Florida State University Libraries
Electronic Theses, Treatises and Dissertations The Graduate School
Secant Indices, Duality Defect, and Graeysnone S.r Joargleinzsoantions of the Segre Zeta Function
Follow this and additional works at the DigiNole: FSU's Digital Repository. For more information, please contact [email protected] FLORIDA STATE UNIVERSITY
COLLEGE OF ARTS AND SCIENCES
SECANT INDICES, DUALITY DEFECT, AND
GENERALIZATIONS OF THE SEGRE ZETA FUNCTION
By
GRAYSON S. JORGENSON
A Dissertation submitted to the Department of Mathematics in partial fulfillment of the requirements for the degree of Doctor of Philosophy
2020
Copyright c 2020 Grayson S. Jorgenson. All Rights Reserved. Grayson S. Jorgenson defended this dissertation on March 30, 2020. The members of the supervisory committee were:
Paolo Aluffi Professor Directing Dissertation
Svetlana Pevnitskaya University Representative
Ettore Aldrovandi Committee Member
Mark van Hoeij Committee Member
Kathleen Petersen Committee Member
The Graduate School has verified and approved the above-named committee members, and certifies that the dissertation has been approved in accordance with university requirements.
ii ACKNOWLEDGMENTS
I wish to thank Paolo Aluffi for all his guidance, encouragement, and support over the duration of my doctoral work. I always look forward to meeting with him and I have greatly enjoyed hearing his insights. His passion for math and for science and the arts as a whole has been an enormous inspiration to me, helping to drive my own work and curiosity. I am extremely grateful to have been able to work under his tutelage. Throughout my degree I felt fortunate to have been able to get to know many of the other great people in the FSU math department. In particular, I would like to thank the professors who taught me in a variety of courses in the earlier stages of my studies: Ettore Aldrovandi, Phil Bowers, Sergio Fenley, Wolfgang Heil, Alec Kercheval, Eric Klassen, Craig Nolder, and Mark van Hoeij. Thanks also to my fellow grad students, especially Arun, Ben, Braulio, Michael, Minfa, and Yi for many fun times and discussions. I would also like to express here my gratitude to Benjamin Hutz to whom I owe a large debt for playing a pivotal role in introducing me to the world of pure math during my undergraduate studies and for inviting me into several great opportunities over the years. Finally, I wish to thank my parents and sister for all of their love and support.
iii TABLE OF CONTENTS
List of Figures ...... vi Abstract ...... vii
1 Introduction 1 1.1 Secant indices ...... 1 1.2 Duality defect ...... 2 1.3 Relative Segre zeta functions ...... 5 1.4 Segre Zeta functions in the equivariant setting ...... 6
2 Secant Indices 9 2.1 Introduction ...... 9 2.2 General properties ...... 10 2.2.1 Strictness of growth ...... 12 2.3 Questions and a guiding principle ...... 16 2.4 Veronese varieties ...... 23 2.5 Segre varieties ...... 36 2.6 Lines on surfaces ...... 41
3 Linear Recurrence Sequences and the Duality Defect Conjecture 45 3.1 Introduction ...... 45 3.2 Preliminaries ...... 48 3.3 Dual varieties and duality defect ...... 49 3.4 A generalized algorithm and bounds for the degrees of counterexamples ...... 54 3.5 Duality defect in codimension 3 ...... 60
4 A Relative Segre Zeta Function 64 4.1 Introduction ...... 64 4.1.1 Motivation ...... 64 4.1.2 Statement of the results ...... 66 4.2 Proof of the main result ...... 70 4.3 Extension of bundles and sections ...... 75 4.4 The case of subbundles ...... 77 4.4.1 When F = E ⊕ L ...... 77 4.4.2 Reduction of the ambient space dimension ...... 79 4.5 Examples ...... 80 4.5.1 Subschemes of projective space ...... 80 4.5.2 Subschemes of products of projective spaces ...... 81
5 Equivariant Realization of the Segre Zeta Function 84 5.1 Introduction ...... 84 5.2 Segre classes ...... 86
iv 6 Conclusion 90
Bibliography ...... 92 Biographical Sketch ...... 96
v LIST OF FIGURES
(2) 2 2.1 Successive hyperplane sections of v3 (P )...... 25 2.2 Successive monomial additions to J for n = 2, d =4...... 26
vi ABSTRACT
This thesis consists of three independent projects in the field of algebraic geometry. The first of these is the focus of Chapter 2. There we define a family of integer sequences we n refer to as the family of secant indices of projective subvarieties. If X ⊆ P is a subvariety, its sequence of secant indices is denoted L(X) and records the maximal finite intersections of X with n linear subvarieties of P . We prove several properties about these sequences, develop a method to compute term-wise lower bounds and compute these lower bounds for Veronese and Segre varieties. We state several questions and conjectures regarding these sequences which to our knowledge are open. Chapter 3 concerns a project studying the so-called duality defect of projective subvarieties. n Given a projective subvariety X ⊆ P , one may consider its dual subvariety in dual projective space n ∨ (P ) . For low codimension smooth X, it would be a consequence of Hartshorne’s conjecture that the dual of X is always a hypersurface. This expectation is known as the duality defect conjecture, and in Chapter 3 we push a combinatorial approach of Holme and Oaland further to verify that the conjecture holds in the codimension 3 case for projective spaces of odd dimension, and also to derive a degree bound that any counterexamples would have to satisfy. Chapters 4 and 5 concern Segre classes of closed subschemes of projective bundles. By extending a given homogeneous ideal to progressively larger polynomial rings, one obtains an infinite sequence of closed subschemes of progressively higher dimensional projective spaces which are cones over the preceding subschemes. The Segre classes of these subschemes are related by the Segre zeta function defined by Aluffi. In Chapter 4, we generalize this Segre zeta function to subschemes of projective bundles over a smooth base variety. In Chapter 5 we explore a realization of the Segre zeta function in the context of equivariant algebraic geometry.
vii CHAPTER 1
INTRODUCTION
Here we summarize the content of the subsequent chapters while introducing some of the necessary prerequisite material.
First, some conventions we will fix throughout. We will take our base field to be k = C, though all the results of Chapter 4 hold when k is any algebraically closed field of characteristic zero. All schemes are assumed to be of finite type over k. A variety for us will always refer to an integral scheme, so in particular is assumed to be irreducible. All points are closed points.
1.1 Secant indices
n A fundamental observation in basic algebraic geometry is that if one has a subvariety X of P , n then for any hyperplane H ⊆ P , the dimension of X ∩ H is at least dim(X) − 1. Furthermore, it is always possible to find hyperplanes so that dim(X ∩ H) = dim(X) − 1 [32, Section I.7]. A n consequence of this is that if one has a linear subvariety L ⊆ P of dimension r, then dim(X ∩ L) ≥ dim(X) − n + r. Thus, if n − r ≤ dim(X), then the intersection X ∩ L must be nonempty. The case where dim(X) = n − r is special in that the smallest possible intersection X ∩ L will be a finite set of points. For “most” such L, X ∩ L will be finite. More precisely, the set of linear subvarieties of dimension r is parameterized by the Grassmannian G(r, n). When n − r = dim(X) there is a dense open subset of G(r, n) of linear subvarieties that all meet X in the same finite number of points. This maximal number of points that can appear in such an intersection is known as the degree deg(X) of X. By the same token, when dim(X) < n − r, then most L do not meet X at all; there is a dense open subset of G(r, n) of linear subvarieties with empty intersection with X. Despite this, nonempty intersections can still easily occur. A fundamental question is then to ask about the maximal possible number of points that can occur in such an intersection. Because most linear subvarieties of these smaller dimensions do not meet X, traditional intersection theory techniques
1 for addressing enumerative geometry problems cannot be used to produce exact answers. This question is an instance of a so-called quasi-enumerative geometry problem [20]. n In Chapter 2 we study this problem of special intersections. For a variety X ⊆ P and an integer 0 ≤ i ≤ codim(X) we define an integer
Li(X) := max{|X ∩ L|} where this maximum is taken over all dimension i linear subvarieties L with X ∩ L finite and reduced. We call the collection of these integers as a sequence
L(X) := (L0(X),..., Lcodim(X)(X)) the sequence of secant indices of X. We will derive several properties that these sequences satisfy. In particular, the sequence is always nondecreasing from left to right and the first and last numbers of the sequence are 1 and n deg(X), respectively. When X is not a minimal degree subvariety of P , that is, when deg(X) > codim(X) + 1, there must be at least one gap of size ≥ 2 in the sequence. Because of this, the sequences of secant indices for higher degree varieties seem to possess interesting combinatorial properties. We will investigate the cases where X is a Veronese or Segre variety in detail, obtaining term-wise lower bounds for L(X) that are conjecturally in fact equal to L(X). In the case of Veronese embeddings, the secant indices are related to the Cayley-Bacharach theorem and the open Eisenbud-Green-Harris conjecture.
1.2 Duality defect
N Given a variety X ⊆ P of dimension n, we say X is a scheme-theoretic complete intersection if the homogeneous ideal I(X) of X is generated by N − n homogeneous polynomials. A presently open conjecture was made over four decades ago by Hartshorne [31] that any smooth subvariety of N sufficiently small codimension in P should be a complete intersection. 2N More precisely, the conjecture states that a smooth X with n > 3 is a complete intersection. As of today, only partial progress toward this conjecture has been made. It is not even known whether all varieties of a fixed codimension ≥ 2 must be complete intersections when the dimension constraint is satisfied.
2 In Chapter 3 we study a related question concerning the duals of smooth subvarieties of pro- N jective spaces. Given a smooth n-dimensional subvariety X ⊆ P , one may define its dual in dual N ∨ N projective space (P ) . This dual projective space parameterizes the hyperplanes in P , and we ∨ N ∨ define the dual X ⊆ (P ) of X to be the collection of all “tangent hyperplanes” to X: those that contain at least one of the embedded tangent spaces of X. It is known that n ≤ dim(X∨) ≤ N − 1 always. The duality defect def(X) = N − 1 − dim(X∨) of X is designed to be a simple measure of how much X∨ fails to be a hypersurface. When X is a smooth complete intersection, it can be shown that X∨ is a hypersurface, and thus has duality 2N defect zero. Therefore, a reasonable conjecture to make is that when n > 3 , the duality defect of X is zero. A priori, this is a weaker version of Hartshorne’s conjecture. The conjecture has been attributed to Alan Landman and is referred to as the duality defect conjecture. This conjecture is notably weaker than Hartshorne’s conjecture in the sense that it has been fully proven in the codimension 2 case, see [15, Theorem 3.4] and [37, Corollary 6.4]. However, it still remains open in the higher codimension cases. One approach to proving the conjecture that we find particularly interesting is that of Holme [36] and his student Oaland [49].
