Matrix Factorizations Of The Classical Discriminant
by
Bradford Hovinen
A thesis submitted in conformity with the requirements for the degree of Doctor of Philosophy Graduate Department of Mathematics University of Toronto
Copyright c 2009 by Bradford Hovinen Abstract
Matrix Factorizations Of The Classical Discriminant
Bradford Hovinen Doctor of Philosophy Graduate Department of Mathematics University of Toronto 2009
The classical discriminant Dn of degree n polynomials detects whether a given univariate polynomial f has a repeated root. It is itself a polynomial in the coefficients of f which, according to several classical results by B´ezout,Sylvester, Cayley, and others, may be expressed as the determinant of a matrix whose entries are much simpler polynomials in the coefficients of f. This thesis is concerned with the construction and classification of such determinantal formulae for Dn.
In particular, all of the formulae for Dn appearing in the classical literature are equivalent in the sense that the cokernels of their associated matrices are isomorphic as modules over the associated polynomial ring. This begs the question of whether there exist formulae which are not equivalent to the classical formulae and not trivial in the sense of having the same cokernel as the 1 × 1 matrix (Dn). The results of this thesis lie in two directions. First, we construct an explicit non-classical formula: the presentation matrix of the open swallowtail first studied by Arnol’d and Givental. We study the properties of this formula, contrasting them with the properties of the classical formulae.
4 2 Second, for the discriminant of the polynomial x + a2x + a3x + a4 we embark on a classification of determinantal formulae which are homogeneous in the sense that the cokernels of their associated matrices are graded modules with respect to the grading deg ai = i. To this end, we use deformation theory: a moduli space of such modules arises from the base spaces of versal deformations of certain
4 3 modules over the E6 singularity {x − y }. The method developed here can in principle be used to classify determinantal formulae for all discriminants, and, indeed, for all singularities.
ii Acknowledgements
The author greatly thanks his thesis advisor, Ragnar-Olaf Buchweitz, for all of his generous support over the past four years. He would like to thank Hubert Flenner, Nagat Karroum, Michael Kunte, Paul Cadman, David Mond, David Eisenbud, Greg Smith, Steve Kudla, Jan Christophersen, and Klaus Altmann for their valuable discussions. Thanks also go to Mike Stillman and Dan Grayson for their assistance with the use of Macaulay 2.
iii Contents
1 The Classical Discriminant 1
1.1 Introduction ...... 1
1.2 A local perspective: Unfoldings ...... 4
1.3 A global perspective: Hilbert schemes ...... 7
1.4 The geometry of ∆n ...... 8 1.5 Structure of this thesis ...... 11
2 Matrix Factorizations 13
2.1 Classical determinantal formulae ...... 14
2.2 Eisenbud’s Theorem ...... 20
2.3 The Cayley Method ...... 24
2.4 On the ranks of the presentation matrices of Σ¯ n,k ...... 35
3 The Open Swallowtail 38
3.1 An algebraic definition ...... 38
3.2 Construction of the presentation matrix ...... 42
3.3 The construction of Arnol’d ...... 51
3.4 The conductor of the open swallowtail ...... 52
3.5 Application to the root structure of a univariate polynomial ...... 56
4 Deformations of Modules 57
4.1 Deformation theory of modules ...... 58
4.2 Computing versal deformations ...... 60
4.2.1 Liftings and obstructions ...... 62
4.2.2 The first-order deformation ...... 65
4.2.3 Further deformations ...... 68
iv 4.3 Rank one MCM modules over xn − yn−1 ...... 71
4.4 Example: Graded modules over ∆4 ...... 76
4.4.1 A classification of rank one graded MCM modules over ∆˜ 4 ...... 86 4.5 Structure of Maximal Cohen-Macaulay Modules ...... 87
5 Conclusion 91
A Macaulay 2 package for Module Deformations 93
Bibliography 100
v Chapter 1
The Classical Discriminant
1.1 Introduction
Fix a field K of characteristic not equal to two and let
n n−1 n−2 f(x) := x + a1x + a2x + ··· + an−1x + an
be a monic polynomial of degree n ≥ 1 over K with roots α1, . . . , αn in some splitting field of f. The quantity
Y 2 D(f) = (αi − αj) (1.1) 1≤i IV]). The formula (1.1) is a symmetric polynomial in the roots α1, . . . , αn and is hence a polynomial in the symmetric functions of those roots, which are the coefficients of f. We call the universal version of this polynomial the classical discriminant of degree n polynomials, denoted by Dn. The utility of D(f) suggests that one might wish to evaluate it on either a single polynomial or a family of polynomials — for example, a polynomial f(x, t) ∈ K[x, t], which may be written n n−1 f(x, t) = x + f1(t)x + ··· + fn(t). Evaluation of D(f) (with respect to x) in this case is just a matter of specializing Dn to the coefficients a1 = f1(t), . . . , an = fn(t). One may then identify, say, for which values t0 of t the specialized polynomial f(x, t0) has a repeated root by computing the roots of the resulting polynomial in t. However, the direct 1 Chapter 1. The Classical Discriminant 2 Figure 1.1: The “swallowtail” approach of writing the polynomial Dn and substituting values of the coefficients is intractable, since the number of terms in Dn grows very quickly with the degree. This motivates the provision of efficient formulae for Dn. Formulae which allow for evaluation of D(f) on a family of polynomials as above are particularly desirable. We begin with an essential property of Dn. Proposition 1.1.1. The universal discriminant Dn is irreducible as an element of K[a1, . . . , an]. Proof. Suppose that Dn factors nontrivially into D1D2. Then the factorization must also hold in the n n−1 splitting field E of the universal polynomial f(x) = x + a1x + ··· + an. Namely, it must be that some nontrivial factor of Y 2 D(f) = (αi − αj) 1≤i 2 they are not symmetric. In the second case, there is some factor (αi − αj) which divides D1 and not D2 2 and some factor (αk − αl) which divides D2 and not D1. Let σ be the permutation which reverses αi and αk, reverses αj and αl, and leaves the remaining roots fixed. Then σ swaps the two aforementioned factors, again contradicting the premise that D1 and D2 be symmetric. The only possibility is that one of D1 or D2 be trivial, and the claim follows. Proposition 1.1.1 implies that the ideal (Dn) ⊆ K[a1, . . . , an] is prime. In particular, the algebraic properties of Dn may be understood directly by studying the geometry of the vanishing locus Z(Dn), which is a hypersurface in the affine space defined by the coordinates a1, . . . , an. Abusing language slightly, we call this hypersurface again the discriminant and denote it ∆n, or ∆ when n is understood. 4 2 Figure 1.1 shows the discriminant ∆˜ 4 of the polynomial x +a2x +a3x+a4, also known as the swallowtail. Chapter 1. The Classical Discriminant 3 The “true” degree four discriminant ∆4 is the product of ∆˜ 4 and an affine line, as will be explained in Section 1.2. In analogy, discriminants of higher degree are sometimes called generalized swallowtails. This thesis is concerned primarily with the study of the geometry of ∆n. In particular, we develop the philosophy that the geometry of ∆n is intimately related to the existence of determinantal formulae for D(f), formulae which represent Dn as the determinant of a matrix. The universal polynomial n n−1 n−2 f(x) = x + a1x + a2x + ··· + an is quasihomogeneous of degree n with weights deg x = 1 and deg ai = i for i = 1, . . . , n. The formula n Y f(x) = (x − αi) i=1 suggests that the roots (treated for the moment as indeterminates) should also have degree 1. In that case, (1.1) indicates that the D(f) is also quasihomogeneous of degree n · 2 = n(n − 1). 