under the M 0 H Hilbert-Chow morphism. H,x be the fiber of x under M 0 H → Symn−3(K(cid:48) The explicit description of the family is as follows. For simplicity, we assume that {q1, . . . , qn−3} ⊂ P1 \ {∞}. If the apparent singularities are distinct, then the explicit description is the following (z − q1)··· (z − qn−3) −(an−4zn−4 + ··· + a0) 12 f (n−1) (z) (cid:19) (cid:40) (13) where on U0 on U∞ z ⊗ ωz)R0 z ⊗ ωz A0 R−1 0 (A0 (cid:18) an−4zn−4 + ··· + a0 (cid:18)1 defined by the relations lij = 0 for i < j, l11 = 1 and lij = (cid:81)i ui+1,jljkpk, R0 = n−3(cid:88) n−3(cid:88) , A0 ai = z = (14) k=1 j=1 0 (cid:19) 0 1 z . Here (q1, p1), . . . , (qn−3, pn−3) are pairs of the apparent singularities and their duals, the element lij is k=1,k(cid:54)=j 1/(qj − qk) otherwise, and the element uij is defined by the relations uii = 1, ui1 = 0 and uij = ui−1,j−1 − ui,j−1qj−1 otherwise, that is, the matrix (uijljk)ik is the inverse of the Vandermonde matrix. Here we set u0j = 0. We omit the description of f (n−1) (z), which is a polynomial in z of degree n − 1, since the description is lengthened and is not necessary below. Next, we consider the case where the apparent singularities and their duals have multiplicities. Let nj (j = 1, . . . , r) be integers such that n1 + ··· + nr = n − 3, nj ≥ 1. For each k=1 nk), that is, the pair of the apparent singularities and their duals is {n1(x1, y1), . . . , nr(xr, yr)}. In this case, for each j (j = 1, . . . , r) we substitute k=1 nk + 1,··· ,(cid:80)j j (j = 1, . . . , r), we assume that qi = xj ∈ P1 \ {∞} and pi = yj ∈ C (i =(cid:80)j−1 j(cid:88) j−1(cid:88) 12 0 + λ(j) 1 (qi − xj) + ··· + λ(j) nj−2(qi − xj)nj−2), pi = yj + (qi − xj)(λ(j) nk + 1,··· , nk, i = for the coefficient (14). Here λ(j) the Hilbert-Chow morphism. Then we can define the coefficients ai for this case. nj−2 are coordinates of the affine open set Cnj−1 of the fiber of 0 , . . . , λ(j) k=1 k=1 10 Proof of Theorem 3.1. Put and M 00 H := M 000 H := (E, Φ, ϕ) ∈ M 0 H (E, Φ, ϕ) ∈ M 0 H (cid:26) (cid:26) (cid:26) Image(M 000 H ) := Let (15) A. KOMYO AND M.-H. SAITO qi (cid:54)= qj (i (cid:54)= j) (cid:12)(cid:12)(cid:12)(cid:12) {q1, . . . , qn−3} : the apparent singularities of (E, Φ, ϕ) (cid:12)(cid:12)(cid:12)(cid:12) {q1, . . . , qn−3} : the apparent singularities of (E, Φ, ϕ) (cid:27) qi (cid:54)= qj (i (cid:54)= j) and tk /∈ {q1, . . . , qn−3} for any k. (cid:12)(cid:12)(cid:12)(cid:12) qi (cid:54)= qj (i (cid:54)= j), qi /∈ {t1, . . . , tn} (i = 1, . . . , n − 3) (cid:27) (cid:27) , . Step 1. In this step, we show that the restriction M 000 H → Symn−3(Kn) is injective. The image of this restriction is the following {(q1, p1), . . . , (qn−3, pn−3)} ⊂ Symn−3(Kn). {([s1 : 1], u1), . . . , ([si : 1], ui), ([1 : qi+1], pi+1), . . . , ([1 : qn−3], qn−3)} H ) where 0 ≤ i ≤ n − 3. We show that the entries f (n−3) (z), (z) of the description (5) are determined by the element (15) up to automorphisms of O ⊕ (z). By the definition of the apparent singularities and an be an element of Image(M 000 and f (n−1) O(−1). First, we consider the entry f (n−3) automorphism of O ⊕ O(−1), we can put (z), f (n−2) 21 12 21 11 f (n−3) 21 (z) = (s1z − 1)··· (siz − 1)(z − qi+1)··· (z − qn−3). Second, we consider the entry f (n−2) (z). Since s1 ··· si (cid:54)= 0, the coefficient zn−3 in f (n−3) nonzero. Then, by the automorphism of O ⊕ O(−1), we can put f (n−2) the description (9). By the definition of the dual parameters, we have that uj = sn−2 1 ≤ j ≤ i and pi = f (n−2) (qj) for i + 1 ≤ j ≤ n − 3. Then we have the following system 11 j 11 (z) is (z) = an−4zn−4 + ··· + a0 as in (1/sj) for f (n−2) 21 11 11  =   u1 ... ui pi+1 ... pn−3 1 ... 1 qn−4 ... qn−4 n−3 i+1 ··· ··· ··· ··· 1 sn−3 ... sn−3 i qi+1 ... qn−3 1 sn−4 ... sn−4 1 ... 1 i    . an−4 ... a1 a0 We can determine the coefficients an−4, . . . a0 by an element (15). Third, we consider the entry f (n−1) 12 solve the equations (z). We put f (n−1) 12 (z) := bn−1zn−1 + bn−2zn−2 + ··· + b0. We (16) Then we have f (n−1) 12 (ti) = det (resti Φ) = −ν2 i , i = 1, . . . , n. (for i = 1, . . . , n − 1), and b0 = −(wn−2f (n−2) 11 (1/w))2w=0 + ν2 n (wn−3f (n−3) 21 (1/w))w=0 . −f (n−2) 11 (ti)2 + ν2 i f (n−3) 21 (ti) Since f (n−2) As the result, we obtain that the map M 000 and f (n−3) 21 11 Step 2. In this step, we extend the map M 000 that the extended map M 00 H → Symn−3(K(cid:48) are determined by the element (15), we can determine the coefficients b0, . . . , bn−1. H → Symn−3(Kn) is injective. H → Symn−3(Kn) to M 00 H → Symn−3(K(cid:48) n) is injective. For the element (15), we can put n), and we show f (n−2) 11 (z) := pj + (z − qj) f11(z), and f (n−3) 21 (z) := (s1z − 1)··· (siz − 1)(z − qi+1)··· (z − qn−3) EXPLICIT DESCRIPTION OF JUMPING PHENOMENA 11 where f11(z) is a polynomial of degree at most n − 3 in z. The polynomial f11(z) is determined by the element (15) up to automorphisms of O ⊕ O(−1). By the condition (16), we have (17) f (n−1) 12 (ti) = = −(pj + (ti − qj) f11(ti))2 + ν2 (ti − q1)··· (ti − qn−3) −(ti − qj)( f11(ti)2 + 2pj i (ti − q1)··· (ti − qn−3) f11(ti)) − p2 j + ν2 i . We consider the blowing-up K(cid:48) (ti,
νi) ∈ Kn as pj −
νi = vi,
we have n → Kn. Let
∈ {+,−}. We define the blowing-up parameters vi,
j (qj − ti). We substitute pj = vi,
f11(ti)) − (vi,
−(ti − qj)( f11(ti)2 + 2pj at j (qj − ti) +
νi for the formula (17). Then j (qj − ti))2 −
2νivi,
j (qj − ti) j (18) f (n−1) 12 (ti) = = (ti − q1)··· (ti − qn−3) f11(ti)) − (vi,
j )2(qj − ti) +
2νivi,
−( f11(ti)2 + 2pj (ti − q1)··· (ti − qj−1)(ti − qj+1)··· (ti − qn−3) . (ti) as qj → ti. The limit limqj→ti f (n−1) j We consider the behavior of f (n−1) (ti) is convergence, and the convergence value is determined by the apparent singularities, the dual parameters, and the blowing-up parameters vi,
(z). Then we obtain the map M 00 j . By the same argument as in Step 1, we can determine the all coefficients of f (n−1) n), and this map is injective. H → Symn−3(K(cid:48) 12 12 12 H → Symn−3(K(cid:48) n) to M 0 H → Hilbn−3(K(cid:48) n), and we show Step 3. In this step, we extend the map M 00 that the extended map M 0 H → Hilbn−3(K(cid:48) (19){([s1 : 1], u1), . . . , ([sn0 : 1], un0), ([1 : q1 n) is injective. Let be an element of Image(M 000 we can determine the entries f (n−3) be a point of Symn−3(K(cid:48) 21 n), denoted by 1], p1 1), . . . , ([1 : q1 n1 ], p1 n1 ), . . . , ([1 : qr 1], pr 1), . . . , ([1 : qr nr ], pr nr H ) where n0 + ··· + nr = n − 3, nj ≥ 1 (j = 1, . . . , r). By the element (19), (z) of the description (5) (Step 1). Let x (z), and f (n−1) (z), f (n−2) 11 12 )} x = {n0([0 : 1], y0, ai,± 0 ), n1([1 : x1], y1, ai,± 1 ), . . . , nr([1 : xr], yr, ai,± r )} ∈ Symn−3(K(cid:48) n) where yl +
νi = ai,
sj → 0 (j = 1, . . . , n0), ql Let I be an ideal contained in the fiber of x by the Hilbert-Chow morphism Hilbn−3(K(cid:48) We show that the entries f11, f12, f21 are determined by the ideal I up to automorphisms. l (xl − ti) for l = 0, . . . , r and
∈ {+,−}. We consider the behavior of the entries as k → yl (k = 1, . . . , nl and l = 1, . . . , r) where xl1 (cid:54)= xl2 (l1 (cid:54)= l2). n) → Symn−3(K(cid:48) n). k → xl, and pl On a neighborhood of I, the Hilbert scheme of points Hilbn−3(K(cid:48) n). We denote by (I n0 Hilbnr (K(cid:48) , . . . ,I nr where fxl,yl (q, p) := (p − yl) − (q − xl)(λ(l) Theorem 1.13]. x0,y0 xr,yr 0 + λ(l) ) the image of I. We put I nl 1 (q − xj) + ··· + λ(l) n) is isomorphic to Hilbn0(K(cid:48) n)×···× = ((q − xl)nl, fxl,yl (q, p)) nl−2(q − xl)nl−2) as in the proof of [10, xl,yl We consider the entry f (n−2) (z). For a neighborhood of the ideal I, we can assume that the coefficient of zn−n0−3 in f (n−3) f (n−2) 21 (z) is nonzero. Then we can normalize f (n−2) (z) := an−2zn−2 + ··· + an−n0−1zn−n0−1 + an−n0−4zn−n0−4 + . . . + a0 (z) as 11 (20) by automorphisms of O ⊕ O(−1). By the definition of dual parameters, we have the following system 11 11  =   u1 ... un0 p1 1 ... pr nr 1 ... 1 1)n1−2 (q1 ... )n−2 (qr nr (21) s1 ... sn0 1)n−3 (q1 ... )n−3 (qr nr ··· ··· ··· ··· sn0−1 1 ... sn0−1 1)n−n0−1 (q1 n0 ... )n−n0−1 (qr nr sn0+2 1 ... sn0+2 n0 1)n−n0−4 (q1 ... )n−n0−4 (qr nr ··· ··· ··· ··· 1 sn−2 ... sn−2 n0 1 ... 1    . an−2 ... an−n0−1 an−n0−4 ... a0 12 We substitute and A. KOMYO AND M.-H. SAITO uk = y0 + sk(λ(k) 0 + λ(k) 1 sk + ··· + λ(k) n0−2sn0−2 k ), k = 1, . . . , n0, pl k = yl + (ql k − xl)(λ(l) 0 + λ(l) 1 (ql k − xl) + ··· + λ(l) nl−2(ql k − xl)nl−2), k = 1, . . . , nl, where j = 1, . . . , r, in the system (21). Then we can show that the coefficients of the polynomial (20) are defined when sk = 0 and ql k = xl. Moreover, these coefficients are determined by the apparent singularities, the dual parameters, and the parameters λ(i) 0 , λ(i) 1 , . . . , λ(i) ni−2. We consider the entry f (n−1) k → xl for any k = 1, . . . , nl. By the values f (n−1) ql of f (n−1) consider the ideal I nl (t1), . . . , f (n−1) (z). Next, we consider the behavior of the value f (n−1) = ((q − ti)nl , fxl,ai,
(q, vi,
)) where (z). 12 12 12 12 12 l xl,ai,
l If xl /∈ {t1, . . . , tn}, then we have the value f (n−1) (ti) ∈ C as (tn), we can determine all coefficients (ti) as xl → ti. Let
∈ {+,−}. We 12 fxl,ai,
For the ideal I nl l (q, vi,
) := (vl,
− ai,
) − (q − xl)(λ(l) 0 + λ(l) l xl,ai,
, we can describe the dual parameter pl (cid:16) k − ti) k =
νi + vi,
k (ql pl k − xl)(λ(l) ai,
l + (ql l (xl − ti) + λ1(ql j−2 + λ(l) 0 + λ(l) k − xl) + λ2(ql =
νi + ai,
k − xl) + ··· + λ(l) =
νi + 1 (ql 1 (q − xl) + ··· + λ(l) nl−2(q − xl)nl−2). k (k = 1, . . . , nl) as follows: (cid:17) k − xl)nl−2) nl−2(ql k − ti); (ql k − xl)nl−1 + λ(l) k − xl)2 + ··· + λnl−1(ql nl−2(ql k − xl)nl j−1(xl − ti) for j = 1, . . . , ni − 1. Here we put λ(l)−1 := ai,
l . For the ideal I nl xl,ai,
l , we 11 where λj := λ(l) put (22) f (n−2) f (n−3) We substitute z = ti for f (n−2) consider the value of f (n−1) 21 12 (z) :=
νi + ai,
(z) := (z − xl)nl (z − qni+1)··· (z − qn−3). f (n−1) 12 (ti) = where α :=
2νi( f11(qj)− λ(l) xl → ti. By the values f (n−1) 12 We obtain an extended map M 0 (cid:32) 11 (ti) as follows: −f (n−2) 11 (ti)2 + ν2 i (ti − xl)nl (ti − qnl+1)··· (ti − qn−3) nl−2) and β := ( f11(qj)− λ(l) (t1), . . . , f (n−1) (cid:33) H → Hilbn−3(K(cid:48) 12 l (xl − ti) + λ1(z − xl) + λ2(z − xl)2 + ··· + λnl−1(z − xl)nl−1 + (z − xl)nl f11(z); (z). Then we have f (n−2) 11 (ti) =
νi + (ti − xl)nl ( f11(ti) − λ(l) nl−2). We α + β(ti − xl)nl (ti − qnl+1)··· (ti − qn−3) = nl−2)2. Then we have the finite value f (n−1) (tn) as xl → ti, we can determine all coefficients of f (n−1) (ti) as (z). 12 12 n) by the assignment of I nl = ((q−ti)nl , fxl,ai,
l (q, vi,
)) xl,ai,
l l 12 12 21 22 11 f (n−1) f (n−2) f (n−2) f (n−3) (ti) determines the data fxl,ai,
. This procedure is confirmed by the implication that the existence of to the matrix limxl→ti f (n−1) injective. 3.2. Geometric description of (cid:99)MH for n = 5. Suppose that n = 5. By the apparent singularities and the dual parameters of (E, Φ, ϕ, [σ]) ∈ (cid:99)MH , we have the following map (q, vi,
) from the argument above. The extended map is (cid:3) (cid:99)MH −→ Sym2(K5) to qi for i = 1, 2. There exists a stratification (cid:99)MH = (cid:99)M 0 where {q1, q2} are apparent singularities and {p1, p2} are their dual parameter where pi corresponds H is the locus such that H ∪ (cid:99)M 1 H where (cid:99)M i (E, Φ, ϕ, [σ]) (cid:55)−→ {(q1, p1), (q2, p2)} (23) (E, Φ, ϕ, [σ]) ∈ (cid:99)M i EXPLICIT DESCRIPTION OF JUMPING PHENOMENA H satisfies E ∼= O(i) ⊕ O(−i − 1). The image of (cid:99)M 1 Image((cid:99)M 1 H is the following H ) := {{(q1, p1), (q2, p2)} q1 = q2, p1 = −p2} ⊂ Symn−3(K5). 13 5) as follows. Let Z ⊂ Sym2(K(cid:48) (cid:103)Hilb 5) → Sym2(K5). Let (cid:101)Z ⊂ Hilb2(K(cid:48) 5) −→ Hilb2(K(cid:48) 5) 5). We denote by (K(cid:48) 2 5) be the proper pre-image of 5) be the proper pre- We take a blowing-up of Hilb2(K(cid:48) 5) → Sym2(K(cid:48) {(q1, p1), (q1,−p1)} ⊂ Sym2(K5) for Sym2(K(cid:48) image of Z for Hilb2(K(cid:48) (24) (25) Theorem 3.2. The map (23) is extended to the blowing-up along (cid:101)Z. Then we have the following (cid:99)MH −→ (cid:103)Hilb H , we have the injective map (cid:99)M 0 Proof. Since (cid:99)M 0 (cid:26) (E, Φ, ϕ, [σ]) ∈ (cid:99)M 1 We have the extended map (cid:99)M 0 The map is injective. (cid:99)M 10 H = M 0 H := 2 (K(cid:48) 5). H → Hilb2(K(cid:48) (cid:12)(cid:12)(cid:12)(cid:12) the apparent singularities {q1, q2} of (E, Φ, ϕ, [σ]) satisfy H ∪(cid:99)M 10 q1, q2 /∈ {t1, . . . , tn} H → Hilb2(K(cid:48) 5) by (cid:99)M 10 H (cid:51) (E, Φ, ϕ, [σ]) (cid:55)−→ ({(q, p), (q,−p)}, λ+ = ∞) ∈ Hilb2(K(cid:48) 5) 5) by Theorem 3.1. Set H (cid:27) . where (q, p), (q,−p) are the pairs of the apparent singularities and the dual parameters, and λ+ is the parameter of the fiber of the Hilbert-Chow morphism, that is, p2 − p1 = λ+(q2 − q1). H ∪(cid:99)M 10 H → Hilb2(K(cid:48) 5) to (cid:99)MH → Hilb2(K(cid:48) 5). For (E, Φ, ϕ, [σ]) ∈ (cid:99)M 1 H , we describe the Higgs field Φ as follows: Next, we extend the map (cid:99)M 0 z ⊗ ωz A1 R−1 0 (A1 (cid:40) Φ = z ⊗ ωz)R0 on U0 on U∞ where A1 z := (cid:19) By the automorphism of E ∼= O(1) ⊕ O(−2), we can normalize A1 (cid:18) 0 1 A1 z = f (6) 12 (z) 0 (26) (cid:32) f (3) f (6) 11 (z) 12 (z) 21 (z) −f (3) f (0) 11 (z) z as follows: (cid:33) . where we put f (6) 12 (z) = 0. The curve Cs passes through the points (t1, ν1), (t1,−ν1), . . . , (tn, νn), (tn,−νn), and (q, p), (q,−p). Here, (q, p), (q,−p) are the pairs of the apparent singularities and the dual parameters. We define the map 12 (z) := b0 + b1z + ··· + b6z6. The spectral curve Cs ⊂ K5 is defined by η2 − f (6) (cid:99)M 1 H (cid:51) (E, Φ, ϕ, [σ]) \(cid:99)M 10 H → Hilb2(K(cid:48) H (cid:55)−→ {(ti, νi, v), (ti,−νi,−v)} ∈ Hilb2(K(cid:48) 5) (cid:99)M 1 H \(cid:99)M 10 where {(ti, νi), (ti,−νi)} are the apparent singularities and the dual parameters of (E, Φ, ϕ, [σ]), and 5) by p − νi q − ti 1 2νi d dz f (6) 12 (ti). 2 = (K(cid:48) v = lim q→ti Then we have the natural extended map (cid:99)MH → Hilb2(K(cid:48) Let (cid:103)Hilb mined by the point of (cid:103)Hilb 5) be the blowing-up of Hilb2(K(cid:48) 5) along Z. We show that the spectral curves are deter- (K(cid:48) 5) where p2 − p1 = λ+(q2 − q1) and p2 + p1 = λ−(q2 − q1). First, any spectral curves pass through the points (ti, νi) and (ti,−νi) for i = 1, . . . , n, that is, the polynomial f (6) 12 (z) satisfies the condition ν2 12 (ti) = 0. By the equations, we can determine the coefficients b1, b2, b3, b6 by b4, b5. Second, the spectral curves 12 (qi) = 0 for i = 1, 2. If q1 (cid:54)= q2, then the passes through the points {(q1, p1), (q2, p2)}, that is, p2 i − f (6) coefficients b4 and b5 are determined by (q1, p1) and (q2, p2), that is, the spectral curve is determined by )}, λ+, λ−) be a point of (cid:103)Hilb i −f (6) 5). Let x := ({(q1, p1, vi,± ), (q2, p2, vi,± (K(cid:48) 5). 2 1 2 2 14 A. KOMYO AND M.-H. SAITO the apparent singularities and the dual parameters. We consider the behavior of the spectral curve as q2 → q1 and p2 → −p1. We consider the following equations (cid:40) 1 − f (6) p2 (−p1 + λ−(q2 − q1))2 − f (6) 12 (q1) = 0 12 (q2) = 0. 2 (K(cid:48) (cid:40)vi,+ 5) → Hilb2(K(cid:48) as q2 → q1 and p2 → −p1. When q1, q2 /∈ {t1, . . . , t5}, we can determine the coefficients b4 and b5 by q1, p1 and λ−. We consider the case q1 = q2 = ti for some i. Let λi− be the parameter such that 2 + vi,+ vi,+ 5). When we take q1 → ti and q2 → ti, we have the following equations 1 = λi−(q2 − q1), which is a blowing-up parameter of (cid:103)Hilb 12 (ti))2(cid:17) (cid:16) d2 curves are determined by the point of (cid:103)Hilb O(1) ⊕ O(−2), then Higgs fields are determined by points of (cid:103)Hilb q1(= q2). Then we have that the restriction map (cid:99)M 1 the restriction map (cid:99)MH → Hilb2(K(cid:48) H → Hilb2(K(cid:48) and λ−. Therefore, spectral 5). If the underlying vector bundles of Higgs bundles are 5) by the normalization (26). On the other hand, cyclic vectors of bundle type O(1) ⊕ O(−2) are determined by the apparent singularities 5) is injective. Finally, we obtain that (cid:3) By these equations, we can determine the coefficients b4 and b5 by q1, vi,+ f (6) 12 (ti) 12 (ti) − 1 f (6) 1 = 1 2νi 1 λ− = 1 2 2νi 5) is injective. (K(cid:48) (K(cid:48) f (6) ( d dz 2ν2 i d dz dz2 1 2 2 . We can describe the image of the map (25) as follows. We define parameters (vj,±)1≤j≤5 by L → L. The moduli space (cid:99)MH is stratified H satisfies E ∼= O(i)⊕O(−i− 1). ± i 2 H 1 2 1 2 (K(cid:48) p − ν± 5) are the following ), (q2, p2, vi,± ), (q2, p2, vi,± ({(q1, p1, vi,± )}) q1 (cid:54)= q2, vi,± H ∪(cid:99)M 1 as (cid:99)MH = (cid:99)M 0 H where (cid:99)M i The parameters (vj,±)1≤j≤5 are blowing-up parameters of Blν Then the images of (cid:99)M 0 H and (cid:99)M 1 (27)(cid:99)M 0 (cid:91)(cid:26) ∼= {({(q1, p1, vi,± ({(q1, p1, vi,± (cid:91) ({(q1, p1, vi,± ({(q1, p1, vi,± (cid:91) j (tj − tk1 )(tj − tk2 )(tj − tk3 ) = vj,±(q − tj). H is the locus such that (E, Φ, ϕ, [σ]) ∈ (cid:99)M i H in (cid:103)Hilb (cid:12)(cid:12)(cid:12)(cid:12) q1, q2 /∈ {t1, . . . , t5}, (cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) q1 = q2 = tj, (cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) q1, q2 (cid:54)= tj for any j = 1, . . . , 5, (cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) j ∈ C}∪ q1 − q2 = p1 − p2 = 0, λ+ ∈ C q1 − q2 = p1 + p2 = 0, λ+ = ∞, λ− ∈ C q1 = q2 = tj, p1 = −p2 = ν
λ+ = ∞ 1 = −vj,−
vj,
p1 = p2 = ν
vj,
1 = vj,
)}, λ+, λ−, λi−) 2 = λ+ ∈ C ), (q2, p2, vi,± ), (q2, p2, vi,± ), (q2, p2, vi,± )}, λ+, λi +) )}, λ+, λ−) (28)(cid:99)M 1 )}, λ+)
=±
=± j=1,...,5 ∼= H j=1,...,5 1 2 1 2 1 2 2 = λ− ∈ C (cid:27) ∪ ∪ Here, λ± and λj± satisfy the following relations Remark 3.3. We consider the map p1 − p2 = λ+(q1 − q2), p1 + p2 = λ−(q1 − q2), +(q1 − q2) 2 = λi 2 = λi−(q1 − q2). 1 − vi,
vi,
1 + vi,
vi,
(cid:99)M 1 H −→ P1 (E, Φ, ϕ, [σ]) (cid:55)−→ q1 j (tj − tk1 )(tj − tk2 )(tj − tk3 ) j (tj − tk1)(tj − tk2)(tj − tk3)   . given by the description (28) and the natural projection. Any fiber of this map is C2, which is isomorphic to M 1 H . Note that q1 is a coordinate of PH 0(P, E), which is a blowing-up parameter of (cid:99)MH → MH . EXPLICIT DESCRIPTION OF JUMPING PHENOMENA 15 In this section, we give an explicit description of the universal family of the moduli space (cid:99)MH for 4. Jumping families for Higgs bundles n = 5. For the purpose, we need give a description of jumping family, which is a family of Higgs fields such that for generic parameters, the underlying vector bundles are O⊕O(−1) and for special parameters, the underlying vector bundles are O(1)⊕O(−2). Descriptions of jumping families are given by the lower and upper modifications. In 4.2, we apply the description of jumping families to the case n ≥ 4. Then we obtain explicit descriptions of jumping families for the case n ≥ 4. 4.1. Jumping family for n = 5. Suppose that n = 5. We consider the following covering of (cid:99)MH : H ⊂ (cid:99)MH , V0 := (cid:99)M 0 V1 := (cid:8)(E, Φ, ϕ) ∈ M 00 H (cid:12)(cid:12) {p1, p2} (cid:54)⊂ Sym2[ branch points of Cs ] (cid:9) ∪(cid:99)M 1 H ⊂ (cid:99)MH , where Cs is the spectral curve of (E, Φ, ϕ) and pi = (qi, pi) ∈ Cs is an apparent singularity and its dual of (E, Φ, ϕ). Since we consider sl2-Higgs bundles, pi = 0 implies that pi = (qi, pi) ∈ Cs is a branch point of Cs. By Theorem 3.1, we have an explicit description of the universal family (EV0, ΦV0) on V0 × P1. an explicit description of the universal family ( E, Φ, [σ]) on (cid:99)MH × P1: Now we give an explicit description of the universal family (EV1 , ΦV1, [σV1]) on V1 × P1. Then we have (EV1, ΦV1, [σV1]) ⊂ / ( E, Φ, [σ]) ⊃o (EV0, ΦV0) (cid:19) , Let (cid:40) V1 × P1 ⊂ ⊃ V0 × P1. / (cid:99)MH × P1 (cid:18) ΦV0 = (29) be the family on V0 obtained by Theorem 3.1 for n = 5. Here we set z ⊗ ωz)R0 z = , where A0 on U0 on U∞ z ⊗ ωz A0 R−1 0 (A0 a1z + a0 (z − q1)(z − q2) −(a1z + a0) f (4) 12 (z) p1 − p2 q1 − q2 , a0 := − p1q2 − p2q1 q1 − q2 and we assume that q1, q2 (cid:54)= ∞ for simplicity. Set ((q1, p1), (q2, p2), λ) ∈ (K(cid:48) X := (30) 5)2 × C a1 := (cid:26) (cid:98)X := X ∪(cid:8)((q1, p1), (q2, p2), λ) ∈ (K(cid:48) (cid:18) 1 (cid:18) 1 0 (cid:19) P1 := z − q1 0 , P2 := Let P1, P2, and P3 be the following matrices and (31) (32) 0 0 1/z , R0 = (cid:18)1 (cid:19) (cid:12)(cid:12)(cid:12)(cid:12) p2 − p1 = λ(q2 − q1), (cid:27) 5)2 × C(cid:12)(cid:12) p2 − p1 = q2 − q1 = 0(cid:9) . (cid:19) p1 (cid:54)= 0, and q2 − q1 (cid:54)= 0 (cid:18) 1 (cid:19) , P3 := z−q1 0 0 1 . q1−q2 2p1 0 1 (cid:40) Proposition 4.1. We define a family of ν-sl2-parabolic Higgs bundle (EX , ΦX , ϕX , [σX ]) with a cyclic vector on X × P1 as (33) ∼= O(1)⊕O(−2). We can extend this family on X × P1 to the family on (cid:98)X × P1, naturally. For where U q1∞ = Spec C [w, 1/(q1w − 1)]. Here σX is the element of H 0(P1, EX ) such that the zero of σX is q1 when EX ΦX = (P1P2P3)−1(A0 z ⊗ ωw)R0 R−1 0 (A0 z ⊗ ωz)P1P2P3 on U0 on U q1∞ /       o / o o 16 A. KOMYO AND M.-H. SAITO the extended family, we have the following. If q1 (cid:54)= q2, then the underlying vector bundle is O ⊕ O(−1). If q1 = q2, then the underlying vector bundle is O(1) ⊕ O(−2). Proof. We describe the construction of a family of ν-sl2-parabolic connections with a cyclic vector on X × P1 so that this family satisfies the assertions of the proposition. As the result, we obtain the family which has the description (33). Then this proposition follows from this construction. By the natural map X → V0, the family ΦV0 on V0 × P1 induces the family on X × P1, denoted by X . We consider the lower and upper modifications (q1, lp1 )up ◦ (q1, lp1)low(Φ0 Φ0 X ), denoted by ΦX , where lp1 is a one dimensional subspace of Eq1 which corresponds to the eigenspace of the residue of Φ0 X at q1 with the eigenvalue p1. Explicitly, the modifications are described as follows. We consider the following diagram U0 × C2 R0 U q1∞ × C2 P1 where In particular, T := T (q2) U0 × C2 P2◦P3 U0 × C2 (cid:18) z − q1 − q1−q2 (cid:19) . 0 z(z−q1) 1 2p1 (cid:18) z − q1 (cid:19) 0 z(z−q1) 1 (34) Here, the transformation P1 implies the lower modification (q1, lp1)low and the transformation P2 ◦ P3 implies the upper modification (q1, lp1 )up. Namely, we describe ΦX as = lim q2→q1 T = 0 . T 0 q1 (35) = (cid:40) ΦX = (q1, lp1)up ◦ (q1, lp1)low(ΦX ) (P1P2P3)−1(A0 z ⊗ ωw)R0 R−1 0 (A0 z ⊗ ωz)P1P2P3 on U0 on U q1∞ . Taking the limit q2 → q1, we have Higgs bundles of bundle type O(1) ⊕ O(−2). Then we have the description of the family of the Higgs fields (33), and the family satisfies the assertion of the proposition. (cid:3) By the transition function (34), the zero of cyclic vectors is q1 when EX ∼= O(1) ⊕ O(−2). By this proposition, we have a map (cid:98)X → V1 ⊂ (cid:99)MH . Next we compute the apparent singularities and the dual parameters of ΦX when q1 (cid:54)= q2. Put (cid:32) z − q1 − q1−q2 2p1 2p1 q1−q2 0 (cid:33) (cid:32) (cid:33) . 1 − 1 0 2p1w2 (q1−q2)(wq1−1) (36) Q1 := , Q2 := Then we have the following diagram U0 × C2 Q1 U0 × C2 T (q2) U q1∞ × C2 Q2 U q1∞ × C2. R0 By the transformation by Q1 and Q2, we can describe ΦX as in the description (5). The apparent singularities and the dual parameters of ΦX is {(q1,−p1), (q2, p2)} (Figure 2). Then we have a map o o

O O O O o o O O o o O O 2 (K(cid:48) X → (cid:103)Hilb parameter of (cid:103)Hilb EXPLICIT DESCRIPTION OF JUMPING PHENOMENA 5). Moreover, we can extend this map to (cid:98)X → (cid:103)Hilb (cid:98)X \ X (cid:51) ((q1, p1), (q1, p1), λ) (cid:55)−→ ((q1,−p1), (q1, p1),∞, λ) ∈ (cid:103)Hilb 5) by (K(cid:48) 2 2 (K(cid:48) 5) 17 where (∞, λ) are values of a parameter of pre-images of the Hilbert-Chow morphism and of a blowing-up 2 (K(cid:48) 5) → Hilb2(K(cid:48) 5). We obtain the commutative diagram (cid:98)X V1 / (cid:103)Hilb 2 (K(cid:48) 5). By this diagram, the explicit description of the family (33) induces an explicit description of the universal family on V1 × P1 which is parametrized by the apparent singularities and their duals. Figure 2. The apparent singularities and the dual parameters 4.2. Jumping families for n ≥ 5. In this section, we give an explicit description of jumping families for n ≥ 5 as in the previous section. (cid:33) , (cid:40) Let (37) Φ0 = be the family on M 0 z ⊗ ωz A0 R−1 0 (A0 (z) H obtained by Theorem 3.1. We assume that q1, . . . , qn−3 (cid:54)= ∞ for simplicity. f (n−1) (z) (z − q1)··· (z − qn−3) −f (n−2) z ⊗ ωw)R0 on U0 on U∞ , where A0 f (n−2) z = (z) 11 12 11 (cid:32) First, we construct a family having Higgs bundles of bundle type O(1) ⊕ O(−2) from the family Φ0 by lower and upper modifications as in 4.1. Fix (q1, p1) ∈ K(cid:48) If we take the limit q2 → q1 of Φ1, then we have Higgs bundles of bundle type O(1) ⊕ O(−2): (38) Φ1 := (q1, lp1)up ◦ (q1, lp1)low(Φ0). n and assume that q1 (cid:54)= q2. Put (cid:32) (cid:48)f (n+1) (z) (z − q3)··· (z − qn−3) −(cid:48)f (n−2) (cid:19) (cid:48)f (n−2) (z) 11 12 11 (z) (cid:33) (cid:40) lim q2→q1 Φ1 = and z ⊗ ωz A1 1 )−1(A1 (Rq1 z ⊗ ωw)Rq1 1 on U0 on U q1∞ , where A1 z = (cid:18) z − q1 0 Rq1 1 = 0 z(z−q1) 1 . pi − (cid:48)f (n−2) 11 (qi) = 0 for i = 3, . . . , n − 3, zz=q1 ) = 0, 1 − det(A1 p2 det (resz=ti limq2→q1 Φ1) − ν2 i = 0 for i = 1, . . . , n. Here, the entries of A1 z satisfy the following equations: Note that q1 is the zero of the corresponding cyclic vector σ ∈ H 0(P1, E). $ $    / 18 A. KOMYO AND M.-H. SAITO Next, we construct a family having Higgs bundles of bundle type O(2) ⊕ O(−3) from the family Φ1. n and Here, we assume that q1, . . . , qn−3 (cid:54)= ∞ for simplicity. For the Higgs field (29), fix (q3, p3) ∈ K(cid:48) assume that q3 (cid:54)= q4. Put Φ2 := (q3, lp3)up ◦ (q3, lp3)low lim q2→q1 Φ1 If we take the limit q4 → q3 of Φ2, then we have Higgs bundles of bundle type O(2) ⊕ O(−3): (39) (cid:33) (cid:48)(cid:48)f (n−2) (cid:48)(cid:48)f (n+3) (z) (z − q5)··· (z − qn−3) −(cid:48)(cid:48)f (n−2) (z) 11 12 11 (z) (cid:40) lim q2→q1 Φ2 = and z ⊗ ωz A2 2 )−1(A2 (Rq1 z ⊗ ωw)Rq1 2 Here, the entries of A2 z satisfy the following equations: (cid:18) (cid:32) (cid:19) (cid:19) . where A2 z = on U0 on U q1∞ (cid:18) (z − q1)(z − q3) 0 0 z(z−q1)(z−q3) 1 Rq1 2 = pi − (cid:48)(cid:48)f (n−2) 11 (qi) = 0 for i = 5, . . . , n − 3, zz=qi) = 0 for i = 1, 3, i − det(A2 p2 det (resz=ti limq2→q1 Φ2) − ν2 i = 0 for i = 1, . . . , n. Note that q1, q3 is the zeros of the corresponding cyclic vector σ ∈ H 0(P1, E). We continue this process. Then we have family having Higgs bundles of bundle type O(k) ⊕ O(−k − 1) for k = 1, . . . , [(n − 3)/2]. 5. Geometric description for connection cases Suppose that ν satisfies the condition (3) and ν1 ··· νn (cid:54)= 0. We put (t1, . . . , tn) := (0, 1, x1, . . . , xn−3,∞), n−1, ν+ n , ν− (ν± 1 , . . . , ν± n ) := (±ν0,±ν1, . . . ,±νn−1, νn, 1 − νn), and νi := νi(ti − t1)··· (ti − ti−1)(ti − ti+1)··· (ti − tn−1) for i = 1, . . . , n − 1. Let M k (resp. (cid:99)M k) be the subvariety of M (resp. (cid:99)M ) where E ∼= O(k) ⊕ O(−k − 1). First, we define the apparent singularities and the dual parameters of (E,∇, ϕ) ∈ M 0, and we compute the apparent singularities and the dual parameters for n ≥ 4. Then we have Theorem 1.3 (Theorem 5.2). Second, we assume n = 5. We construct a jumping family. Third, we compute the apparent singularities of this jumping family on the locus of bundle type O ⊕ O(−1), and we analyze the behavior of the apparent singularities of the jumping families when the parameter closes to the jumping locus. Then we obtain a map from (cid:99)M to the Hilbert scheme of points on some surface. Moreover, we take some sequence of blowing-ups of the Hilbert scheme. Then we obtain an injective map from (cid:99)M to the blowing-ups. 5.1. Geometric description of M 0 for n ≥ 4. Let (E,∇, ϕ) ∈ M . We can define the apparent singularities of (E,∇, ϕ) ∈ M as follows. We fix a section σ ∈ H 0(P1, E). For the section σ, we define the following composition OP1 σ−−→ E ∇−−→ E ⊗ L −→ (E/OP1) ⊗ L. The composition OP1 → (E/OP1 ) ⊗ L is an OP1 -morphism, which is injective. Then we can define a subsheaf F 0 ⊂ E such that OP1 → (F 0/OP1 )⊗ L is an isomorphism. By the isomorphism F 0/OP1 ∼= L−1, we have F 0 ∼= OP1 ⊕ L−1. Therefore, we have the following exact sequence (40) 0 −→ OP1 ⊗ L−1 −→ E −→ TA −→ 0 where TA is a torsion sheaf. By the Riemann-Roch theorem, we have that the torsion sheaf TA is length n − 3. Definition 5.1. For (E,∇, ϕ) ∈ M and a nonzero section σ ∈ H 0(P1, E), we call the support of TA apparent singularities of a ν-sl2-parabolic connection with a cyclic vector (E,∇, ϕ, [σ]). For (E,∇, ϕ) ∈ M 0, we define dual parameters as follows. Since E ∼= O ⊕ O(−1), we can denote the EXPLICIT DESCRIPTION OF JUMPING PHENOMENA 19 (cid:32) (cid:33) connection ∇ by (cid:40) z ⊗ ωz 0 dR0 + R−1 ∇ = 0 (A0 d + A0 d + R−1 where A0 on U0 z ⊗ ωz)R0 on U∞ Note that the zeros of the polynomial f (n−3) (z) are the apparent singularities of (E,∇, ϕ). We denote by {q1, . . . , qn−3} the apparent singularities. We put pi := f (n−2) (qi) ∈ Lqi. We call {p1, . . . , pn−3} the dual parameters of (E,∇, ϕ) ∈ M 0. The definition of the apparent singularities and the dual parameters Let (cid:101)K(cid:48) is already given by Oblezin [12, Section 3]. and (cid:101)Kn be the contraction (cid:101)K(cid:48) n be the Zariski open set of the blowing-up of Hirzebruch surface of degree n − 2 defined in 2.2, n → (cid:101)Kn. Then we can define the following map z := (z) 21 11 21 11 f (n−2) f (n−3) 11 f (n−1) (z) (z) (z) −f (n−2) 12 . M 0 −→ Symn−3((cid:101)Kn) (41) (E,∇, ϕ) (cid:55)−→ {(q1, p1), . . . , (qn−3, pn−3)}, which is already constructed in [12, Section 3]. We consider the composite of the Hilbert-Chow morphism and the blowing-up n → (cid:101)Kn is the blowing up defined in 2.2. By the same argument as in the proof of Theorem 3.1, n) −→ Symn−3((cid:101)Kn). n) −→ Symn−3((cid:101)K(cid:48) Hilbn−3((cid:101)K(cid:48) where (cid:101)K(cid:48) we obtain the following Theorem 5.2. We can extend the map (41) to M 0 −→ Hilbn−3((cid:101)K(cid:48) n). This map is injective. Moreover, we can give an explicit description of the universal family ( E(0), ∇(0)) → M 0 × P1. 5.2. Jumping family for n = 5. Suppose that n = 5. In this section, we give an explicit description of a jumping family of connections. Let K(cid:48) n be the Zariski open set of the blowing-up of Hirzebruch surface of degree n − 2 corresponding n → Kn. to the moduli space of parabolic Higgs bundles (defined in 2.2), and Kn be the contraction K(cid:48) Fix (e0, e1) ∈ C2. Set (cid:26) X := ((q1, p1), (q2, p2), λ) ∈ (K(cid:48) 5)2 × C (cid:98)X := X ∪(cid:8)((q1, p1), (q2, p2), λ) ∈ (K(cid:48) (cid:32) (cid:33) (cid:32) Q1 := , Q2 := z − q1 − q1−q2 2p1 2p1 q1−q2 0 p1 (cid:54)= 0, and q2 − q1 (cid:54)= 0 (cid:12)(cid:12)(cid:12)(cid:12) p2 − p1 = λ(q2 − q1), (cid:27) 5)2 × C(cid:12)(cid:12) p2 − p1 = q2 − q1 = 0(cid:9) , (cid:18)1 (cid:33) 2p1w2 (q1−q2)(wq1−1) , and R0 = 0 1 − 1 0 which are defined in 4.1, and let Q1, Q2 and R0 be the following matrices (cid:19) , 0 1/z (42) and (43) (cid:40)d + Q1dQ−1 1 + Q1(cid:101)AU0 Q−1 respectively. (Q1 and Q2 was defined in 4.1). Proposition 5.3. We define a family of ν-sl2-parabolic connections (EX ,∇X , ϕX , [σX ]) with a cyclic vector on X × P1 as (44) ∇X = on U0 = Spec C [z] on U q1∞ = Spec C [w, 1/(q1w − 1)] . Here (cid:101)AU0 is defined below, and σX is the element of H 0(P1, EX ) such that the zero of σX is q1 when ∼= O(1) ⊕ O(−2). We can extend this family on X × P1 to the family on (cid:98)X × P1, naturally. For the EX extended family, we have the following. If q1 (cid:54)= q2, then the underlying vector bundle is O ⊕ O(−1). If q1 = q2, then the underlying vector bundle is O(1) ⊕ O(−2). 0 (cid:101)AU0R0 R−1 0 dR0 + R−1 d + Q2dQ−1 2 + Q2 Q−1 (cid:17) (cid:16) 1 2 20 We define a family of connection matrices (cid:101)AU0 on X × U0 as A. KOMYO AND M.-H. SAITO (cid:101)AU0 = (cid:18) F11(z) F12(z) F21(z) −F11(z) (cid:19) dz z(z − 1)(z − x1)(z − x2) (45) where F11(z) := 2p1(z − q1) q2 − q1 (cid:16)−2(p1 − ν5q2 + + (z − q1)λ + (1 − 2ν5e0(q2 − q1))z3 + (−2ν5e1(q2 − q1) + q1 − x1 − x2 − 1)z2 1(q2 − q1))e0 + 2ν5q1(q2 − q1)e1 − ν5(q2 − q1) q1(q1 − 1)(q1 − x1)(q1 − x2) p1 + q2 1 − (1 + x1 + x2)q1 + x1 + x2 + x1x2 z − 2p1(1 + e1 + e0q1) + (q1 − 1)(q1 − x1)(q1 − x2) , (cid:19) (cid:18) 1 2 (cid:17) F21(z) := (1 + e0(q1 − q2))z2 − (q1 + q2 + e1(q2 − q1))z + q1q2 + (q1 − q2)q1 2p1 (−2e1p1 − 2e0p1q1 + (q1 − 1)(q1 − x1)(q1 − x2)). We omit the description of F12(z), since the description is lengthened and is not necessary for computation of the apparent singularities and their duals. By this proposition, we have a map (cid:98)X → (cid:99)M . On the other hand, we have a map X → Hilbn−3((cid:101)K(cid:48) 1, p(cid:48) 1), (q(cid:48) 2, p(cid:48) 2)}, λ) where q(cid:48) 1, q(cid:48) 2 are the zero of F21(z), and p(cid:48) n) 1, p(cid:48) 2 defined by ((q1, p1), (q2, p2), λ) (cid:55)→ ({(q(cid:48) are their duals, that is, p(cid:48) i = F11(q(cid:48) i). Then we have the following diagram (46) (cid:98)X (cid:99)M / / Hilb2((cid:101)K(cid:48) 5). Proof of Proposition 5.3. We describe the construction of a family of ν-sl2-parabolic connections with a cyclic vector on X × P1 so that this family satisfies the assertions of the proposition. As the result, we obtain the family which has the description (44). Then this proposition follows from this construction. z ⊗ ωz be the Higgs field (29) on X × U0 defined in 4.1. Let ΦX be the Higgs field on X × P1 Let A0 defined by (47) z ⊗ ωz)P on U0, where the eigenvalues of the residue matrices are P −1(A0 and R−1 0 (A0 z ⊗ ωz)R0 on U∞ res0 ΦX res1 ΦX rest1 ΦX rest2 ΦX res∞ ΦX 1,−ν(cid:48) 5,−ν(cid:48) ν(cid:48) ν(cid:48) 5 4,−ν(cid:48) ν(cid:48) 2,−ν(cid:48) ν(cid:48) 3,−ν(cid:48) ν(cid:48) 1 3 2 4 . Here we put P := P1P2P3 where P1, P2, and P3 are the matrices (32). Note that the underlying vector bundles have bundle type O(1) ⊕ O(−2) when q1 = q2. When q1 (cid:54)= q2, by Q1 and Q2 which are the matrices (36), we denote the Higgs field ΦX by z ⊗ ωz)P Q1 on U0, z ⊗ ωz)R0Q2 on U∞ Q−1 1 P −1(A0 and Q−1 2 R−1 0 (A0 (48) as in the description (5). Now we construct a family of initial connections ∇0 on X × P1 and we determine ν(cid:48) 1, . . . , ν(cid:48) 5 so that ∇0 + ΦX is the desired family ∇X . We put (49) Then we have T −1∞ (restn ΦV0)T∞ = T −1∞ T∞ = (cid:18) 1 (cid:18) 0 −ν(cid:48)2 ν(cid:48) 5 1 0 5 (cid:19) (cid:19) . −1 0 (cid:18) ν(cid:48) 0−1 −ν(cid:48) 5 5 (cid:19) . T∞ =   % % 21 (cid:19) 0,wT∞(Bw ⊗ ωw)T −1∞ R0,w on U0 on U∞ Let ∇0 : O ⊕ O(−1) → (O ⊕ O(−1)) ⊗ Ω1P1 (D) be the connection defined by EXPLICIT DESCRIPTION OF JUMPING PHENOMENA (50) where we put R0,w := Bw := (cid:40) (cid:18) 1 0 and 0 1/w 0,wdR0,w + R−1 d + R−1 (cid:19) d + T∞(Bw ⊗ ωw)T −1∞ (cid:18) f3w3 + f2w2 + f1w + f0 (cid:40) d + R−1 d + T∞(Bw ⊗ ωw)T −1∞ + Q−1 (q1 − q2)(e2w2 + e1w + e0) 0,wdR0,w + R−1 d4w4 + d3w3 + d2w2 + d1w (w − 1)(x1w − 1)(x2w − 1) − (f3w3 + f2w2 + f1w + f0) where f0, . . . , f3, d1, . . . , d4, e0, e1, and e2 are parameters. When q1 (cid:54)= q2, we define the connection ∇0+ΦX as (51) Then the eigenvalues of the residue matrix of ∇0 + ΦX at ∞ is the following z ⊗ ωz)R0Q2 2 R−1 0 (A0 0,wT∞(Bw ⊗ ωw)T −1∞ R0,w + Q−1 1 P −1(A0 z ⊗ ωz)P Q1 on U0 on U∞. res∞ ΦX 5,−ν(cid:48) ν(cid:48) res∞(∇0 + ΦX ) 5 −f0 + ν(cid:48) 5, 1 − (−f0 + ν(cid:48) 5). We put (∇0)U0 = R−1 so that the limit limq2→q1 the parameters f0, . . . , f3, d0, . . . , d4, e1, e2, and e3 such that the following polynomial 0,wT∞(Bw ⊗ ωw)T −1∞ R0,w. First, we determine the parameters of Bw 1 ) + Q1(∇0)U0Q−1 (cid:1) is convergence. We claim that if we determine (cid:0)Q1d(Q−1 0,wdR0,w + R−1 1 (2f0 − 1)z4 + (1 + 2f1 − 2f0q1 + x1 + x2)z3 + (2f2 + 2e0p1 − 2f1q1 − x1 − x2 − x1x2)z2 + (2f3 + 2e1p1 − 2f2q1 + x1x2)z + 2(e2p1 − f3q1) is identically zero, then the limit is convergence. We solve the simultaneous linear equations as follows: f0 = f3 = e2 = 1 2 (q1 − x1 − x2 − 1), f2 = 1 , f1 = 2 (−2e1p1 − 2e0p1q1 + (q1 − 1)(q1 − x1)(q1 − x2)), and 1 2 q1 2p1 (−2e1p1 − 2e0p1q1 + (q1 − 1)(q1 − x1)(q1 − x2)). (−2e0p1 + q2 1 2 1 + x1 + x2 + x1x2 − q1(1 + x1 + x2)), (cid:40) Then we can define the limit limq2→q1(∇0 + ΦX ) as limq2→q1 (d + Q1d(Q−1 limq2→q1 (d + Q2d(Q−1 1 ) + Q1((cid:101)∇0)U0Q−1 2 ) + Q2T∞(Bw ⊗ ωw)T −1∞ Q−1 1 + P (q2)−1(A0 z ⊗ ωz)P (q2)) 0 (A0 z ⊗ ωz)R0) 2 + R−1 on U0 on U∞, which has bundle type O(1) ⊕ O(−2). (v1 Secondly, we determine the remainder parameters of Bw so that the residue matrix of ∇0 + ΦX at ti has the eigenvalues (νi,−νi) (resp. (νi, 1 − νi)) for i = 1, . . . , 4 (resp. i = 5). We take eigenvectors of the residue matrices of ΦX at ti as follows: , v2 (cid:48)+ ν i , v2 (cid:48)− ν i ) = (f + ) = (f− i (q1, p1, λ), (q1 − q2)(q1 − ti)(q2 − ti)) i (q1, p1, λ), (q1 − q2)(q1 − ti)(q2 − ti)) (cid:48)+ ν i (v1 (cid:48)− ν i i (q1, p1, λ) = p1(q2− ti) + p2(q1− ti)−
ν(cid:48) associated to ν(cid:48) associated to −ν(cid:48) i. i (cid:19) i(q1− q2)(ti− tj)(ti− tk)(ti− tl) where
∈ {+,−}. i = 1, . . . , 4,
∈ {+,−}. T
ti = f
i (q1, p1, λ) 0 (q1 − q2)(q1 − ti)(q2 − ti) Here, we put f
Set (cid:18) 1 Then we have (T
ti )−1(resti(Q−1 1 P −1(A0 z ⊗ ωz)P Q1))T
ti = (cid:18) −
ν(cid:48) i 0 (cid:19) . ∗
ν(cid:48) i (53) where h
i i (q1, p1, λ) = h
1 −1 q2 1 q2 2 d2 d3 d4 1 (q1,p1,λ)   d1 0 0 1 1 1 x3 1 x2 1 x1 x3 2 x2 2 x2   =  0 (q1 − ti)(cid:0)−2e1p1 − 2e0p1(q1 + ti) + (q1 − tj)(q1 − tk)(q1 − tl)(cid:1)(cid:16) i(ti − tj)(ti − tk)(ti − tl)(cid:1) − 2(q1 − ti)(cid:0)(q1 − ti)λ + ν(cid:48) i(ti − tj)(ti − tk)(ti − tl)(cid:1)2(cid:17) (cid:0)(q1 − ti)λ + ν(cid:48) 5ti(q1 − ti)(q2 − ti) −
iν(cid:48) 5ti(q1 − ti)(q2 − ti) −
iν(cid:48) h
2 2 (q1,p1,λ) (q1−1)2(q2−1)2 h
3 3 (q1,p1,λ) (q1−x1)2(q2−x1)2 (q1−x2)2(q2−x2)2 q1 − q2 4 (q1,p1,λ) 1 1 1 1 h
4 + 1 2 . p1 p1(q1 − ti) + p1(q2 − ti) 22 A. KOMYO AND M.-H. SAITO We fix a tuple
= (
1, . . . ,
4) where
i ∈ {+,−}. For the tuple
, we solve the following four linear equations (52) ((T
i ti )−1(resti(R−1 0,wdR0,w + R−1 0,wT∞(Bw ⊗ ωw)T −1∞ R0,w))T
i ti )21 = 0 i = 1, . . . , 4 where Aij is the (i, j)-entry of a (2 × 2)-matrix A. Explicitly, we can describe these equations as follows: (q1 − q2)(q1 − ti)2(q2 − ti)2(cid:0)t3 + 2(q1 − ti)(q2 − ti)gi(q1, p1, λ)(cid:0)t3 i (q1, p1, λ))2(cid:0)t2 − (g
i i d1 + t2 i e0 + tie1 + e2 i d2 + tid3 + d4 i f0 + t2 (cid:1) (cid:1) = 0 i = 1, . . . , 4 i f1 + tif2 + f3 (cid:1) where we put i (q1, p1, λ) := p1(q2 − ti) + p2(q1 − ti) − ν(cid:48) g
i = p1(q1 − ti) + p1(q2 − ti) − (q1 − q2)((q1 − ti)λ + ν(cid:48) 5ti(q1 − ti)(q2 − ti) −
iν(cid:48) i(ti − tj)(ti − tk)(ti − tl)). 5ti(q1 − q2)(q1 − ti)(q2 − ti) +
iν(cid:48) i(q1 − q2)(ti − tj)(ti − tk)(ti − tl) Then the solution of the equations (52) is the following h
i i (q1,p1,λ) We consider the case q1 = ti for some i. For fixed
= (
1, . . . ,
5),
i ∈ {+,−}, the domain of the function (q1−ti)2(q2−ti)2 is extended to q1 = ti, p1 =
i νi and λ = vi,
i is the blowing-up parameter of K(cid:48) 5 → K5 at (ti,
i νi). Therefore we can extend the family ∇0 to q1 = ti, p1 =
i νi and λ = vi,
i 1 when we substitute the solution d1, . . . , d4 associated to
. . Here, vi,
1 1 We compute the eigenvalue of the residues of ∇0 at ti for i = 1, . . . , 5. We put −1(resti(R−1 α
i i (q1, q2, p1, λ) := (T
ti β(1) i (q1, p1, λ)β(2) i (54) 0,wdR0,w + R−1 0,wT∞(Bw ⊗ ωw)T −1∞ R0,w)))T
ti )22 = 2p1(q2 − ti)(ti − tj)(ti − tk)(ti − tl) (
i, q1, p1, λ) i = 1, . . . , 4, where β(1) i (
i, q1, p1, λ) := (p1(q1 − ti) − (q1 − q2)((q1 − ti)λ + ν(cid:48) (q1, p1, λ) := −2e1p1 − 2e0p1(q1 + ti) + (q1 − tj)(q1 − tk)(q1 − tl), β(2) i 5ti(q1 − ti)(q2 − ti) −
iν(cid:48) Then we have ti )−1(resti((∇0)U0)T
i (T
i ti = for i = 1, . . . , 4. Here we put (∇0)U0 = R−1 0,wdR0,w + R−1 0 i (q1, q2, p1, λ) ∗ α
i i (q1, q2, p1, λ) 0,wT∞(Bw ⊗ ωw)T −1∞ R0,w. (cid:18) −α
i i(ti − tj)(ti − tk)(ti − tl))). (cid:19) We define the family of connections ∇X as ∇0 + ΦX . The limit limq2→q1 ∇X has bundle type O(1) ⊕ O(−2). The eigenvalues of the residues of ∇X at 0, 1, x1, x2, and ∞ are the following res0 ∇X −α
1 1 +
1ν(cid:48) 1 −
1ν(cid:48) α
1 res1 ∇X 1 −α
2 2 +
2ν(cid:48) 2 −
2ν(cid:48) α
2 rest1 ∇X 2 −α
3 3 +
3ν(cid:48) 3 −
3ν(cid:48) α
3 rest2 ∇X 3 −α
4 4 +
4ν(cid:48) 4 −
4ν(cid:48) α
4 1 2 3 4 4 − 1 res∞ ∇X 2 + ν(cid:48) 2 − ν(cid:48) 5 5 3 EXPLICIT DESCRIPTION OF JUMPING PHENOMENA 23 where we put α
i (ν(cid:48) 1, . . . , ν(cid:48) 5) of Higgs field ΦX as follows. We solve the following equations for ν(cid:48) i: i (q1, q2, p1, λ). For the fixed tuple (ν1, . . . , ν5), we determine the eigenvalues := α
(55) Then ν(cid:48) residues of the family ∇X at ti (i = 1, . . . , 5) are ν± + ν(cid:48) 5 = ν5, and − α
i − 1 2 i = νi for i = 1, . . . , 4. i +
iν(cid:48) i . Let (cid:101)AU0 be the connection matrix z ⊗ ωz)P Q1 1 P −1(A0 R−1 0,wdR0,w + R−1 0,wT∞(Bw ⊗ ωw)T −1∞ R0,w + Q−1 i is determined by νi and ν5 (when at least q2 is close to q1 enough). Then the eigenvalues of (see (51)). Then, we obtain the family (44), which satisfies the assertion of Proposition 5.3 by the (cid:3) construction. Remark 5.4. The connection ∇X is parametrized by e0 and e1. We can take e0, e1 ∈ C freely. 5.3. The apparent singularities and the dual parameters of ∇X . Let C∞ be the ∞-section of the n) be the birational map constructed by the apparent singularities and the dual parameters. For n = 5, by analyzing the behavior of the apparent singularities and their duals of ∇X as q2 → q1, we obtain the following ((cid:101)K(cid:48) 5 ∪ C∞). The moduli space (cid:99)M is biregular to its image Image φ((cid:99)M ) ⊂ 5 ∪ C∞), we have the 2 ((cid:101)K(cid:48) 5 ∪ C∞) → Hilb2((cid:101)K(cid:48) Hirzebruch surface of n − 2. Let φ : (cid:99)M (cid:57)(cid:57)(cid:75) Hilb2((cid:101)K(cid:48) Theorem 5.5. By taking a sequence of blowing-ups (cid:103)Hilb injective map φ : (cid:99)M → (cid:103)Hilb (cid:103)Hilb Proof. First, we construct a map (cid:99)M → Hilb2((cid:101)K(cid:48) ((cid:101)K(cid:48) 5 ∪ C∞). ities of ∇X is the zero of the polynomial 2 2 F21(z) = (1 + e0(q1 − q2))z2 − (q1 + q2 + e1(q2 − q1))z 5 ∪ C∞) by the birational map φ. The apparent singular- + q1q2 + (−2e1p1 − 2e0p1q1 + (q1 − 1)(q1 − x1)(q1 − x2)). 2 be the solution of the equation F21(z) = 0: 1 = q1 + a1(q2 − q1) − 2 = q1 + a1(q2 − q1) + q(cid:48) q(cid:48) , (cid:112)a3(q2 − q1) a2 Let q(cid:48) 1 and q(cid:48) where we set (q1 − q2)q1 2p1 (cid:112)a3(q2 − q1) a2 1 + e1 + 2e0q1 2 − 2e0(q2 − q1) a1 := a3 := p1(−p1(1 + e1 + 2e0q1)2(q1 − q2) + 2q1(q1 − 1)(1 + e0(q1 − q2))(q1 − x1)(q1 − x2)). , a2 := 2p1(−1 + e0(q2 − q1)), 2 = q1. The dual parameter of q(cid:48) If i), then the limit diverges (Figure 3). Let s be the parameter such that i = 1/pi for i = 1, 2. By explicit computation, we have i) for i = 1, 2. i = F11(q(cid:48) i is p(cid:48) 2 − q(cid:48) 1 = s(q(cid:48) 1 = limq2→q1 q(cid:48) Note that limq2→q1 q(cid:48) we consider limq2→q1 F11(q(cid:48) 2 − ¯p(cid:48) 1) where ¯p(cid:48) ¯p(cid:48) Then we can extend the map X → Hilb2((cid:101)K(cid:48) (56)(cid:98)X \ X (cid:51) ((q1, p1), (q1, p1), λ) (cid:55)−→ ({(q1,∞), (q1,∞)}, We can define a map φ : (cid:99)M → Hilb2((cid:101)K(cid:48) lim q2→q1 s = 1 q1(q1 − 1)(q1 − x1)(q1 − x2) 5) to (cid:98)X → Hilb2((cid:101)K(cid:48) 5 ∪ C∞) by . ) ∈ Hilb2((cid:101)K(cid:48) 1 q1(q1 − 1)(q1 − x1)(q1 − x2) 5 ∪ C∞). 5 ∪ C∞) so that φM 0 is the map in Theorem 5.2 and the diagram (cid:98)X (cid:99)M / Hilb2((cid:101)K(cid:48) 5 ∪ C∞). φ / is commutative.   ' ' 24 A. KOMYO AND M.-H. SAITO Figure 3. The apparent singularities and the dual parameters Now, we construct a sequence of blowing-ups (cid:103)Hilb ((cid:101)K(cid:48) 5 ∪ C∞) of Hilb2((cid:101)K(cid:48) (cid:103)Hilb ((cid:101)K(cid:48) 5 ∪ C∞) Hilb2((cid:101)K(cid:48) 5 ∪ C∞) (cid:99)M φ 2 2 φ 5 ∪ C∞) such that i := 1 p(cid:48) . Let ((q(cid:48) where φ is injective. Set ¯p(cid:48) ((cid:101)K(cid:48) 5 ∪ C∞) × ((cid:101)K(cid:48) First, we take the blowing-up of ((cid:101)K(cid:48) Bl1. Note that Hilb2((cid:101)K(cid:48) 1, ¯p(cid:48) 5∪ C∞) = Bl1/S2. We defined the blowing-up parameter s as ¯p(cid:48) 5 ∪ C∞) \ (C0 ∪ π−1(∞)) × (C0 ∪ π−1(∞)). 5∪ C∞)× ((cid:101)K(cid:48) 2− q(cid:48) 5∪ C∞) along the ideal I1 = (q(cid:48) 2)) be coordinates on 1), (q(cid:48) 2, ¯p(cid:48) 1, ¯p(cid:48) i 2− ¯p(cid:48) 2− ¯p(cid:48) 1), denoted by 2− q(cid:48) 1 = s(q(cid:48) 1), and we obtained (cid:18) lim q2→q1 s = q1(q1 − 1)(q1 − x1)(q1 − x2) 1 . (cid:19) Second, we take the blowing-up of Bl1 along the ideal I2 = 2 − q(cid:48) q(cid:48) 1, ¯p(cid:48) 2 − ¯p(cid:48) 1, s − q(q − 1)(q − x1)(q − x2) denoted by Bl2. We define the blowing-up parameters t1, t2 as 1 , ¯p(cid:48) 2 + ¯p(cid:48) 1 where q := q(cid:48) 1 + q(cid:48) 2 2 , s − 1 q(q − 1)(q − x1)(q − x2) = t1(q(cid:48) 2 − q(cid:48) 1), ¯p(cid:48) 2 + ¯p(cid:48) 1 = t2(q(cid:48) (cid:18) By explicit calculations, we have limq2→q1 t1 = limq2→q1 t2 = 0. Third, we take the blowing-up of Bl2 along the ideal I3 = 2 − q(cid:48) q(cid:48) 1, ¯p(cid:48) 2 − ¯p(cid:48) 1, s − 1 q(q − 1)(q − x1)(q − x2) , ¯p(cid:48) denoted by Bl3. We define the blowing-up parameters u1, u2 as t1 = u1(q(cid:48) explicit calculations, we have 1, t1, t2 2 + ¯p(cid:48) 2 − q(cid:48) 1), , t2 = u2(q(cid:48) 2 − q(cid:48) 1). By 2 − q(cid:48) 1). (cid:19) lim q2→q1 u1 = −λ 1(q1 − 1)2(q1 − x1)2(q1 − x2)2 + 4q2 u(1) 1 (q1)p1 1(q1 − 1)3(q1 − x1)3(q1 − x2)3 4q3 (57) lim q2→q1 u2 = u(0) 1 (q1) + −4q3 16q3 1 + 3q2 1(q1 − 1)3(q1 − x1)3(q1 − x2)3 , 1(1 + x1 + x2) − 2q1(x1 + x2 + x1x2) + x1x2 1(q1 − 1)2(q1 − x1)2(q1 − x2)2 4q2   / / 8 8 EXPLICIT DESCRIPTION OF JUMPING PHENOMENA 25 1 (q1) are polynomials in q1, which are independent of λ and p1. Fourth, we take (cid:16) 1 (q1) and u(0) where u(1) the blowing-up of Bl3 along the ideal 2 − ¯p(cid:48) 1,¯p(cid:48) u2 − −4q3 + 3q2(1 + x1 + x2) − 2q(x1 + x2 + x1x2) + x1x2 q(q − 1)(q − x1)(q − x2) 2 − q(cid:48) q(cid:48) 1, s − 2 + ¯p(cid:48) 1, t1, t2, I4 = , ¯p(cid:48) 1 4q2(q − 1)2(q − x1)2(q − x2)2 (cid:17) , denoted by Bl4. We put blowing-up parameters v as follows: By explicit calculations, we have limq2→q1 v = 0. Finally, we take the blowing-up of Bl4 along the ideal u2 − −4q3 + 3q2(1 + x1 + x2) − 2q(x1 + x2 + x1x2) + x1x2 (cid:16) 4q2(q − 1)2(q − x1)2(q − x2)2 I5 = 2 − q(cid:48) q(cid:48) 1, s − 2 − ¯p(cid:48) 1,¯p(cid:48) u2 − −4q3 + 3q2(1 + x1 + x2) − 2q(x1 + x2 + x1x2) + x1x2 q(q − 1)(q − x1)(q − x2) 2 + ¯p(cid:48) 1, t1, t2, , ¯p(cid:48) 1 (cid:17) , , v = v(q(cid:48) 2 − q(cid:48) 1). denoted by Bl5. We define the blowing-up parameter w as v = w(q(cid:48) have 4q2(q − 1)2(q − x1)2(q − x2)2 2 − q(cid:48) 1). By explicit calculations, we λ + 1 − 3q2 where w(1) −w(1)(q1)p1 1 (q1) and w(0) 1(1 + x1 + x2) + 2q1(x1 + x2 + x1x2) − x1x2) lim q2→q1 w = (58) 1(q1 − 1)3(q1 − x1)3(q1 − x2)3 (4q3 8q3 1(q1 − 1)4(q1 − x1)4(q1 − x2)4 + 8q4 ((cid:101)K(cid:48) 5) of Hilb2((cid:101)K(cid:48) 1 (q1) are polynomials in q1, which are independent of λ and p1. We define the 5) = Bl5/S2. Here, S2 is the symmetric group, which blowing-ups (cid:103)Hilb We consider the family limq2→q1 ∇X , which has bundle type O(1)⊕O(−2). The family is parameter- ized by (λ, p1, q1) ∈ C3. We fix the parameter q1. We can regard the parameters (λ, p1) as coordinates of M 1, which is isomorphic to C2. The family limq2→q1 ∇X where q1 is fixed is a universal family of M 1. (cid:3) The theorem follows form (57) and (58). 1(q1 − 1)4(q1 − x1)4(q1 − x2)4 5) as (cid:103)Hilb acts on Bl5 naturally. ((cid:101)K(cid:48) w(0)(q1) 64q4 2 2 References [1] D. Arinkin, S. Lysenko, On the moduli of SL(2)-bundles with connections on P1 \ {x1, . . . , x4}, Internat. Math. Res. Notices (1997), no. 19, 983 -- 999. [2] A. Beauville, M. S. Narasimhan, S. Ramanan, Spectral curves and the generalised theta divisor. J. Reine Angew. Math. 398 (1989), 169 -- 179. [3] V. G. Drinfeld, Elliptic modules and their applications to the Langlands and to the Peterson conjectures for GL(2) over functional field (in Russian), Ph. D. Thesis, Moscow State University, 1977. [4] B. Dubrovin, M. Mazzocco, Canonical structure and symmetries of the Schlesinger equations. Comm. Math. Phys. 271 (2007), no. 2, 289 -- 373. [5] M.-A. Inaba, Moduli of parabolic connections on curves and the Riemann-Hilbert correspondence. J. Algebraic Geom. 22 (2013), no. 3, 407 -- 480. [6] M. Inaba, K. Iwasaki, M.-H. Saito, Moduli of stable parabolic connections, Riemann- Hilbert correspondence and geometry of Painlev´e equation of type VI. I , Publ. Res. Inst. Math. Sci. (2006), no. 4, 987-1089. [7] M. Inaba, K. Iwasaki, M.-H. Saito, Moduli of stable parabolic connections, Riemann-Hilbert correspondence and ge- ometry of Painlev´e equation of type VI. II. Moduli spaces and arithmetic geometry, 387-432, Adv. Stud. Pure Math., 45, Math. Soc. Japan, Tokyo, 2006. [8] M. Inaba, M.-H. Saito, Moduli of regular singular parabolic connections of spectral type on smooth projective curves. In preparation. [9] F. Loray, M.-H. Saito, Lagrangian fibrations in duality on moduli spaces of rank 2 logarithmic connections over the projective line. Internat. Math. Res. Notices (2015), no. 4, 995 -- 1043. [10] H. Nakajima, Lectures on Hilbert schemes of points on surfaces. University Lecture Series, 18. American Mathematical Society, Providence, RI, 1999. [11] K. Okamoto, Isomonodromic deformation and Painlev´e equations, and the Garnier system. J. Fac. Sci. Univ. Tokyo Sect. IA Math. 33 (1986), no. 3, 575 -- 618. 26 A. KOMYO AND M.-H. SAITO [12] S. Oblezin, Isomonodromic deformations of sl(2) Fuchsian systems on the Riemann sphere. Mosc. Math. J. 5 (2005), no. 2, 415 -- 441, 494 -- 495. [13] M.-H. Saito, S. Szabo, Apparent singularities and canonical coordinates of moduli spaces of parabolic connections and parabolic Higgs bundles. in preparation. [14] C. Simpson, Harmonic bundles on noncompact curves. J. Amer. Math. Soc. 3 (1990), no. 3, 713 -- 770 [15] E. K. Sklyanin, Separation of variables -- new trends, Progr. Theoret. Phys. Suppl. (1995), no. 118, 35 -- 60. [16] H. Yoshida, On the structure of strata of the moduli space of parabolic connections with 5-regular singular points on P1, (Japanese), master thesis, Kobe university, (2015). Department of Mathematics, Graduate School of Science, Kobe University, 1-1 Rokkodai-cho, Nada-ku, Kobe, 657-8501, Japan E-mail address: [email protected] Department of Mathematics, Graduate School of Science, Kobe University, 1-1 Rokkodai-cho, Nada-ku, Kobe, 657-8501, Japan E-mail address: [email protected] 

1812.05470

2

1812

Rational curves on a smooth Hermitian surface

We study the set $R$ of nonplanar rational curves of degree $d<q+2$ on a smooth Hermitian surface $X$ of degree $q+1$ defined over an algebraically closed field of characteristic $p>0$, where $q$ is a power of $p$. We prove that $R$ is the empty set when $d<q+1$. In the case where $d=q+1$, we count the number of elements of $R$ by showing that the group of projective automorphisms of $X$ acts transitively on $R$ and by determining the stabilizer subgroup. In the special case where $X$ is the Fermat surface, we present an element of $R$ explicitly.

math.AG

math

RATIONAL CURVES ON A SMOOTH HERMITIAN SURFACE NORIFUMI OJIRO Abstract. We study the set R of nonplanar rational curves of degree d < q + 2 on a smooth Hermitian surface X of degree q + 1 defined over an algebraically closed field of characteristic p > 0, where q is a power of p. We prove that R is the empty set when d < q + 1. In the case where d = q + 1, we count the number of elements of R by showing that the group of projective automorphisms of X acts transitively on R and by determining the stabilizer subgroup. In the special case where X is the Fermat surface, we present an element of R explicitly. 1. Introduction Let q be a power of a prime p, and k an algebraic closure of the finite field Fq. For a matrix m with entries in k, we denote by m(q) the matrix whose entries are the q-th power of those of m. We denote by a column vector x = t(x0, x1, x2, x3) a point in the k-projective space P3. Let A be a nonzero 4-by-4 matrix with entries in k. A k-Hermitian surface XA is defined by XA := {x ∈ P3 txAx(q) = 0}. q2 and tA = A(q), If A is a Hermitian matrix, namely A has the entries in F the surface XA is called a Hermitian surface. It is easily shown that XA is smooth if and only if A is invertible. The geometry of Hermitian varieties was systematically investigated by B. Segre in [8]. Especially, the number of linear spaces lying on a Hermitian variety and their configuration were considered. It was shown that the numbers of points and lines on a smooth Hermitian surface in P3(F q2) are equal to (q3 + 1)(q2 + 1) and (q3 + 1)(q + 1) respectively, and no plane is contained. Further, the set of points and lines on a smooth Hermitian surface forms a block design, see also [3]. In recent years, the number of rational normal curves totally tangent to a smooth Hermitian variety X has been determined in [10] by considering the action of the automorphism group of X on the set of the curves. In [11], non-singular conics totally tangent to the smooth Hermitian curve of degree 6 in characteristic 5 were utilized for a geometric construction of strongly regular graphs. On the other hand, projective isomorphism classes of degenerate Hermitian varieties of 2010 Mathematics Subject Classification. 51E20, 14M99, 14N99. Key words and phrases. rational curve, Hermitian surface, positive characteristic. 1 2 NORIFUMI OJIRO corank 1 and the automorphism group of each isomorphism class have been determined in [7]. Let A be an invertible 4-by-4 matrix with entries in k. We will be con- cerned with rational curves of degree > 1 on a smooth k-Hermitian surface XA. Let d be a positive integer and F a 4-by-(d + 1) matrix of rank(F ) ≥ 2 with entries in k. A rational curve CF of degree d in P3 is the image of a rational map (1) P1 ∋ t(s, t) 7−→ F t(sd, sd−1t, . . . , std−1, td) ∈ P3. We call rank(F ) the rank of the curve CF . If rank(F ) = 2, then CF degen- erates to a line. If rank(F ) = 3, then CF degenerates to a plane curve of degree ≥ 2. When rank(F ) = 4, the curve CF is nondegenerate and is a space curve of degree ≥ 3. Then CF is said to be nonplanar, namely CF is not contained in any plane. Thus the study of rational curves of rank 2 on XA is reduced to that of lines on XA. Further, an algebraic curve of rank 3 on XA is a smooth k-Hermitian curve of degree q + 1, which is of genus q(q − 1)/2 > 0. Hence we may restrict ourselves to the case of rank 4. Our results are as follows: Theorem 1.1. There is no nonplanar rational curve of degree ≤ q on a smooth k-Hermitian surface. Let R be the set of nonplanar rational curves of degree q + 1 on a smooth k-Hermitian surface XA. As will be seen later, the set R is nonempty and each element is projectively isomorphic over k to the smooth curve C0 := (cid:8)t(sq+1, sqt, stq, tq+1) ∈ P3 t(s, t) ∈ P1(cid:9) . We denote by Aut(XA) the group of projective automorphisms of XA. q2) defined by Let n be a positive integer. We deal with the group PGUn(F {Q ∈ GLn(F q2) tQQ(q) = I}/µq+1I, where µq+1 denotes the group of (q + 1)-th roots of unity and I denotes the unit matrix. As is well-known, the group Aut(XA) is isomorphic to PGU4(F q2). Then we shall prove the following theorem. Theorem 1.2. The group Aut(XA) acts transitively on the set R, and the stabilizer subgroup is isomorphic to PGU2(F q4). By Theorem 1.2, the cardinality of R is equal to PGU4(F q2)/PGU2(F q4). We know by [6, pp.64-65] that PGU4(F q2) = q6(q4 − 1)(q3 + 1)(q2 − 1) and PGU2(F q4) = q2(q4 − 1). Thus we have the following. Corollary 1.3. R = q4(q3 + 1)(q2 − 1). The number R is 432, 18144, 249600, 1890000, 39645312, 383162400, . . . as q = 2, 3, 4, 5, 7, 9, . . . . RATIONAL CURVES ON A SMOOTH HERMITIAN SURFACE 3 In the special case where A = I, that is, where the surface XA is the Fermat surface, we can explicitly give an element CFJ of R such as (cid:8)t(cid:0)η−qξqsq+1 − η−qtq+1, sqt, stq, ωη−1ξsq+1 + ωη−1tq+1(cid:1) ∈ P3(cid:12)(cid:12) t(s, t) ∈ P1(cid:9) , q2 satisfying ωq+1 = −1, ξq+1 = 1 with where ω, ξ, and η are elements of F ξ2 6= −1, and ηq+1 = ξq + ξ. Note that η 6= 0 because ξ2 6= 0, −1. The curve CFJ is smooth since it is projectively isomorphic to the smooth curve C0. On the other hand, a complete set of representatives for Aut(XI ) can q2) (see Lemma 4.1). Therefore we have the following. be taken from GL4(F Corollary 1.4. All nonplanar rational curves of degree q + 1 on XI are projectively isomorphic over F q2 to the smooth curve CFJ . q2) = 45 where XI (F In the case where q = 2, we have XI (F q2) denotes q2-rational points of XI , and Aut(XI ) is of order 25920. Then the set of F CF (F q2) = 5 for each nonplanar cubic CF on XI . We can actually obtain by computation 432 nonplanar cubics on XI and the stabilizer subgroup of q2), we can verify Aut(XI ) fixing CFJ of order 60. By restricting XI to XI (F that each cubic intersects 150 other cubics at a single point, 40 other cubics at two points and another cubic at five points. Here, when we say two cubics q2) = n. We can also q2)∩CF ′(F CF , CF ′ intersect at n points we mean CF (F q2) and the stabilizer subgroup verify that Aut(XI ) acts transitively on XI (F is of order 576, and furthermore, there are 48 cubics passing through each q2). These computational data files obtained by using GAP [4] point of XI (F are available upon request addressed to the author. We give a brief outline of our paper. In the next section, we prove The- orem 1.1. By the same argument, we show directly that each irreducible conic, which is a rational curve of rank 3, is not contained in XA. In section 3, we give a bijection between the set R and the quotient of certain sets con- sisting of invertible 4-by-4 matrices, by showing basic lemmas. In section 4, we first prove two lemmas which are necessary for our proof of Theorem 1.2. We prove Theorem 1.2 in the last of the section. The author is grateful to Professor Ichiro Shimada for his encouragement during the course of the work and helpful suggestions on drafts. 2. Proof of Theorem 1.1 Proof of Theorem 1.1. Suppose that a nonplanar rational curve CF defined by (1) is contained in a smooth k-Hermitian surface XA. Denoting by bi,j the entries of the (d + 1)-by-(d + 1) matrix tF AF (q), one has the identity (2) d Xi,j=0 bi,jsd−i+q(d−j)ti+qj ≡ 0. Therefore if d < q, all the coefficients bi,j must vanish because the exponents (i + qj)'s are all different. This implies that tF AF (q) = O, but it is a In fact, since rank(F ) = 4 by definition, we can take an contradiction. 4 NORIFUMI OJIRO invertible matrix F ∗ consisting of linearly independent 4 column vectors of F . Then, however, tF ∗AF ∗(q) must be O. If d = q, the coefficients bi,j must vanish except for bq,l−1 = −b0,l with 1 ≤ l ≤ q. This implies that rank(tF AF (q)) ≤ 2, but it is a contradiction by the argument above. Hence we conclude that CF 6⊂ XA. (cid:3) Remark 2.1. We can similarly give a proof for the case of irreducible conics. In fact, since an irreducible conic CF is of rank 3, we can make an invertible matrix F ∗ consisting of linearly independent 3 column vectors of F and a vector linearly independent to those vectors. Suppose that CF ⊂ XA. Since d = 2 ≤ q, one has rank(tF AF (q)) ≤ 2 in the same argument as the above proof. Therefore the 4-by-4 matrix tF ∗AF ∗(q) must be of rank 3 at the most, but tF ∗AF ∗(q) is of rank 4 by definition. This is a contradiction. As we have seen, this proof is valid for rational curves which are of rank ≥ 3 and degree ≤ q. 3. Basic lemmas In this section, we will prove some basic lemmas to prepare for our proof of Theorem 1.2. The following lemma gives a necessary and sufficient condition for a nonplanar rational curve of degree q + 1 to be on a smooth k-Hermitian surface. Lemma 3.1. Let CF be a nonplanar rational curve of degree q + 1 defined by (1). The curve CF is contained in a smooth k-Hermitian surface XA if and only if the (q + 2)-by-(q + 2) matrix tF AF (q) is of the form 0 0 0 ... 0   −b0,1 −b1,1 b0,1 0, . . . , 0 b1,1 0, . . . , 0 0 0, . . . , 0 ... 0 0 0 ... 0, . . . , 0 0, . . . , 0 −b0,q+1 0, . . . , 0 −b1,q+1 0 0 0 ... 0 b0,q+1 b1,q+1 0 ... 0 0 0 .   If the above condition is satisfied, the matrix F is of the form (f0, f1, 0, . . . , 0, fq, fq+1). Proof. As was seen above, the curve CF is contained in XA if and only if one has (2). In the present case where d = q + 1, if CF ⊂ XA then the coefficients bi,j must vanish except for bq,l−1 = −b0,l, bq+1,l−1 = −b1,l with 1 ≤ l ≤ q +1. Since rank(F ) = 4, there are 4 column vectors fx, fy, fz, fw of F with 0 ≤ x < y < z < w ≤ q+1 such that the matrix F ∗ := (fx, fy, fz, fw) is invertible. Then none of x, y, z, w is from 2 to q − 1 because tF ∗AF ∗(q) is also invertible, and thus x = 0, y = 1, z = q, w = q + 1. Let fi be the i-th RATIONAL CURVES ON A SMOOTH HERMITIAN SURFACE 5 column vector with 2 ≤ i ≤ q − 1 of F . Then one has tfiAF ∗(q) = (bi,0, bi,1, bi,q, bi,q+1) = (0, 0, 0, 0), and thus fi = 0. Hence F and tF AF (q) are of the form described above. The converse is obvious since (2) holds automatically. (cid:3) A rational curve CF defined by (1) is also obtained by replacing F by λF ϕ(g), where λ is an element of the multiplicative group k× and ϕ is a homomorphism from GL2(k) to GLd+1(k) defined by the following: for each t(s, t) ∈ k2 with t(s, t) 6= t(0, 0) and g ∈ GL2(k), put t(u, v) := g t(s, t), then ϕ : −→ GLd+1(k) GL2(k) ∈ ∈ (cid:0)g : t(s, t) 7→ t(u, v)(cid:1) 7−→ (cid:0)ϕ(g) : t(sd, sd−1t, . . . , td) 7→ t(ud, ud−1v, . . . , vd)(cid:1) . Indeed, it is obvious by definition that ϕ(I) = I. Putting t(x, y) := h t(u, v) for each h ∈ GL2(k), one has ϕ(hg) t(sd, sd−1t, . . . , td) = t(xd, xd−1y, . . . , yd) = ϕ(h) t(ud, ud−1v, . . . , vd) = ϕ(h)ϕ(g) t(sd, sd−1t, . . . , td). Hence ϕ(hg) = ϕ(h)ϕ(g), and thus ϕ(g) ∈ GLd+1(k). Conversely if there is a matrix F ′ such that CF = CF ′, then one has F t(sd, sd−1t, . . . , std−1, td) = F ′ t(ud, ud−1v, . . . , uvd−1, vd) ∈ P3. This implies that there are homogeneous polynomials f , f ′ of degree d such that f (s, t) = f ′(u, v). Therefore there is an element g of GL2(k) such that t(s, t) = g t(u, v) ∈ P1, and thus F ′ = λF ϕ(g) for some λ ∈ k×. Hence, denoting by Im(ϕ) the image of ϕ, we see that the set k×F Im(ϕ) corresponds one-to-one with CF . Let S be the set of matrices F such that tF AF (q) satisfies the condition of Lemma 3.1. Then by Lemma 3.1, for each F ∈ S the set k×F Im(ϕ) corre- sponds one-to-one with the nonplanar rational curve CF on XA. Therefore one has the following bijection (3) k×\S/Im(ϕ) ∋ k×F Im(ϕ) 7−→ CF ∈ R. By Lemma 3.1, we define the map ∗ : S ∋ F = (f0, f1, 0, . . . , 0, fq, fq+1) 7−→ F ∗ = (f0, f1, fq, fq+1) ∈ S∗, where S∗ is written as S∗ = {F ∗ ∈ GL4(k) tF ∗AF ∗(q) = DB, B ∈ GL2(k)}, and DB is a matrix defined by DB := (cid:18) 0 0 −b1 0 −b2 b1 b2 0(cid:19) ∈ GL4(k) for B = (b1, b2) ∈ GL2(k). 6 NORIFUMI OJIRO Further, we define the map ∗ from Im(ϕ) ⊂ GLq+2(k) to Im(ϕ)∗ ⊂ GL4(k) as follows: for every g = (cid:18)α β δ(cid:19) ∈ GL2(k), γ ϕ(g)=   αq+1 αqβ , . . . , αβq βq+1 δβq αqγ αqδ ... ... ... βδq βγq αγq δq+1 γq+1 δγq , . . . , γβq ... , . . . , αδq , . . . , γδq ...   where Im(ϕ)∗ is written as 7→ ϕ(g)∗=  αq+1 αqβ αβq βq+1 δβq αqγ αqδ βδq αγq γq+1 δq+1 γβq βγq αδq δγq γδq ,   Im(ϕ)∗ = (cid:26)(cid:18)αqg βqg γqg δqg(cid:19) ∈ GL4(k) (cid:12)(cid:12)(cid:12)(cid:12) g ∈ GL2(k)(cid:27) . Indeed, it is easy to see that det(ϕ(g)∗) = det(g)2q+2 for every g ∈ GL2(k), and thus ϕ(g)∗ ∈ GL4(k). We denote by ϕ∗ the composition of ϕ and ∗, namely ϕ∗(g) = ϕ(g)∗ for every g ∈ GL2(k). Lemma 3.2. The map ϕ∗ is a homomorphism from GL2(k) to GL4(k). There is the following natural bijection k×\S/Im(ϕ) −→ k×\S∗/Im(ϕ)∗. Proof. For each one has Therefore g = (cid:18)α β δ(cid:19) , h = (cid:18)x y z w(cid:19) ∈ GL2(k), γ gh = (cid:18)αx + βz αy + βw γy + δw(cid:19) . γx + δz ϕ∗(gh) = (cid:18)(αx + βz)qgh (αy + βw)qgh (γx + δz)qgh (γy + δw)qgh(cid:19) . On the other hand, ϕ∗(g)ϕ∗(h) = (cid:18)αqg βqg δqg(cid:19)(cid:18)xqh yqh zqh wqh(cid:19) γqg = (cid:18)αqxqgh + βqzqgh αqyqgh + βqwqgh γqxqgh + δqzqgh γqyqgh + δqwqgh(cid:19) = (cid:18)(αqxq + βqzq)gh (αqyq + βqwq)gh (γqxq + δqzq)gh (γqyq + δqwq)gh(cid:19) . Since the q-th power is an automorphism of k, one has ϕ∗(gh) = ϕ∗(g)ϕ∗(h) and thus ϕ∗ is a homomorphism from GL2(k) to GL4(k). RATIONAL CURVES ON A SMOOTH HERMITIAN SURFACE 7 For each F ∈ S, g ∈ GL2(k), denoting by ai,j the entries of ϕ(g), we can write the j-th column vector gj with j ∈ {0, 1, q, q + 1} of F ϕ(g) as gj = Xi∈{0,1,q,q+1} ai,jfi, since fi = 0 for 2 ≤ i ≤ q − 1. Then it is immediate from definition that F ∗ϕ∗(g) = (g0, g1, gq, gq+1), and thus (F ϕ(g))∗ = F ∗ϕ∗(g). This implies that there is the natural map from k×\S/Im(ϕ) to k×\S∗/Im(ϕ)∗. The bijectivity is obvious since by definition the map S → S∗ is bijective. (cid:3) By (3) and Lemma 3.2, one has the bijection (4) k×\S∗/Im(ϕ)∗ ∋ k×F ∗Im(ϕ)∗ 7−→ CF ∈ R. The following well-known proposition is useful. The readers may find a proof for example in [2] and [9, Proposition 2.5.]. Proposition 3.3. For each element A of GLn(k), there is an element B of GLn(k) such that A = tBB(q). If A is a Hermitian matrix, then the matrix B can be taken from GLn(F q2). By Proposition 3.3, it follows immediately that a smooth k-Hermitian q2) to (resp. Hermitian) surface is projectively isomorphic over k (resp. F the Fermat surface XI . We define the set M := (cid:26) DB := (cid:18) 0 −b1 b1 0 0 −b2 Then the following map is surjective: b2 0(cid:19) ∈ GL4(k)(cid:12)(cid:12)(cid:12)(cid:12) B = (cid:0)b1 b2(cid:1) ∈ GL2(k)(cid:27) . (5) S∗ ∋ F ∗ 7−→ tF ∗AF ∗(q) ∈ M. In fact, by Proposition 3.3 there is an element D of GL4(k) such that DB = tDD(q) for each DB ∈ M . Similarly there is an element A′ of GL4(k) such that A = tA′A′(q). Hence putting F ∗ := A′−1D, one has tF ∗AF ∗(q) = DB, and thus F ∗ ∈ S∗. Lemma 3.4. The set R is nonempty, and each element of R is projectively isomorphic over k to the smooth curve C0 := (cid:8)t(sq+1, sqt, stq, tq+1) ∈ P3 t(s, t) ∈ P1(cid:9) . Proof. The set S∗ is nonempty by the surjectivity of the map (5). Hence by (4) the set R is nonempty. For each element CF of R, it is obvious by definition that F ∗−1F = (e1, e2, 0, . . . , 0, e3, e4) with (e1, e2, e3, e4) = I. This implies that CF is projectively isomorphic over k to C0. Then by definition, the curve C0 is smooth clearly. 8 NORIFUMI OJIRO Remark 3.5. It is known that each nonplanar nonreflexive curve of degree q + 1 is projectively isomorphic to the curve C0 (cf. [1, Theorem 2]). For nonreflexive curves, see also [5]. Hence by Lemma 3.4, each element of R is projectively isomorphic to each nonplanar nonreflexive curve of degree q + 1. Remark 3.6. In the case where A = I, we can find an element of R. We put (cid:3) J := (cid:18)0 −1 0 (cid:19) . 1 ∗ such as q2) such that tFJ η−qξq 0 0 Then the matrix DJ is a Hermitian matrix. Hence by Proposition 3.3, there ∗(q) = DJ . Actually taking ∗FJ is an element FJ FJ   ∗ of GL4(F   ωη−1ξ 0 0 ωη−1 0 0 −η−q 1 0 0 1 0 0 for ω, ξ and η as mentioned in Introduction, one has by (4) the corresponding curve CFJ lying on XI . 4. Proof of Theorem 1.2 The group Aut(XA) of projective automorphisms of XA is equal to {Q ∈ GL4(k) tQAQ(q) = λA, λ ∈ k×}/k×I. By Proposition 3.3, the group Aut(XA) is conjugate to Aut(XI ) in PGL4(k). We prove the following lemma on matrix groups of arbitrary rank because we need the lemma to our proof of Theorem 1.2. Lemma 4.1. Let n be a positive integer. The group PGUn(F phic to q2) is isomor- G := {Q ∈ GLn(k) tQQ(q) = λI, λ ∈ k×}/k×I. Proof. We consider the map G ∋ Qk× 7−→ ξλQµq+1 ∈ PGUn(F q2), where λ is the element of k× satisfying tQQ(q) = λI and ξλ is an element q+1 = λ−1. Then the map is well-defined. In fact, it is of k× satisfying ξλ obvious that t(ξλQ)(ξλQ)(q) = I, and the matrix ξλQ has the entries in F q2 q2). because I is a Hermitian matrix. Hence ξλQµq+1 belongs to PGUn(F Further, putting P := αQ for each α ∈ k×, one has tP P (q) = αq+1λI. It is easily shown by definition that ξαq+1λµq+1 = ξαq+1ξλµq+1 and αξαq+1 µq+1 = µq+1. Therefore we conclude that ξαq+1λP µq+1 = ξλQµq+1. RATIONAL CURVES ON A SMOOTH HERMITIAN SURFACE 9 Thus the map is independent of the choice of representatives for G. Let Q′k× be an element of G with tQ′Q′(q) = ηI for some η ∈ k×. Then one has (ξηQ′µq+1)(ξλQµq+1) = ξηλQ′Qµq+1, since ξηξλµq+1 = ξηλµq+1. Hence the map is a homomorphism from G to PGUn(F q2). The injectivity and the surjectivity are immediate from definition. (cid:3) By Lemma 4.1, the group Aut(XA) isomorphic to PGU4(F The following lemma is a key ingredient in our proof of Theorem 1.2. q2). Lemma 4.2. For every g, B ∈ GL2(k), one has tϕ∗(g)DBϕ∗(g)(q) = det(g)qDtgBg(q2 ). Proof. The proof is due to straightforward computation. We put g := (cid:18)α β δ(cid:19) , B := (b1, b2). γ Then one has tϕ∗(g)DB ϕ∗(g)(q) = (cid:18)αq tg γq tg βq tg δq tg(cid:19)(cid:18) 0 −b1 b1 = (cid:18)−γq tgb1 αq tgb1 −γq tgb2 αq tgb2 −δq tgb1 βq tgb1 −δq tgb2 βq tgb2(cid:19) 0 b2 0 −b2 0(cid:19)(cid:18)αq2 γq2   g(q) g(q) βq2 g(q)(cid:19) δq2 g(q) αq2+q αq2 γq αq2 αq2 αqγq2 γq2+q βq αqβq2 γqβq2 δq αqδq2 βqγq2 γqδq2 δqγq2 βq2+q δqβq2 βqδq2 δq2+q   . Putting tϕ∗(g)DBϕ∗(g)(q) := (cid:18)c1 c2 c3 c4 c5 c6 c7 c8(cid:19) , 10 one has NORIFUMI OJIRO c1 = −αq2+qγq tgb1 + αq2 γqαq tgb1 − αqγq2 γq tgb2 + γq2+qαq tgb2 = 0, c2 = −αq2 βqγq tgb1 + αq2 δqαq tgb1 − βqγq2 γq tgb2 + δqγq2 αq tgb2 = det(g)q(αq2 tgb1 + γq2 tgb2) , γq2 = det(g)q tg(b1, b2) t(αq2 ), αq tgb1 − αqδq2 γq tgb1 + γqβq2 c3 = −αqβq2 = 0, γq tgb2 + γqδq2 αq tgb2 c4 = −βq2+qγq tgb1 + δqβq2 αq tgb1 − βqδq2 γq tgb2 + δq2+qαq tgb2 = det(g)q(βq2 tgb1 + δq2 tgb2) , δq2 = det(g)q tg(b1, b2) t(βq2 ), c5 = −αq2+qδq tgb1 + αq2 γqβq tgb1 − αqγq2 δq tgb2 + γq2+qβq tgb2 = −det(g)q(αq2 tgb1 + γq2 tgb2) = −det(g)q tg(b1, b2) t(αq2 , γq2 ), c6 = −αq2 βqδq tgb1 + αq2 δqβq tgb1 − βqγq2 δq tgb2 + δqγq2 βq tgb2 = 0, c7 = −αqβq2 δq tgb1 + γqβq2 βq tgb1 − αqδq2 δq tgb2 + γqδq2 βq tgb2 = −det(g)q(βq2 tgb1 + δq2 tgb2) = −det(g)q tg(b1, b2) t(βq2 , δq2 ), c8 = −βq2+qδq tgb1 + δqβq2 βq tgb1 − βqδq2 δq tgb2 + δq2+qβq tgb2 = 0. Hence one has (c2, c4) = det(g)q tgBg(q2) = −(c5, c7), c1 = c3 = c6 = c8 = 0. This completes the proof. (cid:3) Proof of Theorem 1.2. We define an equivalence relation ∼ on the set M as follows: DB ∼ DB ′ for DB, DB ′ ∈ M if there is an element g ∈ GL2(k) such that DB ′ = tϕ∗(g)DBϕ∗(g)(q). We denote by DB ϕ∗ an equivalence class con- taining DB. On the other hand, the group Aut(XA) acts on k×\S∗/Im(ϕ)∗ by multiplication from the left. Then the following map is bijective: Aut(XA)k×\S∗/Im(ϕ)∗ −→ ∈ k×\M/ ∼ ∈ Aut(XA)k×F ∗Im(ϕ)∗ 7−→ k×(tF ∗AF ∗(q))ϕ∗ . Indeed, the surjectivity is obvious since the map (5) is surjective. assume that k×(tF ∗AF ∗(q))ϕ∗ = k×(tF1 ∗(q))ϕ∗ for some F1 ∗AF1 If we ∗ ∈ S∗, then RATIONAL CURVES ON A SMOOTH HERMITIAN SURFACE 11 we have ∗ϕ∗(g)F ∗−1) (q) t(F1 ∗ϕ∗(g)F ∗−1)A(F1 = λA ∗ϕ∗(g)F ∗−1 belongs to for some g ∈ GL2(k) and λ ∈ k×. Therefore k×F1 Aut(XA). This implies the injectivity, and thus bijectivity. By Proposition 3.3, there is an element B′ of GL2(k) such that B = tB′B′(q2) for each DB ∈ M . Then by Lemma 4.2, one has tϕ∗(B′−1)DBϕ∗(B′−1) (q) q = det(B′−1) DI . ϕ∗. Hence k×\M/ ∼ = 1 and thus This implies that k×DB Aut(XA)k×\S∗/Im(ϕ)∗ = 1, and by (4) one has Aut(XA)\R = 1. This proves half of our theorem. ϕ∗ = k×DI Let Γ/k×I be the stabilizer subgroup of Aut(XA) fixing the element ∗(q) = DI . Then it fol- ∗Im(ϕ)∗ of k×\S∗/Im(ϕ)∗ such that tFI ∗AFI k×FI lows immediately that ∗Im(ϕ)∗FI Γ = FI ∗−1 ∩ {Q ∈ GL4(k) tQAQ(q) = λA, λ ∈ k×}. Hence each element of Γ can be written as FI g of GL2(k) satisfying ∗ϕ∗(g)FI ∗−1)A(FI ∗ϕ∗(g)FI t(FI ∗−1)(q) = λA for λ ∈ k×, ∗ϕ∗(g)FI ∗−1 for some element or equivalently, tϕ∗(g)DI ϕ∗(g)(q) = λDI for λ ∈ k×. By Lemma 4.2, this equality is equivalent to tgg(q2) = λI for λ ∈ k×. Con- sequently, one has the following isomorphism: {g ∈ GL2(k) tgg(q2) = λI, λ ∈ k×}/k×I −→ Γ/k×I ∈ gk× By Lemma 4.1, we conclude that PGU2(F ∈ ∗ϕ∗(g)FI ∗−1k×. 7−→ FI q4) ≃ Γ/k×I. (cid:3) References [1] E. Ballico and A. Hefez, Nonreflexive projective curves of low degree, Manuscripta Math. 70(4):385-396, 1991. [2] A. Beauville, Sur les hypersurfaces dont les sections hyperplanes sont `a module constant, With an appendix by D. Eisenbud and C. Huneke, The Grothendieck Festschrift, Vol. I, (ed. P. Cartier, L. Illusie, N. M. Katz, G. Laumon, Yu. Manin and K. A. Ribet), Progress in Mathematics. 86 (Birkhauser, Boston) 121-133, 1990. [3] R. C. Bose and I. M. Chakravarti, Hermitian varieties in a finite projective space PG(N, q2), Canad. J. Math. 18:1161-1182, 1966. [4] The GAP Group, GAP - Groups, Algorithms, and Programming, Version 4.8.8; 2017. (http://www.gap-system.org). [5] A. Hefez, Nonreflexive curves, Compositio Math. 69(1):3-35, 1989. 12 NORIFUMI OJIRO [6] J. W. P. Hirschfeld and J. A. Thas, General Galois geometries, Oxford Math- ematical Monographs, Oxford Science Publications, The Clarendon Press, Ox- ford University Press, New York, 1991. [7] T. H. Hoang, Degeneration of Fermat hypersurfaces in positive characteristic, Hiroshima Math. J. 46(2):195-215, 2016. [8] B. Segre, Forme e geometrie hermitiane, con particolare riguardo al caso finito, Ann. Mat. Pura Appl. (4), 70:1-201, 1965. [9] I. Shimada, Lattices of algebraic cycles on Fermat varieties in positive charac- teristics, Proc. London Math. Soc. (3), 82(1):131-172, 2001. [10] I. Shimada, A note on rational normal curves totally tangent to a Hermitian variety, Des. Codes Cryptogr. 69(3):299-303, 2013. [11] I. Shimada, The graphs of Hoffman-Singleton, Higman-Sims and McLaughlin, and the Hermitian curve of degree 6 in characteristic 5, Australas. J. Combin. 59:161-181, 2014. Department of Mathematics, Graduate School of Science, Hiroshima Uni- versity, 1-3-1, Kagamiyama, Higashi-Hiroshima, Hiroshima, 739-8526, Japan E-mail address: [email protected] 

1710.05412

1

1710

The Containment Poset of Type $A$ Hessenberg Varieties

Flag varieties are well-known algebraic varieties with many important geometric, combinatorial, and representation theoretic properties. A Hessenberg variety is a subvariety of a flag variety identified by two parameters: an element $X$ of the Lie algebra $\mathfrak{g}$ and a Hessenberg subspace $H\subseteq \mathfrak{g}$. This paper considers when two Hessenberg spaces define the same Hessenberg variety when paired with $X$. To answer this question we present the containment poset $\mathcal{P}_X$ of type $A$ Hessenberg varieties with a fixed first parameter $X$ and prove directly that if $X$ is not a multiple of the element $\bf 1$ then the Hessenberg spaces containing the Borel subalgebra determine distinct Hessenberg varieties. Lastly we give a natural involution on $\mathcal{P}_X$ that induces a homeomorphism of varieties and prove additional properties of $\mathcal{P}_X$ when $X$ is a regular nilpotent element.

math.AG

math

THE CONTAINMENT POSET OF TYPE A HESSENBERG VARIETIES ELIZABETH DRELLICH Abstract. Flag varieties are well-known algebraic varieties with many im- portant geometric, combinatorial, and representation theoretic properties. A Hessenberg variety is a subvariety of a flag variety identified by two param- eters: an element X of the Lie algebra g and a Hessenberg subspace H ⊆ g. This paper considers when two Hessenberg spaces define the same Hessenberg variety when paired with X. To answer this question we present the contain- ment poset PX of type A Hessenberg varieties with a fixed first parameter X and prove directly that if X is not a multiple of the element 1 then the Hes- senberg spaces containing the Borel subalgebra determine distinct Hessenberg varieties. Lastly we give a natural involution on PX that induces a homeomor- phism of varieties and prove additional properties of PX when X is a regular nilpotent element. 1. Introduction A Hessenberg variety is a subvariety of a flag variety G/B determined by two parameters: X, an element of the underlying Lie algebra g, and a Hessenberg sub- space H ⊆ g (see Definition 3 ). Any such pair of parameters defines a Hessenberg variety Hess(X, H) = {gB ∈ G/B : Ad(g−1)X ∈ H}. Hessenberg varieties appear in a number of mathematical fields under a number of names. For example Springer fibers, originally defined by Springer to construct representations of Weyl groups, can be expressed as Spr(X) = Hess(X, b) [12]. Peterson varieties first appeared in the construction of the quantum cohomology of the Grassmannian [9]. Wachs and Shareshian's work on quasi-symmetric functions uses Hessenberg varieties to geometrically encode combinatorial data [11]. Of course Hessenberg varieties are also interesting objects in their own right. People have studied the structure of certain subfamilies, like the Peterson varieties [8], as well as how Hessenberg varieties intersect with certain Schubert varieties [7]. More generally the structure of Hessenberg varieties can be studied by paving them with affines [10, 13, 15]. Some Hessenberg varieties also carry non-trivial torus actions [3], leading to the study of the torus equivariant cohomology of some Hessenberg varieties, such as those where X is either a regular nilpotent element[1, 4, 6] or a regular semisimple element of g [2]. Despite their usefulness, it can be quite difficult to describe the structure of Hessenberg varieties directly. Even something as basic as determining when two Hessenberg varieties are equal can be non-trivial. Definition. [13, Def 1.6] Two Hessenberg spaces H1 and H2 are X-equivalent if Hess(X, H1) = Hess(X, H2). 1 2 ELIZABETH DRELLICH While a few X-equivalent spaces are immediate, for example all Hessenberg spaces are 0- and 1-equivalent, non-trivial examples can be harder to identify. To determine where such examples exist, we define the containment poset PX of Hessenberg varieties with a fixed X. The containment poset on Schubert varieties is well known to be the Bruhat order on the corresponding Weyl group. The posets PX vary widely depending on the choice of X (see Examples 12, 13, and Figure 1). Nevertheless an interval at the top of the poset is the same for most choices of X. Our main theorem (Theorem 14) shows that if X ∈ g is not a multiple of 1 then no two Hessenberg spaces containing b are X-equivalent. In Section 2 we present the necessary definitions and background on Hessen- berg varieties. Section 3 presents our motivating question on the existence of X- equivalent Hessenberg spaces and defines the poset PX of Hessenberg varieties. We also present some of the diversity of structures PX can have. The main theorem, that if b ⊆ H and X 6= a1 then Hess(X, H) is uniquely determined by H, appears in Section 4 where we also give a constructive proof. The last two sections of this paper present additional homeomorphisms between non-equal Hessenberg varieties and the implications of those homeomorphisms to the poset PX . Section 5 proves that the bilateral symmetry of the containment poset of Hessenberg spaces induces a homeomorphism of the corresponding vari- eties. In Section 6 we discuss which Hessenberg varieties are homeomorphic to products of other Hessenberg varieties and conclude with a condition for when regular nilpotent Hessenberg varieties are decomposable in this sense. All of the indecomposable regular nilpotent Hessenberg varieties are contained in a closed interval of PX from the Peterson variety to the full flag variety. The author would like to thank Julianna Tymoczko and Stephen Oloo for their helpful comments at multiple stages of this work, and Martha Precup for several inspiring and informative conversations. 2. definitions Throughout this paper we will refer to a complex reductive linear algebraic group G, a fixed Borel subgroup B ⊆ G, and their corresponding Lie algebras g and b, and a root system Φ. This construction also results in a flag variety G/B and a Weyl group W . Many of the proofs and examples in this paper rely on explicit constructions of individual full flags in the flag variety. A (full) flag F• is sequence of nested subspaces in an ambient n dimensional vector space V {0} = F0 ( F1 ( · · · ( Fn = V. It will be convenient to think of G, B, and F• as explicitly as possible. Throught this paper when discussing type An−1 Hessenberg varieties we will assume G = Gln(C) and B the upper triangular matrices in G. Furthermore a flag F• will be represented as a matrix where the subspace Fi is the span of the first i columns. Definition 1. A strict Hessenberg space is a subspace H ⊆ g such that b ⊆ H and [H, b] ⊆ H. This definition can be relaxed to allow for spaces that do not contain the Borel sub-Lie algebra and as we will see in Theorem 14, having b contained in H has significant implications for what Hessenberg varieties can occur. There is a natural poset on the collection of strict Hessenberg spaces ordered by inclusion. The unique maximal element of this poset is g, the unique minimal THE CONTAINMENT POSET OF TYPE A HESSENBERG VARIETIES 3 element is b. Any strict Hessenberg space H can be defined using a subset MH of the negative roots in the corresponding root system, i.e. H = b ⊕ Mα∈MH gα. The space H defined as above is a Hessenberg space if and only if whenever α ∈ MH and another negative root β is above α in the root lattice, β is also in MH . In type An−1 we use two additional ways to describe Hessenberg spaces: Hes- senberg functions and matrix shape. A strict Hessenberg function h : [n] → [n] has h(i) ≥ max{i, h(i − 1)} for all i, and the root −αi − αi+1 − · · · − αj−1 is in MH if and only if h(i) ≥ j. As a space of matrices strict Hessenberg spaces have a staircase below the diagonal, below which all entries must be zero. A zero in the (i, j)th entry means that −αi − αi+1 − · · · − αj−1 is not in MH . Example 2. There are five strict Hessenberg spaces in g in type A2. H1 = b H2 H3 H4 H5 = g ∗ ∗ 0 ∗ 0 0   ∗ ∗ ∗  ∗ ∗ 0   ∗ ∗ ∗ ∗ 0 ∗  ∗ 0 0   ∗ ∗ ∗ ∗ ∗ ∗  ∗ ∗ ∗ ∗ 0 ∗   ∗ ∗ ∗  ∗ ∗ ∗   ∗ ∗ ∗ ∗ ∗ ∗  h MH 17→1 27→2 37→3 ∅ 17→2 27→2 37→3 17→1 27→3 37→3 17→2 27→3 37→3 {−α1} {−α2} {−α1, −α2} 17→3 27→3 37→3 Φ− The introduction of the Hessenberg function h leads to a broader concept of a Hessenberg space. Definition 3. Given by an non-decreasing function h : [n] → [n], a (generalized) Hessenberg space is a subspace of H ⊆ g given by H = Mh(i)≤j gi,j. We will refer to the spaces defined in the previous two definitions as Hessenberg spaces, using the descriptors primarily to emphasize when strict Hessenberg spaces have special properties. Any one of these definitions of a Hessenberg space can be used to define a Hessenberg variety. Definition 4. Fix a Hessenberg space H and an operator X ∈ g. The Hessenberg variety Hess(X, H) is (1) Hess(X, H) = {gB ∈ G/B : Ad(g−1)X ∈ H}. The Hessenberg varieties form a very large family of algebraic varieties and many well-know varieties can be expressed as Hessenbergs. For example the full flag variety G/B = Hess(X, g) for any X ∈ g but if X is a regular nilpotent element in g, then Hess(X, b) is a single point. 4 ELIZABETH DRELLICH Remark 5. An immediate consequence of the definition of a Hessenberg variety is that if Y = P XP −1 is similar to X then for any Hessenberg space H, Hess(X, H) is homeomorphic to to Hess(Y, H) [14, Prop. 2.7]. While the homeomorphism described above will be used in Theroem 18, unless otherwise specified all operators X represent their entire similarity classes. It will be useful for the purpose of explicit calculations to fix a particular regular nilpotent operator. Definition 6. Fix a basis {Eα} for each subspace gα ⊆ g. Define a regular nilpotent operator N as (2) Eα. N = Xα∈∆ In type An−1 using the standard basis {E(i,j)}, the operator N = the Jordan normal form of the nilpotent matrix with only one Jordan block. E(i,i+1) is n−1 Pi=1 Example 7. Let G = GL3(C), B be the upper triangular matrices in G, and H be the Hessenberg space defined by h(1) = 2, h(2) = h(3) = 3 (H4 from Example 2). There are number of ways to think about this variety: • From the definition, Hess(N, H) = {gB ∈ G/B : Ad(g−1)N ∈ H}. • As flags, Hess(N, H) = {F• ∈ G/B : N Fi ∈ Fi+1}. • Explicitly, these are the flags in the set 1 0 0 1 0 0     0 0 1  ,  a 1 0 1 0 0 0 0 1  ,  0 1 0 0 a 1 0 1 0  ,  a b 1 b 1 0 . 1 0 0    • Geometrically Hess(N, H) consists of two copies of C and one copy of C2 joined together at a single point. The variety Hess(N, H) in the previous example is the type A2 Peterson variety. Peterson varieties are the best studied regular nilpotent Hessenberg varieties and their singularities [8] and cohomology [4] are understood. In general Lie type, the Peterson variety is the regular nilpotent Hessenberg variety corresponding to the Hessenberg space In the sense that will be discussed in Section 6 the Peterson variety is the smallest regular nilpotent Hessenberg variety in type A. H = b ⊕ Mα∈−∆ gα. 3. Posets of Hessenberg Varieties Tymoczko gave the following definition for X-equivalence of Hessenberg spaces. Definition 8. [13, Def 1.6] Two Hessenberg spaces H1 and H2 are called X- equivalent if Hess(X, H1) = Hess(X, H2). There are two immediate occurrences of X-equivalence: all strict Hessenberg spaces are 1-equivalent since Hess(1, H) = G/B for any strict Hessenberg space H. Similarly all Hessenberg spaces are 0-equivalent. The search for non-trivial exam- ples of X-equivalent strict Hessenberg spaces led the author to the main theorem THE CONTAINMENT POSET OF TYPE A HESSENBERG VARIETIES 5 of this paper, namely that no such examples exist. However if we loosen our study to include all (generalized) Hessenberg spaces many non-trivial examples appear. Example 9. Consider the operator X and subspaces H1 and H2 defined by 1 0 0 0 X =  0 0 1 0 0 0 0 0 0 0 0 0   , H1 =  0 0 0 0 ∗ ∗ 0 ∗ 0 ∗ 0 ∗ ∗ ∗ ∗ ∗   , H2 =  0 0 0 0 0 0 0 0 ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗ .   Consider a flag F• ∈ Hess(X, H1). Using standard basis notation, F1 must be in the span of {e3, e4} since it must be in the null space of X. But since XF2 must also be a subspace of the span of {e3, e4}, the subspace F2 must itself be contained in that same span. There are no constraints on F3 or F4. This flag F• is, however, also contained in Hess(X, H2) since F2 is in the null space of X. Therefore Hess(X, H1) = Hess(X, H2). The collection of Hessenberg subspaces of any Lie algebra naturally form a poset ordered by containment. In type An−1 this poset is isomorphic to the closed interval in Young's lattice from ∅ to nn with the strict Hessenberg spaces making up the interval from ∅ to (n, n − 1, ..., 1). This can be seen using the notation in Example 2 and reading each of the zeros as a box in the Young diagram using the French notation: ∗ 0 0   ∗ ∗ 0 ∗ ∗ ∗  It will sometimes be notationally convenient to identify a Hessenberg space Hλ by the corresponding Young diagram λ. This identification leads immediately to an ordering on families of Hessenberg varieties. Definition 10. For any X ∈ g define the poset PX on the Hessenberg vari- eties Hess(X, H) by containment, i.e. Hess(X, H1) (cid:22) Hess(X, H2) if the variety Hess(X, H1) is contained in the variety Hess(X, H2). This poset of Hessenberg varieties inherits some of the structure of the contain- ment poset on Hessenberg spaces. Lemma 11. If H1 ⊆ H2 are Hessenberg spaces then Hess(X, H1) (cid:22) Hess(X, H2) for every X ∈ g. Proof. if H1 ⊆ H2 then Hess(X, H1) (cid:22) Hess(X, H2) since if Ad(g−1)X ∈ H1 and H1 ⊆ H2 then Ad(g−1)X ∈ H2. (cid:3) So while it may be clear when one Hessenberg variety precedes another in PX it is less clear when, or if, the poset PX will be isomorphic to the interval [∅, nn] of Young's lattice. We demonstrate with two examples. Example 12. Let X = (cid:20)1 0 0 0(cid:21). Let eB and s1B be the elements in GL2(C)/B given by the identity matrix and the inversion (1, 2) respectively. Then the five (generalized) Hessenberg spaces give the varieties: • Hess(X, H∅) = Gl2(C)/B • Hess(X, H(1)) = {eB, s1B} 6 ELIZABETH DRELLICH Figure 1. Two posets PX . Here λ denotes the Hessenberg variety Hess(X, Hλ). ∅ = = = ∅ = (a) X = (cid:20)1 0 0 0(cid:21). (b) X = (cid:20)0 0 1 0(cid:21). • Hess(X, H(1,1)) = {s1B} • Hess(X, H(2)) = {eB} • Hess(X, H(2,1)) = Hess(X, H(2,2)) = ∅ Example 13. Let X = (cid:20)0 1 0 0(cid:21). Again let eB be the element in GL2(C)/B de- scribed above. Then the five (generalized) Hessenberg spaces give the varieties: • Hess(X, H∅) = Gl2(C)/B • Hess(X, H(1)) = Hess(X, H(1,1)) = Hess(X, H(2)) = Hess(X, H(2,1)) = {eB} • Hess(X, H(2,2)) = ∅ We visualize the posets PX for these two examples in Figure 1. For aesthetic reasons we identify the variety Hess(X, Hλ) to the diagram λ. While these examples are small, observe that no two varieties in the interval from ∅ to are equal. These are the varieties defined by strict Hessenberg spaces. 4. Main Theorem In this section we prove our main theorem, that no two strict Hessenberg spaces can be X-equivalent unless X is a multiple of the identity element of g. Theorem 14. Let X ∈ g be an element that is not equal to a1, and H1, H2 be distinct strict Hessenberg spaces. Then Hess(X, H1) 6= Hess(X, H2). We will prove this directly by showing that for each negative root α ∈ Φ− we can explicitly construct a flag that is in Hess(X, H) if and only if α ∈ MH. Since we are in type A, root α = −αi − αi+1 − · · · − αj ∈ MH is equivalent to H having Hessenberg function with h(i) ≥ j. Before constructing a general proof, we state a useful lemma and give an illustrative example. Lemma 15. Let X ∈ g be an operator, i ≤ j ≤ n, and F• a flag in Cn with the following properties: (1) XFk ⊆ Fk whenever k < i or k > j THE CONTAINMENT POSET OF TYPE A HESSENBERG VARIETIES 7 (2) XFi ⊆ Fj (3) XFi is not a subspace of Fj−1. Then the flag F• is in Hess(X, H) if and only if the Hessenberg function for H has h(i) ≥ j. Proof. If h(i) ≥ j then the flag F• as defined above is in Hess(X, H) by definition. If the flag F• is in Hess(X, H) then because XFi is not a subspace of Fj−1, we must have h(i) ≥ j. (cid:3) The following example illustrates a flag that satisfies the properties of Lemma 15. Example 16. Consider the operator X = λ1   1 λ1 1 λ1 λ2 1 λ2 λ3 .   To construct a flag which is contained in Hess(X, H) if and only if the Hessenberg function has h(2) ≥ 4, we need a flag for which XF2 is in F4 but not F3, and has XFi ⊂ Fi for i = 1, 4, 5, 6. One such example is A2,4 =(cid:2)e4, e2, e5, e1, e6, e3(cid:3) . Note that e1, e4, and e6 are eigenvectors of X but e2, which is contained in F2, is not. The flag A2,4 is in Hess(X, H) if and only if h(2) ≥ 4. Using the same X,but h(5) = 6 we can build the flag which is in Hess(X, H) if and only if h(5) ≥ 6. A5,6 =(cid:2)e4, e5, e6, e1, e3, e2(cid:3) We now give a method for constructing such flags Ai,j for any given operator X 6= a1. This construction proves the main theorem of this paper. Proof of Theorem 14. Without loss of generality we may assume that X is an n × n matrix in Jordan normal form with blocks of size µ1 ≥ µ2 ≥ · · · µm ≥ 1. Let {ei}i≤n be the standard basis of Cn. Since X is in Jordan normal form, Xei ∈ span{ei, ei−1} for all i. In order to construct our flags explicitly, we need to consider two separate cases: when µ1 > 1 and when µ1 = 1. The operator X has a Jordan block of size µ1 > 1. If the first Jordan block of X has size µ1 > 1 then Xe2 6∈ span{e2} but Xe2 is in the span of {e1, e2}. Similarly for all k = 2, 3, ..., µ1, Xek is in the span of {ek−1, ek} but not {ek}. To build a flag F• that has the properties required by Lemma 15, first we reindex the basis vectors as {vk}k≤n by ek+µ1 ek−n+µ1+3 e2 e1 if k ≤ n − µ1 if k ≥ n − µ1 if k = n − 1 if k = n. vk =  8 ELIZABETH DRELLICH This re-ordering has the property that unless k = n − µ1 or n − 1, then Xvk is in the span of {vk, vk−1}. If i ≤ n − µ1 define F• by the matrix A ∈ Gln(C) with kth column Ak given by: Ak = vk vn−1 = e2 vk−1 vn = e1 vk−2 if k < i if k = i if i < k < j if k = j if j < k.   Since i is less than n − µ1, we have that XFk ⊂ Fk unless i ≤ k ≤ j − 1, in which case XFk is contained in Fj but not Fj−1. If i > n − µ1 define F• by the matrix A ∈ Gln(C) with kth column Ak given by: Ak = vk ek−n+µ1 ek−n+µ1+1 ei−n+µ1 ek−n+µ1 if k ≤ n − µ1 if n − µ1 < k < i if i ≤ k =< j if k = j if j < k.   Again the flag defined by A has all the properties required by Lemma 15 and thus is contained in Hess(X, H) if and only if the Hessenberg function of H has h(i) ≥ j. If all Jordan blocks are size 1 and X is not All Jordan blocks of X are size 1. a multiple of the identity matrix, then there are two standard basis vectors that have different eigenvalues λ1 and λ2. Without loss of generality we may assume that those are vectors e1 and e2. In a similar fashion to the case where X has a Jordan block of size greater than one, we reorder the standard basis as follows and create a flag A using a matrix in which the kth column Ak is given by: Ak = ek+2 e1 + e2 ek+1 e1 ek   if k < i if k = i if i < k < j if k = j if j < k. This flag is contained in Hess(X, H) if and only if h(i) ≥ j, completing the proof. (cid:3) Unsurprisingly, the constructions above point to methods for constructing large families of flags that are in a variety Hess(X, H) if and only if h(i) ≥ j, however the particular flags constructed have several advantages: if X has a nilpotent part, then the flag A is a permutation matrix. Furthermore if X is actually a nilpotent operator then A is a fixed point of the one dimensional torus action on Hess(X, H). THE CONTAINMENT POSET OF TYPE A HESSENBERG VARIETIES 9 5. An Involution on PX The bilateral symmetry that the poset of type A Hessenberg spaces inherits from Young's lattice leads to a natural question: does that bilateral symmetry have implications for the Hessenberg varieties? The answer is yes, the respective varieties are homeomorphic to each other. Proposition 17. If H is a Hessenberg space and T H ⊆ g is the set of matrices in g obtained by flipping the matrix form of H across its antidiagonal, then T H is also a Hessenberg space. Proof. If H = Hλ for some Young diagram λ then T H = HλT which is also a Hessenberg space. (cid:3) Matrix manipulation shows that, when w0 is the permutation matrix with ones on the antidiagonal, which corresponds to the longest word in W , any n × n matrix M can be flipped along its antidiagonal by taking its transpose and conjugating by the longest word: (3) T M = w0M T w0 and so for Hessenberg spaces: (4) T H = w0H T w0. The fact that both taking the transpose and conjugating by an element of the Weyl group are homeomorphisms of the Lie algebra will be useful in the proofs of this section's theorems. Theorem 18. Let Hλ ⊆ GLn(C) be a Hessenberg space and let T H = HλT be the Hessenberg space obtained by flipping along the antidiagonal. Then for any X ∈ g Hess(X, H) ∼= Hess(X, T H). Proof. Consider the homeomorphism (5) φ : G/B → G/B gB 7→ w0(gT )−1w0B. Direct computation shows that g−1Xg ∈ Hλ if and only if w0gT X T (gT )−1w0 is in w0H T w0 = HλT so the map φ restricts to a homeomorphism of Hessenberg varieties (6) Hess(X, Hλ) ∼= Hess(w0X T w0, HλT ). But we know from Remark 5 and the fact that X and w0X T w0 are similar matrices (if X was in Jordan normal form this is just a rearrangement of the blocks) that Hess(w0X T w0, HλT ) ∼= Hess(X, HλT ). Therefore Hess(X, HλT ) ∼= Hess(X, HλT ) for any X, Hλ and the natural involution on PX induces a homeomorphism of algebraic varieties. (cid:3) These two varieties are unlikely to be equal if λ 6= λT and Hess(X, Hλ) is non- empty. In Example 12 we saw that while both Hess(X, H(2)) and Hess(X, H(1,1)) are a single point, they are different points in G/B. 10 ELIZABETH DRELLICH 6. Decomposability: Regular Nilpotent Hessenberg Varieties The product of two type A flag varieties can be expressed as a Hessenberg variety inside a larger flag variety. For example if F1 = Gl3(C)/B and F2 = Gl2(C)/B and N is the regular nilpotent element given in Definition 6 then F1 × F2 ∼= Hess(N, H) ⊂ Gl5(C)/B where H has Hessenberg function h(1) = h(2) = h(3) = 3 and h(4) = h(5) = 5. In such a situation the structure of the Hessenberg variety is fully determined by the structures of F1 and F2. Definition 19. Fix a flag variety G/B and its corresponding Lie algebras g and b. A Hessenberg variety Hess(X, H) is decomposable if for some g1, g2 ( g • there exist X1 ∈ g1 and X2 ∈ g2 • there exist H1 ⊆ g1 and H2 ⊆ g2 such that Hess(X, H) ∼= Hess(X1, H1) × Hess(X2, H2). Note that in the example above where the two components into which the Hes- senberg variety decomposes are in fact full flag varieties, H1 = g1 and H2 = g2. Definition 19 does not preclude the possibility that one of the components is a single point. Theorem 20 ([5]). Let Hess(N, H) be a type An−1 regular nilpotent Hessenberg va- riety. If for some i < n, h(i) = i in the Hessenberg function of H then Hess(N, H) is decomposable. Proof. Let h : [n] → [n] be a Hessenberg function with h(j) = j for some j < n and let Hess(N, H) be the corresponding regular nilpotent Hessenberg variety. We define two new type-A Lie algebras and root systems by letting G1 = GLj(C) and G2 = GLn−j(C) and g1, g2 be their respective Lie algebras. For each, we define a Hessenberg function: (7) h1 : [j] → [j] 7→ h(i) i and h2 : [n − j] → [n − j] i 7→ h(i + j) − j These determine two regular nilpotent Hessenberg varieties Hess(N1, H1) and Hess(N2, H2) where N1 = N g1 and N2 = N g2 are regular nilpotent operators in g1 and g2 re- spectively. We will show that (8) Define a map (9) Hess(N, H) ∼= Hess(N1, H1) × Hess(N2, H2). Hess(N1, H1) × Hess(N2, H2) → Hess(N, H) • ⊕ V (2) 7→ V (1) , V (2) (V (1) ) • • • and V (2) If V (1) flag in Hess(N, H) where • • are flags in the two smaller Hessenberg varieties, then V• is the (10) Vi =(V (1) i j ⊕ V (2) V (1) i−j if i ≤ j if i > j THE CONTAINMENT POSET OF TYPE A HESSENBERG VARIETIES 11 In matrix notation V• ="V (1) • 0 ∗ V (2) • #. To see that V• ∈ Hess(N, H) we observe that (11) N Vi = N1V (1) N Vi ⊆ V (1) i ⊂ V (1) j ⊕ N2V (2) h1(i) = Vh(i) i−j ⊂ V (2) j ⊕ V (2) h2(i−j) = Vh(i) if i ≤ j if i > j. As a direct sum of linear operators, this map is injective. It remains to be shown that every flag V• in Hess(N, H) has this form. Let V• = V1 ( V2 ( · · · ( Vn = Cn be a flag in Hess(N, H). Let v ∈ Vj be a vector. Without loss of generality say that v = (v1, v2, . . . , vp, 0, . . . , 0) where vp is the last non-zero entry. By the definition of Hess(N, H) the vector N v is also in Vj as are all the vectors N kv for k a non-negative integer. Since vp is non-zero, the vectors N kv are non-zero when k is less than p. The collection of vectors {N kv : k = 0, 1, . . . , p − 1} is a linearly independent set in the space Vj . We know Vj has dimension j by definition of the flag. Therefore p is less than or equal to j. This means any vector v in Vj must be in the span of the first j basis elements. We conclude that Vj is equal to the span of the first j basis elements and that the matrix form of V• looks like (12) V• ="V (a) • 0 ∗ V (b) • # . Here we define the two smaller flags to be V (a) i = Vi+j /Vj. By the definition of h1 the flag V (a) • we have that N2V (b) i = N2(Vi+j /Vj) which is equal to (N Vi+j )/Vj as a subspace of Cn−j. Similarly V (b) h2(i) is equal to Vh(i+j)/Vj as a subspace of Cn−j. Therefore for any V• ∈ Hess(N, H) (13) N Vi+j ⊂ Vh(i+j) ⇐⇒ (N Vi+j )/Vj ⊂ Vh(i+j)/Vj ⇐⇒ N2V (b) is in Hess(N1, H1). For V (b) i = Vi and V (b) i ⊂ V (b) • h2(i). Thus every flag V• in Hess(N, H) is the product of a flag in Hess(N1, H1) and a flag in Hess(N2, H2). This process of decomposing the regular nilpotent Hessenberg variety into the product of smaller varieties can be repeated until each Hessenberg function preserves only the largest element of its domain. (cid:3) The Peterson variety, which is the regular nilpotent Hessenberg variety correspond- ing to the Hessenberg function h(i) = i+1 for i < n, is the smallest indecomposable regular nilpotent Hessenberg variety in type A. Proposition 21. The type A regular nilpotent Hessenberg varieties that are inde- composable are contained in a closed interval in the poset PN . The bottom element of this interval is the Peterson variety and the top element is the full flag variety. Proof. This is an immediate consequence of Theorem 20. (cid:3) References [1] Hiraku Abe, Megumi Harada, Tatsuya Horiguchi, and Mikiya Masuda. The co- in lie type A. preprint: regular nilpotent hessenberg varieties homology rings of https://arxiv.org/abs/1512.09072, 2016. 12 ELIZABETH DRELLICH [2] Hiraku Abe, Tatsuya Horiguchi, hessenberg of https://arxiv.org/abs/1704.00934, 2017. semisimple regular and Mikiya Masuda. The varieties for h = (h(1), n, ?, n). cohomology rings preprint: [3] F. De Mari, C. Procesi, and M. A. Shayman. Hessenberg varieties. Trans. Amer. Math. Soc., 332(2):529 -- 534, 1992. [4] Elizabeth Drellich. Monk's rule and Giambelli's formula for Peterson varieties of all Lie types. J. Algebraic Combin., 41(2):539 -- 575, 2015. [5] Elizabeth J Drellich. Combinatorics of Equivariant Cohomology: Flags and Regular Nilpotent Hessenberg Varieties. PhD thesis, University of Massachusetts, Amherst, 2015. [6] Megumi Harada and Julianna Tymoczko. Poset pinball, GKM-compatible subspaces, and Hessenberg varieties. J. Math. Soc. Japan, 69(3):945 -- 994, 2017. [7] Erik Insko and Julianna Tymoczko. Intersection theory of the Peterson variety and certain singularities of Schubert varieties. Geom. Dedicata, 180:95 -- 116, 2016. [8] Erik Insko and Alexander Yong. Patch ideals and Peterson varieties. Transform. Groups, 17(4):1011 -- 1036, 2012. [9] Dale Peterson. Quantum cohomology of g/p. Lecture Course, M. I. T., Spring Term 1997. [10] Martha Precup. Affine pavings of Hessenberg varieties for semisimple groups. Selecta Math. (N.S.), 19(4):903 -- 922, 2013. [11] John Shareshian and Michelle L. Wachs. Chromatic quasisymmetric functions and Hessenberg varieties. In Configuration spaces, volume 14 of CRM Series, pages 433 -- 460. Ed. Norm., Pisa, 2012. [12] T.A. Springer. A construction of representations of weyl groups. Inventiones mathematicae, 44:279 -- 293, 1978. [13] Julianna S. Tymoczko. Hessenberg varieties are not pure dimensional. Pure Appl. Math. Q., 2(3, Special Issue: In honor of Robert D. MacPherson. Part 1):779 -- 794, 2006. [14] Julianna S. Tymoczko. Linear conditions imposed on flag varieties. Amer. J. Math., 128(6):1587 -- 1604, 2006. [15] Julianna S. Tymoczko. Paving Hessenberg varieties by affines. Selecta Math. (N.S.), 13(2):353 -- 367, 2007. 

1108.5367

2

1108

Isospectral commuting variety, the Harish-Chandra D-module, and principal nilpotent pairs

Let g be a complex reductive Lie algebra with Cartan algebra h. Hotta and Kashiwara defined a holonomic D-module M, on g x h, called Harish-Chandra module. We relate gr(M), an associated graded module with respect to a canonical Hodge filtration on M, to the isospectral commuting variety, a subvariety of g x g x h x h which is a ramified cover of the variety of pairs of commuting elements of g. Our main result establishes an isomorphism of gr(M) with the structure sheaf of X_norm, the normalization of the isospectral commuting variety. It follows, thanks to the theory of Hodge modules, that the normalization of the isospectral commuting variety is Cohen-Macaulay and Gorenstein, confirming a conjecture of M. Haiman. We deduce, using Saito's theory of Hodge D-modules, that the scheme X_norm is Cohen-Macaulay and Gorenstein. This confirms a conjecture of M. Haiman. Associated with any principal nilpotent pair in g, there is a finite subscheme of X_norm. The corresponding coordinate ring is a bigraded finite dimensional Gorenstein algebra that affords the regular representation of the Weyl group. The socle of that algebra is a 1-dimensional vector space generated by a remarkable W-harmonic polynomial on h x h. In the special case where g=gl_n the above algebras are closely related to the n!-theorem of Haiman and our W-harmonic polynomial reduces to the Garsia-Haiman polynomial. Furthermore, in the gl_n case, the sheaf gr(M) gives rise to a vector bundle on the Hilbert scheme of n points in C^2 that turns out to be isomorphic to the Procesi bundle. Our results were used by I. Gordon to obtain a new proof of positivity of the Kostka-Macdonald polynomials established earlier by Haiman.