To every scheme X one can associate the Chow group, A∗(X). This the quotient of the group L Z∗(X) = i≥0 Zi(X), where each Zi(X) is the free abelian group generated by the subvarieties of dimension i of X, by rational equivalence, see [23, Section 1.3]. Ai(X) denotes the corresponding i quotient of Zi(X) and we will often write A (X) := Adim(X)−i(X) when we wish to specify the groups by codimension. A proper map f : X → Y induces push-forward homomorphisms f∗ :
Ai(X) → Ai(Y ). When X is a smooth variety, there is an intersection product
Ai(X) × Aj(X) → Ai+j(X)
making A∗(X) into a ring sometimes denoted A(X) [23, Section 8.3]. For general schemes X, Y , flat morphisms f : X → Y induce pullback homomorphisms f ∗ : Ai(Y ) → Ai(X). Such pullback homomorphisms are defined also when f is only assumed to be a morphism between two smooth varieties X,Y [20, Theorem 1.23], and in this case they are also ring homomorphisms with respect to the intersection products defining A(X),A(Y ).
3 Given an algebraic vector bundle E on a scheme X, one may define the ith Chern class of E which is a homomorphism
ci(E): Ak(X) → Ak−i(X).
[23, Section 3.1]. Intuitively, once applied to the fundamental class of a variety, the Chern classes represent the rational equivalence classes of the loci where a set number of sections of E become linearly dependent, though to make this rigorous one must assume E has enough sections [20,
Section 5.3]. To denote ci(E) applied to an element α ∈ Ak(X) we write ci(E) ∩ α. The zeroth
Chern class is the identity endomorphism of A∗(X) and ci(E) is the zero map when i exceeds the rank of E. Thus one may formally invert the Chern polynomial
r ct(E) = 1 + c1(E)t + ... + crank(E)(E)t
dim(X) to obtain the Segre polynomial st(E) = 1 + s1(E)t + ... + sdim(X)(E)t of E. The map si(E): Ak(X) → Ak−i is called the ith Segre class of E. The combinatorial approach of Holme to address the duality defect conjecture is to consider ∨ the Segre classes of the bundle NX/PN (−1) , where NX/PN denotes the normal bundle of a smooth N N variety X in P . When applied to the fundamental class [X] and pushed forward to A(P ) via the N closed embedding i : X,→ P , the total Segre class s(E) ∩ [X] = (1 + s1(E) + ... + sn(E)) ∩ [X] becomes a class represented by an integer polynomial. The duality defect of X may be calculated from these integer coefficients and the coefficients in turn can be recursively expressed in terms of the Chern classes of NX/PN (−1). By capitalizing on numerical constraints that these Chern classes must satisfy, and the possibil- ities for homogeneous recurrence sequences with a prescribed set of zeros, in Chapter 3 we extend the work of Holme to prove that the duality defect conjecture is true for smooth codimension 3 sub- N varieties of P when N is odd. Along the way, using positivity properties of the Schur polynomials in these Chern classes, we prove also that if X is a smooth nonlinear codimension m subvariety of N 3N−2 P with positive duality defect r > 0 and N − m ≥ 4 , then
m j X N − m − r deg(X) ≤ . 2 j=0
4 1.3 Relative Segre zeta functions
n For a closed subscheme i : Z,→ P , there is another type of Segre class, the Segre class of Z n n in P , s(Z, P ) ∈ A∗(Z). When Z is a smooth variety, this Segre class agrees with the one defined n n above for bundles in the sense that s(Z, P ) = s(NZ P ) ∩ [Z]. n n The Segre class s(Z, P ) contains information about the embedding of Z into P . While A∗(Z) is often a group without a simple presentation, by considering the pushforward
n n ∼ n+1 i∗s(Z, P ) ∈ A(P ) = Z[h]/(h ),
n one may still recover much information about s(Z, P ). Understanding the integer coefficients of n the lowest degree polynomial representing i∗s(Z, P ) is thus an important task. n A closed subscheme i : Z,→ P (x0 : ... : xn) corresponds to a homogeneous ideal I ⊆ k[x0, . . . , xn]. One can extend this ideal to a larger polynomial ring k[x0, . . . , xn] ⊆ k[x0, . . . , xN ], N N ≥ n. If we denote by IN the extended ideal, then we obtain a closed subscheme iN : ZN ,→ P N that acts as a cone over Z in P ; at least set-theoretically, Z is the image of ZN of the linear N n projection of P onto P away from the linear subvariety V (x0, . . . , xn).
In this way a given homogeneous ideal I defines an infinite sequence of cones ZN over Z.A reasonable question to ask then is whether the Segre classes of these subschemes are related. Recent work of Aluffi [5] answered this question by showing all these Segre classes could be obtained from a rational power series, known as the Segre zeta function ζI (t) of I. This function is defined to be the power series X j ζI (t) = ajt ∈ Z[[t]] j≥0 satisfying N ζI (HN ) = (iN )∗s(ZN , P )
N N for every N ≥ n, where HN ∈ A∗(P ) is the rational equivalence class of a hyperplane in P .
More specifically, if IN has homogeneous generators F1,...,Fr and if Xj denotes the closed subscheme cut out by Fj for each j, then we have the following fiber square:
N ZN P δ ∆ Qr Qr N j=1 Xj j=1 P
5 Qr N The normal bundle of the bottom regular embedding is a restriction of a bundle on j=1 P . N Denote by E the pullback of this bundle to P . Then tautologically we may always write
N N −1 N c(E) ∩ (iN )∗s(ZN , P ) (iN )∗s(ZN , P ) = c(E) ∩ c(E) ∩ (iN )∗s(ZN , P ) =: . c(E) ∩ [PN ]
The key point in showing that ζI (t) is a rational power series is showing that this “numerator”
N c(E) ∩ (iN )∗s(ZN , P ) has a fixed representation as an integer polynomial even as N increases. In Chapter 4 we define a relative version of this Segre zeta function that generalizes Aluffi’s Segre zeta function to describe the Segre classes of cones over closed subschemes of projective bundles with smooth varieties as bases. Let E,F be bundles over a smooth variety X, and suppose we have a surjection F → E of bundles. This surjection induces a rational map P (F ) 99K P (E). Given a closed subscheme i : Z,→ P (E) that is suitably defined, we can use this rational map to define the cone ˆi : Z,ˆ → P (F ) over Z in P (F ). The precise requirement is that Z is the zero scheme of a section s of a bundle G of rank < rank(E) on P (E), and there exists a sufficiently “compatible” bundle and section on P (F ). We define a formal power series with coefficients in A∗(X) denoted ζG,s(t) which is rational, equal to
P (t) , Q(t) for some polynomials P (t),Q(t) ∈ A∗(X)[t] depending only on the representations of the classes c(G) ∩ i∗s(Z,P (E)) and c(G) ∩ [P (E)] in A∗(P (E)). We show that this same rational expression describes the Segre class ˆi∗s(Z,Pˆ (F )) of Zˆ in P (F ) when such a cone Zˆ exists.
1.4 Segre Zeta functions in the equivariant setting
n Consider again the setup described above for a subscheme i : Z,→ P cut out by a homogeneous ideal I = (F1,...,Fr), where we have the fiber diagram:
n Z P δ ∆ Qr Qr n j=1 Xj j=1 P
6 As before, Xj denotes the zero scheme of Fj. One can define the scheme n M n n+1 C := CZ P = Spec I /I n≥0
n n called the normal cone to Z in P [23, B.6]. Here I denotes the ideal sheaf of Z in P . By definition, n X j s(Z, P ) = p∗ c1(O(1)) ∩ [P (C)] j≥0 [23, Section 4.2], where M n n+1 p : P (C) = Proj I /I → Z n≥0 is the projection map and O(1) denotes the hyperplane bundle on P (C). The normal cone embeds into the bundle N that is the pullback of the normal bundle of the bottom map in the fiber square to Z. If we denote by s : Z,→ N the zero section embedding, then ∗ there is a Gysin homomorphism s : A∗(N) → A∗(Z). Then by [23, Proposition 6.1 (a)], we see that ∗ n s [C] = {c(N) ∩ s(Z, P )}n−r, where [C] is the fundamental class of the normal cone C inside N, and the curly braces notation
{·}k specifies that we are taking the dimension k piece of the class within. n n If we push forward the class c(N) ∩ s(Z, P ) to A(P ) we will obtain the numerator of the n Segre zeta function for I. What this result is saying is that the term {c(N) ∩ s(Z, P )}n−r of the numerator corresponds directly to the fundamental class of the normal cone C in N. This term also n may be thought of as the intersection product X1 ·...·Xr of the divisors X1,...,Xr in P [5, Section 2]. In other words, the class s∗[C] only represents a single piece of the information of the numerator of ζI (t). In Chapter 5 we will examine the analogous situation in equivariant algebraic geometry ∗ ∗ where N is endowed with the fiberwise action of C and Z with the trivial C action. Then as it is a subcone of N, C is invariant under this action and therefore defines an equivariant fundamental class
∗ C [C]C∗ ∈ A∗ (E).
7 We will show there, in slightly greater generality, that this equivariant class of the normal cone captures the entirety of the information of the numerator of the Segre zeta function, rather than just that of one term.
8 CHAPTER 2
SECANT INDICES
2.1 Introduction
n Given a pure-dimensional closed subscheme X of P of codimension m, and a choice of integer n 0 ≤ j ≤ m, let Λj(X) ⊆ G(j, n) denote the subset of dimension j linear subvarieties of P with finite and reduced intersection with X. Then one may define
Lj(X) := max{|X ∩ L| | L ∈ Λj(X)}.
Here | · | is used to denote the set-theoretic count of the points of the scheme within. For our purposes, we refer to this integer Lj(X) as the jth secant index of X, and together these numbers form a sequence of length m + 1 starting at 1 and ending at deg(X), which we denote by L(X). Similar numbers have been studied independently, such as in [10] [48] [47] [43] [2]. It is known that the presence of a m-multisecant line, a line meeting X finitely in at least m points, implies that the Castelnuovo-Mumford regularity reg(X) of X is at least m [43, Proposition 1.1]. The cited references focus on classifying varieties with extremal secant subspaces with one of the goals being to provide examples of varieties with near maximal regularity. Such work provides evidence for the Eisenbud-Goto regularity conjecture [16] that reg(X) ≤ deg(X)−m+1 when X is a nondegenerate n codimension m subvariety of P . In this work, we are instead concerned with the properties that these indices satisfy collectively, as sequences. The sequences that occur in general seem to admit interesting combinatorial descrip- tions. For instance, given a smooth variety X, L(X) is always strictly increasing. Yet if X is not a n subvariety of minimal degree of P , then this sequence must contain gaps and these gaps need not occur only in one place in the sequence. One of the simplest examples of this is when X is the image 2 9 (2) 2 of the degree 3 Veronese embedding of P into P ; the sequence L(v3 (P )) is this case has 8 terms, (2) 2 but there are 9 numbers in the sequence 1, 2,..., 9. Here we have L(v3 (P )) = (1, 2, 3, 4, 5, 6, 7, 9) and the observation that the gap occurs between the final two terms of the sequence is tantamount to the classical Cayley-Bacharach theorem.