2 Indeed, when D(f) is written in terms of the coefficients a1, . . . , an according to the above degrees, that is the case (c.f. [Bou03, Proposition A.IV.6.10]). This implies further that O∆n = K[a1, . . . , an]/(Dn) is a graded ring with respect to the grading specified above. Quasihomogeneity allows the properties of ∆n to be connected with its formal germ ∆ˆ n at 0. Namely, ∗ there is a group action K on ∆n given by 2 n λ · (a1, . . . , an) 7→ (λa1, λ a2, . . . , λ an). Therefore, any property which is true in a neighbourhood of the origin is true on all of ∆n. In the next section, we shall show how the formal germ of ∆n is constructed using the theory of unfoldings of maps. This allows us to bring the powerful tools of deformation theory to bear on the study of ∆n. For the remainder of this thesis, we assume that the field K has characteristic zero. Chapter 1. The Classical Discriminant 4 1.2 A local perspective: Unfoldings In this section, we construct the formal germ of ∆n at 0 using the theory of unfoldings of maps. We roughly follow the development in [Tei77]. To avoid unnecessary technical complications, we develop this theory in a very restricted context, namely, unfoldings of analytic maps C → C. In particular, all future results which use the theory of unfoldings will require that K = C.1 The reader may consult, e.g., [BF96] for an extensive treatment of unfoldings from a deformation-theoretic standpoint. Roughly speaking, an unfolding of a map C → C is a family of maps which varies continuously in the parameters of some base space S. More precisely: Definition 1.2.1. Let ψ :(C, 0) → (C, 0) be the germ of a holomorphic map. A (smooth) unfolding of ψ is a pair ((S, 0), Ψ) where (S, 0) is the germ of a smooth analytic space and Ψ : (S × C, 0) → (C, 0) is the germ of a map such that Ψ |{0}×C= ψ. We shall normally refer to the deformation ((S, 0), Ψ) merely by Ψ, suppressing the base space (S, 0). Given a holomorphic map ψ : C → C, an unfolding of ψ is a pair (S, Ψ) where S =∼ Ck for some k ≥ 0 and Ψ : S × C → C is a holomorphic map such that the germ of Ψ at 0 is an unfolding, in the above sense, of the germ of ψ at 0. Unfolding a map reveals what happens under small perturbations of the map. A map is called stable if no small perturbation fundamentally changes the properties of the map. The precise meaning of stability is dependent on the sense in which two germs are considered to be equivalent: the properties of interest are precisely those which are preserved in an equivalence class of germs. There are several senses of equivalence relations of germs which appear in the literature, c.f. [BF96]. We shall be concerned with the geometry of the fibre of the map over zero, so, for us, a stable map is one whose fibre over zero, up to isomorphism, does not change under small perturbations. The natural relation for such study is extended contact equivalence, and we frame our definitions thus. A condition closely related to stability is versality: an unfolding is versal if it represents all nontrivial ways of perturbing the map. If an unfolding be versal, then no further unfolding thereof can change the properties of the map in a nontrivial way, so versal unfoldings are stable maps. We define versality precisely as follows: Definition 1.2.2. Let ((S, 0), Ψ) and ((T, 0), Ψ0) be unfoldings of a germ ψ :(C, 0) → (C, 0). We say that ((T, 0), Ψ0) is induced from ((S, 0), Ψ) if there exists a holomorphic germ Φ : (T × C, 0) → (S × C, 0) 0 and an invertible element u ∈ O(T ×C,0) such that Ψ and (t, z) 7→ u(t, z)(Ψ ◦ Φ)(t, z) define the same 1This is not a severe restriction in practise, thanks to the Lefschetz principle. Chapter 1. The Classical Discriminant 5 germ. We say that ((S, 0), Ψ) is versal if every unfolding of ψ is induced from ((S, 0), Ψ). We say that ((S, 0), Ψ) is semiuniversal or miniversal if ((S, 0), Ψ) is versal and the dimension of S is minimal. An unfolding Ψ0 of a holomorphic map ψ is induced from an unfolding Ψ of ψ if the germ at zero of Ψ0 is induced from the germ at zero of Ψ. The unfolding Ψ is versal or miniversal if its germ at zero is. Our basic example is the map x 7→ xn, from which arises our object of study. We first construct a miniversal unfolding thereof. n n−2 Proposition 1.2.3. The map Ξ:(˜ x, a2, . . . , an) 7→ x + a2x + ··· + an is a miniversal unfolding of the map x 7→ xn. 0 n Proof. Suppose (T, Ξ ) is another unfolding of the map x 7→ x . Write T = Spec C b1, . . . , bk . We may J K expand Ξ0 as a Taylor series about the origin in T × C: 0 2 Ξ (x, b1, . . . , bk) = g0(b1, . . . , bk) + g1(b1, . . . , bk)x + g2(b1, . . . , bk)x + ··· . 0 n That Ξ is an unfolding of x 7→ x implies that gn(0,..., 0) = 1 and gi(0,..., 0) = 0 for all i 6= n. 0 0 0 The Weierstraß Preparation Theorem implies that Ξ = uξ , where u is a unit in OT ×C and ξ = n−1 n hn + hn−1x + ··· + h1x + x , hi ∈ OT for each i. 0 We now define the map Φ : T × X → S × X in stages. First, let S = Spec C a1, . . . , an and define J K 0 0 0 00 0 Φ : T × X → S × X via Φ (b1, . . . , bk, x) = (h0, . . . , hn−1, x). Now define Φ : S × X → S × X via 00 1 1 n−i Φ (a1, . . . , an, x) = (k2, . . . , kn, x + n a1), where ki ∈ OS0×X is the coefficient of z + n a1 in the n n−1 1 Taylor expansion of z + a1z + ··· + an about the point z = − n a1. Note that the coefficient of 1 n−1 1 n z + n a1 in this Taylor expansion is zero and that the coefficient of z + n a1 is one. Finally, let Φ := Φ00 ◦ Φ0. It is an easy verification (left to the reader) that Ξ0 and (t, z) 7→ u(t, z)(Ξ˜ ◦ Φ)(t, z) define the same germ, and thus (T, Ξ0) is induced from Ξ.˜ Since (T, Ξ0) was arbitrary, Ξ˜ is versal. Our initial focus of study is the fibre over zero of the unfolding of x 7→ xn. We may think of it as the germ of an incidence variety: points on this fibre are pairs (s0, x0) such that x0 is a zero of the perturbed map ψs0 := (x 7→ Ξ(s0, x)). We define two germs which are fundamental. Definition 1.2.4. Let (S, Ψ) be an unfolding of ψ : C → C. Let p :Ψ−1(0) → S is the restriction of −1 1 the projection S × X → S to Ψ (0). The critical locus C(Ψ) is the support of the module ΩΨ−1(0)/S of 1 relative differentials of the map p. The discriminant ∆(Ψ) is the support of the direct image p∗ΩΨ−1(0)/S. Suppose the unfolding Ψ is given by a polynomial vanishing at zero, which is certainly the case for the unfolding Ψ of x 7→ xn in Proposition 1.2.3. Then p is a finite map: the preimages of s ∈ S correspond Chapter 1. The Classical Discriminant 6 −1 to the roots of the perturbed polynomial ψs. Furthermore, the fibre Ψ (0) is the vanishing locus of n n−1 x + a1x + ··· + an, which is a smooth manifold, as can be seen from the Jacobian criterion. In this case, the critical locus C(Ψ) is just the branch locus: the locus of points on Ψ−1(0) in a neighbourhood of which p is not bijective. The discriminant of a miniversal unfolding contains all of the essential geometric content of the discriminant of any versal unfolding. This is important since we are studying primarily the discrimi- n n−1 nant of the versal unfolding (x, a1, . . . , an) 7→ x + a1x + ··· + an, not of the miniversal unfolding n n−2 (x, a2, . . . , an) 7→ x + a2x + ··· + an. The following proposition makes this precise. Proposition 1.2.5. Let ((S, 0), Ψ) be a semiuniversal unfolding of ψ :(C, 0) → (C, 0) and let ((S0, 0), Ψ0) be a versal unfolding of ψ. Then ∆(Ψ0) is analytically isomorphic to the product of ∆(Ψ) and a smooth factor. Proof. Versality of Ψ0 implies that there exists a map Φ : (S0 × C, 0) → (S × C, 0) and a unit u ∈ 0 0 OS ×C,0 such that Ψ and (s, z) 7→ u(s, z)(Ψ ◦ Φ)(s, z) define the same germ. In particular, Φ restricts 0 ∼ to a map ∆(Ψ ) → ∆(Ψ) and thus O∆(Ψ0) is a finitely-generated algebra over O∆(Ψ), say O∆(Ψ0) = O∆(Ψ) t1, . . . , tk /I, whose presentation is chosen so that k is minimal. J K 00 Let us form a new unfolding ((T, 0), Ψ ) of ψ. The germ (T, 0) is S × Spec C t1, . . . , tk . Since J K k is minimal, the Zariski tangent spaces at zero of T and of S0 are isomorphic as C-vector spaces. Geometrically, T is just the product of S and the germ of k-dimensional affine space. The map Ψ00 : T × C → C is just (s, t, z) 7→ Ψ(s, z). Clearly Ψ may be induced from Ψ00, so ((T, 0), Ψ00) is versal. 