math.AG

math

ISOSPECTRAL COMMUTING VARIETY, THE HARISH-CHANDRA D-MODULE, AND PRINCIPAL NILPOTENT PAIRS VICTOR GINZBURG ABSTRACT. Let g be a complex reductive Lie algebra with Cartan algebra t. Hotta and Kashi- wara defined a holonomic D-module M, on g × t, called Harish-Chandra module. We relate gr M, an associated graded module with respect to a canonical Hodge filtration on M, to the isospectral commuting variety, a subvariety of g × g × t × t which is a ramified cover of the vari- ety of pairs of commuting elements of g. Our main result establishes an isomorphism of gr M with the structure sheaf of Xnorm, the normalization of the isospectral commuting variety. We deduce, using Saito's theory of Hodge D-modules, that the scheme Xnorm is Cohen-Macaulay and Gorenstein. This confirms a conjecture of M. Haiman. Associated with any principal nilpotent pair in g, there is a finite subscheme of Xnorm. The corresponding coordinate ring is a bigraded finite dimensional Gorenstein algebra that af- fords the regular representation of the Weyl group. The socle of that algebra is a 1-dimensional space generated by a remarkable W -harmonic polynomial on t × t. In the special case where g = gln the above algebras are closely related to the n!-theorem of Haiman and our W - harmonic polynomial reduces to the Garsia-Haiman polynomial. Furthermore, in the gln case, the sheaf gr M gives rise to a vector bundle on the Hilbert scheme of n points in C2 that turns out to be isomorphic to the Procesi bundle. Our results were used by I. Gordon to obtain a new proof of positivity of the Kostka-Macdonald polynomials established earlier by Haiman. 1 1 0 2 p e S 7 ] . G A h t a m [ 2 v 7 6 3 5 . 8 0 1 1 : v i X r a CONTENTS Introduction 1. 2. Analysis of the Harish-Chandra module 3. Springer resolutions 4. Proof of the main theorem 5. A generalization of a construction of Beilinson and Kazhdan 6. Geometry of the commuting scheme 7. Principal nilpotent pairs 8. Relation to work of M. Haiman 9. Some applications References 1 11 18 22 27 34 40 51 56 58 1. INTRODUCTION 1.1. Notation. We work over the ground field C of complex numbers and we write ⊗ = ⊗C.1 By a scheme X we mean a scheme of finite type over C. We write Xred for the corre- sponding reduced scheme and ψ : Xnorm → Xred for the normalization map (if Xred is irreducible). Let OX denote the structure sheaf of X, resp. KX the canonical sheaf (if X is Cohen-Macaulay), and DX the sheaf of algebraic differential operators on X (if X is smooth). Write C[X] = Γ(X,OX ), resp. D(X) = Γ(X, DX ), for the algebra of global sec- tions. Let T ∗X denote (the total space of) the cotangent bundle on a smooth variety X. 1A more complete Index of Notation is given at the end of the paper. 1 Given an algebraic group K and a K-action on E, we write EK for the set of K-fixed In particular, for a K-variety X, one has the subalgebra C[X]K ⊂ C[X], resp. points. D(X)K ⊂ D(X), of K-invariants. Throughout the paper, we fix a connected complex reductive group G with Lie algebra g. Let T ⊂ G be a maximal torus, t = Lie T the corresponding Cartan subalgebra of g, and r = dim t the rank of g. Write N (T ) for the normalizer of T in G so, W = N (T )/T is the Weyl group. The group W acts on t via the reflection representation and it acts on ∧rt by the sign character w 7→ sign(w). We write E sign for the sign-isotypic component of a W -module E. 1.2. Definition of the Harish-Chandra module. We will use a special notation D := Dg×t for the sheaf of differential operators on g× t. We have Γ(g× t, D) = D(g× t) = D(g)⊗ D(t) where D(g), resp. D(t), is the algebra of polynomial differential operators on the vector space g, resp. on t. The subalgebra of D(g), resp. of D(t), formed by the differential operators with constant coefficients may be identified with Sym g, resp. with Sym t, the corresponding symmetric algebra. Let the group G act on g by the adjoint action. Harish-Chandra [HC] defined a 'radial part' map rad : D(g)G → D(t)W . This is an algebra homomorphism such that its re- striction to the subalgebra of G-invariant polynomials, resp. of G-invariant constant co- efficient differential operators, reduces to the Chevalley isomorphism C[g]G ∼→ C[t]W , resp. (Sym g)G ∼→ (Sym t)W . Given a ∈ g, one may view the map ad a : g → g, x 7→ [a, x] as a (linear) vector field on g, that is, as a first order differential operator on g. The assignment a 7→ ad a gives a linear map ad : g → D(g) with image ad g. Thus, one can form a left ideal D (ad g ⊗ 1) ⊂ D. Definition 1.2.1. The Harish-Chandra module is a left D-module defined as follows M := D(cid:14)(cid:0)D (ad g ⊗ 1) + D {u ⊗ 1 − 1 ⊗ rad(u), u ∈ D(g)G}(cid:1). Remark 1.2.3. For a useful interpretation of this formula see also (2.4.1). (1.2.2) ♦ According to an important result of Hotta and Kashiwara [HK1], the Harish-Chandra module is a simple holonomic D-module of 'geometric origin', cf. Lemma 2.4.3(ii) below. This implies that M comes equipped with a natural structure of Hodge module in the sense of M. Saito [Sa]. In particular, there is a canonical Hodge filtration on M, see §2.5. Taking an associated graded sheaf with respect to the Hodge filtration produces a coherent sheaf scheme of the Lie algebra g, see Theorem 1.3.3 below. The main idea of the paper is to ex- ploit the powerful theory of Hodge modules to deduce new results concerning commuting egrHodge M on T ∗(g × t). The support of the sheaf egrHodge M turns out to be closely related to the commuting schemes using information about the sheaf egrHodge M. Remark 1.2.4. Definition 1.2.1 was motivated by, but is not identical to, the definition of Hotta and Kashiwara, see [HK1], formula (4.5.1). The equivalence of the two definitions follows from Remark 4.1.2, of §4.1 below. 1.3. Main results. Put G = g × g and let G act diagonally on G. The commuting scheme C is defined as the scheme-theoretic zero fiber of the commutator map κ : G → g, (x, y) 7→ [x, y]. Thus, C is a G-stable closed subscheme of G; set-theoretically, one has C = {(x, y) ∈ G [x, y] = 0}. The scheme C is known to be generically reduced and irreducible, cf. Proposition 2.1.1 below. It is a long standing open problem whether or not this scheme is reduced. 2 Let T := t × t ⊂ G. It is clear that T is an N (T )-stable closed subscheme of C and the resulting N (T )-action on T factors through the diagonal action of the Weyl group W = N (T )/T . Therefore, restriction of polynomial functions gives algebra maps The isospectral commuting variety is defined to be the algebraic set: res : C[G]G ։ C[C]G → C[T]W . X = {(x1, x2, t1, t2) ∈ C × T P (x1, x2) = (res P )(t1, t2), ∀P ∈ C[C]G}. We view X as a reduced closed subscheme of G × T, cf. also Definition 2.1.4. To proceed further, we fix an invariant bilinear form h−,−i : g × g → C. This gives an isomorphism g ∼→ g∗, x 7→ hx,−i, resp. t ∼→ t∗, t 7→ −ht,−i (the minus sign in the last formula is related to the minus sign that appears in the anti-involution v 7→ v⊤ considered coherent sheaf on G × T. in §2.4). Thus, one gets an identification T ∗(g × t) = G × T so one may view egrHodge M as a One of the main results of the paper, whose proof will be completed in §4.5, reads (1.3.1) (1.3.2) Theorem 1.3.3. There is a natural OG×T-module isomorphism ψ∗OXnorm This theorem combined with some deep results of Saito [Sa], to be reviewed in §2.3, yields the following theorem that confirms a conjecture of M. Haiman, [Ha3, Conjecture 7.2.3]. ∼→ egrHodge M. Theorem 1.3.4. Xnorm is a Cohen-Macaulay and Gorenstein variety with trivial canonical sheaf. Theorem 1.3.4 will be deduced from Therem 1.3.3 in §2.6. Corollary 1.3.5 (§6.4). The scheme Cnorm is Cohen-Macaulay. Corollary 1.3.6 (§9.2). The Dg-module Dg/Dg·ad g comes equipped with a canonical filtration F such that one has an isomorphism egrF (Dg/Dg·ad g) ∼= ψ∗OCnorm , of OG-modules. The last corollary implies (see §9.2) the following result that has been proved earlier by Levasseur and Stafford [LS2, Theorem 1.2] in a totally different way. Corollary 1.3.7. (ii) The natural right action of the algebra C[g]G makes D(g)/D(g) ad g a flat right C[g]G-module. (i) D(g)/D(g) ad g is a Cohen-Macaulay (non-holonomic) left D(g)-module. 1.4. Group actions. One has a natural G × W -action on g × t, resp. on G × T, where the group G acts on the first factor and the group W acts on the second factor. There is also a C×-action on g, resp. on g× t, by dilations. So, we obtain a C×× C×-action on G = T ∗g, resp. on G × T = (g × t) × (g × t) = T ∗(g × t), such that the standard C×-action by dilations along the fibers of the cotangent bundle corresponds, via the above identification, to the action of the subgroup {1}× C× ⊂ C×× C×. Thus, we have made the space g×t a G×W × C×-variety, resp. the space G × T a G × W × C× × C×-variety. The scheme X is clearly G× W × C× × C×-stable. The resulting G× W × C× × C×-action on X induces one on Xnorm since a reductive group action can always be lifted canonically to the normalization, [Kr], §4.4. On the other hand, the group G × W × C× acts on g × t and the Harish-Chandra module M has the natural structure of a G × W × C×-equivariant D-module. The Hodge filtration on M is canonical, therefore, this G × W × C×-action respects the filtration. Hence, the group G × W × C× acts naturally on egrHodge M. There is also an additional C×-action on egrHodge M that comes from the grading. Thus, combining all these actions together, one may view egrHodge M as a G × W × C× × C×-equivariant coherent sheaf on G × T. 3 The isomorphism of Theorem 1.3.3 respects the G× W × C× × C×-equivariant structures. The equivariant structure on the sheaf OXnorm makes the coordinate ring C[Xnorm] a bigraded locally finite G × W -module. In §§3.4,4.5, we will construct a "DG resolution" of Xnorm, a (derived) "double analogue" of the Grothendieck-Springer resolution, cf. Remark 4.5.4. We show that the DG algebra of global sections of our resolution is acyclic in nonzero degrees. Using this, a standard application of the Atiyah-Bott-Lefschetz fixed point theorem yields the following result. Theorem 1.4.1 (see [BG, §2.4]). The bigraded T -character of C[Xnorm] is given by the formula (1 − e−α)(1 − q eα)(1 − t eα) . χC××C××T (OXnorm ) = the formula is understood as a formal power series of the formPm,n≥0 am,n · qmtn where Here, R+ denotes the set of positive roots of g and the product in the right hand side of the coefficients am,n are viewed as elements of the representation ring of the torus T . w Yα∈R+ 1 (1 − q)r(1 − t)r · Xw∈W 1 − qt eα We refer to [BG] for the proof and some combinatorial applications of the above theorem. 1.5. A coherent sheaf on the commuting scheme. The first projection G × T → G restricts to a map p : X → C. Therefore, the composite Xnorm → X → C factors through a morphism pnorm : Xnorm → Cnorm. It is immediate to see that pnorm is a finite G × C× × C×-equivariant morphism and that the group W acts along the fibers of this morphism. Let R := (pnorm)∗OXnorm. This is a G × C× × C×-equivariant coherent sheaf of OCnorm- algebras. The action of the group W along the fibers of the map pnorm gives a W -action on R by OCnorm-algebra automorphisms. Therefore, for any finite dimensional W -representation E, one has a coherent sheaf RE := (E ⊗ R)W , the E-isotypic component of R. Equivalently, in terms of the contragredient representation E∗, we have RE = HomW (E∗, R). The sheaf R enjoys the following properties, see §6.6. Corollary 1.5.1. (i) The sheaf R is Cohen-Macaulay and we have (ii) There is a G× W × C× × C×-equivariant isomorphism R ∼= Hom OCnorm (R, KCnorm). Further- more, for any finite dimensional W -module E, this gives an isomorphism OCnorm ∼= RW , resp. KCnorm ∼= R sign. RE∗⊗ sign ∼= Hom OCnorm (RE, KCnorm ). Given x ∈ g, let gx denote the centralizer of x in g. Similarly, write gx,y = gx ∩ gy for the centralizer of a pair (x, y) ∈ C in g. We call an element x ∈ g, resp. a pair (x, y) ∈ C, regular if we have dim gx = r, resp. dim gx,y = r. Let gr, resp. Cr, be the set of regular elements of g, resp. of C. One shows that the set Cr is a Zariski open and dense subset of C which is equal to the smooth locus of the scheme C, cf. Proposition 2.1.1 below. There is a coherent sheaf ggg on Cnorm, the "universal stabilizer sheaf", such that the geometric fiber of ggg at each point is the Lie algebra of the isotropy group of that point under the G- action, cf. §5.4 for a more rigorous definition. Any pair (x, y) ∈ Cr may be viewed as a point of Cnorm. The sheaf gggCr is locally free; its fiber at any point (x, y) ∈ Cr equals, by definition, the vector space gx,y. Part (i) of the following theorem says that the sheaf R gives an algebraic vector bundle on Cr that has very interesting structures. Part (ii) of the theorem provides a description of the isotypic components R∧ t, s ≥ 0, of the reflection representation of W , in terms of the sheaf ggg. t⊗ R)W , coresponding to the wedge powers ∧s t = (∧s s 4 (i) The restriction of the sheaf R to Cr is a locally free sheaf. Each fiber Theorem 1.5.2 (§§6.4,6.5). of the corresponding algebraic vector bundle is a finite dimensional algebra that affords the regular representation of the group W . (ii) For any s ≥ 0, there is a natural G× C× × C×-equivariant isomorphism R∧stCr ∼= ∧sgggCr . 1.6. Small representations. Let L be a finite dimensional rational G-representation. Given a Lie subalgebra a ⊂ g, we put La := {v ∈ L av = 0, ∀a ∈ a}. In particular, we have that Lt is the zero weight space of L. In §5.1, we introduce a coherent sheaf Lggg on Cnorm such that the geometric fiber of Lggg at any sufficiently general point (x, y) ∈ Cnorm is the vector space Lgx,y . Following A. Broer [Br], we call L small if the set of weights of L is contained in the root lattice of g and, moreover, 2α is not a weight of L, for any root α. Part (i) of our next theorem provides a description of W -isotypic components of the sheaf R which correspond to the W -representation in the zero weight space of a small G-representation. Theorem 1.6.1. Let E be a W -representation, resp. L be a small G-representation. Then, we have , of G × C× × C×-equivariant OCnorm-sheaves. (i) There is a canonical isomorphism Lggg ∼= RLt (ii) Restriction to T ⊂ Cnorm induces bigraded C[T]W -module isomorphisms: Γ(Cnorm, RE)G ∼→ (E ⊗ C[T])W , resp. (L ⊗ C[Cnorm])G ∼→ (Lt ⊗ C[T])W . (iii) For any s ≥ 0, one has an isomorphism Γ(Cr, ∧sggg)G ∼= (∧st ⊗ C[T])W . Part (i) of the theorem will be proved in §6.5 and parts (ii)-(iii) will be proved in §6.7. Remark 1.6.2. (a) The adjoint representation L = g is small. In this case, one has Lggg = ggg. So, the above theorem yields a sheaf isomorphism ggg ∼= Rt and a bigraded C[T]W -module isomorphism (g ⊗ C[Cnorm])G ∼= (t ⊗ C[T])W . (b) Let G = P GL(V ) and let n = dim V . The natural GL(V )-action on (V ∗)⊗n ⊗ ∧nV descends to G and the resulting G-representation L is known to be small. Furthermore, the zero weight space of that representation is isomorphic, as a W -module, to the regular representation of the Symmetric group W = Sn. Hence, for L = (V ∗)⊗n ⊗ ∧nV , we have R ∼= RLt Using the tautological injective morphism u : Lggg ֒→ L ⊗ OCnorm and writing u∗ for the . Therefore, Theorem 1.6.1(i) yields an isomorphism R ∼= ((V ∗)⊗n ⊗ ∧nV )ggg. dual morphism one obtains the following selfdual diagram of sheaves on Cnorm: V ⊗n ⊗ ∧nV ∗ ⊗ KCnorm where the isomorphism in the middle is due to Corollary 1.5.1(ii). / Hom OCnorm (R,KCnorm ) ∼= R   u / u∗ / (V ∗)⊗n ⊗ ∧nV ⊗ OCnorm , The above diagram is closely related to formula (40) in [Ha1, Proposition 3.7.2]. 1.7. Principal nilpotent pairs. Given a regular point x = (x1, x2) ∈ C, let Rx be the fiber at x of (the algebraic vector bundle on Cr corresponding to) the locally free sheaf R. By definition, one has Rx = C[p−1 norm(x), the scheme theoretic fiber of the morphism pnorm over x, is a W -stable (not necessarily reduced) finite subscheme of Xnorm. Thus, Rx is a finite dimensional algebra equipped with a W -action. norm(x)] where p−1 The W -module Rx is isomorphic to the regular representation of W, by Theorem 1.5.2(i). In particular, one has dim RW x = 1. The line RW x is clearly spanned by the unit of the algebra Rx. Further, one has a canonical map Rx → R sign , r 7→ r sign, the W -equivariant projection to the isotypic component of the sign representation. This map gives, thanks to the isomorphism KCnorm ∼= R sign of Corollary 1.5.1(i), a nondegenerate trace on the algebra x = dim R sign x 5 / Rx. In other words, the assignment r1×r2 7→ (r1·r2) sign provides a nondegenerate symmetric bilinear pairing on Rx. The most interesting fibers of the sheaf R are, in a sense, the fibers over principal nilpotent pairs. Following [Gi], we call a regular pair e = (e1, e2) ∈ Cr a principal nilpotent pair if there exists a rational homorphism C× × C× → G, (τ1, τ2) 7→ g(τ1, τ2) such that one has (1.7.1) Given a principal nilpotent pair e, we introduce a 'twisted' C× × C×-action on G given, τi · ei = Ad g(τ1, τ2)(ei) i = 1, 2, ∀τ1, τ2 ∈ C×. for (τ1, τ2) ∈ C× × C×, by the formula (x, y) 7→ (τ1, τ2) •e (x, y) =(cid:0)τ1 · Ad g(τ1, τ2)−1(x), τ2 · Ad g(τ1, τ2)−1(y)(cid:1). Equations (1.7.1) force e1, e2 be nilpotent elements of g. The scheme p−1 The •e-action on G, combined with the usual C× × C×-action on T by dilations of the two factors t, gives a C× × C×-action on G × T. The subvarieties C and X are clearly •e- stable. Lifting the actions to normalizations, one gets a C× × C×-action (to be referred to as a •e-action again) on Cnorm, resp. on Xnorm. The map pnorm : Xnorm → Cnorm is •e-equivariant. norm(e) is nonre- duced and it has a single closed point, the element (e, 0) ∈ G × T. The point e ∈ Cr is, by construction, a fixed point of the •e-action on Cnorm. Hence, p−1 norm(e) is a •e-stable sub- scheme of Xnorm. We conclude that the coordinate ring Re = C[p−1 norm(e)] is a local algebra norm(e) gives a Z2-grading Re =Lm,n∈Z and the C× × C×-action on p−1 on that algebra. Associated with the nilpotent pair e = (e1, e2) there is a pair of commuting semisimple elements of g defined by hs := ∂g(τ1,τ2) that the indices are "flipped"). Thus, gs, s = 1, 2, is a Levi subalgebra of g such that es ∈ gs. The Lie algebra gh1,h2 = g1 ∩ g2 is known to be a Cartan subalegebra of g. So, we may (and will) put t := gh1,h2. Let R+ s be the set of positive roots, resp. Ws the Weyl group, of the reductive Lie algebra gs. (cid:12)(cid:12)τ1=τ2=1, s = 1, 2. Let g1 = gh2, resp. g2 = gh1 (note Rm,n ∂τs The pair h = (h1, h2) is regular, [Gi, Theorem 1.2]; furthermore, the fiber p−1 norm(h) is a reduced finite subscheme of G × T. Specifically, writing W ·h for the W -orbit of the ele- ment h ∈ T, one has a bijection W ·h ∼→ p−1 norm(h), w(h) 7→ (h, w(h)). Thus, the algebra Rh = C[p−1 norm(h)] is a semisimple algebra isomorphic to C[W ·h], the coordinate ring of the set W ·h. Let C≤m[t], m = 0, 1, 2, . . . , be the space of polynomials on t of degree ≤ m. We intro- duce a pair of ascending filtrations on the algebra C[T] = C[t] ⊗ C[t] defined by ′FmC[T] = C≤m[t] ⊗ C[t], resp. ′′FmC[T] = C[t] ⊗ C≤m[t]. The algebra C[W ·h] is a quotient of the alge- bra C[T] hence it inherits from C[T] a pair of quotient filtrations ′F qC[W ·h] and ′′F qC[W ·h], respectively. We further define bifiltrations e Fm,nC[T] := ′FmC[T] ∩ ′′FnC[T], Fm,nC[W ·h] := ′FmC[W ·h] ∩ ′′FnC[W ·h], m, n ≥ 0, on the algebras C[T] and C[W ·h], respectively. Let grF C[T], resp. grF C[W ·h], be an associ- ated bigraded algebra, see §7 for more details. One of the central results of the paper is the following theorem motivated, in part, by [Ha3, §4.1]. Part (i) of the theorem describes how C[W ·h] = Rh, a semisimple Gorenstein algebra, degenerates to the bigraded Gorenstein algebra Re. Theorem 1.7.2. (i) There is a W -equivariant Z2-graded algebra isomorphism Re ∼= grF C[W ·h]. e = 0 unless 0 ≤ i ≤ d1 & 0 ≤ j ≤ d2, where ds := #R+ (ii) We have Ri,j The proof of this theorem occupies §§7.2-7.4. The main idea of the proof is to use the semisimple pair h to produce a 2-parameter deformation of the point e inside Cr. One s , s = 1, 2. 6 may pull-back the sheaf RCr to the parameter space of the deformation. This way, we ob- tain a locally free sheaf on C2. On the other hand, a construction based on Rees algebras gives another locally free sheaf on C2 such that its fiber over the origin is the vector space grF C[W ·h]. Using a careful analysis (Lemma 7.2.1) based on the theory of W -invariant polynomials on t we obtain an isomorphism between the restrictions of the two sheaves in question to the complement of the origin in C2 (Proposition 7.3.7). We then exploit the fact that an isomorphism between the restrictions to the punctured plane of two locally free sheaves automatically extends across the puncture. From Theorem 1.7.2 we deduce, cf. §7.6, Corollary 1.7.3. There is a natural W -module isomorphism (cf. [Ha3, Proposition 4.1.2]): Lm≥0 e ∼= C[W/W1], R0,m e ∼= C[W/W2]. Rn,0 resp. Ln≥0 One has the following criteria for the variety X be normal (hence, by Theorem 1.3.4, also Cohen-Macaulay and Gorenstein) at the point (e, 0), see §7.6: Corollary 1.7.4. The following properties of a principal nilpotent pair e are equivalent: (i) The restriction map C[T] = Γ(Cnorm, R)G → Re is surjective; (ii) The natural projection grF C[T] → grF C[W ·h] is surjective; (iii) The variety X is normal at the point (e, 0). In the special case of the group G = GLn, Theorem 1.7.2 follows from the work of M. Haiman, [Ha3, §4.1]. Moreover, in this case, thanks to Haiman's result on the normality of the isospectral Hilbert scheme ([Ha1], Proposition 3.8.4) and to the classification of principal nilpotent pairs in the Lie algebra gln (see [Gi, Theorem 5.6]), one knows that the equivalent properties of Corollary 1.7.4 hold true for any principal nilpotent pair. We remark also that Haiman shows that the validity of the Cohen-Macaulay property of the scheme X at each principal nilpotent pair in the Lie algebra gln is actually equivalent to the validity of the n!-theorem, see [Ha1, Proposition 3.7.3]. On the other hand, Haiman produced an example, see [Ha3, §7.2.1], of a principal nilpo- tent pair e for the Lie algebra g = sp6 where the analogue of the n!-theorem fails, hence the corresponding homomorphism gr C[T] → gr C[W ·h] is not surjective and the scheme X is not normal at the point (e, 0). Another example is provided by the exceptional principal nilpotent pair in the simple Lie algebra of type E7 discussed below. 1.8. The polynomial ∆e. From the isomorphism Re = grF C[W ·h], of Theorem 1.7.2, we see that grF C[W ·h] is a Gorenstein algebra and that the line (grF C[W ·h]) sign is the socle of that algebra. In most cases, one can actually obtain a more explicit description of the socle. To explain the meaning of the words "most cases", we recall that the simple Lie algebra of type E7 has one 'exceptional' conjugacy class of principal nilpotent pairs e = (e1, e2) such 101010(cid:3) . Following [Gi, that each of the nilpotent elements e1 and e2 has Dynkin labels (cid:2) Definition 4.1], we call a principal nilpotent pair e = (e1, e2), of an arbitrary reductive Lie algebra g, non-exceptional provided none of the components of e corresponding to the simple factors of g of type E7 belong to the exceptional conjugacy class of principal nilpotent pairs. Let hs = hhs,−i, s = 1, 2, be a linear function on t that corresponds to the element hs via The following result provides a simple description of the socle of the algebra grF C[W ·h] the invariant form. in the case of nonexceptional nilpotent pairs. Theorem 1.8.1 (see §7.5). For any non-exceptional principal nilpotent pair e, we have 0 (grF C[W ·h]) sign = grd1,d2 C[W ·h]. 7 Morover, a base vector of the 1-dimensional vector space on the right is provided by the image under the map grF C[T] → grF C[W ·h] of the class of the following bihomogeneous polynomial: sign(w) · w(hd1 1 ⊗ hd2 2 ) ∈ C[T] sign. (1.8.2) ∆e :=Xw∈W Remark 1.8.3. The polynomial ∆e was first introduced in [Gi]. It is a W -harmonic polynomial on T that provides a natural generalization to the case of arbitrary reductive Lie algebras of the Garsia-Haiman polynomial on Cn × Cn. The proof of Theorem 1.8.1 is based on the properties of ∆e established in [Gi, §4]. The relevance of the notion of non-exceptional pair is due to [Gi, Theorem 4.4], one of the main results of loc cit, which says that ∆e 6= 0 holds if and only if the principal nilpotent pair e is non-exceptional. ♦ One may pull-back the function ∆e via the projection Xnorm → T. The resulting W - alternating function on Xnorm gives a G-invariant section, se, of the sheaf R sign. Let se(e) ∈ R sign denote the value of the section se at the point e. e Corollary 1.8.4 (see §7.6). For a non-exceptional principal nilpotent pair e, one has: ∆e(h) 6= 0 and se(e) 6= 0. Moreover, we have that R sign = C · se(e) is the socle of the algebra Re. e = Rd1,d2 e hhhx, yiii := hx1, y1i + hx2, y2i, for any x = (x1, x2), y = (y1, y2) ∈ T. Lethhh−,−iii : T×T → C be the natural W ×W -invariant bilinear form given by the formula We define a holomorphic function E on T × T as follows sign(w) · ehhhx, w(y)iii, x, y ∈ T. E(x, y) :=Xw∈W It is clear that, for any x, y, we have E(x, y) = E(y, x). Observe also that the function E is W -invariant with respect to the diagonal action on T × T and, for any fixed x ∈ G, the function E(x,−) = E(−, x) is a W -alternating holomorphic function on T. Similarly to the above, one can pull-back the function E via the projection Xnorm×Xnorm → T × T. The resulting function on Xnorm × Xnorm gives a G × G-invariant holomorphic section, S, of the coherent sheaf R sign ⊠ R sign on C × C. The following result shows that the section S provides a canonical holomorphic interpola- tion between the algebraic sections se associated with various, not necessarily G-conjugate, principal nilpotent pairs e of the Lie algebra g. Proposition 1.8.5 (§7.6). For any non-exceptional principal nilpotent pair e, in R sign ⊠ R sign in R sign e ⊠ R sign, one has e , resp. SC×{e} = c · se ⊠ se(e), resp. S{e}×C = c · se(e) ⊠ se where c = 1 d1!·d2!·∆e(h) . n n (C2) is a normal, Cohen-Macaulay and Gorenstein scheme. 1.9. Relation to work of M. Haiman. Let Hilbn(C2) be the Hilbert scheme of n points in C2. In his work on the n!-theorem, Haiman introduced a certain isospectral Hilbert scheme (C2), a reduced finite scheme over Hilbn(C2), see [Ha1]. The main result of loc cit says gHilb that gHilb Now, let G = GLn. It turns out that there is a Zariski open dense subset C◦ ⊂ Cr such that the projection pnorm : p−1 (C2) → Hilbn(C2), see §8. Using this relation, we are able to deduce from our Theorem 1.3.4 that the normalization of the isospectral Hilbert scheme is Cohen-Macaulay and Gorenstein, see Proposition 8.2.4. Unfortunately, our approach does not seem to yield an independent proof of normality of the isospectral Hilbert scheme, while the proof of normality given in [Ha1, norm(C◦) → C◦ is closely related to the projection gHilb n 8 Proposition 3.8.4] is based on the 'polygraph theorem' [Ha1, Theorem 4.1], a key technical result of [Ha1]. Nonetheless, we are able to use the locally free sheaf RCr to construct a rank n! algebraic vector bundle P on Hilbn(C2) whose fibers afford the regular representation of the Symmet- ric group. The results of Haiman [Ha1] then insure that the vector bundle P is isomorphic, a posteriori, to the Procesi bundle, cf. Corollary 8.2.5. We note that the properties of the vector bundle P that we construct are often sufficient (without the knowledge of the isomorphism with the Procesi bundle) for applications. This is so, for instance, in the new proof of the positivity of Kostka-Macdonald polynomials found recently by I. Gordon [Go]. Another situation of this kind is provided by Theorem 1.9.1 below. See also various combinatorial results established in [BG]. Let V = Cn be the fundamental representation of G = GLn and let V be the tautological rank n vector bundle on Hilbn(C2), see §8.3. Given an integer m ≥ 0, let Cm[V ] be the space of degree m homogeneous polynomials on V , and let Sm be the symmetric group on m letters. The group Sm acts naturally on V ⊗m, resp. on V⊗m, hence also on C[Xnorm] ⊗ Cm[V ] ⊗ V ⊗m, resp. on P ⊗ V⊗m, via the action on the last tensor factor. We also have the group W = Sn and a natural W × C× × C× action on Xnorm, resp. on P, hence on C[Xnorm] ⊗ Cm[V ] ⊗ V ⊗m, resp. on P ⊗ V⊗m, via the action on the first factor. Finally, we let the subgroup SLn ⊂ GLn act diagonally on C[Xnorm] ⊗ Cm[V ] ⊗ V ⊗m. The SLn-action commutes with the previously defined actions of other groups. Thus, the space(cid:0)C[Xnorm]⊗ Cm[V ]⊗ V ⊗m(cid:1)SLn, of SLn-invariants, acquires the natural structure of an Sm × W -equivariant bigraded C[T]-module. Theorem 1.9.1 (see §8.3). There is a W × Sm-equivariant bigraded C[T]-module isomorphism (cid:0)C[Xnorm] ⊗ Cm[V ] ⊗ V ⊗m(cid:1)SLn ∼→ Γ(Hilbn(C2), P ⊗ V⊗m), We remark that the space of global sections that appears on the right hand side of the isomorphism above is the "polygraph space" that played a key role in the work of Haiman. In [Ha4, Theorem 3.1], Haiman gives a bigraded character formula for the polygraph space in terms of Macdonald polynomials. On the other hand, our Theorem 1.4.1 may be used to find the bigraded character of the vector space(cid:0)C[Xnorm] ⊗ Cm[V ] ⊗ V ⊗m(cid:1)SLn. Combin- ing these results with Theorem 1.9.1, one obtains a certain explicit identity that relates the formal power series in the right hand side of the formula of Theorem 1.4.1 to Macdonald polynomials, see [BG] for more details. A particularly nice special case occurs in the situation where the integer m, in Theorem 1.9.1, is a multiple of n, i.e., such that m = kn for some integer k ≥ 1. Let Ak ⊂ C[T] be the vector space spanned by the products of k-tuples of elements of C[T]sign, the space of W -alternating polynomials. The bigrading on C[T] induces one on Ak and we have Proposition 1.9.2. For any k ≥ 1, there is a natural bigraded isomorphism (see §8.5): ∀m ≥ 0. (cid:0)C[Cred] ⊗ Ckn[V ](cid:1)SLn ∼→ Ak. This result may be viewed as an example of a situation involving the isomorphism of Theorem 1.9.1 where taking normalizations is unnecessary. 1.10. Layout of the paper. In section 2 we begin with basic properties of commuting and isospectral varieties. We then review the material involving holonomic D-modules and (po- larizable) Hodge modules that will be used in the paper. We show that the Harish-Chandra 9 module M is a simple holonomic D-module and give M the structure of a Hodge mod- ule. We explain that Saito's results on Hodge modules imply easily that the coherent sheaf grHodge M is Cohen-Macaulay and Gorenstein. This allows us to reduce the proof of The- orem 1.3.3 to an isomorphism grHodge MXrr ∼= OXrr, where Xrr is a certain open subset of X, and to a codimension estimate involving Xrr. We finish the section by showing how Theorem 1.3.3 implies Theorem 1.3.4. In section 3, we recall the defininition of the Grothendieck-Springer resolution and con- struct its 'double analogue'. We introduce a DG algebra A such that its homogeneous com- ponents are locally free sheaves on the double analogue of the Grothendieck-Springer res- olution. The spectrum of the DG algebra A may be thought of as a DG resolution of the isospectral commuting variety. In section 4, we recall the Hotta-Kashiwara construction of the Harish-Chandra module M as a direct image of the structure sheaf of the Grothendieck-Springer resolution. We use the Hotta-Kashiwara construction to obtain a description of the sheaf grHodge M in terms of the DG algebra A. From that description, we deduce the above mentioned isomorphism grHodge MXrr ∼= OXrr. In section 5, we give the definition of the 'universal stabilizer' sheaf. Further, we associate with any finite dimensionsal representation L of the Lie algebra g a coherent sheaf Lggg, on gr. In the case where the representation L is small, we obtain, using a refinement of an idea due to Beilinson and Kazhdan, a description of the sheaf Lggg in terms the 'universal zero weight space' of L. At the end of the section we indicate how to modify the construction in order to obtain a similar description of Lggg for a not necessarily small representation L. In section 6, we introduce a stratification of the commutating variety and use it to prove the above mentioned dimension estimate involving the set Xrr which is required for our proof of Theorem 1.3.3. We then prove Theorems 1.5.2 and 1.6.1 by adapting the construction of §5 to a 'double' setting. We also prove Corollary 1.5.1. Section 7 is devoted to the proofs of all the results stated in §§1.7-1.8 of the Introduction. In section 8, we consider the case of the group GLn. We apply our results about commut- ing varieties to the geometry of the isospectral Hilbert scheme. We discuss relations with the work of M. Haiman and prove the results stated in §1.9. Finally, in section 9, we prove Corollary 1.3.6 and Corollary 1.3.7. We remark that the proof of the main results, Theorem 1.3.3 and Theorem 1.3.4, is rather long. The argument involves several ingredients of different nature. The discussion of these igredients is spread over a number of sections of the paper. Our division into sections has been made according to the nature of the material rather than to the order in which that material is actually used in the proof. To make the logical structure of our arguments more clear, below is an outline of the main steps of the proof of Theorem 1.3.3 and Theorem 1.3.4, listed in the logical order: codimension ≥ 2 in X (Corollary 2.6.5). that it is an isomorphism over Xrr (Corollary 3.3.5). §§2.6, 6.3: Define a Zariski open smooth subset Xrr ⊂ X and show that the set X r Xrr has §§3.1, 3.3: Introduce the 'double analogue' of the Grothendieck-Springer resolution and show §§2.4, 2.5: Show that M is a simple D-module with the natural structure of a Hodge module. §§4.2, 4.4: Use the Hotta-Kashiwara construction of M to get a description (see Corollary 4.4.6) of egrHodge M in terms of the double analogue of the Grothendieck-Springer resolution. §4.5: Establish an isomorphism: egrHodge MXrr ∼= OXrr (see Proposition 2.6.6 and §4.5). §2.5: Deduce from Saito's theory that the sheaf egrHodge M is Cohen-Macaulay and selfdual (Corollary 2.5.1). 10 §2.6: Deduce Theorem 1.3.3 from Lemma 2.6.1, using Proposition 2.6.6 and Corollary 2.6.5 (which were proved at an earlier stage). Deduce Theorem 1.3.4 from Theorem 1.3.3 and Corollary 2.5.1. 1.11. Acknowledgements. I am grateful to Dima Arinkin and Gwyn Bellamy for useful discussions. Special thanks are due to Iain Gordon for showing me a draft of [Go] well before that paper was finished, for many stimulating discussions, and for a carefull reading of (many) preliminary versions of the present paper. I thank Eric Vasserot for pointing out a connection of Theorem 1.4.1 with a result in [SV], cf. also §3.4. Vladimir Drinfeld and Mark Haiman have helped me with some questions of algebraic geometry. This work was supported in part by an NSF award DMS-1001677. 2. ANALYSIS OF THE HARISH-CHANDRA MODULE 2.1. Isospectral varieties. Let grs ⊂ gr be the set of regular semisimple elements, resp. Crs ⊂ Cr be the set of pairs (x, y) ∈ Cr such that both x and y are semisimple. Basic properties of the commuting scheme C may be summarized as follows. Proposition 2.1.1. and irreducible scheme; moreover, we have dim C = dim g + r. (i) The set Crs is Zariski open and dense in C. Thus, C is a generically reduced (ii) The smooth locus of the scheme C equals Cr; for any (x, y) ∈ C one has dim gx,y ≥ r. Here, part (i) is due to Richardson [Ri1]. For the proof of part (ii) see e.g. [Po, Lemma 2.3], or Remark 6.6.2 of section 6 below. Let g//G := Spec C[g]G, resp. C//G := Spec C[C]G, be the categorical quotient. The nat- (1.3.1), gives ural restriction homomorphism C[g]G → C[t]W , resp. C[C]G → C[T]W , cf. morphisms of schemes t → t/W → g//G, resp. T → T/W → C//G. Remark 2.1.2. The morphism t/W → g//G is an isomorphism by the Chevalley restriction theorem. One can show, using that the map C[C]G → C[T]W is surjective by a theorem of Joseph [Jo], that the morphism T/W → C//G induces an isomorphism T/W ∼→ [C//G]red. It is expected that the scheme C//G is in fact reduced. This is known to be so in the special case of the group G = GLn, see [GG, Theorem 1.3]. ♦ Next, we form a fiber product x := g ×g//G t, resp. C ×C//G T, a closed G × W -stable subscheme of g× t, resp. of G× T. It is clear that the set xrs := grs ×g//G t is a smooth Zariski open subset of x, resp. Xrs := Crs ×C//G T is a smooth Zariski open subset of C ×C//G T. The first projection x → g, resp. C ×C//G T → C, is a G-equivariant finite morphism. The group W acts along the fibers of this morphism. Part (i) of the following lemma is Corollary 6.2.3 of §6 below, and part (ii) is clear. Lemma 2.1.3. (i) The set xrs, resp. Xrs, is an irreducible dense subset of x, resp. of C ×C//G T. (ii) The first projection xrs → grs, resp. Xrs ։ Crs, is a Galois covering with Galois group W . The scheme x is known to be a reduced normal complete intersection in g × t, cf. eg [BB]. On the contrary, the scheme C ×C//G T is not reduced already in the case g = sl2. This motivated M. Haiman, [Ha2, §8], [Ha3, §7.2], to introduce the following definition, which is equivalent to formula (1.3.2) of the Introduction. Definition 2.1.4. The isospectral commuting variety is defined as X := [C ×C//G T]red, a reduced fiber product. Let p : X → C, resp. pT : X → T, denote the first, resp. second, projection. 11 Lemma 2.1.3(i) shows that X is an irreducible variety and that we may (and will) identify the set Xrs with a smooth Zariski open and dense subset of X. Let r := {(x, t) ∈ x x ∈ gr}. This is a Zariski open and dense subset of x which is contained in the smooth locus of x (since the differential of the adjoint quotient map g → g//G is known to have maximal rank at any point of gr). Let Nr be the total space of the conormal bundle on r in g × t and let Nr be the closure of Nr in T ∗(g × t). In general, let X be a smooth variety. An irreducible reduced subvariety Λ ⊂ T ∗X is said to be Lagrangian if the tangent space to Λ at any smooth point of Λ is a Lagrangian subspace of the tangent space to T ∗X with respect to the canonical symplectic 2-form on T ∗X. Lemma 2.1.5. In T ∗(g × t) = G × T, we have Nr = X. In particular, X is a C× × C×-stable Lagrangian subvariety and Nr is a smooth Zariski open and dense subset of X. Proof. Let t′ := t ∩ gr and X′ := {(x, y, t1, t2) ∈ X x ∈ grs}. We claim that X′ ⊂ Nr. To see this we observe that the assignment gT × (t1, t2) 7→(cid:0) Ad g(t1), Ad g(t2), t1, t2(cid:1) yields an isomorphism (G/T ) × t′ × t ∼→ X′, see Lemma 6.2.2. Thus, by G-equivariance, it suffices to show that every point of X′ of the form (t1, t2, t1, t2) is contained in Nr. Any tangent vector to r at the point (t1, t1) ∈ g× t has the form ξ = ([a, t1] + t, t) ∈ g× t, for some a ∈ g and some t ∈ t. Further, the covector that corresponds to the element (t1, t2, t1, t2) via our identification G × T ∼→ T ∗g × T ∗t = T ∗(g × t) is the linear function ξ∗ : g × t : (u, v) 7→ ht2, ui − ht2, vi. Therefore, using the invariance of the bilinear form h−,−i, we find hξ∗, ξi = ht2, [a, t1] + ti − ht2, ti = ht2, [a, t1]i = ha, [t1, t2]i = 0. Thus, we have proved that X′ ⊂ Nr. Moreover, it is clear that X′ is an open subset of X ∩ Nr and, we have dim X′ = dim(G/T × t′ × t) = dim(g × t) = 1 2 dim T ∗(g × t) = dim Nr. Now, Lemma 2.1.3(i) implies that X is reduced and irreducible. Hence, X′ is a dense subset both in X and in Nr. It follows that X = X′ = Nr. (cid:3) independent of the choice of a good filtration on M , cf. [Bo]. Let M be a DX -module. An ascending filtration F qM such that F ord i 2.2. Let X be a smooth variety and q : T ∗X → X the cotangent bundle. The sheaf DX comes equipped with an ascending filtration F ord DX by the order of differential operator. For the associated graded sheaf, one has a canonical isomorphism grord DX ∼= q∗OT ∗X. DX ·Fj M ⊂ Fi+jM, ∀i, j, is said to be good if grF M , an associated graded module, is a coherent q∗OT ∗X -module. In that case, there is a canonically defined coherent sheaf egrF M on T ∗X such that one has an isomorphism grF M = q∗ egrF M of q∗OT ∗X -modules. We write [Supp(egrF M )] for the support cycle of the sheaf egrF M , a linear combination of the irreducible components of the support of egrF M counted with multiplicites. This is an algebraic cycle in T ∗X which is known to be Recall that M is called holonomic if one has dim Supp(egrF M ) = dim X. In such a case, each irreducible component of Supp(egrF M ), viewed as a reduced variety, is a Lagrangian M 7→ [Supp(egrF M )] is additive on short exact sequences of holonomic modules, cf. eg. [Bo], Lemma 2.2.1. If M is a DX -module such that the cycle [Supp(egrF M )] equals the fundamental cycle Given a morphism f : X → Y of smooth varieties, let DX→Y := OXNf −1OY Assuming f is proper, there is a direct image functorR R subvariety of T ∗X. Holonomic DX -modules form an abelian category and the assignment of a Lagrangian subvariety, taken with multiplicity 1, then M is a simple holonomic DX -module. f = Rf∗(− DX→Y ) between [HTT]. From this, one obtains LNDX q 12 f −1DY . see [HTT, pp.22-23, 69] for more details. bounded derived categories of coherent right D-modules on X and Y , respectively. The cor- M ), responding functor on left D-modules is then defined byRf M = K−1 Y NOY R R In the special case where f : X ֒→ Y is a closed immersion, one has OXNf −1OY DY X so, we getR R f KX(cid:1) = f∗(cid:0)KX ⊗OX (F ord ural order filtration defined by the formula F ord 0. Thus, writing q : T ∗Y → Y for the cotangent bundle, in this case one obtains f −1DY = f KX = f∗(KX⊗OX (DY X)). This DY -module comes equipped with a nat- DY )X(cid:1), m ≥ f (KXNOX m (2.2.2) m (cid:0)R R f KY(cid:17) = q∗(f∗KX ). egrord(cid:16)R R 2.3. Filtered D-modules and duality. Below, we will use rudiments of the formalism of filtered derived categories. Let E : . . . dk−2 / Ek−1 dk−1 / Ek dk / Ek+1 dk+1 / . . . , be a filtered complex in an abelian category. One has the following useful Definition 2.3.1. The filtered complex E is said to be strict if the morphism dk : FjEk → Im dk ∩ FjEk+1 is surjective, for any k, j ∈ Z. This notion only depends on the quasi-isomorphism class of E in an appropriate filtered derived category. Given a filtered complex E, there is an induced filtration on each cohomology group H k(E), k ∈ Z. It is immediate from the construction of the standard spectral sequence of a filtered complex that the filtered complex E is strict if and only if the spectral sequence H q(gr E) ⇒ gr H q(E) degenerates at the first page, cf. [La, Lemma 3.3.5]. In that case, one has a canonical isomorphism gr H q(E) ∼= H q(gr E). Following G. Laumon and M. Saito, for any smooth algebraic variety X, one has an ex- act (not abelian) category of filtered left DX-modules and also the corresponding derived category. Thus, let FDX-mod be an additive category whose objects are DX -modules M equipped with a good filtration F . Further, abusing the notation slightly, we let Db coh(FDX ) be the triangulated category whose objects are isomorphic to bounded complexes (M, F ), of filtered DX -modules, such that each cohomology group Hi(M, F ) is an object of FDX -mod, cf. [La], [Sa]. In his work [Sa], M. Saito defines a semisimple abelian category HM(X) of polarizable Hodge modules, see [Sa, §5.2.10]. The data of a polarizable Hodge module includes, in particular, a holonomic DX -module M with regular singularities and a good filtration F on M called Hodge filtration.2 Thus, (M, F ) is a filtered DX -module; abusing notation, we write (M, F ) ∈ HM(X) and let egr(M, F ) denote the corresponding coherent sheaf on T ∗X. Let D(−) = R Hom DX (−, DX ⊗OX K−1 X )[dim X] be the standard Verdier duality func- tor on D-modules, cf. [HTT, §2.6]. Laumon and Saito have upgraded Verdier duality to a triangulated contravariant duality functor FD on the category Db [La, §4], [Sa, §2]. Furthermore, Saito showed that, for any (M, F ) ∈ HM(X), the filtered complex FD(M, F ) ∈ Db coh(Z) denote the corresponding bounded derived category. Let Coh Z denote the abelian category of coherent sheaves on a scheme Z and let Db coh(FDX ) is strict, see [Sa], Lemma 5.1.13. coh(FDX ), cf. 2Part of the data involving "polarization" will play no role in the present work. For this reason, from now on, we will refer to polarizable Hodge modules as 'Hodge modules', for short. 13 / / / / [La], (4.0.6). coh(T ∗X), cf. coh(FDX ) → Db It is essentially built into the construction of duality on filtered derived categories that the func- coh(FDX ) with the Grothendieck-Serre duality on The assignment (M, F ) 7→ egrF M gives a functor FDX-mod → Coh T ∗X that can be ex- tended to a triangulated functor gdgr : Db torgdgr intertwines the duality FD on Db coh(T ∗X), [La, §4], [Sa, §2]. Db Now let (M, F ) ∈ HM(X). Then M is a holonomic D-module. Hence D(M ), viewed as an object of the derived category, is quasi-isomorphic to its 0th cohomology. Therefore, the strictness of the filtered complex FD(M, F ) implies that the complexgdgr(FD(M, F )) has Grothendieck-Serre dual of the object egr(M, F ) ∈ Db to egr(D(M )). In particular, one concludes that egr(M, F ) is a Cohen-Macaulay sheaf on T ∗X, nonvanishing cohomology in at most one degree. It follows, in view of the above, that the coh(T ∗X) has nonvanishing cohomology in at most one degree again. Moreover, the cohomology sheaf in that degree is isomorphic for any (M, F ) ∈ HM(X), see [Sa], Lemma 5.1.13. 2.4. The order filtration on the Harish-Chandra module. Observe that D(g) · ad g, a left ideal of the algebra D(g), is stable under the G-action on D(g) induced by the adjoint action on g. Multiplication in the algebra D(g) gives the quotient A := D(g)G/[D(g) · ad g(cid:3)G a nat- ural algebra structure called quantum Hamiltonian reduction, cf. eg. [GG, §7.1]. Furthermore, it gives D(g)/D(g) · ad g the natural structure of an (D(g), A)-bimodule, [LS3], [GG, §7.1]. It is immediate from definitions that the radial part map rad considered in §1.2 descends to a well defined map A → D(t)W . According to an important result, due to Levasseur and Stafford [LS1]-[LS2] and Wallach [Wa], the latter map is an algebra isomorphism. Thus, one may view D(g)/D(g)· ad g as an (D(g), D(t)W )-bimodule. It is easy to see that, taking global sections on each side of formula (1.2.2), yields an isomorphism, cf. [LS3, p. 1109], Γ(g × t,M) Ξ [D(g)/D(g)·ad g]ND(t)W D(t). (2.4.1) ∂t 7→ − ∂ ∂t . Here, the object on the left has the structure of a left D(g) ⊗ D(t)-module and the object on the right has the structure of a (D(g), D(t))-bimodule. These structures are related via the isomorphism Ξ as follows Ξ[(u ⊗ v)m] = u Ξ(m) v⊤, for any m ∈ Γ(g × t,M) and any u ⊗ v ∈ D(g) ⊗ D(t) where v 7→ v⊤ is an anti-involution of the algebra D(t) given by t 7→ t, ∂ According to formula (1.2.2) the Harish-Chandra module has the form M = D/I where I is a left ideal of D. The order filtration on D restricts to a filtration on I and it also q M on D/I. Using the identifications T ∗(g × t) = G × T induces a quotient filtration F ord ideal, is a subsheaf of ideals of OG×T, not necessarily reduced, in general. A relation between is provided by part (i) of the lemma below. and egrord D = OG×T, we get egrord M = OG×T/egrord I where egrord I, the associated graded egrord I and J ⊂ OG×T, the ideal sheaf of the (non reduced) subscheme C ×C//G T ⊂ G × T, Notation 2.4.2. We put δt := Qα∈R+ α and tr := t ∩ gr. Let dx ∈ Kg, resp. dt ∈ Kt, be a Write j : r ֒→ g×t for the locally closed imbedding and let j!∗Or be the minimal extension, constant volume form on g, resp. on t. Thus, dx dt is a section of Kg×t. see [Bo], of the structure sheaf Or viewed as a Dr-module. Lemma 2.4.3. (i) One has inclusions J ⊂ egrord I ⊂ √J . (ii) The Harish-Chandra module M is a simple holonomic D-module, specifically, one has an isomorphism M ∼= j!∗Or and an equality [Supp(egrord M)] = [X] of algebraic cycles in G × T. 14 Proof. To simplify notation, it will be convenient below to work with spaces of global sec- tions rather than with sheaves. Thus, we put I = Γ(g × t, I) ⊂ D(g × t), resp. J = Γ(G × T,J ) ⊂ C[G × T]. The vector space g × t being affine, it is sufficient to prove an ana- logue of the lemma for the ideals gr I and J where throughout the proof we put gr := grord. Equip the space D(g)/D(g) · ad g with the quotient filtration. Then, taking associated graded spaces on each side of isomorphism (2.4.1) yields the following chain of graded maps, where pr stands for the natural projection gr D(t) / gr"(cid:18) D(g) gr[D(g)·ad g](cid:19) Ngr D(t)W (cid:18) gr D(g) We have gr D(g) = C[G]. Observe that the map gr(ad) : g → gr D(g) may be identified with the map κ∗ : g = g∗ → C[G], the pull-back morphism induced by the commutator map κ. By definition, one has C[G]/C[G]· κ∗(g) = C[C]. Therefore, using the inclusion gr D(g)·gr(ad)(g) ⊂ gr[D(g)·ad g] we obtain a chain of graded maps D(g)·ad g(cid:19) ND(t)W D(t)# gr Ξ gr Γ(g × t,M). pr C[C] = C[G]/C[G]·κ∗(g) = gr D(g)/ gr D(g)·gr(ad)(g) ։ gr D(g)/ gr[D(g)·ad g]. (2.4.4) We let a denote the composite map in (2.4.4). Now, the radial part map rad : D(g)G → D(t)W is known to respect the order filtrations, [Wa, §3]. Moreover, gr(rad) : gr D(g)G → gr D(t)W , the associated graded map, is nothing but the algebra map res : C[G]G → C[T]W in (1.3.1). Thus, combining the maps in the two displayed formulas above and writing b : C[T] ∼→ gr D(t) for the standard isomorphism, we obtain the following graded surjective maps a⊗b gr Ξ ◦ pr gr D(t) gr Ξ ◦ pr ◦ (a⊗b) gr[D(g)·ad g](cid:17) Ngr D(t)W / (cid:16) gr D(g) / gr Γ(g × t,M). / gr Γ(g × t,M) = gr D(g × t)/ gr I. composite map in the last formula gives a graded surjective morphism C[C]NC[T]W C[T] Further, by definition, we have C[G × T]/J = C[C ×C//G T] = C[C]NC[C]G C[T]. Thus, the C[G × T]/J = C[C]NC[C]G C[T] This proves the inclusion J ⊂ gr I. Hence, set theoretically, we have Supp(egrord M) ⊂ X. We know that Xrs is an open dense smooth subset of C ×C//G T, by Lemma 2.1.3(i). Hence, the ideal J is generically reduced. Further, we know that X is a Lagrangian subvariety (Lemma 2.1.5) and that the dimension of any irreducible component of the support of the sheaf egr M is ≥ dim X. We conclude that either M = 0 or else we have that [Supp(egr M)] = [X], so X is the only irreducible component of Supp(egr M) and this component occurs with multiplicity 1. In the latter case, M must be a simple holonomic D-module, by Lemma 2.2.1. Below, we mimic the proof of [HK1, Proposition 4.7.1] to show that M 6= 0. Let prg and prt denote the natural projections of xrs to grs and to tr, respectively. The restriction of the imbedding j to the open dense subset xrs ⊂ x gives a closed imbedding jrs : xrs ֒→ grs × tr. According to parts (i) and (iii) of Proposition 3.1.1 (of §3.1 below), there is a nowhere vanishing G-invariant section ω ∈ Kxrs of the canonical bundle. The element (dx dt)−1 ⊗ (jrs)∗ω is therefore a nonzero G-invariant section of the D-moduleRjrs Oxrs on grs × tr, see §2.2 and [HTT, p.22]. The radial part map rad is known to have the following property, [HC]: ∀u ∈ D(g)G, f ∈ C[g]G. δt · u(f )t = (rad u)(δt · ft), 15 (2.4.5) / / / / / / / / / / / / Using the above formula and an equation pr∗ t δt) · ω, see Proposition 3.1.1(iii) of §3 below, one verifies easily that the section (dx dt)−1 ⊗ (jrs)∗ω is annihilated by the ideal Igrs×tr , Therefore, similarly to [HK1, §4.7], we conclude that the assignment Dgrs×tr → Rjrs Oxrs, 1 7→ (dx dt)−1 ⊗ (jrs)∗ω, descends to a well defined nonzero D-module morphism g(dx) = (pr∗ (2.4.6) Mgrs×tr = (D/I)grs×tr −→ RjrsOxrs. (cid:3) We conclude that M 6= 0. It follows, as we have shown above, that M is a simple D- module. Hence, the map (2.4.6) must be an isomorphism. Moreover, we have M ∼= jrs !∗Oxrs = j!∗Or. Part (ii) of Lemma 2.4.3 follows. To complete the proof of part (i), we observe that the section ω provides a trivialization of the canonical bundle Kxrs. This implies, thanks to the isomorphism in (2.4.6) and formula (2.2.2), that we have (egr M)Xrs ∼= ONxrs . Hence, any function f ∈ gr I viewed as a function on G × T vanishes on the set Xrs. The set Xrs being Zariski dense in C ×C//G T, by Lemma 2.1.3(i), we deduce that the function f vanishes on the zero set of the ideal J. Hence, f ∈ √J by Hilbert's Nullstellensatz. The inclusion gr I ⊂ √J follows. 2.5. Hodge filtration on the Harish-Chandra module. The minimal extension j!∗Or has a canonical structure of Hodge D-module, [Sa], p.857, Corollary 2. This makes the Harish- Chandra module M a Hodge module via the isomorphism of Lemma 2.4.3(ii). Let F HodgeM be the Hodge filtration on M and egrHodge M be an associated graded sheaf. Observe further that the D-module j!∗Or is isomorphic to its Verdier dual so, we have D(M) ∼= M. Corollary 2.5.1. The sheaf egrHodge M is a Cohen-Macaulay coherent OG×T-module which is iso- morphic to its Grothendieck-Serre dual, up to a shift. In addition, we have Supp(egrHodge M) = X. X ֒→ Y , of smooth varieties, the Hodge filtration on the right DY -moduleR R Our normalization is determined by the requirement that, for any closed immersion f : f KX be equal Remark 2.5.2. The normalization of the Hodge filtration that we use in this paper differs by a degree shift from the one used by Saito [Sa]. to the order filtration introduced at the end of section 2.2. Thus, we have F Hodge 0, which is not the case in Saito's normalization, see [Sa, formula (3.2.2.3) and Lemma 5.1.9] and [HTT, p. 222]. Thus, from the discussion of §2.3 and Lemma 2.4.3(ii), we conclude (cid:0)R R f KX(cid:1) = −1 Degree shifts clearly do not affect the validity of Corollary 2.5.1. ♦ In the previous subsection, we have considered the order filtration on the Harish-Chandra module M. The isomorphism in (2.4.6) implies, in view of Remark 2.5.2, that the order and the Hodge filtrations agree on the open dense subset grs × tr. We do not know if these two filtrations agree on the whole of g × t. The following result, to be proved in §4.6, provides a partial answer. Proposition 2.5.3. With our normalization of the Hodge filtration on M, for any k ≥ 0, one has an inclusion F ord k M ⊂ F Hodge k M. According to Lemma 2.4.3, the isomorphism of Theorem 1.3.3 fits into the following chain of morphisms of coherent sheaves on G × T: OC×C//G T / egrordM / OX / ψ∗OXnorm 16 Theorem 1.3.3 egrHodge M. / / / / / /   / 2.6. Outline of proof of Theorems 1.3.3 and 1.3.4. We begin with the following standard result. Lemma 2.6.1. Let X be an irreducible scheme and ψ : Xnorm → Xred the normalization map. Let j : U ֒→ X be a Zariski open imbedding such that U is smooth and the set X r U has codimension ≥ 2 in X. Let F be a Cohen-Macaulay sheaf on X such that j∗F ∼= OU . Then, there are natural isomorphisms of OX-modules Proof. Let Z = X r U and let Hk Z. One has a standard long exact sequence of cohomology with support F ∼= j∗OU ∼= ψ∗OXnorm . (2.6.2) Z (F) denote the kth cohomology sheaf of F with support in 0 −→ H0 Z (F) −→ F a−→ j∗j∗F −→ H1 Z (F) −→ . . . where j∗ stands for a (nonderived) sheaf theoretic push-forward and a is the canonical ad- junction morphism. A maximal Cohen-Macaulay sheaf has no nonzero torsion subsheaves, [E, §21.4]. There- fore, the sheaf F is actually an OXred-module. In addition, we have H0 Z (F) = 0. Further, a well known general result says that, for any maximal Cohen-Macaulay sheaf F on X and any closed subscheme Z ⊂ X, one has a vanishing Hk Z (F) = 0 for all k < dim X − dim Z (the result can be reduced to the case of a local ring where it follows e.g. from [E, Theorem A.4.3]). Applying this result in our case and using the codimension ≥ 2 assumption of the lemma, we get that H1 Z (F) = 0. Thus, the morphism a in the long exact sequence above is an isomorphism. We deduce that F ∼= j∗j∗F ∼= j∗OU , the first isomorphism in (2.6.2). Observe next that the algebra structure on OU makes j∗OU an OXred -algebra. Let Y := Spec(j∗OU ) be the relative spectrum of that algebra. Thus, Y is a scheme equipped with a morphism f : Y → Xred that restricts to an isomorphism f : f −1(U ) ∼→ U ; by definition, we have j∗OU = f∗OY . The scheme Y is reduced and irreducible, since the algebra j∗OU clearly has no zero divisors. The codimension ≥ 2 assumption implies that j∗OU is a co- herent OXred -module hence f is a finite birational morphism. Hence, the set Y r f −1(U ) has codimension ≥ 2 in Y . We conclude that the scheme Y is smooth in codimension 1 and that it is Cohen-Macaulay, thanks to the isomorphism F ∼= j∗OU . It follows, by Serre's criterion, that Y is a normal variety; moreover, f = ψ is the normalization map so, we have j∗OU = f∗OY = ψ∗OXnorm. (cid:3) We need the following definition, motivated in part by [Ha1, Lemma 3.6.2]. Definition 2.6.3. Let Ci, i = 1, 2, be the set of pairs (x1, x2) ∈ C such that xi is a regular element of g. Put Crr = C1 ∪ C2. Clearly, each of the sets Ci, i = 1, 2, is an open subset of Cr. Thus, Crr is an open subset of C which is contained in the smooth locus of C. Furthermore, in §6.3 we will prove Lemma 2.6.4. The set C r Crr has codimension ≥ 2 in C. Corollary 2.6.5. The set Xrr := p−1(Crr) is a smooth Zariski open subset of X; furthermore, the set X r Xrr has codimension ≥ 2 in X. Proof of Corollary. Let Xi := p−1(Ci), i = 1, 2. According to Lemma 2.1.5, we have X1 = Nr, a Zariski open subset contained in the smooth locus of X. By symmetry, the set X2 is contained in the smooth locus of X as well. Therefore, Xrr = X1 ∪ X2 is a smooth Zariski open subset of X. Furthermore, the map p being finite, it follows from Lemma 2.6.4 that the set X r Xrr has codimension ≥ 2 in X. (cid:3) 17 Write j : Xrr ֒→ X for the open imbedding. A key role in the proof of Theorem 1.3.3 is played by the following result Proposition 2.6.6. There is a natural isomorphism j∗(egrHodge M) ∼= OXrr. The proof of this proposition will occupy most of sections 3 and 4. Our approach is based (x1, x2) × (t1, t2) ↔ (x2, x1) × (t2, t1). Therefore, it would be tempting to try to deduce Proposition 2.6.6 from the isomorphism on the Hotta-Kashiwara construction of M via the Springer resolution, see §4.5. Remark 2.6.7. We observe that the isomorphism M = j!∗Or of Lemma 2.4.3 and the fact that the canonical bundle Kr is trivial (see Proposition 3.1.1(i), (iii)) imply an isomorphism where σ is the following involution σ : G × T ↔ G × T, (egrHodge M)Nr ∼= ONr of coherent sheaves, see (2.2.2). Note further that we have X2 = σ(X1) (egrHodge M)Nr ∼= ONr by proving an isomorphism σ∗(egrHodge M) ∼= egrHodge M. Such an approach is motivated by the observation that the D-module M is isomorphic to its Fourier transform in the sense of D-modules, which is an immediate consequence of Definition 1.2.1. The functor σ∗ : Coh T ∗(g × t) → Coh T ∗(g × t) may be viewed as a 'classical analogue' of the Fourier transform of D-modules on g × t. Thus, Proposition 2.6.6 would follow from the invariance of M under the Fourier transform, had we known a general result saying mutes with Fourier transform. Unfortunately, such a result is not available at the moment of writing of this paper. Indeed, this seems to be a very difficult question (we are grateful to ♦ C. Sabbah for information on this subject). that the functor egrHodge(−) on (C×-monodromic) Hodge D-modules on a vector space, com- (2.6.8) Proposition 2.6.6 and Lemma 2.6.4 imply Theorem 1.3.3. Let J Hodge ⊂ OG×T be the annihilator of egrHodge M viewed as an OG×T-module and let X ⊂ G × T be a closed subscheme defined by the ideal J Hodge. By Corollary 2.5.1, the sheaf egrHodge M is Cohen-Macaulay and, set theoretically, one has X = X. Further, the isomorphism M ∼= j!∗Or of Lemma 2.4.3 implies that the ideal J Hodge is generically reduced. Hence, X is reduced and we have X = X, since X is reduced by definition. Thus, thanks to Corollary 2.6.5, we are in the setting of Lemma 2.6.1, where we take ֒→ X = X be the open imbedding. We see that Theorem 1.3.3 is a direct consequence of Lemma 2.6.1 combined with Proposition 2.6.6. (cid:3) F = egrHodge M and let j : U = Xrr Theorem 1.3.3 implies Theorem 1.3.4. By Theorem 1.3.3, we have egrHodge M ∼= ψ∗OXnorm. Fur- ther, thanks to Corollary 2.5.1, we know that the sheaf egrHodge M is Cohen-Macaulay and, moreover, it is isomorphic to its Grothendieck-Serre dual, up to a shift. Thus, the sheaf ψ∗OXnorm has similar properties. Since, Grothendieck's duality commutes with finite mor- phisms, we deduce that OXnorm is a Cohen-Macaulay sheaf which is, moreover, isomorphic to its Grothendieck-Serre dual, that is, to the dualizing sheaf KXnorm, up to a shift. Therefore, we have OXnorm ∼= KXnorm, completing the proof. (cid:3) 3. SPRINGER RESOLUTIONS 3.1. An analogue of the Grothendieck-Springer resolution. Let B be the flag variety, the variety of all Borel subalgebras b ⊂ g. Motivated by Grothendieck and Springer, we intro- duce the following incidence varieties eg := {(b, x) ∈ B × g x ∈ b}, resp. eG := {(b, x, y) ∈ B × g × g x, y ∈ b}. 18 B× g → B, resp. B× G → B. Given a Borel subgroup B ⊂ G with Lie algebra b = Lie B, we The first projection makeseg, resp. eG, a sub vector bundle of the trivial vector bundle have B ∼= G/B. That gives a G-equivariant vector bundle isomorphismeg ∼= G ×B b, resp. eG ∼= G ×B (b × b). Thus,eg andeG are smooth connected varieties. Recall that, for any pair b, b′ of Borel subalgebras of g, there is a canonical isomorphism b/[b, b] ∼= b′/[b′, b′], cf. eg. [CG, Lemma 3.1.26]. Given a Cartan subalgebra t ⊂ b, the ֒→ b ։ b/[b, b] yields an isomorphism t ∼→ b/[b, b]. Therefore, the assign- composite t ment (b′, x) 7→ x mod [b′, b′] ∈ b′/[b′, b′], resp. (b′, x, y) 7→ (x mod [b′, b′], y mod [b′, b′]) ∈ b′/[b′, b′] × b′/[b′, b′], gives a well defined smooth morphism ν : eg → t, resp. ν : eG → T. Finally, we introduce a projective G-equivariant morphism µ : eg → g, (b, x) 7→ x, resp. µ : eG → G, (b, x, y) 7→ (x, y). The map µ is known as the Grothendieck-Springer resolution. (i) The image of the map µ × ν :eg → g × t is contained in x = g ×g//G t. The resulting morphism π :eg → x is a resolution of singularities, so π−1(r) ∼→ r is an isomorphism. (ii) The image of the map µ × ν : eG → G × T is contained in G ×G//G T. The resulting map eG ։ [G ×G//G T]red is a proper birational morphism. (iii) The canonical bundle oneg has a natural trivialization by a nowhere vanishing G-invariant section ω ∈ Keg such that one has µ∗(dx) = (ν∗δt) · ω, cf. Notation 2.4.2. (iv) We have KeG = ∧dim B q∗TB (here TB is the tangent sheaf and q : eG → B is the projection). Proposition 3.1.1. Part (i) of Proposition 3.1.1 is well-known, cf. [BB], [CG]. The descriptions of canonical bundles in parts (iii)-(iv) are straightforward. The equation of part (iii) that involves the section ω appears eg. in [HK1], formula (4.1.4). This equation is, in essence, nothing but Weyl's classical integration formula. Part (ii) of Proposition 3.1.1 is an immediate consequence of Proposition 6.1.2, of §6.1 below. Lemma 6.1.1 implies, in particular, that the image of the map µ equals the set of pairs (x, y) ∈ G such that x and y generate a solvable Lie subalgebra of g. We remark also that the statement of Proposition 3.1.1(ii) is a variation of results concerning the null-fiber of the adjoint quotient map G → G//G, see [Ri2], [KW]. (cid:3) Question 3.1.2. Is the variety [G ×G//G T]red normal, resp. Cohen-Macaulay ? 3.2. Symplectic geometry interpretation. The map (b, x) 7→ (b, x, ν(b, x)) gives a closed immersion ǫ : eg ֒→ B × g × t. The image of this immersion is a smooth subvariety ǫ(eg) ⊂ B × g × t. Let Λ be the total space of the conormal bundle of that subvariety. Thus, Λ is a smooth C×-stable Lagrangian subvariety in T ∗(B × g × t) = T ∗B × (G × T). Let prΛ→T ∗B : Λ → T ∗B, resp. prΛ→G×T : Λ → G × T, denote the restriction to Λ of the projection of T ∗B × G × T to the first, resp. along the first, factor. Let N be the variety of nilpotent elements of g and put eN := {(b, x) ∈ B × g x ∈ [b, b]}. There is a natural isomorphism T ∗B ∼= eN of G-equivariant vector bundles on B that identifies the cotangent space at a point b ∈ B with the vector space (g/b)∗ ∼= [b, b], cf. [CG, ch. 3]. Let Φ : eN ∼→ T ∗B be an isomorphism obtained by composing the above isomorphism with the sign involution along the fibers of the vector bundle T ∗B. eκ : eG → eN , (b, x, y) 7→ (b, [x, y]). Let Ψ denote the map (Φ ◦eκ)× µ× ν : eG → T ∗B× G× T. Restricting the commutator map κ to each Borel subalgebra b ⊂ g yields a morphism 19 Proposition 3.2.1. The map Ψ yields an isomorphism eG ∼→ Λ that fits into a commutative diagram (3.2.2) µ×ν eκ G × T Id G × T prΛ→G×T eG Λ Ψ prΛ→T ∗ B eN Φ T ∗B the triple u = (b, x, x mod [b, b]). The fiber of the tangent bundle T (B × g × t) at the point u may be identified with the vector space Tu(B × g × t) = (g/b) × g × t. Hence, the tangent Proof. Fix b ∈ B and x ∈ b. So, (b, x) ∈eg and the corresponding point in B× g× t is given by space to the submanifold ǫ(eg) equals Now, write h−,−i for an invariant bilinear form on g and use it to identify the fiber of the cotangent bundle T ∗(B×g×t) at u with the vector space [b, b]×g×t. Let (n, y, h) ∈ [b, b]×g×t be a point of that vector space. Such a point belongs to Λu, the fiber at u of the conormal bundle on the subvariety ǫ(eg), if and only if the following equation holds Tu(ǫ(eg)) =(cid:8)(α mod b, [α, x] + β, β mod [b, b]) ∈ (g/b) × g × t (cid:12)(cid:12) α ∈ g, β ∈ b(cid:9). hα, ni + h[α, x] + β, yi + hβ mod [b, b], hi = 0 ∀α ∈ g, β ∈ b. (3.2.3) Taking α = 0 and applying equation (3.2.3) we get h[b, b], yi = 0. Hence, y ∈ b and h = −y mod [b, b]. Next, for any α ∈ g, we have h[α, x], yi = hα, [x, y]i. Hence, for β = 0 and any n ∈ [b, b], y ∈ b, equation (3.2.3) gives It follows that n + [x, y] = 0. We conclude that 0 = hα, ni + h[α, x], yi = hα, n + [x, y]i ∀α ∈ g. Λu = {(−[x, y], y, y mod [b, b]) ∈ [b, b] × g × t, y ∈ b}. (3.2.4) space in right hand side of (3.2.4) is equal to the image of the set −1(b, x) under the map We have a projection : eG → eg, (b, x, y) → (b, x) along the last factor. The vector Ψ = (Φ ◦eκ) × µ × ν. We conclude that the map Ψ gives an isomorphism of vector bundles eG →eg and Λ → ǫ(eg), respectively. 3.3. The schemeeX. We use the notation ı : X ֒→ T ∗X for the zero section of the cotangent We consider the following commutative diagram where the vertical map µfN is known as bundle on a variety X. (cid:3) the Springer resolution, ν / T (3.3.1) / G ×G//G T B = µ−1(0) ı: b 7→(b,0) µ / eN µfN {0}  / N  eκ / g κo eG µ G one has LeteX :=eκ−1(cid:0)ı(B)(cid:1) ⊂ eG, a scheme theoretic preimage of the zero section. Set theoretically, Diagram (3.3.1) shows that the morphism µ mapseX to C, resp. µ × ν mapseX to C ×C//G T. It follows from Proposition 3.2.1 that the map Ψ induces an isomorphism of schemes [x, y] = 0}. eX = {(b, x, y) ∈ B × g × g x, y ∈ b, eX =eκ−1(cid:0)ı(B)(cid:1) Ψ∼= pr−1 Λ→T ∗B(cid:0)ı(B)(cid:1) = Λ ∩ (B × G × T) (3.3.2) 20 / / o o / / o o     /     o o   /    /  / o / where the scheme structure on each side is that of a scheme theoretic preimage. LeteXrr := µ−1(Crr), a Zariski open subset of the schemeeX =eκ−1(cid:0)ı(B)(cid:1). Lemma 3.3.3. The differential of the morphismeκ : eG → eN is surjective at any point (b, x, y) ∈eXrr; in particular, the seteXrr is contained in the smooth locus of the schemeeX. Proof. Let x ∈ b and write adb x for the map b 7→ [b, b], u 7→ [x, u]. One has dim b − dim gx ≤ dim b − dim Ker(adb x) = dim Im(adb x) ≤ dim[b, b]. For x ∈ b ∩ gr, we have dim gx = dim t and the above inequalities yield dim[b, b] = dim b − dim gx ≤ dim(Im adb x) ≤ dim[b, b]. Thus, in this case, Im(adb x) = [b, b] i.e. the map adb x is surjective. Now, let (b, x, y) ∈ eXrr. Without loss of generality, we may assume that x is a regular element of g. The differential of the commutator map κ : b × b → [b, b] at the point (x, y) ∈ b× b is a linear map db,x,yκ : b⊕ b → [b, b] given by the formula db,x,yκ : (u, v) 7→ adb x(u)− adb y(v). We see that Im(adb x) ⊆ Im(db,x,yκ). By the preceding paragraph, we deduce that the map db,x,yκ is surjective. The lemma follows from this by G-equivariance. (cid:3) ♦ Recall the notation Xrr = p−1(Crr). Corollary 3.3.5. is a smooth Zariski open subset of Λ ∩ (B × G × T). isomorphism of algebraic varieties. Remark 3.3.4. The schemeeX is not irreducible, in general, cf. [Ba]. So, the open seteXrr is not necessarily dense ineX. We use Proposition 3.2.1 to identify the seteXrr ⊂eX ⊂ eG with a subset of Λ ∩ (B × G × T). (i) The varieties Λ and B × G × T meet transversely at any point ofeXrr, soeXrr (ii) The morphism π : eXrr → Xrr, the restriction of the morphism µ × ν to the set eXrr, is an Proof. Part (i) is equivalent to the statement of Lemma 3.3.3 . To prove (ii), let X denote the preimage of Crr under the first projection C ×C//G T → C, so we have Xred = Xrr. It is clear that the morphism µ × ν maps eXrr to X. We claim that the resulting map π : eXrr → X is a set theoretic bijection. Indeed, it is surjective, by Lemma 2.1.3(i), since the image of this map contains the set Crs ×C//G T and µ × ν, hence also π, is a proper morphism. To prove injectivity, we interpret the map (b, x, y) 7→ (x, y) as a composition of the imbedding eX ֒→ eg × g, (b, x, y) 7→ (b, x) × y and the map π × Idg : eg × g → x × g. This last map gives a bijection between the set r × g and its preimage ineg × g, thanks to Proposition 3.1.1(i). Our Recall thateXrr is a smooth scheme by Lemma 3.3.3. Hence, the scheme theoretic image of eXrr under the morphism π is actually contained in Xrr = Xred. The reduced scheme Xrr is smooth by Corollary 2.6.5. Thus, π :eXrr → Xrr is a morphism of smooth varieties, which is There is an alternative proof that the morphism π :eXrr → Xrr is ´etale based on symplectic In more detail, put X = B and Y = g × t and let ǫ : eg ֒→ X × Y be the imbedding, cf. §3.2. We have smooth locally closed subvarieties r ⊂ g × t, π−1(r) ⊂eg and Z := ǫ(π−1(r)) ⊂ X × Y , respectively. We use Proposition 2.1.5, resp. Proposition 3.2.1, to identify X1 with Nr, resp. π−1(gr × g) ∩ Λ with NZ . We know that the projection X × Y → Y induces an isomorphism Z ∼→ r, by Proposition 3.1.1(i). Now, one can prove a general result saying that, in this case, the map (X × T ∗Y ) ∩ NZ → Nr induced by the projection T ∗X × T ∗Y → T ∗Y is ´etale at any point where NZ meets the subvariety X × T ∗Y ⊂ T ∗(X × Y ) transversely. The latter condition holds in our a set theoretic bijection. Such a morphism is necessarily an isomorphism, by Zariski's main theorem, and part (ii) follows. claim follows. geometry. 21 case thanks to part (i) of the Corollary. This implies the isomorphism π : π−1(C1) ∼→ X1. The isomorphism π : π−1(C2) ∼→ X2 then follows by symmetry. 3.4. A DG algebra. In this subsection, we construct a sheaf A of DG OeG-algebras such that H0(A), the zero cohomology sheaf, is isomorphic to the structure sheaf of the closed sub- (cid:3) schemeeX ⊂ eG. To explain the construction, let T := TB and write ı : B ֒→ T ∗B for the zero section, resp. q : T ∗B → B for the projection. The sheaf q∗T ∗ on T ∗B comes equipped with a canonical Euler section eu such that, for each covector ξ ∈ T ∗B, the value of eu at the point ξ is equal to ξ. Further, there is a standard Koszul complex . . . → ∧3q∗T → ∧2q∗T → q∗T → 0 with differential ∂eu given by contraction with eu. The complex (∧ qq∗T , ∂eu) is a locally free resolution of the sheaf ı∗OB on T ∗B. We may use the isomorphism T ∗B ∼= eN to view the morphismeκ as a map eG → T ∗B. Hence, the pull-back of the Koszul complex above via the mapeκ is a complex of locally free sheaves oneG that represents the object Leκ∗(ı∗OB) ∈ Db Let q : eG → B denote the first projection, so we haveeκ∗q∗T = q∗T and one may identify eκ∗eu, the pull-back of the section eu, with a section of q∗T ∗. For each n ≥ 0, let An = ∧n q∗T = ∧neκ∗q∗T , where A0 := OeG. Contraction with eκ∗eu gives a differential ∂eκ∗eu : A q → A q−1. Thus, we may (and will) view A := Ln≥0 An as a sheaf of coherent DG OeG- algebras, with multiplication given by the wedge product and with the differential ∂eκ∗eu, a graded derivation of degree (−1). Notation 3.4.1. Write Hj(F) ∈ Coh X for the jth cohomology sheaf of an object F ∈ Db coh(X). By construction, one has H0(A, ∂eκ∗eu) = OeX, cf. (3.3.2). Thus, one may view A as the structure sheaf of a certain DG scheme, a "derived analogue" of the schemeeX. The DG algebra A has also appeared, in an implicit form, in a calculation in [SV]. Remark 3.4.2. The DG algebra A is concentrated in degrees 0 ≤ i ≤ d := dim B and we have Ad = KeG, by Proposition 3.1.1(iv). It follows that A is a self-dual DG algebra in the sense that multiplication in A yields an isomorphism of complexes coh(eG). Ad− q (A q, Ad) = Hom ∼→ Hom 4. PROOF OF THE MAIN THEOREM O eG O eG (A q,KeG). Hotta and Kashiwara proved the following important result, [HK1, Theorem 4.2]. 4.1. The Hotta-Kashiwara construction. The structure sheaf Oeg has an obvious structure of eg-module via the holonomic left D eg-module. So, one hasRµ×ν Oeg, the direct image of this D D-module set theoretically supported on the variety x ⊂ g × t. projective morphism µ × ν, cf. §3.1. Each cohomology group Hk(Rµ×ν Oeg) is a holonomic Theorem 4.1.1. For any k 6= 0, we have Hk(Rµ×ν Oeg) = 0; furthermore, there is a natural isomor- phism H0(Rµ×ν Oeg) ∼= j!∗Or of D-modules. The D-module H0(Rµ×ν Oeg) has a canonical nonzero section (dx dt)−1N(µ × ν)∗ω, cf. I ′ := D · (ad g ⊗ 1) + D · {P − rad(P ) P ∈ C[g]G} + D · {Q − rad(Q) Q ∈ (Sym g)G} Proposition 3.1.1(iii). Hotta and Kashiwara [HK1, §4.7] show that this section is annihilated by the left ideal (we have exploited their argument in the proof of Lemma 2.4.3). 22 Furthermore, it is proved in [HK1, §4.7] that D/I ′ is a simple D-module and the assign- ment u 7→ u[(dx dt)−1N(µ × ν)∗ω] yields a D-module isomorphism D/I ′ ∼→ H0(Rµ×ν Oeg). An alternative, more direct, proof of an analogous isomorphism may be found in [HK2]. Remark 4.1.2. Hotta and Kashiwara defined the Harish-Chandra module as the quotient D/I ′. Comparing formula (1.2.2) with the definition of the ideal I ′ we see that one has I ′ ⊆ I. This inclusion of left ideals yields a surjection D/I ′ ։ D/I = M. Since D/I ′ is a simple D-module, this surjection must be an isomorphism. So, Definition 1.2.1 is equivalent, a posteriori, to the one used by Hotta and Kashiwara. 4.2. Direct image for filtered D-modules. Let f : X → Y a proper morphism. The di- coh(FDY ), (M, F ) 7→ rect image functorRf can be upgraded to a functor Db Rf (M, F ) between filtered derived categories, cf. [La, §4], [Sa, §2.3]. The latter functor is known to commute with the associated graded functorgdgr(−). We will only need a special case of this result for maps of the form f : X × Y → Y, the projection along a proper variety X. In such a case, one has a diagram coh(FDX ) → Db T ∗Y X × T ∗Y ς=ı×IdT ∗Y / T ∗X × T ∗Y = T ∗(X × Y ). (4.2.1) Here, denotes the second projection and ı is the zero section. coh(FDX×Y ), in Db coh(T ∗Y ) there is a functorial isomorphism The relation between the functorsgdgr(−) andRf is provided by the following result, see gdgr(cid:16)Rf (M, F )(cid:17) = R∗(cid:2)(KX ⊠ OT ∗Y )NOX×T ∗Y [La], formula (5.6.1.2). Theorem 4.2.2. Let f : X × Y → Y be the second projection where X is proper. Then, for any (M, F ) ∈ Db The cohomology sheaves H q(cid:0)Rf (M, F )(cid:1) of the filtered complexRf (M, F ) come equipped ing egr H q(cid:0)Rf (M, F )(cid:1), the associated graded sheaves. A theorem of Saito stated below says coh(FDX ), so there is a well defined objectRf (M, F ) ∈ be viewed as an object (M, F ) ∈ Db Db coh(FDY ). One of the main results of Saito's theory reads, see [Sa, Theorem 5.3.1]: Let f : X → Y be a proper morphism of smooth varieties. A Hodge module on X may with an induced filtration. However, Theorem 4.2.2 is not sufficient, in general, for describ- Lς ∗gdgr(M, F )(cid:3). ✷ that, in the case of Hodge modules, Theorem 4.2.2 is indeed sufficient for that. coherent sheaf on T ∗Y . Similar notation will be used for right D-modules. Theorem 4.2.3. For any (M, F ) ∈ HM(X) and any projective morphism f : X → Y , the filtered natural structure of a Hodge module on Y . complexRf (M, F ) is strict, cf. Definition 2.3.1, and each cohomology group Hj(Rf (M, F )) has the In the situation of the theorem, we refer to the induced filtration on Hj(Rf (M, F )), j = 0, 1, . . ., as the Hodge filtration and let egrHodge Hj(Rf (M, F )) denote the associated graded 4.3. Key result. We recall the setup of §3.2. Thus, we have the immersion ǫ : eg ֒→ B× g× t, (b, x) 7→ (b, x, ν(b, x)) and we write Λ ⊂ T ∗(B × g × t) for the total space of the conormal bundle on ǫ(eg). We will view the structure sheaf OΛ as a coherent sheaf on T ∗B × G × T supported on Λ. 23 o o o o   / In the special case where X = B and Y = g × t diagram (4.2.1) takes the form Λ ⊂ T ∗B × G × T ς B × G × T / / G × T. (4.3.1) Theorem 4.3.2. All nonzero cohomology sheaves of the object R∗Lς∗ OΛ ∈ Db and, we have coh(G × T) vanish H0(R∗Lς ∗ OΛ) = egrHodge M. B×g×t ⊗ E, the corresponding left DB×g×t-module, we find B×g×t⊗ǫ∗Oeg) = Lς ∗q∗(K−1 B ⊗ (−) = (K−1 B ⊠Og×t) = K−1 B ⊠ OG×T)NOB×G×T We factor the map µ×ν as a composition of the closed imbedding ǫ and a proper projection Proof. Let E :=R R ǫ Keg, a right Hodge D-module on B × g × t. Using the notation of Remark 2.5.2, from (2.2.2), we obtain grHodge E = q∗(ǫ∗Keg) = q∗(ǫ∗Oeg), where in the last equality we have used that the canonical bundle oneg is trivial, see Proposition 3.1.1(iii). Therefore, for Rǫ Oeg = K−1 Lς ∗(cid:0)egrHodge(cid:0)Rǫ Oeg(cid:1)(cid:1) = Lς∗q∗(K−1 B ⊗Lς ∗OΛ, (4.3.3) where we have used simplified notation K−1 f : B × g × t → g × t along the first factor. We getRµ×ν Oeg =Rf(cid:0)Rǫ Oeg(cid:1). Hence, applying Theorem 4.2.2 to the D-module M = Rǫ Oeg and using (4.3.3), we obtaingdgr(cid:0)Rµ×ν Oeg(cid:1) = gdgr(cid:0)Rf (Rǫ Oeg)(cid:1) = R∗Lς∗OΛ. Thus, by Theorem 4.2.3 applied to the Hodge module Oeg, we get egrHodge Hj (Rµ×ν Oeg) = Hj(gdgrRµ×ν Oeg) = Hj(R∗Lς ∗OΛ) for any j ∈ Z. We conclude that Hj (R∗Lς∗OΛ) = 0 for any j 6= 0, thanks to Theorem 4.1.1. is determined by the Hodge structure on Or. The morphism µ × ν : eg → x is generi- structures. Therefore, the isomorphism M ∼= H0(Rµ×ν Oeg), which is based on the isomor- phism M ∼= j!∗Or of Lemma 2.4.3(ii), also respects the Hodge filtrations. We deduce that egrHodge M = H0(R∗Lς∗OΛ). Finally, we observe that the Hodge structure on the minimal extention D-module j!∗Or cally an isomorphism. It follows that the isomorphism of Theorem 4.1.1 respects the Hodge 4.4. Let X be a smooth variety and let iY : Y ֒→ X, resp. imbeddings of smooth subvarieties. Below, we will use the following simple iZ : Z ֒→ X, be closed (−). (cid:3) Lemma 4.4.1. In Db coh(X), there are canonical quasi-isomorphisms (iY )∗Li∗ Y [(iZ )∗OZ ] ∼= (iY )∗OY (4.4.2) Proof. Let ∆X : X → X × X be the diagonal imbedding and define a map ∆Y X : Y → Y × X, y 7→ (y, iY (y)). We have a cartesian diagram of closed imbeddings (iZ )∗OZ ∼= (iZ )∗Li∗ Z [(iY )∗OY ]. LNOX ∆Y X ∆X Y iY X / Y × X iY ×IdX / X × X For any F ∈ Db coh(X), we have [(iY )∗OY ] ⊠ F = (iY × IdX)∗f ∗F, where f : Y × X → X is the second projection. Therefore, we obtain [(iY )∗OY ] LNOX F = ∆∗ X([(iY )∗OY ] ⊠ F) = ∆∗ = (iY )∗[∆∗ 24 X(iY × IdX )∗(f ∗F) Y X(f ∗F)] = (iY )∗(f ◦ ∆Y X )∗F = (iY )∗i∗ Y F, ? _ o o / /   /   / where the third isomorphism is a consequence of proper base change with respect to the cartesian square above. Taking here F := (iZ )∗OZ yields the isomorphism on the left of (4.4.2). The isomorphism on the right of (4.4.2) is proved similarly. (cid:3) Associated with diagram (3.3.1), there are derived functors R(µ×ν)∗ Leκ∗ ı∗ / Db / Db / Db Db coh(B) coh(eG) coh(eN ) (4.4.3) In §3.4, we considered a Koszul complex (∧ qq∗T , ∂eu), a locally free resolution of the coh(eG) may be represented by the sheaf ı∗OB on T ∗B. Therefore, the object Leκ∗(ı∗OB) ∈ Db DG OeG-algebra (A, ∂eκ∗eu) =eκ∗(∧ qq∗T , ∂eu) introduced in §3.4. R∗Lς ∗((iΛ)∗OΛ) ≃ R(µ × ν)∗A. The following result provides a link between the DG algebra A and Theorem 4.3.2. coh(G × T), there is a natural isomorphism Lemma 4.4.4. In Db coh(G × T). Proof. To simplify notation, we put Z := B × G × T. Further, write prT ∗B, resp. prG×T, for the projection of the variety T ∗B × G × T to the first, resp. along the first, factor. Clearly, we have ς∗OZ = pr∗ T ∗B(ı∗OB). Also, using the notation of diagram (3.2.2), we get prΛ→T ∗B = prT ∗B ◦ iΛ, resp. prΛ→G×T = iΛ ◦ prG×T. We deduce a chain of isomorphisms Li∗ Λ(ς ∗OZ ) = Li∗ Λ[pr∗ T ∗B(ı∗OB)] = L(iΛ ◦ prT ∗B)∗(ı∗OB) = L pr∗ Λ→T ∗B(ı∗OB). Next, we use the isomorphism Φ : eN ∼→ T ∗B, see §3.2, to identify the imbedding B ֒→ eN with the zero section ı : B ֒→ T ∗B. Thus, from the above chain of isomorphisms, using commutative diagram (3.2.2), we get Λ→T ∗B(ı∗OB) ∼= R(µ × ν)∗Leκ∗(ı∗OB). (4.4.5) R(prΛ→G×T)∗Li∗ Now, we apply Lemma 4.4.1 in the case where X = T ∗B× G× T and Y = Λ. So, we have Λ[ς ∗OZ ] = ς∗Lς∗[(iΛ)∗OΛ]. iZ = ς and the composite isomorphism in (4.4.2) yields (iΛ)∗Li∗ Note that = ς ◦ prG×T, so R∗ = (RprG×T)∗ς ∗. Hence, we obtain Λ(ς ∗OZ ) = R(prΛ→G×T)∗L pr∗ R∗Lς∗[(iΛ)∗OΛ] = R(prG×T)∗ς ∗ς ∗[(iΛ)∗OΛ] (4.4.2) == R(prG×T)∗(iΛ)∗Li∗ Λ[ς ∗OZ ] (4.4.5) == R(prΛ→G×T)∗Li∗ Here, the object on the right is R(µ × ν)∗A, and the lemma is proved. Λ(ς ∗OZ ) = R(µ × ν)∗Leκ∗(ı∗OB). Corollary 4.4.6. The sheaves Hk(cid:0)R(µ × ν)∗A(cid:1) vanish for all k 6= 0 and there is an OG×T-module isomorphism H0(cid:0)R(µ × ν)∗A(cid:1) ∼= egrHodge M. Proof. This is an immediate consequence of Theorem 4.3.2 and Lemma 4.4.4. (cid:3) (cid:3) 4.5. Completing the proof of Theorem 1.3.3. We recall that in order to complete the proof of Theorem 1.3.3 it remains to prove Proposition 2.6.6 and Lemma 2.6.4. The proof of the lemma will be given later, in §6.3. Proof of Proposition 2.6.6. Let A0 be the degree zero homogeneous component of A. We may view A0 as a DG algebra equipped with zero differential and concentrated in degree zero. Thus, one has a natural DG algebra imbedding f : (A0, 0) ֒→ (A, ∂eκ∗eu). OeG = π∗OG×T, we obtain the following chain of DG algebra morphisms To simplify notation, put π := µ × ν. Applying the functor Rπ∗ and using that A0 = adjunction OG×T / Rπ∗π∗OG×T = Rπ∗A0 25 Rπ∗(f ) / Rπ∗A. (4.5.1) / / / / / Let Grr ⊂ G be a Zariski open subset such that (Grr × T) ∩ X = Xrr. From Lemma 3.3.3 and Corollary 3.3.5(ii), we deduce that the composite morphism in (4.5.1) induces an iso- ∼→ Rπ∗AGrr×T. Combining the latter isomorphism with the isomorphism morphism OXrr H0(cid:0)Rπ∗A(cid:1) ∼= egrHodge M of Corollary 4.4.6 yields the required isomorphismOXrr The statement of the next result was suggested to me by Dmitry Arinkin. ∼→ ∗(egrHodge M). (cid:3) supported on X, see Corollary 2.5.1. coh(G × T) is an To establish the DG OG×T-algebra quasi-isomorphism, one can argue as follows. Let F := Theorem 4.5.2. There is G × C× × C×-equivariant DG OG×T-algebra quasi-isomorphism: R(µ × ν)∗A ≃ ψ∗OXnorm. Proof. The fact that R(µ× ν)∗A and ψ∗OXnorm are isomorphic as objects of Db immediate consequence of Theorem 1.3.3 and Corollary 4.4.6. H0(cid:0)R(µ×ν)∗A(cid:1). We claim that the sheaf F is Cohen-Macaulay. This follows from Corollary 4.4.6 since we know that egrHodge M is a Cohen-Macaulay coherent sheaf set theoretically There is also an alternative proof of the claim that does not use the Cohen-Macaulay property of associated graded sheaves arising from Hodge modules. That alternative proof is based instead on the self duality property of the DG algebra A, see Remark 3.4.2. The latter property, combined with the fact that the morphism µ × ν is proper, implies that the object R(µ × ν)∗A ∈ Db coh(G × T) is isomorphic to its Grothendieck-Serre dual, up to a shift. Therefore, the cohomology vanishing from Corollary 4.4.6 forces the sheaf F = H0(cid:0)R(µ × ν)∗A(cid:1) be Cohen-Macaulay. The claim follows. Now, according to the proof of Proposition 2.6.6 given above, the composite morphism in (4.5.1) induces a G× C× × C×-equivariant OG×T-algebra isomorphism OXrr ∼→ ∗F. Thus, Lemma 2.6.1 provides a G × C× × C×-equivariant algebra isomorphism ψ∗OXnorm ∼→ F. (cid:3) We observe next that H q(eG, A), the hyper-cohomology of the DG algebra (A, ∂eκ∗eu), ac- quires the canonical structure of a graded commutative algebra. From Theorem 4.5.2, for any k ∈ Z, we deduce Hk(eG, A) = Hk(G × T, R(µ × ν)∗A) = H k(G × T, ψ∗OXnorm). The group on the right vanishes for any k 6= 0 since the scheme G× T is affine. So, we obtain Corollary 4.5.3. The hyper-cohomology groups Hk(eG, A) vanish for all k 6= 0 and there is a G- equivariant bigraded C[G × T]-algebra isomorphism H0(eG, A) ∼= C[Xnorm]. 'derived analogue' of the schemeeX, cf. §3.3. Then, Corollary 4.5.3 says that the morphism Remark 4.5.4. Write X := Spec A for the DG scheme associated with the DG algebra A in the sense of derived algebraic geometry, cf. [TV]. The DG scheme X may be thought of as a µ × ν induces a DG-algebra quasi-isomorphism C[Xnorm] → RΓ(X , OX ). This may be interpreted as saying that the DG scheme Spec A provides, in a sense, a 'DG resolution' of the variety Xnorm. (cid:3) 4.6. Proof of Proposition 2.5.3. Let f : B × (g × t) → g × t be the second projection. DB×g×t vanish and one has an f Lemma 4.6.1. All nonzero cohomology sheaves of the complexR R isomorphism H0(R R DB×g×t) ∼= Dg×t of right Dg×t-modules. 26 f RΓ(B, K q) ⊗C Dg×t = Dg×t. Proof. There is a standard Koszul type complex K q with terms K j = DB ⊗OB ∧−jTB, j = 0,−1, . . . ,− dim B, that gives a resolution of the structure sheaf OB, cf. [HTT, Lemma 1.5.27]. Using that H 0(B,OB) = C and H k(B,OB) = 0 for any k 6= 0 we deduce that RΓ(B, K q) = RΓ(B,OB) = C. Therefore, an explicit construction of direct image for right D-modules (see [HTT, Proposition 1.5.28] for its left D-module counterpart) shows thatR R DB×g×t = (cid:3) Next, we recall the setting of §4.3. We observe that E = R R ǫ Keg is a cyclic right DB×g×t- module generated by the section ǫ∗(ω ⊗ 1) where 1 ∈ Deg→B×g×t = DB×g×teg. Therefore, the assignment 1 7→ ǫ∗(ω ⊗ 1) can be extended to a surjective morphism γ : DB×g×t ։ E of right DB×g×t-modules. The quotient filtration on E induced by the projection γ is equal, by our normalization of the Hodge filtration, to the Hodge filtration F HodgeE, see Remark DB×g×t, to γ, a morphism of filtered D-modules, one 2.5.2. This filtration on E, resp. the order filtration on DB×g×t, makesR R a filtered complex. Applying the functorR R DB×g×t →R R obtains a morphismR R g×t via the trivialization of the canonical bundle on g× t provided by the section dx dt. Thus, using Lemma 4.6.1, we obtain a chain of morphisms of left Dg×t-modules f f E of filtered complexes. f E, resp.R R We identify the sheaf Dg×t with Dop f γ : R R f f f R R f γ f f (4.6.2) DB×g×t) g×t ⊗ H0(R R g×t ⊗ H0(R R / K−1 f E) = H0(Rµ×ν Oeg) = M. the natural projection Dg×t ։ Dg×t/I = M. Dg×t = K−1 It is straightforward to see, using the explicit formula u 7→ u[(dx dt)−1N(µ× ν)∗ω] for the isomorphism M ∼→ H0(Rµ×ν Oeg), cf. §4.1, that the composite morphism in (4.6.2) is equal to The proof of Lemma 4.6.1 shows that the filtration on the D-module H0(R R duced by the filtered structure onR R DB×g×t) in- DB×g×t goes, under the isomorphism of the Lemma, to the standard order filtration on the sheaf Dg×t. It follows that all the maps in (4.6.2) re- spect the filtrations. Thus, writing ¯γ : Dg×t ։ M for the composite map in (4.6.2), we get ¯γ(F ord k M for any k ∈ Z. Proposition 2.6.6 follows from this since the order filtration F ordM was defined as the quotient filtration on Dg×t/I. Dg×t) ⊂ F Hodge k f 5. A GENERALIZATION OF A CONSTRUCTION OF BEILINSON AND KAZHDAN 5.1. The universal stabilizer sheaf. Given a G-action on an irreducible scheme X, one defines the "universal stabilizer" scheme GX as a scheme theoretic preimage of the diag- onal in X × X under the morphism G × X → X × X, (g, x) 7→ (gx, x). Set theoreti- cally, one has GX = {(g, x) ∈ G × X gx = x}. The group G acts naturally on GX by g1 : (g, x) 7→ (g1gg−1 1 , g1x). The second projection GX → X, (g, x) → x gives GX the natural structure of a G-equivariant group scheme over X. The Lie algebra of the group scheme GX is a coherent sheaf gggX := Ker(g⊗OX → TX), the kernel of the natural "infinitesimal action" morphism. Let ggg∗ Let L be a finite dimensional g-representation. There are natural morphisms of sheaves X := Hom OX (gggX,OX ) be the (nonderived) dual of gggX . L ⊗ OX −→ HomC(g, L) ⊗ OX = L ⊗ g∗ ⊗ OX −→ L ⊗ ggg∗ X. (5.1.1) Here, the first morphism is induced by the linear map L → HomC(g, L) resulting from the g-action on L, and the second morphism is induced by the sheaf imbedding gggX → g ⊗ OX . 27 / Let LgggX denote the kernel of the composite morphism in (5.1.1). Thus, LgggX is a G- equivariant coherent subsheaf of L ⊗ OX . The geometric fiber of the sheaf LgggX at a suf- ficiently general point x ∈ X equals Lgx . Lemma 5.1.2. For an irreducible normal variety X with a G-action, we have provided general points of X have connected stabilizers. (i) The sheaf imbedding LgggX → L⊗OX induces an isomorphism Γ(X, LgggX )G ∼→ (L⊗ C[X])G, (ii) Let j : U ֒→ X be an imbedding of a Zariski open subset such that dim(X rU ) ≤ dim X−2. Then, the canonical morphism LgggX → j∗j∗(LgggX ), resp. gggX → j∗j∗(gggX ), is an isomorphism. Proof. For any x ∈ X and any G-equivariant morphism f : X → L, the element f (x) ∈ L is fixed by the group Gx. This implies (i). ∼→ j∗j∗(g ⊗ OX ), due to normality of X. To prove (ii), we use the isomorphism g ⊗ OX Thus, we deduce gggX = Ker[j∗(g ⊗ OU ) → TX] = j∗(cid:0) Ker[g ⊗ OU → j∗TX](cid:1) = j∗gggU . X) and, by the definition of the sheaf LgggX , we have s′U = 0. Next, let V ⊂ X be an open subset and let s ∈ Γ(V, j∗j∗(LgggX )). One may view s as a morphism V ∩ U → L. This morphism can be extended to a regular map ¯s : V → L, since X is normal. Let s′ be the image of ¯s under the composite morphism in (5.1.1). Thus, s′ ∈ Γ(V, L ⊗ ggg∗ X is torsion free. Hence, s′V ∩U = 0 implies that s′ = 0. We deduce that ¯s is actually a section (over V ) of the sheaf LgggX , the kernel of the map (5.1.1). Since ¯sU = s, we have proved that the morphism LgggX → j∗j∗(LgggX ) is surjective. This morphism is injective since LgggX , being a subsheaf of a locally free sheaf L ⊗ OX , is a torsion free sheaf. Now, the dual of a coherent sheaf is a torsion free sheaf. Therefore, the sheaf L ⊗ ggg∗ (cid:3) Fix a finite dimensional g-module L and a Borel subalgebra b. Let L = ⊕λ∈X Lλ 5.2. A canonical filtration. Let X be the weight lattice of g. Given a Borel subalgebra b, we identify elements of X with linear functions on b/[b, b]. We let R (= the set of roots), resp. R+ (= the set of positive roots) be the set of nonzero weights of the ad b-action on g, resp. on g/b. We introduce a partial order on X by setting λ′ 2 λ if λ − λ′ is a sum of positive roots. t be the weight decomposition of L with respect to a Cartan subalgebra t ⊂ b. For any λ ∈ X, we put L2λ t . This is a b-stable subspace of L that does not depend on the choice of a Cartan subalgebra t ⊂ b. Therefore, the above construction associates with each Borel subalgebra b a canonically defined b-stable filtration, to be denoted by L2 q := Lλ′2λ, λ′6=λ Lλ′ := Lλ′2λ Lλ′ b , labeled by the partially ordered set X. , resp. L≺λ t t t One may let the Borel subalgebra b vary inside the flag variety. The family {L2 q of filtrations on L, then gives a filtration L2 q B = L2λ locally free OB-subsheaves. We put Lλ graded sheaf. This is a G-equivariant locally free sheaf on B. b , b ∈ B}, B of the trivial sheaf L ⊗ OB by G-equivariant B /L≺λ B, an associated B and let LB :=Lλ∈X Lλ In the special case L = g, the adjoint representation, one has L20 b = [b, b], for any b ∈ B. Furthermore, the above construction yields, for each root α ∈ R, a G- equivariant line bundle gα b = b, resp. L≺0 B . Thus, we have B := g2α B /g≺α gB := g0 B L (cid:0) ⊕α∈R gα B(cid:1). (5.2.1) The Lie bracket on g induces a natural fiberwise bracket on the locally free sheaf gB. Further, it is easy to check that the invariant bilinear form on g induces an isomorphism g−α B)∗, of line bundles on B. B ∼= (gα 28 B ⊗ Lλ B → Lλ+α Given a finite dimensional representation L, the action map g ⊗ L → L induces, for any B , of G-equivariant locally free α ∈ R and λ ∈ X, a well defined morphism gα sheaves on B. This way, one obtains a natural 'action' morphism gB ⊗ LB → LB. Note that any Borel subgroup B ⊂ G acts trivially on the vector space L20 b . It fol- lows that there is a canonical isomorphism L20 b′ , for any Borel subalgebras b and b′. In other words, the OB-sheaf L0 B comes equipped with a canonical G-equivariant trivialization. In the case of the adjoint representation the canonical trivialization reads g0 B = h ⊗ OB, where the vector space h is referred to as the universal Cartan algebra of g. More generally, for an arbitrary representation L, one has the canonical G-equivariant iso- B ∼= Lh ⊗ OB, where the vector space Lh may be called the 'universal zero weight morphism L0 space' of L. The group G acts on Lh ⊗ OB through its action on OB. Remark 5.2.2. In the above discussion, we have implicitly viewed elements of the weight lattice X as linear functions on h, the universal Cartan algebra. b = L20 b′ /L≺0 b /L≺0 b /L≺0 be played by the following commutative diagram 5.3. W -action. Letegr = µ−1(gr), a Zariski open subset ofeg. An important role below will egr π ∼ / r = gr ×t/W t (RRRRRRRRRRRRRRRRRR µ pr gr eγ t ϑ (5.3.1) γ / gr/G = t/W, In this diagram, the map pr is the first projection, ϑ is the quotient map, eγ is the second projection and γ is the adjoint quotient morphism. The map π, in the diagram, is the iso- morphism from Proposition 3.1.1(i). Note that the square on the right of diagram (5.3.1) is cartesian, by definition. via the vector bundle projectioneg → B. We obtain a filtration L2 q Given a g-representation L, we pull-back the canonical filtration on the trivial sheaf L ⊗ OB eg , of the sheaf L ⊗ Oeg, by are locally free, and so is Leg = eg , an associated graded sheaf. In particular, for λ = 0 and L = g, the adjoint repre- The Weyl group W acts on the fibers of the map pr in (5.3.1). We may transport this action G-equivariant subsheaves. The sheaves Lλ ⊕λ∈X Lλ sentation, one gets the sheaf g0 eg = h ⊗ Oeg. eg := L2λ eg /L≺λ eg there is a unique Borel subalgebra b′ such that we have w(b, x) = (b′, x). via the isomorphismegr ∼→ r to get a W -action onegr. Thus, for any w ∈ W and (b, x) ∈egr Assume now that the element x is regular semisimple. Then, gx is a Cartan subalgebra of g contained in b ∩ b′. Thus, one has the following chain of isomorphisms ∼→ b′/[b′, b′] = g0 (5.3.2) Here, the composite map from the copy of h on the left to the copy of h on the right is the map h → h, h 7→ w(h). One also has the dual map h∗ → h∗, λ 7→ w(λ). It follows that, given λ ∈ X, there is an analogoue of diagram (5.3.2) for any g-representation L; it reads: h = g0 eg(cid:12)(cid:12)(b,x) = b/[b, b] ∼← gx eg(cid:12)(cid:12)w(b,x) = h. (cid:12)(cid:12)w(b,x). = Lw(λ) eg Lλ B(cid:12)(cid:12)(b,x) = L2λ a canonical isomorphism b /L≺λ b ∼← Lλ gx ∼→ L2w(λ) b′ /L≺w(λ) b′ Now, we let the Borel subalgebra b vary. We conclude that the above construction yields w∗(Lλ eg )(cid:12)(cid:12)egrs ∼= Lw(λ) eg (cid:12)(cid:12)egrs, 29 ∀λ ∈ X, w ∈ W, (5.3.3) ( / / /     / of G-equivariant locally free sheaves onegrs := µ−1(grs): In the special case where λ = 0, we have that L0 eg = Lh ⊗ Oeg is a trivial sheaf. Hence, egr , canonically. Thus, the isomorphism in (5.3.3) yields a canonical W -action on w∗(L0 the universal zero weight space Lh. egr ) = L0 5.4. The morphisms λk and λL. In this section, we are interested in the universal stabilizer contruction of §5.1 in the special case where the group G acts on X = gr by the adjoint action. Below, we will use simplified notation ggg := ggggr . Thus, ggg is a rank r locally free sheaf on gr. The geometric fiber of the sheaf ggg at any point x ∈ gr equals gx, the centralizer of x. From now on, we let L be a finite dimensional rational G-representation such that the set of weights of L is contained in the root lattice. In this case, the results of Kostant [Ko] insure that Lggg is a locally free sheaf on gr. gx ⊂ L20 egr so, for any Borel subalgebra b and any x ∈ egr , with the corre- egr are G-equivariant sub vector Lgx = L0 point (b, x) is contained in the corresponding fiber of the vector bundle L20 Let µ∗(Lggg) be the pullback of Lggg via the map µ :egr → gr. Thus, both µ∗(Lggg) and L20 G-equivariant locally free subsheaves of the trivial sheaf L ⊗ Oegr . Lemma 5.4.1. The sheaf µ∗(Lggg) is a subsheaf of L20 gr ∩ b, we have an inclusion Lgx ⊂ L20 b . Proof. We may (and will) identify the locally free sheaf µ∗(Lggg), resp. L20 sponding vector bundle onegr. Thus, both µ∗(Lggg) and L20 bundles of the trivial vector bundle onegr with fiber L. Let (b, x) ∈egrs. Then, gx is a Cartan subalgebra contained in b and, by definition, we have b . Hence, for any (b, x) ∈ egrs, the fiber of the vector bundle µ∗(Lggg) at the of the lemma follows from this by continuity since the setegrs is dense inegr. cally free G-equivariant sheaves onegr: µ∗(Lggg) −→ L20 egr = Lh ⊗ Oegr . We are now going to transport various sheaves onegr to r via the isomorphism π :egr ∼→ r of diagram (5.3.1). This way, one obtains a filtration L2 q r of the trivial sheaf L ⊗ Or and an associated graded sheaf Lr = ⊕λ∈X Lλ r . Further, we factor the morphism µ in (5.3.1) as a composition µ = pr ◦ π. Thus, we have µ∗(Lggg) = π∗pr∗(Lggg), where we put Lggg r := pr∗(Lggg). Transporting the composite morphism in (5.4.2) via the isomorphism π yields a morphism Thanks to the above lemma, one has the following chain of canonical morphisms of lo- egr . The statement egr = L0 egr ։ L20 egr /L≺0 egr are (cid:3) (5.4.2) r = Lh ⊗ Or. pr∗(Lggg) −→ L0 (5.4.3) Applying adjunction to the above morphism, one obtains a morphism λL : Lggg → Lh⊗pr∗Or, of locally free G-equivariant sheaves on gr. An important special case of the above setting is the case where L = g, the adjoint repre- sentation. The weights of the adjoint representation are clearly contained in the root lattice. Furthermore, for L = g one has Lgx = gx, hence, we have Lggg = ggg. Since g20 b = b, we see that Lemma 5.4.1 reduces in the case L = g to a well known result saying that, for any x ∈ b ∩ gr, one has an inclusion gx ⊂ b. Further, the composite morphism in (5.4.2) becomes a morphism pr∗ggg → h ⊗ Or. We take exterior powers of that morphism. This gives, for each k ≥ 0, an induced morphism pr∗(∧kggg) → ∧kh⊗Or. Finally, applying adjunction one obtains a morphism λk : ∧kggg → ∧kh ⊗ pr∗Or. 30 Associated with any W -module E, there is a G-equivariant coherent sheaf (E ⊗ pr∗Or)W on gr. In particular, one may let E = Lh, the universal zero weight space of a G-module L Lemma 5.4.4. (i) The sheaf (E ⊗ pr∗Or)W is locally free for any W -module E. W -invariants, so the morphism in question factors through an injective morphism (ii) The image of the morphism λL, resp. λk, k ≥ 0, is contained in the corresponding sheaf of λL : Lggg → (Lh ⊗ pr∗Or)W , resp. λk : ∧kggg → (∧kh ⊗ pr∗Or)W , k ≥ 0. (iii) On the open dense set grs ⊂ gr, each of the morphisms in (5.4.5) is an isomorphism. (5.4.5) Sketch of Proof. The statement of part (i) is well known but we recall the proof, for complete- ness. First, we apply base change for the cartesian square in diagram (5.3.1). This yields a chain of isomorphisms pr∗Or = pr∗eγ∗Ot = γ∗ϑ∗Ot. Here, the sheaf on the right is locally free and, moreover, the Weyl group W acts on the geometric fibers of that sheaf via the regular representation. We deduce that the sheaf pr∗Or has similar properties. Part (i) follows from this. The proof of part (iii) is straightforward. Finally, part (ii) follows from part (iii) 'by conti- (cid:3) nuity', using that the sheaves Lggg and ∧kggg are locally free. The theorem below, which is the main result of this section, is inspired by an idea due to Beilinson and Kazhdan. In [BK], the authors considered the isomorphism of part (i) of Theorem 5.4.6 in the special case k = 1 (equivalently, of part (ii) in the case L = g). Theorem 5.4.6. phism. (i) For any k ≥ 1, the morphism λk : ∧kggg ∼→ (∧kh ⊗ pr∗Or)W is an isomor- (ii) For any small representation L, the morphism λL : Lggg ∼→ (Lh ⊗ pr∗Or)W is an isomor- phism. The proof of Theorem 5.4.6 will be completed in §5.6. In §5.7 we sketch a generalization of part (ii) of the above theorem to the case of an arbi- trary, not necessarily small, representation L. 5.5. Proof of Theorem 5.4.6(ii). Let CohG(gr) be the category of G-equivariant coherent sheaves on gr. We begin with an easy Lemma 5.5.1. The functor Γ : CohG(gr) ∼→ C[t]W -mod, F 7→ Γ(gr, F)G is an equivalence. Proof. Recall the adjoint quotient morphism γ, cf. (5.3.1). We have a diagram of functors CohG(gr) F 7→ (γ∗F )G γ∗ / Coh(t/W ) Γ(t/W,−) / C[t]W -mod. It is known, thanks to results of Kostant [Ko], that γ is a smooth morphism, moreover, each fiber of that morphism is a single G-orbit. It follows by equivariant descent that the functor F 7→ (γ∗F)G is an equivalence, with γ∗ being its quasi-inverse. The functor Γ(t/W,−), in the above diagram, is an equivalence since the variety t/W is affine. Finally, the functor Γ is isomorphic to the composite functor F 7→ Γ(t/W, (γ∗F)G) = Γ(gr, F)G. It follows that Γ is also an equivalence. (cid:3) It is useful to transport the morphisms in (5.4.5) via the equivalence Γ of Lemma 5.5.1. To this end, for a G-module L, we compute Γ(Lggg) = Γ(gr, Lggg)G = Γ(g, Lgggg) = (L ⊗ C[g])G, 31 (5.5.2) / o o / where the second equality holds by part (ii) of Lemma 5.1.2 and the third equality holds by part (i) of the same lemma. Recall next that x is known to be a normal variety and the set x r r has codimension ≥ 2 in x. This yields natural isomorphisms C[g] ⊗C[t]W C[t] = C[x] ∼→ C[r]. In particular, we deduce C[r]G = (C[g] ⊗C[t]W C[t])G = C[t]. Thus, for any W -representation E, we find Γ((E ⊗ pr∗Or)W ) = Γ(gr, (E ⊗ pr∗Or)W )G = (E ⊗ Γ(gr, pr∗Or))W ×G (5.5.3) = (E ⊗ Γ(r, Or))W ×G = (E ⊗ C[r]G)W = (E ⊗ C[t])W . Now, fix t ⊂ b ⊂ g and write i : ֒→ g for the resulting imbedding of the Cartan subalgebra. This gives, for any G-representation L, an identification Lt ∼→ Lh and also a natural restriction map i∗ : (L ⊗ C[g])G → (Lt ⊗ C[t])W . Lemma 5.5.4. The following diagram commutes t Γ(Lggg) Γ(λL) (5.5.2) Γ((Lh ⊗ pr∗Or)W ) (5.5.3) (Lh ⊗ C[t])W (L ⊗ C[g])G i∗ (Lt ⊗ C[t])W (5.5.5) Proof. To prove commutativity of the diagram it suffices to check this after localization to the open dense set grs ⊂ gr of regular semisimple elements. The latter is straightforward and is left for the reader. (cid:3) By commutativity of diagram (5.5.5), using the equivalence of Lemma 5.5.1 we conclude that the morphism λL, in (5.4.5), is an isomorphism if and only if so is the map i∗ on the right of diagram (5.5.5). At this point, the proof of Theorem 5.4.6(ii) is completed by the following result of B. Broer [Br]. Proposition 5.5.6. The restriction map i∗ : (L ⊗ C[g])G → (Lt ⊗ C[t])W is an isomorphism, for any small representation L. (cid:3) g , resp. ∧kt ⊗ Ot ∼= Ωk t . 5.6. Proof of Theorem 5.4.6(i). Write Ωk X for the sheaf of k-forms on a smooth variety X. Thus, using the identification g∗ ∼= g, resp. t∗ ∼= t, for each k ≥ 0, we have an isomorphism ∧kg ⊗ Og ∼= Ωk Let f1, . . . , fr be a set of homogeneous generators of C[g]G, a free polynomial algebra. Thus, one may think of the 1-forms dfi, i = 1, . . . , r, as being sections of the sheaf g ⊗ Og ∼= Ω1 g. A result of Kostant says that the values (df1)x, . . . , (dfr)x, of these sections at any point x ∈ gr, give a basis of the vector space gx = gggx. Hence, the r-tuple (df1, . . . , dfr) provides a basis of sections of the sheaf ggg. Next, let ¯fi be the image of fi under the Chevalley isomorphism C[g]G ∼→ C[t/W ]. The map h 7→ ( ¯f1(h), . . . , ¯fr(h)) yields an isomorphism t/W ∼→ Cr. Hence, the r-tuple d ¯f1, . . . , d ¯fr, of 1-forms on t/W , provides a basis of sections of the sheaf Ω1 t/W . Therefore, the r-tuple γ∗(d ¯f1), . . . , γ∗(d ¯fr) provides a basis of sections of the sheaf γ∗Ω1 t/W , on gr. γ∗Ω1 γ∗Ωk Thus, the assignment γ∗(d ¯fi) 7→ dfi, i = 1, . . . , r, gives an isomorphism of sheaves ∼→ ggg. Taking exterior powers, one obtains, for each k ≥ 0, an isomorphism Ψ : ∼→ ∧k ggg, of locally free sheaves on gr. On the other hand, we recall that the square on the right of diagram (5.3.1) is cartesian and the morphism ϑ in the diagram is finite and flat. Hence, by flat base change, for any t/W t/W 32     k ≥ 0, we get W -equivariant isomorphisms pr∗Or = pr∗eγ∗Ot = γ∗ϑ∗Ot. Tensoring here each term by the W -module ∧kt and taking W -invariants in the resulting sheaves, yields a natural sheaf isomorphism Φ : (∧kt ⊗ pr∗Or)W ∼→ γ∗[(ϑ∗(∧kt ⊗ Ot))W ] = γ∗(ϑ∗Ωk Combining all the above morphisms together, we obtain the following diagram t )W . γ∗Ωk t/W Ψ ∼ / ∧kggg λk (5.4.5) / (∧kt ⊗ pr∗Or)W Φ ∼ / γ∗((ϑ∗Ωk t )W ) (5.6.1) Lemma 5.6.2. The composite morphism in (5.6.1) equals the pull back via the map γ of the canonical morphism of sheaves Ωk t/W → (ϑ∗Ωk t )W . Proof. The statement amounts to showing that, for each i, the composite morphism (5.6.1) sends the section dfi to the section γ∗(d(ϑ∗ ¯fi)), where d(ϑ∗ ¯fi) ∈ ϑ∗Ω1 t is viewed as a 1-form on t. Furthermore, it suffices to check this on the open dense set of regular semisimple (cid:3) elements where it is clear. At this point, we apply a result due to L. Solomon [So] saying that the canonical morphism t )W is in fact an isomorphism. It follows, thanks to Lemma 5.6.2, that the Ωk t/W → (ϑ∗Ωk morphism λk in the middle of diagram (5.6.1) must be an isomorphism. This completes the (cid:3) proof of Theorem 5.4.6(i). Remark 5.6.3. There is an alternative approach to the proof that the morphisms in (5.4.5) are isomorphisms as follows. First of all, on the open set of regular semisimple elements our claim amounts to Lemma 5.4.4(ii). Next, one verifies the result in the case g = sl2. It follows that the result holds for any reductive Lie algebra of semisimple rank 1. Using this, one deduces that the morphisms of locally free sheaves in (5.4.5) are isomorphisms outside a codimension 2 subset. The result follows. A similar strategy can also be used to obtain direct proofs of the above mentioned results of Broer and Solomon, respectively. 5.7. The case of a not necessarily small representation. The sheaf morphism λL : Lggg → (Lh ⊗ pr∗Or)W , in (5.4.5), may fail to be surjective in the case where L is a not necessar- ily small representation. Nonetheless, using a result by Khoroshkin, Nazarov, and Vinberg [KNV], we will give a description of the image of that morphism for an arbitrary represen- tation L. To this end, we first apply the construction of sections 5.3 and §5.4 in the special case of the adjoint representation L = g. Thus, starting from the sheaf gB on B, see (5.2.1), the sheaf on r that comes equipped with a fiberwise Lie bracket. construction produces a sheaf gr := h ⊗ Or L (cid:0) ⊕α∈R r(cid:1). This is a G-equivariant locally free For any representation L, there is a natural action morphism gr⊗ Lr → Lr. Taking adjoints r )∗ yields, for any α ∈ R and λ ∈ X, an induced . Iterating the latter morphism k times we obtain and using the isomorphism g−α coaction morphism eα : Lλ morphisms gα r → Lλ+α α : Lλ ek r ∼= (gα r ⊗ g−α r → Lλ+k·α r r ⊗ g−k·α r , k = 1, 2, . . . . Let ker α ⊂ t be the root hyperplane corresponding to a root α ∈ R+. The inverse image of this hyperplane via the mapeγ : r → t, cf. (5.3.1), is a smooth G-stable irreducible divisor Dα ⊂ r. Given k ≥ 1, we use the standard notation Lk·α (−k · Dα) for a subsheaf of r ⊗ g−k·α Lk·α Our description of the image of the morphism λL is provided by the following result inspired by [KNV]. formed by the sections which have a k-th order zero at the divisor Dα. r ⊗ g−k·α r r 33 / / / Theorem 5.7.1. Let L be a finite dimensional g-representation such that the weights of L are con- tained in the root lattice. Then the morphism λL : Lggg → (Lh ⊗ pr∗Or)W , in (5.4.5), yields an isomorphism of Lggg with a subsheaf L of the sheaf (Lh ⊗ pr∗Or)W defined as follows α(pr∗(s)) ∈ Lk·α r ⊗ g−k·α r (−k · Dα), ∀α ∈ R+, k ≥ 1}. L := {s ∈ (Lh ⊗ pr∗Or)W (cid:12)(cid:12) ek Remark 5.7.2. It is not difficult to check that, in the case where L is a small representation, one has L = (Lh ⊗ pr∗Or)W . So, Theorem 5.7.1 reduces in this case to Theorem 5.4.6(ii). ♦ The proof of Theorem 5.7.1 is parallel to the arguments used in §5.5. There is an analogue of commutative diagram (5.5.5). The space (Lh ⊗ C[t])W in that diagram is replaced by its subspace defined in [KNV, p.1169], resp. the space Γ((Lh ⊗ pr∗Or)W ) is replaced by Γ(L). The role of Broer's result from Proposition 5.5.6 is played by [KNV, Theorem 2]. Theorem 5.7.1 will not be used in the rest of the paper; so, details of the proof will be given elsewhere. 6. GEOMETRY OF THE COMMUTING SCHEME 6.1. Another isospectral variety. We write x = hx + nx for the Jordan decomposition of an element x ∈ g. We say that a pair (x, y) ∈ G is semisimple if both x and y are semisimple elements of g. Let Gb be the set of pairs (x, y) ∈ G such that there exists a Borel subalgebra that contains both x and y. (i) Let (x, y) ∈ G. Then (x, y) ∈ Gb if and only if there exists a semisimple pair (ii) If (x, y) ∈ C then the G-diagonal orbit of the pair (hx, hy) is the unique closed G-orbit Lemma 6.1.1. (h1, h2) ∈ T such that, for any polynomial f ∈ C[G]G, we have f (x, y) = f (h1, h2). contained in the closure of the G-diagonal orbit of (x, y). Proof. Let t ⊂ b be Cartan and Borel subalgebras of g and let T be the maximal torus cor- responding to t. Clearly, there exists a suitable one parameter subgroup γ : C× → T such that, for any t ∈ t and n ∈ [b, b], one has lim Ad γ(z)(t + n) = t. To prove (i), let (x, y) ∈ b × b. We can write x = h1 + n1, y = h2 + n2 where hi ∈ t and ni ∈ [b, b]. We see from the above that the pair (h1, h2) is contained in the closure of the G-orbit of the pair (x, y). Hence, for any f ∈ C[G]G, we have f (x, y) = f (h1, h2). Conversely, let (h1, h2) ∈ T and let (x, y) ∈ G be such that f (x, y) = f (h1, h2) holds for any f ∈ C[G]G. The group Gh1,h2 is reductive. Hence, the G-diagonal orbit of (h1, h2) is closed in G, cf. eg. [GOV]. Moreover, this G-orbit is the unique closed G-orbit contained in the closure of the G-orbit of the pair (x, y) since G-invariant polynomials on G separate closed G-orbits. By the Hilbert-Mumford criterion, we deduce that there exists a suitable one parameter subgroup γ : C× → G such that, conjugating the pair (h1, h2) if necessary, on gets lim z→0 Ad γ(z)(x, y) = (h1, h2). z→0 has lim z→0 Now, let a be the Lie subalgebra of g generated by the elements x and y and let t be a Ad γ(z)(x, y) ∈ t × t. We deduce from the above that one Cartan subalgebra such that lim z→0 Ad γ(z)(cid:0)[a, a](cid:1) ⊂ [t, t] = 0. This implies that any element of [a, a] is nilpotent. Hence, [a, a] is a nilpotent Lie algebra, by Engel's theorem. We conclude that a is a solvable Lie algebra. Hence, there exists a Borel subalgebra b such that a ⊂ b and (i) is proved. To prove (ii), observe that the elements hx, hy, nx, ny generate an abelian Lie subalgebra of g. Hence, there exists a Borel subalgebra b that contains all of them. Choose a Cartan subalgebra t ⊂ b such that hx, hy ∈ t. Then, the argument at the beginning of the proof shows that the pair (hx, hy) is contained in the closure of the G-diagonal orbit of (x, y). (cid:3) 34 Note that, by definition, we have Gb = [µ(eG)]red, the reduced image of the morphism µ : eG → G, see §3.1. The following result implies Proposition 3.1.1(ii). Proposition 6.1.2. The morphism µ × ν : eG → [G ×G//G T]red is birational and proper. Proof. Let G♥ be the set of pairs (x, y) ∈ G such that the vector space Cx + Cy ⊂ g spanned by x and y is 2-dimensional and, moreover, any nonzero element of that vector space is regular in g. It is clear that G♥ is a G-stable Zariski open and dense subset of G. The first projection G × T → G induces a birational finite morphism [G ×G//G T]red → Gb. According to [CM], Lemma 6(i), the set (b× b)∩ G♥ is nonempty, for any Borel subalgebra b. Hence, this set is Zariski open and dense in b × b. Furthermore, it follows from [CM], Lemma 7(i) and Lemma 8(i), that for any pair (x, y) ∈ G♥ there is at most one Borel subal- gebra that contains both x and y (in [CM], this result is attributed to Bolsinov [Bol]). Hence, the map µ restricts to a bijection µ−1(G♥) ∼→ Gb ∩ G♥. Both statements of the proposition follow from this since the map µ is proper. (cid:3) 6.2. A stratification of the isospectral commuting variety. Below, we will have to consider several reductive Lie algebras at the same time. To avoid confusion, we write C(l) for the commuting scheme of a reductive Lie algebra l and use similar notation for other objects associated with l. Let N(l) be the nilpotent commuting variety of l, the variety of pairs of commuting nilpotent elements of l, equiped with reduced scheme structure. It is clear that we have N(l) = N(l′), where l′ := [l, l], the derived Lie algebra of l. According to [Pr], the irreducible components of N(l) are parametrized by the conjugacy classes of distinguished nilpotent elements of l′. The irreducible component corresponding to such a conjugacy class is equal to the closure in l′ × l′ = T ∗(l′) of the total space of the conormal bundle on that conjugacy class. It follows, in particular, that the dimension of each irreducible component of the variety N(l) equals dim l′. Fix a reductive connected group G and a Cartan subalgebra t ⊂ g = Lie G. Recall that the centralizer of an element of t is called a standard Levi subalgebra of g. Let S be the set Tl ⊂ T, of standard Levi subalgebras. Given a standard Levi subalgebra l, let denote the set of elements h ∈ t, resp. (h1, h2) ∈ T, such that we have gh = l, resp. gh1,h2 = l. tl is an irreducible Zariski open dense subset Let tl denote the center of l. It is clear that Tl is an irreducible Zariski open dense subset of tl × tl. We get a stratification of tl, resp. t = ⊔l∈S Let X = X(g), the isospectral commuting variety of g, and let pT : X(g) → T be the projection. For each standard Levi subalgebra l of g, we put Xl(g) := (pT)−1( Tl). Thus, we get a partition X(g) = ⊔l∈S Xl(g) by G-stable locally closed, not necessarily smooth, subvarieties. tl, resp. T = ⊔l∈S tl ⊂ t, resp. Tl. ◦ ◦ ◦ ◦ ◦ ◦ ◦ We consider a map G × G × T → G × T given by the following assignment (6.2.1) Lemma 6.2.2. For any standard Levi subalgebra l with Levi subgroup of L ⊂ G, the map (6.2.1) induces a G-equivariant isomorphism g × (y1, y2) × (t1, t2) 7→ (cid:0) Ad g(y1 + t1), Ad g(y2 + t2)(cid:1) × (t1, t2). The second projection(cid:0)G×L N(l)(cid:1) × Tl goes, under the isomorphism, to the map pT : Xl → Tl. Thus, all irreducible components of the set Xl have the same dimension equal to dim g + dim tl Tl → ◦ ◦ (cid:0)G ×L N(l)(cid:1) × ◦ ◦ Tl ∼→ Xl. 35 and are in one-to-one correspondence with the distinguished nilpotent conjugacy classes in the Lie algebra l. Proof. Let (x1, x2, t1, t2) ∈ Xl. Lemma 6.1.1(ii) implies that, for any polynomial f ∈ C[C]G, we have f (x1, x2) = f (hx1, hx2). We know that the semisimple pair (hx1, hx2) is G-conjugate to an element of T, that W -invariant polynomials separate W -orbits in T, and that the re- striction map (1.3.1) is surjective, [Jo]. It follows that the pair (hx1, hx2 ) is G-conjugate to the pair (t1, t2). Hence, the pair (x1, x2) is G-conjugate to a pair of the form (t1 + y1, t2 + y2) for some (y1, y2) ∈ N(l). The isomorphism of the lemma easily follows from this. above, we find Now, for any irreducible component V of N(l), using the result of Premet mentioned dim(cid:0)(G ×L V ) × ◦ Tl(cid:1) = dim G − dim L + dim V + 2 dim tl = dim g − dim(tl + [l, l]) + dim[l, l] + 2 dim tl = dim g + dim tl. This proves the dimension formula for irreducible components of the variety Xl. (cid:3) We are now ready to complete the proof of Lemma 2.1.3. Corollary 6.2.3. The set Xrs is an irreducible and Zariski dense subset of X. Proof. Let Css be the set of semisimple pairs (x, y) ∈ C and let Xss 6.2.2 implies that, for any Levi subalgebra l, any element (x, y, t1, t2) ∈ Xss element (x′, y′, t1, t2) for some (x′, y′) ∈ Hence, Lemma 6.2.2 implies that Xrs is irreducible and, moreover, we have Xss any standard Levi subalgebra l. := Xl ∩ p−1(Css). Lemma is conjugate to the Tl. In the special case where l = t, we have Xrs = Xt. l ⊂ Xrs, for Thus, using the isomorphism of Lemma 6.2.2 one more time, we see that proving the Corollary reduces to showing that N(l), the nilpotent commuting variety of l, is contained in the closure of the set Css(l). But we have Css(l) ⊇ Crs(l) and the set Crs(l) is dense in C(l), by Proposition 2.1.1(i); explicitly, this is Corollary 4.7 from [Ri1]. The result follows. (cid:3) ◦ l l 6.3. Proof of Lemma 2.6.4. We have the projection p : X(g) → C(g) and, for any standard Levi subalgebra l ⊂ g, put Cl(g) = p(Xl(g)) where we use the notation of Lemma 6.2.2. Thus, one has C(g) = ∪l∈S Cl(g). Note that, for a pair of standard Levi subalgebras l1, l2 ⊂ g, the corresponding pieces Cl1(g) and Cl2(g) are equal whenever the Levi subalgebras l1 and l2 are conjugate in g; otherwise these two pieces are disjoint. Let l be a standard Levi subalgebra in g. We are interested in the dimension of the set Cl(g) r Crr. The dimension formula of Lemma 6.2.2 shows that the codimension of the set Cl(g) in C(g) equals dim(t/tl). We see that to prove Lemma 2.6.4 it suffices to show that, for any standard Levi subalgebra l ⊂ g such that the codimension of tl in t equals either 0 or 1, the set Cl(g) r Crr has codimension ≥ 2 in C(g). ◦ In the first case we have t = tl = l. Thus, one has N(l) = {0}, so N(l) + ◦ we have used simplified notation T such that neither h1 nor h2 is regular. Therefore, each of these two elements belongs to some root hyperplane in t, that is, belongs to a finite union of codimension 1 subspaces in t. We T r Crr consists of the pairs (h1, h2) ∈ Tt. The set T, where Tl = ◦ ◦ ◦ ◦ T := conclude that the set ◦ Tl, as required. ◦ T r Crr has codimension ≥ 2 in N(l) + Next, let dim(t/tl) = 1. In that case, l is a minimal Levi subalgebra of g. Thus, there is a root α ∈ t∗ in the root system of (g, t) such that tl = Ker α is a codimension 1 hyperplane in t. We have l = tl ⊕ [l, l] where [l, l] is an sl2-subalgebra of g associated with the root α. It is easy to see that N(sl2) is an irreducible variety formed by the pairs of nilpotent elements 36 proportional to each other (the zero element is declared to be proportional to any element). Thus, N(l) + Tl is an irreducible variety. ◦ ◦ Tl) ∩ Crr has codimension ≥ 1 in N(l) + To complete the proof of the lemma in this case, we must show that the complement of Tl. The set U is a Zariski open the set U = (N(l) + subset in an irreducible variety. Thus, it suffices to show that the set U is nonempty. For this, pick h ∈ tl and let n ∈ sl2 be any nonzero nilpotent element. Then, n is a regular element of the Lie algebra [l, l] = sl2; hence h + n is a regular element of g. Therefore, we have (h + n, h + n) ∈ U , and we are done. (cid:3) ◦ ◦ 6.4. Proof of Theorem 1.5.2(i). The property of a coherent sheaf be Cohen-Macaulay is sta- ble under taking direct images by finite morphisms and taking direct summands. Thus, Theorem 1.3.4 implies that the sheaf R, as well as the isotypic components RE for any W - module E, is Cohen-Macaulay. The scheme Xnorm/W is reduced and integrally closed, as a quotient of an integrally closed reduced scheme by a finite group action, [Kr], §3.3. Further, by Lemma 2.1.3(ii), the map pnorm is generically a Galois covering with W being the Galois group. It follows that the induced map Xnorm/W → Cnorm is a finite and birational morphism of normal varieties. Hence it is an isomorphism, that is, the canonical morphism OCnorm → ((pnorm)∗OXnorm)W = RW is an isomorphism. This yields an isomorphism Xnorm/W ∼= Cnorm and implies Corol- lary 1.3.5. Observe next that the sheaf RCr is a Cohen-Macaulay sheaf on a smooth variety, hence it is locally free. The fiber of the corresponding algebraic vector bundle over any point of the open set Crs ⊂ Cr affords the regular representation of the Weyl group W , by Lemma 2.1.3(ii). The statement of Theorem 1.5.2(i) follows from this by continuity since the set Crs is dense in Cr. 6.5. Proof of Theorem 1.5.2(ii) and of Theorem 1.6.1(i). We use the notation of Definition 2.6.3 and let eX1 := µ−1(C1). Write eq : eX1 → egr, (b, x, y) 7→ (b, x), resp. q1 : C1 → gr, (x, y) 7→ x, for the natural projection, and let π be the map from Corollary 3.3.5. Thus, one has a commutative diagram π ∼ π ∼ / X1 q / r p pr / C1 q1 / gr (6.5.1) eq eX1 egr Lemma 6.5.2. The right square in diagram (6.5.1) is cartesian. from the diagram is a set theoretic bijection. Further, all the varieties involved in diagram (6.5.1) are smooth and the map q1 is a smooth morphism (the vector bundle projection Nr → Proof. It is immediate from Lemma 5.4.1 that the mapeq×(p ◦ π) : eX1 → egr ×gr C1 that results r, cf. Lemma 2.1.5). It follows thategr ×gr C1 is a smooth variety and, moreover, the bijection above gives an isomorphism eX1 ∼→ egr ×gr C1, of algebraic varieties. We deduce that the To complete the proof, we observe that the map π in (6.5.1) is an isomorphism by Corol- lary 3.3.5, resp. the map π is an isomorphism by Proposition 3.1.1(i). We conclude that the (cid:3) right square in diagram (6.5.1) is cartesian as well. rectangle along the perimeter of diagram (6.5.1) is a cartesian square. 37   /   /   / / The morphism pr in diagram (6.5.1) is finite and the morphism q1 is smooth. So, thanks to Lemma 6.5.2, we may apply smooth base change for the cartesian square on the right of (6.5.1). Combining this with smooth base change for the cartesian square on the right of diagram (5.3.1), yields a chain of natural G× W × C× × C×-equivariant sheaf isomorphisms (6.5.3) RC1 = p∗OX1 = p∗q∗Or = q∗ 1pr∗Or = q∗ Therefore, for any W -representation E, we deduce 1(cid:0)(E ⊗ pr∗Or)W(cid:1) = q∗ REC1 = (E ⊗ q∗ 1pr∗Or)W = q∗ (6.5.4) Observe next that, for any (x, y) ∈ C1, the Lie algebra gx is abelian, hence we have gx,y = gx∩ gy = gx. From this, writing ggg1 := gggC1 for short, we get a natural isomorphism ggg1 = q∗ 1ggggr , of sheaves on C1. We deduce that the sheaf ggg1 is a locally free. Furthermore, applying to functor q∗ 1(−) to the isomorphism of Theorem 5.4.6(i) and using (6.5.4) one obtains the following isomorphisms 1γ∗(ϑ∗Ot). 1γ∗(cid:0)(E ⊗ ϑ∗Ot)W(cid:1). 1pr∗(eγ∗Ot) = q∗ ∧k ggg1 = q∗ 1(∧kggggr ) ∼→ q∗ 1(cid:0)(∧kt ⊗ pr∗Or)W(cid:1) = R∧ktC1. (6.5.5) Now, let L be a rational G-module such that the set of weights of L is contained in the root 1(Lggggr ). Therefore 1(−) to the isomorphism of lattice. Then, a similar argument yields a natural isomorphism Lggg1 = q∗ the sheaf Lggg1 is locally free. Moreover, applying the functor q∗ Theorem 5.4.6(ii) and using (6.5.4) again, one similarly obtains an isomorphism Lggg1 ∼→ RLh C1. (6.5.6) Similar considerations apply, of course, in the case where the set C1 is replaced by the set C2. It follows, in particular, that gggrr := gggCrr and Lgggrr := LgggCrr are locally free coherent sheaves on Crr = C1 ∪ C2. However, it is not clear a priori, that the 'C2-counterparts' of morphisms (6.5.5)-(6.5.6) agree with those in (6.5.5)-(6.5.6) on the overlap C1 ∩ C2. To overcome this difficulty, we now produce an independent direct construction of canon- ical morphisms Crr . resp. λL rr : Lgggrr → RLh rr : ∧kgggCrr → R∧ktCrr , λk being a pull-back of a locally free coherent sheaf on Crr. (6.5.7) This will be done by adapting the strategy of §5.3 as follows. Let Xrr = p−1(Crr) and eXrr = µ−1(Crr). We know by the above that the sheaf µ∗gggrr, resp. µ∗(Lgggrr ), is locally free, Now let (x, y, b) ∈ eXrr. Then, x, y ∈ b and, moreover, we have that either gx,y = gx or gx,y = gy. In each of the two cases, applying Lemma 5.4.1, we deduce an inclusion Lgx,y ⊂ L20. Therefore, one gets, as in §5.3, a well defined morphism µ∗(Lgggrr ) → Lh ⊗ OeXrr . We may further transport this morphism via π, the isomorphism of Corollary 3.3.5(ii). This way, one constructs a canonical morphism f : p∗(Lgggrr ) → Lh ⊗ OXrr . The morphisms in (6.5.7) are now defined from the morphism f , by adjunction, mimicing the construction of §5.3. Lemma 6.5.8. The restriction of the morphism λk reduces to isomorphism (6.5.5), resp. (6.5.6). rr, in (6.5.7), to the open set C1 ⊂ Crr rr, resp. λL Similar claim holds in the case of the set C2 ⊂ Crr. Proof. All the sheaves involved are G-equivariant and locally free. Hence, it suffices to check the statement of the lemma fiberwise, and only at the points of the form (x, y) ∈ tr × tr. In that case verification is straightforward and is left for the reader. (cid:3) 38 Lemma 6.5.8 implies that each of the morphisms in (6.5.7) is an isomorphism of locally free sheaves on Crr. To complete the proof of Theorem 1.6.1(i) write j : Crr ֒→ Cnorm be the open imbedding. We know that the sheaf RE is Cohen-Macaulay for any W -representation E. It follows, that the canonical morphism RE → j∗(RECr ) is an isomorphism, cf. Lemma 2.6.1. Similarly, by Lemma 5.1.2(ii), we have a canonical isomorphism LgggCnorm ∼→ j∗(LgggCr ). Thus, applying the functor j∗(−) to the isomorphism LgggCr ∼= RLhCr proved earlier, we deduce LgggCnorm = j∗(LgggCr ) ∼= j∗(RLhCr ) ∼= RLh and Theorem 1.6.1(i) follows. A similar (even simplier) argument proves Theorem 1.5.2(ii). Remark 6.5.9. In the case of a not necessarily small representation L, one can still obtain a description of the image of the corresponding morphism λL rr, in (6.5.7), in the spirit of Theorem 5.7.1. That description is not very useful, however, since the sheaf LgggCCr turns out to be not locally free, in general, already for g = sl2. 6.6. Proof of Corollary 1.5.1. The isomorphism OCnorm ∼= RW , in Corollary 1.5.1(i), has been already established at the beginning of §6.4. Proof of the isomorphism KCr ∼= R signCr . Given a point (x, y) ∈ C, let Tx,yC, resp. T ∗ x,yC, be the Zariski tangent, resp. cotangent, space to C at (x, y). Let κ∗ : g ⊕ g → g be the differential of the commutator map κ at the point (x, y). Then, one has an exact sequence of vector spaces 0 / Tx,yC / g ⊕ g κ∗ / g / Coker(κ∗) / 0. We use an invariant form on g to identify g∗ with g and write κ⊤ ∗ for the linear map dual to the map κ∗. Then, dualizing the exact sequence above yields an exact sequence 0 T ∗ x,yC κ⊤ ∗ g ⊕ g g Ker(κ⊤ ∗ ) 0 (6.6.1) Now, the map κ∗ is given by the formula κ∗ : (u, v) 7→ [x, u] − [y, v]. Using the invariance of the bilinear form one easily finds that the dual map is given by the formula κ⊤ ∗ : a 7→ [x, a] ⊕ [y, a]. We conclude that Ker(κ⊤ Write det for the top exterior power of a vector space. From (6.6.1), we deduce a canonical isomorphism det T ∗ x,yC = det(g⊕ g)⊗ (det g)−1⊗ det gx,y. For (x, y) ∈ Cr, we have det gx,y = ∧rgx,y = ∧rgggC(x,y). Therefore, a choice of base vector in the 1-dimensional vector space x,yC ∼= ∧rgggC(x,y). This yields an det g determines, for all (x, y) ∈ Cr, an isomorphism det T ∗ isomorphism KCr ∼= ∧rgggCr of locally free sheaves. ∗ ) = gx,y. (cid:3) We can now complete the proof of Corollary 1.5.1. We know that Cnorm is a Cohen- Macaulay variety and that the sheaf R on Cnorm is isomorphic to its Grothendieck dual RHom OCnorm (R, KCnorm). It follows that for any j 6= 0 one has Rj Hom OCnorm (R, KCnorm) = 0 and, moreover, there is an isomorphism R ∼= Hom OCnorm (R, KCnorm). Further, since OCnorm is a direct summand of R the sheaf KCnorm is a direct summand of Hom OCnorm (R, KCnorm) ∼= R. Hence, KCnorm is a Cohen-Macaulay sheaf. Similarly, the sheaf R sign is also Cohen-Macaulay. Recall that two Cohen-Macaulay sheaves are isomorphic if and only if they have iso- morphic restrictions to a complement of a closed subset of codimension ≥ 2. The isomor- phism KCnorm ∼= R sign of Corollary 1.5.1(i) now follows from the chain of isomorphisms KCnormCr ∼= ∧rgggCr ∼= (∧rt ⊗ R)WCr = R signCr where the second isomorphism holds by Theorem 1.5.2(ii). 39 / / / / / o o o o o o o o o o Finally, the isomorphism of part (ii) of Corollary 1.5.1 follows, thanks to the isomorphism KCnorm ∼= R sign, by equating the corresponding W -isotypic components on each side of the self-duality isomorphism Hom OCnorm (R, R sign) ∼= R proved earlier. Remark 6.6.2. The short exact sequence (6.6.1) implies that, for any (x, y) ∈ C, one has (cid:3) r + dim g = dim C ≤ dim T ∗ x,yC = 2 dim g − (dim g − dim Ker κ⊤ ∗ ) = dim g + dim gx,y, where in the first equality we have used Proposition 2.1.1(i). We deduce an inequality r ≤ dim gx,y. Moreover, we see that this inequality becomes an equality if and only if (x, y) is a smooth point of the scheme C. This proves Proposition 2.1.1(ii). 6.7. Proof of Theorem 1.6.1(ii)-(iii). We have the following chain of natural W -equivariant algebra maps C[T] = C[T]W ⊗C[T]W C[T] = C[Cred]G ⊗C[Cred]G C[T] Let f be the composition of the above maps. = C[Cred ×Cred//G T]G ։ C[X]G ֒→ C[Xnorm]G. It follows from the isomorphism Crs ∼= G ×N (T ) Tt that the map f induces an isomorphism between the fields of fractions of the algebras C[T] and C[Xnorm]G, respectively. The algebra C[Xnorm]G is, by definition, a finitely generated C[X]G-module, hence, also a finitely generated module over the image of f . ◦ Thus, since the algebra C[T] is integrally closed, we obtain Γ(Cnorm, R)G = C[Xnorm]G = C[T]. Equating W -isotypic components on each side of this isomorphism yields the first isomorphism of Theorem 1.6.1(ii). To prove the second isomorphism, we compute (L ⊗ C[Cnorm])G = Γ(Cnorm, L ⊗ OCnorm )G = Γ(Cnorm, LgggCnorm )G by Lemma 5.1.2(i) = Γ(Cnorm, RLh = (Lh ⊗ C[T])W by the previous paragraph. )G by Theorem 1.6.1(i) It is immediate to check, by restricting to the open set of regular semisimple pairs, that the composition of the chain of isomorphism above goes, via the identification Lh = Lt induced ֒→ g, to the restriction homomorphism i∗ : (L ⊗ C[Cnorm])G → by the imbedding i : t (Lt ⊗ C[T])W . This proves part (ii) of Theorem 1.6.1. To prove part (iii) we use Theorem 1.5.2(ii). From that theorem, we deduce Γ(Cr, R∧st) ∼= Γ(Cr, ∧sgggCr ), for any s ≥ 0. Further, we know that the set Cnorm r Cr has codimension ≥ 2 in Cnorm and that R∧st is a Cohen-Macaulay sheaf on Cnorm. It follows that the natural restriction map induces an isomorphism Γ(Cnorm, R∧st) ∼→ Γ(Cr, R∧st). The proof is now completed by the following chain of isomorphisms: Γ(Cr, ∧sgggCr )G ∼= Γ(Cr, R∧st)G ∼= Γ(Cnorm, R∧st)G ∼= (∧st ⊗ C[T])W . ✷ 7. PRINCIPAL NILPOTENT PAIRS 7.1. Filtrations and Rees modules. Given a vector space E we refer to a direct sum de- composition E = Li,j≥0 Ei,j as a bigrading on E. Similarly, a collection of subspaces 40 Fi,jE ⊂ E, i, j ≥ 0 such that Fi,jE ⊂ Fi′,j ′E whenever i ≤ i′ and j ≤ j′ will be re- ferred to as a bifiltration on E. Canonically associated with a bifiltration Fi,jE, there is a pair of bigraded vector spaces gr E := Mi,j≥0 gri,j E, gri,j E := Fi,jE Fi−1,jE + Fi,j−1E , resp. RE := Mi,j≥0 Fi,jE. (7.1.1) 1τ j We view the polynomial algebra C[τ1, τ2] as bigraded algebra such that deg τ1 = (1, 0) and deg τ2 = (0, 1). Below, we will often make no distinction between RE and the Rees module 2 · Fi,jE ⊂ C[τ1, τ2] ⊗ E, a bigraded C[τ1, τ2]-submodule of a free C[τ1, τ2]-module with generators E. There is a canonical bigraded space isomorphism of E defined asPi,j τ i Given a vector space E equipped with a pair of ascending filtrations ′F qE and ′′F qE, there are Rees modules ′RE :=Pi τ i 2 · ′′FjE ⊂ C[τ2]⊗ E, as well as associated graded spaces ′gr q E, resp. ′′gr q E. One may further define a bifiltration on E by the formula Fi,jE := ′FiE∩ ′′FjE, i, j ≥ 0. One has the corresponding Rees C[τ1, τ2]- module RE. There are canonical isomorphisms gr E ∼= RE/(τ1 · RE + τ2 · RE). 1 · ′FiE ⊂ C[τ1]⊗ E, resp. ′′RE :=Pj τ j (7.1.2) C[τ ±1 2 ]NC[τ2] RE ∼= C[τ ±1 2 ]N ′RE, resp. gr E ∼= ′gr(′′gr E) ∼= ′′gr(′gr E), of graded C[τ1, τ ±1 2 ]-modules, resp. bigraded vector spaces. Specializing the first of the above isomorphisms at the point (τ1, τ2) = (0, 1) and using that ′RE/τ1 · ′RE ∼= ′gr E we deduce a canonical isomorphism of graded vector spaces RE/(cid:0)τ1·RE + (τ2 − 1)·RE(cid:1) ∼= ′gr E. (7.1.3) responding bifiltration. Further, equip the C[τ1, τ2]-module C[τ1, τ2] ⊗ E with a standard Now let E = Li,j Ei,j be a bigraded vector space itself. Associated naturally with the bigrading, there are two filtrations on E defined by ′FmE := L{i,j i≤m} Ei,j and ′′FnE := L{i,j j≤n} Ei,j, respectively. Let Fm,nE := ′FmE ∩ ′′FnE =L{i,j i≤m, j≤n} Ei,j be the cor- tensor product bigrading (C[τ1, τ2] ⊗ E)p,q :=P{0≤i≤p, 0≤j≤q} Lemma 7.1.4. For a bigraded vector space E =Li,j Ei,j, the assignment yields a bigraded C[τ1, τ2]-module isomorphism ℵ : C[τ1, τ2] ⊗ E ∼→ RE. This lemma is clear. Later on, we will use the following simple result ui,j ∈ Ei,j, i, j, m, n ≥ 0, Ci,j[τ1, τ2] ⊗ Ep−i,q−j. 2 ⊗ ui,j 7−→ τ m+i ℵ : τ m · ui,j, τ n+j 2 1 τ n (cid:3) 1 Corollary 7.1.5. Let E be a finite dimensional vector space equipped with a pair of ascending filtra- tions ′F qE and ′′F qE, respectively. Then RE is a finite rank free C[τ1, τ2]-module. Proof. The filtrations ′F qE and ′′F qE form a pair of partial flags in E. Hence, applying the Bruhat lemma for pairs of flags, we deduce that there exists a bigrading E = ⊕m,n Em,n such that the original filtrations are associated, as has been explained before Lemma 7.1.4, (cid:3) with that bigrading. The corollary now follows from the lemma. Remark 7.1.6. For a general bifiltration on a finite dimensional vector space E that does not come from a pair of filtrations one may have dim(gr E) > dim E, so the C[τ1, τ2]-module RE ♦ is not necessarily flat, in general. 41 Next, let I ⊂ E be a subspace of a vector subspace E and put A = E/I. Write V for the image of a vector subspace V ⊂ E under the projection E ։ E/I. A filtration on E induces a quotient filtration on A. Therefore, given a pair of filtrations ′F qE and ′′F qE on E one has the corresponding quotient filtrations ′F qA = (′F qE + I)/I and ′′F qA = (′′F qE + I)/I on A. There are two, potentially different, ways to define a bifiltration on A as follows F min F max i,j A := Fi,jE = (cid:2)(′FiE ∩ ′′FjE) + I(cid:3)/I ∼= (′FiE ∩ ′′FjE)/(′FiE ∩ ′′FjE ∩ I), i,j A := FiE ∩ FjE = (cid:2)(′FiE + I) ∩ (′′FjE + I)(cid:3)/I = ′FiA ∩ ′′FjA. i,j A ⊂ F max Clearly, for any i, j, one has F min i,j A where the inclusion is strict, in general. Therefore, writing RminA, resp. RmaxA, for the Rees module associated with the bifiltration F min i,j A, resp. F max i,j A, one obtains a canonical, not necessarily injective, bigraded C[τ1, τ2]- module homomorphism can : RminA → RmaxA. 7.2. A flat scheme over C2. We have the standard grading C[t] = Li≥0 Ci[t], resp. filtra- tion Fm = C≤m[t] = Li≤m Ci[t], where Ci[t] denotes the space of degree i homogeneous polynomials on t. Similarly, one has a bigrading C[T] = Li,j := Ci[t] ⊗ Cj[t]. Associated with this bigrading, we have the pair of filtrations ′F qC[T] and ′′F qC[T], respectively, and the corresponding bifiltration Fm,nC[T] = ′FmC[T] ∩ ′′FnC[T] = L{i≤m, j≤n} Recall the setting of §1.7. Thus, we have the semisimple pair h = (h1, h2) ∈ T associated with the principal nilpotent pair e = (e1, e2). The coordinate ring of the finite subscheme W ·h ⊂ T has the form C[W ·h] = C[T]/Ih where Ih ⊂ C[T] is an ideal generated by the elements {f − f (h), f ∈ C[T]W}. The quotient algebra C[W ·h] = C[T]/Ih inherits a pair, ′F qC[W ·h] and ′′F qC[W ·h] of quotient filtrations. Associated with these filtrations, there are bifiltrations F max We put Y := Spec Rmax C[W ·h]. This is an affine scheme that comes equipped with a W × C× × C×-action and with a C× × C×-equivariant morphism ℘ : Y → C2 induced by the canonical algebra imbedding C[τ1, τ2] ֒→ Rmax C[W ·h]. Let ϑ : t → t/W be the quotient morphism and write C[ϑ−1(W · h2)] for the coordinate ring of the scheme theoretic fiber of ϑ over the orbit W · h2 viewed as a closed point of t/W . Lemma 7.2.1. (i) The scheme Y is a reduced, flat and finite scheme over C2. Ci,j[T], m, n ≥ 0, cf. §7.1. C[W ·h], resp. F min Ci,j[T] where Ci,j[T] i,j i,j C[W ·h]. There are natural W -equivariant algebra isomorphisms C[℘−1(0, 0)] ∼= grmax C[W ·h] = Mm,n≥0 F max C[℘−1(0, 1)] ∼= ′gr C[W ·h] ∼= C[ϑ−1(W · h2)], C[℘−1(1, 1)] ∼= C[W ·h]. m−1,n F max m,n C[W ·h] C[W ·h] + F max m,n−1 , C[W ·h] (7.2.2) (7.2.3) (7.2.4) (ii) We have F max d1,d2 C[W ·h] = C[W ·h], where ds = #R+ s , s = 1, 2, see §1.7. Proof. The isomorphism in (7.2.2) follows from (7.1.2), resp. the isomorphism in (7.2.4) fol- lows from definitions. The latter isomorphism implies, by C× × C×-equivariance, that we have ℘−1(C× × C×) ∼= C× × C× × W ·h. Further, the map ℘ is flat by Corollary 7.1.5. Therefore, Y is a Cohen-Macaulay scheme such that ℘−1(C× × C×) is a reduced scheme. It follows that the scheme Y is reduced. 42 To complete the proof of the lemma, we consider the Levi subalgebra g1 = gh2 and its Weyl group W1, the subgroup of W generated by reflections with respect to the set R+ 1 ⊂ t∗ formed by the roots α ∈ R+ such that α(h2) = 0. Let I1 ⊂ C[t] denote the ideal of the orbit W1·h1 ⊂ t viewed as a reduced finite subscheme of t. The element h1 has trivial isotropy group under the W1-action, [Gi, Proposition 3.2]. Therefore, a standard argument based on the fact that C[t] is a free C[t]W1-module shows that the ideal grF I1 equals (C[t]W1 + ), the ideal generated by W1-invariant homogeneous polyno- mials of positive degree. We deduce a chain of graded algebra isomorphisms + ) = C[(ϑ1)−1(0)] grF C[W1 · h1] ∼= grF C[t]/ grF I1 = C[t]/(C[t]W1 (7.2.5) where we have used the notation ϑ1 : t → t/W1 for the quotient morphism. Thus, there is an isomorphism Spec grF C[W1 · h1] ∼= (ϑ1)−1(0) of (not necessarily reduced) schemes. Let pr2 : W ·h → W · h2, w(h) 7→ w(h2) be the second projection. We may view the fiber 2 (t), t ∈ W · h2, as a subset of t × {t} ∼= t. Fix m ≥ 0 and let f ∈ ′FmC[W ·h]. Then, by pr−1 definition of the filtration ′F qC[W ·h], for any t ∈ W ·h2, there exists a polynomial eft ∈ C≤m[t] such that we have eftpr−1 2 (t). Conversely, let f ∈ C[W ·h] be an element such that, for any t ∈ W · h2, there exists a polynomial eft ∈ C≤m[t] such that eftpr−1 2 (t). Then, using Langrange interpolation formula, one shows that f ∈ ′FmC[W ·h]. Thus, we have established a natural W × C×-equivariant isomorphism Spec(′gr C[W ·h]) ∼= W ×W1(cid:0) Spec grF C[pr−1 of (not necessarily reduced) schemes. 2 (h2)](cid:1), 2 (t) = fpr−1 2 (t) = fpr−1 Note further that we have pr−1 2 (h2) = W1 · h1. Moreover, the scheme Spec grF C[W1 · h1] is isomorphic to (ϑ1)−1(0) thanks to (7.2.5). Thus, one obtains W × C×-equivariant isomor- phisms (7.2.6) Spec(′gr C[W ·h]) ∼= W ×W1(cid:0)(ϑ1)−1(0)(cid:1) ∼= ϑ−1(W · h2). The isomorphisms in (7.2.3) follow since we have C[℘−1(0, 1)] = = ′gr C[W ·h] = C[ϑ−1(W · h2)]. Here, the first isomorphism holds by definition, the second isomorphism is (7.1.3), and the last isomorphism is (7.2.6). τ1 · Rmax C[W ·h] + (τ2 − 1) · Rmax C[W ·h] Rmax C[W ·h] To prove part (ii), we recall that the coinvariant algebra C[t]/(C[t]W1 + ) is a graded algebra which is known to be concentrated in degrees ≤ d1. We deduce, using isomorphisms (7.2.5) and (7.2.6), that the graded algebra ′gr C[W ·h] is also concentrated in degrees ≤ d1. Thus, we have ′Fd1 C[W ·h] = C[W ·h]. Part (ii) of the lemma follows. C[W ·h] = C[W ·h]. By symmetry, we get ′′Fd2 (cid:3) ♮ ⊂ C[τ1, τ2] ⊗ C[T] be a bihomogeneous ideal generated by the set 2 · fm,n(h) − fm,n fm,n ∈ (Cm,n[T])W , m, n ≥ 0}, The group C× × C× acts on C2 and on T. This makes C[C2 × T] = C[τ1, τ2] ⊗ C[T] a bigraded algebra with respect to the natural bigrading on a tensor product. Let I♮ = Li,j I i,j of bihomogeneous elements. It is clear that we have C[C2 × T]/I♮ = C[C2 ×T/W T], where the fiber product on the right involves the map C2 → T/W, (τ1, τ2) 7→ (τ1·h1, τ2·h2) mod W . According to the definition of the map ℵ of Lemma 7.1.4, for any fm,n ∈ Cm,n[T], we find ℵ(τ m 2 · (fm,n(h) − fm,n). The right hand side here clearly belongs to the subspace τ m 2 · (Ih ∩ Fm,nC[T]). Hence, for any i, j ≥ 0, one has an inclusion 2 · fm,n(h) − fm,n) = τ m 1 τ n {τ m 1 τ n 1 τ n 1 τ n 43 ℵ(I i,j defined map ℵ : C[C2 × T]/I♮ → RC[T]/RIh = Rmin C[W ·h]. 1τ j ♮ ) ⊂ τ i Thus, we obtain the following chain of bigraded W -equivariant C[τ1, τ2]-algebra maps 2 · (Ih ∩ Fi,j C[T]). These inclusions insure that the map ℵ descends to a well C[C2 ×T/W T] = C[C2 × T]/I♮ ℵ / / Rmin C[W ·h] can / / Rmax C[W ·h] = C[Y]. (7.2.7) 7.3. A 2-parameter deformation of e. It follows from definitions that the nilpotent elements e1, e2 and the semisimple elements h1, h2 satisfy the following commutation relations, cf. [Gi, Theorem 1.2] [e1, e2] = 0 = [h1, h2], (7.3.1) The above relations imply that the elements e1 + τ1 · h1 and e2 + τ2 · h2 commute for any τ1, τ2 ∈ C. Therefore, one can define the following map that will play an important role in the arguments below [hi, ej] = δij · ei, i, j ∈ {1, 2}. κ : C2 → C, (τ1, τ2) 7→ κ(τ1, τ2) = (e1 + τ1·h1, e2 + τ2·h2). (7.3.2) Lemma 7.3.3. For any f ∈ C[C]G and τ1, τ2 ∈ C, one has f (κ(τ1, τ2)) = f (τ1·h1, τ2·h2). Proof. This easily follows from Lemma 6.1.1. An alternative way of proving the lemma is based on a useful formula exp ad(τ · ei) (hj ) = hj − δi,j · τ · ei, i = 1, 2. Define a holomorphic (non-algebraic) map γ : C× × C× → G as follows γ(τ1, τ2) = exp( 1 e2). Then, using the formula, we find (we note that hi = 0 holds only if ei = 0): τ·hi = Ad exp( 1 ∀τ 6= 0, i = 1, 2; τ ei)(ei + τ·hi) e1 + 1 τ2 τ1 hence, we have (τ1·h1, τ2·h2) = Ad γ(τ1, τ2)(κ(τ1, τ2)), (7.3.4) (τ1, τ2) ∈ C× × C×. The last equation in (7.3.4) clearly implies the lemma. From Lemma 7.3.3 we deduce that, for any (τ1, τ2) ∈ C2, in C//G = T/W one has κ(τ1, τ2) mod G = (τ1 · h1, τ2 · h2) mod W . Thus, one can introduce a map eκ that fits into C2 ×T/W T the following diagram of cartesian squares C2 ×C (C ×C//G T) / C ×C//G T / T pr2 eκ (cid:3) C2 κ proj / C / C//G = T/W. The composite of the maps in (7.2.7) is a homomorphism of coordinate rings that gives a certain W ×C× × C×-equivariant morphism Y → C2×T/W T. Thus, we obtain the following chain of W × C× × C×-equivariant morphisms Y / C2 ×T/W T C2 ×C (C ×C//G T) eκ / C ×C//G T. (7.3.5) The composite morphism in (7.3.5) factors through the reduced scheme X, since the scheme Y is reduced by Lemma 7.2.1. Thus, we have constructed a W × C× × C×-equivariant mor- phism Υ that fits into a commutative diagram Y ℘ C2 Υ κ 44 / X p / Cred (7.3.6)   /   / / /   / / / / / /     /     / The proof of Theorem 1.7.2 is based on the following result concerning the structure of the map Υ over C2 r {0}, the complement of the origin 0 = (0, 0) ∈ C2. Proposition 7.3.7. (ii) For any (x, y) ∈ κ(C2 r {0}) we have: (i) The image of the map κ is contained in Cr. • The fiber p−1(x, y) is contained in the smooth locus of the variety X • The map p is flat over (x, y). (iii) The map ℘ × Υ yields the following isomorphism of schemes over C2 r {0}: ℘−1(C2 r {0}) ℘×Υ ∼ / (C2 r {0}) ×C X. Proof. Part (i) is clear for (τ1, τ2) = (0, 0) since we have κ(0) = (e1, e2) = e ∈ Cr. Thus, for the rest of the proof we may assume that (x, y) = κ(τ1, τ2) for some (τ1, τ2) ∈ C2 r {0}. One has an open covering C2 r 0 = U1 ∪ U2 where Ui = {(τ1, τ2) ∈ C2 τi 6= 0}. We will consider the case where (τ1, τ2) ∈ U2, the other case being totally similar. Thus, we have (x, y) = (e1 + τ1 · h1, e2 + τ2 · h2) for some τ2 6= 0. If τ1 6= 0 then the element (x, y) is G-conjugate to the element (τ1 · h1, τ2 · h2), by formula (7.3.4). In that case, one has (x, y) ∈ Crs and all the statements of parts (i)-(ii) of Proposition 7.3.7 are clear. It remains to consider the case τ1 = 0. Thus, we have (x, y) = (e1, e2 + τ2 · h2) where τ2 6= 0. Applying the formula in the first line of (7.3.4), we deduce that the element (x, y) is G-conjugate to the element (e1, τ2 · h2). Further, using the C×-action, we may assume without loss of generality that τ2 = 1. To complete the proof of parts (i)-(ii), we write G = C2 ⊗ g, resp. G × T = C2 ⊗ (g × t). The natural GL2-action on C2 induces, via the action on the first tensor factor, a GL2-action on G, resp. on G × T, such that the C× × C×-action considered earlier corresponds to the action of the maximal torus in GL2 formed by diagonal matrices. The scheme C ⊂ G, resp. X ⊂ G×T, is clearly GL2-stable. Further, it is clear that there is an element g ∈ GL2 that takes the point (e1, h2) ∈ C to the point (e1 + h2, h2). Now, according to [Gi, Proposition 3.2.3], the element e1 is a principal nilpotent in the Levi subalgebra gh2. It follows that e1 + h2 is a regular element of g. Hence, we have (e1 + h2, h2) ∈ C1, cf. Definition 2.6.3. It follows that (e1 + h2, h2) is a smooth point of C. Moreover, the fiber p−1(e1 + h2, h2) is contained in the smooth locus of X and the map p is flat over the point (e1 + h2, h2), by Lemma 2.1.5. All statements of Proposition 7.3.7(i)-(ii) follow from this using the GL2-action. We see from part (ii) that U2 ×C X is a reduced scheme, flat over U2. Proving part (iii) amounts to showing that the map ℘−1(U2) → U2×C X induced by ℘× Υ is an isomorphism. The map in question is a C× × C×-equivariant morphism of flat schemes over U2 = C× C×. Therefore, this map is an isomorphism if and only if it induces an isomorphism Ψ : ℘−1(0, 1) ∼→ {(0, 1)} ×C X = p−1(e1, e2 + h2), (7.3.8) of the corresponding fibers over a single point (0, 1) ∈ C × C×. Observe next that the argument used in the proof of part (ii), based on the GL2-action, yields an isomorphism of schemes p−1(e1, e2 + h2) ∼= p−1(e1 + h2, h2). Further, we know that e1 + h2 is a regular element of g. Therefore, taking the fibers at e1 + h2 of the locally free sheaves involved in the chain of isomorphism (6.5.3), yields an algebra isomorphism (e1+h2,h2) ∼= C[ϑ−1(W · h2)]. On the other hand, by Lemma 7.2.1, we C[p−1(e1 + h2, h2)] = R have C[℘−1(0, 1)] ∼= C[ϑ−1(W · h2)]. Thus, we obtain the following chain of W -equivariant algebra isomorphisms C[{(0, 1)} ×C X] = C[p−1(e1, e2 + h2)] ∼= C[p−1(e1 + h2, h2)] ∼= C[ϑ−1(W · h2)] ∼= C[℘−1(0, 1)]. 45 / We claim that the composite of the above isomorphisms is equal to the algebra map Ψ∗ : C[{(0, 1)} ×C X] → C[℘−1(0, 1)] induced by the morphism Ψ in (7.3.8). To see this, one observes that the algebra C[{(0, 1)} ×C X] is a quotient of the algebra C[T]. Hence, it suffices to verify our claim for linear functions on T, the generators of the algebra C[T]. Checking the latter is straightforward and is left to the reader. We conclude that the morphism Ψ in (7.3.8) is itself an isomorphism, and part (iii) of (cid:3) Proposition 7.3.7 follows. 7.4. Proof of Theorem 1.7.2. Part (i) of Proposition 7.3.7 implies that the map κ gives a closed imbedding κ : C2 ֒→ Cr. Therefore, we may view κ as a map C2 → Cnorm and form an associated fiber product C2 ×Cr Xnorm. From diagram (7.3.6) we obtain, by base change, the following commutative diagram of W × C× × C×-equivariant morphisms of schemes over C2 Y ℘ C2 ℘×Υ / C2 ×Cr X IdC2 ×ψ C2 ×Cr Xnorm (7.4.1) Id C2 Id pκ C2. Restricting this diagram to the open set C2 r {0} one obtains a diagram of isomorphisms ℘−1(C2 r {0}) (7.4.2) (C2 r {0}) ×Cr Xnorm. / (C2 r {0}) ×Cr X IdC2 ×ψ ℘×Υ ∼ ∼ Here the map ℘ × Υ on the left is an isomorphism thanks to part (iii) of Proposition 7.3.7 and the map IdC2 × ψ on the right induced by the normalization map ψ : Xnorm → X is an isomorphism thanks to part (ii) of Proposition 7.3.7. The map pnorm : Xnorm → Cnorm being flat, by flat base change we get that (pκ)∗OC2×Cr Xnorm ∼= κ∗[(pnorm)∗OXnorm] = κ∗R is a locally free sheaf on C2. We conclude that ℘∗OY and κ∗R are W × C× × C×-equivariant locally free sheaves of OC2-algebras and we have a W × C× × C×-equivariant algebra isomorphism (℘∗OY)C2 r{0} ∼= (κ∗R)C2r{0} induced by dia- gram (7.4.2). Therefore, the isomorphism can be extended (uniquely) across the origin 0 ∈ C2. The resulting isomorphism κ∗R ∼→ ℘∗OY is automatically W × C× × C×-equivariant and respects the OC2-algebra structures. Restricting the latter isomorphism to the fibers at ∼→ C[℘−1(0)]. On the the origin yields a bigraded W -equivariant algebra isomorphism Re other hand, by (7.2.2), we have C[℘−1(0)] = grmax C[W ·h]. The isomorphism of Theorem 1.7.2 follows. The last claim in the theorem is a consequence of Lemma 7.2.1(ii). (cid:3) 7.5. Proof of Theorem 1.8.1. First of all, we introduce some notation and define several vector spaces associated with the pair ′F qC[T] and ′′F qC[T] of filtrations on C[T]. We may (and will) view elements of the vector space Sym T as constant coefficient differ- ential operators on T = t × t. Thus, an element u ∈ Symi t ⊗ Symj t is a bihomogeneous differential operator of order i with respect to the first, resp. of order j with respect to the second, factor in the cartesian product t× t = T. For any m ≥ 0, we write Sm := Symm t and letbS :=Qi,j≥0 (Symi t ⊗ Symj t). For any i, j ≥ 0, the assignment u × f 7→ u(f )(0) gives a perfect pairing (Si ⊗ Sj) × Ci,j[T] → C. We obtain canonical isomorphisms (cid:0)C[T](cid:1)∗ Ci,j[T](cid:1)∗ ∼= Yi,j≥0 (cid:0)Ci,j[T](cid:1)∗ = Yi,j≥0 = (cid:0)Mi,j (Si ⊗ Sj) = bS . (7.5.1) 46     /     o o     / o o Thus, one has a perfect pairingbS×C[T] → C. It is instructive to think of an element u ∈bS as a constant coefficient differential operator on T "of infinite order". Accordingly, we write the above pairing as u × f 7→ u(f )(0) similarly to the finite order case. Given a vector subspace H ⊂ C[T], let H ⊥ ⊂bS denote the annihilator of H with respect the pairing. Below, we will use the following simplified notation ′Fi := ′FiC[T], resp. ′′Fj := ′′Fj C[T], and Fi,j = Fi,j C[T] = ′Fi ∩ ′′Fj. Further, let I = Ih so, we have C[T]/I = C[W ·h]. Then, for any m, n ≥ 0, there are natural imbeddings Fm,n = ′Fm ∩ ′′Fn / [′Fm + I] ∩ [′′Fn + I]  ′Fm + ′′Fn + I  / C[T]. These imbeddings induce the following chain of natural linear maps grm,n C[T] = Fm,n Fm,n−1 + Fm−1,n = a / grmax m,n C[W ·h] = ′Fm ∩ ′′Fn ′Fm−1 ∩ ′′Fn + ′Fm ∩ ′′Fn−1 [′Fm + I] ∩ [′′Fn + I] [′Fm−1 + I] ∩ [′′Fn + I] + [′Fm + I] ∩ [′′Fn−1 + I] b C[T] ′Fm−1 + ′′Fn−1 + I . (7.5.2) :=Y{i≥0, j>n} >j . Hence, we get F ⊥ Si ⊗ Sj. i,j = (′Fi ∩ ′′Fj)⊥ = ∩ I ⊥. Further, we find >j−1 ∩ I ⊥; ) >n−1 = Sm ⊗ Sn. (7.5.3) ′bS For each m, n ≥ 0, let >i >j >n >m >i >i resp. :=Y{i>m, j≥0} ′bS Clearly, we have (′Fi)⊥ = ′bS and [′Fi + I]⊥ = ′bS + ′′bS ′Fi−1 + ′′Fj−1 + I(cid:19)∗ (cid:18) (cid:18) Fm,n−1 + Fm−1,n(cid:19)∗ ′′bS Si ⊗ Sj, , resp. (′′Fj)⊥ = ′′bS ∩ I ⊥, resp. [′′Fj + I]⊥ = ′′bS ∩ ′′bS = (′Fi−1 + ′′Fj−1 + I)⊥ = ′bS ) ∩ (′bS + ′′bS + ′′bS (′bS Fm,n C[T] >m−1 >i−1 >m >m >n = >j >n + ′′bS ′bS C[W ·h])∗ Therefore, dualizing the maps in (7.5.2) one gets the following linear maps a∗ >m−1 >n−1 b∗ s / (grm,n ∩ I ⊥ / (grmax m,n C[T])∗ = Sm ⊗ Sn. ′bS ∩ ′′bS Proof of Theorem 1.8.1. We write α ∈ t for the coroot corresponding to a root α ∈ R+. Recall s ⊂ R+ for the set of positive roots of the Levi subalgebra gs, s = 1, 2, and the notation R+ put δs :=Qα∈R+ α ∈ Sds. Further, let Vs = C[W ]· δs ⊂ Sds, s = 1, 2, be the W -submodule generated by the element δs. It is known, that Vs is a simple W -module and, moreover, this W -module occurs in Sds with multiplicity one, see [Mc]. Therefore, there is a canonically defined copy of the simple W × W -submodule V1 ⊗ V2 inside Sd1 ⊗ Sd2. Dually, there is a canonically defined copy of the simple W × W -submodule V1 ⊗ V2 inside Cd1,d2[T] where Vs ⊂ Cds[t], s = 1, 2, stands for the contragredient W -module (in fact, one has Vs ∼= Vs, since any simple W -module is known to be selfdual). (7.5.4) Following [Gi, §4], we consider the element ∇ := Pw∈W sign(w) · ew(h) ∈ bS. More ∇ = Xi,j≥0 i! · j! Xw∈W ∇i,j where ∇i,j := 2) ∈ Si ⊗ Sj. sign(w) · w(hi 1 ⊗ hj (7.5.5) explicitly, we let 1 47   /  / /  / / / / / / Using the Taylor formula, for any polynomial f on T, we find ew(h)(f )(0) = f (w(h)). Hence, we get ∇(f )(0) =Pw∈W sign(w)·f (w(h)). It follows that the linear function C[T] → C, f 7→ ∇(f )(0) annihilates the ideal I = Ih in other words, we have ∇ ∈ I ⊥. Further, by [Gi, Lemma 4.3], one has ∇i,j = 0 whenever i < d1 or j < d2. We conclude that ∇ ∈ ′bS ∩ I ⊥. Thus, there is a well defined element a∗(b∗(∇)) ∈ Sd1 ⊗ Sd2, cf. (7.5.4). In addition, it is clear from (7.5.3)-(7.5.5) that we have an equation: ∩ ′′bS >d1−1 >d2−1 ∇d1,d2 = a∗(b∗(∇)) mod (′bS >d1 + ′′bS >d2). From now on, we assume that e is a non-exceptional principal nilpotent pair. Then, accord- ing to [Gi, Theorem 4.4], one has an isomorphism V2 ∼= V1 ⊗ sign of W -modules. It follows that (V1 ⊗ V2) sign is a 1-dimensional vector space, moreover, according to loc cit, the element ∇d1,d2 is a nonzero element of that vector space. Dually, ( V1 ⊗ V2) sign is a 1-dimensional vector space and ∆e is a nonzero element of that vector space. The canonical perfect pairing (V1 ⊗ V2) × ( V1 ⊗ V2) → C is W -invariant with respect to the W -diagonal action. Therefore, this pairing yields a perfect pairing between (V1 ⊗ V2) sign and ( V1 ⊗ V2) sign, the corresponding sign-isotypic components. These are 1-dimensional vector spaces, with ∆e and ∇d1,d2 being respective base vectors. Hence, one must have ∇d1,d2(∆e)(0) 6= 0. Thus, writing hh−,−ii : (grmax C[W ·h]) → C for the canonical pairing and using (7.5.6), we deduce C[W ·h])∗ × (grmax d1,d2 d1,d2 (7.5.6) hhb∗(∇), a(∆e)ii = hha∗(b∗(∇)), ∆eii = ∇d1,d2(∆e)(0) 6= 0. C[W ·h]) sign. In particular, we get (grmax We know that grmax C[W ·h] is a graded algebra such that grmax The map a in (7.5.2) is clearly W -equivariant (with respect to the action induced by the W - diagonal action on C[t] ⊗ C[t]). Thus, we have shown that a(∆e) is a nonzero element of the vector space (grmax d1,d2 C[W ·h] = 0 whenever i > d1 or j > d2, by Theorem 1.7.2. Also, we have dim(grmax C[W ·h]) sign = 1 and Theo- rem 1.7.2 says that the W -equivariant projection grmax C[W ·h] ։ (grmax C[W ·h]) sign gives a nondegenerate trace on the algebra grmax C[W ·h]. Therefore, since (grmax C[W ·h]) sign 6= 0, C[W ·h]) sign = (grmax C[W ·h]) sign. In C[W ·h] = (grmax it follows that we must have grmax d1,d2 particular, the vector space grmax C[W ·h] is 1-dimensional, with a(∆e) being a base vector. d1,d2 The theorem follows. C[W ·h]) sign 6= 0. d1,d2 d1,d2 d1,d2 (cid:3) i,j 7.6. The algebra Re. We use the notation of the previous section and keep the assumption that e is a non-exceptional principal nilpotent pair. Proof of Corollary 1.7.4. Let Oe,0 denote the local ring of the algebra C[X] at the point (e, 0) ∈ G × T. The variety X is normal at (e, 0) if and only if the canonical imbedding Oe,0 → Oe,0 ⊗C[X] C[Xnorm] is an isomorphism. The latter holds, by the Nakayama lemma, if and only if the natural map C[p−1(e)] → C[p−1 norm(e)] = Re is surjective. This yields the equiva- lence (i) ⇔ (ii) of the Corollary, since the restriction map C[T] → C[p−1(e)] is surjective. The equivalence (ii) ⇔ (iii) is a direct consequence of Theorem 1.7.2. (cid:3) Proof of Corollary 1.7.3. Let I = Ih be the ideal of the reduced subscheme W ·h ⊂ T and choose d >> 0 such that ′′FdC[W ·h] = C[W ·h] (by Lemma 7.2.1(ii), one can take d = d2). Thus, we have ′′FdC[T] + I = C[T]. Also, by definition, one has ′F0C[T] = 1⊗ C[t]. Therefore, we find F0,dC[W ·h] = ((1 ⊗ C[t]) + I)) ∩ (′′FdC[T] + I) I 48 = (1 ⊗ C[t]) + I I ∼= 1 ⊗ C[t] (1 ⊗ C[t]) ∩ I . The ideal (1 ⊗ C[t]) ∩ I corresponds, via the identification 1 ⊗ C[t] = C[t], to the ideal in C[t] of the image of the orbit W ·h under the second projection T → t. This image equals the set W · h2 ⊂ t. We conclude that the rightmost space in the displayed formula above is isomorphic to C[W · h2]. Thus, using that formula, we obtain W -module isomorphisms ′′gr(F0,dC[W ·h]) ∼= F0,dC[W ·h] ∼= C[W · h2] ∼= C[W/W1]. This yields the first isomorphism of the corollary. The second isomorphism is proved (cid:3) similarly. As an immediate consequence of the (proof of) Theorem 1.8.1, we have Lemma 7.6.1. There is a nonzero constant c ∈ C such that ∆e(wh) = c · sign(w), ∀w ∈ W . Proof. Any W -alternating function on the orbit W ·h is a constant multiple of the function signW ·h : w(h) 7→ sign(w). The function ∆eW ·h is W -alternating, hence it must be propor- tional to the function signW ·h. We know that the class of ∆e in grmax C[W ·h] is nonzero, by d1,d2 Theorem 1.8.1. Hence, the function ∆eW ·h is not identically zero and the result follows. (cid:3) Let f be a holomorphic function on T. The pull-back of f via the composite Xnorm → X → T is a G-invariant holomorphic function on Xnorm. One may view that function as a holomorphic section, f ♯, of the sheaf R, resp. view the restriction of that function to the subscheme p−1 norm(e)], the value of the section f ♯ at the point e. norm(e) ֒→ Xnorm as an element rese(f ) ∈ Re = C[p−1 We know, thanks to Theorem 1.8.1 and Lemma 7.6.1, that the class in grmax C[W ·h] of the function signW ·h gives a base vector of the 1-dimensional vector space (grmax C[W ·h]) sign = grmax be the image of that base vector under the isomorphism d1,d2 (grmax C[W ·h]) sign ∼→ R sign vector space R sign , of Theorem 1.7.2. Thus, u is a base vector of the 1-dimensional C[W ·h]. Let u ∈ R sign e e . e It is clear that, for any W -alternating function f , on T, there is a constant c(f ) ∈ C such that one has rese(f ) = c(f ) · u. Proof of Corollary 1.8.4. Let prsign W ·h : C[W ·h] ։ C[W ·h] sign, be the projection to the sign isotypic component. Further, we will use a natural identification Ci,j[T] ζ : C[T] = gr C[T], as bigraded algebras, resulting from the bigrading on C[T] =Li,j on the algebra C[T] itself. We have a diagram : Re ։ R sign , resp. prsign R e C[T] ζ rese / Re prsign R / R sign e (7.6.2) Theorem 1.7.2 Theorem 1.8.1 gr C[T] a / grmax C[W ·h] prsign W ·h / grmax d1,d2 C[W ·h]. Here, the map a is induced by the projection C[T] ։ C[W ·h], cf. (7.5.2). Going through the construction of the isomorphism of Theorem 1.7.2, see esp. formulas (7.2.7) and (7.3.5), shows that the left square in diagram (7.6.2) commutes. The right square commutes trivially. For any W -alternating function f on T we have fW ·h = f (h) · signW ·h. Assume, in addition, that f ∈ (Cd1,d2[T]) sign, a homogeneous polynomial of bidegree (d1, d2). Then, the equation fW ·h = f (h) · signW ·h, in C[W ·h], implies an equation a(ζ(f )) = f (h) · signW ·h, in grmax d1,d2 C[W ·h]. 49 / / / / We transport the latter equation via the isomorphism of Theorem 1.7.2. Thus, using com- mutativity of diagram (7.6.2) and our definition of the base vector u ∈ R sign, we deduce rese(f ) = f (h) · u, ∀f ∈ (Cd1,d2[T]) sign. (7.6.3) Next, by definition one has se = ∆♯ e. Hence, se(e) = rese(∆e). Thus, applying formula (7.6.3) to the polynomial ∆e, we get se(e) = ∆e(h) · u. Now, Lemma 7.6.1 implies that ∆e(h) 6= 0. It follows that se(e) 6= 0. (cid:3) Remark 7.6.4. We have shown in the course of the proof of Theorem 1.8.1 that, we have ∇d1,d2(∆e)(0) 6= 0. This provides an alternative proof that ∆e(h) 6= 0, thanks to a simple formula ∇d1,d2(∆e)(0) = (d1!d2!)2 · ∆e(h). Proof of Propositions 1.8.5. For each x ∈ T, we let Ex(−) := E(x,−), a W -alternating holo- morphic function on T. Unraveling definitions, one finds that SC×{e} = s♯ ⊠ u, where s♯ is a G-invariant holomorphic section of the sheaf R sign that corresponds to the holomorphic function s : T → C, x 7→ c(Ex). bidegree (i, j) by the formula Fix x = (x1, x2) ∈ T. For each i, j ≥ 1, we define a homogeneous polynomial on T of Writing y = (y1, y2) ∈ T, one obtains an obvious expansion Eij x (y1, y2) = 1 i!j! Xw∈W sign(w)·ehhhx,w(y)iii = Xw∈W sign(w) · (hx1, w(y1)i)i(hx2, w(y2)i)j . sign(w)·ehx1,w(y1)i·ehx2,w(y2)i = Xi,j≥0 Ex(y) = Xw∈W Eij x (y1, y2). Observe next that the composite map pr sign W ·h ◦ a, in the bottom row of diagram (7.6.2), annihilates the homogeneous component Ci,j[T] unless we have (i, j) = (d1, d2). From the commutativity of the diagram we deduce that the map pr sign ◦ rese, in the top row of the diagram, annihilates all the components Ci,j[T], (i, j) 6= (d1, d2). It follows that we have pr sign and Ex is W - alternating, we deduce rese(Ex) = rese(Ed1,d2 ). Hence, applying formula (7.6.3), we obtain )). Since each of the functions Ed1,d2 R (rese(Ex)) = pr sign R (rese(Ed1,d2 x x x R where u is the base vector of R sign (h) · u, defined earlier in this section. Now, using the definition of the polynomial ∆e, we compute rese(Ex) = rese(Ed1,d2 ) = Ed1,d2 x x e Ed1,d2 x (h) = 1 sign(w)·(hx1, w(h1)i)d1 (hx2, w(h2)i)d2 = 1 d1!d2!·∆e(x). d1!d2! Xw∈W Recall finally the function s : x 7→ c(Ex), where the quantity c(Ex) is determined from the equation rese(Ex) = c(Ex) · u. Thus, using the above formulas, we find s(x) = c(Ex) = Ed1,d2 (h) = 1 x d1!d2! · ∆e(x). Since se = ∆♯ d1!d2! · se. Also, according to the proof of Corollary 1.8.4, we have se(e) = ∆e(h) · u. e, for the corresponding sections of the sheaf R we deduce an equation s♯ = 1 Combining everything together, we obtain SC×{e} = s♯ ⊠ u = 1 d1!·d2! · se ⊠ u = d1!·d2!·∆e(h) · se ⊠ se(e), 1 proving the first equation of Proposition 1.8.5. The proof of the second equation is similar. (cid:3) 50 8. RELATION TO WORK OF M. HAIMAN 8.1. A vector bundle on the Hilbert scheme. We keep the notation of §1.9, so G = GLn and V = Cn. We have g = EndC V . For any (x, y) ∈ C, one may view the vector space gx,y as an associative subalgebra of EndC V . Let Gx,y ⊂ G be the isotropy group of the pair (x, y) under the G-diagonal action. The group Gx,y may be identified with the group of invertible elements of the associative algebra gx,y hence, this group is connected. Let Chx, yi denote an associative subalgebra of EndC V generated by the elements x and y. Recall that a vector v ∈ V is said to be a cyclic vector for a pair (x, y) ∈ C if one has Chx, yiv = V . Let C◦ be the set of pairs (x, y) ∈ C which have a cyclic vector. Part (i) of the following result is due to Neubauer and Saltman and part (ii) is well-known, cf. [NS, Theorem 1.1]. Lemma 8.1.1. (i) A pair (x, y) ∈ C is regular if and only if one has Chx, yi = gx,y. (ii) The set C◦ is a Zariski open subset of Cr. For (x, y) ∈ C◦, all cyclic vectors for (x, y) form a single Gx,y-orbit, which is the unique Zariski open dense Gx,y-orbit in V . (cid:3) Let G act on C × V by g : (x, y, v) 7→ (gxg−1, gyg−1, gv). We introduce a variety of triples S = {(x, y, v) ∈ g × g × V (cid:12)(cid:12) [x, y] = 0, Chx, yiv = V }. (8.1.2) This is a Zariski open G-stable subset of C × V . Lemma 8.1.1 shows that S is smooth and that the G-action on S is free. It is known that there exists a universal geometric quotient morphism ρ : S → S/G. Furthermore, the variety S/G may be identified with Hilb := Hilbn(C2), the Hilbert scheme of n points in C2, see [Na]. The first projection C × V → C restricts to a G-equivariant map δ : S → C◦. We let GC◦, the universal stabilizer group scheme on C◦, cf. §5.4, act along the fibers of δ by (γ, x, y) : (x, y, v) 7→ (x, y, γv). Lemma 8.1.1(i) implies that GC◦ is a smooth commutative group scheme, furthermore, the GC◦-action makes S → C◦ a G-equivariant GC◦-torsor over C◦. In particular, δ is a smooth morphism. We use the notation of §1.5 and put P = (ρ∗δ∗(RC◦))G. There is a W -action on P inherited from the one on R. By equivariant descent, one has a canonical isomorphism ρ∗P = δ∗(RC◦), cf. diagram (8.2.1) below. The sheaf RC◦ is locally free (Theorem 1.5.2) and self-dual (Theo- rem 1.3.4). Thus, we deduce the following result Corollary 8.1.3. The sheaf P is a locally free coherent sheaf of commutative OHilb-algebras equipped with a natural W -action by algebra automorphisms. The fibers of the corresponding algebraic vector bundle on Hilb afford the regular representation of the group W and the projection P ։ P sign induces a nondegenerate trace on each fiber. (cid:3) 8.2. The isospectral Hilbert scheme. Let W := SpecHilb P be the relative spectrum of P, a sheaf of algebras on the Hilbert scheme. By Corollary 8.1.3, the scheme W comes equipped with a flat and finite morphism η : W → Hilb and with a W -action along the fibers of η. We conclude that W is a reduced Cohen-Macaulay and Gorenstein variety. One can interpret the construction of the scheme W in more geometric terms as follows. norm(C◦), an open subvariety in Xnorm. We form a fiber product Xnorm ×C◦ S and consider the following commutative diagram whereep denotes the second projection, cf. also [Ha2, §8], Let X◦ := p−1 51 eδo X◦ X◦ ×C◦ S %KKKKKKKKKK pnorm C◦ ep &MMMMMMMMMMM &MMMMMMMMMM (1.3.1) δ h ∼= S / S ×Hilb W / W eρ yrrrrrrrrrr η eη xppppppppppp xqqqqqqqqqqq ρ Hilbert-Chow / Hilb (8.2.1) We let G act on S ×Hilb W through the first factor, resp. act diagonally on X◦ ×C◦ S. The morphisms δ and ρ in diagram (8.2.1) are smooth, resp. the morphisms pnorm and η are finite T/W C◦ S , base change. Hence, flat base change yields resp. ρ∗P = ρ∗η∗OW =eη∗OS×HilbW . and flat. The morphismeδ, resp.eρ,ep, andeη, is obtained from δ, resp. from ρ, pnorm, and η, by δ∗(RC◦) = δ∗(pnorm)∗OX◦ =ep∗OX◦ × C◦ S ∼= eη∗OS×HilbW of Since δ∗(RC◦) = ρ∗P we deduce a canonical isomorphism ep∗OX◦× G-equivariant sheaves of OS -algebras. This means that there is a canonical G-equivariant isomorphism of schemes h : X◦ ×C◦ S ∼→ S ×Hilb W, the dotted arrow in diagram (8.2.1). The map eρ makes S ×Hilb W a G-torsor over W. Therefore, the composite eρ ◦ h makes X◦ ×C◦ S a G-torsor over W hence, we have W = (X◦ ×C◦ S)/G, a geometric quotient by G. From (8.2.2), we obtain ρ∗δ∗(RC◦) = ρ∗ep∗OX◦×C◦ S = ρ∗eη∗OS×HilbW = η∗eρ∗OS×HilbW. Therefore, taking G-invariants, we get P = [ρ∗δ∗(RC◦)]G = [η∗eρ∗OS×HilbW]G = η∗OW. The W -action on X◦ induces one on X◦ ×C◦ S. This W -action commutes with the G- diagonal action, hence, descends to (X◦ ×C◦ S)/G. The resulting W -action may be identified with the one that comes from the W -action on the sheaf P, see Corollary 8.1.3. The composite map X◦ ×C◦ S → X◦ → [C◦ ×C◦//G T]red → T descends to a W -equivariant map σ : W → T. Thus, we have a diagram (8.2.2) eδo X◦ X◦ ×C◦ S eρ ◦ h (X◦ ×C◦ S)/G = W η×σ / Hilb×T/W T. (8.2.3) Following Haiman, one defines the isospectral Hilbert scheme to be [Hilb ×T/W T]red, a re- duced fiber product. Proposition 8.2.4. The map η × σ on the right of (8.2.3) factors through an isomorphism W ∼→ [Hilb×T/W T]norm. In particular, the normalization of the isospectral Hilbert scheme is Cohen-Macaulay and Gorenstein. Proof. It is clear that eρ ◦ h(eδ−1(Xrs)) is a Zariski open subset of W. Lemma 2.1.3(ii) implies readily that the map η × σ restricts to an isomorphism eρ ◦ h(eδ−1(Xrs)) ∼→ Hilb×T/W Tr; in particular, η × σ is a birational isomorphism. The image of the map η × σ is contained in [Hilb ×T/W T]red since the scheme W is reduced. Further, Corollary 8.1.3 and Lemma 2.6.4 imply that the scheme W is Cohen-Macaulay and smooth in codimension 1. We conclude that W is a normal scheme which is birational and finite over [Hilb ×T/W T]red. This yields the isomorphism of the proposition. (cid:3) As a consequence of Haiman's work, since P = η∗OW we deduce Corollary 8.2.5. The vector bundle P on Hilb is isomorphic to the Procesi bundle, cf. [Ha1]. (cid:3) 52 % o o o & & & / x / / / y y y & & & o o o o / / / x x x o o o / / / / / Remark 8.2.6. For any (x, y) ∈ C, one has a surjective evaluation homomorphism C[z1, z2] ։ Chx, yi, P 7→ P (x, y). The kernel of this homomorphism is an ideal Ix,y ⊂ C[z1, z2]. If (x, y) ∈ Cr then we have C[z1, z2]/Ix,y = Chx, yi = gx,y, by Lemma 8.1.1(i). Hence, Ix,y has codimension n in C[z1, z2]. Therefore, the assignment (x, y) 7→ Ix,y gives a well defined morphism : Cr ։ Hilb such that one has (C◦ ) ◦ δ = ρ. It follows that ∗P, the pull back of the sheaf P, is a G × W -equivariant locally free sheaf of OCr -algebras. Furthermore, we have canonical isomorphisms δ∗(∗P) = ρ∗P = δ∗R of G × W × GC◦-equivariant sheaves of OS -algebras. By equivariant descent, this yields a canonical isomorphism (∗P)C◦ ∼= RC◦. Since Cr r C◦ is a codimension ≥ 2 subset in the smooth variety Cr and the sheaves ∗P and RCr are both locally free, we deduce that the above isomorphism extends canonically to an isomorphism ∗P ∼= RCr of G× W -equivariant locally free sheaves of OCr -algebras. In other words, there is a G × W -equivariant isomorphism Cr ×Hilb W ∼= p−1 norm(Cr) of schemes over Cr. we put MapG(X, E) = (cid:0)C[X] ⊗ E(cid:1)G 8.3. Proof of Theorem 1.9.1. Given a G-variety X and a rational representation E of G , the vector space of G-equivariant polynomial maps X → E. If X → Y is a G-torsor, we let LY (E) be the sheaf of sections of an associated vector bundle X ×G E → Y on Y . By definition, one has Γ(Y, LY (E)) = MapG(X, E). We may apply the above to the geometric quotient morphism ρ : S → Hilb. The tauto- logical bundle on the Hilbert scheme is defined to be V := LHilb(V ), a rank n vector bundle associated with the vector representation of G = GL(V ) in V . Remark 8.3.1. According to Lemma 8.1.1, for any (x, y, v) ∈ S, one has a canonical vector C◦ ) ∼→ ρ∗V, space isomorphism gx,y → V, a 7→ a(v). This gives a canonical isomorphism δ∗(ggg of G-equivariant sheaves on S. ♦ For any m ≥ 0, one has a vector bundle LW(V ⊗m) on W associated with the G-repre- η∗(LHilb(V ⊗m)) = η∗V⊗m where η : W → Hilb is the map from (8.2.1). Hence, the projection formula yields sentation V ⊗m and with the G-torsoreρ ◦ h : X◦ ×C◦ S → W. Clearly, we have LW(V ⊗m) = MapG(X◦ ×C◦ S, V ⊗m) = Γ(W, LW(V ⊗m)) = Γ(Hilb, η∗η∗V⊗m) = Γ(Hilb, P ⊗ V⊗m), where we have used that P = η∗OW. Let Φ1 be the composite of the above isomorphisms. Write u 7→ (xu, yu), xu, yu ∈ g for the projection Xnorm → Cred. Then, by definition, we have X◦ ×C◦ S = {(u, v), u ∈ Xnorm, v ∈ V C[xu, yu]v = V }. Thus, there is a natural open imbedding α : X◦ ×C◦ S ֒→ Xnorm × V and one has the corresponding restriction morphism (8.3.2) Let C× ⊂ GLn be the group of scalar matrices. The scalar matrix z·Id ∈ GLn, z ∈ C×, acts trivially on Cred hence, for any ℓ ≥ 0, the element z· Id acts on the subspace C[Cred]⊗ Cℓ[V ] ⊂ C[Cred × V ] via multiplication by z−ℓ. It follows that the natural inclusion Cm[V ] ֒→ C[V ] induces an isomorphism Cm[V ] ⊗ V ⊗m ∼→(cid:0)C[V ] ⊗ V ⊗m(cid:1)C× . The group G = GLn being generated by the subgroups C× and SLn, we obtain a chain of isomorphisms (cid:0)C[Xnorm]⊗ Cm[V ]⊗ V ⊗m(cid:1)SLn ∼→ MapG(Xnorm, C[V ]⊗ V ⊗m) ∼→ MapG(Xnorm × V, V ⊗m). α∗ : MapG(Xnorm × V, V ⊗m) −→ MapG(X◦ ×C◦ S, V ⊗m). Write Φ2 for the composite isomorphism. 53 The isomorphism of Theorem 1.9.1(i) is defined as the following composition α∗ (cid:0)C[Xnorm] ⊗ Cm[V ] ⊗ V ⊗m(cid:1)SLn Φ2−→∼ MapG(Xnorm × V, V ⊗m) −→ MapG(X◦ ×C◦ S, V ⊗m) Φ1−→∼ Γ(Hilb, P ⊗ V⊗m). All the above maps are clearly W × Sm-equivariant bigraded C[T]-module morphisms. We see that proving the theorem reduces to the following result Lemma 8.3.3. The restriction map α∗ in (8.3.2) is an isomorphism. Proof. The map α∗ is injective since the set X◦ ×C◦ S is Zariski dense in Xnorm × V . To prove surjectivity of α∗ recall that a linear operator x : V → V has a cyclic vector if and only if x ∈ gr. Also, for any (x1, x2) ∈ C and v ∈ V , either of the two equations, C[x1]v = V or C[x2]v = V , implies C[x1, x2]v = V . It follows that we have Crr = C1 ∪ C2 ⊂ C◦ where Ci, i = 1, 2, are the open sets introduced in Definition 2.6.3. Next, let Si, i = 1, 2, be the open subset of S formed by the triples (x1, x2, v) ∈ C× V such that we have C[xi]v = V . Restricting the map δ : S → C◦, (x1, x2, v) 7→ (x1, x2) gives well defined maps Si → Ci. This way, we get a commutative diagram of open imbeddings X◦ ×C◦ S α (8.3.4) / Xnorm × V Xi ×Ci Si αi Xrr ×Crr S  αrr (Xi × V )  / Xrr × V  βS βV Recall that the morphism δ : S → C◦ is smooth and the set Xnorm r Xrr has codimension ≥ 2 in Xnorm, by Lemma 2.6.4. Hence, the complement of the set Xrr ×Crr S has codimension ≥ 2 in X◦ ×C◦ S, resp. the complement of the set Xrr × V has codimension ≥ 2 in Xnorm × V . The varieties X◦ ×C◦ S and Xnorm × V are normal. Therefore, restriction of functions yields the following bijections: (βS )∗ : C[X◦ ×C◦ S] ∼→ C[Xrr ×Crr S], resp. (βV )∗ : C[Xnorm × V ] ∼→ C[Xrr × V ]. Similarly, the inclusion αi in diagram (8.3.4) induces a restriction map (8.3.5) α∗ i : MapG(Xi × V, V ⊗m) −→ MapG(Xi ×Ci Si, V ⊗m), i = 1, 2. (8.3.6) We will show in Lemma 8.4.1 below that this map is surjective. To complete the proof of Lemma 8.3.3, let f : Xrr ×Crr S → V ⊗m be a G-equivariant map. Then, thanks to the surjectivity of (8.3.6), there exist maps fi ∈ MapG(Xi× V, V ⊗m), i = 1, 2, such that one has fXi×Ci Si = fiXi×Ci Si. It follows that the restrictions of the maps f1 and f2 to the set (X1 × V ) ∩ (X2 × V ) agree. Therefore, the map f can be extended to a regular map Xrr × V = (X1 × V ) ∪ (X2 × V ) → V ⊗m. Thus, we have proved (modulo Lemma 8.4.1) the surjectivity of the restriction map (αrr)∗ induced by the vertical imbedding αrr in the middle of diagram (8.3.4). This, combined with isomorphisms (8.3.5), implies the surjectivity of the map α∗ and Lemma 8.3.3 follows. (cid:3) 8.4. To complete the proof of the theorem, it remains to prove the following Lemma 8.4.1. The map α∗ i , i = 1, 2, in (8.3.6) is surjective. Proof. We may restrict ourselves to the case i = 1, the case i = 2 being similar. Let x ⊂ g × t be the closed subvariety considered in §2.1, and let r = {(x, t) ∈ x x ∈ gr}. Write q : X → x, (x, y, t1, t2) 7→ (x, t1) for the projection. According to Lemma 2.1.5 we 54   / / _     / / _    _     /  / have that q−1(r) = Nr, the total space of the conormal bundle on r. Thus, we obtain X1 = (cid:8)(x, y, t1, t2) ∈ X, x ∈ gr(cid:9) = q−1(r) = Nr. It will be convenient to introduce the following set Y := {(x, t, v) ∈ g × t × V (x, t) ∈ x & C[x]v = V }. Note that, for any (x, t, v) ∈ Y , the element x is automatically regular, as has been already observed earlier. Therefore, we have Y ⊂ r × V and the projection (x, t, v) 7→ (x, t) gives a smooth and surjective morphism Y → r. Unraveling the definitions we obtain X1 ×C1 S1 = {(x1, x2, t1, t2, v) ∈ X × V C[x1]v = V } = q−1(r) ×r Y = Nr ×r Y. Using these identifications, we see that the map α1 from (8.3.4) fits into a cartesian square α1 / Nr × V qV :=q×IdV (8.4.2) Nr ×r Y prY Y  / r × V Let D := (r×V )rY . This is an irreducible divisor in r×V , the principal divisor associated with the function (x, t, v) 7→ hvol, v∧ x(v)∧ x2(v)∧ . . .∧ xn−1(v)i where vol ∈ ∧nV ∗ is a fixed nonzero element. Using the cartesian square in (8.4.2), we see that (Nr × V ) r (Nr ×r Y ) = (qV )−1(D) is an irreducible divisor in Nr × V . Now, view D as a subset of g × t × V . Then, Drs := D ∩ (grs × t × V ) is an open dense subset of D hence (qV )−1(Drs) is an open dense subset of (qV )−1(D). Let T be the maximal torus of diagonal matrices in GLn. Thus, the set t ∩ gr may be identified with the set of matrices t = diag(z1, . . . , zn) ∈ Cn such that zi 6= zj ∀i 6= j. Let x = t = diag(z1, . . . , zn) ∈ t ∩ gr. For v = (v1, . . . , vn) ∈ Cn = V , the triple (x, t, v) is contained in Drs if and only if v is not a cyclic vector for the linear operator x which holds if and only if there exists an i ∈ [1, n] such that vi = 0. Let Vi ⊂ V denote the hyperplane formed by the elements with the vanishing i-th coordinate. To complete the proof of the lemma, let f ∈ MapG(X◦ ×C◦ S, V ⊗m). We must show that (α1)∗(f ), viewed as a map Nr ×r Y → V ⊗m, has no singularities at (Nr × V ) r (Nr ×r Y ), a divisor in a smooth variety. The set (qV )−1(Drs) is an open dense subset of that divisor. Hence, by G-equivariance, it suffices to prove that, for x = t ∈ t ∩ grs and a fixed element (x, y, t, t′) ∈ q−1(x, t) = Nr, the rational map fV : V → V ⊗m given by the assignment v 7→ f (x, y, t, t′, v) has no poles at the divisor ∪i Vi ⊂ V . It will be convenient to choose a T -weight basis {uγ} of the vector space V ⊗m. Thus, for each γ and any diagonal matrix diag(z1, . . . , zn) ∈ T , we have diag(z1, . . . , zn)uγ = zm(γ,1) · uγ where m(γ, i) ∈ Z≥0. Expanding the function fV in the basis {uγ} one can write fV (v) =Pγ f γ 1 v = (v1, . . . , vn) 7→ f γ Now, the map f hence also the map (α1)∗(f ) is G-equivariant. Since (x, y) ∈ T we deduce that fV is a T -equivariant map. Thus, for each γ and all diag(z1, . . . , zn) ∈ T , we must have Xk1,...,kn∈Z V are Laurent polynomials of the form aγ k1,...,kn ∈ C. V (v) · uγ where f γ V (v) = Xk1,...,kn∈Z aγ k1,...,kn · (z1v1)k1 ··· (znvn)kn · uγ = f γ aγ k1,...,kn · vk1 V [diag(z1, . . . , zn)v] ··· zm(γ,n) n 1 ··· vkn n , aγ k1,...,kn · vk1 1 ··· vkn n · [zm(γ,1) 1 ··· zm(γ,n) n · uγ]. = diag(z1, . . . , zn)[f γ V (v)] = Xk1,...,kn∈Z 55     /    / The above equation yields ki = m(γ, i) whenever aγ takes the following simplified form k1,...,kn 6= 0. Therefore, the function fV aγ · vm(γ,1) 1 ··· vm(γ,n) n · uγ, aγ ∈ C. (8.4.3) fV (v1, . . . , vn) =Xγ (cid:3) Since the weights of the T -action in V ⊗m are nonnegative, i.e., m(γ, i) ≥ 0, the right hand side of (8.4.3) has no poles at the divisor ∪i Vi and the lemma follows. 8.5. Proof of Proposition 1.9.2. Let vo = (1, 1, . . . , 1) ∈ Cn = V . Restriction of polynomial functions via the imbedding ζ : T = T×{vo} ֒→ Cred × V gives an algebra homomorphism (8.5.1) Given a G-variety X and an integer k ≥ 0, let C[X]detk = {f ∈ C[X] g∗(f ) = (det g)k · f, ∀g ∈ G} be the subspace of detk-semi-invariants of the group G = GLn. Looking at the action of the group C× ⊂ GLn, of scalar matrices, shows that one has C[Cred × V ]detk = (cid:0)C[Cred] ⊗ Ckn[V ](cid:1)SLn. ζ ∗ : C[Cred] ⊗ C[V ] = C[Cred × V ] −→ C[T], Following [GG], we introduce an affine variety (8.5.2) The assignment (x, y, v) 7→ (x, y, v, 0) gives a G-equivariant closed imbedding ι : Cred × ¯S := {(x, y, v, v∗) ∈ g × g × V × V ∗ [x, y] + v ⊗ v∗ = 0}. V ֒→ ¯S. Pull-back of functions via the imbedding ζ, resp. ι, yields linear maps f 7→ fT×{vo}. / C[Cred × V ]detk =(cid:0)C[Cred] ⊗ Ckn[V ](cid:1)SLn C[ ¯S]detk The composite map in (8.5.3) was considered in [GG]. According to Proposition A2 from [GG, Appendix], the image of the map ζ ∗ ◦ ι∗ is equal to Ak and, moreover, this map yields an isomorphism C[ ¯S]detk ∼→ Ak. Note further that the restriction map ι∗ in (8.5.3) is surjective since the group G is reductive. It follows that each of the two maps in (8.5.3) must be an (cid:3) isomorphism. This yields the statement of the proposition. k ≥ 0. / Ckn[T], (8.5.3) ζ ∗ ι∗ 9.1. For each Ad G-orbit O ⊂ g, the closure NO of the total space of the conormal bundle on O is a Lagrangian subvariety in T ∗g. We define 9. SOME APPLICATIONS Cnil := {(x, y) ∈ G [x, y] = 0, x ∈ N} = ∪O⊂N NO, (9.1.1) where the union on the right is taken over the (finite) set of nilpotent Ad G-orbits in g. Let γ : g → g//G = t/W be the adjoint quotient morphism. We introduce the following maps θC : Cred → g//G = t/W, (x, y) 7→ γ(y) The proposition below may be viewed as an analogue of the crucial flatness result in Haiman's proof of the n! theorem, see Proposition 3.8.1 and Corollary 3.8.2 in [Ha1]. Propo- sition 9.1.3 was also obtained independently by I. Gordon [Go] who used it in his proof of positivity of the Kostka-Macdonald polynomials. θX : X → t, (x, y, t1, t2) 7→ t2. (9.1.2) resp. Proposition 9.1.3. The composite morphism eθC : Cnorm → Cred θX→ t is flat. Furthermore, the scheme theoretic zero fibereθ−1 X scheme, not necessarily reduced in general. C (0), resp. eθ−1 56 θC→ g//G, resp. eθX : Xnorm → X (0), is a Cohen-Macaulay / / / / Proof. The dilation action of C× on t descends to a contracting C×-action on t/W . This makes dim(eθC)−1(t) = dim(θ−1 the mapeθC a C×-equivariant morphism. Hence, for any t ∈ t/W , one obtains C (t)) ≤ dim(θC)−1(0) = dim Cnil = dim g = dim C − dim t/W, Here, the inequality holds thanks to the semi-continuity of fiber dimension, the second equality is a consequence of the (set theoretic) equation (θC)−1(0) = Cnil, and the third equal- ity is a consequence of (9.1.1). The scheme Cnorm being Cohen-Macaulay, we conclude that the mapeθC is flat and each scheme theoretic fiber of that map is Cohen-Macaulay, cf. [Ma, The proof of the corresponding statements involving the variety X is similar and is left to (cid:3) §16A, Theorem 30]. the reader. 9.2. We take W -invariant global sections of the sheaves on each side of formula (2.4.1). The functor Γ(g × t,−)W being exact, one obtains an isomorphism of left D(g)-modules Γ(g × t,M)W = D(g)/D(g)·ad g. (9.2.1) Proof of Corollary 1.3.6. The Hodge filtration on M is W -stable and it induces, by restriction to W -invariants, a filtration F q on Γ(g × t,M)W . The functor Γ(g × t,−)W clearly commutes with taking an associated graded module. Therefore, we deduce grF [D(g)/D(g)·ad g] = grF [Γ(g × t, M)W ] =(cid:2) grHodge Γ(g × t, M)(cid:3)W = Γ(g × t, grHodge M)W = C[Xnorm]W = C[Cnorm], where the fourth equality holds by Theorem 1.3.3 and the fifth equality is a consequence of the isomorphism C[Cnorm] = C[Xnorm]W , see Corollary 1.5.1(i). (cid:3) We recall that the vector space D(g)/D(g)· ad g has a natural right action of the algebra A, cf. §2.4. The isomorphism in (9.2.1) intertwines the natural left D(t)W -action on Γ(g×t,M)W and the right A-action on D(g)/D(g)· ad g via the isomorphism D(t)W ∼= Aop induced by the map Ξ, see (2.4.1). Therefore, the filtration F q makes D(g)/D(g) · ad g a filtered (Dg, A)- bimodule. In particular, one may view D(g)/D(g) · ad g as a right filtered (Sym g)G-module via the map (Sym g)G → A given by the composition of natural maps (Sym g)G ֒→ D(g)G ։ D(g)G/[D(g)·ad g]G. Proof of Corollary 1.3.7. By Corollary 1.3.6, we have grF(cid:2)D(g)/D(g)·ad g(cid:3) = C[Cnorm]. There- fore, Corollary 1.3.5 implies that grF(cid:2)D(g)/D(g)·ad g(cid:3) is a Cohen-Macaulay C[G]-module; moreover, this module is flat over the algebra (Sym g)G, by Proposition 9.1.3. Now, a re- sult of Bjork [Bj] insures that D(g)/D(g)·ad g is a Cohen-Macaulay D(g)-module which is, moreover, flat over (Sym g)G. (cid:3) INDEX OF NOTATION 1.1 T ∗X, Xred, Xnorm, ψ, OX , DX , KX , G, T, g, t, W, r, sign D, rad, ad, M 1.2 1.2, 2.5, 4.2. G, C, T, κ, res, h−,−i v⊤ 1.2, 2.4 X 1.2, 2.1.4 R+ 1.4, 5.2 egrHodge 1.2 1.1 57 1.5, 5.1 1.5, 2.1.4 p pnorm, R, RE, E∗, gx, gx,y, Rx, gr, Cr ggg Small representation La, Lggg Sn e, •e, hs, h, gs, R+ C≤m[t], ′F q, ′′F q, s , Ws, ds 1.7, 7.1, 7.2 1.6 1.6, 1.9 1.6 1.7 1.5 1.8 1.9, 8.1 1.7 grF C[W ·h] Exceptional principal nilpotent pair hs, ∆e, se, se(e), E, S 1.8 n 2.1.4 (C2), C◦, P, Ak V 1.9, 8.3.1 grs, Crs, Xrs, g//G, C//G, x, xrs, r, Nr pT Hilbn(C2), gHilb q, F ord, grord, egrF , DX→Y , Rf , R R coh(FDX ), egr(M, F ) A, Ξ, I, J , F ordM, J tr, δt, dx, dh, j, j!∗Or FDX -mod, Db Hodge module, Hodge filtration HM(X), D, FD, Coh Z, Db F Hodge Hk coh(Z), gdgr, Z , Ci, Crr, Xi, Xrr, 2.4 2.4.2 2.2 2.3 2.3 2.5, 4.2 2.6 f 2.1 2.3 3.1 3.1.1 3.1.1, 3.4 π, ω q ǫ, Λ 3.2, 4.3 B, b, eg, eG, ν, ν, µ, µ, TX , prΛ→T ∗ B, prΛ→G×T, N , eN , Φ, Ψ, eκ eX, eXrr 3.3 3.2 3.3, 4.2 ı π 3.3.5, 4.5 A, Hj, eu, ∂eκ∗eu , ς (4.2.1) GX , gggX , LgggX R, X, 2, L 2λ b , L 3.4 5.1 2λ B , Lλ eg, gα eg B, gα 5.3 B, h, Lh 5.2 pr, egr, γ, eγ, ϑ, Lλ r , λL, λk Lλ 5.4 r , gα r , ek Lλ 5.7 α hx, nx, Gb 6.1 C(l), N(l), S, tl, Cl(g) RE, ′R, ′′R, ′gr, ′′gr, ℵ, F min, F max, can Ci,j[T], Ih, Y, ℘, I♮ Tl, Xl(g) 6.3 6.2 7.2 t l, ◦ ◦ R , pr sign f ♯, rese, u, pr sign W ·h, Ex Gx,y, Chx, yi, S, Hilb, ρ, δ 8.2 κ, eκ, Υ 7.3 Sm, bS, H ⊥, δs, Vs, Vs, ∇, ∇i,j W, X◦, η, eδ, eρ, ep, h, Cnil, θC, θX, eθC, eθX MapG, LY (E), α, αi, Φi, βS, βV Y, D, 8.4 8.1 9.1 7.5 7.6 8.3.1 7.1 REFERENCES [Ba] R. Basili, Some remarks on varieties of pairs of commuting upper triangular matrices and an interpretation of commuting varieties. Preprint 2008. arXiv:0803.0722. [BK] A. Beilinson, D. Kazhdan, Flat projective connections. Unpublished manuscript, 1991. [BG] G. Bellamy, V. Ginzburg, Some combinatorial identities related to commuting varieties and Hilbert schemes. [Bj] Preprint 2010. arXiv:1011.5957. J.-E. Bjork, The Auslander condition on Noetherian rings. S´eminaire d'Alg`ebre Paul Dubreil et Marie-Paul Malliavin, (Paris, 1987/1988), 137–173, Lect. Notes in Math., 1404, Springer, Berlin, 1989. [Bol] A. Bolsinov, Commutative families of functions related to consistent Poisson brackets. Acta Appl. Math. 24 (1991), 253-274. [Bo] A. Borel, Algebraic D-modules. Perspectives in Mathematics, 2. Academic Press, Inc., Boston, MA, 1987. [BB] W. Borho, J.-L. Brylinski, Differential operators on homogeneous spaces. II. Relative enveloping algebras. Bull. Soc. Math. France 117 (1989), 167–210. [Br] A. Broer, The sum of generalized exponents and Chevalley's restriction theorem for modules of covariants. Indag. [CM] Math. (N.S.) 6 (1995), 385-396. J.-Y. Charbonnel, A. Moreau, Nilpotent bicone and characteristic submodule of a reductive Lie algebra. Trans- form. Groups 14 (2009), 319-360. [CG] N. Chriss, V. Ginzburg, Representation theory and complex geometry. Birkhauser Boston, 1997. [E] D. Eisenbud, Commutative algebra. With a view toward algebraic geometry. Graduate Texts in Mathematics, 150. Springer-Verlag, New York, 1995. [GG] W. L. Gan, V. Ginzburg, Almost-commuting variety, D-modules, and Cherednik algebras, IMRP 2006:2, 1. [Gi] [GOV] V. Gorbatsevich, A. Onishchik, and E. Vinberg, Foundations of Lie theory and Lie transformation groups. V. Ginzburg, Principal nilpotent pairs in a semisimple Lie algebra. Invent. Math. 140 (2000), 511–561. [Go] Encyclopaedia Math. Sci., 20, Springer, Berlin, 1993. I. Gordon, Macdonald positivity via the Harish-Chandra D-module. Preprint 2010. to appear in Invent. Mathem. [Ha1] M. Haiman, Hilbert schemes, polygraphs and the Macdonald positivity conjecture, J. Amer. Math. Soc., 14 (2001) 941-1006. [Ha2] , Macdonald polynomials and geometry. New perspectives in algebraic combinatorics (Berkeley, CA, 1996–97), 207–254, Math. Sci. Res. Inst. Publ., 38, Cambridge Univ. Press, Cambridge, 1999. 58 [Ha3] [Ha4] , Combinatorics, symmetric functions and Hilbert schemes. Current Developments in Mathematics 2002, no. 1 (2002), 39-111. , Vanishing theorems and character formulas for the Hilbert scheme of points in the plane. Invent. Math. 149 (2002), 371-407. [HC] Harish-Chandra, Invariant differential operators and distributions on a semisimple Lie algebra. Amer. J. Math, 86 (1964), 534-564. [HK1] R. Hotta, M. Kashiwara, The invariant holonomic system on a semisimple Lie algebra. Invent. Math. 75 (1984), 327–358. , [HK2] , Quotients of the Harish-Chandra system by primitive ideals. Geometry today (Rome, 1984), 185–205, Progr. Math., 60, Birkhauser Boston, Boston, MA, 1985. [HTT] , K. Takeuchi, T. Tanisaki, D-modules, perverse sheaves, and representation theory. Progress in Mathe- matics, 236. Birkhauser Boston, Inc., Boston, MA, 2008. A. Joseph, On a Harish-Chandra homomorphism. C.R.Acad. Sci. Paris, 324 (1997), 759-764. [Jo] [KNV] S. Khoroshkin, M. Nazarov, E. Vinberg, A generalized Harish-Chandra isomorphism. Adv. Math. 226 (2011), 1168-1180. B. Kostant, Lie group representations on polynomial rings, Amer. J. Math., 85 (1963), 327–404. [Ko] [Kr] H. Kraft, Geometrische Methoden in der Invariantentheorie. Aspects of Mathematics, D1. Friedr. Vieweg & Sohn, Braunschweig, 1984. , N. Wallach, On the nullcone of representations of reductive groups. Pacif. J. Math. 224 (2006), 119–139. [KW] [La] G. Laumon, Sur la cat´egorie d´eriv´ee des D-modules filtr´es. Algebraic geometry (Tokyo/Kyoto, 1982), 151– 237, Lecture Notes in Math., 1016, Springer, Berlin, 1983. [LS1] T. Levasseur, J. Stafford, Invariant differential operators and a homomorphism of Harish-Chandra. J. Amer. Math. Soc. 8 (1995), 365–372. [LS2] [LS3] , 385–397. , , The kernel of an homomorphism of Harish-Chandra. Ann. Sci. ´Ecole Norm. Sup. 29 (1996), , Semi-simplicity of invariant holonomic systems on a reductive Lie algebra. Amer. J. Math. 119 (1997), 1095–1117. [Ma] H. Matsumura, Commutative algebra. Second edition. Mathematics Lecture Note Series, 56. Ben- jamin/Cummings Publishing Co., Inc., Reading, Mass., 1980. I. Macdonald, Some irreducible representations of Weyl groups. Bull. London Math. Soc. 4 (1972), 148-150. [Mc] [Na] H. Nakajima, Lectures on Hilbert schemes of points on surfaces, University Lecture Series, 18, Amer. Math. Soc., Providence, RI, 1999. [NS] M. Neubauer, D. Saltman, Two-generated commutative subalgebras of Mn(F ). J. Algebra 164 (1994), 545–562. [Po] V. Popov, Irregular and singular loci of commuting varieties. Transform. Groups 13 (2008), 819–837. [Pr] A. Premet, Nilpotent commuting varieties of reductive Lie algebras. Invent. Math. 154 (2003), 653–683. [Ri1] R. W. Richardson, Commuting varieties of semisimple Lie algebras and algebraic groups. Compositio Math. 38 (1979), 311–327. [Ri2] , Irreducible components of the nullcone. Invariant theory (Denton, TX, 1986), 409–434, Contemp. Math., 88, Amer. Math. Soc., Providence, RI, 1989. [Sa] M. Saito, Modules de Hodge polarisables. Publ. Res. Inst. Math. Sci. 24 (1988), 849–995 (1989). [So] [TV] B. Toen, G. Vezzosi, Brave new algebraic geometry and global derived moduli spaces of ring spectra. Elliptic L. Solomon, Invariants of finite reflection groups. Nagoya Math. J. 22, 57–64 (1963). cohomology, 325–359, London Math. Soc. Lecture Note Ser., 342, Cambridge Univ. Press, 2007. [SV] O. Schiffmann, E. Vasserot, Hall algebras of curves, commuting varieties and Langlands duality. Preprint 2010. arXiv:1009.0678. [Wa] N. Wallach, Invariant differential operators on a reductive Lie algebra and Weyl group representations. J. Amer. Math. Soc. 6 (1993), 779–816. DEPARTMENT OF MATHEMATICS, UNIVERSITY OF CHICAGO, CHICAGO, IL 60637, USA. E-mail address: [email protected] 59 

1806.03794

1

1806

Classification of Lipschitz simple function germs

It is known that the bi-Lipschitz right classification of function germs admit moduli. In this article we introduce a notion called the Lipschitz simple function germs and present a full classification in the complex case. A surprising consequence of our result is that a function germ is Lipschitz modal if and only if it deforms to the smooth unimodal family of singularities called $J_{10}$ in Arnold's list.

math.AG

math

CLASSIFICATION OF LIPSCHITZ SIMPLE FUNCTION GERMS NHAN NGUYEN, MARIA RUAS AND SAURABH TRIVEDI Abstract. It is known that the bi-Lipschitz right classification of func- tion germs admit moduli. In this article we introduce a notion called the Lipschitz simple function germs and present a full classification in the complex case. A surprising consequence of our result is that a function germ is Lipschitz modal if and only if it deforms to the smooth unimodal family of singularities called J10 in Arnold's list. 1. Introduction After Thom and Mather's foundational works on classification of ele- mentary catastrophes and singularities of mappings, Arnold [2] presented a complete classification of simple, unimodal and bimodal function germs under the right equivalence. The simple singularities include two series Ak, k ≥ 1, Dk, k ≥ 4 and three exceptional germs called E6, E7, E8, together called the ADE-singularities. Among the modal germs the one with the low- est codimension is T3,3,3 of corank 3 and codimension 8. For the corank 2 modal germs X9 with codimension 9 has the lowest codimension. In positive characteristics, the simple singularities were classified recently by Greuel– Nguyen [11] and unimodal and bimodal by Nguyen [17]. It is well known from a result of Mostowski [16] that the bi-Lipschitz right equivalence of complex analytic set germs does not admit moduli. This result was extended by Parusi´nski [20] for subanalytic sets and by Valette and the first author [18] for definable sets in polynomially bounded o- minimal structures. However, this is not true for function germs. Henry and Parusi´nski [12] showed that the bi-Lipschitz right classification of function germs does admit moduli. They constructed bi-Lipschitz invariants that vary continuously in the family fλ(x, y) = x3 − λ2xy4 + y6 of corank 2 function germs. This led us to ask whether there exists a classification for the bi-Lipschitz equivalence of function germs analogous to the smooth case. In this article we introduce the notion of Lipschitz simple function germs and give a complete classification of Lipschitz simple singularities. Roughly speaking, a finitely determined germ f is said to be Lipschitz simple if there is a neighbourhood around a sufficiently high k-jet of f that intersects only Date: June 12, 2018. 1 2 NHAN NGUYEN, MARIA RUAS AND SAURABH TRIVEDI finitely many bi-Lipschitz equivalence classes. A germ is said to be modal if it is not simple. Notice that a smooth simple germ is also Lipschitz simple and so the ADE-singularities are Lipschitz simple. We observe that a germ is Lipschitz modal if it deforms to a Lipschitz modal family. This observation provides us with a strategy to list all Lipschitz simple germs which we describe below. In Section 4 we prove that the rank and corank of function germs are Lipschitz invariants. We would like to emphasize that these seemingly simple results about Lipschitz equivalence were not known before and their proofs use some very recent work of Sampaio [23]. While bi-Lipschitz invariance of rank and corank allows us to distinguish between the bi-Lipschitz type of certain germs, there are several germs in Arnold's list with same corank and codimensions that, a priori, could be bi-Lipschitz equivalent. In Section 5, we prove that the multiplicity and the singular set of the algebraic tangent cone of the non-quadratic part of a germ obtained after applying the splitting lemma are also bi-Lipschitz invariants. This allows us to differentiate the bi-Lipschitz type of many germs. However, it is still not enough to write the Lipschitz normal forms as in Section 8. The only pair left in our analysis is (Q11, S11). This pair is shown to be of different bi-Lipschitz type by using a result of Bivi`a-Ausina and Fukui [3] on the bi-Lipschitz invariance of the log canonical threshold of ideals. We show in Section 6 that J10 : x3 + txy4 + y6 + Q where Q is a non- degenerate quadratic form, is Lipschitz modal. We would like to remark that the result does not immediately follow from Henry–Parusi´nski's [12] result. This result implies that any germ that deforms to J10 is Lipschitz modal. It is worth mentioning that the result follows trivially if there were a Lipschitz version of Splitting lemma. Unfortunately, this is not known. We observe later that J10 is the Lipschitz modal singularity of the smallest codimension. The main result is contained in Section 8 where we present a complete list of Lipschitz simple germs (Theorem 8.4 and Theorem 8.5). This uses the fact that most singularities in Arnold's list deform to J10 (see Lemma 8.2). For example, all smooth bimodal singularities deform to J10. Therefore, they are Lipschitz modal. The rest of the singularities are shown to be Lips- chitz simple by proving that they belong to a bi-Lipschitz trivial family (see Section 7) and are not adjacent to any class of Lipschitz modal singularities. The bi-Lipschitz triviality is proved by constructing vector fields satisfying the Thom-Levine criterion. Our method is inspired by and improves upon the works of Abderrahmane [1], Fernandes and Ruas [8], Saia et al. [6] and Humberto and Fernandes [9]. The non-adjacencies follow from a result of Brieskorn [4]. We would like to remark that in the paper we only consider the Lipschitz classification of complex simple germs. The real case turned out to be more CLASSIFICATION OF LIPSCHITZ SIMPLE FUNCTION GERMS 3 delicate and requires more careful analysis. For example, in the real case the family J10 is Lipschitz modal if t ≤ 0 (see Section 5) and is Lipschitz simple when t > 0 (see [13]), and this must be taken into account while checking deformations. Moreover, the Milnor number is not a bi-Lipschitz invariant in the real case; see example in Section 2. Throughout the paper, k.k denotes the Euclidean norm, and . denotes the absolute value of real numbers or the modulus of complex numbers. Let K = C or R, and let f, g : Kn ⊃ U → K be two functions. We write f . g (or g & f ) if there is a C > 0 such that f (x) ≤ Cg(x) for all x ∈ U . We write f ∼ g if f . g and g . f . Suppose that U is a neighbourhood of f (x) g(x) = 0. And, a point x0, we write f = o(g) or f ≪ g at x0 if limx→x0 f = O(g) or f & g at x0 if limx→x0 f (x) g(x) is bounded. 2. Preliminaries Let K = R or C. Given a smooth germ f : (Kn, 0) → (K, 0), we denote by Df (x) the derivative of f at the point x, by Vf the zero set of f , and by Σf = {x ∈ Kn : Df (x) = 0} the singular set of f . We write the Taylor expansion of f at 0 as T0f = fk + fk+1 + . . ., where fi is the homogeneous polynomial of degree i, fk 6= 0. We denote by Hf = fk the lowest degree homogeneous polynomial and by mf = k the multiplicity of f at 0. Given a semialgebraic set X ∈ Kn, for x ∈ X where X denotes the closure of X, the tangent cone of X at the point x is defined as follows C(X, x) = {λu : λ ∈ R≥0, u = lim m→∞ xm − x kxm − xk , X ⊃ {xm} → x}. We call Cg(f, 0) = C(Vf , 0) the geometric tangent cone of f . The algebraic tangent cone Ca(f, 0) of f at 0 is defined by Ca(f, 0) = {x ∈ Kn : Hf (x) = 0}. By the singular set of the algebraic tangent cone of f we mean the set ΣHf . For K = C, it is known that Ca(f, 0) = Cg(f, 0) (see Theorem 10.6 in Whitney [25] or, Proposition 2.6 in [14]). This is not true in the real case. For instance, if f (x, y) = x2 + y2k where k is an integer greater than or equal to 2, then Ca(f, 0) = {x = 0} and Cg(f, 0) = {(0, 0)}. Two smooth function germs f, g : (Kn, 0) → (K, 0) are said to be bi- Lipschitz right equivalent if there exists a germ of bi-Lipschitz homeomor- phism ϕ : (Kn, 0) → (Kn, 0) such that f ◦ ϕ = g. More precisely, there exist a neighbourhood U of 0 ∈ Kn, a homeomorphism ϕ : U → ϕ(U ) with ϕ(0) = 0 and 1 L kx − yk ≤ kϕ(x) − ϕ(y)k ≤ Lkx − yk (1) 4 NHAN NGUYEN, MARIA RUAS AND SAURABH TRIVEDI for any x, y ∈ U and for some L positive such that f ◦ ϕ = g. The least L satisfying (1) is said to be the bi-Lipschitz constant of ϕ. In the rest of the article by "equivalence" we mean the right equivalence. It follows from the main result of Sampaio [23] that if f and g are bi- Lipschitz equivalent then Cg(f, 0) and Cg(g, 0) are also bi-Lipschitz homeo- morphic as sets. We denote by En the set of all smooth function germs at 0 ∈ Kn and by mn the set of germs in En vanishing at 0. Notice that En is a local ring and mn is the maximal ideal of En. Denote by Rn the group of all smooth diffeomorphism germs H : (Kn, 0) → (Kn, 0). Then Rn acts on En by composition H.f = f ◦ H. Two germs f, g ∈ En are called smoothly equivalent, denoted by f ∼R g, if they lie in the same orbit of this action. Given a germ f ∈ En, denote by Jf the Jacobean ideal of f , i.e. the ideal in En generated by the partial derivatives of f . The codimension (also called the Milnor number of f ) of a germ f ∈ En is defined to be dimK En/Jf . It is well-known that in the complex case the Milnor number is a topological invariant (Milnor [15]), and is therefore a bi-Lipschitz invariant, i.e. if f, g ∈ En are bi-Lipschitz equivalent then their codimensions are equal. We would like to remark that the Milnor number is not a bi-Lipschitz invariant in the real case. Consider the family ft(x, y) = x4 + y4 + tx2y2 + y6, t ≥ 0. By Remark 7.9, this family is bi-Lipschitz trivial, however, µ(ft) = 9, t 6= 2 and µ(f2) = 13. Denote by J k(n, 1) the k-jet space of mn, by Rk n the set of k-jet space of Rn. A germ f ∈ mn is called k-determined if for any g ∈ mn such that jkg(0) = jkf (0) then f ∼R g; f is called finitely determined if f is k-determined for some k ∈ N. Recall that if a germ f : (Kn, 0) → (K, 0) is finitely determined then it has an isolated singularity at 0. In addition, if a germ f is finitely determined, then for a sufficiently large k there is a neighbourhood of jkf (0) in J k(n, 1) such that every germ in this neighbourhood is k-determined. We call the number k sufficiently large for f . The action of Rk n on J k(n, 1) is defined by taking the composition and then truncating the Taylor expansion. A germ f ∈ mn is said to be smooth simple if for k ∈ N sufficiently large for f , there exist a neighbourhood of jkf (0) in J k(n, 1) that meets only finitely many Rk n-orbits in J k(n, 1). The germs that are not smooth simple are called smooth modal. Two k-jets z, w ∈ J k 0 (n, 1) are bi-Lipschitz equivalent if there exists a bi-Lipschitz homeomorphism germ φ : (Kn, 0) → (Kn, 0) such that z ◦ φ = w. By Lipschitz orbits we mean the equivalence classes of the bi-Lipschitz equivalence. A germ f ∈ mn is Lipschitz simple if for k ∈ N sufficiently large for f , there is a neighbourhood of jkf (0) in J k(n, 1) that meets only finitely many Lipschitz orbits. The germs that are not Lipschitz simple are CLASSIFICATION OF LIPSCHITZ SIMPLE FUNCTION GERMS 5 called Lipschitz modal. We remark that if a germ is smoothly simple then it is also Lipschitz simple. The converse is not true in general. Given f ∈ En, the corank of f at 0 is defined to be the nullity (dimension the Splitting lemma; see Ebeling [7] or Gibson [10] for example. (0)(cid:17). The following result is known as of the kernel) of the Hessian (cid:16) ∂2f ∂xi∂xj Lemma 2.1. Let f ∈ m2 there exists g ∈ m3 c such that n be a finitely determined germ of corank c. Then f (x1, . . . , xn) ∼R g(x1, . . . , xc) ± x2 c+1 ± . . . ± x2 n. The codimension of f and g are equal. Moreover, g is uniquely determined up to smooth equivalence, i.e. if g + Q ∼R h + Q then g ∼R h, where Q is the non-degenerate quadratic form. 3. Arnold's results on classification In this section, we recall some results of Arnold about the classification of complex analytic germs that will be used in the later sections. It is known that the germs in mn of codimension 0 (or non-singular germs) are equivalent to the germ (x1 . . . , xn) 7→ x1, hence they are smooth simple. Arnold classified germs in m2 n with isolated singularity into classes of con- stant Milnor numbers, represented by a germ or family of germs called the 'normal form' of the class of singularities. More precisely, a class D has a normal form Ft(x1, . . . , xk), if D contains all germs smoothly equivalent to Ft + x2 n for some t. k+1 + · · · + x2 Recall that a smooth germ F : (Cn × Ck, 0) → (C, 0), F (x, t) = ft(x) is said to be a k-parameter deformation of a germ f : (Cn, 0) → (C, 0) if f0(x) = f (x). A class of singularities C is said to be adjacent to a class D, denoted C → D, if every f ∈ C can deform to D. That is, for every f ∈ C, there exists a smooth deformation ft of f such that f0 = f and ft belongs to D for every t near 0. The following results are due to Arnold [2]. Theorem 3.1. A germ f ∈ m2 to one of the germs in the following table: n is smooth simple if and only if it is equivalent Name Normal form Codimension Ak Dk E6 E7 E8 xk+1 x2y + yk−1 x3 + y4 x3 + xy3 x3 + y5 k ≥ 1 k ≥ 4 6 7 8 Theorem 3.2. The adjacencies of all smooth modal germs can be described by the diagrams below. The modal germs that do not appear in the diagram deform to one of the classes inside the brackets. 6 NHAN NGUYEN, MARIA RUAS AND SAURABH TRIVEDI Corank 2: J10 = (T2,3,6) ←− · · · X9 = T2,4,4 · · · / (T2,4,6) / T2,4,5 Z11 (Z12) · · · · · · / (T2,5,6) / T2,5,5 W12 (W13) · · · Corank 3: P8 = T3,3,3 · · · / (T3,3,6) / T3,3,5 / T3,3,4 Q10 Q11 (Q12) · · · · · · / (T3,4,6) · · · / (T3,5,6) / T3,5,5 ■ ■ ■ ■ ■ $■ / T3,4,5 / T3,4,4 S11 S12 (S1,0) · · · · · · / (T4,5,6) / T4,5,5 / T4,4,5 / T4,4,4 (U12) · · · · · · / (T5,5,6) / T5,5,5 (T4,4,6) (O) where the normal forms of the above singularities are ... ... 4a3 + 27 6= 0 a2 6= 4 a3 + 27 6= 0 a 6= 0, 1 p + 1 q + 1 r < 1 x3 + ax2y2 + y6 J10 x4 + y4 + ax2y2 X9 x3 + y3 + z3 + axyz, P8 Tp,q,r xp + yq + zr + axyz, Z11 Z12 W12 W13 Q10 Q11 Q12 S11 x3y + y5 + axy4 x3y + xy4 + ax2y3 x4 + y5 + ax2y3 x4 + xy4 + ay6 x3 + y4 + yz2 + axy3 x3 + y2z + xz3 + az5 x3 + y5 + yz2 + axy4 x4 + y2z + xz2 + ax3z / / O O o o o o o o / O O / O O O O o o O O o o o o / / / O O o o o o o o o o / $ / / / O O / O O o o O O o o O O o o o o / O O / O O / O O / O O o o O O o o / O O / O O O O O O O O O O CLASSIFICATION OF LIPSCHITZ SIMPLE FUNCTION GERMS 7 S12 U12 S1,0 O x2y + y2z + xz3 + az5 x3 + y3 + z4 + axyz2 x2z + yz2 + y5 + azy3 + bzy4 germs of corank ≥ 4 Notice that by changing x 7→ x + λy2, the normal form of J10 becomes x3 + axy4 + y6. This is exactly the family considered by Henry–Parusi´nski in [12]. For convenience, we will take x3 + axy4 + y6 as the normal form of J10. 4. Bi-Lipschitz invariance of rank and corank The aim of this section is to prove that the rank and corank of smooth germs are bi-Lipschitz invariants. We use a construction of a map associated to a bi-Lipschitz map ϕ due to Sampaio [23], which in some sense works as the derivative of ϕ. Let us give the details of the construction. Let ϕ : (Kn, 0) → (Kn, 0) be a bi-Lipschitz map germ with ψ as its inverse. By definition, there exist r, L ∈ R+ with L ≥ 1 such that for all x, y ∈ B(0, r), where B(0, r) denotes the closed ball of radius r centered at 0, we have: 1 L For m ∈ N, put kx − yk ≤ kϕ(x) − ϕ(y)k ≤ Lkx − yk. ϕm(x) = mϕ( x m ) and ψm(x) = mψ( x m ). It is easily checked that ϕm and ψm are a bi-Lipschitz germs of the same Lipschitz constant for any m. By the Arzela–Ascoli theorem there exists a subsequence {mi} ⊂ N such that {ϕmi } and {ψmi } converges uniformly to a bi-Lipschitz germs dϕ and dψ respectively. It was proved in [23] that dψ is the inverse of dϕ. Notice that the domains of ϕmi and ψmi grow bigger as mj increases and the limits dϕ and dψ are well-defined on the whole of Kn. We first prove the bi-Lipschitz invariance of the rank. Theorem 4.1. Let f, g : (Kn, 0) → (Kp, 0) be two smooth map germs. If f and g are bi-Lipschitz equivalent then the rank of f is equal to the rank of g at 0. Proof. Since f and g are bi-Lipschitz equivalent, there exists a germ of a bi-Lipschitz homeomorphism ϕ : (Kn, 0) → (Kn, 0) such that f ◦ ϕ = g. Then, there exists a sequence {mi} in N such that the sequence of maps ) converges uniformly ϕmi to a bi-Lipschitz germ dϕ on a neighbourhood of 0 as i tends to ∞. Write : (Kn, 0) → (Kn, 0) defined by ϕmi = miϕ( x mi 8 NHAN NGUYEN, MARIA RUAS AND SAURABH TRIVEDI f (x) = Df (0)(x)+o(kxk) and g(y) = Dg(0)(y)+o(kyk) on a neighbourhood of 0. Since f ◦ ϕ(y) = g(y) =⇒ mi(Df (0)(ϕ( )) + miokϕ( )k = miDg(0)( y mi mif ◦ ϕ( ) = mig( y mi ) y mi y mi y mi )) + lim i→∞ miokϕ( )k = Dg(0)(y) + lim i→∞ miok y mi ) + miok y mi y mi k k. =⇒ lim i→∞ Df (0)(miϕ( y mi Hence, Df (0)(dφ(y)) = Dg(0)(y) for y in a small neighbourhood of 0. This implies that the rank of f and g are equal. (cid:3) In the rest of the section we fix f, g : (Kn, 0) → (K, 0) to be bi-Lipschitz equivalent smooth function germs, and fix ϕ : (Kn, 0) → (Kn, 0) to be a germ of a bi-Lipschitz homeomorphism such that f ◦ ϕ = g. To show that the corank of finitely determined smooth germs is also a bi-Lipschitz invariant, we need the following lemmas. It has been known that the multiplicity is a bi-Lipschitz invariant (see [21], [8]). We present another proof below. Lemma 4.2. The multiplicities mf and mg of f and g are equal. Proof. Write ϕ = (ϕ1, . . . , ϕn) where ϕi : (Kn, 0) → (K, 0). Assume that mf = k and mg = l. In a neighbourhood of 0 we can write f (x) = Hf + o(kxkk) and g(x) = Hg + o(kxkl). By a change of coordinates, we may assume that Hf (1, 0, . . . , 0) 6= 0 and Hg(1, 0, . . . , 0) 6= 0. Then, we can write g in two ways: g(x) = xl 1 + . . . + o(kxkl) and g(x) = f (ϕ(x)) = ϕk 1(x) + . . . + o(kxkk). Since ϕ is bi-Lipschitz, each ϕi, i = 1, . . . , n is Lipschitz. Thus, kϕk 1(x) + . . . + o(kxkk)k . kxkk. Restricting to the x1-axis, we have 1 = lim x→0 kϕk 1(x) + . . . + o(kxkk)k kxl 1 + . . . + o(kxkl)k . lim x1→0 x1k x1l = lim x1→0 x1k−l. This implies that l ≥ k. Therefore, k = l. Interchanging f with g, we have also l ≤ k. (cid:3) Lemma 4.3. The lowest degree homogeneous parts Hf and Hg of f and g are bi-Lipschitz equivalent. mkhHf (ϕ( x m )) + o(kϕ( x m ). ))i = mkg( )kk)i = mkhHg( x m ) + o(cid:16)k x m kk(cid:17)i CLASSIFICATION OF LIPSCHITZ SIMPLE FUNCTION GERMS 9 Proof. We can write f and g at 0 as f (x) = Hf (x) + o(kxkk) and g(x) = Hg(x) + o(kxkk) (2) Then, for x ∈ Kn and m ∈ N big enough we have x m x m mkhf (ϕ( Substituting the expansions of f and g in the equation (2), we have Since the multiplicity is a bi-Lipschitz invariant, Hf and Hg are homoge- neous polynomials of the same degree k = mf = mg. Thus, Hf (mϕ( x m )) + mko(kϕ( x m )kk) = Hg(x) + mko(cid:16)k x m kk(cid:17) . Then, by the Arzela–Ascoli theorem, there exists a sequence {mi} of pos- itive integers such that as i tends to ∞, the sequence of maps {miϕ( x )} mi converges to a bi-Lipschitz map germ dϕ. By taking limits on both sides of the above equation as mi tends to ∞, we get Hf (dϕ(x)) = Hg(x). This completes the proof. (cid:3) Remark 4.4. (i) Consider the sequence {mi} associated with the map ϕ as in the proof of Lemma 4.3. Let f : Kn, 0 → K, 0 be a smooth function germ with the multiplicity mf = k. We always have lim i→∞ mk i f (ϕ( x mi )) → Hf (dϕ(x)). 0(g) for l > k, where T l (ii) The result in Lemma 4.3 is sharp. In general, T l 0(f ) is not bi-Lipschitz equivalent to T l 0(f ) denotes the Taylor expansion of f at 0 up to degree l. For example, consider f (x, y) = x3 + y6 and g(x, y) = (x + y2)3 + y6. It is easy to see that f and g are smooth equivalent 0 (f ) = x3 under the change of coordinates ϕ(x, y) = (x + y2, y). However, T 4 is not bi-Lipschitz equivalent to T 4 0 (g) = x3 + 3x2y2. Lemma 4.5. Let L be the bi-Lipschitz constant of ϕ. Then, L−1kDf (ϕ(x))k ≤ kDg(x))k ≤ LkDf (ϕ(x))k. Proof. The proof follows directly from the definition of the derivative. First, Dg(x) = lim ∆x→0 g(x + ∆x) − g(x) k∆xk = lim ∆x→0 = lim ∆x→0 f ◦ ϕ(x + ∆x) − f ◦ ϕ(x) k∆xk f (ϕ(x) + ϕ(x + ∆x) − ϕ(x)) − f (ϕ(x)) kϕ(x + ∆x) − ϕ(x)k kϕ(x + ∆x) − ϕ(x)k k∆xk 10 NHAN NGUYEN, MARIA RUAS AND SAURABH TRIVEDI Since ∆x → 0 implies that ϕ(x + ∆x) − ϕ(x) → 0, we have lim ∆x→0 f (ϕ(x) + ϕ(x + ∆x) − ϕ(x)) − f (ϕ(x)) kϕ(x + ∆x) − ϕ(x)k = Df (ϕ(x)). On the other hand, Therefore, L−1 ≤(cid:13)(cid:13)(cid:13)(cid:13) ϕ(x + ∆x) − ϕ(x) ∆x ≤ L. (cid:13)(cid:13)(cid:13)(cid:13) L−1kDf (ϕ(x))k ≤ kDg(x)k ≤ LkDf (ϕ(x))k. Lemma 4.6. Σf and Σg are bi-Lipschitz equivalent. Proof. By Lemma 4.5, we have kDf (ϕ(x))k ∼ kDg(x)k and kDf (x)k ∼ kDg(ϕ−1(x))k. This shows that ϕ maps Σg onto the Σf . Theorem 4.7. The corank of f is equal to the corank of g at 0. (cid:3) (cid:3) Proof. If the corank of f is n, then mf ≥ 3. We know from Lemma 4.2 that the multiplicity is a bi-Lipschitz invariant, so mg ≥ 3. This implies that the corank of g is n. Now we assume that the corank of f is r and that of g is k with (0 ≤ r < k < n). By the Splitting lemma, f ∼R Rf (x1, . . . , xr) + Qf and g ∼R Rg(x1, . . . , xk) + Qg where Qf and Qg are quadratic forms in further variables. By Lemma 4.3, Qf and Qg are bi-Lipschitz equivalent. Then, their sin- gular sets ΣQf and ΣQg are bi-Lipschitz equivalent (see Lemma 4.6). Notice that ΣQf = Kr × {0} and ΣQg = Kl × {0}. It is known that the dimension of a semialgebraic set is a bi-Lipschitz invariant, therefore r must be equal to k. This completes the proof. (cid:3) 5. Bi-Lipschitz type of the non-quadratic parts Given a finitely determined germ F : (Kn × Kp, 0) → (K, 0) of corank n, we know by the splitting lemma that F (x, y) ∼R f (x) + Qf (y) where Q is a quadratic form in p variables. Moreover, this splitting of F is unique in the sense that if F ∼R g + Qg then f ∼R g. In this section we prove a weaker version of this result for bi-Lipschitz equivalence. Consider two smooth germs F, G : (Kn × Kp, 0) → (K, 0) of the following forms: F (x, y) = f (x) + Qf (y) and G(x, y) = g(x) + Qg(y) where f, g : (Kn, 0) → (K, 0) are smooth germs of multiplicities ≥ 3 and Qf , Qg are of the form ±y2 p. We prove the following result. 1 ± . . . ± y2 CLASSIFICATION OF LIPSCHITZ SIMPLE FUNCTION GERMS 11 Theorem 5.1. If F and G are bi-Lipschitz equivalent, then (i) mf = mg; (ii) ΣHf and ΣHg are bi-Lipschitz equivalent. Proof. Let ϕ = (ϕ1, ϕ2) : (Kn × Kp, 0) → (Kn × Kp, 0) be a germ of a bi- Lipschitz homeomorphism such that F ◦ ϕ = G. Let ψ = (ψ1, ψ2) be the inverse of ϕ. For m ∈ N, set ϕm = (ϕ1,m, ϕ2,m) = mϕ(cid:18) x ψm = (ψ1,m, ϕ2,m) = mψ(cid:18) x m m y m(cid:19) m(cid:19). y , , and We know that there exists a subsequence {mi} ⊂ N such that ϕmi and ψmi uniformly converge to bi-Lipschitz homeomorphisms which we denote by dϕ and dψ respectively. Moreover, dψ is the inverse of dϕ. In fact, dϕ = (dϕ1, dϕ2) and dψ = (dψ1, dψ2) where dϕ1 (reps. dϕ2, dψ1, dψ2) is the limit of the sequence {ϕ1,mi} (resp. {ϕ2,mi}, {ψ1,mi }, {ψ2,mi}). By Lemma 4.3 and Lemma 4.6, we have dϕ(ΣQg ) = ΣQf . Since ΣQg = ΣQf = Kn × {0}, dϕ(x, 0) = (dϕ1(x, 0), 0). The map dϕ is a bi-Lipschitz homeomorphism, so is the map dϕ1(., 0) : Kn, 0 → Kn, 0. It is easy to check that the inverse of dϕ1(., 0) is the map dψ1(., 0). By Lemma 4.5, for all (x, y) ∈ Kn × Kp near 0 we have and This yields that and kDF (ϕ(x, y))k . kDG(x, y)k kDG(ψ(x, y))k . kDF (x, y)k. kDF (ϕ(x, 0))k . kDG(x, 0)k, kDG(ψ(x, 0))k . kDF (x, 0)k. (3) (4) Notice that DF (ϕ(x, 0) = (Df (ϕ1(x, 0)), 2ϕ2(x, 0)) and DG(x, 0) = Dg(x). The inequality (3) implies Similarly, (4) implies kDf (ϕ1(x, 0))k . kDg(x)k. kDg(ψ1(x, 0))k . kDf (x)k. (5) (6) 12 NHAN NGUYEN, MARIA RUAS AND SAURABH TRIVEDI We now prove (i). First we prove that mg ≤ mf . Let us consider the subsequence {mi} mentioned at the beginning of the proof. It follows from (5) that kmmg −1 i Df (ϕ1( , 0))k . kmmg−1 i Dg( x mi x mi )k. (7) The right-hand side of (7) tends to DHg(x) as mi → ∞ (see Remark 4.4, (i)), therefore, it is bounded when x is bounded. Recall that Hg is the lowest degree homogeneous polynomial of g. Let x∗ be a point near 0 such that DHf (dϕ1(x∗, 0)) 6= 0. If mg > mf , then kmmg −1 , 0))k tends to infinity as mi → ∞, which violates (5). Thus, mg ≤ mf . By the same arguments applied to (6), we get mg ≥ mf . Df (ϕ1( x∗ mi i To prove (ii), it is enough to show that dϕ1(., 0)(ΣHf ) = Hg, or equiva- lently, dϕ1(., 0)(ΣHg ) ⊂ ΣHf , and dψ1(., 0)(ΣHf ) ⊂ ΣHg. Since mf = mg, the left-hand side of (7) tends to kDHf (dϕ1(x, 0))k as mi → ∞. Thus, kDHf (dϕ1(x, 0))k . kDHg(x)k. This shows that dϕ1(., 0) maps ΣHg into ΣHf . Analogously, dψ1(., 0) maps ΣHf into ΣHg . Therefore, ΣHf and ΣHg are bi-Lipschitz equivalent. (cid:3) Example 5.2. Consider the families J10 and T2,4,5 as in Arnold's list (see Section 3). Recall that germs in J10 have the form x3 + txy4 + y6 + Q and germs in T2,4,5 have the form x4 +y5 +tx2y2 +Q where Q is a quadratic form in further variables. Put f (x, y) = x3+txy4+y6 and g(x, y) = x4+y5+tx2y2. It is obvious that mf 6= mg. By Theorem 5.1, germs in J10 and germs in T2,4,5 have different bi-Lipschitz types. 6. Lipschitz modality of J10 Henry and Parusi´nski [12] proved the the non-quadratic part of J10 which is x3 + txy4 + y6 is Lipschitz modal. However, this does not imply that J10 is Lipschitz modal. Following the idea of Henry–Parusi´nski [12], [13] we prove in this section that J10 is also Lipschitz modal. We fix K = C or R and consider two bi-Lipschitz equivalent smooth function germs f, g : Kn, 0 → K, 0. Assume that ϕ : Kn, 0 → Kn, 0 is a germ of a bi-Lipschitz homeomorphism such that f ◦ ϕ = g. Let L be the bi-Lipschitz constant of h. It is shown in Lemma 4.5 that L−1kDf (ϕ(x))k ≤ kDg(x))k ≤ LkDf (ϕ(x))k. (8) Given δ > 1, we define V δ(f ) = {x : f (x) 6= 0, δ−1kxkkDf (x)k ≤ kf (x)k ≤ δkxkkDf (x)k}. It follows from (8) that V L−2δ(f ) ⊂ ϕ(V δ(g)) ⊂ V L2δ(f ). (9) CLASSIFICATION OF LIPSCHITZ SIMPLE FUNCTION GERMS 13 Given σ > 0, d > 0, we define W σ(f, d) = {x ∈ Kn : f (x) ≤ σkxkd}. It is obvious that W σL−d (f, d) ⊂ ϕ(W σ(g, d)) ⊂ W σLd (f, d). (10) Denote by Ωδ σ,d(f ) = V δ(f ) ∩ W σ(f, d). It follows from (9) and (10) that ΩδL−2 σL−d,d(f ) ⊂ ϕ(Ωδ σ,d(g)) ⊂ ΩδL2 σLd,d(f ). (11) Now we consider J10 that consists of germs of the following form fλ(u, v, w) = u3 − 3λ2uv4 + v6 + w2 where (u, v, w) ∈ K2 × Kp, w = (w1, . . . , wp) and w2 := w2 1 + . . . + w2 p. Fix ξ = fλ with 1 ± 2λ3 6= 0. Given δ, σ > 0. Notice that Ωδ semialgebraic set. Suppose that the germ at 0 of Ωδ Then, we have the following results. σ,6(ξ) is a σ,6(ξ) has dimension ≥ 1. Lemma 6.1. The tangent cone at 0 of Ωδ σ,6(ξ) is contained in the v-axis. σ,6(ξ) at 0. There exists a real analytic curve γ : [0, ε) → Ωδ Proof. Let µ = (a, b, c) ∈ K2+p be a unit vector of the tangent cone of Ωδ σ,6(ξ) such γ(t) kγ(t)k . Write γ(t) = (γu(t), γv(t), γw(t)). that γ(0) = 0 and µ = limt→0 Reparametrizing γ if necessary, we may assume that kγ(t)k ∼ t. It suffices to show that a = c = 0. Indeed, if c 6= 0, we have γu(t) ∼ tα, γv(t) ∼ tβ and γw(t) ∼ t where α, β ≥ 1 (put t∞ = 0 as a convention). This implies ξ(γ(t)) ∼ t2. Hence γ(t) 6⊂ W σ(ξ, 6). If c = 0, a 6= 0, we have γu(t) ∼ t, γv(t) ∼ tα and γw(t) ∼ tβ where α ≥ 1 and β > 1. Recall that Dξ(u, v, w) = (3u2 − 3λ2v4, −12λ2uv3 + 6v5, 2w). There are two possibilities • if 2β ≤ 3 then g(γ(t)) . t2β and Dξ(γ(t)) ∼ tβ. Thus, kγ(t)kkDξ(γ(t))k ∼ tβ+1 ≫ t2β ∼ ξ(γ(t)) (since β > 1). This implies γ 6⊂ V δ(ξ). • if 2β > 3 then ξ(γ(t)) ∼ t3. Hence, γ(t) 6⊂ W σ(ξ, 6). All the cases above give contradiction, therefore a and c must be equal to 0. (cid:3) Lemma 6.2. On the set Ωδ σ,6(ξ), we have (1) u = ±v2 + O(v3), (2) g(u, v, w) = (1 ∓ 2λ3)v6 + O(v7). 14 NHAN NGUYEN, MARIA RUAS AND SAURABH TRIVEDI Proof. We have ∂ξ/∂u = 3(u2−λ2v4), ∂ξ/∂v = 6v3(v2−2λ2u), ∂ξ/∂w = 2w. σ,6(ξ), then kDξ(x)kkxk ∼ ξ(x) and ξ(x) . kxk6. Let x = (u, v, w) ∈ Ωδ By Lemma 6.1, we have kxk ∼ v. This implies that kDξ(x)k . kxk5 ∼ v5. Therefore, 3(u2 − λ2v4) . v5, 6v3(v2 − 2λ2u) . v5 2w . v5   This implies that u = ±λv2 + O(v3), w = O(v5), and hence we have formula (1). The second formula then follows directly from (1). (cid:3) Lemma 6.3. Let {ak}, {bk} be two sequences of points in Ωδ to 0 such that kak − bkk . kakk1+ε for some ε > 0. Then, σ,6(ξ) converging lim k→∞ ξ(ak) ξ(bk) = c, where c is one of the following values {1, 1+2λ3 k, v′ g(ak) = (1 ∓ 2λ3)v6 1+2λ3 }. 1−2λ3 , 1−2λ3 k, w′ k + O(v7 k) Proof. Let ak = (uk, vk, wk) and b = (u′ k). By Lemma 6.2, and g(bk) = (1 ∓ 2λ3)v′ k 6 + O(v′ k 7). Since ak ∈ Ωδ vk − v′ The result then follows. σ,6(ξ), by Lemma 6.1, ak ∼ vk. Moreover, k ≤ ak − bk . ak1+ε ∼ vk1+ε. (cid:3) We now prove the main result of the section. Theorem 6.4. Assume that 1 ± 2λ3 6= 0 and 1 ± 2λ′3 6= 0. If fλ and fλ′ are bi-Lipschitz equivalent then λ3 = ±λ′3. Proof. For 1 ± 2λ3 6= 0, we consider the polar curve of fλ Γλ = {fλ/∂u = ∂fλ/∂w = 0} = {u = ±λv2, w = 0}. The restrictions of fλ and Dfλ to Γλ are and fλΓλ = (1 ± 2λ3)v6 DfλΓλ = (0, ∓12λ3v5 + 6v5, 0). We observe that they satisfy the following conditions: (i) fλ(x) . kxk6; (ii) kxkkDf (x)k ∼ fλ(x). Therefore, for a given L ≥ 1 we can choose δ, σ > 0 big enough such that the set σL−d,6(fλ) = V δL−2 ΩδL−2 (fλ) ∩ W σL−6 (fλ, 6) contains the polar curve Γλ. CLASSIFICATION OF LIPSCHITZ SIMPLE FUNCTION GERMS 15 Now, since fλ and fλ′ are bi-Lipschitz equivalent, there is a germ of a bi- Lipschitz homeomorphism h : (K2+p, 0) → (K2+p, 0) such that fλ ◦ h = fλ′. We assume that L ≥ 1 is the bi-Lipschitz constant of h. Take d = 6. By (11), we have ΩδL−2 σL−6,6(fλ) ⊂ h(Ωδ σ,6(fλ′)) ⊂ ΩδL2 σL6,6(fλ). (12) Choose δ, σ > 0 large enough such that ΩδL−2 Γλ and ΩδL−2 σL−6,6(fλ′) contains the polar curve Γλ′. σL−6,6(fλ) contains the polar curve The polar curve Γλ ⊂ ΩδL−2 σL−6,6(fλ) has two branches γ1(v) = (λv2, v, 0) and γ2(v) = (−λv2, v, 0). The restriction of fλ to these branches are fλ(γ1(v)) = (1 − 2λ3)v6 and fλ(γ2(v)) = (1 + 2λ3)v6. Thus, lim v→0 fλ(γ1(v)) fλ(γ2(v)) = 1 − 2λ3 1 + 2λ3 . k, vk, 0) ∈ γ2 where limk→∞ vk = Take xk = (λv2 0. Obviously, k, vk, 0) ∈ γ1 and x′ k = (−λv2 lim k→∞ fλ(xk) fλ(x′ k) = 1 − 2λ3 1 + 2λ3 . k = h−1(x′ Put xk = h−1(xk) and x′ Since xk − x′ a subsequence if necessary we may assume limk→∞ the limit is one of the following values {1, 1−2λ′3 k . xk2 and h−1 is bi-Lipschitz, xk − x′ fλ′ (xk) fλ′ (x′ 1+2λ′3 , 1+2λ′3 σ,6(fλ′). k . xk2. Taking k) exists, and hence 1−2λ′3 } (see Lemma 6.3). k). By (12), xk and x′ k belong to Ωδ On the other hand, lim k→∞ fλ′(xk) fλ′(x′ k) 1+2λ′3 , 1+2λ′3 1+2λ3 ∈ {1, 1−2λ′3 = lim k→∞ fλ(xk) fλ(x′ k) = 1 − 2λ3 1 + 2λ3 . This implies 1−2λ3 we have 1−2λ′3 1+2λ′3 ∈ {1, 1−2λ3 1+2λ3 , 1+2λ3 1−2λ′3 }. The same arguments applied to Γλ′ (cid:3) 1−2λ3 }. Therefore, λ3 = ±λ′3. 7. Bi-Lipschitz determinacy of smooth function germs This section is devoted to proving that all families of singularities in The- orem 3.2 which are not enclosed in brackets are bi-Lipschitz trivial. Our idea is to prove the existence of vector fields that satisfy the Thom–Levine criterion (see Theorem 7.2). 7.1. A criterion for Lipschitz functions. We recall a classical criterion to verify if a given function is Lipschitz. Lemma 7.1. Let U be a convex open subset of Rn and f : U → R be a continuous semialgebraic function. If Df (x) exists and kDf (x)k ≤ M except finitely many points in U , then f is Lipschitz with the Lipschitz constant also bounded by M . 16 NHAN NGUYEN, MARIA RUAS AND SAURABH TRIVEDI Proof. We first prove the statement for n = 1. Without loss of generality we can assume that U = (a, b) is an open interval. Then, Df (x) is C 1 everywhere except finitely many points a1 < . . . < ak. Let x1, x2 ∈ (a, b). Assume that x1 ≤ ai < ai+1 < . . . < al ≤ x2. We have f (x1) − f (x2) ≤ Z ai ≤ M Z x2 x1 x1 This implies that f is Lipschitz. Df (x)dx +Z ai+2 ai+1 Df (x) + . . .Z x2 al Df (x)dx dx ≤ M x1 − x2. For n > 1 fix x, y ∈ U . Set v = y−x kx−yk . Define g : U ⊃ [0, kx − yk] → R by g(t) = f (x + tv) . Notice that g is a semialgebraic function and Dg(t) = Dvf (x + tv) when- ever Dv(f, x + tv exists, where Dvf is the directional derivative of f in the direction v. Since kvk = 1, Dvf (.) ≤ kDf (.)k. From the hypothesis that kDf (.)k ≤ M is bounded except finitely many points, so is Dg(t). Then by the arguments as in the case n = 1, g is a Lipschitz function with the Lipschitz constant bounded by M . Therefore, g(0) − g(kx − yk) ≤ M kx − yk f (x) − f (y) ≤ M kx − yk. This implies that f is Lipschitz. (cid:3) The Thom–Levine criterion for bi-Lipschitz triviality is as follows: Theorem 7.2. Let K = C or R and F : (Kn × K, 0) → (K, 0) be a one- parameter deformation of a germ f : (Kn, 0) → (K, 0). If there exists a germ of a continuous vector field of the form n X(x, t) = + Xi(x, t) ∂ ∂xi ∂ ∂t Xi=1 Lipschitz in x, (i.e. there exists a number C > 0 with kX(x1, t) − X(x2, t)k ≤ Ckx1 − x2 for all t), such that X.F = 0, then F is a bi-Lipschitz trivial deformation of f . Proof. The existence of the flow of the vector field X(x, t), denoted Φ(x, t), follows from the classical Picard–Lindelof theorem. Applying the Gronwall lemma, we can show that Φt : Kn, 0 → Kn, 0 is a bi-Lipschitz homeomor- phism. Therefore, this flow induces the bi-Lipschitz triviality of F . (cid:3) It seems to be unknown whether the converse of the above statement holds. For this reason we call the deformation F in Theorem 7.2 strongly bi-Lipschitz trivial. Fernandes and Ruas [8] gave sufficient conditions for CLASSIFICATION OF LIPSCHITZ SIMPLE FUNCTION GERMS 17 a deformation of a quasihomogeneous polynomial (in the real case) to be strongly bi-Lipschitz trivial. In this section we first establish a similar result for the complex case. From now on we always work with complex function germs. of A is defined as the convex hull of the set Sα∈A(α + Rn 7.2. Bi-Lipschitz determinacy. Let A ⊂ Qn ≥0 be a finite set of points. For r ∈ Q+, denote by rA = {rα : α ∈ A}. The Newton polyhedron Γ+(A) ≥0). The union of compact faces of Γ+(A), denoted Γ(A), is called the Newton diagram of A. We say that A is a Newton set if the intersection of Γ+(A) with each coordinate axis is non-empty and A coincides with the set of vertices of Γ+(A). Given a Newton set A, the control function associated with A is defined as follows ρA(x) = Xα=(α1,...,αn)∈A x1α1 . . . xnαn. We denote by ¯x = (¯x1, . . . , ¯xn) the complex conjugate of x. Consider a polynomial f in the variables (x, ¯x) given by f (x, ¯x) =Xν,µ cν,µxν ¯xµ 1 . . . xνn n for ν = (ν1, . . . , νn) ∈ Nn and ¯xµ = ¯xµ1 where xν = xν1 n for µ = (µ1, . . . , µn) ∈ Nn. Such a polynomial is called a mixed polynomial of x (see also [19]). The support of f is supp(f ) = {(ν1 + µ1, . . . , νn + µn) : cν,µ 6= 0}. The Newton polyhedron and the Newton diagram of supp(f ), denoted Γ(f ) and Γ+(f ) respectively, are called the Newton polyhedron and Newton diagram of f . 1 . . . ¯xµn For w = (w1, . . . , wn) ∈ Qn +, the filtration (with respect to w) of a mixed monomial M = xν ¯xµ is defined by f ilw(M ) = wj(νj + µj). n Xj=1 The filtration of a mixed polynomial f , denoted f ilw(f ) (or f il(f ) if w is clear from the context), is the minimum of the filtrations of the mixed monomials appearing in f . Let w = (w1, . . . , wn) ∈ Qn + and d ∈ Q+. A polynomial h(z) is called quasihomogeneous of the type (w1, . . . , wn; d) if h(λw1x1, . . . , λwnxn) = λdh(x), ∀λ ∈ C∗. A polynomial g =Pα aαxα can always be written as where gj = Pf ilw(xα)=j aαxα. We call gd the initial part (w.r.t. the weight g = gd + gd+1 + . . . w) of g. 18 if NHAN NGUYEN, MARIA RUAS AND SAURABH TRIVEDI A mixed polynomial f is called quasihomogeneous of the type (w1, . . . , wn; d) f (λw1x1, . . . , λwnxn, ¯λw1 ¯x1, . . . , ¯λwn ¯xn) = λdf (x, ¯x), ∀λ ∈ C∗. Newton set A. We say that 0 is a Γ+(A)-isolated point of f if for all compact faces σ of Γ+(A), f (x)σ = 0 has no solution in (C∗)n where f (x)σ = Suppose that f (x, ¯x) = Pν,µ cν,µxν ¯xµ and supp(f ) ⊂ Γ+(A) for some Pν+µ∈σ cν,µxν ¯xµ is the restriction of f to the compact face σ. From now on we assume that A is a Newton set. Lemma 7.3. Let f be a mixed polynomial with supp(f ) ⊂ Γ+(A). Then, there exists a constant C > 0 such that in a neighbourhood of the origin In particular, if supp(f ) lies entirely in the interior of Γ+(A) then ρA(x) ≥ Cf (x). lim x→0 f (x) ρA(x) = 0. Proof. For simplicity we may assume f is a mixed monomial xν ¯xµ. Put u = ν + µ. We have u ∈ Γ+(A) and xν ¯xµ = xu = x1u1 . . . xnun. Now assume on the contrary that the germ at (0, 0) of the set X = {(x, c) ∈ Cn × R+ : ρ(x) < cxu} is non-empty. Notice that X is a semialgebraic germ of dimension ≥ 1. Since (0, 0) ∈ X, there exists a real analytic curve γ(t) = (γ1(t), . . . , γn(t), γn+1(t)) : [0, ε) → X, with γ(0) = 0, γ1(t) ∼ ta1, . . . , γn(t) ∼ tan, γn+1(t) ∼ tb and where γ(t) := (γ1(t), . . . , γn(t)). This implies that ρA(γ(t)) . tbγ(t)u, ta1α1 . . . tαnαn . tbta1u1 . . . tanun. Xα∈A For t small enough, we have tha,αi < tha,ui. Xα∈A Hence, ha, ui < inf α∈Aha, αi. Since ha, ui is a linear function on Γ+(A), it attains a minimum at one of the vertices of Γ+(A). So ha, ui = inf α∈Aha, αi. This gives a contradiction. If u lies in the interior of Γ+(A), then there is a δ = (δ1, . . . , δn) ∈ Qn + such that u − δ still lies in Γ+(A). This implies, xu ρA(x) = xu−δxδ ρA(x) ≤ Cxδ which tends to 0 when x tends to 0. (cid:3) and ρA(γ(t)) =Xα∈σ thα,βi + o(tm), CLASSIFICATION OF LIPSCHITZ SIMPLE FUNCTION GERMS 19 Lemma 7.4. Let f be a mixed polynomial such that 0 is a Γ+(A)-isolated point of f . Then, there is a C > 0 such that in a neighbourhood of the origin f (x) ≥ CρA(x). Proof. Suppose that f (x) = Pν,µ cν,µxν ¯xµ. Assume on the contrary that [0, ε) → Cn, γ(0) = 0 such that ρA(γ(t)) = 0. Write γ = (γ1, . . . , γn) where γj(t) = (bjtβj +o(tβj ))eiθj (t), there exists a real analytic curve γ : limt→0 j = 1, . . . , n. Set b = (b1, . . . , bn), β = (β1, . . . , βn) and θ = (θ1, . . . , θn). f (γ(t)) Consider the linear function h(α) = hα, βi defined on Γ+(A). Then, h(α) attains its minimum value m at a compact face, say σ, of Γ+(A). We can write f (γ(t)) = Xν+µ∈supp(f )∩σ cν,µbν+µthα,βieihθ,ν−µi) + o(tm). f (γ(t)) Since limt→0 ρA(t) = 0,Pν+µ∈supp(f )∩σ cν,µbν+µthα,βieihθ,ν−µi) = 0. This implies that γ∗(t) := (b1tβ1eiθ1, . . . , bntβneiθn) is a solution of f σ(x). Since bj > 0 for all j = 1, . . . , n, γ∗(t) is contained in (C∗)n. This contradicts the hypothesis that 0 is a Γ+(A)-isolated point of f . (cid:3) Denote by KA the set of (n − 1)-dimensional compact faces of Γ+(A). Let σ ∈ KA. Then, there is d ∈ Q+ such that σ belongs to some plane ∆σ given by wσ,1x1+. . .+wσ,nxn = d, where wσ,i ∈ Q≥0. We call wσ = (wσ,1, . . . , wσ,n) the weight corresponding to σ with total weight d. Notice that wσ,i and d can be chosen to be integers. The following result is a consequence of Lemma 7.3. Lemma 7.5. Let B = rA, where r ∈ Q+. If minσ∈KA{f ilwσ (f )} ≥ rd then In particular, if minσ∈KA{f ilwσ } > rd then f (x) . ρB(x). lim x→0 f (x) ρB(x) = 0. Let (w1, . . . , wn; 2k) be integers. We call ρk(x) = x12α1 + . . . + xn2αn where αi = k/wi the standard control function of type (w1, . . . , wn; 2k). Lemma 7.6. Let f be a mixed polynomial and {ft}t∈U be a smooth defor- mation of f . Suppose that ft is a quasihomogenous mixed polynomial of type (w1, . . . , wn; 2k) for every t ∈ U and f −1 (0) = {0}. If U is compact then there are 0 < C1 < C2 independent of t such that t C1ρk(x) ≤ ft(x) ≤ C2ρk(x). 20 NHAN NGUYEN, MARIA RUAS AND SAURABH TRIVEDI Proof. Let A be the set of vertices of Γ+(ρk). It is obvious that Γ+(A) = Γ+(ρk) and ρk(x) = ρA(x). Since supp(f ) ⊂ Γ+(A), there is C2(t) > 0 such that ft ≤ C2(t)ρA (see Lemma 7.3). Since f −1 (0) = 0, 0 is a Γ+(A)- isolated point of ft. Thus, there is exists C1(t) > 0 such that ft ≥ C1(t)ρA (see Lemma 7.4). By the continuity in t of ft, C1(t) and C2(t) can be chosen to be continuous in t. Then, C1 = mint∈U C1(t) and C2 = maxt∈U C2(t) the desired inequality. (cid:3) t Lemma 7.7. Let ρk be the standard control function of the type (w1, . . . , wn; 2k) with w1 ≤ . . . ≤ wn. Let U ⊂ C be a compact set, and {ft}t∈U be a contin- uous family of mixed polynomials. If f il(ft) ≥ 2k + wn for every t ∈ U , then ft ρk Lipschitz constant independent of t. is a germ of a Lipschitz function at 0 with the Proof. Set gt(x) = ft , ∀x 6= 0 and gt(0) = 0. Note that g is a mixed rational ρk function. By Lemma 7.1, it suffices to show that gt is continuous and has the partial derivatives bounded everywhere except 0. Here, by the partial derivatives we mean the derivatives in variables xi and ¯xi. Let A be the set of vertices of Γ+(ρk). Then, ρk = ρA and ρ2 k ∼ ρ2A (see Lemma 7.5). Since ρk is quasihomogeneous of the type (w1, . . . , wn; 2k), Γ+(A) has only one (n − 1)-dimensional compact face. The weight this face is w = (w1, . . . , wn) and the total weight is d = 2k. Since f il(ft) > 2k, limx→0 function. And, ft ρk = 0 (see Lemma 7.3), gt is a continuous ∂gt ∂xi = 1 ρ2 k (ρk ∂ft ∂xi − ft ∂ρk ∂xi ) ∼ 1 ρ2A (ρk ∂ft ∂xi − ft ∂ρk ∂xi ). It is clear that f il(ρk ∂ft ∂xi − ft ∂ρk ∂xi ) ≥ f il(ρk) + f il(ft) − wn ≥ 4k. By Lemma 7.5, ∂gt on ∂xi the compact set U , the bound can be chosen such that it does not depend on t. is bounded near the origin By the continuity of ∂gt ∂xi The boundedness of ∂g ∂ ¯xi proof. is proved in the same way. This completes the (cid:3) Theorem 7.8. Let f : Cn, 0 → C, 0 be a germ of a quasihomogeneous polynomial of the type (w1, . . . , wn; d) with isolated singularity at 0. Let ft(x) = f (x) + tθ(x, t) be a smooth deformation of f such that the initial part of ft w.r.t (w1, . . . , wn) has an isolated singularity at 0 for every t. If f il(θ) ≥ d + maxi,j{wi − wj} then ft is strongly bi-Lipschitz trivial over U . CLASSIFICATION OF LIPSCHITZ SIMPLE FUNCTION GERMS 21 Proof. Notice that ∂f ∂xi si = d − wi. Set is quasihomogeneous of type (w1, . . . , wn : si) where N ∗f =Xi , k = lcm(si). (cid:12)(cid:12)(cid:12)(cid:12) ∂f ∂xi(cid:12)(cid:12)(cid:12)(cid:12) 2αi =Xi (cid:18) ∂f ∂xi ∂f ∂xi(cid:19)αi where αi = k si t Let us denote by f int It is obvious that N ∗f int the initial part of ft w.r.t the weight w = (w1, . . . , wn). is a quasihomogeneous mixed polynomial of the type (w1, . . . , wn; 2k). Since f int has isolated singularity at 0, by Lemma 7.6, N ∗f int t + tR(x, t) where f il(R(x, t)) > 2k. It follows from Lemma 7.5 that R(x, t)/ρk(x) → 0 when x → 0. Thus, N ∗ft ∼ ρk. t ∼ ρk. We have N ∗ft = N ∗f int t t Now, consider the vector field W =P Wi ∂xi(cid:19)αi(cid:18) ∂ft ∂t (cid:18) ∂ft Wi := ∂ft ∂xi(cid:19)αi−1 where . ∂ ∂xi We have ∂ft ∂t N ∗ft = Dft(W ). Without loss of generality, we may assume that w1 ≤ . . . ≤ wn. We have f il( ∂ft ∂t ) ≥ d + wn − w1 and f il(cid:18)(cid:18) ∂ft ∂xi(cid:19)αi(cid:18) ∂ft ∂xi(cid:19)αi−1(cid:19) = αif il(cid:18)(cid:18) ∂ft ∂xi(cid:19)(cid:19) + (αi − 1)f il(cid:18)(cid:18) ∂ft ∂xi(cid:19)(cid:19) = (2αi − 1)f il(cid:18)(cid:18) ∂ft ∂xi(cid:19)(cid:19) = (2αi − 1)(d − wi) = 2k − d + wi ≥ 2k − d + w1. Thus, f il(Wi) ≥ (d + wn − w1) + (2k − d + w1) = 2k + wn. By Lemma 7.7, Wt ρk(x) is Lipschitz. Since ρk ∼ N ∗ft, vt := Wt is also a Lipschitz vector field. The flow generated by v gives the required bi-Lipschitz triviality (see Theorem 7.2). (cid:3) N ∗ft Remark 7.9. (1) Notice that Theorem 3.3 in [8] has a gap when applied to homogeneous germs. The statement of the theorem only holds for sufficiently small t. Consider for example ft(x, y) = x2 + y2 + t(xy + y3). This family satisfies the hypothesis of the theorem but it is not bi-Lipschitz trivial. It is easy to see that f2(x, y) = (x + y)2 + 2y3 cannot be bi-Lipschitz equivalent to ft (t 6= 2), for their tangent cones are different. It seems the mistake first occurs in Lemma 3 in [22] which is used to prove the theorem. (2) Theorem 7.8 also holds in the real case. An application of Theorem 7.8 is that 22 NHAN NGUYEN, MARIA RUAS AND SAURABH TRIVEDI Corollary 7.10. The singularities X9, Z11, W12, P8, S11, S12, U12 in Arnold's list are bi-Lipschitz trivial. In the following, we show that certain families are bi-Lipschitz trivial. Theorem 7.8 does not apply to these families. Lemma 7.11. T2,4,5 : ft(x, y) = x4 + y5 + tx2y2(t 6= 0) is strongly bi- Lipschitz trivial. Proof. It suffices to prove that for t0 6= 0 there is a neighbourhood U of t0 such that ft is bi-Lipschitz trivial along U . We will show that there is continuous vector field X(x, y, t) = − ∂ ∂y which is Lipschitz in (x, y) such that X.ft = 0. The conclusion will then follow from the Thom-Levine criterion. ∂x + X2(x, y, t) ∂ ∂t + X1(x, y, t) ∂ We have = 4x3 + 2txy2, ∂ft ∂y = 5y4 + 2tx2y2, ∂ft ∂t = x2y2. ∂ft ∂x 2 = 4x3 +2txy22 +5y4 +2tx2y2. It is a mixed polyno- mial. Denote by A the set of vertices of Γ+(ρ), i.e. A = {(6, 0), (2, 4), (0, 8)}. Put ρ =(cid:12)(cid:12)(cid:12) ∂ft ∂x(cid:12)(cid:12)(cid:12) 2 +(cid:12)(cid:12)(cid:12) ∂ft ∂y(cid:12)(cid:12)(cid:12) y σ2 σ1 x Figure 1. Newton diagram of Γ+ρ The control function associated with A is ρA = x6 + x2y4 + y8. Since supp(A) ⊂ Γ+A, ρ . ρA. It is easy to check that 0 is a Γ+(A)-isolated point of ρ, hence by Lemma 7.4, ρA . ρ. Thus, ρ ∼ ρA. Notice that This implies, ρ. ∂ft ∂t ∂t =(cid:18) ∂ft = (cid:16) ∂ft ∂t ∂x ∂ft ∂x(cid:19) ∂ft ∂x(cid:17)ρ ∂ft ∂x ∂ft . (13) ∂t ∂y ∂ft +(cid:18) ∂ft ∂y(cid:19) ∂ft + (cid:16) ∂ft ∂y(cid:17)ρ ∂ft ∂y ∂ft ∂t ∂ft ∂t CLASSIFICATION OF LIPSCHITZ SIMPLE FUNCTION GERMS 23 Put a = ∂ft ∂t and ∂ft ∂x and b = ∂ft ∂t ∂ft ∂y . Then, set a(x, y, t) =( a ρ , 0 (x, y) 6= (0, 0) (x, y) = (0, 0) b(x, y, t) =( b ρ , 0 (x, y) 6= (0, 0) (x, y) = (0, 0) . By the equation (13), the vector field X(x, y, t) = −1 ∂ ∂t + a ∂ ∂x + b ∂ ∂y satisfies X.ft = 0. We will now show that X is continuous. The New- ton diagram of A has two 1-dimensional compact faces, denoted σ1 and σ2 (see the Figure 1). The weights corresponding to σ1 and σ2 are wσ1 = (4, 4) and wσ2 = (6, 3) with the total weight d = 24. Computation gives mini=1,2{f ilwσi (a)} = 28 > d and mini=1,2{f ilwσi (b)} = 30 > d. By Lemma tend to 0 when (x, y) tends to (0, 0). Since ρ ∼ ρA, a and b 7.5, a and b ρa ρA are continuous. In order to prove that a and b are Lipschitz in (x, y), it suffices to show that the derivatives with respect to x, x, y and y are bounded everywhere except 0. We will show that ∂A ∂x is bounded, the boundedness of the other partial derivatives can be shown similarly. For (x, y) 6= (0, 0), ∂a ∂x = 1 ρ2 (ρ ∂a ∂x − a ∂ρ ∂x ) ∼ 1 ρ2A (ρ ∂a ∂x − a ∂ρ ∂x ) ∂x −a ∂ρ Notice that f il(ρ ∂a faces σ1 and σ2, we get ∂x ) ≥ f il(a)+f il(ρ)−wx. Restricting to the compact f ilwσ1 ((ρ ∂a ∂x − a ∂ρ ∂x ) ≥ f ilwσ1 (a) + f ilwσ1 (ρ) − w1,x = 28 + 24 − 4 = 48, ((ρ − a ∂a ∂x f ilwσ2 ∂ρ ∂x Thus, mini=1,2 f ilwσi bounded. ) ≥ f ilwσ2 (a) + f ilwσ2 (ρ) − w1,x = 30 + 24 − 6 = 48. ((ρ ∂a ∂x − a ∂ρ ∂x )) = 48 ≥ 2d. By Lemma 7.5, ∂A ∂x is (cid:3) Remark 7.12. (i) Using the method of Lemma 7.11, one can easily prove that the following families are strongly bi-Lipschitz trivial : (1) T2,5,5, take ρ =(cid:12)(cid:12)(cid:12) 2 ∂ft ∂x(cid:12)(cid:12)(cid:12) +(cid:12)(cid:12)(cid:12) 2 ∂ft ∂y(cid:12)(cid:12)(cid:12) , 24 NHAN NGUYEN, MARIA RUAS AND SAURABH TRIVEDI (2) Tp,q,r, 3 ≤ p, q, r ≤ 5, take 2 2 2 , ∂ft ∂ft ∂ft +(cid:12)(cid:12)(cid:12)(cid:12) +(cid:12)(cid:12)(cid:12)(cid:12) ∂x(cid:12)(cid:12)(cid:12)(cid:12) ∂y(cid:12)(cid:12)(cid:12)(cid:12) ∂z (cid:12)(cid:12)(cid:12)(cid:12) ρ =(cid:12)(cid:12)(cid:12)(cid:12) ρ =(cid:12)(cid:12)(cid:12)(cid:12) + y2(cid:12)(cid:12)(cid:12)(cid:12) +(cid:12)(cid:12)(cid:12)(cid:12) ∂z (cid:12)(cid:12)(cid:12)(cid:12) ∂y(cid:12)(cid:12)(cid:12)(cid:12) ∂x(cid:12)(cid:12)(cid:12)(cid:12) ∂ft ∂ft ∂ft 4 4 4 . ρ = ∂ft ∂t = n Xi=1 n Xi=1 ai ∂ft ∂xi ∂ft ∂xi . ∂ft ∂xi ai ∂ft ∂t ρ ∂ft ∂xi . Notice that all the control functions we have used so far are of the form . It is easy to see that ρ ∈ Jf since we can write (3) Q10, take 2 ρ =Pn i=1 ai(cid:12)(cid:12)(cid:12) ∂ft ∂xi(cid:12)(cid:12)(cid:12) This implies that ∂ft ∂xi ai ∂ft ∂t ρ If the vector field W = (W1, . . . , Wn) where Wi = is Lipschitz then it satisfies the Thom–Levine criterion. Checking the Lipschitzness of W is quite easy because we know exactly what the Wi's are. However, the vector field W induced by the control function ρ of the form above is not always Lipschitz. For example, in the case Q11, we could not find a control function of the same type. For this reason, in the following we suggest another control function for Q11 that belongs to Jf . This control function does not give precisely what the vector field W is, but it enough for us to prove the bi-Lipschitz triviality of Q11. Lemma 7.13. f (x, y, z) = x3+y2z+xz3+tz5 is strongly bi-Lipschitz trivial. Proof. The partial derivatives of f are ∂f ∂z = y2 + 3xz2 + 5tz4 and ∂f ∂t = z5. Set ρ = x16 + y16 + z16. It is obvious that ρ is a standard control function of type (1, 1, 1; 16). We will show that there exist germs of mixed polynomials εi(x, y, z, t) (i = 1, 2, 3) satisfying ∂x = 3x2 + z3, ∂f ∂y = 2yz, ∂f ρ = ε1 ∂f ∂y + ε2 ∂f ∂y + ε3 ∂f ∂z , (14) such that εiz5 ρ i = 1, 2, 3. Multiplying the equation (14) by ∂f is a Lipschitz function germ in the variables (x, y, z) for each ∂t = z5 we get, ∂f ∂t =(cid:18) ε1z5 ρ (cid:19) ∂f ∂x +(cid:18) ε2z5 ρ (cid:19) ∂f ∂y +(cid:18) ε3z5 ρ (cid:19) ∂f ∂z . (15) The lemma will then follow from the Thom-Levine criterion. CLASSIFICATION OF LIPSCHITZ SIMPLE FUNCTION GERMS 25 Observe that, for εiz5 ρ to be a Lipschitz function germ, it is enough to show that its partial derivatives in the variables x, ¯x, y, ¯y, z and ¯z are bounded. Set ε′ i = εz5. We have ∂ ∂u(cid:18) εiz5 ρ (cid:19) = ∂ε′ ∂u ρ − ε′ i i ∂ρ ∂u ρ2 , where u ∈ {x, ¯x, y, ¯y, z, ¯z}. Put N = ∂ε′ i ∂u ρ − ε′ i ∂ρ ∂u . Notice then that f il(N ) ≥ f il(ρ) + f il(ε′ i) − wu, where wu is the the weight of u. By Lemma 7.3, if f il(ρ) + f il(ε′ i) − wu ≥ f il(ρ2) = 2f il(ρ), ∀u ∈ {x, ¯x, y, ¯y, z, ¯z} (16) then ∂ ∂u(cid:18) εiz5 ρ (cid:19) is bounded. Since the weights of x, ¯x, y, ¯y, z, ¯z are equal to 1, the inequality (16) is equivalent to f il(ε′ i) ≥ 17, or f il(εi) ≥ 12. Since ρ = x8 ¯x8 + y8 ¯y8 + z8 ¯z8, to prove (14) it suffices to prove that x8, y8 and z8 belong to Jf and their coefficients have filtrations ≥ 4. Here, ∂f ∂z ∈ Jf we mean by coefficients of a polynomial h = h1 (h1, h2, h3). ∂f ∂x + h2 ∂f ∂y + h3 Observe that, This gives ∂f ∂x · x6 = 3x8 + x6z3 ∂f ∂x · x4z3 = 3x6z3 + x4z6 3 1 0 3 ! x8 x6z3 ! = ∂f ∂x · x6 ∂x · x4z3 − x4z6 ! ∂f (17) It is easy to check that z6 ∈ Jf (see the Groebner basis of Jf ). Then, x4z6 ∈ Jf and its coefficients have filtrations ≥ 4. Since the system (17) has unique solution, x8 ∈ Jf . The filtrations of x6 and x4z3 are ≥ 4, this implies that the coefficients of x8 have filtrations ≥ 4. Similarly, ∂f ∂f ∂x · z5 = z8 + 3x2z5 ∂z · xz3 = ∂z · z5 = ∂f 3x2z5 + 5txz7 + xz3y2 3xz7 + 5tz9 + z5y2 This gives   1 3 0 0 3 5t 3 0 0 z8 x2z5 xz7       =  ∂f ∂x · z5 ∂f ∂z · xz3 − xz3y2 ∂f ∂z · z5 − 5tz9 − z5y2 (18)   26 NHAN NGUYEN, MARIA RUAS AND SAURABH TRIVEDI Put g(x, z) = f (x, 0, z). We have ∂g ∂z − y2. Checking the Groebner basis of Jg, we have z5 ∈ Jg. Therefore, there exist polynomials η1 and η2 such that ∂x and ∂g ∂x = ∂f ∂z = ∂f Thus, z5 = η1 ∂g ∂x + η2 ∂g ∂z = η1 ∂f ∂x + η2( ∂f ∂z − y2) z9 = η1 · z4. = η1 · y4. ∂f ∂x ∂f ∂x + η2 · z4 · + η2 · y4 · ∂f ∂z ∂f ∂z − η2 · z4y2 − 1 2 η2 · z3y · ∂f ∂y . This shows that z9 ∈ Jf and its coefficients have filtrations ≥ 4. Also, xz3y2 = 1 ∂y , therefore, they belong to Jf and the filtrations of their coefficients ≥ 4. The system (18) has unique solution, this implies that z8 ∈ Jf with coefficients of filtrations ≥ 4. ∂y , z5y2 = 1 2 xz2y · ∂f 2 z4y · ∂f Finally y8 = y6 · ∂f have filtrations ≥ 6. ∂z − 1 2 (3xy5z + 5ty5z3) · ∂f ∂x , y8 ∈ Jf and its coefficients (cid:3) 8. Classification of Lipschitz simple germs 8.1. Adjacencies. Consider the list of adjacencies given in Theorem 3.2. We know from a result of Brieskorn [4] that the singularities not appearing in the brackets do not deform to J10. In this section we prove that the singularities enclosed in the brackets deform to J10. Let ft(x, y) = x3 + txy4 + y6 ∈ m2. Give (x, y) the weight w = (2, 1). Then, f ilw(ft) = 6. It follows from Arnold's result [2] that Lemma 8.1. (i) Any polynomial g(x, y) ∈ m2 with isolated singularity at 0 and f ilw(g) = 6 is smoothly equivalent to ft for some t; (ii) If g(x, y) is a monomial such that f ilw(g) ≥ 7 then ft + g is smoothly equivalent to ft. We prove further that, Lemma 8.2. Let f ∈ m2 n be a polynomial with isolated singularity at 0. Let w1 = 2, w2 = 1, wj = 3, ∀j ≥ 3, and let w = (w1, . . . , wn). If f ilw(f ) ≥ 6, then f deforms to J10. Proof. First, we take the deformation as follows F = f +Xj≥3 ajx2 j . By the Splitting lemma, we know that F is smooth equivalent to g(x1, x2) + j . Notice that f ilw(g) ≥ f ilw(f ) (see Gibson [10], page 126, the It suffices to prove that algorithm in the proof of the Splitting lemma). Pj≥3 ajx2 CLASSIFICATION OF LIPSCHITZ SIMPLE FUNCTION GERMS 27 j deforms to J10. Indeed, consider the following defor- g(x1, x2) +Pj≥3 ajx2 mation: G = g(x1, x2) +Xj≥3 ajx2 j + b1x3 1 + b2x6 2 + b3x1x4 2. It follows from Lemma 8.1 that g(x1, x2) + b1x3 2 + b3x1x4 equivalent to ft(x1, x2) for some t. This implies that G is J10. 1 + b2x6 2 is smoothly (cid:3) Theorem 8.3. All germs belonging to the families enclosed in the brackets in the diagram in Theorem 3.2 deform to J10. Proof. (1) That Tp,q,6, p ≥ 2, q ≥ 3, Z12 and W13 deform to J10 is a direct consequence of Lemma 8.2 by giving (x, y, z) the weight w = (2, 1, 3). (2) Q12 → J10. We have Q12 : x3 + y5 + yz2 + axy4. By the change of coordinates of type z 7→ z + iy2 and leaving other coordi- nates fixed, we can eliminate y5. That is, Q12 is smooth equivalent to g(x, y, z) = x3 + yz2 + 2iy3z + axy4. Let w = (2, 1, 3) be the weight of (x, y, z). Then, f ilw(g) = 6. This implies that g deforms to J10 (see Lemma 8.2). (3) S1,0 → J10. Recall that S1,0 has the normal form x2z + yz2 + y5 + a1zy3 + a2zy4. First, by the change of coordinate z 7→ z + Ay2, we have S1,0 is smooth equivalent to g(x, y, z) = Ay2x2 + zx2 + bAy6 + (A2 + Aa + 1)y5 + bzy4 + (2A + a)zy3 + z2y Choose A such that A2 + Aa + 1 = 0 to eliminate y5. Give (x, y, z) the weight w = (2, 1, 3). Obviously, f ilw(g) = 6. By Lemma 8.2, g deforms to J10. (4) U12 → J10. Recall that U12 has the normal form x3 + y3 + z4 + axyz2. Take the deformation as suggested in [4]: x3 − y3 + z4 + axyz2 + t2(x + y)2 + 2t(x + y)z2. By the change of coordinate x 7→ x − y, we have x3 − 2y3 + 3xy2 + z4 + axyz2 − ay2z2 + axyz2 + 2txz2 + t2x2. By changing x 7→ x− 1 2t2 (x2−3xy+3y2+2tz2), the term z4 will be eliminated, and the polynomial obtained has filtration with respect to the weight w = (3, 2, 1) is equal to 6, so it deforms to J10 (see Lemma 8.2). (5) (O) → J10. Since every germ of corank ≥ 4 can deform to a germ of corank 4, it suffices to prove that all corank 4 germs deform to J10. Let g be a corank 4 germ. By the Splitting lemma, we may assume (up to addition of a non-degenerate quadratic form of further variables) that 4. Consider a small deformation of g of the form hλ = g + λ(x3 + y3 + g ∈ m3 z3 + w3). Notice that for λ 6= 0 near 0, hλ = λ1x3 + λ2y3 + λ3z3 + λ4w3 + p 28 NHAN NGUYEN, MARIA RUAS AND SAURABH TRIVEDI for some λi 6= 0, i = 1, . . . , 4 and p is a polynomial of order greater than or equal to 3 which does not contain the terms {x3, y3, z3, w3}. By a change of coordinates we may assume λi = 1, ∀i = 1, . . . , 4. Transversal unfolding shows that j3(hλ) is smooth equivalent to fa,b,c,d(x, y, z, w) = x3 + y3 + z3 + w3 + axyz + bxyw + cxzw + dyzw, for some a, b, c, d (see [10], Chapter III). 4, by the complete transversal method (see [5]), every germ whose 3-jet is equivalent to fa,b,c,d has 4-jet equivalent to fa,b,c,d,e := fa,b,c,d + exyzw for some e. It is easy to check that hλ is 4-determined, thus it is equivalent to fa,b,c,d,e for some a, b, c, d, e. 4 + hxyzwi = m4 Since Jf m2 4 + m5 It remains to show that fa,b,c,d,e deforms to J10. First, we take the change of coordinate as follows: x 7→ x − z + mw + a 3 y, z 7→ x + z, w 7→ w + lx for m, l ∈ C. Then, deform the obtained germ with tx2 + tw2. Next, we use the algorithm of the Splitting lemma to eliminate the terms of degree 3 contain- ing either x or w or both. The resulting polynomial, let say ga,b,c,d,e, includes monomials of filtration (with respect to the weight (3, 2, 1, 3)) greater than or equal to 6 except z4 and zy3. Since the coefficients of z4 and zy3 de- pend on two parameters m and l, there is a closed algebraic set S ⊂ C5 of dimension less than 5 such that for all (a, b, c, d, e) ∈ C5 \ S we can choose m, l such that the coefficients of z4 and zy3 equal to zero. By Lemma 8.2, ga,b,c,d,e deforms to J10 for (a, b, c, d, e) ∈ C5 \ S. Since dim S < 5, thus ga,b,c,d,e deforms to J10 for all a, b, c, d, e. (cid:3) 8.2. Classification of Lipschitz simple germs. We are now in position to state and prove our main result. Observe that (1) smooth simple germs are Lipschitz simple; (2) J10 is Lipschitz modal (see Theorem 6.4); (3) if a germ deforms to a class of Lipschitz modal singularities, then it is also Lipschitz modal; Theorem 8.4. A germ f ∈ m2 only if it is bi-Lipschitz equivalent to one of the germs in the table below: n of corank 1 or 2 is Lipschitz simple if and Table 2: Corank 1 and 2 Lipschitz simple germs Name Smooth normal form Lipschitz normal form Codimension Corank Ak Dk E6 E7 E8 X9 T2,4,5 T2,5,5 Z11 xk+1 x2y + yk−1 x3 + y4 x3 + xy3 x3 + y5 x4 + y4 + ax2y2 x4 + y5 + ax2y2) x5 + y5 + ax2y2 x3y + y5 + axy4 xk+1 x2y + yk−1 x3 + y4 x3 + xy3 x3 + y5 x4 + y4 + x2y2 x4 + y5 + x2y2 x5 + y5 + x2y2 x3y + y5 + xy4 1 2 k ≥ 1 k ≥ 4 6 7 8 9 10 11 11 CLASSIFICATION OF LIPSCHITZ SIMPLE FUNCTION GERMS 29 W12 x4 + y5 + ax2y3 x4 + y5 + x2y3 12 Proof. By Theorem 8.3, germs of corank 2 which are not equivalent to germs in the list deform to J10 and hence are not Lipschitz simple. The singularities Ak, (k ≥ 1), Dk, (k ≥ 4), E6, E7 and E8 are smooth simple germs (see Theorem 3.1), they are obviously Lipschitz simple germs. Now we prove that the germs in X9 are Lipschitz simple. Let f ∈ X9 and k ∈ N be sufficiently large for f . Since corank and codimension of singularities are upper semicontinuous, the only singularities contained in a small enough neighbourhood U of jkf (0) are X9 or ADE-singularities. It is shown in Corollary 7.10 that X9 is a bi-Lipschitz trivial family. We know, moreover, that ADE-singularities belong to only finitely many bi-Lipschitz classes. This implies that f is a Lipschitz simple germ. The proof for T2,4,5 is as follows. Let g ∈ T2,4,5 and m ∈ N be sufficently large for g. Then, again the only singularities contained in the small neigh- bourhood V of jmg(0) are those with corank ≤ 2 and codimension ≤ 10. By Brieskorn's result [4], T2,4,5 cannot deform to J10. Shrinking V if necessary, we can assume that V does not contain germs of J10. Thus, singularities in V are either contained in T2,4,5 or X9 or are ADE-singularities. It is proved in Lemma 7.11 that T2,4,5 is bi-Lipschitz trivial. We also know that all germs in X9 and ADE-singularities belong to finitely many bi-Lipschitz classes. Therefore, g is Lipschitz simple. The proof for the other germs is similar by induction on the codimen- sion and using the fact that they belong to bi-Lipschitz trivial families (see Corollary 7.10, Lemma 7.2 and Remark 7.12) and none of them deform to a class of Lipschitz modal singularities (see Brieskorn's result [4]). It remains to prove that every normal form in the third column of Table 2 represents a unique bi-Lipschitz class. Notice that except T2,5,5 and Z11, they all have different codimensions. The normal forms of T2,5,5 and Z11 are given by f = x5 + y5 + x2y2 and g = x3y + y5 + xy4. It is easy to check that ΣHf = {(x, y) ∈ C2 : xy = 0} while ΣHg = {(x, y) ∈ C2 : x = 0}. Therefore, by Theorem 5.1, T2,5,5 and Z11 are of different bi-Lipschitz types. (cid:3) Theorem 8.5. A germ f ∈ m2 if it is bi-Lipschitz equivalent to one of the germs in the table below: n of corank 3 is Lipschitz simple if and only Table 3: Corank 3 Lipschitz simple germs Name P8 = T3,3,3 T3,3,4 T3,3,5 T3,4,5 Smooth normal form Lipschitz normal form Codimension Corank x3 + y3 + z3 + axyz x3 + y3 + z4 + axyz x3 + y3 + z5 + axyz x3 + y4 + z5 + axyz x3 + y3 + z3 + xyz x3 + y3 + z4 + xyz x3 + y3 + z5 + xyz x3 + y4 + z5 + xyz 8 9 10 11 30 NHAN NGUYEN, MARIA RUAS AND SAURABH TRIVEDI T3,5,5 T4,4,4 T4,4,5 T4,5,5 T5,5,5 Q10 Q11 S11 S12 x3 + y5 + z5 + axyz x4 + y4 + z4 + axyz x4 + y4 + z5 + axyz x4 + y5 + z5 + axyz x5 + y5 + z5 + axyz x3 + y4 + yz2 + axy3 x3 + y2z + xz3 + az5 x4 + y2z + xz2 + ax3z x2y + y2z + xz3 + az5 x3 + y5 + z5 + xyz x4 + y4 + z4 + xyz x4 + y4 + z5 + xyz x4 + y5 + z5 + xyz x5 + y5 + z5 + xyz x3 + y4 + yz2 + xy3 x3 + y2z + xz3 + z5 x4 + y2z + xz2 + x3z x2y + y2z + xz3 + z5 12 11 12 13 14 10 11 11 12 3 Proof. By Theorem 8.3, all germs of corank 3 which are not in the above list deform to J10, hence they are Lipschitz modal. We know from Corollary 7.10, Lemma 7.13 and Remark 7.12 that every singularity class in the above list is bi-Lipschitz trivial. Moreover, by the result of Brieskorn [4] none of the singularities are adjacent to any class of Lipschitz modal singularities so by arguments as in the proof of Theorem 8.4, they are Lipschitz simple. Now let us prove that every normal form in the third column of Table 3 represents a unique bi-Lipschitz class. Notice that except Q11 and S11, all germs have different bi-Lipschitz type, for one of the bi-Lipschitz invariants, namely either the codimension or the singular part of the tangent cone, associated with them are different. We list the invariants for these germs below (see Table 4). Table 4: Singular sets of the algebraic tangent cone of normal forms Name Lipschitz normal form (f ) ΣHf Codimension T3,3,5 Q10 T3,4,5 T4,4,4 Q11 S11 T3,5,5 T4,4,5 S12 x3 + y3 + z5 + xyz x3 + y4 + yz2 + xy3 x3 + y4 + z5 + xyz x4 + y4 + z4 + xyz x3 + y2z + xz3 + z5 x4 + y2z + xz2 + x3z x3 + y5 + z5 + xyz x4 + y4 + z5 + xyz x2y + y2z + xz3 + z5 y-axis ∪ z-axis y-axis y-axis ∪ z-axis x-axis ∪ y-axis ∪ z-axis y-axis x-axis y-axis ∪ z-axis x-axis ∪ y-axis ∪ z-axis z-axis 10 10 11 11 11 11 12 12 12 Lemma 8.6 shows that the bi-Lipschitz types of Q11 and S11 are different. (cid:3) Lemma 8.6. Q11 and S11 are not in the same bi-Lipschitz class. Proof. Set f (x, y, z) = x3 + y2z + xz3 + z5 and g(x, y, z) = x4 + y2z + xz2 + x3z. Recall that the germs in Q11 are bi-Lipschitz equivalent to f + Q, CLASSIFICATION OF LIPSCHITZ SIMPLE FUNCTION GERMS 31 and germs in S11 are bi-Lipschitz equivalent to g + Q where Q = w2 1 + . . . + w2 n−3. We need to show that f + Q and g + Q are not bi-Lipschitz equivalent. In [3], Theorem 4.2, Bavi`a-Ausina and Fukui proved that the log canonical threshold of the Jacobian ideal, let us denote by lct, is a bi-Lipschitz invariant. We show that lct(J(f + Q)) 6= lct(J(g + Q)). By Lemma 3.2 in Totaro [24], lct(J(f + Q)) = lct(Jf ) + lct(JQ) and lct(J(g + Q)) = lct(Jg) + lct(JQ). Thus, it suffices to prove that lct(Jf ) 6= lct(Jg). To calculate the log canonical threshold of a Jacobian ideal we follow the method suggested in [3] using the Newton polyhedron. Let us start with computing lct(Jf ). Set I = Jf = h3x2 + z3, 2yz, y2 + 3xz2 + 5z4i. Let I 0 denotes the monomial ideal associated with the Newton polyhedron Γ+(I) of I (see the definitions in [3], page 8.1). Then, I 0 = hx2, yz, y2, z3i. By formula (4.1) in [3], we have lct(I 0) = 3 2 . Consider the map α = (α1, α2, α3) =(cid:18) ∂f ∂x , ∂f ∂y , ∂f ∂z(cid:19). It is obvious that α1, α2, α3 ∈ I and Γ+(α) = Γ+(I 0). Moreover, α is strongly non-degenerate (see [3], Definition 4.6). By Corollary 4.9 of [3], lct(I) = lct(I 0) = 3 2 . The calculation of lct(Jg) is similar. Set J = Jg = h4x3 + z2 + 3x2z, 2yz, y2 + 2xz + x3i. Then the monomial ideal associated with the Newton polyhedron Γ+(J) is J 0 = hx3, yz, y2, z2i. It is obvious by formula (4.1) in [3] that lct(J 0) = 4 3 . Consider the map β = (β1, β2, β3) =(cid:18) ∂g ∂x , ∂g ∂y , ∂g ∂z(cid:19). Since β1, β2, β3 ∈ J, Γ+(β) = Γ+(J 0) and β is strongly non-degenerate, by Corollary 4.9 of [3] we have lct(J) = lct(J 0) = 4 (cid:3) 3 Theorem 8.7. Every germ of corank ≥ 4 is Lipschitz modal. Proof. By Theorem 8.3, all germ of corank ≥ 4 deform to J10. Since J10 are Lipschitz modal, the result follows. (cid:3) A direct consequence of the above theorems is Corollary 8.8. (1) A germ in m2 to J10. n is Lipschitz modal if and only if it deforms (2) Germs in J10 are Lipschitz modal germs of the smallest codimension. (3) Every germ of smooth modality greater than or equal to 2 is Lipschitz modal. 32 NHAN NGUYEN, MARIA RUAS AND SAURABH TRIVEDI Acknowledgements. We would like to thank D. Trotman for suggesting the problem, and A. Fruhbis-Kruger for her useful suggestion on the proof that singularities of corank ≥ 4 are Lipschitz modal. We would also like to thank H. D. Nguyen and A. Fernandes for their interest and helpful discussion. The first author was supported by FAPESP grant 2016/14330 − 0. The second author was supported by FAPESP grant 2014/00304 − 2 and CNPq grant 306306/2015 − 8. The third author was supported by FAPESP grant 2015/12667 − 5. References [1] O. M. Abderrahmane. Poly`edre de Newton et trivialit´e en famille. J. Math. Soc. Japan, 54(3):513–550, 2002. [2] V. I. Arnold. Local normal forms of functions. Invent. Math., 35:87–109, 1976. [3] C. Bivi`a-Ausina and T. Fukui. Invariants for bi-Lipschitz equivalence of ideals. Q. J. Math., 68(3):791–815, 2017. [4] E. Brieskorn. Die Hierarchie der 1-modularen Singularitaten. Manuscripta Math., 27(2):183–219, 1979. [5] J. W. Bruce, N. P. Kirk, and A. A. du Plessis. Complete transversals and the classi- fication of singularities. Nonlinearity, 10(1):253–275, 1997. [6] J. C. F. Costa, M. J. Saia, and C. H. Soares J´unior. Bi-Lipschitz G -triviality and Newton polyhedra, G = f R, C , K , RV , CV , KV . In Real and complex singularities, volume 569 of Contemp. Math., pages 29–43. Amer. Math. Soc., Providence, RI, 2012. [7] W. Ebeling. Functions of several complex variables and their singularities, volume 83 of Graduate Studies in Mathematics. American Mathematical Society, Providence, RI, 2007. [8] A. Fernandes and M. A. S. Ruas. Bi-Lipschitz determinacy of quasihomogeneous germs. Glasg. Math. J., 46(1):77–82, 2004. [9] A. Fernandes and C. H. Soares J´unior. On the bilipschitz triviality of families of real maps. In Real and complex singularities, volume 354 of Contemp. Math., pages 95–103. Amer. Math. Soc., Providence, RI, 2004. [10] C. G. Gibson. Singular points of smooth mappings, volume 25 of Research Notes in Mathematics. Pitman (Advanced Publishing Program), Boston, Mass.-London, 1979. [11] G-M Greuel and H. D. Nguyen. Right simple singularities in positive characteristic. J. Reine Angew. Math., 712:81–106, 2016. [12] J-P. Henry and A. Parusi´nski. Existence of moduli for bi-Lipschitz equivalence of analytic functions. Compositio Math., 136(2):217–235, 2003. [13] J-P. Henry and A. Parusi´nski. Invariants of bi-Lipschitz equivalence of real analytic functions. In Geometric singularity theory, volume 65 of Banach Center Publ., pages 67–75. Polish Acad. Sci. Inst. Math., Warsaw, 2004. [14] C. T. Le and V. C. Luong. On tangent cones of analytic sets and lojasewicz exponents. Preprint, 2018. [15] J. Milnor. Singular points of complex hypersurfaces. Annals of Mathematics Studies, No. 61. Princeton University Press, Princeton, N.J., 1968. [16] T. Mostowski. Lipschitz equisingularity. Dissertationes Math. (Rozprawy Mat.), 243:46, 1985. [17] H. D. Nguyen. Right unimodal and bimodal singularities in positive characteristic. Internat. Math. Res. Notices, 2017. [18] N. Nguyen and G. Valette. Lipschitz stratifications in o-minimal structures. Ann. Sci. ´Ec. Norm. Sup´er. (4), 49(2):399–421, 2016. [19] M. Oka. Non-degenerate mixed functions. Kodai Math. J., 33(1):1–62, 2010. CLASSIFICATION OF LIPSCHITZ SIMPLE FUNCTION GERMS 33 [20] A. Parusi´nski. Lipschitz stratification of subanalytic sets. Ann. Sci. ´Ecole Norm. Sup. (4), 27(6):661–696, 1994. [21] J. J. Risler and D. J. A Trotman. Bi-Lipschitz invariance of the multiplicity. Bull. London Math. Soc., 29(2):200–204, 1997. [22] M. A. S. Ruas and M. J. Saia. C l-determinacy of weighted homogeneous germs. Hokkaido Math. J., 26(1):89–99, 1997. [23] J. E. Sampaio. Bi-Lipschitz homeomorphic subanalytic sets have bi-Lipschitz home- omorphic tangent cones. Selecta Math. (N.S.), 22(2):553–559, 2016. [24] B. Totaro. The ACC conjecture for log canonical thresholds (after de Fernex, Ein, Mustat¸a, Koll´ar). Ast´erisque, (339):Exp. No. 1025, ix, 371–385, 2011. S´eminaire Bour- baki. Vol. 2009/2010. Expos´es 1012–1026. [25] H. Whitney. Tangents to an analytic variety. Ann. of Math. (2), 81:496–549, 1965. Departamento de Matem´atica, ICMC - Universidade de Sao Paulo, 134560- 970, Sao Carlos, S.P, Brasil E-mail address: [email protected] E-mail address: [email protected] E-mail address: [email protected] 