9 We attempt the development of a method to compute the secant indices for an arbitrary smooth variety that involves defining two accessory sequences to L(X), denoted by RLG(X) and RL(X), that are term-wise lower bounds for L(X). In the specific cases where X is a Veronese or Segre variety, we show that computing these accessory sequences is equivalent to solving two attractive combinatorial problems. These accessory sequences then produce conjectural values for L(X) for those varieties. In the case of Veronese varieties, we will show that the truth of the Eisenbud-Green- Harris conjecture [17] would imply the sequences RLG(X), RL(X) are indeed equal to L(X). The definitions of RLG(X) and RL(X) work for arbitrary smooth varieties, and the hope is that for every variety X there is a tractable combinatorial problem associated to computing these sequences, reflecting the nature of intersections of hyperplane sections of X which are as reducible as possible. This chapter is organized as follows. In Section 2.2 we describe properties that the sequence of secant indices satisfy in general, including the strictness of their growth in the case of a smooth n nondegenerate subvariety of P . We state a number of questions about the secant indices in Section 2.3 which to our knowledge are open, and define the integer sequences RLG(X), RL(X). In Section 2.4 we illustrate the accessory sequences and show they are equal when X is a Veronese variety. We prove a method that computes them and show that it would follow from the Eisenbud-Green-Harris conjecture that the sequences agree with L(X). In Section 2.5, we derive an algorithm to compute RLG(X) when X is a Segre variety and provide several computations. In Section 2.6 we discuss our original motivation of the problem of counting lines on surfaces, where our hope is that one can compute these maximal numbers by extrapolation from a related sequence of indices.
2.2 General properties
The most basic form of the main question we study in this chapter is as follows: what is the maximum number of points at which a linear subvariety of a given dimension can meet a pure- dimensional projective algebraic set X? This is only interesting when the linear subvariety has dimension small enough to meet X in finitely many points, thus we ask about the numbers
n max{deg(X ∩ L) | L ⊆ P linear, dim(L) = i, dim(X ∩ L) = 0}, for 0 ≤ i ≤ codim(X). The first of these, the maximum number of points that a single point can meet X, is clearly just 1.
10 However, for the other numbers there are a couple technicalities that require clarification. First, the exact method of counting the points in the intersections must be made precise. It would be nice to know that the last number in this sequence is always deg(X). If we count the points of a zero-dimensional scheme Y by letting deg(Y ) denote the scheme-theoretic degree rather than the set-theoretic count of the distinct closed points of the support of Y , this is not always the case.
4 Example 2.2.1. Consider the union X of two planes in P meeting at a single point. The degree of X is 2, but it is well-known that deg(X ∩ L) = 3 for any plane L meeting X at only its singular point.
To avoid more delicate issues like this, we opt to instead use the naive set-theoretic count, considering the numbers
n max{|X ∩ L| | L ⊆ P linear, dim(L) = i, dim(X ∩ L) = 0}, for 0 ≤ i ≤ codim(X). We still reserve deg(·) to denote scheme-theoretic degree, and instead denote by |Y | the cardinality of the set of closed points of a zero-dimensional scheme Y . The resulting sequence will always be nondecreasing.
Proposition 2.2.2. Suppose L is a linear subvariety of dimension < codim(X) with X ∩ L zero- dimensional. Then there exists a linear subvariety L0 of dimension dim(L) + 1 with X ∩ L0 zero- dimensional, and L ⊆ L0.
This version of the question fell out from our original motivation which is discussed in Section 2.6. One of the properties that seems reasonable to expect is that for nondegenerate X this sequence should in fact be strictly increasing. To prove such a result we introduce one last refinement: we require the intersections we are counting to be reduced, and we restrict our attention to smooth projective varieties. There is then no distinction between using the set-theoretic count or the scheme-theoretic degree. In this section and Sections 2.3, 2.4, and 2.5, the sequences sporting these additional properties are our objects of study.
Example 2.2.3. Consider the rational normal curve C that is the image of the Veronese map (1) 1 d vd : P ,→ P , a degree d smooth curve. The sequence L(C) has d terms, starting at 1 and ending at d. It is indeed the only strictly increasing sequence of that length connecting those two numbers,
11 L(C) = (1, 2, 3, . . . , d). Images of Veronese maps will be discussed further in Section 2.4. But the computation of the secant indices of rational normal curves can be dealt with much more easily: it d is well-known that any d + 1 points on C are linearly independent, that is, they span all of P . In particular, any r points on C are linearly independent, and so the most number of points a linear subvariety of dimension r can meet C is r +1. The intersection is reduced if it contains r +1 points.
0 2 2 Example 2.2.4. Suppose C is a smooth curve of degree d in P , and embed P as a plane in 3 3 P . Thus we obtain a degenerate degree d curve in P . It is clear that its secant indices are L(C) = (1, d, d). n More generally, if Y is a smooth subvariety of P , and Y is contained in a linear subvariety of dimension r, then the last n − r terms of L(Y ) are all deg(Y ). For another specific example, if Y is itself linear, then L(Y ) = (1, 1,..., 1).
2.2.1 Strictness of growth
n The sequence of secant indices of a smooth variety X in P is always nondecreasing. If X is nondegenerate, then the sequence is in fact strictly increasing. If X is degenerate, then L(X) has repeated terms as in Example 2.2.4, but is otherwise strictly increasing. We derive this property below using an elementary argument revolving around Bertini’s theorem. Note this argument is similar in essence to the elementary approach seen in [43] for deriving the upper bound on the secant indices, and indeed this upper bound is an immediate consequence of the strictness of growth we derive here. Many of the lemmas used below are well-known results, but for lack of appropriate references we give complete proofs for most of them here. The specific instance of Bertini’s theorem we will invoke throughout is the following.
Lemma 2.2.5. Suppose X is a smooth projective variety. Then a general element of a positive- dimensional linear system on X is smooth away from the base locus of the system.
As an immediate consequence, we have:
n Lemma 2.2.6. Let L be a linear subvariety and let X be a smooth subvariety of P . The set of all n n ∨ hyperplanes in P containing L forms a linear subvariety T of dual projective space (P ) . There is a nonempty open subset T of hyperplanes containing L and having smooth intersection with X outside of X ∩ L.
12 n For any closed subscheme Y of P (x0 : ... : xn) and a closed point p ∈ Y , we denote by Tp(Y ) the projective tangent space to Y at p. If Y = V (I) for a homogeneous ideal I with generators
F1,...,Fr ∈ k[x0, . . . , xn], then the tangent space is the linear subvariety cut out by the polynomials
∂Fi ∂Fi (p)x0 + ... + (p)xn, ∂x0 ∂xn for i = 1, . . . , r. The subscheme Y is smooth at p if and only if
dim(Tp(Y )) = dim(Y ) by the Jacobian criterion for singularities. Following immediately from this definition:
n Lemma 2.2.7. Let Y be a closed subscheme of P , and let L be a linear subvariety. Suppose p is a closed point of the scheme-theoretic intersection Y ∩ L. Then Tp(Y ∩ L) = Tp(Y ) ∩ L.
n Lemma 2.2.8. Let Y be a closed subscheme of P and let L be a linear subvariety of dimension < n − 1. Suppose L has reduced zero-dimensional intersection with Y , and the intersection consists n ∨ of the points p1, . . . , pr. Then if T is the linear subvariety of (P ) consisting of all hyperplanes containing L, there is an nonempty open subset of T of hyperplanes H with Y ∩ H smooth at the points p1, . . . , pr.
Proof. By hypothesis, dim(Tpi (Y )) = dim(Y ) for each i. Since Y ∩L is reduced, the points p1, . . . , pr are smooth points of Y ∩ L. By the Lemma 2.2.7 this means that for each i, the intersection
Tpi (Y ) ∩ L is a single point.
Fix an i. For a hyperplane H from T , the only way that Y ∩ H can be non-reduced at pi is n if H contains Tpi (Y ). The set of all hyperplanes in P that contain Tpi (Y ) is a linear subvariety n ∨ of (P ) , and so the set of all hyperplanes from T that contain Tpi (Y ) is a closed subset of T . It T must in fact be a proper closed subset, since L ∩ Tpi (Y ) is a single point, and H∈T H = L. Thus its complement in T is a nonempty open subset. The intersection of any finite collection of nonempty open subsets of a variety is nonempty and open, and so ends the proof.
Verifying that L(X) is strictly increasing is not difficult when X is a curve.
n Lemma 2.2.9. Let C be a nondegenerate, smooth, and irreducible curve in P . Then for any hyperplane H with C ∩ H reduced, the points of C ∩ H must span H.
13 Proof. Note we know that C ∩ H consists of exactly deg(C) points since C is nondegenerate and C ∩H is reduced. Suppose these points do not span H. Then their span is a proper linear subvariety n n L of H, so dim(L) ≤ n − 2. Then the set of all hyperplanes in P containing L covers all of P . Therefore, as C \ L is nonempty, we can choose a point p in C \ L, and there exists a hyperplane H0 containing L and also p. So C ∩ H0 consists of more than deg(C) points, which is impossible since C is nondegenerate.
n Lemma 2.2.10. Let C be a nondegenerate, smooth, and irreducible curve in P . Then L(X) is strictly increasing.
Proof. Let L be a linear subvariety with reduced intersection with C. If L is a hyperplane, there is nothing to show, so assume dim(L) < n − 1. n ∨ The set of all hyperplanes containing L is a linear subvariety of (P ) which induces a positive- dimensional linear system on C. By Bertini’s theorem and Lemma 2.2.8, there is a nonempty open subset of T consisting of hyperplanes with reduced intersection with C. We thus can pick such a hyperplane H. But by Lemma 2.2.9, the points of C ∩ L can only span at most L, which is properly contained in H. Thus C ∩ H must contain at least one additional point p.
If dim(L) = n − 2, then we are done; if L is taken to be linear subvariety realizing Ln−2(C), then we have shown Ln−2(C) < Ln−1(C). Otherwise, we can pick a linear subvariety L0 of dimension one greater than dim(L) and with 0 0 L ⊆ L ⊆ H, and p ∈ L . Thus Ldim(L)(C) < Ldim(L)+1(C).
The idea to get the general result is to reduce to the case of a curve when dealing with a higher dimensional variety. First, note that a reduced hyperplane section of a nondegenerate variety is nondegenerate, inside of the hyperplane, a sort of extension of Lemma 2.2.9.
n Lemma 2.2.11. Let X be a nondegenerate subvariety of P of dimension > 1 and suppose H is a ∼ n−1 hyperplane such that X ∩ H is reduced. Then X ∩ H is nondegenerate in H = P .
∼ n−1 Proof. Suppose to the contrary that there is a hyperplane L in H = P with X ∩ H ⊆ L. Since 0 n X is nondegenerate, there exists a point p ∈ X \ H. Thus we can find a hyperplane H of P containing both L and p. Further, H ∩ H0 = L.
14 The intersection X ∩ H0 is pure-dimensional of dimension dim(X) − 1, thus p is a point on an irreducible component Y of X ∩ H0 of that dimension. Then H ∩ Y must be a proper closed subset of Y , thus of smaller dimension, and therefore contained in one of the irreducible components of X ∩ H. 0 Let X1,...,Xr,Y,Z1,...,Zm be the irreducible components of X ∩ H , all considered with reduced scheme structures. Here X1,...,Xr are the irreducible components of X ∩ H. Since X ∩ H is reduced, we have
r X deg(X ∩ H) = deg(Xi), i=1 while
r m 0 X 0 0 X 0 deg(X ∩ H ) = mXi (X,H ) deg(Xi) + mY (X,H ) deg(Y ) + mZi (X,H ) deg(Zi). i=1 i=1
The notation mZ (A, B) stands for the intersection multiplicity of the intersection of two varieties A, B along an irreducible component Z of A ∩ B, as in [20]. This is a contradiction since the intersection multiplicities are positive and we must have deg(X∩ H) = deg(X ∩ H0), see [32, Theorem I.7.7].
n Theorem 2.2.12. Suppose X is a smooth nondegenerate subvariety of P . Then L(X) is strictly increasing.