00 ∼ 00 ∼ 0 Clearly also ∆(Ψ ) = ∆(Ψ) × Spec C t1, . . . , tk . It therefore suffices to show that ∆(Ψ ) = ∆(Ψ ). J K 00 0 0 By versality of Ψ , there exists a map Φ :(T × C, 0) → (S × C, 0) and a unit v ∈ OT ×C,0 such that Ψ00 and (t, z) 7→ v(t, z)(Ψ0 ◦ Φ0)(t, z) define the same germ. Thus Φ0 maps ∆(Ψ00) to ∆(Ψ0). Now let φ : O∆(Ψ0) → O∆(Ψ00) be the associated map of algebras. Let π be the natural projection ∼ 0 O∆(Ψ00) = O∆(Ψ) t1, . . . , tk → O∆(Ψ0). Then π ◦ φ is an automorphism of O∆ 0 , so ∆(Ψ ) is a retract J K Ψ of ∆(Ψ00). Since ∆(Ψ0) and ∆(Ψ00) have isomorphic tangent spaces at zero, they are isomorphic, as claimed. Now we are ready to connect the classical discriminant ∆n with the discriminant of a versal unfolding of x 7→ xn. n n−1 Proposition 1.2.6. The map Ξ:(x, a1, . . . , an) 7→ x + a1x + ··· + an is a versal unfolding of the n map x 7→ x . Its discriminant is the classical discriminant ∆n. In particular, ∆n is isomorphic to the ˜ ˜ product of the the discriminant of Ξ, which we call the reduced discriminant ∆n, and Spec C[a1]. Chapter 1. The Classical Discriminant 7 Proof. That the above map is versal follows immediately from Proposition 1.2.3, since Ξ,˜ and hence every unfolding of x 7→ xn, can be induced from it. Observe that Ξ−1(0) is just the incidence variety n n−1 {(a1, . . . , an, x) ∈ S ×C | x +a1x +···+an = 0}. The preimages of (a1, . . . , an) under the projection −1 n n−1 Ξ (0) → S are the points (a1, . . . , an, α) where α is a root of f(x) = x + a1x + ··· + an. Thus a generic point over (a1, . . . , an) has n distinct preimages. The critical locus of the projection is the branch locus, and its image ∆(Ξ) is just the locus of points whose associated polynomial has less than n distinct roots, namely, the classical discriminant ∆n. 1.3 A global perspective: Hilbert schemes We now give a projective version of the discriminant. Consider the space X = (P1)×n. There is a natural action of the symmetric group Sn on X given by permutation of the factors. The quotient of X by 1 this action is called the Hilbert scheme of n points on P , denoted Hn, or just H when n is understood. An orbit in X under the action by Sn may be viewed as the set of roots of some nonzero homogeneous polynomial n n−1 n−2 2 n−1 n F (x, y) = a0x + a1x y + a2x y + ··· + an−1xy + any . This sets a natural one-to-one correspondence up between those orbits and polynomials F (x, y), up to nonzero scalar multiples of the latter. Thus the Hilbert scheme is isomorphic to Proj [a , . . . , a ] = n . K 0 n PK Now consider the critical locus C of the quotient map q : X → Hn. This locus consists of points on X where the action of Sn is not faithful. Its image under q is the discriminant locus ∆n in Hn, which we identify with the set of homogeneous polynomials F (x, y) with a repeated linear factor, modulo nonzero scalar action. Clearly this homogeneous realization of ∆n agrees, on the affine piece {y 6= 0 and a0 6= 0}, with the affine version defined above. Consider now the product H × P1. Let n n−1 n−1 n F := a0x + a1x y + ··· + an−1xy + any ∈ K[a0, . . . , an, x, y] ∂F be the universal homogeneous polynomial of degree n in x and y and view its partial derivatives Fx := ∂x ∂F 2 1 ¯ and Fy := ∂y as sections of OH×P (1, n − 1). Consider the incidence variety ∆ defined by the sections 1 Fx and Fy. It is a smooth codimension-two subvariety of H × P . The Euler identity nF = xFx + yFy 2 ∂g ∂g In general, for a section g of O 1 (j, k), j, k ≥ 0, we write gx for and gy for . H×P ∂x ∂y Chapter 1. The Classical Discriminant 8 shows that on the affine pieces Uy := {y 6= 0} and Ux := {x 6= 0}, respectively, ∆¯ coincides with the n n 1 1 varieties defined by the sections { y F,Fx} ⊆ Γ(Uy, OH×P (1, n − 1)) and { x F,Fy} ⊆ Γ(Ux, OH×P (1, n − 1)). Thus points on ∆¯ are pairs (f, t) ∈ H × P1 such that t is a repeated root of f, the latter viewed as a 1 ¯ homogeneous polynomial. In particular, the projection map pH : H × P → H, restricted to ∆, defines a map π : ∆¯ → ∆. We now define a family of subvarieties of ∆n which is of fundamental importance: the caustics Σn,k for 2 ≤ k ≤ n. They are a particular subset of the Thom-Boardman stratification, c.f. [Tei77]. Let ¯ 1 Σn,k be the subvariety of H × P defined by the kth-order partial derivatives of F , viewed as sections 1 ¯ of OH×P (1, n − k). It is again an easy computation to see that Σn,k is smooth. Furthermore, since the defining ideal of Σ¯ n,k on a given affine set is quasihomogeneous with respect to the weights given in Section 1.1, Σ¯ n,k is connected. Definition 1.3.1. The kth caustic of ∆n, denoted Σn,k, is the image of Σ¯ n,k under the projection pH . For n ≥ k ≥ 2, the caustic Σn,k is the locus of polynomials of degree n with a root of multiplicity at least k. The caustics ∆n = Σn,2 ⊇ Σn,3 ⊇ · · · ⊇ Σn,n form a stratification of ∆n. For each k > 2, Σn,k is a subvariety of Σn,k−1 of codimension one. Proposition 1.3.2. The restriction of pH to Σ¯ n,k is the normalization of Σn,k. In particular, π : ∆¯ → ∆ is the normalization of ∆, and the normalizations of each of the caustics Σn,k are smooth. Proof. Fix n ≥ k ≥ 2. The map πk is finite since the number of preimages of a point on Σn,k is the number of roots of multiplicity k of the associated degree n polynomial, which is clearly a finite number. In addition, a generic polynomial of degree n with a root of multiplicity k has exactly one root of multiplicity exactly k, so πk is generically one-to-one. Finally, Σ¯ n,k is smooth, hence normal. Thus the map πk is the normalization of Σn,k, as claimed. 1.4 The geometry of ∆n Here we collect some basic results about the geometry of the discriminant ∆n and its caustics Σn,k. The first theorem, whose proof we defer to Chapter 2, characterizes the singular loci of the caustics Σn,k as the locus of points corresponding to polynomials with greater degeneracies than are required to be on Σn,k. Theorem 1.4.1. Let n ≥ k ≥ 2. The singular locus of the k-th caustic Σn,k is the locus of polynomials with either a root of multiplicity strictly greater than k or more than one root of multiplicity at least Chapter 1. The Classical Discriminant 9 k. The former locus has codimension one and the latter has codimension k − 1 if 2k ≤ n and is empty otherwise. Definition 1.4.2. The self-intersection locus of Σn,k (sometimes called the Maxwell set in the case k = 2) is the locus of polynomials with more than one root of multiplicity at least k. The next corollary will be useful in Chapter 3. Corollary 1.4.3. The singular locus of the caustic Σn,3 is contained in its intersection with the self- intersection locus in ∆n. Proof. By Theorem 1.4.1, the singular locus of Σn,3 consists of Σn,4 and the locus of polynomials of degree n with more than one triple root. The former locus is contained in the self-intersection locus since a single root of multiplicity four is two pairs of double roots, while the latter locus is clearly contained in the self-intersection locus. We now use the theory of unfoldings to describe the geometry of ∆n near certain