1507.05710

3

1507

The uniformization of the moduli space of principally polarized abelian 6-folds

Starting from a beautiful idea of Kanev, we construct a uniformization of the moduli space A_6 of principally polarized abelian 6-folds in terms of curves and monodromy data. We show that the general ppav of dimension 6 is a Prym-Tyurin variety corresponding to a degree 27 cover of the projective line having monodromy the Weyl group of the E_6 lattice. Along the way, we establish numerous facts concerning the geometry of the Hurwitz space of such E_6-covers, including: (1) a proof that the canonical class of the Hurwitz space is big, (2) a concrete geometric description of the Hodge-Hurwitz eigenbundles with respect to the Kanev correspondence and (3) a description of the ramification divisor of the Prym-Tyurin map from the Hurwitz space to A_6 in the terms of syzygies of the Abel-Prym-Tyurin curve.

math.AG

math

THE UNIFORMIZATION OF THE MODULI SPACE OF PRINCIPALLY POLARIZED ABELIAN 6-FOLDS VALERY ALEXEEV, RON DONAGI, GAVRIL FARKAS, ELHAM IZADI, AND ANGELA ORTEGA Contents Introduction 1. Kanev's construction and Prym-Tyurin varieties of E6-type 2. The E6 lattice 3. Degenerations of Jacobians and Prym varieties 4. Degenerations of Prym-Tyurin-Kanev varieties 5. Admissible covers and semiabelian Prym-Tyurin-Kanev varieties 6. Positivity properties of the Hurwitz space of E6-covers 7. The Prym-Tyurin map along the boundary components of Hur 8. Ordinary Prym varieties regarded as Prym-Tyurin-Kanev varieties 9. The Weyl-Petri realization of the Hodge eigenbundles 10. The ramification divisor of the Prym-Tyurin map 11. A Petri theorem on Hur References 1 6 9 10 12 16 19 26 32 34 39 42 44 Introduction It is a classical idea that general principally polarized abelian varieties (ppavs) and their moduli spaces are hard to understand, and that one can use algebraic curves to study some special classes, such as Jacobians and Prym varieties. This works particularly well in small dimension, where in this way one reduces the study of all abelian varieties to the rich and concrete theory of curves. For g ≤ 3, a general ppav is a Jacobian, and the Torelli map Mg → Ag between the moduli spaces of curves and ppavs respectively, is birational. For g ≤ 5, a general ppav is a Prym variety by a classical result of Wirtinger [Wir95]. In particular, for g = 5, this gives a uniformization of A5 by curves, as follows. We denote by Rg the Prym moduli space of pairs [C, η] consisting of a smooth curve C of genus g and a non-trivial 2-torsion point η ∈ Pic0(C). By Donagi-Smith [DS81], the Prym map P : R6 → A5 is generically of degree 27, with fibers corresponding to the configuration of the 27 lines on a cubic surface. The uniformization of Ag for g ≤ 5 via the Prym map P : Rg+1 → Ag has been used for many problems concerning ppav of small dimension. Important applications of the Prym uniformization include the proof of Clemens and Griffiths [CG72] respectively Mumford [M74] of the irrationality of smooth cubic threefolds, which rely on the distinctions between Pryms and Jacobians, the proofs of the general Hodge conjecture for the theta divisors of general ppav, see [IvS95] and [ITW16], or the detailed study of the cohomology and stratification of A5 in terms of singularities of theta divisors, see for instance [CF05] or [FGSMV]. The Prym map P : R6 → A5 has been also used 1 pencil contains precisely 24 singular cubic surfaces, each having exactly one node. by the curve of lines in the cubic surfaces of a Lefschetz pencil of hyperplane sections of a cubic By the Hurwitz formula, we find that each such E6-cover C has genus 46. Furthermore, C is π is given by a reflection in W (E6). A prominent example of such a covering π : C → P1 is given threefold X ⊂ P4, see [Kan89a], as well as Section 1 of this paper. Since deg(X∨) = 24, such a endowed with a symmetric correspondence (cid:101)D of degree 10, compatible with the covering π and x (cid:54)= y and π(x) = π(y) belongs to (cid:101)D if and only if the lines corresponding to the points x and y are incident. The correspondence (cid:101)D is disjoint from the diagonal of C × C. The associated defined using the intersection form on a cubic surface. Precisely, a pair (x, y) ∈ C × C with 2 V. ALEXEEV, R. DONAGI, G. FARKAS, E. IZADI, AND A. ORTEGA to determine the birational type of A5. It has been proven in [Don84] that R6 (and hence A5) is unirational. Other proofs followed in [MM83] and [Ver84]. The purpose of this paper is to prove a similar uniformization result for the moduli space A6 of principally polarized abelian varieties of dimension 6. The idea of this construction is due to Kanev [Kan89b] and it uses the geometry of the 27 lines on a cubic surface. Suppose π : C → P1 is a cover of degree 27 whose monodromy group equals the Weyl group W (E6) ⊂ S27 of the E6 lattice. In particular, each smooth fibre of π can be identified with the set of 27 lines on an abstract cubic surface and, by monodromy, this identification carries over from one fibre to another. Assume furthermore that π is branched over 24 points and that over each of them the local monodromy of endomorphism D : JC → JC of the Jacobian satisfies the quadratic relation (D − 1)(D + 5) = 0. Using this, Kanev [Kan87] showed that the associated Prym-Tyurin-Kanev or PTK variety P T (C, D) := Im(D − 1) ⊂ JC of this pair is a 6-dimensional ppav of exponent 6. Thus, if ΘC denotes the Riemann theta divisor on JC, then ΘCP (C,D) ≡ 6 · Ξ, where Ξ is a principal polarization on P (C, D). Since the map π has 24 branch points corresponding to choosing 24 roots in E6 specifying the local monodromy at each branch point, the Hurwitz scheme Hur parameterizing degree 27 covers π : C → P1 with W (E6) monodromy as above is 21-dimensional (and also irreducible, see [Kan06]). The geometric construction described above induces the Prym-Tyurin map P T : Hur → A6 between two moduli spaces of the same dimension. The following theorem answers a conjecture raised by Kanev, see also [LR08, Remark 5.5]: Theorem 0.1. The Prym-Tyurin map P T : Hur → A6 is generically finite. It follows that the general principally polarized abelian variety of dimension 6 is a Prym-Tyurin-Kanev (PTK) variety of exponent 6 corresponding to a W (E6)-cover C → P1. This result, which is the main achievement of this paper, gives a structure theorem for general abelian varieties of dimension 6 and offers a uniformization for A6 by curves with additional discrete data. Just like the classical Prym map P : R6 → A5, it is expected that the Prym-Tyurin map P T will open the way towards a systematic study of abelian 6-folds and their moduli space. What is essential is less the fact that a general 6-dimensional ppav is a PTK variety, but rather the rich geometric structure that Theorem 0.1 provides, which is then of use for other applications presented in Sections 5-11. An immediate consequence of Theorem 0.1 is the following: Corollary 0.2. For every ppav [A, Θ] ∈ A6, the class 6 · θ5/5! ∈ H 10(A, Z) is represented by an effective curve. THE UNIFORMIZATION OF A6 3 It is expected that for a general [A, Θ] ∈ A6, the minimal cohomology class θ5/5! is not even algebraic. Coupled with Corollary 0.2, this would mean that [A, Θ] should not admit any Prym- Tyurin realization of exponent relatively prime to 6. The main idea of the proof of Theorem 0.1 is to study degenerations of PTK varieties as the branch locus (P1, p1 +··· + p24) of the cover π : C → P1 approaches a maximally degenerate point of M0,24. The map P T becomes toroidal and its essential properties can be read off a map of fans. Then, to show that P T is dominant, it is sufficient to show that the rays in the fan describing the image span a 21-dimensional vector space, i.e. that a certain (21 × 21)-matrix has full rank. This can be done by an explicit computation, once the general theory is in place. The theory of degenerations of Jacobians [Ale04] and Prym varieties in [ABH02] is known. One of the main goals of the present paper is an extension of the theory to the case of PTK varieties. For our purposes we do not require the answer to the more delicate problem of understanding the indeterminacy locus of the period map. The remainder of this work focuses on several birational problems that are related to the geom- etry of A6 by Theorem 0.1, and on several quite non-obvious parallels between the Prym map and the Prym-Tyurin map P T . Consider the space H classifying E6-covers [π : C → P1, p1, . . . , p24] together with a labeling of the set of their 24 branch points. In view of the structure Theorem 0.1, it is of compelling interest to understand the birational geometry of this space. It admits a compactification H which is the moduli space of twisted stable maps from curves of genus zero into the classifying stack BW (E6), that is, the normalization of the stack of admissible covers with monodromy group W (E6) having as source a nodal curve of genus 46 and as target a stable 24-pointed curve of genus 0 (see Section 5 for details). One has a finite morphism In Section 6, we show that the canonical class of H is big (Theorem 6.22). From the point of view of A6, it is more interesting to study the global geometry of the quotient space b : H → M0,24. Hur := H/S24, g compactifying the Hurwitz space Hur of E6-covers (without a labeling of the branch points). The Prym-Tyurin map P T extends to a regular morphism P T Sat : Hur → ASat to the Satake compact- ification ASat of A6. Denoting by Ag := Aperf the perfect cone (first Voronoi) compactification of Ag, we establish the following result on the birational geometry of Hur, which we regard as a compact master space for ppav of dimension 6: 6 6 Theorem 0.3. There exists a boundary divisor E of Hur that is contracted by the Prym-Tyurin map P T : Hur (cid:57)(cid:57)(cid:75) A6, such that KHur + E is a big divisor class. The proof of Theorem 0.3 is completed after numerous preliminaries at the end of Section 9. In the course of proving Theorem 0.3, we establish numerous facts concerning the geometry of the space Hur. One of them is a surprising link between the splitting of the rank 46 Hodge bundle E on Hur into Hodge eigenbundles and the Brill-Noether theory of E6-covers, see Theorem 9.3. For a point [π : C → P1] ∈ Hur, we denote by D : H 0(C, ωC) → H 0(C, ωC) the map induced at the level of cotangent spaces by the Kanev endomorphism and by H 0(C, ωC) = H 0(C, ωC)(+1) ⊕ H 0(C, ωC)(−5), the decomposition into the (+1) and the (−5)-eigenspaces of holomorphic differentials respectively. Setting L := π∗(OP1(1)) ∈ W 1 27(C), for a general point [π : C → P1] ∈ Hur, we show that the 4 V. ALEXEEV, R. DONAGI, G. FARKAS, E. IZADI, AND A. ORTEGA following canonical identifications hold: and H 0(C, ωC)(+1) = H 0(C, L) ⊗ H 0(C, ωC ⊗ L∨) H 0(C, ωC)(−5) = (cid:18) H 0(C, L⊗2) (cid:19)∨ 2(cid:94) Sym2H 0(C, L) H 0(C, L). ⊗ In particular, the (+1)-Hodge eigenbundle is fibrewise isomorphic to the image of the Petri map happens generically along Hur, see Theorem 9.2). The identifications above are instrumental in µ(L) : H 0(C, L) ⊗ H 0(C, ωC ⊗ L∨) → H 0(C, ωC), whenever the Petri map is injective (which expressing in Section 9 the class of the (−5)-Hodge eigenbundle E(−5) on a partial compactification GE6 of Hur in terms of boundary divisors. The moduli space GE6 differs from Hur only along divisors that are contracted under the Prym-Tyurin map. Note that the class λ(−5) = c1(E(−5)) is equal to the pull-back P T ∗(λ1) of the Hodge class λ1 on A6. The explicit realization of the class λ(−5) is then used to establish positivity properties of the canonical class KHur. An obvious question is to what extent the geometry of Hur can be used to answer the notorious problem on the Kodaira dimension of A6. Recalling that P T : Hur (cid:57)(cid:57)(cid:75) A6 denotes the extension of the Prym-Tyurin map outside a codimension 2 subvariety of Hur, the pull-back divisor P T ∗(∂A6) contains a unique boundary divisor DE6 of Hur that is not contracted by P T . The statement that A6 is of general type is then equivalent to the bigness of the divisor class 7λ(−5) − [DE6] on Hur (see Corollary 6.3 for a more precise statement). Theorem 0.1 implies that λ(−5) is a big class on Hur, which is a weaker result. Note that it has been established in [FV16] that the boundary divisor ∂A6 of the perfect cone compactification A6 is unirational. geometry of the Abel-Prym-Tyurin curve ϕ(−5) = ϕH 0(ωC )(−5) : C → P5 given by the linear system of (−5)-invariant holomorphic forms on C. Theorem 0.4. An E6-cover [π : C → P1] ∈ Hur such that the Petri map µ(L) is injective lies in the ramification divisor of the map P T : Hur → A6 if and only if the Abel-Prym-Tyurin curve ϕ(−5)(C) ⊂ P5 lies on a quadric. We are also able to describe the ramification divisor of the Prym-Tyurin map in terms of the The conclusion of Theorem 0.4 can be equivalently formulated as saying that the map Sym2H 0(C, ωC)(−5) −→ H 0(C, ω⊗2 C ) given by multiplication of sections is not injective. Note the striking similarity between this description of the ramification divisor of the Prym-Tyurin map and that of the classical Prym map P : Rg+1 → Ag, see [Bea77]: A point [C, η] ∈ Rg+1 lies in the ramification divisor of P if and only if the multiplication map for the Prym-canonical curve Sym2H 0(C, ωC ⊗ η) → H 0(C, ω⊗2 C ) is not injective. An important difference must however be noted. While the general Prym-canonical map ϕωC⊗η : C → Pg−2 is an embedding when g ≥ 5, the Abel-Prym-Tyurin map ϕ(−5) : C → P5 sends the ramification points lying over a branch point of the cover π : C → P1 to the same point of P5 (see Section 10 below). It is natural to ask in what way the Prym-Tyurin-Kanev (PTK) varieties considered in this paper generalize classical Prym varieties. It is classical [Wir95] that the Prym variety of the Wirtinger (cid:48)(cid:48) 0 ⊂ Rg+1 is the cover of a 1-nodal curve of genus g is the Jacobian of its normalization. Thus, if ∆ boundary divisor of such covers and P : Rg+1 (cid:57)(cid:57)(cid:75) Ag is the extension of the Prym map outside THE UNIFORMIZATION OF A6 5 (cid:48)(cid:48) 0) contains the closure of the Jacobian locus in Ag. a codimension 2 subvariety of Rg, then P (∆ In particular, Jacobians arise as limits of Prym varieties. We generalize this situation and explain how ordinary Prym varieties appear as limits of PTK varieties. Via the Riemann Existence Theorem, a general E6-cover π : C → P1 is determined by a branch divisor p1 +··· + p24 ∈ Sym24(P1) and discrete data involving a collection of roots r1, . . . , r24 ∈ E6 which describe the local monodromy of π at the points p1, . . . , p24. Letting two branch points, say p23 and p24, coalesce such that r23 = r24, whereas the reflections in the remaining roots r1, . . . , r22 span the Weyl group W (D5) ⊂ W (E6), gives rise to a boundary divisor DD5 of Hur. We show in Section 8 that the general point of DD5 corresponds to the following geometric data: (i) A genus 7 Prym curve [Y, η] ∈ R7, together with a degree 5 pencil h : Y → P1 branched simply along the divisor p1 +··· + p22; the unramified double cover F1 → Y gives rise to a degree 10 map π1 : F1 → P1 from a curve of genus 13. (ii) A genus 29 curve F2 ⊂ F (5) , which is pentagonally related to F1, and is thus completely determined by F1. Precisely, F2 is one of the two irreducible components of the locus 1 inside the symmetric power F (5) (iii) A distinguished point q1 + ··· + q5 ∈ F2, which determines 5 further pairs of points of F1. One has a degree 16 cover π2 : F2 → P1 induced by π1. : π1(x1) = ··· = π1(x5)(cid:9) (cid:8)x1 + ··· + x5 ∈ F (5) (cid:0)qi, q1 + ··· + ι(qi) + ··· + q5 1 1 (cid:1) ∈ F1 × F2 for i = 1, . . . , 5, which get identified. To F2 we attach a rational curve F0 at the point q1 +··· + q5. The resulting nodal curve C1 = F0 ∪ F1 ∪ F2 has genus 46 and admits a map π : C1 → P1 of degree 27 with πFi = πi for i = 0, 1, 2, where π0 is an isomorphism. The map π can easily be turned into an E6-admissible cover having as source a curve stably equivalent to C1. A general point of the divisor DD5 is realized in this way. We show in Section 8 that P T ([C1, π]) = P ([F1/Y ]) = P ([Y, η]) ∈ A6; furthermore, the general 6-dimensional Prym variety from P (R7) ⊂ A6 appears in this way. We summarize the above discussion, showing that the restriction P TDD5 of the Prym-Tyurin map factors via the (generically injective) Prym map P : R7 (cid:57)(cid:57)(cid:75) A6 in the following way. Theorem 0.5. If DD5 ⊂ Hur is the boundary divisor of W (D5)-covers defined above, one has the following commutative diagram: (0.1) DD5 Hur (cid:0)P [F1/Y ](cid:1) of the Prym-Tyurin map P TD5 : DD5 The fibre P T −1 Prym curve [F1/Y ] ∈ R7 is the fibration over the curve W 1 over a pencil A ∈ W 1 / A6 D5 P P TD5  R7 P T 5 (Y ) the curve F2 obtained by applying the 5-gonal construction to A. We close the introduction by discussing the structure of the paper. In Section 1 we discuss Kanev's construction, whereas in Section 2 we collect basic facts about the E6 lattice and the group W (E6) that are used throughout the paper. After recalling the theory of degenerations of Jacobians and ordinary Prym varieties in Section 3, we complete the proof of Theorem 0.1 in Section 4, by describing the Prym-Tyurin map in the neighborhood of a maximally degenerate (cid:57)(cid:57)(cid:75) R7 over a general genus 7 5 (Y ) of degree 5 pencils on Y with fibre / /    / 6 V. ALEXEEV, R. DONAGI, G. FARKAS, E. IZADI, AND A. ORTEGA point of the space Hur of E6-admissible covers. Sections 5 and 6 are devoted to the birational geometry of this Hurwitz space. The most important result is Theorem 6.17 describing the Hodge class λ on Hur in terms of boundary divisors. In Section 7 we completely describe the extended Prym-Tyurin map P T : Hur (cid:57)(cid:57)(cid:75) A6 to the perfect cone (first Voronoi) toroidal compactification of A6 at the level of divisors and show that only three boundary divisors of Hur, namely DE6, Dsyz and Dazy are not contracted by the map PT (Theorem 7.17). After proving Theorem 0.5 in Section 8, we complete in Section 9 the proof of Theorem 0.3 after a detailed study of the divisors Dazy and Dsyz of azygetic and syzygetic E6-covers respectively on a partial compactification GE6 of Hur. The ramification divisor of the Prym-Tyurin map is described in Section 10. Finally, in Section 11, we prove by degeneration a Petri type theorem on Hur. Acknowledgments: We owe a great debt to the work of Vassil Kanev, who first constructed the Prym-Tyurin map P T and raised the possibility of uniformizing A6 in this way. The authors acknowledge partial support by the NSF: VA under grant DMS 1200726, RD under grant DMS 1603526, EI under grant DMS-1103938/1430600. The work of GF and AO has been partially supported by the DFG Sonderforschungsbereich 647 "Raum-Zeit-Materie". 1. Kanev's construction and Prym-Tyurin varieties of E6-type Consider a cubic threefold X ⊂ P4 and a smooth hyperplane section S ⊂ X. The cubic surface S contains a set of 27 lines Λ := {(cid:96)s}1≤s≤27 forming a famous classical configuration, which we shall review below in Section 2. Consider the lattice ZΛ = Z27 with the standard basis corresponding to 1.1. By assigning to each line (cid:96)s the sum(cid:80){s(cid:48): (cid:96)s·(cid:96)s(cid:48) =1} (cid:96)s(cid:48) of the 10 lines on S intersecting (cid:96)s, we (cid:96)s's, and let deg : ZΛ → Z be the degree homomorphism, so that deg((cid:96)s) = 1 for all s = 1, . . . , 27. define a homomorphism D(cid:48) following quadratic equation: Λ : Z27 → Z27 of degree 10. It is easy to check that D(cid:48) Λ satisfies the (D(cid:48) Λ + 5)(D(cid:48) Λ − 1) = 5 (cid:33) (cid:32) 27(cid:88) s=1 (cid:96)s · deg The restriction DΛ of D(cid:48) the lines lying on St; Consider a generic pencil {St}t∈P1 of cubic hyperplane sections of X. This defines: Λ to the subgroup Ker(deg) satisfies the equation (DΛ + 5)(DΛ − 1) = 0. • a degree 27 smooth curve cover π : C → P1; the points in the fiber π−1(t) correspond to • a symmetric incidence correspondence (cid:101)D ⊂ C × C. Let pi : (cid:101)D → C denote the two projections. Then (cid:101)D has degree deg(p1) = deg(p2) = 10; • a homomorphism D(cid:48) = p2∗ ◦ p∗ • the restriction D of D(cid:48) to JC = Pic0(C), satisfying (D + 5)(D − 1) = 0. 1 : Pic(C) → Pic(C) satisfying the following quadratic equa- For a generic such pencil the map π : C → P1 has 24 branch points on P1, corresponding to singular cubic surfaces in the pencil, each with one node. Over each of the 24 points, the fibre consists of 6 points of multiplicity two and 15 single points. By the Riemann-Hurwitz formula, we compute g(C) = 46. − 1) = 5π−1(0) · deg; tion (see also [Kan89b]): (D(cid:48) + 5)(D(cid:48) 1.2. We refer to [Kan89b, LR08] for the following facts. The cover π : C → P1 is not Galois. The Galois group of its Galois closure is W (E6), the reflection group of the E6 lattice. As we shall review in Section 2, the lattice E6 appears as the lattice K⊥ S ⊂ Pic(S). The 27 lines can be identified with THE UNIFORMIZATION OF A6 7 the W (E6)-orbit of the fundamental weight ω6, and one has a natural embedding W (E6) ⊂ S27. The intermediate non-Galois cover C → P1 is associated with the stabilizer subgroup of ω6 in W (E6), that is, with the subgroup W (E6) ∩ S26 ∼= W (D5). 1.3. By Riemann's Existence Theorem, a 27-sheeted cover C → P1 ramified over 24 points is defined by a choice of 24 elements wi ∈ S27 satisfying w1 ··· w24 = 1. For a cover coming from a pencil of cubic surfaces, each wi ∈ W (E6) is a reflection in a root of the E6. It is a double-six, that is, viewed as an element of S27, it is a product of 6 disjoint transpositions. 1.5. Note that points in the space Hur correspond to covers where we do not choose a labeling of Definition 1.4. Let Hur be the Hurwitz space parametrizing irreducible smooth Galois W (E6)- covers (cid:101)C → P1 ramified in 24 points, such that the monodromy over each point is a reflection in the branch points. The data for the cover (cid:101)C consists of the branch divisor p1 + . . . + p24 on P1, and, for each of these points, the monodromy wi ∈ W (E6) given by a reflection in a root, once a base point p0 ∈ P1 and a system of arcs γi in π1(P1 \ {p1, . . . , p24}, p0) with γ1 ··· γ24 = 1 has been chosen. The elements {wi}24 i=1 generate W (E6) and satisfy the relation w1 ··· w24 = 1. The monodromy data being finite, the space Hur comes with a finite unramified cover a root of the E6 lattice. to the moduli space of 24 unordered points on P1. Thus dim(Hur) = 21. An important fact about this space is the following result of Kanev [Kan06]: br : Hur → M0,24/S24 each branch point is a reflection of W (E6). set of 27 lines on a cubic surface. The incidence of lines, in the same way as for the correspondence Theorem 1.6. For any irreducible root system R, the Hurwitz scheme parameterizing Galois W (R)-covers such that the monodromy around any branch point is a reflection in W (R), is irre- ducible. 1.7. In particular, the space Hur is irreducible. If [(cid:101)π : (cid:101)C → P1] ∈ Hur, let π : C → P1 be an intermediate non-Galois cover of degree 27, that is, the quotient of (cid:101)C by a subgroup W (E6)∩S26 ∼= W (D5) in S27. Since W (E6) acts transitively on the set {1, . . . , 27}, the 27 subgroups S26 ⊂ S27 are conjugate, and the corresponding curves C are isomorphic. Thus, Hur is also a coarse moduli space for degree 27 non-Galois covers π : C → P1, branched over 24 points such that the monodromy at 1.8. Let π : C → P1 be an E6-cover as above. Each fiber of π can be identified consistently with the DΛ in 1.1, induces a symmetric correspondence (cid:101)D ⊂ C × C of degree 10, which is disjoint from the diagonal ∆ ⊂ C × C. In turn, (cid:101)D induces a homomorphism D(cid:48) : Pic(C) → Pic(C), whose component of the identity P T (C, D) :=(cid:0)Ker(D + 5)(cid:1)0 = Im(D − 1) ⊂ JC. 0 = (cid:101)D · ∆ = 2deg((cid:101)D) − 2tr(cid:8)D : H 0(C, ωC) → H 0(C, ωC)(cid:9), 1.10. Using [Kan87], Equation (1.1) implies that the restriction of the principal polarization ΘC of JC to P T (C, D) is a multiple of a principal polarization. Precisely, ΘCP T (C,D) = 6 · Ξ, where (P T (C, D), Ξ) is a ppav. Since restriction D : JC → JC to the degree zero part JC := Pic0(C) satisfies the quadratic relation (1.1) Definition 1.9. The Prym-Tyurin-Kanev (PTK) variety P T (C, D) is defined as the connected (D − 1)(D + 5) = 0 ∈ End(JC). 8 we obtain that V. ALEXEEV, R. DONAGI, G. FARKAS, E. IZADI, AND A. ORTEGA (1.2) dim P T (C, D) = 1 6 = 1 6 (46 − 10) = 6, see also [LR08, Proposition 5.3]. We have the morphism of moduli stacks (cid:16) (cid:17) g(C) − deg((cid:101)D) P T : Hur −→ [C, D] A6 (cid:55)−→ [P T (C, D), Ξ]. Both stacks are irreducible and 21-dimensional. The main result of this paper (Theorem 0.1) is that P T is a dominant, i.e., generically finite, map. 1.11. Our main concrete examples of E6-covers of P1 are the curves of lines in Lefschetz pencils of cubic surfaces. The subvariety T ⊂ Hur corresponding to pencils {St}t∈P1 of hyperplane sections of cubic 3-folds X ⊂ P4 has expected dimension (cid:18)7 (cid:19) − 1 + dim Gr(2, 5) − dim PGL5 = (35 − 1) + 6 − (25 − 1) = 16. 3 1.12. We now describe the restriction of the map P T to the locus T ⊂ Hur parametrizing such covers. Let V be a 5-dimensional vector space over C whose projectivization contains X and let F ∈ Sym3(V ∨) be a defining equation for X. Denote by F := F(X) the Fano variety of lines in X. Let JX := H 2,1(X)∨/H3(X, Z) be the intermediate Jacobian of X. It is well known [CG72] that the Abel-Jacobi map defines an isomorphism JX ∼= AlbF, where AlbF is the Albanese variety of F. Let Λ be a Lefschetz pencil of hyperplane sections of X and denote by E its base curve. The curve C classifying the lines lying on the surfaces contained in Λ lives naturally in F. The map sending a line to its point of intersection with E induces a degree 6 cover C → E. Furthermore, the choice of a base point of C defines a map C → JX. So we obtain a well-defined induced map JC → E × JX. The transpose E × Pic0(F) = E × JX → JC of this map is given by pull-back on divisors on each of the factors, using the map C → E and the embedding C (cid:44)→ F respectively. On the locus T we can explicitly determine the PTK variety: Lemma 1.13. The map JC → E× JX (or its transpose E× JX → JC) induces an isomorphism of ppav P T (C, D) Proof. We first show that the correspondence D restricts to multiplication by (−5) on both factors by H(cid:96) the hyperplane spanned by E and (cid:96) and put S(cid:96) := H(cid:96) ∩ X. The lines incident to E and (cid:96) form 5 pairs ((cid:96)1, (cid:96)(cid:48) E and JX. For (cid:96) ∈ C, let (cid:101)D((cid:96)) be the sum of the lines incident to (cid:96) and E inside X. We denote ∼=→ E × JX. 5), with (cid:96) + (cid:96)i + (cid:96)(cid:48) 1), . . . , ((cid:96)5, (cid:96)(cid:48) i ∈ − KS(cid:96) for i = 1, . . . , 5. Consider first the intermediate Jacobian JX. We have (cid:101)D((cid:96)) = 5(cid:88) i=1 ((cid:96)i + (cid:96)(cid:48) i) ≡ 5 − KS(cid:96) − 5(cid:96), Consider the elliptic curve E. Then (cid:101)D((cid:96)) in E is the sum of the intersection points of (cid:96)i, (cid:96)(cid:48) (cid:80)5 of X from (cid:96)). Therefore(cid:80)5 where ≡ denotes linear equivalence in S(cid:96). Since − KS(cid:96) is constant as (cid:96) varies, it follows that D restricts to multiplication by (−5) on JX. i with E. Note that ((cid:96) + (cid:96)i + (cid:96)(cid:48) i(cid:105) with E. Hence i)E is the intersection of the 5 planes Π1, . . . , Π5 with E. Projecting from (cid:96), we see that the union of these planes is the intersection of H(cid:96) with the inverse image Q of the plane quintic in P2 = P(V /(cid:96)) parametrizing singular conics (the discriminant curve for the projection i)E is contained in the intersection Q ∩ E and since the i)E is also the intersection of the plane Πi := (cid:104)(cid:96), (cid:96)i, (cid:96)(cid:48) i=1((cid:96) + (cid:96)i + (cid:96)(cid:48) i=1((cid:96) + (cid:96)i + (cid:96)(cid:48) 9 two divisors have the same degree, we obtain that(cid:80)5 THE UNIFORMIZATION OF A6 i=1((cid:96) + (cid:96)i + (cid:96)(cid:48) i)E = Q ∩ E is constant. This implies that D is multiplication by (−5) on E as well. So the PTK variety is isogenous to E × JX. To show that they are isomorphic, we show that the pull-back of the polarization of JC to E × JX is 6 times a principal polarization. This is immediate on the factor E, since the map C → E has degree 6. To see it on the JX factor as well, we again use the Abel-Jacobi embedding C (cid:44)→ F (cid:44)→ JX and recall the fact [CG72] that one model of the theta divisor in JX is the image of the degree 6 difference map ψ : F × F → AlbF = JX, (cid:3) defined by ψ((cid:96), (cid:96)(cid:48)) = (cid:96) − (cid:96)(cid:48). We denote by IJ 5 the closure in A5 of the moduli space of intermediate Jacobians of cubic threefolds. We have the following result: Corollary 1.14. We have the following equality of 11-dimensional irreducible cycles in A6: P T (T ) = IJ 5 × A1 ⊂ A5 × A1 ⊂ A6, where the closure on the left hand side is taken inside A6. 2. The E6 lattice Chapters 8,9]. 2.1. Let I 1,6 be the standard Lorenzian lattice with the quadratic form x2 In this section we recall basic facts about the E6 lattice. Our reference for these is [Dol12, (cid:80)6 i=1 x2 i . The negative definite E6 lattice is identified with k⊥, where k = (−3, 1, . . . , 1). Its dual E∨ 6 is identified with (cid:1) + 1 = 36 pairs of roots I 1,6/Zk. Let us denote the standard basis of I 1,6 by f0, f1, . . . , f6, to avoid confusion with the edges ei in a graph. The roots of E6 are the vectors with square −2. There are (cid:0)6 corresponding to αij = fi − fj, αijk = f0 − fi − fj − fk and αmax = 2f0 − f1 − . . . − f6. Obviously, if r ∈ E6 is a root then −r is a root as well. The simple roots, corresponding to the E6 Dynkin diagram can be chosen to be r1 = α123, r2 = α12, r3 = α23, r4 = α34, r5 = α45 and r6 = α56. (cid:1) +(cid:0)6 0− 2 3 2.2. The Weyl group W (E6) is the group generated by the reflections in the roots. It has 51,840 elements. The fundamental weights ω1, . . . , ω6 are the vectors in E∨ 6 with (ri, ωj) = δij. The exceptional vectors are the vectors in the W (E6)-orbit of ω6. They can be identified with vectors (cid:96) in I 1,6 satisfying (cid:96)2 = k(cid:96) = −1. There are 6 + 6 + 15 = 27 of them, namely: for ai = fi, i = 1, . . . , 6; bi = 2f0 − f1 − ··· − f6 + fi, cij = f0 − fi − fj, for 1 ≤ i < j ≤ 6. for i = 1, . . . , 6; 2.3. For each root r ∈ E6, there are 15 exceptional vectors that are orthogonal to it, 6 exceptional vectors with r · (cid:96) = 1 and 6 vectors with r · (cid:96) = −1. The collections of the 6 pairs of exceptional vectors non-orthogonal to a root vector are called double-sixes. The elements in each pair are exchanged by the reflection wr ∈ W (E6) in the root r. There are 36 double-sixes, one for each pair ±r of roots. For example, the double-six for the root r = αmax is {a1, a2, . . . , a6}, {b1, b2, . . . , b6}. The reflection group acts transitively on the set of the exceptional vectors. This gives rise to an embedding W (E6) ⊂ S27. Under this embedding, each reflection corresponds to a product of 6 transpositions. For example, the reflection in the root r = αmax is the permutation (a1, b1)··· (a6, b6) ∈ S27. 10 V. ALEXEEV, R. DONAGI, G. FARKAS, E. IZADI, AND A. ORTEGA Note that the choice of a root is equivalent to an ordering of a pair: when we write the same element of W (E6) as a product (b1, a1)··· (b6, a6), it corresponds to the root −αmax. The W (E6)- action by conjugation is transitive on the set of reflections, i.e., double sixes, so to study their properties it is usually sufficient to make computations for one representative. 2.4. For a smooth cubic surface S, the above objects have the following incarnation: • I 1,6 = Pic(S) together with the intersection form, • k = KS and E6 = K⊥ • the exceptional vectors are identified with the lines (cid:96)1, . . . , (cid:96)27 on S, • a sixer is a set of 6 mutually disjoint lines, a double-six is the set of two sixers corresponding S ⊂ Pic(S), to the opposite roots. The relationship between the W (E6)-action and the correspondence given by the line incidence is as follows. Definition 2.5. The correspondence on the set of exceptional vectors is defined by setting (cid:88) D((cid:96)) := (cid:96)(cid:48). {(cid:96)(cid:48): (cid:96)(cid:48)·(cid:96)=1} Remark 2.6. For further use, we retain the following computation: D(a1) = b2 + ··· + b6 + c12 + ··· + c16 D(b1) = a2 + ··· + a6 + c12 + ··· + c16 D(a1 − b1) = (b2 − a2) + . . . (b6 − a6). 2.7. The group W (E6) has 25 irreducible representations corresponding to its 25 conjugacy classes, which will appear several times in this paper. For conjugacy classes we use the ATLAS or GAP notation 1a, 2a, 2b, 2c, . . . , 12a, (command 'CharacterTable("U4(2).2")'). The number refers to the order of the elements in the conjugacy class. For instance, the reflections in W (E6) (products of six transpositions) belong to the conjugacy class 2c, the product of two syzygetic reflections belongs to the class 2b, whereas the product of two azygetic reflections belongs to the class 3b (see Section 5 for precise definitions). 3. Degenerations of Jacobians and Prym varieties 3.1. By a theorem of Namikawa and Mumford, the classical Torelli map Mg → Ag sending a smooth curve to its Jacobian extends to a regular morphism Mg → Avor from the Deligne- Mumford compactification of Mg to the toroidal compactification of Ag for the second Voronoi fan. See [AB12] for a transparent modern treatment of this result, and extension results for other toroidal compactifications of Ag. The result applies equally to the stacks and to their coarse moduli spaces. Here, we will work with stacks, so that we have universal families over them. g 3.2. At the heart of the result of Namikawa and Mumford lies the Picard-Lefschetz formula for the monodromy of Jacobians in a family of curves, see e.g. [Nam73, Proposition 5]. The map of fans for the toroidal morphism Mg → Avor is described as follows. Fix a stable curve [C] ∈ Mg, and let Γ be its dual graph, with a chosen orientation. Degenerations of Jacobians are described in terms of the groups g Ze, H1(Γ, Z) = Ker(cid:8)∂ : C1(Γ, Z) → C0(Γ, Z)(cid:9). C0(Γ, Z) = Zv, C1(Γ, Z) = (cid:77) (cid:77) vertices v edges e THE UNIFORMIZATION OF A6 11 (3.1) The Jacobian JC = Pic0(C) is a semiabelian group variety, that is, an extension 1 → H 1(Γ, C∗) → Pic0(C) → Pic0((cid:101)C) → 0, The monodromy of a degenerating family of Jacobians is described as follows. Fix a lattice where (cid:101)C is the normalization of C. In particular, Pic0(C) is a multiplicative torus if and only if (cid:101)C is a union of P1's, or equivalently, if b1 = h1(Γ) = g. Λ (cid:39) Zg and a surjection Λ (cid:16) H1(Γ, Z). The rational polyhedral cone for a neighborhood of [C] ∈ Mg lives in the space Λ∨ b1 = h1(Γ) with the rays e∗ H1(Γ, Z) ⊂ C1(Γ, Z) taking the value δij on the edge ej ∈ C1(Γ, Z). The rational polyhedral cone corresponding to a neighborhood of [JC] ∈ Avor Γ2(Λ∨) ⊗ R = (Sym2(Λ) ⊗ R)∨, where the lattice Γ2(Λ∨) is the second divided power of Λ∨. It is a simplicial cone with the rays (e∗ (cid:54)= 0, which means that ei is not a bridge of the graph Γ. We explain what this means in down to earth terms. In an open analytic neighborhood U of [C], one can choose local analytic coordinates z1, . . . , z3g−3 so that the first N coordinates correspond to smoothing the nodes of C, labeled by the edges ei of the graph Γ. Thus, we have a family of smooth curves over the open subset V = U − It is a simplicial cone of dimension i is the linear function on i corresponding to the edges of Γ. Here, e∗ ⊗ R with the lattice Λ∨. Then a complex-analytic map V → Hg to the Siegel upper half-plane is given by a formula (see i=1{zi = 0}. i )2 for all e∗ i lives in the space (cid:83)N g [Nam73, Thm.2] or [Nam76, 18.7]) N(cid:88) i=1 (zi) (cid:55)→ Mi · 1 2π√−1 log zi + (a bounded holomorphic function), where Mi are the g× g matrices with integral coefficients corresponding to the quadratic functions i )2 on Λ (cid:16) H1(Γ, Z). After applying the coordinatewise exponential map (e∗ C g(g+1) 2 → (C∗) g(g+1) 2 , uij (cid:55)→ exp(2π√−1 uij), g . the matrices Mi · (log zi/2π√−1) become Laurent monomials in zi. This monomial map describes a complex-analytic map from a small complex-analytic neighborhood U of [C] ⊂ Mg to an ap- propriate ´etale neighborhood of Ag. For the arguments below the above two formulas suffice. In particular, we do not need to know the indeterminacy locus of the extended maps. Thus, we will not need explicit coordinates near a boundary of Avor 3.3. The following weak form of Torelli's theorem is a sample of our degeneration technique. This is far from being the easiest way to prove the Torelli theorem, but it gives a good illustration of our method which we later apply to PTK varieties. Lemma 3.4. The image of the Torelli map Mg → Ag has full dimension 3g − 3. Proof. For every g, there exists a 3-edge connected trivalent graph Γ of genus g (exercise in graph theory). By Euler's formula, it has 3g − 3 edges. Recall that a connected graph is 2-edge connected if it has no bridges, i.e., the linear functions e∗ i on H1(Γ, Z) are all nonzero, and it is 3-edge connected if for i (cid:54)= j one has e∗ i )2 (cid:54)= (e∗ j )2. Let C be a stable curve whose dual graph is Γ and whose normalization is a disjoint union of P1's. Then the 3g − 3 matrices Mi in Formula (3.2), i.e. the functions (e∗ i )2, are linearly independent in Sym2(Zg), cf. [AB12, Remark 3.6]. By looking at the leading terms as zi → 0, this easily implies that the image has full dimension 3g − 3. i (cid:54)= ±e∗ j , i.e., (e∗ 12 V. ALEXEEV, R. DONAGI, G. FARKAS, E. IZADI, AND A. ORTEGA After applying the exponential function, the map becomes (z1, . . . , z3g−3) (cid:55)→ (monomial map) × (invertible function). Since the monomial part is given by monomials generating an algebra of transcendence degree (cid:3) 3g − 3, the image is full-dimensional. Remark 3.5. Note that the regularity of the extended Torelli map Mg → Avor g played no role in the proof of Lemma 3.4. All we need for the conclusion is the fact that the monodromy matrices Mi are linearly independent. 3.6. The theory for Jacobians was extended to the case of Prym varieties in [ABH02]. We briefly recall it. Let Rg be the stack of Prym curves of genus g, classifying admissible pairs [C, ι] consisting of a stable curve with involution ι : C → C, so that C/ι is a stable curve of genus g and the map C → C/ι is an admissible map of stable curves. We refer to [Bea77] and [FL10] for background on Rg. Consider one pair [C, ι] ∈ Rg and a small analytic neighborhood U of it. As before, Γ is the dual graph of C. Then the space H1(C, Z) of the Jabobian case is replaced by the lattice H1/H + 1 and H− 1 are the (+1)- and the (−1)-eigenspaces of the involution action ι∗ on H1(C, Z) respectively. Via the natural projection H1 (cid:16) H1/H + 1 with a finite index sublattice of H1/H + 1 . ι∗ : Pic0(C) → Pic0(C). P (C, ι) = Ker(1 + ι∗)0 = Im(1 − ι∗), The degeneration of Prym varieties as groups is 1 , we identify H− 1 . Here, H + The monodromy of a degenerating family of Prym varieties is obtained by restricting the mon- odromy map for JC to the (−1)-eigenspace. Combinatorially, it works as follows: For every edge ei of Γ we have a linear function e∗ 1 , the restriction of the linear function on H1(C, Z). For the divisor {zi = 0} on U corresponding to smoothing the node Pi of C, the mon- i )2 restricted to H1(Γ, Z)−. Similarly to Lemma 3.4, this odromy is given by the quadratic form (e∗ can be used to prove various facts about the Prym-Torelli map, but we will not pursue it here. i on the group H− 4. Degenerations of Prym-Tyurin-Kanev varieties We choose a concrete boundary point in a compactification of the Hurwitz scheme Hur. We start with a single cubic surface S and the set {(cid:96)1, . . . , (cid:96)27} of 27 lines on it. Sometimes we shall use the Schlafli notation {ai, bi, cij} for them, as in Section 2. We fix an embedding of W (E6) into the symmetric group S27 permuting the 27 lines on S. 4.1. We choose 12 roots ri which generate the root system E6. Let wi ∈ W (E6) be the reflections in ri; they generate W (E6). As we saw in Section 2, each wi is a double-six. Fixing the root ri gives it an orientation. 4.2. Consider a nodal genus 0 curve E whose normalization is a union of P1's and whose dual graph is the tree T shown in the left half of Figure 1. The 24 ends of this tree correspond to 24 points p1, . . . , p24 on E. We label the points by roots r1, . . . , r12. Each of the outside vertices has two ends, we use the same label ri for both of them. Definition 4.3. Let π : C → E be an admissible 27 : 1 cover ramified at the point pi with monodromy wi for i = 1, . . . , 24. For every irreducible component of E, the product of the monodromy elements equals 1; this count includes the nodes. Since we required that for every component on the boundary the two wi's are the same, the map is unramified at the nodes. Thus, π is ´etale over E \ {p1, . . . , p24}. THE UNIFORMIZATION OF A6 13 Figure 1. The tree T for the target curve E of genus 0 4.4. Here is a concrete description of the dual graph Γ of C. It has 10 × 27 + 12 × (6 + 15) vertices and 21 × 27 edges Each vertex v of T in the ´etale part has 27 vertices over it. Over each of the outside 12 vertices, there are 6 vertices, where the map P1 → P1 is 2 : 1 and ramified at a pair of the points pi and pi+12, and 15 other vertices where the map P1 → P1 is 1 : 1. All the nodes of E lie in the ´etale part, so for each internal edge e of the tree T there are 27 edges of Γ. 4.5. The graph Γ is homotopically equivalent to the following much simpler graph Γ(cid:48). It has: genus 46. s=1, labeled by the lines on S. (Here, s stands for "sheets".) (1) 27 vertices {vs}27 (2) 12 × 6 edges eij. For each of the twelve roots ri, there are 6 edges. For example, for r = rmax, the edges are (a1, b1), . . . , (a6, b6). The first edge is directed from a1 to b1, etc. The graph Γ(cid:48) is obtained from Γ by contracting the tree in each sheet to a point, and removing the middle vertex of degree 2 for each of the 12 × 6 paths corresponding to the double-sixes. The process is illustrated in the right half of Figure 1. By Euler's formula, the genus of Γ is 12 × 6 − 27 + 1 = 46. Thus, the curve C has arithmetic 4.6. Next we define a symmetric correspondence (cid:101)D ⊂ C × C of degree 10, as follows. To each This defines the curve (cid:101)D0 ⊂ C 0 × C 0, where C 0 = C \ π−1{p1, . . . , p24}. The correspondence (cid:101)D ⊂ C × C is the closure of (cid:101)D0. Let pi be a ramification point with monodromy wi. Without loss point Q ∈ C over the ´etale part in the sheet labeled (cid:96)i, associate 10 points in the same fiber of π that are labeled (cid:96)ij by the lines that intersect (cid:96)i. of generality, we may assume w = wmax. The points in the fiber π−1(pi) are labeled a1b1, . . . , a6b6 and cij for i (cid:54)= j. Then the correspondence is described by: a1b1 (cid:55)→ (aibi + c1i), c12 (cid:55)→ a1b1 + a2b2 + cij, etc. 6(cid:88) i=2 (cid:88) i,j(cid:54)=1,2 r1r1r2r2r3r3r4r4r5r5r6r6r7r7r8r8r9r9r10r10r11r11r12r12...−→a1b1a6b6...cij...a1b1a6b6...cij14 V. ALEXEEV, R. DONAGI, G. FARKAS, E. IZADI, AND A. ORTEGA Lemma 4.7. There exists an analytic neighborhood U ⊂ M0,24 of the point [E, p1, . . . , p24] and a family of covers πt : Ct → Et together with correspondences (cid:101)Dt ⊂ Ct × Ct over U , which extends π : C → E and (cid:101)D. Proof. Since the map π is ´etale over each node of E, the families Ct and (cid:101)Dt extend naturally. Lemma 4.8. The correspondence (cid:101)D ⊂ C × C induces an endomorphism of the homology group D : H1(Γ, Z) → H1(Γ, Z) satisfying the relation (D − 1)(D + 5) = 0. The (−5)-eigenspace H (−5) The monodromy data determine the Ct's as topological spaces. Then the finite map Ct → Et (cid:3) determines a unique structure of an algebraic curve on Ct. can be naturally identified with Ker(φ), where 1 12(cid:77) i=1 φ : ZRi → E6, Ri (cid:55)→ ri. canonically identified. The group C0(Γ(cid:48), Z) of vertices is(cid:76)27 Here, Ri is a basis vector for the (−5)-eigenspace for the action of D on the rank 6 lattice generated by the edges of Γ(cid:48) above the root ri. Since the vectors ri generate E6, one has rk H (−5) Proof. We will work with the graph Γ(cid:48) defined in 4.5, since the homology groups of Γ and Γ(cid:48) are Zvi. The endomorphism D0 on it is defined in the same way as the correspondence on the 27 lines. The induced endomorphism D1 on C1(Γ(cid:48), Z) is the following. Pick one of the roots ri. Without loss of generality, let us assume r = αmax. Then 1 = 6. i=1 D1(a1, b1) = −(a2, b2) − . . . − (a6, b6). By Remark 2.6, D commutes with ∂, so defines an endomorphism on H1(Γ(cid:48), Z). The endomorphism D1 on C1(Γ(cid:48), Z) splits into 12 blocks each given by the (6 × 6)-matrix N This gives an identification C1(Γ(cid:48), Z)(−5) = (cid:76)12 such that Nii = 0 and Nij = −1 for i (cid:54)= j. It is easy to see that (N − 1)(N + 5) = 0 and that the (−5)-eigenspace of N is 1-dimensional and is generated by the vector (a1, b1) + . . . + (a6, b6). ZRi. The homomorphism ∂ : C1 → C0 is s=1(ri, es)vs, where es are the 27 exceptional vectors. Since the bilinear form defined by Ri (cid:55)→ (cid:16) 12(cid:88) on E6 is nondegenerate and es span E∨ (cid:16) 12(cid:88) (cid:16) 12(cid:88) (cid:80)27 6 , one has (cid:33) (cid:32) (cid:17) (cid:17) i=1 φ niRi , es i=1 = 0 for s = 1, . . . , 27 ⇐⇒ φ It is an elementary linear algebra exercise to pick an appropriate basis in Ker(φ), which becomes 1 especially easy if r1, . . . , r6 form a basis in E6. Theorem 4.9. The limit of PTK varieties P (Ct, Dt) as a group is the torus (C∗)6 with the character group H (−5) . For each of the 21 internal edges ei of the tree T , the monodromy around the i )∗)2 divisor {zi = 0} in the neighborhood U ⊂ M0,24 is given by the quadratic form Mi =(cid:80)27 on H (−5) Proof. The first statement is immediate: the limit of the Jacobians as a group is a torus with the character group H1(Γ, Z), and the PTK varieties are obtained by taking the (−5)-eigenspace. Every internal edge ei of T corresponds to a node of the curve E. Over it, there are 27 nodes of i for the smoothings of these nodes can be the curve C. The map is ´etale, so the local coordinates zs s=1((es . 1 (cid:17) = 0 ⇐⇒ 1 = C (−5) 1 ∂ niRi i=1 Therefore, H (−5) ∩ Ker(∂) = Ker(φ). niRi = 0. i=1 (cid:3) THE UNIFORMIZATION OF A6 15 identified with the local coordinate zi. By Section 3, the matrix for the monodromy around zs is ((es together and restricting to the (−5)-eigenspace. i = 0 i )∗)2. The monodromy matrix for PTK varieties is obtained by adding these 27 matrices (cid:3) i )∗ on H1(Γ, Z), we unwind the identification H1(Γ, Z) = H1(Γ(cid:48), Z). Zej be the map which associates to Ri the oriented path in the tree T of Figure 1 from the central point O to an end labeled ri. Via the identification To compute the linear forms (es Lemma 4.10. Let p : (cid:76)12 (cid:76)12 ZRi → H1(Γ, Z)(−5) = Ker(φ) ⊂ i )∗ are defined by the formula (cid:76)21 k=1 j=1 i=1 ZRk, the linear functions (es (es i )∗(Rk) = (cid:104)rk, (cid:96)s(cid:105) · (cid:104)p(Rk), e∗ i(cid:105), 6 → Z, and for the second one (cid:104)ej, e∗ where the first pairing is E6 × E∗ Proof. Let (vs1, vs2) be an edge in Γ(cid:48). To it, we associate the path in the graph Γ going from the center of level s1 to the center of level s2: i(cid:105) = δij. (cid:76)12 path(Os1, rs1 1 ) + path(vs1, vs2) − path(Os2, rs2 1 ). This rule gives an identification H1(Γ(cid:48), Z) = H1(Γ, Z). For each of the 12 roots rk, we have 6 edges in the graph Γ(cid:48) going from the vertices s with (cid:104)rk, (cid:96)s(cid:105) = 1 to the vertices s with (cid:104)rk, (cid:96)s(cid:105) = −1. The contribution of Rk to the adjusted cycle therefore is 27(cid:88) s=1(cid:104)rk, (cid:96)s(cid:105) · path(Os, rk) = 27(cid:88) s=1(cid:104)rk, (cid:96)s(cid:105) · p(Rk)(cid:12)(cid:12)ei=es i The value of the linear function es i on it is therefore given by the formula in the statement. (cid:3) To complete the computation, we have to do the following: (1) Choose a basis of the 6-dimensional space H1(Γ, Z)(−5) = Ker(φ) ⊂ (4) And finally compute the 21 monodromy matrices Mi =(cid:80)27 (2) Compute the 21 × 27 linear functions (es (3) Compute the 21×27 quadratic functions ((es i )∗ on this 6-dimensional space. i )∗)2, each of which is a symmetric 6×6-matrix. i )∗)2 of Theorem 4.9. ZRk. s=1((es k=1 Theorem 4.11. There exist collections of E6 roots r1, . . . , r12 generating the lattice E6 for which the 21 symmetric (6 × 6)-matrices Mi of Theorem 4.9 are linearly independent. Proof. A concrete example is r1 = α135, r2 = α12, r3 = α23, r4 = α34, r5 = α45, r6 = α56, r7 = α456, r8 = α26, r9 = α123, r10 = α125, r11 = α256, r12 = α15. An explicit computation using the the formula in Lemma 4.10, aided by a computer algebra system, shows that (1) The monodromy matrices Mi are all divisible by 6. This corresponds to the fact that the restriction of the principal polarization from the Jacobian to the PTK variety is 6 times a principal polarization. (2) For the normalized forms M(cid:48) matrix is 212 (cid:54)= 0. i = Mi/6, the determinant of the corresponding (21 × 21)- (cid:3) A Mathematica notebook with an explicit computation is available at [Web15]. Corollary 4.12. Theorem 0.1 holds. Proof. By the same argument as in the proof of Lemma 3.4, the image of the complex-analytic (cid:3) map U → A6 has full dimension 21. Thus, the map P T : Hur → A6 is dominant. 16 V. ALEXEEV, R. DONAGI, G. FARKAS, E. IZADI, AND A. ORTEGA Remark 4.13. Computer experimentation shows that for a very small portion of random choices of the roots r1, . . . , r12, the matrices Mi are linearly independent. In most of these cases the determinant is 212 but in some cases it is 213. A necessary condition is for the roots r1, r2 to be non-orthogonal, and similarly for the pairs r3, r4, etc. Experimentation also shows that there is nothing special about the graph in Figure 1. Any other trivalent graph with 12 vertices of degree one works no worse and no better. 5. Admissible covers and semiabelian Prym-Tyurin-Kanev varieties In this section, we introduce the space H of admissible E6-covers and define semiabelian Prym- Tyurin-Kanev varieties of E6-admissible pairs. Then we study extensions of the Prym-Tyurin map to the Satake compactification Asat . 5.1. The Hurwitz space. 6 and the perfect cone toroidal compactification A6 := Aperf 6 5.1. We denote by H the Hurwitz space of E6-covers π : C → P1 together with a labeling (p1, . . . , p24) of its branch points. Let H be the compactification of H by admissible W (E6)-covers. By [ACV03], the stack H is isomorphic to the stack of balanced twisted stable maps into the classifying stack BW (E6) of W (E6), that is, H := M0,24 BW (E6) . (cid:16) (cid:17) (cid:0) BW (E6)(cid:1) is a divisor with normal crossings. Points of HME6 By a slight abuse of notation, we shall use the same symbol H both for the stack and for the associated coarse moduli space. For details concerning the local structure of spaces of admissible coverings, we refer to [ACV03]. Note that H is a smooth stack isomorphic to the normalization of the Harris-Mumford moduli space HME6 defined (in the case of covers with Sn-monodromy) in [HM82]. The boundary H \ M0,24 are E6-admissible coverings [π : C → R, p1, . . . , p24], where C and R are nodal curves of genus 46 and 0 respectively, and p1, . . . , p24 ∈ Rreg are the branch points of π. The local monodromy of π around pi ∈ P1 is given by a reflection wi ∈ W (E6), for i = 1, . . . , 24. Let b : H → M0,24 be (cid:102)M0,n := M0,n/Sn, we consider the induced branch and source maps the branch morphism and ϕ : H → M46 be the source morphism. Obviously, S24 acts on H and the projection q : H → Hur is a principal S24-bundle. Passing to the S24-quotient and denoting respectively. For 2 ≤ i ≤ 12, let Bi := (cid:80)T=i δ0:T ∈ Pic(M0,24) be the boundary class, where points lying on one component are precisely those labeled by T . Let (cid:101)Bi be the reduced boundary divisor on (cid:102)M0,24 which pulls-back to Bi under the quotient map M0,24 → (cid:102)M0,24. For each E6-cover [π : C → P1] ∈ Hur, there is an induced Kanev endomorphism at the level of Jacobians D : JC → JC and at the level of differentials D : H 0(C, ωC) → H 0(C, ωC), which we denote by the same symbol. This induces a splitting the sum runs over all subsets T ⊂ {1, . . . , 24} of cardinality i. Recall that δ0:T is the class of the closure of the locus of pointed curves consisting of two rational components, such that the marked br : Hur → (cid:102)M0,24 and ϕ : Hur → M46, H 0(C, ωC) = H 0(C, ωC)(+1) ⊕ H 0(C, ωC)(−5) into (+1) and (−5)-eigenspaces respectively. From (1.2), it follows that dim H 0(C, ωC)(−5) = dim P T (C, D) = 5, THE UNIFORMIZATION OF A6 17 hence dim H 0(C, ωC)(+1) = 40. We have a decomposition of the rank 46 Hodge bundle E := ϕ∗(E) pulled-back from M46 into eigenbundles where, as we pointed out, rk(E(+1)) = 40 and rk(E(−5)) = 6. We set λ(+1) := c1(E(+1)) and λ(−5) := c1(E(−5)), therefore λ := ϕ∗(λ) = λ(+1) + λ(−5). We summarize the discussion in the following diagram: E = E(+1) ⊕ E(−5), (5.1) ϕ / M46 H b M0,24 q Hur br  / (cid:102)M0,24. (D(cid:48) By the above identification, this correspondence satisfies the identity correspondence (cid:101)D ⊂ C × C inducing an endomorphism D : JC → JC of the semiabelian group 5.2. Semiabelian Prym-Tyurin-Kanev varieties. Lemma 5.2. Let [π : C → R, p1, . . . , p24] ∈ H be an E6-admissible cover. Then it comes with a variety JC = Pic0(C), which satisfies the same identity (D−1)(D +5) = 0 as for covers of smooth curves. The group variety P T (C, D) := Im(D − 1) is a semiabelian group subvariety of JC. Proof. Consider any one-parameter family πs : Cs → Rs of E6-admissible covers over a smooth base (S, 0) such that πs is smooth for s (cid:54)= 0 and π0 = π. It gives an identification of the smooth fibers of π with the nearby fibers of πs up to the monodromy W (E6). Thus, we have a symmetric The target curve R is a tree of P1s, so Pic0(R) = 0. Any element of Pic0(C) can be represented degree 10 correspondence (cid:101)D0 on the smooth locus of π, and it clearly does not depend on a chosen one-parameter family. We take (cid:101)D ⊂ C × C to be the closure of (cid:101)D0. − 1)(D(cid:48) + 5) = 5 · Trace(π) . by a divisor on C of multidegree (0, . . . , 0) supported on the smooth locus of C. Thus, (cid:101)D in- duces a correspondence D : Pic0 C → Pic0(C) satisfying (D − 1)(D + 5) = 0. The image of a (cid:3) homomorphism (D − 1) : JC → JC is a semiabelian variety, which finishes the proof. Definition 5.3. Any semiabelian variety G has a unique extension 1 → T → G → A → 0. We call tor(G) = T and ab(G) = A the toric and abelian parts of G respectively, and we call their dimensions the toric and abelian ranks of G. In particular, we shall talk about the toric rank kP T of a Prym-Tyurin-Kanev (PTK) variety P T . Lemma 5.4. Let 1 → T → G → A → 0 be a semiabelian variety with an endomorphism satisfying (D − 1)(D + 5) = 0. Then: (1) P := Im(D − 1) coincides with Ker(D + 5)0, the connected component of identity. (2) D induces homomorphisms DT of the toric part T and DA of the abelian part A. (3) The toric part of P coincides with PT := (DT − 1)T , and the abelian part of P is isogenous to PA := (DA − 1)A. Proof. (1) One has the inclusions Im(D − 1) ⊂ Ker(D + 5)0, Im(D + 5) ⊂ Ker(D − 1)0, and Ker(D − 1) ∩ Ker(D + 5) ⊂ G[6], / /    / / 18 V. ALEXEEV, R. DONAGI, G. FARKAS, E. IZADI, AND A. ORTEGA which is a finite group. Since G is also 6-divisible, Im(D − 1) are Im(D + 5) span G, and so they are semiabelian subvarieties of complementary dimensions. The quotient Ker(D + 5)/ Im(D − 1) is the kernel of a surjective homomorphism (D + 5) : G/ Im(D − 1) → Im(D + 5) of varieties of the same dimension, so it is finite. Thus, Im(D − 1) = Ker(D + 5)0. (2) The homomorphism from the affine variety T to the projective variety A is constant, so D(T ) ⊂ T and we get DT = DT , which in turn induces an endomorphism DA of A = G/T . (3) Clearly, Im(DT − 1) ⊂ Im(D − 1) and Im(D − 1) (cid:16) Im(DA − 1). The kernel of the homomorphism Im(D − 1) → Im(DA − 1) ⊂ A is T ∩ Im(D − 1). One has Im(DT − 1) ⊂ T ∩ Im(D − 1) ⊂ Ker(DT + 5). By (1) applied to T the difference between the last and the first groups is finite. Therefore, the difference between the middle and the first groups is finite. In other words, the homomophism Ker(P/PT → PA) has finite kernel. Thus, P/PT is an abelian variety that is the abelian part of (cid:3) P , it is isogenous to PA, and PT is the toric part of P . The structure of an E6-admissible map π : C → R makes the computation of P T (C, D) espe- cially easy. The target curve R is a tree {Ri (cid:39) P1}n i=1 of smooth rational curves. Then we have induced maps πi : Ci → Ri and correspondences Di for Ci satisfying the same quadratic identity. The curves Ci are smooth; however, they may be disconnected. (1) The abelian part of P T (C, D) is isogenous to(cid:81)n (2) The correspondence (cid:101)D induces correspondences DC0, DC1 and DH1 on the cycle groups C0(Γ, Z), C1(Z) and on the homology group H1(Γ, Z) = Ker(cid:8)C1(Γ, Z) → C0(Γ, Z)(cid:9) of the Lemma 5.5. One has the following: dual graph Γ of C. The toric part of the semiabelian variety P T (C, D) has the character group Im(DH1 − 1). i=1 P (Ci, Di). Proof. (1) follows from Lemma 5.4. The definition and properties of the correspondences on C0(Γ, Z), C1(Γ, Z) and H1(Γ, Z) are immediate. For a homomorphism φ : T1 → T2 of tori with dual homomorphism φ∗ : X2 → X1 of character lattices, Im φ is a torus with character lattice Im φ∗. We apply this to the toric part of JC whose character lattice is H1(Γ, Z) and use Lemma (cid:3) 5.4. 5.3. Extensions of the Prym-Tyurin map. g 5.6. There are several natural targets to consider for the extended Prym-Tyurin map. The easiest one is the Satake-Baily-Borel compactification Asat g = Ag (cid:116) Ag−1 (cid:116) ··· (cid:116) A0 (g = 6 in our case). Other natural targets are the toroidal compactifications Ag = Aperf for the perfect cone, respectively 2nd Voronoi fans. The space Aperf has the advantage of having only one boundary divisor. The space Avor is modular: by [Ale02] it is the normalization of the main component of the moduli space of principally polarized stable semiabelic varieties. All toroidal compactifications of Ag contain the same open subset (cid:101)Ag introduced by Mumford g . Moreover, (cid:101)Ag = Ag (cid:116) (cid:101)Dg, where (cid:101)Dg is the universal Kummer variety over Ag−1. The closure of (cid:101)Dg in Aperf [Mum83], the moduli space of principally polarized abelian varieties of dimension g together with their degenerations of toric rank 1, This is a partial compactification of Ag isomorphic to the blow-up of the open subset ASat divisor Dg. g,tor.rk≤1 = Ag (cid:116) Ag−1 in Asat is the unique boundary g g and Avor g g THE UNIFORMIZATION OF A6 19 On the other hand, by [Mum83], any semiabelian variety G with a principally polarized abelian part of dimension g − 1 has a canonical compactification X, a rank-1 degeneration of ppav. In the language of [Ale02], in this case there exists a unique principally polarized stable semiabelic pair (G (cid:121) X ⊃ Θ). So the moduli of toric rank ≤ 1 semiabelian varieties and the moduli of toric rank ≤ 1 stable semiabelic pairs are the same. Lemma 5.7. Let C → R → S be a family of admissible covers parameterized by H over a reduced scheme S such that the restrictions of C → S, R → S are smooth over an open dense subset. Then there is a semiabelian group scheme PT → S whose fiber over s ∈ S is P T (Cs, Ds) as defined above. Proof. By taking the closure of the correspondence over U , we obtain a correspondence on C whose fibers were described in Lemma 5.2. Thus, we have a semiabelian group scheme Pic0C/S together with an endomorphism D over S giving an endomorphism as in Lemma 5.2 fiberwise. (cid:3) Then PT := Im(D − id) satisfies the conditions of the statement. Lemma 5.8. The map P T : Hur → A6 extends to a regular map P T sat : Hur → Asat [C → R] maps to the abelian part of P T (C, D). Proof. As discussed in the previous section, Hur is a smooth stack and the boundary Hur \ Hur is a divisor with normal crossings. The extension exists by Borel's extension theorem [Bor72]. To find the image of [C → R] it is sufficient to consider a one-parameter family of covers parameterized by (S, 0) with covers of smooth curves for s (cid:54)= 0. It is known that for any family of dimension g semiabelian varieties G → S whose restriction to S \ 0 is a family of principally polarized abelian varieties, the image of 0 ∈ S of the extended map to ASat is the abelian part of G0. We apply this to the family PT → S (cid:3) of the previous lemma. Theorem 5.9. The rational map P T : Hur (cid:57)(cid:57)(cid:75) Aperf map is regular and proper, and it factors through (cid:101)A6. has an indeterminacy locus of codimension at least 2. On the open subset Hurtor.rk≤1 ⊂ Hur where PTK varieties have toric rank ≤ 1, the have a morphism P T : Hurtor.rk≤1 → (cid:101)A6 ⊂ Aperf Proof. The indeterminacy locus has codimension ≥ 2 simply because Hur (viewed as a vari- ety) is normal and Aperf is proper. By the discussion in Paragraph 5.6, we have a regular map Hurtor.rk≤1 → Avor g ◦ P T : Hurtor.rk≤1 → ASat 6,tor.rk≤1 are proper, it follows that P T is proper. 6 . A point 6 6 and Lemma 5.8 implies that the image is contained in (cid:101)A6. Thus, we . Since both maps g : (cid:101)A6 → ASat 6,tor.rk≤1 and (cid:3) 6 6 g 6. Positivity properties of the Hurwitz space of E6-covers In this section we study in detail the divisor theory on the Hurwitz space H. In particular, we express explicitly the Hodge class λ in terms of boundary classes and we prove that the canonical class of H is big. We recall that the pull-back of the Hodge class under the morphism ϕ : Hur → M46, is the sum λ := ϕ∗(λ) = λ(+1) + λ(−5) := c1(E(+1)) + c1(E(−5)) of the first Chern classes of the Hodge eigenbundles for the +1 and the −5 eigenvalues, see Paragraph 5.1. 20 V. ALEXEEV, R. DONAGI, G. FARKAS, E. IZADI, AND A. ORTEGA Lemma 6.1. Both Hodge eigenclasses λ(+1) and λ(−5) ∈ CH 1(Hur) are nef. Proof. Koll´ar [Kol90] showed that the Hodge bundle E is semipositive, therefore the eigenbundles E(+1) and E(−5) as quotients of E are semipositive as well. Therefore det(E(+1)) and det(E(−5)) are (cid:3) nef line bundles. 6.2. The conclusion of Theorem 0.1 can be restated in terms of the positivity of λ(−5). The fact that the map P T : Hur → A6 is dominant implies that the class λ(−5) ∈ CH 1(Hur) is big. One can formulate a necessary and sufficient condition for A6 to be of general type in similar terms. Corollary 6.3. Let Di be the irreducible divisors supported on Hur \ Hurtor.rk≤1. Then to prove that A6 is of general type, it suffices to show that there exist some integers ai such that the divisor P T ∗(KA6 Proof. If this divisor is big on Hur, then its corresponding linear system has maximal Iitaka dimension. Then the linear system P T∗P T ∗(KA6 ) has maximal Iitaka dimension as well. Since (cid:16) = (cid:3) ) +(cid:80) aiDi on Hur is big. all boundary divisors Di are contracted under the Prym-Tyurin map, we write P T∗ P T ∗(KA6 ) +(cid:80) is big. (cid:17) (cid:16) (cid:17) ) = deg(P T )KA6 . So KA6 P T ∗(KA6 i aiDi P T∗ We now turn to describing the geometry of the Hurwitz space H. We make the following: 1 µ(cid:96) 6.5. We now describe a way of indexing the boundary divisors of Hur. We fix the following combinatorial data: Definition 6.4. For a partition µ = (µ1, . . . , µ(cid:96)) (cid:96) n, we define lcm(µ) := lcm(µ1, . . . , µ(cid:96)) and µ := 1 . For i = 2, . . . , 12, we denote by Pi the set of partitions µ (cid:96) 27 associated to µ1 the cycle decompositions of the conjugacy classes of products of i reflections in W (E6) ⊂ S27. + ··· + 1 The possible partitions of 27 corresponding to products of reflections can be read off Table 1. For instance, we find that P2 =(cid:8)(210, 17), (36, 19)(cid:9). (2) Reflections {wi}i∈I and {wj}j∈J in W (E6), such that(cid:81) (cid:3) t :=(cid:2)π : C → R, p1, . . . , p24 (1) A partition I (cid:116) J = {1, . . . , 24}, such that I ≥ 2, J ≥ 2. j∈J wj = u−1, for some u ∈ W (E6). The sequence w1, . . . , w24 is defined up to conjugation by the same element g ∈ W (E6). To this data, we associate the locus of E6-admissible covers with labeled branch points i∈I wi = u,(cid:81) ∈ HME6, where [R = R1 ∪q R2, p1, . . . , p24] ∈ BI ⊂ M0,24 is a pointed union of two smooth rational curves meeting at the point q. The marked points lying on R1 are precisely those labeled by the set I. Over q, the map π is ramified according to u, that is, the points in π−1(q) correspond to cycles in the permutation u considered as an element of S27. Let µ := (µ1, . . . , µ(cid:96)) (cid:96) 27 be the partition induced by u ∈ S27 and denote by Ei:µ the boundary divisor on H classifying E6-twisted stable maps with underlying admissible cover as above, with π−1(q) having partition type µ, and precisely i of the points p1, . . . , p24 lying on R1. Only partitions from the set Pi introduced in Definition 6.4 are considered. In Table 1 we give the list of partitions of 27 appearing as products of reflections in W (E6) (using the GAP notation for the conjugacy classes). For future use, we also record the invariants 1 µ, for each partition µ. If one partition from this list appears in Pi, it will appear in Pi+2j, for all i + 2j ≤ 12. THE UNIFORMIZATION OF A6 21 Minimal number of reflections Partition µ of 27 Conjugacy class 127 (212, 13) (39) (61, 34, 23, 13) (46, 13) (45, 23, 1) (55, 12) 0 1 2 2 3 3 3 4 4 4 4 4 4 4 5 5 5 5 5 5 6 6 6 6 6 (26, 115) (210, 17) (36, 19) (212, 13) (45, 21, 15) 1a 2c 2b 3b 2d 4d 6e 2a 3c 4a 4b 5a 6b 6d 4c 6f 6g 8a 10a 12b 3a 6a 6c 9a 12a Table 1. Products of reflections in W (E6) (45, 23, 1) (62, 35) (64, 31) (83, 21, 1) (101, 53, 21) (121, 61, 42, 1) (63, 23, 13) (62, 32, 24, 1) (122, 31) (64, 31) (63, 23, 13) (39) (93) 1 µ 27 18 12 11 9 27 4 6 9 3 9 2 15 4 3 5 4 15 4 2 1 15 8 6 5 7 4 3 1 5 1 3 1 2 (cid:98)Ot,HME6 6.6. We recall the local structure of the morphism b : H → M0,24, over the point t, see also [HM82] p.62. The (non-normalized) space HME6 is locally described by its local ring (6.1) where t1 is the local parameter on M0,24 corresponding to smoothing the node q ∈ R. By passing to the normalization ν : H → HME6, we deduce that over each point t(cid:48) ∈ ν−1(t) the map The local ring at t(cid:48) is then given by (cid:98)Ot(cid:48),H = C[[τ, t2, . . . , tb−3]], and the normalization map ν is b : H → M0,24 is ramified with index lcm(µ). The fibre ν−1(t) consists of µ1 ··· µ(cid:96)/lcm(µ) points. = C[[t1, . . . , t21, s1, . . . , s(cid:96)]]/sµ1 1 = ··· = sµ(cid:96) (cid:96) = t1, given in local coordinates by lcm(µ) lcm(µ) t1 = τ lcm(µ), s1 = ζ µ1 µ1 , . . . , s(cid:96) = ζ µ(cid:96) µ(cid:96) , where ζµi is a µi-th root of unity for i = 1, . . . , (cid:96). This description implies that for each i = 2, . . . , 12, we have a decomposition (cid:88) µ∈Pi b∗(Bi) = lcm(µ)Ei:µ. In view of applications to the Kodaira dimension of H, we discuss in detail the pull-back b∗(B2). We pick a point t = [π : C = C1 ∪ C2 → R = R1 ∪q R2, p1, . . . , p24] ∈ b∗(B2) as in in 6.5, where Ci = π−1(Ri). Without loss of generality, we assume that I = {1, . . . , 22}, thus p1, . . . , p22 ∈ R1 and p23, p24 ∈ R2. The group G = (cid:104)w1, . . . , w22(cid:105) generated by the reflections in the remaining 22 roots r1, . . . , r22 ∈ E6 is the Weyl group for a lattice L = LG ⊂ E6. Since(cid:81)24 V. ALEXEEV, R. DONAGI, G. FARKAS, E. IZADI, AND A. ORTEGA that w23 · w24 ∈ G, hence rk(L) ≥ rk(E6) − 1 = 5. 6.7. Assume that the reflections w23 and w24 corresponding to the coalescing points p23 and p24 In this case, the corresponding partition is µ = (127) and we set are equal, hence w23 = w24. E0 := E2:127. We denote by EL the boundary divisor of admissible covers in E2:(127) corresponding to the lattice L. The map b is unramified along each divisor EL and we have i=1 wi = 1, it follows (cid:88) L⊂E6 E0 = EL ⊂ H. The general cover t corresponding to each divisor EL carries no automorphism preserving all branch points p1, . . . , p24, that is, Aut(t) = {Id}. Suppose now that the reflections w23 and w24 are distinct. Following [Dol12, Section 9.1], we distinguish two possibilities depending on the relative position of the two double-sixes, described in terms of a general admissible cover t =(cid:2)π : C = C1 ∪ C2 → R1 ∪q R2, p1, . . . , p24 6.8. The reflections w23 and w24 form an azygetic pair, that is, the corresponding roots r23 and r24 satisfy r23 · r24 (cid:54)= 0. In this case, (cid:104)w23, w24(cid:105) = W (A2) and r23 + r24 or r23 − r24 is again a root that is azygetic to both r23 and r24. The double-sixes associated to w23 and w24 share 6 points and the permutation w23 · w24 decomposes into 6 disjoint three cycles, therefore µ = (36, 19) (cid:96) 27. Accordingly, C2 = π−1(R2) decomposes into six rational components mapping 3 : 1, respectively 9 components mapping isomorphically onto R2. If (cid:3). Eazy := E2:(36,19) ⊂ H is the boundary divisor parametrizing such points, then b is triply ramified along Eazy. The general point of Eazy has no non-trivial automorphisms preserving all the branch points. 6.9. The reflections w23 and w24 form a syzygetic pair, that is, r23 · r24 = 0. We have (cid:104)w23, w24(cid:105) = W (A2 1). The two associated double-sixes share 4 points and w23 · w24 ∈ S27 decomposes into a product of 10 disjoint transpositions, therefore µ = (210, 17). Eight of these transpositions are parts of the double-sixes corresponding to w23 and w24 that remain disjoint respectively. Note that C2 consists of 8 rational components mapping 2 : 1 onto R2, as well as a smooth rational component, say Z, mapping 4 : 1 onto R2. The fibers π−1 Z (p24) each consist of two ramification points. We denote by Z (p23) and π−1 Z (q), π−1 the boundary divisor of admissible syzygetic covers. For a general cover t ∈ Esyz, note that Aut(t) = Z2, see Remark 6.12. 6.10. To summarize the discussion above, we have the following relation: Esyz := E2:(210,17) ⊂ H (6.2) b∗(B2) = E0 + 3Eazy + 2Esyz. In opposition to E0, we show in Theorem 7.15 that the boundary divisors Eazy or Esyz have fewer components. Precisely, for a general element t ∈ Eazy or t ∈ Eazy, we have G = W (L) = W (E6), hence the subcurve C1 = π−1(R1) is irreducible. 6.11. The Hurwitz formula applied to the ramified cover b : H → M0,24, coupled with the expression KM0,24 ≡ (6.3) (cid:0) i(24−i) 23 − 2(cid:1)Bi to be found, e.g., in [KM13], yields KH = b∗KM0,24 (cid:80)12 [Eazy] + N, [Esyz] + [E0] + i=2 + Ram(b) = − 2 23 19 23 40 23 THE UNIFORMIZATION OF A6 Eazy, with the coefficient of [Ei:µ] for i = 3, . . . , 12 being equal to lcm(µ)(cid:0) i(24−i) 23 − 1(cid:1) where N is the effective combination of the boundary divisors of H disjoint from E0, Esyz and The ramification divisor of the projection q : H → Hur is contained in the pull-back b∗(B2) map M0,24 → (cid:102)M0,24. The general point of each of the components of E0 and Eazy admits an (recall the commutative diagram 5.1). Note that B2 is the ramification divisor of the quotient involution compatible with the involution of the rational curve R2 preserving q and interchanging the branch points p23 and p24 respectively. No such automorphism exists for a general point of the divisor Esyz (see Remark 6.12), thus − 1 > 0. 23 (cid:3) Ram(q) = E0 + Eazy. general point t :=(cid:2)π : C → R = R1 ∪q R2, p1, . . . , p24 Remark 6.12. We illustrate the above statement in the case of the divisor Esyz. We choose a ∈ Esyz, and denote by πZ : Z → R2, the degree 4 cover having as source a smooth rational curve Z and such that π∗ Z(q) = 2u + 2v, and π∗ Z(pi) = 2xi + 2yi, for i = 23, 24. Then Aut(t) = Z2. Indeed, there exists a unique automorphism σ ∈ Aut(Z) with σ(u) = u, σ(v) = v, σ(x23) = y23 and σ(x24) = y24 and such that πZ ◦ σ = πZ. Note that σ induces the unique non-trivial automorphism of t fixing all the branch points. In contrast, the general point of Eazy corresponds to an admissible cover which has no automorphisms fixing all the branch points. Definition 6.13. On the space Hur of unlabeled E6-covers, we introduce the reduced boundary divisors D0, Dsyz, Dazy, as well as the boundary divisors(cid:8)Di:µ : 3 ≤ i ≤ 12, µ ∈ Pi (cid:9) which pull-back under the map q : H → Hur to the corresponding divisors indexed by E's, that is, q∗(D0) = 2E0, q∗(Dazy) = 2Eazy, q∗(Dsyz) = Esyz and q∗(Di:µ) = Ei:µ, for 3 ≤ i ≤ 12 and µ ∈ Pi. More generally, for each sublattice L ⊂ E6, we denote by DL ⊂ Hur the reduced divisor which pulls back to EL under the map q. If D is an irreducible divisor on Hur, we denote as usual by [D] := [D]Q ∈ CH 1(Hur)Q its Q- class, that is, the quotient of its usual class by the order of the automorphism group of a general point from D. Theorem 6.14. The canonical class of the Hurwitz space Hur is given by the formula: KHur = − 25 46 [D0] + 19 23 [Dsyz] + 17 46 [Dazy] + lcm(µ) − 1 − 1 [Di:µ]. (cid:16) 12(cid:88) (cid:88) i=3 µ∈Pi (cid:16) i(24 − i) 23 (cid:17) (cid:17) Proof. We apply the Riemann-Hurwitz formula to the map q : H → Hur and we find q∗(KHur) = KH − [E0] − [Eazy] = − 25 23 [E0] + [Esyz] + 19 23 17 23 [Eazy] + ··· ∈ CH 1(H). (cid:3) 6.1. The Hodge class on the space of admissible E6-covers. 6.15. We describe the Hodge class on the Hurwitz space H in terms of boundary divisors. Let ψ1, . . . , ψ24 ∈ Pic(M0,24) be the cotangent tautological classes corresponding to the marked points. The universal curve over M0,24 is the morphism π := π25 : M0,25 → M0,24, forgetting the marked point labeled by 25. The following formulas are well-known, see e.g., [FG03]: (cid:80)24 24 V. ALEXEEV, R. DONAGI, G. FARKAS, E. IZADI, AND A. ORTEGA Proposition 6.16. (1) c1(ωπ) = ψ25 − 24(cid:88) 12(cid:88) (2) ψi = i=1 i=2 i=1 δ0:i,25 ∈ CH 1(M0,25). (3) κ1 = i(24 − i) 23 [Bi] ∈ CH 1(M0,24); (i − 1)(23 − i) 23 [Bi]. We now find a boundary expression for the Hodge class at the level of H. Theorem 6.17. The Hodge class at the level of H is given by the following formula: 12(cid:88) (cid:88) i=2 µ∈Pi λ = (cid:16)9i(24 − i) 23 1 12 lcm(µ) − 27 + 1 µ [Ei:µ]. 12(cid:88) i=2 (cid:17) Note that a boundary formula for λ in the case of Sn-covers appeared first in [KKZ11] and was confirmed later with algebraic methods in [vdGK12]. Proof. Over the Hurwitz space H we consider the universal E6-admissible cover f : C → P , where P := H ×M0,24 M0,25 is the universal orbicurve of genus zero over H. Over a general point t = [C → R, p1, . . . , p24] of a boundary divisor Ei:µ, where µ = (µ1, . . . , µ(cid:96)) ∈ Pi corresponds to the local description (6.1), even though P has a singularity of type Alcm(µ)−1, the space C has singularities of type Alcm(µ)/µi−1 at the (cid:96) points corresponding to the inverse image of Rsing. Let φ : P → H and q : P → M0,25 be the two projections and put v := φ ◦ f : C → H and f := q ◦ f : C → M0,25. The ramification divisor of f decomposes as R1 + ··· + R24 = R ⊂ C, where a general point of Ri is of the form [C → R, p1, . . . , p24, x], with x ∈ C one of the six ramification points lying over the branch point pi. In particular f∗([Ri]) = 6[Bi], where Bi ⊂ P is the corresponding branch divisor. We apply the Riemann-Hurwitz formula for f and write: c1(ωv) = f∗q∗c1(ωπ) + [R]. We are going to push-forward via v the square of this identity and describe all the intervening terms in the process. Over H we have the identity: ∗ c2 1(ωπ) + 2f 1(ωv) = v∗ v∗c2 f ∗ . (cid:16) ∗ f (cid:16) (cid:17) (cid:17) . ψi i=1 c1(ωπ) · [∆0:i,25] c1(ωπ) · [R] + [R]2(cid:17) (cid:17) (cid:17) φ∗q∗(cid:16) = 6b∗(cid:16) 24(cid:88) 24(cid:88) i ]) = 3φ∗(q∗(cid:0)δ2 0:i,25)(cid:1) = −3b∗(ψi). [Ri]2(cid:17) 12(cid:88) (cid:17) 0:i,25) = −ψi, to write: = 27b∗π∗ i(24 − i) 24(cid:88) b∗(Bi). ≡ −3 (cid:17)2 (cid:16) δ0:i,25 = 23 i=1 i=2 ψ25 − i=1 (cid:16) 24(cid:88) i=1 (cid:16) Using Proposition 6.16, we find that v∗([R]2) = v∗ We use Proposition 6.16, and the relation π∗(δ2 ∗ v∗f c2 1(ωπ) = φ∗ 27q∗c2 1(ωπ) We evaluate each term: v∗ c1(ωπ) · [R] = (cid:16) 24(cid:88) i=1 φ∗ q∗c1 (ωπ) · 6[Bi] = 6 Furthermore, we write f∗(Bi) = 2Ri + Ai, where the residual divisor Ai defined by the previous equality maps 15 : 1 onto Bi. Note that Ai and Ri are disjoint, hence f∗([Bi])·Ri = 2R2 i , therefore v∗([Ri]2) = 3φ∗([B2 We find the following expression for the pull-back of the Mumford κ class to H: κ1 − 27b∗(cid:16) (cid:16)9i(24 − i) 23 12(cid:88) i=2 THE UNIFORMIZATION OF A6 (cid:17) ψi 24(cid:88) (cid:17) i=1 ≡ −27b∗(cid:16) 12(cid:88) (cid:88) 12(cid:88) i=2 i=2 µ∈Pi (cid:17) . Bi 25 (cid:16)9i(24 − i) 23 (cid:17) − 27 Ei:µ. lcm(µ) (6.4) v∗c2 1(ωv) ≡ Using Mumford's GRR calculation in the case of the universal genus 46 curve v : C → H, coupled with the local analysis of the fibres of the map b, we have that b∗(Bi) ≡ (cid:88) − 27 12(cid:88) i=2 µ∈Pi 12ϕ∗(λ) ≡ v∗c2 1(ωv) + lcm(µ1, . . . , µ(cid:96)) Substituting in (6.4), we finish the proof. (cid:16) 1 µ1 (cid:17) + ··· + 1 µ(cid:96) Ei:µ. (cid:3) Remark 6.18. Using Definition 6.13, we spell out Theorem 6.17 at the level of Hur: (6.5) λ = 33 46 [D0] + 7 46 [Dazy] + 17 46 [Dsyz] + ··· ∈ CH 1(Hur). (cid:1) (cid:0)ϕ(Γ) · δ0 Proposition 6.19. The morphism ϕ : H → M46 has ramification of order 12 along the divisor E0. In particular, the class ϕ∗(δ0) − 12[E0] − 2[Esyz] ∈ CH 1(H) is effective. Proof. The morphism ϕ factors via Hur, that is, ϕ = ϕ◦ q, where we recall that q : H → Hur is the projection map and ϕ : Hur → M46. We have observed that q is ramified along E0. Furthermore, since the general element of ϕ(E0) is an irreducible 6-nodal curve, the local intersection number ϕ(t), for any curve Γ ⊂ Hur passing through a point t ∈ q(E0), is at least equal to 6. Finally, [Esyz] appears with multiplicity 2 because, as pointed out in Remark 6.12, each point of (cid:3) Esyz has an automorphism of order 2. 6.2. The positivity of the canonical class of H. 6.20. To establish the bigness of the class KH, we use Moriwaki's class [Mor98] mo := (8g + 4)λ − gδ0 − 4i(g − i)δi ∈ CH 1(Mg). 2(cid:99)(cid:88) (cid:98) g i=1 It is shown in [Mor98] that mo non-negatively intersects all complete curves in Mg whose members are stable genus g curves with at most one node. Furthermore, the rational map φn·mo : Mg (cid:57)(cid:57)(cid:75) Pν defined by a linear system n·mo with n (cid:29) 0, induces a regular morphism on Mg. In our situation when g = 46, this implies that the pull-back ϕ∗(mo) is an effective Q-divisor class on H, which we shall determine. In what follows, if D1 and D2 are divisors on a normal variety X, we write D1 ≥ D2 if D1 − D2 is effective. Proposition 6.21. The following divisor class on the Hurwitz space H is effective: 2 23 − E0 + 523 2415 Esyz + 62 115 Eazy + 93 1610 i(24 − i)lcm(µ)Ei:µ 12(cid:88) (cid:88) i=3 µ∈Pi 26 V. ALEXEEV, R. DONAGI, G. FARKAS, E. IZADI, AND A. ORTEGA Proof. We give a lower bound for the coefficient of Ei:µ in the expression ϕ∗(λ) of Theorem 6.17, by observing that for a partition (µ1, . . . , µ(cid:96)) (cid:96) 27, the inequality 1 µ(cid:96) ≤ 27 holds. Using this estimate together with Theorem 6.17 ϕ∗(λ) = 33 23[Eazy] + ··· , as well as Proposition 6.19, we write 1 210 + ··· + 1 46[Esyz] + 7 ϕ∗(mo) ≤ 23[E0] + 17 46 · 2 210 [Esyz] = 372 210 0 ≤ µ1 2 23 − [E0] + 523 2415 [Esyz] + 62 115 i(24 − i)lcm(µ)[Ei:µ]. ϕ∗(λ) − 12(cid:88) 46 · 12 210 (cid:88) [Eazy] + i=3 µ∈Pi [E0] − 93 1610 12(cid:88) (cid:88) i=2 µ∈Pi Ei:µ. (cid:80) The scaling has been chosen to match the negative E0 coefficient in the class KH of (6.3). As a step towards determining the Kodaira dimension of Hur we establish the following: Theorem 6.22. The canonical class of H is big. Proof. Recalling that b : H → M0,24, for each 0 < α < 1, using (6.3) we write the equality (cid:3) KH = (1 − α)b∗(κ1) + αb∗(κ1) − Since the class κ1 ∈ CH 1(M0,24) is well-known to be ample, in order to establish that KH is big, it suffices to show that for α sufficiently close to 1, the class αb∗(κ1) − [Ei:µ] is effective. (cid:3) After brief inspection, this turns out to be a consequence of Proposition 6.21. i,µ∈Pi 7. The Prym-Tyurin map along the boundary components of Hur In this section we study the extended Prym-Tyurin map and refine the analysis of the boundary divisors of Hur. In particular we identify the divisors that are not contracted by the extended Prym-Tyurin map P T : Hur (cid:57)(cid:57)(cid:75) A6. 7.1. Following 6.5, we denote by EI:L1,L2,µ the divisor of H of E6-admissible covers t := [π : C := C1 ∪ C2 → R1 ∪q R2, p1, . . . , p24], where I ∪ J = {1, . . . , 24}, R1 contains the branch points {pi}i∈I, with roots {ri}i∈I generating and reflections generating the group H := W (L2) ⊂ W (E6) respectively. We set u := (cid:81) the lattice L1 ⊂ E6 and the corresponding reflections generating the group G := W (L1) ⊂ W (E6), therefore u−1 = (cid:81) whereas R2 contains the branch points {pj}j∈J , with roots {rj}j∈J generating the lattice L2 ⊂ E6 i∈I wi, j∈J wj. As before, µ (cid:96) 27 is the partition corresponding to the cycle type of u ∈ S27 which describes the fibre π−1(q). Let OG (respectively OH) denote the set of orbits of the notation in 6.5, for µ ∈ Pi we write Ei:µ = (cid:80)I=i,L1,L2 G (respectively H) on the set 27 := {1, . . . , 27}. In particular, there is a bijection between OG (respectively OH) and the set of irreducible components of C1 (respectively C2). Returning to EI:L1,L2,µ, the sum being taken over sublattices L1 and L2 of E6 as above. Definition 7.2. Let u ∈ W (E6) and let A = A1(cid:116). . .(cid:116)Aa and B = B1(cid:116). . .(cid:116)Bb be two u-invariant partitions of the set 27. We define the graph Γ(u, A, B) to be the following bipartite graph: (1) The vertices are A1, . . . , Aa and B1, . . . , Bb respectively. (2) The edges correspond to cycles Ck in the cyclic representation of u ∈ S27, including cycles of length 1. THE UNIFORMIZATION OF A6 27 (3) For each cycle Ck, there exist unique vertices Ai and Bj containing the set ck. Then the edge Ck joins Ai and Bj. When both partitions A and B are trivial, that is each consists of the single set 27, we set H1(Γ1, Z) (cid:39) Z26 consists of elements(cid:80)27 Γu := Γ(u, 27, 27) and Γ1 := Γ(1, 27, 27) respectively. Example 7.3. The graph Γ1 has 2 vertices and 27 edges. One has C1(Γ1, Z) = Z27, and s=1 ns = 0. There is a natural degree 10 homomorphism DC1 : C1(Γ1, Q) → C1(Γ1, Q) with eigenvalues 10, 1,−5, which induces a homomor- phism DH1 : H1(Γ1, Q) → H1(Γ1, Q) with (+1)-eigenspace of dimension 20 and (−5)-eigenspace of In particular, for the dual graph Γ of C, the group H1(Γ, Z) comes with an endomorphism DH1 s=1 nses with(cid:80)27 dimension 6 respectively. by Lemma 5.5. Theorem 7.4. Let t ∈ EI,J:L1,L2 be a general point in a boundary divisor of H corresponding to the above data. Then the toric rank of the PTK variety P T (C, D) equals the dimension of the (−5)-eigenspace H1 Proof. By Lemma 5.5, the toric rank of P T (C, D) equals (cid:0)Γ(u, OG, OH), Q(cid:1)(−5) of the endomorphism DH1 on H1 rank Im(DH1 − 1) = rank ker(DH1 + 5) = dim(H1 ⊗ Q)(−5). (cid:0)Γ(u, OG, OH), Q(cid:1). (cid:3) In case both curves C1 and C2 are irreducible, the above result simplifies considerably. Corollary 7.5 agrees with the result of [LR08, p.236] concerning the abelian part of P T (C, D). Lemma 7.6. For u ∈ W (E6), the following statements hold: Corollary 7.5. Assume that OG = OH = 1, that is, both groups G and H act transitively on the set 27. Then the toric rank of P T (C, D) equals the dimension of invariant subspace of u in the 6-dimensional representation E6 ⊗ Q of W (E6). (1) H1(Γu, Q)(−5) =(cid:0)H1(Γ1, Q)(−5)(cid:1)u (that is, the u-invariant subspace), and (2) H1(Γ(u, A, B), Q)(−5) ⊂ H1(Γu, Z)(−5). n := ord(u) and (cid:96)(Ci) denote the length of Ci. We write C1(Γu, Z) = (cid:76)k (cid:80)n−1 Proof. Suppose we have the following cycle decomposition u = C1 · C2 ··· · Ck ∈ S27 and let ZeCi. Then one has an orthogonal projection C1(Γ1, Z) (cid:16) C1(Γu, Z) given by e (cid:55)→ 1 i=0 ui · e for an edge e, which identifies C1(Γu, Z) with a sublattice in C1(Γ1, Q) via the injection eCi ej. (−5)-eigenspace in C1(Γu, Q) and that H1(Γu, Q)(−5) =(cid:0)H1(Γ1, Q)(−5)(cid:1)u. This induces a surjection from H1(Γ1, Z) to H1(Γu, Z), which clearly commutes with D, that is, D(C1(Γu, Z)) ⊂ C1(Γu, Z). It follows that H1(Γu, Z)(−5) is the projection of H1(Γ1, Z)(−5) to the The graph Γ(u, A, B) is obtained from Γu by splitting the two vertices into a + b new ver- tices. This can be obtained by inserting in place of the two vertices two trees with a and b vertices -- without changing H1 -- and then removing the edges of these trees. Thus, one has an inclusion H1(Γ(u, A, B), Z) ⊂ H1(Γu, Z), commuting with D, which gives an inclusion (cid:3) H1(Γ(u, A, B), Z)(−5) ⊂ H1(Γu, Z)(−5). Lemma 7.7. The (−5)-eigenspace H1(Γ(u, A, B), Q)(−5) is a subspace of the u-invariant subspace (E6 ⊗ Q)u in the standard 6-dimensional W (E6)-representation. (cid:80) i=1 n (cid:55)→ 1 (cid:96)(Ci) j∈Ci 28 V. ALEXEEV, R. DONAGI, G. FARKAS, E. IZADI, AND A. ORTEGA Proof. Indeed, H1(Γ1, Q)(−5) = E6 ⊗ Q, therefore (H1(Γ1, Q)u)(−5) = (E6 ⊗ Q)u. r2 = α12, . . . , r6 = α56, and r1 = α123. In the extended Dynkin diagram (cid:101)E6 there is an additional 7.8. In order to illustrate Theorem 7.4 in concrete situations, we classify all root sublattices of E6. Recall that with the notation of Section 2, the standard roots in the E6 Dynkin diagram are (cid:3) root r0 = −αmax, so that 3r4 + 2r1 + 2r3 + 2r5 + r2 + r6 + r0 = 0. Lemma 7.9. The following is the complete list of root sublattices L ⊂ E6: (1) If dim(L) = 6, then L is either E6, or isomorphic to A5A1, or A3 2. 1, or A2 (2) If dim(L) = 5, then L is isomorphic to A5, D5, A4A1, A3A2 (3) If dim(L) = 4, then L is isomorphic to A4, D4, A2 (4) If dim(L) = 3, then L is isomorphic to A3, A2A1, or A3 1. (5) If dim(L) = 2, then L is isomorphic to A2, or A2 1. (6) If dim(L) = 1, then L is isomorphic to A1. 2, A3A1, A2A2 2A1. 1, or A4 1. Furthermore, all the above sublattices (and the associated subgroups) can be obtained by removing vertices from the extended E6 diagram (cid:101)E6: r3• r2• r5• r6• r4• •r1 (cid:63)r0. If the root lattices L, L(cid:48) corresponding to reflections subgroups G and G(cid:48) of W (E6) are isomorphic, then they differ by an automorphism of the E6 lattice, and the corresponding subgroups G and G(cid:48) are conjugate in W (E6). Proof. We first note that there is a natural bijection between root sublattices L of E6 and subgroups G generated by reflections of W (E6). One has Aut(E6) = W (E6) ⊕ Z2, with Z2 acting on E6 by multiplication by ±1. Any automorphism of E6 induces an automorphism of W (E6), and the kernel of φ : Aut(E6) → Aut W (E6) is Z2. Finally, by [Fra01, Section 2.3]), all automorphisms of W (E6) are inner, so that Aut W (E6) = W (E6) and φ is surjective. Thus, the proof reduces to showing that all root sublattices of E6 are of the above types, and that if L, L(cid:48) are isomorphic as abstract root lattices, then they differ by an element of Aut(E6). Dynkin diagram (cid:101)E6 is an a posteriori observation. The statement that all such root sublattices correspond to proper subdiagrams of the extended The standard method for finding all root sublattices of a given root lattice is described in [BdS49, Dyn52]. A modern treatment can be found in [DL11, Theorem 1]. The method is to repeatedly apply the following two procedures to Dynkin diagrams Γ, starting from Γ = E6: (1) remove a node, and/or (2) replace one of the connected components Γs of Γ by an extended Dynkin diagram(cid:101)Γs and remove a node from it. Applying the above method repeatedly, we obtain all the lattices listed above. The fact that isomorphic root sublattices differ by an automorphism of E6 (cid:3) is a case by case computation. This can also be found in [Osh06, Table 10.2]. 7.10. Table 2 lists the orbits for one choice of roots (the other choices being similar) for each type of lattice. The last column describes the degrees of the maps from the irreducible components of THE UNIFORMIZATION OF A6 29 C1 to R1. We keep the Schlafli notation ai, bi, cij for the elements of the set 27, which is being identified with the set of lines of a cubic surface. The smooth (possibly disconnected) curve C1 is a 27-sheeted cover of R1 = P1, with branch points {pi}i∈I with local monodromy given by the reflection wi, and an additional branch point q, with local monodromy u−1, where u = Πi∈Iwi. Degrees Sublattice Roots Orbits {ai, bi, cij} {ai, bi}, {cij} 27 15, 12 93 1, 10, 16 62, 15 2, 5, 102 82, 6, 4, 1 9, 62, 32 10, 53, 12 83, 13 9, 36 8, 6, 42, 22, 1 62, 4, 32, 22, 1 46, 13 6, 44, 15 6, 34, 23, 13 43, 26, 13 36, 19 4, 28, 17 26, 115 E6 A5A1 A3 2 D5 A5 r1, . . . , r6 r0, r2, . . . , r6 ri, i (cid:54)= 4 r1, . . . , r5 r2, . . . , r6 A4A1 r0, r2, . . . , r5 A3A2 1 A2 2A1 A4 D4 A2 2 A3A1 A2A2 1 A4 1 A3 r0, r2, r3, r4, r6 ri, i (cid:54)= 0, 4 r2, . . . , r5 r1, r3, r4, r5 r2, r3, r5, r6 r2, r3, r4, r6 r1, r2, r3, r5 r0, r2, r4, r6 r2, r3, r4 A2A1 r1, r2, r3 A3 1 A2 A2 1 A1 r2, r4, r5 r2, r3 r2, r4 r0 {ai, bi, cij 1 ≤ i, j ≤ 3}, {ai, bi, cij 4 ≤ i, j ≤ 6}, {a6}, {ai, b6, cij 1 ≤ i, j ≤ 5},{bi, ci6 1 ≤ i ≤ 5} {cij 1 ≤ i ≤ 3, 4 ≤ j ≤ 6} {ai},{bi},{cij} {ai, cij 1 ≤ i, j ≤ 4},{bi, c56 1 ≤ i ≤ 4} {a5, a6},{bj, cij 1 ≤ i ≤ 4, 5 ≤ j ≤ 6} {a5, a6, c15, c16},{bj, cij 2 ≤ i ≤ 4, 5 ≤ j ≤ 6} {a1, bi, cij, c56 2 ≤ i, j ≤ 4},{b1},{ai, cij 2 ≤ i, j ≤ 4} {ai, cij 1 ≤ i, j ≤ 3},{bi, cij 4 ≤ i, j ≤ 6},{a4, a5, a6} {ai, cij 1 ≤ i, j ≤ 4},{bi, c56 1 ≤ i ≤ 4},{a5},{a6} {a1, cij, b6 2 ≤ i, j ≤ 5},{ai, c1i 2 ≤ i ≤ 5},{a6} {b1, b2, b3},{cij 1 ≤ i ≤ 3, 4 ≤ j ≤ 6} {b5, ci6 1 ≤ i ≤ 4},{b6, ci5 1 ≤ i ≤ 4} {b1},{bi, ci6 2 ≤ i ≤ 5},{c16} {a1, a2, a3},{b1, b2, b3},{a4, a5, a6},{b4, b5, b6} {c12, c13, c23},{c45, c46, c56},{cij 1 ≤ i ≤ 3, 4 ≤ j ≤ 6} {c56},{a5, a6},{b5, b6},{a1, . . . , a4},{b1, . . . , b4} {cij 1 ≤ i, j ≤ 4},{cij 1 ≤ i ≤ 4, 5 ≤ j ≤ 6} {c12, c13, c14, c23, c24, c34},{c15, c25, c35, c45},{c16, c26, c36, c46} {a6},{b6, c45},{b1, b2, b3},{c16, c26, c36},{b5, b6, c64, c65} {a4, a5},{ai, cij 1 ≤ i, j ≤ 3},{cij 1 ≤ i ≤ 3, 4 ≤ j ≤ 5} {a6},{b1},{a1, b6, c23, c45},{a2, a3, c12, c13},{a4, a5, c14, c15} {b2, b3, c26, c36},{b4, b5, c46, c56},{c24, c34, c25, c35},{c16} {a1, a2, a3, a4},{b1, b2, b3, b4},{ai},{bi}, 5 ≤ i ≤ 6,{c56} {c16, c26, c36},{bj, ckl},{j, k, l} = {4, 5, 6},{ai}, 4 ≤ i ≤ 6 {c13, c14, c23, c24},{c15, c16, c25, c26},{c35, c36, c45, c46} {b1, b2, b3},{c14, c24, c34},{c15, c25, c35},{a1, a2, a3, c12, c13, c23} {ai, ai+1},{bi, bi+1},{cii+1}, i = 1, 3, 5 {a1, a2, a3},{b1, b2, b3},{c12, c13, c23},{c14, c24, c34} {c15, c25, c35},{c16, c26, c36},{ai},{bi},{cij}, 4 ≤ i, j ≤ 6 {c13, c23, c14, c24},{c56},{ai, ai+1},{bi, bi+1}, {cij, ci+1j},{aj},{bj},{cii+1}, i = 1, 3, j = 5, 6 {ai, bi},{cij}, 1 ≤ i, j ≤ 6 Table 2. Sublattices and Orbits We apply Theorem 7.4 to compute the toric ranks associated to the divisors (cid:88) I=22 EL := EI;L,A1,(127) ⊂ H. Since dim(L) ≥ 5, using Lemma 7.9, we have the following possibilities: 2A1}. L ∈ {E6, A5A1, A3 2, A5, D5, A4A1, A3A2 1, A2 30 V. ALEXEEV, R. DONAGI, G. FARKAS, E. IZADI, AND A. ORTEGA Assume OG = 1, which, by Table 2, is the case if and only if L = E6. Then H1(Γ, Z) =(cid:76)6 Proposition 7.11. The toric rank of each boundary divisor EL with L (cid:54)= E6 is equal to zero. The toric rank of EE6 is equal to 1. Proof. Note that there are OH = 21 vertices on the right, of which 15 vertices are ends and thus can be removed without changing the homology of the graph. The remaining 6 vertices each have degree 2. Contracting unnecessary edges, we reduce the calculation to a graph with OG vertices and 6 edges. The 6 edges correspond to the 6 transpositions appearing in the decomposition of w23 = w24 ∈ S27. Zei j(cid:54)=i ej. Therefore H1(Γ, Q)(−5) is 1-dimensional and generated by the element and D(ei) = − (cid:3) e1 + ··· + e6. The other cases follow similarly by direct calculation. Remark 7.12. For more details concerning the calculation of the toric rank in the case L = D5, see Paragraph 8.2. 7.13. Although the divisor theory of H is quite complicated, we now show that most of these divisors are contracted under the Prym-Tyurin map. We first establish the following: (cid:80) i=1 6 6 Theorem 7.14. Assume that the image of a component B of EI:L1,L2 under the rational Prym- . Then {I,I c} = {2, 22}. has codimension 1 in Aperf Tyurin map H (cid:57)(cid:57)(cid:75) Aperf Proof. With notation as in 7.1, denote Pi := P T (Ci, Di) the PTK varieties for the two parts of R. Recall that by Lemma 5.4 the abelian part ab(P T ) is isogenous to P1 × P2. Without loss of generality, we may assume that i := I ≥ 12. Since codim(Aperf the irreducible components Ri, i = 1, 2, of R image of B has codimension 1, then for a general point of B, the toric rank kP of the corresponding PTK variety P is either 0 or 1. Suppose first that kP = 0. In this case in fact P ∼= P1 × P2. If both P1 and P2 have positive dimension, then P belongs to a subvariety of A6 parametrizing products and each such subvariety has codimension greater than 1. So one of the Pi is zero. The parameter space of P1 has dimension at most i − 2, that of P2 has dimension at most 22 − i. Since the parameter space of P is 20- dimensional and i ≥ 12, we have P2 = 0, and dim(P1) = 6. We deduce i = 22. \ (cid:101)A6) ≥ 2, if Now assume kP = 1. In this case the image of B is the boundary divisor D6 of Aperf . Then P1 × P2 must be a general abelian variety of dimension 5. Once again, one of the Pi is zero. The assumption i ≥ 12 implies P2 = 0, hence dim(P1) = 5. The parameter space of P1 is 15- dimensional, which implies i ≥ 17. Let {p1, . . . , p(cid:96)} = C1 ∩ C2 be the set of the nodes of C, which also label the edges of Γ. For each i, let p(cid:48) i be the points of C1 and C2 respectively that we identify to obtain pi on C. Choose the orientation of Γ in such a way that each edge pi is oriented from C1 to C2. The extension i and p(cid:48)(cid:48) 6 6 0 −→ H 1(Γ, C∗) −→ JC −→ JC1 × JC2 −→ 0 is given by the map φ : H1(Γ, Z) → JC1 × JC2 sending the edge pi to p(cid:48)(cid:48) i − p(cid:48) 0 −→ C∗ = (DT − 1)H 1(Γ, C∗) −→ P −→ ab(P ) −→ 0 the extension is given by the map φP : (DH1 − 1)H1(Γ, Z) → ab(P ). Clearly the composition of φP with the isogeny ab(P ) → P1× P2 is the restriction of φ to (DH1 − 1)H1(Γ, Z) composed with the projection JC1 × JC2 → (DA − 1)(JC1 × JC2) = P1 × P2. Since P2 = 0, this means that the extension class of P does not depend on C2 (up to a finite set). Therefore the moduli of C2 does not produce positive moduli for the extension class of P . i. We observe that THE UNIFORMIZATION OF A6 31 It follows that the cover C1 → R1 depends on 20 moduli, hence R1 contains at least 22 branch (cid:3) points, therefore i = 22. The following result shows that the boundary divisors have many fewer irreducible components than one would a priori expect. Recall that in 6.8 and 6.9 we introduced the divisors Eazy := (cid:80)I=22 EI:L1,A2,(36,19) and Esyz :=(cid:80)I=22 EI:L1,A2 1,(210,17) respectively. Theorem 7.15. Assume that I = 22 and L2 = A2 or A2 In other words, for a general E6-admissible cover (cid:2)π : C = C1 ∪ C2 → R1 ∪q R2, p1, . . . , p24 (cid:3) the curve C1 is irreducible with monodromy W (E6) over R1. ∈ Eazy or Esyz, 1. Then EI:L1,L2 is empty unless L1 = E6. Proof. Consider first the azygetic case L2 = A2. Then, as we saw in 6.8, the curve C2 has 15 components, each of which intersects C1 in exactly one point. Therefore, no component of C2 can connect two components of C1 and C1 is irreducible. In the syzygetic case, as we saw in 6.9, the curve C2 has 16 components. Of the components of C2, only the 4-sheeted cover (denoted by Z in 6.9) intersects C1 in two points. All other components of C2 intersect C1 in exactly one point. It follows that C1 has at most two irreducible components. Looking now at Table 2, we see that there are only two possibilities for the lattice L1, namely L1 = E6 or L1 = A5A1. We now eliminate the possibility L1 = A5A1 in the syzygetic case. It is a consequence of Lemma 7.9 that the A5 summand of L1 is the orthogonal complement of the A1 summand. Hence the lattice L1 and the group G1 generated by the reflections w1, . . . , w22 are determined by the A1 sublattice. Since all reflections are conjugate, we can assume that the A1 summand is generated by the reflection w0 in the root r0 (see 7.10). Since (cid:104)G1, w23(cid:105) = (cid:104)G1, w24(cid:105) = W (E6), the reflections w23, w24 do not belong to G1. Therefore the pairs (w0, w23), (w0, w24) are azygetic. After a permutation of the indices {1, . . . , 6}, we can assume that w23 is the reflection in the root α123 and w24 is the reflection in the root α145 (see, e.g., [Dol12, Section 9.1]). The composition w23·w24 ∈ S27, contains the double transposition (a1, b6)(c23, c45) which acts on the 4-sheeted cover Z. However, w23 · w24 also contains the transposition (a2, c13) which acts on a degree 2 component of C2, i.e., the points corresponding to a2 and c13 come together over the node. Looking at the orbits of C1 in Table 2 in the A5A1 case, we see that the points a2 and c13 belong to two different (cid:3) components of C1 and cannot come together over the node, which is a contradiction. We now consider the components of the divisor E0 introduced in 6.7. Recall that (cid:88) I=22 EL := EI;L,A1,(127) ⊂ H. Theorem 7.16. For L (cid:40) E6 the divisor EL is contracted by P T . Proof. Let [π : C := C1 ∪ C2 → R1 ∪q R2] be a general element of a component B of EL with L (cid:40) E6. By Proposition 7.11, the toric rank of P := P T (C, D) is 0. As in the proof of Theorem 7.14 and with the notation there, we have P = P1 × P2 = P1 = P T (C1, D1) because all the components of C2 are rational. Furthemore, since C1 → R1 is not ramified at q, the isomorphism class of C1 and hence also of P1 is independent of the choice of the point q. It follows that P = P1 (cid:3) depends on at most 19 = dim(M0,22) parameters, hence B is contracted by P T . We summarize the results of this section in terms of the Hurwitz space Hur := H/S24: 32 V. ALEXEEV, R. DONAGI, G. FARKAS, E. IZADI, AND A. ORTEGA Theorem 7.17. The only boundary divisors of Hur that are not contracted under the Prym-Tyurin map P T : Hur (cid:57)(cid:57)(cid:75) A6 are DE6, Dsyz and Dazy. The divisor DE6 maps onto the boundary divisor D6 of A6, whereas Dsyz and Dazy map onto divisors not supported on the boundary of A6. 8. Ordinary Prym varieties regarded as Prym-Tyurin-Kanev varieties The aim of this section is to illustrate how 6-dimensional Prym varieties appear as PTK varieties of type E6 and thus prove Theorem 0.5. The Prym moduli space R7 has codimension 3 inside A6, where we identify R7 with the image of the generically injective Prym map P : R7 → A6. We shall show that the boundary divisor DD5 of Hur is an irreducible component of P T −1(R7) and we shall explicitly describe the 2-dimensional fibres of the restriction P TDD5 8.1. Consider an admissible cover [π : C = C1 ∪ C2 → R1 ∪ R2] in the divisor DD5 of Hur. We choose such a cover as follows. The cover C1 → R1 has D5-monodromy generated by the roots r1, . . . , r5, and it is ramified at 22 distinct points. The local monodromy at each branch point is given by one of the reflections wi ∈ W (D5) associated to ri, choosing the ordering such that i=1 wi = 1. The cover C2 → R2 has A1-monodromy generated by the root r0, and is branched at 2 points. Both covers are unramified at the point q ∈ R1 ∩ R2. As listed in Table 2, we have the following irreducible components and orbits for C1: (cid:57)(cid:57)(cid:75) R7. (cid:81)22 : DD5 F1 : F2 : F0 : {b1, b2, b3, b4, b5, c16, c26, c36, c46, c56} {a1, a2, a3, a4, a5, c12, c13, c14, c15, c23, c24, c25, c34, c35, c45, b6} {a6}, and the following irreducible components and orbits for C2: 1 ≤ i ≤ 6 Hi : {ai, bi} 7 ≤ i ≤ 21 Hi : {ck(i)(cid:96)(i)}, for some choice of integers k(i) < (cid:96)(i), between 1 and 6. One computes that the three components F1, F2 and F0 of C1 have genera 13, 29 and 0, and map onto R1 with degree 10, 16 and 1 respectively. The components of C2 are all rational with H1, . . . , H6 mapping 2 : 1 to R2 and H7, . . . , H21 mapping isomorphically. The description of the orbits given above also specifies the points of intersection Fi and Hj. For instance, H6 intersects F2 at a point corresponding to b6 and it intersects F0 at a point corresponding to a6. 8.2. In order to compute the toric rank of the Prym-Tyurin variety P := (D − 1)(JC1), we apply the correspondence D to the homology group H1(Γ(cid:48), Z), where Γ(cid:48) denotes the simplified dual and H1(Γ(cid:48), Z) =(cid:76)4 graph of the stable curve C1 ∪ C2 (see Section 4.5 for the notation). The graph Γ(cid:48) consists of 2 vertices joined by 5 edges: e1 := (b1, a1), e2 := (b2, a2), e3 := (b3, a3), e4 := (b4, a4), e5 := (b5, a5) Z (ei − ei+1) (see 3.2). One computes i=1 D(∂(e1 − e2)) = D(a1 − b1) − D(a2 − b2) = b2 − a2 − (b1 − a1) = ∂(e1 − e2). By Remark 2.6, D commutes with ∂, hence D(e1 − e2) = (e1 − e2). Similarly, one checks that D(ei−ei+1) = (ei−ei+1), for i = 1, . . . , 4, hence (D−1)H1(Γ(cid:48), Z) = 0. Therefore, the Prym-Tyurin variety P := (D − 1)(JC1) has toric rank 0 and it is contained in JC1, since JC2 = {0}. 8.3. As is apparent from the description of the orbits, the correspondence D restricts to a fixed- point-free involution ι : F1 → F1, a correspondence D2 of valence 5 on F2, a correspondence D12 : F1 → F2 and its transpose D21 : F2 → F1 of degree 8 over F1 and of degree 5 over F2. The variety P is the image of the following endomorphism of JC1 = JF1 × JF2: THE UNIFORMIZATION OF A6 (cid:18) ι − 1 D21 D12 D2 − 1 (cid:19) . 33 8.4. Let f : F1 → Y be the induced unramified double cover on the curve Y := F1/(cid:104)ι(cid:105) of genus 7. Note that the degree 10 map π1 : F1 → R1 factors through a degree 5 map h : Y → R1. The image Q1 := (ι − 1)JF1 ⊂ JF1 is the ordinary Prym variety P (F1, ι) associated to the double cover [f : F1 → Y ] ∈ R7. → Y h→ R1 and the 8.5. The relationship between the curves F1 and F2 (or between the tower F1 map π2 : F2 → R1) is an instance of the pentagonal construction ([Don92, Section 5.17]). This is the n = 5 case of the n-gonal construction, see [Don92, Section 2] or [ILS09a, Section 1], which → Y h→ R1 whose Galois group is the Weyl group W (Dn). The idea is to applies to covers F1 consider the following curve inside the symmetric product F (n) : f f 1 h∗F1 :=(cid:8)G ∈ F (n) 1 : Nmf (G) = h−1(t), for some t ∈ R1 (cid:9). The induced map h∗F1 → R1 is of degree 2n = 32 and one checks that above a branch point t ∈ R1 there are exactly 2n−2 = 8 simple ramification points in h∗F1. Proposition 8.6. h∗F1 is the union of two isomorphic components h∗F1 = X0(cid:116)X1, with X0 (cid:39) X1 being smooth curves of genus 1 + 2n−3(n + g(Y ) − 5) = 29. Proof. The splitting is explained in [Don92, Section 2.2] and [ILS09a, Section 1]. The smoothness (cid:3) is proved in [ILS09a, Lemma 1.1]. The genus calculation follows from the Hurwitz formula. share an even number of points of F1. Two divisors G1, G2 ∈ h∗F1 with Nmf (G1) = Nmf (G2) belong to the same component if they We specialize to the case n = 5. Let X = X0 be the component of h∗F1 whose fiber over a point t ∈ R1 can be identified with the class of the divisor c16 + ··· + c56. The proof of the following result is immediate. Proposition 8.7. The map ψ : F2 → X given by x (cid:55)→ D21(x) ∈ h∗F1 is an isomorphism. Remark 8.8. Under the above identification, the restriction D2 of the Kanev (incidence) corre- spondence coincides with the correspondence D defined in [ILS09a, Section 2]. Also, the restriction D21 of the Kanev correspondence coincides with the correspondence S defined in [IL12, Section 2]. It follows from [ILS09a, Corollary 6.2] that the image Q2 of the ordinary Prym variety Q1 in JF2 by D21 is the eigen-abelian variety of D2 for the eigenvalue −n + 2 = −3. It also follows from [ILS09a, Section 6.6] and [Kan87, Theorem 3.1] that in this case Q2 is a Prym-Tyurin variety of dimension 6 and exponent −(−3) + 1 = 4 for the correspondence D2. The restriction ρ of the correspondence D − 1 to Q1 ⊂ JF1 gives the sequence of isogenies of principally polarized abelian varieties ρ Q1 x1 P −→ (cid:55)−→ ((ι − 1)x1, D12x1) −→ Q2 (cid:55)−→ D12x1. Proposition 8.9. The map ρ factors through multiplication by 2 to induce an isomorphism Q1 := P (F1, ι) (cid:39) P and a surjection Q1 → Q2 := P (F2, D2) whose kernel is a maximal isotropic subgroup H (with respect to the Weil pairing) of the group of points of order 2 in Q1. 34 V. ALEXEEV, R. DONAGI, G. FARKAS, E. IZADI, AND A. ORTEGA Proof. For an abelian variety A, we denote by nA : A → A the morphism given by multiplication by n ∈ Z. It follows from [IL12, Corollary 2.3] that D21◦ D12 = 8Q1. A straightforward generalization of the proof of [IL12, Proposition 3.3] implies that D12 = ϕ ◦ 2Q1, for an isogeny ϕ : Q1 → Q2 such that ϕ∗ΘQ2 = ΘQ1, where ΘQi is the polarization of Qi. Therefore we have ϕ ◦ ϕt = 2Q2. It follows that the kernel of ϕ is a maximal isotropic subgroup H of the group Q1[2] of points of order 2 in Q1. Since the restriction of ι − 1 to Q1 is −2Q1, its kernel is the subgroup of points of (cid:3) order 2. Therefore ρ = ψ ◦ 2Q1, where now ψ : Q1 → P is injective, hence an isomorphism. 9. The Weyl-Petri realization of the Hodge eigenbundles Since the pull-back of the Hodge class from A6 is precisely the class λ(−5) on Hur, describing it in terms intrinsic to the Hurwitz space is of obvious importance. Here we show that, at least on an open dense subset of Hur, both Hodge eigenbundles E(+1) and E(−5) admit a Petri-like incarnation, which makes them amenable to intersection-theoretic calculations. 9.1. For a smooth E6-cover [π : C → P1] ∈ Hur, set L := π∗(OP1(1)) ∈ W 1 µ(L) : H 0(C, L) ⊗ H 0(C, ωC ⊗ L∨) → H 0(C, ωC) be the Petri map given by multiplication of global sections. Assume h0(C, L) = 2, so that h0(C, ωC ⊗ L∨) = 20. If, as expected, for a general choice of [C, L] ∈ Hur, the map µ(L) is injective, then Im µ(L) is a codimension 6 subspace of H 0(C, ωC). Remarkably, this is the (+1)- eigenspace of H 0(C, ωC). At the end of this paper, we shall establish that a general covering from Hur is Petri general: Theorem 9.2. For a general point [C, L] ∈ Hur, the multiplication map µ(L) is injective. 27(C) and let Postponing the proof, we have the following description of the Hodge eigenbundles. Theorem 9.3. Let [C, L] ∈ Hur be an element corresponding to a nodal curve of genus 46 and a 27(C), such that h0(C, L) = 2 and the Petri map µ(L) is injective. base point free line bundle L ∈ W 1 One has the following canonical identifications: (i) H 0(C, ωC)(+1) = H 0(C, L) ⊗ H 0(C, ωC ⊗ L∨). (ii) H 0(C, ωC)(−5) = (cid:18) H 0(C, L⊗2) (cid:19)∨ H 0(C, L). 2(cid:94) ⊗ S2H 0(C, L) 0 −→ OC ·s−→ L −→ OΓ(Γ) −→ 0 0 −→ H 0(C,OC) −→ H 0(C, L) −→ H 0(OΓ(Γ)) α−→ H 1(C,OC) −→ H 1(C, L) −→ 0. Proof. Let L be an E6-pencil on C. Consider a general divisor Γ ∈ L and the exact sequence (9.1) induced by a section s ∈ H 0(C, L) with div(s) = Γ, and its cohomology sequence (9.2) There is an action of W (E6) on H 0(OΓ(Γ)) compatible with the trivial action on H 0(C, L), because L and H 0(L) are pull-backs from P1. We identify the space H 0(OΓ(Γ)) with the vector space generated by the 27 lines on a smooth cubic surface; each line is represented by a point of Γ and the incidence correspondence of lines is the Kanev correspondence D. Therefore the representation of W (E6) on H 0(OΓ(Γ)) splits into the sum of three irreducible representations: the trivial 1- dimensional one, the 6-dimensional one which coincides with the representation on the primitive cohomology of a cubic surface and the 20-dimensional one, which coincides with the one on the space of linear equivalences on a cubic surface, see for instance [AV12]. THE UNIFORMIZATION OF A6 ∈ End(cid:0)H 1(C,OC)(cid:1) via the cohomology sequence (9.2). On a cubic surface, how the action on H 0(OΓ(Γ)) can be described. It follows that the image α(cid:0)H 0(OΓ(Γ))(cid:1) contains The Kanev correspondence D induces an endomorphism on H 0(OΓ(Γ)) compatible with the endomorphism D∨ the action of the incidence correspondence on the primitive cohomology is equal to multiplication by −5 and its action on the space of rational equivalences is the identity. Therefore this is also the (−5)-eigenspace, that is, we have an inclusion 35 (cid:0)H 0(C, ωC)(−5)(cid:1)∨ ⊆ α(cid:0)H 0(OΓ(Γ))(cid:1) = (cid:18) H 0(C, ωC) H 0(C, L) ⊗ H 0(C, ωC ⊗ L∨) (cid:19)∨ . When the Petri map µ(L) is injective, the two spaces appearing in this inclusion have the same dimension and the inclusion becomes an equality, which establishes the first claim. To prove the second claim, we start by observing that the Base Point Free Pencil Trick yields → H 0(C, L) ⊗ OC → L → 0. After tensoring with L and the sequence 0 → taking cohomology, we arrive at the following exact sequence (cid:86)2 H 0(C, L) ⊗ L∨ 2(cid:94) H 0(C, L⊗2) Sym2H 0(C, L) −→ 0 −→ H 0(C, L) ⊗ H 1(C,OC) u−→ H 0(C, L) ⊗ H 1(C, L) −→ 0. To describe the map u in this sequence, let us choose a basis s1, s2 ∈ H 0(C, L). Then, u(s1 ∧ s2 ⊗ f ) = s1 ⊗ (s2 · f ) − s2 ⊗ (s1 · f ) ∈ H 0(C, L) ⊗ H 1(C, L). It follows via Serre duality, that Ker(u) consists of all linear maps v : H 0(C, ωC) → C vanishing (cid:3) on H 0(C, L) ⊗ H 0(C, ωC ⊗ L∨) ⊂ H 0(C, ωC), which proves the claim. 27, where L is locally free and base point free with h0(C, L) = 2 and the monodromy The identifications provided by Theorem 9.3 extend to isomorphisms of vector bundles over a partial compactification of Hur which we shall introduce now. This allows us to express to Hodge classes λ(+1) and λ(−5) in terms of certain tautological classes and define the Petri map globally at the level of the moduli stack. G1 curve) and L is a torsion free sheaf of degree 27 on C with h0(C, L) ≥ 2. Note that G1 [C, L] ∈ G1 σ([C, L]) := [C]. 9.4. Let (cid:102)M46 be the open subvariety of M46 parametrizing irreducible curves and denote by 27 → (cid:102)M46 the stack parametrizing pairs [C, L], where [C] ∈ (cid:102)M46 (in particular C is a stable closed substack of the universal Picard stack of degree 27 over (cid:102)M46. Let GE6 be the locus of pairs of the pencil L is equal to W (E6). We denote by σ : GE6 → (cid:102)M46 the projection map given by 9.5. One has a birational map β : Hur (cid:57)(cid:57)(cid:75) GE6 ⊂ G1 27 which can be extended over each boundary divisor of Hur not contracted under the Prym-Tyurin map (see Theorem 7.17 for a description of of π specialize to R2. We assign to t the point(cid:2)st(C), st(f∗ these divisors). Let t := [π : C = C1 ∪ C2 → R1 ∪q R2] be a general point of one of the divisors DE6, Dazy or Dsyz, where we recall that Ci := π−1(Ri) for i = 1, 2 and only two branch points ∈ GE6, where st is the map assigning to a nodal curve X its stable model st(X) and to a line bundle L on X the line bundle st(L) on st(C) obtained by adding base points to each destabilizing component of X which is contracted. Geometrically, for the general point of each of the divisors DE6, Dazy or Dsyz, the OR1∪R2(1, 0))(cid:3) map β : Hur (cid:57)(cid:57)(cid:75) GE6 contracts the curve C2. We still denote by Dazy, Dsyz and DE6 the images under β of the boundary divisors denoted by the same symbols on Hur. 27 is a locally 36 V. ALEXEEV, R. DONAGI, G. FARKAS, E. IZADI, AND A. ORTEGA C1(OR1(1)) ∈ W 1 We now describe the effect of β along each of the boundary divisors in question. If t ∈ Dazy is a general point, then C1 is smooth and L := π∗ 27(C1) has 6 triple ramification points over the branch point q ∈ R1. Then β(t) = [C1, L] ∈ GE6. If, on the other hand, t is a general point of Dsyz, then retaining the notation of Remark 6.12, C1 is a smooth curve of genus 45, meeting the smooth rational component Z in two points u, v ∈ π−1(q). Then β(t) = [C(cid:48), L], where C(cid:48) := C1/u ∼ v is an irreducible 1-nodal curve of genus 46 and L ∈ W 1 27(C(cid:48)) is the pencil inducing the map π. Finally, if t ∈ DE6, then β(t) = [C(cid:48), L], where C(cid:48) is a 6-nodal curve obtained from C1 by identifying the points of π−1(q) which belong to the same component of C2. 46[D0] + 7 We record the formula λ = 33 92[Dsyz] +··· ∈ CH 1(GE6) for the Hodge class at the level of GE6. The factor of 1 2 in front of [Dsyz] compared to (6.5) is explained by the fact that the general point of Dsyz ⊂ Hur has an automorphism of order 2, whereas its image under β has only trivial automorphsms. 9.6. At the level of GE6 one can introduce several tautological classes along the lines of [Far09]. We denote by f : CE6 → GE6 the universal genus 46 curve and choose a universal line bundle L ∈ Pic(CE6) satisfying the property Lf−1([C,L]) = L ∈ W 1 27(C), for each [C, L] ∈ GE6. We then define the following tautological classes: 46[Dazy] + 17 A := f∗(cid:0)c2 1(ωf )(cid:1) 1(L)(cid:1), B := f∗(cid:0)c1(L) · c1(ωf )(cid:1), κ := f∗(cid:0)c2 ∈ CH 1(GE6). Via Grauert's theorem, we observe that V := f∗L is a locally free sheaf of rank two on GE6. Similarly, the sheaf V2 := f∗(L ⊗2) is locally free of rank 9 over GE6. Globalizing at the level of moduli the multiplication map of global sections Sym2H 0(C, L) → H 0(C, L⊗2), we define the rank 6 vector bundle E2 over GE6 via the following exact sequence: 0 −→ Sym2(V) −→ V2 −→ E2 −→ 0. 9.7. The choice of L is not unique; replacing L by L denoting the corresponding tautological classes by A(cid:48), B(cid:48) relations (cid:48) := L ⊗ f∗(α), where α ∈ Pic(GE6) and ∈ CH 1(GE6) respectively, we find the It follows that B(cid:48) (9.3) A(cid:48) = A + 2 · 27 · α and B(cid:48) = B + (2 · 46 − 2) · α. 3 A(cid:48) = B − 5 − 5 3 A, that is, the class γ := B − 5 3 A ∈ CH 1(GE6) is well-defined and independent of the choice of a Poincar´e bundle L. We now describe in a series of calculations the Chern classes of the vector bundles we have just introduced. Proposition 9.8. The following relations hold in CH 1(GE6): ∨) A 2 − Proof. We apply Grothendieck-Riemann-Roch to f : CE6 → GE6 and write c1(V2) = λ − B + 2A and B 2 − c1(V). = λ + (cid:17) (cid:16) c1 R1f∗(ωf ⊗ L (cid:16) (cid:17) f )(cid:1) = 12λ, see [HM82] p. 49, and conclude. f ) + c2(Ω1 f ) (cid:17)(cid:105) c2 1(Ω1 c1(ωf ) 1 − 12 . 2 + 2 · Now use Mumford's formula f∗(cid:0)c2 c1(V2) = f∗ 1 + 2c1(L) + 2c2 1(L) f ) + c2(Ω1 1(Ω1 (cid:104)(cid:16) (cid:3) THE UNIFORMIZATION OF A6 37 9.9. Theorem 9.2 (to be proved in Section 10) shows that the Petri map µ(L) is injective for a general point of [C, L] ∈ GE6. However, we cannot rule out the (unlikely) possibility that µ(L) is not injective along a divisor N on GE6. We denote by n := [N] ∈ CH 1(GE6). This (possibly zero) class is effective. Globalizing Theorem 9.3, we obtain isomorphisms of vector bundles over GE6 − N: Extending this to GE6, there exists an injection of vector bundles R1f∗(cid:0)ωf ⊗ L E(+1) = R1f∗(cid:0)ωf ⊗ L ⊗ V and E(−5) = E ⊗ V (cid:44)→ E(+1), ∨ 2 ⊗ det(V). with quotient a sheaf supported on N and on possibly other higher codimension cycles. Proposition 9.10. The following formulas hold at the level of GE6: ∨(cid:1) ∨(cid:1) (cid:16) λ(+1) = 2λ − γ + n and λ(−5) = −λ + γ − n. (cid:17) R1f∗(cid:0)ωf⊗L ∨(cid:1) ⊗V −[N] ∈ CH 1(GE6) and the rest is a consequence (cid:3) Proof. We have that λ(+1) = c1 of Theorem 9.3 coupled with Proposition 9.8. Proposition 9.11. We have that A = 27c1(V) ∈ CH 1(GE6). Proof. Recall that GE6 has been defined as a locus of pairs [C, L] such that L is a base point free pencil. In particular, the image under f of the codimension 2 locus in CE6 where the morphism of vector bundles f∗(V) → L is not surjective is empty, hence by Porteous' formula = −27c1(V) + A. V) · c1(L) + c2 V) − c1(f∗ c2(f∗ 1(L) 0 = f∗ (cid:16) (cid:17) (cid:3) Essential in all the ensuing calculations is the following result expressing the divisor Dazy in terms of Hodge eigenbundles and showing that its class is quite positive: Theorem 9.12. The following relation holds: [Dazy] = 5λ + λ(−5) − 3[DE6] − 5 6 [Dsyz] + n ∈ CH 1(GE6). Proof. The idea is to represent Dazy as the push-forward of the codimension two locus in the universal curve CE6 of the locus of pairs [C, L, p] such that h0(C, L(−3p)) ≥ 1. We form the fibre product of the universal curve CE6 together with its projections: CE6 π1←−−− CE6 ×GE6 CE6 For each k ≥ 1, we consider the locally free jet bundle Jk(L) defined, e.g., in [Est96], as a locally free replacement of the sheaf of principal parts P k on CE6. Note that P k f (L) is not locally free along the codimension two locus in CE6 where f is not smooth. To remedy this problem, we consider the wronskian locally free replacements J k f (L), which are related by the following commutative diagram for each k ≥ 1: f (L) := (π2)∗ π2−−−→ CE6. (cid:16) π∗ 1(L) ⊗ I(k+1)∆ (cid:17) 0 0 / Ωk f ⊗ L P k f (L) / P k−1 f (L) / 0 / ω⊗k f ⊗ L / J k f (L) / J k−1 f (L) / 0 Here Ωk canonical map Ωk f denotes the OGE6 f → ω⊗k -module Ik∆/I(k+1)∆. The first vertical row here is induced by the f , relating the sheaf of relative Kahler differentials to the relative dualizing / / /     /   / / / / / 38 V. ALEXEEV, R. DONAGI, G. FARKAS, E. IZADI, AND A. ORTEGA f (L) and J k sheaf of the family f . The sheaves P k f (L) differ only along the codimension two singular locus of f . Furthermore, for each integer k ≥ 0 there is a vector bundle morphism νk : f∗(V) → J k f (L), which for points [C, L, p] ∈ GE6 such that p ∈ Creg, is just the evaluation morphism H 0(C, L) → H 0(L(k+1)p). We specialize now to the case k = 2 and consider the codimension two locus Z ⊂ CE6 where ν2 : f∗(V) → J 2 f (L) is not injective. Then, at least over the locus of smooth curves, Dazy is the set-theoretic image of Z. A simple local analysis shows that the morphism ν2 is simply degenerate for each point [C, L, p], where p ∈ Csing. Taking into account that a general point of Dazy corresponds to a pencil with six triple points aligned over one branch point, and that the stable model of a general element of the divisor Dsyz corresponds to a curve with one node, whereas that of a general point of DE6 to a curve with six nodes and so on, we obtain the formula: 6[Dazy] = f∗c2 − 6[DE6] − 3[Dsyz] ∈ CH 1(GE6). The fact that Dsyz appears with multiplicity 3 is a result of the following local computation. We choose a family F : X → B of curves of genus 46 over a smooth 1-dimensional base B, such that X is smooth, and there is a point b0 ∈ B such that Xb := F −1(b) is smooth for b ∈ B \{b0}, whereas Xb0 has a unique node N ∈ X. Assume also that L ∈ Pic(X) is a line bundle such that Lb := LXb is a pencil with E6-monodromy on Xb for each b ∈ B, and furthermore [Xb0, Lb0] ∈ Dsyz. Choose a local parameter t ∈ OB,b0 and x, y ∈ OX,N , such that xy = t represents the local equation of X x = − dy around the point N . Then ωF is locally generated by the meromorphic differential τ = dx y . We choose two sections s1, s2 ∈ H 0(X, L), where s1 does not vanish at N and s2 vanishes with order 2 at N along both branches of Xb0. Then we have the relation s2,N = (x2 + y2)s1,N between the germs of the two sections s1 and s2 at N . We compute In local coordinates, the map H 0(cid:0)Xb0, Lb0 (cid:18) 1 of the following matrix: d(s2) = 2xdx + 2ydy = 2(x2 − y2)τ, and d(x2 − y2) = 2(x2 + y2)τ. → H 0(cid:0)Xb0, Lb03N (cid:1) is then given by the 2 × 2 minors (cid:19) (cid:1) 0 0 (cid:19) (cid:18) J 2 f (L) f∗(V) x2 + y2 x2 − y2 x2 + y2 . We compute: c1(J 2 hence f∗c2 (cid:19) (cid:18) J 2 f (L) f∗(V) This completes the proof that [Dsyz] appears with multiplicity 3 in the degeneracy locus. f (L)) = 3c1(L) + 3c1(ωf ) and c2(J 2 f (L)) = 3c2 1(L) + 6c1(L) · c1(ωf ) + 2c2 1(ωf ), = 3A + 6B − 3(d + 2g − 2)c1(V) + 2κ1 = 6γ + 2κ1. Furthermore, κ1 = 12λ − 6[DE6] − [Dsyz] − ··· , hence after applying Proposition 9.10, we obtain (cid:3) the claimed formula. We can also express the divisors Dsyz and Dazy in terms of the Hodge eigenclasses. Proposition 9.13. The following formulas hold in CH 1(GE6): 33 8 n and [Dsyz] = λ(−5) + [Dazy] = [DE6] + 25 16 51 16 51 16 λ + 3 4 21 8 λ(−5) − 9 2 λ − [DE6] − 21 8 n. Proof. Combine Theorem 9.12 with the expression of the Hodge class λ in terms of the boundary (cid:3) divisor classes on GE6. Corollary 9.14. We have [Dsyz] ≤ 33 8 λ(−5) − 9 8 λ − 21 2[DE6]. THE UNIFORMIZATION OF A6 39 We are now in a position to determine the class of the ramification divisor of the Prym-Tyurin map in terms of the classes already introduced. Recall that D6 := A6 \ A6 is the irreducible boundary divisor of the perfect cone compactification of A6 and λ1 ∈ CH 1(A6) denotes the Hodge class. Note that KA6 Theorem 9.15. The ramification divisor of the map P T : GE6 = 7λ1 − [D6], see [Mum83]. (cid:57)(cid:57)(cid:75) A6 is given by (cid:2)Ram(P T )(cid:3) = 73 32 λ − 221 32 λ(−5) − 9 8 [DE6] + 3 32 n. for all the other boundary divisors in br∗((cid:101)B2) the corresponding torus rank is zero. Moreover Proof. The general point of DE6 corresponds to a semi-abelian variety of torus rank 1, whereas P T ∗(D6) = DE6 (recall that in this proof, the map P T is defined on the partial compactification GE6, the formula above does not hold on Hur). Via the Hurwitz formula, we obtain that (cid:2)Ram(P T )(cid:3) = KGE6 − P T ∗(cid:0)7λ1 − [D6](cid:1) = KGE6 − 7λ(−5) + [DE6]. Recall that the canonical class KHur has been expressed in terms of boundary divisors on Hur. Using Theorem 9.12, we can pass to a new basis in CH 1(GE6) involving the Hodge eigenbundles and one boundary divisor, namely DE6. After simple manipulations we obtain (9.4) KGE6 = 73 32 λ + 3 32 λ(−5) − 17 8 [DE6] + 3 32 n, which then leads to the claimed formula. (cid:3) We now complete the proof of Theorem 0.3. In what follows, we revert to the notation of the introduction and P T : Hur (cid:57)(cid:57)(cid:75) A6 denotes the extended Prym-Tyurin map. Theorem 9.16. The canonical class of the partial compactification GE6 of Hur is big. It follows that there exists a divisor E on Hur with P T∗(E) = 0, such that KHur + E is big. Proof. The varieties GE6 and Hur differ in codimension one only along boundary divisors that are collapsed under the Prym-Tyurin map. Showing that KGE6 is big implies therefore the second half of the claim, and thus Theorem 0.3. Using Theorem 6.14 (note the caveat about the already mentioned factor 1 2 in front of the coefficient of [Dsyz] when passing from Hur to GE6), coupled with Proposition 9.13, we write: (cid:16)25 16 17 46 (cid:17) (cid:17) λ + λ(−5) + 51 16 3 4 [DE6] KGE6 25 46 = − [DE6] + 19 46 [Dsyz] + 17 46 [Dazy] ≥ − 867 λ(−5) + 736 = (cid:16) 25 46 425 736 19 46 196 425 [DE6] + [Dsyz] + λ − [DE6] . Putting Proposition 6.19 together with the fact that λ(−5) is big, the conclusion follows by com- paring the ratio of the λ and [DE6]-coefficients of the last expression. Indeed, it is shown in (6.20) by pulling-back the Moriwaki class from M46 that the Q-class λ − 6 31[DE6] is is big. (cid:3) effective on GE6. It follows that λ − 196 425[DE6] is then also an effective class, hence KGE6 [DE6] = λ − 23 8+ 4 g 10. The ramification divisor of the Prym-Tyurin map Theorem 0.4. The aim of this section is to describe the differential of the Prym-Tyurin map P T and prove In this section, the tangent spaces we consider are those of the corresponding moduli stacks. As in the previous section, we fix a smooth E6-cover π : C → P1 with branch divisor B := p1 + ··· + p24 and denote L := π∗(OP1(1)). (cid:88) x∈f−1(y) tr(ϕ)(y) = ϕ(x), 40 V. ALEXEEV, R. DONAGI, G. FARKAS, E. IZADI, AND A. ORTEGA Via the ´etale map br : Hur → M0,24/S24, we identify the cotangent space T ∨ H 0(P1, ω⊗2 [P (C,f )](A6) is identified with Sym2 H 0(C, ω⊗2 P1 (B)). The cotangent space T ∨ [C,π](Hur) with C )(−5). Definition 10.1. Let R and A be the ramification and antiramification divisors of π, that is, the effective divisors of C defined by the formulas π∗(B) = 2R + A, KC = π∗(KP1) + R, 2KC + A = π∗(2KP1 + B). Definition 10.2. Let tr : π∗OC(−A) → OP1 be the trace map on regular functions. For an open affine subset U ⊂ P1, a regular function ϕ ∈ Γ(U,OC(−A)), and a point y ∈ U , one has (cid:0)2KP1 + B(cid:1) be the (cid:0)Hur(cid:1) is given by the fol- A6 → T ∨ C ) Tr−→ H 0(ω⊗2 counted with multiplicities. Note that tr is surjective. Let π∗OC(2KC) → OP1 induced trace map at the level of quadratic differentials. We denote the corresponding map on global sections by Tr : H 0(C, ω⊗2 Theorem 10.3. The codifferential (dP T )∨ lowing composition of maps: C ) → H 0(P1, ω⊗2 [C,π] : T ∨ P1 (B)). [P T (C,π)] (cid:0) (cid:1) [C,π] Sym2 H 0(ωC)(−5) (cid:44)→ Sym2 H 0(ωC) mul−−→ H 0(ω⊗2 P1 (p1 + ··· + p24)). Proof. The second map is the codifferential of the Torelli map M46 → A46. The first map is the codifferential of the map from the moduli space of ppav of dimension 46 together with an endomorphism D having eigenvalues (+1) and (−5) (with eigenspaces of dimensions 20 and 6 (cid:3) respectively) to A6. The third map is the codifferential of the map Hur → M46. 10.4. We now analyze the differential dP T at a point [C, π] ∈ Hur in detail. For each of the 24 branch points pi ∈ P1, let {rij}6 j=1 ⊂ C be the ramification points lying over pi. The formal neighborhoods of the points rij are naturally identified, so that we can choose a single local parameter x and write any quadratic differential γ ∈ H 0(C, ω⊗2 C ) as γ = ϕij(x) · (dx)⊗2 near rij ∈ C. Choose a local parameter y at the point pi, so that π is given locally by the map y = x2. We can use the same local parameter at the remaining 15 antiramification points {qik}15 k=1 over pi at which π is unramified, and write γ = ψik(y) · (dy)⊗2 near qik ∈ C, for k = 1, . . . , 15. Lemma 10.5. The kernel of the trace map Tr : H 0(C, ω⊗2 dratic differentials γ such that P1 (B)(cid:1) consists of qua- C ) → H 0(cid:0)P1, ω⊗2 Proof. From y = x2, we get dy = 2x dx and (dx)⊗2 = (dy)⊗2/4y. We have that j=1 ϕij(rij) = 0, 6(cid:88) 6(cid:88) 15(cid:88) (cid:0)ϕij(x) + ϕij(−x)(cid:1) + (cid:80)6 the 24 expressions are zero. Then Tr(γ) ∈ H 0(P1, ω⊗2 Suppose Tr(γ) = 0. Then the leading coefficient 1 2 Tr(γ) = (cid:32) 1 4y k=1 j=1 for i = 1, . . . , 24. (cid:33) ψik(y) · (dy)⊗2 near pi. j=1 ϕij(rij) is zero. Conversely, assume that (cid:3) P1 ) = 0. THE UNIFORMIZATION OF A6 41 In order to understand the condition in Lemma 10.5, we recall the action of the endomorphism D : H 0(C, ωC) → H 0(C, ωC) induced by the Kanev correspondence in local coordinates at the 10.6. The unramified case. Suppose that π is unramified at p, thus Γ := π−1(p) =(cid:80)27 points p ∈ C and q ∈ π−1(p), see also Theorem 9.3. Assume p = [0 : 1] ∈ P1. One has(cid:80)27 correspondence on (αs) is described by an endomorphism of O26 =(cid:8)(cid:80)27 s=1 qs. Since π is ´etale, we can use the same local parameter y at p, as well as at each qs ∈ C. Let α ∈ H 0(C, ωC). In a formal neighborhood of each point qs, we write locally α = αs(y)dy. s=1 αs(qs) = 0. The action of the ⊂ O27, where O = OP1,p. This endomorphism is given by the same integral (26 × 26)-matrix as the action of D on H 0(OΓ(Γ))/H 0(C, L), as in the proof of Theorem 9.3. Thus, D has two eigenvalues (+1) and (−5) with the eigenspaces of dimensions 20 and 6 respectively. Choose a basis {vm}6 m=1 of the (−5)-eigenspace in Z26. Then an element α(−5) ∈ H 0(C, ωC)(−5) can be locally written uniquely as ) = 0, so(cid:80)27 s=1 αs = 0(cid:9) s=1 Resqs(α· x0 x1 6(cid:88) m=1 α(−5) = δmvm ∈ O27, for some δm ∈ O. 10.7. The ramified case. Suppose π is branched at p and π−1(p) consists of ramification points r1, . . . , r6 and 15 antiramification points qk. The points ri correspond to the ordered pairs (ai, bi) of sheets coming together. On the sheets, the correspondence is defined by (cid:88) j(cid:54)=i ai (cid:55)→ (cid:88) j(cid:54)=i bi (cid:55)→ (bj + cij) and (aj + cij), for i = 1, . . . , 6. As above, we use a local coordinate y for p ∈ P1 and the 15 points qk ∈ C, and a local coordinate x for the ramification points ri, with y = x2. Thus, we write locally α = αri(x)dx near ri, and α = αqk(y)dy near qk. The local involution x (cid:55)→ −x splits the differential form into the odd and even parts: αri(x)dx = αodd αri(−x)d(−x) = −αodd ri 2αev ri ri (x2)dx + αev ri (x2)dx + αev ri (x2)xdx, (x2)xdx. The even part can be written in terms of y as 1 and we claim that they do not mix with the 15 sheets on which π is ´etale: (y)dy. The odd parts have no such interpretation Lemma 10.8. The correspondence D induces an endomorphism on the 6-dimensional Oy-module of odd parts (cid:104)αodd dx(cid:105). It is given by a matrix which has 0 on the main diagonal and (−1) else- where. The (−5) eigenspace is 1-dimensional with generator (1, . . . , 1), and, for every element α ∈ H 0(C, ωC)(−5), one has ri (αodd r1 , . . . , αodd r6 ) = (φ, . . . , φ), for some φ = φ(y) independent of i = 1, . . . , 6. Proof. This case is obtained by taking a limit of the unramified case. We work in the complex- analytic topology. A ramification point r ∈ C is a limit of two points qai, qbi ∈ C on the sheets ai and bi respectively. The local parameters x, x(cid:48) at qai, qbi are identified as x(cid:48) = −x. One has y = x2 = (x(cid:48))2 and we look at the limit as x tends to 0. From dx(cid:48) = −dx it follows αqai αqbi (x)dx = αodd (x)dx = −αodd ri ri (x2)dx + αev ri (x2)dx + αev ri (x2)xdx, (x2)xdx. 42 V. ALEXEEV, R. DONAGI, G. FARKAS, E. IZADI, AND A. ORTEGA (cid:80) Under the correspondence between the 27 sheets, the αodd contributions from ai and bi to the 5 sheets cij (j (cid:54)= i) cancel out. Conversely, the terms αcij contribute to αev but not to αodd at the point ri. It follows that the homomorphism D sends the O6 block of the odd parts αodd to itself. The matrix of this linear map is the same as the matrix of an endomorphism of Z6 with the basis of vectors ai − bi, that is, ai − bi (cid:55)→ − j(cid:54)=i(aj − bj). It is easy to see that this linear map has eigenvalues (+1) and (−5) and that the (−5)-eigenspace is one-dimensional and is generated by (cid:3) the vector (1, . . . , 1). The statement now follows. Corollary 10.9. Let β ∈ Sym2 H 0(C, ωC)(−5) and let γ = mul(β) be its image in H 0(C, ω⊗2 C ). Then in the notation of Lemma 10.5, one has ϕij(rij) = ϕij(cid:48)(rij(cid:48)), for all i = 1, . . . , 24 and all 1 ≤ j, j(cid:48) Proof. Let α, α(cid:48) mul(α ⊗ α(cid:48))(rij) = αodd (rij) = φ(0)φ(cid:48)(0), which is independent of j = 1, . . . , 6. Lemma 10.8 has consequences for the geometry of the Abel-Prym-Tyurin canonical curve Then in the notation of Lemma 10.8, one has (cid:3) ∈ H 0(C, ωC)(−5). (rij) · (α(cid:48))odd ≤ 6. rij rij ri In stark contrast with the case of ordinary Prym-canonical curves, the map ϕ(−5) is far from being an embedding. ϕ(−5) = ϕH 0(C,ωC )(−5) : C → P5. i = 1, . . . , 24. Proposition 10.10. For an E6-cover π : C → P1, we have ϕ(−5)(ri1) = ··· = ϕ(−5)(ri6), for each Proof. This is a consequence of Lemma 10.8: the condition that α ∈ H 0(C, ωC)(−5) vanishes along the divisor ri1 + ··· + ri6 is expressed by a single condition φ(0) = 0, therefore (cid:3) dim(cid:12)(cid:12)H 0(C, ωC)(−5)(−ri1 − ··· − ri6)(cid:12)(cid:12) = 4. Finally we are in a position to describe the ramification divisor of the map P T . Like in the classical Prym case, it turns out that the infinitesimal study of the Prym-Tyurin map can be reduced to the projective geometry of he Abel-Prym-Tyurin curve: Proof of Theorem 0.4. Using Lemma 10.5 and Corollary 10.9, it follows that the map P T is ramified at a point [C, π] ∈ Hur, if and only if there exists 0 (cid:54)= β ∈ Sym2 H 0(C, ωC)(−5) such that mul(β) ∈ H 0(C, 2KC − R) = H 0(C, KC − 2L) = Ker(µ(L)), where the last equality follows from the Base Point Free Pencil Trick applied to the Petri map (cid:3) µ(L). If now µ(L) is injective, it follows that mul(β) = 0, which finishes the proof. 11. A Petri theorem on Hur We now prove the Petri-like Theorem 9.2, using a degeneration similar to the one used to establish the dominance of the map P T . We start with a cover πt : Ct → P1 ramified in 24 points such that the local monodromy elements are reflections wi in 12 pairs of roots r1, . . . , r12 generating the lattice E6. We consider a degeneration in which the 12 pairs of roots with the same label come together. The degenerate cover π : C → P1 is branched at 12 points q1, . . . , q12 ∈ P1. Over each point qi there are 6 simple ramification points. The curve C is nodal with 12 × 6 = 72 ordinary double points. We record the following fact: Lemma 11.1. The curve C has 27 irreducible components isomorphic to P1 and the restriction of π to each of them is an isomorphism. THE UNIFORMIZATION OF A6 43 Note that the cover π is not admissible in the sense of Section 5. The corresponding admissible points pi, pi+12, and modifying the curve C accordingly. cover is obtained by replacing each point qi ∈ P1 by an inserted P1 with two additional marked 11.2. The 27 irreducible components {Xs (cid:39) P1}27 s=1 of C are in bijection with the lines {(cid:96)s}27 s=1 on a cubic surface. Let Γ be the dual graph of C. For each root ri with i = 1, . . . , 12, there are 6 pairs of lines (aij, bij) such that ri · aij = 1, bij = aij + ri, hence ri · bij = −1. To each pair we associate an edge (aij, bij) of Γ directed from the vertex aij to the vertex bij. We also fix 12 ramification points qi ∈ P1 \ {0,∞} = C∗ and denote by {psi}ns i=1 the nodes of C lying on Xs. Clearly, π(psi) ∈ {q1, . . . , q12}, for all s and i. 27(cid:77) Lemma 11.3. The space H 0(C, ωC) is naturally identified with C(aij, bij) → (cid:1)(cid:17) To an edge (aij, bij) over a root ri one associates a differential form ωij equal to dz z−qi − dz z−qi on Xbij and 0 on Xs, for s (cid:54)= aij, bij. Then H 0(C, ωC) is the subspace of of the forms ω = (cid:80) cijωij, such that for 1 ≤ s ≤ 27 the sum of residues of ω on Xs is zero. H1(Γ, C) = Ker (cid:110) 12(cid:77) H 0(cid:16) (cid:0) ns(cid:88) 6(cid:77) 27(cid:77) on Xaij , to Xs, KXs (cid:111) C(cid:96)s psi Equivalently, for each 1 ≤ s ≤ 27, one considers a space of forms s=1 s=1 i=1 i=1 j=1 . (cid:88) ωs = s∈{aij ,bij} ci dz z − qi (cid:88) i , such that ci = 0. Then a form ω ∈ H 0(C, ωC) is equivalent to a collection of log forms {ωs}27 conditions Resqi(ωaij ) + Resqi(ωbij ) = 0, for each edge (aij, bij) of Γ. Proof. This follows by putting together two well known facts: (1) Let C be a nodal curve with normalization ν : (cid:101)C → C and nodes pi ∈ C such that ν−1(pi) = ω ∈ H 0(cid:16)(cid:101)C, K(cid:101)C( (cid:88) i qi)(cid:1) is a linear combination(cid:80) (2) A section of H 0(cid:0)P1, KP1((cid:80) i }. Then H 0(C, ωC) is identified with the space of sections , with(cid:80) i + p− (p+ i )) satisfying Resp+ s=1 satisfying the 72 (ω) + Resp− i ci = 0. i , p− (ω) = 0. {p+ (cid:17) i ci (cid:3) i i dz z−qi In practice, assume that H 0(Xs, ωCXs) is identified with the space of fractions (cid:81)ns Ps(x) i=1(x − π(psi)) dx, (cid:76)27 (cid:81) i(cid:54)=j Ps(π(psj)) (cid:0)π(psj) − π(psi)(cid:1) s=1 H 0(Xs, ωCXs) is charac- where Ps(x) is a polynomial of degree ns − 2. Then H 0(C, ωC) ⊆ terized by the condition that for every node of psj = ps(cid:48)j(cid:48) ∈ C joining components Xs and Xs(cid:48), the sum of the residues and Ress(cid:48) respectively is 0. Ress := 44 V. ALEXEEV, R. DONAGI, G. FARKAS, E. IZADI, AND A. ORTEGA 11.4. We wish to show the injectivity of the Petri map µ(L) : H 0(C, L) ⊗ H 0(C, ωC ⊗ L∨) → H 0(C, ωC) for a 72-nodal curve C corresponding to a cover π : C → P1 as above. Lemma 11.5. Let H 0(P1,O(1)) = (cid:104)x0, x1(cid:105) ⊂ H 0(C, L). Then the subspace s=1 as above, satisfying for each 1 ≤ s ≤ 27 the additional condition H 0(C, ωC ⊗ L∨) ⊗ (cid:104)x0(cid:105) ⊂ H 0(C, ωC ⊗ L∨) ⊗ H 0(C, L) ⊂ H 0(C, ωC) consists of elements {ωs}27 {ωs}27 Proof. We identify H 0(C, ωC ⊗ L∨) with H 0(C, ωC(−π∗ space of forms {ωs}27 (cid:80) ciqi = 0. Similarly, the subspace H 0(C, ωC ⊗ L∨) ⊗ (cid:104)x1(cid:105) ⊆ H 0(C, ωC) consists of elements s=1 as above, satisfying for 1 ≤ s ≤ 27 an additional condition(cid:80) ci (cid:88) (cid:1) = − ∞)), hence H 0(C, ωC ⊗ L∨) ⊗ (cid:104)x0(cid:105) is the s=1 such that for each s = 1, . . . , 27, they satisfy the equality 0 = Res∞(cid:0)ωs · (cid:0)ωs · = 0. qi (cid:88) (cid:1) = − x1 x0 ciqi. Resqi x1 x0 (cid:3) To prove Theorem 9.2, it is sufficient to find one degeneration π : C → P1 such that (1) h0(C, L) = 2. generate the vector space H 0(C, ωC). (2) The linear subspaces H 0(C, ωC ⊗ L∨) ⊗ (cid:104)x0(cid:105), H 0(C, ωC ⊗ L∨) ⊗ (cid:104)x1(cid:105) and H 0(C, ωC)(−5) The initial input consists of 12 points qi ∈ P1 \ {0,∞} = C∗, and 12 roots ri generating the lattice E6. We obtain a system of linear equations in the 72 variables xij = Resqi(ωaij ) = −Resqi(ωbij ), for i = 1, . . . , 12 and j = 1, . . . , 6. For each of the spaces H 0(C0, ωC ⊗ L∨) ⊗ (cid:104)x0(cid:105), respectively H 0(C, ωC ⊗ L∨) ⊗ (cid:104)x1(cid:105), we get 2 × 27 equations. By Lemma 10.8, H 0(C, ωC)(−5) is the subspace of H 0(C, ωC) of forms {ωs}27 s=1 satisfying xij = xij(cid:48) for all 1 ≤ j, j(cid:48) ≤ 6 and i = 1, . . . , 12. This gives a system of 27 + 12 × 5 equations. Lemma 11.6. The above conditions are satisfied for the following choices of roots and ramification points: (1) r1 = α135, r2 = α12, r3 = α23, r4 = α34, r5 = α45, r6 = α56, r7 = αmax, r8 = α124, r9 = α234, r10 = α35, r11 = α13, r12 = α36. (2) qi = i, for i = 1, . . . , 12. Proof. This is now a straightforward linear algebra computation, which we performed in Mathe- (cid:3) matica. It can be found at [Web15]. References [ACV03] Dan Abramovich, Alessio Corti and Angelo Vistoli, Twisted bundles and admissible covers, Comm. [Ale02] [Ale04] [AB12] Algebra 31 (2003), 3547 -- 3618. Valery Alexeev, Complete moduli in the presence of semiabelian group action, Annals of Math. 155 (2002), 611 -- 708. Valery Alexeev, Compactified Jacobians and Torelli map, Publ. Res. Inst. Math. Sci. 40 (2004), 1241 -- 1265. Valery Alexeev and Adrian Brunyate, Extending the Torelli map to toroidal compactifications of Siegel space, Invent. Math. 188 (2012), 175 -- 196. [ABH02] Valery Alexeev, Christina Birkenhake, and Klaus Hulek, Degenerations of Prym varieties, J. Reine Angew. Math. 553 (2002), 73 -- 116. THE UNIFORMIZATION OF A6 45 [ACGH85] Enrico Arbarello, Maurizio Cornalba, Philip Griffiths and Joseph Harris, Geometry of Algebraic Curves, [AV12] [Bea77] [BdS49] [Bor72] [CF05] [CG72] [Dol12] [DS81] [Don84] [Don92] [DL11] [Dyn52] [Est96] [Far09] [FG03] [FL10] Volume I, Springer, Berlin, 1985. Maksym Arap and Robert Varley, On algebraic equivalences among the 27 Abel-Prym curves on a generic abelian 5-fold, arXiv:1206.6517, to appear in International Math. Research Notices. Arnaud Beauville, Prym varieties and the Schottky problem, Invent. Math. 41 (1977), 149 -- 196. Armand Borel and J. de Siebethal, Les sous-groupes ferm´es de rang maximum des groupes de Lie clos, Comment. Math. Helv. 23, (1949), 200 -- 221. Armand Borel, Some metric properties of arithmetic quotients of symmetric spaces and an extension theorem, J. Diff. Geom. 6 (1972), 543 -- 560. Sebastian Casalaina-Martin and Robert Friedman, Cubic threefolds and abelian varieties of dimension five, J. Algebraic Geom. 14 (2005), 295 -- 326. Herbert Clemens and Phillip Griffiths, The intermediate Jacobian of the cubic threefold, Annals of Math. 95 (1972), 281 -- 356. Igor V. Dolgachev, Classical algebraic geometry: A modern view, Cambridge University Press, Cam- bridge, 2012. Ron Donagi and Roy Smith, The structure of the Prym map, Acta Math. 146 (1981), 25 -- 102. Ron Donagi, The unirationality of A5, Annals of Math. 119 (1984), 269 -- 307. Ron Donagi, The fibers of the Prym map, Curves, Jacobians, and abelian varieties (Amherst, MA, 1990) Contemp. Math., vol. 136, Amer. Math. Soc., Providence, RI, 1992, 55 -- 125. M. J. Dyer and G. I. Lehrer, Reflection subgroups of finite and affine Weyl groups, Transactions Amer- ican Math. Soc. 363 (2011), 5971 -- 6005. E. Dynkin, Semisimple subalgebras of semisimple Lie algebras, Mat. Sbornik N.S. 30 (1952), 349 -- 462. Eduardo Esteves, Wronski algebra systems on families of singular curves, Ann. Scient. ´Ec. Norm. Sup. 29 (1996), 107 -- 134. Gavril Farkas, Koszul divisors on moduli spaces of curves, American J. Math. 131 (2009), 819 -- 869. Gavril Farkas and Angela Gibney, The Mori cones of moduli spaces of pointed curves of small genus, Transactions American Math. Soc. 355 (2003), 1183 -- 1199. Gavril Farkas and Katharina Ludwig, The Kodaira dimension of the moduli space of Prym varieties, J. European Math. Society 12 (2010), 755 -- 795. [FGSMV] Gavril Farkas, Sam Grushevsky, Riccardo Salavati Manni and Alessandro Verra, Singularities of theta [FV16] divisors and the geometry of A5, J. European Math. Society 16 (2014), 1817 -- 1848. Gavril Farkas and Alessandro Verra, The universal abelian variety over A5, Ann. Scient. ´Ec. Norm. Sup. 49 (2016), 521 -- 543. [Fra01] William N. Franzsen, Automorphisms of Coxeter groups, PhD thesis, University of Sydney, 2001. http: //www.maths.usyd.edu.au/u/PG/Theses/franzsen.pdf [vdGK12] Gerard van der Geer and Alexis Kouvidakis, The Hodge bundle on Hurwitz spaces, Pure Appl. Math. [ILS09a] [ITW16] [HM82] [IL12] Quarterly 7 (2011), 1297 -- 1308. Joe Harris and David Mumford, On the Kodaira dimension of Mg, Invent. Math. 67 (1982), 23 -- 88. Elham Izadi and Herbert Lange, Counter-examples of high Clifford index to Prym-Torelli, J. Algebraic Geom. 21 (2012), no. 4, 769 -- 787. Elham Izadi, Herbert Lange, and Volker Strehl, Correspondences with split polynomial equations, J. Reine Angew. Math. 627 (2009), 183 -- 212. Elham Izadi, Csilla Tamas and Jie Wang, The primitive cohomology of the theta divisor of an abelian fivefold, J. Algebraic Geom., to appear. Elham Izadi and Duco van Straten, The intermediate Jacobians of the theta divisors of four-dimensional principally polarized abelian varieties, J. Algebraic Geom. 4 (1995), no. 3, 557 -- 590. [Kan87] Vassil Kanev, Principal polarizations on Prym-Tyurin varieties, Compositio Math. 64 (1987), 243 -- 270. [Kan89a] Vassil Kanev, Intermediate Jacobians and Chow groups of threefolds with a pencil of del Pezzo surfaces, [IvS95] Annali Mat. Pura ed Applicata 154 (1989), 13 -- 48. [Kan89b] Vassil Kanev, Spectral curves, simple Lie algebras, and Prym-Tjurin varieties, Theta functions -- [Kan06] Bowdoin 1987, Proc. Sympos. Pure Math., vol. 49, Amer. Math. Soc. 1989, 627 -- 645. Vassil Kanev, Hurwitz spaces of Galois coverings of P1, whose Galois groups are Weyl groups, J. Algebra 305 (2006), 442 -- 456. 46 V. ALEXEEV, R. DONAGI, G. FARKAS, E. IZADI, AND A. ORTEGA [KM13] Sean Keel and James McKernan, Contractible extremal rays on M0,n, in Handbook of Moduli Vol. 2, 115 -- 130, International Press 2013. [KKZ11] Alexey Kokotov, Dmitry Korotkin and Peter Zograf, Isomonodromic tau function on the space of ad- [Kol90] [LR08] [Mor98] [MM83] missible covers, Advances Math. 227 (2011), 586 -- 600. J´anos Koll´ar, Projectivity of complete moduli, J. Differential Geometry 32 (1990), 235 -- 268. Herbert Lange and Anita M. Rojas, A Galois-theoretic approach to Kanev's correspondence, Manuscripta Math. 125 (2008), 225 -- 240. Atsushi Moriwaki, Relative Bogomolov inequality and the cone of positive divisors on the moduli space of stable curves, J. Amer. Math. Soc. 11 (1998), 569 -- 600. Shigeyumi Mori and Shigeru Mukai, The uniruledness of the moduli space of curves of genus 11, Springer Lecture Notes in Mathematics 1016 (1983), 334 -- 353. [Mum83] David Mumford, On the Kodaira dimension of the Siegel modular variety, Algebraic geometry -- open [M74] problems (Ravello, 1982), Lecture Notes in Math., vol. 997, Springer, 1983, 348 -- 375. David Mumford, Prym varieties I, Contributions to Analysis (a collection of papers dedicated to Lipman Bers), 325350, New York, Academic Press 1974, [Nam73] Yukihiko Namikawa, On the canonical holomorphic map from the moduli space of stable curves to the Igusa monoidal transform, Nagoya Math. J. 52 (1973), 197 -- 259. [Nam76] Yukihiko Namikawa, A new compactification of the Siegel space and degeneration of Abelian varieties II, Math. Ann. 221 (1976), no. 3, 201 -- 241. Toshio Oshima, A classification of subsystems of a root system, arXiv:math/0611904 (2006). Alessandro Verra, A short proof of the unirationality of A5, Indagationes Math. 46 (1984), 339 -- 355. [Osh06] [Ver84] [Web15] Website with supporting computer computations, http://alpha.math.uga.edu/~valery/a6. [Wir95] Wilhelm Wirtinger, Untersuchungen uber Thetafunktionen, Teubner, Leipzig, 1895. Valery Alexeev: Department of Mathematics, University of Georgia Athens GA 30602, USA E-mail address: [email protected] Ron Donagi: Department of Mathematics, University of Pennsylvania 209 South 33rd Street, Philadelphia, PA 19104-6395, USA E-mail address: [email protected] Gavril Farkas: Institut fur Mathematik, Humboldt-Universitat zu Berlin Unter den Linden 6, 10099 Berlin, Germany E-mail address: [email protected] Elham Izadi: Department of Mathematics, University of California, San Diego La Jolla, CA 92093-0112, USA E-mail address: [email protected] Angela Ortega: Institut fur Mathematik, Humboldt-Universitat zu Berlin Unter den Linden 6, 10099 Berlin, Germany E-mail address: [email protected] 

1009.0121

6

1009

Construction of schemes over $F_1$, and over idempotent semirings: towards tropical geometry

In this paper, we give some categorical description of the general spectrum functor, defining it as an adjoint of a global section functor. The general spectrum functor includes that of $F_1$ and of semirings.