Proof. By Lemma 2.2.10, we may assume dim(X) > 1. Suppose that L is a linear subvariety of dimension r < codim(X) so that X ∩ L is reduced and zero-dimensional, and |X ∩ L| = Lr(X). By Lemma 2.2.6 and Lemma 2.2.8, there is a hyperplane H containing L so that X ∩ H is smooth. The Fulton-Hansen connectedness theorem [24] implies any hyperplane section of X is connected, so the hypothesis that X ∩ H is smooth implies it is also irreducible. The previous lemma then shows X ∩ H is nondegenerate as a subvariety of H. Thus by induction, we may assume the existence of a linear subvariety T containing L of dimension n − dim(X) + 1 so that T ∩ X is a smooth, irreducible, nondegenerate curve in T . By Lemma 2.2.10, there exists a linear subvariety L ⊆ L0 ⊆ T of dimension r + 1 such that
|X ∩ T ∩ L0| > |X ∩ T ∩ L|.
15 Therefore
Lr+1(X) > Lr(X).
Note, as a minor consequence, this gives a slightly different way to think about the degree lower bound that all nondegenerate projective subvarieties satisfy, see for instance [19]. For smooth nondegenerate X, the fact that L(X) is strictly increasing forces deg(X) ≥ codim(X) + 1. That X is a variety is also essential. Secant indices of smooth, pure-dimensional, nondegenerate, and reduced subschemes do not necessarily form strictly increasing sequences.
3 Example 2.2.13. Consider the smooth curve C in P that is the union of three skew lines L1,L2,L3 all passing through another line L. Then L1(C) = 3, as L ∩ C consists of three distinct points, but any plane that contains L and meets C at points outside of L ∩ C must contain one of L1,L2,L3. So L(C) = (1, 3, 3) in this case.
One other immediate and basic consequence of the strictly increasing property is recovering the known upper bound for the cardinality of intersections with extremal secant spaces.
Proposition 2.2.14. Tautologically,
codim(X)−1 X Li(X) = deg(X) − (Lj+1(X) − Lj(X)), j=i for i = 0,..., codim(X) − 1. Thus in particular,
Li(X) ≤ deg(X) − codim(X) + i for each i.
See also Kwak [43]. When i = 1, this bound has been used as evidence for the Eisenbud-Goto regularity conjecture.
2.3 Questions and a guiding principle
n One of our main interests is a means of computing these numbers for a given subvariety X ⊆ P but this seems difficult for arbitrary subvarieties. It is clear that one method of obtaining a term- n wise lower bound for L(X) is to take a linear subvariety L ⊆ P of dimension codim(X) so that
16 X ∩ L is finite and reduced, and compute the sequence L(X ∩ L) where X ∩ L is considered as a ∼ codim(X) subscheme of L = P . r r r r First, for two integer sequences (aj)j=1, (bj)j=1 of the same length r, we write (aj)j=1 (bj)j=1 if aj ≤ bj for each j. This is a partial order on the set of all integer sequences of the same length. Additionally, we can define a total order on the set of integer sequences of the same length r as r r follows: (aj)j=1 ≤ (bj)j=1 if and only if either the sequences are equal, or there is a 1 ≤ k ≤ r such that ak < bk, and aj ≤ bj for every j > k.
n n Proposition 2.3.1. Let X ⊆ P be any smooth subvariety, and let H ⊂ P be any hyperplane not containing X so that X ∩ H is also smooth. Then
L(X ∩ H) L(X)
∼ n−1 where X ∩ H is considered as a subvariety of H = P .
A natural question then is to ask when is this all that needs to be done.
n Question 2.3.2. For what subvarieties X of P is L(X) realized by L(X ∩L) for a linear subvariety L of dimension codim(X) such that X ∩ L is finite and reduced? Further, when is L(X) realized as a sequence of the form
(|X ∩ (H1 ∩ ... ∩ Hn)|,..., |X ∩ (H1 ∩ ... ∩ Hdim(X))|), for linearly independent hyperplanes H1,...,Hn?
Question 2.3.3. Consider Y := X ∩L for a linear subvariety L of dimension codim(X) with X ∩L reduced and finite. We can define two integer sequences:
1.
(max{|Y ∩ L0|},..., max{|Y ∩ Lcodim(X)|}),
2.
max{(|Y ∩ L0|,..., |Y ∩ Lcodim(X)|)}.
Here each Lj denotes a linear subvariety of dimension j contained inside L. The maximums in (1) are taken over all possible Lj, and the maximum in (2) is taken over all chains L0 ⊆ ... ⊆ Lcodim(X) using the total order ≤ defined above. Is it always the case that these two sequences are the same?
17 n Note that the answer to Question 2.3.3 is negative for arbitrary finite subsets Y of P .
Example 2.3.4. Let Y be a finite set consisting of 3 points p1, p2, p3 on a line T and 5 points 3 q1, . . . , q5 on a plane H in L = P . Suppose T is not contained in H, and p1, p2, p3 are not the point of intersection T ∩ H. Suppose also that the five points on H are arranged so that no three of them are collinear, and that no two of them lie on a line containing T ∩ H. Then the sequence from (1) of Question 2.3.3 is (1, 3, 5, 8), and the sequence from (2) is (1, 2, 5, 8).
At least for the case of Veronese varieties X considered in Section 2.4, the truth of the Eisenbud- Green-Harris conjecture would positively answer Question 2.3.2. In the same section we will also prove that the two sequences of Question 2.3.3 are equal, as a consequence of the Clements- Lindstr¨omtheorem. n For any finite set Y of points in P , computing L(Y ) is equivalent to considering the dimensions of the spans of all subsets of of Y , considered as subsets of points in the vector space kn+1. Another n question then becomes to ask for a subvariety X of P about what possible linear dependences between the deg(X) points of a reduced and finite intersection X ∩L where L is linear of dimension codim(X) occur as L is varied among all such linear subvarieties. One could phrase this in terms of matroids.
n Question 2.3.5. Let X ⊆ P be a variety. For each linear subvariety L of dimension codim(X) and X ∩ L finite and reduced, the subsets of linearly independent points of X ∩ L considered inside kn+1 form a matroid. What matroids can be realized in this way?
Finally, what sequences of integers can be realized as a sequence of secant indices?
n Question 2.3.6. If X is a smooth, nondegenerate subvariety of P , L(X) is a strictly increasing sequence of integers from 1 to deg(X) of length codim(X) + 1. What strictly increasing sequences of this length from 1 to deg(X) occur in this way?
Question 2.3.7. For X for which deg(X) exceeds codim(X) + 1, in what positions and in what sizes do the gaps in L(X) occur?
18 To our knowledge, the above questions have received little prior study, if any at all. For this last question, some of the related work on classifying varieties with extremal secant spaces provides a partial result about the existence of a gap between the last two terms of L(X). One such result is the following due to Kwak [43, Proposition 3.2].
n Proposition 2.3.8. Let X be a nondegenerate subvariety of P of dimension ≥ 1, and codimension ≥ 2. If X has an extremal curvilinear secant subspace of dimension 1,..., codim(X) − 1, then X is
5 1. a Veronese surface in P ,
4 2. a projected Veronese surface in P ,
3. a rational scroll.
As an immediate corollary of this result, we can classify all varieties that do not have a gap between the final two terms of L(·).
n Corollary 2.3.9. Let X be a nondegenerate smooth subvariety of P . Then unless X is a rational 5 4 scroll, the Veronese surface in P , or a projected Veronese surface in P ,
Lcodim(X)(X) − Lcodim(X)−1(X) ≥ 2.
Proof. The only work that needs to be done is reconcile our language with that used in the cited reference. A linear subvariety L is said to be a curvilinear secant subspace to X if X ∩ L if finite with each point of X ∩ L locally contained in a smooth curve on X. This last criterion is equivalent to specifying that dim(Tp(X) ∩ L) ≤ 1 for each point p ∈ X ∩ L. Such an L is called extremal if its intersection contains the maximal possible number of points, counted with appropriate multiplicity, length(X ∩ L) = deg(X) − codim(X) + dim(L). Our point of view in this chapter predominantly takes the more naive route of considering only reduced intersections; we only consider linear subvarieties L with X ∩ L reduced and finite. This ensures that dim(L ∩ Tp(X)) = 0 for each p ∈ X ∩ L and thus, in particular, such an L is a curvilinear secant subspace to X. Therefore, the result of Kwak implies that the bound seen in Proposition 2.2.14 is in fact strict for each i = 0,..., codim(X) − 1.
19 Kwak’s result along with similar work [48] on bounding the maximal possible lengths can be used in this way to treat the question of whether there is a gap between the final two terms of L(X) for all smooth nondegenerate varieties. But these results do not provide lower bounds for the terms of the sequence, and cannot be used to say more about the size of the penultimate gap, nor about the presence of other gaps in the sequence. Returning to the goal of computing L(X), a natural attempt to reduce the complexity of this computation is to compute the indices that arise when we only use a subset of the possible linear subvarieties. In particular, it seems reasonable to expect that the secant indices are the same if we were to only consider linear subvarieties cut out by hyperplanes that meet X in the most “reducible way” possible. This leads us to define two additional accessory sequences to L(X) which in some cases, such as those considered in Sections 2.4 and 2.5, become tractable to compute. To get a precise notion that generalizes beyond Veronese and Segre varieties we need to make several definitions.
n Definition 2.3.10. Let X be a smooth nondegenerate subvariety of P of dimension r.
• We say X is p-reducible if there exists a collection of hyperplanes H1,...,Hn so that their
common intersection is a single point, H1 ∩ ... ∩ Hr ∩ X is finite and reduced, each Hj ∩ X for j = 1, . . . , r is reduced and has exactly p distinct irreducible components, and finally
|H1 ∩ ... ∩ Hj ∩ X| > |H1 ∩ ... ∩ Hj+1 ∩ X| for each j = r, . . . , n − 1. Any such sequence of hyperplanes is said to satisfy the conditions of p-reducibility.
• We call the maximal p such that X is p-reducible the reducibility of X.
R • Suppose X has reducibility p. Denote by Λj (X) ⊆ Λj(X) the subset of linear subvarieties of dimension j cut out by hyperplanes H1,...,Hn−j such that the H1,...,Hn−j are the initial part of a sequence of hyperplanes satisfying the conditions of p-reducibility.
Note that by Bertini’s theorem, every smooth nondegenerate subvariety is at least 1-reducible.
n Proposition 2.3.11. Let X ⊆ P be a smooth nondegenerate subvariety. Then X is 1-reducible.
Proof. First, we can find hyperplanes H1,...,Hdim(X) so that each X ∩Hj is smooth and irreducible and so that H1 ∩ ... ∩ Hdim(X) ∩ X is finite and reduced by the classical Bertini theorem.