points other than the origin. Theorem 1.4.4. Let p ∈ Σn,3 be a point which is not on the self-intersection locus of ∆n. Then the germ of ∆n at p is formally isomorphic to the product of ∆˜ 3 and a smooth factor. Proof. Let p ∈ Σn,3 be as given and write p = (a1, . . . , an). We denote by p(x) the polynomial corre- sponding to p. We claim that the germ of Ψ at p is a versal unfolding of x 7→ p(x). This is a consequence of the following general result in deformation theory, for whose proof the reader may consult, e.g., [Fle81]. Theorem 1.4.5 (Openness of Versality). Let Ψ:(S × C, 0) → C be a versal unfolding of ψ : C → C. Then there is an open neighbourhood U of 0 in S ×C such that Ψ is a versal unfolding of ψs : z 7→ Ψ(s, z) for every s ∈ U. ∗ The C action on ∆n described in Section 1.1 implies that, in fact, U may be taken to be the entire discriminant. Applying Proposition 1.2.5, we find that the germ of ∆n at p is formally isomorphic to the product of a versal deformation of p(x) and a smooth factor. The choice of p implies that its corresponding polynomial is of the form p(x) = g(x)(x − α)3, where g(x) has distinct roots which do not include α. But then g(x) is invertible in a neighbourhood of α, so a versal unfolding of p(x) is just 3 2 (x, a1, a2, a3) 7→ g(x)((x − α) + a1(x − α) + a2(x − α) + a3), which is equivalent to a versal unfolding of x 7→ x3. The claim now follows from Proposition 1.2.6. Chapter 1. The Classical Discriminant 10 We now restrict to the affine subsets {y 6= 0 and a0 6= 0} of the caustics Σn,k. It will be convenient to work in different coordinates, namely, those associated to the divided powers of x. For 1 ≤ i ≤ n, set si := (n!/(n − i)!)ai. Then s1, . . . , sn are identified with the coefficients of the polynomial (n) (n−1) (n−2) x + s1x + s2x + ··· + sn, where x(k) := xk/k! is the kth divided power of x. For n ≥ k > 0, differentiation of a polynomial in x with respect to x defines a finite map Σn,k → Σn−1,k−1, which in turn defines a tower of varieties terminating at the discriminant ∆n = Σn,2. With respect to the coordinates s1, . . . , sn, this map is just projection (s1, . . . , sn) 7→ (s1, . . . , sn−1). This tower is central to the study of the affine discriminant. We give here a few general geometric results which will be useful in Chapter 3. Proposition 1.4.6. For a fixed n, the varieties Σ¯ n,k, k ≥ 2, share the same normalization in the sense that the normalization map Σ¯ n−1,k−1 → Σn−1,k−1 factors through Σn,k. Proof. We construct a parametrization of Σn,k in the following manner: For a polynomial f with a root t of multiplicity k, its Taylor series around t has the form (n) (n−1) (k) f(x) = (x − t) + u1(x − t) + ··· + un−k(x − t) , (1.2) Expansion of (1.2) defines a map Spec K[t, u1, . . . , un−k] → Σn,k which is clearly a normalization. Dif- ferentiation of (1.2) yields 0 (n−1) (n−2) (k−1) f (x) = (x − t) + u1(x − t) + ··· + un−k(x − t) , which is of the same form as that corresponding to the parametrization of Σn−1,k−1. The claim follows. The following result of Givental shows how the functions sn−2+i on Σn+k,k+2 for k ≥ i > 0 embed ∼ in the normalization Σ¯ n+k,k+2 = ∆¯ n. This will be useful for computations later on. R x 00 (i−1) Lemma 1.4.7. For i > 0, sn−2+i = ± 0 f (t)t dt. Proof. See [Giv82], Lemma 2. The following application of Lemma 1.4.7 will be useful in Chapter 3. Proposition 1.4.8. For n ≥ 2 and k ≥ 0, Ω1 =∼ Ω1 . Σ¯ n+k,k+2/Σn+k,k+2 Σ¯ n,2/Σn,2 Chapter 1. The Classical Discriminant 11 Proof. The modules Ω1 and Ω1 are the cokernels of the Jacobian matrices of Σ¯ n+k,k+2/Σn+k,k+2 Σ¯ n,2/Σn,2 0 the normalization maps π : Σ¯ n+k,k+2 → Σn+k,k+2 and π : Σ¯ n,2 → Σn,2. We use the local coordi- ∼ nates x, s1, . . . , sn−2 for Σ¯ n+k,k+2 = Σ¯ n,2 and local coordinates s1, . . . , sn+k, respectively s1, . . . , sn, for Σn+k,k+2, respectively Σn,2. It suffices to show that, for i > 1, the form dsn+i−2 is a Σ¯ n,2-linear com- bination of ds1, . . . , dsn−1. For i > 0, differentiating the formula for sn+i−2 given in Lemma 1.4.7 with respect to the local coordinates on Σ¯ n,2, we obtain the form n + i − 4 i − 1 ds = ± f 00(x)x(i−1) dx + x(n+i−3) ds + ··· + x(i) ds . n+i−2 i − 1 1 i − 1 n−2 Setting i = 1, we obtain the form 00 (n−2) dsn−1 = ± f (x) dx + x ds1 + ··· + x dsn−2 . (i−1) Therefore, for i > 1, dsn+i−2 is the sum of ±x dsn−1 and a suitable Σ¯ n,2-linear combination of ds1, . . . , dsn−2. The claim follows. 1.5 Structure of this thesis The remainder of this thesis is organized as follows. Chapter 2 introduces determinantal formulae and describes three such formulae for Dn from the classical literature. It then describes the formalism of matrix factorizations, with which we study such formulae. There is then a detailed study of the Cayley method described in [GKZ94]. The main original contribution in Chapter 2 is a generalization of the Cayley method to construct presentation matrices of OΣ¯ n,k over OH. In the final section of this chapter, this construction is used to prove Theorem 1.4.1. Chapter 3 contains the main original results of this thesis, namely, the construction of a new determi- nantal formula which is not equivalent to the classical ones. The principal contribution is Theorem 3.1.2, 1 which characterizes Ω∆¯ /∆. Based on that, the open swallowtail is defined as the kernel of the universal 1 derivation O∆¯ → Ω∆¯ /∆ and, in view of Theorem 3.1.2, a presentation matrix thereof is constructed in Section 3.2. In Section 3.3, the algebraic definition of Section 3.1 is proved to be equivalent to the geometric version introduced by Arnol’d in [Arn81]. The remainder of Chapter 3 is devoted to further study of the properties of the open swallowtail and its presentation matrix, making ample use of the tools provided by the algebraic construction. Chapter 4 discusses the theory of deformations of modules and describes how to use its tools to construct moduli spaces of determinantal formulae for Dn. The first part is devoted to a detailed Chapter 1. The Classical Discriminant 12 description of an algorithm for computing versal deformations of matrix factorizations. This algorithm slightly generalizes the Massey Product Algorithm described in detail in [Siq01] in that the algorithm described in this thesis works in the relative case where the singularity is deformed along with the module over it. This generalization is necessary to use the algorithm to construct a moduli space of liftings of a module over a curve singularity to ∆n. The middle part of Chapter 4 is an elementary treatise on the structure of graded rank one maximal Cohen-Macaulay modules over the curve singularity n n−1 Γn := {x − y = 0}. As proved at the beginning of Chapter 4, all graded rank one MCM modules over ∆n restrict to graded rank one MCM modules over Γn, so the latter modules are starting points for the classification of the former. The next section contains the main contribution of this chapter: an exposition of the use of this algorithm with the computer algebra system Macaulay 2 [GS] to classify graded rank one MCM modules over ∆˜ 4. The last section is devoted to some theoretical results which describe the theory of these deformations and may, in principle, be used to construct deformations without resorting to computer calculation. Chapter 5 includes some concluding remarks and some musings about possible directions for future research along the lines of this thesis. Chapter 2 Matrix Factorizations In this chapter, we introduce our source of formulae for the classical discriminant, namely, represen- tations of the discriminant as the determinant of a matrix whose entries are in the ring of coefficients K[a1, . . . , an]. These so-called