math.AG

math

Construction of schemes over F1, and over idempotent semirings: towards tropical geometry Satoshi Takagi October 25, 2018 Abstract In this paper, we give some categorical description of the general spec- trum functor, defining it as an adjoint of a global section functor. The general spectrum functor includes that of F1 and of semirings. Contents 0 Introduction 0.1 Notation and conventions . . . . . . . . . . . . . . . . . . . . . . 0.2 Preliminaries . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1 Complete algebras 1.1 Complete sets . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1.2 Semirings . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2 Modules over a semiring 2.1 R-modules . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2.2 Tensor products . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3 Congruence relations Idealic semirings . . . . . . . . . . . . . . . . . . . . . . . . . . . 3.1 3.2 Congruence relations . . . . . . . . . . . . . . . . . . . . . . . . . 4 Topological spaces 4.1 The spectrum functor . . . . . . . . . . . . . . . . . . . . . . . . 4.2 Localization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4.3 Comparison with the Stone- Cech compactification . . . . . . . . 5 Schemes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5.1 Sheaves Idealic schemes . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5.2 5.3 A -schemes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5.4 Comparison with classical schemes . . . . . . . . . . . . . . . . . 2 3 3 4 4 10 11 11 14 15 15 15 17 17 20 23 25 25 27 28 34 1 0 Introduction The goal of this paper is to provide a explicit description of the mechanism of the spectrum functor; how we construct a topological space and a structure sheaf from an algebraic object, and why it behaves so nicely. At first, we were only planning to give some definitions of schemes con- structed from idempotent semirings motivated from the tropical geometry (cf. [IME]), since still there is no concrete definition of higher dimensional tropical varieties: in [Mik], Mikhalkin is giving a definition, but is very complicated and not used in full strength yet, so we cannot say that the definition is appropriate yet. Also, the algebraic theories of semirings are still inadequate, compared to those of normal algebras, though there are already many demands from various workfields. (cf. [Lit]). Even the tensor products has been constructed only re- cently (cf. [LMS]). So, we had to start from the beginning: giving a foundation of idempotent algebras. Later, we noticed that there should be a comprehensive theory including the [CC], theory of schemes over F1, of which field is now in fashion (cf. [PL], etc.). [TV], We chose the definition of commutative idempotent monoid to be infinitely additive. There are two reasons. First, if we restrict our attention to finite additivity, then the existence of all adjoint functors are already assured, by the general theory of algebraic systems. Secondly, we can handle topological spaces in this framework: a sober space can be regarded as a semiring with idempotent multiplication. Also, we can make it clear how we obtain a infinite operation (i.e. the intersection of closed subsets) from algebras with only finitary operators. This fact shows that, there are adjoints between the three categories: that of algebraic complete idealic semirings, that of algebraic idealic schemes, and that of algebraic sober spaces: (alg.IRng†) ⇆ (alg.ISch)op ⇆ (alg.Sob)op. This indicates that, the genuine theme of algebraic geometry and arithmetic geometry (including schemes over F1) is no longer the investigation of the link between the category of algebras and the category of topological spaces, but that of the link between the category of algebras and the category of idem- potent semirings. Roughly speaking, the whole workfield may be thought as contained in the tropical geometry, if you take the tropical geometry as a study of semirings. After a while, we realized that there are strong resemblance between the theory of lattices and the argument in the first half of this article. We must admit that there are no new ideas contained in this article; most of the techniques and theorems have been already established decades ago. Therefore, readers who attempt to study in this workfield is strongly recommended to overlook the lattice theory (for example, [MMT], [JH]). However, we decided to contain definitions and propositions as self-contained as possible, since there seems to be some discrepancy of languages between each workfields, and one of the purpose 2 of this paper is to introduce connections between the lattice theory, tropical geometry, and schemes over F1. We were irresolute for a while, of which definition of schemes would be the most general and natural one. For this, we decided to give the weakest (in other words, the most general) definition in which the global section functor Γ becomes an adjoint. This clarifies the condition of which algebra could give a topological space endowed with a structure sheaf. article by Also, we know that we can further extend the arguments and constructions in symmetric this monoidal categories, but we dared not, for it would be too abstract for the present demands. commutative monoids with replacing §1 is a summary of algebras and complete algebras. The definition and arguments are all standard -- in other words, preliminary -- in the lattice theory. In §2, we deal with R-modules, where R is a complete semiring. The results cannot be deduced from §1, so we had to argue seperately. In §3, we focus on congruence relations, which is important when considering localizations. In §4 we deal with topological spaces, defining the spectrum functor by an adjoint. This is also standard in the lattice theory. §5 may be the most valuable part, in which we define A -schemes, and define the spectrum functor (not the same one of §4, but also endows the structure sheaf to the underlying space) as an adjoint. All the adjoints we have constructed are illustrated in the end of 5.3. Acknowledgements. We would like to thank Professor Moriwaki for giving me chances to work on this subject. Also, the author is grateful to Doctor Ikoma, for giving useful advises and encouragements. 0.1 Notation and conventions Throughout this paper, all monoids are assumed associative, unital, commuta- tive. We fix a universe, and all sets and algebraic objects are elements of this universe. For any set X, we denote the power set of X by P(X). We frequently make use of the notation P<∞: this means that given a infinite sum, there exists a finite subindex which preserve the equality and in- equality. For example, x ≤ X λ aλ ⇒ x ≤ <∞ X λ aλ means that, if x < Pλ aλ, then there are finitely many λ1,··· , λn such that x < Pn i=1 λi. We also make use of the notation ∪<∞. 0.2 Preliminaries (1) We frequently make use of the categorical languages in [CWM], especially that of adjoint functors: Let U : A → C , F : C → A be two functors 3 between two categories A and C . We say F and U are adjoint, or F is the left adjoint of U , if there is a natural isomorphism of functors HomC (c, U a) ≃ HomA (F c, a) : C op × A → (Set). This is equivalent to saying that there are natural transformations ǫ : IdC ⇒ U F (the unit ), and η : F U ⇒ IdA (the counit ) which satisfy ηF ◦ F ǫ = IdF and U η ◦ ǫU = IdU : EEEEEEEEE EEEEEEEEE EEEEEEEEE EEEEEEEEE U ǫU / U F U U η F F ǫ F U F ηF / F U (2) We will briefly overview the general theory of algebraic varieties (in the lattice theoritic sense). Here we also use the language of [CWM]. An algebraic type σ is a pair hΩ, Ei, where Ω is a set of finitary operators and E is a set of identities. Let D be the set of derived operators, i.e. the minimal set of operators including Ω, and closed under compositions and substitutions. A σ-algebra is a set A, with an action of Ω on A, satisfying the identities. A homomorphism of σ-algebras is a map f : A → B between two σ-algebras, preserving the actions of Ω. We denote by (σ-alg) the category of small σ-algebras. Here are some facts about algebraic types: (a) The category (σ-alg) is small complete and small co-complete. (b) Let f : σ → τ be a homomorphism of algebraic types, i.e. f is a map Ωσ → Dτ of operators which satisfies f (Eσ) ⊂ Eτ . We say that τ is stronger than σ in this case, and denote by σ ≤ τ for brevity. Then, the underlying functor U : (τ -alg) → (σ-alg) has a left adjoint. In particular, there is a functor F : (Set) → (σ-alg) for any algebraic type σ, and the σ-algebra F (S) is called the free σ-algebra generated by S. 1 Complete algebras Almost all the results are already established in the lattice theory. Here, we are just translating it into a somewhat algebro-geometric language. The reader who is well familiar with the lattice theory may skip. 1.1 Complete sets Definition 1.1.1. (1) A multi-subset of a set X is simply, a map f : Λ → X from any set Λ. We denote by P(X) the set of multi-subset of X modulo isomorphism. Often, a multi-subset is identified with its image. (2) A map sup : P(X) → X is a supremum map if it satisfies: 4   / /   (a) The map factors through the power set of X, namely, there is a map P(X) → X making the following diagram commutative: P(X) X sup <zzzzzzzzz Im P(X) where Im sends a map Λ → X to its image. (b) It is (infinitely) idempotent: It sends a constant map (from a non- empty set) to the unique element in its image. (c) It is (infinitely) associative: If f : Λ → X and fi : Si → X are multi- subsets of X satisfying Imf = ∪iImfi, then sup f = sup{sup fi}i. We often write x ⊕ y instead of sup{x, y}. Note that when given a (not neccesarily infinite) associative commutative idempotent operator ⊕ on X, we can define a preorder on X, by defining a ≤ b ⇔ a ⊕ b = b. Conversely, the supremum map gives the supremum with respect to this preorder. The maximal element 1 of X is the absorbing element of X, i.e. sup(S ∪ {1}) = 1 for any subset S of X. Also, there is an infimum of any subset S of X: inf S = sup{x ∈ X x ≤ s for any s ∈ S}, and the minimal element 0 of X is the unit, i.e. sup S ∪ {0} = sup S for any subset S of X. (3) A complete idempotent monoid is a set endowed with a supremum map. (4) Let A, B be two complete idempotent monoids. A map f : A → B is a homomorphism of complete idempotent monoids if it commutes with the supremum map: f∗ f P(A) sup A P(B) sup / B and sending 1 to 1. Note that when this holds, f sends 0 to 0, since 0 = sup∅. We denote by (CIM) the category of complete idempotent monoids. The notion "complete idempotent monoid" is used in [LMS], but in the lattice theory, this is refered to as a complete lattice. We will use the former notation. 5 / /   < / /     / Definition 1.1.2. The algebraic type σ with a binomial operator + is pre- complete if: (1) + is an associative, commutative idempotent operator with a unit 0 and absorbing element 1, (2) There are infimums for any two elements a, b: x ≤ a, b then x ≤ inf(a, b). inf(a, b) ≤ a, b, and if (3) If φ is another n-ary operator, then it is n-linear with respect to +: φ(x1,··· , xi + x′ i,··· , xn) = φ(x1,··· , xi,··· , xn) + φ(x1,··· , x′ i,··· , xn). The condition (1) and (2) are equivalent to saying that a σ-algebra is a distributive lattice with respect to + and inf. Note that any complete idempotent monoid is already pre-complete, if we restrict the supremum map to finite subsets. Also, see that inf is defined alge- braically, i.e. (a) a + inf(a, b) = a. (b) inf(x + a, x + b) + x = inf(x + a, x + b). Definition 1.1.3. Let σ be a pre-complete algebraic type with respect to +. (1) We denote by (σ-alg) the category of σ-algebras. (2) A complete σ-algebra is a σ-algebra A satisfying: (a) A is a complete idempotent monoid, and the supremum map coin- cides with +, when restricted to any finite subset of A. (b) Any n-ary operator φ : An → A is n-linear with respect to +: φ(x1,··· , xij ,··· , xn). xij ,··· , xn) = X φ(x1,··· , X j j (3) A homomorphism f : A → B of complete σ-algebras is a homomorphism of σ-algebras and homomorphism of complete idempotent monoids. (4) We denote by (σ†-alg) the category of complete σ-algebras. We refer to σ† as a complete algebraic type. We will see that the underlying functor U : (σ†-alg) → (σ-alg) has a left adjoint, but we will analyse this left adjoint more precisely, for future references. Definition 1.1.4. Let A be a complete σ-algebra. (1) Let a be an element of A. A subset S of A is a covering of a, if a ≤ sup S. 6 (2) An element a of A is compact, if any covering of a has a finite subcovering of a. (3) We say that A is algebraic, if the following holds: (a) For any element a of A is algebraic, i.e. a has a covering S which consists of compact elements. (b) Any operator (including the infimum of finite elements) φ : An → A preserves compactness, i.e. φ(x1,··· , xn) is compact if xi's are compact elements. (4) A homomorphism f : A → B of algebraic complete σ-algebras is a ho- momorphism of complete σ-algebras, sending any compact element to a compact element. (5) We denote by (alg.σ†-alg) the category of algebraic complete σ-algebras. These notation come from the lattice theory. Note that any constant (regarded as a 0-ary operator) in an algebraic com- plete σ-algebra is compact, by definition. In particular, the absorbing element 1 is compact. Proposition 1.1.5. Let σ be a pre-complete algebraic type. Then, the under- lying functor U : (σ†-alg) → (σ-alg) has a left adjoint comp. Further, comp factors through (alg.σ†-alg): (σ-alg) (σ†-alg) comp 8qqqqqqqqqq U ′ comp′ (alg.σ-alg) Proof. Let A be a σ-algebra. A filter F on A is a non-empty subset of A satisfying: x, y ∈ F ⇔ x + y ∈ F. We denote by hxi the filter generated by an element x ∈ A. Let A† be the set of all filters on A. Given a family of filters F = {Fλ}λ, the supremum of F is the filter generated by Fλ's, i.e. ξ ∈ sup F if and only if there are finite number of xλ ∈ Fλ's such that Pλ xλ ≥ ξ. Then, the unit is {0} and the absorbing element is A. We can easily see that that 1 is compact, and any element is algebraic: a compact filter is precisely, a finitely generated filter. Let φ : An → A be another operator of A. The operator φ† : (A†)n → A† associated to φ is defined by φ†(F1,··· , Fn) = X xi∈Fi hφ(x1,··· , xn)i. This sends a n-uple of compact filters to a compact filter. This gives a structure of an algebraic complete σ-algebra on A†. Given a homomorphism f : A → B 7 / /   8 of σ-algebras, a homomorphism f † : A† → B† of complete σ-algebras is given by f †(F ) = X a∈F hf (a)i. Finally, we will show that comp is the left adjoint of U : It is easy to see that f † sends a compact filter to a compact filter. Hence, we have a functor comp′ : (σ-alg) → (alg.σ†-alg). Set comp = U ′ ◦ comp′, where U ′ : (σ†-alg) → (alg.σ†-alg) is the underlying functor. the unit ǫ : Id(σ-alg) ⇒ U ◦ comp is given by A ∋ a 7→ hai ∈ A†. η : comp◦U ⇒ Id(alg.σ†-alg) is given by B† ∋ F 7→ sup F ∈ B. Remark 1.1.6. The functor comp′ constructed above is not the left adjoint of the underlying functor (alg.σ†-alg) → (σ-alg): the counit η is not algebraic, i.e. it does not necessarily preserve compactness. However, it is an equivalence of categories: see below. Proposition 1.1.7. Let σ be a pre-complete algebraic type. Then the above functor comp′ gives an equivalence between the category of σ-algebras and the category of algebraic complete σ-algebras. Proof. We will construct a functor Ucpt : (alg.σ†-alg) → (σ-alg) as follows: for an algebraic complete σ-algebra R, let Rcpt be the set of compact elements of R. This set has the natural induced structure of a σ-algebra, since all the operators are algebraic. Also, given a homomorphism f : A → B of algebraic complete σ-algebras, we obtain a homomorphism fcpt : Acpt → Bcpt of σ-algebras. Hence, sending R to Rcpt gives a functor Ucpt : (alg.σ-alg) → (σ-alg). We will see that Ucpt is the inverse of comp′. The unit ǫ : Id(σ-alg) ⇒ Ucpt ◦ comp′ is given by A ∋ a 7→ hai ∈ (A†)cpt. This is an isomorphism: the inverse is given by F 7→ sup F . Note that this is well defined, since a compact filter is finitely generated, hence only finitely many elements is involved when taking its supremum. The counit η : comp′ ◦Ucpt ⇒ Id(alg.σ†-alg) is given by (Bcpt)† ∋ F 7→ sup F ∈ B. This is well defined, since all the filters are generated by compact elements of B, hence η is algebraic. The inverse of η is given by B ∋ b 7→ X hb′i ∈ (Bcpt)†, b′≤b where b′ runs through all the compact elements smaller than b. It is clear that η−1 preserves compactness. Also, η−1 preserves any operator φ of σ, since φ preserves compactness. Hence, η−1 is well defined as a homomorphism of algebraic complete σ-algebras. 8 Corollary 1.1.8. Let σ be a pre-complete algebraic type. (1) The underlying functor U ′ : (alg.σ†-alg) → (σ†-alg) has a right adjoint: it is alg = comp′ ◦U , where U : (σ†-alg) → (σ-alg) is the underlying functor. (2) The category (alg. σ†-alg) is small complete and small co-complete. (3) Let τ be another pre-complete algebraic type stronger than σ. Then we have a following natural commutative diagram of functors: (τ -alg) ≃ / (alg.τ †-alg) (σ-alg) ≃ / / (alg.σ†-alg) where the downward arrows are the underlying functors. Hence, the un- derlying functor U : (alg.τ †-alg) → (alg.σ†-alg) has a left adjoint. Proposition 1.1.9. Let σ be a pre-complete algebraic type. (1) The category (σ†-alg) is small complete. (2) The category (σ†-alg) is small co-complete. Proof. (1) We only need to verify that small products and equalizers exist, but this is just the analogue of the case of usual algebras. (2) We need to verify that small co-products and co-equalizers exist. First, f,g we will verify the existence of co-equalizers. Let A ⇒ B be two homo- morphisms between two σ-algebras A and B. Let a be the congruence relation generated by f and g, i.e. the minimal equivalence relation on B satisfying: (a) (f (a), g(a)) ∈ a for any a ∈ A. (b) If (ai, bi) ∈ a for all i, then (φ(a1,··· , an), φ(b1,··· , bn)) ∈ a for any n-ary operator φ. (c) If (aλ, bλ) ∈ a for all λ, then (P aλ, P bλ) ∈ a. Then, it is clear that B/a is the co-equalizer of f and g. Next, we will show that small co-products exist. Let A = {Aλ}λ be a small family of complete σ-algebras. Let S be the epic undercategory of A , i.e. its objects are pairs hB,{fλ : Aλ → B}i satisfying: (a) B is a complete σ-algebra. (b) fλ is a homomorphism of complete σ-algebras. (c) The image of fλ's generate B as a complete σ-algebra. 9 /     Then, we see that the set S is small, and complete when regarded as a category, since (σ†-alg) is complete. The coproduct ∐Aλ is then defined by the limit of S. Proposition 1.1.10. Let σ, τ be two pre-complete algebraic types, and assume τ is stronger than σ. Then, the underlying functor Uv : (τ †-alg) → (σ†-alg) has a left adjoint. Proof. Let Fσ : (σ-alg) ⇆ (σ†-alg) : Uσ, Fτ : (τ -alg) ⇆ (τ †-alg) : Uτ , and Fσ : (σ-alg) ⇆ (τ -alg) : Uσ be adjoints, respectively: (σ-alg) o σ (σ†-alg) o u v (τ -alg) τ / (τ †-alg) Given a complete σ-algebra A, set A′ = FuUσA ∈ (τ -alg). A filter F on A′ is infinite with respect to A, if xλ ∈ F ∩ A implies P xλ ∈ F . Let Fv(A) be the set of infinite filters on A′. When given a family {Fλ} of infinite fil- ters, the supremum filter is the infinite filter generated by Fλ's. The rest of the construction of the complete σ-algebra structure on Fv(A) is analogous to Proposition 1.1.5. Hence, we have a functor Fv : (σ†-alg) → (τ †-alg). We will show that this is the left adjoint of Uv. The unit ǫ : Id(σ†-alg) ⇒ UvFv is given by A ∋ a 7→ hai ∈ Fv(A). Note that this preserves the supremum map, from the advantage of using infinite filters. The counit η : FvUv ⇒ Id(τ †-alg) is given by F 7→ sup F . 1.2 Semirings Definition 1.2.1. (1) An algebraic system (R, +,×) is a semiring if: (a) R is a monoid with respect to + and ×. (b) The distribution law holds: (a + b) · c = (a · c) + (b · c). (2) A semiring R is pre-complete, if it is pre-complete with respect to +, and if the multiplicative unit is the absorbing element with respect to +. (3) We denote by (SRng) (resp. (PSRng)), the category of semirings (resp. pre-complete semirings). Definition 1.2.2. (1) The initial object of (PSRng) consists of two ele- ments: 1 and 0. We refer to this semiring as F1, the field with one element. Here in the category of semirings, F1 is no longer a makeshift concept. 10 o / / O O   O O   o / (2) The terminal object of (SRng) exists, and consists of a unique element 1 = 0. We refer to this semiring as the zero semiring. (3) Let R be a complete semiring. Let Um : (R/SRng†) → (Mnd) be the underlying functor, sending a complete R-algebra to its multiplicative monoid. Then, Um has a left adjoint Fm. For a monoid M , R[M ] := Fm(M ) is the complete monoid semiring with coefficient R. In particular, when M is the free monoid generated by a set S, R[xs]s∈S = F (M ) is the complete polynomial semiring with coefficient R. (4) Let A, B, R be complete semirings, with homomorphisms R → A and R → B. Then we can define the tensor product of A and B over R which we denote by A ⊗R B. Of course, we have infinite tensor products too. Remark 1.2.3. The constructions of various algebras are also valid in the al- gebraic complete case. The arguments and notations will be just the repetition, so we will skip it. 2 Modules over a semiring We will construct a general fundamental theorems on R-modules, where R is a complete semiring. Note that (R-mod) is a little beyond from what we have been calling complete algebras: the definition of the scalar action is not algebraic, since we must consider infinite sums of elements of the coefficient ring R. However, the arguments are analogous. We just modify the definition of the filters when necessary. 2.1 R-modules Definition 2.1.1. Let R be a complete semiring. (1) An R-module M is a complete idempotent monoid endowed with a scalar operation R × M → M which satisfies the following: (a) 1 · x = x, 0 · x = 0, a · 0 = 0 for any x ∈ M and a ∈ R. (b) The distribution law holds: a(Pλ xλ) = Pλ(axλ) for any a ∈ R, xλ ∈ M . (Pλ aλ)x = Pλ(aλx) for any aλ ∈ R, x ∈ M . (c) (ab)x = a(bx) for any a, b ∈ R and x ∈ M . (2) A map f : M → N between two R-modules is a homomorphism of R- modules if it is a homomorphism of complete idempotent monoids, and preserves the scalar operation. (3) We denote by (R-mod) the category of R-modules. Proposition 2.1.2. (1) The category of F1-modules coincides with the cat- egory of complete idempotent monoids. 11 (2) Let R be a complete semiring. For any two R-modules M and N , Hom(R-mod)(M, N ) has the natural structure of a R-module. The proofs are obvious. Proposition 2.1.3. Let R be a complete semiring. (1) The category (R-mod) is complete. (2) The category (R-mod) is co-complete. Proof. (1) It is just the same argument of classical modules. (2) It suffices to show that there is small co-products and coequalizers. The construction of a coequalizer is easy. We will construct a co-product for a small family {Mλ} of R-modules. Set M∞ = Q Mλ: the set theoritic product of Mλ's. We will show that this is also the co-product. The inclusion Mλ → M∞ is given by x 7→ [µ 7→ xδµ,λ] where δ is the Kronecker delta. Given a family of homomorphisms fλ : Mλ → N , define f : M∞ → N by (xλ)λ 7→ P fλ(xλ). We see that this is the unique map. Definition 2.1.4. For a complete semiring R, set R′ be the pre-complete semir- ing, obtained from R by forgetting the infinite sum. (1) A (pre-complete) R′-module is a pre-complete monoid M equipped with a scalar product R′ × M → M of R′, satisfying: (a) 1 · x = x, 0 · x = 0, a · 0 = 0 for any x ∈ M , a ∈ R. (b) The finite distribution law holds: (i) a(x + y) = ax + ay for any a ∈ R, x, y ∈ M . (ii) (a + b)x = ax + bx for any a, b ∈ R, x ∈ M . (c) (ab)x = a(bx) for any a, b ∈ R, x ∈ M . (2) A map f : M → N between two R′-modules is a homomorphism of R′- modules, if it is a homomorphism of pre-complete monoids, and preserves the scalar operation. (3) We denote by (R′-pmod) the category of pre-complete R′-modules. Proposition 2.1.5. The underlying functor U : (R-mod) → (R′-pmod) has a left adjoint. Proof. The proof is analogous to that of Proposition 1.1.5. Let M be a pre- complete R′-module. A R-filter on M is a non-empty subset F of M satisfying: (1) x, y ∈ F ⇔ x + y ∈ F . (2) If aλx ∈ F for any λ, then (P aλ)x ∈ F . 12 Let M ‡ be the set of R-filters on M . Then, M ‡ becomes a R-module. Hence, we can define a functor compR : (R′-pmod) → (R-mod) by M 7→ M ‡. The rest is the repetition of 1.1.5. Remark 2.1.6. Note that compR(M ) is not algebraic, since a infinite sum appears in the definition of the R-filter. However, the situation is different when R is algebraic. Definition 2.1.7. Let R be an algebraic complete semiring. (1) A R-module M is algebraic if: (a) M is algebraic as a complete idempotent monoid. (b) If a ∈ R, x ∈ M are both compact, then so is ax. (2) A homomorphism f : M → N of R-modules is algebraic if it preserves compactness. (3) We denote by (alg.R-mod) the category of algebraic R-modules. If R is an algebraic complete semiring and M is an algebraic R-module, then the pre-complete monoid Mcpt consisting of compact elements of M becomes a Rcpt-module. Hence we have a functor Ucpt : (alg.R-mod) → (Rcpt-pmod). Proposition 2.1.8. The above functor Ucpt gives an equivalence of categories Ucpt : (alg.R-mod) ≃ (Rcpt-pmod). Proof. This is the analogue of Proposition 1.1.7. Let M be a pre-complete Rcpt-module and M ‡ be the set of filters (not the set of R-filters) of M . The scalar operation on M ‡ is given by F · X = X a∈F,x∈X haxi, where F (resp. X ) is a filter on Rcpt (resp. M ). Hence, we obtain a functor comp′ : (Rcpt-pmod) → (alg.R-mod). It is easy to see that this functor gives the inverse of Ucpt. Definition 2.1.9. Let R be a complete semiring, and R′ be the pre-complete semiring obtained from R by forgetting the infinite sum. Let σ be an alge- braic type stronger than R′-module. We denote by σ‡ the complete algebraic type induced by σ, extending the finite scalar operation to the infinite scalar operation. Proposition 2.1.10. Let R be a complete semiring, and R′ be the pre-complete semiring obtained from R by forgetting the infinite sum. Let σ, τ be an algebraic type stronger than R′-module, and suppose τ is stronger than σ. Then, there is a left adjoint functor of the underlying functor Uv : (τ ‡-alg) → (σ‡-alg). 13 Proof. This is the analogue of Proposition 1.1.10. Let Fσ : (σ-alg) ⇆ (σ‡-alg) : Uσ, Fτ : (τ -alg) ⇆ (τ ‡-alg) : Uτ , and Fσ : (σ-alg) ⇆ (τ -alg) : Uσ be adjoints, respectively: (σ-alg) o σ (σ‡-alg) o u v (τ -alg) τ / (τ ‡-alg) Given a σ‡-algebra A, set A′ = FuUσA ∈ (τ -alg). Let Fv(A) be the set of R-filters on A′ which is infinite with respect to A. Then, Fv(A) becomes a τ ‡-algebra, and we obtain the functor Fv : (σ‡-alg) → (τ ‡-alg). This becomes the left adjoint of Uv. Corollary 2.1.11. Let R be a complete semiring. The underlying functor (R-alg) → (R-mod) has a left adjoint S. For any R-module M , S(M ) is the symmetric algebra generated by M . 2.2 Tensor products Throughout this subsection, we fix a complete semiring R. Proposition 2.2.1. For any R-module N , the functor Hom(R-mod)(N,−) : (R-mod) → (R-mod) has a left adjoint. Proof. Let M be another R-module. A non-empty subset F of M × N is a filter if it satisfies the followings: (1) Let x, y and r be elements of M , N and R, respectively. Then (rx, y) ∈ F if and only if (x, ry) ∈ F . (2) Let xλ be elements of M , and y be an element of N . Then, (xλ, y) ∈ F for any λ if and only if (Pλ xλ, y) ∈ F . (3) Let yλ be elements of N , and x be an element of M . Then, (x, yλ) ∈ F for any λ if and only if (x, Pλ yλ) ∈ F . We will denote by Pλ xλ ⊗ yλ the filter generated by {(xλ, yλ)}λ. We define M ⊗R N as the set of all filters. The supremum of {Fλ} is the filter generated by Fλ's. For any scalar r ∈ R and any filter F , the scalar operation is defined by r · F = {(rx, y)}(x,y)∈F . Given a homomorphism f : M → M ′ of R-modules, the homomorphism f ⊗ N : M ⊗ N → M ′ ⊗ N is defined by Pλ xλ ⊗ yλ 7→ Pλ f (xλ) ⊗ yλ. Thus, the functor − ⊗ N : (R-mod) → (R-mod) is well defined. We will prove that this is the left adjoint of Hom(R-mod)(N,−). The unit ǫ : Id ⇒ Hom(N,−⊗RN ) is given by x 7→ [y 7→ x⊗y]. The counit Hom(N,−)⊗RN ⇒ Id is defined by f ⊗ y 7→ f (y). Proposition 2.2.2. Let R be a semiring, and A be a R-algebra. Then, the underlying functor U : (R-mod) → (A-mod) has a left adjoint A ⊗R −. 14 o / / O O   O O   o / Proof. The proof is the analogy of classical algebras. Let M be a R-module. The scalar multiplication of A on A⊗RM is given by a(P bλ⊗xλ) = P abλ⊗xλ. This gives the structure of A-module on A ⊗R M . The unit ǫ : Id(R-mod) ⇒ U ◦ (A ⊗R −) is given by x 7→ 1 ⊗ x. The counit η : (A ⊗R −) ◦ U ⇒ Id(A-mod) is given by P aλ ⊗ yλ 7→ P aλyλ. Definition 2.2.3. Let F be the left adjoint of the underlying functor U : (Set) → (R-mod). For any set S, RS = F (S) is the free R-module generated by S. 3 Congruence relations 3.1 Idealic semirings Definition 3.1.1. (1) A semiring R is idealic, if the multiplicative unit 1 is the maximal element. (2) We denote by (IRng†) the full subcategory of (SRng†) consisting of com- plete idealic semirings. Definition 3.1.2. Let f : A → B a homomorphism of complete idealic semir- ings. For an element b of B, π−1(b) = sup{x ∈ A π(x) ≤ b} is the inverse image of b. Note that π−1 neither preserves supremums nor multipications. 3.2 Congruence relations Definition 3.2.1. Let R be a complete semiring. A congruence relation a of R is an equivalence relation on R satisfying the following conditions: (1) If (aλ, bλ) ∈ a, then (P aλ, P bλ) ∈ a. (2) If (ai, bi) ∈ a for i = 1, 2, then (a1a2, b1b2) ∈ a. If a is a congruence relation of R, then R/a has a natural structure of a complete semiring, and there is a surjective homomorphism π : R → R/a of complete semirings. We denote by R the set of congruence relations of R. The next proposition is obvious. Proposition 3.2.2. Let R be a complete semiring. Then R parametrizes sur- jective homomorphisms of complete semirings from R, i.e. for any surjective homomorphism f : R → A of complete semirings, there exists a unique congru- ence relation a of A and a natural isomorphism R/a ≃ A making the following diagram commutative: f R R/a A = ≃ 15 / /   = Proposition 3.2.3. The set R has a natural structure of an idealic R-algebra. Proof. Let a and b be two congruence relations of R. We define the multi- plication a · b as the congruence relation generated by {(ab + a′b′, ab′ + a′b)}, where (a, a′) ∈ a and (b, b′) ∈ b. The supremum of congruence relations is the congruence relation generated by those. We can easily verify that this gives the structure of idealic semiring on R: The multiplicative unit is 1 R = R × R, and the additive unit is 0 R = ∆, the diagonal subset of R × R. Finally, the homomorphism R → R of semirings is given by a 7→ h(a, 0)i, where h(a, 0)i is the congruence relation generated by (a, 0). Remark 3.2.4. We must notice that a localization of a idealic semiring (to be appeared in 4.2) is also surjective, hence in the category of idealic semirings we cannot distinguish localizations from quotient rings, using only congruence relations. Definition 3.2.5. Let R be a complete semiring, a be a congruence relation on R, and π : R → R/a be a natural map. For an element a of R, a = π−1π(a) is the a-saturation of a. a is a-saturated if a = a. Lemma 3.2.6. Let R be a complete idealic semiring, and a be a congruence relation on R. (1) For any aλ ∈ R, P aλ = P aλ. (2) For any a, b ∈ R, a · b = ab. Proof. (1) Note that π(π−1(x)) = x for any x ∈ R/a, since π is surjective. P aλ = π−1(π(P aλ)) = π−1(P π(aλ)) = π−1(P π(aλ)) = π−1(π P(aλ)) = P aλ (2) Similar argument. Lemma 3.2.7. Let R be an algebraic complete semiring, and a be a congruence relation of a. Then, the following are equivalent: (1) R/a is algebraic, and the natural map π : R → R/a is algebraic (When this happens, we call a algebraic.) (2) Let a ∈ R be compact, and bλ ∈ R. Then, (a + P bλ, P bλ) ∈ a implies (a + P<∞ bλ, P<∞ bλ) ∈ a. The proof is straightforward. Definition 3.2.8. Let R be a complete semiring. A semiorder ≺ on R is a complete (resp. finite) idealic semiorder if: (1) aλ ≺ bλ for any λ ⇒ P aλ ≺ P bλ (resp. a1 + a2 ≺ b1 + b2). 16 (2) a ≤ b ⇒ a ≺ b. (3) ai ≺ bi (i = 1, 2) ⇒ a1a2 ≺ b1b2. If ≺ is a complete idealic semiorder, then the equivalence relation a defined by (a, b) ∈ a ⇔ a ≺ b, b ≺ a becomes a congruence relation on R. Proposition 3.2.9. Let R be an algebraic complete semiring. Let ≺f be a finite idealic semiorder satisfying: (*) if x is compact and x ≺f b, then there is a compact b′ ≤ b such that x ≺f b′. Let ≺ be the complete idealic semiorder generated by ≺f . Then, the follow- ing are equivalent: (i) a ≺ b (ii) x ≺f b holds for any compact x ≤ a. Furthermore, if a is a congruence relation generated by ≺, then a is algebraic, and a ≤ b in R/a if and only if a ≺ b. Proof. (i) ⇒ (ii) is clear. We will show (ii) ⇒ (i). Let ≪ be the relation defined by a ≪ b if the condition (ii) holds. It suffices to show that ≪ is actually a complete semiorder. Firstly, ≪ is a semiorder: indeed, if a ≪ b and b ≪ c, then x ≺f b holds for any compact x ≤ a. The given condition (*) implies that there is a compact b′ ≤ b such that x ≺ b′. b ≪ c implies that b′ ≺f c, hence the result follows. It is easy to see that ≪ is finite idealic. It remains to show that ≪ is complete. Suppose aλ ≪ bλ. Then, for any compact x ≤ P aλ, <∞ <∞ x ≤ X aλ ≺f X bλ ≺f X bλ, hence P aλ ≪ P bλ. The rest are straightforward. 4 Topological spaces In this section, we will see that the sober spaces are the most appropriate for considering spectrum functors. However, most topics in this section has been already done decades ago in the lattice theory. 4.1 The spectrum functor In the sequel, all semirings are complete. Definition 4.1.1. (1) We denote by (Top), the category of topological spaces. (2) A topological space X is sober, if any (nonempty) irreducible closed subset of X has a unique generic point. We denote by (Sob) the full subcategory of (Top) consisting of sober topological spaces. 17 Proposition 4.1.2. The category (Sob) is a full co-reflective subcategory of (Top), i.e. the underlying functor U : (Sob) → (Top) has a left adjoint. Proof. This proposition is standard. For a given topological space X, Let sob(X) be the set of irreducible closed subsets of X. Closed sets of sob(X) are of forms V (z) = {c ∈ sob(X) c ⊂ z}, where z is a closed subset of X. Given a continuous map f : X → Y between topological spaces, sob(f ) : sob(X) → sob(Y ) is defined by c 7→ f (c), where f (c) is the closure of f (c). This gives a functor sob : (Top) → (Sob). The unit ǫ : Id(Top) ⇒ U ◦ sob is given by x 7→ {x}. The counit η : sob◦U ⇒ Id(Sob) is given by z 7→ ξz, where ξz is the unique generic point of z. Remark 4.1.3. The counit η in the above proof is actually an isomorphism: the inverse is the unit ǫ. Definition 4.1.4. (1) Given a sober space X, the set C(X) of closed sets of X becomes a complete idealic semiring with idempotent multiplica- tion: namely, the supremum of closed sets are their intersections, and the multiplication of two closed sets is the union of the two. (2) Given a continuous map f : Y → X between two sober spaces, we can define C(f ) : C(X) → C(Y ) by Z 7→ f −1(Z). Thus, we obtain a con- travariant functor C : (Sob)op → (IRng†). Definition 4.1.5. Let R be an idealic semiring. (1) An element p of R is prime, if the subset {a ∈ R a (cid:2) p} is a multiplicative monoid. (2) The spectrum Spec R of R is the set of all prime elements of R. For any a ∈ R, define V (a) ⊂ Spec R as the subset consisting of all prime greater than a: V (a) = sup{p ∈ Spec R a ≤ p}. The subset of the form V (a) satisfies the axiom of closed sets, hence gives a topology on Spec R. It is easy to see that Spec R is a sober space. (3) Let f : A → B be a homomorphism of idealic semirings. We define a continuous map Spec f : Spec B → Spec A by y 7→ f −1(y) = sup{x ∈ R f (x) ≤ y}. Hence, we obtain a contravariant functor Spec : (IRng†) → (Sob)op, to which we refer as the spectrum functor. Note that Spec f is well defined. Indeed, let p ∈ B be a prime element. Let a, b ∈ A be two elements satisfying a, b (cid:2) f −1(p). This means that f (a), f (b) (cid:2) p, hence f (ab) (cid:2) p since p is prime. Thus we obtained ab (cid:2) f −1(p). Proposition 4.1.6. The spectrum functor Spec is the left adjoint of C : (Sob)op → (IRng†). 18 Proof. The unit ǫ : Id(IRng†) ⇒ C ◦ Spec is given by A ∋ a 7→ V (a). The counit η : Spec◦C ⇒ Id(Sob)op is given by X ∋ x 7→ {x} ∈ Spec◦C(X). Remark 4.1.7. The above η is actually a homeomorphism: the inverse is given by sending an irreducible closed set Z to its unique generic point. This shows that (Sob)op is a reflective full subcategory of (IRng†). Definition 4.1.8. (1) Let X be a sober space. We call X algebraic if: (a) X is quasi-compact. (b) X is quasi-seperated, i.e. for any two quasi-compact open subsets U, V of X, U ∩ V is quasi-compact (cf. [EGA],Chap I. 1.2). (c) Any open subset of X is a union of some quasi-compact open subsets. (2) A morphism f : X → Y of algebraic sober spaces is a continuous map which is quasi-compact : the inverse image of a quasi-compact open subset is again quasi-compact. (3) We denote by (alg.Sob) the category of algebraic sober spaces. A sober space X is algebraic if and only if C(X) is algebraic. An algebraic sober space is a quasi-compact coherent space, and vice versa. Proposition 4.1.9. Let R be an algebraic complete idealic semiring, and X = Spec R. Denote by D(a) the open subset X \ V (a) for any element a ∈ R. (1) Let U be an open subset of X. Then, U is quasi-compact if and only if there is a compact element a of R such that U = D(a). (2) X is algebraic. (3) If ϕ : B → A is a homomorphism of algebraic complete idealic semirings, then Spec(ϕ) : Spec B → Spec A is algebraic. Proof. (1) Suppose U is quasi-compact. There is an element b ∈ R such that U = D(b). There is a covering b = P bλ of b by compact elements. Then U = ∪D(bλ), and since U is quasi-compact, there is a finite subcover U = ∪<∞D(bλ). This means that U = D(P<∞ bλ), and P<∞ bλ is compact. The converse is easy. (2) First, quasi-compactness of X follows from (1) since the unit 1 is compact. Let U1, U2 be two quasi-compact open subsets of X. Then, there are compact elements ai ∈ R such that Ui = D(ai). Then a1a2 is compact, so U1 ∩ U2 = D(a1a2) is quasi-compact. Thus, X is quasi-seperated. Any open subset of X is a union of quasi-open subsets, since any element of R is algebraic. Thus, X is algebraic. 19 (3) Easy. Corollary 4.1.10. The spectrum functor gives a functor Spec : (alg.IRng†) → (alg.Sob)op when restricted to (alg.IRng†), and is the left adjoint of C : (alg.Sob)op → (alg.IRng†). Proof. To see the adjointness, it suffices to show that the unit ǫ and the counit η in Proposition 4.1.6 are algebraic. The above proposition shows that ǫ is algebraic, and η is algebraic since it is a homeomorphism. We will strengthen this result later. (See Proposition 4.3.2.) 4.2 Localization In this subsection, we gather some basic facts on localization of idealic semirings. Most of the results are similar to those of ordinary rings. Definition 4.2.1. Let R be a complete idealic semiring. (1) A subset Σ of R is a multiplicative system of R if it is a submonoid of R. Σ is compact, if it consists of compact elements. (2) The localization of R along Σ is the homomorphism R → R/a of complete idealic semirings, where a is a congruence relation on R generated by pairs (1, s), where s ∈ Σ. We write Σ−1R instead of R/a. (3) If Σf = {f n}n∈N for an element f ∈ R, we set Rf = Σ−1 (4) If Σp = {x x (cid:2) p} for a prime element p of R, we set Rp = Σ−1 p R, and refer to this semiring the localization of R along p. If R is algebraic, then we set Σp = {x x is compact and x (cid:2) p} instead. f R. (5) Given an element a of R, a = π−1π(a) is the Σ-saturation of a, where π : R → Σ−1R is the natural map. a is Σ-saturated if a = a. (b) For any compact x smaller than f , there exists s ∈ Σ such that sx ≤ g. Lemma 4.2.2. Let R be an algebraic complete idealic semiring, and Σ be a compact multiplicative system. Let a be the congruence relation defined by Σ. Then (f, g) ∈ a if and only if: For any compact y smaller than g, there exists t ∈ Σ such that ty ≤ f . Proof. We make use of Proposition 3.2.9. Let ≺f be a relation defined by a ≺f b if: For any compact x ≤ a, there exists s ∈ Σ such that sx ≤ b. We can easily see that ≺f is a finite idealic semiorder, and a is the congruence relation generated by ≺f . It remains to show that if x ≺f b is compact, then there exists a compact b′ ≤ b such that x ≺f b′. Since sx ≤ b for some s ∈ Σ, and sx is compact, there exists a compact b′ ≤ b satisfying sx ≤ b′. Hence, the result follows. 20 Corollary 4.2.3. Let R be an algebraic complete idealic semiring, and Σ be a compact multiplicative system. Then: (1) Σ−1R is algebraic and the natural map π : R → Σ−1R is also algebraic. (2) More generally, let Σ1, Σ2 be compact multiplicative systems. If there is a homomorphism f : Σ−1 2 R of R-algberas, then f is algebraic. (3) For any a ∈ R, a is the supremum of all compact x ∈ R satisfying sx ≤ a 1 R → Σ−1 for some s ∈ Σ. Proof. These are direct consequences of Lemma 3.2.6 and Lemma 4.2.2. Let us prove (2). Suppose x ∈ Σ−1R is a compact element. Then, we can find a compact a ∈ R such that π1(a) = x, where πi : R → Σ−1 i R is the natural map: R π1 π2 $HHHHHHHHH Σ−1 1 R / Σ−1 2 R f Then, f (x) = π2(a), hence compact. Corollary 4.2.4. Let R be an algebraic complete idealic semiring. (1) Let f be a compact element in R. The localization map R → Rf induces Spec Rf → Spec R, which is an open immersion: Spec Rf ≃ D(f ) = Spec R \ V (f ). (2) The open sets of the form D(f ) with f compact gives an open basis of Spec R. Corollary 4.2.5. Let R be an algebraic complete idealic semiring, and p a prime element of R. Then, Rp is an algebraic complete idealic local semiring, with pRp being the unique maximal ideal. Here, pRp is the image of p via the natural map π : R → Rp. Lemma 4.2.6. Let R be a complete idealic semiring. (1) Any maximal element (of R \ {1}) is prime. (2) Suppose R is algebraic. Then, for any non-unit element a 6= 1, there exists a maximal element larger than a. Proof. (1) Let m be a maximal element of R, and suppose a (cid:2) m and b (cid:2) m. then a + m = b + m = 1 since m is maximal. Hence, m + ab ≥ (m + a)(m + b) = 1 which shows that ab (cid:2) m. Thus, m is prime. (2) Let S be a set of non-unit elements which are larger than a. Then S is a inductively ordered set, since 1 is compact. Hence, there is a maximal element of S by Zorn's lemma. 21   $ / Lemma 4.2.7. Let R be an algebraic complete idealic ring, and a, b ∈ R. Then, the following are equivalent: (i) For any compact x ≤ a, there exists a natural number n such that xn ≤ b. (ii) √a ≤ √b, where √a = sup{x xn ≤ a for some n}. (iii) V (a) ⊃ V (b) on Spec R. Further, if a is compact, these are also equivalent to: (1, b)). (iv) a ≥ b, where a (resp. b) is a congruence relation generated by (1, a) (resp. Proof. (i)⇒ (ii): If x ≤ √a is any compact element, then there exists a natural number m such that xm ≤ a. Then, (i) implies that xmn ≤ b for some n, hence x ≤ √b. This shows that √a ≤ √b. (ii)⇒ (iii): If p ∈ V (b), then p ≥ b. Let x ≤ a be any compact element. Then x ≤ √a ≤ √b implies xm ≤ b for some m. Since xm ≤ p and p is prime, we have x ≤ p. Therefore, a ≤ p and p ∈ V (a). (iii)⇒ (i): Suppose there exists a compact x ≤ a such that xn (cid:2) b for any n. Set A = R/b: this is the quotient ring of R divided by a congruence relation generated by (b, 0). Then xn (cid:2) 0 for all n in A. This shows that Ax 6= 0. Since Ax is algebraic, we can find a prime element p of Ax, by Lemma 4.2.6. Let ϕ : R → Ax be the canonical map. Then, ϕ−1(p) is contained in V (b), but ϕ−1(p) /∈ V (a), since ϕ−1(p) (cid:3) x. This is a contradiction. Suppose a is compact. (i) ⇒ (iv): an ≤ b holds for some n, and since (an, 1) ∈ a, we have (b, 1) ∈ a. (iv) ⇒ (i): (1, b) ∈ b, so (1, b) ∈ a by assumption. This is equivalent to an ≤ b for some n. Corollary 4.2.8. Let R be an algebraic complete idealic semiring, and U be a quasi-compact open subset of X = Spec R. Set Z = X \ U . Let f, g ∈ R be compact elements satisfying V (f ) = V (g) = Z. Then Rf ≃ Rg. Definition 4.2.9. Let R be an algebraic complete idealic semiring. By the above corollary, we may define a localization RZ of R along any compact Z ∈ C(X), by RZ = Rf for any compact f ∈ R such that V (f ) = Z. Also, for a non-compact Z ∈ C(X), set R[Z] = lim←−Z′≤ZRZ′ , where Z ′ runs through all the compact elements smaller than Z, and the limit is taken within the category of algebraic semirings. Remark 4.2.10. Note that R[Z] is not isomorphic to RZ in general: if Z is not compact, RZ may not be even algebraic. The next is the key lemma, which is indispensable when constructing idealic schemes. 22 Lemma 4.2.11. Let R be an algebraic complete idealic semiring, and s, s1,··· , sn be compact elements satisfying s = P si. If there are elements fi ∈ Ri = Rsi satisfying fi = fj in Rij = Rsisj , then there is a unique element f ∈ Rs such that f = fi in Rsi . Proof. First, we will show the uniqueness of f . Let f and g be elements of Rs, such that f = fi = g in Ri. For any compact x ≤ f , there exist natural numbers mi such that smi i x ≤ g in R for any i. Set m = maxi mi. Then, P si = s implies X smi i ≥ X sm i ≥ (X si)nm = snm so that snmx ≤ g. This means that x ≤ g in Rs, so f ≤ g. g ≤ f can be shown in a similar way. Next, we will show the existence. Set f = P fi. To show f = fi in Ri, it suffices to prove fj ≤ fi in Ri. Since fi = fj in Rij, there exists natural numbers mij such that (sisj)mij fj ≤ fi for any i, j. Set m = maxi,j mij. Then X (sisj)mij ≥ X j (sisj)m ≥ sm i (X j sj)mn = sm i smn j i smnfj ≤ fi. Since si ≤ s, this means that sm(n+1) so that sm fj ≤ fi in Ri. Remark 4.2.12. Note that this lemma also holds for pre-complete idealic semirings. fj ≤ fi, hence i 4.3 Comparison with the Stone- Cech compactification Definition 4.3.1. We denote by (alg.IIRng†) the full subcategory of (alg.IRng†) consisting of algebraic complete idealic semirings with idempotent multiplications. Proposition 4.3.2. The spectrum functor and the functor C introduced in category (alg.IIRng†) and 4.1.10 gives an equivalence between the (alg.Sob)op, the (opposite) category of algebraic sober spaces. Proof. We only need to prove that the unit ǫ : Id(alg.IIRng†) ⇒ C ◦ Spec of Corollary 4.1.10 is a natural isomorphism. Let R be an algebraic complete idealic semiring, and a, b ∈ R. It suffices to show that V (a) ≤ V (b) implies a ≤ b. If V (a) ≤ V (b), then Lemma 4.2.7 says that for any compact x, xn ≤ b for some b, but since the multiplication is idempotent, this means that x ≤ b, hence a ≤ b. Making use of Corollary 1.1.8, we have the following: Proposition 4.3.3. The underlying functor U : (alg.Sob) → (Sob) has a left adjoint alg. 23 Proof. The functor alg preserves the idempotency of multiplication, (Sob) is a subcategory of (IIRng†), and (alg.Sob) coincides with (alg.IIRng†), which shows that alg sends a sober space to an algebraic sober space. This is enough for a proof, but let us write it down this functor explicitly, for future references. Given a sober space X, let alg(X) be the set of prime filters on C(X): a prime filter F is a non-empty subset of C(X) satisfying: (1) C1, C2 ∈ F ⇔ C1 ∩ C2 ∈ F . (2) C1, C2 /∈ F ⇒ C1 ∪ C2 /∈ F . A closed set of alg(X) is of a form V (a) = {F a ⊂ F}, where a is a filter on C(X). Given a continuous map f : X → Y of sober spaces, a continuous map alg(f ) : alg(X) → alg(Y ) is given by alg(X) ∋ p 7→ X hzi ∈ alg(Y ), f −1(z)∈p where z runs through all the closed subsets of Y , satisfying f −1(z) ∈ p. Thus, we have a functor alg : (Sob) → (alg.Sob). The unit ǫ : Id(Sob) ⇒ U alg of the adjoint is given by x 7→ h{x}i, and the counit η : alg U ⇒ Id(alg.Sob) is given by p 7→ ∩c∈pc. Proposition 4.3.4. If X ∈ (Top) is a Hausdorff space, then alg(X) is also Hausdorff. Proof. The topological space X is Hausdorff if and only if the diagonal functor ∆ : X → X × X is a closed immersion. It is obvious that the functor alg preserves closed immersion, hence alg(X) → alg(X) × alg(X) is also a closed immersion. Definition 4.3.5. Hausdorff algebraic sober spaces are called Stone spaces. The category of Stone spaces and continuous maps is denoted by (Stone). Remark 4.3.6. Note that a continuous map between Stone spaces are already quasi-compact: an open subset U of a Stone space X is quasi-compact if and only iff U is clopen, since X is Hausdorff. Definition 4.3.7. An II-Ring R is a Boolean algebra, if there is a unary operator ¬ : R → R such that a + ¬a = 1, a · ¬a = 0. Proposition 4.3.8. (1) The category (Bool) of Boolean algebras is equiva- lent to the opposite category of (Stone). (2) The underlying functor (Stone) → (alg.Sob) has a left adjoint and a right adjoint. Proof. (1) This is obvious from the above remark. 24 (2) The existence of the right adjoint follows from the fact that we have a left adjoint of the underlying functor (alg.IIRng†) → (Bool). We will construct the left adjoint. Let X be an algebraic sober space, and A = C(X)cpt be the corresponding pre-complete idealic semiring. Let B be the subring of A consisting of elements with negation, namely elements x ∈ A which satisfies xy = 0 and x + y = 1 for some y ∈ A. Note that y is uniquely determined by x. Therefore, we see that B becomes a Boolean algebra. This correspondence gives a functor G : (alg.IIRng†) → (Bool), and hence Gop : (alg.Sob) → (Stone). For any Boolean algebra C, any morphism C → A of pre-complete idempotent semirings factors through C → B. This shows that Gop is the left adjoint of the underlying functor. Remark 4.3.9. Recall the Stone- Cech compactification: Theorem 4.3.10. Let U : (CptHaus) → (Haus) be the underlying functor from the category (CptHaus) of compact Hausdorff spaces to the category (Haus) of Hausdorff spaces. Then, U has a left adjoint β. The unit morphism X → βX is an open immersion if and only if X is locally compact Hausdorff. Note that the underlying functor (Stone) → (Haus) factors through (CptHaus). This implies that, X → alg(X) factors through β(X) for any Hausdorff space X. Usually, alg(X) does not coincide with β(X). Example 4.3.11. (1) If X is a discrete space, then β(X) coincides with alg(X): indeed, C(X) = Qx∈X F1, and there is a bijection between the set of points of Spec C(X)cpt and set of ultrafilters on X. This also coincides with the set of points of β(X). We can also see that β(X) → alg(X) is in fact, a homeomorphism. (2) The real line R is a locally compact Hausdorff space, hence it becomes an open subset of βR. On the other hand, alg(R) is a one point set, since R is connected, but Stone spaces are totally disconnected. 5 Schemes 5.1 Sheaves When considering sheaves of complete-algebra valued, we need some special care, for there are some obstructions when applying the sheaf theory to algbras admitting infinite operations (in our case, the supremum map.): we don't have sheafifications in general, and the stalk of such sheaves do not admit a natural induced infinite operators. Thus, the notion of algebraicity is essential in the following. Definition 5.1.1. Let X be a topological space. (1) We regard the preordered set C(X) as a category. 25 (2) Let A be a category. A functor F : C(X)op → A is a A -valued presheaf on X. A morphism of presheaves on X is a natural transformation. A presheaf is a sheaf if F is a continuous functor. (3) We denote by (A -PSh/X) (resp. (A -Sh/X)) the category of A -valued presheaves (resp. sheaves) on X. Here, we chose the definition of a sheaf to be a contravariant functor from the category C(X) of closed subsets of X: this is just for convenience sake. Definition 5.1.2. Let σ be a pre-complete algebraic type and X be a topologi- cal space. Then, the underlying functor U : ((alg.σ†-alg)-Sh/X) → ((alg.σ†-alg)-PSh/X) has a left adjoint S, which we call the sheafification. Proof. This is obvious, since (alg.σ†-alg) is equivalent to (σ-alg) by Proposition 1.1.7, and (σ-alg)-valued presheaves admit sheafifications. Next, we will see what happens when the underlying topological space is algebraic. Definition 5.1.3. Let X be an algebraic sober space. Then, C(X)cpt becomes a pre-complete idealic semiring with idempotent multiplication. Let A be any category. (1) We call a functor C(X)op sheaf if it is finite continuous, i.e. preserves fiber products. cpt → A a A -valued presheaf on X cpt. It is a (2) We denote by (A -PSh/X cpt) (resp. (A -Sh/X cpt)) the category of A - valued presheaves (resp. sheaves) on X cpt. Proposition 5.1.4. Let X be an algebraic sober space, and A be a small com- plete category. Then, the underlying functor U cpt : (A -Sh/X) → (A -Sh/X cpt) is an equivalence of categories. Proof. Let F be a A -valued sheaf on X cpt. We define a A -valued sheaf F † on X by where Z ′ runs through all the compact elements smaller than Z. We claim that F † is indeed a sheaf. F †(Z) = lim←−Z′≤Z F (Z ′), For any Z ∈ C(X), any covering Z = P Zλ and aλ ∈ F †(Zλ) satisfying aλZλZµ = aµZλZµ , we will show that there exists a unique a ∈ F †(Z) such that aZλ = aλ. We have aλW CλCµ = aµW CλCµ for any compact W ≤ Z and Cλ ≤ Zλ. Since F is a sheaf and W is covered by finitely many W Cλ's, there is a unique aW ∈ F †(W ) such that aWW Cλ = aλW Cλ. By the definition of F †, the aW 's patch together to give a section a ∈ F †(Z). It is clear that aZλ = aλ. Uniqueness of a is also clear from the construction. Given a morphism f : F → G of sheaves on X cpt, we define f † : F † → G † by F †(Z) = lim←−Z′ F (Z ′) f → lim←−Z′ G (Z ′) = G †(Z). 26 This is well defined. Hence, we have a functor comp : (A -Sh/X cpt) → (A -Sh/X). We will see that this is the left adjoint of U cpt. The unit ǫ : Id(A -Sh/X cpt) ⇒ U cpt comp is given by the natural isomorphism F (Z) ≃ F †(Z) for any compact Z. The counit η : comp U cpt → Id(A -Sh/X) is given by the natural isomorphism (G cpt)†(Z) = lim←−Z′≤Z:cptG (Z ′) ≃ G (Z). 5.2 Idealic schemes Here, we will introduce a notion of idealic schemes. The significant difference between idealic schemes and the usual scheme is that we can construct a universal idealic scheme from its underlying space. Proposition 5.2.1. Let X be a topological space, Z be a closed subset of X, and U = X \ Z be the complement of Z. Then, the restriction homomorphism π : C(X) → C(U ) induces an isomorphism C(X)Z ≃ C(U ), where C(X)Z is the localization along Z. Proof. Since π(Z) = 1 in C(U ), π factors through C(X)Z: C(X) π C(X)Z C(U ) :uuuuuuuuu π It is clear that π is surjective, hence so is π. It remains to show that π is injective. Let a be the congruence relation generated by (1, Z). Then we see that a = {(a, b) ∈ C(X) × C(X) aZ = bZ}. Indeed, the righthand side contains (1, Z), and it is easy to see that this is a congruence relation, since the multiplication is idempotent. If π(a) = π(b), then aZ = a ∪ Z = b ∪ Z = bZ, hence a coincides with b in C(X)Z . Definition 5.2.2. A functor τX : C(X)op → (IRng†) is a sheaf on X defined by Z 7→ C(X)Z ≃ C(X \ Z). Definition 5.2.3. Let α : (IRng†) → (IIRng†) be the left adjoint of the underlying functor (IIRng†) → (IRng†). (1) A triple X = (X, OX , βX ) is an idealic scheme if: (a) X is a sober space. (b) OX is (IRng†)-valued sheaf. (c) βX : αOX → τX is a morphism of (IIRng†)-valued presheaves. 27 / /   : (d) The restriction maps reflect localization: for any W ≤ Z of closed subsets of X, and any a ∈ OX (Z) satisfying βX (α(a)) ≥ W , the restriction functor OX (Z) → OX (W ) factors through the localization OX (Z)a. res OX (W ) 9ssssssssss OX (Z) OX (Z)a (2) A morphism (X, OX , βX ) → (Y, OY , βY ) of idealic schemes is a pair (f, f #) such that: (a) f : X → Y is a continuous map. (b) f # : OY → f∗OX is a morphism of (IRng†)-valued sheaves on Y . (c) The following diagram of (IIRng†)-valued presheaves is commuta- tive: αOY αf # / αf∗OX βY τY f∗βX / f∗τX C(f ) (3) We denote by (ISch) the category of idealic schemes. Proposition 5.2.4. The underlying functor U : (ISch) → (Sob) has a right adjoint. Proof. We will construct a functor C+ : (Sob) → (ISch). Given a sober space X, set OX = τX , and let βX be the identity. Then C+(X) = (X, τX , Id) becomes an idealic scheme. Given a continuous map f : X → Y of sober spaces, set f # = C(f ) : τY → f∗τX . This gives a morphism (f, f #) : C+(X) → C+(Y ) of idealic schemes. Hence, a functor C+ : (Sob) → (ISch) is well defined. We will show that this is the right adjoint of the underlying functor U . The unit ǫ : Id(ISch) ⇒ C+ ◦ U is defined by ǫ(X) = (IdX , βX α). The counit U ◦ C+ → Id(Sob) is given by the identity. Remark 5.2.5. Our next goal is to construct a left adjoint of the global section functor Γ : (ISch)op → (IRng†). However, this seems to be impossible in general, due to what we mentioned in the beginning of the previous subsection. 5.3 A -schemes We will introduce a notion of A -schemes, generalizing the classical schemes in algebraic geometry. This is because several different type of schemes began to appear lately, and we thought there should be a general theory how to construct these schemes. 28 / /   9   /   / Definition 5.3.1. Let σ be an algebraic type with a multiplicative monoid structure. (1) A σ-algebra with a multiplicative system is a pair (R, S), where R is a σ-algebra and S is a multiplicative submonoid of R. (2) A homomorphism (A, S) → (B, T ) of σ-algebras with multiplicative sys- tems is a homomorphism f : A → B of σ-algebras satisfying f (S) ⊂ T . (3) We denote by (σ+Mnd) the category of σ-algebras with multiplicative systems. (4) For an object (R, S) ∈ (σ+Mnd), define L(R) = RS as the localization of R along S. Note that localization exists in (σ-alg) for any algebraic type σ. Given a homomorphism f : (A, S) → (B, T ) of σ-algebras with multiplicative systems, we have a natural homomorphism L(f ) : AS → BT . Hence we obtain a functor L : (σ+Mnd) → (σ-alg), to which we refer as the localization functor. Definition 5.3.2. We will fix a quadruple A = (σ, α1, α2, γ) in the following. Here, (1) σ is an algebraic type, with a multiplicative monoid structure: equiva- lently, there is an underlying functor U1 : (σ-alg) → (Mnd) preserving multiplications. (2) α1 : (σ-alg) → (PIIRng) is a functor, where (PIIRng) is the category of pre-complete idealic semirings with idempotent multiplications. (3) α2 : U1 ⇒ U2α1 is a natural transformation, where U2 : (PIIRng) → (Mnd) is the underlying functor, preserving multiplications. (4) The pair α = (α1, α2) gives a functor α : (σ+Mnd) → (PIIRng+Mnd), namely: α(R, S) = (α1(R), α2(S)), and if f : (A, S) → (B, T ) is a homo- then morphism of (α1f )(α2(S)) ⊂ α2(T ). σ-algebras with multiplicative systems, (5) γ : Lα1 ⇒ αL is a natural isomorphism: (σ+Mnd) L (σ-alg) α / γ =⇒ α1 (PIIRng+Mnd) L / (PIIRng) We will refer to A as a "schematizable algebraic type". Definition 5.3.3. Let A be as above. 29 /     / (1) For any σ-algebra R and any element Z ∈ α1(R), we denote by RZ = Rα−1 2 (Z) the localization of R along Z, where α−1 2 (Z) = {x ∈ R α2(x) ≥ Z}. Note that α−1 2 (Z) is a monoid. (2) A schematizable algebraic type A satisfies the strong patching condition if the following holds: (i) For any σ-algebra R, α2(R) generates α1(R) as a pre-complete idealic semiring. (ii) Let R be any σ-algebra, and s, s1,··· , sn be elements of R satisfying α2(s) = P α2(si) in α1(R). If there are elements ai ∈ Rsi such that ai = aj in Rsisj , then there exists a unique a ∈ Rs such that a = ai in Rsi . (3) A satisfies the weak patching condition, if (i) holds, and (ii) holds for s = 1. Example 5.3.4. Here are some examples of schematizable algebraic types: for any of them, the functor α1 factors through (PIRng) (the category of pre- complete idealic semirings; but in this subsection, we will not assume the exis- tence of infimum operators), so we will just describe α′ 1 : (σ-alg) → (PIRng) for (1) and (2). (1) The algebraic type σ is that of rings, and α′ 1 : (Rng) → (PIRng) sends a ring R to the set of finitely generated ideals on R. Note that α1(R) = C(Spec R)cpt. A homomorphism f : A → B gives a homomor- phism α′ 1(B) defined by 1(f ) : α′ 1(A) → α′ a 7→ f (a)B. The map α2 : R → α1(R) sends an element a ∈ R to V (a), namely the support of a. This gives a natural transformation α2 : U1 ⇒ U2α1, and a functor α = (α1, α2) : (Rng+Mnd) → (PIIRng+Mnd). For any multiplicative system S of a ring R, there is a natural isomorphism γ : Spec RS ≃ α1(RS) ≃ α1(R)α2(S). In this case, the strong patching condition holds. This schematizable algebraic type is what we use in classical algebraic geometry. (2) The algebraic type σ is that of monoids, and α′ 1 sends a monoid M to the set of its finitely generated ideals: an ideal of M is a non-empty set a of M satisfying ax ∈ a for any a ∈ a, x ∈ M . The natural transformation α2 : M → α1(M ) is defined by the unit of the adjoint (Mnd) ⇆ (PIRng), which sends a ∈ M to the saturated ideal a generated by a. 30 There is a natural isomorphism γ : α1(MS) ≃ α1(M )α2(S), since α1(M ) = α1(M/M ×) for any M . The weak patching condition is automatically satisfied, since any monoid is local, i.e. it admits a unique maximal ideal consisting of all non-unit elements. See [Deit]. This schematizable algebraic type is what we use in schemes over F1. Note that, given a ring R, we can construct a scheme of this type by regarding R as a multiplicative monoid. Hence, we must fix the schematizable algebraic type to specify how we make schemes. (3) The algebraic type σ is that of pre-complete idealic semirings, and α1 : (PIRng) → (PIIRng) is the left adjoint of the underlying functor. Note that α1(R) = C(Spec R)cpt. α2 : R → α1(R) is the unit of the adjoint (PIRng) ⇆ (PIIRng), defined by a 7→ V (a). There is a natural isomor- phism γ : C(Spec RS) ≃ C(Spec R)α2(S). The strong patching condition is satisfied, which we already verified in Lemma 4.2.11. This schematizable algebraic type is used in tropical geometry. Proposition 5.3.5. Let R be a pre-complete idealic semiring, and X = Spec R be its spectrum. Define a (PIRng)-valued presheaf OX : C(X)op cpt → (PIRng) on X cpt by Z 7→ RZ . Then OX is a sheaf. This is a direct consequence of Lemma 4.2.11. Definition 5.3.6. Let X be an algebraic sober space. Define a (PIIRng)- valued sheaf τ ′ X on C(X)cpt by Z 7→ C(X)cpt,Z ≃ C(X \ Z)cpt. This is indeed a sheaf, by the above proposition. We will give a definition of A -schemes. Note that we are defining the struc- ture sheaves on X cpt, where X is an algebraic sober space. This is sufficient, by Proposition 5.1.4. Definition 5.3.7. Let A be a schematizable algebraic type. (1) A triple X = (X, OX , βX ) is a A -scheme if: (a) X is an algebraic sober space. (b) OX is a (σ-alg)-valued sheaf on C(X)op cpt. (c) βX : α1OX → τ ′ α1OX is the sheafification of Z 7→ α1(OX (Z)). X is a morphism of (PIIRng)-valued sheaves, where (d) The restriction map reflects localization: Let W ≤ Z be any two ele- ments of C(X)cpt, and a ∈ OX (Z) be a section satisfying βX (α2(a)) ≥ 31 W . Then, the restriction map OX (Z) → OX (W ) factors through OX (Z)a: res OX (W ) 9ssssssssss OX (Z) OX (Z)a (2) A morphism (X, OX , βX ) → (Y, OY , βY ) of A -schemes is a pair (f, f #) such that: (a) f : X → Y is a continuous map. (b) f # : OY → f∗OX is a morphism of (σ-alg)-valued sheaves on Y . (c) The following diagram of (PIIRng)-valued sheaves is commutative: α1OY α1f # / α1f∗OX βY τ ′ Y f∗βX / f∗τ ′ X C(f )cpt (3) We denote by (A -Sch) the category of A -schemes. Definition 5.3.8. We will construct a functor SpecA : (σ-alg) → (A -Sch)op as follows: (1) Given a σ-algebra R, set X = Spec α1(R)†. Note that C(X)cpt = α1(R). Define the structure sheaf OX : α1(R)op → (σ-alg) as the sheafification of a presheaf O ′ We will define a morphism βX : α1OX → τ ′ For any element Z of α1(R), set X of (PIIRng)-valued sheaves: X : Z 7→ RZ . S = S(Z) = {x ∈ R α2(x) ≥ Z}. Then, we have a map α1O ′ X (Z) = α1(RS) γ ≃ α1(R)α2(S) → α1(R)Z = τ ′ X (Z). Sheafifying the lefthand side, we obtain a morphism βX : α1OX → τ ′ X . It is obvious that the restriction map reflects localization. Hence, we obtain a A -scheme SpecA R = (X, OX , βX ). (2) Let ϕ : B → A be a homomorphism of σ-algebras. Set X = SpecA A and Y = SpecA B. We will construct a morphism (f, f #) : X → Y as follows. The quasi-compact continuous map f is Spec(ϕ) : X → Y . Note that C(f )cpt = α1(ϕ). Let us define the morphism f # : OY → f∗OX of sheaves. For any Z ∈ C(Y ), the homomorphism ϕ : B → Aα1(ϕ)(Z) 32 / /   9   /   / sends any element of S(Z) to an invertible element, since α2 is a natural transformation. Hence, it gives rise to O ′ Y (Z) = BZ → Aα1(ϕ)(Z) = O ′ X (f −1Z). Sheafifying the both sides, we obtain f # : OY → f∗OX . It is easy to see that the following diagram commutes: α1OY α1f # α1f∗OX βY τ ′ Y f∗βX / f∗τ ′ X α1(ϕ) Hence, we obtain a functor SpecA : (σ-alg) → (A -Sch), to which we refer to as the spectrum functor. Theorem 5.3.9. Suppose A satisfies the weak patching condition. Then, the spectrum functor SpecA is the left adjoint of the global section functor Γ : (A -Sch)op → (σ-alg). Proof. First, we will define the unit ǫ : Id(σ-alg) ⇒ Γ SpecA . Given a σ-algebra R, set X = SpecA R. Then define ǫ by R ≃ Γ(X, O ′ X ) → Γ(X, OX ). Next, we will define the counit η : SpecA Γ → Id(A -Sch)op . Given a A -scheme X = (X, OX , βX), set Y = SpecA Γ(X). The continuous map η : X → SpecA Γ(X) between the underlying spaces is induced by Γ(βX ) : α1Γ(X) → C(X)cpt. The morphism η# : OY → η∗OX is defined as follows: for a given Z ∈ α1(Γ(X)), the restriction map Γ(X) → OX (η−1Z) = OX (βX (Z)) gives Y (Z) = Γ(X)Z → OX (η−1Z), since the restriction rise to a homomorphism O ′ map reflects localization. Sheafifying the lefthand side, we obtain the required η#. Thanks to the weak patching condition, all four natural transformations ǫΓ, SpecA ǫ, Γη, and η SpecA becomes isomorphisms. Here, we will focus on to the case when A is induced from idealic semirings. This case has a special feature that, we can construct the structure sheaf from the underlying space. Definition 5.3.10. Let A be the schematizable algebraic type induced from pre-complete idealic semirings, introduced in Example 5.3.4(3). (1) We call A -schemes "algebraic idealic schemes". We denote by (alg.ISch) the categories of algebraic idealic schemes. (2) We define a functor C++ : (alg.Sob) → (alg.ISch) as follows: For an algebraic sober space X, set C++(X) = (X, τ ′ X , Id). For a morphism f : X → Y of algebraic sober spaces, set C++(f ) = (f, C(f )cpt) : C++(X) → C++(Y ), where C(f )cpt : τ ′ X is the morphism induced by f . Y → f∗τ ′ 33 / /     / Proposition 5.3.11. The functor C++ is the right adjoint of the underlying functor U : (alg.ISch) → (alg.Sob). Proof. The unit ǫ : Id(alg.ISch) ⇒ C++U is given by (IdX , βX α1) for any alge- braic idealic scheme (X, OX , βX ). The counit η : U C++ ⇒ Id(alg.Sob) is given by the identity. Remark 5.3.12. (1) When the schematizable algebraic type A is the one we introduced in Example 5.3.4(3), then the structure sheaf OX of X = SpecA R is the functor defined by Z 7→ R[Z] for Z ∈ C(X). This follows from Lemma 4.2.11. (2) Note that we don't have a natural underlying functor (alg.ISch) → (ISch): The sheaves τ and τ ′ which represent the topological structure, are different. Let us summarize all the categories and functors we have obtained so far: Let A = (σ, α1, α2, γ) be a schematizable algebraic type. Then the functors are illustrated below: The pairs of arrows are adjoints. The equal signs are equivalences. (alg.σ†-alg) (σ-alg) Ucpt comp′ 8qqqqqqqqqqq xqqqqqqqqqqq comp U alg U (σ†-alg) SpecA Γ / (A -Sch)op (alg.IRng†) U ′ U / (alg.IIRng†) alg U (alg.ISch)op C ++ U (alg.Sob)op alg U Spec Γ Spec C Spec (IIRng†) (ISch)op C U 6nnnnnnnnnnnn vnnnnnnnnnnnn C + / (Sob)op sob U (Top)op As we see, the right half can be regarded as the categories of geometric objects, while the left half are those of algebraic objects. The most mysterious part is the functor U : (A -Sch) → (alg.IIRng†) in the middle, and it has been (and will continue to be) the central subject of algebraic geometry, arithmetic geometry, and tropical geometry. If A is one of the example in 5.3.4, then U factor through (alg.IRng†). 5.4 Comparison with classical schemes Next, we will see the relation between the category of classical schemes and the category of new schemes which we have introduced. In the proceedings, A is the schematizable algebraic type induced from rings, introduced in Example 5.3.4(1). 34 / /   o o     x / o o / 7 7 O O   O O   O O 8 O O / O O v o o   6 O O Proposition 5.4.1. Let (QSch) be the category of quasi-compact, quasi- seperated schemes and quasi-compact morphisms, in the classical sense. Then, there is a fully faithful functor I : (QSch) → (A -Sch) which satisfies I Specop ≃ SpecA ,op. (Rng)op Specop (QSch) (A -Sch) &LLLLLLLLLL SpecA ,op I Proof. Let X = (X, OX ) be a quasi-compact, quasi-seperated scheme in the classical sense. We must construct a morphism βX : α1OX → τ ′ X of sheaves. Since X is locally isomorphic to an affine scheme, we already have a local iso- morphism α1OX ≃ τ ′ X , which patch up to give a global morphism βX . Hence we obtain a A -scheme (X, OX , βX ). For any quasi-compact morphism f : X → Y of quasi-compact, quasi- seperated schemes, f commutes with β, since f is locally induced by a ho- momorphism of rings. This means that f naturally becomes a morphism of A -schemes. Hence, we have a functor I : (QSch) → (A -Sch). We claim that I is fully faithful. Let X, Y be quasi-compact, quasi-seperated schemes, and f : I(X) → I(Y ) be a morphism of A -schemes. We already know that I(X) and I(Y ) are locally ringed spaces, and it suffices to show that f is a morphism of locally ringed spaces, i.e. OY,f (x) → OX,x is a local homomorphism for any x ∈ X. This is a local argument, hence we may assume X and Y are both affine, say X = Spec A and Y = Spec B. Using Theorem 5.3.9, we have Hom(A -Sch)(I(X), I(Y )) ≃ Hom(Rng)(B, A) ≃ Hom(QSch)(X, Y ), which shows that f is induced locally (and hence globally) from a morphism of quasi-compact, quasi-seperated schemes. It is clear from the construction of SpecA that I Specop ≃ SpecA ,op. Remark 5.4.2. Let (M -Sch) be the category of quasi-compact, quasi-seperated schemes over F1, in the sense of Deitmar [Deit]. and A be the schematizable algebraic type of Example 5.3.4(2). Then, by the same arguments as above, we obtain a fully faithful functor (M -Sch) → (A -Sch). References [CC] A. Connes, C. Consani, Schemes over F1 and zeta functions, (arXiv:math.AG/0903.2024) [Deit] A. Deitmar, Schemes over F1. Number fields and function fields -- two parallel worlds, Progr. Math., vol. 239, (2005) [IME] I. Itenberg, G. Mikhalkin, E. Shustin: Tropical algebraic geometry, Ober- wolfach Seminars vol.35 (2007), Birkhauser Verlag 35 / / &   [Lit] G. L. Litvinov: The Maslov dequantization, idempotent and tropical math- ematics: a brief introduction, Zap. Nauchn. Sem. S.-Peterburg. Otdel. Mat. Inst. Steklov. (POMI) 326 (2005) Teor. Predst. Din. Sist. Komb. Algoritm. Metody. B, pp. 145-182 [LMS] G. L. Litvinov, V. P. Maslov, G. B. Shpiz: Tensor products of idempotent semimodules. An algebraic approach, Mat. Zametki 65 (1999) no.4 573-586 [CWM] S. Maclane: Categories for the working mathmatician, Graduate Texts in Mathematics, Vol. 5 (1971). [EGA] A. Grothendieck: El´ements de g´eometrie alg´ebriques IV, Inst.Hautes ´Etudes Sci. Publ. Math. (1964), no.20, 259 [JH] P. Jipsen, H. Rose: Varieties of Lattices, Lecture Notes in Math. 1533 (1992), Springer Verlag [Mik] G. Mikhalkin: Tropical geometry and its applications, Sanz-Sol´e, Marta (ed.) et al., Proceedings of the ICM, Madrid, Spain, August 22-30, (2006). Volume II: Invited lectures. Zrich, European Math Soc., (2006), pp. 827-852 [MMT] R. N. McKenzie, G. F. McNulty, W. F. Taylor: Algebras, lattices, vari- eties, volume I, Wadsworth, California (1987) [PL] J. L. Pena, O. Lorscheid: Mapping F1-land: an overview of geometries over the field with one element, (arXiv: math.AG/0909.0069) [TV] B. Toen, M. Vaqui´e: Au-dessous de Spec Z, J. K-theory 3 (2009), no.3 pp. 437-500 S. Takagi: Department of Mathematics, Faculty of Science, Kyoto University, Kyoto, 606-8502, Japan E-mail address: [email protected] 36 

0908.1831

2

0908

Integral Models of Extremal Rational Elliptic Surfaces

Miranda and Persson classified all extremal rational elliptic surfaces in characteristic zero. We show that each surface in Miranda and Persson's classification has an integral model with good reduction everywhere (except for those of type X_{11}(j), which is an exceptional case), and that every extremal rational elliptic surface over an algebraically closed field of characteristic p > 0 can be obtained by reducing one of these integral models mod p.

math.AG

math

INTEGRAL MODELS OF EXTREMAL RATIONAL ELLIPTIC SURFACES TYLER J. JARVIS, WILLIAM E. LANG, AND JEREMY R. RICKS Dedicated to the memory of Jeremy R. Ricks and Steven Galovich Abstract. Miranda and Persson classified all extremal rational elliptic surfaces in characteristic zero. We show that each surface in Miranda and Persson's classifi- cation has an integral model with good reduction everywhere (except for those of type X11( j), which is an exceptional case), and that every extremal rational elliptic surface over an algebraically closed field of characteristic p > 0 can be obtained by reducing one of these integral models mod p. Introduction An extremal rational elliptic surface is a smooth projective rational elliptic sur- face X over an algebraically closed field k together with a morphism f : X → P1 with a section, such that the Mordell-Weil group of the geometric general fibre of f is finite. Such surfaces were classified by R. Miranda and U. Persson over the complex numbers [MP] and by W. E. Lang over algebraically closed fields of characteristic p > 0 in [L1] and [L2]. A summary of these classifications is provided in the appendix. During his work on [L2], the senior author became convinced that the extremal rational elliptic surfaces in characteristic p were closely linked to their characteristic zero counterparts. For instance, there is no surface X3333 with four singular fibres of type I3 in characteristic three, but one might guess that the surface of type III in characteristic three comes from reducing X3333 mod 3 in such a way that three of the singular fibres come together. There are several such situ- ations throughout the classification, but at the time, the senior author was unable to formulate a precise theorem covering all cases. Later, the senior author began to wonder if the characteristic zero extremal rational elliptic surfaces had integral models with good reduction everywhere. Since a rational elliptic surface is the projective plane with nine points (possibly infinitely near) blown up, this author guessed that such models might exist, and realized that if enough such models could be produced, they would provide the link between characteristic zero and characteristic p discussed in the last paragraph. This paper completes the program outlined above. We show that each surface found by Miranda and Persson has an integral model with good reduction every- where (except for those of type X11(j), which is an exceptional case), and that every extremal rational elliptic surface over an algebraically closed field of characteristic p > 0 can be obtained by reducing one of these integral models mod p. An interesting feature of this theory is that the integral models of the Miranda- Persson surfaces are not unique, and in fact, we need to use more than one integral model of some of the Miranda-Persson surfaces to get a full set of extremal rational Date: June 10, 2018. 1 2 JARVIS, LANG, AND RICKS elliptic surfaces in characteristic p. Note that while the Miranda-Persson models are defined over Z, when they are normalized using the conventions in [Ta], the discriminants tend to be divisible by 2 and 3. When this happens the reduction is no longer an elliptic surface. Much of the work in this paper deals with getting these powers of 2 and 3 out of the discriminant. Also, note that the Miranda-Persson classification works in all characteristics except 2, 3, and 5, so that this paper is mostly about the difficulties involving these small primes. We should emphasize that the models produced here are integral Weierstrass models. We know by Artin's theory of simultaneous resolution of rational double points [A] that these models can be transformed into integral smooth models, but this requires extension of the base ring, and we do not know how to control how big an extension is needed. We have confined our investigation to showing the existence of integral models and finding enough integral models to get all the surfaces in characteristic p found in [L1] and [L2]. We have not tried to study extremal rational elliptic surfaces over non-algebraically closed fields, so that our lifting results are not as strong as those available for elliptic curves. It would be interesting to try to relate the results found here to the universal elliptic curves studied by Katz and Mazur in [KM]. We have not attempted to do this here. It would also be interesting to look at the implications of the existence of different integral models for the same surface for moduli of rational elliptic surfaces, but we have no ideas about this at this time. Here is a plan of the paper. In Section 0, we review some results on elliptic curves and surfaces that we will need in the rest of the paper, and set terminology. In Section 1, we discuss integral models of extremal rational elliptic surfaces over the ring of integers Z. We find that three types of surfaces do not have integral models over Z, and we find integral models for these over rings of integers in number fields in Section 2. In Section 3, we look at the reduction mod p of the models found in Sections 1 and 2, and show that these reductions give us a full set of extremal rational elliptic surfaces over an algebraically closed field of characteristic p. 0. Preliminaries and terminology An elliptic surface over a field k is a smooth projective surface X over k together with a morphism f : X → C, where C us a smooth projective curve over k, and such that all but finitely many fibres are elliptic curves. We will always assume f has a fixed section (which we call zero) and that there are no exceptional curves of the first kind in any fibre of f . In this paper, we are studying rational elliptic surfaces, which implies C = P1. Our surfaces are also extremal, which means that the Mordell-Weil group of the generic fibre is finite. A complete classification of these surfaces over the complex field C has been given by Miranda and Persson [MP], and we will follow the notations of that paper for the most part. As is well known, an elliptic surface over P1 has a Weierstrass model, which is obtained from the smooth model by contracting those components of each singular fibre which do not meet the zero section. The resulting contracted surface has at most rational double points as singularities. The Weierstrass model is birationally isomorphic to the surface defined by the Weierstrass equation y2 + a1xy + a3y = x3 + a2x2 + a4x + a6, INTEGRAL MODELS 3 where the ai are polynomials in t, an affine coordinate on the base P1. There is a global minimal Weierstrass equation for our surface (see [Sil]1 , Ch. VII, Sect. 1, and Ch. VIII, Sect. 8, for discussion of the minimal Weierstrass equation), and the requirement that our surface is rational forces degree ai ≤ i for this minimal equation. We are interested in finding integral models of extremal rational elliptic surfaces. This means we want diagrams X f ❄ P1 ❄ g .......................... ......................... ....................... ...................... .................... .................... ..................... ...................... ...................... ....................... . ❘ Spec(R), where f and g are flat and proper, R is an integral domain (almost always the ring of integers of an algebraic number field), and such that the geometric general fibre of g is the Weierstrass model of an extremal rational elliptic surface. This setup will be called an integral model in the weak sense. An integral model in the strong sense will satisfy the extra condition that it will have good reduction at all primes of R, meaning that all fibres of g will be Weierstrass models of elliptic surfaces, and that all the Weierstrass equations will be minimal. If we use the term integral model without qualification, we will mean integral model in the strong sense. Our strategy will be to start with the Miranda-Persson models for each type of extremal rational elliptic surface, which are integral models in the weak sense, and modify them so they become integral in the strong sense. Therefore we will review briefly how the Weierstrass equations transform under various substitutions. Here we follow Tate's "formulaire" [Ta]. If we have two Weierstrass equations over R, y2 + a1xy + a3y = x3 + a2x2 + a4x + a6 and y′2 + a′1x′y′ + a′3y′ = x′3 + a′2x′2 + a′4x′ + a′6 and an isomorphism of the underlying elliptic curves preserving the origin, then this isomorphism is obtained by a substitution of the form x = u2x′ + r y = u3 y′ + su2x′ + q, where u is a unit of R, and r, s, and q are elements of R. (We use q instead of Tate's t since we are using t as a coordinate on P1.) The coefficients of the equations are 4 JARVIS, LANG, AND RICKS related by ua′1 u2a′2 u3a′3 u4a′4 u6a′6 = a1 + 2s = a2 − sa1 + 3r − s2 = a3 + ra1 + 2q = a4 − sa3 + 2ra2 − (q + rs)a1 + 3r2 − 2st = a6 + ra4 + r2a2 + r3 − qa3 − q2 − rqa1. In practice, we will restrict ourselves to substitutions of two types: a.) Those where u = 1, which will be called rsq substitutions, and b.) Those where r = s = q = 0. These will be called u-substitutions. We will be using the auxiliary quantities + 4a2 b2 = a2 1 b4 = a1a3 + 2a4 b6 = a2 3 b8 = a2 c4 = b2 c6 = −b3 ∆ = b2 j = c3 4 2b8 − 8b3 /∆. 3 − a2 4 + 4a6 1a6 − a1a3a4 + 4a2a6 + a2a2 2 − 24b4 2 + 36b2b4 − 216b6 4 − 27b2 6 + 9b2b4b6 We know that our Weierstrass equation defines an elliptic curve if and only if ∆ , 0. 1. Integral models over Z §1.A. X11(j). This surface is different from all the others in several ways, so we deal with it first. Here we have a family of surfaces, one for each value of j. The Miranda- Persson model has Weierstrass equation y2 = x3 + rt2x + st3, where r and s are chosen so that j = 6912r3/(4r3 + 27s2). (Note that we use t as an affine coordinate on P1, while Miranda-Persson use u and v as projective coordinates. Also our j differs from theirs by a factor of 1728 and our ∆ differs from theirs by a factor of 16). We observe also that this surface is a quadratic twist of the constant elliptic curve defined by y2 = x3 + rx + s. Since there is no constant elliptic curve over Z with good reduction everywhere, we do not try to find an integral model in the strong sense for X11(j). Instead, we construct our model as a quadratic twist of a constant elliptic curve in such a way that our model has good reduction everywhere that the original curve did. We begin with an elliptic curve E over a ring R where 2 is not a zero-divisor. We let E have Weierstrass equation y2 + g1xy + g3y = x3 + g2x2 + g4x + g6. Now we perform an rsq-substitution with r = 0, s = −(1/2)g1, and q = −(1/2)g3. 1)x2 + 3. Now we perform a quadratic twist over R[1/2, (t2 + This produces an elliptic curve over R[1/2] with equation y2 = x3 +(g2 +(1/4)g2 (g4 + (1/2)g1g3)x + g6 + (1/4)g2 INTEGRAL MODELS 5 4t)1/2] and obtain an elliptic surface with equation y2 = x3 + (g2 + (1/4)g2 (g4 + (1/2)g1g3)(t2 + 4t)2x + (g6 + (1/4)g2 3)(t2 + 4t)3. 1)(t2 + 4t)x2 + Finally, we perform an rsq-substitution with r = 0, s = (1/2)g1t, and q = (1/2)g3t3 to get the surface with equation y2 + g1txy + g3t3y = x3 + (g2t2 + 4g2t + g2 1)x2 + (g4t4 + 8g4t3 + 16g4t2 + 4g1g3t3 + 8g1g3t2)x + (g6t6 + 12g6t5 + 48g6t4 + 64g6t3 + 3g3 3t5 + 12g2 3t4 + 16g2 3t3). This is the desired model. Its discriminant is ∆E(t2 + 4t)6, where ∆E is the discriminant of the original elliptic curve E, and its j-invariant is the same as the j-invariant of E. Our surface is an integral model in the weak sense, and it has good reduction at all points of Spec(R) where the original constant curve has good reduction. Note that we could not twist over R[1/2, √t] as in [MP], for then we would have obtained a model where the singular fibres are at 0 and ∞. This model would have reduced mod 2 to a surface with two singular fibres of a sort which is not permitted by [L2]. §1.B. X22, X33, and X44. We will show that these surfaces do not have integral models over Z. Note that X22 and X44 have constant j-invariant 0, while X33 has constant j-invariant 123. The non-existence of integral models over Z for these surfaces will follow easily from the following lemma, which is well known to the experts (at least in the number field case). Lemma 1.1. Suppose E is an elliptic curve over a discrete valuation ring R, and suppose Fract(R) has characteristic 0. a) Suppose 3 is a uniformizer of R, and suppose j(E) = 0. Then E has bad reduction. b) Suppose 2 is a uniformizer of R, and suppose j(E) = 123. Then E has bad reduction. 4 − c2 4 − c2 Proof. a.) Suppose E has good reduction. Since j(E) = 0, and j = c3 4 6 and v3(1728∆) = v(1728) = 3, while v3(c2 c4 = 0. Since 1728∆ = c3 have a contradiction. /∆, we get 6) is even, we 4 6 and j = c3 b.) Since 1728∆ = c3 /∆, the fact that j = 1728 implies c6 = 0. If we have good reduction at 2, then 4 exactly divides c4. Since c4 = b2 2 − 24b4, this means 2 exactly divides b2. But b2 = a2 + 4a2, which gives a contradiction. Now an 1 integral model over Z can be considered as an elliptic curve over Z[t] with good reduction at all prime ideals generated by rational primes. Applying the lemma to the cases where R equals Z[t] localized at (2) or (3), we see integral models of the (cid:3) three surfaces given cannot exist. We will see in Section 2 that these surfaces have integral models over rings of integers in number fields. §1.C. X3333. While it would not be useful to show all the details of the way we find the integral model, a sketch of the method may be useful. Some of the other integral models were found using a similar technique. 6 JARVIS, LANG, AND RICKS The Miranda-Persson model of X3333 has Weierstrass equation y2 = x3 + (−3t4 + 24t)x + (2t6 + 40t3 − 16). The discriminant of this is ∆ = 21233(t3 + 1)3. This is an integral model in the weak sense, and to produce an integral model in the strong sense, we must eliminate the factors of 2 and 3 which appear in ∆. For this, we use Tate's algorithm at the prime ideals (2) and (3) of Z[t]. Roughly speaking, the algorithm begins by making a series of rsq-substitutions. When these are done, either the algorithm stops, or a u-substitution is made, which has the effect of dividing the discriminant by u12. The above steps are then repeated. In order to get rid of the unwanted factors, it is natural to start by making the discriminant divisible by 312. This is done by replacing t by 9t + 8. After making this change, we ran Tate's algorithm separately at the primes (2) and (3). At the prime (3), the algorithm went through to the end, so that the factor 312 could be eliminated. At the prime (2), the algorithm terminated before the end. Thus, the minimal version of our modified Miranda-Persson model had good reduction at 3, but bad reduction at 2. To deal with this, we extended scalars to the Gaussian integers, and ran the algorithm again. We obtained a new surface whose Weierstrass equation had integer coefficients, had good reduction at 2, and was isomorphic to the Miranda-Persson model over Q(i). Finally, we used the Chinese remainder theorem to get substitutions that achieved the desired goals simultaneously at 2 and 3. The final integral model is y2 + a1xy + a3y = x3 + a2x2 + a4x + a6, with the ai defined below. a1 = 171t a2 = 16 − 7353t a3 = −3 a4 = 594t4 − 54t3 − 528t2 + 214t + 76 a6 = −2700t6 + 648t5 + 3924t4 − 3t3 − 1682t2 − 304t + 88 The discriminant is ∆ = −(t + 1)3(27t2 + 45t + 19)3. §1.D. X222. This is the only model found by a computer search, which was carried out with the assistance of Jason Grout. The singular fibres of this surface are of types I∗2, I2, and I2. So we begin by writing down a Weierstrass equation for a surface with a singular fibre of type I∗2 at t = 0. This equation has coefficients a1 = tc0 a2 = t(a + bt) a3 = t3d0 a4 = t3(ct + d) a6 = t5(et + f ). We then let the computer vary c0, d0, a, b, c, d, e, and f , until it found a surface whose discriminant was t8 times a perfect square. Again, we are grateful to Jason Grout for help with the computer programming and with selecting the search region. An integral model yielded by this search has Weierstrass coefficients INTEGRAL MODELS 7 a1 = t a2 = t + t2 a3 = 4t3 a4 = 2t4 a6 = 4t5 + t6. The discriminant is ∆ = −t8(16 + 40t + 89t2)2. §1.E. X321. 321A, which has Weierstrass coefficients We found two integral models for this surface. The first integral model we call a1 = 1 a2 = −1 a3 = 6t + 4 a4 = −4t − 2 a6 = −9t2 − 12t − 4 ∆ = t2(64t + 9). A second integral model of this surface will be needed in Section 3. This integral model we call 321B, with Weierstrass coefficients a1 = t a2 = 0 a3 = 0 a4 = t3 a6 = 0 ∆ = t9(t − 64). (This is the simplest possible surface with a fibre of type III* at t = 0.) Notice that these surfaces are not isomorphic over Z. The reduction of model 321A mod 3 has two singular fibres, while the reduction of model 321B mod 3 has three singular fibres. This example shows that integral models of extremal rational elliptic surfaces are not unique. In the remaining cases we simply give the Weierstrass coefficients of the integral models. In some of the cases, we will need two models of the same surface. §1.F. X9111. a1 = t a2 = 0 a3 = −1 a4 = 0 a6 = 0 ∆ = −(t + 3)(t2 − 3t + 9). 8 JARVIS, LANG, AND RICKS §1.G. X5511. §1.H. X8211. §1.I. X6321. §1.J. X4422. a1 = 5t + 1 a2 = −6t2 − 4t − 3 a3 = 1 a4 = 2 a6 = −t − 1 ∆ = t5(t2 − 11t − 1). Model 8211A a1 = 1 a2 = 32t2 a3 = 0 a4 = 256t4 a6 = −t6 − 64t4 ∆ = t2(1 + 16t2). Model 8211B a1 = t a2 = 128 a3 = 0 a4 = 21t2 + 5461 a6 = 441t2 + 77568 ∆ = t2(t2 + 16). Model 6321A a1 = 1 a2 = 4t2 + 2t a3 = t a4 = 2t6 a6 = 0 ∆ = t3(t + 1)(−1 + 8t)2. Model 6321B a1 = t a2 = 1 + t a3 = 2t2 + t a4 = t − t3 a6 = −t3 − t4 ∆ = (t + 8)(t − 1)2t3. a1 = 1 a2 = 4t2 a3 = 4t2 a4 = −t2 a6 = −4t4 ∆ = t4(4t − 1)2(4t + 1)2. §1.K. X211. §1.L. X431. §1.M. X411. §1.N. X141. INTEGRAL MODELS 9 a1 = 1 a2 = 0 a3 = 0 a4 = −72t(432t + 1)(373248t2 + 864t + 1) a6 = 557256278016t5 + 2257403904t4 + 2985984t3 + 864t2 − t ∆ = t(432t + 1)(864t + 1)10. a1 = 1 a2 = −27t a3 = 0 a4 = 243t2 a6 = −729t3 − 27t2 + t ∆ = −t(27t + 1)3. a1 = 1 a2 = 32t + 3 a3 = 3 a4 = 256t2 + 64t + 2 a6 = 192t2 + 31t − 1 ∆ = t(16t + 1). a1 = 1 a2 = −256t2 + 24t a3 = 10t2 + 20t a4 = −117t2 a6 = −25t4 − 100t3 − 112t2 + t ∆ = t(16t − 1)7. 2. Integral models over larger rings In this section, we find integral models over rings of integers of algebraic number fields of the three surfaces which do not have integral models over Z. §2.A. X33. This surface has j-invariant 123 and two singular fibres of types III and III*. Our approach to finding an integral model was to start with a constant elliptic curve over Z[i] with j-invariant 123 and with good reduction at (1 + i), the unique prime ideal of Z[i] lying over 2. We then replaced this with an elliptic surface with singular fibres of the desired types which is (1 + i)-adically close to our constant It was necessary choose the new surface to be sufficiently close (1 + i)- curve. adically to the constant curve to have good reduction at (1 + i), but not so close as to have constant reduction (that is, reduction not depending on t) at (1 + i). Our starting curve was defined by y2 = x3 + (−1 + 2i)x. (This is an interesting curve over Z[i], since its minimal model only has bad reduction at one Gaussian prime, namely (−1 + 2i). We replaced this by the surface with Weierstrass equation 10 JARVIS, LANG, AND RICKS y2 = x3 + (8t− 1 + 2i)x. After running Tate's algorithm at (1 + i), we found an integral model in the strong sense over Z[i] of X33. Its Weierstrass coefficients are a1 = 1 − i a2 = −i a3 = −i a4 = −2t a6 = it ∆ = (8t − 1 + 2i)3. We found integral models in the remaining cases over Z[31/4]. According to the computer package pari, this is the full ring of integers in Q(31/4), and this ring has class number one. We omit the method of construction of the models, which was similar to the previous case. §2.B. X22. §2.C. X44. a1 = 0 a2 = −16 √3 a3 = 27 · 31/4 a4 = 256 a6 = t − 637 ∆ = −169 − 312t √3 − 432t2. a1 = 0 a2 = −16 √3 a3 = 27 · 31/4 a4 = 256 a6 = −376 √3 + 97t + 3 √3t2 ∆ = −(1/9)(18t + 97 √3)4. Note that the constant term of ∆ is not divisible by 31/4 in Z[31/4]. 3. Reductions In this section, we find the reductions modulo primes of the surfaces found in the first two sections. The list of characteristic p surfaces produced will be sufficient to show that all extremal rational elliptic surfaces in characteristic p come from characteristic zero. Proofs will not be given, since checking that the reduced surfaces are of the stated types is an easy exercise for the reader. However, we will prove the following lemma, which may speed the task of verifying the results. INTEGRAL MODELS 11 Lemma 3.1. Suppose X → P1 R → Spec R is a minimal Weierstrass model of an extremal rational elliptic surface, where R is the ring of integers of an algebraic number field. Suppose p is a prime of R and suppose X has good reduction at p, which means that the geometric general fibre Xp → R/p ≡ k is the minimal Weierstrass model of an elliptic surface. Then Xp is also an extremal rational elliptic surface. Proof. By enlarging R to a bigger ring S and using Artin's theory [A] of simulta- neous resolution of rational double points, we get a smooth model Y of X such that Y → Spec S is smooth. By enlarging S further if necessary, we may assume that all components of singular fibres are rational over S. Since Y is extremal, the sublattice of NS(Y) (the Neron-Severi group) generated by those components of singular fibres not meeting the zero solution is of rank 8. Therefore the sublattice of NS(Yp) generated by components of fibres not meeting the zero solution also (cid:3) has rank 8 and we see Yp (and hence also Xp) is extremal. A slight extension of the above argument gives K, where K is an algebraically closed field.) Lemma 3.2. The order of the Mordell-Weil group of Xp divides the order of the Mordell- Weil group. (Here we use Mordell-Weil group to mean the order of the Mordell-Weil of the general fibre of X → P1 Lemma 3.3. Let p be a prime, p , 2, 3. Then the reduction mod p of each model of an extremal rational elliptic surface found above is an extremal rational elliptic surface of the same type, with one exception. The reduction at p = 5 of the given model of X5511 is the unique extremal rational elliptic surface in characteristic 5 with singular fibres of types II, I5, and I5. The existence of this characteristic 5 surface was pointed out by Chad Schoen, who should have been thanked in [L2]. We will call this surface the Schoen Surface. Now we give a table of the reduction behavior of the surfaces found in the previous sections at characteristics 2 and 3. For the surfaces not defined over Z, reduction mod 2 or 3 means reduction at primes lying over these rational primes. The numbers labeling the mod 2 and mod 3 reductions come from [L2]. The explanation of the reduction behavior of surfaces of type X11(j) comes after the table. Table 1: Reduction of Integral Models mod 2 and 3 Integral Reduction Reduction Model mod 2 X11(j) I II II II VII X3333 I V V X9111 X33 X22 X44 X3333 X222 X321A X321B X9111 mod 3 VI VI bis V I I III IX V XI II continuedonnextpage 12 JARVIS, LANG, AND RICKS continued frompreviouspage Integral Reduction Reduction Model mod 2 X5511 X8211A X8211B X6321A X6321B X4422 X211 X431 X411 X141 mod 3 X5511 X8211 X8211 VII VII X4422 IV IV X VIII X5511 V III IX VIII IV VI IX VI VI Theorem 3.4. All extremal rational elliptic surfaces over algebraically closed fields of characteristic p lift to characteristic zero. Proof. The only cases which are not obvious from the above table are those related to X11(j). Suppose we have a surface of type X11(j) in characteristic p, p , 2, 3. Choose an elliptic curve E over some ring R with good reduction at some prime ideal p containing p, and such that j(E) ≡ j (mod p). Then perform a quadratic twist as in Section §1.A. We get an elliptic surface over R with good reduction at p of type X11(j). Further explanation of the reduction behavior of X11(j) at 2 and 3 is needed. For characteristic 2, if we produce a model of X11(j) as in Section §1.A, where E/R is chosen to have good reduction at a prime p containing 2, the reduction of our model at p will have discriminant ∆Et12. Thus, it will be of Type I or Type II. The Weierstrass equation of a surface of Type I is y2 + txy + a3y = x3 + tx2 + kt6, k , 0, and the Weierstrass equation of Type II is y2 + t3y = x3 + t5. The j-invariant of the general fibre of type I is 1/k. So if we choose such that j(E) ≡ 1/k, then the reduction of our model of X11(j) will be isomorphic (over the algebraic closure of R/p) to the surface of type I with k as specified. The j-invariant of the general fibre of the surface of type II is 0. So if we choose E so that j(E) ≡ 0 (mod p), we get a surface of type II. So all surfaces of type I or type II are liftable. Incidentally the above discussion shows that type II in characteristic 2 is a specialization of type I, which may not have been obvious from [L2]. The situation in characteristic 3 is similar. If we have a model of X11(j) as in Section §1.A (we might as well choose g1 = g3 = 0 for simplicity), its reduction at a prime containing 3 will be of type VI or VI bis. The equations for surfaces of types VI and VI bis are y2 = x3 + tx2 + kt3 (k , 0) and y2 = x3 + t2x respectively. The j-invariants are −1/k and 0 respectively. So if we choose E/R and the prime p of R containing 3 such that j(E) ≡ −1/k (mod p), then the reduction of our model of X11(j) will be isomorphic (over the algebraic closure of R/p) to a surface of type VI with k as specified. If our setup is such that j(E) ≡ 0 (mod p), the reduction will be of type VI bis. So all surfaces of type VI or type VI bis are liftable. The above (cid:3) discussion shows that type VI bis is a specialization of type VI. In this appendix we briefly summarize the results of [MP] and [L2]. A. Appendix INTEGRAL MODELS 13 A.1. Characteristic 0 and Charachteristic≥ 5. In [MP] Miranda and Persson prove that over C every extremal rational elliptic surface is one of the following, and for each of the following configurations of singular fibers except I∗0 I∗0, and II I5 I5 there is a unique surface with that configuration. For the configuration I∗0 I∗0 there is precisely one surface X11(j) for each j ∈ C, and the configuration II I5 I5 does not occur in characteristic 0. Table A.1: Extremal Rational Elliptic Surfaces in Char ≥ 5 Singular Char 0 and Fibers Char > 5 II II∗ III III∗ IV IV∗ X X X Surface Name X22 X33 X44 X11(j) j ∈ C I∗0 I∗0 X211 X321 X431 X411 X141 X222 X9111 X8211 X5511 Schoen X6321 X4422 X3333 II∗ I1 I1 III∗ I2 I1 IV∗ I3 I1 I∗4 I1 I1 I∗1 I4 I1 I∗2 I2 I2 I9 I1 I1 I1 I8 I2 I1 I1 I5 I5 I1 I1 II I5 I5 I6 I3 I2 I1 I4 I4 I2 I2 I3 I3 I3 I3 Char 5 X X X X X X X X X X X X dne X X X X X X X X X X X X X X dne X X X In [L1, L2] it is shown that in characteristics not equal to 2, 3, or 5 the results of [MP] also hold. In characteristic 5 the results of [MP] hold except that the surface X5511 is replaced by a (unique) surface with singular fibers II I5 I5, which we call the Schoen surface. A.2. Chararacteristics 2 and 3. Characteristics 2 and 3 are quite different from the others. A.2.1. Characteristic 3. In [L1, L2] it is shown that in characteristic 3 every ex- tremal rational elliptic surface is one of the following, and for each of the following configurations of singular fibers except I∗0 I∗0, there is a unique surface with that configuration. For the configuration I∗0 I∗0, there is a family of surfaces VI(k) param- eterized by non-zero values of k and one additional surface VIbis. 14 JARVIS, LANG, AND RICKS Table A.2: Extremal Rational Elliptic Surfaces in Char 3 Singular Fibers II∗ II I9 IV∗ I3 II∗ I1 III∗ III Surface Name I II III IV V VI(k) k , 0 I∗0 I∗0 I∗0 I∗0 VI bis III I3 I6 VII I∗1 I1 I4 VIII I∗2 I2 I2 IX I∗4 I1 I1 X III∗ I1 I2 XI I8 I2 I1 I1 X8211 X5511 I5 I5 I1 I1 I4 I4 I2 I2 X4422 A.2.2. Characteristic 2. Finally, in [L1, L2] it is shown that in characteristic 2 ev- ery extremal rational elliptic surface is one of the following, and for each of the following configurations of singular fibers except I∗4, there is a unique surface with that configuration. For the configuration I∗4, there is a family of surfaces I(k) parameterized by non-zero values of k. Remark A.1. Note that the the surfaces in characteristic 2 do not necessarily corre- spond to surfaces with the same name in characteristic 3. Table A.3: Extremal Rational Elliptic Surfaces in Char 2 Surface Name I(k) k , 0 II III IV V VI VII VIII IX X9111 X5511 X3333 Singular Fibers I∗4 II∗ III I8 I∗1 I4 III∗ I2 II∗ I1 IV IV∗ IV I2 I6 IV∗ I1 I3 I9 I1 I1 I1 I5 I5 I1 I1 I3 I3 I3 I3 INTEGRAL MODELS 15 Acknowledgments This paper began as a research project by Jeremy Ricks, a BYU undergraduate student, under the direction of Tyler Jarvis. Unfortunately, Jeremy died in a tragic accident in September of 2001, before his work on the project was completed. Both surviving authors were greatly impressed by Jeremy's mathematical ability and his dedication to this project. He produced many of the models given here in a remarkably short amount of time. We are extremely grateful to Jeremy's wife, Melinda, for providing us with Jeremy's notes. Without them, this paper would not exist. We also wish to remember Professor Steven Galovich, who passed away in 2006. Steve was the senior author's mentor at Carleton College (long before the concept of undergraduate mentoring became fashionable) and provided him with an outstanding introduction to the world of algebra and number theory. The senior author would also like to thank C. Schoen, D. Doud, J. Grout, and K. Rubin for useful conversations on elliptic curves and surfaces, and Brigham Young University for computer support. Finally, we would like to thank M. Sch utt and the referee for their comments on a previous version of this paper. References [A] Artin, M. Algebraic construction of Brieskorn's resolutions. J. Algebra 29 (1974), 330 -- 348. [KM] Katz, N.; Mazur, B. Arithmetic moduli of elliptic curves. Annals of Mathematics Studies, 108. Princeton University Press, Princeton, NJ, 1985. [L1] Lang, W. E. Extremal rational elliptic surfaces in characteristic p. I. Beauville surfaces. Math. Z. 207 (1991), 429-437. [L2] Lang, W. E. Extremal rational elliptic surfaces in characteristic p. II. Surfaces with three or fewer singular fibres. Ark. Mat. 32 (1994), 423 -- 448. [MP] Miranda, R.; Persson, U. On extremal rational elliptic surfaces. Math. Z. 193 (1986), 537-558. [Sil] Silverman, Joseph H. The Arithmetic of Elliptic Curves, Graduate Texts in Math. 106, Berlin- Heidelberg-New York, Springer-Verlag 1975. [Ta] Tate, J. Algorithm for determining the type of a singular fibre in an elliptic pencil, in Modular functions of one variable IV (Birch, B. J., Kuyk, W., eds.), Graduate Texts in Math. 476, Berlin-Heidelberg-New York, Springer-Verlag 1975. 

1209.4404

1

1209

Regularity of curves in abelian varieties

Inspired by a theorem of Gruson-Lazarsfeld-Peskine bounding the Castelnuovo-Mumford regularity of curves in projective spaces, we bound the Theta-regularity of curves in polarized abelian varieties.

math.AG

math

REGULARITY OF CURVES IN ABELIAN VARIETIES LUIGI LOMBARDI AND WENBO NIU Abstract. Inspired by a theorem of Gruson-Lazarsfeld-Peskine bounding the Castelnuovo- Mumford regularity of curves in projective spaces, we bound the Theta-regularity of curves in polarized abelian varieties. 1. Introduction A coherent sheaf F on a projective space Pn is Castelnuovo-Mumford k-regular if for all i > 0 the spaces H i(Pn, F(k − i)) = 0. One can read off some geometric properties of a subvariety Y ⊂ Pn just by looking at the regularity of its ideal sheaf IY . For instance, if IY is Castelnuovo-Mumford k-regular, then Y is cut-out by hypersurfaces of degree k (cf. [La] Theorem 1.8.3). [GLP] Theorem 1.1), answering and extending a classical question of Castelnuovo, turns out to be very useful: if C is a reduced, irreducible, non-degenerate curve of degree d in Pn, then IC is Castelnuovo-Mumford (d + 2 − n)-regular. In this direction, a theorem of Gruson-Lazarsfeld-Peskine (cf. In analogy to Castelnuovo-Mumford regularity, Pareschi and Popa introduced a notion of regularity for sheaves on a polarized abelian variety (X, Θ), the so called Θ-regularity (cf. [PP1] Definition 6.1). It is defined as follows. Given a coherent sheaf F on X, we denote by (1) V i(F) := {α ∈ Pic0(X) hi(X, F ⊗ α) > 0} the cohomological support loci of F. Then we say that F is Mukai-regular (or M-regular for short) if codimPic0(X) V i(F) > i for all i > 0 and that F is k-Θ-regular def⇐⇒ F ⊗ Θ⊗(k−1) is M -regular. The systematic study of Pareschi and Popa on Θ-regularity shows that Θ-regular sheaves enjoy analogous properties to Castelnuovo-Mumford regular sheaves on projective spaces (cf. [PP1] Theorem 6.3). In particular, if IC is k-Θ-regular, then C is cut-out by k-Θ-equations. The aim of these notes is to provide a bound for the Θ-regularity of curves in a polarized abelian variety. Theorem 1.1. Let X be a complex abelian variety of dimension n and Θ be an ample and globally generated line bundle on X. Let ι : C ֒→ X be a reduced and irreducible curve and let ν : eC −→ C be its normalization. Set f := ι ◦ ν and d := deg f ∗Θ. Then the ideal sheaf (cid:0)n + d + 1(cid:1)-Θ-regular. IC is Since the square of an ample line bundle on an abelian variety is globally generated, we obtain the bound (2n + 4d + 1) in case the polarization Θ is not globally generated. As in the bound of Gruson-Lazarsfeld-Peskine, we note that our bound is linear and depends only on 1 2 LUIGI LOMBARDI AND WENBO NIU the dimension of the ambient space and on the degree of the curve. On the other hand it is not sharp, as previous computations of Θ-regularity show that both Abel-Jacobi and Abel-Prym curves are 3-Θ-regular ([PP1] Theorem 4.1, [PP2] Theorem 7.17 and [CMLV] Corollary B). We point out that other bounds for Θ-regularity have been worked out in [PP1] Theorem 6.5 for subvarieties of a polarized abelian variety (X, Θ) defined by d-Θ-equations. The proof of Theorem 1.1 goes as follows. As in [GLP] Theorem 1.1, we use Eagon- Northcott complexes to resolve IC with a complex of locally free sheaves such that: 1) it is exact away from C; 2) the Theta-regularity of its terms is easily computable. However, the methods to establish this resolution differ considerably from the ones used by Gruson- Lazarsfeld-Peskine as they mainly rely on the generic vanishing theory developed in [PP1] and [PP2]. 2. Setting and Proof Throughout the paper we work in the following setting. Let X be a complex abelian variety of dimension n and Θ be an ample and globally generated line bundle on X. Let ι : C ֒→ X be a reduced and irreducible curve of geometric genus g and ν : eC −→ C be its normalization. Set f = ι ◦ ν and d := deg f ∗Θ. Let Γ ⊂ eC × X be the graph of f and p and q be the projections from eC × X onto the first and second factor respectively. Finally we denote by IC and IΓ the ideal sheaves of C and Γ respectively. We recall that a coherent sheaf F on X is continuously globally generated if there exists a positive integer N such that for any general α1, . . . , αN ∈ Pic0(X) the sum of the twisted evaluation maps NM i=1 H 0(X, F ⊗ αi) ⊗ α−1 i −→ F is surjective. For instance, ample line bundles and, in general, M -regular sheaves are continu- ously globally generated (cf. [PP1] Proposition 2.13). In order to compute the Θ-regularity of IC, we will show that it is enough to check the continuous global generation of sheaves of type q∗(IΓ⊗p∗A)⊗Θ where A is a globally generated line bundle on eC (cf. in Proposition 2.2). To begin with we present a simple lemma. Lemma 2.1. Let E • : · · · −→ E2 d2−→ E1 d1−→ E0 d0−→ J −→ 0 be a complex of coherent sheaves on a smooth projective irregular variety Y such that d0 is a surjective morphism. If E • is exact away from an algebraic set of dimension ≤ 1, then we have an inclusion of cohomological support loci ( cf. (1)) V i(J ) ⊂ V i(E0) ∪ V i+1(E1) ∪ . . . ∪ V dim Y (Edim Y −i) for any i ≥ 1. Proof. We set Ki dim Supp Hi ≤ 1 for any i ≥ 0. Therefore := ker di, Ii := im di and Hi := Ki/Ii+1 for i ≥ 0. By assumption By looking at the exact sequence 0 −→ K0 −→ E0 −→ J −→ 0, we have V j(Hi) = ∅ for any i ≥ 0 and j > 1. V i(J ) ⊂ V i(E0) ∪ V i+1(K0). REGULARITY OF CURVES IN ABELIAN VARIETIES 3 In addition, the exact sequence 0 −→ I1 −→ K0 −→ H0 −→ 0 yields V i+1(K0) ⊂ V i+1(I1) ∪ V i+1(H0) = V i+1(I1). Finally, by looking at the exact sequence 0 −→ K1 −→ E1 −→ I1 −→ 0, we deduce that and therefore that V i+1(I1) ⊂ V i+1(E1) ∪ V i+2(K1) V i(J ) ⊂ V i(E0) ∪ V i+1(E1) ∪ V i+2(K1). At this point it is enough to iterate the previous argument to obtain the lemma. (cid:3) Proposition 2.2. Let A be a globally generated line bundle on eC. If q∗(IΓ ⊗ p∗A) ⊗ Θ is continuously globally generated, then IC is (cid:0)h0(eC, A) + n(cid:1)-Θ-regular. Proof. Consider the exact sequence defining the graph Γ (2) 0 −→ IΓ −→ O eC×X −→ OΓ −→ 0. By tensoring (2) by p∗A and then by pushing forward to X via q, we obtain the exact sequence (3) 0 −→ q∗(p∗A ⊗ IΓ) −→ H 0(eC, A) ⊗ OX ev−→ f∗A. Let G be the image of the evaluation map ev. By hypotheses, there exists a positive integer N and line bundles α1, . . . , αN ∈ Pic0(X) such that the map (4) NM i=1 H 0(X, q∗(p∗A ⊗ IΓ) ⊗ Θ ⊗ αi) ⊗ α−1 i −→ q∗(p∗A ⊗ IΓ) ⊗ Θ is surjective. We set Wi := H 0(X, q∗(p∗A ⊗ IΓ) ⊗ Θ ⊗ αi) for i = 1, . . . , N . From (3) and (4), we deduce a presentation of G: (5) We set NM i=1 Wi ⊗ α−1 i ⊗ Θ−1 ϕ −→ H 0(eC, A) ⊗ OX −→ G −→ 0. E := NM i=1 Wi ⊗ α−1 i ⊗ Θ−1 and h := dim H 0(eC, A). Let J be the 0-th Fitting ideal of G. Since A is globally generated and IC is a radical ideal, J coincides with IC away the singular points of C (see [GLP] p.496). By applying the Eagon- Northcott complex (cf. [GLP] (0.4)) to the morphism ϕ in (5), we get a complex E • : · · · −→ E1 −→ E0 −→ J −→ 0 which is exact away from C and whose terms are copies of Vh+i E for all i ≥ 0. Therefore ∼= M Ei Θ⊗(−h−i) ⊗ βit for some βit ∈ Pic0(X). Moreover we have t V j(Ei ⊗ Θ⊗(h+n−1)) ⊂ V j(Θ⊗(n−i+1)) = ∅ for any j > 0 and i < n − 1 and V j(En−1 ⊗ Θ⊗(h+n−1)) ⊂ V j(OX ) = {OX } for any j > 0. 4 LUIGI LOMBARDI AND WENBO NIU Thus, by Lemma 2.1 we obtain inclusions V 1(J ⊗ Θ⊗(h+n−1)) ⊂ V n(En−1 ⊗ Θ⊗(h+n−1)) ⊂ {OX } and Finally, by noting that V j(J ⊗ Θ⊗(h+n−1)) = ∅ for any j > 1. V j(IC ⊗ Θ⊗(h+n−1)) ⊂ V j(J ⊗ Θ⊗(h+n−1)) for any j > 0, we conclude then that IC ⊗ Θ⊗(h+n−1) is an M -regular sheaf. (cid:3) In the next proposition we will give sufficient conditions for the hypotheses of Proposition 2.2 to be satisfied. Proposition 2.3. If B is a line bundle on eC such that f∗B is M -regular on X, then q∗(IΓ ⊗ p∗(B ⊗ f ∗Θ)) ⊗ Θ is M -regular on X. Proof. The sheaf f∗B ⊗ Θ⊗m is globally generated for any m ≥ 1 by [PP1] Proposition 2.12. Moreover, by [PP2] Proposition 3.1, f∗B ⊗ Θ⊗m is an I.T.0 sheaf for any m ≥ 1, i.e. V i(f∗B ⊗ Θ⊗m) = ∅ for all i, m ≥ 1. By (3) we obtain exact sequences for any α ∈ Pic0(X) (6) 0 −→ q∗(IΓ ⊗p∗(B ⊗f ∗Θ))⊗Θ⊗α −→ H 0(eC, B ⊗f ∗Θ)⊗Θ⊗α −→ f∗B ⊗Θ⊗2 ⊗α −→ 0. We set H := q∗(IΓ⊗p∗(B⊗f ∗Θ)) so that we only need to check the conditions codimPic0(X)V i(H⊗ Θ) > i for all i > 0. By the Kodaira Vanishing Theorem and by (6), we have Furthermore V i(H ⊗ Θ) = ∅ for any i ≥ 3. V 2(H ⊗ Θ) ∼= V 1(f∗B ⊗ Θ⊗2) = ∅. Now we study the dimension of V 1(H ⊗ Θ). By denoting by (7) mα : H 0(X, f∗B ⊗ Θ) ⊗ H 0(X, Θ ⊗ α) −→ H 0(X, f∗B ⊗ Θ⊗2 ⊗ α) the multiplication map on global sections induced by (6), we have an identification (8) We claim that the inverse morphism (−1) : Pic0(X) −→ Pic0(X) taking α to α−1 maps V 1(H ⊗ Θ) = {α ∈ Pic0(X) mα is not surjective}. (9) V 1(H ⊗ Θ) 7→ V 1(f∗B). This finishes the proof since it implies dim V 1(H ⊗ Θ) ≤ dim V 1(f∗B) ≤ n − 2. Now we show (9). Since Θ ⊗ α is globally generated for any α ∈ Pic0(X), we have exact sequences 0 −→ MΘ⊗α −→ H 0(X, Θ ⊗ α) ⊗ OX −→ Θ ⊗ α −→ 0. Tensoring by Θ and then restricting to C and finally tensoring by ν∗B, we get exact sequences 0 −→ ι∗(MΘ⊗α ⊗ Θ) ⊗ ν∗B −→ H 0(X, Θ ⊗ α) ⊗ ι∗Θ ⊗ ν∗B evα−→ ι∗(Θ⊗2 ⊗ α) ⊗ ν∗B −→ 0. REGULARITY OF CURVES IN ABELIAN VARIETIES 5 We note that the map on global sections induced by evα coincides with the multiplication map (7). Hence by (8) (10) α ∈ V 1(H ⊗ Θ) =⇒ H 1(C, ι∗(MΘ⊗α ⊗ Θ) ⊗ ν∗B) 6= 0. Now pick an arbitrary element α ∈ V 1(H⊗Θ) and set W := Im(cid:0)H 0(X, Θ⊗α) −→ H 0(C, ι∗(Θ⊗ α))(cid:1). We note that W generates ι∗(Θ ⊗ α) and hence dim W ≥ 2 since Θ is not trivial. Moreover, the preimages s1 and s2 in H 0(X, Θ ⊗ α) of two general sections in W generate ι∗(Θ ⊗ α). We have then a commutative diagram 0 0 / ι∗(Θ−1 ⊗ α−1) / OC ⊕ OC (s1,s2) / ι∗(Θ ⊗ α) / 0 / ι∗MΘ⊗α / H 0(X, Θ ⊗ α) ⊗ OC / ι∗(Θ ⊗ α) / 0. By defining eV := H 0(X, Θ⊗α)/hs1, s2i and by the Snake Lemma, we obtain the exact sequence 0 −→ ι∗(Θ−1 ⊗ α−1) −→ ι∗MΘ⊗α −→ eV ⊗ OC −→ 0 and hence the exact sequence 0 −→ ι∗α−1 ⊗ ν∗B −→ ι∗(MΘ⊗α ⊗ Θ) ⊗ ν∗B −→ eV ⊗ ι∗Θ ⊗ ν∗B −→ 0. Finally, the projection formula yields isomorphisms H 1(C, ν∗B ⊗ ι∗α−1) ∼= H 1(X, f∗B ⊗ α−1) and H 1(C, ν∗B ⊗ ι∗Θ) ∼= H 1(X, f∗B ⊗ Θ) = 0 which in turn show that α−1 ∈ V 1(f∗B) by (10). (cid:3) Before proceeding with the proof of Theorem 1.1, we prove a lemma giving a necessary condition for a sheaf of the form f∗B to be M -regular on X. Lemma 2.4. Let D be a smooth and irreducible curve of genus g and ϕ : D −→ X be a morphism to a complex abelian variety X of dimension n. If B is a general line bundle on D of degree b, then dim V 1(ϕ∗B) ≤ n + g − b − 2. Proof. Without loss of generality, we can assume that ϕ in non-constant since in this case V 1(ϕ∗B) = ∅ for all B ∈ Pic0(D). The algebraic set V 1(B) is irreducible as Serre duality yields an isomorphism V 1(B) ∼= W2g−2−b(D) (here W2g−2−b(D) is the image of the Abel- Jacobi map Sym2g−2−b(D) −→ Pic2g−2−b(D)). The algebraic group Pic0(D) acts on itself via translations. For any γ ∈ Pic0(D), we write γV 1(B) for the image of V 1(B) under the action of γ. Then, by Kleiman's Transversality Theorem [Kl] Theorem 2, there exists an open dense subset V ⊂ Pic0(D) such that for all γ ∈ V the fiber product γV 1(B) ×Pic0(D) Pic0(X) is either empty or of dimension dim V 1(B) + dim Pic0(X) − dim Pic0(D) = dim V 1(B) + n − g. We note the isomorphisms of algebraic sets γV 1(B) ∼= V 1(B ⊗ γ−1) for any γ ∈ V . Moreover, by the universal property of the fiber product and by the projection formula, we obtain closed /   /   / / / / / / 6 LUIGI LOMBARDI AND WENBO NIU immersions V 1(ϕ∗(B ⊗ γ−1)) ֒→ V 1(B ⊗ γ−1) ×Pic0(D) Pic0(X). Hence for any γ ∈ V we have dim V 1(ϕ∗(B ⊗ γ−1)) ≤ dim(cid:0)γV 1(B) ×Pic0(D) Pic0(X)(cid:1) = dim V 1(B) + n − g = dim W2g−2−b(D) + n − g ≤ n + g − b − 2. (cid:3) At this point the proof of Theorem 1.1 easily follows by the previous lemmas and propo- sitions. Proof of Theorem 1.1. By Lemma 2.4 we can pick a line bundle B of degree g on eC such that f∗B is M -regular on X and h1(eC, B ⊗ f ∗Θ) = 0. Hence, the sheaf q∗(IΓ ⊗ p∗(B ⊗ f ∗Θ)) ⊗ Proposition 2.2 after having noted that h0(eC, B ⊗ f ∗Θ) = d + 1. Θ is continuously globally generated by Proposition 2.3 and we conclude then by applying (cid:3) Acknowledgements. We thank Lawrence Ein, Angela Ortega, Giuseppe Pareschi and Mih- nea Popa for stimulating and useful conversations. We also thank the Mathematics Research Communities (a program of the AMS) for having supported a visit of the second author by the first. References [CMLV] S. Casalaina-Martin, M. Lahoz, F. Viviani Cohomological support loci for Abel-Prym curves, 2008, Le Matematiche, LXIII [GLP] L. Gruson, R. Lazarsfeld, C. Peskine On a theorem of Castelnuovo, and the equations defining space curves, 1983, Invent. Math., 72, 491-506 [Kl] S. L. Kleiman The transversality of a general translate, 1974, Compos. Math., 28, 287-297 [La] R. Lazarsfeld Positivity in Algebraic Geometry Vol. 1, Ergebnisse der Mathematik und ihrer Grenzge- biete, Springer-Verlag, Berlin, 48, 2004. [PP1] G. Pareschi, M. Popa Regularity on abelian varieties I, 2003, J. Amer. Math. Soc, 16, 285-302 [PP2] G. Pareschi, M. Popa Regularity on abelian varieties III: relationship with generic vanishing and ap- plications, 2011, in Grassmannians, Moduli Spaces and Vector Bundles, Clay Mathematics Proceedings 14, Amer. Math. Soc., Providence, RI, 141-167 Department of Mathematics, Statistics, and Computer Science, University of Illinois at Chicago, 851 S. Morgan Street, Chicago, IL, 60607 E-mail address: [email protected] Department of Mathematics, Purdue University, 150, N. University Street, West Lafayette, IN, 47907 E-mail address: [email protected] 

1205.4577

2

1205

$p^{-1}$-linear maps in algebra and geometry

In this article we survey the basic properties of $p^{-e}$-linear endomorphisms of coherent $\O_X$-modules, i.e. of $\O_X$-linear maps $F_* \sF \to \sG$ where $\sF,\sG$ are $\O_X$-modules and $F$ is the Frobenius of a variety of finite type over a perfect field of characteristic $p > 0$. We emphasize their relevance to commutative algebra, local cohomology and the theory of test ideals on the one hand, and global geometric applications to vanishing theorems and lifting of sections on the other.

math.AG

math

p−1-LINEAR MAPS IN ALGEBRA AND GEOMETRY MANUEL BLICKLE AND KARL SCHWEDE Abstract. In this article we survey the basic properties of p−e-linear endomorphisms of coher- ent OX -modules, i.e. of OX -linear maps F∗F −→ G where F , G are OX -modules and F is the Frobenius of a variety of finite type over a perfect field of characteristic p > 0. We emphasize their relevance to commutative algebra, local cohomology and the theory of test ideals on the one hand, and global geometric applications to vanishing theorems and lifting of sections on the other. Contents Introduction 1. 2. Preliminaries on Frobenius 3. p−e-linear maps: definition and examples 4. Connections with divisors 5. Frobenius splittings 6. Frobenius non-splittings 7. Change of variety 8. Cartier modules 9. Applications to local cohomology and test ideals Appendix A. Reflexification of sheaves and Weil divisors References 1 3 5 11 16 23 31 36 45 56 59 1. Introduction In this survey we study the basic properties of p−1-linear morphisms between coherent sheaves on a scheme X over a perfect field of positive characteristic p. If F : X −→ X is the Frobenius morphism (i.e. the pth power map on the structure sheaf) we denote by F∗ the restriction functor along F (cf. Section 2.2). A p−1-linear map is then an OX-linear map ϕ : F∗F −→ G for two OX -modules F and G . The name stems from the fact that if we view ϕ as a map on the underlying sheaves of Abelian groups, ϕ satisfies the condition ϕ(rpf ) = rϕ(f ) for local sections r ∈ OX and f ∈ F . In particular, if r has a pth root, then we may write this relation as ϕ(rf ) = rp−1 ϕ(f ). As an example for a p−1-linear map, we start with a splitting of the Frobenius map, that is an OX -linear map ϕ : F∗OX −→ OX such that the composition OX F −−→ F∗OX ϕ −−→ OX is equal to the identity. The mere existence of such a ϕ has strong implications for the local geometry of X (it is reduced, for example). Furthermore, it immediately implies a highly effective version of Serre vanishing: the higher cohomology of any ample line bundle vanishes. In the light 2010 Mathematics Subject Classification. 13A35, 13D45, 14B05, 14B15, 14C20, 14F17, 14F18. The first author was partially supported by a Heisenberg Fellowship and the SFB/TRR45. The second author was partially supported by the NSF grant DMS #1064485. 1 2 MANUEL BLICKLE AND KARL SCHWEDE of such strong implications, it is somewhat surprising that there are varieties of interest that are Frobenius split. For example, regular affine varieties, projective spaces, normal toric varieties, and most prominently flag- and Schubert varieties are Frobenius split. And it was precisely for the latter varieties where the above vanishing yields a simple proof of Kempf's vanishing theorem [MR85], see also [Hab80]. Frobenius split varieties have been extensively studied [BK05] and in Section 5 we give a detailed account of their theory, explaining some of the more delicate vanishing and extension results, and discussing criteria to decide if a given variety is Frobenius split. In Section 6 we show how some of the results and techniques for Frobenius splittings can be extended to more general contexts (where the variety is not F -split) to derive similar conclusions (vanishing and extension results). For example, a systematic use of certain p−1-linear maps can replace Kodaira and Kawamata-Viehweg vanishing theorems [Kaw82, Vie82] in some applica- tions, see Section 6.1. These techniques rely on an explicit connection between p−e-linear maps L , OX ) and Q-divisors ∆ such that OX((pe − 1)(KX + ∆)) ∼= L −1 which is ϕ ∈ HomOX (F e ∗ explained in detail in Section 4. Indeed, this correspondence between p−e-maps and Q-divisors pervades much of the paper. This correspondence also provides us with valuable geometric intuition in working with p−e-linear maps. In Section 7 we state a number of general results on the behavior of p−e-linear maps under certain functorial operations, such as pullback along closed immersions, localization, pushfor- ward along a birational map, and finally pullback along a finite map. In all these cases, viewing p−e-linear maps as Q-divisors and performing operations on divisors is the guiding principle. A second key example of a p−1-linear map is the classical Cartier operator C : F∗ωX −→ ωX introduced in [Car57]. There are various guises in which this operator on the dualizing sheaf appears, but most generally one may view it as the trace of Frobenius under the duality for finite morphisms, see Section 3.1. The Cartier operator has been extensively studied in connection to residues of differentials in positive characteristic, and plays a crucial role in Deligne and Illusie's [DI87] algebraic proof of Kodaira vanishing. In the final two Sections 8 and 9 we describe the category of Cartier modules introduced in [BB11]. This category consists of coherent OX -modules F equipped with a p−e-linear endomor- phism, i.e. a OX -linear map F e F −→ F . We show that the Abelian category of Cartier modules ∗ satisfies some remarkable properties. Most importantly, Cartier modules have finite length up to nilpotence1. Furthermore, Cartier modules are related to a number of other categories which have been extensively used in the study of local cohomology in positive characteristic. Hence the finiteness results about Cartier modules imply and generalize previous finiteness results about local cohomology, see Section 9.1 where we indicate how results of Hartshorne-Speiser [HS77], Lyubeznik [Lyu97] and Enescu and Hochster [EH08] can be derived easily. In the final section we explain a certain degree-reducing property of pe-linear maps and show how this property yields a completely elementary approach to the above mentioned finiteness result. In the last subsection we finally close the gap to the theory of tight closure [HH90, Hun96, Hoc07a], which im- and explicitly heavily relied on p−e-linear maps since its be- ginnings, in showing how the test ideals of Hara and Yoshida [HY03] are obtained from certain generalizations of Cartier modules. We include as another demonstration of the utility of this viewpoint a quick proof of the discreteness of the jumping numbers of the test ideal. The target audience for this article is a researcher or student who is familiar with commuta- tive algebra and algebraic geometry and who wishes to learn how to use p−1-linear maps in a wide variety of contexts. We do not assume the reader has one particular background (i.e. rep- resentation theory/Frobenius splitting, tight closure theory, D-modules, or higher dimensional complex algebraic geometry). Because we view this article as a place where material can be 1A coherent Cartier module F is nilpotent is some power of the structural map is zero. p−1-LINEAR MAPS IN ALGEBRA AND GEOMETRY 3 learned, at the end of each section there are many many exercises. The more difficult exercises are decorated with a *. The exercises are a fundamental part of this document. Acknowledgements. The authors are deeply indebted to Alberto Fernandez Boix, Lance Miller, Claudiu Raicu, Kevin Tucker, Wenliang Zhang and the referee for innumerable valu- able comments on previous drafts of this paper. 2. Preliminaries on Frobenius In this section we introduce our conventions on notation -- in particular with regards to the Frobenius morphism. 2.1. Prerequisites and Notation. We assume that the reader is familiar with the basics of commutative algebra and algebraic geometry, all of which is covered in the standard refer- ence works [Har77] and [Mat80]. Beyond this, a familiarity with Grothendieck duality, [Har66b, Con00], will be particularly helpful. Explicitly, Serre vanishing, canonical modules, dualizing and Serre duality, and the connection between divisors and line bundles, will appear frequently, see also [BH93]. The notion of Q-divisors will be used extensively (see [KM98] or [Laz04a, Laz04b]). The process of reflexification of sheaves on normal varieties and its relation to Weil divisors will be recalled in Appendix A for the convenience of the reader, also see [Har94] where the same theory is worked out in substantially greater generality. Throughout this paper all rings and schemes are assumed to be of finite type over a perfect field k of characteristic p > 0, or they are a localization or completion of such at a prime. This implies that our schemes are excellent and possess canonical modules and dualizing complexes [Mat80, Har66b]. We further assume that all schemes are separated. 2.2. Frobenius and push-forward. We begin by reviewing the most basic notation (since it varies wildly in the literature). The key structure in algebra and geometry over a field of positive characteristic p > 0 is the (absolute) Frobenius endomorphism. For a ring R this is just the pth power ring endomorphism given by sending r ∈ R to rp. F = FR : R −→ R Since the Frobenius is canonical it induces a morphism for any scheme X over a field k of characteristic p > 0, also called the Frobenius endomorphism and also denoted by F = FX : X −→ X. Supposing that k is perfect and X is a k-variety (or a scheme according to our convention) then FX is a finite map2 by Exercise 2.2. Note that FX is in general not a morphism of k-schemes -- however this point can be rectified by changing the k-structure on the first copy of X, if desired. We denote by F e the e-fold self composition of Frobenius. Even in the affine situation X = Spec R we use geometric notation and denote the Frobenius on R by F : R −→ F∗R to remind us that it is not R-linear. This has the added benefit that we now can distinguish the source and target of F = FR. Given an ideal I = hf1, . . . , fmi ⊆ R, we define its peth Frobenius power to be I [pe] = hf pe 1 , . . . , f pe m i. This is independent of the choice of generators fj, see Exercise 2.3. The formation of I [pe] commutes with localization and so for any ideal sheaf I ⊆ OX , we can define I [pe] in the obvious way. Note F e namely, OX acts on F e that F e ∗ ∗ OX is isomorphic to OX as a sheaf of rings -- but as OX-modules they are distinct: ∗ OX via peth powers. More generally, for any OX -module M , one observes M is isomorphic to M as a sheaf of Abelian groups but the OX -module structure is 2An abstract scheme with a finite Frobenius is called F -finite. 4 MANUEL BLICKLE AND KARL SCHWEDE m for a local section r ∈ OX and m ∈ F e ∗ M also has ∗ OX -module structure, which coincides with M 's original OX-module structure. We also ∗ M in the affine case X = Spec R to denote an R-module with the twisted given by r.m = rpe an F e use the notation F e (restriction of scalars) Frobenius structure. M . Of course, F e ∗ One immediately verifies that F∗fM coincides with ]F∗M as OX -modules, where fM denotes the OX -module associated to the R-module M . However, we caution the reader that the same identification does not hold in the graded case with respect to Proj, see for example [SS10, Lemma 5.6] and Exercise 2.7. Notation 2.2.1. Given an element m ∈ M , we will sometimes use F e sponding element of F e ∗ M . Likewise for sheaves of OX-modules M on X. ∗ m to denote the corre- 2.3. Frobenius pull-back and the projection formula. Let X be a scheme over a perfect field k of characteristic p > 0 and let F be a coherent sheaf and L a line bundle on X. Since the Frobenius is an isomorphism on the underlying topological space, the pullback F e∗F (as an OX -module) can be identified with F ⊗OX F e ∗ OX is isomorphic with OX as sheaves of rings. If the line bundle L is given by the datum of a local trivialization and transition functions, then the line bundle F e∗L is given by the peth powers of the transition functions in that datum for L . This shows that ∗ OX -module, again using that F e ∗ OX as an F e (1) i.e. the pullback along the Frobenius of a line bundle just raises that line bundle to the peth tensor power. Combining this observation with the projection formula [Har77, Chapter II, Exercise 5.1(d)] we obtain , (F e)∗L ∼= L pe (2) (F e ∗ F ) ⊗OX L ∼= F e ∗ (F ⊗OX F e∗L ) ∼= F e ∗ (F ⊗OX L pe ) . This basic equality is used frequently throughout the theory and will be referred to as the projection formula. 2.4. Exercises. Exercise 2.1. Set X = Spec k[x1, . . . , xn] for some perfect field k. Show that F e OX -module with basis F e holds for power series Spec kJx1, . . . , xnK. ∗ OX is a free n where 0 ≤ λi ≤ pe − 1. Show that the same result also 1 · · · xλn ∗ xλ1 Exercise 2.2. Suppose that k is a perfect field and that X is scheme of (essentially) finite type over k. Prove that the Frobenius map on X is a finite map. Exercise 2.3. Suppose that I ⊆ R is an ideal in a ring R of characteristic p > 0. Show that I [pe] can be identified with Image(F e∗I −→ F e∗R) ⊆ F e∗R ∼= R where the last isomorphism is ∗ r. Conclude that I [pe], the Frobenius the canonical one identifying R with F e ∗ R sending r to F e power of I, is independent of the choice of generators of I. Exercise 2.4. Suppose that X is a smooth d-dimensional variety and L is a vector bundle of rank m on X. Prove that F∗L is also a vector bundle and find its rank. Hint: Complete, use Cohen structure theorem [Mat89, Theorem 28.3], and use Exercise 2.1. Exercise 2.5. Suppose that E is a locally free sheaf of finite rank on X. Is E ⊗pe (F e)∗E ? isomorphic to Exercise 2.6. Suppose that R is (essentially) of finite type over a perfect field. (a) If W ⊆ R is any multiplicatively closed set, then show that W −1(F e ∗ R) ∼= F e ∗ (W −1R). Here the first F e ∗ R means as an R-module, and the second is as an W −1R-module. (b) If m ⊆ R is a maximal ideal, prove that F e ∗ R where denotes completion along m. Again, the first F e ∗ is the Frobenius for R, and the second is that of R-modules. ∗bR ∼= dF e p−1-LINEAR MAPS IN ALGEBRA AND GEOMETRY 5 to F . Verify that F e is a graded F e Exercise 2.7. Suppose that X is a projective variety with ample line bundle L and suppose that F is a coherent sheaf on X. Set S =Li∈Z H 0(X, L i) to be the section ring with respect to L and set M =Li∈Z H 0(X, F ⊗ L i) to be the saturated graded S-module corresponding Li∈Z H 0(X, (F e ∗ M is not in general isomorphic toLi∈Z H 0(X, (F e F )⊗L i). Instead, prove that ∗ M . The summand F ) ⊗ L i) is isomorphic to a (graded) direct summand of F e pe · Z)-graded ring3, the natural map S −→ F e ∗ S-module. Of course, F e ∗ M is also a graded S-module. whose terms have integral gradings. ∗ S is graded, and F e ∗ M Show that F e ∗ ∗ S is a ( 1 ∗ Exercise 2.8. A ring R (or scheme X) such that the Frobenius map F : R −→ F∗R is a finite map is called F -finite. Essentially all rings considered in this paper are F -finite but not all rings are. Find an example of a field which is not F -finite. If X is a smooth variety, then we have already seen that F∗OX is a locally free (in other words flat) OX -module. In this exercise, you will prove the converse. First we introduce a definition. Definition 2.4.1. Suppose that (R, m) is a local ring. A sequence of elements f1, . . . , fn ∈ m ⊆ R is called Lech-independent if for any a1, . . . , an ∈ R such that a1f1 + · · · + anfn = 0, then each ai ∈ hf1, . . . , fni. Now we come to the exercise. Exercise* 2.9 (Kunz's regularity criterion [Kun69]). Suppose that (R, m) is a local ring. We will show that if F∗R is flat, then R is regular. We need some Lemmas due to Lech [Lec64]. (i) [Lec64, Lemma 3]. If f1, . . . , fn are Lech-independent elements and f1 ∈ gR for some g ∈ R, then g, f2, . . . , fn is also Lech-independent. Furthermore, hf2, . . . , fni : g ⊆ hf1, . . . , fni. (ii) [Lec64, Lemma 4]. If f1, . . . , fn are Lech-independent and f1 = gh. Then (iii) Now we return to the proof of the theorem of Kunz. Show that m lR (R/hf1, . . . , fni) = lR (R/hg, f2, . . . , fni) + lR (R/hh, f2, . . . , fni) . [pe]/(m [pe])2 is a free [pe]-module. Conclude that if m = hx1, . . . , xni is generated by a minimal set of R/m generators, then xpe 1 , . . . , xpe n is Lech-independent. (iv) Use the previous parts of the exercise to conclude that lR(R/m (v) Reduce to the case that R is complete and write R = S/a = k[[x1, . . . , xn]]/a using the Cohen structure theorem [Mat89, Theorem 28.3] where k = R/m. Then notice that lS(S/m [pe] S ) = pne for all e ≥ 0. Complete the proof of Kunz' regularity criterion. [pe]) = pne. Remark 2.4.2. A simpler proof of Kunz' result using the Buchsbaum-Eisenbud acyclicity crite- rion can be found on page 12 of [Hun96]. Alberto Fernandez Boix pointed out to us that another short proof can be found in [MR10, Theorem 4.4.2]. 3. p−e-linear maps: definition and examples In this section we introduce p−e-linear maps and give a number of examples which will be discussed in more detail throughout the rest of the paper. Definition 3.0.3 (p−e-linear map). Suppose that X is a scheme and M and N are OX - modules. A p−e-linear map is an additive map ϕ : M −→ N such that (3) ϕ(rpe m) = rϕ(m) for all local sections r ∈ OX and m ∈ M . 3Here 1 pe · Z is the subgroup of Q generated by 1 pe . 6 MANUEL BLICKLE AND KARL SCHWEDE Equivalently, we may specify a p−e-linear map by the data of an OX -linear map ϕ : F e ∗ M −→ N . We will frequently and freely switch between these two points of view, depending on the context. If R is a ring, then a p−e-linear map between R-modules M and N is simply an additive map between them satisfying the rule from (3). If k is a perfect field, then p−e-linearity for an additive map ϕ : k −→ k just means ϕ(λx) = λ1/pe ϕ(x) for all x, λ ∈ k. In particular, such a map is completely determined by where it sends any nonzero element. Example 3.0.4. Consider R = k[x]. Then F∗R is a free module with basis {F∗1, F∗x, F∗x2, . . . , F∗xp−1}, see Exercise 2.1 above. Therefore any p−1-linear map from k[x] to any other R-module N is simply a choice of where to send these basis elements. Example 3.0.5. Consider R = k[x1, . . . , xn] then as we saw in Exercise 2.1, F∗R is a free R- module with basis F∗xλ1 n for 0 ≤ λi ≤ p − 1. A map F∗R −→ R is uniquely determined by where it sends the elements of this basis. 1 · · · xλn Consider the R-linear map Φ : F∗R −→ R which sends F∗xp−1 · · · xp−1 n 1 to 1 and all other basis elements to zero. In other words: Φ(cid:16)F∗(xλ1 1 · · · xλn n )(cid:17) = λ1−(p−1) p x 1 λn−(p−1) · · · x n p , if all λi−(p−1) p ∈ Z 0, otherwise For each tuple λ = (λ1, . . . , λn) ∈ {0, 1, . . . , p − 1}n, consider the map ϕλ : F∗R −→ R defined ) = Φ(F∗(xp−1−λ1 )). It is easy to see that ϕλ sends xλ to 1 by the rule ϕλ(F∗ and all other basis monomials to zero. · · · xp−1−λn n 1 · Because we can thus obtain all of the projections this way, it follows that the map F∗R −→ )) is surjective as a map of HomR(F∗R, R) which sends F∗z to the map ϕz(F∗ F∗R-modules. On the other hand, it is clearly injective as well and so it is an isomorphism. In other words, we just showed that HomR(F∗R, R) is a free F∗R -- module generated by Φ. In other words, Φ generates HomR(F∗R, R) as an F∗R-module. ) = Φ(F∗(z · The most pervasive type of p−1-linear maps are maps ϕ : R −→ R. Of course, for fixed e, the set of p−e-linear maps {ϕ : R −→ R ϕ is p−e-linear} form a group under addition. However, as we vary e, we have a multiplication of these maps as well. Indeed, suppose that ϕ : R −→ R is p−e-linear and ψ : R −→ R is p−d-linear. Then both ϕ ◦ ψ and ψ ◦ ϕ are p−e−d-linear. However, they need not be equal as the following example shows. It follows that Me≥0(cid:8)ϕ : R −→ R ϕ is p−e-linear(cid:9) forms a noncommutative graded ring. This graded ring will be studied more in Section 9.3. Example 3.0.6. Suppose that R = Fp[x]. We will describe two p−1-linear maps, ϕ, ψ presented as in Example 3.0.4. ◦ ϕ : R −→ R satisfies ϕ(xp−1) = 1 and ϕ(xi) = 0 for 0 ≤ i < p − 1. ◦ ψ : R −→ R satisfies ψ(1) = 1 and ψ(xi) = 0 for 0 < i ≤ p − 1. Then ψ ◦ ϕ and ϕ ◦ ψ are p−2-linear maps. However, notice that but that ϕ(ψ(xp−1)) = ϕ(0) = 0 ψ(ϕ(xp−1)) = ψ(1) = 1. p−1-LINEAR MAPS IN ALGEBRA AND GEOMETRY 7 In particular, ψ ◦ ϕ 6= ϕ ◦ ψ. An important class of examples of p−1-linear maps are the splittings of Frobenius. Example 3.0.7. [MR85, RR85] Let X be a scheme. A Frobenius splitting is any p−1-linear map ϕ : OX −→ OX that sends 1 to 1. Equivalently, it is an OX -linear map ϕ : F∗OX −→ OX that sends F∗1 to 1. This, in particular, implies that the composition (4) OX F −→ F∗OX ϕ −→ OX is an isomorphism, hence ϕ "splits" the Frobenius. If X has a Frobenius splitting, then it satisfies many remarkable properties as we shall discuss in detail in Section 5. Let us just mention two of them to taste. If X is a scheme that has some Frobenius splitting ϕ : F∗OX −→ OX (we call such X Frobenius split), then X is reduced: Indeed, if x ∈ Γ(U, OX ) is such that 0 = xpe = F e(x) for some e ≥ 0, then 0 = ϕe(F e(x)) = x, simply by that fact that ϕ ◦ F = id. This is a simple but important local property of Frobenius split varieties. A similarly fundamental global result is the following vanishing theorem: Suppose that L is a line bundle and that H i(X, L p) = 0 for some i > 0 (for example, H i(X, L pe ) = 0 holds for e ≫ 0 for ample L by Serre vanishing), then H i(X, L ) = 0 as well since we have the following isomorphism obtained by tensoring (4) by L , using the projection formula and taking cohomology: H i(X, L ) F −→ H i(X, F∗L p) = 0 ϕ −→ H i(X, L ). If e > 1, rinse and repeat. We will study vanishing theorems for Frobenius split varieties in much greater detail in Theorem 5.2.4. 3.1. The Cartier isomorphism. We now come to the most important example of a p−1-linear map, coming from the Cartier operator. Suppose that X is a smooth variety over a perfect field k of characteristic p > 0. Consider the de-Rham complex, Ω q X. This is not a complex of OX - modules (the differentials are not OX-linear). However, the complex is a complex of OX -modules (notice that d(xp) = 0). We now state the Cartier isomorphism. We take this presentation from [Car57], [Kat70], [EV92], and [BK05]. Definition-Proposition 3.1.1. There is a natural isomorphism (of OX -modules): F∗Ω q X C −1 : Ωi X → hi(F∗Ω q X) Remark 3.1.2. It might strike the reader as odd that we put an inverse on C. This is because the isomorphism in the other direction is called the Cartier operator and represented by C. It is just more convenient for us to define C −1 than it is to define C. We will not use the details of this isomorphism later in the paper. However, the map T we obtain from it below in Section 3.2 will be indispensable. Let us explain how to construct this isomorphism C −1. We follow [EV92, 9.13] and [Kat70]. We begin with C −1 in the case that i = 1. We work locally on X (which we assume is affine) and we define C −1 by its action on dx ∈ Ωi X, x ∈ OX ; C −1(dx) := F∗xp−1dx, or rather, the image of F∗xp−1dx in cohomology. In order for this to make sense, we observe that d(xp−1dx) = 0, in other words, that C −1(dx) is in the kernel of d. We now must show that C −1 is additive. df = d p−1Xi=1 = p−1Xi=1 = p−1Xi=1 p(cid:0)p i(cid:1) = (p−1)(p−2)···1 γixiyp−i! γiixi−1yp−i! dx + p−1Xi=1 γiix(p−1)−(p−i)yp−i! dx + p−1Xi=1 i (cid:1). Thus i(cid:0)p−1 p−i(cid:1) = 1 p−i(cid:0)p−1 i!(p−i)! = 1 γi(p − i)xiyp−i−1! dy γi(p − i)xiy(p−1)−i! dy where γi = 1 8 MANUEL BLICKLE AND KARL SCHWEDE Now C −1(d(x) + d(y)) = C −1(d(x + y)) = F∗(x + y)p−1d(x + y), we need to compare this to p· is a F∗xp−1dx + F∗yp−1dy = C −1(dx) + C −1(dy). Write f = 1 formal operation that simply cancels ps from the binomial coefficients. Then p ((x + y)p − xp − yp). Here the 1 df = (x + y)p−1(dx + dy) − xp−1dx − yp−1dy. Therefore, xp−1dx + yp−1dy and (x + y)p−1d(x + y) are the same in Ω1 that C −1 is additive. Finally, we extend by p-linearity to obtain that X(cid:14)d(cid:0)Ω0 X(cid:1). This proves C −1(f dx) = F∗f pxp−1dx. We should also show that C −1 is an isomorphism. We only show that this initial C −1 is injective -- in a special case. Set X = Spec Fp[x, y] (see for example [EV92, Theorem 9.14] for how to reduce to the polynomial ring case in general). Suppose that C −1(f dx + gdy) = 0. Let h ∈ OX be such that we have f pxp−1dx + gpyp−1dy = dh = ∂h ∂x dx + ∂h ∂y dy. Therefore if f =P λi,jyixj we see that X λi,jyipxjp+p−1 = f pxp−1 = ∂h ∂x . However, this is ridiculous unless f dx + gdy = 0. If you take a derivative of some non-zero polynomial in x with respect to x, no output can ever have xjp+p−1 in it. This completes the proof of injectivity of C −1 : Ω1 X) in the case that X = Spec Fp[x, y]. The general case is similar. X → h1(F∗Ω q The surjectivity of C −1 is more involved. See for example, [EV92, Theorem 9.14(d)] or [BK05, Theorem 1.3.4] or do Exercise* 3.1. At this point, we have only defined We define C −1 : Ωi definition for any X. X → hi(F∗Ω q X) for i > 1 using wedge powers of C −1 for i = 1. We make this Ω1 X → h1(F∗Ω q X). Example 3.1.3 (Cartier isomorphism A2). Returning again to X = A2 = Fp[x, y], we explicitly compute C −1 : Ω2 X) at the top cohomology. X → h2(F∗Ω q By definition C −1(f dxdy) = C −1(f dx ∧ dy) := F∗(cid:0)f p(xp−1dx) ∧ (yp−1dy)(cid:1) = F∗f pxp−1yp−1dxdy or rather its image in cohomology. Again, this map is an isomorphism, Exercise 3.2. 3.2. Grothendieck-trace of Frobenius. Suppose that X is a smooth n-dimensional variety over a perfect field k of characteristic p > 0. Then coming from the Cartier isomorphism, Theorem 3.1.1, we have an exact sequence: F∗Ωn−1 X/k d −→ F∗Ωn X/k T −→ Ωn X/k −→ 0 p−1-LINEAR MAPS IN ALGEBRA AND GEOMETRY 9 The surjective map T : F∗Ωn map/operator. X/k =: F∗ωX → ωX := Ωn X/k is often called the trace map or Cartier This map can be constructed in other ways. With X as above, again set ωX = Ωn X/k. Then ωX is a dualizing/canonical module in the sense of [Har77, Chapter III, Section 7] or more generally, [Har66b, Chapter V]. For any finite dominant map π : Y −→ X with Y and X smooth, it is a fact (black- boxed for now [Har66b, Chapter V, Proposition 2.4], [KM98, Proposition 5.68]) that π∗ωY ∼= H omOY (π∗OY , ωX ) as a π∗OY -module. This is described in greater generality on the next page, see the diagram (7). Note that this completely determines ωY as well, since π is finite and so the data of a coherent π∗OY -module on X is equivalent to the data of a coherent OY -module on Y . Now, we also have the following map: (5) This is the map which sends a section ϕ ∈ ωY ∼= Γ(U, H omOY (π∗OY , ωX )) to the element ϕ(1) ∈ Γ(U, ωX). π∗ωY ∼= H omOX (π∗OY , ωX) eval @ 1 −−−−−→ ωX Now we specialize to the case that Y = X and π = F the Frobenius map. Theorem 3.2.1. The map described in (5) is the map T described above (up to choice of isomorphism). Sketch of proof. We only show this for X = Spec Fp[x, y] = A2. By considering Example 3.1.3, we see that the map T sends F∗f pxp−1yp−1dxdy 7→ f dxdy and everything not of that form to zero. So we then consider H omOX (F∗OX , ωX) o ≃ / F∗ωX o F∗dxdy o ≃ / F∗OX . / F∗1 Now, we identify the Φ ∈ HomOX (F∗OX, ωX ) which generates H omOX (F∗OX , ωX) as an F∗OX - module just as in Example 3.0.5. Since ωX = OX · (dxdy) ∼= OX, we notice that Φ sends F∗f pxp−1yp−1 7→ f dxdy and Φ sends things not of this form, to zero. Choosing then ϕ(F∗ map (5) sends ϕ to Φ(F e ) = Φ(F∗c · ∗ c). Making the identification ) ∈ H omOX (F∗OX , ωX ), we see that the evaluation-at-1 (F∗OX) · (F∗dxdy) = F∗ωX ∼= H omOX (F∗OX , ωX) = (F∗OX ) · Φ we immediately observe that T and the evaluation-at-1 map (5) coincide. The general case for X 6= A2 is similar but slightly more technical to write down. Both the map T and the evaluation-at-1 map can be shown to be a local generator of the same H om-sheaf. Thus they coincide up to multiplication by a unit of Γ(X, OX ). (cid:3) 3.3. The Trace map for singular varieties. Suppose that X is a normal variety with U ⊆ X the regular locus. Consider the map T : F e ∗ ωU −→ ωU as described above. This is an element of HomU (F e ∗ ωX, ωX ) since X \ U is a codimension 2 subset of X and X is normal, see Appendix A. Therefore we obtain the following proposition. ∗ ωU , ωU ). However, there is an isomorphism HomOU (F e ∗ ωU , ωU ) ∼= HomOX (F e Proposition 3.3.1. Given any normal variety X there is a trace map T : F e ∗ ωX −→ ωX which agrees with, and is completely determined by the map T described in terms of the Cartier isomorphism on the regular locus U ⊆ X. o / o / o / 10 MANUEL BLICKLE AND KARL SCHWEDE Even for non-normal schemes, we can do something similar, we need to work in the derived category. Suppose that X and Y are schemes of finite type over a field k with a map f : X −→ Y . Then there is functor f ! from D+ coh(X) (bounded below complexes of OY -modules, respectively OX -modules, with coherent cohomology). For a precise definition of f !, please see [Har66b]. Its abstract existence can nowadays be shown quite formally from general principles, cf. [Lip06]. Its key property is that it is right adjoint to Rf∗ in the case that f is proper, see Exercise* 3.5. We will define f ! in two cases which will suffice for our purposes. coh(Y ) to D+ Finite: If f is finite (for example, Frobenius or a closed immersion), then F ∈ Db coh(X) we have an isomorphism of f∗OX -complexes f∗f !F = R H om q OY (f∗OX , F ). (6) where R H om q of F and applying H om q finite so that f∗ is harmless. OY (f∗OX , F ) is the complex obtained by taking an injective resolution ). Note that this completely describes f ! since f is OY (f∗OX , Smooth: If f is smooth of relative dimension n, the for any F ∈ Db coh(X) we have an isomorphism f !F = (Lf ∗F ) ⊗ (∧nΩ1 Y /X)[n]. If f : X −→ Spec k is itself the structural map, then we define the dualizing complex of X, coh(Spec k) as the complex which is trivial except in degree X := f !k to be the dualizing complex on X. X as follows. View k ∈ Db denoted ω q zero where it is k. Then we define ω q Consider the following diagram (7) X f F e / X f Spec k / Spec k F e where the top row is the absolute e-iterated Frobenius on X and the bottom row is the e-iterated Frobenius on k. Notice that the bottom row is an isomorphism (although not the identity) and so (F e)!k ∼= k. The fact that (f ◦ g)! = f ! ◦ g! then implies that ω q X is independent of the choice of Frobenius-variant of the k-structure on X. In particular, we see that (8) ω q X ∼= f !k ∼= (f ◦ F e)!k ∼= (F e ◦ f )!k ∼= (F e)!ω q X. Now we will apply the duality functor R H om q ∗ OX . This operation yields F e OX ( , ω q X) to the Frobenius map OX −→ ω q X ∼= R H om q OX (OX , ω q X) ← R H om q OX (F e ∗ OX , ω q X) ∼= F e ∗ (F e)!ω q X ∼= F e ∗ ω q X where the isomorphisms are in the derived category and the final two isomorphisms are due to Equations (8) and (6) respectively. Taking cohomology of this map of complexes gives us maps (9) hiω q X ← F e ∗ hiω q X ∼= hiF e ∗ ω q X for each integer i ∈ Z. For any equidimensional scheme X over k with dualizing complex ω q X := f !k, we define X) and call it the canonical module of X. It follows that (9) induces a map ωX = h− dim X (ω q F e ∗ ωX −→ ωX. As expected, we then have: Proposition 3.3.2. The map F e regular locus of X. ∗ ωX −→ ωX coincides with the map T defined previously on the   /   / p−1-LINEAR MAPS IN ALGEBRA AND GEOMETRY 11 X → h1(F∗Ω q 3.4. Exercises. Exercise* 3.1. Suppose that k is a perfect field and that X = Spec k[x, y] = A2, prove that C −1 : Ω1 Hint: First prove the result for A1 = Spec Fp[x]. Now considerPj yj(αj +βjxbj dy) = α ∈ Ω1 such that dα = 0 where αj ∈ Ω1 is not divisible by p. Use this to rewrite α and then use the result for A1. A1 and βj ∈ Fp[x]. Deduce that yj+1αj+1 + yjβjdy ∈ dΩ0 X) is surjective. X X if j + 1 This method can be used to do the general proof by induction, see [BK05, Theorem 1.3.4]. Exercise 3.2. Suppose that k is a perfect field and that X = Spec k[x, y] = A2, prove that C −1 : Ω2 X) is an isomorphism. X → h2(F∗Ω q Exercise 3.3. Suppose that R is a regular local ring. We have seen that F∗R is a flat R-module by Exercise* 2.9. Consider the evaluation-at-1 map HomR(F∗R, R) e / R ϕ ✤ / ϕ(F∗1) Fix an isomorphism γ : F∗R −→ HomR(F∗R, R) and consider the composition e ◦ γ : F∗R −→ R. Prove that (e ◦ γ) generates HomR(F∗R, R) as an F∗R-module. ∼= ωX[dim X] is a complex with coho- Exercise 3.4. A variety X is called Cohen-Macaulay if ω q X mology only in degree − dim X. Suppose that H is a Cartier divisor on a Cohen-Macaulay scheme X. Prove that H is also Cohen-Macaulay. Conversely, suppose that H is Cohen-Macaulay, prove that X is Cohen-Macaulay in a neighborhood of H. Hint: Apply the duality functor to the short exact sequence 0 −→ OX(−H) −→ OX −→ H by (6). For the converse statement, use X ) ∼= ω q OX (OH , ω q OH −→ 0 and observe that R H om q Nakayama's Lemma. Exercise* 3.5 (Grothendieck duality). (For those who wish to learn some homological algebra) Grothendieck duality says the following: Theorem. If f : X −→ Y is a proper map of schemes of finite type over a field k, then we have an isomorphism in Db coh(Y ) R H om q OY (Rf∗F , G ) ∼= Rf∗R H om q OX (F , f !G ) for F ∈ Db coh(X) and G ∈ Db coh(Y ). Set Y = Spec k and learn enough about the symbols above to deduce the variant of Serre duality found in Hartshorne, [Har77, Chapter III, Section 7]. 4. Connections with divisors In this section we explain why maps ϕ ∈ HomOX (F e ∗ OX , OX ) contain roughly the same information as a Q-divisor ∆ such that KX + ∆ is Q-Cartier (i.e., such that there exists an integer n such that n∆ is integral and nKX + n∆ is Cartier). These ideas go back at least to the original papers on Frobenius splittings [MR85, RR85]. The difference between this section and those original papers is that we normalize our divisors with respect to iterates of Frobenius and thus obtain Q-divisors.4 The statements in this section are somewhat technical. Therefore, the reader may wish to skim this section for the main idea, and refer back to the numbered bijections as needed throughout the remainder of the article. 4Formal sums of codimension 1 subvarieties with rational coefficients. / / 12 MANUEL BLICKLE AND KARL SCHWEDE Fix X to be a smooth variety of finite type over a perfect field. Consider an element ϕ ∈ HomOX (F e (10) ∗ OX, OX ). We claim that HomOX (F e ∗ OX, OX ) ∼= F e ∗ OX ((1 − pe)KX ). Let us prove this claim. By applying the projection formula as in Equation (2), taking global sections and using the fact that H omOX (F e ∗ OX , OX (KX )) ∼= F e ∗ OX (KX ), we have (11) HomOX (F e ∗ OX, OX ) ∼= HomOX ((F e ∼= HomOX (F e ∼= HomF e ∼= F e ∼= F e ∗ OX (F e ∗ HomOX ((OX (peKX)), OX (KX )) ∗ OX((1 − pe)KX ). ∗ OX) ⊗ OX (KX ), OX (KX )) ∗ (OX (peKX )), OX (KX )) ∗ (OX (peKX )), H omOX (F e ∗ OX, OX (KX ))) See (8), [KM98, Proposition 5.68] or [Har66b]. Alternately, it follows from Grothendieck duality for the finite map F : X −→ X (see Exercise 4.1 below). Therefore, any nonzero map ϕ : F e pe)KX ). By using the fact that F e groups, we see that there is a bijective correspondence: ∗ OX −→ OX induces a nonzero global section of F e ∗ OX((1 − ∗ does not change the underlying structure of sheaves of Abelian (cid:26) nonzero elements ϕ ∈ HomOX (F e ∗ OX , OX ) (cid:27) ←→(cid:26) nonzero elements z ∈ Γ(cid:0)X, OX ((1 − pe)KX)(cid:1) (cid:27) . the following bijection: Note every nonzero global section of OX ((1 − pe)KX ) induces an effective Weil divisor 0 ≤ D ∼ (1 − pe)KX, see Theorem A.2.6. We notice also that two non-zero elements z1, z2 ∈ Γ(cid:0)X, OX ((1 − pe)KX )(cid:1) induce the same divisor if and only if there exists a unit u ∈ Γ(cid:0)U, OX(cid:1) such that uz1 = z2. Therefore, we have  Likewise Γ(cid:0)X, OX ((1 − pe)KX)(cid:1) = Γ(cid:0)U, OX ((1 − pe)KX )(cid:1) since X \ U has codimension ≥ 2 cf. Now suppose that X is normal but not necessarily smooth. Of course, the previous argument works fine on U = Xreg ⊆ X. However, Weil divisors are determined off a set of codimension 2. Γ(X, OX )  ←→ ∗ OX , OX ) (cid:27), [Har94, Proposition 2.9]. In particular, we see that effective divisors linearly equivalent multiplication by units in to (1 − pe)KX (cid:26) nonzero ϕ ∈ HomOX (F e (12) . (12) holds on normal varieties. We continue now to work with normal X. Given an effective Weil divisor D = Dϕ ∼ (1 − pe)KX corresponding to ϕ, set ∆ = ∆ϕ = 1 pe−1 Dϕ. This is an effective Q-divisor. Notice that KX + ∆ = KX + 1 pe − 1 D ∼ KX + 1 pe − 1 (1 − pe)KX = KX − KX = 0 In particular, we obtain a bijective correspondence: (13) (cid:26) nonzero ϕ ∈ HomOX (F e ∗ OX , OX ) (cid:27), multiplication by units in Γ(X, OX )  ←→ At this point, it is natural to ask why should one divide by pe − 1. This division is a normalizing factor as described below. Q-divisors ∆ ≥ 0 such that (pe − 1)(KX + ∆) is an integral Weil divisor linearly equivalent to 0   p−1-LINEAR MAPS IN ALGEBRA AND GEOMETRY 13 Suppose that ϕ : F e ∗ and obtain: F e ∗ OX−−−−→ F e ∗ ϕ : F 2e F e ∗ OX −→ OX. We use ϕ2 to denote this map (note if we view ϕ as an honest p−e linear map, then this is really just ϕ composed with itself). More generally, for each n ≥ 1, we obtain maps ∗ OX −→ OX is an OX -linear map. We apply the functor F e ∗ OX . Composing this with ϕ we obtain ϕ ◦ (F e ∗ ϕ) : F 2e ∗ OX (14) in the same way. ϕn : F ne ∗ OX −→ OX Lemma 4.0.1. [Sch09, Theorem 3.11(e)] Suppose that X is a normal variety. Then the map ϕ ∈ HomOX (F e ∗ OX , OX ) induces the same Q-divisor ∆ via (13) as does the map ϕn ∈ HomOX (F ne ∗ OX , OX ) for any n ≥ 1. Proof. The divisor section correspondence is determined in codimension 1, and so we may assume that X = Spec R where (R, m) is a DVR with m = hri. We will simply verify the claim in the Lemma for n = 2 and leave the general case to the reader Exercise 4.3. Now, since R is regular (and so Gorenstein) and local, KX ∼ 0. Thus HomOX (F e ∗ OX , OX ) ∼= Γ(X, F e ∗ OX ((1 − pe)KX )) ∼= F e ∗ R, ∗ OX , OX ) corresponding to F e we fix Φ ∈ HomOX (F e DΦ = 0 (note that the section 1 doesn't vanish anywhere). It is an exercise left to the reader that Φ2 = Φ ◦ (F e ∗ -module, Exercise 4.2. This is the key point though! Now consider ϕ(F e ∗ ) = Φ(F e ∗ ura · F 2e immediately that Dϕ = a div(r) and so ∆ϕ = a pe−1 div(r). Now we consider ϕ2. We observe that ∗ 1. In other words we pick Φ such that, ∗ Φ) generates HomR(F 2e ∗ R, R) as an ) for some unit u ∈ R and integer a ≥ 0. It follows ) = Φ(F e ϕ2(F 2e ∗ Thus Dϕ2 = a(pe +1) div(r) and so that ∆ϕ2 = a(pe+1) ∗ uraΦ(F e )) = Φ(F e ∗ Φ(F e ∗ ura · ∗ upe+1ra(pe+1) · p2e−1 div(r) = a )) = Φ2(F 2e ∗ upe+1ra(pe+1) · ). pe−1 div(r) = ∆ϕ as desired. (cid:3) Therefore, we obtain a bijection: (15)  nonzero ϕ ∈ HomOX (F e ∗ OX , OX ) as e ≥ 0 varies  , relation generated by multiplication by units in Γ(X, OX ) and by composition in (14) ←→   Q-divisors ∆ ≥ 0 such that n(KX + ∆) ∼ 0 for some n > 0 with p not dividing n.  Here we notice that (pe − 1)(KX + ∆) ∼ 0 for some e > 0 is equivalent to requiring that n(KX + ∆) ∼ 0 for some n > 0 which is not divisible by p, Exercise 4.5. Example 4.0.2. Consider X = An = Spec k[x1, . . . , xn] = Spec R over a perfect field k. Con- sider Φ : F e ∗ R −→ R defined by the following action on monomials Φ(cid:16)F e ∗ (xλ1 1 · · · xλn n )(cid:17) = ∗ (xpe−1 λ1−(pe−1) pe x 1 λn−(pe−1) pe · · · x n 0, , if all λi−(pe−1) pe ∈ Z otherwise In other words, Φ sends F e sis of F e HomR(F e ∗ R over R to zero. We already saw in Example 3.0.5 that Φ : F e ∗ R, R) as an F e ∗ R-module (at least when e = 1, but the general case is no different). ) to 1 and all other elements of the obvious ba- ∗ R −→ R generates n 1 · · · xpe−1 14 MANUEL BLICKLE AND KARL SCHWEDE But then it immediately follows that the divisor DΦ is the zero divisor. In particular, DΦ corresponds to the element in HomR(F e ∗ R, R) ∼= F e ∗ R that doesn't vanish anywhere. ∗ OX, OX ). In this subsection, we generalize this to maps ϕ ∈ HomOX (F e ∗ 4.1. A generalization with line bundles. Previously we considered nonzero maps ϕ ∈ HomOX (F e L , OX ) where L is a line bundle on X. This generality actually simplifies some of the statements con- sidered in the previous section. Indeed, just as in (11), it is easy to see that for a smooth variety X HomOX (F e ∗ L , OX ) ∼= F e ∗ L −1((1 − pe)KX ) Just as before, this extends to normal varieties as well. Therefore for any line bundle on a normal variety X, we have a bijection of sets. multiplication by units in Γ(X, OX )  ←→ L , OX ) induce Q-divisors ∆ϕ = 1 effective Weil divisors D such that OX (D) ∼= L −1((1 − pe)KX ) pe−1 D such that OX((pe − .  (16) (cid:26) nonzero ϕ ∈ HomOX (F e ∗ L , OX ) (cid:27), Thus just as before, ϕ ∈ HomOX (F e ∗ 1)(KX + ∆)) ∼= L −1. Definition 4.1.1. Given ϕ : F e ∗ as above. L −→ OX , we use ∆ϕ to denote the Q-divisor associated to ϕ Finally, consider the data of a line bundle L and an OX-linear map ϕ : F e ∗ L −→ OX . We will compose ϕ with itself in the following way. We tensor ϕ with L and use the projection formula to obtain: ∗ (L pe+1) −→ L F e pushing forward by F e ∗ we obtain ∗ (L pe+1) −→ F e F 2e ∗ L . Composing with ϕ again we obtain a map ∗ (L pe+1) −→ OX F 2e which we denote by ϕ2. Continuing in this way, we obtain maps ∗ (L p(n−1)e+···+pe+1) −→ OX ϕn : F ne (17) for all n ≥ 1. It is then straightforward to verify that: Lemma 4.1.2. The Q-divisor ∆ϕ induced by ϕ : F e ∗ induced by ϕn : F ne ∗ (L p(n−1)e+···+pe+1) −→ OX. Proof. Left as an exercise to the reader Exercise 4.7. L −→ OX is equal to the Q-divisor ∆ϕn (cid:3) In other words, forming the Q-divisor ∆ = 1 pe−1 D normalizes the divisor with respect to self composition just as in the case that L = OX. Given two line bundles L , M , we declare maps ϕ : F e ∗ equivalent if there exists a commutative diagram: L −→ OX and ψ : F e ∗ M −→ OX F e ∗ L α / F e ∗ M ϕ OX ψ / OX id   /   / maps ϕ : F e ∗ Line bundles L and L −→ OX modulo equivalence   ←→ Effective Q-divisors ∆ OX ((pe − 1)(KX + ∆)) ∼= L −1  such that . (cid:3) p−1-LINEAR MAPS IN ALGEBRA AND GEOMETRY 15 where α is an isomorphism. We also declare ϕ and ϕn to be equivalent. These relations generate an equivalence relation ∼ on pairs (L , ϕ : F e ∗ L −→ OX). Theorem 4.1.3. For a normal variety X over a field of characteristic p > 0, there is a bijection of sets Proof. Left to the reader Exercise* 4.8. We compute a final example. Example 4.1.4. Set X = Pn 1)(p − 1)). Then k and consider the line bundle L = OX ((1 − p)KX ) = OX((n + H omOX (F∗L , OX ) = H omOX (F∗OX ((1 − p)KX ), OX ) = H omOX (F∗OX (KX ), OX (KX )) = F∗ H omOX (OX (KX ), OX (KX )) = F∗OX . In particular, there is only one non-zero element ϕ ∈ HomOX (F∗L , OX ) up to scaling by elements of k. In particular, it follows that Dϕ = ∆ϕ = 0 since a non-zero global section of F∗OX doesn't vanish anywhere. On the affine charts, this element is easily seen to coincide with the map described in Example 4.0.2 (at least for e = 1). On the other hand, there is an obvious map ψ : F∗OX −→ OX defined by the rule ψ(F∗y) = y1/p, for y ∈ Γ(U, OX ), if y1/p ∈ Γ(U, OX ) and ψ(F∗y) = 0 otherwise. It is an exercise left to the reader that Dψ = (p − 1)F where F is the union of the various coordinate axes in Pn. For example, if n = 2 and X = Proj k[x, y, z], then F = V (xyz). 4.2. Exercises. Exercise 4.1 (Grothendieck duality for a finite map). Suppose that R ⊆ S is a finite inclusion of Cohen-Macaulay local rings and M is an S-module. Grothendieck duality for this inclusion says that there is an isomorphism of S-modules: HomR(M, ωR) ∼= HomS(M, ωS). Here ωR and ωS are canonical modules for R and S respectively. Verify that this is an easy consequence of the formula HomR(S, ωR) = ωS, a formula which was given to you (8). Exercise 4.2. Suppose that R is a ring and S is an R-algebra such that HomR(S, R) ∼= S as S-modules. Suppose that M is any S-module and prove that the natural map: HomS(M, S) × HomR(S, R) −→ HomR(M, R) defined by (ψ, ϕ) 7→ ϕ ◦ ψ is surjective. In particular, every map in HomR(M, R) can be factored through a map in HomR(S, R). A solution can be found in [Kun86, Appendix F] Exercise 4.3. Prove the general case of Lemma 4.0.1. Exercise 4.4. Suppose that R ⊆ S is a finite extension of Gorenstein local rings. Prove that HomR(S, R) is a rank-1 free S-module. Conclude that if R is Gorenstein and local, HomR(F e ∗ R, R) is isomorphic to F e ∗ R as an F e ∗ R-module. Hint: Since R is Gorenstein and local (semi-local is good enough), ωR ∼= R. 16 MANUEL BLICKLE AND KARL SCHWEDE Exercise 4.5. Suppose we are given an integer n > 0 such that p does not divide n, prove that n(pe − 1) for some integer e > 0. Conclude that (pe − 1)(KX + ∆) ∼ 0 for some e > 0 if and only if n(KX + ∆) ∼ 0 for some n > 0 which is not divisible by p. Exercise 4.6. Compute Dψ and ∆ψ where ψ is as in Example 4.1.4. Exercise 4.7. Prove Lemma 4.1.2. See [Sch09, Theorem 3.11(e)]. Exercise* 4.8. Prove Theorem 4.1.3. Exercise 4.9. Suppose that X is a smooth (or Gorenstein) variety and T : F∗ωX −→ ωX is the trace map as described in Section 3.3. By twisting by −KX and reflexifying, we obtain a map Φ : F∗OX ((1 − p)KX ) −→ OX . Prove that Φ corresponds to the zero divisor by (16). Exercise 4.10. A normal variety X is called Q-Gorenstein if OX (nKX) is a line bundle for some n > 0 (in other words, nKX is Cartier). Note that we do not require Q-Gorenstein varieties to be Cohen-Macaulay. In this case, the index of KX is the smallest n > 0 such that nKX is a Cartier divisor. Suppose that X is Q-Gorenstein with index not divisible by p. Suppose that R = OX,x is the stalk of R at some point x ∈ X. Prove that we have an isomorphism of R-modules, ∗ R ∼= HomR(F e F e Exercise 4.11. Suppose that R is a normal domain and that ϕ : F e corresponding to a divisor ∆ϕ as in Definition 4.1.1. Fix a non-zero g ∈ R. Form a new map ∗ R, R), for all sufficiently divisible e. ∗ R −→ R is an R-linear map Prove that ∆ψ = ∆ϕ + 1 pe−1 div(g). ψ(F e ∗ ) = ϕ(F e ∗ (g · )). Exercise* 4.12. Suppose that ϕ : F e ∗ Form the twisted composition ϕ ◦ (F e L . Now pushforward by F e L −→ OX and ψ : F f ∗ ψ′) as follows. Twist ψ by L to get ψ′ : F f ∗ M −→ OX are two OX -linear maps. ) −→ ∗ (M ⊗ L pf ∗ and compose with ϕ and obtain: ϕ ◦ (F e ∗ ψ′) : F f +e ∗ (M ⊗ L pf ) ∗ ψ′ F e −−−→ F e ∗ L ϕ −→ OX . Find a relation between ∆ϕ, ∆ψ and ∆ϕ◦(F e Definition 4.1.1. For a solution, see the proof of [SS10, Lemma 4.9(i)]. ∗ ψ′) where the ∆ are Q-divisors defined as in Exercise* 4.13 (Non-effective divisors). Fix a line bundle L on a variety X. There is a bijection between non-zero elements of HomOX (F e L , K (X)) and (not necessarily effective) Weil divisors ∗ D such that OX (D) ∼= L −1((1 − pe)KX ). Indeed, suppose that ϕ ∈ HomOX (F e ∗ L , K (X)). Then, working locally if needed, for some L ((1 − pe)E)) ⊆ OX . Set ψ : L ((1 − pe)E) −→ OX to be the restriction map. Then ψ induces a divisor Dψ > 0. Set sufficiently large Cartier divisor E ≥ 0, we have that ϕ(F e ∗ F e ∗ Dϕ = Dψ + (1 − pe)E and prove that Dϕ is independent of the choice of E. 5. Frobenius splittings In this section we give a brief introduction to Frobenius splittings. A more complete treatment can be found in [BK05, Chapter 1]. Suppose that X is a scheme over a perfect field of characteristic p > 0. Definition 5.0.1. We say that X is Frobenius split (or F -split) if the map splits as a map of OX -modules. In this case the splitting map ϕ : F∗OX −→ OX is called a Frobenius splitting. Of course, there may be multiple different Frobenius splittings ϕ ∈ HomOX (F∗OX, OX ). OX −→ F∗OX p−1-LINEAR MAPS IN ALGEBRA AND GEOMETRY 17 Likewise, we say that a ring R is Frobenius split (or F -pure) if the map R −→ F∗R splits as a map of R-modules. A scheme X is said to be F -pure (or locally F -split) if every point x ∈ X has a neighborhood which is F -split. Remark 5.0.2. Frobenius split varieties were formally introduced in [MR85] (also see [RR85]), although very closely related concepts were studied in [PS73, HR76, HS77, Hab80]. Indeed, Frobenius split affine varieties (i.e. rings) had been heavily studied by Hochster and his students in the 1970s and 1980s cf. [Fed83]. We shall see below that every regular variety is F -pure Proposition 5.1.2 but not every regular variety is F -split Lemma 5.2.2. Lemma 5.0.3. A variety X is Frobenius split if and only if (a) the e-iterated Frobenius OX −→ F e (b) the e-iterated Frobenius OX −→ F e ∗ OX splits for some e, or ∗ OX splits for all e. Proof. This is left as an exercise to the reader, see Exercise 5.1. (cid:3) Suppose that X is a variety, we will look for Frobenius splittings inside HomOX (F e Indeed, notice that for any c ∈ Γ(X, OX ), we have a map HomOX (F e defined by evaluation at c, in other words, ϕ 7→ ϕ(F e ∗ c). Now we observe that: ∗ OX , OX ). ∗ OX , OX ) −→ Γ(X, OX ) Lemma 5.0.4. A variety X is Frobenius split if and only if the evaluation-at-1 map HomOX (F e ∗ OX , OX ) −→ Γ(X, OX ) is surjective. Proof. Left as an exercise to the reader in Exercise 5.2 below. (cid:3) Finally, we observe that a normal X is Frobenius split if and only if the regular locus of X is Frobenius split. Lemma 5.0.5. Suppose that X is normal and U ⊆ X is the regular locus. Then X is Frobenius split if and only if U is Frobenius split. Proof. The natural restriction map HomOX (F e ∗ OU , OU ) is an isomor- phism since X \ U has codimension ≥ 2 and the H om sheaves are reflexive. See Appendix A and [BK05, Lemma 1.1.7] for additional discussion. (cid:3) ∗ OX , OX ) −→ HomOU (F e 5.1. Local properties of Frobenius split varieties. The easiest property to prove about Frobenius split varieties is that they are reduced. Lemma 5.1.1. Suppose that a scheme X is F -pure, then X is reduced. Proof. Without loss of generality we may assume that X = Spec R is affine and Frobenius split. Suppose that x ∈ R is such that xn = 0. Then xpe = 0 for some e > 0 (where p is the characteristic of R). Therefore x = xϕ(F e (cid:3) ) = ϕ(F e ∗ 0) = 0. ∗ 1) = ϕ(F e ∗ xpe First we identify some Frobenius split varieties. Proposition 5.1.2 (Regular affine varieties are Frobenius split). Suppose that X = Spec R is a regular affine variety. Then X is Frobenius split. 18 MANUEL BLICKLE AND KARL SCHWEDE Proof. We prove the result for Rm = OX,x, the stalk of X at a closed point x ∈ X. The global case is Exercise 5.6. Let R denote the completion of Rm at the maximal ideal m. Now consider the evaluation-at-1 map Φ : HomRm(F e ∗ Rm, Rm) −→ Rm. Tensoring with R gives us a map Φ : Hom R(F e ∗ R, R) ∼= HomRm(F e ∗ Rm, Rm) ⊗Rm R −→ Rm ⊗Rm R ∼= R. Here we have used Exercise 2.6. Note that by the Cohen-structure theorem, [Mat89, Theorem 28.3], we have that R = kJx1, . . . , xnK. It follows then from the argument of Exercise 2.1 that R. In particular, F e ∗ Φ is surjective. But therefore Φ is surjective as well since tensoring with R is faithfully flat. Thus by Lemma 5.0.4, we are done. (cid:3) R is free as an R-module and in particular, that there is a splitting of R −→ F e ∗ Of course, not all Frobenius split varieties are regular. Lemma 5.1.3 (Simple normal crossings are F -split). The ring R = k[x1, . . . , xn]/hx1 · x2 · · · xni = S/J is Frobenius split. 1 . . . xλn Proof. Observe we have an "obvious" Frobenius splitting ϕ : F e coming from Exercise 2.1, which sends the basis element corresponding to F e all the other basis elements xλ1 ideal hx1 · x2 · · · xni = J. Consider any monomial in xα = xα1 ϕ(F e with each βi ≥ 1. Therefore, ϕ(F e we have that ϕ(F e ∗ k[x1, . . . , xn] −→ k[x1, . . . , xn] ∗ 1 to 1 and sends n to 0. We want to consider what this map does to the n ∈ hx1 · x2 · · · xni = J. Then 1 · · · xβn n ∗ xα) ∈ J. Since every element of J is a sum of such monomials, 1 · · · xαn ∗ xα) 6= 0 if and only if peαi for each i. In particular, this means that ϕ(F e ∗ xα) = xβ1 ∗ J) ⊆ J. But now consider the commutative diagram: (18) ϕJ ϕ F e ∗ J F e ∗ R J / R F e ∗ (R/J) ϕ/J / R/J Since ϕ sends 1 to 1, so does ϕ/J. (cid:3) In the next section, we will introduce a highly effective tool, based upon similar analysis, which can be used to test whether an affine variety is Frobenius split -- Fedder's criterion. Definition 5.1.4. Suppose that ϕ : F e J ⊆ OX is called compatibly (ϕ-)split if ϕ(F e also say that Y is compatibly (ϕ-)split. ∗ OX −→ OX is a Frobenius splitting, then an ideal sheaf ∗ J) ⊆ J. If the subscheme Y = V (J) ⊆ X, then we Note that in Lemma 5.1.3, we showed that the coordinate hyperplanes were compatibly split with the obvious Frobenius splitting on X = Spec k[x1, . . . , xn]. Indeed, consider the following proposition: Proposition 5.1.5 (Properties of compatibly split varieties). Suppose that ϕ : F e is a Frobenius splitting. Then: ∗ OX −→ OX (a) If J ⊆ OX is compatibly ϕ-split, then V (J) is Frobenius split as well. In particular, J is a radical ideal. (b) If J ⊆ OX is compatibly ϕ-split, then ϕ(F e ∗ J) = J (instead of just contained in). / /  _   _       /     / p−1-LINEAR MAPS IN ALGEBRA AND GEOMETRY 19 (c) If I, J ⊆ OX is compatibly ϕ-split, then so are I + J and I ∩ J. (d) If Q is a minimal prime over J, then Q is also compatibly ϕ-split. (e) If I ⊆ OX is compatibly ϕ-split, then so is I : K for any ideal sheaf K ⊆ OX . (f ) A prime ideal sheaf Q is compatibly ϕ-split if and only if Q · OX,Q is compatibly ϕQ split where ϕQ is the map induced on the stalk ϕQ : F e ∗ OX,Q −→ OX,Q. Proof. This is left as an exercise to the reader in Exercise 5.9. (cid:3) Remark 5.1.6. Suppose that ϕ : F e ∗ OX −→ OX is a Frobenius splitting. It is easy to see that a sort of converse to Proposition 5.1.5(a) holds. In particular, suppose there is a commutative diagram F e ∗ OX ϕ / OX F e ∗ (OX /J) ϕ/J / OX /J then J is ϕ-compatibly split (simply take the kernel of the vertical arrows). One important point about Frobenius splittings are that compatibly split subvarieties inter- sect normally. In particular: Corollary 5.1.7. If ϕ : F e then I + J is a radical ideal. ∗ OX −→ OX is a Frobenius splitting, if I and J are compatibly ϕ-split, Proof. Combine properties (a) and (c) from Proposition 5.1.5. (cid:3) Also see Exercise 5.3 below. 5.2. Global properties of Frobenius split varieties. Now we turn to projective (or more generally complete) Frobenius split varieties. First we introduce another definition. Definition 5.2.1. Suppose that D is an effective Weil divisor on a normal variety X. Then we say that X is e-Frobenius split relative to D if the composition: OX −→ F e ∗ OX ֒→ F e ∗ (OX (D)) is split. Notice that if X is e-iterated Frobenius split relative to D, then X is Frobenius split. We mentioned earlier that regular affine varieties are Frobenius split Proposition 5.1.2, but not every smooth projective variety is Frobenius split. We prove that now. Lemma 5.2.2. If X is proper, Frobenius split and normal, then H 0(X, OX (−nKX )) 6= 0 for some n > 0. In particular X is not of general type. Even more, if X is e-Frobenius split relative to an ample divisor A, then −KX is big. Proof. The fact that X is Frobenius split implies that there is some non-zero element ϕ ∈ HomX (F e ∗ OX((1 − ∗ OX ((1 − pe)KX) is isomorphic to OX((1 − pe)KX ) as an Abelian group pe)KX )) 6= 0. But F e and so the result follows for n = (pe − 1). ∗ OX ((1 − pe)KX )) by Section 4. In particular, H 0(X, F e ∗ OX , OX ) ∼= H 0(X, F e For the second statement, we notice that we have a section ϕ ∈ HomX (F e ∗ OX (D), OX ) ∼= ∗ OX((1 − pe)KX − A)) and so there is an effective divisor H ∼ (1 − pe)KX − A and (cid:3) H 0(X, F e thus (1 − pe)KX ∼ A + H = "ample + effective" and so KX is big5. 5On a projective variety X, you can take the definition of big to be a divisor which has a multiple which is linearly equivalent to an ample divisor plus an effective divisor [Laz04a, Corollary 2.2.7].     /     / 20 MANUEL BLICKLE AND KARL SCHWEDE Our next goal is to prove vanishing theorems for Frobenius split varieties. First however, we need the following Lemma. Lemma 5.2.3. If X is e-Frobenius split relative to D, then for any integer n > 0, X is ne- Frobenius split relative to (p(n−1)e + · · · + pe + 1)D. Proof. Suppose that OX −→ F e this with D, taking the reflexification of the sheaves, and applying the functor F e splitting ϕ −→ OX is the Frobenius splitting. By tensoring ∗ , we obtain a ∗ OX −→ F e ∗ OX (D) F e ∗ (OX (D)) −→ F e ∗ (OX (peD)) −→ F 2e ∗ (OX (D + peD)) ∗ ϕ(D) F e −−−−−→ F e ∗ (OX (D)). But now composing with Frobenius and ϕ on the left and right sides respectively, we obtain our desired splitting OX −→ F e ∗ (OX (peD)) −→ F 2e Continuing in this way yields the desired result. ∗ (OX (D)) −→ F e ∗ (OX (D + peD)) ∗ ϕ(D) F e −−−−−→ F e ∗ (OX (D)) ϕ −→ OX. (cid:3) Theorem 5.2.4 (Vanishing Theorems for Frobenius split varieties). Suppose that X is a pro- jective Frobenius split variety. Then: (a) H i(X, L ) = 0 for any ample line bundle L and any i > 0. (b) H i(X, L ⊗ ωX) = 0 for any ample line bundle L and any i > 0. (c) If X is normal and e-Frobenius split relative to an ample Cartier divisor D, then we have H i(X, L ) = 0 for any nef line bundle L and any i > 0. (d) If X is normal and e-Frobenius split relative to an ample Cartier divisor D such that X \ D is regular, then H i(X, L ⊗ ωX) = 0 for any big and nef line bundle L and any i > 0. Proof. For (a), notice that we have a splitting of L ∼= OX ⊗ L −→ (F e Thus H i(X, L ) ֒→ H i(X, F e ∗ Abelian groups, and the latter vanishes for i > 0 and e ≫ 0 by Serre vanishing. ) injects. On the other hand H i(X, F e ∗ ∗ OX ) ⊗ L ∼= F e ∗ ) ∼= H i(X, L pe L pe L pe L pe . ) as For (b), notice that an application of H omOX ( , ωX) to the splitting OX −→ F e ∗ OX −→ OX induces a splitting: Twisting by L and applying the projection formula gives us ωX T ←− F e ∗ ωX ←֓ ωX. Taking cohomology for i > 0 we obtain maps ωX ⊗ L T ←− F e ∗ (ωX ⊗ L pe ) ←֓ ωX ⊗ L . H i(X, ωX ⊗ L ) T ←− H i(X, F e ∗ (ωX ⊗ L pe )) ←֓ H i(X, ωX ⊗ L ) whose composition is an isomorphism. But the middle term vanishes by Serre vanishing since we may take e ≫ 0. For (c), we first notice that by using Lemma 5.2.3 we may assume that D is as ample as we wish (at the expense of increasing e). Thus, using the same strategy as in (a), it is sufficient to prove that H i(X, OX (D) ⊗ L pe ) = 0 for all i > 0. But this follows from Fujita's vanishing theorem [Fuj83]. Part (d) is left as a somewhat involved exercise to the reader Exercise* 5.10. (cid:3) Finally, we notice that sections on Frobenius split subvarieties often extend to sections on the ambient spaces. p−1-LINEAR MAPS IN ALGEBRA AND GEOMETRY 21 Theorem 5.2.5. Suppose that Y ⊆ X is compatibly Frobenius split. Then the natural maps: are surjective for any ample line bundle L . H 0(X, L ) −→ H 0(Y, L Y ) Proof. By composition of the Frobenius splitting with itself, we have the following diagram for any e > 0. H 0(X, F e ∗ (L pe )) H 0(X, L ) β α / H 0(X, F e ∗ (L pe Y )) H 1(X, F e ∗ (IY ⊗ L pe )) = 0 / H 0(X, L Y ) Note we have the top-right vanishing by Serre vanishing which implies that β is surjective. The vertical maps are surjective because they are obtained from twisting the Frobenius splitting F e ∗ OX −→ OX by L . The diagram then implies that α is surjective, this completes the proof. (cid:3) 5.3. Tools for proving proper varieties are Frobenius split. There are two common tools for proving that proper varieties are Frobenius split. The first involves a study of the singularities of sections of H 0(X, OX ((1 − pe)KX )). The second is a general fact that images of Frobenius split varieties often remain Frobenius split. In many applications, these tools are combined. Theorem 5.3.1. [MR85] [BK05, Section 1.3] Suppose X is a proper normal d-dimensional variety of finite type over an algebraically closed field of characteristic p > 0. Further suppose that there is an effective divisor D, linearly equivalent to (1 − pe)KX for some e, that satisfies the following condition: ◦ There exists a smooth point x ∈ X and divisors D1, D2, . . . , Dd intersecting in a simple normal crossings divisor at x ∈ X such that D = (pe − 1)D1 + · · · + (pe − 1)Dd + G for some effective divisor G not passing through x ∈ X. Then X is Frobenius split by a map ϕ : F e ∗ OX −→ OX which corresponds to D as in (12). Proof. There are two main ideas in this proof. (a) D corresponds to some map, ϕ : F e ∗ OX −→ OX by (12). Thus ϕ(F e ∗ 1) = λ ∈ H 0(X, OX ) = k is a constant. If we can show that λ 6= 0, then by rescaling ϕ we are done. (b) The value of ϕ(F e ∗ 1) can be detected at any point. In particular, we can try to compute it at the stalk of x ∈ X. For simplicity, we denote the stalk at x by R := OX,x and we use m to denote the maximal ideal. Fix ϕ corresponding to D as in (12) and consider ϕx : F e ∗ R −→ R. Suppose that DSpec R = V (f pe−1 1 · · · f pe−1 d ) = V (f ) where the fi are the local equations for Di near x. factored as: Set bR to be the completion of R = OX,x. We know that ϕ corresponds to D, so it can be ∗ OX ((1 − pe)KX − D) ֒→ F e F e ∗ OX((1 − pe)KX ) −→ OX. Taking the completion of this factorization, we obtain: ·(F e ∗ f ) F e ∗bR / F e ∗bR bϕ ψ / bR     / / / / /     / 7 7 7 7   / / 22 MANUEL BLICKLE AND KARL SCHWEDE By construction, ψ, viewed as an element of M = Hom(F e (use Exercise 4.9). On the other hand, bR = kJf1, . . . , fdK and so the map Ψ : F e to 1 and the other basis monomials {f a1 f pe−1 1 also generates M as an F e 1 · · · f ad · · · f pe−1 d ∗bR,bR), generates M as an F e ∗bR-module ∗bR −→ bR which sends f = )) for some invertible element c ∈ bR. But notice that c is 6= f 0 ≤ ai ≤ pe − 1} to zero d invertible, so it has a non-zero constant term c0 ∈ k where c = c0 + c′, c′ ∈ hf1, . . . , fdi bR. Thus ∗bR-module by Example 4.0.2. ) = Ψ(F e ∗ (c · It follows that ψ(F e ∗ = bϕ(F e λ = ϕx(F e ∗ 1) ∗ 1) ∗ f ) ∗ (c · f )) ∗ (c0 · f )) + Ψ(F e ∗ (c′ · f )). = ψ(F e = Ψ(F e = Ψ(F e = c1/pe 0 + Ψ(F e ∗ (c′ · f )) But Ψ(F e c1/pe 0 + Ψ(F e desired. ∗ (c′ · f )) ∈ hf1, . . . , fdi bR by our choice of Ψ (note that c′ · f ∈ hf pe ∗ (c′ · f )) = λ ∈ k is a constant, we see that Ψ(F e 1 , . . . , f pe ∗ (c′ · f )) = 0. Thus λ = c1/pe d i). Since 6= 0 as (cid:3) 0 Remark 5.3.2. A more general, simpler and more conceptual version of the above result is described in Exercise 6.12 in the next section. We lack the language to describe it here however. Now we study the behavior of Frobenius splittings under maps between varieties. We will study some complementary constructions later in Section 7. Theorem 5.3.3. [HR76, MR85] Suppose that π : Y −→ X is a map of varieties such that OX −→ π∗OY splits as a map of OX -modules (for example, if π∗OY = OX ). Then if Y is Frobenius split, so is X. Before proving the theorem, we point out just how common the condition that OX −→ π∗OY splits is. Indeed, if π : Y −→ X is a proper surjective map between normal varieties with connected fibers, then π∗OY = OX . Alternately, if π : Y −→ X is proper, dominant, generically finite, Y and X are normal, and p does not divide [K (Y ) : K (X)] = n, then the normalized field trace 1 Proof of Theorem 5.3.3. Set ϕ : F e π∗OY −→ OX to be the splitting of i : OX −→ π∗OY . Pushing down ϕ we obtain: n Tr : K (Y ) −→ K (X) restricts to a map π∗OY −→ OX which sends 1 to 1. ∗ OY −→ OY to be the Frobenius splitting of Y and fix α : Now, we simply form the composition: (π∗ϕ) : π∗F e ∗ OY −→ π∗OY ∗ OY By chasing through the composition, we see that F e ∗ π∗OY = π∗F e ∗ OX F e F e ∗ i ֒−−→ F e π∗ϕ −−→ π∗OY α −→ OX . ∗ 1 is sent to 1 and that X is F -split. (cid:3) 5.4. Exercises. Exercise 5.1. Prove Lemma 5.0.3. Hint: Compose Frobenius and Frobenius splittings by using the functor F e ∗ . Exercise 5.2. Prove Lemma 5.0.4. Exercise 5.3. A domain R containing a field of characteristic p > 0 is said to be weakly normal if any r ∈ K(R) satisfying rp ∈ R also satisfies r ∈ R as well, see [Yan85, Lemma 3] and [Vit11, Section 3]. Show that any F -pure/split R is weakly normal. You can find a solution in [BK05, Proposition 1.2.5], cf. [HR76, Proposition 5.31]. p−1-LINEAR MAPS IN ALGEBRA AND GEOMETRY 23 Exercise 5.4. Suppose that X is Frobenius split relative to a Cartier divisor D such that X \ D is Cohen-Macaulay. Prove that X is Cohen-Macaulay. Hint: Working locally we may assume that X = Spec R and D = V (f ). Fix a maximal ideal m ∈ Spec R and consider the composition H i ∗ R) recalling that a variety can be proven to be Cohen-Macaulay by examining its local cohomology modules as in [Har77, Chapter III, Exercises 3.3 and 3.4]. m(R) −→ H i m(F e m(F e ∗ R) ∗ f ) ·(F e −−−−→ H i ∗ R ∼= R⊕M as R-modules where M is some arbitrary R-module. Exercise 5.5. Suppose that F e Prove that R is Frobenius split. More generally, prove the same result if there is any surjective map F e ∗ R −→ R. Exercise 5.6. Suppose that X = Spec R is an affine variety and suppose that for every maximal ideal m ∈ Spec R, we have that Rm is F -split. Prove that X is F -split. Hint: The given splittings definitely do not glue. However consider the evaluation-at-1 map HomR(F∗R, R) −→ R. Exercise 5.7 (Toric varieties). Suppose that X is a normal toric variety. Consider the map Ψ : F∗OX −→ OX defined as follows. We define Ψ(F∗xλ) =(cid:26) xλ/p 0 if λ/p has integer entries otherwise acting on each affine toric chart (where xλ is a monomial). Show that this induces a Frobenius splitting on X which compatibly splits all the torus invariant divisors. What is the ∆Ψ (as defined as in (13))? Exercise 5.8 (Affine section rings). Suppose that X is a projective algebraic variety with ample line bundle A . Consider S :=Mi∈Z H 0(X, A i), the section right with respect to A . Prove that X is Frobenius split if and only if S is Frobenius split. For additional discussion of related topics, see [Smi00]. Exercise 5.9. Prove Proposition 5.1.5. Hint: For part (a), use a diagram similar to the one in Lemma 5.1.3. For solutions, see [BK05, Chapter 1]. Exercise* 5.10. Prove Theorem 5.2.4(d). Hint: This is somewhat involved. There exists a Cartier divisor B such that L n(−B) is ample for all n ≫ 0 since L is big and nef. For some m ≫ 0, we also know that mD + B is still ample. First show that X is r-Frobenius split relative to mD + B for some integer r ≫ 0 (this is hard). Then notice we have a composition ωX ⊗ L T←− F r ∗ (ωX ⊗ L ) ←֓ F r ∗ (ωX (−B) ⊗ L ) ←֓ F r ∗ (ωX (−mD − B) ⊗ L ) ← ωX which is an isomorphism (we type this with the arrows going backwards to suggest that this arises by duality). Now, by composing the map ωX ← F r ∗ (ωX (−B) ⊗ L ) with itself as in (17), we can obtain the desired vanishing. For a solution, see [SS10, Theorem 6.8]. 6. Frobenius non-splittings Our goal in this section is to develop a theory for p−1-linear maps generalizing the theory of Frobenius split varieties demonstrated in the previous section. First we start with a definition. 24 MANUEL BLICKLE AND KARL SCHWEDE Definition 6.0.1. Suppose that we are given a line bundle L on a variety X. Consider an OX -linear map ϕ : F e L −→ OX . We say that an ideal J is ϕ-compatible if we have that ∗ ϕ(F e ∗ (J · L )) ⊆ J. If Y = V (J) ⊆ X, then we say that Y is ϕ-compatible if J is. For example, if L = OX and ϕ is a Frobenius splitting, then any ϕ-compatibly split ideal is ϕ-compatible. We also have a slight variation on this definition. Definition 6.0.2. Given ∆ corresponding to ϕ : F e L −→ OX as in (16). A subvariety Y ⊆ X ∗ is called an F -pure center of (X, ∆) if Y is ϕ-compatible and ϕη is surjective where η is the generic point of Y . Lemma 6.0.3. If J ⊆ OX is an ideal sheaf, then J is ϕ : F e ∗ if ϕ induces a map ϕY : F e ∗ (L Y ) −→ OY . L −→ OX compatible if and only Proof. Left as an exercise to the reader Exercise 6.1. (cid:3) We explore compatibility after composing maps as in (17). Lemma 6.0.4. Suppose that J ⊆ OX is ϕ : F e ∗ L −→ OX -compatible. Then J is ϕn : F ne ∗ L pe(n−1)+···+1 −→ OX compatible for all n > 0. Conversely, suppose that ϕ is surjective. If J is ϕn-compatible then J is also ϕ-compatible. Proof. The statement is local so we may as well only check this at the stalks OX,x and in particular assume that L ∼= OX,x. The first statement is obvious and will be left to the reader. For the second statement, we sketch the idea of the proof. The first step is to show that any J ⊆ OX,x which is ϕ : F e ∗ OX,x ։ OX,x-compatible is also radical, see Exercise* 6.3. One can then show it is sufficient to verify the statement at the minimal primes of J. In particular, we can assume that J is the maximal ideal of OX,x by localizing. Now then, suppose that J is ϕn compatible but not ϕ-compatible. Then ϕ(F e ∗ J) = OX,x (since otherwise, it would be in the maximal ideal, which coincides with J). But then it is easy to see that ϕ2(F 2e ∗ J)) = OX,x as well. Continuing in this way, we obtain a contradiction. (cid:3) ∗ J) = ϕ(F e ∗ ϕ(F e We also generalize the notion of F -pure to non-Frobenius splittings and to pairs. Definition 6.0.5. Suppose that X is a normal variety and that ∆ is a Q-divisor such that KX + ∆ is a Q-Cartier divisor with index not divisible by p. (†) We say that (X, ∆) is sharply F -pure if the map ϕ : F e ∗ (15) is surjective as a map of OX-modules. L −→ OX, corresponding to ∆ as in If we do not satisfy (†), then we say that (X, ∆) is sharply F -pure if for every point x ∈ X, there exists a neighborhood U of x ∈ X and a divisor ∆U on U such that ∆U ≥ ∆U and such that (U, ∆U ) is sharply F -pure in the above sense. It is an exercise below, Exercise 6.5, that the definition of sharply F -pure above and the definition given in Definition 5.0.1 coincide. 6.1. Global considerations. In this subsection, we briefly demonstrate that some of the global methods from the Frobenius splitting section can still bear fruit, even if the actual vanishing theorems do not hold. Our first goal is to consider a generalization of a proof due to D. Keeler [Kee08] (also inde- pendently obtained by N. Hara [unpublished]). Related results were first proven by [Smi97] and also [Har05]. Before doing that, we recall a Definition and a Lemma. p−1-LINEAR MAPS IN ALGEBRA AND GEOMETRY 25 Definition 6.1.1 (Castelnuovo-Mumford Regularity). [Laz04a, Section 1.8] Suppose that F is a coherent sheaf on a projective variety X and that A is a globally generated ample divisor on X. Then F is called 0-regular (with respect to A ) if H i(X, F ⊗ A −i) = 0 for all i > 0. Lemma 6.1.2 (Mumford's Theorem). [Laz04a, Theorem 1.8.5] If F is 0-regular with respect to a globally generated ample line bundle A , then F is globally generated. Now we are in a position to prove that certain sheaves are globally generated. Theorem 6.1.3. [Kee08, Sch11a] Suppose that ϕ : F e L −→ OX is a surjective OX-linear map ∗ and L is a line bundle. Additionally suppose that A is a globally generated ample line bundle and that M is any other line bundle such that L ⊗ M pe−1 is ample (for example, if L is itself ample, then we may take M = OX ). In this case, the line bundle M ⊗ A dim X is globally generated. Proof. Choose n ≫ 0. Then we have a surjective map: L p(n−1)e+···+pe+1 −→ OX from (17). Twisting by M ⊗ A dim X we obtain a surjective map: ϕn : F ne ∗ ∗ (L p(n−1)e+···+pe+1 ⊗ M pne F ne ⊗ A pne dim X) → M ⊗ A dim X. It is sufficient to show that the left side is globally generated as an OX -module since then the right side is a quotient of a globally generated module and thus globally generated itself. Note it is definitely not sufficient to show that the left side is globally generated as an F ne ∗ OX -module. We will proceed by proving that the left side is 0-regular as an OX -module. Note L p(n−1)e+···+pe+1 ⊗ M pne = (L ⊗ M pe−1)p(n−1)e+···+pe+1 ⊗ M . But now we have H i(cid:16)X, F ne = H i(cid:16)X, F ne = H i(cid:16)X, F ne ∗ (cid:0)(L ⊗ M pe−1)p(n−1)e+···+pe+1 ⊗ M ⊗ A pne dim X(cid:1) ⊗ A −i(cid:17) ∗ (cid:0)(L ⊗ M pe−1)p(n−1)e+···+pe+1 ⊗ M ⊗ A pne(dim X−i)(cid:1)(cid:17) ∗ (cid:0)(L ⊗ M pe−1 ⊗ A (dim X−i)(pe−1))p(n−1)e+···+pe+1 ⊗ (M ⊗ A dim X−i)(cid:1)(cid:17). We already have the vanishing for i > dim X. Now the F ne does not effect the vanishing or non-vanishing of the cohomology since it doesn't change the underlying sheaf of Abelian groups. Therefore, the above cohomology groups vanish by Serre vanishing, since L ⊗ M pe−1 ⊗ A (dim X−i)(pe−1) is ample and each of the finitely many M ⊗ A dim X−i are coherent sheaves. (cid:3) ∗ Example 6.1.4. If X is smooth (or even F -pure), then there is always a surjective map ∗ OX ((1 − pe)KX ) −→ OX . It follows that if M is a divisor such that M − KX is ample, and F e A is any globally generated ample line bundle, then OX (M ) ⊗ A dim X is globally generated. Remark 6.1.5. It is worth pointing out that not only is M ⊗ A dim X globally generated, one even has that it is globally generated by the image of the map H 0(cid:16)X, F ne ∗ (cid:0)L p(n−1)e+···+pe+1 ⊗ M pne ⊗ A pne dim X(cid:1)(cid:17) → H 0(cid:0)X, M ⊗ A dim X(cid:1). This special sub-vector space of global sections also behaves well with respect to restriction to compatible subvarieties as we shall see shortly. Similar arguments to those in the proof Theorem 6.1.3 also yield the following result. 26 MANUEL BLICKLE AND KARL SCHWEDE Proposition 6.1.6. [Smi97, Har05, Kee08] If X is any F -pure variety, A is a globally generated ample line bundle and M is any other ample line bundle then is globally generated. ωX ⊗ A dim X ⊗ M Proof. The proof is left to the reader in Exercise 6.7. (cid:3) Finally, we also remark that compatible ideals also play a special role with regards to lifting of sections. Theorem 6.1.7. Suppose that ϕ : F e L −→ OX is an OX -linear map and that J ⊆ OX is ∗ ϕ-compatible. Set Y = V (J) and set ϕY : F e ∗ (L Y ) −→ OY to be the map ϕ restricted to Y as in Lemma 6.0.3. Suppose that H is a line bundle on X such that H pe−1 ⊗ L is ample and also such that the map induced by ϕY (19) is non-zero for some n ≫ 0. Then H 0(X, H ) 6= 0 as well. Even more, the sections in the image of γ all extend to sections on H 0(X, H ). ∗ ((L p(n−1)e+···+pe+1 ⊗ H pne −→ H 0(Y, H Y ) H 0(cid:0)Y, F ne )Y )(cid:1) γ Before starting the proof, let us note some conditions under which the map γ is non-zero. For example, if L Y = OY and ϕY is a Frobenius splitting, then γ is in fact surjective (for example, if Y is a point and ϕY is non-zero). Alternately, if ϕY is surjective and also H Y = A dim Y ⊗ M where A is a globally generated ample line bundle on Y and M pe−1 ⊗ L Y is ample on Y , then we can apply Theorem 6.1.3. In the case that Y is a curve, see Exercise 6.2 Proof. We fix n ≫ 0, for simplicity of notation set η = p(n−1)e + · · · + pe + 1 and consider the following diagram. / H 0(cid:0)Y, F ne ∗ ((L η ⊗ H pne ϕY )Y )(cid:1) / H 1(cid:0)X, F ne ∗ (J ⊗ L η ⊗ H pne / H 0(Y, H Y ) / H 1(X, J ⊗ H ). However, note that by Serre vanishing since the F ne ∗ (J ⊗ L η ⊗ H pne )(cid:1) = H 1(cid:0)X, F ne ∗ (J ⊗ H ⊗ (L ⊗ H pe−1)η(cid:1) = 0 ∗ does not effect the underlying sheaf of Abelian groups. 6.2. Fedder's Lemma. We now delve into the local theory of p−e-linear maps and in particular state Fedder's Lemma. This is a particularly effective tool for explicitly writing down these maps and also for identifying which of them are surjective. Suppose that S = k[x1, . . . , xn] and R = S/I for some ideal I ⊆ R. The point is that if ∗ S −→ S, which Fedder's Lemma ∗ S, S) as an F e ∗ S- . . . xpe−1 ∗ (xpe−1 ) n R = S/I, then maps ¯ϕ : F e ∗ S −→ S to be the map which generates HomS(F e precisely identifies. Set ΦS : F e module as identified in Example 4.0.2. Recall that ΦS sends the monomial F e to 1 and all other basis monomials to zero. ∗ R −→ R come from maps ϕ : F e 1 Lemma 6.2.1 (Fedder's Lemma). [Fed83, Lemma 1.6] With S ⊇ I, R and ΦS as above, then (cid:26) Maps ϕ ∈ HomS(F e compatible with I ∗ S, S) More generally, there is an isomorphism of S-modules: ∗ ∗ (z · ) = ΦS(F e (cid:27) =(cid:8) ϕ ϕ(F e ∗ (I [pe] : I)(cid:1) · ΦS ∗ R, R) ←→ (cid:0)F e (cid:0)F e ∗ I [pe](cid:1) · ΦS HomR(F e )), for some z ∈ I [pe] : I (cid:9) . H 0(cid:0)X, F ne ϕ ∗ (L η ⊗ H pne(cid:1) H 0(cid:0)X, H ) H 1(cid:0)X, F ne )(cid:1) (cid:3)   /   /   / / p−1-LINEAR MAPS IN ALGEBRA AND GEOMETRY 27 induced by restricting ψ ∈ (F e ∗ (I [pe] : I)) · ΦS ⊆ HomS(F e Finally, for any point q ∈ V (I) ⊆ Spec S, there exists a map ϕ ∈ HomR(F e ∗ S, S) to R = S/I as in Lemma 6.0.3. ∗ R, R) which is [pe]. In other words, R is F -pure in a surjective at q/I ∈ Spec R if and only if I [pe] : I * q neighborhood of q if and only if I [p] : I * q [p]. Proof. There are a lot of statements here. First we notice that any map of the form ϕ(F e ∗ ΦS(F e )) for some z ∈ I [pe] : I is clearly compatible with I since ΦS(F e ∗ (I [pe] : I) · I) ⊆ ΦS(F e ∗ (z · I)) ⊆ ΦS(F e ∗ I [pe]) = I · ΦS(F e ∗ S) = I. ∗ (z · ) = This gives us the containment ⊇ in the first equality. For the other containment, we first prove the following claim. Claim 6.2.2. For ideals I, J ⊆ S we have ΦS(F e ∗ J) ⊆ I if and only if J ⊆ I [pe]. Proof of claim. Certainly the if direction is obvious, so suppose then that ΦS(F e implies that ϕ(F e module. But F e ∗ S is a free S-module of rank pen, so we see that ∗ J) ⊆ I for every ϕ ∈ HomS(F e ∗ S, S) since ΦS generates that set as an F e ∗ J) ⊆ I. This ∗ S- since we could take the ϕ as the various projections. Now, I ⊕ · · · ⊕ I = I · (F e This proves the claim. ∗ S) = F e ∗ I [pe]. (cid:3) F e ∗ J ⊆ I ⊕ · · · ⊕ I pne−times {z } Now we return to the proof of Fedder's Lemma. We observe that if ϕ(F e ∗ )) is I-compatible, then z · I ⊆ I [pe] by the claim, which proves that z ∈ I [pe] : I and so the equality is proven. ) = ΦS(F e ∗ (z · Now we come to the bijection. We certainly have a natural map Λ : (F e ∗ (I [pe] : I)) · ΦS −→ HomR(F e ∗ z) · ) = ΦS((F e ∗ R, R) induced by sending F e map in HomR(F e so we only need to show that this map is surjective. ∗ z first to (F e ∗ R, R) as in Lemma 6.0.3. The kernel of Λ is (F e ∗ z) · ΦS( )) and then second, inducing a ∗ I [pe]) · ΦS by the claim, and Given ϕ ∈ HomR(F e ∗ R, R) = HomS(F e ∗ R, R), consider the following diagram of S-linear maps where the horizontal maps are the canonical surjections: F e ∗ S / F e ∗ (R/I) ∃ψ S ϕ / (R/I) Because F e ∗ S is a free (and so projective) S-module, the dotted map ψ exists. By construction, ψ is compatible with I. By the earlier parts of the theorem, ψ corresponds to a z ∈ I [pe] : I which restricts to ϕ, completing the proof of the bijection. The last part of the theorem is left as an exercise to the reader. (cid:3) Remark 6.2.3 (Regular local rings are fine). The proof given above goes through without change if one assumes that S is a regular local6 ring instead of assuming that S is a polynomial ring. One of the most important corollaries of this is the following. 6or even semilocal   / / /   / / / 28 MANUEL BLICKLE AND KARL SCHWEDE Corollary 6.2.4. Given f ∈ k[x1, . . . , xn] = S, then S/hf i is F -split in a neighborhood of the origin if and only if f p−1 /∈ hxp 1, . . . , xp ni. Proof. Note that S/hf i = R is F -split if and only if there exists a surjective map ϕ ∈ HomR(F e by Exercise 5.5. The result then follows from Fedder's Lemma since hf pi : hf i = hf p−1i. ∗ R, R) (cid:3) We now apply Fedder's Lemma in a number of examples of hypersurface singularities: Example 6.2.5. We consider S to be a polynomial ring in the following examples. Node: Consider the ring S = k[x, y] and R = k[x, y]/hxyi. Then R is F -split near the origin since (xy)p−1 = xp−1yp−1 /∈ hxp, ypi. Cusp: Consider the ring S = k[x, y] and R = k[x, y]/hx3 − y2i. Then we claim that R is not F -split near the origin since (for odd primes). To see this observe that for some constant c (x3 − y2)p−1 = x3(p−1) + · · · + cx3(p−1)/2yp−1 + · · · + x2(p−1) ∈ hxp, ypi. The computation for p = 2 is similar (or follows from the work below). Pinch point: Consider the ring S = k[x, y, z] and R = k[x, y, z]/hxy2 − z2i. If p 6= 2, this is F -split near the origin since (xy2 − z2)p−1 /∈ hxp, yp, zpi, (p−1)/2(cid:1)(−1)(p−1)/2x(p−1)/2yp−1zp−1 + · · · + z(p−1)/2 = xp−1y2(p−1) + · · · +(cid:0) p−1 noting that p does not divide(cid:0) p−1 (p−1)/2(cid:1). Characteristic 2: If R = k[x1, . . . , xn]/hf i and char k = 2, then R is F -split near the ni. In particular, it is immediate that the cusp and the origin if and only if f /∈ hx2 pinch point are also not F -split near the origin in characteristic 2. 1, . . . , x2 Characteristic 3: Just like characteristic 2, if R = k[x1, . . . , xn]/hf i and char k = 3, then R is F -split near the origin if and only if f 2 /∈ hx3 1, . . . , x3 ni. Finally, we point out that complete intersection singularities are nearly as easy to compute as hypersurfaces. Proposition 6.2.6. Suppose that f1, . . . , fm ⊆ hx1, . . . , xni ⊆ k[x1, . . . , xn] = S is a regular sequence.7 Set I = hf1, . . . , fmi. Then (I [pe] : I) = hf pe−1 1 · · · f pe−1 n i + I [pe] In particular, S/I is F -split near the origin m = hx1, . . . , xni if and only if the product for some e > 0. f pe−1 1 · · · f pe−1 n [pe] /∈ m Proof. The containment ⊇ is trivial. The converse direction is left as Exercise 6.15. (cid:3) 7This means that fi is not a zero divisor in S/hf1, . . . , fi−1i for all i > 0. p−1-LINEAR MAPS IN ALGEBRA AND GEOMETRY 29 6.3. Exercises. Exercise 6.1. Prove Lemma 6.0.3. Exercise 6.2. Suppose that C is a smooth curve and that L is a line bundle of degree ≥ 2. Prove that the image of the map H 0(cid:0)X, F e ∗ (ωX ⊗ L pe )(cid:1) −→ H 0(X, ωX ⊗ L ) globally generates ωX ⊗ L for any e ≫ 0. Hint: Mimic the proof in [Har77, Chapter IV, Proposition 3.1]. For a solution, see [Sch11a, Theorem 3.3]. Exercise* 6.3. Consider a map ϕ : F e L −→ OX for some line bundle L and e > 0. Formulate ∗ analogs of the properties from Proposition 5.1.5 and Corollary 5.1.7 for such a map (and ϕ- compatible ideals / subvarieties). Which of these properties hold for all ϕ? Which hold for surjective ϕ? Prove those that do and give counterexamples to those that do not. Some of the answers can be found in [Sch10, Sch09]. Exercise* 6.4. Suppose that X is a Frobenius split normal variety. Suppose that X embeds into Pn as a closed subvariety. Prove that X is compatibly F -split by a Frobenius splitting of Pn if and only if the embedding X ⊆ Pn is projectively normal, cf. [Har77, Chapter II, Exercise 5.14]. Hint: Projective normality can be detected by the difference between the affine cone and the section ring as in Exercise 5.8. Develop then a "graded variant" of Fedder's Lemma that will allow you to prove the result. Exercise 6.5. We can define X to be F -pure if (X, 0) is sharply F -pure in the sense of Definition 6.0.5. Show that this coincides with the definition of F -pure given in Definition 5.0.1. Exercise 6.6. Suppose that L is an ample line bundle on a smooth variety X. Prove that ∗ (ωX ⊗ L mpe )) −→ H 0(X, ωX ⊗ L m) is surjective for all m ≫ 0. For one solution, see H 0(X, F e [Sch11a, Lemma 3.1]. Exercise 6.7. Use the method of Theorem 6.1.3 to prove Proposition 6.1.6. Hint: Dualize a local splitting OU −→ F∗OU −→ OU to obtain a surjective map T : F∗ωU −→ ωU . Use T instead of ϕ in the proof of Theorem 6.1.3. Exercise 6.8. Consider F5[x, y, z] = S and f = x4 + y4 + z4. Consider the map ΦS : F∗S −→ S which sends F∗x4y4z4 to 1 and sends all the other monomials xiyjzk to 0 for 0 ≤ i, j, k ≤ 4 as in Example 4.0.2. Consider the map ϕ : F∗S −→ S defined by )). ) = ΦS(F∗(f 4 · ϕ(F∗ (a) Prove that hf i is ϕ-compatible and let ϕ : F∗R −→ R be the induced map on R = S/hf i as in Lemma 6.0.3. (b)* Set m = hx, y, zi ∈ S. Fix a, b, c ∈ F52 \ F5. Show that J = m ϕ2-compatible. However, show that J is not ϕ-compatible. 2 + hax + by + czi is ∗ (f · ) = ΦS(F e Exercise 6.9. With ΦS as in Section 6.2, fix f ∈ S and consider the map ϕ defined by the rule ϕ(F e )). Show that ϕ is compatible with an ideal J ⊆ S if and only if ∗ f ∈ J [pe] : J. Exercise 6.10. Complete the proof of Fedder's Lemma by proving the following. For any point q ∈ V (I) ⊆ Spec S, there exists a map ϕ ∈ HomR(F e ∗ R, R) which is surjective at q/I ∈ Spec R if and only if I [pe] : I * q [pe]. In other words, R is F -pure in a neighborhood of q if and only if I [pe] : I * q [pe]. Hint: Note that a map ϕq : F e ∗ (Rq) −→ Rq is surjective if and only if Image(ϕq) * qRq. 30 MANUEL BLICKLE AND KARL SCHWEDE Exercise 6.11. Suppose that X = Spec R is a regular ring and ∆ = 1 pe−1 divX(f ) is a Q- divisor on X. Show that (X, ∆) is sharply F -pure near a point m ∈ Spec R = X if and only if f pe−1 /∈ m [pe]. Hint: Use Fedder's lemma in the form of Remark 6.2.3. Exercise 6.12. Suppose that X is a proper variety and that ϕ : F e ∗ OX −→ OX is a map that is compatible with m, the ideal of a closed point x ∈ X. Further suppose that (X, ∆ϕ) is sharply F -pure in a neighborhood of m. Prove that 0 6= ϕ(F e ∗ 1) ∈ k and so in particular X is F -split. This generalizes Theorem 5.3.1 by the following argument. Given a D = (pe − 1)D1 + · · · + (pe − 1)Dd + G and ϕ as in Theorem 5.3.1, set ∆ = 1 pe−1 D. Observe that mx, the maximal ideal of x is ϕ-compatible since each Di is ϕ-compatible, cf. Lemma 5.1.3. Use exercise Exercise 6.11 to conclude that ϕ is surjective in a neighborhood of x ∈ X. Obtain a new proof of Theorem 5.3.1 by combining the above. Exercise 6.13. Suppose that X = Spec kJx, yK where k has characteristic 7 and that ∆ = 1 2 divX (y2 − x3) + 1 2 divX (y). Prove that (X, ∆) is sharply F -pure at the origin m and also that if ϕ corresponds to ∆, then m is ϕ-compatible. 3 divX (x) + 1 Now suppose that Y is a smooth projective variety with a Q-divisor Θ ≥ 0 such that ◦ (p − 1)(KY + Θ) ∼ 0 and, ◦ (Y, Θ) has a point y ∈ Y analytically isomorphic to (X, ∆) above. Show that Y is Frobenius split using Exercise 6.12. Exercise 6.14. Suppose that R is an integral domain with normalization RN in K(R), the field of fractions of R. In this exercise, we will prove that every map ϕ : F e ∗ R −→ R induces an ∗ RN −→ RN which is compatible with the conductor ideal c := AnnR(RN), RN-linear map ϕN : F e an ideal in both R and RN. We do this in two steps. (a) Prove that ϕ is compatible with c (when viewed as an ideal in R). (b) Notice that ϕ induces a map ϕ0 : F e ∗ K(R) −→ K(R) by localization. Prove that ϕ0(RN) ⊆ RN which proves that we can take ϕN = ϕ0RN . Hint: Recall that x ∈ K(R) is integral over R if there exists a non-zero c ∈ R such that cxn ∈ R for all n ≫ 0, see [HS06, Exercise 2.26]. Exercise 6.15. Prove Proposition 6.2.6. Hint: A very easy proof (pointed out to us by Alberto Fernandez Boix), follows from [Har66a, Corollary 1]. Alternately, the ⊇ containment is easy. For the reverse proceed by induction on ∗ S/I, S/I) is a free F e the number of fi. Notice that HomS/I (F e ∗ S/I-module of rank 1. Thus a generator of that module corresponds to an element h ∈ I [pe] : I. For a generalization to Gorenstein rings (instead of just complete intersections), see [Sch09, Corollary 7.5]. Exercise 6.16 (Macaulay2 Fedder's criterion). The following Macaulay2 code, written by Mordechai Katzman and available at http://katzman.staff.shef.ac.uk/FSplitting/ can be quite useful. frobeniusPower=method(); frobeniusPower(Ideal,ZZ) := (I,e) ->( R:=ring I; p:=char R; p−1-LINEAR MAPS IN ALGEBRA AND GEOMETRY 31 local u; local answer; G:=first entries gens I; if (#G==0) then answer=ideal(0 R) else answer=ideal(apply(G, u->u^(p^e))); answer ); This takes an ideal I and raises it to the peth Frobenius power, I 7→ I [pe]. Using this as a starting place, implement within Macaulay2 a method which determines whether a given ring is F -pure near the origin. Check your method against the following examples: (a) R = k[x, y, z]/hxy, xz, yzi in whatever characteristics you feel like. (b) R = k[w, x, y, z]/hxy, z2 + wx2, yzi in characteristic 2 and 3. (c) R = k[x, y, z]/hx3 + y3 + z3i in characteristics 7, 11 and 13. (d) R = k[x, y, z]/hx2 + y3 + z5i in characteristics 2, 3, 5, 7 and 11. Exercise* 6.17. Use Fedder's criterion to determine for which p > 0, the ring k[x, y, z]/hx3 + y3 + z3i is F -pure near the origin. For some related computations, see [Sil09, Chapter V, Section 4]. Exercise* 6.18. If (R, m) is a regular local ring and 0 6= f ∈ m, then the F -pure threshold cm(f ) of f ∈ k[x1, . . . , xn], at the origin m = hx1, . . . , xni, is defined as follows: max{l f l /∈ m [pe]} pe . lim e−→∞ Prove that this limit exists in general and then show that cm(x3 −y2) = 5 for solutions, cf. [TW04] . 6 if p = 7. See [MTW05] 7. Change of variety In this section, we describe how p−e-linear maps change under common change of variety operations. 7.1. Closed subschemes. We have already studied the behavior of p−e-linear maps for sub- schemes extensively. Indeed, suppose that ϕ : F e ∗ R −→ R is an R-linear map which is compatible with an ideal I ⊆ R. Then we have an induced map ϕR/I : F e ∗ (R/I) −→ (R/I). It is natural to ask what the divisor associated to R/I is. Lemma 7.1.1. Suppose that R is a normal Gorenstein local ring, and that D = V (f ) is a normal Cartier divisor on X = Spec R. Fix Φ : F e ∗ R, R) as an F e )). Then ϕ is compatible with D and furthermore, ϕD generates HomR/hf i(F e ∗ R/hf i-module. It follows that the Q-divisor ∆ on D associated to ϕD, as in (13) is the zero divisor. ) = Φ(F e ∗ (R/hf i), R/hf i) as an F e ∗ R −→ R to be map generating HomR(F e ∗ R-module as in Exercise 4.4. Set ϕ(F e ∗ ∗ (f pe−1 · Proof. See Exercise* 7.2. (cid:3) However, things are not always nearly so nice. In particular the divisor associated to ϕD need not always be zero. Example 7.1.2. Consider S = k[x, y, z] with p = chark 6= 2, set R := k[x, y, z]/hxy −z2i and fix D = V (hx, zi). Set ΦS ∈ HomS(F e ∗ S-module generator as in Example 4.0.2. We notice that by Fedder's Lemma, Lemma 6.2.1, that Ψ(F e )) ∗ ∗ R, R) by restriction. Notice that OX (−2nD) = hxni and induces the generator of HomR(F e consider the map ∗ S, S) to be the F e ∗ ((xy − z2)pe−1 · ) = Φ(F e ϕ(F e ∗ ) = Ψ(F e ∗ (x pe−1 2 · )) = Φ(F e ∗ (x pe−1 2 (xy − z2)pe−1 · )). 32 MANUEL BLICKLE AND KARL SCHWEDE If we set X = Spec R, then the induced map ϕX ∈ HomR(F e (pe − 1)D. ∗ R, R) corresponds to the divisor However, it is easy to see that ϕX also is compatible with D. Thus we obtain ϕD. To compute the divisor associated to D, we need only read off the term containing xpe−1zpe−1 in pe−1 2 zpe−1 + · · · + z2(pe−1) )) induces the generator on ∗ k[y] −→ k[y] is the map generating 3(pe−1) pe−1 2 (x )(xy − z2)pe−1 = x Again, the reason this works is because the map ΦS(F e HomOD (F e HomOD (F e ∗ OD, OD). But(cid:0)pe−1 2 (cid:19)xpe−1y ype−1 + · · · +(cid:18)pe − 1 2 (cid:1) 6= 0 mod p and so if ΦD : F e ∗ (xpe−1zpe−1· pe−1 2 pe−1 ∗ OD, OD), then ϕD (which is just ϕX restricted to D) is defined by the rule at least up to multiplication by an element of k. Thus, in the terminology of (13), ϕD(F e ∗ ) = ΦD(F e ∗ y pe−1 2 · ) ∆ϕD = 1 pe − 1 div(y pe−1 2 ) = 1 2 div(y). In particular, in contrast with Lemma 7.1.1, ∆ϕD 6= 0. Theorem 7.1.3 (F -adjunction). If X is a normal variety, ∆ ≥ 0 is a Q-divisor on X such that KX + ∆ is Q-Cartier with index not divisible by p. Suppose that Y is an F -pure center (see Definition 6.0.2) of (X, ∆) and that ϕ corresponds to ∆ as in (15). Then there exists a canonically determined Q-divisor ∆Y ≥ 0 such that: (a) (KY + ∆)Y ∼Q KY + ∆Y (b) (X, ∆) is sharply F -pure near Y if and only if (Y, ∆Y ) is sharply F -pure. Proof. Set ϕY to be the restriction of ϕ to Y as in Lemma 6.0.3. Set ∆Y to be the Q-divisor associated to ϕY as in (15). The first result then follows easily. The second follows since ϕ is surjective near Y if and only if ϕY is surjective. (cid:3) Remark 7.1.4. The previous result should be compared with subadjunction and inversion of adjunction in birational geometry. See for example [Kaw98, Kaw07, Hac12] and [KM98, Chapter 5, Section 4]. 7.2. Birational maps. Suppose that X is a normal variety, L is a line bundle on X and ϕ : F e L −→ OX is an OX -linear map corresponding to the Q-divisor ∆ as in (15). Suppose ∗ π : eX −→ X is a birational map with eX normal. Fix K eX and KX which agree wherever π is an isomorphism. We can write K eX + ∆ eX = π∗(KX + ∆) where now ∆ eX is uniquely determined. Notice that ∆ eX need not be effective. The main result of this section is the following: Lemma 7.2.1. The map ϕ : F e ∗ divisors via Exercise* 4.13, we have that ∆ eϕ = ∆ eX. Even more, using the fact that maps to the fraction field correspond to possibly non-effective L −→ OX induces a map eϕ : F e ∗ (π∗L ) −→ K (eX) where K (eX) is the fraction field sheaf of eX (which we can also identify with the fraction field on X since π is birational). Furthermore, eϕ agrees with ϕ wherever π is an isomorphism. Proof. We construct eϕ as follows. We note that L = OX ((1 − pe)(KX + ∆)) by (15), and so V ⊆ eX, an embedding Γ(V, π∗L ) ⊆ K(eX) = K(X). ∗ K(X) −→ K(X) (note that our embedding of L ⊆ K (X) fixes the isomorphism Lη ∼= K(X)). But we after fixing KX, we obtain an embedding of L ⊆ K (X). In particular, for each affine open set U , we have an embedding Γ(U, L ) ⊆ K(X). But then we also obtain for each affine open set L −→ OX at the generic point η of X, we obtain ϕη : F e Now, by taking the map F e ∗ p−1-LINEAR MAPS IN ALGEBRA AND GEOMETRY 33 have a map ϕeη : F e identify η with the generic pointeη of eX (since they have isomorphic neighborhoods) and so we ∗ K(eX) −→ K(eX). By restricting ϕeη to Γ(V, π∗L ) for each open set V , we ∗ π∗L −→ K (eX). obtain a map eϕ : F e By construction, eϕ agrees with ϕ wherever π is an isomorphism. For the statement ∆ eϕ = ∆ eX we proceed as follows. We notice that ∆ eϕ and ∆ eX already agree wherever π is an isomorphism so that ∆ eϕ − ∆ eX is π-exceptional. Furthermore, by the construction done in Exercise* 4.13 O eX ((1 − pe)(K eX + ∆ eϕ)) ∼= π∗L ∼= π∗OX ((1 − pe)(KX + ∆)). Thus ∆ eϕ ∼Q ∆ eX and so ∆ eϕ − ∆ eX ∼Q 0 is π-exceptional. Therefore ∆ eϕ = ∆ eX as desired, cf. [KM98]. (cid:3) We now come to the definition of log canonical singularities (in arbitrary characteristic). Definition 7.2.2. Suppose that X is a normal variety and that ∆ is a Q-divisor such that KX + ∆ is Q-Cartier. Then we say that (X, ∆) is log canonical if the following condition holds. For every proper birational map π : eX −→ X with eX normal, when we write X aiEi = K eX − π∗(KX + ∆) each ai is ≥ −1. Theorem 7.2.3. [HW02, Main Theorem] If (X, ∆) is sharply F -pure, then (X, ∆) is log canon- ical. Proof. The statement is local on X and so we may assume that L = OX and that X = Spec R is affine. We only prove the case where the index of KX + ∆ is not divisible by p. To reduce to this case, use Exercise 7.6 below. Set ϕ : F e ∗ R −→ R to be a map corresponding to ∆. Thus there exists an element c ∈ R = Γ(X, L ) such that ϕ(F e ∗ c) = 1 since (X, ∆) is sharply F -pure. Set π : eX −→ X a proper birational map with eX normal and writeP aiEi = K eX −π∗(KX +∆). Suppose that some ai < −1 (with corresponding fixed Ei). Then in particular ai ≤ 0. Set ηi to be the generic point of Ei. It follows that −ai, the Ei-coefficient of ∆ eϕ, is positive and so we have a factorization: F e ∗ O eX,ηi ⊆ F e ∗ O eX,ηi ((1 − pe)aiEi) eϕ −→ O eX,ηi , that we have the factorization where eϕ is as in Lemma 7.2.1. But now it is easy to see that if ai < −1, then (1 − pe)ai ≥ pe so F e ∗ O eX,ηi ⊆ F e ∗ O eX,ηi (peEi) ∗ O eX,ηi eϕF e −−−−−−−−−−→ O eX,ηi (pe Ei) . which sends F e ∗ c ∈ F e ∗ R ⊆ F e parameter for Ei, then eϕ sends F e ∗ O eX,ηi to 1. But that is impossible since if d ∈ O eX,ηi ∗ (c/dpe ) ∈ F e ∗ O eX,ηi (peEi) to 1/d /∈ O eX,ηi . is the local (cid:3) 7.3. Finite maps. Finally, suppose that π : Y −→ X is a finite surjective map of normal varieties. Then there is an inclusion OX ⊆ π∗OY . Given a line bundle L on X and a map ∗ π∗(π∗L ) −→ π∗OY . ϕ : F e ∗ Since π is finite, the π∗ is harmless and so we can ask when ϕ can be extended to a map ϕY : F e L −→ OX , it is natural to ask when ϕ can be extended to a map F e ∗ π∗L −→ OY . The local version of this statement is as follows. Suppose that R ⊆ S is a finite extension of ∗ R −→ R is a finite map. Then when does there semi-local normal rings and suppose that ϕ : F e 34 MANUEL BLICKLE AND KARL SCHWEDE exist a commutative diagram as follows? ϕS F e ∗ S / S F e ∗ R ϕ / R It is easy to see that the answer is not always. Example 7.3.1. Consider k[x2] ⊆ k[x] with p = chark 6= 2. Consider the map ϕ : F∗k[x2] −→ k[x2] which sends F∗x2(p−1) to 1 and other monomials F∗x2i, for 0 ≤ i < p − 1 to zero. Note ∆ϕ = 0. Suppose this map extended to a map ψ : F∗k[x] −→ k[x]. Then ϕ(F∗x2(p−1)) = 1 and so since ϕ and ψ are the same on k[x2], we have which implies that x is a unit. But that is a contradiction. 1 = ψ(F∗x2(p−1)) = ψ(F∗xpxp−2) = xψ(F∗xp−2) On the other hand, consider the map α : F∗k[x2] −→ k[x2] which sends F∗x2(p−1)/2 = F∗xp−1 to 1 and the other monomials F∗x2i to 0 for 0 ≤ i ≤ p − 1 to zero. Note ∆α = 1 2 div(x2). We will show that α extends to a map β : F∗k[x] −→ k[x]. It is in fact easy to show that α extends to a map on the fraction field β : F∗k(x) −→ k(x), see Exercise 7.7. Therefore, it is enough to show that β(F∗xj) ∈ k[x] for each 0 ≤ j ≤ p − 1. Fix such a j. If j is even, then there is nothing to do since β(F∗xj) = α(F∗xj) ∈ k[x2] ⊆ k[x]. Therefore, we may suppose that j is odd. But then j + p is even and p ≤ j + p ≤ 2(p − 1). Thus β(F∗xj) = 1 x β(F∗xj+p) = 1 x α(F∗xj+p) = 1 x · 0 = 0 ∈ k[x]. This proves that β exists and is well defined. Theorem 7.3.2. [ST10a] Fix π : Y −→ X as above. Fix a nonzero map ϕ : F e L −→ OX as ∗ above. If π is inseparable then ϕ never extends to ϕY . If π is separable, then there exists a map ϕY : F e ∗ π∗L −→ OY extending ϕ if and only if ∆ϕ is bigger than or equal to the ramification divisor of π : Y −→ X. Proof. We won't prove this but we will sketch the main steps and leave the details as an exercise. We first work in the separable case. Step 1: The statement is local on X and so we may suppose that X = Spec R, Y = Spec S and L = OX . In fact, we may even assume that R is a DVR and that S is a Dedekind domain. Step 2: There is a map ϕS and a commutative diagram: ϕS F e ∗ S / S ∗ R? F e ϕ / R? if and only if there exists a map ϕS and a commutative diagram. ϕS F e ∗ S / S F e ∗ Tr Tr F e ∗ R ϕ / R where Tr : S −→ R is simply the restriction of the field trace Tr : K(S) −→ K(R) to S. / ?  O O / ?  O O /  O O /  O O   /   / p−1-LINEAR MAPS IN ALGEBRA AND GEOMETRY 35 Step 3: HomR(S, R) is isomorphic to S as an S-module. The map Tr : S −→ R is a section of this and so corresponds to a divisor D on Spec R. This divisor is the ramification divisor Ramπ of π : Spec S −→ Spec R. Step 4: Supposing ϕS exists, compute the divisor corresponding to Tr ◦ϕS = ϕ ◦ (F e ∗ Tr). This gives one direction of the if and only if. Working with the fraction fields, as in Exercise* 4.13, yields the other direction. For the inseparable case, it turns out that the only map that can extend is the zero map, see (cid:3) Exercise 7.9. 7.4. Exercises. Exercise 7.1. In the setting of Lemma 7.1.1, prove that the divisor D is F -pure (as a variety) if and only if ϕ is surjective. Exercise* 7.2. Prove Lemma 7.1.1. Hint: Consider the map hϕiF e ∗ R −→ HomR/hf i(F e ∗ R/hf i, R/hf i) and prove it is surjective at the codimension 1 points of R/hf i. For a solution, see [Sch09, Proposition 7.2]. Exercise** 7.3 (The F -different). Suppose that X is a normal variety and D is an effective normal Weil divisor such that KX + D is Q-Cartier with index not divisible by p. Thus there exists a map ϕD : F e L −→ OX as in (15) corresponding to D for any e such that (pe−1)(KX +D) ∗ is Cartier. It is easy to see that this map is compatible with D and so it induces a map: ϕD : F e ∗ L D −→ OD. This map corresponds to a Q-divisor ∆D on D, again by (15), which is called the F -different. Verify all the statements made above. It is an open question whether or not the F -different always coincides with the different, as described in [Kc92, Chapter 17] or [Sho92, 10.6]. Prove that it either does or does not and write a paper about it, and then tell the authors of this survey paper what you found (this is why the problem gets ∗∗). For more discussion see [Sch09, Remark 7.6]. Exercise* 7.4. Consider the family of cones over elliptic curves: X = Spec k[x, y, z, t]/hy2 − x(x − 1)(x − t)i −→ A1 = Spec k[t] Set Φ ∈ HomX (F∗OX , OX ) to be the map generating HomX(F∗OX, OX ) as an F∗OX -module. Show that Φ is compatible with the ideal J = hx, y, zi. Consider ΦJ = Φ/J, the map obtained by restricting Φ to V (J) ∼= A1. Show that ∆ΦJ is supported exactly at those points whose fibers correspond to supersingular elliptic curves. Exercise 7.5. Using the notation of Lemma 7.2.1, suppose that ∆ϕ is the effective divisor associated to ϕ. Show that there is a map ϕ′ : F e ∗(cid:0)(π∗L )(⌈K eX − π∗(KX + ∆ϕ)⌉)(cid:1) −→ O eX (⌈K eX − π∗(KX + ∆ϕ)⌉) that agrees with ϕ wherever π is an isomorphism. H int: It is sufficient to show that there is a map ϕ′′ : F e ∗(cid:0)(π∗L )(⌈K eX − π∗(KX + ∆ϕ)⌉ − pe⌈K eX − π∗(KX + ∆ϕ)⌉)(cid:1) −→ O eX. Now, use the roundings to your advantage and the fact that π∗L = O eX(π∗(1 − pe)(KX + ∆ϕ)). Exercise 7.6. Suppose that (X, ∆) is sharply F -pure. Prove that for every point x ∈ X there exists a divisor ∆U on a neighborhood U of x such that ∆U ≥ ∆U , such that (U, ∆U ) is sharply F -pure and such that KU + ∆U has index not divisible by p. Conclude that Theorem 7.2.3 holds in full generality. For a solution, see [SS10, Theorem 4.3(ii)]. 36 MANUEL BLICKLE AND KARL SCHWEDE Exercise 7.7. Suppose that R ⊆ S is an extension of integral domains with induced separable extension of fraction fields K(R) ⊆ K(S). Fix ϕ : F e ∗ R −→ R to be an R-linear map. Prove that there is always a map ψ : F e Hint: First form ϕη : F e ∗ K(S) −→ K(S) such that ψR = ϕ. ∗ K(R) −→ K(R) by localization. Then tensor this map with K(S) and use the fact that K(R) ⊆ K(S) is separable (unlike K(R) ⊆ F e ∗ K(R) ∼= (K(R))1/pe ). Exercise* 7.8. Prove the separable case of Theorem 7.3.2 by filling in the details of the Steps 1 through 4. Step 3 is somewhat involved, see for example [Mor53, SS75, dS97]. On the other hand, see [ST10a] for a complete proof. Exercise 7.9. Prove the inseparable case of Theorem 7.3.2 as follows. First suppose that K ⊆ L is a purely inseparable extension of fields. Suppose that ϕ : F∗K −→ K is a K-linear map that extends to an L-linear map ϕL : F∗L −→ L. Prove that ϕ = 0. Use the above to prove that now if L ⊇ K is any inseparable map, the only map ϕ : F∗K −→ K that extends to ϕL : F∗L −→ L is the zero map. For a complete solution, see [ST10a, Proposition 5.2]. 8. Cartier modules Perhaps the most natural example of a p−e-linear map is the trace of the Frobenius F∗ωX −→ ωX on the canonical sheaf of a normal variety as discussed in detail in Section 3.2. In generalizing one is lead to consider the category consisting of (coherent) OX -modules F equipped with a p−e-linear map κ : F e F −→ F . We will outline here the resulting theory in a slightly more ∗ general setting than considered in [BB11]. Definition 8.0.1. If L is a line bundle on X, then a (L , pe) -- Cartier module is a coherent OX -module F equipped with an OX -linear map κ : F e ∗ (F ⊗ L ) −→ F . (or equivalently, equipped with a p−e linear map F ⊗ L −→ F ). If L ∼= OX , we call these objects mostly just Cartier modules. Remark 8.0.2. Cartier modules as originally defined in the work of [BB11] were always defined with L ∼= OX. The addition of the L adds little to the complication of the basic theory (which generally reduces to the local case where L is trivialized). Although admittedly, it does add some notational complications. However, this generalization does show up naturally. Regardless, little will be lost if the reader always assumes that L = OX . A morphism of (L , pe) -- Cartier modules (F , κF ) and (G , κG ) is an OX-linear map ϕ : F −→ G such that the diagram F e ∗ (F ⊗ L ) κF F F e ∗ (ϕ⊗id) ϕ F e ∗ (G ⊗ L ) / G κG commutes. If (F , κ) is a (L , pe) -- Cartier module, then we can apply F e using the projection formula -- a map ∗ to κ ⊗ L to obtain -- κ2 : F 2e ∗ (F ⊗ L ⊗ L pe ) ∼= F e ∗ (F e ∗ (F ⊗ L ) ⊗ L ) ∗ (κ⊗L ) F e −−−−−−−→ F e ∗ (F ⊗ L ) κ−−→ F which equips F with the structure of a (L 1+pe in the obvious way (similar to (17)) we obtain morphisms , p2e) -- Cartier module. Iterating this construction κe : F ne ∗ (cid:16)F ⊗ L 1+pe+p2e+···+p(n−1)e(cid:17) −→ F / /     / p−1-LINEAR MAPS IN ALGEBRA AND GEOMETRY 37 for all n ≥ 1, making F into a (L pne−1 pe−1 , pne)-Cartier module. Proposition 8.0.3. The category of (coherent) (L , pe)-Cartier modules is an Abelian cate- gory. The kernel and cokernel of the underlying quasi-coherent sheaves carry an obvious Cartier module structure and are the kernel and cokernel in the category of Cartier modules. Proof. This is easy to verify since ⊗L as well as F e are exact functors. Alternatively, we may ∗ view (L , pe)-Cartier modules as the right module category over a certain (non-commutative) sheaf of rings, see Exercise 8.4 below, which immediately implies that the category is Abelian. (cid:3) Compared to a Frobenius splitting, which is nothing but a Cartier module structure on the coherent sheaf OX, the advantages of working in this larger category of Cartier modules are manifold. For one, there are a number of natural examples of Cartier modules, most promi- nently the canonical sheaf ωX together with the trace of Frobenius as Cartier module structure. Furthermore one has in this category methods to construct new Cartier modules by functorial operations. Most notably there is the notion of a push-forward for proper maps (in the case that L ∼= OX), localization and ´etale pullback, and even an extraordinary pullback f ! can be defined [BB06, BB11]. We conclude this subsection by illustrating some of these concepts in special cases. First however, we state some examples. Examples 8.0.4 (Examples of Cartier modules). (a) The canonical sheaf ωX is a Cartier module with structural map κ : F∗ωX −→ ωX given X is the dualizing complex of X, then the trace of X. This induces for each i the by the trace map. More generally, if ω q Frobenius is a map (in the derived category) F∗ω q structure of a Cartier module on the cohomology hiω q X. (b) Suppose that D is a Cartier divisor on X, then the map X −→ ω q F e ∗ (ωX(peD)) Tr −→ ωX(D) equips ωX(D) with the structure of an OX((pe − 1)D)-Cartier module. (c) Suppose that D is an effective integral divisor on X, then the composition F∗ωX(D) ֒→ F∗ωX(pD) Tr−→ ωX(D) equips ωX(D) with the structure of a Cartier module as well. (d) Suppose that π : Y −→ X is a proper map of varieties. Then Riπ∗ωY is a Cartier module for any i ≥ 0. This is because F∗Riπ∗ωY = Riπ∗F∗ωY . (e) Set X = A2 and let π : Y −→ X be the blowup at the origin with exceptional divisor E. Thus we have TrY : F∗ωY −→ ωY as the trace on Y . Now, ωY ∼= OY (E). Thus by twisting by −E ,we have a OY ((1 − p)E)-Cartier module structure on OY . Namely, a map Tr : F∗(OY ((1 − p)E)) −→ OY . Since localization at any multiplicative set commutes with pushforward along the Frobenius (see Exercise 2.6 and Exercise 8.3) we observe that localization preserves the Cartier module structure. Lemma 8.0.5. Let S ⊆ R be a multiplicative system and F a (L , pe) -- Cartier module on X = Spec R. Then the map S−1R∗(S−1F ⊗S−1R S−1L ) ∼= S−1F e F e is a (S−1L , pe) -- Cartier module structure on S−1F . ∗ (F ⊗R L ) S−1κF −−−−−−→ S−1F 38 MANUEL BLICKLE AND KARL SCHWEDE In particular, if j : U ⊆ X = Spec R is the inclusion of a basic open subset U = Spec Rf for some f ∈ R, then the pullback j∗ induces a functor from L -- Cartier modules on X to j∗L -- Cartier modules on U . Using a Cech-complex construction, this globalizes to an arbitrary open immersion U ⊆ X. Even more generally this holds for any essentially ´etale8 morphism j : U −→ X, see [BB06]. Proposition 8.0.6. Let j : U −→ X be essentially ´etale and let F be a L -Cartier module on X. Then the pullback j∗F carries a natural functorial structure of a j∗L -Cartier module on U . The structural map is given by FU ∗(j∗F ⊗ j∗L ) ∼= FU ∗j∗(F ⊗ L ) ∼= j∗FX∗(F ⊗ L ) j ∗κ −−−→ j∗F . Proof. The key point is the fact that for an essentially ´etale morphism j : U −→ X the diagram FY U U j j X FX / X is Cartesian and that the base change morphism j∗FX∗ ∼= FU∗j∗ is an isomorphism since j is flat, see [HH90]. This justifies the definition of the Cartier structure on j∗F . (cid:3) For a closed immersion i : Y −→ X, the pullback i∗ does not give a functor on Cartier modules. The reason is precisely that the above diagram is not Cartesian in this case. However there is an exotic restriction functor one can define. For concreteness, let X = Spec R be affine and let Y = Spec R/I for some ideal I ⊆ R. Then, for an R-module M , the R/I submodule i♭(M ) := HomR(R/I, M ) = {m ∈ M Im = 0} is just the I-torsion submodule M [I] ⊆ M . Note that F∗(M [I]) ⊆ F∗(M [I [p]]) = (F∗M )[I] which shows that F∗(M [I]), is contained in the I-torsion submodule (F∗M )[I] of F∗M . Hence, if κ : F∗M −→ M is a Cartier module structure on M , then we have that this restricts to a map κ : F∗(M [I]) −→ M [I] giving M [I] a natural Cartier module structure. The same construction works globally and more generally for (L , pe)-Cartier modules: Proposition 8.0.7. Let i : Y ֒−→ X be a closed immersion given by a sheaf of ideals I of OX , and let F be a (L , pe)-Cartier module on X. Then the OY -module (via action on the first argument) i♭(F ) = HomOX (i∗OY , F ) = F [I] carries a natural functorial structure of a (L Y , pe)-Cartier module on Y . The structural map is given by F e ∗ (F [I] ⊗OY L Y ) ⊆ F e ∗ ((F ⊗OX L )[I [pe]]) = (F e ∗ (F ⊗OX L ))[I] κF−−−→ F [I] = i♭F . Finally, let us consider a proper morphism of varieties π : Y −→ X. Since the Frobenius X∗ ◦ π∗ ∼= π∗ ◦ F e commutes with any morphism one has a natural isomorphism of functors F e which implies that the pushforward induces a functor on Cartier modules as well. Y ∗ Proposition 8.0.8. Let π : Y −→ X be a proper morphism and κ : F e ∗ on Y . Then the map F −→ F a Cartier module ∗ (π∗(F )) ∼= π∗(F e F e ∗ F ) π∗(κ) −−−−→ π∗F is a Cartier module structure on π∗F . The same construction also holds for the higher derived images Riπ∗F . 8essentially ´etale means essentially of finite type and formally ´etale, i.e. a morphism that can be factored as a localization followed by a finite type ´etale morphism / /     / p−1-LINEAR MAPS IN ALGEBRA AND GEOMETRY 39 Note, however, that if F is a (L , pe) -- Cartier module there is no obvious way to equip π∗F with such a structure unless L is of the form π∗L ′ for some invertible sheaf on X. In this case, using the projection formula, one obtains ∗ (π∗F ⊗ L ′) ∼= F e F e ∗ (π∗(F ⊗ π∗L ′) ∼= π∗(F e ∗ (F ⊗ L ) π∗(κ) −−−−→ π∗F as a Cartier structure on π∗F . Example 8.0.9. Let κ : F e ∗ Frobenius (which is an affine map) equips F∗F with the Cartier module structure F −→ F be a Cartier module, then the pushforward along the F e ∗ κ : F e ∗ (F e ∗ (F )) −→ F e ∗ F making κ : F e ∗ F −→ F into a map of Cartier modules. 8.1. Finiteness results for Cartier modules. In this section we state, and outline the proofs of two key structural results which make the category of Cartier modules interesting. But first we introduce the basic concept of nilpotence of a Cartier module and recall some elementary constructions, starting with the following simple Lemma whose verification we leave to the reader in Exercise 8.1. Lemma 8.1.1. Let κ : F e κn(F ne ∗ (F ⊗ L 1+pe+···+p(n−1)e ∗ (F ⊗ L ) −→ F be a Cartier module. Then the images Fn := ) ⊆ F are Cartier submodules of F , and satisfy the properties: (a) Fn ⊇ Fn+1. (b) κ(F e (c) If S ⊆ OX is a multiplicative set, then S−1Fn = (S−1F )n. (d) The sequence of closed subsets Yn := Supp Fn/Fn+1 is descending. ∗ (Fn ⊗ L )) = Fn+1. An important notion in the theory of Cartier modules, and in particular, for its applications to finiteness results for local cohomology for local rings, is the notion of nilpotence. Definition 8.1.2. Let F be a coherent Cartier module on X. We say that F is nilpotent if for some n ≥ 0 the nth power κn of the structural map κ is zero. Some basic properties of this notion are collected in the following Lemma: Lemma 8.1.3. Let κ : F e submodule of F consisting of all local sections s such that κn(F ne Then ∗ (F ⊗ L ) −→ F be a Cartier module. Denote by F n ⊆ F the Cartier ) = 0. ∗ (OC ·s⊗L 1+pe+···+p(n−1)e (a) F n ⊆ F n+1 for all n ≥ 0. ∗ (F n+1 ⊗ L )) ⊆ F n. (b) κ(F e (c) If S ⊆ OX is a multiplicative set, then S−1F n = (S−1F )n. (d) If F is coherent, then the ascending sequence stabilizes and the stable member Fnil = F n is the maximal nilpotent Cartier submodule of F . Sn Nilpotent Cartier modules form a Serre subcategory of all coherent Cartier modules, i.e. they form an Abelian subcategory which is closed under extension. The only non-trivial part here is the non-closedness under extensions, see Exercise 8.5. The first structural result for Cartier modules we will show is that the descending sequence of iterated images stabilizes. This result was first proved in [Gab04, Lemma 13.1]. In fact, this result is essentially Matlis dual to a famous result of Hartshorne and Speiser [HS77, Proposition 1.11] and generalized by G. Lyubeznik [Lyu97], also cf. [Sha06, Sha07b] and [Bli08]. Proposition 8.1.4. Let (F , κ) be a coherent (L , pe) -- Cartier module. Then the descending sequence of images Fn := κn(F ne ∗ (F ⊗ L 1+pe+···+p(n−1)e ) ⊆ F 40 MANUEL BLICKLE AND KARL SCHWEDE stabilizes. In particular, the stable image σ(F ) ⊆ F is the largest (L , pe)-Cartier submodule with the property that the structural map κ is surjective. Proof. To show the stabilization of a sequence of subsheaves on a Noetherian scheme X can be done on an affine open cover. Choosing the open sets of the cover sufficiently small we may assume that L is trivial. Hence we may assume that X = Spec R and M is a finitely generated R module equipped with a p−e-linear map κ : M −→ M . And we have to show that the descending sequence of Cartier submodules of M stabilizes. The sets M ⊇ κ(M ) ⊇ κ2(M ) ⊇ · · · Yn := Supp(κn(M )/κ(κn(M ))) form a descending sequence of closed subsets of X, by Lemma 8.1.1. Since X is Noetherian, the descending sequence must stabilize. After truncating we may assume that Y = Yn = Yn+1 for all n. We have to show that Y is empty. Assuming otherwise, let p be the generic point of a component of Y . Localizing at p we may assume that R is local with maximal ideal p and that Y = {p} = Supp(κn(M )/κ(κn(M ))) for all n. In particular, for e = 0 we obtain that there is an integer k such that p kM ⊆ κ(M ). Hence, for any x ∈ p k kM ⊆ xκ(M ) = κ(xpe x2M ⊆ xp M ) ⊆ κ(x2M ) and iterating we get x2M ⊆ κn(M ) for all n. Hence p is the number of generators of p κn(M )/p k(b − 1) ⊆ κn(M ) for all e where b k. Hence the original chain stabilizes if and only if the chain (cid:3) k(b−1)M does. But the latter is a chain in the finite length module M/p k(b−1)M . A characterization of this stable image is as follows. σ(F ) ⊆ F is the smallest Cartier submodule of F such that on the quotient F /σ(F ) some power of the structural map is zero. If this property is satisfied for some Cartier submodule N ⊆ F , then it is also satisfied for its image. The minimality now implies that for σ(F ) the structural map F e ∗ (σ(F ) ⊗ L ) −→ σ(F ) is surjective. The Cartier modules with surjective structural map play an important role in the theory. For example one can see immediately (Exercise 8.2), that for such Cartier module κ : F e ∗ (F ⊗ L )−→→F its annihilator Ann F is a sheaf of radical ideals, i.e. F has reduced support. This may be viewed as a generalization of the reduced-ness of Frobenius split varieties alluded to earlier. It is also a key ingredient in the following Kashiwara-type equivalence which will be used repeatedly below (the easy but rewarding proof is left to the reader as Exercise 8.10, see also [BB11, Proposition 2.6 and Section 3.3]): Proposition 8.1.5. Let F be a coherent Cartier module on X with surjective structural map κF F is a sheaf of radical ideals and hence F = F [I] = i♭(F ) (i.e. σ(F ) = F ). Then I = AnnOX is a Cartier module on Y = Supp F , the closed reduced subset of X given by I. More precisely, if i : Y −→ X denotes a closed immersion, then the functors i♭ and i∗ induce a (inclusion preserving) bijection between  coherent L -- Cartier modules on X with surjective structural map and Supp F ⊆ Y ←→(cid:26) coherent L Y -Cartier modules on Y with surjective structural map (cid:27) The most important structural result for Cartier modules is the following theorem which asserts that for a coherent Cartier module F , the lattice of Cartier submodules with surjective structural map satisfies the ascending and descending chain conditions.  p−1-LINEAR MAPS IN ALGEBRA AND GEOMETRY 41 Theorem 8.1.6. Let X be a scheme and κ : F e Then any chain of Cartier submodules ∗ (F ⊗ L ) −→ F a coherent Cartier module. · · · Fi ⊇ Fi+1 ⊇ Fi+2 ⊇ · · · each of whose structural map κFi is surjective, is eventually constant (in both directions). Proof. The ascending chain stabilizes simply because the underlying OX-module is coherent and our schemes are Noetherian. So it remains to show the descending chain condition. One way to proof this result is to show that there is a unique smallest Cartier submodule τ (F ) ⊆ F which agrees with σ(F ) on each generic point of X, i.e. τ (F )η = σ(F )η for each η the generic point of an irreducible component of X. This is a generalization of the notion of a test ideal which will be discussed in some detail below Section 9.3. Assuming the existence of τ (F ) for now, the proof can be outlined as follows: We show that the chain F0 ⊇ F1 ⊇ F2 ⊇ · · · stabilizes by induction on dim X, the case dim X = 0 being clear. Since a chain stabilizes if it stabilizes after restriction to each of the finitely many irreducible components of X, we may assume that X is irreducible. Since X is Noetherian, the descending sequence of supports Supp Fi stabilizes. After truncating we may assume that Y = Supp Fi for all i. Since the structural map of each Fi is surjective, we have by Proposition 8.1.5 that Fi is annihilated by the ideal sheaf defining the reduced structure of Y . Hence we may view the Fi as (L Y , pe) Cartier modules on Y . If dim Y < dim X then we are done by induction. So let us assume otherwise, that dim X = dim Y . Further truncating the sequence Fi we may assume that all Fi's have the same generic rank. Now, by definition, τ (F0) is contained in Fi for all i (in fact τ (F0) = τ (Fi)) such that it is enough to show the stabilization of the sequence F0/τ (F0) ⊇ F1/τ (F0) ⊇ F2/τ (F0) ⊇ · · · . But since τ (F0) generically agrees with each Fi this is a sequence of Cartier modules Fi/τ (F0) whose entries have strictly smaller support than X. As above, we are done by induction. (cid:3) A corollary of the proof is the following result. Proposition 8.1.7. Let F be a coherent Cartier module on X with surjective structural map. Then the set {supp F /G G ⊆ F a Cartier submodule} is a finite set of reduced subschemes that is closed under finite unions and taking irreducible components. Proof. We only prove the finiteness and leave the rest as an exercise Exercise* 8.6. We proceed by induction on dim X. By Proposition 8.1.5 we may view F as a Cartier module on supp F , hence we may assume that supp F = X. Since X is Noetherian it has only finitely many irreducible components so we may assume that X itself is irreducible. If supp F /G 6= X then F and G agree on the generic point of X. Hence the test module τ (F ) ⊆ G . Therefore supp F /G ⊆ supp F /τ (F ) =: Y and Y is a proper closed subset of X. Again using Proposition 8.1.5 we can apply the induction hypothesis to the Cartier module F /τ (F ) on Y whose dimension is strictly less than dim X. (cid:3) This yields the following corollary which was obtained in [KM09] and also independently obtained by the second author in [Sch09]. In the case that X = Spec R and R is local, proofs of this fact were first obtained in [Sha07a] and [EH08]. Corollary 8.1.8. Let X be Frobenius split, then OX has only finitely many ideals with are compatible with the splitting. 42 MANUEL BLICKLE AND KARL SCHWEDE Proof. If ϕ : F e ∗ OX −→ OX is the splitting of Frobenius, note that ϕ is surjective. The Cartier submodules of OX are just the ideals which are ϕ-compatible. Since Ann(OX /I) = I there is a one-to-one correspondence between the set of ϕ-compatible ideals, and the set supp OX /I for I a Cartier submodule of OX. The latter set is finite by the preceding proposition. (cid:3) 8.2. Cartier Crystals. The finiteness results for Cartier modules of the preceding section re- ceive a more natural formulation if one deals with the notion of nilpotence in a more systematic manner. This is done by localizing the category of coherent Cartier modules at its Serre subcat- egory9 of nilpotent Cartier modules. That is, we invert morphisms which are nil-isomorphisms, i.e. maps of Cartier modules ϕ : F −→ G whose kernel and cokernel are nilpotent. For the formal definition, see [Miy91, Gab62], but roughly speaking the localization is defined as follows: Definition 8.2.1. Let X be a scheme. The category of L -- Cartier crystals has as objects the coherent Cartier modules on X. A morphism ϕ : F −→ G of Cartier crystals is an equivalence class (left fraction) of diagrams of morphisms of the underlying Cartier modules ϕ : F ← F ′ ϕ′ −−→ G where F ′ is some Cartier module and F ← F ′ is a nil-isomorphism. More precisely, where F ′ −→ G ranges over all nil-isomorphisms. HomCrys(F , G ) = colimF ′−→F HomCart(F ′, F ) It follows from general principles that the category of Cartier crystals on X is again Abelian. Using this point of view the preceding result can be phrased (and extended) as follows, see [BB11, Theorem 4.17 and Corollay 4.7]: Theorem 8.2.2. Let X be a scheme. (a) Each Cartier crystal F has finite length in the category of Cartier crystals. (b) Hom-sets in the category of Cartier crystals are finite sets (finite dimensional Fpe vector spaces). (c) Each Cartier crystal F has only finitely many Cartier sub-crystals. Proof. The first statement follows from Theorem 8.1.6 above by noting that F and σ(F ) are isomorphic as Cartier crystals (i.e. nil-isomorphic as Cartier modules). The second statement is shown in [BB11, Theorem 4.17] (but see Exercise 8.11 below for an idea why such a statement may hold), and the last one follows formally from the other two. (cid:3) In [BB06] the category of Cartier crystals (for L ∼= OX) is thoroughly studied on an arbitrary Noetherian scheme such that F : X −→ X is finite. In particular it is shown that half of Grothendieck's six operations, namely f !, Rf∗ and an exotic tensor product, can be defined on a suitable derived category of Cartier crystals. In particular the construction of the functors f ! and Rf∗ is rather subtle and bears some interesting insights. This greatly extends the examples of the pullback for open and closed immersions and the proper push-forward that was discussed in the preceding section. If f : Y −→ X is a proper morphism, then Rif∗ induces a functor on (coherent) Cartier modules, which can be shown to preserve nilpotence. Hence it descends to a functor on Cartier crystals. However, if f is not proper, then already f∗F of a coherent sheaf is no longer coherent. It is a crucial observation in [BB06] that if F is a coherent Cartier crystal on Y , then Rif∗F is a locally nil-coherent Cartier crystal on X. Nil-coherent for a Cartier module F means that F has a coherent Cartier submodule E ⊆ F such that the quotient F /E is locally nilpotent, i.e. is the union of nilpotent Cartier submodules. This implies the following result: 9i.e. a full Abelian subcategory which is closed under extensions, see [BB06]. p−1-LINEAR MAPS IN ALGEBRA AND GEOMETRY 43 Theorem 8.2.3. For an arbitrary finite type morphism f : Y −→ X, the usual push-forward functor Rf∗ on quasi-coherent sheaves induces an exact functor Rf∗ : Db crys(QCrys(Y )) −→ Db crys(QCrys(X)) , where Db whose cohomology is locally nil-coherent. crys(QCrys( )) denotes the bounded derived category of quasi-coherent Cartier crystals The proof of this result, though not difficult, is somewhat subtle, so we won't attempt it here but instead refer to [BB06]. However the basic idea is already present in Exercise* 8.7 The situation with the functor f ! is similar but more subtle. As we have already seen in Section 3.3, on quasi-coherent sheaves the construction of the functor f ! is generally quite in- volved. Already in the finite case, in particular for a closed immersion Y ⊆ X with X not smooth, one sees that f ! does not have bounded cohomological dimension, hence does not pre- serve the bounded derived category. However, in [BB06] it is shown quite generally that f ! preserves local nilpotence, and hence induces a functor on quasi-coherent Cartier crystals. The induced functor on Cartier crystals preserves boundedness up to local nilpotence. Theorem 8.2.4. If f : Y −→ X is essentially of finite type, then the twisted inverse image functor f ! on quasi-coherent sheaves induces an exact functor f ! : Db crys(QCrys(X)) −→ Db crys(QCrys(Y )) . of bounded cohomological dimension. Besides a number of obvious compatibilities between these functors which are induced from the corresponding ones of the underlying quasi-coherent sheaves, there are two adjointness statements which are important in the theory. Proposition 8.2.5. (a) Let f : Y −→ X be a proper morphism. Then as functors on cate- gories Db crys(QCrys( )) the functor Rf∗ is naturally left adjoint to f !. (b) If j : Y −→ X is an open immersion, then j∗ is naturally right adjoint to j! = j∗. For an open immersion j : U ֒−→ X and a closed complement i : Z ֒−→ X the above adjunction yields natural isomorphisms i∗i! −→ id and id −→ j∗j∗. This yields the following technically important result regarding their combination: Theorem 8.2.6. In Db crys(QCrys(X)), there is a natural exact triangle i∗i! −→ id −→ Rj∗j∗ +1 −−−→ This in turn yields a very general form of the Kashiwara equivalence that was alluded to in Proposition 8.1.5 above. Theorem 8.2.7. Let i : Y −→ X be a closed immersion. Then i! and i∗ define natural isomor- phisms Db crys(QCrys(Y )) i∗ i! / Db crys,Y (QCrys(X)) where the right hand category consists of bounded complexes of quasi-coherent Cartier crystals on X whose cohomology is coherent and supported in Y . 8.3. Arithmetic aspects of p−e-linear maps. We conclude with a brief discussion of con- nections between Cartier crystals and more arithmetic constructions. What follows is much less explicit than previous sections of the paper, so if the terms used are not familiar to you, we suggest the reader use this as a place jump off for further reading. The finite length result for Cartier crystals in Theorem 8.2.2 suggests -- in analogy with the Riemann-Hilbert correspondence for D-modules (i.e.modules of the ring of differential operators) / o o 44 MANUEL BLICKLE AND KARL SCHWEDE on smooth complex manifolds -- a connection of Cartier crystals with a category of constructible sheaves. Indeed, in [Gab04] Gabber introduces a family of t-structures on the derived category of bounded complexes of constructible Fp-vector spaces on the ´etale site of X. He shows that for the middle perversity the heart of this t-structure (i.e. the perverse sheaves with respect to this t-structure) form an Abelian category which also is Noetherian and Artinian. The connection between Cartier crystals and constructible Fp-vector spaces is a combination of [BP09] and [BB06] and yields an equivalence of derived categories: ∼=−−→ Db Db crys(QCrys(X)) c(Xet, Fp) where the right hand side is the category of constructible sheaves of Fp-vector spaces on Xet. This correspondence is a two step procedure: First is a Grothendieck-Serre duality between Cartier crystals (coherent OX -modules with a right action of Frobenius) with the category of τ -crystals (coherent OX -modules with a left Frobenius action) of [BP09] and was largely motivated by our desire to understand the precise connection of the theory in [BP09] with the work of Emerton and Kisin [EK04] and Lyubeznik [Lyu97]. This Grothendieck-Serre duality is the main result of [BB06]. The step from τ -crystals to constructible sheaves is just by taking Frobenius fix-points, i.e. the Artin-Schreier sequence, see [BP09]. The first author's PhD student Tobias Schedlmeier has shown in his upcoming thesis that the equivalence is given directly by the functor Sol( X) and proved that the image of the Abelian subcategory of Cartier crystals under Sol is precisely Gabbers category of perverse sheaves Perv(Xet, Fp) for the middle perversity. ) := RH omcrys( , ω q 8.4. Exercises. Exercise 8.1. Prove Lemma 8.1.1. Exercise 8.2. Show that the annihilator of any coherent Cartier module F on X with surjective structural map is a sheaf of radical ideals, i.e. its support is reduced. Exercise 8.3. Let R be a ring and S ⊆ R a multiplicative set. Then for any module M show that S−1(FR)∗M ∼= (FS−1R)∗S−1M . Hint: Localize with respect to the multiplicative set Sp is the as with respect to S. This gener- alizes Exercise 2.6. Exercise 8.4. Let X be a scheme and L a line bundle. We define a sheaf of rings OL X [F e] as OX ⊕ (L · F e) ⊕ (L 1+pe · F 2e) ⊕ (L 1+pe+p2e · F 3e) ⊕ · · · where F ne are formal symbols and the multiplication of homogeneous elements lF ne and l′F ne′ is defined as lF nel′F n′e = l(l′)pn F (n+n′)e. e ′ (a) Show that this defines the structure of a sheaf of rings on OL (b) Show that the category of (L , pe) -- Cartier modules is equivalent to the category of X [F e]. (sheaves of) right OL X [F e]-modules. Hint: Do the case of L ∼= OX first and then attempt the general case. Exercise 8.5. If 0 −→ F ′ −→ F −→ F ′′ −→ 0 is an exact sequence of coherent Cartier modules. Show that F ′, F ′′ are nilpotent (of order ≤ e, e′) if and only if F is nilpotent (of order ≤ e+ e′). Exercise* 8.6. Let F be a quasi-coherent Cartier module with surjective structural map. Show that the collection {supp(F /G ) G ⊆ F a Cartier submodule} is a collection of reduced subschemes that is closed under finite unions and taking irreducible components. p−1-LINEAR MAPS IN ALGEBRA AND GEOMETRY 45 Exercise* 8.7. Let X = Spec R be an affine scheme and U = Spec Rf a basic open subset with f ∈ R, and denote the open inclusion U ⊆ X by j. Let F be a coherent Cartier module on U . Show that j∗F has a coherent Cartier submodule F such that the quotient j∗F /F is locally nilpotent, i.e. the union of its nilpotent Cartier submodules. Exercise 8.8. Let F be a coherent Cartier module on X. The test submodule τ (F ) is defined as the smallest Cartier submodule G ⊆ F which agrees with σ(F ) for each generic point of X. Show that Theorem 8.1.6 implies the existence and uniqueness of τ (F ). Exercise 8.9. Suppose that R is a ring and (M, ϕ) is a Cartier module on M . Suppose further that R −→ S is a finite ring homomorphism. Prove that HomR(S, M ) has the structure of a Cartier module induced by ϕ and by the Frobenius map S −→ F∗S. Exercise 8.10. Prove Proposition 8.1.5. Exercise 8.11. Let R be a regular F -finite ring with dualizing sheaf ωR with its standard Cartier structure T : F∗ωR −→ ωR (see Section 3.2). Show that the homomorphisms of Cartier modules HomCart(ωR, ωR) = RF = Fp is just the Frobenius fixed points of the action of F on R. In particular, this Hom-set is finite. Exercise* 8.12. Suppose that R = k[x1, . . . , x4]hx1,...,x4i and that ϕ : F e ∗ R −→ R is a Frobenius splitting. In Corollary 8.1.8, it was shown that there are at most finitely many ϕ-compatible ideals. d(cid:1) prime ideals Q which are compatibly split by ϕ such that Prove that there at most (cid:0)4 dim(R/Q) = d. Hint: Prove it for d = 0 first (very easy), then d = 1 (use the fact that compatibly split subvarieties must intersect normally, Corollary 5.1.7, but we only have 4 "directions" in Spec R, which is just the origin in A4). For d = 2, 3, simply consider all possibilities exhaustively (keeping in mind the normal intersections). For a complete proof for any An (not just n = 4), see [ST10b]. Exercise 8.13. Suppose that (F , κ) is an (L , pe)-Cartier module on a projective variety X such that the structural map κ : F e ∗ (L ⊗ F ) −→ F is surjective. Further suppose that A is a globally generated ample line bundle and that N is another line bundle such that N pe−1 ⊗ L is ample. Prove that is a globally generated sheaf. Hint: Use the same strategy as in Theorem 6.1.3. F ⊗ A dim X ⊗ N 9. Applications to local cohomology and test ideals In this section we discuss in detail the relation of the theory of Cartier modules to other theo- ries of modules with a Frobenius action, with an emphasize on applications to local cohomology. Then we discuss a simple but interesting degree-reducing property of Cartier linear maps, which allows an elementary treatment of the theory of Cartier modules in the case that X is of finite type over a perfect field. We use this approach to study the test ideals and show the discreteness of their jumping numbers. 9.1. Cartier modules and local cohomology. The category of Cartier modules, besides en- joying some extraordinary finiteness conditions, is useful due to its connection to other categories which are studied, in particular in connection with local cohomology. Besides the connection to constructible p-torsion sheaves that we hinted at above, we show the relation to two further categories which are particularly important in the study of the local cohomology of rings in positive characteristic. Our goal is to explain the following diagram of categories and to derive 46 MANUEL BLICKLE AND KARL SCHWEDE a number of finiteness results for local cohomology from the above finiteness result for Cartier modules. with left Frobenius action cofinite R-modules (R complete local ring) ) ↔n coherent Cartier modules on X (X Noetherian and F -finite) o −→n Lyubeznik's F -modules over R (R regular, Noetherian ring) o ( The parenthetical parts indicate in what generality the categories are defined and the arrows are defined when both assumption holds, for example the first double arrow holds for complete local and F -finite rings. The left double arrow is an equivalence of categories given by Matlis , ER/m) where ER/m is an injective hull of the perfect residue field of R. The duality HomR( right arrow is a functor which gives an equivalence after inverting Cartier modules at nil- isomorphisms, that is it induces an equivalence of categories from Cartier crystals to F -finite modules. Lyubeznik's F -finite modules and this equivalence will be explained in detail below. Let us begin with Matlis duality. Let (R, m) be complete and local and denote by E = ER ∗ F !ER := an injective hull of the prefect residue field of R. Since R is F -finite one has that F e HomR(F∗R, ER) ∼= EF∗R which we identify with ER since R and F∗R are isomorphic as rings. We fix hence an isomorphism F !E ∼= E. If we denote by ( , ER) the Matlis duality functor, we have the following lemma whose proof we leave as Exercise 9.2. )∨ = HomR( Lemma 9.1.1. For (R, m) local and F -finite there is a (functorial) isomorphism F∗( (F∗ )∨. )∨ ∼= This immediately implies the first of the equivalences above, also cf. [SY11]. Proposition 9.1.2. Let (R, m) be complete, local and F -finite. Then Matlis duality induces an equivalence between the categories of (cid:26) co-finite R-modules with left Frobenius action (cid:27) ↔(cid:26) finitely generated R-modules with right Frobenius action (cid:27) Of course the R-modules with right Frobenius action are just the coherent Cartier modules on X = Spec R. The equivalence preserves nilpotence. Proof. A left action of Frobenius on M is an R-linear map ϕ : M −→ F∗M . Applying Matlis duality and the preceding lemma this yields a map F∗(M ∨) ∼= (F∗M )∨ ϕ∨ −−−→ M ∨ which is the desired Cartier structure (=right Frobenius action) on the dual M ∨. The same construction works in the opposite direction and one immediately checks that this induces an equivalence of categories. (cid:3) With this result we can translate the finiteness theorems for Cartier modules obtained above to the setting of cofinite R-modules with a left Frobenius action. In particular the results hold for local cohomology modules H i m(R) with support in the maximal ideal m of R. Theorem 9.1.3. Let N be a cofinite R module equipped with a p-linear map F : N −→ N (i.e. F is additive and F (rm) = rpF (m)). (a) The ascending chain of submodules ker F ⊆ ker F 2 ⊆ ker F 3 ⊆ · · · stabilizes ([HS77, Proposition 1.1]). (b) Any chain · · · ⊆ Ni ⊆ Ni+1 ⊆ Ni+2 ⊆ · · · of submodules Ni ⊆ N which are stable under F (i.e. F (Ni) ⊆ Ni) has eventually F -nilpotent quotients ([Lyu97, Theorem 4.7]) (c) N has up to nilpotent action of F , only finitely many F -stable submodules. Concretely, there are only finitely many F stable submodules N ′ for which the action of F on the quotient N/N ′ is injective. p−1-LINEAR MAPS IN ALGEBRA AND GEOMETRY 47 Proof. These are just the Matlis dual statements of Proposition 8.1.4, Theorem 8.1.6, and Theorem 8.2.2 part (c). (cid:3) An immediate consequence of these observations is the following result originally obtained by Enescu and Hochster [EH08], see [Ma12] for a recent extension showing that F -split alone is sufficient in the assumptions below. Proposition 9.1.4. If R is quasi-Gorenstein (i.e. H d cohomology module H d submodules. m(R) ∼= ER) and F -split, then the top local m(R) with its left action of the Frobenius has only finitely many F -stable Proof. The existence of a splitting ϕ : R −→ S implies that the Cartier module (R, ϕ) has only m(R), ϕ∨) finitely many Cartier submodules. Hence, by the above duality result its dual (r∨ = H d has only finitely many submodules stable under the action of ϕ∨. But H d m(R) also has a natural Frobenius action FH induced by the Frobenius on R by functoriality of H d m( ). One can show (Exercise 9.7) that there is a r ∈ R such that ϕ∨ = rFH. Hence all submodules which are stable under FH are also stable under ϕ∨, but of the latter there are only finitely many as just argued. (cid:3) The connection of Cartier modules with Lyubeznik's F -finite modules also relies on a certain commutation of functors which we recall first. Lyubeznik's theory [Lyu97] is phrased for a regular ring R, and even though there is an extension to schemes by Emerton and Kisin [EK04], we will stick to this setting and assume from now on that X = Spec R, with R regular (and such that the Frobenius morphism F : R −→ R is finite). Lemma 9.1.5. Let f : Y −→ X be a finite flat morphism and M a OX module, then there is a functorial isomorphism f !OX ⊗OY f ∗M ∼= f !M , where f !( ) = HomOX (f∗OX , ). Proof. See Exercise 9.4. (cid:3) Applying this to the case of the Frobenius on the regular scheme X and M = ωX the dualizing sheaf (which is invertible!) we obtain an isomorphism F !OX ∼= F !ωX ⊗ F ∗ω−1 X . Further using that the adjoint of the map F∗ωX −→ ωX coming from the Cartier isomorphism in Theorem 3.1.1, is an isomorphism ωX −→ F !ωX we obtain F !M ⊗ ω−1 X ∼= F ∗M ⊗ F !ωX ⊗ ω−1 X ⊗ F ∗ω ∼= F ∗(M ⊗ ω−1 X ) which allows us to describe the functor from Cartier modules to Lyubeznik's F -finite modules. Starting with a Cartier module M with structural map κ : F∗M −→ M we first consider its adjoint κ′ : M −→ F !M and tensor it with ω−1 γ : M ⊗ ω−1 X X to obtain κ′⊗id −−−−→ F !(M ) ⊗ ω−1 X ∼= F ∗(M ⊗ ω−1 X ) where the final isomorphism is the one derived above. Let us pause for a moment to recall the definition of Lyubeznik's F -finite modules, which we phrase in a way convenient for our purpose: Definition 9.1.6. Let R be regular. Given a finitely generated R-module N together with a map γ : N −→ F ∗N , then an F -finite module is the limit N of the directed system γ −−→ F ∗N N together with the induced map ϑ : N phism. F ∗γ −−−−→ F 2∗N ∼=−−→ F ∗N which is immediately verified to be an isomor- F 2∗γ −−−−→ F 3∗N −→ · · · 48 MANUEL BLICKLE AND KARL SCHWEDE Phrased differently, an F -finite module is a (not necessarily finitely generated) R-module N ∼=−−→ F ∗N which arises in the above described manner together with an isomorphism ϑ : N from a finitely generated R-module N . It is shown in [Lyu97] that F -finite modules are an Abelian category which is closed under extensions, that local cohomology modules H i I(R) are F -finite modules, and that F -finite mod- ules enjoy a number of important finiteness results. For example they have only finitely many associated primes, and all Bass numbers are finite. From this definition it is immediate how to connect the Cartier modules with F -finite modules. The F -finite module attached to a Cartier module M is just the limit of X −→ F ∗(M ⊗ ω−1 One obtains the following Proposition [BB11]. M ⊗ ω−1 X ) −→ F 2∗(M ⊗ ω−1 X ) −→ · · · Proposition 9.1.7. For a regular ring R, the just described construction assigning to a coherent Cartier module M on R an F -finite module is an essentially surjective functor {coherent Cartier modules} −→ {F -finite modules} which sends nilpotent Cartier modules to zero. The induced functor {coherent Cartier crystals} ∼=−−→ {F -finite modules} is an equivalence of categories. Proof. All statements are shown in [BB11] but with the above preparations none of them is particularly difficult. (cid:3) Hence we obtain as an immediate consequence of Theorem 8.2.2 the following finiteness result for F -finite modules, which partially extends one of the main results of [Lyu97]: Theorem 9.1.8. Let R be regular and F -finite, then (a) F -finite modules over R-have finite length. (b) The Hom-sets in the category of F -finite modules are finite. (c) An F -finite module has only finitely many F -finite submodules. Part (a) of the theorem has been proven for R regular and of finite type over a regular local ring in [Lyu97], and for arbitrary F -finite schemes X in [BB06]. The latter results also are shown for regular rings (part (b) even without the F -finiteness assumption) in [Hoc07b]. Finally, let us state the aforementioned finiteness result for local cohomology modules. Theorem 9.1.9. Let M be an F -finite module, I ⊆ R an ideal in a regular ring, then H j is an F -finite R-module and hence has only finitely many associated primes. Proof. We only have to show that H j I (M ) is an F -finite module. The crucial step is to show that for f ∈ R we have that the localization Mf is also F -finite (cf. Exercise* 8.7). Once this is established, the Cech-complex finishes the proof. (cid:3) I (M ) 9.2. Contracting property of p−e-linear maps. In this section we point out a simple fact about p−e-linear map which has a number of interesting consequences. In particular we give an elementary proof of the finite length result for Cartier modules. The idea goes back at least to a paper of Anderson [And00] and says that a p−e-linear endomorphism reduces the degree in a graded context. For this we consider X = Spec S with S = k[x1, . . . , xn] a polynomial ring over a perfect field. Then we consider the filtration of S given by the finite-dimensional vector spaces Sd := khxi1 1 · · · xin n 0 ≤ ij ≤ d for j = 1, . . . , ni . p−1-LINEAR MAPS IN ALGEBRA AND GEOMETRY 49 Hence Sd is the k -- subspace of S freely generated by the monomials with degree ≤ d in each variable. One immediately verifies that S−∞ := 0, S0 = k, SdSd′ ⊆ Sd+d′, and Sd + S′ d ⊆ Smax d+d′ . For each choice of a set of generators m1, . . . , mk of an S module M we define the induced filtration on M given by M−∞ := 0 and Md = Sdhm1, . . . , mki. For m ∈ M we write δ(m) = d if and only if m ∈ Md \Md−1 and call δ = δM a gauge for M . One should think of the gauge δ as a substitute for a degree on M , and the contracting property of p−e-linear maps on M is measured in terms of the gauge δ. Spelling out the definition we see that δ(m) ≤ d if m can be written as a S-linear combination of the mi such that all coefficients are in Sd. S itself has a gauge, induced by the generator 1. We summarize the immediate properties of a gauge (the proof is left to the reader in Exercise 9.3): Lemma 9.2.1. Let M be finitely generated over S = k[x1, . . . , xn], and δ a gauge corresponding to some generators m1, . . . , mk of M . Then (a) δ(m) = −∞ if and only if m = 0. (b) Each Md is finite dimensional over k (since Sd is). (c) Sd Md = M (since the mi generate M ). (d) δ(m + m′) ≤ max{δ(m), δ(m′)} (e) δM (f m) ≤ δS(f ) + δM (m) Proposition 9.2.2 ([And00], Proposition 3). Let M be a finitely generated S-module, and δ = δM a gauge corresponding to some generators m1, . . . , mk of M and let ϕ : M −→ M be a p−e-linear map. Then there is a constant K such that for all m ∈ M : Furthermore, for all n ≥ 0 we have δ(ϕ(m)) ≤ δ(m) pe + K pe δ(ϕn(m)) ≤ δ(m) pne + K pe − 1 . l=1 flml with δS(fl) ≤ δ(m). For each l write uniquely n . Then Exercise 9.1 shows that δS(rl,i) ≤ ⌊δ(m)/pe⌋. 1 · · · xin Proof. By definition, we may write m =Pk fl =Pxi∈Sδ(m) rpe l,i xi with xi = xi1 Writing this out ϕ(m) = we consequently obtain kXl=1 Xxi∈Sδ(m) rl,iϕ(ximl) δ(ϕ(m)) ≤ max l,i {δS (rl,i) + δ(ϕ(ximl))} ≤ ⌊ δ(m) pe ⌋ + K pe taking for K = pe · maxl,i{δ(ximl)} we obtain the claimed inequality. The final inequality follows by applying the first inequality iteratively and then to use the (cid:3) geometric series (exercise!). This proposition has an important consequence about the generators of the images of a submodule under a p−e-linear map. 50 MANUEL BLICKLE AND KARL SCHWEDE Lemma 9.2.3. Let ϕ : M −→ M be a p−e-linear map on the finitely generated S-module M with gauge δ and bound K as in Proposition 9.2.2. Suppose that the S-submodule N ⊆ M is generated by elements with gauge ≤ d. Then ϕn(N ) ⊆ M is generated by elements of gauge at most d/pne + K/(pe − 1) + 1. Proof. If N is generated by n1, . . . , nt, then ϕn(N ) is generated by ϕ(xinj) where 0 ≤ i1, . . . , in ≤ pne − 1 and j = 1, . . . , t. Now, if each δ(nj) ≤ d then δ(ϕ(xinj)) ≤ δ(xinj) pne + K pe − 1 ≤ (pne − 1) + d pne + K pe − 1 ≤ 1 + d pne + K pe − 1 (cid:3) Corollary 9.2.4. Let ϕ : M −→ M be a Cartier module with gauge δ and bound K (as in Proposition 9.2.2). Then every Cartier submodule N ⊆ M with surjective structural map pe−1 +1 (indepen- ϕ : N −→→N is generated by elements in the finite dimensional k-vector space M K dently of N ). Proof. If N has surjective structural maps, then for each n we have ϕn(N ) = N . Since N is finitely generated it is generated by elements of some gauge ≤ d. By the above Lemma, we have hence for all n that N is generated by elements of gauge ≤ d/pne + K/(pe − 1) + 1. But for n big enough the first term is irrelevant (less than 1), and the result follows. (cid:3) Corollary 9.2.5. In a coherent Cartier module M there are no infinite proper chains of Cartier submodules Ni each with surjective structural map. Proof. Each Ni has generators in the finite dimensional k-vector space M K cannot be any infinite proper chains. pe−1 +1, hence there (cid:3) As we have alluded to (in Exercise 8.8) before, the fact that there are no infinite chains of Cartier submodules with surjective structural maps implies the existence of the test module τ (M ). By definition of being the smallest Cartier submodule of M which generically agrees with σ(M ) it is clear that τ (M ) has surjective structural map (since the image under the structural map would again be of that type). The intersection of two Cartier submodules agreeing generically with σ(M ) clearly also has this property. Now, the existence of τ (M ) follows from the stabilization of any chain of submodules generically agreeing with σ(M ) and with surjective structural maps, which we just showed. In the next section, Section 9.3, we will show this approach to Cartier modules via gauges also gives an elementary proof of the discreteness of jumping numbers for test ideals. We conclude this section with pointing out that our restriction to the polynomial ring S = k[x1, . . . , xn] is not very restrictive after all. The case of an arbitrary scheme X we may reduce to the affine case by considering an affine cover. Then any finite type k-algebra R = S/I is the quotient of a polynomial ring. Then we can use the Kashiwara-equivalence Proposition 8.1.5 to reduce to the case of the polynomial ring itself. Remark 9.2.6 (Historical discussion). The major source of inspiration to explore the contracting property of p−e-linear maps in [Bli09] came from a paper of Anderson [And00] where he uses this property to study L-functions mod p on varieties over Fp. The key observation there is that if ϕ : M −→ M is a p−e-linear map of R-modules (say R of finite type over Fp) then there is a finite dimensional Fp-subspace into which every element of M is eventually contracted by iterated application of ϕ. This allows him, inspired by Tate's work [Tat68] to develop a trace calculus for these operators. This is then used to show the rationality of L-functions mod p attached to a finitely generated R-module M with a left action of Frobenius F on M . In fact, he shows that if R is the polynomial ring and M is projective, this L-function is equal to p−1-LINEAR MAPS IN ALGEBRA AND GEOMETRY 51 the characteristic polynomial (or its inverse) of the action of the dual of F on M ∨. This dual is a Cartier linear endomorphism of M ∨ and the characteristic polynomial is defined via the important contracting property of Cartier linear maps. 9.3. Algebras of maps and the test ideal. Suppose that X = Spec R is an affine variety (for simplicity). Previously we considered finitely generated R-modules M and p−e-linear maps ϕ : M −→ M . Unless M = ωR (or is obtained functorially from T : ωY −→ ωY from some other variety Y ), there probably is no natural choice of ϕ. The obvious solution is to choose all possible ϕ, see [Sch11b, Bli09] and cf. [LS01] for a dual formulation. For any finitely generated module M , we set Ende(M ) to be the set of p−e-linear maps from M to M . In other words, Ende(M ) is just HomR(F e ∗ M, M ). Of course Ende(M ) has an R-module structure via both the source and target R-module structures. Notice that if ϕ ∈ Ende(M ) and ψ ∈ Endd(M ), then we can form the composition ψ ◦ ϕ ∈ Ende+d(M ). Thus End∗(M ) = ⊕e≥0 Ende(M ) forms a non-commutative graded ring. Unfortunately, the ring End0(M ) is often too big and so we set C M to denote the image of R inside End0(M ) via the 0 natural map that sends r ∈ R to the multiplication by r map on M . Definition 9.3.1 (Cartier Algebras). An (abstract) Cartier algebra over R 10 is a N-graded for all ϕe ∈ Ce and r ∈ R and furthermore Ce satisfying the rule r · ϕe = ϕe · rpe ring C =Le≥0 such that C0 ∼= R/I for some ideal I. Example 9.3.2. Suppose that M is a finitely generated R-module. The total Cartier algebra on M , denoted C M , is the following graded subring of End∗(M ). C M := C M 0 ⊕ Me>0 Ende(M )! =Me≥0 C M e . It is obviously a Cartier algebra. A Cartier-subalgebra (on M) is any graded subring C ⊆ C M such that [C ]0 = C M 0 . With the above definitions, if C is an (abstract) Cartier algebra, and M is any left-C -module, then there is a natural map C −→ C M , the image of which is a Cartier-subalgebra on M . Conversely, note that any Cartier-subalgebra C ⊆ C M acts on M by the application of functions. In particular, M is also a C -module. Remark 9.3.3. Most commonly, we will consider C R, in which case C R End0(R) automatically. 0 = HomR(R, R) = Now suppose that C is a Cartier-algebra and that M is a left C -module (or that C is a Cartier-submodule on M ), we use C+ to denote ⊕e>0Ce. It is easy to see that C+ is a 2-sided ideal. For any C -submodule N ⊆ M , we define C+N := hϕ(x) x ∈ N, ϕ ∈ Ce for some e > 0iR ⊆ N to be the submodule generated by all ϕ(x) for homogeneous ϕ ∈ C+ and n ∈ N . We set (C+)nN := C+(C+(· · · C+ (N ))) ⊆ N. n-times {z } A crucial step from dealing with an algebra of Cartier linear operators as opposed to a single one, is to establish the right notion of nilpotence. With following definition the theory develops in surprising analogy to the single operator case dealt with above. Definition 9.3.4. We say that N is C -nilpotent if (C+)nN = 0 for some n > 0. 10It is important to note that while we call it an algebra, it is not generally an R-algebra because R is not central. 52 MANUEL BLICKLE AND KARL SCHWEDE It is obvious we have a chain of inequalities: (20) N ⊇ C+N ⊇ (C+)2N ⊇ · · · ⊇ (C+)iN ⊇ (C+)i+1N ⊇ · · · The following remarkable theorem about this chain generalizes Proposition 8.1.4 above. Theorem 9.3.5. [Bli09, Proposition 2.14] Suppose that M is a finitely generated R-module that is also a left C -module for some Cartier algebra C . Then (C+)nM = (C+)n+1M for all n ≫ 0. In other words, the chain of submodules in (20) eventually stabilizes. Proof. The proof is similar to that of Proposition 8.1.4 and left to the reader in Exercise* 9.6. (cid:3) As an immediate corollary we obtain: Corollary 9.3.6. [Bli09, Corollary 2.14] Let M be a finitely generated R-module that is also an C -module for some Cartier algebra C . Then there is a unique C -submodule σ(M ) ⊆ M such that (a) the quotient M/σ(M ) is nilpotent, and (b) C+σ(M ) = σ(M ) and so σ(M ) does not have nilpotent quotients. Proof. Set σ(M ) = (C+)nM for n ≫ 0, then verify the statements in Exercise 9.8. (cid:3) Suppose that M is a finitely generated R-module and a left C -module. We can now define a notion of the test ideal on M . Definition 9.3.7. Suppose that M and C are as above. Then we define the test submodule τ (M, C ) to be the unique smallest submodule N of M which (a) is a C -module, (b) which satisfies (σ(M ))η = Nη for every minimal prime of R.11 if it exists. The existence of τ (M, C ) is known in many important cases, but not in all generality. It is known to exist if R is of finite type over a field (or a localization of such), or if C is generated by a single operator, see [Bli09, Theorem 4.13, Corollary 3.18]. It is also known to exist if M = R by the same argument as Proposition 9.3.10 below. For the rest of the section, we consider C R, the total Cartier algebra on R, and subalgebras of it. Indeed, a common way to construct a Cartier algebra is as follows. Definition 9.3.8. Suppose that R is a normal domain with X = Spec R. Suppose further that ∆ ≥ 0 is an effective Q-divisor, a ⊆ R is a nonzero ideal and t ≥ 0 is a real number. Then we define the following Cartier subalgebra of C R. For each e ≥ 0 first identify HomR(F e ∗ R, R) with C R to be the subset e and fix C ∆ e HomR(F e ∗ R(⌈(pe − 1)∆⌉), R) ⊆ HomR(F e ∗ R, R) = C R e . Here R(⌈(pe − 1)∆⌉) = Γ(X, OX (⌈(pe − 1)∆⌉)). It follows that C ∆ :=Me≥0 C ∆ e is a Cartier subalgebra of C R (the details will be left as Exercise 9.9). 11This definition differs slightly from the original one given in [Bli09] where one requires equality for every minimal prime of σ(M ) instead of R. Though this yields different results in general, in light of the Kashiwara equivalence Proposition 8.1.5 the respective theories imply each other. p−1-LINEAR MAPS IN ALGEBRA AND GEOMETRY 53 Furthermore, we can form C ∆,at composition, in other words C ∆,at ). Again the direct sum e e e · a := C ∆ is identified with HomR(F e ⌈t(pe−1)⌉ (where multiplication on the right is pre- ⌈t(pe−1)⌉) ∗ R(⌈(pe − 1)∆⌉), R) · (F e ∗ a t C ∆,a :=Me≥0 t C ∆,a e is a Cartier-subalgebra of C R, see Exercise 9.9. With these definitions, we can now define the test ideal τ (R; ∆, a t) := τ (R, C ∆,at ) [Sch11b, Bli09]. Remark 9.3.9. Test ideals (with ∆ = 0 and a = R) were originally introduced by Hochster and Huneke in their theory of tight closure [HH90]. In fact, what we call the test ideal is often called τ coincide in general [LS99, LS01]. the big test ideal [Hoc07a] and is denoted byeτ or τb. This object though is better behaved with respect to geometric operations (such as localization Exercise 9.11). It is conjectured thateτ and was originally defined in [HY03] (andeτ (R; a t) t) was studied in [HT04]). For ∆ 6= 0, τ (R; ∆) was Even with ∆ 6= 0 and a 6= R, this definition we gave is not the original one. For a 6= R, τ (R; a Proposition 9.3.10. Suppose R is a normal domain. The test ideal τ (R, C ∆,a exists. introduced in [Tak04]. ) = τ (R; ∆, a t) t Proof. The main point is the following Lemma, which is a generalization of a result of Hochster and Huneke. Lemma 9.3.11. [HH90, Section 6] [Sch11b, Lemma 3.21] There exists an element 0 6= c ∈ R such that for every 0 6= d ∈ R, there exists e > 0 such that c ∈ C ∆,a (dR). t e Now choose c as in the Lemma, and it follows that c ∈ I for any non-zero C ∆,at -submodule I ⊆ R. However, is evidently the smallest C ∆,at C ∆,at e (Rc) Xe≥0 submodule containing c. (cid:3) One of the aspects of the test ideal which has attracted the most interest over the past few t) changes as t varies. First we mention the following lemma years is how the test ideal τ (R; ∆, a which serves as a baseline for how the test ideal behaves. Lemma 9.3.12. [MTW05, Remark 2.12], [BMS08, Proposition 2.14], [BSTZ10, Lemma 3.23] With notation as above, for every real number t ≥ 0, there exists an ε > 0 such that τ (R; ∆, a t) = τ (R; ∆, a s) for every s ∈ [t, t + ε]. Proof. The containment ⊇ is obvious. A substantial hint is given in Exercise* 9.12. (cid:3) Because of this, we make the following definition: Definition 9.3.13 (F -jumping numbers). Suppose that (R, ∆, a t > 0 is called an F -jumping number if t) are as above. Then a number for all 1 ≫ ε > 0. τ (R; ∆, a t) 6= τ (R; ∆, a t−ε) 54 MANUEL BLICKLE AND KARL SCHWEDE Based on the above Lemma, and a connection between test ideals and multiplier ideals [HY03, Tak04] it is natural to expect that the set of jumping numbers for the test ideal is discrete. In the case that X is smooth this was shown to be the case in [BMS08] and [BMS09]. The singular case was obtained in [BSTZ10], see also [Har06, KLZ09, ST08, TT08, STZ12, `AMBZ12]. We will outline here an elementary proof based on the contracting property of p−e linear maps that was investigated in the preceding section. In order to be able to handle not only a single p−e-linear map but a whole Cartier algebra, we need to generalize the results on gauge bounds obtained above slightly. To keep things simple we will consider a Cartier algebra of the type C =M R · ϕn ⌈t(pne−1)⌉ a where a is an ideal in R, t ≥ 0 is a real number and ϕ is a single p−e-linear operator on R. This for (pe − 1)(KR + ∆) is a Cartier divisor (i.e. the pair (R, ∆) is essentially the case C = C ∆,a is Q-Cartier with index not divisible by p). We first state a generalization of Lemma 9.2.3 to this context. t Lemma 9.3.14. Let C be the Cartier subalgebra of C R generated by ϕ, a p−e-linear map on R = k[x1, . . . , xn]/I. Let M be a coherent C module and suppose that for all m ∈ M and n > 0 one has δ(ϕn(m)) ≤ δ(m) pne + K pe − 1 for some bound K ≥ 0 as in Proposition 9.2.2. Then, if a ⊆ R is an ideal generated by element of gauge ≤ d, and N ⊆ M is a R-submodule generated by elements of gauge ≤ D, then (C at + )n(N ) ⊆ N is generated by elements of gauge ≤ D pne + K pe−1 + td + 1. Proof. Note that R has a set of generators over Rpne−1 each of gauge ≤ pne −1 (the images of the ⌈t(pne−1)⌉ relevant monomials of k[x1, . . . , xn] in R will do fine). Next, it is easy to check that a is generated by element with gauge ≤ tdpne + 1. Hence, as in the proof of Lemma 9.2.3 the t + )n is generated as a left R-modules by elements ψ of the form ψ = ϕl · b · a where ideal (C a ⌈t(pne−1)⌉). l ≥ n and b (resp. a) is one of the just described generators of R over Rpln (resp. of a Ranging over all such ψ and a set of R-generators m of N we see that (C at + )n(N ) is generated by elements of the form ψ(m). Now we just compute δ(ψ(m)) = δ(ϕl · b · a · m) ≤ ≤ δ(bam) ple + δ(m) pne − 1 + (ple − 1) + (tdple + 1) ple + K pe − 1 K pe − 1 ≤ δ(m) ple + K pe − 1 + td + 1 This shows the claim. (cid:3) Corollary 9.3.15. With notation as in the Lemma, let N be an R-submodule of M such that C at + (N ) = N . Then N is generated by elements of gauge ≤ K t For T ≥ 0 there no infinite chains of R-submodules N of M for which C a pe−1 + td + 1. + (N ) = N for some t < T . Proof. Clearly, C at + )n(N ) = N for all n and hence the first claim follows from the preceding Lemma. The second claim follows from the first one since each such pe−1 +T d+1, hence there N is generated by elements in the finite dimensional vector space M≤ K cannot be infinite chains. + (N ) = N implies that (C at (cid:3) Now, the discreteness of the jumping numbers for the test ideal is an immediate consequence. Theorem 9.3.16. For (R, ∆, a t) as above, the F -jumping numbers form a discrete subset of Q. p−1-LINEAR MAPS IN ALGEBRA AND GEOMETRY 55 Proof. In the case that (pe − 1)(KX + ∆) is Cartier, the Cartier algebra C ∆ is of the form t′ t) ⊇ τ (R, ∆, a ) considered above. Since each test ideals τ (R, ∆, a for t′ ≥ t and C ∆,at + τ (R, ∆, at) = τ (R, ∆, at), The preceding corollary shows that there are only finitely for t below a fixed bound T . Hence the jumping numbers must be discrete. The general case is similar or can be reduced to this case by using the methods of Section 7.3, see [ST10a, STZ12]. (cid:3) t) has the properties τ (R, ∆, a 9.4. Exercises. Exercise 9.1. Let f ∈ S = k[x1, . . . , xn] with k perfect and with gauge δ corresponding to the generator 1. Show that, if δ(f ) ≤ d and writing uniquely spe i xi f = Xxi∈Spe−1 one has δ(si) ≤ ⌊d/pe⌋. (Here we used multi-exponent notation xi as shorthand for xi1 1 · · · xin n ) Exercise 9.2. Use the duality for finite morphisms to prove Lemma 9.1.1. Exercise 9.3. Prove Lemma 9.2.1. Exercise 9.4. Let R ֒−→ S be a module-finite and flat ring extension. Show that the natural map is an isomorphism. Derive from this the statement of Lemma 9.1.5. HomR(S, R) ⊗R M ϕ⊗n7→(r7→ϕ(r)n) −−−−−−−−−−−−→ HomR(S, M ) Exercise 9.5. Consider the example of a Cartier structure κ on the polynomial ring k[x] given by sending 1 7→ xt and x, x2, . . . , xp−1 7→ 0. Show that δ(κ(f )) ≤ δ(f )/p + t where δ is the gauge on k[x] induced by the generator 1 ∈ k[x]. Exercise* 9.6. Prove Theorem 9.3.5 by using the same strategy as in Proposition 8.1.4. Exercise 9.7. Let (R, m) be complete local of dimension d and denote by F : H d the natural Frobenius action. Show that any left action ϕ on H d form ϕ = r · F for some r ∈ R. m(R) m(R) of the Frobenius is of the m(R) −→ H d Exercise 9.8. Prove Corollary 9.3.6. Exercise 9.9. With notation as in Definition 9.3.8, show that C ∆ and C ∆,a algebras of C R. For a proof, see [Sch11b, Remark 3.10]. t are Cartier sub- Exercise 9.10. Suppose that R is a normal local domain and that ∆ ≥ 0 is a Q-divisor on X = Spec R such that KX + ∆ is Q-Cartier with index not divisible by p > 0. Prove that C ∆ is a finitely generated ring over C ∆ 0 = R. Hint: Show that HomR(F e ∗ R(⌈(pe−1)∆⌉), R) ∼= F e ∗ R for some e > 0 and then use Exercise 4.2. For additional discussion see [Sch11b, Section 4]. Exercise 9.11. Suppose that R is a normal domain, W ⊆ R is a multiplicative system, ∆ ≥ 0 is a Q-divisor on X = Spec R, a ⊆ R is a nonzero ideal and t ≥ 0 is a real number. Set U = Spec(W −1R) ⊆ Spec R = X. Prove that W −1τ (R; ∆, a t) = τ (W −1R; ∆U , (W −1 a)t). Exercise* 9.12. Prove Lemma 9.3.12. Hint: Use the description of τ (R; ∆, a t) from the proof of Proposition 9.3.10. Also use the fact that R is Noetherian to see that the sum from Proposition 9.3.10 is a finite sum (e = 0 to m). Now notice that if c works in that sum, then so does bc where 0 6= b ∈ a. Set ε = 1 pm . 56 MANUEL BLICKLE AND KARL SCHWEDE Exercise* 9.13. Suppose that R is a normal ring and that X = Spec R. Consider the anti- canonical ring K :=Mn≥0 OX(−nKX ). Set KF :=Le≥0 OX ((1 − pe)KX ) to be the summand of K made up of terms of degree pe − 1 for some e ≥ 0. This is not a subring of K. However, define a non-commutative multiplication on KF as follows. If α ∈ OX ((1 − pe)KX ) and β ∈ OX ((1 − pd)KX ) then define α ⋆ β = αpd β ∈ OX (((1 − pe)pd + pe)KX ) = OX ((1 − pe+d)KX ). With this ring operation, prove that KF is isomorphic to C R. Appendix A. Reflexification of sheaves and Weil divisors In this section, we briefly recall basic properties of reflexive sheaves and Weil divisors on normal varieties. This material is all "well known" but there isn't a good source for it in the literature (we note that it is certainly assumed in [KM98]). We note that substantial generaliza- tions of all this material (and complete proofs) can be found in [Har94]. As before, all schemes are of finite type over a field (or localizations or completions of such schemes). We assume the reader is familiar with the basic notion of depth and Sn (Serre's nth condition) and the connec- tions with local cohomology / cohomology with support. See for example [Har77, Chapter III, Exercises in Section 3], [BH93] or [Har67]. A.1. Reflexive sheaves. Given a coherent sheaf F on any scheme X, there is the following (dualizing) operation: F ∨ = H omOX (F , OX ). Furthermore there is a natural map from F to the double-dual, F → (F ∨)∨. Definition A.1.1. If this map is an isomorphism, we say that F is reflexive (or more specifically that it is OX -reflexive). Note that if a sheaf is reflexive, it is also coherent (by definition). If X = Spec R and M is a coherent R-module, we say that M is reflexive if the corresponding sheaf is reflexive (equivalently, if M −→ HomR(HomR(M, R), R) is an isomorphism). Notice first that any locally free sheaf is reflexive. But there are other reflexive sheaves as well. If one is careful, one can check that hx, zi ⊆ k[x, y, z]/(xy − z2) corresponds to a reflexive ideal sheaf after taking Spec, Exercise A.1. There are a few basic facts about reflexive sheaves that should be mentioned. We now limit ourselves to varieties (i.e.integral schemes) which makes dealing with torsion much easier. One can do analogues of the following in more general situations (say for reduced schemes), but the statements become much more involved. Lemma A.1.2. Suppose that X is a variety and suppose that F is a coherent sheaf on X. Then F ∨ is torsion-free. (That is, if U ⊂ X is open and 0 6= r ∈ OX(U ) and 0 6= z ∈ F ∨(U ), then rz 6= 0). In particular, a reflexive sheaf is torsion-free. Note that a torsion-free sheaf is necessarily S1 (any nonzero element makes up a rather short regular sequence). Lemma A.1.3. Suppose that X is a variety and that F is a torsion-free coherent sheaf. Then the natural map α : F → F ∨∨ is injective. Lemma A.1.4. [Har80, Proposition 1.1] A coherent sheaf F on a quasi-projective variety X is reflexive if and only if it can be included in an exact sequence where E is locally free and G is torsion-free. We note that the OX-dual of any coherent sheaf is always reflexive. 0 → F → E → G → 0 p−1-LINEAR MAPS IN ALGEBRA AND GEOMETRY 57 Theorem A.1.5. If F is a coherent sheaf on a variety X, then F ∨ is reflexive. More generally, if F is coherent and G is reflexive, then H omOX (F , G ) is reflexive. We now come to a very useful criterion for checking whether a sheaf is reflexive. Theorem A.1.6. [Har94, Theorem 1.9] Suppose that X is a normal (not necessarily quasi- projective) variety and that F is a coherent sheaf on X such that Supp(F ) = X. Then F is S2 if and only if F is reflexive. The key reason why the previous criterion is so useful is the Hartog's phenomenon associated with S2 sheaves. Corollary A.1.7. Let X be a integral, normal (not necessarily quasi-projective) variety and suppose that F is a reflexive sheaf on X (defined as above). Let Y ⊂ X be a closed subset of codimension ≥ 2 and set U = X\Y . Then if i : U → X is the natural inclusion, then the natural map F → i∗F U is an isomorphism. Corollary A.1.8. Suppose that F is a reflexive sheaf on U ⊆ X (where X is as above) such that X − U is codimension two. Let us denote by i : U → X the inclusion. Then i∗F is a reflexive sheaf on X. A.2. Divisors. Let X be a normal variety of finite type over a field. By a Weil divisor on X, we mean a formal sum of integral codimension 1 subschemes (prime divisors). Recall that a divisor D is called effective if the coefficients of D are nonnegative. Just like in the regular case, each prime divisor D corresponds to some discrete valuation vD of the fraction field of X (although the reverse direction is not true). Definition A.2.1. Choose f ∈ K (X), f 6= 0. We define the principal divisor div(f ) as in the regular case: div(f ) = ΣivDi(f )Di. Likewise, we say that two Weil divisors D1 and D2 are linearly equivalent, if D1 − D2 is principal. Definition A.2.2. Given a divisor D, we define OX(D) be the sheaf associated to the following rule: Γ(V, OX (D)) = {f ∈ K (X) div(f )V + DV ≥ 0} A divisor D is called Cartier if OX (D) is an invertible sheaf. It is called Q-Cartier if nD is Cartier for some n > 0. Note that D is effective if and only if OX(D) ⊇ OX . Proposition A.2.3. Suppose that D is a prime divisor, then OX (−D) = ID, the ideal sheaf defining D. Furthermore, if D is any divisor, then OX (D) is reflexive. Proof. We first show the equality. The object defined above is clearly a sheaf. We will prove the equality of the sheaves in the setting where U is affine. Then Γ(U, OX (D)) is just the functions in OX which vanish to order at least 1 along D, in other words the ideal of D. We now want to show that this sheaf is reflexive (or equivalently, that it is S2). First notice that clearly if U is the regular locus of X, then Γ(V ∩ U, OX (D)) ∼= Γ(V, OX (D)) for any open set V . This is because V ∩ U = U \{non-regular locus}, the non-regular locus is codimension 2, and the sections of OX(D) obviously do not change when removing a codimension 2 subset. This implies that the natural map OX (D) → i∗OX (D)U is an isomorphism, but then we notice that OX(D)U is reflexive (since it is invertible) and thus, by corollary Corollary A.1.8, OX (D) is also reflexive. (cid:3) We now list some basic properties of rank-1 reflexive sheaves which completely link their behavior to divisors. Proposition A.2.4. Suppose that X is a normal variety. Then: 58 MANUEL BLICKLE AND KARL SCHWEDE (a) If X is regular, then every reflexive rank-1 sheaf F on X is invertible. [Har80, Proposi- tion 1.9] (b) Every rank one reflexive sheaf F on a normal scheme X embeds as a subsheaf of K (X). (c) Any reflexive rank 1 subsheaf of K (X) is OX (D) for some (uniquely determined) divisor D. Proof. Left to the reader in Exercise A.4. (cid:3) The addition operations for divisors translates into the tensor of the associated sheaves, up to reflexification. Proposition A.2.5. Suppose that X is a normal variety and D and E are divisors on X. Then (a) If E is Cartier, then OX (D) ⊗ OX (E) ∼= OX (D + E) (b) In general, OX(D + E) ∼= (OX (D) ⊗ OX (E))∨∨ (c) OX(−D) = H omOX (OX (D), OX ) = OX (−D)∨ Proof. Left to the reader, see Exercise A.5 (cid:3) Finally, we mention a result relating sections and linearly equivalent divisors, which will be a key part of this paper. Theorem A.2.6. Suppose that X is a normal variety and D is a Weil divisor on X. Then there is a bijection between the following two sets (cid:26) Effective divisors E linearly equivalent to D (cid:27) ←→(cid:26) Nonzero sections γ ∈ H 0(X, OX (D)) modulo equivalence (cid:27) where we define γ and γ′ in H 0(X, OX (D)) to be equivalent if there exists a unit u ∈ H 0(X, OX ) such that uγ = γ′. Proof. Set M = OX (D). The choice γ induces an embedding iγ : M ֒→ K (X) which sends γ to 1. Thus γ induces a divisor via Proposition A.2.4. It follows from the same argument that γ and γ′ induce the same divisor if and only if iγ and iγ ′ have the same image in K (X). But this happens if and only if γ and γ′ are unit multiplies of one another. (cid:3) A.3. Exercises. Exercise A.1. Show that hx, zi ∈ k[x, y, z]/hxy − z2i corresponds to a reflexive ideal sheaf after taking Spec. Exercise A.2. Which of the following k[x, y] = R-modules are reflexive? If a module is not reflexive, compute its double dual M ∨∨. (a) The ideal hxi. (b) The ideal hx, yi. (c) The module R/hx, yi. (d) The module R/hxi. (e) The ideal hx2, xyi = hx, yi2 ∩ hyi. Exercise A.3. Suppose that π : Y −→ X is a finite dominant map of normal varieties and F is a coherent sheaf on Y . Then F is reflexive on Y if and only if π∗F is reflexive on X. Hint: Use the fact that you can check whether a sheaf is reflexive by checking whether it is S2. Then use the criterion for checking depth via local cohomology. Exercise A.4. Prove Proposition A.2.4. Exercise A.5. Prove Proposition A.2.5. p−1-LINEAR MAPS IN ALGEBRA AND GEOMETRY 59 [ `AMBZ12] J. `Alvarez Montaner, A. Boix, and S. Zarzuela: Frobenius and Cartier algebras of Stanley- Reisner rings, J. Algebra 358 (2012), no. 15, 162 -- 177. [And00] G. W. Anderson: An elementary approach to L-functions mod p, J. Number Theory 80 (2000), References [Bli08] [Bli09] [BB11] no. 2, 291 -- 303. M. Blickle: Minimal γ -- sheaves, Algebra and Number Theory 2 (2008), no. 3, 347 -- 368. M. Blickle: Test ideals via algebras of p−e-liner maps, arXiv:0912.2255, to appear in J. Algebraic Geom. M. Blickle and G. Bockle: Cartier modules: finiteness results, J. Reine Angew. Math. 661 (2011), 85 -- 123. 2863904 M. Blickle and G. Bockle: Cartier Crystals, manuscript in preparation, started 2006. [BB06] [BMS08] M. Blickle, M. Mustat¸a, and K. Smith: Discreteness and rationality of F-thresholds, Michigan Math. J. 57 (2008), 43 -- 61. [BMS09] M. Blickle, M. Mustat¸a, and K. E. Smith: F -thresholds of hypersurfaces, Trans. Amer. Math. Soc. 361 (2009), no. 12, 6549 -- 6565. 2538604 (2011a:13006) [BSTZ10] M. Blickle, K. Schwede, S. Takagi, and W. Zhang: Discreteness and rationality of F -jumping [BP09] [BK05] numbers on singular varieties, Math. Ann. 347 (2010), no. 4, 917 -- 949. 2658149 G. Bockle and R. Pink: Cohomological Theory of crystals over function fields, European Mathe- matical Society, Zurich, 2009, Tracts in Mathematics. M. Brion and S. Kumar: Frobenius splitting methods in geometry and representation theory, Progress in Mathematics, vol. 231, Birkhauser Boston Inc., Boston, MA, 2005. MR2107324 (2005k:14104) [BH93] W. Bruns and J. Herzog: Cohen-Macaulay rings, Cambridge Studies in Advanced Mathematics, [Car57] [Con00] [dS97] [DI87] [EK04] [EH08] [EV92] [Fed83] [Fuj83] [Gab04] [Gab62] vol. 39, Cambridge University Press, Cambridge, 1993. MR1251956 (95h:13020) P. Cartier: Une nouvelle op´eration sur les formes diff´erentielles, C. R. Acad. Sci. Paris 244 (1957), 426 -- 428. 0084497 (18,870b) B. Conrad: Grothendieck duality and base change, Lecture Notes in Mathematics, vol. 1750, Springer- Verlag, Berlin, 2000. MR1804902 (2002d:14025) B. de Smit: The different and differentials of local fields with imperfect residue fields, Proc. Edinburgh Math. Soc. (2) 40 (1997), no. 2, 353 -- 365. MR1454030 (98d:11148) P. Deligne and L. Illusie: Relvements modulo $p2$ et dcomposition du complexe de de rham, Inventiones Mathematica 89 (1987), 247 -- 270. M. Emerton and M. Kisin: Riemann -- Hilbert correspondence for unit F -crystals, Ast´erisque 293 (2004), vi+257 pp. F. Enescu and M. Hochster: The Frobenius structure of local cohomology, Algebra Number Theory 2 (2008), no. 7, 721 -- 754. MR2460693 (2009i:13009) H. Esnault and E. Viehweg: Lectures on vanishing theorems, DMV Seminar, vol. 20, Birkhauser Verlag, Basel, 1992. MR1193913 (94a:14017) R. Fedder: F -purity and rational singularity, Trans. Amer. Math. Soc. 278 (1983), no. 2, 461 -- 480. MR701505 (84h:13031) T. Fujita: Vanishing theorems for semipositive line bundles, Algebraic geometry (Tokyo/Kyoto, 1982), Lecture Notes in Math., vol. 1016, Springer, Berlin, 1983, pp. 519 -- 528. 726440 (85g:14023) O. Gabber: Notes on some t-structures, Geometric aspects of Dwork theory. Vol. I, II, Walter de Gruyter GmbH & Co. KG, Berlin, 2004, pp. 711 -- 734. P. Gabriel: Des cat´egories ab´eliennes, Bull. Soc. Math. France 90 (1962), 323 -- 448. 0232821 (38 #1144) [Hab80] W. J. Haboush: A short proof of the Kempf vanishing theorem, Invent. Math. 56 (1980), no. 2, [Hac12] [Har05] [Har06] [HT04] [HW02] 109 -- 112. 558862 (81g:14008) C. D. Hacon: On log canonical inversion of adjunction, arXiv:1202.0491, to appear in the PEMS, volume in honour of V. Shokurov. N. Hara: A characteristic p analog of multiplier ideals and applications, Comm. Algebra 33 (2005), no. 10, 3375 -- 3388. MR2175438 (2006f:13006) N. Hara: F-pure thresholds and F-jumping exponents in dimension two, Math. Res. Lett. 13 (2006), no. 5-6, 747 -- 760, With an appendix by Paul Monsky. MR2280772 N. Hara and S. Takagi: On a generalization of test ideals, Nagoya Math. J. 175 (2004), 59 -- 74. MR2085311 (2005g:13009) N. Hara and K.-I. Watanabe: F-regular and F-pure rings vs. log terminal and log canonical singu- larities, J. Algebraic Geom. 11 (2002), no. 2, 363 -- 392. MR1874118 (2002k:13009) 60 [HY03] MANUEL BLICKLE AND KARL SCHWEDE N. Hara and K.-I. Yoshida: A generalization of tight closure and multiplier ideals, Trans. Amer. Math. Soc. 355 (2003), no. 8, 3143 -- 3174 (electronic). MR1974679 (2004i:13003) [Har66a] R. Hartshorne: A property of A-sequences, Bull. Soc. Math. France 94 (1966), 61 -- 65. 0209279 (35 #181) [Har77] [Har67] [Har66b] R. Hartshorne: Residues and duality, Lecture notes of a seminar on the work of A. Grothendieck, given at Harvard 1963/64. With an appendix by P. Deligne. Lecture Notes in Mathematics, No. 20, Springer-Verlag, Berlin, 1966. MR0222093 (36 #5145) R. Hartshorne: Local cohomology, A seminar given by A. Grothendieck, Harvard University, Fall, vol. 1961, Springer-Verlag, Berlin, 1967. MR0224620 (37 #219) R. Hartshorne: Algebraic geometry, Springer-Verlag, New York, 1977, Graduate Texts in Mathe- matics, No. 52. MR0463157 (57 #3116) R. Hartshorne: Stable reflexive sheaves, Math. Ann. 254 (1980), no. 2, 121 -- 176. 597077 (82b:14011) R. Hartshorne: Generalized divisors on Gorenstein schemes, Proceedings of Conference on Algebraic Geometry and Ring Theory in honor of Michael Artin, Part III (Antwerp, 1992), vol. 8, 1994, pp. 287 -- 339. MR1291023 (95k:14008) R. Hartshorne and R. Speiser: Local cohomological dimension in characteristic p, Ann. of Math. (2) 105 (1977), no. 1, 45 -- 79. MR0441962 (56 #353) [Har80] [Har94] [HS77] [Hoc07a] M. Hochster: Foundations of tight closure theory, lecture notes from a course taught on the Univer- sity of Michigan Fall 2007 (2007). [HH90] [Hun96] [Hoc07b] M. Hochster: Some finiteness properties of Lyubeznik's F -modules, Algebra, geometry and their interactions, Contemp. Math., vol. 448, Amer. Math. Soc., Providence, RI, 2007, pp. 119 -- 127. 2389238 (2009c:13010) M. Hochster and C. Huneke: Tight closure, invariant theory, and the Brian¸con-Skoda theorem, J. Amer. Math. Soc. 3 (1990), no. 1, 31 -- 116. MR1017784 (91g:13010) M. Hochster and J. L. Roberts: The purity of the Frobenius and local cohomology, Advances in Math. 21 (1976), no. 2, 117 -- 172. MR0417172 (54 #5230) C. Huneke: Tight closure and its applications, CBMS Regional Conference Series in Mathematics, vol. 88, Published for the Conference Board of the Mathematical Sciences, Washington, DC, 1996, With an appendix by Melvin Hochster. MR1377268 (96m:13001) C. Huneke and I. Swanson: Integral closure of ideals, rings, and modules, London Mathemati- cal Society Lecture Note Series, vol. 336, Cambridge University Press, Cambridge, 2006. MR2266432 (2008m:13013) N. M. Katz: Nilpotent connections and the monodromy theorem: Applications of a result of Turrittin, Inst. Hautes ´Etudes Sci. Publ. Math. (1970), no. 39, 175 -- 232. 0291177 (45 #271) [Kat70] [HR76] [HS06] [KLZ09] M. Katzman, G. Lyubeznik, and W. Zhang: On the discreteness and rationality of F -jumping coefficients, J. Algebra 322 (2009), no. 9, 3238 -- 3247. 2567418 (2011c:13005) [Kaw07] M. Kawakita: Inversion of adjunction on log canonicity, Invent. Math. 167 (2007), no. 1, 129 -- 133. MR2264806 (2008a:14025) [Kaw82] Y. Kawamata: A generalization of Kodaira-Ramanujam's vanishing theorem, Math. Ann. 261 (1982), no. 1, 43 -- 46. MR675204 (84i:14022) [Kaw98] Y. Kawamata: Subadjunction of log canonical divisors. II, Amer. J. Math. 120 (1998), no. 5, 893 -- 899. [Kee08] [Kc92] [KM98] [KM09] [Kun69] [Kun86] MR1646046 (2000d:14020) D. S. Keeler: Fujita's conjecture and Frobenius amplitude, Amer. J. Math. 130 (2008), no. 5, 1327 -- 1336. 2450210 (2009i:14006) J. Koll´ar and 14 coauthors: Flips and abundance for algebraic threefolds, Soci´et´e Math´ematique de France, Paris, 1992, Papers from the Second Summer Seminar on Algebraic Geometry held at the University of Utah, Salt Lake City, Utah, August 1991, Ast´erisque No. 211 (1992). MR1225842 (94f:14013) J. Koll´ar and S. Mori: Birational geometry of algebraic varieties, Cambridge Tracts in Mathemat- ics, vol. 134, Cambridge University Press, Cambridge, 1998, With the collaboration of C. H. Clemens and A. Corti, Translated from the 1998 Japanese original. MR1658959 (2000b:14018) S. Kumar and V. B. Mehta: Finiteness of the number of compatibly split subvarieties, Int. Math. Res. Not. IMRN (2009), no. 19, 3595 -- 3597. 2539185 (2010j:13012) E. Kunz: Characterizations of regular local rings for characteristic p, Amer. J. Math. 91 (1969), 772 -- 784. MR0252389 (40 #5609) E. Kunz: Kahler differentials, Advanced Lectures in Mathematics, Friedr. Vieweg & Sohn, Braun- schweig, 1986. MR864975 (88e:14025) [Laz04a] R. Lazarsfeld: Positivity in algebraic geometry. I, Ergebnisse der Mathematik und ihrer Grenzge- biete. 3. Folge. A Series of Modern Surveys in Mathematics [Results in Mathematics and Related p−1-LINEAR MAPS IN ALGEBRA AND GEOMETRY 61 Areas. 3rd Series. A Series of Modern Surveys in Mathematics], vol. 48, Springer-Verlag, Berlin, 2004, Classical setting: line bundles and linear series. MR2095471 (2005k:14001a) [Lec64] [LS99] [LS01] [Lip06] [Lyu97] [Laz04b] R. Lazarsfeld: Positivity in algebraic geometry. II, Ergebnisse der Mathematik und ihrer Grenzge- biete. 3. Folge. A Series of Modern Surveys in Mathematics [Results in Mathematics and Related Areas. 3rd Series. A Series of Modern Surveys in Mathematics], vol. 49, Springer-Verlag, Berlin, 2004, Positivity for vector bundles, and multiplier ideals. MR2095472 (2005k:14001b) C. Lech: Inequalities related to certain couples of local rings, Acta Math. 112 (1964), 69 -- 89. 0161876 (28 #5080) J. Lipman: Notes on derived categories and derived functors, A preprint of a manuscript, 2006. G. Lyubeznik: F -modules: applications to local cohomology and D-modules in characteristic p > 0, J. Reine Angew. Math. 491 (1997), 65 -- 130. MR1476089 (99c:13005) G. Lyubeznik and K. E. Smith: Strong and weak F -regularity are equivalent for graded rings, Amer. J. Math. 121 (1999), no. 6, 1279 -- 1290. MR1719806 (2000m:13006) G. Lyubeznik and K. E. Smith: On the commutation of the test ideal with localization and comple- tion, Trans. Amer. Math. Soc. 353 (2001), no. 8, 3149 -- 3180 (electronic). MR1828602 (2002f:13010) L. Ma: Finiteness property of local cohomology for F -pure local rings, arXiv:1204.1539 (2012). J. Majadas and A. G. Rodicio: Smoothness, regularity and complete intersection, London Mathe- matical Society Lecture Note Series, vol. 373, Cambridge University Press, Cambridge, 2010. 2640631 (2011m:13028) H. Matsumura: Commutative algebra, second ed., Mathematics Lecture Note Series, vol. 56, Ben- jamin/Cummings Publishing Co., Inc., Reading, Mass., 1980. MR575344 (82i:13003) H. Matsumura: Commutative ring theory, second ed., Cambridge Studies in Advanced Mathematics, vol. 8, Cambridge University Press, Cambridge, 1989, Translated from the Japanese by M. Reid. MR1011461 (90i:13001) V. B. Mehta and A. Ramanathan: Frobenius splitting and cohomology vanishing for Schubert varieties, Ann. of Math. (2) 122 (1985), no. 1, 27 -- 40. MR799251 (86k:14038) J.-I. Miyachi: Localization of triangulated categories and derived categories, J. Algebra 141 (1991), no. 2, 463 -- 483. 1125707 (93b:18016) [Ma12] [MR10] [Mat80] [Mat89] [Miy91] [MR85] [Mor53] M. Moriya: Theorie der Derivationen und Korperdifferenten, Math. J. Okayama Univ. 2 (1953), 111 -- 148. MR0054643 (14,952f) [PS73] [MTW05] M. Mustat¸a, S. Takagi, and K.-i. Watanabe: F-thresholds and Bernstein-Sato polynomials, Euro- pean Congress of Mathematics, Eur. Math. Soc., Zurich, 2005, pp. 341 -- 364. MR2185754 (2007b:13010) C. Peskine and L. Szpiro: Dimension projective finie et cohomologie locale. Applications `a la d´emonstration de conjectures de M. Auslander, H. Bass et A. Grothendieck, Inst. Hautes ´Etudes Sci. Publ. Math. (1973), no. 42, 47 -- 119. MR0374130 (51 #10330) S. Ramanan and A. Ramanathan: Projective normality of flag varieties and Schubert varieties, Invent. Math. 79 (1985), no. 2, 217 -- 224. MR778124 (86j:14051) G. Scheja and U. Storch: Uber Spurfunktionen bei vollstandigen Durchschnitten, J. Reine Angew. Math. 278/279 (1975), 174 -- 190. MR0393056 (52 #13867) K. Schwede: F -adjunction, Algebra Number Theory 3 (2009), no. 8, 907 -- 950. K. Schwede: Centers of F -purity, Math. Z. 265 (2010), no. 3, 687 -- 714. 2644316 [RR85] [SS75] [Sch09] [Sch10] [Sch11a] K. Schwede: A canonical linear system associated to adjoint divisors in characteristic p > 0, arXiv:1107.3833. [Sch11b] K. Schwede: Test ideals in non-Q-Gorenstein rings, Trans. Amer. Math. Soc. 363 (2011), no. 11, [SS10] [ST08] [ST10a] [ST10b] 5925 -- 5941. 2817415 (2012c:13011) K. Schwede and K. E. Smith: Globally F -regular and log Fano varieties, Adv. Math. 224 (2010), no. 3, 863 -- 894. 2628797 (2011e:14076) K. Schwede and S. Takagi: Rational singularities associated to pairs, Michigan Math. J. 57 (2008), 625 -- 658. K. Schwede and K. Tucker: On the behavior of test ideals under finite morphisms, arXiv:1003.4333, to appear in J. Algebraic Geom. K. Schwede and K. Tucker: On the number of compatibly Frobenius split subvarieties, prime F - ideals, and log canonical centers, Ann. Inst. Fourier (Grenoble) 60 (2010), no. 5, 1515 -- 1531. 2766221 (2012d:13007) [STZ12] K. Schwede, K. Tucker, and W. Zhang: Test ideals via a single alteration and discreteness and [Sha06] rationality of f -jumping numbers, Math. Res. Lett. 19 (2012), no. 01, 191 -- 197. R. Y. Sharp: Tight closure test exponents for certain parameter ideals, Michigan Math. J. 54 (2006), no. 2, 307 -- 317. 2252761 (2007e:13008) 62 MANUEL BLICKLE AND KARL SCHWEDE [Sha07a] R. Y. Sharp: Graded annihilators of modules over the Frobenius skew polynomial ring, and tight closure, Trans. Amer. Math. Soc. 359 (2007), no. 9, 4237 -- 4258 (electronic). MR2309183 (2008b:13006) [Sha07b] R. Y. Sharp: On the Hartshorne-Speiser-Lyubeznik theorem about Artinian modules with a Frobenius [SY11] [Sho92] [Sil09] [Smi97] [Smi00] [Tak04] [TT08] [TW04] [Tat68] [Vie82] [Vit11] [Yan85] action, Proc. Amer. Math. Soc. 135 (2007), no. 3, 665 -- 670 (electronic). 2262861 (2007f:13008) R. Y. Sharp and Y. Yoshino: Right and left modules over the Frobenius skew polynomial ring in the F -finite case, Math. Proc. Cambridge Philos. Soc. 150 (2011), no. 3, 419 -- 438. 2784768 (2012e:13009) V. V. Shokurov: Three-dimensional log perestroikas, Izv. Ross. Akad. Nauk Ser. Mat. 56 (1992), no. 1, 105 -- 203. MR1162635 (93j:14012) J. H. Silverman: The arithmetic of elliptic curves, second ed., Graduate Texts in Mathematics, vol. 106, Springer, Dordrecht, 2009. 2514094 (2010i:11005) K. E. Smith: Fujita's freeness conjecture in terms of local cohomology, J. Algebraic Geom. 6 (1997), no. 3, 417 -- 429. MR1487221 (98m:14002) K. E. Smith: Globally F-regular varieties: applications to vanishing theorems for quotients of Fano varieties, Michigan Math. J. 48 (2000), 553 -- 572, Dedicated to William Fulton on the occasion of his 60th birthday. MR1786505 (2001k:13007) S. Takagi: An interpretation of multiplier ideals via tight closure, J. Algebraic Geom. 13 (2004), no. 2, 393 -- 415. MR2047704 (2005c:13002) S. Takagi and R. Takahashi: D-modules over rings with finite F -representation type, Math. Res. Lett. 15 (2008), no. 3, 563 -- 581. MR2407232 (2009e:13003) S. Takagi and K.-i. Watanabe: On F-pure thresholds, J. Algebra 282 (2004), no. 1, 278 -- 297. MR2097584 (2006a:13010) J. Tate: Residues of differentials on curves, Ann. Sci. ´Ecole Norm. Sup. (4) 1 (1968), 149 -- 159. 0227171 (37 #2756) E. Viehweg: Vanishing theorems, J. Reine Angew. Math. 335 (1982), 1 -- 8. MR667459 (83m:14011) M. A. Vitulli: Weak normality and seminormality, Commutative algebra -- Noetherian and non- Noetherian perspectives, Springer, New York, 2011, pp. 441 -- 480. 2762521 (2012d:13040) H. Yanagihara: On an intrinsic definition of weakly normal rings, Kobe J. Math. 2 (1985), no. 1, 89 -- 98. MR811809 (87d:13007) Institut fur Mathematik, Johannes Gutenberg-Universitat Mainz, 55099 Mainz, Germany E-mail address: [email protected] Department of Mathematics, The Pennsylvania State University, University Park, PA, 16802, USA E-mail address: [email protected] 