Let L = H1 ∩ ... ∩ Hdim(X) and note that the points of X ∩ L span L since X ∩ L is reduced so consists of deg(X) distinct points. By Lemma 2.2.8, if this were not the case, we would be able
20 to find a linear subvariety of dimension codim(X) meeting X at more than deg(X) points, despite the intersection being finite. Choose a subset of the points of X ∩L that span a linear subvariety L0 of dimension dim(L)−1.
Then by Lemmas 2.2.6, 2.2.8, we may find a hyperplane Hdim(X)+1 not containing L but containing 0 0 L so that X ∩ Hdim(X)+1 is smooth and irreducible. Since L has smaller dimension than L, we see that
|H1 ∩ ... ∩ Hdim(X) ∩ X| > |H1 ∩ ... ∩ Hdim(X)+1 ∩ X|.
Repeating this process completes the needed sequence of hyperplanes satisfying the conditions of 1-reducibility.
In general, it seems to be an independently interesting question about what the reducibility of a particular variety is. However, in the cases studied in Sections 4, 5, there is no mystery about n m the reducibility of the varieties in consideration. The reducibility of the Segre variety σ(P × P ) (n) n is 2 when n, m > 0, and that of the Veronese variety vd (P ) is d. At this point, one could define a new sequence which is a lower bound for L(X) at each term R by modifying the definition of Li(X) to only use linear subvarieties from Λ (X) instead of from Λ(X). This sequence seems interesting, but still appears challenging to compute. One difficulty in computing this new sequence is controlling how many irreducible components must be considered. To define the two accessory sequences to L(X) we will restrict the number of components.
n Definition 2.3.12. Let X ⊆ P be a smooth subvariety with reducibility p. Consider the set of all sequences of hyperplanes H1,...,Hn satisfying the conditions of p-reducibility. S • For each, consider the number of irreducible components in the union (Hj ∩ X) which have
nonempty intersection with H1 ∩ ... ∩ Hdim(X) ∩ X. Denote by µ(X) the minimal number of such irreducible components attained by the union of one of these sequences.
• Denote by H(X) the set of all sequences of hyperplanes (H1,...,Hn) satisfying the conditions S of p-reducibility and with the number of irreducible components of (Hj ∩ X), which each
have nonempty intersection with H1 ∩ ... ∩ Hdim(X) ∩ X, equal to µ(X).
Note that µ(X) ≥ p dim(X) always. If this were not the case, then this would contradict the p-reducibility properties of any sequence of hyperplanes H1,...,Hn realizing µ(X): either some of the hyperplane sections H1 ∩ X,...,Hdim(X) ∩ X would have to share an irreducible component
21 that meets the intersection H1 ∩ ... ∩ Hdim(X) ∩ X, which would violate the condition that H1 ∩
... ∩ Hdim(X) ∩ X is finite, or some of the Hj ∩ X would have to be nonreduced, for j ≤ dim(X). Once more, it seems to be an interesting question for arbitrary X what the value of µ(X) is. However, the for both the Veronese and Segre varieties we will consider, the minimal possible value p dim(X) is attained. We now are able to define the first of the accessory sequences to L(X).
Definition 2.3.13. Let X be as in the previous definition. Then if codim(X) = m, we define for each 0 ≤ j ≤ m an integer
n−j \ RLj(X) := max{|X ∩ L| | L = Hi, (H1,...,Hn) ∈ H(X)}, i=1 called the jth reducible secant index. As for the original secant indices, we denote these numbers collectively by
RL(X) := (RL0(X),..., RLm(X)).
n With this new terminology, for a given smooth subvariety X of P with reducibility p, we can now define the final sequence we are interested in.
Definition 2.3.14. We define the sequence of greedy reducible secant indices of X as
G RL (X) := max{(|X ∩ (H1 ∩ ... ∩ Hn)|,..., |X ∩ (H1 ∩ ... ∩ Hdim(X))|)}, where this maximum is taken over all (H1,...,Hn) ∈ H(X) and the sequences of integers are compared using the total order ≤ defined previously.
Altogether, we see that RLG(X) RL(X) L(X).
It seems conceivable that for many subvarieties X these three sequences are in fact equal. In Section 2.4 we will compute RLG(X), RL(X) when X is a Veronese variety, and will show they are equal. There we show also that the truth of the Eisenbud-Green-Harris conjecture would imply all three sequences are equal. With so little evidence however, we refer to the expectation in general that these sequences should all be equal as a “guiding principle” rather than a conjecture.
n Guiding principle. For any smooth nondegenerate subvariety X ⊆ P ,
RLG(X) = RL(X) = L(X).
22 2.4 Veronese varieties
In this section we consider the secant indices of the images of the Veronese embeddings. (n) (n) n Nd −1 n Throughout we will denote by vd : P ,→ P the degree d Veronese embedding of P into (n) (n) n+d (1) Nd −1 1 P , where Nd = n . In Example 2.2.3, we saw that L(vd (P )) = (1, 2, . . . , d) for every d > 0. (2) 2 5 Similarly, by Theorem 2.2.12, since the Veronese surface v2 (P ) has degree 4 in P , that is, it (2) 2 is a minimal degree subvariety, we see that L(v2 (P )) = (1, 2, 3, 4). The degree of the image of the (n) n Veronese embedding vd is d , and these sequences begin to become interesting when the degree exceeds the codimension of the image by more than one. (2) 2 The first case where this happens is for X = v3 (P ). This is a degree 9 subvariety of codi- 9 mension 7 in P . Thus L(X) has 8 terms, but there are 9 numbers in the sequence 1, 2, 3,..., 9. Therefore, the sequence of secant indices of X must contain exactly one gap, of size 2. The only remaining question is where in the sequence does this gap occur. In this case, the position of the gap is explained by the classical Cayley-Bacharach theorem [18].
2 Theorem 2.4.1. Let C1,C2 be two cubic curves in P not sharing any irreducible component and meeting in 9 distinct points. Suppose C3 is another cubic curve containing 8 of those 9 points.
Then C3 contains all 9 points.
Together with Theorem 2.2.12, this proves that L(X) = (1, 2, 3, 4, 5, 6, 7, 9). From this point of view, the content of the Cayley-Bacharach theorem is that the gap occurs between the final two terms of L(X). (n) n The correspondence in use here is that the hyperplane sections of vd (P ) are exactly the degree n d hypersurfaces in P . The maximum number of irreducible components such a hypersurface can (n) n have is d, and it is clear that the reducibility of vd (P ) is always d. G (n) n In what follows we give a method to compute RL (vd (P )) for all n, d > 1. We conjecture (n) n that this is in fact equal to L(vd (P )) in all cases. Some evidence in support of this is that the equality would follow from the truth of the currently open Eisenbud-Green-Harris conjecture, a conjecture which has received a large amount of attention in the last few decades admitting only partial progress [1] [12] [17] [28].
23 Our motivation in defining the sequences of greedy reducible secant indices becomes clear in G (n) n this context since the problem of computing RL (vd (P )) can be converted into a tractable combinatorial question. We illustrate the guiding heuristic to computing L(X) with the example (2) 2 X = v3 (P ).
9 Example 2.4.2. Consider two fully reducible hyperplane sections of X ⊆ P whose intersection 2 consists of exactly 9 distinct points. These hyperplane sections are curves in P (x : y : z), say
C1 = V ((x − z)(x − 2z)(x − 3z)) and C2 = V ((y − z)(y − 2z)(y − 3z)). We will compute RLG(X). The top term is 9, so then to find the next term we need to pick a 9 G hyperplane in P independent to the first two. The “greediness” of the sequence RL (X) is from the fact that we form it by finding hyperplane sections which remove the fewest points from the remaining finitely many points in the intersection at each step.
The number of irreducible components in the union C1 ∪ C2 is 6 = µ(X). To compute the next G term of RL (X) we must find an independent hyperplane section C3 which removes the fewest amount of points from C1 ∩ C2, but does not violate the µ(X) condition. That is, the way to interpret the defining constraint of the hyperplane sections used in the construction of RLG(X) is that they are those which do not introduce any additional irreducible components that meet the points we are working with; in other words, we must form C3 out of some choice of at most three distinct irreducible components from C1,C2.
These three components cannot all be from the same Cj, so it is clear that the best we can do is take, for example, C3 = V ((x − z)(x − 2z)(y − z)). Then C1 ∩ C2 ∩ C3 is reduced and consists of 7 points. Note also that so long as the number of points in the intersection decreases, the latest hyperplane section cannot be a linear combination of the previous. This process is continued to compute the remaining numbers. One possible continuation is illustrated in Figure 2.1. In Figure 2.1 the images are ordered left to right, top to bottom. The first image is of the 9 points in the intersection of C1 and C2. The curve C1 consists of the union of the three vertical lines, and the curve C2 of the three horizontal lines. These curves are all visualized by working 2 within the affine chart P \ V (z) and over the real numbers. The next image depicts C3 and the
7 points remaining after intersecting it with C1 ∩ C2. From there, the hyperplane sections chosen are V ((x − z)(x − 2z)z),V ((x − z)(y − z)(y − 2z)),V ((x − z)(y − z)z),V ((x − z)z2),V ((y − z)(y − 2z)z),V ((y − z)z2), in that order. Note the presence of the irreducible component V (z) in some of
24 (2) 2 Figure 2.1: Successive hyperplane sections of v3 (P ) these hyperplane sections, used to ensure the curve is of the correct degree. As V (z) does not meet
C1 ∩ C2, these additional hyperplane sections satisfy the conditions of d-reducibility.
This example was too simple to illustrate some questions that need to be resolved when com- puting RLG(·) in more complicated situations. One is, at a given step of the computation, if there are multiple possible new hyperplane sections to use, which each take away the same number of points, does it matter which one is used? Additionally, how can one determine whether the greedy sequence matches the sequence RL(·)? For Veronese varieties, answering these questions becomes easier if we first move the compu- tations to a simpler, more combinatorial setting. The heuristic guiding the transformation in this section is that the number of points we are getting after adding hyperplane sections is the same if we were to “collapse” the distinct irreducible components into one irreducible component, counted with a multiplicity. 3 So in the above example, for the purposes of the computation, we would collapse C1 into V (x ), 3 2 C2 into V (y ), and C3 into V (x y). The scheme-theoretic degree of their intersection is
3 2 2 deg(C1 ∩ C2 ∩ C3) = deg(V (x , y , x y)) = 7.
Therefore, a combinatorial problem we could consider is the following.
Combinatorial problem 2.4.3. Determine the maximal dimension of k[x0, . . . , xn−1]/I where I d d is the dehomogenization with respect to xn of a monomial ideal (x0, . . . , xn−1) + J in k[x0, . . . , xn] generated by r linearly independent degree d monomials, for each r ≥ n.
25 Example 2.4.4. We address this new problem for n = 2, d = 4. When r = 2, we obtain the 4 4 vector space k[x0, x1]/(x0, x1), of dimension 16. We see that the sequence of maximal dimensions is (1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 12, 13, 16), organized from r = 12 to r = 2. Figure 2.2 illustrates how this sequence can be obtained by sequentially adding monomials to the ideal J using a staircase diagram.