determinantal formulae allow us to realize our goal of efficiently comput- ing the discriminant both of individual polynomials and of families thereof, since the problem is reduced to computing the determinant of a matrix defined over a suitable polynomial ring. Furthermore, we shall see that the matrices involved in these formulae provide far richer information about the root structure of the polynomial than the discriminant alone. We begin with a treatment of the formulae discovered in the 18th and 19th centuries. We then state the well-known result that all of these formulae are, in a sense which we shall make precise, equivalent. This motivates the main question to which this thesis is devoted: Are there any nontrivial determinantal formulae for Dn which are inequivalent to the classical ones? We then introduce Eisenbud’s [Eis80] construction, which connects determinantal formulae with the homological properties of the coordinate ring O∆n of the discriminant hypersurface. This indispensable tool permits us to view determinantal formulae from a geometric viewpoint and provides a detailed understanding of the properties of the matrices of the formulae. In the last section of this chapter, we introduce the Cayley method, which can be used to derive each of the classical determinantal formulae. We generalize this method slightly to construct resolutions of the normalizations of the caustics introduced in Chapter 1. This is of interest in its own right, since the presentation matrices reveal information about the root structure of the polynomial, but it will also be important in Chapter 3 when constructing the presentation matrix of the so-called open swallowtail. 13 Chapter 2. Matrix Factorizations 14 2.1 Classical determinantal formulae There are three specific classical formula for Dn, two of which come in both affine and projective versions. We describe them in turn. The Vandermonde formula. Consider the Vandermonde matrix in the roots α1, . . . , αn: 1 1 ··· 1 1 α1 α2 ··· αn−1 αn V := α2 α2 ··· α2 α2 . 1 2 n−1 n . . . . ...... . . . . . n−1 n−2 n−1 n−1 α1 α2 ··· αn−1 αn Q T It is a classical result in linear algebra that det(V ) = 1≤i T Y 2 det(VV ) = (αi − αj) = D(f). 1≤i Furthermore, the (i, j) entry of VV T is n n X i j X i+j αkαk = αk . k=1 k=1 This is a power sum symmetric function, which is in particular symmetric in the roots α1, . . . , αn. Thus T the entries of VV may be written in terms of the elementary symmetric functions a1, . . . , an. The T formula Dn = det VV is called the Vandermonde formula for Dn. The Sylvester formula. For k > 0, let K[x]≤k be the (k + 1-dimensional) vector space of polynomials of degree at most k. Consider the map φ : K[x]≤n−2 ⊕ K[x]≤n−1 → K[x]≤2n−2 (g, h) 7→ gf + hf 0. The polynomial f lies on the discriminant if and only if f and f 0 share a common factor, which is the case if and only if φ is not surjective. Since φ is a map of finite-dimensional vector spaces of the same dimension, this is the case if and only if φ is not an isomorphism. Hence the determinant of φ equals the discriminant of f up to some nonzero constant factor. The resulting matrix has the form shown in Figure 2.1. Chapter 2. Matrix Factorizations 15 1 n a 1 (n − 1)a n 1 1 . . a a .. (n − 2)a (n − 1)a .. 2 1 2 1 ...... ...... . . 1 . . n an an−1 a1 an−1 2an−2 (n − 1)a1 . . . . ...... an . an−1 . .. .. . an−1 . 2an−2 an an−1 Figure 2.1: Affine Sylvester Matrix na a 0 1 (n − 1)a na 2a a 1 0 2 1 . . (n − 2)a (n − 1)a .. 2a 2a .. 2 1 3 2 ...... ...... . . na0 . . a1 an−1 2an−2 (n − 1)a1 nan (n − 1)an−1 2a2 . . . . ...... an−1 . nan . .. .. . 2an−2 . (n − 1)an−1 an−1 nan Figure 2.2: Projective Sylvester Matrix There is also a projective version of the Sylvester formula. Namely, let n n−1 n f(x, y) = a0x + a1x y + ··· + any be a homogeneous polynomial in two variables. An argument analogous to that above shows that the map Φ: K[x, y]n−2 ⊕ K[x, y]n−2 → K[x, y]2n−3 (g1, g2) 7→ fxg1 + fyg2, where K[x, y]l is the degree l piece of K[x, y], is not an isomorphism of vector spaces if and only if fx and fy have a factor in common, which is the case if and only if f has a repeated linear factor. Thus a matrix representing Φ constitutes a determinantal formula for the homogeneous discriminant ∆n = ∆n(a0, . . . , an). The matrix is shown in Figure 2.2. The Sylvester formula can more generally be used to compute the resultant of two univariate poly- nomials; see, e.g., [vzGG03, Chapter 6] for a readable treatment. Chapter 2. Matrix Factorizations 16 The B´ezoutformula. Consider the rational function f 0(x)f(y) − f 0(y)f(x) B(x, y) = . (2.1) y − x Since the numerator vanishes when x = y, B(x, y) is in fact a polynomial which my be written n−1 X i j B(x, y) = bijx y . i,j=0 The coefficients bij are polynomials in the variables a1, . . . , an. Let B be the matrix whose (i, j) entry n is bi+1,j+1. We may view B as a bilinear form on the vector space K . The following proposition shows that B is a determinantal formula for D(f). Proposition 2.1.1. We have det B = D(f). Proof. (See also [BEvB06, Theorem 5.6].) We show firstly that the bilinear form B is nondegenerate if and only if the roots of f are distinct. Let α be a root of f. We claim first that α is a repeated root of f if and only if (x − α) divides B(x, y). One direction is obvious from (2.1). Suppose now that α is a root but not a repeated root of f. Specializing B, we obtain f 0(α)f(y) − f 0(y)f(α) B(α, y) = y − α f(y) = f 0(α) . y − α Since α is not a repeated root of f, f 0(α) 6= 0, so B(α, y) is some nonzero polynomial in y, whence the claim. For α ∈ K, denote by v(α) the vector (1, α, α2, . . . , αn−1). Suppose now that α is a repeated root. Then (x−α) divides B(x, y), so, for all β ∈ K, v(β)T Bv(α) = 0. But vectors of the form v(β) span Kn, so T n w Bv(α) = 0 for all vectors w ∈ K , and B is degenerate. Suppose now that the roots (α1, . . . , αn) of f are distinct. Then (x−αi) does not divide B(x, y) for any i = 1, . . . , n. In particular, for i = 1, . . . , n, the T map β 7→ v(β) Bv(αi) is not identically zero. Since α1, . . . , αn are distinct, the vectors v(α1), . . . , v(αn) span Kn. Thus B is nondegenerate. We have thus shown that det B = uD(f) for some nonzero constant u. We now show that u = 1 by testing on the polynomial f(x) = xn − 1. Let ξ be a primitive nth root of unity in some extension field of K. Then the roots of f are 1, ξ, . . . , ξn−1. We first compute D(f) via the Vandermonde formula. This requires computing the power sum symmetric functions of the roots of f, for which we appeal to Chapter 2. Matrix Factorizations 17 the following well-known formula (see, e.g., [vzGG03, Lemma 8.7]): n−1 X n, if n | l, (ξj)l = j=0 0, otherwise. Thus, letting V be the Vandermonde matrix in the roots 1, ξ, . . . , ξn−1, the matrix VV T has the form n n T .. VV = . . n n Thus D(f) = det(VV T ) = (−1)b(n−1)/2cnn. Now we compute the determinant of the B´ezoutmatrix. We have f 0(x)f(y) − f 0(y)f(x) (yn − 1)(nxn−1) − (xn − 1)(nyn−1) = y − x y − x xn−1yn−1(y − x) + yn−1 − xn−1 = n y − x (y − x) xn−1yn−1 + xn−2 + xn−3y + ··· + yn−2 = n y − x = n xn−1yn−1 + xn−2 + xn−3y + ··· + yn−2 . Thus the matrix B has the form n . .. B = n . n n Its determinant is therefore (−1)b(n−1)/2cnn = D(f), as claimed. The formula described above is an affine version of the B´ezoutformula, since the polynomial f is n n−1 n not homogeneous. There is also a projective version. Let f(x, y) = a0x + a1x y + ··· + any be a Chapter 2. Matrix Factorizations 18 homogeneous polynomial in the indeterminates x and y. We set fx(x0, y0)fy(x1, y1) − fy(x0, y0)fx(x1, y1) BP (x0, y0, x1, y1) := . x0y1 − x1y0 Then BP (x0, y0, x1, y1) is a bihomogeneous polynomial