1409.6100

4

1409

A constructive approach to the module of twisted global sections on relative projective spaces

The ideal transform of a graded module $M$ is known to compute the module of twisted global sections of the sheafification of $M$ over a relative projective space. We introduce a second description motivated by the relative BGG-correspondence. However, our approach avoids the full BGG-correspondence by replacing the Tate resolution with the computationally more efficient purely linear saturation and the Castelnuovo-Mumford regularity with the often enough much smaller linear regularity. This paper provides elementary, constructive, and unified proofs that these two descriptions compute the (truncated) modules of twisted global sections. The main argument relies on an established characterization of Gabriel monads.

math.AG

math

A CONSTRUCTIVE APPROACH TO THE MODULE OF TWISTED GLOBAL SECTIONS ON RELATIVE PROJECTIVE SPACES MOHAMED BARAKAT AND MARKUS LANGE-HEGERMANN ABSTRACT. The ideal transform of a graded module M is known to compute the module of twisted global sections of the sheafification of M over a relative projective space. We introduce a second description motivated by the relative BGG-correspondence. However, our approach avoids the full BGG-correspondence by replacing the Tate resolution with the computationally more efficient purely linear saturation and the Castelnuovo-Mumford regu- larity with the often enough much smaller linear regularity. This paper provides elementary, constructive, and unified proofs that these two descriptions compute the (truncated) modules of twisted global sections. The main argument relies on an established characterization of Gabriel monads. 1. INTRODUCTION We consider coherent sheaves over the projective space Pn B for a suitable ring B. Any such coherent sheaf F can be described by a graded module over the polynomial ring S := B[x0, . . . , xn]. Even though this representation is not unique, among the different graded S-modules representing F there is the distinguished representative H 0 • (F ), the module of twisted global sections1. In general the module of twisted global sections is not finitely generated, but any of its truncations H 0 ≥d(F ) is. ≥d constructively. In this paper we treat the functor H 0 It is well-known that the ideal transform functor computes H 0 ≥d, which we state as Theorem 4.6. Furthermore, in Theorem 6.7 we present a new recursive algorithm, inspired by [EFS03, ES08], to compute H 0 ≥d. The Tate resolution in loc. cit. incorporates all higher cohomologies2, whereas our new algorithm introduces a smaller complex, called the purely linear saturation, which is tailored to H 0 • and computationally more efficient as it discards all higher cohomologies. This paper presents a categorical setup which yields unified elementary proofs of both theorems. A central notion in this paper is that of the linear regularity. We use it in Corollary 4.7 for the convergence analysis of the inductive limit defining the ideal transform and in Corol- lary 6.8 to give the number of recursion steps in our new algorithm. Like the Tate resolution, the Castelnuovo-Mumford regularity incorporates all higher cohomologies. And again, the linear regularity is an adaption to H 0 • which discards all higher cohomologies. It follows that the linear regularity is smaller or at most equal to the Castelnuovo-Mumford regularity. 2010 Mathematics Subject Classification. 13D02, 13D07, 13D45, 13P10, 13P20, 18E10, 18E35, 18A40 68W30, 14Q99 . Key words and phrases. Serre quotient category, reflective localization of Abelian categories, Gabriel monad, (truncated) module of twisted global sections, direct image functor, linear regularity, Gröbner bases, saturation . 1H 0 2See the introduction to Section 6 and Remark 7.4. • (F ) :=Lp∈Z H 0(Pn B, F (p)). 1 2 MOHAMED BARAKAT AND MARKUS LANGE-HEGERMANN This provides another reason why computing the purely linear saturation is more efficient than computing the Tate resolution. B and π : Pn One application is the computation of global sections. More precisely, let F be a co- herent sheaf on Pn B → Spec B the natural projection.3 The direct image sheaf H 0(F ) := π∗F over Spec B is the degree zero part (cf. Algorithm 3.1) of H 0 ≥0(F ). For ex- ample, if OX ⊂ OPn B, then computing π∗OX is the geometric form of eliminating all n + 1 homogeneous coordinates x0, . . . , xn from the defining equations of X ⊂ Pn B. denotes the structure sheaf of a subscheme X ⊂ Pn B A computer model of the Abelian category Coh PB of coherent sheaves on Pn corporate the objects and the morphisms4. We represent an object F ∈ Coh Pn B must in- B by a finitely −→ℓ −→ℓ Extq • (e· ) ≃ lim Another description of the cohomology functors H q The ideal transform is defined to be the inductive limit Dm(M) of the graded mod- ules Hom•(mℓ, M), where m = hx0, . . . , xni ✁ S is the irrelevant ideal. The equiva- lence H 0 H q presented graded S-modules M , such that F := fM is the sheafification of M . • (e· ) ≃ lim Hom•(cid:0)mℓ, ·(cid:1), reproved elementarily as Theorem 4.6, implies that •(cid:0)mℓ, ·(cid:1) for the higher derived cohomology functors [BS98, 20.4.4]. • arose from the BGG-correspondence [BGG78]. It is a triangle equivalence between the bounded derived category of Coh Pn B (originally over a base field B) and the stable category of finitely generated graded E- modules over the exterior algebra E, the Koszul dual B-algebra of S. Since E is a Frobe- nius algebra, this stable category is easily seen to be triangle equivalent to the homotopy category of so-called Tate complexes. A constructive description of the composition of these two triangle equivalences was given in [EFS03, DE02]. The treatment of the relative BGG-correspondence in [ES08] does not only describe the coherent sheaf cohomologies 1 The bottom complex E≥d,0 H q(fM ) = Rqπ∗fM as B-modules, but also provides a concrete realization of the direct im- age complex Rπ∗fM . However, in this approach even the computation of global sections H 0(fM) in the relative case relies a priori on the entire Tate resolution. (cid:0)T≥d(M)(cid:1) on the first page of the spectral sequence of the ≥d(fM ). We define (cid:0)T≥d(M)(cid:1). The point is that we Tate resolution T≥d(M) is a linear complex which corresponds to H 0 a new so-called purely linear saturation functor S≥d, which is computationally far more economic than the Tate functor T≥d. In Theorem 6.7, Proposition 7.2, and Corollary7.3 we prove that S≥d computes H 0 can compute S≥d without computing T≥d. This statement is not obvious in the relative case (cf. Remark 7.4). Furthermore, the linear regularity of M gives the precise number of recursion steps needed to achieve saturation. Since computing S≥d relies on Gröbner bases over the exterior algebra E of finite rank over B the involved algorithms are, for many examples, faster than the ones for the ideal transform. The latter involve Gröbner bases over the polynomial ring S of infinite rank over B. ≥d(fM ), and hence E≥d,0 1 In order to develop a unified proof that both functors Dm,≥d and S≥d compute H 0 need an appropriate categorical setup. Abstractly, the category Coh Pn B of coherent sheaves on Pn B is equivalent to the Serre quotient category A/C of the Abelian category A of finitely presented graded S-modules modulo a certain subcategory C. The necessary categorical language is summarized in Section 2. In Section 7 we show that the categories A and C can ≥d(e· ) we 3The base Spec B might even serve as the ambient space of a geometric quotient, e.g., if B is the Cox ring of a toric variety. 4The more involved modeling of the morphisms is of no relevance for this paper (cf. [BLH14b]). A CONSTRUCTIVE APPROACH TO THE MODULE OF TWISTED GLOBAL SECTIONS 3 be replaced by their respective full subcategories of modules which vanish in degrees < d for an arbitrary but fixed d ∈ Z (cf. Proposition 7.1). The A-endofunctor M 7→ H 0 is a special case of what we call a Gabriel monad, which we characterized in [BLH13] by a short set of properties. By verifying this short list of properties for the two functors Dm,≥d and S≥d we prove that they compute H 0 ≥d(fM ) ≥d(e· ). Two further applications rely on constructivity of the Gabriel monad, and now become algorithmically accessible for the category Coh Pn B: First, the Serre quotient category A/C becomes constructively Abelian once A is constructively Abelian [BLH14b, Appendix D].5 Second, the computability of the bivariate Hom and Exti functors in A/C now reduces to the computability of Hom and Exti in A (modulo a directed colimit process if i > 0) [BLH14c]. 2. PRELIMINARIES ON SERRE QUOTIENT CATEGORIES A non-empty full subcategory C of an Abelian category A is called thick if it is closed under passing to subobjects, factor objects, and extensions. In this case the Serre quotient category A/C is a category with the same objects as A and Hom-groups defined by the directed colimit HomA/C(M, N) := lim −→ M ′≤M,N ′≤N, M/M ′,N ′∈C HomA(M ′, N/N ′). The canonical functor Q : A → A/C is defined to be the identity on objects and maps a morphism ϕ ∈ HomA(M, N) to its class in the directed colimit HomA/C(M, N). The category A/C is Abelian and the canonical functor Q : A → A/C is exact. An object A(C, M) ∼= 0 for all C ∈ C, i.e., M has M ∈ A is called C-saturated if Ext0 no nonzero subobjects in C and every extension of an object C ∈ C by M is trivial. Denote by SatC(A) ⊂ A the full subcategory of C-saturated objects and by ι : SatC(A) ֒→ A its full embedding. A complex F in SatC(A) is exact if and only if ι(F ) has homology in C. A(C, M) ∼= Ext1 A thick subcategory C ⊂ A is called localizing if the canonical functor Q : A → A/C admits a right adjoint S : A/C → A, called the section functor of Q. In this case, the image of S is contained in SatC(A) and A/C −→ S (A/C) ֒→ SatC(A) are equivalences of categories. The Hom-adjunction HomA/C(Q(M), Q(N)) ∼= HomA(M, (S ◦ Q)(N)) allows to compute Hom-groups in A/C if they are computable in A and the monad S ◦ Q is computable. In particular, this avoids computing the directed colimit in the definition of HomA/C. We call any monad equivalent to S ◦ Q a Gabriel monad (of A w.r.t. C). The following theorem characterizes Gabriel monads. S Theorem 2.1 ([BLH13, Thm. 3.6]6). Let C ⊂ A be a localizing subcategory of the Abelian category A and ι : SatC(A) ֒→ A the full embedding. An endofunctor W : A → A together with a natural transformation η : IdA → W is a Gabriel monad (of A w.r.t. C) if and only if the following five conditions hold: (a) C ⊂ ker W , (b) W (A) ⊂ SatC(A), (c) the corestriction co-resSatC(A) W of W to SatC(A) is exact, (d) ηW = W η : W → W 2, and 5However, the approach using Gabriel morphisms in [BLH14b] seems computationally faster. 6Thm. 4.6 in arXiv version. 4 MOHAMED BARAKAT AND MARKUS LANGE-HEGERMANN (e) ηι : ι → W ι is a natural isomorphism. In Sections 4 and 6 we utilize this theorem to prove that certain functors are Gabriel monads of the category of coherent sheaves on the relative projective space Pn B, and thus compute the (truncated) module of twisted global sections. However, this theorem, abstract as it is, can be applied to categories of coherent sheaves of more general schemes. 3. GRADED MODULES OVER THE FREE POLYNOMIAL RING For the rest of the paper let B denote a Noetherian commutative ring with 1, V a free B- module of rank n+1, W := V ∗ = HomB(V, B) its B-dual, and x0, . . . , xn a free generating set of the B-module W . Set S := SymB(W ) = B[V ] = B[x0, . . . , xn] to be the free polynomial ring over B in the n + 1 indeterminates x0, . . . , xn. Setting deg(xj) = 1 turns S into a positively graded ring S = Li≥0 Si where Si is the set of homogeneous polynomials of degree i in S. Define the irrelevant ideal m := S>0 = hx0, . . . , xni ✁ S. The isomorphism B = S0 ∼= S/m endows B with a natural graded S-module structure. To make the statements of this paper constructive, the ring S needs to have a Gröbner bases algorithm. This is the case if B has effective coset representatives [AL94, §4.3], i.e., for every ideal I ⊂ B we can determine a set T of coset representatives of B/I, such that for every b ∈ B we can compute a unique t ∈ T with b + I = t + I. We denote by S-mod the category of (non-graded) finitely presented S-modules and by S-grmod the category of finitely presented graded S-modules. Further we denote by S-grmod≥d ⊂ S-grmod the full subcategory of all modules M with M = M≥d. Define the shift autoequivalence on S-grmod by M(i)j := Mi+j for all i ∈ Z; it induces an endofunctor on the subcategory S-grmod≥d for i ≤ 0. Algorithm 3.1. We briefly describe how to compute the i-th homogeneous part of an M ∈ S-grmod: Such a module is realized on the computer as the cokernel of a graded free S-presentation graded submodule hMii ≤ M , which we compute as the kernel of the cokernel of the The image of the graded submodule Lk,i+gk≥0 S≥i+gk(gk) ≤ Lk S(gk) under π is the restricted map M ← Lk,i+gk≥0 S≥i+gk(gk). Computing a free S-presentation hMii և Lk S(i) mi← Lℓ S(r′ ℓ) of hMii thus involves two successive syzygy computations as ex- plained in [BLH11, (10) in the proof of Theorem 3.4]. To get a free B-presentation of Mi we just need to tensor the last exact sequence with B = S/m over S, which corresponds to extracting the degree 0 relations in the reduced Gröbner basis of the S-matrix of relations mi. M π ևMk S(gk) m←Mℓ S(rℓ). A CONSTRUCTIVE APPROACH TO THE MODULE OF TWISTED GLOBAL SECTIONS 5 3.1. Internal and external Hom functors. Let M, N ∈ S-grmod. Then the Hom-group HomS-mod(M, N) of their underlying modules in S-mod is again naturally graded. This induces internal Hom functors Hom• : S-grmodop × S-grmod → S-grmod in the category S-grmod and Hom≥d : S-grmodop ≥d × S-grmod≥d → S-grmod≥d in S-grmod≥d. These internal Hom functors are algorithmically computable if B has effec- tive coset representatives (cf., e.g., [AL94, §4.3] and [BLH11, §3.3]). The (external) Hom-groups of the category S-grmod are finitely generated B-modules. They can be recovered as the graded part of degree 0 of the corresponding internal Hom's: Hom(M, N) := HomS-grmod(M, N) ∼= Hom•(M, N)0, HomS-grmod≥d(M, N) ∼= Hom≥d(M, N)0 for d ≤ 0. In particular, HomS-grmod(S, M) ∼= M0 and HomS-grmod≥d(S, M) ∼= M0 for d ≤ 0. Dealing with d > 0 would enforce further case distinctions. For example, B ∼= S/m lies in S-grmod≥d only if d ≤ 0. Till the end of Section 4 we assume that d ≤ 0. Remark 3.2. Applying Hom•(−, M) to the short exact sequence S/mℓ և S ←֓ mℓ yields (ηℓ M ) Hom•(S/mℓ, M) ֒→ M ηℓ M−−→ Hom•(mℓ, M) ։ Ext1 •(S/mℓ, M) as part of the long exact contravariant Ext•-sequence. We will repeatedly refer to this exact sequence as well as to the ℓ = 1 case M ) η1 M−−→ Hom•(m, M) ։ Ext1 •(B, M). Hom•(B, M) ֒→ M (η1 3.2. Quasi-zero modules. Let S-grmod0 denote the thick subcategory of quasi-zero mod- ules, i.e., those with M≥ℓ = 0 for ℓ large enough. Analogously, we denote by S-grmod0 ≥d the localizing (cf. Theorem 4.6) subcategory of quasi-zero modules in S-grmod≥d. Remark 3.3. For M ∈ S-grmod. Then for all ℓ ≥ 0 i (S/mℓ, M)• ∈ S-grmod0 for all i ≥ 0. •(S/mℓ, M) ∈ S-grmod0 for all j ≥ 0. •(mℓ, M) ∈ S-grmod0 for all j ≥ 1. (a) TorS (b) Extj (c) Extj Proof. The existence of a finitely generated free resolution of the first argument S/mℓ (and hence of mℓ) implies that all the above derived modules lie in S-grmod. By applying S/mℓ ⊗S − to a projective resolution of M and Hom•(S/mℓ, −) to an injective resolution of M shows that the ideal mℓ ✁ S annihilates TorS •(S/mℓ, M), which implies that they are also finitely generated S/mℓ-modules, proving (a) and (b). The exis- tence of the connecting isomorphisms Extj (S/mℓ, M) (j ≥ 1) finally implies (c). (cid:3) i (S/mℓ, M)• and Extj •(mℓ, M) ∼= Extj+1 • Remark 3.4. The use of the nonconstructive injective resolution in the previous proof is an example of an admissible use of nonconstructive arguments in an otherwise constructive setup to prove statements which neither involve existential quantifiers nor disjunctions (so- called negative formulae): Ext•(S/mℓ, M) has two descriptions. The nonconstructive one in 6 MOHAMED BARAKAT AND MARKUS LANGE-HEGERMANN the proof and the constructive one in which Hom(−, M) is applied to a finite free resolution of S/mℓ. Although the isomorphism between the two descriptions is not constructive it is "good enough" for transferring the property we want to establish. 3.3. Regularity, linear regularity, and relation to Tor and Ext. For convenience of the reader we recall the definition of the Castelnuovo-Mumford regularity in the relative case from [ES08, §2]. For any quasi-zero graded S-module N define reg N := max{d ∈ Z Nd 6= 0}. The regularity of the zero module is set to −∞. Then, for M ∈ S-grmod the S-module TorS i (B, M)• is quasi-zero and reg M := max{reg TorS i (B, M)• − i i = 0, . . . , n + 1}. Equivalently, one can define reg M := max{reg H j using the local cohomology modules H j •(S/mℓ, M) (cf., e.g., [ES08, Prop. 2.1]).7 In fact only ℓ = 1 in this sequential colimit is relevant for us. To see this we need the following result, which we also use in the proof of our key Lemma 5.3. m(M) + j j = 0, . . . , n + 1} m(M) = lim −→ℓ Extj Lemma 3.5. There exists a natural isomorphism ∼= Extn+1−i −p(Extq Proof. The Tor-Ext spectral sequence TorS collapses since Extq i (B, M)• TorS • •(∧n+1V, S) = 0 for q 6= n + 1 and Extn+1 • (∧n+1V, M). •(∧n+1V, S), M)• ⇒ Extp+q (∧n+1V, S) = B. • (∧n+1V, M) (cid:3) When B = k is a field this Lemma becomes the intrinsic and rather generalizable form i (B, M)j and the graded of the equality between the graded Betti numbers βij := dimk TorS Bass numbers: µn+1−i,j−n−1 := dimk Extn+1−i • (∧n+1V, M)j . Remark 3.6. Lemma 3.5 and the noncanonical isomorphism ∧n+1V ∼= B(n + 1) yield reg M = max{reg Extj •(B, M) + j j = 0, . . . , n + 1}. The value of the following definition will start to become obvious in Proposition 3.9 in the next subsection. Definition 3.7. Define the linear regularity of M ∈ S-grmod to be linreg M = max{reg Extj •(B, M) j = 0, 1} ∈ Z ∪ {−∞}. Analogously, the d-th truncated linear regularity of M ∈ S-grmod≥d is defined by linreg≥d M = max{reg Extj for d ≤ 0 where Extj ≥d := Extj S-grmod≥d ≥d(B, M) j = 0, 1} ∈ Z≥d ∪ {−∞}, ≃ (Extj •)≥d. Note that linreg = reg on S-grmod0 and linreg ≤ reg on S-grmod. Example 3.8. linreg S/mℓ+1 = reg S/mℓ+1 = ℓ = linreg mℓ+1 < reg mℓ+1 = ℓ + 1. 7This definition clarifies the relation to two other regularity notions: The geometric regularity is defined by g-reg M := max{reg H j max{reg H j m(M ) + j j = 2, . . . , n + 1}. m(M ) + j j = 1, . . . , n + 1} and the regularity of the sheafification regfM := A CONSTRUCTIVE APPROACH TO THE MODULE OF TWISTED GLOBAL SECTIONS 7 The motivation behind introducing linreg is that it offers an upper bound in the saturation algorithms discussed below, where the use of the (often enough much larger) regularity would be a waste of computational resources. 3.4. Saturated modules. The equivalent conditions (d) and (e) in the following proposition are computationally effective characterizations of saturated modules. Proposition 3.9. For M ∈ S-grmod the following are equivalent: (a) M is saturated w.r.t. S-grmod0; (b) Ext0 (c) The natural map ηℓ phism for all ℓ ≥ 0; •(S/mℓ, M) = 0 for all ℓ ≥ 0; •(S/mℓ, M) = 0 and Ext1 M := Hom•(S ←֓ mℓ, M) : M → Hom•(mℓ, M) is an isomor- (d) Ext0 (e) The natural map η1 •(B, M) = 0 and Ext1 •(B, M) = 0;8 M := Hom•(S ←֓ m, M) : M → Hom•(m, M) is an isomor- phism; (f) TorS (g) linreg M = −∞. n+1(B, M)• = 0 and TorS n(B, M)• = 0; And if the base ring B is a field the above is also equivalent to: (h) The projective dimension pd M ≤ n − 1. In the proof of this proposition, we use the following simple remark. Remark 3.10. The kernel K of the epimorphism mℓ և ⊗ℓm is concentrated in degree ℓ. To see this note that any homogeneous element in ⊗ℓm of degree m > ℓ which is the tensor product of monomials can be brought to the normal form xi1 ⊗B · · · ⊗B xiℓ−1 ⊗B xµ with i1 ≤ · · · ≤ im−1 ≤ min{i µi 6= 0} and µ = m − ℓ + 1. This kernel K is free over B of rank (n + 1)ℓ −(cid:0)n+ℓ n (cid:1) as the kernel of the B-epimorphism Symℓ W և ⊗ℓW . Proof of Proposition 3.9. (b) ⇔ (c): The claim is obvious from the (ηℓ M )-sequence in Remark 3.2. (d) ⇔ (e): This is a special case of the equivalence (b) ⇔ (c) for ℓ = 1. (d) ⇔ (f): This is the statement of Lemma 3.5 for i = n + 1 and i = n. (d) ⇔ (g): By definition of linreg. (a) ⇒ (d): This follows directly from the definition of saturated objects (cf. Section 2), as B ∈ C = S-grmod0. (e) ⇒ (c): Applying the ℓ-th power of Hom•(S ←֓ m, −) to M and taking the diagonal in the ℓ-dimensional cube yields the isomorphism ϕ := M ∼−→ Hom•(⊗ℓm, M) by the adjunction between ⊗ and Hom•. This isomorphism can be written as the composi- tion Hom•(S ←֓ mℓ և ⊗ℓm, M) =(cid:16)M ψ −→ Hom•(mℓ, M) χ −→ Hom•(⊗ℓm, M)(cid:17) . 8Conditions (d) and (e) are in their use of Gröbner bases algorithmically equivalent. Computing them only involves the first two morphisms in the Koszul resolution of B (and then tensoring their duals with M ). One might be tempted to expect that (d) is always algorithmically superior to condition (f), which seem to involve an n + 1-term resolution of either B or of M . However, one can easily construct examples of M ∈ S-grmod, where condition (f) is algorithmically superior, e.g., if the resolution of M terminates after few steps, long before reaching step n. 8 MOHAMED BARAKAT AND MARKUS LANGE-HEGERMANN The homomorphism χ is a monomorphism since Hom• is left exact and an epimorphism since χ ◦ ψ = ϕ is an isomorphism. Hence, χ is isomorphism and thus ψ is an isomorphism. (b) ⇒ (a): It is clear that any N ∈ S-grmod0 is an epimorphic image ofLi∈I(S/mai)(bi) for a finite set I and suitable ai and bi. Denote the kernel of N և Li(S/mai)(bi) by K. Applying Hom•(−, M) to N և Li(S/mai)(bi) ←֓ K yields as parts of the long exact sequence and Hom•(N, M) ֒→ Hom• Mi Hom•(K, M) → Ext1 •(N, M) → Ext1 , ∼=0 (S/mai)(bi), M! } {z • Mi . (S/mai)(bi), M! } {z ∼=0 The first part implies that Hom•(−, M) = 0 on S-grmod0. In particular Hom•(K, M) = 0 since K ∈ S-grmod0. Combining this and the second part implies that Ext1 •(−, M) vanishes on S-grmod0. (f) ⇔ (h): If B is a field then there exists a finite free (and not merely relatively free) presen- tation M և F• with TorS (cid:3) i (B, M)• isomorphic to the head of Fi. Corollary 3.11. For M ∈ S-grmod≥d the following are equivalent (recall, d ≤ 0): (a) M is saturated w.r.t. S-grmod0 (b) Ext0 ≥d(S/mℓ, M) = 0 and Ext1 (c) The natural map ηℓ ≥d; ≥d(S/mℓ, M) = 0 for all ℓ ≥ 0; M := Hom≥d(cid:0)S ←֓ mℓ, M(cid:1) : M → Hom≥d(mℓ, M) is an iso- morphism for all ℓ ≥ 0; ≥d(B, M) = 0 and Ext1 (d) Ext0 (e) The natural map η1 ≥d(B, M) = 0; M := Hom≥d (S ←֓ m, M) : M → Hom≥d(m, M) is an isomor- phism; n+1(B(n + 1), M)≥d = 0 and TorS (f) TorS (g) linreg≥d M = −∞. n(B(n + 1), M)≥d = 0; 4. IDEAL TRANSFORMS Recall, the m-transform of M ∈ S-grmod is the (not necessarily finitely generated) graded S-module defined by the sequential colimit Dm := lim −→ ℓ Hom•(mℓ, −) : S-grmod → S-grMod. On S-grmod≥d the d-truncated m-transform (recall, d ≤ 0) Dm,≥d := lim −→ ℓ Hom≥d(mℓ, −) : S-grmod≥d → S-grmod≥d is an endofunctor. This is a simple corollary of the Lemma 4.2 below. Definition 4.1. We define the saturation interval of M ∈ S-grmod≥d to be I≥d(M) := [δ0 M,d, δ1 M,d] ∩ Z ⊂ Z≥0, where δ0 M,d := max{reg Hom≥d(B, M) − d + 1, 0} and δ1 M,d := max{linreg≥d −d + 1, 0}. A CONSTRUCTIVE APPROACH TO THE MODULE OF TWISTED GLOBAL SECTIONS 9 The saturation interval plays a role in the following convergence analysis and the defini- tion of its upper bound is a further motivation for the linear regularity. Lemma 4.2. For each M ∈ S-grmod≥d the sequential colimit defining the m-transform is finite. More precisely, there exists a nonnegative integer δM,d ∈ I≥d(M) such that the induced maps Hom≥d(mℓ, M) → Hom≥d(mℓ+1, M) are isomorphisms for all ℓ ≥ t iff t ≥ δM,d. In particular, the natural map Hom≥d(mℓ, M) → Dm,≥d(M) is an isomorphism iff ℓ ≥ δM,d. Proof. The short exact sequence B(−ℓ)⊕? ∼= mℓ/mℓ+1 և mℓ ←֓ mℓ+1 induces for M ∈ S-grmod the exact contravariant Ext•-sequence of which the first four terms are Hom•(B, M)⊕?(ℓ) ֒→ Hom•(mℓ, M) → Hom•(mℓ+1, M) → Ext1 •(B, M)⊕?(ℓ). By Remark 3.3.(b) both Hom•(B, M) and Ext•(B, M) are quasi-zero. Hence, there exists a δM,d ∈ I≥d(M) such that the truncated morphisms Hom≥d(mℓ, M) → Hom≥d(mℓ+1, M) become isomorphisms in S-grmod≥d for ℓ ≥ t iff t ≥ δM,d. (cid:3) In particular, once B has effective coset representatives, Dm,≥d is algorithmically com- putable on objects and morphisms, since the internal Hom functor Hom≥d is. Definition 4.3. We call δM,d ∈ I≥d(M) from Lemma 4.2 the defect of saturation of M . Example 4.4. Note that 1 = δm(−t),0 ∈ I≥0(m(−t)) = [0, linreg≥0 m(−t)+1] = [0, t+1] for all t ∈ Z≥0. In other words, the maximum of I≥d(M) can be an arbitrarily bad upper bound for δM,d. Example 4.5. For M = S ⊕ B(−t) and t ≥ 0 we compute Hom•(B, M) = B(−t) and •(B, M) = B(−t + 1)n+1 (for n > 0). Hence δ0 M,0 = δM,0 is the defect ≥d-torsion part of M a Ext1 M,0 = t + 1 = δ1 of saturation. Thus, for certain examples factoring out the S-grmod0 priori could be beneficial. The natural transformation ηM := lim −→ ℓ (cid:0)ηℓ M : M → Hom≥d(mℓ, M)(cid:1) : M → Dm,≥d(M) is induced by applying Hom≥d(−, M) to the the embeddings (S ←֓ mℓ)≥d. Now we reprove that the ideal transform computes the module of twisted global sections (cf., e.g., [Vas98, §C.3]). Theorem 4.6. The d-truncated m-transform Dm,≥d together with the natural transforma- tion η : IdA → Dm,≥d is a Gabriel monad of A = S-grmod≥d w.r.t. C := S-grmod0 ≥d. In fact, the theorem holds for all d ∈ Z. The proof below assumes d ≤ 0 to avoid case distinctions. Corollary 4.7. Hom≥d(mℓ, M) is S-grmod0 Before proving the theorem we state some simple facts about ideal transforms. ≥d-saturated iff ℓ ≥ δM,d. Remark 4.8. (a) Any N ∈ S-grmod0 vanishes in degrees greater than reg N . Thus, Hom•(L≥ℓ, N)≥reg N +1−ℓ = 0 10 MOHAMED BARAKAT AND MARKUS LANGE-HEGERMANN for all ℓ ∈ Z and L ∈ S-grmod. (b) The embedding M≥t ֒→ M ∈ S-grmod≥d induces (by simple degree considerations) an isomorphism Hom≥d(L≥ℓ, M≥t) ∼→ Hom≥d(L≥ℓ, M) for all d ≤ t ≤ ℓ + d. In particular, Dm,≥d(M) ∼= Dm,≥d(M≥t) for any t ≥ d and we are allowed to replace M by any of its truncations. (c) For M ∈ S-grmod≥d take t ≥ d large enough such that the submodule M≥t has no S-grmod0 ≥d-torsion. Then Hom≥d(mℓ, M) ∼= Hom≥d(mℓ, M≥t) ∼= Hom≥d(⊗ℓm, M≥t) for all ℓ ≥ t−d by (b) and Remark 3.10. An admissible choice is t := linreg≥d M +1, then ℓ ≥ t − d ≥ δM,d (cf. Lemma 4.2). In particular, after replacing M by a high enough truncation we can always assume that Hom≥d(mℓ, M) ∼= Hom≥d(⊗ℓm, M). (d) Since the shift functor (1) : S-grmod≥d → S-grmod≥d+1, M 7→ M(1), ϕ 7→ ϕ(1) is (quasi-)inverse to the shift functor (−1) : S-grmod≥d+1 → S-grmod≥d and Dm,≥d ◦ (−1) = (−1) ◦ Dm,≥d+1 we can restrict the following proofs to Dm,≥0. Proof of Theorem 4.6. We use Theorem 2.1. Due to Remark 4.8.(d) we only need to con- sider the case d = 0. 2.1.(a) C ⊂ ker Dm,≥0: Applying Remark 4.8.(a) with L = S (and L≥L = S≥ℓ = mℓ) we conclude that Dm vanishes9 on S-grmod0 and Dm,≥0 on S-grmod0 ≥0. 2.1.(b) Dm,≥0(A) ⊂ SatC(A): For any M ∈ A, the map Hom≥0 (S ←֓ m, Dm,≥0(M)) = Hom≥0(cid:0)S ←֓ m, Hom≥0(mδM,0, M)(cid:1) = Hom≥0(cid:0)S ←֓ m, Hom≥0(⊗δM,0 m, M)(cid:1) = Hom≥0(cid:0)⊗δM,0 m ←֓ ⊗δM,0+1m, M(cid:1) = Hom≥0(cid:0)⊗δM,0 m, M(cid:1) → Hom≥0(cid:0)⊗δM,0+1m, M(cid:1) = Hom≥0(cid:0)mδM,0, M(cid:1) → Hom≥0(cid:0)mδM,0+1, M(cid:1) is an isomorphism by Lemma 4.2 proving statement (e) of Corollary 3.11. We have repeatedly used Remark 4.8.(c) and the adjunction between ⊗ and Hom≥0. 2.1.(c) G := co-resSatC (A) Dm,≥0 is exact: Applying Hom≥0(mℓ, −) to the short exact sequence L ֒→ M ։ N in S-grmod≥0 yields the exact sequence Hom≥0(mℓ, L) ֒→ Hom≥0(mℓ, M) → Hom≥0(mℓ, N) → Ext1 ≥0(mℓ, L) as part of the long exact covariant Ext≥0-sequence. Since Ext1 by Remark 3.3.(c) the sequence is exact up to defects in S-grmod0 ≥0. ≥0(mℓ, L) is quasi-zero 2.1.(d) ηDm,≥0 = Dm,≥0η: We repeatedly use the adjunction between ⊗ and Hom≥0 and Lemma 4.2 to inter- change the involved sequential colimits over ℓ′ and ℓ′′ by a common ℓ ≥ ℓ′, ℓ′′, high 9For N ∈ C all modules in the sequential colimit defining Dm,≥0(N ) vanish for ℓ ≥ δN,0 < ∞.  ∼=0 {z } ∼=0 {z } A CONSTRUCTIVE APPROACH TO THE MODULE OF TWISTED GLOBAL SECTIONS 11 enough to stabilize both colimits: ηDm,≥0(M ) = lim −→ ℓ′ Hom≥0(S ←֓ mℓ′ , lim −→ ℓ′′ Hom≥0(mℓ′′ , M)) = Hom≥0(S ←֓ mℓ, Hom≥0(mℓ, M)) = Hom≥0((S ←֓ mℓ) ⊗S mℓ, M) = Hom≥0(mℓ, Hom≥0(S ←֓ mℓ, M)) = lim −→ ℓ′′ Hom≥0(mℓ′′ , lim −→ ℓ′ = Dm,≥0(ηM ). Hom≥0(S ←֓ mℓ′ , M)) The proof implicitly uses commuting diagrams of morphisms in S-grmod≥0 to jus- tify the equality signs.10 2.1.(e) ηι is a natural isomorphism: Let M ∈ S-grmod≥0 be saturated w.r.t. S-grmod0 short exact sequence S/mℓ և S ←֓ mℓ yields ≥0. Applying Hom≥0(−, M) to the Hom≥0(S/mℓ, M) ֒→ M ηℓ M−−→ Hom≥0(mℓ, M) ։ Ext1 ≥0(S/mℓ, M) since S/mℓ ∈ S-grmod0 ≥0. In other words, ηℓ M is an isomorphism for all ℓ. (cid:3) Remark 4.9. The saturation process of M ∈ S-grmod conducted by Dm brings linreg to −∞, whereas reg is only brought down to the regularity of the sheafification. 0, . . . , xℓ Since the Frobenius powers m[ℓ] := hxℓ ni satisfy mℓ ≥ m[ℓ] ≥ m(n+1)ℓ we can use them instead of mℓ them in the above sequential colimits. They are computationally superior since their number of generators does not increase with ℓ. In other words, the module Hom≥d(m[δM,d], M) is S-grmod0 ≥d-saturated. Alternatively, one could iteratively (δM,d times) apply Hom≥d(m, −) to (the S-grmod0 ≥d-torsion-free factor of) M . It depends on the example which algorithm is faster. 5. GRADED S-MODULES AND LINEAR E-COMPLEXES In this section we describe how to translate the module structure of M ∈ S-grmod into the structure of a linear complex R(M) over the exterior algebra E := ∧V , which is Koszul dual to S = Sym V ∗. This translation turns out to be functorial, algorithmic, and an adjoint equivalence of categories. We denote the category of finitely generated graded E-modules by E-grmod. Let e0, . . . , en denote a B-basis V of which the indeterminates x0, . . . , xn of S form the dual B-basis of W = V ∗ = HomB(V, B). We set deg(ei) = −1 for all i = 0, . . . , n. 10We could have used the fact that Dm,≥0 = lim −→ℓ Hom≥0(mℓ, −) commutes with directed colimits. How- ever, the general form of the second statement is not quite trivial [BS98, Coro. 3.4.11] (the directed colimit is called direct limit in [BS98, Terminology 3.4.1]). Note that although the ideal transform commutes with directed colimits, it does not generally commute with arbitrary finite colimits, for otherwise it would be right exact and hence exact. 12 MOHAMED BARAKAT AND MARKUS LANGE-HEGERMANN 5.1. The functor R. The B-linear maps µi(xj) : Mi → Mi+1, m 7→ xjm, for j = 0, . . . , n, and i ∈ Z induced by the indeterminates xj encode the graded S-module structure of an M ∈ S-grmod (cf. Algorithm 3.1 for an algorithm to compute Mi). Example 5.1. For S := B[x0, x1] consider the free S-module M := S = S(0) of rank 1. Each graded part Mi is a free B-module for which we fix a basis of monomials, e.g., M0 = h1iB, M1 = hx0, x1iB, M2 = hx2 0, x0x1, x2 1iB. Then the matrices µ0(x0) = ( 1 0 ) , µ0(x1) = ( 0 1 ) , 0 1 0 ) , µ1(x1) = ( 0 1 0 µ1(x0) = ( 1 0 0 0 0 1 ) , 0 : 1 : ... represent the maps µi(xj). Using the B-basis (e0, . . . , en) of V define for each i ∈ Z the map µi as the composition µi :(cid:26) Mi → EndB(V ) ⊗B Mi → V ⊗B Mi+1, idV ⊗m m 7→ i=0 ej ⊗ xjm . 7→ Pn Using the natural isomorphism HomB(Mi, V ⊗B Mi+1) ∼= HomE-grmod(E ⊗B Mi, E ⊗B Mi+1) each µi can equally be understood as a map of graded E-modules µi : E ⊗B Mi → E ⊗B Mi+1, where the B-module Mj is considered as a graded B-module concentrated in degree j and, therefore, E ⊗B Mj is generated by (a generating set of) Mj in degree j. For a better functorial behavior we replace E by its B-dual [Lam99, §16C] ωE := HomB(E, B) ∼= ∧W ∼= ∧n+1W ⊗B E in the above maps.11 In particular, ωE lives in the degree interval 0, . . . , n + 1 and its socle (ωE)0, which is naturally isomorphic to B, is concentrated in degree 0. We denote the distinguished generator of the socle corresponding to 1B by 1ωE . This change of language is justified by reinterpreting µi : Mi → V ⊗B Mi+1 as a map µi : W ⊗B Mi → Mi+1 using the adjunction HomB(W ⊗B X, Y ) ∼= HomB(X, HomB(W, Y )) ∼= HomB(X, W ∗ ⊗B Y ). The graded E-module ωE ⊗B Mj has (compared with E ⊗B Mj) the advantage of having the B-module Mj as its socle interpreted as a graded E-module concentrated in degree j. The commutativity of S implies that the composed map µi+1 ◦ µi : ωE ⊗B Mi → ωE ⊗B Mi+2 is zero. Thus, the sequence of µi's yields the so-called R-complex (cf. [EFS03, §2] and [ES08, §2]) R(M) : · · · → ωE ⊗B Mi µi −→ ωE ⊗B Mi+1 µi+1 −−→ ωE ⊗B Mi+2 → · · · Example 5.1 (continued). For M = S(0) we obtain the R-complex 0 ωE(0)1 ( e0 e1 ) ωE(−1)2 ( e0 e1 · · e0 e1 ) ωE(−2)3 (cid:16) e0 e1 · · e0 e1(cid:17) ωE(−3)4 · · e0 e1 · · · · · 11It is again a free graded E-module which is nonnaturally isomorphic to E(−n − 1). A CONSTRUCTIVE APPROACH TO THE MODULE OF TWISTED GLOBAL SECTIONS 13 Lemma 5.2 ([EFS03, Prop. 2.3]). There exists a natural isomorphism H a(R(M))a+i ∼= TorS i (B, M)a+i. Proof. The idea is to interpret the bigraded differential module ωE ⊗B M either as R(M) or as the Koszul resolution of B tensored with M over S. (cid:3) Lemmas 5.2 and 3.5 imply the following lemma, an important technical insight for the rest of this paper. Lemma 5.3 (Key Lemma). There exists a natural isomorphism (∧n+1V, M)a+i. H a(R(M))a+i ∼= Extn+1−i • Hence, there is a noncanonical isomorphism H a(R(M))a+n+1−j ∼= Extj •(B, M)a−j. Let A be either S or E. An epimorphism in A-grmod is said to be B-split if it splits as a morphism over B. A graded module P ∈ A-grmod is said to be relatively projective (with respect to B) if HomS(P, −) sends B-split epis to surjections. Any module of the form A ⊗B M , where M is a B-module, is called relatively free (with respect to B). By [ES08, Proposition 1.1], an N ∈ A-grmod is relatively projective if and only if it is relatively free. A complex C = C • of graded E-modules is called linear if each C i is relatively free (with respect to B) with socle concentrated in degree i.12 The regularity of a linear complex C is defined as reg C := sup{a ∈ Z H a(C) 6= 0} ∈ Z ∪ {−∞, ∞}. Lemma 5.2 or 5.3 connects the regularity of a graded module with that of its R-complex. Corollary 5.4. For M ∈ S-grmod the equality reg M = reg R(M) holds. These definitions allow us to describe the image of R. Definition 5.5. We denote by E- grlin the full subcategory of complexes C of graded E-modules satisfying (a) C is linear; (b) each C i is finitely generated; (c) C is left bounded; (d) reg C < ∞. By E- grlin0 we denote the thick subcategory of bounded complexes. Finally, for any d ∈ Z, denote by E- grlin≥d the full subcategory of complexes in E- grlin with C <d = 0 and by E- grlin≥d,0 := E- grlin≥d ∩ E- grlin0. Remark 5.6. An object C ∈ E- grlin≥d can be represented on a computer by the finite com- plex C d → C d+1 → . . . → C j−1 → C j provided that j > reg C. The part C >j of C can be recovered by an injective resolution of coker(C j−1 → C j). This relatively injec- tive resolution is isomorphic to Hom•(−, E) applied to a relatively projective resolution of Hom•(coker(C j−1 → C j), E). A morphisms in E- grlin can be represented on the com- puter by a chain morphism between two such finite complexes, one only needs to extend these complexes to equal cohomological degrees. Again, the part of the morphism in higher cohomological degrees can be computed by an injective resolution. Proposition 5.7. The construction R induces two fully faithful functors R : S-grmod → E- grlin and R≥d : S-grmod≥d → E- grlin≥d for all d ∈ Z. 12and hence C i is generated in degree i + n + 1. 14 MOHAMED BARAKAT AND MARKUS LANGE-HEGERMANN Proof. As M ∈ S-grmod is finitely generated, R(M) is left bounded. By definition, each R(M)i = ωE ⊗B Mi is a finitely generated graded relatively free module with socle Mi concentrated in degree i. Furthermore reg R(M) = reg M < ∞ by Corollary 5.4. A graded morphism ϕ : M → N induces morphisms ϕi : Mi → Ni for all i ∈ Z. Tensoring with ωE yields morphisms R(ϕ)i : R(M)i → R(N)i. These morphisms are chain morphisms, as xj ◦ ϕi = ϕi+1 ◦ xj and the µi are induced by the xj for all i ∈ Z and all 0 ≤ j ≤ n. Restricting R to S-grmod≥d corestricts to E- grlin≥d by construction. These functors are obviously faithful. The fullness R and R≥d follows directly from the below Proposition 5.8 and Corollary 5.9, respectively. (cid:3) 5.2. The functor R induces an equivalence. The functor R is an equivalence S-grmod ∼−→ E- grlin by [EFS03, Prop. 2.1]. In this section we explicitly construct the left adjoint quasi- inverse M of R and thus show constructively that R is an adjoint equivalence. Proposition 5.8. There exists a functor M : E- grlin → S-grmod such that M ⊣ R is an adjoint equivalence of categories which sends S-grmod0 to E- grlin0. Proof. Let (C, µ) ∈ E- grlin. For a preparatory step, assume that (A) H r(C) is the only nonvanishing cohomology (this implies that C <r = 0). r . Define M(C) as its cokernel (with relatively free presentation πr : S ⊗B C r r r → C r+1 r+1 and extend ker(µr) κ−→ W ⊗B C r r to a map S ⊗B ker(µr) −→ ։ Consider µr : W ⊗B C r S ⊗B C r M(C)). and such that R(M) satisfies assumption (A). The natural isomorphism eδ−1 : Mr To justify the correctness of this preparatory step let M ∈ S-grmod with M<r = 0 ∼−→ M(R(M))r : m 7→ 1S ⊗B 1ωE ⊗B m identifies a minimal set of generators M with one of M(R(M)). The assumption (A) for R(M) is equivalent, by Lemma 5.2, to M being gener- ated in degree r and having a relatively free resolution which is linear in the xi. In particular, the only relations involving the indeterminates xi of the finite set of generators of M in Mr are linear relations. All these linear relations are encoded in the map R(M)r → R(M)r+1. The construction of M above just imposes these linear relations of the generators of M to the generators of M(R(M)). In particular,eδ induces an isomorphism δM : M(R(M)) → M . Similarly, there exists an isomorphism ηC : C → R(M(C)) for any C ∈ E- grlin satisfying assumption (A). For a general (C, µ) ∈ E- grlin, there is a bound r (e.g., any r > reg(C)) such that the preparatory step applies to (C ≥r, µ≥r). Then, we inductively define M(C) by decreasing the cohomological degree d. Let (C, µ) ∈ E- grlin be a complex and d < r such that M(C ≥d+1) is defined by the induction hypothesis with relatively free presentation πd+1 : S ⊗B (C d+1 r ) ։ M(C ≥d+1). We define M(C ≥d) as a pushout of the span of β and γ defined as follows: Let α : S ⊗B W → S : p ⊗ xi → xip and ι : C d+1 d+1 ⊕ . . . ⊕ C r r d and γ := πd+1 ◦ (S ⊗B (ι ◦ µd)) be the embedding in the direct sum. Now set β := α ⊗B C d d → C d+1 with common source S ⊗B W ⊗ C d d (recall, µd : W ⊗B C d d+1 ). This inductive step of the construction of M is the reverse construction of R. d+1 ⊕. . .⊕C r d+1 ֒→ C d+1 To apply M to a morphism ϕ : C → D in E- grlin we use the identification of M(C)i with C i i , map C i i using ϕi to Di i, which we finally identify with M(D)i. A CONSTRUCTIVE APPROACH TO THE MODULE OF TWISTED GLOBAL SECTIONS 15 This equivalence of categories is an adjoint equivalence. We already have constructed the unit η and counit δ as natural isomorphisms in the preparatory step. This unit and counit nat- ∼−→ urally extends into lower cohomological degrees using the natural B-isomorphisms C i i i : m 7→ 1ωE ⊗B m. The triangle identities are M(C)i : c 7→ 1S ⊗B c and Mi easily verified. (cid:3) ∼−→ R(M)i Corollary 5.9. The restriction-corestriction M≥d : E- grlin≥d → S-grmod≥d of M and ≥d to the functor R≥d form an adjoint equivalence M≥d ⊣ R≥d, which sends S-grmod0 E- grlin≥d,0. 5.3. Saturated linear complexes. We now give a characterization of saturated linear com- plexes corresponding to the one we gave for graded modules. Definition 5.10. The linear regularity of a linear complex C ∈ E- grlin is defined as linreg C := max{a ∈ Z H a(C)a+n+1 6= 0 or H a(C)a+n 6= 0} ∈ Z ∪ {−∞}. We get a further characterization of E- grlin0-saturated linear complexes. Corollary 5.11. A complex C ∈ E- grlin is E- grlin0-saturated iff linreg C = −∞. Proof. By Proposition 3.9, the module M(C) is S-grmod0-saturated if Extj •(B, M(C)) = 0 for j ∈ {0, 1}. This is equivalent to H a(R(M(C)))a+n+1−j = 0 for j ∈ {0, 1} by the key Lemma 5.3. The claim follows from C ∼= R(M(C)). (cid:3) The key Lemma 5.3 also implies: Corollary 5.12. linreg C = linreg M(C) for all C ∈ E- grlin. The localizing subcategory S-grmod0 ≥d of S-grmod≥d corresponds via the adjoint equiv- alence M ⊣ R to the full localizing subcategory E- grlin≥d,0 of right bounded complexes in E- grlin≥d, i.e., of those complexes C ∈ E- grlin≥d with C ≥ℓ = 0 for ℓ large enough. A module M ∈ S-grmod≥d is then S-grmod0 ≥d-saturated if and only if R(M) is E- grlin≥d,0- saturated, i.e., the adjoint equivalence M≥d ⊣ (R≥d : S-grmod≥d → E- grlin≥d) re- stricts to an adjoint equivalence between the full subcategories of S-grmod0 ≥d-saturated resp. E- grlin≥d,0-saturated objects. The definition of the linear regularity of complexes in E- grlin≥d and the characterization of E- grlin≥d,0-saturated linear complexes is a little bit more subtle and is therefore deferred to the next section. The reason is that the lowest cohomology H d(C) has to be treated separately. 6. SATURATION OF LINEAR COMPLEXES The ideal transform in Section 4 leads to an algorithm for the saturation of graded S- modules. In this section, we present an algorithm to saturate linear complexes. The adjoint equivalence M ⊣ R translates this to a second algorithm for the saturation of graded S- modules. Corollary 5.11 indicates that one has to modify a linear complex C until H a(C)a+n+1 = 0 and H a(C)a+n = 0. Our purely linear saturation is similar to that of the Tate resolution in that one truncates C in cohomological degree high enough and then computes a suitable part in lower cohomological degrees. In contrast to the Tate resolution, our approach remains in the category of linear complexes, as we do not take free presentations of kernels to compute 16 MOHAMED BARAKAT AND MARKUS LANGE-HEGERMANN the part of lower cohomological degree, but so-called purely linear kernels. We can also truncate above the linear regularity, a lower bound of the Castelnuovo-Mumford regularity. For the relation between the purely linear saturation and the Tate resolution see Remark 7.4. 6.1. Purely linear kernels. Let C i, C i+1 ∈ E-grmod be relatively free with socle concen- trated in degree i and i + 1, respectively13. We call a morphism ϕi : C i → C i+1 purely linear (of degree i) if its kernel vanishes in the top degree i + n + 1. Definition 6.1. Let ϕi : C i → C i+1 be purely linear of degree i. A purely linear morphism κ : K i−1 → C i of degree i − 1 with ϕi ◦ κ = 0 is called purely linear kernel if for any purely linear λ : Li−1 → C i of degree i − 1 with ϕi ◦ λ = 0 there exists a unique morphism ψ : Li−1 → K i−1 with κ ◦ ψ = λ. ϕi C i+1 κ λ C i 0 K i−1 ψ Li−1 Lemma 6.2. Each purely linear morphism has a purely linear kernel, which, by the uni- versal property, is unique up to a unique isomorphism. Proof. We denote the restriction of any morphism β to the graded part of degree i + n by βi+n. Let ϕi : M i → M i+1 be purely linear of degree i and ν : N i−1 ֒→ M i be its kernel. i+n ⊗B E and by λ : K i−1 → N i−1 the map induced by the identity Denote by K i−1 := N i−1 on N i−1 i+n. We show that κ := ν ◦ λ : K i−1 → M i is a purely linear kernel of ϕi. By definition, K i−1 and M i are relatively free generated in degree i + n and degree i + n + 1, respectively. As ϕi is purely linear, N i−1 lives in the degree interval i, . . . , i + n. i+n. By definition, λi+n : K i−1 In particular, νi+n is a kernel of ϕi i+n is an isomorphism and thus also κi+n is a kernel of ϕi i+n. In particular, the kernel of κ lives in the degree interval i − 1, . . . , i + n − 1. Thus, κ is purely linear of degree i − 1. i+n → N i−1 The composition ϕi ◦ ν is zero and κ factors over ν by construction. Thus, ϕi ◦ κ = 0. N i−1 λ K i−1 ψ M i−1 ν κ ϕi M i M i+1 ϕi−1 0 To show the universal property of κ let ϕi−1 : M i−1 → M i be purely linear with ϕi ◦ ϕi−1 = 0. From the universal property of κi+n as a kernel, there exists a unique ψi+n : M i−1 i+n = 0. We define i+n → K i−1 i+n, since ϕi−1 i+n ◦ ϕi i+n with κi+n ◦ ψi+n = ϕi−1 ψ := ψi+n ⊗B E : M i−1 ∼= M i−1 i+n ⊗B E −→ K i−1 ∼= K i−1 i+n ⊗B E, which extends ψi+n to a morphism of graded E-modules. Finally, ϕi−1 = κ ◦ ψ since κi+n ◦ ψi+n = ϕi−1 (cid:3) i+n and ϕi−1 is uniquely determined by ϕi−1 i+n. Note that all steps in the proof of this last lemma are constructive. We can now state the definition of linear regularity for complexes in E- grlin≥d. 13or, equivalently, freely generated in degree i + n + 1 and i + n + 2, respectively. A CONSTRUCTIVE APPROACH TO THE MODULE OF TWISTED GLOBAL SECTIONS 17 Definition 6.3. Define for any d ≤ 0 the d-th truncated linear regularity linreg≥d C ∈ Z≥d ∪ {−∞} of C ∈ E- grlin≥d as follows: If there exists an a ∈ Z>d such that H a(C)a+n+1 6= 0 or H a(C)a+n 6= 0 then linreg≥d C := max{a ∈ Z>d H a(C)a+n+1 6= 0 or H a(C)a+n 6= 0} ∈ Z>d. Otherwise, if the lowest morphism C d → C d+1 is a purely linear kernel (of C d+1 → C d+2) then linreg≥d C := −∞ else linreg≥d C := d. Corollary 6.4. C ∈ E- grlin≥d is E- grlin≥d,0-saturated iff linreg≥d C = −∞. Proof. The claim follows from Corollary 5.11 and Corollary 3.11 (by Remark 4.8.(d) we only need to consider the case d = 0). (cid:3) Corollary 6.5. linreg≥d C = linreg≥d Proposition 6.6. A (C, µ) ∈ E- grlin≥d, is E- grlin≥d,0-saturated if and only if µi is the M≥d(C) for all C ∈ E- grlin≥d. purely linear kernel of µi+1 for all i ≥ d. Proof. By the proof of Lemma 6.2, a morphism κ : K i−1 → M i is a purely linear kernel of a purely linear morphism ϕi : M i → M i+1 of degree i if and only if it is purely linear and K i−1 κ−→ M i ϕi −→ M i+1 is a complex which is exact in degrees i + n and i + n + 1. Now, the claim follows from the characterizations of saturated linear complexes in Corollary 6.4. (cid:3) 6.2. Saturation of a linear complex. In this subsection we algorithmically saturate linear complexes by iteratively computing purely linear kernels. Let (C, µ) ∈ E- grlin≥d with regularity r ∈ Z. Define the purely linear saturation (truncated in degree d) functor S≥d : E- grlin≥d → E- grlin≥d as follows. The idea is to truncate the complex above the linear regularity and then to "saturate" it recursively by purely linear kernels, more precisely: For cohomological degrees greater than the linear regularity r = linreg≥d C define S≥r+1 by setting S≥r+1(C, µ) := (C ≥r+1, µ≥r+1). Assume that (D≥i, τ ≥i) = S≥i(C, µ) is defined for some i > d. Let τ i−1 : Di−1 → Di be the purely linear kernel of τ i. Define S≥i−1(C, µ) by adding τ i−1 to (D≥i, τ ≥i) in cohomological degree i − 1. The morphism part S≥d(ϕ) for ϕ : (C, µC) → (C ′, µC ′) in E- grlin≥d is induced by the identity in high degrees. The universal property of the purely linear kernels implies a unique completion of the square S≥d(C)ℓ S≥d(C)ℓ+1 S≥d(ϕ)ℓ+1 S≥d(C ′)ℓ S≥d(C ′)ℓ+1 and thus iteratively constructs the chain morphisms in lower degrees. Theorem 6.7. Let A = E- grlin≥d and C := E- grlin≥d,0. There exists a natural trans- formation η : IdA → S≥d such that the purely linear saturation S≥d truncated in degree d together with this natural transformation η is a Gabriel monad of A w.r.t. C. Again, the statement of the theorem is valid for all d ∈ Z. The statement of the following immediate corollary and the proof the theorem assume d ≤ 0. 18 MOHAMED BARAKAT AND MARKUS LANGE-HEGERMANN Corollary 6.8. The nonnegative integer max{linreg≥d C − d + 1, 0} is the precise count of recursion steps needed to achieve saturation. Thus, the linear regularity yields a better bound for computing zeroth cohomologies and saturation than the Castelnuovo-Mumford regularity does. However, the data structure for E- grlin≥d suggested in Remark 5.6 still requires the Castelnuovo-Mumford regularity. Proof of Theorem 6.7. First, we construct the natural transformation ηC for the complex (C, µ) ∈ E- grlin≥d. Consider the cochain-isomorphism ηC : C ≥r → S≥r(C) induced by the identity for r > linreg≥d C. Assume that ηC is lifted to a cochain morphism C ≥ℓ+1 → S≥ℓ+1(C). The universal property of the purely linear kernels implies a completion of the square C ℓ ηℓ C C ℓ+1 ηℓ+1 C S≥d(C)ℓ S≥d(C)ℓ+1 by a morphism ηℓ S≥d(C). C : C ℓ → S≥d(C)ℓ. Iteratively, we get the cochain-morphism ηC : C → Now, we use Theorem 2.1 to show that S≥d together with η is a Gabriel monad. 2.1.(a) C ⊂ ker S≥d: As ηC is an isomorphism in high cohomological degrees, its kernel is contained in C. 2.1.(b) S≥d(A) ⊂ SatC(A): S≥d(C) has only trivial cohomologies above the regularity of C. Below the regularity we use Proposition 6.6. 2.1.(c) G := co-resSatC (A) S≥d is exact: As S≥d is the identity on objects and morphism in high cohomological degree, apply- ing it to a short exact sequence in A yields a new sequence with A-defects, which are bounded by the maximum of the regularities of said short exact sequence. Thus, the A-defects are contained in C. In particular, this sequence is exact when considered in SatA(C). 2.1.(d) ηS≥d = S≥dη: Truncated at cohomological degree ℓ above the regularity this is clear, since both natural transformations are induced by the identity. For lower degrees, this follows from the uniqueness of the universal morphism ψ in the definition of purely linear kernels. 2.1.(e) ηι is a natural isomorphism: Let C ∈ A be C-saturated. We need to show that ηC is a cochain isomorphism. This is clear in high cohomolical degrees, as S≥d is the identity on objects and morphism there. Assume that ηC restricted to C ≥ℓ+1 → S≥ℓ+1(C) for some ℓ ∈ Z is a cochain isomorphism. Then, the morphism ηℓ C from the completion of the square C ℓ ηℓ C S≥d(C)ℓ C ℓ+1 ηℓ+1 C S≥d(C)ℓ+1 A CONSTRUCTIVE APPROACH TO THE MODULE OF TWISTED GLOBAL SECTIONS 19 is an isomorphism, since both C and S≥d(C) are saturated and, by Proposition 6.6 C ℓ and S≥d(C)ℓ are purely linear kernels of µℓ+1 and the morphism in cohomological degree ℓ+1 of S≥d(C), respectively. The uniqueness of purely linear kernels implies that ηC restricted to C ≥ℓ → S≥ℓ(C) is a cochain isomorphism, and so is ηC by induction. (cid:3) We stress that the above functors M, M≥d, R, R≥d, and S≥d are constructive functors between constructively Abelian categories. We furthermore note that computing the natural transformation η is constructive. 7. THE GABRIEL MONAD OF THE CATEGORY OF COHERENT SHEAVES In this section we prove that for any d ∈ Z the quotient category S-grmod≥d/S-grmod0 ≥d B and that the corresponding Gabriel monad computes is equivalent to the category Coh Pn the (truncated) module of twisted global sections. Proposition 7.1. Coh Pn B ≃ S-grmod≥d/S-grmod0 ≥d for all d ∈ Z. Proof. First, the definitions directly imply S-grmod0 ∩ S-grmod≥d = S-grmod0 ≥d. Second, a preimage of M ∈ S-grmod/S-grmod0 under S-grmod≥d → S-grmod/S-grmod0 is given by M≥d, since M ∼= M≥d in S-grmod/S-grmod0. Now the second isomorphism theorem for Abelian categories [BLH14a, Prop. 3.2] implies the equivalence S-grmod≥d/S-grmod0 ≥d ≃ S-grmod/S-grmod0. The latter category is equivalent to Coh Pn B by [BLH14a, Coro. 4.2]. (cid:3) A graded S-modules M is called quasi finitely generated if each truncation M≥d is finitely generated. We denote by S-qfgrmod ⊂ S-grMod the full subcategory of such modules. The functor computing the module of twisted global sections is right adjoint to the sheafification functor Sh : S-qfgrmod → Coh Pn [Ser55, 59] and later by Grothendieck for the relative case. H 0(Pn B, F (p)) B → S-qfgrmod : F 7→Mp∈Z B, M 7→ fM . This was proved by Serre in the absolute case B → S-grmod≥d : F 7→ Mp∈Z≥d B the restriction of Sh to S-grmod≥d and by B, F (p)) H 0(Pn Denote by Sh≥d : S-grmod≥d → Coh Pn H 0 • : Coh Pn H 0 ≥d : Coh Pn the functor computing the truncated module of twisted global sections. It follows that H 0 ≥d is the right adjoint of Sh≥d. Proposition 7.2. The monad H 0 w.r.t. the localizing subcategory S-grmod0 the truncated module of twisted global sections. ≥d(e· ) = H 0 ≥d ◦ Sh≥d is a Gabriel monad of S-grmod≥d ≥d. In particular, any Gabriel monad computes Proof. Let Q≥d : S-grmod≥d → S-grmod≥d/S-grmod0 equivalence in Proposition 7.1 is constructed as a functor α≥d : S-grmod≥d/S-grmod0 Coh Pn given by S≥d := H 0 ≥d be the canonical functor. The ≥d → B with α≥d ◦ Q≥d ≃ Sh≥d. An easy calculation shows that a right adjoint of Q≥d is ≥d ◦ α≥d. In particular, S≥d ◦ Q≥d is a Gabriel monad of S-grmod≥d 20 MOHAMED BARAKAT AND MARKUS LANGE-HEGERMANN w.r.t. S-grmod0 H 0 ≥d by [BLH13, Lemma 4.3]. Now the claim follows, as S≥d ◦ Q≥d = (cid:3) ≥d ◦ α≥d ◦ Q≥d ≃ H 0 Corollary 7.3. There exist natural isomorphisms ≥d ◦ Sh≥d = H 0 ≥d(e· ). and R(cid:16)H 0 H 0(fM (i)) ∼= (Dm,≥d(M))i ≥d(fM )(cid:17) ∼= S≥d(R(M)), ∼=(cid:0)S≥d(R(M))(cid:1)i ωE ⊗k H q(cid:16)fM (i − q)(cid:17) , S≥d(R(M))i = ωE ⊗B H 0(cid:16)fM (i)(cid:17) . min{n,i−d}Mq=0 T≥d(M)i = . i Remark 7.4. In the absolut case, i.e., when B = k is a field, the (objects of the truncated) Tate resolution T≥d(M) relate to the higher cohomology modules H q ≥d(fM ) by [EFS03] (*) while the (truncated) purely linear saturation directly extracts H 0: in particular for i ≥ d H 0 ≥d(fM ) ∼= Dm,≥d(M) In the relative case the analogue of (*) is more subtle: The Tate resolution T≥d(M) is by its bi-graded structure in fact a multi-complex T≥d,•(M). Since each multi-complex is a filtered complex and hence induces a spectral sequence14 Ep,q(T≥d,•(M)) =⇒ 0, where for each row-complex on the first page the following isomorphism holds This is implicit in [ES08], see also [EFS03, Corollary 3.6]. Thus, the relation between the purely linear saturation and the Tate resolution is given by 1 E≥d,q (cid:0)T≥d,•(M)(cid:1) ∼= R(cid:16)H q ≥d(fM )(cid:17) . (cid:0)T≥d,•(M)(cid:1) ∼= S≥d(R(M)) ∼= R(cid:16)H 0 E≥d,0 1 ≥d(fM )(cid:17) . The first isomorphism is not a priori obvious in the relative case and gives a direct way to compute R(cid:16)H 0 ≥d(fM)(cid:17) via S≥d(R(M)) without computing (most of) the Tate resolution.15 REFERENCES [AL94] William W. Adams and Philippe Loustaunau, An introduction to Gröbner bases, Graduate Studies in Mathematics, vol. 3, American Mathematical Society, Providence, RI, 1994. MR 1287608 (95g:13025) 4, 5 I. N. Bernšteın, I. M. Gel′fand, and S. I. Gel′fand, Algebraic vector bundles on Pn and prob- lems of linear algebra, Funktsional. Anal. i Prilozhen. 12 (1978), no. 3, 66 -- 67. MR MR509387 (80c:14010a) 2 [BGG78] 14The vertical morphisms of this multi-complex are the morphisms between the graded summands rep- resented by scalar matrices (i.e., degree zero in E). This differs from the MACAULAY2 convention used in [ES08], where these morphisms are arranged "diagonally up and to the right" (cf. [ES08, Chapter 3]). Hence, we do not arrange the direct summands of the modules in the Tate resolution vertically, but diagonally up and to the left. 1 15In absolute case B = k the isomorphism easily follows from the fact that the bottom complex (cid:0)T≥d,•(M )(cid:1) of the first spectral sequence is already the subcomplex of the Tate resolution consisting E≥d,0 of those direct summands of the objects in T≥d(M ) where the degree of the socle equals the cohomological degree. A CONSTRUCTIVE APPROACH TO THE MODULE OF TWISTED GLOBAL SECTIONS 21 [BLH11] Mohamed Barakat and Markus Lange-Hegermann, An axiomatic setup for algorithmic homolog- ical algebra and an alternative approach to localization, J. Algebra Appl. 10 (2011), no. 2, 269 -- 293, (arXiv:1003.1943). MR 2795737 (2012f:18022) 4, 5 [BLH13] Mohamed Barakat and Markus Lange-Hegermann, On monads of exact reflective local- izations of Abelian categories, Homology Homotopy Appl. 15 (2013), no. 2, 145 -- 151, (arXiv:1202.3337). MR 3138372 3, 20 [BLH14a] Mohamed Barakat and Markus Lange-Hegermann, Characterizing Serre quotients with no section functor and applications to coherent sheaves, Appl. Categ. Structures 22 (2014), no. 3, 457 -- 466, (arXiv:1210.1425). MR 3200455 19 [BLH14b] Mohamed Barakat and Markus Lange-Hegermann, Gabriel morphisms and the computability of Serre quotients with applications to coherent sheaves, (arXiv:1409.2028), 2014. 2, 3 [BLH14c] Mohamed Barakat and Markus Lange-Hegermann, On the Ext-computability of Serre quotient categories, J. Algebra 420 (2014), 333 -- 349, (arXiv:1212.4068). MR 3261464 3 [BS98] M. P. Brodmann and R. Y. Sharp, Local cohomology: an algebraic introduction with geometric applications, Cambridge Studies in Advanced Mathematics, vol. 60, Cambridge University Press, Cambridge, 1998. MR 1613627 (99h:13020) 2, 11 [DE02] Wolfram Decker and David Eisenbud, Sheaf algorithms using the exterior algebra, Computations in algebraic geometry with Macaulay 2, Algorithms Comput. Math., vol. 8, Springer, Berlin, 2002, pp. 215 -- 249. MR MR1949553 2 [EFS03] David Eisenbud, Gunnar Fløystad, and Frank-Olaf Schreyer, Sheaf cohomology and free resolu- tions over exterior algebras, Trans. Amer. Math. Soc. 355 (2003), no. 11, 4397 -- 4426 (electronic). MR MR1990756 (2004f:14031) 1, 2, 12, 13, 14, 20 David Eisenbud and Frank-Olaf Schreyer, Relative Beilinson monad and direct image for families of coherent sheaves, Trans. Amer. Math. Soc. 360 (2008), no. 10, 5367 -- 5396, (arXiv:math/0506391). MR 2415078 (2009f:14030) 1, 2, 6, 12, 13, 20 [ES08] [Lam99] T. Y. Lam, Lectures on modules and rings, Graduate Texts in Mathematics, vol. 189, Springer- [Ser55] Verlag, New York, 1999. MR MR1653294 (99i:16001) 12 Jean-Pierre Serre, Faisceaux algébriques cohérents, Ann. of Math. (2) 61 (1955), 197 -- 278. MR MR0068874 (16,953c) 19 [Vas98] Wolmer V. Vasconcelos, Computational methods in commutative algebra and algebraic geometry, Algorithms and Computation in Mathematics, vol. 2, Springer-Verlag, Berlin, 1998, With chap- ters by David Eisenbud, Daniel R. Grayson, Jürgen Herzog and Michael Stillman. MR 1484973 (99c:13048) 9 DEPARTMENT OF MATHEMATICS, UNIVERSITY OF SIEGEN, 57068 SIEGEN, GERMANY E-mail address: [email protected] E-mail address: [email protected] 

1111.2372

4

1111

Galois coverings of moduli spaces of curves and loci of curves with symmetry

Let $\ccM_{g,[n]}$, for $2g-2+n>0$, be the stack of genus $g$, stable algebraic curves, endowed with $n$ unordered marked points. Looijenga introduced the notion of Prym level structures in order to construct smooth projective Galois coverings of the stack $\ccM_{g}$. In \S 2 of this paper, we introduce the notion of Looijenga level structure which generalizes Looijenga construction and provides a tower of Galois coverings of $\ccM_{g,[n]}$ equivalent to the tower of all geometric level structures over $\ccM_{g,[n]}$. In \S 3, Looijenga level structures are interpreted geometrically in terms of moduli of curves with symmetry. A byproduct of this characterization is a simple criterion for their smoothness. As a consequence of this criterion, it is shown that Looijenga level structures are smooth under mild hypotheses. The second part of the paper, from \S 4, deals with the problem of describing the D-M boundary of level structures. In \S 6, a description is given of the nerve of the D-M boundary of abelian level structures. In \S 7, it is shown how this construction can be used to "approximate" the nerve of Looijenga level structures. These results are then applied to elaborate a new approach to the congruence subgroup problem for the Teichm\"uller modular group.

math.AG

math

Galois coverings of moduli spaces of curves and loci of curves with symmetry Marco Boggi October 23, 2018 Abstract Let Mg,[n], for 2g − 2 + n > 0, be the stack of genus g, stable algebraic curves over C, endowed with n unordered marked points. In [15], Looijenga introduced the notion of Prym level structures in order to construct smooth projective Galois coverings of the stack Mg. In §2 of this paper, we introduce the notion of Looijenga level structure which generalizes Looijenga construction and provides a tower of Galois coverings of Mg,[n] equivalent to the tower of all geometric level structures over Mg,[n]. In §3, Looijenga level structures are interpreted geometrically in terms of moduli of curves with symmetry. A byproduct of this characterization is a simple criterion for their smoothness. As a consequence of this criterion, it is shown that Looijenga level structures are smooth under mild hypotheses. The second part of the paper, from §4, deals with the problem of describing the In §6, a description is given of the D -- M boundary of Looijenga level structures. nerve of the D -- M boundary of abelian level structures. In §7, it is shown how this construction can be used to "approximate" the nerve of Looijenga level structures. These results are then applied to elaborate a new approach to the congruence sub- group problem for the Teichmuller modular group along the lines of [6]. MSC2010: 14H10, 14H15, 14H30, 32G15, 30F60. 1 Level structures over moduli of curves In this first section we give a preliminary exposition of the theory of level structures over moduli of curves, mostly needed in order to fix notation. So, let Mg,n, for 2g − 2 + n > 0, be the stack of n -- pointed, genus g, stable algebraic curves over C. It is a regular connected proper D -- M stack over C of dimension 3g − 3 + n, and it contains, as an open substack, the stack Mg,n of n -- pointed, genus g, smooth algebraic curves over C. We will keep the same notations to denote the respective underlying analytic and topological stacks. There is a natural action of the symmetric group Σn on the n labels which mark the curves parametrized by the stack Mg,n. This induces an action of Σn on Mg,n which is 1 2 1 LEVEL STRUCTURES OVER MODULI OF CURVES stack-theoretically free. The quotient by this action is denoted by Mg,[n] and is the stack of genus g, stable algebraic curves over C, endowed with n unordered marked points. The open substack parametrizing smooth curves is then denoted by Mg,[n]. Let Sg,n be an n-punctured, genus g Riemann surface. Then, the universal cover of Mg,n, the Teichmuller space Tg,n, is the stack of n -- pointed, genus g, smooth complex analytic curves (E → U , s1, . . . , sn) endowed with a topological trivialization Φ : Sg,n × U ∼→ E r n[ i=1 si(U ) over U , where two such trivializations are considered equivalent when they are homotopic over U . We then denote by (E → U , s1, . . . , sn, Φ) the corresponding object of Tg,n or, when U is just one point, simply by (E, Φ). From the existence of Kuranishi families, it follows that the complex analytic stack Tg,n is representable by a complex manifold (cf. [2], for more details on this approach). Then, it is not hard to prove that the complex manifold Tg,n is contractible and that the natural map of complex analytic stacks Tg,n → Mg,[n] is a universal cover. Its deck transformation group is described as follows. Let Hom+(Sg,n) be the group of orientation preserving self-homeomorphisms of Sg,n and by Hom0(Sg,n) the subgroup consisting of homeomorphisms homotopic to the iden- tity. The mapping class group Γg,[n] is defined to be the group of homotopy classes of homeomorphisms of Sg,n which preserve the orientation: Γg,[n] := Hom+(Sg,n)(cid:14) Hom0(Sg,n), where Hom0(Sg,n) is the connected component of the identity in the topological group of homeomorphisms Hom+(Sg,n). This group then is the deck transformation group of the covering Tg,n → Mg,[n]. There is a short exact sequence: 1 → Γg,n → Γg,[n] → Σn → 1, where Σn denotes the symmetric group on the set of punctures of Sg,n and Γg,n is the deck transformation group of the covering Tg,n → Mg,n. There is a natural way to define homotopy groups for topological D -- M stacks, as done, for instance, by Noohi in [19] and [20]. Therefore, the choice of a point a = [C] ∈ Mg,[n] and a homeomorphism φ : Sg,n → C identifies the topological fundamental group π1(Mg,[n], a) with the Teichmuller modular group Γg,[n]. The following notation will turn out to be useful later. For a given oriented Rie- mann surface S of negative Euler characteristic, let us denote by Γ(S) the mapping class group of S and by M(S) and M(S), respectively, the D -- M stack of smooth complex curves homeomorphic to S and its D -- M compactification. In particular, Γg,[n] := Γ(Sg,n), Mg,[n] := M(Sg,n) and Mg,[n] := M(Sg,n). Let Mg,[n]+1 be the moduli stack of genus g, stable algebraic curves over C, endowed with n unordered labels and one distinguished label Pn+1. The morphism Mg,[n]+1→Mg,n, 3 induced by forgetting Pn+1, is naturally isomorphic to the universal n-punctured, genus g curve p : C → Mg,[n]. Since p is a Serre fibration and π2(Mg,[n]) = π2(Tg,n) = 0, there is a short exact sequence on fundamental groups 1 → π1(Ca, a) → π1(C , a) → π1(Mg,[n], a) → 1, where a is a point in the fiber Ca. By a standard argument this defines a monodromy representation: ρg,[n] : π1(Mg,[n], a) → Out(π1(Ca, a)), called the universal monodromy representation. Let us then fix a homeomorphism φ : Sg,n → Ca and let Πg,n be the fundamental group of Sg,n based in φ−1(a). Then, the representation ρg,[n] is identified with the faithful representation Γg,[n] ֒→ Out(Πg,n), induced by the action, modulo homotopy, of Γg,[n] on the Riemann surface Sg,n. Let us give Πg,n the standard presentation: Πg,n = hα1, . . . αg, β1, . . . , βg, u1, . . . , un gY i=1 [αi, βi] · un · · · u1i, where ui, for i = 1, . . . , n, is a simple loop around the puncture Pi. For n ≥ 1, let A(g, n) be the group of automorphisms of Πg,n which fix the set of conjugacy classes of all ui. For n = 0, let instead A(g, 0) be the image of A(g, 1) in the automorphism group of Πg := Πg,0. Finally, let I(g, n) be the group of inner automorphisms of Πg,n. With these notations, the Nielsen realization Theorem says that the representation ρg,[n] induces an isomorphism Γg,[n] ∼= A(g, n)/I(g, n). In this paper, a level structure Mλ is a finite, connected, Galois, ´etale covering of the stack Mg,[n] (by ´etale covering, we mean here an ´etale, surjective, representable morphism of algebraic stacks), therefore it is also a regular D -- M stack. The level associated to Mλ is the finite index normal subgroup Γλ ∼= π1(Mλ, a′) of the Teichmuller group Γg,[n]. A level structure Mλ′ dominates Mλ, if there is a natural ´etale morphism Mλ′ → Mλ or, equivalently, Γλ′ ≤ Γλ. To stress the fact that Mλ is a level structure over Mg,[n] (respectively, M(S)), we will sometimes denote it by Mλ g,[n] (respectively, M(S)λ). For a given level Γλ of Γg,[n], the intersection Γλ ∩ Γg,n is also a level of Γg,[n] which we g,n. Equivalently, the g,n. The corresponding level structure is denoted by Mλ g,n is the pull-back over Mg,n → Mg,[n] of the level structure Mλ. denote by Γλ level structure Mλ The most natural way to define levels is provided by the universal monodromy repre- sentation ρg,[n]. In general, for a subgroup K ≤ Πg,n, which is invariant under A(g, n) (in such case, we simply say that K is invariant), let us define the representation: ρK : Γg,[n] → Out(Πg,n/K), whose kernel we denote by ΓK. When K has finite index in Πg,n, then ΓK has finite index in Γg,[n] and is called the geometric level associated to K. The corresponding level structure is denoted by MK. 4 2 LOOIJENGA LEVEL STRUCTURES Of particular interest are the levels defined by the kernels of the representations: ρ(m) : Γg,[n] → Sp(H1(Sg, Z/m)), for m ≥ 2. They are denoted by Γ(m) and called abelian levels of order m. The corresponding level structures are then denoted by M(m). A result by Serre asserts that an automorphism of a smooth curve of genus ≥ 1 act- ing trivially on its first homology group with Z/m coefficients, for m ≥ 3, is trivial (cf. Lemma 2.9, Ch. XVII, [2]). This implies that, for g ≥ 1, any level structure Mλ over Mg,[n], which dominates an abelian level structure M(m), with m ≥ 3, is representable in the category of algebraic varieties. 2 Looijenga level structures There is another way to define levels of Γg,[n] which turn out to be more tractable in a series of applications. They were introduced by Looijenga in [15]. Here, we generalize his definition and, in Section 3, we will give a geometric interpretation of the construction. The idea, which underlies this construction and is more or less implicit in Looijenga's definition, is that the information present in the algebraic fundamental group of an hyper- bolic curve can be recovered from the first homology groups of all its finite coverings with the induced Galois actions. This idea was further developed by Mochizuki [17] in his proof of the anabelian conjecture for hyperbolic curves. Here, this idea is fully implemented to the study of Galois coverings of moduli spaces of curves. Let K be a normal finite index subgroup of Πg,n and let pK : SK → Sg,n be the ´etale Galois covering with deck transformation group GK, associated to such subgroup. There is then a natural monomorphism GK ֒→ Γ(SK) and the quotient of the normalizer NΓ(SK )(GK) by GK identifies with a finite index subgroup of Γg,[n]. If we assume, moreover, that K is invariant for the action of Γg,[n], then any homeomorphism f : Sg,n → Sg,n lifts to a homeomorphism f : SK → SK. So, there is a natural short exact sequence: 1 → GK → NΓ(SK )(GK) → Γg,[n] → 1. (1) Let us assume that K is a proper invariant subgroup of Πg,n. Then, the covering pK : SK → Sg,n ramifies non-trivially over all punctures of Sg,n. From Hurwitz's formula, for g ≥ 1 or n ≥ 4, it follows that the genus of the compact Riemann surface SK obtained filling in the punctures of SK is at least one. For m ≥ 2, let us then consider the natural representation ρ(m) : Γ(SK) → Sp(H1(SK, Z/m)). Remark 2.1. For m ≥ 3 and m = 0, the restriction of ρ(m) to GK is faithful. For m = 2, the restriction ρ(m)GK is faithful unless GK contains a hyperelliptic involution ι (cf. Lemma 2.9, Ch. XVII, [2]), in which case ι generates the kernel of ρ(m)GK . In particular, for g ≥ 1 or, for g = 0, if [Π0,n : K] > 2, the group GK does not contain a hyperelliptic involution and so the restriction of ρ(2) to GK is also faithful. 5 Let us assume that the restriction ρ(m)GK is faithful and let us denote by GK its image. For m ≥ 2, there is a natural representation: ρK,(m) : Γg,[n] → NSp(H1(SK ,Z/m))(GK). GK. Let us denote the kernel of ρK,(m) by ΓK,(m) and call it the Looijenga level associated to the subgroup K of Πg,n. The corresponding level structure is denoted by MK,(m). Geometric levels can also be described in terms of the exact sequence (1). The geometric level ΓK associated to K is indeed the set of elements of Γg,[n] which admit a lift, through the natural epimorphism NΓ(SK )(GK) ։ Γg,[n], to the centralizer ZΓ(SK )(GK). Thus, there is a short exact sequence: 1 → Z(GK) → ZΓ(SK )(GK) → ΓK → 1, (2) where Z(GK) is the center of the group GK. If this center is trivial, the geometric level ΓK is then identified with the centralizer of GK inside Γ(SK). Let us observe that a normal subgroup H of NΓ(SK )(GK) such that H ∩ GK = {1} centralizes GK, since, for x ∈ H and y ∈ GK, it holds xyx−1y−1 ∈ H ∩ GK. Therefore, if Z(GK) = {1}, then the centralizer ZΓ(SK )(GK) is the maximal normal subgroup of the normalizer NΓ(SK )(GK) which intersects trivially the group GK. For a given finite index invariant subgroup K E Πg,n, the subgroup K ′ := [K, K]K m of K, generated by commutators and m-th powers, is invariant and of finite index in Πg,n and the associated geometric level ΓK ′ is contained in the Looijenga level ΓK,(m). Conversely, it holds: Theorem 2.2. For 2g − 2 + n > 0, let K be a finite index invariant subgroup of Πg,n. Then, for m ≥ 3, or, for m ≥ 2, if the quotient group GK does not contain a hyperelliptic involution (cf. Remark 2.1), there is an inclusion of levels ΓK,(m) E ΓK. Proof. The action of the group GK on the surface SK induces, for m ≥ 2, a faithful action on H1(SK, Z/m)). Let us denote by ΓK,(m) the kernel of the natural representation ρK,(m) : NΓ(SK )(GK) → Sp(H1(SK, Z/m)). It follows that ΓK,(m) ∩ GK = {1} and then ΓK,(m) is contained in the centralizer of GK. Hence, the Looijenga level ΓK,(m) is contained in the geometric level ΓK. Corollary 2.3. For any fixed integer m ≥ 2, the set of Looijenga levels {ΓK,(m)}K EΠg,n forms an inverse system of finite index normal subgroups of Γg,[n] which defines the same profinite topology as the tower of all geometric levels {ΓK}K EΠg,n. Corollary 2.3 vindicates the claim we made at the beginning of this section. For in- stance, in the study of the congruence topology of the Teichmuller group Γg,n, Looijenga 6 3 LEVEL STRUCTURES AND LOCI OF CURVES WITH SYMMETRY levels can replace geometric levels. In Section 7, the utility of this approach will emerge more clearly. In the hypotheses of Theorem 2.2, the group GK acts faithfully on the homology group H1(SK, Z/m). So, the intersection of GK with the level Γ(m) of Γ(SK) is trivial and, by definition of the Looijenga level ΓK,(m), the natural epimorphism NΓ(SK )(GK) ։ Γg,[n] induces, for all m ≥ 2, an isomorphism: ΓK,(m) = NΓ(SK )(GK) ∩ Γ(m) ∼= ΓK,(m). (3) More explicitly, an f ∈ ΓK,(m) has a unique lift f : SK → SK which acts trivially on H1(SK, Z/m). Hence, by Theorem 2.2, for every m ≥ 2, the Looijenga level ΓK,(m) has an associated Torelli representation: tK,(m) : ΓK,(m) → ZSp(H1(SK ,Z))(GK). A natural question is whether the image of tK,(m) has finite index in its codomain for any invariant finite index subgroup K of Πg,n as above, and m ≥ 2. This question was addressed positively by Looijenga in [16], for n = 0 and the levels associated to the subgroup Π2 (the so-called Prym levels). 3 Level structures and loci of curves with symmetry In this section, we give a geometric construction which describes Looijenga level structures in terms of moduli of curves with a given group of automorphisms and of abelian level structures. This has some immediate applications to the problem of describing locally the D -- M compactification of a Looijenga level structure. As above, let K be an invariant finite index subgroup of Πg,n and let pK : SK → Sg,n be the ´etale Galois covering with deck transformation group GK, associated to such subgroup. Let C be a smooth n-punctured, genus g curve. Since the subgroup K is invariant for the action of Γg,[n], it determines the same finite index subgroup of π1(C) for any given marking Sg,n → C. Let then CK → C be the corresponding Galois covering. A marking φ : SK ∼→ CK identifies the group of automorphims Aut(CK) of the curve CK with a finite subgroup of Γ(SK). In this way, the Galois group of the covering CK → C is identified with some conjugate of GK in Γ(SK). The theory of Riemann surfaces with symmetry (cf. [11]) describes the locus of M(SK), parametrizing curves which have a group of automorphisms conjugated to GK as an irre- ducible closed substack MGK of M(SK) with at most normal crossing singularities, whose normalization M′ GK is a smooth GK-gerbe over the moduli stack Mg,[n]. In particular, there is a natural short exact sequence: 1 → GK → π1(M′ GK ) → Γg,[n] → 1. A connected and analytically irreducible component of the inverse image of MGK , via the covering map T (SK) → M(SK), is obtained as the fixed point set TGK for the action of the subgroup GK < Γ(SK) on the Teichmuller space T (SK). 7 The submanifold TGK of the Teichmuller space T (SK) is described as the set (D, φ) of Teichmuller points of T (SK) such that the group of automorphisms of the curve D contains a subgroup which is topologically conjugated to GK by means of the homeomorphism ∼→ D (cf. [11], Theorem A and B). There is then a natural isomorphism of complex φ : SK manifolds TGK ∼= Tg,n. From this description, it follows that there is an isomorphism NΓ(SK )(GK) ∼= π1(M′ GK ) and then we recover the short exact sequence (1) of Section 2. The short exact sequence (2) of Section 2 and the subsequent remarks have a geometric If the center of GK is trivial, then the geometric level structure interpretation as well. MK → Mg,[n] is the universal trivializing covering for the gerbe M′ GK → Mg,[n], i.e. the pull-back of this gerbe along a morphism X → Mg,[n] is the trivial GK-gerbe over X if and only if this morphism factors through the geometric level structure MK → Mg,[n]. In the hypotheses of Theorem 2.2, the group GK acts faithfully on the homology group H1(SK, Z/m). Therefore, independently from the triviality of the center of GK, the pull- GK → Mg,[n] along the Looijenga level structure MK,(m) → Mg,[n] is back of the gerbe M′ the trivial GK-gerbe over MK,(m). In conclusion, the isomorphism (3) of Section 2 yields a simple geometric interpretation of Loojenga level structures. It follows from (3) that the Looijenga level structure MK,(m) is isomorphic to any of the irreducible components of the pull-back of the abelian level structure M(SK)(m) → M(SK) over the natural morphism M′ GK → M(SK): MK,(m) −→ M(SK)(m) y M′ GK (cid:3) −→ M(SK). y The substack MGK of M(SK) has normal crossing singularities if and only if the following occurs. Let CK and an embedding of Aut(CK) in Γ(SK) be defined as above. Then, the finite subgroup Aut(CK) < Γ(SK) may contain, besides GK, a Γ(SK)-conjugate of GK distinct from GK. If this happens, the stack MGK self-intersects transversally inside the moduli stack M(SK), in the point parametrizing CK. The situation is different for level structures dominating an abelian level. Proposition 3.1. i.) For m ≥ 3, an irreducible component of the inverse image of the closed substack MGK of M(SK), in the abelian level structure M(SK)(m), is smooth and isomorphic to the Looijenga level structure MK,(m) over Mg,[n]. ii.) For m = 2, a self-intersection of such irreducible component may occur only in the locus parametrizing hyperelliptic curves and, if the group GK does not contain a hy- perelliptic involution, its normalization is isomorphic to the Looijenga level structure MK,(2) over Mg,[n]. Proof. The image of TGK in the abelian level structure M(SK)(m) has self-intersections in the points parameterizing the curve CK, if and only if its automorphism group Aut(CK) contains two distinct Γ(m)-conjugates of GK. 8 3 LEVEL STRUCTURES AND LOCI OF CURVES WITH SYMMETRY For m ≥ 3 and, in case CK is not hyperelliptic, for m = 2 as well, the restriction of the natural representation ρ(m) : Γ(SK) → Sp(H1(SK, Z/m) to Aut(CK) is faithful. Let then f be an element of the abelian level ker ρ(m) of Γ(SK) such that both GK and f GKf −1 are contained in the finite subgroup Aut(CK). It holds: ρ(m)(f GKf −1) = ρ(m)(f )ρ(m)(GK)ρ(m)(f −1) = ρ(m)(GK). Therefore, it follows that f GKf −1 = GK. The most obvious way to compactify a level structure Mλ over Mg,[n] is to take the of Mg,[n] in the function field of Mλ. This definition can be formulated normalization M more functorially in the category of regular log schemes as done by Mochizuki in [18]. It is λ log easy to see that the natural morphism of logarithmic stacks (M g,[n] is log-´etale, where ( )log denotes the logarithmic structure associated to the respective D -- M boundaries ∂Mλ := M r Mλ and ∂Mg,[n] := Mg,[n] r Mg,[n]. Viceversa, by the log purity Theorem, any finite, connected, log ´etale Galois covering of M log g,[n] is of the above type. )log → M λ λ A basic property of (compactified) level structures is the following: Proposition 3.2. If a level Γλ is contained in an abelian level of order m, for some m ≥ 3, λ then the level structure M g,n is represented by a projective variety. Proof. Even though this result is well known (cf. (ii), Proposition 2 [10]), its proof is rather technical and we prefer to give a sketch from the point of view of Teichmuller theory whose ideas will be useful later. For a given stable n-pointed, genus g curve C, let N be its singular set and P its set of marked points. A degenerate marking φ : Sg,n → C is a continuous map such that C r Im φ = P, the inverse image φ−1(x), for all x ∈ N , is a simple closed curve (briefly, s.c.c.) on Sg,n and the restriction of the marking φ : Sg,n r φ−1(N ) → C r (N ∪ P) is a homeomorphism. The Bers bordification T g,n of the Teichmuller space Tg,n is the real analytic space which parametrizes pairs (C, φ), consisting of a stable n-pointed, genus g curve C and the homotopy class of a degenerate marking φ : Sg,n → C (cf. §3, Ch. II in [1] for more details on this construction). The natural action of Γg,[n] on Tg,n extends to T g,n. However, this action is not anymore proper and discontinuous, since a boundary stratum has for inertia group the free abelian group generated by the Dehn twists along the simple closed curves of Sg,n which are collapsed on such boundary stratum. The geometric quotient T g,n/Γg,[n] identifies with the real-analytic space underlying the coarse moduli space M g,[n] of stable n-pointed, genus g curves. Instead, the quotient stack [T g,n/Γg,[n]] only admits a non-representable natural map to the D -- M stack Mg,[n], because of the extra-inertia at infinity. We can assume that the given level Γλ is contained in the pure mapping class group Γg,n. From the universal property of the normalization, it then follows that the level structure 9 λ λ is the relative moduli space of the morphism [T g,n/Γλ] → Mg,n. In order to prove is representable, we have to show that for all x = (C, φ) ∈ ∂Tg,n, the stabilizer Γλ x x ∩ Ix, where Ix is the free abelian subgroup of Γg,n generated M that M equals its normal subgroup Γλ by the Dehn twists along the s.c.c.'s on Sg,n which are contracted by the map φ. It is easy to see that Γx/Ix is naturally isomorphic to the automorphism group of the complex stable curve C and that elements in Γ(m) ∩ Γx project to automorphisms acting trivially on H1(C, Z/m). Therefore, the claim and then the proposition follows from the fact that, for m ≥ 3, the only such automorphism is the identity (cf. Lemma 4 [10]). Remarks 3.3. i.) It is interesting to notice that, for n ≥ 2 and m ≥ 3, the abelian level (m) structure M(m) g,[n] is representable while its D -- M compactification M g,[n] is not, because stable curves with double pointed rational tails have a non-trivial automorphism which acts trivially on the homology. This shows that Proposition 3.2 formulates a non-trivial property of level structures. ii.) For 2g − 2 + n > 0, let K be a finite index invariant subgroup of Πg,n with the properties that the quotient Πg,n/K surjects on H1(Sg, Z/ℓ), for some ℓ ≥ 3, and the inverse image of any non-peripheral s.c.c. γ on Sg,n, via the covering map pK : SK → Sg,n, is a union of non-separating curves. Then, it holds ΓK,(m) ≤ Γg,n and, by Theorem 2.2, it holds as well ΓK,(m) ≤ Γ(ℓ), for all m ≥ 2. Therefore, it follows that (ℓ) the corresponding compactified Looijenga level structure M g,n, for ℓ ≥ 3, and then, by Proposition 3.2, is representable. dominates M K,(m) Returning to moduli of curves with symmetry, let us consider the Zariski closure MGK of the locus MGK in the D -- M compactification M(SK). The closed substack MGK consists of the points [C] ∈ M(SK) such that the automorphisms group of the curve C contains a subgroup conjugated to GK via a degenerate marking φ : SK → C. This makes sense because an automorphism of C preserves the sets N and P of nodes and labels of C and is determined by its restriction to C r (N ∪ P), a self-homeomorphism of SK which preserves the closed submanifold φ−1(N ) is determined by its restriction to SK r φ−1(N ) and the restriction φ : SK r φ−1(N ) → C r (N ∪ P) is a homeomorphism. By means of a degenerate marking φ : Sg,n → C, the finite group Aut(C) then pulls back to a, not necessarily finite, subgroup Hφ of the mapping class group Γ(SK). The closed substack MGK of the moduli stack M(SK) has a self-intersection in the point parameterizing the curve C, if and only if, for some (and then for all) degenerate marking φ : Sg,n → C, the associated subgroup Hφ < Γ(SK) contains two distinct Γ(SK)- conjugates of GK. The GK-equivariant universal deformation of a stable curve endowed with a GK-action is formally smooth (cf. §5.1.1 [5]). It follows that the stack MGK has at most normal crossing singularities and its normalization M GK is a smooth GK-gerbe over the moduli stack Mg,[n]. ′ 10 3 LEVEL STRUCTURES AND LOCI OF CURVES WITH SYMMETRY Let us describe the inverse image of the closed substack MGK in the compactified abelian level structure M(SK)(m). The natural morphism of stacks M(SK)(m) → M(SK) is log-´etale with respect to the logarithmic structures associated to the repsective D -- M boundaries. By base-change via ′ the natural morphism M GK is also log-´etale with respect to the logarithmic structures associated to the respective D -- M boundaries. From the log-purity Theorem, it then follows that the connected components GK ×M(SK ) M(SK)(m) are normal and hence isomorphic to the level of the fiber product M GK → M(SK), the morphism M ′ ′ GK ×M(SK ) M(SK)(m) → M ′ structure M K,(m) . Let us now describe locally the image of the level structure M abelian level structure M(SK)(m). K,(m) in the compactified Lemma 3.4. For 2g − 2 + n > 0, let K be a proper invariant finite index subgroup of Πg,n. Let [C] ∈ ∂MGK be such that the natural representation Aut(C) → GL(H1(C, Z/m)) is injective for a given m ≥ 2. Then, an irreducible component of the inverse image of MGK ⊂ M(SK) in the abelian level structure M(SK)(m) is normal in a neighborhood of a point parametrizing the curve C. Proof. A connected and analytically irreducible component of the inverse image of MGK in the Bers bordification T (SK) is given by the fixed point set T GK for the action of ∼= Tg,n extends to an the subgroup GK of Γ(SK). Then, the natural isomorphism TGK isomorphism of real-analytic spaces T GK ∼= T g,n. A degenerate marking φ : SK → C, for a point [C] ∈ ∂M(SK), determines a point P ∈ ∂T (SK). The stabilizer Γ(SK)P of P , for the action of the Teichmuller group Γ(SK) on the Bers bordification T (SK), is then described by the short exact sequence: 1 → IP → Γ(SK)P → Aut(C) → 1, where IP is the free abelian group generated by the Dehn twists along the s.c.c.'s on SK which are contracted by the marking φ. A self-intersection of MGK occurs in the boundary point [C] ∈ ∂MGK , if and only if the subgroup Γ(SK)P of Γ(SK) contains two conjugates of GK which project to distinct subgroups of Aut(C). Similarly, the image of T GK in the abelian level structure M(SK)(m) has self-intersections in the points parameterizing the curve C, if and only if Γ(SK)P contains two Γ(m)-conjugates of GK which project to distinct subgroups of Aut(C). If, for the given m ≥ 2, the natural representation Aut(C) → GL(H1(C, Z/m)) is injective, then two finite subgroups of Γ(SK)P , which differ by conjugation by an element of the abelian level Γ(m), project to the same subgroup of Aut(C). This proves the lemma. Let us then show that, under suitable hypotheses, the compactified Looijenga level structure M K,(m) fits in the commutative diagram: 11 K,(m) M y M ′ GK ֒→ M(SK)(m) y (cid:3) → M(SK). Theorem 3.5. For 2g − 2 + n > 0, let K be a finite index invariant subgroup of Πg,n satisfying the hypotheses of ii.) Remarks 3.3. Then, the closed substack MGK of M(SK) does not meet the hyperelliptic locus and, for m ≥ 2, the associated Looijenga level structure M over Mg,[n] is representable and isomorphic to any of the irreducible components of the inverse image of MGK in the abelian level structure M(SK)(m). K,(m) is representable for all m ≥ 2. In Proof. By ii.) Remarks 3.3, the level structure M ′ GK ×M(SK ) M(SK)(2) → M(SK)(2) does not particular, the image of the finite morphism M meet the hyperelliptic locus of M(SK)(2), which has a generic non-trivial automorphism. But then MGK does not meet the hyperelliptic locus of M(SK) either. K,(m) If the hypotheses of ii.) Remarks 3.3, are satisfied, from the proof of Proposition 3.2, it follows that, for all m ≥ 2 and all points [C] ∈ ∂MGK ⊂ M(SK), the natural representation Aut(C) → GL(H1(C, Z/m)) is injective. Therefore, by Lemma 3.4 and Proposition 3.1, for m ≥ 2, the irreducible components of the inverse image of the closed substack MGK of M(SK) in the abelian level structure M(SK)(m) have no self-intersections and so they are isomorphic to M K,(m) . , for m ≥ 2, Remark 3.6. From the above description of a Looijenga level structure M if the inverse image of the closed substack we can derive a simple smoothness criterion: MGK of M(SK) in the abelian level structure M(SK)(m) avoids its singular locus, then the compactified Looijenga level structure M is smooth. This easily follows from the fact that the closed substack MGK meets transversally the branch locus of the covering M(SK)(m) → M(SK), which is contained in the D -- M boundary of M(SK). K,(m) K,(m) There is a very elementary and effective method to describe the compactifications M λ locally in the analytic topology. The natural morphism M D -- M boundary. Therefore, we need just to consider the case of a point x ∈ ∂Mλ. , → Mg,[n] is ´etale outside the λ Let us observe first that, since the natural morphism of stacks Mg,n → Mg,[n] is ´etale, λ it is not restrictive to consider only level structures M over Mg,n. Let then B → Mg,n be an analytic neighborhood of the image y of x in Mg,n such that: • local coordinates z1, ..., z3g−3+n embeds B in C3g−3+n as an open ball; • the curve C := π−1(y) is maximally degenerate inside the pull-back C π→ B of the universal family over B; • an ´etale groupoid representing Mg,n trivializes over B to Aut(C) × B ⇒ B. 12 3 LEVEL STRUCTURES AND LOCI OF CURVES WITH SYMMETRY Let {Q1, . . . , Qs} be the set of singular points of C and let zi, for i = 1, . . . , s, parametrize curves where the singularity Qi subsists. The discriminant locus ∂B ⊂ B of π has then equation z1 · · · zs = 0. Let U = B r ∂B and suppose that a ∈ U. The natural morphism U → Mg,n induces a homomorphism of fundamental groups: ψU : π1(U, a) → π1(Mg,n, a) ≡ Γg,n. λ A connected component U λ of U ×Mg,n M of the group π1(U, a). is then determined by the subgroup ψ−1 U (Γλ) The neighborhood U is homotopic to the s -- dimensional torus (S1)s. Therefore, the fundamental group π1(U, a) is abelian and freely generated by simple loops γi around the divisors zi, for i = 1, . . . , s. Since we have fixed a homeomorphism Sg,n ∼→ Ca, the loops γi, for i = 1, . . . , s, can be lifted to disjoint, not mutually homotopic loops γi on Sg,n, which do not bound a disc with less than 2 punctures. The loop γi is mapped by ψU exactly in the Dehn twist τγi of Γg,n, for i = 1, . . . , s. In particular, the representation ψU is faithful. Let EΣ(C) be the free abelian group generated by the edges of the dual graph Σ(C) of the stable curve C. The edges of the dual graph correspond to the isotopy classes of the s.c.c. γe in Sg,n which become isotrivial specializing to C. Let us then identify the group EΣ(C) both with the free abelian group generated in Γg,n by the set of Dehn twists {τγe} and with the fundamental group π1(U, a). For a given level Γλ, let us denote by ψλ epimorphism Γg,n → Γg,n/Γλ. Hence, the kernel of ψλ we call the local monodromy kernel in U of the level Γλ. U the composition of ψU with the natural U is identified with EΣ(C) ∩ Γλ, which The local monodromy kernels for the abelian levels Γ(m) of Γg,n are easy to compute. Let NΣ(C) and SΣ(C) be respectively the subgroups of EΣ(C) generated by edges correspond- ing to non-separating s.c.c. and by edges corresponding to separating s.c.c.. Let instead PΣ(C) be the subgroup generated by elements of the form e1 −e2, where {e1, e2} corresponds to a cut pair on C. Theorem 3.7. With the above notations, the kernel of ψ(m) is given by: U m NΣ(C) + PΣ(C) + SΣ(C). Therefore, the singular locus of the abelian level structure M in the strata parametrized by cut pairs on Sg,n. (m) g,[n], for m ≥ 2, is contained γ0 ·. . .·τ hk Proof. Let σ := {γ0, . . . , γk} be a set of disjoint, non-mutually homotopic non-peripheral s.c.c.'s on Sg,n and let us prove that τ h0 γk ∈ Γ(m) if and only if it maps to the identity in the quotient of the free abelian group E generated by the set of Dehn twists {τγ}γ∈σ by its subgroup N + P + S generated by m-th powers of Dehn twists τγ, for γ ∈ σ, bounding pair maps τγτ −1 γ ′ , for γ, γ′ ∈ σ a cut pair, and Dehn twists τγ, for γ ∈ σ a separating s.c.c.. Choosing an adapted system of generators for the homology of the compact Riemann γk ∈ Γ(m) surface Sg, it is easy to see that N +P +S < Γ(m). So let us assume that τ h0 and let us show that this element maps to the identity in the quotient E/(N + P + S). γ0 ·. . .·τ hk 13 γ0 · . . . · τ hk Replacing τ h0 γk by an element which is congruent modulo N + P + S, we can assume that the set σ does not contain cut pairs or separating s.c.c.'s and that it holds 0 ≤ hi < m, for all i = 0, . . . , k. From a simple topological argument, it then follows that, for two given distinct s.c.c.'s γi and γj in the set σ, there is a non-separating s.c.c. γ on Sg,n which is disjoint from the other s.c.c.'s in σ and intersects γi and γj transversally in one point. For a given s.c.c. α on Sg,n, let us denote by α ∈ H1(Sg, Z/m) the cycle associated to a given orientation of α. For suitable orientations of γ, γi and γj, it holds: γ0 · . . . · τ hk τ h0 γk (γ) = γ + γi hi + γj hj ∈ H1(Sg, Z/m). Since the cycles γi and γj are primitive and linearly independent in the homology group H1(Sg, Z/m), the above identity implies that hi = hj = 0. Let K be a finite index invariant subgroup of Πg,n, satisfying the hypotheses of ii.) be the associated Looijenga level structure over Mg,[n], which Remarks 3.3, and let M then is representable and dominates the abelian level M (ℓ) g,n, for some ℓ ≥ 3. K,(m) Let us denote by ψ : M(SK)(m) → M(SK) the natural finite morphism. rem 3.5, for m ≥ 2, we identified the level structure M of the inverse image ψ−1(MGK ), which is a normal and proper subvariety of M(SK)(m). It is easy to determine the boundary strata of M(SK)(m) which are met by the inverse image ψ−1(MGK ), in terms of the covering pK : SK → Sg,n and systems of s.c.c.'s on Sg,n. K,(m) In Theo- with an irreducible component Definition 3.8. We say that a set σ = {γ0, . . . , γk} of disjoint, non-mutually homotopic non-peripheral s.c.c.'s on Sg,n is admissible. Let us also denote by the same letter σ their union in Sg,n. For an admissible set σ of s.c.c.'s on Sg,n, the inverse image p−1 set of s.c.c.'s on SK and so determines a closed boundary stratum B(m) p−1 K (σ) (cf. Section 4 for details on this correspondence). It is clear that: K (σ) is also an admissible of M(SK)(m) ψ−1(MGK ) ∩ ∂M(SK)(m) ⊂ [ σ⊂Sg,n B(m) p−1 K (σ) . Therefore, from Remark 3.6 and Theorem 3.7, it follows: Theorem 3.9. For 2g − 2 + n > 0, let K be a finite index invariant subgroup of Πg,n such that the covering pK : SK → Sg,n has the property that, for all admissible sets σ of s.c.c.'s, the inverse image p−1 K (σ) does not contain cut pairs. Then, the compact Looijenga level structure M over Mg,[n] is smooth for all m ≥ 2. K,(m) It is now possible both to clarify and improve substantially the main result of [15]. The only technical result needed from [15] is Proposition 2, which states that, for g ≥ 2 and 14 3 LEVEL STRUCTURES AND LOCI OF CURVES WITH SYMMETRY K = Π2 Sg, the inverse image p−1 argument to invariant subgroups of Πg,n satisfying more general hypotheses: g, the covering pK : SK → Sg is such that, for any admissible set σ of s.c.c.'s on K (σ) does not contain cut pairs. It is not difficult to extend the Lemma 3.10. For 2g − 2 + n > 0, let K be a finite index invariant subgroup of Πg,n such that the covering map pK : SK → Sg,n ramifies non-trivially over all punctures of Sg,n and let Kℓ := [K, K]K ℓ, for an integer ℓ ≥ 2. This is also a finite index invariant subgroup of Πg,n and the associated covering map pKℓ : SKℓ → Sg,n is such that, for an admissible set σ of s.c.c.'s on Sg,n, the inverse image p−1 (σ) contains no separating curves or cut pairs. Kℓ Proof. The condition on the covering map pK : SK → Sg,n implies, by Hurwitz Theorem, that, for an admissible set σ of s.c.c.'s on Sg,n, the inverse image p−1 K (σ) determines an admissible set of s.c.c.'s on the closed Riemann surface SK and that this is of genus at least 2 (except, possibly, for the case g = 0 and n = 3, which is irrelevant). Let us denote by φℓ : SKℓ → SK the abelian covering defined by the subgroup Kℓ of K. It is enough to prove that the dual graph ΣKℓ of the stable curve obtained collapsing on SKℓ the s.c.c.'s contained in φ−1 ℓ (σ) stays connected if we remove two edges, where σ is an admissible set of s.c.c.'s on the closed Riemann surface SK. Let ΣK be the dual graph of the stable curve obtained collapsing on SK the s.c.c.'s in σ. As in the proof of Proposition 2 [15], it is enough to consider the case in which the cardinality of σ is ≤ 2. Let us observe in the first place that the fact that Dehn twists on SK lift to homeomor- phisms of the covering surface SKℓ implies that each s.c.c. contained in φ−1 ℓ (σ) bounds two distinct connected components of SK. Then, the graph ΣKℓ does not contain loops and there is a natural action without inversions of the deck transformation group H1(K, Z/ℓ) on ΣKℓ with quotient the graph ΣK. A vertex v of the graph ΣK corresponds to a connected component S ′ of SK r σ The vertices of ΣKℓ lying over v have the same stabilizer for the action of H1(K, Z/ℓ) and this is the image of H1(S ′) in H1(K, Z/ℓ) for the homomorphism induced by inclusion. Similarly, the edges of ΣKℓ lying over an edge e of ΣK, corresponding to the s.c.c. γ of the set σ, have for stabilizer the associated cyclic subgroup Ie of H1(K, Z/ℓ). The number of vertices (resp. edges) of ΣKℓ lying over v (resp. e) is given by the index of the stabilizer of one of these vertices (resp. edges) in the group H1(K, Z/ℓ). A simple count and symmetry show that each vertex of ΣKℓ is connected to the rest of the graph by at least three edges and that, when two vertices are connected by an edge, they are connected by at least two edges. Therefore, the graph cannot be disconnected removing two edges. Looijenga's results then generalize to all pairs (g, n), for 2g − 2 + n > 0: Theorem 3.11. For 2g − 2 + n > 0, let Kℓ be a subgroup of Πg,n of the type defined in Lemma 3.10. Then, the associated compact Looijenga level structure M over Mg,[n] is smooth for all m ≥ 2 and representable for ℓ ≥ 3 or m ≥ 3. Kℓ,(m) Proof. The statement about smoothness is a straightforward consequence of Theorem 3.9 and Lemma 3.10. The hypotheses of Lemma 3.10 on the subgroup K also imply that ΓKℓ,(m) < Γg,n, for all m ≥ 2. Moreover, the quotient Πg,n/Kℓ surjects onto H1(Sg, Z/ℓ). Therefore, if ℓ ≥ 3 or m ≥ 3, it holds ΓKℓ,(m) < Γg,n ∩ Γ(k), for some k ≥ 3, and the associated compact level structure is representable. 15 It is not difficult to determine explicitly the local monodromy coefficients for the levels ΓK,(m) satisfying the hypotheses of Theorem 3.9. For a non-peripheral s.c.c. γ on Sg,n, let cγ be the order in the quotient group GK = Πg,n/K of an element of Πg,n, whose free homotopy class contains γ. By the invariance of K, the positive integer cγ depends only on the topological type of Sg,n r γ. It then holds: Proposition 3.12. Let K be a finite index invariant subgroup of Πg,n satisfying the hy- potheses of Theorem 3.9. With the same notations of Theorem 3.7, for m ≥ 2, the kernel of ψK,(m) is given by: U X e∈Σ(C) mcγee. Proof. For a non-peripheral s.c.c. γ on Sg,n, let us denote by kγ the order of the image of the Dehn twist τγ in the quotient group Γg,n/ΓK,(m) and let ∐h i=1Si := SK r p−1 K (γ). Let us show first that cγ divides kγ. This follows from the fact that any lift to SK γ , where cγ does not divide k, either switches the connected components of K (γ), and does not act trivially on H1(SK, Z/m) for m ≥ 2, or restricts to a non- K (γ), which of a power τ k SK r p−1 trivial finite order homeomorphism of some connected component Si of SK r p−1 then acts non-trivially on the image of H1(Si, Z/m) in H1(SK, Z/m), for m ≥ 2. γ ∈ Γg,[n] of the given Dehn twist lifts to a product ξγ := Qδ∈σγ τδ ∈ Γ(SK) of Dehn twists along the set σγ of disjoint s.c.c.'s on SK contained in p−1 K (γ). This lift is contained in the centralizer of GK. Therefore, another lift is of the form ξγ · α, for some α ∈ GK, where ξγ and α are commuting elements of Γ(SK). The power τ cγ Any power of ξγ fixes the subsurface Si of SK and acts trivially on the image of H1(Si, Z/m) in H1(SK, Z/m), for i = 1, . . . , h, while no non-trivial element of GK has this property. Hence, the cyclic subgroup generated by the image of ξγ in the group Sp(H1(SK, Z/m) intersects trivially the image of the group GK. It then follows that the image of ξγ in Sp(H1(SK, Z/m) is of minimal order between those of all possible lifts of τ cγ γ to Γ(SK). By hypothesis, the set σγ consists of non-separating s.c.c. and does not contain cut pairs. Hence, by Theorem 3.7, the image of ξγ in Sp(H1(SK, Z/m) has order m. Therefore, it holds kγ = mcγ. 16 4 THE BOUNDARY OF LEVEL STRUCTURES 4 The boundary of level structures and the complex of curves For simplicity, we restrict here to level structures over Mg,n. Most of the result of this section are well known or have already appeared elsewhere (§3 of [6]) but there are also some rectifications to [6] (notably, in Proposition 4.2 and Proposition 4.3). For a point [C] ∈ ∂Mg,n, let Cg1,n1 ∐. . .∐Cgh,nh be the normalization of C, where Cgi,ni, for i = 1, . . . , h, is a genus gi smooth curve with ni labels on it (the labels also include the inverse images of singularities in C). Then, there is a natural morphism, which we call boundary map: βC : Mg1,n1 × · · · × Mgh,nh −→ Mg,n. The image of Mg1,n1 × · · · × Mgh,nh by βC parametrizes curves homeomorphic to C and is called a stratum. We denote the restriction of βC to Mg1,n1 × · · · × Mgh,nh by βC and call it a stratum map. In general, these morphisms are not embeddings. A variant of the Bers bordification T g,n of the Teichmuller space Tg,n is the Harvey bordification bTg,n. This can be constructed by means of the D -- M compactification Mg,n. Let cMg,n be the real oriented blow-up of Mg,n along the D -- M boundary. Its boundary ∂ cMg,n := cMg,n r Mg,n is homeomorphic to a deleted tubular neighborhood of the D -- M boundary of Mg,n and the natural projection cMg,n → Mg,n restricts over each codimension k stratum to a bundle in k-dimensional tori. The inclusion Mg,n ֒→ cMg,n is a homotopy equivalence and then induces an inclusion of the respective universal covers Tg,n ֒→ bTg,n. From Proposition 3.2, it follows that bTg,n is representable and, therefore, is a real analytic manifold with corners containing Tg,n as an open dense submanifold. The ideal boundary of Teichmuller space is defined to be ∂bTg,n := bTg,n r Tg,n. The strata of ∂bTg,n lying above the boundary map βC, are isomorphic to the product (R+)k × bTg1,n1 × · · · × bTgh,nh. Thus, they are all contractible. Hence, ∂bTg,n is homotopy equivalent to the geometric realization of the nerve of its cover by irreducible components. This nerve is described by the simplicial complex whose simplices consist of sets of distinct, non-trivial isotopy classes of non-peripheral s.c.c. on Sg,n, such that they admit a set of disjoint representatives. This is the complex of curves C(Sg,n) of Sg,n. It is easy to check that the combinatorial dimension of C(Sg,n) is one less the complex dimension of the moduli space Mg,n, i.e.: n − 4 for g = 0 and 3g − 4 + n for g ≥ 1. The natural action Let cMλ λ g,n be the oriented real blow-up of M of Γg,n on ∂bTg,n then induces a simplicial action of Γg,n on C(Sg,n). ∼= [bTg,n/Γλ] ∼= [(∂bTg,n)/Γλ]. Therefore, the quotient C(Sg,n)/Γλ is a cellular and then also ∂ cMλ complex which realizes the nerve of the cover of ∂ cMλ by irreducible components and, g,n along ∂Mλ. It holds cMλ thus, the nerve of the cover of ∂Mλ by irreducible components. g,n g,n Definition 4.1. For 2g − 2 + n > 0, let Γλ be a level of Γg,n. Then, we define C λ(Sg,n)• to be the finite simplicial set associated to the cellular complex C(Sg,n)/Γλ. This simplicial set parameterizes the nerve of the cover of ∂Mλ by its irreducible components. Let σ be a non-degenerate k-simplex of C(Sg,n)λ • , such that, for some lift σ of σ to the curve complex C(Sg,n), it holds Sg,n r σ := Sg1,n1 ∐ . . . ∐ Sgh,nh. Then, to the simplex σ, is associated a boundary map βλ which is the restriction to an irreducible component of the pull-back β ′ of the boundary map β associated to the topological type of the Riemann surface Sg,n r σ: λ σ → M σ : δλ 17 X −→ Mg1,n1 × · · · × Mgh,nh yβ ′ λ M (cid:3) −→ yβ Mg,n. The image of βλ σ is the closed boundary stratum associated to σ. Similarly, we call λ λ the restriction βλ δλ g,n of βλ σ over Mg1,n1 × · · · × Mgh,nh the stratum map of M σ → M σ : associated to σ and its image is the boundary stratum associated to σ. By log-´etale base change, the natural morphism δλ σ → Mg1,n1 × · · · × Mgh,nh is ´etale. Similar definitions can be given with cMλ σ : ∆λ we define the ideal boundary map βλ g,n in place of M λ g,n. For a simplex σ ∈ C(Sg,n), σ via the blow-up g,n, as the pull-back of βλ σ → cMλ g,n → M λ g,n. map cMλ The fundamental group of ∆λ σ along the divisor δλ σ of δλ is a bundle over δλ and a short exact sequence: r δλ σ is described as follows. Let δλ σ. The embedding δλ σ ֒→ δλ σ be the real oriented blow-up σ is a homotopy equivalence and ∆λ σ σ) ∼= π1( δλ σ) σ in k-dimensional tori. Hence, there is an isomorphism π1(δλ 1 → M γ∈σ Z · γ → π1( ∆λ σ) → π1(δλ σ) → 1. A connected component of the fibred product ∆σ × cMg,n bTg,n is naturally isomorphic to (R+)k × bTg1,n1 × · · · × bTgh,nh. Therefore, the fundamental group of ∆σ is isomorphic to the subgroup of elements in Γg,n which stabilize one of these connected components, preserving, moreover, the order of its factors. So, if we let Γ~σ := {f ∈ Γg,n f (~γ) = ~γ, ∀γ ∈ σ}, where ~γ is the s.c.c. γ endowed with an orientation, it holds π1( ∆σ) ∼= Γ~σ and, for the trivial level, the above short exact sequence takes the more familiar form 1 → M γ∈σ Z · τγ → Γ~σ → Γg1,n1 × · · · × Γgh,nh → 1. By the same argument, more generally, it holds π1( ∆λ σ) ∼= Γλ ∩ Γ~σ. The fundamental group of βσ( ∆σ) is isomorphic to the stabilizer Γσ of σ, for the action of Γg,n on C(Sg,n) and more generally, it holds π1( βλ σ ( ∆λ σ)) ∼= Γλ ∩ Γσ. 18 4 THE BOUNDARY OF LEVEL STRUCTURES Let us observe that an f ∈ Γσ can switch the s.c.c. in σ as well as their orientations. Hence, denoting by Σσ{±} the group of signed permutations of the set σ, there is an exact sequence 1 → Γλ ∩ Γ~σ → Γλ ∩ Γσ → Σσ{±} and the Galois group of the ´etale covering δλ in Σσ{±}. Thus, the stratum map βλ σ → Im βλ σ is isomorphic to the image of Γλ ∩Γσ σ is injective if and only if Γλ ∩ Γ~σ = Γλ ∩ Γσ. In Proposition 3.1 in [6], it is wrongly claimed that all boundary maps δ(m) σ → M (m) are injective for m ≥ 3. What is actually proved there is the weaker statement (which is also an immediate consequence of Corollary 1.8 [12]): Proposition 4.2. Let Γλ be a level of Γg,n contained in some abelian level Γ(m), for m ≥ 3. Then, the group Γλ operates without inversions on the curve complex C(Sg,n). Proof. Let us show first that, given two disjoint oriented separating s.c.c.'s ~γ0 and ~γ1 which either have distinct free homotopy classes or differ in their orientations, there is no f ∈ Γλ such that f (~γ0) = ~γ1. Let us denote by S0 and S1 two disjoint subsurfaces of Sg,n such that γ0 = ∂S0 and γ1 = ∂S1, respectively. If there is an f ∈ Γλ such that f (~γ0) = ~γ1, then the homeomorphism f maps S0 on S1. If S0 or S1 are punctured, this is not possible. If they are unpunctured, then their genus is at least 1 and f acts non-trivially on the homology of Sg, with any system of coefficients. Hence f /∈ Γλ ≤ Γ(m), for m ≥ 2. The only inversions, for the action of Γλ on C(Sg,n), may occur on non-separating s.c.c.'s. But then the only possibility is that an f ∈ Γλ swops two disjoint homologous s.c.c.'s, i.e. a cut pair {γ0, γ1}. Let Sg,n r {γ0, γ1} ∼= S1 ∐ S2. If f swops γ0 and γ1, but not S1 and S2, it is easy to see that f (~γ0) = −~γ0 in the homology group H1(Sg,n, Z/m). Therefore, f /∈ Γ(m), for m ≥ 3. If instead f swops S1 and S2, arguing as above, we conclude that f /∈ Γ(m), for m ≥ 2. The statement of Proposition 3.1 in [6] holds with the following additional hypotheses: Proposition 4.3. For 2g − 2 + n > 0, let M smooth level structure M all boundary maps δλ inversions on C(Sg,n). be a level structure which dominates a over Mg,n, such that Γµ ≤ Γ(m), for some m ≥ 2. Then, λ are embeddings. In particular, the level Γλ operates without σ → M µ λ Proof. By arguments similar to those in the proof of Proposition 4.2, one is reduced to show that, for any two disjoint, non-isotopic, non-separating s.c.c.'s γ0 and γ1 on Sg,n, there is no f ∈ Γλ such that f (γ0) = γ1. µ By Theorem 3.7, the natural morphism M → Mg,n ramifies over the divisor of nodal irreducible curves. But, in the proof of Proposition 2.1 [9], it was shown that an f ∈ Γµ with the above properties would then yield a singularity in the stratum of the D -- M boundary µ of M , corresponding to the simplex {γ0, γ1}, against our hypotheses. Hence, there is no such an f ∈ Γµ and, a fortiori, in Γλ. 19 Remark 4.4. Let Γλ be a level of Γg,n satisfying the hypotheses of either Proposition 4.2 or Proposition 4.3. Then, the simplicial set C λ(Sg,n)• of Definition 4.1 can be naturally realized as a quotient in the category of simplicial sets. Let us order the vertices of the simplicial complex C(Sg,n) compatibly with the action of Γλ and let C(Sg,n)• be the sim- plicial set associated to the simplicial complex C(Sg,n) and this ordering of its vertices. The finite simplicial set C λ(Sg,n)• is then naturally isomorphic to the quotient of C(Sg,n)• by the simplicial action of the level Γλ. 5 The D -- M boundary of abelian level structures In this section, we are going to describe the boundary components of abelian level structures in terms of other simple geometric levels. Let us recall the statement of Lemma 3.4 [7]: Lemma 5.1. λ i.) Let Γλ be a level of Γg,n such that the branch locus of the covering → Mg,n is contained in the boundary divisor of nodal irreducible curves. Let M p : Γg,n → Γg,n−1 be the epimorphism induced filling in Pn on Sg,n and let M be the level structure over Mg,n−1 associated to the level p(Γλ) of Γg,n−1. Then, it holds λ π1(M ) = π1(M ). λ′ λ′ ii.) Let Γλ1 ≤ Γλ2 be two levels of Γg,n whose associated level structures satisfy the hy- λ2 is ´etale and, with pothesis in the above item. If the natural morphism M the same notations as above, it holds p(Γλ1) = p(Γλ2), then Γλ1 = Γλ2. λ1 → M Let N be the kernel of the epimorphism Πg−1,n+2 ։ Πg−1,2, induced filling in the punc- tures P1, . . . , Pn. Let us then define the Γg−1,n+2-invariant normal subgroup of Πg−1,n+2: Let Γ(m)0 E Γg−1,n+2 be the associated geometric level. For n = 0, it holds Γ(m)0 = Γ[2],m. Π(m)0 := N · Π[2],m g−1,n+2. Theorem 5.2. For 2g − 2 + n > 0, let M (m) g,n be an abelian level structure of order m ≥ 2. i.) Let α be a separating circle on Sg,n, such that Sg,n r α ∼= Sg1,n1+1 ∐ Sg2,n2+1, and let is an g,n be the associated boundary map. Then, the morphism β(m) β(m) α embedding and there is a natural isomorphism δ(m) (m) g1,n1+1 × M (m) g2,n2+1. α ∼= M α → M : δ(m) (m) α ii.) Let γ be a non-separating circle on Sg,n and β(m) γ boundary map. Then, the morphism β(m) with the above notations, there is a natural isomorphism δ(m) γ : δ(m) (m) g,n be the associated , for g > 2, is never an embedding, and, γ → M γ ∼= M (m)0 g−1,n+2. Proof. The homology H1(C, Z/m) of a stable n-pointed genus g curve of compact type C has a natural structure of non-degenerate symplectic free Z/m-module of rank 2g and the 20 5 THE D -- M BOUNDARY OF ABELIAN LEVEL STRUCTURES decomposition of C in irreducible components determines a symplectic decomposition of this module. (m) g,n the substack of M Let fM(m) g,n parametrizing curves of compact type. Then, fM(m) g,n is the moduli space of stable, n-pointed, genus g curves of compact type endowed with a symplectic isomorphism H1(C, Z/m) ∼→ ((Z/m)2g, h , i), where h , i is the standard symplectic form. It is also clear that the clutching morphism defines an embedding of the (m) product fM(m) g,n is also an embedding, by the proof of Proposition 4.2, the first part of the theorem follows. g2,n2+1 inside fM(m) g,n . Since the boundary map δ(m) g1,n1+1 × fM(m) α → M As to ii.), from Proposition 6.7 in [23], it follows that, if γ and γ′ form a cut pair on Sg,n, then there is an f ∈ Γ(m) such that f (γ) = γ′. By the description of the D -- M boundary of level structures given in Section 4, we then know that such cut pair determines a self-intersection on the boundary component of M (m) g,n parametrized by γ. Let us then prove the second part of ii.). The D -- M stack δ(m) is normal, hence naturally isomorphic to the compactification of the level structure defined by the ´etale Galois covering δ(m) γ → Mg−1,n+2. By Section 4 and Theorem 3.7, there is a short exact sequence γ 1 → Z · τ m γ → Γ(m) ∩ Γ~γ q→ π1( δ(m) γ ) → 1. Fix a homeomorphism h : Sg,n rγ → Sg−1,n+2 mapping the punctures of Sg,n in the first n punctures of Sg−1,n+2. This induces a monomorphism h♯ : Πg−1,n+2 ֒→ Πg,n and then a monomorphism h♯ : Πg−1,n+2/ Π(m)0 ֒→ H1(Sg, Z/m). Let f ∈ Γ(m) ∩ Γ~γ then f acts trivially on H1(Sg, Z/m). Therefore, f acts trivially also ) is contained in on the subgroup Πg−1,n+2/Π(m)0, hence q(f ) ∈ Γ(m)0 and the level π1( δ(m) γ Γ(m)0. To prove that the two levels are the same, let us use Lemma 5.1. Denote by p : Γg−1,n+2 → Γg−1,n+1 the epimorphism induced filling in Pn+2 and by Γλ the level in Γg−1,n+1 given by the subgroup p(π1( δ(m) γ )). We claim that Γλ = p(Γ(m)0). It is easy to check that p(Γ(m)0) = Γ(m). In particular, it holds Γλ ≤ Γ(m). For the contains a reverse inclusion, let us observe that, by the first part of the theorem, δ(m) stratum of M g,n isomorphic to M0,3 × M(m) (m) γ g−1,n+1 → Mλ morphism M(m) Therefore, the natural morphism δ(m) It remains to prove that the morphism δ(m) λ g−1,n+1 induces by restriction an ´etale g−1,n+1. This implies that Γ(m) ≤ Γλ and then Γλ = p(Γ(m)0). (m)0 g−1,n+2 is ´etale and that the branch loci of both levels over Mg−1,n+2 are contained in the divisor β0(Mg−2,n+4). By the first part of the theorem, it is enough to prove both statements for n = 0. Since, as noticed above, in this case it holds Γ(m)0 = Γ[2],m, the conclusion follows from the computation of the local monodromy coefficients for these levels (cf. Theorem 3.3.3 in [21]). γ → M g−1,n+1. γ → M 21 6 The nerve of the D -- M boundary of abelian level structures There is no obvious way to extend the description of codimension 1 strata given in Theo- rem 5.2 to higher codimension. At any rate, in this section, we are going to describe the nerve of the D -- M boundary of the abelian levels in an explicit if complicated way. (m) g,n is the In Section 4, we saw that, for all m ≥ 2, the nerve of the boundary of M simplicial set C (m)(Sg,n)• associated to the quotient cellular complex C(Sg,n)/Γ(m). As a preliminary step, we need to describe the quotient of the curve complex C(Sg,n) by the action of the Torelli group Tg,n, which is the kernel of the natural representation Γg,n → Sp2g(Z). Since the Torelli group Tg,n is contained in the abelian level Γ(3), it acts without inversions on the curve complex. So, let C(Sg,n)• be the simplicial set associated to the curve complex and an order of its vertices compatible with the action of the Torelli group Tg,n. We then define C (0)(Sg,n)• to be the quotient simplicial set C(Sg,n)•/Tg,n. Let us recall the definition of a graph of groups and of the fundamental group of a graph of groups (cf. Ch. I, §5.1 [26]). For a graph Y , let us denote by v(Y ) its vertex set, by e(Y ) its set of unoriented edges and by ~e(Y ) its set of oriented edges. For an oriented edge ~e ∈ ~e(Y ) we denote by ~e ∈ ~e(Y ) the same edge with the opposite orientation and by e ∈ e(Y ) the underlying unoriented edge. A graph of groups (G, Y ) is given by a graph Y and the data of groups Gv and Ge, for every vertex v ∈ v(Y ) and every unoriented edge e ∈ e(Y ), together with monomorphisms s~e : Ge ֒→ Gs(~e) and t~e : Ge ֒→ Gt(~e), for every oriented edge ~e ∈ ~e(Y ), where s(~e) and t(~e) are the vertices of e in the source and in the target, respectively, of the given orientation. The fundamental group π1(G, Y ) of the graph of groups (G, Y ) is the group with pre- sentation hP Ri, where P is the free product of all vertex groups Gv and of the free group on the oriented edges ~e(Y ), while R is the set of relations ~e = ~e−1 and ~e · t~e(a) · ~e = s~e(a), for all e ∈ e(Y ) and a ∈ Ge. Definition 6.1. For m 6= 1 a non-negative integer and g ≥ 1, let Hm := H1(Sg, Z/m) (in particular, H0 := H1(Sg, Z)), endowed with the symplectic form h , i induced by intersec- tion of cycles on Sg. Let Pn be the set of subsets of the set {1, . . . , n}. For 2g−2+n > 0, an n-marked graph decomposition (G, Y, d) of Hm is a graph of groups (G, Y ), whose vertex and edge groups {Gσ} are primitive subgroups of Hm, endowed with a marking d : v(Y ) → Pn. We require that the following properties be satisfied. i.) There is an isomorphism φ : π1(G, Y )ab ⊗ Z/m ∼→ Hm such that it holds φGσ = idGσ , for all simplices σ of the graph Y , where the superscript "ab" is for abelianization. ii.) It holds Sv∈v(Y ) d(v) = {1, . . . , n} and d(v) ∩ d(v′) = ∅, for v 6= v′ ∈ v(Y ). iii.) Let E be the subgroup of Hm generated by all edge groups of (G, Y ). Then, for all vertices v of Y , it holds hx, yi = 0, for all x ∈ E and y ∈ Gv. In particular, E is an isotropic sub-lattice of Hm. 22 6 THE NERVE OF THE BOUNDARY OF ABELIAN LEVEL STRUCTURES iv.) All edge groups are cyclic. Let {ei}i∈I be a set of unoriented edges. Then, if Y r∪i∈I ei is connected, the associated set of edge groups {Gei}i∈I spans a primitive subgroup of Hm of rank I. If instead {ei}i∈I is minimal for the property that Y r ∪i∈I ei is disconnected, there is a set {ui}i∈I of generators of these edge groups such that it holds Pi∈I ui = 0. v.) For v a vertex of Y , let star(v) be its star in Y and let star(v) be the number of edges contained in the star. Then, all vertex groups Gv, such that star(v) + d(v) ≤ 2, are non trivial and of rank ≥ 2. vi.) For v a vertex of Y , let E(v) be the subgroup of Gv generated by the edge groups it v is a symplectic, contains. Then, there is a decomposition Gv = E(v) ⊕ G′ non-degenerate, possibly trivial, sub-module of Hm. v, where G′ The rank of an n-marked graph decomposition (G, Y, d) is the cardinality e(Y ) of the set of unoriented edges of Y . We say that two n-marked graph decompositions (G, Y, d) and (G′, Y ′, d′) of Hm are equivalent if there is an isomorphism f : Y → Y ′ of the underlying graphs such that, for all v ∈ v(Y ), it holds d′(f (v)) = d(v) and, for every simplex σ of Y , it holds G′ f (σ) = Gσ. Let us denote by [G, Y, d] the equivalence class of the n-marked graph decomposition (G, Y, d). If n = 0, we simply say that (G, Y ) is a graph decomposition of Hm and denote by [G, Y ] the corresponding equivalence class. A k-simplex σ ∈ C(Sg,n) determines a partition ∐v∈V Sv := Sg,n r σ of the surface Sg,n. We can associate to σ the graph (G, Y ) of abelian subgroups of H1(Sg, Z/m) whose vertex groups consist of the images of the natural homomorphisms H1(Sv, Z/m) → H1(Sg, Z/m), for v ∈ V , and whose edge groups are the cyclic subgroups of H1(Sg, Z/m) determined by the s.c.c.'s in σ. The marking d : v(Y ) = V → Pn is then defined assigning to a vertex v of the graph Y the indices of the punctures on the corresponding subsurface Sv of Sg,n. In this way, it is associated to every simplex σ ∈ C(Sg,n) an n-marked graph decompo- sition (G, Y, d) of rank k + 1. Conversely, it is easy to check that, given an n-marked graph decomposition (G, Y, d) of Hm of rank k + 1, there is a k-simplex σ ∈ C(Sg,n) which has (G, Y, d) for associated n-marked graph decomposition. The partition of Sg,n determined by a simplex σ ∈ C(Sg,n) is refined by the partition determined by a simplex σ′ such that σ ⊂ σ′. Definition 6.2. Let us define a partial order on the set of all n-marked graph decompo- sitions of Hm, letting (G′, Y ′, d′) ≤ (G, Y, d) if the following conditions are satisfied: i.) the set of edge groups of (G′, Y ′, d′) is a subset of the set of edge groups of (G, Y, d); ii.) for each vertex v′ of Y ′, there is a connected subgraph of groups (K, Z) of (G, Y ) such that d′(v′) = Sv∈v(Z) d(v) and there is an isomorphism ψ : π1(K, Z)ab ⊗ Z/m ∼→ G′ with the property that ψKσ = idKσ for all simplices σ of the subgraph Z. v′ 23 The partial order so defined on the set of n-marked graph decompositions of Hm is compatible with the equivalence relation defined above and thus induces a partial order on the set of equivalence classes of n-marked graph decompositions of Hm. Definition 6.3. For m 6= 1 a non-negative integer, let G n(Hm)• be the simplicial set whose semi-simplicial set of non-degenerate simplices is defined as follows. i.) The set of k-simplices is the set of equivalence classes of n-marked graph decompo- sitions of Hm of rank k + 1. ii.) The (k − 1)-faces of a k-simplex [G, Y, d] are the equivalence classes of n-marked graph decompositions (G′, Y ′, d′) ≤ (G, Y, d) of rank k. A face map ∂e : [G, Y, d] → [G′, Y ′, d′] is assigned to the unoriented edge e ∈ e(Y ), if the set of edge groups of [G′, Y ′, d′] coincides with the set of edge groups associated to the edges of Y r e. Theorem 6.4. For 2g − 2 + n > 0 and g ≥ 1, there is a natural isomorphism of simplicial sets Ψ• : C (0)(Sg,n)• → G n(H0)•. Proof. By definition, C (0)(Sg,n)k = C(Sg,n)k/Tg,n. From the above remarks, it follows that, for k ≥ 0, there is a natural surjective map eΨk : C(Sg)k → G n(H)k. It is easy to check the compatibility of eΨ• with the face maps so that we get a map of simplicial sets. In order to prove that eΨ• is an isomorphism, we have to show that its fibers coincide with the orbits of H0, for all k ≥ 0, the map eΨk factors through the map: Since the Torelli group Tg,n acts trivially on the set of n-marked graph decompositions of the action of Tg,n on C(Sg,n)•. Ψk : C(Sg,n)k/ Tg,n → G n(H0)k. Let us then show that two simplices σ and σ′ which define equivalent n-marked graph decompositions of H of rank k +1 are in the same Tg,n-orbit. Let us choose a representative (G, Y, d) of the equivalence class and let us identify the two isomorphic n-marked graph decompositions of H with it. Let σ = {γ0, . . . , γk} and σ′ = {γ′ k}. With suitable orientations, the s.c.c.'s in the simplices σ and σ′ determine the same set of not necessarily linearly independent homology classes {a0, . . . , ak} of H0. Let us then extend this set to a generating set {a0, . . . , as} of H0 with the following properties. 0, . . . , γ′ • Each vertex group of (G, Y, d) is naturally isomorphic to the homology of a subsurface of Sg, whose topological type is a punctured surface Sg′,n′, with 2g′ − 2 + n′ > 0. Let us then define a standard set of generators {a1, b1, . . . , ag′, bg′, w1, . . . , wn′} for the group H1(Sg′,n′, Z), where, for n′ = 1, we omit w1, to be a set of generators which can be lifted to a standard set of generators for the fundamental group (cf. §1). The intersection Gy ∩ {a0, . . . , as} is then a standard set of generators for the group Gy, for every vertex y of Y . 24 6 THE NERVE OF THE BOUNDARY OF ABELIAN LEVEL STRUCTURES • There is a natural epimorphism r : π1(G, Y )ab ։ H1(Y, Z). The set {ai1, . . . , ait} of elements in {a0, . . . , as} which do not belong to any vertex group of (G, Y, d) is then projected by r to a basis of H1(Y, Z) such that r(aij ), for j = 1, . . . , t, is the cycle associated to a circuit in the graph Y which does not pass twice for the same vertex. Moreover, it holds haij , ali = 0, for all j = 1, . . . , t and l > k. By the second property, the set of homology classes {ai1, . . . , ait} can be lifted to a set it}) such that, for j = 1, . . . , t, each ij ) intersects a curve in σ (respectively, in σ′) at most once. Let s} which of disjoint s.c.c.'s {γi1, . . . , γit} (respectively, {γ′ γij (respectively, each γ′ us extend these sets of disjoint s.c.c.'s to sets of s.c.c.'s {γ0, . . . , γs} and {γ′ moreover satisfy: i1, . . . , γ′ 0, . . . , γ′ • for the associated integral homology classes, it holds [γi] = [γ′ i] = ai, for i = 0, . . . , s; • it holds γi ∩g γj = γ′ i ∩g γ′ j = hai, aji, for i, j = 0, . . . , s, where "∩g" denotes geometric intersection; For a given vertex y of Y , let Sy and S ′ y be, respectively, the closures in Sg,n of the connected components of Sg,n rσ and Sg,n rσ′ whose homology groups map onto the vertex group Gy. There is then a homeomorphism φy : Sy → S ′ y such that it holds φy(γi ∩ Sy) = γ′ i ∩ S ′ y, for i = 0, . . . , s. Glueing along σ and σ′ these homeomorphisms, for all connected components of Sg,n rσ i, for i = 0, . . . , s. and Sg,nrσ′, we get a homeomorphism φ : Sg,n → Sg,n such that φ(γi) = γ′ This homeomorphism then acts trivially on H0 and is such that φ(σ) = σ′. The situation is a bit more complicated when we want to describe the finite simplicial sets C (m)(Sg,n)•, for m ≥ 2. In this case, the set of equivalence classes of n-marked graph decompositions of Hm of rank k + 1, in general, is not in bijective correspondence with the set of non-degenerate k-simplices of C (m)(Sg,n)•. The set C (m)(Sg,n)k, for k ≥ 0, is the quotient of the set C (0)(Sg,n)k ∼= G n(H)k, described in Theorem 6.4, by the action of the congruence subgroup Sp2g(Z, m) ∼= Γ(m)/Tg,n of the symplectic group Sp2g(Z) and there is a natural surjective map G n(H)• ։ G (Hm)•, which is invariant under the action of Sp2g(Z, m). Therefore, there is at least a natural surjective map of simplicial sets: C (m)(Sg,n)• ։ G n(Hm)•. Let then (G, Y, d) and (G′, Y ′, d′) be two n-marked graph decompositions of H0 which induce, reducing modulo m, equivalent n-marked graph decompositions of Hm, i.e. such that there is an isomorphism of graphs f : Y → Y ′ with the property that d(v) = d′(f (v)), for all v ∈ v(Y ), and it holds Gσ/m = G′ f (σ)/m, for all simplices σ of Y , where, for a subgroup K of H0, we denote by K/m its image via the natural epimorphism H0 ։ Hm. The corresponding equivalence classes [G, Y, d] and [G′, Y ′, d′] of n-marked graph de- compositions of H0 are then in the same Sp2g(Z, m)-orbit if and only if there is an F ∈ Sp2g(Z, m) such that F (Gσ) = G′ f (σ). This happens if and only if there are stan- dard symplectic basis {a1, . . . , a2g} and {a′ 2g} of H0 with the following properties: 1, . . . , a′ a) it holds ai ≡ a′ i mod m, for i = 1, . . . , 2g; 25 b) the set {a1, . . . , a2g} (resp. {a′ 1, . . . , a′ 2g}) contains generators for the edge groups in a maximal set of linearly independent edge groups of (G, Y ) (resp. of (G′, Y ′)); c) for every vertex v ∈ Y , there is a set of indices {i1, . . . ik} such that it holds, simultane- ik}, and f (v)/E(f (v)), ously, Gv∩{a1, . . . , a2g} = {ai1, . . . , aik} and G′ the projections of these intersections to the quotients Gv/E(v) and G′ see vi.) Definition 6.1, generate these groups. 2g} = {a′ i1, . . . , a′ f (v)∩{a′ 1, . . . , a′ By property a), the assignment ai Properties b) and c) make sure that F (Gσ) = G′ 7→ a′ i, for i = 1, . . . , 2g, defines an F ∈ Sp2g(Z, m). f (σ), for σ a vertex or an edge of Y . Let, as above, (G, Y, d) and (G′, Y ′, d′) be two n-marked graph decompositions of H0 which induce, reducing modulo m, equivalent n-marked graph decompositions of Hm, re- lated by the isomorphism of graphs f : Y → Y ′. Reduction modulo m then defines, for every simplex σ of Y , epimorphisms µσ : Gσ → Gσ/m and µ′ f (σ)/m of Z-modules, where Gσ/m = G′ f (σ)/m is a free primitive Z/m-submodule of Hm. f (σ) → G′ f (σ) : G′ Let us observe that, given a vertex v of Y , any basis for a maximal symplectic non- f (v)/m can be lifted at the same time, via f (v), to bases for maximal symplectic non-degenerate free Z- degenerate free Z/m-submodule of Gv/m = G′ the epimorphisms µv and µ′ submodules of Gv and G′ f (v), respectively. On the other hand, given an edge e of Y , for m > 3, it is not necessarily the case that there is a generator for the edge group Ge/m = G′ f (e)/m which can be lifted both to a generator of the edge group Ge, via the epimorphism µe, and to a generator of the edge group G′ f (e), via the epimorphism µ′ f (e). However, if the last-mentioned condition is fulfilled for all edges, then there are bases {a1, . . . , a2g} and {a′ 2g} of H0 which satisfy properties a), b) and c) with respect to the two given n-marked graph decompositions of H0. This is the motivation for the following definitions: 1, . . . , a′ Definition 6.5. A framed, n-marked graph decomposition (G, Y, d, {µe}) of Hm is the data of an n-marked graph decomposition (G, Y, d) of Hm and, for every unoriented edge e of Y , an epimorphism µe : Z → Ge (a frame for the edge group Ge). Two framed, n-marked graph decompositions (G, Y, d, {µe}) and (G′, Y ′, d′, {µ′ e}) of Hm are equivalent if there is an isomorphism f : Y → Y ′ of the underlying graphs such that it holds d′(f (v)) = d(v), for all v ∈ v(Y ), it holds Gσ = G′ f (σ), for all simplices σ of Y , and it holds µf (e)(1) = ±µe(1), for every e ∈ e(Y ). We then denote by [G, Y, d, {µe}] the equivalence class of (G, Y, d, {µe}). It is easy to check that, given a framed, n-marked graph decomposition (G, Y, d, {µe}) of Hm, there is an n-marked graph decomposition ( G, Y, d) of H0 such that its reduction modulo m is the n-marked graph decomposition (G, Y, d) and the corresponding natural epimorphisms of edge groups Ge ։ Ge define frames equivalent to the µe, for all e ∈ e(Y ). 26 6 THE NERVE OF THE BOUNDARY OF ABELIAN LEVEL STRUCTURES Definition 6.6. Let G n degenerate simplices is defined as follows. ∗ (Hm)• be the simplicial set whose semi-simplicial set of non- i.) The set of k-simplices is the set of equivalence classes of framed, n-marked graph decompositions of Hm of rank k + 1. ii.) The (k −1)-faces of a k-simplex [G, Y, d, {µe}] are the equivalence classes of n-marked graph decompositions (G′, Y ′, d′) ≤ (G, Y, d) of rank k, endowed with the frames {µe}e∈e(Y ′). Face maps are then defined in the same way as in Definition 6.3. From the above discussion, it follows: Theorem 6.7. For all m ≥ 2 and g ≥ 1, there is a natural isomorphism of simplicial sets Ψ(m) : C (m)(Sg,n)• → G n ∗ (Hm)•. • Remark 6.8. Let (G, Y, d) be an n-marked graph decomposition of Hm. For m = 2, 3, there is clearly, for each unoriented edge e of Y , only one equivalence class of frames Z ։ Ge. Therefore, for m = 2, 3, it holds G n ∗ (Hm)• ≡ G n(Hm)•. The description given in Theorem 6.4 and Theorem 6.7 of the simplicial sets C (m)(Sg,n)•, for g ≥ 1 and m 6= 1, can be made more precise, for g ≥ 2 and n = 0, restricting either to the complex of non-separating curves C0(Sg) or to the complex of separating curves Csep(Sg). The first is the sub-complex of C(Sg) which consists of simplices σ such that Sg r σ is connected. The second is the full sub-complex of C(Sg) generated by 0-simplices γ such that Sg r γ is disconnected. In Proposition 6.7 of [23], the quotient C (m) (Sg)• := C0(Sg)•/Γ(m) is described, for m ≥ 2, as the complex of lax isotropic bases (cf. Definition 6.6 [23]). In our terminology, the simplices of G∗(Hm) contained in C (m) (Sg)• are determined by their framed edge groups. In particular, C (m) (Sg)• is the simplicial set associated to a genuine simplicial complex, equivalently, each simplex is determined by its vertices. 0 0 0 In [13], the quotient C (0) sep(Sg)• := Csep(Sg)•/Tg is described by a complex of tree de- compositions of H0. The simplicial set G n(H0)• is a natural generalization of this notion, except that the complex of tree decompositions is directly realized as the simplicial complex associated to the poset of tree decompositions with the partial order ≤ defined above. In [23] and [13], the above descriptions of the complexes C (m) sep(Sg) are introduced in order to prove that, in analogy with the curve complexes C0(Sg) and Csep(Sg), they are highly connected. More precisely, they prove that C (m) (Sg) is (g − 2)-connected, 0 for m ≥ 2, and that C (0) sep(Sg) is (g − 3)-connected, respectively. (Sg) and C (0) 0 Thus, it is a natural question whether similar results hold for the simplicial sets C (m)(Sg,n)•, for m 6= 1. From Proposition 3.3 [9] and Proposition 6.4 [23], it follows: Proposition 6.9. For g ≥ 2 and m 6= 1, the simplicial set C (m)(Sg,n)• is simply connected and (g − 2)-connected. 27 Proof. By Proposition 3.3 in [9], the abelian level Γ(m) of Γg,n, for g ≥ 2 and m 6= 1, is generated by Dehn twists along separating maps, m-th powers of Dehn twists and cut pair maps. All these elements stabilize some simplex of the curve complex C(Sg,n) and then some point of its geometric realization C(Sg,n). From Theorem 3 in [3], it follows that the quotient space C(Sg,n)/Γ(m) is simply connected for g ≥ 2 and m 6= 1. As A. Putman pointed out to me, from Proposition 6.4 [23] and the fact that, for g ≥ 2, the (g − 1)-skeleton of the geometric realization C (m)(Sg,n)• can be deformed, inside this complex, to the sub-complex C (m) (Sg), it follows that the simplicial set C (m)(Sg,n)• is also (g − 2)-connected. 0 7 The congruence subgroup problem In this section, we are going to connect all the topics of the previous sections with the congruence subgroup problem for the modular Teichmuller group Γg,n. This is the problem of determining whether the tower of geometric levels {ΓK}K EΠg,n is cofinal inside the tower of all finite index subgroups of Γg,n. This is known to be true in genus ≤ 2 (cf. [4] and [8]) but it is still an open problem in general. The solution proposed in [6] is flawed by a mistake in the proof of Theorem 5.4. However, the strategy elaborated there is still viable and will be explained shortly. In [6], to every profinite completion Γ′ mild technical conditions (cf. Definition 5.2 [6]), is associated the Γ′ of the curve complex C(Sg,n). g,n of the Teichmuller group Γg,n satisfying some g,n-completion C ′(Sg,n)• Let {Γλ}λ∈Λ′ be the set of levels of Γg,n obtained as the inverse images of open normal g,n, and contained in an abelian g,n-completion of the curve complex C(Sg,n) is defined subgroups of Γ′ level of order at least 3. Then, the Γ′ to be the inverse limit of finite simplicial sets (cf. Remark 4.4): g,n, via the natural homomorphism Γg,n → Γ′ C ′(Sg,n)• := lim ←− λ∈Λ′ C λ(Sg,n)•. Thus, it is an object in the category of simplicial profinite sets. In [24] and [25], Gereon Quick developed a homotopy theory for simplicial profinite sets, which turns out to be particularly useful here. For a simplicial profinite set X•, let us then denote by π1(X•) its fundamental group as defined by Quick. Quick's profinite homotopy theory would not be so useful for us, there were not a way g,n-completion C ′(Sg,n)• with those of its finite quotients to relate homotopy groups of the Γ′ C λ(Sg,n)•. However, this issue can be solved positively thanks to the results of [25]: Proposition 7.1. Let {X λ let X• := lim ←− λ∈Λ X λ • }λ∈Λ be a cofiltering inverse system of simplicial finite sets and • . For all q ≥ 0, there is then a natural isomorphism: πq(X•) ∼= lim ←− λ∈Λ πq(X λ • ). 28 7 THE CONGRUENCE SUBGROUP PROBLEM In particular, the fundamental group of X• is the inverse limit of the profinite completions of the fundamental groups of the geometric realizations {X λ • }λ∈Λ. Proof. In §3.5 of [25], an explicit fibrant replacement functor is defined for the model S , constructed in §2 [24]. For the structure of the category of simplicial profinite sets given simplicial profinite set X•, it is described as follows. Let GX λ • , for λ ∈ Λ, be the free simplicial profinite loop groupoid associated to the simplicial finite set X λ • , for λ ∈ Λ, is fibrant in S and is equipped with a natural map ηλ : X λ • , which is a trivial cofibration in S and then makes of ¯W GX λ • , for λ ∈ Λ, an explicit functorial fibrant replacement of the simplicial finite set X•. Again by Theorem 3.11 [25], the inverse limit • . By Theorem 3.11 [25], the profinite classifying space ¯W GX λ • → ¯W GX λ ¯W GX• := lim ←− λ∈Λ ¯W GX λ • is equipped with a canonical map η : X• → ¯W GX• which is a trivial cofibration in S and makes of ¯W GX• an explicit functorial fibrant replacement of the simplicial profinite set X•. According to Lemma 2.18 [24], for all q ≥ 0, there is then a series of natural isomorphisms: πq(X•) ∼= πq( ¯W GX•) ∼= lim ←− λ∈Λ πq( ¯W GX λ • ) ∼= lim ←− λ∈Λ πq(X λ • ). By Proposition 2.1 [24], the fundamental group π1(X λ • ) is just the profinite completion of the topological fundamental group of the geometric realization of the simplicial finite set X λ • and this implies the last claim in the proposition. In particular, for the Γ′ g,n-completion C ′(Sg,n)• of the curve complex C(Sg,n), there is a natural isomorphism, for all q ≥ 0: πq(C ′(Sg,n)•) ∼= lim ←− λ∈Λ′ πq(C λ(Sg,n)•). Let Γg,n be the profinite completion of the Teichmuller group and let us define the profinite curve complex C(Sg,n)• to be the Γg,n-completion of the complex of curves. The following result, with a different formalism, was already proved in [6]: Theorem 7.2. For 2g − 2 + n > 0 and 3g − 3 + n > 2, the simplicial profinite set C(Sg,n)• is simply connected. Proof. Let {Γλ}λ∈Λ be the set of all levels Γλ of Γg,n which contain an abelian level of order at least 3. By definition, the simplicial profinite set C(Sg,n)• is the inverse limit, over Λ, of the simplicial finite sets C λ(Sg,n)•. By what observed above, there is a natural isomorphism: π1( C(Sg,n)•) ∼= lim π1(C λ(Sg,n)•). ←− λ∈Λ 29 For 3g − 3 + n > 2, the simplicial set C λ(Sg,n)• is the quotient of a simply connected simplicial set by the simplicial action of Γλ. By Theorem 3 in Armstrong [3], there is then a natural epimorphism Γλ ։ π1(C λ(Sg,n)•). Therefore, taking profinite completions, we get a natural epimorphism Γλ ։ π1(C λ(Sg,n)•) and, passing to inverse limits, an epimorphism: Γλ ։ π1( C(Sg,n)•). lim ←− λ∈Λ But now lim ←− λ∈Λ Γλ = ∩λ∈Λ Γλ = {1} and the theorem follows. Let us now denote by Γg,n the profinite completion of the Teichmuller group Γg,n induced by the tower of geometric levels {ΓK}K EΠg,n and by C(Sg,n)• the Γg,n-completion of the complex of curves. They are called, respectively, the procongruence Teichmuller group and the procongruence curve complex. An immediate corollary of Theorem 7.2 is then: Corollary 7.3. The congruence subgroup property for the Teichmuller group Γg,n holds for all pairs (g, n) such that g ≤ k and 2g − 2 + n > 0, if and only if it holds π1( C(Sg)•) = {1} for all 2 ≤ g ≤ k. Proof. One implication is immediate from Theorem 7.2. To prove the other, let us pro- ceed by induction on k. The congruence subgroup property holds in genus ≤ 2. So, the statement of the corollary holds at least for k = 1, 2. For k ≥ 3, in order to complete the induction, we have to show that the assumptions, that the congruence subgroup property holds for all pairs (g′, n) such that g′ < g and that π1( C(Sg)•) = {1}, imply the congruence subgroup property for all pairs (g, n). Let σ be a simplex of the curve complex C(Sg) and let us denote with the same letter its image in the simplicial profinite sets C(Sg)• and C(Sg)•. Each Γg-orbit of simplices in C(Sg)• and C(Sg)• contains a "discrete" simplex σ. By Proposition 6.5 and Proposition 6.6 in [6], the assumption that the congruence subgroup property holds for all pairs (g′, n) such that g′ < g, implies that the stabilizer Γσ for the action of Γg on C(Sg)• is the profinite completion of the discrete stabilizer Γσ. Therefore, the natural epimorphism Γσ ։ Γσ is actually an isomorphism. This is then true for any simplex σ ∈ C(Sg)• and its image in C(Sg)•. Thus, if Ψg : Γg → Γg denotes the natural epimorphism, the profinite group ker Ψg acts freely on C(Sg)• with quotient naturally isomorphic to C(Sg)•. From Corollary 2.3 in [24], ∼= π1( C(Sg)•) = {1}, i.e. the congruence subgroup property holds it follows that ker Ψg for the Teichmuller group Γg. From Theorem 2 in [4], it then follows that the congruence subgroup property holds for all pairs (g, n), completing the induction step. Remark 7.4. With the above notations, Kent (cf. Theorem 12 [14]) has recently proved that it holds π1( C(Sg,n)•) = {1} for g ≥ 2 and n ≥ 2. However, the induction argument of Corollary 7.3 applies only for n ≤ 1. 30 7 THE CONGRUENCE SUBGROUP PROBLEM The main difficulty in trying to prove the simple connectivity of the procongruence curve complexes C(Sg,n)•, for g > 2, is that, for a given geometric level ΓK of Γg,n, there is no combinatorial description of the finite simplicial set C(Sg,n)/ΓK. So, it is natural to try to remedy to this situation by replacing the tower of geometric levels {ΓK}K EΠg,n of the Techmuller group Γg,n with the equivalent tower {ΓK,(m)}K EΠg,n of Looijenga levels, for some m ≥ 2. This was actually one of the main motivations to introduce these levels in the first place. In Section 3, we saw that, for m ≥ 2, the natural map T GK → M(SK)(m) factors → M(SK)(m) which, according to Theorem 3.5, is an through a morphism ιK,(m) : M embedding whenever K satisfies the hypotheses of ii.) Remarks 3.3. K,(m) The morphism ιK,(m) then induces a map from the set of simplices of the simplicial set C K,(m)(Sg,n)•, which describes the nerve of the D -- M boundary of M , to the set of simplices of the simplicial set C (m)(SK)•, which describes the nerve of the D -- M boundary of M(SK)(m). So, for every k ≥ 0, to the morphism ιK,(m), is associated a map: K,(m) bk : C K,(m)(Sg,n)k → a h≥0 C (m)(SK)h. Taking into account (cf. Theorem 6.7) the natural isomorphism of C (m)(SK)• with (H1(SK, Z/m))• of framed, nK-marked graph decompositions, the semi-simplicial set G nK where nK is the number of punctures on SK, the map bk is described as follows. ∗ For a given k-simplex σ ∈ C K,(m)(Sg,n)•, let σ ∈ C(Sg,n) be a lift to the curve complex. Then, to the simplex σ, we associate the simplex bk(σ) of G nK (H1(SK, Z/m))• defined by the equivalence class of the framed, nK-marked graph decomposition of the homology group H1(SK, Z/m) determined by the partition SK r p−1 K (σ) of the Riemann surface SK. It is not clear whether the map bk, even in case the morphism ιK,(m) is an embedding, is injective and if its image can be described in a simple way. In fact, it is injective only if intersects each GK-invariant stratum of the abelian level M(SK)(m) the image of M in a single irreducible component and this is not obvious. K,(m) ∗ The equivalence class [D, Y, d, {µe}], of a framed, nK-marked graph decomposition of H1(SK, Z/m) in the image of the map bk, for some k ≥ 0, admits a natural action of the group GK (in the sense of Definition 7.5). Conversely, the equivalence class [D, Y, d, {µe}] of a framed, nK-marked graph decom- position of H1(SK, Z/m), which admits GK as a group of automorphisms, via the natural symplectic action of GK on H1(SK, Z/m), parameterizes a stratum of M(SK)(m) which is GK-invariant and thus intersects the fixed point locus in M(SK)(m) for the action of GK. Therefore, the union of the images of the maps bk, for k ≥ 0, includes all framed, nK-marked graph decomposition endowed with a natural action of GK, if and only if, each GK-invariant stratum of M(SK)(m) intersects all connected components of the fixed point locus of GK in M(SK)(m), which is extremely unclear. The above geometric picture suggests at least that the theory of graph decompositions can be used to "approximate" the nerve of Looijenga level structures in a sense to be made clear by the following definition and Corollary 7.10. 31 Definition 7.5. For 2g − 2 + n ≥ 0, let K be a finite index characteristic subgroup of Πg,n and let pK : SK → Sg,n be the associated covering with covering transformation group GK. Let nK be the number of punctures on the Riemann surface SK. For an integer m ≥ 2, let then G nK (H1(SK, Z/m))• be the simplicial set whose semi-simplicial set of non-degenerate GK simplices is defined as follows. i.) The set of k-simplices is the set of equivalence classes [D, Y, d, {µe}] of framed, nK- marked graph decompositions of H1(SK, Z/m) endowed with a natural action of GK on the graph Y such that e(Y )/GK = k + 1. By natural action, it is meant that, for all v ∈ v(Y ) and f ∈ GK, it holds Df (v) = f (Dv) and d(f (v) = d(v), and that, for all e ∈ e(Y ) and f ∈ GK, it holds µf (e)(1) = ±f (µe(1)). ii.) The faces of a k-simplex [D, Y, d, {µe}] are the (k − 1)-simplices [D′, Y ′, d′, {µ′ e}] such e(1) = ±µe(1), for all e ∈ e(Y ′). A that (D′, Y ′, d′) ≤ (D, Y, d) and it holds µ′ face map ∂o : [D, Y, d, {µe}] → [D′, Y ′, d′, {µ′ e}] is then assigned to every GK-orbit o of unoriented edges of Y such that, for some representative (D′, Y ′, d′, {µ′ e}) of the equivalence class, its edge groups are the edge groups associated to the edges of Y ro. Remarks 7.6. For 2g − 2 + n ≥ 0 and m ≥ 2, let K be a finite index characteristic subgroup of Πg,n which satisfies the hypotheses of Theorem 3.9. Then, it holds: i.) For a simplex σ ∈ C(Sg,n)•, let (D, Y, d, {µe}) be a framed, nK-marked graph decom- position of H1(SK, Z/m) induced by the partition SK rp−1 K (σ). Then, the vertex and edge groups of (D, Y, d, {µe}) are all distinct. Therefore, there is only one possible action of the group GK on the graph Y compatible with its natural action on the homology group H1(SK, Z/m). Moreover, there are no non-trivial self-equivalences of (D, Y, d, {µe}), so that its equivalence class is determined modulo unique isomor- phisms of the underlying graphs of its representatives. ii.) A framed nK-marked graph decomposition (D, Y, d, {µe}) of H1(SK, Z/m), endowed with a natural action of GK, determines a GK-invariant stratum of M(SK)(m) which then parameterizes some curve endowed with a GK-action. This implies that there is a conjugate G′ K of GK inside Γ(SK) such that the nK-marked graph decomposition (D, Y, d, {µe}) is induced by a partition SK r q−1(σ), where q : SK → SK/G′ K is the quotient map, for some σ ∈ C(SK/G′ K ≡ GK mod Γ(m). In particular, also the nK-marked graph decomposition (D, Y, d, {µe}) has the properties stated in i.). K)•, and it holds G′ iii.) It follows from i.) and ii.) that all non-degenarate simplices of G nK GK are determined by their faces. In other words, G nK GK set associated to a combinatorial simplicial complex. (H1(SK, Z/m))• (H1(SK, Z/m))• is the simplicial iv.) For m ≥ 2, the covering pK : SK → Sg,n induces a map of finite simplicial sets gK,(m) : C K,(m)(Sg,n)• → G nK GK (H1(SK, Z/m))• 32 7 THE CONGRUENCE SUBGROUP PROBLEM and the group NSp(H1(SK ,Z/m))(GK) acts naturally on G nK (H1(SK, Z/m))•, with its subgroup GK acting trivially. Therefore, there is a natural action of the Teichmuller group Γg,[n] on this simplicial set, via the representation ρK,(m) of Section 2, and the GK map gK,(m) is Γg,[n](cid:14)ΓK,(m) -equivariant. By Theorem 2.2, the Looijenga level ΓK,(m) is contained in the geometric level ΓK. Therefore, there is also a natural surjective map πK : C K,(m)(Sg,n)• ։ C K(Sg,n)•. The maps gK,(m) and πK are related in the following way: Theorem 7.7. For 2g − 2 + n ≥ 0, let K be a finite index characteristic subgroup of Πg,n which satisfies the hypotheses of Theorem 3.9. Then, for m ≥ 2, the natural sur- jective map πK : C K,(m)(Sg,n)• ։ C K(Sg,n)• factors through the natural surjective map gK,(m) : C K,(m)(Sg,n)• ։ Im gK,(m) and a surjective map Im gK,(m) ։ C K(Sg,n)•. Proof. The map gK,(m) is induced by the map C(Sg,n)• → G nK (H1(SK, Z/m))• which GK associates to a simplex σ ∈ C(Sg,n)• the equivalence class of the framed, nK-marked graph decomposition of H1(SK, Z/m) induced by the decomposition SK r p−1 K (σ) of SK, where pK : SK → Sg,n is the Galois covering associated to the subgroup K of Πg,n. By Remarks 7.6, the hypotheses of Theorem 3.9 on K imply that, if two framed, nK- marked graph decompositions of H1(SK, Z/m), endowed with a natural GK-action, are equivalent, there is a unique isomorphism of the underlying graphs which induces the equivalence. Therefore, we can identify them by means of this isomorphism. Let then be given two simplices σ, σ′ ∈ C(Sg,n)• inducing the same framed, nK-marked graph decomposition (D, Y, d, {µe}) of H1(SK, Z/m). In order to prove the theorem, we have to show that there is an f ∈ ZΓ(SK )(GK) such that f (p−1 K (σ′). By the short exact sequence (2) in Section 2, the image f of f , by the natural epimorphism NΓ(SK )(GK) ։ Γg,[n], is then contained in the level ΓK and is such that f (σ) = σ′. The nK-marked graph Y determines the topological types of both SK r p−1 K (σ) and K (σ′) and then of Sg,n r σ and Sg,n r σ′ as well. So, at least, we know, that there K (σ)) = p−1 K (σ) = ∪k i. Let us assume that γi and γ′ i define the same cycle in H1(SK, Z/m), for i = 0, . . . , k. The hypothesis on the subgroup K implies that γi and γj define distinct homology classes whenever i 6= j and then that the action of the group GK on H1(SK, Z/m) determines its action on the closed submanifolds p−1 K (σ) and p−1 K (σ) → p−1 K (σ′) which sends γi to γ′ i, for i = 0, . . . , k. We claim that δ can be extended to a GK-equivariant homeomorphism χ : SK → SK. K (σ′). So, there is a unique GK-equivariant homeomorphism δ : p−1 By Theorem 6.7, the simplices σ and σ′ determine the same stratum of the abelian level structure M(SK)(m). Therefore, there is an element φ in the abelian level Γ(m) of Γ(SK) such that it holds φ(p−1 K (σ′) and then, by the above assumptions, such that φ(γi) = γ′ i, for i = 0, . . . , k. K (σ)) = p−1 Let G′ K be the isomorphism defined by the assignment α 7→ φαφ−1. The homeomorphism φ then satisfies φ(α·x) = inn φ(α)·φ(x), K := φGKφ−1 < Γ(SK) and let inn φ : GK → G′ SK r p−1 is an element f ∈ NΓ(SK )(GK) such that f (p−1 i=0γ′ i=0γi and p−1 K (σ′) = ∪k Let p−1 K (σ)) = p−1 K (σ′). 33 for all x ∈ SK and α ∈ GK. The group G′ circles p−1 Let S ′ K (σ′). More precisely, it holds α · x = inn φ(α) · x, for all x ∈ p−1 K := SK rp−1 K, like the group GK, preserves the union of K (σ′) and α ∈ GK. K/GK K are equivalent Galois ´etale coverings of homeomorphic surfaces. Thus, K, which K are conjugated by a self-homeomorphism of S ′ K (σ′). By the definition of φ and G′ K, the natural maps S ′ K → S ′ K/G′ K → S ′ and S ′ the actions of GK and G′ lifts a homeomorphism S ′ K on S ′ K/GK → S ′ K/G′ K. The GK-covering SK → SK/GK and the G′ K are determined by K (σ′), modulo the actions by conjugation, respectively, K) on SK. Therefore, the identity map on the K (σ′) extends to a homeomorphism ψ : SK → SK which lifts a homeomorphism their restrictions on the subspace p−1 of the groups NΓ(SK )(GK) and NΓ(SK )(G′ subspace p−1 SK/GK → SK/G′ K and satisfies ψ(α · x) = inn φ(α) · ψ(x), for all x ∈ SK and α ∈ GK. K-covering SK → SK/G′ The composition χ := ψ−1 ◦ φ : SK → SK is then a GK-equivariant homeomorphism with the required property that χ(p−1 K (σ)) = p−1 K (σ′). Corollary 7.8. For 2g −2+n ≥ 0, let {K} be a cofinal system of finite index characteristic subgroups of Πg,n satisfying the hypotheses of Theorem 3.9 and let {pK : SK → Sg,n} be the associated set of coverings. For any fixed integer m ≥ 2, there is then a continuous injective Γg,n-equivariant map of simplicial profinite sets: Y K gK,(m) : C(Sg,n)• ֒→ Y K G nK GK (H1(SK, Z/m))•. Proof. The map gK,(m) : C(Sg,n)• → G nK (H1(SK, Z/m))• is defined by the composition GK of the natural projection C(Sg,n)• ։ C K,(m)(Sg,n)• with the map gK,(m) associated to the covering pK : SK → Sg,n. The corollary then follows from Corollary 2.3 and Theorem 7.7. Remark 7.9. By Lemma 3.10, a cofinal system of finite index characteristic subgroups of Πg,n satisfying the hypotheses of Theorem 3.9 can be constructed from any given cofinal system of finite index characteristic subgroups of Πg,n. It is not clear whether the set {G nK GK (H1(SK, Z/m))•}, where {K} is a cofinal system of finite index characteristic subgroups of Πg,n and m ≥ 2 a fixed integer, forms an inverse system of finite simplicial sets. This is partially fixed by the following formal consequence of Corollary 7.8: Corollary 7.10. For 2g−2+n ≥ 0, let {K} be a cofinal system of finite index characteristic subgroups of Πg,n and let {pK : SK → Sg,n} be the associated set of coverings. Then, for any fixed integer m ≥ 2, there is a cofinal sub-system {Kλ}λ∈Λ of {K} such that the set {Im gKλ,(m)}λ∈Λ can be organized in an inverse system of finite simplicial sets and there is a Γg,n-equivariant isomorphism of simplicial profinite sets: C(Sg,n)• ∼= lim ←− λ∈Λ Im gKλ,(m). 34 REFERENCES Remark 7.11. For K as in the hypotheses of Theorem 3.9, by i.) Remarks 7.6, the sim- plicial set Im gK,(m) can be realized as the simplicial set associated to an ordered simplicial complex. The same property then holds for the simplicial profinite set C(Sg,n)•, that is to say, there is an ordered simplicial complex C(Sg,n) whose sets of k-simplices are profinite, for all k ≥ 0, and whose associated simplicial set is C(Sg,n)•. By Corollary 7.10 and Corollary 7.3, the congruence subgroup property would follow from a positive answer to the following question in combinatorial topology: Question 7.12. For g ≥ 2 and some integer m ≥ 2, is there a cofinal system {K} of finite index characteristic subgroups of Πg such that the finite simplicial sets Im gK,(m) are simply connected? Of course, it would be much easier to approach Question 7.12 if a more explicit descrip- tion of the images of the maps gK,(m) were available, for instance, if the maps gK,(m) were known to be surjective for some cofinal system {K} and some integer m ≥ 2. If the answer to Question 7.12 is positive, by Corollary 7.10, there is a cofinal system {Kλ}λ∈Λ of finite index characteristic subgroups of Πg, refining the cofinal system {K}, such that {Im gKλ,(m)} forms an inverse system of simply connected finite simplicial sets and, then, it holds: π1( C(Sg)•) ∼= lim π1(Im gKλ,(m)) = {1}. ←− λ∈Λ For the moment, the only scarce evidence for a positive answer to Question 7.12 is provided by Proposition 6.9. Compared to the approach to the subgroup congruence problem of [6], this has at least the advantage that it can be dealt with standard techniques of combinatorial topology in the spirit of [13] and [22]. Acknowledgements This work was begun during my stay at the Department of Mathematics of the University of Costa Rica in San Jos´e and completed during my stay at the Department of Mathematics of the University of los Andes in Bogot´a. I thank both institutions for their support, financial and otherwise. I also thank A. Putman for suggesting the second part of Proposition 6.9. References [1] W. Abikoff. The Real Analytic Theory of Teichmuller space. Lecture Notes in Math. 820, Springer-Verlag. [2] [3] E. Arbarello, M. Cornalba, P. A. Griffiths. Geometry of algebraic curves, II. Grundlehren der mathematischen Wissenschaften 268. Springer-Verlag (2011). M. A. Armstrong. On the fundamental group of an orbit space. Proc. Cambridge Philos. Soc. 61 (1965), 639 -- 646. REFERENCES 35 [4] [5] [6] [7] [8] [9] M. Asada. The faithfulness of the monodromy representations associated with cer- tain families of algebraic curves. Journal of Pure and Applied Algebra 159, (2001), 123 -- 147. J. Bertin, M. Romagny. Champs de Hurwitz. M´emoires de la Soci´et´e Math´ematique de France 125/126 (2011). M. Boggi. Profinite Teichmuller theory. Math. Nach. 279, n. 9-10 (2006), 953 -- 987. M. Boggi. Fundamental groups of moduli stacks of stable curves of compact type. Geometry & Topology 13 (2009), 247 -- 276. M. Boggi. The congruence subgroup property for the hyperelliptic Teichmuller mod- ular group. arXiv:0803.3841v4. M. Boggi, M. Pikaart. Galois covers of moduli of curves. Comp. Math. 120 (2000), 171 -- 191. [10] J. L. Brylinski. Propri´et´es de ramification `a l'infini du groupe modulaire de Te- ichmuller. Ann. Sc. ´Ec. Norm. Sup. 12, 4e s´erie, (1979), 295 -- 333. [11] G. Gonz´alez-D´ıez, W.J. Harvey. Moduli of Riemann surfaces with symmetry. Ap- peared in Discrete Groups and Geometry. Ed. by W.J. Harvey and C. Maclachlan. LMSLNS 173 Cambridge University press (1992), 75 -- 93. [12] N. V. Ivanov. Subgroups of Teichmuller modular groups. Translations of Mathe- matical Monographs 115. AMS (1992). [13] W. van der Kallen, E. Looijenga. Spherical complexes attached to symplectic lat- tices. arXiv:1001.0883v1 (2010). [14] R. P. Kent IV. Congruence kernels around affine curves. arXiv:1109.1267v1 (2011). [15] [16] [17] [18] [19] E. Looijenga. Smooth Deligne-Mumford compactifications by means of Prym level structures. J. Algebraic Geometry 3 (1994), 283 -- 293. E. Looijenga. Prym Representations of Mapping Class Groups. Geometriae Dedi- cata 64 (1997), 69 -- 83. S. Mochizuki. The Local Pro-p Anabelian Geometry of Curves. Invent. math. 138 (1999), 319 -- 423. S. Mochizuki. Extending families of curves over log regular schemes. J. reine angew. Math. 511 (1999), 43 -- 71. B. Noohi. Fundamental groups of algebraic stacks. J. Inst. Math. Jussieu 3, n. 1 (2004), 69 -- 103. 36 REFERENCES [20] B. Noohi. Foundations of topological stacks I. arXiv:math/0503247v1 (2005). [21] M. Pikaart. Moduli spaces of curves: stable cohomology and Galois covers. Ph.D. thesis, Utrecht University (1997). [22] [23] A. Putman. Cutting and pasting in the Torelli group. Geometry & Topology 11 (2007), 829 -- 865. A. Putman. The second rational homology group of the moduli space of curves with level structure. Adv. Math. 229 (2012), 1205 -- 1234. [24] G. Quick. Profinite homotopy theory. Documenta Mathematica 13 (2008), 585 -- 612. [25] G. Quick. Some remarks on profinite completion of spaces. In Galois-Teichmller theory and Arithmetic Geometry (H. Nakamura, F. Pop, L. Schneps, and A. Tama- gawa, eds.), Advanced Studies in Pure Mathematics, vol. 63, Mathematical Society of Japan (2012), 413 -- 448. [26] J.P. Serre. Arbres, Amalgames, SL2. Ast´erisque no. 46, Soc. Math. France (1977). Address: Departamento de Matem´aticas, Universidad de los Andes, Carrera 1a No 18A-10, Bogot´a, Colombia. E -- mail: [email protected] 