4 4 4 x1 x1 x1
3 3 3 x1 x1 x1
2 2 2 x1 x1 x1
x1 x1 x1
x0 2 3 4 x0 2 3 4 x0 2 3 4 x0 x0 x0 x0 x0 x0 x0 x0 x0
4 4 4 x1 x1 x1
3 3 3 x1 x1 x1
2 2 2 x1 x1 x1
x1 x1 x1
x0 2 3 4 x0 2 3 4 x0 2 3 4 x0 x0 x0 x0 x0 x0 x0 x0 x0
Figure 2.2: Successive monomial additions to J for n = 2, d = 4
This figure is read from left to right, top to bottom like Figure 2.1. The first image is that of the 4 4 16 monomials representing the basis elements of k[x0, x1]/(x0, x1). These monomials correspond to points in the diagram in accordance to their exponent vectors. Adding a monomial to J amounts to removing that monomial and all others it divides from the diagram; the number of monomials leftover is the dimension of the new polynomial ring quotient. We may only add monomials to J that are on or below the pictured diagonal line, as these are exactly those of degree ≤ 4. This setup leads to a “game” wherein one tries to remove as few monomials as possible from the diagram with each successive addition to J. The pictured sequence is part of the sequence of additions 3 3 2 2 2 2 3 2 3 2 2 x0x1, x0, x0x1, x0x1, x0, x0x1, x0x1, x0x1, x0, x1, x1, x1. The additions past x0 are not pictured as there are no more gaps larger than 1 that appear in the sequences of dimensions. Note that this sequence of monomials is part of the lexicographic sequence of degree 4 monomials in x0, x1, x2, but dehomogenized with respect to x2. A priori, this lexicographic sequence only gives a “greedy” sequence of removals, removing as few monomials from the diagram at each step as possible. This example is simple enough to check by brute force that this greedy sequence is in fact the term-wise
26 maximal sequence. That this is true for the more complicated examples when n or d is larger is more difficult.
G (n) n We will show that this combinatorial problem is equivalent to computing both RL (vd (P )), (n) n (2) 2 RL(vd (P )). In particular, in this example, we computed RL(v4 (P )). The two gaps observed in the sequence can be explained by a modern formulation of the Cayley-Bacharach theorem [17]:
r Theorem 2.4.5. Let Γ ⊆ P be a complete intersection of hypersurfaces X1,...,Xr of degrees 0 00 d1, . . . , dr, and let Γ , Γ ⊆ Γ be closed subschemes residual to one another. Set X m = −r − 1 + .
di Then for any ` ≥ 0, we have
0 r 0 r 1 r h (P , IΓ0 (`)) − h (P , IΓ(`)) = h (P , IΓ00 (m − `)).
As n and d grow, enumerating the possibilities for the dimensions of the involved cohomology groups becomes difficult, and so it does not seem feasible to use this result as a way of computing (n) n gaps in the sequences L(vd (P )). G (n) n (n) n We now show that the problem of computing RL (vd (P )), RL(vd (P )) is equivalent to computing the maximal dimensions of the polynomial ring quotients above. First we show that any dimension obtained there can be realized as the number of points left in the intersection of n hypersurfaces in P that are each the union of d hyperplanes. Qd Qd We know that the ideal I = ( i=1(x0 −ixn),..., i=1(xn−1 −ixn)) defines a complete intersec- n n tion in P . The support of this intersection is the collection of d distinct points {(a0 : ... : an−1 :
1) | ai ∈ {1, . . . , d}}, and so the complete intersection must be reduced. Adding any equation to n the set of generators of the ideal I gives an ideal which cuts out a reduced subscheme of P . Next, observe there is a divisibility-preserving bijection
R φn :Ωn → Ωn
R between Ωn and the set of monomials Ωn in k[x0, . . . , xn−1], defined
n−1 n−1 di Y di Y Y xi 7→ (xi − j), i=0 i=0 j=1 that is, for a, b ∈ Ωn, a divides b if and only if φn(a) divides φn(b). With this notation, we have the following key fact.
27 d d Lemma 2.4.6. Let I = (x0, . . . , xn−1, m1, . . . , mr) be a monomial ideal such that deg(mj) ≤ d for each j. Then
dimk k[x0, . . . , xn−1]/I
d d = dimk k[x0, . . . , xn−1]/(φn(x0), . . . , φn(xn−1), φn(m1), . . . , φn(mr)), as finite-dimensional vector spaces.
Proof. The dimension dimk k[x0, . . . , xn−1]/(φn(m1), . . . , φn(mr)) equals the degree of the reduced d d scheme V (φn(x0), . . . , φn(xn−1), φn(m1), . . . , φn(mr)). d d n We know that the support of the reduced scheme Y := V (φn(x0), . . . , φn(xn−1)) ⊆ A is the collection of dn distinct points
n {(a0, . . . , an−1) | ai ∈ {1, . . . , d}} ⊆ A
Qn−1 Qdi as discussed above. The effect of intersecting Y with a hypersurface of the form V ( i=0 j=1(xi − j)) is to remove all points (a0, . . . , an−1) of Y that satisfy ai > di for each i.
Qn−1 ai On the other hand, each point (a0, . . . , an−1) corresponds to a monomial i=0 xi . The quotient d d ring k[x0, . . . , xn−1]/(x0, . . . , xn−1) has as a basis the monomials corresponding to the points of Y .
Qn−1 Qdi The effect of intersecting Y with the hypersurface V ( i=0 j=1(xi − j)) is analogous to forming the quotient ring n−1 d d Y di k[x0, . . . , xn−1]/(x0, . . . , xn−1, xi ); i=0 the number of points left in Y is equal to the dimension of this new quotient ring. By repeating this process r times, we obtain the desired result.
Note the similarity between the patterns of Figures 1 and 2, accounting for the difference in dimension. Making rigorous this similarity is all that Lemma 2.4.6 is serving to do. The quotient (n) n ring problem leads to a clean proof that µ(vd (P )) = nd.
(n) n Lemma 2.4.7. µ(vd (P )) = nd.
(n) Proof. In the polynomial ring k[x0, . . . , xn−1] there are Nd monomials of degree ≤ d. d d n The quotient ring k[x0, . . . , xn−1]/(x0, . . . , xn−1) has dimension d as a k-vector space, and has d d as a basis the monomials from k[x0, . . . , xn−1] of degree ≤ d other than the pure powers x0, . . . , xn−1.
28 (n) It suffices to choose a sequence (aj)j of length Nd − n − 1 of these monomials so that aj does not divide ai for every i > j. This condition ensures that
d d dimk(k[x0, . . . , xn−1]/((x0, . . . , xn−1) + (a1, . . . , aj)))
d d > dimk(k[x0, . . . , xn−1]/((x0, . . . , xn−1) + (a1, . . . , aj+1))) for each j. N (n)−1 By Lemma 2.4.6, this sequence corresponds to a sequence of hyperplanes (H1,...,Hn) of P d so that d (n) n Y Hj ∩ vd (P ) = V ( (xi − ix0)) i=1
(n) n for each j = 1, . . . , n, and every other Hj ∩ vd (P ) for j > n is formed as the union of irreducible components of the preceding hyperplane sections, possibly along with the extraneous component (n) n V (xn) which does not meet H1 ∩ ... ∩ Hn ∩ vd (P ). Thus the union of these hyperplane sections has exactly nd irreducible components. Further, these hyperplanes all satisfy the requirements of d-reducibility.
G (n) n (n) n To compute RL (vd (P )), RL(vd (P )) we note a symmetry that occurs in these computa- tions.
(n) (n) n Nd −1 0 R Lemma 2.4.8. Let X = vd (P ) ⊆ P for n, d > 0. Then for any L, L ∈ Λcodim(X)(X), then for any sequence (H1,...,H (n) ) ∈ H(X) such that L = H1 ∩ ... ∩ Hn, there is a sequence Nd −1 0 0 0 0 0 (H1,...,H (n) ) ∈ H(X) with L = H1 ∩ ... ∩ Hn and Nd −1
0 0 |X ∩ (H1 ∩ ... ∩ Hj)| = |X ∩ (H1 ∩ ... ∩ Hj)| for each j ≥ n.
Proof. The key observation here is that since the reducibility of X is d and µ(X) = nd, each (j) (j) n hyperplane section Hj ∩ X for j = 1, . . . , n is a union of d distinct hyperplanes T1 ,...,Td in P .
The rest of the hyperplane sections Hj ∩ X are unions of the hyperplanes forming the preceding hyperplane sections. The intersection L ∩ X is finite and reduced, consisting of dn points. These points are exactly
Tn (jk) those of the form k=1 Tk where each 1 ≤ jk ≤ d.
29 0 n We may realize L ∩ X in the same way as an intersection of unions of hyperplanes in P . The formation of new hyperplane sections from unions of these constituent hyperplanes for the purpose of cutting away points from L ∩ X or L0 ∩ X can be done in the same way for both L and L0.
G (n) n (n) n Because of this symmetry, for the purposes of computing RL (vd (P )), RL(vd (P )) we may consider only sequences (H1,...,H (n) ) ∈ H(X) with Nd −1
d Y n Hj ∩ X = V ( (xj−1 − ixn)) ⊆ P , i=1 for j = 1, . . . , n. For any j ≥ n, we have that
|X ∩ (H1 ∩ ... ∩ Hj)| = deg(X ∩ (H1 ∩ ... ∩ Hj)) = dimk(k[x0, . . . , xn−1]/I) where I is the homogeneous ideal of X ∩ (H1 ∩ ... ∩ Hj) dehomogenized with respect to xn. By Lemma 2.4.6, we achieve our objective:
Lemma 2.4.9. We have
1.
G (n) n RL (vd (P )) d d = max{(dimk(k[x0, . . . , xn−1]/(x0, . . . , xn−1, a1, . . . , a (n) )), Nd −n−1 d d . . . , k[x0, . . . , xn−1]/(x0, . . . , xn−1))},
2.
(n) n RL (n) (vd (P )) Nd −n−1−j d d = max{dimk(k[x0, . . . , xn−1]/(x0, . . . , xn−1, a1, . . . , aj))},
d d where both maximums are taken over all monomials ak of degree ≤ d and so that x0, . . . , xn−1, a1, . . . , ak are all linearly independent, and the first maximum is done using the total order ≤ on integer sequences.
The question of whether these two sequences of secant indices are equal is then equivalent to asking whether for the integer sequences over which the maximum is taken in the expression for
30 G (n) n RL (vd (P )) above, is a maximal element with respect to the total order ≤ also a maximal element with respect to the partial order ? Our reduction of the problem to the quotient ring setup helps with the visualization of the combinatorics behind this question. In the n = 2 case, as in Example 2.4.4, the diagram is simple enough that the expectation that these two sequences are the same seems reasonable. But this is a fact which requires a careful proof. We can think of the combinatorial problem of computing the maximal dimensions of these quotient rings as a specific instance of a more general family of problems. Consider a poset which is decomposed into the disjoint union of two finite sets C = A ∪ B. Denote by ≤ the partial order on C and let m = |A|. The problem is to consider all sequences of length m−1 obtained by picking an element of A, and then removing it and all elements greater than it according to ≤ from C, subject to the condition that the chosen element does not remove any further elements from A.