which we write n−2 X i n−i−2 j n−j−2 BP (x0, y0, x1, y1) = bijx0y0 x1y1 . i,j=0 The matrix BP with entries bij then has determinant equal to the homogeneous discriminant of f. This matrix is (n − 1) × (n − 1), as opposed to the affine B´ezoutmatrix B described above which is n × n. As with the Sylvester formula, the B´ezoutformula may be easily generalized to compute the resultant of two univariate polynomials. The classical formulae introduced above are related to one another. To explain how, we introduce a definition which generalizes the ordinary notion of equivalence of matrices. Definition 2.1.2. Let A and B be two matrices over a commutative ring R. We say that A and B are weakly equivalent if they have isomorphic cokernels when viewed as maps of free modules over R. Remark 2.1.3. If R is a local ring, then the condition that A and B are weakly equivalent is equivalent to the existence of invertible matrices U1,V1,U2,V2 over R such that A0 0 A0 0 U1AV1 = and U2BV2 = , 0 U 0 0 V 0 with U 0 and V 0 invertible. An analogous condition holds when R is graded and A and B are presentation matrices of graded R-modules. Theorem 2.1.4. The matrices of the Vandermonde, affine Sylvester, and affine B´ezoutformulae are pairwise weakly equivalent. The projective versions of the Sylvester and B´ezoutformulae restricted to the affine piece {a0 6= 0} are equivalent to the affine versions. Proof. See [Bou03, A.IV.78] for a proof of the equivalence of the affine Sylvester and Vandermonde formulae. The equivalence of the projective Sylvester and B´ezoutformulae on an affine set follows from Examples 2.3.2 and 2.3.3 in Section 2.3 below. It remains to show that the affine Sylvester formula is equivalent to the projective Sylvester formula and that the affine B´ezout formula is equivalent to the projective B´ezoutformula. We address each in turn. Chapter 2. Matrix Factorizations 19 First, we consider the Sylvester matrix. It is necessary to specify row and column operations on the affine Sylvester matrix of Figure 2.1 so that the resulting matrix is of a block diagonal form in which one block is invertible and the other agrees with the projective Sylvester matrix of Figure 2.2 with a0 = 1. To do this, first scale the left-hand block of the affine matrix by a factor of n and then subtract the first n − 1 columns of the right-hand block from the left-hand block. Then switch the order of the two blocks and apply row operations to eliminate the entries below the (1, 1) entry. The resulting matrix is of the desired form. Now we consider the B´ezoutmatrix. Let B¯P (u, v) := BP (u, 1, v, 1). In view of the Euler identity nf = xfx + yfy, we have f (u, 1)f(v, 1) − f(u, 1)f (v, 1) nB(u, v) = n x x v − u f (u, 1)(vf (v, 1) + f (v, 1)) − (uf (u, 1) + f (u, 1))f (v, 1) = x x y x y x v − u = B¯P (u, v) + fx(u, 1)fx(v, 1). (2.2) Pn−1 i j Write fx(u, 1)fx(v, 1) = i,j=0 ciju v and let C be the matrix whose (i, j) entry is ci+1,j+1. Then, for i 1 ≤ i < n, the i-th row, respectively i-th column, of C equals the product of n an−i and the last row, respectively last column, of C. In particular, there exists an invertible matrix U such that U T CU is a matrix with a unit in the (n, n) entry and zero elsewhere. Furthermore, denoting by B˜P the n × n matrix whose upper-left (n − 1) × (n − 1) submatrix is the specialization BP |a0=1 and whose last row T and column are both zero, we have U B˜P U = B˜P . In view of (2.2), nB = B˜P + C. Thus T T nU BU = B˜P + U CU. The matrix on the left-hand side is clearly equivalent to B, while the matrix on the right-hand side is a block-diagonal form where one block is BP and the other is invertible. The claim follows. Within this equivalence class, the projective B´ezoutformula is a minimal presentation in the sense that no matrix of size strictly less than (n−1)×(n−1) can be weakly equivalent to the classical formulae. Chapter 2. Matrix Factorizations 20 One can see this already by examining the Vandermonde formula: the matrix is of size n × n, and the only entry with a unit is the (1, 1) entry. Theorem 2.1.4 begs the question of whether there exist determinantal formulae for the classical discriminant which are inequivalent to the classical ones. Of course, there exists a formula in which Dn is placed in a 1 × 1 matrix. This is inequivalent to the classical formulae for the reason stated above. This formula, and any other formula weakly equivalent to it, is trivial. We now restate our main question given at the beginning of this chapter. Are there any nontrivial determinantal formulae for Dn which are not weakly equivalent to the classical ones? 2.2 Eisenbud’s Theorem We now introduce matrix factorizations of a polynomial, which generalize determinantal formulae. In this formalism Eisenbud frames his celebrated results which connect determinantal formulae with homological algebra. Definition 2.2.1. Let R be a commutative ring and f ∈ R \{0}.A matrix factorization of f is a pair of square matrices (A, B) such that AB = f · id = BA, where id is the identity matrix of suitable size. If f is a non-zerodivisor of R, then one of the two equations suffices. Example 2.2.2. Let f ∈ R and suppose that A is a k × k matrix such that det A = f. Let B = adj A, the adjugate (or classical adjoint) of A. That is, k+1 det A11 − det A21 ··· (−1) det Ak1 k+2 − det A12 det A22 ··· (−1) det Ak2 B := . . . . , . . .. . . . . k+1 k+2 2k (−1) det A1k (−1) det A2k ··· (−1) det Akk where Aij is the matrix resulting from removing the ith row and jth column from A. Then (A, B) is a matrix factorization of f. This shows how matrix factorizations generalize determinantal formulae. Let f ∈ R be a non-zerodivisor. We define a morphism of matrix factorizations (A, B) → (A0,B0) of f in the following way. View A, respectively A0, as a map A : G → F , respectively a map A0 : G0 → F 0, Chapter 2. Matrix Factorizations 21 of free R-modules and B, respectively B0, as a map B : F → G, respectively B0 : F 0 → G0. A morphism (A, B) → (A0,B0) consists of a pair (α, β) of maps α : F → F 0 and β : G → G0 making the diagram G - F - G B A β α β (2.3) ? ? ? G0 - F 0 - G0 B0 A0 commute. Matrix factorizations of a non-zerodivisor f ∈ R form a category MF(f) in the following way. The objects of MF(f) are matrix factorizations of f. The morphisms of MF(f) are morphisms of matrix factorizations modulo those morphisms which factor through the trivial matrix factorization (1, f). In this category, two matrix factorizations (A, B) and (A0,B0) of f such that A and A0 are weakly equivalent are isomorphic. If (A, B) is a matrix factorization of f, then it follows from Cramer’s rule that f annihilates coker A, so coker A is naturally a module over R/(f). Furthermore, provided that f is a non-zerodivisor, A : G → F is injective as a map of R-modules. Hence the complex A 0 - G - F - coker A - 0 is a resolution of coker A as an R-module. A morphism (α, β):(A, B) → (A0,B0) of matrix factoriza- tions by definition defines a morphism from the associated resolution of coker A to that of coker A0, and therefore a module homomorphism coker A → coker A0. Thus (A, B) 7→ coker A defines a functor, also denoted coker, from MF(f) to the category mod(R/(f)) of finitely-generated R/(f)-modules. Further- more, if R is Cohen-Macaulay and f ∈ R is a regular element, it follows from the Auslander-Buchsbaum formula that depth coker A = dim R − 1 = dim R/(f), so coker A is a maximal Cohen-Macaulay mod- ule over R/(f). Denote by MCM(R/(f)) the full subcategory of mod(R/(f)) consisting of maximal Cohen-Macaulay modules over R/(f). The following result due to Eisenbud is fundamental. Theorem 2.2.3 ([Eis80]). Suppose that R is a regular local ring