1205.5463

2

1205

On stringy cohomology spaces

We modify the definition of the families of $A$ and $B$ stringy cohomology spaces associated to a pair of dual reflexive Gorenstein cones. The new spaces have the same dimension as the ones defined in the joint paper with Mavlyutov \cite{BM}, but they admit natural flat connections with respect to the appropriate parameters. This solves a longstanding question of relating GKZ hypergeometric system to stringy cohomology. We construct products on these spaces by vertex algebra techniques. In the process, we fix a minor gap in \cite{BM} and prove a statement on intersection cohomology of dual cones that may be of independent interest.

math.AG

math

ON STRINGY COHOMOLOGY SPACES LEV A. BORISOV Abstract. We modify the definition of the families of A and B stringy cohomology spaces associated to a pair of dual reflexive Gorenstein cones. The new spaces have the same dimension as the ones defined in the joint paper with Mavlyutov [BoM], but they admit natural flat connections with respect to the appropriate parameters. This solves a longstanding question of relating GKZ hypergeometric system to stringy cohomology. We construct prod- ucts on these spaces by vertex algebra techniques. In the process, we fix a minor gap in [BoM] and prove a statement on intersection cohomology of dual cones that may be of independent interest. 1. Introduction In the early history of mirror symmetry, the duality of Hodge num- bers played an important role. Specifically, for three-dimensional mir- ror Calabi-Yau varieties X and X ∨ one has h1,1(X) = h1,2(X ∨) and h1,2(X) = h1,1(X ∨). In general, for n-dimensional mirror Calabi-Yau varieties one may expect (1.1) hp,q(X) = hn−p,q(X ∨), however there are complications. The initial obstacle is that higher dimensional singular Calabi-Yau varieties often do not admit crepant resolutions, and if one tries to verify (1.1) for singular mirrors, it sim- ply fails. In this sense, usual Hodge numbers are ill-suited for higher dimensional mirror symmetry. An elegant solution to this difficulty has been obtained in [BaD, Ba3]. Namely, for any X with log-terminal singularities one introduces the stringy E-function E(u, v) as a certain weighted sum over strata of a log resolution of X and shows that it is independent of the resolution. In the case when X is toroidal, the stringy E-function turns out to be a polynomial, which allows one to define stringy Hodge numbers hp,q st (X) as coefficients of E. These numbers satisfy several nice properties. If X admits a crepant resolution Y , then hp,q st (X) = hp,q(Y ). If X admits a crepant resolution by an orbifold (smooth DM stack) then hp,q(X) equals the dimension of the corresponding orbifold cohomology of Y . Perhaps the best justification for working with hp,q st comes from the 1 2 LEV A. BORISOV large class of mirror symmetry examples provided by hypersurfaces and complete intersections in Gorenstein toric Fano varieties. For these examples, the duality (1.2) st (X) = hn−p,q hp,q st (X ∨) has been proved in [BaBo]. Once the appropriate stringy Hodge numbers are defined, the next natural question to ask is what are the stringy cohomology vector spaces whose dimensions are given by hp,q st . This is already a mean- ingful question in dimension three. While one can consider cohomol- ogy of a crepant resolution to be a sensible answer, the non-uniqueness of such resolution presents a problem. One expects to have not one stringy cohomology space but a family of such spaces that interpolates between cohomology of different crepant (orbifold) resolutions, if any exist. While in general the problem of constructing stringy cohomology is wide open, reasonable definitions can be made in the setting of toric mirror symmetry. Specifically, Definition 1.1 below was proposed in [BoM]. To explain the ingredients of this definition we first recall the combinatorial data of toric mirror symmetry. Let M and N be dual free abelian groups (often referred to as lat- tices). Let K and K ∨ be dual maximum-dimensional rational polyhe- dral cones in M and N. These dual cones are called reflexive Gorenstein if the lattice generators of the one-dimensional faces of both K and K ∨ lie in a hyperplane at distance one from the origin in M and N re- spectively. This condition on the cones K and K ∨ is equivalent to the statement that both semigroup rings C[K] and C[K ∨] are Gorenstein. Dual reflexive Gorenstein cones K and K ∨ uniquely determine a pair of lattice points deg ∈ M and deg∨ ∈ N which define the aforemen- tioned hyperplanes by {• · deg∨ = 1} and {deg ·• = 1}. We denote by ∆ (respectively ∆∨) the set of lattice points in M (respectively N) that lie in these hyperplanes. For generic choices of coefficient func- tions f : ∆ → C and g : ∆∨ → C, one can typically1 define mirror families of singular Calabi-Yau varieties (Xf ) and (X ∨ g ). Definition 1.1 of stringy cohomology vector spaces in the setting of dual Gorenstein cones has been first suggested by Mavlyutov. Their dimensions have been consequently calculated in [BoM] to coincide with the stringy Hodge numbers of Xf and X ∨ g whenever they exist. Here is the description of the construction. For a face θ of K we consider 1There are some obstructions to this in deg · deg∨ > 1 case, see [BaN]. However, stringy cohomology vector spaces can still be constructed in the absence of Calabi- Yau varieties. This suggests that the underlying N = (2, 2) superconformal field theories can also be constructed. ON STRINGY COHOMOLOGY SPACES 3 the ideal If,θ in the semigroup ring C[θ] generated by the elements Pm∈θ f (m)µ(m)[m] for all linear functions µ on the span of θ. Here the summation is taken over elements in ∆ that lie in θ. Then the space R1(f, θ) is defined as the image of C[θ◦]/If,θC[θ◦] → C[θ]/If,θC[θ] (see [BoM]) where θ◦ is the interior of θ and thus C[θ◦] is an ideal in C[θ]. To a face θ of K we associate a face θ∗ := Ann(θ) ∩ K ∨ of K ∨ (we use θ∗ to avoid the confusion with the dual cone θ∨) and define R1(g, θ∗) similarly. Definition 1.1. Stringy cohomology space of (Xf , X ∨ g ) is given by M {0}⊆θ⊆K R1(f, θ) ⊗ R1(g, θ∗) ⊗ Λdim θ∗ (Cθ∗). The space above possesses a natural double grading, see [BoM]2 such that the stringy Hodge numbers can be recovered as dimensions of double-graded components. While it is satisfying to have natural spaces with the appropriate double grading, Definition 1.1 raises two important questions. • One expects to have two different graded super-commutative associative products on the stringy cohomology vector spaces. How can one construct these products? • How to relate these spaces to the GKZ hypergeometric system of partial differential equations? The first issue has been partially resolved in [BoM] and [Bo2] by identifying the spaces of Definition 1.1 with the chiral rings of certain N = 2 vertex algebras under an additional technical assumption that f and g are strongly nondegenerate, which is a somewhat smaller open set than that of nondegenerate coefficient functions. This interpretation of the stringy cohomology vector space as a chiral ring hinges on the following construction of this space as cohomology of a natural complex of vector spaces. Consider the quotient C[(K ⊕ K ∨)0] of the semigroup ring C[K ⊕ K ∨] by the ideal spanned by [m, n], m · n > 0. Define V = C[(K ⊕ K ∨)0] ⊗ Λ∗NC and an endomorphism df,g : V → V by df,g = X f (m)[m] ⊗ (contr.m) + X m∈∆ n∈∆∨ g(n)[n] ⊗ (n∧). 2For exposition reasons, only the case of deg · deg∨ = 1 case was considered in [BoM] but the arguments apply to any pair of reflexive Gorenstein cones. 4 LEV A. BORISOV It is easy to see that d2 f,g = 0. Moreover, d increases by one the natural grading on V given by [m, n] ⊗ P → m · deg∨ + deg · n. The following result was first claimed in [BoM] and then used in [Bo2] to identify the space of Definition 1.1 with the chiral ring of a vertex algebra. Theorem 3.1. The cohomology of V with respect to df,g is naturally isomorphic to the direct sum over faces θ of K M {0}⊆θ⊆K R1(f, θ) ⊗ R1(g, θ∗) ⊗ Λdim θ∗ (Cθ∗). Unfortunately, the proof of Theorem 3.1 presented in [BoM] was in- correct, and the correction is given in this paper as a consequence of a result in intersection cohomology of polyhedral fans. The second issue with Definition 1.1, namely the relation to the GKZ hypergeometric system has eluded understanding for too long now. Finally, we are able to resolve it satisfactorily by modifying the complex of [BoM] slightly. Specifically, the main results of this paper are the following Definition 4.6, Theorem 4.7 and Corollary 5.10. Definition 4.6. Let K and K ∨ be dual reflexive Gorenstein cones. Define as before the space V = C[(K ⊕ K ∨)0] ⊗ Λ∗NC. Consider the differential bdf,g on it given by bdf,g([m1 ⊕ n1] ⊗ P ) = df,g([m1 ⊕ n1] ⊗ P ) + [m1 ⊕ n1] ⊗ (n1 ∧ P ) = X + X g(n)[m1 ⊕ (n + n1)] ⊗ (n ∧ P ) + [m1 ⊕ n1] ⊗ (n1 ∧ P ). f (m)[(m + m1) ⊕ n1] ⊗ (contr.m)(P ) m∈∆ n∈∆∨ Then we define stringy cohomology B-space HB(Xf , X ∨ g ) as the coho- mology of V with respect to bdf,g, with the grading given by [m ⊕ n] ⊗ P 7→ 2 m · deg∨ + deg(P ). We will also call this the A-space of the mirror pair (X ∨ HA(Xf , X ∨ C[(K ⊕ K ∨)0] ⊗ Λ∗MC by the differential that maps g , Xf ). Similarly, g , Xf )) is defined as the cohomology of g ) (equal to HB(X ∨ [m1 ⊕ n1] ⊗ P 7→ X m∈∆ f (m)[(m + m1) ⊕ n1] ⊗ (m ∧ P ) + X n∈∆∨ g(n)[m1 ⊕ (n + n1)] ⊗ (contr.n)(P ) + [m1 ⊕ n1] ⊗ (m1 ∧ P ). ON STRINGY COHOMOLOGY SPACES 5 Theorem 4.7. For nondegenerate f and g, the B-space HB(Xf , X ∨ g ) is naturally isomorphic to R1(f, θ) ⊗ \R1(g, θ∗) ⊗ Λdim θ∗ (Cθ∗). M {0}⊆θ⊆K The bundle of HB(Xf , X ∨ natural flat connection. g ) over the space of nondegenerate g has a Corollary 5.10. For strongly nondegenerate f and g the stringy co- homology spaces HB(Xf , X ∨ g , Xf ) and HA(Xf , X ∨ g ) = HA(X ∨ g ) = HB(X ∨ g , Xf ) are equipped with natural structures of graded associative super-com- mutative algebras. The flat connection in Theorem 4.7 is in fact explicitly seen to be re- lated to the GKZ hypergeometric system, or rather its better-behaved version considered in [BoH]. Corollary 5.10 is obtained by vertex alge- bra techniques. It requires a technical strong non-degeneracy condition developed in [Bo2]. We also remark that our notation has Xf and X ∨ g in it, even though the varieties Xf and X ∨ g may not exist or be uniquely defined. One might in fact use notation HA/B(K, K ∨, f, g) instead. However, we prefer the current notation for the following reason. If deg · deg∨ = 1, then we may interpret Xf and X ∨ g as singular Calabi-Yau hypersur- faces in Gorenstein Fano toric varieties Proj(C[K]) and Proj(C[K ∨]) respectively. In deg · deg∨ > 1 case we prefer to keep this notation to stress the relation to usual cohomology when it is applicable. The paper is organized as follows. In Section 2 we prove a result on intersection cohomology of polyhedral fans which is key to the ar- gument and may be of independent interest. In Section 3 we prove Theorem 3.1 as an easy consequence of the results of Section 2. In Section 4 we modify our complex so that the new version of stringy cohomology comes equipped with a flat connection. The discussion culminates in Theorem 4.7. The first four sections are not particularly technical. Section 5 is devoted to the construction of the product struc- tures on the stringy cohomology defined in Section 4. We rely heavily on the results of [Bo1] and [Bo2]. Finally, in Section 6 we outline some open problems related to the construction. Acknowledgements. Joint work with Paul Horja [BoH] was one of the key motivations behind this paper. The author thanks Nick Addington for useful comments. The author has been partially sup- ported by the NSF grants DMS-1003445 and DMS-1201466. 6 LEV A. BORISOV 2. Intersection cohomology of dual cones In this section we prove a technical result on intersection cohomology of dual polyhedral cones which is the cornerstone of the paper. We ex- pect the reader to be familiar with the approach of [BrL] and [BBFK] to the intersection cohomology of fans. In these papers the authors de- fine first a combinatorial analog of equivariant intersection cohomology of toric varieties as global sections of a minimum locally free (in the appropriate sense) and flabby sheaf on the topological space defined by the fan. Intersection cohomology is then defined as the quotient of the equivariant intersection cohomology by the action of linear functions. Consider dual strictly convex polyhedral cones C and C ∨ in dual vector spaces MR and NR of dimension r. We do not assume that C and C ∨ are rational, although this is the case for our primary applications. Consider the finite polyhedral fan Φ in MR ⊕ NR whose cones are direct sums of faces θ ⊕ σ with θ ⊆ C, σ ⊆ C ∨ and θ · σ = 0. Consider the minimum flabby graded locally free sheaf L = L{0}⊕{0} (in the sense of [BrL]) on Φ. Let W = Γ(Φ, L) be the space of global sections of L. Note that W is a module over Sym∗(MC ⊕ NC) which is annihilated by the quadratic equation that gives the bilinear pairing. Consider dual bases (mi) and (ni) of MR and NR and the corresponding linear functions yi and xi on NR and MR respectively. Then W is a mod- i=1 xiyi. We ule over the ring C[x1, . . . , xr, y1, . . . , yr] annihilated by Pr define the differential on W ⊗ Λ∗NC by xi ⊗ (contr.mi) + yi ⊗ (ni∧). rX i=1 d = rX i=1 Clearly, d2 = 0 because W is annihilated by Pr i=1 xiyi. In addition, d increases the degree by 1. It is also easy to see that d is in fact independent of the choice of dual bases in MR and NR. The main result of this section is the following. Theorem 2.1. Cohomology of W ⊗ Λ∗NC with respect to d is 0 for any C in dimension r > 0. Remark 2.2. For r = 0, the space W ⊗ Λ∗NC is one-dimensional with d = 0. finite projective dimension over the ring C[x, y]/hPr Remark 2.3. Theorem 2.1 is equivalent to the statement that W has i=1 xiyii. We thank Nick Addington for this observation. It would be interesting to produce a free resolution of W explicitly, but it is not directly related to the goals of this paper. ON STRINGY COHOMOLOGY SPACES 7 Proof. We can construct L = L{0}⊕{0} as a restriction to Φ of product of pullbacks of L{0} sheaves on K and K ∨. As a consequence, the space W is double graded by degree in x and degree in y. The differential d preserves the grading by (2.1) degx − degy + deg(Λ∗) which consequently passes to cohomology. We will show that for r > 0 this grading on the cohomology is at once < 1 2r to conclude that cohomology is 0. To accomplish this, we estimate the cohomology in two different ways by means of spectral sequences associated to two different filtrations of W . Consider the filtration 2r and > 1 W = F −1W ⊇ F 0W ⊇ F 1W ⊇ . . . ⊇ F rW = 0 on W defined by F kW = {w ∈ W, such that wθ⊕σ = 0 for all θ of dim θ ≤ k} and the corresponding filtration (F kW ⊗ Λ∗NC) on W ⊗ Λ∗NC. Clearly (F kW ⊗ Λ∗NC) is preserved by d and is compatible with the grad- ing by degx − degy + deg(Λ∗), in the sense that it induces filtration on each graded component. We will now consider the spectral sequence associated to this filtration. Note that F kW is the kernel of the restriction map from the space of global sections of L on Φ to the space of its sections on the open set Φk (in the fan topology) described by dim θ ≤ k. Since L is flabby, these restriction maps are surjective. Therefore, the quotient of F k−1W by F kW is the kernel of the restriction map Γ(Φk, L) → Γ(Φk−1, L). Note that L is built inductively by expanding the open sets. The fan Φk is obtained from the fan Φk−1 by attaching cells θ ⊕ θ∗ with dim θ = k to the boundary ∂θ ⊕ θ∗. We thus have F k−1W/F kW = M dim θ=k Γ(θ, ∂θ, L) ⊗ Γ(θ∗, L). where Γ(θ, ∂θ, L) is the kernel of the map Γ(θ, L) → Γ(∂θ, L). The associated graded complex is then given by M θ Γ(θ, ∂θ, L) ⊗ Γ(θ∗, L) ⊗ Λ∗NC. For each direct summand we decompose (2.2) Λ∗NC = Λ∗((Cθ)∨) ⊗ Λ∗(Cθ∗) and see that the θ component of the associated graded complex in question is isomorphic to the tensor product of two Koszul complexes. 8 LEV A. BORISOV Its cohomology is given by M θ IH(θ, ∂θ) ⊗ IH(θ∗) ⊗ Λdim θ∗ (Cθ∗). where the intersection cohomology spaces above are defined in [BrL] as quotients of Γ(θ, ∂θ, L) and Γ(θ∗, L) by the ideals of linear functions. We will now use the inequalities on the grading of IH(θ, ∂θ) and IH(θ) which follow from the strong Lefschetz theorem [Ka]. Specifically, the graded dimensions of these spaces can be given in terms of the G-polynomials of Stanley [St]. We see that degx ≥ 1 2k, degy ≤ 1 2 (r − k). So the degree in the sense of (2.1) is at least 2 dim θ = 1 2 dim θ∗ = 1 1 2 1 2 k − (r − k) + (r − k) = 1 2 r. Note that if r > 0 then at least one of θ, θ∗ is of positive dimension, so the corresponding inequality is strict. Thus, we get that the degree is > 1 2r. By considering the analogous filtration by dimension of σ, we see that there is a spectral sequence that converges to the cohomology of W ⊗ Λ∗NC and starts from M θ IH(θ) ⊗ IH(θ∗, ∂θ∗) ⊗ Λdim θ∗ (Cθ∗). The corresponding inequalities degx ≤ 1 2 dim θ∗ (as before, at least one of them is sharp) lead to the grading (2.1) being less than 1 (cid:3) 2 dim θ and degy ≥ 1 2r. This finishes the proof. One of the main results of [BrL] is that every locally free flabby sheaf on Φ is (non-canonically) isomorphic to a direct sum of copies of minimal locally flabby sheaves that originate at various cones of Φ. Theorem 2.1 is a statement about global sections of such minimal sheaf that originates at the zero cone. In the following proposition we extend this result to other minimal flabby locally free sheaves on Φ. Proposition 2.4. Let C and Φ be as in Theorem 2.1. Fix (θ0, σ0) with σ0 ⊆ θ∗ 0 and consider the minimal flabby locally free sheaf of modules L(θ0⊕σ0). If σ0 ( θ∗ 0 then the cohomology of d = rX i=1 xi ⊗ (contr.mi) + rX yi ⊗ (ni∧) on Γ(Φ, L(θ0⊕σ0))⊗Λ∗NC is zero. If σ0 = θ∗ dimensional and can be identified with Λdim θ∗ space. i=1 0 then the cohomology is one- 0) as a graded vector 0 (Cθ∗ ON STRINGY COHOMOLOGY SPACES 9 Proof. We repeat the same argument as in the proof of Theorem 2.1, with the same filtrations as before. We have two spectral sequences which converge to the cohomology and start from M θ,θ0⊆θ⊆σ∗ 0 IH(θ/θ0, ∂θ/θ0) ⊗ IH(θ∗/σ0) ⊗ Λdim θ∗ (Cθ∗) and M θ,θ0⊆θ⊆σ∗ IH(θ/θ0) ⊗ IH(θ∗/σ0, ∂θ∗/σ0) ⊗ Λdim θ∗ (Cθ∗) where taking the quotient means considering the image of the appro- priate cones modulo the span of a face. The grading is estimated to be at least 1 2 (dim θ∗−dim σ0)+dim θ∗ = (dim θ−dim θ0)− (dim θ0+dim σ0) 1 2 1 2 r+ 1 2 from one sequence and to be at most the same quantity by the other. If σ0 ( θ∗ 0, then the inequalities are strict, because at least one of the inclusions θ0 ⊆ θ and σ0 ⊆ θ∗ is proper, which proves the claim. If σ0 = θ∗ 0, then only θ = θ0 term appears. It remains to observe that while the tensor product description of (2.2) is not canonical, the 0 (Cθ∗ embedding Cθ∗ 0) is a natural subspace of Λ∗NCC. Also, the inherent ambiguity in con- struction of L(θ0⊕σ0) does not occur at (θ0⊕σ0) itself, where the sections are canonically isomorphic to C. (cid:3) 0 ⊆ NC is. Thus the one-dimensional space Λdim θ∗ 3. Double Koszul complexes for dual reflexive Gorenstein cones In this section we use Theorem 2.1 and Proposition 2.4 to prove a result on cohomology of a certain complex built from a pair of dual reflexive Gorenstein cones. This fills in a gap in the proof in [BoM]. The reader should be warned however, that this is ultimately not the right complex to consider. In the next section we will modify the complex somewhat so that the cohomology admits a flat connection. Let K and K ∨ be dual reflexive Gorenstein cones in lattices M and N as in Section 1, deg and deg∨ their degree elements, ∆ and ∆∨ the sets of lattice elements of degree one and f and g nondegenerate coefficient functions. As in Section 1, we consider the quotient C[(K ⊕ K ∨)0] of the ring C[K ⊕ K ∨] by the ideal spanned by [m, n], m · n > 0. Define V = C[(K ⊕ K ∨)0] ⊗ Λ∗NC 10 LEV A. BORISOV and an endomorphism d : V → V by (3.1) df,g = X f (m)[m] ⊗ (contr.m) + X m∈∆ n∈∆∨ g(n)[n] ⊗ (n∧). It is easy to see that d2 f,g = 0. Moreover, df,g increases by one the natural grading on V given by [m, n] ⊗ P 7→ m · deg∨ + deg · n. An explicit description of the cohomology of V with respect to differential df,g has been claimed in [BoM], but the proof presented there was incorrect. Recall that for a face of θ ⊆ K one can define R1(f, θ) as the image of C[θ◦]/If,θC[θ◦] → C[θ]/If,θC[θ] where If,θ is the ideal generated by the logarithmic derivatives of θ, see Section 1. Similarly, one defines R1(g, θ∗) for the dual face θ∗ = Ann(θ) ∩ K ∨. Theorem 3.1. The cohomology of V with respect to df,g is naturally isomorphic to the direct sum over faces θ of K M {0}⊆θ⊆K R1(f, θ) ⊗ R1(g, θ∗) ⊗ Λdim θ∗ (Cθ∗). Proof. Consider the fan of facets of K ⊕K ∨. Consider the flabby locally free sheaf F on it given by (θ, σ) → C[θ⊕σ]. It can be naturally given a structure of the sheaf of modules over the ring of polynomial functions via a ring homomorphism Sym∗(MC ⊕ NC) → C[K ⊕ K ∨] defined on generators by (a, b) 7→ X (m · b)f (m)[m ⊕ 0] + X m∈∆ n∈∆∨ (a · n)g(n)[0 ⊕ n]. In other words, linear functions on M (resp. N) act as multiplications by the logarithmic partial derivatives of f (resp. g). This induces on F the structure of the module over polynomial functions on the faces θ ⊕ σ of K ⊕ K ∨. In fact, this sheaf is a product of pullbacks of two similarly defined sheaves FM and FN on K and K ∨. The key observation of [BoM] is that the non-degeneracy of f and g implies that F is locally free as a sheaf of modules over the rings of polynomials, and similarly for FM and FN . In addition, FM and FN are graded by degree. Thus, the machinery of [BrL] applies. We can write (non-canonically) these sheaves as direct sums of copies of graded locally free modules Lθ and Lσ which originate at θ and σ respectively. We have (3.2) FM = M 0⊆θ⊆K Hθ ⊗ Lθ, FN = M 0⊆σ⊆K ∨ Hσ ⊗ Lσ ON STRINGY COHOMOLOGY SPACES 11 where Hθ and Hσ are finite-dimensional graded vector spaces. The space C[(K ⊕ K ∨)0] is the space of global sections of F on the open set Φ that corresponds to Sθ(θ, θ∗) so by (3.2) the space V = C[(K ⊕ K ∨)0] ⊗ Λ∗NC decomposes as M θ,σ Hθ ⊗ Hσ ⊗ Γ(Φ, Lθ × Lσ) ⊗ Λ∗NC The differential df,g preserves the direct sum, and in fact acts on each Γ(Φ, Lθ × Lσ) ⊗ Λ∗NC as the differential of Proposition 2.4. Indeed, for any pair of dual bases (mi) of M and (ni) of N we have m∈∆ X = X f (m)[m] ⊗ (contr.m) = X (X f (m)[m] ⊗X f (m)(m · ni)[m]) ⊗ (contr.mi) = X m∈∆ i i i and similarly for the other part of the differential. m∈∆ (m · ni)(contr.mi) xi ⊗ (contr.mi) Let us calculate the cohomology. If σ 6⊆ θ∗ then Γ(Φ, Lθ × Lσ) If σ ⊆ θ∗ then by Proposition 2.4 the is 0, as is the cohomology. cohomology is only nonzero for σ = θ∗, in which case it can be identified with Λdim θ∗(Cθ∗). Thus, as a graded vector space, the cohomology of V ⊗ Λ∗NC is given by M θ Hθ ⊗ Hθ∗ ⊗ Λdim θ∗ (Cθ∗). It has been observed in [BoM] that R1(f, θ) is isomorphic to Hθ as as a graded vector space (and similar R1(g, θ∗) is isomorphic to Hθ∗). Indeed, in the decomposition (3.2), the space C[θ◦] is the kernel of the restriction map from C[θ] to the boundary, so it is Hθ ⊗Sym∗(Cθ)∨. We have thus established the isomorphism of the statement of the theorem, but we have not showed that it is canonical, since the decomposition into direct sum of copies of minimal locally free flabby sheaves is not. In order to define this isomorphism canonically, we will use the ar- gument of [BoM]. Namely, there are maps of complexes (3.3) M θ C[θ◦ ⊕ (θ∗)◦] ⊗ Λ∗NC → V → M θ C[θ ⊕ θ∗] ⊗ Λ∗NC with differentials df,g defined by the same formula. This induces maps in cohomology, and the composition of these maps has image isomor- phic to M {0}⊆θ⊆K R1(f, θ) ⊗ R1(g, θ∗) ⊗ Λdim θ∗ (Cθ∗) 12 LEV A. BORISOV see [BoM]. Thus, this space is a sub-quotient of the cohomology of V , and equality of dimensions implies that it is isomorphic to the coho- mology of V , i.e. the first map of (3.3) leads to a surjective map in cohomology and the second one leads to an injective map in cohomol- ogy. (cid:3) Remark 3.2. The arguments of Theorem 3.1 are applicable to the partial (called deformed in earlier papers) lattice algebras C[K ⊕ K ∨]Σ. In the next section that we will see that it can also be used for a slightly modified differential. 4. Modified double Koszul complexes and GKZ hypergeometric system In this section we modify the double Koszul complex of the previous section so that the cohomology is naturally endowed with a flat connec- tion. The issue at hand is that while the spaces R1(g, θ) form bundles over the space of nondegenerate coefficient functions g, there is no natu- ral connection on them. To create such connection, one needs to modify the spaces slightly by altering the action of logarithmic derivatives, as in [BoH]. Definition 4.1. Consider the sheaf of abelian groups \C[K ∨] on the fan of faces of K ∨ whose sections over σ ⊆ K ∨ are given by C[σ]. We give it a structure of the sheaf of modules over Sym∗((Cσ)∨) by declaring for each c ∈ σ, µ ∈ (Cσ)∨ µb[c] = X n∈∆∨∩σ g(n)µ(n)\[n + c] + µ(c)b[c]. considered in the previous section and [BoM]. Throughout the rest Remark 4.2. If not for the term µ(c)b[c], the above is just the structure of the paper, we use b[c] as opposed to [c] to signify this new module structure. From now on we will also use the notation dC[σ] for the space of sections of \C[K ∨] on the open subset that corresponds to σ. Note that \C[K ∨] is no longer naturally graded. However, it is natu- rally filtered, and the associated graded object is naturally isomorphic to C[K ∨] as a graded module. We can use results about C[K ∨] to infer statements about \C[K ∨] in view of the following theorem, proved in [BoH]. ON STRINGY COHOMOLOGY SPACES 13 Theorem 4.3. [BoH] There exists a non-canonical isomorphism of sheaves of locally free modules \C[K ∨] ≃ C[K ∨]. Moreover, this iso- morphism can be chosen to act as identity on the associated graded objects. The following definition mimics the one for R1(g, σ). Definition 4.4. For any σ ⊆ K ∨ we have a natural inclusion [C[σ◦] → dC[σ]. We define \R1(g, σ) as the image of where I is the irrelevant ideal in Sym∗((Cσ)∨). [C[σ◦]/I [C[σ◦] → dC[σ]/I dC[σ] As g varies, the spaces \R1(g, σ) can be given a natural flat connection based on the following result of [BoH]. Proposition 4.5. [BoH] The space \R1(g, K ∨) is naturally isomorphic to the dual of the space of solutions to the better-behaved GKZ hyper- geometric system bbGKZ(∆∨, (K ∨)◦, β = 0) that can be extended to the solutions of bbGKZ(∆∨, K ∨, β = 0), in the neighborhood of g. The space has a natural filtration by the order of vanishing at g and the associated graded space is naturally isomorphic to R1(g, K ∨). Indeed, we can define the connection by declaring the solutions to the aforementioned equations to be flat. For σ ( K ∨ we can can simply consider the solutions to be independent of the parameter g(v) for v 6∈ σ. We remark that these spaces have a geometric interpretation in terms of cohomology of hypersurfaces in tori with the Gauss-Manin connection, see [Ba1] but we will not focus on this. We are now ready to define and calculate the main objects of interest, namely the stringy cohomology spaces associated to dual Gorenstein cones. Definition 4.6. Let K and K ∨ be dual reflexive Gorenstein cones. Define as before the space V = C[(K ⊕ K ∨)0] ⊗ Λ∗NC. Consider the differential bdf,g on it given by bdf,g([m1 ⊕ n1] ⊗ P ) = df,g([m1 ⊕ n1] ⊗ P ) + [m1 ⊕ n1] ⊗ (n1 ∧ P ) = X + X g(n)[m1 ⊕ (n + n1)] ⊗ (n ∧ P ) + [m1 ⊕ n1] ⊗ (n1 ∧ P ). f (m)[(m + m1) ⊕ n1] ⊗ (contr.m)(P ) m∈∆ n∈∆∨ 14 LEV A. BORISOV Then we define stringy cohomology B-space HB(Xf , X ∨ g ) as the coho- mology of V with respect to bdf,g, with the grading given by [m ⊕ n] ⊗ P 7→ 2 m · deg∨ + deg(P ). We will also call this the A-space of the mirror pair (X ∨ HA(Xf , X ∨ C[(K ⊕ K ∨)0] ⊗ Λ∗MC by the differential that maps g , Xf ). Similarly, g , Xf )) is defined as the cohomology of g ) (equal to HB(X ∨ [m1 ⊕ n1] ⊗ P 7→ X m∈∆ f (m)[(m + m1) ⊕ n1] ⊗ (m ∧ P ) n∈∆∨ g(n)[m1 ⊕ (n + n1)] ⊗ (contr.n)(P ) + [m1 ⊕ n1] ⊗ (m1 ∧ P ). + X It is straightforward to see that bdf,g is a differential. Moreover, we can describe its cohomology explicitly. Theorem 4.7. For nondegenerate f and g, the B-space HB(Xf , X ∨ g ) is naturally isomorphic to M {0}⊆θ⊆K R1(f, θ) ⊗ \R1(g, θ∗) ⊗ Λdim θ∗ (Cθ∗). The bundle of HB(Xf , X ∨ natural flat connection. g ) over the space of nondegenerate g has a Proof. We have the maps of complexes M C[θ◦] ⊗ \C[(θ∗)◦] ⊗ Λ∗NC → V → M θ θ C[θ] ⊗ [C[θ∗] ⊗ Λ∗NC similar to (3.3). Then, as in the proof of Theorem 3.1 the space L{0}⊆θ⊆K R1(f, θ) ⊗ \R1(g, θ∗) ⊗ Λdim θ∗(Cθ∗) is a sub-quotient of the cohomology of V with respect to bdf,g. In view of Theorem 4.3, there is an isomorphism between (3.3) and the above maps of complexes. This implies that the left map induces a surjection in the cohomology and the right map induces an injection. This proves the first asser- tion of the theorem, and the flat connection statement follows from Proposition 4.5. (cid:3) Remark 4.8. The stringy cohomology spaces HA/B(Xf , X ∨ g ) are only single-graded rather than double-graded. However, they possess natu- ral filtrations such that the associated graded spaces are those described in [BoM]. Thus, spaces HA/B(Xf , X ∨ g ) are stringy analogs of de Rham cohomology while the spaces of [BoM] are stringy analogs of Dolbeault cohomology. ON STRINGY COHOMOLOGY SPACES 15 5. Product structure on the stringy cohomology spaces The previous candidates for the stringy cohomology spaces that were considered in [BoM] have been given a structure of super-commutative algebra (in fact, two structures depending on whether A or B version are considered) by identifying them with chiral rings of a certain N = 2 vertex algebra. We would like to extend these results of [Bo1] and [Bo2] to the newly constructed spaces HB(Xf , X ∨ g ). What we find in the process is that contrary to the situation we have encoun- g ) and HA(Xf , X ∨ tered so far in this paper, the b analog of the vertex algebra of mirror symmetry is drastically different from the one previously considered. In fact, it is just the stringy cohomology space itself.3 Working knowledge of vertex algebras is assumed for this more technical section. Let us first recall the construction of [Bo1] and [Bo2], focusing on the B space. Let M, N, K, K ∨, ∆, ∆∨, f and g be as before. Consider first the lattice vertex algebra FockM ⊕N . It is generated by the fields mbos(z), nbos(z), eR mbos(z), eR nbos(z), mf erm(z), nf erm(z) in the world-sheet variable z. Consider the differential Df,g on FockM ⊕N Df,g := Resz=0(cid:16) X m∈∆ f (m)mf erm(z)eR mbos(z)+X n∈∆∨ g(n)nf erm(z)eR nbos(z)(cid:17). We denote by Vf,g the cohomology of FockM ⊕N with respect to Df,g. It carries a natural structure of N = 2 vertex algebra, induced from FockM ⊕N by G+(z) = Pk(nk)bos(z)(mk)f erm(z) − ∂z degf erm(z) G−(z) = Pk(mk)bos(z)(nk)f erm(z) − ∂z(deg∨)f erm(z) J(z) = Pk(mk)f erm(z)(nk)f erm(z) + degbos(z) − (deg∨)bos(z) L(z) = Pk(mk)bos(z)(nk)bos(z) + 1 2 Pk(mk)f erm(z)∂z(nk)f erm(z) − 1 2 Pk ∂z(mk)f erm(z)(nk)f erm(z) 2 ∂z(deg∨)bos(z) 2∂z degbos(z) − 1 The operators L[0] = Resz=0(zL(z)) and J[0] = Resz=0J(z) play a special role and are called conformal weight and fermion numbers re- spectively. The main result of [Bo2] is the following. − 1 Theorem 5.1. [Bo2] For strongly nondegenerate f and g, the above N = 2 structure on Vf,g is of sigma model type. This means that the gradings by HA = L[0] − 1 2J[0] are integer and nonnegative. 2J[0] and HB = L[0] + 1 3This should be viewed as analogous to the statement of [MSV] that chiral de Rham complex is quasi-isomorphic to the usual de Rham complex. 16 LEV A. BORISOV Remark 5.2. We refer the reader to [Bo2] for the technical definition of strong non-degeneracy. We will not need the precise statement here. Every vertex algebra of sigma model type has two natural subspaces, called A and B chiral rings characterized as kernels of HA and HB re- spectively. In the case of Vf,g these have been computed to coincide with the stringy cohomology spaces suggested in [BoM]. The compu- tation is based on the following result. Theorem 5.3. [Bo2] Let f and g be strongly non-degenerate. Then Df,g-cohomology of FockM ⊕N has only nonnegative integer eigenvalues of HA. Moreover, the HA = 0 eigenspace comes from FockK⊕(K ∨−deg∨). The operator HB also has only nonnegative integer eigenvalues on Vf,g and its kernel comes from Fock(K−deg)⊕K ∨. As a corollary, the kernel of HB is isomorphic to the cohomology of Df,g of the HB = 0 subspace of Fock(K−deg)⊕K ∨. This subspace is seen to be spanned by elements that correspond to fields of the form4 (5.1) nf erm 1 (z) · · · nf erm k (z)eR (m−deg)bos(z)+nbos(z) for ni ∈ NC, m ∈ K, n ∈ K ∨, m · n = 0. The space of these fields is naturally isomorphic to C[(K ⊕ K ∨)0] ⊗ Λ∗NC, with the action of Df,g given by df,g from (3.1). This leads to the following result. Corollary 5.4. [Bo2] For strongly nondegenerate f and g, the coho- mology of C[(K ⊕ K ∨)0] ⊗ Λ∗NC with respect to df,g is naturally iden- tified with the B chiral ring of Vf,g. It thus carries a natural structure of graded associative algebra. Remark 5.5. It is currently very unclear how to write the product implied by the above corollary without the full force of vertex algebra machinery. The trouble is that as one takes the operator product ex- pansions of the fields of (5.1), one typically lands in Fock(K−2 deg)⊕K ∨ rather than Fock(K−deg)⊕K ∨. It then takes delicate vertex algebra con- siderations to prove that these fields can be reduced to fields from Fock(K−deg)⊕K ∨ in Df,g cohomology. We would like to extend the results of Theorem 5.3 and Corollary 5.4 to the new version of the differential bdf,g on C[(K ⊕K ∨)0] ⊗Λ∗NC. The first observation is that the additional term in it is actually a residue of a very natural field. For an element [m⊕n]⊗P of C[(K ⊕K ∨)0]⊗Λ∗NC we denote the corresponding field by P (N f erm)(z)eR (m−deg)bos(z)+nbos(z). 4with the normal ordering implicitly assumed ON STRINGY COHOMOLOGY SPACES 17 Proposition 5.6. The action of Resz=0G−(z) on a field P (N f erm)(z)eR (m−deg)bos(z)+nbos(z) is given by (n ∧ P )(N f erm)(z)eR (m−deg)bos(z)+nbos(z). Proof. This is a routine lattice vertex algebra calculation, which we nonetheless include for the benefit of the reader. The ∂z(deg∨)f erm(z) in G−(z) does not contribute to the residue at zero and can be ignored. Then we have the operator product expansion (mk)bos(z)(nk)f erm(z)P (N f erm)(w)eR (m−deg)bos(w)+nbos(w) X k ∼ (z − w)−1X k (mk · n)(nk)f erm(w)P (N f erm)(w)eR (m−deg)bos(w)+nbos(w) = (z − w)−1nf erm(w)P (N f erm)(w)eR (m−deg)bos(w)+nbos(w) = (z − w)−1(n ∧ P )(N f erm)(z)eR (m−deg)bos(z)+nbos(z), which verifies the statement of the proposition. (cid:3) Remark 5.7. The operator Resz=0G−(z) has an important geometric meaning. In terms of X ∨ g it roughly corresponds to the de Rham differ- ential. Also note that cohomology of Vf,g with respect to Resz=0G−(z) is the B chiral ring of Vf,g. It is therefore reasonable to add the residue of G−(z) to the definition of Df,g. Definition 5.8. We denote by bVf,g;B the cohomology of FockM ⊕N by the differential bDf,g;B given by Resz=0(cid:16) X +X m∈∆ f (m)mf erm(z)eR mbos(z) + X (mk)bos(z)(nk)f erm(z) − ∂z(deg∨)f erm(z)(cid:17) n∈∆∨ g(n)nf erm(z)eR nbos(z) k for dual bases (mk) and (nk) of M and N. We will now show that bVf,g;B is naturally isomorphic to the stringy cohomology B-space HB(Xf , X ∨ g ). Theorem 5.9. For strongly nondegenerate f and g the cohomology bVf,g;B of Definition 5.8 is naturally isomorphic to HB(Xf , X ∨ g ). 18 LEV A. BORISOV Proof. The proof proceeds in several steps. First, we show that the cohomology of FockM ⊕N with respect to bDf,g;B is naturally isomorphic to the cohomology of FockM ⊕K ∨ with respect to the same operator. The analogous fact for the cohomology with respect to Df,g is proved in [Bo1, Proposition 8.2]. The idea of [Bo1, Proposition 8.2] is to construct operators R such that the anticommutator of R with bDf,g;B is equal to identity plus an operator that increases a grading with respect to a ray of K. It remains to observe that the operators R constructed in [Bo1, Proposition 8.1] anti-commute with G−(z), because the former are made from eR M bos(z) and N f erm(z) fields only, and the latter is made from M bos(z) and N f erm(z). Thus the same fields R can be used in the new argument. Now we are dealing with cohomology of FockM ⊕K ∨ with respect to bDf,g;B. The method of [Bo2] was to employ a spectral sequence for the double complex, with the f and g parts of the differential acting as horizontal and vertical arrows. This spectral sequence does not converge on the whole FockM ⊕N , but does converge on FockM ⊕K ∨. The situation is a bit different in the current setting, because the additional term in bDf,g;B destroys the double grading by degree in M and N, however there is still a spectral sequence which we will describe below if we use a more subtle double grading. Consider the operator HA on FockM ⊕K ∨ that comes from the N = 2 structure. Note that Df,g preserves eigenvalues of HA and Resz=0G−(z) increases HA by one, as part of the commutator relations of the N = 2 algebra. Thus, we can think of bDf,g;B as a differential of the total complex of the double complex with the double grading (HA + • · deg∨, deg ·•) and d1 = Resz=0(cid:16) X m∈∆ f (m)mf erm(z)eR mbos(z) + G−(z)(cid:17) d2 = Resz=0 X n∈∆∨ g(n)nf erm(z)eR nbos(z). The corresponding spectral sequence that starts with Hd2(FockM ⊕K ∨) then converges, since the second grading is nonnegative on FockM ⊕K ∨. The cohomology with respect to d2 has been studied in [Bo2]. In particular, the HB = 0 part of the cohomology is zero for Fockc⊕K ∨ for c 6∈ K − deg, provided g is strongly nondegenerate. Then, as in [Bo2] we conclude that the HB = 0 part of the cohomology of FockM ⊕K ∨) with respect to bDf,g;B is the HB = 0 part of the coho- mology of FockK−deg ⊕K ∨). As a corollary of Proposition 5.6, we see ON STRINGY COHOMOLOGY SPACES 19 g ). that the HB = 0 part of the cohomology of FockM ⊕N with respect to It remains to show that HB 6= 0 part of the cohomology vanishes, bDf,g;B is naturally isomorphic to HB(Xf , X ∨ which is only a feature of bVf,g;B rather than Vf,g. We use Resz=0zG+(z) which anti-commutes with the f and g terms of bDf,g;B. As part of the N = 2 algebra structure, the anticommutator of Resz=0zG+(z) with the Resz=0G−(z) is HB. This shows that the cohomology occurs only at HB = 0, since in other eigenspaces of HB the operator Resz=0zG+(z) provides homotopy to the identity. (cid:3) Corollary 5.10. For strongly nondegenerate f and g the stringy co- homology spaces HB(Xf , X ∨ g , Xf ) and HA(Xf , X ∨ g ) = HB(X ∨ g ) = HA(X ∨ g , Xf ) are equipped with natural structures of graded associative super-com- mutative algebras. Proof. It is sufficient to consider HB(Xf , X ∨ g ) = bVf,g;B, the other is obtained by switching M and N. Take two bu and bv elements in bVf,g;B, and lift them to u, v ∈ FockM ⊕N . Then take the operator product expansion of the corresponding fields u(z)v(w) = X k≥−l (z − w)kck(w). Moreover, only c0 has HB = 0 and can be nonzero in the cohomology. Each ck is annihilated by bDf,g;B and thus descends to cohomology. We define the product by bu ∗ bv := bc0. Standard vertex algebra con- siderations then assure that this product is (super)commutative and associative. (cid:3) Remark 5.11. The HB ring is in fact graded by what was called in [Bo2] the sum of fermion number and cohomology grading. It coincides with the total grading of the double complex in the proof of Theorem 5.9. It also coincides with the grading of Definition 4.6. 6. Concluding remarks and open questions One can glean from the argument that in order to define the product structure on HB(Xf , X ∨ g ) one does not need both f and g to be strongly nondegenerate. Rather, one needs g to be strongly nondegenerate while usual non-degeneracy suffices for f . The spaces themselves are defined and are of constant dimension with just a simple non-degeneracy as- sumption on both f and g. It is possible that strong non-degeneracy is an artifact of the argument, rather than a consequence of product 20 LEV A. BORISOV structure acquiring poles at nondegenerate by not strongly nondegen- erate g. Can one describe the product on the stringy cohomology in terms of commutative algebra, thus avoiding vertex algebra machinery? This has been a recurring question for the last ten years or so. However, it is possible that the extra term Resz=0G−(z) in the differential may allow for an easier reduction from Fock(K−2 deg)⊕K ∨ to Fock(K−deg)⊕K ∨ which lies at the heart of the matter. The structures described in this paper correspond to small quantum cohomology. What is the big quantum cohomology for these examples? In other words, we need to construct Frobenius manifolds based on the stringy cohomology. Similarly one can ask whether the construction of this paper can lead to an algebraic construction of the A∞ categories of boundary condi- tions in the open string theory. When dual Gorenstein cones (K, K ∨) lead to complete intersection Calabi-Yau varieties, this would amount to an alternative description of Fukaya category and derived category of coherent sheaves, and might lead to a verification of Homological Mirror Symmetry. The dual Gorenstein cones construction has been recently generalized in two directions. In [Bo3], the duality condition has been relaxed a bit, to include the examples of Berglund and Hubsch. It appears that the arguments of this paper should extend to the setting of almost dual Gorenstein cones, but this needs to be checked. Another generalization, studied in [BoK], modifies the differential Df,g by allowing more general linear combinations of the fermionic fields. The resulting algebra no longer has N = 2 structure and is related to N = (0, 2) nonlinear sigma models. It would be interesting to see to what extent the definitions of this paper extend to this more general setting. References [Ba1] [Ba3] V.V. Batyrev, Variations of the mixed Hodge structure of affine hypersur- faces in algebraic tori. Duke Math. J. 69 (1993), no. 2, 349 -- 409. V.V. Batyrev, Stringy Hodge numbers of varieties with Gorenstein canoni- cal singularities. Integrable systems and algebraic geometry (Kobe/Kyoto, 1997), 1 -- 32, World Sci. Publ., River Edge, NJ, 1998. [BaBo] V.V. Batyrev, L.A. Borisov, Mirror duality and string-theoretic Hodge [BaD] numbers. Invent. Math. 126 (1996), no. 1, 183 -- 203. V.V. Batyrev, D.I. Dais, Strong McKay correspondence, string-theoretic Hodge numbers and mirror symmetry. Topology 35 (1996), no. 4, 901 -- 929. ON STRINGY COHOMOLOGY SPACES 21 [BaN] [Bo1] [Bo2] [Bo3] [BoH] [BoK] [BoM] [BrL] V.V. Batyrev, B. Nill, Combinatorial aspects of mirror symmetry. Integer points in polyhedra -- geometry, number theory, representation theory, al- gebra, optimization, statistics, 35 -- 66, Contemp. Math., 452, Amer. Math. Soc., Providence, RI, 2008. L.A. Borisov, Vertex algebras and mirror symmetry. Comm. Math. Phys. 215 (2001), no. 3, 517 -- 557. L.A. Borisov, Chiral rings of vertex algebras of mirror symmetry. Math. Z. 248 (2004), no. 3, 567 -- 591. L.A. Borisov, Berglund-Hubsch mirror symmetry via vertex algebras. preprint arXiv:1007.2633. L.A. Borisov, R.P. Horja, On the better behaved version of the GKZ hy- pergeometric system. preprint arXiv:1011.5720. L.A. Borisov, R.M. Kaufmann, On CY-LG correspondence for (0,2) toric models. preprint arXiv:1011.5720, to appear in Advances in Mathematics. L.A. Borisov, A.R. Mavlyutov, String cohomology of Calabi-Yau hyper- surfaces via mirror symmetry, Adv. Math. 180 (2003), no. 1, 355 -- 390. P. Bressler, V. Lunts, Intersection cohomology on nonrational polytopes, Compositio Math. 135 (2003), no. 3, 245 -- 278. [BBFK] G. Barthel, J.-P. Brasselet, K.-H. Fieseler, L. Kaup, Combinatorial inter- section cohomology for fans. Tohoku Math. J. (2) 54 (2002), no. 1, 1 -- 41. K. Karu, Hard Lefschetz theorem for nonrational polytopes. Invent. Math. 157 (2004), no. 2, 419 -- 447. [Ka] [MSV] F. Malikov, V. Schechtman, A. Vaintrob, Chiral de Rham complex. Comm. [St] Math. Phys. 204 (1999), no. 2, 439 -- 473. R. Stanley, Generalized H-vectors, Intersection Cohomology of Toric Va- rieties, and Related Results. Adv. Stud. in Pure Math. 11 (1987), 187 -- 213. Mathematics Department, Rutgers University, 110 Frelinghuysen Rd, Piscataway, NJ 08540, USA E-mail address: [email protected] 

1802.01311

2

1802

Non-surjective Gaussian maps for singular curves on K3 surfaces

Let $(S,L)$ be a polarized K3 surface with $\mathrm{Pic}(S) = \mathbb{Z}[L]$ and $L\cdot L=2g-2$, let $C$ be a nonsingular curve of genus $g-1$ and let $f:C\to S$ be such that $f(C) \in \vert L \vert$. We prove that the Gaussian map $\Phi_{\omega_C(-T)}$ is non-surjective, where $T$ is the degree two divisor over the singular point $x$ of $f(C)$. This generalizes a result of Kemeny with an entirely different proof. It uses the very ampleness of $C$ on the blown-up surface $\widetilde S$ of $S$ at $x$ and a theorem of L'vovski.

math.AG

math

NON-SURJECTIVE GAUSSIAN MAPS FOR SINGULAR CURVES ON K3 SURFACES CLAUDIO FONTANARI AND EDOARDO SERNESI Abstract. Let (S, L) be a polarized K3 surface with Pic(S) = Z[L] and L·L = 2g − 2, let C be a nonsingular curve of genus g − 1 and let f : C → S be such that f (C) ∈ L. We prove that the Gaussian map ΦωC (−T ) is non-surjective, where T is the degree two divisor over the singular point x of f (C). This generalizes a result of Kemeny with an entirely different proof. It uses the very ampleness of C on the blown-up surface eS of S at x and a theorem of L'vovski. . G A h t a m [ 2 v 1 1 3 1 0 . 2 0 8 1 : v i X r a 1. Introduction Let C be a complex projective nonsingular curve of genus g. Let L, A be invert- ible sheaves on C and let R(L, A) := ker[H 0(C, L) ⊗ H 0(C, A) −→ H 0(C, LA)] Then we can define a Gaussian map ΦL,A : R(L, A) −→ H 0(C, ωC LA) in a well-known way that will be recalled in §2. If L = A the map ΦL,L has the same image as its restriction to V2 H 0(C, L) ⊂ R(L, L), which is denoted by: 2^ H 0(C, L) −→ H 0(C, ωC L2) ΦL : If we take L = ωC then the map: 2^ H 0(C, ωC ) −→ H 0(C, ω3 C ) ΦωC : is called the Wahl map. The following result, due to Wahl (see [W] and also [BM] for a different proof), gives a necessary condition for a nonsingular curve to be hyperplane section of a K3 surface: Theorem 1 (Wahl). Every nonsingular curve in a very ample linear system L on a K3 surface S has non-surjective Wahl map. This result has been generalized by L'vovski (see [L] and also [BF] for an ele- mentary detailed proof) in the following form: 2010 Mathematics Subject Classification. Primary 14J28, 14H10; Secondary 14H51. Key words and phrases. K3 surface, nodal curve, Gaussian map, Wahl map. This research has been partially supported by GNSAGA of INdAM, by PRIN 2015 "Geometria delle variet`a algebriche", and by FIRB 2012 "Moduli spaces and Applications". 1 2 CLAUDIO FONTANARI AND EDOARDO SERNESI Theorem 2 (L'vovski). Let C be a smooth curve of genus g > 0 and let A be a very ample line bundle on C embedding C in Pn, n ≥ 3. If C ⊂ Pn is scheme- theoretically a hyperplane section of a smooth surface X ⊂ Pn+1 then the Gaussian map ΦωC ,A is non-surjective. In this paper we will focus on singular curves on a K3 surface S. Our starting point is the recent work [Ke] by Kemeny. Let V 1 g be the moduli space of triples (S, L, f : C → S), where (S, L) is a polarized K3 surface with L · L = 2g − 2, C is a smooth curve of genus g − 1 and f is an unramified stable map, birational onto its image, such that f (C) ∈ L. Then the following holds: Theorem 3 ([Ke], Theorem 1.7). Fix an integer g ≥ 14. Then there is an irre- ducible component I 0 ⊆ V 1 g such that for a general triple (S, L, f : C → S) ∈ I 0 the Gaussian map ΦωC (−T ) is non-surjective, where T = P + Q ⊆ C is the divisor over the node of f (C). The component I 0 appearing in the statement might a priori include all 1-nodal curves in L, but this is not proved in [Ke]. The proof is rather indirect and relies on the fact that H 0(C, f ∗TS) = 0 for the general triple (S, L, f : C → S) ∈ I 0 (see [Ke], Lemma 3.17). Our main result is the following more general statement, which is an exact ana- logue of Theorem 1 for singular curves: Theorem 4. Fix an integer g ≥ 9. Let (S, L) be a polarized K3 surface such that Pic(S) = Z[L] and L · L = 2g − 2. Let C be a smooth curve of genus g − 1 endowed with a morphism f : C → S birational onto its image and such that f (C) ∈ L. If T = P + Q ⊆ C is the divisor over the singular point of f (C), then the Gaussian map ΦωC (−T ) is non-surjective. Note that we are not making any generality assumption on the triple (S, L, f : C → S). In particular we are not assuming that f (C) is a nodal curve: the hypothesis that C has genus g − 1 just implies that f (C) is either 1-nodal or has an ordinary cusp. Note also, by contrast, that the normalization C of a 1-nodal curve on a K3 surface tends to have surjective Wahl map ΦωC , and therefore, by Theorem 1, not to be embeddable in any K3 surface. In fact, the following result holds: Theorem 5 (Sernesi [S]). Let (S, L) be a general primitively polarized K3 surface of genus g + 1. Assume that g = 40, 42 or ≥ 44. Then the Wahl map of the normalization of any 1-nodal curve in L is surjective. In outline, the proof of Theorem 4 goes as follows. We prove that on the blow-up σ : eS → S at the singular point of f (C) the line bundle H := σ∗L(−2E) is very ample, where E denotes the exceptional divisor of the blow-up σ : eS → S. This fact is a special case of Theorem 10, which gives a more general very-ampleness criterion of independent interest. Hence we can apply Theorem 2 to A := H · C = C · C = (C + E) · C − E · C = ωC (−T ) and obtain that the Gaussian map ΦωC ,ωC (−T ) is non-surjective. Finally, we prove that coker(ΦωC (−T )) surjects onto coker(ΦωC ,ωC(−T )) (Theorem 8). In particular, also ΦωC (−T ) is non-surjective. We work over the field C of complex numbers. NON-SURJECTIVE GAUSSIAN MAPS FOR SINGULAR CURVES ON K3 SURFACES 3 2. Conormal sheaves and projections Let C be as in the Introduction. It turns out to be natural, for our purposes, to introduce the so-called syzygy sheaves. We define the syzygy sheaf ML of a globally generated invertible sheaf L by the exact sequence: 0 → ML → H 0(C, L) ⊗ OC → L → 0 If L is very ample the above sequence is a twist of the dualized Euler sequence, and we get (1) ML = Ω1 PC ⊗ L where P ∼= PH 0(C, L)∨. Therefore the conormal sequence of C ⊂ P twisted by L takes the following form: (2) 0 / N ∨ C/P ⊗ L / ML / ωC L / 0 Now let A be another invertible sheaf on C and tensor (2) by A: (3) 0 / N ∨ C/P ⊗ LA / ML ⊗ A ρ / ωC LA / 0 Then H 0(C, ML ⊗ A) = R(L, A) and the map induced by ρ on global sections: ΦL,A : H 0(C, ML ⊗ A) −→ H 0(C, ωC LA) is the Gaussian map of L, A. When L = A we have: where H 0(C, ML ⊗ L) = I2 M 2^ H 0(C, L) I2 = ker[S2H 0(C, L) → H 0(C, L2)] Since I2 ⊂ ker(ΦL,L) the map ΦL,L has the same image as its restriction to V2 H 0(C, L) ⊂ R(L, L), which is denoted by: 2^ H 0(C, L) −→ H 0(C, ωC L2) ΦL : Now let L be very ample, P ∈ C and assume that L(−P ) is also very ample. Then we have embeddings: ϕL : C → Pr, ϕL(−P ) : C → Pr−1 where h0(C, L) = r + 1. The following proposition relates the conormal sheaves N ∨ C/Pr and N ∨ C/Pr−1. Proposition 6. There is an exact sequence: (4) 0 → N ∨ C/Pr−1 ⊗ L(−P ) → N ∨ C/Pr ⊗ L → OC (−2P ) → 0 Proof. There is an exact sequence 0 → ML(−P ) → ML → OC (−P ) → 0 / / / / / / / / 4 CLAUDIO FONTANARI AND EDOARDO SERNESI induced by the inclusion H 0(C, L(−P )) ⊂ H 0(C, L). Recalling (1) we get a com- mutative and exact diagram whose first two rows are twisted conormal sequences: 0 0 0 0 0 0 / N ∨ C/Pr−1 ⊗ L(−P ) / ML(−P ) ωC L(−P ) / N ∨ C/Pr ⊗ L ML / ωC L / OC (−2P ) / OC (−P ) / ωC L ⊗ OP 0 0 0 / 0 0 / 0 The first column gives the sequence (4). (cid:3) Corollary 7. Let C, L be as before and suppose that P, Q ∈ C are points such that L(−P − Q) is very ample. Then there is an exact sequence: (5) 0 → N ∨ C/Pr−2 ⊗ L(−P − Q) → N ∨ C/Pr ⊗ L → OC (−2P )M OC (−2Q) → 0 Proof. Left to the reader. (cid:3) 3. A comparison result between Gaussian maps After the preliminaries collected in the previous section, we are ready to prove the following result: Theorem 8. Let C be a projective nonsingular curve of genus g, T = P + Q an effective divisor of degree 2 on C. Assume that Cliff(C) ≥ 3. Then there is a surjection: coker(ΦωC (−T )) −→ coker(ΦωC ,ωC (−T )) −→ 0 In particular, if ΦωC ,ωC (−T ) is not surjective then ΦωC (−T ) is not surjective. Proof. The hypothesis Cliff(C) ≥ 3 implies that ω(−T ) is very ample and maps C ⊂ Pg−3. We have an exact sequence ([L89], Lemma 1.4.1): 0 → MωC (−T ) → MωC → OC (−P )M OC (−Q) → 0       /   / / /     / / / /     / / /   /   /   /   / NON-SURJECTIVE GAUSSIAN MAPS FOR SINGULAR CURVES ON K3 SURFACES 5 which, twisted by ωC(−T ), appears as the middle column in the following diagram: 0 0 0 0 0 0 / N ∨ C/Pg−3 ⊗ ω2 C (−2T ) / N ∨ C/Pg−1 ⊗ ω2 C (−T ) / MωC (−T ) ⊗ ωC(−T ) f / MωC ⊗ ωC (−T ) a b ω3 C(−2T ) / ω3 C (−T ) / ωC(−T − 2P ) L ωC(−T − 2Q) g / ωC(−T − P ) L ωC(−T − Q) / ω3 C (−T ) ⊗ OT 0 0 0 where the first column is (5) for L = ωC , twisted by ωC (−T ). The homomorphisms a and b induce ΦωC (−T ),ωC (−T ) and ΦωC ,ωC(−T ) respectively on global sections. Therefore, taking cohomology, we obtain the following diagram: 0 0 0 / coker(ΦωC (−T )) ζ / coker(ΦωC ,ωC (−T )) H 1(N ∨ C/Pg−3 ⊗ ω2 C (−2T )) / H 1(N ∨ C/Pg−1 ⊗ ω2 C (−T )) H 1(Ω1 Pg−3C ⊗ ω2 C(−2T )) H 1(f ) / H 1(Ω1 Pg−1 C ⊗ ω2 C(−T )) / ker(H 1(g)) / H 1(ωC (−T −2P )) H1 (ωC (−T −2Q)) L H 1(g) H 1(ωC (−T −P )) L H1 (ωC (−T −Q)) / 0 / 0 0 0 / 0 0 0 0 Since Cliff(C) ≥ 3 it follows that H 1(g) is an isomorphism, thus ζ is surjective. (cid:3) Theorem 8 can be generalized in several ways. for example, using similar methods and induction on n one can also prove the following: Theorem 9. Let C be a projective nonsingular curve of genus g, T = P1 + · · · + Pn an effective divisor of degree n ≥ 1. Let L be an invertible sheaf on C of degree d ≥ 2g + 1 + n. Then there is a surjection: coker(ΦL(−T )) −→ coker(ΦL,L(−T )) −→ 0 We will not pursue this here. 4. Very ampleness on blown-up surfaces For the proof of Theorem 4 we will need a special case of the following result of independent interest: Theorem 10. Let S be a K3 surface such that Pic(S) = Z[L] for some ample invertible sheaf L. Assume that L · L = 2g − 2 ≥ (ℓ + 1)2 + 3 for some ℓ ≥ 1. Let       /   / / /     / /   /   /   / /   /   / / /   / / /   / /   / /   /   /   /   / / / / /   / /   6 CLAUDIO FONTANARI AND EDOARDO SERNESI x ∈ S, σ : eS → S the blow-up of S at x and E ⊂ eS the exceptional curve. Then the sheaf H := σ∗L(−ℓE) is very ample on eS. Proof. We follow closely the application of Reider's method in [V], proof of Theorem H) = 0. 16. We must prove that for each subscheme Z ⊂ eS of length two we have H 1(eS, IZ ⊗ By contradiction, assume that H 1(eS, IZ ⊗ H) 6= 0 for some Z. By Serre duality we have: and therefore there is a non-split exact sequence: H 1(eS, IZ ⊗ H)∨ = Ext1(IZ , E − H) (6) 0 / σ∗L−1((ℓ + 1)E) / E / IZ / 0 where E is torsion free. We have: c1(E)2 = 2g − 2 − (ℓ + 1)2, c2(E) = 2 χ(Hom(E, E)) = 4χ(OeS) + c2 1(E) − 4c2(E) = 2g − 2 − (ℓ + 1)2 ≥ 3 Therefore by Serre duality: 2h0(Hom(E, E(E))) ≥ h0(Hom(E, E)) + h0(Hom(E, E(E)) ≥ 3 It follows that there is a homomorphism φ : E −→ E(E) which is not proportional to the identity. By the usual trick we can assume that φ is generically of rank one (see [V], proof of Proposition 15). ker(φ) and im(φ) are torsion free rank one, therefore of the form A ⊗ IW and B ⊗ IW ′ respectively, for some invertible sheaves A, B which are of the form: A = σ∗Lα(βE), B = σ∗L−1−α((ℓ + 1 − β)E) From the exact sequence 0 / A ⊗ IW / E / B ⊗ IW ′ / 0 we compute: (7) 2 = c2(E) = deg(W ) + deg(W ′) − β(ℓ + 1 − β) ≥ −β(ℓ + 1 − β) Indeed, since A ⊗ IW ⊂ E, from (6) we see that we must have α ≤ 0. Similarly −α − 1 ≤ 0 because B ⊗ IW ′ ⊂ E(E). Therefore: proving (7). −1 ≤ α ≤ 0, Suppose α = 0. Then we have an inclusion OeS(βE) ⊗ IW ⊂ IZ , which implies If β = 0 we get an inclusion If β < 0 then (7) gives a contradiction. β ≤ 0. IW ⊂ IZ . This implies that the pullback homomorphism: ψ : Ext1(IZ , E − H) −→ Ext1(IW , E − H) maps (6) to zero, thus ψ is not injective. But ψ is dual to: ψ∨ : H 1(eS, IW ⊗ H) −→ H 1(eS, IZ ⊗ H) / / / / / / / / NON-SURJECTIVE GAUSSIAN MAPS FOR SINGULAR CURVES ON K3 SURFACES 7 which is henceforth not surjective. On the other hand the diagram: (8) 0 T 0 0 / IW ⊗ H / H / HW / IZ ⊗ H / H / HZ / 0 0 shows that we have an exact sequence 0 0 / IW ⊗ H / IZ ⊗ H / T / 0 with T torsion: therefore ψ∨ is surjective, and we have a contradiction. So the case α = 0 cannot occur. Suppose α = −1. In this case we use the inclusion B ⊗ IW ′ ⊂ E(E). We obtain an inclusion IW ′ ((ℓ − β)E) ⊂ E which implies IW ′ ((ℓ − β)E) ⊂ IZ and therefore ℓ − β ≤ 0. The case ℓ − β = 0 gives an inclusion IW ′ ⊂ IZ and is treated using a diagram analogous to (8), leading to a contradiction as before. If ℓ − β < 0 then β ≥ ℓ + 1. If β > ℓ + 1 then (7) gives a contradiction. If β = ℓ + 1 then (7) gives: deg(W ) + deg(W ′) = 2 If deg(W ) > 0 then A ⊗ IW = σ∗L−1((ℓ + 1)E) ⊗ IW ⊂ σ∗L−1((ℓ + 1)E) and therefore φ(σ∗L−1((ℓ + 1)E)) is a torsion subsheaf of E(E), a contradiction. Then W = ∅ and W ′ = Z. This gives IZ ⊂ E(E), viz. IZ (−E) ⊂ E. This implies that the pullback θ : Ext1(IZ , E − H) −→ Ext1(IZ (−E), E − H) maps (6) to zero, thus the dual map: θ∨ : H 1(eS, IZ ⊗ (H − E)) −→ H 1(eS, IZ ⊗ H) is not surjective. But coker(θ∨) ⊂ H 1(E, IZ ⊗ OE(H)) = 0 because IZ ⊗ OE(H) is an invertible sheaf of degree ≥ 0 on E. We have a contradiction and the theorem is proved. (cid:3) Remarks 11. (i) For the first values of ℓ the condition of the theorem gives: ℓ = 1 : ℓ = 2 : ℓ = 3 : g ≥ 5 g ≥ 7 g ≥ 11. (ii) The case ℓ = 1, g = 5 has already been considered in [B]. (iii) An interesting implicit consequence of Theorem 10 is the following existence result:     /   / /   / / / / / /   / / / / 8 CLAUDIO FONTANARI AND EDOARDO SERNESI Under the assumptions of Theorem 10 for a given ℓ ≥ 2, through each point x ∈ S there exist integral curves in L having an ordinary multiple point of multiplicity exactly ℓ at x and no other singularities. (iv) One can combine the main result of [GL] with Theorem 1.4 of [FKP] to deduce, in certain ranges of g, ℓ, that σ∗L(−ℓE) is not only very ample but even embeds eS as an arithmetically Cohen-Macaulay surface. This is the case e.g. for ℓ = 1, g ≥ 5 and for ℓ = 2 and g ≥ 9. In the next section we are going to apply Theorem 10 in the case ℓ = 2. 5. Proof of Theorem 4 Let x ∈ S be the singular point of f (C), let σ : eS → S be the blow-up at x and E ⊂ eS the exceptional curve. Then C ∈ σ∗L(−2E) and Theorem 10 implies that σ∗L(−2E) is very ample, thus C is a hyperplane section of eS ⊂ Pg−2 embedded by σ∗L(−2E). Let T := f ∗(x) = P + Q = C · E. Then ωC (−T ) = (C + E) · C − C · E = C · C Therefore ΦωC ,ωC(−T ) is not surjective, by Theorem 2. Now we use Theorem 1.4 of [FKP]. Since g ≥ 9 we have, in the notation of [FKP]: ρsing(g, 1, 4, g − 1) = ρ(g − 1, 1, 4) + 1 < 0 and ρsing(g, 2, 6, g − 1) = ρ(g − 1, 2, 6) + 1 < 0 Therefore by [FKP], Theorem 1.4, we have W 1 6 (C), thus Cliff(C) ≥ 3. We can then apply Theorem 8 to deduce that ΦωC (−T ) is not surjective either. (cid:3) 4 (C) = ∅ = W 2 [BF] [B] [BM] [FKP] [GL] [Ke] [L] [L89] [S] [V] [W] References Ballico E., Fontanari C.: Gaussian maps, the Zak map and projective extensions of singular varieties. Result. Math. 44 (2003), 29-34. Bauer I.: Inner projections of algebraic surfaces: a finiteness result. J. Reine Angew. Math. 460 (1995), 1-13. Beauville A., Merindol J.Y.: Sections hyperplanes des surfaces K3. Duke Math. J. 55 (1987), 873-878. Flamini F., Knutsen A.L., Pacienza G.: Singular curves on a K3 surface and linear series on their normalizations. International J. of Math. 18 (2007), 671-693. Green M., Lazarsfeld R: On the projective normality of complete linear series on an algebraic curve, Inventiones Math. 83 (1986), 73-90. Kemeny M.: The moduli of singular curves on K3 surfaces. J. de Math. Pures et appliqu´ees 104 (2015), 882-920. L'vovsky S.: Extensions of projective varieties and deformations. I, II. Michigan Math. J., 39 (1992), 41–51, 65–70. Lazarsfeld R.: A sampling of vector bundle techniques in the study of linear series. In Riemann Surfaces - Proceedings of the College on Riemann Surfaces, ICTP Trieste, 1987. World Scientific (1989). Sernesi E.: The Wahl map of one-nodal curves on K3 surfaces. To appear on Contem- porary Mathematics. arXiv:1701.04801. Voisin C.: Segre classes of tautological bundles on Hilbert schemes of surfaces. arXiv:1708.06325. Wahl J.: The jacobian algebra of a graded Gorenstein singularity. Duke Math. J. 55 (1987), 843-872. NON-SURJECTIVE GAUSSIAN MAPS FOR SINGULAR CURVES ON K3 SURFACES 9 Dipartimento di Matematica, Universit`a di Trento, Via Sommarive 14, 38123 Trento, Italy. E-mail address: [email protected] Dipartimento di Matematica e Fisica, Universit`a Roma Tre, L.go S.L. Murialdo 1, 00146 Roma, Italy. E-mail address: [email protected] 

1501.04304

1

1501

Ineffective descent of genus one curves

Raynaud proved in 1968 that etale descent of genus one curves is not effective in general. In this paper we provide an alternative, simplified construction of this phenomenon. Our counterexample is fully explicit.

math.AG

math

Ineffective descent of genus one curves Wouter Zomervrucht April 22, 2018 Abstract. Raynaud proved in 1968 that étale descent of genus one curves is not effective in general. In this paper we provide an alternative, simplified construction of this phenomenon. Our counterexample is fully explicit. 1. Introduction Let U = {Ui : i ∈ I} be a cover of a scheme S, in some topology. A descent datum of schemes relative to U consists of schemes Xi over Ui for all i ∈ I and isomorphisms φji : Xi ×Ui Uij → Xj ×Uj Uij over Uij for all i, j ∈ I, satisfying the cocycle condition φki = φkjφji on Uijk. In favorable situations the descent datum is effective, i.e. descends to a scheme X over S. For instance, if U is a Zariski open cover, descent data are better known as gluing data and always effective. In larger topologies, such as the étale topology, ineffective descent data occur. An example can be found in [6, 03FN]. A natural next step is to determine classes of schemes for which étale descent is effective. In this paper we consider the case of genus one curves. First, let us once and for all fix our notion of a (relative) curve. Definition 1.1. Let S be a scheme. A curve of genus g over S is a proper smooth scheme X/S of relative dimension one, whose geometric fibers are connected curves of genus g. For genus one curves, the result is negative. Theorem 1.2. There exist ineffective étale descent data of genus one curves. This theorem is due to Raynaud [5, XIII 3.2]. The aim of the current paper is to provide a simplified counterexample. In the case of genus g 6= 1 curves, it is well-known that étale, or even fpqc, descent is effective. Indeed, outside genus one the canonical bundle (or its dual) is ample, and the descent comes from descent of quasi-coherent sheaves. See e.g. [7, 4.39] for more details. On genus one curves the canonical bundle is fiberwise trivial and the argument does not apply. Note however that for elliptic curves the zero section provides an ample line bundle, and fpqc descent is again effective. Theorem 1.2 can also be interpreted as follows. Let F be the fibered category over Sch that assigns to a scheme S the groupoid F (S) of genus one curves over S. Then F is not an étale stack. Instead one works with the fibered category M1 where M1(S) is the groupoid of proper smooth algebraic spaces over S whose geometric fibers are genus one curves. Now M1 is an étale (even fppf) stack; it is the fppf stackification of F . 1 Organization. The next section reduces the proof of theorem 1.2 to one concerning torsors under elliptic curves. Section 3 contains the actual construction. In section 4 we compare it with the original counterexample by Raynaud. Acknowledgements. The research in this paper is part of the author's master's thesis. I would like to thank Lenny Taelman and Bas Edixhoven for their valuable contributions. 2. Torsors Attached to a genus one curve X/S is its Jacobian E = Jac(X/S). It is an elliptic curve over S, endowed with a natural action on X that makes X into an étale E-torsor. Now let U be an étale cover of S. A descent datum of genus one curves relative to U descends to a (not necessarily representable) sheaf of sets X on S. Also, by functoriality of Jac we obtain a descent datum of elliptic curves relative to U. The latter descends to an elliptic curve E/S. Again the natural action of E on X makes X into an étale E-torsor. Conversely, any étale E-torsor gives rise to a descent datum of genus one curves relative to some étale cover. So the problem at hand is really to find a non-representable torsor under some elliptic curve. The following theorem from [5, XIII 2.6] will be useful. Theorem 2.1. Let S be a local scheme and E/S an elliptic curve. • If S is normal, an étale E-torsor is representable if and only if it has finite order in H1(S, E). • If S is regular, all étale E-torsors are representable. Here and further on, H1(S, E) always denotes sheaf cohomology on (Sch/S)et. Recall that normal means all local rings of S are integrally closed domains. 3. Construction We shall now construct a noetherian normal local scheme S, an elliptic curve E/S, and a class τ ∈ H1(S, E) of infinite order. Necessarily S is irregular, so not of dimension 0 or 1. Let k be a field of characteristic not 2. Let E/k be an elliptic curve with a given embedding in P2. Let C ⊂ A3 be the affine cone over E. Let s ∈ C be the top of the cone, and S the localization of C at s. It is normal by Serre's criterion [3, 23.8] since S is a complete intersection with its singularity in codimension 2. Lemma 3.1. There exists a finite étale cover π : S′ → S of degree 2 whose fiber over s consists of two distinct k-rational points s0, s1, such that E(S′ \ {s1}) = E(S′ \ {s0}) = E(k). Proof. Let E ⊂ P2 be given by some cubic f ∈ k[x, y, z]. The ring of S is the integral domain R = k[x, y, z](x,y,z)/( f ). Since 2 is invertible in k, a degree 2 finite étale cover of S may be constructed by adjoining to R the square root of a unit u ∈ R×. Set R′ = R[t]/(t2 − u) and S′ = Spec R′. Then π : S′ → S is split above s if u(s) ∈ κ(s) is a square. We choose u = 1 + x. For the second part it suffices to prove that all k-morphisms α : S′ \ {s1} → E are constant. Take a point (a : b : c) ∈ E(k). Let L ⊂ C be the corresponding ray, and η ∈ S the generic point of L. As long as a is non-zero, u is not a square at η. Then the fiber of π at η is a single point η′ with rational function field. Therefore, α must send η′ to some closed point p ∈ E. By continuity we have α(s0) = p as well. After passage to an algebraic closure of k, the points η′ as above lie dense in S′ \ {s1}. So α maps a dense subset of S′ \ {s1} to p. Since p is closed, α is constant. (cid:4) 2 Remark 3.2. The preceding lemma is still true in characteristic 2, where one may construct a suitable cover by means of an Artin–Schreier extension. Therefore, the restriction on k is not necessary; it is imposed only to simplify the exposition. ◭ Write U = S \ {s}, U0 = S′ \ {s1}, U1 = S′ \ {s0}, and U01 = U0 ∩ U1. Then U = {U0, U1} is an open cover of S′. The associated first Cech cohomology is given by the Mayer–Vietoris exact sequence 0 E(S′) E(U0) × E(U1) E(U01) H1(U, E) 0. (3.3) By construction we have E(S′) = E(U0) = E(U1) = E(k). So (3.3) reduces to 0 E(k) E(U01) H1(U, E) 0. The Galois group G = Aut(S′/S) acts on this sequence: the involution σ ∈ G acts on E(k) by inversion, on E(U01) by a 7→ −σ∗a, and on H1(U, E) by [X] 7→ [σ−1X]. In fact this action comes from the natural G-action on (3.3). Taking anti-invariants yields 0 E(k) E(U) H1(U, E)−σ where for any G-module M we denote by M−σ = {m ∈ M : σm = −m} its subgroup of anti-invariants. Let E = S′ ⊗Aut(S′/S) ES be the −1-twist of ES along π. In other words, E is the quotient of S′ × ES by Aut(S′/S), where σ acts as (x, a) 7→ (σx, −a). Then E is an elliptic curve over S with a canonical S′-isomorphism ES′ ∼= ES′. Lemma 3.4. There is a natural map H1(S′, E)−σ → H1(S, E) whose kernel is 2-torsion. Proof. Let A = π∗ES′ be the Weil restriction of ES′ to S. By [1, VIII 5.6] the pushforward map π∗ : H1(S′, E) → H1(S, A) is an isomorphism. It is also equivariant for the natural G-action on H1(S, A), so π∗ restricts to an isomorphism H1(S′, E)−σ → H1(S, A)−σ. We have E = A−σ, or more precisely E(T) = A(T)−σ for all schemes T/S. Consider the map id − σ : A → E and the inclusion E → A. The induced composition H1(S, A) H1(S, E) H1(S, A). is again id − σ. Restricted to H1(S, A)−σ this is simply multiplication by 2. Hence the kernel of H1(S, A)−σ → H1(S, E) is 2-torsion. (cid:4) The map U → C \ {s} → E is an element of E(U), no multiple of which is constant. Via the injection E(U)/E(k) → H1(S′, E)−σ and lemma 3.4 we obtain a non-torsion class τ ∈ H1(S, E). The corresponding E-torsor on S is not representable by theorem 2.1. As we have explained in section 2, this proves theorem 1.2. 4. Raynaud's approach In this section we briefly compare our construction with that by Raynaud in [5, XIII 3.2]. We sketch his approach. Let R be a discrete valuation ring in which 2 is invertible. (As before, the characteristic condition is imposed only to simplify the exposition.) Let E/R be an elliptic curve, and let V ⊂ E be the complement of the zero section. 3 Lemma 4.1. There exist a normal scheme Z over R and an R-morphism f : V → Z that is an isomor- phism on the generic fiber and constant on the special fiber. Proof. (This proof is different from Raynaud's in [5, XIII 3.2 b].) Embedding E in the projective plane, V is isomorphic to the spectrum of R[x, y]/(y2 − x3 − ax2 − bx − c) for suitable a, b, c ∈ R. Let t ∈ R be a uniformizer. Let Z be the spectrum of R[u, v]/(v2 − u3 − t2au2 − t4bu − t6c) and define f : V → Z on rings by u 7→ t2x, v 7→ t3y. Then Z is regular in codimension 1, hence normal [3, 23.8]. On the generic fiber, t is a unit so f is an isomorphism. On the special fiber f is constant with image (0, 0). (cid:4) Let s ∈ Z be the image of f on the special fiber. Let S be the localization of Z at s. It is normal by construction. We may choose R such that it admits a connected finite étale cover Spec R′ → Spec R of degree 2 that is split on the special fiber. Set S′ = SR′ , then π : S′ → S is finite étale of degree 2 as well. Let s0, s1 be the lifts of s to S′. Write U = S \ {s}, U0 = S′ \ {s1}, U1 = S′ \ {s0}, and U01 = U0 ∩ U1. Lemma 4.2. E(U0) = E(U1) = E(R′). Proof. Throughout the proof, primes indicate base change along R → R′, e.g. E′ = ER′ . By symmetry, it suffices to prove that all R′-morphisms α : U0 → E′ factor over Spec R′. Note that α extends to an R′-morphism W → E′ for some open W ⊆ Z′ containing U0. Consider the R′-rational map h : E′ V′ Z′ W E′. On the generic fiber, h is a rational map of elliptic curves over a field, hence extends to a morphism. Since E′ is the Néron model of its generic fiber [2, 1.2.8], h actually extends to an R′-morphism E′ → E′. Note that h is constant on the fiber over the closed point of Spec R′ under s0. Hence h comes from E(R′) by rigidity [4, 6.1] because Spec R′ is connected. (cid:4) Let U be the open cover {U0, U1} of S′. The Mayer–Vietoris sequence (3.3) now reduces to 0 E(R′) E(U01) H1(U, E) 0 and taking anti-invariants for the action of the involution σ ∈ Aut(S′/S) yields 0 E(R) E(U) H1(U, E)−σ. Let E be the −1-twist of ES along π. As in lemma 3.4 we have a map H1(S′, E)−σ → H1(S, E) with 2-torsion kernel. Let η be the generic point of Spec R. We have a map Uη → Zη → Eη by inverting fη. Since U is normal of dimension 1 and E is proper, it extends uniquely to a map U → E. This map is an element of E(U), no multiple of which comes from E(R). We obtain a non-torsion class τ ∈ H1(S, E). By theorem 2.1 it corresponds to a non-representable E-torsor on S, proving theorem 1.2. Remark 4.3. It is worthwile to observe that Raynaud proves slightly less. He constructs a non- torsion element in H1(S′, E) as above. Let A = π∗ES′ be the Weil restriction of ES′ to S. The 4 pushforward map π∗ : H1(S′, E) → H1(S, A) is an isomorphism by [1, VIII 5.6], so we find a non-torsion element γ ∈ H1(S, A). Consider the short exact sequence 0 ES A E 0 of abelian schemes over S. In the long exact sequence of cohomology, either γ maps to a non-torsion class in H1(S, E), or γ lifts to a non-torsion class in H1(S, E). This proves that there exists a non-representable torsor under either E or ES, hence theorem 1.2. However, this shorter proof does not permit us to explicitly write down a counterexample. ◭ References [1] M. Artin, A. Grothendieck, J.L. Verdier, Théorie des topos et cohomologie étale des schémas (SGA 4), Tome 2. Lecture Notes in Mathematics 270, Springer, 1972. [2] S. Bosch, W. Lütkebohmert, M. Raynaud, Néron models. Ergebnisse der Mathematik und ihrer Grenzgebiete (3rd series) 21, Springer, 1990. [3] H. Matsumura, Commutative ring theory. Cambridge Studies in Advanced Mathematics 8, Cambridge University Press, 1986. [4] D. Mumford, J. Fogarty, F. Kirwan, Geometric invariant theory. Ergebnisse der Mathematik und ihrer Grenzgebiete (2nd series) 34, 3rd ed., Springer, 1994. [5] M. Raynaud, Faisceaux amples sur les schémas en groupes et les espaces homogènes. Lecture Notes in Mathematics 119, Springer, 1970. [6] The Stacks Project. Available at http://stacks.math.columbia.edu, 2015. [7] A. Vistoli, Grothendieck topologies, fibered categories and descent theory. In: Fundamental Al- gebraic Geometry: Grothendieck's FGA Explained, pp. 1–104, Mathematical Surveys and Monographs 123, American Mathematical Society, 2005. 5 

1603.08693

4

1603

Global spectra, polytopes and stacky invariants

Given a convex polytope, we define its geometric spectrum, a stacky version of Batyrev's stringy E-functions, and we prove a stacky version of a formula of Libgober and Wood about the E-polynomial of a smooth projective variety. As an application, we get a closed formula for the variance of the geometric spectrum and a Noether's formula for two dimensional Fano polytopes (polytopes whose vertices are primitive lattice points). We also show that this geometric spectrum is equal to the algebraic spectrum of the polytope (the spectrum at infinity of a tame Laurent polynomial whose Newton polytope is the polytope alluded to). This gives an explanation and some positive answers to Hertling's conjecture about the variance of the spectrum of tame regular functions.

math.AG

math

Global spectra, polytopes and stacky invariants Antoine Douai ∗ Universit´e Cote d'Azur, CNRS, LJAD, France Parc Valrose, F-06108 Nice Cedex 2, France Email address: [email protected] December 12, 2018 Abstract Given a convex polytope, we define its geometric spectrum, a stacky version of Batyrev's stringy E-functions, and we prove a stacky version of a formula of Libgober and Wood about the E-polynomial of a smooth projective variety. As an application, we get a closed formula for the variance of the geometric spectrum and a Noether's formula for two dimensional Fano polytopes (polytopes whose vertices are primitive lattice points; a Fano polytope is not necessarily smooth). We also show that this geometric spectrum is equal to the algebraic spectrum (the spectrum at infinity of a tame Laurent polynomial whose Newton polytope is the polytope alluded to). This gives an explanation and some positive answers to Hertling's conjecture about the variance of the spectrum of tame regular functions. 1 Introduction Let X be a smooth projective variety of dimension n with Hodge numbers hp,q(X). Using Hirzebruch- Riemann-Roch theorem, Libgober and Wood show in [20] the formula d2 du2 E(X; u, 1)u=1 = n(3n − 5) 12 cn(X) + 1 6 c1(X)cn−1(X) (1) where E(X; u, v) =Pp,q(−1)p+qhp,q(X)upvq is the Hodge-Deligne polynomial of X. By duality, it follows that (−1)p+qhp,q(X)(p − )2 = cn(X) + c1(X)cn−1(X). (2) Xp,q n 2 n 12 1 6 If X is a n-dimensional projective variety with at most log-terminal singularities (we will focus in this paper on the toric case), Batyrev proves in [2] the "stringy" version of formula (1) d2 du2 Est(X; u, 1)u=1 = n(3n − 5) 12 est(X) + 1 6 c1,n st (X) (3) ∗Partially supported by the grant ANR-13-IS01-0001-01 of the Agence nationale de la recherche. Mathematics Subject Classification 32S40, 14J33, 34M35, 14C40. Key words and phrases: Mirror symmetry, toric varieties, polytopes, orbifold cohomology, spectrum of regular tame functions. 1 where Est is the stringy E-function of X, est is the stringy Euler number and c1,n version of c1(X)cn−1(X). st (X) is a stringy On the singularity theory side (the B-side), the expected mirror partners of toric varieties are the Givental-Hori-Vafa models (see [15], [18]), in general a class of Laurent polynomials. One associates to such functions their spectrum at infinity, namely a sequence α1, · · · , αµ of rational numbers, suitable logarithms of the eigenvalues of the monodromy at infinity of the function involved (see [24]; the main features are recalled in Section 5). A specification of mirror symmetry is that the spectrum at infinity of a given Givental-Hori-Vafa model is related to the degrees of the (orbifold) cohomology groups of its mirror variety (orbifold). So one can expect a formula similar to (2) involving the spectrum at infinity of any regular function: the aim of this text is to look for such a counterpart. The key observation is that the spectrum at infinity of a Laurent polynomial can be described (under a tameness condition due to Kouchnirenko [19], see Section 5) with the help of the Newton filtration of its Newton polytope. Since a polytope determines a stacky fan in the sense of [5], we are led to show a "stacky" version of formula (1). Given a Laurent polynomial f with Newton polytope P , with global Milnor number µ (the number of critical points with multiplicities) and with spectrum at infinity α1, · · · , αµ, the program is thus as follows: • to construct a stacky version of the E-polynomial, the geometric spectrum of P : we define P (z) := (z − 1)nPv∈N z−ν(v) where ν is the Newton function of the polytope P , see Specgeo Section 4. This geometric spectrum is closely related to the Ehrhart series and to the δ- vector of the polytope P , more precisely to their twisted versions studied by Stapledon [25] and Mustat¸a-Payne [21]; it is also an orbifold Poincar´e series (see Corollary 4.5), thanks to the description of the orbifold cohomology given by Borisov, Chen and Smith [5, Proposition 4.7], • to show that this geometric spectrum is equal to the (generating function of the) spectrum at infinity of f , and this is done by showing that both functions are Hilbert-Poincar´e series of isomorphic graded rings (see Corollary 6.2): this gives the identification between the spectrum at infinity and the orbifold degrees, • to show a "stacky" version of (1), namely d2 dz2 Specgeo P (z)z=1 = n(3n − 5) 12 µP + 1 6bµP where µP is the normalized volume of P (see equation (11)) and bµP is a linear combination of intersection numbers, see Theorem 7.3. At the end, given a tame Laurent polynomial f , we get in Theorem 7.5 the formula µXi=1 (αi − n 2 )2 = n 12 µP + 1 6bµP (4) where P is the Newton polytope of f and α1, · · · , αµ is the spectrum at infinity of f . In order to enlighten formula (4), assume that N = Z2 and that P is a full dimensional reflexive lattice polytope in NR. Then we have the following well-known Noether's formula 12 = µP + µP ◦ 2 (5) where P ◦ is the polar polytope of P and µP (resp. µP ◦) is the normalized volume of P (resp. P ◦). We show in Section 8 that bµP = µP ◦ if P is a Fano polytope (that is if its vertices are primitive lattice points). From formula (4), we then get µXi=1 (αi − 1)2 = 1 6 µP + 1 6 µP ◦ (6) spectrum satisfies Pµ which is a generalization of formula (5): indeed, a reflexive polytope P is Fano and its geometric i=1(αi − 1)2 = 2 (after a preprint version of this paper was written [8], I have been informed that an analogous result was proposed independently by Batyrev and Schaller in [3]). Last, notice that, because the mean value of α1, · · · , αµ is n i=1(αi − n 2 , we can use formula (4) in order 2 )2 of the spectrum at infinity of a Laurent polynomial f . to compute the variance 1 µP Pµ For instance, assume that bµP ≥ 0: it follows from equation (4) that αmax − αmin 1 µP µXi=1 (αi − )2 ≥ n 2 12 (7) where αmax (resp. αmin) denotes the maximal (resp. minimal) spectral value of f (indeed, if f is a Laurent polynomial we have αmax − αmin = n). This inequality is expected to be true for any tame regular function: this is the global version of Hertling's conjecture about the variance of the spectrum, see Section 9. For instance, formula (6) shows that this will be the case in the two dimensional case if the Newton polytope of f is Fano. This paper is organized as follows: in Section 2 we recall the basic facts on polytopes and toric varieties that we will use. In Section 3, we define the spectrum of a polytope. The geometric spectrum is defined in Section 4 and the algebraic spectrum is defined in Section 5: both are compared in Section 6. The previous results are used in Section 7 in order to get formula (4). We show Noether's formula (6) for Fano polytopes in Section 8. Last, we use our results in order to motivate (and to prove in some cases) the conjecture about the variance of the spectrum at infinity of a regular function in Section 9. This text owes much to Batyrev's work [1], [2]. The starting point was [1, Remark 3.13] and its close resemblance with Hertling's conjecture about the variance of the spectrum of an isolated singularity [16]: this link is previously alluded to in [17]. 2 Polytopes and toric varieties (framework) 2.1 Polytopes and reflexive polytopes Let N be the lattice Zn, let M be its dual lattice and let h , i be the pairing between NR = N ⊗Z R and MR = M ⊗Z R. A full dimensional lattice polytope P ⊂ NR is the convex hull of a finite set of N such that dim P = n. If P is a full dimensional lattice polytope containing the origin in its interior, there exists, for each facet (face of dimension n − 1) F of P , uF ∈ MQ such that P ⊂ {n ∈ NR, huF , ni ≤ 1} and F = P ∩ {n ∈ NR, huF , ni = 1}. This gives the hyperplane presentation P = ∩F {n ∈ NR, huF , ni ≤ 1}. 3 (8) (9) We define, for v ∈ NR, νF (v) := huF , vi and ν(v) := maxF νF (v) where the maximum is taken over the facets of P . Definition 2.1 The function ν : NR → R is the Newton function of P . Let P be a full dimensional lattice polytope in NR containing the origin. The polytope P ◦ = {m ∈ MR, hm, ni ≤ 1 for all n ∈ P } is the polar polytope of P . The vertices of P ◦ are in correspondence with the facets of P via uF vertex of P ◦ ↔ F = P ∩ {x ∈ NR, huF , xi = 1}. (10) A lattice polytope P is reflexive if it contains the origin and if P ◦ is a lattice polytope. All the polytopes considered in this paper are full dimensional lattice polytopes containing the origin in their interior Int P . For such a polytope P we define its normalized volume where the volume vol(P ) is normalized such that the volume of the unit cube is equal to 1. µP := n! vol(P ) (11) 2.2 Ehrhart polynomial and Ehrhart series Let Q be a full dimensional lattice polytope in NR. The function ℓ 7→ EhrQ(ℓ) := Card((ℓQ) ∩ N ), ℓ ∈ N, is a polynomial of degree n := dim NR. This is the Ehrhart polynomial of Q. We have Xm≥0 EhrQ(m)zm = δ0 + δ1z + · · · + δnzn (1 − z)n+1 where the δj's are non-negative integers [4, Theorem 3.12]: the generating function FQ(z) := Xm≥0 EhrQ(m)zm is the Ehrhart series of Q and the vector δ := (δ0, · · · , δn) ∈ Nn+1 (12) (13) is the δ-vector of Q. The δ-vector gives a characterization of reflexive polytopes: the polytope Q is reflexive if and only if δi = δn−i for i = 0, · · · , n, see for instance [4, Theorem 4.6]. 2.3 Toric varieties Let ∆ be a fan in NR and let ∆(i) be the set of its cones of dimension i. The rays of ∆ are its one-dimensional cones. Let X := X∆ be the toric variety of the fan ∆: X is simplicial if each cone of ∆ is generated by independent vectors of NR, complete if the support of its fan (the union of its cones) is NR. A full dimensional lattice polytope Q in MR yields a toric variety XQ, associated with the normal fan ΣQ of Q. Alternatively, if P ⊂ NR is a full dimensional lattice polytope containing the 4 origin in its interior we get a complete fan ∆P in NR by taking the cones over the proper faces of P and we will denote by X∆P the associated toric variety. Both constructions are dual, see for instance [6, Exercise 2.3.4]. Recall that a projective normal toric variety X is Fano (resp. weak Fano) if the anticanonical divisor −KX is Q-Cartier and ample (resp. nef and big). See [6, Theorem 6.1.7] for a characteriza- tion of weak Fano toric varieties. We will say that a full dimensional lattice polytope P containing the origin in its interior is Fano if its vertices are primitive lattice points of N , smooth Fano if each of its facets has exactly n vertices forming a basis of the lattice N . It should be emphasized that a Fano polytope is not necessarily smooth. Otherwise stated, all toric varieties that we will consider are complete and simplicial. 2.4 Stacky fans and orbifold cohomology Let ∆ be a complete simplicial fan and let ρ1, · · · , ρr be its rays, generated respectively by the primitive vectors v1, · · · , vr of N . Choose b1, · · · , br ∈ N whose images in NQ generate the rays ρ1, · · · , ρr: the data ∆ = (N, ∆, {bi}) is a stacky fan, see [5]. In particular, let P be a lattice polytope containing the origin such that ∆ := ∆P is simplicial: there are ai such that bi := aivi ∈ ∂P ∩ N and we will call the stacky fan ∆ = (N, ∆, {bi}) the stacky fan of P . One associates to this stacky fan a (separated) Deligne-Mumford stack X (∆), see [5, Proposition 3.2]. We will denote by H ∗ orb(X (∆))(= H 2∗ orb(X (∆), Q) its orbifold cohomology (with rational coefficients) and by A∗ orb(X (∆), Q)) its orbifold Chow ring (with rational coefficients). In this situation, we define, for a cone σ ∈ ∆, • Nσ the subgroup generated by bi, ρi ⊆ σ, • N (σ) = N/Nσ, • the fan ∆/σ in N (σ)Q: this is the set {τ = τ + (Nσ)Q, σ ⊆ τ, τ ∈ ∆}, • Box(σ) := {Pρi⊆σ λibi, λi ∈]0, 1[}. By [5, Proposition 4.7] we have H 2i orb(X (∆), Q) = ⊕σ∈∆ ⊕v∈Box(σ)∩N H 2(i−ν(v))(X∆/σ, Q) (14) where ν is the Newton function of P (see Definition 2.1). 2.5 Batyrev's stringy functions Let X∆ be a normal Q-Gorenstein toric variety and let ρ : Y → X∆ be a toric resolution defined by a refinement ∆′ of ∆, see for instance [6, Proposition 11.2.4]. The irreducible components of the exceptional divisor of ρ are in one-to-one correspondence with the primitive generators v′ 1, · · · , v′ q of the rays of ∆′(1) of Y that do not belong to ∆(1) and in the formula KY = ρ∗KX∆ + qXi=1 aiDi (15) we have ai = ϕ(v′ In our toric situation we have ai > −1 because ϕ(v′ i) > 0. i) − 1 where ϕ is the support function of the divisor KX∆, see [6, Lemma 11.4.10]. 5 The E-polynomial of a smooth variety X is defined by E(X, u, v) := nXp,q=0 (−1)p+qhp,q(X)upvq (16) where the hp,q(X)'s are the Hodge numbers of X. It is possible to extend this definition to singu- lar spaces having log-terminal singularities (and to get stringy invariants that extend topological let ρ : Y → X be a resolution of X := X∆ as above, invariants of smooth varieties) as follows: I ′ = {1, · · · , q} and put, for any subset J ⊂ I ′, DJ := ∩j∈J Dj if J 6= ∅, DJ := Y if J = ∅ and D◦ J = DJ − [j∈I ′−J Dj. The following definition is due to Batyrev [1] (we assume that the product over ∅ is 1; recall that ai > −1): Definition 2.2 Let X be a toric variety. The function Est(X, u, v) := XJ⊂I ′ E(D◦ J , u, v)Yj∈J uv − 1 (uv)aj +1 − 1 is the stringy E-function of X. The number est(X) := lim u,v→1 Est(X, u, v) is the stringy Euler number. (17) (18) The stringy E-function can be defined using motivic integrals, see [1] and [26]. By [1, Theorem 3.4], Est(X, u, v) does not depend on the resolution. In our setting, Est depends only on the variable z := uv, and we will write Est(X, z) instead of Est(X, u, v). In Section 4.2.2 we will use a modified version of the stringy E-function in order to compute the geometric spectrum of a polytope. 3 The spectrum of a polytope Let P be a full dimensional lattice polytope in NR. In this text, a spectrum SpecP of P is a priori an ordered sequence of rational numbers α1 ≤ · · · ≤ αµ that we will identify with the generating i=1 zαi. The specifications are the following (d(αi) denotes the multiplicity function SpecP (z) :=Pµ of αi in the SpecP ): • Rationality: the αi's are rational numbers, • Positivity: the αi's are non-negative numbers, • Poincar´e duality: SpecP (z) = zn SpecP (z−1), • Volume: limz→1 SpecP (z) = n! vol(P ) =: µP , • Normalisation: d(α1) = 1. In particular, SpecP is contained in [0, n] and Pµ Fano polytope in Rn, the Poincar´e polynomial Pn i=1 αi = n i=1 b2i(X∆P )zi is a spectrum of P . 2 µP . Basic example: if P is a smooth 6 4 Geometric spectrum of a polytope We define here the geometric spectrum of a polytope and we give several methods in order to com- pute it. Recall that the toric varieties considered here are assumed to be complete and simplicial. 4.1 The geometric spectrum Let P be a full dimensional lattice polytope in NR, containing the origin in its interior. Recall the Newton function ν of P of Definition 2.1. Definition 4.1 The function Specgeo P (z) := (z − 1)n Xv∈N z−ν(v) is the geometric spectrum of the polytope P . The number eP := limz→1 Specgeo Euler number of P . P (z) is the geometric It will follow from Proposition 4.3 that Specgeo βeP of non-negative rational numbers. P (z) =PeP i=1 zβi for an ordered sequence β1 ≤ · · · ≤ 4.2 Various interpretations We give three methods in order to compute Specgeo P , showing that it yields finally a spectrum of P in the sense of Section 3. The first one and the third one are inspired by the works of Mustat¸a-Payne [21] and Stapledon [25]. The second one is inspired by Batyrev's stringy E-functions. 4.2.1 First interpretation: fundamental domains Let P be a full dimensional lattice polytope in NR, containing the origin in its interior and let ∆ := ∆P be the corresponding complete fan as in Section 2.3. We assume in this section that ∆ is simplicial. We identify each vertex of P with an element bi ∈ N . If σ ∈ ∆(r) is generated by b1, · · · , br, define ✷(σ) := { qibi, qi ∈ [0, 1[, i = 1, · · · , r} rXi=1 rXi=1 and Lemma 4.2 We have Box(σ) := { qibi, qi ∈]0, 1[, i = 1, · · · , r}. Specgeo P (z) = nXr=0 (z − 1)n−r Xσ∈∆(r) Xv∈✷(σ)∩N zν(v) (19) and eP = n! vol(P ) =: µP . Proof. Let σ ∈ ∆(r). A lattice element v ∈ ◦ σ has one of the following decompositions: 7 i=1 λibi with w ∈ Box(σ) ∩ N and λi ∈ N for all i, i=1 λibi with w ∈ Boxc(σ) ∩ N − {0}, λi ≥ 0 for all i ≥ 2 and λ1 > 0 (up to • v = w +Pr • v = w +Pr • v =Pr renumbering), i=1 λibi where λi > 0 for all i where Boxc(σ) is the complement of Box(σ) in ✷(σ). We get (z − 1)r Xv∈ ◦ σ∩N z−ν(v) = Xv∈Box(σ)∩N zr−ν(v) + Xv∈Boxc(σ)∩N −{0} zr−1−ν(v) + 1 (20) because • Pλ1,··· ,λr≥0 z−ν(w)z−λ1 · · · z−λr = zr−ν(w) • Pλ1>0,λ2,··· ,λr≥0 z−ν(w)z−λ1 · · · z−λr = zr−1−ν(w) • Pλ1,··· ,λr>0 z−λ1 · · · z−λr = 1 (z−1)r (z−1)r (z−1)r (and we use the fact that ν(bi) = 1). Moreover, if w ∈ Box(σ) ∩ N , if w ∈ Boxc(σ) ∩ N − {0}, • α ∈ ν(Box(σ)) := {ν(v), v ∈ Box(σ)} if and only if r − α ∈ ν(Box(σ)), • α ∈ ν(Boxc(σ)) := {ν(v), v ∈ Boxc(σ)} if and only if r − 1 − α ∈ ν(Boxc(σ)) because qi ∈]0, 1[ if and only if 1 − qi ∈]0, 1[. We then deduce from (20) that (z − 1)n Xv∈ ◦ σ∩N z−ν(v) = (z − 1)n−r Xv∈✷(σ)∩N zν(v) (21) for any σ ∈ ∆(r). Equality (19) follows because the relative interiors of the cones of the complete fan ∆ give a partition of its support. For the assertion about the Euler number, notice that lim z→1 Specgeo P (z) = Xσ∈∆(n) Xv∈✷(σ)∩N 1 = n! vol(P ) because the normalized volume of σ ∩ {v ∈ NR, ν(v) ≤ 1} is equal to the number of lattice points in ✷(σ). ✷ Proposition 4.3 Let P be a full dimensional simplicial lattice polytope in NR containing the origin in its interior. Then Specgeo P (z) = Xσ∈∆ Xv∈Box(σ)∩N hσ(z)zν(v) where hσ(z) :=Pσ⊆τ (z − 1)n−dim τ . 8 Proof. Follows from equation (19). ✷ It turns out that hσ(z) is the Hodge-Deligne polynomial of the orbit closure V (σ) (as defined for instance in [6, page 121]) of the orbit O(σ). Because V (σ) is a toric variety, the coefficients of hσ(z) are non-negative integers (see for instance [25, Lemma 2.4] and the references therein) and we get Specgeo i=1 zβi for a sequence β1, · · · , βµP of non-negative rational numbers. We will also call this sequence the geometric spectrum of P . P (z) = PµP Corollary 4.4 The geometric spectrum of P satisfies zn Specgeo P (z−1) = Specgeo P (z). Proof. Follows from Proposition 4.3 because zn−dim σhσ(z−1) = hσ(z), see [25, Lemma 2.4], and ✷ Pi(1 − mi)bi ∈ Box(σ) if Pi mibi ∈ Box(σ). Corollary 4.5 Let P be a full dimensional simplicial lattice polytope in NR containing the origin in its interior and let ∆ = (N, ∆P , {bi}) be its stacky fan. Then Specgeo P (z) = Xα∈Q dimC H 2α orb(X (∆), C)zα. The geometric spectrum of P is the Hilbert-Poincar´e series of the graded vector space H 2∗ orb(X (∆), C). Proof. 0, · · · , n − dim σ. We thus get By formula (14), the orbifold degrees are α = j + ν(v) where v ∈ Box(σ) and j = orb(X (∆), Q)zα = Xσ∈∆ Xv∈Box(σ)∩N n−dim σXj=0 dimQ H 2j(X∆/σ, Q)zjzν(v). dimQ H 2α Xα Now, Pn−dim σ j=0 dimQ H 2j(X∆/σ, Q)zj = hσ(z) because the orbit closure V (σ) and the toric variety X∆/σ are isomorphic (see [6, Proposition 3.2.7]) and we get Xα dimQ H 2α orb(X (∆), Q)zα = Xσ∈∆ Xv∈Box(σ)∩N hσ(z)zν(v). The assertion then follows from Proposition 4.3. ✷ To sum up, the geometric spectrum of a simplicial polytope is a spectrum in the sense of Section 3. Rationality, positivity and the volume property are given by Lemma 4.2 and Proposition 4.3 and symmetry (Poincar´e duality) by Corollary 4.4. 4.2.2 Second interpretation: stacky E-function of a polytope (resolution of singular- ities) Let P be a full dimensional lattice polytope in NR, containing the origin in its interior. Let ρ : Y → X be a resolution of X := X∆P as in Section 2.5 and let ρ1, · · · , ρr be the rays of 9 Y , with primitive generators v1, · · · , vr and associated divisors D1, · · · , Dr. Put, for a subset J ⊂ I := {1, · · · , r}, DJ := ∩j∈J Dj if J 6= ∅ and DJ := Y if J = ∅ and define Est,P (z) :=XJ⊂I E(DJ , z)Yj∈J z − zνj zνj − 1 (22) where νj = ν(vj) and ν is the Newton function of P of Definition 2.1. Proposition 4.6 We have Specgeo resolution ρ. P (z) = Est,P (z). In particular, Est,P (z) does not depend on the Proof. Using the notations of Section 2.5, we have E(D◦ J , z) =PJ ′⊂J (−1)J−J ′E(DJ ′, z) and Est,P (z) =XJ⊂I E(D◦ J , z)Yj∈J z − 1 zνj − 1 as in [1, Proof of Theorem 3.7]. Let σ be a smooth cone of ∆′, the fan of Y , generated by vi1, · · · , vir ◦ σ: we have v = a1vi1 + · · · + arvir for a1, · · · , ar > 0 and ν(v) = a1ν(vi1) + · · · + arν(vir ). and v ∈ Thus Xv∈ ◦ σ∩N z−ν(v) = 1 · · · 1 . zν(vi1 ) − 1 zν(vir ) − 1 With these two observations in mind, the proof of the proposition is similar to the one of [1, Theorem 4.3]. ✷ Remark 4.7 Applying Poincar´e duality to the smooth subvarieties DJ , we get once again the symmetry relation zn Specgeo P (z) of Corollary 4.4. P (z−1) = Specgeo 4.2.3 Third interpretation: twisted δ-vector Let P be a full dimensional simplicial lattice polytope in NR containing the origin in its interior. Following [25], we define F 0 P (z) = Xm≥0 Xv∈mP ∩N zν(v)−⌈ν(v)⌉+m. This is a twisted version of the Ehrhart series FP (z) defined in Section 2.2. Proposition 4.8 We have Specgeo P (z) = (1 − z)n+1F 0 P (z). Proof. Notice first that v ∈ mP if and only if ν(v) ≤ m: this follows from the presentation (9) and the definition of the Newton function ν. We thus have F 0 P (z−1) = Xm≥0 Xν(v)≤m z−ν(v)+⌈ν(v)⌉−m = Xv∈N X⌈ν(v)⌉≤m z−ν(v)+⌈ν(v)⌉−m = 1 1 − z−1 Xv∈N z−ν(v) and this gives (z − 1)n(1 − z−1)F 0 P (z−1) = Specgeo P (z). Thus (1 − z)n+1F 0 P (z) = zn Specgeo P (z−1) = Specgeo P (z) where the last equality follows from Corollary 4.4. ✷ 10 Corollary 4.9 The part of the geometric spectrum contained in [0, 1[ is the sequence ν(v), v ∈ Int P ∩ N . In particular, the multiplicity of 0 in the geometric spectrum is equal to one. Moreover the multiplicity of 1 in Specgeo P is equal to Card(∂P ∩ N ) − n. Proof. Scrutinization of the coefficients of za for a ≤ 1 in the formula of Proposition 4.8 (see also [25, Lemma 3.13]). ✷ If P is reflexive, we have the following link between the δ-vector of P from Section 2.2 and its geometric spectrum, see also [21]: Corollary 4.10 Let P be a reflexive full dimensional simplicial lattice polytope in NR containing the origin in its interior. Then Specgeo P (z) = δ0 + δ1z + · · · + δnzn where δ = (δ0, · · · , δn) is the δ-vector of P . Proof. By (10), we have ν(v) ∈ N for all v ∈ N because P is reflexive and we get F 0 where FP (z) is the Ehrhart series of P of Section 2.2 because P (z) = FP (z) FP (z) = Xm≥0 Card(mP ∩ N )zm = Xm≥0 Xv∈mP ∩N zm. By Proposition 4.8 we thus have Specgeo formula (12). P (z) = (1 − z)n+1FP (z) and we get the assertion using ✷ 5 Algebraic spectrum of a polytope Singularity theory associates to a (tame) Laurent polynomial function f a spectrum at infinity, see [24]. We recall its definition and its main properties in Section 5.3. We can shift this notion to the Newton polytope P of f and we get in this way the algebraic spectrum of P . In order to motivate the next sections, we describe the Givental-Hori-Vafa models [15], [18] which are the expected mirror partners of toric varieties. 5.1 Tameness (Kouchnirenko's framework) We briefly recall the setting of [19]. Let f : (C∗)n → C be a Laurent polynomial, f (u) =Pa∈Zn caua where ua := ua1 n . The Newton polytope P of f is the convex hull of the multi-indices a such that ca 6= 0. We say that f is convenient if P contains the origin in its interior, nondegenerate if, for any face F of P , the system 1 · · · uan u1 ∂fF ∂u1 ∂fF ∂un = · · · = un = 0 has no solution on (C∗)n where fF (u) =Pa∈F ∩P caua, the sum being taken over the multi-indices a such that ca 6= 0. A convenient and nondegenerate Laurent polynomial f has only isolated critical points and its global Milnor number µf (the number of critical points with multiplicities) is equal to the normalized volume n! vol(P ), see [19, Th´eor`eme III]. Moreover, f is tame in the sense that the set outside which f is a locally trivial fibration is made from critical values of f , and these critical values belong to this set only because of the critical points at finite distance. 11 5.2 Givental-Hori-Vafa models and mirror symmetry Let N = Zn and let M be the dual lattice. Let ∆ be a complete and simplicial fan and let v1, · · · , vr be the primitive generators of its rays. For the sake of simplicity, we assume here that the vi's generate the lattice N . Consider the exact sequence 0 −→ Zr−n ψ −→ Zr ϕ −→ Zn −→ 0 where ϕ(ei) = vi for i = 1, · · · , r ((ei) denotes the canonical basis of Zr) and ψ describes the relations between the vi's. Applying HomZ(−−, C∗) to this exact sequence, we get 1 −→ (C∗)n −→ (C∗)r π−→ (C∗)r−n −→ 1 where and the integers ai,j satisfy Pr the toric variety X∆ is the function π(u1, · · · , ur) = (q1, · · · , qr−n) = (ua1,1 1 · · · uar,1 r , · · · , ua1,r−n 1 · · · uar,r−n r ) (23) j=1 aj,ivj = 0 for i = 1, · · · , r − n. The Givental-Hori-Vafa model of u1 + · · · + ur restricted to U := π−1(q1, · · · , qr−n) We will denote it by fX∆. If ∆ contains a smooth cone, the function fX∆ is easily described: Proposition 5.1 Assume that (v1, · · · , vn) is the canonical basis of N . Then fX∆ is the Laurent polynomial defined on (C∗)n by fX∆(u1, · · · , un) = u1 + · · · + un + rXi=n+1 qiu vi 1 · · · uvi 1 n n if vi = (vi 1, · · · , vi n) ∈ Zn for i = n + 1, · · · , r. Proof. We have vi =Pn j=1 vi jvj for i = n + 1, · · · , r and the result follows from (23). ✷ Above a convenient and nondegenerate Laurent polynomial f we make grow a differential system (see [12] and also [11] for explicit computations on weighted projective spaces) and we say that f is a mirror partner of a variety X if this differential system is isomorphic to the one associated with the (small quantum, orbifold) cohomology of X (see [10], [23]). It is expected that the Givental- Hori-Vafa model fX∆ provides a mirror partner of the toric variety X∆. If f is the mirror partner of a smooth toric variety X the following properties are in particular expected (non-exhaustive list): • the Milnor number of f is equal to the rank of the cohomology of X, • the spectrum at infinity of f (see Section 5.3 below) is equal to half of the degrees of the cohomology groups of X, • multiplication by f on its Jacobi ring yields the cup-product by c1(X) on the cohomology algebra of X. 12 In particular, we have to restrict to (weak) Fano varieties: the rank of the cohomology of X is the number of maximal cones while the Milnor number of f is the normalized volume of its Newton polytope by [19, Th´eor`eme III]. See Example 5.5 for a picture of the situation. In the singular simplicial case (i.e X is an orbifold), cohomology should be replaced by orbifold cohomology, see Section 6: in this situation, the rank of the orbifold cohomology is not a number of cones but a normalized volume, and the cohomology degrees are the orbifold degrees. 5.3 The spectrum at infinity of a tame Laurent polynomial We assume in this section that f is a convenient and nondegenerate Laurent polynomial, defined on U := (C∗)n. Let µ be its global Milnor number. We use in this section the notations of [12, 2.c]. Let G be the Fourier-Laplace transform of the Gauss-Manin system of f , let G0 be its Brieskorn lattice and let V• be the V -filtration of G at infinity, that is along θ−1 = 0. Because f is convenient and nondegenerate, G0 is a free C[θ]-module of rank µ and G = C[θ, θ−1] ⊗ G0, see [12, Remark 4.8]. From these data we get by projection a V -filtration on the µ-dimensional vector space Ωf := Ωn(U )/df ∧ Ωn−1(U ) = G0/θG0. The spectrum at infinity of f is the spectrum of this filtration, that is the (ordered) sequence α1, α2, · · · , αµ of rational numbers with the following property: the frequency of α in the sequence is equal to dimC GrV α Ωf . We will denote it by Specf and we will i=1 zαi. By [24] we have αi ≥ 0 for all i and Specf (z) = zn Specf (z−1). If f is convenient and nondegenerate, its spectrum at infinity can be computed using the Newton function of its Newton polytope: let us define the Newton filtration N• on Ωn(U ) by write Specf (z) =Pµ NαΩn(U ) = {Xc∈N acωc, ν(ωc) ≤ α for all c such that ac 6= 0} where ν(ωc) := ν(c) if ωc = uc1 and c = (c1, · · · , cn) ∈ N . This filtration induces a filtration on Ωf by projection and by [12, Corollary 4.13] the spectrum at infinity of f is equal to the spectrum of this filtration. 1 · · · ucn n du1∧···∧dun u1···un 5.4 The algebraic spectrum of a polytope We define the algebraic spectrum Specalg P of a simplicial full dimensional lattice polytope P con- taining the origin in its interior to be the spectrum at infinity of the Laurent polynomial fP (u) = Pb∈V(P ) ub where V(P ) denotes the set of the vertices of P . Notice that fP is a convenient and nondegenerate Laurent polynomial and that its Milnor number is µfP = µP : indeed, fP is conve- nient by definition because P contains the origin in its interior and it is nondegenerate because of the simpliciality assumption. The assertion about the Milnor number then follows from [19]. We will identify the algebraic spectrum with its generating function Specalg P (z) :=Pµ i=1 zαi. Proposition 5.2 Let P be a full dimensional simplicial lattice polytope in NR containing the origin in its interior. Then the part of the algebraic spectrum contained in [0, 1[ is the sequence ν(v), v ∈ Int P ∩ N where ν is the Newton function of P . In particular, the multiplicity of 0 in Specalg P is equal to one. Proof. The assertion for SpecfP follows from [12, Lemma 4.6] and [12, Example 4.17]. ✷ In the two dimensional case, we deduce the following description of the algebraic spectrum: 13 Proposition 5.3 Let P be a full dimensional lattice polytope in NR = R2 containing the origin in its interior. Then Specalg P (z) = (Card(∂P ∩ N ) − 2)z + Xv∈Int P ∩N (zν(v) + z2−ν(v)) where ν is the Newton function of P . Proof. Let fP (u) =Pb∈V(P ) ub be as above. From Proposition 5.2, and because of the symmetry property SpecfP (z) = z2 SpecfP (z−1), we get SpecfP (z) = (Card(∂P ∩ N ) − 2)z + Xv∈Int P ∩N (zν(v) + z2−ν(v)) The coefficient of z is computed using Pick's formula (see for instance [4, Theorem 2.8]) because µfP = 2 vol(P ) by [19, Th´eor`eme III]. ✷ In any dimension, we also have the following description for reflexive polytopes: Proposition 5.4 Let P be a full dimensional reflexive simplicial polytope in NR = Rn containing the origin in its interior. Then: • Specalg P (z) =Pn i=0 d(i)zi where d(i) ∈ N, • d(i) = d(n − i) for i = 0, · · · , n with d(0) = d(n) = 1, • Pn i=0 d(i) = µP Proof. Because P is reflexive, the Newton function takes integer values at the lattice points, see (10). This gives the first point because Specalg P ⊂ [0, n]. For the second one, use the symmetry and the fact that 0 is in the spectrum with multiplicity one. ✷ Example 5.5 (Hirzebruch surfaces and their Givental-Hori-Vafa models) Let m be a positive integer. The fan ∆Fm of the Hirzebruch surface Fm is the one whose rays are generated by the vectors v1 = (1, 0), v2 = (0, 1), v3 = (−1, m), v4 = (0, −1), see for instance [14]. The surface Fm is Fano if m = 1, weak Fano if m = 2. Its Givental-Hori-Vafa model is the Laurent polynomial q1 u2 + q2 um 2 u1 . fm(u1, u2) = u1 + u2 + We have 1. µf1 = 4 and Specf1(z) = 1 + 2z + z2, 2. µf2 = 4 if q2 6= 1 4 and Specf2(z) = 1 + 2z + z2, 14 3. µfm = m + 2 if m ≥ 3 and Specfm(z) = 1 + 2z + z2 + z 1 p + z 2 p + · · · + z if m = 2p and p ≥ 2, p−1 p + z2− 1 p + z2− 2 p + · · · + z2− p−1 p Specfm(z) = 1 + z + z2 + z 2 m + z 4 m + · · · + z if m = 2p + 1 and p ≥ 1. 2p m + z2− 2 m + z2− 4 m + · · · z2− 2p m Indeed, for m 6= 2 we have µfm = 2! vol(P ), where P is the Newton polytope of fm, because fm is convenient and nondegenerate for all non zero value of the parameters, see Section 5.1. For m = 2, f2 is nondegenerate if and only if q2 6= 1 4 and the previous argument applies in this case (if q2 = 1/4 the Milnor number is 2). The spectrum is given by Proposition 5.3. The function f2 is a genuine mirror partner of the surface F2, see [9], [23]. If m ≥ 3, we have µfm > 4 and the model fm has too many critical points. 6 Geometric spectrum vs algebraic spectrum We show in this section (Corollary 6.2) the equality Specalg P (z). The idea is to show that these functions are Hilbert-Poincar´e series of isomorphic graded rings (Theorem 6.1). This is achieved using the description of the orbifold Chow ring given in [5]. Notice that in the two dimensional case, the expected equality follows immediately from Corollary 4.9, Proposition 5.2 and the symmetry property. P (z) = Specgeo 6.1 An isomorphism of graded rings Let P be a full dimensional simplicial polytope containing the origin in its interior and let fP (u) = j=1 ubj be the corresponding convenient and nondegenerate Laurent polynomial as in Section 5.4. 1 , · · · , un, u−1 n ]. n if c = (c1, · · · , cn) ∈ N and K := Q[u1, u−1 1 · · · ucn Pr In what follows, we will write uc := uc1 Let AfP = K hu1 ∂fP ∂u1 , · · · , un ∂fP ∂un i be the jacobian ring of fP . Define on K the Newton filtration N• by NαK = {Xc∈N acuc ∈ K, ν(c) ≤ α for all c such that ac 6= 0} where ν is the Newton function of P . The vector space N<αK is defined similarly: replace the condition ν(c) ≤ α by ν(c) < α. This filtration induces by projection a filtration, also denoted by N• on AfP , and we get the graded ring GrN AfP = ⊕αNαAfP /N<αAfP (see equation (25) below). Theorem 6.1 There is an isomorphism of graded rings A∗ orb(X (∆)) ∼= GrN AfP where ∆ is the stacky fan associated with P as in Section 2.4. 15 Proof. We first recall the setting of [5, Section 5]. Let ∆ = (N, ∆, {bi}) be the stacky fan of P . We define its deformed group ring Q[N ]∆ as follows: • as a Q-vector space, Q[N ]∆ = ⊕c∈N Q yc where y is a formal variable, • the ring structure is given by yc1.yc2 = yc1+c2 if c1 and c2 belong to the same cone, yc1.yc2 = 0 otherwise, • the grading is defined as follows: if c =Pρi⊆σ(c) mibi then deg(yc) =P mi ∈ Q (σ(c) is the minimal cone containing c). Observe that deg(yc) = ν(c) where ν is the Newton function of P . Because ∆ is simplicial, we have, by [5, Theorem 1.1], an isomorphism of Q-graded rings where σ(v) is the minimal cone containing v, ✷(σ) := {Pρi⊂σ λibi, λi ∈ [0, 1[} and ✷(∆) is the union of ✷(σ) for all n-dimensional cones σ ∈ ∆. −→ ⊕v∈✷(∆)A∗(X (∆/σ(v)))[deg(yv)] (24) Q[N ]∆ i=1hm, biiybi, m ∈ M i hPr On the other side, we have AfP = because ui the ring ∂fP ∂ui = Pr j=1hm, bjiubj , m ∈ M i K K g hPr hPr AfP = j=1hm, bjiubj , m ∈ M i j=1he∗ i , bjiubj where (e∗ i ) is the dual basis of the canonical basis of N . Define where K g = K as a vector space and the multiplication on K g is defined as follows: uc1.uc2 =(cid:26) uc1+c2 0 if c1 and c2 belong to the same cone σ of ∆P , otherwise Define a grading on K g by deg(uc) = ν(c). Because ν(c1 + c2) = ν(c1) + ν(c2) if and only if c1 and c2 belong to the same cone σ of ∆P and because ν(bj) = 1, the ring AfP is graded. Moreover, because uv.uw ∈(cid:26) Nα+βK if uv ∈ NαK and uw ∈ NβK, N<α+βK if v and w do not belong to the same cone of ∆P , the graded ring AfP is isomorphic to GrN AfP . Last, the rings hPr isomorphic, and the theorem follows now from the isomorphism (24). Q[N ]∆ i=1hm,biiybi , m∈M i (25) and AfP are ✷ Corollary 6.2 Assume that P is a simplicial polytope containing the origin in its interior. Then Specalg P (z) = Specgeo P (z). Proof. As explained in Section 5.3, Specalg P (z) is the Hilbert-Poincar´e series of the graded vector space GrN AfP . Now, Theorem 6.1 shows that the latter coincide with the Hilbert-Poincar´e series of A∗ orb(X (∆)) and we get the assertion by Corollary 4.5. ✷ This corollary can also be shown using [19, Th´eor`eme 2.8]. 16 6.2 Application to weighted projective spaces Let (λ0, · · · , λn) ∈ (N∗)n+1 be such that gcd(λ0, · · · , λn) = 1 and let X be the weighted projective space P(λ0, · · · , λn). We assume from now on that λ0 = 1. The (stacky) fan of X is given by: 1. the lattice N = Zn, 2. the morphism β : Zn+1 → N , which sends e′ i to ei for i = 1, · · · , n and e′ 0 to −λ1e1−· · ·−λnen, where (e′ i) is the canonical basis of Zn+1 and (ei) is the canonical basis of N , 3. the fan ∆, which is the complete and simplicial fan whose rays are generated by bi := β(e′ i), i = 0, · · · , n. In this situation, we will call the convex hull P of b0, · · · , bn the polytope of X. We have µP = 1 + λ1 + · · · + λn and µP ◦ = (1+λ1+···+λn)n . λ1···λn Let F :=(cid:26) ℓ λi 0 ≤ ℓ ≤ λi − 1, 0 ≤ i ≤ n(cid:27) and let f1, · · · , fk be the elements of F arranged by increasing order. Define Sfi := {j λjfi ∈ Z} ⊂ {0, · · · , n} and di := Card Sfi and let c0, c1, · · · , cµ−1 be the sequence f1, · · · , f1 , f2, · · · , f2 , · · · , fk, · · · , fk {z d1 } {z d2 } {z dk } arranged by increasing order. By [13, Theorem 1], the spectrum at infinity of f is the sequence α0, α1, · · · , αµ−1 where αk := k − µck for k = 0, · · · , µ − 1. Notice that this spectrum is integral if and only if the polytope P of X is reflexive. We have fP (u1, · · · , un) = u1 + · · · + un + 1 uλ1 1 · · · uλn n . By Proposition 5.1, this is also the Givental-Hori-Vafa model of X. A mirror theorem is shown in [11]. Example 6.3 We test Corollary 6.2 on weighted projective spaces. The computation of the geo- metric spectrum is done using Proposition 4.6. The rays are numbered clockwise. 1. Let a be a positive integer and let P be the convex hull of (1, 0), (−1, −a) and (0, 1): this is the polytope of P(1, 1, a). We consider the resolution obtained by adding the ray generated by (0, −1). Using the notations of Proposition 4.6, we have ν1 = 1, ν2 = 2 a , ν3 = 1 and ν4 = 1 and we get Specgeo P (z) = E(Fa, z) + E(P1, z) z − z2/a z2/a − 1 = 1 + 2z + z2 + (1 + z)( z − 1 z2/a − 1 − 1) = 1 + z + z2 + z2/a + z4/a + · · · + z2(a−1)/a = Specalg P (z) where Fa is the Hirzebruch surface. 17 2. Let P be the convex hull of (1, 0), (−2, −5) and (0, 1): P is the polytope of P(1, 2, 5). We con- sider the resolution obtained by adding the rays generated by (0, −1), (−1, −3) and (−1, −2). Using the notations of Proposition 4.6, we have ν1 = 1, ν2 = 3 5 , ν4 = 1, ν5 = 1 and ν6 = 1 and we get 5 , ν3 = 4 Specgeo P (z) = z2 + 4z + 1 + (z + 1) z − z3/5 z3/5 − 1 + (z + 1) z − z4/5 z4/5 − 1 + z − z3/5 z3/5 − 1 . z − z4/5 z4/5 − 1 = z2 + 2z + 1 + z3/5 + z4/5 + z6/5 + z7/5 = Specalg P (z). 3. Let ℓ ∈ N∗ and P be the convex hull of (1, 0), (−ℓ, −ℓ) and (0, 1): P is the polytope of P(1, ℓ, ℓ). The variety X∆P is P2, generated by the rays v0 = (−1, −1), v1 = (1, 0) and v2 = (0, 1). Because ν(v0) = 1 ℓ , ν(v1) = 1 and ν(v2) = 1, we get Specgeo P (z) = E(P2, z) + E(P1, z) z − z1/ℓ z1/ℓ − 1 = 1 + z + z2 + (1 + z) z − z1/ℓ z1/ℓ − 1 = z2 + z + 1 + z1/ℓ + · · · + z(ℓ−1)/ℓ + z1+1/ℓ + · · · + z1+(ℓ−1)/ℓ = Specalg P (z). 4. Let P be the convex hull of (−2, −2, −2), (1, 0, 0), (0, 1, 0) and (0, 0, 1): P is the polytope of P(1, 2, 2, 2). We have X∆P = P3, generated by the rays v0 = (−1, −1, −1), v1 = (1, 0, 0), v2 = (0, 1, 0) and v3 = (0, 0, 1). Because ν(v0) = 1 2 , we get Specgeo P (z) = E(P3, z) + E(P2, z) z − z1/2 z1/2 − 1 = z3 + z2 + z + 1 + z1/2 + z3/2 + z5/2 = Specalg P (z). 7 A formula for the variance of the spectra We now prove formula (4) of the introduction. We first show it for the geometric spectrum of a full dimensional simplicial lattice polytope. 7.1 Libgober-Wood's formula for the spectra In order to get first a stacky version of the Libgober-Wood formula (1), we give the following definition, inspired by Batyrev's stringy number c1,n−1 (X), see [2, Definition 3.1]: st Definition 7.1 Let P be a full dimensional simplicial lattice polytope in N containing the origin in its interior and let ρ : Y → X be a resolution of X := X∆P . We define the rational number bµP := c1(Y )cn−1(Y ) + XJ⊂I, J6=∅ − XJ⊂I, J6=∅ (Xj∈J c1(DJ )cn−J−1(DJ )Yj∈J 1 − νj νj (26) νj)cn−J(DJ )Yj∈J 1 − νj νj where the notations in the right hand term are the ones of Section 4.2.2 (convention: cr(DJ ) = 0 if r < 0). 18 Remark 7.2 We have bµP = c1(Y )cn−1(Y ) if νi = 1 for all i (crepant resolutions) and bµP = c1(X)cn−1(X) if X is smooth. Theorem 7.3 Let P be a full dimensional simplicial lattice polytope in N containing the origin in its interior and let Specgeo i=1 zβi be its geometric spectrum. Then P (z) =PeP ePXi=1 (βi − n 2 )2 = n 12 µP + 1 6bµP (27) z−zνj zνj −1 , see formula (22). (28) where bµP is defined by formula (26) and µP := n! vol(P ). Proof. Recall the stacky E-function Est,P (z) := PJ⊂I E(DJ , z)Qj∈J Then we have E′′ st,P (1) = n(3n − 5) 12 eP + 1 6bµP where eP is the geometric Euler number of P of Definition 4.1. The proof of this formula is a straightforward computation and is similar to the one of [2, Theorem 3.8]: if V is a smooth variety of dimension n, we have the Libgober-Wood formula E′′(V, 1) = n(3n − 5) 12 cn(V ) + 1 6 c1(V )cn−1(V ) (29) where E is the E-polynomial of V , see [20, Proposition 2.3]; in order to get (28), apply this formula to the components E(DJ , z) of Est,P (z) and use the equalities E(DJ , 1) = cn−J(DJ ), d dz (E(DJ , z))z=1 = n − J 2 cn−J(DJ ) (the first one follows from the fact that the value at z = 1 of the Poincar´e polynomial is the Euler characteristic and we get the second one using Poincar´e duality for DJ ) and d dz ( z − zν zν − 1 )z=1 = 1 − ν 2ν , d2 dz2 ( z − zν zν − 1 )z=1 = (ν − 1)(ν + 1) 6ν if ν is a positive rational number. By Proposition 4.6, we have Specgeo P (z) = Est,P (z) and we get d2 dz2 (Specgeo P (z))z=1 = n(3n − 5) 12 eP + 1 6bµP . (30) Because the geometric spectrum is symmetric with respect to n dz (Specgeo d 2 eP and we deduce that P (z))z=1 = n 2 (see Corollary 4.4), we have d2 dz2 (Specgeo P (z))z=1 = ePXi=1 (βi − n 2 )2 + n(n − 2) 4 eP . Now, formulas (30) and (31) give equality (27) because eP = µP by Lemma 4.2. Corollary 7.4 The number bµP does not depend on the resolution ρ. 19 (31) ✷ ✷ The version for singularities is straightforward: Theorem 7.5 Let f be a convenient and nondegenerate Laurent polynomial with global Milnor i=1 zαi. Let P be its Newton polytope (assumed number µ and spectrum at infinity Specf (z) = Pµ to be simplicial). Then 6bµP where bµP is defined by formula (26) and µP := n! vol(P ) = µ. µP + (αi − n 2 )2 = n 12 1 µXi=1 Proof. By [22] we have Specalg and Corollary 6.2. Last, because f is convenient and nondegenerate we have µ = µP by [19]. P (z) = Specf (z) and the assertion thus follows from Theorem 7.3 ✷ (32) 7.2 The number bµP in the two dimensional case In this section we give an explicit formula for bµP in the two dimensional case. Let P be a full dimensional polytope in NR = R2 containing the origin, and let ρ : Y → X be a resolution of X := X∆P as in Section 4.2.2. We assume that the primitive generators v1, · · · , vr of the rays of Y are numbered clockwise and we consider indices as integers modulo r so that νr+1 := ν1 (recall that νi := ν(vi) where ν is the Newton function of P ). Proposition 7.6 We have 1(Y ) − 2r + bµP = c2 rXi=1 ( νi νi+1 + νi+1 νi ) = ( rXi=1 νiDi)( rXj=1 1 νj Dj). Proof. By definition, we have bµP := c1(Y )c1(Y ) + XJ⊂I, J6=∅ c1(DJ )c1−J(DJ )Yj∈J 1 − νj νj (33) and thus It follows that − XJ⊂I, J6=∅ rXi=1 rXi=1 1 νi ( 1(Y ) + 2 bµP = c2 bµP − c2 (Xj∈J νj)c2−J(DJ )Yj∈J rXi=1 (νi + νi+1) − 1 − νj νj (1 − νi) (1 − νi+1) νi νi+1 . (1 − νi)2 νi 1(Y ) = + νi − 1 νi+1 − νi+1 + νi νi+1 + νi+1 νi − 2) = −2r + rXi=1 ( νi νi+1 + νi+1 νi ) 20 and this gives the first equality. For the second one, notice that νiDi)( rXj=1 1 νj Dj) = rXi=1 (D2 i + νi νi+1 DiDi+1 + νi νi−1 DiDi−1) ( rXi=1 rXi=1 = (D2 i + νi νi+1 + νi+1 νi ) = c2 1(Y ) − 2r + rXi=1 ( νi νi+1 + νi+1 νi ) = bµP . ✷ Corollary 7.7 Let P be a full dimensional lattice polytope in NR = R2 containing the origin in its interior and let Specgeo i=1 zβi be its geometric spectrum. Then P (z) =PeP µPXi=1 (βi − 1)2 = µP 6 + bµP 6 (34) 1(Y ) for any resolution ρ : Y → X∆P . where bµP ≥ c2 Proof. The equality follows from Theorem 7.3. By Proposition 7.6 we have bµP ≥ c2 ν ≥ 2 for all real positive numbers ν. ν + 1 1(Y ) because ✷ 7.3 Examples 7.3.1 Weighted projective spaces We test Theorem 7.5 on the weighted projective spaces considered in Example 6.3. Let P be the i=0 ubi where the i=1 zαi the spectrum at infinity of f and we polytope of the weighted projective space X = P(1, λ1, · · · , λn) and let f (u) =Pn bi's are defined as in Section 6.2. We will denote by Pµ will write V (α) :=Pµ i=1(αi − n 2 )2. • The polytope P is Fano (see Section 2.3): X µ V (α) µn/12 P(1, 1, a) P(1, 2, 5) a + 2 (2a2 + 6a + 4)/6a (a + 2)/6 8 12/5 4/3 bµP a (a+2)2 32/5 For P(1, 1, a), the polytope P is the convex hull of (1, 0), (0, 1) and (−1, −a) and we consider the resolution obtained by adding the ray generated by (0, −1). We use Proposition 7.6 in a , ν3 = 1, ν4 = 1 and c2 For P(1, 2, 5), the polytope P is the convex hull of (1, 0), (0, 1) and (−2, −5) and we consider the resolution obtained by adding the rays generated by (0, −1), (−1, −3) and (−1, −2). We 5 , ν4 = 1, ν5 = 1, order to compute bµP , with ν1 = 1, ν2 = 2 use Proposition 7.6 in order to compute bµP , with ν1 = 1, ν2 = 3 Notice that in these examples we have bµP = µP ◦ where µP ◦ is the volume of the polar polytope: this is not a coincidence, see Section 8 below. ν6 = 1 and c2 5 , ν3 = 4 1(Y ) = 6. 1(Y ) = 8. 21 • The polytope P is not Fano: X µ V (α) µn/12 P(1, ℓ, ℓ) 1 + 2ℓ P(1, 2, 2, 2) 7 2 + (ℓ−1)(2ℓ−1) 3ℓ 7 (2ℓ + 1)/6 7/4 9 + 2 (ℓ−1)2 bµP ℓ 63/2 For P(1, ℓ, ℓ), ℓ ≥ 2, the polytope P is the convex hull of (1, 0), (0, 1) and (−ℓ, −ℓ). Formula (26) gives bµP = c1(P2)c1(P2) + c1(P1)(ℓ − 1) − 1 ℓ c1(P1)(ℓ − 1). Notice that bµP 6= µP ◦. For P(1, 2, 2, 2), P is the convex hull of (−2, −2, −2), (1, 0, 0), (0, 1, 0) and (0, 0, 1). Formula (26) gives 7.3.2 Miscellaneous bµP = c1(P3)c2(P3) + c1(P2)c1(P2) − 1 2 c2(P2). In order to complete the panaorama, let us now consider somewhat different situations: • let P1,2,2 be the polytope with vertices b1 = (1, 0), b2 = (0, 2) and b3 = (−2, −2). • Let Pℓ,ℓ,ℓ be the polytope with vertices b1 = (ℓ, 0), b2 = (0, ℓ) and b3 = (−ℓ, −ℓ) where ℓ is a positive integer. We have the following table: This agrees with formula (32) µ 8 3ℓ2 V (α) µn/12 3 (ℓ2 + 3)/2 4/3 ℓ2/2 bµP 10 9 P1,2,2 Pℓ,ℓ,ℓ 8 A Noether's formula for two dimensional Fano polytopes In this section, we still focus on the two dimensional case: P is full dimensional polytope in NR = R2. Remark 7.2. The anticanonical divisor of Y is nef and c2 Assume that P is a reflexive polytope. Then X∆P has a crepant resolution and bµP = c2 finally bµP = µP ◦. By Corollary 4.10, the geometric spectrum of P satisfies Pµ 1(Y ) by 1(Y ) = µP ◦ by [6, Theorem 13.4.3], so that i=1(βi − 1)2 = 2 and we thus get from Corollary 7.7 the well-known formula 12 = µP + µP ◦ (35) for a reflexive polytope P . We have the following generalization of equation (35) for the Fano polytopes defined in Section 2.3 (a reflexive polytope is Fano): 22 Theorem 8.1 Assume that P is a Fano polytope in R2 and let Specgeo geometric spectrum. Then P (z) = PµP i=1 zβi be its µPXi=1 (βi − 1)2 = µP 6 + µP ◦ 6 (36) where P ◦ is the polar polytope of P . Proof. Notice first that, because of the Fano assumption, the support function of the Q-Cartier i=1 νiDi since ρ∗(−KX ) divisor KX is equal to the Newton function of P and thus ρ∗(−KX) = Pr and −KX have the same support function. We shall show that bµP = ρ∗(−KX )ρ∗(−KX) (37) Because (ρ∗(−KX ))2 = (−KX )2 = µP ◦ (for the first equality see [6, Lemma 13.4.2] and for the second one see the Q-Cartier version of [6, Theorem 13.4.3]), equation (36) will follow from Theorem 7.3. By Proposition 7.6, we have bµP = and, as noticed above, rXi=1 (D2 i + νi νi+1 + νi+1 νi ) = rXi=1 D2 i + rXi=1 ( νi−1 νi + νi+1 νi ) ρ∗(−KX)ρ∗(−KX ) = ( rXi=1 νiDi)2 = rXi=1 ν2 i D2 i + rXi=1 (νiνi+1 + νiνi−1) so (37) reads Notice the following: rXi=1 (ν2 i − 1)D2 i = rXi=1 (ν2 i − 1)(− νi−1 νi − νi+1 νi ). (38) • if vi−1, vi and vi+1 are primitive generators of rays of Y inside the same cone of the fan of X, we have ν(vi−1 + vi+1) = ( νi−1 νi + νi+1 νi )ν(vi) because ν(vi−1) = νi−1, ν(vi) = νi and ν(vi+1) = νi+1 and the Newton function is linear on each cone of the fan of X. Because Y is smooth and complete, it follows that vi−1 + vi+1 = ( νi−1 νi + νi+1 νi )vi and we get D2 i = − νi−1 νi − νi+1 νi . • Otherwise, νi = 1 due to the Fano condition. Equation (38), hence equation (37), follows from these two observations. ✷ 23 Example 8.2 Let P be the convex hull of (1, 0), (−λ1, −λ2) and (0, 1) where λ1 and λ2 are rela- tively prime integers. Then µPXi=1 (βi − 1)2 = 1 6 [(1 + λ1 + λ2) + (1 + λ1 + λ2)2 λ1λ2 ] Notice that P is the polytope of the weighted projective space P(1, λ1, λ2). Remark 8.3 Theorem 8.1 is not true if we drop the Fano assumption: for instance, if P is the 4 . Also, it follows from Proposition 7.6 that polytope of P(1, 2, 2), we have bµP = 10 and µP ◦ = 25 bµℓP = bµP if ℓ is an integer greater or equal than one, and thus bµP can't be seen as a volume in general. Finally, we get the following statement for singularities: Corollary 8.4 Let f be a nondegenerate and convenient Laurent polynomial on (C∗)2 with spec- trum at infinity α1, · · · , αµ. Assume that the Newton polytope P of f is Fano. Then µXi=1 (αi − 1)2 = µP 6 + µP ◦ 6 where P ◦ is the polar polytope of P . In particular, 1 i=1(αi − 1)2 ≥ 1 6 . µPµ (39) ✷ 9 Hertling's conjecture for regular functions From Theorem 7.5 we get: Proposition 9.1 Let f be a convenient and nondegenerate Laurent polynomial and let P be its Newton polytope (assumed to be simplicial). Assume that bµP ≥ 0. Then (40) 1 µ µXi=1 (αi − )2 ≥ n 2 n 12 were µ is the global Milnor number of f and Specf (z) =Pµ Corollary 7.7, examples of Section 7.3 and Corollary 8.4 give some cases for which bµP ≥ 0 and we expect that it is a general rule. Notice that, if true, inequality (40) is the best possible: for instance, in the situation of Example 8.2, we have i=1 zαi is its spectrum at infinity. ✷ 1 µ µXi=1 (αi − 1)2 = 1 6 + (1 + λ1 + λ2) 6λ1λ2 and the last term on the right can be as small as possible. Equation (40) motivates the following conjecture (by Proposition 5.2, we have α1 = 0 and αµ = n if f is a convenient and nondegenerate Laurent polynomial) which has been already stated 24 without any further comments in [12, Remark 4.15] as a global counterpart of C. Hertling's con- jecture for germs of holomorphic functions (see [16], where the equality is inversed). The tameness assumption is discussed in Section 5.1. Conjecture on the variance of the spectrum (global version) Let f be a regular, tame function on a smooth n-dimensional affine variety U . Then 1 µ µXi=1 (αi − n 2 )2 ≥ 1 12 (αµ − α1) where α1 ≤ · · · ≤ αµ is the ordered spectrum of f at infinity. (41) ✷ This is another story, but one should expect equality in formula (41) if f belongs to the ideal generated by its partial derivatives (this is the case for quasi-homogeneous polynomials, see [7] and [16]) because mirror symmetry predicts that the multiplication by f on its Jacobi ring corresponds to the cup-product by c1(X) on the cohomology algebra of the mirror partner X of f . References [1] [2] [3] [4] [5] Batyrev, V.: Stringy Hodge numbers of varieties with Gorenstein canonical singularities, in: Integrable Systems and Algebraic Geometry (ed. M-H Saito et al.), Proceedings of the 41st Taniguchi Symposium, Japan 1997, World Scientific, 1-32 (1998). Batyrev, V.: Stringy Hodge numbers and Virasoro algebra, Math. Res. Lett., 7, 155-164 (2000). Batyrev, V., Schaller, K.: Stringy Chern classes of singular toric varieties and their applica- tions, arXiv:1607.04135. Beck, M. , Robbins, S.: Computing the continuous discretely, Springer, New York (2007). Borisov, L., Chen, L., Smith, G.: The orbifold Chow ring of toric Deligne-Mumford stacks, J. Amer. Soc. , 18(1), 193-215 (2005). [6] Cox, D., Little, J., Schenck, A.: Toric varieties, American Math. Soc., 124 (2010). [7] Dimca, A.: Monodromy and Hodge theory of regular functions, in: New Developments in Singularity Theory (ed. D. Siersma et al.), Kluwer Acad. Publ., Dordrecht-Boston-London, 257-278 (2001). [8] Douai, A.: Global spectra, polytopes and stacky invariants, arXiv:1603.08693. [9] Douai, A.: Quantum differential systems and some applications to mirror symmetry, arXiv:1203.5920. [10] Douai, A.: Quantum differential systems and rational structures, Manuscripta Mathematica, 145(3), 285-317 (2014). 25 [11] Douai, A., Mann, E.: The small quantum cohomology of a weighted projective space, a mirror D-module and their classical limits, Geometriae Dedicata, 164, 187-226 (2013). [12] Douai, A., Sabbah, C.: Gauss-Manin systems, Brieskorn lattices and Frobenius structures I, Ann. Inst. Fourier, 53(4), 1055-1116 (2003). [13] Douai, A., Sabbah, C.: Gauss-Manin systems, Brieskorn lattices and Frobenius structures II, in: Frobenius Manifolds, C. Hertling and M. Marcolli (Eds.), Aspects of Mathematics E 36 (2004). [14] Fulton, W.: Introduction to toric varieties, Ann. of Math. Stud., 131, Princeton University Press, Princeton, NJ (1993). [15] Givental, A.: Homological geometry and mirror symmetry, Talk at ICM-94, in: Proceedings of ICM-94 Zurich. Birkhauser, Basel, 472-480 (1995). [16] Hertling, C.: Frobenius manifolds and moduli spaces for singularities, Cambridge Tracts in Mathematics, 151, Cambridge University Press, Cambridge (2002). [17] Hertling, C.: Frobenius manifolds and variance of the spectral numbers, in: New Developments in Singularity Theory (ed. D. Siersma et al.), Kluwer Acad. Publ., Dordrecht-Boston-London, 235-255 (2001). [18] Hori, K., Vafa, C.: Mirror symmetry, arXiv:hep-th/0002222 [19] Kouchnirenko, A.G.: Poly`edres de Newton et nombres de Milnor, Invent. Math. 32, 1-31 (1976). [20] Libgober, A.S., Wood, J.W.: Uniqueness of the complex structure on Kahler manifolds of certain homotopy types, J. Diff. Geom., 32(1), 139-154 (1990). [21] Mustat¸a, M., Payne, S.: Ehrhart polynomials and stringy Betti numbers, Mathematische Annalen, 333(4), 787-795 (2005). [22] Nemethi, A. , Sabbah, C.: Semicontinuity of the spectrum at infinity, Abh. Math. Sem. Univ. Hamburg, 69, 25-35 (1999). [23] Reichelt, T., Sevenheck, C.: Logarithmic Frobenius manifolds, hypergeometric systems and quantum D-modules, J. Algebraic Geom., 24, 201-281 (2015). [24] Sabbah, C.: Hypergeometric periods for a tame polynomial, Portugaliae Mathematica, Nova Serie, 63(2), 173-226 (2006). [25] Stapledon, A.: Weighted Ehrhart Theory and Orbifold Cohomology, Adv. Math., 219, 63-88 (2008). [26] Veys, W.: Arc spaces, motivic integration and stringy invariants, Adv. Stud. Pure Math., 43, Math. Soc. Japan, 529-572 (2006). 26 

1102.1150

4

1102

Derived algebraic geometry, determinants of perfect complexes, and applications to obstruction theories for maps and complexes

We show how a quasi-smooth derived enhancement of a Deligne-Mumford stack X naturally endows X with a functorial perfect obstruction theory in the sense of Behrend-Fantechi. This result is then applied to moduli of maps and perfect complexes on a smooth complex projective variety. For moduli of maps, we consider X=S an algebraic K3-surface, $g\geq 0$, and $\beta$ a curve class, and we construct a derived stack whose truncation is the usual stack of pointed stable maps from curves of genus g to S hitting the class $\beta$, and such that the inclusion of the trunaction induces on a perfect obstruction theory whose tangent and obstruction spaces coincide with the corresponding reduced spaces of Okounkov-Maulik-Pandharipande-Thomas. We give two further applications to moduli of complexes. For a K3-surface S we show that the stack of simple perfect complexes on S is smooth. This result was proved with different methods by Inaba for the corresponding coarse moduli space. Finally, we construct a map from the derived stack of stable embeddings of curves (into a smooth complex projective variety X) to the derived stack of simple perfect complexes on X with vanishing negative Ext's, and show how this map induces a morphism of the corresponding obstruction theories when X is a Calabi-Yau threefold. An important ingredient of our construction is a perfect determinant map from the derived stack of perfect complexes to the derived stack of line bundles whose tangent morphism is, pointwise, Illusie's trace map for perfect complexes. We expect that this determinant map might be useful in other contexts as well.