That is, one considers all sequences (a1, . . . , am−1) of distinct elements of A such that aj 6≤ ai for all i > j, and then considers the sequence of cardinalities obtained by removing a1, . . . , am−1 one at a time, in that order, along with all elements larger than each. In our computation of the secant indices, C is the set of monomials that represent the generators d d of the quotient k[x0, . . . , xn−1]/(x0, . . . , xn−1), and the set A is that consisting of all monomials of d d degree ≤ d aside from the pure powers x0, . . . , xn−1 in k[x1, . . . , xd]. It is simple to construct examples of such posets in general where the greedy sequence is not maximal with respect to .
Example 2.4.10. Consider the poset {1, a, b, b2, c, a2, a2, a3, a4, a5}, where α ≤ β if and only if α divides β. Let A = {1, a, b, c, a2, a3}, and let B = {b2, a4, a5}. Choosing the elements of A in the order c, b, a3, a2, a yields the “greedy” sequence (9, 8, 6, 3, 2, 1). Whereas removing the elements in the order a3, a2, a, c, b yields the sequence (9, 6, 5, 4, 3, 1), which surpasses the greedy sequence in the fourth term.
This poset seems uncomfortably close to the setup we work with to compute the reducible secant indices. However, no such discrepancy arises in our computations due to the following consequence [45, Proposition 3.12] of the Clements-Lindstr¨omtheorem.
Lemma 2.4.11. Let R = {f1, . . . , fr} ⊆ k[x0, . . . , xn−1] be a regular sequence of monomials, with degrees ej = deg(fj), e1 ≤ ... ≤ er. Let N be any homogeneous ideal containing R. Then there
31 e1 er exists a lex ideal L ⊆ k[x0, . . . , xn−1] such that N and (x0 , . . . , xn−1) + L have the same Hilbert series.
By a lex ideal, we mean a monomial ideal I of k[x0, . . . , xn−1] such that the degree d piece of
I, Id, is generated by an initial segment of the lexicographic sequence of degree d monomials, for every d. This solves the quotient ring dimension problem for us because of the following.
(n) Lemma 2.4.12. Let a1, . . . , ar, 0 ≤ r ≤ Nd − n − 1 be any sequence of monomials of degree d d d in k[x0, . . . , xn] so that x0, . . . , xn−1, a1, . . . , ar are all linearly independent. Let b1, . . . , br be the initial segment of the lexicographic sequence of degree d monomials in k[x0, . . . , xn], excluding the d d d d 0 0 d d pure powers x0, . . . , xn. Then letting I = (x0, . . . , xn−1) + (a1, . . . , ar) and J = (x0, . . . , xn−1)) +
(b1, . . . , br), we have
0 0 dimk(k[x0, . . . , xn−1]/I ) ≤ dimk(k[x0, . . . , xn−1]/J ),
0 0 where I ,J denote I,J dehomogenized with respect to xn, respectively.
d d Proof. By Lemma 2.4.11, since x0, . . . , xn−1 is a regular sequence, we know that there is a lex ideal d d L such that I has the same Hilbert series as the ideal (x0, . . . , xn−1) + L. This Hilbert series is d d smaller at each term than that of the ideal (x0, . . . , xn−1) + Ld. Since in particular the degree d d d parts of the ideals I and (x0, . . . , xn−1) + L must have the same dimensions as k-vector spaces, we know that Ld is generated by the initial segment of the lexicographic sequence of degree d monomials, excluding the pure powers, of length r.
Denoting by b1, . . . , br these monomials, in lexicographic order, we arrive at the desired result.
This lemma proves that the two sequences of reducible secant indices and greedy indices are the same, and together with our previous observations in this section, shows their terms are identical to those that arise from the poset problem associated to the polynomial ring quotients above.
(n) n Theorem 2.4.13. Let X = vd (P ), n, d > 1. Then
RL(X) = RLG(X),
(n) n (n,d) (n,d) and RL (n) (vd (P )) = dimk(k[x1, . . . , xn])/Ij where the ideal Ij is the dehomoge- Nd −1−n−j d d nization with respect to xn of the sum of the ideal (x0, . . . , xn−1) with the ideal generated by the
32 (n) lexicographic sequence of degree d monomials excluding pure powers of length Nd − n − j in k[x0, . . . , xn].
It would be interesting to have a simple formula that produces these sequences. For n = 2 or for d = 2, the corresponding sequences as the other number varies follow simple patterns. But as n, d both increase these patterns become increasingly complex, as we will soon illustrate with several examples. Our expectation is that these two sequences are in fact also equal to L(X). We leave this as a conjecture.
Conjecture 2.4.14. For each n, d > 1,
RLG(X) = RL(X) = L(X),
(n) n where X = vd (P ). In other words, the maximal number of points that can be contained in the intersection of r linearly independent degree d hypersurfaces in n is RLG (X), given that the P Nd−r−1 intersection is finite and reduced.
If true, Conjecture 2.4.14 could be thought of a quasi-enumerative version of B´ezout’stheorem. There is some evidence for it from the so-called Eisenbud-Green-Harris conjecture.
Conjecture 2.4.15 (EGH [17]). Let I be a homogeneous ideal in the polynomial ring k[x0, . . . , xn−1] containing a regular sequence f1, . . . , fn of degrees deg(fi) = ai, where 2 ≤ a1 ≤ ... ≤ an. Then I
a1 an has the same Hilbert function as an ideal containing x0 , . . . , xn .
Proposition 2.4.16. The truth of the EGH-conjecture would imply that of Conjecture 2.4.14.
n Proof. Suppose we have any complete intersection of degree d hypersurfaces T1,...,Tn in P , and let Tn+1,...,Tn+r be degree d hypersurfaces so that T1,...,Tn+r are linearly independent and
T1 ∩ ... ∩ Tn+r is reduced. Write Tj = V (Fj), Fj ∈ k[x0, . . . , xn] for each j. So the F1,...,Fn form a regular sequence. Suppose the EGH-conjecture is true. Then by [12, Proposition 9] and Lemma 2.4.11 see that d d the ideal (F1,...,Fn+r) has the same Hilbert series as an ideal of the form (x0, . . . , xn−1)+L, where L is a lex ideal. The degree d part of L must then be generated by the lexicographic sequence of monomials of degree d excluding the pure powers and of length r. Thus
dimk(k[x0, . . . , xn−1]/I) ≤ dimk(k[x0, . . . , xn−1]/J),
33 d d where I,J are the ideals (F1,...,Fn+r), (x0, . . . , xn−1)+Ld dehomogenized with respect to xn.
We conclude this section with several computations of the greedy sequence of secant indices and G (n) n a specific instance of Conjecture 2.4.14. For a given n, d, computing RL (vd (P )) can be done easily using a computer algebra system such as SageMath [52].
(n) n Example 2.4.17. Let X = vd (P ). n = 2
• d = 2, RLG(X) = (1, 2, 3, 4),
• d = 3, RLG(X) = (1, 2, 3, 4, 5, 6, 7, 9),
• d = 4, RLG(X) = (1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 12, 13, 16),
• d = 5, RLG(X) = (1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 15, 16, 17, 20, 21, 25),
• d = 6,
RLG(X) =(1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 18, 19, 20, 21, 24, 25, 26, 30, 31, 36),
• d = 7,
RLG(X) =(1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 21, 22, 23, 24, 25, 28, 29, 30, 31, 35, 36, 37, 42, 43, 49).
n = 3
• d = 2, RLG(X) = (1, 2, 3, 4, 5, 6, 8),
• d = 3, RLG(X) = (1, 2, 3, 4, 5, 6, 7, 9, 10, 11, 12, 13, 15, 18, 19, 21, 27),
• d = 4,
RLG(X) =(1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 12, 13, 16, 17, 18, 19, 20, 21, 22, 24, 25, 28, 32, 33, 34, 36, 37, 40, 48, 49, 52, 64),
n = 4
• d = 2, RLG(X) = (1, 2, 3, 4, 5, 6, 8, 9, 10, 12, 16),
34 • d = 3,
RLG(X) =(1, 2, 3, 4, 5, 6, 7, 9, 10, 11, 12, 13, 15, 18, 19, 21, 27, 28, 29, 30, 31, 33, 36, 37, 39, 45, 54, 55, 57, 63, 81),
n = 5
• d = 2, RLG(X) = (1, 2, 3, 4, 5, 6, 8, 9, 10, 12, 16, 17, 18, 20, 24, 32),
• d = 3,
RLG(X) =(1, 2, 3, 4, 5, 6, 7, 9, 10, 11, 12, 13, 15, 18, 19, 21, 27, 28, 29, 30, 31, 33, 36, 37, 39, 45, 54, 55, 57, 63, 81, 82, 83, 84, 85, 87, 90, 91, 93, 99, 108, 109, 111, 117, 135, 162, 163, 165, 171, 189, 243),
Note when d = 2, the gaps in these sequences are similar to those observed in [17, pg. 193].
There they consider the maximal possible dimensions of quotients k[x0, . . . , xn−1]/I where I is a homogeneous ideal generated by r linearly independent quadrics. Because lower degree polyno- mials are not allowed as generators, their sequences are shorter. This difference becomes more pronounced if one considers the same numbers for ideals generated by linearly independent degree d homogeneous polynomials rather than allowing lower degree generators as we do here. G (n) n When n = 2, the pattern determining RL (vd (P )) is straightforward and yields a particularly attractive incarnation of Conjecture 2.4.14. It is the continuation of the patterns observed in the above sequences for n = 2.
35 (2) Conjecture 2.4.18. Fix d > 0 and set N = Nd . Consider the sequence (a1, . . . , aN−3) with terms (organized first to last)
d − 1
1
d − 2
1, 1
...
4
1,..., 1, (repeated d − 4 times)
3
1,..., 1, (repeated d − 3 times)
2
1,..., 1, (repeated 3d − 3 times)
Let (b1, . . . , bN−2) be the sequence defined by
j−1 n X bj = d − ai. i=1 Then the maximal number of points that could be contained in the intersection of r linearly inde- pendent degree d curves is br−1, given that the intersection is finite and reduced.
Note it is also an implication of the EGH-conjecture, so any counterexample would also suffice to disprove the EGH-conjecture.
2.5 Segre varieties
In this section we will devise a method for computing the greedy reducible secant indices for n m N the images of Segre embeddings. Let n, m > 0 and throughout denote by σn,m : P × P ,→ P the n m N Segre embedding, where N = (n + 1)(m + 1) − 1, and let X = σn,m(P × P ) ⊆ P . The degree n+m of X is n , so this is the top term of L(X).
36 Example 2.5.1. When m = 1 is fixed and n is allowed to vary (or vice versa), X has dimension N n + 1 inside P where N = 2(n + 1) − 1 = 2n + 1. Thus X has codimension n and its degree is n+1 N n = n + 1. Therefore X is a minimal degree subvariety of P and so by Theorem 2.2.12 its sequence of secant indices is L(X) = (1, 2, . . . , n + 1).