and that f ∈ R is a regular element. Then the functor coker is an equivalence of the categories MF(f) and MCM(R/(f)). Proof sketch. We show here only how to define a pseudoinverse Υ for coker. For a proof that Υ is well-defined and that it is indeed a pseudoinverse of coker, see, e.g. [Yos90, Theorem 7.4]. Let M be a maximal Cohen-Macaulay module over R/(f). We first resolve M over R. Since depth M = dim R/(f) = Chapter 2. Matrix Factorizations 22 depth R − 1, pdR M = 1, so the resolution has the following form: 0 - G - F - M - 0. (2.4) A We pull (2.4) back through the map given by multiplication by f, which is zero on M since M is a module over R/(f): 0 - G - F - M - 0 6 A 6 6 f f f = 0 0 - G - F - M - 0 A The map f : F → F therefore takes values in the kernel of the projection F → M. Thus it lifts uniquely to a map B : F → G as follows: - - - - 0 G F M 0 6 A 6 6 f B f f = 0 . (2.5) 0 - G - F - M - 0 A Commutativity of (2.5) implies that B ◦ A = f = A ◦ B. Given two maximal Cohen-Macaulay modules M and M 0 with a morphism φ : M → M 0, we may define a morphism of associated matrix factorizations (α, β) by lifting φ to a map of minimal resolutions of M and M 0, as in the following diagram: 0 - G - F - M - 0 A β α φ . ? ? ? 0 - G0 - F 0 - M 0 - 0 A The morphism (α, β) is not uniquely determined by φ. However, since morphisms in MF(f) are taken modulo those which factor through the matrix factorization (1, f), the class of (α, β) as a morphism in MF(f) is uniquely determined. These constructions define a functor Υ from the category MCM(R/(f)) of maximal Cohen-Macaulay modules over R/(f) to the category of matrix factorizations of f. The data in a matrix factorization (A, B) of f also completely describe a resolution of coker A over R/(f), as the following theorem shows. Chapter 2. Matrix Factorizations 23 Theorem 2.2.4 ([Eis80]). Let R be a regular local ring, f ∈ R a regular element, and (A, B) a matrix factorization of f. Denote by A¯ : G¯ → F¯ and B¯ : F¯ → G¯ the maps resulting from reducing the entries of A and B modulo f. Then coker A has a two-periodic resolution A¯ B¯ A¯ ··· - G¯ - F¯ - G¯ - F¯ - coker A - 0. (2.6) Proof. See, e.g., [Yos90, Proposition 7.2]. The next proposition shows precisely which maximal Cohen-Macaulay modules correspond to deter- minantal formulae, at least when f is irreducible. Proposition 2.2.5 ([Eis80]). With the same hypotheses as in Theorem 2.2.3, assume now that f ∈ R is irreducible and let (A, B) be a matrix factorization of f by k × k matrices. Then det A = uf i and det B = u−1f k−i, where i is the rank of coker A as an R/(f)-module and u ∈ R is a unit. Proof. Since f is irreducible in the regular local ring R, the ideal generated by f is prime. Consider the localization R(f). Over this localized ring, A is equivalent to a diagonal matrix with diagonal entries α1 αk P i f , . . . , f , where i := j αj ≤ k. Clearly then det A is associate to f , and this remains true over R. Furthermore, in R(f), the length of the cokernel of A i, and, reducing modulo f, we see that this length is precisely the dimension of the cokernel of A over K(R/(f)), which is the rank of A. The claim follows. In particular, matrix factorizations associated to determinantal formulae for f correspond to rank one maximal Cohen-Macaulay modules over R/(f). A natural question is to which module the classi- cal determinantal formulae introduced in Section 2.1 correspond. The following theorem, which is an immediate consequence of Examples 2.3.2 and 2.3.3 below combined with Proposition 1.3.2, answers this. Theorem 2.2.6. The cokernels of the classical determinantal formulae introduced in Section 2.1 are all locally isomorphic to the normalization O∆¯ of O∆. If R is a graded regular ring and f ∈ R is a homogeneous element, then one may define the category GrMF(f) of graded matrix factorizations in analogy with the above definition. Namely, the objects of GrMF(f) are pairs of matrices (A, B) which define homomorphisms of graded free modules such that the compositions AB and BA are multiplication by f. The morphisms are pairs of homomorphisms α, β of graded free modules making (2.3) commute. In this case, the functors coker and Υ induce inverse Chapter 2. Matrix Factorizations 24 equivalences GrMF(f) and GrMCM(R/(f)), the category of graded maximal Cohen-Macaulay modules over the graded ring R/(f). 2.3 The Cayley Method In this section, we develop a method for resolving over the ambient space H of ∆ modules which are locally isomorphic to the normalizations OΣ¯ n,k of the caustics Σn,k. Originally due to Cayley, this method was developed to construct determinantal formulae for the equation of the dual variety X∨ of a given projective variety X. In our case, the variety X is P1, embedded via the nth Veronese embedding n ∨ in P , and X is ∆n. Our treatment omits many of the technical points required for the general case. See [GKZ94, Chapter 2] for a thorough, modern treatment of this method. ¯ 1 The normalization Σn,k of the caustic Σn,k is of codimension k in H × P , while its vanishing ideal Ik 1 is defined by the k − 1-order partial derivatives of F viewed as sections of OH×P (1, n − k + 1), of which there are k. Thus it is a global complete intersection and its structure sheaf is resolved over the ambient k 1 space via the Koszul complex K• on the aforementioned sections of OH×P (1, n − k + 1) defining Ik. For ∼ example, Figure 2.3 shows resolutions by locally free sheaves of OΣ¯ n,2 = O∆¯ and OΣ¯ n,3 as modules over 1 OH×P . One might hope to construct a resolution of pH∗OΣ¯ n,k over H by constructing the maps of the spectral p,q q • k sequence associated to the double complex E0 := Ip , where Ip is an injective resolution of Kp for each k 0 ≤ p ≤ k. Indeed, since K• is quasi-isomorphic to OΣ¯ n,k , the resulting spectral sequence converges to l l 0,0 E := R pH∗OΣ¯ n,k , which vanishes for l 6= 0. However, the resulting spectral sequence has nonzero E∞ −1,1 and E∞ . Thus such a computation can produce only a resolution of the associated graded module of a filtration of pH∗OΣ¯ n,k , which is not exactly what we need. k To correct this problem, we twist K• by such a line bundle as to ensure that, in the spectral sequence p,q associated to the resulting double complex, E∞ = 0 except for p = q = 0. We first select a pivot k k 1 j ∈ {1, . . . , k} and twist K• by OH×P (0, j(n − k + 1) − 1). The resulting complex K• (j(n − k + 1) − 1) is then a resolution of the twisted structure sheaf OΣ¯ n,k (0, j(n − k + 1) − 1), which is isomorphic to OΣ¯ n,k 1 1 on any subset of H ×P over which OH×P (0, j(n−k +1)−1) is trivial — in particular any affine subset. l Since the choice of twist is positive, we still have R pH∗OΣ¯ n,k (0, j(n − k + 1) − 1) = 0 when l 6= 0. As we shall see, our choice of twist not only yields a spectral sequence with the desired properties, but greatly facilitates the computation. We now describe how a resolution of OΣ¯ n,k (0, j(n − k + 1) − 1) arises from the maps of the spectral Chapter 2. Matrix Factorizations 25 0 0 6 6 OΣ¯ n,2 OΣ¯ n,3 6 6 O O 6 6 8 9 8 9 :Fx Fy ; :Fxx Fxy Fyy ; O(−1, 1 − n)⊕2 O(−1, 2 − n)⊕3 . 6 6 8 9 8 9 0 Fyy −Fxy Fy > > > > >−Fyy 0 Fxx > :−Fx; > > : Fxy −Fxx 0 ; O(−2, 2 − 2n) O(−2, 4 − 2n)⊕3 6 6 8 9 Fxx > > >Fxy > > > :Fyy ; 0 O(−3, 6 − 3n) 6 0 Figure 2.3: Koszul complexes resolving OΣ¯ n,2 and OΣ¯ n,3 Chapter 2. Matrix Factorizations 26 0 1 1 OH ⊗K H P , OP (j(n − k + 1) − 1) 6 −1,0 d1 . . 6 −j+1,0 d1 0 1 ⊕( k ) 1 j−1 OH(−j + 1) ⊗K H P , OP (n − k) 1 1 ⊕( k ) 1 j+1 OH(−j − 1) ⊗K H P , OP (k − n − 2) 6 −j−2,1 d1 . . 6 −k,1 d1 1 1 1 OH(−k) ⊗K H P , OP ((j − k)(n − k + 1) − 1) . Figure 2.4: First page of the spectral sequence sequence. It follows from the projection formula that, for l, p, q ∈ Z, l l 1 1 ∼ 1 R pH∗OH×P (p, q) = OH(p) ⊗K H (P , OP (q)). 