math.AG

math

Derived algebraic geometry, determinants of perfect complexes, and applications to obstruction theories for maps and complexes Timo Schurg Max-Planck-Institut fur Mathematik Bertrand Toen I3M UMR 5149 Bonn - Germany Universit´e de Montpellier2 - France Montpellier - France 1 1 0 2 v o N 4 ] . G A h t a m [ 4 v 0 5 1 1 . 2 0 1 1 : v i X r a Gabriele Vezzosi Dipartimento di Sistemi ed Informatica Sezione di Matematica Universit`a di Firenze Firenze - Italy June 2011 Abstract red red We show how a quasi-smooth derived enhancement of a Deligne-Mumford stack X naturally endows X with a functorial perfect obstruction theory in the sense of Behrend-Fantechi. This result is then applied to moduli of maps and perfect complexes on a smooth complex projective variety. For moduli of maps, for X = S an algebraic K3-surface, g ∈ N, and β 6= 0 in H2(S, Z) a curve class, we construct a derived stack RM g,n(S; β) whose truncation is the usual stack Mg,n(S; β) of pointed stable maps from curves of genus g to S hitting the class β, and such that the inclusion Mg(S; β) ֒→ RM g (S; β) induces on Mg(S; β) a perfect obstruction theory whose tangent and obstruction spaces coincide with the corresponding reduced spaces of Okounkov-Maulik-Pandharipande-Thomas [O-P2, M-P, M-P-T]. The approach we present here uses derived algebraic geometry and yields not only a full rigorous proof of the existence of a reduced obstruction theory - not relying on any result on semiregularity maps - but also a new global geometric interpretation. We give two further applications to moduli of complexes. For a K3-surface S we show that the stack of simple perfect complexes on S is smooth. This result was proved with different methods by Inaba ([In]) for the corresponding coarse moduli space. Finally, we construct a map from the derived stack of stable embeddings of curves (into a smooth complex projective variety X) to the derived stack of simple perfect complexes on X with vanishing negative Ext’s, and show how this map induces a morphism of the corresponding obstruction theories when X is a Calabi-Yau threefold. An important ingredient of our construction is a perfect determinant map from the derived stack of perfect complexes to the derived stack of line bundles whose tangent morphism is, pointwise, Illusie’s trace map for perfect complexes. We expect that this determinant map might be useful in other contexts as well. 1 Mathematics Subject Classification (2010): 14J10, 14A20, 14J28, 14N35. Contents 1 Derived extensions, obstruction theories and their functoriality 1.1 Derived extensions induce obstruction theories . . . . . . . . . . . . . . . . 1.2 Functoriality of deformation theories induced by derived extensions . . . . . 7 7 8 2 Derived stack of stable maps and derived Picard stack 9 2.1 The derived Picard stack . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9 2.2 The derived stack of stable maps . . . . . . . . . . . . . . . . . . . . . . . . 10 3 The derived determinant morphism 14 3.1 The perfect determinant map . . . . . . . . . . . . . . . . . . . . . . . . . . 14 3.2 The map RMg(X) −→ RPerf (X) . . . . . . . . . . . . . . . . . . . . . . . 16 4 The reduced derived stack of stable maps to a K3-surface 19 4.1 Review of reduced obstruction theory . . . . . . . . . . . . . . . . . . . . . 19 4.1.1 Deformation and obstruction spaces of the reduced theory according 4.2 The canonical projection RPic(S) −→ RSpec(Sym(H 0(S, KS )[1])) 4.3 The reduced derived stack of stable maps RM to O-M-P-T . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20 . . . . . 23 red g (S; β) . . . . . . . . . . . . 25 4.4 Quasi-smoothness of RM red g (S; β) and comparison with O-M-P-T reduced obstruction theory. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27 5 Moduli of perfect complexes 31 5.1 On K3 surfaces . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 32 . . . . . . . . . . . . . . . . . . . . . . . . . . . . 34 5.2 On Calabi-Yau threefolds A Derived stack of perfect complexes and Atiyah classes 37 Introduction It is well known in Algebraic Geometry - e.g. in Gromov-Witten and Donaldson-Thomas theories - the importance of endowing a Deligne-Mumford moduli stack with a (perfect) obstruction theory, as defined in [B-F]: such an obstruction theory gives a virtual fun- damental class in the Chow group of the stack. If the stack in question is the stack of pointed stable maps to a fixed smooth projective variety ([B-M]), then integrating appro- priate classes against this class produces all versions of Gromov-Witten invariants ([Be]). Now, it is a distinguished feature of Derived Algebraic Geometry (as exposed, e.g. in [HAG-II]) that any quasi-smooth derived extension of such a stack F , i.e. a derived stack whose underived part or truncation is the given stack F , induces a canonical obstruction theory on F : we have collected these results in §1 below. A morphisms of derived stacks induces naturally a morphism between the induced obstruction theories - so that functo- riality results like [B-F, Prop. 5.10] or the so-called virtual pullback result in [Man] follow immediately. Moreover the functoriality of obstruction theories induced by morphisms of 2 derived extensions is definitely richer than the usual one in [B-F], that is restricted to special situations (e.g. [B-F, Prop. 5.10]), and requires the axiomatics of compatible ob- struction theories. In other words, a suitable reformulation of a moduli problem in derived algebraic geometry, immediately gives us a canonical obstruction theory, in a completely geometric way, with no need of clever choices. And, under suitable conditions, also the converse is expected to hold. The present paper is an application of this feature of derived algebraic geometry, not re- lying on the mentioned conjectural general equivalence between a class of derived stacks and a class of underived stacks endowed with a properly structured obstruction theory. However, as a matter of fact, all the obstruction theories we are aware of indeed arise from derived extensions - and the cases covered in this paper simply add to this list. In this paper we apply this ability of derived algebraic geometry in producing obstruc- tion theories - functorial with respect to maps of derived stacks - to the cases of moduli of maps and moduli of perfect complexes on a complex smooth projective variety X. Moduli of maps. For moduli of maps, we show how the standard obstruction theory yielding Gromov-Witten invariants comes from a natural derived extension of the stack of pointed stable maps to X. Then we concentrate on the first geometrically interest- ing occurrence of two different obstruction theories on a given stack, the stack Mg(S; β) of stable maps of type (g, β) to a smooth projective complex K3-surface S. The stack Mg(S; β) has a standard obstruction theory, yielding trivial Gromov-Witten invariants in the n-pointed case, and a so-called reduced obstruction theory, first considered by Okounkov-Maulik-Pandharipande-Thomas (often abbreviated to O-M-P-T in the text), giving interesting - and extremely rich in structure - curve counting invariants in the n- pointed case (see [P1, M-P, M-P-T], and §4.1 below, for a detailed review). In this paper we use derived algebraic geometry to give a construction of a global reduced obstruction theory on Mg(S; β), and compare its deformation and obstruction spaces with those of Okounkov-Maulik-Pandharipande-Thomas. More precisely, we use a perfect determinant map form the derived stack of perfect complexes to the derived stack of line bundles, and exploit the peculiarities of the derived stack of line bundles on a K3-surface, to produce red a derived extension RM g (S; β) arises as the canonical homotopy fiber over the unique derived factor of the derived stack of line bundles on S, so it is, in a very essential way, a purely derived geometrical object. We red g (S; β), and this immediately gives us a global reduced prove quasi-smoothness of RM obstruction theory on Mg(S; β). Our proof is self contained (inside derived algebraic ge- ometry), and does not rely on any previous results on semiregularity maps. red g (S; β) of Mg(S; β). The derived stack RM Moduli of complexes. We give two applications to moduli of perfect complexes on smooth projective varieties. In the first one we show that the moduli space of simple per- fect complexes on a K3-surface is smooth. Inaba gave a direct proof of this result in [In], by generalizing methods of Mukai ([Mu]). Our proof is different and straightforward. We use the perfect determinant map, and the peculiar structure of the derived Picard stack of a K3-surface, to produce a derived stack of simple perfect complexes. Then we show that this derived stack is actually underived (i.e trivial in the derived direction) and smooth. The moduli space studied by Inaba is exactly the coarse moduli space of this stack. In the second application, for X an arbitrary smooth complex projective scheme X, we 3 L first construct a map C from the derived stack RMg,n(X)emb consisting of pointed stable maps which are closed immersions, to the derived stack RPerf (X)si,>0 of simple per- fect complexes with no negative Ext’s and fixed determinant L (for arbitrary L). Then we show that, if X is a Calabi-Yau threefold, the derived stack RPerf (X)si,>0 is actually quasi-smooth, and use the map C to compare (according to §5.2) the canonical obstruction theories induced by the source and target derived stacks on their truncations. Finally, we relate this second applications to a baby, open version of the Gromov-Witten/Donaldson- Thomas conjectural comparison. In such a comparison, one meets two basic problems. The first, easier, one is in producing a map enabling one to compare the obstruction the- ories - and derived algebraic geometry, as we show in the open case, is perfectly suited for this (see §1 and §5.2). Such a comparison would induce a comparison (via a virtual pullback construction as in [Sch, Thm 7.4]) between the corresponding virtual fundamental classes, and thus a comparison between the GW and DT invariants. The second problem, certainly the most difficult one, is to deal with problems arising at the boundary of the compactifications. For this second problem, derived algebraic methods unfortunately do not provide at the moment any new tool or direction. L One of the main ingredients of all the applications given in this paper is the construc- tion of a perfect determinant map detPerf : Perf → Pic, where Perf is the stack of perfect complexes, Pic the stack of line bundles, and both are viewed as derived stacks (see §3.1 for details), whose definition requires the use of a bit of Waldhausen K-theory for simplicial commutative rings, and whose tangent map can be identified with Illusie’s trace map of perfect complexes ([Ill, Ch. 5]). We expect that this determinant map might be useful in other moduli contexts as well. An important remark - especially for applications to Gromov-Witten theory - is that, in order to simplify the exposition, we have chosen to write the proofs only in the non-pointed case, since obviously no substantial differences except for notational ones are involved. The relevant statements are however given in both the unpointed and the n-pointed case. To summarize, there are four main points in our paper. The first one is that derived algebraic geometry is a natural world where to get completely functorial obstruction the- ories. This is proved in §1, and explained through many examples in the rest of the text. The second main point concerns the application to reduced obstruction theory for stable maps to a K3 surface. More precisely, we give a rigorous proof of the existence of a global reduced obstruction theory on the stack of pointed stable maps to a K3-surface. The most complete, among the previous attempts in the literature, is the unpublished [M-P, §2.2], that only leads - after some further elaboration - to a uniquely defined “obstruction theory” with target just the truncation in degrees ≥ −1 of the cotangent complex of the stack of maps (from a fixed domain curve). Moreover, in order to reach this, the authors invoke some results on semiregularity maps, whose validity does not seem to be completely satisfactorily established. Nevertheless, there is certainly a clean and complete descrip- tion of the corresponding expected tangent and obstruction spaces in several papers (see [O-P2, M-P, M-P-T]), and we prove that our obstruction theory has exactly such tangent and obstruction spaces. Our approach does not only establish rigorously such a reduced global obstruction theory - with values in the full cotangent complex of the stack of stable maps - but also endows such an obstruction theory with all the functoriality properties that sometimes are missing in the underived, purely obstruction-theoretic approach. This 4 might prove useful in getting similar results in families or even relative to the whole mod- uli stack of K3-surfaces. Our definition of the reduced obstruction theory on the stack of stable maps to a K3-surface comes together with a clear (derived) geometrical picture - half of which is valid for any smooth complex projective variety. This should be compared to the other existing, partial approaches, where the construction of obstruction spaces arises from local linear algebra manipulations whose global geometrical interpretation is a bit obscure. In order to get a satisfying global geometric picture underlying this reduced obstruction theory, we are forced to move to the world of derived algebraic geometry; so, in some sense, our picture describes and explains the geometry underlying those local computations. This is another example of what seems to be a fairly general principle: some constructions on moduli stacks, that happen to be ad hoc inside algebraic geometry, become canonical and gain a neater geometrical interpretation in derived algebraic geom- etry. The third main point of our paper is another application of our perfect determinant map. We show that the stack of simple perfect complexes on a K3-surface is smooth. In the fourth and final application, for X a Calabi-Yau threefold, we use the perfect determi- nant map and the functoriality of obstruction theories arising from derived extensions, to produce a map from a derived stack of stable maps to X to the derived stack of simple perfect complexes with fixed determinant on X. We prove that the target derived stack is quasi-smooth, and show how this map induces, by functoriality, a comparison map be- tween the associated obstruction theories. Finally, let us observe that the emerging picture seems to suggest that most of the nat- ural maps of complexes arising in moduli problems can be realized as tangent or cotangent maps associated to morphisms between appropriate derived moduli stacks. This sugges- tion is confirmed in the present paper for the standard obstruction theories associated to the stack of maps between a fixed algebraic scheme and a smooth projective target, to the stack of stable maps to a smooth projective scheme or to the Picard stack of a smooth projective scheme, for the trace map, the Atiyah class map, and the first Chern class map for perfect complexes ([Ill, Ch. 5]), and for the map inducing O-M-P-T’s reduced obstruction theory. Description of contents. The first three sections and the beginning of the fifth are written for an arbitrary smooth complex projective scheme X. We explain how a derived extension induces an obstruction theory on its truncation (§1), how to define the standard derived extensions of the Picard stack of X, and of the stack of stable maps to X (§2), and finally define the perfect determinant map (§3). In section (§4), we specialize to the case where X = S is a smooth complex projective K3 surface. We first give a self-contained description of O-M-P-T’s pointwise reduced tangent and obstruction spaces (§4.1). Then, by exploiting the features of the derived Picard stack of S (§4.2), we define in §4.3 a derived red g (S; β) of the usual stack Mg(S; β) of stable maps of type (g, β 6= 0) to S, extension RM red having the property that, for the canonical inclusion jred : Mg(S; β) ֒→ RM g (S; β), the induced map j∗ red L RM red g (S;β) −→ L Mg(S;β) is a perfect obstruction theory with the same tangent and obstruction spaces as the re- duced theory introduced by Maulik-Okounkov-Pandharipande-Thomas (§4.4, Theorem. 4.8). 5 In §5, for a complex smooth projective variety X, we define the derived stack RMg,n(X)emb of pointed stable maps to X that are closed embeddings, the derived stack MX ≡ RPerf (X)si,>0 of simple perfect complexes on X with vanishing negative Ext’s, and the derived stack MX,L ≡ RPerf (X)si,>0 of simple perfect complexes on X with vanishing negative Ext’s and fixed determinant L, and we define a morphism C : RMg,n(X)emb −→ MX,L. When X is a K3-surface, we show that the truncation stack of MX is smooth. When X is a Calabi-Yau threefold, we prove that MX,L is quasi-smooth, and that the map C induces a map between the obstruction theories on the underlying underived stacks. L In an Appendix we give a derived geometrical interpretation of the Atiyah class map and the first Chern class map for a perfect complex E on a scheme Y , by relating them to the tangent of the corresponding map ϕE : Y → Perf ; then we follow this reinterpretation to prove some properties used in the main text. Acknowledgments. Our initial interest in the possible relationships between reduced obstruction theories and derived algebraic geometry was positively boosted by comments and questions by B. Fantechi, D. Huybrechts and R. Thomas. We are grateful to R. Pandharipande for pointing out a useful classical statement, and to H. Flenner for some important remarks. We especially thanks A. Vistoli for generously sharing his expertise on stable maps with us, and R. Thomas for his interest and further comments on this paper. The second and third authors acknowledge financial support from the french ANR grant HODAG (ANR-09-BLAN-0151). Notations. For background and basic notations in derived algebraic geometry we refer the reader to [HAG-II, Ch.2.2] and to the overview [To-2, §4.2, 4.3]. In particular, StC (respectively, dStC) will denote the (homotopy) category of stacks (respectively, of derived stacks) on Spec C with respect to the ´etale (resp., strongly ´etale) topology. We will most often omit the inclusion functor i : StC → dStC from our notations, since it is fully faith- ful; its left adjoint, the truncation functor, will be denoted t0 : dStC → StC ([HAG-II, In particular, we will write t0(F ) ֒→ F for the adjunction morphism Def. 2.2.4.3]). it0(F ) ֒→ F . However recall that the inclusion functor i does not commute with taking internal HOM (derived) stacks nor with taking homotopy limits. All fibered products of derived stacks will be implicitly derived (i.e. they will be homotopy fibered products in the model category of derived stacks). When useful, we will freely switch back and forth between (the model category of) sim- plicial commutative k-algebras and (the model category of) commutative differential non- positively graded k-algebras, where k is a field of characteristic 0 ([To-Ve, App. A]). All complexes will be cochain complexes and, for such a complex C •, either C≤n or C ≤n (depending on typographical convenience) will denote its good truncation in degrees ≤ n. Analogously for either C≥n or C ≥n ([Wei, 1.2.7]). To ease notation we will often write ⊗ for the derived tensor product ⊗L, whenever no confusion is likely to arise. X will denote a smooth complex projective scheme while S a smooth complex projective K3-surface. As a purely terminological remark, for a given obstruction theory, we will call its de- formation space what is usually called its tangent space (while we keep the terminology 6 obstruction space). We do this to avoid confusion with tangent spaces, tangent complexes or tangent cohomologies of related (derived) stacks. We will often abbreviate the list of authors Okounkov-Maulik-Pandharipande-Thomas to O-M-P-T. 1 Derived extensions, obstruction theories and their func- toriality We briefly recall here the basic observation that a derived extension of a given stack X induces an obstruction theory (in the sense of [B-F]) on X , and deduce a richer functori- ality with respect to the one known classically. Everything in this section is true over an arbitrary base ring, though it will be stated for the base field C. 1.1 Derived extensions induce obstruction theories Let t0 : dStC → StC be the truncation functor between derived and underived stacks over C for the ´etale topologies ([HAG-II, Def. 2.2.4.3]). It has a left adjoint i : StC → dStC which is fully faithful (on the homotopy categories), and is therefore usually omitted from our notations. Definition 1.1 Given a stack X ∈ Ho(StC), a derived extension of X is a derived stack X der together with an isomorphism X ≃ t0(X der). Proposition 1.2 Let X der be a derived geometric stack which is a derived extension of the (geometric) stack X . Then, the closed immersion induces a morphism j : X ≃ t0(X der) ֒→ X der j∗(L X der) −→ LX which is 2-connective, i.e. its cone has vanishing cohomology in degrees ≥ −1. Proof. The proof follows easily from the remark that if A is a simplicial commutative C-algebra and A → π0(A) is the canonical surjection, then the cotangent complex Lπ0(A)/A is 2-connective, i.e. has vanishing cohomology in degrees ≥ −1. ✷ The previous Proposition shows that a derived extension always induces an obstruc- tion theory (whenever such a notion is defined by [B-F, Def. 4.4], e.g. when X is a Deligne-Mumford stack). In particular, recalling that a derived stack is quasi-smooth if its cotangent complex is perfect of amplitude in [−1, 0], we have the following result Corollary 1.3 Let X der be a quasi-smooth derived Deligne-Mumford stack which is a derived extension of a (Deligne-Mumford) stack X . Then is a [−1, 0]-perfect obstruction theory as defined in [B-F, Def. 5.1]. j∗(L X der) −→ LX 7 Remark 1.4 We expect that also the converse is true, i.e. that given any stack X locally of finite presentation over a field k, endowed with a map of co-dg-Lie algebroids E → LX, there should exist a derived extension inducing the given obstruction theory. We will come back to this in a future work and will not use it in the rest of this paper, although it should be clear that it was exactly such an expected result that first led us to think about the present work. Moreover, as a matter of fact, all the obstruction theories we are aware of indeed come from derived extensions, and the cases covered below simply add to this list. 1.2 Functoriality of deformation theories induced by derived extensions If f : X → Y is a morphism of (Deligne-Mumford) stacks, and oX : EX → LX and oY : EY → LY are ([−1, 0]-perfect) obstruction theories, the classical theory of obstructions ([B-F]) does not provide in general a map f ∗EY → EX such that the following square f ∗EY f ∗(oY ) f ∗LY EX oX / LX is commutative (or commutative up to an isomorphism) in the derived category D(X) of complexes on X, where f ∗LY → LX is the canonical map on cotangent complexes induced by f ([Ill, Ch. 2, (1.2.3.2)’]). On the contrary, if jX : X ֒→ RX and jY : Y ֒→ RY are quasi-smooth derived (Deligne-Mumford) extensions of X and Y , respectively, and F : RX → RY is a morphism of derived stacks t0F X Y , RX F / / RY LRX → LX and j∗ Y then j∗ LRY → LY are ([−1, 0]-perfect) obstruction theories by Cor. X 1.3, and moreover there is indeed a canonical morphism of triangles in D(X) (we denote t0(F ) by f ) f ∗j∗ Y LRY f ∗LY f ∗LRY /Y j∗ X LRX / LX / LRX/X (see [HAG-II, Prop. 1.2.1.6] or [Ill, Ch. 2, (2.1.1.5)]). This map relates the two induced obstruction theories and may be used to relate the corresponding virtual fundamental classes, too (when they exist, e.g. when X and Y are proper over k). We will not do this here since we will not need it for the results in this paper. However, the type of result we are referring to is the following Proposition 1.5 [Sch, Thm. 7.4] Let F : RX → RY be a quasi-smooth morphism between quasi-smooth Deligne-Mumford stacks, and f : X → Y the induced morphism on the truncations. Then, there is an induced virtual pullback (as defined in [Man]) f ! : A∗(Y ) → A∗(X), between the Chow groups of Y and X, such that f !([Y ]vir) = [X]vir, where [X]vir (respectively, [Y ]vir) is the virtual fundamental class ( [B-F]) on X (resp., on Y ) induced by the [−1, 0] perfect obstruction theory j∗ X LRX → LX (resp., by j∗ Y LRY → LY ). 8 / /     / / /  _   _   / /   / /     / / 2 Derived stack of stable maps and derived Picard stack In this section we prove a correspondence between derived open substacks of a derived stack and open substacks of its truncation, and use it to construct the derived Picard stack RPic(X; β) of type β ∈ H 2(X, Z), for any complex projective smooth variety X. After recalling the derived version of the stack of (pre-)stable maps to X, possibly pointed, the same correspondence will lead us to defining the derived stack RMg(X; β) of stable maps of type (g, β) to X and its pointed version. Throughout the section X will denote a smooth complex projective scheme, g a nonneg- ative integer, c1 a class in H 2(X, Z) (which, for our purposes, may be supposed to belong to the image of Pic(X) ≃ H 1(X, O∗ X ) → H 2(X, Z), i.e. belonging to H 1,1(X) ∩ H 2(X, Z)), and β ∈ H2(X, Z) an effective curve class. We will frequently use of the following Proposition 2.1 Let F be a derived stack and t0(F ) its truncation. There is a bijective correspondence φF : {Zariski open substacks of t0(F )} −→ {Zariski open derived substacks of F }. For any Zariski open substack U0 ֒→ t0(F ), we have a homotopy cartesian diagram in dStC U0 t0(F ) φF (U0)  / F where the vertical maps are the canonical closed immersions. Proof. The statement is an immediate consequence of the fact that F and t0(F ) have the same topology ([HAG-II, Cor. 2.2.2.9]). More precisely, let us define φF as follows. If U0 ֒→ t0(F ) is an open substack, φF (U0) is the functor SAlgC −→ SSets : A 7−→ F (A) ×t0(F )(π0(A)) U0(π0(A)) where F (A) maps to t0(F )(π0(A)) via the morphism (induced by the truncation functor t0) F (A) ≃ RHomdStC(RSpec(A), F ) −→ RHomStC(t0(RSpec(A)), t0(F )) ≃ t0(F )(π0(A)). The inverse to φF is simply induced by the truncation functor t0. ✷ 2.1 The derived Picard stack Definition 2.2 The Picard stack of X/C is the stack The derived Picard stack of X/C is the derived stack Pic(X) := RHOMStC(X, BGm). RPic(X) := RHOMdStC(X, BGm). 9   / /      / By definition we have a natural isomorphism t0(RPic(X)) ≃ Pic(X) in Ho(dStC). Note that even though Pic(X) is smooth, it is not true that RPic(X) ≃ Pic(X), if dim(X) > 1; this can be seen on tangent spaces since TLRPic(X) ≃ C•(X, OX )[1] for any global point xL : Spec(C) → RPic(X) corresponding to a line bundle L over X. Given c1 ∈ H 2(X, Z), we denote by Pic(X; c1) the open substack of Pic(X) classifying line bundles with first Chern class c1. More precisely, for any R ∈ AlgC, let us denote by Vect1(R; c1) the groupoid of line bundles L on Spec(R) × X such that, for any point x : Spec(C) → Spec(R) the pullback line bundle Lx on X has first Chern class equal to c1. Then, Pic(X; c1) is the stack: AlgC −→ SSets : R 7−→ Nerve(Vect1(R; c1)) where Nerve(C) is the nerve of the category C. Note that we have Pic(X) = a Pic(X; c1). c1∈H 2(X,Z) Definition 2.3 Let c1 ∈ H 2(X, Z). The derived Picard stack of type c1 of X/C is the derived stack RPic(X; c1) := φRPic(X)(Pic(X; c1)). In particular, we have a natural isomorphism t0(RPic(X; c1)) ≃ Pic(X; c1), and a homotopy cartesian diagram in dStC Pic(X; c1) Pic(X) RPic(X; c1)  / RPic(X) 2.2 The derived stack of stable maps We recall from [To-2, 4.3 (4.d)] the construction of the derived stack RMpre g (X) (respec- tively, RMpre g,n(X)) of prestable maps (resp., of n-pointed prestable maps) of genus g to X, and of its open derived substack RMg(X) (respectively, RMg,n(X)) of stable maps (resp., of n-pointed stable maps) of genus g to X. Then we move to define the derived version of the stack of (pointed) stable maps of type (g, β) to X. Let Mpre g genus g, and Cpre (respectively, Mpre g −→ Mpre (resp. Cpre g,n) be the stack of (resp. n-pointed) pre-stable curves of g,n) its universal family (see e.g. [Be, O-P1]). g,n −→ Mpre g Definition 2.4 defined as • The derived stack RMpre g (X) of prestable maps of genus g to X is RMpre g (X) := RHOMdStC/Mpre g (Cpre g , X × Mpre g ). 10  / /      / g (X) is then canonically a derived stack over Mpre RMpre derived universal family RCpre g , and the corresponding g; X is defined by the following homotopy cartesian square RCpre g; X / RMpre g (X) Cpre g / Mpre g • The derived stack RMpre g,n(X) of n-pointed prestable maps of genus g to X is defined as RMpre g,n(X) := RHOMdStC/Mpre g,n(Cpre g,n, X × Mpre g,n). g,n(X) is then canonically a derived stack over Mpre RMpre rived universal family RCpre g,n, and the corresponding de- g,n; X is defined by the following homotopy cartesian square RCpre g,n; X / RMpre g,n(X) Cpre g,n / Mpre g,n Note that, by definition, RCpre g; X comes also equipped with a canonical map RCpre g (X)) ≃ Mpre g; X −→ RMpre We also have t0(RMpre g (X) (the stack of prestable maps of genus g to X), and t0(RCpre g; X (the universal family over the stack of pre-stable maps of genus g to X), since the truncation functor t0 commutes with homotopy fibered products. The same is true for the pointed version. g; X ) ≃ Cpre g (X) × X. We can now use Proposition 2.1 to define the derived stable versions. Let Mg(X) (respectively, Mg,n(X) ) be the open substack of Mpre g,n(X)) con- sisting of stable maps of genus g to X (resp. n-pointed stable maps of genus g to X), and Cg; X −→ Mpre g,n(X)) the (induced) universal family ([Be, O-P1]). g (X) (resp. Cg,n; X −→ Mpre g (X) (resp. of Mpre Definition 2.5 • The derived stack RMg(X) of stable maps of genus g to X is de- fined as The derived stable universal family RMg(X) := φRMpre g (X)(Mg(X)). is the derived restriction of RCpre RCg; X −→ RMg(X) g; X → RMpre g (X) to RMg(X). • The derived stack RMg,n(X) of n-pointed stable maps of genus g to X is defined as The derived stable universal family RMg,n(X) := φRMpre g,n(X)(Mg,n(X)). is the derived restriction of RCpre RCg,n; X −→ RMg,n(X) g,n; X → RMpre g,n(X) to RMg,n(X). 11   /   /   /   / Recall that • t0(RMg(X)) ≃ Mg(X); • t0(RCg; X ) ≃ Cg; X ; • RCg; X comes equipped with a canonical map π : RCg; X −→ RMg(X) × X; • we have a homotopy cartesian diagram in dStC Mg(X)  Mpre g (X) RMg(X)  / RMpre g (X) With the obvious changes, this applies to the pointed version too. Let g a non-negative integer, β ∈ H2(X, Z), and Mg(X; β) (respectively, Mg,n(X; β)) be the stack of stable maps (resp. of n-pointed stable maps) of type (g, β) to X (see e.g. [Be] or [O-P1]); its derived version is given by the following Definition 2.6 • The derived stack of stable maps of type (g, β) to X is defined as the open substack of RMg(X) RMg(X; β) := φRMg(X)(Mg(X; β)). The derived stable universal family of type (g; β), RCg,β; X −→ RMg(X; β), is the (derived) restriction of RCg; X −→ RMg(X) to RMg(X; β). • The derived stack of n-pointed stable maps of type (g, β) to X is defined as the open substack of RMg,n(X) RMg,n(X; β) := φRMg,n(X)(Mg,n(X; β)). The derived stable universal family of type (g; β), RCg,n,β; X −→ RMg,n(X; β), is the (derived) restriction of RCg,n; X −→ RMg,n(X) to RMg,n(X; β). Note that, by definition, t0(RMg(X; β)) ≃ Mg(X; β), therefore RMg(X; β) is a proper derived Deligne-Mumford stack ([HAG-II, 2.2.4]). Moreover, the derived stable universal family RCg,β; X comes, by restriction, equipped with a natural map π : RCg,β; X −→ RMg(X; β) × X. 12  / /      / We have a homotopy cartesian diagram in dStC Mg(X; β)  Mg(X) . RMg(X; β)  / RMg(X) Analogous remarks are valid in the pointed case. The tangent complex of RMg(X; β) at a stable map (f : C → X) of type (g, β) (corresponding to a classical point xf : Spec(C) → RMg(X; β)) is given by1 T (f :C→X) ≃ RΓ(C, Cone(TC → f ∗TX)), g where TC is the tangent complex of C and TX is the tangent sheaf of X. The canonical map RMg(X; β) → Mpre is quasi-smooth. In fact, the fiber at a geomet- ric point, corresponding to prestable curve C, is the derived stack RHOMβ(C, X) whose tangent complex at a point f : C → X is RΓ(C, f ∗TS) which, obviously, has cohomology only in degrees [0, 1]. But Mpre is smooth, and any derived stack quasi-smooth over a smooth base is quasi-smooth (by the corresponding exact triangle of tangent complexes). Therefore the derived stack RMg(X; β) is quasi-smooth. Proposition 1.2 then recovers the standard (absolute) perfect obstruction theory on Mg(X; β) via the canonical map g induced by the closed immersion j : Mg(X; β) ֒→ RMg(X; β). j∗(L RMg(X;β)) −→ L Mg(X;β) In the pointed case, the tangent complex of RMg,n(X; β) at a pointed stable map (f : (C; x1, . . . , xn) → X) of type (g, β) (corresponding to a classical point xf : Spec(C) → RMg,n(X; β)) is likewise given by T(f :(C;x1,...,xn)→X) ≃ RΓ(C, Cone(TC(−X xi) → f ∗TX)). i The same argument as above proves that also the canonical map RMg,n(X; β) → Mpre g,n is quasi-smooth, and Proposition 1.2 then recovers the standard absolute perfect obstruction theory on Mg,n(X; β) via the canonical map j∗(L RMg,n(X;β)) −→ L Mg,n(X;β) induced by the closed immersion j : Mg,n(X; β) ֒→ RMg,n(X; β). Note that, as ob- served in [O-P1, 5.3.5], this obstruction theory yields trivial Gromov-Witten invariants on Mg,n(X; β) for X = S a K3 surface. Hence the need for another obstruction theory carrying more interesting curves counting invariants on a K3-surface: this will be the so-called reduced obstruction theory (see §4.1, §4.3, and Theorem 4.8). 1The [1] shift in [CF-K, Thm. 5.4.8] is clearly a typo: their proof is correct and yields no shift. 13  / /      / 3 The derived determinant morphism In this section we start by defining a quite general perfect determinant map of derived stacks detPerf : Perf −→ Pic = BGm whose construction requires a small detour into Waldhausen K-theory. We think this perfect determinant might play an important role in other contexts as well, e.g. in a general GW/DT correspondence. Using the perfect determinant together with a natural perfect complex on RMg(X; β), we will be able to define a map δ1(X) : RMg(X) −→ RPic(X) which will be one of the main ingredients in the construction of the reduced derived stack of stable maps RM red g (S; β), for a K3-surface S, given in the next section. 3.1 The perfect determinant map The aim of this subsection is to produce a determinant morphism detPerf : Perf −→ Pic in Ho(dStC) extending the natural determinant morphism Vect −→ Pic. To do this, we will have to pass through Waldhausen K-theory. We start with the classical determinant map in Ho(StC), det : Vect −→ Pic, induced by the map sending a vector bundle to its top exterior power. Consider the following simplicial stacks B•Pic : ∆op ∋ [n] 7−→ (Pic)n (with the simplicial structure maps given by tensor products of line bundles, or equiva- lently, induced by the product in the group structure of BGm ≃ Pic), and B•Vect : ∆op ∋ [n] 7−→ wSnVect, where, for any commutative C-algebra R, wSnVect(R) is the nerve of the category of sequences of split monomorphisms 0 → M1 → M2 → . . . → Mn → 0 with morphisms the obvious equivalences, and the simplicial structure maps are the natural ones described in [Wal, 1.3]. Similarly, we define the simplicial object in stacks B•Perf : ∆op ∋ [n] 7−→ wSnPerf (see [Wal, 1.3] for the definition of wSn in this case). Now, B•Pic and B•Vect, and B•Perf are pre-∆op-stacks according to Def. 1.4.1 of [To-1], and the map det extends to a morphism det• : B•Vect −→ B•Pic in the homotopy category of pre-∆op-stacks. By applying the functor i : Ho(StC) → Ho(dStC) (that will be, according to our conventions, omitted from notations), we get a determinant morphism (denoted in the same way) det• : B•Vect −→ B•Pic 14 in the homotopy category of pre-∆op-derived stacks. We now pass to Waldhausen K−theory, i.e. apply K := Ω ◦ − (see [To-1, Thm 1.4.3], where the loop functor Ω is denoted by RΩ∗, and the realization functor − by B), and observe that, by [To-1, Thm 1.4.3 (2)], there is a canonical isomorphism in Ho(dStC) K(B•Pic) ≃ Pic since Pic is group-like (i.e. an H∞-stack in the parlance of [To-1, Thm 1.4.3]). This gives us a map in Ho(dStC) K(det•) : K(B•Vect) −→ Pic. Now, consider the map u : KVect := K(B•Vect) −→ K(B•Perf ) := KPerf in Ho(dStC), induced by the inclusion Vect ֒→ Perf . By [Wal, Thm. 1.7.1], u is an isomorphism in Ho(dStC). Therefore, we get a diagram in Ho(dStC) KVect K(det•) Pic u Perf / KPerf 1st-level where u is an isomorphism. This allows us to give the following Definition 3.1 The induced map in Ho(dStC) is called the perfect determinant morphism. detPerf : Perf −→ Pic For any complex scheme (or derived stack) X, the perfect determinant morphism detPerf : Perf −→ Pic induces a map in Ho(dStC) detPerf (X) : RPerf (X) := RHOMdStC(X, Perf ) −→ RHOMdStC(X, Pic) =: RPic(X). As perhaps not totally unexpected (e.g. perfect determinant map is given by the trace for perfect complexes [Ill, Rem. 5.3.3]), the tangent morphism to the Proposition 3.2 Let X be a complex quasi-projective scheme, and detPerf(X) : RPerf (X) → RPic(X) the induced perfect determinant map. For any complex point xE : Spec C → RPerf (X), corresponding to a perfect complex E over X, the tangent map TxE detPerf (X) : TxE RPerf (X) ≃ RHom(E, E)[1] −→ RHom(OS , OS)[1] ≃ TxE RPic(X) is given by trE[1], where trE is the trace map for the perfect complex E of 3.7.3]. [Ill, Ch. 5, Proof. Let RPerf strict(X) := RHOMdStC(X, Perf strict) be the derived stack of strict perfect complexes on X ([SGA6, Exp. I, 2.1]). Since X is quasi-projective, the canon- ical map RPerf strict(X) → RPerf (X) is an isomorphism in Ho(dStC). Therefore (e.g. [SGA6, Exp. I, 8.1.2]), the comparison statement is reduced to the case where E is a vector bundle on X, which is a direct computation and is left to the reader. ✷ 15 / /   / 3.2 The map RMg(X) −→ RPerf(X) A map RMg(X) −→ RPerf (X) = RHOMdStC(X, Perf ) in Ho(dStC) is, by adjunction, the same thing as a map RMg(X) × X −→ Perf i.e. a perfect complex on RMg(X)× X; so, it is enough to find such an appropriate perfect complex. Let be the derived stable universal family (§2.2). π : RCg; X −→ RMg(X) × X Proposition 3.3 Rπ∗(ORCg; X ) is a perfect complex on RMg(X) × X. Proof. The truncation of π is proper, hence it is enough to prove that π is also quasi- smooth. To see this, observe that both RCg; X and RMg(X)×X are smooth over RMg(X). Then we conclude, since any map between derived stacks smooth over a base is quasi- smooth. ✷ Remark 3.4 If we fix a class β ∈ H2(X, Z), the corresponding β-decorated version of Proposition 3.3 obviously holds. We may therefore give the following Definition 3.5 We will denote by AX : RMg(X) −→ RPerf (X) the map induced by the perfect complex Rπ∗(ORCg; X ). Note that, by definition, AX sends a complex point of RMg(X), corresponding to a stable map f : C → X to the perfect complex Rf∗OC on X. The tangent morphism of AX . The tangent morphism of AX is related to the Atiyah class of Rπ∗(ORCg; X ), and pointwise on RMg(X) to the Atiyah class map of the perfect complex Rf∗OC: this is explained in detail in Appendix A, so we will just recall here the results and the notations we will need in the rest of the main text. Let us write E := Rπ∗(ORCg; X ); since this is a perfect complex on RMg(X)× X, its Atiyah class map (see Appendix A) corresponds uniquely, by adjunction, to a map, denoted in the same way, atE : E −→ L RMg(X)×X ⊗ E[1] atE : T RMg(X)×X −→ E ∨ ⊗ E[1]. 16 Let x be a complex point x of RMg(X) corresponding to a stable map f : C → X, and let p : C → Spec C and q : X → Spec C denote the structural morphisms, so that p = q ◦ f . Correspondingly, we have a ladder of homotopy cartesian diagrams ιf x C f X q RCg; X π RMg(X) × X prX / / X pr q Spec C / RMg(X) x / Spec C Let us consider the perfect complex E := Rf∗OC on X. By [Ill, Ch. 4, 2.3.7], the complex E has an Atiyah class map atE : E −→ E ⊗ Ω1 X[1] which corresponds uniquely (E being perfect) by adjunction to a map (denoted in the same way) atE : TX −→ REndX(Rf∗OC)[1]. Proposition 3.6 In the situation and notations above, we have that • the tangent map of AX fits into the following commutative diagram T RMg(X) can TAX A∗ X TRPerf (X) ∼ / Rpr∗(E ∨ ⊗ E)[1] Rpr∗(atE) Rpr∗pr∗T RMg(X) can / / Rpr∗(pr∗T RMg(X) ⊕ pr∗ XTX) ∼ / / Rpr∗T RMg(X)×X where can denote obvious canonical maps, and E := Rπ∗(ORCg; X ). • The tangent map to AX at x = (f : C → X), is the composition TxAX : TxRMg(X) ≃ RΓ(C, Cone(TC → f ∗TX)) / RΓ(X, x∗T RMg(X)×X ) RΓ(X,x∗atE ) / REndX(Rf∗OC )[1] ≃ TRf∗OC RPerf (X) where E := Rf∗OC • The composition RΓ(X, TX ) can / / RΓ(X, Rf∗f ∗TX) can / / RΓ(X, Cone(Rf∗TC → Rf∗f ∗TX )) ≃ TxRMg(X) TxAX / / x∗A∗ X TRPerf(X) ≃ TRf∗OC RPerf (X) ≃ REndX(Rf∗OC )[1] coincides with RΓ(X, atE), where E := Rf∗OC. Proof. See Appendix A. ✷ 17 / /     / /       / / / /   / O O / / / / / / Definition 3.7 We denote by δ1(X) the composition RMg(X) AX / / RPerf (X) detPerf (X) / RPic(X), and, for a complex point x of RMg(X) corresponding to a stable map f : C → X, by Θf := Tf δ1(X) : T (f :C→X) RMg(X) TxAX / / TRf∗OC RPerf (X) trX / / T det(Rf∗OC ) RPic(X). Note that, as a map of explicit complexes, we have Θf : RΓ(C, Cone(TC → f ∗TX )) TxAX / / RHomX(Rf∗OC, Rf∗OC )[1] trX / / RΓ(X, OX )[1] Remark 3.8 - First Chern class of Rf∗OC and the map Θf . Using Proposition 3.6, we can relate the map Θf above to the first Chern class of the perfect complex Rf∗OC ([Ill, Ch. V]). With the same notations as in Prop. 3.6, the following diagram is commutative Rq∗(atRf∗OC ) Rq∗TX / Rq∗REndX(Rf∗OC )[1] TxAX Rq∗Rf∗f ∗TX ≃ Rp∗f ∗TX / Rp∗Cone(TC → f ∗TX) tr Θf / Rq∗OX[1] . id / Rq∗OX[1] In this diagram, the composite upper row is the image under Rq∗ of the first Chern class c1(Rf∗OC) ∈ Ext1 X(TX , OX ) ≃ H 1(X, Ω1 X ). Pointed case - In the pointed case, if π : RCg,n; X −→ RMg,n(X) × X is the derived stable universal family (§2.2), the same argument as in Proposition 3.3 shows that Rπ∗(ORCg; X ) is a perfect complex on RMg,n(X) × X. And we give the analogous Definition 3.9 • We denote by A(n) X : RMg,n(X) −→ RPerf (X) the map induced by the perfect complex Rπ∗(ORCg,n; X ). • We denote by δ(n) 1 (X) the composition RMg,n(X) A(n) X / / RPerf (X) detPerf (X) / RPic(X), and, for a complex point x of RMg(X) corresponding to a pointed stable map f : (C; x1, . . . , xn) → X, by Θ(n) f := Tf δ(n) 1 (X) : Tf RMg,n(X) TxA(n) X / / TRf∗OC RPerf (X) trX / / T det(Rf∗OC ) RPic(X). And again, if we fix a class β ∈ H2(X, Z), we have the corresponding β-decorated version of Definition 3.9. 18 /   / / / O O / O O / 4 The reduced derived stack of stable maps to a K3-surface In this section we specialize to the case of a K3-surface S, with a fixed nonzero curve class β ∈ H2(S; Z) ≃ H 2(S; Z). After recalling in some detail the reduced obstruction theory of O-M-P-T, we first identify canonically the derived Picard stack RPic(S) with Pic(S) × RSpec(Sym(H 0(S, KS )[1])) where KS is the canonical sheaf of S. This result is red then used to define the reduced version RM g (S; β) of the derived stack of stable maps red g,n(S; β)), and to show that this in- of type (g, β) to S (and its n-pointed variant RM duces, via the canonical procedure available for any algebraic derived stack, a modified obstruction theory on its truncation Mg(S; β) whose deformation and obstruction spaces are then compared with those of the reduced theory of O-M-P-T. As a terminological remark, given an obstruction theory, we will call deformation space what is usually called its tangent space (while we keep the terminology obstruction space). We do this to avoid confusion with tangent spaces, tangent complexes or tangent cohomologies of possibly re- lated (derived) stacks. 4.1 Review of reduced obstruction theory For a K3-surface S, the moduli of stable maps of genus g curves to S with non-zero effective class β ∈ H 1,1(S, C) ∩ H 2(S, Z) (note that Poincar´e duality yields a canonical isomorphism H2(S; Z) ≃ H 2(S; Z) between singular (co)homologies) carries a relative perfect obstruction theory. This obstruction theory is given by (Rπ∗F ∗TS)∨ → L Mg(S;β)/Mpre g . Here π : Cg,β; S → Mg(S; β) is the universal curve, F : Cg,β; S → S is the universal morphism from the universal curve to S, and Mpre denotes the Artin stack of prestable curves. A Riemann-Roch argument along with the fact that a K3-surface has trivial canonical bundle yields the expected dimension of Mg(S; β): g exp dim Mg(S; β) = g − 1. We thus expect no rational curves on a K3-surface. This result stems from the deformation invariance of Gromov-Witten invariants. A K3-surface admits deformations such that the homology class β is no longer of type (1, 1), and thus can not be the class of a curve. This is unfortunate, given the rich literature on enumerative geometry of K3-surfaces, and is in stark contrast to the well-known conjecture that a projective K3-surface over an algebraically closed field contains infinitely many rational curves. Further evidence that there should be an interesting Gromov-Witten theory of K3-surfaces are the results of Bloch, Ran and Voisin that rational curves deform in a family of K3-surfaces provided their homology classes remain of type (1, 1). The key ingredients in the proof is the semi- regularity map. We thus seek a new kind of obstruction theory for Mg(S; β) which is deformation invariant only for such deformations of S which keep β of type (1, 1). Such a new obstruction theory, called the reduced obstruction theory, was introduced in [O-P2, M-P, M-P-T]. Sticking to the case of moduli of morphisms from a fixed curve C to S, the obstruction space at a fixed morphism f is H 1(C, f ∗TS). This obstruction space admits a map 19 H 1(C, f ∗TS) ∼ / / H 1(C, f ∗ΩS) H 1(df ) / H 1(C, Ω1 C ) / H 1(C, ωC ) ≃ C , where the first isomorphism is induced by the choice of a holomorphic symplectic form on S. The difficult part is to prove that all obstructions for all types of deformations of f (and not only curvilinear ones) lie in the kernel of this map. Once this is proven, Mg(S; β) car- ries a reduced obstruction theory which yields a virtual class, called the reduced class. This reduced class is one dimension larger that the one obtained from the standard perfect ob- struction theory and leads to many interesting enumerative results (see [P1, M-P, M-P-T]). We will review below the construction of the reduced deformation and obstruction spaces giving all the details will be needed in our comparison result (Thm. 4.8). 4.1.1 Deformation and obstruction spaces of the reduced theory according to O-M-P-T For further reference, we give here a self-contained treatment of the reduced deformation and reduced obstruction spaces on Mg(S; β) according to Okounkov-Maulik-Pandharipande- Thomas. Let us fix a stable map f : C → S of class β 6= 0 and genus g; p : C → Spec C and q : S → Spec C will denote the structural morphisms. Let ωC ≃ p!OSpec C be the dualizing complex of C, and ωC = ωC[−1] the corresponding dualizing sheaf. First of all, the deformation spaces of the standard (i.e. unreduced) and reduced theory, at the stable map f , coincide with where TC is the cotangent complex of the curve C. H 0(C, Cone(TC → f ∗TS)) Let’s recall now ([P1, §3.1]) the construction of the reduced obstruction space. We give here a canonical version, independent of the choice of a holomorphic symplectic form on S. Consider the canonical isomorphism2 ϕ : TS ⊗ H 0(S, KS ) ∼ / / Ω1 S. By tensoring this by H 0(S, KS )∨ ≃ H 2(S, OS ) (this isomorphism is canonical by Serre duality) which is of dimension 1 over C, we get a canonical sequence of isomorphisms of OS-Modules TS ∼o TS ⊗ H 0(S, KS ) ⊗ H 2(S, OS ) ∼ / / Ω1 S ⊗ H 2(S, OS ). We denote by ψ : TS → Ω1 we get a canonical isomorphism of OC -Modules S ⊗ H 2(S, OS ) the induced, canonical isomorphism. Form this, f ∗ψ : f ∗TS ∼ / / f ∗(Ω1 S) ⊗ H 2(S, OS). 2We use throughout the standard abuse of writing F ⊗V for F ⊗OX p∗V , for any scheme p : X → Spec C, any OX -Module F, and any C-vector space V . 20 / / o Now consider the canonical maps f ∗Ω1 S s / Ω1 C t / ωC ≃ p!OSpec C[−1] where ωC ≃ ωC[−1] is the dualizing sheaf of C and ωC = p!OSpec C the dualizing complex of C (see [Ha-RD, Ch. V]). We thus obtain a map ev : f ∗TS −→ ωC ⊗ H 2(S, OS )[−1] ≃ ωC ⊗ H 2(S, OS ). By the properties of dualizing complexes, we have ωC ⊗ H 2(S, OS )[−1] = ωC ⊗ p∗(H 2(S, OS ))[−1] ≃ p!(H 2(S, OS )[−1]), so we get a canonical morphism f ∗TS −→ p!(H 2(S, OS)[−1]) which induces, by applying Rp∗ and composing with the adjunction map Rp∗p! → Id, a canonical map eα : RΓ(C, f ∗TS) ≃ Rp∗(f ∗TS) Rp∗(ev) / Rp∗(ωC ⊗ H 2(S, OS )) ≃ Rp∗p!(H 2(S, OS)[−1]) / H 2(S, OS )[−1] . Since RΓ is a triangulated functor, to get a unique induced map α : RΓ(C, Cone(TC → f ∗TS)) −→ H 2(S, OS )[−1] it will be enough to observe that HomD(C)(Rp∗TC[1], H 2(S, OS )[−1]) = 0 (which is obvious since Rp∗TC[1] lives in degrees [−1, 0], while H 2(S, OS )[−1] in degree 1), and to prove the following Lemma 4.1 The composition Rp∗TC / Rp∗f ∗TS Rp∗(ev) / Rp∗(ωC ⊗ H 2(S, OS )) vanishes in the derived category D(C). Proof. If C is smooth, the composition TC f ∗ψ / f ∗TS / f ∗Ω1 S ⊗ H 2(S, OS ) s⊗id / Ω1 C ⊗ H 2(S, OS ) is obviously zero, since TC ≃ TC in this case, and a curve has no 2-forms. For a general prestable C, we proceed as follows. Let’s consider the composition θ : TC f ∗ψ / f ∗TS / f ∗Ω1 S ⊗ H 2(S, OS ) s⊗id / Ω1 C ⊗ H 2(S, OS ) t⊗id / ωC ⊗ H 2(S, OS ) := L. On the smooth locus of C, H0(θ) is zero (by the same argument used in the case C smooth), hence the image of H0(θ) : H0(TC) ≃ TC → L is a torsion subsheaf of the line bundle L. But C is Cohen-Macaulay, therefore this image is 0, i.e. H0(θ) = 0; and, obviously, Hi(θ) = 0 for any i (i.e. for i = 1). Now we use the hypercohomology spectral sequences H p(C, Hq(TC)) ⇒ Hp+q(C, TC ) ≃ H p+q(RΓ(C, TC )), 21 / / / / / / / / / / / / / H p(C, Hq(L[0])) ⇒ Hp+q(C, L[0]) ≃ H p+q(RΓ(C, L[0])) ≃ H p+q(C, L), to conclude that the induced maps H i(RΓ(θ)) : H i(RΓ(C, TC )) −→ H i(RΓ(C, L)) ≃ H i(C, L) are zero for all i’s. Since C is a field, we deduce that the map RΓ(θ) = Rp∗(θ) is zero in D(C) as well. ✷ By the Lemma above, we have therefore obtained an induced map α : RΓ(C, Cone(TC → f ∗TS)) −→ H 2(S, OS )[−1]. Now, O-M-P-T reduced obstruction space is defined as ker H 1(α). Moreover, again by Lemma 4.1, we have an induced map v : Rp∗Cone(TC → f ∗TS) −→ Rp∗(ωC ⊗ H 2(S, OS )), and, since the canonical map Rp∗(ωC ⊗ H 2(S, OS)) −→ H 2(S, OS )[−1] obviously induces an isomorphism on H 1, we have that O-M-P-T reduced obstruction space is also the kernel of the map H 1(v) : H 1(RΓ(C, Cone(TC → f ∗TS)) −→ H 1(C, ωC ⊗ H 2(S, OS )). The following result proves the nontriviality of O-M-P-T reduced obstruction space. Proposition 4.2 If, as we are supposing, β 6= 0,the maps H 1(v), H 1(α), H 1(eα), and H 1(Rp∗(ev)) are all nontrivial, hence surjective. Proof The non vanishing of H 1(Rp∗(ev)) obviously implies all other nonvanishing state- ments, and the nonvanishing of H 1(Rp∗(ev)) is an immediate consequence of the following3. Lemma 4.3 Since the curve class β 6= 0, the map H 1(t ◦ s) : H 1(C, f ∗Ω1 S) −→ H 1(C, ωC ) is nonzero (hence surjective). Proof of Lemma. By [B-M, Cor. 2.3], β 6= 0 implies nontriviality of the map df : f ∗Ω1 Ω1 S → C. But S is a smooth surface and C a prestable curve, hence in the short exact sequence f ∗Ω1 S s / Ω1 C / Ω1 C/S → 0 the sheaf of relative differentials Ω1 points and thus its H 1 vanishes. Therefore the map C/S is concentrated at the (isolated, closed) singular H 1(s) : H 1(C, f ∗Ω1 S) −→ H 1(C, Ω1 C ) is surjective. The same argument yields surjectivity, hence nontriviality (since H 1(C, ωC ) has dimension 1 over C), of the map H 1(t) : H 1(C, Ω1 C ) → H 1(C, ωC ), by observing that, on the smooth locus of C, Ω1 C ≃ ωC and H 1(t) is the induced isomorphism. In particular, H 1(C, Ω1 C) 6= 0. Therefore both H 1(s) and H 1(t) are non zero and surjective, so the same is true of their composition. ♦ ✷ 3We thank R. Pandharipande for pointing out this statement, of which we give here our proof. 22 / / 4.2 The canonical projection RPic(S) −→ RSpec(Sym(H 0(S, KS)[1])) In this subsection we identify canonically the derived Picard stack RPic(S) of a K3-surface with Pic(S) × RSpec(Sym(H 0(S, KS )[1])), where KS := Ω2 S is the canonical sheaf of S; this allows us to define the canonical map prder : RPic(S) −→ RSpec(Sym(H 0(S, KS)[1])) red which is the last ingredient we will need to define the reduced derived stack RM g (S; β) of stable maps of genus g and class β to S in the next subsection. In the proof of the next Proposition, we will need the following elementary result (which holds true for k replaced by any semisimple ring, or k replaced by a hereditary commutative ring and E by a bounded above complex of free modules) Lemma 4.4 Let k be a field and E be a bounded above complex of k-vector spaces. Then there is a canonical map E → E<0 in the derived category D(k), such that the obvious composition E<0 −→ E −→ E<0 is the identity. Proof. Any splitting p of the map of k-vector spaces ker(d0 : E−1 → E0) ֒→ E−1 yields a map p : E → E<0 in the category Ch(k) of complexes of k-vector spaces. To see that different splittings p and q gives the same map in the derived category D(k), we consider the canonical exact sequences of complexes 0 → E<0 / E / E≥0 → 0 and apply Ext0(−, E<0), to get an exact sequence Ext0(E≥0, E<0) a / Ext0(E, E<0) b / Ext0(E<0, E<0). Now, the the class of the difference (p − q) in HomD(k)(E, E<0) = Ext0(E, E<0) is in the kernel of b, so it is enough to show that Ext0(E≥0, E<0) = 0. But E≥0 is a bounded above complex of projectives, therefore (e.g. [Wei, Cor. 10.4.7]) Ext0(E≥0, E<0) = 0 is a quotient of HomCh(k)(E≥0, E<0) which obviously consists of the zero morphism alone. ✷ Proposition 4.5 Let G be a derived group stack locally of finite presentation over a field k, e : Spec k → G its identity section, and g := TeG. Then there is a canonical map in Ho(dStk) γ(G) : t0(G) × RSpec(A) −→ G where A := k⊕(g is the trivial square zero extension of k by the complex of k-vector spaces (g ∨)<0 is the commutative differential non-positively graded k-algebra which ∨)<0. Proof. First observe that RSpec(A) has a canonical k-point x0 : Spec k → RSpec(A), corresponding to the canonical projection A → k. By definition of the derived cotangent 23 / / / / complex of a derived stack ([HAG-II, 1.4.1]), giving a map α such that α RSpec(A) fLLLLLLLLLL x0 Spec k / G <yyyyyyyyy e is equivalent to giving a morphism in the derived category of complex of k-vector spaces α′ : LG, e ≃ g ∨ −→ (g ∨)<0. Since k is a field, we may take as α′ the canonical map provided by Lemma 4.4, and define γ(G) as the composition t0(G) × RSpec(A) j×id / G × RSpec(A) id×α′ / G × G µ / G where µ is the product in G. ✷ Proposition 4.6 Let S be a K3 surface over k = C, and G := RPic(S) its derived Picard group stack. Then the map γ(G) defined in (the proof of ) Proposition 4.5 is an isomorphism γS := γ(RPic(S)) : Pic(S) × RSpec(Sym(H 0(S, KS )[1])) ∼ / / RPic(S) in Ho(dStC), where KS denotes the canonical bundle on S. Proof. Since G := RPic(S) is a derived group stack, γ(G) is an isomorphism if and only if it induces an isomorphism on truncations, and it is ´etale at e, i.e. the induced map T (t0(e),x0)(γ(G)) : T (t0(e),x0)(t0(G) × RSpec(A)) −→ Te(G) is an isomorphism in the derived category D(k), where x0 is the canonical canonical k-point Spec C → RSpec(A), corresponding to the canonical projection A → C. Since π0(A) ≃ C, t0(γ(G)) is an isomorphism of stacks. So we are left to showing that γ(G) induces an isomorphism between tangent spaces. Now, g ≡ Te(G) = Te(RPic(S)) ≃ RΓ(S, OS )[1], and, S being a K3-surface, we have g ≃ RΓ(S, OS)[1] ≃ H 0(S, OS )[1] ⊕ H 2(S, OS )[−1] so that (g ∨)<0 ≃ H 2(S, OS)∨[1] ≃ H 0(S, KS)[1] (where we have used Serre duality in the last isomorphism). But H 0(S, KS ) is free of dimension 1, so we have a canonical isomorphism C ⊕ (g ∨)<0 ≃ C ⊕ H 0(S, KS )[1] ≃ Sym(H 0(S, KS )[1]) 24 / f < / / / in the homotopy category of commutative simplicial C-algebras. Therefore T(t0(e),x0)(Pic(S) × RSpec(A)) ≃ g≤0 ⊕ g>0 ≃ H 0(S, OS )[1] ⊕ H 2(S, OS )[−1] and T(t0(e),x0)(γ(G)) is obviously an isomorphism (given, in the notations of the proof of Prop. 4.5, by the sum of the dual of α′ and the canonical map g≤0 → g). ✷ Using Prop. 4.6, we are now able to define the projection prder of RPic(S) onto its full derived factor as the composite RPic(S) / Pic(S) × RSpec(Sym(H 0(S, KS )[1])) pr2 γ(S)−1 / RSpec(Sym(H 0(S, KS)[1])). Note that prder yields on tangent spaces the canonical projection4 Te(RPic(S; β)) = g −→ g>0 = Tx0(RSpec(Sym(H 0(S, KS)[1]))) ≃ H 2(S, OS )[−1], where x0 is the canonical canonical k-point Spec C → Spec(Sym(H 0(S, KS)[1])), and g ≃ H 0(S, OS)[1] ⊕ H 2(S, OS )[−1]. 4.3 The reduced derived stack of stable maps RM red g (S; β) In this subsection we define the reduced version of the derived stack of stable maps of type (g, β) to S and describe the obstruction theory it induces on its truncation Mg(S; β). Let us define δder 1 (S, β) (respectively, δ(n), der 1 (S, β)) as the composition (see Def. 3.7 and Def. 3.9) RMg(S; β)  / RMg(S) δ1(S) / RPic(S) prder / / RSpec(Sym(H 0(S, KS )[1])) (resp. as the composition RMg,n(S; β)  / RMg,n(S) δ(n) 1 (S) / RPic(S) prder / / RSpec(Sym(H 0(S, KS)[1])) ). Definition 4.7 • The reduced derived stack of stable maps of genus g and class β to S RM red g (S; β) is defined by the following homotopy-cartesian square in dStC RM red g (S; β) / RMg(S; β) δder 1 (S,β) Spec C / RSpec(Sym(H 0(S, KS )[1])) • The reduced derived stack of n-pointed stable maps of genus g and class β to S RM red g,n(S; β) is defined by the following homotopy-cartesian square in dStC RM red g,n(S; β) / RMg,n(S; β) Spec C / RSpec(Sym(H 0(S, KS)[1])) δ(n),der 1 (S,β) 4Recall that, if M is a C-vector space, Tx0 (RSpec(Sym(M [1]))) ≃ M ∨[−1]. 25 / /  / /  / /   /   /   /   / Since the truncation functor t0 commutes with homotopy fiber products and t0(RSpec(Sym(H 0(S, KS )[1]))) ≃ Spec C, we get t0(RM red g (S; β)) ≃ Mg(S; β) red g (S; β) is a derived extension (Def. 1.1) of the usual stack of stable maps of type i.e. RM (g, β) to S, different from RMg(S; β). Similarly in the pointed case. We are now able to compute the obstruction theory induced, according to §1, by the red g (S; β). We leave to the reader the straightfor- closed immersion jred : Mg(S; β) ֒→ RM ward modifications for the pointed case. By applying Proposition 1.2 to the derived extension RM red g (S; β) of Mg(S; β), we get an obstruction theory j∗ red L RM red g (S;β) −→ L Mg(S;β) that we are now going to describe. Let ρ : RM red g (S; β) −→ RMg(S; β) be the canonical map. Since RM red g (S; β) is defined by the homotopy pullback diagram in Def. 4.7, we get an isomorphism in the derived category of RM red g (S; β) ρ∗(L RMg (S;β)/RSpec(Sym(H 0(S,KS)[1]))) ≃ L RM red g (S;β) . We will show below that RM red g (S; β) is quasi-smooth so that, by Corollary 1.3, j∗ red L RM red g (S;β) −→ L Mg(S;β) is indeed a perfect obstruction theory on Mg(S; β). Now, for any C-point Spec C → RMg(S; β), corresponding to a stable map (f : C → S) of type (g, β), we get a distin- guished triangle LRSpec(Sym(H 0(S,KS)[1])), x0 −→ L RMg(S;β), (f :C→S) −→ L RM red g (S;β), (f :C→S) red (where we have denoted by (f : C → S) also the induced C-point of RM g (S; β): recall that a derived stack and its truncation have the same classical points, i.e. points with values in usual commutative C-algebras) in the derived category of complexes of C-vector spaces. By dualizing, we get that the tangent complex Tred (f :C→S) := T(f :C→S)(RM red g (S; β)) of RM red g (S; β) at the C-point (f : C → S) of type (g, β), sits into a distinguished triangle Tred (f :C→S) / RΓ(C, Cone(TC → f ∗TS)) Θf / RΓ(S, OS)[1] pr / H 2(S, OS )[−1] , 26 / / / where Θf is the composite Θf : RΓ(C, Cone(TC → f ∗TS)) TxAX / / RHomS(Rf∗OC, Rf∗OC )[1] trS / / RΓ(S, OS)[1], and pr denotes the tangent map of prder taken at the point δ1(S)(f : C → S). Note that the map pr obviously induces an isomorphism on H 1. 4.4 Quasi-smoothness of RM g duced obstruction theory. red (S; β) and comparison with O-M-P-T re- In the case β 6= 0 is a curve class in H 2(S, Z), we will prove quasi-smoothness of the derived red stack RM g (S; β), and compare the induced obstruction theory with that of Okounkov- Maulik-Pandharipande-Thomas (see §4.1.1 or [M-P, §2.2] and [P1]). Theorem 4.8 Let β 6= 0 be a curve class in H 2(S, Z) ≃ H2(S, Z), f : C → S a stable map of type (g, β), and Tred (f :C→S) := T (f :C→S)(RM red g (S; β)) / RΓ(C, Cone(TC → f ∗TS)) / H 2(S, OS )[−1] the corresponding distinguished triangle. Then, 1. the rightmost arrow in the triangle above induces on H 1 a map H 1(Θf ) : H 1(C, Cone(TC → f ∗TS)) −→ H 2(S, OS ) which is nonzero (hence surjective, since H 2(S, OS) has dimension 1 over C). There- fore the derived stack RM red g (S; β) is everywhere quasi-smooth; 2. H 0(Tred (f :C→S)) (resp. H 1(Tred (f :C→S))) coincides with the reduced deformation space (resp. the reduced obstruction space) of O-M-P-T. Proof. First Proof of quasi-smoothness – Let us prove quasi-smoothness first. It is clearly enough to prove that the composite H 1(C, f ∗TS) / H 1(C, Cone(TC → f ∗TS)) H 1(TxAX ) / Ext2 S(Rf∗OC , Rf∗OC) H 2(trS ) / H 2(S, OS ) is non zero (hence surjective). Recall that p : C → Spec C and q : S → Spec C denote the structural morphisms, so that p = q ◦ f . Now, the map Rq∗TS −→ Rq∗Rf∗f ∗TS induces a map H 1(S, TS) → H 1(C, f ∗TS), and by Proposition 3.6 and Remark 3.8, the following diagram commutes H 1(S, TS ) <−,atRf∗OC > / Ext2 S(Rf∗OC , Rf∗OC) . H 1(TxAX ) H 1(C, f ∗TS) / H 1(C, Cone(TC → f ∗TS)) 27 / / / / /   / / O O So, we are reduced to proving that the composition a : H 1(S, TS) <−,atRf∗OC > / Ext2 S(Rf∗OC, Rf∗OC ) H 2(tr) / H 2(S, OS ) does not vanish. But, since the first Chern class is the trace of the Atiyah class, this composition acts as follows (on maps in the derived category of S) (ξ : OS → TS[1]) / (a(ξ) : OS c1⊗ξ / Ω1 S ⊗ TS[2] <−,−> / OS[2]) where c1 := c1(Rf∗OC) : OS −→ Ω1 S[1] is the first Chern class of the perfect complex Rf∗OC. What we have said so far, is true for any smooth complex projective scheme X in place of S. We now use the fact that S is a K3-surface. Choose a non zero section σ : OS → Ω2 S of the canonical bundle, and denote by ϕσ : Ω1 / TS the induced isomorphism. A straightforward linear algebra S computation shows then that the composition ∼ / OS ((ϕσ ◦c1)∧ξ)⊗σ / (TS ∧ TS ⊗ Ω2 S)[2] <−,−>[2] / OS[2] coincides with a(ξ). But, since β 6= 0, we have that c1 6= 0. σ is nondegenerate, so this composition cannot vanish for all ξ, and we conclude. Second Proof of quasi-smoothness – Let us give an alternative proof of quasi-smoothness. By Serre duality, passing to dual vector spaces and maps, we are left to proving that the composite H 0(S, Ω2 S) tr∨ / Ext0 S(RHom(Rf∗OC , Rf∗OC), Ω2 S) τ ∨ / Ext0(Rf∗f ∗TS[−1], Ω2 S) is non zero. So it is enough to prove that the map obtained by further composing to the left with the adjunction map Ext0(Rf∗f ∗TS[−1], Ω2 S ) −→ Ext0(TS[−1], Ω2 S) is nonzero. But this new composition acts as follows H 0(S, Ω2 S) ∋ (σ : OS → Ω2 S) 7→ (σ◦tr) 7→ (σ◦tr◦at) = (σ◦c1(Rf∗OC )) ∈ Ext0(TS [−1], Ω2 S) where at : TS[−1] → RHom(Rf∗OC, Rf∗OC ) is the Atiyah class of Rf∗OC (see Proposition 3.6 and Remark 3.8). Since β 6= 0, we have c1(Rf∗OC) 6= 0, and we conclude. Proof of the comparison – Let us move now to the second point of Thm. 4.8, i.e. the comparison statement about deformations and obstructions spaces. First of all it is clear that, for any β, H 0(T red, (f :C→S)) ≃ H 0(T RMg(S;β), (f :C→S)) ≃ H 0(C, Cone(TC → f ∗TS)) therefore our deformation space is the same as O-M-P-T’s one. Let us then concentrate on obstruction spaces. We begin by noticing the following fact 28 / / / / / / / / / Lemma 4.9 There is a canonical morphism in D(C) ν : Rp∗ωC ⊗L H 2(S, OS ) −→ Rq∗OS[1] inducing an isomorphism on H 1. Proof of Lemma. To ease notation we will simply write ⊗ for ⊗L. Recall that p : C → Spec C and q : S → Spec C denote the structural morphisms, so that p = q ◦ f . Since S is a K3-surface, the canonical map OS ⊗ H 0(S, Ω2 S) −→ Ω2 S is an isomorphism. Since f ! preserves dualizing complexes, ωS ≃ Ω2 we have S[2] and ωC ≃ ωC[1], ωC ≃ f !Ω2 S[1] ≃ f !(OS ⊗ H 0(S, Ω2 S))[1]. By applying Rp∗ and using the adjunction map Rf∗f ! → Id, we get a map Rp∗ωC ≃ Rq∗Rf∗ωC ≃ Rq∗Rf∗f !(OS [1]⊗H 0(S, Ω2 S)) → Rq∗(OS[1]⊗H 0(S, Ω2 S )) ≃ Rq∗OS[1]⊗H 0(S, Ω2 S ) (the last isomorphism being given by projection formula). Tensoring this map by H 0(S, Ω2 S)∨ ≃ H 2(S, OS ) (a canonical isomorphism by Serre duality), and using the canonical evalua- tion map V ⊗ V ∨ → C for a C-vector space V , we get the desired canonical map ν : Rp∗ωC ⊗ H 2(S, OS ) −→ Rq∗OS[1]. The isomorphism on H 1 is obvious since the trace map R1p∗ωC → C is an isomorphism (C is geometrically connected). ♦ If σ : OS ∼ / / Ω2 S is a nonzero element in H 0(S, Ω2 S), and ϕσ : TS ≃ Ω1 S the induced isomorphism, the previous Lemma gives us an induced map ν(σ) : Rp∗ωC −→ Rq∗OS [1], and an induced isomorphism H 1(ν(σ)) =: νσ : H 1(C, ωC ) ∼ / / H 2(S, OS ). Using the same notations as in §4.1.1, to prove that our reduced obstruction space ker(H 1(Θf ) : H 1(C, Cone(TC → f ∗TS)) −→ H 2(S, OS )) coincides with O-M-P-T’s one, it will be enough to show that the following diagram is commutative H 1(C, f ∗TS) can / H 1(RΓ(C, Cone(TC → f ∗TS))) can H 1(RΓ(C, Cone(TC → f ∗TS))) H 1(Θf ) H 1(v) H 1(C, ωC ) / H 2(S, OS ). ∼ νσ 29   /     / But this follows from the commutativity of Rp∗f ∗TS[−1] Rp∗(ϕσ) / Rp∗f ∗Ω1 S[−1] Rp∗(s) / Rp∗Ω1 C[−1] Rp∗(t) / Rp∗ωC[−1] id Rp∗f ∗TS[−1] TxAX / Rq∗RHomS(Rf∗OC, Rf∗OC ) tr whose verification is left to the reader. ν(σ)[−1] / Rq∗OS ✷ Remark 4.10 Note that by Lemma 4.2, the second assertion of Theorem 4.8 implies the first one. Nonetheless, we have preferred to give an independent proof of the quasismooth- red g (S; β) because we find it conceptually more relevant than the comparison ness of RM with O-M-P-T, meaning that quasi-smoothness alone would in any case imply the exis- tence of some perfect reduced obstruction theory on Mg(S; β), regardless of its comparison with the one introduced and studied by O-M-P-T. Moreover, we could only find in the literature a definition of O-M-P-T global reduced obstruction theory (relative to Mpre g ) with values in the τ≥−1-truncation of the cotangent complex of the stack of stable maps5, that uses a result on the semiregularity map whose proof is not completely convincing ([M-P, 2.2, formula (14)]); on the other hand there is a clean and complete description of the corresponding pointwise tangent and obstruction spaces. Therefore, our comparison is necessarily limited to these spaces. And our con- struction might also be seen as establishing such a reduced global obstruction theory - in the usual sense, i.e. with values in the full cotangent complex, and completely independent from any result on semiregularity maps. Theorem 4.8 shows that the distinguished triangle Tred (f :C→S) := T(f :C→S)(RM red g (S; β)) −→ RΓ(C, Cone(TC → f ∗TS)) −→ H 2(S, OS )[−1] induces isomorphisms H i(Tred (f :C→S)) ≃ H i(C, Cone(TC → f ∗TS)), for any i 6= 1, while in degree 1, it yields a short exact sequence 0 → H 1(Tred (f :C→S)) / H 1(C, Cone(TC → f ∗TS)) / H 2(S, OS ) → 0. red g (S; β) and RMg(S; β) (hence our induced reduced and So, the tangent complexes of RM the standard obstruction theories) only differ at the level of H 1 where the former is the kernel of a 1-dimensional quotient of the latter: this is indeed the distinguished feature of a reduced obstruction theory. The pointed case - In the pointed case, a completely analogous proof as that of Theorem 4.8 (1), yields 5The reason being that the authors use factorization through the cone, and therefore the resulting obstruction theory is only well-defined, without further arguments, if one considers it as having values in such a truncation. 30   / / /   / / / / Theorem 4.11 Let β 6= 0 be a curve class in H 2(S, Z) ≃ H2(S, Z). The derived stack red g,n(S; β) of n-pointed stable maps of type (g, β) is everywhere quasi-smooth, and there- RM fore the canonical map is a [−1, 0] perfect obstruction theory on Mg,n(X; β). j∗(L RMg,n(X;β)) −→ L Mg,n(X;β) 5 Moduli of perfect complexes In this Section we will define and study derived versions of various stacks of perfect com- plexes on a smooth projective variety X. If X is a K3-surface, by using the determinant map and the structure of RPic(X), we deduce that the derived stack of simple perfect complexes on X is smooth. This result was proved with different methods by Inaba in [In]. When X is a Calabi-Yau 3-fold we prove that the derived stack of simple perfect com- plexes (with fixed determinant) is quasi-smooth, and then use an elaboration of the map A(n) X : RMg,n(X) −→ RPerf (X) to compare the obstruction theories induced on the trun- cation stacks. This might be seen as a derived geometry approach to a baby, open version of the Gromov-Witten/Donaldson-Thomas comparison. Definition 5.1 Let X be a smooth complex projective variety, L a line bundle on X, and xL : Spec C → RPic(X) the corresponding point. • The derived stack RPerf (X)L of perfect complexes on X with fixed determinant L is defined by the following homotopy cartesian diagram in dStC RPerf (X)L RPerf (X) det Spec C / RPic(X) xL We will write RPerf (X)0 for RPerf (X)OX , the derived stack of perfect complexes on X with trivial determinant. • If Perf (X)≥0 denotes the open substack of Perf (X) consisting of perfect com- plexes F on X such that Exti(F, F ) = 0 for i < 0, we define RPerf (X)≥0 := φRPerf (X)(Perf (X)≥0) (as a derived open substack of RPerf (X), see Prop. 2.1). • If Perf (X)si,>0 denotes the open substack of Perf (X) consisting of perfect com- plexes F on X for which Exti(F, F ) = 0 for i < 0, and the trace map Ext0(F, F ) → H 0(X, OX ) ≃ C is an isomorphism, we define RPerf (X)si,>0 := φRPerf (X)(Perf (X)si,>0) (as a derived open substack of RPerf (X), see Prop. 2.1). • The derived stack RPerf (X)≥0 L is defined by the following homotopy cartesian dia- gram in dStC RPerf (X)≥0 L RPerf (X)≥0 . det Spec C / RPic(X) xL 31 / /     / / /     / As above, we will write RPerf (X)≥0 0 for RPerf (X)≥0 OX . • The derived stack MX ≡ RPerf (X)si,>0 L is defined by the following homotopy carte- sian diagram in dStC RPerf (X)si,>0 L RPerf (X)si,>0 . det Spec C xL / RPic(X) We will write RPerf (X)si,>0 0 for RPerf (X)si,>0 OX . Proposition 5.2 Let E be a perfect complex on X with determinant L, and xE : Spec C → RPerf (X)L the corresponding point. The tangent complex of RPerf (X)L at xE is REnd(E)0[1], the shifted traceless derived endomorphisms complex of E ([Hu-Le, Def. 10.1.4]) (so that H i(REnd(E)0) = ker(tr : Exti(E, E) → H i(X, OX )), for any i). Proof. Let T denote the tangent complex of RPerf (X)L at xE. By definition of RPerf (X)L, we have an exact triangle in the derived category D(C) of C-vector spaces T / REnd(E)[1] tr / RΓ(X, OX )[1] . By using the canonical map RΓ(X, OX ) → REnd(E) we can split this triangle, and we conclude. ✷ L and RPerf (X)si,>0 Remark 5.3 Since RPerf (X)≥0 Proposition 5.2 holds for their tangent complexes too. L are derived open substacks of RPerf (X)L, 5.1 On K3 surfaces By using the derived determinant map and the derived stack of perfect complexes, we are able to give another proof of a result by Inaba ([In, Thm. 3.2]) that generalizes an earlier work by Mukai ([Mu]). For simplicity, we prove this result for K3 surfaces, the result for a general Calabi-Yau surface being similar. Let S be a a smooth projective K3 surface, and let RPerf (S)si,>0 (Def. 5.1) be the open derived substack of RPerf (S) consisting of perfect complexes F on S for which Exti S(F, F ) → H 0(S, OS ) ≃ C is an iso- morphism. The truncation Perf (S)si,>0 of RPerf (S)si,>0 is a stack whose coarse moduli space Perf(S)si,>0 is exactly the moduli space Inaba calls Splcpx´et S(F, F ) = 0 for i < 0, and the trace map Ext0 S/C in [In, §3]. Coming back to Section 4.2, right after Prop. 4.6, we consider the projection prder of RPic(S) onto its full derived factor RPic(S) prder / / RSpec(Sym(H 0(S, KS)[1])). 32 / /     / / / Definition 5.4 The reduced derived stack RPerf (S)si,red of simple perfect complexes on S is defined by the following homotopy pullback diagram RPerf (S)si,red RPerf (S)si,>0 detS RPic(S) prder Spec C x0 / RSpec(Sym(H 0(S, KS )[1])) Since the truncation functor commutes with homotopy pullbacks, the truncation of RPerf (S)si,red is the same as the truncation of RPerf (S)si,>0, i.e. Perf (S)si,>0, therefore its coarse moduli space is again Inaba’s Splcpx´et S/C ([In, §3]). Theorem 5.5 The composite map RPerf (S)si,>0 detS / / RPic(S) prder / / RSpec(Sym(H 0(S, KS)[1])) is smooth. Therefore the derived stack RPerf (S)si,red is actually a smooth, usual (i.e. underived) stack, and RPerf (S)si,red ≃ t0(RPerf (S)si,red) ≃ Perf (S)si,>0. Under these identifications, the canonical map RPerf (S)si,red → RPerf (S)si,>0 becomes isomorphic to the inclusion of the truncation Perf (S)si,>0 → RPerf (S)si,>0. Proof. Let E be a perfect complex on S such that Exti trace map Ext0 defining RPerf (S)si,red yields a distinguished triangle of tangent complexes S(E, E) = 0 for i < 0, and the S(E, E) → H 0(S, OS ) ≃ C is an isomorphism. The homotopy fiber product TE RPerf (S)si,red / TE RPerf (S)si,>0 / H 0(S, KS )∨[−1] . Since TE RPerf (S)si,>0 ≃ REndS(E)[1], this complex is cohomologically concentrated in degrees [−1, 1]. Therefore, to prove the theorem, it is enough to show that the map (induced by the above triangle on H 1) α : Ext2 S(E, E) ≃ H 1(TE RPerf (S)si,>0) −→ H 0(S, KS)∨ is an isomorphism. If we denote by α′ : Ext2 S(E, E) α / / H 0(S, KS )∨ s ∼ / H 2(S, OS ) (the isomorphism s given by Serre duality), the following diagram Ext2 S(E, E) α′ H 2(S, OS ) Ext0 S(E, E)∨ tr∨ E / H 0(S, OS )∨ s s 33 / /       / / / / / /     / (where again, the s isomorphisms are given by Serre duality on S) is commutative. But, by hypothesis, trE is an isomorphism and we conclude. ✷ The following corollary was first proved by Inaba [In, Thm. 3.2]. Corollary 5.6 The coarse moduli space Perf(S)si,>0 of simple perfect complexes on a smooth projective K3 surface S is a smooth algebraic space. Proof. The stack RPerf (S)si,red ≃ Perf (S)si,>0 is a Gm-gerbe so its smoothness is equiv- alent to the smoothness of its coarse moduli space. ✷ Remark 5.7 Inaba shows in [In, Thm. 3.3 ] (again generalizing earlier results by Mukai in [Mu]) that the coarse moduli space Perf(S)si,>0 also carries a canonical symplectic structure. In [P-T-V-V] it is shown that in fact the whole derived stack RPerf(S) carries a natural derived symplectic structure of degree 0, and that this induces on Perf(S)si,>0 the symplectic structure defined by Inaba. Remark 5.8 The same argument used in proving Theorem 5.5 works by replacing S by the de Rham moduli space MDR(C), the Dolbeault moduli space MDol(C), for C a complex smooth projective curve, or by the Betti moduli space MB(S) represent- ing the homotopy type of an oriented compact topological surface S (see [Si], for the definitions of these moduli spaces). This yields smoothness of Perf(X)si,>0 for X = MDR(C), MDol(C), MB(S). 5.2 On Calabi-Yau threefolds In this section, for X an arbitrary complex smooth projective variety, we first elaborate on the map X : RMg,n(X) −→ RPerf (X) from Def. 3.9. This elaboration will give us a map C(n) X,L from a derived substack of RMg,n(X) to the derived stack RPerf (X)si,>0 , L being a line bundle on X (see Def. 5.1). When we specialize to the case where X is a projective smooth Calabi-Yau 3-fold Y , we prove that RPerf (Y )si,>0 Y,L allows us to compare the induced obstruction theories on the truncations of its source and target. is quasi-smooth (Proposition 5.11), and that the map C(n) L A(n) L To begin with, let X be a smooth complex projective variety. First of all, observe that taking tensor products of complexes induces an action of the derived group stack RPic(X) on RPerf (X) µ : RPic(X) × RPerf (X) −→ RPerf (X). Let xL : Spec C → RPic(X) be the point corresponding to a line bundle L on X. Definition 5.9 Let σL : RPic(X) → RPic(X) be the composite RPic(X) / RPic(X) × RPic(X) (inv,xL) × / RPic(X) where × (resp. σL(L1) = L ⊗ L−1 inv) denotes the product (resp. 1 ). the inverse) map in RPic(X) (shortly, 34 / / • Define A(n) X,L : RMg,n(X) → RPerf (X)L via the composite (det ◦A(n) X ,A(n) X ) RMg,n(X) / RPic(X) × RPerf (X) σL×id / RPic(X) × RPerf (X) µ / RPerf (X) (shortly, A(n) X,L(E) = E ⊗ (det E)−1 ⊗ L). • Define the derived open substack RMg,n(X)emb ֒→ RMg,n(X) as φRMg,n(X)(Mg,n(X)emb) (see Prop. 2.1) where Mg,n(X)emb is the open substack of Mg,n(X) consisting of pointed stable maps which are closed immersions. • Define C(n) X,L : RMg,n(X)emb → RPerf (X)si,>0 L via the composite RMg,n(X)emb  / RMg,n(X) A(n) X,L / RPerf (X)L (note that this composite indeed factors through RPerf (X)si,>0 L , since tr : Ext0(Rf∗OC , Rf∗OC) ≃ C, if the pointed stable map f is a closed immersion). Remark 5.10 The map C(n) X,L is also defined on the a priori larger open derived substack consisting (in the sense of Prop. 2.1) of pointed stable maps f such that the trace map tr : Ext0(Rf∗OC , Rf∗OC) → H 0(X, OX ) ≃ C is an isomorphism. We would like to use the map C(n) X,L to induce a comparison map between the induced obstruction theories on the truncations of RMg,n(X)emb and of RPerf (X)si,>0 This is possible when we take X to be a Calabi-Yau 3-fold Y. In fact: L . Theorem 5.11 If Y is a smooth complex projective Calabi-Yau 3-fold, then the derived stack RPerf (Y )si,>0 ֒→ RPerf (Y )si,>0 . L induces a [−1, 0]-perfect obstruction theory j∗T is quasi-smooth. Therefore, the closed immersion j : Perf (Y )si,>0 → T L RPerf (Y )si,>0 L Perf (Y )si,>0 L L L Proof. This is a corollary of Proposition 5.2. Let TE be the tangent complex of RPerf (Y )si,>0 at a point corresponding to the perfect complex E. Now Y is Calabi- Y ≡ KY ≃ OY ; but E is simple (i.e. the trace map Ext0(F, F ) → Yau of dimension 3, so Ω3 H 0(X, OX ) ≃ C is an isomorphism), so Serre duality implies Exti(E, E)0 = 0 for i ≥ 3 (and all i ≤ 0). Therefore the perfect complex TE is concentrated in degrees [0, 1], and RPerf (Y )si,>0 is quasi-smooth. The second assertion follows immediately from Prop. 1.2. L ✷ L Remark 5.12 Note that the stack Perf (Y )si,>0 is not proper over Spec C. However it receives maps from both Thomas moduli space In(Y ; β) of ideal sheaves (whose subschemes have Euler characteristic n and fundamental class β ∈ H2(Y, Z)) - see [Th] - and from Pandharipande-Thomas moduli space Pn(Y ; β) of stable pairs - see [P-T]. For example, the map from Pn(Y ; β) sends a pair to the pair itself, considered as a complex on Y . Moreover, at the points in the image of such maps, the tangent and obstruction spaces of these spaces are the same as those induced from the cotangent complex of our RPerf (Y )si,>0 (see e.g. [P-T, §2.1]). L 35 / / /  / / As showed in §1.2, the map C(n) induces a compar- ison map between the two obstruction theories. More precisely, the commutative diagram in dStC Y,L : RMg,n(Y )emb −→ RPerf (Y )si,>0 L Mg,n(Y )emb t0C(n) Y,L jGW Perf (Y )si,>0 L jDT RMg,n(Y )emb / RPerf (Y )si,>0 L C(n) Y,L (where each j is the closed immersion of the truncation of a derived stack into the full derived stack), induces a morphism of triangles (t0C(n) Y,L)∗j∗ DT L RPerf (Y )si,>0 L (t0C(n) Y,L)∗L Perf (Y )si,>0 L (t0C(n) Y,L)∗L RPerf (Y )si,>0 L /Perf (Y )si,>0 L j∗ GW L RMg,n(Y )emb / L Mg,n(Y )emb / L RMg,n(Y )emb/Mg,n(Y )emb - in the derived category of perfect complexes on Mg,n(Y )emb - i.e. a morphism relating the two obstruction theories induced on the truncations stacks Mg,n(Y )emb and Perf (X)si,>0 . Note that, for the object in the upper left corner of the above diagram, we have a natural isomorphism L (t0C(n) Y,L)∗j∗ DT L RPerf (Y )si,>0 L ≃ j∗ GW (C(n) Y,L)∗L RPerf (Y )si,>0 L . Remark 5.13 The motivation for constructing the above comparison morphism between the induced obstruction theories comes from the so called GW/DT comparison. The con- jectural comparison between Gromov-Witten and Donaldson-Thomas or Pandharipande- Thomas invariants for a general Calabi-Yau 3-fold Y (stated in [M-N-O-P], and proved in special cases, e.g. [B-B] and [M-O-O-P]) might be in principle approached by producing a map from the stack of pointed stable maps to Y to Thomas moduli space of ideal sheaves ([Th]) or to the Pandharipande-Thomas stack of stable pairs on Y ([P-T]), together with an induced map relating the corresponding obstruction theories. It is worth remarking that simply producing a map between these moduli spaces is not enough to relate the standard obstruction theories (the ones yielding the GW and DT or PT invariants, re- spectively): the notion of obstruction theory is not canonically attached to a stack, and therefore has not enough functoriality. This problem is completely solved in when the ob- struction theories come from derived extensions, as explained in §1. Therefore, one could hope to produce instead a map from the derived stack of pointed stable maps to Y to a derived extension of the moduli space of ideal sheaves on Y or of the stack of stable pairs on Y . We are not able to do this, at present. What we did above was to produce a map of derived stacks from the derived open substack RMg,n(Y )emb of RMg,n(Y ) consisting of stable maps which are closed immersions, to the derived stack RPerf (Y )si,>0 of simple perfect complexes with no negative Ext’s and fixed determinant L (for any L). As shown above, such a map automatically induces a morphism between the corresponding obstruc- tion theories, with no need of further data, and one might think of this as a baby, open L 36 / /  _   _   / / /   / /     / / version of the GW/DT comparison. In the presence of suitable compactifications RN of RMg,n(Y )emb and RP of RPerf (Y )si,>0 (or better of some derived substacks thereof, cut out by suitable cohomological or K- theoretic numerical conditions), to which the map C(n) Y,L extends as a quasi-smooth map F : RN → RP, the corresponding morphism between obstruction theories would give a canonical way of comparing the induced virtual fundamental classes, and therefore, the two counting invariants. More precisely, if f denotes the truncation of the morphism F , we would get ([Sch, Thm. 7.4] or Prop. 1.5) L f !([P]vir) = [N ]vir, where f ! : A∗(P) → A∗(N ) denotes the virtual pullback between Chow groups, as defined in [Man]. Of course the real technical heart of the GW/DT comparison lies exactly in the fine analysis of what happens at the boundary of the compactifications of the two stacks involved (in the above picture, the existence of an extension F : RN → RP), and for this our methods from derived algebraic geometry does not supply, at the moment, any new tool or direction. Remark 5.14 The group AutPerf(X) of self-equivalences of the derived category DPerf(X) of perfect complexes on X, acts on the derived stack RPerf (X). This action preserves, in an obvious sense, the induced obstruction theory on the stack Perf (X). It would be interesting to study how this action affects Bridgeland stability conditions on DPerf(X), and use it to draw consequences on the comparison of various counting invariants (e.g. Donaldson-Thomas and Pandharipande-Thomas ones). This will be the object of a future paper. A Derived stack of perfect complexes and Atiyah classes We explain here the relationship between the tangent maps associated to morphisms to the stack of perfect complexes and Atiyah classes (of perfect complexes) used in the main text (see §3.2). As in the main text, we work over C, even if most of what we say below holds true over any field of characteristic zero. As usual, all tensor products and fiber products will be implicitly derived. If Y is a derived geometric stack having a perfect cotangent complex ([HAG-II, §1.4]), and E is a perfect complex on Y, then we will implicitly identify the Atiyah class map of E with the corresponding map via the bijection atE : E −→ LY ⊗ E[1] TY −→ E∨ ⊗ E[1] [TY , E∨ ⊗ E[1]] ≃ [TY ⊗ E, E[1]] ≃ [E, LY ⊗ E[1]] given by the adjunction (⊗, RHom), and perfectness of E and LY (where [−, −] denotes the Hom set in the derived category of perfect complexes on Y). 37 We start with a quite general situation. Let Y be a derived geometric stack having a perfect cotangent complex, and Perf the stack of perfect complexes (viewed as a derived stack). Then, giving a map of derived stacks φE : Y → Perf is the same thing as giving a perfect complex E on Y, and • φ∗ E TPerf ≃ REndY (E)[1] • the tangent map to φE TφE : TY / φ∗ E TPerf ≃ REndY (E)[1] ≃ E∨ ⊗ E[1] is the Atiyah class map atE of E. Remark A.1 The second point above might be considered as a definition when Y is a derived stack, and it coincides with Illusie’s definition ([Ill, Ch. 4, 2.3.7]) when Y = Y is a quasi-projective scheme. In fact, in this case, the map ΦE factors through the stack Perf strict of strict perfect complexes; thus the proof reduces immediately to the case where E is a vector bundle on Y , which is straightforward. The above description applies in particular to a map of derived stacks of the form ΦE : Y := X × X −→ Perf where X is a smooth projective scheme, X is a derived geometric stack having a perfect cotangent complex, and E is a perfect complex on X × X : in the main text we are interested in X = RMg(X). Such a map corresponds, by adjunction to a map ΨE : X −→ RHOM(X, Perf ) =: RPerf (X). The tangent map of ΨE fits into the following commutative diagram TX can TΨE Ψ∗ E TRPerf (X) ∼ / RprX ,∗(E∨ ⊗ E)[1] RprX ,∗(TΦE ) RprX ,∗pr∗ X TX can / / RprX ,∗(pr∗ X TX ⊕ pr∗ X TX) / RprX ,∗TX ×X ∼ where can denote obvious canonical maps, and we can identify RprX ,∗(TΦE) with RprX ,∗(atE), in the sense explained above. In other words, TΨE is described in terms of the relative Atiyah class map atE/X : pr∗ X TX ≃ TX ×X/X → E∨ ⊗L E[1] of E relative to X, as the composition TΨE : TX can / / RprX ,∗pr∗ X TX ∼ / / RprX ,∗TX ×X/X RprX ,∗(atE/X ) / RprX ,∗(E∨ ⊗ E)[1]. Remark A.2 TΨE might be viewed at as a generalization of what is sometimes called the Kodaira-Spencer map associated to the X -family E of perfect complexes over X (e.g. [Ku-Ma, formula (14)]). 38 / / /   / / O O / In the main text, we are interested in the case X = RMg(X), pr := prX , and E perfect of the form Rπ∗E, where π : RCg; X −→ RMg(X) × X is the universal map and E is a complex on RCg; X, namely E = ORCg; X . In such cases, if we call (f : C → X) the stable map corresponding to the complex point x, we have a ladder of homotopy cartesian diagrams ιf x C f X q RCg; X π RMg(X) × X prX / / X pr q Spec C / RMg(X) x / Spec C and the base-change isomorphism (true in derived algebraic geometry with no need of flatness) gives us x∗E = x∗Rπ∗E ≃ Rf∗ι∗ f E. For E = ORCg; X , we then get x∗E = x∗Rπ∗ORCg; X ≃ Rf∗OC . Again by base-change formula, we get x∗Rpr∗ ≃ Rq∗x∗, and therefore the tangent map to AX := ΨRπ∗ORCg; X at x = (f : C → X), is the composition TxAX : TxRMg(X) ≃ RΓ(C, Cone(TC → f ∗TX )) / RΓ(X, x∗T RMg(X)×X ) RΓ(X,x∗atE ) / REndX (Rf∗OC)[1] ≃ TRf∗OC RPerf (X) The following is the third assert in Proposition 3.6, §3.2. Proposition A.3 The composition RΓ(X, TX ) can / / RΓ(X, Rf∗f ∗TX) can / / RΓ(X, Cone(Rf∗TC → Rf∗f ∗TX)) ≃ TxRMg(X) TxAX / / x∗A∗ X TRPerf (X) ≃ TRf∗OC RPerf (X) ≃ REndX (Rf∗OC)[1] coincides with RΓ(X, atRf∗OC ). Proof. We first observe that if F is perfect complex on X, and RAut(X) is the derived stack of automorphisms of X, there are obvious maps of derived stacks ρx : RAut(X) −→ RHOMdStC(C, X) σF : RAut(X) −→ RPerf (X) 39 / /     / /       / / / / / / / / induced by the natural action of RAut(X) by composition on maps and by pullbacks on perfect complexes, respectively. Moreover, the tangent map to σF at the identity Spec C-point of RAut(X) TidXσF : RΓ(X, TX ) ≃ TidX RAut(X) −→ TF RPerf (X) ≃ REndX (F)[1] is RΓ(X, atF ), where atF is the Atiyah class map of F. Then we observe that, by taking F := x∗Rπ∗ORCg; X - which is, by base-change formula, isomorphic to Rf∗OC - we get that the composition kx : RAut(X) ρx / / RHOMdStC(C, X) can / / RMg(X) AX / / RPerf (X) coincides with σF . But the map in the statement of the proposition is just TidX kx, and we conclude. ✷ References [Be] K. Behrend, Gromov-Witten invariants in algebraic geometry, Invent. Math. 127 (1997), 601–617. [B-B] K. Behrend, J. Bryan, Super-rigid Donaldson–Thomas invariants, Math. Res. Lett., vol. 14, pp. 559-571 (2007). [B-F] K. Behrend, B. Fantechi, The intrinsic normal cone, Invent. Math. (1) 128 (1997), 45-88. [B-M] K. Behrend, Y. Manin, Stacks of stable maps and Gromov-Witten invariants, Duke Math. Journal, Volume 85, Number 1 (1996), 1-60. [Bu-Fl] R.-O. Buchweitz, H. Flenner, A semiregularity map for modules and applications to deformations, Compositio Math. 137, 135-210 (2003). [Co] B. Conrad, Grothendieck Duality and Base Change, Lecture Notes in Mathematics, Vol. 1750, 2001, Berlin; New York; Springer-Verlag. [Hu-Le] D. Huybrechts, M. Lehn, The geometry of moduli spaces of sheaves, Second edi- tion, Cambridge University Press, Cambridge, 2010. [Ha-RD] R. Hartshorne, Residues and duality, Lecture Notes of a Seminar on the Work of A. Grothendieck, Lecture Notes in Mathematics, Vol. 20, 1966, Berlin; New York; Springer-Verlag. [Ill] L. Illusie, Complexe cotangent et d´eformations I, Lecture Notes in Mathematics 239, Springer Verlag, Berlin, 1971. [In] M. Inaba, Smoothness of the moduli space of complexes of coherent sheaves on an abelian or projective K3 surface, Advances in Mathematics, Volume 227, Issue 4, 10 July 2011, 1399-1412. 40 [CF-K] I. Ciocan-Fontanine, M. Kapranov, The derived Hilbert scheme, JAMS 15 (2002), 787-815. [Kle] S. Kleiman, The Picard scheme, in B. Fantechi, L. Gottsche, L. Illusie, S. L. Kleiman, N. Nitsure, A. Vistoli, Fundamental algebraic geometry. Grothendieck’s FGA ex- plained, Mathematical Surveys and Monographs 123, Amer. Math. Soc. 2005. [Ku-Ma] A. Kuznetsov, D. Markushevich, Symplectic structures on moduli spaces of sheaves via the Atiyah class, J. Geom. and Phys., 59(2009) 843-860. [O-P1] A. Okounkov and R. Pandharipande, Gromov-Witten theory, Hurwitz numbers, and Matrix models, In Algebraic geometry II, Seattle 2005. Part 2, volume 80 of Proc. Sympos. Pure Math., Amer. Math. Soc., Providence, RI, 2009. [O-P2] A. Okounkov and R. Pandharipande, Quantum cohomology of the Hilbert scheme of points of the plane, Invent. Math. 179 (2010), 523–557. [Ma] M. A. Mandell, An Inverse K-Theory Functor, Documenta Math. 15 (2010), 765-791. [Man] C. Manolache, Virtual pullbacks, Preprint arXiv:0805.2065. [M-N-O-P] D. Maulik, N. Nekrasov, A. Okounkov, R. Pandharipande, Gromov-Witten theory and Donaldson-Thomas theory I, Compositio Math. 142 (2006), 1263–1285. [M-O-O-P] D. Maulik, A. Oblomkov, A. Okounkov, R. Pandharipande, Gromov- Witten/Donaldson-Thomas correspondence for toric 3- folds, Invent. Math. (to ap- pear). [M-P] D. Maulik, R. Pandharipande, Gromov-Witten theory and Noether-Lefschetz theory, Preprint arXiv:0705.1653. [M-P-T] D. Maulik, R. Pandharipande, R. P. Thomas (with an appendix by A. Pixton), Curves on K3-surfaces and modular forms, Journal of Topology 3, 937-996, 2010. [Mu] S. Mukai, Symplectic structure of the moduli space of sheaves on an abelian or K3 surface, Inv. Math. Volume 77, Number 1, (1984), 101-116. [P1] R. Pandharipande, Maps, sheaves and K3-surfaces, Clay Research Lectures (to ap- pear). [P-T] R. Pandharipande, R. P. Thomas, Curve counting via stable pairs in the derived category, Inventiones Math 178, 407-447, 2009. [P-T-V-V] T. Pantev, B. Toen, G. Vezzosi, M. Vaqui´e, Derived symplectic structures, Preprint, November 2011. [Ran] Z. Ran, Semiregularity, obstructions and deformations of Hodge Classes, Ann. Scuola Norm. Pisa, Vol. XXVIII (1999), 809-821. [Sch] T. Schurg, Deriving Deligne-Mumford stacks with perfect obstruction theories, PhD Thesis, Johannes Gutenberg Universitat Mainz, July 2011. 41 [SGA6] P. Berthelot, A. Grothendieck, L. Illusie, eds. (1971) S´eminaire de G´eom´etrie Alg´ebrique du Bois Marie - 1966-67 - Th´eorie des intersections et th´eor`eme de Riemann-Roch, Lecture notes in mathematics 225, Berlin; New York: Springer- Verlag. xii+700. [Si] C. Simpson, Moduli of representations of the fundamental group of a smooth projective variety I, Publications Math´ematiques de l’IHES, 79 (1994), p. 47-129. [Th] R. P. Thomas, A holomorphic Casson invariant for Calabi-Yau 3-folds, and bundles on K3 fibrations, Jour. Diff. Geom. 54, no. 2, 367-438, 2000. [To-1] B. Toen, Champs affines, Selecta Math. (N.S.) 12 (2006), no. 1, 39–135. [To-2] B. Toen, Higher and derived stacks: a global overview, In Algebraic geometry I, Seattle 2005. Part 1, volume 80 of Proc. Sympos. Pure Math., pages 435-487. Amer. Math. Soc., Providence, RI, 2009. [HAG-II] B. Toen, G. Vezzosi, Homotopical algebraic geometry II: Geometric stacks and applications, Mem. Amer. Math. Soc. 193 (2008), no. 902, x+224 pp. [To-Ve] B. Toen, G. Vezzosi, Alg`ebres simpliciales S1-´equivariantes, th´eorie de de Rham et th´eor`emes HKR multiplicatifs, to appear in Compositio Math. [Wal] F. Waldhausen, Algebraic K-theory of spaces, Algebraic and geometric topology (New Brunswick, N.J., 1983), 318–419, Lecture Notes in Math., 1126, Springer, Berlin, 1985 [Wei] C. Weibel, An introduction to homological algebra, Cambridge Studies in Advanced Mathematics, Cambridge University Press, Cambridge, 1995. 42 

1310.6684

5

1310

Tropical approach to Nagata's conjecture in positive characteristic

Suppose that there exists a hypersurface with the Newton polytope $\Delta$, which passes through a given set of subvarieties. Using tropical geometry, we associate a subset of $\Delta$ to each of these subvarieties. We prove that a weighted sum of the volumes of these subsets estimates the volume of $\Delta$ from below. As a particular application of our method we consider a planar algebraic curve $C$ which passes through generic points $p_1,\dots,p_n$ with prescribed multiplicities $m_1,\dots,m_n$. Suppose that the minimal lattice width $\omega(\Delta)$ of the Newton polygon $\Delta$ of the curve $C$ is at least $\max(m_i)$. Using tropical floor diagrams (a certain degeneration of $p_1,\dots, p_n$ on a horizontal line) we prove that $$\mathrm{area}(\Delta)\geq \frac{1}{2}\sum_{i=1}^n m_i^2-S,\ \ \text{where } S=\frac{1}{2}\max \left(\sum_{i=1}^n s_i^2 \Big| s_i\leq m_i, \sum_{i=1}^n s_i\leq \omega(\Delta)\right).$$ In the case $m_1=m_2=\ldots =m\leq \omega(\Delta)$ this estimate becomes $\mathrm{area}(\Delta)\geq \frac{1}{2}(n-\frac{\omega(\Delta)}{m})m^2$. That rewrites as $d\geq (\sqrt{n}-\frac{1}{2}-\frac{1}{2\sqrt n})m$ for the curves of degree $d$. We consider an arbitrary toric surface (i.e. arbitrary $\Delta$) and our ground field is an infinite field of any characteristic, or a finite field large enough. The latter constraint arises because it is not {\it \`a priori} clear what is {\it a collection of generic points} in the case of a small finite field. We construct such collections for fields big enough, and that may be also interesting for the coding theory.

math.AG

math

TROPICAL APPROACH TO NAGATA'S CONJECTURE IN POSITIVE CHARACTERISTIC NIKITA KALININ Abstract. Suppose that there exists a hypersurface with the Newton polytope ∆, which passes through a given set of subvarieties. Using tropical geometry, we associate a subset of ∆ to each of these subvarieties. We prove that a weighted sum of the volumes of these subsets estimates the volume of ∆ from below. As a particular application of our method we consider a planar algebraic curve C which passes through generic points p1, . . . , pn with prescribed multiplicities m1, . . . , mn. Suppose that the mini- mal lattice width ω(∆) of the Newton polygon ∆ of the curve C is at least max(mi). Using tropical floor diagrams (a certain degeneration of p1, . . . , pn on a horizontal line) we prove that n si ≤ ω(∆)! . 2 (n − ω(∆) m )m2. That n m2 1 2 area(∆) ≥ Xi=1 In the case m1 = m2 = . . . = m ≤ ω(∆) this estimate becomes area(∆) ≥ 1 Xi=1 rewrites as d ≥ (√n − 1 2 − 1 2√n )m for the curves of degree d. i − S, where S = si ≤ mi, s2 i(cid:12)(cid:12)(cid:12) 1 2 max n Xi=1 We consider an arbitrary toric surface (i.e. arbitrary ∆) and our ground field is an infinite field of any characteristic, or a finite field large enough. The latter constraint arises because it is not `a priori clear what is a collection of generic points in the case of a small finite field. We construct such collections for fields big enough, and that may be also interesting for the coding theory. 1. Main Theorem and a discussion around Nagata's conjecture It is easy to find a polynomial in one variable with prescribed values at given points. Then, it is not difficult to find a polynomial in many variables with prescribed values at given points, or to find a polynomial in one variable with prescribed higher derivatives at given points. Each of the conditions appearing above imposes one linear constraint on the polynomial's coefficients. When we have only linear constraints, it is natural to ask whether they are mutually independent. In the above problems it is indeed the case, but in general it is not always true and can be a source of major difficulties. Consider the following general question: given natural numbers m1, m2, . . . , mn and a set of varieties X1, X2, . . . , Xn ⊂ Fk (where F is a field of any characteristic), we are wondering if there exists a hypersurface Y ⊂ Fk (with a given Newton polytope ∆) which passes through each of the Xi with the multiplicity mi respectively. This procedure (defining a variety by incidence and tangence relations) helps in constructing concrete examples and counter-examples in the realm of singular varieties. This paper promotes the tropical point of view on the above problem. We define the subsets Infl(Xi) of ∆, "influenced" by each of the Xi. These subsets can overlap, but no more than k at once (Corollary 4.17). We mainly concentrate on the case k = dim Y + 1 = 2, i.e. Y is a planar algebraic curve and each of the Xi is a point. 1.1. Main Theorem. A vector (u1, u2) ∈ Z2 is primitive if gcd(u1, u2) = 1. In particular, a primitive vector (u1, u2) cannot be (0, 0). Denote by P (Z2) the set of primitive vectors in Z2. The lattice width ωu(∆) of a polygon ∆ ⊂ Z2 in a direction u = (u1, u2) ∈ P (Z2) is max (u1, u2)· (x− y). x,y∈∆ Definition 1.1. The minimal lattice width ω(∆) of a polygon ∆ ⊂ Z2 is min ωu(∆). u∈P (Z2) Date: July 24, 2018. Key words and phrases. Nagata's conjecture, m-fold point, floor diagrams, tropical geometry. 1 2 N. KALININ The following theorem is an application of general methods developed in this article, bred with results of [15]. Theorem 1.2. Let F be an infinite field or a field big enough (for details see Lemma 4.23). If ω(∆) ≥ max(mi) and for each set of points p1, p2, . . . , pn ∈ (F∗)2 there exists an algebraic curve C ⊂ (F∗)2 with the Newton polygon ∆, passing through p1, p2, . . . , pn with multiplicities m1, m2, . . . , mn correspondingly, then (1) area(∆) ≥ 1 2 nXi=1 m2 i! − S(cid:0)m1, m2, . . . , mn, ω(∆)(cid:1). The correction term S in Eq.(1) is given by the following formula. Definition 1.3. For a set m1, m2, . . . , mn ∈ Z>0 we define (2) S(m1, . . . , mn, M ) = 1 2 max nXi=1 i! s2 i=1 such that 0 ≤ si ≤ mi,Pn where we maximize by all sets of numbers {si}n i=1 si ≤ M . Example 1.4. If m1 = m2 = ··· = mn = m, M = [√n]m, then S(m1, . . . , mn, M ) = 1 Indeed, if for a, b ≥ 0 the sum a + b is fixed, then a2 + b2 is bigger when a and b are maximally Corollary 1.5. Under the hypothesis of Theorem 1.2, if m1 = m2 = ··· = mn = m, ω(∆) = [√n]m, then area(∆) ≥ 1 Corollary 1.6. Under the hypothesis of Theorem 1.2, suppose that far from each other. Using this example we obtain the following corollary of Theorem 1.2. 2 (n − [√n]) m2. 2 [√n]m2. m1 = m2 = ··· = mn = m ≤ ω(∆). 2 (n − ω(∆) m )m2. using the argument in Example 1.4, that the minimum is attained when Then, the following inequality takes place: area(∆) ≥ 1 Proof. We are seeking for the minimum of nm2 − 2S(m, . . . , m, ω(∆)) = Pn conditionsP si ≤ ω(∆), si ≤ m. Choose k, k′ ∈ Z such that ω(∆) = mk + k′, 0 ≤ k′ < m. We see, HencePn sj = m, if 1 ≤ j ≤ k, and 0 ≤ sk+1 = k′ < m, and sj = 0 for j > k + 1. i=1(m2 − s2 i=1(m2 − s2 i ) under ω(∆) 1 i ) ≥ nm2 − km2 − k′2. Therefore, 2(cid:0)nm2 − km2 − k′2(cid:1) ≥ area(∆) ≥ 1 2(cid:18)n − m (cid:19) m2, because k + k′2 m2 ≤ k + k′ m < k + 1. The equality in the right hand side takes place if k′ = 0. (cid:3) Corollary 1.7. If ∆ is the triangle ConvHull(cid:0)(0, 0), (d, 0), (0, d)(cid:1), then the above corollary gives d ≥ (√n − 1 2 . So, we have d2 ≥ (n − d/m)m2. If d ≥ m√n, then we are done. Proof. Indeed, area(∆) = d2 Suppose that d < m√n, then 2 − 1 2√n )m. d2 ≥ (n − d/m)m2 ≥ (n − √n)m2 ≥(cid:18)√n − 2√n(cid:19)2 2√n )2 = n − √n − (1 − 1/4 − 1/4n − 1/2√n) ≤ n − √n. 2 − 1 1 2 − 1 because (√n − 1 m2, (cid:3) TROPICAL APPROACH TO NAGATA'S CONJECTURE IN POSITIVE CHARACTERISTIC 3 1.2. The idea of proof. Let K be the field of rational functions over F. Since each element of K, except zero, can be written as cktkm ∈ Z, ck ∈ F) , we can define a valuation map val : K → T by the rule val(cid:18) ∞Pm ( ∞Xm cktk(cid:19) := − min{k, ck 6= 0} and In other words, val(f ) is minus the order of vanishing of f at zero. There are val(0) := −∞. different extensions of K, algebraically closed, with surjective map val; one can take a field of power series whose elements converge near zero, etc; see [19, 23]. of multiplicity mi splits into two parts mi = si + ri, such that Pn We prove that Theorem 1.2 holds over this valuation field K. We use the nature of a singular point's influence on the Newton polygon of a curve (this "influence" means that a part Infl(pi) of ∆ corresponds to each point pi, see Section 3) and tropical floor diagrams [6, 7]. Tropical floor diagrams illustrate the process of a degeneration of the points p1, . . . , pn on a horizontal line, in a sense it is a tropical version of the Horace method [12]. While degenerating p1, p2, . . . , pn onto a line, we see the following behavior (Figure 3) of the points on the tropical picture. Each point i=1 si ≤ ω(∆). Furthermore, we choose a part of Infl(pi) ⊂ ∆ for each i = 1, . . . , n; these parts do not intersect, and the area of such a part for a point pi is at least 1 Then we prove Detropicalization lemma. It says that if Theorem 1.2 holds over K and does not hold over F, then there exists a non-zero polynomial of bounded degree which has all points of F as its roots. This implies that there exists a constant N ∈ N such that if the cardinality of F is at least N (which is always the case if F is infinite), then Theorem 1.2 holds for F. This is a natural restriction, because in small fields we cannot find a sufficiently generic collection of points. The constant N , then, depends on max(mi), ∆, and char(F). This reasoning could be of a particular interest to coding theory, see Section 6. i − s2 i ). 2 (m2 1.3. Nagata's conjecture. Let us fix a field F. For a point p = (x1, y1) ∈ F2 we denote by Ip the ideal of the point p, namely Ip = hx − x1, y − y1i. Definition 1.8. Consider an algebraic curve C given by an equation F (x, y) = 0, F ∈ F[x, y]. We say that p is of multiplicity at least m for C (and write µp(C) ≥ m), if F ∈ (Ip)m. In the most non-degenerate case, "p is a point of multiplicity m on C" means that there are m branches of C passing through p. For fields of characteristic zero, two following conditions are equivalent: 1) F ∈ (Ip)m, and 2) all the partial derivatives of F up to order m − 1 vanish at p. Example 1.9. Consider a planar algebraic curve C of degree d given by an equation F (x, y) = 0, where F (x, y) = Xi,j≥0,i+j≤d aijxiyj. The point p = (0, 0) is of multiplicity m for C if and only if for all i, j ≥ 0 with i + j < m we have aij = 0. As a consequence, for each point p ∈ F2 the condition µp(C) ≥ m can be rewritten as a system of m(m+1) linear equations in the coefficients {aij} of F . 2 Let p1, . . . , pn be a collection of n > 9 points in F2 and m1, . . . , mn ∈ N. We are looking for the minimal degree dmin of an algebraic curve passing through p1, . . . , pn with multiplicities at least m1, . . . , mn respectively. One can naively calculate the expected dimension edim(d, m1, . . . , mn) of the space S of the curves constraints of degree d satisfying the hypothesis above. Indeed, each singular point imposes m(m+1) on the coefficients of the curve equation, therefore, 2 edim(d, m1, . . . , mn) = max −1, d(d + 3) 2 − nXi=1 mi(mi + 1) 2 ! . 4 N. KALININ The actual dimension of S is always at least the expected one, because all the constraints are linear. However, sometimes even for a generic choice of the set of points p1, p2, . . . , pn the actual dimension is strictly greater than the expected one. As a reasonable estimate for dmin, Nagata's conjecture claims: mi. nPi=1 Conjecture 1. If points p1, . . . , pn ∈ P2, n > 9 are chosen generically and d ≤ 1√n dim S = −1. In other words, dmin > 1√n Example 1.10. Let us consider two points p1, p2. The minimal degree of a curve passing through p1, p2 with multiplicities m1, m2 is m1, if m1 ≥ m2: it is the line passing through p1 and p2 taken with multiplicity m1. So the inequality dmin ≥ m1+m2√2 in Nagata's conjecture is not satisfied as long as m2 > m1(√2 − 1). For five generic points pi with multiplicities mi, i = 1, . . . , 5, one can draw a conic through p1, p2, p3, p4, p5, and take it with multiplicity max(mi). In the case all mi = m, this example also violates the inequality in Nagata's conjecture. A classification for the case of less than ten points can be found in ([25], Example 2.4, [8], Proposition 5.8, Theorem 5.9). mi, then nPi=1 The case n = l2 had been proven by Nagata himself [21]. These days, even the case n = 10 and m1 = m2 = ··· = m10 = m is under exhaustive study ([10]), but has not yet been proven. The similar questions in higher dimensions are widely open (cf. [3],[11]). The pictures appeared in our approach are somewhat similar to those in [22], though the relation is not direct. Historically Nagata's conjecture appeared as a tool (with n = 16) to disprove Hilbert 14th prob- lem. Related to Nagata's conjecture is the Segre-Harbourne-Hirschowitz's conjecture which basically says that if the expected dimension edim of S is not equal to the actual one, then the linear system S contains a rational curve in its base locus. The reader is kindly referred to look into surveys [8, 9, 14, 20] for an introduction to Nagata's conjecture and related topics. In view of Theorem 1.2 the following three results should be mentioned: Theorem ([29], Xu). If C is a reduced and irreducible curve of degree d passing through generically chosen points p1, p2, . . . , pn ∈ CP 2 with multiplicities m1, m2, . . . , mn respectively, then the estimate d2 ≥Pn Xu's theorem can be verbatim extended to curves with arbitrary Newton polygons, and to re- ducible and non-reduced curves. But unlike Xu's theorem, which requires characteristic zero, in Theorem 1.2 we consider curves defined over fields of any characteristic. i − min(mi) holds. i=1 m2 Theorem ([1], Alexander, Hirschowitz). The dimension of the space of degree d > 2 hypersurfaces in CP k(k ≥ 3), passing through generic points p1, p2, . . . , pn with multiplicities m1 = ··· = mn = 2, is the expected one except for the cases (k, d, n) = (2, 4, 5), (3, 4, 9), (4, 4, 14), (4, 3, 7). Using the classification of tropical singular surfaces in [18], we give a sketch of a proof that the volume V of the Newton polytope of a surface in CP 3 with n two-fold points in general position satisfies n ≤ 2V . Using the above theorem we can obtain a better estimate, see Remark 4.20 for details. Theorem ([2], Alexander, Hirschowitz). For each field F, the dimension of degree d hypersurfaces in FP k passing through generic points p1, p2, . . . , pn with multiplicities m1, m2, . . . , mn is the expected one if d ≫ max mi. will lead to an explicit degree estimate in these cases. We expect that our approach can be extended to the cases k ≥ 3 and mi > 2. Such an extension Research is supported by the grant 168647 (PostDoc.Mobility) of the Swiss National Science Foun- dation. I would like to thank the anonymous reviewers for numerous suggestions which improved this paper a lot. 2. Preliminaries in tropical geometry In this section we recall some definitions and set up the notation. We refer the reader to [5],[17] for a general introduction to tropical geometry. TROPICAL APPROACH TO NAGATA'S CONJECTURE IN POSITIVE CHARACTERISTIC 5 Let T denote R ∪ {−∞}, and K be a field with a valuation map val : K → T. We use the convention val(a + b) ≤ max(val(a), val(b)), val(0) = −∞. Usually T is called the tropical semi-ring. Consider a hypersurface Y ⊂ Kk. Let Y be given by an equation F (x1, x2, . . . , xk) = 0, cI xI , I = (i1, i2, . . . , ik), cI 6= 0. In such case ∆ = ConvexHull(A) is called the Newton polytope of Y . consider the extended Newton polytope of Y , F =XI∈A The Newton polytope of F is provided with a subdivision deined by the coefficients of F . Namely, a subdivision of ∆. e∆ = ConvexHull{(I, x) ∈ Zk × TI ∈ A, x ≤ val(cI ))}. The projection of the faces of the extended Newton polytope e∆ onto the Newton polytope ∆ defines We give a definition of the tropicalization of Y , based on its equation F (x) =PI∈A weight ω = (w1, w2, . . . , wk) ∈ Tk we consider the weight function cI xI . For a ω(cxi1 1 xi2 2 . . . xik k ) := val(c) + i1w1 + i2w2 + ··· + ikwk. Then we define initial part inω(F ) of F as the ω-maximal part of F . Namely, we find W , the maximal weight (with respect to ω) among monomials of F . Then, by definition, inω(F ) is the sum of all monomials of F with weight W with respect to ω. Finally, we define Trop(Y ) ⊂ Tk to be the set of all weights ω such that inω(F ) is not a monomial. This is same as define Trop(Y ) as the corner locus of the function Trop(F )(ω) = max I∈A (val(cI + I · ω)). The set Trop(Y ) is a polyhedral complex. Indeed, for each subset S of the set of monomials of F we can consider the set S∗ of ω ∈ Tk such that the set of monomials in inω(F ) is S. One can prove that each of sets S∗ is convex. Therefore this defines a polyhedral subdivision of Tk, whose cells are parametrized by subsets S of the set of monomials of F . Adopting this point of view we see that Trop(Y ) is the union of cells in the above subdivision, which correspond to S with S ≥ 2. This (Grobner) subdivision of Tk is related with the subdivision (described earlier) of the Newton polytope ∆ in the following way. A point I ∈ ∆ is a vertex of the subdivision of ∆ if there exists such a weight ω0 ∈ Tk that inω0(F ) = cI xI . So we say that the cell {ω′ ∈ Tkinω0(F ) = inω′(F ) = cI xI} is dual to the vertex I in the subdivision of ∆. An interval I1I2 between two vertices I1, I2 ∈ ∆ is an edge of the subdivision of ∆ if there exists a weight ω0 such that inω0(F ) =PI∈J cI xI where the convex hull of J is the interval I1I2. Again, we say that the cell {ω′ ∈ Tkinω′(F ) = inω0(F )} is dual to the interval IJ of the subdivision of ∆. We further define this duality between the cells of dimension i in the subdivision of ∆ and cells of dimension k − i in the subdivision of Tk similarly. We can summarize the above arguments as follows. Remark 2.1. Each cell of the subdivision of ∆ is of the type ∆ω = ConvexHull(support(inω(F ))) for some ω ∈ Tk. We recall that the support of a polynomial is the set of multipowers of its monomials. For example, support(x2 + t−2xy + 3y3) is the set {(2, 0), (1, 1), (0, 3)}. Remark 2.2. If Y ⊂ Kk is a hypersurface, then Trop(Y ) ⊂ Tk is a polyhedral complex of codi- mension one. For each cell ∆ω ⊂ ∆ of the subdivision of the Newton polytope ∆ of Y we define d(∆ω) = {ω′ ∈ Tk∆ω = ∆ω′}. This map d provides the following correspondence: the vertices of the subdivision of ∆ correspond to the connected components of the complement of Trop(Y ) in Tk, the edges of the subdivision correspond to the faces of Trop(Y ) of the maximal dimension, the two-cells of the subdivision correspond to the faces of codimension one in Trop(Y ), etc. A particular example of such a duality 6 N. KALININ for k = 2 is presented in Figure 1. Since d is a bijection between cells, abusing notation we also write ∆ω = d({ω′ ∈ Tk∆ω = ∆ω′}). Definition 2.3. If X ⊂ Kn is a variety of higher codimension, we define its tropicalization Trop(X) as follows. Let I be the ideal of X. Let inω(I) be the ideal generated by the elements inω(f ), f ∈ I. Then, by definition, ω ∈ Trop(X) if and only if inω(I) is monomial free. A proof that Trop(X) is a polyhedral complex repeats the above arguments for the case of hypersurface, see [17] for details. 3. An estimate of a singular point's influence on the Newton polygon of a curve In this section we cover facts from [15] that we need. Let C be a curve over a valuation field K with the Newton polygon ∆ such that ω(∆) ≥ m. Let Q ∈ Trop(C). Q • Q • (a) if Q is not a vertex (b) if Q is a vertex Figure 1. If Q is not a vertex of Trop(C) (left column), then the collection I(Q) of vertices consists of all the vertices of Trop(C) lying on the extension of the edge through Q. If Q is a vertex of Trop(C) (right column), then I(Q) is the set of the vertices on the extensions of all the edges through Q. In each case the corresponding set Infl(Q) of faces of the subdivision of ∆, the "region of influence" of Q, is drawn at the top. Definition 3.1. Let lQ(u) be the line through Q in the direction u ∈ P (Z2). Take the connected component, containing Q, of the intersection Trop(C)∩ lQ(u). We call this component the long edge through Q in the direction u and denote it by EQ(u). Definition 3.2. For each u ∈ P (Z2) we denote by IQ(u) the set of vertices of Trop(C) which belong to the long edge EQ(u). Define I(Q) =Su∈P (Z2) IQ(u). Note that I(Q) is not a multiset; it contains only one copy of Q. Examples of I(Q) are presented in Figure 1. On the left we see one long edge EQ((1, 0)) and I(Q) consists of 7 vertices, and above we see 7 corresponding faces in the subdivision of ∆. On the right, we see long edges EQ((1, 0)), EQ((0, 1)), EQ((−1, 1)). Each of the long edges EQ((1, 2)) and EQ((−3,−2)) consists of only one edge. Definition 3.3 ([15]). For a point Q ∈ Trop(C) we define the region of influence of Q Infl(Q) = [V ∈I(Q) d(V ), TROPICAL APPROACH TO NAGATA'S CONJECTURE IN POSITIVE CHARACTERISTIC 7 the union of the faces of the Newton polygon of Trop(C), dual to the vertices in I(Q). Definition 3.4. For a point Q ∈ Trop(C) which is not a vertex of Trop(C), we define area(Infl(Q)) = XF∈Infl(Q) area(F ), the sum of the areas of faces F in the region of influence of Q, see Figure 1. Note that area(Infl(Q)) depends only on Trop(C) and does not depend on a particular choice of an equation defining C. Also, if Q belongs to an edge E of Trop(C) and Q is not a vertex of Trop(C), then IQ(u) = I(Q) where u is the direction of E. Indeed, for any other direction v not collinear to u, the connected component of Q in the intersection Trop(C) ∩ lQ(v) is just Q. Recall that if Q is a vertex of Trop(C), then d(Q) is a face dual to Q in the subdivision of ∆. Definition 3.5. If Q is a vertex of Trop(C), we define area(Infl(Q)) = XF∈Infl(Q) area∗(Infl(Q)) = XF∈Infl(Q) area(F ) + area(d(Q)), area(F ). Theorem 3.6 ([15], Lemma 2.8, Theorems 1,2). Suppose that a point p = (p1, p2) ∈ (K∗)2 is of multiplicity m for this curve C, P = Val(p) = (val(p1), val(p2)). Suppose also that the Newton polygon ∆ of C satisfies ω(∆) ≥ m. Then, (3) . area(Infl(P )) ≥ m2 2 2 m2 in (A) and at least 3 If the point Q in Figure 1 is as in this theorem, then the sum of the areas of the faces is at 8 m2(= area∗(Infl(Q))) in (B) (but we will not consider area∗ in this least 1 article). Example 3.7. Consider a curve C given by the equation (x − 1)k(y − 1)m−k = 0, take p = (1, 1). Clearly, µp(C) = m, but the Newton polygon ∆ of C violates the condition ω(∆) ≥ m, and the inequality area(Infl(Val(p))) = 2k(m − k) ≥ m2 Lemma 3.8 ([15], Lemma 1.25). If µ(1,1)(C) = m and for the Newton polygon ∆ of C we have ωu(∆) = m− a for some a > 0, u = (u1, u2) ∈ P (Z2), then C contains a rational component through (1, 1) parametrized as (su1, su2). 2 does not hold except for the case k = m/2. Note that the tropicalization of such a component is the straight line though (0, 0) in the direction (u1, u2). We will use this lemma in Corollary 4.19. Lemma 3.9 ([15], Lemma 2.8, Lemma 5.20). Let µp(C) ≥ m, p = (x1, x2) ∈ (K∗)2, denote P = (val(x1), val(x2)). Suppose that P is a vertex of Trop(C) and ωu(d(P )) = a ≤ m for some direction u. Then both sides of d(P ), perpendicular to u, have length at least m − a, and (4) area(d(V )) ≥ 1 2 (m − a)2. XV ∈IP (u),V 6=P Example 3.10. Suppose that u = (1, 0). Note that in this case a is the length of the projection of d(P ) onto the x-axis. Recall that IP ((1, 0)) is the set of vertices of Trop(C), lying in the connected component of P in the intersection of Trop(C) with the straight horizontal line through P . Figure 2 illustrates the set of dual faces d(V ) to the vertices V in IP ((1, 0)). Since the long edge through P is horizontal, all the edges separating faces in Infl(P ) in Figure 2 are vertical. Remark 3.11. Lemma 3.9 holds in the degenerate case (a = 0), too: if P belongs to an edge E of Trop(C), then the dual edge d(E) of the subdivision of the Newton polygon of C has lattice width at least m (Theorem 1 in [15]). 8 N. KALININ M ≥ m − a N a L d(P ) ≥ m − a K Figure 2. Dual picture to a singular point P on an edge. Since ω(1,0)(d(P )) = a, the lengths of LM and N K are at least m − a (Lemma 3.9). The set S d(V ) for V ∈ IP ((1, 0)), V 6= P is colored. The sum of the areas of the colored faces is at least 2 (m − a)2. 1 The same reasoning works in any dimension. Remark 3.12. Consider a planar tropical curve H. Suppose that this curve is not a usual line. Then the Newton polygon ∆ of H is two-dimensional. The intersection of H with a usual line in direction u ∈ P (Z2) is equal to ωu(∆). Hence, if the intersection of H with each usual line of rational direction is at least m > 0, then its self-intersection H · H is at least 3 4 m2 which is the minimal double area of a polytope ∆ with ω(∆) ≥ m, for details and references see [15]. If the intersection of a hypersurface H ⊂ TP k with any usual line of rational direction is at least m, then the minimal lattice width ω(∆) of the Newton polytope ∆ of H is at least m and therefore the self-intersection H k is at least cmk where c is a constant which depends only on k. However, the best value of c in the inequality Volume(∆) ≥ cω(∆)k is not known. The following question in codimension two, therefore, is the simplest one possessing no estimate at all (because there is no notion like Newton polytope which keeps track of degree and self-intersection at the same time). Conjecture 2. Consider a tropical (i.e. balanced along dimension one faces) two-dimensional fan L in R4. Suppose that L is not an affine Euclidean plane. Suppose that the stable tropical intersection of L with each plane of rational slope is at least m. Then there exists a constant c such that L · L ≥ cm2 in this case, and c does not depend on m. 4. Influenced subsets in the Newton polytope In this section we generalize the definitions of influenced sets in the Newton polygon, given in the article [15] (see Section 3 for recap). Also we discuss here the notion of a set of points in Zk in tropical general position with respect to a polytope ∆. Let Y be a hypersurface in Kk with Newton polytope ∆. In this subsection, for a given subvariety X ⊂ Y , we define the set I∆(Trop(X)) of vertices of Trop(Y ) and the subset Infl(Trop(X)) ⊂ ∆. Definition 4.1. We denote by P (Zk) the set of all primitive non-zero vectors in Zk. An affine hyperplane with a normal direction u ∈ P (Zk) is a set {x ∈ Rku · x = c} with some c ∈ R. Let Q be a subset of Trop(Y ). Definition 4.2. Let lQ(u) be the affine hyperplane in Rk with normal direction u, containing the set Q, if such a hyperplane exists, and lQ(u) = ∅, otherwise. Let P (∆) ⊂ P (Zk) be the set of the primitive vectors {IJI, J ∈ ∆} between the lattice points in ∆. Define the star of Q (with respect to ∆) as Star∆(Q) = [u∈P (∆) lQ(u). The connected component of Q in the intersection Trop(Y ) ∩ Star∆(Q) is called the star Star∆ of Q in Trop(Y ). Y (Q) TROPICAL APPROACH TO NAGATA'S CONJECTURE IN POSITIVE CHARACTERISTIC 9 Example 4.3. Let Y ⊂ (K∗)2 be an algebraic curve whose Newton polygon is ∆. • If Q is a vertex of Trop(Y ), then Star∆ Y (Q) is the connected component of Q in the inter- section of Trop(Y ) with the union of the lines spanned by the edges of Trop(Y ) through Q, see example in Figure 1, (B). • If Q ∈ Trop(Y ) is not a vertex of Trop(Y ), then Star∆ Y (Q) is the connected component of Q in the intersection of Trop(Y ) with the line spanned by the unique edge of Trop(Y ) through Q, Figure 1, (A). Definition 4.4. Let I∆(Q) be the set of the vertices of Trop(Y ) which belong to Star∆ Y (Q). We provide each point in Star∆(Q) with a multiplicity corresponding to the codimension of its stratum. Definition 4.5. For a point V ∈ Star∆(Q) the natural number multQ(V ) is the dimension of the linear span of the directions u ∈ P (∆) such that the affine hyperplane through V with the normal direction u contains Q. Example 4.6. If ∆ ⊂ Z2 and Q is a point, then Star∆(Q) is a union of intervals emanating from Q. In this case multQ(Q) = 2 and multQ(V ) = 1 for V ∈ Star∆(Q), V 6= Q. Trop(X) is a subcomplex of the Grobner complex (see Section 2). So, we present Trop(X) =S X i as a union of cells which we denote by X i. Recall that if X is a hypersurface, then each cell of Trop(X) is an equivalence class of some ω ∈ Trop(X), with the equivalence relation ω ∼ ω′ iff ∆ω = ∆ω′, see Remark 2.1. Let X ⊂ Y , then Trop(X) ⊂ Trop(Y ). Definition 4.7. Define I∆(Trop(X)) =S I∆(X i). We define the star of the variety Trop(X) as Each tropical variety Trop(X) is naturally decomposed into vertices, edges, faces, etc, because Y (X i). Star∆(Trop(X)) =[ Star∆(X i), Star∆ Y (Trop(X)) =[ Star∆ So, we take all the cells of X, draw the star for each of them, and take the union of these stars. Definition 4.8. For a vertex V ∈ I∆(Trop(X)) we define its multiplicity multTrop(X)(V ) as multTrop(X)(V ) = max X i multX i(V ), i.e. we take the maximum of the multiplicities of V with respect to the cells in the natural cell decomposition of Trop(X). Definition 4.9. Let X ⊂ Y ⊂ Kk be algebraic varieties, Y be a hypersurface with the Newton polytope ∆. The distinguished domain Infl(X) in ∆, corresponding to X, is Infl(X) = d(V ), [V ∈I∆(Trop(X)) where d(V ) is the cell (of the maximal dimension) of ∆, dual to the vertex V of Trop(Y ). For I∆(Trop(X)) see Definitions 4.4, 4.7. Note that Infl(X) depends only on Trop(X), so we will write Infl(Trop(X)). Definition 4.10. By Volume(Infl(Trop(X))) we denote the sum of volumes (with multiplicities, see Definition 4.8) of the cells in the subdivision of ∆, dual to the vertices in I∆(Trop(X)), i.e. Volume(Infl(Trop(X))) = XV ∈I∆(Trop(X)) multTrop(X)(V ) · Volume(d(V )). Example 4.11. Refer to Example 4.6. Consider the two-dimensional case, X = (x1, x2) ∈ (K∗)2 is a point such that Trop(X) = P = (val(x1), val(x2)) ∈ T2. If P is a vertex of Trop(Y ), then area(Infl(P )) = 2 · area(d(P )) + XV ∈I∆(P ), V 6=P 1 · area(d(V )), 10 N. KALININ which coincides with the definition of area(Infl(P )) in Definition 3.5. In Figure 1 (B) area(Infl(P )) is the area of the depicted part of the Newton polygon, with the area of the central face counted twice. Remark 4.12. The dual object for a hypersurface is its Newton polytope. The dual objects for the varieties of higher codimension are so-called generalized Newton polytopes or valuations in the McMullen polytope algebra [4, 24]. Even though for X ⊂ Y ⊂ Kk with codim(Y ) > 1 we can define I(Trop(X)) in a similar fashion (by intersecting the stars of cells of Trop(X) with Y ), it is not clear what is the right substitute for Volume(Infl(Trop(X))) in this case. 4.1. General position of points with respect to the Newton polytope. Definition 4.13. A collection of tropical subvarieties Z1, Z2, . . . , Zn ⊂ Tk is in general position with respect to a polytope ∆ ⊂ Zk if for each m = 1, 2, . . . k + 1 for each collection of indices i1 < i2 < ··· < im the intersection Star∆(Zi1) ∩ Star∆(Zi2) ∩ ··· ∩ Star∆(Zim) (Definition 4.7) has codimension at least m in Tk. In particular, the intersection of any k + 1 stars is empty. Proposition 4.14. Let v ∈ P (Z2) be a primitive vector such that v /∈ P (∆) (Definition 4.2). Let l be the line {t·vt ∈ R}. Then for any n ∈ Z>0 there exists a collection of points P1, P2, . . . , Pn ∈ l∩Z2 in general position with respect to ∆. Proof. Let P1 = 0. Take any point P2 ∈ l ∩ Z2. Note that the vector P1P2 /∈ P (∆), therefore P1, P2 are in general position. Draw Star∆(P1), Star∆(P2) and all stars Star∆ for the points of intersection of lines in Star∆(P1), Star∆(P2). It is a finite collection of lines, none of them is l. Therefore we can find a point P3 ∈ l ∩ Z2 which does not belong to these lines. Then we draw Star∆(P3) and starts Star∆ of all points of intersection between lines in Star∆(P3) and lines in Star∆(P1), Star∆(P2). Again, we see a finite collection of lines and we can choose P4 ∈ l ∩ Z2 not on these lines. We can 2 (cid:1)P (∆)2 points of continue to choose points P5, . . . , Pn as above. When we choose Pn we have(cid:0)n−1 intersections of lines in Star∆(P1), Star∆(P2), . . . , Star∆(Pn−1), when we draw stars Star∆ for them, 2 (cid:1)P (∆)3. (cid:3) we have(cid:0)n−1 Let Tv be the translation Tk → Tk by the vector v ∈ Rk, i.e. x → x + v. Proposition 4.15. For a polytope ∆ ⊂ Zk and a given set Z1, Z2, . . . , Zn ∈ Tk of tropical varieties there exists a set of vectors v1, v2, . . . , vn ∈ Zk such that the tropical varieties Tvi(Zi) are in general position with ∆. 2 (cid:1)P (∆)3 lines, therefore we can choose Pn ∈ l∩[0, R]2 where R = v·2(cid:0)n−1 Proof. Indeed, each star Star∆(Zi) is a finite union of hyperplanes (Definition 4.2). We argue as in Proposition 4.14. We can choose v1 = 0 and v2 ∈ Zk such that the intersection of each two hyperplanes L1, L2 from the collections Star∆(Z1) and Star∆(Tv2(Z2)) respectively is a linear subspace of dimension at most k − 2. Then we choose a vector v3 ∈ Zk such that the intersection of each pair of hyperplanes from different collections Star∆(Tvi(Zi)), i = 1, 2, 3 is of dimension at most k− 2 and the intersection of a triple of hyperplanes from different collections is of dimension at most k − 3, etc. Each time we see that the set of vi such that Star∆(Tvi Zi) violates some transversality condition is a finite union of hyperplanes, and we need to take a vi ∈ Zk outside of it. Corollary 4.16. For each n, k ∈ N, ∆ there exists a set of points (taken as tropical varieties of dimension zero) P1, P2 . . . , Pn ∈ Zk ⊂ Tk in general position with respect to ∆. Corollary 4.17. For a collection of tropical varieties Z1, Z2, . . . , Zn ⊂ Y ⊂ Tk in general position with respect to the Newton polytope ∆ of a tropical hypersurface Y , the sumPn i=1 Volume(Infl(Zi)) is at most k · Volume(∆). Proof. This follows from the definitions of a general position (Definition 4.13) and multiplicities in the volume of Infl (Definition 4.10). (cid:3) (cid:3) TROPICAL APPROACH TO NAGATA'S CONJECTURE IN POSITIVE CHARACTERISTIC 11 4.2. First applications. Consider an algebraci curve C ⊂ (K∗)2 passing through p1, p2, . . . , pn ∈ (K∗)2 with multiplicities m1, m2, . . . , mn respectively. Suppose that n ≥ 2 and the minimal lattice width ω(∆) of the Newton polygon ∆ of C satisfies ω(D) ≥ max(mi). Lemma 4.18. If the points Val(pi) ∈ Z2, i = 1, . . . , n are in general position with respect to ∆ (see Proposition 4.15 and its corollaries), then the area of ∆ satisfies the inequality (5) area(∆) ≥ Proof. Theorem 3.6 and Corollary 4.17 imply that 1 4 nXi=1 m2 i . m2 i 2 ≤ nXi=1 nXi=1 area(Infl(Pi)) ≤ 2 · area(∆). (cid:3) Corollary 4.19. Suppose that a curve of degree d passes through the points p1, p2, . . . , pn ∈ (K∗)2 with multiplicities m1, m2, . . . , mn respectively, d ≥ max(mi), and n ≥ 2. Suppose also that the 2Pn points Val(pi) are in general position with respect to ∆. Then, we have d2 ≥ 1 Proof. The equation of a curve of degree d may contain some monomials with zero coefficients. So, if the minimal lattice width of the actual Newton polygon of C is at least max(mi), then we conclude by Lemma 4.18. If it is not the case, we apply Lemma 3.8. i=1 m2 i . If C has a rational component passing through a point p = (x1, y1) parametrized by s as (x1sk, y1sl), then C is reducible, and we can perturb this component, because it does not pass through other pi by genericity (recall that the tropicalization of this component is a straight line in the direction (k, l) ∈ P (∆)). After that this component is no longer of the type (x1sk, y1sl), and this perturbation does not change the degree of the curve. After repeating this cycle of arguments necessary number of times we can apply Lemma 4.18. (cid:3) Remark 4.20. Consider a hypersurface H in (K∗)3 passing through generic points of multiplicity two. The classification of possible combinatorial neighborhoods of a two-fold point P in a tropical surface in T3 ([18]) allows us to produce an estimates for the volume and the shape of Infl(P ). We can prove that Volume(Infl(P )) ≥ 2 if P = (val(p1), val(p2)) for a point (p1, p2) of multiplicity two in H. With a few more work (one should check possible intersections Infl(P ) ∩ Infl(Q) for different points P, Q in general position) the author obtained a proof of an estimate n ≤ d3 3 for the degree d of a surface with n two-fold points, as we did in Lemma 4.18. However, the theorem of Alexander and Hirschowitz provides a better estimate n ≤ (d+1)(d+2)(d+3) . Nevertheless, we expect that this research paradigm can be carried out for points of multiplicity m on hypersurfaces in (K∗)3, with a conjectural estimate Volume(Infl(P )) ≥ cm3 for some constant c. Remark 4.21. We expect that for a line L of multiplicity m inside a surface Y of degree d in KP 3 the estimate Volume(Infl(Trop(L))) ≥ cm2d holds with some constant c. This will give an estimate for the degree of a surface with multiple two-fold points and m-fold lines. The idea is how it could work is as follows. Consider Trop(L) ⊂ Trop(Y ). Intersect it with a plane {Z = const} ⊂ T3. In the intersection we will see the same picture as for planar curves with an m-fold point. Varying the constant and the normal vector of such a plane we can make a conjecture as follows. 24 Conjecture 3. Under the above hypothesis, Infl(Trop(L)) is a a connected subset of the Newton polytope ∆ of Y , which intersects the faces of ∆, perpendicular to the directions of the rays of Trop(L), and whose sections by planes with a primitive normal vector in Z3 have area at least 3 8 m2. Hence, Volume(Infl(Trop(L))) ≥ 3 4.3. Detropicalization Lemma. An algebraic statement over an algebraically closed field some- times implies the same statement over all fields of the same characteristic. Tropical geometry may help in such a situation, see [27]. Another application of tropical geometry in number theory is [16]. This section describes a particular application of this principle to our estimate. 8 m2d. 12 N. KALININ Recall that our field K is the field of fractions f (t) g(t) where f, g ∈ F[t]. Note that we can substitute t = a if g(a) 6= 0. Let us recall how to tropicalize the problem of curves' counting. We would like to count plane complex algebraic curves of given genus and degree, these curves are required to pass through a number of generic points q1, q2, . . . , ql ∈ CP 2 (l is chosen in such a way that the number of curves is expected to be finite). Since the points are generic, we can force them to go to infinity with some asymptotics, say qi = (txi, tyi), (xi, yi) ∈ Z2. Then we consider the limits of these curves Ct under the function logt(z) : C2 → R2. This is more or less the same as if we consider a curve C over K passing through (txi, tyi) ∈ (K∗)2 and then take its tropicalization Trop(C). Hence we started from C, lifted to K, and finally descended to T. Detropicalization is the opposite process: we prove something in T, then lift the construction to K, and return to F using such a substitution for an appropriate a. We establish the following lemma. Lemma 4.22. Let m1, m2, . . . , mn be non-negative integers. Let ∆ be a lattice polygon such that area(∆) < m2 i 4 . nXi=1 Then, if a set of points P1, . . . , Pn ∈ Z2 ⊂ T2 is in general position with respect to ∆ (Definition 4.13), then for each valuation field K and points p1, p2, . . . , pn ∈ (K∗)2 such that Val(pi) = Pi there exists no curve C over K with the Newton polygon ∆, with µpi(C) ≥ mi, i = 1, . . . , n. Proof. Suppose that such a curve C exists. Then, consider Trop(C). We know that in this case for i = 1, . . . , n and, therefore,Pn we arrive at a contradiction. area(cid:0)Infl(Pi)(cid:1) ≥ nPi=1 i=1 area(Infl(Pi)) ≥ m2 i 2 m2 i 4 ≥ 2 · area(∆). So, using Corollary 4.17 (cid:3) Lemma 4.23 (Detropicalization lemma). Let K be the field of fractions of F[t]. Suppose that there exists no curve C over K with the Newton polygon ∆ such that µ(t−xi ,t−yi )(C) ≥ mi, for given different points (xi, yi) ∈ Z2, i = 1, . . . , n and given numbers mi ∈ Z>0, i = 1, . . . , n. Then, there exists a constant N depending on m1, m2, . . . , mn, ∆, max xi, max yi with the following property. If F ≥ N , then there exists a ∈ F such that there is no curve C over F with the Newton polygon ∆, satisfying µ(a−xi ,a−yi )(C) ≥ mi for each i = 1, . . . , n. Proof. Suppose the contrary. Take any b ∈ F. All the constraints imposed by the fact µp(C) ≥ m are linear equations in the coefficients of the equation of C. Therefore the only reason why there is no solution for this system over K and there exists a solution over F is that some minor of the matrix of the equations turns out to be 0 after substitution t = b. Thus, let us compute all the considered minors before, they reveal to be polynomials in t with degrees depending on our data. Therefore, b is a root of this fixed polynomial of a certain bounded degree. Obviously, if F is big enough, then there exists a which is not a root of this polynomial. Therefore, this a satisfies the statement of lemma. (cid:3) Remark 4.24. In a similar way we can "detropicalize" in other situations, if the conditions imposed on C reveal to be algebraic conditions on the coefficients of the equation of C. 5. Degeneration of tropical points to a line In this section, using tropical floor diagrams (see [5, 7]), we construct a special collection of points in T2, which are in general position with respect to the Newton polygon ∆; this construction gives another estimate for area(∆). TROPICAL APPROACH TO NAGATA'S CONJECTURE IN POSITIVE CHARACTERISTIC 13 Consider a tropical curve H given by Trop(F )(X, Y ) = max(i,j)(Aij + iX + jY ) where (i, j) runs over lattice points in a fixed Newton polygon ∆. After a toric change of coordinates we may assume that the minimal lattice width ω(∆) of ∆ is attained in the horizontal direction. Let ∆ be contained in the strip {(x, y)0 ≤ y ≤ N}. Let us choose points P1, P2, . . . , Pn on the line l = {(X, Y )Y = 1 N +1 is less than any possible slope of a non-horizontal edge of a tropical curve with the given Newton polygon ∆. N +1 X} which is almost horizontal, namely, its slope 1 Proposition 5.1. Suppose that each of the points P1, P2, . . . , Pn is not a vertex of H, and each Pi belongs to a horizontal edge Ei of H. In this case, for each 1 ≤ i < j ≤ n we have Infl(Pi)∩Infl(Pj) = ∅. Proof. Indeed, in this case the vertices in I(Pi) belong to the horizontal line through Pi for each i = 1, 2, . . . , n, and all Pi have different y-coordinates. (cid:3) Corollary 5.2. In the above case,Pn i=1 area(Infl(Pi)) ≤ area(∆). In general, a correction term (Eq. 2) appears by the following reasons. The line l is subdivided by intersections with H, each connected component of l \ (l ∩ H) corresponds to a monomial in Trop(F ), i.e. to a lattice point in ∆. Moving along l from left to right and marking corresponding lattice points in ∆ we obtain a lattice path in ∆, which possesses the following property: each edge in this path is either vertical (and has positive projection on the vertical line), or has positive projection on the horizontal line. Indeed, if Aij + iX + jY > Ai′j′ + i′X + j′Y , but for small ε > 0 Aij + i(X + ε) + j(Y + 1 N + 1 ε) < Ai′j′ + i′(X + ε) + j′(Y + 1 N + 1 ε), then the vector (i′ − i, j′ − j) has the described above properties. If Pi is not a vertex of H, and Pi belongs to an edge Ei of H, then denote by si the length of the horizontal projection of d(Ei). If Pi is a vertex of H, then denote by si the length of the horizontal projection of d(Pi). Above considerations show that nPi=1 si ≤ ω(∆), see Figure 3 for illustration. ω(∆) 4 • 3 • ∈ Infl(P3) ∈ Infl(P2) • 2 Infl(P1) ∋ • 1 s1 = 0 s2 s3 = 0 P1 • 2 1 P2 • 3 P3 • 4 Figure 3. The left picture represents a part of a tropical curve through points P1, P2, P3 on an almost horizontal line. The second picture is dual to the first picture, we see parts of the regions of influence of the points P1, P2, P3. The marked points 1, 2, 3, 4 represent the monomials which are maximal on the parts of the dotted line on the left picture. The lattice path 1, 2, 3, 4 is non-decreasing by the x-coordinate, i=1 si ≤ ω(1,0)(∆). The vertical projections of the intervals 12, 23, 34 in the right picture are s1 = 0, s2, s3 = 0 respectively. therefore Pn 14 N. KALININ Proposition 5.3. In the above notation, and using Definition 3.2 for IPi((1, 0)) we have 1 2 nXi=1 (m2 i − s2 i ) ≤ nXi=1  XV ∈IPi ((1,0)) area(d(V )) ≤ area(∆). Proof. The right inequality is trivial, because the sets IPi((1, 0)) do not intersect each other. The left inequality follows from the estimate XV ∈IPi ((1,0)) area(d(V )) ≥ 1 2 (m2 i − s2 i ) for each i = 1, . . . , n. Indeed, if Pi is not a vertex of H, then Pi belongs to an edge Ei of H. If Ei is horizontal (the case for i = 1, 3 in Figure 3), then d(Ei) is vertical, and si = 0. By Lemma 3.9 (or Theorem 3.6 for the case when Pi belongs to an edge) XV ∈IPi ((1,0)) area(d(V )) ≥ 1 2 m2 i . If Ei is not horizontal, then si ≥ mi (because si is at least the lattice length of d(Ei) and we use Remark 3.11 in this case) and our inequality becomes trivial. If Pi is a vertex of H, then the inequality follows from Lemma 3.9, because in this case 1 2 Indeed, the term 1 (mi − si)2 = area(d(V )) ≥ (mi − si) · si + XV ∈IPi ((1,0)) 2 (mi − si)2 comes from Lemma 3.9, and (mi − si)· si estimates the area of d(Pi) from below, because ω(1,0)(d(Pi)) = si by our assumption and d(Pi) has two vertical sides of length at least mi − si by Lemma 3.9. (cid:3) Proof of Theorem 1.2. By Proposition 4.14 there exists N such that there exists a generic with ω(∆)+1 x with xi,yi < N . Then, Propo- respect to ∆ collection of lattice points on the line y = sition 5.3 implies the statement over K (as in Lemma 4.22), and Lemma 4.23, "detropicalizing", concludes the proof. (cid:3) m2 i − s2 2 i 1 . 6. Speculations destined to coding theory In informatics, (error-correcting) coding-theory deals with subsets C ⊂ An (A is a finite set) which are as big as possible, and the Hamming distance d between the elements of C is also as big as possible, i.e. we maximize δ = mina,b∈C,a6=b d(a, b). Such a subset C is called a code and If, it is suitable for the following problem. We transmit a message which is an element of C. during the transmission procedure, the message does change in at most δ−1 2 positions, then we can uniquely repair it back, that is why this is called an error-correcting code. As an introductory book, which relates this subject to algebraic geometry, see [26]. Studying of singular varieties is related with code-theory ([28]), for the relation of this topic with Seshadri constants (which is a relative of Nagata's conjecture), see [13]. Finding such subsets C is a hard combinatorial problem. A particular source for codes is the set of linear subspaces of Fn q (linear codes), mostly because they have a comparatively simple description. A common construction is the following. We choose a set of points p = {p1, p2, . . . , pn} ⊂ Fm q and consider a set Vd ⊂ Fq[x1, x2, . . . , xm] of polynomials of degree at most d (one can consider any linear system on a toric variety and points on it as well). Then we take the evaluation map: evp : Vd → Fn q , evp(f ) = (f (p1), f (p2), . . . , f (pn)). The image of evp is a linear code which is quite simple to calculate, but the problem is how to choose the points pi such that there is no polynomial which vanishes at the chosen points (otherwise we have to deal with the kernel of evp) and how to estimate the minimal distance δ. For example, one TROPICAL APPROACH TO NAGATA'S CONJECTURE IN POSITIVE CHARACTERISTIC 15 may take all the points with all non-zero coordinates as p, p = Fm minimize the codimension of the code in Fn q this choice is far from optimal. q . However, since we are trying to Following a suggestion of Joaquim Ro´e, we mention here a way we can exploit the main ideas of this article to construct a linear code, which uses not too many points and provides a map, similar to evp, without a kernel. In the previous sections, for a given polygon ∆ and numbers m1, m2, . . . , mn we constructed the set of points p = {p1, p2, . . . , pn} ⊂ (F∗q)2 such that there is no curve C with Newton polygon ∆, possessing the property µpi(C) ≥ mi for each i. Recall that for this construction we should carefully choose points (xi, yi) ∈ Z2, i = 1, . . . , n, then, for q big enough there exists an a ∈ Fq, such that the points pi = (a−xi, a−yi) possess the required properties. Example 6.1. Consider ∆ = [0, 1, . . . , d] × [0, 1 . . . , N ] ⊂ Z2. It follows from the proof of Theo- rem 1.2 that if we put n points p1, p2, . . . , pn of the same multiplicity m ≤ min(N, d) along the curve yω(∆)+1 = x, then there is no algebraic curve C with the Newton polygon ∆ and µpi(C) ≥ m if the inequality dN < 1 2 (n − d/m)m2 holds. Therefore, taking N < (n−d/m)m2 2d we construct the evaluation map ev : FdN with a , which we treat as a polynomial F with the q → F q trivial kernel. For this map, we take a point f ∈ FdN Newton polygon ∆, then take the coefficients of(cid:0)F mod I m numbers gives the image ev(f ). In this construction we immediately see that the minimal non-zero Hamming distance δ is at most n, because the image of f ≡ 1 under the map ev contains exactly n non-zero elements. pi(cid:1) for each i = 1, . . . , n. This bunch of q nm(m+1) 2 Conjecture 4. This estimate is sharp, i.e. δ = n for this code. References [1] J. Alexander and A. Hirschowitz. Polynomial interpolation in several variables. J. Algebraic Geom., 4(2):201 -- 222, 1995. (page 4). [2] J. Alexander and A. Hirschowitz. An asymptotic vanishing theorem for generic unions of multiple points. Inven- tiones mathematicae, 140(2):303 -- 325, 2000. (page 4). [3] C. Bocci. Special effect varieties in higher dimension. Collect. Math., 56(3):299 -- 326, 2005. (page 4). [4] M. Brion. Piecewise polynomial functions, convex polytopes and enumerative geometry. In Parameter spaces (Warsaw, 1994), volume 36 of Banach Center Publ., pages 25 -- 44. Polish Acad. Sci., Warsaw, 1996. (page 10). [5] E. Brugall´e, I. Itenberg, G. Mikhalkin, and K. Shaw. Brief introduction to tropical geometry. Proceedings of 21st Gokova Geometry-Topology Conference, arXiv:1502.05950, 2015. (pages 4, 12). [6] E. Brugall´e and G. Mikhalkin. Enumeration of curves via floor diagrams. C. R. Math. Acad. Sci. Paris, 345(6):329 -- 334, 2007. (page 3). [7] E. Brugall´e and G. Mikhalkin. Floor decompositions of tropical curves: the planar case. In Proceedings of Gokova Geometry-Topology Conference 2008, pages 64 -- 90. Gokova Geometry/Topology Conference (GGT), Gokova, 2009. (pages 3, 12). [8] C. Ciliberto. Geometric aspects of polynomial interpolation in more variables and of Waring's problem. In Euro- pean Congress of Mathematics, Vol. I (Barcelona, 2000), volume 201 of Progr. Math., pages 289 -- 316. Birkhauser, Basel, 2001. (page 4). [9] C. Ciliberto, B. Harbourne, R. Miranda, and J. Ro´e. Variations of Nagata's conjecture. In A celebration of algebraic geometry, volume 18 of Clay Math. Proc., pages 185 -- 203. Amer. Math. Soc., Providence, RI, 2013. (page 4). [10] C. Ciliberto and R. Miranda. Homogeneous interpolation on ten points. J. Algebraic Geom., 20(4):685 -- 726, 2011. (page 4). [11] M. Dumnicki, B. Harbourne, T. Szemberg, and H. Tutaj-Gasi´nska. Linear subspaces, symbolic powers and Nagata type conjectures. Adv. Math., 252:471 -- 491, 2014. (page 4). [12] L. Evain. Computing limit linear series with infinitesimal methods. Ann. Inst. Fourier (Grenoble), 57(6):1947 -- 1974, 2007. (page 3). [13] S. H. Hansen. Error-correcting codes from higher-dimensional varieties. Finite fields and their applications, 7(4):530 -- 552, 2001. (page 14). [14] B. Harbourne. Problems and progress: a survey on fat points in P2. In Zero-dimensional schemes and applications (Naples, 2000), volume 123 of Queen's Papers in Pure and Appl. Math., pages 85 -- 132. Queen's Univ., Kingston, ON, 2002. (page 4). [15] N. Kalinin. The Newton polygon of a planar singular curve and its subdivision. Journal of Combinatorial Theory, Series A, 137:226 -- 256, 2016. (pages 2, 6, 7, 8). 16 N. KALININ [16] E. Katz, J. Rabinoff, and D. Zureick-Brown. Uniform bounds for the number of rational points on curves of small mordell -- weil rank. arXiv:1504.00694, 2015. (page 11). [17] D. Maclagan and B. Sturmfels. Introduction to tropical geometry, volume 161 of Graduate Studies in Mathematics. American Mathematical Society, Providence, RI, 2015. (pages 4, 6). [18] H. Markwig, T. Markwig, and E. Shustin. Tropical surface singularities. Discrete Comput. Geom., 48(4):879 -- 914, 2012. (pages 4, 11). [19] T. Markwig. A field of generalised Puiseux series for tropical geometry. Rend. Semin. Mat. Univ. Politec. Torino, 68(1):79 -- 92, 2010. (page 3). [20] R. Miranda. Linear systems of plane curves. Notices AMS, 46(2):192 -- 202, 1999. (page 4). [21] M. Nagata. On the 14-th problem of Hilbert. Amer. J. Math., 81:766 -- 772, 1959. (page 4). [22] S. Paul. New methods for determining speciality of linear systems based at fat points in Pn. J. Pure Appl. Algebra, 217(5):927 -- 945, 2013. (page 4). [23] J. M. Ruiz. The basic theory of power series. Advanced Lectures in Mathematics. Friedr. Vieweg & Sohn, Braun- schweig, 1993. (page 3). [24] R. Steffens and T. Theobald. Combinatorics and genus of tropical intersections and Ehrhart theory. SIAM J. Discrete Math., 24(1):17 -- 32, 2010. (page 10). [25] B. Strycharz-Szemberg and T. Szemberg. Remarks on the Nagata conjecture. Serdica Math. J., 30(2-3):405 -- 430, 2004. (page 4). [26] M. Tsfasman, S. Vladut¸, and D. Nogin. Algebraic geometric codes: basic notions, volume 139 of Mathematical Surveys and Monographs. American Mathematical Society, Providence, RI, 2007. (page 14). [27] I. Tyomkin. On Zariski's theorem in positive characteristic. J. Eur. Math. Soc. (JEMS), 15(5):1783 -- 1803, 2013. (page 11). [28] J. Wahl. Nodes on sextic hypersurfaces in P3. J. Differential Geom., 48(3):439 -- 444, 1998. (page 14). [29] G. Xu. Curves in P2 and symplectic packings. Math. Ann., 299(4):609 -- 613, 1994. (page 4). Departamento de Matem´aticas, CINVESTAV Apartado Postal: 14-740, C.P. 07000,Ciudad de M´exico, Mexico E-mail address: nikaanspb{dog}gmail.com 

1802.02735

1

1802

A new presentation of the plane Cremona group

We give a presentation of the plane Cremona group over an algebraically closed field with respect to the generators given by the Theorem of Noether and Castelnuovo. This presentation is particularly simple and can be used for explicit calculations.

math.AG

math

A NEW PRESENTATION OF THE PLANE CREMONA GROUP CHRISTIAN URECH AND SUSANNA ZIMMERMANN Abstract. We give a presentation of the plane Cremona group over an algebraically closed field with respect to the generators given by the Theorem of Noether and Castelnuovo. This presentation is particularly simple and can be used for explicit calculations. 1. Introduction 1.1. Main result. Let k be an algebraically closed field. The plane Cremona group Bir(P2) is the group of birational transformations of the projective plane P2 = P2 k. It was intensely studied by classical algebraic geometers and has attracted again considerable attention in the last decades. If we fix homogeneous coordinates [x : y : z] of P2, every element f ∈ Bir(P2) is given by f : [x : y : z] (cid:55)(cid:57)(cid:57)(cid:75) [f0(x, y, z) : f1(x, y, z) : f2(x, y, z)], where the fi are homogeneous polynomials of the same degree, the degree of f. We will write f = [f0(x, y, z) : f1(x, y, z) : f2(x, y, z)]. One of the main classical results is the Theorem of Noether and Castelnuovo ([C1901]), which states that Bir(P2) is generated by the linear group Aut(P2) = PGL3(k) and the standard quadratic involution σ := [yz : xz : xy]. Throughout the last century, various presentations of the plane Cremona group have been given (see Section 1.2). Each one of these displays a particularly interesting and beautiful aspect of the Cremona group. The aim of this article is to give a presentation of the plane Cremona group that uses the generators given by the Theorem of Noether and Castelnuovo. Another advantage of our presentation is that it is particularly simple and in many cases easy to use for specific calculations. We will illustrate this in Section 1.3 with an example that is due to Gizatullin. Denote by D2 ⊂ PGL3(k) the subgroup of diagonal automorphisms and by S3 ⊂ PGL3(k) the symmetric group of order 6 acting on P2 by coordinate permutations. Main Theorem. Let k be an algebraically closed field. The Cremona group Bir(P2) is isomorphic to Bir(P2) (cid:39)(cid:10)σ, PGL3(k) (1) − (5)(cid:11) 3 = id for all g1, g2, g3 ∈ PGL3(k) such that g1g2 = g3. (1) g1g2g−1 (2) σ2 = id, (3) στ (τ σ)−1 = id for all τ ∈ S3, (4) σdσd = id for all diagonal automorphisms d ∈ D2, (5) (σh)3 = id, where h = [z − x : z − y : z]. Observe that the relations (2) to (4) occur in the group Aut(k∗ × k∗), which is given by the group of monomial transformations GL2(Z) (cid:110) D2. Relation (5) is a relation from the group Aut(P1 × P1)0 (cid:39) PGL2(k) × PGL2(k) which we consider as a subgroup of Bir(P2) by conjugation with the birational equivalence
: P1 × P1 (cid:57)(cid:57)(cid:75) P2, given by ([u0 : u1], [v0 : v1]) (cid:55)(cid:57)(cid:57)(cid:75) [u1v0 : u0v1 : u1v1]. Date: February 9, 2018. 2010 Mathematics Subject Classification. 14E07, 20F05. During this work, both authors were partially supported by the Swiss National Science Foundation. 1 With respect to affine coordinates (x, y), the de Jonquières can be described by: J = (x, y) (cid:55)(cid:57)(cid:57)(cid:75) ∈ PGL2(k), , α(x)y + β(x) γ(x)y + δ(x) (cid:18)α(x) β(x) (cid:19) γ(x) δ(x) (cid:19) (cid:19)(cid:12)(cid:12)(cid:12)(cid:12)(cid:18)a b c d (cid:27) ∈ PGL2(k(x)) (cid:26) (cid:18) ax + b cx + d (cid:39) PGL2(k) (cid:110) PGL2(k(x)) 2 CHRISTIAN URECH AND SUSANNA ZIMMERMANN The idea of the proof of the Main Theorem is the same as in [I1984, B2012, Z2016]. We study linear systems of compositions of birational transformations and use the presentation of the Cremona group given by Blanc in [B2012]. 1.2. Previous presentations. The aim of this section is to give an overview over previous presen- tations of the plane Cremona group. The first presentation was given by Gizatullin in 1982 ([G1982], see also [G1990]): Theorem 1.1 ([G1982, Theorem 10.7]). The Cremona group Bir(P2) is generated by the set Q of all quadratic transformations and the relations in Bir(P2) are consequences of relations of the form q1q2q3 = id, where q1, q2, q3 are quadratic transformations, i.e. we have the presentation: Bir(P2) = (cid:104)Q q1q2q3 = id for all q1, q2, q3 ∈ Q such that q1q2q3 = id in Bir(P2)(cid:105). Definition 1.2. We denote by J ⊂ Bir(P2) the group of transformations preserving the pencil of lines through [1 : 0 : 0] and call it the de Jonquières group. By the Theorem of Noether-Castelnuovo, Bir(P2) is generated by J and PGL3(k), since σ is an element of J . The following theorem shows that, in a certain way, most of the relations in Bir(P2) are consequences of relations within these groups: Theorem 1.3 ([B2012]). The Cremona group Bir(P2) over an algebraically closed field is the amalga- mated product of PGL3(k) and J along their intersection, divided by the relation where σ is the standard involution and τ ∈ PGL3(k) the transposition [y : x : z]. στ = τ σ, Note that the Cremona group does not have the structure of an amalgamated product [C2013]. Theorem 1.3 was proceeded by the following statement shown by Iskovskikh in [I1984]: Theorem 1.4 ([I1984]). The Cremona group Bir(P1 × P1) is generated by τ : (x, y) (cid:55)→ (y, x) and the group B of birational transformations preserving the fibration given by the first projection and the following relations form a complete system of relations: • relations inside the groups Aut(P1 × P1) and B, • (τ ◦ ( 1 • (τ ◦ (−x, y − x))3 = id. x ))3 = id, y , y A gap in the original proof of Theorem 1.4 had been detected and closed by Lamy ([L2010]). A presentation of Bir(P2) in the form of a generalised amalgam was given in the following statement: Theorem 1.5 ([W1992]). The group Bir(P2) is the free product of PGL3(k), Aut(P1 × P1) and J amalgamated along their pairwise intersections in Bir(P2). In [BF2013] the authors introduced the Euclidean topology on the Cremona group over a locally compact local field. With respect to this topology, Bir(P2) is a Hausdorff topological group and the restriction of the Euclidean topology to any algebraic subgroup is the classical Euclidean topology. In order to show that Bir(P2) is compactly presentable with respect to the Euclidean topology, Zimmer- mann proved the following: Theorem 1.6 ([Z2016]). The Cremona group Bir(P2) is isomorphic to the amalgamated product of Aut(P2), Aut(F2), Aut(P1×P1) along their pairwise intersection in Bir(P2) modulo the relation τ στ σ, where τ ∈ Aut(P2) is the coordinate permutation τ : [x : y : z] (cid:55)→ [z : y : x]. A NEW PRESENTATION OF THE PLANE CREMONA GROUP 3 We would also like to mention the paper [IKT94] by Iskovskikh, Kabdykairov and Tregub, in which the authors present a list of generators and relations of Bir(P2) over arbitrary perfect fields. 1.3. Gizatullin-homomorphisms between Cremona groups. In this Section we recall a result of Gizatullin in order to illustrate how our presentation can be used for explicit calculations. Throughout Section 1.3 we assume k to be algebraically closed and of characteristic (cid:54)= 2. In [G1990], Gizatullin considers the following question: Can a given group-homomorphism ϕ : PGL3(k) → PGLn+1(k) be extended to a group-homomorphism Φ : Cr2(k) → Crn(k)? He answers this question positively if ϕ is the projective representation induced by the regular action of PGL3(k) on the space of plane conics, plane cubics or plane quartics. In order to construct these homomorphisms he uses the following construction. Let Symn be the k-algebra of symmetric n × n-matrices and define the variety S2(n) to be the quotient (Symn)3// GLn(k) where the regular action of GLn(k) is given by C · (A1, A2, A3) = (CA1C T , CA2C T , CA3C T ). Lemma 1.7. The variety S2(n) is a rational variety of dimension (n + 1)(n + 2)/2 − 1. Proof. The dimension of (Symn)3 is 3n(n + 1)/2. Since the dimension of GLn(k) is n2 and the action of GLn(k) on (Symn)3 has finite kernel, the dimension of S2(n) is (n + 1)(n + 2)/2 − 1. Define the subvariety X ⊂ (Symn)3 given by triplets of the form (id, d, A), where d ∈ D2 and consider the following maps: X (cid:44)→ (Symn)3 π−→ S2(n). Let U ⊂ (Symn)3 be the open dense subset of triplets of the form (A1, A2, A3) such that A1 and A2 are invertible. We note that π(X) = π(U ). Indeed, every symmetric element A1 ∈ GLn(k) is of the form CC T for some C ∈ GLn(k) and for every symmetric matrix A2 ∈ GLn(k) there exists an orthogonal matrix S ∈ On(k) such that SA2ST is diagonal. Consider the open dense subset V ⊂ X of elements of the form (id, d, A), where A ∈ Symn and d = (dij) an element in D2 such that dii (cid:54)= djj for all i (cid:54)= j. This condition ensures that the centralizer G of d in the orthogonal group On(k) consists only of diagonal elements. The group G is therefore an abelian group of exponent 2. A Theorem of Fischer ([F1915]) states that the quotient of AN by G is a rational variety. The image π(V ) is still dense in S2(n) and the restriction of π to V is the quotient map of the action of G on V given by conjugation. (cid:3) The theorem of Fischer therefore implies that S2(n) is rational. Note that Lemma 1.7 implies that S2(n) has the same dimension as the space of all plane curves of A linear transformation g = [g1 : g2 : g3] ∈ PGL3(k) induces an automorphism on (Symn)3 by g(A1, A2, A3) := (g1(A1, A2, A3), g2(A1, A2, A3), g3(A1, A2, A3)). This automorphism commutes with the action of GLn(k), so we obtain a regular action of PGL3(k) on S2(n). Our main theorem allows now to give a short proof of the following result of Gizatullin: Proposition 1.8 ([G1990]). This regular action of PGL3(k) extends to a rational action of Bir(P2) on S2(n). Proof. We define the birational action of the element σ on S2(n) by 2 , A−1 3 ). (A1, A2, A3) (cid:55)(cid:57)(cid:57)(cid:75) (A−1 1 , A−1 degree n. In order to see that this indeed defines a rational action of Cr2(k) on S2(n) it is enough, by our main theorem, to check that the relations (1) to (5) are satisfied. The relations (1) to (4) are straightforward to check. For relation (5) we calculate for general A1, A2, A3 ∈ Symn: 3 − A−1 σhσ(A1, A2, A3) = ((A−1 1 )−1, (A−1 2 )−1, A3) 3 − A−1 and hσh(A1, A2, A3) = (A−1 3 − (A3 − A1)−1, A−1 3 − (A3 − A2)−1, A−1 3 ) = (A3 − A3(A3 − A1)−1A3, A3 − A3(A3 − A2)−1A3, A3) 4 CHRISTIAN URECH AND SUSANNA ZIMMERMANN One calculates that the two expressions are the same. Hence relation (5) is satisfied. (cid:3) Lemma 1.7 and Proposition 1.8 imply that there is a rational Cr2(k)-action on PN for all N = (n + 1)(n + 2)/2 − 1. In [G1990] Gizatullin shows using classical geometry that, for n = 2, 3 and 4, the PGL3(k)-action on S2(n) is conjugate to the PGL3(k)-action on the space of plane conics, plane cubics and plane quartics. It is an interesting question whether the PGL3(k)-actions on S2(n) are conjugate to the PGL3(k)-actions on the space of plane curves of degree n. A positive answer would give a rational action of Cr2(k) on the space of all plane curves preserving degrees. It would also be interesting to look at the geometrical properties of these rational actions. The case of the rational action of Cr2(k) on the space of plane conics has been studied in [U2016]. Acknowledgements: The authors would like to express their warmest thanks to Serge Cantat and Jérémy Blanc for many interesting discussions and for pointing out a mistake in the first version of the proof. They are also grateful to Marat Gizatullin, Mattias Hemmig and Anne Lonjou for many discussions. Throughout all the sections we work over an algebraically closed field k. 2. Useful relations 2.1. Preliminaries. There is a general formula for the degree of a composition of two Cremona transformations [A2002, Corollary 4.2.12], but the multiplicities of the base-points of the composition are hard to compute in general. However, if we compose an arbitrary birational map with a map of degree 2 it is a rather straight forward calculation [A2002, Proposition 4.2.5]. We will only use the formula for de Jonquières maps. For a given transformation f ∈ Bir(P2), we denote by mp(f ) the multiplicity of f in the point p. Lemma 2.1. Let τ, f ∈ J be transformations of degree 2 and d respectively. Let p1, p2 be the base- points of τ different from [1 : 0 : 0] and q1, q2 the base-points of τ−1 different from [1 : 0 : 0] such that the pencil of lines through pi is sent by τ onto the pencil of lines through qi. Then deg(f τ ) = d + 1 − mq1(f ) − mq2 (f ) m[1:0:0](f τ ) = d − mq1(f ) − mq2 (f ) = deg(f τ ) − 1 mpi(f τ ) = 1 − mqj (f ), i (cid:54)= j. Proof. A de Jonquières map of degree d has multiplicity d − 1 in [1 : 0 : 0] and multiplicity 1 in every (cid:3) other of its base-points. The claim now follows from [A2002, Proposition 4.2.5]. Remark 2.2. Keep the notation of Lemma 2.1. Let Λ be the linear system of f. Then Lemma 2.1 translates to: deg(f τ ) = deg((τ−1)(Λ)) = d + 1 − mq1 (Λ) − mq2 (Λ) m[1:0:0]((τ−1)(Λ)) = d − mq1(Λ) − mq2(Λ) = deg((τ−1)(Λ)) − 1 In particular, as the multiplicity of Λ in a point different from [1 : 0 : 0] is zero or one, we obtain mpi((τ−1)(Λ)) = 1 − mqj (Λ), i (cid:54)= j. deg(Λ) + 1, mq1(Λ) = mq2(Λ) = 0 deg(Λ), mq1(Λ) + mq2(Λ) = 1 deg(Λ) − 1, mq1(Λ) = mq2(Λ) = 1. deg((τ−1)(Λ)) = Note as well that Bézout theorem implies that [1 : 0 : 0] and any two other base-points of f are not collinear since [1 : 0 : 0] is a base-point of multiplicity d − 1 all other base-points are of multiplicity 1 and a general member of Λ intersects a line in d points counted with multiplicity. A NEW PRESENTATION OF THE PLANE CREMONA GROUP 5 Notation 2.3. We use the following picture to work with relations. If f1, . . . , fn ∈ {σ} ∪ PGL3(k) and f1 ··· fn = 1 in the group (cid:104)σ, PGL3(k) (1) − (5)(cid:105), we say that the commutative diagram fn / fn−1 / f1 or fn / fn−1 / f2 Id −1 f 1 is generated by relations (1)–(5) and that the expression f1 ··· fn = 1 is generated by relations (1)–(5). 2.2. Relations. When dealing with quadratic maps, some relations among them appear quite natu- rally. We now take a closer look at a few of them and show that they are in fact generated by relations (1)–(5). Lemma 2.4. Let g ∈ PGL3(k) ∩ J and suppose that the map g(cid:48) := σgσ is linear. Then g(cid:48) = σgσ is generated by relations (1)–(4). In other words, the commutative diagram is generated by relations (1)–(4). σ σ g α(cid:48) Proof. The map σgσ being linear means by Lemma 2.1 that the base points of σg are the same as the base points of σ, which are {[1 : 0 : 0], [0 : 1 : 0], [0 : 0 : 1]}. As g ∈ J , it fixes [1 : 0 : 0] and thus permutes the points [0 : 1 : 0], [0 : 0 : 1]. Hence g = dτ for some τ ∈ S3 ∩ J and d ∈ D2. Using our relations we obtain σgσ (1) = σdτ σ (3),(4) = d−1τ = g(cid:48). (cid:3) Lemma 2.5. Let g ∈ PGL3(k) ∩ J such that deg(σgσ) = 2. Suppose that no three of the base-points of σ and σg are collinear. Then there exist g(cid:48), g(cid:48)(cid:48) ∈ PGL3(k) ∩ J such that σgσ = g(cid:48)σg(cid:48)(cid:48) and this expression is generated by relations (1)–(5). In other words, the commutative diagram g σ σ g(cid:48)(cid:48) σ g(cid:48) is generated by relations (1)–(5). Proof. The assumption deg(σgσ) = 2 implies by Lemma 2.1 that σ and σg have exactly two common base-points, among them [1 : 0 : 0] because σg and σ are de Jonquières maps. Up to coordinate permutations, the second point is [0 : 1 : 0]. More precisely, there exist two coordinate permutations τ1, τ2 ∈ S3∩J such that τ1gτ2 fixes the points [1 : 0 : 0] and [0 : 1 : 0]. Hence there are a1, a2, b1, b2, c ∈ k such that τ1gτ2 = [a1x + a2z : b1y + b2z : cz]. The assumption that no three of the base-points of σ and σg are collinear implies that a2b2 (cid:54)= 0. So there exist two elements d1, d2 ∈ D2 such that τ1gτ2 = d1hd2. Using relations (1)–(5) we obtain 2 τ−1 2 . 1 τ1gτ2τ−1 2 σ σgσ = στ−1 1 σhσd−1 2 τ−1 1 d−1 1 hσhd−1 (1) (5) = τ−1 2 The claim follows with g(cid:48) := τ−1 Lemma 2.6. Let g1, . . . , g4 ∈ PGL3(k) ∩ J such that σg2σg1 = g4σg3σ is of degree 3 and (cid:3) = στ−1 1 d−1 1 d1hd2τ−1 2 σ 1 h and g(cid:48)(cid:48) := hd−1 = τ−1 1 d−1 2 τ−1 2 . (3),(4) • no three of the base-points of σg2 and σ are collinear • no three of the base-points of σg1 and σ are collinear. Then the relation σg2σg1 = g4σg3σ is generated by relations (1)–(5). / 5 5 / / / / / / 5 5 / / / / / / / 7 7 / / / / / /   ? ? / / / / / / 6 CHRISTIAN URECH AND SUSANNA ZIMMERMANN Proof. Let p0 := [0 : 0 : 1]. The maps g1, . . . , g4 ∈ PGL3(k) ∩ J fix it. By Lemma 2.1, the assumption deg(σg2σg1) = 3 is equivalent to the maps σg2 and g−1 1 σ having exactly one common base-point, namely p0. For the same reason, the maps g4σ and σg−1 3 have only p0 as a common base-point. The assumptions on the base-points imply that the maps σg2σg1 and g3σg−1 1 σ have only proper base-points. Therefore, the maps g4σg3σ and σg−1 4 σg2 have only proper base-points and we obtain that for (f, g) ∈ {(g−1 1 σ, σg2), (σg1, g3σ), (σg−1 3 , g4σ), (g−1 4 )} 2 σ, σg−1 no three base-points of the maps f and g are collinear. Hence there exists a quadratic map τ1 that has exactly two base-points in common with σg1 and with g3σ. We write τ1 = h1σh2 for some h1, h2 ∈ PGL3(k) ∩ J . Then σg1τ−1 are quadratic again. Let h3, . . . , h8 ∈ PGL3(k) ∩ J and define the quadratic maps τi := h2iσh2i−1 ∈ J , i = 2, . . . , 4. We can chose h3, h4 such that τ−1 2 τ1 = h−1 3 σh−1 • the diagram below commutes • each of the τi has exactly two common base-points (one of them p0) with each of the two other 4 h2σh3 = σg1. Analogously we chose h4, . . . , h8 such that: and g3στ−1 1 1 maps departing from the same corner. σg1 τ1 g4σ g3σ τ3 τ2 τ4 σg2 Lemma 2.5 applied to the triangles yields that each of them is generated by relations (1)–(5). In (cid:3) particular, the above commutative diagram is generated by relations (1)–(5). Remark 2.7. Let a1, a2, b1, b2, c ∈ k. The map g = [a1x + a2z : b1y + b2z : cz] ∈ PGL3(k) ∩ J is a de Jonquiéres map. Furthermore, observe that if a2 = b2 = 0, then σgσ is linear. Otherwise, σgσ is a quadratic map. In fact, if a2b2 (cid:54)= 0, all its base-points are in P2; this corresponds to the case that is treated in Lemma 2.5. If a2 = 0, the map σgσ has a base-point infinitely near to [1 : 0 : 0], and if b2 = 0 it has a base-point infinitely near to [0 : 1 : 0]. Figure 1 displays the constellation of the base-points of σg and σ. [0 : 0 : 1] [0 : 0 : 1] [0 : 1 : 0] [0 : 1 : 0] g−1([0 : 0 : 1]) = [0 : − b2 b1c : 1 c ] g−1([0 : 0 : 1]) = [− a2 a1c : 0 : 1 c ] [1 : 0 : 0] [1 : 0 : 0] Figure 1. The base-points of σg if a2 (cid:54)= 0, b2 = 0 (left) and if a2 = 0, b2 (cid:54)= 0 (right). The following two lemmas provide a solution of how to treat these two cases. Definition 2.8. For a de Jonquières transformation f ∈ J and a line l ⊂ P2 we define the discrepancy N (f, l) as follows: If f (l) is a line, we set N (f, l) := 0. If f (l) = {p} is a point, then there exists a sequence of K blow-ups π : S → P2 and an induced birational map f : P2 (cid:57)(cid:57)(cid:75) S, such that the following diagram commutes: f f P2 S π P2 / /         / / ? ? _ _  / / > > A NEW PRESENTATION OF THE PLANE CREMONA GROUP 7 and such that the strict transform f (l) ⊂ S is a curve. We define N (f, l) to be the least number K of blow-ups necessary for this construction. Remark 2.9. Let f ∈ J and let l ⊂ P2 be a line. If N (f, l) ≥ 1, then the point P := f (l) is a base- point of f−1. More precisely, if π and f are as above, then f−1 has a base-point in the (N (f, l)− 1)-th neighbourhood of f (l) and the exceptional divisor of this base-point is f (l). Lemma 2.10. Let f ∈ J be a de Jonquières transformation and l ⊂ P2 a line such that N (f, l) ≥ 1, and let g ∈ PGL3(k) ∩ J . Denote by η : S → P2 the blow-up of the base-points of σg. Then (1) N (σgf, l) = N (f, l) + 1 if and only if the point f (l) ∈ P2 is on a line contracted by σg but is not a base-point of σg. (2) N (σgf, l) = N (f, l) − 1 if and only if the point f (l) ∈ P2 is a base-point of σg and (η−1f )(l) is a line or a point not on the strict transform of a line contracted by σg. (3) N (σgf, l) = N otherwise. Proof. If p := f (l) is not a base-point of σg, then its image by σg is a base-point of (σgf )−1 with N (σgf, l) ≥ N (f, l). The inequality is strict if and only if p is on a line contracted by σg. If p is a base-point of σg the discrepancy decreases strictly unless η−1f contracts l onto an inter- (cid:3) section point of the exceptional divisor of p and the strict transform of a line contracted by σg. Lemma 2.11. Let g1, . . . , gm ∈ PGL3(k)∩J such that gmσgm−1σ ··· σg1 = id. Then there exist linear maps h1, . . . , hm ∈ PGL3(k) ∩ J and a de Jonquières transformation ϕ ∈ J such that ϕgmσgm−1σ ··· σg1ϕ−1 = hmσhm−1σ ··· h1σ, and such that the following properties are satisfied: (1) the above relation is generated by relations (1)–(5), (2) deg(σhiσhi−1 ··· σh1) = deg(σgiσgi−1 ··· σg1) for all i = 1, . . . , m, (3) (σhiσhi−1 ··· σh1)−1 does not have any infinitely near base-points above [1 : 0 : 0] for all i = 1, . . . , m. Proof. Define fi := σgi ··· σg1 for i = 1, . . . , m − 1. We will construct the hi such that (3) is satisfied and then show that the other properties hold as well. If none of the fi has a base-point that is infinitely near to [1 : 0 : 0], we set hi := gi for all i and all the claims are trivially satisfied. Otherwise, there exist k ≥ 1 such that fi has a base-point in the k-th neigbourhood of [1 : 0 : 0] for some i. Assume k to be maximal with this property. We will lower k and then proceed by induction. For this we pick two general points s0, t0 ∈ P2 and define si := fi(s0) = σgiσ ··· σg1(s0), sm−1 := gmσgm−1(sm−2), tm−1 := gmσgm−1(tm−2). ti := fi(t0) = σgiσ ··· σg1(t0), for i = 1, . . . , m − 2, Since s0 and t0 are general points of P2, the si and ti are general as well. For all i = 1, . . . , m− 1, there exist αi ∈ PGL3(k) ∩ J that send p0, p1, p2 onto p0, si, ti respectively. The map τi := αiσα−1 i ∈ J is a quadratic involution with base-points p0, si, ti. The assumption gmσ ··· σg1 = Id implies that sm−1 = s0 and tm−1 = t0, so we may choose αm−1 = α0 and obtain τm−1 = τ0. We define m−2 ∈ J θm−1 := τm−1(gmσgm−1)τ−1 i = 1, . . . , m − 2, i−1 ∈ J , θi := τiσgiτ−1 and consider the sequence Id = τ0gmσgm−1σ ··· σg1τ−1 The situation is visualised in the following commutative diagram: m−1τm−1σgm−1τ−1 0 = τ0gmτ−1 m−2τm−2σgm−2 . . . σg1τ−1 0 = θm−1θm−2 ··· θ1 8 CHRISTIAN URECH AND SUSANNA ZIMMERMANN σg1 / σg2 / / σgn−1 / σgn / / gmσgm−1/ τ0 τ1 τ2 τn−2 τn−1 τn τm−2 τm−1=τ0 θ1 θ2 θn−1 θn θm−1 We claim that for each i = 1, . . . , m − 1, the map θi ∈ J is a quadratic map with only proper base- points in P2. The only common base-point of the transformations σgi and τi is p0, hence by Lemma 2.1 the map τi−1g−1 i σ is of degree 3 and has p0, p1, p2, si+1, ti+1 as base-points. The common base-points of the two transformations τi−1g−1 i σ and τi are the points p0, si+1, ti+1. Therefore, the composition θi = τiσgiτ−1 i−1 is quadratic and its base-points are p0 and the image under τi−1 of the two base-points of σgi different from p0. [p0,p1,p2] σgi [p0,si−1,ti−1] τi−1 [p0,p1,p2,si,ti] [p0,si,ti] τi θi Since si−1 and ti−1 are general points, these images are proper points of P2. So for each i = 1, . . . , m−1, we find βi, γi ∈ PGL3(k) ∩ J such that θi = βiσγi. / σgn−1 / / gmσgm−1/ σgn / σg1 / σg2 / τ0 τ1 τ2 τn−2 τn−1 τn τm−1 τm=τ0 β1σγ1 β2σγ2 βn−1σγn−1 βnσγn βm−1σγm−1 We write h1 := γ1, hm := βm−1, hi := γiβi−1, i = 2, . . . , m − 1. By Remark 2.2, we have deg(τiσgi ··· σg1) = deg(σgi ··· σg1) + 1 and therefore deg(θi ··· θ1) = deg(τiσgi ··· σg1τ−1 0 ) = deg(σgi ··· σg1). Hence we finally obtain deg(σhiσ ··· σh1) = deg(β−1 i+1θi ··· θ1) = deg(σgi ··· σg1). This is part (2) of the lemma. By Lemma 2.6, all squares in the above diagram are generated by relations (1)–(5), hence the whole diagram is generated by these relations. This is part (1). Let us look at part (3). The base-points of (θi ··· θ1)−1 are [1 : 0 : 0] and the proper images of the base-points of (σgi ··· σg1)−1 different from [1 : 0 : 0]. By construction of the τi, none of the base-points of (σgi ··· σg1)−1 different from [1 : 0 : 0] are on a line contracted by τi. Moreover, the base-points of (σgi ··· σg1)−1 in the first neighbourhood of [1 : 0 : 0] are sent by τi onto proper points of P2. In conclusion, (θi ··· θ1)−1 has less base-points infinitely near [1 : 0 : 0]. Induction step done. The same (cid:3) argument yields part (4). Lemma 2.12. Let g1, . . . , gm ∈ PGL3(k)∩J be linear de Jonquières transformations. Then there exist linear de Jonquières transformations h1, . . . , hm ∈ PGL3(k) ∩ J , and a de Jonquières transformation ϕ ∈ J such that ϕgmσgm−1σ ··· σg1ϕ−1 = hmσhm−1σ ··· σh1 and such that the following properties hold: (1) the above relation is generated by relations (1)–(5), (2) deg(σhi ··· σh1) = deg(σgi ··· σg1) for all i = 1, . . . , m, (3) (σhiσhi−1 ··· σh1)−1 does not have any infinitely near base-points for all i = 1, . . . , m. /   /     / /   /     / /     / / / / / / / / / / / / / / / /     t t / / /   /     / /   /     / /     / / / / / / / / / / / / / / A NEW PRESENTATION OF THE PLANE CREMONA GROUP 9 Proof. Define fi := σgiσgi−1 ··· σg1 for i = 1, . . . , m. After applying Lemma 2.11 we may assume that none of the transformations f−1 has base-points infinitely near to [1 : 0 : 0]. We define i N := max{N (fi, l) i = 1, . . . , m, k := max{i N (fi, l) = N}, # := number of lines l ⊂ P2 with N (fk, l) = N l ⊂ P2 line} In other words, N is the maximal discrepancy, k is the maximal number of all i such that fi has N as a discrepancy and # ≥ 1 is the number of lines contracted by fk with discrepancy N. We do induction over the lexicographically ordered triple (N, k, #). If N = 0 there are no infinitely near base-points and we are done. Suppose that N ≥ 1. We denote by lk ⊂ P2 a line such that N = N (fk, lk) and define pk := fk(lk) ∈ P2. Let s ∈ P2 be a general point. Then there exists a linear de Jonquières map h ∈ PGL3(k)∩J such that σh has [1 : 0 : 0], pk and s as base-points. By definition of k and N we have N > N (fk+1, lk). Since fk+1 = σgk+1fk, Lemma 2.10 implies that pk is also a base-point of σgk+1. It follows from Lemma 2.1 that deg(σgk+1(σh)−1) = 2. Since s is a general point, we find a1, a2 ∈ PGL3(k) ∩ J such that σgk+1(σh)−1 = a2σa1. Note that N (σhσgk ··· σg1, lk) = N − 1 by Lemma 2.10. If N (σgk−1 ··· σg1, lk) < N, we find, by similar arguments as above, b1, b2 ∈ PGL3(k)∩J such that σhσgk = b2σb1 σg1 / σg2 / / ··· ··· σgk−1 / σgk / σgk+1 / σgk+2/ / ··· b2σb1 σh a2σa1 The new sequence gmσgm−1 ··· σgk+2a2σ(a1b2)σ(b1gk−1)σgk−2 ··· σg1 has maximal discrepancy at most N. If it is N, that means that there is either another line l ⊂ P2 such that N (σgk ··· σg1, l) = N, in which case # has decreased. Or that N (σgk ··· σg1, l) < N for all lines l and N (σgi ··· σg1, li) = N for some i < k and some line li. In any case, the triple (N, k, #) decreases. Note as well that deg(σhσgk ··· σg1) = deg(σgk ··· σg1). If N (σgk−1 ··· σg1, lk) = N, there are two options by Lemma 2.10: (a) The point pk is a base-point of (σgk)−1 and, if η : S → P2 is the blow-up of the base-points of (σgk)−1, then (η−1fk)−1(lk) is the intersection point of an exceptional divisor and the strict transform of a line contracted by (σgk)−1. (b) The point pk is not a base-point of (σgk)−1 and is not on its contracted lines. In case (a) we can proceed analogously as in the case above, where N (σgk−1 ··· σg1, lk) < N. So assume now that we are in case (b). The image p(cid:48) k of pk by (σgk)−1 is a proper point of P2. The point s ∈ P2 is general, hence also its image s(cid:48) by (σgk)−1 is a general point of P2. In particular, there exists h(cid:48) ∈ PGL3(k) ∩ J such that σh(cid:48) has base-points [1 : 0 : 0], p(cid:48) k and s(cid:48). The map σhσgk(σh(cid:48))−1 is of degree 2, and since the points s and s(cid:48) are in general position, we find c1, c2 ∈ PGL3(k) ∩ J such that σhσgk(σh(cid:48))−1 = c2σc1. Moreover, we have N (σh(cid:48)σgk−1 ··· σg1, lk) = N − 1. Also note that deg(σh(cid:48)σgk−1 ··· σg1) = deg(σgk−1 ··· σg1). We proceed like this until we find an index i < k such that N (σgi ··· σg1, lk) < N. Such an index has to exist because σgm ··· σg1 = id. There, the preimage of pk is a base-point of (σgi ··· σg1)−1, and we repeat the construction of the former case. It yields a new sequence with maximal discrepancy at most N. If it is equal to N then either at the index k and another line or at an index smaller than k. In other words, the triple (N, k, #) has decreased. / /   /   / @ @ 10 CHRISTIAN URECH AND SUSANNA ZIMMERMANN The commutative diagram below illustrates the procedure if N (σgk−2 ··· σg1, lk) < N. σg1 / σg2 / / ··· ··· σgk−1 / σgk / σgk+1 / σgk+2/ / ··· b2σb1 σh a2σa1 σh(cid:48) c2σc1 It remains to remark that all the squares and triangles are generated by relations (1)–(5) by (cid:3) Lemma 2.5 and Lemma 2.6. 3. Proof of Theorem 1.1 Proposition 3.1. Let g1, . . . , gm ∈ PGL3(k) ∩ J and suppose that Id = gmσgm−1σ ··· σg1 as maps. Then this expression is generated by relations (1)–(5). Proof. Let Λ0 be the complete linear system of lines in P2 and define i = 1, . . . , m. Λi := σgi−1σ ··· σg1(Λ0), Let δi := deg(Λi), D := max{δi i = 1, . . . , m} and n := max{i δi = D}. We do induction on the lexicographically ordered set of pairs of positive integers (D, n). If D = 1, then m = 1 and there is nothing to prove. Let D > 1. By Lemma 2.12, we can suppose that for each i = 1, . . . , m the transformation (giσgi−1σ ··· σg1)−1 does not have any infinitely near base-points, and we can do this without in- creasing the pair (D, n). Equivalently, each Λi does not have any infinitely near base-points. All the gi are de Jonquières maps and therefore fix [1 : 0 : 0], the maps σgi and σ always have [1 : 0 : 0] as common base-point. In particular, Lemma 2.1 yields deg(σgiσ) ≤ 3 for all i = 1, . . . , m. We will look at the three distinct cases deg(σgnσ) = 1, 2 and 3. (a) If deg(σgnσ) = 1, Lemma 2.4 implies that we can replace the word σgnσ by the linear map n = σgnσ using only relations (1)–(5). We obtain a new pair (D(cid:48), n(cid:48)) where D(cid:48) ≤ D and if D = D(cid:48) g(cid:48) then n < n(cid:48). (b) If deg(σgnσ) = 2, then σ and σgn have exactly two common base-points, among them [1 : 0 : 0]. Up to coordinate permutations, these two points are [1 : 0 : 0] and [0 : 1 : 0]. More precisely, there exist two permutations of coordinates τ1, τ2 ∈ S3 ∩ J such that τ1gnτ2 fixes the points [1 : 0 : 0] and [0 : 1 : 0]. Hence there are a1, a2, b1, b2, c ∈ k such that Using relations (1) and (3) we get τ1gnτ2 = [a1x + a2z : b1y + b2z : cz]. gmσ ··· σgn+1στ−1 1 τ1gnτ2τ−1 2 σ ··· g1 = gmσ ··· σ(gn+1τ−1 1 )σ(τ1gnτ2)σ(τ−1 2 gn−1)σ ··· g1. This replacement does not change the pair (D, n). So, we may assume that τ1 = τ2 = id and gn = [a1x + a2z : b1y + b2z : cz]. By assumption, for i = 1, . . . , n, the maps giσgi−1σ ··· σg1 have no infinitely near base-points. It follows that Λn has no infinitely near base-points. (b1) If a2b2 (cid:54)= 0, then no three of the base-points of σ and σgn are collinear. By Lemma 2.5 there exist g(cid:48), g(cid:48)(cid:48) ∈ PGL3(k) such that we can replace the word σgnσ with the word g(cid:48)σg(cid:48)(cid:48) using relations (1)–(5). This yields a new pair (D(cid:48), n(cid:48)) where D(cid:48) ≤ D and if D = D(cid:48) then n(cid:48) < n. (b2) We want to see that a2b2 = 0 is impossible. Suppose that a2b2 = 0. Then q := g−1 n ([0 : 0 : 1]) is a base-point of σgn on a line that is contracted by (σgn−1)−1 (see Remark 2.7). Remark 2.2 implies that D − 1 = δn+1 = D + 1 − m[0:1:0](Λn) − mq(Λn) / / ! !   /   / / / @ @ A NEW PRESENTATION OF THE PLANE CREMONA GROUP 11 In particular, we have mq(Λn) = 1 Since it is not a base-point of σ, it is not a base-point of (σgn−1)−1. Hence its proper image by (σgn−1)−1 is a base-point of Λn−1. Because of a2b2 = 0, this point is an infinitely near point, a contradiction to our assumption that the Λi do not have any infinitely near base-points. (c) Assume that deg(σgnσ) = 3, so σ and σgn have one common base point, which is [1 : 0 : 0]. Denote by p0 = [1 : 0 : 0], p1 = [0 : 1 : 0], p2 = [0 : 0 : 1] the base-points of σ and by p0, q1, q2 the base-points of σgn. Remark 2.2 implies δn ≥ deg(Λn−1) = deg((σgn−1)−1(Λn)) = δn + 1 − mp1(Λn) − mp2 (Λn) δn > deg(Λn+1) = deg(σgn(Λn)) = δn + 1 − mq1(Λn) − mq2(Λn). Therefore, we obtain 1 = mq1 (Λn) = mq2 (Λn), 1 ≤ mp1(Λn) + mp2 (Λn) Choose i ∈ {1, 2} such that mpi(Λn) = 1. Then the points [1 : 0 : 0], pi and q1 are not collinear (because mp0(Λn) + mp1(Λn) + mq1(Λn) > δn) and there exists a g(cid:48) ∈ PGL3(k) ∩ J∗ such that σg(cid:48) has base-points [1 : 0 : 0], pi and q1. Consider Figure 2; we claim that the left pair and the right pair of maps satisfy the assumptions of case (b1), which implies that σg(σg(cid:48) n−1)−1 and σgnσg(cid:48) can be replaced by maps of the form hσh(cid:48) and h(cid:48)(cid:48)σh(cid:48)(cid:48)(cid:48) respectively for some linear maps h, h(cid:48), h(cid:48)(cid:48), h(cid:48)(cid:48)(cid:48) and the pair (D, n) decreases. [p0,p1,p2] Λn [p0,q1,q2] σgn [p0,pi,q1] σg(cid:48) σg(cid:48)(Λn) Λn+1 (σgn−1)−1 Λn−1 Figure 2. The two maps on the left and the two maps on the right satisfy the assumptions of case (b1). The maps σ and σg(cid:48) have two common base-points, namely p0 and pi. The maps σg(cid:48) and σgn have the base-points p0 and q1 in common. Remark 2.2 tells us that deg(σg(cid:48)(Λn)) = δn + 1 − mpi(Λn) − mq1(Λn) < δn = D. It remains to check that no three of the four points p0, p1, p2, q1 are collinear and that no three of the four points p0, pi, q1, q2 are collinear. Indeed, in the latter case all four points are base-points of Λn and the image of pi by σgn is a base-point of Λn+1, which does not have any infinitely near base-points. Therefore, no three of p0, pi, q1, q2 are collinear. In the first case, at least p0, pi, q1 are base-points of Λn and therefore not collinear by Remark 2.2. The points p1, p2, q1 and the points p0, p2, q1 are not collinear because Λn−1 has no infinitely near base-points. Thus, we can apply case (b1) to the maps σ (cid:3) and σg(cid:48) and to the maps σ and σgng(cid:48)−1. The pair (D, n) decreases. So far we have been working with de Jonquières maps only. Now we will use the structure of the plane Cremona group as an almost amalgamated product from Theorem 1.3 to prove the Main Theorem: Proof of Main Theorem. Let G = (cid:104)σ, PGL3(k) (1) − (5)(cid:105) be the group generated by σ and PGL3(k) divided by the relations (1)–(5), and let π : G → Bir(P2) be the canonical homomorphism that sends generators onto generators. It follows from Proposition 3.1 that sending an element of J onto its corresponding word in G is well defined. This yields a homomorphism w : J −→ G that satisfies z z   ( ( 12 CHRISTIAN URECH AND SUSANNA ZIMMERMANN π ◦ w = IdJ and is therefore injective. Consider the commutative diagram G w J? PGL3(k) PGL3(k) ∩ J where all unnamed homomorphisms are the canonical inclusions. The universal property of the amal- gamated product implies that there exists a unique homomorphism ϕ : PGL3(k) ∗PGL3(k)∩J J → G such that the following diagram commutes: G ∃ ! ϕ PGL3(k) ∗PGL3(k)∩J J w J PGL3(k) PGL3(k) ∩ J By Theorem 1.3, the group Bir(P2) is isomorphic to PGL3(k) ∗PGL3(k)∩J J divided by the relation τ στ σ, where τ = [y : x : z], which is a relation that holds as well in G. So ϕ factors through the quotient PGL3(k) ∗PGL3(k)∩J J /(cid:104)τ στ σ(cid:105) and we thus obtain a homomorphism ¯ϕ : Bir(P2) → G. In fact, the homomorphisms π and ¯ϕ both send generators to generators: G π−→ Bir(P2), σ (cid:55)−→ σ, α (cid:55)−→ α ∀ α ∈ PGL3 (cid:55)−→ σ, α (cid:55)−→ α ∀ α ∈ PGL3. Bir(P2) ¯ϕ−→ G, σ Hence ¯ϕ and π are isomorphisms that are inverse to each other. (cid:3) References [A2002] M. Albert-Carramiñana: Geometry of the Plane Cremona Maps, Springer Verlag Berlin Heidelberg, 2002. [B2012] J. Blanc: Simple Relations in the Cremona Group, Proceedings of the American Mathematical Society, 140 (2012), 1495–1500. [BF2013] J. Blanc, J.-P. Furter: Topologies and structures of the Cremona groups, Ann. of Math. 178 (2013), no. 3, 1173–1198. [C2013] Yves de Cornulier: The Cremona Group is not an amalgam, appendix of the article Serge Cantat, Stéphane Lamy: Normal subgroups in the Cremona group, Acta Math., 210 (2013), 31–94. [C1901] G. Castelnuovo: Le trasformazioni generatrici del gruppo cremoniano nel piano, Atti della R. Accad. delle Scienze di Torino, 36:861–874, 1901. [F1915] E. Fischer: Die Isomorphie der Invariantenkörper der endlichen abelschen Gruppen linearer Transformatio- nen, Nachrichten von der Gesellschaft der Wissenschaften zu Göttingen, Mathematisch-Physikalische Klasse, 77–80, 1915. [G1982] M. Gizatullin: Defining relations for the Cremona group of the plane (Russian), Izv. Akad. Nauk SSSR Ser. Mat. 46 (1982), no. 5, 909–970, 1134; english translation in Math. USSR 21 (1983), no.2, 211–268. [G1990] M. Gizatullin: On some tensor representations of the Cremona group of the projective plane, New trends in algebraic geometry (Warwick, 1996). London Math. Soc. Lecture Note Ser. Vol. 264, 111–150, 1999. [I1984] V.A. Iskovskikh: Proof of a theorem on relations in the two-dimensional Cremona group, Uspekhi Mat. Nauk 40 (1985), no. 5(245), 255–256; english translation: Comm. of the Moscow Math. Soc. 1984. [IKT94] V. A. Iskovskikh, F. K. Kabdykairov, S. L. Tregub: Relationsin the two dimensional Cremona group over a perfect field, Russian Acad. Sci. Izv. Math 42 (3) (1994), 427–478. [L2010] S. Lamy: Groupes de transformations birationnelles de surfaces, Mémoire d'habilitation à diriger des recherches, l'université Claude Bernarde Lyon 1, 2010. [U2016] C. Urech: On homomorphisms between Cremona groups, Annales de l'Institut Fourier to appear. [W1992] D. Wright: Two-dimensional Cremona groups acting on simplicial complexes, Trans. Amer. Math. Soc. 331 (1992), no. 1, 281–300. _ o o ?  O O ? _ o o  ? O O h h o o o o ?  O O T T ? _ o o  ? O O A NEW PRESENTATION OF THE PLANE CREMONA GROUP 13 [Z2016] S. Zimmermann: The Cremona group is compactly presented, J. Lond. Math. Soc. (2) 93 (2016), no. 1, 25–46. Christian Urech, Departement Mathematik und Informatik, Spiegelgasse 1, 4051 Basel, Switzerland E-mail address: [email protected] Susanna Zimmermann, Faculté des sciences, 2 Bd de Lavoisier, 49045 Angers cedex 01, France E-mail address: [email protected] 

1507.06891

1

1507

Wall divisors and algebraically coisotropic subvarieties of irreducible holomorphic symplectic manifolds

Rational curves on Hilbert schemes of points on $K3$ surfaces and generalised Kummer manifolds are constructed by using Brill-Noether theory on nodal curves on the underlying surface. It turns out that all wall divisors can be obtained, up to isometry, as dual divisors to such rational curves. The locus covered by the rational curves is then described, thus exhibiting algebraically coisotropic subvarieties that deform to general small deformations of the manifold. This provides strong evidence for a conjecture by Voisin concerning the Chow ring of irreducible holomorphic symplectic manifolds. Some general results concerning the birational geometry of irreducible holomorphic symplectic manifolds are also proved, such as a non-projective contractibility criterion for wall divisors.