Therefore the question of computing the sequences of secant indices only becomes interesting 2 2 8 for n, m > 1. The first such example is P × P which is a codimension 4 subvariety of P of degree 6. So its sequence of secant indices must contain exactly one gap of size 2. In this section we will describe a method which can be used to compute RLG(X). These computations show that
G 2 2 RL (σ2,2(P × P )) = (1, 2, 3, 4, 6),
2 2 suggesting that the gap in the true sequence L(σ2,2(P × P )) occurs between the final two terms, 2 9 like for the example of the degree 3 Veronese embedding of P into P . One can verify that in fact
2 2 L(σ2,2(P × P )) = (1, 2, 3, 4, 6).
The example is simple enough that we can conclude by Theorem 2.2.12 and Proposition 2.3.9, as 2 2 σ2,2(P × P ) is smooth and nondegenerate, and is not a rational scroll.
Proposition 2.5.2. 2 2 L(σ2,2(P × P )) = (1, 2, 3, 4, 6).
(2) 2 Proposition 2.3.9 could also be used to prove L(v3 (P )) = (1, 2, 3, 4, 5, 6, 7, 9) in place of the classical Cayley-Bacharach theorem. In this sense, this classification result, and results such as that of Noma [48], fulfill similar roles to the Cayley-Bacharach theorem but in the case of arbi- trary smooth nondegenerate projective subvarieties. However, they only concern extremal secant subspaces and do not provide lower bounds for the secant indices, so do give us a means to answer whether L(X) = RLG(X) for the more complicated instances of Segre and Veronese varieties.
37 Because of this, we focus here on computing instead RLG(X) to obtain conjectural values for n m n L(X), for X = σn,m(P × P ). To derive this method, note that if we choose coordinates, P (x0 : m N ... : xn), P (y0 : ... : ym), P (z0 : ... : zN ) then for any hyperplane H = V (a0z0 + ... + aN zN ) of N P , we see that X ∩ H is the zero locus of the polynomial a0x0y0 + a1x0y1 + ... + aN xnym. This quadric can split into at most two factors. On the other hand, we can choose any collection of n + m hyperplanes
n H1,...,Hn+m ⊆ P with the property that any subset of n of these hyperplanes has only a single point in common, and any n + 1 do not have any point in common. Likewise we can choose an analogous collection 0 0 m of hyperplanes H1,...,Hn+m ⊆ P with the property that any subset of m of those hyperplanes meet at a single point and no m + 1 have a point in common. Each subscheme
0 n m Tj := Hj ∪ Hj ⊆ P × P
n m is then a reducible and reduced hyperplane section of P × P via the Segre embedding, and we Tn+m n+m n m have j=1 Tj is finite and reduced, realizing the degree n of P ×P in cardinality. Any n+m N hyperplanes of P meeting at a dimension N − n − m-dimension linear subvariety that meets X in a finite and reduced collection of points will be of this form. As in Section 2.4, there is symmetry here in the sense that the points of the intersection
T1 ∩ ... ∩ Tn+m are exactly those of the form
0 0 (Ha1 ∩ ... ∩ Han ) × (Hb1 ∩ ... ∩ Hbm ) where {1, 2, . . . , n + m} = {a1, . . . , an} ∪ {b1, . . . , bm}. Therefore, we can strip the essence of the computation of RLG(X) from the context of the Segre embedding and find it equivalent to the following problem.
Combinatorial problem 2.5.3. Let Y be the set of all tuples (A, B) where A is a set of size n and B a set of size m so that A ∪ B = {1, 2, . . . , n + m}. Let S be the set of all sequences of length nm consisting of distinct tuples (a, b) where a, b ∈ {1, 2, . . . , n + m} and a 6= b. Given a tuple (a, b), we say we are cutting Y by (a, b) if we replace Y with the subset of elements (A, B) of
Y for which either a ∈ A or b ∈ B. For each sequence ((a1, b1),..., (anm, bnm)), form an integer
38 sequence (c0, . . . , cnm) where cj is the cardinality of the set obtained by cutting Y by each (ai, bi) for i = 1, . . . , nm − j. Note cnm = |Y |. The problem is then to compute the maximal (with respect to the total order ≤ of Section 2.3) possible integer sequence arising in this manner that is also strictly increasing.
The combinatorial problem gives us a clearer way to compute the reducibility of X and find µ(X).
n m Proposition 2.5.4. Let X = σn,m(P ×P ). Then the reducibility of X is 2, and µ(X) = 2 dim(X).
Proof. Because of our observations above, we see that the most a hyperplane section of X can split is into two components, so the reducibility of X is at most 2. All that must be done is exhibit a sequence of hyperplane sections satisfying the conditions of 2-reducibility that also has the minimum total number of irreducible components. To help simplify the notation, we use the notation of Combinatorial problem 2.5.3. Let p =
(A, B) be any point of Y , and write A = {a1, . . . , an}, B = {b1, . . . , bm}. Then consider the set T consisting of the nm points obtained from swapping one element of A with one element of B.
Use pij to denote the point where ai was swapped with bj. For each such i, j, cutting by the tuple
Hi,j = (ai, bj) removes pij from Y but does not remove any of the other elements of T .
Each Hi,j corresponds to a hyperplane section of X, and the sequence
H1,1,H2,1,...,Hn,m, taken in any order, is a sequence of hyperplane sections satisfying the conditions of 2-reducibility, with 2 dim(X) total irreducible components.
Thus altogether we have the following.
Theorem 2.5.5. For a given n, m > 1, the answer to Combinatorial problem 2.5.3 is the sequence RLG(X).
It is straightforward to write an algorithm that solves this problem for a given n, m > 1. We have implemented such an algorithm using the SageMath computer algebra system [52], and have used it to compute the greedy sequence of reducible secant indices for several values of n, m. It
39 would be interesting to know if there is an analog of Lemma 2.4.11 that would work in this context to show that RLG(X) = RL(X).
Verifying this without additional theoretical support requires a brute-force check of every possible sequence of elements of the form (A, B) using the notation of Combinatorial problem 2.5.3, and this becomes impractical for even small n, m.
Example 2.5.6. m = n • n = 2, m = 2, RLG(X) = (1, 2, 3, 4, 6),
• n = 3, m = 3, RLG(X) = (1, 2, 3, 4, 6, 7, 10, 11, 14, 20),
• n = 4, m = 4, RLG(X) = (1, 2, 3, 4, 6, 7, 10, 11, 14, 20, 21, 25, 35, 36, 40, 50, 70),
m = n − 1 • n = 3, m = 2, RLG(X) = (1, 2, 3, 4, 6, 7, 10),
• n = 4, m = 3, RLG(X) = (1, 2, 3, 4, 6, 7, 10, 11, 14, 20, 21, 25, 35),
m = n − 2 • n = 4, m = 2, RLG(X) = (1, 2, 3, 4, 6, 7, 10, 11, 15),
• n = 5, m = 3, RLG(X) = (1, 2, 3, 4, 6, 7, 10, 11, 14, 20, 21, 25, 35, 36, 41, 56),
m = n − 3 • n = 5, m = 2, RLG(X) = (1, 2, 3, 4, 6, 7, 10, 11, 15, 16, 21).
It seems reasonable to expect that for this problem, like for that of the Veronese varieties, the greedy sequence is equal to the sequence of secant indices. We leave this as a conjecture.
n m Conjecture 2.5.7. For any n, m > 1, X = σn,m(P × P ), we have
RLG(X) = L(X).
While it may be difficult to write down a closed form formula for RLG(X), one simple compu- tation is determining the size of the last gap in RLG(X). A special case of the above conjecture is then that the same gap must be present in the sequence of secant indices.
Conjecture 2.5.8. n + m − 2 L (X) − L (X) = . nm nm−1 n − 1
40 2.6 Lines on surfaces
Our original motivation comes from the classical problem of determining the maximal number 3 m`(d) of lines that can be contained in a degree d smooth surface S of P (x : y : z : w). As all surfaces of degree d ≤ 2 are ruled, this question is only relevant for surfaces of degree d ≥ 3. When d = 3, the Cayley-Salmon theorem ensures that every such S must contain exactly 27 distinct lines. However, for d ≥ 4, the general degree d surface contains no lines at all; the problem of determining m`(d) is a so-called quasi-enumerative problem [20].
Over a century ago, Clebsch [13] formulated the bound m`(d) ≤ d(11d − 24) and Segre [51] later proved m`(d) ≤ (d − 2)(11d − 6), for all d ≥ 3. These bounds have been improved slightly in modern times and the methods of Segre made rigorous using modern intersection theory. Bauer and Rams [9, Theorem 1.1] have proven
2 m`(d) ≤ 11d − 32d + 24 for all d ≥ 3. To our knowledge, this is the best known bound for m`(d) when d ≥ 6. However, this latest bound is still known not to be sharp. It fails to be so already for the case of smooth quartic surfaces, where it is known that any irreducible quartic surface not ruled by lines contains at most 64 lines, see for instance [26, Theorem 4.5]. Furthermore, quartics achieving this bound exist. To find a lower bound for m`(d) one only needs to provide an example of a surface with lines. For general d, the best known example of a surface with many lines is the smooth degree d Fermat surface V (xd + yd + zd + wd), which contains exactly 3d2 distinct lines.
The first of the numbers m`(d) which is currently unknown is that for d = 5. By [50, Theorem
1.2], together with the example of the Fermat quintic, we see that 75 ≤ m`(5) ≤ 127. To our knowledge, no example is currently known of a smooth quintic surface with more than 75 lines.
Interestingly, because m`(d) is bounded between two quadratic polynomials, if it turns out that m`(5) ≤ 101, then by interpolation using the examples m`(3) = 27, m`(4) = 64, there can be no polynomial function f(d) that agrees with m`(d) for all d ≥ 3. If this were true, then it would imply that any method which can only produce polynomial bounds is doomed to fail to produce an exact formula for m`(d). Our work presented in this chapter began by looking for an alternate description of the numbers m`(d). In what follows, we relax the constraint on the surfaces we consider to allow for nonreduced,
41 reducible, and singular surfaces. Let m`(d) now denote the maximal possible number of lines that 3 can be contained in a degree d surface of P , given that the surface contains only finitely many lines. There is a generalization of the Veronese embedding for Grassmannians [30]; in particular, given a d > 1, one may define an embedding
d + 3 d + 3 v : (1, 3) = G(2, 4) ,→ G − d − 1, d G d d by
L 7→ I(L)d,
3 for each line L ⊆ P , where I(L)d denotes the degree d part of the homogeneous ideal I(L) of L. Given a nonzero homogeneous polynomial F ∈ k[x, y, z, w] of degree d, one can consider the d+3 d+3 d+3 subset of G( d − d − 1, d ) consisting of all d − d − 1-planes of k[x, y, z, w]d containing d+3 d+3 F . This is a special type of Schubert subvariety of G( d − d − 1, d ), a sub-Grassmannian, d+3 d+3 isomorphic to G( d − d − 2, d − 1). Then note that we can express m`(d) as the maximal
finite intersection that can occur by intersecting vd(G(1, 3)) such sub-Grassmannians,
m`(d) = max{|vd(G(1, 3)) ∩ G| | |vd(G(1, 3))| < ∞},