0 1 1 1 In particular, R pH∗OH×P (p, q) = 0 when q < 0 and R pH∗OH×P (i, j) when q > −2. The first page of the spectral sequence is shown in Figure 2.4. If j = 1, the second page of the spectral sequence is 0 1 O ⊗ H , O 1 (n − k) H K P P − d 2 2 ,1 −3,1 coker d1 . −2,1 The cokernel of d2 is pH∗OΣ¯ n,k (n − k), and the spectral sequence degenerates after this step. If j > 1, Chapter 2. Matrix Factorizations 27 then the second page of the spectral sequence is pH∗OΣ¯ n,k (j(n − k + 1) − 1) . . −j+1,0 ker d1 d − j− 2 1 ,1 −j−2,1 coker d1 . k In this case, since the original complex K• is exact except at homological degree zero and since l R pH∗OΣ¯ n,k (j(n − k + 1) − 1) = 0 −j−1,1 for l < 0, d2 must be an isomorphism. Again, the spectral sequence degenerates after this point. k k 0 1 1 ⊕( ) 0 1 ⊕( ) 1 j+1 1 j−1 Let d : OH(−j − 1) ⊗K H P , OP (k − n − 2) → OH(−j + 1) ⊗K H P , OP (n − k) −j−1,1 be a lifting of d2 . For 0 < i < k, let −i,0 d1 , if i < j, ∂i := d0, if i = j, −i−1,1 d1 , if i > j. The maps ∂i for 0 < i < k are the differentials of a complex C• which, in view of the above observations, ∼ is exact except at the right. We have coker ∂1 = pH∗OΣ¯ n,k (j(n − k + 1) − 1). In this manner the maps of the spectral sequence give rise to a resolution of pH∗OΣ¯ n,k (j(n − k + 1) − 1) over H by locally free sheaves. To describe the maps explicitly, it is convenient to compute in the fibre over a fixed point [a0 : ··· : an] ∈ H. The restriction of C• to this fibre is a map of vector spaces whose differentials vary • k polynomially in the coordinates a0, . . . , an. In doing so, we replace the injective resolutions Il of Kl ˇ k with Cech complexes which compute the cohomology of Kl , as shown in Figure 2.5. The first page of the spectral sequence is shown in Figure 2.6. Recall Serre’s theorem characterizing Ha( b, O(c)) for a, b, c ∈ . Set b = Proj [x , . . . , x ] and P Z PK K 0 b let V denote the degree one part of K[x0, . . . , xb]. Chapter 2. Matrix Factorizations 28 Γ(Ux, O(j(n − k + 1) − 1)) - L - - 0 Γ(Uxy, O(j(n − k + 1) − 1)) 0 d0,0 Γ(Uy, O(j(n − k + 1) − 1)) H 6 6 −1,0 −1,1 dV dV . . . . 6 6 −j+1,0 −j+1,1 dV dV ⊕( k ) Γ Ux, O(n − k) j−1 ⊕ k 0 - L - Γ U , O(n − k) (j−1) - 0 −j+1,0 xy k ⊕( ) dH Γ Uy, O(n − k) j−1 6 6−j,0 −j,1 dV dV ⊕(k) Γ Ux, O(−1) j ⊕ k 0 - L - Γ U , O(−1) (j) - 0 −j,0 xy k ⊕( ) dH Γ Uy, O(−1) j 6 6−j−1,0 −j−1,1 dV dV ⊕( k ) Γ Ux, O(k − n − 2) j+1 ⊕ k 0 - L - Γ U , O(k − n − 2) (j+1) - 0 −j−1,0 xy k ⊕( ) dH Γ Uy, O(k − n − 2) j+1 6 6−j−2,0 −j−2,1 dV dV . . . . 6 6 −k,0 −k,1 dV dV Γ(Ux, O((j − k)(n − k + 1) − 1)) 0 - L - Γ(U , O((j − k)(n − k + 1) − 1)) - 0 −k,0 xy Γ(Uy, O((j − k)(n − k + 1) − 1)) dH k Figure 2.5: Double complex for K• over [a0 : ··· : an]. Chapter 2. Matrix Factorizations 29 0 1 1 H P , OP (j(n − k + 1) − 1) 6 −1,0 d1 . . 6 −j+1,0 d1 0 1 ⊕( k ) 1 j−1 H P , OP (n − k) 1 1 ⊕( k ) 1 j+1 H P , OP (k − n − 2) 6 −j−2,1 d1 . . 6 −k,1 d1 1 1 1 H P , OP ((j − k)(n − k + 1) − 1) . Figure 2.6: First page of the spectral sequence on a fibre over a point Chapter 2. Matrix Factorizations 30 Theorem 2.3.1 (Serre). We have Sym (V ), if k ≥ 0 and a = 0, c a b H (P , O(c)) =∼ (Sym (V ))∗, if k < −j and a = b, −c−b−1 0, otherwise. ∗ Here Symc(−) denotes the cth symmetric power of the argument and (−) is the K-dual. Proof. See [Har77, Theorem 5.1]. −i,0 −i,1 k In view of these identifications, the differentials d1 and d1 are just the maps of K• restricted to −1,0 −k,1 the given fibre. In particular, the matrices of the first and last maps, d1 and, respectively, d1 are generalized Sylvester matrices whose structure we now describe. −1,0 Assume that j > 1. Then the map d1 is ∂k−1F ∂k−1F (g , . . . , g ) 7→ g + ··· + g , 1 k ∂xk−1 1 ∂yk−1 k where each gi is a homogeneous polynomial in K[x, y] of degree (j − 1)(n − k + 1) − 1. The associated matrix is therefore divided vertically into k blocks of (j − 1)(n − k + 1) − 1 columns each. Each block is associated with a given k − 1-order partial derivative of F . Each block is of the form α0 α α 1 0 .. α2 α1 . . . . . . .. α 0 , αn−k+1 αn−k α1 . . α .. . n−k+1 .. . αn−k αn−k+1 where α0, . . . , αn−k are the coefficients of the associated partial derivative of F . In particular, the entries are linear in the coefficients a0, . . . , an. −k,1 Assume now that j < k − 1. Then the map d1 is ∂k−1F ∂k−1F g 7→ g, . . . , g , ∂xk−1 ∂yk−1 Chapter 2. Matrix Factorizations 31 where g is a homogeneous element of K[x−1, y−1] of degree (j −k)(n−k +1)−1 < 0. Its matrix then has a form similar to the transpose of the matrix described above: it is divided horizontally into k blocks of −(j −k +1)(n−k +1) rows each. The i-th row of a given block contains the coefficients of the associated k − 1-order partial derivative of F shifted to the right i − 1 places. The dimension of the vector space of homogeneous elements of K[x−1, y−1] of degree (j − k)(n − k + 1) − 1 is (k − j)(n − k + 1), so the matrix is of size k(k − j − 1)(n − k + 1) × (k − j)(n − k + 1). It remains to characterize the map 1 1 ⊕( k ) 0 1 ⊕( k ) ∂j : H P , O(k − n − 2) j+1 → H P , O(n − k) j−1 . Since Hi(P1, O(−1)) = 0 for all i, the map k k k −j,0 ⊕(j) ⊕(j) ⊕(j) dH :Γ Ux, O(−1) ⊕ Γ Uy, O(−1) → Γ Uxy, O(−1) is an isomorphism. Thus the map k k k −j,0 −1 ⊕(j) ⊕(j) ⊕(j) (dH ) :Γ Uxy, O(−1) → Γ Ux, O(−1) ⊕ Γ Uy, O(−1) is well-defined and ∂j is the composition −j−1,1 k d k ⊕( ) V ⊕( ) Γ Uxy, O(k − n − 2) j+1 −→ Γ Uxy, O(−1) j −j,0 −1 d k k ( H ) ⊕( ) ⊕( ) −→ Γ Ux, O(−1) j ⊕ Γ Uy, O(−1) j −j,0 d k k V ⊕( ) ⊕( ) −→ Γ Ux, O(n − k) j−1 ⊕ Γ Uy, O(n − k) j−1 k 1 1 ⊕( ) restricted to H P , O(k − n − 2) j+1 . The matrix of ∂j is B´ezout-type, in the sense that its entries are quadratic binominals in the variables a0, . . . , an, just as in the B´ezoutmatrix. Example 2.3.2. By applying this method to ∆ = Σn,2 and setting the pivot j = 2, we recover the classical k projective Sylvester formula for the discriminant. Namely, the complex K• is twisted by 2n − 3. The Chapter 2. Matrix Factorizations 32 first page of the spectral sequence is 0 1 1 OH ⊗K H P , OP (2n − 3) 6 −1,0 d1 0 1 ⊕2 1 OH(−1) ⊗K H P , OP (n − 2) . −1,0 0 1 ⊕2 ⊕2 1 ∼ The map d1 takes g1, g2 ∈ H P , OP (n − 2) = K[x, y]n−2 to 0 1 1 ∼ Fxg1 + Fyg2 ∈ H P , OP (2n − 3) = K[x, y]2n−3. See [GKZ94, Example 2.11] for more information. Example 2.3.3. The projective B´ezoutformula for the discriminant arises when when we apply this method to ∆ = Σn,2 with pivot j = 1. The first page of the spectral sequence in this case is 0 1 1 OH ⊗K H P , OP (n − 2) 1 1 1 OH(−2) ⊗K H P , OP (−n) . There are no nonzero differentials. The second page of the spectral sequence therefore has the same −2,1 1 1 0 1 1 1 entries as the first page, with a map d2 : OH(−2)⊗K H P , OP (−n) → OH ⊗K H P , OP (n − 2) . −2,1 See [GKZ94, Proposition 5.4] for a proof that the B´ezoutformula is a matrix representing d2 . Example 2.3.4. We construct a resolution of a module M locally isomorphic to OΣ¯ 4,3 over the ambient space of ∆4. We set the pivot j = 1. The resolution will have the form - ⊕4 - ⊕6 - ⊕2 - - 0 OH(−3) OH(−2) OH M 0. ∂2 ∂1 We compute explicit formulae for the matrices representing ∂1 and ∂2. Removing from the double complex modules which have no effect on the computation, we obtain the diagram shown in Figure 2.7. For n = 4, we obtain the diagram shown in Figure 2.8. A basis of H1(P1, O(−5)) is {x−4y−1, x−3y−2, x−2y−3, x−1y−4} and a basis for H1(P1, O(−3)) is {x−2y−1, x−1y−2}. Chapter 2. Matrix Factorizations 33 Γ(Ux, O(n − 3)) ⊕ Γ(Uy, O(n − 3)) 6 −1,0 dV - ⊕3 ⊕3 - ⊕3 - 0 Γ Ux, O(−1) ⊕ Γ Uy, O(−1) Γ Uxy, O(−1) 0 d−1,0 H 6 −2,1 dV