math.AG

math

WALL DIVISORS AND ALGEBRAICALLY COISOTROPIC SUBVARIETIES OF IRREDUCIBLE HOLOMORPHIC SYMPLECTIC MANIFOLDS ANDREAS LEOPOLD KNUTSEN, MARGHERITA LELLI-CHIESA, AND GIOVANNI MONGARDI Abstract. Rational curves on Hilbert schemes of points on K3 surfaces and generalised Kummer manifolds are constructed by using Brill-Noether theory on nodal curves on the underlying surface. It turns out that all wall divisors can be obtained, up to isometry, as dual divisors to such rational curves. The locus covered by the rational curves is then described, thus exhibiting algebraically coisotropic subvarieties. This provides strong evidence for a conjecture by Voisin concerning the Chow ring of irreducible holomorphic symplectic manifolds. Some general results concerning the birational geometry of irreducible holomorphic symplectic manifolds are also proved, such as a non- projective contractibility criterion for wall divisors. 0. Introduction Rational curves play a pivotal role in the study of the birational geometry and the Chow ring of algebraic varieties. The present paper concerns a specific class of varieties, namely, irreducible holomorphic symplectic (IHS) manifolds and, more precisely, Hilbert schemes of points on K3 surfaces and generalised Kummer manifolds (cf. §1), and is focused on some special rational curves arising from the Brill-Noether theory of normalisations of curves lying on K3 and abelian surfaces. In order to treat the two cases simultaneously, we introduce the following notation: we set ε = 0 (respectively, ε = 1) when S is a K3 (resp., abelian) surface, and we denote by S[k] the Hilbert scheme of k points ε on S when ε = 0 and the 2k-dimensional generalised Kummer variety on S when ε = 1. In the last few years, some classical results concerning (−2)-curves on K3 surfaces have been generalised to higher dimension and in particular it was shown that rational curves fully control the birational geometry of IHS manifolds. More precisely, Ran [Ra] proved that extremal rational curves can be deformed together with the ambient IHS manifold, and this was exploited by Bayer, Hassett and Tschinkel [BHT] in order to determine the structure of the ample cone. The same result was independently obtained by the third named author [Mo1] using intrinsic properties of IHS manifolds and a deformation invariant class of divisors, the so-called wall divisors (cf. Definition 2.2), which contains all divisors dual to extremal rays. This class of divisors was also studied by Amerik and Verbitsky [AV], who investigated fibres of extremal contractions. Indeed, the MBM classes in [AV] turn out to be precisely the dual curve classes to wall divisors, cf. Remark 2.4. By results of Bayer and Macrì [BM, BM2] and Yoshioka [Yo3], moduli spaces of stable objects in the bounded derived category of a K3 or abelian surface S provide examples of deformations of S[k] ε and the space of stability conditions can be used towards computing their ample cones. In this paper we use Brill-Noether theory of nodal curves on abelian and K3 surfaces in order to exhibit rational curves in S[k] ε and describe, in many cases, the locus they cover. Our construction proceeds as follows. Let (S, L) be a general primitively polarized K3 or abelian surface of genus g1 k+ε. Existence of a family of such curves having the expected dimension (and satisfying certain additional properties) has been proved in [CK, KLM] under suitable conditions on the triple (p, k, δ), p := pa(L) and let C ∈ L be a δ-nodal curve whose normalization eC has a linear series of type cf. Theorem 3.1. Any pencil of degree k + ε on eC defines a rational curve in S[k] Rp,δ,k := L − (p − δ + k − 1 + ε)rk, ε , whose class is 1 2 ANDREAS LEOPOLD KNUTSEN, MARGHERITA LELLI-CHIESA, AND GIOVANNI MONGARDI in terms of the canonical decomposition N1(S[k] particular, its Beauville-Bogomolov square is easily computed to be ε ) ≃ N1(S) ⊕ Z[rk], cf. (8) and Lemma 3.3. In q(Rp,δ,k) = 2(p − 1) − (p − δ + k − 1 + ε)2 2(k − 1 + 2ε) , cf. (18). An important additional feature of the rational curves obtained in this way is that they move in a family of dimension precisely 2k − 2 in S[k] ε and thus survive in all small deformations of S[k] ε that keep Rp,δ,k algebraic. We prove the following result concerning the dual (in the sense of the lattice duality induced by the Beauville-Bogomolov form) divisor Dp,δ,k to the class Rp,δ,k. Theorem 0.1. (cf. Theorem 4.1) The divisor Dp,δ,k is a wall divisor if and only if q(Rp,δ,k) < 0. By comparison with [BM2, Yo3], we show that all wall divisors are realized as Dp,δ,k for some integers p, δ and k, up to isometry (in the sense of lattice theory), cf. Proposition 4.6. This is rather striking, as it shows that the birational geometry of S[k] ε can be recovered from classical Brill-Noether theory of curves on the underlying surface, at least when the monodromy group is maximal. We mention that some wall divisors have also been recently constructed by Hassett and Tschinkel [HT3], using a different approach. Under opportune assumptions, we explicitly construct the locus in S[k] ε covered by our rational curves of class Rp,δ,k. When Dp,δ,k is a wall divisor this locus may be described abstractly using only lattice theoretic properties, as in [BM, Yo3] and in the more recent [HT3]. However, our constructions only rely on the definition of our curves of class Rp,δ,k and are thus very concrete. The first type of construction goes as follows. Let M be the component of the moduli space of (Gieseker) L-stable torsion free sheaves on S with Mukai vector v = (2, c1(L), χ+2(ε−1)) containing the Lazarsfeld-Mukai bundle associated with the pushforward to a δ-nodal curve in S of a g1 k+ε on its normalization. As soon as χ := p − δ − k + 3 − 5ε ≥ 2δ + 2, we construct a variety P → M × S[δ] which is generically a projective bundle. The fibre of P over a point ([E], τ ) ∈ M × S[δ] is the projectivization of the space of global sections of E vanishing along τ . We then define a rational map g : P 99K S[k] ε and denote by T the closure of the image of g, which is an irreducible component of the locus covered by curves of class Rp,δ,k. We show that g is birational, thus obtaining the following: Theorem 0.2. (cf. Theorem 6.1) Let (S, L) be a very general primitively polarized K3 or abelian surface of genus p ≥ 2. Let k ≥ 2 and 0 ≤ δ ≤ p − ε be integers such that max{2δ + 2, 4ε} ≤ χ := p − δ − k + 3 − 5ε ≤ δ + k + 1. Then, there is a subscheme T ⊂ S[k] birational to a Pχ−2δ−1-bundle on a holomorphic symplectic ε manifold W of dimension 2(k + 1 + 2δ − χ). Furthermore, the lines contained in any fibre of the rational projection T 99K W have class Rp,δ,k. The resulting uniruled subvarieties are contractible (up to birational equivalence) when the curve Rp,δ,k has negative square. In the case where δ = 0 and Rp,δ,k has the minimal possible Beauville- Bogomolov square, namely, −(k + 3 − 2ε)/2, we use Theorem 0.2 in order to construct a Lagrangian k-plane Pk ⊂ S[k] such that Rp,δ,k is the class of its lines, cf. Example 6.5 and Proposition 6.6. This ε agrees with Bakker's result [Ba, Thm. 3] stating that, in the case ε = 0, a primitive class generating an extremal ray is the line in a Lagrangian k-plane if and only if its square is −(k +3)/2, and suggests that the analogous statement should hold for ε = 1. Note that very few examples of Lagrangian planes are explicitly described in the literature, cf. [Ba, Ex. 8, 9, 10]. Our rational curves have applications to the Chow ring of IHS manifolds, too. In the recent paper [Vo], Voisin stated the following: WALL DIVISORS AND ALGEBRAICALLY COISOTROPIC SUBVARIETIES OF IHS MANIFOLDS 3 Conjecture 0.3. (cf. [Vo, Conj. 0.4]) Let X be a projective IHS manifold of dimension 2k and let Sr(X) be the set of points in X whose orbit under rational equivalence has dimension at least r. Then Sr(X) has dimension 2k − r. The above sets Sr(X) are countable unions of closed algebraic subsets of X and endow the Chow group CH0(X) of 0-cycles with a filtration S• which is conjecturally connected with the Bloch- Beilinson filtration and its splitting predicted by Beauville [Be2]. The question about non-emptiness of Sr(X) is still open and related to the existence problem for algebraically coisotropic subvarieties of X. If X has dimension 2k and σ is its symplectic form, a subvariety Y ⊂ X of codimension r is algebraically coisotropic if there exist a (2k − 2r)-dimensional variety B and a surjective rational map Y 99K B such that σY is the pullback of a 2-form on B. The subvarieties T ⊂ S[k] ε of Theorem 0.2 are algebraically coisotropic by construction and they are components of Sr(S[k] ε ) of dimension 2k − r, with r := χ − 2δ − 1 (cf. Corollary 6.2). Starting from T and then applying the natural rational map S[k+ε] × S[l−k] 99K S[l+ε], one obtains a component of Sr(S[l] ε ) for any l ≥ k. We use this observation in Theorem 6.3 in order to construct components of Sr(S[k] ε ), with k fixed, for several values of r. Our second construction of uniruled subvarieties of S[k] ε is obtained by considering the Severi variety V{L},δ of δ-nodal curves in the continuous system {L} with δ big enough; the assumptions on k+ε. For any integer δ ensure, in particular, that the normalization eC of any curve in V{L},δ has a g1 k′ satisfying suitable conditions, the symmetric product Symk′+ε(eC) is generically a Pr-bundle on Pick′+ε(eC), where r depends on the integers δ and k′. By varying them, we exhibit (2k − r)- dimensional components of Sr(S[k] the following: ε ) for any r, except r = k when ε = 1. More precisely, we prove Theorem 0.4. (cf. Theorem 6.4) Let (S, L) be a general primitively polarized K3 or abelian surface of genus p ≥ 2 and fix an integer k ≥ 2. Then for any integer r such that 1 ≤ r ≤ k − ε, and any integer k′ such that r + ε ≤ k′ ≤ min{k, p + r − ε}, the set Sr(S[k] ε ) has an irreducible component Wr,k′ satisfying the following: (i) dim Wr,k′ = 2k − r; (ii) Wr,k′ is birational to a Pr-bundle and hence algebraically coisotropic; (iii) the class of the lines in the Pr-fibres is L − [2(k′ + ε) − r − 1]rk; (iv) the maximal rational quotient of the desingularization of Wr,k′ has dimension 2(k − r). Point (iv) positively answers, in the case of S[k] [CP, Question 1.2]) concerning existence of subvarieties of an IHS manifold whose maximal rational quo- tients have the minimal possible dimension. ε , a question by Charles and Pacienza (cf. For ε = 0, examples of (2k − r)-dimensional components of Sr(S[k] ε ) for any r were already provided in [Vo, §4.1 Ex. 1 and Lemma 4.3] by considering fibres of the Hilbert-Chow morphism µk : S[k] → Symk(S). However, our components Wr,k′ are not contained in the exceptional locus of µ and thus provide much stronger evidence for Conjecture 0.3. In developing techniques towards proving the above theorems, we obtain some general results on IHS manifolds. First of all, in Proposition 2.13 we provide a criterion to tell whether a deformation of S[k] for some surface S′. This appears to be related to ideas from [Ad] and ε [MW]. Secondly, we prove that wall divisors can be contracted under general assumptions: is isomorphic to S′[k] ε Theorem 0.5. (cf. Theorem 2.5) Let X be a projective IHS manifold and let D be a wall divisor on X. Then one of the following holds: • There exists a curve R dual to D that moves in a divisor and a birational map X 99K Y contracting R. Moreover Y is singular symplectic. • For a general deformation (X ′, D′) of (X, D), there is a birational map X ′ 99K X ′′ with X ′′ IHS and a contraction X ′′ → Y that contracts all curves dual to D′. 4 ANDREAS LEOPOLD KNUTSEN, MARGHERITA LELLI-CHIESA, AND GIOVANNI MONGARDI This result holds in particular for general nonprojective deformations of (X, D), where a proof of the contraction theorem was, as yet, unavailable. The paper is organized as follows. Section 1 contains background material concerning IHS mani- folds and in particular varieties of the form S[k] ε . In Section 2 we recall known results on the birational geometry of IHS manifolds and use them in order to prove Theorem 0.5. We then specialize to de- formations of S[k] and prove that −(k + 3 − 2ε)/2 is a lower bound for the self-intersection of a ε primitive generator of an extremal ray of the Mori cone, cf. Proposition 2.11; the result is new for ε = 1, while it had already appeared in [BHT, Mo1] for ε = 0. Section 3 summarises the results from [CK, KLM] concerning the Brill-Noether theory of nodal curves on symplectic surfaces. Classes Rp,δ,k are computed. Proposition 3.6 proves the existence of a family of rational curves of class Rp,δ,k having the expected dimension and surviving in any small deformation of S[k] that keeps the class algebraic. In Section 4 we prove Theorem 0.1 and exhibit a ε collection of wall divisors that we later show to be essentially "complete" in Proposition 4.6. Section 5 proves several results concerning vector bundle techniques associated with nodal curves, which are essential in the proof of Theorem 0.2. We believe that these results are of independent interest, due to the recent activity in the study of nodal curves on K3 and abelian surfaces. In particular, Proposition 5.3 extends a result by Pareschi [Pa, Lemma 2] to possibly nodal curves on symplectic surfaces; Proposition 5.5 and Lemma 5.6 describe properties of general (semistable) sheaves lying in a specific component M of their moduli space. The main results Theorems 0.2, and 0.4 are finally proved in Section 6. Note. After this paper was completed, a paper by H. Y. Lin [Li] appeared on the arXiv, where the author also constructs components of the locus Sr for generalised Kummer manifolds. Our constructions are different from the Lin's and the spirit of the two papers is quite distant. Acknowledgements We are grateful to C. Ciliberto and K. O'Grady for interesting conversations on this topic. More- over, we thank the Max Planck Institute for mathematics and the Hausdorff Center for Mathematics in Bonn, the University of Bonn and the Universities of Roma La Sapienza, Roma Tor Vergata and Roma Tre, for hosting one or more of the authors at different times enabling this collaboration. The second named author was supported by the Centro di Ricerca Matematica Ennio De Giorgi in Pisa and the third named author by "Firb 2012, Spazi di moduli ed applicazioni". A compact Kähler manifold X is called hyperkähler or irreducible holomorphic symplectic (IHS) if 1. Generalities on IHS manifolds it is simply connected and H 0(Ω2 X ) is generated by a symplectic form. The symplectic form implies the existence of a canonical quadratic form q( ) on H 2(X, Z), called the Beauville-Bogomolov form, and of a constant c, the Fujiki constant, such that for every α ∈ H 2(X, Z) one has: (1) q(α)n = c · α2n, where dim(X) = 2n. We will denote by b( , ) the bilinear form associated with q. This endows H 2(X, Z) with the structure of a lattice of signature (3, b2(X) − 3) and provides an embedding of H2(X, Z) in H 2(X, Q) as the usual lattice embedding L∨ ֒→ L ⊗ Q. Fora any D ∈ H 2(X, Z) denote by div(D) the positive generator of the ideal b(D, H 2(X, Z)); then the elements D/ div(D), with D running among all primitive elements in H 2(X, Z), generate H2(X, Z). The quadratic form and the symplectic form also allow to define a period domain for IHS manifolds, much as in the case of K3 surfaces, as follows. For any lattice L, one defines the period domain ΩL := {ω ∈ P(L ⊗ C) q(ω) = 0, b(ω, ω) > 0}. WALL DIVISORS AND ALGEBRAICALLY COISOTROPIC SUBVARIETIES OF IHS MANIFOLDS 5 Any isometry f : H 2(X, Z) → L is called a marking and there is a natural map, the period map P, sending a marked IHS manifold (X, f ) to P(X, f ) := [f (σX)] ∈ ΩL, where σX is any symplectic form on X. Let ML be the moduli space of deformation equivalent marked IHS manifolds with H 2 isometric to L. The period map P : ML → ΩL is surjective [Hu2, Thm. 8.1] and it is a local isomorphism [Be1, Thm. 5]. There are singular analogues of IHS manifolds, called symplectic varieties. A normal variety Y is symplectic if it has a unique (up to scalars) nondegenerate symplectic form on its smooth locus and a resolution of singularities π : eX → X such that the pullback of this form is everywhere defined, but possibly degenerate. If it is nondegenerate, then eX is IHS and we say that π is a symplectic resolution. Symplectic varieties share many properties with IHS manifolds, especially when they admit a symplectic resolution. In this case it is indeed possible to define a quadratic form on their second cohomology group and the following results hold. covering. Moreover, X has a flat deformation to an IHS manifold. Any smoothing of X is an IHS Theorem 1.1. (Namikawa [Na, Thm. 2.2]) Let π : eX → X be a symplectic resolution of a projective symplectic variety X. Then the Kuranishi spaces Def(X) and Def(eX) are both smooth and of the same dimension. There exists a natural map π∗ : Def(eX) → Def(X) which is a finite manifold obtained as a flat deformation of eX. Theorem 1.2. (Kirchner) Let X be a normal symplectic variety admitting a symplectic resolution of singularities and such that codim(Sing X) ≥ 4. Let Def(X)lt denote the Kuranishi space of locally trivial deformations of X. Then there is a well defined period map P : Def(X)lt → ΩL, where L ≃ H 2(X, Z), having injective tangent map. Proof. Locally trivial deformations are parametrized by a locally closed subset of Def(X). The latter is smooth by Theorem 1.1. After replacing X with a small locally trivial deformation, we can suppose that Def(X)lt is smooth, therefore [Ki, Cor. 3.4.2] applies and first order locally trivial deformations are parametrised by H 1(X − Sing X, Ω1 X) ≃ H 1,1(X). Now [Ki, Thm. 3.4.4] provides the period map as stated above. (cid:3) Remark 1.3. Keep notation as above and let R1, . . . , Ri be the curve classes that span the classes of curves contracted by the resolution of singularities eX → X. The above theorem implies that first order locally trivial deformations of X are parametrised by H 1,1(eX) ∩ hR1, . . . , Rii⊥ ≃ H 1,1(X). Very few examples of IHS manifolds are known. The present paper will focus on the two infinite families of examples introduced by Beauville [Be1], namely, Hilbert schemes of points on K3 surfaces and generalised Kummer manifolds. Let S be a K3 or abelian surface. Throughout the paper we will let (2) ε = εS :=(1 0 if S is abelian, if S is K3. It was proved by Beauville [Be1] that the Hilbert scheme S[k+ε] of 0-dimensional subschemes of S of length k + ε , where k ≥ 2, inherits a symplectic form from S and is smooth. When S is K3, it is simply connected and thus an IHS manifold of dimension 2k. When S is abelian, S[k+1] is not simply connected, but any fibre of the Albanese map Σk : S[k+1] → Alb S[k+1] ≃ S is a 2k- dimensional IHS manifold K k(S), which is called a generalised Kummer manifold. We recall that Σk is the composition of the Hilbert-Chow morphism µk : S[k+ε] → Symk+ε(S) and the summation map + : Symk+ε(S) → S. In order to handle the two families simultaneously, we set (3) S[k] ε :=(K k(S) S[k] if ε = 1 (i.e., S is abelian), if ε = 0 (i.e., S is K3). 6 ANDREAS LEOPOLD KNUTSEN, MARGHERITA LELLI-CHIESA, AND GIOVANNI MONGARDI ε = 2k in both cases, even though S[k] Note that dim S[k] 1 ⊂ S[k+1]. By abuse of notation, in the latter case we will still use the same symbol µk and the same name for the restriction of the Hilbert-Chow morphism to S[k] 1 . There are natural embeddings (4) (5) NS(S) ֒→ Pic(S[k] N1(S) ֒→ N1(S[k] ε ), ε ). The former is given by associating with the class of a prime divisor D in S the divisor (6) {Z ⊂ S[k] ε Supp(Z) ∩ D 6= ∅} and the latter is given by fixing a set of general points {x1, . . . , xk+ε−1} ⊂ S and associating with the class of an effective curve C ⊂ S the class of the curve {Z ⊂ S[k] ε Supp(Z) ∩ C 6= ∅, {x1, . . . , xk+ε−1} ⊂ Supp(Z)}. The exceptional divisor ∆k of the Hilbert-Chow morphism µk has class 2ek and one has an orthogonal decomposition with respect to b( , ): such that b( , ) restricts to the usual cup product on S and q(ek) = −2(k − 1 + 2ε). The above isometry restricts to the embedding (4) on the algebraic part, whence H 2(S[k] ε , Z) ≃ H 2(S, Z) ⊕⊥ Z[ek], (7) Under the embedding H2(S[k] ε , Z) is generated by H 2(S, Z) and rk := ek/2(k − 1 + 2ε). Here rk is the class of a general rational curve lying in the exceptional divisor ∆k of the Hilbert-Chow morphism, that is, rk is the inverse image under µk of a cycle in Symk+ε(S) supported at precisely k − 1 + ε points. Hence, div(ek) = 2(k − 1 + 2ε) and ε ) ≃ NS(S) ⊕ Z[ek]. ε , Q) given by lattice duality, H2(S[k] Pic(S[k] ε , Z) ֒→ H 2(S[k] Any smooth Kähler deformation of S[k] ε type if ε = 0. N1(S[k] ε ) ≃ N1(S) ⊕ Z[rk]. is called a manifold of Kummer type if ε = 1 and of K3[k] (8) (9) Remark 1.4. The manifold S[k] can also be defined by means of moduli spaces of stable sheaves on ε the underlying surface. There is a natural map Coh(S) → H 2∗(S, Z) sending a sheaf F to its Mukai vector v(F) := ch(F)ptd(S) = (rk F, c1(F), χ(F) + (ε − 1) rk F). In order to construct a moduli space of sheaves, one needs also a choice of a polarization L and, for most choices of v (see [Yo1, Thm. 0.1]), a general ample L gives a smooth moduli space M(v) of Gieseker L-semistable torsion free sheaves with Mukai vector v. Moreover, the fibre of M(v) under the Albanese map is deformation equivalent to S[k] ε . If v := (1, 0, 1 − 2ε − k), every element [F] ∈ M(v) can be written as F = H0 ⊗ IZ with H0 ∈ Pic0(S) and [Z] ∈ S[k+ε]. Hence, one has M(v) ≃ S[k] in the K3 case, while in the abelian case K k(S) is the fibre over 0 of the Albanese map of M(v), cf. [Yo2, Thm. 0.1]. For k ≥ 2, we have a canonical Hodge isometry H 2(S[k] ε , Z) ≃ H 2(S, Z) ⊕⊥ Z[ek] ≃ v⊥ ⊂ H 2∗(S, Z) = Λ := U ⊕4 ⊕ E8(−1)⊕2−2ε, such that ek is sent to (1, 0, k − 1 + 2ε) and the second cohomology of S is sent back to itself, cf. [Yo2, Thm. 0.2]. In particular, one has v + ek v − ek (10) ∈ Λ and 2 2(k − 1 + 2ε) ∈ Λ. WALL DIVISORS AND ALGEBRAICALLY COISOTROPIC SUBVARIETIES OF IHS MANIFOLDS 7 2. Birational geometry and wall divisors of IHS manifolds Having trivial canonical bundle, IHS manifolds are minimal in the sense of MMP. Therefore, maps between IHS manifolds are rather rigid, as the following shows: Proposition 2.1. Let X and X ′ be two IHS manifolds and let f : X 99K X ′ be a birational map. Then the following hold: (i) The manifolds X and X ′ are deformation equivalent and H 2(X, Z) ≃ H 2(X ′, Z) as Hodge structures. (ii) The map f has indeterminacy locus of codimension at least 2. (iii) If X is projective, there exists a klt divisor D such that the map f is a sequence of flips obtained by running the minimal model program for the pair (X, D). Proof. Item (i) is the content of [Hu2, Thm. 4.6], and (ii) is proved in [Hu2, Rem. 4.4] and holds true for all manifolds with nef canonical divisor. For (iii), any (sufficiently small) multiple of an effective divisor on a IHS manifold is klt (see [HT2, Rem. 12]). Therefore, if we take an ample divisor A on X ′ and set D = ǫf ∗(A), for ǫ << 1, we have a klt pair (X, D). As A is ample and f is well defined on divisors, D is positive on all curves C such that Locus(R+[C])1 is a divisor. Therefore, by running the MMP for (X, D) we do not encounter any divisorial contraction. As f∗D is ample, (X ′, A) is a minimal model for (X, D). (cid:3) We refer to [LP, Thm. 4.1] for the termination of the log-MMP for IHS manifolds. Being well-defined on divisors, any birational map between two IHS manifolds induces a pullback map between their second cohomology groups. This allows to define a birational invariant called the birational Kähler cone of an IHS manifold X. We recall that the positive cone CX is the connected component containing a Kähler class of the cone of positive classes inside H 1,1(X, R). It contains the Kähler cone KX , which is the cone containing all Kähler classes. The birational Kähler cone BKX is the union ∪f −1KX ′, where f runs through all birational maps between X and any IHS manifold X ′. If X is projective, then the closure of the algebraic part of the birational Kähler cone is just the movable cone, that is, the closure of the cone of divisors whose linear systems have no divisorial base components. We recall that an isomorphism H 2(X, Z) ≃−→ H 2(Y, Z), where X and Y are two IHS manifolds, is called a parallel transport operator if it is induced by the parallel transport in the local system R2π∗Z along a path of smooth deformations π : X → D over a disc D such that X and Y are two fibres. The group of parallel self-operators is called the monodromy group and denoted Mon2(X). Definition 2.2. ([Mo1, Def. 1.2]) Let X be an IHS manifold and let D be a divisor on X. Then D is called a wall divisor if q(D) < 0 and f (D)⊥ ∩ BKX = ∅ for all Hodge isometries f ∈ Mon2(X). The set of wall divisors on X is denoted by WX . The ample cone is one of the connected components of CX − ∪D∈WX D⊥. Wall divisors are closely related to extremal rays of the Mori cone, as was analised independently in [BHT] and [Mo1]. In particular, dual divisors to generators of rational extremal rays of negative square are wall divisors by [Mo1, Lemma 1.4]. Notice that the extremal rays needed to determine the Kähler cone are indeed rational since the part of the Mori cone of curves of negative square is locally a finite rational polyhedron [HT2, Cor. 18]. The analogy runs deeper: Proposition 2.3. Let D be a divisor and let R be the primitive class D/ div(D) ∈ H2(X, Z) ⊂ H 2(X, Q). Then D is a wall divisor if and only if there exists a Hodge isometry f ∈ Mon2(X) such that f (R) generates an extremal ray of the Mori cone on some IHS manifold X ′ birational to X. 1We recall that the locus of V ⊂ N1(X) is the closure of the locus in X covered by curves of class lying in V , that is, Locus(V ) := {x ∈ Γ ⊂ X : [Γ] ∈ V }. 8 ANDREAS LEOPOLD KNUTSEN, MARGHERITA LELLI-CHIESA, AND GIOVANNI MONGARDI Proof. Let D be a wall divisor. As q(D) < 0, we have D⊥ ∩ CX 6= ∅. Therefore, if X is projective, there is a Hodge isometry f ∈ Mon2(X) such that f (D)⊥ ∩ BKX 6= ∅ by [Ma, Thm. 6.18 (2)]. If X is not projective, the same result is a direct consequence of [Hu2, Cor. 5.2 and Rem. 5.4], where the cycle Γ in the mentioned results is of parallel transport and acts as a Hodge isometry on H 2(X, Z). By definition of wall divisor, f (D)⊥ supports a component of the boundary of BKX. Up to taking a different birational model X ′ of X, we can suppose f (D)⊥ ∩ KX 6= ∅. As the ample cone is locally rationally polyhedral by [HT2, Prop. 13], we can also suppose that f (D)⊥ supports a face of this cone (again, if needed, by changing birational model). This implies that R is an extremal ray. The converse is the content of [Mo1, Lemma 1.4] (see also [BHT, Prop. 3]). (cid:3) Remark 2.4. The above result is also implied by [BHT, Cor. 6] and can be used in order to give an equivalent definition of wall divisors, i.e., divisors dual to extremal rays up to the action of parallel transport Hodge isometries. In other words, the MBM classes defined in [AV] are exactly the classes of curves dual to wall divisors. A different characterisation of wall divisors can be given in terms of contractions: Theorem 2.5. Let R be a primitive rational curve on a projective IHS manifold X such that the dual divisor D is a wall divisor. Then one the following cases occurs: (i) Locus(R+[R]) contains a divisor of class a multiple of D. Furthermore, there exists a bira- tional map f : X 99K Y with Y singular symplectic such that f contracts R. (ii) For a general small deformation (Xt, Rt) of (X, R) the locus Locus(R+[Rt]) is not a divisor t and a t along with a birational map ft and there exists an IHS manifold X ′ morphism X ′ t → Yt contracting ft(Rt). : Xt 99K X ′ Proof. Let X ′′ be an IHS manifold deformation of X such that the parallel transport R′′ of R is an effective rational curve generating the algebraic classes of H2(X ′′, Z) (cf. [Mo1, Thm. 1.3] for the existence of such an X ′′). Let D′′ be the dual divisor to R′′. Suppose that Locus(R+[R′′]) has codimension one (thus, the same holds for Locus(R+[R]) by semicontinuity) and let bD′′ be the class of its closure. As we deform back to X, the divisor bD′′ deforms to bD, which is thus effective and is contained in Locus(R+[R]). As D · R < 0, the MMP for the pair (X, D) yields the existence of a birational map f as in item (i). Let us suppose now that Locus(R+[R′′]) has codimension at least two and show that we fall in case (ii). Under this assumption X ′′ contains no effective divisor. Then, by the wall and chamber decomposition of the positive cone given in [Ma, §5], the closure of the birational Kähler cone of X ′′ coincides with its positive cone. On the other hand, as the curve R′′ is effective, the Kähler cone is the intersection of the positive cone with the half space of real (1, 1)-classes intersecting R′′ positively. By the definition of the birational Kähler cone, this yields the existence of an IHS manifold Z ′′ along with a birational map X ′′ 99K Z ′′, the indeterminacy locus of which is Locus(R+[R′′]). In particular, the class −R′′ is effective on Z ′′ as proved in [Hu1, Cor. 2.4]. We now deform X ′′ (hence, also X) to a projective IHS manifold where the class of R′′ is still effective; this is possible as, by [BHT, Prop. 3], all small deformations of X where D stays of type (1, 1) have R or −R effective and projective deformations are dense. In particular, we can choose a projective deformation X ′′′ where the parallel transport of R is effective and extremal; indeed, up to changing birational model, R is an extremal ray on all deformations (X0, R0) belonging to the Zariski open set where CX0 = BKX0. Therefore, the Contraction Theorem yields a contraction X ′′′ → Y ′′′ and the conclusion follows from the next lemma. (cid:3) Lemma 2.6. Let Z be a projective IHS manifold and let R be a curve generating an extremal ray such that Locus(R+[R]) has codimension at least 2. Let Z → Y be the contraction of this extremal ray. Then for all small locally trivial deformations Yt of Y there is a symplectic resolution Zt → Yt contracting exactly Locus(R+[Rt]), where (Zt, Rt) is a small deformation of (Z, R). WALL DIVISORS AND ALGEBRAICALLY COISOTROPIC SUBVARIETIES OF IHS MANIFOLDS 9 Proof. By [Wi, Thm. 1.3], the singular locus of Y has codimension at least four. Let Yt be a locally trivial small deformation of Y . Then Yt has the same Beauville-Bogomolov form of that of Y (and also the same second Betti number) and it has a symplectic resolution Zt, which is a small deformation of Z by Theorem 1.1. Remark 1.3 ensures that the deformation [Rt] of [R] is algebraic. As R is extremal, small deformations [Rt] of its class are represented by curves Rt [BHT, Prop. 3]; the Rigidity Lemma then implies that Rt is contracted by Zt → Yt. By Remark 1.3, b2(Z) = b2(Y ) + 1. Hence, b2(Zt) = b2(Yt) + 1 and the map contracts precisely Locus(R+[Rt]). (cid:3) Remark 2.7. The first item of Theorem 2.5 is slightly stronger than [Ma, Prop. 6.1] as it ensures that exceptional divisors, as defined in [Ma, Def. 5.1], are contractible, up to birational equivalence. This should be regarded as the higher dimensional analogue of the contractability of effective divisors with self-intersection −2 on K3 surfaces. Notice that, when R is reducible, the contraction does not necessarily have relative Picard rank one. The contraction map f : X 99K Y is a composition of flops and divisorial contractions and therefore is only rational. The second item of the proposition cannot be strengthened and in particular it might not hold for (X, R). Indeed, one has to take into account the action of the subgroup Wexc of Mon2 generated by the reflections on reduced and irreducible exceptional divisors. The general deformations in the statement are precisely those manifolds where Wexc is the identity. Note that this set strictly contains the open set of manifolds with an irreducible Hodge structure and it is Zariski open as the set of generators of Wexc is finite up to the monodromy action. Wall divisors on S[k] ε can be determined lattice-theoretically using results of Yoshioka [Yo3] and Bayer and Macrì [BM]. In the following, we use the same notation as in Remark 1.4. Remark 2.8. In [BM] and [Yo3], Bayer, Macrì and Yoshioka determine a decomposition of the space of stability conditions Stab0(S, v) given by walls and chambers. Any stability condition σ in a chamber gives a smooth moduli space M (v, S, σ) of stable objects in Db(S) with Mukai vector v, whereas any condition lying on a wall gives a singular space and conditions on nearby chambers give its symplectic resolution. Moreover, for every σ in a chamber of Stab0(S, v), [BM, Thm 1.2] gives a map from Stab0(S, v) to the positive part of the movable cone BKM (v,S,σ), and every chamber lands in BKM (v,S,σ). By Proposition 2.3, this implies that all walls of Stab0(S, v) are dual to wall divisors and, up to the action of Wexc (defined in Remark 2.7), we obtain all wall divisors of M (v, S, σ) in this way. By Remark 1.4 along with the fact that Mumford's stability lies in Stab0(S, v) for any v, the ordinary moduli spaces of Mumford's stable sheaves with Mukai vector v is obtained as M (v, S, σ) for a σ ∈ Stab0(S, v). In particular, S[k] ε is the Albanese fibre of some M (v, S, σ). Theorem 2.9. Let D be a divisor of S[k] ε with q(D) < 0 and let T ⊂ Λ := H 2∗(S, Z) be the saturated lattice generated by v := (1, 0, 1 − 2ε − k) and D. Then D is a wall divisor if and only if there is an s ∈ T such that (i) 0 ≤ q(s) < b(s, v) ≤ (q(v) + q(s))/2; or, (ii) ε = 0, q(s) = −2 and 0 ≤ b(s, v) ≤ q(v)/2. Proof. Remark 2.8 implies that all wall divisors of S[k] ε correspond to walls in the space Stab0(S, v). For ε = 0 we can thus apply [BM, Thms. 5.7 and 12.1] with a := s and b := v − s; our inequalities are equivalent to imposing that both a and b are in the positive cone of T (cf. [BM, Def. 5.4]), i.e., q(a) ≥ 0 and b(v, a) > 0 and the same for b. For ε = 1 the statement follows from[Yo3, Prop. 1.3]. Indeed, the conditions in [Yo3, Def. 1.2] can be rephrased by asking that a := s and b := v − s are in the positive cone of T as before. The additional condition b(s, v)2 > q(v)q(s) in [Yo3, Prop. 1.3] is equivalent to the requirement that T is indefinite, which is implied by q(D) < 0. (cid:3) Remark 2.10. A lattice T as in the above theorem can contain several elements s satisfying (i) and (ii), and abstractly isometric lattices can even correspond to different kinds of wall divisors, as 10 ANDREAS LEOPOLD KNUTSEN, MARGHERITA LELLI-CHIESA, AND GIOVANNI MONGARDI the following example illustrates (cf. also [HT3, Sec. 4]). Let k − 1 + 2ε = 2rt, where r and t are relatively prime integers. Let S be a symplectic surface and let M be the moduli space of stable sheaves with Mukai vector v := (r, 0, −t). Let Γ ∈ H 1,1(M, Z) be the image of (r, 0, t) under the natural Hodge isometry H 2(M, Z) ≃ v⊥ ⊂ H 2∗(S, Z). The saturated lattice generated by v and Γ is isometric to U and contains no elements s such that q(s) = 0 and b(s, v) = 1, unless either r or t are 1. Note that v+Γ satisfy the conditions of the above theorem, and hence Γ is a wall divisor. The lattice U is also associated with the exceptional divisor ∆k of S[k] ε , but in the saturated lattice generated by v and ek there is an element s such that b(s, v) = 1 and q(s) = 0. However, isometric lattices as in Theorem 2.9 give rise to isometric wall divisors. 2r and v−Γ 2t Theorem 2.9 enables us to extend to manifolds of Kummer type a result obtained by Bayer, Hassett and Tschinkel, and independently by the third author, in the case of manifolds of K3[k] type. Proposition 2.11. Let R be a primitive generator of an extremal ray of the Mori cone of a manifold X deformation of S[k] ε . Then q(R) ≥ −(k + 3 − 2ε)/2. Proof. For ε = 0 this is the content of [Mo1, Cor. 2.7] or [BHT, Prop. 2]. Let ε = 1 and q(R) < 0. Then the dual divisor D to R, namely, R = D/ div(D), is a wall divisor by Proposition 2.3. As wall divisors are invariant under deformation, we can assume X = S[k] for 1 some abelian surface S. Let T, v, s be as in Theorem 2.9. Let a := GCD(q(v), b(s, v)). We have aD = b(s, v)v − q(v)s and div(D) = q(v)/a. Then we have q(D) = (q(v)2q(s) − q(v)b(s, v)2)/a2 ≥ ≥ 4q(v)2q(s) − q(v)3 − q(v)q(s)2 − 2q(v)2q(s) 4a2 where we have used the inequality b(s, v) ≤ (q(v) + q(s))/2. ≥ − q(v)3 4a2 = − (k + 1) div(D)2 2 , (cid:3) The above statement in the K3 case is part of a conjecture by Hassett and Tschinkel [HT1, Conj. 1.2], who predicted that the class R of a primitive 1-cycle in a manifold of K3[k]-type is effective if and only if the inequality in Proposition 2.11 holds. Counterexamples to the if part are known, cf. [BM2, Rem. 10.4] and [CK, Rem. 8.10]. The analogous conjecture for manifolds X of Kummer type was stated only in the four-dimensional case [HT1, Conj. 1.4]. Proposition 2.11 shows that the only if part holds independently of the dimension of X; on the other hand, the if part fails as soon as dim X > 4, as the following example shows. Example 2.12. Let S be an abelian surface with an order four symplectic group automorphism ϕ. Such an automorphism induces an automorphism ϕ of order four on all the generalised Kum- mer manifolds arising from S. There exists a primitive non-effective class F ⊂ NS(S) such that ϕ(F ) = −F and F 2 = −2, cf. 1 ) that is orthogonal to any ϕ-invariant ample class (hence, it is not effective) and has square −2. This shows that the inequality in Proposition 2.11 is not sufficient for the effectivity of a 1-cycle. [Fu, Table 15]. This class gives a 1-cycle class in N1(S[k] We now state a criterion for determining whether a projective manifold of K3[k] or Kummer type for some S. is isomorphic to S[k] ε Proposition 2.13. Let X be a projective manifold of K3[k] or Kummer type. Then X is isomorphic to S[k] k for some ǫ nef divisor D ∈ NS(X). for some S if and only if there is a birational map f : S[k] ε 99K X and f ∗[D] ∈ e⊥ Proof. The only if part is trivial and we prove the converse implication. S[k] ǫ ∩ e⊥ S[k] ǫ ∩ e⊥ We first claim that BK k = K this, the divisor class f ∗[D] ∈ NS(S[k] the pullback under f of a small ample modification of D is ample on S[k] global Torelli Theorem [Ma, Thms. 1.2 and 1.3]. k are nef. Granting ǫ ) lies in the image of (4) and is movable, hence nef. Moreover, ε by the k , that is, all movable divisors on e⊥ ε and thus X ≃ S[k] WALL DIVISORS AND ALGEBRAICALLY COISOTROPIC SUBVARIETIES OF IHS MANIFOLDS 11 It remains to prove the claim. Let E ∈ C be a divisor such that b(E, ek) = 0. In particular, the class [E] lies in the image of the restriction of (4) to the closure of the positive cone CS and we will denote by ES an effective divisor on S representing its preimage. Let us assume that [E] is not nef. Any irreducible curve Γ ⊂ S[k] such that Γ · E < 0 is not contained in ∆k. The image of such a Γ ǫ under the projection to S of the incidence variety S[k] ǫ (11) I := {(P, [Z]) ∈ S × S[k] ǫ P ∈ Supp(Z)} is an effective curve ΓS ⊂ S, whose class is sent to [Γ] by (5). Since ES · ΓS < 0, the divisor ES is not nef. In the abelian case this is impossible and hence [E] is nef and we are done. Let us show that in the K3 case [E] is not movable. Let R ⊂ S be a (−2)-curve such that ES · R < 0 and denote by DR ⊂ S[k] the corresponding uniruled divisor defined as in (6). Then b(E, DR) < 0, whence E is not movable by [Ma, Prop. 5.6]. (cid:3) Remark 2.14. In the above proposition the condition that X is birational to S[k] is equivalent to ε asking that there is a parallel transport Hodge isometry between the two manifolds, cf. [Ma, Thm. 1.3]. If S is K3, there is a topological way of recognizing a parallel transport Hodge isometry, cf. [Ma, Cor. 9.5]. By the computation of the monodromy group in the Kummer case [Mo2, Thm. 2.3], it is highly expected that a similar characterisation holds if S is abelian. We end this section with a result that will be used in the proof of Theorem 0.2. Proposition 2.15. Let X be a holomorphic symplectic manifold, i. e., there is an étale cover eX := Πi∈I Mi → X, where every Mi is either IHS or abelian. For every subset J ⊂ I, denote by FJ the image in X of a general fibre of the projection eX → Πj∈J Mj. Let q : P → X be generically Assume that g : P 99K Y is a rational map to an IHS manifold Y such that: a Pr-bundle. • dim Y = 2r + dim X; • g is well-defined in codimension one; • g is injective on general fibres of q; • for all J, the map g is generically injective when restricted to PFJ ; • the image of g is an irreducible component of the locus covered by the rational curves of class [g(ℓ)], where ℓ is a line in a fibre of q. Then g is finite. Proof. Let T denote the closure of the image of g and h : eT → T be its desingularization. We consider the maximal rationally connected fibration π : eT 99K B of T . We denote by g : P 99K eT the rational map induced by g and assume that a general fibre of g (or, equivalently, of g) has dimension α. By [AV, Thm. 4.4] along with the equality dim Y = 2r + dim X, a general fibre F of π has dimension equal to codimX T = r + α and g−1(F ) has dimension r + 2α. As g is injective on a general fibre of q, the locus q(g−1(F )) is 2α-dimensional. Let σ be a symplectic form on Y . As in [AV, Pf. of Thm. 4.4], one shows that the form h∗(σT ) is degenerate precisely on the fibres of π, which are rationally connected and hence have no 2-forms. g∗(σT ) is well-defined in codimension one, it extends to a 2-form on P that is degenerate along By definition of g, the 2-form g∗(σT ) coincides with eg∗(h∗(σT )) where the latter is defined. Since eg−1(F ). On the other hand, any form on P is the pullback of a form on X and forms on X can be degenerate only along the FJ 's. Therefore, if α > 0, then the closure of q(g−1(F )) coincides with FJ for some J ⊂ I. This contradicts the injectivity of the restriction of g to PFJ . (cid:3) 3. Curves on symplectic surfaces and their pencils For a polarized surface (S, L), we denote by {L} the continuous system of L, that is, the connected component of Hilb(S) containing the linear system L. If S is a K3 surface, then L = {L}. If 12 ANDREAS LEOPOLD KNUTSEN, MARGHERITA LELLI-CHIESA, AND GIOVANNI MONGARDI S is an abelian surface, then {L} is obtained translating curves in L by points of S. We denote by VL,δ(S) (respectively, V{L},δ(S)) the Severi variety of δ-nodal curves in {L} (resp. L), and by {L}1 δ,d) the Brill-Noether locus parametrizing the nodal curves whose normalization carries a g1 d. We recall (2) and the following result. δ,d (resp. L1 Theorem 3.1. Let (S, L) be a general polarized K3 or abelian surface of genus p := pa(L). Let δ and k be integers satisfying 0 ≤ δ ≤ p − 2ε and k + ε ≥ 2. Then the following hold: (i) {L}1 δ,k+ε 6= ∅ if and only if (12) (13) where δ ≥ α(cid:16)p − δ − ε − (k − 1 + 2ε)(α + 1)(cid:17), 2(k − 1 + 2ε)k; α =j p − δ − ε (ii) whenever nonempty, {L}1 δ,k+ε is equidimensional of dimension min{p − δ, 2(k − 1 + ε)} and a g := p − δ such that dim G1 (iii) there is at least one component Yδ,k+ε of {L}1 k+ε(eC) = max{0, ρ(g, 1, k + ε) = 2(k − 1 + ε) − g}; general element in each component is an irreducible curve C with normalization eC of genus δ,k+ε where, for C and eC as in (ii), when k+ε on eC has simple k+1(eC) is reduced. g ≥ 2(k − 1 + ε) (respectively g < 2(k − 1 + ε)), any (resp. a general) g1 ramification and all nodes of C are non-neutral with respect to it. Furthermore, when S is abelian, for general C in this component the Brill-Noether variety G1 Proof. This is [KLM, Thm. 1.6] when S is abelian and [CK, Thm. 0.1], combined with [KLM, Rem. 5.6], when S is K3. (cid:3) Remark 3.2. (i) The condition (12) is equivalent to (14) ρ(p, l, (k + ε)l + δ) + εl(l + 2) ≥ 0 for all integers l ≥ 0. Indeed, the left hand side of (14) attains its minimum for l = α as in (13) and (12) is a rewrite of (14) with l = α. (ii) The condition (12) is also necessary for the existence of a curve in {L} with partial normalization of arithmetic genus g := p − δ carrying a g1 k+ε. This follows from [KLM, Thm. 5.9 and Rem. 5.11] in the abelian case and [CK, Thm. 3.1] in the K3 case, by remarking that the proofs go through replacing the normalization of the curve with a partial normalization, as remarked in [CK, Rem. 3.2(b)]. Let g be a linear series of type gr where A is a line bundle of degree k + ε and V ⊆ H 0(A) is an (r + 1)-dimensional subspace. If g is base point free, we have a natural rational map k+ε on the normalization eC of a curve C ⊂ S, that is, g = (A, V ), (15) ιg : Pr := P(V ) 99K S[k+ε] obtained from the composition P(V ) ⊆ A ⊂ Symk+ε(eC) → Symk+ε(C) ⊂ Symk+ε(S), whose image does not lie in the exceptional locus ∆k of the Hilbert-Chow morphism. Thus, g defines a rational r-fold inside the Hilbert scheme S[k+ε]. In particular, when r = 1, we obtain a rational curve. When S is an abelian surface, the Albanese map Σk restricted to the image of ιg is constant, because otherwise we would get a rational curve in S. Therefore, up to translating the curve C, we may assume that (15) lands into the generalised Kummer variety K k(S). Let us now specialise to ε be the rational curve image of ιg, recalling (3). (The same construction can be performed for any linear series on a partial normalization of C.) the case r = 1, denote by ν : eC → C the normalization, and let RC,ν∗g ⊂ S[k] WALL DIVISORS AND ALGEBRAICALLY COISOTROPIC SUBVARIETIES OF IHS MANIFOLDS 13 Lemma 3.3. Let C ∈ {L}1 g1 k+ε with simple ramification and such that all nodes of C are non-neutral with respect to it. Then the class of the rational curve RC,ν∗g in H2(S[k] δ,k+ε be a curve whose normalization possesses a linear series g of type ε , Z) with respect to the decomposition (8) is (16) Rp,δ,k := L − (p − δ + k − 1 + ε)rk and its dual divisor class is (17) Dp,δ,k := L − (p − δ + k − 1 + ε) 2(k − 1 + 2ε) ek, Proof. In the K3 case, this is [CK, Lemma 2.1]. The proof in the abelian case is similar. (cid:3) In particular, one has (with α as in (13)): (18) q(Rp,δ,k) = 2(p − 1) − (p − δ + k − 1 + ε)2 2(k − 1 + 2ε) . Observe that, for fixed values of k and p, the minimum in (18), as well as the maximal "slope" p − δ + k − 1 + ε of the class Rp,δ,k, is reached for a curve with the minimal number of nodes. Remark 3.4. Under the same hypotheses as in Lemma 3.3, one may rewrite (16) as q(Rp,δ,k) = 2(cid:16)ρ + εα(α + 2) + ε − 1(cid:17) − β2 2(k − 1 + 2ε) , with ρ := ρ(p, α, (k + ε)α + δ) and β := (2α + 1)(k − 1 + 2ε) − p + δ + ε. In particular, we have −(k − 1 + 2ε) < β ≤ k − 1 + 2ε, and (12), or equivalently (14) with l = α, says that ρ + εα(α + 2) ≥ 0. From these inequalities one reobtains the bound from Proposition 2.11: q(Rp,δ,k) ≥ − k + 3 − 2ε 2 , with equality if and only if p = α(α + 1)(k − 1 + 2ε) + ε and δ = α(α − 1)(k − 1 + 2ε) (see also [CK, Cor. 3.4]). Proposition 2.11 yields the following extension of [CK, Cor. 8.6] to Kummer manifolds. Corollary 3.5. Assume that NS(S) ≃ Z[L]. Let n ∈ Z>0 and set p := n(n + 1)(k − 1 + 2ε) + ε and δ := n(n − 1)(k − 1 + 2ε). Then the rational curves in S[k] ε obtained from the component Yδ,k of Theorem 3.1 generate extremal rays of S[k] ε . We also have: Proposition 3.6. The rational curves in S[k] ε Noether variety G1 the normalization of C move in a family of rational curves of dimension precisely 2k − 2. obtained from any component of the relative Brill- k+ε on δ,k+ε) parametrizing pairs (C, g) such that C ∈ {L}1 k({L}1 δ,k+ε and g is a g1 Any small deformation Xt of X0 = S[k] ε keeping the class of the rational curves algebraic contains a (2k − 2)-dimensional family of rational curves that are deformations of the rational curves in S[k] ε . Proof. Any irreducible family of rational curves in S[k] containing our family yields, by the incidence ε (11), a family of pairs (C, g) with C ∈ {L} and g a linear series of type g1 k+ε on the normalization of C. By [KLM, Thm. 5.3], the rational curves will therefore move in a family of dimension precisely 2k − 2, which is the expected dimension of any family of rational curves on a (2k)-dimensional IHS manifold [Ra, Cor. 5.1]. Hence, as a consequence of [Ra, Cor. 3.2-3.3], the rational curves will deform to any Xt as in the statement, cf. [BHT, Pf. of Prop. 3]. (cid:3) 14 ANDREAS LEOPOLD KNUTSEN, MARGHERITA LELLI-CHIESA, AND GIOVANNI MONGARDI 4. Examples of wall divisors Let (S, L) be a general abelian or K3 surface, and fix integers p, k and δ satisfying (12). Let Rp,δ,k be as in (16) and denote by Dp,δ,k its dual (class) divisor. Theorem 4.1. The divisor Dp,δ,k is a wall divisor if and only if q(Rp,δ,k) < 0. Proof. By Proposition 3.6, the family of rational curves with class Rp,δ,k has a component of dimen- sion 2k − 2 and deforms in all small deformations Xt of S[k] ε where the class Rp,δ,k remains algebraic. Let (Xt, Rt) be a very general such deformation. The class Rt spans N1(Xt), hence it is extremal. As it has negative square, its dual is a wall divisor. Since wall divisors are invariant under deformation, Dp,δ,k is a wall divisor on S[k] (cid:3) ε , too. Remark 4.2. If Dp,δ,k is a wall divisor, we can recover the lattice T associated with it in Theorem 2.9. Set a := GCD(2k − 2 + 4ε, g + k − 1 + ε), ab := g + k − 1 + ε and ac := 2k − 2 + 4ε. The saturation of the lattice generated by v and Dp,δ,k is T := hv, wi, where w = b c (v − ek) + L − v. Note that q(w) = 2δ − 2 + 2ε and b(w, v) = g − k + 1 − 3ε. The element w does not necessarily satisfy the inequalities (i) or (ii) in Theorem 2.9 for s. However, this occurs in some special cases, e.g., in the examples below. 2 k − 1 + 2ε k − 1 + 2ε and the lattice T associated Example 4.3. Let p = 2k −2+5ε and δ = 0. Then q(Rp,δ,k) = − k+3−2ε 2k − 2 + 4ε (cid:19), cf. Remark 4.2. with Rp,δ,k is isometric to (cid:18) −2 + 2ε Example 4.4. Let p = 2k − 2 + 5ε − a, a ≤ k − 1 + 2ε, and δ = 0. Then q(Rp,δ,k) < 0 and the lattice T associated with Rp,δ,k is isometric to (cid:18) Example 4.5. Let p = 2k − 2 + 5ε and 0 ≤ δ ≤ k−1+2ε associated with Rp,δ,k is isometric to (cid:18) 2δ − 2 + 2ε Proposition 4.6. Let k ≥ 2 be an integer and set ε = 0 (respectively, ε = 1). Let v := (1, 0, 1−2ε−k) and let s ∈ Λ = U ⊕4 ⊕ E8(−1)⊕2−2ε be an element satisfying the inequalities (i) or (ii) in Theorem 2.9. Let T = hv, si. Then there exists a primitively polarized K3 (resp. abelian) surface (S, L) of genus p and an integer 0 ≤ δ ≤ p − 2ε such that p, δ, k satisfy (12) and the following hold: 2k − 2 + 4ε (cid:19), cf. Remark 4.2. 2k − 2 + 4ε (cid:19), cf. Remark 4.2. . Then q(Rp,δ,k) < 0 and the lattice T k − 1 + 2ε − a k − 1 + 2ε − a k − 1 + 2ε k − 1 + 2ε −2 + 2ε 2 (a) the divisor Dp,δ,k is a wall divisor; (b) the saturation of the lattice generated by v and Dp,δ,k in Λ is isometric to T . δ,k+ε is non-empty, then {L}1 Proof. By Theorem 3.1(i), as soon as {L}1 δ+1,k+ε is non-empty, too. If 2k − 2 + 4ε (cid:19), the saturation of the lattice generated by Dp,δ,k and v is isometric to (cid:18) 2δ − 2 + 2ε 2k − 2 + 4ε (cid:19). the saturation of the lattice generated by Dp,δ+1,k and v is isometric to (cid:18) 2δ + 2ε Analogously, if (S, L) has genus p and {L}1 δ,k+ε is non-empty for every primitively polarized (S′, L′) of genus p − 1. If the corresponding lattice in the genus p case is 2k − 2 + 4ε (cid:19). These (cid:18) 2δ − 2 + 2ε remarks along Example 4.3 give us all possible isometry classes of lattices T as in the statement. (cid:3) 2k − 2 + 4ε (cid:19), the lattice in the genus (p − 1) case is (cid:18) 2δ − 2 + 2ε δ,k+ε is non-empty, then {L′}1 b − 1 b − 1 b − 1 b − 1 b b b b Remark 4.7. As explained in Remark 2.10, the above proposition does not give all wall divisors up to the monodromy action. However, when k − 1 + 2ε is a prime power, we have that T determines and is determined by the monodromy orbit of D as all isometries of H 2(S[k] ε ) can be extended to isometries of Λ fixing v. Hence the above proposition gives a full list of wall divisors up to monodromy in these cases. WALL DIVISORS AND ALGEBRAICALLY COISOTROPIC SUBVARIETIES OF IHS MANIFOLDS 15 5. Lazarsfeld-Mukai bundles associated with nodal curves complete g1 k+ε, that is, a globally generated line bundle A of degree k + ε such that h0(A) = 2. We Let C be a nodal curve on an abelian or K3 surface S such that its normalization eC possesses a denote by ν : eC → C the normalization map, and by N the 0-dimensional subscheme of the nodes Let f : eS → S be the blow up of S at N , so that we have a commutative diagram of C. By standard facts, ν∗A is a torsion free sheaf of rank one on C that fails to be locally free precisely at N . (19) ❃ ❃ eC ⊂ ϕ ❃ ❃ ❃ ν eS f ❃ ❃ ❃ C ⊂ S We denote by E eC,A and EC,ν∗A the so-called Lazarsfeld-Mukai bundles associated with the line bundle A on eC and the torsion free sheaf ν∗A on C, respectively; the duals of these bundles are the kernels of the evaluation maps regarding A and ν∗A as torsion sheaves on the surfaces, that is, we have the following short exact sequences: (20) and (21) 0 / E ∨ eC,A / H 0(eC, A) ⊗ O eS ev eS,A / A / 0, 0 / E ∨ C,ν∗A / H 0(C, ν∗A) ⊗ OS evS,ν∗A / ν∗A. The right arrow in (21) might be non-surjective, as ν∗A is not necessarily globally generated (cf. Lemma 5.1). Pushing forward (20) to eS and using the isomorphisms H 0(eC, A) ≃ H 0(C, ν∗A) and f∗O eS ≃ OS, one shows that (22) E ∨ C,ν∗A ≃ f∗(E ∨ ). eC,A The following result establishes when (21) is exact on the right. Proof. We can assume that C has only one node P , since the general case is analogous using partial a complete, base point free pencil on C. Then the sheaf ν∗A is globally generated except precisely at the nodes of C that are neutral with respect to A. Lemma 5.1. Let C be a nodal curve and denote by ν : eC → C the normalization map. Let A be normalizations. Let φA : eC → P1 be the morphsm defined by A. Assume that P is a neutral node with respect to A. Then φA factors through a morphism φ : C → P1 having the same degree as φA. Having set A′ := φ∗OP1(1), one has ν∗A′ ≃ A and ν∗A ≃ ν∗ν∗A′ ≃ A′ ⊗ ν∗OC , hence ν∗A sits in the following short exact sequence: 0 −→ A′ −→ ν∗A −→ OP → 0. Since h0(C, ν∗A) = h0(eC, A) = h0(C, A′) = 2, the sheaf ν∗A cannot be globally generated. Conversely, assume that ν∗A is not globally generated at P , that is, the evaluation map ev : H 0(ν∗A) ⊗ OC → ν∗A is not surjective. Since A is globally generated and ν is a finite map, we have a surjection Using (32), one can easily show that the cokernel A1 of ev sits in a short exact sequence H 0(ν∗A) ⊗ ν∗O eC ≃ H 0(A) ⊗ ν∗O eC ։ ν∗A. 0 −→ A1 −→ ν∗A −→ OP −→ 0.      / / / / / / / 16 ANDREAS LEOPOLD KNUTSEN, MARGHERITA LELLI-CHIESA, AND GIOVANNI MONGARDI In particular, A1 is a line bundle and ν∗A1 = A. Hence, the morphism φA factors through a morphism φA1 : C → P1, which means that P is a neutral node with respect to A. (cid:3) The above lemma implies that if C is a general curve of the nice component Yδ,k+ε in Theorem 3.1 and A is a general g1 is exact on the right. By dualizing it, we obtain: k+ε on eC, then ν∗A is globally generated and the short exact sequence (21) (23) 0 / H 0(C, ν∗A)∨ ⊗ OS / EC,ν∗A / ext1(ν∗A, OS ) / 0, where ext1(ν∗A, OS ) is a rank one torsion free sheaf on C. This defines a subspace V ≃ H 0(C, ν∗A)∨ ∈ G(2, H 0(S, EC,ν∗A)) such that the evaluation map V ⊗ OS → EC,ν∗A is injective, drops rank along C and has rank 0 exactly at the nodes of C. As a consequence, every section in V vanishes along a 0-dimensional subscheme of length k + ε + δ always containing the subscheme N of the δ nodes of C. We want to understand whether the pair (C, ν∗A) univocally determines the subspace V ; this (ν∗A, OS)). To achieve this goal we is equivalent to computing the dimension of Hom(EC,ν∗A, ext1 need some technical results. OS For any torsion free rank one sheaf A on C we denote by AD the dual sheaf homOC (A, OC ). We prove the following: Lemma 5.2. Let F be any rank one torsion free sheaf on a curve C ⊂ S, where S is a K3 or abelian surface. Then one has: (24) ωC ⊗ F D ≃ ext1 OS (F, OS ). Assume furthermore that F is the cokernel of an injective map V ⊗ OS → E, where E is a rank two bundle on S and V ∈ G(2, H 0(S, E)). If Z is a zero-dimensional scheme that is the vanishing locus of a s ∈ V , one obtains the isomorphisms: (25) and (26) F ≃ ωC ⊗ JZ/C ext1 OS (F, OS ) ≃ homOC (JZ/C, OC ). Proof. Applying homOS (F, −) to the short exact sequence (27) we obtain 0 / JC/S α / OS / OC / 0, 0 = homOS (F, OS ) / homOS (F, OC ) / ext1 OS (F, JC/S ) α′ / ext1 OS (F, OS ). Since α and α′ are given by multiplication with the local equation of C and F is supported precisely at C, the map α′ is zero. Hence, we get homOC (F, OC ) ≃ homOS (F, OC ) ≃ ext1 and (24) is obtained by tensoring with OS (C). OS (F, JC/S ) ≃ ext1 OS (F, OS (−C)) ≃ ext1 OS (F, OS )⊗OS(−C), / / / / / / / / / / / WALL DIVISORS AND ALGEBRAICALLY COISOTROPIC SUBVARIETIES OF IHS MANIFOLDS 17 Concerning the second part of the statement, we set L := det E and consider the following com- mutative diagram: (28) 0 OS i L ⊗ JZ/S / F 0 0 0 0 / OS s / V ⊗ OS E E ≃ 0 0 OS 0 The short exact sequence of ideals 0 / JC/S / JZ/S / JZ/C / 0 yields the isomorphism JZ/C ≃ JZ/S ⊗ OC . Hence, (25) is obtained by restricting the vertical exact sequence on the right in (28) to C and using that i is given by multiplication with the local equation of C. Combining (25) and (24), we obtain (26). (cid:3) We now extend [Pa, Lemma 2] to possibly nodal curves on K3 or abelian surfaces (cf. also [LC2, Pf. of Prop. 3.2] concerning other types of irregular surfaces); we are confident that this result of independent interest. Proposition 5.3. Let C be a nodal curve on a K3 or abelian surface S and denote by ν : eC → C the normalization map. Let A be a globally generated line bundle on eC satisfying h0(eC, A) = 2. Then there is a natural short exact sequence 0 / O eC / E ∨ eC,A ⊗ ω eC ⊗ A∨ / ω eC ⊗ (A∨)⊗2 / 0 (29) whose coboundary map Q : H 0(eC, ω eC ⊗ (A∨)⊗2) → H 1(eC, O eC ) coincides, up to multiplication by a nonzero scalar factor, with the composition of the Gaussian map and the dual of the Kodaira-Spencer map κ : H 1(C, ν∗O eC)∨ ≃ T[C]V{L},δ µ1,A : ker µ0,A → H 0(eC, ω⊗2 eC ) / T[ eC]Mpg(C) ≃ H 0(eC, ω⊗2 eC )∨, via the canonical isomorphism H 1(C, ν∗O eC) ≃ H 1(eC, O eC ). Proof. The exact sequence (29) is obtained as [Pa, (4)] since ωS ≃ OS. We denote by N ′ morphism ϕ in (19). We recall that N ′ [Ta, p. 111]. In particular, we have: C/S the equisingular normal sheaf of C in S, and by Nϕ the normal sheaf to the C/S ≃ ν∗Nϕ ≃ ωC ⊗ (ν∗O eC )D by, e.g., [Se, Lemma 3.4.15] and (30) T[C]V{L},δ ≃ H 0(C, N ′ C/S) ≃ H 0(eC, Nϕ) ≃ H 1(C, ν∗O eC)∨.       / / /   / /   / /   / / /   / / /     / / / / / / / / / 18 ANDREAS LEOPOLD KNUTSEN, MARGHERITA LELLI-CHIESA, AND GIOVANNI MONGARDI The Kodaira-Spencer map κ is the first coboundary map of the normal sequence (31) 0 / T eC / ϕ∗TS / Nϕ ≃ ω eC / 0, where the isomorphism on the right again follows from the triviality of ωS. Applying ν∗ and again using the isomorphism ν∗ω eC ≃ ωC ⊗(ν∗O eC )D, we obtain the exact sequence 0 / ν∗(ω∨ eC ) / TS ⊗ ν∗O eC / ωC ⊗ (ν∗O eC)D / 0, and its dual 0 / ω∨ C ⊗ (ν∗O eC ) / ΩS ⊗ (ν∗O eC)D / (ν∗(ω∨ eC ))D / 0. The right exactness of the latter is due to the fact that ext1 be verified using the standard exact sequence OC ((ν∗O eC)D, OC ) = 0, which can easily (32) 0 −→ OC −→ ν∗O eC −→ ON −→ 0. Tensoring with ωC, we obtain (33) 0 / ν∗O eC / ΩS ⊗ ωC ⊗ (ν∗O eC)D / ωC ⊗ (ν∗(ν∨ eC ))D ≃ ν∗(ω⊗2 eC ) / 0, where the last isomorphism follows from [BP, Lemma 4.6]. By construction, the first coboundary map of (33) is κ∨. We now follow [Pa, Pf. of Lemma 1]. Tensoring the derivation operator d : O eC → ω eC with the evaluation map ev eC,A : H 0(A)⊗ O eC → A, we obtain a map H 0(A)⊗ O eC → ω eC ⊗ A, whose restriction to ker ev eC,A ≃ A∨ is O eC-linear. Tensoring the restricted map with ω eC ⊗ A∨, we obtain a map of O eC -modules (34) s : ω eC ⊗ (A∨)⊗2 / ω⊗2 eC , whose associated map at the global section level is the Gaussian map µ1,A (remember the standard isomorphism ker µ0,A ≃ H 0(eC, ω eC ⊗ (A∨)⊗2)). Similarly, tensoring the pullback under f of the derivation operator d : OS → ΩS with the evaluation map ev eC,A : H 0(A) ⊗ O eS → A, we obtain a map H 0(A) ⊗ O eS → f ∗ΩS ⊗ A, whose is O eS-linear. Tensoring the restricted map with ω eC ⊗ A∨, we obtain a map of restriction to E ∨ OC -modules eC,A (35) t : E ∨ eC,A ⊗ ω eC ⊗ A∨ / f ∗ΩS ⊗ ω eC. The sequences and maps (29), (33), (34) and (35) combine into: 0 0 / ν∗O eC / ν∗(E ∨ eC,A ⊗ ω eC ⊗ A∨) ν∗(ω eC ⊗ (A∨)⊗2) ν∗t / ν∗O eC / ΩS ⊗ ωC ⊗ (ν∗O eC)D ν∗s / ν∗(ω⊗2 eC ) 0 / 0 and the result follows. We are now ready to prove the following: (cid:3) Proposition 5.4. If C is a general element of the component Yδ,k+ε in Theorem 3.1 and A is a general g1 k+ε on eC, then dim Hom(EC,ν∗A, ext1 OS (ν∗A, OS )) = 1. / / / / / / / / / / / / / / / / / / / / / /   / /   / / / / WALL DIVISORS AND ALGEBRAICALLY COISOTROPIC SUBVARIETIES OF IHS MANIFOLDS 19 C,ν∗A ⊗ ωC ⊗ (ν∗A)D) and, by (23) Proof. By (24), we have Hom(EC,ν∗A, ext1(ν∗A, OS)) ≃ H 0(E ∨ tensored with E ∨ C,ν∗A ⊗ EC,ν∗A) ≃ C, where the isomorphism follows from the fact that EC,ν∗A is simple when (S, L) is general as in Theorem 3.1 (this is standard, cf., e.g., [Pa, CK, KLM], and follows from the fact that {L} does not contain reducible or nonreduced members). C,ν∗A, the latter contains H 0(E ∨ If S is K3 the result is well-known and due to the fact that h1(OS ) = 0 yields h1(E ∨ C,ν∗A) = 0 by (21). It remains to treat the case where S is abelian. We have ωC ⊗ (ν∗A)D ≃ ν∗(ω eC ⊗ A∨) by [BP, Lemma 4.6]. Hence, by (22), there is a natural morphism E ∨ C,ν∗A ⊗ ωC ⊗ (ν∗A)D ≃ f∗(E ∨ eC,A ) ⊗ ν∗(ω eC ⊗ A∨) / ν∗(E ∨ eC,A ⊗ ω eC ⊗ A∨), which is injective as the left hand side is torsion free on C and the map is an isomorphism outside of N . In particular, we get an inclusion H 0(C, E ∨ C,ν∗A ⊗ ωC ⊗ (ν∗A)D) ⊆ H 0(C, ν∗(E ∨ eC,A ⊗ ω eC ⊗ A∨)) ≃ H 0(eC, E ∨ eC,A ⊗ ω eC ⊗ A∨), and thus it is enough to prove the injectivity of the first coboundary map Q of (29). automatically injective. If ρ(p − δ, 1, k + 1) < 0, then dim ker µ0,A = −ρ(p − δ, 1, k + 1) because, by Theorem 3.1(iii), A If ρ(p − δ, 1, k + 1) ≥ 0, then Theorem 3.1 yields ker µ0,A ≃ H 0(eC, ω eC ⊗ (A∨)⊗2) = 0 and Q is k+1(eC). By Proposition 5.3, we need to show that κ∨ ◦µ1,A defines an isolated and reduced point of G1 is injective, or equivalently, µ∨ and its dual sits in the following short exact sequence: 1,A ◦ κ is surjective. Since A is a pencil, then the map µ1,A is injective 0 / T[ eC]M1 g,k+1 / T[ eC]Mg µ∨ 1,A / NM1 g,k+1/Mg [ eC] / 0, cf. [ACG, pp. 807 -- 824]. By Theorem 3.1, Yδ,k+1 has codimension −ρ(p − δ, 1, k + 1) in the Severi variety V{L},δ(S), hence the image of the natural map ψ : V{L},δ(S) → Mg is transversal to M1 around [C]. This forces µ∨ g,k+1 (cid:3) 1,A ◦ κ to be surjective. We now fix notation that will be used in the construction of a component of the locus in S[k] ε covered by rational curves of class Rp,δ,k. Let C and A be as in Proposition 5.4. Since ν∗A is globally generated in this case (by Theorem 3.1 and Lemma 5.1), the Mukai vector of the Lazarsfeld-Mukai bundle EC,ν∗A is (36) vp,δ,k := v(EC,ν∗A) = (2, c1(L), χ + 2(ε − 1)), with χ := χ(EC,ν∗A) = p − δ − k + 3 − 5ε. If Pic(S) ≃ Z[L], which holds off a countable union of proper closed subvarieties of the moduli space of pairs (S, L), then EC,ν∗A is stable with respect to the polarization L (as in, e.g., [KLM, Proposition A.2]). Let M be the component of the moduli space of Gieseker L-semistable torsion free sheaves on S with Mukai vector vp,δ,k as in (36) that contains EC,ν∗A. We recall that M is a projective symplectic manifold admitting a symplectic resolution of dimension: (37) with χ as in (36). dim M = 2p − 4χ + (1 − ε)8, Every torsion free sheaf [E] ∈ M satisfies h2(E) = 0 because of Serre duality and inequality µL(E) > 0. Furthermore, as soon as h0(E) ≥ 2, then for all V ∈ G(2, H 0(E)) the evaluation map ev : V ⊗ OS → E is injective. Indeed, if this were not the case, its kernel would be isomorphic to OS (−D) for an effective divisor D and we would find a short exact sequence: 0 −→ OS(D) −→ E −→ det E(−D) ⊗ Iξ −→ 0, / / / / / 20 ANDREAS LEOPOLD KNUTSEN, MARGHERITA LELLI-CHIESA, AND GIOVANNI MONGARDI where ξ ⊂ S is a 0-dimensional subscheme. As det E is indecomposable and h2(E) = 0, then D = 0 and this contradicts the fact that V is generated by 2 linearly independent sections of E. The following two results determine properties of M that will play a fundamental role in our ε . They can also be seen as applications of Proposition construction of uniruled subvarieties of S[k] 5.4 to moduli spaces of stable sheaves on S and are therefore interesting in themselves. Proposition 5.5. Let p, δ, k be integers such that (12) is satisfied and let v and χ be as in (36). Then there is a (nonempty) irreducible component M of the moduli space of L-stable sheaves on S with Mukai vector vp,δ,k such that the following properties are satisfied: (i) If χ ≥ 2δ + 2, then h1(E) = h1(E ⊗ Iτ ) = 0 for a general pair ([E], τ ) ∈ M × S[δ]. (ii) If χ < 2δ + 2, the locus X := {([E], τ ) ∈ M × S[δ] h0(E ⊗ Iτ ) ≥ 2} is nonempty, with an irreducible component X0 whose general point satisfies h0(E ⊗ Iτ ) = 2 and h1(E ⊗ Iτ ) = 2 + 2δ − χ. Furthermore, X0 is birational to a component of the relative Brill-Noether variety G1 k+ε(cid:16)V k+ε {L},δ(cid:17) := {([C], g) [C] ∈ V k+ε having the expected dimension 2(k − 1 + ε). {L},δ, g ∈ G1 k+ε(eC), with eC the normalization of C}, Proof. Let M contain EC,ν∗A with C and A as in Proposition 5.4 and let X ⊂ M × S[δ] parametrize pairs ([E], τ ) such that h0(E ⊗ Iτ ) ≥ 2. Trivially, X coincides with M × S[δ] as soon as χ ≥ 2δ + 2. If instead χ < 2δ + 2, then X is a closed subscheme which can be defined using fitting ideals and its expected codimension is 2(2δ − χ + 2), whence its expected dimension is 2(k − 1 + ε). Furthermore, X is nonempty because ([EC,ν∗A], N ) lies in it. Let G be the family of triples ([E], τ, V ) with ([E], τ ) ∈ X and V ∈ G(2, H 0(E ⊗ Iτ )), and let p : G → X × S[δ] be the natural projection. Existence of G follows from existence of a moduli space of OS-stable coherent systems; indeed, since M parametrizes torsion free sheaves of rank 2 and indecomposable first Chern class, then any pair ([E], V ) with [E] ∈ M and V ∈ G(2, H 0(E)) is automatically an OS-stable coherent system (proceed as in [LC1, §3]). Since ([EC,ν∗A], N, H 0(C, ν∗A)∨) lies in G, then for a general element ([E], τ, V ) ∈ G the evaluation map ev : V ⊗ OS → E is injective and drops rank along a δ-nodal curve Γ, which is singular precisely at the support of τ . Furthermore, the cokernel B of ev is torsion free of rank 1 on Γ and is not (38) locally free exactly along Sing(Γ). We set B1 := ext1(B, OS). If η :eΓ → Γ is the normalization map, we can write B1 = η∗A1 for some A1 ∈ Pick+ε(eΓ). From the short exact sequence k+ε on eΓ. We we get that H 0(A1)∨ ≃ H 0(B1)∨ ≃ H 1(B) ։ V and hence g = (A1, V ∨) defines a g1 thus obtain a rational map h : G 99K G1 that (C, ν∗A) lies in a component Z of the image of h of dimension: {L},δ). Our hypotheses, along with Theorem 3.1, ensure 0 −→ V ⊗ OS −→ E −→ B −→ 0, k+ε(V k+ε (39) dim Z = dim V k+ε {L},δ + max{0, ρ(p − δ, 1, k + ε)} = 2(k − 1 + ε). On the other hand, Proposition 5.4 implies that h is generically injective and hence, for a general ([E], τ ) ∈ X, we have: (40) dim Z = dim X + 2(h0(E ⊗ Iτ ) − 2). If χ ≥ 2δ + 2, then X = M × S[δ] and (37), (39) and (40)yield: 2(k − 1 + ε) = dim Z ≥ dim M + dim S[δ] + 2(χ(E ⊗ Iτ ) − 2) = 2(k − 1 + ε); (41) thus, equality holds and a general ([E], τ ) ∈ M × S[δ] satisfies h1(E) = h1(E ⊗ Iτ ) = 0. WALL DIVISORS AND ALGEBRAICALLY COISOTROPIC SUBVARIETIES OF IHS MANIFOLDS 21 If instead χ < 2δ + 2, then expdim X = 2(k − 1 + ε) and, by (39) and (40), one obtains dim X = (cid:3) expdimX and h0(E ⊗ Iτ ) = 2 for a general ([E], τ ) ∈ X. Lemma 5.6. Under the same hypotheses as in Proposition 5.5, assume moreover that S is abelian and χ ≥ 4. Then a general [E] ∈ M satisfies h1(E ⊗ L0) = 0 for all L0 ∈ Pic0(S). Proof. It is enough to show that the locus F := {[E] ∈ M h1(E) 6= 0} has codimension greater than two in M. We perform a parameter count as in [LC3, Prop. 4.2 and §7]. Let G1 be the parameter space of extensions (42) 0 −→ OS −→ E ′ β −→ E −→ 0, with [E] ∈ F. Since Hom(E, OS ) = 0 for all E ∈ M, the fibre of the natural map π1 : G1 → F over [E] is isomorphic to P(H 1(E)). A sheaf E ′ as in (42) is stable; let v′ := vp,δ,k + (1, 0, 0) be its Mukai vector and denote by π2 : G1 → M(v′) the natural projection mapping (42) to [E ′]. The fibre of π2 over [E ′] is the Quot-scheme QuotS(E ′, P ), where P is the Hilbert polynomial of E. The following upper bound for the dimension of QuotS(E ′, P ) at [β : E ′ → E] is well-known: dim[β] QuotS(E ′, P ) ≤ dim Hom(OS, E) = h0(E). It follows that: dim M − 2χ = dim M(v′) ≥ dim Imπ2 ≥ dim F − χ − 1, and codimM F ≥ χ − 1 > 2 as soon as χ ≥ 4. (cid:3) The next lemma concerns every component of any moduli space of rank-2 Gieseker semistable torsion free sheaves on S. Lemma 5.7. Let S be an abelian or K3 surface and let M be a component of the moduli space of rank-2 Gieseker semistable torsion free sheaves on S with Mukai vector v = (2, c1, χ + 2(ε − 1)). For every [E] ∈ M, denote by S(E) the cokernel of the injection E ֒→ E ∨∨ and by lE its length. Then every irreducible component of the locus (43) Mq := {[E] ∈ M lE = q} has codimension at least q in M. Proof. Assume that Mq is non-empty. One defines a map α : Mq → M(vq), where vq := v +(0, 0, q), which maps [E] ∈ M to [E ∨∨] ∈ M(vq). The fibre of α over a general vector bundle F ∈ Imα is isomorphic to the Quot-scheme QuotS(F, q) of zero dimensional quotient sheaves of F of length q; by [OG, Prop. 6.0.1], this Quot-scheme has dimension at most 3q. Hence, (44) dim M − 4q = dim M(vq) ≥ dim Imα = dim Mq − QuotS(F, q) ≥ dim Mq − 3q, and this concludes the proof. (cid:3) 6. Algebraically coisotropic subvarieties of IHS manifolds We are now ready to prove our results concerning existence of uniruled subvarieties of S[k] ε . Let (S, L) be a very general primitively polarized K3 or abelian surface of genus p, in the sense that it satisfies Theorem 3.1 and Pic(S) ≃ Z[L]. For p, δ, k satisfying (12), let [C] ∈ Yδ,k+ε be general and let A be a general complete g1 k+ε on its normalization. Let M be the component of M(vp,δ,k) as in the previous section, with [EC,ν∗A] ∈ M. Assume χ ≥ 2δ + 2 and set X := M × S[δ]. We denote by P the parameter space for triples ([E], τ, [s]) with ([E], τ ) ∈ X and [s] ∈ P(H 0(E ⊗ Iτ )); existence of P again follows from existence of a moduli space of OS-stable coherent systems and there is a natural map q : P → X. For any ([E], τ, [s]) ∈ P, the section s vanishes along a finite set because otherwise we would have Hom(det E, E ∨∨) 6= 0 and this would contradict the µL-stability of E. Hence, we have a short exact sequence: (45) 0 / OS s / E / det E ⊗ IWs / 0, / / / / 22 ANDREAS LEOPOLD KNUTSEN, MARGHERITA LELLI-CHIESA, AND GIOVANNI MONGARDI where Ws is a 0-dimensional subscheme of S of length k + ǫ + δ containing τ . This provides a short exact sequence (46) 0 / ηs / OWs / Oτ / 0, defining ηs, which is a torsion sheaf whose support is contained in that of Ws. If E is locally free, the scheme Ws is a local complete intersection as it is the zero scheme of s. It (OWs, OS ) ≃ OWs by, e.g., [Fr, p. 36]. Applying the functor homOS (−, OS ) to (46) follows that ext2 we therefore obtain a surjection OS OWs / ext2 OS (ηs, OS ), which yields the existence of a subscheme Zs ⊂ Ws of length k + ε such that ext2 This defines a rational map OS (ηs, OS ) ≃ OZs. (47) g′ : P 99K S[k+ε] mapping a point ([E], τ, [s]), with E locally free, to Zs. We want to extend g′ in codimension 1. Pick a sheaf [E] ∈ M that is not locally free and such that S(E) has length one, i.e., S(E) ≃ OP for some P ∈ S. A section s of E gives a section s′ of E ∨∨ and, having denoted by Ws′ the vanishing locus of s′, one has a short exact sequence: 0 −→ OP −→ OWs −→ OWs′ −→ 0. If both [E] ∈ M1 (cf. (43)) and s ∈ H 0(E) are general, then Ws′ does not contain P and P is a general point on S. Indeed, H 0(E) is not contained in H 0(E ∨∨ ⊗ IP ) as soon as H 1(E ∨∨ ⊗ IP ) = 0; the vanishing follows from the fact that a general [F ] ∈ M(v(E ∨∨)) is generically generated by global sections. For s ∈ H 0(E ⊗ Iτ ) with τ ∈ S[δ] such that P 6∈ Supp(τ ), we get OWs = OWs′ ⊕ OP , whence Ws is a local complete intersection, and we may find a subscheme Zs ⊂ Ws of length k + ε as before and set g′([E], τ, [s]) = Zs also in this case. We set M◦ := {[E] ∈ M \ ∪q≥2Mq h1(E ∨∨ ⊗ IP ) = 0 if S(E) ≃ OP }, with Mq as in (43); then M◦ is open in M and its complement has codimension at least two by Lemma 5.7. We define X ◦ := {([E], τ ) ∈ X [E] ∈ M◦, Supp(τ ) ∩ Supp(S(E)) = ∅}; the complement of X ◦ in X has codimension at least two. Let P ◦ be the open subscheme of q−1(X ◦) ⊂ P consisting of those triples ([E], τ, s) such that either E is locally free, or the vanishing locus Ws′ as above does not contain the singular locus of E. Then, the complement of P ◦ in P has codimension at least two and, by the above discussion, the rational map g′ in (47) is well-defined on the whole P ◦. We denote by (48) g : P ◦ / S[k+ε] the restriction of g′ to P ◦. We prove the following result. Theorem 6.1. Let (S, L) be a very general primitively polarized K3 or abelian surface of genus p ≥ 2. Let 0 ≤ δ ≤ p − 2ε and k ≥ 2 be integers satisfying (49) max{2δ + 2, 4ε} ≤ χ := p − δ − k + 3 − 5ε ≤ δ + k + 1. Then the morphism g is generically injective. In particular, the locus Locus(R+Rp,δ,k) ⊂ S[k] ε , with Rp,δ,k as in (16), has an irreducible component that is birational to a Pχ−2δ−1-bundle on a holomorphic symplectic manifold of dimension 2(k + 1 + 2δ − χ). / / / / / / WALL DIVISORS AND ALGEBRAICALLY COISOTROPIC SUBVARIETIES OF IHS MANIFOLDS 23 Proof. First of all, note that when χ ≥ 2δ + 2, condition (12) is equivalent to χ ≤ δ + k + 1. Indeed, the latter inequality is a rewrite of (14) with l = 1; on the other hand, if 2δ + 2 ≤ χ ≤ δ + k + ε, then δ ≤ k + ε − 2 and α = 1 in (12). Hence, (49) ensures that the hypotheses of Proposition 5.5 are satisfied and a general fibre of q : P ◦ → X ◦ is isomorphic to Pχ−2δ−1. We denote by T the closure of the image of the morphism g in (48). The proof proceeds by steps. STEP I: The morphism g is injective when restricted to a general fibre of the projection q. Let Z = g(([E], τ, [s])) ∈ T with ([E], τ ) ∈ X ◦ general, and denote by Ws the zero scheme of s. The fibre of gq−1([E],τ ) over Z is contained in P(Hom(E, det E ⊗ IWs)). This projective space is a point because Hom(E, det E ⊗ IWs) ≃ H 0(E ⊗ E ∨) ≃ C by stability, exact sequence (45) and Proposition 5.5(i), which yields h1(E) = 0. STEP II: The map g is generically injective when restricted to a general fibre of the projection p1 : P ◦ → S[δ]. If δ = 0, then g is injective. Let Z = g(([E1], τ, [s1])) ∈ T with ([E1], τ, [s1]) ∈ P ◦ general; in particular, Supp(τ ) is disjoint from Supp(Z) and E1 satisfies h1(E1⊗L0) = 0 for all L0 ∈ Pic0(S) by Lemma 5.6. By contradiction, assume the existence of ([E2], τ, [s2]) ∈ P ◦ with E2 6≃ E1 such that g(([E2], τ, [s2])) = Z. In particular, one has Ws1 = Ws2. We remark that det E1 6≃ det E2 because otherwise we would have h1(det Ei ⊗ IWsi ) > 1 and, by considering the long exact sequence in cohomology associated with (45) for E1, we would get a contradiction with h1(E1) = 0. We tensor (45) for E1 with (det E1)∨ ⊗ det E2, thus obtaining: 0 → (det E1)∨ ⊗ det E2 −→ E1 ⊗ (det E1)∨ ⊗ det E2 −→ det E2 ⊗ IWs −→ 0. This yields the contradiction H 1(E1 ⊗ (det E1)∨ ⊗ det E2) ≃ H 1(det E2 ⊗ IWs) 6= 0. STEP III: The map g is generically injective when restricted to a general fibre of the projection p2 : P ◦ → M◦ ⊂ M. Let Z = g(([E], τ1, [s1])) ∈ T for a general ([E], τ1, [s1]). By contradiction, assume the existence of a subscheme τ2 ∈ S[δ] different from τ1 and a section s2 ∈ H 0(E ⊗ Iτ2) such that g(([E], τ2, [s2])) = Z. The evaluation map ev : hs1, s2i ⊗ OS → E is injective and drops rank along an integral curve Γ ∈ {L} of geometric genus ≤ p − k − ε (indeed, Γ is singular along Z). If B is the cokernel of ev, then B1 := ext1(B, OS ) = n∗A1, where n : eΓ → Γ is a partial normalization of Γ with δ−h on eΓ. As a consequence, there is a pa(eΓ) = p − k − ε − h for some h ≥ 0 and A1 a complete g1 subscheme τ ′ 1]. Starting from R′, one easily constructs a rational curves in S[δ] passing through τ1 in contradiction with the generality of τ1. 1 ⊂ τ1 of length δ − h and a rational curve R′ in S[δ−h] passing through [τ ′ STEP IV: The map g is generically finite. This follows from the previous steps and Proposition 2.15, which can be applied because the complement of P ◦ in P has codimension at least two and X is holomorphic symplectic. STEP V: Let Pi = P(H 0(Ei ⊗ Iτi)) for i = 1, 2 with ([E1], τ1), ([E2], τ2) ∈ X ◦ distinct points such that h1(Ei) = 0 for i = 1, 2. Then g does not identify P1 and P2. By contradiction, assume that g(P1) = g(P2). As gPi is injective for i = 1, 2 by Step I, it is easy to verify that, if ℓ1 ⊂ P1 is a general line, then ℓ2 := g−1(g(ℓ1)) ∩ P2 is a line, too. By generality, ℓ1 = PV1 corresponds to a pair (C1, ν∗A1), where C1 is a δ-nodal curve and A1 is a g1 k+ε on its normalization. Having set ℓ2 = PV2, the evaluation map V2 ⊗ OS → E2 drops rank along a curve C2 singular along τ2. Then, C1 = C2 as both coincide with the image in S of the incidence variety I := {(Z, P ) ∈ g(ℓ1) × S ⊂ Sk+ε × S P ∈ Supp(Z)}. As a consequence, τ2 = τ1 and E2 ≃ EC,ν∗A1 ≃ E1. STEP VI: The morphism g is generically injective. 24 ANDREAS LEOPOLD KNUTSEN, MARGHERITA LELLI-CHIESA, AND GIOVANNI MONGARDI By contradiction, assume that for a general Z ∈ T there exist at least two distinct points ([E1], τ1), ([E2], τ2) ∈ X ◦ such that Z ∈ g(P1) ∩ g(P2), where Pi = P(H 0(Ei ⊗ Iτi)); we may as- sume that h1(E1) = h1(E2) = 0. The previous step then implies that g(P1) 6= g(P2). We denote by π : eT 99K B the maximal rational quotient of the desingularization eT of T , and by eZ the inverse image of Z in eT . Since a general fibre of π is irreducible and π−1(π(eZ)) contains the strict transforms of both g(P1) and g(P2), then dim π−1(π(eZ)) ≥ χ − 2δ > codimS[k+ε] T ; this contradicts [AV, Thm. 4.4] (when ε = 1, in order to apply the mentioned result, one needs to pass to the fibers of the Albanese map and use that g(P1) ∪ g(P2) is contained in such a fiber). When ε = 0, Step VI concludes the proof. If ε = 1, consider the composition of g with the Albanese map Σk : S[k+1] → S. Since Σk ◦ g is constant when restricted to any fibre of q, it induces a morphism F : X ◦ → S that factors, by the universal property of the Albanese variety, through a map f : Alb(X) → S. One can easily show that both F and f are surjective. The inverse image g−1(S[k] ε ) ⊂ P ◦ is generically a Pχ−2δ−1-bundle on the inverse image (albX X ◦)−1(ker f ), where albX is the Albanese map of X. Since Alb(X) is the product of copies of S and S∨, the same holds for any connected component of ker f . Therefore, any component of (albX X ◦)−1(ker f ) is holomorphic symplectic of dimension equal to dim X − 2 = 2(k + 1 + 2δ − χ), and the statement follows. (cid:3) We now give a first application to Conjecture 0.3. Corollary 6.2. With the same hypotheses and notation as in Theorem 6.1, set r := χ − 2δ − 1. Then, Sr(S[k] ε ) has a (2k − r)-dimensional component which is an algebraic coisotropic subvariety of S[k] ε covered by curves of class Rp,δ,k. Proof. This follows directly from Theorem 6.1 and [Vo, Thm. 1.3], which states that a closed subvariety of an IHS manifold X contained in Sr(X) has codimension at least r. (cid:3) Starting from the closure of Im g ⊂ S[k+ε], with g as in 48, and then applying the natural rational for any l ≥ k. We use this observation map S[k+ε] × S[l−k] 99K S[l+ε], one obtains subvarieties of S[l] ε in order to construct subvarieties of S[k] ε , with k fixed, of codimension r for several values of r: Theorem 6.3. Let (S, L) be a very general primitively polarized K3 or abelian surface of genus p ≥ 2 and fix an integer k ≥ 2. Then, for any integer r satisfying 1 ≤ r ≤ min(cid:26)2k − 5 − p − 5ε 2 , p − 5ε 2 + 1(cid:27) , p ≥ 9 if (ε, r) = (1, 1); p ≥ 11 if (ε, r) = (1, 2), (50) and (51) (52) there is an algebraically coisotropic subvariety of codimension r in S[k] ε and is birational to a Pr-bundle over a holomorphic symplectic manifold. that is a component of Sr(S[k] ε ) More precisely, for any integer δ satisfying max(cid:26)0, p − 5ε + 2 − r − k 3 p − 5ε + 2 − 2r 4 , and δ > 0 if r ≤ 2 and ε = 1, (cid:27) ≤ δ ≤ there is such a subvariety Σr,δ whose lines have class L − [2(p − 2δ − 2ε) − r + 1]rk. Proof. We note that if δ and k are integers satisfying the conditions in Theorem 6.1, then, for any integer l ≥ k, the image of Im g × S[l−k] ⊂ S[k+ε] × S[l−k] under the natural birational map S[k+ε] × S[l−k] 99K S[l+ε] has the same codimension r as Im g ⊂ S[k+ε]. The intersection of this image with any fibre of the Albanese map is birational to a Pr-bundle over a holomorphic symplectic WALL DIVISORS AND ALGEBRAICALLY COISOTROPIC SUBVARIETIES OF IHS MANIFOLDS 25 manifold and the coefficients in the class of the lines with respect to the canonical decomposition (8) remain unchanged. Hence, Theorem 6.1 yields that for any pair of integers (δ, k′) such that (53) and 2 ≤ k′ ≤ k (54) there is a subscheme of codimension r := p − 5ε − 3δ − k′ + 2 satisfying the desired conditions. Rewriting (53) and (54) in terms of r instead of k′ yields, respectively, max{2δ + 2, 4ε} ≤ p − 5ε − δ − k′ + 3 ≤ δ + k′ + 1, δ ≥ 0 (55) and (56) p − 5ε + 2 − r − k 3 ≤ δ ≤ p − 5ε + r 3 1 2 max{0, 4ε − r − 1} ≤ δ ≤ p − 5ε + 2 − 2r 4 , One easily verifies that these two conditions are equivalent to (52). It only remains to prove that for each r satisfying (50) and (51), the conditions (52) are nonempty. The inequality r ≤ 2k − 5 − p−5ε 2 ensures that thus guaranteeing that the interval h p−5ε+2−r−k 3 3 i contains an integer. The condition p − 5ε + 2 − r − k p − 5ε + 2 − 2r ≤ 4 , p−5ε+2−2r 4 − 3 4 , r ≤ p−5ε 2 + 1 is equivalent to p − 5ε + 2 − 2r ≥ 0 and thus guarantees that the above interval contains a nonnegative integer. We are therefore done, except for the cases ε = 1 and r ≤ 2, where the requirement is that δ > 0 by (52), that is, that p − 5ε + 2 − 2r ≥ 4, which is precisely (51). (cid:3) We now perform a different construction in order to exhibit uniruled subvarieties of S[k] ε of any allowed codimension, except codimension k for ε = 1. This provides very strong evidence for Con- jecture 0.3. Theorem 6.4. Let (S, L) be a general primitively polarized K3 or abelian surface of genus p ≥ 2 and fix an integer k ≥ 2. Then for any integer r such that 1 ≤ r ≤ k − ε there is an algebraically coisotropic subvariety of codimension r in S[k] ε ) and is birational to a ε Pr-bundle; furthermore, the maximal rational quotient of it desingularization has dimension 2(k − r). More precisely, for any integer k′ such that r+ε ≤ k′ ≤ min{k, p+r−ε}, there is such a subvariety that is a component of Sr(S[k] Wr,k′ whose lines have class L − [2(k′ + ε) − r − 1]rk. Proof. For any r and k as in the statement, set g := k′ − r + ε and δ := p − g. Note that 2ε ≤ g ≤ p. Since ρ(g, 1, k + ε) ≥ 0, the Brill-Noether locus {L}1 δ,k+ε coincides with the g-dimensional Severi variety V{L},δ of genus g nodal curves. For any component V of V{L},δ, we denote by C → V the universal family and by eC → V the simultaneous desingularization of all curves in C (the latter exists, as V is smooth). Let Symk′+ε(eC) → V be the relative (k′ + ε)-symmetric product, with fibre over a point [C] ∈ V equal to Symk′+ε(eC), where eC is the normalization of C. By surjectivity of the Abel map Symk′+ε(eC) → Pick′+ε(eC), the line bundle O eC(x1 + · · · + xk′+ε) is nonspecial if x1, . . . , xk′+ε ∈ eC are general . Therefore, dim O eC (x1 + · · · + xk′+ε) = k′ + ε − g = r ≥ 1 and Symk′+ε(eC) is generically a Pr-bundle over a dense, open subset of Pick′+ε(eC), which has dimension g. It follows that Symk′+ε(eC) is generically a Pr-bundle over a scheme of dimension g + dim V = 2g. We first prove that f is generically injective. Clearly, for each curve [C] ∈ V with normalization eC, the restriction of f to Symk′+ε(eC) is generically injective. Hence, it is enough to show that, if [Z] ∈ Imf is f : Symk′+ε(eC) → Symk′+ε(C) → Symk′+ε(S). Consider the natural composed morphism 26 ANDREAS LEOPOLD KNUTSEN, MARGHERITA LELLI-CHIESA, AND GIOVANNI MONGARDI general, then the scheme Z is contained in a unique curve parametrized by V . A general [Z] ∈ Imf consists of k′ + ε general points on a general curve C0 parametrized by V . For a general point x1 ∈ C0, the set {[C] ∈ V x1 ∈ C} has codimension one in V , as C0 is not a common component of all curves parametrized by V . Proceeding inductively, assume that for a fixed 1 ≤ j ≤ g − 1 we have chosen j distinct points x1, . . . , xj ∈ C0 such that the set {[C] ∈ V x1 . . . , xj ∈ C} has codimension j in V . Again, as C0 is not a component of any curve in this set different from C0, for a general xj+1 ∈ C0 the set {[C] ∈ V x1 . . . , xj+1 ∈ C} has codimension j + 1 in V . It follows that dim{[C] ∈ V x1 . . . , xg ∈ C} = 0 for general points x1, . . . , xg ∈ C0. Hence, for general points x1, . . . , xg+1 ∈ C0, we have {[C] ∈ V x1 . . . , xg+1 ∈ C} = {C0}, and the generic injectivity of f follows since k′ + ε = g + r ≥ g + 1. The image of f does not lie in the singular locus of Symk′+ε(S). Hence, its inverse image under the Hilbert-Chow morphism is a (k′ + ε + g)-dimensional subvariety of S[k′+ε] that is birational to a Pr-bundle. As above, the natural rational map S[k′+ε] × S[k−k′] 99K S[k+ε] maps Im f to a codimension r subvariety of S[k+ε]. Since the Albanese map is constant on each rational subvariety of S[k+ε], we obtain a subvariety Wr,k′ ⊂ S[k] ε of codimension r that is birational to a Pr-bundle and the maximal rational quotient of the desingularization of Wr,k′ has dimension 2(k − r) by [AV, Thm. 4.4]. The coefficients in the class of the lines in the Pr-fibres of Wr,k′ with respect to the canonical decomposition (8) are the same as the ones of Rp,δ,k′ = Rp,p+r−k′−ε,k′ = L − [2(k′ + ε) − r − 1]rk′. This concludes the proof. (cid:3) We conclude with an interesting example, where Theorem 6.1 provides an immersion Pk ֒→ S[k] ε . Example 6.5. By Remark 3.4 , when δ = 0 the class Rp,0,k has the minimal possible self-intersection (i.e., q(Rp,0,k) = −(k + 3 − 2ε)/2) if and only if α = 1 and p = 2(k − 1) + 5ε. We assume these numerical conditions are satisfied and show, by explicit construction, that Rp,0,k is the class of a line moving in a Pk ⊂ S[k] ε . Note that in this case dim M = 2ε and condition (49) is satisfied. With the same notation introduced just before Theorem 6.1, P ◦ = P and any component of g−1(S[k] ε ) is isomorphic to PH 0(E) for some vector bundle [E] ∈ M. We consider the restricted morphism g := gPH 0(E) : Pk = PH 0(E) / S[k] ε ⊆ S[k+ε], which is injective by Theorem 6.1 and, more precisely, an embedding by the result below. Proposition 6.6. Let (S, L) be a general primitively polarized symplectic surface of genus p = 2(k − 1) + 5ε for an integer k ≥ 2. Then the map g is an embedding. In particular, the class Rp,0,k is the class of a line in a Pk ⊂ S[k] ε . Proof. It is enough to show that for all [s] ∈ PH 0(E) the differential dg[s] : T[s]P(H 0(E)) → T[Z]S[k+ε] is injective, where Z is the zero scheme of s. First of all, we recall the isomorphisms T[s]PH 0(E) ≃ Hom(C, H 0(E)/hsi) ≃ H 0(E)/hsi and T[Z]S[k+ε] = H 0(NZ/S), and use them in order to describe dg[s]. Given t ∈ H 0(E)/hsi, the evaluation map ev : ht, si ⊗ OS → E is injective and drops rank along a curve Γt ∈ {L} containing Z. We denote by B the cokernel of ev, which is supported on Γt. Since h2(E) = h1(E) = 0 by Proposition 5.5, we obtain isomorphisms: (57) where B1 := ext1 the duals of the images of s and t under (57), and consider the following short exact sequence: (B, OS). By (26), one has B1 = homOΓt ht, si ≃ H 1(B) ≃ H 0(B1)∨, OS (IZ/Γt , OΓt ). Let σs, σt ∈ H 0(B1) denote 0 / OΓt σs / homOΓt (IZ/Γt, OXt ) αt / homOΓt (IZ/Γt, OZ ) / 0. / / / / / WALL DIVISORS AND ALGEBRAICALLY COISOTROPIC SUBVARIETIES OF IHS MANIFOLDS 27 Then we have 0 6= H 0(αt)σt ∈ H 0(homOΓt inclusion is given by taking cohomology in (IZ/Γt , OZ )) = H 0(NZ/Γt) ֒→ H 0(NZ/S), where the last (58) 0 / NZ/Γt ιt / NZ/S / NΓt/SZ / 0. By construction, dg[s](t) = H 0(ιt) ◦ H 0(αt)σt ∈ H 0(NZ/S). We are now able to prove the injectivity of dg[s]. Let t, t′ ∈ H 0(E)/hsi such that dg[s](t) = dg[s](t′). We first assume that Γt ≃ Γt′. Since h1(E) = 0, then the natural map h : G(2, H 0(E)) → G1 k+1(L) is injective; indeed, any V ∈ G(2, H 0(E)) defines a short exact sequence like (38) and and it is enough to tensor it with E ∨ in order to conclude that P(Hom(E, B)) is a point. Since h(hs, ti) = h(hs, t′i) = (Γt, OΓt (Z)), we have hs, ti = hs, t′i, that is, t′ = λt for some λ ∈ C. Then, σλt = λσt 6= σt 6≃ Γt; it is then clear from (58) that the section unless λ = 1. Therefore, we can assume Γt′ H 0(ιt′) ◦ H 0(αt′ )σt′ ∈ H 0(NZ/S) does not lie in the image of H 0(ιt). This concludes the proof. (cid:3) References [Ad] [AV] N. Addington, On two rationality conjectures for cubic fourfolds, arXiv:1405.4902 E. Amerik, M. Verbitsky, Rational curves on hyperkahler manifolds, Int. Math. Res. Notices (2015), doi: 10.1093/imrn/rnv133. [ACG] E. Arbarello, M. Cornalba, P. A. Griffiths, Geometry of algebraic curves. Volume II. With a contribution by Joseph Daniel Harris, Grundlehren der mathematischen Wissenschaften 267, Springer-Verlag, Berlin (2011). B. Bakker, A classification of Lagrangian planes in holomorphic symplectic varieties, arXiv:1310.6341. A. Bayer and E. Macrì, MMP for moduli of sheaves on K3s via wall crossing: Nef and movable cones, Lagrangian fibrations, Invent. Math 198 (2014), 505 -- 590. [Ba] [BM] [BM2] A. Bayer, E. Macrì, Projectivity and birational geometry of Bridgeland moduli spaces, J. Amer. Math. Soc 27 (2014), 707 -- 752. [BHT] A. Bayer, B. Hassett and Y. Tschinkel, Mori cones of holomorphic symplectic varieties of K3 type, [Be1] [Be2] [BP] [CP] [CK] [Fu] arXiv:1307.2291. A. Beauville, Variétés Kählériennes dont la première classe de Chern est nulle, J. Diff. Geom. 18 (1983), 755 -- 782. A. Beauville, On the splitting of the Bloch-Beilinson filtration, in Algebraic cycles and motives. Vol. 2, Lond. Math. Soc. Lecture Notes 344 (2007), Cambridge University Press, 38 -- 53. U. N. Bhosle, A. J. Parameswaren Picard bundles and Brill-Noether loci in the compactified Jacobian of a nodal curve, Int. Math. Res. Not., http://dx.doi.org: 10.1093/imrn/rnt069 (2013). F. Charles, G. Pacienza, Families of rational curves on holomorphic symplectic varieties, arXiv:1401.4071 C. Ciliberto, A. L. Knutsen, On k-gonal loci in Severi varieties on general K3 surfaces, J. Math. Pures Appl. 101 (2014), 473 -- 494. A. Fujiki, Finite Automorphism Groups of Complex Tori of Dimension Two, Publ. RIMS, Kyoto Univ. 24 (1988), 1 -- 97. R. Friedman, Algebraic surfaces and holomorphic vector bundles, Springer-Verlag, New York (1998). [Fr] [HT1] B. Hassett, Y. Tschinkel, Intersection numbers of extremal rays on holomorphic symplectic varieties, Asian J. Math. 14 (2010), 303 -- 322. [HT2] B. Hassett, Y. Tschinkel, Moving and ample cones of holomorphic symplectic fourfolds, Geom. Funct. Anal. 19 (2009), 1065 -- 1080. [HT3] B. Hassett, Y. Tschinkel, Extremal rays and automorphisms of holomorphic symplectic varieties, arXiv:1506.08153 [Hu2] D. Huybrechts, Compact hyperkähler manifolds: basic results, Invent. Math. 135 (1999), 63 -- 113. Erratum: "Compact hyperkähler manifolds: basic results", Invent. Math. 152 (2003), 209 -- 212. [Hu1] D. Huybrechts, The Kähler cone of a compact hyperkähler manifold, Math. Ann. 326 (2003), 499 -- 513. [Ki] T. Kirschner, Irreducible symplectic complex spaces, PhD thesis, Universität Bayreuth, 2012. Available at epub.uni-bayreuth.de/209/1/diss-kirschner-bibprint.pdf. [KLM] A. L. Knutsen, M. Lelli-Chiesa, and G. Mongardi Severi varieties and Brill-Noether theory of curves on abelian surfaces, arXiv:1503.04465. [LP] C. Lehn, G. Pacienza On the log minimal model program for irreducible symplectic varieties, arXiv:1405.5649. [LC1] M. Lelli-Chiesa, generalized Lazarsfeld-Mukai bundles and a conjecture of Donagi and Morrison, Adv. Math. 268 (2015), 529 -- 563. [LC2] M. Lelli-Chiesa, Green's Conjecture for curves on rational surfaces with an anticanonical pencil, Math. Z. 275 (2013), 899 -- 910. / / / / / / 28 ANDREAS LEOPOLD KNUTSEN, MARGHERITA LELLI-CHIESA, AND GIOVANNI MONGARDI [LC3] M. Lelli-Chiesa, Stability of rank-3 Lazarsfeld-Mukai bundles on K3 surfaces, Proc. Lond. Math. Soc. 107 [Li] [Ma] (2013), 451 -- 479. H. Y. Lin, On the Chow group of zero-cycles of a generalized Kummer variety, arXiv:1507.05155 E. Markman, A survey of Torelli and monodromy results for Hyperkähler manifolds, Proc. of the conference "Complex and Differential Geometry", Springer Proceedings in Math. (2011), Volume 8, 257 -- 322. [Mo1] G. Mongardi, A note on the Kähler and Mori cones of hyperkähler manifolds, to appear in Asian. J. Math. [Mo2] G. Mongardi, On the monodromy of irreducible symplectic manifolds, arXiv:1407.4269 [MW] G. Mongardi and M. Wandel, Induced automorphisms on irreducible symplectic manifolds, J. London Math. [Na] [OG] [Pa] [Ra] [Se] [Ta] [Vo] [Yo1] [Yo2] [Yo3] [Wi] Soc. (2015), doi: 10.1112/jlms/jdv012 Y. Namikawa, Deformation theory of singular symplectic n-folds, Math. Ann. 319 (2001), 597 -- 623. K. O'Grady, The irreducible components of moduli spaces of vector bundles on surfaces, Invent. Math. 112 (1993), 585 -- 613. G. Pareschi, A proof of Lazarsfeld's Theorem on curves on K3 surfaces, J. Alg. Geom. 4 (1995), 195 -- 200. Z. Ran, Hodge theory and deformations of maps, Compos. Math. 97 (1995), 309 -- 328. E. Sernesi, Deformations of algebraic schemes, Grundlehren der mathematischen Wissenschaften 334, Springer -- Verlag, Berlin Heidelberg (2006). A. Tannenbaum, Families of algebraic curves with nodes, Compos. Math. 41 (1980), 107 -- 126. C. Voisin, Remarks and questions on coisotropic subvarieties and 0-cycles of hyper-Kähler varieties, arXiv:1501.02984. K. Yoshioka, Irreducibility of moduli spaces of vector bundles on K3 surfaces, arXiv:9907001. K. Yoshioka, Moduli spaces of stable sheaves on abelian surfaces, Math. Ann. 321 (2001), 818 -- 884. K. Yoshioka, Bridgeland's stability and the positive cone of the moduli spaces of stable objects on an abelian surface, arXiv:1206.4838. J. Wierzba, Contractions of symplectic varieties, J. Alg. Geom. 12 (2003), 507 -- 534. Andreas Leopold Knutsen, Department of Mathematics, University of Bergen, Postboks 7800, 5020 BERGEN, Norway E-mail address: [email protected] Margherita Lelli-Chiesa, Centro di Ricerca Matematica Ennio De Giorgi, Scuola Normale Supe- riore, Piazza dei Cavalieri 3, 56100 Pisa, Italy E-mail address: [email protected] Giovanni Mongardi, Department of Mathematics, University of Milan, via Cesare Saldini 50, 20133 Milan, Italy E-mail address: [email protected] 

1811.01865

2

1811

Identifiability for a class of symmetric tensors

We use methods of algebraic geometry to find new, effective methods for detecting the identifiability of symmetric tensors. In particular, for ternary symmetric tensors T of degree 7, we use the analysis of the Hilbert function of a finite projective set, and the Cayley-Bacharach property, to prove that, when the Kruskal's rank of a decomposition of T are maximal (a condition which holds outside a Zariski closed set of measure 0), then the tensor T is identifiable, i.e. the decomposition is unique, even if the rank lies beyond the range of application of both the Kruskal's and the reshaped Kruskal's criteria.

math.AG

math

IDENTIFIABILITY FOR A CLASS OF SYMMETRIC TENSORS ELENA ANGELINI, LUCA CHIANTINI, AND ANDREA MAZZON Abstract. We use methods of algebraic geometry to find new, effective meth- ods for detecting the identifiability of symmetric tensors. In particular, for ternary symmetric tensors T of degree 7, we use the analysis of the Hilbert function of a finite projective set, and the Cayley-Bacharach property, to prove that, when the Kruskal's ranks of a decomposition of T are maximal (a condi- tion which holds outside a Zariski closed set of measure 0), then the tensor T is identifiable, i.e. the decomposition is unique, even if the rank lies beyond the range of application of both the Kruskal's and the reshaped Kruskal's criteria. 1. Introduction 1 + · · · + Ld In the study of symmetric tensors T of type (n + 1) × · · · × (n + 1) (d times), which can be identified with polynomial forms of degree d in n + 1 variables, one of the main aspects concerns their Waring decompositions, i.e. decompositions of type T = Ld r, where each Li is a linear form. The (Waring) rank of T , which is the minimum r for which the decomposition exists, turns out to be a good measure for the complexity of T . Many recent results are devoted to the computation of the (Waring) rank of forms. Effective methods for the computation of decompositions are available, though their practical implementation is often out of reach (see e.g. [14], [4]). An important, open question, however, is related to the uniqueness of a decom- position with r minimal, i.e. the uniqueness of a decomposition which computes the rank. The existence of a unique decomposition that computes the rank of T is a property of T denoted as identifiability. Uniqueness is important for the applica- tions, at least for two reasons. First, the computation of an effective decomposition is often faster, when the target is uniquely determined. Second, if we are looking for a specific decomposition, the uniqueness ensures us that if we get close to one of them, then that one is exactly what we were looking for. Here, with uniqueness of the decomposition we mean uniqueness up to trivialities, like permutation of the summands or their rescaling. We mention the papers [2], [3], [6] for examples of applications. When T is a sufficiently general form of rank r, and rn + r + n + 1 < (cid:0)d+n n (cid:1) (technically, when we are in the range of subgeneric ranks), then the identifiability of T is known to hold, with the exclusion of a short list of cases for (n, d, r), which are completely described (see [11]). On the other hand, for specific forms T , the question of the computation of the rank and the identifiability of T is still widely open, except for small values of the rank. More in details, given a form T , which arises either from theoretical reasons 2010 Mathematics Subject Classification. 14J70, 14C20, 14N05, 15A69, 15A72. Key words and phrases. Symmetric tensors, Waring identifiability, Kruskal's criterion, Hilbert function, Cayley-Bacharach property. 1 2 E. ANGELINI, L. CHIANTINI, AND A. MAZZON 1 + · · · + Ld or from experimental data, it is often possible to determine a specific decomposition T = Ld r, but it remains unclear whether the decomposition computes the rank or it is unique, i.e. it is unclear if we can exclude the existence of a second decomposition with r′ ≤ r summands. For specific forms, of which we know a decomposition T = Ld r, the most famous criterion to exclude the existence of a second decomposition with r′ ≤ r summands goes back to the 70's, to the celebrated paper by Kruskal [15]. The criterion, which is valid even for non-symmetric tensors but holds only for small values of r, is based on the notion of Kruskal's ranks of a decomposition (see Definition 2.2 below): if r satisfies an inequality in terms of the Kruskal's ranks, then r is the rank of T and the decomposition is unique. Derksen [12] proved that the inequality is sharp: one can not hope to improve Kruskal's criterion by using just the Kruskal's ranks of the decomposition. 1 + · · · + Ld In the specific case of symmetric tensors, however, other methods have been introduced to study the uniqueness of a decomposition. One of them, based on the study of the kernels of catalecticant maps, can be found in [16]. Another method, based on the Kruskal's original criterion, is described in [10] and it is based on the computation of the Kruskal's ranks of a reshaping of a decomposition, i.e. the Kruskal's ranks of the points obtained by raising the Li's to some partial powers d1, d2, d3, with d1 + d2 + d3 = d. Both methods usually work in a range wider than the range in which the original Kruskal's criterion can be applied. Thus, e.g., the determination of the reshaped Kruskal's rank can exclude Derksen's examples in the symmetric case (see [7], [8] for other methods). Yet, all the previous methods will give no answer when r grows enough. For instance, for the case of ternary septics, i.e. the case n = 2, d = 7, subgeneric ranks (for which identifiability holds generically) are ranks r = 1, . . . , 11, but the original Kruskal's criterion will give no answer for r > 7, while both the catalecticant method or the reshaped Kruskal's method will work only for r ≤ 10. The aim of this paper is an extension of a method, introduced in [10] and in [5] for the case of forms of degree 4, which can determine that r is the rank of T and the decomposition is unique, even in a range in which both the catalecticant method or the reshaped Kruskal's method do not work. The method is still based on the computation of the reshaped Kruskal's ranks of the decomposition, but the conclusion is obtained with advanced tools of algebraic geometry: the analysis of the Hilbert function of the set of projective points associated with a decomposition. We apply the method to ternary forms of degree 7, and we find that if T has a decomposition T = Ld r whose reshaped Kruskal's ranks are general, then T has rank r and the decomposition is unique (up to trivialities). We obtain thus an effective criterion for the identifiability of specific symmetric tensors, which definitively improves our previous knowledge. The proof of the criterion is based on results for the analysis of the postulation of finite sets in projective spaces (e.g. on the Cayley-Bacharach property) which, we believe, have an independent theoretical interest for investigators in the field. In particular, our result also proves that, for a general choice of the linear forms Li's, the span of Ld r, which is a subset of the r-th secant variety S of the Veronese image of degree 7 of P2 = P(C3), does not contain singular points of S, outside those generated by a proper subset of the powers Ld i . 1 + · · · + Ld 1, . . . , Ld IDENTIFIABILITY FOR A CLASS OF SYMMETRIC TENSORS 3 In section 6 we consider forms of any degree, with a decomposition in which the Kruskal's ranks are not maximal, but are maximal modulo the fact that the points representing the decomposition lie in a plane cubic curve. We show also in this case that, with the same methods, one can prove the identifiability of the form T . We end by noticing that ternary septics T are the last numerical case in which the analysis of the Kruskal's rank of a decomposition is sufficient to conclude, outside a (Zariski closed) set of measure 0, that a given decomposition is the unique one that computes the rank of T . In a forthcoming series of papers we will prove that, already for ternary optics, the linear span of every general decomposition contains both identifiable and not identifiable points. Acknowledgements. The first two authors are members of the Italian GNSAGA- INDAM and are supported by the Italian PRIN 2015 - Geometry of Algebraic Varieties (B16J15002000005). 2. Notation We work over the complex field C. For any finite subset A of the projective space Pn = P(Cn+1), we denote by ℓ(A) the cardinality of A. We call vd : Pn → PN , N = (cid:0)n+d d (cid:1) − 1, the Veronese embedding of degree d. The space PN can be identified with P(Symd(Cn+1)). So, the points of PN can be identified, up to scalar multiplication, with forms T of degree d in n + 1 variables. By abuse, we will denote by T both the form in Symd(Cn+1) and the point in PN which represents T . The form T belongs to the Veronese variety vd(Pn) if and only if T = Ld, for some linear form L. With the above notations we give the following definitions. Definition 2.1. Let A ⊂ Pn be a finite set, A = {P1, . . . , Pr}. A is a decomposition of T ∈ P(Symd(Cn+1)) if T ∈ hvd(A)i, the linear space spanned by the points of vd(A). In other words, for a choice of scalars ai's, T = a1vd(P1) + · · · + arvd(Pr). The number ℓ(A) is the length of the decomposition. The decomposition A is minimal or non-redundant if T is not contained in the span of vd(A′), for any proper subset A′ ⊂ A. In particular, if vd(A) is not linearly independent, then A can not be minimal. If we identify points P ∈ Pn = P(Cn+1) with linear forms L in n + 1 variables, then the image vd(L) corresponds to the power Ld, so that a decomposition of T ∈ P(Symd(Cn+1)) corresponds to a Waring decomposition of a form of degree d. Definition 2.2. For a finite set A ⊂ Pn, the Kruskal's rank k(A) of A is the maximum k for which any subset of cardinality ≤ k of A is linearly independent. The Kruskal's rank k(A) is bounded above by n + 1 and ℓ(A). If A is general enough, then k(A) = min{n + 1, ℓ(A)}. For instance, if A ⊂ P3 is a set of cardinality 5, with a subset of 4 points on a plane and no three points aligned, then k(A) = 3. 4 E. ANGELINI, L. CHIANTINI, AND A. MAZZON Remark 2.3. It is straightforward that k(A) attains the maximum min{n+1, ℓ(A)} when all subsets of A cardinality at most n + 1 are linearly independent. In this case, for any subset A′ ⊂ A one has k(A′) = min{n + 1, ℓ(A′)}. Notice also that for any subset A′ ⊂ A, we have kA′ ≥ min{ℓ(A′), kA}. We can use the Veronese maps to define the higher Kruskal's ranks of a finite set A. Definition 2.4. For a finite set A ⊂ Pn, the d-th Kruskal's rank kd(A) of A is the Kruskal's rank of the image of A in the Veronese map vd. Thus, the Kruskal's rank k(A) coincides with the first Kruskal's rank k1(A). The d-th Kruskal's rank kd(A) is thus bounded by min{ℓ(A), (cid:0)n+d n (cid:1)}. Remark 2.5. Since the projective spaces are irreducible, then any subset of a general finite set A is general. Thus one can prove that, for a sufficiently general subset A ⊂ Pn, then all the Kruskal's ranks kd(A) are maximal. 3. Preliminary results We collect in this section the main technical tools for the investigation of the identifiability of symmetric tensors. For the proofs of most of them, we refer to Sect. 2 of [9] and Sect. 2.2 of [5]. Definition 3.1. Let Y ⊂ Cn+1 be an ordered, finite set of cardinality ℓ of vectors. Fix an integer d ∈ N. The evaluation map of degree d on Y is the linear map evY (d) : Symd(Cn+1) → Cℓ which sends F ∈ Symd(Cn+1) to the evaluation of F at the vectors of Y . Let Z ⊂ Pn be a finite set. Choose a set of homogeneous coordinates for the points of Z. We get an ordered set of vectors Y ⊂ Cn+1, for which the evaluation map evY (d) is defined for every d. If we change the choice of the homogeneous coordinates for the points of the fixed set Z, the evaluation map changes, but all the evaluation maps have the same rank. So, we can define the Hilbert function of Z as the map: hZ : Z → N hZ(d) = rank(evY (d)). We point out that the Hilbert function does not depend on the choice of the coor- dinates, as well as it does not vary after a change of coordinates in Pn. We define the first difference of the Hilbert function DhZ of Z as: DhZ (j) = hZ(j) − hZ(j − 1), j ∈ Z. Let vd : Pn → PN be the d-th Veronese embedding of Pn. For any finite set Z ⊂ Pn, and for any d ≥ 0, the value hZ(d) determines the dimension of the span of vd(Z). Indeed we have the following straightforward fact: Proposition 3.2. Let vd : Pn → PN be the d-th Veronese embedding of Pn. hZ (d) = dim(hvd(Z)i) + 1. From the previous proposition it follows that the Kruskal's rank kd(Z) is maximal n (cid:1) of vd(Z) is linearly if and only if every subset of cardinality at most N + 1 = (cid:0)n+d independent. IDENTIFIABILITY FOR A CLASS OF SYMMETRIC TENSORS 5 Several properties of the Hilbert functions and their differences are well known in Algebraic Geometry. We will need in particular the following facts (for the proofs, see Sect. 2 of [9]). Proposition 3.3. If Z ′ ⊂ Z, then for every d we have hZ ′ (d) ≤ hZ(d) and DhZ ′(d) ≤ DhZ(d). Proposition 3.4. Assume that for some j > 0 we have DhZ(j) ≤ j. Then: DhZ(j) ≥ DhZ (j + 1). In particular, if for some j > 0, DhZ(j) = 0, then DhZ(i) = 0 for all i ≥ j. We introduce the following notation, which comes from a cohomological inter- pretation of the Hilbert function of Z. We will use it as a mere symbolism. Definition 3.5. For every finite set Z and for every degree d, define: h1 Z(d) = ℓ(Z) − hZ(d) = ∞ X j=d+1 DhZ(j). In particular, from Proposition 3.2 we get: (1) dim(hvd(Z)i) = ℓ(Z) − 1 − h1 Z(d). Often, we will use the properties of the Hilbert function of the union of two different decompositions of a form. Thus, for our application the following property, which is a consequence of a straightforward application of the Grassmann formula, has particular interest. Proposition 3.6. Let A, B ⊂ Pn be disjoint finite sets and set Z = A ∪ B. Then for any d, the spans of vd(A) and vd(B) satisfy the following formula: dim(hvd(A)i ∩ hvd(B)i) + 1 = h1 Z(d). Proof. See [5], Sect. 6. (cid:3) Next, we define the Cayley-Bacharach property for a finite set in a projective space. Definition 3.7. A finite set Z ⊂ Pn satisfies the Cayley-Bacharach property in degree d, abbreviated as CB(d), if, for any P ∈ Z, every form of degree d vanishing at Z \ {P } also vanishes at P . Example 3.8. Let Z be a set of 6 points in P2. If the 6 points are general, then DhZ(i) = i + 1 for i ∈ {0, 1, 2}. Moreover Z satisfies CB(1). Since hZ(2) = 6, Z does not satisfy CB(2). If Z is a general subset of an irreducible conic, then DhZ(2) = 2 and DhZ(3) = 1, moreover Z satisfies CB(2), hence it satisfies CB(1). If Z has 5 points on a line plus one point off the line, then DhZ is given by the following table: d DhZ (d) 0 1 1 2 2 3 1 1 4 5 1 0 . . . . . . Moreover Z does not satisfy CB(1). 6 E. ANGELINI, L. CHIANTINI, AND A. MAZZON Remark 3.9. If Z satisfies CB(d), then it satisfies CB(d − 1) too. Otherwise, one could find P ∈ Z and a hypersurface F ⊂ Pn of degree d − 1 such that Z \ {P } ⊂ F and P /∈ F . Therefore, if H ⊂ Pn is a hyperplane missing P , then F ∪ H contains Z \ {P } and misses P , contradicting the CB(d) hypothesis. The main reason to introduce the Cayley-Bacharach properties relies in the fol- lowing observation. Proposition 3.10. Z satisfies CB(d) if and only if for all P ∈ Z and for all j ≤ d we have hZ(j) = hZ\{P }(j). I.e. Z satisfies CB(d) if and only if DhZ(j) = DhZ\{P }(j) ∀j ≤ d. Proof. Fix coordinates for the points of Z and consider the corresponding evaluation map C[x0, . . . , xn] → Cℓ(Z) in degree j. The kernel is the space of forms of degree j vanishing on Z. Since Z satisfies CB(j) for all j ≤ d, then, for all P ∈ Z, every form of degree j ≤ d that vanishes on Z \ {P } also vanishes on Z. Thus, in any degree j ≤ d the kernel of the evaluation map does not change if we forget any point P ∈ Z. Hence hZ(j) = hZ\{P }(j). (cid:3) Remark 3.11. We will use the previous proposition as follows. Assume that Z does not satisfy CB(d). Then there exists a points P ∈ Z and an integer j0 ≤ d such that DhZ(j0) > DhZ\{P }(j0). Since P DhZ(i) = ℓ(Z) = ℓ(Z \ {P }) + 1 = P DhZ\{P }(i), it follows that DhZ (i) = DhZ\{P }(i) for all i 6= j0. In particular, the equality holds for i ≥ d + 1. Thus there exists P ∈ Z such that h1 Z(d) = ∞ X i=d+1 DhZ (i) = ∞ X i=d+1 DhZ\{P }(i) = h1 Z\{P }(d). The following proposition, which gives a strong constraint on the Hilbert func- tion of sets with a Cayley-Bacharach property, is a refinement of a result due to Geramita, Kreuzer, and Robbiano (see Corollary 3.7 part (b) and (c) of [13]). Theorem 3.12. If a finite set Z ⊂ Pn satisfies CB (i), then for any j such that 0 ≤ j ≤ i + 1 we have DhZ(0) + DhZ (1) + · · · + DhZ(j) ≤ DhZ (i + 1 − j) + · · · + DhZ(i + 1). Proof. See Theorem 4.9 of [5]. (cid:3) 4. Kruskal's criterion for symmetric tensors The symmetric version of the Kruskal's criterion for the identifiability of tensors can be found in [10]. We recall it for the reader's convenience. Theorem 4.1. (Kruskal's criterion) Let T be a form of degree d and let A be a minimal decomposition of T , with ℓ(A) = r. Fix a partition a, b, c of d and call ka, kb, kc the Kruskal's ranks of va(A), vb(A), vc(A) respectively. If: r ≤ ka + kb + kc − 2 2 , then T has rank r and it is identifiable. IDENTIFIABILITY FOR A CLASS OF SYMMETRIC TENSORS 7 5. The extension In this section, we prove the main result, i.e. an extension of the Kruskal's criterion for the case of septics in 3 variables. Let T be a septic and consider a minimal decomposition of T with 11 elements A = {P1, . . . , P11}. We assume that the points of A are general, and more precisely: • the Kruskal's rank of A is 3, i.e. no three points of A are aligned; • the Kruskal's rank of v3(A) is 10, i.e. no 10 points of A are contained in a cubic or, equivalently, for any subset A′ ⊂ A with ℓ(A′) ≤ 10, the set v3(A′) is linearly independent. With these assumptions, we get: Theorem 5.1. T has rank 11 and it is identifiable. Notice that the original Kruskal's criterion 4.1 does not cover this case. Namely if we take a partition 7 = 3+3+1, then under our hypothesis the given decomposition A of T satisfies k3 = 10, k1 = 3, but: 11 6≤ 10 + 10 + 3 − 2 2 . A similar computation holds for any other partition of 7. Indeed in section 4 of [10] it is pointed out that the partition 7 = 3 + 3 + 1 is the one that covers the widest range for the application of Kruskal's criterion. We will prove the extension of the Kruskal's criterion by contradiction. So, from now on consider the set Z = A ∪ B, where B is a second minimal decomposition of T with at most 11 points. Since Z contains A, then the difference DhZ of the Hilbert function of Z is: j DhZ(j) 0 1 1 2 4 2 3 8 3 4 a4 a5 a6 a7 a8 7 5 6 . . . . . . with a8 > 0, since T ∈ hv7(A)i ∩ hv7(B)i and A, B are minimal, thus the points of v7(Z) can not be linearly independent. It follows that h1 Z(7) > 0, as in Proposition 3.6. Hence also a4, a5, a6, a7 > 0, by Proposition 3.4. Proposition 5.2. Z can not satisfy CB(7). Proof. Assume that Z satisfies CB(7). Then, by Proposition 3.10, a5+a6+a7+a8 ≥ 1 + 2 + 3 + 4 = 10. Since ℓ(Z) ≤ 22, this implies a4 ≤ 2. But then, by Proposition 3.4, a5 + a6 + a7 + a8 ≤ 2 + 2 + 2 + 2 = 8, a contradiction. (cid:3) Next, we need a result which is the application of the argument of Remark 3.11 to the existence of two decompositions of a tensor T . Since we will use it in several contests, we prove it in a general setting. Lemma 5.3. Let U be a symmetric tensor and consider two minimal decomposi- tions A1, A2 of U . Set W = A1 ∪ A2. If A1 ∩ A2 = ∅, then W has the Cayley- Bacharach property CB(d). Proof. Suppose that CB(d) does not hold for W . Hence there is a point P ∈ W and a and a hypersurface F ⊂ Pn of degree d such that F contains W \ {P } but misses P . Just as in Remark 3.11, we get that: W (d) = h1 h1 W \{P }(d). 8 E. ANGELINI, L. CHIANTINI, AND A. MAZZON We do not know in principle if P belongs either to A1 or to A2. In any case, by Proposition 3.6: dimhvd(A1)i ∩ hvd(A2)i = h1 W (d) − 1 = h1 W \{P }(d) − 1 = dimhvd(A1 \ {P })i ∩ hvd(A2 \ {P })i. Thus U belongs to both vd(A1 \ {P }) and vd(A2 \ {P }). If P ∈ A1, we get a contradiction with the minimality of A1. Similarly, if P ∈ A2, we get a contradiction with the minimality of A2. (cid:3) As a consequence of Lemma 5.3 and Proposition 5.2, applied to the decomposi- tions A, B of T , in our setting we find: Proposition 5.4. The intersection A ∩ B is non empty. So we must have ℓ(A ∩ B) = i > 0. After rearranging the points of A, B, we may assume A = {P1, . . . , Pi, Pi+1, . . . , P11} B = {P1, . . . , Pi, P ′ i+1, . . . , P ′ i+1, . . . , P ′ q} is disjoint from A, q}, where q = ℓ(B) ≤ 11, i < q, and the set B0 = {P ′ i.e. B0 = B \ A. Then we know that Z = A ∪ B0 has not the property CB(7). For a choice of the scalars, we can write: a1v7(P1) + · · · + a11v7(P11) = T = b1v7(P1) + · · · + biv7(Pi) + bi+1v7(P ′ i+1) + · · · + bqv7(P ′ q). The minimality of A, B implies that none of the coefficients ai, bi is 0. Write: T0 = (a1 − b1)v7(P1) + · · · + (ai − bi)v7(Pi) + ai+1v7(Pi+1) + · · · + a11v7(P11) = bi+1v7(P ′ i+1) + · · · + bqv7(P ′ q) so that T = T0 + b1v7(P1) + · · · + biv7(Pi). We are now ready to prove that: Proposition 5.5. The existence of B yields a contradiction. Proof. Since Z = A ∪ B0 does not satisfy CB(7), as in the proof of Lemma 5.3, we know that there exists a point P ∈ Z such that hv7(A \ {P })i ∩ hv7(B0 \ {P })i = hv7(A)i ∩ hv7(B0)i, hence T0 belongs to hv7(A \ {P })i ∩ hv7(B0 \ {P })i. If P belongs to B0, then T is spanned by v7(B \{P }), which contradicts the minimality of B. Similarly, if P = Pj with j > i, then T is spanned by v7(A \ {P }), which contradicts the minimality of A. Thus P is a point in A ∩ B, and, after rearranging, we may assume P = P1. Thus T0 = c2v7(P2) + · · · + c11v7(P11). If a1 − b1 6= 0, this implies that v7(A) is linearly dependent, which contradicts the minimality of A. Thus a1 = b1, so that T0 is spanned by v7(P2), . . . , v7(P11). It may happen that some other coefficient (ai − bi) is 0. Nevertheless T0 has a minimal decomposition A′ ⊂ A whose length ℓ(A′) satisfies by 11 − i ≤ ℓ(A′) ≤ 10, namely: T0 = (a2 − b2)v7(P2) + · · · + (ai − bi)v7(Pi) + ai+1v7(Pi+1) + · · · + a11v7(P11). Moreover T0 has a decomposition, B0, of length ℓ(B0) = q − i ≤ 11 − i ≤ ℓ(A′). Since A′ consists of at most 10 points of A, then v3(A′) is linearly independent. IDENTIFIABILITY FOR A CLASS OF SYMMETRIC TENSORS 9 Thus, the Kruskal's rank k′ 1 of A′ is min{ℓ(A′), 3}. Moreover ℓ(A′) > 1, otherwise i = 10, q = 11 and T0 = a11v7(P11) = b11v7(P ′ 11, a contradiction. Thus one computes: 11), which means that, as projective points, P11 = P ′ 3 of v3(A′) is ℓ(A′), while the Kruskal's rank k′ ℓ(A′) ≤ k′ 3 + k′ 3 + k′ 2 1 − 2 . It follows that the Kruskal's criterion contradicts the existence of the two different decompositions of T0. (cid:3) Example 5.6. To give an example, consider the linear forms L1, . . . , L11 in 3 variables associated to the points (1, 0, 0), (0, 1, 0), (0, 0, 1), (1, 1, 1), (1, −1, 2), (1, 3, −1), (1, 2, 3), (2, −1, 1), (−2, −1, 3), (−1, 3, 4), (3, −1, 4). One computes easily that the set A of the points has maximal Kruskal's ranks k1 = 3 and k3 = 10. It turns out that any linear combination of the 7th powers of the forms Li's, with no zero coefficients, has rank 11 and it is identifiable. For instance, this holds for the sum T = L7 11, i.e. 1 + · · · + L7 y y + 45x 3 4 3 T = 2191x 7 − 849x 6 y + 249x 5 2 y − 51x 4 + 501x 2 5 y + 69xy 6 + 4500y 7 + 6 3181x z − 918x 5 yz + 430x 4 2 y z − 204x 3 3 y z + 346x 2 4 y z − 1128xy 5 z + 2390y 6 z + 3631x 5 2 z −1390x 4 2 yz + 274x 3 2 2 z y + 344x 2 2 y3z − 1034xy 4 2 z + 4390y 5 2 z + 5731x 4 3 z − 1668x 3 3 yz +1372x 2 2 3 z y − 1686xy 3 3 z + 5636y 4 3 z + 6115x 3 4 z − 1714x 2 4 yz − 1346xy 2 4 z +7234y 3 4 z + 11491x 2 5 z − 5208xyz 5 + 11480y 2 5 z + 7531xz 6 + 8860yz 6 + 37272z 7 . 6. Decompositions on cubics In this section we will prove results on the identifiability of a particular set of tensors. From now on, consider symmetric tensors T of degree 7 + 2q in three variables, which satisfy the following assumptions: • Fix the integer q ≥ 0 and let T be a symmetric tensor of degree d = 7 + 2q in 3 variables, with a decomposition A = {P1, . . . , Pr} ⊂ P2, of length r = ℓ(A) ≤ 10 + 3q such that A is contained in a plane cubic curve C. • Assume that the Kruskal's rank of A is k1 = min{3, r} and the (q + 3)-th Kruskal's rank kq+3 of A is equal to min{r, 3q + 9}. The second assumption implies that no 3 points of A are aligned. Remark 6.1. Assume that r = ℓ(A) is smaller than 3q + 10. Then the second assumption implies that the (q + 3)-th Kruskal's rank of A is r. Take the partition d = 2q + 7 = (q + 3) + (q + 3) + 1. Then r satisfies: r ≤ kq+3 + kq+3 + k1 − 2 2 , thus by Kruskal's Theorem 4.1, we get that T has rank r and A is the unique decomposition of T (up to trivialities). Thus, from now on we will assume that A has cardinality 3q + 10. Notice that, when r = 3q +10, the Kruskal's criterion gives no information (e.g. for the partition d = 2q + 7 = (q + 3) + (q + 3) + 1, we have 3q + 10 > ((3q + 9) + (3q + 9) + 3 − 2)/2). Hence we need to apply another strategy. 10 E. ANGELINI, L. CHIANTINI, AND A. MAZZON Remark 6.2. Assume r = 3q + 10. Since A is contained in a cubic curve, the difference of the Hilbert function of A satisfies DhA(0) = 1, DhA(1) ≤ 2, DhA(2) ≤ 3 and also DhA(3) ≤ 3. Thus DhA(i) ≤ 3 for i > 3, by Proposition 3.4. Assume that DhA(1) ≤ 1. Then hA(1) < 3, which contradicts the assumption k1 = 3. Thus DhA(1) = 2. Assume that DhA(i) < 3 for some i in the range 2 ≤ i ≤ q + 3. Then: hA(q + 3) = q+3 X j=0 DhA(j) ≤ 1 + 2 + 3(q + 1) + 2 < 3q + 9, which contradicts the assumption kq+3 = 3q+9. Thus DhA(i) = 3 for i = 2, . . . , q+ 3. It follows that Pq+3 otherwise DhA(j) = 0 for j ≥ q + 4, by Proposition 3.4, hence P∞ 3q + 10, a contradiction. j > q + 4. j=0 DhA(j) = 3q + 9. Moreover, DhA(q + 4) can not be 0, j=0 DhA(j) < It follows that DhA(q + 4) = 1 and DhA(j) = 0 for In particular, we get hA(2) = 6 and hA(3) = 9, which means that A is contained in no conics and in exactly one cubic curve. The function DhA is thus the one displayed below. i 0 DhA(i) 1 1 2 2 3 3 . . . 3 . . . q + 2 q + 3 q + 4 q + 5 . . . . . . 3 0 3 1 We want to exclude the existence of a second decomposition B of T of length ℓ(B) ≤ 3q + 10. So assume, by contradiction, that B exists, and assume, as above, that B is minimal. Define as above Z = A ∪ B, so that ℓ(Z) ≤ 6q + 20. Using Lemma 5.3, we can prove that the intersection A ∩ B can not be empty. Proposition 6.3. The Cayley-Bacharach property CB(d) can not hold for Z = A ∪ B. Thus A ∩ B 6= ∅. Proof. Assume that CB(d) holds for Z = A ∪ B. Thus, by Theorem 3.12 we have: q+3 X j=0 DhZ(j) ≤ d+1 X j=q+5 DhZ(j). Since Pq+3 j=0 DhZ(j) ≥ Pq+3 j=0 DhA(j) = 3q + 9, we get: 6q + 20 ≥ ℓ(Z) ≥ 2q+8 X j=0 DhZ (j) = q+3 X j=0 DhZ(j) + DhZ(q + 4) + 2q+8 X j=q+5 DhZ(j) ≥ 3q + 9 + DhZ (q + 4) + 3q + 9, thus DhZ (q + 4) ≤ 2. But then, by Proposition 3.4, DhZ(j) ≤ 2 for j ≥ q + 4, thus: 3q + 9 ≤ 2q+8 X j=q+5 DhZ (j) ≤ 2(q + 4), a contradiction. The second claim follows by Lemma 5.3 applied to A, B and T . (cid:3) Now we can prove the main results of this section. Proposition 6.4. The existence of B yields a contradiction. IDENTIFIABILITY FOR A CLASS OF SYMMETRIC TENSORS 11 Proof. Suppose there is another decomposition B of T , of cardinality ℓ(B) = k ≤ 10 + 3q. We know from Proposition 6.3 that A ∩ B 6= ∅. Thus we can write, without loss of generality, B = {P1, . . . , Pi, P ′ r}, i.e. we may assume that A ∩ B = {P1, . . . , Pi}, i > 0. Then there are coefficients a1, . . . , a3q+10, b1, . . . , bk such that: i+1, . . . , P ′ T = a1vd(P1) + · · · + aivd(Pi) + ai+1vd(Pi+1) · · · + a3q+10vd(P3q+10) = b1vd(P1) + · · · + bivd(Pi) + bi+1vd(P ′ i+1) + · · · + bkvd(P ′ k). Consider the tensor T0 = (a1 − b1)vd(P1) + · · · + (ai − bi)vd(Pi) + ai+1vd(Pi+1) + · · · + a3q+10vd(P3q+10), i+1) + · · · + bkvd(P ′ which is also equal to bi+1vd(P ′ k). Thus T0 has the two decom- k}, which are disjoint. Thus, if A and B′ are positions A and B′ = {P ′ both minimal, as decompositions of T0, then by Lemma 5.3 applied to A, B′ and T0, we get that A ∪ B′ satisfies CB(d). Since A ∪ B′ = A ∪ B = Z, and we know by Proposition 6.3 that Z does not satisfies CB(d), we find that either A or B′ are not minimal. i+1, . . . , P ′ Assume that B′ is not minimal. Then we can find a point of B′, say P ′ T0 belongs to the span of vd(B′ \ {P ′ this would mean that T belongs to the span of vd(B \ {P ′ minimality of B. k, such that k}). Since T = T0 + b1vd(P1) + · · · + bivd(Pi), k}), which contradicts the Assume that A is not minimal, and T0 belongs to the span of vd(A \ {Pj}), for some j > i. That is, as above, since T = T0 + b1vd(P1) + · · · + bivd(Pi), this would mean that T belongs to the span of vd(A \ {Pj}), which contradicts the minimality of A. Assume that A is not minimal, and T0 belongs to the span of vd(A \ {Pj}), for some j ≤ i, say j = 1. Then T0 = γ2vd(P2) + · · · + γ3q+10vd(P3q+10), for some choice of the coefficients γj. Since vd(A) is linearly independent, because A is minimal for T , this is only possible if a1 − b1 = 0. So there exists a proper subset A′ ⊂ A which provides a minimal decomposition of T0, together with B′. Moreover ℓ(A′) ≥ ℓ(B′). Since A has (q + 3)-th Kruskal's rank 3q + 9 ≥ ℓ(A′), by Remark 2.5 q+3 of A′ is ℓ(A′). Similarly, the Kruskal's rank k′ the (q + 3)-th Kruskal's rank k′ 1 of A′ is min{3, ℓ(A′)}. Moreover ℓ(A′) > 1, otherwise i = 3q + 9, k = 3q + 10 and T0 = a3q+10vd(P3q+10) = b3q+10vd(P ′ 3q+10), which means that, as projective points, P3q+10 = P ′ 3q+10, a contradiction. Thus one computes: ℓ(A′) ≤ k′ q+3 + k′ q+3 + k′ 2 1 − 2 . hence the existence of a second decomposition B′ of T0, with ℓ(A′) ≥ ℓ(B′), con- tradicts Kruskal's Theorem 4.1. (cid:3) Remark 6.5. We can apply verbatim the previous procedure even for the case q = −1. We get ternary forms of degree 7 + 2q = 5 and rank r ≤ 10 + 3q = 7. In particular, notice that 7 is the rank of a generic quintic form, by [1]. Notice also that, for q = −1, the assumption that A is contained in a cubic is unnecessary, for any set of cardinality ≤ 7 in P2 lies in a cubic. Thus, we get back, from our procedure, the classically well known fact that a general ternary form T of degree 5 has a unique decomposition with 7 powers of linear forms. 12 E. ANGELINI, L. CHIANTINI, AND A. MAZZON Indeed, we can give a more precise notion of generality for T : the uniqueness holds if a decomposition A of T has Kruskal's ranks k1 = 3 and k2 = 6. References 1. J. Alexander and A. Hirschowitz, Polynomial interpolation in several variables, J. Algebraic Geom. 4 (1995), 201 -- 222. 2. E.S. Allman, C. Matias, and J.A. Rhodes, Identifiability of parameters in latent structure models with many observed variables, Ann. Statistics 37 (2009), 3099 -- 3132. 3. A. Anandkumar, R. Ge, D. Hsu, S.M. Kakade, and M. Telgarsky, Tensor decompositions for learning latent variable models, J. Machine Learn. Res. 15 (2014), 2773 -- 2832. 4. E. Angelini, Waring decompositions and identifiability via Bertini and Macaulay2 software, J. Symbolic Comput. 91 (2019), 200 -- 212. 5. E. Angelini, L. Chiantini, and N. Vannieuwenhoven, Identifiability beyond Kruskal's bound for symmetric tensors of degree 4, Rend. Lincei Mat. Applic. 29 (2018), 465 -- 485. 6. C.J. Appellof and E.R. Davidson, Strategies for analyzing data from video fluorometric mon- itoring of liquid chromatographic effluents, Anal. Chem. 53 (1981), 2053 -- 2056. 7. E. Ballico and L. Chiantini, A criterion for detecting the identifiability of symmetric tensors of size three, Diff. Geom. Applic. 30 (2012), 233 -- 237. 8. , Sets computing the symmetric tensor rank, Mediterranean J. Math. 10 (2013), 643 -- -654. 9. L. Chiantini, Hilbert functions and tensor analysis, Available online arXiv:1807.00642, 2018. 10. L. Chiantini, G. Ottaviani, and N. Vannieuwenhoven, Effective criteria for specific identifia- bility of tensors and forms, SIAM J. Matrix Anal. Appl. 38 (2017), 656 -- 681. 11. , On generic identifiability of symmetric tensors of subgeneric rank, Trans. Amer. Math. Soc. 369 (2017), 4021 -- 4042. 12. H. Dersken, Kruskal's uniqueness inequality is sharp, Linear Alg. Applic. 438 (2013), 708 -- 712. 13. A.V. Geramita, M. Kreuzer, and L. Robbiano, Cayley-Bacharach schemes and their canonical modules, Trans. Amer. Math. Soc. 339 (1993), 443 -- 452. 14. J. Hauenstein, L. Oeding, G. Ottaviani, and A.J. Sommese, Homotopy techniques for tensor decomposition and perfect identifiability, J. Reine Angew. Math. (Crelles Journal) (2016), Available online https://doi.org/10.1515/crelle-2016-0067. 15. J.B. Kruskal, Three-way arrays: rank and uniqueness of trilinear decompositions, with appli- cation to arithmetic complexity and statistics, Linear Algebra Appl. 18 (1977), 95 -- 138. 16. A. Massarenti, M. Mella, and G. Staglian´o, Effective identifiability criteria for tensors and polynomials, J. Symbolic Comput. 87 (2018), 227 -- 237. Elena Angelini. Dipartimento di Ingegneria dell'Informazione e Scienze Matem- atiche, Universit`a di Siena, Italy E-mail address: [email protected] Luca Chiantini. Dipartimento di Ingegneria dell'Informazione e Scienze Matematiche, Universit`a di Siena, Italy E-mail address: [email protected] Andrea Mazzon. Dipartimento di Ingegneria dell'Informazione e Scienze Matem- atiche, Universit`a di Siena, Italy E-mail address: [email protected] 

1707.02602

2

1707

The stringy Euler number of Calabi-Yau hypersurfaces in toric varieties and the Mavlyutov duality

We show that minimal models of nondegenerated hypersufaces defined by Laurent polynomials with a $d$-dimensional Newton polytope $\Delta$ are Calabi-Yau varieties $X$ if and only if the Fine interior of $\Delta$ consists of a single lattice point. We give a combinatorial formula for computing the stringy Euler number of $X$. This formula allows to test mirror symmetry in cases when $\Delta$ is not a reflexive polytope. In particular we apply this formula to pairs of lattice polytopes $(\Delta, \Delta^{\vee})$ that appear in the Mavlyutov's generalization of the polar duality for reflexive polytopes. Some examples of Mavlyutov's dual pairs $(\Delta, \Delta^{\vee})$ show that the stringy Euler numbers of the corresponding Calabi-Yau varieties $X$ and $X^{\vee}$ may not satisfy the expected topological mirror symmetry test: $e_{\rm st}(X) = (-1)^{d-1} e_{\rm st}(X^{\vee})$. This shows the necessity of an additional condition on Mavlyutov's pairs $(\Delta, \Delta^\vee)$.

math.AG

math

THE STRINGY EULER NUMBER OF CALABI-YAU HYPERSURFACES IN TORIC VARIETIES AND THE MAVLYUTOV DUALITY VICTOR BATYREV Dedicated to Yuri Ivanovich Manin on the occasion of his 80-th birthday Abstract. We show that minimal models of nondegenerated hypersufaces de- fined by Laurent polynomials with a d-dimensional Newton polytope ∆ are Calabi- Yau varieties X if and only if the Fine interior of the polytope ∆ consists of a single lattice point. We give a combinatorial formula for computing the stringy Euler number of such Calabi-Yau variety X via the lattice polytope ∆. This for- mula allows to test mirror symmetry in cases when ∆ is not a reflexive polytope. In particular, we apply this formula to pairs of lattice polytopes (∆, ∆∨) that ap- pear in the Mavlyutov's generalization of the polar duality for reflexive polytopes. Some examples of Mavlyutov's dual pairs (∆, ∆∨) show that the stringy Euler numbers of the corresponding Calabi-Yau varieties X and X ∨ may not satisfy the expected topological mirror symmetry test: est(X) = (−1)d−1est(X ∨). This shows the necessity of an additional condition on Mavlyutov's pairs (∆, ∆∨). 1. Introduction Many examples of pairs of mirror symmetric Calabi-Yau manifolds X and X ∗ can be obtained using Calabi-Yau hypersurfaces in Gorenstein toric Fano varieties corresponding to pairs (∆, ∆∗) of d-dimensional reflexive lattice polytopes ∆ and ∆∗ that are polar dual to each other [Bat94]. A d-dimensional convex polytope ∆ ⊂ Rd is called a lattice polytope if ∆ = Conv(∆ ∩ Zd), i. If a d- dimensional polytope ∆ contains the zero 0 ∈ Zd in its interior, one defines the polar polytope ∆∗ ⊂ Rd as e., all vertices of ∆ belong to the lattice Zd ⊂ Rd. ∆∗ := {y ∈ Rd : hx, yi ≥ −1 ∀x ∈ ∆}, where h∗, ∗i : Rd × Rd → R is the standard scalar product on Rd. A d-dimensional lattice polytope ∆ ⊂ Rd containing 0 in its interior is called reflexive if the polar polytope ∆∗ is also a lattice polytope. If ∆ is reflexive then ∆∗ is also reflexive and one has (∆∗)∗ = ∆. For a d-dimensional reflexive polytope ∆ one considers the familiy of Laurent polynomials f (x) = Xm∈Zd∩∆ amxm ∈ C[x±1 1 , . . . , x±1 d ] with sufficiently general coefficients am ∈ C. Using the theory of toric varieties [Ful93, CLS11]), one can prove that the affine hypersurface Z ⊂ Td := (see e.g. (C∗)d defined by f = 0 is birational to a (d − 1)-dimensional Calabi-Yau variety 1 2 VICTOR BATYREV X. In the same way one obtains another (d − 1)-dimensional Calabi-Yau variety X ∗ corresponding to the polar reflexive polytope ∆∗. The polar duality ∆ ↔ ∆∗ between the reflexive polytopes ∆ and ∆∗ defines a duality between their proper faces Θ ↔ Θ∗ (Θ ≺ ∆, Θ∗ ≺ ∆∗) satisfying the condition dim Θ + dim Θ∗ = d − 1, where the dual face Θ∗ ≺ ∆∗ is defined as Θ∗ := {y ∈ ∆∗ : hx, yi = −1 ∀x ∈ Θ}. There is a simple combinatorial formula for computing the stringy Euler number of the Calabi-Yau manifold X [BD96, Corollary 7.10]: (1.1) estr(X) = d−2Xk=1 (−1)k−1 XΘ(cid:22)∆ dim Θ=k v(Θ) · v (Θ∗) , where v(P ) := (dim P )!V ol(P ) ∈ Z denotes the integral volume of a lattice polytope P . An alternative proof of the formula (1.1) together with its generalizations for Calabi-Yau complete intersections is contained in [BS17]. The formula (1.1) and the duality Θ ↔ Θ∗ between faces of reflexive polytopes ∆ and ∆∗ immediately imply the equality estr(X) = (−1)d−1estr(X ∗) which is a consequence of a stronger topological mirror symmetry test for the stringy Hodge numbers [BB96]: str (X) = hd−1−p,q hp,q (X ∗), 0 ≤ p, q ≤ d − 1. str It is important to mention another combinatorial mirror construction suggested by Berglung and Hübsch [BHü93] and generalized by Krawitz [Kra09]. This mir- ror construction considers (d − 1)-dimensional Calabi-Yau varieties X which are birational to affine hypersurfaces Z ⊂ Td defined by Laurent polynomials f (x) = Xm∈Zd∩∆ amxm ∈ C[x±1 1 , . . . , x±1 d ], whose Newton polytope ∆ ⊂ Rd is a lattice simplex, but it is important to stress that this simplex may be not a reflexive simplex. The mirror duality for the stringy (or orbifold) Hodge numbers of Calabi-Yau varieties obtained by Berglung-Hübsch- Krawitz mirror construction was proved by Chiodo and Ruan [CR11] and Borisov [Bor13]. The Batyrev mirror construction [Bat94] and the Berglund-Hübsch-Krawitz mir- ror construction [BHü93, Kra09] can be applied to different classes of hypersurfaces in toric varieties, but they coincide for Calabi-Yau hypersurfaces of Fermat-type. So it is natural to expect that there must be a generalization of two mirror constructions that includes both as special cases (see [AP15, Bor13, ACG16, Pum11, BHü16]). Moreover, it is natural to expect the existence of a generalization of combinatorial formula (1.1) for the stringy Euler number estr(X) of Calabi-Yau varieties X which holds true for projective varieties coming from a wider class of nondegenerate affine hypersurfaces Z ⊂ Td defined by Laurent polynomials. Recall that the stringy Euler number estr(X) can be defined for an arbitrary n- dimensional normal projective Q-Gorenstein variety X with at worst log-terminal THE STRINGY EULER NUMBER AND THE MAVLYUTOV DUALITY 3 singularities using a desingularization ρ : Y → X whose exceptional locus is a union of smooth irreducible divisors D1, . . . , Ds with only normal crossings [Bat98]. For this purpose, one sets I := {1, . . . , s}, D∅ := Y and for any nonempty subset J ⊆ I one defines DJ :=Tj∈J . Using the rational coefficients a1, . . . , as from the formula KY = ρ∗KX + one defines the stringy Euler number aiDi, sXi=1 estr(X) := X∅⊆J⊆I aj + 1! . e(DJ )Yj∈J −aj Using methods of a nonarchimedean integration (see e.g. [Bat98, Bat99]), one can show that estr(X) is independent of the choice of the desingularization ρ : Y → X and one has estr(X) = estr(X ′) if two projective Calabi-Yau varieties X and X ′ with at worst canonical singularities are birational. More generally, the stringy Euler number estr(X) of any minimal projective algebraic variety X does not depend on the choice of this model and coincides with the stringy Euler number of its canonical model, because all these birational models are K-equivalent to each other. There exist some versions of the stringy Euler number that are conjectured to have minimum exactly on minimal models in a given birational class [BG18]. We remark that in general the stringy Euler number may not be an integer, and so far no example of mirror symmetry is known if the stringy Euler number estr(X) of a Calabi-Yau variety X is not an integer. In Section 2 we give a review of results of Ishii [Ish99] on minimal models of nondegenerate hypersurfaces and give a combinatorial criterion that describes all d-dimensional lattice polytopes ∆ such that minimal models of ∆-nondegenerate hypersurfaces Z ⊂ Td are Calabi-Yau varieties. We show that a ∆-nondegenerate hypersurface Z ⊂ Td is birational to a Calabi-Yau variety X with at worst Q- factorial terminal singularities if and only if the Fine interior ∆F I of its Newton polytope ∆ consists of a single lattice point (Theorem 2.26). We remark that there exist many d-dimensional lattice polytopes ∆ such that ∆F I = 0 ∈ Zd which are not reflexive if d ≥ 3. In Section 3 we discuss the generalized combinatorial duality suggested by Mav- lyutov in [Mav11] . Lattice polytopes ∆ that appear in the Mavlyutov duality satisfy not only the condition ∆F I = 0, but also the additional condition [[∆∗]∗] = ∆, where P ∗ denotes the polar polytope of P and [P ] denotes the convex hull of all lattice points in P . The lattice polytopes ∆ with ∆F I = 0 satisfying the condition [[∆∗]∗] = ∆ we call pseudoreflexive. A lattice polytope ∆ with ∆F I = 0 may not be a pseudoreflexive, but its Mavlyutov dual polytope ∆∨ := [∆∗] and the lattice polytope [[∆∗]∗] are always pseudoreflexive. Moreover, if ∆F I = 0 then [[∆∗]∗] is the smallest pseudoreflexive polytope containing ∆. For this reason we call d-dimensional lattice polytopes ∆ with the only condition ∆F I = 0 almost pseudoreflexive. If the lattice polytope ∆ is pseudoreflexive, then one has ((∆∨)∨) = ∆. Any reflexive polytope ∆ is pseudoreflexive, because in this case one has ∆∨ = [∆∗] = ∆∗. 4 VICTOR BATYREV Therefore the Mavlyutov duality ∆ ↔ ∆∨ is a generalization of the polar duality ∆ ↔ ∆∗ for reflexive polytopes. Unfortunately Mavlyutov dual pseudoreflexive polytopes ∆ and ∆∨ are not neces- sarily combinatorially dual to each other. For this reason we can not expect a natural duality between k-dimensional faces of pseudoreflexive polytope ∆ and (d − 1 − k)- dimensional faces of its dual pseudoreflexive polytope ∆∨. Mavlyutov observed that a natural duality can be obtained if ones restricts attention to some part of faces of ∆ [Mav13]. A proper k-dimensional face Θ ≺ ∆ of a pseudoreflexive polytope ∆ will be called regular if dim[Θ∗] = d − k − 1, where Θ∗ is the dual face of the polar polytope ∆∗. If Θ ≺ ∆ is a regular face of a pseudoreflexive polytope ∆ then Θ∨ := [Θ∗] is a regular face of the Mavlyutov dual pseudoreflexive polytope ∆∨ and one has (Θ∨)∨ = Θ, so that one obtains a natural duality between k-dimensional regular faces of ∆ and (d −k −1)-dimensional regular faces of ∆∨. Mavlyutov hoped that this duality could help to find a mirror symmetric generalization of the formula (1.1) for arbitrary pairs (∆, ∆∨) of pseudoreflexive polytopes [Mav13]. In Section 4 we are interested in a combinatorial formula for the stringy E-function Estr(X; u, v) of a canonical Calabi-Yau model X of a ∆-nondegenerated hypersurface for an arbitrary d-dimensional almost pseudoreflexive polytope ∆. Using the results of Danilov and Khovanskii [DKh86], we obtain such a combinatorial formula for the stringy function Estr(X; u, 1) (Theorem 4.10) and for the stringy Euler number estr(X) := Estr(X; 1, 1) (Theorem 4.11): (1.2) est(X) = d−1Xk=0 XΘ⊆∆ dim Θ=d−k (−1)d−1−kv(Θ) · v(σΘ ∩ ∆∗). In this formula the polar polytope ∆∗ is in general a rational polytope, the integer v(Θ) denotes the integral volume of a (d − k)-dimensional face Θ (cid:22) ∆ and the rational number v(σΘ∩∆∗) denotes the integral volume of the k-dimensional rational polytope σΘ ∩ ∆∗ contained in the k-dimensional normal cone σΘ corresponding to the face Θ (cid:22) ∆ in the normal fan of the polytope ∆. One can easily see that the formula (1.1) can be considered as a particular case of the formula (1.2) if ∆ is a reflexive polytope. In Section 5 we consider examples of Mavlyutov pairs (∆, ∆∨) of pseudoreflex- ive polytopes obtained from Newton polytopes of polynomials defining Calabi-Yau hypersurfaces X of degree a + d in the d-dimensional weighted projective spaces P(a, 1, . . . , 1) of dimension d ≥ 5 such that the weight a does not divide the degree a + d and a < d/2. These pseudoreflexive polytopes ∆ and ∆∨ are not reflexive. If d = ab + 1 for an integer b ≥ 2 then X is quasi-smooth and one can apply the Berglund-Hübsch-Krawitz mirror construction. We compute the stringy Euler numbers of Calabi-Yau hypersurfaces X and their mirrors X ∨. In particular, we show that the equality estr(X) = (−1)d−1estr(X ∨) holds if d = ab + 1 and in this case one obtains quasi-smooth Calabi-Yau hypersurfaces. However, if d = ab + l (2 ≤ l ≤ a − 1), then the Calabi-Yau hypersurfaces X ⊂ P(a, 1, . . . , 1) are not quasi- smooth. Using our formulas for the stringy Euler numbers estr(X) and estr(X ∨) we show that the equality estr(X) = (−1)d−1estr(X ∨) can not be satisfied if e.g. d = ab + 2, where a, b are two distinct odd prime numbers (Theorem 5.5). THE STRINGY EULER NUMBER AND THE MAVLYUTOV DUALITY 5 In Section 6 we investigate the Mavlyutov duality ∆ ↔ ∆∨ together with an additional condition on singular facets of the pseudoreflexive polytopes ∆ and ∆∨. This condition can be considered as a version of a quasi-smoothness condition on Mavlyutov's pairs (∆, ∆∨) suggested in some form by Borisov [Bor13]. For Calabi- Yau varieties X and X ∨ corresponding to Mavlyutov's pairs (∆, ∆∨) satisfying this additional condition we prove another generalization of the formula (1.1) such that the equality estr(X) = (−1)d−1estr(X ∨) holds (Theorem 6.3). Acknowledgements. It is a pleasure to express my thanks to Anvar Mavlyutov, Lev Borisov, Duco van Straten, Harald Skarke, Shihoko Ishii, Makoto Miura, Karin Schaller and Alexander Kasprzyk for useful stimulation discussions. 2. Minimal models of nondegenerate hypersurfaces Let M ∼= Zd be a free abelian group of rank d and MR = M ⊗ R. Denote by NR the dual space Hom(M, R) with the natural pairing h∗, ∗i : MR × NR → R. Definition 2.1. Let P = Conv(x1, . . . , xk) ⊂ MR be a convex polytope obtained as the convex hull of a finite subset {x1, . . . , xk} ⊂ MR. Define the piecewise linear function as ordP : NR → R ordP (y) := min x∈P hx, yi = k min i=1 hxi, yi. We associate with P its normal fan ΣP which is finite collection of normal cones σQ in the dual space NR parametrized by faces Q (cid:22) P . The the cone σQ is defined as σQ := {y ∈ NR : ordP (y) = hx, yi, ∀x ∈ Q}. The zero 0 ∈ NR is considered as the normal cone to P . One has NR = [Q(cid:22)P σQ. If P ⊂ MR is a d-dimensional polytope containing 0 ∈ M in its interior, then we call P ∗ := {y ∈ NR : ordP (y) ≥ −1} the polar polytope of P . The polar polytope P ∗ is the union over all proper faces Q ≺ P of the subsets P ∗ ∩ σQ = {y ∈ σQ : hx, yi ≥ −1, ∀x ∈ Q}, i.e., P ∗ = [Q≺P(cid:16)P ∗ ∩ σQ(cid:17) . Definition 2.2. Let n ∈ N be a primitive lattice vector and let l ∈ Z. We consider an affine hyperplane Hn(l) ⊂ MR defined by the equation hx, ni = l. If m ∈ M then the nonnegative integer is called the integral distance between m and the hyperplane Hn(l). hm, ni − l 6 VICTOR BATYREV Definition 2.3. Let ∆ ⊂ MR be a lattice polytope, i.e., all vertices of ∆ belong to M. Then ord∆ has integral values on N. For any nonzero lattice point n ∈ N one defines the following two half-spaces in MR: Γ∆ Γ∆ 0 (n) := {x ∈ MR : hx, ni ≥ ord∆(n)}, 1 (n) := {x ∈ MR : hx, ni ≥ ord∆(n) + 1}. 1 (n) ⊂ Γ∆ For all n ∈ N we have obvious the inclusion Γ∆ 0 (n) and the lattice polytope ∆ can be written as intersection ∆ = \06=n∈N Γ∆ 0 (n). Definition 2.4. Let ∆ be a d-dimensional lattice polytope. The Fine interior of ∆ is defined as ∆F I := \06=n∈N Γ∆ 1 (n). Remark 2.5. It is clear that ∆F I is a convex subset in the interior of ∆. We remark that the interior of a d-dimensional polytope ∆ is always nonempty, but the Fine interior of ∆ may be sometimes empty. I was told that the subset ∆F I ⊂ ∆ first has appeared in the PhD thesis of Jonathan Fine [Fine83]. Remark 2.6. Since ∆F I is defined as an intersection of countably many half-spaces Γ∆ 1 (n) it is not immediately clear that the polyhedral set ∆F I has only finitely many faces. The latter follows from the fact that for any proper face Θ ≺ ∆ the semigroup SΘ := N ∩ σΘ of all lattice points in the cone σΘ is finitely generated (Gordan's lemma). One can show that ∆F I can be obtained as a finite intersection of those half-spaces Γ∆ 1 (n) such that the lattice vector n appears as a minimal generator of the semigroup SΘ for some face Θ ≺ ∆. Indeed, if n′, n′′ ∈ σΘ, i.e., if two lattice 1 (n′′), vectors n′, n′′ are in the same cone σΘ, and if x ∈ ∆ is a point in Γ∆ then we have hx, n′ + n′′i = hx, n′i + hx, n′′i ≥ ord∆(n′) + 1 + ord∆(n′′) + 1 > ord∆(n′ + n′′) + 1, i. e. , Γ∆ 1 (n′′) is contained in Γ∆ 1 (n′) ∩ Γ∆ 1 (n′) ∩ Γ∆ 1 (n′ + n′′) Remark 2.7. Let ∆ ⊂ MR be a d-dimensional lattice polytope. If m ∈ ∆ is an interior lattice point, then m ∈ ∆F I. Indeed, if m ∈ ∆ is an interior lattice point, then for any lattice point n ∈ N one has hm, ni > ord∆(n). Since both numbers hm, ni and ord∆(n) are integers, we obtain i.e., m belongs to ∆F I. This implies the inclusion hm, ni ≥ ord∆(n) + 1, ∀n ∈ N, Conv(Int(∆) ∩ M) ⊆ ∆F I, i.e., the Fine interior of ∆ contains the convex hull of the set interior lattice points in ∆. We see below that for 2-dimensional lattice polytopes this inclusion is equality. In order to find the Fine interior of an arbitrary 2-dimensional lattice polytope we will use the following well-known fact: THE STRINGY EULER NUMBER AND THE MAVLYUTOV DUALITY 7 Proposition 2.8. Let ∆ ⊂ R2 be a lattice triangle such that ∆ ∩ Z2 consists of vertices of ∆. Then ∆ is isomorphic to the standard triangle with vertices (0, 0), (1, 0), (0, 1). In particular, the integral distance between a vertex of ∆ and its opposite side of ∆ is always 1. Proposition 2.9. If ∆ is a 2-dimensional lattice polytope, then ∆F I is exactly the convex hull of interior lattice points in ∆. Proof. Let ∆ be a 2-dimensional lattice polytope. If ∆ has no interior lattice points, then ∆ is isomorphic to either a lattice polytope in R2 contained in the strip 0 ≤ x1 ≤ 1, or to the lattice triangle with vertices (0, 0), (2, 0), (0, 2) (see e. g. [Kho97]). In both cases one can easily check that ∆F I = ∅. If ∆ has exactly one interior lattice point then ∆ is isomorphic to one of 16 reflexive polygons and one can check that this interior lattice point is exactly the Fine interior of ∆, because, by 2.8, this interior lattice point has integral distance 1 to its sides. If ∆ is a 2-dimensional lattice polytope with at least two interior lattice points then we denote ∆′ := Conv(Int(∆) ∩ M). One has dim ∆′ ∈ {1, 2}. If dim ∆′ = 1 then by 2.8 the integral distance from the affine line L containing ∆′ and any lattice vertex of ∆ outside of this line must be 1. This implies that the Fine interior ∆F I is contained in L. By 2.7, if A und B are two vertices of the segment ∆′ then A, B ∈ ∆F I. By 2.8, there exist a side of ∆ with the integral distance 1 from A having nonempty intersection with the line L. Therefore, A is vertex of ∆F I. Analogously, B is also a vertex of ∆F I. Assume now that dim ∆′ = 2, i.e., ∆′ is k-gon. Then ∆′ is an intersection of k half-planes Γ1, . . . , Γk whose boundaries are k lines L1, . . . , Lk through the k sides of ∆′. By 2.8, for any 1 ≤ i ≤ k all vertices of ∆ outside the half-plane Γi must have integral distance 1 from Li. Therefore Γi contains the Fine interior ∆F I and (cid:3) i=1 Γi = ∆′. The opposite inclusion ∆′ ⊆ ∆F I follows from 2.7. ∆F I ⊆Tk Remark 2.10. The convex hull of all interior lattice points in a lattice polytope ∆ of dimension d ≥ 3 must not coincide with the Fine interior ∆F I in general. For example there exist 3-dimensional lattice polytopes ∆ without interior lattice points such that ∆F I is not empty [Tr08]. The simplest well-known example of such a situation is the 3-dimensional lattice simplex corresponding to Newton polytope of the Godeaux surface obtained as a free cyclic group of order 5 quotient of the Fermat surface w5 + x5 + y5 + z5 = 0 by the mapping (w : x : y : z) → (w : ρx : ρ2y : ρ3z), where ρ is a fifth root of 1. Definition 2.11. Assume that the lattice polytope ∆ has a nonempty Fine interior ∆F I. We define the support of ∆F I as Supp(∆F I) := {n ∈ N : hx, ni = ord∆(n) + 1 for some x ∈ ∆F I} ⊂ N. The convex rational polytope ∆can := \n∈Supp(∆F I ) Γ∆ 0 (n) containing ∆ we call the canonical hull of ∆. 8 VICTOR BATYREV Remark 2.12. The support of ∆F I is a finite subset in the lattice N, because it is contained in the union over all faces Θ ≺ ∆ of all minimal generating subsets for the semigroups N ∩ σΘ. In particular, Supp(∆F I) always consists of finitely many primitive nonzero lattice vectors v1, . . . , vl ∈ N such that Pl i=1 R≥0vi = NR. Let us now identify the lattice M ∼= Zd with the set of monomials xm = xm1 · · · xmd d d ] ∼= C[M]. For simplicity we will 1 in the Laurent polynomial ring C[x±1 consider only algebraic varieties X over the algebraically closed field C. 1 , . . . , x±1 Pm amxm ∈ C[x±1 Definition 2.13. The Newton polytope ∆(f ) of a Laurent polynomial f (x) = d ] is the convex hull of all lattice points m ∈ M such that am 6= 0. For any face Θ ⊆ ∆(f ) one defines the Θ-part of the Laurent polyno- mial f as 1 , . . . , x±1 fΘ(x) := Xm∈M ∩Θ amxm. A Laurent polynomial f ∈ C[M] with a Newton polytope ∆ is called ∆-nondegenerate or simply nondegenerate if for every face Θ ⊆ ∆ the zero locus ZΘ := {x ∈ Td : fΘ(x) = 0} of the Θ-part of f is empty or a smooth affine hypersurface in the d-dimensional algebraic torus Td. The theory of toric varieties allows to construct a smooth projective algebraic desingularization of Z ∆. Zariski open subset. For this pupose one first considers the closure Z ∆ of Z∆ in the projective toric variety P∆ associated with the normal fan Σ∆. Then one chooses a variety bZ∆ that contains the affine ∆-nondegenerate hypersurface Z∆ ⊂ T as a regular simplicial subdivision bΣ of the fan Σ∆ and obtains a projective morphism ρ : PbΣ → P∆ from a smooth toric variety PbΣ to P∆ such that, by Bertini theorem, its restriction to the Zariski closure bZ∆ of Z∆ in PbΣ is a smooth and projective hypersurface bZ∆ (see e.g. [Mat02]). One can show that for ∆-nondegenerate hyper- surfaces a minimal model of bZ∆ can be obtained via the toric Mori theory due to Miles Reid [Reid83, Wi02, Fu03, FS04] applied to pairs (V, D) consisting of a pro- jective toric variety V and the Zariski closure D of the nondegenerate hypersurface Z∆ in V [Ish99]. Therefore, minimal models of nondegenerate hypersurfaces Z∆ can be constructed by combinatorial methods. Now one can apply the Minimal Model Program of Mori to the smooth projective Recall the following standard definitions from the Minimal Model Program [Ko13]. Definition 2.14. Let X be a normal projective algebraic variety over C, and let KX be its canonical class. A birational morphism ρ : Y → X is called a log- desingularization of X if Y is smooth and the exceptional locus of ρ consists of smooth irreducible divisors D1, . . . , Dk with simple normal crossings. Assume that X is Q-Gorenstein, i.e., some integral multiple of KX is a Cartier divisor on X. We set I := {1, . . . , k}, D∅ := Y , and for any nonempty subset J ⊆ I we denote by DJ the intersection of divisors Tj∈J Dj, which is either empty or a smooth projective subvariety in Y of codimension J. Then the canonical classes of X and Y are related by the formula KY = ρ∗KX + aiDi, kXi=1 THE STRINGY EULER NUMBER AND THE MAVLYUTOV DUALITY 9 where a1, . . . , ak are rational numbers which are called discrepancies. The singular- ities of X are said to be log-terminal canonical terminal if if if ai > −1 ∀i ∈ I; ∀i ∈ I; ai ≥ 0 ∀i ∈ I. ai > 0 It is known that Q-Gorenstein toric varieties always have at worst log-terminal singularities [Reid83]. Definition 2.15. A projective normal Q-Gorenstein algebraic variety Y is called canonical model of X if Y is birationally equivalent to X, Y has at worst canonical singularities and the linear system mKY is base point free for sufficiently large integer m ∈ N. Definition 2.16. A projective algebraic variety Y is called minimal model of X if Y is birationally equivalent to X, Y has at worst terminal Q-factorial singularities, the canonical class KY is numerically effective, and the linear system mKY is base point free for sufficiently large integer m ∈ N. The main result of the Minimal Model Program for nondegenerate hypersurfaces in toric varieties in [Ish99] can be reformulated using combinatorial interpretations of Zariski decompositions of effective divisors on toric varieties [OP91] (see also [HKP06, Appendix A]) and their applications to log minimal models of polarized pairs [BiH14] Definition 2.17. Let PΣ be a d-dimensional projective toric variety defined by a fan Σ whose 1-dimensional cones σi = R≥0vi ∈ Σ(1) are generated by primitive lattice vectors v1, . . . , vs ∈ N. Denote by Vi (1 ≤ i ≤ s) torus invariant divisors on i=1 aiVi be an arbitrary torus invariant Q-divisor P corresponding to vi. Let D =Ps on P such that the rational polytope ∆D := {x ∈ MR : hx, vii ≥ −ai, 1 ≤ i ≤ s} is not empty. Then the rational numbers satisfy the inequalities ord∆D(vi) := min x∈∆D hx, vii, 1 ≤ i ≤ r, ord∆D(vi) + ai ≥ 0, 1 ≤ i ≤ r. Without loss of generality we can assume that the equality ord∆D(vi) + ai = 0 holds if and only if 1 ≤ i ≤ r (r ≤ s). We define the support of the polytope ∆D as Supp(∆D) := {v1, . . . , vr} and we write the divisor D =Pl sXi=1 D = P + N, P = i=1 aiVi as the sum (−ord∆D (vi))Vi, N := sXi=r+1 (ord∆D(vi) + ai)Vi, where ord∆D(vi) + ai > 0 for all r + 1 ≤ i ≤ s. Then there exists a d-dimensional Q- factorial projective toric variety P′ defined by a simplicial fan Σ′ whose 1-dimensional cones are generated by the lattice vectors vi ∈ Supp(∆D) together with a birational 10 VICTOR BATYREV toric morphism ϕ : P → P′ that contracts the divisors Vr+1, . . . , Vs such the nef divisor P on P is the pull back of the nef divisor P ′ := Xvi∈Supp(∆D) (−ord∆D(vi))V ′ i on P′. The decomposition D = P + N together with the nef divisor P ′ on the Q-factorial projective toric variety P′ we call toric Zariski decomposition of D. Theorem 2.18. A ∆-nondegenerate hypersuface Z∆ ⊂ Td has a minimal model if and only if the Fine interior ∆F I is not empty. In the latter case, a canonical model of the nondegenerate hypersurface Z∆ is its closure X in the toric variety P∆can associated to the canonical hull ∆can of the lattice polytope ∆. The birational isomorphism between Z ∆ and X is induced by the birational isomorphism of toric varieties α : P∆ 99K P∆can, it can be included in a diagram PbΣ α ρ1 ~⑥⑥⑥⑥⑥⑥⑥⑥ ρ2 "❊❊❊❊❊❊❊❊ P∆ /❴❴❴❴❴❴❴ P∆can ρ2 Z ∆ /❴❴❴❴❴❴❴❴ X bZ∆ ~⑥⑥⑥⑥⑥⑥⑥⑥ ρ1 α and Σ∆can where bΣ denotes the common regular simplicial subdivision of the normal fans Σ∆ . In particular, one obtains two birational morphisms ρ1 : bZ∆ → Z ∆, ρ2 : bZ∆ → X in the diagram Proof. Let L be the ample Cartier divisor on the d-dimensional toric variety P∆ corresponding to a d-dimensional lattice polytope ∆. We apply the toric Zariski for some toric desingularization ❅❅❅❅❅❅❅❅ where bZ∆ denotes the Zariski closure of Z∆ in PbΣ. decomposition to the adjoint divisor D := ρ∗L+KPbΣ ρ : PbΣ → P∆ defined by a fan bΣ which is a regular simplicial subdivision of the 1-dimensional cones in bΣ. i=1 Vi and ρ∗L =Ps = −Ps Since one has KPbΣ rational polytope ∆D corresponding to the adjoint divisor on PbΣ D = ρ∗L + KPbΣ normal fan Σ∆. Let {v1, . . . , vs} be the set of primitive lattice vectors in N generating i=1(−ord∆(vi))Vi, we obtain that the is exactly the Fine interior ∆F I of ∆. We can assume that Supp(∆F I) = {v1, . . . , vr} (r ≤ s) and the first l lattice vectors v1, . . . , vl (l ≤ r) form the set of generators of 1-dimensional cones Rvi (1 ≤ i ≤ l) in the normal fan Σ∆can so that one has (−ord∆(vi) − 1)Vi sXi=1 = ∆can = r\i=1 Γ∆ 0 (vi) = Γ∆ 0 (vi) l\i=1 ~ " / ~  / THE STRINGY EULER NUMBER AND THE MAVLYUTOV DUALITY 11 and ord∆(vi) + 1 = ord∆F I (vi), ∀i = 1, . . . , r. the sum P + N where The toric Zariski decomposition of D = ρ∗L + KPbΣ (−ord∆F I (vi))Vi, P = = Ps i=1(−ord∆(vi) − 1)Vi is sXi=1 sXi=r+1 N = (ord∆F I (vi) − ord∆(vi) − 1)Vi where (ord∆F I (vi) − ord∆(vi) − 1) > 0 for all i > r. Moreover, there exists a projective Q-factorial toric variety P′ such that v1, . . . , vr is the set of primitive lattice generators of 1-dimensional cones in the fan Σ′ defining the toric variety P′ and (−ord∆F I (vi))V ′ i rXi=1 is a nef Q-Cartier divisor on P′ i=1 Vi where V1, . . . , Vl the set of torus invariant divisors on P∆can. Let X be the Zariski closure of the affine ∆- nondegenerated hypersurface Z∆ in P∆can. Then X is linearly equivalent to a linear Therefore the canonical divisor KP∆can equals −Pl combination Pl i=1 biVi, where bi = −ord∆(vi) (1 ≤ i ≤ l). So we obtain KP∆can + X ∼ (−ord∆(vi) − 1)Vi = (−ord∆F I (vi) − 1)Vi. lXi=1 On the other hand, we have lXi=1 l\i=1 ∆F I = \n∈Supp(∆F I ) Γ∆ 1 (n) = Γ∆ 1 (vi). So KP∆can + X is a semiample Q-Cartier divisor on the projective toric variety P∆can corresponding to the rational convex polytope ∆F I. For nondegenerate hypersufaces one can apply the adjunction and obtain that the canonical class KX is the restriction to X of the semiample Q-Cartier divi- sor KP∆can + X. The log-discrepancies of the toric pair (P∆can, X) are equal to the discrepancies of X, because of inversion of the anjunction for non-degenerate hypersurfaces [Amb03]. (cid:3) Corollary 2.19. For the above birational morphism ρ2 : PbΣ → P∆can one has 2 (KP∆can + X) + aiVi sXi=l+1 and + bZ∆ = ρ∗ = ρ∗ 2KX + KbPΣ KbZ∆ sXi=l+1 aiDi, Di := Vi ∩ bZ∆, vi ∈ N and where Vi denotes the torus invariant divisor on PbΣ corresponding to the lattice point ai = −ord∆(vi) + ord∆F I (vi) − 1 ≥ 0. By 2.7, one immediately obtains 12 VICTOR BATYREV Corollary 2.20. It a d-dimensional lattice polytope ∆ contains an interior lattice point, then a ∆-nondegenerated affine hypersurface Z∆ ⊂ Td has a minimal model. Example 2.21. As we have already mentioned in 2.10 there exist 3-dimensional lattice polytopes ∆ without interior lattice points such that ∆F I is not empty. One of such examples is the 3-dimensional lattice simplex ∆ such that ∆-nondegenerated hypersurface is birational to the Godeaux surface obtained as a quotient of the 3 = 0 in P3 by the action of a cyclic group of order 5 Fermat quintic z5 2 + z5 0 + z5 1 + z5 (z0 : z1 : z2 : z3) 7→ (z0 : ρz1 : ρ2z2 : ρ3z3) where ρ is a 5-th root of unity. We apply the above general results to ∆-nondegenerate hypersurfaces whose min- imal models are Calabi-Yau varieties. It is known that the number of interior lat- tice points in ∆ equals the geometric genus of the ∆-nondegenerate hypersurface [Kho78]. Therefore, if a nondegenerate hypersurface Z∆ is birational to a Calabi- Yau variety, then ∆ must contain exactly one interior lattice point. However, this condition for ∆ is not sufficient. Example 2.22. In [CG11] Corti and Golyshev gave 9 examples of 3-dimensional lattice simplices ∆ with a single interior lattice point 0 such that the corresponding nondegenerate hypersurfaces Z∆ are not birational to a K3-surface. For example they consider hypersurfaces of degree 20 in the weighted projective space P(1, 5, 6, 8). The corresponding 3-dimensional lattice simplex ∆ is the convex hull of the lattice points (1, 0, 0), (0, 1, 0), (0, 0, 1), (−5, −6, −8). The Fine interior ∆F I of ∆ is a 1- dimensional polytope on the ray generated by the lattice vector (−1, −1, −2). Theorem 2.23. A canonical model of ∆-nondegenerate affine hypersuface Z∆ ⊂ Td is birational to a Calabi-Yau variety X with at worst Gorenstein canonical singu- larities if and only if the Fine interior ∆F I of the lattice polytope ∆ consists of a single lattice point. If ∆F I = 0 then X can be obtained as a Zariski closure of Z∆ in the toric Q-Fano variety P∆can so that X is an anticanonical divisor on P∆can. There exists an embedded desingularization ρ2 : bZ∆ → X and = ρ∗ 2KX + aiDi, KbZ∆ sXi=l+1 where the discrepancy ai of the exceptional divisor Di := Vi ∩ bZ∆ on bZ∆ can be computed by the formula ai = −ord∆(vi) − 1 ≥ 0. Proof. First of all we remark that this statement has been partially proved in [ACG16, Prop. 2.2.], but the application of Mori theory for nondegenerate hyper- surfaces (see 2.18 and 2.19) imply stronger statements. The above formula for the discrepancies ai is not new and it has appeared already in [CG11] for resolutions of canonical singularities of Calabi-Yau hypersurfaces X in weighted projective spaces. In general case, one can make a direct computation of ai using the global nowhere vanishing differential (d − 1)-form ω obtained as the Poincaré residue ω = Res Ω of THE STRINGY EULER NUMBER AND THE MAVLYUTOV DUALITY 13 the rational differential form [Bat93]: 1 f Ω := dx1 x1 ∧ · · · ∧ dxd xd . Since ∆ is the Newton polytope of the Laurent polynomial f , the order of zero of ω along the exceptional divisor Ei corresponding to the lattice point vi ∈ N equals −ord∆(vi) − 1. (cid:3) Remark 2.24. We note that a d-dimensional lattice polytope ∆ with ∆F I = 0 is reflexive if and only if ∆ = ∆can. If ∆ is reflexive then the canonical singularities of the projective Calabi-Yau hy- persurface Z ∆ ⊂ P∆ have a MPCP (maximal projective crepant partial) resolution obtained from a simplicial fan bΣ whose generators of 1-dimensional cones are lattice points on the boundary of the polar reflexive polytope ∆∗ [Bat94]. This fact can be generalized to an arbitrary d-dimensional lattice polytope ∆ ⊂ MR such that ∆F I = 0. For this we need the following statement: Proposition 2.25. Let ∆ be a d-dimensional polytope with ∆F I = 0. Then one has where ∆∗ is the polar polytope. Supp(∆F I) = {∆∗ ∩ N} \ {0}, Proof. By Definition 2.11, a lattice point n ∈ N belongs to the support of the Fine interior ∆F I = 0 if and only if ord∆(n) = −1. The polar polytope ∆∗ ⊂ NR is defined by the condition ord∆(x) ≥ −1. Therefore, we obtain Supp(∆F I) ⊂ ∆∗. Since 0 is an interior lattice point of ∆ one has 0 > ord∆(n) ∈ Z for any nonzero lattice vector n ∈ N. In particular, one has ord∆(n) = −1 for any nonzero lattice point n ∈ ∆∗, i.e., {∆∗ ∩ N} \ {0} ⊂ Supp(∆F I). (cid:3) Theorem 2.26. A minimal model of a ∆-nondegenerate affine hypersuface Z∆ ⊂ Td is birational to a Calabi-Yau variety X ′ with at worst Q-factorial Gorenstein terminal singularities if and only if the Fine interior of ∆ is 0. Proof. By Theorem 2.23, it remains to explain how to construct a maximal pro- jective crepant partial resolution ρ′ : X ′ → X. We consider the finite set Supp(∆F I) = {v1, . . . , vr} ⊂ N consisting of all nonzero lattice points in ∆∗ ⊂ NR. (see 2.25). Denote by Σ the fan of cones over all faces of the lattice polytope [∆∗] obtained as convex hull of all lattice points in Supp(∆F I). The fan Σ ad- mits a maximal simplicial projective subivision Σ′ which consits of simplicial cones whose generators are nonzero lattice vectors in ∆∗. Thus we obtain a projective crepant toric morphism ρ′ : PΣ′ → PΣ. Since Σ is the normal fan to the polytope ∆can = [∆∗]∗, the morphism ρ′ induces a projective crepant morphism of Calabi-Yau varieties ρ′ : X ′ → X, where X ′ is the Zariski closure of Z∆ in PΣ′. Since the toric singularities of PΣ′ are Q-factorial and terminal, the same is true for the singularities of X ′. (cid:3) 3. The Mavlyutov duality In [Mav11] Mavlyutov has proposed a generalization the Batyrev-Borisov dual- ity [BB97]. In particular, his generalization includes the polar duality for reflexive 14 VICTOR BATYREV polytopes [Bat94]. We reformulate the ideas of Mavlyutov about Calabi-Yau hyper- surfaces in toric varieties in some equivalent more convenient form. For simplicity we denote by [P ] the convex hull Conv(P ∩ Zd) for any subset P ⊂ Rd. As above, we denote by P ∗ the polar set of P if 0 is an interior lattice point of P . Let ∆ ⊂ MR be d-dimensional lattice polytope such that the Fine interior of ∆ is zero, i.e., ∆F I = 0 ∈ M. By 2.25, the support of the Fine interior Supp(∆F I) is equal to the set of nonzero lattice points in the polar polytope ∆∗ ⊂ NR and the zero lattice point 0 ∈ N is the single interior lattice point of [∆∗]. Therefore, the inclusion [∆∗] ⊆ ∆∗ implies the inclusions and ∆ = (∆∗)∗ ⊆ [∆∗]∗ = ∆can ∆ ⊆ [[∆∗]∗] = [∆can], because ∆ is a lattice polytope. Definition 3.1. We call a d-dimensional lattice polytope ∆ ⊂ MR with ∆F I = 0 pseudoreflexive if one has the equality ∆ = [[∆∗]∗]. Remark 3.2. The above defintion of pseudoreflexive polytopes has been discovered by Mavlyutov in 2004 (see [Mav05, Remark 4.7]). In the paper [Mav11] Mavlyu- tov called these polytopes Z-reflexive (or integrally reflexive). Independently, this definition has been discovered by Kreuzer [Kr08, Definition 3.11] who called such polytopes IP C-closed. Remark 3.3. Every reflexive polytope ∆ is pseudoreflexive, because for reflexive polytopes ∆ we have ∆ = [∆] and ∆∗ = [∆∗]. The converse is not true if dim ∆ ≥ 5. For instance the convex hull Conv(e0, e1, . . . , e5) of the standard basis e1, . . . , e5 in Z5 and the lattice vector e0 = −e1 − e2 − e3 − e4 − 2e5 is a 5-dimensional pseudoreflexive simplex which is not reflexive. There exist a close connection between lattice polytopes ∆ with ∆F I = 0 and pseudoreflexive polytopes: Proposition 3.4. Let ∆ ⊂ MR a d-dimensional lattice polytope. Then the following conditions are equivalent: (i) ∆F I = 0; (ii) the polytopes ∆ and [∆∗] contain 0 in their interior; (iii) ∆ contains 0 in its interior and ∆ is contained in a pseudoreflexive polytope. Proof. (i) ⇒ (ii). Assume that ∆F I = 0. Then 0 is an interior lattice point of ∆ and the support of the Fine interior Supp(∆F I) is exactly the set of nonzero lattice points in the polar polytope ∆∗. Moreover, one has 0 = {x ∈ MR : hx, vi ≥ 0 ∀v ∈ Supp(∆F I)}. Hence, [∆∗] also contains 0 in its interior. (ii) ⇒ (i). If ∆ contains 0 in its interior, then 0 ∈ ∆F I. For any nonzero lattice point v ∈ ∆∗ the minimum of h∗, vi on ∆ equals −1. If [∆∗] contains 0 in its interior, THE STRINGY EULER NUMBER AND THE MAVLYUTOV DUALITY 15 then there exists lattice points v1, . . . , vl ∈ [∆∗] generating MR such that for some positive numbers λi (1 ≤ i ≤ l) one has λ1v1 + · · · + λlvl = 0. On the other hand, ∆F I is contained in the intersection of the half-spaces hx, vii ≥ 0 1 ≤ i ≤ l. Therefore, one has ∆F I = 0. (iii) ⇒ (ii). Assume that ∆ is contained in a pseudoreflexive lattice polytope e∆. Then we obtain the inclusions e∆∗ ⊆ ∆∗ and [e∆∗] ⊆ [∆∗] . Since e∆ is pseudoreflexive, its Fine interior is zero and it follows from (i) ⇒ (ii) that [e∆∗] contains 0 in its interior. Therefore, the lattice polytope [∆∗] also contains 0 in its interior. (i ⇒ (iii). Assume that ∆F I = 0. Then we obtain the inclusion ∆ ⊆ [[∆∗]∗]. It is sufficient to show that [[∆∗]∗] is pseudoreflexive. The latter follows from the equality [[[∆∗]∗]∗] = [∆∗]. Indeed, the inclusion ∆ ⊆ [[∆∗]∗] implies the inclusions [[∆∗]∗]∗ ⊆ ∆∗ and [[[∆∗]∗]∗] ⊆ [∆∗]. On the other hand, the Fine interior of [∆∗] is also zero, because ∆ contains 0 in its interior. This implies the opposite inclusion [∆∗] ⊆ [[[∆∗]∗]∗]. (cid:3) Corollary 3.5. Let ∆ ⊂ MR be a d-dimensional lattice polytope with ∆F I = 0. Then the following statements hold. (i) The lattice polytopes [∆∗] and [[∆∗]∗] are pseudoreflexive; (ii) [[∆∗]∗] is the smallest pseudoreflexive polytope containing ∆. (cid:3) Proof. The statement (i) follows from the equality [[[∆∗]∗]∗] = [∆∗] in the proof of 3.4. If e∆ is a pseudoreflexive polytope containing ∆, then the inclusion ∆ ⊆ e∆ implies the sequence of inclusions e∆∗ ⊆ ∆∗, [e∆∗] ⊆ [∆∗], [∆∗]∗ ⊆ [e∆∗]∗, [[∆∗]∗] ⊆ [[e∆∗]∗] = e∆. This implies (ii). The statements in 3.4 and 3.5 motivate another names for lattice polytopes ∆ with ∆F I = 0: Definition 3.6. A d-dimensional lattice polytope is called almost pseudoreflexive if ∆F I = 0. If ∆ is almost pseudoreflexive then we call the lattice polytope [[∆∗]∗] the pseudoreflexive closure of ∆ and the lattice polytope [∆∗] the pseudoreflexive dual of ∆. The polytope ∆ is called almost reflexive if its pseudoreflexive closure [[∆∗]∗] (or, equivalently, its pseudoreflexive dual [∆∗]) is reflexive. In this case, we will call the lattice polytope [[∆∗]∗] = [∆∗]∗ = ∆can also the canonical reflexive closure of ∆. Example 3.7. The 3-dimensional lattice polytope ∆ obtained as the convex hull of (1, 0, 0), (0, 1, 0), (0, 0, 1), (−1, −1, −2) ∈ Z3 is a 3-dimensional almost reflexive simplex which is not reflexive. The canonical reflexive closure [[∆∗]∗] = [∆∗]∗ of ∆ is a reflexive lattice polytope obtained from ∆ by adding one more vertex (0, 0, −1). Remark 3.8. If ∆ is pseudoreflexive, then ∆∨ := [∆∗] is also pseudoreflexive. In particular, one obtains a natural duality ∆ ↔ ∆∨ for pseudoreflexive polytopes that generalizes the polar duality for reflexive polytopes. This duality was suggested by Mavlyutov in [Mav11] for unifying different combinatorial mirror constructions. Remark 3.9. Unfortunately almost pseudoreflexive polytopes ∆ do not have a nat- ural duality, although they appear in the Berglung-Hübsch-Krawitz mirror construc- tion. Nevertheless, the pseudoreflexive duals [∆∗] of almost pseudoreflexive poly- topes ∆ allow to connect the Mavlyutov duality with the Berglung-Hübsch-Krawitz 16 VICTOR BATYREV mirror construction. For instance, it may happen that two different almost pseu- doreflexive polytopes ∆1 6= ∆2 have the same pseudoreflexive duals, i.e., [∆∗ 1] = [∆∗ 2]. This equality is the key observation for the birationality of BHK-mirrors investigated in [Ke13, Cla14, Sh14]. Definition 3.10. Let Θ be a k-dimensional face of a d-dimensional pseudoreflexive polytope ∆ ⊂ MR. We call Θ ordinary if the following equality holds:  \l∈Z≥0 lΘ ∩ M = R≥0Θ ∩ M, in other words, if all lattice points in the (k + 1)-dimensional cone σΘ = R≥0Θ over the face Θ ≺ ∆ are contained in the multiples lΘ (l ∈ Z≥0). Proposition 3.11. Let Θ ≺ ∆ be a k-dimensional face of a lattice polytope ∆ with ∆F I = 0. Assume that [Θ∗] is nonempty. Then Θ is ordinary and [Θ∗] is a face of dimension ≤ d − 1 − k of the pseudoreflexive polytope [∆∗]. Proof. Let x ∈ Θ be a point in the relative interior of Θ. The minimum of the linear function hx, ∗i on ∆∗ equals −1 and it is attained exactly on the polar face Θ∗ ≺ ∆∗ of the rational polar polytope ∆∗. The minimum µx of hx, ∗i on the lattice polytope [∆∗] is attained on some lattice face F ≺ [∆∗] of the lattice polytope [∆∗] such that F := {y ∈ [∆∗] : hx, yi = µx}. The minimum µx must be at least −1, because the polytope [∆∗] is contained in ∆∗. If [Θ∗] is not empty then the linear function hx, ∗i has the constant value −1 on [Θ∗]. This implies that the minimum µx must be −1. Therefore F must be contained in Θ∗ and F = [F ] ⊆ [Θ∗]. Since [Θ∗] ⊆ {y ∈ [∆∗] hx, yi = µx = −1} = F , we conclude F = [Θ∗]. (cid:3) : The next statement is a slight generalization of the results of Skarke in [Sk96]. Theorem 3.12. [Mav13] Let Θ is a face of dimension k ≤ 3 of a d-dimensional pseudoreflexive polytope ∆. Then Θ is ordinary. In particular, any pseudoreflexive lattice polytope ∆ of dimension ≤ 4 is reflexive. Proof. Consider the (k + 1)-dimensional subspace L := RΘ generated by Θ. Then Θ is contained in the k-dimensional affine hyperplane HΘ in L. It is enough to show that the integral distance between HΘ and 0 equals 1. Assume that this distance is larger than 1. Consider the pyramid ΠΘ := Conv(Θ, 0). By lemma of Skarke [Sk96, Lemma 1], there exists an interior lattice point u0 ∈ M in 2ΠΘ which is not contained in ΠΘ. Therefore, the lattice point u0 is an interior lattice point in the polytope 2∆ ⊆ 2[∆∗]∗. If {v1, . . . , vl} ⊂ N is the set of vertices of [∆∗] then the polytope [∆∗]∗ is determined by the inequalities hx, vii ≥ −1 (1 ≤ i ≤ l). The interior lattice point u0 ∈ 2[∆∗]∗ must safisfy the inequalities hu0, vii > −2 (1 ≤ i ≤ l). Since hu0, vii >∈ Z (1 ≤ i ≤ l), we obtain hu0, vii ≥ −1 (1 ≤ i ≤ l), i.e., u0 is a nonzero lattice point in [∆∗]∗. Since ∆ = [[∆∗]∗], u0 is a nonzero lattice point contained in ∆ and in the k-dimensional cone R≥0Θ over Θ. On the other hand, ∆ ∩ R≥0Θ = Θ ⊂ ΠΘ. Contradiction. (cid:3) THE STRINGY EULER NUMBER AND THE MAVLYUTOV DUALITY 17 Proposition 3.13. [Mav13] Let Θ ≺ ∆ be an ordinary k-dimensional face of a pseudoreflexive polytope ∆ such dim[Θ∗] = dim Θ∗ = d − 1 − k ≥ 0. Then one has [[Θ∗]∗] = Θ. Proof. If dim Θ∗ = dim[Θ∗] = d − k − 1 then there exists a point y ∈ ∆∗ which is contained in the relative interior of [Θ∗] and in the relative interior of Θ∗. In particular, y ∈ Θ∗ is contained in the relative interior of the (d − k)-dimensional normal cone σΘ and therefore the minimum of the linear function h∗, yi on ∆ equals −1 and it is attained on Θ = {x ∈ ∆ : hx, yi = −1}. By definition of the polar polytope [∆∗]∗, the minimum of h∗, yi on [∆∗]∗ also equals −1, and it is attained on the k-dimensional dual face [Θ∗]∗ ≺ [∆∗]∗. Hence, [Θ∗]∗ contains the lattice face Θ and [[Θ∗]∗] also contains Θ. By 3.11, the lattice polytope [[Θ∗]∗] is face of [[∆∗]∗] = ∆ of dimension ≤ k. Since [[Θ∗]∗] contains the k-dimensional face Θ ≺ ∆, the face [[Θ∗]∗] ≺ ∆ must be Θ. (cid:3) Definition 3.14. We call a k-dimensional face Θ of a pseudoreflexive polytope ∆ ⊂ MR regular, if A k-dimensional face Θ is called singular if it is not regular. dim[Θ∗] = d − k − 1. By 3.13, we immediately obtain: Corollary 3.15. Let ∆ ⊂ MR be a d-dimensional pseudoreflexive polytope. Then there exists a natural bijection Θ ↔ Θ∨ := [Θ∗] between the set of k-dimensional regular faces of ∆ and (d − k − 1)-dimensional regular faces of ∆∗. Remark 3.16. Pseudoreflexive lattice polytopes ∆ satisfy a combinatorial duality ∆ ↔ ∆∨ that extends the polar duality for reflexive lattice polytopes . However, in contrast to polar duality for reflexive polytopes there is no natural bijection between arbitrary k-dimensional faces of a pseudoreflexive polytope ∆ and (d − k − 1)- dimensional faces of its dual ∆∨. Such a natural bijection exists only for regular k-dimensional faces Θ ⊂ ∆. Remark 3.17. By 3.11, every regular face Θ ≺ ∆ is ordinary. It is easy to see that for a (d − 1)-dimensional face Θ ≺ ∆ the following conditions are equivalent: (i) Θ is regular; (ii) Θ is ordinary; (iii) the integral distance from 0 ∈ M to Θ is 1. Remark 3.18. Reflexive polytopes of dimension 3 and 4 have been classified by Kreuzer und Skarke [KS98, KS00]. It is natural task to extend these classification to lattice polytopes with Fine interior 0. By 3.4, a lattice polytope ∆ of dimension 3 or 4 has Fine interior 0 if and only if ∆ contains 0 in its interior and ∆ is contained in some reflexive polytope ∆′. All 3-dimensional lattice polytopes with the single interior lattice point 0 have been classified by Kasprzyk [Kas10]. There exists exactly 674,688 3-dimensional lattice polytopes ∆ with only a single interior lattice point. However, not all these polytopes ∆ have Fine interior 0. I was informed by Kasprzyk that among these 674, 688 lattice polytopes there exist exactly 9, 089 lattice polytopes whose Fine 18 VICTOR BATYREV interior has dimension ≥ 1. These polytopes correspond to elliptic surfaces, Todorov surfaces and some other interesting algebraic surfaces. According to Kreuzer und Skarke [KS98, KS00], there exist exactly 4, 319 3- dimensional reflexive polytopes. We remark that canonical models of K3-surfaces coming from 3-dimensional reflexive polytopes have at worst toroidal quotient sin- gularities of type An. However, the canonical models of K3-surfaces coming from 3-dimensional lattice polytopes ∆ with the weaker condition ∆F I = 0 may have more general Gorenstein canonical singularities of types Dn and En. Analogously, we remark that canonical singularities of 3-dimensional Calabi-Yau varieties obtained as hypersurfaces in 4-dimensional Gorenstein toric Fano vari- eties defined by 4-dimensional reflexive polytopes are toroidal. They admit smooth crepant resolutions, because any 3-dimensional Q-factorial terminal Gorenstein toric variety is smooth. Singularities of 3-dimensional Calabi-Yau varieties X obtained as minimal models of ∆-nondegenerate hypersurfaces with ∆F I = 0 generally can not be resolved crepantly, because Q-factorial Gorenstein terminal singularities in dimension 3 are cDV -points that may cause that the stringy Euler number of X will be a rational number [DR01]. So the classification of 4-dimensional lattice polytopes ∆ with Fine interior 0 would give many new examples of 3-dimensional Calabi-Yau varieties with isolated terminal cDV -points that need additionally to be smoothed by a deformation [Na94] in order to get a smooth Calabi-Yau 3-fold. It would be very interesting to know what rational numbers can appear as stringy Euler numbers of minimal 3-dimensional Calabi-Yau varieties coming from 4-dimensional lattice polytopes ∆ with ∆F I = 0. 4. The stringy Euler number Definition 4.1. If V is a smooth projective algebraic variety over C, then its E- polynomial (or Hodge polynomial) is defined as E(V ; u, v) := X0≤p,q≤dim V (−1)p+qhp,q(V )upvq, where hp,q(V ) are Hodge numbers of V . For any quasi-projective variety W one can use the mixed Hodge structure in k- c (W ) with compact supports and define E(W ; u, v) by the th cohomology group H k formula where the coefficients E(W ; u, v) :=Xp,q ep,q(W ) =Xk ep,q(W )upvq, (−1)khp,q(H k c (W )) are called Hodge-Deligne numbers of W [DKh86]. Definition 4.2. Let X be a normal projective variety over C with at worst Q- Gorenstein log-terminal singularities. Denote by r the minimal positive integer such that rKX is a Cartier divisor. Let ρ := Y → X be a log-desingularization together with smooth irreducible divisors D1, . . . , Dk with simple normal crossings whose THE STRINGY EULER NUMBER AND THE MAVLYUTOV DUALITY 19 support covers the exceptional locus of ρ. We can uniquely write KY = ρ∗KX + aiDi. kXi=1 for some rational numbers ai ∈ 1 r Z satisfying the additional condition ai = 0 if Di is not in the exceptional locus of ρ. We set I := {1, . . . , k} and, for any ∅ ⊆ J ⊆ I, we define Y if J = ∅, if J 6= ∅, Tj∈J Dj DJ := Yj∈J Estr(X; u, v) := X∅⊆J⊆I Yj∈J = X∅⊆J⊆I D◦ J := DJ \ [j∈I\J Dj. uv − 1 (uv)aj+1 − 1 · E(D◦ uv − 1 (uv)aj+1 − 1 J ; u, v) − 1 · E(DJ ; u, v), where E(DJ ; u, v) = Pp,q(−1)p+qhp,q(DJ )upvq is the E-polynomial of the smooth projective variety DJ . The rational function Estr(X; u, v) is called stringy E-function of the algebraic variety X. Let x ∈ X be a point on X. We define the local stringy E-function of X at x ∈ X by the formula In particular, we define the local stringy Euler number of X at point x ∈ X as Estr(X, x; u, v) := X∅⊆J⊆I Yj∈J Yj∈J estr(X, x) := X∅⊆J⊆I uv − 1 (uv)aj+1 − 1 −aj aj + 1 − 1 · E(ρ−1(x) ∩ DJ ; u, v). − 1 · e(ρ−1(x) ∩ DJ). Our goal is to derive a combinatorial formula for the stringy E-function Estr(X; u, v) of a minimal Calabi-Yau model X of an affine ∆-nondegenerate hypersurface Z ⊂ Td corresponding to a d-dimensional lattice polytope ∆ ⊂ MR such that ∆F I = 0. For this purpose we need a rational function R(C, m, t) associated with an arbitrary d- dimensional rational polyhedral cone C ⊂ NR and a primitive lattice point m ∈ M in the interior of the dual cone C ∗ ⊂ MR. Definition 4.3. Let C ⊂ NR be an arbitrary d-dimensional rational polyhedral cone with vertex 0 = C ∩ (−C) and let m ∈ M be a primitive lattice point such that C(1) := {y ∈ C : hm, yi ≤ 1} is a d-dimensional compact polytope with rational vertices. Let C ◦ be the interior of the cone C. We define two power series and R(C, m, t) := Xn∈C∩N R(C ◦, m, t) := Xn∈C ◦∩N thm,ni thm,ni. 20 VICTOR BATYREV Example 4.4. If M = N = Zd, C = Rd ≥0 ⊂ Rd and m = (1, . . . , 1). Then C(1) is a d-dimensional simplex in Rd defined by the conditions xi ≥ 0 (1 ≤ i ≤ d), i=1 xi ≤ 1. We have Pd and R(C, m, t) = ∞Xk=0 tk!d R(C ◦, m, t) = ∞Xk=1 tk!d = 1 (1 − t)d = td (1 − t)d Proposition 4.5. The power series R(C, m, t) and R(C ◦, m, t) are rational func- tions satisfying the the equation R(C, m, t) = (−1)dR(C ◦, −m, t). Moreover, two limits (1 − t)dR(C, m, t), lim t→1 lim t→1 (t − 1)dR(C ◦, −m, t) equal the integral volume v(C(1)) = d!V ol(C(1)), where V ol(C(1)) denotes the usual volume of the d-dimensional compact set C(1). Proof. First we remark that R(C, m, t) is a rational function, because R(C, m, t) can be considered as the Poincaré series of the graded finitely graded commutative semigroup algebra C[C ∩ N] such that the degree of an element n ∈ C ∩ N equals hm, ni. The function R(C ◦, m, t) is also rational, because it is the Poincaré series of a graded homogeneous ideal in C[C ∩ N]. In order to compute the rational functions R(C, m, t) and R(C ◦, m, t) explicitly we use a regular simplicial subdivision of the cone C defined by a finite fan Σ = {σ} consisting of cones σ generated by parts of Z-bases of N. Denote by σ◦ the relative interior of a cone σ ∈ Σ. Then we obtain (4.1) and (4.2) R(C, m, t) = Xσ∈Σ R(C ◦, m, t) = Xσ∈Σ σ◦⊆C◦ R(σ◦, m, t) R(σ◦, m, t). If σ ∈ Σ is a k-dimensional cone, the semigroup σ ∩ N is freely generate by some elements v1, . . . , vk ∈ N such that hm, vii = ci ∈ Z>0 (1 ≤ i ≤ k). Therefore, we obtain and R(σ, m, t) = 1 1 − tci kYi=1 R(σ◦, −m, t) = t−ci 1 − t−ci kYi=1 = kYi=1 1 tci − 1 = (−1)k 1 1 − tci kYi=1 = (−1)kR(σ, m, t). THE STRINGY EULER NUMBER AND THE MAVLYUTOV DUALITY 21 In order to prove the equation R(C, m, t) = (−1)dR(C ◦, −m, t) for the whole d- dimensional cone C we note that for any σ ∈ Σ one has (−1)dim σR(τ ◦, −m, t). (−1)dim σR(τ ◦, −m, t) = Using the equalities (4.1) and (4.2), we get (−1)dim σ−dim τ R(τ, m, t) = Xτ (cid:22)σ σ◦⊆C◦ Xτ (cid:22)σ R(σ◦, m, t) = Xσ∈Σ R(τ ◦, −m, t) Xτ (cid:22)σ∈Σ R(σ◦, m, t) = Xτ (cid:22)σ R(C ◦, m, t) = Xσ∈Σ =Xτ ∈Σ =(−1)d Xτ ⊆C σ◦⊆C◦ (−1)dim σ = R(τ ◦, −m, t) = (−1)dR(C, −m, t), We note that the limit because for any cone τ ∈ Σ one has Pτ (cid:22)σ(−1)dim σ = (−1)d. kYi=1 (1 − t)dR(σ◦, m, t) = lim t→1 (1 − t)d lim t→1 tci 1 − tci is zero if k = dim σ < d. If σ ∈ Σ(d) is a d-dimensional cone, then lim t→1 (1 − t)dR(σ◦, m, t) = lim t→1 (1 − t)d tci 1 − tci = dYi=1 dYi=1 1 ci = d!V ol(σ(1)), because σ(1) is a d-dimensional simplex which is the convex hull of vectors 1 ci where v1, . . . , vd is a Z-basis of M. Using (4.1), we get lim t→1 (1 − t)dR(C, m, t) = Xσ∈Σ lim t→1 (1 − t)dR(σ◦, m, t) = Xσ∈Σ(d) v(σ(1)) = v(C(1)), because V ol(C(1)) = Xσ∈Σ(d) V ol(σ(1)). It follows now from the equation R(C, m, t) = (−1)dR(C ◦, −m, t) that (t − 1)dR(C ◦, −m, t) = v(C(1)). lim t→1 vi, (cid:3) The following results of Danilov and Khovanskii allow us to compute the poly- nomial E(Z; u, v) for any (d − 1)-dimensional ∆-nondegenerate affine hypersurface Z∆ ⊂ T [DKh86, Remark 4.6]. Theorem 4.6. Let ∆ ⊂ MR be a d-dimensional lattice polytope. The power series P (∆, t) := k∆ ∩ Mtk ∞Xk=0 is a rational function of the form P (∆, t) = ψ0(∆) + ψ1(∆)t + · · · + ψd(∆)td (1 − t)d+1 , 22 VICTOR BATYREV where ψi(∆) (0 ≤ i ≤ d) are nonnegative integers satisfying the conditions ψ0(∆) = i=1 ψi(∆) = v(∆). 1, Pd hypersurface Z∆ ⊂ Td. Then one has Let E(Z∆; u, 1) be the E-polynomial of a (d − 1)-dimensional ∆-nondegenerate E(Z∆; u, 1) = (u − 1)d − (−1)d u + (−1)d−1 ψi(∆)ui−1. dXi=1 In particular, the Euler number e(Z∆) = E(Z∆; 1, 1) equals (−1)d−1v(∆). In particular, one obtains Corollary 4.7. [Kho78] The Euler number e(Z∆) = E(Z∆; 1, 1) of a (d − 1)- dimensional affine ∆-nondegenerate hypersuface Z∆ equals (−1)d−1v(∆). Definition 4.8. Let ∆ ⊂ MR is an arbitrary d-dimensional almost pseudoreflexive polytope. For any k-dimensional face Θ ≺ ∆ we define the polynomial E(Θ, u) := (u − 1)dim Θ u + (u − 1)dim Θ+1 u lΘ ∩ Mul = (u − 1)k − (−1)k u + (−1)k−1 ψi(Θ)ui−1. Xl∈Z≥0 kXi=1 Definition 4.9. Let ∆ ⊂ MR be an arbitrary d-dimensional almost pseudoreflexive polytope and let σΘ be the (d − k)-dimensional cone in the normal fan Σ∆ that correspond to the k-dimensional face Θ ≺ ∆. We choose an arbitrary lattice point m ∈ Θ ∩ M and set R(σΘ, u) := R(σΘ, −m, u) where the rational function R(σΘ, −m, u) defined in 4.3. It is easy to see that the rational function R(σΘ, −m, u) does not depend on the choice of the lattice point m ∈ Θ ∩ M. We prove the following theorem: Theorem 4.10. Let ∆ be an arbitrary d-dimensional almost pseudoreflexive poly- tope. Denote by X a canonical (d − 1)-dimensional Calabi-Yau model of a ∆- nondegenerate hypersurface Z∆ ⊂ T in the d-dimensional algebraic torus T . Then the stringy function Est(X; u, 1) can be computed as follows: Est(X; u, 1) = XΘ(cid:22)∆ dim Θ≥1 E(Θ, u) · R(σΘ, u) · (1 − u)d−dim Θ. Proof. Let bΣ be a common regular simplicial subdivision of the normal fans Σ∆ and Σ∆can. As in Theorem 2.18 we obtain birational morphisms ρ1 : bZ∆ → Z ∆, ρ2 : bZ∆ → X in the diagram bZ∆ ~⑥⑥⑥⑥⑥⑥⑥⑥ ρ1 α ρ2 ❅❅❅❅❅❅❅❅ Z ∆ /❴❴❴❴❴❴❴❴ X ~  / THE STRINGY EULER NUMBER AND THE MAVLYUTOV DUALITY 23 where bZ∆ is a smooth variety obtained as Zariski closure of Z∆ in the smooth projective toric variety bP defined by the fan bΣ. For computing the stringy E- function Estr(X, u, 1) we use the formula (4.2) uaj+1 − 1 · E(D◦ where the strata D◦ Estr(X; u, 1) = X∅⊆J⊆I J are intersections of the hypersurface bZ∆ ⊂ bP with torus orbits corresponding to cones σ ∈ bΣ. Let σ ∈ bΣ(k) be a k-dimensional simplicial cone generated by primitive lattice vectors vi1, . . . , vik. Then the relative interior σ◦ of σ is contained in the relative interior σΘ ◦ of some cone σΘ of the normal fan Σ∆ corresponding to a face Θ ≺ ∆, dim Θ ≤ d − k. We set J := {i1, . . . , ik}. By 2.23, the discrepancy coefficients aj of Yj∈J J ; u, 1) u − 1 smooth divisors Dj (j ∈ J) on bZ∆ can be computed by the formula aj = −hm, vji−1 (j ∈ J) where m ∈ M is any lattice point in the face Θ. Since the fiber ρ−1 1 (p) of the birational toric morphism ρ1 over every point p ∈ P∆ consists of torus orbits, the codimension k stratum D◦ J is isomorphic to the product of a torus (C∗)d−k−dim Θ and a Θ-nondegenerate hypersurface ZΘ ⊂ (C∗)dimΘ. By 4.6 and 4.8 we have E(D◦ J ; u, 1) = E(Θ, u) · (u − 1)d−k−dim Θ and Yj∈J u − 1 uaj+1 − 1 · E(D◦ J ; u, 1) =Yj∈J u − 1 uh−m,vj i − 1 · E(Θ, u) · (u − 1)d−k−dim Θ = =(−1)dim σR(σ, −m, u) · E(Θ, u) · (u − 1)d−dim Θ. Using 4.5, we obtain (−1)dim σR(σ, −m, u) = R(σ◦, m, u) and R(σ◦, m, u) = R(σΘ ◦ , m, u) = (−1)dim σΘ R(σΘ, −m, u). σ◦⊆σΘ ◦ Xσ∈bΣ Estr(X; u, 1) = XΘ(cid:22)∆ = XΘ(cid:22)∆ Since dim σΘ = d − dim Θ, we conclude E(Θ, u) · (u − 1)d−dim Θ · Xσ∈bΣ E(Θ, u) · (1 − u)d−dim ΘR(σΘ, −m, u). σ◦⊆σΘ ◦ R(σ◦, m, u) = (cid:3) Theorem 4.11. Let ∆ ⊂ MR be an arbitrary d-dimensional almost pseudoreflexive polytope. Denote by X a canonical Calabi-Yau model of a ∆-nondegenerate affine hypersurface Z∆ ⊂ Td in the d-dimensional algebraic torus Td. Then est(X) = XΘ(cid:22)∆ dim Θ≥1 (−1)dim Θ−1v(Θ) · v(σΘ ∩ ∆∗), where σΘ is the cone in the normal fan of the polytope ∆ and ∆∗ ⊂ NR is the polar polytope of ∆. 24 VICTOR BATYREV Proof. One has est(X) = lim u→1 Est(X; u, 1) = XΘ(cid:22)∆ dim Θ≥1 It remains to apply Corollary 4.7 E(Θ, 1) lim u→1 R(σΘ, u) · (1 − u)d−dim Θ. and Proposition 4.5 E(Θ, 1) = (−1)dim Θ−1v(Θ) R(σΘ, u) · (1 − u)d−dim Θ = v(σΘ ∩ ∆∗). lim u→1 (cid:3) Remark 4.12. It is easy to see that the formula for the stringy Euler number in Theorem 4.11 is a generalization of the formula (1.1) in the case when ∆ is a reflexive polytope. If Θ ≺ ∆ is a (d − k)-dimensional face of reflexive polytope ∆ then the k-dimensional polytope σΘ ∩ ∆∗ is a lattice pyramid with height 1 over the (k − 1)-dimensional dual face Θ∗ of the polar reflexive polytope ∆∗. Therefore, v(σΘ ∩ ∆∗) = v(Θ∗). On the other hand, the d-dimensional reflexive polytope ∆ is the union of d-dimensional pyramids over all (d − 1)-dimensional faces Θ ≺ ∆. So we have v(∆) = XΘ(cid:22)∆ dim Θ=d−1 v(Θ). Thus, we obtain (−1)dim Θ−1v(Θ) · v(σΘ ∩ ∆∗) = XΘ(cid:22)∆ dim Θ≥1 d−2Xk=1 (−1)k−1 XΘ(cid:22)∆ dim Θ=k v(Θ) · v (Θ∗) . We consider below several examples illustrating applications of our formula to non-reflexive polytopes ∆. Example 4.13. The Newton polytope ∆ of a general 3-dimensional quintic X in P4 containing the point (1 : 0 : 0 : 0 : 0) ∈ P4 is the almost reflexive polytope ∆ = {(x1, x2, x3, x4) ∈ R4 ≥0 : 1 ≤ x1 + x2 + x3 + x4 ≤ 5.}, which is a set-theoretic difference of two 4-dimensional simplices. The Fine interior of ∆ consists of the single lattice point p = (1, 1, 1, 1). The integral distance between p and the 3-dimensional face Θ defined by the equation x1 + x2 + x3 + x4 = 1 equals 3. Therefore, ∆ is not a reflexive polytope. One has v(∆) = 54 − 14 = 624. The polytope ∆ has 6 faces Θ of codimension 1 (4 faces Θ with v(Θ) = 53 − 13 = 124, one face Θ with v(Θ) = 53 and one face Θ with v(Θ) = 13). There also 14 faces of dimension 2 and 16 faces of dimension 1 in ∆. Our formula for the stringy Euler number gives: estr(X) = − (54 − 14) + 4(53 − 13) + 53 + 13 · 1 3 −4 · 52 − 6 · (52 − 12) − 4 · 12 · 1 3 + 6 · 5 + 4 · (5 − 1) + 6 · 1 · 1 3 = −200. This is a well-known fact, since X is a smooth quintic 3-fold. THE STRINGY EULER NUMBER AND THE MAVLYUTOV DUALITY 25 Example 4.14. The Newton polytope ∆ of a general 3-dimensional quintic in P4 having an isolated quadratic (conifold) singularity at point x = (1 : 0 : 0 : 0 : 0) ∈ P4 is the almost pseudoreflexive polytope ∆ = {(x1, x2, x3, x4) ∈ R4 ≥0 : 2 ≤ x1 + x2 + x3 + x4 ≤ 5, } which is again a set-theoretic difference of two 4-dimensional simplices. The Fine interior of ∆ consists of the single lattice point p = (1, 1, 1, 1). The integral distance between p and the 3-dimensional face Θ defined by the equation x1 +x2 +x3 +x4 = 2 equals 2. Therefore, ∆ is not a reflexive polytope. One has v(∆) = 54 − 24 = 609. The polytope ∆ has 6 faces Θ of codimension 1 (4 faces Θ with v(Θ) = 53−23 = 117, one face Θ with v(Θ) = 53 and one face Θ with v(Θ) = 23). There also 14 faces of dimension 2 and 16 faces of dimension 1 in ∆. Our formula for the stringy Euler number gives: estr(X) = − (54 − 24) + 4(53 − 23) + 53 + 23 · 1 2 −4 · 52 − 6 · (52 − 22) − 4 · 22 · 1 2 + 6 · 5 + 4 · (5 − 2) + 6 · 2 · 1 2 = −198. Unfortunatly, this singular Calabi-Yau 3-fold X does not have a projective small resolution of its singularity. So X has no a smooth projective birational Calabi-Yau model. Remark 4.15. If a log-desingularization ρ : Y → X of a projective variety X contains only one smooth exceptional divisor D such that KY = ρ∗KX + aD, then est(X) = e(Y ) + e(D)(cid:18) −a a + 1(cid:19) . In particular, est(X) is not an integer, if (a + 1) does not divide e(D). Example 4.16. One can generalize Example 4.14 and compute the stringy Euler number of a general d-dimensional Calabi-Yau hypersurface X ′ ⊂ Pd+1 of degree d + 2 with a single quadratic singularity at point x = (1 : 0 : · · · : 0). The blow up of this singular point is a desingularization ρ : cX ′ → X ′ such that the exceptional divisor D ⊂ cX ′ is isomorphic to a (d − 1)-dimensional quadric. One has By 4.15, we obtain KcX ′ = KX ′ + (d − 2)D. estr(X ′) = e(cX ′) − d − 2 d − 1 e(D). Therefore, the local stringy Euler number of the singular point x ∈ X ′ equals estr(X ′, x) = e(D) − d − 2 d − 1 e(D) = e(D) d − 1 If the dimension d ≥ 4 is an even number then e(D) = d and estr(X ′) = c for some coprime numbers c, d − 1. In particular, estr(X ′) is not an integer. d−1 ∈ Q \ Z So far no mirror manifolds have been known for singular Calabi-Yau varieties X with non-integral stringy Euler number est(X) ∈ Q \ Z. 26 VICTOR BATYREV 5. Calabi-Yau hypersurfaces in P(a, 1, . . . , 1) Let a, b ∈ N be two integers a, b ≥ 2. We put d := ab + l for some integer 1 ≤ l ≤ a − 1 and consider Calabi-Yau hypersurfaces of degree a + d in the weighted projective space P(a, 1d) := P(a, 1, . . . , 1 ) d {z } of dimension d ≥ 5. The space of quasihomogeneous polynomials of degree a + d in d+1 variables z0, z1, . . . , zd (deg z0 = a, deg zi = 1, 1 ≤ i ≤ d) has the monomial basis · · · zmd 0 zm1 zm0 ≥0 satisfying 1 the condition d determined by the lattice points (m0, m1, . . . , md) ∈ Zd+1 The convex set am0 + m1 + · · · md = a + d. Sd := {(x0, x1, . . . , xd) ∈ Rd+1 ≥0 : ax0 + x1 + · · · xd = a + d}. {z }i is a d-dimensional simplex having the unique interior lattice point p := (1, . . . , 1) ∈ Zn+1, d integral vertices νi = (0, . . . , a + d , . . . , 0) 1 ≤ i ≤ d and one rational vertex ν0 := ( a+d a , 0, . . . , 0). The convex hull of the set Sd ∩ Zd+1 is the lattice polytope ∆ which is the intersection of the simplex Sd with the half-space define by the inequality x0 ≤ b + 1. The lattice polytope ∆ has 2d vertices, first d vertices of ∆ belong to the hyperplane x0 = 0 and the remaining d vertices of ∆ belong to the hyperplane x0 = b + 1. The lattice polytop ∆ is not reflexive, because the integral distance between its single interior lattice point p and the (d − 1)-dimensional face in the hyperplane x0 = b + 1 is equal to b ≥ 2. However, it is easy to show that ∆ is a pseudoreflexive polytope. Its dual pseudoreflexive polytope ∆∨ = [∆∗] is a d-dimensional lattice simplex whose vertices v0, v1, . . . , vd ∈ N satisfy the relation i=1 vi = 0. The polar polytope (∆∨)∗ can be identified with the rational d-dimensional simplex Sd ⊂ Rd+1 ≥0 lattice vectors v0, v1, . . . , vd ∈ N are exactly the generators of all 1-dimensional cones in the d-dimensional fan describing the weighted projective space P(a, 1d) as a d-dimensional toric variety. in the hyperplane ax0 +Pi=1 xi = a + d. The av0 +Pd The combinatorial structure of the pseudoreflexive polytope ∆ is rather simple, because ∆ is combinatorially equivalent to the product of (d − 1)-dimensional and 1-dimensional simplices. Its polar polytope ∆∗ is simplicial and it has d + 2 vertices: the lattice vertices v0, v1, . . . vd and the rational vertex vd+1 := − 1 i=1 vi. The vertices v0, v1, . . . , vd−1 can be chosen as a basis of the lattice N. b v0 = 1 abPd Remark 5.1. If l = 1, i.e., d = ab + 1 then the weighted projective space P(a, 1d) contains d different quasi-smooth Calabi-Yau hypersurfaces Xi ⊂ P(a, 1d) (1 ≤ i ≤ d) defined respectively by the invertible polynomials Fi(z0, z1, . . . , zd) := za+d 1 + · · · + za+d d + zb+1 0 zi, 1 ≤ i ≤ d, such that Berglund-Hübsch-Krawitz mirror construction can be applied to every hypersurface Xi (1 ≤ i ≤ d) [BHü93, Kra09]. THE STRINGY EULER NUMBER AND THE MAVLYUTOV DUALITY 27 Proposition 5.2. Assume that l = 1. Then the stringy Euler number estr(X) of a general quasi-smooth Calabi-Yau hypersurface X in P(a, 1d) equals a − 1 a + (−1)d−2 (−1)i d d−2Xi=0 i!(a + d)d−1−i + (−1)d−1 (−1)i n d−1Xi=0 i! (a + d)d−i a . Proof. The polytope ∆ is the difference of two d-dimensional simplices. Therefore, all proper faces Θ of ∆ are either simlices or differences of two simplices. For any k−1(cid:17) simplicial (d − k)-dimensional faces of ∆: 1 ≤ k ≤ d − 1 there exist exactly 2(cid:16) d k−1(cid:17) of these (d − k)-dimensional simplicial faces are contained in the hyperplane (cid:16) d x0 = b + 1 and the other (cid:16) d k−1(cid:17) simplicial (d − k)-dimensional faces of ∆ contained in the hyperplane x0 = 0. There exist exactly(cid:16)d k(cid:17) nonsimplicial (d − k)-dimensional faces Θ of ∆ which are differencies of two simplices. A (d − k)-dimensional face Θ ≺ ∆ is singular if and only if it is contained in the hyperplane x0 = b + 1, and in this case Θ∗ is a simplex with the rational vertex vd+1 and for the corresponding rational polytope σΘ ∩ ∆∗ one has v(σΘ ∩ ∆∗) = 1/b. For all regular (d − k)- dimensional faces Θ ≺reg ∆ one has v(σΘ ∩ ∆∗) = 1. Now we can apply the formula (4.11) for computing the stringy Euler number of a generic Calabi-Yau hypersurface X ⊂ P(a, 1d): v(Θ) · v(σΘ ∩ ∆∗) = 1 a!! + − a dim Θ=d−k a b 1 − est(X) = (−1)d−1v(∆) + 1! (a + d)d−1 d−1Xk=1 (−1)d−1−k XΘ≺∆ =(−1)d−1 (a + d)d a! + 0!(a + d)d−1 + d + d 0! 1 +(−1)d−2 d k − 1! 1 · · · + (−1)d−1−k d k − 1!(a + d)d−k + d + d d − 2! 1 +(−1)0 d d − 1! a + d d − 2!(a + d) + d + d i! + (−1)d 1 i!+ (−1)i d (−1)i d d−1Xi=0 d−2Xi=0 =(−1)d−2 1 b i! (a + d)d−i (−1)i d i!(a + d)d−1−i + (−1)d−1 (−1)i d -.2Xi=0 d−1Xi=0 (−1)i n i!(a + d)d−1−i + (−1)d−1 (−1)i d d−1Xi=0 d−2Xi=0 k! (a + d)d−k a!! = + (−1)d−2 +(−1)d−2 =a − 1 − 1 a a a − a b b 1 a!! + · · · = a i! (a + d)d−i a . (cid:3) Proposition 5.3. For any l (1 ≤ l ≤ a − 1) the stringy Euler number estr(X ∨) of a canonical Calabi-Yau model X ∨ of a ∆∨-nondegenerated affine hypersurface in Td 28 VICTOR BATYREV defined by the Laurent polynomial equals (−1)d−1(cid:18)a − 1 a(cid:19) − d−2Xi=1 (−1)i d f (x) = x−a d xi x−1 i + d−1Yi=1 dXi=1 i!(a + d)d−i−1 + (−1)i d d−1Xi=2 i! (a + d)d−i a . Proof. For the dual pseudoreflexive polytope ∆∨ one has v(∆∨) = a + d. All faces Θ (cid:22) ∆∨ are lattice simplices. A (n − k)-dimensional face Θ ≺ ∆∨ is singular if and only if it is contained in the (d − 1)-simplex with vertices v1, . . . , vd. Now we apply the formula (4.11) for computing the stringy Euler number of the canonical Calabi-Yau model X ∨: est(X ∨) =(−1)d−1v(∆∨) + v(Θ) · v(σΘ ∩ (∆∨)∗) = 1 dim Θ=d−k d−1Xk=1 (−1)d−1−k XΘ≺∆∨ a(cid:19) + d − 1! (a + d) ! + ! + d − 2!(a + d)2 d − 3!(a + d)3 ! + · · · ! = 2!(a + d)d−2 (−1)d−1(a + d) + (−1)d−2(cid:18)d + d − 2!(a + d) + d +(−1)d−3 d d − 3!(a + d)2 + d +(−1)d−4 d d − 4!(a + d)3 + d +(−1)d−5 d 1!(a + d)d−2 + d +(−1)0 d =(−1)d−1(cid:18)a − (−1)i d d−2Xi=1 a(cid:19) − i!(a + d)d−i−1 + (−1)i d d−1Xi=2 i!(a + d)d−i a a a a a 1 − . (cid:3) Corollary 5.4. If l = 1 then one has est(X) = (−1)d−1est(X ∨). Proof. Comparing the formulas for est(X) and est(X ∨) in 5.2 and 5.3, we see that (−1)d−1est(X ∨) − (d + a)d−1 + (a + d)d a Therefore, we get − d 1!(d + a)d−1 a (−1)d−1est(X ∨) = est(X). = est(X). (cid:3) THE STRINGY EULER NUMBER AND THE MAVLYUTOV DUALITY 29 v0, v1, . . . , vd satisfying the relation av0 +Pd The weighted projective space V = P(a, 1d) is a toric variety defined by a sim- plicial d-dimensional fan whose 1-dimensional cones a generated by lattice vectors i=1 vi = 0. It is easy to show that again the convex hull of {v0, v1, . . . , vd} is a pseudoreflexive simplex ∆∨ which is dual to ∆. There is a toric desingularization ρ : V ′ → V having one exceptional divisor E ∼= Pd−1 that corresponds to the lattice point −v0 so that the set of lattice vectors {−v0, v0, v1, . . . , vd} can be identified with the set of inner normal vectors to facets of ∆. The toric desingularization ρ : V ′ → V induces a desingularization ρ : X ′ → X of the generic Calabi-Yau hypersurface X ⊂ V such that the fiber D := ρ−1(p) over the unique singular point x ∈ X is isomorphic to a generic hypersurface of degree k in Pd−1. One has KX ′ = ρ∗KX + bD, i=1 Rvi with ϕ(v1) = . . . = ϕ(vd) = 1 because the linear function ϕ on the cone Pd has value b + 1 on the lattice vector −v0 = 1/aPd i=1 vi. Theorem 5.5. Let X be a generic ∆-nondegenerated Calabi-Yau hypersurface in the d-dimensional weighted projective space P(a, 1d) (d = ab+l, l ≥ 2) and let X ∨ be a canonical Calabi-Yau model of the ∆∨-nondegenerated affine hypersurface Z ⊂ Td defined by the Laurent polynomial f (x) = x−a d x−1 i + d−1Yi=1 xi. dXi=1 Then estr(X) ∈ 1 integer. In particular, the equality b Z and estr(X ∨) ∈ 1 a Z. Moreover, if l = 2 then estr(X) is not an estr(X) = (−1)d−1estr(X ∨) can not be satisfied if l = 2 and (a, b) is a pair of distinct odd prime numbers. Proof. We compute the stringy Euler number of a generic Calabi-Yau hypersurface X ⊂ P(a, 1d) as in 5.2: ld a! + − b a est(X) = (−1)d−1 (a + d)d 0!ld−1 1 0!(a + d)d−1 + d + d 1!(d + a)d−2 + d + d 1!ld−2 1 2!ld−3 1 2!(d + a)d−3 + d + d b b +(−1)d−2 n +(−1)d−3 n +(−1)d−4 d − a 1! (d + a)d−1 2! (d + a)d−2 3! (d + a)d−3 a a − − ld−2 ld−1 a !! a !! + a !! + ld−3 +(−1)0 d d − 2!l 1 b + d d − 2!(d + a) + d d − 1! d + a a l a!! . − · · · 30 VICTOR BATYREV Since a divides (d + a)i − li = (ab + a + l)i − li for any i ∈ N, we obtain that estr(X) ∈ 1 b Z. The terms in estr(X) having the denominator b sum up to A := (−1)d−2 1 lb (−1)i d d−2Xi=0 i!ld−i = (−1)d−2 1 lb(cid:16)(l − 1)d − (−1)d − (−1)d−1dl(cid:17) . In particular, for l = 2 and odd integers a, b the dimension d = ab + 2 is odd, A = 1 − d b 6∈ Z and estr(X) is not an integer. On the other hand, we did already the computation for estr(X ∨) in 5.3 and ob- tained estr(X ∨) = (−1)d−1(cid:18)a − 1 a(cid:19) − (−1)i d d−2Xi=1 i!(a + d)d−i−1 + d−1Xi=2 (−1)i d i!(a + d)d−i a . This shows that estr(X ∨) ∈ 1 a Z. Therefore, if a and b two distinct odd prime numbers the equality estr(X) = (−1)d−1estr(X ∨) can hold only if the stringy Euler numbers are integers, but for l = 2 this is not the case. (cid:3) 6. An additional condition on singular facets Let ∆ ⊂ MR be a d-dimensional pseudoreflexive polytope and let ∆∨ := [∆∗] be the Mavlyutov dual pseudoreflexive polytope. We consider also two additional d-dimensional almost pseudoreflexive polytopes ∆1 ⊆ ∆ and ∆2 ⊆ ∆∨ such that one has the inclusions ∆1 ⊆ [∆can 1 ∆2 ⊆ [∆can ] = ∆, ] = ∆∨. 2 A generalization of Berglund-Hübsch-Krawitz mirror construction suggested by Artebani, Comparin and Guilbot [ACG16] needs an additional condition that guar- antees that the Zariski closure of an affine ∆1-nondegenerated hypersurface Z1 in the toric variety P∆∗ 2 will be quasi-smooth. The same condition is demanded for the Zariski closure of a ∆2- associated with the nondegenerated affine hypersurface Z2 in the toric variety P∆∗ rational polar polytope ∆∗ 1. The quasi-smoothness condition implies that the singu- larities of Calabi-Yau hypersurfaces are locally quotient singularities. In particular, the stringy Euler number of such Calabi-Yau hypersurfaces is always an integer. 2 associated with the rational polar polytope ∆∗ 1 In [Bor13, Def. 7.1.1, Prop. 7.1.3] Borisov suggested to generalize the quasi- smoothness condition using some versions of Jacobian rings. It is not quite clear how Borisov's condition can be described by purely combinatorial properties of convex polytopes, but it is satisfied in two cases: 1) for reflexive polytopes and 2) for almost pseudoreflexive simplices ∆1 and ∆2 that appear in the Berglund-Hübsch-Krawitz mirror construction. Our purpose is to describe a new another condition on Calabi-Yau varieties X and X ∨ that must be added to the Mavlyutov duality for pairs of d-dimensional pseudoreflexive polytopes ∆ and ∆∨ such that the stringy Euler numbers estr(X) and estr(X ∨) will be integers satisfying the equation estr(X) = (−1)d−1estr(X ∨). THE STRINGY EULER NUMBER AND THE MAVLYUTOV DUALITY 31 We remark that a pseudoreflexive lattice polytope ∆ is not reflexive if and only if there exists at least one singular facet Θ ≺sing ∆. Our additional condition on a pseudoreflexive polytope ∆ is exactly an additional condition on its singular facets of ∆. By 3.17, there exist a natural bijection between singular facets Θ of pseudoreflexive polytope ∆ and non-integral vertices νΘ of the polar polytope ∆∗. Let Z ⊂ Td be a ∆∨-nondegenerated hypersurface and let X ∨ be its canonical Zariski closure in the toric Q-Fano variety P∆∗ corresponding to the polar polytope ∆∗. Then for any singular facet Θ ≺sing ∆ the Calabi-Yau hypersurface X ∨ ⊂ P∆∗ contains the torus fixed point xΘ ∈ X ∨ corresponding to the rational vertex νΘ ∈ ∆∗. Definition 6.1. Let Θ ≺sing ∆ be a singular facet of a d-dimensional pseudoreflexive polytope ∆. Denote by nΘ (nΘ ≥ 2) the integral distance from 0 ∈ M to the facet Θ. We call the facet Θ quasi-regular if the local stringy Euler number estr(X ∨, xΘ) is an integer that can be computed by the formula: estr(X ∨, xΘ) = nΘ · v(Θ). We illustrate this definition with the examples of the Mavlyutov pairs (∆, ∆∨) from the previous section. Pd−1 Example 6.2. We consider ∆∨ to be a d-dimensional pseudoreflexive simplex which is the convex hull of a basis e1, . . . , ed of the lattice M and a point e0 = −aed + i=1 , where a does not divide d and a ≤ d/2. Let d = ab + l for some integers 1 ≤ l < a and b ≥ 2. Then the dual pseudoreflexive polytope ∆ = (∆∨)∨ is the Newton polytope of a Calabi-Yau hypersurface X of degree a + d in the d- dimensional weighted projective space P(a, 1d) that contains a torus fixed point x := (1 : 0 : . . . : 0). A desingularization of P(a, 1d) at x contains a single exceptional divisor isomorphic to Pd−1. The induced birational morphism ρ : Y → X has a single exceptional divisor D ⊂ Pd−1 which is a hypersurface of degree l and one has KY = ρ∗KX + (b − 1)D. By 4.15, we obtain estr(X, x) = e(D)/b. On the other hand, we have nΘ · v(Θ) = a, where Θ is a singular facet with vertices v1, . . . , vd of ∆∨ corresponding to the point x. The equality estr(X, x) = nΘ · v(Θ) is equivalent to e(D) = ab = d−l. This can happen for a smooth (d−2)-dimensional hypersurface D of degree l in Pd−1 if and only if l = 1, i.e., only if X is a quasi-smooth hypersurface in P(a, 1d). Theorem 6.3. Let (∆, ∆∨) be a Mavlyutov pair of d-dimensional pseudoreflexive polytopes ∆ and ∆∨. Assume all singular facets Θ′ ≺sing ∆∨ are quasi-regular. Then the stringy Euler number estr(X) of a canonical Calabi-Yau model X of a ∆-nondegenerate hypersurface can be computed by the following formula: XΘ′≺sing∆∨ dim Θ′=d−1 nΘ′ · v(Θ′) + XΘ≺reg∆ 1≤dim Θ≤d−2 (−1)dim Θ−1v(Θ) · v(Θ∨) + (−1)d−1 XΘ≺sing∆ dim θ=d−1 nΘ · v(Θ). In particular, if all singular facets of ∆ are also quasi-regular then for canonical Calabi-Yau models X ∨ of a ∆∨-nondegenerate hypersuface one obtains the equality estr(X) = (−1)d−1estr(X ∨). Proof. Let Z ⊂ Td be a ∆-nondegenerate affine hypersurface. There are two pro- jective closures of Z: the closure Z in the toric variety P∆ and the canonical model 32 VICTOR BATYREV X obtained as the Zariski closure of Z in the toric Q-Fano variety corresponding to the rational polytope ∆can = (∆∨)∗ ⊂ MR. We choose a regular simplicial fan bΣ which is a common subdivision of two rational polyhedral fans: the normal fan Σ∆ and the normal fan Σ∆can. So we obtain two birational toric morhisms ρ1 and ρ2: PbΣ f ρ1 ~⑥⑥⑥⑥⑥⑥⑥⑥ ρ2 "❊❊❊❊❊❊❊❊ P∆ /❴❴❴❴❴❴❴ P∆can together with the induced birational morphisms bZ f ρ1 ρ2 ❃❃❃❃❃❃❃❃ Z /❴❴❴❴❴❴❴ X The canonical Calabi-Yau hypersurface X ⊂ P∆can is a disjoint union of locally closed strata XF := X ∩ TF where TF is a torus orbit in the projective toric variety P∆can and F runs over all faces F (cid:22) ∆can of the rational polytope ∆can: X = [F (cid:22)∆can XF . Let v1, . . . , vs be the set of primitive lattice generators of 1-dimensional cones in the fan bΣ. We set I := {1, . . . , s}. Then k-dimensional cone σ ∈ bΣ is determined by a subset J ⊂ I such that J = k and σ is generated by vj (j ∈ J). For any face F (cid:22) ∆can we define the stringy Euler number estr(X, XF ) := X∅⊆J⊆I e(D◦ J ∩ ρ−1 2 (TF ))Yj∈J 1 aj + 1 , where D◦ J are either empty or a locally closed stratum on the smooth projective hypersurface bZ in the toric variety PbΣ corresponding to a cone σ ∈ bΣ of dimension J. By the additivity of the Euler number, we obtain estr(X) = XF (cid:22)∆can estr(X, XF ). So it remains to compute estr(X, XF ) for any face F (cid:22) ∆can. We consider the following 4 possibilities for a face F (cid:22) ∆can: • dim[F ] = dim F = k ≥ 1, i. e., F = Θ∗ for some regular (d − k − 1)- dimensional face Θ (cid:22) ∆∨ • dim[F ] < dim F = k ≥ 1, i. e., F = Θ∗ for some singular (d − k − 1)- dimensional face Θ (cid:22) ∆∨ • dim[F ] = dim F = 0, i. e., F = Θ∗ is a lattice vertex of ∆can corresponding to some regular (d − 1)-dimensional face Θ (cid:22) ∆∨ • dim F = 0 and [F ] = ∅, i. e., F = Θ∗ is a rational vertex of ∆can corre- sponding to some singular (d − 1)-dimensional face Θ (cid:22) ∆∨. ~ " /  / THE STRINGY EULER NUMBER AND THE MAVLYUTOV DUALITY 33 If dim[F ] = dim F = k ≥ 1, then [F ] = Θ for some k-dimensional face Θ (cid:22) ∆. For a generic ∆-nondegenerate hypersurface Z the affine hypersuface XF ⊂ TF is [F ]-nondegenerate and its Euler number equals (−1)k−1v([F ]) (see 4.7). Moreover, X has Gorenstein toroidal singularities along XF corresponding to the (d − k)- dimensional cone over the dual regular (d − k − 1)-dimensional face T heta∨ of ∆. So one has estr(X, XF ) = (−1)k−1v(Θ) · v(Θ∨). If dim[F ] < dim F = k ≥ 1, then the affine hypersuface XF ⊂ TF is isomophic to a product of (C∗)k−dim[F ] and [F ]-nondegenerated affine hypersurface. Therefore e(XF ) = 0 and one has estr(X, XF ) = 0. If dim[F ] = dim F = 0, then XF is empty and one has estr(X, XF ) = 0. If dim F = 0 and [F ] = ∅, then XF is a torus fixed point xΘ′ ∈ P∆can estr(X, XF ) equals to local stringy Euler number estr(X, xΘ) = n(Θ) · v(Θ) for some singular facet Θ ≺sing ∆∨. Thus, we obtain estr(X) = XΘsing≺∆∨ dim Θsing=d−1 nΘ · v(Θ) + dXk=1 (−1)k−1 XΘord≺∆ dim Θord=k v(Θord) · v(Θ∗ ord). Since the d-dimensional lattice polytope ∆ is the union over all (d−1)-dimensional faces Θ ≺ ∆ of d-dimensional pyramids ΠΘ := Conv(0, Θ) with vertex 0, one has v(∆) = XΘ≺∆ dim Θ=d−1 v(ΠΘ). On the other hand, v(ΠΘ) = v(Θ) · nΘ, where nΘ is the integral distance from Θ to 0 ∈ M. The equality nΘ = 1 holds if and only if Θ ≺ ∆ is a regular (d − 1)- dimensional face. This implies the equality dXk=d−1 (−1)k−1 XΘ≺reg∆ dim Θ=k v(Θ) · v(Θ∨) =(−1)d−1v(∆) − XΘ≺reg∆ v(Θ) = v(Θ) · nΘ =(−1)d−1 XΘ≺sing∆ dim Θ=d−1 dim Θ=d−1 that proves the required formula for estr(X). The equality estr(X) = (−1)d−1estr(X ∨) follows now from the duality Θ ↔ Θ∨ between k-dimensional regular faces Θ ≺reg ∆ and (d − k − 1)-dimensional regular faces Θ∨ ≺reg ∆∨ and from the equality d−2Xk=1 (−1)k−1 XΘ≺reg∆ dim Θ=k v(Θ) · v(Θ∨) = (−1)d−1 d−2Xk=1 (−1)k−1 XΘ∨≺reg∆∨ dim Θ∨=k v(Θ) · v(Θ∨) (cid:3) 34 VICTOR BATYREV References [Amb03] F. Ambro, Inversion of adjunction for non-degenerated hypersufaces, Manuscripta Math. 111 (2003), 43-49. [ACG16] M. Artebani, P. Comparin, R. Guilbot, Families of Calabi-Yau hypersurfaces in [AP15] [Bat93] [Bat94] [Bat98] [Bat99] [BB96] [BB97] [BD96] [BG18] [BS17] [BHü93] [BHü16] [BiH14] [Bor13] [CR11] [Cla14] [CG11] [CLS11] [DR01] [DKh86] [Fine83] [Fu03] [FS04] Q-Fano toric varieties, J. Math. Pures Appl. (9) 106 (2016), no. 2, 319–341. P. S. Aspinwall, M. R. Plesser General mirror pairs for gauged linear sigma models, JHEP 11 (2015) 029, arXiv:1507.00301. V. V. Batyrev, Variations of the mixed Hodge structure of affine hypersurfaces in algebraic tori, Duke Math. J. 69 (1993), no. 2, 349–409. V. V. Batyrev, Dual Polyhedra and Mirror Symmetry for Calabi-Yau Hypersurfaces in Toric Varieties, Journal of Algebraic Geometry 3 (1994), no. 3, 493–535. V. V. Batyrev, Stringy Hodge numbers of varieties with Gorenstein canonical singu- larities, Integrable systems and algebraic geometry (Kobe/Kyoto 1997), World Sci. Publ., River Edge, NJ (1998), 1–31. V. V. Batyrev, Birational Calabi–Yau n-folds have equal Betti numbers, in: New Trends in Algebraic Geometry, Warwick, 1996, 1999, pp. 1–11. V. V. Batyrev and L. A. Borisov, Mirror duality and string-theoretic Hodge numbers, Invent. Math. 126 (1) (1996) 183–203. V. V. Batyrev and L. A. Borisov, Dual cones and mirror symmetry for generalized Calabi-Yau manifolds, Mirror Symmetry II, AMS/IP Stud. Adv. Math. 1, Amer. Math. Soc., Providence, RI (1997), 71–86. V. V. Batyrev and D. I. Dais, Strong McKay correspondende, string-theoretic Hodge numbers and mirror symmetry, Topology 35 (1996), no. 4, 901–929. V. V. Batyrev, G. Gagliardi, On the algebraic stringy Euler number, Proc. Amer. Math. Soc. 146 (2018), no. 1, 29–41. V. V. Batyrev, K. Schaller, Stringy Chern classes of singular toric varieties and their applications, Communications in Number Theory and Physics 11 (2017), 1–40. P. Berglund, T. Hübsch, A generalized construction of mirror manifolds, Nucl. Phys. B 393(1993), 377–391. P. Berglund, T. Hübsch, A generalized construction of Calabi-Yau models and mirror symmetry, arXiv:1611.10300. C. Birkar, Z. Hu, Polarized pairs, log minimal models, and Zariski decompositions, Nagoya Math. J. 215 (2014), 203–224. L. A. Borisov, Berglund–Hübsch mirror symmetry via vertex algebras, Commun. Math. Phys. 320 (1) (2013) 73–99. A. Chiodo, Y. Ruan, LG/CY correspondence: the state space isomorphism, Adv. Math. 227 (2011), no. 6, p. 2157–2188. P. Clarke, A proof of the birationality of certain BHK-mirrors, Complex Manifolds 1 (2014), 45–51. A. Corti, V. Golyshev, Hypergeometric equations and weighted projective spaces, Sci. China Math. 54 (8) (2011) 1577–1590. D. A. Cox, J. B. Little and H. K. Schenck, Toric varieties, Graduate Studies in Math- ematics, 124, Amer. Math. Soc., Providence, RI, 2011. D. Dais, M. Roczen, On the String-Theoretic Euler Number of 3-dimensional A-D-E Singularities, Advances in Geometry 1 (2001), 373–426. V. I. Danilov, A. G. Khovanski, Newton polyhedra and an algorithm for calculating Hodge-Deligne numbers, Izv. Akad Nauk SSSR Ser Mat 50 (1986), 925–945. J. Fine, Resolution and Completion of Algebraic Varieties, Ph.D. University of War- wick 1983. O. Fujino, Notes on toric varieties from Mori theoretic viewpoint, Tohoku Math. J. 55 (2003), 551–564. O. Fujino, H. Sato, Introduction to the toric Mori theory, Michigan Math. J. 52 (2004), no. 3, 649–665. THE STRINGY EULER NUMBER AND THE MAVLYUTOV DUALITY 35 [Ful93] [HKP06] M. Hering, A. Küronya, S. Payne, Asymptotic cohomological functions of toric divisors, W. Fulton, Introduction to Toric Varieties, Princeton University Press, 1993. [Ish99] [Kas10] [Ke13] [Kho78] [Kho97] [Ko13] [Kra09] [KS98] [KS00] [Kr08] [Mat02] [Mav05] [Mav11] [Mav13] [Na94] [OP91] Adv. in Math. 207 (2006), 634–645. S. Ishii, The minimal model theorem for divisors of toric varieties, Tohoku Math. J. (2) 51 (1999), no. 2, 213–226. A. M. Kasprzyk, Canonical toric Fano threefolds, Can. J. Math. 62 (6) (2010) 1293– 1309. T. Kelly, Berglund-Hübsch-Krawitz Mirrors via Shioda Maps, Adv. Theor. Math. Phys. 17 (2013), 1425–1449. A. G. Khovanskii, Newton polyhedra and the genus of complete intersections, Funct. Anal. Appl. 12 (1978), 38–46. A. G. Khovanskii, Newton polytopes, curves on toric surfaces, and inversion of Weil's theorem, Russian Math. Surveys 52 (1997), 1251–1279. J. Kollár, Singularities of the minimal model program, Cambridge Tracts in Mathe- matics 200, Cambridge University Press, 2013. M. Krawitz, FJRW rings and Landau-Ginzburg Mirror Symmetry, Thesis (Ph.D.)- University of Michigan, ProQuest LLC, Ann Arbor, MI, 2010, arXiv:0906.0796. M. Kreuzer and H. Skarke, Classification of reflexive polyhedra in three dimensions, Adv. Theor. Math. Phys. 2 (4) (1998) 853–871. M. Kreuzer and H. Skarke, Complete classification of reflexive polyhedra in four di- mensions, Adv. Theor. Math. Phys. 4 (2000), 1209–1230. M. Kreuzer, On the Statistics of Lattice Polytopes, arXiv:0809.1188. K. Matsuki, Introduction to the Mori program, Universitext, Berlin, New York (2002). A. A. Mavlytov, Embedding of Calabi-Yau deformations into toric varieties, Math. Ann., 333 (2005), 45–65. A. A. Mavlytov, Mirror Symmetry for Calabi-Yau complete intersections in Fano toric varieties, arXiv:1103.2093. A. A. Mavlytov, Private communication, January 2013. Y. Namikawa, On deformations of Calabi-Yau threefolds with terminal singularities, Topology 33 (1994), 429–446. T. Oda, H. S. Park, Linear Gale transforms and Gelfand-Kapranov-Zelevinskij decom- position, Tohoku Math. J., II Ser., 43 (1991), 375–399. [Pum11] M. Pumperla, Unifying constructions in toric mirror symmetry, Diss. Hamburg (2011) http://ediss.sub.uni-hamburg.de/volltexte/2012/5837/pdf/Dissertation.pdf [Reid83] M. Reid, Decomposition of toric morphisms, in Arithmetic and geometry, Progr. Math. [Sh14] [Sk96] [Tr08] [Wi02] 36, Birkhäuser Boston, Boston, MA (1983), 395–418. M. Shoemaker, Birationality of Berglund-Hübsch-Krawitz mirrors, Commun. Math. Phys. 331 (2) (2014) 417–429. H. Skarke, Weight systems for toric Calabi-Yau varieties and reflexivity of Newton polyhedra. Modern Phys. Lett. A 11 (1996), no. 20, 1637–1652. J. Treutlein, 3-Dimensional Lattice Polytopes Without Interior Lattice Points, arXiv:0809.1787. J. Wiśniewski, Toric Mori theory and Fano manifolds, Séminair & Congés, 6 (2002), 249–272. Fachbereich Mathematik, Universität Tübingen, Auf der Morgenstelle 10, 72076 Tübingen, Germany E-mail address: [email protected]